Re: [Intel-gfx] [PATCH RFC] drm/i915/gt: Retry RING_HEAD reset until it sticks

[Rodrigo Vivi <rodrigo.vivi@intel.com>]

On Fri, Jul 15, 2022 at 09:26:16AM +0100, Mauro Carvalho Chehab wrote:
> From: Chris Wilson <chris@chris-wilson.co.uk>
> 
> On Haswell, in particular, we see an issue where resets fails because

Can we then make this platform specific?
Only because some older hw doesn't behave like expected we shouldn't
make this a default & global workaround.

> the engine resumes from an incorrect RING_HEAD. Since the RING_HEAD
> doesn't point to the remaining requests to re-run, but may instead point
> into the uninitialised portion of the ring, the GPU may be then fed
> invalid instructions from a privileged context, oft pushing the GPU into
> an unrecoverable hang.
> 
> If at first the write doesn't succeed, try, try again.
> 
> References: https://gitlab.freedesktop.org/drm/intel/-/issues/5432
> Testcase: igt/i915_selftest/hangcheck
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: Andrzej Hajda <andrzej.hajda@intel.com>
> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
> Signed-off-by: Mauro Carvalho Chehab <mchehab@kernel.org>
> ---
>  .../gpu/drm/i915/gt/intel_ring_submission.c   | 44 +++++++++++++------
>  drivers/gpu/drm/i915/i915_utils.h             | 10 +++++
>  2 files changed, 40 insertions(+), 14 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/gt/intel_ring_submission.c b/drivers/gpu/drm/i915/gt/intel_ring_submission.c
> index d5d6f1fadcae..cc53feb1f8ed 100644
> --- a/drivers/gpu/drm/i915/gt/intel_ring_submission.c
> +++ b/drivers/gpu/drm/i915/gt/intel_ring_submission.c
> @@ -190,6 +190,7 @@ static bool stop_ring(struct intel_engine_cs *engine)
>  static int xcs_resume(struct intel_engine_cs *engine)
>  {
>  	struct intel_ring *ring = engine->legacy.ring;
> +	ktime_t kt;
>  
>  	ENGINE_TRACE(engine, "ring:{HEAD:%04x, TAIL:%04x}\n",
>  		     ring->head, ring->tail);
> @@ -228,9 +229,20 @@ static int xcs_resume(struct intel_engine_cs *engine)
>  	set_pp_dir(engine);
>  
>  	/* First wake the ring up to an empty/idle ring */
> -	ENGINE_WRITE_FW(engine, RING_HEAD, ring->head);
> +	until_timeout_ns(kt, 2 * NSEC_PER_MSEC) {
> +		ENGINE_WRITE_FW(engine, RING_HEAD, ring->head);
> +		if (ENGINE_READ_FW(engine, RING_HEAD) == ring->head)
> +			break;
> +	}
> +
>  	ENGINE_WRITE_FW(engine, RING_TAIL, ring->head);
> -	ENGINE_POSTING_READ(engine, RING_TAIL);
> +	if (ENGINE_READ_FW(engine, RING_HEAD) != ENGINE_READ_FW(engine, RING_TAIL)) {
> +		ENGINE_TRACE(engine, "failed to reset empty ring: [%x, %x]: %x\n",
> +			     ENGINE_READ_FW(engine, RING_HEAD),
> +			     ENGINE_READ_FW(engine, RING_TAIL),
> +			     ring->head);
> +		goto err;
> +	}

commit message mentions until this point I'm afraid... everything below
(except the new until_timeout_ns) looks like a different patch to me,
or deserves some mention in the commit msg.

>  
>  	ENGINE_WRITE_FW(engine, RING_CTL,
>  			RING_CTL_SIZE(ring->size) | RING_VALID);
> @@ -239,12 +251,16 @@ static int xcs_resume(struct intel_engine_cs *engine)
>  	if (__intel_wait_for_register_fw(engine->uncore,
>  					 RING_CTL(engine->mmio_base),
>  					 RING_VALID, RING_VALID,
> -					 5000, 0, NULL))
> +					 5000, 0, NULL)) {
> +		ENGINE_TRACE(engine, "failed to restart\n");
>  		goto err;
> +	}
>  
> -	if (GRAPHICS_VER(engine->i915) > 2)
> +	if (GRAPHICS_VER(engine->i915) > 2) {
>  		ENGINE_WRITE_FW(engine,
>  				RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
> +		ENGINE_POSTING_READ(engine, RING_MI_MODE);
> +	}
>  
>  	/* Now awake, let it get started */
>  	if (ring->tail != ring->head) {
> @@ -257,16 +273,16 @@ static int xcs_resume(struct intel_engine_cs *engine)
>  	return 0;
>  
>  err:
> -	drm_err(&engine->i915->drm,
> -		"%s initialization failed; "
> -		"ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
> -		engine->name,
> -		ENGINE_READ(engine, RING_CTL),
> -		ENGINE_READ(engine, RING_CTL) & RING_VALID,
> -		ENGINE_READ(engine, RING_HEAD), ring->head,
> -		ENGINE_READ(engine, RING_TAIL), ring->tail,
> -		ENGINE_READ(engine, RING_START),
> -		i915_ggtt_offset(ring->vma));
> +	ENGINE_TRACE(engine,
> +		     "initialization failed; "
> +		     "ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
> +		     ENGINE_READ(engine, RING_CTL),
> +		     ENGINE_READ(engine, RING_CTL) & RING_VALID,
> +		     ENGINE_READ(engine, RING_HEAD), ring->head,
> +		     ENGINE_READ(engine, RING_TAIL), ring->tail,
> +		     ENGINE_READ(engine, RING_START),
> +		     i915_ggtt_offset(ring->vma));
> +	GEM_TRACE_DUMP();
>  	return -EIO;
>  }
>  
> diff --git a/drivers/gpu/drm/i915/i915_utils.h b/drivers/gpu/drm/i915/i915_utils.h
> index c10d68cdc3ca..717fb6b9cc15 100644
> --- a/drivers/gpu/drm/i915/i915_utils.h
> +++ b/drivers/gpu/drm/i915/i915_utils.h
> @@ -256,6 +256,16 @@ wait_remaining_ms_from_jiffies(unsigned long timestamp_jiffies, int to_wait_ms)
>  	}
>  }
>  
> +/**
> + * until_timeout_ns - Keep retrying (busy spin) until the duration has passed
> + * @end: temporary var to be used to track the spent time
> + * @timeout_ns: Maximum timeout, in nanosseconds
> + */
> +#define until_timeout_ns(end, timeout_ns) \
> +	for ((end) = ktime_get() + (timeout_ns); \
> +	     ktime_before(ktime_get(), (end)); \
> +	     cpu_relax())
> +

why do we need yet another timeout macro and cannot use any of the existent ways?

>  /**
>   * __wait_for - magic wait macro
>   *
> -- 
> 2.36.1
> 
>