Re: [PATCH v3] shmem: fix a race between shmem_unused_huge_shrink and shmem_evict_inode

[Muchun Song <songmuchun@bytedance.com>]

On Wed, Nov 24, 2021 at 5:43 PM Gang Li <ligang.bdlg@bytedance.com> wrote:
>
> This patch fixes a data race in commit 779750d20b93 ("shmem: split huge pages
> beyond i_size under memory pressure").
>
> Here are call traces:
>
> Call Trace 1:
>         shmem_unused_huge_shrink+0x3ae/0x410
>         ? __list_lru_walk_one.isra.5+0x33/0x160
>         super_cache_scan+0x17c/0x190
>         shrink_slab.part.55+0x1ef/0x3f0
>         shrink_node+0x10e/0x330
>         kswapd+0x380/0x740
>         kthread+0xfc/0x130
>         ? mem_cgroup_shrink_node+0x170/0x170
>         ? kthread_create_on_node+0x70/0x70
>         ret_from_fork+0x1f/0x30
>
> Call Trace 2:
>         shmem_evict_inode+0xd8/0x190
>         evict+0xbe/0x1c0
>         do_unlinkat+0x137/0x330
>         do_syscall_64+0x76/0x120
>         entry_SYSCALL_64_after_hwframe+0x3d/0xa2
>
> The simultaneous deletion of adjacent elements in the local list (@list)
> by shmem_unused_huge_shrink and shmem_evict_inode will break the list.
>
> Image there are 3 items in the local list (@list).
> In the first traversal, A is not deleted from @list.
>
>   1)    A->B->C
>         ^
>         |
>         pos (leave)
>
> In the second traversal, B is deleted from @list. Concurrently, A is
> deleted from @list through shmem_evict_inode() since last reference counter of
> inode is dropped by other thread. Then the @list is corrupted.
>
>   2)    A->B->C
>         ^  ^
>         |  |
>      evict pos (drop)
>
> Fix:
>
> We should make sure the item is either on the global list or deleted from
> any local list before iput().
>
> Fixed by moving inodes that are on @list and will not be deleted back to
> global list before iput.
>
> Fixes: 779750d20b93 ("shmem: split huge pages beyond i_size under memory pressure")
> Signed-off-by: Gang Li <ligang.bdlg@bytedance.com>
>
> ---
>
> Changes in v3:
> - Add more comment.
> - Use list_move(&info->shrinklist, &sbinfo->shrinklist) instead of
>   list_move(pos, &sbinfo->shrinklist) for consistency.
>
> Changes in v2: https://lore.kernel.org/all/20211124030840.88455-1-ligang.bdlg@bytedance.com/
> - Move spinlock to the front of iput instead of changing lock type
>   since iput will call evict which may cause deadlock by requesting
>   shrinklist_lock.
> - Add call trace in commit message.
>
> v1: https://lore.kernel.org/lkml/20211122064126.76734-1-ligang.bdlg@bytedance.com/
>
> ---
>  mm/shmem.c | 35 ++++++++++++++++++++---------------
>  1 file changed, 20 insertions(+), 15 deletions(-)
>
> diff --git a/mm/shmem.c b/mm/shmem.c
> index 9023103ee7d8..ab2df692bd58 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -569,7 +569,6 @@ static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
>                 /* inode is about to be evicted */
>                 if (!inode) {
>                         list_del_init(&info->shrinklist);
> -                       removed++;
>                         goto next;
>                 }
>
> @@ -577,15 +576,16 @@ static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
>                 if (round_up(inode->i_size, PAGE_SIZE) ==
>                                 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
>                         list_move(&info->shrinklist, &to_remove);
> -                       removed++;
>                         goto next;
>                 }
>
>                 list_move(&info->shrinklist, &list);
>  next:
> +               removed++;
>                 if (!--batch)
>                         break;
>         }
> +       sbinfo->shrinklist_len -= removed;
>         spin_unlock(&sbinfo->shrinklist_lock);
>
>         list_for_each_safe(pos, next, &to_remove) {
> @@ -602,7 +602,7 @@ static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
>                 inode = &info->vfs_inode;
>
>                 if (nr_to_split && split >= nr_to_split)
> -                       goto leave;
> +                       goto move_back;
>
>                 page = find_get_page(inode->i_mapping,
>                                 (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT);
> @@ -616,38 +616,43 @@ static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
>                 }
>
>                 /*
> -                * Leave the inode on the list if we failed to lock
> -                * the page at this time.
> +                * Move the inode on the list back to shrinklist if we failed
> +                * to lock the page at this time.
>                  *
>                  * Waiting for the lock may lead to deadlock in the
>                  * reclaim path.
>                  */
>                 if (!trylock_page(page)) {
>                         put_page(page);
> -                       goto leave;
> +                       goto move_back;
>                 }
>
>                 ret = split_huge_page(page);
>                 unlock_page(page);
>                 put_page(page);
>
> -               /* If split failed leave the inode on the list */
> +               /* If split failed move the inode on the list back to shrinklist */
>                 if (ret)
> -                       goto leave;
> +                       goto move_back;
>
>                 split++;
>  drop:
>                 list_del_init(&info->shrinklist);
> -               removed++;
> -leave:
> +               goto put;
> +move_back:
> +               /* inodes that are on @list and will not be deleted must be moved back to
> +                * global list before iput for two reasons:
> +                * 1. iput in lock: iput call shmem_evict_inode, then cause deadlock.
> +                * 2. iput before lock: shmem_evict_inode may grab the inode on @list,
> +                *    which will cause race.
> +                */

Multi-line comment is like the following format.

/*
 * Comment here.
 */

And I also suggest reworking the comments here. Something like:

/*
 * Make sure the inode is either on the global list or deleted from
 * any local list before iput() since it could be deleted in another
 * thread once we put the inode (then the local list is corrupted).
 */

With that.

Reviewed-by: Muchun Song <songmuchun@bytedance.com>

> +               spin_lock(&sbinfo->shrinklist_lock);
> +               list_move(&info->shrinklist, &sbinfo->shrinklist);
> +               sbinfo->shrinklist_len++;
> +               spin_unlock(&sbinfo->shrinklist_lock);
> +put:
>                 iput(inode);
>         }
>
> -       spin_lock(&sbinfo->shrinklist_lock);
> -       list_splice_tail(&list, &sbinfo->shrinklist);
> -       sbinfo->shrinklist_len -= removed;
> -       spin_unlock(&sbinfo->shrinklist_lock);
> -
>         return split;
>  }
>
> --
> 2.20.1
>