xref: /linux/drivers/gpu/drm/drm_mm.c (revision a1ff5a7d78a036d6c2178ee5acd6ba4946243800)
1 /**************************************************************************
2  *
3  * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA.
4  * Copyright 2016 Intel Corporation
5  * All Rights Reserved.
6  *
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:
14  *
15  * The above copyright notice and this permission notice (including the
16  * next paragraph) shall be included in all copies or substantial portions
17  * of the Software.
18  *
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.
26  *
27  *
28  **************************************************************************/
29 
30 /*
31  * Generic simple memory manager implementation. Intended to be used as a base
32  * class implementation for more advanced memory managers.
33  *
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.
38  *
39  * Aligned allocations can also see improvement.
40  *
41  * Authors:
42  * Thomas Hellström <thomas-at-tungstengraphics-dot-com>
43  */
44 
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>
50 
51 #include <drm/drm_mm.h>
52 
53 /**
54  * DOC: Overview
55  *
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.
60  *
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
65  * themselves.
66  *
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.
73  *
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).
80  *
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
86  * graphics TT.
87  *
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.
91  *
92  * Finally iteration helpers to walk all nodes and all holes are provided as are
93  * some basic allocator dumpers for debugging.
94  *
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.
99  */
100 
101 #ifdef CONFIG_DRM_DEBUG_MM
102 #include <linux/stackdepot.h>
103 
104 #define STACKDEPTH 32
105 #define BUFSZ 4096
106 
save_stack(struct drm_mm_node * node)107 static noinline void save_stack(struct drm_mm_node *node)
108 {
109 	unsigned long entries[STACKDEPTH];
110 	unsigned int n;
111 
112 	n = stack_trace_save(entries, ARRAY_SIZE(entries), 1);
113 
114 	/* May be called under spinlock, so avoid sleeping */
115 	node->stack = stack_depot_save(entries, n, GFP_NOWAIT);
116 }
117 
show_leaks(struct drm_mm * mm)118 static void show_leaks(struct drm_mm *mm)
119 {
120 	struct drm_mm_node *node;
121 	char *buf;
122 
123 	buf = kmalloc(BUFSZ, GFP_KERNEL);
124 	if (!buf)
125 		return;
126 
127 	list_for_each_entry(node, drm_mm_nodes(mm), node_list) {
128 		if (!node->stack) {
129 			DRM_ERROR("node [%08llx + %08llx]: unknown owner\n",
130 				  node->start, node->size);
131 			continue;
132 		}
133 
134 		stack_depot_snprint(node->stack, buf, BUFSZ, 0);
135 		DRM_ERROR("node [%08llx + %08llx]: inserted at\n%s",
136 			  node->start, node->size, buf);
137 	}
138 
139 	kfree(buf);
140 }
141 
142 #undef STACKDEPTH
143 #undef BUFSZ
144 #else
save_stack(struct drm_mm_node * node)145 static void save_stack(struct drm_mm_node *node) { }
show_leaks(struct drm_mm * mm)146 static void show_leaks(struct drm_mm *mm) { }
147 #endif
148 
149 #define START(node) ((node)->start)
150 #define LAST(node)  ((node)->start + (node)->size - 1)
151 
INTERVAL_TREE_DEFINE(struct drm_mm_node,rb,u64,__subtree_last,START,LAST,static inline,drm_mm_interval_tree)152 INTERVAL_TREE_DEFINE(struct drm_mm_node, rb,
153 		     u64, __subtree_last,
154 		     START, LAST, static inline, drm_mm_interval_tree)
155 
156 struct drm_mm_node *
157 __drm_mm_interval_first(const struct drm_mm *mm, u64 start, u64 last)
158 {
159 	return drm_mm_interval_tree_iter_first((struct rb_root_cached *)&mm->interval_tree,
160 					       start, last) ?: (struct drm_mm_node *)&mm->head_node;
161 }
162 EXPORT_SYMBOL(__drm_mm_interval_first);
163 
drm_mm_interval_tree_add_node(struct drm_mm_node * hole_node,struct drm_mm_node * node)164 static void drm_mm_interval_tree_add_node(struct drm_mm_node *hole_node,
165 					  struct drm_mm_node *node)
166 {
167 	struct drm_mm *mm = hole_node->mm;
168 	struct rb_node **link, *rb;
169 	struct drm_mm_node *parent;
170 	bool leftmost;
171 
172 	node->__subtree_last = LAST(node);
173 
174 	if (drm_mm_node_allocated(hole_node)) {
175 		rb = &hole_node->rb;
176 		while (rb) {
177 			parent = rb_entry(rb, struct drm_mm_node, rb);
178 			if (parent->__subtree_last >= node->__subtree_last)
179 				break;
180 
181 			parent->__subtree_last = node->__subtree_last;
182 			rb = rb_parent(rb);
183 		}
184 
185 		rb = &hole_node->rb;
186 		link = &hole_node->rb.rb_right;
187 		leftmost = false;
188 	} else {
189 		rb = NULL;
190 		link = &mm->interval_tree.rb_root.rb_node;
191 		leftmost = true;
192 	}
193 
194 	while (*link) {
195 		rb = *link;
196 		parent = rb_entry(rb, struct drm_mm_node, rb);
197 		if (parent->__subtree_last < node->__subtree_last)
198 			parent->__subtree_last = node->__subtree_last;
199 		if (node->start < parent->start) {
200 			link = &parent->rb.rb_left;
201 		} else {
202 			link = &parent->rb.rb_right;
203 			leftmost = false;
204 		}
205 	}
206 
207 	rb_link_node(&node->rb, rb, link);
208 	rb_insert_augmented_cached(&node->rb, &mm->interval_tree, leftmost,
209 				   &drm_mm_interval_tree_augment);
210 }
211 
212 #define HOLE_SIZE(NODE) ((NODE)->hole_size)
213 #define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE))
214 
rb_to_hole_size(struct rb_node * rb)215 static u64 rb_to_hole_size(struct rb_node *rb)
216 {
217 	return rb_entry(rb, struct drm_mm_node, rb_hole_size)->hole_size;
218 }
219 
insert_hole_size(struct rb_root_cached * root,struct drm_mm_node * node)220 static void insert_hole_size(struct rb_root_cached *root,
221 			     struct drm_mm_node *node)
222 {
223 	struct rb_node **link = &root->rb_root.rb_node, *rb = NULL;
224 	u64 x = node->hole_size;
225 	bool first = true;
226 
227 	while (*link) {
228 		rb = *link;
229 		if (x > rb_to_hole_size(rb)) {
230 			link = &rb->rb_left;
231 		} else {
232 			link = &rb->rb_right;
233 			first = false;
234 		}
235 	}
236 
237 	rb_link_node(&node->rb_hole_size, rb, link);
238 	rb_insert_color_cached(&node->rb_hole_size, root, first);
239 }
240 
RB_DECLARE_CALLBACKS_MAX(static,augment_callbacks,struct drm_mm_node,rb_hole_addr,u64,subtree_max_hole,HOLE_SIZE)241 RB_DECLARE_CALLBACKS_MAX(static, augment_callbacks,
242 			 struct drm_mm_node, rb_hole_addr,
243 			 u64, subtree_max_hole, HOLE_SIZE)
244 
245 static void insert_hole_addr(struct rb_root *root, struct drm_mm_node *node)
246 {
247 	struct rb_node **link = &root->rb_node, *rb_parent = NULL;
248 	u64 start = HOLE_ADDR(node), subtree_max_hole = node->subtree_max_hole;
249 	struct drm_mm_node *parent;
250 
251 	while (*link) {
252 		rb_parent = *link;
253 		parent = rb_entry(rb_parent, struct drm_mm_node, rb_hole_addr);
254 		if (parent->subtree_max_hole < subtree_max_hole)
255 			parent->subtree_max_hole = subtree_max_hole;
256 		if (start < HOLE_ADDR(parent))
257 			link = &parent->rb_hole_addr.rb_left;
258 		else
259 			link = &parent->rb_hole_addr.rb_right;
260 	}
261 
262 	rb_link_node(&node->rb_hole_addr, rb_parent, link);
263 	rb_insert_augmented(&node->rb_hole_addr, root, &augment_callbacks);
264 }
265 
add_hole(struct drm_mm_node * node)266 static void add_hole(struct drm_mm_node *node)
267 {
268 	struct drm_mm *mm = node->mm;
269 
270 	node->hole_size =
271 		__drm_mm_hole_node_end(node) - __drm_mm_hole_node_start(node);
272 	node->subtree_max_hole = node->hole_size;
273 	DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
274 
275 	insert_hole_size(&mm->holes_size, node);
276 	insert_hole_addr(&mm->holes_addr, node);
277 
278 	list_add(&node->hole_stack, &mm->hole_stack);
279 }
280 
rm_hole(struct drm_mm_node * node)281 static void rm_hole(struct drm_mm_node *node)
282 {
283 	DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
284 
285 	list_del(&node->hole_stack);
286 	rb_erase_cached(&node->rb_hole_size, &node->mm->holes_size);
287 	rb_erase_augmented(&node->rb_hole_addr, &node->mm->holes_addr,
288 			   &augment_callbacks);
289 	node->hole_size = 0;
290 	node->subtree_max_hole = 0;
291 
292 	DRM_MM_BUG_ON(drm_mm_hole_follows(node));
293 }
294 
rb_hole_size_to_node(struct rb_node * rb)295 static inline struct drm_mm_node *rb_hole_size_to_node(struct rb_node *rb)
296 {
297 	return rb_entry_safe(rb, struct drm_mm_node, rb_hole_size);
298 }
299 
rb_hole_addr_to_node(struct rb_node * rb)300 static inline struct drm_mm_node *rb_hole_addr_to_node(struct rb_node *rb)
301 {
302 	return rb_entry_safe(rb, struct drm_mm_node, rb_hole_addr);
303 }
304 
best_hole(struct drm_mm * mm,u64 size)305 static struct drm_mm_node *best_hole(struct drm_mm *mm, u64 size)
306 {
307 	struct rb_node *rb = mm->holes_size.rb_root.rb_node;
308 	struct drm_mm_node *best = NULL;
309 
310 	do {
311 		struct drm_mm_node *node =
312 			rb_entry(rb, struct drm_mm_node, rb_hole_size);
313 
314 		if (size <= node->hole_size) {
315 			best = node;
316 			rb = rb->rb_right;
317 		} else {
318 			rb = rb->rb_left;
319 		}
320 	} while (rb);
321 
322 	return best;
323 }
324 
usable_hole_addr(struct rb_node * rb,u64 size)325 static bool usable_hole_addr(struct rb_node *rb, u64 size)
326 {
327 	return rb && rb_hole_addr_to_node(rb)->subtree_max_hole >= size;
328 }
329 
find_hole_addr(struct drm_mm * mm,u64 addr,u64 size)330 static struct drm_mm_node *find_hole_addr(struct drm_mm *mm, u64 addr, u64 size)
331 {
332 	struct rb_node *rb = mm->holes_addr.rb_node;
333 	struct drm_mm_node *node = NULL;
334 
335 	while (rb) {
336 		u64 hole_start;
337 
338 		if (!usable_hole_addr(rb, size))
339 			break;
340 
341 		node = rb_hole_addr_to_node(rb);
342 		hole_start = __drm_mm_hole_node_start(node);
343 
344 		if (addr < hole_start)
345 			rb = node->rb_hole_addr.rb_left;
346 		else if (addr > hole_start + node->hole_size)
347 			rb = node->rb_hole_addr.rb_right;
348 		else
349 			break;
350 	}
351 
352 	return node;
353 }
354 
355 static struct drm_mm_node *
first_hole(struct drm_mm * mm,u64 start,u64 end,u64 size,enum drm_mm_insert_mode mode)356 first_hole(struct drm_mm *mm,
357 	   u64 start, u64 end, u64 size,
358 	   enum drm_mm_insert_mode mode)
359 {
360 	switch (mode) {
361 	default:
362 	case DRM_MM_INSERT_BEST:
363 		return best_hole(mm, size);
364 
365 	case DRM_MM_INSERT_LOW:
366 		return find_hole_addr(mm, start, size);
367 
368 	case DRM_MM_INSERT_HIGH:
369 		return find_hole_addr(mm, end, size);
370 
371 	case DRM_MM_INSERT_EVICT:
372 		return list_first_entry_or_null(&mm->hole_stack,
373 						struct drm_mm_node,
374 						hole_stack);
375 	}
376 }
377 
378 /**
379  * DECLARE_NEXT_HOLE_ADDR - macro to declare next hole functions
380  * @name: name of function to declare
381  * @first: first rb member to traverse (either rb_left or rb_right).
382  * @last: last rb member to traverse (either rb_right or rb_left).
383  *
384  * This macro declares a function to return the next hole of the addr rb tree.
385  * While traversing the tree we take the searched size into account and only
386  * visit branches with potential big enough holes.
387  */
388 
389 #define DECLARE_NEXT_HOLE_ADDR(name, first, last)			\
390 static struct drm_mm_node *name(struct drm_mm_node *entry, u64 size)	\
391 {									\
392 	struct rb_node *parent, *node = &entry->rb_hole_addr;		\
393 									\
394 	if (!entry || RB_EMPTY_NODE(node))				\
395 		return NULL;						\
396 									\
397 	if (usable_hole_addr(node->first, size)) {			\
398 		node = node->first;					\
399 		while (usable_hole_addr(node->last, size))		\
400 			node = node->last;				\
401 		return rb_hole_addr_to_node(node);			\
402 	}								\
403 									\
404 	while ((parent = rb_parent(node)) && node == parent->first)	\
405 		node = parent;						\
406 									\
407 	return rb_hole_addr_to_node(parent);				\
408 }
409 
DECLARE_NEXT_HOLE_ADDR(next_hole_high_addr,rb_left,rb_right)410 DECLARE_NEXT_HOLE_ADDR(next_hole_high_addr, rb_left, rb_right)
411 DECLARE_NEXT_HOLE_ADDR(next_hole_low_addr, rb_right, rb_left)
412 
413 static struct drm_mm_node *
414 next_hole(struct drm_mm *mm,
415 	  struct drm_mm_node *node,
416 	  u64 size,
417 	  enum drm_mm_insert_mode mode)
418 {
419 	switch (mode) {
420 	default:
421 	case DRM_MM_INSERT_BEST:
422 		return rb_hole_size_to_node(rb_prev(&node->rb_hole_size));
423 
424 	case DRM_MM_INSERT_LOW:
425 		return next_hole_low_addr(node, size);
426 
427 	case DRM_MM_INSERT_HIGH:
428 		return next_hole_high_addr(node, size);
429 
430 	case DRM_MM_INSERT_EVICT:
431 		node = list_next_entry(node, hole_stack);
432 		return &node->hole_stack == &mm->hole_stack ? NULL : node;
433 	}
434 }
435 
436 /**
437  * drm_mm_reserve_node - insert an pre-initialized node
438  * @mm: drm_mm allocator to insert @node into
439  * @node: drm_mm_node to insert
440  *
441  * This functions inserts an already set-up &drm_mm_node into the allocator,
442  * meaning that start, size and color must be set by the caller. All other
443  * fields must be cleared to 0. This is useful to initialize the allocator with
444  * preallocated objects which must be set-up before the range allocator can be
445  * set-up, e.g. when taking over a firmware framebuffer.
446  *
447  * Returns:
448  * 0 on success, -ENOSPC if there's no hole where @node is.
449  */
drm_mm_reserve_node(struct drm_mm * mm,struct drm_mm_node * node)450 int drm_mm_reserve_node(struct drm_mm *mm, struct drm_mm_node *node)
451 {
452 	struct drm_mm_node *hole;
453 	u64 hole_start, hole_end;
454 	u64 adj_start, adj_end;
455 	u64 end;
456 
457 	end = node->start + node->size;
458 	if (unlikely(end <= node->start))
459 		return -ENOSPC;
460 
461 	/* Find the relevant hole to add our node to */
462 	hole = find_hole_addr(mm, node->start, 0);
463 	if (!hole)
464 		return -ENOSPC;
465 
466 	adj_start = hole_start = __drm_mm_hole_node_start(hole);
467 	adj_end = hole_end = hole_start + hole->hole_size;
468 
469 	if (mm->color_adjust)
470 		mm->color_adjust(hole, node->color, &adj_start, &adj_end);
471 
472 	if (adj_start > node->start || adj_end < end)
473 		return -ENOSPC;
474 
475 	node->mm = mm;
476 
477 	__set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
478 	list_add(&node->node_list, &hole->node_list);
479 	drm_mm_interval_tree_add_node(hole, node);
480 	node->hole_size = 0;
481 
482 	rm_hole(hole);
483 	if (node->start > hole_start)
484 		add_hole(hole);
485 	if (end < hole_end)
486 		add_hole(node);
487 
488 	save_stack(node);
489 	return 0;
490 }
491 EXPORT_SYMBOL(drm_mm_reserve_node);
492 
rb_to_hole_size_or_zero(struct rb_node * rb)493 static u64 rb_to_hole_size_or_zero(struct rb_node *rb)
494 {
495 	return rb ? rb_to_hole_size(rb) : 0;
496 }
497 
498 /**
499  * drm_mm_insert_node_in_range - ranged search for space and insert @node
500  * @mm: drm_mm to allocate from
501  * @node: preallocate node to insert
502  * @size: size of the allocation
503  * @alignment: alignment of the allocation
504  * @color: opaque tag value to use for this node
505  * @range_start: start of the allowed range for this node
506  * @range_end: end of the allowed range for this node
507  * @mode: fine-tune the allocation search and placement
508  *
509  * The preallocated @node must be cleared to 0.
510  *
511  * Returns:
512  * 0 on success, -ENOSPC if there's no suitable hole.
513  */
drm_mm_insert_node_in_range(struct drm_mm * const mm,struct drm_mm_node * const node,u64 size,u64 alignment,unsigned long color,u64 range_start,u64 range_end,enum drm_mm_insert_mode mode)514 int drm_mm_insert_node_in_range(struct drm_mm * const mm,
515 				struct drm_mm_node * const node,
516 				u64 size, u64 alignment,
517 				unsigned long color,
518 				u64 range_start, u64 range_end,
519 				enum drm_mm_insert_mode mode)
520 {
521 	struct drm_mm_node *hole;
522 	u64 remainder_mask;
523 	bool once;
524 
525 	DRM_MM_BUG_ON(range_start > range_end);
526 
527 	if (unlikely(size == 0 || range_end - range_start < size))
528 		return -ENOSPC;
529 
530 	if (rb_to_hole_size_or_zero(rb_first_cached(&mm->holes_size)) < size)
531 		return -ENOSPC;
532 
533 	if (alignment <= 1)
534 		alignment = 0;
535 
536 	once = mode & DRM_MM_INSERT_ONCE;
537 	mode &= ~DRM_MM_INSERT_ONCE;
538 
539 	remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
540 	for (hole = first_hole(mm, range_start, range_end, size, mode);
541 	     hole;
542 	     hole = once ? NULL : next_hole(mm, hole, size, mode)) {
543 		u64 hole_start = __drm_mm_hole_node_start(hole);
544 		u64 hole_end = hole_start + hole->hole_size;
545 		u64 adj_start, adj_end;
546 		u64 col_start, col_end;
547 
548 		if (mode == DRM_MM_INSERT_LOW && hole_start >= range_end)
549 			break;
550 
551 		if (mode == DRM_MM_INSERT_HIGH && hole_end <= range_start)
552 			break;
553 
554 		col_start = hole_start;
555 		col_end = hole_end;
556 		if (mm->color_adjust)
557 			mm->color_adjust(hole, color, &col_start, &col_end);
558 
559 		adj_start = max(col_start, range_start);
560 		adj_end = min(col_end, range_end);
561 
562 		if (adj_end <= adj_start || adj_end - adj_start < size)
563 			continue;
564 
565 		if (mode == DRM_MM_INSERT_HIGH)
566 			adj_start = adj_end - size;
567 
568 		if (alignment) {
569 			u64 rem;
570 
571 			if (likely(remainder_mask))
572 				rem = adj_start & remainder_mask;
573 			else
574 				div64_u64_rem(adj_start, alignment, &rem);
575 			if (rem) {
576 				adj_start -= rem;
577 				if (mode != DRM_MM_INSERT_HIGH)
578 					adj_start += alignment;
579 
580 				if (adj_start < max(col_start, range_start) ||
581 				    min(col_end, range_end) - adj_start < size)
582 					continue;
583 
584 				if (adj_end <= adj_start ||
585 				    adj_end - adj_start < size)
586 					continue;
587 			}
588 		}
589 
590 		node->mm = mm;
591 		node->size = size;
592 		node->start = adj_start;
593 		node->color = color;
594 		node->hole_size = 0;
595 
596 		__set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
597 		list_add(&node->node_list, &hole->node_list);
598 		drm_mm_interval_tree_add_node(hole, node);
599 
600 		rm_hole(hole);
601 		if (adj_start > hole_start)
602 			add_hole(hole);
603 		if (adj_start + size < hole_end)
604 			add_hole(node);
605 
606 		save_stack(node);
607 		return 0;
608 	}
609 
610 	return -ENOSPC;
611 }
612 EXPORT_SYMBOL(drm_mm_insert_node_in_range);
613 
drm_mm_node_scanned_block(const struct drm_mm_node * node)614 static inline bool drm_mm_node_scanned_block(const struct drm_mm_node *node)
615 {
616 	return test_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
617 }
618 
619 /**
620  * drm_mm_remove_node - Remove a memory node from the allocator.
621  * @node: drm_mm_node to remove
622  *
623  * This just removes a node from its drm_mm allocator. The node does not need to
624  * be cleared again before it can be re-inserted into this or any other drm_mm
625  * allocator. It is a bug to call this function on a unallocated node.
626  */
drm_mm_remove_node(struct drm_mm_node * node)627 void drm_mm_remove_node(struct drm_mm_node *node)
628 {
629 	struct drm_mm *mm = node->mm;
630 	struct drm_mm_node *prev_node;
631 
632 	DRM_MM_BUG_ON(!drm_mm_node_allocated(node));
633 	DRM_MM_BUG_ON(drm_mm_node_scanned_block(node));
634 
635 	prev_node = list_prev_entry(node, node_list);
636 
637 	if (drm_mm_hole_follows(node))
638 		rm_hole(node);
639 
640 	drm_mm_interval_tree_remove(node, &mm->interval_tree);
641 	list_del(&node->node_list);
642 
643 	if (drm_mm_hole_follows(prev_node))
644 		rm_hole(prev_node);
645 	add_hole(prev_node);
646 
647 	clear_bit_unlock(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
648 }
649 EXPORT_SYMBOL(drm_mm_remove_node);
650 
651 /**
652  * DOC: lru scan roster
653  *
654  * Very often GPUs need to have continuous allocations for a given object. When
655  * evicting objects to make space for a new one it is therefore not most
656  * efficient when we simply start to select all objects from the tail of an LRU
657  * until there's a suitable hole: Especially for big objects or nodes that
658  * otherwise have special allocation constraints there's a good chance we evict
659  * lots of (smaller) objects unnecessarily.
660  *
661  * The DRM range allocator supports this use-case through the scanning
662  * interfaces. First a scan operation needs to be initialized with
663  * drm_mm_scan_init() or drm_mm_scan_init_with_range(). The driver adds
664  * objects to the roster, probably by walking an LRU list, but this can be
665  * freely implemented. Eviction candidates are added using
666  * drm_mm_scan_add_block() until a suitable hole is found or there are no
667  * further evictable objects. Eviction roster metadata is tracked in &struct
668  * drm_mm_scan.
669  *
670  * The driver must walk through all objects again in exactly the reverse
671  * order to restore the allocator state. Note that while the allocator is used
672  * in the scan mode no other operation is allowed.
673  *
674  * Finally the driver evicts all objects selected (drm_mm_scan_remove_block()
675  * reported true) in the scan, and any overlapping nodes after color adjustment
676  * (drm_mm_scan_color_evict()). Adding and removing an object is O(1), and
677  * since freeing a node is also O(1) the overall complexity is
678  * O(scanned_objects). So like the free stack which needs to be walked before a
679  * scan operation even begins this is linear in the number of objects. It
680  * doesn't seem to hurt too badly.
681  */
682 
683 /**
684  * drm_mm_scan_init_with_range - initialize range-restricted lru scanning
685  * @scan: scan state
686  * @mm: drm_mm to scan
687  * @size: size of the allocation
688  * @alignment: alignment of the allocation
689  * @color: opaque tag value to use for the allocation
690  * @start: start of the allowed range for the allocation
691  * @end: end of the allowed range for the allocation
692  * @mode: fine-tune the allocation search and placement
693  *
694  * This simply sets up the scanning routines with the parameters for the desired
695  * hole.
696  *
697  * Warning:
698  * As long as the scan list is non-empty, no other operations than
699  * adding/removing nodes to/from the scan list are allowed.
700  */
drm_mm_scan_init_with_range(struct drm_mm_scan * scan,struct drm_mm * mm,u64 size,u64 alignment,unsigned long color,u64 start,u64 end,enum drm_mm_insert_mode mode)701 void drm_mm_scan_init_with_range(struct drm_mm_scan *scan,
702 				 struct drm_mm *mm,
703 				 u64 size,
704 				 u64 alignment,
705 				 unsigned long color,
706 				 u64 start,
707 				 u64 end,
708 				 enum drm_mm_insert_mode mode)
709 {
710 	DRM_MM_BUG_ON(start >= end);
711 	DRM_MM_BUG_ON(!size || size > end - start);
712 	DRM_MM_BUG_ON(mm->scan_active);
713 
714 	scan->mm = mm;
715 
716 	if (alignment <= 1)
717 		alignment = 0;
718 
719 	scan->color = color;
720 	scan->alignment = alignment;
721 	scan->remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
722 	scan->size = size;
723 	scan->mode = mode;
724 
725 	DRM_MM_BUG_ON(end <= start);
726 	scan->range_start = start;
727 	scan->range_end = end;
728 
729 	scan->hit_start = U64_MAX;
730 	scan->hit_end = 0;
731 }
732 EXPORT_SYMBOL(drm_mm_scan_init_with_range);
733 
734 /**
735  * drm_mm_scan_add_block - add a node to the scan list
736  * @scan: the active drm_mm scanner
737  * @node: drm_mm_node to add
738  *
739  * Add a node to the scan list that might be freed to make space for the desired
740  * hole.
741  *
742  * Returns:
743  * True if a hole has been found, false otherwise.
744  */
drm_mm_scan_add_block(struct drm_mm_scan * scan,struct drm_mm_node * node)745 bool drm_mm_scan_add_block(struct drm_mm_scan *scan,
746 			   struct drm_mm_node *node)
747 {
748 	struct drm_mm *mm = scan->mm;
749 	struct drm_mm_node *hole;
750 	u64 hole_start, hole_end;
751 	u64 col_start, col_end;
752 	u64 adj_start, adj_end;
753 
754 	DRM_MM_BUG_ON(node->mm != mm);
755 	DRM_MM_BUG_ON(!drm_mm_node_allocated(node));
756 	DRM_MM_BUG_ON(drm_mm_node_scanned_block(node));
757 	__set_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
758 	mm->scan_active++;
759 
760 	/* Remove this block from the node_list so that we enlarge the hole
761 	 * (distance between the end of our previous node and the start of
762 	 * or next), without poisoning the link so that we can restore it
763 	 * later in drm_mm_scan_remove_block().
764 	 */
765 	hole = list_prev_entry(node, node_list);
766 	DRM_MM_BUG_ON(list_next_entry(hole, node_list) != node);
767 	__list_del_entry(&node->node_list);
768 
769 	hole_start = __drm_mm_hole_node_start(hole);
770 	hole_end = __drm_mm_hole_node_end(hole);
771 
772 	col_start = hole_start;
773 	col_end = hole_end;
774 	if (mm->color_adjust)
775 		mm->color_adjust(hole, scan->color, &col_start, &col_end);
776 
777 	adj_start = max(col_start, scan->range_start);
778 	adj_end = min(col_end, scan->range_end);
779 	if (adj_end <= adj_start || adj_end - adj_start < scan->size)
780 		return false;
781 
782 	if (scan->mode == DRM_MM_INSERT_HIGH)
783 		adj_start = adj_end - scan->size;
784 
785 	if (scan->alignment) {
786 		u64 rem;
787 
788 		if (likely(scan->remainder_mask))
789 			rem = adj_start & scan->remainder_mask;
790 		else
791 			div64_u64_rem(adj_start, scan->alignment, &rem);
792 		if (rem) {
793 			adj_start -= rem;
794 			if (scan->mode != DRM_MM_INSERT_HIGH)
795 				adj_start += scan->alignment;
796 			if (adj_start < max(col_start, scan->range_start) ||
797 			    min(col_end, scan->range_end) - adj_start < scan->size)
798 				return false;
799 
800 			if (adj_end <= adj_start ||
801 			    adj_end - adj_start < scan->size)
802 				return false;
803 		}
804 	}
805 
806 	scan->hit_start = adj_start;
807 	scan->hit_end = adj_start + scan->size;
808 
809 	DRM_MM_BUG_ON(scan->hit_start >= scan->hit_end);
810 	DRM_MM_BUG_ON(scan->hit_start < hole_start);
811 	DRM_MM_BUG_ON(scan->hit_end > hole_end);
812 
813 	return true;
814 }
815 EXPORT_SYMBOL(drm_mm_scan_add_block);
816 
817 /**
818  * drm_mm_scan_remove_block - remove a node from the scan list
819  * @scan: the active drm_mm scanner
820  * @node: drm_mm_node to remove
821  *
822  * Nodes **must** be removed in exactly the reverse order from the scan list as
823  * they have been added (e.g. using list_add() as they are added and then
824  * list_for_each() over that eviction list to remove), otherwise the internal
825  * state of the memory manager will be corrupted.
826  *
827  * When the scan list is empty, the selected memory nodes can be freed. An
828  * immediately following drm_mm_insert_node_in_range_generic() or one of the
829  * simpler versions of that function with !DRM_MM_SEARCH_BEST will then return
830  * the just freed block (because it's at the top of the free_stack list).
831  *
832  * Returns:
833  * True if this block should be evicted, false otherwise. Will always
834  * return false when no hole has been found.
835  */
drm_mm_scan_remove_block(struct drm_mm_scan * scan,struct drm_mm_node * node)836 bool drm_mm_scan_remove_block(struct drm_mm_scan *scan,
837 			      struct drm_mm_node *node)
838 {
839 	struct drm_mm_node *prev_node;
840 
841 	DRM_MM_BUG_ON(node->mm != scan->mm);
842 	DRM_MM_BUG_ON(!drm_mm_node_scanned_block(node));
843 	__clear_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
844 
845 	DRM_MM_BUG_ON(!node->mm->scan_active);
846 	node->mm->scan_active--;
847 
848 	/* During drm_mm_scan_add_block() we decoupled this node leaving
849 	 * its pointers intact. Now that the caller is walking back along
850 	 * the eviction list we can restore this block into its rightful
851 	 * place on the full node_list. To confirm that the caller is walking
852 	 * backwards correctly we check that prev_node->next == node->next,
853 	 * i.e. both believe the same node should be on the other side of the
854 	 * hole.
855 	 */
856 	prev_node = list_prev_entry(node, node_list);
857 	DRM_MM_BUG_ON(list_next_entry(prev_node, node_list) !=
858 		      list_next_entry(node, node_list));
859 	list_add(&node->node_list, &prev_node->node_list);
860 
861 	return (node->start + node->size > scan->hit_start &&
862 		node->start < scan->hit_end);
863 }
864 EXPORT_SYMBOL(drm_mm_scan_remove_block);
865 
866 /**
867  * drm_mm_scan_color_evict - evict overlapping nodes on either side of hole
868  * @scan: drm_mm scan with target hole
869  *
870  * After completing an eviction scan and removing the selected nodes, we may
871  * need to remove a few more nodes from either side of the target hole if
872  * mm.color_adjust is being used.
873  *
874  * Returns:
875  * A node to evict, or NULL if there are no overlapping nodes.
876  */
drm_mm_scan_color_evict(struct drm_mm_scan * scan)877 struct drm_mm_node *drm_mm_scan_color_evict(struct drm_mm_scan *scan)
878 {
879 	struct drm_mm *mm = scan->mm;
880 	struct drm_mm_node *hole;
881 	u64 hole_start, hole_end;
882 
883 	DRM_MM_BUG_ON(list_empty(&mm->hole_stack));
884 
885 	if (!mm->color_adjust)
886 		return NULL;
887 
888 	/*
889 	 * The hole found during scanning should ideally be the first element
890 	 * in the hole_stack list, but due to side-effects in the driver it
891 	 * may not be.
892 	 */
893 	list_for_each_entry(hole, &mm->hole_stack, hole_stack) {
894 		hole_start = __drm_mm_hole_node_start(hole);
895 		hole_end = hole_start + hole->hole_size;
896 
897 		if (hole_start <= scan->hit_start &&
898 		    hole_end >= scan->hit_end)
899 			break;
900 	}
901 
902 	/* We should only be called after we found the hole previously */
903 	DRM_MM_BUG_ON(&hole->hole_stack == &mm->hole_stack);
904 	if (unlikely(&hole->hole_stack == &mm->hole_stack))
905 		return NULL;
906 
907 	DRM_MM_BUG_ON(hole_start > scan->hit_start);
908 	DRM_MM_BUG_ON(hole_end < scan->hit_end);
909 
910 	mm->color_adjust(hole, scan->color, &hole_start, &hole_end);
911 	if (hole_start > scan->hit_start)
912 		return hole;
913 	if (hole_end < scan->hit_end)
914 		return list_next_entry(hole, node_list);
915 
916 	return NULL;
917 }
918 EXPORT_SYMBOL(drm_mm_scan_color_evict);
919 
920 /**
921  * drm_mm_init - initialize a drm-mm allocator
922  * @mm: the drm_mm structure to initialize
923  * @start: start of the range managed by @mm
924  * @size: end of the range managed by @mm
925  *
926  * Note that @mm must be cleared to 0 before calling this function.
927  */
drm_mm_init(struct drm_mm * mm,u64 start,u64 size)928 void drm_mm_init(struct drm_mm *mm, u64 start, u64 size)
929 {
930 	DRM_MM_BUG_ON(start + size <= start);
931 
932 	mm->color_adjust = NULL;
933 
934 	INIT_LIST_HEAD(&mm->hole_stack);
935 	mm->interval_tree = RB_ROOT_CACHED;
936 	mm->holes_size = RB_ROOT_CACHED;
937 	mm->holes_addr = RB_ROOT;
938 
939 	/* Clever trick to avoid a special case in the free hole tracking. */
940 	INIT_LIST_HEAD(&mm->head_node.node_list);
941 	mm->head_node.flags = 0;
942 	mm->head_node.mm = mm;
943 	mm->head_node.start = start + size;
944 	mm->head_node.size = -size;
945 	add_hole(&mm->head_node);
946 
947 	mm->scan_active = 0;
948 
949 #ifdef CONFIG_DRM_DEBUG_MM
950 	stack_depot_init();
951 #endif
952 }
953 EXPORT_SYMBOL(drm_mm_init);
954 
955 /**
956  * drm_mm_takedown - clean up a drm_mm allocator
957  * @mm: drm_mm allocator to clean up
958  *
959  * Note that it is a bug to call this function on an allocator which is not
960  * clean.
961  */
drm_mm_takedown(struct drm_mm * mm)962 void drm_mm_takedown(struct drm_mm *mm)
963 {
964 	if (WARN(!drm_mm_clean(mm),
965 		 "Memory manager not clean during takedown.\n"))
966 		show_leaks(mm);
967 }
968 EXPORT_SYMBOL(drm_mm_takedown);
969 
drm_mm_dump_hole(struct drm_printer * p,const struct drm_mm_node * entry)970 static u64 drm_mm_dump_hole(struct drm_printer *p, const struct drm_mm_node *entry)
971 {
972 	u64 start, size;
973 
974 	size = entry->hole_size;
975 	if (size) {
976 		start = drm_mm_hole_node_start(entry);
977 		drm_printf(p, "%#018llx-%#018llx: %llu: free\n",
978 			   start, start + size, size);
979 	}
980 
981 	return size;
982 }
983 /**
984  * drm_mm_print - print allocator state
985  * @mm: drm_mm allocator to print
986  * @p: DRM printer to use
987  */
drm_mm_print(const struct drm_mm * mm,struct drm_printer * p)988 void drm_mm_print(const struct drm_mm *mm, struct drm_printer *p)
989 {
990 	const struct drm_mm_node *entry;
991 	u64 total_used = 0, total_free = 0, total = 0;
992 
993 	total_free += drm_mm_dump_hole(p, &mm->head_node);
994 
995 	drm_mm_for_each_node(entry, mm) {
996 		drm_printf(p, "%#018llx-%#018llx: %llu: used\n", entry->start,
997 			   entry->start + entry->size, entry->size);
998 		total_used += entry->size;
999 		total_free += drm_mm_dump_hole(p, entry);
1000 	}
1001 	total = total_free + total_used;
1002 
1003 	drm_printf(p, "total: %llu, used %llu free %llu\n", total,
1004 		   total_used, total_free);
1005 }
1006 EXPORT_SYMBOL(drm_mm_print);
1007