On 23/03/23 18:41, Peter Zijlstra wrote:
> On Thu, Mar 23, 2023 at 04:25:25PM +0000, Valentin Schneider wrote:
>> On 22/03/23 15:04, Peter Zijlstra wrote:
>> > @@ -798,14 +794,20 @@ static void smp_call_function_many_cond(
>> > }
>> >
>> > /*
>> > + * Trace each smp_function_call_*() as an IPI, actual IPIs
>> > + * will be traced with func==generic_smp_call_function_single_ipi().
>> > + */
>> > + trace_ipi_send_cpumask(cfd->cpumask_ipi, _RET_IP_, func);
>>
>> I just got a trace pointing out this can emit an event even though no IPI
>> is sent if e.g. the cond_func predicate filters all CPUs in the argument
>> mask:
>>
>> ipi_send_cpumask: cpumask= callsite=on_each_cpu_cond_mask+0x3c callback=flush_tlb_func+0x0
>>
>> Maybe something like so on top?
>>
>> ---
>> diff --git a/kernel/smp.c b/kernel/smp.c
>> index ba5478814e677..1dc452017d000 100644
>> --- a/kernel/smp.c
>> +++ b/kernel/smp.c
>> @@ -791,6 +791,8 @@ static void smp_call_function_many_cond(const struct cpumask *mask,
>> }
>> }
>>
>> + if (!nr_cpus)
>> + goto local;
>
> Hmm, this isn't right. You can get nr_cpus==0 even though it did add
> some to various lists but never was first.
>
Duh, glanced over that.
> But urgh, even if we were to say count nr_queued we'd never get the mask
> right, because we don't track which CPUs have the predicate matched,
> only those we need to actually send an IPI to :/
>
> Ooh, I think we can clear those bits from cfd->cpumask, arguably that's
> a correctness fix too, because the 'run_remote && wait' case shouldn't
> wait on things we didn't queue.
>
Yeah, that makes sense to me. Just one tiny suggestion below.
> Hmm?
>
>
> --- a/kernel/smp.c
> +++ b/kernel/smp.c
> @@ -728,9 +728,9 @@ static void smp_call_function_many_cond(
> int cpu, last_cpu, this_cpu = smp_processor_id();
> struct call_function_data *cfd;
> bool wait = scf_flags & SCF_WAIT;
> + int nr_cpus = 0, nr_queued = 0;
> bool run_remote = false;
> bool run_local = false;
> - int nr_cpus = 0;
>
> lockdep_assert_preemption_disabled();
>
> @@ -772,8 +772,10 @@ static void smp_call_function_many_cond(
> for_each_cpu(cpu, cfd->cpumask) {
> call_single_data_t *csd = per_cpu_ptr(cfd->csd, cpu);
>
> - if (cond_func && !cond_func(cpu, info))
> + if (cond_func && !cond_func(cpu, info)) {
> + __cpumask_clear_cpu(cpu, cfd->cpumask);
> continue;
> + }
>
> csd_lock(csd);
> if (wait)
> @@ -789,13 +791,15 @@ static void smp_call_function_many_cond(
> nr_cpus++;
> last_cpu = cpu;
> }
> + nr_queued++;
> }
>
> /*
> * Trace each smp_function_call_*() as an IPI, actual IPIs
> * will be traced with func==generic_smp_call_function_single_ipi().
> */
> - trace_ipi_send_cpumask(cfd->cpumask_ipi, _RET_IP_, func);
> + if (nr_queued)
With your change to cfd->cpumask, we could ditch nr_queued and make this
if (!cpumask_empty(cfd->cpumask))
since cfd->cpumask now only contains CPUs that have had a CSD queued.
> + trace_ipi_send_cpumask(cfd->cpumask, _RET_IP_, func);
>
> /*
> * Choose the most efficient way to send an IPI. Note that the
