drm/i915: add workarounds to gen7_render_ring_flush
From Bspec, Vol 2a, Section 1.9.3.4 "PIPE_CONTROL", intro section
detailing the various workarounds:
"[DevIVB {W/A}, DevHSW {W/A}]: Pipe_control with CS-stall bit
set must be issued before a pipe-control command that has the State
Cache Invalidate bit set."
Note that public Bspec has different numbering, it's Vol2Part1,
Section 1.10.4.1 "PIPE_CONTROL" there.
There's also a second workaround for the PIPE_CONTROL command itself:
"[DevIVB, DevVLV, DevHSW] {WA}: Every 4th PIPE_CONTROL command, not
counting the PIPE_CONTROL with only read-cache-invalidate bit(s) set,
must have a CS_STALL bit set"
For simplicity we simply set the CS_STALL bit on every pipe_control on
gen7+
Note that this massively helps on some hsw machines, together with the
following patch to unconditionally set the CS_STALL bit on every
pipe_control it prevents a gpu hang every few seconds.
This is a regression that has been introduced in the pipe_control
cleanup:
commit 6c6cf5aa9c583478b19e23149feaa92d01fb8c2d
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Fri Jul 20 18:02:28 2012 +0100
drm/i915: Only apply the SNB pipe control w/a to gen6
It looks like the massive snb pipe_control workaround also papered
over any issues on ivb and hsw.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
[danvet: squashed both workarounds together, pimped commit message
with Bsepc citations, regression commit citation and changed the
comment in the code a bit to clarify that we unconditionally set
CS_STALL to avoid being hurt by trying to be clever.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
index 42a4b85..55cdb4d 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.c
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
@@ -263,6 +263,25 @@
}
static int
+gen7_render_ring_cs_stall_wa(struct intel_ring_buffer *ring)
+{
+ int ret;
+
+ ret = intel_ring_begin(ring, 4);
+ if (ret)
+ return ret;
+
+ intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
+ intel_ring_emit(ring, PIPE_CONTROL_CS_STALL |
+ PIPE_CONTROL_STALL_AT_SCOREBOARD);
+ intel_ring_emit(ring, 0);
+ intel_ring_emit(ring, 0);
+ intel_ring_advance(ring);
+
+ return 0;
+}
+
+static int
gen7_render_ring_flush(struct intel_ring_buffer *ring,
u32 invalidate_domains, u32 flush_domains)
{
@@ -271,6 +290,16 @@
u32 scratch_addr = pc->gtt_offset + 128;
int ret;
+ /*
+ * Ensure that any following seqno writes only happen when the render
+ * cache is indeed flushed.
+ *
+ * Workaround: 4th PIPE_CONTROL command (except the ones with only
+ * read-cache invalidate bits set) must have the CS_STALL bit set. We
+ * don't try to be clever and just set it unconditionally.
+ */
+ flags |= PIPE_CONTROL_CS_STALL;
+
/* Just flush everything. Experiments have shown that reducing the
* number of bits based on the write domains has little performance
* impact.
@@ -278,11 +307,6 @@
if (flush_domains) {
flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
- /*
- * Ensure that any following seqno writes only happen
- * when the render cache is indeed flushed.
- */
- flags |= PIPE_CONTROL_CS_STALL;
}
if (invalidate_domains) {
flags |= PIPE_CONTROL_TLB_INVALIDATE;
@@ -295,6 +319,11 @@
* TLB invalidate requires a post-sync write.
*/
flags |= PIPE_CONTROL_QW_WRITE;
+
+ /* Workaround: we must issue a pipe_control with CS-stall bit
+ * set before a pipe_control command that has the state cache
+ * invalidate bit set. */
+ gen7_render_ring_cs_stall_wa(ring);
}
ret = intel_ring_begin(ring, 4);