If we quickly switch from writing through the GTT to a read of the
physical page directly with the CPU (e.g. performing relocations through
the GTT and then running the command parser), we can observe that the
writes are not visible to the CPU. It is not a coherency problem, as
extensive investigations with clflush have demonstrated, but a mere
timing issue - we have to wait for the GTT to complete it's write before
we start our read from the CPU.
The issue can be illustrated in userspace with:
gtt = gem_mmap__gtt(fd, handle, 0, OBJECT_SIZE, PROT_READ | PROT_WRITE);
cpu = gem_mmap__cpu(fd, handle, 0, OBJECT_SIZE, PROT_READ | PROT_WRITE);
gem_set_domain(fd, handle, I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
for (i = 0; i < OBJECT_SIZE / 64; i++) {
int x = 16*i + (i%16);
gtt[x] = i;
clflush(&cpu[x], sizeof(cpu[x]));
assert(cpu[x] == i);
}
Experimenting with that shows that this behaviour is indeed limited to
recent Atom-class hardware.
Testcase: igt/gem_exec_flush/basic-batch-default-cmd #byt
Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@xxxxxxxxxxxxxxx>
---
drivers/gpu/drm/i915/i915_gem.c | 12 +++++++++++-
1 file changed, 11 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index e8c5f9748852..a33c7cb0549a 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -3164,20 +3164,30 @@ i915_gem_clflush_object(struct drm_i915_gem_object *obj,
static void
i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
{
+ struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
uint32_t old_write_domain;
if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
return;
/* No actual flushing is required for the GTT write domain. Writes
- * to it immediately go to main memory as far as we know, so there's
+ * to it "immediately" go to main memory as far as we know, so there's
* no chipset flush. It also doesn't land in render cache.
*
* However, we do have to enforce the order so that all writes through
* the GTT land before any writes to the device, such as updates to
* the GATT itself.
+ *
+ * We also have to wait a bit for the writes to land from the GTT.
+ * An uncached read (i.e. mmio) seems to be ideal for the round-trip
+ * timing. This issue has only been observed when switching quickly
+ * between GTT writes and CPU reads from inside the kernel on recent hw,
+ * and it appears to only affect discrete GTT blocks (i.e. on LLC
+ * system agents we cannot reproduce this behaviour).
*/
wmb();
+ if (INTEL_GEN(dev_priv) >= 6 && !HAS_LLC(dev_priv))
+ POSTING_READ(RING_ACTHD(dev_priv->engine[RCS].mmio_base));
old_write_domain = obj->base.write_domain;
obj->base.write_domain = 0;
--
2.8.1
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
