[igt-dev] [PATCH i-g-t v2] lib/hang_ctx: Make use of dummyload library to create recursive batch

[Antonio Argenziano]

On 10/07/18 08:59, Chris Wilson wrote:
> Quoting Antonio Argenziano (2018-07-10 16:48:38)
>> On 10/07/18 08:38, Chris Wilson wrote:
>>> Quoting Antonio Argenziano (2018-07-10 16:31:55)
>>> Overall I am still dubious there is any reason to set this flag. The
>>> ratelimiting step here isn't the speed at which the batch spins, but the
>>> time for hangcheck to kick in. Where a fast spinner is interesting is
>>> the latency in responding to terminating the spinner (which never
>>> happens for a hang!).
>>
>> The flag makes sense to me because hangcheck will always sample the same
>> head on a very tight loop like the fast spinner. The current
>> implementation of hangcheck will also declare the engine hung much
>> quicker.
> 
> True, but I don't think much more than 1 or 2 extra ticks. In the grand
> scheme of things, using the pollable spinner + manual hang injection is
> far far far far far far quicker and precise.

Agree. Maybe this should be our preferred method for triggering hangs.

> 
>> In general I think we should expose different hanging batch to
>> the driver to make sure it can react to the different situations.
> 
> Harder to call (where do you draw the line between meaningful and
> hanging batch), but such construction is better done inside the test
> than inside a general purpose library (as it is no longer general
> purpose).

Makes sense, I was thinking of just some basic tests anyway.

Antonio

> -Chris
>