[PATCH] drm/amd/powerplay/smu7_hwmgr: replace blocking delay with non-blocking

[Alex Deucher]

On Tue, Jun 4, 2019 at 4:22 PM Yrjan Skrimstad <yrjan at skrimstad.net> wrote:
>
> On Thu, May 30, 2019 at 02:08:21AM +0200, Yrjan Skrimstad wrote:
> > This driver currently contains a repeated 500ms blocking delay call
> > which causes frequent major buffer underruns in PulseAudio. This patch
> > fixes this issue by replacing the blocking delay with a non-blocking
> > sleep call.
>
> I see that I have not explained this bug well enough, and I hope that is
> the reason for the lack of replies on this patch. I will here attempt to
> explain the situation better.
>
> To start with some hardware description I am here using an AMD R9 380
> GPU, an AMD Ryzen 7 1700 Eight-Core Processor and an AMD X370 chipset.
> If any more hardware or software specifications are necessary, please
> ask.
>
> The bug is as follows: When playing audio I will regularly have major
> audio issues, similar to that of a skipping CD. This is reported by
> PulseAudio as scheduling delays and buffer underruns when running
> PulseAudio verbosely and these scheduling delays are always just under
> 500ms, typically around 490ms. This makes listening to any music quite
> the horrible experience as PulseAudio constantly will attempt to rewind
> and catch up. It is not a great situation, and seems to me to quite
> clearly be a case where regular user space behaviour has been broken.
>
> I want to note that this audio problem was not something I experienced
> until recently, it is therefore a new bug.
>
> I have bisected the kernel to find out where the problem originated and
> found the following commit:
>
> # first bad commit: [f5742ec36422a39b57f0256e4847f61b3c432f8c] drm/amd/powerplay: correct power reading on fiji
>
> This commit introduces a blocking delay (mdelay) of 500ms, whereas the
> old behaviour was a smaller blocking delay of only 1ms. This seems to me
> to be very curious as the scheduling delays of PulseAudio are always
> almost 500ms. I have therefore with the previous patch replaced the
> scheduling delay with a non-blocking sleep (msleep).
>
> The results of the patch seems promising as I have yet to encounter any
> of the old <500ms scheduling delays when using it and I have also not
> encountered any kernel log messages regarding sleeping in an atomic
> context.

The patch is fine and I can apply it (I don't think there are any
restrictions on sleeping in sysfs), but this code only gets executed
when you actually read the power status from the card (e.g., via sysfs
or debugfs).  Presumably you have something in userspace polling one
of those files on a regular basis?

Alex