lkml.org 
[lkml]   [2017]   [May]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v2 1/6] mm, page_alloc: fix more premature OOM due to race with cpuset update
On Wed 17-05-17 10:11:35, Vlastimil Babka wrote:
> Commit e47483bca2cc ("mm, page_alloc: fix premature OOM when racing with cpuset
> mems update") has fixed known recent regressions found by LTP's cpuset01
> testcase. I have however found that by modifying the testcase to use per-vma
> mempolicies via bind(2) instead of per-task mempolicies via set_mempolicy(2),
> the premature OOM still happens and the issue is much older.
>
> The root of the problem is that the cpuset's mems_allowed and mempolicy's
> nodemask can temporarily have no intersection, thus get_page_from_freelist()
> cannot find any usable zone. The current semantic for empty intersection is to
> ignore mempolicy's nodemask and honour cpuset restrictions. This is checked in
> node_zonelist(), but the racy update can happen after we already passed the
> check. Such races should be protected by the seqlock task->mems_allowed_seq,
> but it doesn't work here, because 1) mpol_rebind_mm() does not happen under
> seqlock for write, and doing so would lead to deadlock, as it takes mmap_sem
> for write, while the allocation can have mmap_sem for read when it's taking the
> seqlock for read. And 2) the seqlock cookie of callers of node_zonelist()
> (alloc_pages_vma() and alloc_pages_current()) is different than the one of
> __alloc_pages_slowpath(), so there's still a potential race window.
>
> This patch fixes the issue by having __alloc_pages_slowpath() check for empty
> intersection of cpuset and ac->nodemask before OOM or allocation failure. If
> it's indeed empty, the nodemask is ignored and allocation retried, which mimics
> node_zonelist(). This works fine, because almost all callers of
> __alloc_pages_nodemask are obtaining the nodemask via node_zonelist(). The only
> exception is new_node_page() from hotplug, where the potential violation of
> nodemask isn't an issue, as there's already a fallback allocation attempt
> without any nodemask. If there's a future caller that needs to have its specific
> nodemask honoured over task's cpuset restrictions, we'll have to e.g. add a gfp
> flag for that.
>
> Signed-off-by: Vlastimil Babka <vbabka@suse.cz>

Do we want this backported to the stable tree?

OK I do agree this makes some sense as a quick and easy to backport
workaround.

Acked-by: Michal Hocko <mhocko@suse.com?

> ---
> mm/page_alloc.c | 51 ++++++++++++++++++++++++++++++++++++++-------------
> 1 file changed, 38 insertions(+), 13 deletions(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index beb2827fd5de..43aa767c3188 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -3661,6 +3661,39 @@ should_reclaim_retry(gfp_t gfp_mask, unsigned order,
> return false;
> }
>
> +static inline bool
> +check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
> +{
> + /*
> + * It's possible that cpuset's mems_allowed and the nodemask from
> + * mempolicy don't intersect. This should be normally dealt with by
> + * policy_nodemask(), but it's possible to race with cpuset update in
> + * such a way the check therein was true, and then it became false
> + * before we got our cpuset_mems_cookie here.
> + * This assumes that for all allocations, ac->nodemask can come only
> + * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
> + * when it does not intersect with the cpuset restrictions) or the
> + * caller can deal with a violated nodemask.
> + */
> + if (cpusets_enabled() && ac->nodemask &&
> + !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
> + ac->nodemask = NULL;
> + return true;
> + }
> +
> + /*
> + * When updating a task's mems_allowed or mempolicy nodemask, it is
> + * possible to race with parallel threads in such a way that our
> + * allocation can fail while the mask is being updated. If we are about
> + * to fail, check if the cpuset changed during allocation and if so,
> + * retry.
> + */
> + if (read_mems_allowed_retry(cpuset_mems_cookie))
> + return true;
> +
> + return false;
> +}
> +
> static inline struct page *
> __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> struct alloc_context *ac)
> @@ -3856,11 +3889,9 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> &compaction_retries))
> goto retry;
>
> - /*
> - * It's possible we raced with cpuset update so the OOM would be
> - * premature (see below the nopage: label for full explanation).
> - */
> - if (read_mems_allowed_retry(cpuset_mems_cookie))
> +
> + /* Deal with possible cpuset update races before we start OOM killing */
> + if (check_retry_cpuset(cpuset_mems_cookie, ac))
> goto retry_cpuset;
>
> /* Reclaim has failed us, start killing things */
> @@ -3879,14 +3910,8 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> }
>
> nopage:
> - /*
> - * When updating a task's mems_allowed or mempolicy nodemask, it is
> - * possible to race with parallel threads in such a way that our
> - * allocation can fail while the mask is being updated. If we are about
> - * to fail, check if the cpuset changed during allocation and if so,
> - * retry.
> - */
> - if (read_mems_allowed_retry(cpuset_mems_cookie))
> + /* Deal with possible cpuset update races before we fail */
> + if (check_retry_cpuset(cpuset_mems_cookie, ac))
> goto retry_cpuset;
>
> /*
> --
> 2.12.2

--
Michal Hocko
SUSE Labs

\
 
 \ /
  Last update: 2017-05-19 13:52    [W:0.260 / U:2.688 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site