Re: [RFC PATCH] mm/huge_memory: Do not clobber swp_entry_t during THP split

[Yang Shi]

On Wed, Oct 19, 2022 at 6:42 AM Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx> 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@xxxxxxxxxx/

>
> Fixes: b653db77350c ("mm: Clear page->private when splitting or migrating a page")
> Signed-off-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
> ---
>  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();
>