@@ -753,6 +753,64 @@ static int eb_select_context(struct i915_execbuffer *eb)
return 0;
}
+static struct i915_request *__eb_wait_for_ring(struct intel_ring *ring)
+{
+ struct i915_request *rq;
+
+ if (intel_ring_update_space(ring) >= PAGE_SIZE)
+ return NULL;
+
+ /*
+ * Find a request that after waiting upon, there will be at least half
+ * the ring available. The hystersis allows us to compete for the
+ * shared ring and should mean that we sleep less often prior to
+ * claiming our resources, but not so long that the ring completely
+ * drains before we can submit our next request.
+ */
+ list_for_each_entry(rq, &ring->request_list, ring_link) {
+ if (__intel_ring_space(rq->postfix,
+ ring->emit, ring->size) > ring->size / 2)
+ break;
+ }
+ if (&rq->ring_link == &ring->request_list)
+ return NULL; /* weird, we will check again later for real */
+
+ return i915_request_get(rq);
+}
+
+static int eb_wait_for_ring(const struct i915_execbuffer *eb)
+{
+ const struct intel_context *ce;
+ struct i915_request *rq;
+ int ret = 0;
+
+ /*
+ * Apply a light amount of backpressure to prevent excessive hogs
+ * from blocking waiting for space whilst holding struct_mutex and
+ * keeping all of their resources pinned.
+ */
+
+ ce = to_intel_context(eb->ctx, eb->engine);
+ if (!ce->ring) /* first use, assume empty! */
+ return 0;
+
+ rq = __eb_wait_for_ring(ce->ring);
+ if (rq) {
+ mutex_unlock(&eb->i915->drm.struct_mutex);
+
+ if (i915_request_wait(rq,
+ I915_WAIT_INTERRUPTIBLE,
+ MAX_SCHEDULE_TIMEOUT) < 0)
+ ret = -EINTR;
+
+ i915_request_put(rq);
+
+ mutex_lock(&eb->i915->drm.struct_mutex);
+ }
+
+ return ret;
+}
+
static int eb_lookup_vmas(struct i915_execbuffer *eb)
{
struct radix_tree_root *handles_vma = &eb->ctx->handles_vma;
@@ -2291,6 +2349,10 @@ i915_gem_do_execbuffer(struct drm_device *dev,
if (err)
goto err_rpm;
+ err = eb_wait_for_ring(&eb); /* may temporarily drop struct_mutex */
+ if (unlikely(err))
+ goto err_unlock;
+
err = eb_relocate(&eb);
if (err) {
/*
@@ -2435,6 +2497,7 @@ i915_gem_do_execbuffer(struct drm_device *dev,
err_vma:
if (eb.exec)
eb_release_vmas(&eb);
+err_unlock:
mutex_unlock(&dev->struct_mutex);
err_rpm:
intel_runtime_pm_put(eb.i915, wakeref);
@@ -49,19 +49,6 @@ static inline u32 intel_hws_seqno_address(struct intel_engine_cs *engine)
I915_GEM_HWS_INDEX_ADDR);
}
-static unsigned int __intel_ring_space(unsigned int head,
- unsigned int tail,
- unsigned int size)
-{
- /*
- * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
- * same cacheline, the Head Pointer must not be greater than the Tail
- * Pointer."
- */
- GEM_BUG_ON(!is_power_of_2(size));
- return (head - tail - CACHELINE_BYTES) & (size - 1);
-}
-
unsigned int intel_ring_update_space(struct intel_ring *ring)
{
unsigned int space;
@@ -831,6 +831,18 @@ intel_ring_set_tail(struct intel_ring *ring, unsigned int tail)
return tail;
}
+static inline unsigned int
+__intel_ring_space(unsigned int head, unsigned int tail, unsigned int size)
+{
+ /*
+ * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
+ * same cacheline, the Head Pointer must not be greater than the Tail
+ * Pointer."
+ */
+ GEM_BUG_ON(!is_power_of_2(size));
+ return (head - tail - CACHELINE_BYTES) & (size - 1);
+}
+
void intel_engine_write_global_seqno(struct intel_engine_cs *engine, u32 seqno);
int intel_engine_setup_common(struct intel_engine_cs *engine);
Apply backpressure to hogs that emit requests faster than the GPU can process them by waiting for their ring to be less than half-full before proceeding with taking the struct_mutex. This is a gross hack to apply throttling backpressure, the long term goal is to remove the struct_mutex contention so that each client naturally waits, preferably in an asynchronous, nonblocking fashion (pipelined operations for the win), for their own resources and never blocks another client within the driver at least. (Realtime priority goals would extend to ensuring that resource contention favours high priority clients as well.) This patch only limits excessive request production and does not attempt to throttle clients that block waiting for eviction (either global GTT or system memory) or any other global resources, see above for the long term goal. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: John Harrison <John.C.Harrison@Intel.com> --- Acquire the struct_mutex first then drop it for the wait. This means that we keep a rpm wakeref across the sleep reducing subsequent latency. Oh for splitting up i915_request_alloc() so that we can do the global operations like reserve_gt() and keep the GT awake while we sleep on it being busy etc. -Chris --- drivers/gpu/drm/i915/i915_gem_execbuffer.c | 63 ++++++++++++++++++++++ drivers/gpu/drm/i915/intel_ringbuffer.c | 13 ----- drivers/gpu/drm/i915/intel_ringbuffer.h | 12 +++++ 3 files changed, 75 insertions(+), 13 deletions(-)