Re: [PATCH 2/2] clocksource: Defer preempt_disable() after clocksource_verify_choose_cpus()

[Waiman Long <llong@redhat.com>]

On 1/24/25 4:37 PM, Paul E. McKenney wrote:
> On Fri, Jan 24, 2025 at 03:41:45PM -0500, Waiman Long wrote:
>> On 1/24/25 3:37 PM, Paul E. McKenney wrote:
>>> On Fri, Jan 24, 2025 at 12:25:19PM -0800, Paul E. McKenney wrote:
>>>> On Fri, Jan 24, 2025 at 01:59:23PM -0500, Waiman Long wrote:
>>>>> The following bug report happened in a PREEMPT_RT kernel.
>>>>>
>>>>> [   30.957705] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:48
>>>>> [   30.957711] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 2012, name: kwatchdog
>>>>> [   30.962673] preempt_count: 1, expected: 0
>>>>> [   30.962676] RCU nest depth: 0, expected: 0
>>>>> [   30.962680] 3 locks held by kwatchdog/2012:
>>>>> [   30.962684]  #0: ffffffff8af2da60 (clocksource_mutex){+.+.}-{3:3}, at: clocksource_watchdog_kthread+0x13/0x50
>>>>> [   30.967703]  #1: ffffffff8aa8d4d0 (cpu_hotplug_lock){++++}-{0:0}, at: clocksource_verify_percpu.part.0+0x5c/0x330
>>>>> [   30.972774]  #2: ffff9fe02f5f33e0 ((batched_entropy_u32.lock)){+.+.}-{2:2}, at: get_random_u32+0x4f/0x110
>>>>> [   30.977827] Preemption disabled at:
>>>>> [   30.977830] [<ffffffff88c1fe56>] clocksource_verify_percpu.part.0+0x66/0x330
>>>>> [   30.982837] CPU: 33 PID: 2012 Comm: kwatchdog Not tainted 5.14.0-503.23.1.el9_5.x86_64+rt-debug #1
>>>>> [   30.982843] Hardware name: HPE ProLiant DL385 Gen10 Plus/ProLiant DL385 Gen10 Plus, BIOS A42 04/29/2021
>>>>> [   30.982846] Call Trace:
>>>>> [   30.982850]  <TASK>
>>>>> [   30.983821]  dump_stack_lvl+0x57/0x81
>>>>> [   30.983821]  __might_resched.cold+0xf4/0x12f
>>>>> [   30.983824]  rt_spin_lock+0x4c/0x100
>>>>> [   30.988833]  get_random_u32+0x4f/0x110
>>>>> [   30.988833]  clocksource_verify_choose_cpus+0xab/0x1a0
>>>>> [   30.988833]  clocksource_verify_percpu.part.0+0x6b/0x330
>>>>> [   30.993894]  __clocksource_watchdog_kthread+0x193/0x1a0
>>>>> [   30.993898]  clocksource_watchdog_kthread+0x18/0x50
>>>>> [   30.993898]  kthread+0x114/0x140
>>>>> [   30.993898]  ret_from_fork+0x2c/0x50
>>>>> [   31.002864]  </TASK>
>>>>>
>>>>> It is due to the fact that get_random_u32() is called in
>>>>> clocksource_verify_choose_cpus() with preemption disabled.
>>>>> If crng_ready() is true by the time get_random_u32() is called, The
>>>>> batched_entropy_32 local lock will be acquired. In PREEMPT_RT kernel,
>>>>> it is a rtmutex and we can't acquire it with preemption disabled.
>>>>>
>>>>> To avoid this problem, we can't call get_random_u32() with preemption
>>>>> disabled. However, smp_processor_id() has to be called with preemption
>>>>> disabled though and we have to exclude the current CPU from the
>>>>> cpus_chosen list to be tested.
>>>>>
>>>>> Extract current CPU removal code out from
>>>>> clocksource_verify_choose_cpus() and defer the preempt_disable()
>>>>> call to after clocksource_verify_choose_cpus() and before
>>>>> current CPU removal. Also use raw_smp_processor_id() in
>>>>> clocksource_verify_choose_cpus().
>>>>>
>>>>> Fixes: 7560c02bdffb ("clocksource: Check per-CPU clock synchronization when marked unstable")
>>>>> Signed-off-by: Waiman Long <longman@redhat.com>
>>>> Good catch!
>>>>
>>>> But we don't need cryptographically secure random numbers (or blistering
>>>> speed) here.  Substituting something like torture_random()%nr_cpu_ids
>>>> for get_random_u32_below() work?
>>> I suppose I should add my concern...  If we don't have preemption disabled
>>> across this code, we cannot reliably avoid attempting to IPI ourselves.
>> Does the CPU choosing process itsself needs to have preemption disabled? I
>> thought it was because of need to use smp_processor_id() and have the
>> current CPU excluded. Preemption is still disabled after that.
> If the function is migrated from one CPU to another, the check against
> raw_smp_processor_id() doesn't mean much.  At that point, you might
> as well just rely in the calling function clearing it.
>
> And the extra preemption happens only in an error condition were extra
> debugging is enabled.  Plus clocksource_verify_choose_cpus() should be
> pretty quick.  So do we really care about the additional disabling of
> preemption?

I understand your concern about using raw_smp_processor_id().

Anyway, I took your suggestion of using another less random 
pseudo-random generator and call it a day. Unfortunately, I can't use 
torture_random() as it is only available if CONFIG_TORTURE_TEST is set 
:-(. Please let me know if you are OK with the v2 patch.

Cheers,
Longman