drm/i915: Implement pwrite without struct-mutex
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 28 Oct 2016 12:58:40 +0000 (13:58 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 28 Oct 2016 19:53:48 +0000 (20:53 +0100)
We only need struct_mutex within pwrite for a brief window where we need
to serialise with rendering and control our cache domains. Elsewhere we
can rely on the backing storage being pinned, and forgive userspace any
races against us.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161028125858.23563-17-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem.c

index d122658..fad1487 100644 (file)
@@ -1116,72 +1116,50 @@ out:
  * page faults in the source data
  */
 
-static inline int
-fast_user_write(struct io_mapping *mapping,
-               loff_t page_base, int page_offset,
-               char __user *user_data,
-               int length)
+static inline bool
+ggtt_write(struct io_mapping *mapping,
+          loff_t base, int offset,
+          char __user *user_data, int length)
 {
-       void __iomem *vaddr_atomic;
        void *vaddr;
        unsigned long unwritten;
 
-       vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
        /* We can use the cpu mem copy function because this is X86. */
-       vaddr = (void __force*)vaddr_atomic + page_offset;
-       unwritten = __copy_from_user_inatomic_nocache(vaddr,
+       vaddr = (void __force *)io_mapping_map_atomic_wc(mapping, base);
+       unwritten = __copy_from_user_inatomic_nocache(vaddr + offset,
                                                      user_data, length);
-       io_mapping_unmap_atomic(vaddr_atomic);
-       return unwritten;
-}
-
-static inline unsigned long
-slow_user_access(struct io_mapping *mapping,
-                unsigned long page_base, int page_offset,
-                char __user *user_data,
-                unsigned long length, bool pwrite)
-{
-       void __iomem *ioaddr;
-       void *vaddr;
-       unsigned long unwritten;
-
-       ioaddr = io_mapping_map_wc(mapping, page_base, PAGE_SIZE);
-       /* We can use the cpu mem copy function because this is X86. */
-       vaddr = (void __force *)ioaddr + page_offset;
-       if (pwrite)
-               unwritten = __copy_from_user(vaddr, user_data, length);
-       else
-               unwritten = __copy_to_user(user_data, vaddr, length);
+       io_mapping_unmap_atomic(vaddr);
+       if (unwritten) {
+               vaddr = (void __force *)
+                       io_mapping_map_wc(mapping, base, PAGE_SIZE);
+               unwritten = copy_from_user(vaddr + offset, user_data, length);
+               io_mapping_unmap(vaddr);
+       }
 
-       io_mapping_unmap(ioaddr);
        return unwritten;
 }
 
 /**
  * This is the fast pwrite path, where we copy the data directly from the
  * user into the GTT, uncached.
- * @i915: i915 device private data
- * @obj: i915 gem object
+ * @obj: i915 GEM object
  * @args: pwrite arguments structure
- * @file: drm file pointer
  */
 static int
-i915_gem_gtt_pwrite_fast(struct drm_i915_private *i915,
-                        struct drm_i915_gem_object *obj,
-                        struct drm_i915_gem_pwrite *args,
-                        struct drm_file *file)
+i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
+                        const struct drm_i915_gem_pwrite *args)
 {
+       struct drm_i915_private *i915 = to_i915(obj->base.dev);
        struct i915_ggtt *ggtt = &i915->ggtt;
-       struct drm_device *dev = obj->base.dev;
-       struct i915_vma *vma;
        struct drm_mm_node node;
-       uint64_t remain, offset;
-       char __user *user_data;
+       struct i915_vma *vma;
+       u64 remain, offset;
+       void __user *user_data;
        int ret;
-       bool hit_slow_path = false;
 
-       if (i915_gem_object_is_tiled(obj))
-               return -EFAULT;
+       ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
+       if (ret)
+               return ret;
 
        intel_runtime_pm_get(i915);
        vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
@@ -1198,21 +1176,17 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_private *i915,
        if (IS_ERR(vma)) {
                ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
                if (ret)
-                       goto out;
-
-               ret = i915_gem_object_pin_pages(obj);
-               if (ret) {
-                       remove_mappable_node(&node);
-                       goto out;
-               }
+                       goto out_unlock;
+               GEM_BUG_ON(!node.allocated);
        }
 
        ret = i915_gem_object_set_to_gtt_domain(obj, true);
        if (ret)
                goto out_unpin;
 
+       mutex_unlock(&i915->drm.struct_mutex);
+
        intel_fb_obj_invalidate(obj, ORIGIN_CPU);
-       obj->mm.dirty = true;
 
        user_data = u64_to_user_ptr(args->data_ptr);
        offset = args->offset;
@@ -1243,92 +1217,36 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_private *i915,
                 * If the object is non-shmem backed, we retry again with the
                 * path that handles page fault.
                 */
-               if (fast_user_write(&ggtt->mappable, page_base,
-                                   page_offset, user_data, page_length)) {
-                       hit_slow_path = true;
-                       mutex_unlock(&dev->struct_mutex);
-                       if (slow_user_access(&ggtt->mappable,
-                                            page_base,
-                                            page_offset, user_data,
-                                            page_length, true)) {
-                               ret = -EFAULT;
-                               mutex_lock(&dev->struct_mutex);
-                               goto out_flush;
-                       }
-
-                       mutex_lock(&dev->struct_mutex);
+               if (ggtt_write(&ggtt->mappable, page_base, page_offset,
+                              user_data, page_length)) {
+                       ret = -EFAULT;
+                       break;
                }
 
                remain -= page_length;
                user_data += page_length;
                offset += page_length;
        }
-
-out_flush:
-       if (hit_slow_path) {
-               if (ret == 0 &&
-                   (obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) {
-                       /* The user has modified the object whilst we tried
-                        * reading from it, and we now have no idea what domain
-                        * the pages should be in. As we have just been touching
-                        * them directly, flush everything back to the GTT
-                        * domain.
-                        */
-                       ret = i915_gem_object_set_to_gtt_domain(obj, false);
-               }
-       }
-
        intel_fb_obj_flush(obj, false, ORIGIN_CPU);
+
+       mutex_lock(&i915->drm.struct_mutex);
 out_unpin:
        if (node.allocated) {
                wmb();
                ggtt->base.clear_range(&ggtt->base,
                                       node.start, node.size);
-               i915_gem_object_unpin_pages(obj);
                remove_mappable_node(&node);
        } else {
                i915_vma_unpin(vma);
        }
-out:
+out_unlock:
        intel_runtime_pm_put(i915);
+       mutex_unlock(&i915->drm.struct_mutex);
        return ret;
 }
 
-/* Per-page copy function for the shmem pwrite fastpath.
- * Flushes invalid cachelines before writing to the target if
- * needs_clflush_before is set and flushes out any written cachelines after
- * writing if needs_clflush is set. */
 static int
-shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
-                 char __user *user_data,
-                 bool page_do_bit17_swizzling,
-                 bool needs_clflush_before,
-                 bool needs_clflush_after)
-{
-       char *vaddr;
-       int ret;
-
-       if (unlikely(page_do_bit17_swizzling))
-               return -EINVAL;
-
-       vaddr = kmap_atomic(page);
-       if (needs_clflush_before)
-               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);
-
-       return ret ? -EFAULT : 0;
-}
-
-/* Only difference to the fast-path function is that this can handle bit17
- * and uses non-atomic copy and kmap functions. */
-static int
-shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
+shmem_pwrite_slow(struct page *page, int offset, int length,
                  char __user *user_data,
                  bool page_do_bit17_swizzling,
                  bool needs_clflush_before,
@@ -1339,124 +1257,114 @@ shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
 
        vaddr = kmap(page);
        if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
-               shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
-                                            page_length,
+               shmem_clflush_swizzled_range(vaddr + offset, length,
                                             page_do_bit17_swizzling);
        if (page_do_bit17_swizzling)
-               ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
-                                               user_data,
-                                               page_length);
+               ret = __copy_from_user_swizzled(vaddr, offset, user_data,
+                                               length);
        else
-               ret = __copy_from_user(vaddr + shmem_page_offset,
-                                      user_data,
-                                      page_length);
+               ret = __copy_from_user(vaddr + offset, user_data, length);
        if (needs_clflush_after)
-               shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
-                                            page_length,
+               shmem_clflush_swizzled_range(vaddr + offset, length,
                                             page_do_bit17_swizzling);
        kunmap(page);
 
        return ret ? -EFAULT : 0;
 }
 
+/* Per-page copy function for the shmem pwrite fastpath.
+ * Flushes invalid cachelines before writing to the target if
+ * needs_clflush_before is set and flushes out any written cachelines after
+ * writing if needs_clflush is set.
+ */
 static int
-i915_gem_shmem_pwrite(struct drm_device *dev,
-                     struct drm_i915_gem_object *obj,
-                     struct drm_i915_gem_pwrite *args,
-                     struct drm_file *file)
+shmem_pwrite(struct page *page, int offset, int len, char __user *user_data,
+            bool page_do_bit17_swizzling,
+            bool needs_clflush_before,
+            bool needs_clflush_after)
 {
-       ssize_t remain;
-       loff_t offset;
-       char __user *user_data;
-       int shmem_page_offset, page_length, ret = 0;
-       int obj_do_bit17_swizzling, page_do_bit17_swizzling;
-       int hit_slowpath = 0;
+       int ret;
+
+       ret = -ENODEV;
+       if (!page_do_bit17_swizzling) {
+               char *vaddr = kmap_atomic(page);
+
+               if (needs_clflush_before)
+                       drm_clflush_virt_range(vaddr + offset, len);
+               ret = __copy_from_user_inatomic(vaddr + offset, user_data, len);
+               if (needs_clflush_after)
+                       drm_clflush_virt_range(vaddr + offset, len);
+
+               kunmap_atomic(vaddr);
+       }
+       if (ret == 0)
+               return ret;
+
+       return shmem_pwrite_slow(page, offset, len, user_data,
+                                page_do_bit17_swizzling,
+                                needs_clflush_before,
+                                needs_clflush_after);
+}
+
+static int
+i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj,
+                     const struct drm_i915_gem_pwrite *args)
+{
+       struct drm_i915_private *i915 = to_i915(obj->base.dev);
+       void __user *user_data;
+       u64 remain;
+       unsigned int obj_do_bit17_swizzling;
+       unsigned int partial_cacheline_write;
        unsigned int needs_clflush;
-       struct sg_page_iter sg_iter;
+       unsigned int offset, idx;
+       int ret;
 
-       ret = i915_gem_obj_prepare_shmem_write(obj, &needs_clflush);
+       ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
        if (ret)
                return ret;
 
-       obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
-       user_data = u64_to_user_ptr(args->data_ptr);
-       offset = args->offset;
-       remain = args->size;
+       ret = i915_gem_obj_prepare_shmem_write(obj, &needs_clflush);
+       mutex_unlock(&i915->drm.struct_mutex);
+       if (ret)
+               return ret;
 
-       for_each_sg_page(obj->mm.pages->sgl, &sg_iter, obj->mm.pages->nents,
-                        offset >> PAGE_SHIFT) {
-               struct page *page = sg_page_iter_page(&sg_iter);
-               int partial_cacheline_write;
+       obj_do_bit17_swizzling = 0;
+       if (i915_gem_object_needs_bit17_swizzle(obj))
+               obj_do_bit17_swizzling = BIT(17);
 
-               if (remain <= 0)
-                       break;
+       /* 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 = 0;
+       if (needs_clflush & CLFLUSH_BEFORE)
+               partial_cacheline_write = boot_cpu_data.x86_clflush_size - 1;
 
-               /* Operation in this page
-                *
-                * shmem_page_offset = offset within page in shmem file
-                * page_length = bytes to copy for this page
-                */
-               shmem_page_offset = offset_in_page(offset);
-
-               page_length = remain;
-               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 & CLFLUSH_BEFORE &&
-                       ((shmem_page_offset | page_length)
-                               & (boot_cpu_data.x86_clflush_size - 1));
-
-               page_do_bit17_swizzling = obj_do_bit17_swizzling &&
-                       (page_to_phys(page) & (1 << 17)) != 0;
-
-               ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
-                                       user_data, page_do_bit17_swizzling,
-                                       partial_cacheline_write,
-                                       needs_clflush & CLFLUSH_AFTER);
-               if (ret == 0)
-                       goto next_page;
-
-               hit_slowpath = 1;
-               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 & CLFLUSH_AFTER);
+       user_data = u64_to_user_ptr(args->data_ptr);
+       remain = args->size;
+       offset = offset_in_page(args->offset);
+       for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
+               struct page *page = i915_gem_object_get_page(obj, idx);
+               int length;
 
-               mutex_lock(&dev->struct_mutex);
+               length = remain;
+               if (offset + length > PAGE_SIZE)
+                       length = PAGE_SIZE - offset;
 
+               ret = shmem_pwrite(page, offset, length, user_data,
+                                  page_to_phys(page) & obj_do_bit17_swizzling,
+                                  (offset | length) & partial_cacheline_write,
+                                  needs_clflush & CLFLUSH_AFTER);
                if (ret)
-                       goto out;
-
-next_page:
-               remain -= page_length;
-               user_data += page_length;
-               offset += page_length;
-       }
-
-out:
-       i915_gem_obj_finish_shmem_access(obj);
+                       break;
 
-       if (hit_slowpath) {
-               /*
-                * Fixup: Flush cpu caches in case we didn't flush the dirty
-                * cachelines in-line while writing and the object moved
-                * out of the cpu write domain while we've dropped the lock.
-                */
-               if (!(needs_clflush & CLFLUSH_AFTER) &&
-                   obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
-                       if (i915_gem_clflush_object(obj, obj->pin_display))
-                               needs_clflush |= CLFLUSH_AFTER;
-               }
+               remain -= length;
+               user_data += length;
+               offset = 0;
        }
 
-       if (needs_clflush & CLFLUSH_AFTER)
-               i915_gem_chipset_flush(to_i915(dev));
-
        intel_fb_obj_flush(obj, false, ORIGIN_CPU);
+       i915_gem_obj_finish_shmem_access(obj);
        return ret;
 }
 
@@ -1472,7 +1380,6 @@ int
 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
                      struct drm_file *file)
 {
-       struct drm_i915_private *dev_priv = to_i915(dev);
        struct drm_i915_gem_pwrite *args = data;
        struct drm_i915_gem_object *obj;
        int ret;
@@ -1485,13 +1392,6 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
                       args->size))
                return -EFAULT;
 
-       if (likely(!i915.prefault_disable)) {
-               ret = fault_in_pages_readable(u64_to_user_ptr(args->data_ptr),
-                                                  args->size);
-               if (ret)
-                       return -EFAULT;
-       }
-
        obj = i915_gem_object_lookup(file, args->handle);
        if (!obj)
                return -ENOENT;
@@ -1513,11 +1413,9 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
        if (ret)
                goto err;
 
-       intel_runtime_pm_get(dev_priv);
-
-       ret = i915_mutex_lock_interruptible(dev);
+       ret = i915_gem_object_pin_pages(obj);
        if (ret)
-               goto err_rpm;
+               goto err;
 
        ret = -EFAULT;
        /* We can only do the GTT pwrite on untiled buffers, as otherwise
@@ -1532,23 +1430,16 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
                 * pointers (e.g. gtt mappings when moving data between
                 * textures). Fallback to the shmem path in that case.
                 */
-               ret = i915_gem_gtt_pwrite_fast(dev_priv, obj, args, file);
+               ret = i915_gem_gtt_pwrite_fast(obj, args);
 
        if (ret == -EFAULT || ret == -ENOSPC) {
                if (obj->phys_handle)
                        ret = i915_gem_phys_pwrite(obj, args, file);
                else
-                       ret = i915_gem_shmem_pwrite(dev, obj, args, file);
+                       ret = i915_gem_shmem_pwrite(obj, args);
        }
 
-       i915_gem_object_put(obj);
-       mutex_unlock(&dev->struct_mutex);
-       intel_runtime_pm_put(dev_priv);
-
-       return ret;
-
-err_rpm:
-       intel_runtime_pm_put(dev_priv);
+       i915_gem_object_unpin_pages(obj);
 err:
        i915_gem_object_put_unlocked(obj);
        return ret;