[Intel-gfx] [RFC] Page table sharing and bufferless execbuf

[Jesse Barnes]

On 09/04/2015 10:23 AM, Chris Wilson wrote:
> On Fri, Sep 04, 2015 at 09:58:54AM -0700, Jesse Barnes wrote:
>> A few things to think about:
>>   - how to signal GPU hangs with the new execbuf (a signal might be more
>>     natural as the execution appears more CPU-like?  what state do we
>>     have to worry about restoring for bufferless contexts?
> 
> GPU hangs are not signalled via execbuf. The only time userspace is told
> via execbuf is when the kernel rejects the cmdbuffer (EIO when it can't
> program the hardware queues). EIO notification currently is at the wait
> points, and with manually querying. Having a signal would fit in nicely
> with the async nature of cmdbuffers and be useful today - though I don't
> know how practical it is to extend the signal ABI.

Right; there's signalfd, but maybe a new type of event would be even
better.  I'm more worried about handling the internals of GPU hangs with
bufferless contexts though; I definitely haven't tested that much (only
accidentally :).

Jesse