Re: [RFC PATCH v2 8/8] sched, smp: Trace smp callback causing an IPI

From: Valentin Schneider
Date: Thu Nov 17 2022 - 09:46:43 EST


On 17/11/22 15:12, Peter Zijlstra wrote:
> On Wed, Nov 02, 2022 at 06:33:36PM +0000, Valentin Schneider wrote:
> *yuck*

:-)

>
> How about something like so?
>
> ---
> --- a/kernel/irq_work.c
> +++ b/kernel/irq_work.c
> @@ -24,6 +24,8 @@
>
> #include <trace/events/ipi.h>
>
> +#include "sched/smp.h"
> +
> static DEFINE_PER_CPU(struct llist_head, raised_list);
> static DEFINE_PER_CPU(struct llist_head, lazy_list);
> static DEFINE_PER_CPU(struct task_struct *, irq_workd);
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -3763,16 +3763,17 @@ void sched_ttwu_pending(void *arg)
> rq_unlock_irqrestore(rq, &rf);
> }
>
> -void send_call_function_single_ipi(int cpu)
> +bool send_call_function_single_ipi(int cpu)
> {
> struct rq *rq = cpu_rq(cpu);
>
> if (!set_nr_if_polling(rq->idle)) {
> - trace_ipi_send_cpumask(cpumask_of(cpu), _RET_IP_, NULL);
> arch_send_call_function_single_ipi(cpu);
> - } else {
> - trace_sched_wake_idle_without_ipi(cpu);
> + return true;
> }
> +
> + trace_sched_wake_idle_without_ipi(cpu);
> + return false;
> }
>
> /*
> --- a/kernel/sched/smp.h
> +++ b/kernel/sched/smp.h
> @@ -6,7 +6,7 @@
>
> extern void sched_ttwu_pending(void *arg);
>
> -extern void send_call_function_single_ipi(int cpu);
> +extern bool send_call_function_single_ipi(int cpu);
>
> #ifdef CONFIG_SMP
> extern void flush_smp_call_function_queue(void);
> --- a/kernel/smp.c
> +++ b/kernel/smp.c
> @@ -163,7 +163,6 @@ void __init call_function_init(void)
> static inline void
> send_call_function_ipi_mask(const struct cpumask *mask)
> {
> - trace_ipi_send_cpumask(mask, _RET_IP_, func);
> arch_send_call_function_ipi_mask(mask);
> }
>
> @@ -438,11 +437,16 @@ static void __smp_call_single_queue_debu
> struct cfd_seq_local *seq = this_cpu_ptr(&cfd_seq_local);
> struct call_function_data *cfd = this_cpu_ptr(&cfd_data);
> struct cfd_percpu *pcpu = per_cpu_ptr(cfd->pcpu, cpu);
> + struct __call_single_data *csd;
> +
> + csd = container_of(node, call_single_data_t, node.llist);
> + WARN_ON_ONCE(!(CSD_TYPE(csd) & (CSD_TYPE_SYNC | CSD_TYPE_ASYNC)));
>
> cfd_seq_store(pcpu->seq_queue, this_cpu, cpu, CFD_SEQ_QUEUE);
> if (llist_add(node, &per_cpu(call_single_queue, cpu))) {
> cfd_seq_store(pcpu->seq_ipi, this_cpu, cpu, CFD_SEQ_IPI);
> cfd_seq_store(seq->ping, this_cpu, cpu, CFD_SEQ_PING);
> + trace_ipi_send_cpumask(cpumask_of(cpu), _RET_IP_, csd->func);
> send_call_function_single_ipi(cpu);
> cfd_seq_store(seq->pinged, this_cpu, cpu, CFD_SEQ_PINGED);
> } else {
> @@ -487,6 +491,27 @@ static __always_inline void csd_unlock(s
>
> static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
>
> +static __always_inline
> +bool raw_smp_call_single_queue(int cpu, struct llist_node *node)
> +{
> + /*
> + * The list addition should be visible to the target CPU when it pops
> + * the head of the list to pull the entry off it in the IPI handler
> + * because of normal cache coherency rules implied by the underlying
> + * llist ops.
> + *
> + * If IPIs can go out of order to the cache coherency protocol
> + * in an architecture, sufficient synchronisation should be added
> + * to arch code to make it appear to obey cache coherency WRT
> + * locking and barrier primitives. Generic code isn't really
> + * equipped to do the right thing...
> + */
> + if (llist_add(node, &per_cpu(call_single_queue, cpu)))
> + return send_call_function_single_ipi(cpu);
> +
> + return false;
> +}
> +
> void __smp_call_single_queue(int cpu, struct llist_node *node)
> {
> #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
> @@ -503,19 +528,28 @@ void __smp_call_single_queue(int cpu, st
> #endif
>
> /*
> - * The list addition should be visible to the target CPU when it pops
> - * the head of the list to pull the entry off it in the IPI handler
> - * because of normal cache coherency rules implied by the underlying
> - * llist ops.
> - *
> - * If IPIs can go out of order to the cache coherency protocol
> - * in an architecture, sufficient synchronisation should be added
> - * to arch code to make it appear to obey cache coherency WRT
> - * locking and barrier primitives. Generic code isn't really
> - * equipped to do the right thing...
> - */
> - if (llist_add(node, &per_cpu(call_single_queue, cpu)))
> - send_call_function_single_ipi(cpu);
> + * We have to check the type of the CSD before queueing it, because
> + * once queued it can have its flags cleared by
> + * flush_smp_call_function_queue()
> + * even if we haven't sent the smp_call IPI yet (e.g. the stopper
> + * executes migration_cpu_stop() on the remote CPU).
> + */
> + if (trace_ipi_send_cpumask_enabled()) {
> + call_single_data_t *csd;
> + smp_call_func_t func;
> +
> + csd = container_of(node, call_single_data_t, node.llist);
> +
> + func = sched_ttwu_pending;
> + if (CSD_TYPE(csd) != CSD_TYPE_TTWU)
> + func = csd->func;
> +
> + if (raw_smp_call_single_queue(cpu, node))
> + trace_ipi_send_cpumask(cpumask_of(cpu), _RET_IP_, func);

So I went with the tracepoint being placed *before* the actual IPI gets
sent to have a somewhat sane ordering between trace_ipi_send_cpumask() and
e.g. trace_call_function_single_entry().

Packaging the call_single_queue logic makes the code less horrible, but it
does mix up the event ordering...


> + return;
> + }
> +
> + raw_smp_call_single_queue(cpu, node);
> }
>
> /*
> @@ -983,10 +1017,13 @@ static void smp_call_function_many_cond(
> * number of CPUs might be zero due to concurrent changes to the
> * provided mask.
> */
> - if (nr_cpus == 1)
> + if (nr_cpus == 1) {
> + trace_ipi_send_cpumask(cpumask_of(last_cpu), _RET_IP_, func);
> send_call_function_single_ipi(last_cpu);

This'll yield an IPI event even if no IPI is sent due to the idle task
polling, no?

> - else if (likely(nr_cpus > 1))
> - send_call_function_ipi_mask(cfd->cpumask_ipi);
> + } else if (likely(nr_cpus > 1)) {
> + trace_ipi_send_cpumask(mask, _RET_IP_, func);
> + send_call_function_ipi_mask(mask);
> + }
>
> cfd_seq_store(this_cpu_ptr(&cfd_seq_local)->pinged, this_cpu, CFD_SEQ_NOCPU, CFD_SEQ_PINGED);
> }