(In reply to comment #6)
> Hm, just noticed this (same as last time, but I guess I wasn't paying enough
> attention):
> 
> nouveau E[  PGRAPH][0000:05:00.0] ch 4 [0x00178fb000 mplayer[1645]] subc 3
> class 0x8297 mthd 0x1b0c data 0x1000f010
> nouveau E[     PFB][0000:05:00.0] trapped write at 0x00204050c0 on channel
> 0x000178fb [mplayer[1645]] PGRAPH/PROP/RT0 reason: PAGE_NOT_PRESENT
> 
> method 1b0c == QUERY_GET. However the data is 0x1000f010, not 0xf010, which
> means it's not coming from the video stuff or the screen fence stuff, but
> rather is a PIPE_QUERY_GPU_FINISHED query. Which is interesting, because
> nothing appears to actually use that PIPE_QUERY type...
> 
> Could you stick a assert(0) right before
> 
>       nv50_query_get(push, q, 0, 0x1000f010);
> 
> in nv50_query.c and see if it triggers (make sure to use a debug build... or
> make it something nastier than assert), and get a backtrace from gdb? (Or
> actually just set a breakpoint on that line and see if gdb hits it...)

No, I tried  to add assert :

diff --git a/src/gallium/drivers/nouveau/nv50/nv50_query.c
b/src/gallium/drivers/nouveau/nv50/nv50_query.c
index a373dc6..23f5eaf 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_query.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_query.c
@@ -222,6 +222,7 @@ nv50_query_end(struct pipe_context *pipe, struct pipe_query
*pq)
       break;
    case PIPE_QUERY_GPU_FINISHED:
       q->sequence++;
+      assert(0);
       nv50_query_get(push, q, 0, 0x1000f010);
       break;
    case NVA0_QUERY_STREAM_OUTPUT_BUFFER_OFFSET:

on top of mesa-git-454aab4 and it doesn't trigger  during mplayer's run
(resulted in usual hang).