Re: HEADS UP: Merging drm update

[Ryo ONODERA <ryo%tetera.org@localhost>]

Hi,

Taylor R Campbell <riastradh%NetBSD.org@localhost> writes:

>> Date: Tue, 28 Dec 2021 11:34:43 +0900
>> From: Ryo ONODERA <ryo%tetera.org@localhost>
>> 
>> intel_gt_pm_fini() at netbsd:intel_gt_pm_fini+0x18
>> intel_gt_init() at netbsd:intel_gt_init+0x6ad
>> i915_gem_init() at netbsd:i915_gem_init+0x14d
>> i915_driver_probe() at netbsd:i915_driver_probe+0x949
>> i915drmkms_attach_real() at netbsd:i915drmkms_attach_real+0x4c
>> config_mountroot_thread() at netbsd:config_mountroot_thread+0x60
>
> So intel_gt_init is failing on boot, and the driver has decided to
> give up -- and proximate cause of the crash is that one of the error
> branches is screwy, but while it would be nice to fix the error
> branches it's more important to find why we're reaching them in the
> first place.
>
> Can you get a line number for intel_gt_init+0x6ad, and can you also
> insert prints into every error branch of intel_gt_init to find out
> which one it is and how it fails?  And maybe do that recursively in
> whichever branch does fail?

In sys/external/bsd/drm2/dist/drm/i915/gt/intel_gt.c: intel_gt_init(),
__engines_record_defaults(gt) failed and went to err_gt label,
then the panic happened.
"intel_gt_init+0x6ad" is err_uc_init's intel_uc_fini(&gt->uc).

(snip)
        err = __engines_record_defaults(gt);
        if (err)
                goto err_gt;
(snip)
err_gt:
        __intel_gt_disable(gt);
        intel_uc_fini_hw(&gt->uc);
err_uc_init:
        intel_uc_fini(&gt->uc);
err_engines:
        intel_engines_release(gt);
        i915_vm_put(fetch_and_zero(&gt->vm));
err_pm:
        intel_gt_pm_fini(gt);
        intel_gt_fini_scratch(gt);
out_fw:
        if (err)
(snip)


And I have added some printfs to __engines_record_defaults() and
the other functions invoked from __engines_record_defaults() as follows.

__engines_record_defaults
intel_gt_wait_for_idle
intel_gt_retire_requests_timeout
dma_fence_wait_timeout
i915_fence_wait (via *fence->ops->wait)
i915_request_wait

In i915_request_wait, DRM_SPIN_TIMED_WAIT_UNTIL sets timeout=0
and i915_request_wait returns timeout=-ETIME.

#ifdef __NetBSD__
        spin_lock(rq->fence.lock);
#define C       (i915_request_completed(rq) ? 1 :                             \
                    (spin_unlock(rq->fence.lock),                             \
                        intel_engine_flush_submission(rq->engine),            \
                        spin_lock(rq->fence.lock),                            \
                        i915_request_completed(rq)))
        if (flags & I915_WAIT_INTERRUPTIBLE) {
                DRM_SPIN_TIMED_WAIT_UNTIL(timeout, &wait.wq,
                    rq->fence.lock, timeout,
                    C);
        } else {
                DRM_SPIN_TIMED_WAIT_NOINTR_UNTIL(timeout, &wait.wq,
                    rq->fence.lock, timeout,
                    C);
        }
#undef  C
        if (timeout > 0) {      /* succeeded before timeout */
                KASSERT(i915_request_completed(rq));
                dma_fence_signal_locked(&rq->fence);
        } else if (timeout == 0) {      /* timed out */
                timeout = -ETIME;
        }
        spin_unlock(rq->fence.lock);
        DRM_DESTROY_WAITQUEUE(&wait.wq);
#else


Thank you.

-- 
Ryo ONODERA // ryo%tetera.org@localhost
PGP fingerprint = 82A2 DC91 76E0 A10A 8ABB  FD1B F404 27FA C7D1 15F3