drm/i915: Pin backing pages for pwrite
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 4 Sep 2012 20:02:55 +0000 (21:02 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Thu, 20 Sep 2012 12:22:56 +0000 (14:22 +0200)
By using the recently introduced pinning of pages, we can safely drop
the mutex in the knowledge that the pages are not going to disappear
jeneath us, and so we can simplify the code for iterating over the pages.

Note: The old code had such complicated page refcounting since it used
obj->pages as a micro-optimization if it's there, but that could
(before this patch) disappear when we drop the dev->struct_mutex.
Hence some manual page refcounting was required for the slow path,
complicated by the fact that pages returned by shmem_read_mapping_page
already have a pageref, which needs to be dropped again.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Added note to explain the question Ben raised in review.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_gem.c

index ea3a5e1791e9f3dcd17deddf0333a3f4757a2ff6..73702e583eb986c61ab2fe1dee4ca63d8584e7df 100644 (file)
@@ -690,7 +690,7 @@ shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
                                       page_length);
        kunmap_atomic(vaddr);
 
-       return ret;
+       return ret ? -EFAULT : 0;
 }
 
 /* Only difference to the fast-path function is that this can handle bit17
@@ -724,7 +724,7 @@ shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
                                             page_do_bit17_swizzling);
        kunmap(page);
 
-       return ret;
+       return ret ? -EFAULT : 0;
 }
 
 static int
@@ -733,7 +733,6 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
                      struct drm_i915_gem_pwrite *args,
                      struct drm_file *file)
 {
-       struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
        ssize_t remain;
        loff_t offset;
        char __user *user_data;
@@ -742,7 +741,6 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
        int hit_slowpath = 0;
        int needs_clflush_after = 0;
        int needs_clflush_before = 0;
-       int release_page;
 
        user_data = (char __user *) (uintptr_t) args->data_ptr;
        remain = args->size;
@@ -768,6 +766,12 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
            && obj->cache_level == I915_CACHE_NONE)
                needs_clflush_before = 1;
 
+       ret = i915_gem_object_get_pages(obj);
+       if (ret)
+               return ret;
+
+       i915_gem_object_pin_pages(obj);
+
        offset = args->offset;
        obj->dirty = 1;
 
@@ -793,18 +797,7 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
                        ((shmem_page_offset | page_length)
                                & (boot_cpu_data.x86_clflush_size - 1));
 
-               if (obj->pages) {
-                       page = obj->pages[offset >> PAGE_SHIFT];
-                       release_page = 0;
-               } else {
-                       page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
-                       if (IS_ERR(page)) {
-                               ret = PTR_ERR(page);
-                               goto out;
-                       }
-                       release_page = 1;
-               }
-
+               page = obj->pages[offset >> PAGE_SHIFT];
                page_do_bit17_swizzling = obj_do_bit17_swizzling &&
                        (page_to_phys(page) & (1 << 17)) != 0;
 
@@ -816,26 +809,20 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
                        goto next_page;
 
                hit_slowpath = 1;
-               page_cache_get(page);
                mutex_unlock(&dev->struct_mutex);
-
                ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
                                        user_data, page_do_bit17_swizzling,
                                        partial_cacheline_write,
                                        needs_clflush_after);
 
                mutex_lock(&dev->struct_mutex);
-               page_cache_release(page);
+
 next_page:
                set_page_dirty(page);
                mark_page_accessed(page);
-               if (release_page)
-                       page_cache_release(page);
 
-               if (ret) {
-                       ret = -EFAULT;
+               if (ret)
                        goto out;
-               }
 
                remain -= page_length;
                user_data += page_length;
@@ -843,6 +830,8 @@ next_page:
        }
 
 out:
+       i915_gem_object_unpin_pages(obj);
+
        if (hit_slowpath) {
                /* Fixup: Kill any reinstated backing storage pages */
                if (obj->madv == __I915_MADV_PURGED)