[Intel-gfx] [PATCH 2/2] drm/i915: Remove impossibe checks from i915_gem_request_add_to_client

Tvrtko Ursulin tvrtko.ursulin at linux.intel.com
Thu Mar 24 09:49:25 UTC 2016

On 23/03/16 15:31, Chris Wilson wrote:
> On Wed, Mar 23, 2016 at 03:15:03PM +0000, Tvrtko Ursulin wrote:
>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>> There is only one caller of this functions which calls it under
>> strictly defined conditions. Error checking and return value
>> can be removed.
>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>> ---
>> BTW it looks like the whole point of this request list is to implement
>> the throttle ioctl? Looks like a strange ioctl with a fixed 20ms criteria,
>> could it be re-implemented somehow without having to have this list?
>> Does it have to be per-client or could it wait for any requests on any
>> engine emitted 20ms ago?
> It is per-client, and I long wished the 20ms wasn't defined by the ABI.

Is it useful and/or widely used though?

With multiple clients submitting work I don't see the semantics are 
deterministic. I was thinking if it could be replaced with a simpler 
implementation for the similar random effect, losing the list altogether?

Especially wonder how the scheduler will affect it.

> What I proposed doing is this:
> https://cgit.freedesktop.org/~ickle/linux-2.6/commit/?h=breadcrumbs&id=e33947505b54582964c8c202b22f88fc5705f484

How is it safe? list_add modifies various pointers in multiple steps so 
I didn't know that is safe against concurrent iteration.

RCU flavoured list API can do such things but in this case switching to 
that would only enable lockless throttle and not gain anything on the 
add_to_client path.

> Please note that this also fixes a race condition I've seen in the wild.

What is the race?



More information about the Intel-gfx mailing list