drm/i915: implement inline clflush for pwrite

In micro-benchmarking of the usual pwrite use-pattern of alternating
pwrites with gtt domain reads from the gpu, this yields around 30%
improvement of pwrite throughput across all buffers size. The trick is
that we can avoid clflush cachelines that we will overwrite completely
anyway.

Furthermore for partial pwrites it gives a proportional speedup on top
of the 30% percent because we only clflush back the part of the buffer
we're actually writing.

v2: Simplify the clflush-before-write logic, as suggested by Chris
Wilson.

v3: Finishing touches suggested by Chris Wilson:
- add comment to needs_clflush_before and only set this if the bo is
  uncached.
- s/needs_clflush/needs_clflush_after/ in the write paths to clearly
  differentiate it from needs_clflush_before.

Tested-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drivers/gpu/drm/i915/i915_gem.c

@@ -570,23 +570,39 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
 	int shmem_page_offset, page_length, ret = 0;
 	int obj_do_bit17_swizzling, page_do_bit17_swizzling;
 	int hit_slowpath = 0;
+	int needs_clflush_after = 0;
+	int needs_clflush_before = 0;
 	int release_page;
 
-	ret = i915_gem_object_set_to_cpu_domain(obj, 1);
-	if (ret)
-		return ret;
-
 	user_data = (char __user *) (uintptr_t) args->data_ptr;
 	remain = args->size;
 
 	obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
 
+	if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
+		/* If we're not in the cpu write domain, set ourself into the gtt
+		 * write domain and manually flush cachelines (if required). This
+		 * optimizes for the case when the gpu will use the data
+		 * right away and we therefore have to clflush anyway. */
+		if (obj->cache_level == I915_CACHE_NONE)
+			needs_clflush_after = 1;
+		ret = i915_gem_object_set_to_gtt_domain(obj, true);
+		if (ret)
+			return ret;
+	}
+	/* Same trick applies for invalidate partially written cachelines before
+	 * writing.  */
+	if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)
+	    && obj->cache_level == I915_CACHE_NONE)
+		needs_clflush_before = 1;
+
 	offset = args->offset;
 	obj->dirty = 1;
 
 	while (remain > 0) {
 		struct page *page;
 		char *vaddr;
+		int partial_cacheline_write;
 
 		/* Operation in this page
 		 *
@@ -599,6 +615,13 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
 		if ((shmem_page_offset + page_length) > PAGE_SIZE)
 			page_length = PAGE_SIZE - shmem_page_offset;
 
+		/* If we don't overwrite a cacheline completely we need to be
+		 * careful to have up-to-date data by first clflushing. Don't
+		 * overcomplicate things and flush the entire patch. */
+		partial_cacheline_write = needs_clflush_before &&
+			((shmem_page_offset | page_length)
+				& (boot_cpu_data.x86_clflush_size - 1));
+
 		if (obj->pages) {
 			page = obj->pages[offset >> PAGE_SHIFT];
 			release_page = 0;
@@ -616,9 +639,15 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
 
 		if (!page_do_bit17_swizzling) {
 			vaddr = kmap_atomic(page);
+			if (partial_cacheline_write)
+				drm_clflush_virt_range(vaddr + shmem_page_offset,
+						       page_length);
 			ret = __copy_from_user_inatomic(vaddr + shmem_page_offset,
 							user_data,
 							page_length);
+			if (needs_clflush_after)
+				drm_clflush_virt_range(vaddr + shmem_page_offset,
+						       page_length);
 			kunmap_atomic(vaddr);
 
 			if (ret == 0)
@@ -630,6 +659,9 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
 		mutex_unlock(&dev->struct_mutex);
 
 		vaddr = kmap(page);
+		if (partial_cacheline_write)
+			drm_clflush_virt_range(vaddr + shmem_page_offset,
+					       page_length);
 		if (page_do_bit17_swizzling)
 			ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
 							user_data,
@@ -638,6 +670,9 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
 			ret = __copy_from_user(vaddr + shmem_page_offset,
 					       user_data,
 					       page_length);
+		if (needs_clflush_after)
+			drm_clflush_virt_range(vaddr + shmem_page_offset,
+					       page_length);
 		kunmap(page);
 
 		mutex_lock(&dev->struct_mutex);
@@ -671,6 +706,9 @@ out:
 		}
 	}
 
+	if (needs_clflush_after)
+		intel_gtt_chipset_flush();
+
 	return ret;
 }