cpuset: adjust the lock sequence when rebuilding the sched domains.
Subsystems (2)
PATCH 1/1 cpuset: adjust the lock sequence when rebuilding the sched domains.
From: Vincent Wang <[email protected]>
A deadlock issue is found when executing a cpu hotplug stress test on android phones with cpuset and scheduil enabled.
When CPUx is plugged out, the hotplug thread that calls cpu_down() will hold cpu_hotplug_lock and wait the thread cpuhp/x to finish hotplug. If the core is the last one in a cluster, cpuhp/x have to call cpuhp_cpufreq_offline() and the kernel thread sugov need to exit for schedutil governor. The exit of sugov need to hold cgroup_threadgroup_rwsem in exit_signals(). For example:
PID: 150 TASK: ffffffc0b9cad080 CPU: 0 COMMAND: "sprdhotplug"
#0 [ffffff8009fcb9d0] __switch_to at ffffff80080858f0 #1 [ffffff8009fcb9f0] __schedule at ffffff80089f185c #2 [ffffff8009fcba80] schedule at ffffff80089f1b84 #3 [ffffff8009fcbaa0] schedule_timeout at ffffff80089f5124 #4 [ffffff8009fcbb40] wait_for_common at ffffff80089f2944 #5 [ffffff8009fcbbe0] wait_for_completion at ffffff80089f29a4 #6 [ffffff8009fcbc00] __cpuhp_kick_ap at ffffff80080ab030 #7 [ffffff8009fcbc20] cpuhp_kick_ap_work at ffffff80080ab154 #8 [ffffff8009fcbc70] _cpu_down at ffffff80089ee19c #9 [ffffff8009fcbcd0] cpu_down at ffffff80080ac144
PID: 26 TASK: ffffffc0bbe22080 CPU: 3 COMMAND: "cpuhp/3"
#0 [ffffff8009693a30] __switch_to at ffffff80080858f0 #1 [ffffff8009693a50] __schedule at ffffff80089f185c #2 [ffffff8009693ae0] schedule at ffffff80089f1b84
#3 [ffffff8009693b00] schedule_timeout at ffffff80089f5124 #4 [ffffff8009693ba0] wait_for_common at ffffff80089f2944 #5 [ffffff8009693c40] wait_for_completion at ffffff80089f29a4 #6 [ffffff8009693c60] kthread_stop at ffffff80080ccd2c #7 [ffffff8009693c90] sugov_exit at ffffff8008102134 #8 [ffffff8009693cc0] cpufreq_exit_governor at ffffff80086c03bc #9 [ffffff8009693ce0] cpufreq_offline at ffffff80086c0634
PID: 13819 TASK: ffffffc0affb6080 CPU: 0 COMMAND: "sugov:3"
#0 [ffffff800ee73c30] __switch_to at ffffff80080858f0 #1 [ffffff800ee73c50] __schedule at ffffff80089f185c #2 [ffffff800ee73ce0] schedule at ffffff80089f1b84
#3 [ffffff800ee73d00] rwsem_down_read_failed at ffffff80089f49d0 #4 [ffffff800ee73d80] __percpu_down_read at ffffff8008102ebc #5 [ffffff800ee73da0] exit_signals at ffffff80080bbd24 #6 [ffffff800ee73de0] do_exit at ffffff80080ae65c
#7 [ffffff800ee73e60] kthread at ffffff80080cc550
Sometimes cgroup_threadgroup_rwsem is hold by another thread, for example Binder:681_2 on android, it wants to hold cpuset_mutex:
PID: 732 TASK: ffffffc09668b080 CPU: 2 COMMAND: "Binder:681_2"
#0 [ffffff800cb7b8c0] __switch_to at ffffff80080858f0
#1 [ffffff800cb7b8e0] __schedule at ffffff80089f185c
#2 [ffffff800cb7b970] schedule at ffffff80089f1b84
#3 [ffffff800cb7b990] schedule_preempt_disabled at ffffff80089f205c #4 [ffffff800cb7b9a0] __mutex_lock at ffffff80089f3118 #5 [ffffff800cb7ba40] __mutex_lock_slowpath at ffffff80089f3230 #6 [ffffff800cb7ba60] mutex_lock at ffffff80089f3278
#7 [ffffff800cb7ba80] cpuset_can_attach at ffffff8008152f84 #8 [ffffff800cb7bae0] cgroup_migrate_execute at ffffff800814ada8
On android, a thread kworker/3:0 will hold cpuset_mutex in rebuild_sched_domains() and want to hold cpu_hotplug_lock which is already hold by the hotplug thread.
PID: 4847 TASK: ffffffc031a6a080 CPU: 3 COMMAND: "kworker/3:0"
#0 [ffffff8016fd3ad0] __switch_to at ffffff80080858f0
#1 [ffffff8016fd3af0] __schedule at ffffff80089f185c
#2 [ffffff8016fd3b80] schedule at ffffff80089f1b84
#3 [ffffff8016fd3ba0] rwsem_down_read_failed at ffffff80089f49d0 #4 [ffffff8016fd3c20] __percpu_down_read at ffffff8008102ebc #5 [ffffff8016fd3c40] cpus_read_lock at ffffff80080aa59c
#6 [ffffff8016fd3c50] rebuild_sched_domains_locked at ffffff80081522a8
In order to fix the deadlock, this patch will adjust the lock sequence when rebuilding sched domains. After stress tests, it works well.
--- kernel/cgroup/cpuset.c | 9 ++++----- 1 file changed, 4 insertions(+), 5 deletions(-)
kernel/cgroup/cpuset.c +4 -5
@@ -912,7 +912,6 @@ static void rebuild_sched_domains_locked(void)int ndoms;lockdep_assert_held (&cpuset_mutex); -get_online_cpus ();/* * We have raced with CPU hotplug. Don't do anything to avoid
@@ -921,19 +920,17 @@ static void rebuild_sched_domains_locked(void)*/ if (!top_cpuset.nr_subparts_cpus && !cpumask_equal (top_cpuset.effective_cpus, cpu_active_mask)) - goto out; +return ;if (top_cpuset.nr_subparts_cpus && !cpumask_subset (top_cpuset.effective_cpus, cpu_active_mask)) - goto out; +return ;/* Generate domain masks and attrs */ ndoms =generate_sched_domains (&doms, &attr);/* Have scheduler rebuild the domains */ partition_sched_domains(ndoms, doms, attr); -out: - put_online_cpus(); }#else /* !CONFIG_SMP */ static void rebuild_sched_domains_locked (void )
@@ -943,9 +940,11 @@ static void rebuild_sched_domains_locked(void)void rebuild_sched_domains (void ) { +get_online_cpus ();mutex_lock (&cpuset_mutex);rebuild_sched_domains_locked ();mutex_unlock (&cpuset_mutex); +put_online_cpus (); }/**
Signature · 1 line
2.20.1
Hi,
Quoted 117 lines · Chunyan Zhang
From: Vincent Wang <[email protected]>
A deadlock issue is found when executing a cpu hotplug stress test on android phones with cpuset and scheduil enabled.
When CPUx is plugged out, the hotplug thread that calls
cpu_down()will holdcpu_hotplug_lockand wait the thread cpuhp/x to finish hotplug. If the core is the last one in a cluster, cpuhp/x have to callcpuhp_cpufreq_offline()and the kernel thread sugov need to exit for schedutil governor. The exit of sugov need to holdcgroup_threadgroup_rwseminexit_signals(). For example:PID: 150 TASK:
ffffffc0b9cad080CPU: 0 COMMAND: "sprdhotplug"
#0 [ffffff8009fcb9d0]__switch_toatffffff80080858f0#1 [ffffff8009fcb9f0]__scheduleatffffff80089f185c#2 [ffffff8009fcba80] schedule atffffff80089f1b84#3 [ffffff8009fcbaa0]schedule_timeoutatffffff80089f5124#4 [ffffff8009fcbb40]wait_for_commonatffffff80089f2944#5 [ffffff8009fcbbe0]wait_for_completionatffffff80089f29a4#6 [ffffff8009fcbc00]__cpuhp_kick_apatffffff80080ab030#7 [ffffff8009fcbc20]cpuhp_kick_ap_workatffffff80080ab154#8 [ffffff8009fcbc70]_cpu_downatffffff80089ee19c#9 [ffffff8009fcbcd0]cpu_downatffffff80080ac144PID: 26 TASK:
ffffffc0bbe22080CPU: 3 COMMAND: "cpuhp/3"
#0 [ffffff8009693a30]__switch_toatffffff80080858f0#1 [ffffff8009693a50]__scheduleatffffff80089f185c#2 [ffffff8009693ae0] schedule atffffff80089f1b84
#3 [ffffff8009693b00]schedule_timeoutatffffff80089f5124#4 [ffffff8009693ba0]wait_for_commonatffffff80089f2944#5 [ffffff8009693c40]wait_for_completionatffffff80089f29a4#6 [ffffff8009693c60]kthread_stopatffffff80080ccd2c#7 [ffffff8009693c90]sugov_exitatffffff8008102134#8 [ffffff8009693cc0]cpufreq_exit_governoratffffff80086c03bc#9 [ffffff8009693ce0]cpufreq_offlineatffffff80086c0634PID: 13819 TASK:
ffffffc0affb6080CPU: 0 COMMAND: "sugov:3"
#0 [ffffff800ee73c30]__switch_toatffffff80080858f0#1 [ffffff800ee73c50]__scheduleatffffff80089f185c#2 [ffffff800ee73ce0] schedule atffffff80089f1b84
#3 [ffffff800ee73d00]rwsem_down_read_failedatffffff80089f49d0#4 [ffffff800ee73d80]__percpu_down_readatffffff8008102ebc#5 [ffffff800ee73da0]exit_signalsatffffff80080bbd24#6 [ffffff800ee73de0]do_exitatffffff80080ae65c
#7 [ffffff800ee73e60] kthread atffffff80080cc550Sometimes
cgroup_threadgroup_rwsemis hold by another thread, for example Binder:681_2 on android, it wants to holdcpuset_mutex:PID: 732 TASK:
ffffffc09668b080CPU: 2 COMMAND: "Binder:681_2"
#0 [ffffff800cb7b8c0]__switch_toatffffff80080858f0
#1 [ffffff800cb7b8e0]__scheduleatffffff80089f185c
#2 [ffffff800cb7b970] schedule atffffff80089f1b84
#3 [ffffff800cb7b990]schedule_preempt_disabledatffffff80089f205c#4 [ffffff800cb7b9a0]__mutex_lockatffffff80089f3118#5 [ffffff800cb7ba40]__mutex_lock_slowpathatffffff80089f3230#6 [ffffff800cb7ba60]mutex_lockatffffff80089f3278
#7 [ffffff800cb7ba80]cpuset_can_attachatffffff8008152f84#8 [ffffff800cb7bae0]cgroup_migrate_executeatffffff800814ada8On android, a thread kworker/3:0 will hold
cpuset_mutexinrebuild_sched_domains()and want to holdcpu_hotplug_lockwhich is already hold by the hotplug thread.PID: 4847 TASK:
ffffffc031a6a080CPU: 3 COMMAND: "kworker/3:0"
#0 [ffffff8016fd3ad0]__switch_toatffffff80080858f0
#1 [ffffff8016fd3af0]__scheduleatffffff80089f185c
#2 [ffffff8016fd3b80] schedule atffffff80089f1b84
#3 [ffffff8016fd3ba0]rwsem_down_read_failedatffffff80089f49d0#4 [ffffff8016fd3c20]__percpu_down_readatffffff8008102ebc#5 [ffffff8016fd3c40]cpus_read_lockatffffff80080aa59c
#6 [ffffff8016fd3c50]rebuild_sched_domains_lockedatffffff80081522a8In order to fix the deadlock, this patch will adjust the lock sequence when rebuilding sched domains. After stress tests, it works well.
Signed-off-by Vincent Wang <[email protected]>Signed-off-by Chunyan Zhang <[email protected]>--- kernel/cgroup/cpuset.c | 9 ++++----- 1 file changed, 4 insertions(+), 5 deletions(-)kernel/cgroup/cpuset.c +2
-5@@ -912,7 +912,6 @@ static void rebuild_sched_domains_locked(void)int ndoms;lockdep_assert_held (&cpuset_mutex); -get_online_cpus ();/* * We have raced with CPU hotplug. Don't do anything to avoid @@ -921,19 +920,17 @@ static void rebuild_sched_domains_locked(void)*/ if (!top_cpuset.nr_subparts_cpus && !cpumask_equal (top_cpuset.effective_cpus, cpu_active_mask)) - goto out; +return ;if (top_cpuset.nr_subparts_cpus && !cpumask_subset (top_cpuset.effective_cpus, cpu_active_mask)) - goto out; +return ;/* Generate domain masks and attrs */ ndoms =generate_sched_domains (&doms, &attr);/* Have scheduler rebuild the domains */ partition_sched_domains(ndoms, doms, attr); -out: - put_online_cpus(); }#else /* !CONFIG_SMP */ static void rebuild_sched_domains_locked (void )
kernel/cgroup/cpuset.c static void rebuild_sched_domains_locked(void) · line 948
Quoted 9 lines
kernel/cgroup/cpuset.c +2
-0@@ -943,9 +940,11 @@ static void rebuild_sched_domains_locked(void)void rebuild_sched_domains (void ) { +get_online_cpus ();mutex_lock (&cpuset_mutex);rebuild_sched_domains_locked ();mutex_unlock (&cpuset_mutex); +put_online_cpus (); }
This looks a subset of d74b27d63a8b ("cgroup/cpuset: Change cpuset_rwsem and hotplug lock order") from
git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git sched/core
right?
Cc-ing Peter for reference.
Best,
Juri
Quoted 134 lines · Juri Lelli
Hi,
Quoted 117 lines · Chunyan Zhang
From: Vincent Wang <[email protected]>
A deadlock issue is found when executing a cpu hotplug stress test on android phones with cpuset and scheduil enabled.
When CPUx is plugged out, the hotplug thread that calls
cpu_down()will holdcpu_hotplug_lockand wait the thread cpuhp/x to finish hotplug. If the core is the last one in a cluster, cpuhp/x have to callcpuhp_cpufreq_offline()and the kernel thread sugov need to exit for schedutil governor. The exit of sugov need to holdcgroup_threadgroup_rwseminexit_signals(). For example:PID: 150 TASK:
ffffffc0b9cad080CPU: 0 COMMAND: "sprdhotplug"
#0 [ffffff8009fcb9d0]__switch_toatffffff80080858f0#1 [ffffff8009fcb9f0]__scheduleatffffff80089f185c#2 [ffffff8009fcba80] schedule atffffff80089f1b84#3 [ffffff8009fcbaa0]schedule_timeoutatffffff80089f5124#4 [ffffff8009fcbb40]wait_for_commonatffffff80089f2944#5 [ffffff8009fcbbe0]wait_for_completionatffffff80089f29a4#6 [ffffff8009fcbc00]__cpuhp_kick_apatffffff80080ab030#7 [ffffff8009fcbc20]cpuhp_kick_ap_workatffffff80080ab154#8 [ffffff8009fcbc70]_cpu_downatffffff80089ee19c#9 [ffffff8009fcbcd0]cpu_downatffffff80080ac144PID: 26 TASK:
ffffffc0bbe22080CPU: 3 COMMAND: "cpuhp/3"
#0 [ffffff8009693a30]__switch_toatffffff80080858f0#1 [ffffff8009693a50]__scheduleatffffff80089f185c#2 [ffffff8009693ae0] schedule atffffff80089f1b84
#3 [ffffff8009693b00]schedule_timeoutatffffff80089f5124#4 [ffffff8009693ba0]wait_for_commonatffffff80089f2944#5 [ffffff8009693c40]wait_for_completionatffffff80089f29a4#6 [ffffff8009693c60]kthread_stopatffffff80080ccd2c#7 [ffffff8009693c90]sugov_exitatffffff8008102134#8 [ffffff8009693cc0]cpufreq_exit_governoratffffff80086c03bc#9 [ffffff8009693ce0]cpufreq_offlineatffffff80086c0634PID: 13819 TASK:
ffffffc0affb6080CPU: 0 COMMAND: "sugov:3"
#0 [ffffff800ee73c30]__switch_toatffffff80080858f0#1 [ffffff800ee73c50]__scheduleatffffff80089f185c#2 [ffffff800ee73ce0] schedule atffffff80089f1b84
#3 [ffffff800ee73d00]rwsem_down_read_failedatffffff80089f49d0#4 [ffffff800ee73d80]__percpu_down_readatffffff8008102ebc#5 [ffffff800ee73da0]exit_signalsatffffff80080bbd24#6 [ffffff800ee73de0]do_exitatffffff80080ae65c
#7 [ffffff800ee73e60] kthread atffffff80080cc550Sometimes
cgroup_threadgroup_rwsemis hold by another thread, for example Binder:681_2 on android, it wants to holdcpuset_mutex:PID: 732 TASK:
ffffffc09668b080CPU: 2 COMMAND: "Binder:681_2"
#0 [ffffff800cb7b8c0]__switch_toatffffff80080858f0
#1 [ffffff800cb7b8e0]__scheduleatffffff80089f185c
#2 [ffffff800cb7b970] schedule atffffff80089f1b84
#3 [ffffff800cb7b990]schedule_preempt_disabledatffffff80089f205c#4 [ffffff800cb7b9a0]__mutex_lockatffffff80089f3118#5 [ffffff800cb7ba40]__mutex_lock_slowpathatffffff80089f3230#6 [ffffff800cb7ba60]mutex_lockatffffff80089f3278
#7 [ffffff800cb7ba80]cpuset_can_attachatffffff8008152f84#8 [ffffff800cb7bae0]cgroup_migrate_executeatffffff800814ada8On android, a thread kworker/3:0 will hold
cpuset_mutexinrebuild_sched_domains()and want to holdcpu_hotplug_lockwhich is already hold by the hotplug thread.PID: 4847 TASK:
ffffffc031a6a080CPU: 3 COMMAND: "kworker/3:0"
#0 [ffffff8016fd3ad0]__switch_toatffffff80080858f0
#1 [ffffff8016fd3af0]__scheduleatffffff80089f185c
#2 [ffffff8016fd3b80] schedule atffffff80089f1b84
#3 [ffffff8016fd3ba0]rwsem_down_read_failedatffffff80089f49d0#4 [ffffff8016fd3c20]__percpu_down_readatffffff8008102ebc#5 [ffffff8016fd3c40]cpus_read_lockatffffff80080aa59c
#6 [ffffff8016fd3c50]rebuild_sched_domains_lockedatffffff80081522a8In order to fix the deadlock, this patch will adjust the lock sequence when rebuilding sched domains. After stress tests, it works well.
Signed-off-by Vincent Wang <[email protected]>Signed-off-by Chunyan Zhang <[email protected]>--- kernel/cgroup/cpuset.c | 9 ++++----- 1 file changed, 4 insertions(+), 5 deletions(-)kernel/cgroup/cpuset.c +2
-5@@ -912,7 +912,6 @@ static void rebuild_sched_domains_locked(void)int ndoms;lockdep_assert_held (&cpuset_mutex); -get_online_cpus ();/* * We have raced with CPU hotplug. Don't do anything to avoid @@ -921,19 +920,17 @@ static void rebuild_sched_domains_locked(void)*/ if (!top_cpuset.nr_subparts_cpus && !cpumask_equal (top_cpuset.effective_cpus, cpu_active_mask)) - goto out; +return ;if (top_cpuset.nr_subparts_cpus && !cpumask_subset (top_cpuset.effective_cpus, cpu_active_mask)) - goto out; +return ;/* Generate domain masks and attrs */ ndoms =generate_sched_domains (&doms, &attr);/* Have scheduler rebuild the domains */ partition_sched_domains(ndoms, doms, attr); -out: - put_online_cpus(); }#else /* !CONFIG_SMP */ static void rebuild_sched_domains_locked (void )kernel/cgroup/cpuset.c static void rebuild_sched_domains_locked(void) · line 948
Quoted 9 lines
kernel/cgroup/cpuset.c +2
-0@@ -943,9 +940,11 @@ static void rebuild_sched_domains_locked(void)void rebuild_sched_domains (void ) { +get_online_cpus ();mutex_lock (&cpuset_mutex);rebuild_sched_domains_locked ();mutex_unlock (&cpuset_mutex); +put_online_cpus (); }This looks a subset of
d74b27d63a8b("cgroup/cpuset: Changecpuset_rwsemand hotplug lock order") from
Quoted 3 lines
git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.gitsched/coreright?
Yes, seems that your patch is a more thorough solution :-)
Quoted 6 lines
Cc-ing Peter for reference.
Best,
Juri