[Intel-gfx] [PATCH] drm/i915: Treat WC a separate cache domain
Chris Wilson
chris at chris-wilson.co.uk
Wed Apr 5 21:07:41 UTC 2017
When discussing a new WC mmap, we based the interface upon the
assumption that GTT was fully coherent. How naive! Commits 3b5724d702ef
("drm/i915: Wait for writes through the GTT to land before reading
back") and ed4596ea992d ("drm/i915/guc: WA to address the Ringbuffer
coherency issue") demonstrate that writes through the GTT are indeed
delayed and may be overtaken by direct WC access. To be safe, if
userspace is mixing WC mmaps with other potential GTT access (pwrite,
GTT mmaps) it should use set_domain(WC).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96563
Testcase: igt/gem_pwrite/small-gtt*
Testcase: igt/drv_selftest/coherency
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
drivers/gpu/drm/i915/i915_drv.h | 5 +-
drivers/gpu/drm/i915/i915_gem.c | 77 ++++++++++++++++++++--
drivers/gpu/drm/i915/intel_guc_log.c | 6 +-
.../gpu/drm/i915/selftests/i915_gem_coherency.c | 10 +--
drivers/gpu/drm/i915/selftests/i915_gem_request.c | 2 +-
include/uapi/drm/i915_drm.h | 2 +
6 files changed, 86 insertions(+), 16 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index c9b0949f6c1a..37d2dc7308e7 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -3447,8 +3447,9 @@ int i915_gem_object_wait_priority(struct drm_i915_gem_object *obj,
#define I915_PRIORITY_DISPLAY I915_PRIORITY_MAX
int __must_check
-i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj,
- bool write);
+i915_gem_object_set_to_wc_domain(struct drm_i915_gem_object *obj, bool write);
+int __must_check
+i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write);
int __must_check
i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write);
struct i915_vma * __must_check
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 4ca88f2539c0..497c669d0096 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -1591,10 +1591,12 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
if (err)
goto out_unpin;
- if (read_domains & I915_GEM_DOMAIN_GTT)
- err = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
+ if (read_domains & I915_GEM_DOMAIN_WC)
+ err = i915_gem_object_set_to_wc_domain(obj, write_domain);
+ else if (read_domains & I915_GEM_DOMAIN_GTT)
+ err = i915_gem_object_set_to_gtt_domain(obj, write_domain);
else
- err = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
+ err = i915_gem_object_set_to_cpu_domain(obj, write_domain);
/* And bump the LRU for this access */
i915_gem_object_bump_inactive_ggtt(obj);
@@ -1737,6 +1739,9 @@ static unsigned int tile_row_pages(struct drm_i915_gem_object *obj)
* into userspace. (This view is aligned and sized appropriately for
* fenced access.)
*
+ * 2 - Recognise WC as a separate cache domain so that we can flush the
+ * delayed writes via GTT before performing direct access via WC.
+ *
* Restrictions:
*
* * snoopable objects cannot be accessed via the GTT. It can cause machine
@@ -1764,7 +1769,7 @@ static unsigned int tile_row_pages(struct drm_i915_gem_object *obj)
*/
int i915_gem_mmap_gtt_version(void)
{
- return 1;
+ return 2;
}
static inline struct i915_ggtt_view
@@ -3390,6 +3395,70 @@ void i915_gem_object_flush_if_display(struct drm_i915_gem_object *obj)
}
/**
+ * Moves a single object to the WC read, and possibly write domain.
+ * @obj: object to act on
+ * @write: ask for write access or read only
+ *
+ * This function returns when the move is complete, including waiting on
+ * flushes to occur.
+ */
+int
+i915_gem_object_set_to_wc_domain(struct drm_i915_gem_object *obj, bool write)
+{
+ int ret;
+
+ lockdep_assert_held(&obj->base.dev->struct_mutex);
+
+ ret = i915_gem_object_wait(obj,
+ I915_WAIT_INTERRUPTIBLE |
+ I915_WAIT_LOCKED |
+ (write ? I915_WAIT_ALL : 0),
+ MAX_SCHEDULE_TIMEOUT,
+ NULL);
+ if (ret)
+ return ret;
+
+ if (obj->base.write_domain == I915_GEM_DOMAIN_WC)
+ return 0;
+
+ /* Flush and acquire obj->pages so that we are coherent through
+ * direct access in memory with previous cached writes through
+ * shmemfs and that our cache domain tracking remains valid.
+ * For example, if the obj->filp was moved to swap without us
+ * being notified and releasing the pages, we would mistakenly
+ * continue to assume that the obj remained out of the CPU cached
+ * domain.
+ */
+ ret = i915_gem_object_pin_pages(obj);
+ if (ret)
+ return ret;
+
+ i915_gem_object_flush_cpu_write_domain(obj);
+ i915_gem_object_flush_gtt_write_domain(obj);
+
+ /* Serialise direct access to this object with the barriers for
+ * coherent writes from the GPU, by effectively invalidating the
+ * WC domain upon first access.
+ */
+ if ((obj->base.read_domains & I915_GEM_DOMAIN_WC) == 0)
+ mb();
+
+ /* It should now be out of any other write domains, and we can update
+ * the domain values for our changes.
+ */
+ GEM_BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_WC) != 0);
+ obj->base.read_domains |= I915_GEM_DOMAIN_WC;
+ if (write) {
+ obj->base.read_domains = I915_GEM_DOMAIN_WC;
+ obj->base.write_domain = I915_GEM_DOMAIN_WC;
+ obj->mm.dirty = true;
+ }
+
+ i915_gem_object_unpin_pages(obj);
+ return 0;
+}
+
+/**
* Moves a single object to the GTT read, and possibly write domain.
* @obj: object to act on
* @write: ask for write access or read only
diff --git a/drivers/gpu/drm/i915/intel_guc_log.c b/drivers/gpu/drm/i915/intel_guc_log.c
index 6fb63a3c65b0..16d3b8719cab 100644
--- a/drivers/gpu/drm/i915/intel_guc_log.c
+++ b/drivers/gpu/drm/i915/intel_guc_log.c
@@ -359,12 +359,16 @@ static int guc_log_runtime_create(struct intel_guc *guc)
void *vaddr;
struct rchan *guc_log_relay_chan;
size_t n_subbufs, subbuf_size;
- int ret = 0;
+ int ret;
lockdep_assert_held(&dev_priv->drm.struct_mutex);
GEM_BUG_ON(guc_log_has_runtime(guc));
+ ret = i915_gem_object_set_to_wc_domain(guc->log.vma->obj, true);
+ if (ret)
+ return ret;
+
/* Create a WC (Uncached for read) vmalloc mapping of log
* buffer pages, so that we can directly get the data
* (up-to-date) from memory.
diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_coherency.c b/drivers/gpu/drm/i915/selftests/i915_gem_coherency.c
index f08d0179b3df..95d4aebc0181 100644
--- a/drivers/gpu/drm/i915/selftests/i915_gem_coherency.c
+++ b/drivers/gpu/drm/i915/selftests/i915_gem_coherency.c
@@ -138,10 +138,7 @@ static int wc_set(struct drm_i915_gem_object *obj,
typeof(v) *map;
int err;
- /* XXX GTT write followed by WC write go missing */
- i915_gem_object_flush_gtt_write_domain(obj);
-
- err = i915_gem_object_set_to_gtt_domain(obj, true);
+ err = i915_gem_object_set_to_wc_domain(obj, true);
if (err)
return err;
@@ -162,10 +159,7 @@ static int wc_get(struct drm_i915_gem_object *obj,
typeof(v) map;
int err;
- /* XXX WC write followed by GTT write go missing */
- i915_gem_object_flush_gtt_write_domain(obj);
-
- err = i915_gem_object_set_to_gtt_domain(obj, false);
+ err = i915_gem_object_set_to_wc_domain(obj, false);
if (err)
return err;
diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_request.c b/drivers/gpu/drm/i915/selftests/i915_gem_request.c
index 98b7aac41eec..6664cb2eb0b8 100644
--- a/drivers/gpu/drm/i915/selftests/i915_gem_request.c
+++ b/drivers/gpu/drm/i915/selftests/i915_gem_request.c
@@ -580,7 +580,7 @@ static struct i915_vma *recursive_batch(struct drm_i915_private *i915)
if (err)
goto err;
- err = i915_gem_object_set_to_gtt_domain(obj, true);
+ err = i915_gem_object_set_to_wc_domain(obj, true);
if (err)
goto err;
diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
index 3554495bef13..9ee06ec8a2d6 100644
--- a/include/uapi/drm/i915_drm.h
+++ b/include/uapi/drm/i915_drm.h
@@ -666,6 +666,8 @@ struct drm_i915_gem_relocation_entry {
#define I915_GEM_DOMAIN_VERTEX 0x00000020
/** GTT domain - aperture and scanout */
#define I915_GEM_DOMAIN_GTT 0x00000040
+/** WC domain - uncached access */
+#define I915_GEM_DOMAIN_WC 0x00000080
/** @} */
struct drm_i915_gem_exec_object {
--
2.11.0
More information about the Intel-gfx
mailing list