On 10/26/2021 23:36, Thomas Hellström wrote:
Hi, John,
On 10/26/21 21:55, John Harrison wrote:
On 10/21/2021 23:23, Thomas Hellström wrote:
On 10/21/21 22:37, Matthew Brost wrote:
On Thu, Oct 21, 2021 at 08:15:49AM +0200, Thomas Hellström wrote:
Hi, Matthew,
On Mon, 2021-10-11 at 16:47 -0700, Matthew Brost wrote:
The hangcheck selftest blocks per engine resets by setting magic
bits
in
the reset flags. This is incorrect for GuC submission because if the
GuC
fails to reset an engine we would like to do a full GT reset. Do no
set
these magic bits when using GuC submission.
Side note this lockless algorithm with magic bits to block resets
really
should be ripped out.
Lockless algorithm aside, from a quick look at the code in
intel_reset.c it appears to me like the interface that falls back
to a
full GT reset is intel_gt_handle_error() whereas intel_engine_reset()
is explicitly intended to not do that, so is there a discrepancy
between GuC and non-GuC here?
With GuC submission when an engine reset fails, we get an engine reset
failure notification which triggers a full GT reset
(intel_guc_engine_failure_process_msg in intel_guc_submission.c). That
reset is blocking by setting these magic bits. Clearing the bits in
this
function doesn't seem to unblock that reset either, the driver
tries to
unload with a worker blocked, and results in the blow up. Something
with
this lockless algorithm could be wrong as clear of the bit should
unlblock the reset but it is doesn't. We can look into that but in the
meantime we need to fix this test to be able to fail gracefully and
not
crash CI.
Yeah, for that lockless algorithm if needed, we might want to use a
ww_mutex per engine or something,
but point was that AFAICT at least one of the tests that set those
flags explicitly tested the functionality that no other engines than
the intended one was reset when the intel_engine_reset() function
was used, and then if GuC submission doesn't honor that, wouldn't a
better approach be to make a code comment around
intel_engine_reset() to explain the differences and disable that
particular test for GuC?. Also wouldn't we for example we see a
duplicated full GT reset with GuC if intel_engine_reset() fails as
part of the intel_gt_handle_error() function?
Re-reading this thread, I think there is a misunderstanding.
The selftests themselves have already been updated to support GuC
based engine resets. That is done by submitting a hanging context and
letting the GuC detect the hang and issue a reset. There is no
mechanism available for i915 to directly issue or request an engine
based reset (because i915 does not know what is running on any given
engine at any given time, being disconnected from the scheduler).
So the tests are already correctly testing per engine resets and do
not go anywhere near either intel_engine_reset() or
intel_gt_handle_error() when GuC submission is used. The problem is
what happens if the engine reset fails (which supposedly can only
happen with broken hardware). In that scenario, there is an
asynchronous message from GuC to i915 to notify us of the failure.
The KMD receives that notification and then (eventually) calls
intel_gt_handle_error() to issue a full GT reset. However, that is
blocked because the selftest is not expecting it and has vetoed the
possibility.
This is where my understanding of the discussion differs. According to
Matthew, the selftest actually proceeds to clear the bits, but the
worker that calls into intel_gt_handle_error() never wakes up. (and
that's probably due to clear_bit() being used instead of
clear_and_wake_up_bit()).
Hmm, missed that point. Yeah, sounds like the missing wake_up suffix is
what is causing the deadlock. I can't see any other reason why the reset
handler would not proceed once the flags are cleared. And it looks like
the selftest should timeout out waiting for the request and continue on
to clear the bits just fine.
And my problem with this particular patch is that it adds even more
"if (!guc_submission)" which is already sprinkled all over the place
in the selftests to the point that it becomes difficult to see what
(if anything) the tests are really testing.
I agree with this. Fixing the problem at source seems like a better
solution than hacking lots of different bits in different tests.
For example igt_reset_nop_engine() from a cursory look looks like it's
doing something but inside the engine loop it becomes clear that the
test doesn't do *anything* except iterate over engines. Same for
igt_reset_engines() in the !TEST_ACTIVE case and for
igt_reset_idle_engine(). For some other tests the reset_count checks
are gone, leaving only a test that we actually do a reset.
The nop_engine test is meant to be a no-op. It is testing what happens
when you reset an idle engine. That is not something we can do with GuC
based engine resets - there are no debug hooks into GuC specifically to
trigger an engine reset when there is no hang. The same applies to the
!TEST_ACTIVE case (which is igt_reset_idle_engine). Hence those are
skipped for GuC.
The reset_count tests are still present. They are looking for global
resets occurring when they should not. It is the reset_engine_count
check that is disabled for GuC. Again, because GuC is handling the reset
not i915. GuC tells us when a context is reset, it does not report any
specifics about engines. So this is not a count we can maintain when
using GuC submission.
So if possible, as previously mentioned, I think a solution without
adding more of this in the selftests is preferrable. To me the best
option is probably be the one you suggest in your previous email:
Don't wait on the I915_RESET_ENGINE bits with GuC in
intel_gt_handle_error(), (or perhaps extract what's left in a separate
function called from the GuC handler).
I don't like the idea of splitting intel_gt_handle_error() apart. It is
meant to be the top level reset handler that manages all resets - per
engine and full GT (and potentially FLR in the future). Certainly, none
of it should be pulled out into GuC specific code.
One less than ideal aspect of the current reset support is that
'intel_has_reset_engine()' is ambiguous. With GuC submission, we have
per engine reset. It's just not managed by i915. So code that wants to
know if engines can be reset individual wants a true response. But code
such as intel_gt_handle_error() needs to know *who* is doing that reset.
Hence it needs the GuC specific check on top to differentiate. Ideally,
there would be separate intel_has_reset_engine() and
intel_does_i915_do_engine_reset() queries. Then the reset code itself
would not need to query GuC submission. However, that seems like
overkill to remove a couple of GuC tests. Maybe it could be done as
future work, but it is certainly not in the scope of making this
selftest not hang the system.
John.
Thanks,
Thomas