On Thu, Jun 29, 2017 at 09:57:27AM -0600, Ross Zwisler wrote:
> On Thu, Jun 29, 2017 at 09:35:38AM -0600, Ross Zwisler wrote:
> > Our validation team noticed that in some configurations mkfs.ext4 with the
> > -D option creates a filesystem that can't be mounted:
The -D option just means that we're doing the I/O using Direct I/O (as
opposed to buffered I/O). It shouldn't make any difference to what
gets written, so this very much smells like a bug in how /dev/pmem
supports Direct I/O...
> One more bit of info - this seems to be strongly tied to the size of the
> block device. With a 32 GB block device it works fine, with 248 GB you get
> overlap messages for groups 1 through 63, and with a 250 GB device you get
> overlaps for groups 1 through 1999.
This very much sounds like Direct I/O is just getting completely
botched for the pmem device, and writes to a block group descriptor
block is affecting the wrong place on the storage device.
- Ted
