[Intel-gfx] [RFC] drm/i915/guc: capture GuC logs if FW fails to load

[Daniele Ceraolo Spurio]

On 04/05/17 14:31, Chris Wilson wrote:
> On Thu, May 04, 2017 at 09:26:35PM +0000, Srivatsa, Anusha wrote:
>>> +void i915_guc_load_error_log_capture(struct drm_i915_private *i915) {
>>> +	void *log, *buf;
>>> +	struct i915_vma *vma = i915->guc.log.vma;
>>> +
>>> +	if (i915->gpu_error.guc_load_fail_log || !vma)
>>> +		return;
>>> +
>>> +	/*
>>> +	 * the vma should be already pinned and mapped for log runtime
>>> +	 * management but let's play safe
>>> +	 */
>>> +	log = i915_gem_object_pin_map(vma->obj, I915_MAP_WC);
>>> +	if (IS_ERR(log)) {
>>> +		DRM_ERROR("Failed to pin guc_log vma\n");
>>> +		return;
>>> +	}
>>> +
>>> +	buf = kzalloc(GUC_LOG_SIZE, GFP_KERNEL);
>>> +	if (buf) {
>>> +		memcpy(buf, log, GUC_LOG_SIZE);
>>> +		i915->gpu_error.guc_load_fail_log = buf;
>>> +	} else {
>>> +		DRM_ERROR("Failed to copy guc log\n");
>>> +	}
>>> +
>>> +	i915_gem_object_unpin_map(vma->obj);
>
> You are trading a swappable object for unswappable kernel memory. If you
> want to have the guc log after guc is disabled, just keep the log object
> around.
> -Chris
>

I had considered that, but in the end I wasn't sure if that was 
acceptable in case we end up modifying the code to recycle the object 
for future load attempts, although that is very unlikely. I was however 
unconvinced myself of using kzalloc and that's mainly why this was an RFC :)
I'll flip it to take an extra reference on the object.

Thanks,
Daniele