oprofile: Fix locking dependency in sync_start()
authorRobert Richter <robert.richter@amd.com>
Thu, 26 May 2011 16:39:35 +0000 (18:39 +0200)
committerRobert Richter <robert.richter@amd.com>
Tue, 31 May 2011 14:33:34 +0000 (16:33 +0200)
This fixes the A->B/B->A locking dependency, see the warning below.

The function task_exit_notify() is called with (task_exit_notifier)
.rwsem set and then calls sync_buffer() which locks buffer_mutex. In
sync_start() the buffer_mutex was set to prevent notifier functions to
be started before sync_start() is finished. But when registering the
notifier, (task_exit_notifier).rwsem is locked too, but now in
different order than in sync_buffer(). In theory this causes a locking
dependency, what does not occur in practice since task_exit_notify()
is always called after the notifier is registered which means the lock
is already released.

However, after checking the notifier functions it turned out the
buffer_mutex in sync_start() is unnecessary. This is because
sync_buffer() may be called from the notifiers even if sync_start()
did not finish yet, the buffers are already allocated but empty. No
need to protect this with the mutex.

So we fix this theoretical locking dependency by removing buffer_mutex
in sync_start(). This is similar to the implementation before commit:

 750d857 oprofile: fix crash when accessing freed task structs

which introduced the locking dependency.

Lockdep warning:

oprofiled/4447 is trying to acquire lock:
 (buffer_mutex){+.+...}, at: [<ffffffffa0000e55>] sync_buffer+0x31/0x3ec [oprofile]

but task is already holding lock:
 ((task_exit_notifier).rwsem){++++..}, at: [<ffffffff81058026>] __blocking_notifier_call_chain+0x39/0x67

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #1 ((task_exit_notifier).rwsem){++++..}:
       [<ffffffff8106557f>] lock_acquire+0xf8/0x11e
       [<ffffffff81463a2b>] down_write+0x44/0x67
       [<ffffffff810581c0>] blocking_notifier_chain_register+0x52/0x8b
       [<ffffffff8105a6ac>] profile_event_register+0x2d/0x2f
       [<ffffffffa00013c1>] sync_start+0x47/0xc6 [oprofile]
       [<ffffffffa00001bb>] oprofile_setup+0x60/0xa5 [oprofile]
       [<ffffffffa00014e3>] event_buffer_open+0x59/0x8c [oprofile]
       [<ffffffff810cd3b9>] __dentry_open+0x1eb/0x308
       [<ffffffff810cd59d>] nameidata_to_filp+0x60/0x67
       [<ffffffff810daad6>] do_last+0x5be/0x6b2
       [<ffffffff810dbc33>] path_openat+0xc7/0x360
       [<ffffffff810dbfc5>] do_filp_open+0x3d/0x8c
       [<ffffffff810ccfd2>] do_sys_open+0x110/0x1a9
       [<ffffffff810cd09e>] sys_open+0x20/0x22
       [<ffffffff8146ad4b>] system_call_fastpath+0x16/0x1b

-> #0 (buffer_mutex){+.+...}:
       [<ffffffff81064dfb>] __lock_acquire+0x1085/0x1711
       [<ffffffff8106557f>] lock_acquire+0xf8/0x11e
       [<ffffffff814634f0>] mutex_lock_nested+0x63/0x309
       [<ffffffffa0000e55>] sync_buffer+0x31/0x3ec [oprofile]
       [<ffffffffa0001226>] task_exit_notify+0x16/0x1a [oprofile]
       [<ffffffff81467b96>] notifier_call_chain+0x37/0x63
       [<ffffffff8105803d>] __blocking_notifier_call_chain+0x50/0x67
       [<ffffffff81058068>] blocking_notifier_call_chain+0x14/0x16
       [<ffffffff8105a718>] profile_task_exit+0x1a/0x1c
       [<ffffffff81039e8f>] do_exit+0x2a/0x6fc
       [<ffffffff8103a5e4>] do_group_exit+0x83/0xae
       [<ffffffff8103a626>] sys_exit_group+0x17/0x1b
       [<ffffffff8146ad4b>] system_call_fastpath+0x16/0x1b

other info that might help us debug this:

1 lock held by oprofiled/4447:
 #0:  ((task_exit_notifier).rwsem){++++..}, at: [<ffffffff81058026>] __blocking_notifier_call_chain+0x39/0x67

stack backtrace:
Pid: 4447, comm: oprofiled Not tainted 2.6.39-00007-gcf4d8d4 #10
Call Trace:
 [<ffffffff81063193>] print_circular_bug+0xae/0xbc
 [<ffffffff81064dfb>] __lock_acquire+0x1085/0x1711
 [<ffffffffa0000e55>] ? sync_buffer+0x31/0x3ec [oprofile]
 [<ffffffff8106557f>] lock_acquire+0xf8/0x11e
 [<ffffffffa0000e55>] ? sync_buffer+0x31/0x3ec [oprofile]
 [<ffffffff81062627>] ? mark_lock+0x42f/0x552
 [<ffffffffa0000e55>] ? sync_buffer+0x31/0x3ec [oprofile]
 [<ffffffff814634f0>] mutex_lock_nested+0x63/0x309
 [<ffffffffa0000e55>] ? sync_buffer+0x31/0x3ec [oprofile]
 [<ffffffffa0000e55>] sync_buffer+0x31/0x3ec [oprofile]
 [<ffffffff81058026>] ? __blocking_notifier_call_chain+0x39/0x67
 [<ffffffff81058026>] ? __blocking_notifier_call_chain+0x39/0x67
 [<ffffffffa0001226>] task_exit_notify+0x16/0x1a [oprofile]
 [<ffffffff81467b96>] notifier_call_chain+0x37/0x63
 [<ffffffff8105803d>] __blocking_notifier_call_chain+0x50/0x67
 [<ffffffff81058068>] blocking_notifier_call_chain+0x14/0x16
 [<ffffffff8105a718>] profile_task_exit+0x1a/0x1c
 [<ffffffff81039e8f>] do_exit+0x2a/0x6fc
 [<ffffffff81465031>] ? retint_swapgs+0xe/0x13
 [<ffffffff8103a5e4>] do_group_exit+0x83/0xae
 [<ffffffff8103a626>] sys_exit_group+0x17/0x1b
 [<ffffffff8146ad4b>] system_call_fastpath+0x16/0x1b

Reported-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Cc: Carl Love <carll@us.ibm.com>
Cc: <stable@kernel.org> # .36+
Signed-off-by: Robert Richter <robert.richter@amd.com>
drivers/oprofile/buffer_sync.c

index 04250aa16f513a5823669bf4026fe4c112fad374..f34b5b29fb955cb3129dd38f15296583ce64ff5e 100644 (file)
@@ -155,8 +155,6 @@ int sync_start(void)
        if (!zalloc_cpumask_var(&marked_cpus, GFP_KERNEL))
                return -ENOMEM;
 
-       mutex_lock(&buffer_mutex);
-
        err = task_handoff_register(&task_free_nb);
        if (err)
                goto out1;
@@ -173,7 +171,6 @@ int sync_start(void)
        start_cpu_work();
 
 out:
-       mutex_unlock(&buffer_mutex);
        return err;
 out4:
        profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
@@ -190,14 +187,13 @@ out1:
 
 void sync_stop(void)
 {
-       /* flush buffers */
-       mutex_lock(&buffer_mutex);
        end_cpu_work();
        unregister_module_notifier(&module_load_nb);
        profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
        profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
        task_handoff_unregister(&task_free_nb);
-       mutex_unlock(&buffer_mutex);
+       barrier();                      /* do all of the above first */
+
        flush_cpu_work();
 
        free_all_tasks();