Message ID | 20240221085036.105621-1-21cnbao@gmail.com (mailing list archive) |
---|---|
State | New |
Headers | show |
Series | madvise:madvise_cold_or_pageout_pte_range(): allow split while folio_estimated_sharers = 0 | expand |
On 21.02.24 09:50, Barry Song wrote: > From: Barry Song <v-songbaohua@oppo.com> > > The purpose is stopping splitting large folios whose mapcount are 2 or > above. Folios whose estimated_shares = 0 should be still perfect and > even better candidates than estimated_shares = 1. > > Consider a pte-mapped large folio with 16 subpages, if we unmap 1-15, > the current code will split folios and reclaim them while madvise goes > on this folio; but if we unmap subpage 0, we will keep this folio and > break. This is weird. > > For pmd-mapped large folios, we can still use "= 1" as the condition > as anyway we have the entire map for it. So this patch doesn't change > the condition for pmd-mapped large folios. > This also explains why we had been using "= 1" for both pmd-mapped and > pte-mapped large folios before commit 07e8c82b5eff ("madvise: convert > madvise_cold_or_pageout_pte_range() to use folios"), because in the > past, we used the mapcount of the specific subpage, since the subpage > had pte present, its mapcount wouldn't be 0. > > The problem can be quite easily reproduced by writing a small program, > unmapping the first subpage of a pte-mapped large folio vs. unmapping > anyone other than the first subpage. > > Fixes: 2f406263e3e9 ("madvise:madvise_cold_or_pageout_pte_range(): don't use mapcount() against large folio for sharing check") > Cc: Yin Fengwei <fengwei.yin@intel.com> > Cc: Yu Zhao <yuzhao@google.com> > Cc: Ryan Roberts <ryan.roberts@arm.com> > Cc: David Hildenbrand <david@redhat.com> > Cc: Kefeng Wang <wangkefeng.wang@huawei.com> > Cc: Matthew Wilcox <willy@infradead.org> > Cc: Minchan Kim <minchan@kernel.org> > Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com> > Cc: Yang Shi <shy828301@gmail.com> > Signed-off-by: Barry Song <v-songbaohua@oppo.com> > --- > mm/madvise.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/mm/madvise.c b/mm/madvise.c > index cfa5e7288261..abde3edb04f0 100644 > --- a/mm/madvise.c > +++ b/mm/madvise.c > @@ -453,7 +453,7 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd, > if (folio_test_large(folio)) { > int err; > > - if (folio_estimated_sharers(folio) != 1) > + if (folio_estimated_sharers(folio) > 1) > break; > if (pageout_anon_only_filter && !folio_test_anon(folio)) > break; That's also what I do in https://lkml.kernel.org/r/20231124132626.235350-4-david@redhat.com I'll revive that soon.
On 21.02.24 10:15, David Hildenbrand wrote: > On 21.02.24 09:50, Barry Song wrote: >> From: Barry Song <v-songbaohua@oppo.com> >> >> The purpose is stopping splitting large folios whose mapcount are 2 or >> above. Folios whose estimated_shares = 0 should be still perfect and >> even better candidates than estimated_shares = 1. >> >> Consider a pte-mapped large folio with 16 subpages, if we unmap 1-15, >> the current code will split folios and reclaim them while madvise goes >> on this folio; but if we unmap subpage 0, we will keep this folio and >> break. This is weird. >> >> For pmd-mapped large folios, we can still use "= 1" as the condition >> as anyway we have the entire map for it. So this patch doesn't change >> the condition for pmd-mapped large folios. >> This also explains why we had been using "= 1" for both pmd-mapped and >> pte-mapped large folios before commit 07e8c82b5eff ("madvise: convert >> madvise_cold_or_pageout_pte_range() to use folios"), because in the >> past, we used the mapcount of the specific subpage, since the subpage >> had pte present, its mapcount wouldn't be 0. >> >> The problem can be quite easily reproduced by writing a small program, >> unmapping the first subpage of a pte-mapped large folio vs. unmapping >> anyone other than the first subpage. >> >> Fixes: 2f406263e3e9 ("madvise:madvise_cold_or_pageout_pte_range(): don't use mapcount() against large folio for sharing check") >> Cc: Yin Fengwei <fengwei.yin@intel.com> >> Cc: Yu Zhao <yuzhao@google.com> >> Cc: Ryan Roberts <ryan.roberts@arm.com> >> Cc: David Hildenbrand <david@redhat.com> >> Cc: Kefeng Wang <wangkefeng.wang@huawei.com> >> Cc: Matthew Wilcox <willy@infradead.org> >> Cc: Minchan Kim <minchan@kernel.org> >> Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com> >> Cc: Yang Shi <shy828301@gmail.com> >> Signed-off-by: Barry Song <v-songbaohua@oppo.com> >> --- >> mm/madvise.c | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/mm/madvise.c b/mm/madvise.c >> index cfa5e7288261..abde3edb04f0 100644 >> --- a/mm/madvise.c >> +++ b/mm/madvise.c >> @@ -453,7 +453,7 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd, >> if (folio_test_large(folio)) { >> int err; >> >> - if (folio_estimated_sharers(folio) != 1) >> + if (folio_estimated_sharers(folio) > 1) >> break; >> if (pageout_anon_only_filter && !folio_test_anon(folio)) >> break; > > That's also what I do in > > https://lkml.kernel.org/r/20231124132626.235350-4-david@redhat.com > > I'll revive that soon. Forgot to add: we can pull this in early. Reviewed-by: David Hildenbrand <david@redhat.com> (I'll do the simple folio_estimated_sharers() to folio_mapped_shared() conversion first and optimize with total mapcount separately)
On Wed, Feb 21, 2024 at 12:50 AM Barry Song <21cnbao@gmail.com> wrote: > > From: Barry Song <v-songbaohua@oppo.com> > > The purpose is stopping splitting large folios whose mapcount are 2 or > above. Folios whose estimated_shares = 0 should be still perfect and > even better candidates than estimated_shares = 1. > > Consider a pte-mapped large folio with 16 subpages, if we unmap 1-15, > the current code will split folios and reclaim them while madvise goes > on this folio; but if we unmap subpage 0, we will keep this folio and > break. This is weird. > For pmd-mapped large folios, we can still use "= 1" as the condition > as anyway we have the entire map for it. So this patch doesn't change > the condition for pmd-mapped large folios. > This also explains why we had been using "= 1" for both pmd-mapped and > pte-mapped large folios before commit 07e8c82b5eff ("madvise: convert > madvise_cold_or_pageout_pte_range() to use folios"), because in the > past, we used the mapcount of the specific subpage, since the subpage > had pte present, its mapcount wouldn't be 0. > The problem can be quite easily reproduced by writing a small program, > unmapping the first subpage of a pte-mapped large folio vs. unmapping > anyone other than the first subpage. > > Fixes: 2f406263e3e9 ("madvise:madvise_cold_or_pageout_pte_range(): don't use mapcount() against large folio for sharing check") > Cc: Yin Fengwei <fengwei.yin@intel.com> > Cc: Yu Zhao <yuzhao@google.com> > Cc: Ryan Roberts <ryan.roberts@arm.com> > Cc: David Hildenbrand <david@redhat.com> > Cc: Kefeng Wang <wangkefeng.wang@huawei.com> > Cc: Matthew Wilcox <willy@infradead.org> > Cc: Minchan Kim <minchan@kernel.org> > Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com> > Cc: Yang Shi <shy828301@gmail.com> > Signed-off-by: Barry Song <v-songbaohua@oppo.com> Reviewed-by: Vishal Moola (Oracle) <vishal.moola@gmail.com> > --- > mm/madvise.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/mm/madvise.c b/mm/madvise.c > index cfa5e7288261..abde3edb04f0 100644 > --- a/mm/madvise.c > +++ b/mm/madvise.c > @@ -453,7 +453,7 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd, > if (folio_test_large(folio)) { > int err; > > - if (folio_estimated_sharers(folio) != 1) > + if (folio_estimated_sharers(folio) > 1) > break; > if (pageout_anon_only_filter && !folio_test_anon(folio)) > break; > -- > 2.34.1 >
On 21/02/2024 08:50, Barry Song wrote: > From: Barry Song <v-songbaohua@oppo.com> > > The purpose is stopping splitting large folios whose mapcount are 2 or > above. Folios whose estimated_shares = 0 should be still perfect and > even better candidates than estimated_shares = 1. > > Consider a pte-mapped large folio with 16 subpages, if we unmap 1-15, > the current code will split folios and reclaim them while madvise goes > on this folio; but if we unmap subpage 0, we will keep this folio and > break. This is weird. > > For pmd-mapped large folios, we can still use "= 1" as the condition > as anyway we have the entire map for it. So this patch doesn't change > the condition for pmd-mapped large folios. > This also explains why we had been using "= 1" for both pmd-mapped and > pte-mapped large folios before commit 07e8c82b5eff ("madvise: convert > madvise_cold_or_pageout_pte_range() to use folios"), because in the > past, we used the mapcount of the specific subpage, since the subpage > had pte present, its mapcount wouldn't be 0. > > The problem can be quite easily reproduced by writing a small program, > unmapping the first subpage of a pte-mapped large folio vs. unmapping > anyone other than the first subpage. > > Fixes: 2f406263e3e9 ("madvise:madvise_cold_or_pageout_pte_range(): don't use mapcount() against large folio for sharing check") > Cc: Yin Fengwei <fengwei.yin@intel.com> > Cc: Yu Zhao <yuzhao@google.com> > Cc: Ryan Roberts <ryan.roberts@arm.com> > Cc: David Hildenbrand <david@redhat.com> > Cc: Kefeng Wang <wangkefeng.wang@huawei.com> > Cc: Matthew Wilcox <willy@infradead.org> > Cc: Minchan Kim <minchan@kernel.org> > Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com> > Cc: Yang Shi <shy828301@gmail.com> > Signed-off-by: Barry Song <v-songbaohua@oppo.com> > --- > mm/madvise.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/mm/madvise.c b/mm/madvise.c > index cfa5e7288261..abde3edb04f0 100644 > --- a/mm/madvise.c > +++ b/mm/madvise.c > @@ -453,7 +453,7 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd, > if (folio_test_large(folio)) { > int err; > > - if (folio_estimated_sharers(folio) != 1) > + if (folio_estimated_sharers(folio) > 1) > break; > if (pageout_anon_only_filter && !folio_test_anon(folio)) > break; I wonder if we should change all the instances: folio_estimated_sharers() != 1 -> folio_estimated_sharers() > 1 folio_estimated_sharers() == 1 -> folio_estimated_sharers() <= 1 It shouldn't cause a problem for the pmd case, and there are definitely other cases where it will help. e.g. madvise_free_pte_range(). Regardless: Reviewed-by: Ryan Roberts <ryan.roberts@arm.com>
On 26.02.24 14:46, Ryan Roberts wrote: > On 21/02/2024 08:50, Barry Song wrote: >> From: Barry Song <v-songbaohua@oppo.com> >> >> The purpose is stopping splitting large folios whose mapcount are 2 or >> above. Folios whose estimated_shares = 0 should be still perfect and >> even better candidates than estimated_shares = 1. >> >> Consider a pte-mapped large folio with 16 subpages, if we unmap 1-15, >> the current code will split folios and reclaim them while madvise goes >> on this folio; but if we unmap subpage 0, we will keep this folio and >> break. This is weird. >> >> For pmd-mapped large folios, we can still use "= 1" as the condition >> as anyway we have the entire map for it. So this patch doesn't change >> the condition for pmd-mapped large folios. >> This also explains why we had been using "= 1" for both pmd-mapped and >> pte-mapped large folios before commit 07e8c82b5eff ("madvise: convert >> madvise_cold_or_pageout_pte_range() to use folios"), because in the >> past, we used the mapcount of the specific subpage, since the subpage >> had pte present, its mapcount wouldn't be 0. >> >> The problem can be quite easily reproduced by writing a small program, >> unmapping the first subpage of a pte-mapped large folio vs. unmapping >> anyone other than the first subpage. >> >> Fixes: 2f406263e3e9 ("madvise:madvise_cold_or_pageout_pte_range(): don't use mapcount() against large folio for sharing check") >> Cc: Yin Fengwei <fengwei.yin@intel.com> >> Cc: Yu Zhao <yuzhao@google.com> >> Cc: Ryan Roberts <ryan.roberts@arm.com> >> Cc: David Hildenbrand <david@redhat.com> >> Cc: Kefeng Wang <wangkefeng.wang@huawei.com> >> Cc: Matthew Wilcox <willy@infradead.org> >> Cc: Minchan Kim <minchan@kernel.org> >> Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com> >> Cc: Yang Shi <shy828301@gmail.com> >> Signed-off-by: Barry Song <v-songbaohua@oppo.com> >> --- >> mm/madvise.c | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/mm/madvise.c b/mm/madvise.c >> index cfa5e7288261..abde3edb04f0 100644 >> --- a/mm/madvise.c >> +++ b/mm/madvise.c >> @@ -453,7 +453,7 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd, >> if (folio_test_large(folio)) { >> int err; >> >> - if (folio_estimated_sharers(folio) != 1) >> + if (folio_estimated_sharers(folio) > 1) >> break; >> if (pageout_anon_only_filter && !folio_test_anon(folio)) >> break; > > I wonder if we should change all the instances: > > folio_estimated_sharers() != 1 -> folio_estimated_sharers() > 1 > folio_estimated_sharers() == 1 -> folio_estimated_sharers() <= 1 I'll send out something that wraps that in folio_mapped_shared() later today or tomorrow.
On Tue, Feb 27, 2024 at 2:46 AM Ryan Roberts <ryan.roberts@arm.com> wrote: > > On 21/02/2024 08:50, Barry Song wrote: > > From: Barry Song <v-songbaohua@oppo.com> > > > > The purpose is stopping splitting large folios whose mapcount are 2 or > > above. Folios whose estimated_shares = 0 should be still perfect and > > even better candidates than estimated_shares = 1. > > > > Consider a pte-mapped large folio with 16 subpages, if we unmap 1-15, > > the current code will split folios and reclaim them while madvise goes > > on this folio; but if we unmap subpage 0, we will keep this folio and > > break. This is weird. > > > > For pmd-mapped large folios, we can still use "= 1" as the condition > > as anyway we have the entire map for it. So this patch doesn't change > > the condition for pmd-mapped large folios. > > This also explains why we had been using "= 1" for both pmd-mapped and > > pte-mapped large folios before commit 07e8c82b5eff ("madvise: convert > > madvise_cold_or_pageout_pte_range() to use folios"), because in the > > past, we used the mapcount of the specific subpage, since the subpage > > had pte present, its mapcount wouldn't be 0. > > > > The problem can be quite easily reproduced by writing a small program, > > unmapping the first subpage of a pte-mapped large folio vs. unmapping > > anyone other than the first subpage. > > > > Fixes: 2f406263e3e9 ("madvise:madvise_cold_or_pageout_pte_range(): don't use mapcount() against large folio for sharing check") > > Cc: Yin Fengwei <fengwei.yin@intel.com> > > Cc: Yu Zhao <yuzhao@google.com> > > Cc: Ryan Roberts <ryan.roberts@arm.com> > > Cc: David Hildenbrand <david@redhat.com> > > Cc: Kefeng Wang <wangkefeng.wang@huawei.com> > > Cc: Matthew Wilcox <willy@infradead.org> > > Cc: Minchan Kim <minchan@kernel.org> > > Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com> > > Cc: Yang Shi <shy828301@gmail.com> > > Signed-off-by: Barry Song <v-songbaohua@oppo.com> > > --- > > mm/madvise.c | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > diff --git a/mm/madvise.c b/mm/madvise.c > > index cfa5e7288261..abde3edb04f0 100644 > > --- a/mm/madvise.c > > +++ b/mm/madvise.c > > @@ -453,7 +453,7 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd, > > if (folio_test_large(folio)) { > > int err; > > > > - if (folio_estimated_sharers(folio) != 1) > > + if (folio_estimated_sharers(folio) > 1) > > break; > > if (pageout_anon_only_filter && !folio_test_anon(folio)) > > break; > > I wonder if we should change all the instances: > > folio_estimated_sharers() != 1 -> folio_estimated_sharers() > 1 > folio_estimated_sharers() == 1 -> folio_estimated_sharers() <= 1 > > It shouldn't cause a problem for the pmd case, and there are definitely other > cases where it will help. e.g. madvise_free_pte_range(). right. My test case covered PAGEOUT only and I agree madvise_free and others have exactly the same issue. for pmd case, it doesn't matter whether we change the condition or not because we have already pmd-mapped in the page table. And good to know David will have a wrapper in folio_mapped_shared() to more widely address this issue. > > Regardless: > > Reviewed-by: Ryan Roberts <ryan.roberts@arm.com> > Thanks though we might have missed your tag as this one has been in mm-stable. Best regards, Barry
On 26/02/2024 21:17, Barry Song wrote: > On Tue, Feb 27, 2024 at 2:46 AM Ryan Roberts <ryan.roberts@arm.com> wrote: >> >> On 21/02/2024 08:50, Barry Song wrote: >>> From: Barry Song <v-songbaohua@oppo.com> >>> >>> The purpose is stopping splitting large folios whose mapcount are 2 or >>> above. Folios whose estimated_shares = 0 should be still perfect and >>> even better candidates than estimated_shares = 1. >>> >>> Consider a pte-mapped large folio with 16 subpages, if we unmap 1-15, >>> the current code will split folios and reclaim them while madvise goes >>> on this folio; but if we unmap subpage 0, we will keep this folio and >>> break. This is weird. >>> >>> For pmd-mapped large folios, we can still use "= 1" as the condition >>> as anyway we have the entire map for it. So this patch doesn't change >>> the condition for pmd-mapped large folios. >>> This also explains why we had been using "= 1" for both pmd-mapped and >>> pte-mapped large folios before commit 07e8c82b5eff ("madvise: convert >>> madvise_cold_or_pageout_pte_range() to use folios"), because in the >>> past, we used the mapcount of the specific subpage, since the subpage >>> had pte present, its mapcount wouldn't be 0. >>> >>> The problem can be quite easily reproduced by writing a small program, >>> unmapping the first subpage of a pte-mapped large folio vs. unmapping >>> anyone other than the first subpage. >>> >>> Fixes: 2f406263e3e9 ("madvise:madvise_cold_or_pageout_pte_range(): don't use mapcount() against large folio for sharing check") >>> Cc: Yin Fengwei <fengwei.yin@intel.com> >>> Cc: Yu Zhao <yuzhao@google.com> >>> Cc: Ryan Roberts <ryan.roberts@arm.com> >>> Cc: David Hildenbrand <david@redhat.com> >>> Cc: Kefeng Wang <wangkefeng.wang@huawei.com> >>> Cc: Matthew Wilcox <willy@infradead.org> >>> Cc: Minchan Kim <minchan@kernel.org> >>> Cc: Vishal Moola (Oracle) <vishal.moola@gmail.com> >>> Cc: Yang Shi <shy828301@gmail.com> >>> Signed-off-by: Barry Song <v-songbaohua@oppo.com> >>> --- >>> mm/madvise.c | 2 +- >>> 1 file changed, 1 insertion(+), 1 deletion(-) >>> >>> diff --git a/mm/madvise.c b/mm/madvise.c >>> index cfa5e7288261..abde3edb04f0 100644 >>> --- a/mm/madvise.c >>> +++ b/mm/madvise.c >>> @@ -453,7 +453,7 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd, >>> if (folio_test_large(folio)) { >>> int err; >>> >>> - if (folio_estimated_sharers(folio) != 1) >>> + if (folio_estimated_sharers(folio) > 1) >>> break; >>> if (pageout_anon_only_filter && !folio_test_anon(folio)) >>> break; >> >> I wonder if we should change all the instances: >> >> folio_estimated_sharers() != 1 -> folio_estimated_sharers() > 1 >> folio_estimated_sharers() == 1 -> folio_estimated_sharers() <= 1 >> >> It shouldn't cause a problem for the pmd case, and there are definitely other >> cases where it will help. e.g. madvise_free_pte_range(). > > right. My test case covered PAGEOUT only and I agree madvise_free and > others have > exactly the same issue. for pmd case, it doesn't matter whether we > change the condition > or not because we have already pmd-mapped in the page table. > > And good to know David will have a wrapper in folio_mapped_shared() to more > widely address this issue. > >> >> Regardless: >> >> Reviewed-by: Ryan Roberts <ryan.roberts@arm.com> >> > > Thanks though we might have missed your tag as this one has been > in mm-stable. No problem! I've been out on holiday so a bit behind on where everything is. > > Best regards, > Barry
diff --git a/mm/madvise.c b/mm/madvise.c index cfa5e7288261..abde3edb04f0 100644 --- a/mm/madvise.c +++ b/mm/madvise.c @@ -453,7 +453,7 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd, if (folio_test_large(folio)) { int err; - if (folio_estimated_sharers(folio) != 1) + if (folio_estimated_sharers(folio) > 1) break; if (pageout_anon_only_filter && !folio_test_anon(folio)) break;