linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* [PATCH v7 00/15] Introduce Data Access MONitor (DAMON)
@ 2020-03-18 11:27 SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 01/15] scripts/spelling: Add a few more typos SeongJae Park
                   ` (15 more replies)
  0 siblings, 16 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

Introduction
============

Memory management decisions can be improved if finer data access information is
available.  However, because such finer information usually comes with higher
overhead, most systems including Linux forgives the potential benefit and rely
on only coarse information or some light-weight heuristics.  The pseudo-LRU and
the aggressive THP promotions are such examples.

A number of data access pattern awared memory management optimizations (refer
to 'Appendix A' for more details) consistently say the potential benefit is not
small.  However, none of those has successfully merged to the mainline Linux
kernel mainly due to the absence of a scalable and efficient data access
monitoring mechanism.  Refer to 'Appendix B' to see the limitations of existing
memory monitoring mechanisms.

DAMON is a data access monitoring subsystem for the problem.  It is 1) accurate
enough to be used for the DRAM level memory management (a straightforward
DAMON-based optimization achieved up to 2.55x speedup), 2) light-weight enough
to be applied online (compared to a straightforward access monitoring scheme,
DAMON is up to 94,242.42x lighter) and 3) keeps predefined upper-bound overhead
regardless of the size of target workloads (thus scalable).  Refer to 'Appendix
C' if you interested in how it is possible, and 'Appendix F' to know how the
numbers collected.

DAMON has mainly designed for the kernel's memory management mechanisms.
However, because it is implemented as a standalone kernel module and provides
several interfaces, it can be used by a wide range of users including kernel
space programs, user space programs, programmers, and administrators.  DAMON
is now supporting the monitoring only, but it will also provide simple and
convenient data access pattern awared memory managements by itself.  Refer to
'Appendix D' for more detailed expected usages of DAMON.


Visualized Outputs of DAMON
===========================

For intuitively understanding of DAMON, I made web pages[1-8] showing the
visualized dynamic data access pattern of various realistic workloads, which I
picked up from PARSEC3 and SPLASH-2X bechmark suites.  The figures are
generated using the user space tool in 10th patch of this patchset.

There are pages showing the heatmap format dynamic access pattern of each
workload for heap area[1], mmap()-ed area[2], and stack[3] area.  I splitted
the entire address space to the three area because there are huge unmapped
regions between the areas.

You can also show how the dynamic working set size of each workload is
distributed[4], and how it is chronologically changing[5].

The most important characteristic of DAMON is its promise of the upperbound of
the monitoring overhead.  To show whether DAMON keeps the promise well, I
visualized the number of monitoring operations required for each 5
milliseconds, which is configured to not exceed 1000.  You can show the
distribution of the numbers[6] and how it changes chronologically[7].

[1] https://damonitor.github.io/reports/latest/by_image/heatmap.0.png.html
[2] https://damonitor.github.io/reports/latest/by_image/heatmap.1.png.html
[3] https://damonitor.github.io/reports/latest/by_image/heatmap.2.png.html
[4] https://damonitor.github.io/reports/latest/by_image/wss_sz.png.html
[5] https://damonitor.github.io/reports/latest/by_image/wss_time.png.html
[6] https://damonitor.github.io/reports/latest/by_image/nr_regions_sz.png.html
[7] https://damonitor.github.io/reports/latest/by_image/nr_regions_time.png.html


Data Access Monitoring-based Operation Schemes
==============================================

As 'Appendix D' describes, DAMON can be used for data access monitoring-based
operation schemes (DAMOS).  RFC patchsets for DAMOS are already available
(https://lore.kernel.org/linux-mm/20200218085309.18346-1-sjpark@amazon.com/).

By applying a very simple scheme for THP promotion/demotion with a latest
version of the patchset (not posted yet), DAMON achieved 18x lower memory space
overhead compared to THP while preserving about 50% of the THP performance
benefit with SPLASH-2X benchmark suite.

The detailed setup and number will be posted soon with the next RFC patchset
for DAMOS.  The posting is currently scheduled for tomorrow.


Frequently Asked Questions
==========================

Q: Why DAMON is not integrated with perf?
A: From the perspective of perf like profilers, DAMON can be thought of as a
data source in kernel, like the tracepoints, the pressure stall information
(psi), or the idle page tracking.  Thus, it is easy to integrate DAMON with the
profilers.  However, this patchset doesn't provide a fancy perf integration
because current step of DAMON development is focused on its core logic only.
That said, DAMON already provides two interfaces for user space programs, which
based on debugfs and tracepoint, respectively.  Using the tracepoint interface,
you can use DAMON with perf.  This patchset also provides a debugfs interface
based user space tool for DAMON.  It can be used to record, visualize, and
analyze data access patterns of target processes in a convenient way.

Q: Why a new module, instead of extending perf or other tools?
A: First, DAMON aims to be used by other programs including the kernel.
Therefore, having dependency to specific tools like perf is not desirable.
Second, because it need to be lightweight as much as possible so that it can be
used online, any unnecessary overhead such as kernel - user space context
switching cost should be avoided.  These are the two most biggest reasons why
DAMON is implemented in the kernel space.  The idle page tracking subsystem
would be the kernel module that most seems similar to DAMON.  However, its own
interface is not compatible with DAMON.  Also, the internal implementation of
it has no common part to be reused by DAMON.

Q: Can 'perf mem' provide the data required for DAMON?
A: On the systems supporting 'perf mem', yes.  DAMON is using the PTE Accessed
bits in low level.  Other H/W or S/W features that can be used for the purpose
could be used.  However, as explained with above question, DAMON need to be
implemented in the kernel space.


Evaluations
===========

We evaluated DAMON's overhead, monitoring quality and usefulness using 25
realistic workloads on my QEMU/KVM based virtual machine.

DAMON is lightweight.  It consumes only -0.08% more system memory and up to 1%
CPU time.  It makes target worloads only 0.76% slower.

DAMON is accurate and useful for memory management optimizations.  An
experimental DAMON-based operation scheme for THP removes 83.66% of THP memory
overheads while preserving 40.67% of THP speedup.  Another experimental
DAMON-based 'proactive reclamation' implementation reduced 22.42% of system
memory usage and 88.86% of residential sets while incurring only 3.07% runtime
overhead in best case.

NOTE that the experimentail THP optimization and proactive reclamation are not
for production, just only for proof of concepts.

Please refer to 'Appendix E' for detailed evaluation setup and results.


References
==========

Prototypes of DAMON have introduced by an LPC kernel summit track talk[1] and
two academic papers[2,3].  Please refer to those for more detailed information,
especially the evaluations.  The latest version of the patchsets has also
introduced by an LWN artice[4].

[1] SeongJae Park, Tracing Data Access Pattern with Bounded Overhead and
    Best-effort Accuracy. In The Linux Kernel Summit, September 2019.
    https://linuxplumbersconf.org/event/4/contributions/548/
[2] SeongJae Park, Yunjae Lee, Heon Y. Yeom, Profiling Dynamic Data Access
    Patterns with Controlled Overhead and Quality. In 20th ACM/IFIP
    International Middleware Conference Industry, December 2019.
    https://dl.acm.org/doi/10.1145/3366626.3368125
[3] SeongJae Park, Yunjae Lee, Yunhee Kim, Heon Y. Yeom, Profiling Dynamic Data
    Access Patterns with Bounded Overhead and Accuracy. In IEEE International
    Workshop on Foundations and Applications of Self- Systems (FAS 2019), June
    2019.
[4] Jonathan Corbet, Memory-management optimization with DAMON. In Linux Weekly
    News (LWN), Feb 2020. https://lwn.net/Articles/812707/


Sequence Of Patches
===================

The patches are organized in the following sequence.  The first two patches are
preparation of DAMON patchset.  The 1st patch adds typos found in previous
versions of DAMON patchset to 'scripts/spelling.txt' so that the typos can be
caught by 'checkpatch.pl'.  The 2nd patch exports 'lookup_page_ext()' to GPL
modules so that it can be used by DAMON even though it is built as a loadable
module.

Next four patches implement the core of DAMON and it's programming interface.
The 3rd patch introduces DAMON module, it's data structures, and data structure
related common functions.  Each of following three patches (4nd to 6th)
implements the core mechanisms of DAMON, namely regions based sampling,
adaptive regions adjustment, and dynamic memory mapping chage adoption,
respectively, with programming interface supports of those.

Following four patches are for low level users of DAMON.  The 7th patch
implements callbacks for each of monitoring steps so that users can do whatever
they want with the access patterns.  The 8th one implements recording of access
patterns in DAMON for better convenience and efficiency.  Each of next two
patches (9th and 10th) respectively adds a debugfs interface for privileged
people and/or programs in user space, and a tracepoint for other tracepoints
supporting tracers such as perf.

Two patches for high level users of DAMON follows.  To provide a minimal
reference to the debugfs interface and for high level use/tests of the DAMON,
the next patch (11th) implements an user space tool.  The 12th patch adds a
document for administrators of DAMON.

Next two patches are for tests.  The 13th and 14th patches provide unit tests
(based on kunit) and user space tests (based on kselftest), respectively.

Finally, the last patch (15th) updates the MAINTAINERS file.

The patches are based on the v5.5.  You can also clone the complete git
tree:

    $ git clone git://github.com/sjp38/linux -b damon/patches/v7

The web is also available:
https://github.com/sjp38/linux/releases/tag/damon/patches/v7


Patch History
=============

Changes from v6
(https://lore.kernel.org/linux-mm/20200224123047.32506-1-sjpark@amazon.com/)
 - Wordsmith cover letter (Shakeel Butt)
 - Cleanup code and commit messages (Jonathan Cameron)
 - Avoid kthread_run() under spinlock critical section (Jonathan Cameron)
 - Use kthread_stop() (Jonathan Cameron)
 - Change tracepoint to trace regions (Jonathan Cameron)
 - Implement API from the beginning (Jonathan Cameron)
 - Fix typos (Jonathan Cameron)
 - Fix access checking to properly handle regions smaller than single page
   (Jonathan Cameron)
 - Add found typos to 'scripts/spelling.txt'
 - Add recent evaluation results including DAMON-based Operation Schemes

Changes from v5
(https://lore.kernel.org/linux-mm/20200217103110.30817-1-sjpark@amazon.com/)
 - Fix minor bugs (sampling, record attributes, debugfs and user space tool)
 - selftests: Add debugfs interface tests for the bugs
 - Modify the user space tool to use its self default values for parameters
 - Fix pmg huge page access check

Changes from v4
(https://lore.kernel.org/linux-mm/20200210144812.26845-1-sjpark@amazon.com/)
 - Add 'Reviewed-by' for the kunit tests patch (Brendan Higgins)
 - Make the unit test to depedns on 'DAMON=y' (Randy Dunlap and kbuild bot)
   Reported-by: kbuild test robot <lkp@intel.com>
 - Fix m68k module build issue
   Reported-by: kbuild test robot <lkp@intel.com>
 - Add selftests
 - Seperate patches for low level users from core logics for better reading
 - Clean up debugfs interface
 - Trivial nitpicks

Changes from v3
(https://lore.kernel.org/linux-mm/20200204062312.19913-1-sj38.park@gmail.com/)
 - Fix i386 build issue
   Reported-by: kbuild test robot <lkp@intel.com>
 - Increase the default size of the monitoring result buffer to 1 MiB
 - Fix misc bugs in debugfs interface

Changes from v2
(https://lore.kernel.org/linux-mm/20200128085742.14566-1-sjpark@amazon.com/)
 - Move MAINTAINERS changes to last commit (Brendan Higgins)
 - Add descriptions for kunittest: why not only entire mappings and what the 4
   input sets are trying to test (Brendan Higgins)
 - Remove 'kdamond_need_stop()' test (Brendan Higgins)
 - Discuss about the 'perf mem' and DAMON (Peter Zijlstra)
 - Make CV clearly say what it actually does (Peter Zijlstra)
 - Answer why new module (Qian Cai)
 - Diable DAMON by default (Randy Dunlap)
 - Change the interface: Seperate recording attributes
   (attrs, record, rules) and allow multiple kdamond instances
 - Implement kernel API interface

Changes from v1
(https://lore.kernel.org/linux-mm/20200120162757.32375-1-sjpark@amazon.com/)
 - Rebase on v5.5
 - Add a tracepoint for integration with other tracers (Kirill A. Shutemov)
 - document: Add more description for the user space tool (Brendan Higgins)
 - unittest: Improve readability (Brendan Higgins)
 - unittest: Use consistent name and helpers function (Brendan Higgins)
 - Update PG_Young to avoid reclaim logic interference (Yunjae Lee)

Changes from RFC
(https://lore.kernel.org/linux-mm/20200110131522.29964-1-sjpark@amazon.com/)
 - Specify an ambiguous plan of access pattern based mm optimizations
 - Support loadable module build
 - Cleanup code

SeongJae Park (15):
  scripts/spelling: Add a few more typos
  mm/page_ext: Export lookup_page_ext() to GPL modules
  mm: Introduce Data Access MONitor (DAMON)
  mm/damon: Implement region based sampling
  mm/damon: Adaptively adjust regions
  mm/damon: Apply dynamic memory mapping changes
  mm/damon: Implement callbacks
  mm/damon: Implement access pattern recording
  mm/damon: Add debugfs interface
  mm/damon: Add tracepoints
  tools: Add a minimal user-space tool for DAMON
  Documentation/admin-guide/mm: Add a document for DAMON
  mm/damon: Add kunit tests
  mm/damon: Add user space selftests
  MAINTAINERS: Update for DAMON

 .../admin-guide/mm/data_access_monitor.rst    |  428 +++++
 Documentation/admin-guide/mm/index.rst        |    1 +
 MAINTAINERS                                   |   12 +
 include/linux/damon.h                         |   79 +
 include/trace/events/damon.h                  |   43 +
 mm/Kconfig                                    |   22 +
 mm/Makefile                                   |    1 +
 mm/damon-test.h                               |  604 +++++++
 mm/damon.c                                    | 1437 +++++++++++++++++
 mm/page_ext.c                                 |    1 +
 scripts/spelling.txt                          |    4 +
 tools/damon/.gitignore                        |    1 +
 tools/damon/_dist.py                          |   36 +
 tools/damon/bin2txt.py                        |   64 +
 tools/damon/damo                              |   37 +
 tools/damon/heats.py                          |  358 ++++
 tools/damon/nr_regions.py                     |   89 +
 tools/damon/record.py                         |  212 +++
 tools/damon/report.py                         |   45 +
 tools/damon/wss.py                            |   95 ++
 tools/testing/selftests/damon/Makefile        |    7 +
 .../selftests/damon/_chk_dependency.sh        |   28 +
 tools/testing/selftests/damon/_chk_record.py  |   89 +
 .../testing/selftests/damon/debugfs_attrs.sh  |  139 ++
 .../testing/selftests/damon/debugfs_record.sh |   50 +
 25 files changed, 3882 insertions(+)
 create mode 100644 Documentation/admin-guide/mm/data_access_monitor.rst
 create mode 100644 include/linux/damon.h
 create mode 100644 include/trace/events/damon.h
 create mode 100644 mm/damon-test.h
 create mode 100644 mm/damon.c
 create mode 100644 tools/damon/.gitignore
 create mode 100644 tools/damon/_dist.py
 create mode 100644 tools/damon/bin2txt.py
 create mode 100755 tools/damon/damo
 create mode 100644 tools/damon/heats.py
 create mode 100644 tools/damon/nr_regions.py
 create mode 100644 tools/damon/record.py
 create mode 100644 tools/damon/report.py
 create mode 100644 tools/damon/wss.py
 create mode 100644 tools/testing/selftests/damon/Makefile
 create mode 100644 tools/testing/selftests/damon/_chk_dependency.sh
 create mode 100644 tools/testing/selftests/damon/_chk_record.py
 create mode 100755 tools/testing/selftests/damon/debugfs_attrs.sh
 create mode 100755 tools/testing/selftests/damon/debugfs_record.sh

-- 
2.17.1

================================== >8 =========================================
Appendix A: Related Works
=========================

There are a number of researches[1,2,3,4,5,6] optimizing memory management
mechanisms based on the actual memory access patterns that shows impressive
results.  However, most of those has no deep consideration about the monitoring
of the accesses itself.  Some of those focused on the overhead of the
monitoring, but does not consider the accuracy scalability[6] or has additional
dependencies[7].  Indeed, one recent research[5] about the proactive
reclamation has also proposed[8] to the kernel community but the monitoring
overhead was considered a main problem.

[1] Subramanya R Dulloor, Amitabha Roy, Zheguang Zhao, Narayanan Sundaram,
    Nadathur Satish, Rajesh Sankaran, Jeff Jackson, and Karsten Schwan. 2016.
    Data tiering in heterogeneous memory systems. In Proceedings of the 11th
    European Conference on Computer Systems (EuroSys). ACM, 15.
[2] Youngjin Kwon, Hangchen Yu, Simon Peter, Christopher J Rossbach, and Emmett
    Witchel. 2016. Coordinated and efficient huge page management with ingens.
    In 12th USENIX Symposium on Operating Systems Design and Implementation
    (OSDI).  705–721.
[3] Harald Servat, Antonio J Peña, Germán Llort, Estanislao Mercadal,
    HansChristian Hoppe, and Jesús Labarta. 2017. Automating the application
    data placement in hybrid memory systems. In 2017 IEEE International
    Conference on Cluster Computing (CLUSTER). IEEE, 126–136.
[4] Vlad Nitu, Boris Teabe, Alain Tchana, Canturk Isci, and Daniel Hagimont.
    2018. Welcome to zombieland: practical and energy-efficient memory
    disaggregation in a datacenter. In Proceedings of the 13th European
    Conference on Computer Systems (EuroSys). ACM, 16.
[5] Andres Lagar-Cavilla, Junwhan Ahn, Suleiman Souhlal, Neha Agarwal, Radoslaw
    Burny, Shakeel Butt, Jichuan Chang, Ashwin Chaugule, Nan Deng, Junaid
    Shahid, Greg Thelen, Kamil Adam Yurtsever, Yu Zhao, and Parthasarathy
    Ranganathan.  2019. Software-Defined Far Memory in Warehouse-Scale
    Computers.  In Proceedings of the 24th International Conference on
    Architectural Support for Programming Languages and Operating Systems
    (ASPLOS).  ACM, New York, NY, USA, 317–330.
    DOI:https://doi.org/10.1145/3297858.3304053
[6] Carl Waldspurger, Trausti Saemundsson, Irfan Ahmad, and Nohhyun Park.
    2017. Cache Modeling and Optimization using Miniature Simulations. In 2017
    USENIX Annual Technical Conference (ATC). USENIX Association, Santa
    Clara, CA, 487–498.
    https://www.usenix.org/conference/atc17/technical-sessions/
[7] Haojie Wang, Jidong Zhai, Xiongchao Tang, Bowen Yu, Xiaosong Ma, and
    Wenguang Chen. 2018. Spindle: Informed Memory Access Monitoring. In 2018
    USENIX Annual Technical Conference (ATC). USENIX Association, Boston, MA,
    561–574.  https://www.usenix.org/conference/atc18/presentation/wang-haojie
[8] Jonathan Corbet. 2019. Proactively reclaiming idle memory. (2019).
    https://lwn.net/Articles/787611/.


Appendix B: Limitations of Other Access Monitoring Techniques
=============================================================

The memory access instrumentation techniques which are applied to
many tools such as Intel PIN is essential for correctness required cases such
as memory access bug detections or cache level optimizations.  However, those
usually incur exceptionally high overhead which is unacceptable.

Periodic access checks based on access counting features (e.g., PTE Accessed
bits or PG_Idle flags) can reduce the overhead.  It sacrifies some of the
quality but it's still ok to many of this domain.  However, the overhead
arbitrarily increase as the size of the target workload grows.  Miniature-like
static region based sampling can set the upperbound of the overhead, but it
will now decrease the quality of the output as the size of the workload grows.

DAMON is another solution that overcomes the limitations.  It is 1) accurate
enough for this domain, 2) light-weight so that it can be applied online, and
3) allow users to set the upper-bound of the overhead, regardless of the size
of target workloads.  It is implemented as a simple and small kernel module to
support various users in both of the user space and the kernel space.  Refer to
'Evaluations' section below for detailed performance of DAMON.

For the goals, DAMON utilizes its two core mechanisms, which allows lightweight
overhead and high quality of output, repectively.  To show how DAMON promises
those, refer to 'Mechanisms of DAMON' section below.


Appendix C: Mechanisms of DAMON
===============================


Basic Access Check
------------------

DAMON basically reports what pages are how frequently accessed.  The report is
passed to users in binary format via a ``result file`` which users can set it's
path.  Note that the frequency is not an absolute number of accesses, but a
relative frequency among the pages of the target workloads.

Users can also control the resolution of the reports by setting two time
intervals, ``sampling interval`` and ``aggregation interval``.  In detail,
DAMON checks access to each page per ``sampling interval``, aggregates the
results (counts the number of the accesses to each page), and reports the
aggregated results per ``aggregation interval``.  For the access check of each
page, DAMON uses the Accessed bits of PTEs.

This is thus similar to the previously mentioned periodic access checks based
mechanisms, which overhead is increasing as the size of the target process
grows.


Region Based Sampling
---------------------

To avoid the unbounded increase of the overhead, DAMON groups a number of
adjacent pages that assumed to have same access frequencies into a region.  As
long as the assumption (pages in a region have same access frequencies) is
kept, only one page in the region is required to be checked.  Thus, for each
``sampling interval``, DAMON randomly picks one page in each region and clears
its Accessed bit.  After one more ``sampling interval``, DAMON reads the
Accessed bit of the page and increases the access frequency of the region if
the bit has set meanwhile.  Therefore, the monitoring overhead is controllable
by setting the number of regions.  DAMON allows users to set the minimal and
maximum number of regions for the trade-off.

Except the assumption, this is almost same with the above-mentioned
miniature-like static region based sampling.  In other words, this scheme
cannot preserve the quality of the output if the assumption is not guaranteed.


Adaptive Regions Adjustment
---------------------------

At the beginning of the monitoring, DAMON constructs the initial regions by
evenly splitting the memory mapped address space of the process into the
user-specified minimal number of regions.  In this initial state, the
assumption is normally not kept and thus the quality could be low.  To keep the
assumption as much as possible, DAMON adaptively merges and splits each region.
For each ``aggregation interval``, it compares the access frequencies of
adjacent regions and merges those if the frequency difference is small.  Then,
after it reports and clears the aggregated access frequency of each region, it
splits each region into two regions if the total number of regions is smaller
than the half of the user-specified maximum number of regions.

In this way, DAMON provides its best-effort quality and minimal overhead while
keeping the bounds users set for their trade-off.


Applying Dynamic Memory Mappings
--------------------------------

Only a number of small parts in the super-huge virtual address space of the
processes is mapped to physical memory and accessed.  Thus, tracking the
unmapped address regions is just wasteful.  However, tracking every memory
mapping change might incur an overhead.  For the reason, DAMON applies the
dynamic memory mapping changes to the tracking regions only for each of an
user-specified time interval (``regions update interval``).


Appendix D: Expected Use-cases
==============================

A straightforward usecase of DAMON would be the program behavior analysis.
With the DAMON output, users can confirm whether the program is running as
intended or not.  This will be useful for debuggings and tests of design
points.

The monitored results can also be useful for counting the dynamic working set
size of workloads.  For the administration of memory overcommitted systems or
selection of the environments (e.g., containers providing different amount of
memory) for your workloads, this will be useful.

If you are a programmer, you can optimize your program by managing the memory
based on the actual data access pattern.  For example, you can identify the
dynamic hotness of your data using DAMON and call ``mlock()`` to keep your hot
data in DRAM, or call ``madvise()`` with ``MADV_PAGEOUT`` to proactively
reclaim cold data.  Even though your program is guaranteed to not encounter
memory pressure, you can still improve the performance by applying the DAMON
outputs for call of ``MADV_HUGEPAGE`` and ``MADV_NOHUGEPAGE``.  More creative
optimizations would be possible.  Our evaluations of DAMON includes a
straightforward optimization using the ``mlock()``.  Please refer to the below
Evaluation section for more detail.

As DAMON incurs very low overhead, such optimizations can be applied not only
offline, but also online.  Also, there is no reason to limit such optimizations
to the user space.  Several parts of the kernel's memory management mechanisms
could be also optimized using DAMON. The reclamation, the THP (de)promotion
decisions, and the compaction would be such a candidates.  DAMON will continue
its development to be highly optimized for the online/in-kernel uses.


A Future Plan: Data Access Monitoring-based Operation Schemes
-------------------------------------------------------------

As described in the above section, DAMON could be helpful for actual access
based memory management optimizations.  Nevertheless, users who want to do such
optimizations should run DAMON, read the traced data (either online or
offline), analyze it, plan a new memory management scheme, and apply the new
scheme by themselves.  It must be easier than the past, but could still require
some level of efforts.  In its next development stage, DAMON will reduce some
of such efforts by allowing users to specify some access based memory
management rules for their specific processes.

Because this is just a plan, the specific interface is not fixed yet, but for
example, users will be allowed to write their desired memory management rules
to a special file in a DAMON specific format.  The rules will be something like
'if a memory region of size in a range is keeping a range of hotness for more
than a duration, apply specific memory management rule using madvise() or
mlock() to the region'.  For example, we can imagine rules like below:

    # format is: <min/max size> <min/max frequency (0-99)> <duration> <action>

    # if a region of a size keeps a very high access frequency for more than
    # 100ms, lock the region in the main memory (call mlock()). But, if the
    # region is larger than 500 MiB, skip it. The exception might be helpful
    # if the system has only, say, 600 MiB of DRAM, a region of size larger
    # than 600 MiB cannot be locked in the DRAM at all.
    na 500M 90 99 100ms mlock

    # if a region keeps a high access frequency for more than 100ms, put the
    # region on the head of the LRU list (call madvise() with MADV_WILLNEED).
    na na 80 90 100ms madv_willneed

    # if a region keeps a low access frequency for more than 100ms, put the
    # region on the tail of the LRU list (call madvise() with MADV_COLD).
    na na 10 20 100ms madv_cold

    # if a region keeps a very low access frequency for more than 100ms, swap
    # out the region immediately (call madvise() with MADV_PAGEOUT).
    na na 0 10 100ms madv_pageout

    # if a region of a size bigger than 2MB keeps a very high access frequency
    # for more than 100ms, let the region to use huge pages (call madvise()
    # with MADV_HUGEPAGE).
    2M na 90 99 100ms madv_hugepage

    # If a regions of a size bigger than > 2MB keeps no high access frequency
    # for more than 100ms, avoid the region from using huge pages (call
    # madvise() with MADV_NOHUGEPAGE).
    2M na 0 25 100ms madv_nohugepage

An RFC patchset for this is available:
https://lore.kernel.org/linux-mm/20200218085309.18346-1-sjpark@amazon.com/


Appendix E: Evaluations
=======================

Setup
-----

On my personal QEMU/KVM based virtual machine on an Intel i7 host machine
running Ubuntu 18.04, I measure runtime and consumed system memory while
running various realistic workloads with several configurations.  I use 13 and
12 workloads in PARSEC3[3] and SPLASH-2X[4] benchmark suites, respectively.  I
personally use another wrapper scripts[5] for setup and run of the workloads.
On top of this patchset, we also applied the DAMON-based operation schemes
patchset[6] for this evaluation.

Measurement
~~~~~~~~~~~

For the measurement of the amount of consumed memory in system global scope, I
drop caches before starting each of the workloads and monitor 'MemFree' in the
'/proc/meminfo' file.  To make results more stable, I repeat the runs 5 times
and average results.  You can get stdev, min, and max of the numbers among the
repeated runs in appendix below.

Configurations
~~~~~~~~~~~~~~

The configurations I use are as below.

orig: Linux v5.5 with 'madvise' THP policy
rec: 'orig' plus DAMON running with record feature
thp: same with 'orig', but use 'always' THP policy
ethp: 'orig' plus a DAMON operation scheme[6], 'efficient THP'
prcl: 'orig' plus a DAMON operation scheme, 'proactive reclaim[7]'

I use 'rec' for measurement of DAMON overheads to target workloads and system
memory.  The remaining configs including 'thp', 'ethp', and 'prcl' are for
measurement of DAMON monitoring accuracy.

'ethp' and 'prcl' is simple DAMON-based operation schemes developed for
proof of concepts of DAMON.  'ethp' reduces memory space waste of THP by using
DAMON for decision of promotions and demotion for huge pages, while 'prcl' is
as similar as the original work.  Those are implemented as below:

# format: <min/max size> <min/max frequency (0-100)> <min/max age> <action>
# ethp: Use huge pages if a region >2MB shows >5% access rate, use regular
# pages if a region >2MB shows <5% access rate for >1 second
2M null    5 null    null null    hugepage
2M null    null 5    1s null      nohugepage

# prcl: If a region >4KB shows <5% access rate for >5 seconds, page out.
4K null    null 5    5s null      pageout

Note that both 'ethp' and 'prcl' are designed with my only straightforward
intuition, because those are for only proof of concepts and monitoring accuracy
of DAMON.  In other words, those are not for production.  For production use,
those should be tuned more.


[1] "Redis latency problems troubleshooting", https://redis.io/topics/latency
[2] "Disable Transparent Huge Pages (THP)",
    https://docs.mongodb.com/manual/tutorial/transparent-huge-pages/
[3] "The PARSEC Becnhmark Suite", https://parsec.cs.princeton.edu/index.htm
[4] "SPLASH-2x", https://parsec.cs.princeton.edu/parsec3-doc.htm#splash2x
[5] "parsec3_on_ubuntu", https://github.com/sjp38/parsec3_on_ubuntu
[6] "[RFC v4 0/7] Implement Data Access Monitoring-based Memory Operation
    Schemes",
    https://lore.kernel.org/linux-mm/20200303121406.20954-1-sjpark@amazon.com/
[7] "Proactively reclaiming idle memory", https://lwn.net/Articles/787611/


Results
-------

Below two tables show the measurement results.  The runtimes are in seconds
while the memory usages are in KiB.  Each configurations except 'orig' shows
its overhead relative to 'orig' in percent within parenthesises.

runtime                 orig     rec      (overhead) thp      (overhead) ethp     (overhead) prcl     (overhead)
parsec3/blackscholes    106.586  107.160  (0.54)     106.535  (-0.05)    107.393  (0.76)     114.543  (7.47)    
parsec3/bodytrack       78.621   79.220   (0.76)     78.678   (0.07)     79.169   (0.70)     80.793   (2.76)    
parsec3/canneal         138.951  142.258  (2.38)     123.555  (-11.08)   133.588  (-3.86)    143.239  (3.09)    
parsec3/dedup           11.876   11.918   (0.35)     11.767   (-0.92)    11.957   (0.68)     13.235   (11.44)   
parsec3/facesim         207.761  208.159  (0.19)     204.735  (-1.46)    207.172  (-0.28)    208.663  (0.43)    
parsec3/ferret          190.694  192.004  (0.69)     190.345  (-0.18)    190.453  (-0.13)    192.081  (0.73)    
parsec3/fluidanimate    210.189  212.511  (1.10)     208.695  (-0.71)    210.843  (0.31)     213.379  (1.52)    
parsec3/freqmine        289.000  289.483  (0.17)     287.724  (-0.44)    289.761  (0.26)     297.878  (3.07)    
parsec3/raytrace        118.482  119.346  (0.73)     118.861  (0.32)     119.151  (0.56)     136.566  (15.26)   
parsec3/streamcluster   323.338  328.431  (1.58)     285.039  (-11.85)   296.830  (-8.20)    331.670  (2.58)    
parsec3/swaptions       155.853  156.826  (0.62)     154.089  (-1.13)    156.332  (0.31)     155.422  (-0.28)   
parsec3/vips            58.864   59.408   (0.92)     58.450   (-0.70)    58.976   (0.19)     61.068   (3.74)    
parsec3/x264            69.201   69.208   (0.01)     68.795   (-0.59)    71.501   (3.32)     71.766   (3.71)    
splash2x/barnes         81.140   80.869   (-0.33)    74.734   (-7.90)    79.859   (-1.58)    108.875  (34.18)   
splash2x/fft            33.442   33.579   (0.41)     22.949   (-31.38)   27.055   (-19.10)   40.261   (20.39)   
splash2x/lu_cb          85.064   85.441   (0.44)     84.688   (-0.44)    85.868   (0.95)     88.949   (4.57)    
splash2x/lu_ncb         92.606   93.615   (1.09)     90.484   (-2.29)    93.368   (0.82)     93.279   (0.73)    
splash2x/ocean_cp       44.672   44.826   (0.34)     43.024   (-3.69)    43.671   (-2.24)    45.889   (2.72)    
splash2x/ocean_ncp      81.360   81.434   (0.09)     51.157   (-37.12)   66.711   (-18.00)   91.611   (12.60)   
splash2x/radiosity      91.374   91.568   (0.21)     90.406   (-1.06)    91.609   (0.26)     103.790  (13.59)   
splash2x/radix          31.330   31.509   (0.57)     25.145   (-19.74)   26.296   (-16.07)   31.835   (1.61)    
splash2x/raytrace       84.715   85.274   (0.66)     82.034   (-3.16)    84.458   (-0.30)    84.967   (0.30)    
splash2x/volrend        86.625   87.844   (1.41)     86.206   (-0.48)    87.851   (1.42)     87.809   (1.37)    
splash2x/water_nsquared 231.661  233.817  (0.93)     221.024  (-4.59)    228.020  (-1.57)    236.306  (2.01)    
splash2x/water_spatial  89.101   89.616   (0.58)     88.845   (-0.29)    89.710   (0.68)     103.370  (16.01)   
total                   2992.490 3015.330 (0.76)     2857.950 (-4.50)    2937.610 (-1.83)    3137.260 (4.84)    


memused.avg             orig         rec          (overhead) thp          (overhead) ethp         (overhead) prcl         (overhead)
parsec3/blackscholes    1822704.400  1833697.600  (0.60)     1826160.400  (0.19)     1833316.800  (0.58)     1657871.000  (-9.04)   
parsec3/bodytrack       1417677.600  1434893.200  (1.21)     1420652.200  (0.21)     1431637.000  (0.98)     1433359.800  (1.11)    
parsec3/canneal         1044807.000  1056496.200  (1.12)     1037582.400  (-0.69)    1050845.200  (0.58)     1051668.200  (0.66)    
parsec3/dedup           2408896.200  2433019.000  (1.00)     2403343.200  (-0.23)    2421191.800  (0.51)     2461284.400  (2.17)    
parsec3/facesim         541808.200   554404.200   (2.32)     545591.600   (0.70)     553669.600   (2.19)     553910.600   (2.23)    
parsec3/ferret          319697.200   331642.400   (3.74)     320722.000   (0.32)     332126.000   (3.89)     330581.800   (3.40)    
parsec3/fluidanimate    573267.400   587376.200   (2.46)     574660.200   (0.24)     596108.600   (3.98)     538974.600   (-5.98)   
parsec3/freqmine        986872.400   998956.200   (1.22)     992037.800   (0.52)     989680.800   (0.28)     765626.800   (-22.42)  
parsec3/raytrace        1749641.800  1761473.200  (0.68)     1743617.800  (-0.34)    1753105.600  (0.20)     1580514.800  (-9.67)   
parsec3/streamcluster   125165.400   149479.600   (19.43)    122082.000   (-2.46)    140484.200   (12.24)    132027.000   (5.48)    
parsec3/swaptions       15515.400    29577.200    (90.63)    15692.000    (1.14)     26733.200    (72.30)    28423.000    (83.19)   
parsec3/vips            2954233.800  2970852.400  (0.56)     2954338.800  (0.00)     2959100.200  (0.16)     2951979.600  (-0.08)   
parsec3/x264            3174959.000  3191900.200  (0.53)     3192736.200  (0.56)     3201927.200  (0.85)     3194867.400  (0.63)    
splash2x/barnes         1215064.400  1209725.600  (-0.44)    1215945.600  (0.07)     1212294.600  (-0.23)    937605.800   (-22.83)  
splash2x/fft            9429331.600  9187727.600  (-2.56)    9290976.600  (-1.47)    9036430.800  (-4.17)    9409815.800  (-0.21)   
splash2x/lu_cb          512744.800   521964.600   (1.80)     521795.800   (1.77)     522445.600   (1.89)     346352.200   (-32.45)  
splash2x/lu_ncb         516623.000   523673.200   (1.36)     520129.200   (0.68)     522398.800   (1.12)     522246.200   (1.09)    
splash2x/ocean_cp       3325422.200  3287326.200  (-1.15)    3381646.400  (1.69)     3294803.400  (-0.92)    3287401.800  (-1.14)   
splash2x/ocean_ncp      3894128.600  3868638.400  (-0.65)    7065137.400  (81.43)    4844981.600  (24.42)    3811968.400  (-2.11)   
splash2x/radiosity      1471464.000  1470680.800  (-0.05)    1481054.600  (0.65)     1472332.200  (0.06)     521064.000   (-64.59)  
splash2x/radix          1698164.400  1707518.400  (0.55)     1385276.800  (-18.42)   1415885.000  (-16.62)   1717103.600  (1.12)    
splash2x/raytrace       45334.200    59478.400    (31.20)    52893.400    (16.67)    62366.000    (37.57)    53765.800    (18.60)   
splash2x/volrend        151118.400   167429.800   (10.79)    151600.000   (0.32)     163950.800   (8.49)     162873.800   (7.78)    
splash2x/water_nsquared 46839.000    61947.000    (32.26)    49173.600    (4.98)     58301.200    (24.47)    56678.400    (21.01)   
splash2x/water_spatial  666960.000   674851.600   (1.18)     668957.600   (0.30)     673287.400   (0.95)     463938.800   (-30.44)  
total                   40108199.000 40074800.000 (-0.08)    42933800.000 (7.04)     40569300.000 (1.15)     37972000.000 (-5.33)   


DAMON Overheads
~~~~~~~~~~~~~~~

In total, DAMON recording feature incurs 0.76% runtime overhead (up to 2.38% in
worst case with 'parsec3/canneal') and -0.08% memory space overhead.

For convenience test run of 'rec', I use a Python wrapper.  The wrapper
constantly consumes about 10-15MB of memory.  This becomes high memory overhead
if the target workload has small memory footprint.  In detail, 19%, 90%, 31%,
10%, and 32% overheads shown for parsec3/streamcluster (125 MiB),
parsec3/swaptions (15 MiB), splash2x/raytrace (45 MiB), splash2x/volrend (151
MiB), and splash2x/water_nsquared (46 MiB)).  Nonetheless, the overheads are
not from DAMON, but from the wrapper, and thus should be ignored.  This fake
memory overhead continues in 'ethp' and 'prcl', as those configurations are
also using the Python wrapper.


Efficient THP
~~~~~~~~~~~~~

THP 'always' enabled policy achieves 4.5% speedup but incurs 7.04% memory
overhead.  It achieves 37.12% speedup in best case, but 81.43% memory overhead
in worst case.  Interestingly, both the best and worst case are with
'splash2x/ocean_ncp').

The 2-lines implementation of data access monitoring based THP version ('ethp')
shows 1.83% speedup and 1.15% memory overhead.  In other words, 'ethp' removes
83.66% of THP memory waste while preserving 40.67% of THP speedup in total.

In case of the 'splash2x/ocean_ncp', which is best for speedup but worst for
memory overhead of THP, 'ethp' removes 70% of THP memory space overhead while
preserving 48.49% of THP speedup.


Proactive Reclamation
~~~~~~~~~~~~~~~~~~~~

As same to the original work, I use 'zram' swap device for this configuration.

In total, our 1 line implementation of Proactive Reclamation, 'prcl', incurred
4.84% runtime overhead in total while achieving 5.33% system memory usage
reduction.

Nonetheless, as the memory usage is calculated with 'MemFree' in
'/proc/meminfo', it contains the SwapCached pages.  As the swapcached pages can
be easily evicted, I also measured the residential set size of the workloads:

rss.avg                 orig         prcl         (overhead)
parsec3/blackscholes    589633.600   329611.400   (-44.10)
parsec3/bodytrack       32217.600    21652.200    (-32.79)
parsec3/canneal         840411.600   838931.000   (-0.18)
parsec3/dedup           1223907.600  835473.600   (-31.74)
parsec3/facesim         311271.600   311070.200   (-0.06)
parsec3/ferret          99635.600    89290.800    (-10.38)
parsec3/fluidanimate    531760.000   484945.600   (-8.80)
parsec3/freqmine        552609.400   61583.600    (-88.86)
parsec3/raytrace        896446.600   317792.000   (-64.55)
parsec3/streamcluster   110793.600   108061.600   (-2.47)
parsec3/swaptions       5604.600     2694.400     (-51.93)
parsec3/vips            31779.600    28422.200    (-10.56)
parsec3/x264            81943.800    81874.600    (-0.08)
splash2x/barnes         1219389.600  619038.600   (-49.23)
splash2x/fft            9597789.600  7264542.200  (-24.31)
splash2x/lu_cb          510524.000   327813.600   (-35.79)
splash2x/lu_ncb         510131.200   510146.800   (0.00)
splash2x/ocean_cp       3406968.600  3341620.400  (-1.92)
splash2x/ocean_ncp      3919926.800  3670768.800  (-6.36)
splash2x/radiosity      1474387.800  254678.600   (-82.73)
splash2x/radix          1723283.200  1763916.000  (2.36)
splash2x/raytrace       23194.400    17454.000    (-24.75)
splash2x/volrend        43980.000    32524.600    (-26.05)
splash2x/water_nsquared 29327.200    23989.200    (-18.20)
splash2x/water_spatial  656323.200   381068.600   (-41.94)
total                   28423300.000 21719000.000 (-23.59)

In total, 23.59% of residential sets were reduced.

With parsec3/freqmine, 'prcl' reduced 22.42% of system memory usage and 88.86%
of residential sets while incurring only 3.07% runtime overhead.


Appendix F: Prototype Evaluations
=================================

A prototype of DAMON has evaluated on an Intel Xeon E7-8837 machine using 20
benchmarks that picked from SPEC CPU 2006, NAS, Tensorflow Benchmark,
SPLASH-2X, and PARSEC 3 benchmark suite.  Nonethless, this section provides
only summary of the results.  For more detail, please refer to the slides used
for the introduction of DAMON at the Linux Plumbers Conference 2019[1] or the
MIDDLEWARE'19 industrial track paper[2].

[1] SeongJae Park, Tracing Data Access Pattern with Bounded Overhead and
    Best-effort Accuracy. In The Linux Kernel Summit, September 2019.
    https://linuxplumbersconf.org/event/4/contributions/548/
[2] SeongJae Park, Yunjae Lee, Heon Y. Yeom, Profiling Dynamic Data Access
    Patterns with Controlled Overhead and Quality. In 20th ACM/IFIP
    International Middleware Conference Industry, December 2019.
    https://dl.acm.org/doi/10.1145/3366626.3368125


Quality
-------

We first traced and visualized the data access pattern of each workload.  We
were able to confirm that the visualized results are reasonably accurate by
manually comparing those with the source code of the workloads.

To see the usefulness of the monitoring, we optimized 9 memory intensive
workloads among them for memory pressure situations using the DAMON outputs.
In detail, we identified frequently accessed memory regions in each workload
based on the DAMON results and protected them with ``mlock()`` system calls by
manually modifying the source code.  The optimized versions consistently show
speedup (2.55x in best case, 1.65x in average) under artificial memory
pressures.  We use cgroups for the pressure.


Overhead
--------

We also measured the overhead of DAMON.  The upperbound we set was kept as
expected.  Besides, it was much lower (0.6 percent of the bound in best case,
13.288 percent of the bound in average).  This reduction of the overhead is
mainly resulted from its core mechanism called adaptive regions adjustment.
Refer to 'Appendix D' for more detail about the mechanism.  We also compared
the overhead of DAMON with that of a straightforward periodic PTE Accessed bit
checking based monitoring.  DAMON's overhead was smaller than it by 94,242.42x
in best case, 3,159.61x in average.

The latest version of DAMON running with its default configuration consumes
only up to 1% of CPU time when applied to realistic workloads in PARSEC3 and
SPLASH-2X and makes no visible slowdown to the target processes.

^ permalink raw reply	[flat|nested] 28+ messages in thread

* [PATCH v7 01/15] scripts/spelling: Add a few more typos
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 02/15] mm/page_ext: Export lookup_page_ext() to GPL modules SeongJae Park
                   ` (14 subsequent siblings)
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit adds typos found from DAMON patchset.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 scripts/spelling.txt | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/scripts/spelling.txt b/scripts/spelling.txt
index 672b5931bc8d..2fcad5dbd9c6 100644
--- a/scripts/spelling.txt
+++ b/scripts/spelling.txt
@@ -57,6 +57,7 @@ actualy||actually
 acumulating||accumulating
 acumulative||accumulative
 acumulator||accumulator
+acutally||actually
 adapater||adapter
 addional||additional
 additionaly||additionally
@@ -241,6 +242,7 @@ calender||calendar
 calescing||coalescing
 calle||called
 callibration||calibration
+callser||caller
 calucate||calculate
 calulate||calculate
 cancelation||cancellation
@@ -988,6 +990,7 @@ partiton||partition
 pased||passed
 passin||passing
 pathes||paths
+pattrns||patterns
 pecularities||peculiarities
 peformance||performance
 peforming||performing
@@ -1338,6 +1341,7 @@ thead||thread
 therfore||therefore
 thier||their
 threds||threads
+threee||three
 threshhold||threshold
 thresold||threshold
 throught||through
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 02/15] mm/page_ext: Export lookup_page_ext() to GPL modules
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 01/15] scripts/spelling: Add a few more typos SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 03/15] mm: Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (13 subsequent siblings)
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit exports 'lookup_page_ext()' to GPL modules.  This will be
used by DAMON.

This reverts commit a50179d0e22931032aa56718cd069b0ec74974c6.
---
 mm/page_ext.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/mm/page_ext.c b/mm/page_ext.c
index 4ade843ff588..71169b45bba9 100644
--- a/mm/page_ext.c
+++ b/mm/page_ext.c
@@ -131,6 +131,7 @@ struct page_ext *lookup_page_ext(const struct page *page)
 					MAX_ORDER_NR_PAGES);
 	return get_entry(base, index);
 }
+EXPORT_SYMBOL_GPL(lookup_page_ext);
 
 static int __init alloc_node_page_ext(int nid)
 {
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 03/15] mm: Introduce Data Access MONitor (DAMON)
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 01/15] scripts/spelling: Add a few more typos SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 02/15] mm/page_ext: Export lookup_page_ext() to GPL modules SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 04/15] mm/damon: Implement region based sampling SeongJae Park
                   ` (12 subsequent siblings)
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit introduces a kernel module named DAMON.  Note that this
commit is implementing only the stub for the module load/unload, basic
data structures, and simple manipulation functions of the structures to
keep the size of commit small.  The core mechanisms of DAMON will be
implemented one by one by following commits.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 include/linux/damon.h |  38 +++++++++
 mm/Kconfig            |  11 +++
 mm/Makefile           |   1 +
 mm/damon.c            | 189 ++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 239 insertions(+)
 create mode 100644 include/linux/damon.h
 create mode 100644 mm/damon.c

diff --git a/include/linux/damon.h b/include/linux/damon.h
new file mode 100644
index 000000000000..7117bb7e7544
--- /dev/null
+++ b/include/linux/damon.h
@@ -0,0 +1,38 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * DAMON api
+ *
+ * Copyright 2019 Amazon.com, Inc. or its affiliates.  All rights reserved.
+ *
+ * Author: SeongJae Park <sjpark@amazon.de>
+ */
+
+#ifndef _DAMON_H_
+#define _DAMON_H_
+
+#include <linux/random.h>
+#include <linux/types.h>
+
+/* Represents a monitoring target region on the virtual address space */
+struct damon_region {
+	unsigned long vm_start;
+	unsigned long vm_end;
+	unsigned long sampling_addr;
+	unsigned int nr_accesses;
+	struct list_head list;
+};
+
+/* Represents a monitoring target task */
+struct damon_task {
+	unsigned long pid;
+	struct list_head regions_list;
+	struct list_head list;
+};
+
+struct damon_ctx {
+	struct rnd_state rndseed;
+
+	struct list_head tasks_list;	/* 'damon_task' objects */
+};
+
+#endif
diff --git a/mm/Kconfig b/mm/Kconfig
index ab80933be65f..79da00d15604 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -739,4 +739,15 @@ config ARCH_HAS_HUGEPD
 config MAPPING_DIRTY_HELPERS
         bool
 
+config DAMON
+	tristate "Data Access Monitor"
+	depends on MMU
+	help
+	  Provides data access monitoring.
+
+	  DAMON is a kernel module that allows users to monitor the actual
+	  memory access pattern of specific user-space processes.  It aims to
+	  be 1) accurate enough to be useful for performance-centric domains,
+	  and 2) sufficiently light-weight so that it can be applied online.
+
 endmenu
diff --git a/mm/Makefile b/mm/Makefile
index 1937cc251883..2911b3832c90 100644
--- a/mm/Makefile
+++ b/mm/Makefile
@@ -108,3 +108,4 @@ obj-$(CONFIG_ZONE_DEVICE) += memremap.o
 obj-$(CONFIG_HMM_MIRROR) += hmm.o
 obj-$(CONFIG_MEMFD_CREATE) += memfd.o
 obj-$(CONFIG_MAPPING_DIRTY_HELPERS) += mapping_dirty_helpers.o
+obj-$(CONFIG_DAMON) += damon.o
diff --git a/mm/damon.c b/mm/damon.c
new file mode 100644
index 000000000000..d7e6226ab7f1
--- /dev/null
+++ b/mm/damon.c
@@ -0,0 +1,189 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Data Access Monitor
+ *
+ * Copyright 2019 Amazon.com, Inc. or its affiliates.  All rights reserved.
+ *
+ * Author: SeongJae Park <sjpark@amazon.de>
+ */
+
+#define pr_fmt(fmt) "damon: " fmt
+
+#include <linux/damon.h>
+#include <linux/mm.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+
+#define damon_get_task_struct(t) \
+	(get_pid_task(find_vpid(t->pid), PIDTYPE_PID))
+
+#define damon_next_region(r) \
+	(container_of(r->list.next, struct damon_region, list))
+
+#define damon_prev_region(r) \
+	(container_of(r->list.prev, struct damon_region, list))
+
+#define damon_for_each_region(r, t) \
+	list_for_each_entry(r, &t->regions_list, list)
+
+#define damon_for_each_region_safe(r, next, t) \
+	list_for_each_entry_safe(r, next, &t->regions_list, list)
+
+#define damon_for_each_task(ctx, t) \
+	list_for_each_entry(t, &(ctx)->tasks_list, list)
+
+#define damon_for_each_task_safe(ctx, t, next) \
+	list_for_each_entry_safe(t, next, &(ctx)->tasks_list, list)
+
+/* Get a random number in [l, r) */
+#define damon_rand(ctx, l, r) (l + prandom_u32_state(&ctx->rndseed) % (r - l))
+
+/*
+ * Construct a damon_region struct
+ *
+ * Returns the pointer to the new struct if success, or NULL otherwise
+ */
+static struct damon_region *damon_new_region(struct damon_ctx *ctx,
+				unsigned long vm_start, unsigned long vm_end)
+{
+	struct damon_region *region;
+
+	region = kmalloc(sizeof(*region), GFP_KERNEL);
+	if (!region)
+		return NULL;
+
+	region->vm_start = vm_start;
+	region->vm_end = vm_end;
+	region->nr_accesses = 0;
+	region->sampling_addr = damon_rand(ctx, vm_start, vm_end);
+	INIT_LIST_HEAD(&region->list);
+
+	return region;
+}
+
+/*
+ * Add a region between two other regions
+ */
+static inline void damon_insert_region(struct damon_region *r,
+		struct damon_region *prev, struct damon_region *next)
+{
+	__list_add(&r->list, &prev->list, &next->list);
+}
+
+static void damon_add_region(struct damon_region *r, struct damon_task *t)
+{
+	list_add_tail(&r->list, &t->regions_list);
+}
+
+static void damon_del_region(struct damon_region *r)
+{
+	list_del(&r->list);
+}
+
+static void damon_free_region(struct damon_region *r)
+{
+	kfree(r);
+}
+
+static void damon_destroy_region(struct damon_region *r)
+{
+	damon_del_region(r);
+	damon_free_region(r);
+}
+
+/*
+ * Construct a damon_task struct
+ *
+ * Returns the pointer to the new struct if success, or NULL otherwise
+ */
+static struct damon_task *damon_new_task(unsigned long pid)
+{
+	struct damon_task *t;
+
+	t = kmalloc(sizeof(*t), GFP_KERNEL);
+	if (!t)
+		return NULL;
+
+	t->pid = pid;
+	INIT_LIST_HEAD(&t->regions_list);
+
+	return t;
+}
+
+/* Returns n-th damon_region of the given task */
+struct damon_region *damon_nth_region_of(struct damon_task *t, unsigned int n)
+{
+	struct damon_region *r;
+	unsigned int i = 0;
+
+	damon_for_each_region(r, t) {
+		if (i++ == n)
+			return r;
+	}
+
+	return NULL;
+}
+
+static void damon_add_task(struct damon_ctx *ctx, struct damon_task *t)
+{
+	list_add_tail(&t->list, &ctx->tasks_list);
+}
+
+static void damon_del_task(struct damon_task *t)
+{
+	list_del(&t->list);
+}
+
+static void damon_free_task(struct damon_task *t)
+{
+	struct damon_region *r, *next;
+
+	damon_for_each_region_safe(r, next, t)
+		damon_free_region(r);
+	kfree(t);
+}
+
+static void damon_destroy_task(struct damon_task *t)
+{
+	damon_del_task(t);
+	damon_free_task(t);
+}
+
+static unsigned int nr_damon_tasks(struct damon_ctx *ctx)
+{
+	struct damon_task *t;
+	unsigned int nr_tasks = 0;
+
+	damon_for_each_task(ctx, t)
+		nr_tasks++;
+
+	return nr_tasks;
+}
+
+static unsigned int nr_damon_regions(struct damon_task *t)
+{
+	struct damon_region *r;
+	unsigned int nr_regions = 0;
+
+	damon_for_each_region(r, t)
+		nr_regions++;
+
+	return nr_regions;
+}
+
+static int __init damon_init(void)
+{
+	return 0;
+}
+
+static void __exit damon_exit(void)
+{
+	return;
+}
+
+module_init(damon_init);
+module_exit(damon_exit);
+
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("SeongJae Park <sjpark@amazon.de>");
+MODULE_DESCRIPTION("DAMON: Data Access MONitor");
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 04/15] mm/damon: Implement region based sampling
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (2 preceding siblings ...)
  2020-03-18 11:27 ` [PATCH v7 03/15] mm: Introduce Data Access MONitor (DAMON) SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-31 16:02   ` Jonathan Cameron
  2020-04-02 15:07   ` SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 05/15] mm/damon: Adaptively adjust regions SeongJae Park
                   ` (11 subsequent siblings)
  15 siblings, 2 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit implements DAMON's basic access check and region based
sampling mechanisms.  This change would seems make no sense, mainly
because it is only a part of the DAMON's logics.  Following two commits
will make more sense.

Basic Access Check
------------------

DAMON basically reports what pages are how frequently accessed.  Note
that the frequency is not an absolute number of accesses, but a relative
frequency among the pages of the target workloads.

Users can control the resolution of the reports by setting two time
intervals, ``sampling interval`` and ``aggregation interval``.  In
detail, DAMON checks access to each page per ``sampling interval``,
aggregates the results (counts the number of the accesses to each page),
and reports the aggregated results per ``aggregation interval``.  For
the access check of each page, DAMON uses the Accessed bits of PTEs.

This is thus similar to common periodic access checks based access
tracking mechanisms, which overhead is increasing as the size of the
target process grows.

Region Based Sampling
---------------------

To avoid the unbounded increase of the overhead, DAMON groups a number
of adjacent pages that assumed to have same access frequencies into a
region.  As long as the assumption (pages in a region have same access
frequencies) is kept, only one page in the region is required to be
checked.  Thus, for each ``sampling interval``, DAMON randomly picks one
page in each region and clears its Accessed bit.  After one more
``sampling interval``, DAMON reads the Accessed bit of the page and
increases the access frequency of the region if the bit has set
meanwhile.  Therefore, the monitoring overhead is controllable by
setting the number of regions.

Nonetheless, this scheme cannot preserve the quality of the output if
the assumption is not kept.  Following commit will introduce how we can
make the guarantee with best effort.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 include/linux/damon.h |  24 ++
 mm/damon.c            | 553 ++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 577 insertions(+)

diff --git a/include/linux/damon.h b/include/linux/damon.h
index 7117bb7e7544..f1945df6e6b4 100644
--- a/include/linux/damon.h
+++ b/include/linux/damon.h
@@ -11,6 +11,8 @@
 #define _DAMON_H_
 
 #include <linux/random.h>
+#include <linux/mutex.h>
+#include <linux/time64.h>
 #include <linux/types.h>
 
 /* Represents a monitoring target region on the virtual address space */
@@ -29,10 +31,32 @@ struct damon_task {
 	struct list_head list;
 };
 
+/*
+ * For each 'sample_interval', DAMON checks whether each region is accessed or
+ * not.  It aggregates and keeps the access information (number of accesses to
+ * each region) for each 'aggr_interval' time.
+ *
+ * All time intervals are in micro-seconds.
+ */
 struct damon_ctx {
+	unsigned long sample_interval;
+	unsigned long aggr_interval;
+	unsigned long min_nr_regions;
+
+	struct timespec64 last_aggregation;
+
+	struct task_struct *kdamond;
+	struct mutex kdamond_lock;
+
 	struct rnd_state rndseed;
 
 	struct list_head tasks_list;	/* 'damon_task' objects */
 };
 
+int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids);
+int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
+		unsigned long aggr_int, unsigned long min_nr_reg);
+int damon_start(struct damon_ctx *ctx);
+int damon_stop(struct damon_ctx *ctx);
+
 #endif
diff --git a/mm/damon.c b/mm/damon.c
index d7e6226ab7f1..018016793555 100644
--- a/mm/damon.c
+++ b/mm/damon.c
@@ -10,8 +10,14 @@
 #define pr_fmt(fmt) "damon: " fmt
 
 #include <linux/damon.h>
+#include <linux/delay.h>
+#include <linux/kthread.h>
 #include <linux/mm.h>
 #include <linux/module.h>
+#include <linux/page_idle.h>
+#include <linux/random.h>
+#include <linux/sched/mm.h>
+#include <linux/sched/task.h>
 #include <linux/slab.h>
 
 #define damon_get_task_struct(t) \
@@ -171,6 +177,553 @@ static unsigned int nr_damon_regions(struct damon_task *t)
 	return nr_regions;
 }
 
+/*
+ * Get the mm_struct of the given task
+ *
+ * Caller should put the mm_struct after use, unless it is NULL.
+ *
+ * Returns the mm_struct of the task on success, NULL on failure
+ */
+static struct mm_struct *damon_get_mm(struct damon_task *t)
+{
+	struct task_struct *task;
+	struct mm_struct *mm;
+
+	task = damon_get_task_struct(t);
+	if (!task)
+		return NULL;
+
+	mm = get_task_mm(task);
+	put_task_struct(task);
+	return mm;
+}
+
+/*
+ * Size-evenly split a region into 'nr_pieces' small regions
+ *
+ * Returns 0 on success, or negative error code otherwise.
+ */
+static int damon_split_region_evenly(struct damon_ctx *ctx,
+		struct damon_region *r, unsigned int nr_pieces)
+{
+	unsigned long sz_orig, sz_piece, orig_end;
+	struct damon_region *piece = NULL, *next;
+	unsigned long start;
+
+	if (!r || !nr_pieces)
+		return -EINVAL;
+
+	orig_end = r->vm_end;
+	sz_orig = r->vm_end - r->vm_start;
+	sz_piece = sz_orig / nr_pieces;
+
+	if (!sz_piece)
+		return -EINVAL;
+
+	r->vm_end = r->vm_start + sz_piece;
+	next = damon_next_region(r);
+	for (start = r->vm_end; start + sz_piece <= orig_end;
+			start += sz_piece) {
+		piece = damon_new_region(ctx, start, start + sz_piece);
+		damon_insert_region(piece, r, next);
+		r = piece;
+	}
+	/* complement last region for possible rounding error */
+	if (piece)
+		piece->vm_end = orig_end;
+
+	return 0;
+}
+
+struct region {
+	unsigned long start;
+	unsigned long end;
+};
+
+static unsigned long sz_region(struct region *r)
+{
+	return r->end - r->start;
+}
+
+static void swap_regions(struct region *r1, struct region *r2)
+{
+	struct region tmp;
+
+	tmp = *r1;
+	*r1 = *r2;
+	*r2 = tmp;
+}
+
+/*
+ * Find the three regions in an address space
+ *
+ * vma		the head vma of the target address space
+ * regions	an array of three 'struct region's that results will be saved
+ *
+ * This function receives an address space and finds three regions in it which
+ * separated by the two biggest unmapped regions in the space.  Please refer to
+ * below comments of 'damon_init_regions_of()' function to know why this is
+ * necessary.
+ *
+ * Returns 0 if success, or negative error code otherwise.
+ */
+static int damon_three_regions_in_vmas(struct vm_area_struct *vma,
+		struct region regions[3])
+{
+	struct region gap = {0,}, first_gap = {0,}, second_gap = {0,};
+	struct vm_area_struct *last_vma = NULL;
+	unsigned long start = 0;
+
+	/* Find two biggest gaps so that first_gap > second_gap > others */
+	for (; vma; vma = vma->vm_next) {
+		if (!last_vma) {
+			start = vma->vm_start;
+			last_vma = vma;
+			continue;
+		}
+		gap.start = last_vma->vm_end;
+		gap.end = vma->vm_start;
+		if (sz_region(&gap) > sz_region(&second_gap)) {
+			swap_regions(&gap, &second_gap);
+			if (sz_region(&second_gap) > sz_region(&first_gap))
+				swap_regions(&second_gap, &first_gap);
+		}
+		last_vma = vma;
+	}
+
+	if (!sz_region(&second_gap) || !sz_region(&first_gap))
+		return -EINVAL;
+
+	/* Sort the two biggest gaps by address */
+	if (first_gap.start > second_gap.start)
+		swap_regions(&first_gap, &second_gap);
+
+	/* Store the result */
+	regions[0].start = start;
+	regions[0].end = first_gap.start;
+	regions[1].start = first_gap.end;
+	regions[1].end = second_gap.start;
+	regions[2].start = second_gap.end;
+	regions[2].end = last_vma->vm_end;
+
+	return 0;
+}
+
+/*
+ * Get the three regions in the given task
+ *
+ * Returns 0 on success, negative error code otherwise.
+ */
+static int damon_three_regions_of(struct damon_task *t,
+				struct region regions[3])
+{
+	struct mm_struct *mm;
+	int rc;
+
+	mm = damon_get_mm(t);
+	if (!mm)
+		return -EINVAL;
+
+	down_read(&mm->mmap_sem);
+	rc = damon_three_regions_in_vmas(mm->mmap, regions);
+	up_read(&mm->mmap_sem);
+
+	mmput(mm);
+	return rc;
+}
+
+/*
+ * Initialize the monitoring target regions for the given task
+ *
+ * t	the given target task
+ *
+ * Because only a number of small portions of the entire address space
+ * is acutally mapped to the memory and accessed, monitoring the unmapped
+ * regions is wasteful.  That said, because we can deal with small noises,
+ * tracking every mapping is not strictly required but could even incur a high
+ * overhead if the mapping frequently changes or the number of mappings is
+ * high.  Nonetheless, this may seems very weird.  DAMON's dynamic regions
+ * adjustment mechanism, which will be implemented with following commit will
+ * make this more sense.
+ *
+ * For the reason, we convert the complex mappings to three distinct regions
+ * that cover every mapped areas of the address space.  Also the two gaps
+ * between the three regions are the two biggest unmapped areas in the given
+ * address space.  In detail, this function first identifies the start and the
+ * end of the mappings and the two biggest unmapped areas of the address space.
+ * Then, it constructs the three regions as below:
+ *
+ *     [mappings[0]->start, big_two_unmapped_areas[0]->start)
+ *     [big_two_unmapped_areas[0]->end, big_two_unmapped_areas[1]->start)
+ *     [big_two_unmapped_areas[1]->end, mappings[nr_mappings - 1]->end)
+ *
+ * As usual memory map of processes is as below, the gap between the heap and
+ * the uppermost mmap()-ed region, and the gap between the lowermost mmap()-ed
+ * region and the stack will be two biggest unmapped regions.  Because these
+ * gaps are exceptionally huge areas in usual address space, excluding these
+ * two biggest unmapped regions will be sufficient to make a trade-off.
+ *
+ *   <heap>
+ *   <BIG UNMAPPED REGION 1>
+ *   <uppermost mmap()-ed region>
+ *   (other mmap()-ed regions and small unmapped regions)
+ *   <lowermost mmap()-ed region>
+ *   <BIG UNMAPPED REGION 2>
+ *   <stack>
+ */
+static void damon_init_regions_of(struct damon_ctx *c, struct damon_task *t)
+{
+	struct damon_region *r;
+	struct region regions[3];
+	int i;
+
+	if (damon_three_regions_of(t, regions)) {
+		pr_err("Failed to get three regions of task %lu\n", t->pid);
+		return;
+	}
+
+	/* Set the initial three regions of the task */
+	for (i = 0; i < 3; i++) {
+		r = damon_new_region(c, regions[i].start, regions[i].end);
+		damon_add_region(r, t);
+	}
+
+	/* Split the middle region into 'min_nr_regions - 2' regions */
+	r = damon_nth_region_of(t, 1);
+	if (damon_split_region_evenly(c, r, c->min_nr_regions - 2))
+		pr_warn("Init middle region failed to be split\n");
+}
+
+/* Initialize '->regions_list' of every task */
+static void kdamond_init_regions(struct damon_ctx *ctx)
+{
+	struct damon_task *t;
+
+	damon_for_each_task(ctx, t)
+		damon_init_regions_of(ctx, t);
+}
+
+static bool damon_pte_pmd_young(pte_t *pte, pmd_t *pmd)
+{
+	if (pte && pte_young(*pte))
+		return true;
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+	if (pmd && pmd_young(*pmd))
+		return true;
+#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
+	return false;
+}
+
+static void damon_pte_pmd_mkold(pte_t *pte, pmd_t *pmd)
+{
+	if (pte) {
+		if (pte_young(*pte)) {
+			clear_page_idle(pte_page(*pte));
+			set_page_young(pte_page(*pte));
+		}
+		*pte = pte_mkold(*pte);
+		return;
+	}
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+	if (pmd) {
+		if (pmd_young(*pmd)) {
+			clear_page_idle(pmd_page(*pmd));
+			set_page_young(pmd_page(*pmd));
+		}
+		*pmd = pmd_mkold(*pmd);
+	}
+#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
+}
+
+/*
+ * Check whether the region accessed and prepare for next check
+ *
+ * mm	'mm_struct' for the given virtual address space
+ * r	the region to be checked
+ */
+static void kdamond_check_access(struct damon_ctx *ctx,
+			struct mm_struct *mm, struct damon_region *r)
+{
+	static struct mm_struct *last_mm;
+	static unsigned long last_addr;
+	static int last_page_sz = PAGE_SIZE;
+	static bool last_accessed;
+
+	pte_t *pte = NULL;
+	pmd_t *pmd = NULL;
+	spinlock_t *ptl;
+
+	/* If the region is in the last checked page, reuse the result */
+	if (mm == last_mm && (ALIGN_DOWN(last_addr, last_page_sz) ==
+				ALIGN_DOWN(r->sampling_addr, last_page_sz))) {
+		if (last_accessed)
+			r->nr_accesses++;
+		return;
+	}
+
+	if (follow_pte_pmd(mm, r->sampling_addr, NULL, &pte, &pmd, &ptl))
+		goto prepare_next_check;
+
+	/* Read the page table access bit of the page */
+	if (damon_pte_pmd_young(pte, pmd)) {
+		last_accessed = true;
+		r->nr_accesses++;
+	}
+	spin_unlock(ptl);
+
+prepare_next_check:
+	last_mm = mm;
+	last_addr = r->sampling_addr;
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+	last_page_sz = pte ? PAGE_SIZE : ((1UL) << HPAGE_PMD_SHIFT);
+#endif
+
+	r->sampling_addr = damon_rand(ctx, r->vm_start, r->vm_end);
+	pte = NULL, pmd = NULL;
+	if (follow_pte_pmd(mm, r->sampling_addr, NULL, &pte, &pmd, &ptl))
+		return;
+
+	damon_pte_pmd_mkold(pte, pmd);
+	spin_unlock(ptl);
+}
+
+/*
+ * damon_check_reset_time_interval() - Check if a time interval is elapsed.
+ * @baseline:	the time to check whether the interval has elapsed since
+ * @interval:	the time interval (microseconds)
+ *
+ * See whether the given time interval has passed since the given baseline
+ * time.  If so, it also updates the baseline to current time for next check.
+ *
+ * Return:	true if the time interval has passed, or false otherwise.
+ */
+static bool damon_check_reset_time_interval(struct timespec64 *baseline,
+		unsigned long interval)
+{
+	struct timespec64 now;
+
+	ktime_get_coarse_ts64(&now);
+	if ((timespec64_to_ns(&now) - timespec64_to_ns(baseline)) <
+			interval * 1000)
+		return false;
+	*baseline = now;
+	return true;
+}
+
+/*
+ * Check whether it is time to flush the aggregated information
+ */
+static bool kdamond_aggregate_interval_passed(struct damon_ctx *ctx)
+{
+	return damon_check_reset_time_interval(&ctx->last_aggregation,
+			ctx->aggr_interval);
+}
+
+/*
+ * Reset the aggregated monitoring results
+ */
+static void kdamond_reset_aggregated(struct damon_ctx *c)
+{
+	struct damon_task *t;
+	struct damon_region *r;
+
+	damon_for_each_task(c, t) {
+		damon_for_each_region(r, t)
+			r->nr_accesses = 0;
+	}
+}
+
+/*
+ * Check whether current monitoring should be stopped
+ *
+ * If users asked to stop, need stop.  Even though no user has asked to stop,
+ * need stop if every target task has dead.
+ *
+ * Returns true if need to stop current monitoring.
+ */
+static bool kdamond_need_stop(struct damon_ctx *ctx)
+{
+	struct damon_task *t;
+	struct task_struct *task;
+	bool stop;
+
+	stop = kthread_should_stop();
+	if (stop)
+		return true;
+
+	damon_for_each_task(ctx, t) {
+		task = damon_get_task_struct(t);
+		if (task) {
+			put_task_struct(task);
+			return false;
+		}
+	}
+
+	return true;
+}
+
+/*
+ * The monitoring daemon that runs as a kernel thread
+ */
+static int kdamond_fn(void *data)
+{
+	struct damon_ctx *ctx = data;
+	struct damon_task *t;
+	struct damon_region *r, *next;
+	struct mm_struct *mm;
+
+	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
+	kdamond_init_regions(ctx);
+	while (!kdamond_need_stop(ctx)) {
+		damon_for_each_task(ctx, t) {
+			mm = damon_get_mm(t);
+			if (!mm)
+				continue;
+			damon_for_each_region(r, t)
+				kdamond_check_access(ctx, mm, r);
+			mmput(mm);
+		}
+
+		if (kdamond_aggregate_interval_passed(ctx))
+			kdamond_reset_aggregated(ctx);
+
+		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
+	}
+	damon_for_each_task(ctx, t) {
+		damon_for_each_region_safe(r, next, t)
+			damon_destroy_region(r);
+	}
+	pr_debug("kdamond (%d) finishes\n", ctx->kdamond->pid);
+	mutex_lock(&ctx->kdamond_lock);
+	ctx->kdamond = NULL;
+	mutex_unlock(&ctx->kdamond_lock);
+
+	return 0;
+}
+
+/*
+ * Controller functions
+ */
+
+static bool damon_kdamond_running(struct damon_ctx *ctx)
+{
+	bool running;
+
+	mutex_lock(&ctx->kdamond_lock);
+	running = ctx->kdamond != NULL;
+	mutex_unlock(&ctx->kdamond_lock);
+
+	return running;
+}
+
+/*
+ * Start or stop the kdamond
+ *
+ * Returns 0 if success, negative error code otherwise.
+ */
+static int damon_turn_kdamond(struct damon_ctx *ctx, bool on)
+{
+	int err = -EBUSY;
+
+	mutex_lock(&ctx->kdamond_lock);
+	if (!ctx->kdamond && on) {
+		err = 0;
+		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond");
+		if (IS_ERR(ctx->kdamond))
+			err = PTR_ERR(ctx->kdamond);
+	} else if (ctx->kdamond && !on) {
+		mutex_unlock(&ctx->kdamond_lock);
+		kthread_stop(ctx->kdamond);
+		while (damon_kdamond_running(ctx))
+			usleep_range(ctx->sample_interval,
+					ctx->sample_interval * 2);
+		return 0;
+	}
+	mutex_unlock(&ctx->kdamond_lock);
+
+	return err;
+}
+
+/*
+ * damon_start() - Starts monitoring with given context.
+ * @ctx:	monitoring context
+ *
+ * Return: 0 on success, negative error code otherwise.
+ */
+int damon_start(struct damon_ctx *ctx)
+{
+	return damon_turn_kdamond(ctx, true);
+}
+
+/*
+ * damon_stop() - Stops monitoring of given context.
+ * @ctx:	monitoring context
+ *
+ * Return: 0 on success, negative error code otherwise.
+ */
+int damon_stop(struct damon_ctx *ctx)
+{
+	return damon_turn_kdamond(ctx, false);
+}
+
+/*
+ * damon_set_pids() - Set monitoring target processes.
+ * @ctx:	monitoring context
+ * @pids:	array of target processes pids
+ * @nr_pids:	number of entries in @pids
+ *
+ * This function should not be called while the kdamond is running.
+ *
+ * Return: 0 on usccess, negative error code otherwise.
+ */
+int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids)
+{
+	ssize_t i;
+	struct damon_task *t, *next;
+
+	damon_for_each_task_safe(ctx, t, next)
+		damon_destroy_task(t);
+
+	for (i = 0; i < nr_pids; i++) {
+		t = damon_new_task(pids[i]);
+		if (!t) {
+			pr_err("Failed to alloc damon_task\n");
+			return -ENOMEM;
+		}
+		damon_add_task(ctx, t);
+	}
+
+	return 0;
+}
+
+/*
+ * damon_set_attrs() - Set attributes for the monitoring.
+ * @ctx:		monitoring context
+ * @sample_int:		time interval between samplings
+ * @aggr_int:		time interval between aggregations
+ * @min_nr_reg:		minimal number of regions
+ *
+ * This function should not be called while the kdamond is running.
+ * Every time interval is in micro-seconds.
+ *
+ * Return: 0 on success, negative error code otherwise.
+ */
+int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
+		unsigned long aggr_int, unsigned long min_nr_reg)
+{
+	if (min_nr_reg < 3) {
+		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
+				min_nr_reg);
+		return -EINVAL;
+	}
+
+	ctx->sample_interval = sample_int;
+	ctx->aggr_interval = aggr_int;
+	ctx->min_nr_regions = min_nr_reg;
+
+	return 0;
+}
+
 static int __init damon_init(void)
 {
 	return 0;
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 05/15] mm/damon: Adaptively adjust regions
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (3 preceding siblings ...)
  2020-03-18 11:27 ` [PATCH v7 04/15] mm/damon: Implement region based sampling SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-31 16:08   ` Jonathan Cameron
  2020-03-18 11:27 ` [PATCH v7 06/15] mm/damon: Apply dynamic memory mapping changes SeongJae Park
                   ` (10 subsequent siblings)
  15 siblings, 1 reply; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

At the beginning of the monitoring, DAMON constructs the initial regions
by evenly splitting the memory mapped address space of the process into
the user-specified minimal number of regions.  In this initial state,
the assumption of the regions (pages in same region have similar access
frequencies) is normally not kept and thus the monitoring quality could
be low.  To keep the assumption as much as possible, DAMON adaptively
merges and splits each region.

For each ``aggregation interval``, it compares the access frequencies of
adjacent regions and merges those if the frequency difference is small.
Then, after it reports and clears the aggregated access frequency of
each region, it splits each region into two regions if the total number
of regions is smaller than the half of the user-specified maximum number
of regions.

In this way, DAMON provides its best-effort quality and minimal overhead
while keeping the bounds users set for their trade-off.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 include/linux/damon.h |   6 +-
 mm/damon.c            | 148 ++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 145 insertions(+), 9 deletions(-)

diff --git a/include/linux/damon.h b/include/linux/damon.h
index f1945df6e6b4..7562b85b1ec0 100644
--- a/include/linux/damon.h
+++ b/include/linux/damon.h
@@ -42,6 +42,7 @@ struct damon_ctx {
 	unsigned long sample_interval;
 	unsigned long aggr_interval;
 	unsigned long min_nr_regions;
+	unsigned long max_nr_regions;
 
 	struct timespec64 last_aggregation;
 
@@ -54,8 +55,9 @@ struct damon_ctx {
 };
 
 int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids);
-int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
-		unsigned long aggr_int, unsigned long min_nr_reg);
+int damon_set_attrs(struct damon_ctx *ctx,
+		unsigned long sample_int, unsigned long aggr_int,
+		unsigned long min_nr_reg, unsigned long max_nr_reg);
 int damon_start(struct damon_ctx *ctx);
 int damon_stop(struct damon_ctx *ctx);
 
diff --git a/mm/damon.c b/mm/damon.c
index 018016793555..23c0de3b502e 100644
--- a/mm/damon.c
+++ b/mm/damon.c
@@ -342,9 +342,12 @@ static int damon_three_regions_of(struct damon_task *t,
  * regions is wasteful.  That said, because we can deal with small noises,
  * tracking every mapping is not strictly required but could even incur a high
  * overhead if the mapping frequently changes or the number of mappings is
- * high.  Nonetheless, this may seems very weird.  DAMON's dynamic regions
- * adjustment mechanism, which will be implemented with following commit will
- * make this more sense.
+ * high.  The adaptive regions adjustment mechanism will further help to deal
+ * with the noises by simply identifying the unmapped areas as a region that
+ * has no access.  Moreover, applying the real mappings that would have many
+ * unmapped areas inside will make the adaptive mechanism quite complex.  That
+ * said, too huge unmapped areas inside the monitoring target should be removed
+ * to not take the time for the adaptive mechanism.
  *
  * For the reason, we convert the complex mappings to three distinct regions
  * that cover every mapped areas of the address space.  Also the two gaps
@@ -533,6 +536,121 @@ static void kdamond_reset_aggregated(struct damon_ctx *c)
 	}
 }
 
+#define sz_damon_region(r) (r->vm_end - r->vm_start)
+
+/*
+ * Merge two adjacent regions into one region
+ */
+static void damon_merge_two_regions(struct damon_region *l,
+				struct damon_region *r)
+{
+	l->nr_accesses = (l->nr_accesses * sz_damon_region(l) +
+			r->nr_accesses * sz_damon_region(r)) /
+			(sz_damon_region(l) + sz_damon_region(r));
+	l->vm_end = r->vm_end;
+	damon_destroy_region(r);
+}
+
+#define diff_of(a, b) (a > b ? a - b : b - a)
+
+/*
+ * Merge adjacent regions having similar access frequencies
+ *
+ * t		task that merge operation will make change
+ * thres	merge regions having '->nr_accesses' diff smaller than this
+ */
+static void damon_merge_regions_of(struct damon_task *t, unsigned int thres)
+{
+	struct damon_region *r, *prev = NULL, *next;
+
+	damon_for_each_region_safe(r, next, t) {
+		if (!prev || prev->vm_end != r->vm_start ||
+		    diff_of(prev->nr_accesses, r->nr_accesses) > thres) {
+			prev = r;
+			continue;
+		}
+		damon_merge_two_regions(prev, r);
+	}
+}
+
+/*
+ * Merge adjacent regions having similar access frequencies
+ *
+ * threshold	merge regions havind nr_accesses diff larger than this
+ *
+ * This function merges monitoring target regions which are adjacent and their
+ * access frequencies are similar.  This is for minimizing the monitoring
+ * overhead under the dynamically changeable access pattern.  If a merge was
+ * unnecessarily made, later 'kdamond_split_regions()' will revert it.
+ */
+static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold)
+{
+	struct damon_task *t;
+
+	damon_for_each_task(c, t)
+		damon_merge_regions_of(t, threshold);
+}
+
+/*
+ * Split a region into two small regions
+ *
+ * r		the region to be split
+ * sz_r		size of the first sub-region that will be made
+ */
+static void damon_split_region_at(struct damon_ctx *ctx,
+		struct damon_region *r, unsigned long sz_r)
+{
+	struct damon_region *new;
+
+	new = damon_new_region(ctx, r->vm_start + sz_r, r->vm_end);
+	r->vm_end = new->vm_start;
+
+	damon_insert_region(new, r, damon_next_region(r));
+}
+
+static void damon_split_regions_of(struct damon_ctx *ctx, struct damon_task *t)
+{
+	struct damon_region *r, *next;
+	unsigned long sz_left_region;
+
+	damon_for_each_region_safe(r, next, t) {
+		/*
+		 * Randomly select size of left sub-region to be at least
+		 * 10 percent and at most 90% of original region
+		 */
+		sz_left_region = (prandom_u32_state(&ctx->rndseed) % 9 + 1) *
+			(r->vm_end - r->vm_start) / 10;
+		/* Do not allow blank region */
+		if (sz_left_region == 0)
+			continue;
+		damon_split_region_at(ctx, r, sz_left_region);
+	}
+}
+
+/*
+ * splits every target regions into two randomly-sized regions
+ *
+ * This function splits every target regions into two random-sized regions if
+ * current total number of the regions is smaller than the half of the
+ * user-specified maximum number of regions.  This is for maximizing the
+ * monitoring accuracy under the dynamically changeable access patterns.  If a
+ * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
+ * it.
+ */
+static void kdamond_split_regions(struct damon_ctx *ctx)
+{
+	struct damon_task *t;
+	unsigned int nr_regions = 0;
+
+	damon_for_each_task(ctx, t)
+		nr_regions += nr_damon_regions(t);
+	if (nr_regions > ctx->max_nr_regions / 2)
+		return;
+
+	damon_for_each_task(ctx, t)
+		damon_split_regions_of(ctx, t);
+}
+
 /*
  * Check whether current monitoring should be stopped
  *
@@ -571,21 +689,29 @@ static int kdamond_fn(void *data)
 	struct damon_task *t;
 	struct damon_region *r, *next;
 	struct mm_struct *mm;
+	unsigned int max_nr_accesses;
 
 	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
 	kdamond_init_regions(ctx);
 	while (!kdamond_need_stop(ctx)) {
+		max_nr_accesses = 0;
 		damon_for_each_task(ctx, t) {
 			mm = damon_get_mm(t);
 			if (!mm)
 				continue;
-			damon_for_each_region(r, t)
+			damon_for_each_region(r, t) {
 				kdamond_check_access(ctx, mm, r);
+				max_nr_accesses = max(r->nr_accesses,
+						max_nr_accesses);
+			}
 			mmput(mm);
 		}
 
-		if (kdamond_aggregate_interval_passed(ctx))
+		if (kdamond_aggregate_interval_passed(ctx)) {
+			kdamond_merge_regions(ctx, max_nr_accesses / 10);
 			kdamond_reset_aggregated(ctx);
+			kdamond_split_regions(ctx);
+		}
 
 		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
 	}
@@ -702,24 +828,32 @@ int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids)
  * @sample_int:		time interval between samplings
  * @aggr_int:		time interval between aggregations
  * @min_nr_reg:		minimal number of regions
+ * @max_nr_reg:		maximum number of regions
  *
  * This function should not be called while the kdamond is running.
  * Every time interval is in micro-seconds.
  *
  * Return: 0 on success, negative error code otherwise.
  */
-int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
-		unsigned long aggr_int, unsigned long min_nr_reg)
+int damon_set_attrs(struct damon_ctx *ctx,
+			unsigned long sample_int, unsigned long aggr_int,
+			unsigned long min_nr_reg, unsigned long max_nr_reg)
 {
 	if (min_nr_reg < 3) {
 		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
 				min_nr_reg);
 		return -EINVAL;
 	}
+	if (min_nr_reg >= ctx->max_nr_regions) {
+		pr_err("invalid nr_regions.  min (%lu) >= max (%lu)\n",
+				min_nr_reg, max_nr_reg);
+		return -EINVAL;
+	}
 
 	ctx->sample_interval = sample_int;
 	ctx->aggr_interval = aggr_int;
 	ctx->min_nr_regions = min_nr_reg;
+	ctx->max_nr_regions = max_nr_reg;
 
 	return 0;
 }
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 06/15] mm/damon: Apply dynamic memory mapping changes
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (4 preceding siblings ...)
  2020-03-18 11:27 ` [PATCH v7 05/15] mm/damon: Adaptively adjust regions SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 07/15] mm/damon: Implement callbacks SeongJae Park
                   ` (9 subsequent siblings)
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

Only a number of parts in the virtual address space of the processes is
mapped to physical memory and accessed.  Thus, tracking the unmapped
address regions is just wasteful.  However, tracking every memory
mapping change might incur an overhead.  For the reason, DAMON applies
the dynamic memory mapping changes to the tracking regions only for each
of a user-specified time interval (``regions update interval``).

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 include/linux/damon.h | 10 +++--
 mm/damon.c            | 92 +++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 96 insertions(+), 6 deletions(-)

diff --git a/include/linux/damon.h b/include/linux/damon.h
index 7562b85b1ec0..64e327400749 100644
--- a/include/linux/damon.h
+++ b/include/linux/damon.h
@@ -34,17 +34,21 @@ struct damon_task {
 /*
  * For each 'sample_interval', DAMON checks whether each region is accessed or
  * not.  It aggregates and keeps the access information (number of accesses to
- * each region) for each 'aggr_interval' time.
+ * each region) for 'aggr_interval' time.  DAMON also checks whether the memory
+ * mapping of the target tasks has changed (e.g., by mmap() calls from the
+ * application) and applies the changes. for each 'regions_update_interval'.
  *
  * All time intervals are in micro-seconds.
  */
 struct damon_ctx {
 	unsigned long sample_interval;
 	unsigned long aggr_interval;
+	unsigned long regions_update_interval;
 	unsigned long min_nr_regions;
 	unsigned long max_nr_regions;
 
 	struct timespec64 last_aggregation;
+	struct timespec64 last_regions_update;
 
 	struct task_struct *kdamond;
 	struct mutex kdamond_lock;
@@ -55,8 +59,8 @@ struct damon_ctx {
 };
 
 int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids);
-int damon_set_attrs(struct damon_ctx *ctx,
-		unsigned long sample_int, unsigned long aggr_int,
+int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
+		unsigned long aggr_int, unsigned long regions_update_int,
 		unsigned long min_nr_reg, unsigned long max_nr_reg);
 int damon_start(struct damon_ctx *ctx);
 int damon_stop(struct damon_ctx *ctx);
diff --git a/mm/damon.c b/mm/damon.c
index 23c0de3b502e..535337529d20 100644
--- a/mm/damon.c
+++ b/mm/damon.c
@@ -651,6 +651,87 @@ static void kdamond_split_regions(struct damon_ctx *ctx)
 		damon_split_regions_of(ctx, t);
 }
 
+/*
+ * Check whether it is time to check and apply the dynamic mmap changes
+ *
+ * Returns true if it is.
+ */
+static bool kdamond_need_update_regions(struct damon_ctx *ctx)
+{
+	return damon_check_reset_time_interval(&ctx->last_regions_update,
+			ctx->regions_update_interval);
+}
+
+static bool damon_intersect(struct damon_region *r, struct region *re)
+{
+	return !(r->vm_end <= re->start || re->end <= r->vm_start);
+}
+
+/*
+ * Update damon regions for the three big regions of the given task
+ *
+ * t		the given task
+ * bregions	the three big regions of the task
+ */
+static void damon_apply_three_regions(struct damon_ctx *ctx,
+		struct damon_task *t, struct region bregions[3])
+{
+	struct damon_region *r, *next;
+	unsigned int i = 0;
+
+	/* Remove regions which isn't in the three big regions now */
+	damon_for_each_region_safe(r, next, t) {
+		for (i = 0; i < 3; i++) {
+			if (damon_intersect(r, &bregions[i]))
+				break;
+		}
+		if (i == 3)
+			damon_destroy_region(r);
+	}
+
+	/* Adjust intersecting regions to fit with the three big regions */
+	for (i = 0; i < 3; i++) {
+		struct damon_region *first = NULL, *last;
+		struct damon_region *newr;
+		struct region *br;
+
+		br = &bregions[i];
+		/* Get the first and last regions which intersects with br */
+		damon_for_each_region(r, t) {
+			if (damon_intersect(r, br)) {
+				if (!first)
+					first = r;
+				last = r;
+			}
+			if (r->vm_start >= br->end)
+				break;
+		}
+		if (!first) {
+			/* no damon_region intersects with this big region */
+			newr = damon_new_region(ctx, br->start, br->end);
+			damon_insert_region(newr, damon_prev_region(r), r);
+		} else {
+			first->vm_start = br->start;
+			last->vm_end = br->end;
+		}
+	}
+}
+
+/*
+ * Update regions for current memory mappings
+ */
+static void kdamond_update_regions(struct damon_ctx *ctx)
+{
+	struct region three_regions[3];
+	struct damon_task *t;
+
+	damon_for_each_task(ctx, t) {
+		if (damon_three_regions_of(t, three_regions))
+			continue;
+		damon_apply_three_regions(ctx, t, three_regions);
+	}
+}
+
 /*
  * Check whether current monitoring should be stopped
  *
@@ -713,6 +794,9 @@ static int kdamond_fn(void *data)
 			kdamond_split_regions(ctx);
 		}
 
+		if (kdamond_need_update_regions(ctx))
+			kdamond_update_regions(ctx);
+
 		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
 	}
 	damon_for_each_task(ctx, t) {
@@ -826,6 +910,7 @@ int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids)
  * damon_set_attrs() - Set attributes for the monitoring.
  * @ctx:		monitoring context
  * @sample_int:		time interval between samplings
+ * @regions_update_int:	time interval between vma update checks
  * @aggr_int:		time interval between aggregations
  * @min_nr_reg:		minimal number of regions
  * @max_nr_reg:		maximum number of regions
@@ -835,9 +920,9 @@ int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids)
  *
  * Return: 0 on success, negative error code otherwise.
  */
-int damon_set_attrs(struct damon_ctx *ctx,
-			unsigned long sample_int, unsigned long aggr_int,
-			unsigned long min_nr_reg, unsigned long max_nr_reg)
+int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
+		unsigned long aggr_int, unsigned long regions_update_int,
+		unsigned long min_nr_reg, unsigned long max_nr_reg)
 {
 	if (min_nr_reg < 3) {
 		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
@@ -852,6 +937,7 @@ int damon_set_attrs(struct damon_ctx *ctx,
 
 	ctx->sample_interval = sample_int;
 	ctx->aggr_interval = aggr_int;
+	ctx->regions_update_interval = regions_update_int;
 	ctx->min_nr_regions = min_nr_reg;
 	ctx->max_nr_regions = max_nr_reg;
 
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 07/15] mm/damon: Implement callbacks
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (5 preceding siblings ...)
  2020-03-18 11:27 ` [PATCH v7 06/15] mm/damon: Apply dynamic memory mapping changes SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 08/15] mm/damon: Implement access pattern recording SeongJae Park
                   ` (8 subsequent siblings)
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit implements callbacks for DAMON.  Using this, DAMON users can
install their callbacks for each step of the access monitoring so that
they can do something interesting with the monitored access patterns
online.  For example, callbacks can report the monitored patterns to
users or do some access pattern based memory management such as
proactive reclamations or access pattern based THP promotions/demotions
decision makings.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 include/linux/damon.h | 4 ++++
 mm/damon.c            | 4 ++++
 2 files changed, 8 insertions(+)

diff --git a/include/linux/damon.h b/include/linux/damon.h
index 64e327400749..cad3ee3d78db 100644
--- a/include/linux/damon.h
+++ b/include/linux/damon.h
@@ -56,6 +56,10 @@ struct damon_ctx {
 	struct rnd_state rndseed;
 
 	struct list_head tasks_list;	/* 'damon_task' objects */
+
+	/* callbacks */
+	void (*sample_cb)(struct damon_ctx *context);
+	void (*aggregate_cb)(struct damon_ctx *context);
 };
 
 int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids);
diff --git a/mm/damon.c b/mm/damon.c
index 535337529d20..2e8af1ce4389 100644
--- a/mm/damon.c
+++ b/mm/damon.c
@@ -787,9 +787,13 @@ static int kdamond_fn(void *data)
 			}
 			mmput(mm);
 		}
+		if (ctx->sample_cb)
+			ctx->sample_cb(ctx);
 
 		if (kdamond_aggregate_interval_passed(ctx)) {
 			kdamond_merge_regions(ctx, max_nr_accesses / 10);
+			if (ctx->aggregate_cb)
+				ctx->aggregate_cb(ctx);
 			kdamond_reset_aggregated(ctx);
 			kdamond_split_regions(ctx);
 		}
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 08/15] mm/damon: Implement access pattern recording
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (6 preceding siblings ...)
  2020-03-18 11:27 ` [PATCH v7 07/15] mm/damon: Implement callbacks SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 09/15] mm/damon: Add debugfs interface SeongJae Park
                   ` (7 subsequent siblings)
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit implements the recording feature of DAMON. If this feature
is enabled, DAMON writes the monitored access patterns in its binary
format into a file which specified by the user. This is already able to
be implemented by each user using the callbacks.  However, as the
recording is expected to be used widely, this commit implements the
feature in the DAMON, for more convenience and efficiency.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 include/linux/damon.h |   7 +++
 mm/damon.c            | 120 ++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 124 insertions(+), 3 deletions(-)

diff --git a/include/linux/damon.h b/include/linux/damon.h
index cad3ee3d78db..47fb0ec03030 100644
--- a/include/linux/damon.h
+++ b/include/linux/damon.h
@@ -50,6 +50,11 @@ struct damon_ctx {
 	struct timespec64 last_aggregation;
 	struct timespec64 last_regions_update;
 
+	unsigned char *rbuf;
+	unsigned int rbuf_len;
+	unsigned int rbuf_offset;
+	char *rfile_path;
+
 	struct task_struct *kdamond;
 	struct mutex kdamond_lock;
 
@@ -66,6 +71,8 @@ int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids);
 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
 		unsigned long aggr_int, unsigned long regions_update_int,
 		unsigned long min_nr_reg, unsigned long max_nr_reg);
+int damon_set_recording(struct damon_ctx *ctx,
+				unsigned int rbuf_len, char *rfile_path);
 int damon_start(struct damon_ctx *ctx);
 int damon_stop(struct damon_ctx *ctx);
 
diff --git a/mm/damon.c b/mm/damon.c
index 2e8af1ce4389..59d49b0990cf 100644
--- a/mm/damon.c
+++ b/mm/damon.c
@@ -41,6 +41,8 @@
 #define damon_for_each_task_safe(ctx, t, next) \
 	list_for_each_entry_safe(t, next, &(ctx)->tasks_list, list)
 
+#define MAX_RFILE_PATH_LEN	256
+
 /* Get a random number in [l, r) */
 #define damon_rand(ctx, l, r) (l + prandom_u32_state(&ctx->rndseed) % (r - l))
 
@@ -523,16 +525,81 @@ static bool kdamond_aggregate_interval_passed(struct damon_ctx *ctx)
 }
 
 /*
- * Reset the aggregated monitoring results
+ * Flush the content in the result buffer to the result file
+ */
+static void damon_flush_rbuffer(struct damon_ctx *ctx)
+{
+	ssize_t sz;
+	loff_t pos;
+	struct file *rfile;
+
+	while (ctx->rbuf_offset) {
+		pos = 0;
+		rfile = filp_open(ctx->rfile_path, O_CREAT | O_RDWR | O_APPEND,
+				0644);
+		if (IS_ERR(rfile)) {
+			pr_err("Cannot open the result file %s\n",
+					ctx->rfile_path);
+			return;
+		}
+
+		sz = kernel_write(rfile, ctx->rbuf, ctx->rbuf_offset, &pos);
+		filp_close(rfile, NULL);
+
+		ctx->rbuf_offset -= sz;
+	}
+}
+
+/*
+ * Write a data into the result buffer
+ */
+static void damon_write_rbuf(struct damon_ctx *ctx, void *data, ssize_t size)
+{
+	if (!ctx->rbuf_len || !ctx->rbuf)
+		return;
+	if (ctx->rbuf_offset + size > ctx->rbuf_len)
+		damon_flush_rbuffer(ctx);
+
+	memcpy(&ctx->rbuf[ctx->rbuf_offset], data, size);
+	ctx->rbuf_offset += size;
+}
+
+/*
+ * Flush the aggregated monitoring results to the result buffer
+ *
+ * Stores current tracking results to the result buffer and reset 'nr_accesses'
+ * of each regions.  The format for the result buffer is as below:
+ *
+ *   <time> <number of tasks> <array of task infos>
+ *
+ *   task info: <pid> <number of regions> <array of region infos>
+ *   region info: <start address> <end address> <nr_accesses>
  */
 static void kdamond_reset_aggregated(struct damon_ctx *c)
 {
 	struct damon_task *t;
-	struct damon_region *r;
+	struct timespec64 now;
+	unsigned int nr;
+
+	ktime_get_coarse_ts64(&now);
+
+	damon_write_rbuf(c, &now, sizeof(struct timespec64));
+	nr = nr_damon_tasks(c);
+	damon_write_rbuf(c, &nr, sizeof(nr));
 
 	damon_for_each_task(c, t) {
-		damon_for_each_region(r, t)
+		struct damon_region *r;
+
+		damon_write_rbuf(c, &t->pid, sizeof(t->pid));
+		nr = nr_damon_regions(t);
+		damon_write_rbuf(c, &nr, sizeof(nr));
+		damon_for_each_region(r, t) {
+			damon_write_rbuf(c, &r->vm_start, sizeof(r->vm_start));
+			damon_write_rbuf(c, &r->vm_end, sizeof(r->vm_end));
+			damon_write_rbuf(c, &r->nr_accesses,
+					sizeof(r->nr_accesses));
 			r->nr_accesses = 0;
+		}
 	}
 }
 
@@ -803,6 +870,7 @@ static int kdamond_fn(void *data)
 
 		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
 	}
+	damon_flush_rbuffer(ctx);
 	damon_for_each_task(ctx, t) {
 		damon_for_each_region_safe(r, next, t)
 			damon_destroy_region(r);
@@ -910,6 +978,52 @@ int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids)
 	return 0;
 }
 
+/*
+ * damon_set_recording() - Set attributes for the recording.
+ * @ctx:	target kdamond context
+ * @rbuf_len:	length of the result buffer
+ * @rfile_path:	path to the monitor result files
+ *
+ * Setting 'rbuf_len' 0 disables recording.
+ *
+ * This function should not be called while the kdamond is running.
+ *
+ * Return: 0 on success, negative error code otherwise.
+ */
+int damon_set_recording(struct damon_ctx *ctx,
+				unsigned int rbuf_len, char *rfile_path)
+{
+	size_t rfile_path_len;
+
+	if (rbuf_len > 4 * 1024 * 1024) {
+		pr_err("too long (>%d) result buffer length\n",
+				4 * 1024 * 1024);
+		return -EINVAL;
+	}
+	rfile_path_len = strnlen(rfile_path, MAX_RFILE_PATH_LEN);
+	if (rfile_path_len >= MAX_RFILE_PATH_LEN) {
+		pr_err("too long (>%d) result file path %s\n",
+				MAX_RFILE_PATH_LEN, rfile_path);
+		return -EINVAL;
+	}
+	ctx->rbuf_len = rbuf_len;
+	kfree(ctx->rbuf);
+	kfree(ctx->rfile_path);
+	ctx->rfile_path = NULL;
+	if (!rbuf_len) {
+		ctx->rbuf = NULL;
+	} else {
+		ctx->rbuf = kvmalloc(rbuf_len, GFP_KERNEL);
+		if (!ctx->rbuf)
+			return -ENOMEM;
+	}
+	ctx->rfile_path = kmalloc(rfile_path_len + 1, GFP_KERNEL);
+	if (!ctx->rfile_path)
+		return -ENOMEM;
+	strncpy(ctx->rfile_path, rfile_path, rfile_path_len + 1);
+	return 0;
+}
+
 /*
  * damon_set_attrs() - Set attributes for the monitoring.
  * @ctx:		monitoring context
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 09/15] mm/damon: Add debugfs interface
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (7 preceding siblings ...)
  2020-03-18 11:27 ` [PATCH v7 08/15] mm/damon: Implement access pattern recording SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 10/15] mm/damon: Add tracepoints SeongJae Park
                   ` (6 subsequent siblings)
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit adds a debugfs interface for DAMON.

DAMON exports four files, ``attrs``, ``pids``, ``record``, and
``monitor_on`` under its debugfs directory, ``<debugfs>/damon/``.

Attributes
----------

Users can read and write the ``sampling interval``, ``aggregation
interval``, ``regions update interval``, and min/max number of
monitoring target regions by reading from and writing to the ``attrs``
file.  For example, below commands set those values to 5 ms, 100 ms,
1,000 ms, 10, 1000 and check it again::

    # cd <debugfs>/damon
    # echo 5000 100000 1000000 10 1000 > attrs
    # cat attrs
    5000 100000 1000000 10 1000

Target PIDs
-----------

Users can read and write the pids of current monitoring target processes
by reading from and writing to the ``pids`` file.  For example, below
commands set processes having pids 42 and 4242 as the processes to be
monitored and check it again::

    # cd <debugfs>/damon
    # echo 42 4242 > pids
    # cat pids
    42 4242

Note that setting the pids doesn't starts the monitoring.

Record
------

DAMON support direct monitoring result record feature.  The recorded
results are first written to a buffer and flushed to a file in batch.
Users can set the size of the buffer and the path to the result file by
reading from and writing to the ``record`` file.  For example, below
commands set the buffer to be 4 KiB and the result to be saved in
'/damon.data'.

    # cd <debugfs>/damon
    # echo 4096 /damon.data > pids
    # cat record
    4096 /damon.data

Turning On/Off
--------------

You can check current status, start and stop the monitoring by reading
from and writing to the ``monitor_on`` file.  Writing ``on`` to the file
starts DAMON to monitor the target processes with the attributes.
Writing ``off`` to the file stops DAMON.  DAMON also stops if every
target processes is be terminated.  Below example commands turn on, off,
and check status of DAMON::

    # cd <debugfs>/damon
    # echo on > monitor_on
    # echo off > monitor_on
    # cat monitor_on
    off

Please note that you cannot write to the ``attrs`` and ``pids`` files
while the monitoring is turned on.  If you write to the files while
DAMON is running, ``-EINVAL`` will be returned.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 mm/damon.c | 354 ++++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 352 insertions(+), 2 deletions(-)

diff --git a/mm/damon.c b/mm/damon.c
index 59d49b0990cf..b77e537e2ffe 100644
--- a/mm/damon.c
+++ b/mm/damon.c
@@ -10,6 +10,7 @@
 #define pr_fmt(fmt) "damon: " fmt
 
 #include <linux/damon.h>
+#include <linux/debugfs.h>
 #include <linux/delay.h>
 #include <linux/kthread.h>
 #include <linux/mm.h>
@@ -46,6 +47,15 @@
 /* Get a random number in [l, r) */
 #define damon_rand(ctx, l, r) (l + prandom_u32_state(&ctx->rndseed) % (r - l))
 
+/* A monitoring context for debugfs interface users. */
+static struct damon_ctx damon_user_ctx = {
+	.sample_interval = 5 * 1000,
+	.aggr_interval = 100 * 1000,
+	.regions_update_interval = 1000 * 1000,
+	.min_nr_regions = 10,
+	.max_nr_regions = 1000,
+};
+
 /*
  * Construct a damon_region struct
  *
@@ -1062,14 +1072,354 @@ int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
 	return 0;
 }
 
-static int __init damon_init(void)
+static ssize_t debugfs_monitor_on_read(struct file *file,
+		char __user *buf, size_t count, loff_t *ppos)
+{
+	struct damon_ctx *ctx = &damon_user_ctx;
+	char monitor_on_buf[5];
+	bool monitor_on;
+	int len;
+
+	monitor_on = damon_kdamond_running(ctx);
+	len = snprintf(monitor_on_buf, 5, monitor_on ? "on\n" : "off\n");
+
+	return simple_read_from_buffer(buf, count, ppos, monitor_on_buf, len);
+}
+
+static ssize_t debugfs_monitor_on_write(struct file *file,
+		const char __user *buf, size_t count, loff_t *ppos)
+{
+	struct damon_ctx *ctx = &damon_user_ctx;
+	ssize_t ret;
+	char cmdbuf[5];
+	int err;
+
+	ret = simple_write_to_buffer(cmdbuf, 5, ppos, buf, count);
+	if (ret < 0)
+		return ret;
+
+	if (sscanf(cmdbuf, "%s", cmdbuf) != 1)
+		return -EINVAL;
+	if (!strncmp(cmdbuf, "on", 5))
+		err = damon_start(ctx);
+	else if (!strncmp(cmdbuf, "off", 5))
+		err = damon_stop(ctx);
+	else
+		return -EINVAL;
+
+	if (err)
+		ret = err;
+	return ret;
+}
+
+static ssize_t damon_sprint_pids(struct damon_ctx *ctx, char *buf, ssize_t len)
+{
+	struct damon_task *t;
+	int written = 0;
+	int rc;
+
+	damon_for_each_task(ctx, t) {
+		rc = snprintf(&buf[written], len - written, "%lu ", t->pid);
+		if (!rc)
+			return -ENOMEM;
+		written += rc;
+	}
+	if (written)
+		written -= 1;
+	written += snprintf(&buf[written], len - written, "\n");
+	return written;
+}
+
+static ssize_t debugfs_pids_read(struct file *file,
+		char __user *buf, size_t count, loff_t *ppos)
+{
+	struct damon_ctx *ctx = &damon_user_ctx;
+	ssize_t len;
+	char pids_buf[320];
+
+	len = damon_sprint_pids(ctx, pids_buf, 320);
+	if (len < 0)
+		return len;
+
+	return simple_read_from_buffer(buf, count, ppos, pids_buf, len);
+}
+
+/*
+ * Converts a string into an array of unsigned long integers
+ *
+ * Returns an array of unsigned long integers if the conversion success, or
+ * NULL otherwise.
+ */
+static unsigned long *str_to_pids(const char *str, ssize_t len,
+				ssize_t *nr_pids)
+{
+	unsigned long *pids;
+	const int max_nr_pids = 32;
+	unsigned long pid;
+	int pos = 0, parsed, ret;
+
+	*nr_pids = 0;
+	pids = kmalloc_array(max_nr_pids, sizeof(pid), GFP_KERNEL);
+	if (!pids)
+		return NULL;
+	while (*nr_pids < max_nr_pids && pos < len) {
+		ret = sscanf(&str[pos], "%lu%n", &pid, &parsed);
+		pos += parsed;
+		if (ret != 1)
+			break;
+		pids[*nr_pids] = pid;
+		*nr_pids += 1;
+	}
+	if (*nr_pids == 0) {
+		kfree(pids);
+		pids = NULL;
+	}
+
+	return pids;
+}
+
+static ssize_t debugfs_pids_write(struct file *file,
+		const char __user *buf, size_t count, loff_t *ppos)
 {
+	struct damon_ctx *ctx = &damon_user_ctx;
+	char *kbuf;
+	unsigned long *targets;
+	ssize_t nr_targets;
+	ssize_t ret;
+	int err;
+
+	kbuf = kmalloc(count, GFP_KERNEL);
+	if (!kbuf)
+		return -ENOMEM;
+
+	ret = simple_write_to_buffer(kbuf, count, ppos, buf, count);
+	if (ret < 0)
+		goto out;
+
+	targets = str_to_pids(kbuf, ret, &nr_targets);
+	if (!targets) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	mutex_lock(&ctx->kdamond_lock);
+	if (ctx->kdamond) {
+		ret= -EINVAL;
+		goto unlock_out;
+	}
+
+	err = damon_set_pids(ctx, targets, nr_targets);
+	if (err)
+		ret = err;
+unlock_out:
+	mutex_unlock(&ctx->kdamond_lock);
+	kfree(targets);
+out:
+	kfree(kbuf);
+	return ret;
+}
+
+static ssize_t debugfs_record_read(struct file *file,
+		char __user *buf, size_t count, loff_t *ppos)
+{
+	struct damon_ctx *ctx = &damon_user_ctx;
+	char record_buf[20 + MAX_RFILE_PATH_LEN];
+	int ret;
+
+	ret = snprintf(record_buf, ARRAY_SIZE(record_buf), "%u %s\n",
+			ctx->rbuf_len, ctx->rfile_path);
+	return simple_read_from_buffer(buf, count, ppos, record_buf, ret);
+}
+
+static ssize_t debugfs_record_write(struct file *file,
+		const char __user *buf, size_t count, loff_t *ppos)
+{
+	struct damon_ctx *ctx = &damon_user_ctx;
+	char *kbuf;
+	unsigned int rbuf_len;
+	char rfile_path[MAX_RFILE_PATH_LEN];
+	ssize_t ret;
+	int err;
+
+	kbuf = kmalloc(count + 1, GFP_KERNEL);
+	if (!kbuf)
+		return -ENOMEM;
+	kbuf[count] = '\0';
+
+	ret = simple_write_to_buffer(kbuf, count, ppos, buf, count);
+	if (ret < 0)
+		goto out;
+	if (sscanf(kbuf, "%u %s",
+				&rbuf_len, rfile_path) != 2) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	mutex_lock(&ctx->kdamond_lock);
+	if (ctx->kdamond) {
+		ret = -EBUSY;
+		goto unlock_out;
+	}
+
+	err = damon_set_recording(ctx, rbuf_len, rfile_path);
+	if (err)
+		ret = err;
+unlock_out:
+	mutex_unlock(&ctx->kdamond_lock);
+out:
+	kfree(kbuf);
+	return ret;
+}
+
+
+static ssize_t debugfs_attrs_read(struct file *file,
+		char __user *buf, size_t count, loff_t *ppos)
+{
+	struct damon_ctx *ctx = &damon_user_ctx;
+	char kbuf[128];
+	int ret;
+
+	ret = snprintf(kbuf, ARRAY_SIZE(kbuf), "%lu %lu %lu %lu %lu\n",
+			ctx->sample_interval, ctx->aggr_interval,
+			ctx->regions_update_interval, ctx->min_nr_regions,
+			ctx->max_nr_regions);
+
+	return simple_read_from_buffer(buf, count, ppos, kbuf, ret);
+}
+
+static ssize_t debugfs_attrs_write(struct file *file,
+		const char __user *buf, size_t count, loff_t *ppos)
+{
+	struct damon_ctx *ctx = &damon_user_ctx;
+	unsigned long s, a, r, minr, maxr;
+	char *kbuf;
+	ssize_t ret;
+	int err;
+
+	kbuf = kmalloc(count, GFP_KERNEL);
+	if (!kbuf)
+		return -ENOMEM;
+
+	ret = simple_write_to_buffer(kbuf, count, ppos, buf, count);
+	if (ret < 0)
+		goto out;
+
+	if (sscanf(kbuf, "%lu %lu %lu %lu %lu",
+				&s, &a, &r, &minr, &maxr) != 5) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	mutex_lock(&ctx->kdamond_lock);
+	if (ctx->kdamond) {
+		ret = -EBUSY;
+		goto unlock_out;
+	}
+
+	err = damon_set_attrs(ctx, s, a, r, minr, maxr);
+	if (err)
+		ret = err;
+unlock_out:
+	mutex_unlock(&ctx->kdamond_lock);
+out:
+	kfree(kbuf);
+	return ret;
+}
+
+static const struct file_operations monitor_on_fops = {
+	.owner = THIS_MODULE,
+	.read = debugfs_monitor_on_read,
+	.write = debugfs_monitor_on_write,
+};
+
+static const struct file_operations pids_fops = {
+	.owner = THIS_MODULE,
+	.read = debugfs_pids_read,
+	.write = debugfs_pids_write,
+};
+
+static const struct file_operations record_fops = {
+	.owner = THIS_MODULE,
+	.read = debugfs_record_read,
+	.write = debugfs_record_write,
+};
+
+static const struct file_operations attrs_fops = {
+	.owner = THIS_MODULE,
+	.read = debugfs_attrs_read,
+	.write = debugfs_attrs_write,
+};
+
+static struct dentry *debugfs_root;
+
+static int __init damon_debugfs_init(void)
+{
+	const char * const file_names[] = {"attrs", "record",
+		"pids", "monitor_on"};
+	const struct file_operations *fops[] = {&attrs_fops, &record_fops,
+		&pids_fops, &monitor_on_fops};
+	int i;
+
+	debugfs_root = debugfs_create_dir("damon", NULL);
+	if (!debugfs_root) {
+		pr_err("failed to create the debugfs dir\n");
+		return -ENOMEM;
+	}
+
+	for (i = 0; i < ARRAY_SIZE(file_names); i++) {
+		if (!debugfs_create_file(file_names[i], 0600, debugfs_root,
+					NULL, fops[i])) {
+			pr_err("failed to create %s file\n", file_names[i]);
+			return -ENOMEM;
+		}
+	}
+
+	return 0;
+}
+
+static int __init damon_init_user_ctx(void)
+{
+	int rc;
+
+	struct damon_ctx *ctx = &damon_user_ctx;
+
+	ktime_get_coarse_ts64(&ctx->last_aggregation);
+	ctx->last_regions_update = ctx->last_aggregation;
+
+	rc = damon_set_recording(ctx, 1024 * 1024, "/damon.data");
+	if (rc)
+		return rc;
+
+	mutex_init(&ctx->kdamond_lock);
+
+	prandom_seed_state(&ctx->rndseed, 42);
+	INIT_LIST_HEAD(&ctx->tasks_list);
+
 	return 0;
 }
 
+static int __init damon_init(void)
+{
+	int rc;
+
+	rc = damon_init_user_ctx();
+	if (rc)
+		return rc;
+
+	rc = damon_debugfs_init();
+	if (rc)
+		pr_err("%s: debugfs init failed\n", __func__);
+
+	return rc;
+}
+
 static void __exit damon_exit(void)
 {
-	return;
+	damon_stop(&damon_user_ctx);
+	debugfs_remove_recursive(debugfs_root);
+
+	kfree(damon_user_ctx.rbuf);
+	kfree(damon_user_ctx.rfile_path);
 }
 
 module_init(damon_init);
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 10/15] mm/damon: Add tracepoints
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (8 preceding siblings ...)
  2020-03-18 11:27 ` [PATCH v7 09/15] mm/damon: Add debugfs interface SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 11/15] tools: Add a minimal user-space tool for DAMON SeongJae Park
                   ` (5 subsequent siblings)
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit adds a tracepoint for DAMON.  It traces the monitoring
results of each region for each aggregation interval.  Using this, DAMON
will be easily integrated with any tracepoints supporting tools such as
perf.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 include/trace/events/damon.h | 43 ++++++++++++++++++++++++++++++++++++
 mm/damon.c                   |  5 +++++
 2 files changed, 48 insertions(+)
 create mode 100644 include/trace/events/damon.h

diff --git a/include/trace/events/damon.h b/include/trace/events/damon.h
new file mode 100644
index 000000000000..bec3501f2b05
--- /dev/null
+++ b/include/trace/events/damon.h
@@ -0,0 +1,43 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM damon
+
+#if !defined(_TRACE_DAMON_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_DAMON_H
+
+#include <linux/types.h>
+#include <linux/tracepoint.h>
+
+TRACE_EVENT(damon_aggregated,
+
+	TP_PROTO(unsigned long pid, unsigned int nr_regions,
+		unsigned long vm_start, unsigned long vm_end,
+		unsigned int nr_accesses),
+
+	TP_ARGS(pid, nr_regions, vm_start, vm_end, nr_accesses),
+
+	TP_STRUCT__entry(
+		__field(unsigned long, pid)
+		__field(unsigned int, nr_regions)
+		__field(unsigned long, vm_start)
+		__field(unsigned long, vm_end)
+		__field(unsigned int, nr_accesses)
+	),
+
+	TP_fast_assign(
+		__entry->pid = pid;
+		__entry->nr_regions = nr_regions;
+		__entry->vm_start = vm_start;
+		__entry->vm_end = vm_end;
+		__entry->nr_accesses = nr_accesses;
+	),
+
+	TP_printk("pid=%lu nr_regions=%u %lu-%lu: %u", __entry->pid,
+			__entry->nr_regions, __entry->vm_start,
+			__entry->vm_end, __entry->nr_accesses)
+);
+
+#endif /* _TRACE_DAMON_H */
+
+/* This part must be outside protection */
+#include <trace/define_trace.h>
diff --git a/mm/damon.c b/mm/damon.c
index b77e537e2ffe..25c961fabdf4 100644
--- a/mm/damon.c
+++ b/mm/damon.c
@@ -9,6 +9,8 @@
 
 #define pr_fmt(fmt) "damon: " fmt
 
+#define CREATE_TRACE_POINTS
+
 #include <linux/damon.h>
 #include <linux/debugfs.h>
 #include <linux/delay.h>
@@ -20,6 +22,7 @@
 #include <linux/sched/mm.h>
 #include <linux/sched/task.h>
 #include <linux/slab.h>
+#include <trace/events/damon.h>
 
 #define damon_get_task_struct(t) \
 	(get_pid_task(find_vpid(t->pid), PIDTYPE_PID))
@@ -608,6 +611,8 @@ static void kdamond_reset_aggregated(struct damon_ctx *c)
 			damon_write_rbuf(c, &r->vm_end, sizeof(r->vm_end));
 			damon_write_rbuf(c, &r->nr_accesses,
 					sizeof(r->nr_accesses));
+			trace_damon_aggregated(t->pid, nr,
+					r->vm_start, r->vm_end, r->nr_accesses);
 			r->nr_accesses = 0;
 		}
 	}
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 11/15] tools: Add a minimal user-space tool for DAMON
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (9 preceding siblings ...)
  2020-03-18 11:27 ` [PATCH v7 10/15] mm/damon: Add tracepoints SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 12/15] Documentation/admin-guide/mm: Add a document " SeongJae Park
                   ` (4 subsequent siblings)
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit adds a shallow wrapper python script, ``/tools/damon/damo``
that provides more convenient interface.  Note that it is only aimed to
be used for minimal reference of the DAMON's debugfs interfaces and for
debugging of the DAMON itself.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 tools/damon/.gitignore    |   1 +
 tools/damon/_dist.py      |  36 ++++
 tools/damon/bin2txt.py    |  64 +++++++
 tools/damon/damo          |  37 ++++
 tools/damon/heats.py      | 358 ++++++++++++++++++++++++++++++++++++++
 tools/damon/nr_regions.py |  89 ++++++++++
 tools/damon/record.py     | 212 ++++++++++++++++++++++
 tools/damon/report.py     |  45 +++++
 tools/damon/wss.py        |  95 ++++++++++
 9 files changed, 937 insertions(+)
 create mode 100644 tools/damon/.gitignore
 create mode 100644 tools/damon/_dist.py
 create mode 100644 tools/damon/bin2txt.py
 create mode 100755 tools/damon/damo
 create mode 100644 tools/damon/heats.py
 create mode 100644 tools/damon/nr_regions.py
 create mode 100644 tools/damon/record.py
 create mode 100644 tools/damon/report.py
 create mode 100644 tools/damon/wss.py

diff --git a/tools/damon/.gitignore b/tools/damon/.gitignore
new file mode 100644
index 000000000000..96403d36ff93
--- /dev/null
+++ b/tools/damon/.gitignore
@@ -0,0 +1 @@
+__pycache__/*
diff --git a/tools/damon/_dist.py b/tools/damon/_dist.py
new file mode 100644
index 000000000000..9851ec964e5c
--- /dev/null
+++ b/tools/damon/_dist.py
@@ -0,0 +1,36 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: GPL-2.0
+
+import os
+import struct
+import subprocess
+
+def access_patterns(f):
+    nr_regions = struct.unpack('I', f.read(4))[0]
+
+    patterns = []
+    for r in range(nr_regions):
+        saddr = struct.unpack('L', f.read(8))[0]
+        eaddr = struct.unpack('L', f.read(8))[0]
+        nr_accesses = struct.unpack('I', f.read(4))[0]
+        patterns.append([eaddr - saddr, nr_accesses])
+    return patterns
+
+def plot_dist(data_file, output_file, xlabel, ylabel):
+    terminal = output_file.split('.')[-1]
+    if not terminal in ['pdf', 'jpeg', 'png', 'svg']:
+        os.remove(data_file)
+        print("Unsupported plot output type.")
+        exit(-1)
+
+    gnuplot_cmd = """
+    set term %s;
+    set output '%s';
+    set key off;
+    set xlabel '%s';
+    set ylabel '%s';
+    plot '%s' with linespoints;""" % (terminal, output_file, xlabel, ylabel,
+            data_file)
+    subprocess.call(['gnuplot', '-e', gnuplot_cmd])
+    os.remove(data_file)
+
diff --git a/tools/damon/bin2txt.py b/tools/damon/bin2txt.py
new file mode 100644
index 000000000000..d5ffac60e02c
--- /dev/null
+++ b/tools/damon/bin2txt.py
@@ -0,0 +1,64 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: GPL-2.0
+
+import argparse
+import os
+import struct
+import sys
+
+def parse_time(bindat):
+    "bindat should be 16 bytes"
+    sec = struct.unpack('l', bindat[0:8])[0]
+    nsec = struct.unpack('l', bindat[8:16])[0]
+    return sec * 1000000000 + nsec;
+
+def pr_region(f):
+    saddr = struct.unpack('L', f.read(8))[0]
+    eaddr = struct.unpack('L', f.read(8))[0]
+    nr_accesses = struct.unpack('I', f.read(4))[0]
+    print("%012x-%012x(%10d):\t%d" %
+            (saddr, eaddr, eaddr - saddr, nr_accesses))
+
+def pr_task_info(f):
+    pid = struct.unpack('L', f.read(8))[0]
+    print("pid: ", pid)
+    nr_regions = struct.unpack('I', f.read(4))[0]
+    print("nr_regions: ", nr_regions)
+    for r in range(nr_regions):
+        pr_region(f)
+
+def set_argparser(parser):
+    parser.add_argument('--input', '-i', type=str, metavar='<file>',
+            default='damon.data', help='input file name')
+
+def main(args=None):
+    if not args:
+        parser = argparse.ArgumentParser()
+        set_argparser(parser)
+        args = parser.parse_args()
+
+    file_path = args.input
+
+    if not os.path.isfile(file_path):
+        print('input file (%s) is not exist' % file_path)
+        exit(1)
+
+    with open(file_path, 'rb') as f:
+        start_time = None
+        while True:
+            timebin = f.read(16)
+            if len(timebin) != 16:
+                break
+            time = parse_time(timebin)
+            if not start_time:
+                start_time = time
+                print("start_time: ", start_time)
+            print("rel time: %16d" % (time - start_time))
+            nr_tasks = struct.unpack('I', f.read(4))[0]
+            print("nr_tasks: ", nr_tasks)
+            for t in range(nr_tasks):
+                pr_task_info(f)
+                print("")
+
+if __name__ == '__main__':
+    main()
diff --git a/tools/damon/damo b/tools/damon/damo
new file mode 100755
index 000000000000..58e1099ae5fc
--- /dev/null
+++ b/tools/damon/damo
@@ -0,0 +1,37 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: GPL-2.0
+
+import argparse
+
+import record
+import report
+
+class SubCmdHelpFormatter(argparse.RawDescriptionHelpFormatter):
+    def _format_action(self, action):
+        parts = super(argparse.RawDescriptionHelpFormatter,
+                self)._format_action(action)
+        # skip sub parsers help
+        if action.nargs == argparse.PARSER:
+            parts = '\n'.join(parts.split('\n')[1:])
+        return parts
+
+parser = argparse.ArgumentParser(formatter_class=SubCmdHelpFormatter)
+
+subparser = parser.add_subparsers(title='command', dest='command',
+        metavar='<command>')
+subparser.required = True
+
+parser_record = subparser.add_parser('record',
+        help='record data accesses of the given target processes')
+record.set_argparser(parser_record)
+
+parser_report = subparser.add_parser('report',
+        help='report the recorded data accesses in the specified form')
+report.set_argparser(parser_report)
+
+args = parser.parse_args()
+
+if args.command == 'record':
+    record.main(args)
+elif args.command == 'report':
+    report.main(args)
diff --git a/tools/damon/heats.py b/tools/damon/heats.py
new file mode 100644
index 000000000000..48e966c5ca02
--- /dev/null
+++ b/tools/damon/heats.py
@@ -0,0 +1,358 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: GPL-2.0
+
+"""
+Transform binary trace data into human readable text that can be used for
+heatmap drawing, or directly plot the data in a heatmap format.
+
+Format of the text is:
+
+    <time> <space> <heat>
+    ...
+
+"""
+
+import argparse
+import os
+import struct
+import subprocess
+import sys
+import tempfile
+
+class HeatSample:
+    space_idx = None
+    sz_time_space = None
+    heat = None
+
+    def __init__(self, space_idx, sz_time_space, heat):
+        if sz_time_space < 0:
+            raise RuntimeError()
+        self.space_idx = space_idx
+        self.sz_time_space = sz_time_space
+        self.heat = heat
+
+    def total_heat(self):
+        return self.heat * self.sz_time_space
+
+    def merge(self, sample):
+        "sample must have a space idx that same to self"
+        heat_sum = self.total_heat() + sample.total_heat()
+        self.heat = heat_sum / (self.sz_time_space + sample.sz_time_space)
+        self.sz_time_space += sample.sz_time_space
+
+def pr_samples(samples, time_idx, time_unit, region_unit):
+    display_time = time_idx * time_unit
+    for idx, sample in enumerate(samples):
+        display_addr = idx * region_unit
+        if not sample:
+            print("%s\t%s\t%s" % (display_time, display_addr, 0.0))
+            continue
+        print("%s\t%s\t%s" % (display_time, display_addr, sample.total_heat() /
+            time_unit / region_unit))
+
+def to_idx(value, min_, unit):
+    return (value - min_) // unit
+
+def read_task_heats(f, pid, aunit, amin, amax):
+    pid_ = struct.unpack('L', f.read(8))[0]
+    nr_regions = struct.unpack('I', f.read(4))[0]
+    if pid_ != pid:
+        f.read(20 * nr_regions)
+        return None
+    samples = []
+    for i in range(nr_regions):
+        saddr = struct.unpack('L', f.read(8))[0]
+        eaddr = struct.unpack('L', f.read(8))[0]
+        eaddr = min(eaddr, amax - 1)
+        heat = struct.unpack('I', f.read(4))[0]
+
+        if eaddr <= amin:
+            continue
+        if saddr >= amax:
+            continue
+        saddr = max(amin, saddr)
+        eaddr = min(amax, eaddr)
+
+        sidx = to_idx(saddr, amin, aunit)
+        eidx = to_idx(eaddr - 1, amin, aunit)
+        for idx in range(sidx, eidx + 1):
+            sa = max(amin + idx * aunit, saddr)
+            ea = min(amin + (idx + 1) * aunit, eaddr)
+            sample = HeatSample(idx, (ea - sa), heat)
+            samples.append(sample)
+    return samples
+
+def parse_time(bindat):
+    sec = struct.unpack('l', bindat[0:8])[0]
+    nsec = struct.unpack('l', bindat[8:16])[0]
+    return sec * 1000000000 + nsec
+
+def apply_samples(target_samples, samples, start_time, end_time, aunit, amin):
+    for s in samples:
+        sample = HeatSample(s.space_idx,
+                s.sz_time_space * (end_time - start_time), s.heat)
+        idx = sample.space_idx
+        if not target_samples[idx]:
+            target_samples[idx] = sample
+        else:
+            target_samples[idx].merge(sample)
+
+def __pr_heats(f, pid, tunit, tmin, tmax, aunit, amin, amax):
+    heat_samples = [None] * ((amax - amin) // aunit)
+
+    start_time = 0
+    end_time = 0
+    last_flushed = -1
+    while True:
+        start_time = end_time
+        timebin = f.read(16)
+        if (len(timebin)) != 16:
+            break
+        end_time = parse_time(timebin)
+        nr_tasks = struct.unpack('I', f.read(4))[0]
+        samples_set = {}
+        for t in range(nr_tasks):
+            samples = read_task_heats(f, pid, aunit, amin, amax)
+            if samples:
+                samples_set[pid] = samples
+        if not pid in samples_set:
+            continue
+        if start_time >= tmax:
+            continue
+        if end_time <= tmin:
+            continue
+        start_time = max(start_time, tmin)
+        end_time = min(end_time, tmax)
+
+        sidx = to_idx(start_time, tmin, tunit)
+        eidx = to_idx(end_time - 1, tmin, tunit)
+        for idx in range(sidx, eidx + 1):
+            if idx != last_flushed:
+                pr_samples(heat_samples, idx, tunit, aunit)
+                heat_samples = [None] * ((amax - amin) // aunit)
+                last_flushed = idx
+            st = max(start_time, tmin + idx * tunit)
+            et = min(end_time, tmin + (idx + 1) * tunit)
+            apply_samples(heat_samples, samples_set[pid], st, et, aunit, amin)
+
+def pr_heats(args):
+    binfile = args.input
+    pid = args.pid
+    tres = args.tres
+    tmin = args.tmin
+    ares = args.ares
+    amin = args.amin
+
+    tunit = (args.tmax - tmin) // tres
+    aunit = (args.amax - amin) // ares
+
+    # Compensate the values so that those fit with the resolution
+    tmax = tmin + tunit * tres
+    amax = amin + aunit * ares
+
+    with open(binfile, 'rb') as f:
+        __pr_heats(f, pid, tunit, tmin, tmax, aunit, amin, amax)
+
+class GuideInfo:
+    pid = None
+    start_time = None
+    end_time = None
+    lowest_addr = None
+    highest_addr = None
+    gaps = None
+
+    def __init__(self, pid, start_time):
+        self.pid = pid
+        self.start_time = start_time
+        self.gaps = []
+
+    def regions(self):
+        regions = []
+        region = [self.lowest_addr]
+        for gap in self.gaps:
+            for idx, point in enumerate(gap):
+                if idx == 0:
+                    region.append(point)
+                    regions.append(region)
+                else:
+                    region = [point]
+        region.append(self.highest_addr)
+        regions.append(region)
+        return regions
+
+    def total_space(self):
+        ret = 0
+        for r in self.regions():
+            ret += r[1] - r[0]
+        return ret
+
+    def __str__(self):
+        lines = ['pid:%d' % self.pid]
+        lines.append('time: %d-%d (%d)' % (self.start_time, self.end_time,
+                    self.end_time - self.start_time))
+        for idx, region in enumerate(self.regions()):
+            lines.append('region\t%2d: %020d-%020d (%d)' %
+                    (idx, region[0], region[1], region[1] - region[0]))
+        return '\n'.join(lines)
+
+def is_overlap(region1, region2):
+    if region1[1] < region2[0]:
+        return False
+    if region2[1] < region1[0]:
+        return False
+    return True
+
+def overlap_region_of(region1, region2):
+    return [max(region1[0], region2[0]), min(region1[1], region2[1])]
+
+def overlapping_regions(regions1, regions2):
+    overlap_regions = []
+    for r1 in regions1:
+        for r2 in regions2:
+            if is_overlap(r1, r2):
+                r1 = overlap_region_of(r1, r2)
+        if r1:
+            overlap_regions.append(r1)
+    return overlap_regions
+
+def get_guide_info(binfile):
+    "Read file, return the set of guide information objects of the data"
+    guides = {}
+    with open(binfile, 'rb') as f:
+        while True:
+            timebin = f.read(16)
+            if len(timebin) != 16:
+                break
+            monitor_time = parse_time(timebin)
+            nr_tasks = struct.unpack('I', f.read(4))[0]
+            for t in range(nr_tasks):
+                pid = struct.unpack('L', f.read(8))[0]
+                nr_regions = struct.unpack('I', f.read(4))[0]
+                if not pid in guides:
+                    guides[pid] = GuideInfo(pid, monitor_time)
+                guide = guides[pid]
+                guide.end_time = monitor_time
+
+                last_addr = None
+                gaps = []
+                for r in range(nr_regions):
+                    saddr = struct.unpack('L', f.read(8))[0]
+                    eaddr = struct.unpack('L', f.read(8))[0]
+                    f.read(4)
+
+                    if not guide.lowest_addr or saddr < guide.lowest_addr:
+                        guide.lowest_addr = saddr
+                    if not guide.highest_addr or eaddr > guide.highest_addr:
+                        guide.highest_addr = eaddr
+
+                    if not last_addr:
+                        last_addr = eaddr
+                        continue
+                    if last_addr != saddr:
+                        gaps.append([last_addr, saddr])
+                    last_addr = eaddr
+
+                if not guide.gaps:
+                    guide.gaps = gaps
+                else:
+                    guide.gaps = overlapping_regions(guide.gaps, gaps)
+    return sorted(list(guides.values()), key=lambda x: x.total_space(),
+                    reverse=True)
+
+def pr_guide(binfile):
+    for guide in get_guide_info(binfile):
+        print(guide)
+
+def region_sort_key(region):
+    return region[1] - region[0]
+
+def set_missed_args(args):
+    if args.pid and args.tmin and args.tmax and args.amin and args.amax:
+        return
+    guides = get_guide_info(args.input)
+    guide = guides[0]
+    if not args.pid:
+        args.pid = guide.pid
+    for g in guides:
+        if g.pid == args.pid:
+            guide = g
+            break
+
+    if not args.tmin:
+        args.tmin = guide.start_time
+    if not args.tmax:
+        args.tmax = guide.end_time
+
+    if not args.amin or not args.amax:
+        region = sorted(guide.regions(), key=lambda x: x[1] - x[0],
+                reverse=True)[0]
+        args.amin = region[0]
+        args.amax = region[1]
+
+def plot_heatmap(data_file, output_file):
+    terminal = output_file.split('.')[-1]
+    if not terminal in ['pdf', 'jpeg', 'png', 'svg']:
+        os.remove(data_file)
+        print("Unsupported plot output type.")
+        exit(-1)
+
+    gnuplot_cmd = """
+    set term %s;
+    set output '%s';
+    set key off;
+    set xrange [0:];
+    set yrange [0:];
+    set xlabel 'Time (ns)';
+    set ylabel 'Virtual Address (bytes)';
+    plot '%s' using 1:2:3 with image;""" % (terminal, output_file, data_file)
+    subprocess.call(['gnuplot', '-e', gnuplot_cmd])
+    os.remove(data_file)
+
+def set_argparser(parser):
+    parser.add_argument('--input', '-i', type=str, metavar='<file>',
+            default='damon.data', help='input file name')
+    parser.add_argument('--pid', metavar='<pid>', type=int,
+            help='pid of target task')
+    parser.add_argument('--tres', metavar='<resolution>', type=int,
+            default=500, help='time resolution of the output')
+    parser.add_argument('--tmin', metavar='<time>', type=lambda x: int(x,0),
+            help='minimal time of the output')
+    parser.add_argument('--tmax', metavar='<time>', type=lambda x: int(x,0),
+            help='maximum time of the output')
+    parser.add_argument('--ares', metavar='<resolution>', type=int, default=500,
+            help='space address resolution of the output')
+    parser.add_argument('--amin', metavar='<address>', type=lambda x: int(x,0),
+            help='minimal space address of the output')
+    parser.add_argument('--amax', metavar='<address>', type=lambda x: int(x,0),
+            help='maximum space address of the output')
+    parser.add_argument('--guide', action='store_true',
+            help='print a guidance for the min/max/resolution settings')
+    parser.add_argument('--heatmap', metavar='<file>', type=str,
+            help='heatmap image file to create')
+
+def main(args=None):
+    if not args:
+        parser = argparse.ArgumentParser()
+        set_argparser(parser)
+        args = parser.parse_args()
+
+    if args.guide:
+        pr_guide(args.input)
+    else:
+        set_missed_args(args)
+        orig_stdout = sys.stdout
+        if args.heatmap:
+            tmp_path = tempfile.mkstemp()[1]
+            tmp_file = open(tmp_path, 'w')
+            sys.stdout = tmp_file
+
+        pr_heats(args)
+
+        if args.heatmap:
+            sys.stdout = orig_stdout
+            tmp_file.flush()
+            tmp_file.close()
+            plot_heatmap(tmp_path, args.heatmap)
+
+if __name__ == '__main__':
+    main()
diff --git a/tools/damon/nr_regions.py b/tools/damon/nr_regions.py
new file mode 100644
index 000000000000..fcc2ce13e5f5
--- /dev/null
+++ b/tools/damon/nr_regions.py
@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: GPL-2.0
+
+"Print out distribution of the number of regions in the given record"
+
+import argparse
+import struct
+import sys
+import tempfile
+
+import _dist
+
+def set_argparser(parser):
+    parser.add_argument('--input', '-i', type=str, metavar='<file>',
+            default='damon.data', help='input file name')
+    parser.add_argument('--range', '-r', type=int, nargs=3,
+            metavar=('<start>', '<stop>', '<step>'),
+            help='range of percentiles to print')
+    parser.add_argument('--sortby', '-s', choices=['time', 'size'],
+            help='the metric to be used for sorting the number of regions')
+    parser.add_argument('--plot', '-p', type=str, metavar='<file>',
+            help='plot the distribution to an image file')
+
+def main(args=None):
+    if not args:
+        parser = argparse.ArgumentParser()
+        set_argparser(parser)
+        args = parser.parse_args()
+
+    percentiles = [0, 25, 50, 75, 100]
+
+    file_path = args.input
+    if args.range:
+        percentiles = range(args.range[0], args.range[1], args.range[2])
+    nr_regions_sort = True
+    if args.sortby == 'time':
+        nr_regions_sort = False
+
+    pid_pattern_map = {}
+    with open(file_path, 'rb') as f:
+        start_time = None
+        while True:
+            timebin = f.read(16)
+            if len(timebin) != 16:
+                break
+            nr_tasks = struct.unpack('I', f.read(4))[0]
+            for t in range(nr_tasks):
+                pid = struct.unpack('L', f.read(8))[0]
+                if not pid in pid_pattern_map:
+                    pid_pattern_map[pid] = []
+                pid_pattern_map[pid].append(_dist.access_patterns(f))
+
+    orig_stdout = sys.stdout
+    if args.plot:
+        tmp_path = tempfile.mkstemp()[1]
+        tmp_file = open(tmp_path, 'w')
+        sys.stdout = tmp_file
+
+    print('# <percentile> <# regions>')
+    for pid in pid_pattern_map.keys():
+        # Skip firs 20 regions as those would not adaptively adjusted
+        snapshots = pid_pattern_map[pid][20:]
+        nr_regions_dist = []
+        for snapshot in snapshots:
+            nr_regions_dist.append(len(snapshot))
+        if nr_regions_sort:
+            nr_regions_dist.sort(reverse=False)
+
+        print('# pid\t%s' % pid)
+        print('# avr:\t%d' % (sum(nr_regions_dist) / len(nr_regions_dist)))
+        for percentile in percentiles:
+            thres_idx = int(percentile / 100.0 * len(nr_regions_dist))
+            if thres_idx == len(nr_regions_dist):
+                thres_idx -= 1
+            threshold = nr_regions_dist[thres_idx]
+            print('%d\t%d' % (percentile, nr_regions_dist[thres_idx]))
+
+    if args.plot:
+        sys.stdout = orig_stdout
+        tmp_file.flush()
+        tmp_file.close()
+        xlabel = 'runtime (percent)'
+        if nr_regions_sort:
+            xlabel = 'percentile'
+        _dist.plot_dist(tmp_path, args.plot, xlabel,
+                'number of monitoring target regions')
+
+if __name__ == '__main__':
+    main()
diff --git a/tools/damon/record.py b/tools/damon/record.py
new file mode 100644
index 000000000000..a547d479a103
--- /dev/null
+++ b/tools/damon/record.py
@@ -0,0 +1,212 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: GPL-2.0
+
+"""
+Record data access patterns of the target process.
+"""
+
+import argparse
+import copy
+import os
+import signal
+import subprocess
+import time
+
+debugfs_attrs = None
+debugfs_record = None
+debugfs_pids = None
+debugfs_monitor_on = None
+
+def set_target_pid(pid):
+    return subprocess.call('echo %s > %s' % (pid, debugfs_pids), shell=True,
+            executable='/bin/bash')
+
+def turn_damon(on_off):
+    return subprocess.call("echo %s > %s" % (on_off, debugfs_monitor_on),
+            shell=True, executable="/bin/bash")
+
+def is_damon_running():
+    with open(debugfs_monitor_on, 'r') as f:
+        return f.read().strip() == 'on'
+
+def do_record(target, is_target_cmd, attrs, old_attrs):
+    if os.path.isfile(attrs.rfile_path):
+        os.rename(attrs.rfile_path, attrs.rfile_path + '.old')
+
+    if attrs.apply():
+        print('attributes (%s) failed to be applied' % attrs)
+        cleanup_exit(old_attrs, -1)
+    print('# damon attrs: %s' % attrs)
+    if is_target_cmd:
+        p = subprocess.Popen(target, shell=True, executable='/bin/bash')
+        target = p.pid
+    if set_target_pid(target):
+        print('pid setting (%s) failed' % target)
+        cleanup_exit(old_attrs, -2)
+    if turn_damon('on'):
+        print('could not turn on damon' % target)
+        cleanup_exit(old_attrs, -3)
+    if is_target_cmd:
+        p.wait()
+    while True:
+        # damon will turn it off by itself if the target tasks are terminated.
+        if not is_damon_running():
+            break
+        time.sleep(1)
+
+    cleanup_exit(old_attrs, 0)
+
+class Attrs:
+    sample_interval = None
+    aggr_interval = None
+    regions_update_interval = None
+    min_nr_regions = None
+    max_nr_regions = None
+    rbuf_len = None
+    rfile_path = None
+
+    def __init__(self, s, a, r, n, x, l, f):
+        self.sample_interval = s
+        self.aggr_interval = a
+        self.regions_update_interval = r
+        self.min_nr_regions = n
+        self.max_nr_regions = x
+        self.rbuf_len = l
+        self.rfile_path = f
+
+    def __str__(self):
+        return "%s %s %s %s %s %s %s" % (self.sample_interval, self.aggr_interval,
+                self.regions_update_interval, self.min_nr_regions,
+                self.max_nr_regions, self.rbuf_len, self.rfile_path)
+
+    def attr_str(self):
+        return "%s %s %s %s %s " % (self.sample_interval, self.aggr_interval,
+                self.regions_update_interval, self.min_nr_regions,
+                self.max_nr_regions)
+
+    def record_str(self):
+        return '%s %s ' % (self.rbuf_len, self.rfile_path)
+
+    def apply(self):
+        ret = subprocess.call('echo %s > %s' % (self.attr_str(), debugfs_attrs),
+                shell=True, executable='/bin/bash')
+        if ret:
+            return ret
+        return subprocess.call('echo %s > %s' % (self.record_str(),
+            debugfs_record), shell=True, executable='/bin/bash')
+
+def current_attrs():
+    with open(debugfs_attrs, 'r') as f:
+        attrs = f.read().split()
+    attrs = [int(x) for x in attrs]
+
+    with open(debugfs_record, 'r') as f:
+        rattrs = f.read().split()
+    attrs.append(int(rattrs[0]))
+    attrs.append(rattrs[1])
+    return Attrs(*attrs)
+
+def cmd_args_to_attrs(args):
+    "Generate attributes with specified arguments"
+    sample_interval = args.sample
+    aggr_interval = args.aggr
+    regions_update_interval = args.updr
+    min_nr_regions = args.minr
+    max_nr_regions = args.maxr
+    rbuf_len = args.rbuf
+    if not os.path.isabs(args.out):
+        args.out = os.path.join(os.getcwd(), args.out)
+    rfile_path = args.out
+    return Attrs(sample_interval, aggr_interval, regions_update_interval,
+            min_nr_regions, max_nr_regions, rbuf_len, rfile_path)
+
+def cleanup_exit(orig_attrs, exit_code):
+    if is_damon_running():
+        if turn_damon('off'):
+            print('failed to turn damon off!')
+    if orig_attrs:
+        if orig_attrs.apply():
+            print('original attributes (%s) restoration failed!' % orig_attrs)
+    exit(exit_code)
+
+def sighandler(signum, frame):
+    print('\nsignal %s received' % signum)
+    cleanup_exit(orig_attrs, signum)
+
+def chk_update_debugfs(debugfs):
+    global debugfs_attrs
+    global debugfs_record
+    global debugfs_pids
+    global debugfs_monitor_on
+
+    debugfs_damon = os.path.join(debugfs, 'damon')
+    debugfs_attrs = os.path.join(debugfs_damon, 'attrs')
+    debugfs_record = os.path.join(debugfs_damon, 'record')
+    debugfs_pids = os.path.join(debugfs_damon, 'pids')
+    debugfs_monitor_on = os.path.join(debugfs_damon, 'monitor_on')
+
+    if not os.path.isdir(debugfs_damon):
+        print("damon debugfs dir (%s) not found", debugfs_damon)
+        exit(1)
+
+    for f in [debugfs_attrs, debugfs_record, debugfs_pids, debugfs_monitor_on]:
+        if not os.path.isfile(f):
+            print("damon debugfs file (%s) not found" % f)
+            exit(1)
+
+def chk_permission():
+    if os.geteuid() != 0:
+        print("Run as root")
+        exit(1)
+
+def set_argparser(parser):
+    parser.add_argument('target', type=str, metavar='<target>',
+            help='the target command or the pid to record')
+    parser.add_argument('-s', '--sample', metavar='<interval>', type=int,
+            default=5000, help='sampling interval')
+    parser.add_argument('-a', '--aggr', metavar='<interval>', type=int,
+            default=100000, help='aggregate interval')
+    parser.add_argument('-u', '--updr', metavar='<interval>', type=int,
+            default=1000000, help='regions update interval')
+    parser.add_argument('-n', '--minr', metavar='<# regions>', type=int,
+            default=10, help='minimal number of regions')
+    parser.add_argument('-m', '--maxr', metavar='<# regions>', type=int,
+            default=1000, help='maximum number of regions')
+    parser.add_argument('-l', '--rbuf', metavar='<len>', type=int,
+            default=1024*1024, help='length of record result buffer')
+    parser.add_argument('-o', '--out', metavar='<file path>', type=str,
+            default='damon.data', help='output file path')
+    parser.add_argument('-d', '--debugfs', metavar='<debugfs>', type=str,
+            default='/sys/kernel/debug', help='debugfs mounted path')
+
+def main(args=None):
+    global orig_attrs
+    if not args:
+        parser = argparse.ArgumentParser()
+        set_argparser(parser)
+        args = parser.parse_args()
+
+    chk_permission()
+    chk_update_debugfs(args.debugfs)
+
+    signal.signal(signal.SIGINT, sighandler)
+    signal.signal(signal.SIGTERM, sighandler)
+    orig_attrs = current_attrs()
+
+    new_attrs = cmd_args_to_attrs(args)
+    target = args.target
+
+    target_fields = target.split()
+    if not subprocess.call('which %s > /dev/null' % target_fields[0],
+            shell=True, executable='/bin/bash'):
+        do_record(target, True, new_attrs, orig_attrs)
+    else:
+        try:
+            pid = int(target)
+        except:
+            print('target \'%s\' is neither a command, nor a pid' % target)
+            exit(1)
+        do_record(target, False, new_attrs, orig_attrs)
+
+if __name__ == '__main__':
+    main()
diff --git a/tools/damon/report.py b/tools/damon/report.py
new file mode 100644
index 000000000000..c661c7b2f1af
--- /dev/null
+++ b/tools/damon/report.py
@@ -0,0 +1,45 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: GPL-2.0
+
+import argparse
+
+import bin2txt
+import heats
+import nr_regions
+import wss
+
+def set_argparser(parser):
+    subparsers = parser.add_subparsers(title='report type', dest='report_type',
+            metavar='<report type>', help='the type of the report to generate')
+    subparsers.required = True
+
+    parser_raw = subparsers.add_parser('raw', help='human readable raw data')
+    bin2txt.set_argparser(parser_raw)
+
+    parser_heats = subparsers.add_parser('heats', help='heats of regions')
+    heats.set_argparser(parser_heats)
+
+    parser_wss = subparsers.add_parser('wss', help='working set size')
+    wss.set_argparser(parser_wss)
+
+    parser_nr_regions = subparsers.add_parser('nr_regions',
+            help='number of regions')
+    nr_regions.set_argparser(parser_nr_regions)
+
+def main(args=None):
+    if not args:
+        parser = argparse.ArgumentParser()
+        set_argparser(parser)
+        args = parser.parse_args()
+
+    if args.report_type == 'raw':
+        bin2txt.main(args)
+    elif args.report_type == 'heats':
+        heats.main(args)
+    elif args.report_type == 'wss':
+        wss.main(args)
+    elif args.report_type == 'nr_regions':
+        nr_regions.main(args)
+
+if __name__ == '__main__':
+    main()
diff --git a/tools/damon/wss.py b/tools/damon/wss.py
new file mode 100644
index 000000000000..890deee5b9be
--- /dev/null
+++ b/tools/damon/wss.py
@@ -0,0 +1,95 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: GPL-2.0
+
+"Print out the distribution of the working set sizes of the given trace"
+
+import argparse
+import struct
+import sys
+import tempfile
+
+import _dist
+
+def set_argparser(parser):
+    parser.add_argument('--input', '-i', type=str, metavar='<file>',
+            default='damon.data', help='input file name')
+    parser.add_argument('--range', '-r', type=int, nargs=3,
+            metavar=('<start>', '<stop>', '<step>'),
+            help='range of wss percentiles to print')
+    parser.add_argument('--sortby', '-s', choices=['time', 'size'],
+            help='the metric to be used for the sort of the working set sizes')
+    parser.add_argument('--plot', '-p', type=str, metavar='<file>',
+            help='plot the distribution to an image file')
+
+def main(args=None):
+    if not args:
+        parser = argparse.ArgumentParser()
+        set_argparser(parser)
+        args = parser.parse_args()
+
+    percentiles = [0, 25, 50, 75, 100]
+
+    file_path = args.input
+    if args.range:
+        percentiles = range(args.range[0], args.range[1], args.range[2])
+    wss_sort = True
+    if args.sortby == 'time':
+        wss_sort = False
+
+    pid_pattern_map = {}
+    with open(file_path, 'rb') as f:
+        start_time = None
+        while True:
+            timebin = f.read(16)
+            if len(timebin) != 16:
+                break
+            nr_tasks = struct.unpack('I', f.read(4))[0]
+            for t in range(nr_tasks):
+                pid = struct.unpack('L', f.read(8))[0]
+                if not pid in pid_pattern_map:
+                    pid_pattern_map[pid] = []
+                pid_pattern_map[pid].append(_dist.access_patterns(f))
+
+    orig_stdout = sys.stdout
+    if args.plot:
+        tmp_path = tempfile.mkstemp()[1]
+        tmp_file = open(tmp_path, 'w')
+        sys.stdout = tmp_file
+
+    print('# <percentile> <wss>')
+    for pid in pid_pattern_map.keys():
+        # Skip first 20 snapshots as regions may not adjusted yet.
+        snapshots = pid_pattern_map[pid][20:]
+        wss_dist = []
+        for snapshot in snapshots:
+            wss = 0
+            for p in snapshot:
+                # Ignore regions not accessed
+                if p[1] <= 0:
+                    continue
+                wss += p[0]
+            wss_dist.append(wss)
+        if wss_sort:
+            wss_dist.sort(reverse=False)
+
+        print('# pid\t%s' % pid)
+        print('# avr:\t%d' % (sum(wss_dist) / len(wss_dist)))
+        for percentile in percentiles:
+            thres_idx = int(percentile / 100.0 * len(wss_dist))
+            if thres_idx == len(wss_dist):
+                thres_idx -= 1
+            threshold = wss_dist[thres_idx]
+            print('%d\t%d' % (percentile, wss_dist[thres_idx]))
+
+    if args.plot:
+        sys.stdout = orig_stdout
+        tmp_file.flush()
+        tmp_file.close()
+        xlabel = 'runtime (percent)'
+        if wss_sort:
+            xlabel = 'percentile'
+        _dist.plot_dist(tmp_path, args.plot, xlabel,
+                'working set size (bytes)')
+
+if __name__ == '__main__':
+    main()
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 12/15] Documentation/admin-guide/mm: Add a document for DAMON
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (10 preceding siblings ...)
  2020-03-18 11:27 ` [PATCH v7 11/15] tools: Add a minimal user-space tool for DAMON SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-18 11:27 ` [PATCH v7 13/15] mm/damon: Add kunit tests SeongJae Park
                   ` (3 subsequent siblings)
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit adds a simple document for DAMON under
`Documentation/admin-guide/mm`.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 .../admin-guide/mm/data_access_monitor.rst    | 428 ++++++++++++++++++
 Documentation/admin-guide/mm/index.rst        |   1 +
 2 files changed, 429 insertions(+)
 create mode 100644 Documentation/admin-guide/mm/data_access_monitor.rst

diff --git a/Documentation/admin-guide/mm/data_access_monitor.rst b/Documentation/admin-guide/mm/data_access_monitor.rst
new file mode 100644
index 000000000000..947d66713fbd
--- /dev/null
+++ b/Documentation/admin-guide/mm/data_access_monitor.rst
@@ -0,0 +1,428 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+==========================
+DAMON: Data Access MONitor
+==========================
+
+Introduction
+============
+
+Memory management decisions can be improved if finer data access information is
+available.  However, because such finer information usually comes with higher
+overhead, most systems including Linux forgives the potential benefit and rely
+on only coarse information or some light-weight heuristics.
+
+A number of data access pattern awared memory management optimizations
+consistently say the potential benefit is not small (2.55x speedup).  However,
+none of those has successfully adopted into into the Linux kernel mainly due to
+the absence of a scalable and efficient data access monitoring mechanism.
+
+DAMON is a data access monitoring subsystem for the problem.  It is 1) accurate
+enough to be used for the DRAM level memory management, 2) light-weight enough
+to be applied online, and 3) keeps predefined upper-bound overhead
+regardless of the size of target workloads (thus scalable).
+
+DAMON is implemented as a standalone kernel module and provides several simple
+interfaces.  Owing to that, though it has mainly designed for the kernel's
+memory management mechanisms, it can be also used for a wide range of user
+space programs and people.
+
+
+Frequently Asked Questions
+==========================
+
+Q: Why not integrated with perf?
+A: From the perspective of perf like profilers, DAMON can be thought of as a
+data source in kernel, like tracepoints, pressure stall information (psi), or
+idle page tracking.  Thus, it can be easily integrated with those.  However,
+this patchset doesn't provide a fancy perf integration because current step of
+DAMON development is focused on its core logic only.  That said, DAMON already
+provides two interfaces for user space programs, which based on debugfs and
+tracepoint, respectively.  Using the tracepoint interface, you can use DAMON
+with perf.  This patchset also provides the debugfs interface based user space
+tool for DAMON.  It can be used to record, visualize, and analyze data access
+pattern of target processes in a convenient way.
+
+Q: Why a new module, instead of extending perf or other tools?
+A: First, DAMON aims to be used by other programs including the kernel.
+Therefore, having dependency to specific tools like perf is not desirable.
+Second, because it need to be lightweight as much as possible so that it can be
+used online, any unnecessary overhead such as kernel - user space context
+switching cost should be avoided.  These are the two most biggest reasons why
+DAMON is implemented in the kernel space.  The idle page tracking subsystem
+would be the kernel module that most seems similar to DAMON.  However, it's own
+interface is not compatible with DAMON.  Also, the internal implementation of
+it has no common part to be reused by DAMON.
+
+Q: Can 'perf mem' provide the data required for DAMON?
+A: On the systems supporting 'perf mem', yes.  DAMON is using the PTE Accessed
+bits in low level.  Other H/W or S/W features that can be used for the purpose
+could be used.  However, as explained with above question, DAMON need to be
+implemented in the kernel space.
+
+
+Expected Use-cases
+==================
+
+A straightforward usecase of DAMON would be the program behavior analysis.
+With the DAMON output, users can confirm whether the program is running as
+intended or not.  This will be useful for debuggings and tests of design
+points.
+
+The monitored results can also be useful for counting the dynamic working set
+size of workloads.  For the administration of memory overcommitted systems or
+selection of the environments (e.g., containers providing different amount of
+memory) for your workloads, this will be useful.
+
+If you are a programmer, you can optimize your program by managing the memory
+based on the actual data access pattern.  For example, you can identify the
+dynamic hotness of your data using DAMON and call ``mlock()`` to keep your hot
+data in DRAM, or call ``madvise()`` with ``MADV_PAGEOUT`` to proactively
+reclaim cold data.  Even though your program is guaranteed to not encounter
+memory pressure, you can still improve the performance by applying the DAMON
+outputs for call of ``MADV_HUGEPAGE`` and ``MADV_NOHUGEPAGE``.  More creative
+optimizations would be possible.  Our evaluations of DAMON includes a
+straightforward optimization using the ``mlock()``.  Please refer to the below
+Evaluation section for more detail.
+
+As DAMON incurs very low overhead, such optimizations can be applied not only
+offline, but also online.  Also, there is no reason to limit such optimizations
+to the user space.  Several parts of the kernel's memory management mechanisms
+could be also optimized using DAMON. The reclamation, the THP (de)promotion
+decisions, and the compaction would be such a candidates.
+
+
+Mechanisms of DAMON
+===================
+
+
+Basic Access Check
+------------------
+
+DAMON basically reports what pages are how frequently accessed.  The report is
+passed to users in binary format via a ``result file`` which users can set it's
+path.  Note that the frequency is not an absolute number of accesses, but a
+relative frequency among the pages of the target workloads.
+
+Users can also control the resolution of the reports by setting two time
+intervals, ``sampling interval`` and ``aggregation interval``.  In detail,
+DAMON checks access to each page per ``sampling interval``, aggregates the
+results (counts the number of the accesses to each page), and reports the
+aggregated results per ``aggregation interval``.  For the access check of each
+page, DAMON uses the Accessed bits of PTEs.
+
+This is thus similar to the previously mentioned periodic access checks based
+mechanisms, which overhead is increasing as the size of the target process
+grows.
+
+
+Region Based Sampling
+---------------------
+
+To avoid the unbounded increase of the overhead, DAMON groups a number of
+adjacent pages that assumed to have same access frequencies into a region.  As
+long as the assumption (pages in a region have same access frequencies) is
+kept, only one page in the region is required to be checked.  Thus, for each
+``sampling interval``, DAMON randomly picks one page in each region and clears
+its Accessed bit.  After one more ``sampling interval``, DAMON reads the
+Accessed bit of the page and increases the access frequency of the region if
+the bit has set meanwhile.  Therefore, the monitoring overhead is controllable
+by setting the number of regions.  DAMON allows users to set the minimal and
+maximum number of regions for the trade-off.
+
+Except the assumption, this is almost same with the above-mentioned
+miniature-like static region based sampling.  In other words, this scheme
+cannot preserve the quality of the output if the assumption is not guaranteed.
+
+
+Adaptive Regions Adjustment
+---------------------------
+
+At the beginning of the monitoring, DAMON constructs the initial regions by
+evenly splitting the memory mapped address space of the process into the
+user-specified minimal number of regions.  In this initial state, the
+assumption is normally not kept and thus the quality could be low.  To keep the
+assumption as much as possible, DAMON adaptively merges and splits each region.
+For each ``aggregation interval``, it compares the access frequencies of
+adjacent regions and merges those if the frequency difference is small.  Then,
+after it reports and clears the aggregated access frequency of each region, it
+splits each region into two regions if the total number of regions is smaller
+than the half of the user-specified maximum number of regions.
+
+In this way, DAMON provides its best-effort quality and minimal overhead while
+keeping the bounds users set for their trade-off.
+
+
+Applying Complex and Dynamic Memory Mappings
+--------------------------------------------
+
+Only a number of small parts in the super-huge virtual address space of the
+processes is mapped to physical memory and accessed.  Thus, tracking the
+unmapped address regions is just wasteful.  However, because DAMON can deal
+with some level of noises using the adaptive regions adjustment mechanism,
+tracking every mapping is not strictly required but could even incur a high
+overhead in somce cases.  That said, too huge unmapped areas inside the
+monitoring target should be removed to not take the time for the adaptive
+mechanism.
+
+For the reason, DAMON converts the complex mappings to three distinct regions
+that cover every mapped areas of the address space.  Also the two gaps between
+the three regions are the two biggest unmapped areas in the given address
+space.  The two biggest unmapped areas might be the gap between the heap and
+the uppermost mmap()-ed region, and the gap between the lowermost mmap()-ed
+region and the stack will be two biggest unmapped regions.  Because these gaps
+are exceptionally huge areas in usual address space, excluding these two
+biggest unmapped regions will be sufficient to make a trade-off.
+
+To further minimize dynamic mapping changes applying overhead, DAMON check the
+dynamic memory mapping changes and applies it to the abstracted target area
+only for each of an user-specified time interval (``regions update interval``).
+
+
+``debugfs`` Interface
+=====================
+
+DAMON exports four files, ``attrs``, ``pids``, ``record``, and ``monitor_on``
+under its debugfs directory, ``<debugfs>/damon/``.
+
+Attributes
+----------
+
+Users can read and write the ``sampling interval``, ``aggregation interval``,
+``regions update interval``, and min/max number of monitoring target regions by
+reading from and writing to the ``attrs`` file.  For example, below commands
+set those values to 5 ms, 100 ms, 1,000 ms, 10, 1000 and check it again::
+
+    # cd <debugfs>/damon
+    # echo 5000 100000 1000000 10 1000 > attrs
+    # cat attrs
+    5000 100000 1000000 10 1000
+
+Target PIDs
+-----------
+
+Users can read and write the pids of current monitoring target processes by
+reading from and writing to the ``pids`` file.  For example, below commands set
+processes having pids 42 and 4242 as the processes to be monitored and check it
+again::
+
+    # cd <debugfs>/damon
+    # echo 42 4242 > pids
+    # cat pids
+    42 4242
+
+Note that setting the pids doesn't starts the monitoring.
+
+Record
+------
+
+DAMON support direct monitoring result record feature.  The recorded results
+are first written to a buffer and flushed to a file in batch.  Users can set
+the size of the buffer and the path to the result file by reading from and
+writing to the ``record`` file.  For example, below commands set the buffer to
+be 4 KiB and the result to be saved in ``/damon.data``.
+
+    # cd <debugfs>/damon
+    # echo "4096 /damon.data" > record
+    # cat record
+    4096 /damon.data
+
+Turning On/Off
+--------------
+
+You can check current status, start and stop the monitoring by reading from and
+writing to the ``monitor_on`` file.  Writing ``on`` to the file starts DAMON to
+monitor the target processes with the attributes.  Writing ``off`` to the file
+stops DAMON.  DAMON also stops if every target processes is be terminated.
+Below example commands turn on, off, and check status of DAMON::
+
+    # cd <debugfs>/damon
+    # echo on > monitor_on
+    # echo off > monitor_on
+    # cat monitor_on
+    off
+
+Please note that you cannot write to the ``attrs`` and ``pids`` files while the
+monitoring is turned on.  If you write to the files while DAMON is running, an
+error code such as ``-EBUSY`` will be returned.
+
+
+User Space Tool for DAMON
+=========================
+
+There is a user space tool for DAMON, ``/tools/damon/damo``.  It provides
+another user interface which more convenient than the debugfs interface.
+Nevertheless, note that it is only aimed to be used for minimal reference of
+the DAMON's debugfs interfaces and for tests of the DAMON itself.  Based on the
+debugfs interface, you can create another cool and more convenient user space
+tools.
+
+The interface of the tool is basically subcommand based.  You can almost always
+use ``-h`` option to get help of the use of each subcommand.  Currently, it
+supports two subcommands, ``record`` and ``report``.
+
+
+Recording Data Access Pattern
+-----------------------------
+
+The ``record`` subcommand records the data access pattern of target process in
+a file (``./damon.data`` by default) using DAMON.  You can specifies the target
+as either pid or a command for an execution of the process.  Below example
+shows a command target usage::
+
+    # cd <kernel>/tools/damon/
+    # ./damo record "sleep 5"
+
+The tool will execute ``sleep 5`` by itself and record the data access patterns
+of the process.  Below example shows a pid target usage::
+
+    # sleep 5 &
+    # ./damo record `pidof sleep`
+
+You can set more detailed attributes and path to the recorded data file using
+optional arguments to the subcommand.  Use the ``-h`` option for more help.
+
+
+Analyzing Data Access Pattern
+-----------------------------
+
+The ``report`` subcommand reads a data access pattern record file (if not
+explicitly specified, reads ``./damon.data`` file if exists) and generates
+reports of various types.  You can specify what type of report you want using
+sub-subcommand to ``report`` subcommand.  For supported types, pass the ``-h``
+option to ``report`` subcommand.
+
+
+raw
+~~~
+
+``raw`` sub-subcommand simply transforms the record, which is storing the data
+access patterns in binary format to human readable text.  For example::
+
+    $ ./damo report raw
+    start_time:  193485829398
+    rel time:                0
+    nr_tasks:  1
+    pid:  1348
+    nr_regions:  4
+    560189609000-56018abce000(  22827008):  0
+    7fbdff59a000-7fbdffaf1a00(   5601792):  0
+    7fbdffaf1a00-7fbdffbb5000(    800256):  1
+    7ffea0dc0000-7ffea0dfd000(    249856):  0
+
+    rel time:        100000731
+    nr_tasks:  1
+    pid:  1348
+    nr_regions:  6
+    560189609000-56018abce000(  22827008):  0
+    7fbdff59a000-7fbdff8ce933(   3361075):  0
+    7fbdff8ce933-7fbdffaf1a00(   2240717):  1
+    7fbdffaf1a00-7fbdffb66d99(    480153):  0
+    7fbdffb66d99-7fbdffbb5000(    320103):  1
+    7ffea0dc0000-7ffea0dfd000(    249856):  0
+
+The first line shows recording started timestamp (nanosecond).  Records of data
+access patterns are following this.  Each record is sperated by a blank line.
+Each record first specifies the recorded time (``rel time``), number of
+monitored tasks in this record (``nr_tasks``).  Multiple number of records of
+data access pattern for each task continue.  Each data access pattern for each
+task shows first it's pid (``pid``) and number of monitored virtual address
+regions in this access pattern (``nr_regions``).  After that, each line shows
+start/end address, size, and number of monitored accesses to the region for
+each of the regions.
+
+
+heats
+~~~~~
+
+The ``raw`` type shows detailed information but it is exhaustive to manually
+read and analyzed.  For the reason, ``heats`` plots the data in heatmap form,
+using time as x-axis, virtual address as y-axis, and access frequency as
+z-axis.  Also, users set the resolution and start/end point of each axis via
+optional arguments.  For example::
+
+    $ ./damo report heats --tres 3 --ares 3
+    0               0               0.0
+    0               7609002         0.0
+    0               15218004        0.0
+    66112620851     0               0.0
+    66112620851     7609002         0.0
+    66112620851     15218004        0.0
+    132225241702    0               0.0
+    132225241702    7609002         0.0
+    132225241702    15218004        0.0
+
+This command shows the recorded access pattern of the ``sleep`` command using 3
+data points for each of time axis and address axis.  Therefore, it shows 9 data
+points in total.
+
+Users can easily converts this text output into heatmap image or other 3D
+representation using various tools such as 'gnuplot'.  ``raw`` sub-subcommand
+also provides 'gnuplot' based heatmap image creation.  For this, you can use
+``--heatmap`` option.  Also, note that because it uses 'gnuplot' internally, it
+will fail if 'gnuplot' is not installed on your system.  For example::
+
+    $ ./damo report heats --heatmap heatmap.png
+
+Creates ``heatmap.png`` file containing the heatmap image.  It supports
+``pdf``, ``png``, ``jpeg``, and ``svg``.
+
+For proper zoom in / zoom out, you need to see the layout of the record.  For
+that, use '--guide' option.  If the option is given, it will provide useful
+information about the records in the record file.  For example::
+
+    $ ./damo report heats --guide
+    pid:1348
+    time: 193485829398-198337863555 (4852034157)
+    region   0: 00000094564599762944-00000094564622589952 (22827008)
+    region   1: 00000140454009610240-00000140454016012288 (6402048)
+    region   2: 00000140731597193216-00000140731597443072 (249856)
+
+The output shows monitored regions (start and end addresses in byte) and
+monitored time duration (start and end time in nanosecond) of each target task.
+Therefore, it would be wise to plot only each region rather than plotting
+entire address space in one heatmap because the gaps between the regions are so
+huge in this case.
+
+
+wss
+~~~
+
+The ``wss`` type shows the distribution or time-varying working set sizes of
+the recorded workload using the records.  For example::
+
+    $ ./damo report wss
+    # <percentile> <wss>
+    # pid   1348
+    # avr:  66228
+    0       0
+    25      0
+    50      0
+    75      0
+    100     1920615
+
+Without any option, it shows the distribution of the working set sizes as
+above.  Basically it shows 0th, 25th, 50th, 75th, and 100th percentile and
+average of the measured working set sizes in the access pattern records.  In
+this case, the working set size was zero for 75th percentile but 1,920,615
+bytes in max and 66,228 in average.
+
+By setting the sort key of the percentile using '--sortby', you can also see
+how the working set size is chronologically changed.  For example::
+
+    $ ./damo report wss --sortby time
+    # <percentile> <wss>
+    # pid   1348
+    # avr:  66228
+    0       0
+    25      0
+    50      0
+    75      0
+    100     0
+
+The average is still 66,228.  And, because we sorted the working set using
+recorded time and the access is very short, we cannot show when the access
+made.
+
+Users can specify the resolution of the distribution (``--range``).  It also
+supports 'gnuplot' based simple visualization (``--plot``) of the distribution.
diff --git a/Documentation/admin-guide/mm/index.rst b/Documentation/admin-guide/mm/index.rst
index 11db46448354..d3d0ba373eb6 100644
--- a/Documentation/admin-guide/mm/index.rst
+++ b/Documentation/admin-guide/mm/index.rst
@@ -27,6 +27,7 @@ the Linux memory management.
 
    concepts
    cma_debugfs
+   data_access_monitor
    hugetlbpage
    idle_page_tracking
    ksm
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 13/15] mm/damon: Add kunit tests
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (11 preceding siblings ...)
  2020-03-18 11:27 ` [PATCH v7 12/15] Documentation/admin-guide/mm: Add a document " SeongJae Park
@ 2020-03-18 11:27 ` SeongJae Park
  2020-03-18 11:38 ` [PATCH v7 14/15] mm/damon: Add user space selftests SeongJae Park
                   ` (2 subsequent siblings)
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:27 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit adds kunit based unit tests for DAMON.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
Reviewed-by: Brendan Higgins <brendanhiggins@google.com>
---
 mm/Kconfig      |  11 +
 mm/damon-test.h | 604 ++++++++++++++++++++++++++++++++++++++++++++++++
 mm/damon.c      |   2 +
 3 files changed, 617 insertions(+)
 create mode 100644 mm/damon-test.h

diff --git a/mm/Kconfig b/mm/Kconfig
index 79da00d15604..ba5ee66f5c2f 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -750,4 +750,15 @@ config DAMON
 	  be 1) accurate enough to be useful for performance-centric domains,
 	  and 2) sufficiently light-weight so that it can be applied online.
 
+config DAMON_KUNIT_TEST
+	bool "Test for damon"
+	depends on DAMON=y && KUNIT
+	help
+	  This builds the DAMON Kunit test suite.
+
+	  For more information on KUnit and unit tests in general, please refer
+	  to the KUnit documentation.
+
+	  If unsure, say N.
+
 endmenu
diff --git a/mm/damon-test.h b/mm/damon-test.h
new file mode 100644
index 000000000000..498c637b78ff
--- /dev/null
+++ b/mm/damon-test.h
@@ -0,0 +1,604 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Data Access Monitor Unit Tests
+ *
+ * Copyright 2019 Amazon.com, Inc. or its affiliates.  All rights reserved.
+ *
+ * Author: SeongJae Park <sjpark@amazon.de>
+ */
+
+#ifdef CONFIG_DAMON_KUNIT_TEST
+
+#ifndef _DAMON_TEST_H
+#define _DAMON_TEST_H
+
+#include <kunit/test.h>
+
+static void damon_test_str_to_pids(struct kunit *test)
+{
+	char *question;
+	unsigned long *answers;
+	unsigned long expected[] = {12, 35, 46};
+	ssize_t nr_integers = 0, i;
+
+	question = "123";
+	answers = str_to_pids(question, strnlen(question, 128), &nr_integers);
+	KUNIT_EXPECT_EQ(test, (ssize_t)1, nr_integers);
+	KUNIT_EXPECT_EQ(test, 123ul, answers[0]);
+	kfree(answers);
+
+	question = "123abc";
+	answers = str_to_pids(question, strnlen(question, 128), &nr_integers);
+	KUNIT_EXPECT_EQ(test, (ssize_t)1, nr_integers);
+	KUNIT_EXPECT_EQ(test, 123ul, answers[0]);
+	kfree(answers);
+
+	question = "a123";
+	answers = str_to_pids(question, strnlen(question, 128), &nr_integers);
+	KUNIT_EXPECT_EQ(test, (ssize_t)0, nr_integers);
+	KUNIT_EXPECT_PTR_EQ(test, answers, (unsigned long *)NULL);
+
+	question = "12 35";
+	answers = str_to_pids(question, strnlen(question, 128), &nr_integers);
+	KUNIT_EXPECT_EQ(test, (ssize_t)2, nr_integers);
+	for (i = 0; i < nr_integers; i++)
+		KUNIT_EXPECT_EQ(test, expected[i], answers[i]);
+	kfree(answers);
+
+	question = "12 35 46";
+	answers = str_to_pids(question, strnlen(question, 128), &nr_integers);
+	KUNIT_EXPECT_EQ(test, (ssize_t)3, nr_integers);
+	for (i = 0; i < nr_integers; i++)
+		KUNIT_EXPECT_EQ(test, expected[i], answers[i]);
+	kfree(answers);
+
+	question = "12 35 abc 46";
+	answers = str_to_pids(question, strnlen(question, 128), &nr_integers);
+	KUNIT_EXPECT_EQ(test, (ssize_t)2, nr_integers);
+	for (i = 0; i < 2; i++)
+		KUNIT_EXPECT_EQ(test, expected[i], answers[i]);
+	kfree(answers);
+
+	question = "";
+	answers = str_to_pids(question, strnlen(question, 128), &nr_integers);
+	KUNIT_EXPECT_EQ(test, (ssize_t)0, nr_integers);
+	KUNIT_EXPECT_PTR_EQ(test, (unsigned long *)NULL, answers);
+	kfree(answers);
+
+	question = "\n";
+	answers = str_to_pids(question, strnlen(question, 128), &nr_integers);
+	KUNIT_EXPECT_EQ(test, (ssize_t)0, nr_integers);
+	KUNIT_EXPECT_PTR_EQ(test, (unsigned long *)NULL, answers);
+	kfree(answers);
+}
+
+static void damon_test_regions(struct kunit *test)
+{
+	struct damon_region *r;
+	struct damon_task *t;
+
+	r = damon_new_region(&damon_user_ctx, 1, 2);
+	KUNIT_EXPECT_EQ(test, 1ul, r->vm_start);
+	KUNIT_EXPECT_EQ(test, 2ul, r->vm_end);
+	KUNIT_EXPECT_EQ(test, 0u, r->nr_accesses);
+	KUNIT_EXPECT_TRUE(test, r->sampling_addr >= r->vm_start);
+	KUNIT_EXPECT_TRUE(test, r->sampling_addr < r->vm_end);
+
+	t = damon_new_task(42);
+	KUNIT_EXPECT_EQ(test, 0u, nr_damon_regions(t));
+
+	damon_add_region(r, t);
+	KUNIT_EXPECT_EQ(test, 1u, nr_damon_regions(t));
+
+	damon_del_region(r);
+	KUNIT_EXPECT_EQ(test, 0u, nr_damon_regions(t));
+
+	damon_free_task(t);
+}
+
+static void damon_test_tasks(struct kunit *test)
+{
+	struct damon_ctx *c = &damon_user_ctx;
+	struct damon_task *t;
+
+	t = damon_new_task(42);
+	KUNIT_EXPECT_EQ(test, 42ul, t->pid);
+	KUNIT_EXPECT_EQ(test, 0u, nr_damon_tasks(c));
+
+	damon_add_task(&damon_user_ctx, t);
+	KUNIT_EXPECT_EQ(test, 1u, nr_damon_tasks(c));
+
+	damon_destroy_task(t);
+	KUNIT_EXPECT_EQ(test, 0u, nr_damon_tasks(c));
+}
+
+static void damon_test_set_pids(struct kunit *test)
+{
+	struct damon_ctx *ctx = &damon_user_ctx;
+	unsigned long pids[] = {1, 2, 3};
+	char buf[64];
+
+	damon_set_pids(ctx, pids, 3);
+	damon_sprint_pids(ctx, buf, 64);
+	KUNIT_EXPECT_STREQ(test, (char *)buf, "1 2 3\n");
+
+	damon_set_pids(ctx, NULL, 0);
+	damon_sprint_pids(ctx, buf, 64);
+	KUNIT_EXPECT_STREQ(test, (char *)buf, "\n");
+
+	damon_set_pids(ctx, (unsigned long []){1, 2}, 2);
+	damon_sprint_pids(ctx, buf, 64);
+	KUNIT_EXPECT_STREQ(test, (char *)buf, "1 2\n");
+
+	damon_set_pids(ctx, (unsigned long []){2}, 1);
+	damon_sprint_pids(ctx, buf, 64);
+	KUNIT_EXPECT_STREQ(test, (char *)buf, "2\n");
+
+	damon_set_pids(ctx, NULL, 0);
+	damon_sprint_pids(ctx, buf, 64);
+	KUNIT_EXPECT_STREQ(test, (char *)buf, "\n");
+}
+
+/*
+ * Test damon_three_regions_in_vmas() function
+ *
+ * DAMON converts the complex and dynamic memory mappings of each target task
+ * to three discontiguous regions which cover every mapped areas.  However, the
+ * three regions should not include the two biggest unmapped areas in the
+ * original mapping, because the two biggest areas are normally the areas
+ * between 1) heap and the mmap()-ed regions, and 2) the mmap()-ed regions and
+ * stack.  Because these two unmapped areas are very huge but obviously never
+ * accessed, covering the region is just a waste.
+ *
+ * 'damon_three_regions_in_vmas() receives an address space of a process.  It
+ * first identifies the start of mappings, end of mappings, and the two biggest
+ * unmapped areas.  After that, based on the information, it constructs the
+ * three regions and returns.  For more detail, refer to the comment of
+ * 'damon_init_regions_of()' function definition in 'mm/damon.c' file.
+ *
+ * For example, suppose virtual address ranges of 10-20, 20-25, 200-210,
+ * 210-220, 300-305, and 307-330 (Other comments represent this mappings in
+ * more short form: 10-20-25, 200-210-220, 300-305, 307-330) of a process are
+ * mapped.  To cover every mappings, the three regions should start with 10,
+ * and end with 305.  The process also has three unmapped areas, 25-200,
+ * 220-300, and 305-307.  Among those, 25-200 and 220-300 are the biggest two
+ * unmapped areas, and thus it should be converted to three regions of 10-25,
+ * 200-220, and 300-330.
+ */
+static void damon_test_three_regions_in_vmas(struct kunit *test)
+{
+	struct region regions[3] = {0,};
+	/* 10-20-25, 200-210-220, 300-305, 307-330 */
+	struct vm_area_struct vmas[] = {
+		(struct vm_area_struct) {.vm_start = 10, .vm_end = 20},
+		(struct vm_area_struct) {.vm_start = 20, .vm_end = 25},
+		(struct vm_area_struct) {.vm_start = 200, .vm_end = 210},
+		(struct vm_area_struct) {.vm_start = 210, .vm_end = 220},
+		(struct vm_area_struct) {.vm_start = 300, .vm_end = 305},
+		(struct vm_area_struct) {.vm_start = 307, .vm_end = 330},
+	};
+	vmas[0].vm_next = &vmas[1];
+	vmas[1].vm_next = &vmas[2];
+	vmas[2].vm_next = &vmas[3];
+	vmas[3].vm_next = &vmas[4];
+	vmas[4].vm_next = &vmas[5];
+	vmas[5].vm_next = NULL;
+
+	damon_three_regions_in_vmas(&vmas[0], regions);
+
+	KUNIT_EXPECT_EQ(test, 10ul, regions[0].start);
+	KUNIT_EXPECT_EQ(test, 25ul, regions[0].end);
+	KUNIT_EXPECT_EQ(test, 200ul, regions[1].start);
+	KUNIT_EXPECT_EQ(test, 220ul, regions[1].end);
+	KUNIT_EXPECT_EQ(test, 300ul, regions[2].start);
+	KUNIT_EXPECT_EQ(test, 330ul, regions[2].end);
+}
+
+/* Clean up global state of damon */
+static void damon_cleanup_global_state(void)
+{
+	struct damon_task *t, *next;
+
+	damon_for_each_task_safe(&damon_user_ctx, t, next)
+		damon_destroy_task(t);
+
+	damon_user_ctx.rbuf_offset = 0;
+}
+
+/*
+ * Test kdamond_reset_aggregated()
+ *
+ * DAMON checks access to each region and aggregates this information as the
+ * access frequency of each region.  In detail, it increases '->nr_accesses' of
+ * regions that an access has confirmed.  'kdamond_reset_aggregated()' flushes
+ * the aggregated information ('->nr_accesses' of each regions) to the result
+ * buffer.  As a result of the flushing, the '->nr_accesses' of regions are
+ * initialized to zero.
+ */
+static void damon_test_aggregate(struct kunit *test)
+{
+	struct damon_ctx *ctx = &damon_user_ctx;
+	unsigned long pids[] = {1, 2, 3};
+	unsigned long saddr[][3] = {{10, 20, 30}, {5, 42, 49}, {13, 33, 55} };
+	unsigned long eaddr[][3] = {{15, 27, 40}, {31, 45, 55}, {23, 44, 66} };
+	unsigned long accesses[][3] = {{42, 95, 84}, {10, 20, 30}, {0, 1, 2} };
+	struct damon_task *t;
+	struct damon_region *r;
+	int it, ir;
+	ssize_t sz, sr, sp;
+
+	damon_set_recording(ctx, 256, "damon.data");
+	damon_set_pids(ctx, pids, 3);
+
+	it = 0;
+	damon_for_each_task(ctx, t) {
+		for (ir = 0; ir < 3; ir++) {
+			r = damon_new_region(ctx,
+					saddr[it][ir], eaddr[it][ir]);
+			r->nr_accesses = accesses[it][ir];
+			damon_add_region(r, t);
+		}
+		it++;
+	}
+	kdamond_reset_aggregated(ctx);
+	it = 0;
+	damon_for_each_task(ctx, t) {
+		ir = 0;
+		/* '->nr_accesses' should be zeroed */
+		damon_for_each_region(r, t) {
+			KUNIT_EXPECT_EQ(test, 0u, r->nr_accesses);
+			ir++;
+		}
+		/* regions should be preserved */
+		KUNIT_EXPECT_EQ(test, 3, ir);
+		it++;
+	}
+	/* tasks also should be preserved */
+	KUNIT_EXPECT_EQ(test, 3, it);
+
+	/* The aggregated information should be written in the buffer */
+	sr = sizeof(r->vm_start) + sizeof(r->vm_end) + sizeof(r->nr_accesses);
+	sp = sizeof(t->pid) + sizeof(unsigned int) + 3 * sr;
+	sz = sizeof(struct timespec64) + sizeof(unsigned int) + 3 * sp;
+	KUNIT_EXPECT_EQ(test, (unsigned int)sz, ctx->rbuf_offset);
+
+	damon_set_recording(ctx, 0, "damon.data");
+	damon_cleanup_global_state();
+}
+
+static void damon_test_write_rbuf(struct kunit *test)
+{
+	struct damon_ctx *ctx = &damon_user_ctx;
+	char *data;
+
+	damon_set_recording(&damon_user_ctx, 256, "damon.data");
+
+	data = "hello";
+	damon_write_rbuf(ctx, data, strnlen(data, 256));
+	KUNIT_EXPECT_EQ(test, ctx->rbuf_offset, 5u);
+
+	damon_write_rbuf(ctx, data, 0);
+	KUNIT_EXPECT_EQ(test, ctx->rbuf_offset, 5u);
+
+	KUNIT_EXPECT_STREQ(test, (char *)ctx->rbuf, data);
+	damon_set_recording(&damon_user_ctx, 0, "damon.data");
+}
+
+/*
+ * Test 'damon_apply_three_regions()'
+ *
+ * test			kunit object
+ * regions		an array containing start/end addresses of current
+ *			monitoring target regions
+ * nr_regions		the number of the addresses in 'regions'
+ * three_regions	The three regions that need to be applied now
+ * expected		start/end addresses of monitoring target regions that
+ *			'three_regions' are applied
+ * nr_expected		the number of addresses in 'expected'
+ *
+ * The memory mapping of the target processes changes dynamically.  To follow
+ * the change, DAMON periodically reads the mappings, simplifies it to the
+ * three regions, and updates the monitoring target regions to fit in the three
+ * regions.  The update of current target regions is the role of
+ * 'damon_apply_three_regions()'.
+ *
+ * This test passes the given target regions and the new three regions that
+ * need to be applied to the function and check whether it updates the regions
+ * as expected.
+ */
+static void damon_do_test_apply_three_regions(struct kunit *test,
+				unsigned long *regions, int nr_regions,
+				struct region *three_regions,
+				unsigned long *expected, int nr_expected)
+{
+	struct damon_task *t;
+	struct damon_region *r;
+	int i;
+
+	t = damon_new_task(42);
+	for (i = 0; i < nr_regions / 2; i++) {
+		r = damon_new_region(&damon_user_ctx,
+				regions[i * 2], regions[i * 2 + 1]);
+		damon_add_region(r, t);
+	}
+	damon_add_task(&damon_user_ctx, t);
+
+	damon_apply_three_regions(&damon_user_ctx, t, three_regions);
+
+	for (i = 0; i < nr_expected / 2; i++) {
+		r = damon_nth_region_of(t, i);
+		KUNIT_EXPECT_EQ(test, r->vm_start, expected[i * 2]);
+		KUNIT_EXPECT_EQ(test, r->vm_end, expected[i * 2 + 1]);
+	}
+
+	damon_cleanup_global_state();
+}
+
+/*
+ * This function test most common case where the three big regions are only
+ * slightly changed.  Target regions should adjust their boundary (10-20-30,
+ * 50-55, 70-80, 90-100) to fit with the new big regions or remove target
+ * regions (57-79) that now out of the three regions.
+ */
+static void damon_test_apply_three_regions1(struct kunit *test)
+{
+	/* 10-20-30, 50-55-57-59, 70-80-90-100 */
+	unsigned long regions[] = {10, 20, 20, 30, 50, 55, 55, 57, 57, 59,
+				70, 80, 80, 90, 90, 100};
+	/* 5-27, 45-55, 73-104 */
+	struct region new_three_regions[3] = {
+		(struct region){.start = 5, .end = 27},
+		(struct region){.start = 45, .end = 55},
+		(struct region){.start = 73, .end = 104} };
+	/* 5-20-27, 45-55, 73-80-90-104 */
+	unsigned long expected[] = {5, 20, 20, 27, 45, 55,
+				73, 80, 80, 90, 90, 104};
+
+	damon_do_test_apply_three_regions(test, regions, ARRAY_SIZE(regions),
+			new_three_regions, expected, ARRAY_SIZE(expected));
+}
+
+/*
+ * Test slightly bigger change.  Similar to above, but the second big region
+ * now require two target regions (50-55, 57-59) to be removed.
+ */
+static void damon_test_apply_three_regions2(struct kunit *test)
+{
+	/* 10-20-30, 50-55-57-59, 70-80-90-100 */
+	unsigned long regions[] = {10, 20, 20, 30, 50, 55, 55, 57, 57, 59,
+				70, 80, 80, 90, 90, 100};
+	/* 5-27, 56-57, 65-104 */
+	struct region new_three_regions[3] = {
+		(struct region){.start = 5, .end = 27},
+		(struct region){.start = 56, .end = 57},
+		(struct region){.start = 65, .end = 104} };
+	/* 5-20-27, 56-57, 65-80-90-104 */
+	unsigned long expected[] = {5, 20, 20, 27, 56, 57,
+				65, 80, 80, 90, 90, 104};
+
+	damon_do_test_apply_three_regions(test, regions, ARRAY_SIZE(regions),
+			new_three_regions, expected, ARRAY_SIZE(expected));
+}
+
+/*
+ * Test a big change.  The second big region has totally freed and mapped to
+ * different area (50-59 -> 61-63).  The target regions which were in the old
+ * second big region (50-55-57-59) should be removed and new target region
+ * covering the second big region (61-63) should be created.
+ */
+static void damon_test_apply_three_regions3(struct kunit *test)
+{
+	/* 10-20-30, 50-55-57-59, 70-80-90-100 */
+	unsigned long regions[] = {10, 20, 20, 30, 50, 55, 55, 57, 57, 59,
+				70, 80, 80, 90, 90, 100};
+	/* 5-27, 61-63, 65-104 */
+	struct region new_three_regions[3] = {
+		(struct region){.start = 5, .end = 27},
+		(struct region){.start = 61, .end = 63},
+		(struct region){.start = 65, .end = 104} };
+	/* 5-20-27, 61-63, 65-80-90-104 */
+	unsigned long expected[] = {5, 20, 20, 27, 61, 63,
+				65, 80, 80, 90, 90, 104};
+
+	damon_do_test_apply_three_regions(test, regions, ARRAY_SIZE(regions),
+			new_three_regions, expected, ARRAY_SIZE(expected));
+}
+
+/*
+ * Test another big change.  Both of the second and third big regions (50-59
+ * and 70-100) has totally freed and mapped to different area (30-32 and
+ * 65-68).  The target regions which were in the old second and third big
+ * regions should now be removed and new target regions covering the new second
+ * and third big regions should be crated.
+ */
+static void damon_test_apply_three_regions4(struct kunit *test)
+{
+	/* 10-20-30, 50-55-57-59, 70-80-90-100 */
+	unsigned long regions[] = {10, 20, 20, 30, 50, 55, 55, 57, 57, 59,
+				70, 80, 80, 90, 90, 100};
+	/* 5-7, 30-32, 65-68 */
+	struct region new_three_regions[3] = {
+		(struct region){.start = 5, .end = 7},
+		(struct region){.start = 30, .end = 32},
+		(struct region){.start = 65, .end = 68} };
+	/* expect 5-7, 30-32, 65-68 */
+	unsigned long expected[] = {5, 7, 30, 32, 65, 68};
+
+	damon_do_test_apply_three_regions(test, regions, ARRAY_SIZE(regions),
+			new_three_regions, expected, ARRAY_SIZE(expected));
+}
+
+static void damon_test_split_evenly(struct kunit *test)
+{
+	struct damon_ctx *c = &damon_user_ctx;
+	struct damon_task *t;
+	struct damon_region *r;
+	unsigned long i;
+
+	KUNIT_EXPECT_EQ(test, damon_split_region_evenly(c, NULL, 5), -EINVAL);
+
+	t = damon_new_task(42);
+	r = damon_new_region(&damon_user_ctx, 0, 100);
+	KUNIT_EXPECT_EQ(test, damon_split_region_evenly(c, r, 0), -EINVAL);
+
+	damon_add_region(r, t);
+	KUNIT_EXPECT_EQ(test, damon_split_region_evenly(c, r, 10), 0);
+	KUNIT_EXPECT_EQ(test, nr_damon_regions(t), 10u);
+
+	i = 0;
+	damon_for_each_region(r, t) {
+		KUNIT_EXPECT_EQ(test, r->vm_start, i++ * 10);
+		KUNIT_EXPECT_EQ(test, r->vm_end, i * 10);
+	}
+	damon_free_task(t);
+
+	t = damon_new_task(42);
+	r = damon_new_region(&damon_user_ctx, 5, 59);
+	damon_add_region(r, t);
+	KUNIT_EXPECT_EQ(test, damon_split_region_evenly(c, r, 5), 0);
+	KUNIT_EXPECT_EQ(test, nr_damon_regions(t), 5u);
+
+	i = 0;
+	damon_for_each_region(r, t) {
+		if (i == 4)
+			break;
+		KUNIT_EXPECT_EQ(test, r->vm_start, 5 + 10 * i++);
+		KUNIT_EXPECT_EQ(test, r->vm_end, 5 + 10 * i);
+	}
+	KUNIT_EXPECT_EQ(test, r->vm_start, 5 + 10 * i);
+	KUNIT_EXPECT_EQ(test, r->vm_end, 59ul);
+	damon_free_task(t);
+
+	t = damon_new_task(42);
+	r = damon_new_region(&damon_user_ctx, 5, 6);
+	damon_add_region(r, t);
+	KUNIT_EXPECT_EQ(test, damon_split_region_evenly(c, r, 2), -EINVAL);
+	KUNIT_EXPECT_EQ(test, nr_damon_regions(t), 1u);
+
+	damon_for_each_region(r, t) {
+		KUNIT_EXPECT_EQ(test, r->vm_start, 5ul);
+		KUNIT_EXPECT_EQ(test, r->vm_end, 6ul);
+	}
+	damon_free_task(t);
+}
+
+static void damon_test_split_at(struct kunit *test)
+{
+	struct damon_task *t;
+	struct damon_region *r;
+
+	t = damon_new_task(42);
+	r = damon_new_region(&damon_user_ctx, 0, 100);
+	damon_add_region(r, t);
+	damon_split_region_at(&damon_user_ctx, r, 25);
+	KUNIT_EXPECT_EQ(test, r->vm_start, 0ul);
+	KUNIT_EXPECT_EQ(test, r->vm_end, 25ul);
+
+	r = damon_next_region(r);
+	KUNIT_EXPECT_EQ(test, r->vm_start, 25ul);
+	KUNIT_EXPECT_EQ(test, r->vm_end, 100ul);
+
+	damon_free_task(t);
+}
+
+static void damon_test_merge_two(struct kunit *test)
+{
+	struct damon_task *t;
+	struct damon_region *r, *r2, *r3;
+	int i;
+
+	t = damon_new_task(42);
+	r = damon_new_region(&damon_user_ctx, 0, 100);
+	r->nr_accesses = 10;
+	damon_add_region(r, t);
+	r2 = damon_new_region(&damon_user_ctx, 100, 300);
+	r2->nr_accesses = 20;
+	damon_add_region(r2, t);
+
+	damon_merge_two_regions(r, r2);
+	KUNIT_EXPECT_EQ(test, r->vm_start, 0ul);
+	KUNIT_EXPECT_EQ(test, r->vm_end, 300ul);
+	KUNIT_EXPECT_EQ(test, r->nr_accesses, 16u);
+
+	i = 0;
+	damon_for_each_region(r3, t) {
+		KUNIT_EXPECT_PTR_EQ(test, r, r3);
+		i++;
+	}
+	KUNIT_EXPECT_EQ(test, i, 1);
+
+	damon_free_task(t);
+}
+
+static void damon_test_merge_regions_of(struct kunit *test)
+{
+	struct damon_task *t;
+	struct damon_region *r;
+	unsigned long sa[] = {0, 100, 114, 122, 130, 156, 170, 184};
+	unsigned long ea[] = {100, 112, 122, 130, 156, 170, 184, 230};
+	unsigned int nrs[] = {0, 0, 10, 10, 20, 30, 1, 2};
+
+	unsigned long saddrs[] = {0, 114, 130, 156, 170};
+	unsigned long eaddrs[] = {112, 130, 156, 170, 230};
+	int i;
+
+	t = damon_new_task(42);
+	for (i = 0; i < ARRAY_SIZE(sa); i++) {
+		r = damon_new_region(&damon_user_ctx, sa[i], ea[i]);
+		r->nr_accesses = nrs[i];
+		damon_add_region(r, t);
+	}
+
+	damon_merge_regions_of(t, 9);
+	/* 0-112, 114-130, 130-156, 156-170 */
+	KUNIT_EXPECT_EQ(test, nr_damon_regions(t), 5u);
+	for (i = 0; i < 5; i++) {
+		r = damon_nth_region_of(t, i);
+		KUNIT_EXPECT_EQ(test, r->vm_start, saddrs[i]);
+		KUNIT_EXPECT_EQ(test, r->vm_end, eaddrs[i]);
+	}
+	damon_free_task(t);
+}
+
+static void damon_test_split_regions_of(struct kunit *test)
+{
+	struct damon_task *t;
+	struct damon_region *r;
+
+	t = damon_new_task(42);
+	r = damon_new_region(&damon_user_ctx, 0, 22);
+	damon_add_region(r, t);
+	damon_split_regions_of(&damon_user_ctx, t);
+	KUNIT_EXPECT_EQ(test, nr_damon_regions(t), 2u);
+	damon_free_task(t);
+}
+
+static struct kunit_case damon_test_cases[] = {
+	KUNIT_CASE(damon_test_str_to_pids),
+	KUNIT_CASE(damon_test_tasks),
+	KUNIT_CASE(damon_test_regions),
+	KUNIT_CASE(damon_test_set_pids),
+	KUNIT_CASE(damon_test_three_regions_in_vmas),
+	KUNIT_CASE(damon_test_aggregate),
+	KUNIT_CASE(damon_test_write_rbuf),
+	KUNIT_CASE(damon_test_apply_three_regions1),
+	KUNIT_CASE(damon_test_apply_three_regions2),
+	KUNIT_CASE(damon_test_apply_three_regions3),
+	KUNIT_CASE(damon_test_apply_three_regions4),
+	KUNIT_CASE(damon_test_split_evenly),
+	KUNIT_CASE(damon_test_split_at),
+	KUNIT_CASE(damon_test_merge_two),
+	KUNIT_CASE(damon_test_merge_regions_of),
+	KUNIT_CASE(damon_test_split_regions_of),
+	{},
+};
+
+static struct kunit_suite damon_test_suite = {
+	.name = "damon",
+	.test_cases = damon_test_cases,
+};
+kunit_test_suite(damon_test_suite);
+
+#endif /* _DAMON_TEST_H */
+
+#endif	/* CONFIG_DAMON_KUNIT_TEST */
diff --git a/mm/damon.c b/mm/damon.c
index 25c961fabdf4..4ca8a822c30c 100644
--- a/mm/damon.c
+++ b/mm/damon.c
@@ -1433,3 +1433,5 @@ module_exit(damon_exit);
 MODULE_LICENSE("GPL");
 MODULE_AUTHOR("SeongJae Park <sjpark@amazon.de>");
 MODULE_DESCRIPTION("DAMON: Data Access MONitor");
+
+#include "damon-test.h"
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 14/15] mm/damon: Add user space selftests
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (12 preceding siblings ...)
  2020-03-18 11:27 ` [PATCH v7 13/15] mm/damon: Add kunit tests SeongJae Park
@ 2020-03-18 11:38 ` SeongJae Park
  2020-03-18 11:39 ` [PATCH v7 15/15] MAINTAINERS: Update for DAMON SeongJae Park
  2020-03-30 10:53 ` [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:38 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit adds a simple user space tests for DAMON.  The tests are
using kselftest framework.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 tools/testing/selftests/damon/Makefile        |   7 +
 .../selftests/damon/_chk_dependency.sh        |  28 ++++
 tools/testing/selftests/damon/_chk_record.py  |  89 +++++++++++
 .../testing/selftests/damon/debugfs_attrs.sh  | 139 ++++++++++++++++++
 .../testing/selftests/damon/debugfs_record.sh |  50 +++++++
 5 files changed, 313 insertions(+)
 create mode 100644 tools/testing/selftests/damon/Makefile
 create mode 100644 tools/testing/selftests/damon/_chk_dependency.sh
 create mode 100644 tools/testing/selftests/damon/_chk_record.py
 create mode 100755 tools/testing/selftests/damon/debugfs_attrs.sh
 create mode 100755 tools/testing/selftests/damon/debugfs_record.sh

diff --git a/tools/testing/selftests/damon/Makefile b/tools/testing/selftests/damon/Makefile
new file mode 100644
index 000000000000..cfd5393a4639
--- /dev/null
+++ b/tools/testing/selftests/damon/Makefile
@@ -0,0 +1,7 @@
+# SPDX-License-Identifier: GPL-2.0
+# Makefile for damon selftests
+
+TEST_FILES = _chk_dependency.sh _chk_record_file.py
+TEST_PROGS = debugfs_attrs.sh debugfs_record.sh
+
+include ../lib.mk
diff --git a/tools/testing/selftests/damon/_chk_dependency.sh b/tools/testing/selftests/damon/_chk_dependency.sh
new file mode 100644
index 000000000000..814dcadd5e96
--- /dev/null
+++ b/tools/testing/selftests/damon/_chk_dependency.sh
@@ -0,0 +1,28 @@
+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0
+
+# Kselftest framework requirement - SKIP code is 4.
+ksft_skip=4
+
+DBGFS=/sys/kernel/debug/damon
+
+if [ $EUID -ne 0 ];
+then
+	echo "Run as root"
+	exit $ksft_skip
+fi
+
+if [ ! -d $DBGFS ]
+then
+	echo "$DBGFS not found"
+	exit $ksft_skip
+fi
+
+for f in attrs record pids monitor_on
+do
+	if [ ! -f "$DBGFS/$f" ]
+	then
+		echo "$f not found"
+		exit 1
+	fi
+done
diff --git a/tools/testing/selftests/damon/_chk_record.py b/tools/testing/selftests/damon/_chk_record.py
new file mode 100644
index 000000000000..ef55f478c2af
--- /dev/null
+++ b/tools/testing/selftests/damon/_chk_record.py
@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: GPL-2.0
+
+"Check whether the DAMON record file is valid"
+
+import argparse
+import struct
+import sys
+
+def err_percent(val, expected):
+    return abs(val - expected) / expected * 100
+
+def chk_task_info(f):
+    pid = struct.unpack('L', f.read(8))[0]
+    nr_regions = struct.unpack('I', f.read(4))[0]
+
+    if nr_regions > max_nr_regions:
+        print('too many regions: %d > %d' % (nr_regions, max_nr_regions))
+        exit(1)
+
+    nr_gaps = 0
+    eaddr = 0
+    for r in range(nr_regions):
+        saddr = struct.unpack('L', f.read(8))[0]
+        if eaddr and saddr != eaddr:
+            nr_gaps += 1
+        eaddr = struct.unpack('L', f.read(8))[0]
+        nr_accesses = struct.unpack('I', f.read(4))[0]
+
+        if saddr >= eaddr:
+            print('wrong region [%d,%d)' % (saddr, eaddr))
+            exit(1)
+
+        max_nr_accesses = aint / sint
+        if nr_accesses > max_nr_accesses:
+            if err_percent(nr_accesses, max_nr_accesses) > 15:
+                print('too high nr_access: expected %d but %d' %
+                        (max_nr_accesses, nr_accesses))
+                exit(1)
+    if nr_gaps != 2:
+        print('number of gaps are not two but %d' % nr_gaps)
+        exit(1)
+
+def parse_time_us(bindat):
+    sec = struct.unpack('l', bindat[0:8])[0]
+    nsec = struct.unpack('l', bindat[8:16])[0]
+    return (sec * 1000000000 + nsec) / 1000
+
+def main():
+    global sint
+    global aint
+    global min_nr
+    global max_nr_regions
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('file', metavar='<file>',
+            help='path to the record file')
+    parser.add_argument('--attrs', metavar='<attrs>',
+            default='5000 100000 1000000 10 1000',
+            help='content of debugfs attrs file')
+    args = parser.parse_args()
+    file_path = args.file
+    attrs = [int(x) for x in args.attrs.split()]
+    sint, aint, rint, min_nr, max_nr_regions = attrs
+
+    with open(file_path, 'rb') as f:
+        last_aggr_time = None
+        while True:
+            timebin = f.read(16)
+            if len(timebin) != 16:
+                break
+
+            now = parse_time_us(timebin)
+            if not last_aggr_time:
+                last_aggr_time = now
+            else:
+                error = err_percent(now - last_aggr_time, aint)
+                if error > 15:
+                    print('wrong aggr interval: expected %d, but %d' %
+                            (aint, now - last_aggr_time))
+                    exit(1)
+                last_aggr_time = now
+
+            nr_tasks = struct.unpack('I', f.read(4))[0]
+            for t in range(nr_tasks):
+                chk_task_info(f)
+
+if __name__ == '__main__':
+    main()
diff --git a/tools/testing/selftests/damon/debugfs_attrs.sh b/tools/testing/selftests/damon/debugfs_attrs.sh
new file mode 100755
index 000000000000..d5188b0f71b1
--- /dev/null
+++ b/tools/testing/selftests/damon/debugfs_attrs.sh
@@ -0,0 +1,139 @@
+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0
+
+source ./_chk_dependency.sh
+
+# Test attrs file
+file="$DBGFS/attrs"
+
+ORIG_CONTENT=$(cat $file)
+
+echo 1 2 3 4 5 > $file
+if [ $? -ne 0 ]
+then
+	echo "$file write failed"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+echo 1 2 3 4 > $file
+if [ $? -eq 0 ]
+then
+	echo "$file write success (should failed)"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+CONTENT=$(cat $file)
+if [ "$CONTENT" != "1 2 3 4 5" ]
+then
+	echo "$file not written"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+echo $ORIG_CONTENT > $file
+
+# Test record file
+file="$DBGFS/record"
+
+ORIG_CONTENT=$(cat $file)
+
+echo "4242 foo.bar" > $file
+if [ $? -ne 0 ]
+then
+	echo "$file writing sane input failed"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+echo abc 2 3 > $file
+if [ $? -eq 0 ]
+then
+	echo "$file writing insane input 1 success (should failed)"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+echo 123 > $file
+if [ $? -eq 0 ]
+then
+	echo "$file writing insane input 2 success (should failed)"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+CONTENT=$(cat $file)
+if [ "$CONTENT" != "4242 foo.bar" ]
+then
+	echo "$file not written"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+echo "0 null" > $file
+if [ $? -ne 0 ]
+then
+	echo "$file disabling write fail"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+CONTENT=$(cat $file)
+if [ "$CONTENT" != "0 null" ]
+then
+	echo "$file not disabled"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+echo "4242 foo.bar" > $file
+if [ $? -ne 0 ]
+then
+	echo "$file writing sane data again fail"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+echo $ORIG_CONTENT > $file
+
+# Test pids file
+file="$DBGFS/pids"
+
+ORIG_CONTENT=$(cat $file)
+
+echo "1 2 3 4" > $file
+if [ $? -ne 0 ]
+then
+	echo "$file write fail"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+echo "1 2 abc 4" > $file
+if [ $? -ne 0 ]
+then
+	echo "$file write fail"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+echo abc 2 3 > $file
+if [ $? -eq 0 ]
+then
+	echo "$file write success (should failed)"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+CONTENT=$(cat $file)
+if [ "$CONTENT" != "1 2" ]
+then
+	echo "$file not written"
+	echo $ORIG_CONTENT > $file
+	exit 1
+fi
+
+echo $ORIG_CONTENT > $file
+
+echo "PASS"
diff --git a/tools/testing/selftests/damon/debugfs_record.sh b/tools/testing/selftests/damon/debugfs_record.sh
new file mode 100755
index 000000000000..fa9e07eea258
--- /dev/null
+++ b/tools/testing/selftests/damon/debugfs_record.sh
@@ -0,0 +1,50 @@
+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0
+
+source ./_chk_dependency.sh
+
+restore_attrs()
+{
+	echo $ORIG_ATTRS > $DBGFS/attrs
+	echo $ORIG_PIDS > $DBGFS/pids
+	echo $ORIG_RECORD > $DBGFS/record
+}
+
+ORIG_ATTRS=$(cat $DBGFS/attrs)
+ORIG_PIDS=$(cat $DBGFS/pids)
+ORIG_RECORD=$(cat $DBGFS/record)
+
+rfile=$pwd/damon.data
+
+rm -f $rfile
+ATTRS="5000 100000 1000000 10 1000"
+echo $ATTRS > $DBGFS/attrs
+echo 4096 $rfile > $DBGFS/record
+sleep 5 &
+echo $(pidof sleep) > $DBGFS/pids
+echo on > $DBGFS/monitor_on
+sleep 0.5
+killall sleep
+echo off > $DBGFS/monitor_on
+
+sync
+
+if [ ! -f $rfile ]
+then
+	echo "record file not made"
+	restore_attrs
+
+	exit 1
+fi
+
+python3 ./_chk_record.py $rfile --attrs "$ATTRS"
+if [ $? -ne 0 ]
+then
+	echo "record file is wrong"
+	restore_attrs
+	exit 1
+fi
+
+rm -f $rfile
+restore_attrs
+echo "PASS"
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* [PATCH v7 15/15] MAINTAINERS: Update for DAMON
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (13 preceding siblings ...)
  2020-03-18 11:38 ` [PATCH v7 14/15] mm/damon: Add user space selftests SeongJae Park
@ 2020-03-18 11:39 ` SeongJae Park
  2020-03-30 10:53 ` [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-18 11:39 UTC (permalink / raw)
  To: akpm
  Cc: SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

From: SeongJae Park <sjpark@amazon.de>

This commit updates MAINTAINERS file for DAMON related files.

Signed-off-by: SeongJae Park <sjpark@amazon.de>
---
 MAINTAINERS | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/MAINTAINERS b/MAINTAINERS
index 56765f542244..422c86f64cdd 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -4611,6 +4611,18 @@ F:	net/ax25/ax25_out.c
 F:	net/ax25/ax25_timer.c
 F:	net/ax25/sysctl_net_ax25.c
 
+DATA ACCESS MONITOR
+M:	SeongJae Park <sjpark@amazon.de>
+L:	linux-mm@kvack.org
+S:	Maintained
+F:	Documentation/admin-guide/mm/data_access_monitor.rst
+F:	include/linux/damon.h
+F:	include/trace/events/damon.h
+F:	mm/damon-test.h
+F:	mm/damon.c
+F:	tools/damon/*
+F:	tools/testing/selftests/damon/*
+
 DAVICOM FAST ETHERNET (DMFE) NETWORK DRIVER
 L:	netdev@vger.kernel.org
 S:	Orphan
-- 
2.17.1


^ permalink raw reply related	[flat|nested] 28+ messages in thread

* Re: [PATCH v7 00/15] Introduce Data Access MONitor (DAMON)
  2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
                   ` (14 preceding siblings ...)
  2020-03-18 11:39 ` [PATCH v7 15/15] MAINTAINERS: Update for DAMON SeongJae Park
@ 2020-03-30 10:53 ` SeongJae Park
  15 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-03-30 10:53 UTC (permalink / raw)
  To: SeongJae Park
  Cc: akpm, SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel, shakeelb

On Wed, 18 Mar 2020 12:27:07 +0100 SeongJae Park <sjpark@amazon.com> wrote:

> From: SeongJae Park <sjpark@amazon.de>
> 
> Introduction
> ============
> 
> Memory management decisions can be improved if finer data access information is
> available.  However, because such finer information usually comes with higher
> overhead, most systems including Linux forgives the potential benefit and rely
> on only coarse information or some light-weight heuristics.  The pseudo-LRU and
> the aggressive THP promotions are such examples.
> 
> A number of data access pattern awared memory management optimizations (refer
> to 'Appendix A' for more details) consistently say the potential benefit is not
> small.  However, none of those has successfully merged to the mainline Linux
> kernel mainly due to the absence of a scalable and efficient data access
> monitoring mechanism.  Refer to 'Appendix B' to see the limitations of existing
> memory monitoring mechanisms.
> 
> DAMON is a data access monitoring subsystem for the problem.  It is 1) accurate
> enough to be used for the DRAM level memory management (a straightforward
> DAMON-based optimization achieved up to 2.55x speedup), 2) light-weight enough
> to be applied online (compared to a straightforward access monitoring scheme,
> DAMON is up to 94,242.42x lighter) and 3) keeps predefined upper-bound overhead
> regardless of the size of target workloads (thus scalable).  Refer to 'Appendix
> C' if you interested in how it is possible, and 'Appendix F' to know how the
> numbers collected.
> 
> DAMON has mainly designed for the kernel's memory management mechanisms.
> However, because it is implemented as a standalone kernel module and provides
> several interfaces, it can be used by a wide range of users including kernel
> space programs, user space programs, programmers, and administrators.  DAMON
> is now supporting the monitoring only, but it will also provide simple and
> convenient data access pattern awared memory managements by itself.  Refer to
> 'Appendix D' for more detailed expected usages of DAMON.

There was no review but a few of comments from Shakeel in last week, and
therefore I made no change in this patchset.  Instead, I'm preparing extending
DAMON for physical memory monitoring.

Also, I ran the whole evaluation tests including those for DAMON-based
operation schemes again, because this version (v7) patchset fixed an access
check related bug, thanks to Jonathan's finding, while the attached evaluation
results are measured with the previous version (v6).  Overall, it shows only
subtle changes.

In short, v7 DAMON increases system memory footprint by 0.08%, make the target
workloads 0.25% slower.  The numbers of v6 were -0.08% and 0.76%, respectively.

DAMON-based THP promotion/demotion scheme removes 100% memory overhead
of THP, and even shows 0.11% smaller system memory footprint, compared to THP
disabled case, while preserving 39.67% of THP speedup.  The numbers of v6 were
83.66% and 40.67%, respectively.

DAMON-based proactive reclamation scheme reduced 22.96% of system memory
fooprint and 89.49% of residential sets while incurring only 2.45% runtime
overhead in best case (parsec3/freqmine).  The numbers of v6 were 22.42%,
88.86% and 3.07%, respectively.

The detailed numbers are attached below.  For the detailed numbers of v6, refer
to the CV of v6 DAMON patchset:
https://lore.kernel.org/linux-mm/20200318112722.30143-1-sjpark@amazon.com/

I hope this numbers make more REVIEWS/COMMENTS than my patchsets ;)


Thanks,
SeongJae Park

================================ >8 ===========================================

runtime                 orig     rec      (overhead) thp      (overhead) ethp     (overhead) prcl     (overhead)
parsec3/blackscholes    107.594  107.956  (0.34)     106.750  (-0.78)    107.672  (0.07)     111.916  (4.02)
parsec3/bodytrack       79.230   79.368   (0.17)     78.908   (-0.41)    79.705   (0.60)     80.423   (1.50)
parsec3/canneal         142.831  143.810  (0.69)     123.530  (-13.51)   133.778  (-6.34)    144.998  (1.52)
parsec3/dedup           11.986   11.959   (-0.23)    11.762   (-1.87)    12.028   (0.35)     13.313   (11.07)
parsec3/facesim         210.125  209.007  (-0.53)    205.226  (-2.33)    207.766  (-1.12)    209.815  (-0.15)
parsec3/ferret          191.601  191.177  (-0.22)    190.420  (-0.62)    191.775  (0.09)     192.638  (0.54)
parsec3/fluidanimate    212.735  212.970  (0.11)     209.151  (-1.68)    211.904  (-0.39)    218.573  (2.74)
parsec3/freqmine        291.225  290.873  (-0.12)    289.258  (-0.68)    289.884  (-0.46)    298.373  (2.45)
parsec3/raytrace        118.289  119.586  (1.10)     119.045  (0.64)     119.064  (0.66)     137.919  (16.60)
parsec3/streamcluster   323.565  328.168  (1.42)     279.565  (-13.60)   287.452  (-11.16)   333.244  (2.99)
parsec3/swaptions       155.140  155.473  (0.21)     153.816  (-0.85)    156.423  (0.83)     156.237  (0.71)
parsec3/vips            58.979   59.311   (0.56)     58.733   (-0.42)    59.005   (0.04)     61.062   (3.53)
parsec3/x264            70.539   68.413   (-3.01)    64.760   (-8.19)    67.180   (-4.76)    68.103   (-3.45)
splash2x/barnes         80.414   81.751   (1.66)     73.585   (-8.49)    80.232   (-0.23)    115.753  (43.95)
splash2x/fft            33.902   34.111   (0.62)     24.228   (-28.53)   29.926   (-11.73)   44.438   (31.08)
splash2x/lu_cb          85.556   86.001   (0.52)     84.538   (-1.19)    86.000   (0.52)     91.447   (6.89)
splash2x/lu_ncb         93.399   93.652   (0.27)     90.463   (-3.14)    94.008   (0.65)     93.901   (0.54)
splash2x/ocean_cp       45.253   45.191   (-0.14)    43.049   (-4.87)    44.022   (-2.72)    46.588   (2.95)
splash2x/ocean_ncp      86.927   87.065   (0.16)     50.747   (-41.62)   86.855   (-0.08)    199.553  (129.57)
splash2x/radiosity      91.433   91.511   (0.09)     90.626   (-0.88)    91.865   (0.47)     104.524  (14.32)
splash2x/radix          31.923   32.023   (0.31)     25.194   (-21.08)   32.035   (0.35)     39.231   (22.89)
splash2x/raytrace       84.367   84.677   (0.37)     82.417   (-2.31)    83.505   (-1.02)    84.857   (0.58)
splash2x/volrend        87.499   87.495   (-0.00)    86.775   (-0.83)    87.311   (-0.21)    87.511   (0.01)
splash2x/water_nsquared 236.397  236.759  (0.15)     219.902  (-6.98)    224.228  (-5.15)    238.562  (0.92)
splash2x/water_spatial  89.646   89.767   (0.14)     89.735   (0.10)     90.347   (0.78)     103.585  (15.55)
total                   3020.570 3028.080 (0.25)     2852.190 (-5.57)    2953.960 (-2.21)    3276.550 (8.47)


memused.avg             orig         rec          (overhead) thp          (overhead) ethp         (overhead) prcl         (overhead)
parsec3/blackscholes    1785916.600  1834201.400  (2.70)     1826249.200  (2.26)     1828079.200  (2.36)     1712210.600  (-4.13)
parsec3/bodytrack       1415049.400  1434317.600  (1.36)     1423715.000  (0.61)     1430392.600  (1.08)     1435136.000  (1.42)
parsec3/canneal         1043489.800  1058617.600  (1.45)     1040484.600  (-0.29)    1048664.800  (0.50)     1050280.000  (0.65)
parsec3/dedup           2414453.200  2458493.200  (1.82)     2411379.400  (-0.13)    2400516.000  (-0.58)    2461120.800  (1.93)
parsec3/facesim         541597.200   550097.400   (1.57)     544364.600   (0.51)     553240.000   (2.15)     552316.400   (1.98)
parsec3/ferret          317986.600   332346.000   (4.52)     320218.000   (0.70)     331085.000   (4.12)     330895.200   (4.06)
parsec3/fluidanimate    576183.400   585442.000   (1.61)     577780.200   (0.28)     587703.400   (2.00)     506501.000   (-12.09)
parsec3/freqmine        990869.200   997817.000   (0.70)     990350.400   (-0.05)    997669.000   (0.69)     763325.800   (-22.96)
parsec3/raytrace        1748370.800  1757109.200  (0.50)     1746153.800  (-0.13)    1757830.400  (0.54)     1581455.800  (-9.55)
parsec3/streamcluster   121521.800   140452.400   (15.58)    129725.400   (6.75)     132266.000   (8.84)     130558.200   (7.44)
parsec3/swaptions       15592.400    29018.800    (86.11)    14765.800    (-5.30)    27260.200    (74.83)    26631.600    (70.80)
parsec3/vips            2957567.600  2967993.800  (0.35)     2956623.200  (-0.03)    2973062.600  (0.52)     2951402.000  (-0.21)
parsec3/x264            3169012.400  3175048.800  (0.19)     3190345.400  (0.67)     3189353.000  (0.64)     3172924.200  (0.12)
splash2x/barnes         1209066.000  1213125.400  (0.34)     1217261.400  (0.68)     1209661.600  (0.05)     921041.800   (-23.82)
splash2x/fft            9359313.200  9195213.000  (-1.75)    9377562.400  (0.19)     9050957.600  (-3.29)    9517977.000  (1.70)
splash2x/lu_cb          514966.200   522939.400   (1.55)     520870.400   (1.15)     522635.000   (1.49)     329933.600   (-35.93)
splash2x/lu_ncb         514180.400   525974.800   (2.29)     521420.200   (1.41)     521063.600   (1.34)     523557.000   (1.82)
splash2x/ocean_cp       3346493.400  3288078.000  (-1.75)    3382253.800  (1.07)     3289477.600  (-1.70)    3260810.400  (-2.56)
splash2x/ocean_ncp      3909966.400  3882968.800  (-0.69)    7037196.000  (79.98)    4046363.400  (3.49)     3471452.400  (-11.22)
splash2x/radiosity      1471119.400  1470626.800  (-0.03)    1482604.200  (0.78)     1472718.400  (0.11)     546893.600   (-62.82)
splash2x/radix          1748360.800  1729163.400  (-1.10)    1371463.200  (-21.56)   1701993.600  (-2.65)    1817519.600  (3.96)
splash2x/raytrace       46670.000    60172.200    (28.93)    51901.600    (11.21)    60782.600    (30.24)    52644.800    (12.80)
splash2x/volrend        150666.600   167444.200   (11.14)    151335.200   (0.44)     163345.000   (8.41)     162760.000   (8.03)
splash2x/water_nsquared 45720.200    59422.400    (29.97)    46031.000    (0.68)     61801.400    (35.17)    62627.000    (36.98)
splash2x/water_spatial  663052.200   672855.800   (1.48)     665787.600   (0.41)     674696.200   (1.76)     471052.600   (-28.96)
total                   40077300.000 40108900.000 (0.08)     42997900.000 (7.29)     40032700.000 (-0.11)    37813000.000 (-5.65)


rss.avg                 orig         rec          (overhead) thp          (overhead) ethp         (overhead) prcl         (overhead)
parsec3/blackscholes    592502.000   589764.400   (-0.46)    592132.600   (-0.06)    593702.000   (0.20)     406639.400   (-31.37)
parsec3/bodytrack       32365.400    32195.000    (-0.53)    32210.800    (-0.48)    32114.600    (-0.77)    21537.600    (-33.45)
parsec3/canneal         839904.200   840292.200   (0.05)     836866.400   (-0.36)    838263.200   (-0.20)    837895.800   (-0.24)
parsec3/dedup           1208337.200  1218465.600  (0.84)     1233278.600  (2.06)     1200490.200  (-0.65)    882911.400   (-26.93)
parsec3/facesim         311380.800   311363.600   (-0.01)    315642.600   (1.37)     312573.400   (0.38)     310257.400   (-0.36)
parsec3/ferret          99514.800    99542.000    (0.03)     100454.200   (0.94)     99879.800    (0.37)     89679.200    (-9.88)
parsec3/fluidanimate    531760.800   531735.200   (-0.00)    531865.400   (0.02)     531940.800   (0.03)     440781.000   (-17.11)
parsec3/freqmine        552455.400   552882.600   (0.08)     555793.600   (0.60)     553019.800   (0.10)     58067.000    (-89.49)
parsec3/raytrace        894798.400   894953.400   (0.02)     892223.400   (-0.29)    893012.400   (-0.20)    315259.800   (-64.77)
parsec3/streamcluster   110780.400   110856.800   (0.07)     110954.000   (0.16)     111310.800   (0.48)     108066.800   (-2.45)
parsec3/swaptions       5614.600     5645.600     (0.55)     5553.200     (-1.09)    5552.600     (-1.10)    3251.800     (-42.08)
parsec3/vips            31942.200    31752.800    (-0.59)    32042.600    (0.31)     32226.600    (0.89)     29012.200    (-9.17)
parsec3/x264            81770.800    81609.200    (-0.20)    82800.800    (1.26)     82612.200    (1.03)     81805.800    (0.04)
splash2x/barnes         1216515.600  1217113.800  (0.05)     1225605.600  (0.75)     1217325.000  (0.07)     540108.400   (-55.60)
splash2x/fft            9668660.600  9751350.800  (0.86)     9773806.400  (1.09)     9613555.400  (-0.57)    7951241.800  (-17.76)
splash2x/lu_cb          510368.800   510095.800   (-0.05)    514350.600   (0.78)     510276.000   (-0.02)    311584.800   (-38.95)
splash2x/lu_ncb         509904.800   510001.600   (0.02)     513847.000   (0.77)     510073.400   (0.03)     509905.600   (0.00)
splash2x/ocean_cp       3389550.600  3404466.000  (0.44)     3443363.600  (1.59)     3410388.000  (0.61)     3330608.600  (-1.74)
splash2x/ocean_ncp      3923723.200  3911148.200  (-0.32)    7175800.400  (82.88)    4104482.400  (4.61)     2030525.000  (-48.25)
splash2x/radiosity      1472994.600  1475946.400  (0.20)     1485636.800  (0.86)     1476193.000  (0.22)     262161.400   (-82.20)
splash2x/radix          1750329.800  1765697.000  (0.88)     1413304.000  (-19.25)   1754154.400  (0.22)     1516142.600  (-13.38)
splash2x/raytrace       23149.600    23208.000    (0.25)     28574.400    (23.43)    26694.600    (15.31)    16257.800    (-29.77)
splash2x/volrend        43968.800    43919.000    (-0.11)    44087.600    (0.27)     44224.000    (0.58)     32484.400    (-26.12)
splash2x/water_nsquared 29348.000    29338.400    (-0.03)    29604.600    (0.87)     29779.400    (1.47)     23644.800    (-19.43)
splash2x/water_spatial  655263.600   655097.800   (-0.03)    655199.200   (-0.01)    656282.400   (0.16)     379816.800   (-42.04)
total                   28486900.000 28598400.000 (0.39)     31625000.000 (11.02)    28640100.000 (0.54)     20489600.000 (-28.07)

^ permalink raw reply	[flat|nested] 28+ messages in thread

* Re: [PATCH v7 04/15] mm/damon: Implement region based sampling
  2020-03-18 11:27 ` [PATCH v7 04/15] mm/damon: Implement region based sampling SeongJae Park
@ 2020-03-31 16:02   ` Jonathan Cameron
  2020-04-01  8:22     ` SeongJae Park
  2020-04-02 15:07   ` SeongJae Park
  1 sibling, 1 reply; 28+ messages in thread
From: Jonathan Cameron @ 2020-03-31 16:02 UTC (permalink / raw)
  To: SeongJae Park
  Cc: akpm, SeongJae Park, aarcange, acme, alexander.shishkin, amit,
	brendan.d.gregg, brendanhiggins, cai, colin.king, corbet, dwmw,
	jolsa, kirill, mark.rutland, mgorman, minchan, mingo, namhyung,
	peterz, rdunlap, riel, rientjes, rostedt, shuah, sj38.park,
	vbabka, vdavydov.dev, yang.shi, ying.huang, linux-mm, linux-doc,
	linux-kernel

On Wed, 18 Mar 2020 12:27:11 +0100
SeongJae Park <sjpark@amazon.com> wrote:

> From: SeongJae Park <sjpark@amazon.de>
> 
> This commit implements DAMON's basic access check and region based
> sampling mechanisms.  This change would seems make no sense, mainly
> because it is only a part of the DAMON's logics.  Following two commits
> will make more sense.
> 
> Basic Access Check
> ------------------
> 
> DAMON basically reports what pages are how frequently accessed.  Note
> that the frequency is not an absolute number of accesses, but a relative
> frequency among the pages of the target workloads.
> 
> Users can control the resolution of the reports by setting two time
> intervals, ``sampling interval`` and ``aggregation interval``.  In
> detail, DAMON checks access to each page per ``sampling interval``,
> aggregates the results (counts the number of the accesses to each page),
> and reports the aggregated results per ``aggregation interval``.  For
> the access check of each page, DAMON uses the Accessed bits of PTEs.
> 
> This is thus similar to common periodic access checks based access
> tracking mechanisms, which overhead is increasing as the size of the
> target process grows.
> 
> Region Based Sampling
> ---------------------
> 
> To avoid the unbounded increase of the overhead, DAMON groups a number
> of adjacent pages that assumed to have same access frequencies into a
> region.  As long as the assumption (pages in a region have same access
> frequencies) is kept, only one page in the region is required to be
> checked.  Thus, for each ``sampling interval``, DAMON randomly picks one
> page in each region and clears its Accessed bit.  After one more
> ``sampling interval``, DAMON reads the Accessed bit of the page and
> increases the access frequency of the region if the bit has set
> meanwhile.  Therefore, the monitoring overhead is controllable by
> setting the number of regions.
> 
> Nonetheless, this scheme cannot preserve the quality of the output if
> the assumption is not kept.  Following commit will introduce how we can
> make the guarantee with best effort.
> 
> Signed-off-by: SeongJae Park <sjpark@amazon.de>

Hi.

A few comments inline.

I've still not replicated your benchmarks so may well have some more
feedback once I've managed that on one of our servers.

Thanks,

Jonathan

> ---
>  include/linux/damon.h |  24 ++
>  mm/damon.c            | 553 ++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 577 insertions(+)
> 
> diff --git a/include/linux/damon.h b/include/linux/damon.h
> index 7117bb7e7544..f1945df6e6b4 100644
> --- a/include/linux/damon.h
> +++ b/include/linux/damon.h
> @@ -11,6 +11,8 @@
>  #define _DAMON_H_
>  
>  #include <linux/random.h>
> +#include <linux/mutex.h>
> +#include <linux/time64.h>
>  #include <linux/types.h>
>  
>  /* Represents a monitoring target region on the virtual address space */
> @@ -29,10 +31,32 @@ struct damon_task {
>  	struct list_head list;
>  };
>  
> +/*
> + * For each 'sample_interval', DAMON checks whether each region is accessed or
> + * not.  It aggregates and keeps the access information (number of accesses to
> + * each region) for each 'aggr_interval' time.
> + *
> + * All time intervals are in micro-seconds.
> + */
>  struct damon_ctx {
> +	unsigned long sample_interval;
> +	unsigned long aggr_interval;
> +	unsigned long min_nr_regions;
> +
> +	struct timespec64 last_aggregation;
> +
> +	struct task_struct *kdamond;
> +	struct mutex kdamond_lock;
> +
>  	struct rnd_state rndseed;
>  
>  	struct list_head tasks_list;	/* 'damon_task' objects */
>  };
>  
> +int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids);
> +int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> +		unsigned long aggr_int, unsigned long min_nr_reg);
> +int damon_start(struct damon_ctx *ctx);
> +int damon_stop(struct damon_ctx *ctx);
> +
>  #endif
> diff --git a/mm/damon.c b/mm/damon.c
> index d7e6226ab7f1..018016793555 100644
> --- a/mm/damon.c
> +++ b/mm/damon.c
> @@ -10,8 +10,14 @@
>  #define pr_fmt(fmt) "damon: " fmt
>  
>  #include <linux/damon.h>
> +#include <linux/delay.h>
> +#include <linux/kthread.h>
>  #include <linux/mm.h>
>  #include <linux/module.h>
> +#include <linux/page_idle.h>
> +#include <linux/random.h>
> +#include <linux/sched/mm.h>
> +#include <linux/sched/task.h>
>  #include <linux/slab.h>
>  
>  #define damon_get_task_struct(t) \
> @@ -171,6 +177,553 @@ static unsigned int nr_damon_regions(struct damon_task *t)
>  	return nr_regions;
>  }
>  
> +/*
> + * Get the mm_struct of the given task
> + *
> + * Caller should put the mm_struct after use, unless it is NULL.
> + *
> + * Returns the mm_struct of the task on success, NULL on failure
> + */
> +static struct mm_struct *damon_get_mm(struct damon_task *t)
> +{
> +	struct task_struct *task;
> +	struct mm_struct *mm;
> +
> +	task = damon_get_task_struct(t);
> +	if (!task)
> +		return NULL;
> +
> +	mm = get_task_mm(task);
> +	put_task_struct(task);
> +	return mm;
> +}
> +
> +/*
> + * Size-evenly split a region into 'nr_pieces' small regions
> + *
> + * Returns 0 on success, or negative error code otherwise.
> + */
> +static int damon_split_region_evenly(struct damon_ctx *ctx,
> +		struct damon_region *r, unsigned int nr_pieces)
> +{
> +	unsigned long sz_orig, sz_piece, orig_end;
> +	struct damon_region *piece = NULL, *next;
> +	unsigned long start;
> +
> +	if (!r || !nr_pieces)
> +		return -EINVAL;
> +
> +	orig_end = r->vm_end;
> +	sz_orig = r->vm_end - r->vm_start;
> +	sz_piece = sz_orig / nr_pieces;
> +
> +	if (!sz_piece)
> +		return -EINVAL;
> +
> +	r->vm_end = r->vm_start + sz_piece;
> +	next = damon_next_region(r);
> +	for (start = r->vm_end; start + sz_piece <= orig_end;
> +			start += sz_piece) {
> +		piece = damon_new_region(ctx, start, start + sz_piece);
piece may be n
> +		damon_insert_region(piece, r, next);
> +		r = piece;
> +	}
> +	/* complement last region for possible rounding error */
> +	if (piece)
> +		piece->vm_end = orig_end;

Update the sampling address to ensure it's in the region?

> +
> +	return 0;
> +}
> +
> +struct region {
> +	unsigned long start;
> +	unsigned long end;
> +};
> +
> +static unsigned long sz_region(struct region *r)
> +{
> +	return r->end - r->start;
> +}
> +
> +static void swap_regions(struct region *r1, struct region *r2)
> +{
> +	struct region tmp;
> +
> +	tmp = *r1;
> +	*r1 = *r2;
> +	*r2 = tmp;
> +}
> +
> +/*
> + * Find the three regions in an address space
> + *
> + * vma		the head vma of the target address space
> + * regions	an array of three 'struct region's that results will be saved
> + *
> + * This function receives an address space and finds three regions in it which
> + * separated by the two biggest unmapped regions in the space.  Please refer to
> + * below comments of 'damon_init_regions_of()' function to know why this is
> + * necessary.
> + *
> + * Returns 0 if success, or negative error code otherwise.
> + */
> +static int damon_three_regions_in_vmas(struct vm_area_struct *vma,
> +		struct region regions[3])
> +{
> +	struct region gap = {0,}, first_gap = {0,}, second_gap = {0,};
> +	struct vm_area_struct *last_vma = NULL;
> +	unsigned long start = 0;
> +
> +	/* Find two biggest gaps so that first_gap > second_gap > others */
> +	for (; vma; vma = vma->vm_next) {
> +		if (!last_vma) {
> +			start = vma->vm_start;
> +			last_vma = vma;
> +			continue;
> +		}
> +		gap.start = last_vma->vm_end;
> +		gap.end = vma->vm_start;
> +		if (sz_region(&gap) > sz_region(&second_gap)) {
> +			swap_regions(&gap, &second_gap);
> +			if (sz_region(&second_gap) > sz_region(&first_gap))
> +				swap_regions(&second_gap, &first_gap);
> +		}
> +		last_vma = vma;
> +	}
> +
> +	if (!sz_region(&second_gap) || !sz_region(&first_gap))
> +		return -EINVAL;
> +
> +	/* Sort the two biggest gaps by address */
> +	if (first_gap.start > second_gap.start)
> +		swap_regions(&first_gap, &second_gap);
> +
> +	/* Store the result */
> +	regions[0].start = start;
> +	regions[0].end = first_gap.start;
> +	regions[1].start = first_gap.end;
> +	regions[1].end = second_gap.start;
> +	regions[2].start = second_gap.end;
> +	regions[2].end = last_vma->vm_end;
> +
> +	return 0;
> +}
> +
> +/*
> + * Get the three regions in the given task
> + *
> + * Returns 0 on success, negative error code otherwise.
> + */
> +static int damon_three_regions_of(struct damon_task *t,
> +				struct region regions[3])
> +{
> +	struct mm_struct *mm;
> +	int rc;
> +
> +	mm = damon_get_mm(t);
> +	if (!mm)
> +		return -EINVAL;
> +
> +	down_read(&mm->mmap_sem);
> +	rc = damon_three_regions_in_vmas(mm->mmap, regions);
> +	up_read(&mm->mmap_sem);
> +
> +	mmput(mm);
> +	return rc;
> +}
> +
> +/*
> + * Initialize the monitoring target regions for the given task
> + *
> + * t	the given target task
> + *
> + * Because only a number of small portions of the entire address space
> + * is acutally mapped to the memory and accessed, monitoring the unmapped
> + * regions is wasteful.  That said, because we can deal with small noises,
> + * tracking every mapping is not strictly required but could even incur a high
> + * overhead if the mapping frequently changes or the number of mappings is
> + * high.  Nonetheless, this may seems very weird.  DAMON's dynamic regions
> + * adjustment mechanism, which will be implemented with following commit will
> + * make this more sense.
> + *
> + * For the reason, we convert the complex mappings to three distinct regions
> + * that cover every mapped areas of the address space.  Also the two gaps
> + * between the three regions are the two biggest unmapped areas in the given
> + * address space.  In detail, this function first identifies the start and the
> + * end of the mappings and the two biggest unmapped areas of the address space.
> + * Then, it constructs the three regions as below:
> + *
> + *     [mappings[0]->start, big_two_unmapped_areas[0]->start)
> + *     [big_two_unmapped_areas[0]->end, big_two_unmapped_areas[1]->start)
> + *     [big_two_unmapped_areas[1]->end, mappings[nr_mappings - 1]->end)
> + *
> + * As usual memory map of processes is as below, the gap between the heap and
> + * the uppermost mmap()-ed region, and the gap between the lowermost mmap()-ed
> + * region and the stack will be two biggest unmapped regions.  Because these
> + * gaps are exceptionally huge areas in usual address space, excluding these
> + * two biggest unmapped regions will be sufficient to make a trade-off.
> + *
> + *   <heap>
> + *   <BIG UNMAPPED REGION 1>
> + *   <uppermost mmap()-ed region>
> + *   (other mmap()-ed regions and small unmapped regions)
> + *   <lowermost mmap()-ed region>
> + *   <BIG UNMAPPED REGION 2>
> + *   <stack>
> + */
> +static void damon_init_regions_of(struct damon_ctx *c, struct damon_task *t)
> +{
> +	struct damon_region *r;
> +	struct region regions[3];
> +	int i;
> +
> +	if (damon_three_regions_of(t, regions)) {
> +		pr_err("Failed to get three regions of task %lu\n", t->pid);
> +		return;
> +	}
> +
> +	/* Set the initial three regions of the task */
> +	for (i = 0; i < 3; i++) {
> +		r = damon_new_region(c, regions[i].start, regions[i].end);
> +		damon_add_region(r, t);
> +	}
> +
> +	/* Split the middle region into 'min_nr_regions - 2' regions */
> +	r = damon_nth_region_of(t, 1);
> +	if (damon_split_region_evenly(c, r, c->min_nr_regions - 2))
> +		pr_warn("Init middle region failed to be split\n");
> +}
> +
> +/* Initialize '->regions_list' of every task */
> +static void kdamond_init_regions(struct damon_ctx *ctx)
> +{
> +	struct damon_task *t;
> +
> +	damon_for_each_task(ctx, t)
> +		damon_init_regions_of(ctx, t);
> +}
> +
> +static bool damon_pte_pmd_young(pte_t *pte, pmd_t *pmd)
> +{
> +	if (pte && pte_young(*pte))
> +		return true;
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +	if (pmd && pmd_young(*pmd))
> +		return true;
> +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
> +	return false;
> +}
> +
> +static void damon_pte_pmd_mkold(pte_t *pte, pmd_t *pmd)
> +{
> +	if (pte) {
> +		if (pte_young(*pte)) {
> +			clear_page_idle(pte_page(*pte));
> +			set_page_young(pte_page(*pte));
> +		}
> +		*pte = pte_mkold(*pte);
> +		return;
> +	}
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +	if (pmd) {
> +		if (pmd_young(*pmd)) {
> +			clear_page_idle(pmd_page(*pmd));
> +			set_page_young(pmd_page(*pmd));
> +		}
> +		*pmd = pmd_mkold(*pmd);
> +	}
> +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */

No need to flush the TLBs?

> +}
> +
> +/*
> + * Check whether the region accessed and prepare for next check
> + *
> + * mm	'mm_struct' for the given virtual address space
> + * r	the region to be checked
> + */
> +static void kdamond_check_access(struct damon_ctx *ctx,
> +			struct mm_struct *mm, struct damon_region *r)
> +{
> +	static struct mm_struct *last_mm;
> +	static unsigned long last_addr;
> +	static int last_page_sz = PAGE_SIZE;
> +	static bool last_accessed;
> +
> +	pte_t *pte = NULL;
> +	pmd_t *pmd = NULL;
> +	spinlock_t *ptl;
> +
> +	/* If the region is in the last checked page, reuse the result */
> +	if (mm == last_mm && (ALIGN_DOWN(last_addr, last_page_sz) ==
> +				ALIGN_DOWN(r->sampling_addr, last_page_sz))) {
> +		if (last_accessed)
> +			r->nr_accesses++;
> +		return;
> +	}
> +
> +	if (follow_pte_pmd(mm, r->sampling_addr, NULL, &pte, &pmd, &ptl))
> +		goto prepare_next_check;
> +
> +	/* Read the page table access bit of the page */
> +	if (damon_pte_pmd_young(pte, pmd)) {
> +		last_accessed = true;
> +		r->nr_accesses++;
> +	}
> +	spin_unlock(ptl);
> +
> +prepare_next_check:
> +	last_mm = mm;
> +	last_addr = r->sampling_addr;
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +	last_page_sz = pte ? PAGE_SIZE : ((1UL) << HPAGE_PMD_SHIFT);
> +#endif
> +
> +	r->sampling_addr = damon_rand(ctx, r->vm_start, r->vm_end);
> +	pte = NULL, pmd = NULL;
> +	if (follow_pte_pmd(mm, r->sampling_addr, NULL, &pte, &pmd, &ptl))
> +		return;
> +
> +	damon_pte_pmd_mkold(pte, pmd);
> +	spin_unlock(ptl);
> +}
> +
> +/*
> + * damon_check_reset_time_interval() - Check if a time interval is elapsed.
> + * @baseline:	the time to check whether the interval has elapsed since
> + * @interval:	the time interval (microseconds)
> + *
> + * See whether the given time interval has passed since the given baseline
> + * time.  If so, it also updates the baseline to current time for next check.
> + *
> + * Return:	true if the time interval has passed, or false otherwise.
> + */
> +static bool damon_check_reset_time_interval(struct timespec64 *baseline,
> +		unsigned long interval)
> +{
> +	struct timespec64 now;
> +
> +	ktime_get_coarse_ts64(&now);
> +	if ((timespec64_to_ns(&now) - timespec64_to_ns(baseline)) <
> +			interval * 1000)
> +		return false;
> +	*baseline = now;
> +	return true;
> +}
> +
> +/*
> + * Check whether it is time to flush the aggregated information
> + */
> +static bool kdamond_aggregate_interval_passed(struct damon_ctx *ctx)
> +{
> +	return damon_check_reset_time_interval(&ctx->last_aggregation,
> +			ctx->aggr_interval);
> +}
> +
> +/*
> + * Reset the aggregated monitoring results
> + */
> +static void kdamond_reset_aggregated(struct damon_ctx *c)
> +{
> +	struct damon_task *t;
> +	struct damon_region *r;
> +
> +	damon_for_each_task(c, t) {
> +		damon_for_each_region(r, t)
> +			r->nr_accesses = 0;
> +	}
> +}
> +
> +/*
> + * Check whether current monitoring should be stopped
> + *
> + * If users asked to stop, need stop.  Even though no user has asked to stop,
> + * need stop if every target task has dead.
> + *
> + * Returns true if need to stop current monitoring.
> + */
> +static bool kdamond_need_stop(struct damon_ctx *ctx)
> +{
> +	struct damon_task *t;
> +	struct task_struct *task;
> +	bool stop;
> +
> +	stop = kthread_should_stop();
> +	if (stop)
> +		return true;
> +
> +	damon_for_each_task(ctx, t) {
> +		task = damon_get_task_struct(t);
> +		if (task) {
> +			put_task_struct(task);
> +			return false;
> +		}
> +	}
> +
> +	return true;
> +}
> +
> +/*
> + * The monitoring daemon that runs as a kernel thread
> + */
> +static int kdamond_fn(void *data)
> +{
> +	struct damon_ctx *ctx = data;
> +	struct damon_task *t;
> +	struct damon_region *r, *next;
> +	struct mm_struct *mm;
> +
> +	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
> +	kdamond_init_regions(ctx);

We haven't called mkold on the initial regions so first check will
get us fairly random state.

> +	while (!kdamond_need_stop(ctx)) {
> +		damon_for_each_task(ctx, t) {
> +			mm = damon_get_mm(t);
> +			if (!mm)
> +				continue;
> +			damon_for_each_region(r, t)
> +				kdamond_check_access(ctx, mm, r);
> +			mmput(mm);
> +		}
> +
> +		if (kdamond_aggregate_interval_passed(ctx))
> +			kdamond_reset_aggregated(ctx);
> +
> +		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
> +	}
> +	damon_for_each_task(ctx, t) {
> +		damon_for_each_region_safe(r, next, t)
> +			damon_destroy_region(r);
> +	}
> +	pr_debug("kdamond (%d) finishes\n", ctx->kdamond->pid);
> +	mutex_lock(&ctx->kdamond_lock);
> +	ctx->kdamond = NULL;
> +	mutex_unlock(&ctx->kdamond_lock);
> +
> +	return 0;
> +}
> +
> +/*
> + * Controller functions
> + */
> +
> +static bool damon_kdamond_running(struct damon_ctx *ctx)
> +{
> +	bool running;
> +
> +	mutex_lock(&ctx->kdamond_lock);
> +	running = ctx->kdamond != NULL;
> +	mutex_unlock(&ctx->kdamond_lock);
> +
> +	return running;
> +}
> +
> +/*
> + * Start or stop the kdamond
> + *
> + * Returns 0 if success, negative error code otherwise.
> + */
> +static int damon_turn_kdamond(struct damon_ctx *ctx, bool on)
> +{
> +	int err = -EBUSY;
> +
> +	mutex_lock(&ctx->kdamond_lock);
> +	if (!ctx->kdamond && on) {

Given there is very little shared code between on and off, I would
suggest just splitting it into two functions.

> +		err = 0;
> +		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond");
> +		if (IS_ERR(ctx->kdamond))
> +			err = PTR_ERR(ctx->kdamond);
> +	} else if (ctx->kdamond && !on) {
> +		mutex_unlock(&ctx->kdamond_lock);
> +		kthread_stop(ctx->kdamond);
> +		while (damon_kdamond_running(ctx))
> +			usleep_range(ctx->sample_interval,
> +					ctx->sample_interval * 2);
> +		return 0;
> +	}
> +	mutex_unlock(&ctx->kdamond_lock);
> +
> +	return err;
> +}
> +
> +/*
> + * damon_start() - Starts monitoring with given context.
> + * @ctx:	monitoring context
> + *
> + * Return: 0 on success, negative error code otherwise.
> + */
> +int damon_start(struct damon_ctx *ctx)
> +{
> +	return damon_turn_kdamond(ctx, true);
> +}
> +
> +/*
> + * damon_stop() - Stops monitoring of given context.
> + * @ctx:	monitoring context
> + *
> + * Return: 0 on success, negative error code otherwise.
> + */
> +int damon_stop(struct damon_ctx *ctx)
> +{
> +	return damon_turn_kdamond(ctx, false);
> +}
> +
> +/*
> + * damon_set_pids() - Set monitoring target processes.
> + * @ctx:	monitoring context
> + * @pids:	array of target processes pids
> + * @nr_pids:	number of entries in @pids
> + *
> + * This function should not be called while the kdamond is running.
> + *
> + * Return: 0 on usccess, negative error code otherwise.
> + */
> +int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids)
> +{
> +	ssize_t i;
> +	struct damon_task *t, *next;
> +
> +	damon_for_each_task_safe(ctx, t, next)
> +		damon_destroy_task(t);
> +
> +	for (i = 0; i < nr_pids; i++) {
> +		t = damon_new_task(pids[i]);
> +		if (!t) {
> +			pr_err("Failed to alloc damon_task\n");
> +			return -ENOMEM;
> +		}
> +		damon_add_task(ctx, t);
> +	}
> +
> +	return 0;
> +}
> +
> +/*

Why not make these actual kernel-doc?  That way you can use the
kernel-doc scripts to sanity check them.

/**

> + * damon_set_attrs() - Set attributes for the monitoring.
> + * @ctx:		monitoring context
> + * @sample_int:		time interval between samplings
> + * @aggr_int:		time interval between aggregations
> + * @min_nr_reg:		minimal number of regions
> + *
> + * This function should not be called while the kdamond is running.
> + * Every time interval is in micro-seconds.
> + *
> + * Return: 0 on success, negative error code otherwise.
> + */
> +int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> +		unsigned long aggr_int, unsigned long min_nr_reg)
> +{
> +	if (min_nr_reg < 3) {
> +		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
> +				min_nr_reg);
> +		return -EINVAL;
> +	}
> +
> +	ctx->sample_interval = sample_int;
> +	ctx->aggr_interval = aggr_int;
> +	ctx->min_nr_regions = min_nr_reg;
> +
> +	return 0;
> +}
> +
>  static int __init damon_init(void)
>  {
>  	return 0;



^ permalink raw reply	[flat|nested] 28+ messages in thread

* Re: [PATCH v7 05/15] mm/damon: Adaptively adjust regions
  2020-03-18 11:27 ` [PATCH v7 05/15] mm/damon: Adaptively adjust regions SeongJae Park
@ 2020-03-31 16:08   ` Jonathan Cameron
  2020-04-01  8:22     ` SeongJae Park
  0 siblings, 1 reply; 28+ messages in thread
From: Jonathan Cameron @ 2020-03-31 16:08 UTC (permalink / raw)
  To: SeongJae Park
  Cc: akpm, SeongJae Park, aarcange, acme, alexander.shishkin, amit,
	brendan.d.gregg, brendanhiggins, cai, colin.king, corbet, dwmw,
	jolsa, kirill, mark.rutland, mgorman, minchan, mingo, namhyung,
	peterz, rdunlap, riel, rientjes, rostedt, shuah, sj38.park,
	vbabka, vdavydov.dev, yang.shi, ying.huang, linux-mm, linux-doc,
	linux-kernel

On Wed, 18 Mar 2020 12:27:12 +0100
SeongJae Park <sjpark@amazon.com> wrote:

> From: SeongJae Park <sjpark@amazon.de>
> 
> At the beginning of the monitoring, DAMON constructs the initial regions
> by evenly splitting the memory mapped address space of the process into
> the user-specified minimal number of regions.  In this initial state,
> the assumption of the regions (pages in same region have similar access
> frequencies) is normally not kept and thus the monitoring quality could
> be low.  To keep the assumption as much as possible, DAMON adaptively
> merges and splits each region.
> 
> For each ``aggregation interval``, it compares the access frequencies of
> adjacent regions and merges those if the frequency difference is small.
> Then, after it reports and clears the aggregated access frequency of
> each region, it splits each region into two regions if the total number
> of regions is smaller than the half of the user-specified maximum number
> of regions.
> 
> In this way, DAMON provides its best-effort quality and minimal overhead
> while keeping the bounds users set for their trade-off.
> 
> Signed-off-by: SeongJae Park <sjpark@amazon.de>

A few more edge cases in here, and a suggestion that might be more costly
but lead to simpler code.

Jonathan

> ---
>  include/linux/damon.h |   6 +-
>  mm/damon.c            | 148 ++++++++++++++++++++++++++++++++++++++++--
>  2 files changed, 145 insertions(+), 9 deletions(-)
> 
> diff --git a/include/linux/damon.h b/include/linux/damon.h
> index f1945df6e6b4..7562b85b1ec0 100644
> --- a/include/linux/damon.h
> +++ b/include/linux/damon.h
> @@ -42,6 +42,7 @@ struct damon_ctx {
>  	unsigned long sample_interval;
>  	unsigned long aggr_interval;
>  	unsigned long min_nr_regions;
> +	unsigned long max_nr_regions;
>  
>  	struct timespec64 last_aggregation;
>  
> @@ -54,8 +55,9 @@ struct damon_ctx {
>  };
>  
>  int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids);
> -int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> -		unsigned long aggr_int, unsigned long min_nr_reg);
> +int damon_set_attrs(struct damon_ctx *ctx,
> +		unsigned long sample_int, unsigned long aggr_int,
> +		unsigned long min_nr_reg, unsigned long max_nr_reg);
>  int damon_start(struct damon_ctx *ctx);
>  int damon_stop(struct damon_ctx *ctx);
>  
> diff --git a/mm/damon.c b/mm/damon.c
> index 018016793555..23c0de3b502e 100644
> --- a/mm/damon.c
> +++ b/mm/damon.c
> @@ -342,9 +342,12 @@ static int damon_three_regions_of(struct damon_task *t,
>   * regions is wasteful.  That said, because we can deal with small noises,
>   * tracking every mapping is not strictly required but could even incur a high
>   * overhead if the mapping frequently changes or the number of mappings is
> - * high.  Nonetheless, this may seems very weird.  DAMON's dynamic regions
> - * adjustment mechanism, which will be implemented with following commit will
> - * make this more sense.
> + * high.  The adaptive regions adjustment mechanism will further help to deal
> + * with the noises by simply identifying the unmapped areas as a region that
> + * has no access.  Moreover, applying the real mappings that would have many
> + * unmapped areas inside will make the adaptive mechanism quite complex.  That
> + * said, too huge unmapped areas inside the monitoring target should be removed
> + * to not take the time for the adaptive mechanism.
>   *
>   * For the reason, we convert the complex mappings to three distinct regions
>   * that cover every mapped areas of the address space.  Also the two gaps
> @@ -533,6 +536,121 @@ static void kdamond_reset_aggregated(struct damon_ctx *c)
>  	}
>  }
>  
> +#define sz_damon_region(r) (r->vm_end - r->vm_start)
> +
> +/*
> + * Merge two adjacent regions into one region
> + */
> +static void damon_merge_two_regions(struct damon_region *l,
> +				struct damon_region *r)
> +{
> +	l->nr_accesses = (l->nr_accesses * sz_damon_region(l) +
> +			r->nr_accesses * sz_damon_region(r)) /
> +			(sz_damon_region(l) + sz_damon_region(r));
> +	l->vm_end = r->vm_end;
> +	damon_destroy_region(r);
> +}
> +
> +#define diff_of(a, b) (a > b ? a - b : b - a)
> +
> +/*
> + * Merge adjacent regions having similar access frequencies
> + *
> + * t		task that merge operation will make change
> + * thres	merge regions having '->nr_accesses' diff smaller than this
> + */
> +static void damon_merge_regions_of(struct damon_task *t, unsigned int thres)
> +{
> +	struct damon_region *r, *prev = NULL, *next;
> +
> +	damon_for_each_region_safe(r, next, t) {
> +		if (!prev || prev->vm_end != r->vm_start ||
> +		    diff_of(prev->nr_accesses, r->nr_accesses) > thres) {
> +			prev = r;
> +			continue;
> +		}
> +		damon_merge_two_regions(prev, r);
> +	}
> +}
> +
> +/*
> + * Merge adjacent regions having similar access frequencies
> + *
> + * threshold	merge regions havind nr_accesses diff larger than this
> + *
> + * This function merges monitoring target regions which are adjacent and their
> + * access frequencies are similar.  This is for minimizing the monitoring
> + * overhead under the dynamically changeable access pattern.  If a merge was
> + * unnecessarily made, later 'kdamond_split_regions()' will revert it.
> + */
> +static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold)
> +{
> +	struct damon_task *t;
> +
> +	damon_for_each_task(c, t)
> +		damon_merge_regions_of(t, threshold);
> +}
> +
> +/*
> + * Split a region into two small regions
> + *
> + * r		the region to be split
> + * sz_r		size of the first sub-region that will be made
> + */
> +static void damon_split_region_at(struct damon_ctx *ctx,
> +		struct damon_region *r, unsigned long sz_r)
> +{
> +	struct damon_region *new;
> +
> +	new = damon_new_region(ctx, r->vm_start + sz_r, r->vm_end);
> +	r->vm_end = new->vm_start;

We may well have a sampling address that is in the wrong region.
It should have little effect on the stats as will fix on next sample
but in my view still worth cleaning up.

> +
> +	damon_insert_region(new, r, damon_next_region(r));
> +}
> +
> +static void damon_split_regions_of(struct damon_ctx *ctx, struct damon_task *t)
> +{
> +	struct damon_region *r, *next;
> +	unsigned long sz_left_region;
> +
> +	damon_for_each_region_safe(r, next, t) {
> +		/*
> +		 * Randomly select size of left sub-region to be at least
> +		 * 10 percent and at most 90% of original region
> +		 */
> +		sz_left_region = (prandom_u32_state(&ctx->rndseed) % 9 + 1) *
> +			(r->vm_end - r->vm_start) / 10;
> +		/* Do not allow blank region */
> +		if (sz_left_region == 0)
> +			continue;
> +		damon_split_region_at(ctx, r, sz_left_region);
> +	}
> +}
> +
> +/*
> + * splits every target regions into two randomly-sized regions
> + *
> + * This function splits every target regions into two random-sized regions if
> + * current total number of the regions is smaller than the half of the
> + * user-specified maximum number of regions.  This is for maximizing the
> + * monitoring accuracy under the dynamically changeable access patterns.  If a
> + * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
> + * it.
> + */
> +static void kdamond_split_regions(struct damon_ctx *ctx)
> +{
> +	struct damon_task *t;
> +	unsigned int nr_regions = 0;
> +
> +	damon_for_each_task(ctx, t)
> +		nr_regions += nr_damon_regions(t);
> +	if (nr_regions > ctx->max_nr_regions / 2)
> +		return;
> +
> +	damon_for_each_task(ctx, t)
> +		damon_split_regions_of(ctx, t);
> +}
> +
>  /*
>   * Check whether current monitoring should be stopped
>   *
> @@ -571,21 +689,29 @@ static int kdamond_fn(void *data)
>  	struct damon_task *t;
>  	struct damon_region *r, *next;
>  	struct mm_struct *mm;
> +	unsigned int max_nr_accesses;
>  
>  	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
>  	kdamond_init_regions(ctx);
>  	while (!kdamond_need_stop(ctx)) {
> +		max_nr_accesses = 0;
>  		damon_for_each_task(ctx, t) {
>  			mm = damon_get_mm(t);
>  			if (!mm)
>  				continue;
> -			damon_for_each_region(r, t)
> +			damon_for_each_region(r, t) {
>  				kdamond_check_access(ctx, mm, r);
> +				max_nr_accesses = max(r->nr_accesses,
> +						max_nr_accesses);
> +			}
>  			mmput(mm);
>  		}
>  
> -		if (kdamond_aggregate_interval_passed(ctx))
> +		if (kdamond_aggregate_interval_passed(ctx)) {
> +			kdamond_merge_regions(ctx, max_nr_accesses / 10);
>  			kdamond_reset_aggregated(ctx);
> +			kdamond_split_regions(ctx);
> +		}

I wonder if it would be simpler to split the sampling address setup and
mkold from the access check.  We would have to walk regions twice,
but not have to bother separately dealing with updating some regions
if they are modified in the above block.

Also, the above has some overhead, so will bias that first sample each
time the block above runs.  If we do the mkold afterwards it will make
much less difference.

>  
>  		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
>  	}
> @@ -702,24 +828,32 @@ int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids)
>   * @sample_int:		time interval between samplings
>   * @aggr_int:		time interval between aggregations
>   * @min_nr_reg:		minimal number of regions
> + * @max_nr_reg:		maximum number of regions
>   *
>   * This function should not be called while the kdamond is running.
>   * Every time interval is in micro-seconds.
>   *
>   * Return: 0 on success, negative error code otherwise.
>   */
> -int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> -		unsigned long aggr_int, unsigned long min_nr_reg)
> +int damon_set_attrs(struct damon_ctx *ctx,
> +			unsigned long sample_int, unsigned long aggr_int,
> +			unsigned long min_nr_reg, unsigned long max_nr_reg)
>  {
>  	if (min_nr_reg < 3) {
>  		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
>  				min_nr_reg);
>  		return -EINVAL;
>  	}
> +	if (min_nr_reg >= ctx->max_nr_regions) {
> +		pr_err("invalid nr_regions.  min (%lu) >= max (%lu)\n",
> +				min_nr_reg, max_nr_reg);
> +		return -EINVAL;
> +	}
>  
>  	ctx->sample_interval = sample_int;
>  	ctx->aggr_interval = aggr_int;
>  	ctx->min_nr_regions = min_nr_reg;
> +	ctx->max_nr_regions = max_nr_reg;
>  
>  	return 0;
>  }



^ permalink raw reply	[flat|nested] 28+ messages in thread

* Re: Re: [PATCH v7 04/15] mm/damon: Implement region based sampling
  2020-03-31 16:02   ` Jonathan Cameron
@ 2020-04-01  8:22     ` SeongJae Park
  2020-04-01 14:24       ` Jonathan Cameron
  2020-04-02 13:59       ` SeongJae Park
  0 siblings, 2 replies; 28+ messages in thread
From: SeongJae Park @ 2020-04-01  8:22 UTC (permalink / raw)
  To: Jonathan Cameron
  Cc: SeongJae Park, akpm, SeongJae Park, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

On Tue, 31 Mar 2020 17:02:33 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:

> On Wed, 18 Mar 2020 12:27:11 +0100
> SeongJae Park <sjpark@amazon.com> wrote:
> 
> > From: SeongJae Park <sjpark@amazon.de>
> > 
> > This commit implements DAMON's basic access check and region based
> > sampling mechanisms.  This change would seems make no sense, mainly
> > because it is only a part of the DAMON's logics.  Following two commits
> > will make more sense.
> > 
> > Basic Access Check
> > ------------------
> > 
> > DAMON basically reports what pages are how frequently accessed.  Note
> > that the frequency is not an absolute number of accesses, but a relative
> > frequency among the pages of the target workloads.
> > 
> > Users can control the resolution of the reports by setting two time
> > intervals, ``sampling interval`` and ``aggregation interval``.  In
> > detail, DAMON checks access to each page per ``sampling interval``,
> > aggregates the results (counts the number of the accesses to each page),
> > and reports the aggregated results per ``aggregation interval``.  For
> > the access check of each page, DAMON uses the Accessed bits of PTEs.
> > 
> > This is thus similar to common periodic access checks based access
> > tracking mechanisms, which overhead is increasing as the size of the
> > target process grows.
> > 
> > Region Based Sampling
> > ---------------------
> > 
> > To avoid the unbounded increase of the overhead, DAMON groups a number
> > of adjacent pages that assumed to have same access frequencies into a
> > region.  As long as the assumption (pages in a region have same access
> > frequencies) is kept, only one page in the region is required to be
> > checked.  Thus, for each ``sampling interval``, DAMON randomly picks one
> > page in each region and clears its Accessed bit.  After one more
> > ``sampling interval``, DAMON reads the Accessed bit of the page and
> > increases the access frequency of the region if the bit has set
> > meanwhile.  Therefore, the monitoring overhead is controllable by
> > setting the number of regions.
> > 
> > Nonetheless, this scheme cannot preserve the quality of the output if
> > the assumption is not kept.  Following commit will introduce how we can
> > make the guarantee with best effort.
> > 
> > Signed-off-by: SeongJae Park <sjpark@amazon.de>
> 
> Hi.
> 
> A few comments inline.
> 
> I've still not replicated your benchmarks so may well have some more
> feedback once I've managed that on one of our servers.

Appreciate your comments.  If you need any help for the replication, please let
me know.  I basically use my parsec3 wrapper scripts[1] to run parsec3 and
splash2x workloads and `damo` tool, which resides in the kernel tree at
`/tools/damon/`.

For example, below commands will reproduce ethp applied splash2x/fft run.
    
    $ echo "2M      null    5       null    null    null    hugepage
    2M      null    null    5       1s      null    nohugepage" > ethp
    $ parsec3_on_ubuntu/run.sh splash2x.fft
    $ linux/tools/damon/damo schemes -c ethp `pidof fft`

[1] https://github.com/sjp38/parsec3_on_ubuntu

> 
> Thanks,
> 
> Jonathan
> 
> > ---
> >  include/linux/damon.h |  24 ++
> >  mm/damon.c            | 553 ++++++++++++++++++++++++++++++++++++++++++
> >  2 files changed, 577 insertions(+)
> > 
[...]
> > diff --git a/mm/damon.c b/mm/damon.c
> > index d7e6226ab7f1..018016793555 100644
> > --- a/mm/damon.c
> > +++ b/mm/damon.c
> > @@ -10,8 +10,14 @@
> >  #define pr_fmt(fmt) "damon: " fmt
> >  
> >  #include <linux/damon.h>
> > +#include <linux/delay.h>
> > +#include <linux/kthread.h>
> >  #include <linux/mm.h>
> >  #include <linux/module.h>
> > +#include <linux/page_idle.h>
> > +#include <linux/random.h>
> > +#include <linux/sched/mm.h>
> > +#include <linux/sched/task.h>
> >  #include <linux/slab.h>
> >  
[...]
> > +/*
> > + * Size-evenly split a region into 'nr_pieces' small regions
> > + *
> > + * Returns 0 on success, or negative error code otherwise.
> > + */
> > +static int damon_split_region_evenly(struct damon_ctx *ctx,
> > +		struct damon_region *r, unsigned int nr_pieces)
> > +{
> > +	unsigned long sz_orig, sz_piece, orig_end;
> > +	struct damon_region *piece = NULL, *next;
> > +	unsigned long start;
> > +
> > +	if (!r || !nr_pieces)
> > +		return -EINVAL;
> > +
> > +	orig_end = r->vm_end;
> > +	sz_orig = r->vm_end - r->vm_start;
> > +	sz_piece = sz_orig / nr_pieces;
> > +
> > +	if (!sz_piece)
> > +		return -EINVAL;
> > +
> > +	r->vm_end = r->vm_start + sz_piece;
> > +	next = damon_next_region(r);
> > +	for (start = r->vm_end; start + sz_piece <= orig_end;
> > +			start += sz_piece) {
> > +		piece = damon_new_region(ctx, start, start + sz_piece);
> piece may be n

Yes, that name is short and more intuitive.  I will rename so.

> > +		damon_insert_region(piece, r, next);
> > +		r = piece;
> > +	}
> > +	/* complement last region for possible rounding error */
> > +	if (piece)
> > +		piece->vm_end = orig_end;
> 
> Update the sampling address to ensure it's in the region?

I think `piece->vm_end` should be equal or smaller than `orig_end` and
therefore the sampling address of `piece` will be still in the region.

> 
> > +
> > +	return 0;
> > +}
> > +
[...]
> > +static void damon_pte_pmd_mkold(pte_t *pte, pmd_t *pmd)
> > +{
> > +	if (pte) {
> > +		if (pte_young(*pte)) {
> > +			clear_page_idle(pte_page(*pte));
> > +			set_page_young(pte_page(*pte));
> > +		}
> > +		*pte = pte_mkold(*pte);
> > +		return;
> > +	}
> > +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > +	if (pmd) {
> > +		if (pmd_young(*pmd)) {
> > +			clear_page_idle(pmd_page(*pmd));
> > +			set_page_young(pmd_page(*pmd));
> > +		}
> > +		*pmd = pmd_mkold(*pmd);
> > +	}
> > +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
> 
> No need to flush the TLBs?

Good point!

I have intentionally skipped TLB flushing here to minimize the performance
effect to the target workload.  I also thought this might not degrade the
monitoring accuracy so much because we are targetting for the DRAM level
accesses of memory-intensive workloads, which might make TLB flood frequently.

However, your comment makes me thinking differently now.  By flushing the TLB
here, we will increase up to `number_of_regions` TLB misses for sampling
interval.  This might be not a huge overhead.  Also, improving the monitoring
accuracy makes no harm at all.  I even didn't measured the overhead.

I will test the overhead and if it is not significant, I will make this code to
flush TLB, in the next spin.

> 
> > +}
> > +
[...]
> > +/*
> > + * The monitoring daemon that runs as a kernel thread
> > + */
> > +static int kdamond_fn(void *data)
> > +{
> > +	struct damon_ctx *ctx = data;
> > +	struct damon_task *t;
> > +	struct damon_region *r, *next;
> > +	struct mm_struct *mm;
> > +
> > +	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
> > +	kdamond_init_regions(ctx);
> 
> We haven't called mkold on the initial regions so first check will
> get us fairly random state.

Yes, indeed.  However, the early results will not be accurate anyway because
the adaptive regions adjustment algorithm will not take effect yet.  I would
like to leave this part as is but add some comments about this point to keep
the code simple.

> 
> > +	while (!kdamond_need_stop(ctx)) {
> > +		damon_for_each_task(ctx, t) {
> > +			mm = damon_get_mm(t);
> > +			if (!mm)
> > +				continue;
> > +			damon_for_each_region(r, t)
> > +				kdamond_check_access(ctx, mm, r);
> > +			mmput(mm);
> > +		}
> > +
> > +		if (kdamond_aggregate_interval_passed(ctx))
> > +			kdamond_reset_aggregated(ctx);
> > +
> > +		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
> > +	}
> > +	damon_for_each_task(ctx, t) {
> > +		damon_for_each_region_safe(r, next, t)
> > +			damon_destroy_region(r);
> > +	}
> > +	pr_debug("kdamond (%d) finishes\n", ctx->kdamond->pid);
> > +	mutex_lock(&ctx->kdamond_lock);
> > +	ctx->kdamond = NULL;
> > +	mutex_unlock(&ctx->kdamond_lock);
> > +
> > +	return 0;
> > +}
> > +
[...]
> > +/*
> > + * Start or stop the kdamond
> > + *
> > + * Returns 0 if success, negative error code otherwise.
> > + */
> > +static int damon_turn_kdamond(struct damon_ctx *ctx, bool on)
> > +{
> > +	int err = -EBUSY;
> > +
> > +	mutex_lock(&ctx->kdamond_lock);
> > +	if (!ctx->kdamond && on) {
> 
> Given there is very little shared code between on and off, I would
> suggest just splitting it into two functions.

Good point, I will do so in next spin.

> 
> > +		err = 0;
> > +		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond");
> > +		if (IS_ERR(ctx->kdamond))
> > +			err = PTR_ERR(ctx->kdamond);
> > +	} else if (ctx->kdamond && !on) {
> > +		mutex_unlock(&ctx->kdamond_lock);
> > +		kthread_stop(ctx->kdamond);
> > +		while (damon_kdamond_running(ctx))
> > +			usleep_range(ctx->sample_interval,
> > +					ctx->sample_interval * 2);
> > +		return 0;
> > +	}
> > +	mutex_unlock(&ctx->kdamond_lock);
> > +
> > +	return err;
> > +}
> > +
[...]
> > +
> > +/*
> 
> Why not make these actual kernel-doc?  That way you can use the
> kernel-doc scripts to sanity check them.

Oops, I just forgot that it should start with '/**'.  Will fix it in next spin.


Thanks,
SeongJae Park

> 
> /**
> 
> > + * damon_set_attrs() - Set attributes for the monitoring.
> > + * @ctx:		monitoring context
> > + * @sample_int:		time interval between samplings
> > + * @aggr_int:		time interval between aggregations
> > + * @min_nr_reg:		minimal number of regions
> > + *
> > + * This function should not be called while the kdamond is running.
> > + * Every time interval is in micro-seconds.
> > + *
> > + * Return: 0 on success, negative error code otherwise.
> > + */
> > +int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> > +		unsigned long aggr_int, unsigned long min_nr_reg)
> > +{
> > +	if (min_nr_reg < 3) {
> > +		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
> > +				min_nr_reg);
> > +		return -EINVAL;
> > +	}
> > +
> > +	ctx->sample_interval = sample_int;
> > +	ctx->aggr_interval = aggr_int;
> > +	ctx->min_nr_regions = min_nr_reg;
> > +
> > +	return 0;
> > +}
> > +
> >  static int __init damon_init(void)
> >  {
> >  	return 0;
> 

^ permalink raw reply	[flat|nested] 28+ messages in thread

* Re: Re: [PATCH v7 05/15] mm/damon: Adaptively adjust regions
  2020-03-31 16:08   ` Jonathan Cameron
@ 2020-04-01  8:22     ` SeongJae Park
  0 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-04-01  8:22 UTC (permalink / raw)
  To: Jonathan Cameron
  Cc: SeongJae Park, akpm, SeongJae Park, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

On Tue, 31 Mar 2020 17:08:55 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:

> On Wed, 18 Mar 2020 12:27:12 +0100
> SeongJae Park <sjpark@amazon.com> wrote:
> 
> > From: SeongJae Park <sjpark@amazon.de>
> > 
> > At the beginning of the monitoring, DAMON constructs the initial regions
> > by evenly splitting the memory mapped address space of the process into
> > the user-specified minimal number of regions.  In this initial state,
> > the assumption of the regions (pages in same region have similar access
> > frequencies) is normally not kept and thus the monitoring quality could
> > be low.  To keep the assumption as much as possible, DAMON adaptively
> > merges and splits each region.
> > 
> > For each ``aggregation interval``, it compares the access frequencies of
> > adjacent regions and merges those if the frequency difference is small.
> > Then, after it reports and clears the aggregated access frequency of
> > each region, it splits each region into two regions if the total number
> > of regions is smaller than the half of the user-specified maximum number
> > of regions.
> > 
> > In this way, DAMON provides its best-effort quality and minimal overhead
> > while keeping the bounds users set for their trade-off.
> > 
> > Signed-off-by: SeongJae Park <sjpark@amazon.de>
> 
> A few more edge cases in here, and a suggestion that might be more costly
> but lead to simpler code.

Thank you for finding those!

> 
> Jonathan
> 
> > ---
> >  include/linux/damon.h |   6 +-
> >  mm/damon.c            | 148 ++++++++++++++++++++++++++++++++++++++++--
> >  2 files changed, 145 insertions(+), 9 deletions(-)
> > 
[...]
> > diff --git a/mm/damon.c b/mm/damon.c
> > index 018016793555..23c0de3b502e 100644
> > --- a/mm/damon.c
> > +++ b/mm/damon.c
[...]
> > +
> > +/*
> > + * Split a region into two small regions
> > + *
> > + * r		the region to be split
> > + * sz_r		size of the first sub-region that will be made
> > + */
> > +static void damon_split_region_at(struct damon_ctx *ctx,
> > +		struct damon_region *r, unsigned long sz_r)
> > +{
> > +	struct damon_region *new;
> > +
> > +	new = damon_new_region(ctx, r->vm_start + sz_r, r->vm_end);
> > +	r->vm_end = new->vm_start;
> 
> We may well have a sampling address that is in the wrong region.
> It should have little effect on the stats as will fix on next sample
> but in my view still worth cleaning up.

Good catch!  I will fix this in next spin.

> 
> > +
> > +	damon_insert_region(new, r, damon_next_region(r));
> > +}
> > +
[...]
> > @@ -571,21 +689,29 @@ static int kdamond_fn(void *data)
> >  	struct damon_task *t;
> >  	struct damon_region *r, *next;
> >  	struct mm_struct *mm;
> > +	unsigned int max_nr_accesses;
> >  
> >  	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
> >  	kdamond_init_regions(ctx);
> >  	while (!kdamond_need_stop(ctx)) {
> > +		max_nr_accesses = 0;
> >  		damon_for_each_task(ctx, t) {
> >  			mm = damon_get_mm(t);
> >  			if (!mm)
> >  				continue;
> > -			damon_for_each_region(r, t)
> > +			damon_for_each_region(r, t) {
> >  				kdamond_check_access(ctx, mm, r);
> > +				max_nr_accesses = max(r->nr_accesses,
> > +						max_nr_accesses);
> > +			}
> >  			mmput(mm);
> >  		}
> >  
> > -		if (kdamond_aggregate_interval_passed(ctx))
> > +		if (kdamond_aggregate_interval_passed(ctx)) {
> > +			kdamond_merge_regions(ctx, max_nr_accesses / 10);
> >  			kdamond_reset_aggregated(ctx);
> > +			kdamond_split_regions(ctx);
> > +		}
> 
> I wonder if it would be simpler to split the sampling address setup and
> mkold from the access check.  We would have to walk regions twice,
> but not have to bother separately dealing with updating some regions
> if they are modified in the above block.
> 
> Also, the above has some overhead, so will bias that first sample each
> time the block above runs.  If we do the mkold afterwards it will make
> much less difference.

Agreed, it will make code much more simple and easy to read.  However, I'm not
sure how much of the overhead will be biased because 'aggregate interval' is
usually larger than 'sampling interval'.  Anyway, Will change so in the next
spin!


Thanks,
SeongJae Park

[...]

^ permalink raw reply	[flat|nested] 28+ messages in thread

* Re: [PATCH v7 04/15] mm/damon: Implement region based sampling
  2020-04-01  8:22     ` SeongJae Park
@ 2020-04-01 14:24       ` Jonathan Cameron
  2020-04-01 14:46         ` SeongJae Park
  2020-04-02 13:59       ` SeongJae Park
  1 sibling, 1 reply; 28+ messages in thread
From: Jonathan Cameron @ 2020-04-01 14:24 UTC (permalink / raw)
  To: SeongJae Park
  Cc: akpm, SeongJae Park, aarcange, acme, alexander.shishkin, amit,
	brendan.d.gregg, brendanhiggins, cai, colin.king, corbet, dwmw,
	jolsa, kirill, mark.rutland, mgorman, minchan, mingo, namhyung,
	peterz, rdunlap, riel, rientjes, rostedt, shuah, sj38.park,
	vbabka, vdavydov.dev, yang.shi, ying.huang, linux-mm, linux-doc,
	linux-kernel

On Wed, 1 Apr 2020 10:22:22 +0200
SeongJae Park <sjpark@amazon.com> wrote:

> On Tue, 31 Mar 2020 17:02:33 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:
> 
> > On Wed, 18 Mar 2020 12:27:11 +0100
> > SeongJae Park <sjpark@amazon.com> wrote:
> >   
> > > From: SeongJae Park <sjpark@amazon.de>
> > > 
> > > This commit implements DAMON's basic access check and region based
> > > sampling mechanisms.  This change would seems make no sense, mainly
> > > because it is only a part of the DAMON's logics.  Following two commits
> > > will make more sense.
> > > 
> > > Basic Access Check
> > > ------------------
> > > 
> > > DAMON basically reports what pages are how frequently accessed.  Note
> > > that the frequency is not an absolute number of accesses, but a relative
> > > frequency among the pages of the target workloads.
> > > 
> > > Users can control the resolution of the reports by setting two time
> > > intervals, ``sampling interval`` and ``aggregation interval``.  In
> > > detail, DAMON checks access to each page per ``sampling interval``,
> > > aggregates the results (counts the number of the accesses to each page),
> > > and reports the aggregated results per ``aggregation interval``.  For
> > > the access check of each page, DAMON uses the Accessed bits of PTEs.
> > > 
> > > This is thus similar to common periodic access checks based access
> > > tracking mechanisms, which overhead is increasing as the size of the
> > > target process grows.
> > > 
> > > Region Based Sampling
> > > ---------------------
> > > 
> > > To avoid the unbounded increase of the overhead, DAMON groups a number
> > > of adjacent pages that assumed to have same access frequencies into a
> > > region.  As long as the assumption (pages in a region have same access
> > > frequencies) is kept, only one page in the region is required to be
> > > checked.  Thus, for each ``sampling interval``, DAMON randomly picks one
> > > page in each region and clears its Accessed bit.  After one more
> > > ``sampling interval``, DAMON reads the Accessed bit of the page and
> > > increases the access frequency of the region if the bit has set
> > > meanwhile.  Therefore, the monitoring overhead is controllable by
> > > setting the number of regions.
> > > 
> > > Nonetheless, this scheme cannot preserve the quality of the output if
> > > the assumption is not kept.  Following commit will introduce how we can
> > > make the guarantee with best effort.
> > > 
> > > Signed-off-by: SeongJae Park <sjpark@amazon.de>  
> > 
> > Hi.
> > 
> > A few comments inline.
> > 
> > I've still not replicated your benchmarks so may well have some more
> > feedback once I've managed that on one of our servers.  
> 
> Appreciate your comments.  If you need any help for the replication, please let
> me know.  I basically use my parsec3 wrapper scripts[1] to run parsec3 and
> splash2x workloads and `damo` tool, which resides in the kernel tree at
> `/tools/damon/`.
> 
> For example, below commands will reproduce ethp applied splash2x/fft run.
>     
>     $ echo "2M      null    5       null    null    null    hugepage
>     2M      null    null    5       1s      null    nohugepage" > ethp
>     $ parsec3_on_ubuntu/run.sh splash2x.fft
>     $ linux/tools/damon/damo schemes -c ethp `pidof fft`
> 
> [1] https://github.com/sjp38/parsec3_on_ubuntu


No significant problem, more a case of fitting this in between other things :)
+ some fixes needed for parsec3 to build for arm64.

> 
> > 
> > Thanks,
> > 
> > Jonathan
> >   
> > > ---
> > >  include/linux/damon.h |  24 ++
> > >  mm/damon.c            | 553 ++++++++++++++++++++++++++++++++++++++++++
> > >  2 files changed, 577 insertions(+)
> > >   
> [...]
> > > diff --git a/mm/damon.c b/mm/damon.c
> > > index d7e6226ab7f1..018016793555 100644
> > > --- a/mm/damon.c
> > > +++ b/mm/damon.c
> > > @@ -10,8 +10,14 @@
> > >  #define pr_fmt(fmt) "damon: " fmt
> > >  
> > >  #include <linux/damon.h>
> > > +#include <linux/delay.h>
> > > +#include <linux/kthread.h>
> > >  #include <linux/mm.h>
> > >  #include <linux/module.h>
> > > +#include <linux/page_idle.h>
> > > +#include <linux/random.h>
> > > +#include <linux/sched/mm.h>
> > > +#include <linux/sched/task.h>
> > >  #include <linux/slab.h>
> > >    
> [...]
> > > +/*
> > > + * Size-evenly split a region into 'nr_pieces' small regions
> > > + *
> > > + * Returns 0 on success, or negative error code otherwise.
> > > + */
> > > +static int damon_split_region_evenly(struct damon_ctx *ctx,
> > > +		struct damon_region *r, unsigned int nr_pieces)
> > > +{
> > > +	unsigned long sz_orig, sz_piece, orig_end;
> > > +	struct damon_region *piece = NULL, *next;
> > > +	unsigned long start;
> > > +
> > > +	if (!r || !nr_pieces)
> > > +		return -EINVAL;
> > > +
> > > +	orig_end = r->vm_end;
> > > +	sz_orig = r->vm_end - r->vm_start;
> > > +	sz_piece = sz_orig / nr_pieces;
> > > +
> > > +	if (!sz_piece)
> > > +		return -EINVAL;
> > > +
> > > +	r->vm_end = r->vm_start + sz_piece;

This is the end where it is unlikely the sampling address is
still in region.

(see below)

> > > +	next = damon_next_region(r);
> > > +	for (start = r->vm_end; start + sz_piece <= orig_end;
> > > +			start += sz_piece) {
> > > +		piece = damon_new_region(ctx, start, start + sz_piece);  
> > piece may be n  
> 
> Yes, that name is short and more intuitive.  I will rename so.
> 
> > > +		damon_insert_region(piece, r, next);
> > > +		r = piece;
> > > +	}
> > > +	/* complement last region for possible rounding error */
> > > +	if (piece)
> > > +		piece->vm_end = orig_end;  
> > 
> > Update the sampling address to ensure it's in the region?  
> 
> I think `piece->vm_end` should be equal or smaller than `orig_end` and
> therefore the sampling address of `piece` will be still in the region.

Good point.  The one above however is more of an issue I think..
So the region we modify before adding the new regions.

> 
> >   
> > > +
> > > +	return 0;
> > > +}
> > > +  
> [...]
> > > +static void damon_pte_pmd_mkold(pte_t *pte, pmd_t *pmd)
> > > +{
> > > +	if (pte) {
> > > +		if (pte_young(*pte)) {
> > > +			clear_page_idle(pte_page(*pte));
> > > +			set_page_young(pte_page(*pte));
> > > +		}
> > > +		*pte = pte_mkold(*pte);
> > > +		return;
> > > +	}
> > > +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > > +	if (pmd) {
> > > +		if (pmd_young(*pmd)) {
> > > +			clear_page_idle(pmd_page(*pmd));
> > > +			set_page_young(pmd_page(*pmd));
> > > +		}
> > > +		*pmd = pmd_mkold(*pmd);
> > > +	}
> > > +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */  
> > 
> > No need to flush the TLBs?  
> 
> Good point!
> 
> I have intentionally skipped TLB flushing here to minimize the performance
> effect to the target workload.  I also thought this might not degrade the
> monitoring accuracy so much because we are targetting for the DRAM level
> accesses of memory-intensive workloads, which might make TLB flood frequently.
> 
> However, your comment makes me thinking differently now.  By flushing the TLB
> here, we will increase up to `number_of_regions` TLB misses for sampling
> interval.  This might be not a huge overhead.  Also, improving the monitoring
> accuracy makes no harm at all.  I even didn't measured the overhead.
> 
> I will test the overhead and if it is not significant, I will make this code to
> flush TLB, in the next spin.
> 
> >   
> > > +}
> > > +  
> [...]
> > > +/*
> > > + * The monitoring daemon that runs as a kernel thread
> > > + */
> > > +static int kdamond_fn(void *data)
> > > +{
> > > +	struct damon_ctx *ctx = data;
> > > +	struct damon_task *t;
> > > +	struct damon_region *r, *next;
> > > +	struct mm_struct *mm;
> > > +
> > > +	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
> > > +	kdamond_init_regions(ctx);  
> > 
> > We haven't called mkold on the initial regions so first check will
> > get us fairly random state.  
> 
> Yes, indeed.  However, the early results will not be accurate anyway because
> the adaptive regions adjustment algorithm will not take effect yet.  I would
> like to leave this part as is but add some comments about this point to keep
> the code simple.

I'd argue in favour of it being a low overhead and better to put them
in for 'correctness'.  It's much easier to discuss code that conforms to
a simple model (even if that makes the code more complex!)


> 
> >   
> > > +	while (!kdamond_need_stop(ctx)) {
> > > +		damon_for_each_task(ctx, t) {
> > > +			mm = damon_get_mm(t);
> > > +			if (!mm)
> > > +				continue;
> > > +			damon_for_each_region(r, t)
> > > +				kdamond_check_access(ctx, mm, r);
> > > +			mmput(mm);
> > > +		}
> > > +
> > > +		if (kdamond_aggregate_interval_passed(ctx))
> > > +			kdamond_reset_aggregated(ctx);
> > > +
> > > +		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
> > > +	}
> > > +	damon_for_each_task(ctx, t) {
> > > +		damon_for_each_region_safe(r, next, t)
> > > +			damon_destroy_region(r);
> > > +	}
> > > +	pr_debug("kdamond (%d) finishes\n", ctx->kdamond->pid);
> > > +	mutex_lock(&ctx->kdamond_lock);
> > > +	ctx->kdamond = NULL;
> > > +	mutex_unlock(&ctx->kdamond_lock);
> > > +
> > > +	return 0;
> > > +}
> > > +  
> [...]
> > > +/*
> > > + * Start or stop the kdamond
> > > + *
> > > + * Returns 0 if success, negative error code otherwise.
> > > + */
> > > +static int damon_turn_kdamond(struct damon_ctx *ctx, bool on)
> > > +{
> > > +	int err = -EBUSY;
> > > +
> > > +	mutex_lock(&ctx->kdamond_lock);
> > > +	if (!ctx->kdamond && on) {  
> > 
> > Given there is very little shared code between on and off, I would
> > suggest just splitting it into two functions.  
> 
> Good point, I will do so in next spin.
> 
> >   
> > > +		err = 0;
> > > +		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond");
> > > +		if (IS_ERR(ctx->kdamond))
> > > +			err = PTR_ERR(ctx->kdamond);
> > > +	} else if (ctx->kdamond && !on) {
> > > +		mutex_unlock(&ctx->kdamond_lock);
> > > +		kthread_stop(ctx->kdamond);
> > > +		while (damon_kdamond_running(ctx))
> > > +			usleep_range(ctx->sample_interval,
> > > +					ctx->sample_interval * 2);
> > > +		return 0;
> > > +	}
> > > +	mutex_unlock(&ctx->kdamond_lock);
> > > +
> > > +	return err;
> > > +}
> > > +  
> [...]
> > > +
> > > +/*  
> > 
> > Why not make these actual kernel-doc?  That way you can use the
> > kernel-doc scripts to sanity check them.  
> 
> Oops, I just forgot that it should start with '/**'.  Will fix it in next spin.

cool.

Thanks,

Jonathan

> 
> 
> Thanks,
> SeongJae Park
> 
> > 
> > /**
> >   
> > > + * damon_set_attrs() - Set attributes for the monitoring.
> > > + * @ctx:		monitoring context
> > > + * @sample_int:		time interval between samplings
> > > + * @aggr_int:		time interval between aggregations
> > > + * @min_nr_reg:		minimal number of regions
> > > + *
> > > + * This function should not be called while the kdamond is running.
> > > + * Every time interval is in micro-seconds.
> > > + *
> > > + * Return: 0 on success, negative error code otherwise.
> > > + */
> > > +int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> > > +		unsigned long aggr_int, unsigned long min_nr_reg)
> > > +{
> > > +	if (min_nr_reg < 3) {
> > > +		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
> > > +				min_nr_reg);
> > > +		return -EINVAL;
> > > +	}
> > > +
> > > +	ctx->sample_interval = sample_int;
> > > +	ctx->aggr_interval = aggr_int;
> > > +	ctx->min_nr_regions = min_nr_reg;
> > > +
> > > +	return 0;
> > > +}
> > > +
> > >  static int __init damon_init(void)
> > >  {
> > >  	return 0;  
> >   



^ permalink raw reply	[flat|nested] 28+ messages in thread

* Re: Re: [PATCH v7 04/15] mm/damon: Implement region based sampling
  2020-04-01 14:24       ` Jonathan Cameron
@ 2020-04-01 14:46         ` SeongJae Park
  0 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-04-01 14:46 UTC (permalink / raw)
  To: Jonathan Cameron
  Cc: SeongJae Park, akpm, SeongJae Park, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

On Wed, 1 Apr 2020 15:24:56 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:

> On Wed, 1 Apr 2020 10:22:22 +0200
> SeongJae Park <sjpark@amazon.com> wrote:
> 
> > On Tue, 31 Mar 2020 17:02:33 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:
> > 
> > > On Wed, 18 Mar 2020 12:27:11 +0100
> > > SeongJae Park <sjpark@amazon.com> wrote:
> > >   
> > > > From: SeongJae Park <sjpark@amazon.de>
> > > > 
> > > > This commit implements DAMON's basic access check and region based
> > > > sampling mechanisms.  This change would seems make no sense, mainly
> > > > because it is only a part of the DAMON's logics.  Following two commits
> > > > will make more sense.
> > > > 
> > > > Basic Access Check
> > > > ------------------
> > > > 
> > > > DAMON basically reports what pages are how frequently accessed.  Note
> > > > that the frequency is not an absolute number of accesses, but a relative
> > > > frequency among the pages of the target workloads.
> > > > 
> > > > Users can control the resolution of the reports by setting two time
> > > > intervals, ``sampling interval`` and ``aggregation interval``.  In
> > > > detail, DAMON checks access to each page per ``sampling interval``,
> > > > aggregates the results (counts the number of the accesses to each page),
> > > > and reports the aggregated results per ``aggregation interval``.  For
> > > > the access check of each page, DAMON uses the Accessed bits of PTEs.
> > > > 
> > > > This is thus similar to common periodic access checks based access
> > > > tracking mechanisms, which overhead is increasing as the size of the
> > > > target process grows.
> > > > 
> > > > Region Based Sampling
> > > > ---------------------
> > > > 
> > > > To avoid the unbounded increase of the overhead, DAMON groups a number
> > > > of adjacent pages that assumed to have same access frequencies into a
> > > > region.  As long as the assumption (pages in a region have same access
> > > > frequencies) is kept, only one page in the region is required to be
> > > > checked.  Thus, for each ``sampling interval``, DAMON randomly picks one
> > > > page in each region and clears its Accessed bit.  After one more
> > > > ``sampling interval``, DAMON reads the Accessed bit of the page and
> > > > increases the access frequency of the region if the bit has set
> > > > meanwhile.  Therefore, the monitoring overhead is controllable by
> > > > setting the number of regions.
> > > > 
> > > > Nonetheless, this scheme cannot preserve the quality of the output if
> > > > the assumption is not kept.  Following commit will introduce how we can
> > > > make the guarantee with best effort.
> > > > 
> > > > Signed-off-by: SeongJae Park <sjpark@amazon.de>  
> > > 
> > > Hi.
> > > 
> > > A few comments inline.
> > > 
> > > I've still not replicated your benchmarks so may well have some more
> > > feedback once I've managed that on one of our servers.  
> > 
> > Appreciate your comments.  If you need any help for the replication, please let
> > me know.  I basically use my parsec3 wrapper scripts[1] to run parsec3 and
> > splash2x workloads and `damo` tool, which resides in the kernel tree at
> > `/tools/damon/`.
> > 
> > For example, below commands will reproduce ethp applied splash2x/fft run.
> >     
> >     $ echo "2M      null    5       null    null    null    hugepage
> >     2M      null    null    5       1s      null    nohugepage" > ethp
> >     $ parsec3_on_ubuntu/run.sh splash2x.fft
> >     $ linux/tools/damon/damo schemes -c ethp `pidof fft`
> > 
> > [1] https://github.com/sjp38/parsec3_on_ubuntu
> 
> 
> No significant problem, more a case of fitting this in between other things :)
> + some fixes needed for parsec3 to build for arm64.

Cool :)

> 
> > 
> > > 
> > > Thanks,
> > > 
> > > Jonathan
> > >   
> > > > ---
> > > >  include/linux/damon.h |  24 ++
> > > >  mm/damon.c            | 553 ++++++++++++++++++++++++++++++++++++++++++
> > > >  2 files changed, 577 insertions(+)
> > > >   
> > [...]
> > > > diff --git a/mm/damon.c b/mm/damon.c
> > > > index d7e6226ab7f1..018016793555 100644
> > > > --- a/mm/damon.c
> > > > +++ b/mm/damon.c
> > > > @@ -10,8 +10,14 @@
> > > >  #define pr_fmt(fmt) "damon: " fmt
> > > >  
> > > >  #include <linux/damon.h>
> > > > +#include <linux/delay.h>
> > > > +#include <linux/kthread.h>
> > > >  #include <linux/mm.h>
> > > >  #include <linux/module.h>
> > > > +#include <linux/page_idle.h>
> > > > +#include <linux/random.h>
> > > > +#include <linux/sched/mm.h>
> > > > +#include <linux/sched/task.h>
> > > >  #include <linux/slab.h>
> > > >    
> > [...]
> > > > +/*
> > > > + * Size-evenly split a region into 'nr_pieces' small regions
> > > > + *
> > > > + * Returns 0 on success, or negative error code otherwise.
> > > > + */
> > > > +static int damon_split_region_evenly(struct damon_ctx *ctx,
> > > > +		struct damon_region *r, unsigned int nr_pieces)
> > > > +{
> > > > +	unsigned long sz_orig, sz_piece, orig_end;
> > > > +	struct damon_region *piece = NULL, *next;
> > > > +	unsigned long start;
> > > > +
> > > > +	if (!r || !nr_pieces)
> > > > +		return -EINVAL;
> > > > +
> > > > +	orig_end = r->vm_end;
> > > > +	sz_orig = r->vm_end - r->vm_start;
> > > > +	sz_piece = sz_orig / nr_pieces;
> > > > +
> > > > +	if (!sz_piece)
> > > > +		return -EINVAL;
> > > > +
> > > > +	r->vm_end = r->vm_start + sz_piece;
> 
> This is the end where it is unlikely the sampling address is
> still in region.

Ah, now I got your point!

> 
> (see below)
> 
> > > > +	next = damon_next_region(r);
> > > > +	for (start = r->vm_end; start + sz_piece <= orig_end;
> > > > +			start += sz_piece) {
> > > > +		piece = damon_new_region(ctx, start, start + sz_piece);  
> > > piece may be n  
> > 
> > Yes, that name is short and more intuitive.  I will rename so.
> > 
> > > > +		damon_insert_region(piece, r, next);
> > > > +		r = piece;
> > > > +	}
> > > > +	/* complement last region for possible rounding error */
> > > > +	if (piece)
> > > > +		piece->vm_end = orig_end;  
> > > 
> > > Update the sampling address to ensure it's in the region?  
> > 
> > I think `piece->vm_end` should be equal or smaller than `orig_end` and
> > therefore the sampling address of `piece` will be still in the region.
> 
> Good point.  The one above however is more of an issue I think..
> So the region we modify before adding the new regions.

Yes, you're right.  I will fix it in next spin.

> 
> > 
> > >   
> > > > +
> > > > +	return 0;
> > > > +}
> > > > +  
> > [...]
> > > > +static void damon_pte_pmd_mkold(pte_t *pte, pmd_t *pmd)
> > > > +{
> > > > +	if (pte) {
> > > > +		if (pte_young(*pte)) {
> > > > +			clear_page_idle(pte_page(*pte));
> > > > +			set_page_young(pte_page(*pte));
> > > > +		}
> > > > +		*pte = pte_mkold(*pte);
> > > > +		return;
> > > > +	}
> > > > +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > > > +	if (pmd) {
> > > > +		if (pmd_young(*pmd)) {
> > > > +			clear_page_idle(pmd_page(*pmd));
> > > > +			set_page_young(pmd_page(*pmd));
> > > > +		}
> > > > +		*pmd = pmd_mkold(*pmd);
> > > > +	}
> > > > +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */  
> > > 
> > > No need to flush the TLBs?  
> > 
> > Good point!
> > 
> > I have intentionally skipped TLB flushing here to minimize the performance
> > effect to the target workload.  I also thought this might not degrade the
> > monitoring accuracy so much because we are targetting for the DRAM level
> > accesses of memory-intensive workloads, which might make TLB flood frequently.
> > 
> > However, your comment makes me thinking differently now.  By flushing the TLB
> > here, we will increase up to `number_of_regions` TLB misses for sampling
> > interval.  This might be not a huge overhead.  Also, improving the monitoring
> > accuracy makes no harm at all.  I even didn't measured the overhead.
> > 
> > I will test the overhead and if it is not significant, I will make this code to
> > flush TLB, in the next spin.
> > 
> > >   
> > > > +}
> > > > +  
> > [...]
> > > > +/*
> > > > + * The monitoring daemon that runs as a kernel thread
> > > > + */
> > > > +static int kdamond_fn(void *data)
> > > > +{
> > > > +	struct damon_ctx *ctx = data;
> > > > +	struct damon_task *t;
> > > > +	struct damon_region *r, *next;
> > > > +	struct mm_struct *mm;
> > > > +
> > > > +	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
> > > > +	kdamond_init_regions(ctx);  
> > > 
> > > We haven't called mkold on the initial regions so first check will
> > > get us fairly random state.  
> > 
> > Yes, indeed.  However, the early results will not be accurate anyway because
> > the adaptive regions adjustment algorithm will not take effect yet.  I would
> > like to leave this part as is but add some comments about this point to keep
> > the code simple.
> 
> I'd argue in favour of it being a low overhead and better to put them
> in for 'correctness'.  It's much easier to discuss code that conforms to
> a simple model (even if that makes the code more complex!)

Agreed!  Will do so in next spin.

> 
> 
> > 
> > >   
> > > > +	while (!kdamond_need_stop(ctx)) {
> > > > +		damon_for_each_task(ctx, t) {
> > > > +			mm = damon_get_mm(t);
> > > > +			if (!mm)
> > > > +				continue;
> > > > +			damon_for_each_region(r, t)
> > > > +				kdamond_check_access(ctx, mm, r);
> > > > +			mmput(mm);
> > > > +		}
> > > > +
> > > > +		if (kdamond_aggregate_interval_passed(ctx))
> > > > +			kdamond_reset_aggregated(ctx);
> > > > +
> > > > +		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
> > > > +	}
> > > > +	damon_for_each_task(ctx, t) {
> > > > +		damon_for_each_region_safe(r, next, t)
> > > > +			damon_destroy_region(r);
> > > > +	}
> > > > +	pr_debug("kdamond (%d) finishes\n", ctx->kdamond->pid);
> > > > +	mutex_lock(&ctx->kdamond_lock);
> > > > +	ctx->kdamond = NULL;
> > > > +	mutex_unlock(&ctx->kdamond_lock);
> > > > +
> > > > +	return 0;
> > > > +}
> > > > +  
> > [...]
> > > > +/*
> > > > + * Start or stop the kdamond
> > > > + *
> > > > + * Returns 0 if success, negative error code otherwise.
> > > > + */
> > > > +static int damon_turn_kdamond(struct damon_ctx *ctx, bool on)
> > > > +{
> > > > +	int err = -EBUSY;
> > > > +
> > > > +	mutex_lock(&ctx->kdamond_lock);
> > > > +	if (!ctx->kdamond && on) {  
> > > 
> > > Given there is very little shared code between on and off, I would
> > > suggest just splitting it into two functions.  
> > 
> > Good point, I will do so in next spin.
> > 
> > >   
> > > > +		err = 0;
> > > > +		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond");
> > > > +		if (IS_ERR(ctx->kdamond))
> > > > +			err = PTR_ERR(ctx->kdamond);
> > > > +	} else if (ctx->kdamond && !on) {
> > > > +		mutex_unlock(&ctx->kdamond_lock);
> > > > +		kthread_stop(ctx->kdamond);
> > > > +		while (damon_kdamond_running(ctx))
> > > > +			usleep_range(ctx->sample_interval,
> > > > +					ctx->sample_interval * 2);
> > > > +		return 0;
> > > > +	}
> > > > +	mutex_unlock(&ctx->kdamond_lock);
> > > > +
> > > > +	return err;
> > > > +}
> > > > +  
> > [...]
> > > > +
> > > > +/*  
> > > 
> > > Why not make these actual kernel-doc?  That way you can use the
> > > kernel-doc scripts to sanity check them.  
> > 
> > Oops, I just forgot that it should start with '/**'.  Will fix it in next spin.
> 
> cool.
> 
> Thanks,
> 
> Jonathan

:)


Thanks,
SeongJae Park
[...]

^ permalink raw reply	[flat|nested] 28+ messages in thread

* Re: Re: Re: [PATCH v7 04/15] mm/damon: Implement region based sampling
  2020-04-01  8:22     ` SeongJae Park
  2020-04-01 14:24       ` Jonathan Cameron
@ 2020-04-02 13:59       ` SeongJae Park
  2020-04-02 17:24         ` Jonathan Cameron
  1 sibling, 1 reply; 28+ messages in thread
From: SeongJae Park @ 2020-04-02 13:59 UTC (permalink / raw)
  To: SeongJae Park
  Cc: Jonathan Cameron, akpm, SeongJae Park, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

On Wed, 1 Apr 2020 10:22:22 +0200 SeongJae Park <sjpark@amazon.com> wrote:

> On Tue, 31 Mar 2020 17:02:33 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:
> 
> > On Wed, 18 Mar 2020 12:27:11 +0100
> > SeongJae Park <sjpark@amazon.com> wrote:
> > 
> > > From: SeongJae Park <sjpark@amazon.de>
> > > 
> > > This commit implements DAMON's basic access check and region based
> > > sampling mechanisms.  This change would seems make no sense, mainly
> > > because it is only a part of the DAMON's logics.  Following two commits
> > > will make more sense.
> > > 
> > > Basic Access Check
> > > ------------------
> > > 
> > > DAMON basically reports what pages are how frequently accessed.  Note
> > > that the frequency is not an absolute number of accesses, but a relative
> > > frequency among the pages of the target workloads.
> > > 
> > > Users can control the resolution of the reports by setting two time
> > > intervals, ``sampling interval`` and ``aggregation interval``.  In
> > > detail, DAMON checks access to each page per ``sampling interval``,
> > > aggregates the results (counts the number of the accesses to each page),
> > > and reports the aggregated results per ``aggregation interval``.  For
> > > the access check of each page, DAMON uses the Accessed bits of PTEs.
> > > 
> > > This is thus similar to common periodic access checks based access
> > > tracking mechanisms, which overhead is increasing as the size of the
> > > target process grows.
> > > 
> > > Region Based Sampling
> > > ---------------------
> > > 
> > > To avoid the unbounded increase of the overhead, DAMON groups a number
> > > of adjacent pages that assumed to have same access frequencies into a
> > > region.  As long as the assumption (pages in a region have same access
> > > frequencies) is kept, only one page in the region is required to be
> > > checked.  Thus, for each ``sampling interval``, DAMON randomly picks one
> > > page in each region and clears its Accessed bit.  After one more
> > > ``sampling interval``, DAMON reads the Accessed bit of the page and
> > > increases the access frequency of the region if the bit has set
> > > meanwhile.  Therefore, the monitoring overhead is controllable by
> > > setting the number of regions.
> > > 
> > > Nonetheless, this scheme cannot preserve the quality of the output if
> > > the assumption is not kept.  Following commit will introduce how we can
> > > make the guarantee with best effort.
> > > 
> > > Signed-off-by: SeongJae Park <sjpark@amazon.de>
> > 
> > Hi.
> > 
> > A few comments inline.
> > 
> > I've still not replicated your benchmarks so may well have some more
> > feedback once I've managed that on one of our servers.
> 
> Appreciate your comments.  If you need any help for the replication, please let
> me know.  I basically use my parsec3 wrapper scripts[1] to run parsec3 and
> splash2x workloads and `damo` tool, which resides in the kernel tree at
> `/tools/damon/`.
> 
> For example, below commands will reproduce ethp applied splash2x/fft run.
>     
>     $ echo "2M      null    5       null    null    null    hugepage
>     2M      null    null    5       1s      null    nohugepage" > ethp
>     $ parsec3_on_ubuntu/run.sh splash2x.fft
>     $ linux/tools/damon/damo schemes -c ethp `pidof fft`
> 
> [1] https://github.com/sjp38/parsec3_on_ubuntu
> 
> > 
> > Thanks,
> > 
> > Jonathan
> > 
> > > ---
> > >  include/linux/damon.h |  24 ++
> > >  mm/damon.c            | 553 ++++++++++++++++++++++++++++++++++++++++++
> > >  2 files changed, 577 insertions(+)
> > > 
> [...]
> > > diff --git a/mm/damon.c b/mm/damon.c
> > > index d7e6226ab7f1..018016793555 100644
> > > --- a/mm/damon.c
> > > +++ b/mm/damon.c
> > > @@ -10,8 +10,14 @@
> > >  #define pr_fmt(fmt) "damon: " fmt
> > >  
> > >  #include <linux/damon.h>
> > > +#include <linux/delay.h>
> > > +#include <linux/kthread.h>
> > >  #include <linux/mm.h>
> > >  #include <linux/module.h>
> > > +#include <linux/page_idle.h>
> > > +#include <linux/random.h>
> > > +#include <linux/sched/mm.h>
> > > +#include <linux/sched/task.h>
> > >  #include <linux/slab.h>
> > >  
> [...]
> > > +/*
> > > + * Size-evenly split a region into 'nr_pieces' small regions
> > > + *
> > > + * Returns 0 on success, or negative error code otherwise.
> > > + */
> > > +static int damon_split_region_evenly(struct damon_ctx *ctx,
> > > +		struct damon_region *r, unsigned int nr_pieces)
> > > +{
> > > +	unsigned long sz_orig, sz_piece, orig_end;
> > > +	struct damon_region *piece = NULL, *next;
> > > +	unsigned long start;
> > > +
> > > +	if (!r || !nr_pieces)
> > > +		return -EINVAL;
> > > +
> > > +	orig_end = r->vm_end;
> > > +	sz_orig = r->vm_end - r->vm_start;
> > > +	sz_piece = sz_orig / nr_pieces;
> > > +
> > > +	if (!sz_piece)
> > > +		return -EINVAL;
> > > +
> > > +	r->vm_end = r->vm_start + sz_piece;
> > > +	next = damon_next_region(r);
> > > +	for (start = r->vm_end; start + sz_piece <= orig_end;
> > > +			start += sz_piece) {
> > > +		piece = damon_new_region(ctx, start, start + sz_piece);
> > piece may be n
> 
> Yes, that name is short and more intuitive.  I will rename so.
> 
> > > +		damon_insert_region(piece, r, next);
> > > +		r = piece;
> > > +	}
> > > +	/* complement last region for possible rounding error */
> > > +	if (piece)
> > > +		piece->vm_end = orig_end;
> > 
> > Update the sampling address to ensure it's in the region?
> 
> I think `piece->vm_end` should be equal or smaller than `orig_end` and
> therefore the sampling address of `piece` will be still in the region.
> 
> > 
> > > +
> > > +	return 0;
> > > +}
> > > +
> [...]
> > > +static void damon_pte_pmd_mkold(pte_t *pte, pmd_t *pmd)
> > > +{
> > > +	if (pte) {
> > > +		if (pte_young(*pte)) {
> > > +			clear_page_idle(pte_page(*pte));
> > > +			set_page_young(pte_page(*pte));
> > > +		}
> > > +		*pte = pte_mkold(*pte);
> > > +		return;
> > > +	}
> > > +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > > +	if (pmd) {
> > > +		if (pmd_young(*pmd)) {
> > > +			clear_page_idle(pmd_page(*pmd));
> > > +			set_page_young(pmd_page(*pmd));
> > > +		}
> > > +		*pmd = pmd_mkold(*pmd);
> > > +	}
> > > +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
> > 
> > No need to flush the TLBs?
> 
> Good point!
> 
> I have intentionally skipped TLB flushing here to minimize the performance
> effect to the target workload.  I also thought this might not degrade the
> monitoring accuracy so much because we are targetting for the DRAM level
> accesses of memory-intensive workloads, which might make TLB flood frequently.
> 
> However, your comment makes me thinking differently now.  By flushing the TLB
> here, we will increase up to `number_of_regions` TLB misses for sampling
> interval.  This might be not a huge overhead.  Also, improving the monitoring
> accuracy makes no harm at all.  I even didn't measured the overhead.
> 
> I will test the overhead and if it is not significant, I will make this code to
> flush TLB, in the next spin.

Hmm, it seems like 'page_idle.c' is also modifying the Accessed bit but doesn't
flush related TLB entries.  If I'm not missing something here, I would like to
leave this part as is to make the behavior consistent.


Thanks,
SeongJae Park

> 
> > 
> > > +}
> > > +
> [...]
> > > +/*
> > > + * The monitoring daemon that runs as a kernel thread
> > > + */
> > > +static int kdamond_fn(void *data)
> > > +{
> > > +	struct damon_ctx *ctx = data;
> > > +	struct damon_task *t;
> > > +	struct damon_region *r, *next;
> > > +	struct mm_struct *mm;
> > > +
> > > +	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
> > > +	kdamond_init_regions(ctx);
> > 
> > We haven't called mkold on the initial regions so first check will
> > get us fairly random state.
> 
> Yes, indeed.  However, the early results will not be accurate anyway because
> the adaptive regions adjustment algorithm will not take effect yet.  I would
> like to leave this part as is but add some comments about this point to keep
> the code simple.
> 
> > 
> > > +	while (!kdamond_need_stop(ctx)) {
> > > +		damon_for_each_task(ctx, t) {
> > > +			mm = damon_get_mm(t);
> > > +			if (!mm)
> > > +				continue;
> > > +			damon_for_each_region(r, t)
> > > +				kdamond_check_access(ctx, mm, r);
> > > +			mmput(mm);
> > > +		}
> > > +
> > > +		if (kdamond_aggregate_interval_passed(ctx))
> > > +			kdamond_reset_aggregated(ctx);
> > > +
> > > +		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
> > > +	}
> > > +	damon_for_each_task(ctx, t) {
> > > +		damon_for_each_region_safe(r, next, t)
> > > +			damon_destroy_region(r);
> > > +	}
> > > +	pr_debug("kdamond (%d) finishes\n", ctx->kdamond->pid);
> > > +	mutex_lock(&ctx->kdamond_lock);
> > > +	ctx->kdamond = NULL;
> > > +	mutex_unlock(&ctx->kdamond_lock);
> > > +
> > > +	return 0;
> > > +}
> > > +
> [...]
> > > +/*
> > > + * Start or stop the kdamond
> > > + *
> > > + * Returns 0 if success, negative error code otherwise.
> > > + */
> > > +static int damon_turn_kdamond(struct damon_ctx *ctx, bool on)
> > > +{
> > > +	int err = -EBUSY;
> > > +
> > > +	mutex_lock(&ctx->kdamond_lock);
> > > +	if (!ctx->kdamond && on) {
> > 
> > Given there is very little shared code between on and off, I would
> > suggest just splitting it into two functions.
> 
> Good point, I will do so in next spin.
> 
> > 
> > > +		err = 0;
> > > +		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond");
> > > +		if (IS_ERR(ctx->kdamond))
> > > +			err = PTR_ERR(ctx->kdamond);
> > > +	} else if (ctx->kdamond && !on) {
> > > +		mutex_unlock(&ctx->kdamond_lock);
> > > +		kthread_stop(ctx->kdamond);
> > > +		while (damon_kdamond_running(ctx))
> > > +			usleep_range(ctx->sample_interval,
> > > +					ctx->sample_interval * 2);
> > > +		return 0;
> > > +	}
> > > +	mutex_unlock(&ctx->kdamond_lock);
> > > +
> > > +	return err;
> > > +}
> > > +
> [...]
> > > +
> > > +/*
> > 
> > Why not make these actual kernel-doc?  That way you can use the
> > kernel-doc scripts to sanity check them.
> 
> Oops, I just forgot that it should start with '/**'.  Will fix it in next spin.
> 
> 
> Thanks,
> SeongJae Park
> 
> > 
> > /**
> > 
> > > + * damon_set_attrs() - Set attributes for the monitoring.
> > > + * @ctx:		monitoring context
> > > + * @sample_int:		time interval between samplings
> > > + * @aggr_int:		time interval between aggregations
> > > + * @min_nr_reg:		minimal number of regions
> > > + *
> > > + * This function should not be called while the kdamond is running.
> > > + * Every time interval is in micro-seconds.
> > > + *
> > > + * Return: 0 on success, negative error code otherwise.
> > > + */
> > > +int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> > > +		unsigned long aggr_int, unsigned long min_nr_reg)
> > > +{
> > > +	if (min_nr_reg < 3) {
> > > +		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
> > > +				min_nr_reg);
> > > +		return -EINVAL;
> > > +	}
> > > +
> > > +	ctx->sample_interval = sample_int;
> > > +	ctx->aggr_interval = aggr_int;
> > > +	ctx->min_nr_regions = min_nr_reg;
> > > +
> > > +	return 0;
> > > +}
> > > +
> > >  static int __init damon_init(void)
> > >  {
> > >  	return 0;
> > 

^ permalink raw reply	[flat|nested] 28+ messages in thread

* Re: [PATCH v7 04/15] mm/damon: Implement region based sampling
  2020-03-18 11:27 ` [PATCH v7 04/15] mm/damon: Implement region based sampling SeongJae Park
  2020-03-31 16:02   ` Jonathan Cameron
@ 2020-04-02 15:07   ` SeongJae Park
  1 sibling, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-04-02 15:07 UTC (permalink / raw)
  To: SeongJae Park
  Cc: akpm, SeongJae Park, Jonathan.Cameron, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

On Wed, 18 Mar 2020 12:27:11 +0100 SeongJae Park <sjpark@amazon.com> wrote:

> From: SeongJae Park <sjpark@amazon.de>
> 
> This commit implements DAMON's basic access check and region based
> sampling mechanisms.  This change would seems make no sense, mainly
> because it is only a part of the DAMON's logics.  Following two commits
> will make more sense.
> 
> Basic Access Check
> ------------------
> 
> DAMON basically reports what pages are how frequently accessed.  Note
> that the frequency is not an absolute number of accesses, but a relative
> frequency among the pages of the target workloads.
> 
> Users can control the resolution of the reports by setting two time
> intervals, ``sampling interval`` and ``aggregation interval``.  In
> detail, DAMON checks access to each page per ``sampling interval``,
> aggregates the results (counts the number of the accesses to each page),
> and reports the aggregated results per ``aggregation interval``.  For
> the access check of each page, DAMON uses the Accessed bits of PTEs.
> 
> This is thus similar to common periodic access checks based access
> tracking mechanisms, which overhead is increasing as the size of the
> target process grows.
> 
> Region Based Sampling
> ---------------------
> 
> To avoid the unbounded increase of the overhead, DAMON groups a number
> of adjacent pages that assumed to have same access frequencies into a
> region.  As long as the assumption (pages in a region have same access
> frequencies) is kept, only one page in the region is required to be
> checked.  Thus, for each ``sampling interval``, DAMON randomly picks one
> page in each region and clears its Accessed bit.  After one more
> ``sampling interval``, DAMON reads the Accessed bit of the page and
> increases the access frequency of the region if the bit has set
> meanwhile.  Therefore, the monitoring overhead is controllable by
> setting the number of regions.
> 
> Nonetheless, this scheme cannot preserve the quality of the output if
> the assumption is not kept.  Following commit will introduce how we can
> make the guarantee with best effort.
> 
> Signed-off-by: SeongJae Park <sjpark@amazon.de>
> ---
>  include/linux/damon.h |  24 ++
>  mm/damon.c            | 553 ++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 577 insertions(+)
> 
> diff --git a/include/linux/damon.h b/include/linux/damon.h
> index 7117bb7e7544..f1945df6e6b4 100644
> --- a/include/linux/damon.h
> +++ b/include/linux/damon.h
> @@ -11,6 +11,8 @@
>  #define _DAMON_H_
>  
>  #include <linux/random.h>
> +#include <linux/mutex.h>
> +#include <linux/time64.h>
>  #include <linux/types.h>
>  
>  /* Represents a monitoring target region on the virtual address space */
> @@ -29,10 +31,32 @@ struct damon_task {
>  	struct list_head list;
>  };
>  
> +/*
> + * For each 'sample_interval', DAMON checks whether each region is accessed or
> + * not.  It aggregates and keeps the access information (number of accesses to
> + * each region) for each 'aggr_interval' time.
> + *
> + * All time intervals are in micro-seconds.
> + */
>  struct damon_ctx {
> +	unsigned long sample_interval;
> +	unsigned long aggr_interval;
> +	unsigned long min_nr_regions;
> +
> +	struct timespec64 last_aggregation;
> +
> +	struct task_struct *kdamond;
> +	struct mutex kdamond_lock;
> +
>  	struct rnd_state rndseed;
>  
>  	struct list_head tasks_list;	/* 'damon_task' objects */
>  };
>  
> +int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids);
> +int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> +		unsigned long aggr_int, unsigned long min_nr_reg);
> +int damon_start(struct damon_ctx *ctx);
> +int damon_stop(struct damon_ctx *ctx);
> +
>  #endif
> diff --git a/mm/damon.c b/mm/damon.c
> index d7e6226ab7f1..018016793555 100644
> --- a/mm/damon.c
> +++ b/mm/damon.c
> @@ -10,8 +10,14 @@
>  #define pr_fmt(fmt) "damon: " fmt
>  
>  #include <linux/damon.h>
> +#include <linux/delay.h>
> +#include <linux/kthread.h>
>  #include <linux/mm.h>
>  #include <linux/module.h>
> +#include <linux/page_idle.h>
> +#include <linux/random.h>
> +#include <linux/sched/mm.h>
> +#include <linux/sched/task.h>
>  #include <linux/slab.h>
>  
>  #define damon_get_task_struct(t) \
> @@ -171,6 +177,553 @@ static unsigned int nr_damon_regions(struct damon_task *t)
>  	return nr_regions;
>  }
>  
> +/*
> + * Get the mm_struct of the given task
> + *
> + * Caller should put the mm_struct after use, unless it is NULL.
> + *
> + * Returns the mm_struct of the task on success, NULL on failure
> + */
> +static struct mm_struct *damon_get_mm(struct damon_task *t)
> +{
> +	struct task_struct *task;
> +	struct mm_struct *mm;
> +
> +	task = damon_get_task_struct(t);
> +	if (!task)
> +		return NULL;
> +
> +	mm = get_task_mm(task);
> +	put_task_struct(task);
> +	return mm;
> +}
> +
> +/*
> + * Size-evenly split a region into 'nr_pieces' small regions
> + *
> + * Returns 0 on success, or negative error code otherwise.
> + */
> +static int damon_split_region_evenly(struct damon_ctx *ctx,
> +		struct damon_region *r, unsigned int nr_pieces)
> +{
> +	unsigned long sz_orig, sz_piece, orig_end;
> +	struct damon_region *piece = NULL, *next;
> +	unsigned long start;
> +
> +	if (!r || !nr_pieces)
> +		return -EINVAL;
> +
> +	orig_end = r->vm_end;
> +	sz_orig = r->vm_end - r->vm_start;
> +	sz_piece = sz_orig / nr_pieces;
> +
> +	if (!sz_piece)
> +		return -EINVAL;
> +
> +	r->vm_end = r->vm_start + sz_piece;
> +	next = damon_next_region(r);
> +	for (start = r->vm_end; start + sz_piece <= orig_end;
> +			start += sz_piece) {
> +		piece = damon_new_region(ctx, start, start + sz_piece);
> +		damon_insert_region(piece, r, next);
> +		r = piece;
> +	}
> +	/* complement last region for possible rounding error */
> +	if (piece)
> +		piece->vm_end = orig_end;
> +
> +	return 0;
> +}
> +
> +struct region {
> +	unsigned long start;
> +	unsigned long end;
> +};
> +
> +static unsigned long sz_region(struct region *r)
> +{
> +	return r->end - r->start;
> +}
> +
> +static void swap_regions(struct region *r1, struct region *r2)
> +{
> +	struct region tmp;
> +
> +	tmp = *r1;
> +	*r1 = *r2;
> +	*r2 = tmp;
> +}
> +
> +/*
> + * Find the three regions in an address space
> + *
> + * vma		the head vma of the target address space
> + * regions	an array of three 'struct region's that results will be saved
> + *
> + * This function receives an address space and finds three regions in it which
> + * separated by the two biggest unmapped regions in the space.  Please refer to
> + * below comments of 'damon_init_regions_of()' function to know why this is
> + * necessary.
> + *
> + * Returns 0 if success, or negative error code otherwise.
> + */
> +static int damon_three_regions_in_vmas(struct vm_area_struct *vma,
> +		struct region regions[3])
> +{
> +	struct region gap = {0,}, first_gap = {0,}, second_gap = {0,};
> +	struct vm_area_struct *last_vma = NULL;
> +	unsigned long start = 0;
> +
> +	/* Find two biggest gaps so that first_gap > second_gap > others */
> +	for (; vma; vma = vma->vm_next) {
> +		if (!last_vma) {
> +			start = vma->vm_start;
> +			last_vma = vma;
> +			continue;
> +		}
> +		gap.start = last_vma->vm_end;
> +		gap.end = vma->vm_start;
> +		if (sz_region(&gap) > sz_region(&second_gap)) {
> +			swap_regions(&gap, &second_gap);
> +			if (sz_region(&second_gap) > sz_region(&first_gap))
> +				swap_regions(&second_gap, &first_gap);
> +		}
> +		last_vma = vma;
> +	}
> +
> +	if (!sz_region(&second_gap) || !sz_region(&first_gap))
> +		return -EINVAL;
> +
> +	/* Sort the two biggest gaps by address */
> +	if (first_gap.start > second_gap.start)
> +		swap_regions(&first_gap, &second_gap);
> +
> +	/* Store the result */
> +	regions[0].start = start;
> +	regions[0].end = first_gap.start;
> +	regions[1].start = first_gap.end;
> +	regions[1].end = second_gap.start;
> +	regions[2].start = second_gap.end;
> +	regions[2].end = last_vma->vm_end;
> +
> +	return 0;
> +}
> +
> +/*
> + * Get the three regions in the given task
> + *
> + * Returns 0 on success, negative error code otherwise.
> + */
> +static int damon_three_regions_of(struct damon_task *t,
> +				struct region regions[3])
> +{
> +	struct mm_struct *mm;
> +	int rc;
> +
> +	mm = damon_get_mm(t);
> +	if (!mm)
> +		return -EINVAL;
> +
> +	down_read(&mm->mmap_sem);
> +	rc = damon_three_regions_in_vmas(mm->mmap, regions);
> +	up_read(&mm->mmap_sem);
> +
> +	mmput(mm);
> +	return rc;
> +}
> +
> +/*
> + * Initialize the monitoring target regions for the given task
> + *
> + * t	the given target task
> + *
> + * Because only a number of small portions of the entire address space
> + * is acutally mapped to the memory and accessed, monitoring the unmapped
> + * regions is wasteful.  That said, because we can deal with small noises,
> + * tracking every mapping is not strictly required but could even incur a high
> + * overhead if the mapping frequently changes or the number of mappings is
> + * high.  Nonetheless, this may seems very weird.  DAMON's dynamic regions
> + * adjustment mechanism, which will be implemented with following commit will
> + * make this more sense.
> + *
> + * For the reason, we convert the complex mappings to three distinct regions
> + * that cover every mapped areas of the address space.  Also the two gaps
> + * between the three regions are the two biggest unmapped areas in the given
> + * address space.  In detail, this function first identifies the start and the
> + * end of the mappings and the two biggest unmapped areas of the address space.
> + * Then, it constructs the three regions as below:
> + *
> + *     [mappings[0]->start, big_two_unmapped_areas[0]->start)
> + *     [big_two_unmapped_areas[0]->end, big_two_unmapped_areas[1]->start)
> + *     [big_two_unmapped_areas[1]->end, mappings[nr_mappings - 1]->end)
> + *
> + * As usual memory map of processes is as below, the gap between the heap and
> + * the uppermost mmap()-ed region, and the gap between the lowermost mmap()-ed
> + * region and the stack will be two biggest unmapped regions.  Because these
> + * gaps are exceptionally huge areas in usual address space, excluding these
> + * two biggest unmapped regions will be sufficient to make a trade-off.
> + *
> + *   <heap>
> + *   <BIG UNMAPPED REGION 1>
> + *   <uppermost mmap()-ed region>
> + *   (other mmap()-ed regions and small unmapped regions)
> + *   <lowermost mmap()-ed region>
> + *   <BIG UNMAPPED REGION 2>
> + *   <stack>
> + */
> +static void damon_init_regions_of(struct damon_ctx *c, struct damon_task *t)
> +{
> +	struct damon_region *r;
> +	struct region regions[3];
> +	int i;
> +
> +	if (damon_three_regions_of(t, regions)) {
> +		pr_err("Failed to get three regions of task %lu\n", t->pid);
> +		return;
> +	}
> +
> +	/* Set the initial three regions of the task */
> +	for (i = 0; i < 3; i++) {
> +		r = damon_new_region(c, regions[i].start, regions[i].end);
> +		damon_add_region(r, t);
> +	}
> +
> +	/* Split the middle region into 'min_nr_regions - 2' regions */
> +	r = damon_nth_region_of(t, 1);
> +	if (damon_split_region_evenly(c, r, c->min_nr_regions - 2))
> +		pr_warn("Init middle region failed to be split\n");
> +}
> +
> +/* Initialize '->regions_list' of every task */
> +static void kdamond_init_regions(struct damon_ctx *ctx)
> +{
> +	struct damon_task *t;
> +
> +	damon_for_each_task(ctx, t)
> +		damon_init_regions_of(ctx, t);
> +}
> +
> +static bool damon_pte_pmd_young(pte_t *pte, pmd_t *pmd)
> +{
> +	if (pte && pte_young(*pte))
> +		return true;
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +	if (pmd && pmd_young(*pmd))
> +		return true;
> +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
> +	return false;
> +}
> +
> +static void damon_pte_pmd_mkold(pte_t *pte, pmd_t *pmd)
> +{
> +	if (pte) {
> +		if (pte_young(*pte)) {
> +			clear_page_idle(pte_page(*pte));
> +			set_page_young(pte_page(*pte));
> +		}
> +		*pte = pte_mkold(*pte);
> +		return;
> +	}
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +	if (pmd) {
> +		if (pmd_young(*pmd)) {
> +			clear_page_idle(pmd_page(*pmd));
> +			set_page_young(pmd_page(*pmd));
> +		}
> +		*pmd = pmd_mkold(*pmd);
> +	}
> +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
> +}
> +
> +/*
> + * Check whether the region accessed and prepare for next check
> + *
> + * mm	'mm_struct' for the given virtual address space
> + * r	the region to be checked
> + */
> +static void kdamond_check_access(struct damon_ctx *ctx,
> +			struct mm_struct *mm, struct damon_region *r)
> +{
> +	static struct mm_struct *last_mm;
> +	static unsigned long last_addr;
> +	static int last_page_sz = PAGE_SIZE;
> +	static bool last_accessed;
> +
> +	pte_t *pte = NULL;
> +	pmd_t *pmd = NULL;
> +	spinlock_t *ptl;
> +
> +	/* If the region is in the last checked page, reuse the result */
> +	if (mm == last_mm && (ALIGN_DOWN(last_addr, last_page_sz) ==
> +				ALIGN_DOWN(r->sampling_addr, last_page_sz))) {
> +		if (last_accessed)
> +			r->nr_accesses++;
> +		return;
> +	}
> +
> +	if (follow_pte_pmd(mm, r->sampling_addr, NULL, &pte, &pmd, &ptl))
> +		goto prepare_next_check;
> +
> +	/* Read the page table access bit of the page */
> +	if (damon_pte_pmd_young(pte, pmd)) {
> +		last_accessed = true;
> +		r->nr_accesses++;
> +	}
> +	spin_unlock(ptl);

And, I found that I'm not setting 'last_accessed' to 'false' again.  This will
be fixed in the next spin, by setting it 'false' just before calling
'follow_pte_pmd()' above.


Thanks,
SeongJae Park

> +
> +prepare_next_check:
> +	last_mm = mm;
> +	last_addr = r->sampling_addr;
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +	last_page_sz = pte ? PAGE_SIZE : ((1UL) << HPAGE_PMD_SHIFT);
> +#endif
> +
> +	r->sampling_addr = damon_rand(ctx, r->vm_start, r->vm_end);
> +	pte = NULL, pmd = NULL;
> +	if (follow_pte_pmd(mm, r->sampling_addr, NULL, &pte, &pmd, &ptl))
> +		return;
> +
> +	damon_pte_pmd_mkold(pte, pmd);
> +	spin_unlock(ptl);
> +}
> +
> +/*
> + * damon_check_reset_time_interval() - Check if a time interval is elapsed.
> + * @baseline:	the time to check whether the interval has elapsed since
> + * @interval:	the time interval (microseconds)
> + *
> + * See whether the given time interval has passed since the given baseline
> + * time.  If so, it also updates the baseline to current time for next check.
> + *
> + * Return:	true if the time interval has passed, or false otherwise.
> + */
> +static bool damon_check_reset_time_interval(struct timespec64 *baseline,
> +		unsigned long interval)
> +{
> +	struct timespec64 now;
> +
> +	ktime_get_coarse_ts64(&now);
> +	if ((timespec64_to_ns(&now) - timespec64_to_ns(baseline)) <
> +			interval * 1000)
> +		return false;
> +	*baseline = now;
> +	return true;
> +}
> +
> +/*
> + * Check whether it is time to flush the aggregated information
> + */
> +static bool kdamond_aggregate_interval_passed(struct damon_ctx *ctx)
> +{
> +	return damon_check_reset_time_interval(&ctx->last_aggregation,
> +			ctx->aggr_interval);
> +}
> +
> +/*
> + * Reset the aggregated monitoring results
> + */
> +static void kdamond_reset_aggregated(struct damon_ctx *c)
> +{
> +	struct damon_task *t;
> +	struct damon_region *r;
> +
> +	damon_for_each_task(c, t) {
> +		damon_for_each_region(r, t)
> +			r->nr_accesses = 0;
> +	}
> +}
> +
> +/*
> + * Check whether current monitoring should be stopped
> + *
> + * If users asked to stop, need stop.  Even though no user has asked to stop,
> + * need stop if every target task has dead.
> + *
> + * Returns true if need to stop current monitoring.
> + */
> +static bool kdamond_need_stop(struct damon_ctx *ctx)
> +{
> +	struct damon_task *t;
> +	struct task_struct *task;
> +	bool stop;
> +
> +	stop = kthread_should_stop();
> +	if (stop)
> +		return true;
> +
> +	damon_for_each_task(ctx, t) {
> +		task = damon_get_task_struct(t);
> +		if (task) {
> +			put_task_struct(task);
> +			return false;
> +		}
> +	}
> +
> +	return true;
> +}
> +
> +/*
> + * The monitoring daemon that runs as a kernel thread
> + */
> +static int kdamond_fn(void *data)
> +{
> +	struct damon_ctx *ctx = data;
> +	struct damon_task *t;
> +	struct damon_region *r, *next;
> +	struct mm_struct *mm;
> +
> +	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
> +	kdamond_init_regions(ctx);
> +	while (!kdamond_need_stop(ctx)) {
> +		damon_for_each_task(ctx, t) {
> +			mm = damon_get_mm(t);
> +			if (!mm)
> +				continue;
> +			damon_for_each_region(r, t)
> +				kdamond_check_access(ctx, mm, r);
> +			mmput(mm);
> +		}
> +
> +		if (kdamond_aggregate_interval_passed(ctx))
> +			kdamond_reset_aggregated(ctx);
> +
> +		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
> +	}
> +	damon_for_each_task(ctx, t) {
> +		damon_for_each_region_safe(r, next, t)
> +			damon_destroy_region(r);
> +	}
> +	pr_debug("kdamond (%d) finishes\n", ctx->kdamond->pid);
> +	mutex_lock(&ctx->kdamond_lock);
> +	ctx->kdamond = NULL;
> +	mutex_unlock(&ctx->kdamond_lock);
> +
> +	return 0;
> +}
> +
> +/*
> + * Controller functions
> + */
> +
> +static bool damon_kdamond_running(struct damon_ctx *ctx)
> +{
> +	bool running;
> +
> +	mutex_lock(&ctx->kdamond_lock);
> +	running = ctx->kdamond != NULL;
> +	mutex_unlock(&ctx->kdamond_lock);
> +
> +	return running;
> +}
> +
> +/*
> + * Start or stop the kdamond
> + *
> + * Returns 0 if success, negative error code otherwise.
> + */
> +static int damon_turn_kdamond(struct damon_ctx *ctx, bool on)
> +{
> +	int err = -EBUSY;
> +
> +	mutex_lock(&ctx->kdamond_lock);
> +	if (!ctx->kdamond && on) {
> +		err = 0;
> +		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond");
> +		if (IS_ERR(ctx->kdamond))
> +			err = PTR_ERR(ctx->kdamond);
> +	} else if (ctx->kdamond && !on) {
> +		mutex_unlock(&ctx->kdamond_lock);
> +		kthread_stop(ctx->kdamond);
> +		while (damon_kdamond_running(ctx))
> +			usleep_range(ctx->sample_interval,
> +					ctx->sample_interval * 2);
> +		return 0;
> +	}
> +	mutex_unlock(&ctx->kdamond_lock);
> +
> +	return err;
> +}
> +
> +/*
> + * damon_start() - Starts monitoring with given context.
> + * @ctx:	monitoring context
> + *
> + * Return: 0 on success, negative error code otherwise.
> + */
> +int damon_start(struct damon_ctx *ctx)
> +{
> +	return damon_turn_kdamond(ctx, true);
> +}
> +
> +/*
> + * damon_stop() - Stops monitoring of given context.
> + * @ctx:	monitoring context
> + *
> + * Return: 0 on success, negative error code otherwise.
> + */
> +int damon_stop(struct damon_ctx *ctx)
> +{
> +	return damon_turn_kdamond(ctx, false);
> +}
> +
> +/*
> + * damon_set_pids() - Set monitoring target processes.
> + * @ctx:	monitoring context
> + * @pids:	array of target processes pids
> + * @nr_pids:	number of entries in @pids
> + *
> + * This function should not be called while the kdamond is running.
> + *
> + * Return: 0 on usccess, negative error code otherwise.
> + */
> +int damon_set_pids(struct damon_ctx *ctx, unsigned long *pids, ssize_t nr_pids)
> +{
> +	ssize_t i;
> +	struct damon_task *t, *next;
> +
> +	damon_for_each_task_safe(ctx, t, next)
> +		damon_destroy_task(t);
> +
> +	for (i = 0; i < nr_pids; i++) {
> +		t = damon_new_task(pids[i]);
> +		if (!t) {
> +			pr_err("Failed to alloc damon_task\n");
> +			return -ENOMEM;
> +		}
> +		damon_add_task(ctx, t);
> +	}
> +
> +	return 0;
> +}
> +
> +/*
> + * damon_set_attrs() - Set attributes for the monitoring.
> + * @ctx:		monitoring context
> + * @sample_int:		time interval between samplings
> + * @aggr_int:		time interval between aggregations
> + * @min_nr_reg:		minimal number of regions
> + *
> + * This function should not be called while the kdamond is running.
> + * Every time interval is in micro-seconds.
> + *
> + * Return: 0 on success, negative error code otherwise.
> + */
> +int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> +		unsigned long aggr_int, unsigned long min_nr_reg)
> +{
> +	if (min_nr_reg < 3) {
> +		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
> +				min_nr_reg);
> +		return -EINVAL;
> +	}
> +
> +	ctx->sample_interval = sample_int;
> +	ctx->aggr_interval = aggr_int;
> +	ctx->min_nr_regions = min_nr_reg;
> +
> +	return 0;
> +}
> +
>  static int __init damon_init(void)
>  {
>  	return 0;
> -- 
> 2.17.1
> 

^ permalink raw reply	[flat|nested] 28+ messages in thread

* Re: [PATCH v7 04/15] mm/damon: Implement region based sampling
  2020-04-02 13:59       ` SeongJae Park
@ 2020-04-02 17:24         ` Jonathan Cameron
  2020-04-02 18:00           ` SeongJae Park
  0 siblings, 1 reply; 28+ messages in thread
From: Jonathan Cameron @ 2020-04-02 17:24 UTC (permalink / raw)
  To: SeongJae Park
  Cc: akpm, SeongJae Park, aarcange, acme, alexander.shishkin, amit,
	brendan.d.gregg, brendanhiggins, cai, colin.king, corbet, dwmw,
	jolsa, kirill, mark.rutland, mgorman, minchan, mingo, namhyung,
	peterz, rdunlap, riel, rientjes, rostedt, shuah, sj38.park,
	vbabka, vdavydov.dev, yang.shi, ying.huang, linux-mm, linux-doc,
	linux-kernel

On Thu, 2 Apr 2020 15:59:59 +0200
SeongJae Park <sjpark@amazon.com> wrote:

> On Wed, 1 Apr 2020 10:22:22 +0200 SeongJae Park <sjpark@amazon.com> wrote:
> 
> > On Tue, 31 Mar 2020 17:02:33 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:
> >   
> > > On Wed, 18 Mar 2020 12:27:11 +0100
> > > SeongJae Park <sjpark@amazon.com> wrote:
> > >   
> > > > From: SeongJae Park <sjpark@amazon.de>
> > > > 
> > > > This commit implements DAMON's basic access check and region based
> > > > sampling mechanisms.  This change would seems make no sense, mainly
> > > > because it is only a part of the DAMON's logics.  Following two commits
> > > > will make more sense.
> > > > 
> > > > Basic Access Check
> > > > ------------------
> > > > 
> > > > DAMON basically reports what pages are how frequently accessed.  Note
> > > > that the frequency is not an absolute number of accesses, but a relative
> > > > frequency among the pages of the target workloads.
> > > > 
> > > > Users can control the resolution of the reports by setting two time
> > > > intervals, ``sampling interval`` and ``aggregation interval``.  In
> > > > detail, DAMON checks access to each page per ``sampling interval``,
> > > > aggregates the results (counts the number of the accesses to each page),
> > > > and reports the aggregated results per ``aggregation interval``.  For
> > > > the access check of each page, DAMON uses the Accessed bits of PTEs.
> > > > 
> > > > This is thus similar to common periodic access checks based access
> > > > tracking mechanisms, which overhead is increasing as the size of the
> > > > target process grows.
> > > > 
> > > > Region Based Sampling
> > > > ---------------------
> > > > 
> > > > To avoid the unbounded increase of the overhead, DAMON groups a number
> > > > of adjacent pages that assumed to have same access frequencies into a
> > > > region.  As long as the assumption (pages in a region have same access
> > > > frequencies) is kept, only one page in the region is required to be
> > > > checked.  Thus, for each ``sampling interval``, DAMON randomly picks one
> > > > page in each region and clears its Accessed bit.  After one more
> > > > ``sampling interval``, DAMON reads the Accessed bit of the page and
> > > > increases the access frequency of the region if the bit has set
> > > > meanwhile.  Therefore, the monitoring overhead is controllable by
> > > > setting the number of regions.
> > > > 
> > > > Nonetheless, this scheme cannot preserve the quality of the output if
> > > > the assumption is not kept.  Following commit will introduce how we can
> > > > make the guarantee with best effort.
> > > > 
> > > > Signed-off-by: SeongJae Park <sjpark@amazon.de>  
> > > 
> > > Hi.
> > > 
> > > A few comments inline.
> > > 
> > > I've still not replicated your benchmarks so may well have some more
> > > feedback once I've managed that on one of our servers.  
> > 
> > Appreciate your comments.  If you need any help for the replication, please let
> > me know.  I basically use my parsec3 wrapper scripts[1] to run parsec3 and
> > splash2x workloads and `damo` tool, which resides in the kernel tree at
> > `/tools/damon/`.
> > 
> > For example, below commands will reproduce ethp applied splash2x/fft run.
> >     
> >     $ echo "2M      null    5       null    null    null    hugepage
> >     2M      null    null    5       1s      null    nohugepage" > ethp
> >     $ parsec3_on_ubuntu/run.sh splash2x.fft
> >     $ linux/tools/damon/damo schemes -c ethp `pidof fft`
> > 
> > [1] https://github.com/sjp38/parsec3_on_ubuntu
> >   
> > > 
> > > Thanks,
> > > 
> > > Jonathan
> > >   
> > > > ---
> > > >  include/linux/damon.h |  24 ++
> > > >  mm/damon.c            | 553 ++++++++++++++++++++++++++++++++++++++++++
> > > >  2 files changed, 577 insertions(+)
> > > >   
> > [...]  
> > > > diff --git a/mm/damon.c b/mm/damon.c
> > > > index d7e6226ab7f1..018016793555 100644
> > > > --- a/mm/damon.c
> > > > +++ b/mm/damon.c
> > > > @@ -10,8 +10,14 @@
> > > >  #define pr_fmt(fmt) "damon: " fmt
> > > >  
> > > >  #include <linux/damon.h>
> > > > +#include <linux/delay.h>
> > > > +#include <linux/kthread.h>
> > > >  #include <linux/mm.h>
> > > >  #include <linux/module.h>
> > > > +#include <linux/page_idle.h>
> > > > +#include <linux/random.h>
> > > > +#include <linux/sched/mm.h>
> > > > +#include <linux/sched/task.h>
> > > >  #include <linux/slab.h>
> > > >    
> > [...]  
> > > > +/*
> > > > + * Size-evenly split a region into 'nr_pieces' small regions
> > > > + *
> > > > + * Returns 0 on success, or negative error code otherwise.
> > > > + */
> > > > +static int damon_split_region_evenly(struct damon_ctx *ctx,
> > > > +		struct damon_region *r, unsigned int nr_pieces)
> > > > +{
> > > > +	unsigned long sz_orig, sz_piece, orig_end;
> > > > +	struct damon_region *piece = NULL, *next;
> > > > +	unsigned long start;
> > > > +
> > > > +	if (!r || !nr_pieces)
> > > > +		return -EINVAL;
> > > > +
> > > > +	orig_end = r->vm_end;
> > > > +	sz_orig = r->vm_end - r->vm_start;
> > > > +	sz_piece = sz_orig / nr_pieces;
> > > > +
> > > > +	if (!sz_piece)
> > > > +		return -EINVAL;
> > > > +
> > > > +	r->vm_end = r->vm_start + sz_piece;
> > > > +	next = damon_next_region(r);
> > > > +	for (start = r->vm_end; start + sz_piece <= orig_end;
> > > > +			start += sz_piece) {
> > > > +		piece = damon_new_region(ctx, start, start + sz_piece);  
> > > piece may be n  
> > 
> > Yes, that name is short and more intuitive.  I will rename so.
> >   
> > > > +		damon_insert_region(piece, r, next);
> > > > +		r = piece;
> > > > +	}
> > > > +	/* complement last region for possible rounding error */
> > > > +	if (piece)
> > > > +		piece->vm_end = orig_end;  
> > > 
> > > Update the sampling address to ensure it's in the region?  
> > 
> > I think `piece->vm_end` should be equal or smaller than `orig_end` and
> > therefore the sampling address of `piece` will be still in the region.
> >   
> > >   
> > > > +
> > > > +	return 0;
> > > > +}
> > > > +  
> > [...]  
> > > > +static void damon_pte_pmd_mkold(pte_t *pte, pmd_t *pmd)
> > > > +{
> > > > +	if (pte) {
> > > > +		if (pte_young(*pte)) {
> > > > +			clear_page_idle(pte_page(*pte));
> > > > +			set_page_young(pte_page(*pte));
> > > > +		}
> > > > +		*pte = pte_mkold(*pte);
> > > > +		return;
> > > > +	}
> > > > +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > > > +	if (pmd) {
> > > > +		if (pmd_young(*pmd)) {
> > > > +			clear_page_idle(pmd_page(*pmd));
> > > > +			set_page_young(pmd_page(*pmd));
> > > > +		}
> > > > +		*pmd = pmd_mkold(*pmd);
> > > > +	}
> > > > +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */  
> > > 
> > > No need to flush the TLBs?  
> > 
> > Good point!
> > 
> > I have intentionally skipped TLB flushing here to minimize the performance
> > effect to the target workload.  I also thought this might not degrade the
> > monitoring accuracy so much because we are targetting for the DRAM level
> > accesses of memory-intensive workloads, which might make TLB flood frequently.
> > 
> > However, your comment makes me thinking differently now.  By flushing the TLB
> > here, we will increase up to `number_of_regions` TLB misses for sampling
> > interval.  This might be not a huge overhead.  Also, improving the monitoring
> > accuracy makes no harm at all.  I even didn't measured the overhead.
> > 
> > I will test the overhead and if it is not significant, I will make this code to
> > flush TLB, in the next spin.  
> 
> Hmm, it seems like 'page_idle.c' is also modifying the Accessed bit but doesn't
> flush related TLB entries.  If I'm not missing something here, I would like to
> leave this part as is to make the behavior consistent.

Interesting.  In that usecase, the risk is that the MMU believes
the page still has the accessed bit set when we have cleared it and hence
the accessed bit is not written out to the table in memory.

That will give them a wrong decision so not great and would lead to them
thinking more pages are idle than are.

Here we could have a particular TLB entry for a huge page in which
a region lies entirely.  Because we don't flush the TLB each time
we could end with a count of 0 accesses when it should be the maximum.
A very frequently accessed page might well sit in the TLB for a very
long time (particularly if the TLB is running a clever eviction
strategy).

I think we would want to be test this and see if we get that
pathological case sometimes.  Also worth benchmarking if it actually
costs us very much to do the flushes.

Jonathan

> 
> 
> Thanks,
> SeongJae Park
> 
> >   
> > >   
> > > > +}
> > > > +  
> > [...]  
> > > > +/*
> > > > + * The monitoring daemon that runs as a kernel thread
> > > > + */
> > > > +static int kdamond_fn(void *data)
> > > > +{
> > > > +	struct damon_ctx *ctx = data;
> > > > +	struct damon_task *t;
> > > > +	struct damon_region *r, *next;
> > > > +	struct mm_struct *mm;
> > > > +
> > > > +	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
> > > > +	kdamond_init_regions(ctx);  
> > > 
> > > We haven't called mkold on the initial regions so first check will
> > > get us fairly random state.  
> > 
> > Yes, indeed.  However, the early results will not be accurate anyway because
> > the adaptive regions adjustment algorithm will not take effect yet.  I would
> > like to leave this part as is but add some comments about this point to keep
> > the code simple.
> >   
> > >   
> > > > +	while (!kdamond_need_stop(ctx)) {
> > > > +		damon_for_each_task(ctx, t) {
> > > > +			mm = damon_get_mm(t);
> > > > +			if (!mm)
> > > > +				continue;
> > > > +			damon_for_each_region(r, t)
> > > > +				kdamond_check_access(ctx, mm, r);
> > > > +			mmput(mm);
> > > > +		}
> > > > +
> > > > +		if (kdamond_aggregate_interval_passed(ctx))
> > > > +			kdamond_reset_aggregated(ctx);
> > > > +
> > > > +		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
> > > > +	}
> > > > +	damon_for_each_task(ctx, t) {
> > > > +		damon_for_each_region_safe(r, next, t)
> > > > +			damon_destroy_region(r);
> > > > +	}
> > > > +	pr_debug("kdamond (%d) finishes\n", ctx->kdamond->pid);
> > > > +	mutex_lock(&ctx->kdamond_lock);
> > > > +	ctx->kdamond = NULL;
> > > > +	mutex_unlock(&ctx->kdamond_lock);
> > > > +
> > > > +	return 0;
> > > > +}
> > > > +  
> > [...]  
> > > > +/*
> > > > + * Start or stop the kdamond
> > > > + *
> > > > + * Returns 0 if success, negative error code otherwise.
> > > > + */
> > > > +static int damon_turn_kdamond(struct damon_ctx *ctx, bool on)
> > > > +{
> > > > +	int err = -EBUSY;
> > > > +
> > > > +	mutex_lock(&ctx->kdamond_lock);
> > > > +	if (!ctx->kdamond && on) {  
> > > 
> > > Given there is very little shared code between on and off, I would
> > > suggest just splitting it into two functions.  
> > 
> > Good point, I will do so in next spin.
> >   
> > >   
> > > > +		err = 0;
> > > > +		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond");
> > > > +		if (IS_ERR(ctx->kdamond))
> > > > +			err = PTR_ERR(ctx->kdamond);
> > > > +	} else if (ctx->kdamond && !on) {
> > > > +		mutex_unlock(&ctx->kdamond_lock);
> > > > +		kthread_stop(ctx->kdamond);
> > > > +		while (damon_kdamond_running(ctx))
> > > > +			usleep_range(ctx->sample_interval,
> > > > +					ctx->sample_interval * 2);
> > > > +		return 0;
> > > > +	}
> > > > +	mutex_unlock(&ctx->kdamond_lock);
> > > > +
> > > > +	return err;
> > > > +}
> > > > +  
> > [...]  
> > > > +
> > > > +/*  
> > > 
> > > Why not make these actual kernel-doc?  That way you can use the
> > > kernel-doc scripts to sanity check them.  
> > 
> > Oops, I just forgot that it should start with '/**'.  Will fix it in next spin.
> > 
> > 
> > Thanks,
> > SeongJae Park
> >   
> > > 
> > > /**
> > >   
> > > > + * damon_set_attrs() - Set attributes for the monitoring.
> > > > + * @ctx:		monitoring context
> > > > + * @sample_int:		time interval between samplings
> > > > + * @aggr_int:		time interval between aggregations
> > > > + * @min_nr_reg:		minimal number of regions
> > > > + *
> > > > + * This function should not be called while the kdamond is running.
> > > > + * Every time interval is in micro-seconds.
> > > > + *
> > > > + * Return: 0 on success, negative error code otherwise.
> > > > + */
> > > > +int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> > > > +		unsigned long aggr_int, unsigned long min_nr_reg)
> > > > +{
> > > > +	if (min_nr_reg < 3) {
> > > > +		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
> > > > +				min_nr_reg);
> > > > +		return -EINVAL;
> > > > +	}
> > > > +
> > > > +	ctx->sample_interval = sample_int;
> > > > +	ctx->aggr_interval = aggr_int;
> > > > +	ctx->min_nr_regions = min_nr_reg;
> > > > +
> > > > +	return 0;
> > > > +}
> > > > +
> > > >  static int __init damon_init(void)
> > > >  {
> > > >  	return 0;  
> > >   



^ permalink raw reply	[flat|nested] 28+ messages in thread

* Re: Re: [PATCH v7 04/15] mm/damon: Implement region based sampling
  2020-04-02 17:24         ` Jonathan Cameron
@ 2020-04-02 18:00           ` SeongJae Park
  2020-04-03 10:30             ` SeongJae Park
  0 siblings, 1 reply; 28+ messages in thread
From: SeongJae Park @ 2020-04-02 18:00 UTC (permalink / raw)
  To: Jonathan Cameron
  Cc: SeongJae Park, akpm, SeongJae Park, aarcange, acme,
	alexander.shishkin, amit, brendan.d.gregg, brendanhiggins, cai,
	colin.king, corbet, dwmw, jolsa, kirill, mark.rutland, mgorman,
	minchan, mingo, namhyung, peterz, rdunlap, riel, rientjes,
	rostedt, shuah, sj38.park, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

On Thu, 2 Apr 2020 18:24:01 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:

> On Thu, 2 Apr 2020 15:59:59 +0200
> SeongJae Park <sjpark@amazon.com> wrote:
> 
> > On Wed, 1 Apr 2020 10:22:22 +0200 SeongJae Park <sjpark@amazon.com> wrote:
> > 
> > > On Tue, 31 Mar 2020 17:02:33 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:
> > >   
> > > > On Wed, 18 Mar 2020 12:27:11 +0100
> > > > SeongJae Park <sjpark@amazon.com> wrote:
> > > >   
> > > > > From: SeongJae Park <sjpark@amazon.de>
> > > > > 
> > > > > This commit implements DAMON's basic access check and region based
> > > > > sampling mechanisms.  This change would seems make no sense, mainly
> > > > > because it is only a part of the DAMON's logics.  Following two commits
> > > > > will make more sense.
> > > > > 
> > > > > Basic Access Check
> > > > > ------------------
> > > > > 
> > > > > DAMON basically reports what pages are how frequently accessed.  Note
> > > > > that the frequency is not an absolute number of accesses, but a relative
> > > > > frequency among the pages of the target workloads.
> > > > > 
> > > > > Users can control the resolution of the reports by setting two time
> > > > > intervals, ``sampling interval`` and ``aggregation interval``.  In
> > > > > detail, DAMON checks access to each page per ``sampling interval``,
> > > > > aggregates the results (counts the number of the accesses to each page),
> > > > > and reports the aggregated results per ``aggregation interval``.  For
> > > > > the access check of each page, DAMON uses the Accessed bits of PTEs.
> > > > > 
> > > > > This is thus similar to common periodic access checks based access
> > > > > tracking mechanisms, which overhead is increasing as the size of the
> > > > > target process grows.
> > > > > 
> > > > > Region Based Sampling
> > > > > ---------------------
> > > > > 
> > > > > To avoid the unbounded increase of the overhead, DAMON groups a number
> > > > > of adjacent pages that assumed to have same access frequencies into a
> > > > > region.  As long as the assumption (pages in a region have same access
> > > > > frequencies) is kept, only one page in the region is required to be
> > > > > checked.  Thus, for each ``sampling interval``, DAMON randomly picks one
> > > > > page in each region and clears its Accessed bit.  After one more
> > > > > ``sampling interval``, DAMON reads the Accessed bit of the page and
> > > > > increases the access frequency of the region if the bit has set
> > > > > meanwhile.  Therefore, the monitoring overhead is controllable by
> > > > > setting the number of regions.
> > > > > 
> > > > > Nonetheless, this scheme cannot preserve the quality of the output if
> > > > > the assumption is not kept.  Following commit will introduce how we can
> > > > > make the guarantee with best effort.
> > > > > 
> > > > > Signed-off-by: SeongJae Park <sjpark@amazon.de>  
> > > > 
> > > > Hi.
> > > > 
> > > > A few comments inline.
> > > > 
> > > > I've still not replicated your benchmarks so may well have some more
> > > > feedback once I've managed that on one of our servers.  
> > > 
> > > Appreciate your comments.  If you need any help for the replication, please let
> > > me know.  I basically use my parsec3 wrapper scripts[1] to run parsec3 and
> > > splash2x workloads and `damo` tool, which resides in the kernel tree at
> > > `/tools/damon/`.
> > > 
> > > For example, below commands will reproduce ethp applied splash2x/fft run.
> > >     
> > >     $ echo "2M      null    5       null    null    null    hugepage
> > >     2M      null    null    5       1s      null    nohugepage" > ethp
> > >     $ parsec3_on_ubuntu/run.sh splash2x.fft
> > >     $ linux/tools/damon/damo schemes -c ethp `pidof fft`
> > > 
> > > [1] https://github.com/sjp38/parsec3_on_ubuntu
> > >   
> > > > 
> > > > Thanks,
> > > > 
> > > > Jonathan
> > > >   
> > > > > ---
> > > > >  include/linux/damon.h |  24 ++
> > > > >  mm/damon.c            | 553 ++++++++++++++++++++++++++++++++++++++++++
> > > > >  2 files changed, 577 insertions(+)
> > > > >   
> > > [...]  
> > > > > diff --git a/mm/damon.c b/mm/damon.c
> > > > > index d7e6226ab7f1..018016793555 100644
> > > > > --- a/mm/damon.c
> > > > > +++ b/mm/damon.c
> > > > > @@ -10,8 +10,14 @@
> > > > >  #define pr_fmt(fmt) "damon: " fmt
> > > > >  
> > > > >  #include <linux/damon.h>
> > > > > +#include <linux/delay.h>
> > > > > +#include <linux/kthread.h>
> > > > >  #include <linux/mm.h>
> > > > >  #include <linux/module.h>
> > > > > +#include <linux/page_idle.h>
> > > > > +#include <linux/random.h>
> > > > > +#include <linux/sched/mm.h>
> > > > > +#include <linux/sched/task.h>
> > > > >  #include <linux/slab.h>
> > > > >    
> > > [...]  
> > > > > +/*
> > > > > + * Size-evenly split a region into 'nr_pieces' small regions
> > > > > + *
> > > > > + * Returns 0 on success, or negative error code otherwise.
> > > > > + */
> > > > > +static int damon_split_region_evenly(struct damon_ctx *ctx,
> > > > > +		struct damon_region *r, unsigned int nr_pieces)
> > > > > +{
> > > > > +	unsigned long sz_orig, sz_piece, orig_end;
> > > > > +	struct damon_region *piece = NULL, *next;
> > > > > +	unsigned long start;
> > > > > +
> > > > > +	if (!r || !nr_pieces)
> > > > > +		return -EINVAL;
> > > > > +
> > > > > +	orig_end = r->vm_end;
> > > > > +	sz_orig = r->vm_end - r->vm_start;
> > > > > +	sz_piece = sz_orig / nr_pieces;
> > > > > +
> > > > > +	if (!sz_piece)
> > > > > +		return -EINVAL;
> > > > > +
> > > > > +	r->vm_end = r->vm_start + sz_piece;
> > > > > +	next = damon_next_region(r);
> > > > > +	for (start = r->vm_end; start + sz_piece <= orig_end;
> > > > > +			start += sz_piece) {
> > > > > +		piece = damon_new_region(ctx, start, start + sz_piece);  
> > > > piece may be n  
> > > 
> > > Yes, that name is short and more intuitive.  I will rename so.
> > >   
> > > > > +		damon_insert_region(piece, r, next);
> > > > > +		r = piece;
> > > > > +	}
> > > > > +	/* complement last region for possible rounding error */
> > > > > +	if (piece)
> > > > > +		piece->vm_end = orig_end;  
> > > > 
> > > > Update the sampling address to ensure it's in the region?  
> > > 
> > > I think `piece->vm_end` should be equal or smaller than `orig_end` and
> > > therefore the sampling address of `piece` will be still in the region.
> > >   
> > > >   
> > > > > +
> > > > > +	return 0;
> > > > > +}
> > > > > +  
> > > [...]  
> > > > > +static void damon_pte_pmd_mkold(pte_t *pte, pmd_t *pmd)
> > > > > +{
> > > > > +	if (pte) {
> > > > > +		if (pte_young(*pte)) {
> > > > > +			clear_page_idle(pte_page(*pte));
> > > > > +			set_page_young(pte_page(*pte));
> > > > > +		}
> > > > > +		*pte = pte_mkold(*pte);
> > > > > +		return;
> > > > > +	}
> > > > > +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > > > > +	if (pmd) {
> > > > > +		if (pmd_young(*pmd)) {
> > > > > +			clear_page_idle(pmd_page(*pmd));
> > > > > +			set_page_young(pmd_page(*pmd));
> > > > > +		}
> > > > > +		*pmd = pmd_mkold(*pmd);
> > > > > +	}
> > > > > +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */  
> > > > 
> > > > No need to flush the TLBs?  
> > > 
> > > Good point!
> > > 
> > > I have intentionally skipped TLB flushing here to minimize the performance
> > > effect to the target workload.  I also thought this might not degrade the
> > > monitoring accuracy so much because we are targetting for the DRAM level
> > > accesses of memory-intensive workloads, which might make TLB flood frequently.
> > > 
> > > However, your comment makes me thinking differently now.  By flushing the TLB
> > > here, we will increase up to `number_of_regions` TLB misses for sampling
> > > interval.  This might be not a huge overhead.  Also, improving the monitoring
> > > accuracy makes no harm at all.  I even didn't measured the overhead.
> > > 
> > > I will test the overhead and if it is not significant, I will make this code to
> > > flush TLB, in the next spin.  
> > 
> > Hmm, it seems like 'page_idle.c' is also modifying the Accessed bit but doesn't
> > flush related TLB entries.  If I'm not missing something here, I would like to
> > leave this part as is to make the behavior consistent.
> 
> Interesting.  In that usecase, the risk is that the MMU believes
> the page still has the accessed bit set when we have cleared it and hence
> the accessed bit is not written out to the table in memory.
> 
> That will give them a wrong decision so not great and would lead to them
> thinking more pages are idle than are.
> 
> Here we could have a particular TLB entry for a huge page in which
> a region lies entirely.  Because we don't flush the TLB each time
> we could end with a count of 0 accesses when it should be the maximum.
> A very frequently accessed page might well sit in the TLB for a very
> long time (particularly if the TLB is running a clever eviction
> strategy).
> 
> I think we would want to be test this and see if we get that
> pathological case sometimes.  Also worth benchmarking if it actually
> costs us very much to do the flushes.

Agreed, it wouldn't be late to make a decision after measuring the real cost.
I will share the measurement results soon.  Meanwhile, it would be helpful if
anyone could confirm whether page_idle.c is skipping TLB flushing and explain
why such decision has made.


Thanks,
SeongJae Park

> 
> Jonathan
> 
> > 
> > 
> > Thanks,
> > SeongJae Park
> > 
> > >   
> > > >   
> > > > > +}
> > > > > +  
> > > [...]  
> > > > > +/*
> > > > > + * The monitoring daemon that runs as a kernel thread
> > > > > + */
> > > > > +static int kdamond_fn(void *data)
> > > > > +{
> > > > > +	struct damon_ctx *ctx = data;
> > > > > +	struct damon_task *t;
> > > > > +	struct damon_region *r, *next;
> > > > > +	struct mm_struct *mm;
> > > > > +
> > > > > +	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
> > > > > +	kdamond_init_regions(ctx);  
> > > > 
> > > > We haven't called mkold on the initial regions so first check will
> > > > get us fairly random state.  
> > > 
> > > Yes, indeed.  However, the early results will not be accurate anyway because
> > > the adaptive regions adjustment algorithm will not take effect yet.  I would
> > > like to leave this part as is but add some comments about this point to keep
> > > the code simple.
> > >   
> > > >   
> > > > > +	while (!kdamond_need_stop(ctx)) {
> > > > > +		damon_for_each_task(ctx, t) {
> > > > > +			mm = damon_get_mm(t);
> > > > > +			if (!mm)
> > > > > +				continue;
> > > > > +			damon_for_each_region(r, t)
> > > > > +				kdamond_check_access(ctx, mm, r);
> > > > > +			mmput(mm);
> > > > > +		}
> > > > > +
> > > > > +		if (kdamond_aggregate_interval_passed(ctx))
> > > > > +			kdamond_reset_aggregated(ctx);
> > > > > +
> > > > > +		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
> > > > > +	}
> > > > > +	damon_for_each_task(ctx, t) {
> > > > > +		damon_for_each_region_safe(r, next, t)
> > > > > +			damon_destroy_region(r);
> > > > > +	}
> > > > > +	pr_debug("kdamond (%d) finishes\n", ctx->kdamond->pid);
> > > > > +	mutex_lock(&ctx->kdamond_lock);
> > > > > +	ctx->kdamond = NULL;
> > > > > +	mutex_unlock(&ctx->kdamond_lock);
> > > > > +
> > > > > +	return 0;
> > > > > +}
> > > > > +  
> > > [...]  
> > > > > +/*
> > > > > + * Start or stop the kdamond
> > > > > + *
> > > > > + * Returns 0 if success, negative error code otherwise.
> > > > > + */
> > > > > +static int damon_turn_kdamond(struct damon_ctx *ctx, bool on)
> > > > > +{
> > > > > +	int err = -EBUSY;
> > > > > +
> > > > > +	mutex_lock(&ctx->kdamond_lock);
> > > > > +	if (!ctx->kdamond && on) {  
> > > > 
> > > > Given there is very little shared code between on and off, I would
> > > > suggest just splitting it into two functions.  
> > > 
> > > Good point, I will do so in next spin.
> > >   
> > > >   
> > > > > +		err = 0;
> > > > > +		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond");
> > > > > +		if (IS_ERR(ctx->kdamond))
> > > > > +			err = PTR_ERR(ctx->kdamond);
> > > > > +	} else if (ctx->kdamond && !on) {
> > > > > +		mutex_unlock(&ctx->kdamond_lock);
> > > > > +		kthread_stop(ctx->kdamond);
> > > > > +		while (damon_kdamond_running(ctx))
> > > > > +			usleep_range(ctx->sample_interval,
> > > > > +					ctx->sample_interval * 2);
> > > > > +		return 0;
> > > > > +	}
> > > > > +	mutex_unlock(&ctx->kdamond_lock);
> > > > > +
> > > > > +	return err;
> > > > > +}
> > > > > +  
> > > [...]  
> > > > > +
> > > > > +/*  
> > > > 
> > > > Why not make these actual kernel-doc?  That way you can use the
> > > > kernel-doc scripts to sanity check them.  
> > > 
> > > Oops, I just forgot that it should start with '/**'.  Will fix it in next spin.
> > > 
> > > 
> > > Thanks,
> > > SeongJae Park
> > >   
> > > > 
> > > > /**
> > > >   
> > > > > + * damon_set_attrs() - Set attributes for the monitoring.
> > > > > + * @ctx:		monitoring context
> > > > > + * @sample_int:		time interval between samplings
> > > > > + * @aggr_int:		time interval between aggregations
> > > > > + * @min_nr_reg:		minimal number of regions
> > > > > + *
> > > > > + * This function should not be called while the kdamond is running.
> > > > > + * Every time interval is in micro-seconds.
> > > > > + *
> > > > > + * Return: 0 on success, negative error code otherwise.
> > > > > + */
> > > > > +int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
> > > > > +		unsigned long aggr_int, unsigned long min_nr_reg)
> > > > > +{
> > > > > +	if (min_nr_reg < 3) {
> > > > > +		pr_err("min_nr_regions (%lu) should be bigger than 2\n",
> > > > > +				min_nr_reg);
> > > > > +		return -EINVAL;
> > > > > +	}
> > > > > +
> > > > > +	ctx->sample_interval = sample_int;
> > > > > +	ctx->aggr_interval = aggr_int;
> > > > > +	ctx->min_nr_regions = min_nr_reg;
> > > > > +
> > > > > +	return 0;
> > > > > +}
> > > > > +
> > > > >  static int __init damon_init(void)
> > > > >  {
> > > > >  	return 0;  
> > > >   
> 
> 

^ permalink raw reply	[flat|nested] 28+ messages in thread

* Re: Re: Re: [PATCH v7 04/15] mm/damon: Implement region based sampling
  2020-04-02 18:00           ` SeongJae Park
@ 2020-04-03 10:30             ` SeongJae Park
  0 siblings, 0 replies; 28+ messages in thread
From: SeongJae Park @ 2020-04-03 10:30 UTC (permalink / raw)
  To: SeongJae Park
  Cc: Jonathan Cameron, SeongJae Park, akpm, SeongJae Park, aarcange,
	acme, alexander.shishkin, amit, brendan.d.gregg, brendanhiggins,
	cai, colin.king, corbet, dwmw, jolsa, kirill, mark.rutland,
	mgorman, minchan, mingo, namhyung, peterz, rdunlap, riel,
	rientjes, rostedt, shuah, vbabka, vdavydov.dev, yang.shi,
	ying.huang, linux-mm, linux-doc, linux-kernel

On Thu,  2 Apr 2020 20:00:22 +0200 SeongJae Park <sj38.park@gmail.com> wrote:

> On Thu, 2 Apr 2020 18:24:01 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:
> 
> > On Thu, 2 Apr 2020 15:59:59 +0200
> > SeongJae Park <sjpark@amazon.com> wrote:
> > 
> > > On Wed, 1 Apr 2020 10:22:22 +0200 SeongJae Park <sjpark@amazon.com> wrote:
> > > 
> > > > On Tue, 31 Mar 2020 17:02:33 +0100 Jonathan Cameron <Jonathan.Cameron@Huawei.com> wrote:
> > > >   
> > > > > On Wed, 18 Mar 2020 12:27:11 +0100
> > > > > SeongJae Park <sjpark@amazon.com> wrote:
> > > > >   
> > > > > > From: SeongJae Park <sjpark@amazon.de>
[...]
> > > > > > +static void damon_pte_pmd_mkold(pte_t *pte, pmd_t *pmd)
> > > > > > +{
> > > > > > +	if (pte) {
> > > > > > +		if (pte_young(*pte)) {
> > > > > > +			clear_page_idle(pte_page(*pte));
> > > > > > +			set_page_young(pte_page(*pte));
> > > > > > +		}
> > > > > > +		*pte = pte_mkold(*pte);
> > > > > > +		return;
> > > > > > +	}
> > > > > > +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > > > > > +	if (pmd) {
> > > > > > +		if (pmd_young(*pmd)) {
> > > > > > +			clear_page_idle(pmd_page(*pmd));
> > > > > > +			set_page_young(pmd_page(*pmd));
> > > > > > +		}
> > > > > > +		*pmd = pmd_mkold(*pmd);
> > > > > > +	}
> > > > > > +#endif /* CONFIG_TRANSPARENT_HUGEPAGE */  
> > > > > 
> > > > > No need to flush the TLBs?  
> > > > 
> > > > Good point!
> > > > 
> > > > I have intentionally skipped TLB flushing here to minimize the performance
> > > > effect to the target workload.  I also thought this might not degrade the
> > > > monitoring accuracy so much because we are targetting for the DRAM level
> > > > accesses of memory-intensive workloads, which might make TLB flood frequently.
> > > > 
> > > > However, your comment makes me thinking differently now.  By flushing the TLB
> > > > here, we will increase up to `number_of_regions` TLB misses for sampling
> > > > interval.  This might be not a huge overhead.  Also, improving the monitoring
> > > > accuracy makes no harm at all.  I even didn't measured the overhead.
> > > > 
> > > > I will test the overhead and if it is not significant, I will make this code to
> > > > flush TLB, in the next spin.  
> > > 
> > > Hmm, it seems like 'page_idle.c' is also modifying the Accessed bit but doesn't
> > > flush related TLB entries.  If I'm not missing something here, I would like to
> > > leave this part as is to make the behavior consistent.
> > 
> > Interesting.  In that usecase, the risk is that the MMU believes
> > the page still has the accessed bit set when we have cleared it and hence
> > the accessed bit is not written out to the table in memory.
> > 
> > That will give them a wrong decision so not great and would lead to them
> > thinking more pages are idle than are.
> > 
> > Here we could have a particular TLB entry for a huge page in which
> > a region lies entirely.  Because we don't flush the TLB each time
> > we could end with a count of 0 accesses when it should be the maximum.
> > A very frequently accessed page might well sit in the TLB for a very
> > long time (particularly if the TLB is running a clever eviction
> > strategy).
> > 
> > I think we would want to be test this and see if we get that
> > pathological case sometimes.  Also worth benchmarking if it actually
> > costs us very much to do the flushes.
> 
> Agreed, it wouldn't be late to make a decision after measuring the real cost.
> I will share the measurement results soon.  Meanwhile, it would be helpful if
> anyone could confirm whether page_idle.c is skipping TLB flushing and explain
> why such decision has made.

I just finished implementing TLB flushing in straightforward way (the diff for
this change is at the bottom of this mail) on the version I applied your
recommended changes and my one bug fix (setting 'last_accessed' false).

It shows monitoring accuracy improvement as expected, though it is not so big.
I compared the visualized access patterns of each test workload to check this.
There is no big difference, but the visualized pattern of TLB flushing version
seems a little bit more clear to my human eye.

However, the overhead is clear and significant to some workloads including
parsec3/streamcluster, splash2x/barnes, splash2x/lu_ncb and splash2x/volrend.
The CPU utilization of kdamond, the deamon monitoring the access pattern, never
exceeds 2% of single CPU time for the 4 workloads before the change is applied,
but it frequently exceeds 30% of single CPU time with the TLB flushing.  The
runtimes of the monitoring target workloads also increased.  In case of
parsec3/streamcluster, the TLB flushing changed its runtime from 320 seconds to
470 seconds.

So, it seems the benefit of TLB flushing is smaller than the cost in some
cases.  Thus, I think enabling TLB flushes by default wouldn't be a good
decision.  Rather than that, it could be better to allow users to manually do
TLB flushing for their specific workloads.  This will be easily available by
using the DAMON's sampling callback functions.  Also, I am planning to let
users to configure DAMON with their own access check / sampling setup functions
in future, to support not only virtual memory but also physical memory and some
other special cases.

If I'm missing something or you have other thinking, please let me know.


Thanks,
SeongJae Park

=============================== >8 ============================================

Below is the essential diff for the TLB flushing I implemented.  To double
check the overhead is really from the TLB flushing, I also measured the
overhead of the additional works including the vma searching by commenting out
the 'flush_tlb_range()' call.  After commenting out it, the CPU utilization of
kdamond and runtime of target workloads has restored back.  So, the overhead
seems really came from the TLB flushing.

@@ -408,6 +411,9 @@ static void kdamond_prepare_sampling(struct damon_ctx *ctx,
        pte_t *pte = NULL;
        pmd_t *pmd = NULL;
        spinlock_t *ptl;
+       struct vm_area_struct *vma;
+       unsigned long tlb_start;
+       unsigned long tlb_size = PAGE_SIZE;

        r->sampling_addr = damon_rand(ctx, r->vm_start, r->vm_end);

@@ -420,18 +426,29 @@ static void kdamond_prepare_sampling(struct damon_ctx *ctx,
                        set_page_young(pte_page(*pte));
                }
                *pte = pte_mkold(*pte);
-               return;
        }
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
-       if (pmd) {
+       else if (pmd) {
                if (pmd_young(*pmd)) {
                        clear_page_idle(pmd_page(*pmd));
                        set_page_young(pmd_page(*pmd));
                }
                *pmd = pmd_mkold(*pmd);
+               tlb_size = ((1UL) << HPAGE_PMD_SHIFT);
        }
 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
        spin_unlock(ptl);
+
+       tlb_start = ALIGN(r->sampling_addr, tlb_size);
+
+       down_read(&mm->mmap_sem);
+       vma = find_vma(mm, r->sampling_addr);
+       if (!vma || (r->sampling_addr < vma->vm_start))
+               goto out;
+       flush_tlb_range(vma, tlb_start, tlb_start + tlb_size);
+
+out:
+       up_read(&mm->mmap_sem);
 }


Thanks,
SeongJae Park

^ permalink raw reply	[flat|nested] 28+ messages in thread

end of thread, other threads:[~2020-04-03 10:31 UTC | newest]

Thread overview: 28+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-03-18 11:27 [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park
2020-03-18 11:27 ` [PATCH v7 01/15] scripts/spelling: Add a few more typos SeongJae Park
2020-03-18 11:27 ` [PATCH v7 02/15] mm/page_ext: Export lookup_page_ext() to GPL modules SeongJae Park
2020-03-18 11:27 ` [PATCH v7 03/15] mm: Introduce Data Access MONitor (DAMON) SeongJae Park
2020-03-18 11:27 ` [PATCH v7 04/15] mm/damon: Implement region based sampling SeongJae Park
2020-03-31 16:02   ` Jonathan Cameron
2020-04-01  8:22     ` SeongJae Park
2020-04-01 14:24       ` Jonathan Cameron
2020-04-01 14:46         ` SeongJae Park
2020-04-02 13:59       ` SeongJae Park
2020-04-02 17:24         ` Jonathan Cameron
2020-04-02 18:00           ` SeongJae Park
2020-04-03 10:30             ` SeongJae Park
2020-04-02 15:07   ` SeongJae Park
2020-03-18 11:27 ` [PATCH v7 05/15] mm/damon: Adaptively adjust regions SeongJae Park
2020-03-31 16:08   ` Jonathan Cameron
2020-04-01  8:22     ` SeongJae Park
2020-03-18 11:27 ` [PATCH v7 06/15] mm/damon: Apply dynamic memory mapping changes SeongJae Park
2020-03-18 11:27 ` [PATCH v7 07/15] mm/damon: Implement callbacks SeongJae Park
2020-03-18 11:27 ` [PATCH v7 08/15] mm/damon: Implement access pattern recording SeongJae Park
2020-03-18 11:27 ` [PATCH v7 09/15] mm/damon: Add debugfs interface SeongJae Park
2020-03-18 11:27 ` [PATCH v7 10/15] mm/damon: Add tracepoints SeongJae Park
2020-03-18 11:27 ` [PATCH v7 11/15] tools: Add a minimal user-space tool for DAMON SeongJae Park
2020-03-18 11:27 ` [PATCH v7 12/15] Documentation/admin-guide/mm: Add a document " SeongJae Park
2020-03-18 11:27 ` [PATCH v7 13/15] mm/damon: Add kunit tests SeongJae Park
2020-03-18 11:38 ` [PATCH v7 14/15] mm/damon: Add user space selftests SeongJae Park
2020-03-18 11:39 ` [PATCH v7 15/15] MAINTAINERS: Update for DAMON SeongJae Park
2020-03-30 10:53 ` [PATCH v7 00/15] Introduce Data Access MONitor (DAMON) SeongJae Park

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).