Re: [PATCH] perf: call perf_event_comm under task_lock to fix suspicious rcu usage

[Hannes Frederic Sowa <hannes@stressinduktion.org>]

Ping, this problem still persists in v3.7-rc6. Could someone have a look?

On Sat, Nov 10, 2012 at 06:32:28AM +0100, Hannes Frederic Sowa wrote:
> Following RCU warning showed up while executing a shebang-script under
> perf-record (could even be an empty script) on a 3.7-rc4 stable kernel:
> 
>   [   32.185108] 
>   [   32.185332] ===============================
>   [   32.185602] [ INFO: suspicious RCU usage. ]
>   [   32.185903] 3.7.0-rc4 #1 Not tainted
>   [   32.186021] -------------------------------
>   [   32.186021] include/linux/cgroup.h:566 suspicious rcu_dereference_check() usage!
>   [   32.186021] 
>   [   32.186021] other info that might help us debug this:
>   [   32.186021] 
>   [   32.186021] 
>   [   32.186021] rcu_scheduler_active = 1, debug_locks = 0
>   [   32.186021] 1 lock held by empty.sh/556:
>   [   32.186021]  #0:  (&sig->cred_guard_mutex){+.+.+.}, at: [<ffffffff811d1d16>] prepare_bprm_creds+0x36/0x80
>   [   32.186021] 
>   [   32.186021] stack backtrace:
>   [   32.186021] Pid: 556, comm: empty.sh Not tainted 3.7.0-rc4 #1
>   [   32.186021] Call Trace:
>   [   32.186021]  [<ffffffff810d336d>] lockdep_rcu_suspicious+0xfd/0x130
>   [   32.186021]  [<ffffffff81153de6>] perf_event_comm+0x436/0x610
>   [   32.186021]  [<ffffffff810d181d>] ? trace_hardirqs_off+0xd/0x10
>   [   32.186021]  [<ffffffff810ac37f>] ? local_clock+0x6f/0x80
>   [   32.186021]  [<ffffffff810d217f>] ? lock_release_holdtime.part.26+0xf/0x180
>   [   32.186021]  [<ffffffff811d1bd3>] set_task_comm+0x73/0x180
>   [   32.186021]  [<ffffffff811d24ba>] setup_new_exec+0x9a/0x210
>   [   32.186021]  [<ffffffff81225cf3>] load_elf_binary+0x3e3/0x1ab0
>   [   32.186021]  [<ffffffff810ac0f5>] ? sched_clock_local+0x25/0xa0
>   [   32.186021]  [<ffffffff810ac298>] ? sched_clock_cpu+0xa8/0x120
>   [   32.186021]  [<ffffffff810d181d>] ? trace_hardirqs_off+0xd/0x10
>   [   32.186021]  [<ffffffff810ac37f>] ? local_clock+0x6f/0x80
>   [   32.186021]  [<ffffffff81225910>] ? load_elf_library+0x240/0x240
>   [   32.186021]  [<ffffffff81225910>] ? load_elf_library+0x240/0x240
>   [   32.186021]  [<ffffffff811d14a4>] search_binary_handler+0x194/0x4f0
>   [   32.186021]  [<ffffffff811d136f>] ? search_binary_handler+0x5f/0x4f0
>   [   32.186021]  [<ffffffff81225220>] ? compat_sys_ioctl+0x1510/0x1510
>   [   32.186021]  [<ffffffff812254b4>] load_script+0x294/0x2c0
>   [   32.186021]  [<ffffffff810d217f>] ? lock_release_holdtime.part.26+0xf/0x180
>   [   32.186021]  [<ffffffff81225220>] ? compat_sys_ioctl+0x1510/0x1510
>   [   32.186021]  [<ffffffff811d14a4>] search_binary_handler+0x194/0x4f0
>   [   32.186021]  [<ffffffff811d136f>] ? search_binary_handler+0x5f/0x4f0
>   [   32.186021]  [<ffffffff811d22cb>] do_execve_common.isra.25+0x50b/0x5b0
>   [   32.186021]  [<ffffffff811d1eea>] ? do_execve_common.isra.25+0x12a/0x5b0
>   [   32.186021]  [<ffffffff811d238b>] do_execve+0x1b/0x20
>   [   32.186021]  [<ffffffff811d26c4>] sys_execve+0x54/0x80
>   [   32.186021]  [<ffffffff816d8f99>] stub_execve+0x69/0xc0
> 
> I think this dereference qualifies for the task_lock exception (as noted
> in kernel/cgroup.c), thus this patch ensures calling perf_event_comm
> before giving up the task_lock.
> 
> Changelog -v2 (since <20121103235758.GD18106@order.stressinduktion.org>):
>   1) rebased to 3.7-rc4
>   2) slightly improved/updated commit msg and added more people to Cc
> 
> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
> Cc: Paul Mackerras <paulus@samba.org>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
> Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
> ---
>  fs/exec.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/fs/exec.c b/fs/exec.c
> index 0039055..a961b9d 100644
> --- a/fs/exec.c
> +++ b/fs/exec.c
> @@ -1038,8 +1038,8 @@ void set_task_comm(struct task_struct *tsk, char *buf)
>  	memset(tsk->comm, 0, TASK_COMM_LEN);
>  	wmb();
>  	strlcpy(tsk->comm, buf, sizeof(tsk->comm));
> -	task_unlock(tsk);
>  	perf_event_comm(tsk);
> +	task_unlock(tsk);
>  }
>  
>  static void filename_to_taskname(char *tcomm, const char *fn, unsigned int len)
>