If the preempted context takes too long to relinquish control, e.g. it
is stuck inside a shader with arbitration disabled, evict that context
with an engine reset. This ensures that preemptions are reasonably
responsive, providing a tighter QoS for the more important context at
the cost of flagging unresponsive contexts more frequently (i.e. instead
of using an ~10s hangcheck, we now evict at ~100ms). The challenge of
lies in picking a timeout that can be reasonably serviced by HW for
typical workloads, balancing the existing clients against the needs for
responsiveness.
Note that coupled with timeslicing, this will lead to rapid GPU "hang"
detection with multiple active contexts vying for GPU time.
Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Cc: Mika Kuoppala <mika.kuoppala@xxxxxxxxxxxxxxx>
Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
Reviewed-by: Mika Kuoppala <mika.kuoppala@xxxxxxxxxxxxxxx>
---
drivers/gpu/drm/i915/Kconfig.profile | 12 +++++++
drivers/gpu/drm/i915/gt/intel_lrc.c | 51 ++++++++++++++++++++++++++++
drivers/gpu/drm/i915/i915_params.h | 2 +-
3 files changed, 64 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/Kconfig.profile b/drivers/gpu/drm/i915/Kconfig.profile
index 48df8889a88a..3184e8491333 100644
--- a/drivers/gpu/drm/i915/Kconfig.profile
+++ b/drivers/gpu/drm/i915/Kconfig.profile
@@ -25,3 +25,15 @@ config DRM_I915_SPIN_REQUEST
May be 0 to disable the initial spin. In practice, we estimate
the cost of enabling the interrupt (if currently disabled) to be
a few microseconds.
+
+config DRM_I915_PREEMPT_TIMEOUT
+ int "Preempt timeout (ms)"
+ default 100 # milliseconds
+ help
+ How long to wait (in milliseconds) for a preemption event to occur
+ when submitting a new context via execlists. If the current context
+ does not hit an arbitration point and yield to HW before the timer
+ expires, the HW will be reset to allow the more important context
+ to execute.
+
+ May be 0 to disable the timeout.
diff --git a/drivers/gpu/drm/i915/gt/intel_lrc.c b/drivers/gpu/drm/i915/gt/intel_lrc.c
index 3a23a7bda50e..3e7ebad836ac 100644
--- a/drivers/gpu/drm/i915/gt/intel_lrc.c
+++ b/drivers/gpu/drm/i915/gt/intel_lrc.c
@@ -1005,6 +1005,21 @@ static void record_preemption(struct intel_engine_execlists *execlists)
(void)I915_SELFTEST_ONLY(execlists->preempt_hang.count++);
}
+static unsigned long preempt_expires(void)
+{
+ const unsigned long timeout =
+ msecs_to_jiffies_timeout(CONFIG_DRM_I915_PREEMPT_TIMEOUT);
+
+ /*
+ * Paranoia to make sure the compiler computes the timeout before
+ * loading 'jiffies' as jiffies is volatile and may be updated in
+ * the background by a timer tick. All to reduce the complexity
+ * of the addition and reduce the risk of losing a jiffie.
+ */
+ barrier();
+ return jiffies + timeout;
+}
+
static void execlists_dequeue(struct intel_engine_cs *engine)
{
struct intel_engine_execlists * const execlists = &engine->execlists;
@@ -1345,6 +1360,8 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
execlists->switch_priority_hint =
switch_prio(engine, *execlists->pending);
execlists_submit_ports(engine);
+ if (CONFIG_DRM_I915_PREEMPT_TIMEOUT)
+ mod_timer(&execlists->timer, preempt_expires());
} else {
ring_set_paused(engine, 0);
}
@@ -1604,6 +1621,37 @@ static void __execlists_submission_tasklet(struct intel_engine_cs *const engine)
execlists_dequeue(engine);
}
+static noinline void preempt_reset(struct intel_engine_cs *engine)
+{
+ const unsigned int bit = I915_RESET_ENGINE + engine->id;
+ unsigned long *lock = &engine->gt->reset.flags;
+
+ if (i915_modparams.reset < 3)
+ return;
+
+ if (test_and_set_bit(bit, lock))
+ return;
+
+ /* Mark this tasklet as disabled to avoid waiting for it to complete */
+ tasklet_disable_nosync(&engine->execlists.tasklet);
+
+ intel_engine_reset(engine, "preemption time out");
+
+ tasklet_enable(&engine->execlists.tasklet);
+ clear_and_wake_up_bit(bit, lock);
+}
+
+static bool preempt_timeout(struct intel_engine_cs *const engine)
+{
+ if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
+ return false;
+
+ if (!intel_engine_has_preemption(engine))
+ return false;
+
+ return !timer_pending(&engine->execlists.timer);
+}
+
/*
* Check the unread Context Status Buffers and manage the submission of new
* contexts to the ELSP accordingly.
@@ -1618,6 +1666,9 @@ static void execlists_submission_tasklet(unsigned long data)
spin_lock_irqsave(&engine->active.lock, flags);
__execlists_submission_tasklet(engine);
spin_unlock_irqrestore(&engine->active.lock, flags);
+ } else {
+ if (preempt_timeout(engine))
+ preempt_reset(engine);
}
}
diff --git a/drivers/gpu/drm/i915/i915_params.h b/drivers/gpu/drm/i915/i915_params.h
index d29ade3b7de6..56058978bb27 100644
--- a/drivers/gpu/drm/i915/i915_params.h
+++ b/drivers/gpu/drm/i915/i915_params.h
@@ -61,7 +61,7 @@ struct drm_printer;
param(char *, dmc_firmware_path, NULL) \
param(int, mmio_debug, -IS_ENABLED(CONFIG_DRM_I915_DEBUG_MMIO)) \
param(int, edp_vswing, 0) \
- param(int, reset, 2) \
+ param(int, reset, 3) \
param(unsigned int, inject_load_failure, 0) \
param(int, fastboot, -1) \
param(int, enable_dpcd_backlight, 0) \
--
2.23.0
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
