Re: [Question] set_cpus_allowed_ptr() call failed at cpuset_attach()

[Waiman Long]

On 1/16/22 21:25, Zhang Qiao wrote:
> hello
>
> 在 2022/1/15 4:33, Waiman Long 写道:
> > On 1/14/22 11:20, Tejun Heo wrote:
> > > (cc'ing Waiman and Michal and quoting whole body)
> > >
> > > Seems sane to me but let's hear what Waiman and Michal think.
> > >
> > > On Fri, Jan 14, 2022 at 09:15:06AM +0800, Zhang Qiao wrote:
> > > > Hello everyone
> > > >
> > > >      I found the following warning log on qemu. I migrated a task from one cpuset cgroup to
> > > > another, while I also performed the cpu hotplug operation, and got following calltrace.
> > > >
> > > >      This may lead to a inconsistency between the affinity of the task and cpuset.cpus of the
> > > > dest cpuset, but this task can be successfully migrated to the dest cpuset cgroup.
> > > >
> > > >      Can we use cpus_read_lock()/cpus_read_unlock() to guarantee that set_cpus_allowed_ptr()
> > > > doesn't fail, as follows:
> > > >
> > > > diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> > > > index d0e163a02099..2535d23d2c51 100644
> > > > --- a/kernel/cgroup/cpuset.c
> > > > +++ b/kernel/cgroup/cpuset.c
> > > > @@ -2265,6 +2265,7 @@ static void cpuset_attach(struct cgroup_taskset *tset)
> > > >           guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
> > > >
> > > >           cgroup_taskset_for_each(task, css, tset) {
> > > > +               cpus_read_lock();
> > > >                   if (cs != &top_cpuset)
> > > >                           guarantee_online_cpus(task, cpus_attach);
> > > >                   else
> > > > @@ -2274,6 +2275,7 @@ static void cpuset_attach(struct cgroup_taskset *tset)
> > > >                    * fail.  TODO: have a better way to handle failure here
> > > >                    */
> > > >                   WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
> > > > +               cpus_read_unlock();
> > > >
> > > >
> > > >      Is there a better solution?
> > > >
> > > >      Thanks
> > The change looks OK to me. However, we may need to run the full set of regression test to make sure that lockdep won't complain about potential deadlock.
> I run the test with lockdep enabled, and got lockdep warning like that below.
> so we should take the cpu_hotplug_lock first, then take the cpuset_rwsem lock.
>
> thanks,
> Zhang Qiao
>
> [   38.420372] ======================================================
> [   38.421339] WARNING: possible circular locking dependency detected
> [   38.422312] 5.16.0-rc4+ #13 Not tainted
> [   38.422920] ------------------------------------------------------
> [   38.423883] bash/594 is trying to acquire lock:
> [   38.424595] ffffffff8286afc0 (cpu_hotplug_lock){++++}-{0:0}, at: cpuset_attach+0xc2/0x1e0
> [   38.425880]
> [   38.425880] but task is already holding lock:
> [   38.426787] ffffffff8296a5a0 (&cpuset_rwsem){++++}-{0:0}, at: cpuset_attach+0x3e/0x1e0
> [   38.428015]
> [   38.428015] which lock already depends on the new lock.
> [   38.428015]
> [   38.429279]
> [   38.429279] the existing dependency chain (in reverse order) is:
> [   38.430445]
> [   38.430445] -> #1 (&cpuset_rwsem){++++}-{0:0}:
> [   38.431371]        percpu_down_write+0x42/0x130
> [   38.432085]        cpuset_css_online+0x2b/0x2e0
> [   38.432808]        online_css+0x24/0x80
> [   38.433411]        cgroup_apply_control_enable+0x2fa/0x330
> [   38.434273]        cgroup_mkdir+0x396/0x4c0
> [   38.434930]        kernfs_iop_mkdir+0x56/0x80
> [   38.435614]        vfs_mkdir+0xde/0x190
> [   38.436220]        do_mkdirat+0x7d/0xf0
> [   38.436824]        __x64_sys_mkdir+0x21/0x30
> [   38.437495]        do_syscall_64+0x3a/0x80
> [   38.438145]        entry_SYSCALL_64_after_hwframe+0x44/0xae
> [   38.439015]
> [   38.439015] -> #0 (cpu_hotplug_lock){++++}-{0:0}:
> [   38.439980]        __lock_acquire+0x17f6/0x2260
> [   38.440691]        lock_acquire+0x277/0x320
> [   38.441347]        cpus_read_lock+0x37/0xc0
> [   38.442011]        cpuset_attach+0xc2/0x1e0
> [   38.442671]        cgroup_migrate_execute+0x3a6/0x490
> [   38.443461]        cgroup_attach_task+0x22c/0x3d0
> [   38.444197]        __cgroup1_procs_write.constprop.21+0x10d/0x170
> [   38.445145]        cgroup_file_write+0x6f/0x230
> [   38.445860]        kernfs_fop_write_iter+0x130/0x1b0
> [   38.446636]        new_sync_write+0x120/0x1b0
> [   38.447319]        vfs_write+0x359/0x3b0
> [   38.447937]        ksys_write+0xa2/0xe0
> [   38.448540]        do_syscall_64+0x3a/0x80
> [   38.449183]        entry_SYSCALL_64_after_hwframe+0x44/0xae
> [   38.450057]
> [   38.450057] other info that might help us debug this:
> [   38.450057]
> [   38.451297]  Possible unsafe locking scenario:
> [   38.451297]
> [   38.452218]        CPU0                    CPU1
> [   38.452935]        ----                    ----
> [   38.453650]   lock(&cpuset_rwsem);
> [   38.454188]                                lock(cpu_hotplug_lock);
> [   38.455148]                                lock(&cpuset_rwsem);
> [   38.456069]   lock(cpu_hotplug_lock);

Yes, you need to play around with lock ordering to make sure that 
lockdep won't complain.

Cheers,
Longman