Re: [RFC] block integrity: Fix write after checksum calculation problem

["Darrick J. Wong" <djwong@us.ibm.com>]

On Wed, Apr 06, 2011 at 04:29:38PM -0700, Darrick J. Wong wrote:
> On Mon, Mar 21, 2011 at 05:43:05PM +0100, Jan Kara wrote:
> > On Mon 21-03-11 10:24:41, Chris Mason wrote:
> > > Excerpts from Jan Kara's message of 2011-03-21 10:04:51 -0400:
> > > > On Fri 18-03-11 17:07:55, Darrick J. Wong wrote:
> > > > > > > Ok, here's what I have so far.  I took everyone's suggestions of where to add
> > > > > > > calls to wait_on_page_writeback, which seems to handle the multiple-write case
> > > > > > > adequately.  Unfortunately, it is still possible to generate checksum errors by
> > > > > > > scribbling furiously on a mmap'd region, even after adding the writeback wait
> > > > > > > in the ext4 writepage function.  Oddly, I couldn't break btrfs with mmap by
> > > > > > > removing its wait_for_page_writeback call, so I suspect there's a bit more
> > > > > > > going on in btrfs than I've been able to figure out.
> > > > > 
> > > > > I wonder, is it possible for this to happen:
> > > > > 
> > > > > 1. Thread A mmaps a page and tries to write to it.  ext4_page_mkwrite executes,
> > > > >    but there's no ongoing writeback, so it returns without delay.
> > > > > 2. Thread A starts writing furiously to the page.
> > > > > 3. Thread B runs fsync() or something that results in the page being
> > > > >    checksummed and scheduled for writeout.
> > > > > 4. Thread A continues to write furiously(!) on that same page before the
> > > > >    controller finishes the DMA transfer.
> > > > > 5. Disk gets the page, which now doesn't match its checksum, and *boom*
> > > >   What happens on writepage (see mm/page-writeback.c:write_cache_pages())
> > > > is:
> > > >   lock_page(page)
> > > >   ...
> > > >   clear_page_dirty_for_io() - removes PageDirty, marks page as read-only in
> > > >     PTE
> > > >   ...
> > > >   set_page_writeback() (happens e.g. in __block_write_full_page() called
> > > > from filesystem's writepage implementation).
> > > >   unlock_page(page)
> > > > 
> > > >   So if you compute the checksum after set_page_writeback() is done in the
> > > > writepage() implementation (you cannot use __block_write_full_page() in
> > > > that case)
> >   I should add that if you are computing the checksum in the block layer
> > once the bio is submitted, you obviously are computing it after the page is
> > marked as writeback. So that should be fine...
> > 
> > > > and you call wait_on_page_writeback() in ext4_page_mkwrite()
> > > > under page lock, you should be safe. If you do all this and still see
> > > > errors, something is broken I'd say...
> > > 
> > > Looking at the ext4_page_mkwrite, it does this:
> > > 
> > > lock the page
> > > check for holes
> > > unlock the page
> > > if (no_holes)
> > > 	return;
> > > 
> > > write_begin/write_end
> > > return
> > > 
> > > So, to have page_mkwrite work, you need to wait for writeback with the
> > > page locked in both the no holes case and after the
> > > write_begin/write_end.  write_begin will dirty the page, so someone can
> > > wander in and start the IO while we are still in page_mkwrite.
> >   Oh right, that's a good point.
> > 
> > > This is untested and uncompiled, but it should
> > > do the trick.
> > > 
> > > Jan, did you get rid of all the buffer head based writeback for
> > > data=ordered in ext4?  That's my only other idea, that someone is doing
> > > writeback directly without taking the page lock.
> >   Yes, ext4 shouldn't do any buffer based writeback.
> > 
> > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> > > index 9f7f9e4..8a75e12 100644
> > > --- a/fs/ext4/inode.c
> > > +++ b/fs/ext4/inode.c
> > > @@ -5880,6 +5880,7 @@ int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
> > >  	if (page_has_buffers(page)) {
> > >  		if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
> > >  					ext4_bh_unmapped)) {
> > > +			wait_on_page_writeback(page);
> > >  			unlock_page(page);
> > >  			goto out_unlock;
> > >  		}
> > > @@ -5901,6 +5902,16 @@ int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
> > >  	if (ret < 0)
> > >  		goto out_unlock;
> > >  	ret = 0;
> > > +
> > > +	/*
> > > +	 * write_begin/end might have created a dirty page and someone
> > > +	 * could wander in and start the IO.  Make sure that hasn't
> > > +	 * happened
> > > +	 */
> > > +	lock_page(page);
> > > +	wait_on_page_writeback(page);
> > > +	unlock_page(page);
> > > +
> > >  out_unlock:
> > >  	if (ret)
> > >  		ret = VM_FAULT_SIGBUS;
> > > 
> >   This looks good AFAICT.
> 
> I gave this a spin a couple of weeks ago (and accidentally left the test
> machines running for a full week!)  From what I can tell, with all the various
> wait_for_page_writeback stuff-ins, we've cut the frequency of writeback errors
> down to about 7-8 per day.  Not bad, but not fixed.
> 
> On the odd chance that jbd2 really can provide stable pages during writeback, I
> am now rerunning the test with no patches and data=journal, while noting that
> (a) DIO mode doesn't work with data=journal and (b) the first write failure
> will probably cause the journal to abort == game over.  When that's done I'll

Heh, nope, even with data=journal I still see checksum errors.  So much for
that hare-brained theory. :(

--D