[Intel-gfx] a potential dead loop in intel_lrc_irq_handler

[Chris Wilson]

Quoting Dong, Chuanxiao (2017-08-07 11:31:57)
> > -----Original Message-----
> > From: Chris Wilson [mailto:chris at chris-wilson.co.uk]
> > GPU reset -> clears CSB head/tail
> But the GPU reset will make CSB_head = 0 and CSB_tail = 7.

Experience says otherwise, but the issue of the delayed interrupt is
still a concern.

On the per-engine reset path, interrupts are not disabled, so once we
disable the tasklet and cancel the pending execution across the reset, we
should be fine.

It's the full reset path, where we disable the interrupt first, we need
to be careful about the hw keeping the interrupt around. Certainly we
have no fundamental reason to disable_irq there, the supposition is that
we are safer without spurious interrupts we aren't ready to handle. But
afaics to put your mind at rest all we need is an

	I915_WRITE(GEN8_GT_IIR[engine->irq_idx],
		   GT_CONTEXT_SWITCH_INTERRUPT << engine->irq_shift);
	I915_WRITE(GEN8_GT_IIR[engine->irq_idx],
		   GT_CONTEXT_SWITCH_INTERRUPT << engine->irq_shift);

as we process the reset.
-Chris