Re: [PATCH v2] ext4: fix race condition between buffer write and page_mkwrite

[Baokun Li]

On 2024/4/15 20:34, Jan Kara wrote:
> On Mon 15-04-24 12:28:01, Baokun Li wrote:
> > On 2023/6/5 23:08, Jan Kara wrote:
> > > On Mon 05-06-23 15:55:35, Matthew Wilcox wrote:
> > > > On Mon, Jun 05, 2023 at 02:21:41PM +0200, Jan Kara wrote:
> > > > > On Mon 05-06-23 11:16:55, Jan Kara wrote:
> > > > > > Yeah, I agree, that is also the conclusion I have arrived at when thinking
> > > > > > about this problem now. We should be able to just remove the conversion
> > > > > > from ext4_page_mkwrite() and rely on write(2) or truncate(2) doing it when
> > > > > > growing i_size.
> > > > > OK, thinking more about this and searching through the history, I've
> > > > > realized why the conversion is originally in ext4_page_mkwrite(). The
> > > > > problem is described in commit 7b4cc9787fe35b ("ext4: evict inline data
> > > > > when writing to memory map") but essentially it boils down to the fact that
> > > > > ext4 writeback code does not expect dirty page for a file with inline data
> > > > > because ext4_write_inline_data_end() should have copied the data into the
> > > > > inode and cleared the folio's dirty flag.
> > > > >
> > > > > Indeed messing with xattrs from the writeback path to copy page contents
> > > > > into inline data xattr would be ... interesting. Hum, out of good ideas for
> > > > > now :-|.
> > > > Is it so bad?  Now that we don't have writepage in ext4, only
> > > > writepages, it seems like we have a considerably more benign locking
> > > > environment to work in.
> > > Well, yes, without ->writepage() it might be *possible*. But still rather
> > > ugly. The problem is that in ->writepages() i_size is not stable. Thus also
> > > whether the inode data is inline or not is not stable. We'd need inode_lock
> > > for that but that is not easily doable in the writeback path - inode lock
> > > would then become fs_reclaim unsafe...
> > >
> > > 								Honza
> > Hi Honza!
> > Hi Ted!
> > Hi Matthew!
> >
> > Long time later came back to this, because while discussing another similar
> > ABBA problem with Hou Tao, he mentioned VM_FAULT_RETRY, and then I
> > thought that this could be used to solve this problem as well.
> >
> > The general idea is that if we see a file with inline data in
> > ext4_page_mkwrite(),
> > we release the mmap_lock and grab the inode_lock to convert the inline data,
> > and then return VM_FAULT_RETRY to retry to get the mmap_lock.
> >
> > The code implementation is as follows, do you have any thoughts?
> So the problem with this is that VM_FAULT_RETRY is not always an option -
> in particular the caller has to set FAULT_FLAG_ALLOW_RETRY to indicate it
> is prepared to handle VM_FAULT_RETRY return. See how
> maybe_unlock_mmap_for_io() is carefully checking this.
Yes, at least we need to check for FAULT_FLAG_RETRY_NOWAIT.
> There are callers
> (most notably some get_user_pages() users) that don't set
> FAULT_FLAG_ALLOW_RETRY so the escape through VM_FAULT_RETRY is sadly not a
> reliable solution.
It is indeed sad.  I'm going to go learn more about the code for
FAULT_FLAG_ALLOW_RETRY.
> My long-term wish is we were always allowed to use VM_FAULT_RETRY and that
> was actually what motivated some get_user_pages() cleanups I did couple
> years ago. But dealing with all the cases in various drivers was too
> difficult and I've run out of time. Now maybe it would be worth it to
> revisit this since things have changed noticeably and maybe now it would be
> easier to achive the goal...
>
> 								Honza
That sounds like a great idea. I will try to get the history on it and
then come back.

Thank you very much for your patient explanation!
--
With Best Regards,
Baokun Li
.