| Message ID | 20221102183336.3120536-7-vschneid@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Generic IPI sending tracepoint | expand |
On Wed, Nov 02, 2022 at 06:33:36PM +0000, Valentin Schneider wrote: > The newly-introduced ipi_send_cpumask tracepoint has a "callback" parameter > which so far has only been fed with NULL. > > While CSD_TYPE_SYNC/ASYNC and CSD_TYPE_IRQ_WORK share a similar backing > struct layout (meaning their callback func can be accessed without caring > about the actual CSD type), CSD_TYPE_TTWU doesn't even have a function > attached to its struct. This means we need to check the type of a CSD > before eventually dereferencing its associated callback. > > This isn't as trivial as it sounds: the CSD type is stored in > __call_single_node.u_flags, which get cleared right before the callback is > executed via csd_unlock(). This implies checking the CSD type before it is > enqueued on the call_single_queue, as the target CPU's queue can be flushed > before we get to sending an IPI. > > Furthermore, send_call_function_single_ipi() only has a CPU parameter, and > would need to have an additional argument to trickle down the invoked > function. This is somewhat silly, as the extra argument will always be > pushed down to the function even when nothing is being traced, which is > unnecessary overhead. > > Two options present themselves: > a) Create copies of send_call_function_{single_ipi, ipi_mask}() that take > an extra argument used for tracing, so that codepaths remain unchanged > when tracing isn't in effect (a sort of manual -fipa-sra). > > b) Stash the CSD func in somewhere as a side effect that > the portion of send_call_function_{single_ipi, ipi_mask}() under the > tracepoint's static key can fetch. > > a) creates redundant code, and b) is quite fragile due to requiring extra > care for "reentrant" functions (async SMP calls). > > This implements a). > > Signed-off-by: Valentin Schneider <vschneid@redhat.com> > --- > kernel/irq_work.c | 2 ++ > kernel/sched/core.c | 35 ++++++++++++++++++++++++----------- > kernel/sched/smp.h | 1 + > kernel/smp.c | 42 ++++++++++++++++++++++++++++++++++++++---- > 4 files changed, 65 insertions(+), 15 deletions(-) > > diff --git a/kernel/irq_work.c b/kernel/irq_work.c > index aec38c294ce68..fcfa75c4a5daf 100644 > --- 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); > diff --git a/kernel/sched/core.c b/kernel/sched/core.c > index 02181f8072b5f..41196ca67e913 100644 > --- a/kernel/sched/core.c > +++ b/kernel/sched/core.c > @@ -3743,17 +3743,30 @@ void sched_ttwu_pending(void *arg) > rq_unlock_irqrestore(rq, &rf); > } > > -void 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); > - } > -} > +/* > + * We want a variant that traces the function causing the IPI to be sent, but > + * we don't want the extra argument to cause unnecessary overhead when tracing > + * isn't happening. > + */ > +#define GEN_CFSI(suffix, IPI_EXP, ...) \ > +void send_call_function_single_ipi##suffix(__VA_ARGS__) \ > +{ \ > + struct rq *rq = cpu_rq(cpu); \ > + \ > + if (!set_nr_if_polling(rq->idle)) { \ > + IPI_EXP; \ > + arch_send_call_function_single_ipi(cpu); \ > + } else { \ > + trace_sched_wake_idle_without_ipi(cpu); \ > + } \ > +} > + > +GEN_CFSI(/* nop */, > + /* nop */, > + int cpu) > +GEN_CFSI(_trace, > + trace_ipi_send_cpumask(cpumask_of(cpu), _RET_IP_, func), > + int cpu, smp_call_func_t func) > *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); + 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); - 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); }
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); > }
On Thu, Nov 17, 2022 at 02:45:29PM +0000, Valentin Schneider wrote: > > + 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... Keeps em sharp ;-) > > + 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? Oh, right..
On 11/18/22 10:12, Peter Zijlstra wrote: > On Thu, Nov 17, 2022 at 02:45:29PM +0000, Valentin Schneider wrote: > >>> + 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... > Keeps em sharp ;-) > Having the trace before the IPI avoids the (non ideal) case where the trace stops because of an IPI execution before we have trace about who sent it... :-(. -- Daniel
diff --git a/kernel/irq_work.c b/kernel/irq_work.c index aec38c294ce68..fcfa75c4a5daf 100644 --- 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); diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 02181f8072b5f..41196ca67e913 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -3743,17 +3743,30 @@ void sched_ttwu_pending(void *arg) rq_unlock_irqrestore(rq, &rf); } -void 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); - } -} +/* + * We want a variant that traces the function causing the IPI to be sent, but + * we don't want the extra argument to cause unnecessary overhead when tracing + * isn't happening. + */ +#define GEN_CFSI(suffix, IPI_EXP, ...) \ +void send_call_function_single_ipi##suffix(__VA_ARGS__) \ +{ \ + struct rq *rq = cpu_rq(cpu); \ + \ + if (!set_nr_if_polling(rq->idle)) { \ + IPI_EXP; \ + arch_send_call_function_single_ipi(cpu); \ + } else { \ + trace_sched_wake_idle_without_ipi(cpu); \ + } \ +} + +GEN_CFSI(/* nop */, + /* nop */, + int cpu) +GEN_CFSI(_trace, + trace_ipi_send_cpumask(cpumask_of(cpu), _RET_IP_, func), + int cpu, smp_call_func_t func) /* * Queue a task on the target CPUs wake_list and wake the CPU via IPI if diff --git a/kernel/sched/smp.h b/kernel/sched/smp.h index 2eb23dd0f2856..8075ad5e84181 100644 --- a/kernel/sched/smp.h +++ b/kernel/sched/smp.h @@ -7,6 +7,7 @@ extern void sched_ttwu_pending(void *arg); extern void send_call_function_single_ipi(int cpu); +extern void send_call_function_single_ipi_trace(int cpu, smp_call_func_t func); #ifdef CONFIG_SMP extern void flush_smp_call_function_queue(void); diff --git a/kernel/smp.c b/kernel/smp.c index b96579fe08f09..3b8e6456ac7e7 100644 --- a/kernel/smp.c +++ b/kernel/smp.c @@ -161,12 +161,18 @@ void __init call_function_init(void) } static inline void -send_call_function_ipi_mask(const struct cpumask *mask) +send_call_function_ipi_mask_trace(const struct cpumask *mask, smp_call_func_t func) { trace_ipi_send_cpumask(mask, _RET_IP_, func); arch_send_call_function_ipi_mask(mask); } +static inline void +send_call_function_ipi_mask(const struct cpumask *mask) +{ + arch_send_call_function_ipi_mask(mask); +} + void smp_send_reschedule(int cpu) { /* XXX scheduler_ipi is inline :/ */ @@ -438,12 +444,16 @@ static void __smp_call_single_queue_debug(int cpu, struct llist_node *node) 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); - send_call_function_single_ipi(cpu); + send_call_function_single_ipi_trace(cpu, csd->func); cfd_seq_store(seq->pinged, this_cpu, cpu, CFD_SEQ_PINGED); } else { cfd_seq_store(pcpu->seq_noipi, this_cpu, cpu, CFD_SEQ_NOIPI); @@ -501,6 +511,26 @@ void __smp_call_single_queue(int cpu, struct llist_node *node) } } #endif + /* + * 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 = CSD_TYPE(csd) == CSD_TYPE_TTWU ? + sched_ttwu_pending : csd->func; + + /* See below comment on ordering requirements */ + if (llist_add(node, &per_cpu(call_single_queue, cpu))) + send_call_function_single_ipi_trace(cpu, func); + return; + } /* * The list addition should be visible to the target CPU when it pops @@ -984,9 +1014,13 @@ static void smp_call_function_many_cond(const struct cpumask *mask, * provided mask. */ if (nr_cpus == 1) - send_call_function_single_ipi(last_cpu); + trace_ipi_send_cpumask_enabled() ? + send_call_function_single_ipi_trace(last_cpu, func) : + send_call_function_single_ipi(last_cpu); else if (likely(nr_cpus > 1)) - send_call_function_ipi_mask(cfd->cpumask_ipi); + trace_ipi_send_cpumask_enabled() ? + send_call_function_ipi_mask_trace(mask, 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); }
The newly-introduced ipi_send_cpumask tracepoint has a "callback" parameter which so far has only been fed with NULL. While CSD_TYPE_SYNC/ASYNC and CSD_TYPE_IRQ_WORK share a similar backing struct layout (meaning their callback func can be accessed without caring about the actual CSD type), CSD_TYPE_TTWU doesn't even have a function attached to its struct. This means we need to check the type of a CSD before eventually dereferencing its associated callback. This isn't as trivial as it sounds: the CSD type is stored in __call_single_node.u_flags, which get cleared right before the callback is executed via csd_unlock(). This implies checking the CSD type before it is enqueued on the call_single_queue, as the target CPU's queue can be flushed before we get to sending an IPI. Furthermore, send_call_function_single_ipi() only has a CPU parameter, and would need to have an additional argument to trickle down the invoked function. This is somewhat silly, as the extra argument will always be pushed down to the function even when nothing is being traced, which is unnecessary overhead. Two options present themselves: a) Create copies of send_call_function_{single_ipi, ipi_mask}() that take an extra argument used for tracing, so that codepaths remain unchanged when tracing isn't in effect (a sort of manual -fipa-sra). b) Stash the CSD func in somewhere as a side effect that the portion of send_call_function_{single_ipi, ipi_mask}() under the tracepoint's static key can fetch. a) creates redundant code, and b) is quite fragile due to requiring extra care for "reentrant" functions (async SMP calls). This implements a). Signed-off-by: Valentin Schneider <vschneid@redhat.com> --- kernel/irq_work.c | 2 ++ kernel/sched/core.c | 35 ++++++++++++++++++++++++----------- kernel/sched/smp.h | 1 + kernel/smp.c | 42 ++++++++++++++++++++++++++++++++++++++---- 4 files changed, 65 insertions(+), 15 deletions(-)