[Intel-gfx] [PATCH 14/18] drm/i915/execlists: Direct submission of new requests (avoid tasklet/ksoftirqd)
Chris Wilson
chris at chris-wilson.co.uk
Fri May 25 09:32:02 UTC 2018
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.
In this iteration of the tasklet-vs-direct submission debate, we seek a
compromise where by we submit new requests immediately to the HW but
defer processing the CS interrupt onto a tasklet. We gain the advantage
of low-latency and ksoftirqd avoidance when waking up the HW, while
avoiding the system-wide starvation of our CS irq-storms.
Comparing the impact on the maximum latency observed (that is the time stolen
from an RT process) 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 an actually improve
the worse case.
Maximum latency in microseconds of a third party RT thread
(gem_syslatency -t 120 -f 2)
x Always using tasklets (a couple of >1000us outliers removed)
+ Only using tasklets from CS irq, direct submission of requests
+------------------------------------------------------------------------+
| + |
| + |
| + |
| + + |
| + + + |
| + + + + x x x |
| +++ + + + x x x x x x |
| +++ + ++ + + *x x x x x x |
| +++ + ++ + * *x x * x x x |
| + +++ + ++ * * +*xxx * x x xx |
| * +++ + ++++* *x+**xx+ * x x xxxx x |
| **x++++*++**+*x*x****x+ * +x xx xxxx x x |
|x* ******+***************++*+***xxxxxx* xx*x xxx + x+|
| |__________MA___________| |
| |______M__A________| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 118 91 186 124 125.28814 16.279137
+ 120 92 187 109 112.00833 13.458617
Difference at 95.0% confidence
-13.2798 +/- 3.79219
-10.5994% +/- 3.02677%
(Student's t, pooled s = 14.9237)
However the mean latency is adversely affected:
Mean latency in microseconds of a third party RT thread
(gem_syslatency -t 120 -f 1)
x Always using tasklets
+ Only using tasklets from CS irq, direct submission of requests
+------------------------------------------------------------------------+
| xxxxxx + ++ |
| xxxxxx + ++ |
| xxxxxx + +++ ++ |
| xxxxxxx +++++ ++ |
| xxxxxxx +++++ ++ |
| xxxxxxx +++++ +++ |
| xxxxxxx + ++++++++++ |
| xxxxxxxx ++ ++++++++++ |
| xxxxxxxx ++ ++++++++++ |
| xxxxxxxxxx +++++++++++++++ |
| xxxxxxxxxxx x +++++++++++++++ |
|x xxxxxxxxxxxxx x + + ++++++++++++++++++ +|
| |__A__| |
| |____A___| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 120 3.506 3.727 3.631 3.6321417 0.02773109
+ 120 3.834 4.149 4.039 4.0375167 0.041221676
Difference at 95.0% confidence
0.405375 +/- 0.00888913
11.1608% +/- 0.244735%
(Student's t, pooled s = 0.03513)
However, since the mean latency corresponds to the amount of irqsoff
processing we have to do for a CS interrupt, we only need to speed that
up to benefit not just system latency but our own throughput.
v2: Remember to defer submissions when under reset.
v4: Only use direct submission for new requests
Testcase: igt/gem_exec_latency/*rthog*
References: 27af5eea54d1 ("drm/i915: Move execlists irq handler to a bottom half")
Suggested-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
---
drivers/gpu/drm/i915/i915_gem.h | 5 +
drivers/gpu/drm/i915/i915_irq.c | 11 +-
drivers/gpu/drm/i915/intel_engine_cs.c | 8 +-
drivers/gpu/drm/i915/intel_lrc.c | 147 ++++++++++++++----------
drivers/gpu/drm/i915/intel_ringbuffer.h | 1 -
5 files changed, 98 insertions(+), 74 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_gem.h b/drivers/gpu/drm/i915/i915_gem.h
index 261da577829a..7892ac773916 100644
--- a/drivers/gpu/drm/i915/i915_gem.h
+++ b/drivers/gpu/drm/i915/i915_gem.h
@@ -88,4 +88,9 @@ static inline void __tasklet_enable_sync_once(struct tasklet_struct *t)
tasklet_kill(t);
}
+static inline bool __tasklet_is_enabled(const struct tasklet_struct *t)
+{
+ return likely(!atomic_read(&t->count));
+}
+
#endif /* __I915_GEM_H__ */
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
index f9bc3aaa90d0..3a8eacab5172 100644
--- a/drivers/gpu/drm/i915/i915_irq.c
+++ b/drivers/gpu/drm/i915/i915_irq.c
@@ -1462,14 +1462,10 @@ 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)
{
- struct intel_engine_execlists * const execlists = &engine->execlists;
bool tasklet = false;
- if (iir & GT_CONTEXT_SWITCH_INTERRUPT) {
- if (READ_ONCE(engine->execlists.active))
- tasklet = !test_and_set_bit(ENGINE_IRQ_EXECLIST,
- &engine->irq_posted);
- }
+ if (iir & GT_CONTEXT_SWITCH_INTERRUPT)
+ tasklet = true;
if (iir & GT_RENDER_USER_INTERRUPT) {
notify_ring(engine);
@@ -1477,7 +1473,7 @@ gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir)
}
if (tasklet)
- tasklet_hi_schedule(&execlists->tasklet);
+ tasklet_hi_schedule(&engine->execlists.tasklet);
}
static void gen8_gt_irq_ack(struct drm_i915_private *i915,
@@ -2185,7 +2181,6 @@ static irqreturn_t cherryview_irq_handler(int irq, void *arg)
I915_WRITE(VLV_IER, ier);
I915_WRITE(GEN8_MASTER_IRQ, GEN8_MASTER_IRQ_CONTROL);
- POSTING_READ(GEN8_MASTER_IRQ);
gen8_gt_irq_handler(dev_priv, master_ctl, gt_iir);
diff --git a/drivers/gpu/drm/i915/intel_engine_cs.c b/drivers/gpu/drm/i915/intel_engine_cs.c
index bc0193199a03..43880bf0f529 100644
--- a/drivers/gpu/drm/i915/intel_engine_cs.c
+++ b/drivers/gpu/drm/i915/intel_engine_cs.c
@@ -1329,12 +1329,10 @@ static void intel_engine_print_registers(const struct intel_engine_cs *engine,
ptr = I915_READ(RING_CONTEXT_STATUS_PTR(engine));
read = GEN8_CSB_READ_PTR(ptr);
write = GEN8_CSB_WRITE_PTR(ptr);
- drm_printf(m, "\tExeclist CSB read %d [%d cached], write %d [%d from hws], interrupt posted? %s, tasklet queued? %s (%s)\n",
+ drm_printf(m, "\tExeclist CSB read %d [%d cached], write %d [%d from hws], tasklet queued? %s (%s)\n",
read, execlists->csb_head,
write,
intel_read_status_page(engine, intel_hws_csb_write_index(engine->i915)),
- yesno(test_bit(ENGINE_IRQ_EXECLIST,
- &engine->irq_posted)),
yesno(test_bit(TASKLET_STATE_SCHED,
&engine->execlists.tasklet.state)),
enableddisabled(!atomic_read(&engine->execlists.tasklet.count)));
@@ -1515,11 +1513,9 @@ void intel_engine_dump(struct intel_engine_cs *engine,
spin_unlock(&b->rb_lock);
local_irq_restore(flags);
- drm_printf(m, "IRQ? 0x%lx (breadcrumbs? %s) (execlists? %s)\n",
+ drm_printf(m, "IRQ? 0x%lx (breadcrumbs? %s)\n",
engine->irq_posted,
yesno(test_bit(ENGINE_IRQ_BREADCRUMB,
- &engine->irq_posted)),
- yesno(test_bit(ENGINE_IRQ_EXECLIST,
&engine->irq_posted)));
drm_printf(m, "HWSP:\n");
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index 71693008b5d0..c88ea807945a 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -557,13 +557,15 @@ static void complete_preempt_context(struct intel_engine_execlists *execlists)
{
GEM_BUG_ON(!execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT));
+ __unwind_incomplete_requests(container_of(execlists,
+ typeof(struct intel_engine_cs),
+ execlists));
execlists_cancel_port_requests(execlists);
- execlists_unwind_incomplete_requests(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;
@@ -575,7 +577,11 @@ static void __execlists_dequeue(struct intel_engine_cs *engine)
lockdep_assert_held(&engine->timeline.lock);
- /* Hardware submission is through 2 ports. Conceptually each port
+ GEM_BUG_ON(execlists_is_active(&engine->execlists,
+ EXECLISTS_ACTIVE_PREEMPT));
+
+ /*
+ * 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
@@ -765,15 +771,6 @@ 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)
{
@@ -870,14 +867,6 @@ static void reset_irq(struct intel_engine_cs *engine)
synchronize_hardirq(engine->i915->drm.irq);
clear_gtiir(engine);
-
- /*
- * The port is checked prior to scheduling a tasklet, but
- * just in case we have suspended the tasklet to do the
- * wedging make sure that when it wakes, it decides there
- * is no work to do by clearing the irq_posted bit.
- */
- clear_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
}
static void execlists_cancel_requests(struct intel_engine_cs *engine)
@@ -950,6 +939,12 @@ static void execlists_cancel_requests(struct intel_engine_cs *engine)
local_irq_restore(flags);
}
+static inline bool
+reset_in_progress(const struct intel_engine_execlists *execlists)
+{
+ return unlikely(!__tasklet_is_enabled(&execlists->tasklet));
+}
+
static void process_csb(struct intel_engine_cs *engine)
{
struct intel_engine_execlists * const execlists = &engine->execlists;
@@ -957,10 +952,6 @@ static void process_csb(struct intel_engine_cs *engine)
const u32 * const buf = execlists->csb_status;
u8 head, tail;
- /* Clear before reading to catch new interrupts */
- clear_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
- smp_mb__after_atomic();
-
/* Note that csb_write, csb_status may be either in HWSP or mmio */
head = execlists->csb_head;
tail = READ_ONCE(*execlists->csb_write);
@@ -1090,19 +1081,9 @@ static void process_csb(struct intel_engine_cs *engine)
execlists->csb_head = head;
}
-/*
- * Check the unread Context Status Buffers and manage the submission of new
- * contexts to the ELSP accordingly.
- */
-static void execlists_submission_tasklet(unsigned long data)
+static void __execlists_submission_tasklet(struct intel_engine_cs *const engine)
{
- struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
-
- GEM_TRACE("%s awake?=%d, active=%x, irq-posted?=%d\n",
- engine->name,
- engine->i915->gt.awake,
- engine->execlists.active,
- test_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted));
+ lockdep_assert_held(&engine->timeline.lock);
/*
* We can skip acquiring intel_runtime_pm_get() here as it was taken
@@ -1114,18 +1095,32 @@ static void execlists_submission_tasklet(unsigned long data)
*/
GEM_BUG_ON(!engine->i915->gt.awake);
- /*
- * Prefer doing test_and_clear_bit() as a two stage operation to avoid
- * imposing the cost of a locked atomic transaction when submitting a
- * new request (outside of the context-switch interrupt).
- */
- if (test_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted))
- process_csb(engine);
-
+ process_csb(engine);
if (!execlists_is_active(&engine->execlists, EXECLISTS_ACTIVE_PREEMPT))
execlists_dequeue(engine);
}
+/*
+ * Check the unread Context Status Buffers and manage the submission of new
+ * contexts to the ELSP accordingly.
+ */
+static void execlists_submission_tasklet(unsigned long data)
+{
+ struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
+ unsigned long flags;
+
+ GEM_TRACE("%s awake?=%d, active=%x\n",
+ engine->name,
+ engine->i915->gt.awake,
+ engine->execlists.active);
+
+ spin_lock_irqsave(&engine->timeline.lock, flags);
+
+ __execlists_submission_tasklet(engine);
+
+ spin_unlock_irqrestore(&engine->timeline.lock, flags);
+}
+
static void queue_request(struct intel_engine_cs *engine,
struct i915_sched_node *node,
int prio)
@@ -1134,16 +1129,30 @@ 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)
+{
+ struct intel_engine_execlists * const execlists = &engine->execlists;
+
+ if (reset_in_progress(execlists))
+ return; /* defer until we restart the engine following reset */
+
+ if (execlists->tasklet.func == execlists_submission_tasklet)
+ __execlists_submission_tasklet(engine);
+ else
+ tasklet_hi_schedule(&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);
+ __submit_queue(engine);
+ }
}
static void execlists_submit_request(struct i915_request *request)
@@ -1155,11 +1164,12 @@ 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 +1296,11 @@ 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)) {
+ /* defer submission until after all of our updates */
+ __update_queue(engine, prio);
+ tasklet_hi_schedule(&engine->execlists.tasklet);
+ }
}
spin_unlock_irq(&engine->timeline.lock);
@@ -1831,6 +1844,7 @@ execlists_reset_prepare(struct intel_engine_cs *engine)
{
struct intel_engine_execlists * const execlists = &engine->execlists;
struct i915_request *request, *active;
+ unsigned long flags;
GEM_TRACE("%s\n", engine->name);
@@ -1854,8 +1868,10 @@ execlists_reset_prepare(struct intel_engine_cs *engine)
* and avoid blaming an innocent request if the stall was due to the
* preemption itself.
*/
- if (test_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted))
- process_csb(engine);
+ spin_lock_irqsave(&engine->timeline.lock, flags);
+ synchronize_hardirq(engine->i915->drm.irq);
+
+ process_csb(engine);
/*
* The last active request can then be no later than the last request
@@ -1865,15 +1881,12 @@ execlists_reset_prepare(struct intel_engine_cs *engine)
active = NULL;
request = port_request(execlists->port);
if (request) {
- unsigned long flags;
-
/*
* Prevent the breadcrumb from advancing before we decide
* which request is currently active.
*/
intel_engine_stop_cs(engine);
- spin_lock_irqsave(&engine->timeline.lock, flags);
list_for_each_entry_from_reverse(request,
&engine->timeline.requests,
link) {
@@ -1883,12 +1896,28 @@ execlists_reset_prepare(struct intel_engine_cs *engine)
active = request;
}
- spin_unlock_irqrestore(&engine->timeline.lock, flags);
}
+ spin_unlock_irqrestore(&engine->timeline.lock, flags);
+
return active;
}
+static void reset_csb_pointers(struct intel_engine_execlists *execlists)
+{
+ /*
+ * After a reset, the HW starts writing into CSB entry [0]. We
+ * therefore have to set our HEAD pointer back one entry so that
+ * the *first* entry we check is entry 0. To complicate this further,
+ * as we don't wait for the first interrupt after reset, we have to
+ * fake the HW write to point back to the last entry so that our
+ * inline comparison of our cached head position against the last HW
+ * write works even before the first interrupt.
+ */
+ execlists->csb_head = GEN8_CSB_ENTRIES - 1;
+ WRITE_ONCE(*execlists->csb_write, (GEN8_CSB_ENTRIES - 1) | 0xff << 16);
+}
+
static void execlists_reset(struct intel_engine_cs *engine,
struct i915_request *request)
{
@@ -1918,7 +1947,7 @@ static void execlists_reset(struct intel_engine_cs *engine,
__unwind_incomplete_requests(engine);
/* Following the reset, we need to reload the CSB read/write pointers */
- engine->execlists.csb_head = GEN8_CSB_ENTRIES - 1;
+ reset_csb_pointers(&engine->execlists);
spin_unlock_irqrestore(&engine->timeline.lock, flags);
@@ -2412,7 +2441,6 @@ static int logical_ring_init(struct intel_engine_cs *engine)
upper_32_bits(ce->lrc_desc);
}
- execlists->csb_head = GEN8_CSB_ENTRIES - 1;
execlists->csb_read =
i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine));
if (csb_force_mmio(i915)) {
@@ -2427,6 +2455,7 @@ static int logical_ring_init(struct intel_engine_cs *engine)
execlists->csb_write =
&engine->status_page.page_addr[intel_hws_csb_write_index(i915)];
}
+ reset_csb_pointers(execlists);
return 0;
diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.h b/drivers/gpu/drm/i915/intel_ringbuffer.h
index b92e76c23843..636a1b1ac0dc 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.h
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.h
@@ -358,7 +358,6 @@ struct intel_engine_cs {
atomic_t irq_count;
unsigned long irq_posted;
#define ENGINE_IRQ_BREADCRUMB 0
-#define ENGINE_IRQ_EXECLIST 1
/* Rather than have every client wait upon all user interrupts,
* with the herd waking after every interrupt and each doing the
--
2.17.0
More information about the Intel-gfx
mailing list