On 2026. szeptember 16., szerda 22:23:42 közép-európai nyári idő Alex Deucher wrote: > On Tue, Sep 15, 2026 at 1:23 PM Timur Kristóf <[email protected]> wrote: > > On Friday, September 11, 2026 9:00:07 PM Central European Summer Time Alex > > > > Deucher wrote: > > > On Mon, Sep 7, 2026 at 4:40 PM Timur Kristóf <[email protected]> > > > > wrote: > > > > This series applies on top of my previous series: > > > > "Improve existing SDMA queue resets" (currently under review) > > > > > > > > Prepare the code for implementing recovery for > > > > SDMA v4 and all older SDMA (and SI DMA) versions. > > > > Reorganize the SDMA 4.0 code so that the functions that are > > > > responsible for managing the SDMA engines take an instance ID > > > > as an argument. > > > > > > > > This makes it possible to manage the SDMA instances > > > > independently of each other, which will enable us to > > > > also use these functions to implement resetting and > > > > recovering them independently. (The actual recovery > > > > implementation will be in a follow-up patch series > > > > which I will submit after this one is accepted.) > > > > > > This series looks fine to me, although I think I would prefer to apply > > > it along with the relevant soft reset changes to avoid churning the > > > code until each family is ready. E.g., apply the cik_sdma refactor > > > along with the CIK soft reset support, etc. Unless you have something > > > else in mind. > > > > Thank you Alex. > > > > My plan is the following: > > > > 1. I would like to first refactor the functions for every SDMA IP block > > version to be per-instance. > > 2. Then, I'd like to improve the soft reset code, more specifically the > > multi- ring reset helpers and amdgpu_device_ip_soft_reset() to make them > > aware of IP block instances and change SDMA v4.4.2, v5.0 and v5.2 over to > > use that code. 3. With all of that out of the way, it will be pretty easy > > to hook up the same soft reset mechanism for SI, CIK, VI as well. > > 4. At that point, we should also consider if we want soft reset for SDMA > > v6 > > and v7. Technically these versions support proper queue reset through the > > MES, but in practice I've seen the MES fail so often that I wouldn't > > consider it stable. > > > > This RFC is basically a prototype for step (1) and only takes care of SDMA > > v4 and older just to show how I imagine doing the refactor. If you like > > this direction, I would like to do the same for all IP versions before > > moving forward with soft reset changes and other refactors. I prefer to > > do this first for all IP versions because I think it's important to have > > consistency within the code base regarding how the SDMA is programmed. > > > > What do you think, would that be OK for you? > > Do you have a PoC working on at least one generation based on the code > restructure? I guess I'd rather not restructure everything until you > know the end result actually works for what you are trying to do.
(Sorry for the late reply.) Okay, that sounds reasonable to me. I will get back to you once I got it fully working on multiple HW generations. Thanks & best regards, Timur
