FIO: kjournald blocked for more than 120 seconds
Subsystems (2)
Hi, Jens
When runnig FIO benchmark, kjournald blocked for more than 120 seconds. Detailed root cause analysis and proposed solutions as below.
Any comment is appreciated.
Hardware Environment
---------------------
13 SEAGATE ST373307FC disks in a JBOD, connected by a Qlogic ISP2312 Fibe Channel HBA.
Bug description
----------------
fio vsync random read 4K in 13 disks, 4 processes per disk, fio global paramter as below,
[global]
direct=0
ioengine=vsync
iodepth=256
iodepth_batch=32
size=2G
bs=4k
numjobs=2
loops=5
runtime=1200group_reporting
Tested 4 IO schedulers, issue is only seen in CFQ.
INFO: task kjournald:20558 blocked for more than 120 seconds.
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
kjournald D ffff810010820978 6712 20558 2ffff81022ddb1d10 0000000000000046 ffff81022e7baa10 ffffffff803ba6f2 ffff81022ecd0000 ffff8101e6dc9160 ffff81022ecd0348 000000008048b6cb 0000000000000086 ffff81022c4e8d30 0000000000000000 ffffffff80247537
Call Trace: [<ffffffff803ba6f2>] kobject_get+0x12/0x17 [<ffffffff80247537>] getnstimeofday+0x2f/0x83 [<ffffffff8029c1ac>] sync_buffer+0x0/0x3f [<ffffffff8066d195>] io_schedule+0x5d/0x9f [<ffffffff8029c1e7>] sync_buffer+0x3b/0x3f [<ffffffff8066d3f0>] __wait_on_bit+0x40/0x6f [<ffffffff8029c1ac>] sync_buffer+0x0/0x3f [<ffffffff8066d48b>] out_of_line_wait_on_bit+0x6c/0x78 [<ffffffff80243909>] wake_bit_function+0x0/0x23 [<ffffffff8029e3ad>] sync_dirty_buffer+0x98/0xcb [<ffffffff8030056b>] journal_commit_transaction+0x97d/0xcb6 [<ffffffff8023a676>] lock_timer_base+0x26/0x4b [<ffffffff8030300a>] kjournald+0xc1/0x1fb [<ffffffff802438db>] autoremove_wake_function+0x0/0x2e [<ffffffff80302f49>] kjournald+0x0/0x1fb [<ffffffff802437bb>] kthread+0x47/0x74 [<ffffffff8022de51>] schedule_tail+0x28/0x5d [<ffffffff8020cac8>] child_rip+0xa/0x12 [<ffffffff80243774>] kthread+0x0/0x74 [<ffffffff8020cabe>] child_rip+0x0/0x12
Other people reported the similar issue on LKML.
http://lkml.org/lkml/2008/3/26/494
I did lots of kernel instrumentation and testing, and located the root cause.
Root cause analysis
-------------------
kjournald is blocked at:journal_write_commit_record ->sync_dirty_buffer(bh) ->wait_on_buffer(bh)
Buffer is submitted but it never finished.
git bisect to below patch,
commit cc19747977824ece6aa1c56a29e974fef5ec2b32
Author: Jens Axboe <[email protected]> Date: Fri Apr 20 20:45:39 2007 +0200
cfq-iosched: tighten queue request overlap condition
For tagged devices, allow overlap of requests if the idle window isn't enabled on the current active queue.
block/cfq-iosched.c +0 -0
@@ -989,7 +989,8 @@ static struct cfq_queue *cfq_select_queue(struct
cfq_data *cfqd)
* flight or is idling for a new request, allow either of these
* conditions to happen (or time out) before selecting a new queue. */
- if (cfqq->dispatched || timer_pending(&cfqd->idle_slice_timer)) {
+ if (timer_pending(&cfqd->idle_slice_timer) ||
+ (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
cfqq = NULL;
goto keep_queue;
}
The disks of my testing machine are tagged devices, so the CFQ idle window is disabled. In other words, the active queue of tagged devices(cfqd->hw_tag=1) never idle for a new request.
This causes active queue be expired immediately if it's empty, although it has not run out of time. CFQ will select next queue as active queue. In this testcase, there are thousands of FIO read requests in sync queues, only a few write requests by journal_write_commit_record in async queues.
In the other hand, all processes use the default io class and priority. They share the async queue for the same device, but have their own sync queue, so the sync queue number is 4 while asyn queue number is just 1 for the same device.
So sync queue has much more chances be selected as new active queue than async queue.
Sync queues do not idle and they are dispatched all the time. This leads to many unfinished requests in external queue,
namely, cfqd->sync_flight > 0.
static int cfq_dispatch_requests (...) {
....
while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
....
if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
break;
....
__cfq_dispatch_requests(cfqq)
}
....
}When cfq_select_queue selects the async queue which includes kjournald's write request, this selected async queue will never be dispatched since cfqd->sync_flight > 0, so kjournald is blocked.
Proposed 3 solutions
------------------
1. Do not check cfqd->sync_flight
- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
- break;
2. If we do need to check cfqd->sync_flight, then for tagged devices, we should give a little more chances to async queue to be dispatched.
@@ -1102,7 +1102,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force)
break;
}- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
+ if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq) && !cfqd->hw_tag)
break;
3. Force write request issued by journal_write_commit_record as sync request. As a matter of fact, it looks like most write requests submitted by kjournald is async request. We need convert them to sync requests.
Cheers,
Lin Ming
Quoted 162 lines · Lin Ming
Hi, Jens
When runnig FIO benchmark, kjournald blocked for more than 120 seconds. Detailed root cause analysis and proposed solutions as below.
Any comment is appreciated.
Hardware Environment
---------------------
13 SEAGATE ST373307FC disks in a JBOD, connected by a Qlogic ISP2312 Fibe Channel HBA.Bug description
----------------
fio vsync random read 4K in 13 disks, 4 processes per disk, fio global paramter as below,
[global]
direct=0
ioengine=vsync
iodepth=256
iodepth_batch=32
size=2G
bs=4k
numjobs=2
loops=5
runtime=1200group_reportingTested 4 IO schedulers, issue is only seen in CFQ.
INFO: task kjournald:20558 blocked for more than 120 seconds."echo 0 >
/proc/sys/kernel/hung_task_timeout_secs" disables this message.
kjournald Dffff8100108209786712 20558 2ffff81022ddb1d100000000000000046ffff81022e7baa10ffffffff803ba6f2ffff81022ecd0000ffff8101e6dc9160ffff81022ecd0348000000008048b6cb0000000000000086ffff81022c4e8d300000000000000000ffffffff80247537Call Trace: [<ffffffff803ba6f2>] kobject_get+0x12/0x17 [<ffffffff80247537>] getnstimeofday+0x2f/0x83 [<ffffffff8029c1ac>] sync_buffer+0x0/0x3f [<ffffffff8066d195>] io_schedule+0x5d/0x9f [<ffffffff8029c1e7>] sync_buffer+0x3b/0x3f [<ffffffff8066d3f0>] __wait_on_bit+0x40/0x6f [<ffffffff8029c1ac>] sync_buffer+0x0/0x3f [<ffffffff8066d48b>] out_of_line_wait_on_bit+0x6c/0x78 [<ffffffff80243909>] wake_bit_function+0x0/0x23 [<ffffffff8029e3ad>] sync_dirty_buffer+0x98/0xcb [<ffffffff8030056b>] journal_commit_transaction+0x97d/0xcb6 [<ffffffff8023a676>] lock_timer_base+0x26/0x4b [<ffffffff8030300a>] kjournald+0xc1/0x1fb [<ffffffff802438db>] autoremove_wake_function+0x0/0x2e [<ffffffff80302f49>] kjournald+0x0/0x1fb [<ffffffff802437bb>] kthread+0x47/0x74 [<ffffffff8022de51>] schedule_tail+0x28/0x5d [<ffffffff8020cac8>] child_rip+0xa/0x12 [<ffffffff80243774>] kthread+0x0/0x74 [<ffffffff8020cabe>] child_rip+0x0/0x12Other people reported the similar issue on LKML.
http://lkml.org/lkml/2008/3/26/494I did lots of kernel instrumentation and testing, and located the root cause.
Root cause analysis
-------------------kjournald is blocked at:
journal_write_commit_record->sync_dirty_buffer(bh)->wait_on_buffer(bh)Buffer is submitted but it never finished.
git bisect to below patch,
commit
cc19747977824ece6aa1c56a29e974fef5ec2b32Author: Jens Axboe <[email protected]> Date: Fri Apr 20 20:45:39 2007 +0200cfq-iosched: tighten queue request overlap condition
For tagged devices, allow overlap of requests if the idle window isn't enabled on the current active queue.
Signed-off-by Jens Axboe <[email protected]>block/cfq-iosched.c +2
-1@@ -989,7 +989,8 @@ static struct cfq_queue *cfq_select_queue(structcfq_data *cfqd) * flight or is idling for a new request, allow either of these * conditions to happen (or time out) before selecting a new queue. */ -if (cfqq->dispatched ||timer_pending (&cfqd->idle_slice_timer)) { +if (timer_pending (&cfqd->idle_slice_timer) || + (cfqq->dispatched &&cfq_cfqq_idle_window (cfqq))) { cfqq =NULL ; goto keep_queue;}
The disks of my testing machine are tagged devices, so the CFQ idle window is disabled. In other words, the active queue of tagged
devices(cfqd->hw_tag=1)never idle for a new request.This causes active queue be expired immediately if it's empty, although it has not run out of time. CFQ will select next queue as active queue. In this testcase, there are thousands of FIO read requests in sync queues, only a few write requests by
journal_write_commit_recordin async queues.In the other hand, all processes use the default io class and priority. They share the async queue for the same device, but have their own sync queue, so the sync queue number is 4 while asyn queue number is just 1 for the same device.
So sync queue has much more chances be selected as new active queue than async queue.
Sync queues do not idle and they are dispatched all the time. This leads to many unfinished requests in external queue,
namely,cfqd->sync_flight> 0.static int cfq_dispatch_requests (...) { .... while ((cfqq = cfq_select_queue(cfqd)) != NULL) { .... if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) break; .... __cfq_dispatch_requests(cfqq) } .... }When
cfq_select_queueselects the async queue which includes kjournald's write request, this selected async queue will never be dispatched sincecfqd->sync_flight> 0, so kjournald is blocked.Proposed 3 solutions
------------------
1. Do not checkcfqd->sync_flightblock/cfq-iosched.c +0
-2-if (cfqd->sync_flight && !cfq_cfqq_sync (cfqq)) -break ;2. If we do need to check
cfqd->sync_flight, then for tagged devices, we should give a little more chances to async queue to be dispatched.block/cfq-iosched.c +1
-1@@ -1102,7 +1102,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force) break; } -if (cfqd->sync_flight && !cfq_cfqq_sync (cfqq)) +if (cfqd->sync_flight && !cfq_cfqq_sync (cfqq) && ! cfqd->hw_tag)break ;
Quoted 4 lines
3. Force write request issued by
journal_write_commit_recordas sync request. As a matter of fact, it looks like most write requests submitted by kjournald is async request. We need convert them to sync requests.
Thanks for the very detailed analysis of the problem, complete with suggestions. While I think that any code that does:
submit async io
wait for it
should be issuing sync IO (or, better, automatically upgrade the request from async -> sync), we cannot rely on that.
This problem is similar in nature to device starvation, and a classic solution to that problem is to issue occasional ordered tags to prevent indefinite starvation. Perhaps we can apply some similar logic here.
For 2.6.26, the simple approach of just removing the sync_flight check is probably the safest.
Signature · 1 line
Jens Axboe
Quoted 11 lines
Quoted 11 lines
-----Original Message-----
From: Jens Axboe [mailto:[email protected]]
Sent: Tuesday, June 17, 2008 3:30 AM
To: Lin, Ming M
Cc: Zhang, Yanmin; Linux Kernel Mailing List
Subject: Re: FIO: kjournald blocked for more than 120 secondsQuoted 3 lines · Lin Ming
Hi, Jens
When runnig FIO benchmark, kjournald blocked for more than 120
seconds.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Detailed root cause analysis and proposed solutions as below.
Any comment is appreciated.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Hardware Environment
---------------------
13 SEAGATE ST373307FC disks in a JBOD, connected by a Qlogic ISP2312 Fibe Channel HBA.Bug description
----------------
fio vsync random read 4K in 13 disks, 4 processes per disk, fio
global
Quoted 16 lines
Quoted 16 lines
Quoted 16 lines
paramter as below,
Tested 4 IO schedulers, issue is only seen in CFQ.INFO: task kjournald:20558 blocked for more than 120 seconds."echo 0 >
/proc/sys/kernel/hung_task_timeout_secs" disables this message.
kjournald Dffff8100108209786712 20558 2ffff81022ddb1d100000000000000046ffff81022e7baa10ffffffff803ba6f2ffff81022ecd0000ffff8101e6dc9160ffff81022ecd0348000000008048b6cb0000000000000086ffff81022c4e8d300000000000000000ffffffff80247537Call Trace: [<ffffffff803ba6f2>] kobject_get+0x12/0x17The disks of my testing machine are tagged devices, so the CFQ idle window is disabled. In other words, the active queue of tagged
devices(cfqd->hw_tag=1)never idle for a new request.
Quoted 1 line
Quoted 1 line
Quoted 1 line
This causes active queue be expired immediately if it's empty,
although
Quoted 1 line
Quoted 1 line
Quoted 1 line
it has not run out of time. CFQ will select next queue as active
queue.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
In this testcase, there are thousands of FIO read requests in sync queues, only a few write requests by
journal_write_commit_recordin async queues.In the other hand, all processes use the default io class and
priority.
Quoted 1 line
Quoted 1 line
Quoted 1 line
They share the async queue for the same device, but have their own
sync
Quoted 1 line
Quoted 1 line
Quoted 1 line
queue, so the sync queue number is 4 while asyn queue number is just
1
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
for the same device.
So sync queue has much more chances be selected as new active queue
than
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
async queue.
Sync queues do not idle and they are dispatched all the time. This
leads
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
to many unfinished requests in external queue,
namely,cfqd->sync_flight> 0.static int cfq_dispatch_requests (...) { .... while ((cfqq = cfq_select_queue(cfqd)) != NULL) { .... if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) break; .... __cfq_dispatch_requests(cfqq) } .... }
Quoted 1 line
Quoted 1 line
Quoted 1 line
When
cfq_select_queueselects the async queue which includes
kjournald's
Quoted 1 line
Quoted 1 line
Quoted 1 line
write request, this selected async queue will never be dispatched
since
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
cfqd->sync_flight> 0, so kjournald is blocked.Proposed 3 solutions
------------------
1. Do not checkcfqd->sync_flight(quoted diff) +0
-2- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) - break;2. If we do need to check
cfqd->sync_flight, then for tagged
devices, we
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
should give a little more chances to async queue to be dispatched.
(quoted diff) +0
-0@@ -1102,7 +1102,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force) break; }
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
(quoted diff) +1
-1- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) + if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq) && ! cfqd->hw_tag) break;3. Force write request issued by
journal_write_commit_recordas sync request. As a matter of fact, it looks like most write requests submitted by kjournald is async request. We need convert them to
sync
Quoted 9 lines
Quoted 9 lines
Quoted 1 line
requests.
Thanks for the very detailed analysis of the problem, complete with suggestions. While I think that any code that does:
submit async io
wait for itshould be issuing sync IO (or, better, automatically upgrade the
request
Quoted 1 line
Quoted 1 line
from async -> sync), we cannot rely on that.
[YM] We can talk case by case. We could convert some important async io codes
to sync io codes at least. For example, kjournald callssync_dirty_buffer what
we captured in this case.
Another case is writeback. If processes do mmapped I/O and they might stop in
page fault to wait writeback finishing. Or a buffer write might trigger a dirty
page balance. As the latest kernel is more aggressive to start writeback, it might
be an issue now.
Quoted 3 lines
Quoted 3 lines
This problem is similar in nature to device starvation, and a classic solution to that problem is to issue occasional ordered tags to
prevent
Quoted 4 lines
Quoted 4 lines
indefinite starvation. Perhaps we can apply some similar logic here.
For 2.6.26, the simple approach of just removing the
sync_flightcheck is probably the safest.
Quoted 117 lines · Yanmin
Quoted 11 lines
Quoted 11 lines
-----Original Message-----
From: Jens Axboe [mailto:[email protected]]
Sent: Tuesday, June 17, 2008 3:30 AM
To: Lin, Ming M
Cc: Zhang, Yanmin; Linux Kernel Mailing List
Subject: Re: FIO: kjournald blocked for more than 120 secondsQuoted 3 lines · Lin Ming
Hi, Jens
When runnig FIO benchmark, kjournald blocked for more than 120
seconds.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Detailed root cause analysis and proposed solutions as below.
Any comment is appreciated.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Hardware Environment
---------------------
13 SEAGATE ST373307FC disks in a JBOD, connected by a Qlogic ISP2312 Fibe Channel HBA.Bug description
----------------
fio vsync random read 4K in 13 disks, 4 processes per disk, fioglobal
Quoted 16 lines
Quoted 16 lines
Quoted 16 lines
paramter as below,
Tested 4 IO schedulers, issue is only seen in CFQ.INFO: task kjournald:20558 blocked for more than 120 seconds."echo 0 >
/proc/sys/kernel/hung_task_timeout_secs" disables this message.
kjournald Dffff8100108209786712 20558 2ffff81022ddb1d100000000000000046ffff81022e7baa10ffffffff803ba6f2ffff81022ecd0000ffff8101e6dc9160ffff81022ecd0348000000008048b6cb0000000000000086ffff81022c4e8d300000000000000000ffffffff80247537Call Trace: [<ffffffff803ba6f2>] kobject_get+0x12/0x17The disks of my testing machine are tagged devices, so the CFQ idle window is disabled. In other words, the active queue of tagged
devices(cfqd->hw_tag=1)never idle for a new request.Quoted 1 line
Quoted 1 line
Quoted 1 line
This causes active queue be expired immediately if it's empty,
although
Quoted 1 line
Quoted 1 line
Quoted 1 line
it has not run out of time. CFQ will select next queue as active
queue.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
In this testcase, there are thousands of FIO read requests in sync queues, only a few write requests by
journal_write_commit_recordin async queues.In the other hand, all processes use the default io class and
priority.
Quoted 1 line
Quoted 1 line
Quoted 1 line
They share the async queue for the same device, but have their own
sync
Quoted 1 line
Quoted 1 line
Quoted 1 line
queue, so the sync queue number is 4 while asyn queue number is just
1
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
for the same device.
So sync queue has much more chances be selected as new active queue
than
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
async queue.
Sync queues do not idle and they are dispatched all the time. This
leads
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
to many unfinished requests in external queue,
namely,cfqd->sync_flight> 0.static int cfq_dispatch_requests (...) { .... while ((cfqq = cfq_select_queue(cfqd)) != NULL) { .... if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) break; .... __cfq_dispatch_requests(cfqq) } .... }Quoted 1 line
Quoted 1 line
Quoted 1 line
When
cfq_select_queueselects the async queue which includeskjournald's
Quoted 1 line
Quoted 1 line
Quoted 1 line
write request, this selected async queue will never be dispatched
since
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
cfqd->sync_flight> 0, so kjournald is blocked.Proposed 3 solutions
------------------
1. Do not checkcfqd->sync_flight(quoted diff) +0
-2- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) - break;2. If we do need to check
cfqd->sync_flight, then for taggeddevices, we
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
should give a little more chances to async queue to be dispatched.
(quoted diff) +0
-0@@ -1102,7 +1102,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force) break; }Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
(quoted diff) +1
-1- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) + if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq) && ! cfqd->hw_tag) break;3. Force write request issued by
journal_write_commit_recordas sync request. As a matter of fact, it looks like most write requests submitted by kjournald is async request. We need convert them tosync
Quoted 8 lines
Quoted 8 lines
Quoted 1 line
requests.
Thanks for the very detailed analysis of the problem, complete with suggestions. While I think that any code that does:
submit async io
wait for it
Quoted 8 lines
Quoted 1 line
Quoted 1 line
should be issuing sync IO (or, better, automatically upgrade the
request
Quoted 1 line
Quoted 1 line
from async -> sync), we cannot rely on that.
[YM] We can talk case by case. We could convert some important async io codes
to sync io codes at least. For example, kjournald callssync_dirty_bufferwhat
we captured in this case.
I agree, we should fix the obvious cases. My point was merely that there will probably always be missed cases, so we should attempt to handle it in the scheduler as well. Does the below buffer patch make it any better?
Quoted 7 lines
Another case is writeback. If processes do mmapped I/O and they might stop in
page fault to wait writeback finishing. Or a buffer write might trigger a dirty
page balance. As the latest kernel is more aggressive to start writeback, it might
be an issue now.
Sync process getting stuck in async writeout is another problem of the same variety.
fs/buffer.c +1 -1
@@ -2978,7 +2978,7 @@ int sync_dirty_buffer(struct buffer_head *bh)if (test_clear_buffer_dirty (bh)) {get_bh (bh); bh->b_end_io = end_buffer_write_sync; - ret =submit_bh (WRITE , bh); + ret =submit_bh (WRITE_SYNC , bh);wait_on_buffer (bh);if (buffer_eopnotsupp (bh)) {clear_buffer_eopnotsupp (bh);
Signature · 1 line
Jens Axboe
Quoted 127 lines · Jens Axboe
Quoted 117 lines · Yanmin
Quoted 11 lines
Quoted 11 lines
-----Original Message-----
From: Jens Axboe [mailto:[email protected]]
Sent: Tuesday, June 17, 2008 3:30 AM
To: Lin, Ming M
Cc: Zhang, Yanmin; Linux Kernel Mailing List
Subject: Re: FIO: kjournald blocked for more than 120 secondsQuoted 3 lines · Lin Ming
Hi, Jens
When runnig FIO benchmark, kjournald blocked for more than 120
seconds.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Detailed root cause analysis and proposed solutions as below.
Any comment is appreciated.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Hardware Environment
---------------------
13 SEAGATE ST373307FC disks in a JBOD, connected by a Qlogic ISP2312 Fibe Channel HBA.Bug description
----------------
fio vsync random read 4K in 13 disks, 4 processes per disk, fioglobal
Quoted 16 lines
Quoted 16 lines
Quoted 16 lines
paramter as below,
Tested 4 IO schedulers, issue is only seen in CFQ.INFO: task kjournald:20558 blocked for more than 120 seconds."echo 0 >
/proc/sys/kernel/hung_task_timeout_secs" disables this message.
kjournald Dffff8100108209786712 20558 2ffff81022ddb1d100000000000000046ffff81022e7baa10ffffffff803ba6f2ffff81022ecd0000ffff8101e6dc9160ffff81022ecd0348000000008048b6cb0000000000000086ffff81022c4e8d300000000000000000ffffffff80247537Call Trace: [<ffffffff803ba6f2>] kobject_get+0x12/0x17The disks of my testing machine are tagged devices, so the CFQ idle window is disabled. In other words, the active queue of tagged
devices(cfqd->hw_tag=1)never idle for a new request.Quoted 1 line
Quoted 1 line
Quoted 1 line
This causes active queue be expired immediately if it's empty,
although
Quoted 1 line
Quoted 1 line
Quoted 1 line
it has not run out of time. CFQ will select next queue as active
queue.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
In this testcase, there are thousands of FIO read requests in sync queues, only a few write requests by
journal_write_commit_recordin async queues.In the other hand, all processes use the default io class and
priority.
Quoted 1 line
Quoted 1 line
Quoted 1 line
They share the async queue for the same device, but have their own
sync
Quoted 1 line
Quoted 1 line
Quoted 1 line
queue, so the sync queue number is 4 while asyn queue number is just
1
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
for the same device.
So sync queue has much more chances be selected as new active queue
than
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
async queue.
Sync queues do not idle and they are dispatched all the time. This
leads
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
to many unfinished requests in external queue,
namely,cfqd->sync_flight> 0.static int cfq_dispatch_requests (...) { .... while ((cfqq = cfq_select_queue(cfqd)) != NULL) { .... if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) break; .... __cfq_dispatch_requests(cfqq) } .... }Quoted 1 line
Quoted 1 line
Quoted 1 line
When
cfq_select_queueselects the async queue which includeskjournald's
Quoted 1 line
Quoted 1 line
Quoted 1 line
write request, this selected async queue will never be dispatched
since
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
cfqd->sync_flight> 0, so kjournald is blocked.Proposed 3 solutions
------------------
1. Do not checkcfqd->sync_flight(quoted diff) +0
-2- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) - break;2. If we do need to check
cfqd->sync_flight, then for taggeddevices, we
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
should give a little more chances to async queue to be dispatched.
(quoted diff) +0
-0@@ -1102,7 +1102,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force) break; }Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
(quoted diff) +1
-1- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) + if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq) && ! cfqd->hw_tag) break;3. Force write request issued by
journal_write_commit_recordas sync request. As a matter of fact, it looks like most write requests submitted by kjournald is async request. We need convert them tosync
Quoted 8 lines
Quoted 8 lines
Quoted 1 line
requests.
Thanks for the very detailed analysis of the problem, complete with suggestions. While I think that any code that does:
submit async io
wait for itQuoted 8 lines
Quoted 1 line
Quoted 1 line
should be issuing sync IO (or, better, automatically upgrade the
request
Quoted 1 line
Quoted 1 line
from async -> sync), we cannot rely on that.
[YM] We can talk case by case. We could convert some important async io codes
to sync io codes at least. For example, kjournald callssync_dirty_bufferwhat
we captured in this case.
Quoted 4 lines
I agree, we should fix the obvious cases. My point was merely that there will probably always be missed cases, so we should attempt to handle it in the scheduler as well. Does the below buffer patch make it any better?
Yes, kjournald blocked issue is gone with below patch applied.
Lin Ming
Quoted 26 lines
Quoted 8 lines
Another case is writeback. If processes do mmapped I/O and they might stop in
page fault to wait writeback finishing. Or a buffer write might trigger a dirty
page balance. As the latest kernel is more aggressive to start writeback, it might
be an issue now.Sync process getting stuck in async writeout is another problem of the same variety.
fs/buffer.c +1
-1@@ -2978,7 +2978,7 @@ int sync_dirty_buffer(struct buffer_head *bh)if (test_clear_buffer_dirty (bh)) {get_bh (bh); bh->b_end_io = end_buffer_write_sync; - ret =submit_bh (WRITE , bh); + ret =submit_bh (WRITE_SYNC , bh);wait_on_buffer (bh);if (buffer_eopnotsupp (bh)) {clear_buffer_eopnotsupp (bh);
Quoted 129 lines · Lin Ming
Quoted 127 lines · Jens Axboe
Quoted 125 lines · Yanmin
Quoted 11 lines
Quoted 11 lines
-----Original Message-----
From: Jens Axboe [mailto:[email protected]]
Sent: Tuesday, June 17, 2008 3:30 AM
To: Lin, Ming M
Cc: Zhang, Yanmin; Linux Kernel Mailing List
Subject: Re: FIO: kjournald blocked for more than 120 secondsQuoted 3 lines · Lin Ming
Hi, Jens
When runnig FIO benchmark, kjournald blocked for more than 120
seconds.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Detailed root cause analysis and proposed solutions as below.
Any comment is appreciated.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Hardware Environment
---------------------
13 SEAGATE ST373307FC disks in a JBOD, connected by a Qlogic ISP2312 Fibe Channel HBA.Bug description
----------------
fio vsync random read 4K in 13 disks, 4 processes per disk, fioglobal
Quoted 16 lines
Quoted 16 lines
Quoted 16 lines
paramter as below,
Tested 4 IO schedulers, issue is only seen in CFQ.INFO: task kjournald:20558 blocked for more than 120 seconds."echo 0 >
/proc/sys/kernel/hung_task_timeout_secs" disables this message.
kjournald Dffff8100108209786712 20558 2ffff81022ddb1d100000000000000046ffff81022e7baa10ffffffff803ba6f2ffff81022ecd0000ffff8101e6dc9160ffff81022ecd0348000000008048b6cb0000000000000086ffff81022c4e8d300000000000000000ffffffff80247537Call Trace: [<ffffffff803ba6f2>] kobject_get+0x12/0x17The disks of my testing machine are tagged devices, so the CFQ idle window is disabled. In other words, the active queue of tagged
devices(cfqd->hw_tag=1)never idle for a new request.Quoted 1 line
Quoted 1 line
Quoted 1 line
This causes active queue be expired immediately if it's empty,
although
Quoted 1 line
Quoted 1 line
Quoted 1 line
it has not run out of time. CFQ will select next queue as active
queue.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
In this testcase, there are thousands of FIO read requests in sync queues, only a few write requests by
journal_write_commit_recordin async queues.In the other hand, all processes use the default io class and
priority.
Quoted 1 line
Quoted 1 line
Quoted 1 line
They share the async queue for the same device, but have their own
sync
Quoted 1 line
Quoted 1 line
Quoted 1 line
queue, so the sync queue number is 4 while asyn queue number is just
1
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
for the same device.
So sync queue has much more chances be selected as new active queue
than
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
async queue.
Sync queues do not idle and they are dispatched all the time. This
leads
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
to many unfinished requests in external queue,
namely,cfqd->sync_flight> 0.static int cfq_dispatch_requests (...) { .... while ((cfqq = cfq_select_queue(cfqd)) != NULL) { .... if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) break; .... __cfq_dispatch_requests(cfqq) } .... }Quoted 1 line
Quoted 1 line
Quoted 1 line
When
cfq_select_queueselects the async queue which includeskjournald's
Quoted 1 line
Quoted 1 line
Quoted 1 line
write request, this selected async queue will never be dispatched
since
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
cfqd->sync_flight> 0, so kjournald is blocked.Proposed 3 solutions
------------------
1. Do not checkcfqd->sync_flight(quoted diff) +0
-2- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) - break;2. If we do need to check
cfqd->sync_flight, then for taggeddevices, we
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
should give a little more chances to async queue to be dispatched.
(quoted diff) +0
-0@@ -1102,7 +1102,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force) break; }Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
(quoted diff) +1
-1- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) + if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq) && ! cfqd->hw_tag) break;3. Force write request issued by
journal_write_commit_recordas sync request. As a matter of fact, it looks like most write requests submitted by kjournald is async request. We need convert them tosync
Quoted 9 lines
Quoted 9 lines
Quoted 1 line
requests.
Thanks for the very detailed analysis of the problem, complete with suggestions. While I think that any code that does:
submit async io
wait for itshould be issuing sync IO (or, better, automatically upgrade the
request
Quoted 1 line
Quoted 1 line
from async -> sync), we cannot rely on that.
[YM] We can talk case by case. We could convert some important async io codes
to sync io codes at least. For example, kjournald callssync_dirty_bufferwhat
we captured in this case.
Quoted 6 lines
Quoted 4 lines
I agree, we should fix the obvious cases. My point was merely that there will probably always be missed cases, so we should attempt to handle it in the scheduler as well. Does the below buffer patch make it any better?
Yes, kjournald blocked issue is gone with below patch applied.
I think it's obviously the right thing to do, but I'm also a bit worried about applying it so close to 2.6.26 release. OTOH, we need to do SOMETHING for 2.6.26 release, so...
Quoted 30 lines
Lin Ming
Quoted 26 lines
Quoted 8 lines
Another case is writeback. If processes do mmapped I/O and they might stop in
page fault to wait writeback finishing. Or a buffer write might trigger a dirty
page balance. As the latest kernel is more aggressive to start writeback, it might
be an issue now.Sync process getting stuck in async writeout is another problem of the same variety.
fs/buffer.c +1
-1@@ -2978,7 +2978,7 @@ int sync_dirty_buffer(struct buffer_head *bh)if (test_clear_buffer_dirty (bh)) {get_bh (bh); bh->b_end_io = end_buffer_write_sync; - ret =submit_bh (WRITE , bh); + ret =submit_bh (WRITE_SYNC , bh);wait_on_buffer (bh);if (buffer_eopnotsupp (bh)) {clear_buffer_eopnotsupp (bh);
Signature · 1 line
Jens Axboe
Jens,
How about below patch?
If cfqd->sync_flight > 0, we still allow async queue to be dispached, and its slice time would be decreased. So sync queue never starve async queue.
Introduce a sysfs interface cfq_slice_async_penalty to adjust the percentages of slice that will be decreased.
--- block/cfq-iosched.c | 33 ++++++++++++++++++++++++++++++++- 1 files changed, 32 insertions(+), 1 deletions(-)
block/cfq-iosched.c +7 -0
@@ -26,6 +26,12 @@ static const int cfq_slice_sync = HZ / 10;static int cfq_slice_async =HZ /25 ;static const int cfq_slice_async_rq =2 ;static int cfq_slice_idle =HZ /125 ; +/* +* penalty of dispatching async queue +* if there are unfinished requests in sync queue +* by default, 50% of time slice decreased +*/ +static int cfq_slice_async_penalty =50 ;/* * offset from end of service tree
@@ -114,6 +120,7 @@ struct cfq_data {unsigned int cfq_slice[2 ];unsigned int cfq_slice_async_rq;unsigned int cfq_slice_idle; +unsigned int cfq_slice_async_penalty;struct list_head cic_list; };
@@ -288,6 +295,25 @@ static inline int cfq_slice_used(struct cfq_queue
*cfqq)
}
/*
+ * Get the slice time to be decreased
+ * when async queue is dispatched and
+ * there are unfinished requests in sync queue
+ */
+static inline int
+cfq_async_penalty(struct cfq_data *cfqd, struct cfq_queue *cfqq) +{
+ int penalty;
+
+ if (cfq_cfqq_sync(cfqq))
+ return 0;
+
+ penalty = (cfqq->slice_end - jiffies) *
+ cfqd->cfq_slice_async_penalty / 100;
+
+ return penalty;
+}
+
+/*
* Lifted from AS - choose which of rq1 and rq2 that is best served
now.
* We choose the request that is closest to the head right now.
Distance
* behind the head is penalized and only allowed to a certain extent.
@@ -1103,7 +1129,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force)
}
if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
- break;
+ cfqq->slice_end -= cfq_async_penalty(cfqd, cfqq);
cfq_clear_cfqq_must_dispatch(cfqq); cfq_clear_cfqq_wait_request(cfqq);
@@ -2143,6 +2169,7 @@ static void *cfq_init_queue(struct request_queue *q)
cfqd->cfq_slice[1] = cfq_slice_sync; cfqd->cfq_slice_async_rq = cfq_slice_async_rq; cfqd->cfq_slice_idle = cfq_slice_idle;
+ cfqd->cfq_slice_async_penalty = cfq_slice_async_penalty;
return cfqd;
}
@@ -2211,6 +2238,7 @@ SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
+SHOW_FUNCTION(cfq_slice_async_penalty_show, cfqd->cfq_slice_async_penalty, 0);
#undef SHOW_FUNCTION
#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
@@ -2242,6 +2270,8 @@ STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX,
1);
STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
UINT_MAX, 0);
+STORE_FUNCTION(cfq_slice_async_penalty_store, &cfqd->cfq_slice_async_penalty, 0,
+ 100, 0);
#undef STORE_FUNCTION
#define CFQ_ATTR(name) \
@@ -2257,6 +2287,7 @@ static struct elv_fs_entry cfq_attrs[] = {
CFQ_ATTR(slice_async), CFQ_ATTR(slice_async_rq), CFQ_ATTR(slice_idle),
+ CFQ_ATTR(slice_async_penalty),
__ATTR_NULL
};
Lin Ming
Quoted 172 lines · Jens Axboe
Quoted 129 lines · Lin Ming
Quoted 127 lines · Jens Axboe
Quoted 125 lines · Yanmin
Quoted 11 lines
Quoted 11 lines
-----Original Message-----
From: Jens Axboe [mailto:[email protected]]
Sent: Tuesday, June 17, 2008 3:30 AM
To: Lin, Ming M
Cc: Zhang, Yanmin; Linux Kernel Mailing List
Subject: Re: FIO: kjournald blocked for more than 120 secondsQuoted 3 lines · Lin Ming
Hi, Jens
When runnig FIO benchmark, kjournald blocked for more than 120
seconds.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Detailed root cause analysis and proposed solutions as below.
Any comment is appreciated.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Hardware Environment
---------------------
13 SEAGATE ST373307FC disks in a JBOD, connected by a Qlogic ISP2312 Fibe Channel HBA.Bug description
----------------
fio vsync random read 4K in 13 disks, 4 processes per disk, fioglobal
Quoted 16 lines
Quoted 16 lines
Quoted 16 lines
paramter as below,
Tested 4 IO schedulers, issue is only seen in CFQ.INFO: task kjournald:20558 blocked for more than 120 seconds."echo 0 >
/proc/sys/kernel/hung_task_timeout_secs" disables this message.
kjournald Dffff8100108209786712 20558 2ffff81022ddb1d100000000000000046ffff81022e7baa10ffffffff803ba6f2ffff81022ecd0000ffff8101e6dc9160ffff81022ecd0348000000008048b6cb0000000000000086ffff81022c4e8d300000000000000000ffffffff80247537Call Trace: [<ffffffff803ba6f2>] kobject_get+0x12/0x17The disks of my testing machine are tagged devices, so the CFQ idle window is disabled. In other words, the active queue of tagged
devices(cfqd->hw_tag=1)never idle for a new request.Quoted 1 line
Quoted 1 line
Quoted 1 line
This causes active queue be expired immediately if it's empty,
although
Quoted 1 line
Quoted 1 line
Quoted 1 line
it has not run out of time. CFQ will select next queue as active
queue.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
In this testcase, there are thousands of FIO read requests in sync queues, only a few write requests by
journal_write_commit_recordin async queues.In the other hand, all processes use the default io class and
priority.
Quoted 1 line
Quoted 1 line
Quoted 1 line
They share the async queue for the same device, but have their own
sync
Quoted 1 line
Quoted 1 line
Quoted 1 line
queue, so the sync queue number is 4 while asyn queue number is just
1
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
for the same device.
So sync queue has much more chances be selected as new active queue
than
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
async queue.
Sync queues do not idle and they are dispatched all the time. This
leads
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
to many unfinished requests in external queue,
namely,cfqd->sync_flight> 0.static int cfq_dispatch_requests (...) { .... while ((cfqq = cfq_select_queue(cfqd)) != NULL) { .... if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) break; .... __cfq_dispatch_requests(cfqq) } .... }Quoted 1 line
Quoted 1 line
Quoted 1 line
When
cfq_select_queueselects the async queue which includeskjournald's
Quoted 1 line
Quoted 1 line
Quoted 1 line
write request, this selected async queue will never be dispatched
since
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
cfqd->sync_flight> 0, so kjournald is blocked.Proposed 3 solutions
------------------
1. Do not checkcfqd->sync_flight(quoted diff) +0
-2- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) - break;2. If we do need to check
cfqd->sync_flight, then for taggeddevices, we
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
should give a little more chances to async queue to be dispatched.
(quoted diff) +0
-0@@ -1102,7 +1102,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force) break; }Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
(quoted diff) +1
-1- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) + if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq) && ! cfqd->hw_tag) break;3. Force write request issued by
journal_write_commit_recordas sync request. As a matter of fact, it looks like most write requests submitted by kjournald is async request. We need convert them tosync
Quoted 9 lines
Quoted 9 lines
Quoted 1 line
requests.
Thanks for the very detailed analysis of the problem, complete with suggestions. While I think that any code that does:
submit async io
wait for itshould be issuing sync IO (or, better, automatically upgrade the
request
Quoted 1 line
Quoted 1 line
from async -> sync), we cannot rely on that.
[YM] We can talk case by case. We could convert some important async io codes
to sync io codes at least. For example, kjournald callssync_dirty_bufferwhat
we captured in this case.Quoted 6 lines
Quoted 4 lines
I agree, we should fix the obvious cases. My point was merely that there will probably always be missed cases, so we should attempt to handle it in the scheduler as well. Does the below buffer patch make it any better?
Yes, kjournald blocked issue is gone with below patch applied.
I think it's obviously the right thing to do, but I'm also a bit worried about applying it so close to 2.6.26 release. OTOH, we need to do SOMETHING for 2.6.26 release, so...
Quoted 30 lines
Lin Ming
Quoted 26 lines
Quoted 8 lines
Another case is writeback. If processes do mmapped I/O and they might stop in
page fault to wait writeback finishing. Or a buffer write might trigger a dirty
page balance. As the latest kernel is more aggressive to start writeback, it might
be an issue now.Sync process getting stuck in async writeout is another problem of the same variety.
fs/buffer.c +1
-1@@ -2978,7 +2978,7 @@ int sync_dirty_buffer(struct buffer_head *bh)if (test_clear_buffer_dirty (bh)) {get_bh (bh); bh->b_end_io = end_buffer_write_sync; - ret =submit_bh (WRITE , bh); + ret =submit_bh (WRITE_SYNC , bh);wait_on_buffer (bh);if (buffer_eopnotsupp (bh)) {clear_buffer_eopnotsupp (bh);
Quoted 10 lines · Lin Ming
Jens,
How about below patch?
If
cfqd->sync_flight> 0, we still allow async queue to be dispached, and its slice time would be decreased. So sync queue never starve async queue.Introduce a sysfs interface
cfq_slice_async_penaltyto adjust the percentages of slice that will be decreased.
I don't think it's a good change, for a variety of reasons:
- The reason we don't dispatch async with sync in flight, is to avoid interference of the async IO with the sync IO. Perhaps that is OK in this case, since elv_dispatch_sort() by default doesn't allow crossing read vs write boundaries.
- async queue is typically expired immediately, so the slice length doesn't factor into it so much.
- I don't think more tunables are going to help solve anything.
With the current check, we risk starving async IO in presence of lots of sync IO. Without the check, we risk starving sync process in presence of lots of async IO. So we need some sort of middle ground. Async IO is penalized in the presence of lots of sync IO already. As the number of sync queues grow, async IO gets a continually smaller part of the total disk time. That isn't very fair to begin with.
For 2.6.26, how about adding the WRITE_SYNC change to buffer.c that we discussed earlier? I've been using it here since and it's a less risky change. Then for 2.6.27 we can look into tweaking CFQ.
Quoted 284 lines
Signed-off-by Lin Ming <[email protected]>--- block/cfq-iosched.c | 33 ++++++++++++++++++++++++++++++++- 1 files changed, 32 insertions(+), 1 deletions(-)block/cfq-iosched.c +26
-0@@ -26,6 +26,12 @@ static const int cfq_slice_sync = HZ / 10;static int cfq_slice_async =HZ /25 ;static const int cfq_slice_async_rq =2 ;static int cfq_slice_idle =HZ /125 ; +/* +* penalty of dispatching async queue +* if there are unfinished requests in sync queue +* by default, 50% of time slice decreased +*/ +static int cfq_slice_async_penalty =50 ;/* * offset from end of service tree @@ -114,6 +120,7 @@ struct cfq_data {unsigned int cfq_slice[2 ];unsigned int cfq_slice_async_rq;unsigned int cfq_slice_idle; +unsigned int cfq_slice_async_penalty;struct list_head cic_list; };@@ -288,6 +295,25 @@ static inline int cfq_slice_used(struct cfq_queue *cfqq) }/* +* Get the slice time to be decreased +* when async queue is dispatched and +* there are unfinished requests in sync queue +*/ +static inline int +cfq_async_penalty (struct cfq_data *cfqd,struct cfq_queue *cfqq) +{ +int penalty; + +if (cfq_cfqq_sync (cfqq)) +return 0 ; + + penalty = (cfqq->slice_end - jiffies) * + cfqd->cfq_slice_async_penalty /100 ; + +return penalty; +} + +/* * Lifted from AS - choose which of rq1 and rq2 that is best served now.
* We choose the request that is closest to the head right now.
Distance
* behind the head is penalized and only allowed to a certain extent.block/cfq-iosched.c +1
-1@@ -1103,7 +1129,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force) }if (cfqd->sync_flight && !cfq_cfqq_sync (cfqq)) -break ; + cfqq->slice_end -=cfq_async_penalty (cfqd, cfqq);cfq_clear_cfqq_must_dispatch (cfqq);block/cfq-iosched.c +1
-0cfq_clear_cfqq_wait_request (cfqq);@@ -2143,6 +2169,7 @@ static void *cfq_init_queue(struct request_queue *q) cfqd->cfq_slice[1 ] = cfq_slice_sync; cfqd->cfq_slice_async_rq = cfq_slice_async_rq; cfqd->cfq_slice_idle = cfq_slice_idle; + cfqd->cfq_slice_async_penalty = cfq_slice_async_penalty;return cfqd;}
block/cfq-iosched.c +1
-0@@ -2211,6 +2238,7 @@ SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);SHOW_FUNCTION (cfq_slice_sync_show, cfqd->cfq_slice[1 ],1 );SHOW_FUNCTION (cfq_slice_async_show, cfqd->cfq_slice[0 ],1 );SHOW_FUNCTION (cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq,0 ); +SHOW_FUNCTION (cfq_slice_async_penalty_show, cfqd->cfq_slice_async_penalty,0 );#undef SHOW_FUNCTION#define
STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)\block/cfq-iosched.c +2
-0@@ -2242,6 +2270,8 @@ STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);STORE_FUNCTION (cfq_slice_async_store, &cfqd->cfq_slice[0 ],1 ,UINT_MAX ,1 );STORE_FUNCTION (cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq,1 ,UINT_MAX ,0 ); +STORE_FUNCTION (cfq_slice_async_penalty_store, &cfqd->cfq_slice_async_penalty,0 , +100 ,0 );#undef
STORE_FUNCTION#define
CFQ_ATTR(name)\block/cfq-iosched.c +1
-0@@ -2257,6 +2287,7 @@ static struct elv_fs_entry cfq_attrs[] = {CFQ_ATTR (slice_async),CFQ_ATTR (slice_async_rq),CFQ_ATTR (slice_idle), +CFQ_ATTR (slice_async_penalty), __ATTR_NULL };Lin Ming
Quoted 172 lines · Jens Axboe
Quoted 129 lines · Lin Ming
Quoted 127 lines · Jens Axboe
Quoted 125 lines · Yanmin
Quoted 11 lines
Quoted 11 lines
-----Original Message-----
From: Jens Axboe [mailto:[email protected]]
Sent: Tuesday, June 17, 2008 3:30 AM
To: Lin, Ming M
Cc: Zhang, Yanmin; Linux Kernel Mailing List
Subject: Re: FIO: kjournald blocked for more than 120 secondsQuoted 3 lines · Lin Ming
Hi, Jens
When runnig FIO benchmark, kjournald blocked for more than 120
seconds.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Detailed root cause analysis and proposed solutions as below.
Any comment is appreciated.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Hardware Environment
---------------------
13 SEAGATE ST373307FC disks in a JBOD, connected by a Qlogic ISP2312 Fibe Channel HBA.Bug description
----------------
fio vsync random read 4K in 13 disks, 4 processes per disk, fioglobal
Quoted 16 lines
Quoted 16 lines
Quoted 16 lines
paramter as below,
Tested 4 IO schedulers, issue is only seen in CFQ.INFO: task kjournald:20558 blocked for more than 120 seconds."echo 0 >
/proc/sys/kernel/hung_task_timeout_secs" disables this message.
kjournald Dffff8100108209786712 20558 2ffff81022ddb1d100000000000000046ffff81022e7baa10ffffffff803ba6f2ffff81022ecd0000ffff8101e6dc9160ffff81022ecd0348000000008048b6cb0000000000000086ffff81022c4e8d300000000000000000ffffffff80247537Call Trace: [<ffffffff803ba6f2>] kobject_get+0x12/0x17The disks of my testing machine are tagged devices, so the CFQ idle window is disabled. In other words, the active queue of tagged
devices(cfqd->hw_tag=1)never idle for a new request.Quoted 1 line
Quoted 1 line
Quoted 1 line
This causes active queue be expired immediately if it's empty,
although
Quoted 1 line
Quoted 1 line
Quoted 1 line
it has not run out of time. CFQ will select next queue as active
queue.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
In this testcase, there are thousands of FIO read requests in sync queues, only a few write requests by
journal_write_commit_recordin async queues.In the other hand, all processes use the default io class and
priority.
Quoted 1 line
Quoted 1 line
Quoted 1 line
They share the async queue for the same device, but have their own
sync
Quoted 1 line
Quoted 1 line
Quoted 1 line
queue, so the sync queue number is 4 while asyn queue number is just
1
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
for the same device.
So sync queue has much more chances be selected as new active queue
than
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
async queue.
Sync queues do not idle and they are dispatched all the time. This
leads
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
to many unfinished requests in external queue,
namely,cfqd->sync_flight> 0.static int cfq_dispatch_requests (...) { .... while ((cfqq = cfq_select_queue(cfqd)) != NULL) { .... if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) break; .... __cfq_dispatch_requests(cfqq) } .... }Quoted 1 line
Quoted 1 line
Quoted 1 line
When
cfq_select_queueselects the async queue which includeskjournald's
Quoted 1 line
Quoted 1 line
Quoted 1 line
write request, this selected async queue will never be dispatched
since
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
cfqd->sync_flight> 0, so kjournald is blocked.Proposed 3 solutions
------------------
1. Do not checkcfqd->sync_flight(quoted diff) +0
-2- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) - break;2. If we do need to check
cfqd->sync_flight, then for taggeddevices, we
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
should give a little more chances to async queue to be dispatched.
(quoted diff) +0
-0@@ -1102,7 +1102,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force) break; }Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
(quoted diff) +1
-1- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) + if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq) && ! cfqd->hw_tag) break;3. Force write request issued by
journal_write_commit_recordas sync request. As a matter of fact, it looks like most write requests submitted by kjournald is async request. We need convert them tosync
Quoted 9 lines
Quoted 9 lines
Quoted 1 line
requests.
Thanks for the very detailed analysis of the problem, complete with suggestions. While I think that any code that does:
submit async io
wait for itshould be issuing sync IO (or, better, automatically upgrade the
request
Quoted 1 line
Quoted 1 line
from async -> sync), we cannot rely on that.
[YM] We can talk case by case. We could convert some important async io codes
to sync io codes at least. For example, kjournald callssync_dirty_bufferwhat
we captured in this case.Quoted 6 lines
Quoted 4 lines
I agree, we should fix the obvious cases. My point was merely that there will probably always be missed cases, so we should attempt to handle it in the scheduler as well. Does the below buffer patch make it any better?
Yes, kjournald blocked issue is gone with below patch applied.
I think it's obviously the right thing to do, but I'm also a bit worried about applying it so close to 2.6.26 release. OTOH, we need to do SOMETHING for 2.6.26 release, so...
Quoted 30 lines
Lin Ming
Quoted 26 lines
Quoted 8 lines
Another case is writeback. If processes do mmapped I/O and they might stop in
page fault to wait writeback finishing. Or a buffer write might trigger a dirty
page balance. As the latest kernel is more aggressive to start writeback, it might
be an issue now.Sync process getting stuck in async writeout is another problem of the same variety.
fs/buffer.c +1
-1@@ -2978,7 +2978,7 @@ int sync_dirty_buffer(struct buffer_head *bh)if (test_clear_buffer_dirty (bh)) {get_bh (bh); bh->b_end_io = end_buffer_write_sync; - ret =submit_bh (WRITE , bh); + ret =submit_bh (WRITE_SYNC , bh);wait_on_buffer (bh);if (buffer_eopnotsupp (bh)) {clear_buffer_eopnotsupp (bh);
Signature · 1 line
Jens Axboe
Quoted 31 lines · Jens Axboe
Quoted 10 lines · Lin Ming
Jens,
How about below patch?
If
cfqd->sync_flight> 0, we still allow async queue to be dispached, and its slice time would be decreased. So sync queue never starve async queue.Introduce a sysfs interface
cfq_slice_async_penaltyto adjust the percentages of slice that will be decreased.I don't think it's a good change, for a variety of reasons:
- The reason we don't dispatch async with sync in flight, is to avoid interference of the async IO with the sync IO. Perhaps that is OK in this case, since
elv_dispatch_sort()by default doesn't allow crossing read vs write boundaries.- async queue is typically expired immediately, so the slice length doesn't factor into it so much.
- I don't think more tunables are going to help solve anything.
With the current check, we risk starving async IO in presence of lots of sync IO. Without the check, we risk starving sync process in presence of lots of async IO. So we need some sort of middle ground. Async IO is penalized in the presence of lots of sync IO already. As the number of sync queues grow, async IO gets a continually smaller part of the total disk time. That isn't very fair to begin with.
Quoted 3 lines
For 2.6.26, how about adding the
WRITE_SYNCchange tobuffer.cthat we discussed earlier? I've been using it here since and it's a less risky change. Then for 2.6.27 we can look into tweaking CFQ.
Sounds good.
OK.
Lin Ming
Quoted 288 lines
Quoted 287 lines
Signed-off-by Lin Ming <[email protected]>--- block/cfq-iosched.c | 33 ++++++++++++++++++++++++++++++++- 1 files changed, 32 insertions(+), 1 deletions(-)block/cfq-iosched.c +26
-0@@ -26,6 +26,12 @@ static const int cfq_slice_sync = HZ / 10;static int cfq_slice_async =HZ /25 ;static const int cfq_slice_async_rq =2 ;static int cfq_slice_idle =HZ /125 ; +/* +* penalty of dispatching async queue +* if there are unfinished requests in sync queue +* by default, 50% of time slice decreased +*/ +static int cfq_slice_async_penalty =50 ;/* * offset from end of service tree @@ -114,6 +120,7 @@ struct cfq_data {unsigned int cfq_slice[2 ];unsigned int cfq_slice_async_rq;unsigned int cfq_slice_idle; +unsigned int cfq_slice_async_penalty;struct list_head cic_list; };@@ -288,6 +295,25 @@ static inline int cfq_slice_used(struct cfq_queue *cfqq) }/* +* Get the slice time to be decreased +* when async queue is dispatched and +* there are unfinished requests in sync queue +*/ +static inline int +cfq_async_penalty (struct cfq_data *cfqd,struct cfq_queue *cfqq) +{ +int penalty; + +if (cfq_cfqq_sync (cfqq)) +return 0 ; + + penalty = (cfqq->slice_end - jiffies) * + cfqd->cfq_slice_async_penalty /100 ; + +return penalty; +} + +/* * Lifted from AS - choose which of rq1 and rq2 that is best served now.
* We choose the request that is closest to the head right now.
Distance
* behind the head is penalized and only allowed to a certain extent.block/cfq-iosched.c +1
-1@@ -1103,7 +1129,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force) }if (cfqd->sync_flight && !cfq_cfqq_sync (cfqq)) -break ; + cfqq->slice_end -=cfq_async_penalty (cfqd, cfqq);cfq_clear_cfqq_must_dispatch (cfqq);block/cfq-iosched.c +1
-0cfq_clear_cfqq_wait_request (cfqq);@@ -2143,6 +2169,7 @@ static void *cfq_init_queue(struct request_queue *q) cfqd->cfq_slice[1 ] = cfq_slice_sync; cfqd->cfq_slice_async_rq = cfq_slice_async_rq; cfqd->cfq_slice_idle = cfq_slice_idle; + cfqd->cfq_slice_async_penalty = cfq_slice_async_penalty;return cfqd;}
block/cfq-iosched.c +1
-0@@ -2211,6 +2238,7 @@ SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);SHOW_FUNCTION (cfq_slice_sync_show, cfqd->cfq_slice[1 ],1 );SHOW_FUNCTION (cfq_slice_async_show, cfqd->cfq_slice[0 ],1 );SHOW_FUNCTION (cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq,0 ); +SHOW_FUNCTION (cfq_slice_async_penalty_show, cfqd->cfq_slice_async_penalty,0 );#undef SHOW_FUNCTION#define
STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)\block/cfq-iosched.c +2
-0@@ -2242,6 +2270,8 @@ STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);STORE_FUNCTION (cfq_slice_async_store, &cfqd->cfq_slice[0 ],1 ,UINT_MAX ,1 );STORE_FUNCTION (cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq,1 ,UINT_MAX ,0 ); +STORE_FUNCTION (cfq_slice_async_penalty_store, &cfqd->cfq_slice_async_penalty,0 , +100 ,0 );#undef
STORE_FUNCTION#define
CFQ_ATTR(name)\block/cfq-iosched.c +1
-0@@ -2257,6 +2287,7 @@ static struct elv_fs_entry cfq_attrs[] = {CFQ_ATTR (slice_async),CFQ_ATTR (slice_async_rq),CFQ_ATTR (slice_idle), +CFQ_ATTR (slice_async_penalty), __ATTR_NULL };Lin Ming
Quoted 172 lines · Jens Axboe
Quoted 129 lines · Lin Ming
Quoted 127 lines · Jens Axboe
Quoted 125 lines · Yanmin
Quoted 11 lines
Quoted 11 lines
-----Original Message-----
From: Jens Axboe [mailto:[email protected]]
Sent: Tuesday, June 17, 2008 3:30 AM
To: Lin, Ming M
Cc: Zhang, Yanmin; Linux Kernel Mailing List
Subject: Re: FIO: kjournald blocked for more than 120 secondsQuoted 3 lines · Lin Ming
Hi, Jens
When runnig FIO benchmark, kjournald blocked for more than 120
seconds.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Detailed root cause analysis and proposed solutions as below.
Any comment is appreciated.
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Hardware Environment
---------------------
13 SEAGATE ST373307FC disks in a JBOD, connected by a Qlogic ISP2312 Fibe Channel HBA.Bug description
----------------
fio vsync random read 4K in 13 disks, 4 processes per disk, fioglobal
Quoted 16 lines
Quoted 16 lines
Quoted 16 lines
paramter as below,
Tested 4 IO schedulers, issue is only seen in CFQ.INFO: task kjournald:20558 blocked for more than 120 seconds."echo 0 >
/proc/sys/kernel/hung_task_timeout_secs" disables this message.
kjournald Dffff8100108209786712 20558 2ffff81022ddb1d100000000000000046ffff81022e7baa10ffffffff803ba6f2ffff81022ecd0000ffff8101e6dc9160ffff81022ecd0348000000008048b6cb0000000000000086ffff81022c4e8d300000000000000000ffffffff80247537Call Trace: [<ffffffff803ba6f2>] kobject_get+0x12/0x17The disks of my testing machine are tagged devices, so the CFQ idle window is disabled. In other words, the active queue of tagged
devices(cfqd->hw_tag=1)never idle for a new request.Quoted 1 line
Quoted 1 line
Quoted 1 line
This causes active queue be expired immediately if it's empty,
although
Quoted 1 line
Quoted 1 line
Quoted 1 line
it has not run out of time. CFQ will select next queue as active
queue.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
In this testcase, there are thousands of FIO read requests in sync queues, only a few write requests by
journal_write_commit_recordin async queues.In the other hand, all processes use the default io class and
priority.
Quoted 1 line
Quoted 1 line
Quoted 1 line
They share the async queue for the same device, but have their own
sync
Quoted 1 line
Quoted 1 line
Quoted 1 line
queue, so the sync queue number is 4 while asyn queue number is just
1
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
for the same device.
So sync queue has much more chances be selected as new active queue
than
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
async queue.
Sync queues do not idle and they are dispatched all the time. This
leads
Quoted 15 lines
Quoted 15 lines
Quoted 15 lines
to many unfinished requests in external queue,
namely,cfqd->sync_flight> 0.static int cfq_dispatch_requests (...) { .... while ((cfqq = cfq_select_queue(cfqd)) != NULL) { .... if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) break; .... __cfq_dispatch_requests(cfqq) } .... }Quoted 1 line
Quoted 1 line
Quoted 1 line
When
cfq_select_queueselects the async queue which includeskjournald's
Quoted 1 line
Quoted 1 line
Quoted 1 line
write request, this selected async queue will never be dispatched
since
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
cfqd->sync_flight> 0, so kjournald is blocked.Proposed 3 solutions
------------------
1. Do not checkcfqd->sync_flight(quoted diff) +0
-2- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) - break;2. If we do need to check
cfqd->sync_flight, then for taggeddevices, we
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
should give a little more chances to async queue to be dispatched.
(quoted diff) +0
-0@@ -1102,7 +1102,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force) break; }Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
(quoted diff) +1
-1- if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) + if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq) && ! cfqd->hw_tag) break;3. Force write request issued by
journal_write_commit_recordas sync request. As a matter of fact, it looks like most write requests submitted by kjournald is async request. We need convert them tosync
Quoted 9 lines
Quoted 9 lines
Quoted 1 line
requests.
Thanks for the very detailed analysis of the problem, complete with suggestions. While I think that any code that does:
submit async io
wait for itshould be issuing sync IO (or, better, automatically upgrade the
request
Quoted 1 line
Quoted 1 line
from async -> sync), we cannot rely on that.
[YM] We can talk case by case. We could convert some important async io codes
to sync io codes at least. For example, kjournald callssync_dirty_bufferwhat
we captured in this case.Quoted 6 lines
Quoted 4 lines
I agree, we should fix the obvious cases. My point was merely that there will probably always be missed cases, so we should attempt to handle it in the scheduler as well. Does the below buffer patch make it any better?
Yes, kjournald blocked issue is gone with below patch applied.
I think it's obviously the right thing to do, but I'm also a bit worried about applying it so close to 2.6.26 release. OTOH, we need to do SOMETHING for 2.6.26 release, so...
Quoted 30 lines
Lin Ming
Quoted 26 lines
Quoted 8 lines
Another case is writeback. If processes do mmapped I/O and they might stop in
page fault to wait writeback finishing. Or a buffer write might trigger a dirty
page balance. As the latest kernel is more aggressive to start writeback, it might
be an issue now.Sync process getting stuck in async writeout is another problem of the same variety.
fs/buffer.c +1
-1@@ -2978,7 +2978,7 @@ int sync_dirty_buffer(struct buffer_head *bh)if (test_clear_buffer_dirty (bh)) {get_bh (bh); bh->b_end_io = end_buffer_write_sync; - ret =submit_bh (WRITE , bh); + ret =submit_bh (WRITE_SYNC , bh);wait_on_buffer (bh);if (buffer_eopnotsupp (bh)) {clear_buffer_eopnotsupp (bh);