[Intel-gfx] [PATCH] drm/i915: Disable per-engine reset for Broxton

[Chris Wilson]

Quoting Michel Thierry (2017-07-06 02:24:26)
> On 04/07/17 09:09, Chris Wilson wrote:
> > Triggering a GPU reset for one engine affects another, notably
> > corrupting the context status buffer (CSB) effectively losing track of
> > inflight requests.
> >
> > Adding a few printks:
> > diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
> > index ad41836fa5e5..a969456bc0fa 100644
> > --- a/drivers/gpu/drm/i915/i915_drv.c
> > +++ b/drivers/gpu/drm/i915/i915_drv.c
> > @@ -1953,6 +1953,7 @@ int i915_reset_engine(struct intel_engine_cs *engine)
> >                 goto out;
> >         }
> >
> > +       pr_err("Resetting %s\n", engine->name);
> >         ret = intel_gpu_reset(engine->i915, intel_engine_flag(engine));
> >         if (ret) {
> >                 /* If we fail here, we expect to fallback to a global reset */
> > diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
> > index 716e5c9ea222..a72bc35d0870 100644
> > --- a/drivers/gpu/drm/i915/intel_lrc.c
> > +++ b/drivers/gpu/drm/i915/intel_lrc.c
> > @@ -355,6 +355,7 @@ static void execlists_submit_ports(struct intel_engine_cs *engine)
> >                                 execlists_context_status_change(rq, INTEL_CONTEXT_SCHEDULE_IN);
> >                         port_set(&port[n], port_pack(rq, count));
> >                         desc = execlists_update_context(rq);
> > +                       pr_err("%s: in (rq=%x) ctx=%d\n", engine->name, rq->global_seqno, upper_32_bits(desc));
> >                         GEM_DEBUG_EXEC(port[n].context_id = upper_32_bits(desc));
> >                 } else {
> >                         GEM_BUG_ON(!n);
> > @@ -594,9 +595,23 @@ static void intel_lrc_irq_handler(unsigned long data)
> >                         if (!(status & GEN8_CTX_STATUS_COMPLETED_MASK))
> >                                 continue;
> >
> > +                       pr_err("%s: out CSB (%x head=%d, tail=%d), ctx=%d, rq=%d\n",
> > +                                       engine->name,
> > +                                       readl(csb_mmio),
> > +                                       head, tail,
> > +                                       readl(buf+2*head+1),
> > +                                       port->context_id);
> > +
> >                         /* Check the context/desc id for this event matches */
> > -                       GEM_DEBUG_BUG_ON(readl(buf + 2 * head + 1) !=
> > -                                        port->context_id);
> > +                       if (readl(buf + 2 * head + 1) != port->context_id) {
> > +                               pr_err("%s: BUG CSB (%x head=%d, tail=%d), ctx=%d, rq=%d\n",
> > +                                               engine->name,
> > +                                               readl(csb_mmio),
> > +                                               head, tail,
> > +                                               readl(buf+2*head+1),
> > +                                               port->context_id);
> > +                               BUG();
> > +                       }
> >
> >                         rq = port_unpack(port, &count);
> >                         GEM_BUG_ON(count == 0);
> >
> > Results in:
> >
> > [ 6423.006602] Resetting rcs0
> > [ 6423.009080] rcs0: in (rq=fffffe70) ctx=1
> > [ 6423.009216] rcs0: in (rq=fffffe6f) ctx=3
> > [ 6423.009542] rcs0: out CSB (2 head=1, tail=2), ctx=3, rq=3
> > [ 6423.009619] Resetting bcs0
> > [ 6423.009980] rcs0: BUG CSB (0 head=1, tail=2), ctx=0, rq=3
> >
> 
> It took me a while, but  I was able to replicate this (with your 
> igt_reset_active_engines) a couple of times. I also captured the value 
> of the CSB events at that point and it looked like this.

Ah, that's a separate issue that definitely isn't limited to bxt. In the
bug on my machine the CSB is distinctly zeroed.
 
> [   55.134393] Resetting rcs0
> [   55.134747] bcs0: BUG CSB (3 head=1, tail=2), ctx=10, rq=4
> [   55.134755]  bcs0: HWSP[16-17] Execlist CSB[0]:   0x00000018 _ 0x0000000a
> [   55.134759]  bcs0: HWSP[18-19] Execlist CSB[1]:   0x00000012 _ 0x0000000a
> [   55.134762]  bcs0: HWSP[20-21] Execlist CSB[2]:   0x00008002 _ 0x00000004
> [   55.134765]  bcs0: HWSP[22-23] Execlist CSB[3]:   0x00000014 _ 0x00000004
> [   55.134767]  bcs0: HWSP[24-25] Execlist CSB[4]:   0x00000018 _ 0x0000000a
> [   55.134770]  bcs0: HWSP[26-27] Execlist CSB[5]: 0x00000001 _ 0x00000000
> 
> The problem is ctx 10 finished in CSB[0] (ctx_complete & 
> active_to_idle), but then somehow CSB[1] has the same ctx 10 with 
> 'preempted' & ctx_complete.
> 
> To make things worse, in CSB[2], the hw claims to have lite-restored ctx 4.
> 
> So it seems, we could ignore events like CSB[1], i.e. preempted && 
> ctx_complete && !lite_restore.

I think this bug is fixed in the series I sent last week, this is just
the race in reset vs the tasklet.
-Chris