diff options
author | Chris Wilson <chris@chris-wilson.co.uk> | 2016-07-01 19:23:15 +0300 |
---|---|---|
committer | Chris Wilson <chris@chris-wilson.co.uk> | 2016-07-01 22:58:43 +0300 |
commit | 688e6c7258164de86d626e8e983ca8d28015c263 (patch) | |
tree | ea2040fd06199f2335f7431ca2506cd4e7757fe5 /drivers/gpu | |
parent | 1f15b76f1ec973d1eb5d21b6d98b21aebb9025f1 (diff) | |
download | linux-688e6c7258164de86d626e8e983ca8d28015c263.tar.xz |
drm/i915: Slaughter the thundering i915_wait_request herd
One particularly stressful scenario consists of many independent tasks
all competing for GPU time and waiting upon the results (e.g. realtime
transcoding of many, many streams). One bottleneck in particular is that
each client waits on its own results, but every client is woken up after
every batchbuffer - hence the thunder of hooves as then every client must
do its heavyweight dance to read a coherent seqno to see if it is the
lucky one.
Ideally, we only want one client to wake up after the interrupt and
check its request for completion. Since the requests must retire in
order, we can select the first client on the oldest request to be woken.
Once that client has completed his wait, we can then wake up the
next client and so on. However, all clients then incur latency as every
process in the chain may be delayed for scheduling - this may also then
cause some priority inversion. To reduce the latency, when a client
is added or removed from the list, we scan the tree for completed
seqno and wake up all the completed waiters in parallel.
Using igt/benchmarks/gem_latency, we can demonstrate this effect. The
benchmark measures the number of GPU cycles between completion of a
batch and the client waking up from a call to wait-ioctl. With many
concurrent waiters, with each on a different request, we observe that
the wakeup latency before the patch scales nearly linearly with the
number of waiters (before external factors kick in making the scaling much
worse). After applying the patch, we can see that only the single waiter
for the request is being woken up, providing a constant wakeup latency
for every operation. However, the situation is not quite as rosy for
many waiters on the same request, though to the best of my knowledge this
is much less likely in practice. Here, we can observe that the
concurrent waiters incur extra latency from being woken up by the
solitary bottom-half, rather than directly by the interrupt. This
appears to be scheduler induced (having discounted adverse effects from
having a rbtree walk/erase in the wakeup path), each additional
wake_up_process() costs approximately 1us on big core. Another effect of
performing the secondary wakeups from the first bottom-half is the
incurred delay this imposes on high priority threads - rather than
immediately returning to userspace and leaving the interrupt handler to
wake the others.
To offset the delay incurred with additional waiters on a request, we
could use a hybrid scheme that did a quick read in the interrupt handler
and dequeued all the completed waiters (incurring the overhead in the
interrupt handler, not the best plan either as we then incur GPU
submission latency) but we would still have to wake up the bottom-half
every time to do the heavyweight slow read. Or we could only kick the
waiters on the seqno with the same priority as the current task (i.e. in
the realtime waiter scenario, only it is woken up immediately by the
interrupt and simply queues the next waiter before returning to userspace,
minimising its delay at the expense of the chain, and also reducing
contention on its scheduler runqueue). This is effective at avoid long
pauses in the interrupt handler and at avoiding the extra latency in
realtime/high-priority waiters.
v2: Convert from a kworker per engine into a dedicated kthread for the
bottom-half.
v3: Rename request members and tweak comments.
v4: Use a per-engine spinlock in the breadcrumbs bottom-half.
v5: Fix race in locklessly checking waiter status and kicking the task on
adding a new waiter.
v6: Fix deciding when to force the timer to hide missing interrupts.
v7: Move the bottom-half from the kthread to the first client process.
v8: Reword a few comments
v9: Break the busy loop when the interrupt is unmasked or has fired.
v10: Comments, unnecessary churn, better debugging from Tvrtko
v11: Wake all completed waiters on removing the current bottom-half to
reduce the latency of waking up a herd of clients all waiting on the
same request.
v12: Rearrange missed-interrupt fault injection so that it works with
igt/drv_missed_irq_hang
v13: Rename intel_breadcrumb and friends to intel_wait in preparation
for signal handling.
v14: RCU commentary, assert_spin_locked
v15: Hide BUG_ON behind the compiler; report on gem_latency findings.
v16: Sort seqno-groups by priority so that first-waiter has the highest
task priority (and so avoid priority inversion).
v17: Add waiters to post-mortem GPU hang state.
v18: Return early for a completed wait after acquiring the spinlock.
Avoids adding ourselves to the tree if the is already complete, and
skips the awkward question of why we don't do completion wakeups for
waits earlier than or equal to ourselves.
v19: Prepare for init_breadcrumbs to fail. Later patches may want to
allocate during init, so be prepared to propagate back the error code.
Testcase: igt/gem_concurrent_blit
Testcase: igt/benchmarks/gem_latency
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com>
Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Cc: Dave Gordon <david.s.gordon@intel.com>
Cc: "Goel, Akash" <akash.goel@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18
Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
Diffstat (limited to 'drivers/gpu')
-rw-r--r-- | drivers/gpu/drm/i915/Makefile | 1 | ||||
-rw-r--r-- | drivers/gpu/drm/i915/i915_debugfs.c | 16 | ||||
-rw-r--r-- | drivers/gpu/drm/i915/i915_drv.h | 40 | ||||
-rw-r--r-- | drivers/gpu/drm/i915/i915_gem.c | 143 | ||||
-rw-r--r-- | drivers/gpu/drm/i915/i915_gpu_error.c | 60 | ||||
-rw-r--r-- | drivers/gpu/drm/i915/i915_irq.c | 20 | ||||
-rw-r--r-- | drivers/gpu/drm/i915/intel_breadcrumbs.c | 380 | ||||
-rw-r--r-- | drivers/gpu/drm/i915/intel_lrc.c | 7 | ||||
-rw-r--r-- | drivers/gpu/drm/i915/intel_ringbuffer.c | 12 | ||||
-rw-r--r-- | drivers/gpu/drm/i915/intel_ringbuffer.h | 79 |
10 files changed, 648 insertions, 110 deletions
diff --git a/drivers/gpu/drm/i915/Makefile b/drivers/gpu/drm/i915/Makefile index a3d2b789c9ec..618293c8c9d9 100644 --- a/drivers/gpu/drm/i915/Makefile +++ b/drivers/gpu/drm/i915/Makefile @@ -38,6 +38,7 @@ i915-y += i915_cmd_parser.o \ i915_gem_userptr.o \ i915_gpu_error.o \ i915_trace_points.o \ + intel_breadcrumbs.o \ intel_lrc.o \ intel_mocs.o \ intel_ringbuffer.o \ diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c index f664884ad27f..f522a366223d 100644 --- a/drivers/gpu/drm/i915/i915_debugfs.c +++ b/drivers/gpu/drm/i915/i915_debugfs.c @@ -788,10 +788,22 @@ static int i915_gem_request_info(struct seq_file *m, void *data) static void i915_ring_seqno_info(struct seq_file *m, struct intel_engine_cs *engine) { + struct intel_breadcrumbs *b = &engine->breadcrumbs; + struct rb_node *rb; + seq_printf(m, "Current sequence (%s): %x\n", engine->name, engine->get_seqno(engine)); seq_printf(m, "Current user interrupts (%s): %x\n", engine->name, READ_ONCE(engine->user_interrupts)); + + spin_lock(&b->lock); + for (rb = rb_first(&b->waiters); rb; rb = rb_next(rb)) { + struct intel_wait *w = container_of(rb, typeof(*w), node); + + seq_printf(m, "Waiting (%s): %s [%d] on %x\n", + engine->name, w->tsk->comm, w->tsk->pid, w->seqno); + } + spin_unlock(&b->lock); } static int i915_gem_seqno_info(struct seq_file *m, void *data) @@ -1428,6 +1440,8 @@ static int i915_hangcheck_info(struct seq_file *m, void *unused) engine->hangcheck.seqno, seqno[id], engine->last_submitted_seqno); + seq_printf(m, "\twaiters? %d\n", + intel_engine_has_waiter(engine)); seq_printf(m, "\tuser interrupts = %x [current %x]\n", engine->hangcheck.user_interrupts, READ_ONCE(engine->user_interrupts)); @@ -2415,7 +2429,7 @@ static int count_irq_waiters(struct drm_i915_private *i915) int count = 0; for_each_engine(engine, i915) - count += engine->irq_refcount; + count += intel_engine_has_waiter(engine); return count; } diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 0d0e4ac4dadb..1fefa8c495f2 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -506,6 +506,7 @@ struct drm_i915_error_state { bool valid; /* Software tracked state */ bool waiting; + int num_waiters; int hangcheck_score; enum intel_ring_hangcheck_action hangcheck_action; int num_requests; @@ -551,6 +552,12 @@ struct drm_i915_error_state { u32 tail; } *requests; + struct drm_i915_error_waiter { + char comm[TASK_COMM_LEN]; + pid_t pid; + u32 seqno; + } *waiters; + struct { u32 gfx_mode; union { @@ -1429,7 +1436,7 @@ struct i915_gpu_error { #define I915_STOP_RING_ALLOW_WARN (1 << 30) /* For missed irq/seqno simulation. */ - unsigned int test_irq_rings; + unsigned long test_irq_rings; }; enum modeset_restore { @@ -3064,7 +3071,6 @@ ibx_disable_display_interrupt(struct drm_i915_private *dev_priv, uint32_t bits) ibx_display_interrupt_update(dev_priv, bits, 0); } - /* i915_gem.c */ int i915_gem_create_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv); @@ -3975,4 +3981,34 @@ static inline void i915_trace_irq_get(struct intel_engine_cs *engine, i915_gem_request_assign(&engine->trace_irq_req, req); } +static inline bool __i915_request_irq_complete(struct drm_i915_gem_request *req) +{ + /* Ensure our read of the seqno is coherent so that we + * do not "miss an interrupt" (i.e. if this is the last + * request and the seqno write from the GPU is not visible + * by the time the interrupt fires, we will see that the + * request is incomplete and go back to sleep awaiting + * another interrupt that will never come.) + * + * Strictly, we only need to do this once after an interrupt, + * but it is easier and safer to do it every time the waiter + * is woken. + */ + if (i915_gem_request_completed(req, false)) + return true; + + /* We need to check whether any gpu reset happened in between + * the request being submitted and now. If a reset has occurred, + * the seqno will have been advance past ours and our request + * is complete. If we are in the process of handling a reset, + * the request is effectively complete as the rendering will + * be discarded, but we need to return in order to drop the + * struct_mutex. + */ + if (i915_reset_in_progress(&req->i915->gpu_error)) + return true; + + return false; +} + #endif diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index b5278d117ea0..c9814572e346 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -1343,17 +1343,6 @@ i915_gem_check_wedge(unsigned reset_counter, bool interruptible) return 0; } -static void fake_irq(unsigned long data) -{ - wake_up_process((struct task_struct *)data); -} - -static bool missed_irq(struct drm_i915_private *dev_priv, - struct intel_engine_cs *engine) -{ - return test_bit(engine->id, &dev_priv->gpu_error.missed_irq_rings); -} - static unsigned long local_clock_us(unsigned *cpu) { unsigned long t; @@ -1386,7 +1375,7 @@ static bool busywait_stop(unsigned long timeout, unsigned cpu) return this_cpu != cpu; } -static int __i915_spin_request(struct drm_i915_gem_request *req, int state) +static bool __i915_spin_request(struct drm_i915_gem_request *req, int state) { unsigned long timeout; unsigned cpu; @@ -1401,17 +1390,14 @@ static int __i915_spin_request(struct drm_i915_gem_request *req, int state) * takes to sleep on a request, on the order of a microsecond. */ - if (req->engine->irq_refcount) - return -EBUSY; - /* Only spin if we know the GPU is processing this request */ if (!i915_gem_request_started(req, true)) - return -EAGAIN; + return false; timeout = local_clock_us(&cpu) + 5; - while (!need_resched()) { + do { if (i915_gem_request_completed(req, true)) - return 0; + return true; if (signal_pending_state(state, current)) break; @@ -1420,12 +1406,9 @@ static int __i915_spin_request(struct drm_i915_gem_request *req, int state) break; cpu_relax_lowlatency(); - } - - if (i915_gem_request_completed(req, false)) - return 0; + } while (!need_resched()); - return -EAGAIN; + return false; } /** @@ -1450,18 +1433,14 @@ int __i915_wait_request(struct drm_i915_gem_request *req, s64 *timeout, struct intel_rps_client *rps) { - struct intel_engine_cs *engine = i915_gem_request_get_engine(req); - struct drm_i915_private *dev_priv = req->i915; - const bool irq_test_in_progress = - ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_engine_flag(engine); int state = interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE; DEFINE_WAIT(reset); - DEFINE_WAIT(wait); - unsigned long timeout_expire; + struct intel_wait wait; + unsigned long timeout_remain; s64 before = 0; /* Only to silence a compiler warning. */ - int ret; + int ret = 0; - WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled"); + might_sleep(); if (list_empty(&req->list)) return 0; @@ -1469,7 +1448,7 @@ int __i915_wait_request(struct drm_i915_gem_request *req, if (i915_gem_request_completed(req, true)) return 0; - timeout_expire = 0; + timeout_remain = MAX_SCHEDULE_TIMEOUT; if (timeout) { if (WARN_ON(*timeout < 0)) return -EINVAL; @@ -1477,7 +1456,7 @@ int __i915_wait_request(struct drm_i915_gem_request *req, if (*timeout == 0) return -ETIME; - timeout_expire = jiffies + nsecs_to_jiffies_timeout(*timeout); + timeout_remain = nsecs_to_jiffies_timeout(*timeout); /* * Record current time in case interrupted by signal, or wedged. @@ -1485,55 +1464,32 @@ int __i915_wait_request(struct drm_i915_gem_request *req, before = ktime_get_raw_ns(); } - if (INTEL_INFO(dev_priv)->gen >= 6) - gen6_rps_boost(dev_priv, rps, req->emitted_jiffies); - trace_i915_gem_request_wait_begin(req); - /* Optimistic spin for the next jiffie before touching IRQs */ - ret = __i915_spin_request(req, state); - if (ret == 0) - goto out; + if (INTEL_INFO(req->i915)->gen >= 6) + gen6_rps_boost(req->i915, rps, req->emitted_jiffies); - if (!irq_test_in_progress && WARN_ON(!engine->irq_get(engine))) { - ret = -ENODEV; - goto out; - } + /* Optimistic spin for the next ~jiffie before touching IRQs */ + if (__i915_spin_request(req, state)) + goto complete; - add_wait_queue(&dev_priv->gpu_error.wait_queue, &reset); - for (;;) { - struct timer_list timer; + set_current_state(state); + add_wait_queue(&req->i915->gpu_error.wait_queue, &reset); - prepare_to_wait(&engine->irq_queue, &wait, state); - - /* We need to check whether any gpu reset happened in between - * the request being submitted and now. If a reset has occurred, - * the seqno will have been advance past ours and our request - * is complete. If we are in the process of handling a reset, - * the request is effectively complete as the rendering will - * be discarded, but we need to return in order to drop the - * struct_mutex. + intel_wait_init(&wait, req->seqno); + if (intel_engine_add_wait(req->engine, &wait)) + /* In order to check that we haven't missed the interrupt + * as we enabled it, we need to kick ourselves to do a + * coherent check on the seqno before we sleep. */ - if (i915_reset_in_progress(&dev_priv->gpu_error)) { - ret = 0; - break; - } - - if (i915_gem_request_completed(req, false)) { - ret = 0; - break; - } + goto wakeup; + for (;;) { if (signal_pending_state(state, current)) { ret = -ERESTARTSYS; break; } - if (timeout && time_after_eq(jiffies, timeout_expire)) { - ret = -ETIME; - break; - } - /* Ensure that even if the GPU hangs, we get woken up. * * However, note that if no one is waiting, we never notice @@ -1541,32 +1497,33 @@ int __i915_wait_request(struct drm_i915_gem_request *req, * held by the GPU and so trigger a hangcheck. In the most * pathological case, this will be upon memory starvation! */ - i915_queue_hangcheck(dev_priv); - - timer.function = NULL; - if (timeout || missed_irq(dev_priv, engine)) { - unsigned long expire; + i915_queue_hangcheck(req->i915); - setup_timer_on_stack(&timer, fake_irq, (unsigned long)current); - expire = missed_irq(dev_priv, engine) ? jiffies + 1 : timeout_expire; - mod_timer(&timer, expire); + timeout_remain = io_schedule_timeout(timeout_remain); + if (timeout_remain == 0) { + ret = -ETIME; + break; } - io_schedule(); - - if (timer.function) { - del_singleshot_timer_sync(&timer); - destroy_timer_on_stack(&timer); - } - } - remove_wait_queue(&dev_priv->gpu_error.wait_queue, &reset); + if (intel_wait_complete(&wait)) + break; - if (!irq_test_in_progress) - engine->irq_put(engine); + set_current_state(state); - finish_wait(&engine->irq_queue, &wait); +wakeup: + /* Carefully check if the request is complete, giving time + * for the seqno to be visible following the interrupt. + * We also have to check in case we are kicked by the GPU + * reset in order to drop the struct_mutex. + */ + if (__i915_request_irq_complete(req)) + break; + } + remove_wait_queue(&req->i915->gpu_error.wait_queue, &reset); -out: + intel_engine_remove_wait(req->engine, &wait); + __set_current_state(TASK_RUNNING); +complete: trace_i915_gem_request_wait_end(req); if (timeout) { @@ -2796,6 +2753,12 @@ i915_gem_init_seqno(struct drm_i915_private *dev_priv, u32 seqno) } i915_gem_retire_requests(dev_priv); + /* If the seqno wraps around, we need to clear the breadcrumb rbtree */ + if (!i915_seqno_passed(seqno, dev_priv->next_seqno)) { + while (intel_kick_waiters(dev_priv)) + yield(); + } + /* Finally reset hw state */ for_each_engine(engine, dev_priv) intel_ring_init_seqno(engine, seqno); diff --git a/drivers/gpu/drm/i915/i915_gpu_error.c b/drivers/gpu/drm/i915/i915_gpu_error.c index 34ff2459ceea..250f0b818099 100644 --- a/drivers/gpu/drm/i915/i915_gpu_error.c +++ b/drivers/gpu/drm/i915/i915_gpu_error.c @@ -463,6 +463,18 @@ int i915_error_state_to_str(struct drm_i915_error_state_buf *m, } } + if (error->ring[i].num_waiters) { + err_printf(m, "%s --- %d waiters\n", + dev_priv->engine[i].name, + error->ring[i].num_waiters); + for (j = 0; j < error->ring[i].num_waiters; j++) { + err_printf(m, " seqno 0x%08x for %s [%d]\n", + error->ring[i].waiters[j].seqno, + error->ring[i].waiters[j].comm, + error->ring[i].waiters[j].pid); + } + } + if ((obj = error->ring[i].ringbuffer)) { err_printf(m, "%s --- ringbuffer = 0x%08x\n", dev_priv->engine[i].name, @@ -605,8 +617,9 @@ static void i915_error_state_free(struct kref *error_ref) i915_error_object_free(error->ring[i].ringbuffer); i915_error_object_free(error->ring[i].hws_page); i915_error_object_free(error->ring[i].ctx); - kfree(error->ring[i].requests); i915_error_object_free(error->ring[i].wa_ctx); + kfree(error->ring[i].requests); + kfree(error->ring[i].waiters); } i915_error_object_free(error->semaphore_obj); @@ -892,6 +905,48 @@ static void gen6_record_semaphore_state(struct drm_i915_private *dev_priv, } } +static void engine_record_waiters(struct intel_engine_cs *engine, + struct drm_i915_error_ring *ering) +{ + struct intel_breadcrumbs *b = &engine->breadcrumbs; + struct drm_i915_error_waiter *waiter; + struct rb_node *rb; + int count; + + ering->num_waiters = 0; + ering->waiters = NULL; + + spin_lock(&b->lock); + count = 0; + for (rb = rb_first(&b->waiters); rb != NULL; rb = rb_next(rb)) + count++; + spin_unlock(&b->lock); + + waiter = NULL; + if (count) + waiter = kmalloc_array(count, + sizeof(struct drm_i915_error_waiter), + GFP_ATOMIC); + if (!waiter) + return; + + ering->waiters = waiter; + + spin_lock(&b->lock); + for (rb = rb_first(&b->waiters); rb; rb = rb_next(rb)) { + struct intel_wait *w = container_of(rb, typeof(*w), node); + + strcpy(waiter->comm, w->tsk->comm); + waiter->pid = w->tsk->pid; + waiter->seqno = w->seqno; + waiter++; + + if (++ering->num_waiters == count) + break; + } + spin_unlock(&b->lock); +} + static void i915_record_ring_state(struct drm_i915_private *dev_priv, struct drm_i915_error_state *error, struct intel_engine_cs *engine, @@ -926,7 +981,7 @@ static void i915_record_ring_state(struct drm_i915_private *dev_priv, ering->instdone = I915_READ(GEN2_INSTDONE); } - ering->waiting = waitqueue_active(&engine->irq_queue); + ering->waiting = intel_engine_has_waiter(engine); ering->instpm = I915_READ(RING_INSTPM(engine->mmio_base)); ering->acthd = intel_ring_get_active_head(engine); ering->seqno = engine->get_seqno(engine); @@ -1032,6 +1087,7 @@ static void i915_gem_record_rings(struct drm_i915_private *dev_priv, error->ring[i].valid = true; i915_record_ring_state(dev_priv, error, engine, &error->ring[i]); + engine_record_waiters(engine, &error->ring[i]); request = i915_gem_find_active_request(engine); if (request) { diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c index 6c17596d75dd..a11ab00cdee0 100644 --- a/drivers/gpu/drm/i915/i915_irq.c +++ b/drivers/gpu/drm/i915/i915_irq.c @@ -976,13 +976,10 @@ static void ironlake_rps_change_irq_handler(struct drm_i915_private *dev_priv) static void notify_ring(struct intel_engine_cs *engine) { - if (!intel_engine_initialized(engine)) - return; - - trace_i915_gem_request_notify(engine); - engine->user_interrupts++; - - wake_up_all(&engine->irq_queue); + if (intel_engine_wakeup(engine)) { + trace_i915_gem_request_notify(engine); + engine->user_interrupts++; + } } static void vlv_c0_read(struct drm_i915_private *dev_priv, @@ -1063,7 +1060,7 @@ static bool any_waiters(struct drm_i915_private *dev_priv) struct intel_engine_cs *engine; for_each_engine(engine, dev_priv) - if (engine->irq_refcount) + if (intel_engine_has_waiter(engine)) return true; return false; @@ -3074,13 +3071,14 @@ static unsigned kick_waiters(struct intel_engine_cs *engine) if (engine->hangcheck.user_interrupts == user_interrupts && !test_and_set_bit(engine->id, &i915->gpu_error.missed_irq_rings)) { - if (!(i915->gpu_error.test_irq_rings & intel_engine_flag(engine))) + if (!test_bit(engine->id, &i915->gpu_error.test_irq_rings)) DRM_ERROR("Hangcheck timer elapsed... %s idle\n", engine->name); else DRM_INFO("Fake missed irq on %s\n", engine->name); - wake_up_all(&engine->irq_queue); + + intel_engine_enable_fake_irq(engine); } return user_interrupts; @@ -3124,7 +3122,7 @@ static void i915_hangcheck_elapsed(struct work_struct *work) intel_uncore_arm_unclaimed_mmio_detection(dev_priv); for_each_engine_id(engine, dev_priv, id) { - bool busy = waitqueue_active(&engine->irq_queue); + bool busy = intel_engine_has_waiter(engine); u64 acthd; u32 seqno; unsigned user_interrupts; diff --git a/drivers/gpu/drm/i915/intel_breadcrumbs.c b/drivers/gpu/drm/i915/intel_breadcrumbs.c new file mode 100644 index 000000000000..a3bbf2d90dce --- /dev/null +++ b/drivers/gpu/drm/i915/intel_breadcrumbs.c @@ -0,0 +1,380 @@ +/* + * Copyright © 2015 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + */ + +#include "i915_drv.h" + +static void intel_breadcrumbs_fake_irq(unsigned long data) +{ + struct intel_engine_cs *engine = (struct intel_engine_cs *)data; + + /* + * The timer persists in case we cannot enable interrupts, + * or if we have previously seen seqno/interrupt incoherency + * ("missed interrupt" syndrome). Here the worker will wake up + * every jiffie in order to kick the oldest waiter to do the + * coherent seqno check. + */ + rcu_read_lock(); + if (intel_engine_wakeup(engine)) + mod_timer(&engine->breadcrumbs.fake_irq, jiffies + 1); + rcu_read_unlock(); +} + +static void irq_enable(struct intel_engine_cs *engine) +{ + WARN_ON(!engine->irq_get(engine)); +} + +static void irq_disable(struct intel_engine_cs *engine) +{ + engine->irq_put(engine); +} + +static bool __intel_breadcrumbs_enable_irq(struct intel_breadcrumbs *b) +{ + struct intel_engine_cs *engine = + container_of(b, struct intel_engine_cs, breadcrumbs); + struct drm_i915_private *i915 = engine->i915; + bool irq_posted = false; + + assert_spin_locked(&b->lock); + if (b->rpm_wakelock) + return false; + + /* Since we are waiting on a request, the GPU should be busy + * and should have its own rpm reference. For completeness, + * record an rpm reference for ourselves to cover the + * interrupt we unmask. + */ + intel_runtime_pm_get_noresume(i915); + b->rpm_wakelock = true; + + /* No interrupts? Kick the waiter every jiffie! */ + if (intel_irqs_enabled(i915)) { + if (!test_bit(engine->id, &i915->gpu_error.test_irq_rings)) { + irq_enable(engine); + irq_posted = true; + } + b->irq_enabled = true; + } + + if (!b->irq_enabled || + test_bit(engine->id, &i915->gpu_error.missed_irq_rings)) + mod_timer(&b->fake_irq, jiffies + 1); + + return irq_posted; +} + +static void __intel_breadcrumbs_disable_irq(struct intel_breadcrumbs *b) +{ + struct intel_engine_cs *engine = + container_of(b, struct intel_engine_cs, breadcrumbs); + + assert_spin_locked(&b->lock); + if (!b->rpm_wakelock) + return; + + if (b->irq_enabled) { + irq_disable(engine); + b->irq_enabled = false; + } + + intel_runtime_pm_put(engine->i915); + b->rpm_wakelock = false; +} + +static inline struct intel_wait *to_wait(struct rb_node *node) +{ + return container_of(node, struct intel_wait, node); +} + +static inline void __intel_breadcrumbs_finish(struct intel_breadcrumbs *b, + struct intel_wait *wait) +{ + assert_spin_locked(&b->lock); + + /* This request is completed, so remove it from the tree, mark it as + * complete, and *then* wake up the associated task. + */ + rb_erase(&wait->node, &b->waiters); + RB_CLEAR_NODE(&wait->node); + + wake_up_process(wait->tsk); /* implicit smp_wmb() */ +} + +static bool __intel_engine_add_wait(struct intel_engine_cs *engine, + struct intel_wait *wait) +{ + struct intel_breadcrumbs *b = &engine->breadcrumbs; + struct rb_node **p, *parent, *completed; + bool first; + u32 seqno; + + /* Insert the request into the retirement ordered list + * of waiters by walking the rbtree. If we are the oldest + * seqno in the tree (the first to be retired), then + * set ourselves as the bottom-half. + * + * As we descend the tree, prune completed branches since we hold the + * spinlock we know that the first_waiter must be delayed and can + * reduce some of the sequential wake up latency if we take action + * ourselves and wake up the completed tasks in parallel. Also, by + * removing stale elements in the tree, we may be able to reduce the + * ping-pong between the old bottom-half and ourselves as first-waiter. + */ + first = true; + parent = NULL; + completed = NULL; + seqno = engine->get_seqno(engine); + + /* If the request completed before we managed to grab the spinlock, + * return now before adding ourselves to the rbtree. We let the + * current bottom-half handle any pending wakeups and instead + * try and get out of the way quickly. + */ + if (i915_seqno_passed(seqno, wait->seqno)) { + RB_CLEAR_NODE(&wait->node); + return first; + } + + p = &b->waiters.rb_node; + while (*p) { + parent = *p; + if (wait->seqno == to_wait(parent)->seqno) { + /* We have multiple waiters on the same seqno, select + * the highest priority task (that with the smallest + * task->prio) to serve as the bottom-half for this + * group. + */ + if (wait->tsk->prio > to_wait(parent)->tsk->prio) { + p = &parent->rb_right; + first = false; + } else { + p = &parent->rb_left; + } + } else if (i915_seqno_passed(wait->seqno, + to_wait(parent)->seqno)) { + p = &parent->rb_right; + if (i915_seqno_passed(seqno, to_wait(parent)->seqno)) + completed = parent; + else + first = false; + } else { + p = &parent->rb_left; + } + } + rb_link_node(&wait->node, parent, p); + rb_insert_color(&wait->node, &b->waiters); + GEM_BUG_ON(!first && !b->tasklet); + + if (completed) { + struct rb_node *next = rb_next(completed); + + GEM_BUG_ON(!next && !first); + if (next && next != &wait->node) { + GEM_BUG_ON(first); + b->first_wait = to_wait(next); + smp_store_mb(b->tasklet, b->first_wait->tsk); + /* As there is a delay between reading the current + * seqno, processing the completed tasks and selecting + * the next waiter, we may have missed the interrupt + * and so need for the next bottom-half to wakeup. + * + * Also as we enable the IRQ, we may miss the + * interrupt for that seqno, so we have to wake up + * the next bottom-half in order to do a coherent check + * in case the seqno passed. + */ + __intel_breadcrumbs_enable_irq(b); + wake_up_process(to_wait(next)->tsk); + } + + do { + struct intel_wait *crumb = to_wait(completed); + completed = rb_prev(completed); + __intel_breadcrumbs_finish(b, crumb); + } while (completed); + } + + if (first) { + GEM_BUG_ON(rb_first(&b->waiters) != &wait->node); + b->first_wait = wait; + smp_store_mb(b->tasklet, wait->tsk); + first = __intel_breadcrumbs_enable_irq(b); + } + GEM_BUG_ON(!b->tasklet); + GEM_BUG_ON(!b->first_wait); + GEM_BUG_ON(rb_first(&b->waiters) != &b->first_wait->node); + + return first; +} + +bool intel_engine_add_wait(struct intel_engine_cs *engine, + struct intel_wait *wait) +{ + struct intel_breadcrumbs *b = &engine->breadcrumbs; + bool first; + + spin_lock(&b->lock); + first = __intel_engine_add_wait(engine, wait); + spin_unlock(&b->lock); + + return first; +} + +void intel_engine_enable_fake_irq(struct intel_engine_cs *engine) +{ + mod_timer(&engine->breadcrumbs.fake_irq, jiffies + 1); +} + +static inline bool chain_wakeup(struct rb_node *rb, int priority) +{ + return rb && to_wait(rb)->tsk->prio <= priority; +} + +void intel_engine_remove_wait(struct intel_engine_cs *engine, + struct intel_wait *wait) +{ + struct intel_breadcrumbs *b = &engine->breadcrumbs; + + /* Quick check to see if this waiter was already decoupled from + * the tree by the bottom-half to avoid contention on the spinlock + * by the herd. + */ + if (RB_EMPTY_NODE(&wait->node)) + return; + + spin_lock(&b->lock); + + if (RB_EMPTY_NODE(&wait->node)) + goto out_unlock; + + if (b->first_wait == wait) { + struct rb_node *next; + const int priority = wait->tsk->prio; + + GEM_BUG_ON(b->tasklet != wait->tsk); + + /* We are the current bottom-half. Find the next candidate, + * the first waiter in the queue on the remaining oldest + * request. As multiple seqnos may complete in the time it + * takes us to wake up and find the next waiter, we have to + * wake up that waiter for it to perform its own coherent + * completion check. + */ + next = rb_next(&wait->node); + if (chain_wakeup(next, priority)) { + /* If the next waiter is already complete, + * wake it up and continue onto the next waiter. So + * if have a small herd, they will wake up in parallel + * rather than sequentially, which should reduce + * the overall latency in waking all the completed + * clients. + * + * However, waking up a chain adds extra latency to + * the first_waiter. This is undesirable if that + * waiter is a high priority task. + */ + u32 seqno = engine->get_seqno(engine); + + while (i915_seqno_passed(seqno, to_wait(next)->seqno)) { + struct rb_node *n = rb_next(next); + + __intel_breadcrumbs_finish(b, to_wait(next)); + next = n; + if (!chain_wakeup(next, priority)) + break; + } + } + + if (next) { + /* In our haste, we may have completed the first waiter + * before we enabled the interrupt. Do so now as we + * have a second waiter for a future seqno. Afterwards, + * we have to wake up that waiter in case we missed + * the interrupt, or if we have to handle an + * exception rather than a seqno completion. + */ + b->first_wait = to_wait(next); + smp_store_mb(b->tasklet, b->first_wait->tsk); + if (b->first_wait->seqno != wait->seqno) + __intel_breadcrumbs_enable_irq(b); + wake_up_process(b->tasklet); + } else { + b->first_wait = NULL; + WRITE_ONCE(b->tasklet, NULL); + __intel_breadcrumbs_disable_irq(b); + } + } else { + GEM_BUG_ON(rb_first(&b->waiters) == &wait->node); + } + + GEM_BUG_ON(RB_EMPTY_NODE(&wait->node)); + rb_erase(&wait->node, &b->waiters); + +out_unlock: + GEM_BUG_ON(b->first_wait == wait); + GEM_BUG_ON(rb_first(&b->waiters) != + (b->first_wait ? &b->first_wait->node : NULL)); + GEM_BUG_ON(!b->tasklet ^ RB_EMPTY_ROOT(&b->waiters)); + spin_unlock(&b->lock); +} + +int intel_engine_init_breadcrumbs(struct intel_engine_cs *engine) +{ + struct intel_breadcrumbs *b = &engine->breadcrumbs; + + spin_lock_init(&b->lock); + setup_timer(&b->fake_irq, + intel_breadcrumbs_fake_irq, + (unsigned long)engine); + + return 0; +} + +void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine) +{ + struct intel_breadcrumbs *b = &engine->breadcrumbs; + + del_timer_sync(&b->fake_irq); +} + +unsigned int intel_kick_waiters(struct drm_i915_private *i915) +{ + struct intel_engine_cs *engine; + unsigned int mask = 0; + + /* To avoid the task_struct disappearing beneath us as we wake up + * the process, we must first inspect the task_struct->state under the + * RCU lock, i.e. as we call wake_up_process() we must be holding the + * rcu_read_lock(). + */ + rcu_read_lock(); + for_each_engine(engine, i915) + if (unlikely(intel_engine_wakeup(engine))) + mask |= intel_engine_flag(engine); + rcu_read_unlock(); + + return mask; +} diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c index 339d8041075f..226bba22e4b4 100644 --- a/drivers/gpu/drm/i915/intel_lrc.c +++ b/drivers/gpu/drm/i915/intel_lrc.c @@ -1961,6 +1961,8 @@ void intel_logical_ring_cleanup(struct intel_engine_cs *engine) i915_cmd_parser_fini_ring(engine); i915_gem_batch_pool_fini(&engine->batch_pool); + intel_engine_fini_breadcrumbs(engine); + if (engine->status_page.obj) { i915_gem_object_unpin_map(engine->status_page.obj); engine->status_page.obj = NULL; @@ -1998,7 +2000,6 @@ logical_ring_default_irqs(struct intel_engine_cs *engine, unsigned shift) { engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift; engine->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift; - init_waitqueue_head(&engine->irq_queue); } static int @@ -2025,6 +2026,10 @@ logical_ring_init(struct intel_engine_cs *engine) struct i915_gem_context *dctx = engine->i915->kernel_context; int ret; + ret = intel_engine_init_breadcrumbs(engine); + if (ret) + goto error; + ret = i915_cmd_parser_init_ring(engine); if (ret) goto error; diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c index 24cdc920f4b4..af50aa01bcd9 100644 --- a/drivers/gpu/drm/i915/intel_ringbuffer.c +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c @@ -2394,7 +2394,9 @@ static int intel_init_ring_buffer(struct drm_device *dev, memset(engine->semaphore.sync_seqno, 0, sizeof(engine->semaphore.sync_seqno)); - init_waitqueue_head(&engine->irq_queue); + ret = intel_engine_init_breadcrumbs(engine); + if (ret) + goto error; /* We may need to do things with the shrinker which * require us to immediately switch back to the default @@ -2474,6 +2476,7 @@ void intel_cleanup_engine(struct intel_engine_cs *engine) i915_cmd_parser_fini_ring(engine); i915_gem_batch_pool_fini(&engine->batch_pool); + intel_engine_fini_breadcrumbs(engine); intel_ring_context_unpin(dev_priv->kernel_context, engine); @@ -2676,6 +2679,13 @@ void intel_ring_init_seqno(struct intel_engine_cs *engine, u32 seqno) engine->last_submitted_seqno = seqno; engine->hangcheck.seqno = seqno; + + /* After manually advancing the seqno, fake the interrupt in case + * there are any waiters for that seqno. + */ + rcu_read_lock(); + intel_engine_wakeup(engine); + rcu_read_unlock(); } static void gen6_bsd_ring_write_tail(struct intel_engine_cs *engine, diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.h b/drivers/gpu/drm/i915/intel_ringbuffer.h index 113d5230a6de..6fd70a56e219 100644 --- a/drivers/gpu/drm/i915/intel_ringbuffer.h +++ b/drivers/gpu/drm/i915/intel_ringbuffer.h @@ -148,6 +148,32 @@ struct intel_engine_cs { struct intel_ringbuffer *buffer; struct list_head buffers; + /* Rather than have every client wait upon all user interrupts, + * with the herd waking after every interrupt and each doing the + * heavyweight seqno dance, we delegate the task (of being the + * bottom-half of the user interrupt) to the first client. After + * every interrupt, we wake up one client, who does the heavyweight + * coherent seqno read and either goes back to sleep (if incomplete), + * or wakes up all the completed clients in parallel, before then + * transferring the bottom-half status to the next client in the queue. + * + * Compared to walking the entire list of waiters in a single dedicated + * bottom-half, we reduce the latency of the first waiter by avoiding + * a context switch, but incur additional coherent seqno reads when + * following the chain of request breadcrumbs. Since it is most likely + * that we have a single client waiting on each seqno, then reducing + * the overhead of waking that client is much preferred. + */ + struct intel_breadcrumbs { + spinlock_t lock; /* protects the lists of requests */ + struct rb_root waiters; /* sorted by retirement, priority */ + struct intel_wait *first_wait; /* oldest waiter by retirement */ + struct task_struct *tasklet; /* bh for user interrupts */ + struct timer_list fake_irq; /* used after a missed interrupt */ + bool irq_enabled; + bool rpm_wakelock; + } breadcrumbs; + /* * A pool of objects to use as shadow copies of client batch buffers * when the command parser is enabled. Prevents the client from @@ -296,8 +322,6 @@ struct intel_engine_cs { bool gpu_caches_dirty; - wait_queue_head_t irq_queue; - struct i915_gem_context *last_context; struct intel_ring_hangcheck hangcheck; @@ -483,4 +507,55 @@ static inline u32 intel_hws_seqno_address(struct intel_engine_cs *engine) return engine->status_page.gfx_addr + I915_GEM_HWS_INDEX_ADDR; } +/* intel_breadcrumbs.c -- user interrupt bottom-half for waiters */ +struct intel_wait { + struct rb_node node; + struct task_struct *tsk; + u32 seqno; +}; + +int intel_engine_init_breadcrumbs(struct intel_engine_cs *engine); + +static inline void intel_wait_init(struct intel_wait *wait, u32 seqno) +{ + wait->tsk = current; + wait->seqno = seqno; +} + +static inline bool intel_wait_complete(const struct intel_wait *wait) +{ + return RB_EMPTY_NODE(&wait->node); +} + +bool intel_engine_add_wait(struct intel_engine_cs *engine, + struct intel_wait *wait); +void intel_engine_remove_wait(struct intel_engine_cs *engine, + struct intel_wait *wait); + +static inline bool intel_engine_has_waiter(struct intel_engine_cs *engine) +{ + return READ_ONCE(engine->breadcrumbs.tasklet); +} + +static inline bool intel_engine_wakeup(struct intel_engine_cs *engine) +{ + bool wakeup = false; + struct task_struct *tsk = READ_ONCE(engine->breadcrumbs.tasklet); + /* Note that for this not to dangerously chase a dangling pointer, + * the caller is responsible for ensure that the task remain valid for + * wake_up_process() i.e. that the RCU grace period cannot expire. + * + * Also note that tsk is likely to be in !TASK_RUNNING state so an + * early test for tsk->state != TASK_RUNNING before wake_up_process() + * is unlikely to be beneficial. + */ + if (tsk) + wakeup = wake_up_process(tsk); + return wakeup; +} + +void intel_engine_enable_fake_irq(struct intel_engine_cs *engine); +void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine); +unsigned int intel_kick_waiters(struct drm_i915_private *i915); + #endif /* _INTEL_RINGBUFFER_H_ */ |