xfs_repair fails after trying to format log cycle?

[Dave Chinner]

On Tue, Apr 12, 2016 at 11:02:37PM -0400, Andrew Ryder wrote:
> Is it possible the location its searching for at block
> 
> 02:34:43.887528 pread64(4, 0x7fb8f53e0200, 2097152, 3001552175104) =
> -1 EIO (Input/output error)

so offset is 3001552175104, or roughly around the 3TB mark. Given
the log i always placed int eh middle of the filesystem and you have
a 6TB device, then the above definitely looks like a valid place to
be reading from the log.

> xfs_logprint:
>     data device: 0x902
>     log device: 0x902 daddr: 5860130880 length: 4173824

daddr converted to offset is 5860130880 * 512 = 3001552175104, which
tells us that the above pread64 failure was definitely coming from
an attempt to read the log.

That this is coming from the block device from userspace indicates a
problem below XFS. There is something going wrong with your
underlying block device and/or hardware here; AFAICT it's not
related to XFS at all.

> >GNU Parted 3.2
> >Using /dev/sdk
> >Welcome to GNU Parted! Type 'help' to view a list of commands.
> >(parted) p
> >Model: ATA ST2000DL001-9VT1 (scsi)
> >Disk /dev/sdk: 2000GB
> >Sector size (logical/physical): 512B/512B
> >Partition Table: msdos
> >Disk Flags:
> >
> >Number  Start  End     Size    Type     File system  Flags
> >  1      512B   2000GB  2000GB  primary               raid
> >Number  Start  End          Size         Type     File system  Flags
> >  1      1s     3907029167s  3907029167s  primary               raid

Compared to the other devices, it has a different start sector, a
different size, and an msdos partition table rather than gpt.
Definitely a red flag...

> >>>This all began when the RR2722 driver running under 3.18.15
> >>>complained and

Reported physical IO errors to a write command. Really, this looks
like a hardware issue, not something that can be fixed by running
xfs_repair...

Cheers,

Dave.
-- 
Dave Chinner
david at fromorbit.com