drm/i915: Add a delay between interrupt and inspecting the final seqno (ilk)
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 1 Jul 2016 16:23:21 +0000 (17:23 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 1 Jul 2016 20:00:53 +0000 (21:00 +0100)
On Ironlake, there is no command nor register to ensure that the write
from a MI_STORE command is completed (and coherent on the CPU) before the
command parser continues. This means that the ordering between the seqno
write and the subsequent user interrupt is undefined (like gen6+). So to
ensure that the seqno write is completed after the final user interrupt
we need to delay the read sufficiently to allow the write to complete.
This delay is undefined by the bspec, and empirically requires 75us even
though a register read combined with a clflush is less than 500ns. Hence,
the delay is due to an on-chip buffer rather than the latency of the write
to memory.

Note that the render ring controls this by filling the PIPE_CONTROL fifo
with stalling commands that force the earliest pipe-control with the
seqno to be completed before the command parser continues. Given that we
need a barrier operation for BSD, we may as well forgo the extra
per-batch latency by using a common per-interrupt barrier.

Studying the impact of adding the usleep shows that in both sequences of
and individual synchronous no-op batches is negligible for the media
engine (where the write now is unordered with the interrupt). Converting
the render engine over from the current glutton of pie-controls over to
the per-interrupt delays speeds up both the sequential and individual
synchronous no-ops by 20% and 60%, respectively. This speed up holds
even when looking at the throughput of small copies (4KiB->4MiB), both
serial and synchronous, by about 20%. This is because despite adding a
significant delay to the interrupt, in all likelihood we will see the
seqno write without having to apply the barrier (only in the rare corner
cases where the write is delayed on the last required is the delay
necessary).

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94307
Testcase: igt/gem_sync #ilk
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-12-git-send-email-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_irq.c
drivers/gpu/drm/i915/intel_ringbuffer.c

index 7c379afcff2f7ea74b6ae5f87ff86ea74cff1afb..be7f0b9b27e0c3f971a8bb11c686204a98211664 100644 (file)
@@ -1264,8 +1264,7 @@ static void ivybridge_parity_error_irq_handler(struct drm_i915_private *dev_priv
 static void ilk_gt_irq_handler(struct drm_i915_private *dev_priv,
                               u32 gt_iir)
 {
-       if (gt_iir &
-           (GT_RENDER_USER_INTERRUPT | GT_RENDER_PIPECTL_NOTIFY_INTERRUPT))
+       if (gt_iir & GT_RENDER_USER_INTERRUPT)
                notify_ring(&dev_priv->engine[RCS]);
        if (gt_iir & ILK_BSD_USER_INTERRUPT)
                notify_ring(&dev_priv->engine[VCS]);
@@ -1274,9 +1273,7 @@ static void ilk_gt_irq_handler(struct drm_i915_private *dev_priv,
 static void snb_gt_irq_handler(struct drm_i915_private *dev_priv,
                               u32 gt_iir)
 {
-
-       if (gt_iir &
-           (GT_RENDER_USER_INTERRUPT | GT_RENDER_PIPECTL_NOTIFY_INTERRUPT))
+       if (gt_iir & GT_RENDER_USER_INTERRUPT)
                notify_ring(&dev_priv->engine[RCS]);
        if (gt_iir & GT_BSD_USER_INTERRUPT)
                notify_ring(&dev_priv->engine[VCS]);
@@ -3601,8 +3598,7 @@ static void gen5_gt_irq_postinstall(struct drm_device *dev)
 
        gt_irqs |= GT_RENDER_USER_INTERRUPT;
        if (IS_GEN5(dev)) {
-               gt_irqs |= GT_RENDER_PIPECTL_NOTIFY_INTERRUPT |
-                          ILK_BSD_USER_INTERRUPT;
+               gt_irqs |= ILK_BSD_USER_INTERRUPT;
        } else {
                gt_irqs |= GT_BLT_USER_INTERRUPT | GT_BSD_USER_INTERRUPT;
        }
index 25d616d114e4e84ab9835922de0ce7985f0aa032..3bd598673d8824c5a4051a3ebbede96a3740e91b 100644 (file)
@@ -1593,67 +1593,22 @@ gen6_ring_sync(struct drm_i915_gem_request *waiter_req,
        return 0;
 }
 
-#define PIPE_CONTROL_FLUSH(ring__, addr__)                                     \
-do {                                                                   \
-       intel_ring_emit(ring__, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |                \
-                PIPE_CONTROL_DEPTH_STALL);                             \
-       intel_ring_emit(ring__, (addr__) | PIPE_CONTROL_GLOBAL_GTT);                    \
-       intel_ring_emit(ring__, 0);                                                     \
-       intel_ring_emit(ring__, 0);                                                     \
-} while (0)
-
-static int
-pc_render_add_request(struct drm_i915_gem_request *req)
+static void
+gen5_seqno_barrier(struct intel_engine_cs *ring)
 {
-       struct intel_engine_cs *engine = req->engine;
-       u32 addr = engine->status_page.gfx_addr +
-               (I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
-       u32 scratch_addr = addr;
-       int ret;
-
-       /* For Ironlake, MI_USER_INTERRUPT was deprecated and apparently
-        * incoherent with writes to memory, i.e. completely fubar,
-        * so we need to use PIPE_NOTIFY instead.
+       /* MI_STORE are internally buffered by the GPU and not flushed
+        * either by MI_FLUSH or SyncFlush or any other combination of
+        * MI commands.
         *
-        * However, we also need to workaround the qword write
-        * incoherence by flushing the 6 PIPE_NOTIFY buffers out to
-        * memory before requesting an interrupt.
+        * "Only the submission of the store operation is guaranteed.
+        * The write result will be complete (coherent) some time later
+        * (this is practically a finite period but there is no guaranteed
+        * latency)."
+        *
+        * Empirically, we observe that we need a delay of at least 75us to
+        * be sure that the seqno write is visible by the CPU.
         */
-       ret = intel_ring_begin(req, 32);
-       if (ret)
-               return ret;
-
-       intel_ring_emit(engine,
-                       GFX_OP_PIPE_CONTROL(4) |
-                       PIPE_CONTROL_QW_WRITE |
-                       PIPE_CONTROL_WRITE_FLUSH |
-                       PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE);
-       intel_ring_emit(engine, addr | PIPE_CONTROL_GLOBAL_GTT);
-       intel_ring_emit(engine, req->seqno);
-       intel_ring_emit(engine, 0);
-       PIPE_CONTROL_FLUSH(engine, scratch_addr);
-       scratch_addr += 2 * CACHELINE_BYTES; /* write to separate cachelines */
-       PIPE_CONTROL_FLUSH(engine, scratch_addr);
-       scratch_addr += 2 * CACHELINE_BYTES;
-       PIPE_CONTROL_FLUSH(engine, scratch_addr);
-       scratch_addr += 2 * CACHELINE_BYTES;
-       PIPE_CONTROL_FLUSH(engine, scratch_addr);
-       scratch_addr += 2 * CACHELINE_BYTES;
-       PIPE_CONTROL_FLUSH(engine, scratch_addr);
-       scratch_addr += 2 * CACHELINE_BYTES;
-       PIPE_CONTROL_FLUSH(engine, scratch_addr);
-
-       intel_ring_emit(engine,
-                       GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |
-                       PIPE_CONTROL_WRITE_FLUSH |
-                       PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
-                       PIPE_CONTROL_NOTIFY);
-       intel_ring_emit(engine, addr | PIPE_CONTROL_GLOBAL_GTT);
-       intel_ring_emit(engine, req->seqno);
-       intel_ring_emit(engine, 0);
-       __intel_ring_advance(engine);
-
-       return 0;
+       usleep_range(125, 250);
 }
 
 static void
@@ -2964,6 +2919,7 @@ static void intel_ring_init_irq(struct drm_i915_private *dev_priv,
        } else if (INTEL_GEN(dev_priv) >= 5) {
                engine->irq_get = gen5_ring_get_irq;
                engine->irq_put = gen5_ring_put_irq;
+               engine->irq_seqno_barrier = gen5_seqno_barrier;
        } else if (INTEL_GEN(dev_priv) >= 3) {
                engine->irq_get = i9xx_ring_get_irq;
                engine->irq_put = i9xx_ring_put_irq;
@@ -3012,11 +2968,12 @@ int intel_init_render_ring_buffer(struct drm_device *dev)
 
        intel_ring_default_vfuncs(dev_priv, engine);
 
+       engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
+
        if (INTEL_GEN(dev_priv) >= 8) {
                engine->init_context = intel_rcs_ctx_init;
                engine->add_request = gen8_render_add_request;
                engine->flush = gen8_render_ring_flush;
-               engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
                if (i915_semaphore_is_enabled(dev_priv))
                        engine->semaphore.signal = gen8_rcs_signal;
        } else if (INTEL_GEN(dev_priv) >= 6) {
@@ -3024,12 +2981,8 @@ int intel_init_render_ring_buffer(struct drm_device *dev)
                engine->flush = gen7_render_ring_flush;
                if (IS_GEN6(dev_priv))
                        engine->flush = gen6_render_ring_flush;
-               engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
        } else if (IS_GEN5(dev_priv)) {
-               engine->add_request = pc_render_add_request;
                engine->flush = gen4_render_ring_flush;
-               engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT |
-                                       GT_RENDER_PIPECTL_NOTIFY_INTERRUPT;
        } else {
                if (INTEL_GEN(dev_priv) < 4)
                        engine->flush = gen2_render_ring_flush;
@@ -3048,7 +3001,7 @@ int intel_init_render_ring_buffer(struct drm_device *dev)
        if (ret)
                return ret;
 
-       if (INTEL_GEN(dev_priv) >= 5) {
+       if (INTEL_GEN(dev_priv) >= 6) {
                ret = intel_init_pipe_control(engine, 4096);
                if (ret)
                        return ret;