Message ID | 20180517074055.14638-18-chris@chris-wilson.co.uk (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On 17/05/2018 08:40, Chris Wilson wrote: > Back in commit 27af5eea54d1 ("drm/i915: Move execlists irq handler to a > bottom half"), we came to the conclusion that running our CSB processing > and ELSP submission from inside the irq handler was a bad idea. A really > bad idea as we could impose nearly 1s latency on other users of the > system, on average! Deferring our work to a tasklet allowed us to do the > processing with irqs enabled, reducing the impact to an average of about > 50us. > > We have since eradicated the use of forcewaked mmio from inside the CSB > processing and ELSP submission, bringing the impact down to around 5us > (on Kabylake); an order of magnitude better than our measurements 2 > years ago on Broadwell and only about 2x worse on average than the > gem_syslatency on an unladen system. > > Comparing the impact on the maximum latency observed over a 120s interval, > repeated several times (using gem_syslatency, similar to RT's cyclictest) > while the system is fully laden with i915 nops, we see that direct > submission definitely worsens the response but not to the same outlandish > degree as before. > > x Unladen baseline > + Using tasklet > * Direct submission > > +------------------------------------------------------------------------+ > |xx x ++ +++ + * * * ** *** * *| > ||A| |__AM__| |_____A_M___| | > +------------------------------------------------------------------------+ What are these headers? This one and below, I cannot decipher them at all. > N Min Max Median Avg Stddev > x 10 5 18 10 9.3 3.6530049 > + 10 72 120 108 102.9 15.758243 > * 10 255 348 316 305.7 28.74814 In micro-seconds? so tasklet is 108us median? Direct submission 316us median? > > And with a background load This is IO background load? Regards, Tvrtko > > +------------------------------------------------------------------------+ > |x + * * | > |x + + + + + + +* * ** ++ * * * *| > |A |_______A_____|__|_______A___M______| | > +------------------------------------------------------------------------+ > N Min Max Median Avg Stddev > x 10 4 11 9 8.5 2.1730675 > + 10 633 1388 972 993 243.33744 > * 10 1152 2109 1608 1488.3 314.80719 > > References: 27af5eea54d1 ("drm/i915: Move execlists irq handler to a bottom half") > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > --- > drivers/gpu/drm/i915/i915_irq.c | 11 ++------ > drivers/gpu/drm/i915/intel_lrc.c | 44 +++++++++++++++++--------------- > 2 files changed, 26 insertions(+), 29 deletions(-) > > diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c > index 3f139ff64385..8b61ebf5cb4a 100644 > --- a/drivers/gpu/drm/i915/i915_irq.c > +++ b/drivers/gpu/drm/i915/i915_irq.c > @@ -1462,22 +1462,15 @@ static void snb_gt_irq_handler(struct drm_i915_private *dev_priv, > static void > gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir) > { > - bool tasklet = false; > - > - if (iir & GT_CONTEXT_SWITCH_INTERRUPT) { > + if (iir & GT_CONTEXT_SWITCH_INTERRUPT) > intel_engine_handle_execlists_irq(engine); > - tasklet = true; > - } > > if (iir & GT_RENDER_USER_INTERRUPT) { > if (intel_engine_uses_guc(engine)) > - tasklet = true; > + tasklet_hi_schedule(&engine->execlists.tasklet); > > notify_ring(engine); > } > - > - if (tasklet) > - tasklet_hi_schedule(&engine->execlists.tasklet); > } > > static void gen8_gt_irq_ack(struct drm_i915_private *i915, > diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c > index 954eb3a71051..37839d89e03a 100644 > --- a/drivers/gpu/drm/i915/intel_lrc.c > +++ b/drivers/gpu/drm/i915/intel_lrc.c > @@ -575,7 +575,7 @@ static void complete_preempt_context(struct intel_engine_execlists *execlists) > execlists_clear_active(execlists, EXECLISTS_ACTIVE_PREEMPT); > } > > -static void __execlists_dequeue(struct intel_engine_cs *engine) > +static void execlists_dequeue(struct intel_engine_cs *engine) > { > struct intel_engine_execlists * const execlists = &engine->execlists; > struct execlist_port *port = execlists->port; > @@ -587,7 +587,11 @@ static void __execlists_dequeue(struct intel_engine_cs *engine) > > lockdep_assert_held(&engine->timeline.lock); > > - /* Hardware submission is through 2 ports. Conceptually each port > + if (execlists_is_active(&engine->execlists, EXECLISTS_ACTIVE_PREEMPT)) > + return; > + > + /* > + * Hardware submission is through 2 ports. Conceptually each port > * has a (RING_START, RING_HEAD, RING_TAIL) tuple. RING_START is > * static for a context, and unique to each, so we only execute > * requests belonging to a single context from each ring. RING_HEAD > @@ -777,15 +781,6 @@ static void __execlists_dequeue(struct intel_engine_cs *engine) > !port_isset(engine->execlists.port)); > } > > -static void execlists_dequeue(struct intel_engine_cs *engine) > -{ > - unsigned long flags; > - > - spin_lock_irqsave(&engine->timeline.lock, flags); > - __execlists_dequeue(engine); > - spin_unlock_irqrestore(&engine->timeline.lock, flags); > -} > - > void > execlists_cancel_port_requests(struct intel_engine_execlists * const execlists) > { > @@ -1106,6 +1101,7 @@ void intel_engine_handle_execlists_irq(struct intel_engine_cs *engine) > execlists->csb_read); > execlists->csb_head = head; > > + execlists_dequeue(engine); > spin_unlock(&engine->timeline.lock); > } > > @@ -1122,8 +1118,9 @@ static void execlists_submission_tasklet(unsigned long data) > engine->i915->gt.awake, > engine->execlists.active); > > - if (!execlists_is_active(&engine->execlists, EXECLISTS_ACTIVE_PREEMPT)) > - execlists_dequeue(engine); > + spin_lock_irq(&engine->timeline.lock); > + execlists_dequeue(engine); > + spin_unlock_irq(&engine->timeline.lock); > } > > static void queue_request(struct intel_engine_cs *engine, > @@ -1134,16 +1131,20 @@ static void queue_request(struct intel_engine_cs *engine, > &lookup_priolist(engine, prio)->requests); > } > > -static void __submit_queue(struct intel_engine_cs *engine, int prio) > +static void __update_queue(struct intel_engine_cs *engine, int prio) > { > engine->execlists.queue_priority = prio; > - tasklet_hi_schedule(&engine->execlists.tasklet); > } > > static void submit_queue(struct intel_engine_cs *engine, int prio) > { > - if (prio > engine->execlists.queue_priority) > - __submit_queue(engine, prio); > + if (prio > engine->execlists.queue_priority) { > + __update_queue(engine, prio); > + if (!intel_engine_uses_guc(engine)) > + execlists_dequeue(engine); > + else > + tasklet_hi_schedule(&engine->execlists.tasklet); > + } > } > > static void execlists_submit_request(struct i915_request *request) > @@ -1155,11 +1156,12 @@ static void execlists_submit_request(struct i915_request *request) > spin_lock_irqsave(&engine->timeline.lock, flags); > > queue_request(engine, &request->sched, rq_prio(request)); > - submit_queue(engine, rq_prio(request)); > > GEM_BUG_ON(!engine->execlists.first); > GEM_BUG_ON(list_empty(&request->sched.link)); > > + submit_queue(engine, rq_prio(request)); > + > spin_unlock_irqrestore(&engine->timeline.lock, flags); > } > > @@ -1286,8 +1288,10 @@ static void execlists_schedule(struct i915_request *request, > } > > if (prio > engine->execlists.queue_priority && > - i915_sw_fence_done(&sched_to_request(node)->submit)) > - __submit_queue(engine, prio); > + i915_sw_fence_done(&sched_to_request(node)->submit)) { > + __update_queue(engine, prio); > + tasklet_hi_schedule(&engine->execlists.tasklet); > + } > } > > spin_unlock_irq(&engine->timeline.lock); >
Quoting Tvrtko Ursulin (2018-05-17 14:13:00) > > On 17/05/2018 08:40, Chris Wilson wrote: > > Back in commit 27af5eea54d1 ("drm/i915: Move execlists irq handler to a > > bottom half"), we came to the conclusion that running our CSB processing > > and ELSP submission from inside the irq handler was a bad idea. A really > > bad idea as we could impose nearly 1s latency on other users of the > > system, on average! Deferring our work to a tasklet allowed us to do the > > processing with irqs enabled, reducing the impact to an average of about > > 50us. > > > > We have since eradicated the use of forcewaked mmio from inside the CSB > > processing and ELSP submission, bringing the impact down to around 5us > > (on Kabylake); an order of magnitude better than our measurements 2 > > years ago on Broadwell and only about 2x worse on average than the > > gem_syslatency on an unladen system. > > > > Comparing the impact on the maximum latency observed over a 120s interval, > > repeated several times (using gem_syslatency, similar to RT's cyclictest) > > while the system is fully laden with i915 nops, we see that direct > > submission definitely worsens the response but not to the same outlandish > > degree as before. > > > > x Unladen baseline > > + Using tasklet > > * Direct submission > > > > +------------------------------------------------------------------------+ > > |xx x ++ +++ + * * * ** *** * *| > > ||A| |__AM__| |_____A_M___| | > > +------------------------------------------------------------------------+ > > What are these headers? This one and below, I cannot decipher them at all. Ministat histogram. The headers being the label for the charts; it's a bit flat so hard to tell it's a histogram. > > N Min Max Median Avg Stddev > > x 10 5 18 10 9.3 3.6530049 > > + 10 72 120 108 102.9 15.758243 > > * 10 255 348 316 305.7 28.74814 > > In micro-seconds? so tasklet is 108us median? Direct submission 316us > median? Yup, more runs required so you have prettier graphs and units. > > And with a background load > > This is IO background load? Yes, background writeout. -Chris
On 17/05/2018 18:07, Chris Wilson wrote: > Quoting Tvrtko Ursulin (2018-05-17 14:13:00) >> >> On 17/05/2018 08:40, Chris Wilson wrote: >>> Back in commit 27af5eea54d1 ("drm/i915: Move execlists irq handler to a >>> bottom half"), we came to the conclusion that running our CSB processing >>> and ELSP submission from inside the irq handler was a bad idea. A really >>> bad idea as we could impose nearly 1s latency on other users of the >>> system, on average! Deferring our work to a tasklet allowed us to do the >>> processing with irqs enabled, reducing the impact to an average of about >>> 50us. >>> >>> We have since eradicated the use of forcewaked mmio from inside the CSB >>> processing and ELSP submission, bringing the impact down to around 5us >>> (on Kabylake); an order of magnitude better than our measurements 2 >>> years ago on Broadwell and only about 2x worse on average than the >>> gem_syslatency on an unladen system. >>> >>> Comparing the impact on the maximum latency observed over a 120s interval, >>> repeated several times (using gem_syslatency, similar to RT's cyclictest) >>> while the system is fully laden with i915 nops, we see that direct >>> submission definitely worsens the response but not to the same outlandish >>> degree as before. >>> >>> x Unladen baseline >>> + Using tasklet >>> * Direct submission >>> >>> +------------------------------------------------------------------------+ >>> |xx x ++ +++ + * * * ** *** * *| >>> ||A| |__AM__| |_____A_M___| | >>> +------------------------------------------------------------------------+ >> >> What are these headers? This one and below, I cannot decipher them at all. > > Ministat histogram. The headers being the label for the charts; it's a > bit flat so hard to tell it's a histogram. > >>> N Min Max Median Avg Stddev >>> x 10 5 18 10 9.3 3.6530049 >>> + 10 72 120 108 102.9 15.758243 >>> * 10 255 348 316 305.7 28.74814 >> >> In micro-seconds? so tasklet is 108us median? Direct submission 316us >> median? > > Yup, more runs required so you have prettier graphs and units. Biggest problem for me is that in these tests it is only showing as significantly worse than the current tasklet. So it is kind of difficult the sell the series. :) If it only solves issues when submitting from a RT task then it is not so interesting. RT tasks are rare and which deal with 3d/media/ocl probably even rarer. Or you know of any? Or perhaps you need to add data from gem_latency as well if that shows some wins purely on the i915 side? Regards, Tvrtko >>> And with a background load >> >> This is IO background load? > > Yes, background writeout. > -Chris >
Quoting Tvrtko Ursulin (2018-05-18 09:06:03) > > On 17/05/2018 18:07, Chris Wilson wrote: > > Quoting Tvrtko Ursulin (2018-05-17 14:13:00) > >> > >> On 17/05/2018 08:40, Chris Wilson wrote: > >>> Back in commit 27af5eea54d1 ("drm/i915: Move execlists irq handler to a > >>> bottom half"), we came to the conclusion that running our CSB processing > >>> and ELSP submission from inside the irq handler was a bad idea. A really > >>> bad idea as we could impose nearly 1s latency on other users of the > >>> system, on average! Deferring our work to a tasklet allowed us to do the > >>> processing with irqs enabled, reducing the impact to an average of about > >>> 50us. > >>> > >>> We have since eradicated the use of forcewaked mmio from inside the CSB > >>> processing and ELSP submission, bringing the impact down to around 5us > >>> (on Kabylake); an order of magnitude better than our measurements 2 > >>> years ago on Broadwell and only about 2x worse on average than the > >>> gem_syslatency on an unladen system. > >>> > >>> Comparing the impact on the maximum latency observed over a 120s interval, > >>> repeated several times (using gem_syslatency, similar to RT's cyclictest) > >>> while the system is fully laden with i915 nops, we see that direct > >>> submission definitely worsens the response but not to the same outlandish > >>> degree as before. > >>> > >>> x Unladen baseline > >>> + Using tasklet > >>> * Direct submission > >>> > >>> +------------------------------------------------------------------------+ > >>> |xx x ++ +++ + * * * ** *** * *| > >>> ||A| |__AM__| |_____A_M___| | > >>> +------------------------------------------------------------------------+ > >> > >> What are these headers? This one and below, I cannot decipher them at all. > > > > Ministat histogram. The headers being the label for the charts; it's a > > bit flat so hard to tell it's a histogram. > > > >>> N Min Max Median Avg Stddev > >>> x 10 5 18 10 9.3 3.6530049 > >>> + 10 72 120 108 102.9 15.758243 > >>> * 10 255 348 316 305.7 28.74814 > >> > >> In micro-seconds? so tasklet is 108us median? Direct submission 316us > >> median? > > > > Yup, more runs required so you have prettier graphs and units. > > Biggest problem for me is that in these tests it is only showing as > significantly worse than the current tasklet. So it is kind of difficult > the sell the series. :) Ba! Compared to the previous best state 2 years ago, we're an order magnitude better. (Unbelievable!) I think it's simply a lack of a reference point, we're about 100% faster in some microbenchmark stress igts, about 10% faster in regular stress igts, and will not suffer some of the 200ms timeout -> wedged! Any latency is bad, but at what point do we worry about the trade-off between our latency and the rest of the system? > If it only solves issues when submitting from a RT task then it is not > so interesting. RT tasks are rare and which deal with 3d/media/ocl > probably even rarer. Or you know of any? No, it's not just an issue from RT. We see examples of the driver being so starved we declare the system is wedged, to much more mundane artifacts of our throughput being diminished due to the latency in submitting work. > Or perhaps you need to add data from gem_latency as well if that shows > some wins purely on the i915 side? The patch before (process CSB from interrupt handler) is justifiable just by the bugs it solves. This patch, I think is justifiable by our perf gains and the impact it has for low latency requirements. Or we can wait until the next patch is mature, that tries to reduce the lock contention by splitting the data structures. -Chris
Quoting Tvrtko Ursulin (2018-05-18 09:06:03) > > On 17/05/2018 18:07, Chris Wilson wrote: > > Quoting Tvrtko Ursulin (2018-05-17 14:13:00) > >> > >> On 17/05/2018 08:40, Chris Wilson wrote: > >>> Back in commit 27af5eea54d1 ("drm/i915: Move execlists irq handler to a > >>> bottom half"), we came to the conclusion that running our CSB processing > >>> and ELSP submission from inside the irq handler was a bad idea. A really > >>> bad idea as we could impose nearly 1s latency on other users of the > >>> system, on average! Deferring our work to a tasklet allowed us to do the > >>> processing with irqs enabled, reducing the impact to an average of about > >>> 50us. > >>> > >>> We have since eradicated the use of forcewaked mmio from inside the CSB > >>> processing and ELSP submission, bringing the impact down to around 5us > >>> (on Kabylake); an order of magnitude better than our measurements 2 > >>> years ago on Broadwell and only about 2x worse on average than the > >>> gem_syslatency on an unladen system. > >>> > >>> Comparing the impact on the maximum latency observed over a 120s interval, > >>> repeated several times (using gem_syslatency, similar to RT's cyclictest) > >>> while the system is fully laden with i915 nops, we see that direct > >>> submission definitely worsens the response but not to the same outlandish > >>> degree as before. > >>> > >>> x Unladen baseline > >>> + Using tasklet > >>> * Direct submission > >>> > >>> +------------------------------------------------------------------------+ > >>> |xx x ++ +++ + * * * ** *** * *| > >>> ||A| |__AM__| |_____A_M___| | > >>> +------------------------------------------------------------------------+ > >> > >> What are these headers? This one and below, I cannot decipher them at all. > > > > Ministat histogram. The headers being the label for the charts; it's a > > bit flat so hard to tell it's a histogram. > > > >>> N Min Max Median Avg Stddev > >>> x 10 5 18 10 9.3 3.6530049 > >>> + 10 72 120 108 102.9 15.758243 > >>> * 10 255 348 316 305.7 28.74814 > >> > >> In micro-seconds? so tasklet is 108us median? Direct submission 316us > >> median? > > > > Yup, more runs required so you have prettier graphs and units. > > Biggest problem for me is that in these tests it is only showing as > significantly worse than the current tasklet. So it is kind of difficult > the sell the series. :) As a reference point, on bdw from before we introduced tasklets: gem_syslatency: cycles=20849458, latency mean=678.892us max=8308500us gem_syslatency: cycles=22964600, latency mean=1583.312us max=5991894us gem_syslatency: cycles=21404190, latency mean=1766.220us max=12423925us gem_syslatency: cycles=22779405, latency mean=1698.282us max=9110117us gem_syslatency: cycles=22021655, latency mean=1921.855us max=7193398us Message from syslogd@broadwell at May 18 18:27:47 ... kernel:[ 8991.394488] NMI watchdog: BUG: soft lockup - CPU#2 stuck for 22s! [gem_syslatency:1800] gem_syslatency: cycles=21218381, latency mean=1711.088us max=13932087us gem_syslatency: cycles=22189636, latency mean=1724.968us max=15483237us gem_syslatency: cycles=21941275, latency mean=1697.738us max=6065099us gem_syslatency: cycles=21896529, latency mean=1748.536us max=8007063us gem_syslatency: cycles=22053588, latency mean=1677.465us max=6604678us gem_syslatency: cycles=21657859, latency mean=1570.085us max=10346811us gem_syslatency: cycles=21627270, latency mean=1227.782us max=10489759us gem_syslatency: cycles=22441025, latency mean=1635.776us max=5932223us gem_syslatency: cycles=21890354, latency mean=1790.042us max=7956373us gem_syslatency: cycles=20976620, latency mean=989.215us max=11631632us gem_syslatency: cycles=22087242, latency mean=1723.138us max=12337920us gem_syslatency: cycles=22800746, latency mean=1749.050us max=7080445us * at this point, the sata driver ate itself and took out the fs. -Chris
Quoting Tvrtko Ursulin (2018-05-17 14:13:00) > > On 17/05/2018 08:40, Chris Wilson wrote: > > Back in commit 27af5eea54d1 ("drm/i915: Move execlists irq handler to a > > bottom half"), we came to the conclusion that running our CSB processing > > and ELSP submission from inside the irq handler was a bad idea. A really > > bad idea as we could impose nearly 1s latency on other users of the > > system, on average! Deferring our work to a tasklet allowed us to do the > > processing with irqs enabled, reducing the impact to an average of about > > 50us. > > > > We have since eradicated the use of forcewaked mmio from inside the CSB > > processing and ELSP submission, bringing the impact down to around 5us > > (on Kabylake); an order of magnitude better than our measurements 2 > > years ago on Broadwell and only about 2x worse on average than the > > gem_syslatency on an unladen system. > > > > Comparing the impact on the maximum latency observed over a 120s interval, > > repeated several times (using gem_syslatency, similar to RT's cyclictest) > > while the system is fully laden with i915 nops, we see that direct > > submission definitely worsens the response but not to the same outlandish > > degree as before. > > > > x Unladen baseline > > + Using tasklet > > * Direct submission > > > > +------------------------------------------------------------------------+ > > |xx x ++ +++ + * * * ** *** * *| > > ||A| |__AM__| |_____A_M___| | > > +------------------------------------------------------------------------+ > > What are these headers? This one and below, I cannot decipher them at all. > > > N Min Max Median Avg Stddev > > x 10 5 18 10 9.3 3.6530049 > > + 10 72 120 108 102.9 15.758243 > > * 10 255 348 316 305.7 28.74814 > > In micro-seconds? so tasklet is 108us median? Direct submission 316us > median? Perspective, ivb ringbuf: x syslatency-ringbuf-mean.txt + syslatency-ringbuf-max.txt +------------------------------------------------------------------------+ | xx + | | xx + | | xx + | | xx ++ | | xxxx + + + ++ + ++ | |x xxxx + + + ++ + ++ ++++ ++++ +| | |AM |__________A__M_______| | +------------------------------------------------------------------------+ N Min Max Median Avg Stddev x 30 5.395 17.893 13.626 13.135367 2.2143809 + 30 33 169 143 135.86667 25.609445 Using execlists+tasklet is on par with irqoff (i.e. max) latency of ringbuf. (There should be very little irq disabling for ringbuf.) Normal average latency of execlists+direct_submission is still better than ivb, but that is almost entirely generational differences. -Chris
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c index 3f139ff64385..8b61ebf5cb4a 100644 --- a/drivers/gpu/drm/i915/i915_irq.c +++ b/drivers/gpu/drm/i915/i915_irq.c @@ -1462,22 +1462,15 @@ static void snb_gt_irq_handler(struct drm_i915_private *dev_priv, static void gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir) { - bool tasklet = false; - - if (iir & GT_CONTEXT_SWITCH_INTERRUPT) { + if (iir & GT_CONTEXT_SWITCH_INTERRUPT) intel_engine_handle_execlists_irq(engine); - tasklet = true; - } if (iir & GT_RENDER_USER_INTERRUPT) { if (intel_engine_uses_guc(engine)) - tasklet = true; + tasklet_hi_schedule(&engine->execlists.tasklet); notify_ring(engine); } - - if (tasklet) - tasklet_hi_schedule(&engine->execlists.tasklet); } static void gen8_gt_irq_ack(struct drm_i915_private *i915, diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c index 954eb3a71051..37839d89e03a 100644 --- a/drivers/gpu/drm/i915/intel_lrc.c +++ b/drivers/gpu/drm/i915/intel_lrc.c @@ -575,7 +575,7 @@ static void complete_preempt_context(struct intel_engine_execlists *execlists) execlists_clear_active(execlists, EXECLISTS_ACTIVE_PREEMPT); } -static void __execlists_dequeue(struct intel_engine_cs *engine) +static void execlists_dequeue(struct intel_engine_cs *engine) { struct intel_engine_execlists * const execlists = &engine->execlists; struct execlist_port *port = execlists->port; @@ -587,7 +587,11 @@ static void __execlists_dequeue(struct intel_engine_cs *engine) lockdep_assert_held(&engine->timeline.lock); - /* Hardware submission is through 2 ports. Conceptually each port + if (execlists_is_active(&engine->execlists, EXECLISTS_ACTIVE_PREEMPT)) + return; + + /* + * Hardware submission is through 2 ports. Conceptually each port * has a (RING_START, RING_HEAD, RING_TAIL) tuple. RING_START is * static for a context, and unique to each, so we only execute * requests belonging to a single context from each ring. RING_HEAD @@ -777,15 +781,6 @@ static void __execlists_dequeue(struct intel_engine_cs *engine) !port_isset(engine->execlists.port)); } -static void execlists_dequeue(struct intel_engine_cs *engine) -{ - unsigned long flags; - - spin_lock_irqsave(&engine->timeline.lock, flags); - __execlists_dequeue(engine); - spin_unlock_irqrestore(&engine->timeline.lock, flags); -} - void execlists_cancel_port_requests(struct intel_engine_execlists * const execlists) { @@ -1106,6 +1101,7 @@ void intel_engine_handle_execlists_irq(struct intel_engine_cs *engine) execlists->csb_read); execlists->csb_head = head; + execlists_dequeue(engine); spin_unlock(&engine->timeline.lock); } @@ -1122,8 +1118,9 @@ static void execlists_submission_tasklet(unsigned long data) engine->i915->gt.awake, engine->execlists.active); - if (!execlists_is_active(&engine->execlists, EXECLISTS_ACTIVE_PREEMPT)) - execlists_dequeue(engine); + spin_lock_irq(&engine->timeline.lock); + execlists_dequeue(engine); + spin_unlock_irq(&engine->timeline.lock); } static void queue_request(struct intel_engine_cs *engine, @@ -1134,16 +1131,20 @@ static void queue_request(struct intel_engine_cs *engine, &lookup_priolist(engine, prio)->requests); } -static void __submit_queue(struct intel_engine_cs *engine, int prio) +static void __update_queue(struct intel_engine_cs *engine, int prio) { engine->execlists.queue_priority = prio; - tasklet_hi_schedule(&engine->execlists.tasklet); } static void submit_queue(struct intel_engine_cs *engine, int prio) { - if (prio > engine->execlists.queue_priority) - __submit_queue(engine, prio); + if (prio > engine->execlists.queue_priority) { + __update_queue(engine, prio); + if (!intel_engine_uses_guc(engine)) + execlists_dequeue(engine); + else + tasklet_hi_schedule(&engine->execlists.tasklet); + } } static void execlists_submit_request(struct i915_request *request) @@ -1155,11 +1156,12 @@ static void execlists_submit_request(struct i915_request *request) spin_lock_irqsave(&engine->timeline.lock, flags); queue_request(engine, &request->sched, rq_prio(request)); - submit_queue(engine, rq_prio(request)); GEM_BUG_ON(!engine->execlists.first); GEM_BUG_ON(list_empty(&request->sched.link)); + submit_queue(engine, rq_prio(request)); + spin_unlock_irqrestore(&engine->timeline.lock, flags); } @@ -1286,8 +1288,10 @@ static void execlists_schedule(struct i915_request *request, } if (prio > engine->execlists.queue_priority && - i915_sw_fence_done(&sched_to_request(node)->submit)) - __submit_queue(engine, prio); + i915_sw_fence_done(&sched_to_request(node)->submit)) { + __update_queue(engine, prio); + tasklet_hi_schedule(&engine->execlists.tasklet); + } } spin_unlock_irq(&engine->timeline.lock);
Back in commit 27af5eea54d1 ("drm/i915: Move execlists irq handler to a bottom half"), we came to the conclusion that running our CSB processing and ELSP submission from inside the irq handler was a bad idea. A really bad idea as we could impose nearly 1s latency on other users of the system, on average! Deferring our work to a tasklet allowed us to do the processing with irqs enabled, reducing the impact to an average of about 50us. We have since eradicated the use of forcewaked mmio from inside the CSB processing and ELSP submission, bringing the impact down to around 5us (on Kabylake); an order of magnitude better than our measurements 2 years ago on Broadwell and only about 2x worse on average than the gem_syslatency on an unladen system. Comparing the impact on the maximum latency observed over a 120s interval, repeated several times (using gem_syslatency, similar to RT's cyclictest) while the system is fully laden with i915 nops, we see that direct submission definitely worsens the response but not to the same outlandish degree as before. x Unladen baseline + Using tasklet * Direct submission +------------------------------------------------------------------------+ |xx x ++ +++ + * * * ** *** * *| ||A| |__AM__| |_____A_M___| | +------------------------------------------------------------------------+ N Min Max Median Avg Stddev x 10 5 18 10 9.3 3.6530049 + 10 72 120 108 102.9 15.758243 * 10 255 348 316 305.7 28.74814 And with a background load +------------------------------------------------------------------------+ |x + * * | |x + + + + + + +* * ** ++ * * * *| |A |_______A_____|__|_______A___M______| | +------------------------------------------------------------------------+ N Min Max Median Avg Stddev x 10 4 11 9 8.5 2.1730675 + 10 633 1388 972 993 243.33744 * 10 1152 2109 1608 1488.3 314.80719 References: 27af5eea54d1 ("drm/i915: Move execlists irq handler to a bottom half") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> --- drivers/gpu/drm/i915/i915_irq.c | 11 ++------ drivers/gpu/drm/i915/intel_lrc.c | 44 +++++++++++++++++--------------- 2 files changed, 26 insertions(+), 29 deletions(-)