Re: [PATCH] block_dump: don't put the last refcount when marking inode dirty

["zhangyi (F)" <yi.zhang@huawei.com>]

On 2021/3/5 18:10, Jan Kara wrote:
> On Thu 04-03-21 21:37:42, zhangyi (F) wrote:
>> On 2021/3/1 19:21, Jan Kara wrote:
>>> On Fri 26-02-21 18:31:03, zhangyi (F) wrote:
>>>> There is an AA deadlock problem when using block_dump on ext4 file
>>>> system with data=journal mode.
>>>>
>>>>   watchdog: BUG: soft lockup - CPU#19 stuck for 22s! [jbd2/pmem0-8:1002]
>>>>   CPU: 19 PID: 1002 Comm: jbd2/pmem0-8
>>>>   RIP: 0010:queued_spin_lock_slowpath+0x60/0x3b0
>>>>   ...
>>>>   Call Trace:
>>>>    _raw_spin_lock+0x57/0x70
>>>>    jbd2_journal_invalidatepage+0x166/0x680
>>>>    __ext4_journalled_invalidatepage+0x8c/0x120
>>>>    ext4_journalled_invalidatepage+0x12/0x40
>>>>    truncate_cleanup_page+0x10e/0x1c0
>>>>    truncate_inode_pages_range+0x2c8/0xec0
>>>>    truncate_inode_pages_final+0x41/0x90
>>>>    ext4_evict_inode+0x254/0xac0
>>>>    evict+0x11c/0x2f0
>>>>    iput+0x20e/0x3a0
>>>>    dentry_unlink_inode+0x1bf/0x1d0
>>>>    __dentry_kill+0x14c/0x2c0
>>>>    dput+0x2bc/0x630
>>>>    block_dump___mark_inode_dirty.cold+0x5c/0x111
>>>>    __mark_inode_dirty+0x678/0x6b0
>>>>    mark_buffer_dirty+0x16e/0x1d0
>>>>    __jbd2_journal_temp_unlink_buffer+0x127/0x1f0
>>>>    __jbd2_journal_unfile_buffer+0x24/0x80
>>>>    __jbd2_journal_refile_buffer+0x12f/0x1b0
>>>>    jbd2_journal_commit_transaction+0x244b/0x3030
>>>>
>>>> The problem is a race between jbd2 committing data buffer and user
>>>> unlink the file concurrently. The jbd2 will get jh->b_state_lock and
>>>> redirty the inode's data buffer and inode itself. If block_dump is
>>>> enabled, it will try to find inode's dentry and invoke the last dput()
>>>> after the inode was unlinked. Then the evict procedure will unmap
>>>> buffer and get jh->b_state_lock again in journal_unmap_buffer(), and
>>>> finally lead to deadlock. It works fine if block_dump is not enabled
>>>> because the last evict procedure is not invoked in jbd2 progress and
>>>> the jh->b_state_lock will also prevent inode use after free.
>>>>
>>>> jbd2                                xxx
>>>>                                     vfs_unlink
>>>>                                      ext4_unlink
>>>> jbd2_journal_commit_transaction
>>>> **get jh->b_state_lock**
>>>> jbd2_journal_refile_buffer
>>>>  mark_buffer_dirty
>>>>   __mark_inode_dirty
>>>>    block_dump___mark_inode_dirty
>>>>     d_find_alias
>>>>                                      d_delete
>>>>                                       unhash
>>>>     dput  //put the last refcount
>>>>      evict
>>>>       journal_unmap_buffer
>>>>        **get jh->b_state_lock again**
>>>>
>>>> In most cases of where invoking mark_inode_dirty() will get inode's
>>>> refcount and the last iput may not happen, but it's not safe. After
>>>> checking the block_dump code, it only want to dump the file name of the
>>>> dirty inode, so there is no need to get and put denrty, and dump an
>>>> unhashed dentry is also fine. This patch remove the dget() && dput(),
>>>> print the dentry name directly.
>>>>
>>>> Signed-off-by: zhangyi (F) <yi.zhang@huawei.com>
>>>> Signed-off-by: yebin (H) <yebin10@huawei.com>
>>>
>>> Hrm, ok. Honestly, I wanted to just delete that code for a long time. IMO
>>> tracepoints (and we have one in __mark_inode_dirty) are much more useful
>>> for tracing anyway. This code exists only because it was there much before
>>> tracepoints existed... Do you have a strong reason why are you using
>>> block_dump instead of tracepoint trace_writeback_mark_inode_dirty() for
>>> your monitoring?
>>>
>>
>> Hi, Jan. We just do some stress tests and find this issue, I'm not sure who
>> are still using this old debug interface and gather it may need time. Could
>> we firstly fix this issue, and then delete this code if no opposed?
> 
> I'd do it the other way around :) Delete the code and only fix it if
> someone complains that the feature is still used and so we should not
> delete it. Will you send a patch or should I do it?
> 

OK, I will send a RFC patch to delete this feature.

Thanks,
Yi.