[Devel] [PATCH rh7] vmscan: do not throttle kthreads due to too_many_isolated

[Konstantin Khorenko]

Dima, please review.

--
Best regards,

Konstantin Khorenko,
Virtuozzo Linux Kernel Team

On 11/26/2015 11:48 AM, Vladimir Davydov wrote:
> Ploop hands off bio processing to its kernel thread. This might result
> to a dead lock as described below:
>
>   1. Several processes performing memcg reclaim isolate a bunch of
>      reclaimable pages and all get stuck in shrink_page_list waiting for
>      a page writeback to finish. Page writeback is up to ploop.
>
>   2. Ploop thread tries to allocate some memory and gets stuck in
>      shrink_inactive_list waiting for too_many_isolated to return false.
>      The latter will never happen, because the pages were isolated by
>      processes at step 1 waiting for ploop to complete page writeback.
>
>      PID: 44857  TASK: ffff880020bbf1a0  CPU: 1   COMMAND: "ploop37460"
>      [ffff880079b67228] schedule at ffffffff81607579
>      [ffff880079b67238] schedule_timeout at ffffffff81605365
>      [ffff880079b67270] process_timeout at ffffffff81080d10
>      [ffff880079b672e0] io_schedule_timeout at ffffffff81606c6b
>      [ffff880079b67310] congestion_wait at ffffffff8118ad32
>      [ffff880079b67328] autoremove_wake_function at ffffffff8109b640
>      [ffff880079b67370] shrink_inactive_list at ffffffff8117f7ec
>      [ffff880079b67388] radix_tree_delete_item at ffffffff812d9b76
>      [ffff880079b67458] shrink_lruvec at ffffffff811800f5
>      [ffff880079b67560] shrink_zone at ffffffff811806b1
>      [ffff880079b675f0] do_try_to_free_pages at ffffffff81180c20
>      [ffff880079b67690] try_to_free_pages at ffffffff811810b5
>      [ffff880079b67720] __alloc_pages_nodemask at ffffffff81174c17
>      [ffff880079b67860] alloc_pages_current at ffffffff811b8339
>      [ffff880079b678a8] new_slab at ffffffff811c4273
>      [ffff880079b678e8] __slab_alloc at ffffffff815ffeca
>      [ffff880079b67928] __radix_tree_preload at ffffffff812d859c
>      [ffff880079b67980] __mem_cgroup_commit_charge at ffffffff811d5c42
>      [ffff880079b679c0] kmem_cache_alloc at ffffffff811c6e8c
>      [ffff880079b679c8] __radix_tree_preload at ffffffff812d859c
>      [ffff880079b67a00] __radix_tree_preload at ffffffff812d859c
>      [ffff880079b67a28] radix_tree_maybe_preload at ffffffff812d8659
>      [ffff880079b67a38] __add_to_page_cache_locked at ffffffff8116a57b
>      [ffff880079b67a90] add_to_page_cache_lru at ffffffff8116a7b7
>      [ffff880079b67ac0] grab_cache_page_write_begin at ffffffff8116ac0f
>      [ffff880079b67b00] ext4_da_write_begin at ffffffffa0301054 [ext4]
>      [ffff880079b67b80] pagecache_write_begin at ffffffff81168f1c
>      [ffff880079b67b98] cached_submit at ffffffffa03e432b [pio_direct]
>      [ffff880079b67c80] dio_submit_alloc at ffffffffa03e5866 [pio_direct]
>      [ffff880079b67ca0] dio_submit at ffffffffa03e4eef [pio_direct]
>      [ffff880079b67d20] ploop1_allocate at ffffffffa03ddafc [pfmt_ploop1]
>      [ffff880079b67d38] ploop_entry_request at ffffffffa03b9be4 [ploop]
>      [ffff880079b67db8] ploop_req_state_process at ffffffffa03ba56d [ploop]
>      [ffff880079b67dd8] autoremove_wake_function at ffffffff8109b640
>      [ffff880079b67e28] ploop_thread at ffffffffa03baf13 [ploop]
>      [ffff880079b67ea8] ploop_thread at ffffffffa03bad60 [ploop]
>      [ffff880079b67ec8] kthread at ffffffff8109a4df
>      [ffff880079b67f30] kthread at ffffffff8109a410
>      [ffff880079b67f50] ret_from_fork at ffffffff81612058
>      [ffff880079b67f80] kthread at ffffffff8109a410
>
> Fix this by omitting the too_many_isolated check for kernel threads.
> This sounds reasonable not only for ploop kthread, because any kthread
> might be responsible for reclaimer progress and therefore should not be
> throttled by it.
>
> https://jira.sw.ru/browse/PSBM-41410
>
> Signed-off-by: Vladimir Davydov <vdavydov at virtuozzo.com>
> ---
>   mm/vmscan.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index eadc08b7a5d6..3b96a02eaddd 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -1426,7 +1426,7 @@ static int __too_many_isolated(struct zone *zone, int file,
>   static int too_many_isolated(struct zone *zone, int file,
>   			     struct scan_control *sc)
>   {
> -	if (current_is_kswapd())
> +	if (current->flags & PF_KTHREAD)
>   		return 0;
>
>   	if (!global_reclaim(sc))
>