[PATCH] perf/core: only check cpuctx with cgroup events during cgroup switch

[PATCH] perf/core: only check cpuctx with cgroup events during cgroup switch

[David Carrillo-Cisneros]

Following conversation in CQM/CMT last series about speeding up the
context switch for cgroup events:

https://patchwork.kernel.org/patch/9478617/

This is a low-hanging fruit optimization. It replaces the iteration over
the "pmus" list in cgroup switch by an iteration over a new list that
contains only cpuctxs with at least one cgroup event.

The iteration over pmus has become increasingly costly as the number of
pmus have increased. It is also heavily affected by cache contention.

Below some instrumentation measurements with 10, 50 and 90 percentiles
of the total cost of context switch at distinct levels of cache
contenton for a simple array read/write microbenchark.

Contention
   Level    Nr events      Before (us)            After (us)       Median
 L2    L3     types      (10%, 50%, 90%)       (10%, 50%, 90%     Speedup
--------------------------------------------------------------------------
Low   Low       1       (1.72, 2.42, 5.85)    (1.35, 1.64, 5.46)     29%
High  Low       1       (2.08, 4.56, 19.8)    (1720, 2.20, 13.7)     51%
High  High      1       (2.86, 10.4, 12.7)    (2.54, 4.32, 12.1)     58%

Low   Low       2       (1.98, 3.20, 6.89)    (1.68, 2.41, 8.89)     24%
High  Low       2       (2.48, 5.28, 22.4)    (2150, 3.69, 14.6)     30%
High  High      2       (3.32, 8.09, 13.9)    (2.80, 5.15, 13.7)     36%

where:
 1 event type  = cycles
 2 event types = cycles,intel_cqm/llc_occupancy/

Contetion L2 Low: workset < L2 cache size.
             High:  "    >> L2   "     " .
Contetion L3 Low: workset of task on all sockets < L3 cache size.
             High:   "     "   "   "   "    "   >> L3   "     " .

Median Speed up is (50%ile Before - 50%ile After) /  50%ile Before

Signed-off-by: David Carrillo-Cisneros <davidcc@google.com>
---
 include/linux/perf_event.h |  1 +
 kernel/events/core.c       | 96 +++++++++++++++++++++-------------------------
 2 files changed, 44 insertions(+), 53 deletions(-)

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index 4741ecd..10d2235 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -788,6 +788,7 @@ struct perf_cpu_context {
 	struct pmu			*unique_pmu;
 #ifdef CONFIG_CGROUP_PERF
 	struct perf_cgroup		*cgrp;
+	struct list_head		cgrp_cpuctx_entry;
 #endif
 
 	struct list_head		sched_cb_entry;
diff --git a/kernel/events/core.c b/kernel/events/core.c
index faf073d..35c4fbc 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -678,6 +678,8 @@ perf_cgroup_set_timestamp(struct task_struct *task,
 	info->timestamp = ctx->timestamp;
 }
 
+static DEFINE_PER_CPU(struct list_head, cgrp_cpuctx_list);
+
 #define PERF_CGROUP_SWOUT	0x1 /* cgroup switch out every event */
 #define PERF_CGROUP_SWIN	0x2 /* cgroup switch in events based on task */
 
@@ -690,61 +692,46 @@ perf_cgroup_set_timestamp(struct task_struct *task,
 static void perf_cgroup_switch(struct task_struct *task, int mode)
 {
 	struct perf_cpu_context *cpuctx;
-	struct pmu *pmu;
+	struct list_head *list;
 	unsigned long flags;
 
 	/*
-	 * disable interrupts to avoid geting nr_cgroup
-	 * changes via __perf_event_disable(). Also
-	 * avoids preemption.
+	 * Disable interrupts and preemption to avoid this CPU's
+	 * cgrp_cpuctx_entry to change under us.
 	 */
 	local_irq_save(flags);
 
-	/*
-	 * we reschedule only in the presence of cgroup
-	 * constrained events.
-	 */
+	list = this_cpu_ptr(&cgrp_cpuctx_list);
+	list_for_each_entry(cpuctx, list, cgrp_cpuctx_entry) {
+		WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0);
 
-	list_for_each_entry_rcu(pmu, &pmus, entry) {
-		cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
-		if (cpuctx->unique_pmu != pmu)
-			continue; /* ensure we process each cpuctx once */
+		perf_ctx_lock(cpuctx, cpuctx->task_ctx);
+		perf_pmu_disable(cpuctx->ctx.pmu);
 
-		/*
-		 * perf_cgroup_events says at least one
-		 * context on this CPU has cgroup events.
-		 *
-		 * ctx->nr_cgroups reports the number of cgroup
-		 * events for a context.
-		 */
-		if (cpuctx->ctx.nr_cgroups > 0) {
-			perf_ctx_lock(cpuctx, cpuctx->task_ctx);
-			perf_pmu_disable(cpuctx->ctx.pmu);
-
-			if (mode & PERF_CGROUP_SWOUT) {
-				cpu_ctx_sched_out(cpuctx, EVENT_ALL);
-				/*
-				 * must not be done before ctxswout due
-				 * to event_filter_match() in event_sched_out()
-				 */
-				cpuctx->cgrp = NULL;
-			}
+		if (mode & PERF_CGROUP_SWOUT) {
+			cpu_ctx_sched_out(cpuctx, EVENT_ALL);
+			/*
+			 * must not be done before ctxswout due
+			 * to event_filter_match() in event_sched_out()
+			 */
+			cpuctx->cgrp = NULL;
+		}
 
-			if (mode & PERF_CGROUP_SWIN) {
-				WARN_ON_ONCE(cpuctx->cgrp);
-				/*
-				 * set cgrp before ctxsw in to allow
-				 * event_filter_match() to not have to pass
-				 * task around
-				 * we pass the cpuctx->ctx to perf_cgroup_from_task()
-				 * because cgorup events are only per-cpu
-				 */
-				cpuctx->cgrp = perf_cgroup_from_task(task, &cpuctx->ctx);
-				cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
-			}
-			perf_pmu_enable(cpuctx->ctx.pmu);
-			perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
+		if (mode & PERF_CGROUP_SWIN) {
+			WARN_ON_ONCE(cpuctx->cgrp);
+			/*
+			 * set cgrp before ctxsw in to allow
+			 * event_filter_match() to not have to pass
+			 * task around
+			 * we pass the cpuctx->ctx to perf_cgroup_from_task()
+			 * because cgorup events are only per-cpu
+			 */
+			cpuctx->cgrp = perf_cgroup_from_task(task,
+							     &cpuctx->ctx);
+			cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
 		}
+		perf_pmu_enable(cpuctx->ctx.pmu);
+		perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
 	}
 
 	local_irq_restore(flags);
@@ -889,6 +876,7 @@ list_update_cgroup_event(struct perf_event *event,
 			 struct perf_event_context *ctx, bool add)
 {
 	struct perf_cpu_context *cpuctx;
+	struct list_head *lentry;
 
 	if (!is_cgroup_event(event))
 		return;
@@ -902,15 +890,16 @@ list_update_cgroup_event(struct perf_event *event,
 	 * this will always be called from the right CPU.
 	 */
 	cpuctx = __get_cpu_context(ctx);
-
-	/*
-	 * cpuctx->cgrp is NULL until a cgroup event is sched in or
-	 * ctx->nr_cgroup == 0 .
-	 */
-	if (add && perf_cgroup_from_task(current, ctx) == event->cgrp)
-		cpuctx->cgrp = event->cgrp;
-	else if (!add)
+	lentry = &cpuctx->cgrp_cpuctx_entry;
+	/* cpuctx->cgrp is NULL unless a cgroup event is active in this CPU .*/
+	if (add) {
+		list_add(lentry, this_cpu_ptr(&cgrp_cpuctx_list));
+		if (perf_cgroup_from_task(current, ctx) == event->cgrp)
+			cpuctx->cgrp = event->cgrp;
+	} else {
+		list_del(lentry);
 		cpuctx->cgrp = NULL;
+	}
 }
 
 #else /* !CONFIG_CGROUP_PERF */
@@ -10595,6 +10584,7 @@ static void __init perf_event_init_all_cpus(void)
 		INIT_LIST_HEAD(&per_cpu(pmu_sb_events.list, cpu));
 		raw_spin_lock_init(&per_cpu(pmu_sb_events.lock, cpu));
 
+		INIT_LIST_HEAD(&per_cpu(cgrp_cpuctx_list, cpu));
 		INIT_LIST_HEAD(&per_cpu(sched_cb_list, cpu));
 	}
 }
-- 
2.8.0.rc3.226.g39d4020

Re: [PATCH] perf/core: only check cpuctx with cgroup events during cgroup switch

[kbuild test robot]

[-- Attachment #1: Type: text/plain, Size: 3646 bytes --]

Hi David,

[auto build test ERROR on tip/perf/core]
[also build test ERROR on v4.10-rc1 next-20161224]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system]

url:    https://github.com/0day-ci/linux/commits/David-Carrillo-Cisneros/perf-core-only-check-cpuctx-with-cgroup-events-during-cgroup-switch/20161231-103433
config: i386-randconfig-x005-201652 (attached as .config)
compiler: gcc-6 (Debian 6.2.0-3) 6.2.0 20160901
reproduce:
        # save the attached .config to linux build tree
        make ARCH=i386 

All error/warnings (new ones prefixed by >>):

   In file included from include/asm-generic/percpu.h:6:0,
                    from arch/x86/include/asm/percpu.h:542,
                    from arch/x86/include/asm/preempt.h:5,
                    from include/linux/preempt.h:59,
                    from include/linux/spinlock.h:50,
                    from include/linux/wait.h:8,
                    from include/linux/fs.h:5,
                    from kernel/events/core.c:12:
   kernel/events/core.c: In function 'perf_event_init_all_cpus':
>> kernel/events/core.c:10587:27: error: 'cgrp_cpuctx_list' undeclared (first use in this function)
      INIT_LIST_HEAD(&per_cpu(cgrp_cpuctx_list, cpu));
                              ^
   include/linux/percpu-defs.h:206:47: note: in definition of macro '__verify_pcpu_ptr'
     const void __percpu *__vpp_verify = (typeof((ptr) + 0))NULL; \
                                                  ^~~
   include/linux/percpu-defs.h:250:47: note: in expansion of macro 'VERIFY_PERCPU_PTR'
    #define per_cpu_ptr(ptr, cpu) ({ (void)(cpu); VERIFY_PERCPU_PTR(ptr); })
                                                  ^~~~~~~~~~~~~~~~~
   include/linux/percpu-defs.h:256:29: note: in expansion of macro 'per_cpu_ptr'
    #define per_cpu(var, cpu) (*per_cpu_ptr(&(var), cpu))
                                ^~~~~~~~~~~
>> kernel/events/core.c:10587:19: note: in expansion of macro 'per_cpu'
      INIT_LIST_HEAD(&per_cpu(cgrp_cpuctx_list, cpu));
                      ^~~~~~~
   kernel/events/core.c:10587:27: note: each undeclared identifier is reported only once for each function it appears in
      INIT_LIST_HEAD(&per_cpu(cgrp_cpuctx_list, cpu));
                              ^
   include/linux/percpu-defs.h:206:47: note: in definition of macro '__verify_pcpu_ptr'
     const void __percpu *__vpp_verify = (typeof((ptr) + 0))NULL; \
                                                  ^~~
   include/linux/percpu-defs.h:250:47: note: in expansion of macro 'VERIFY_PERCPU_PTR'
    #define per_cpu_ptr(ptr, cpu) ({ (void)(cpu); VERIFY_PERCPU_PTR(ptr); })
                                                  ^~~~~~~~~~~~~~~~~
   include/linux/percpu-defs.h:256:29: note: in expansion of macro 'per_cpu_ptr'
    #define per_cpu(var, cpu) (*per_cpu_ptr(&(var), cpu))
                                ^~~~~~~~~~~
>> kernel/events/core.c:10587:19: note: in expansion of macro 'per_cpu'
      INIT_LIST_HEAD(&per_cpu(cgrp_cpuctx_list, cpu));
                      ^~~~~~~

vim +/cgrp_cpuctx_list +10587 kernel/events/core.c

 10581			mutex_init(&swhash->hlist_mutex);
 10582			INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu));
 10583	
 10584			INIT_LIST_HEAD(&per_cpu(pmu_sb_events.list, cpu));
 10585			raw_spin_lock_init(&per_cpu(pmu_sb_events.lock, cpu));
 10586	
 10587			INIT_LIST_HEAD(&per_cpu(cgrp_cpuctx_list, cpu));
 10588			INIT_LIST_HEAD(&per_cpu(sched_cb_list, cpu));
 10589		}
 10590	}

---
0-DAY kernel test infrastructure                Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all                   Intel Corporation

[-- Attachment #2: .config.gz --]
[-- Type: application/gzip, Size: 24329 bytes --]

Re: [PATCH] perf/core: only check cpuctx with cgroup events during cgroup switch

[kbuild test robot]

[-- Attachment #1: Type: text/plain, Size: 1760 bytes --]

Hi David,

[auto build test ERROR on tip/perf/core]
[also build test ERROR on v4.10-rc1 next-20161224]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system]

url:    https://github.com/0day-ci/linux/commits/David-Carrillo-Cisneros/perf-core-only-check-cpuctx-with-cgroup-events-during-cgroup-switch/20161231-103433
config: x86_64-randconfig-s1-12311108 (attached as .config)
compiler: gcc-4.4 (Debian 4.4.7-8) 4.4.7
reproduce:
        # save the attached .config to linux build tree
        make ARCH=x86_64 

All error/warnings (new ones prefixed by >>):

   kernel/events/core.c: In function 'perf_event_init_all_cpus':
   kernel/events/core.c:10587: error: 'cgrp_cpuctx_list' undeclared (first use in this function)
   kernel/events/core.c:10587: error: (Each undeclared identifier is reported only once
   kernel/events/core.c:10587: error: for each function it appears in.)
>> kernel/events/core.c:10587: error: type defaults to 'int' in declaration of 'type name'
>> kernel/events/core.c:10587: warning: cast from pointer to integer of different size
>> kernel/events/core.c:10587: error: type defaults to 'int' in declaration of 'type name'

vim +10587 kernel/events/core.c

 10581			mutex_init(&swhash->hlist_mutex);
 10582			INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu));
 10583	
 10584			INIT_LIST_HEAD(&per_cpu(pmu_sb_events.list, cpu));
 10585			raw_spin_lock_init(&per_cpu(pmu_sb_events.lock, cpu));
 10586	
 10587			INIT_LIST_HEAD(&per_cpu(cgrp_cpuctx_list, cpu));
 10588			INIT_LIST_HEAD(&per_cpu(sched_cb_list, cpu));
 10589		}
 10590	}

---
0-DAY kernel test infrastructure                Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all                   Intel Corporation

[-- Attachment #2: .config.gz --]
[-- Type: application/gzip, Size: 23819 bytes --]