[PATCH] f2fs: fix potential deadlock on compressed quota file

From: Chao Yu
Date: Fri Mar 20 2020 - 06:14:53 EST


generic/232 reports below deadlock:

fsstress D 0 96980 96969 0x00084000
Call Trace:
schedule+0x4a/0xb0
io_schedule+0x12/0x40
__lock_page+0x127/0x1d0
pagecache_get_page+0x1d8/0x250
prepare_compress_overwrite+0xe0/0x490 [f2fs]
f2fs_prepare_compress_overwrite+0x5d/0x80 [f2fs]
f2fs_write_begin+0x833/0xb90 [f2fs]
f2fs_quota_write+0x145/0x1e0 [f2fs]
write_blk+0x36/0x80 [quota_tree]
do_insert_tree+0x2ac/0x4a0 [quota_tree]
do_insert_tree+0x26e/0x4a0 [quota_tree]
qtree_write_dquot+0x70/0x190 [quota_tree]
v2_write_dquot+0x43/0x90 [quota_v2]
dquot_acquire+0x77/0x100
f2fs_dquot_acquire+0x2f/0x60 [f2fs]
dqget+0x310/0x450
dquot_transfer+0xb2/0x120
f2fs_setattr+0x11a/0x4a0 [f2fs]
notify_change+0x349/0x480
chown_common+0x168/0x1c0
do_fchownat+0xbc/0xf0
__x64_sys_lchown+0x21/0x30
do_syscall_64+0x5f/0x220
entry_SYSCALL_64_after_hwframe+0x44/0xa9

task PC stack pid father
kworker/u256:0 D 0 103444 2 0x80084000
Workqueue: writeback wb_workfn (flush-251:1)
Call Trace:
schedule+0x4a/0xb0
schedule_timeout+0x15e/0x2f0
io_schedule_timeout+0x19/0x40
congestion_wait+0x7e/0x120
f2fs_write_multi_pages+0x12a/0x840 [f2fs]
f2fs_write_cache_pages+0x48f/0x790 [f2fs]
f2fs_write_data_pages+0x2db/0x330 [f2fs]
do_writepages+0x1a/0x60
__writeback_single_inode+0x3d/0x340
writeback_sb_inodes+0x225/0x4a0
wb_writeback+0xf7/0x320
wb_workfn+0xba/0x470
process_one_work+0x16c/0x3f0
worker_thread+0x4c/0x440
kthread+0xf8/0x130
ret_from_fork+0x35/0x40

fsstress D 0 5277 5266 0x00084000
Call Trace:
schedule+0x4a/0xb0
rwsem_down_write_slowpath+0x29d/0x540
block_operations+0x105/0x360 [f2fs]
f2fs_write_checkpoint+0x101/0x1010 [f2fs]
f2fs_sync_fs+0xa8/0x130 [f2fs]
f2fs_do_sync_file+0x1ad/0x890 [f2fs]
do_fsync+0x38/0x60
__x64_sys_fdatasync+0x13/0x20
do_syscall_64+0x5f/0x220
entry_SYSCALL_64_after_hwframe+0x44/0xa9

The root cause is there is potential deadlock between quota data
update and writeback.

Kworker Thread B Thread C
- f2fs_write_cache_pages
- lock whole cluster --- A
- f2fs_write_multi_pages
- f2fs_write_raw_pages
- f2fs_write_single_data_page
- f2fs_do_write_data_page
- f2fs_setattr
- f2fs_lock_op --- B
- f2fs_write_checkpoint
- block_operations
- f2fs_lock_all --- B
- dquot_transfer
- f2fs_quota_write
- f2fs_prepare_compress_overwrite
- pagecache_get_page --- A
- f2fs_trylock_op failed --- B
- congestion_wait
- goto rewrite

To fix this issue, during quota file writeback, just redirty all pages
left in cluster rather holding pages' lock in cluster and looping retrying
lock cp_rwsem.

Signed-off-by: Chao Yu <yuchao0@xxxxxxxxxx>
---
fs/f2fs/compress.c | 15 ++++++++++-----
1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/fs/f2fs/compress.c b/fs/f2fs/compress.c
index 758ef8183e59..df3e6ad5b6d7 100644
--- a/fs/f2fs/compress.c
+++ b/fs/f2fs/compress.c
@@ -1235,6 +1235,15 @@ static int f2fs_write_raw_pages(struct compress_ctx *cc,
unlock_page(cc->rpages[i]);
ret = 0;
} else if (ret == -EAGAIN) {
+ /*
+ * for quota file, just redirty left pages to
+ * avoid deadlock caused by cluster update race
+ * from foreground operation.
+ */
+ if (IS_NOQUOTA(cc->inode)) {
+ err = 0;
+ goto out_err;
+ }
ret = 0;
cond_resched();
congestion_wait(BLK_RW_ASYNC,
@@ -1244,16 +1253,12 @@ static int f2fs_write_raw_pages(struct compress_ctx *cc,
goto retry_write;
}
err = ret;
- goto out_fail;
+ goto out_err;
}

*submitted += _submitted;
}
return 0;
-
-out_fail:
- /* TODO: revoke partially updated block addresses */
- BUG_ON(compr_blocks);
out_err:
for (++i; i < cc->cluster_size; i++) {
if (!cc->rpages[i])
--
2.18.0.rc1