From 642c2d671ceff40e9453203ea0c66e991e11e249 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Mon, 30 Nov 2015 12:56:15 +0100 Subject: [PATCH 1/4] perf: Fix PERF_EVENT_IOC_PERIOD deadlock Dmitry reported a fairly silly recursive lock deadlock for PERF_EVENT_IOC_PERIOD, fix this by explicitly doing the inactive part of __perf_event_period() instead of calling that function. Reported-by: Dmitry Vyukov Signed-off-by: Peter Zijlstra (Intel) Cc: Cc: Alexander Potapenko Cc: Arnaldo Carvalho de Melo Cc: Arnaldo Carvalho de Melo Cc: Eric Dumazet Cc: Jiri Olsa Cc: Kostya Serebryany Cc: Linus Torvalds Cc: Peter Zijlstra Cc: Sasha Levin Cc: Stephane Eranian Cc: Thomas Gleixner Cc: Vince Weaver Fixes: c7999c6f3fed ("perf: Fix PERF_EVENT_IOC_PERIOD migration race") Link: http://lkml.kernel.org/r/20151130115615.GJ17308@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar --- kernel/events/core.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/kernel/events/core.c b/kernel/events/core.c index 5854fcf7f05a..49a5118f3564 100644 --- a/kernel/events/core.c +++ b/kernel/events/core.c @@ -4225,7 +4225,14 @@ retry: goto retry; } - __perf_event_period(&pe); + if (event->attr.freq) { + event->attr.sample_freq = value; + } else { + event->attr.sample_period = value; + event->hw.sample_period = value; + } + + local64_set(&event->hw.period_left, 0); raw_spin_unlock_irq(&ctx->lock); return 0; From e0fbac1cd49c5ca6cb6a0aa918d0943ce95d906c Mon Sep 17 00:00:00 2001 From: Yuanfang Chen Date: Tue, 24 Nov 2015 12:05:01 -0500 Subject: [PATCH 2/4] perf/x86/intel: Make L1D_PEND_MISS.FB_FULL not constrained on Haswell There was a mistake in the Haswell constraints table. Signed-off-by: Yuanfang Chen Signed-off-by: Peter Zijlstra (Intel) Reviewed-by: Andi Kleen Cc: Arnaldo Carvalho de Melo Cc: Jiri Olsa Cc: Linus Torvalds Cc: Peter Zijlstra Cc: Stephane Eranian Cc: Thomas Gleixner Cc: Vince Weaver Link: http://lkml.kernel.org/r/1448384701-9110-1-git-send-email-cheny@udel.edu Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_event_intel.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/arch/x86/kernel/cpu/perf_event_intel.c b/arch/x86/kernel/cpu/perf_event_intel.c index f63360be2238..e2a430021e46 100644 --- a/arch/x86/kernel/cpu/perf_event_intel.c +++ b/arch/x86/kernel/cpu/perf_event_intel.c @@ -232,7 +232,7 @@ static struct event_constraint intel_hsw_event_constraints[] = { FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ - INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.* */ + INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */ INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ From 169b932a15318e8e9f2f0f12eeb55dda09c8737f Mon Sep 17 00:00:00 2001 From: Jiri Olsa Date: Mon, 9 Nov 2015 10:24:31 +0100 Subject: [PATCH 3/4] perf/x86/intel: Fix INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA macro We need to add rest of the flags to the constraint mask instead of another INTEL_ARCH_EVENT_MASK, fixing a typo. Signed-off-by: Jiri Olsa Signed-off-by: Peter Zijlstra (Intel) Cc: Arnaldo Carvalho de Melo Cc: Jiri Olsa Cc: Linus Torvalds Cc: Paul Mackerras Cc: Peter Zijlstra Cc: Stephane Eranian Cc: Thomas Gleixner Cc: Vince Weaver Link: http://lkml.kernel.org/r/1447061071-28085-1-git-send-email-jolsa@kernel.org Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_event.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/arch/x86/kernel/cpu/perf_event.h b/arch/x86/kernel/cpu/perf_event.h index ab18b8a91583..d0e35ebb2adb 100644 --- a/arch/x86/kernel/cpu/perf_event.h +++ b/arch/x86/kernel/cpu/perf_event.h @@ -387,7 +387,7 @@ struct cpu_hw_events { /* Check flags and event code/umask, and set the HSW N/A flag */ #define INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(code, n) \ __EVENT_CONSTRAINT(code, n, \ - INTEL_ARCH_EVENT_MASK|INTEL_ARCH_EVENT_MASK, \ + INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \ HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_NA_HSW) From 4e93ad601a4308d4a67673c81556580817d56940 Mon Sep 17 00:00:00 2001 From: Jiri Olsa Date: Wed, 4 Nov 2015 16:00:05 +0100 Subject: [PATCH 4/4] perf: Do not send exit event twice In case we monitor events system wide, we get EXIT event (when configured) twice for each task that exited. Note doubled lines with same pid/tid in following example: $ sudo ./perf record -a ^C[ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.480 MB perf.data (2518 samples) ] $ sudo ./perf report -D | grep EXIT 0 60290687567581 0x59910 [0x38]: PERF_RECORD_EXIT(1250:1250):(1250:1250) 0 60290687568354 0x59948 [0x38]: PERF_RECORD_EXIT(1250:1250):(1250:1250) 0 60290687988744 0x59ad8 [0x38]: PERF_RECORD_EXIT(1250:1250):(1250:1250) 0 60290687989198 0x59b10 [0x38]: PERF_RECORD_EXIT(1250:1250):(1250:1250) 1 60290692567895 0x62af0 [0x38]: PERF_RECORD_EXIT(1253:1253):(1253:1253) 1 60290692568322 0x62b28 [0x38]: PERF_RECORD_EXIT(1253:1253):(1253:1253) 2 60290692739276 0x69a18 [0x38]: PERF_RECORD_EXIT(1252:1252):(1252:1252) 2 60290692739910 0x69a50 [0x38]: PERF_RECORD_EXIT(1252:1252):(1252:1252) The reason is that the cpu contexts are processes each time we call perf_event_task. I'm changing the perf_event_aux logic to serve task_ctx and cpu contexts separately, which ensure we don't get EXIT event generated twice on same cpu context. This does not affect other auxiliary events, as they don't use task_ctx at all. Signed-off-by: Jiri Olsa Signed-off-by: Peter Zijlstra (Intel) Cc: Arnaldo Carvalho de Melo Cc: Arnaldo Carvalho de Melo Cc: David Ahern Cc: Jiri Olsa Cc: Linus Torvalds Cc: Namhyung Kim Cc: Peter Zijlstra Cc: Stephane Eranian Cc: Thomas Gleixner Cc: Vince Weaver Link: http://lkml.kernel.org/r/1446649205-5822-1-git-send-email-jolsa@kernel.org Signed-off-by: Ingo Molnar --- kernel/events/core.c | 42 +++++++++++++++++++++++++++++++----------- 1 file changed, 31 insertions(+), 11 deletions(-) diff --git a/kernel/events/core.c b/kernel/events/core.c index 49a5118f3564..39cf4a40aa4c 100644 --- a/kernel/events/core.c +++ b/kernel/events/core.c @@ -5682,6 +5682,17 @@ perf_event_aux_ctx(struct perf_event_context *ctx, } } +static void +perf_event_aux_task_ctx(perf_event_aux_output_cb output, void *data, + struct perf_event_context *task_ctx) +{ + rcu_read_lock(); + preempt_disable(); + perf_event_aux_ctx(task_ctx, output, data); + preempt_enable(); + rcu_read_unlock(); +} + static void perf_event_aux(perf_event_aux_output_cb output, void *data, struct perf_event_context *task_ctx) @@ -5691,14 +5702,23 @@ perf_event_aux(perf_event_aux_output_cb output, void *data, struct pmu *pmu; int ctxn; + /* + * If we have task_ctx != NULL we only notify + * the task context itself. The task_ctx is set + * only for EXIT events before releasing task + * context. + */ + if (task_ctx) { + perf_event_aux_task_ctx(output, data, task_ctx); + return; + } + rcu_read_lock(); list_for_each_entry_rcu(pmu, &pmus, entry) { cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); if (cpuctx->unique_pmu != pmu) goto next; perf_event_aux_ctx(&cpuctx->ctx, output, data); - if (task_ctx) - goto next; ctxn = pmu->task_ctx_nr; if (ctxn < 0) goto next; @@ -5708,12 +5728,6 @@ perf_event_aux(perf_event_aux_output_cb output, void *data, next: put_cpu_ptr(pmu->pmu_cpu_context); } - - if (task_ctx) { - preempt_disable(); - perf_event_aux_ctx(task_ctx, output, data); - preempt_enable(); - } rcu_read_unlock(); } @@ -8803,10 +8817,8 @@ static void perf_event_exit_task_context(struct task_struct *child, int ctxn) struct perf_event_context *child_ctx, *clone_ctx = NULL; unsigned long flags; - if (likely(!child->perf_event_ctxp[ctxn])) { - perf_event_task(child, NULL, 0); + if (likely(!child->perf_event_ctxp[ctxn])) return; - } local_irq_save(flags); /* @@ -8890,6 +8902,14 @@ void perf_event_exit_task(struct task_struct *child) for_each_task_context_nr(ctxn) perf_event_exit_task_context(child, ctxn); + + /* + * The perf_event_exit_task_context calls perf_event_task + * with child's task_ctx, which generates EXIT events for + * child contexts and sets child->perf_event_ctxp[] to NULL. + * At this point we need to send EXIT events to cpu contexts. + */ + perf_event_task(child, NULL, 0); } static void perf_free_event(struct perf_event *event,