drm/i915: Use a radixtree for random access to the object's backing storage
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 28 Oct 2016 12:58:33 +0000 (13:58 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 28 Oct 2016 19:53:45 +0000 (20:53 +0100)
A while ago we switched from a contiguous array of pages into an sglist,
for that was both more convenient for mapping to hardware and avoided
the requirement for a vmalloc array of pages on every object. However,
certain GEM API calls (like pwrite, pread as well as performing
relocations) do desire access to individual struct pages. A quick hack
was to introduce a cache of the last access such that finding the
following page was quick - this works so long as the caller desired
sequential access. Walking backwards, or multiple callers, still hits a
slow linear search for each page. One solution is to store each
successful lookup in a radix tree.

v2: Rewrite building the radixtree for clarity, hopefully.

v3: Rearrange execbuf to avoid calling i915_gem_object_get_sg() from
within an atomic section and so relax the allocation context to a simple
GFP_KERNEL and mutex.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161028125858.23563-10-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_gem_stolen.c
drivers/gpu/drm/i915/i915_gem_userptr.c

index a1b9801296078b7d4d888aae2b99da328ceec8ee..85ae83a0c9374de82926d9281c2d4cbe96a2294a 100644 (file)
@@ -2286,9 +2286,12 @@ struct drm_i915_gem_object {
 
        struct sg_table *pages;
        int pages_pin_count;
-       struct get_page {
-               struct scatterlist *sg;
-               int last;
+       struct i915_gem_object_page_iter {
+               struct scatterlist *sg_pos;
+               unsigned int sg_idx; /* in pages, but 32bit eek! */
+
+               struct radix_tree_root radix;
+               struct mutex lock; /* protects this cache */
        } get_page;
        void *mapping;
 
@@ -2491,6 +2494,14 @@ static __always_inline struct sgt_iter {
        return s;
 }
 
+static inline struct scatterlist *____sg_next(struct scatterlist *sg)
+{
+       ++sg;
+       if (unlikely(sg_is_chain(sg)))
+               sg = sg_chain_ptr(sg);
+       return sg;
+}
+
 /**
  * __sg_next - return the next scatterlist entry in a list
  * @sg:                The current sg entry
@@ -2505,9 +2516,7 @@ static inline struct scatterlist *__sg_next(struct scatterlist *sg)
 #ifdef CONFIG_DEBUG_SG
        BUG_ON(sg->sg_magic != SG_MAGIC);
 #endif
-       return sg_is_last(sg) ? NULL :
-               likely(!sg_is_chain(++sg)) ? sg :
-               sg_chain_ptr(sg);
+       return sg_is_last(sg) ? NULL : ____sg_next(sg);
 }
 
 /**
@@ -3185,45 +3194,21 @@ static inline int __sg_page_count(struct scatterlist *sg)
        return sg->length >> PAGE_SHIFT;
 }
 
-struct page *
-i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj, int n);
-
-static inline dma_addr_t
-i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj, int n)
-{
-       if (n < obj->get_page.last) {
-               obj->get_page.sg = obj->pages->sgl;
-               obj->get_page.last = 0;
-       }
-
-       while (obj->get_page.last + __sg_page_count(obj->get_page.sg) <= n) {
-               obj->get_page.last += __sg_page_count(obj->get_page.sg++);
-               if (unlikely(sg_is_chain(obj->get_page.sg)))
-                       obj->get_page.sg = sg_chain_ptr(obj->get_page.sg);
-       }
+struct scatterlist *
+i915_gem_object_get_sg(struct drm_i915_gem_object *obj,
+                      unsigned int n, unsigned int *offset);
 
-       return sg_dma_address(obj->get_page.sg) + ((n - obj->get_page.last) << PAGE_SHIFT);
-}
-
-static inline struct page *
-i915_gem_object_get_page(struct drm_i915_gem_object *obj, int n)
-{
-       if (WARN_ON(n >= obj->base.size >> PAGE_SHIFT))
-               return NULL;
-
-       if (n < obj->get_page.last) {
-               obj->get_page.sg = obj->pages->sgl;
-               obj->get_page.last = 0;
-       }
+struct page *
+i915_gem_object_get_page(struct drm_i915_gem_object *obj,
+                        unsigned int n);
 
-       while (obj->get_page.last + __sg_page_count(obj->get_page.sg) <= n) {
-               obj->get_page.last += __sg_page_count(obj->get_page.sg++);
-               if (unlikely(sg_is_chain(obj->get_page.sg)))
-                       obj->get_page.sg = sg_chain_ptr(obj->get_page.sg);
-       }
+struct page *
+i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj,
+                              unsigned int n);
 
-       return nth_page(sg_page(obj->get_page.sg), n - obj->get_page.last);
-}
+dma_addr_t
+i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj,
+                               unsigned long n);
 
 static inline void i915_gem_object_pin_pages(struct drm_i915_gem_object *obj)
 {
index 528958d8fa5ac06b16327c7ef4617a0e1f0455fe..aa0de3aa256557306d40c56b86023cbb4153b9ce 100644 (file)
@@ -2330,6 +2330,15 @@ i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
        kfree(obj->pages);
 }
 
+static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj)
+{
+       struct radix_tree_iter iter;
+       void **slot;
+
+       radix_tree_for_each_slot(slot, &obj->get_page.radix, &iter, 0)
+               radix_tree_delete(&obj->get_page.radix, iter.index);
+}
+
 int
 i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
 {
@@ -2362,6 +2371,8 @@ i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
                obj->mapping = NULL;
        }
 
+       __i915_gem_object_reset_page_iter(obj);
+
        ops->put_pages(obj);
        obj->pages = NULL;
 
@@ -2531,8 +2542,8 @@ i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
 
        list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list);
 
-       obj->get_page.sg = obj->pages->sgl;
-       obj->get_page.last = 0;
+       obj->get_page.sg_pos = obj->pages->sgl;
+       obj->get_page.sg_idx = 0;
 
        return 0;
 }
@@ -4337,6 +4348,8 @@ void i915_gem_object_init(struct drm_i915_gem_object *obj,
 
        obj->frontbuffer_ggtt_origin = ORIGIN_GTT;
        obj->madv = I915_MADV_WILLNEED;
+       INIT_RADIX_TREE(&obj->get_page.radix, GFP_KERNEL | __GFP_NOWARN);
+       mutex_init(&obj->get_page.lock);
 
        i915_gem_info_add_obj(to_i915(obj->base.dev), obj->base.size);
 }
@@ -5032,21 +5045,6 @@ void i915_gem_track_fb(struct drm_i915_gem_object *old,
        }
 }
 
-/* Like i915_gem_object_get_page(), but mark the returned page dirty */
-struct page *
-i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj, int n)
-{
-       struct page *page;
-
-       /* Only default objects have per-page dirty tracking */
-       if (WARN_ON(!i915_gem_object_has_struct_page(obj)))
-               return NULL;
-
-       page = i915_gem_object_get_page(obj, n);
-       set_page_dirty(page);
-       return page;
-}
-
 /* Allocate a new GEM object and fill it with the supplied data */
 struct drm_i915_gem_object *
 i915_gem_object_create_from_data(struct drm_device *dev,
@@ -5087,3 +5085,156 @@ fail:
        i915_gem_object_put(obj);
        return ERR_PTR(ret);
 }
+
+struct scatterlist *
+i915_gem_object_get_sg(struct drm_i915_gem_object *obj,
+                      unsigned int n,
+                      unsigned int *offset)
+{
+       struct i915_gem_object_page_iter *iter = &obj->get_page;
+       struct scatterlist *sg;
+       unsigned int idx, count;
+
+       might_sleep();
+       GEM_BUG_ON(n >= obj->base.size >> PAGE_SHIFT);
+       GEM_BUG_ON(obj->pages_pin_count == 0);
+
+       /* As we iterate forward through the sg, we record each entry in a
+        * radixtree for quick repeated (backwards) lookups. If we have seen
+        * this index previously, we will have an entry for it.
+        *
+        * Initial lookup is O(N), but this is amortized to O(1) for
+        * sequential page access (where each new request is consecutive
+        * to the previous one). Repeated lookups are O(lg(obj->base.size)),
+        * i.e. O(1) with a large constant!
+        */
+       if (n < READ_ONCE(iter->sg_idx))
+               goto lookup;
+
+       mutex_lock(&iter->lock);
+
+       /* We prefer to reuse the last sg so that repeated lookup of this
+        * (or the subsequent) sg are fast - comparing against the last
+        * sg is faster than going through the radixtree.
+        */
+
+       sg = iter->sg_pos;
+       idx = iter->sg_idx;
+       count = __sg_page_count(sg);
+
+       while (idx + count <= n) {
+               unsigned long exception, i;
+               int ret;
+
+               /* If we cannot allocate and insert this entry, or the
+                * individual pages from this range, cancel updating the
+                * sg_idx so that on this lookup we are forced to linearly
+                * scan onwards, but on future lookups we will try the
+                * insertion again (in which case we need to be careful of
+                * the error return reporting that we have already inserted
+                * this index).
+                */
+               ret = radix_tree_insert(&iter->radix, idx, sg);
+               if (ret && ret != -EEXIST)
+                       goto scan;
+
+               exception =
+                       RADIX_TREE_EXCEPTIONAL_ENTRY |
+                       idx << RADIX_TREE_EXCEPTIONAL_SHIFT;
+               for (i = 1; i < count; i++) {
+                       ret = radix_tree_insert(&iter->radix, idx + i,
+                                               (void *)exception);
+                       if (ret && ret != -EEXIST)
+                               goto scan;
+               }
+
+               idx += count;
+               sg = ____sg_next(sg);
+               count = __sg_page_count(sg);
+       }
+
+scan:
+       iter->sg_pos = sg;
+       iter->sg_idx = idx;
+
+       mutex_unlock(&iter->lock);
+
+       if (unlikely(n < idx)) /* insertion completed by another thread */
+               goto lookup;
+
+       /* In case we failed to insert the entry into the radixtree, we need
+        * to look beyond the current sg.
+        */
+       while (idx + count <= n) {
+               idx += count;
+               sg = ____sg_next(sg);
+               count = __sg_page_count(sg);
+       }
+
+       *offset = n - idx;
+       return sg;
+
+lookup:
+       rcu_read_lock();
+
+       sg = radix_tree_lookup(&iter->radix, n);
+       GEM_BUG_ON(!sg);
+
+       /* If this index is in the middle of multi-page sg entry,
+        * the radixtree will contain an exceptional entry that points
+        * to the start of that range. We will return the pointer to
+        * the base page and the offset of this page within the
+        * sg entry's range.
+        */
+       *offset = 0;
+       if (unlikely(radix_tree_exception(sg))) {
+               unsigned long base =
+                       (unsigned long)sg >> RADIX_TREE_EXCEPTIONAL_SHIFT;
+
+               sg = radix_tree_lookup(&iter->radix, base);
+               GEM_BUG_ON(!sg);
+
+               *offset = n - base;
+       }
+
+       rcu_read_unlock();
+
+       return sg;
+}
+
+struct page *
+i915_gem_object_get_page(struct drm_i915_gem_object *obj, unsigned int n)
+{
+       struct scatterlist *sg;
+       unsigned int offset;
+
+       GEM_BUG_ON(!i915_gem_object_has_struct_page(obj));
+
+       sg = i915_gem_object_get_sg(obj, n, &offset);
+       return nth_page(sg_page(sg), offset);
+}
+
+/* Like i915_gem_object_get_page(), but mark the returned page dirty */
+struct page *
+i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj,
+                              unsigned int n)
+{
+       struct page *page;
+
+       page = i915_gem_object_get_page(obj, n);
+       if (!obj->dirty)
+               set_page_dirty(page);
+
+       return page;
+}
+
+dma_addr_t
+i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj,
+                               unsigned long n)
+{
+       struct scatterlist *sg;
+       unsigned int offset;
+
+       sg = i915_gem_object_get_sg(obj, n, &offset);
+       return sg_dma_address(sg) + (offset << PAGE_SHIFT);
+}
index f4f6d3a48b050c7ee0e23bae53fab4577acc7235..70e61bc35c60d08687a08576cfa41fdd0ffcbe8a 100644 (file)
@@ -595,8 +595,8 @@ _i915_gem_object_create_stolen(struct drm_device *dev,
        if (obj->pages == NULL)
                goto cleanup;
 
-       obj->get_page.sg = obj->pages->sgl;
-       obj->get_page.last = 0;
+       obj->get_page.sg_pos = obj->pages->sgl;
+       obj->get_page.sg_idx = 0;
 
        i915_gem_object_pin_pages(obj);
        obj->stolen = stolen;
index c49dd95413bd42373863495aec6a06704e432f48..e2fa970bb629fd035c8e823b1243d363aa348ad9 100644 (file)
@@ -530,8 +530,8 @@ __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
                        if (ret == 0) {
                                list_add_tail(&obj->global_list,
                                              &to_i915(dev)->mm.unbound_list);
-                               obj->get_page.sg = obj->pages->sgl;
-                               obj->get_page.last = 0;
+                               obj->get_page.sg_pos = obj->pages->sgl;
+                               obj->get_page.sg_idx = 0;
                                pinned = 0;
                        }
                }