Re: [GIT PULL rcu/next] RCU commits for 3.1

[Li Zefan <lizf@cn.fujitsu.com>]

(I shoud have cced Stephane Eranian instead of Turner..)

Paul E. McKenney wrote:
> On Mon, Oct 31, 2011 at 04:09:19PM +0800, Li Zefan wrote:
>> (Let's cc Peter and Paul Turner for this perf cgroup issue.)
>>
>>> Thank you for the analysis.  Does the following patch fix this problem?
>>>
>>> 							Thanx, Paul
>>>
>>> ------------------------------------------------------------------------
>>>
>>> fs: Add RCU protection in set_task_comm()
>>>
>>> Running "perf stat true" results in the following RCU-lockdep splat:
>>>
>>> ===============================
>>> [ INFO: suspicious RCU usage. ]
>>> -------------------------------
>>> include/linux/cgroup.h:548 suspicious rcu_dereference_check() usage!
>>>
>>> other info that might help us debug this:
>>>
>>> rcu_scheduler_active = 1, debug_locks = 0
>>> 1 lock held by true/655:
>>> #0:  (&sig->cred_guard_mutex){+.+.+.}, at: [<810d1bd7>] prepare_bprm_creds+0x27/0x70
>>>
>>> stack backtrace:
>>> Pid: 655, comm: true Not tainted 3.1.0-tip-01868-g1271bd2-dirty #161079
>>> Call Trace:
>>> [<81abe239>] ? printk+0x18/0x1a
>>> [<81064920>] lockdep_rcu_suspicious+0xc0/0xd0
>>> [<8108aa02>] perf_event_enable_on_exec+0x1d2/0x1e0
>>> [<81063764>] ? __lock_release+0x54/0xb0
>>> [<8108cca8>] perf_event_comm+0x18/0x60
>>> [<810d1abd>] ? set_task_comm+0x5d/0x80
>>> [<81af622d>] ? _raw_spin_unlock+0x1d/0x40
>>> [<810d1ac4>] set_task_comm+0x64/0x80
>>> [<810d25fd>] setup_new_exec+0xbd/0x1d0
>>> [<810d1b61>] ? flush_old_exec+0x81/0xa0
>>> [<8110753e>] load_elf_binary+0x28e/0xa00
>>> [<810d2101>] ? search_binary_handler+0xd1/0x1d0
>>> [<81063764>] ? __lock_release+0x54/0xb0
>>> [<811072b0>] ? load_elf_library+0x260/0x260
>>> [<810d2108>] search_binary_handler+0xd8/0x1d0
>>> [<810d2060>] ? search_binary_handler+0x30/0x1d0
>>> [<810d242f>] do_execve_common+0x22f/0x2a0
>>> [<810d24b2>] do_execve+0x12/0x20
>>> [<81009592>] sys_execve+0x32/0x70
>>> [<81af7752>] ptregs_execve+0x12/0x20
>>> [<81af76d4>] ? sysenter_do_call+0x12/0x36
>>>
>>> Li Zefan noted that this is due to set_task_comm() dropping the task
>>> lock before invoking perf_event_comm(), which could in fact result in
>>> the task being freed up before perf_event_comm() completed tracing in
>>> the case where one task invokes set_task_comm() on another task -- which
>>> actually does occur via comm_write(), which can be invoked via /proc.
>>>
>>
>> This is not true. The caller should ensure @tsk is valid during
>> set_task_comm().
>>
>> The warning comes from perf_cgroup_from_task(). We can trigger this warning
>> in some other cases where perf cgroup is used, for example:
> 
> I must defer to your greater knowledge of this situation.  What patch
> would you propose?
> 

With the following patch, we should see no rcu warning from perf, but as I
don't know the internel of perf, I guess we have to defer to Peter and
Stephane. ;)

I have two doubts:

- in perf_cgroup_sched_out/in(), we retrieve the task's cgroup twice in the function
and it's callee perf_cgroup_switch(), but the task can move to another cgroup between
two calls, so they might return two different cgroup pointers. Does it matter?

- in perf_cgroup_switch():

	 cpuctx->cgrp = perf_cgroup_from_task(task);

but seems the cgroup is not pinned, so cpuctx->cgrp can be invalid in later use.

---
diff --git a/kernel/events/core.c b/kernel/events/core.c
index d1a1bee..f5e05ce 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -302,7 +302,10 @@ static inline void update_cgrp_time_from_event(struct perf_event *event)
 	if (!is_cgroup_event(event))
 		return;
 
+	rcu_read_lock();
 	cgrp = perf_cgroup_from_task(current);
+	rcu_read_unlock();
+
 	/*
 	 * Do not update time when cgroup is not active
 	 */
@@ -325,9 +328,11 @@ perf_cgroup_set_timestamp(struct task_struct *task,
 	if (!task || !ctx->nr_cgroups)
 		return;
 
+	rcu_read_lock();
 	cgrp = perf_cgroup_from_task(task);
 	info = this_cpu_ptr(cgrp->info);
 	info->timestamp = ctx->timestamp;
+	rcu_read_unlock();
 }
 
 #define PERF_CGROUP_SWOUT	0x1 /* cgroup switch out every event */
@@ -406,6 +411,8 @@ 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
 	 */
@@ -418,6 +425,8 @@ static inline void perf_cgroup_sched_out(struct task_struct *task,
 	if (next)
 		cgrp2 = perf_cgroup_from_task(next);
 
+	rcu_read_unlock();
+
 	/*
 	 * only schedule out current cgroup events if we know
 	 * that we are switching to a different cgroup. Otherwise,
@@ -433,6 +442,8 @@ 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
 	 */
@@ -441,6 +452,8 @@ static inline void perf_cgroup_sched_in(struct task_struct *prev,
 	/* prev can never be NULL */
 	cgrp2 = perf_cgroup_from_task(prev);
 
+	rcu_read_unlock();
+
 	/*
 	 * only need to schedule in cgroup events if we are changing
 	 * cgroup during ctxsw. Cgroup events were not scheduled