]> git.baikalelectronics.ru Git - kernel.git/commitdiff
drm/mm: optimize rb_hole_addr rbtree search
authorNirmoy Das <nirmoy.das@amd.com>
Mon, 4 May 2020 15:40:35 +0000 (17:40 +0200)
committerChristian König <christian.koenig@amd.com>
Tue, 5 May 2020 11:39:38 +0000 (13:39 +0200)
Userspace can severely fragment rb_hole_addr rbtree by manipulating
alignment while allocating buffers. Fragmented rb_hole_addr rbtree
would result in large delays while allocating buffer object for a
userspace application. It takes long time to find suitable hole
because if we fail to find a suitable hole in the first attempt
then we look for neighbouring nodes using rb_prev()/rb_next().
Traversing rbtree using rb_prev()/rb_next() can take really long
time if the tree is fragmented.

This patch improves searches in fragmented rb_hole_addr rbtree by
modifying it to an augmented rbtree which will store an extra field
in drm_mm_node, subtree_max_hole. Each drm_mm_node now stores maximum
hole size for its subtree in drm_mm_node->subtree_max_hole. Using
drm_mm_node->subtree_max_hole, it is possible to eliminate a complete
subtree if that subtree is unable to serve a request hence reducing
number of rb_prev()/rb_next() used.

With this patch applied, 1 million bo allocs on amdgpu took ~8 sec,
compared to 50k bo allocs which took 28 sec without it.

partial test code:
int test_fragmentation(void)
{

int i = 0;
        uint32_t  minor_version;
        uint32_t  major_version;

        struct amdgpu_bo_alloc_request request = {};
        amdgpu_bo_handle vram_handle[MAX_ALLOC] = {};
        amdgpu_device_handle device_handle;

        request.alloc_size = 4096;
        request.phys_alignment = 8192;
        request.preferred_heap = AMDGPU_GEM_DOMAIN_VRAM;

        int fd = open("/dev/dri/card0", O_RDWR | O_CLOEXEC);
        amdgpu_device_initialize(fd, &major_version,  &minor_version,
 &device_handle);

        for (i = 0; i < MAX_ALLOC; i++) {
                amdgpu_bo_alloc(device_handle, &request, &vram_handle[i]);
        }

        for (i = 0; i < MAX_ALLOC; i++)
                amdgpu_bo_free(vram_handle[i]);

        return 0;
}

v2:
Use RB_DECLARE_CALLBACKS_MAX to maintain subtree_max_hole
v3:
insert_hole_addr() should be static a function
fix return value of next_hole_high_addr()/next_hole_low_addr()
Reported-by: kbuild test robot <lkp@intel.com>
v4:
Fix commit message.

Signed-off-by: Nirmoy Das <nirmoy.das@amd.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/364341/
Signed-off-by: Christian König <christian.koenig@amd.com>
drivers/gpu/drm/drm_mm.c
include/drm/drm_mm.h

index 8981abe8b7c940e82fe33fa1085994035dd0daf4..f4ca1ff80af9f23ff2ada4dd545917014aa32e74 100644 (file)
@@ -212,20 +212,6 @@ static void drm_mm_interval_tree_add_node(struct drm_mm_node *hole_node,
                                   &drm_mm_interval_tree_augment);
 }
 
-#define RB_INSERT(root, member, expr) do { \
-       struct rb_node **link = &root.rb_node, *rb = NULL; \
-       u64 x = expr(node); \
-       while (*link) { \
-               rb = *link; \
-               if (x < expr(rb_entry(rb, struct drm_mm_node, member))) \
-                       link = &rb->rb_left; \
-               else \
-                       link = &rb->rb_right; \
-       } \
-       rb_link_node(&node->member, rb, link); \
-       rb_insert_color(&node->member, &root); \
-} while (0)
-
 #define HOLE_SIZE(NODE) ((NODE)->hole_size)
 #define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE))
 
@@ -255,16 +241,42 @@ static void insert_hole_size(struct rb_root_cached *root,
        rb_insert_color_cached(&node->rb_hole_size, root, first);
 }
 
+RB_DECLARE_CALLBACKS_MAX(static, augment_callbacks,
+                        struct drm_mm_node, rb_hole_addr,
+                        u64, subtree_max_hole, HOLE_SIZE)
+
+static void insert_hole_addr(struct rb_root *root, struct drm_mm_node *node)
+{
+       struct rb_node **link = &root->rb_node, *rb_parent = NULL;
+       u64 start = HOLE_ADDR(node), subtree_max_hole = node->subtree_max_hole;
+       struct drm_mm_node *parent;
+
+       while (*link) {
+               rb_parent = *link;
+               parent = rb_entry(rb_parent, struct drm_mm_node, rb_hole_addr);
+               if (parent->subtree_max_hole < subtree_max_hole)
+                       parent->subtree_max_hole = subtree_max_hole;
+               if (start < HOLE_ADDR(parent))
+                       link = &parent->rb_hole_addr.rb_left;
+               else
+                       link = &parent->rb_hole_addr.rb_right;
+       }
+
+       rb_link_node(&node->rb_hole_addr, rb_parent, link);
+       rb_insert_augmented(&node->rb_hole_addr, root, &augment_callbacks);
+}
+
 static void add_hole(struct drm_mm_node *node)
 {
        struct drm_mm *mm = node->mm;
 
        node->hole_size =
                __drm_mm_hole_node_end(node) - __drm_mm_hole_node_start(node);
+       node->subtree_max_hole = node->hole_size;
        DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
 
        insert_hole_size(&mm->holes_size, node);
-       RB_INSERT(mm->holes_addr, rb_hole_addr, HOLE_ADDR);
+       insert_hole_addr(&mm->holes_addr, node);
 
        list_add(&node->hole_stack, &mm->hole_stack);
 }
@@ -275,8 +287,10 @@ static void rm_hole(struct drm_mm_node *node)
 
        list_del(&node->hole_stack);
        rb_erase_cached(&node->rb_hole_size, &node->mm->holes_size);
-       rb_erase(&node->rb_hole_addr, &node->mm->holes_addr);
+       rb_erase_augmented(&node->rb_hole_addr, &node->mm->holes_addr,
+                          &augment_callbacks);
        node->hole_size = 0;
+       node->subtree_max_hole = 0;
 
        DRM_MM_BUG_ON(drm_mm_hole_follows(node));
 }
@@ -361,9 +375,90 @@ first_hole(struct drm_mm *mm,
        }
 }
 
+/**
+ * next_hole_high_addr - returns next hole for a DRM_MM_INSERT_HIGH mode request
+ * @entry: previously selected drm_mm_node
+ * @size: size of the a hole needed for the request
+ *
+ * This function will verify whether left subtree of @entry has hole big enough
+ * to fit the requtested size. If so, it will return previous node of @entry or
+ * else it will return parent node of @entry
+ *
+ * It will also skip the complete left subtree if subtree_max_hole of that
+ * subtree is same as the subtree_max_hole of the @entry.
+ *
+ * Returns:
+ * previous node of @entry if left subtree of @entry can serve the request or
+ * else return parent of @entry
+ */
+static struct drm_mm_node *
+next_hole_high_addr(struct drm_mm_node *entry, u64 size)
+{
+       struct rb_node *rb_node, *left_rb_node, *parent_rb_node;
+       struct drm_mm_node *left_node;
+
+       if (!entry)
+               return NULL;
+
+       rb_node = &entry->rb_hole_addr;
+       if (rb_node->rb_left) {
+               left_rb_node = rb_node->rb_left;
+               parent_rb_node = rb_parent(rb_node);
+               left_node = rb_entry(left_rb_node,
+                                    struct drm_mm_node, rb_hole_addr);
+               if ((left_node->subtree_max_hole < size ||
+                    entry->size == entry->subtree_max_hole) &&
+                   parent_rb_node && parent_rb_node->rb_left != rb_node)
+                       return rb_hole_addr_to_node(parent_rb_node);
+       }
+
+       return rb_hole_addr_to_node(rb_prev(rb_node));
+}
+
+/**
+ * next_hole_low_addr - returns next hole for a DRM_MM_INSERT_LOW mode request
+ * @entry: previously selected drm_mm_node
+ * @size: size of the a hole needed for the request
+ *
+ * This function will verify whether right subtree of @entry has hole big enough
+ * to fit the requtested size. If so, it will return next node of @entry or
+ * else it will return parent node of @entry
+ *
+ * It will also skip the complete right subtree if subtree_max_hole of that
+ * subtree is same as the subtree_max_hole of the @entry.
+ *
+ * Returns:
+ * next node of @entry if right subtree of @entry can serve the request or
+ * else return parent of @entry
+ */
+static struct drm_mm_node *
+next_hole_low_addr(struct drm_mm_node *entry, u64 size)
+{
+       struct rb_node *rb_node, *right_rb_node, *parent_rb_node;
+       struct drm_mm_node *right_node;
+
+       if (!entry)
+               return NULL;
+
+       rb_node = &entry->rb_hole_addr;
+       if (rb_node->rb_right) {
+               right_rb_node = rb_node->rb_right;
+               parent_rb_node = rb_parent(rb_node);
+               right_node = rb_entry(right_rb_node,
+                                     struct drm_mm_node, rb_hole_addr);
+               if ((right_node->subtree_max_hole < size ||
+                    entry->size == entry->subtree_max_hole) &&
+                   parent_rb_node && parent_rb_node->rb_right != rb_node)
+                       return rb_hole_addr_to_node(parent_rb_node);
+       }
+
+       return rb_hole_addr_to_node(rb_next(rb_node));
+}
+
 static struct drm_mm_node *
 next_hole(struct drm_mm *mm,
          struct drm_mm_node *node,
+         u64 size,
          enum drm_mm_insert_mode mode)
 {
        switch (mode) {
@@ -372,10 +467,10 @@ next_hole(struct drm_mm *mm,
                return rb_hole_size_to_node(rb_prev(&node->rb_hole_size));
 
        case DRM_MM_INSERT_LOW:
-               return rb_hole_addr_to_node(rb_next(&node->rb_hole_addr));
+               return next_hole_low_addr(node, size);
 
        case DRM_MM_INSERT_HIGH:
-               return rb_hole_addr_to_node(rb_prev(&node->rb_hole_addr));
+               return next_hole_high_addr(node, size);
 
        case DRM_MM_INSERT_EVICT:
                node = list_next_entry(node, hole_stack);
@@ -489,7 +584,7 @@ int drm_mm_insert_node_in_range(struct drm_mm * const mm,
        remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
        for (hole = first_hole(mm, range_start, range_end, size, mode);
             hole;
-            hole = once ? NULL : next_hole(mm, hole, mode)) {
+            hole = once ? NULL : next_hole(mm, hole, size, mode)) {
                u64 hole_start = __drm_mm_hole_node_start(hole);
                u64 hole_end = hole_start + hole->hole_size;
                u64 adj_start, adj_end;
index ee8b0e80ca90b72d34e10dffc50b5562a52144ae..a01bc6fac83cef32f77ac849643b115f033929e8 100644 (file)
@@ -168,6 +168,7 @@ struct drm_mm_node {
        struct rb_node rb_hole_addr;
        u64 __subtree_last;
        u64 hole_size;
+       u64 subtree_max_hole;
        unsigned long flags;
 #define DRM_MM_NODE_ALLOCATED_BIT      0
 #define DRM_MM_NODE_SCANNED_BIT                1