lockdep: annotate epoll
Peter Zijlstra [Tue, 5 Feb 2008 06:27:20 +0000 (22:27 -0800)]
On Sat, 2008-01-05 at 13:35 -0800, Davide Libenzi wrote:

> I remember I talked with Arjan about this time ago. Basically, since 1)
> you can drop an epoll fd inside another epoll fd 2) callback-based wakeups
> are used, you can see a wake_up() from inside another wake_up(), but they
> will never refer to the same lock instance.
> Think about:
>
>  dfd = socket(...);
>  efd1 = epoll_create();
>  efd2 = epoll_create();
>  epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
>  epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
>
> When a packet arrives to the device underneath "dfd", the net code will
> issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
> callback wakeup entry on that queue, and the wake_up() performed by the
> "dfd" net code will end up in ep_poll_callback(). At this point epoll
> (efd1) notices that it may have some event ready, so it needs to wake up
> the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
> that ends up in another wake_up(), after having checked about the
> recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
> avoid stack blasting. Never hit the same queue, to avoid loops like:
>
>  epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
>  epoll_ctl(efd3, EPOLL_CTL_ADD, efd2, ...);
>  epoll_ctl(efd4, EPOLL_CTL_ADD, efd3, ...);
>  epoll_ctl(efd1, EPOLL_CTL_ADD, efd4, ...);
>
> The code "if (tncur->wq == wq || ..." prevents re-entering the same
> queue/lock.

Since the epoll code is very careful to not nest same instance locks
allow the recursion.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Tested-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Acked-by: Davide Libenzi <davidel@xmailserver.org>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

fs/eventpoll.c
include/linux/wait.h

index 81c04ab..a415f42 100644 (file)
@@ -353,7 +353,7 @@ static void ep_poll_safewake(struct poll_safewake *psw, wait_queue_head_t *wq)
        spin_unlock_irqrestore(&psw->lock, flags);
 
        /* Do really wake up now */
-       wake_up(wq);
+       wake_up_nested(wq, 1 + wake_nests);
 
        /* Remove the current task from the list */
        spin_lock_irqsave(&psw->lock, flags);
index 1f4fb0a..33a2aa9 100644 (file)
@@ -162,6 +162,22 @@ wait_queue_head_t *FASTCALL(bit_waitqueue(void *, int));
 #define wake_up_interruptible_all(x)   __wake_up(x, TASK_INTERRUPTIBLE, 0, NULL)
 #define wake_up_interruptible_sync(x)  __wake_up_sync((x), TASK_INTERRUPTIBLE, 1)
 
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+/*
+ * macro to avoid include hell
+ */
+#define wake_up_nested(x, s)                                           \
+do {                                                                   \
+       unsigned long flags;                                            \
+                                                                       \
+       spin_lock_irqsave_nested(&(x)->lock, flags, (s));               \
+       wake_up_locked(x);                                              \
+       spin_unlock_irqrestore(&(x)->lock, flags);                      \
+} while (0)
+#else
+#define wake_up_nested(x, s)           wake_up(x)
+#endif
+
 #define __wait_event(wq, condition)                                    \
 do {                                                                   \
        DEFINE_WAIT(__wait);                                            \