1 /**************************************************************************
3 * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA.
4 * Copyright 2016 Intel Corporation
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
28 **************************************************************************/
31 * Generic simple memory manager implementation. Intended to be used as a base
32 * class implementation for more advanced memory managers.
34 * Note that the algorithm used is quite simple and there might be substantial
35 * performance gains if a smarter free list is implemented. Currently it is
36 * just an unordered stack of free regions. This could easily be improved if
37 * an RB-tree is used instead. At least if we expect heavy fragmentation.
39 * Aligned allocations can also see improvement.
42 * Thomas Hellström <thomas-at-tungstengraphics-dot-com>
45 #include <linux/export.h>
46 #include <linux/interval_tree_generic.h>
47 #include <linux/seq_file.h>
48 #include <linux/slab.h>
49 #include <linux/stacktrace.h>
51 #include <drm/drm_mm.h>
56 * drm_mm provides a simple range allocator. The drivers are free to use the
57 * resource allocator from the linux core if it suits them, the upside of drm_mm
58 * is that it's in the DRM core. Which means that it's easier to extend for
59 * some of the crazier special purpose needs of gpus.
61 * The main data struct is &drm_mm, allocations are tracked in &drm_mm_node.
62 * Drivers are free to embed either of them into their own suitable
63 * datastructures. drm_mm itself will not do any memory allocations of its own,
64 * so if drivers choose not to embed nodes they need to still allocate them
67 * The range allocator also supports reservation of preallocated blocks. This is
68 * useful for taking over initial mode setting configurations from the firmware,
69 * where an object needs to be created which exactly matches the firmware's
70 * scanout target. As long as the range is still free it can be inserted anytime
71 * after the allocator is initialized, which helps with avoiding looped
72 * dependencies in the driver load sequence.
74 * drm_mm maintains a stack of most recently freed holes, which of all
75 * simplistic datastructures seems to be a fairly decent approach to clustering
76 * allocations and avoiding too much fragmentation. This means free space
77 * searches are O(num_holes). Given that all the fancy features drm_mm supports
78 * something better would be fairly complex and since gfx thrashing is a fairly
79 * steep cliff not a real concern. Removing a node again is O(1).
81 * drm_mm supports a few features: Alignment and range restrictions can be
82 * supplied. Furthermore every &drm_mm_node has a color value (which is just an
83 * opaque unsigned long) which in conjunction with a driver callback can be used
84 * to implement sophisticated placement restrictions. The i915 DRM driver uses
85 * this to implement guard pages between incompatible caching domains in the
88 * Two behaviors are supported for searching and allocating: bottom-up and
89 * top-down. The default is bottom-up. Top-down allocation can be used if the
90 * memory area has different restrictions, or just to reduce fragmentation.
92 * Finally iteration helpers to walk all nodes and all holes are provided as are
93 * some basic allocator dumpers for debugging.
95 * Note that this range allocator is not thread-safe, drivers need to protect
96 * modifications with their own locking. The idea behind this is that for a full
97 * memory manager additional data needs to be protected anyway, hence internal
98 * locking would be fully redundant.
101 #ifdef CONFIG_DRM_DEBUG_MM
102 #include <linux/stackdepot.h>
104 #define STACKDEPTH 32
107 static noinline
void save_stack(struct drm_mm_node
*node
)
109 unsigned long entries
[STACKDEPTH
];
112 n
= stack_trace_save(entries
, ARRAY_SIZE(entries
), 1);
114 /* May be called under spinlock, so avoid sleeping */
115 node
->stack
= stack_depot_save(entries
, n
, GFP_NOWAIT
);
118 static void show_leaks(struct drm_mm
*mm
)
120 struct drm_mm_node
*node
;
121 unsigned long *entries
;
122 unsigned int nr_entries
;
125 buf
= kmalloc(BUFSZ
, GFP_KERNEL
);
129 list_for_each_entry(node
, drm_mm_nodes(mm
), node_list
) {
131 DRM_ERROR("node [%08llx + %08llx]: unknown owner\n",
132 node
->start
, node
->size
);
136 nr_entries
= stack_depot_fetch(node
->stack
, &entries
);
137 stack_trace_snprint(buf
, BUFSZ
, entries
, nr_entries
, 0);
138 DRM_ERROR("node [%08llx + %08llx]: inserted at\n%s",
139 node
->start
, node
->size
, buf
);
148 static void save_stack(struct drm_mm_node
*node
) { }
149 static void show_leaks(struct drm_mm
*mm
) { }
152 #define START(node) ((node)->start)
153 #define LAST(node) ((node)->start + (node)->size - 1)
155 INTERVAL_TREE_DEFINE(struct drm_mm_node
, rb
,
157 START
, LAST
, static inline, drm_mm_interval_tree
)
160 __drm_mm_interval_first(const struct drm_mm
*mm
, u64 start
, u64 last
)
162 return drm_mm_interval_tree_iter_first((struct rb_root_cached
*)&mm
->interval_tree
,
163 start
, last
) ?: (struct drm_mm_node
*)&mm
->head_node
;
165 EXPORT_SYMBOL(__drm_mm_interval_first
);
167 static void drm_mm_interval_tree_add_node(struct drm_mm_node
*hole_node
,
168 struct drm_mm_node
*node
)
170 struct drm_mm
*mm
= hole_node
->mm
;
171 struct rb_node
**link
, *rb
;
172 struct drm_mm_node
*parent
;
175 node
->__subtree_last
= LAST(node
);
177 if (drm_mm_node_allocated(hole_node
)) {
180 parent
= rb_entry(rb
, struct drm_mm_node
, rb
);
181 if (parent
->__subtree_last
>= node
->__subtree_last
)
184 parent
->__subtree_last
= node
->__subtree_last
;
189 link
= &hole_node
->rb
.rb_right
;
193 link
= &mm
->interval_tree
.rb_root
.rb_node
;
199 parent
= rb_entry(rb
, struct drm_mm_node
, rb
);
200 if (parent
->__subtree_last
< node
->__subtree_last
)
201 parent
->__subtree_last
= node
->__subtree_last
;
202 if (node
->start
< parent
->start
) {
203 link
= &parent
->rb
.rb_left
;
205 link
= &parent
->rb
.rb_right
;
210 rb_link_node(&node
->rb
, rb
, link
);
211 rb_insert_augmented_cached(&node
->rb
, &mm
->interval_tree
, leftmost
,
212 &drm_mm_interval_tree_augment
);
215 #define RB_INSERT(root, member, expr) do { \
216 struct rb_node **link = &root.rb_node, *rb = NULL; \
217 u64 x = expr(node); \
220 if (x < expr(rb_entry(rb, struct drm_mm_node, member))) \
221 link = &rb->rb_left; \
223 link = &rb->rb_right; \
225 rb_link_node(&node->member, rb, link); \
226 rb_insert_color(&node->member, &root); \
229 #define HOLE_SIZE(NODE) ((NODE)->hole_size)
230 #define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE))
232 static u64
rb_to_hole_size(struct rb_node
*rb
)
234 return rb_entry(rb
, struct drm_mm_node
, rb_hole_size
)->hole_size
;
237 static void insert_hole_size(struct rb_root_cached
*root
,
238 struct drm_mm_node
*node
)
240 struct rb_node
**link
= &root
->rb_root
.rb_node
, *rb
= NULL
;
241 u64 x
= node
->hole_size
;
246 if (x
> rb_to_hole_size(rb
)) {
249 link
= &rb
->rb_right
;
254 rb_link_node(&node
->rb_hole_size
, rb
, link
);
255 rb_insert_color_cached(&node
->rb_hole_size
, root
, first
);
258 static void add_hole(struct drm_mm_node
*node
)
260 struct drm_mm
*mm
= node
->mm
;
263 __drm_mm_hole_node_end(node
) - __drm_mm_hole_node_start(node
);
264 DRM_MM_BUG_ON(!drm_mm_hole_follows(node
));
266 insert_hole_size(&mm
->holes_size
, node
);
267 RB_INSERT(mm
->holes_addr
, rb_hole_addr
, HOLE_ADDR
);
269 list_add(&node
->hole_stack
, &mm
->hole_stack
);
272 static void rm_hole(struct drm_mm_node
*node
)
274 DRM_MM_BUG_ON(!drm_mm_hole_follows(node
));
276 list_del(&node
->hole_stack
);
277 rb_erase_cached(&node
->rb_hole_size
, &node
->mm
->holes_size
);
278 rb_erase(&node
->rb_hole_addr
, &node
->mm
->holes_addr
);
281 DRM_MM_BUG_ON(drm_mm_hole_follows(node
));
284 static inline struct drm_mm_node
*rb_hole_size_to_node(struct rb_node
*rb
)
286 return rb_entry_safe(rb
, struct drm_mm_node
, rb_hole_size
);
289 static inline struct drm_mm_node
*rb_hole_addr_to_node(struct rb_node
*rb
)
291 return rb_entry_safe(rb
, struct drm_mm_node
, rb_hole_addr
);
294 static inline u64
rb_hole_size(struct rb_node
*rb
)
296 return rb_entry(rb
, struct drm_mm_node
, rb_hole_size
)->hole_size
;
299 static struct drm_mm_node
*best_hole(struct drm_mm
*mm
, u64 size
)
301 struct rb_node
*rb
= mm
->holes_size
.rb_root
.rb_node
;
302 struct drm_mm_node
*best
= NULL
;
305 struct drm_mm_node
*node
=
306 rb_entry(rb
, struct drm_mm_node
, rb_hole_size
);
308 if (size
<= node
->hole_size
) {
319 static struct drm_mm_node
*find_hole(struct drm_mm
*mm
, u64 addr
)
321 struct rb_node
*rb
= mm
->holes_addr
.rb_node
;
322 struct drm_mm_node
*node
= NULL
;
327 node
= rb_hole_addr_to_node(rb
);
328 hole_start
= __drm_mm_hole_node_start(node
);
330 if (addr
< hole_start
)
331 rb
= node
->rb_hole_addr
.rb_left
;
332 else if (addr
> hole_start
+ node
->hole_size
)
333 rb
= node
->rb_hole_addr
.rb_right
;
341 static struct drm_mm_node
*
342 first_hole(struct drm_mm
*mm
,
343 u64 start
, u64 end
, u64 size
,
344 enum drm_mm_insert_mode mode
)
348 case DRM_MM_INSERT_BEST
:
349 return best_hole(mm
, size
);
351 case DRM_MM_INSERT_LOW
:
352 return find_hole(mm
, start
);
354 case DRM_MM_INSERT_HIGH
:
355 return find_hole(mm
, end
);
357 case DRM_MM_INSERT_EVICT
:
358 return list_first_entry_or_null(&mm
->hole_stack
,
364 static struct drm_mm_node
*
365 next_hole(struct drm_mm
*mm
,
366 struct drm_mm_node
*node
,
367 enum drm_mm_insert_mode mode
)
371 case DRM_MM_INSERT_BEST
:
372 return rb_hole_size_to_node(rb_prev(&node
->rb_hole_size
));
374 case DRM_MM_INSERT_LOW
:
375 return rb_hole_addr_to_node(rb_next(&node
->rb_hole_addr
));
377 case DRM_MM_INSERT_HIGH
:
378 return rb_hole_addr_to_node(rb_prev(&node
->rb_hole_addr
));
380 case DRM_MM_INSERT_EVICT
:
381 node
= list_next_entry(node
, hole_stack
);
382 return &node
->hole_stack
== &mm
->hole_stack
? NULL
: node
;
387 * drm_mm_reserve_node - insert an pre-initialized node
388 * @mm: drm_mm allocator to insert @node into
389 * @node: drm_mm_node to insert
391 * This functions inserts an already set-up &drm_mm_node into the allocator,
392 * meaning that start, size and color must be set by the caller. All other
393 * fields must be cleared to 0. This is useful to initialize the allocator with
394 * preallocated objects which must be set-up before the range allocator can be
395 * set-up, e.g. when taking over a firmware framebuffer.
398 * 0 on success, -ENOSPC if there's no hole where @node is.
400 int drm_mm_reserve_node(struct drm_mm
*mm
, struct drm_mm_node
*node
)
402 u64 end
= node
->start
+ node
->size
;
403 struct drm_mm_node
*hole
;
404 u64 hole_start
, hole_end
;
405 u64 adj_start
, adj_end
;
407 end
= node
->start
+ node
->size
;
408 if (unlikely(end
<= node
->start
))
411 /* Find the relevant hole to add our node to */
412 hole
= find_hole(mm
, node
->start
);
416 adj_start
= hole_start
= __drm_mm_hole_node_start(hole
);
417 adj_end
= hole_end
= hole_start
+ hole
->hole_size
;
419 if (mm
->color_adjust
)
420 mm
->color_adjust(hole
, node
->color
, &adj_start
, &adj_end
);
422 if (adj_start
> node
->start
|| adj_end
< end
)
427 __set_bit(DRM_MM_NODE_ALLOCATED_BIT
, &node
->flags
);
428 list_add(&node
->node_list
, &hole
->node_list
);
429 drm_mm_interval_tree_add_node(hole
, node
);
433 if (node
->start
> hole_start
)
441 EXPORT_SYMBOL(drm_mm_reserve_node
);
443 static u64
rb_to_hole_size_or_zero(struct rb_node
*rb
)
445 return rb
? rb_to_hole_size(rb
) : 0;
449 * drm_mm_insert_node_in_range - ranged search for space and insert @node
450 * @mm: drm_mm to allocate from
451 * @node: preallocate node to insert
452 * @size: size of the allocation
453 * @alignment: alignment of the allocation
454 * @color: opaque tag value to use for this node
455 * @range_start: start of the allowed range for this node
456 * @range_end: end of the allowed range for this node
457 * @mode: fine-tune the allocation search and placement
459 * The preallocated @node must be cleared to 0.
462 * 0 on success, -ENOSPC if there's no suitable hole.
464 int drm_mm_insert_node_in_range(struct drm_mm
* const mm
,
465 struct drm_mm_node
* const node
,
466 u64 size
, u64 alignment
,
468 u64 range_start
, u64 range_end
,
469 enum drm_mm_insert_mode mode
)
471 struct drm_mm_node
*hole
;
475 DRM_MM_BUG_ON(range_start
> range_end
);
477 if (unlikely(size
== 0 || range_end
- range_start
< size
))
480 if (rb_to_hole_size_or_zero(rb_first_cached(&mm
->holes_size
)) < size
)
486 once
= mode
& DRM_MM_INSERT_ONCE
;
487 mode
&= ~DRM_MM_INSERT_ONCE
;
489 remainder_mask
= is_power_of_2(alignment
) ? alignment
- 1 : 0;
490 for (hole
= first_hole(mm
, range_start
, range_end
, size
, mode
);
492 hole
= once
? NULL
: next_hole(mm
, hole
, mode
)) {
493 u64 hole_start
= __drm_mm_hole_node_start(hole
);
494 u64 hole_end
= hole_start
+ hole
->hole_size
;
495 u64 adj_start
, adj_end
;
496 u64 col_start
, col_end
;
498 if (mode
== DRM_MM_INSERT_LOW
&& hole_start
>= range_end
)
501 if (mode
== DRM_MM_INSERT_HIGH
&& hole_end
<= range_start
)
504 col_start
= hole_start
;
506 if (mm
->color_adjust
)
507 mm
->color_adjust(hole
, color
, &col_start
, &col_end
);
509 adj_start
= max(col_start
, range_start
);
510 adj_end
= min(col_end
, range_end
);
512 if (adj_end
<= adj_start
|| adj_end
- adj_start
< size
)
515 if (mode
== DRM_MM_INSERT_HIGH
)
516 adj_start
= adj_end
- size
;
521 if (likely(remainder_mask
))
522 rem
= adj_start
& remainder_mask
;
524 div64_u64_rem(adj_start
, alignment
, &rem
);
527 if (mode
!= DRM_MM_INSERT_HIGH
)
528 adj_start
+= alignment
;
530 if (adj_start
< max(col_start
, range_start
) ||
531 min(col_end
, range_end
) - adj_start
< size
)
534 if (adj_end
<= adj_start
||
535 adj_end
- adj_start
< size
)
542 node
->start
= adj_start
;
546 __set_bit(DRM_MM_NODE_ALLOCATED_BIT
, &node
->flags
);
547 list_add(&node
->node_list
, &hole
->node_list
);
548 drm_mm_interval_tree_add_node(hole
, node
);
551 if (adj_start
> hole_start
)
553 if (adj_start
+ size
< hole_end
)
562 EXPORT_SYMBOL(drm_mm_insert_node_in_range
);
564 static inline bool drm_mm_node_scanned_block(const struct drm_mm_node
*node
)
566 return test_bit(DRM_MM_NODE_SCANNED_BIT
, &node
->flags
);
570 * drm_mm_remove_node - Remove a memory node from the allocator.
571 * @node: drm_mm_node to remove
573 * This just removes a node from its drm_mm allocator. The node does not need to
574 * be cleared again before it can be re-inserted into this or any other drm_mm
575 * allocator. It is a bug to call this function on a unallocated node.
577 void drm_mm_remove_node(struct drm_mm_node
*node
)
579 struct drm_mm
*mm
= node
->mm
;
580 struct drm_mm_node
*prev_node
;
582 DRM_MM_BUG_ON(!drm_mm_node_allocated(node
));
583 DRM_MM_BUG_ON(drm_mm_node_scanned_block(node
));
585 prev_node
= list_prev_entry(node
, node_list
);
587 if (drm_mm_hole_follows(node
))
590 drm_mm_interval_tree_remove(node
, &mm
->interval_tree
);
591 list_del(&node
->node_list
);
593 if (drm_mm_hole_follows(prev_node
))
597 clear_bit_unlock(DRM_MM_NODE_ALLOCATED_BIT
, &node
->flags
);
599 EXPORT_SYMBOL(drm_mm_remove_node
);
602 * drm_mm_replace_node - move an allocation from @old to @new
603 * @old: drm_mm_node to remove from the allocator
604 * @new: drm_mm_node which should inherit @old's allocation
606 * This is useful for when drivers embed the drm_mm_node structure and hence
607 * can't move allocations by reassigning pointers. It's a combination of remove
608 * and insert with the guarantee that the allocation start will match.
610 void drm_mm_replace_node(struct drm_mm_node
*old
, struct drm_mm_node
*new)
612 struct drm_mm
*mm
= old
->mm
;
614 DRM_MM_BUG_ON(!drm_mm_node_allocated(old
));
618 __set_bit(DRM_MM_NODE_ALLOCATED_BIT
, &new->flags
);
619 list_replace(&old
->node_list
, &new->node_list
);
620 rb_replace_node_cached(&old
->rb
, &new->rb
, &mm
->interval_tree
);
622 if (drm_mm_hole_follows(old
)) {
623 list_replace(&old
->hole_stack
, &new->hole_stack
);
624 rb_replace_node_cached(&old
->rb_hole_size
,
627 rb_replace_node(&old
->rb_hole_addr
,
632 clear_bit_unlock(DRM_MM_NODE_ALLOCATED_BIT
, &old
->flags
);
634 EXPORT_SYMBOL(drm_mm_replace_node
);
637 * DOC: lru scan roster
639 * Very often GPUs need to have continuous allocations for a given object. When
640 * evicting objects to make space for a new one it is therefore not most
641 * efficient when we simply start to select all objects from the tail of an LRU
642 * until there's a suitable hole: Especially for big objects or nodes that
643 * otherwise have special allocation constraints there's a good chance we evict
644 * lots of (smaller) objects unnecessarily.
646 * The DRM range allocator supports this use-case through the scanning
647 * interfaces. First a scan operation needs to be initialized with
648 * drm_mm_scan_init() or drm_mm_scan_init_with_range(). The driver adds
649 * objects to the roster, probably by walking an LRU list, but this can be
650 * freely implemented. Eviction candiates are added using
651 * drm_mm_scan_add_block() until a suitable hole is found or there are no
652 * further evictable objects. Eviction roster metadata is tracked in &struct
655 * The driver must walk through all objects again in exactly the reverse
656 * order to restore the allocator state. Note that while the allocator is used
657 * in the scan mode no other operation is allowed.
659 * Finally the driver evicts all objects selected (drm_mm_scan_remove_block()
660 * reported true) in the scan, and any overlapping nodes after color adjustment
661 * (drm_mm_scan_color_evict()). Adding and removing an object is O(1), and
662 * since freeing a node is also O(1) the overall complexity is
663 * O(scanned_objects). So like the free stack which needs to be walked before a
664 * scan operation even begins this is linear in the number of objects. It
665 * doesn't seem to hurt too badly.
669 * drm_mm_scan_init_with_range - initialize range-restricted lru scanning
671 * @mm: drm_mm to scan
672 * @size: size of the allocation
673 * @alignment: alignment of the allocation
674 * @color: opaque tag value to use for the allocation
675 * @start: start of the allowed range for the allocation
676 * @end: end of the allowed range for the allocation
677 * @mode: fine-tune the allocation search and placement
679 * This simply sets up the scanning routines with the parameters for the desired
683 * As long as the scan list is non-empty, no other operations than
684 * adding/removing nodes to/from the scan list are allowed.
686 void drm_mm_scan_init_with_range(struct drm_mm_scan
*scan
,
693 enum drm_mm_insert_mode mode
)
695 DRM_MM_BUG_ON(start
>= end
);
696 DRM_MM_BUG_ON(!size
|| size
> end
- start
);
697 DRM_MM_BUG_ON(mm
->scan_active
);
705 scan
->alignment
= alignment
;
706 scan
->remainder_mask
= is_power_of_2(alignment
) ? alignment
- 1 : 0;
710 DRM_MM_BUG_ON(end
<= start
);
711 scan
->range_start
= start
;
712 scan
->range_end
= end
;
714 scan
->hit_start
= U64_MAX
;
717 EXPORT_SYMBOL(drm_mm_scan_init_with_range
);
720 * drm_mm_scan_add_block - add a node to the scan list
721 * @scan: the active drm_mm scanner
722 * @node: drm_mm_node to add
724 * Add a node to the scan list that might be freed to make space for the desired
728 * True if a hole has been found, false otherwise.
730 bool drm_mm_scan_add_block(struct drm_mm_scan
*scan
,
731 struct drm_mm_node
*node
)
733 struct drm_mm
*mm
= scan
->mm
;
734 struct drm_mm_node
*hole
;
735 u64 hole_start
, hole_end
;
736 u64 col_start
, col_end
;
737 u64 adj_start
, adj_end
;
739 DRM_MM_BUG_ON(node
->mm
!= mm
);
740 DRM_MM_BUG_ON(!drm_mm_node_allocated(node
));
741 DRM_MM_BUG_ON(drm_mm_node_scanned_block(node
));
742 __set_bit(DRM_MM_NODE_SCANNED_BIT
, &node
->flags
);
745 /* Remove this block from the node_list so that we enlarge the hole
746 * (distance between the end of our previous node and the start of
747 * or next), without poisoning the link so that we can restore it
748 * later in drm_mm_scan_remove_block().
750 hole
= list_prev_entry(node
, node_list
);
751 DRM_MM_BUG_ON(list_next_entry(hole
, node_list
) != node
);
752 __list_del_entry(&node
->node_list
);
754 hole_start
= __drm_mm_hole_node_start(hole
);
755 hole_end
= __drm_mm_hole_node_end(hole
);
757 col_start
= hole_start
;
759 if (mm
->color_adjust
)
760 mm
->color_adjust(hole
, scan
->color
, &col_start
, &col_end
);
762 adj_start
= max(col_start
, scan
->range_start
);
763 adj_end
= min(col_end
, scan
->range_end
);
764 if (adj_end
<= adj_start
|| adj_end
- adj_start
< scan
->size
)
767 if (scan
->mode
== DRM_MM_INSERT_HIGH
)
768 adj_start
= adj_end
- scan
->size
;
770 if (scan
->alignment
) {
773 if (likely(scan
->remainder_mask
))
774 rem
= adj_start
& scan
->remainder_mask
;
776 div64_u64_rem(adj_start
, scan
->alignment
, &rem
);
779 if (scan
->mode
!= DRM_MM_INSERT_HIGH
)
780 adj_start
+= scan
->alignment
;
781 if (adj_start
< max(col_start
, scan
->range_start
) ||
782 min(col_end
, scan
->range_end
) - adj_start
< scan
->size
)
785 if (adj_end
<= adj_start
||
786 adj_end
- adj_start
< scan
->size
)
791 scan
->hit_start
= adj_start
;
792 scan
->hit_end
= adj_start
+ scan
->size
;
794 DRM_MM_BUG_ON(scan
->hit_start
>= scan
->hit_end
);
795 DRM_MM_BUG_ON(scan
->hit_start
< hole_start
);
796 DRM_MM_BUG_ON(scan
->hit_end
> hole_end
);
800 EXPORT_SYMBOL(drm_mm_scan_add_block
);
803 * drm_mm_scan_remove_block - remove a node from the scan list
804 * @scan: the active drm_mm scanner
805 * @node: drm_mm_node to remove
807 * Nodes **must** be removed in exactly the reverse order from the scan list as
808 * they have been added (e.g. using list_add() as they are added and then
809 * list_for_each() over that eviction list to remove), otherwise the internal
810 * state of the memory manager will be corrupted.
812 * When the scan list is empty, the selected memory nodes can be freed. An
813 * immediately following drm_mm_insert_node_in_range_generic() or one of the
814 * simpler versions of that function with !DRM_MM_SEARCH_BEST will then return
815 * the just freed block (because it's at the top of the free_stack list).
818 * True if this block should be evicted, false otherwise. Will always
819 * return false when no hole has been found.
821 bool drm_mm_scan_remove_block(struct drm_mm_scan
*scan
,
822 struct drm_mm_node
*node
)
824 struct drm_mm_node
*prev_node
;
826 DRM_MM_BUG_ON(node
->mm
!= scan
->mm
);
827 DRM_MM_BUG_ON(!drm_mm_node_scanned_block(node
));
828 __clear_bit(DRM_MM_NODE_SCANNED_BIT
, &node
->flags
);
830 DRM_MM_BUG_ON(!node
->mm
->scan_active
);
831 node
->mm
->scan_active
--;
833 /* During drm_mm_scan_add_block() we decoupled this node leaving
834 * its pointers intact. Now that the caller is walking back along
835 * the eviction list we can restore this block into its rightful
836 * place on the full node_list. To confirm that the caller is walking
837 * backwards correctly we check that prev_node->next == node->next,
838 * i.e. both believe the same node should be on the other side of the
841 prev_node
= list_prev_entry(node
, node_list
);
842 DRM_MM_BUG_ON(list_next_entry(prev_node
, node_list
) !=
843 list_next_entry(node
, node_list
));
844 list_add(&node
->node_list
, &prev_node
->node_list
);
846 return (node
->start
+ node
->size
> scan
->hit_start
&&
847 node
->start
< scan
->hit_end
);
849 EXPORT_SYMBOL(drm_mm_scan_remove_block
);
852 * drm_mm_scan_color_evict - evict overlapping nodes on either side of hole
853 * @scan: drm_mm scan with target hole
855 * After completing an eviction scan and removing the selected nodes, we may
856 * need to remove a few more nodes from either side of the target hole if
857 * mm.color_adjust is being used.
860 * A node to evict, or NULL if there are no overlapping nodes.
862 struct drm_mm_node
*drm_mm_scan_color_evict(struct drm_mm_scan
*scan
)
864 struct drm_mm
*mm
= scan
->mm
;
865 struct drm_mm_node
*hole
;
866 u64 hole_start
, hole_end
;
868 DRM_MM_BUG_ON(list_empty(&mm
->hole_stack
));
870 if (!mm
->color_adjust
)
874 * The hole found during scanning should ideally be the first element
875 * in the hole_stack list, but due to side-effects in the driver it
878 list_for_each_entry(hole
, &mm
->hole_stack
, hole_stack
) {
879 hole_start
= __drm_mm_hole_node_start(hole
);
880 hole_end
= hole_start
+ hole
->hole_size
;
882 if (hole_start
<= scan
->hit_start
&&
883 hole_end
>= scan
->hit_end
)
887 /* We should only be called after we found the hole previously */
888 DRM_MM_BUG_ON(&hole
->hole_stack
== &mm
->hole_stack
);
889 if (unlikely(&hole
->hole_stack
== &mm
->hole_stack
))
892 DRM_MM_BUG_ON(hole_start
> scan
->hit_start
);
893 DRM_MM_BUG_ON(hole_end
< scan
->hit_end
);
895 mm
->color_adjust(hole
, scan
->color
, &hole_start
, &hole_end
);
896 if (hole_start
> scan
->hit_start
)
898 if (hole_end
< scan
->hit_end
)
899 return list_next_entry(hole
, node_list
);
903 EXPORT_SYMBOL(drm_mm_scan_color_evict
);
906 * drm_mm_init - initialize a drm-mm allocator
907 * @mm: the drm_mm structure to initialize
908 * @start: start of the range managed by @mm
909 * @size: end of the range managed by @mm
911 * Note that @mm must be cleared to 0 before calling this function.
913 void drm_mm_init(struct drm_mm
*mm
, u64 start
, u64 size
)
915 DRM_MM_BUG_ON(start
+ size
<= start
);
917 mm
->color_adjust
= NULL
;
919 INIT_LIST_HEAD(&mm
->hole_stack
);
920 mm
->interval_tree
= RB_ROOT_CACHED
;
921 mm
->holes_size
= RB_ROOT_CACHED
;
922 mm
->holes_addr
= RB_ROOT
;
924 /* Clever trick to avoid a special case in the free hole tracking. */
925 INIT_LIST_HEAD(&mm
->head_node
.node_list
);
926 mm
->head_node
.flags
= 0;
927 mm
->head_node
.mm
= mm
;
928 mm
->head_node
.start
= start
+ size
;
929 mm
->head_node
.size
= -size
;
930 add_hole(&mm
->head_node
);
934 EXPORT_SYMBOL(drm_mm_init
);
937 * drm_mm_takedown - clean up a drm_mm allocator
938 * @mm: drm_mm allocator to clean up
940 * Note that it is a bug to call this function on an allocator which is not
943 void drm_mm_takedown(struct drm_mm
*mm
)
945 if (WARN(!drm_mm_clean(mm
),
946 "Memory manager not clean during takedown.\n"))
949 EXPORT_SYMBOL(drm_mm_takedown
);
951 static u64
drm_mm_dump_hole(struct drm_printer
*p
, const struct drm_mm_node
*entry
)
955 size
= entry
->hole_size
;
957 start
= drm_mm_hole_node_start(entry
);
958 drm_printf(p
, "%#018llx-%#018llx: %llu: free\n",
959 start
, start
+ size
, size
);
965 * drm_mm_print - print allocator state
966 * @mm: drm_mm allocator to print
967 * @p: DRM printer to use
969 void drm_mm_print(const struct drm_mm
*mm
, struct drm_printer
*p
)
971 const struct drm_mm_node
*entry
;
972 u64 total_used
= 0, total_free
= 0, total
= 0;
974 total_free
+= drm_mm_dump_hole(p
, &mm
->head_node
);
976 drm_mm_for_each_node(entry
, mm
) {
977 drm_printf(p
, "%#018llx-%#018llx: %llu: used\n", entry
->start
,
978 entry
->start
+ entry
->size
, entry
->size
);
979 total_used
+= entry
->size
;
980 total_free
+= drm_mm_dump_hole(p
, entry
);
982 total
= total_free
+ total_used
;
984 drm_printf(p
, "total: %llu, used %llu free %llu\n", total
,
985 total_used
, total_free
);
987 EXPORT_SYMBOL(drm_mm_print
);