nouveau crash due to missing channel (WAS: Re: [ANNOUNCE] 3.12.12-rt19)

Commit Message

Sebastian Andrzej Siewior March 7, 2014, 11:18 a.m. UTC

* Fernando Lopez-Lezcano | 2014-03-01 17:48:29 [-0800]:

>On 02/23/2014 10:47 AM, Sebastian Andrzej Siewior wrote:
>>Dear RT folks!
>>
>>I'm pleased to announce the v3.12.12-rt19 patch set.
>
>Just hit this Oops in my desktop at home:
>
>[22328.388996] BUG: unable to handle kernel NULL pointer dereference
>at 0000000000000008
>[22328.389013] IP: [<ffffffffa011a912>]
>nouveau_fence_wait_uevent.isra.2+0x22/0x440 [nouveau]

This is

| static int
| nouveau_fence_wait_uevent(struct nouveau_fence *fence, bool intr)
| 
| {
|         struct nouveau_channel *chan = fence->channel;
|         struct nouveau_fifo *pfifo = nouveau_fifo(chan->drm->device);

and chan is NULL. 

>[22328.389046] RAX: 0000000000000000 RBX: ffff8807a68f8fa8 RCX:
>0000000000000000
>[22328.389046] RDX: 0000000000000001 RSI: ffff8807a68f8fb0 RDI:
>ffff8807a68f8fa8
>[22328.389047] RBP: ffff8807c09bdca0 R08: 000000000000045e R09:
>000000000000e200
>[22328.389047] R10: ffffffffa0157d80 R11: ffff8807c09bdde0 R12:
>0000000000000001
>[22328.389047] R13: 0000000000000000 R14: ffff8807d8493a80 R15:
>ffff8807a68f8fb0
>[22328.389053] Call Trace:
>[22328.389069]  [<ffffffffa011af56>] nouveau_fence_wait+0x86/0x1a0 [nouveau]
>[22328.389081]  [<ffffffffa011ca35>] nouveau_bo_fence_wait+0x15/0x20
>[nouveau]
>[22328.389084]  [<ffffffffa00867c6>] ttm_bo_wait+0x96/0x1a0 [ttm]
>[22328.389095]  [<ffffffffa0121dac>]
>nouveau_gem_ioctl_cpu_prep+0x5c/0xf0 [nouveau]
>[22328.389101]  [<ffffffffa002cd42>] drm_ioctl+0x502/0x630 [drm]
>[22328.389114]  [<ffffffffa01180a1>] nouveau_drm_ioctl+0x51/0x90 [nouveau]

I can't find any kind of locking so my question is what ensures that chan is
not set to NULL between nouveau_fence_done() and
nouveau_fence_wait_uevent()? There are just a few opcodes in between but
nothing that pauses nouveau_fence_signal().

Fernando, can you please check the patch below and test if the warning
or the crash appears?


>-- Fernando

Sebastian

Comments

Sebastian Andrzej Siewior March 7, 2014, 11:53 a.m. UTC | #1

* Maarten Lankhorst | 2014-03-07 12:36:13 [+0100]:

>>I can't find any kind of locking so my question is what ensures that chan is
>>not set to NULL between nouveau_fence_done() and
>>nouveau_fence_wait_uevent()? There are just a few opcodes in between but
>>nothing that pauses nouveau_fence_signal().
>Absolutely nothing. :-) Worse still, there's no guarantee that channel isn't freed, but hopefully that is less likely to be an issue.

Okay, so I hit the correct spot. What do we do here? Do you want the
patch I posted without the WARN_ON() or do you prefer to fix this in an
other way?

>~Maarten

Sebastian

Patch

diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c b/drivers/gpu/drm/nouveau/nouveau_fence.c
--- a/drivers/gpu/drm/nouveau/nouveau_fence.c
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
@@ -184,14 +184,20 @@  nouveau_fence_wait_uevent(struct nouveau_fence *fence, bool intr)
 
 {
 	struct nouveau_channel *chan = fence->channel;
-	struct nouveau_fifo *pfifo = nouveau_fifo(chan->drm->device);
-	struct nouveau_fence_priv *priv = chan->drm->fence;
+	struct nouveau_fifo *pfifo;
+	struct nouveau_fence_priv *priv;
 	struct nouveau_fence_uevent uevent = {
 		.handler.func = nouveau_fence_wait_uevent_handler,
-		.priv = priv,
 	};
 	int ret = 0;
 
+	if (WARN_ON_ONCE(!chan))
+		return 0;
+
+	pfifo = nouveau_fifo(chan->drm->device);
+	priv = chan->drm->fence;
+	uevent.priv = priv;
+
 	nouveau_event_get(pfifo->uevent, 0, &uevent.handler);
 
 	if (fence->timeout) {