Linux-XFS Archive on
 help / color / Atom feed
From: Amir Goldstein <>
To: Dave Chinner <>
Cc: linux-xfs <>
Subject: Re: [PATCH 00/24] xfs: rework inode flushing to make inode reclaim fully asynchronous
Date: Fri, 22 May 2020 09:18:37 +0300
Message-ID: <> (raw)
In-Reply-To: <>

On Fri, May 22, 2020 at 6:51 AM Dave Chinner <> wrote:
> Hi folks,
> Inode flushing requires that we first lock an inode, then check it,
> then lock the underlying buffer, flush the inode to the buffer and
> finally add the inode to the buffer to be unlocked on IO completion.
> We then walk all the other cached inodes in the buffer range and
> optimistically lock and flush them to the buffer without blocking.
> This cluster write effectively repeats the same code we do with the
> initial inode, except now it has to special case that initial inode
> that is already locked. Hence we have multiple copies of very
> similar code, and it is a result of inode cluster flushing being
> based on a specific inode rather than grabbing the buffer and
> flushing all available inodes to it.
> The problem with this at the moment is that we we can't look up the
> buffer until we have guaranteed that an inode is held exclusively
> and it's not going away while we get the buffer through an imap
> lookup. Hence we are kinda stuck locking an inode before we can look
> up the buffer.
> This is also a result of inodes being detached from the cluster
> buffer except when IO is being done. This has the further problem
> that the cluster buffer can be reclaimed from memory and then the
> inode can be dirtied. At this point cleaning the inode requires a
> read-modify-write cycle on the cluster buffer. If we then are put
> under memory pressure, cleaning that dirty inode to reclaim it
> requires allocating memory for the cluster buffer and this leads to
> all sorts of problems.
> We used synchronous inode writeback in reclaim as a throttle that
> provided a forwards progress mechanism when RMW cycles were required
> to clean inodes. Async writeback of inodes (e.g. via the AIL) would
> immediately exhaust remaining memory reserves trying to allocate
> inode cluster after inode cluster. The synchronous writeback of an
> inode cluster allowed reclaim to release the inode cluster and have
> it freed almost immediately which could then be used to allocate the
> next inode cluster buffer. Hence the IO based throttling mechanism
> largely guaranteed forwards progress in inode reclaim. By removing
> the requirement for require memory allocation for inode writeback
> filesystem level, we can issue writeback asynchrnously and not have
> to worry about the memory exhaustion anymore.
> Another issue is that if we have slow disks, we can build up dirty
> inodes in memory that can then take hours for an operation like
> unmount to flush. A RMW cycle per inode on a slow RAID6 device can
> mean we only clean 50 inodes a second, and when there are hundreds
> of thousands of dirty inodes that need to be cleaned this can take a
> long time. PInning the cluster buffers will greatly speed up inode
> writeback on slow storage systems like this.
> These limitations all stem from the same source: inode writeback is
> inode centric, And they are largely solved by the same architectural
> change: make inode writeback cluster buffer centric.  This series is
> makes that architectural change.
> Firstly, we start by pinning the inode backing buffer in memory
> when an inode is marked dirty (i.e. when it is logged). By tracking
> the number of dirty inodes on a buffer as a counter rather than a
> flag, we avoid the problem of overlapping inode dirtying and buffer
> flushing racing to set/clear the dirty flag. Hence as long as there
> is a dirty inode in memory, the buffer will not be able to be
> reclaimed. We can safely do this inode cluster buffer lookup when we
> dirty an inode as we do not hold the buffer locked - we merely take
> a reference to it and then release it - and hence we don't cause any
> new lock order issues.
> When the inode is finally cleaned, the reference to the buffer can
> be removed from the inode log item and the buffer released. This is
> done from the inode completion callbacks that are attached to the
> buffer when the inode is flushed.
> Pinning the cluster buffer in this way immediately avoids the RMW
> problem in inode writeback and reclaim contexts by moving the memory
> allocation and the blocking buffer read into the transaction context
> that dirties the inode.  This inverts our dirty inode throttling
> mechanism - we now throttle the rate at which we can dirty inodes to
> rate at which we can allocate memory and read inode cluster buffers
> into memory rather than via throttling reclaim to rate at which we
> can clean dirty inodes.
> Hence if we are under memory pressure, we'll block on memory
> allocation when trying to dirty the referenced inode, rather than in
> the memory reclaim path where we are trying to clean unreferenced
> inodes to free memory.  Hence we no longer have to guarantee
> forwards progress in inode reclaim as we aren't doing memory
> allocation, and that means we can remove inode writeback from the
> XFS inode shrinker completely without changing the system tolerance
> for low memory operation.
> Tracking the buffers via the inode log item also allows us to
> completely rework the inode flushing mechanism. While the inode log
> item is in the AIL, it is safe for the AIL to access any member of
> the log item. Hence the AIL push mechanisms can access the buffer
> attached to the inode without first having to lock the inode.
> This means we can essentially lock the buffer directly and then
> call xfs_iflush_cluster() without first going through xfs_iflush()
> to find the buffer. Hence we can remove xfs_iflush() altogether,
> because the two places that call it - the inode item push code and
> inode reclaim - no longer need to flush inodes directly.
> This can be further optimised by attaching the inode to the cluster
> buffer when the inode is dirtied. i.e. when we add the buffer
> reference to the inode log item, we also attach the inode to the
> buffer for IO processing. This leads to the dirty inodes always
> being attached to the buffer and hence we no longer need to add them
> when we flush the inode and remove them when IO completes. Instead
> the inodes are attached when the node log item is dirtied, and
> removed when the inode log item is cleaned.
> With this structure in place, we no longer need to do
> lookups to find the dirty inodes in the cache to attach to the
> buffer in xfs_iflush_cluster() - they are already attached to the
> buffer. Hence when the AIL pushes an inode, we just grab the buffer
> from the log item, and then walk the buffer log item list to lock
> and flush the dirty inodes attached to the buffer.
> This greatly simplifies inode writeback, and removes another memory
> allocation from the inode writeback path (the array used for the
> radix tree gang lookup). And while the radix tree lookups are fast,
> walking the linked list of dirty inodes is faster.
> There is followup work I am doing that uses the inode cluster buffer
> as a replacement in the AIL for tracking dirty inodes. This part of
> the series is not ready yet as it has some intricate locking
> requirements. That is an optimisation, so I've left that out because
> solving the inode reclaim blocking problems is the important part of
> this work.
> In short, this series simplifies inode writeback and fixes the long
> standing inode reclaim blocking issues without requiring any changes
> to the memory reclaim infrastructure.
> Note: dquots should probably be converted to cluster flushing in a
> similar way, as they have many of the same issues as inode flushing.
> Thoughts,

Thank you for this wonderful write up!
Can't wait to deploy this on production machines :)

> comments

Since this is a "long standing issue" I am sure you have tests.
Can you please share the tests you used (simoop?) with some
encouraging performance numbers - flattening the stall curve ;-).

> and improvements

I wonder if the cluster buffer inode counter can be mixed as another
metric for choosing the best inodes to reclaim. If my math is correct
and depending on the specific system page/struct inode size,
reclaiming a cluster buffer with single dirty inode reclaims at least
1 page + 1 xfs_inode with low probability to free a slab page.
Reclaiming a cluster buffer with 8 dirty inodes reclaims at least
1 page + 8 xfs_inode with much higher probability to freeing one or two
slab pages (in the best case where adjacent inodes where allocated


  parent reply index

Thread overview: 91+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-05-22  3:50 Dave Chinner
2020-05-22  3:50 ` [PATCH 01/24] xfs: remove logged flag from inode log item Dave Chinner
2020-05-22  7:25   ` Christoph Hellwig
2020-05-22 21:13   ` Darrick J. Wong
2020-05-22  3:50 ` [PATCH 02/24] xfs: add an inode item lock Dave Chinner
2020-05-22  6:45   ` Amir Goldstein
2020-05-22 21:24   ` Darrick J. Wong
2020-05-23  8:45   ` Christoph Hellwig
2020-05-22  3:50 ` [PATCH 03/24] xfs: mark inode buffers in cache Dave Chinner
2020-05-22  7:45   ` Amir Goldstein
2020-05-22 21:35   ` Darrick J. Wong
2020-05-24 23:41     ` Dave Chinner
2020-05-23  8:48   ` Christoph Hellwig
2020-05-25  0:06     ` Dave Chinner
2020-05-22  3:50 ` [PATCH 04/24] xfs: mark dquot " Dave Chinner
2020-05-22  7:46   ` Amir Goldstein
2020-05-22 21:38   ` Darrick J. Wong
2020-05-22  3:50 ` [PATCH 05/24] xfs: mark log recovery buffers for completion Dave Chinner
2020-05-22  7:41   ` Amir Goldstein
2020-05-24 23:54     ` Dave Chinner
2020-05-22 21:41   ` Darrick J. Wong
2020-05-22  3:50 ` [PATCH 06/24] xfs: call xfs_buf_iodone directly Dave Chinner
2020-05-22  7:56   ` Amir Goldstein
2020-05-22 21:53   ` Darrick J. Wong
2020-05-22  3:50 ` [PATCH 07/24] xfs: clean up whacky buffer log item list reinit Dave Chinner
2020-05-22 22:01   ` Darrick J. Wong
2020-05-23  8:50   ` Christoph Hellwig
2020-05-22  3:50 ` [PATCH 08/24] xfs: fold xfs_istale_done into xfs_iflush_done Dave Chinner
2020-05-22 22:10   ` Darrick J. Wong
2020-05-23  9:12   ` Christoph Hellwig
2020-05-22  3:50 ` [PATCH 09/24] xfs: use direct calls for dquot IO completion Dave Chinner
2020-05-22 22:13   ` Darrick J. Wong
2020-05-23  9:16   ` Christoph Hellwig
2020-05-22  3:50 ` [PATCH 10/24] xfs: clean up the buffer iodone callback functions Dave Chinner
2020-05-22 22:26   ` Darrick J. Wong
2020-05-25  0:37     ` Dave Chinner
2020-05-23  9:19   ` Christoph Hellwig
2020-05-22  3:50 ` [PATCH 11/24] xfs: get rid of log item callbacks Dave Chinner
2020-05-22 22:27   ` Darrick J. Wong
2020-05-23  9:19   ` Christoph Hellwig
2020-05-22  3:50 ` [PATCH 12/24] xfs: pin inode backing buffer to the inode log item Dave Chinner
2020-05-22 22:39   ` Darrick J. Wong
2020-05-23  9:34   ` Christoph Hellwig
2020-05-23 21:43     ` Dave Chinner
2020-05-24  5:31       ` Christoph Hellwig
2020-05-24 23:13         ` Dave Chinner
2020-05-22  3:50 ` [PATCH 13/24] xfs: make inode reclaim almost non-blocking Dave Chinner
2020-05-22 12:19   ` Amir Goldstein
2020-05-22 22:48   ` Darrick J. Wong
2020-05-23 22:29     ` Dave Chinner
2020-05-22  3:50 ` [PATCH 14/24] xfs: remove IO submission from xfs_reclaim_inode() Dave Chinner
2020-05-22 23:06   ` Darrick J. Wong
2020-05-25  3:49     ` Dave Chinner
2020-05-23  9:40   ` Christoph Hellwig
2020-05-23 22:35     ` Dave Chinner
2020-05-22  3:50 ` [PATCH 15/24] xfs: allow multiple reclaimers per AG Dave Chinner
2020-05-22 23:10   ` Darrick J. Wong
2020-05-23 22:35     ` Dave Chinner
2020-05-22  3:50 ` [PATCH 16/24] xfs: don't block inode reclaim on the ILOCK Dave Chinner
2020-05-22 23:11   ` Darrick J. Wong
2020-05-22  3:50 ` [PATCH 17/24] xfs: remove SYNC_TRYLOCK from inode reclaim Dave Chinner
2020-05-22 23:14   ` Darrick J. Wong
2020-05-23 22:42     ` Dave Chinner
2020-05-22  3:50 ` [PATCH 18/24] xfs: clean up inode reclaim comments Dave Chinner
2020-05-22 23:17   ` Darrick J. Wong
2020-05-22  3:50 ` [PATCH 19/24] xfs: attach inodes to the cluster buffer when dirtied Dave Chinner
2020-05-22 23:48   ` Darrick J. Wong
2020-05-23 22:59     ` Dave Chinner
2020-05-22  3:50 ` [PATCH 20/24] xfs: xfs_iflush() is no longer necessary Dave Chinner
2020-05-22 23:54   ` Darrick J. Wong
2020-05-22  3:50 ` [PATCH 21/24] xfs: rename xfs_iflush_int() Dave Chinner
2020-05-22 12:33   ` Amir Goldstein
2020-05-22 23:57   ` Darrick J. Wong
2020-05-22  3:50 ` [PATCH 22/24] xfs: rework xfs_iflush_cluster() dirty inode iteration Dave Chinner
2020-05-23  0:13   ` Darrick J. Wong
2020-05-23 23:14     ` Dave Chinner
2020-05-23 11:31   ` Christoph Hellwig
2020-05-23 23:23     ` Dave Chinner
2020-05-24  5:32       ` Christoph Hellwig
2020-05-23 11:39   ` Christoph Hellwig
2020-05-22  3:50 ` [PATCH 23/24] xfs: factor xfs_iflush_done Dave Chinner
2020-05-23  0:20   ` Darrick J. Wong
2020-05-23 11:35   ` Christoph Hellwig
2020-05-22  3:50 ` [PATCH 24/24] xfs: remove xfs_inobp_check() Dave Chinner
2020-05-23  0:16   ` Darrick J. Wong
2020-05-23 11:36   ` Christoph Hellwig
2020-05-22  4:04 ` [PATCH 00/24] xfs: rework inode flushing to make inode reclaim fully asynchronous Dave Chinner
2020-05-23 16:18   ` Darrick J. Wong
2020-05-23 21:22     ` Dave Chinner
2020-05-22  6:18 ` Amir Goldstein [this message]
2020-05-22 12:01   ` Amir Goldstein

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \ \ \ \ \

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

Linux-XFS Archive on

Archives are clonable:
	git clone --mirror linux-xfs/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 linux-xfs linux-xfs/ \
	public-inbox-index linux-xfs

Example config snippet for mirrors

Newsgroup available over NNTP:

AGPL code for this site: git clone