Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   Chunyan Zhang <zhang.lyra@gmail.com>
To:     Li Zefan <lizefan@huawei.com>, Tejun Heo <tj@kernel.org>,
        Johannes Weiner <hannes@cmpxchg.org>
Cc:     cgroups@vger.kernel.org, linux-kernel@vger.kernel.org,
        Vincent Wang <vincent.wang@unisoc.com>,
        Chunyan Zhang <zhang.lyra@gmail.com>
Subject: [PATCH] cpuset: adjust the lock sequence when rebuilding the sched domains.
Date:   Thu,  5 Sep 2019 10:18:08 +0800
Message-Id: <20190905021808.25130-1-zhang.lyra@gmail.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

From: Vincent Wang <vincent.wang@unisoc.com>

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.

Signed-off-by: Vincent Wang <vincent.wang@unisoc.com>
Signed-off-by: Chunyan Zhang <zhang.lyra@gmail.com>
---
 kernel/cgroup/cpuset.c | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
index 5aa37531ce76..ef10d276da22 100644
--- a/kernel/cgroup/cpuset.c
+++ b/kernel/cgroup/cpuset.c
@@ -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();
 }
 
 /**
-- 
2.20.1