Re: Wrong DIF guard tag on ext2 write

[Nick Piggin <npiggin@suse.de>]

On Tue, Jun 01, 2010 at 03:50:59PM +0200, Christof Schmitt wrote:
> On Tue, Jun 01, 2010 at 09:03:25AM -0400, Chris Mason wrote:
> > On Tue, Jun 01, 2010 at 12:30:42PM +0200, Christof Schmitt wrote:
> > > On Mon, May 31, 2010 at 06:30:05PM +0300, Boaz Harrosh wrote:
> > > > On 05/31/2010 06:01 PM, James Bottomley wrote:
> > > > > On Mon, 2010-05-31 at 10:20 -0400, Martin K. Petersen wrote:
> > > > >>>>>>> "Christof" == Christof Schmitt <christof.schmitt@de.ibm.com> writes:
> > > > >>
> > > > >> Christof> Since the guard tags are created in Linux, it seems that the
> > > > >> Christof> data attached to the write request changes between the
> > > > >> Christof> generation in bio_integrity_generate and the call to
> > > > >> Christof> sd_prep_fn.
> > > > >>
> > > > >> Yep, known bug.  Page writeback locking is messed up for buffer_head
> > > > >> users.  The extNfs folks volunteered to look into this a while back but
> > > > >> I don't think they have found the time yet.
> > > > >>
> > > > >>
> > > > >> Christof> Using ext3 or ext4 instead of ext2 does not show the problem.
> > > > >>
> > > > >> Last I looked there were still code paths in ext3 and ext4 that
> > > > >> permitted pages to be changed during flight.  I guess you've just been
> > > > >> lucky.
> > > > > 
> > > > > Pages have always been modifiable in flight.  The OS guarantees they'll
> > > > > be rewritten, so the drivers can drop them if it detects the problem.
> > > > > This is identical to the iscsi checksum issue (iscsi adds a checksum
> > > > > because it doesn't trust TCP/IP and if the checksum is generated in
> > > > > software, there's time between generation and page transmission for the
> > > > > alteration to occur).  The solution in the iscsi case was not to
> > > > > complain if the page is still marked dirty.
> > > > > 
> > > > 
> > > > And also why RAID1 and RAID4/5/6 need the data bounced. I wish VFS
> > > > would prevent data writing given a device queue flag that requests
> > > > it. So all these devices and modes could just flag the VFS/filesystems
> > > > that: "please don't allow concurrent writes, otherwise I need to copy data"
> > > > 
> > > > From what Chris Mason has said before, all the mechanics are there, and it's
> > > > what btrfs is doing. Though I don't know how myself?
> > > 
> > > I also tested with btrfs and invalid guard tags in writes have been
> > > encountered as well (again in 2.6.34). The only difference is that no
> > > error was reported to userspace, although this might be a
> > > configuration issue.
> > 
> > This would be a btrfs bug.  We have strict checks in place that are
> > supposed to prevent buffers changing while in flight.  What was the
> > workload that triggered this problem?
> 
> I am running an internal test tool that creates files with a known
> pattern until the disk is full, reads the data to verify if the
> pattern is still intact, removes the files and starts over.

How are the checks done? The lock_page and wait_on_page_writeback from
prepare_pages?

Looks OK but AFAIKS you would need to unmap_mapping_range() the range
after taking the locks. You are also screwed if the page is in the
process of being modified by a get_user_pages() caller. You'd have to
add VM_IO to vm_flags to prevent get_user_pages.

Direct-IO is always going to have the same problems, so it's not like
block based solutions can just pretend it will easily go away.