Re: [Lsf-pc] [LSF/MM TOPIC] Badblocks checking/representation in filesystems

[Dan Williams <dan.j.williams@intel.com>]

On Thu, Jan 19, 2017 at 10:59 AM, Vishal Verma <vishal.l.verma@intel.com> wrote:
> On 01/19, Jan Kara wrote:
>> On Wed 18-01-17 21:56:58, Verma, Vishal L wrote:
>> > On Wed, 2017-01-18 at 13:32 -0800, Dan Williams wrote:
>> > > On Wed, Jan 18, 2017 at 1:02 PM, Darrick J. Wong
>> > > <darrick.wong@oracle.com> wrote:
>> > > > On Wed, Jan 18, 2017 at 03:39:17PM -0500, Jeff Moyer wrote:
>> > > > > Jan Kara <jack@suse.cz> writes:
>> > > > >
>> > > > > > On Tue 17-01-17 15:14:21, Vishal Verma wrote:
>> > > > > > > Your note on the online repair does raise another tangentially
>> > > > > > > related
>> > > > > > > topic. Currently, if there are badblocks, writes via the bio
>> > > > > > > submission
>> > > > > > > path will clear the error (if the hardware is able to remap
>> > > > > > > the bad
>> > > > > > > locations). However, if the filesystem is mounted eith DAX,
>> > > > > > > even
>> > > > > > > non-mmap operations - read() and write() will go through the
>> > > > > > > dax paths
>> > > > > > > (dax_do_io()). We haven't found a good/agreeable way to
>> > > > > > > perform
>> > > > > > > error-clearing in this case. So currently, if a dax mounted
>> > > > > > > filesystem
>> > > > > > > has badblocks, the only way to clear those badblocks is to
>> > > > > > > mount it
>> > > > > > > without DAX, and overwrite/zero the bad locations. This is a
>> > > > > > > pretty
>> > > > > > > terrible user experience, and I'm hoping this can be solved in
>> > > > > > > a better
>> > > > > > > way.
>> > > > > >
>> > > > > > Please remind me, what is the problem with DAX code doing
>> > > > > > necessary work to
>> > > > > > clear the error when it gets EIO from memcpy on write?
>> > > > >
>> > > > > You won't get an MCE for a store;  only loads generate them.
>> > > > >
>> > > > > Won't fallocate FL_ZERO_RANGE clear bad blocks when mounted with
>> > > > > -o dax?
>> > > >
>> > > > Not necessarily; XFS usually implements this by punching out the
>> > > > range
>> > > > and then reallocating it as unwritten blocks.
>> > > >
>> > >
>> > > That does clear the error because the unwritten blocks are zeroed and
>> > > errors cleared when they become allocated again.
>> >
>> > Yes, the problem was that writes won't clear errors. zeroing through
>> > either hole-punch, truncate, unlinking the file should all work
>> > (assuming the hole-punch or truncate ranges wholly contain the
>> > 'badblock' sector).
>>
>> Let me repeat my question: You have mentioned that if we do IO through DAX,
>> writes won't clear errors and we should fall back to normal block path to
>> do write to clear the error. What does prevent us from directly clearing
>> the error from DAX path?
>>
> With DAX, all IO goes through DAX paths. There are two cases:
> 1. mmap and loads/stores: Obviously there is no kernel intervention
> here, and no badblocks handling is possible.
> 2. read() or write() IO: In the absence of dax, this would go through
> the bio submission path, through the pmem driver, and that would handle
> error clearing. With DAX, this goes through dax_iomap_actor, which also
> doesn't go through the pmem driver (it does a dax mapping, followed by
> essentially memcpy), and hence cannot handle badblocks.

Hmm, that may no longer be true after my changes to push dax flushing
to the driver. I.e. we could have a copy_from_iter() implementation
that attempts to clear errors... I'll get that series out and we can
discuss there.