[Intel-gfx] [RFC PATCH] drm/i915: Suppress page allocation warnings on engine park

[Janusz Krzysztofik]

On Thu, 2020-04-02 at 11:42 +0100, Chris Wilson wrote:
> Quoting Janusz Krzysztofik (2020-04-02 11:36:21)
> > On Thu, 2020-04-02 at 11:32 +0100, Chris Wilson wrote:
> > > Quoting Janusz Krzysztofik (2020-04-02 11:28:03)
> > > > On Thu, 2020-04-02 at 11:21 +0100, Chris Wilson wrote:
> > > > > Quoting Janusz Krzysztofik (2020-04-02 11:19:06)
> > > > > > On memory constrained systems it may happen that no pages are available
> > > > > > for serving object creation attempt during engine park.  Since we can
> > > > > > and we do ignore that failure, let's suppress possible warnings from
> > > > > > page allocator to avoid confusion and make CI happy.
> > > > > 
> > > > > The effect of ignoring it though is dangerous, hence why I had a
> > > > > warning.
> > > > 
> > > > Then maybe just WARN() from switch_to_kernel_context() on
> > > > __i915_request_create() returning -ENOMEM instead?
> > > 
> > > The warning exists already. The only real question is what to do about
> > > it; the best answer would be to preallocate the final request during
> > > unpark where we can report an error, but that would take a bit more
> > > effort to refactor request allocation. Hence the warning to make it a
> > > futureselves problem.
> > 
> > I meant a warning with a very specific message that could be filtered
> > easily by CI for now.
> 
> It has a very specific stacktrace, 

If CI is able to filter on specific stacktrace then OK.

> and I hope by filtered you mean
> identified and reported as an issue, possibly with multiple causes since
> this is an indication that reclaim is snafu.

Identified and reported as the reclaim issue specifically rather than
something quite general like:
TGL: igt at gem_exec_create@madvise - dmesg-warn - SUCCESS, page allocation failure: order:0, mode:0x40810(GFP_NOWAIT|__GFP_COMP|__GFP_RECLAIMABLE), nodemask=(null)

Thanks,
Janusz

> -Chris