On Thu, Jun 06, 2024 at 12:14:22AM +0300, миша ухин wrote:
> <div><div>Thank you for the comment.<br />It seems there might be a misunderstanding.<br />The commit 00d873c17e29 ("ext4: avoid deadlock in fs reclaim with page writeback") you mentioned introduces the use of memalloc_nofs_save()/memalloc_nofs_restore() when acquiring the EXT4_SB(sb)->s_writepages_rwsem lock.<br />On the other hand the patch we proposed corrects the order of locking/unlocking resources with calls to the functions ext4_journal_start()/ext4_journal_stop() and down_write(&EXT4_I(inode)->i_data_sem)/up_write(&EXT4_I(inode)->i_data_sem).<br />These patches do not appear to resolve the same issue, and the code changes are different.</div><div> </div><div>- <span style="white-space:pre-wrap">Mikhail Ukhin</span></div></div>
PLEASE do not send HTML messages to the linux-kernel mailing list. It
looks like garbage when read on a text mail reader.
In any case, you're correct. I had misremembered the issue with this
patch. The complaint that I had made with the V1 of the patch has not
been corrected, which is that the assertion made in the commit
description "the order of unlocking must be the reverse of the order
of locking" is errant nonsense. It is simply is technically
incorrect; the order in which locks are released doesn't matter. (And
a jbd2 handle is not a lock.)
The syzkaller report which apparntly triggered this failure was
supplied by Artem here[1], and the explanation should include that it
was triggered by an EXT4_IOC_MIGRATE ioctl which was set to require
synchornous update because the file descriptor was opened with O_SYNC,
and this could result in the jbd2_journal_stop() function calling
jbd2_might_wait_for_commit() which could potentially trigger a
deadlock if the EXT4_IOC_MIGRATE call is racing with write(2) system
call.
[1] https://lore.kernel.org/r/1845977.e0hk0VWMCB@cherry
In any case, this is a low priority issue since the only program which
uses EXT4_IOC_MIGRATE is e4defrag, and it doesn't open files with
O_SYNC, so this isn't going to happen in real life. And so why don't
you use this as an opportunity to practice writing a technically valid
and correct commit description, and how to properlty submit patches
and send valid (non-HTML) messages to the Linux kernel mailing list?
Cheers,
- Ted
