Merge tag 'drm-next-2019-05-16' of git://anongit.freedesktop.org/drm/drm

Pull drm fixes from Dave Airlie:
 "A bunch of fixes for the merge window closure, doesn't seem to be
  anything too major or serious in there.

  It does add TU117 turing modesetting to nouveau but it's just an
  enable for preexisting code.

  amdgpu:
   - gpu reset at load crash fix
   - ATPX hotplug fix for when dGPU is off
   - SR-IOV fixes

  radeon:
   - r5xx pll fixes

  i915:
   - GVT (MCHBAR, buffer alignment, misc warnings fixes)
   - Fixes for newly enabled semaphore code
   - Geminilake disable framebuffer compression
   - HSW edp fast modeset fix
   - IRQ vs RCU race fix

  nouveau:
   - Turing modesetting fixes
   - TU117 support

  msm:
   - SDM845 bringup fixes

  panfrost:
   - static checker fixes

  pl111:
   - spinlock init fix.

  bridge:
   - refresh rate register fix for adv7511"

* tag 'drm-next-2019-05-16' of git://anongit.freedesktop.org/drm/drm: (36 commits)
  drm/msm: Upgrade gxpd checks to IS_ERR_OR_NULL
  drm/msm/dpu: Remove duplicate header
  drm/pl111: Initialize clock spinlock early
  drm/msm: correct attempted NULL pointer dereference in debugfs
  drm/msm: remove resv fields from msm_gem_object struct
  drm/nouveau: fix duplication of nv50_head_atom struct
  drm/nouveau/disp/dp: respect sink limits when selecting failsafe link configuration
  drm/nouveau/core: initial support for boards with TU117 chipset
  drm/nouveau/core: allow detected chipset to be overridden
  drm/nouveau/kms/gf119-gp10x: push HeadSetControlOutputResource() mthd when encoders change
  drm/nouveau/kms/nv50-: fix bug preventing non-vsync'd page flips
  drm/nouveau/kms/gv100-: fix spurious window immediate interlocks
  drm/bridge: adv7511: Fix low refresh rate selection
  drm/panfrost: Add missing _fini() calls in panfrost_device_fini()
  drm/panfrost: Only put sync_out if non-NULL
  drm/i915: Seal races between async GPU cancellation, retirement and signaling
  drm/i915: Fix fastset vs. pfit on/off on HSW EDP transcoder
  drm/i915/fbc: disable framebuffer compression on GeminiLake
  drm/amdgpu/psp: move psp version specific function pointers to early_init
  drm/radeon: prefer lower reference dividers
  ...

1 files changed, 59 insertions, 1 deletions

diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index b836721d3b13..f6c78c0fa74b 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -425,6 +425,26 @@ void __i915_request_submit(struct i915_request *request)
 	if (i915_gem_context_is_banned(request->gem_context))
 		i915_request_skip(request, -EIO);
 
+	/*
+	 * Are we using semaphores when the gpu is already saturated?
+	 *
+	 * Using semaphores incurs a cost in having the GPU poll a
+	 * memory location, busywaiting for it to change. The continual
+	 * memory reads can have a noticeable impact on the rest of the
+	 * system with the extra bus traffic, stalling the cpu as it too
+	 * tries to access memory across the bus (perf stat -e bus-cycles).
+	 *
+	 * If we installed a semaphore on this request and we only submit
+	 * the request after the signaler completed, that indicates the
+	 * system is overloaded and using semaphores at this time only
+	 * increases the amount of work we are doing. If so, we disable
+	 * further use of semaphores until we are idle again, whence we
+	 * optimistically try again.
+	 */
+	if (request->sched.semaphores &&
+	    i915_sw_fence_signaled(&request->semaphore))
+		request->hw_context->saturated |= request->sched.semaphores;
+
 	/* We may be recursing from the signal callback of another i915 fence */
 	spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
 
@@ -432,6 +452,7 @@ void __i915_request_submit(struct i915_request *request)
 	set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags);
 
 	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags) &&
+	    !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &request->fence.flags) &&
 	    !i915_request_enable_breadcrumb(request))
 		intel_engine_queue_breadcrumbs(engine);
 
@@ -799,6 +820,39 @@ err_unreserve:
 }
 
 static int
+i915_request_await_start(struct i915_request *rq, struct i915_request *signal)
+{
+	if (list_is_first(&signal->ring_link, &signal->ring->request_list))
+		return 0;
+
+	signal = list_prev_entry(signal, ring_link);
+	if (i915_timeline_sync_is_later(rq->timeline, &signal->fence))
+		return 0;
+
+	return i915_sw_fence_await_dma_fence(&rq->submit,
+					     &signal->fence, 0,
+					     I915_FENCE_GFP);
+}
+
+static intel_engine_mask_t
+already_busywaiting(struct i915_request *rq)
+{
+	/*
+	 * Polling a semaphore causes bus traffic, delaying other users of
+	 * both the GPU and CPU. We want to limit the impact on others,
+	 * while taking advantage of early submission to reduce GPU
+	 * latency. Therefore we restrict ourselves to not using more
+	 * than one semaphore from each source, and not using a semaphore
+	 * if we have detected the engine is saturated (i.e. would not be
+	 * submitted early and cause bus traffic reading an already passed
+	 * semaphore).
+	 *
+	 * See the are-we-too-late? check in __i915_request_submit().
+	 */
+	return rq->sched.semaphores | rq->hw_context->saturated;
+}
+
+static int
 emit_semaphore_wait(struct i915_request *to,
 		    struct i915_request *from,
 		    gfp_t gfp)
@@ -811,11 +865,15 @@ emit_semaphore_wait(struct i915_request *to,
 	GEM_BUG_ON(INTEL_GEN(to->i915) < 8);
 
 	/* Just emit the first semaphore we see as request space is limited. */
-	if (to->sched.semaphores & from->engine->mask)
+	if (already_busywaiting(to) & from->engine->mask)
 		return i915_sw_fence_await_dma_fence(&to->submit,
 						     &from->fence, 0,
 						     I915_FENCE_GFP);
 
+	err = i915_request_await_start(to, from);
+	if (err < 0)
+		return err;
+
 	err = i915_sw_fence_await_dma_fence(&to->semaphore,
 					    &from->fence, 0,
 					    I915_FENCE_GFP);