On 12/1/2017 3:34 PM, Shaohua Li wrote:
On Thu, Nov 30, 2017 at 11:22:23AM -0500, Larkin Lowrey wrote:
Seems locks are involved, and in a not so nice way.
I had been chasing a problem that I thought was with the SAS controllers
where the IO for the drives would stall for extended periods of time. The
data from iostat would show 100% utilization but zero I/O and 1 or 2
avgqu-sz. Once the IO would resume the await value would be exactly the
amount of time the IO had been stalled. I was sure it was the SAS controller
or driver but it turns out it was the raid6 journal! During the periods of
stalled IO, the mdX_raid6 process would go to 100% CPU and when the IO
resumed that process would be back down at ~50%. When I switched back to
resync (disabled the journal), those IO stalls went away. It appears to me
that the bad locking behavior is the culprit.
My priority now is to disable the journals on my two remaining arrays which
boot up with mdX_raid6 at 100% CPU. Because of their state mdadm is unable
to act on them at all. How can I disable the journal of these arrays which
are 'inactive'?
I think we should fix the journal bug. Can you describe how you trigger this?
anything abnormal happens?
Thanks,
Shaohua
My tests were all with jounral_mode set to write-back. I didn't do any
testing in write-through mode since that's not very interesting to me.
The tests I ran were for various failure modes like sudden power off and
physically pulling multiple drives from the array simultaneously all
while writing to the array as fast as possible. The problem would be
apparent on boot. It looks like mdadm would read the journal and use it
to populate the stripe cache with dirty data. Once mdadm was done the
mdX_raid6 process would go to 100% and I would no longer be able to
control the array. All mdadm commands would hang and writes to
/sys/block/mdX/md/XYZ would hang. The array would remain 'inactive'
according to /prod/mdstat.
The mdadm v4.0-218-gfe05dc4 - 2017-11-09.
This happened with a 4.13 kernel. I did not have easy access to a 4.12
kernel or a 4.14. When I booted with 4.11.8 the array assembled just
fine. So, there was regression somewhere between 4.11.8 and 4.13.X.
The two machines I've been using for testing are no longer available to
me but I'll work on creating a VM I can use to try to reproduce this.
--Larkin
--Larkin
On 11/27/2017 9:42 PM, Larkin Lowrey wrote:
The kernel thread is spending a lot of time in handle_active_stripes but
not staying in there since I see it calling handle_stripe and then I
also see calls to __release_stripe then back to handle_stripe again. So,
it's clearly exiting handle_active_stripes but getting dumped right back
in there. Presumably, handle_active_stripes is not completing the
necessary work and the kernel thread keeps calling it expecting the
pending work to be completed.
When the system boots up an mdadm process runs at near 100% CPU with
sporadic reads of the journal device. After ~30s or so, the IO stops,
the mdadm process disappears, and the mdX_raid6 kernel thread starts
taking 100% CPU.
RIP: 0010:do_release_stripe+0x52/0x410 [raid456]
RSP: 0018:ffffbb7d41dcbcc0 EFLAGS: 00000017
RAX: 0000000002000000 RBX: ffff8f10c819bf20 RCX: 0000000000000000
RDX: 0000000000000002 RSI: ffff8f10c819bf20 RDI: ffff8f114a2ad400
RBP: ffffbb7d41dcbce8 R08: ffff8f113e3dd888 R09: 0000000000000000
R10: 00000000ffffffff R11: 0000000000000000 R12: ffff8f10c819bf30
R13: ffff8f113e3ddb58 R14: ffff8f113e3dd800 R15: 0000000000000003
FS: 0000000000000000(0000) GS:ffff8f115ed40000(0000)
knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007fb7675e1000 CR3: 0000000412cf5000 CR4: 00000000000406e0
Call Trace:
__release_stripe+0x16/0x20 [raid456]
handle_active_stripes.isra.60+0x44e/0x5a0 [raid456]
raid5d+0x42e/0x630 [raid456]
--Larkin
On 11/24/2017 2:20 PM, Larkin Lowrey wrote:
After hard-rebooting, this instance (stripe_cache_active: 2)
assembled just fine on boot. The next time I encountered this the
array was 'inactive' on boot. There was a flurry of I/O initially
(which seems to indicate journal re-play, then the array becoming
'active') but the I/O ceased without the array becoming active.
This time...
stripe_cache_active: 2376
md125 : inactive md127p4[9](J) sdk1[2] sdl1[3] sdn1[5] sdo1[6]
sdm1[4] sdj1[1] sdq1[8] sdp1[7]
31258219068 blocks super 1.2
# mdadm -D /dev/md125
/dev/md125:
Version : 1.2
Creation Time : Thu Oct 19 10:11:35 2017
Raid Level : raid6
Used Dev Size : 18446744073709551615
Raid Devices : 8
Total Devices : 9
Persistence : Superblock is persistent
Update Time : Fri Nov 24 13:41:38 2017
State : active, FAILED, Not Started
Active Devices : 8
Working Devices : 9
Failed Devices : 0
Spare Devices : 0
Layout : left-symmetric
Chunk Size : 64K
Consistency Policy : journal
Name : ########:3
UUID : de6a2ce0:1a4c510f:d7c89da4:1215a312
Events : 156844
Number Major Minor RaidDevice State
- 0 0 0 removed
- 0 0 1 removed
- 0 0 2 removed
- 0 0 3 removed
- 0 0 4 removed
- 0 0 5 removed
- 0 0 6 removed
- 0 0 7 removed
- 259 3 - spare /dev/md127p4
- 8 225 5 sync /dev/sdo1
- 8 209 4 sync /dev/sdn1
- 8 193 3 sync /dev/sdm1
- 8 177 2 sync /dev/sdl1
- 8 161 1 sync /dev/sdk1
- 8 145 0 sync /dev/sdj1
- 65 1 7 sync /dev/sdq1
- 8 241 6 sync /dev/sdp1
--Larkin
On 11/23/2017 1:22 PM, Larkin Lowrey wrote:
Sometimes, stopping a raid6 array (with journal) hangs, the
mdX_raid6 process pegs at 100% CPU, and there is no I/O. Looks
like it's stuck in an infinite loop.
Kernel: 4.13.13-200.fc26.x86_64
The stack trace (echo l > /proc/sysrq-trigger) is always the same:
handle_stripe+0x10c/0x2140 [raid456]
? pick_next_task_fair+0x491/0x550
handle_active_stripes.isra.60+0x3e5/0x5a0 [raid456]
raid5d+0x42e/0x630 [raid456]
? prepare_to_wait_event+0x79/0x160
md_thread+0x125/0x170
? md_thread+0x125/0x170
? finish_wait+0x80/0x80
kthread+0x125/0x140
? state_show+0x2f0/0x2f0
? kthread_park+0x60/0x60
? do_syscall_64+0x67/0x140
ret_from_fork+0x25/0x30
The array is healthy, has a journal, and writes were idle for
several minutes prior to running 'mdadm --stop'.
md124 : active raid6 sdt1[6] sds1[5] sdw1[1] sdx1[2] sdy1[3]
sdu1[7] sdv1[8] sdz1[4] md125p4[9](J)
23442092928 blocks super 1.2 level 6, 64k chunk,
algorithm 2 [8/8] [UUUUUUUU]
stripe_cache_active: 2
stripe_cache_size: 32768
array_state: write-pending
journal_mode: write-through [write-back]
consistency_policy: journal
--Larkin
--
To unsubscribe from this list: send the line "unsubscribe
linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
