mm: Throttle allocators when failing reclaim over memory.high
diff mbox series

Message ID 20190201011352.GA14370@chrisdown.name
State Superseded
Headers show
Series
  • mm: Throttle allocators when failing reclaim over memory.high
Related show

Commit Message

Chris Down Feb. 1, 2019, 1:13 a.m. UTC
We're trying to use memory.high to limit workloads, but have found that
containment can frequently fail completely and cause OOM situations
outside of the cgroup. This happens especially with swap space -- either
when none is configured, or swap is full. These failures often also
don't have enough warning to allow one to react, whether for a human or
for a daemon monitoring PSI.

Here is output from a simple program showing how long it takes in μsec
(column 2) to allocate a megabyte of anonymous memory (column 1) when a
cgroup is already beyond its memory high setting, and no swap is
available:

    [root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
    > --wait -t timeout 300 /root/mdf
    [...]
    95  1035
    96  1038
    97  1000
    98  1036
    99  1048
    100 1590
    101 1968
    102 1776
    103 1863
    104 1757
    105 1921
    106 1893
    107 1760
    108 1748
    109 1843
    110 1716
    111 1924
    112 1776
    113 1831
    114 1766
    115 1836
    116 1588
    117 1912
    118 1802
    119 1857
    120 1731
    [...]
    [System OOM in 2-3 seconds]

The delay does go up extremely marginally past the 100MB memory.high
threshold, as now we spend time scanning before returning to usermode,
but it's nowhere near enough to contain growth. It also doesn't get
worse the more pages you have, since it only considers nr_pages.

The current situation goes against both the expectations of users of
memory.high, and our intentions as cgroup v2 developers. In
cgroup-v2.txt, we claim that we will throttle and only under "extreme
conditions" will memory.high protection be breached. Likewise, cgroup v2
users generally also expect that memory.high should throttle workloads
as they exceed their high threshold. However, as seen above, this isn't
always how it works in practice -- even on banal setups like those with
no swap, or where swap has become exhausted, we can end up with
memory.high being breached and us having no weapons left in our arsenal
to combat runaway growth with, since reclaim is futile.

It's also hard for system monitoring software or users to tell how bad
the situation is, as "high" events for the memcg may in some cases be
benign, and in others be catastrophic. The current status quo is that we
fail containment in a way that doesn't provide any advance warning that
things are about to go horribly wrong (for example, we are about to
invoke the kernel OOM killer).

This patch introduces explicit throttling when reclaim is failing to
keep memcg size contained at the memory.high setting. It does so by
applying an exponential delay curve derived from the memcg's overage
compared to memory.high.  In the normal case where the memcg is either
below or only marginally over its memory.high setting, no throttling
will be performed.

This composes well with system health monitoring and remediation, as
these allocator delays are factored into PSI's memory pressure
calculations. This both creates a mechanism system administrators or
applications consuming the PSI interface to trivially see that the memcg
in question is struggling and use that to make more reasonable
decisions, and permits them enough time to act. Either of these can act
with significantly more nuance than that we can provide using the system
OOM killer.

This is a similar idea to memory.oom_control in cgroup v1 which would
put the cgroup to sleep if the threshold was violated, but it's also
significantly improved as it results in visible memory pressure, and
also doesn't schedule indefinitely, which previously made tracing and
other introspection difficult.

Contrast the previous results with a kernel with this patch:

    [root@ktst ~]# systemd-run -p MemoryHigh=100M -p MemorySwapMax=1 \
    > --wait -t timeout 300 /root/mdf
    [...]
    95  1002
    96  1000
    97  1002
    98  1003
    99  1000
    100 1043
    101 84724
    102 330628
    103 610511
    104 1016265
    105 1503969
    106 2391692
    107 2872061
    108 3248003
    109 4791904
    110 5759832
    111 6912509
    112 8127818
    113 9472203
    114 12287622
    115 12480079
    116 14144008
    117 15808029
    118 16384500
    119 16383242
    120 16384979
    [...]

As you can see, in the normal case, memory allocation takes around 1000
μsec. However, as we exceed our memory.high, things start to increase
exponentially, but fairly leniently at first. Our first megabyte over
memory.high takes us 0.16 seconds, then the next is 0.46 seconds, then
the next is almost an entire second. This gets worse until we reach our
eventual 2*HZ clamp per batch, resulting in 16 seconds per megabyte.
However, this is still making forward progress, so permits tracing or
further analysis with programs like GDB.

This patch expands on earlier work by Johannes Weiner. Thanks!

Signed-off-by: Chris Down <chris@chrisdown.name>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: linux-kernel@vger.kernel.org
Cc: cgroups@vger.kernel.org
Cc: linux-mm@kvack.org
Cc: kernel-team@fb.com
---
 mm/memcontrol.c | 118 +++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 117 insertions(+), 1 deletion(-)

Comments

Michal Hocko Feb. 1, 2019, 7:17 a.m. UTC | #1
On Thu 31-01-19 20:13:52, Chris Down wrote:
[...]
> The current situation goes against both the expectations of users of
> memory.high, and our intentions as cgroup v2 developers. In
> cgroup-v2.txt, we claim that we will throttle and only under "extreme
> conditions" will memory.high protection be breached. Likewise, cgroup v2
> users generally also expect that memory.high should throttle workloads
> as they exceed their high threshold. However, as seen above, this isn't
> always how it works in practice -- even on banal setups like those with
> no swap, or where swap has become exhausted, we can end up with
> memory.high being breached and us having no weapons left in our arsenal
> to combat runaway growth with, since reclaim is futile.
> 
> It's also hard for system monitoring software or users to tell how bad
> the situation is, as "high" events for the memcg may in some cases be
> benign, and in others be catastrophic. The current status quo is that we
> fail containment in a way that doesn't provide any advance warning that
> things are about to go horribly wrong (for example, we are about to
> invoke the kernel OOM killer).
> 
> This patch introduces explicit throttling when reclaim is failing to
> keep memcg size contained at the memory.high setting. It does so by
> applying an exponential delay curve derived from the memcg's overage
> compared to memory.high.  In the normal case where the memcg is either
> below or only marginally over its memory.high setting, no throttling
> will be performed.

How does this play wit the actual OOM when the user expects oom to
resolve the situation because the reclaim is futile and there is nothing
reclaimable except for killing a process?
Johannes Weiner Feb. 1, 2019, 4:12 p.m. UTC | #2
On Fri, Feb 01, 2019 at 08:17:57AM +0100, Michal Hocko wrote:
> On Thu 31-01-19 20:13:52, Chris Down wrote:
> [...]
> > The current situation goes against both the expectations of users of
> > memory.high, and our intentions as cgroup v2 developers. In
> > cgroup-v2.txt, we claim that we will throttle and only under "extreme
> > conditions" will memory.high protection be breached. Likewise, cgroup v2
> > users generally also expect that memory.high should throttle workloads
> > as they exceed their high threshold. However, as seen above, this isn't
> > always how it works in practice -- even on banal setups like those with
> > no swap, or where swap has become exhausted, we can end up with
> > memory.high being breached and us having no weapons left in our arsenal
> > to combat runaway growth with, since reclaim is futile.
> > 
> > It's also hard for system monitoring software or users to tell how bad
> > the situation is, as "high" events for the memcg may in some cases be
> > benign, and in others be catastrophic. The current status quo is that we
> > fail containment in a way that doesn't provide any advance warning that
> > things are about to go horribly wrong (for example, we are about to
> > invoke the kernel OOM killer).
> > 
> > This patch introduces explicit throttling when reclaim is failing to
> > keep memcg size contained at the memory.high setting. It does so by
> > applying an exponential delay curve derived from the memcg's overage
> > compared to memory.high.  In the normal case where the memcg is either
> > below or only marginally over its memory.high setting, no throttling
> > will be performed.
> 
> How does this play wit the actual OOM when the user expects oom to
> resolve the situation because the reclaim is futile and there is nothing
> reclaimable except for killing a process?

Hm, can you elaborate on your question a bit?

The idea behind memory.high is to throttle allocations long enough for
the admin or a management daemon to intervene, but not to trigger the
kernel oom killer. It was designed as a replacement for the cgroup1
oom_control, but without the deadlock potential, ptrace problems etc.

What we specifically do is to set memory.high and have a daemon (oomd)
watch memory.pressure, io.pressure etc. in the group. If pressure
exceeds a certain threshold, the daemon kills something.

As you know, the kernel OOM killer does not kick in reliably when
e.g. page cache is thrashing heavily, since from a kernel POV it's
still successfully allocating and reclaiming - meanwhile the workload
is spending most its time in page faults. And when the kernel OOM
killer does kick in, its selection policy is not very workload-aware.

This daemon on the other hand can be configured to 1) kick in reliably
when the workload-specific tolerances for slowdowns and latencies are
violated (which tends to be way earlier than the kernel oom killer
usually kicks in) and 2) know about the workload and all its
components to make an informed kill decision.

Right now, that throttling mechanism works okay with swap enabled, but
we cannot enable swap everywhere, or sometimes run out of swap, and
then it breaks down and we run into system OOMs.

This patch makes sure memory.high *always* implements the throttling
semantics described in cgroup-v2.txt, not just most of the time.
Chris Down Feb. 1, 2019, 7:16 p.m. UTC | #3
Michal Hocko writes:
>How does this play wit the actual OOM when the user expects oom to
>resolve the situation because the reclaim is futile and there is nothing
>reclaimable except for killing a process?

In addition to what Johannes said, this doesn't impede OOM in the case of 
global system starvation (eg. in the case that all major consumers of memory 
are allocator throttling). In that case nothing unusual will happen, since the 
task's state is TASK_KILLABLE rather than TASK_UNINTERRUPTIBLE, and we will 
exit out of mem_cgroup_handle_over_high as quickly as possible.
Michal Hocko Feb. 28, 2019, 9:52 a.m. UTC | #4
[Sorry for a late reply]

On Fri 01-02-19 11:12:33, Johannes Weiner wrote:
> On Fri, Feb 01, 2019 at 08:17:57AM +0100, Michal Hocko wrote:
> > On Thu 31-01-19 20:13:52, Chris Down wrote:
> > [...]
> > > The current situation goes against both the expectations of users of
> > > memory.high, and our intentions as cgroup v2 developers. In
> > > cgroup-v2.txt, we claim that we will throttle and only under "extreme
> > > conditions" will memory.high protection be breached. Likewise, cgroup v2
> > > users generally also expect that memory.high should throttle workloads
> > > as they exceed their high threshold. However, as seen above, this isn't
> > > always how it works in practice -- even on banal setups like those with
> > > no swap, or where swap has become exhausted, we can end up with
> > > memory.high being breached and us having no weapons left in our arsenal
> > > to combat runaway growth with, since reclaim is futile.
> > > 
> > > It's also hard for system monitoring software or users to tell how bad
> > > the situation is, as "high" events for the memcg may in some cases be
> > > benign, and in others be catastrophic. The current status quo is that we
> > > fail containment in a way that doesn't provide any advance warning that
> > > things are about to go horribly wrong (for example, we are about to
> > > invoke the kernel OOM killer).
> > > 
> > > This patch introduces explicit throttling when reclaim is failing to
> > > keep memcg size contained at the memory.high setting. It does so by
> > > applying an exponential delay curve derived from the memcg's overage
> > > compared to memory.high.  In the normal case where the memcg is either
> > > below or only marginally over its memory.high setting, no throttling
> > > will be performed.
> > 
> > How does this play wit the actual OOM when the user expects oom to
> > resolve the situation because the reclaim is futile and there is nothing
> > reclaimable except for killing a process?
> 
> Hm, can you elaborate on your question a bit?
> 
> The idea behind memory.high is to throttle allocations long enough for
> the admin or a management daemon to intervene, but not to trigger the
> kernel oom killer. It was designed as a replacement for the cgroup1
> oom_control, but without the deadlock potential, ptrace problems etc.

Yes, this makes sense. The high limit reclaim is also documented as a
best effort resource guarantee. My understanding is that if the workload
cannot be contained within the high limit then the system cannot do much
and eventually gives up. Having the full memory unreclaimable is such an
example. And there is either the global OOM killer or hard limit OOM
killer to trigger to resolve such a situation.

[...]
Thanks for describing the usecase.

> Right now, that throttling mechanism works okay with swap enabled, but
> we cannot enable swap everywhere, or sometimes run out of swap, and
> then it breaks down and we run into system OOMs.
> 
> This patch makes sure memory.high *always* implements the throttling
> semantics described in cgroup-v2.txt, not just most of the time.

I am not really opposed to the throttling in the absence of a
reclaimable memory. We do that for the regular allocation paths already
(should_reclaim_retry). A swapless system with anon memory is very
likely to oom too quickly and this sounds like a real problem. But I do
not think that we should throttle the allocation to freeze it
completely. We should eventually OOM. And that was my question about
essentially. How much we can/should throttle to give a high limit events
consumer enough time to intervene. I am sorry to still not have time to
study the patch more closely but this should be explained in the
changelog. Are we talking about seconds/minutes or simply freeze each
allocator to death?

Patch
diff mbox series

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 18f4aefbe0bf..1844a88f1f68 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -65,6 +65,7 @@ 
 #include <linux/lockdep.h>
 #include <linux/file.h>
 #include <linux/tracehook.h>
+#include <linux/psi.h>
 #include "internal.h"
 #include <net/sock.h>
 #include <net/ip.h>
@@ -2161,12 +2162,68 @@  static void high_work_func(struct work_struct *work)
 	reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
 }
 
+/*
+ * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
+ * enough to still cause a significant slowdown in most cases, while still
+ * allowing diagnostics and tracing to proceed without becoming stuck.
+ */
+#define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
+
+/*
+ * When calculating the delay, we use these either side of the exponentiation to
+ * maintain precision and scale to a reasonable number of jiffies (see the table
+ * below.
+ *
+ * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
+ *   overage ratio to a delay.
+ * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
+ *   proposed penalty in order to reduce to a reasonable number of jiffies, and
+ *   to produce a reasonable delay curve.
+ *
+ * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
+ * reasonable delay curve compared to precision-adjusted overage, not
+ * penalising heavily at first, but still making sure that growth beyond the
+ * limit penalises misbehaviour cgroups by slowing them down exponentially. For
+ * example, with a high of 100 megabytes:
+ *
+ *  +-------+------------------------+
+ *  | usage | time to allocate in ms |
+ *  +-------+------------------------+
+ *  | 100M  |                      0 |
+ *  | 101M  |                      6 |
+ *  | 102M  |                     25 |
+ *  | 103M  |                     57 |
+ *  | 104M  |                    102 |
+ *  | 105M  |                    159 |
+ *  | 106M  |                    230 |
+ *  | 107M  |                    313 |
+ *  | 108M  |                    409 |
+ *  | 109M  |                    518 |
+ *  | 110M  |                    639 |
+ *  | 111M  |                    774 |
+ *  | 112M  |                    921 |
+ *  | 113M  |                   1081 |
+ *  | 114M  |                   1254 |
+ *  | 115M  |                   1439 |
+ *  | 116M  |                   1638 |
+ *  | 117M  |                   1849 |
+ *  | 118M  |                   2000 |
+ *  | 119M  |                   2000 |
+ *  | 120M  |                   2000 |
+ *  +-------+------------------------+
+ */
+ #define MEMCG_DELAY_PRECISION_SHIFT 20
+ #define MEMCG_DELAY_SCALING_SHIFT 14
+
 /*
  * Scheduled by try_charge() to be executed from the userland return path
  * and reclaims memory over the high limit.
  */
 void mem_cgroup_handle_over_high(void)
 {
+	unsigned long usage, high;
+	unsigned long pflags;
+	unsigned long penalty_jiffies, overage;
 	unsigned int nr_pages = current->memcg_nr_pages_over_high;
 	struct mem_cgroup *memcg = current->memcg_high_reclaim;
 
@@ -2177,9 +2234,68 @@  void mem_cgroup_handle_over_high(void)
 		memcg = get_mem_cgroup_from_mm(current->mm);
 
 	reclaim_high(memcg, nr_pages, GFP_KERNEL);
-	css_put(&memcg->css);
 	current->memcg_high_reclaim = NULL;
 	current->memcg_nr_pages_over_high = 0;
+
+	/*
+	 * memory.high is breached and reclaim is unable to keep up. Throttle
+	 * allocators proactively to slow down excessive growth.
+	 *
+	 * We use overage compared to memory.high to calculate the number of
+	 * jiffies to sleep (penalty_jiffies). Ideally this value should be
+	 * fairly lenient on small overages, and increasingly harsh when the
+	 * memcg in question makes it clear that it has no intention of stopping
+	 * its crazy behaviour, so we exponentially increase the delay based on
+	 * overage amount.
+	 */
+
+	usage = page_counter_read(&memcg->memory);
+	high = READ_ONCE(memcg->high);
+
+	if (usage <= high)
+		goto out;
+
+	overage = ((u64)(usage - high) << MEMCG_DELAY_PRECISION_SHIFT) / high;
+	penalty_jiffies = ((u64)overage * overage * HZ)
+		>> (MEMCG_DELAY_PRECISION_SHIFT + MEMCG_DELAY_SCALING_SHIFT);
+
+	/*
+	 * Factor in the task's own contribution to the overage, such that four
+	 * N-sized allocations are throttled approximately the same as one
+	 * 4N-sized allocation.
+	 *
+	 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
+	 * larger the current charge patch is than that.
+	 */
+	penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
+
+	/*
+	 * Clamp the max delay per usermode return so as to still keep the
+	 * application moving forwards and also permit diagnostics, albeit
+	 * extremely slowly.
+	 */
+	penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
+
+	/*
+	 * Don't sleep if the amount of jiffies this memcg owes us is so low
+	 * that it's not even worth doing, in an attempt to be nice to those who
+	 * go only a small amount over their memory.high value and maybe haven't
+	 * been aggressively reclaimed enough yet.
+	 */
+	if (penalty_jiffies <= HZ / 100)
+		goto out;
+
+	/*
+	 * If we exit early, we're guaranteed to die (since
+	 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
+	 * need to account for any ill-begotten jiffies to pay them off later.
+	 */
+	psi_memstall_enter(&pflags);
+	schedule_timeout_killable(penalty_jiffies);
+	psi_memstall_leave(&pflags);
+
+out:
+	css_put(&memcg->css);
 }
 
 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,