On Fri, Nov 16, 2018 at 09:13:37AM +1100, Dave Chinner wrote:
> On Thu, Nov 15, 2018 at 11:10:36AM +0800, Ming Lei wrote:
> > On Thu, Nov 15, 2018 at 12:22:01PM +1100, Dave Chinner wrote:
> > > On Thu, Nov 15, 2018 at 09:06:52AM +0800, Ming Lei wrote:
> > > > On Wed, Nov 14, 2018 at 08:18:24AM -0700, Jens Axboe wrote:
> > > > > On 11/13/18 2:43 PM, Dave Chinner wrote:
> > > > > > From: Dave Chinner <dchinner@xxxxxxxxxx>
> > > > > >
> > > > > > A discard cleanup merged into 4.20-rc2 causes fstests xfs/259 to
> > > > > > fall into an endless loop in the discard code. The test is creating
> > > > > > a device that is exactly 2^32 sectors in size to test mkfs boundary
> > > > > > conditions around the 32 bit sector overflow region.
> > > > > >
> > > > > > mkfs issues a discard for the entire device size by default, and
> > > > > > hence this throws a sector count of 2^32 into
> > > > > > blkdev_issue_discard(). It takes the number of sectors to discard as
> > > > > > a sector_t - a 64 bit value.
> > > > > >
> > > > > > The commit ba5d73851e71 ("block: cleanup __blkdev_issue_discard")
> > > > > > takes this sector count and casts it to a 32 bit value before
> > > > > > comapring it against the maximum allowed discard size the device
> > > > > > has. This truncates away the upper 32 bits, and so if the lower 32
> > > > > > bits of the sector count is zero, it starts issuing discards of
> > > > > > length 0. This causes the code to fall into an endless loop, issuing
> > > > > > a zero length discards over and over again on the same sector.
> > > > >
> > > > > Applied, thanks. Ming, can you please add a blktests test for
> > > > > this case? This is the 2nd time it's been broken.
> > > >
> > > > OK, I will add zram discard test in blktests, which should cover the
> > > > 1st report. For the xfs/259, I need to investigate if it is easy to
> > > > do in blktests.
> > >
> > > Just write a test that creates block devices of 2^32 + (-1,0,1)
> > > sectors and runs a discard across the entire device. That's all that
> > > xfs/259 it doing - exercising mkfs on 2TB, 4TB and 16TB boundaries.
> > > i.e. the boundaries where sectors and page cache indexes (on 4k page
> > > size systems) overflow 32 bit int and unsigned int sizes. mkfs
> > > issues a discard for the entire device, so it's testing that as
> > > well...
> >
> > Indeed, I can reproduce this issue via the following commands:
> >
> > modprobe scsi_debug virtual_gb=2049 sector_size=512 lbpws10=1 dev_size_mb=512
> > blkdiscard /dev/sde
> >
> > >
> > > You need to write tests that exercise write_same, write_zeros and
> > > discard operations around these boundaries, because they all take
> > > a 64 bit sector count and stuff them into 32 bit size fields in
> > > the bio tha tis being submitted.
> >
> > write_same/write_zeros are usually used by driver directly, so we
> > may need make the test case on some specific device.
>
> My local linux iscsi server and client advertise support for them.
> It definitely does not ships zeros across the wire(*) when I use
> things like FALLOC_FL_ZERO_RANGE, but fstests does not have block
> device fallocate() tests for zeroing or punching...
fstests does (generic/{349,350,351}) but those basic functionality tests
don't include creating a 2^32 block device and seeing if overflows
happen... :/
...I also see that Eryu succeeded in kicking those tests out of the
quick group, so they probably don't run that often either.
--D
>
> Cheers,
>
> Dave.
>
> (*) but the back end storage is a sparse file on an XFS filesystem,
> and the iscsi server fails to translate write_zeroes or
> WRITE_SAME(0) to FALLOC_FL_ZERO_RANGE on the storage side and hence
> is really slow because it physically writes zeros to the XFS file.
> i.e. the client offloads the operation to the server to minimise
> wire traffic, but then the server doesn't offload the operation to
> the storage....
>
> Cheers,
>
> Dave.
> --
> Dave Chinner
> david@xxxxxxxxxxxxx
