[PATCH v3] perf: fix RCU issues with cgroup monitoring mode

[PATCH v3] perf: fix RCU issues with cgroup monitoring mode

[Stephane Eranian]

This patch eliminates RCU violations detected by the RCU
checker (PROVE_RCU). The impact code paths were all related
to cgroup mode monitoring and involved access a task's cgrp.

V2 is updated to include comments from PeterZ to eliminate
some of the warnings without grabbing the rcu_read lock because
we know we are already holding th ctx->lock which prevents
the cgroup from disappearing while we are accessing it.
The trick, as suggested by Peter, is to modify the
perf_cgroup_from_task() to take an extra boolean parameter
to allow bypassing the lockdep test in the task_subsys_cstate()
macros. This patch uses this approach to update all calls the
perf_cgroup_from_task().

In V3, we change the boolean parameter for a pointer to a
perf_event_context so we can check the ctx->lock explicitely.
This is more robust, than passing the boolean to express that
we know the lock is held. The code can change, and thus the
locking assumption, checking lockdep_is_held() ensures,
the proper locking is in place. Patch relative to tip.git
at commit 57ef9fc.

Signed-off-by: Stephane Eranian <eranian@google.com>
---
 arch/x86/kernel/cpu/perf_event_intel_cqm.c |  2 +-
 include/linux/perf_event.h                 |  5 +++--
 kernel/events/core.c                       | 25 +++++++++++++++----------
 3 files changed, 19 insertions(+), 13 deletions(-)

diff --git a/arch/x86/kernel/cpu/perf_event_intel_cqm.c b/arch/x86/kernel/cpu/perf_event_intel_cqm.c
index 377e8f8..a316ca9 100644
--- a/arch/x86/kernel/cpu/perf_event_intel_cqm.c
+++ b/arch/x86/kernel/cpu/perf_event_intel_cqm.c
@@ -298,7 +298,7 @@ static bool __match_event(struct perf_event *a, struct perf_event *b)
 static inline struct perf_cgroup *event_to_cgroup(struct perf_event *event)
 {
 	if (event->attach_state & PERF_ATTACH_TASK)
-		return perf_cgroup_from_task(event->hw.target);
+		return perf_cgroup_from_task(event->hw.target, event->ctx);
 
 	return event->cgrp;
 }
diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index d841d33..94107e4 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -697,9 +697,10 @@ struct perf_cgroup {
  * if there is no cgroup event for the current CPU context.
  */
 static inline struct perf_cgroup *
-perf_cgroup_from_task(struct task_struct *task)
+perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
 {
-	return container_of(task_css(task, perf_event_cgrp_id),
+	bool safe = ctx ? lockdep_is_held(&ctx->lock) : true;
+	return container_of(task_css_check(task, perf_event_cgrp_id, safe),
 			    struct perf_cgroup, css);
 }
 #endif /* CONFIG_CGROUP_PERF */
diff --git a/kernel/events/core.c b/kernel/events/core.c
index ea02109..f611246 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -435,7 +435,7 @@ static inline void update_cgrp_time_from_event(struct perf_event *event)
 	if (!is_cgroup_event(event))
 		return;
 
-	cgrp = perf_cgroup_from_task(current);
+	cgrp = perf_cgroup_from_task(current, event->ctx);
 	/*
 	 * Do not update time when cgroup is not active
 	 */
@@ -458,7 +458,7 @@ perf_cgroup_set_timestamp(struct task_struct *task,
 	if (!task || !ctx->nr_cgroups)
 		return;
 
-	cgrp = perf_cgroup_from_task(task);
+	cgrp = perf_cgroup_from_task(task, ctx);
 	info = this_cpu_ptr(cgrp->info);
 	info->timestamp = ctx->timestamp;
 }
@@ -489,7 +489,6 @@ static void perf_cgroup_switch(struct task_struct *task, int mode)
 	 * we reschedule only in the presence of cgroup
 	 * constrained events.
 	 */
-	rcu_read_lock();
 
 	list_for_each_entry_rcu(pmu, &pmus, entry) {
 		cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
@@ -523,7 +522,7 @@ static void perf_cgroup_switch(struct task_struct *task, int mode)
 				 * event_filter_match() to not have to pass
 				 * task around
 				 */
-				cpuctx->cgrp = perf_cgroup_from_task(task);
+				cpuctx->cgrp = perf_cgroup_from_task(task, NULL);
 				cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
 			}
 			perf_pmu_enable(cpuctx->ctx.pmu);
@@ -531,8 +530,6 @@ static void perf_cgroup_switch(struct task_struct *task, int mode)
 		}
 	}
 
-	rcu_read_unlock();
-
 	local_irq_restore(flags);
 }
 
@@ -542,17 +539,18 @@ static inline void perf_cgroup_sched_out(struct task_struct *task,
 	struct perf_cgroup *cgrp1;
 	struct perf_cgroup *cgrp2 = NULL;
 
+	rcu_read_lock();
 	/*
 	 * we come here when we know perf_cgroup_events > 0
 	 */
-	cgrp1 = perf_cgroup_from_task(task);
+	cgrp1 = perf_cgroup_from_task(task, NULL);
 
 	/*
 	 * next is NULL when called from perf_event_enable_on_exec()
 	 * that will systematically cause a cgroup_switch()
 	 */
 	if (next)
-		cgrp2 = perf_cgroup_from_task(next);
+		cgrp2 = perf_cgroup_from_task(next, NULL);
 
 	/*
 	 * only schedule out current cgroup events if we know
@@ -561,6 +559,8 @@ static inline void perf_cgroup_sched_out(struct task_struct *task,
 	 */
 	if (cgrp1 != cgrp2)
 		perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
+
+	rcu_read_unlock();
 }
 
 static inline void perf_cgroup_sched_in(struct task_struct *prev,
@@ -569,13 +569,14 @@ static inline void perf_cgroup_sched_in(struct task_struct *prev,
 	struct perf_cgroup *cgrp1;
 	struct perf_cgroup *cgrp2 = NULL;
 
+	rcu_read_lock();
 	/*
 	 * we come here when we know perf_cgroup_events > 0
 	 */
-	cgrp1 = perf_cgroup_from_task(task);
+	cgrp1 = perf_cgroup_from_task(task, NULL);
 
 	/* prev can never be NULL */
-	cgrp2 = perf_cgroup_from_task(prev);
+	cgrp2 = perf_cgroup_from_task(prev, NULL);
 
 	/*
 	 * only need to schedule in cgroup events if we are changing
@@ -584,6 +585,8 @@ static inline void perf_cgroup_sched_in(struct task_struct *prev,
 	 */
 	if (cgrp1 != cgrp2)
 		perf_cgroup_switch(task, PERF_CGROUP_SWIN);
+
+	rcu_read_unlock();
 }
 
 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
@@ -9442,7 +9445,9 @@ static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
 static int __perf_cgroup_move(void *info)
 {
 	struct task_struct *task = info;
+	rcu_read_lock();
 	perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
+	rcu_read_unlock();
 	return 0;
 }
 
-- 
2.5.0

Re: [PATCH v3] perf: fix RCU issues with cgroup monitoring mode

[kbuild test robot]

[-- Attachment #1: Type: text/plain, Size: 1955 bytes --]

Hi Stephane,

[auto build test ERROR on: tip/perf/core]
[also build test ERROR on: v4.3 next-20151103]

url:    https://github.com/0day-ci/linux/commits/Stephane-Eranian/perf-fix-RCU-issues-with-cgroup-monitoring-mode/20151104-121512
config: parisc-allyesconfig (attached as .config)
reproduce:
        wget https://git.kernel.org/cgit/linux/kernel/git/wfg/lkp-tests.git/plain/sbin/make.cross -O ~/bin/make.cross
        chmod +x ~/bin/make.cross
        # save the attached .config to linux build tree
        make.cross ARCH=parisc 

All errors (new ones prefixed by >>):

   In file included from include/linux/trace_events.h:9:0,
                    from include/trace/syscall.h:6,
                    from include/linux/syscalls.h:81,
                    from kernel/events/core.c:34:
   include/linux/perf_event.h: In function 'perf_cgroup_from_task':
>> include/linux/perf_event.h:702:2: error: implicit declaration of function 'lockdep_is_held' [-Werror=implicit-function-declaration]
     bool safe = ctx ? lockdep_is_held(&ctx->lock) : true;
     ^
   include/linux/perf_event.h:702:7: warning: unused variable 'safe' [-Wunused-variable]
     bool safe = ctx ? lockdep_is_held(&ctx->lock) : true;
          ^
   cc1: some warnings being treated as errors

vim +/lockdep_is_held +702 include/linux/perf_event.h

   696	 * this function. In other words, we cannot call this function
   697	 * if there is no cgroup event for the current CPU context.
   698	 */
   699	static inline struct perf_cgroup *
   700	perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
   701	{
 > 702		bool safe = ctx ? lockdep_is_held(&ctx->lock) : true;
   703		return container_of(task_css_check(task, perf_event_cgrp_id, safe),
   704				    struct perf_cgroup, css);
   705	}

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

[-- Attachment #2: .config.gz --]
[-- Type: application/octet-stream, Size: 42341 bytes --]

Re: [PATCH v3] perf: fix RCU issues with cgroup monitoring mode

[kbuild test robot]

[-- Attachment #1: Type: text/plain, Size: 1957 bytes --]

Hi Stephane,

[auto build test WARNING on: tip/perf/core]
[also build test WARNING on: v4.3 next-20151103]

url:    https://github.com/0day-ci/linux/commits/Stephane-Eranian/perf-fix-RCU-issues-with-cgroup-monitoring-mode/20151104-121512
config: i386-randconfig-i0-201544 (attached as .config)
reproduce:
        # save the attached .config to linux build tree
        make ARCH=i386 

All warnings (new ones prefixed by >>):

   In file included from include/linux/trace_events.h:9:0,
                    from include/trace/syscall.h:6,
                    from include/linux/syscalls.h:81,
                    from init/main.c:18:
   include/linux/perf_event.h: In function 'perf_cgroup_from_task':
>> include/linux/perf_event.h:702:7: warning: unused variable 'safe' [-Wunused-variable]
     bool safe = ctx ? lockdep_is_held(&ctx->lock) : true;
          ^

vim +/safe +702 include/linux/perf_event.h

   686		u64				timestamp;
   687	};
   688	
   689	struct perf_cgroup {
   690		struct cgroup_subsys_state	css;
   691		struct perf_cgroup_info	__percpu *info;
   692	};
   693	
   694	/*
   695	 * Must ensure cgroup is pinned (css_get) before calling
   696	 * this function. In other words, we cannot call this function
   697	 * if there is no cgroup event for the current CPU context.
   698	 */
   699	static inline struct perf_cgroup *
   700	perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
   701	{
 > 702		bool safe = ctx ? lockdep_is_held(&ctx->lock) : true;
   703		return container_of(task_css_check(task, perf_event_cgrp_id, safe),
   704				    struct perf_cgroup, css);
   705	}
   706	#endif /* CONFIG_CGROUP_PERF */
   707	
   708	#ifdef CONFIG_PERF_EVENTS
   709	
   710	extern void *perf_aux_output_begin(struct perf_output_handle *handle,

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

[-- Attachment #2: .config.gz --]
[-- Type: application/octet-stream, Size: 22679 bytes --]

Re: [PATCH v3] perf: fix RCU issues with cgroup monitoring mode

[Peter Zijlstra]

On Wed, Nov 04, 2015 at 05:12:19AM +0100, Stephane Eranian wrote:
> 
> This patch eliminates RCU violations detected by the RCU
> checker (PROVE_RCU). The impact code paths were all related
> to cgroup mode monitoring and involved access a task's cgrp.
> 
> V2 is updated to include comments from PeterZ to eliminate
> some of the warnings without grabbing the rcu_read lock because
> we know we are already holding th ctx->lock which prevents
> the cgroup from disappearing while we are accessing it.
> The trick, as suggested by Peter, is to modify the
> perf_cgroup_from_task() to take an extra boolean parameter
> to allow bypassing the lockdep test in the task_subsys_cstate()
> macros. This patch uses this approach to update all calls the
> perf_cgroup_from_task().
> 
> In V3, we change the boolean parameter for a pointer to a
> perf_event_context so we can check the ctx->lock explicitely.
> This is more robust, than passing the boolean to express that
> we know the lock is held. The code can change, and thus the
> locking assumption, checking lockdep_is_held() ensures,
> the proper locking is in place. Patch relative to tip.git
> at commit 57ef9fc.

So aside from the reported build fails; this is not suitable Changelog.

Also, please split it at least two patches; as there are at least the
two distinct issues here.

One is the perf_cgroup_sched_{in,out} thing, which requires moving the
rcu_read_lock bits around, the other is the timestamp bits which require
the ctx argument.