[Intel-gfx] [PATCH 05/51] drm/i915: Add return code check to i915_gem_execbuffer_retire_commands()

[John Harrison]

On 05/03/2015 16:14, Daniel Vetter wrote:
> On Thu, Mar 05, 2015 at 03:06:42PM +0000, John Harrison wrote:
>> On 05/03/2015 14:44, Daniel Vetter wrote:
>>> Imo reserving a bit of ring space for each add_request should be solid.
>>> Userspace uses the exact same reservation logic for adding end-of-batch
>>> workarounds. The only thing needed to make this solid is to WARN if
>>> add_request ends up using more ring space than what we've reserved (not
>>> just when it actually runs out, that obviously doesn't happen often
>>> enough for testing).
>> The problem is that there could be multiple requests being processed in
>> parallel. This is especially true with the scheduler. Userland could submit
>> a whole stream of batches that all get queued up in the scheduler. Only
>> later do they get submitted to the hardware. The request must be allocated
>> up front because there is no other means of tracking them. But reserving
>> space at that point won't work because you either end up reserving massive
>> amounts of space if the reserve is cumulative, or not enough if only one
>> slot is reserved.
> At least with execlist we don't have that problem really since writing the
> ringbuffer is done synchronously and directly.
>
> For the legacy scheduler I expect that we won't do any of the ringbuf
> writes directly and instead that's all done by the scheduler
> asynchronously.
>
> So this should just be an issue while we are converting to the scheduler
> or on platforms that will never have one. And imo the request ringbuf
> reservation is the solution with the simplest impact on the design.
I don't understand what you mean here. The scheduler decouples the 
submission of a batch buffer to the driver with the submission of that 
batch buffer to the hardware. The submission code path itself is 
identical. The scheduler does not do any hardware writes - it merely 
adds an extra layer between the user interface code and the hardware 
access code. The request creation must be done at driver submission 
time. The hardware submission could be an arbitrary amount of time 
later. In the intervening time, any number of new batch buffers may be 
submitted to the driver and thus new requests be created. This is true 
for both legacy and execlist mode.

The space reserve would have to be done at the start of the hardware 
submission process not at driver submission time. I.e. it would need to 
be a separate and new i915_gem_request_reserve_space() call at the start 
of the back half of the exec buffer code. It could not be done during 
request creation.


>>> Everything else just readds olr through the backdoor, which is kinda what
>>> we wanted to avoid from an accounting pov. Because then you have again
>>> some random request outstanding which scoops up everything it encounters.
>> Not quite.  The difference is that with something like an outstanding failed
>> request rather than a lazy one, there is still the segregation of work. The
>> failed request will be posted and added to the request list in its entirety
>> before a new request is allocated and used for the new work.
> Well history lesson, but that's exactly how olr started out. Then the
> hydra grew heads to no end. That's why I don't want to go down this road
> again, since I've been on that trip the past 3 years ;-) And since your
> motivation for olr light is exactly the one I provided 3 years ago to get
> my patch in I think history repeating is likely.
> -Daniel
Fair enough. Let's stay away from the evil hydra.

John.