| Message ID | 20240708215537.2630610-1-peterx@redhat.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | mm/migrate: Putback split folios when numa hint migration fails | expand |
On Mon Jul 8, 2024 at 5:55 PM EDT, Peter Xu wrote: > This issue is not from any report yet, but by code observation only. > > This is yet another fix besides Hugh's patch [1] but on relevant code path, > where eager split of folio can happen if the folio is already on deferred > list during a folio migration. > > Here the issue is NUMA path (migrate_misplaced_folio()) may start to > encounter such folio split now even with MR_NUMA_MISPLACED hint applied. > Then when migrate_pages() didn't migrate all the folios, it's possible the > split small folios be put onto the list instead of the original folio. > Then putting back only the head page won't be enough. > > Fix it by putting back all the folios on the list. > > [1] https://lore.kernel.org/all/46c948b4-4dd8-6e03-4c7b-ce4e81cfa536@google.com/ > > Cc: Zi Yan <ziy@nvidia.com> > Cc: Yang Shi <shy828301@gmail.com> > Cc: Hugh Dickins <hughd@google.com> > Cc: Baolin Wang <baolin.wang@linux.alibaba.com> > Cc: Huang Ying <ying.huang@intel.com> > Cc: David Hildenbrand <david@redhat.com> > Fixes: 7262f208ca68 ("mm/migrate: split source folio if it is on deferred split list") > Signed-off-by: Peter Xu <peterx@redhat.com> > --- > > Don't need to copy stable if this can still hit 6.10.. Only smoke tested. > --- > mm/migrate.c | 10 ++-------- > 1 file changed, 2 insertions(+), 8 deletions(-) LGTM. Thanks. Reviewed-by: Zi Yan <ziy@nvidia.com>
On Mon, 8 Jul 2024 17:55:37 -0400 Peter Xu <peterx@redhat.com> wrote: > This issue is not from any report yet, but by code observation only. > > This is yet another fix besides Hugh's patch [1] but on relevant code path, > where eager split of folio can happen if the folio is already on deferred > list during a folio migration. > > Here the issue is NUMA path (migrate_misplaced_folio()) may start to > encounter such folio split now even with MR_NUMA_MISPLACED hint applied. > Then when migrate_pages() didn't migrate all the folios, it's possible the > split small folios be put onto the list instead of the original folio. > Then putting back only the head page won't be enough. > > Fix it by putting back all the folios on the list. mm/migrate.c: In function 'migrate_misplaced_folio': mm/migrate.c:2624:13: error: unused variable 'nr_pages' [-Werror=unused-variable] 2624 | int nr_pages = folio_nr_pages(folio); | ^~~~~~~~ Worrisome. Which kernel version was this tested against? > Don't need to copy stable if this can still hit 6.10.. Only smoke tested. Also worrisome. Are we to take an only-smoke-tested patch which doesn't apply to mainline and which doesn't compile on mm-unstable into mainline based on "only smoke tested"?
On 2024/7/9 05:55, Peter Xu wrote: > This issue is not from any report yet, but by code observation only. > > This is yet another fix besides Hugh's patch [1] but on relevant code path, > where eager split of folio can happen if the folio is already on deferred > list during a folio migration. > > Here the issue is NUMA path (migrate_misplaced_folio()) may start to > encounter such folio split now even with MR_NUMA_MISPLACED hint applied. > Then when migrate_pages() didn't migrate all the folios, it's possible the > split small folios be put onto the list instead of the original folio. > Then putting back only the head page won't be enough. > > Fix it by putting back all the folios on the list. > > [1] https://lore.kernel.org/all/46c948b4-4dd8-6e03-4c7b-ce4e81cfa536@google.com/ > > Cc: Zi Yan <ziy@nvidia.com> > Cc: Yang Shi <shy828301@gmail.com> > Cc: Hugh Dickins <hughd@google.com> > Cc: Baolin Wang <baolin.wang@linux.alibaba.com> > Cc: Huang Ying <ying.huang@intel.com> > Cc: David Hildenbrand <david@redhat.com> > Fixes: 7262f208ca68 ("mm/migrate: split source folio if it is on deferred split list") > Signed-off-by: Peter Xu <peterx@redhat.com> > --- Good catch. With fixing the building issue pointed by Andrew, please feel free to add: Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com> > > Don't need to copy stable if this can still hit 6.10.. Only smoke tested. > --- > mm/migrate.c | 10 ++-------- > 1 file changed, 2 insertions(+), 8 deletions(-) > > diff --git a/mm/migrate.c b/mm/migrate.c > index e10d2445fbd8..20da2595527a 100644 > --- a/mm/migrate.c > +++ b/mm/migrate.c > @@ -2615,14 +2615,8 @@ int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma, > nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio, > NULL, node, MIGRATE_ASYNC, > MR_NUMA_MISPLACED, &nr_succeeded); > - if (nr_remaining) { > - if (!list_empty(&migratepages)) { > - list_del(&folio->lru); > - node_stat_mod_folio(folio, NR_ISOLATED_ANON + > - folio_is_file_lru(folio), -nr_pages); > - folio_putback_lru(folio); > - } > - } > + if (nr_remaining && !list_empty(&migratepages)) Nit: you can drop the '!list_empty(&migratepages)' validation, since putback_movable_pages() can handle this unusual case. > + putback_movable_pages(&migratepages); > if (nr_succeeded) { > count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded); > if (!node_is_toptier(folio_nid(folio)) && node_is_toptier(node))
On Mon, Jul 08, 2024 at 04:04:07PM -0700, Andrew Morton wrote: > On Mon, 8 Jul 2024 17:55:37 -0400 Peter Xu <peterx@redhat.com> wrote: > > > This issue is not from any report yet, but by code observation only. > > > > This is yet another fix besides Hugh's patch [1] but on relevant code path, > > where eager split of folio can happen if the folio is already on deferred > > list during a folio migration. > > > > Here the issue is NUMA path (migrate_misplaced_folio()) may start to > > encounter such folio split now even with MR_NUMA_MISPLACED hint applied. > > Then when migrate_pages() didn't migrate all the folios, it's possible the > > split small folios be put onto the list instead of the original folio. > > Then putting back only the head page won't be enough. > > > > Fix it by putting back all the folios on the list. > > mm/migrate.c: In function 'migrate_misplaced_folio': > mm/migrate.c:2624:13: error: unused variable 'nr_pages' [-Werror=unused-variable] > 2624 | int nr_pages = folio_nr_pages(folio); > | ^~~~~~~~ > > Worrisome. Which kernel version was this tested against? mm-unstable (and on top of a few of my other totally irrelevant patches), and I thought it also applied to mm-stable. Totally missed this warning when still with WERROR=off locally when building against this patch, my apologies. > > > Don't need to copy stable if this can still hit 6.10.. Only smoke tested. > > Also worrisome. Are we to take an only-smoke-tested patch which > doesn't apply to mainline and which doesn't compile on mm-unstable into > mainline based on "only smoke tested"? Hmm so it doesn't apply to mainline.. For the smoke test part, I was not confident to reproduce it, and I just stumbled over it when looking at the real BUG_ON we hit. I thought it might be a good idea to send a patch before everyone forgets about it. I think it is easily overlooked probably because the issue wasn't obvious. IIUC the sympton of hitting it should be that we leak a few of those tail pages even if they're freed in the future from the mappings. I am not sure how much an issue with keep being !lru for them besides the leaked refcounts, perhaps only vmscan won't see them. After all, all these is based on the chance of hitting this case and it should be rare. I don't think I know well enough to say. Considering that nobody yelled after rc5 until now, and this is only something I observed when looking at the more severe issue Hugh fixed.. maybe we should target this for next release, then stablize it and wait for a backport to 6.10? Thanks,
On Tue, Jul 09, 2024 at 09:48:54AM +0800, Baolin Wang wrote: > > > On 2024/7/9 05:55, Peter Xu wrote: > > This issue is not from any report yet, but by code observation only. > > > > This is yet another fix besides Hugh's patch [1] but on relevant code path, > > where eager split of folio can happen if the folio is already on deferred > > list during a folio migration. > > > > Here the issue is NUMA path (migrate_misplaced_folio()) may start to > > encounter such folio split now even with MR_NUMA_MISPLACED hint applied. > > Then when migrate_pages() didn't migrate all the folios, it's possible the > > split small folios be put onto the list instead of the original folio. > > Then putting back only the head page won't be enough. > > > > Fix it by putting back all the folios on the list. > > > > [1] https://lore.kernel.org/all/46c948b4-4dd8-6e03-4c7b-ce4e81cfa536@google.com/ > > > > Cc: Zi Yan <ziy@nvidia.com> > > Cc: Yang Shi <shy828301@gmail.com> > > Cc: Hugh Dickins <hughd@google.com> > > Cc: Baolin Wang <baolin.wang@linux.alibaba.com> > > Cc: Huang Ying <ying.huang@intel.com> > > Cc: David Hildenbrand <david@redhat.com> > > Fixes: 7262f208ca68 ("mm/migrate: split source folio if it is on deferred split list") > > Signed-off-by: Peter Xu <peterx@redhat.com> > > --- > > Good catch. With fixing the building issue pointed by Andrew, please feel > free to add: > Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com> > > > > > Don't need to copy stable if this can still hit 6.10.. Only smoke tested. > > --- > > mm/migrate.c | 10 ++-------- > > 1 file changed, 2 insertions(+), 8 deletions(-) > > > > diff --git a/mm/migrate.c b/mm/migrate.c > > index e10d2445fbd8..20da2595527a 100644 > > --- a/mm/migrate.c > > +++ b/mm/migrate.c > > @@ -2615,14 +2615,8 @@ int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma, > > nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio, > > NULL, node, MIGRATE_ASYNC, > > MR_NUMA_MISPLACED, &nr_succeeded); > > - if (nr_remaining) { > > - if (!list_empty(&migratepages)) { > > - list_del(&folio->lru); > > - node_stat_mod_folio(folio, NR_ISOLATED_ANON + > > - folio_is_file_lru(folio), -nr_pages); > > - folio_putback_lru(folio); > > - } > > - } > > + if (nr_remaining && !list_empty(&migratepages)) > > Nit: you can drop the '!list_empty(&migratepages)' validation, since > putback_movable_pages() can handle this unusual case. Sure, considering that it should normally be !empty when the first check passed. Though to make this simple for now, I assume we can keep what has been queued in Andrew's tree. It isn't so bad either to double check the list to avoid a function call if possible, I think. Thanks for the comment,
On Tue, 9 Jul 2024 10:47:19 -0400 Peter Xu <peterx@redhat.com> wrote: > Considering that nobody yelled after rc5 until now, and this is only > something I observed when looking at the more severe issue Hugh fixed.. > maybe we should target this for next release, then stablize it and wait for > a backport to 6.10? Sure, let's do that.
diff --git a/mm/migrate.c b/mm/migrate.c index e10d2445fbd8..20da2595527a 100644 --- a/mm/migrate.c +++ b/mm/migrate.c @@ -2615,14 +2615,8 @@ int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma, nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio, NULL, node, MIGRATE_ASYNC, MR_NUMA_MISPLACED, &nr_succeeded); - if (nr_remaining) { - if (!list_empty(&migratepages)) { - list_del(&folio->lru); - node_stat_mod_folio(folio, NR_ISOLATED_ANON + - folio_is_file_lru(folio), -nr_pages); - folio_putback_lru(folio); - } - } + if (nr_remaining && !list_empty(&migratepages)) + putback_movable_pages(&migratepages); if (nr_succeeded) { count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded); if (!node_is_toptier(folio_nid(folio)) && node_is_toptier(node))
This issue is not from any report yet, but by code observation only. This is yet another fix besides Hugh's patch [1] but on relevant code path, where eager split of folio can happen if the folio is already on deferred list during a folio migration. Here the issue is NUMA path (migrate_misplaced_folio()) may start to encounter such folio split now even with MR_NUMA_MISPLACED hint applied. Then when migrate_pages() didn't migrate all the folios, it's possible the split small folios be put onto the list instead of the original folio. Then putting back only the head page won't be enough. Fix it by putting back all the folios on the list. [1] https://lore.kernel.org/all/46c948b4-4dd8-6e03-4c7b-ce4e81cfa536@google.com/ Cc: Zi Yan <ziy@nvidia.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Hugh Dickins <hughd@google.com> Cc: Baolin Wang <baolin.wang@linux.alibaba.com> Cc: Huang Ying <ying.huang@intel.com> Cc: David Hildenbrand <david@redhat.com> Fixes: 7262f208ca68 ("mm/migrate: split source folio if it is on deferred split list") Signed-off-by: Peter Xu <peterx@redhat.com> --- Don't need to copy stable if this can still hit 6.10.. Only smoke tested. --- mm/migrate.c | 10 ++-------- 1 file changed, 2 insertions(+), 8 deletions(-)