On Fri, 2021-12-24 at 12:09 +0000, Tvrtko Ursulin wrote:
> Hi,
>
> Somehow I stumbled on this while browsing through the mailing list.
>
> On 23/12/2021 18:54, Teres Alexis, Alan Previn wrote:
> > Revisiting below hunk of patch-7 comment, as per offline discussion with Matt,
> > there is little benefit to even making that guc-id lookup because:
> >
> > 1. the delay between the context reset notification (when the vmas are copied
> > and when we verify we had received a guc err capture dump) may be subjectively
> > large enough and not tethered that the guc-id may have already been re-assigned.
> >
> > 2. I was really looking for some kind of unique context handle to print out that could
> > be correlated (by user inspecting the dump) back to a unique app or process or
> > context-id but cant find such a param in struct intel_context.
> >
> > As part of further reviewing the end to end flows and possible error scenarios, there
> > also may potentially be a mismatch between "which context was reset by guc at time-n"
> > vs "which context's vma buffers is being printed out at time-n+x" if
> > we are experiencing back-to-back resets and the user dumped the debugfs x-time later.
>
> What does this all actually mean, because it sounds rather alarming,
> that it just won't be possible to know which context, belonging to which
> process, was reset? And because of guc_id potentially re-assigned even
> the captured VMAs may not be the correct ones?
>
>
The flow of events are as below:
1. guc sends notification that an error capture was done and ready to take.
- at this point we copy the guc error captured dump into an interim store
(larger buffer that can hold multiple captures).
2. guc sends notification that a context was reset (after the prior)
- this triggers a call to i915_gpu_coredump with the corresponding engine-mask
from the context that was reset
- i915_gpu_coredump proceeds to gather entire gpu state including driver state,
global gpu state, engine state, context vmas and also engine registers. For the
engine registers now call into the guc_capture code which merely needs to verify
that GuC had already done a step 1 and we have data ready to be parsed.
(time elapses)
3. end user triggers the sysfs to dump the error state and all prior information is
printed out in proper format.
Between 2 and 3:
- Looking at existing framework (established by execlist-capture codes),
I believe we only hold on to the first error state capture and drop any
subsequent context reset captures occurring before #3 (i.e. before the end user
triggers the debugfs)
- However, in that same space, guc can send us more and more error-capture logs
long as we have space for it in the buffer.
So the issue was that in my original patch, for every next capture-snaphot we find in
guc-error-capture output buffer, i would find the matching engine and print out all
the VMA data (that was successfully captured in #2). However, i should only do that
for the first dump only since that would correlate exactly with the existing execlist
code behavior. So this fix is actually pretty straight forward to get the right matching
VMA.
WRT to my statement about "getting the context-to->process" lookup, i was initially hoping that
I could "on my own" (within the guc-err-capture module) get that information, but it would be
a stretch (in terms of inter-component information access). More importantly, its totally
unnecessary since existing execlist code already did that in Step 2. That code remains intact
with guc-error-capture.
One open i plan to test before final rev is with shared engines like CCS and RCS where i want to
trigger cascading hangs + resets in quick succession just to see how the overall flow behavior works.
I will attach an output guc error capture based gpu error dump as per the review comment from Matthew
on last rev.
..alan
>
> Regards,
>
> Tvrtko
>
> > (Recap: First, guc notifies capture event, second, guc notifies context reset during
> > which we trigger i915_gpu_coredump. In this second step, the vma's are dumped and we
> > verify that the guc capture happened but don't parse the guc-err-capture-logs yet.
> > Third step is when user triggers the debugfs to dump which is when we parse the error
> > capture logs.)
> >
> > As a fix, what we can do in the guc_error_capture report out is to ensure that
> > we dont re-print the previously dumped vmas if we end up finding multiple
> > guc-error-capture dumps since the i915_gpu_coredump would have only captured the vma's
> > for the very first context that was reset. And with guc-submission, that would always
> > correlate to the "next-yet-to-be-parsed" guc-err-capture dump (since the guc-error-capture
> > logs are large enough to hold data for multiple dumps).
> >
> > The changes (removal of below-hunk and adding of only-print-the-first-vma") is trivial
> > but i felt it warranted a good explanation. Apologies for the inbox noise.
> >
> > ...alan
> >
> > On Tue, 2021-12-07 at 22:32 -0800, Alan Previn Teres Alexis wrote:
> > > Thanks again for the detailed review here.
> > > Will fix all the rest on next rev.
> > > One special response for this one:
> > >
> > >
> > > On Tue, 2021-12-07 at 16:22 -0800, Matthew Brost wrote:
> > > > On Mon, Nov 22, 2021 at 03:04:02PM -0800, Alan Previn wrote:
> > > > > + if (datatype == GUC_CAPTURE_LIST_TYPE_ENGINE_INSTANCE) {
> > > > > + GCAP_PRINT_GUC_INST_INFO(i915, ebuf, data);
> > > > > + eng_inst = FIELD_GET(GUC_CAPTURE_DATAHDR_SRC_INSTANCE, data.info);
> > > > > + eng = guc_lookup_engine(guc, engineclass, eng_inst);
> > > > > + if (eng) {
> > > > > + GCAP_PRINT_INTEL_ENG_INFO(i915, ebuf, eng);
> > > > > + } else {
> > > > > + PRINT(&i915->drm, ebuf, " i915-Eng-Lookup Fail!\n");
> > > > > + }
> > > > > + ce = guc_context_lookup(guc, data.guc_ctx_id);
> > > >
> > > > You are going to need to reference count the 'ce' here. See
> > > > intel_guc_context_reset_process_msg for an example.
> > > >
> > >
> > > Oh crap - i missed this one - which you had explicitly mentioned offline when i was doing the
> > > development. Sorry about that i just totally missed it from my todo-notes.
> > >
> > > ...alan
