When we get a notification that cpu topology changed, we schedule a
work struct which just calls arch_reinit_sched_domains. This function
in turn calls get_online_cpus() which results int the lockdep warning
below.
After all it turnded out that it's not legal to call get_online_cpus()
from the context of a multi-threaded work queue.
It could deadlock this way:
process 0 (events/cpu-x):
-> run_workqueue
-> removes my work_struct from the work queue
-> calls work_struct->fn
-> get_online_cpus()
-> locks on cpu_hotplug.lock since process 1 below is doing cpu hotplug
process 1:
-> cpu_down (for cpu-x)
-> cpu_hotplug_begin (holds cpu_hotplug.lock now)
-> cpu-x dead
-> notifier_call_chain with CPU_DEAD
-> cleanup_workqueue_thread
-> flush_cpu_workqueue (succeeds)
-> kthread_stop for events/cpu-x
-> now kthread_stop waits for my work_struct to complete from within
process 0. -> dead.
A single threaded workqueue wouldn't have such problems, however there is
no such common queue available and it's not worth to create one for the
very rare calls to arch_reinit_sched_domains.
So we just create a kernel thread from our work struct which calls
arch_reinit_sched_domains and are done with it.
Thanks to Oleg Nesterov and Peter Zijlstra for helping me figuring out
that this isn't a false positive lockdep warning:
=======================================================
[ INFO: possible circular locking dependency detected ]
2.6.25-03562-g3dc5063-dirty #12
-------------------------------------------------------
events/3/14 is trying to acquire lock:
(&cpu_hotplug.lock){--..}, at: [<0000000000076094>] get_online_cpus+0x50/0x78
but task is already holding lock:
(topology_work){--..}, at: [<0000000000059cde>] run_workqueue+0x106/0x278
which lock already depends on the new lock.
the existing dependency chain (in reverse order) is:
-> #2 (topology_work){--..}:
[<000000000006fc74>] __lock_acquire+0x1010/0x111c
[<000000000006fe40>] lock_acquire+0xc0/0xf8
[<0000000000059d48>] run_workqueue+0x170/0x278
[<0000000000059edc>] worker_thread+0x8c/0xf0
[<000000000005f5bc>] kthread+0x68/0xa0
[<000000000001a33e>] kernel_thread_starter+0x6/0xc
[<000000000001a338>] kernel_thread_starter+0x0/0xc
-> #1 (events){--..}:
[<000000000006fc74>] __lock_acquire+0x1010/0x111c
[<000000000006fe40>] lock_acquire+0xc0/0xf8
[<000000000005a23c>] cleanup_workqueue_thread+0x60/0xa8
[<00000000003b2ab8>] workqueue_cpu_callback+0xbc/0x170
[<00000000003bba80>] notifier_call_chain+0x5c/0xa4
[<00000000000655a2>] __raw_notifier_call_chain+0x26/0x38
[<00000000000655e2>] raw_notifier_call_chain+0x2e/0x40
[<0000000000075e00>] cpu_down+0x228/0x31c
[<00000000003b1dd8>] store_online+0x64/0xb8
[<00000000001e7128>] sysdev_store+0x48/0x58
[<0000000000121cd2>] sysfs_write_file+0x126/0x1c0
[<00000000000c1944>] vfs_write+0xb0/0x15c
[<00000000000c20e6>] sys_write+0x56/0x88
[<0000000000027a68>] sys32_write+0x34/0x4c
[<0000000000023f70>] sysc_noemu+0x10/0x16
[<0000000077f3f186>] 0x77f3f186
-> #0 (&cpu_hotplug.lock){--..}:
[<000000000006fa84>] __lock_acquire+0xe20/0x111c
[<000000000006fe40>] lock_acquire+0xc0/0xf8
[<00000000003b701c>] mutex_lock_nested+0xd0/0x364
[<0000000000076094>] get_online_cpus+0x50/0x78
[<000000000003a03e>] arch_reinit_sched_domains+0x26/0x58
[<000000000002700e>] topology_work_fn+0x26/0x34
[<0000000000059d4e>] run_workqueue+0x176/0x278
[<0000000000059edc>] worker_thread+0x8c/0xf0
[<000000000005f5bc>] kthread+0x68/0xa0
[<000000000001a33e>] kernel_thread_starter+0x6/0xc
[<000000000001a338>] kernel_thread_starter+0x0/0xc
other info that might help us debug this:
2 locks held by events/3/14:
#0: (events){--..}, at: [<0000000000059cde>] run_workqueue+0x106/0x278
#1: (topology_work){--..}, at: [<0000000000059cde>] run_workqueue+0x106/0x278
stack backtrace:
CPU: 3 Not tainted 2.6.25-03562-g3dc5063-dirty #12
Process events/3 (pid: 14, task: 000000002fb04038, ksp: 000000002fb0bd70)
0400000000000000 000000002fb0ba40 0000000000000002 0000000000000000
000000002fb0bae0 000000002fb0ba58 000000002fb0ba58 0000000000016488
0000000000000000 000000002fb0bd70 0000000000000000 0000000000000000
000000002fb0ba40 000000000000000c 000000002fb0ba40 000000002fb0bab0
00000000003c99e0 0000000000016488 000000002fb0ba40 000000002fb0ba90
Call Trace:
([<00000000000163fc>] show_trace+0x138/0x158)
[<00000000000164e2>] show_stack+0xc6/0xf8
[<0000000000016624>] dump_stack+0xb0/0xc0
[<000000000006cd36>] print_circular_bug_tail+0xa2/0xb4
[<000000000006fa84>] __lock_acquire+0xe20/0x111c
[<000000000006fe40>] lock_acquire+0xc0/0xf8
[<00000000003b701c>] mutex_lock_nested+0xd0/0x364
[<0000000000076094>] get_online_cpus+0x50/0x78
[<000000000003a03e>] arch_reinit_sched_domains+0x26/0x58
[<000000000002700e>] topology_work_fn+0x26/0x34
[<0000000000059d4e>] run_workqueue+0x176/0x278
[<0000000000059edc>] worker_thread+0x8c/0xf0
[<000000000005f5bc>] kthread+0x68/0xa0
[<000000000001a33e>] kernel_thread_starter+0x6/0xc
[<000000000001a338>] kernel_thread_starter+0x0/0xc
INFO: lockdep is turned off.
Cc: Oleg Nesterov <oleg@tv-sign.ru>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
If vertical cpu polarization is active then the hypervisor will
dispatch certain cpus for a longer time than other cpus for maximum
performance. For example if a guest would have three virtual cpus,
each of them with a share of 33 percent, then in case of vertical
cpu polarization all of the processing time would be combined to a
single cpu which would run all the time, while the other two cpus
would get nearly no cpu time.
There are three different types of vertical cpus: high, medium and
low. Low cpus hardly get any real cpu time, while high cpus get a
full real cpu. Medium cpus get something in between.
In order to switch between the two possible modes (default is
horizontal) a 0 for horizontal polarization or a 1 for vertical
polarization must be written to the dispatching sysfs attribute:
/sys/devices/system/cpu/dispatching
The polarization of each single cpu can be figured out by the
polarization sysfs attribute of each cpu:
/sys/devices/system/cpu/cpuX/polarization
horizontal, vertical:high, vertical:medium, vertical:low or unknown.
When switching polarization the polarization attribute may contain
the value unknown until the configuration change is done and the
kernel has figured out the new polarization of each cpu.
Note that running a system with different types of vertical cpus may
result in significant performance regressions. If possible only one
type of vertical cpus should be used. All other cpus should be
offlined.
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Add s390 backend so we can give the scheduler some hints about the
cpu topology.
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Heiko Carstens