drm/i915/ringbuffer: Delay after EMIT_INVALIDATE for gen4/gen5
authorChris Wilson <chris@chris-wilson.co.uk>
Mon, 5 Nov 2018 09:43:05 +0000 (09:43 +0000)
committerJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
Mon, 12 Nov 2018 15:07:12 +0000 (17:07 +0200)
Exercising the gpu reloc path strenuously revealed an issue where the
updated relocations (from MI_STORE_DWORD_IMM) were not being observed
upon execution. After some experiments with adding pipecontrols (a lot
of pipecontrols (32) as gen4/5 do not have a bit to wait on earlier pipe
controls or even the current on), it was discovered that we merely
needed to delay the EMIT_INVALIDATE by several flushes. It is important
to note that it is the EMIT_INVALIDATE as opposed to the EMIT_FLUSH that
needs the delay as opposed to what one might first expect -- that the
delay is required for the TLB invalidation to take effect (one presumes
to purge any CS buffers) as opposed to a delay after flushing to ensure
the writes have landed before triggering invalidation.

Testcase: igt/gem_tiled_fence_blits
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: stable@vger.kernel.org
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20181105094305.5767-1-chris@chris-wilson.co.uk
(cherry picked from commit 55f99bf2a9c331838c981694bc872cd1ec4070b2)
Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
drivers/gpu/drm/i915/intel_ringbuffer.c

index d0ef50b..187bb0c 100644 (file)
@@ -91,6 +91,7 @@ static int
 gen4_render_ring_flush(struct i915_request *rq, u32 mode)
 {
        u32 cmd, *cs;
+       int i;
 
        /*
         * read/write caches:
@@ -127,12 +128,45 @@ gen4_render_ring_flush(struct i915_request *rq, u32 mode)
                        cmd |= MI_INVALIDATE_ISP;
        }
 
-       cs = intel_ring_begin(rq, 2);
+       i = 2;
+       if (mode & EMIT_INVALIDATE)
+               i += 20;
+
+       cs = intel_ring_begin(rq, i);
        if (IS_ERR(cs))
                return PTR_ERR(cs);
 
        *cs++ = cmd;
-       *cs++ = MI_NOOP;
+
+       /*
+        * A random delay to let the CS invalidate take effect? Without this
+        * delay, the GPU relocation path fails as the CS does not see
+        * the updated contents. Just as important, if we apply the flushes
+        * to the EMIT_FLUSH branch (i.e. immediately after the relocation
+        * write and before the invalidate on the next batch), the relocations
+        * still fail. This implies that is a delay following invalidation
+        * that is required to reset the caches as opposed to a delay to
+        * ensure the memory is written.
+        */
+       if (mode & EMIT_INVALIDATE) {
+               *cs++ = GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE;
+               *cs++ = i915_ggtt_offset(rq->engine->scratch) |
+                       PIPE_CONTROL_GLOBAL_GTT;
+               *cs++ = 0;
+               *cs++ = 0;
+
+               for (i = 0; i < 12; i++)
+                       *cs++ = MI_FLUSH;
+
+               *cs++ = GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE;
+               *cs++ = i915_ggtt_offset(rq->engine->scratch) |
+                       PIPE_CONTROL_GLOBAL_GTT;
+               *cs++ = 0;
+               *cs++ = 0;
+       }
+
+       *cs++ = cmd;
+
        intel_ring_advance(rq, cs);
 
        return 0;