Skip to contentSkip to conversation

drm/i915: Implement dma-buf mmap() on systems with LLC

PATCH In review Rob Bradford 2 participants 2 messages Feb 24, 2014 – Mar 5, 2014

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

PATCH 1/1 drm/i915: Implement dma-buf mmap() on systems with LLC

From: Rob Bradford <[email protected]>

The LLC check echoes the check made by i915_gem_fault() which will handle the mapping the pages through the GTT.

The aim is that this is functionally equivalent to using the I915_GEM_MMAP_GTT ioctl and then mmaping using the offset returned.

---
 drivers/gpu/drm/i915/i915_gem_dmabuf.c | 20 +++++++++++++++++++-
 1 file changed, 19 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/i915_gem_dmabuf.c +19 -1
index 9bb533e..55ff592 100644
@@ -197,7 +197,25 @@ static void i915_gem_dmabuf_kunmap(struct dma_buf *dma_buf, unsigned long page_n
 
 static int i915_gem_dmabuf_mmap(struct dma_buf *dma_buf, struct vm_area_struct *vma)
 {
-	return -EINVAL;
+	struct drm_i915_gem_object *obj = dma_buf_to_obj(dma_buf);
+	struct drm_device *dev = obj->base.dev;
+	int ret;
+
+	if (!HAS_LLC(dev))
+		return -EINVAL;
+
+	ret = i915_mutex_lock_interruptible(dev);
+	if (ret)
+		goto unlock;
+
+	ret = drm_gem_mmap_obj(&obj->base, vma_pages(vma) << PAGE_SHIFT, vma);
+
+	if (ret)
+		goto unlock;
+
+unlock:
+	mutex_unlock(&dev->struct_mutex);
+	return ret;
 }
 
 static int i915_gem_begin_cpu_access(struct dma_buf *dma_buf, size_t start, size_t length, enum dma_data_direction direction)
Signature · 1 line
1.8.3.1
#1 lore raw
Quoted 24 lines · Rob Bradford

From: Rob Bradford <[email protected]>

The LLC check echoes the check made by i915_gem_fault() which will handle the mapping the pages through the GTT.

The aim is that this is functionally equivalent to using the I915_GEM_MMAP_GTT ioctl and then mmaping using the offset returned.

---
 drivers/gpu/drm/i915/i915_gem_dmabuf.c | 20 +++++++++++++++++++-
 1 file changed, 19 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/i915_gem_dmabuf.c +2 -1
index 9bb533e..55ff592 100644
@@ -197,7 +197,25 @@ static void i915_gem_dmabuf_kunmap(struct dma_buf *dma_buf, unsigned long page_n
 
 static int i915_gem_dmabuf_mmap(struct dma_buf *dma_buf, struct vm_area_struct *vma)
 {
-	return -EINVAL;
+	struct drm_i915_gem_object *obj = dma_buf_to_obj(dma_buf);
+	struct drm_device *dev = obj->base.dev;

drivers/gpu/drm/i915/i915_gem_dmabuf.c static void i915_gem_dmabuf_kunmap(struct dma_buf *dma_buf, unsigned long page_n · line 211

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_dmabuf.c +10 -0
+	int ret;
+
+	if (!HAS_LLC(dev))
+		return -EINVAL;
+
+	ret = i915_mutex_lock_interruptible(dev);
+	if (ret)
+		goto unlock;
+
+	ret = drm_gem_mmap_obj(&obj->base, vma_pages(vma) << PAGE_SHIFT, vma);

The gtt memory mapping you're forwarding here to is actually always coherent, so the LLC test is redundant. The only issue is if you try to gtt mmap a coherent BO on pre-snb, iirc that can hang.

Another thing is the pte shootdown. I think to make that work you need to frob the vma -> mapping association a bit so that unmap_mapping_range works correctly. A testcase which dma-buf mmaps more data than we have cpu accessible gtt space should easily hit this.

Which brings up the last bit: This needs a prime_mmap testcase in igt ...

Cheers, Daniel

Quoted 17 lines
drivers/gpu/drm/i915/i915_gem_dmabuf.c +7 -0
+
+	if (ret)
+		goto unlock;
+
+unlock:
+	mutex_unlock(&dev->struct_mutex);
+	return ret;
 }
 
 static int i915_gem_begin_cpu_access(struct dma_buf *dma_buf, size_t start, size_t length, enum dma_data_direction direction)
Signature · 6 lines
1.8.3.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch