Re: [PATCH] btrfs: Keep one more workspace around

From: David Sterba
Date: Thu Jun 29 2017 - 10:00:27 EST


On Wed, Jun 28, 2017 at 08:01:51PM -0700, Nick Terrell wrote:
> > Is there a version I should be testing?
>
> Not yet, I'm working on v2 of the patch set, which will be ready soon.
>
> > I got a bunch of those:
> > [10170.448783] kworker/u8:6: page allocation stalls for 60720ms, order:0, mode:0x14000c2(GFP_KERNEL|__GFP_HIGHMEM), nodemask=(null)
> > [10170.448819] kworker/u8:6 cpuset=/ mems_allowed=0
> > [10170.448842] CPU: 3 PID: 13430 Comm: kworker/u8:6 Not tainted 4.12.0-rc7-00034-gdff47ed160bb #1
> > [10170.448846] Hardware name: SAMSUNG EXYNOS (Flattened Device Tree)
> > [10170.448872] Workqueue: btrfs-endio btrfs_endio_helper
> > [10170.448910] [<c010de1c>] (unwind_backtrace) from [<c010adb8>] (show_stack+0x10/0x14)
> > [10170.448925] [<c010adb8>] (show_stack) from [<c0442b00>] (dump_stack+0x78/0x8c)
> > [10170.448942] [<c0442b00>] (dump_stack) from [<c01b0178>] (warn_alloc+0xc0/0x170)
> > [10170.448952] [<c01b0178>] (warn_alloc) from [<c01b0c3c>] (__alloc_pages_nodemask+0x97c/0xe30)
> > [10170.448964] [<c01b0c3c>] (__alloc_pages_nodemask) from [<c01e217c>] (__vmalloc_node_range+0x144/0x27c)
> > [10170.448976] [<c01e217c>] (__vmalloc_node_range) from [<c01e2550>] (__vmalloc_node.constprop.10+0x48/0x50)
> > [10170.448982] [<c01e2550>] (__vmalloc_node.constprop.10) from [<c01e25ec>] (vmalloc+0x2c/0x34)
> > [10170.448990] [<c01e25ec>] (vmalloc) from [<c038f7cc>] (zstd_alloc_workspace+0x6c/0xb8)
> > [10170.448997] [<c038f7cc>] (zstd_alloc_workspace) from [<c038fcb8>] (find_workspace+0x120/0x1f4)
> > [10170.449002] [<c038fcb8>] (find_workspace) from [<c038ff60>] (end_compressed_bio_read+0x1d4/0x3b0)
> > [10170.449016] [<c038ff60>] (end_compressed_bio_read) from [<c0130e14>] (process_one_work+0x1d8/0x3f0)
> > [10170.449026] [<c0130e14>] (process_one_work) from [<c0131a18>] (worker_thread+0x38/0x558)
> > [10170.449035] [<c0131a18>] (worker_thread) from [<c0136854>] (kthread+0x124/0x154)
> > [10170.449042] [<c0136854>] (kthread) from [<c01076f8>] (ret_from_fork+0x14/0x3c)
> >
> > which never happened with compress=lzo, and a 2GB RAM machine that runs 4
> > threads of various builds runs into memory pressure quite often. On the
> > other hand, I used 4.11 for lzo so this needs more testing before I can
> > blame the zstd code.
>
> I'm not sure what is causing the symptom of stalls in vmalloc(), but I
> think I know what is causing vmalloc() to be called so often. Its probably
> showing up for zstd and not lzo because it requires more memory.
>
> find_workspace() allocates up to num_online_cpus() + 1 workspaces.
> free_workspace() will only keep num_online_cpus() workspaces. When
> (de)compressing we will allocate num_online_cpus() + 1 workspaces, then
> free one, and repeat. Instead, we can just keep num_online_cpus() + 1
> workspaces around, and never have to allocate/free another workspace in the
> common case.

That would be much better and probably was the original intention. And I
guess improves performance when we don't have to do the extra alloc/free
rounds.

> I tested on a Ubuntu 14.04 VM with 2 cores and 4 GiB of RAM. I mounted a
> BtrFS partition with -o compress-force={lzo,zlib,zstd} and logged whenever
> a workspace was allocated of freed. Then I copied vmlinux (527 MB) to the
> partition. Before the patch, during the copy it would allocate and free 5-6
> workspaces. After, it only allocated the initial 3. This held true for lzo,
> zlib, and zstd.
>
> > I'm on linus:4.12-rc7 with only a handful of btrfs patches (v3 of Qu's chunk
> > check, some misc crap) -- I guess I should use at least btrfs-for-4.13. Or
> > would you prefer full-blown next?
>
> Whatever is convenient for you. The relevant code in BtrFS hasn't changed
> for a few months, so it shouldn't matter too much.
>
> Signed-off-by: Nick Terrell <terrelln@xxxxxx>
> ---
> fs/btrfs/compression.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/fs/btrfs/compression.c b/fs/btrfs/compression.c
> index 3beb0d0..1a0ef55 100644
> --- a/fs/btrfs/compression.c
> +++ b/fs/btrfs/compression.c
> @@ -874,7 +874,7 @@ static void free_workspace(int type, struct list_head *workspace)
> int *free_ws = &btrfs_comp_ws[idx].free_ws;
>
> spin_lock(ws_lock);
> - if (*free_ws < num_online_cpus()) {
> + if (*free_ws <= num_online_cpus()) {
> list_add(workspace, idle_ws);
> (*free_ws)++;
> spin_unlock(ws_lock);

Please send it as a proper patch, thanks.