Re: [BUG BISECTED] Missing RCU reader in perf_event_setup_cpumask()

[Liang, Kan]

On 2024-09-13 6:47 a.m., Peter Zijlstra wrote:
> On Fri, Sep 13, 2024 at 01:00:44AM -0700, Paul E. McKenney wrote:
>> Hello!
>>
>> On next-20240912 running rcutorture scenario TREE05, I see this
>> deterministically:
>>
>> [   32.603233] =============================
>> [   32.604594] WARNING: suspicious RCU usage
>> [   32.605928] 6.11.0-rc5-00040-g4ba4f1afb6a9 #55238 Not tainted
>> [   32.607812] -----------------------------
>> [   32.609140] kernel/events/core.c:13946 RCU-list traversed in non-reader section!!
>> [   32.611595]
>> [   32.611595] other info that might help us debug this:
>> [   32.611595]
>> [   32.614247]
>> [   32.614247] rcu_scheduler_active = 2, debug_locks = 1
>> [   32.616392] 3 locks held by cpuhp/4/35:
>> [   32.617687]  #0: ffffffffb666a650 (cpu_hotplug_lock){++++}-{0:0}, at: cpuhp_thread_fun+0x4e/0x200
>> [   32.620563]  #1: ffffffffb666cd20 (cpuhp_state-down){+.+.}-{0:0}, at: cpuhp_thread_fun+0x4e/0x200
>> [   32.623412]  #2: ffffffffb677c288 (pmus_lock){+.+.}-{3:3}, at: perf_event_exit_cpu_context+0x32/0x2f0
>> [   32.626399]
>> [   32.626399] stack backtrace:
>> [   32.627848] CPU: 4 UID: 0 PID: 35 Comm: cpuhp/4 Not tainted 6.11.0-rc5-00040-g4ba4f1afb6a9 #55238
>> [   32.628832] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014
>> [   32.628832] Call Trace:
>> [   32.628832]  <TASK>
>> [   32.628832]  dump_stack_lvl+0x83/0xa0
>> [   32.628832]  lockdep_rcu_suspicious+0x143/0x1a0
>> [   32.628832]  perf_event_exit_cpu_context+0x2e5/0x2f0
>> [   32.628832]  ? __pfx_perf_event_exit_cpu+0x10/0x10
>> [   32.628832]  perf_event_exit_cpu+0x9/0x10
>> [   32.628832]  cpuhp_invoke_callback+0x130/0x2a0
>> [   32.628832]  ? lock_release+0xc7/0x290
>> [   32.628832]  ? cpuhp_thread_fun+0x4e/0x200
>> [   32.628832]  cpuhp_thread_fun+0x183/0x200
>> [   32.628832]  smpboot_thread_fn+0xd8/0x1d0
>> [   32.628832]  ? __pfx_smpboot_thread_fn+0x10/0x10
>> [   32.628832]  kthread+0xd4/0x100
>> [   32.628832]  ? __pfx_kthread+0x10/0x10
>> [   32.628832]  ret_from_fork+0x2f/0x50
>> [   32.628832]  ? __pfx_kthread+0x10/0x10
>> [   32.628832]  ret_from_fork_asm+0x1a/0x30
>> [   32.628832]  </TASK>
>>
>> I bisected this to:
>>
>> 4ba4f1afb6a9 ("perf: Generic hotplug support for a PMU with a scope")
>>
>> This adds a perf_event_setup_cpumask() function that uses
>> list_for_each_entry_rcu() without an obvious RCU read-side critical
>> section, so the fix might be as simple as adding rcu_read_lock() and
>> rcu_read_unlock().  In the proper places, of course.  ;-)
> 
> IIRC that condition should be:
> 
>   lockdep_is_held(&pmus_srcu) || lockdep_is_held(&pmus_lock)
> 
> And at this pooint we actually do hold pmus_lock.
> 
> But that all begs the question why we're using RCU iteration here to
> begin with, as this code seems to be only called from this context.

I think I just copied and paste the PMU iterate code here, and forget to
add the srcu_read_lock(). Sorry for it.

> 
> Kan, is the simple fix to do:
> 
> -	list_for_each_entry_rcu(pmu, &pmus, entry, lockdep_is_held(&pmus_srcu)) {
> +	list_for_each_entry(pmu, &pmus, entry) {
> 
> ?
>   

Yes, the &pmus_lock protect is good enough. we don't need the rcu here.
I will post a patch with the suggested fix.

Thanks,
Kan