Message-ID: <20110319000755.GD1110@tux1.beaverton.ibm.com>
Date:	Fri, 18 Mar 2011 17:07:55 -0700
From:	"Darrick J. Wong" <djwong@...ibm.com>
To:	Dave Chinner <david@...morbit.com>
Cc:	Chris Mason <chris.mason@...cle.com>, Jan Kara <jack@...e.cz>,
	Joel Becker <jlbec@...lplan.org>,
	"Martin K. Petersen" <martin.petersen@...cle.com>,
	Jens Axboe <axboe@...nel.dk>,
	linux-kernel <linux-kernel@...r.kernel.org>,
	linux-fsdevel <linux-fsdevel@...r.kernel.org>,
	Mingming Cao <mcao@...ibm.com>,
	linux-scsi <linux-scsi@...r.kernel.org>
Subject: Re: [RFC] block integrity: Fix write after checksum calculation
	problem

On Tue, Mar 08, 2011 at 03:56:26PM +1100, Dave Chinner wrote:
> On Fri, Mar 04, 2011 at 01:07:24PM -0800, Darrick J. Wong wrote:
> > On Mon, Feb 28, 2011 at 07:54:05AM -0500, Chris Mason wrote:
> > > Excerpts from Darrick J. Wong's message of 2011-02-24 13:27:32 -0500:
> > > > On Thu, Feb 24, 2011 at 12:37:53PM -0500, Chris Mason wrote:
> > > > > Excerpts from Jan Kara's message of 2011-02-24 11:47:58 -0500:
> > > > > > On Wed 23-02-11 15:35:11, Chris Mason wrote:
> > > > > > > Excerpts from Joel Becker's message of 2011-02-23 15:24:47 -0500:
> > > > > > > > On Tue, Feb 22, 2011 at 11:45:44AM -0500, Martin K. Petersen wrote:
> > > > > > > > > Also, DIX is only the tip of the iceberg. Many other impending
> > > > > > > > > technologies feature checksums and require pages to be stable during I/O
> > > > > > > > > due to checksumming, encryption and so on.
> > > > > > > > > 
> > > > > > > > > The VM is already trying to do the right thing. We just need the
> > > > > > > > > relevant filesystems to catch up.
> > > > > > > > 
> > > > > > > >     ocfs2 handles stable metadata for its checksums when feeding
> > > > > > > > things to the journal.  If we're doing pagecache-based I/O, is the
> > > > > > > > pagecache going to help here for data?
> > > > > > > 
> > > > > > > Data is much easier than metadata.  All you really need is to wait on
> > > > > > > writeback in file_write, wait on writeback in page_mkwrite, and make
> > > > > > > sure you don't free blocks back to the allocator that are actively under
> > > > > > > IO.
> > > > > > > 
> > > > > > > I expect the hard part to be jbd and metadata in ext34.
> > > > > >   But JBD already has to do data copy if a buffer is going to be modified
> > > > > > before/while it is written to the journal. So we should alredy do all that
> > > > > > is needed for metadata. I don't say there aren't any bugs as they could be
> > > > > > triggered only by crashing at the wrong moment and observing fs corruption.
> > > > > > But most of the work should be there...
> > > > > 
> > > > > Most of it is there, but there are always little bits and pieces.  The
> > > > > ext4 journal csumming code was one semi-recent example where we found
> > > > > metadata changing in flight.
> > > > > 
> > > > > A big part of testing this is getting some way to detect the bugs
> > > > > without dif/dix.  With btrfs I have patches to do set_memory_ro on
> > > > > pages once I've don the crc, hopefully we can generalize that idea or
> > > > > some up with something smarter.
> > > > 
> > > > Right now I'm faking it with modprobe scsi_debug ato=1 guard=1 dif=3 dix=199.
> > > > 
> > > > Hm, would you mind sharing those patches?  I've been working on a second patch
> > > > to do the wait-on-writeback per everyone's suggestions, but I still see the
> > > > occasional corruption error as soon as I enable the mmap write case and covet
> > > > some more debugging tools.  It does seem to be working for the pure pwrite()
> > > > case. :)
> > > 
> > > Here's an ext4 version of the debugging patch.  It's a few years old but
> > > it'll give you the idea.  This only covers metadata pages.
> > > 
> > > Looks like I hacked the btrfs version up and didn't keep the original,
> > > I'll have to rework it, I was trying to use it for the big corruption I
> > > fixed recently and made a bunch of changes.
> > > 
> > > For data if mmap is giving you trouble you need to wait on writeback in
> > > page_mkwrite, with the page locked.  fs/btrfs/inode.c has our
> > > page_mkwrite, which uses wait_on_page_writeback() and also the btrfs
> > > ordered write code.  But for the other filesystems, waiting on writeback
> > > should be enough.
> > 
> > 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*

After letting the stress tool run for a few days, I can say fairly confidently
that the write() case doesn't seem to fail regardless of the O_DIRECT setting.
However, with writes to mmap regions, failures happen about once every 20-40
minutes, even with O_DIRECT set.  To me this suggests some sort of race
condition that we seem to win except once every 20 minutes.

I then thought, if page_mkwrite contains a wait_on_page_writeback, then perhaps
there's something that I could do just prior to calculating the DIF checksum
that would cause any subsequent write attempts to be shuffled back into
page_mkwrite.  I tried the set_memory_ro thing again, though that led to some
recursive lock errors and I noticed that those functions only seem to exist in
arch/x86/.  Next I tried directly mucking with PTEs, in addition to feeling
messy, only seemed to corrupt memory. :)

Is there a "correct" way to take a writeable page and make it so that any
process trying to write to it ends up hitting the page fault handler where we
can then wait for writeback?  Or perhaps I am simply barking up the wrong tree?

(Just FYI I took the old copy-everything-to-bounce-buffers patch that few
people liked for a second spin, and the errors did not surface regardless of
what combination of write/mmap and directio/bufferedio I told it to use.)

--D
> > 
> > The set_memory_ro debugging trick didn't ferret out any write paths that I
> > didn't catch... though it did have the effect of causing occasional fsync()
> > deadlocks.  I suppose I could sprinkle in a few more of those write calls to
> > see what happens.
> > 
> > Either way, I'm emailing to ask everyone's advice since I've run out of ideas.
> > Or: Did I miss something?
> > 
> > Thanks all for the feedback so far!
> > 
> > --
> > fs: Wait for page writeback when rewrite detected
> > 
> > Signed-off-by: Darrick J. Wong <djwong@...ibm.com>
> > ---
> > 
> >  fs/buffer.c     |    4 +++-
> >  fs/ext4/inode.c |    3 +++
> >  mm/filemap.c    |   15 +++++++++++++--
> >  3 files changed, 19 insertions(+), 3 deletions(-)
> > 
> > diff --git a/fs/buffer.c b/fs/buffer.c
> > index 2219a76..39e934c 100644
> > --- a/fs/buffer.c
> > +++ b/fs/buffer.c
> > @@ -2379,8 +2379,10 @@ block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
> >  			ret = VM_FAULT_OOM;
> >  		else /* -ENOSPC, -EIO, etc */
> >  			ret = VM_FAULT_SIGBUS;
> > -	} else
> > +	} else {
> > +		wait_on_page_writeback(page);
> >  		ret = VM_FAULT_LOCKED;
> > +	}
> 
> I think this needs to wait before the __block_write_begin() call,
> not after it. i.e. wait before the page is mapped, not afterwards.
> 
> ....
> > diff --git a/mm/filemap.c b/mm/filemap.c
> > index 83a45d3..f201d80 100644
> > --- a/mm/filemap.c
> > +++ b/mm/filemap.c
> > @@ -2217,8 +2217,8 @@ EXPORT_SYMBOL(generic_file_direct_write);
> >   * Find or create a page at the given pagecache position. Return the locked
> >   * page. This function is specifically for buffered writes.
> >   */
> > -struct page *grab_cache_page_write_begin(struct address_space *mapping,
> > -					pgoff_t index, unsigned flags)
> > +struct page *__grab_cache_page_write_begin(struct address_space *mapping,
> > +					   pgoff_t index, unsigned flags)
> >  {
> >  	int status;
> >  	struct page *page;
> > @@ -2243,6 +2243,17 @@ repeat:
> >  	}
> >  	return page;
> >  }
> > +struct page *grab_cache_page_write_begin(struct address_space *mapping,
> > +					pgoff_t index, unsigned flags)
> > +{
> > +	struct page *p;
> > +
> > +	p = __grab_cache_page_write_begin(mapping, index, flags);
> > +	if (p)
> > +		wait_on_page_writeback(p);
> > +
> > +	return p;
> > +}
> >  EXPORT_SYMBOL(grab_cache_page_write_begin);
> 
> Not much point in add in a wrapper when nothing else calls
> __grab_cache_page_write_begin(), which should also be static....
> 
> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david@...morbit.com
> --
> To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
> the body of a message to majordomo@...r.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists