Re: [PATCH v3] tracing: Guard __DECLARE_TRACE() use of __DO_TRACE_CALL() with SRCU-fast

[Joel Fernandes <joelagnelf@nvidia.com>]

> On Dec 13, 2025, at 8:10 AM, Paul E. McKenney <paulmck@kernel.org> wrote:
> 
> On Fri, Dec 12, 2025 at 09:28:37AM +0000, Joel Fernandes wrote:
>> 
>> 
>>>> On Dec 12, 2025, at 4:50 PM, Paul E. McKenney <paulmck@kernel.org> wrote:
>>> 
>>> On Fri, Dec 12, 2025 at 03:43:07AM +0000, Joel Fernandes wrote:
>>>> 
>>>> 
>>>>>> On Dec 12, 2025, at 9:47 AM, Paul E. McKenney <paulmck@kernel.org> wrote:
>>>>> 
>>>>> On Fri, Dec 12, 2025 at 09:12:07AM +0900, Joel Fernandes wrote:
>>>>>> 
>>>>>> 
>>>>>>> On 12/11/2025 3:23 PM, Paul E. McKenney wrote:
>>>>>>> On Thu, Dec 11, 2025 at 08:02:15PM +0000, Joel Fernandes wrote:
>>>>>>>> 
>>>>>>>> 
>>>>>>>>> On Dec 8, 2025, at 1:20 PM, Paul E. McKenney <paulmck@kernel.org> wrote:
>>>>>>>>> 
>>>>>>>>> The current use of guard(preempt_notrace)() within __DECLARE_TRACE()
>>>>>>>>> to protect invocation of __DO_TRACE_CALL() means that BPF programs
>>>>>>>>> attached to tracepoints are non-preemptible.  This is unhelpful in
>>>>>>>>> real-time systems, whose users apparently wish to use BPF while also
>>>>>>>>> achieving low latencies.  (Who knew?)
>>>>>>>>> 
>>>>>>>>> One option would be to use preemptible RCU, but this introduces
>>>>>>>>> many opportunities for infinite recursion, which many consider to
>>>>>>>>> be counterproductive, especially given the relatively small stacks
>>>>>>>>> provided by the Linux kernel.  These opportunities could be shut down
>>>>>>>>> by sufficiently energetic duplication of code, but this sort of thing
>>>>>>>>> is considered impolite in some circles.
>>>>>>>>> 
>>>>>>>>> Therefore, use the shiny new SRCU-fast API, which provides somewhat faster
>>>>>>>>> readers than those of preemptible RCU, at least on Paul E. McKenney's
>>>>>>>>> laptop, where task_struct access is more expensive than access to per-CPU
>>>>>>>>> variables.  And SRCU-fast provides way faster readers than does SRCU,
>>>>>>>>> courtesy of being able to avoid the read-side use of smp_mb().  Also,
>>>>>>>>> it is quite straightforward to create srcu_read_{,un}lock_fast_notrace()
>>>>>>>>> functions.
>>>>>>>>> 
>>>>>>>>> While in the area, SRCU now supports early boot call_srcu().  Therefore,
>>>>>>>>> remove the checks that used to avoid such use from rcu_free_old_probes()
>>>>>>>>> before this commit was applied:
>>>>>>>>> 
>>>>>>>>> e53244e2c893 ("tracepoint: Remove SRCU protection")
>>>>>>>>> 
>>>>>>>>> The current commit can be thought of as an approximate revert of that
>>>>>>>>> commit, with some compensating additions of preemption disabling.
>>>>>>>>> This preemption disabling uses guard(preempt_notrace)().
>>>>>>>>> 
>>>>>>>>> However, Yonghong Song points out that BPF assumes that non-sleepable
>>>>>>>>> BPF programs will remain on the same CPU, which means that migration
>>>>>>>>> must be disabled whenever preemption remains enabled.  In addition,
>>>>>>>>> non-RT kernels have performance expectations that would be violated by
>>>>>>>>> allowing the BPF programs to be preempted.
>>>>>>>>> 
>>>>>>>>> Therefore, continue to disable preemption in non-RT kernels, and protect
>>>>>>>>> the BPF program with both SRCU and migration disabling for RT kernels,
>>>>>>>>> and even then only if preemption is not already disabled.
>>>>>>>> 
>>>>>>>> Hi Paul,
>>>>>>>> 
>>>>>>>> Is there a reason to not make non-RT also benefit from SRCU fast and trace points for BPF? Can be a follow up patch though if needed.
>>>>>>> 
>>>>>>> Because in some cases the non-RT benefit is suspected to be negative
>>>>>>> due to increasing the probability of preemption in awkward places.
>>>>>> 
>>>>>> Since you mentioned suspected, I am guessing there is no concrete data collected
>>>>>> to substantiate that specifically for BPF programs, but correct me if I missed
>>>>>> something. Assuming you're referring to latency versus tradeoffs issues, due to
>>>>>> preemption, Android is not PREEMPT_RT but is expected to be low latency in
>>>>>> general as well. So is this decision the right one for Android as well,
>>>>>> considering that (I heard) it uses BPF? Just an open-ended question.
>>>>>> 
>>>>>> There is also issue of 2 different paths for PREEMPT_RT versus otherwise,
>>>>>> complicating the tracing side so there better be a reason for that I guess.
>>>>> 
>>>>> You are advocating a change in behavior for non-RT workloads.  Why do
>>>>> you believe that this change would be OK for those workloads?
>>>> 
>>>> Same reasons I provided in my last email. If we are saying SRCU-fast is required for lower latency, I find it strange that we are leaving out Android which has low latency audio usecases, for instance.
>>> 
>>> If Android provides numbers showing that it helps them, then it is easy
>>> to provide a Kconfig option that defaults to PREEMPT_RT, but that Android
>>> can override.  Right?
>> 
>> Sure, but my suspicion is Android or others are not going to look into every PREEMPT_RT specific optimization (not just this one) and see if it benefits their interactivity usecases. They will simply miss out on it without knowing they are.
>> 
>> It might be a good idea (for me) to explore how many such optimizations exist though, that we take for granted. I will look into exploring this on my side. :)
> 
> One workload's optimization is another workload's pessimization, in
> part because there are a lot of different measures of performance that
> different workloads care about..
> 
> But as a practical matter, this is Steven's decision.
> 
> Though if he does change the behavior on non-RT setups, I would thank
> him to remove my name from the commit, or at least record in the commit
> log that I object to changing other workloads' behaviors.

You have a point. I am not saying we should do this for sure but should at least consider / explore it.

Thanks.



> 
>                            Thanx, Paul
> 
>> thanks,
>> 
>> - Joel
>> 
>>> 
>>>                           Thanx, Paul
>>> 
>>>> Thanks,
>>>> 
>>>> - Joel
>>>> 
>>>> 
>>>>> 
>>>>>                          Thanx, Paul