On May 9, 2021 12:12:36 Martin Peres <martin.peres@xxxxxxx> wrote:
Hi,On 06/05/2021 22:13, Matthew Brost wrote:Basic GuC submission support. This is the first bullet point in theupstreaming plan covered in the following RFC [1].At a very high level the GuC is a piece of firmware which sits betweenthe i915 and the GPU. It offloads some of the scheduling of contextsfrom the i915 and programs the GPU to submit contexts. The i915communicates with the GuC and the GuC communicates with the GPU.May I ask what will GuC command submission do that execlist won't/can'tdo? And what would be the impact on users? Even forgetting the troubledhistory of GuC (instability, performance regression, poor level of usersupport, 6+ years of trying to upstream it...), adding this much codeand doubling the amount of validation needed should come with arationale making it feel worth it... and I am not seeing here. Would youmind providing the rationale behind this work?GuC submission will be disabled by default on all current upstreamplatforms behind a module parameter - enable_guc. A value of 3 willenable submission and HuC loading via the GuC. GuC submission shouldwork on all gen11+ platforms assuming the GuC firmware is present.What is the plan here when it comes to keeping support for execlist? Iam afraid that landing GuC support in Linux is the first step towardskilling the execlist, which would force users to use proprietaryfirmwares that even most Intel engineers have little influence over.Indeed, if "drm/i915/guc: Disable semaphores when using GuC scheduling"which states "Disable semaphores when using GuC scheduling as semaphoresare broken in the current GuC firmware." is anything to go by, it meansthat even Intel developers seem to prefer working around the GuCfirmware, rather than fixing it.
Yes, landing GuC support may be the first step in removing execlist support. The inevitable reality is that GPU scheduling is coming and likely to be there only path in the not-too-distant future. (See also the ongoing thread with AMD about fences.) I'm not going to pass judgement on whether or not this is a good thing. I'm just reading the winds and, in my view, this is where things are headed for good or ill.
In answer to the question above, the answer to "what do we gain from GuC?" may soon be, "you get to use your GPU." We're not there yet and, again, I'm not necessarily advocating for it, but that is likely where things are headed.
A firmware-based submission model isn't a bad design IMO and, aside from the firmware freedom issues, I think there are actual advantages to the model. Immediately, it'll unlock a few features like parallel submission (more on that in a bit) and long-running compute because they're implemented in GuC and the work to implement them properly in the execlist scheduler is highly non-trivial. Longer term, it may (no guarantees) unlock some performance by getting the kernel out of the way.
In the same vein, I have another concern related to the impact of GuC onLinux's stable releases. Let's say that in 3 years, a new applicationtriggers a bug in command submission inside the firmware. Given that theLinux community cannot patch the GuC firmware, how likely is it thatIntel would release a new GuC version? That would not be necessarilysuch a big problem if newer versions of the GuC could easily bebackported to this potentially-decade-old Linux version, but given thatthe GuC seems to have ABI-breaking changes on a monthly cadence (we areat major version 60 *already*? :o), I would say that it ishighly-unlikely that it would not require potentially-extensive changesto i915 to make it work, making the fix almost impossible to land in thestable tree... Do you have a plan to mitigate this problem?Patches like "drm/i915/guc: Disable bonding extension with GuCsubmission" also make me twitch, as this means the two commandsubmission paths will not be functionally equivalent, and enabling GuCcould thus introduce a user-visible regression (one app used to work,then stopped working). Could you add in the commit's message a proofthat this would not end up being a user regression (in which case, whyhave this codepath to begin with?).
I'd like to address this one specifically as it's become something of a speciality of mine the past few weeks. The current bonded submission model is bad. It provides a plethora of ways for a client to back itself into a corner and doesn't actually provide the guarantees the media driver needs for its real-time high-resolution decode. It's bad enough we're seriously considering ripping it out, backwards compatibility or not. The good news is that very little that your average desktop user does depends on it: basically just real-time >4K video decode.
The new parallel submit API is much better and should be the path forward. (We should have landed parallel submit the first time around.) It isn't full of corners and does let us provides actual parallel execution guarantees. It also gives the scheduler the information it needs to reliably provide those guarantees.
If we need to support the parallel submit API with the execlist back-end, that's totally possible. The choice to only implement the parallel submit API with GuC is a pragmatic one. We're trying to get upstream back on it's feet and get all the various up-and-coming bits of hardware enabled. Enabling the new API in the execlist back-end makes that pipeline longer.
Finally, could you explain why IGT tests need to be modified to work theGuC [1], and how much of the code in this series is covered byexisting/upcoming tests? I would expect a very solid set of tests tominimize the maintenance burden, and enable users to reproduce potentialissues found in this new codepath (too many users run with enable_guc=3,as can be seen on Google[2]).
The IGT changes, as I understand them, are entirely around switching to the new parallel submit API. There shouldn't be a major effect to most users.
--Jason
