| Message ID | 20221019134156.zjyyn5aownakvztf@techsingularity.net (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [RFC] mm/huge_memory: Do not clobber swp_entry_t during THP split | expand |
On Wed, Oct 19, 2022 at 6:42 AM Mel Gorman <mgorman@techsingularity.net> wrote: > > The following has been observed when running stressng mmap since commit > b653db77350c ("mm: Clear page->private when splitting or migrating a page") > > watchdog: BUG: soft lockup - CPU#75 stuck for 26s! [stress-ng:9546] > CPU: 75 PID: 9546 Comm: stress-ng Tainted: G E 6.0.0-revert-b653db77-fix+ #29 0357d79b60fb09775f678e4f3f64ef0579ad1374 > Hardware name: SGI.COM C2112-4GP3/X10DRT-P-Series, BIOS 2.0a 05/09/2016 > RIP: 0010:xas_descend+0x28/0x80 > Code: cc cc 0f b6 0e 48 8b 57 08 48 d3 ea 83 e2 3f 89 d0 48 83 c0 04 48 8b 44 c6 08 48 89 77 18 48 89 c1 83 e1 03 48 83 f9 02 75 08 <48> 3d fd 00 00 00 76 08 88 57 12 c3 cc cc cc cc 48 c1 e8 02 89 c2 > RSP: 0018:ffffbbf02a2236a8 EFLAGS: 00000246 > RAX: ffff9cab7d6a0002 RBX: ffffe04b0af88040 RCX: 0000000000000002 > RDX: 0000000000000030 RSI: ffff9cab60509b60 RDI: ffffbbf02a2236c0 > RBP: 0000000000000000 R08: ffff9cab60509b60 R09: ffffbbf02a2236c0 > R10: 0000000000000001 R11: ffffbbf02a223698 R12: 0000000000000000 > R13: ffff9cab4e28da80 R14: 0000000000039c01 R15: ffff9cab4e28da88 > FS: 00007fab89b85e40(0000) GS:ffff9cea3fcc0000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 00007fab84e00000 CR3: 00000040b73a4003 CR4: 00000000003706e0 > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 > DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 > Call Trace: > <TASK> > xas_load+0x3a/0x50 > __filemap_get_folio+0x80/0x370 > ? put_swap_page+0x163/0x360 > pagecache_get_page+0x13/0x90 > __try_to_reclaim_swap+0x50/0x190 > scan_swap_map_slots+0x31e/0x670 > get_swap_pages+0x226/0x3c0 > folio_alloc_swap+0x1cc/0x240 > add_to_swap+0x14/0x70 > shrink_page_list+0x968/0xbc0 > reclaim_page_list+0x70/0xf0 > reclaim_pages+0xdd/0x120 > madvise_cold_or_pageout_pte_range+0x814/0xf30 > walk_pgd_range+0x637/0xa30 > __walk_page_range+0x142/0x170 > walk_page_range+0x146/0x170 > madvise_pageout+0xb7/0x280 > ? asm_common_interrupt+0x22/0x40 > madvise_vma_behavior+0x3b7/0xac0 > ? find_vma+0x4a/0x70 > ? find_vma+0x64/0x70 > ? madvise_vma_anon_name+0x40/0x40 > madvise_walk_vmas+0xa6/0x130 > do_madvise+0x2f4/0x360 > __x64_sys_madvise+0x26/0x30 > do_syscall_64+0x5b/0x80 > ? do_syscall_64+0x67/0x80 > ? syscall_exit_to_user_mode+0x17/0x40 > ? do_syscall_64+0x67/0x80 > ? syscall_exit_to_user_mode+0x17/0x40 > ? do_syscall_64+0x67/0x80 > ? do_syscall_64+0x67/0x80 > ? common_interrupt+0x8b/0xa0 > entry_SYSCALL_64_after_hwframe+0x63/0xcd > > The problem can be reproduced with the mmtests config > config-workload-stressng-mmap. It does not always happen and when it > triggers is variable but it has happened on multiple machines. > > The intent of commit b653db77350c patch was to avoid the case where > PG_private is clear but folio->private is not-NULL. However, THP tail > pages uses page->private for "swp_entry_t if folio_test_swapcache()" as > stated in the documentation for struct folio. This patch only clobbers > page->private for tail pages if the head page was not in swapcache and > warns once if page->private had an unexpected value. It looks like the same issue fixed by https://lore.kernel.org/linux-mm/20220906190602.1626037-1-bfoster@redhat.com/ > > Fixes: b653db77350c ("mm: Clear page->private when splitting or migrating a page") > Signed-off-by: Mel Gorman <mgorman@techsingularity.net> > --- > mm/huge_memory.c | 11 ++++++++++- > 1 file changed, 10 insertions(+), 1 deletion(-) > > diff --git a/mm/huge_memory.c b/mm/huge_memory.c > index 1cc4a5f4791e..03fc7e5edf07 100644 > --- a/mm/huge_memory.c > +++ b/mm/huge_memory.c > @@ -2455,7 +2455,16 @@ static void __split_huge_page_tail(struct page *head, int tail, > page_tail); > page_tail->mapping = head->mapping; > page_tail->index = head->index + tail; > - page_tail->private = 0; > + > + /* > + * page->private should not be set in tail pages with the exception > + * of swap cache pages that store the swp_entry_t in tail pages. > + * Fix up and warn once if private is unexpectedly set. > + */ > + if (!folio_test_swapcache(page_folio(head))) { > + VM_WARN_ON_ONCE_PAGE(page_tail->private != 0, head); > + page_tail->private = 0; > + } > > /* Page flags must be visible before we make the page non-compound. */ > smp_wmb(); >
On Wed, 19 Oct 2022 11:17:14 -0700 Yang Shi <shy828301@gmail.com> wrote: > > The intent of commit b653db77350c patch was to avoid the case where > > PG_private is clear but folio->private is not-NULL. However, THP tail > > pages uses page->private for "swp_entry_t if folio_test_swapcache()" as > > stated in the documentation for struct folio. This patch only clobbers > > page->private for tail pages if the head page was not in swapcache and > > warns once if page->private had an unexpected value. > > It looks like the same issue fixed by > https://lore.kernel.org/linux-mm/20220906190602.1626037-1-bfoster@redhat.com/ It is. As I asked earlier this week, what about reverting b653db77350c? Why do we care about the value of ->private for non-PG_private pages?
Adding Brian to cc On Wed, Oct 19, 2022 at 04:18:10PM -0700, Andrew Morton wrote: > On Wed, 19 Oct 2022 11:17:14 -0700 Yang Shi <shy828301@gmail.com> wrote: > > > > The intent of commit b653db77350c patch was to avoid the case where > > > PG_private is clear but folio->private is not-NULL. However, THP tail > > > pages uses page->private for "swp_entry_t if folio_test_swapcache()" as > > > stated in the documentation for struct folio. This patch only clobbers > > > page->private for tail pages if the head page was not in swapcache and > > > warns once if page->private had an unexpected value. > > > > It looks like the same issue fixed by > > https://lore.kernel.org/linux-mm/20220906190602.1626037-1-bfoster@redhat.com/ > > It is. > Yep, based on Brian's changelog, it was the same workload that triggered it as it happens to stress the corner case that hits the bug. > As I asked earlier this week, what about reverting b653db77350c? Why > do we care about the value of ->private for non-PG_private pages? I don't think we do care but based on the changelog of b653db77350c, it's part of an effort to either remove the PG_private bit or is a preparation step for casting page to a meaningful type based on context but only Matthew can tell us his motivation. There at least is some value to identifying cases where a referenced page has valid information in page->private that is not reflected in the flags.
cc Kirill On Thu, Oct 20, 2022 at 09:52:14AM +0100, Mel Gorman wrote: > Adding Brian to cc > > On Wed, Oct 19, 2022 at 04:18:10PM -0700, Andrew Morton wrote: > > On Wed, 19 Oct 2022 11:17:14 -0700 Yang Shi <shy828301@gmail.com> wrote: > > > > > > The intent of commit b653db77350c patch was to avoid the case where > > > > PG_private is clear but folio->private is not-NULL. However, THP tail > > > > pages uses page->private for "swp_entry_t if folio_test_swapcache()" as > > > > stated in the documentation for struct folio. This patch only clobbers > > > > page->private for tail pages if the head page was not in swapcache and > > > > warns once if page->private had an unexpected value. > > > > > > It looks like the same issue fixed by > > > https://lore.kernel.org/linux-mm/20220906190602.1626037-1-bfoster@redhat.com/ > > > > It is. > > > > Yep, based on Brian's changelog, it was the same workload that triggered > it as it happens to stress the corner case that hits the bug. > > > As I asked earlier this week, what about reverting b653db77350c? Why > > do we care about the value of ->private for non-PG_private pages? > > I don't think we do care but based on the changelog of b653db77350c, it's > part of an effort to either remove the PG_private bit or is a preparation > step for casting page to a meaningful type based on context but only Matthew > can tell us his motivation. There at least is some value to identifying > cases where a referenced page has valid information in page->private that > is not reflected in the flags. > Thanks. It would have been nice to have received some feedback on the patch I had posted 6 weeks or so ago ;), but regardless yours is better and includes the comment Kirill asked for (and also appears to be added to the hotfixes tree), so FWIW: Acked-by: Brian Foster <bfoster@redhat.com> > -- > Mel Gorman > SUSE Labs >
diff --git a/mm/huge_memory.c b/mm/huge_memory.c index 1cc4a5f4791e..03fc7e5edf07 100644 --- a/mm/huge_memory.c +++ b/mm/huge_memory.c @@ -2455,7 +2455,16 @@ static void __split_huge_page_tail(struct page *head, int tail, page_tail); page_tail->mapping = head->mapping; page_tail->index = head->index + tail; - page_tail->private = 0; + + /* + * page->private should not be set in tail pages with the exception + * of swap cache pages that store the swp_entry_t in tail pages. + * Fix up and warn once if private is unexpectedly set. + */ + if (!folio_test_swapcache(page_folio(head))) { + VM_WARN_ON_ONCE_PAGE(page_tail->private != 0, head); + page_tail->private = 0; + } /* Page flags must be visible before we make the page non-compound. */ smp_wmb();
The following has been observed when running stressng mmap since commit b653db77350c ("mm: Clear page->private when splitting or migrating a page") watchdog: BUG: soft lockup - CPU#75 stuck for 26s! [stress-ng:9546] CPU: 75 PID: 9546 Comm: stress-ng Tainted: G E 6.0.0-revert-b653db77-fix+ #29 0357d79b60fb09775f678e4f3f64ef0579ad1374 Hardware name: SGI.COM C2112-4GP3/X10DRT-P-Series, BIOS 2.0a 05/09/2016 RIP: 0010:xas_descend+0x28/0x80 Code: cc cc 0f b6 0e 48 8b 57 08 48 d3 ea 83 e2 3f 89 d0 48 83 c0 04 48 8b 44 c6 08 48 89 77 18 48 89 c1 83 e1 03 48 83 f9 02 75 08 <48> 3d fd 00 00 00 76 08 88 57 12 c3 cc cc cc cc 48 c1 e8 02 89 c2 RSP: 0018:ffffbbf02a2236a8 EFLAGS: 00000246 RAX: ffff9cab7d6a0002 RBX: ffffe04b0af88040 RCX: 0000000000000002 RDX: 0000000000000030 RSI: ffff9cab60509b60 RDI: ffffbbf02a2236c0 RBP: 0000000000000000 R08: ffff9cab60509b60 R09: ffffbbf02a2236c0 R10: 0000000000000001 R11: ffffbbf02a223698 R12: 0000000000000000 R13: ffff9cab4e28da80 R14: 0000000000039c01 R15: ffff9cab4e28da88 FS: 00007fab89b85e40(0000) GS:ffff9cea3fcc0000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007fab84e00000 CR3: 00000040b73a4003 CR4: 00000000003706e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: <TASK> xas_load+0x3a/0x50 __filemap_get_folio+0x80/0x370 ? put_swap_page+0x163/0x360 pagecache_get_page+0x13/0x90 __try_to_reclaim_swap+0x50/0x190 scan_swap_map_slots+0x31e/0x670 get_swap_pages+0x226/0x3c0 folio_alloc_swap+0x1cc/0x240 add_to_swap+0x14/0x70 shrink_page_list+0x968/0xbc0 reclaim_page_list+0x70/0xf0 reclaim_pages+0xdd/0x120 madvise_cold_or_pageout_pte_range+0x814/0xf30 walk_pgd_range+0x637/0xa30 __walk_page_range+0x142/0x170 walk_page_range+0x146/0x170 madvise_pageout+0xb7/0x280 ? asm_common_interrupt+0x22/0x40 madvise_vma_behavior+0x3b7/0xac0 ? find_vma+0x4a/0x70 ? find_vma+0x64/0x70 ? madvise_vma_anon_name+0x40/0x40 madvise_walk_vmas+0xa6/0x130 do_madvise+0x2f4/0x360 __x64_sys_madvise+0x26/0x30 do_syscall_64+0x5b/0x80 ? do_syscall_64+0x67/0x80 ? syscall_exit_to_user_mode+0x17/0x40 ? do_syscall_64+0x67/0x80 ? syscall_exit_to_user_mode+0x17/0x40 ? do_syscall_64+0x67/0x80 ? do_syscall_64+0x67/0x80 ? common_interrupt+0x8b/0xa0 entry_SYSCALL_64_after_hwframe+0x63/0xcd The problem can be reproduced with the mmtests config config-workload-stressng-mmap. It does not always happen and when it triggers is variable but it has happened on multiple machines. The intent of commit b653db77350c patch was to avoid the case where PG_private is clear but folio->private is not-NULL. However, THP tail pages uses page->private for "swp_entry_t if folio_test_swapcache()" as stated in the documentation for struct folio. This patch only clobbers page->private for tail pages if the head page was not in swapcache and warns once if page->private had an unexpected value. Fixes: b653db77350c ("mm: Clear page->private when splitting or migrating a page") Signed-off-by: Mel Gorman <mgorman@techsingularity.net> --- mm/huge_memory.c | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-)