Skip to contentSkip to conversation

drm/i915: clflush on pin_to_display after pwrite to UC bo in LLC

PATCH Superseded [email protected] 4 participants 11 messages Aug 11, 2015 – Aug 15, 2015

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/i915: clflush on pin_to_display after pwrite to UC bo in LLC

From: Ville Syrjälä <[email protected]>

Currently we don't clflush on pin_to_display if the bo is already UC/WT and is not in the CPU write domain. This causes problems with pwrite since pwrite doesn't change the write domain, and it avoids clflushing on UC/WT buffers on LLC platforms unless the buffer is currently being scanned out.

Fix the problem by marking the cache dirty and adjusting i915_gem_object_set_cache_level() to clflush when the cache is dirty even if the cache_level doesn't change.

My last attempt [1] at fixing this via write domain frobbing was shot down, but now with the cache_dirty flag we can do things in a nicer way.

[1] http://lists.freedesktop.org/archives/intel-gfx/2014-November/055390.html

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86422
Testcase: igt/kms_pwrite_crc

---
 drivers/gpu/drm/i915/i915_gem.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +7 -3
index 73293b4..73eff2e 100644
@@ -1005,12 +1005,15 @@ out:
 		if (!needs_clflush_after &&
 		    obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
 			if (i915_gem_clflush_object(obj, obj->pin_display))
-				i915_gem_chipset_flush(dev);
+				needs_clflush_after = true;
 		}
 	}
 
 	if (needs_clflush_after)
 		i915_gem_chipset_flush(dev);
+	else if (obj->cache_level == I915_CACHE_NONE ||
+		 obj->cache_level == I915_CACHE_WT)
+		obj->cache_dirty = true;
 
 	intel_fb_obj_flush(obj, false, ORIGIN_CPU);
 	return ret;
@@ -3639,10 +3642,10 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 {
 	struct drm_device *dev = obj->base.dev;
 	struct i915_vma *vma, *next;
-	int ret;
+	int ret = 0;
 
 	if (obj->cache_level == cache_level)
-		return 0;
+		goto out;
 
 	if (i915_gem_obj_is_pinned(obj)) {
 		DRM_DEBUG("can not change the cache level of pinned objects\n");
@@ -3687,6 +3690,7 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 		vma->node.color = cache_level;
 	obj->cache_level = cache_level;
 
+out:
 	if (obj->cache_dirty &&
 	    obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
 	    cpu_write_needs_clflush(obj)) {
Signature · 6 lines
2.4.6

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#1 lore raw

PATCH 1/1 tests/kms_pwrite_crc: Use drmModeSetPlane() instead of igt_plane_set_fb()

From: Ville Syrjälä <[email protected]>

igt_plane_set_fb()+igt_display_commit() have too much overhead, and that causes the cache to get flushed before we flip, making the test useless, at least on machines with small LLC. Switch to drmModeSetPlane() to reduce the chance that the cache gets flushed before we grab the crc.

Still nowhere near 100% reliable on my IVB laptop with 3 MiB LLC, but at least it can now hit the problem occasioanally. My desktop IVB with 8 MiB LLC seems to hit it rather reliably.

---
 tests/kms_pwrite_crc.c | 25 ++++++++++++++++++-------
 1 file changed, 18 insertions(+), 7 deletions(-)
tests/kms_pwrite_crc.c +18 -7
index 05b9e38..990866f 100644
@@ -52,7 +52,6 @@ typedef struct {
 
 static void test(data_t *data)
 {
-	igt_display_t *display = &data->display;
 	igt_output_t *output = data->output;
 	struct igt_fb *fb = &data->fb[1];
 	drmModeModeInfo *mode;
@@ -72,12 +71,20 @@ static void test(data_t *data)
 	cairo_destroy(cr);
 
 	/* flip to it to make it UC/WC and fully flushed */
-	igt_plane_set_fb(data->primary, fb);
-	igt_display_commit(display);
+	drmModeSetPlane(data->drm_fd,
+			data->primary->drm_plane->plane_id,
+			output->config.crtc->crtc_id,
+			fb->fb_id, 0,
+			0, 0, fb->width, fb->height,
+			0, 0, fb->width << 16, fb->height << 16);
 
 	/* flip back the original white buffer */
-	igt_plane_set_fb(data->primary, &data->fb[0]);
-	igt_display_commit(display);
+	drmModeSetPlane(data->drm_fd,
+			data->primary->drm_plane->plane_id,
+			output->config.crtc->crtc_id,
+			data->fb[0].fb_id, 0,
+			0, 0, fb->width, fb->height,
+			0, 0, fb->width << 16, fb->height << 16);
 
 	/* make sure caching mode has become UC/WT */
 	caching = gem_get_caching(data->drm_fd, fb->gem_handle);
@@ -91,8 +98,12 @@ static void test(data_t *data)
 	free(buf);
 
 	/* and flip to it */
-	igt_plane_set_fb(data->primary, fb);
-	igt_display_commit(display);
+	drmModeSetPlane(data->drm_fd,
+			data->primary->drm_plane->plane_id,
+			output->config.crtc->crtc_id,
+			fb->fb_id, 0,
+			0, 0, fb->width, fb->height,
+			0, 0, fb->width << 16, fb->height << 16);
 
 	/* check that the crc is as expected, which requires that caches got flushed */
 	igt_pipe_crc_collect_crc(data->pipe_crc, &crc);
Signature · 6 lines
2.4.6

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Reviewed-by on v1 1/1
#2 lore raw
In reply to [email protected]
Quoted 35 lines · On Tue, Aug 11, 2015 at 04:59:14PM +0300, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

Currently we don't clflush on pin_to_display if the bo is already UC/WT and is not in the CPU write domain. This causes problems with pwrite since pwrite doesn't change the write domain, and it avoids clflushing on UC/WT buffers on LLC platforms unless the buffer is currently being scanned out.

Fix the problem by marking the cache dirty and adjusting i915_gem_object_set_cache_level() to clflush when the cache is dirty even if the cache_level doesn't change.

My last attempt [1] at fixing this via write domain frobbing was shot down, but now with the cache_dirty flag we can do things in a nicer way.

[1] http://lists.freedesktop.org/archives/intel-gfx/2014-November/055390.html

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86422
Testcase: igt/kms_pwrite_crc

---
 drivers/gpu/drm/i915/i915_gem.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +1 -1
index 73293b4..73eff2e 100644
@@ -1005,12 +1005,15 @@ out:
 		if (!needs_clflush_after &&
 		    obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
 			if (i915_gem_clflush_object(obj, obj->pin_display))
-				i915_gem_chipset_flush(dev);
+				needs_clflush_after = true;
 		}

drivers/gpu/drm/i915/i915_gem.c out: · line 1019

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem.c +3 -0
 	}
 
 	if (needs_clflush_after)
 		i915_gem_chipset_flush(dev);
+	else if (obj->cache_level == I915_CACHE_NONE ||
+		 obj->cache_level == I915_CACHE_WT)
+		obj->cache_dirty = true;
 
 	intel_fb_obj_flush(obj, false, ORIGIN_CPU);
 	return ret;

Ok, this seems reasonable.

Quoted 14 lines
drivers/gpu/drm/i915/i915_gem.c +2 -2
@@ -3639,10 +3642,10 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 {
 	struct drm_device *dev = obj->base.dev;
 	struct i915_vma *vma, *next;
-	int ret;
+	int ret = 0;
 
 	if (obj->cache_level == cache_level)
-		return 0;
+		goto out;
 
 	if (i915_gem_obj_is_pinned(obj)) {
 		DRM_DEBUG("can not change the cache level of pinned objects\n");

drivers/gpu/drm/i915/i915_gem.c int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, · line 3696

Quoted 7 lines
drivers/gpu/drm/i915/i915_gem.c +1 -0
@@ -3687,6 +3690,7 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 		vma->node.color = cache_level;
 	obj->cache_level = cache_level;
 
+out:
 	if (obj->cache_dirty &&
 	    obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
 	    cpu_write_needs_clflush(obj)) {

This we can do better (and I know I am guilty of the original sin). It just feels a little too loose that we expect pin-to-display-plane should always call set-cache-level (yes, it has to, but it feels like we are putting pin-to-display-plane specifics into set-cache-level).

I think this chunk is much more understandable if we did:

drivers/gpu/drm/i915/i915_gem.c +4 -7
index 5e7fcf77e57a..fc6bcc19cca1 100644
@@ -3656,13 +3656,6 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
                vma->node.color = cache_level;
        obj->cache_level = cache_level;
 
-       if (obj->cache_dirty &&
-           obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
-           cpu_write_needs_clflush(obj)) {
-               if (i915_gem_clflush_object(obj, true))
-                       i915_gem_chipset_flush(obj->base.dev);
-       }
-
        return 0;
 }
 
@@ -3795,6 +3788,10 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
        WARN_ON(obj->pin_display > vma->pin_count);
 
        i915_gem_object_flush_cpu_write_domain(obj);
+       if (obj->cache_dirty) {
+               if (i915_gem_clflush_object(obj, true))
+                       i915_gem_chipset_flush(obj->base.dev);
+       }
 
        /* It should now be out of any other write domains, and we can update
         * the domain values for our changes.

Maybe even

/* Whilst the object was away from the scanout we may have been eliding the coherent
* writes into the CPU cache. However, the moment it has to be read by the display
* engine, those writes into the CPU cache become inaccessible and so we have to
* clflush them out to main memory. We track elided flushes with obj->cache_dirty
* and hope they are rare.
*/

The other user of set-cache-level (set_caching_ioctl) should not care about the clflush side-effect and the clflush elision should work just fine instead.

Either way,

but I'd prefer a v2 with the comments :)
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#3 lore raw
Quoted 118 lines · Chris Wilson
Quoted 35 lines · On Tue, Aug 11, 2015 at 04:59:14PM +0300, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

Currently we don't clflush on pin_to_display if the bo is already UC/WT and is not in the CPU write domain. This causes problems with pwrite since pwrite doesn't change the write domain, and it avoids clflushing on UC/WT buffers on LLC platforms unless the buffer is currently being scanned out.

Fix the problem by marking the cache dirty and adjusting i915_gem_object_set_cache_level() to clflush when the cache is dirty even if the cache_level doesn't change.

My last attempt [1] at fixing this via write domain frobbing was shot down, but now with the cache_dirty flag we can do things in a nicer way.

[1] http://lists.freedesktop.org/archives/intel-gfx/2014-November/055390.html

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86422
Testcase: igt/kms_pwrite_crc

---
 drivers/gpu/drm/i915/i915_gem.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +1 -1
index 73293b4..73eff2e 100644
@@ -1005,12 +1005,15 @@ out:
 		if (!needs_clflush_after &&
 		    obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
 			if (i915_gem_clflush_object(obj, obj->pin_display))
-				i915_gem_chipset_flush(dev);
+				needs_clflush_after = true;
 		}

drivers/gpu/drm/i915/i915_gem.c out: · line 1019

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem.c +3 -0
 	}
 
 	if (needs_clflush_after)
 		i915_gem_chipset_flush(dev);
+	else if (obj->cache_level == I915_CACHE_NONE ||
+		 obj->cache_level == I915_CACHE_WT)
+		obj->cache_dirty = true;
 
 	intel_fb_obj_flush(obj, false, ORIGIN_CPU);
 	return ret;

Ok, this seems reasonable.

Quoted 14 lines
drivers/gpu/drm/i915/i915_gem.c +2 -2
@@ -3639,10 +3642,10 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 {
 	struct drm_device *dev = obj->base.dev;
 	struct i915_vma *vma, *next;
-	int ret;
+	int ret = 0;
 
 	if (obj->cache_level == cache_level)
-		return 0;
+		goto out;
 
 	if (i915_gem_obj_is_pinned(obj)) {
 		DRM_DEBUG("can not change the cache level of pinned objects\n");

drivers/gpu/drm/i915/i915_gem.c int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, · line 3696

Quoted 7 lines
drivers/gpu/drm/i915/i915_gem.c +1 -0
@@ -3687,6 +3690,7 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 		vma->node.color = cache_level;
 	obj->cache_level = cache_level;
 
+out:
 	if (obj->cache_dirty &&
 	    obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
 	    cpu_write_needs_clflush(obj)) {

This we can do better (and I know I am guilty of the original sin). It just feels a little too loose that we expect pin-to-display-plane should always call set-cache-level (yes, it has to, but it feels like we are putting pin-to-display-plane specifics into set-cache-level).

I think this chunk is much more understandable if we did:

drivers/gpu/drm/i915/i915_gem.c +4 -7
index 5e7fcf77e57a..fc6bcc19cca1 100644
@@ -3656,13 +3656,6 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
                vma->node.color = cache_level;
        obj->cache_level = cache_level;
 
-       if (obj->cache_dirty &&
-           obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
-           cpu_write_needs_clflush(obj)) {
-               if (i915_gem_clflush_object(obj, true))
-                       i915_gem_chipset_flush(obj->base.dev);
-       }
-
        return 0;
 }
 
@@ -3795,6 +3788,10 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
        WARN_ON(obj->pin_display > vma->pin_count);
 
        i915_gem_object_flush_cpu_write_domain(obj);
+       if (obj->cache_dirty) {
+               if (i915_gem_clflush_object(obj, true))
+                       i915_gem_chipset_flush(obj->base.dev);
+       }
 
        /* It should now be out of any other write domains, and we can update
         * the domain values for our changes.

Maybe even

/* Whilst the object was away from the scanout we may have been eliding the coherent
* writes into the CPU cache. However, the moment it has to be read by the display
* engine, those writes into the CPU cache become inaccessible and so we have to
* clflush them out to main memory. We track elided flushes with obj->cache_dirty
* and hope they are rare.
*/

Quoted 3 lines

The other user of set-cache-level (set_caching_ioctl) should not care about the clflush side-effect and the clflush elision should work just fine instead.

Hmm. So what would happen on !LLC if we start with a cached bo, then pwrite it and afterwards make it uncached?

Quoted 8 lines

Either way,

but I'd prefer a v2 with the comments :)
-Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
Signature · 6 lines
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#4 lore raw
Quoted 119 lines · Ville Syrjälä
Quoted 118 lines · Chris Wilson
Quoted 35 lines · On Tue, Aug 11, 2015 at 04:59:14PM +0300, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

Currently we don't clflush on pin_to_display if the bo is already UC/WT and is not in the CPU write domain. This causes problems with pwrite since pwrite doesn't change the write domain, and it avoids clflushing on UC/WT buffers on LLC platforms unless the buffer is currently being scanned out.

Fix the problem by marking the cache dirty and adjusting i915_gem_object_set_cache_level() to clflush when the cache is dirty even if the cache_level doesn't change.

My last attempt [1] at fixing this via write domain frobbing was shot down, but now with the cache_dirty flag we can do things in a nicer way.

[1] http://lists.freedesktop.org/archives/intel-gfx/2014-November/055390.html

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86422
Testcase: igt/kms_pwrite_crc

---
 drivers/gpu/drm/i915/i915_gem.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +1 -1
index 73293b4..73eff2e 100644
@@ -1005,12 +1005,15 @@ out:
 		if (!needs_clflush_after &&
 		    obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
 			if (i915_gem_clflush_object(obj, obj->pin_display))
-				i915_gem_chipset_flush(dev);
+				needs_clflush_after = true;
 		}

drivers/gpu/drm/i915/i915_gem.c out: · line 1019

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem.c +3 -0
 	}
 
 	if (needs_clflush_after)
 		i915_gem_chipset_flush(dev);
+	else if (obj->cache_level == I915_CACHE_NONE ||
+		 obj->cache_level == I915_CACHE_WT)
+		obj->cache_dirty = true;
 
 	intel_fb_obj_flush(obj, false, ORIGIN_CPU);
 	return ret;

Ok, this seems reasonable.

Quoted 14 lines
drivers/gpu/drm/i915/i915_gem.c +2 -2
@@ -3639,10 +3642,10 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 {
 	struct drm_device *dev = obj->base.dev;
 	struct i915_vma *vma, *next;
-	int ret;
+	int ret = 0;
 
 	if (obj->cache_level == cache_level)
-		return 0;
+		goto out;
 
 	if (i915_gem_obj_is_pinned(obj)) {
 		DRM_DEBUG("can not change the cache level of pinned objects\n");

drivers/gpu/drm/i915/i915_gem.c int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, · line 3696

Quoted 7 lines
drivers/gpu/drm/i915/i915_gem.c +1 -0
@@ -3687,6 +3690,7 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 		vma->node.color = cache_level;
 	obj->cache_level = cache_level;
 
+out:
 	if (obj->cache_dirty &&
 	    obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
 	    cpu_write_needs_clflush(obj)) {

This we can do better (and I know I am guilty of the original sin). It just feels a little too loose that we expect pin-to-display-plane should always call set-cache-level (yes, it has to, but it feels like we are putting pin-to-display-plane specifics into set-cache-level).

I think this chunk is much more understandable if we did:

drivers/gpu/drm/i915/i915_gem.c +4 -7
index 5e7fcf77e57a..fc6bcc19cca1 100644
@@ -3656,13 +3656,6 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
                vma->node.color = cache_level;
        obj->cache_level = cache_level;
 
-       if (obj->cache_dirty &&
-           obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
-           cpu_write_needs_clflush(obj)) {
-               if (i915_gem_clflush_object(obj, true))
-                       i915_gem_chipset_flush(obj->base.dev);
-       }
-
        return 0;
 }
 
@@ -3795,6 +3788,10 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
        WARN_ON(obj->pin_display > vma->pin_count);
 
        i915_gem_object_flush_cpu_write_domain(obj);
+       if (obj->cache_dirty) {
+               if (i915_gem_clflush_object(obj, true))
+                       i915_gem_chipset_flush(obj->base.dev);
+       }
 
        /* It should now be out of any other write domains, and we can update
         * the domain values for our changes.

Maybe even

/* Whilst the object was away from the scanout we may have been eliding the coherent
* writes into the CPU cache. However, the moment it has to be read by the display
* engine, those writes into the CPU cache become inaccessible and so we have to
* clflush them out to main memory. We track elided flushes with obj->cache_dirty
* and hope they are rare.
*/

Quoted 6 lines
Quoted 3 lines

The other user of set-cache-level (set_caching_ioctl) should not care about the clflush side-effect and the clflush elision should work just fine instead.

Hmm. So what would happen on !LLC if we start with a cached bo, then pwrite it and afterwards make it uncached?

In fact that would still fail even with my patch, and wouldn't work with current upstream code either. To fix that I'd need to drop the I915_CACHE_NONE/WT checks from pwrite in my patch and always set cache_dirty=true when it didn't clflush. Doing that would seem perfectly reasonable to me.

Signature · 6 lines
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#5 lore raw
Quoted 8 lines · Ville Syrjälä
Quoted 2 lines · Ville Syrjälä

Hmm. So what would happen on !LLC if we start with a cached bo, then pwrite it and afterwards make it uncached?

In fact that would still fail even with my patch, and wouldn't work with current upstream code either. To fix that I'd need to drop the I915_CACHE_NONE/WT checks from pwrite in my patch and always set cache_dirty=true when it didn't clflush. Doing that would seem perfectly reasonable to me.

Yes. I agree. Marking obj->cache_dirty whenever we dirty the cpu cache and clear it upon clflush seems sane. It will only take effect on transition to the display plane, so not going to impact us except when correctness is required. So maybe we should call it obj->display_dirty? -Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#6 lore raw

PATCH v2 1/1 drm/i915: clflush on pin_to_display after pwrite to UC bo in LLC

In reply to [email protected]

From: Ville Syrjälä <[email protected]>

Currently we don't clflush on pin_to_display if the bo is already UC/WT and is not in the CPU write domain. This causes problems with pwrite since pwrite doesn't change the write domain, and it avoids clflushing on UC/WT buffers on LLC platforms unless the buffer is currently being scanned out.

Fix the problem by marking the cache dirty and adjusting i915_gem_object_set_cache_level() to clflush when the cache is dirty even if the cache_level doesn't change.

My last attempt [1] at fixing this via write domain frobbing was shot down, but now with the cache_dirty flag we can do things in a nicer way.

[1] http://lists.freedesktop.org/archives/intel-gfx/2014-November/055390.html

v2: Drop the I915_CACHE_NONE/WT checks from pwrite

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86422
Testcase: igt/kms_pwrite_crc
Testcase: igt/gem_pwrite_snooped

---
 drivers/gpu/drm/i915/i915_gem.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +6 -3
index 73293b4..f828dc7 100644
@@ -1005,12 +1005,14 @@ out:
 		if (!needs_clflush_after &&
 		    obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
 			if (i915_gem_clflush_object(obj, obj->pin_display))
-				i915_gem_chipset_flush(dev);
+				needs_clflush_after = true;
 		}
 	}
 
 	if (needs_clflush_after)
 		i915_gem_chipset_flush(dev);
+	else
+		obj->cache_dirty = true;
 
 	intel_fb_obj_flush(obj, false, ORIGIN_CPU);
 	return ret;
@@ -3639,10 +3641,10 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 {
 	struct drm_device *dev = obj->base.dev;
 	struct i915_vma *vma, *next;
-	int ret;
+	int ret = 0;
 
 	if (obj->cache_level == cache_level)
-		return 0;
+		goto out;
 
 	if (i915_gem_obj_is_pinned(obj)) {
 		DRM_DEBUG("can not change the cache level of pinned objects\n");
@@ -3687,6 +3689,7 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 		vma->node.color = cache_level;
 	obj->cache_level = cache_level;
 
+out:
 	if (obj->cache_dirty &&
 	    obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
 	    cpu_write_needs_clflush(obj)) {
Signature · 6 lines
2.4.6

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Reviewed-by on v2 1/1
#7 lore raw
Quoted 15 lines · On Tue, Aug 11, 2015 at 07:47:10PM +0300, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

Currently we don't clflush on pin_to_display if the bo is already UC/WT and is not in the CPU write domain. This causes problems with pwrite since pwrite doesn't change the write domain, and it avoids clflushing on UC/WT buffers on LLC platforms unless the buffer is currently being scanned out.

Fix the problem by marking the cache dirty and adjusting i915_gem_object_set_cache_level() to clflush when the cache is dirty even if the cache_level doesn't change.

My last attempt [1] at fixing this via write domain frobbing was shot down, but now with the cache_dirty flag we can do things in a nicer way.

Quoted 8 lines

[1] http://lists.freedesktop.org/archives/intel-gfx/2014-November/055390.html

v2: Drop the I915_CACHE_NONE/WT checks from pwrite

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86422
Testcase: igt/kms_pwrite_crc
Testcase: igt/gem_pwrite_snooped

-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#8 lore raw
Quoted 18 lines · Chris Wilson
Quoted 17 lines · On Tue, Aug 11, 2015 at 07:47:10PM +0300, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

Currently we don't clflush on pin_to_display if the bo is already UC/WT and is not in the CPU write domain. This causes problems with pwrite since pwrite doesn't change the write domain, and it avoids clflushing on UC/WT buffers on LLC platforms unless the buffer is currently being scanned out.

Fix the problem by marking the cache dirty and adjusting i915_gem_object_set_cache_level() to clflush when the cache is dirty even if the cache_level doesn't change.

My last attempt [1] at fixing this via write domain frobbing was shot down, but now with the cache_dirty flag we can do things in a nicer way.

[1] http://lists.freedesktop.org/archives/intel-gfx/2014-November/055390.html

Quoted 7 lines
Quoted 6 lines

v2: Drop the I915_CACHE_NONE/WT checks from pwrite

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86422
Testcase: igt/kms_pwrite_crc
Testcase: igt/gem_pwrite_snooped

Queued for -next, thanks for the patch.
-Daniel

Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Tested-by on v2 1/1
#9 lore raw
In reply to [email protected]

Task id: 7152 -------------------------------------Summary-------------------------------------
Platform Delta drm-intel-nightly Series Applied

ILK                                  302/302              302/302
SNB                                  315/315              315/315
IVB                                  336/336              336/336
BYT                                  283/283              283/283
HSW                                  378/378              378/378

-------------------------------------Detailed-------------------------------------
Platform Test drm-intel-nightly Series Applied
Note: You need to pay more attention to line start with '*'
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#10 lore raw

PATCH 1/1 tests/gem_pwrite_snooped: Verify set_caching vs. pwrite clflush behaviour

In reply to [email protected]

From: Ville Syrjälä <[email protected]>

The test does the following
1. set_domain src GTT
2. set_caching src NONE
3. pwrite src
4. set_caching src CACHED
5. blt src->dst
6. pread dst
7. verify data matches

---
 tests/Makefile.sources     |   1 +
 tests/gem_pwrite_snooped.c | 140 +++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 141 insertions(+)
 create mode 100644 tests/gem_pwrite_snooped.c
tests/Makefile.sources +1 -0
index c94714b..b9a4cb4 100644
@@ -47,6 +47,7 @@ TESTS_progs_M = \
 	gem_pread_after_blit \
 	gem_pwrite \
 	gem_pwrite_pread \
+	gem_pwrite_snooped \
 	gem_readwrite \
 	gem_read_read_speed \
 	gem_reloc_overflow \
tests/gem_pwrite_snooped.c new file +140 -0

140 lines not on this page · raw message

Signature · 6 lines
2.4.6

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx