console: implement lockdep support for console_lock
authorDaniel Vetter <daniel.vetter@ffwll.ch>
Sat, 22 Sep 2012 17:52:11 +0000 (19:52 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 22 Oct 2012 23:12:20 +0000 (16:12 -0700)
Dave Airlie recently discovered a locking bug in the fbcon layer,
where a timer_del_sync (for the blinking cursor) deadlocks with the
timer itself, since both (want to) hold the console_lock:

https://lkml.org/lkml/2012/8/21/36

Unfortunately the console_lock isn't a plain mutex and hence has no
lockdep support. Which resulted in a few days wasted of tracking down
this bug (complicated by the fact that printk doesn't show anything
when the console is locked) instead of noticing the bug much earlier
with the lockdep splat.

Hence I've figured I need to fix that for the next deadlock involving
console_lock - and with kms/drm growing ever more complex locking
that'll eventually happen.

Now the console_lock has rather funky semantics, so after a quick irc
discussion with Thomas Gleixner and Dave Airlie I've quickly ditched
the original idead of switching to a real mutex (since it won't work)
and instead opted to annotate the console_lock with lockdep
information manually.

There are a few special cases:
- The console_lock state is protected by the console_sem, and usually
  grabbed/dropped at _lock/_unlock time. But the suspend/resume code
  drops the semaphore without dropping the console_lock (see
  suspend_console/resume_console). But since the same thread that did
  the suspend will do the resume, we don't need to fix up anything.

- In the printk code there's a special trylock, only used to kick off
  the logbuffer printk'ing in console_unlock. But all that happens
  while lockdep is disable (since printk does a few other evil
  tricks). So no issue there, either.

- The console_lock can also be acquired form irq context (but only
  with a trylock). lockdep already handles that.

This all leaves us with annotating the normal console_lock, _unlock
and _trylock functions.

And yes, it works - simply unloading a drm kms driver resulted in
lockdep complaining about the deadlock in fbcon_deinit:

======================================================
[ INFO: possible circular locking dependency detected ]
3.6.0-rc2+ #552 Not tainted
-------------------------------------------------------
kms-reload/3577 is trying to acquire lock:
 ((&info->queue)){+.+...}, at: [<ffffffff81058c70>] wait_on_work+0x0/0xa7

but task is already holding lock:
 (console_lock){+.+.+.}, at: [<ffffffff81264686>] bind_con_driver+0x38/0x263

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #1 (console_lock){+.+.+.}:
       [<ffffffff81087440>] lock_acquire+0x95/0x105
       [<ffffffff81040190>] console_lock+0x59/0x5b
       [<ffffffff81209cb6>] fb_flashcursor+0x2e/0x12c
       [<ffffffff81057c3e>] process_one_work+0x1d9/0x3b4
       [<ffffffff810584a2>] worker_thread+0x1a7/0x24b
       [<ffffffff8105ca29>] kthread+0x7f/0x87
       [<ffffffff813b1204>] kernel_thread_helper+0x4/0x10

-> #0 ((&info->queue)){+.+...}:
       [<ffffffff81086cb3>] __lock_acquire+0x999/0xcf6
       [<ffffffff81087440>] lock_acquire+0x95/0x105
       [<ffffffff81058cab>] wait_on_work+0x3b/0xa7
       [<ffffffff81058dd6>] __cancel_work_timer+0xbf/0x102
       [<ffffffff81058e33>] cancel_work_sync+0xb/0xd
       [<ffffffff8120a3b3>] fbcon_deinit+0x11c/0x1dc
       [<ffffffff81264793>] bind_con_driver+0x145/0x263
       [<ffffffff81264a45>] unbind_con_driver+0x14f/0x195
       [<ffffffff8126540c>] store_bind+0x1ad/0x1c1
       [<ffffffff8127cbb7>] dev_attr_store+0x13/0x1f
       [<ffffffff8116d884>] sysfs_write_file+0xe9/0x121
       [<ffffffff811145b2>] vfs_write+0x9b/0xfd
       [<ffffffff811147b7>] sys_write+0x3e/0x6b
       [<ffffffff813b0039>] system_call_fastpath+0x16/0x1b

other info that might help us debug this:

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(console_lock);
                               lock((&info->queue));
                               lock(console_lock);
  lock((&info->queue));

 *** DEADLOCK ***

v2: Mark the lockdep_map static, noticed by Jani Nikula.

Cc: Dave Airlie <airlied@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
kernel/printk.c

index 2d607f4d1797cb02e0e4fb60453312a8d3491975..ee79f14db358f1f996313739dfc7d9fc47f8a61c 100644 (file)
@@ -87,6 +87,12 @@ static DEFINE_SEMAPHORE(console_sem);
 struct console *console_drivers;
 EXPORT_SYMBOL_GPL(console_drivers);
 
+#ifdef CONFIG_LOCKDEP
+static struct lockdep_map console_lock_dep_map = {
+       .name = "console_lock"
+};
+#endif
+
 /*
  * This is used for debugging the mess that is the VT code by
  * keeping track if we have the console semaphore held. It's
@@ -1914,6 +1920,7 @@ void console_lock(void)
                return;
        console_locked = 1;
        console_may_schedule = 1;
+       mutex_acquire(&console_lock_dep_map, 0, 0, _RET_IP_);
 }
 EXPORT_SYMBOL(console_lock);
 
@@ -1935,6 +1942,7 @@ int console_trylock(void)
        }
        console_locked = 1;
        console_may_schedule = 0;
+       mutex_acquire(&console_lock_dep_map, 0, 1, _RET_IP_);
        return 1;
 }
 EXPORT_SYMBOL(console_trylock);
@@ -2095,6 +2103,7 @@ skip:
                local_irq_restore(flags);
        }
        console_locked = 0;
+       mutex_release(&console_lock_dep_map, 1, _RET_IP_);
 
        /* Release the exclusive_console once it is used */
        if (unlikely(exclusive_console))