On Wed, 4 Mar 2015 16:09:04 -0500 Dave Jones [off-list ref] wrote:
Hi Neil,
According to Documentation/iostats.txt, the 9th column of
/proc/diskstats (and its modern replacement in sysfs) should go to 0
as IO completes.
I assembled a RAID0 stripe using two SSD's, and saw this..
# mdadm --assemble /dev/md0
mdadm: /dev/md0 has been started with 2 drives.
# cat /sys/block/md0/stat
167 0 5656 0 5 0 4096 0 172 3408 582825
# cat /sys/block/md0/stat
167 0 5656 0 5 0 4096 0 172 231469 39809317
The 10th & 11th fields constantly increase, as field 9 remains non-zero.
If I mount and umount a filesystem on that volume, it works as expected,
but the 9th 'IOs inflight' field continues to rise and never decreases even
though the IO has obviously completed.
# umount /mnt/ssd
# cat /sys/block/md0/stat
167 0 5656 0 9 0 4225 0 176 571384 98278615
The underlying disks have their respective stats entries behaving as
expected, it only seems to affect the upper md layer.
Some missing accounting somewhere in md ?
(Only tested on 4.0rc2 so far, and only on RAID0)
Dave
blockdev stats often aren't really a good match for md/raid...
"in_flight" assumes a queue, and raid0 doesn't have one. It just redirects
each request to the relevant device and lets the device handle it.
The only useful thing we could do here is make that value always zero.
I guess I need to add a "generic_end_io_acct()" call to md.c somewhere.
For raid1/5 there is probably somewhere sensible to put it.
For raid0/linear, it probably goes immediately after generic_start_io_acct().
Patches welcome :-)
NeilBrown