Hi, Xiang,
Gao Xiang <hsiangkao@xxxxxxxxxxxxxxxxx> writes:
> Currently, migrate_pages_batch() can lock multiple locked folios
> with an arbitrary order. Although folio_trylock() is used to avoid
> deadlock as commit 2ef7dbb26990 ("migrate_pages: try migrate in batch
> asynchronously firstly") mentioned, it seems try_split_folio() is
> still missing.
>
> It was found by compaction stress test when I explicitly enable EROFS
> compressed files to use large folios, which case I cannot reproduce with
> the same workload if large folio support is off (current mainline).
> Typically, filesystem reads (with locked file-backed folios) could use
> another bdev/meta inode to load some other I/Os (e.g. inode extent
> metadata or caching compressed data), so the locking order will be:
>
> file-backed folios (A)
> bdev/meta folios (B)
>
> The following calltrace shows the deadlock:
> Thread 1 takes (B) lock and tries to take folio (A) lock
> Thread 2 takes (A) lock and tries to take folio (B) lock
>
> [Thread 1]
> INFO: task stress:1824 blocked for more than 30 seconds.
> Tainted: G OE 6.10.0-rc7+ #6
> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> task:stress state:D stack:0 pid:1824 tgid:1824 ppid:1822 flags:0x0000000c
> Call trace:
> __switch_to+0xec/0x138
> __schedule+0x43c/0xcb0
> schedule+0x54/0x198
> io_schedule+0x44/0x70
> folio_wait_bit_common+0x184/0x3f8
> <-- folio mapping ffff00036d69cb18 index 996 (**)
> __folio_lock+0x24/0x38
> migrate_pages_batch+0x77c/0xea0 // try_split_folio (mm/migrate.c:1486:2)
> // migrate_pages_batch (mm/migrate.c:1734:16)
> <--- LIST_HEAD(unmap_folios) has
> ..
> folio mapping 0xffff0000d184f1d8 index 1711; (*)
> folio mapping 0xffff0000d184f1d8 index 1712;
> ..
> migrate_pages+0xb28/0xe90
> compact_zone+0xa08/0x10f0
> compact_node+0x9c/0x180
> sysctl_compaction_handler+0x8c/0x118
> proc_sys_call_handler+0x1a8/0x280
> proc_sys_write+0x1c/0x30
> vfs_write+0x240/0x380
> ksys_write+0x78/0x118
> __arm64_sys_write+0x24/0x38
> invoke_syscall+0x78/0x108
> el0_svc_common.constprop.0+0x48/0xf0
> do_el0_svc+0x24/0x38
> el0_svc+0x3c/0x148
> el0t_64_sync_handler+0x100/0x130
> el0t_64_sync+0x190/0x198
>
> [Thread 2]
> INFO: task stress:1825 blocked for more than 30 seconds.
> Tainted: G OE 6.10.0-rc7+ #6
> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> task:stress state:D stack:0 pid:1825 tgid:1825 ppid:1822 flags:0x0000000c
> Call trace:
> __switch_to+0xec/0x138
> __schedule+0x43c/0xcb0
> schedule+0x54/0x198
> io_schedule+0x44/0x70
> folio_wait_bit_common+0x184/0x3f8
> <-- folio = 0xfffffdffc6b503c0 (mapping == 0xffff0000d184f1d8 index == 1711) (*)
> __folio_lock+0x24/0x38
> z_erofs_runqueue+0x384/0x9c0 [erofs]
> z_erofs_readahead+0x21c/0x350 [erofs] <-- folio mapping 0xffff00036d69cb18 range from [992, 1024] (**)
> read_pages+0x74/0x328
> page_cache_ra_order+0x26c/0x348
> ondemand_readahead+0x1c0/0x3a0
> page_cache_sync_ra+0x9c/0xc0
> filemap_get_pages+0xc4/0x708
> filemap_read+0x104/0x3a8
> generic_file_read_iter+0x4c/0x150
> vfs_read+0x27c/0x330
> ksys_pread64+0x84/0xd0
> __arm64_sys_pread64+0x28/0x40
> invoke_syscall+0x78/0x108
> el0_svc_common.constprop.0+0x48/0xf0
> do_el0_svc+0x24/0x38
> el0_svc+0x3c/0x148
> el0t_64_sync_handler+0x100/0x130
> el0t_64_sync+0x190/0x198
>
> Fixes: 5dfab109d519 ("migrate_pages: batch _unmap and _move")
> Cc: "Huang, Ying" <ying.huang@xxxxxxxxx>
> Signed-off-by: Gao Xiang <hsiangkao@xxxxxxxxxxxxxxxxx>
> ---
> mm/migrate.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/migrate.c b/mm/migrate.c
> index 20cb9f5f7446..a912e4b83228 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -1483,7 +1483,8 @@ static inline int try_split_folio(struct folio *folio, struct list_head *split_f
> {
> int rc;
>
> - folio_lock(folio);
> + if (!folio_trylock(folio))
> + return -EAGAIN;
> rc = split_folio_to_list(folio, split_folios);
> folio_unlock(folio);
> if (!rc)
Good catch! Thanks for the fixing!
The deadlock is similar as the one we fixed in commit fb3592c41a44
("migrate_pages: fix deadlock in batched migration"). But apparently,
we missed this case.
For the fix, I think that we should still respect migrate_mode because
users may prefer migration success over blocking.
@@ -1492,11 +1492,17 @@ static int unmap_and_move_huge_page(new_folio_t get_new_folio,
return rc;
}
-static inline int try_split_folio(struct folio *folio, struct list_head *split_folios)
+static inline int try_split_folio(struct folio *folio, struct list_head *split_folios,
+ enum migrate_mode mode)
{
int rc;
- folio_lock(folio);
+ if (mode == MIGRATE_ASYNC) {
+ if (!folio_trylock(folio))
+ return -EAGAIN;
+ } else {
+ folio_lock(folio);
+ }
rc = split_folio_to_list(folio, split_folios);
folio_unlock(folio);
if (!rc)
--
Best Regards,
Huang, Ying
