[Intel-gfx] [PATCH 0/7] Commit early to GuC

[Daniele Ceraolo Spurio]

On 1/15/20 8:18 AM, Chris Wilson wrote:
> Quoting Daniele Ceraolo Spurio (2020-01-15 15:57:27)
>>
>>
>> On 1/15/2020 12:40 AM, Chris Wilson wrote:
>>> Quoting Daniele Ceraolo Spurio (2020-01-15 01:31:36)
>>>> We currently wait until we attempt to load the GuC to confirm if we're
>>>> in GuC mode or not, at which point a lot of the engine setup has already
>>>> happened and needs to be updated for GuC submission. To allow us to get
>>>> the setup done directly into GuC mode, we need to commit to using GuC
>>>> as soon as possible.
>>> I think this is the wrong direction; as I thought the goal was to allow
>>> delayed loading of firmware, even going as far as allowing the system to
>>> run a browser for the user to get the firmware first. I may be
>>
>> We do indeed want to keep supporting execlists mode even as some HW
>> features move to the GuC to allow the user to get to the binaries, but
>> we don't want to switch between the 2 modes without a reboot. Switching
>> between the 2 modes is not a HW capability that we're committed to; the
>> guc->elsp transition is already not possible, while the elsp->guc one
>> still seems to work, but who knows for how long it will?
>>
>> This series is also not really changing the commitment at the
>> implementation level, just making it "official" and acting based on
>> that. Even without these patches, if the blobs are on the system we will
>> attempt to get into GuC mode unless we get an allocation failure or
>> something similar, in which case it is extremely likely that the
>> fall-back to non-guc will not work either.
>>
>>> completely wrong about that, but imho I never want to have to build
>>> firmware images into the kernel.
>>
>> I do 100% agree with this statement, although I'm not sure how this
>> relates to the series. Are you planning to pull some of the engine setup
>> to an even earlier point?
>>
>>>
>>> The transition from execlists to guc could be just set-wedged; delete
>>> old engines, build guc engines. [This should also work for guc -> guc.]
>>> Throwing away context state is ugly, but simple enough.
>>
>> As mentioned above, we can't switch between elsp and GuC modes so this
>> transition would have to be done before the first submission to HW. Why
>> not go directly in GuC mode then?
> 
> So the problem is if we can't freely switch (we can never power down the
> guc, that seems unlikely?) then we can't make a decision on which mode

AFAIU it's not the GuC itself that's the issue, but some of the other 
units, some of which outside the GT, that can get locked in one mode 
without being resettable (like what happens for the WOPCM regs for example).

> to run (and which engines to initialise) until userspace is active and
> has committed to supplying or not supplying a fw image. Which puts us in
> a catch-22 of wanting to register the driver with userspace before we
> have finalized initialisation.

I think this is the bit I'm missing. Why do you want userspace to 
directly provide the firmware instead of fetching it from /lib/firmware 
like we do now? The distros should pack the correct firmware in their 
linux-firmware packages and it seems reasonable to me to expect the 
system to be rebooted after fetching a binary by hand. We can move the 
fetch to an earlier point in time if we need the info earlier, since it 
does not require the HW to be ready.

> 
> If the transition is impossible, it seems like you have no choice but to
> require the fw image at initialisation. I do not understand why it has
> to be that way, seems such a hindrance.

My understanding from the talk I had with the HW team when we realized 
that the guc->elsp transition was broken on gen11 is that the HW expects 
SW to pick a mode and stick to that. The elsp->guc transition seem to 
still work, but there is no guarantee it will keep doing so in the 
future and therefore it doesn't seem like a good idea to build on that.

Daniele

> -Chris
>