Date:   Fri, 24 May 2019 11:04:06 -0700
From:   Shakeel Butt <shakeelb@...gle.com>
To:     Johannes Weiner <hannes@...xchg.org>
Cc:     Matthew Wilcox <willy@...radead.org>,
        Linus Torvalds <torvalds@...ux-foundation.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Linux MM <linux-mm@...ck.org>,
        Cgroups <cgroups@...r.kernel.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Kernel Team <kernel-team@...com>
Subject: Re: [PATCH] mm: fix page cache convergence regression

On Fri, May 24, 2019 at 10:41 AM Johannes Weiner <hannes@...xchg.org> wrote:
>
> On Fri, May 24, 2019 at 09:04:17AM -0700, Matthew Wilcox wrote:
> > On Fri, May 24, 2019 at 11:31:48AM -0400, Johannes Weiner wrote:
> > > diff --git a/include/linux/xarray.h b/include/linux/xarray.h
> > > index 0e01e6129145..cbbf76e4c973 100644
> > > --- a/include/linux/xarray.h
> > > +++ b/include/linux/xarray.h
> > > @@ -292,6 +292,7 @@ struct xarray {
> > >     spinlock_t      xa_lock;
> > >  /* private: The rest of the data structure is not to be used directly. */
> > >     gfp_t           xa_flags;
> > > +   gfp_t           xa_gfp;
> > >     void __rcu *    xa_head;
> > >  };
> >
> > No.  I'm willing to go for a xa_flag which says to use __GFP_ACCOUNT, but
> > you can't add another element to the struct xarray.
>
> Ok, we can generalize per-tree gfp flags later if necessary.
>
> Below is the updated fix that uses an XA_FLAGS_ACCOUNT flag instead.
>
> ---
> From 63a0dbc571ff38f7c072c62d6bc28192debe37ac Mon Sep 17 00:00:00 2001
> From: Johannes Weiner <hannes@...xchg.org>
> Date: Fri, 24 May 2019 10:12:46 -0400
> Subject: [PATCH] mm: fix page cache convergence regression
>
> Since a28334862993 ("page cache: Finish XArray conversion"), on most
> major Linux distributions, the page cache doesn't correctly transition
> when the hot data set is changing, and leaves the new pages thrashing
> indefinitely instead of kicking out the cold ones.
>
> On a freshly booted, freshly ssh'd into virtual machine with 1G RAM
> running stock Arch Linux:
>
> [root@ham ~]# ./reclaimtest.sh
> + dd of=workingset-a bs=1M count=0 seek=600
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + ./mincore workingset-a
> 153600/153600 workingset-a
> + dd of=workingset-b bs=1M count=0 seek=600
> + cat workingset-b
> + cat workingset-b
> + cat workingset-b
> + cat workingset-b
> + ./mincore workingset-a workingset-b
> 104029/153600 workingset-a
> 120086/153600 workingset-b
> + cat workingset-b
> + cat workingset-b
> + cat workingset-b
> + cat workingset-b
> + ./mincore workingset-a workingset-b
> 104029/153600 workingset-a
> 120268/153600 workingset-b
>
> workingset-b is a 600M file on a 1G host that is otherwise entirely
> idle. No matter how often it's being accessed, it won't get cached.
>
> While investigating, I noticed that the non-resident information gets
> aggressively reclaimed - /proc/vmstat::workingset_nodereclaim. This is
> a problem because a workingset transition like this relies on the
> non-resident information tracked in the page cache tree of evicted
> file ranges: when the cache faults are refaults of recently evicted
> cache, we challenge the existing active set, and that allows a new
> workingset to establish itself.
>
> Tracing the shrinker that maintains this memory revealed that all page
> cache tree nodes were allocated to the root cgroup. This is a problem,
> because 1) the shrinker sizes the amount of non-resident information
> it keeps to the size of the cgroup's other memory and 2) on most major
> Linux distributions, only kernel threads live in the root cgroup and
> everything else gets put into services or session groups:
>
> [root@ham ~]# cat /proc/self/cgroup
> 0::/user.slice/user-0.slice/session-c1.scope
>
> As a result, we basically maintain no non-resident information for the
> workloads running on the system, thus breaking the caching algorithm.
>
> Looking through the code, I found the culprit in the above-mentioned
> patch: when switching from the radix tree to xarray, it dropped the
> __GFP_ACCOUNT flag from the tree node allocations - the flag that
> makes sure the allocated memory gets charged to and tracked by the
> cgroup of the calling process - in this case, the one doing the fault.
>
> To fix this, allow xarray users to specify per-tree flag that makes
> xarray allocate nodes using __GFP_ACCOUNT. Then restore the page cache
> tree annotation to request such cgroup tracking for the cache nodes.
>
> With this patch applied, the page cache correctly converges on new
> workingsets again after just a few iterations:
>
> [root@ham ~]# ./reclaimtest.sh
> + dd of=workingset-a bs=1M count=0 seek=600
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + cat workingset-a
> + ./mincore workingset-a
> 153600/153600 workingset-a
> + dd of=workingset-b bs=1M count=0 seek=600
> + cat workingset-b
> + ./mincore workingset-a workingset-b
> 124607/153600 workingset-a
> 87876/153600 workingset-b
> + cat workingset-b
> + ./mincore workingset-a workingset-b
> 81313/153600 workingset-a
> 133321/153600 workingset-b
> + cat workingset-b
> + ./mincore workingset-a workingset-b
> 63036/153600 workingset-a
> 153600/153600 workingset-b
>
> Cc: stable@...r.kernel.org # 4.20+
> Signed-off-by: Johannes Weiner <hannes@...xchg.org>

Reviewed-by: Shakeel Butt <shakeelb@...gle.com>

Powered by blists - more mailing lists