Re: [PATCH v2] f2fs: fix potential deadlock in f2fs_balance_fs()

[Ruipeng Qi <ruipengqi3@gmail.com>]

On 2026/4/29 15:59, Chao Yu wrote:
> On 4/29/26 11:39, Ruipeng Qi wrote:
>>
>> On 2026/4/27 16:38, Chao Yu wrote:
>>> On 4/26/26 17:32, ruipengqi wrote:
>>>> From: Ruipeng Qi <ruipengqi3@gmail.com>
>>>>
>>>> When the f2fs filesystem space is nearly exhausted, we encounter 
>>>> deadlock
>>>> issues as below:
>>>>
>>>> INFO: task A:1890 blocked for more than 120 seconds.
>>>>        Tainted: G           O       6.12.41-g3fe07ddf05ab #1
>>>> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this 
>>>> message.
>>>> task:A    state:D stack:0     pid:1890  tgid:1626  ppid:1153 
>>>> flags:0x00000204
>>>> Call trace:
>>>>   __switch_to+0xf4/0x158
>>>>   __schedule+0x27c/0x908
>>>>   schedule+0x3c/0x118
>>>>   io_schedule+0x44/0x68
>>>>   folio_wait_bit_common+0x174/0x370
>>>>   folio_wait_bit+0x20/0x38
>>>>   folio_wait_writeback+0x54/0xc8
>>>>   truncate_inode_partial_folio+0x70/0x1e0
>>>>   truncate_inode_pages_range+0x1b0/0x450
>>>>   truncate_pagecache+0x54/0x88
>>>>   f2fs_file_write_iter+0x3e8/0xb80
>>>>   do_iter_readv_writev+0xf0/0x1e0
>>>>   vfs_writev+0x138/0x2c8
>>>>   do_writev+0x88/0x130
>>>>   __arm64_sys_writev+0x28/0x40
>>>>   invoke_syscall+0x50/0x120
>>>>   el0_svc_common.constprop.0+0xc8/0xf0
>>>>   do_el0_svc+0x24/0x38
>>>>   el0_svc+0x30/0xf8
>>>>   el0t_64_sync_handler+0x120/0x130
>>>>   el0t_64_sync+0x190/0x198
>>>>
>>>> INFO: task kworker/u8:11:2680853 blocked for more than 120 seconds.
>>>>        Tainted: G           O       6.12.41-g3fe07ddf05ab #1
>>>> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this 
>>>> message.
>>>> task:kworker/u8:11   state:D stack:0     pid:2680853 tgid:2680853 
>>>> ppid:2      flags:0x00000208
>>>> Workqueue: writeback wb_workfn (flush-254:0)
>>>> Call trace:
>>>>   __switch_to+0xf4/0x158
>>>>   __schedule+0x27c/0x908
>>>>   schedule+0x3c/0x118
>>>>   io_schedule+0x44/0x68
>>>>   folio_wait_bit_common+0x174/0x370
>>>>   __filemap_get_folio+0x214/0x348
>>>>   pagecache_get_page+0x20/0x70
>>>>   f2fs_get_read_data_page+0x150/0x3e8
>>>>   f2fs_get_lock_data_page+0x2c/0x160
>>>>   move_data_page+0x50/0x478
>>>>   do_garbage_collect+0xd38/0x1528
>>>>   f2fs_gc+0x240/0x7e0
>>>>   f2fs_balance_fs+0x1a0/0x208
>>>>   f2fs_write_single_data_page+0x6e4/0x730 //0xfffffe0d6ca08300
>>>>   f2fs_write_cache_pages+0x378/0x9b0
>>>>   f2fs_write_data_pages+0x2e4/0x388
>>>>   do_writepages+0x8c/0x2c8
>>>>   __writeback_single_inode+0x4c/0x498
>>>>   writeback_sb_inodes+0x234/0x4a8
>>>>   __writeback_inodes_wb+0x58/0x118
>>>>   wb_writeback+0x2f8/0x3c0
>>>>   wb_workfn+0x2c4/0x508
>>>>   process_one_work+0x180/0x408
>>>>   worker_thread+0x258/0x368
>>>>   kthread+0x118/0x128
>>>>   ret_from_fork+0x10/0x200
>>>>
>>>> INFO: task kworker/u8:8:2641297 blocked for more than 120 seconds.
>>>>        Tainted: G           O       6.12.41-g3fe07ddf05ab #1
>>>> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this 
>>>> message.
>>>> task:kworker/u8:8    state:D stack:0     pid:2641297 tgid:2641297 
>>>> ppid:2      flags:0x00000208
>>>> Workqueue: writeback wb_workfn (flush-254:0)
>>>> Call trace:
>>>>   __switch_to+0xf4/0x158
>>>>   __schedule+0x27c/0x908
>>>>   rt_mutex_schedule+0x30/0x60
>>>>   __rt_mutex_slowlock_locked.constprop.0+0x460/0x8a8
>>>>   rwbase_write_lock+0x24c/0x378
>>>>   down_write+0x1c/0x30
>>>>   f2fs_balance_fs+0x184/0x208
>>>>   f2fs_write_inode+0xf4/0x328
>>>>   __writeback_single_inode+0x370/0x498
>>>>   writeback_sb_inodes+0x234/0x4a8
>>>>   __writeback_inodes_wb+0x58/0x118
>>>>   wb_writeback+0x2f8/0x3c0
>>>>   wb_workfn+0x2c4/0x508
>>>>   process_one_work+0x180/0x408
>>>>   worker_thread+0x258/0x368
>>>>   kthread+0x118/0x128
>>>>   ret_from_fork+0x10/0x20
>>>>
>>>> INFO: task B:1902 blocked for more than 120 seconds.
>>>>        Tainted: G           O       6.12.41-g3fe07ddf05ab #1
>>>> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this 
>>>> message.
>>>> task:B     state:D stack:0     pid:1902  tgid:1626 ppid:1153 
>>>> flags:0x0000020c
>>>> Call trace:
>>>>   __switch_to+0xf4/0x158
>>>>   __schedule+0x27c/0x908
>>>>   rt_mutex_schedule+0x30/0x60
>>>>   __rt_mutex_slowlock_locked.constprop.0+0x460/0x8a8
>>>>   rwbase_write_lock+0x24c/0x378
>>>>   down_write+0x1c/0x30
>>>>   f2fs_balance_fs+0x184/0x208
>>>>   f2fs_map_blocks+0x94c/0x1110
>>>>   f2fs_file_write_iter+0x228/0xb80
>>>>   do_iter_readv_writev+0xf0/0x1e0
>>>>   vfs_writev+0x138/0x2c8
>>>>   do_writev+0x88/0x130
>>>>   __arm64_sys_writev+0x28/0x40
>>>>   invoke_syscall+0x50/0x120
>>>>   el0_svc_common.constprop.0+0xc8/0xf0
>>>>   do_el0_svc+0x24/0x38
>>>>   el0_svc+0x30/0xf8
>>>>   el0t_64_sync_handler+0x120/0x130
>>>>   el0t_64_sync+0x190/0x198
>>>>
>>>> INFO: task sync:2769849 blocked for more than 120 seconds.
>>>>        Tainted: G           O       6.12.41-g3fe07ddf05ab #1
>>>> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this 
>>>> message.
>>>> task:sync            state:D stack:0     pid:2769849 tgid:2769849 
>>>> ppid:736    flags:0x0000020c
>>>> Call trace:
>>>>   __switch_to+0xf4/0x158
>>>>   __schedule+0x27c/0x908
>>>>   schedule+0x3c/0x118
>>>>   wb_wait_for_completion+0xb0/0xe8
>>>>   sync_inodes_sb+0xc8/0x2b0
>>>>   sync_inodes_one_sb+0x24/0x38
>>>>   iterate_supers+0xa8/0x138
>>>>   ksys_sync+0x54/0xc8
>>>>   __arm64_sys_sync+0x18/0x30
>>>>   invoke_syscall+0x50/0x120
>>>>   el0_svc_common.constprop.0+0xc8/0xf0
>>>>   do_el0_svc+0x24/0x38
>>>>   el0_svc+0x30/0xf8
>>>>   el0t_64_sync_handler+0x120/0x130
>>>>   el0t_64_sync+0x190/0x198
>>>>
>>>> The root cause is a potential deadlock between the following tasks:
>>>>
>>>> kworker/u8:11                Thread A
>>>> - f2fs_write_single_data_page
>>>>   - f2fs_do_write_data_page
>>>>    - folio_start_writeback(X)
>>>>    - f2fs_outplace_write_data
>>>>     - bio_add_folio(X)
>>>>   - folio_unlock(X)
>>>>                     - truncate_inode_pages_range
>>>>                      - __filemap_get_folio(X, FGP_LOCK)
>>>>                      - truncate_inode_partial_folio(X)
>>>>                       - folio_wait_writeback(X)
>>>>   - f2fs_balance_fs
>>>>    - f2fs_gc
>>>>     - do_garbage_collect
>>>>      - move_data_page
>>>>       - f2fs_get_lock_data_page
>>>>        - __filemap_get_folio(X, FGP_LOCK)
>>>>
>>>> Both threads try to access folio X. Thread A holds the lock but waits
>>>> for writeback, while kworker waits for the lock. This causes a 
>>>> deadlock.
>>>>
>>>> Other threads also enter D state, waiting for locks such as gc_lock 
>>>> and
>>>> writepages.
>>>>
>>>> OPU/IPU DATA folio are all affected by this issue. To avoid such
>>>> potential deadlocks, always commit these cached folios before
>>>> triggering f2fs_gc() in f2fs_balance_fs().
>>>>
>>>> v2:
>>>> - Commit cached OPU/IPU folios, not just OPU folios as in v1.
>>>>
>>>> Suggested-by: Chao <chao@kernel.org>
>>>> Signed-off-by: Ruipeng Qi <ruipengqi3@gmail.com>
>>>> ---
>>>>   fs/f2fs/data.c    | 26 ++++++++++++++++++++++++++
>>>>   fs/f2fs/f2fs.h    |  1 +
>>>>   fs/f2fs/segment.c |  9 +++++++++
>>>>   3 files changed, 36 insertions(+)
>>>>
>>>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
>>>> index 338df7a2aea6..fd03366b3228 100644
>>>> --- a/fs/f2fs/data.c
>>>> +++ b/fs/f2fs/data.c
>>>> @@ -939,6 +939,32 @@ void f2fs_submit_merged_ipu_write(struct 
>>>> f2fs_sb_info *sbi,
>>>>       }
>>>>   }
>>>>   +void f2fs_submit_all_merged_ipu_writes(struct f2fs_sb_info *sbi)
>>>> +{
>>>> +    struct bio_entry *be, *tmp;
>>>> +    struct f2fs_bio_info *io;
>>>> +    enum temp_type temp;
>>>> +    LIST_HEAD(list);
>>>> +
>>>> +    for (temp = HOT; temp < NR_TEMP_TYPE; temp++) {
>>>> +        io = sbi->write_io[DATA] + temp;
>>>> +
>>>> +        if (list_empty(&io->bio_list))
>>>> +            continue;
>>>
>>> Needs to be covered w/ bio_list_lock to avoid race condition.
>>
>> Hi,Chao
>>
>> The lockless list_empty() here is intentional and acceptable.
>>
>>
>> If list_empty() returns true but the list becomes non-empty
>> afterwards (due to race), the newly added bio will be submitted
>> by the subsequent write path, so no bio will be lost.
>
> Ah, right, we only need to submit the folios cached by local thread.
>
>>
>>
>> Similar patterns exist in the kernel, e.g.:
>>    net/rfkill/core.c: rfkill_fop_read()
>>      /* since we re-check and it just compares pointers,
>>       * using !list_empty() without locking isn't a problem
>>       */
>>    fs/f2fs/data.c: f2fs_submit_merged_ipu_write()
>>      list_empty() is also used without holding bio_list_lock
>>      as a lockless pre-check
>>
>>
>> If you'd prefer, we can add a comment to make the intent clear:
>>
>>      /* list_empty() without lock is safe here - READ_ONCE()
>>       * ensures pointer read atomicity. A false negative is
>>       * acceptable since any bio added concurrently will be
>>       * submitted by the next write path.
>>       */
>>      if (list_empty(&io->bio_list))
>>          continue;
>>>
>>>> +
>>>> +        f2fs_down_write(&io->bio_list_lock);
>>>> +        list_splice_init(&io->bio_list, &list);
>>>> +        f2fs_up_write(&io->bio_list_lock);
>>>> +
>>>> +        list_for_each_entry_safe(be, tmp, &list, list) {
>>>> +            f2fs_submit_write_bio(sbi, be->bio, DATA);
>>>> +            del_bio_entry(be);
>>>> +        }
>>>> +
>>>
>>> Unnecessary blank line.
>>>
>>> Thanks,
>>
>> Thanks for your correction. Will fix in v3.
>>      v3:
>>      - Fixed minor grammatical issues
>>      - Add comment on lockless list_empty() to explain why it is safe
>>    without holding bio_list_lock
>
> Seems fine.
>
>>
>>
>> Thanks,
>>
>>>
>>>> +    }
>>>> +
>>>> +}
>>>> +
>>>>   int f2fs_merge_page_bio(struct f2fs_io_info *fio)
>>>>   {
>>>>       struct bio *bio = *fio->bio;
>>>> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
>>>> index bb34e864d0ef..e9038ab1b2bd 100644
>>>> --- a/fs/f2fs/f2fs.h
>>>> +++ b/fs/f2fs/f2fs.h
>>>> @@ -4148,6 +4148,7 @@ void f2fs_submit_merged_write_folio(struct 
>>>> f2fs_sb_info *sbi,
>>>>                   struct folio *folio, enum page_type type);
>>>>   void f2fs_submit_merged_ipu_write(struct f2fs_sb_info *sbi,
>>>>                       struct bio **bio, struct folio *folio);
>>>> +void f2fs_submit_all_merged_ipu_writes(struct f2fs_sb_info *sbi);
>>>>   void f2fs_flush_merged_writes(struct f2fs_sb_info *sbi);
>>>>   int f2fs_submit_page_bio(struct f2fs_io_info *fio);
>>>>   int f2fs_merge_page_bio(struct f2fs_io_info *fio);
>>>> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
>>>> index 6a97fe76712b..856ffe91b94f 100644
>>>> --- a/fs/f2fs/segment.c
>>>> +++ b/fs/f2fs/segment.c
>>>> @@ -454,6 +454,15 @@ void f2fs_balance_fs(struct f2fs_sb_info *sbi, 
>>>> bool need)
>>>>           io_schedule();
>>>>           finish_wait(&sbi->gc_thread->fggc_wq, &wait);
>>>>       } else {
>>>> +
>>>> +        /*
>>>> +         * Submit all cached OPU/IPU DATA bios before triggering
>>>> +         * foreground GC to avoid potential deadlocks.
>>>> +         */
>>>> +
>>>> +        f2fs_submit_merged_write(sbi, DATA);
>>>> +        f2fs_submit_all_merged_ipu_writes(sbi);
>
> Can we relocate above code to below the variable definitions?
>
> Thanks,
>
Hi, Chao

Sure, will fix it in V3.

BTW, To avoid potential deadlocks, this patch submits cached OPU/IPU folios
before triggering f2fs_gc() in f2fs_balance_fs(), which changes the
existing IPU/OPU BIO lifecycle.

For OPU, io->io_rwsem provides the necessary synchronization.
For IPU, io->bio_list_lock ensures race-free submission.
In both cases, new BIOs will be allocated as needed after submission.

I may have missed something in the current implementation.
Your professional review would be much appreciated.

Thanks,

>>>> +
>>>>           struct f2fs_gc_control gc_control = {
>>>>               .victim_segno = NULL_SEGNO,
>>>>               .init_gc_type = f2fs_sb_has_blkzoned(sbi) ?
>>>
>