Linux Audio

Check our new training course

Loading...
v5.4
  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
107static 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
118static void show_leaks(struct drm_mm *mm)
119{
120	struct drm_mm_node *node;
121	unsigned long *entries;
122	unsigned int nr_entries;
123	char *buf;
124
125	buf = kmalloc(BUFSZ, GFP_KERNEL);
126	if (!buf)
127		return;
128
129	list_for_each_entry(node, drm_mm_nodes(mm), node_list) {
 
 
 
 
 
130		if (!node->stack) {
131			DRM_ERROR("node [%08llx + %08llx]: unknown owner\n",
132				  node->start, node->size);
133			continue;
134		}
135
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);
140	}
141
142	kfree(buf);
143}
144
145#undef STACKDEPTH
146#undef BUFSZ
147#else
148static void save_stack(struct drm_mm_node *node) { }
149static void show_leaks(struct drm_mm *mm) { }
150#endif
151
152#define START(node) ((node)->start)
153#define LAST(node)  ((node)->start + (node)->size - 1)
154
155INTERVAL_TREE_DEFINE(struct drm_mm_node, rb,
156		     u64, __subtree_last,
157		     START, LAST, static inline, drm_mm_interval_tree)
158
159struct drm_mm_node *
160__drm_mm_interval_first(const struct drm_mm *mm, u64 start, u64 last)
161{
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;
164}
165EXPORT_SYMBOL(__drm_mm_interval_first);
166
167static void drm_mm_interval_tree_add_node(struct drm_mm_node *hole_node,
168					  struct drm_mm_node *node)
169{
170	struct drm_mm *mm = hole_node->mm;
171	struct rb_node **link, *rb;
172	struct drm_mm_node *parent;
173	bool leftmost;
174
175	node->__subtree_last = LAST(node);
176
177	if (hole_node->allocated) {
178		rb = &hole_node->rb;
179		while (rb) {
180			parent = rb_entry(rb, struct drm_mm_node, rb);
181			if (parent->__subtree_last >= node->__subtree_last)
182				break;
183
184			parent->__subtree_last = node->__subtree_last;
185			rb = rb_parent(rb);
186		}
187
188		rb = &hole_node->rb;
189		link = &hole_node->rb.rb_right;
190		leftmost = false;
191	} else {
192		rb = NULL;
193		link = &mm->interval_tree.rb_root.rb_node;
194		leftmost = true;
195	}
196
197	while (*link) {
198		rb = *link;
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;
204		} else {
205			link = &parent->rb.rb_right;
206			leftmost = false;
207		}
208	}
209
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);
213}
214
215#define RB_INSERT(root, member, expr) do { \
216	struct rb_node **link = &root.rb_node, *rb = NULL; \
217	u64 x = expr(node); \
218	while (*link) { \
219		rb = *link; \
220		if (x < expr(rb_entry(rb, struct drm_mm_node, member))) \
221			link = &rb->rb_left; \
222		else \
223			link = &rb->rb_right; \
224	} \
225	rb_link_node(&node->member, rb, link); \
226	rb_insert_color(&node->member, &root); \
227} while (0)
228
229#define HOLE_SIZE(NODE) ((NODE)->hole_size)
230#define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE))
231
232static u64 rb_to_hole_size(struct rb_node *rb)
233{
234	return rb_entry(rb, struct drm_mm_node, rb_hole_size)->hole_size;
235}
236
237static void insert_hole_size(struct rb_root_cached *root,
238			     struct drm_mm_node *node)
239{
240	struct rb_node **link = &root->rb_root.rb_node, *rb = NULL;
241	u64 x = node->hole_size;
242	bool first = true;
243
244	while (*link) {
245		rb = *link;
246		if (x > rb_to_hole_size(rb)) {
247			link = &rb->rb_left;
248		} else {
249			link = &rb->rb_right;
250			first = false;
251		}
252	}
253
254	rb_link_node(&node->rb_hole_size, rb, link);
255	rb_insert_color_cached(&node->rb_hole_size, root, first);
256}
257
258static void add_hole(struct drm_mm_node *node)
259{
260	struct drm_mm *mm = node->mm;
261
262	node->hole_size =
263		__drm_mm_hole_node_end(node) - __drm_mm_hole_node_start(node);
264	DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
265
266	insert_hole_size(&mm->holes_size, node);
267	RB_INSERT(mm->holes_addr, rb_hole_addr, HOLE_ADDR);
268
269	list_add(&node->hole_stack, &mm->hole_stack);
270}
271
272static void rm_hole(struct drm_mm_node *node)
273{
274	DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
275
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);
279	node->hole_size = 0;
280
281	DRM_MM_BUG_ON(drm_mm_hole_follows(node));
282}
283
284static inline struct drm_mm_node *rb_hole_size_to_node(struct rb_node *rb)
285{
286	return rb_entry_safe(rb, struct drm_mm_node, rb_hole_size);
287}
288
289static inline struct drm_mm_node *rb_hole_addr_to_node(struct rb_node *rb)
290{
291	return rb_entry_safe(rb, struct drm_mm_node, rb_hole_addr);
292}
293
294static inline u64 rb_hole_size(struct rb_node *rb)
295{
296	return rb_entry(rb, struct drm_mm_node, rb_hole_size)->hole_size;
297}
298
299static struct drm_mm_node *best_hole(struct drm_mm *mm, u64 size)
300{
301	struct rb_node *rb = mm->holes_size.rb_root.rb_node;
302	struct drm_mm_node *best = NULL;
 
 
 
303
304	do {
305		struct drm_mm_node *node =
306			rb_entry(rb, struct drm_mm_node, rb_hole_size);
307
308		if (size <= node->hole_size) {
309			best = node;
310			rb = rb->rb_right;
311		} else {
312			rb = rb->rb_left;
313		}
314	} while (rb);
315
316	return best;
317}
318
319static struct drm_mm_node *find_hole(struct drm_mm *mm, u64 addr)
320{
321	struct rb_node *rb = mm->holes_addr.rb_node;
322	struct drm_mm_node *node = NULL;
 
323
324	while (rb) {
325		u64 hole_start;
326
327		node = rb_hole_addr_to_node(rb);
328		hole_start = __drm_mm_hole_node_start(node);
329
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;
334		else
335			break;
336	}
337
338	return node;
339}
340
341static struct drm_mm_node *
342first_hole(struct drm_mm *mm,
343	   u64 start, u64 end, u64 size,
344	   enum drm_mm_insert_mode mode)
345{
 
 
 
346	switch (mode) {
347	default:
348	case DRM_MM_INSERT_BEST:
349		return best_hole(mm, size);
350
351	case DRM_MM_INSERT_LOW:
352		return find_hole(mm, start);
353
354	case DRM_MM_INSERT_HIGH:
355		return find_hole(mm, end);
356
357	case DRM_MM_INSERT_EVICT:
358		return list_first_entry_or_null(&mm->hole_stack,
359						struct drm_mm_node,
360						hole_stack);
361	}
362}
363
364static struct drm_mm_node *
365next_hole(struct drm_mm *mm,
366	  struct drm_mm_node *node,
367	  enum drm_mm_insert_mode mode)
368{
369	switch (mode) {
370	default:
371	case DRM_MM_INSERT_BEST:
372		return rb_hole_size_to_node(rb_prev(&node->rb_hole_size));
373
374	case DRM_MM_INSERT_LOW:
375		return rb_hole_addr_to_node(rb_next(&node->rb_hole_addr));
376
377	case DRM_MM_INSERT_HIGH:
378		return rb_hole_addr_to_node(rb_prev(&node->rb_hole_addr));
379
380	case DRM_MM_INSERT_EVICT:
381		node = list_next_entry(node, hole_stack);
382		return &node->hole_stack == &mm->hole_stack ? NULL : node;
383	}
384}
385
386/**
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
390 *
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.
396 *
397 * Returns:
398 * 0 on success, -ENOSPC if there's no hole where @node is.
399 */
400int drm_mm_reserve_node(struct drm_mm *mm, struct drm_mm_node *node)
401{
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;
406
407	end = node->start + node->size;
408	if (unlikely(end <= node->start))
409		return -ENOSPC;
410
411	/* Find the relevant hole to add our node to */
412	hole = find_hole(mm, node->start);
413	if (!hole)
414		return -ENOSPC;
415
416	adj_start = hole_start = __drm_mm_hole_node_start(hole);
417	adj_end = hole_end = hole_start + hole->hole_size;
418
419	if (mm->color_adjust)
420		mm->color_adjust(hole, node->color, &adj_start, &adj_end);
421
422	if (adj_start > node->start || adj_end < end)
423		return -ENOSPC;
424
425	node->mm = mm;
426
427	list_add(&node->node_list, &hole->node_list);
428	drm_mm_interval_tree_add_node(hole, node);
429	node->allocated = true;
430	node->hole_size = 0;
431
432	rm_hole(hole);
433	if (node->start > hole_start)
434		add_hole(hole);
435	if (end < hole_end)
436		add_hole(node);
437
438	save_stack(node);
439	return 0;
440}
441EXPORT_SYMBOL(drm_mm_reserve_node);
442
443static u64 rb_to_hole_size_or_zero(struct rb_node *rb)
444{
445	return rb ? rb_to_hole_size(rb) : 0;
446}
447
448/**
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
458 *
459 * The preallocated @node must be cleared to 0.
460 *
461 * Returns:
462 * 0 on success, -ENOSPC if there's no suitable hole.
463 */
464int drm_mm_insert_node_in_range(struct drm_mm * const mm,
465				struct drm_mm_node * const node,
466				u64 size, u64 alignment,
467				unsigned long color,
468				u64 range_start, u64 range_end,
469				enum drm_mm_insert_mode mode)
470{
471	struct drm_mm_node *hole;
472	u64 remainder_mask;
473	bool once;
474
475	DRM_MM_BUG_ON(range_start > range_end);
476
477	if (unlikely(size == 0 || range_end - range_start < size))
478		return -ENOSPC;
479
480	if (rb_to_hole_size_or_zero(rb_first_cached(&mm->holes_size)) < size)
481		return -ENOSPC;
482
483	if (alignment <= 1)
484		alignment = 0;
485
486	once = mode & DRM_MM_INSERT_ONCE;
487	mode &= ~DRM_MM_INSERT_ONCE;
488
489	remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
490	for (hole = first_hole(mm, range_start, range_end, size, mode);
491	     hole;
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;
497
498		if (mode == DRM_MM_INSERT_LOW && hole_start >= range_end)
499			break;
500
501		if (mode == DRM_MM_INSERT_HIGH && hole_end <= range_start)
502			break;
503
504		col_start = hole_start;
505		col_end = hole_end;
506		if (mm->color_adjust)
507			mm->color_adjust(hole, color, &col_start, &col_end);
508
509		adj_start = max(col_start, range_start);
510		adj_end = min(col_end, range_end);
511
512		if (adj_end <= adj_start || adj_end - adj_start < size)
513			continue;
514
515		if (mode == DRM_MM_INSERT_HIGH)
516			adj_start = adj_end - size;
517
518		if (alignment) {
519			u64 rem;
520
521			if (likely(remainder_mask))
522				rem = adj_start & remainder_mask;
523			else
524				div64_u64_rem(adj_start, alignment, &rem);
525			if (rem) {
526				adj_start -= rem;
527				if (mode != DRM_MM_INSERT_HIGH)
528					adj_start += alignment;
529
530				if (adj_start < max(col_start, range_start) ||
531				    min(col_end, range_end) - adj_start < size)
532					continue;
533
534				if (adj_end <= adj_start ||
535				    adj_end - adj_start < size)
536					continue;
537			}
538		}
539
540		node->mm = mm;
541		node->size = size;
542		node->start = adj_start;
543		node->color = color;
544		node->hole_size = 0;
545
546		list_add(&node->node_list, &hole->node_list);
547		drm_mm_interval_tree_add_node(hole, node);
548		node->allocated = true;
549
550		rm_hole(hole);
551		if (adj_start > hole_start)
552			add_hole(hole);
553		if (adj_start + size < hole_end)
554			add_hole(node);
555
556		save_stack(node);
557		return 0;
558	}
559
560	return -ENOSPC;
561}
562EXPORT_SYMBOL(drm_mm_insert_node_in_range);
563
564/**
565 * drm_mm_remove_node - Remove a memory node from the allocator.
566 * @node: drm_mm_node to remove
567 *
568 * This just removes a node from its drm_mm allocator. The node does not need to
569 * be cleared again before it can be re-inserted into this or any other drm_mm
570 * allocator. It is a bug to call this function on a unallocated node.
571 */
572void drm_mm_remove_node(struct drm_mm_node *node)
573{
574	struct drm_mm *mm = node->mm;
575	struct drm_mm_node *prev_node;
576
577	DRM_MM_BUG_ON(!node->allocated);
578	DRM_MM_BUG_ON(node->scanned_block);
579
580	prev_node = list_prev_entry(node, node_list);
581
582	if (drm_mm_hole_follows(node))
583		rm_hole(node);
584
585	drm_mm_interval_tree_remove(node, &mm->interval_tree);
586	list_del(&node->node_list);
587	node->allocated = false;
588
589	if (drm_mm_hole_follows(prev_node))
590		rm_hole(prev_node);
591	add_hole(prev_node);
592}
593EXPORT_SYMBOL(drm_mm_remove_node);
594
595/**
596 * drm_mm_replace_node - move an allocation from @old to @new
597 * @old: drm_mm_node to remove from the allocator
598 * @new: drm_mm_node which should inherit @old's allocation
599 *
600 * This is useful for when drivers embed the drm_mm_node structure and hence
601 * can't move allocations by reassigning pointers. It's a combination of remove
602 * and insert with the guarantee that the allocation start will match.
603 */
604void drm_mm_replace_node(struct drm_mm_node *old, struct drm_mm_node *new)
605{
606	struct drm_mm *mm = old->mm;
607
608	DRM_MM_BUG_ON(!old->allocated);
609
610	*new = *old;
611
612	list_replace(&old->node_list, &new->node_list);
613	rb_replace_node_cached(&old->rb, &new->rb, &mm->interval_tree);
614
615	if (drm_mm_hole_follows(old)) {
616		list_replace(&old->hole_stack, &new->hole_stack);
617		rb_replace_node_cached(&old->rb_hole_size,
618				       &new->rb_hole_size,
619				       &mm->holes_size);
620		rb_replace_node(&old->rb_hole_addr,
621				&new->rb_hole_addr,
622				&mm->holes_addr);
623	}
624
625	old->allocated = false;
626	new->allocated = true;
627}
628EXPORT_SYMBOL(drm_mm_replace_node);
629
630/**
631 * DOC: lru scan roster
632 *
633 * Very often GPUs need to have continuous allocations for a given object. When
634 * evicting objects to make space for a new one it is therefore not most
635 * efficient when we simply start to select all objects from the tail of an LRU
636 * until there's a suitable hole: Especially for big objects or nodes that
637 * otherwise have special allocation constraints there's a good chance we evict
638 * lots of (smaller) objects unnecessarily.
639 *
640 * The DRM range allocator supports this use-case through the scanning
641 * interfaces. First a scan operation needs to be initialized with
642 * drm_mm_scan_init() or drm_mm_scan_init_with_range(). The driver adds
643 * objects to the roster, probably by walking an LRU list, but this can be
644 * freely implemented. Eviction candiates are added using
645 * drm_mm_scan_add_block() until a suitable hole is found or there are no
646 * further evictable objects. Eviction roster metadata is tracked in &struct
647 * drm_mm_scan.
648 *
649 * The driver must walk through all objects again in exactly the reverse
650 * order to restore the allocator state. Note that while the allocator is used
651 * in the scan mode no other operation is allowed.
652 *
653 * Finally the driver evicts all objects selected (drm_mm_scan_remove_block()
654 * reported true) in the scan, and any overlapping nodes after color adjustment
655 * (drm_mm_scan_color_evict()). Adding and removing an object is O(1), and
656 * since freeing a node is also O(1) the overall complexity is
657 * O(scanned_objects). So like the free stack which needs to be walked before a
658 * scan operation even begins this is linear in the number of objects. It
659 * doesn't seem to hurt too badly.
660 */
661
662/**
663 * drm_mm_scan_init_with_range - initialize range-restricted lru scanning
664 * @scan: scan state
665 * @mm: drm_mm to scan
666 * @size: size of the allocation
667 * @alignment: alignment of the allocation
668 * @color: opaque tag value to use for the allocation
669 * @start: start of the allowed range for the allocation
670 * @end: end of the allowed range for the allocation
671 * @mode: fine-tune the allocation search and placement
672 *
673 * This simply sets up the scanning routines with the parameters for the desired
674 * hole.
675 *
676 * Warning:
677 * As long as the scan list is non-empty, no other operations than
678 * adding/removing nodes to/from the scan list are allowed.
679 */
680void drm_mm_scan_init_with_range(struct drm_mm_scan *scan,
681				 struct drm_mm *mm,
682				 u64 size,
683				 u64 alignment,
684				 unsigned long color,
685				 u64 start,
686				 u64 end,
687				 enum drm_mm_insert_mode mode)
688{
689	DRM_MM_BUG_ON(start >= end);
690	DRM_MM_BUG_ON(!size || size > end - start);
691	DRM_MM_BUG_ON(mm->scan_active);
692
693	scan->mm = mm;
694
695	if (alignment <= 1)
696		alignment = 0;
697
698	scan->color = color;
699	scan->alignment = alignment;
700	scan->remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
701	scan->size = size;
702	scan->mode = mode;
703
704	DRM_MM_BUG_ON(end <= start);
705	scan->range_start = start;
706	scan->range_end = end;
707
708	scan->hit_start = U64_MAX;
709	scan->hit_end = 0;
710}
711EXPORT_SYMBOL(drm_mm_scan_init_with_range);
712
713/**
714 * drm_mm_scan_add_block - add a node to the scan list
715 * @scan: the active drm_mm scanner
716 * @node: drm_mm_node to add
717 *
718 * Add a node to the scan list that might be freed to make space for the desired
719 * hole.
720 *
721 * Returns:
722 * True if a hole has been found, false otherwise.
723 */
724bool drm_mm_scan_add_block(struct drm_mm_scan *scan,
725			   struct drm_mm_node *node)
726{
727	struct drm_mm *mm = scan->mm;
728	struct drm_mm_node *hole;
729	u64 hole_start, hole_end;
730	u64 col_start, col_end;
731	u64 adj_start, adj_end;
732
733	DRM_MM_BUG_ON(node->mm != mm);
734	DRM_MM_BUG_ON(!node->allocated);
735	DRM_MM_BUG_ON(node->scanned_block);
736	node->scanned_block = true;
737	mm->scan_active++;
738
739	/* Remove this block from the node_list so that we enlarge the hole
740	 * (distance between the end of our previous node and the start of
741	 * or next), without poisoning the link so that we can restore it
742	 * later in drm_mm_scan_remove_block().
743	 */
744	hole = list_prev_entry(node, node_list);
745	DRM_MM_BUG_ON(list_next_entry(hole, node_list) != node);
746	__list_del_entry(&node->node_list);
747
748	hole_start = __drm_mm_hole_node_start(hole);
749	hole_end = __drm_mm_hole_node_end(hole);
750
751	col_start = hole_start;
752	col_end = hole_end;
753	if (mm->color_adjust)
754		mm->color_adjust(hole, scan->color, &col_start, &col_end);
755
756	adj_start = max(col_start, scan->range_start);
757	adj_end = min(col_end, scan->range_end);
758	if (adj_end <= adj_start || adj_end - adj_start < scan->size)
759		return false;
760
761	if (scan->mode == DRM_MM_INSERT_HIGH)
762		adj_start = adj_end - scan->size;
763
764	if (scan->alignment) {
765		u64 rem;
766
767		if (likely(scan->remainder_mask))
768			rem = adj_start & scan->remainder_mask;
769		else
770			div64_u64_rem(adj_start, scan->alignment, &rem);
771		if (rem) {
772			adj_start -= rem;
773			if (scan->mode != DRM_MM_INSERT_HIGH)
774				adj_start += scan->alignment;
775			if (adj_start < max(col_start, scan->range_start) ||
776			    min(col_end, scan->range_end) - adj_start < scan->size)
777				return false;
778
779			if (adj_end <= adj_start ||
780			    adj_end - adj_start < scan->size)
781				return false;
782		}
783	}
784
785	scan->hit_start = adj_start;
786	scan->hit_end = adj_start + scan->size;
787
788	DRM_MM_BUG_ON(scan->hit_start >= scan->hit_end);
789	DRM_MM_BUG_ON(scan->hit_start < hole_start);
790	DRM_MM_BUG_ON(scan->hit_end > hole_end);
791
792	return true;
793}
794EXPORT_SYMBOL(drm_mm_scan_add_block);
795
796/**
797 * drm_mm_scan_remove_block - remove a node from the scan list
798 * @scan: the active drm_mm scanner
799 * @node: drm_mm_node to remove
800 *
801 * Nodes **must** be removed in exactly the reverse order from the scan list as
802 * they have been added (e.g. using list_add() as they are added and then
803 * list_for_each() over that eviction list to remove), otherwise the internal
804 * state of the memory manager will be corrupted.
805 *
806 * When the scan list is empty, the selected memory nodes can be freed. An
807 * immediately following drm_mm_insert_node_in_range_generic() or one of the
808 * simpler versions of that function with !DRM_MM_SEARCH_BEST will then return
809 * the just freed block (because it's at the top of the free_stack list).
810 *
811 * Returns:
812 * True if this block should be evicted, false otherwise. Will always
813 * return false when no hole has been found.
814 */
815bool drm_mm_scan_remove_block(struct drm_mm_scan *scan,
816			      struct drm_mm_node *node)
817{
818	struct drm_mm_node *prev_node;
819
820	DRM_MM_BUG_ON(node->mm != scan->mm);
821	DRM_MM_BUG_ON(!node->scanned_block);
822	node->scanned_block = false;
823
824	DRM_MM_BUG_ON(!node->mm->scan_active);
825	node->mm->scan_active--;
826
827	/* During drm_mm_scan_add_block() we decoupled this node leaving
828	 * its pointers intact. Now that the caller is walking back along
829	 * the eviction list we can restore this block into its rightful
830	 * place on the full node_list. To confirm that the caller is walking
831	 * backwards correctly we check that prev_node->next == node->next,
832	 * i.e. both believe the same node should be on the other side of the
833	 * hole.
834	 */
835	prev_node = list_prev_entry(node, node_list);
836	DRM_MM_BUG_ON(list_next_entry(prev_node, node_list) !=
837		      list_next_entry(node, node_list));
838	list_add(&node->node_list, &prev_node->node_list);
839
840	return (node->start + node->size > scan->hit_start &&
841		node->start < scan->hit_end);
842}
843EXPORT_SYMBOL(drm_mm_scan_remove_block);
844
845/**
846 * drm_mm_scan_color_evict - evict overlapping nodes on either side of hole
847 * @scan: drm_mm scan with target hole
848 *
849 * After completing an eviction scan and removing the selected nodes, we may
850 * need to remove a few more nodes from either side of the target hole if
851 * mm.color_adjust is being used.
852 *
853 * Returns:
854 * A node to evict, or NULL if there are no overlapping nodes.
855 */
856struct drm_mm_node *drm_mm_scan_color_evict(struct drm_mm_scan *scan)
857{
858	struct drm_mm *mm = scan->mm;
859	struct drm_mm_node *hole;
860	u64 hole_start, hole_end;
861
862	DRM_MM_BUG_ON(list_empty(&mm->hole_stack));
863
864	if (!mm->color_adjust)
865		return NULL;
866
867	/*
868	 * The hole found during scanning should ideally be the first element
869	 * in the hole_stack list, but due to side-effects in the driver it
870	 * may not be.
871	 */
872	list_for_each_entry(hole, &mm->hole_stack, hole_stack) {
873		hole_start = __drm_mm_hole_node_start(hole);
874		hole_end = hole_start + hole->hole_size;
875
876		if (hole_start <= scan->hit_start &&
877		    hole_end >= scan->hit_end)
878			break;
879	}
880
881	/* We should only be called after we found the hole previously */
882	DRM_MM_BUG_ON(&hole->hole_stack == &mm->hole_stack);
883	if (unlikely(&hole->hole_stack == &mm->hole_stack))
884		return NULL;
885
886	DRM_MM_BUG_ON(hole_start > scan->hit_start);
887	DRM_MM_BUG_ON(hole_end < scan->hit_end);
888
889	mm->color_adjust(hole, scan->color, &hole_start, &hole_end);
890	if (hole_start > scan->hit_start)
891		return hole;
892	if (hole_end < scan->hit_end)
893		return list_next_entry(hole, node_list);
894
895	return NULL;
896}
897EXPORT_SYMBOL(drm_mm_scan_color_evict);
898
899/**
900 * drm_mm_init - initialize a drm-mm allocator
901 * @mm: the drm_mm structure to initialize
902 * @start: start of the range managed by @mm
903 * @size: end of the range managed by @mm
904 *
905 * Note that @mm must be cleared to 0 before calling this function.
906 */
907void drm_mm_init(struct drm_mm *mm, u64 start, u64 size)
908{
909	DRM_MM_BUG_ON(start + size <= start);
910
911	mm->color_adjust = NULL;
912
913	INIT_LIST_HEAD(&mm->hole_stack);
914	mm->interval_tree = RB_ROOT_CACHED;
915	mm->holes_size = RB_ROOT_CACHED;
916	mm->holes_addr = RB_ROOT;
917
918	/* Clever trick to avoid a special case in the free hole tracking. */
919	INIT_LIST_HEAD(&mm->head_node.node_list);
920	mm->head_node.allocated = false;
921	mm->head_node.mm = mm;
922	mm->head_node.start = start + size;
923	mm->head_node.size = -size;
924	add_hole(&mm->head_node);
925
926	mm->scan_active = 0;
927}
928EXPORT_SYMBOL(drm_mm_init);
929
930/**
931 * drm_mm_takedown - clean up a drm_mm allocator
932 * @mm: drm_mm allocator to clean up
933 *
934 * Note that it is a bug to call this function on an allocator which is not
935 * clean.
936 */
937void drm_mm_takedown(struct drm_mm *mm)
938{
939	if (WARN(!drm_mm_clean(mm),
940		 "Memory manager not clean during takedown.\n"))
941		show_leaks(mm);
942}
943EXPORT_SYMBOL(drm_mm_takedown);
944
945static u64 drm_mm_dump_hole(struct drm_printer *p, const struct drm_mm_node *entry)
946{
947	u64 start, size;
948
949	size = entry->hole_size;
950	if (size) {
951		start = drm_mm_hole_node_start(entry);
952		drm_printf(p, "%#018llx-%#018llx: %llu: free\n",
953			   start, start + size, size);
954	}
955
956	return size;
957}
958/**
959 * drm_mm_print - print allocator state
960 * @mm: drm_mm allocator to print
961 * @p: DRM printer to use
962 */
963void drm_mm_print(const struct drm_mm *mm, struct drm_printer *p)
964{
965	const struct drm_mm_node *entry;
966	u64 total_used = 0, total_free = 0, total = 0;
967
968	total_free += drm_mm_dump_hole(p, &mm->head_node);
969
970	drm_mm_for_each_node(entry, mm) {
971		drm_printf(p, "%#018llx-%#018llx: %llu: used\n", entry->start,
972			   entry->start + entry->size, entry->size);
973		total_used += entry->size;
974		total_free += drm_mm_dump_hole(p, entry);
975	}
976	total = total_free + total_used;
977
978	drm_printf(p, "total: %llu, used %llu free %llu\n", total,
979		   total_used, total_free);
980}
981EXPORT_SYMBOL(drm_mm_print);
v4.17
  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 <drm/drmP.h>
 46#include <drm/drm_mm.h>
 47#include <linux/slab.h>
 48#include <linux/seq_file.h>
 49#include <linux/export.h>
 50#include <linux/interval_tree_generic.h>
 
 
 
 
 
 51
 52/**
 53 * DOC: Overview
 54 *
 55 * drm_mm provides a simple range allocator. The drivers are free to use the
 56 * resource allocator from the linux core if it suits them, the upside of drm_mm
 57 * is that it's in the DRM core. Which means that it's easier to extend for
 58 * some of the crazier special purpose needs of gpus.
 59 *
 60 * The main data struct is &drm_mm, allocations are tracked in &drm_mm_node.
 61 * Drivers are free to embed either of them into their own suitable
 62 * datastructures. drm_mm itself will not do any memory allocations of its own,
 63 * so if drivers choose not to embed nodes they need to still allocate them
 64 * themselves.
 65 *
 66 * The range allocator also supports reservation of preallocated blocks. This is
 67 * useful for taking over initial mode setting configurations from the firmware,
 68 * where an object needs to be created which exactly matches the firmware's
 69 * scanout target. As long as the range is still free it can be inserted anytime
 70 * after the allocator is initialized, which helps with avoiding looped
 71 * dependencies in the driver load sequence.
 72 *
 73 * drm_mm maintains a stack of most recently freed holes, which of all
 74 * simplistic datastructures seems to be a fairly decent approach to clustering
 75 * allocations and avoiding too much fragmentation. This means free space
 76 * searches are O(num_holes). Given that all the fancy features drm_mm supports
 77 * something better would be fairly complex and since gfx thrashing is a fairly
 78 * steep cliff not a real concern. Removing a node again is O(1).
 79 *
 80 * drm_mm supports a few features: Alignment and range restrictions can be
 81 * supplied. Furthermore every &drm_mm_node has a color value (which is just an
 82 * opaque unsigned long) which in conjunction with a driver callback can be used
 83 * to implement sophisticated placement restrictions. The i915 DRM driver uses
 84 * this to implement guard pages between incompatible caching domains in the
 85 * graphics TT.
 86 *
 87 * Two behaviors are supported for searching and allocating: bottom-up and
 88 * top-down. The default is bottom-up. Top-down allocation can be used if the
 89 * memory area has different restrictions, or just to reduce fragmentation.
 90 *
 91 * Finally iteration helpers to walk all nodes and all holes are provided as are
 92 * some basic allocator dumpers for debugging.
 93 *
 94 * Note that this range allocator is not thread-safe, drivers need to protect
 95 * modifications with their own locking. The idea behind this is that for a full
 96 * memory manager additional data needs to be protected anyway, hence internal
 97 * locking would be fully redundant.
 98 */
 99
100#ifdef CONFIG_DRM_DEBUG_MM
101#include <linux/stackdepot.h>
102
103#define STACKDEPTH 32
104#define BUFSZ 4096
105
106static noinline void save_stack(struct drm_mm_node *node)
107{
108	unsigned long entries[STACKDEPTH];
109	struct stack_trace trace = {
110		.entries = entries,
111		.max_entries = STACKDEPTH,
112		.skip = 1
113	};
114
115	save_stack_trace(&trace);
116	if (trace.nr_entries != 0 &&
117	    trace.entries[trace.nr_entries-1] == ULONG_MAX)
118		trace.nr_entries--;
119
120	/* May be called under spinlock, so avoid sleeping */
121	node->stack = depot_save_stack(&trace, GFP_NOWAIT);
122}
123
124static void show_leaks(struct drm_mm *mm)
125{
126	struct drm_mm_node *node;
127	unsigned long entries[STACKDEPTH];
 
128	char *buf;
129
130	buf = kmalloc(BUFSZ, GFP_KERNEL);
131	if (!buf)
132		return;
133
134	list_for_each_entry(node, drm_mm_nodes(mm), node_list) {
135		struct stack_trace trace = {
136			.entries = entries,
137			.max_entries = STACKDEPTH
138		};
139
140		if (!node->stack) {
141			DRM_ERROR("node [%08llx + %08llx]: unknown owner\n",
142				  node->start, node->size);
143			continue;
144		}
145
146		depot_fetch_stack(node->stack, &trace);
147		snprint_stack_trace(buf, BUFSZ, &trace, 0);
148		DRM_ERROR("node [%08llx + %08llx]: inserted at\n%s",
149			  node->start, node->size, buf);
150	}
151
152	kfree(buf);
153}
154
155#undef STACKDEPTH
156#undef BUFSZ
157#else
158static void save_stack(struct drm_mm_node *node) { }
159static void show_leaks(struct drm_mm *mm) { }
160#endif
161
162#define START(node) ((node)->start)
163#define LAST(node)  ((node)->start + (node)->size - 1)
164
165INTERVAL_TREE_DEFINE(struct drm_mm_node, rb,
166		     u64, __subtree_last,
167		     START, LAST, static inline, drm_mm_interval_tree)
168
169struct drm_mm_node *
170__drm_mm_interval_first(const struct drm_mm *mm, u64 start, u64 last)
171{
172	return drm_mm_interval_tree_iter_first((struct rb_root_cached *)&mm->interval_tree,
173					       start, last) ?: (struct drm_mm_node *)&mm->head_node;
174}
175EXPORT_SYMBOL(__drm_mm_interval_first);
176
177static void drm_mm_interval_tree_add_node(struct drm_mm_node *hole_node,
178					  struct drm_mm_node *node)
179{
180	struct drm_mm *mm = hole_node->mm;
181	struct rb_node **link, *rb;
182	struct drm_mm_node *parent;
183	bool leftmost;
184
185	node->__subtree_last = LAST(node);
186
187	if (hole_node->allocated) {
188		rb = &hole_node->rb;
189		while (rb) {
190			parent = rb_entry(rb, struct drm_mm_node, rb);
191			if (parent->__subtree_last >= node->__subtree_last)
192				break;
193
194			parent->__subtree_last = node->__subtree_last;
195			rb = rb_parent(rb);
196		}
197
198		rb = &hole_node->rb;
199		link = &hole_node->rb.rb_right;
200		leftmost = false;
201	} else {
202		rb = NULL;
203		link = &mm->interval_tree.rb_root.rb_node;
204		leftmost = true;
205	}
206
207	while (*link) {
208		rb = *link;
209		parent = rb_entry(rb, struct drm_mm_node, rb);
210		if (parent->__subtree_last < node->__subtree_last)
211			parent->__subtree_last = node->__subtree_last;
212		if (node->start < parent->start) {
213			link = &parent->rb.rb_left;
214		} else {
215			link = &parent->rb.rb_right;
216			leftmost = false;
217		}
218	}
219
220	rb_link_node(&node->rb, rb, link);
221	rb_insert_augmented_cached(&node->rb, &mm->interval_tree, leftmost,
222				   &drm_mm_interval_tree_augment);
223}
224
225#define RB_INSERT(root, member, expr) do { \
226	struct rb_node **link = &root.rb_node, *rb = NULL; \
227	u64 x = expr(node); \
228	while (*link) { \
229		rb = *link; \
230		if (x < expr(rb_entry(rb, struct drm_mm_node, member))) \
231			link = &rb->rb_left; \
232		else \
233			link = &rb->rb_right; \
234	} \
235	rb_link_node(&node->member, rb, link); \
236	rb_insert_color(&node->member, &root); \
237} while (0)
238
239#define HOLE_SIZE(NODE) ((NODE)->hole_size)
240#define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE))
241
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
242static void add_hole(struct drm_mm_node *node)
243{
244	struct drm_mm *mm = node->mm;
245
246	node->hole_size =
247		__drm_mm_hole_node_end(node) - __drm_mm_hole_node_start(node);
248	DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
249
250	RB_INSERT(mm->holes_size, rb_hole_size, HOLE_SIZE);
251	RB_INSERT(mm->holes_addr, rb_hole_addr, HOLE_ADDR);
252
253	list_add(&node->hole_stack, &mm->hole_stack);
254}
255
256static void rm_hole(struct drm_mm_node *node)
257{
258	DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
259
260	list_del(&node->hole_stack);
261	rb_erase(&node->rb_hole_size, &node->mm->holes_size);
262	rb_erase(&node->rb_hole_addr, &node->mm->holes_addr);
263	node->hole_size = 0;
264
265	DRM_MM_BUG_ON(drm_mm_hole_follows(node));
266}
267
268static inline struct drm_mm_node *rb_hole_size_to_node(struct rb_node *rb)
269{
270	return rb_entry_safe(rb, struct drm_mm_node, rb_hole_size);
271}
272
273static inline struct drm_mm_node *rb_hole_addr_to_node(struct rb_node *rb)
274{
275	return rb_entry_safe(rb, struct drm_mm_node, rb_hole_addr);
276}
277
278static inline u64 rb_hole_size(struct rb_node *rb)
279{
280	return rb_entry(rb, struct drm_mm_node, rb_hole_size)->hole_size;
281}
282
283static struct drm_mm_node *best_hole(struct drm_mm *mm, u64 size)
284{
285	struct rb_node *best = NULL;
286	struct rb_node **link = &mm->holes_size.rb_node;
287
288	while (*link) {
289		struct rb_node *rb = *link;
290
291		if (size <= rb_hole_size(rb)) {
292			link = &rb->rb_left;
293			best = rb;
 
 
 
 
294		} else {
295			link = &rb->rb_right;
296		}
297	}
298
299	return rb_hole_size_to_node(best);
300}
301
302static struct drm_mm_node *find_hole(struct drm_mm *mm, u64 addr)
303{
 
304	struct drm_mm_node *node = NULL;
305	struct rb_node **link = &mm->holes_addr.rb_node;
306
307	while (*link) {
308		u64 hole_start;
309
310		node = rb_hole_addr_to_node(*link);
311		hole_start = __drm_mm_hole_node_start(node);
312
313		if (addr < hole_start)
314			link = &node->rb_hole_addr.rb_left;
315		else if (addr > hole_start + node->hole_size)
316			link = &node->rb_hole_addr.rb_right;
317		else
318			break;
319	}
320
321	return node;
322}
323
324static struct drm_mm_node *
325first_hole(struct drm_mm *mm,
326	   u64 start, u64 end, u64 size,
327	   enum drm_mm_insert_mode mode)
328{
329	if (RB_EMPTY_ROOT(&mm->holes_size))
330		return NULL;
331
332	switch (mode) {
333	default:
334	case DRM_MM_INSERT_BEST:
335		return best_hole(mm, size);
336
337	case DRM_MM_INSERT_LOW:
338		return find_hole(mm, start);
339
340	case DRM_MM_INSERT_HIGH:
341		return find_hole(mm, end);
342
343	case DRM_MM_INSERT_EVICT:
344		return list_first_entry_or_null(&mm->hole_stack,
345						struct drm_mm_node,
346						hole_stack);
347	}
348}
349
350static struct drm_mm_node *
351next_hole(struct drm_mm *mm,
352	  struct drm_mm_node *node,
353	  enum drm_mm_insert_mode mode)
354{
355	switch (mode) {
356	default:
357	case DRM_MM_INSERT_BEST:
358		return rb_hole_size_to_node(rb_next(&node->rb_hole_size));
359
360	case DRM_MM_INSERT_LOW:
361		return rb_hole_addr_to_node(rb_next(&node->rb_hole_addr));
362
363	case DRM_MM_INSERT_HIGH:
364		return rb_hole_addr_to_node(rb_prev(&node->rb_hole_addr));
365
366	case DRM_MM_INSERT_EVICT:
367		node = list_next_entry(node, hole_stack);
368		return &node->hole_stack == &mm->hole_stack ? NULL : node;
369	}
370}
371
372/**
373 * drm_mm_reserve_node - insert an pre-initialized node
374 * @mm: drm_mm allocator to insert @node into
375 * @node: drm_mm_node to insert
376 *
377 * This functions inserts an already set-up &drm_mm_node into the allocator,
378 * meaning that start, size and color must be set by the caller. All other
379 * fields must be cleared to 0. This is useful to initialize the allocator with
380 * preallocated objects which must be set-up before the range allocator can be
381 * set-up, e.g. when taking over a firmware framebuffer.
382 *
383 * Returns:
384 * 0 on success, -ENOSPC if there's no hole where @node is.
385 */
386int drm_mm_reserve_node(struct drm_mm *mm, struct drm_mm_node *node)
387{
388	u64 end = node->start + node->size;
389	struct drm_mm_node *hole;
390	u64 hole_start, hole_end;
391	u64 adj_start, adj_end;
392
393	end = node->start + node->size;
394	if (unlikely(end <= node->start))
395		return -ENOSPC;
396
397	/* Find the relevant hole to add our node to */
398	hole = find_hole(mm, node->start);
399	if (!hole)
400		return -ENOSPC;
401
402	adj_start = hole_start = __drm_mm_hole_node_start(hole);
403	adj_end = hole_end = hole_start + hole->hole_size;
404
405	if (mm->color_adjust)
406		mm->color_adjust(hole, node->color, &adj_start, &adj_end);
407
408	if (adj_start > node->start || adj_end < end)
409		return -ENOSPC;
410
411	node->mm = mm;
412
413	list_add(&node->node_list, &hole->node_list);
414	drm_mm_interval_tree_add_node(hole, node);
415	node->allocated = true;
416	node->hole_size = 0;
417
418	rm_hole(hole);
419	if (node->start > hole_start)
420		add_hole(hole);
421	if (end < hole_end)
422		add_hole(node);
423
424	save_stack(node);
425	return 0;
426}
427EXPORT_SYMBOL(drm_mm_reserve_node);
428
 
 
 
 
 
429/**
430 * drm_mm_insert_node_in_range - ranged search for space and insert @node
431 * @mm: drm_mm to allocate from
432 * @node: preallocate node to insert
433 * @size: size of the allocation
434 * @alignment: alignment of the allocation
435 * @color: opaque tag value to use for this node
436 * @range_start: start of the allowed range for this node
437 * @range_end: end of the allowed range for this node
438 * @mode: fine-tune the allocation search and placement
439 *
440 * The preallocated @node must be cleared to 0.
441 *
442 * Returns:
443 * 0 on success, -ENOSPC if there's no suitable hole.
444 */
445int drm_mm_insert_node_in_range(struct drm_mm * const mm,
446				struct drm_mm_node * const node,
447				u64 size, u64 alignment,
448				unsigned long color,
449				u64 range_start, u64 range_end,
450				enum drm_mm_insert_mode mode)
451{
452	struct drm_mm_node *hole;
453	u64 remainder_mask;
 
454
455	DRM_MM_BUG_ON(range_start >= range_end);
456
457	if (unlikely(size == 0 || range_end - range_start < size))
458		return -ENOSPC;
459
 
 
 
460	if (alignment <= 1)
461		alignment = 0;
462
 
 
 
463	remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
464	for (hole = first_hole(mm, range_start, range_end, size, mode); hole;
465	     hole = next_hole(mm, hole, mode)) {
 
466		u64 hole_start = __drm_mm_hole_node_start(hole);
467		u64 hole_end = hole_start + hole->hole_size;
468		u64 adj_start, adj_end;
469		u64 col_start, col_end;
470
471		if (mode == DRM_MM_INSERT_LOW && hole_start >= range_end)
472			break;
473
474		if (mode == DRM_MM_INSERT_HIGH && hole_end <= range_start)
475			break;
476
477		col_start = hole_start;
478		col_end = hole_end;
479		if (mm->color_adjust)
480			mm->color_adjust(hole, color, &col_start, &col_end);
481
482		adj_start = max(col_start, range_start);
483		adj_end = min(col_end, range_end);
484
485		if (adj_end <= adj_start || adj_end - adj_start < size)
486			continue;
487
488		if (mode == DRM_MM_INSERT_HIGH)
489			adj_start = adj_end - size;
490
491		if (alignment) {
492			u64 rem;
493
494			if (likely(remainder_mask))
495				rem = adj_start & remainder_mask;
496			else
497				div64_u64_rem(adj_start, alignment, &rem);
498			if (rem) {
499				adj_start -= rem;
500				if (mode != DRM_MM_INSERT_HIGH)
501					adj_start += alignment;
502
503				if (adj_start < max(col_start, range_start) ||
504				    min(col_end, range_end) - adj_start < size)
505					continue;
506
507				if (adj_end <= adj_start ||
508				    adj_end - adj_start < size)
509					continue;
510			}
511		}
512
513		node->mm = mm;
514		node->size = size;
515		node->start = adj_start;
516		node->color = color;
517		node->hole_size = 0;
518
519		list_add(&node->node_list, &hole->node_list);
520		drm_mm_interval_tree_add_node(hole, node);
521		node->allocated = true;
522
523		rm_hole(hole);
524		if (adj_start > hole_start)
525			add_hole(hole);
526		if (adj_start + size < hole_end)
527			add_hole(node);
528
529		save_stack(node);
530		return 0;
531	}
532
533	return -ENOSPC;
534}
535EXPORT_SYMBOL(drm_mm_insert_node_in_range);
536
537/**
538 * drm_mm_remove_node - Remove a memory node from the allocator.
539 * @node: drm_mm_node to remove
540 *
541 * This just removes a node from its drm_mm allocator. The node does not need to
542 * be cleared again before it can be re-inserted into this or any other drm_mm
543 * allocator. It is a bug to call this function on a unallocated node.
544 */
545void drm_mm_remove_node(struct drm_mm_node *node)
546{
547	struct drm_mm *mm = node->mm;
548	struct drm_mm_node *prev_node;
549
550	DRM_MM_BUG_ON(!node->allocated);
551	DRM_MM_BUG_ON(node->scanned_block);
552
553	prev_node = list_prev_entry(node, node_list);
554
555	if (drm_mm_hole_follows(node))
556		rm_hole(node);
557
558	drm_mm_interval_tree_remove(node, &mm->interval_tree);
559	list_del(&node->node_list);
560	node->allocated = false;
561
562	if (drm_mm_hole_follows(prev_node))
563		rm_hole(prev_node);
564	add_hole(prev_node);
565}
566EXPORT_SYMBOL(drm_mm_remove_node);
567
568/**
569 * drm_mm_replace_node - move an allocation from @old to @new
570 * @old: drm_mm_node to remove from the allocator
571 * @new: drm_mm_node which should inherit @old's allocation
572 *
573 * This is useful for when drivers embed the drm_mm_node structure and hence
574 * can't move allocations by reassigning pointers. It's a combination of remove
575 * and insert with the guarantee that the allocation start will match.
576 */
577void drm_mm_replace_node(struct drm_mm_node *old, struct drm_mm_node *new)
578{
579	struct drm_mm *mm = old->mm;
580
581	DRM_MM_BUG_ON(!old->allocated);
582
583	*new = *old;
584
585	list_replace(&old->node_list, &new->node_list);
586	rb_replace_node_cached(&old->rb, &new->rb, &mm->interval_tree);
587
588	if (drm_mm_hole_follows(old)) {
589		list_replace(&old->hole_stack, &new->hole_stack);
590		rb_replace_node(&old->rb_hole_size,
591				&new->rb_hole_size,
592				&mm->holes_size);
593		rb_replace_node(&old->rb_hole_addr,
594				&new->rb_hole_addr,
595				&mm->holes_addr);
596	}
597
598	old->allocated = false;
599	new->allocated = true;
600}
601EXPORT_SYMBOL(drm_mm_replace_node);
602
603/**
604 * DOC: lru scan roster
605 *
606 * Very often GPUs need to have continuous allocations for a given object. When
607 * evicting objects to make space for a new one it is therefore not most
608 * efficient when we simply start to select all objects from the tail of an LRU
609 * until there's a suitable hole: Especially for big objects or nodes that
610 * otherwise have special allocation constraints there's a good chance we evict
611 * lots of (smaller) objects unnecessarily.
612 *
613 * The DRM range allocator supports this use-case through the scanning
614 * interfaces. First a scan operation needs to be initialized with
615 * drm_mm_scan_init() or drm_mm_scan_init_with_range(). The driver adds
616 * objects to the roster, probably by walking an LRU list, but this can be
617 * freely implemented. Eviction candiates are added using
618 * drm_mm_scan_add_block() until a suitable hole is found or there are no
619 * further evictable objects. Eviction roster metadata is tracked in &struct
620 * drm_mm_scan.
621 *
622 * The driver must walk through all objects again in exactly the reverse
623 * order to restore the allocator state. Note that while the allocator is used
624 * in the scan mode no other operation is allowed.
625 *
626 * Finally the driver evicts all objects selected (drm_mm_scan_remove_block()
627 * reported true) in the scan, and any overlapping nodes after color adjustment
628 * (drm_mm_scan_color_evict()). Adding and removing an object is O(1), and
629 * since freeing a node is also O(1) the overall complexity is
630 * O(scanned_objects). So like the free stack which needs to be walked before a
631 * scan operation even begins this is linear in the number of objects. It
632 * doesn't seem to hurt too badly.
633 */
634
635/**
636 * drm_mm_scan_init_with_range - initialize range-restricted lru scanning
637 * @scan: scan state
638 * @mm: drm_mm to scan
639 * @size: size of the allocation
640 * @alignment: alignment of the allocation
641 * @color: opaque tag value to use for the allocation
642 * @start: start of the allowed range for the allocation
643 * @end: end of the allowed range for the allocation
644 * @mode: fine-tune the allocation search and placement
645 *
646 * This simply sets up the scanning routines with the parameters for the desired
647 * hole.
648 *
649 * Warning:
650 * As long as the scan list is non-empty, no other operations than
651 * adding/removing nodes to/from the scan list are allowed.
652 */
653void drm_mm_scan_init_with_range(struct drm_mm_scan *scan,
654				 struct drm_mm *mm,
655				 u64 size,
656				 u64 alignment,
657				 unsigned long color,
658				 u64 start,
659				 u64 end,
660				 enum drm_mm_insert_mode mode)
661{
662	DRM_MM_BUG_ON(start >= end);
663	DRM_MM_BUG_ON(!size || size > end - start);
664	DRM_MM_BUG_ON(mm->scan_active);
665
666	scan->mm = mm;
667
668	if (alignment <= 1)
669		alignment = 0;
670
671	scan->color = color;
672	scan->alignment = alignment;
673	scan->remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
674	scan->size = size;
675	scan->mode = mode;
676
677	DRM_MM_BUG_ON(end <= start);
678	scan->range_start = start;
679	scan->range_end = end;
680
681	scan->hit_start = U64_MAX;
682	scan->hit_end = 0;
683}
684EXPORT_SYMBOL(drm_mm_scan_init_with_range);
685
686/**
687 * drm_mm_scan_add_block - add a node to the scan list
688 * @scan: the active drm_mm scanner
689 * @node: drm_mm_node to add
690 *
691 * Add a node to the scan list that might be freed to make space for the desired
692 * hole.
693 *
694 * Returns:
695 * True if a hole has been found, false otherwise.
696 */
697bool drm_mm_scan_add_block(struct drm_mm_scan *scan,
698			   struct drm_mm_node *node)
699{
700	struct drm_mm *mm = scan->mm;
701	struct drm_mm_node *hole;
702	u64 hole_start, hole_end;
703	u64 col_start, col_end;
704	u64 adj_start, adj_end;
705
706	DRM_MM_BUG_ON(node->mm != mm);
707	DRM_MM_BUG_ON(!node->allocated);
708	DRM_MM_BUG_ON(node->scanned_block);
709	node->scanned_block = true;
710	mm->scan_active++;
711
712	/* Remove this block from the node_list so that we enlarge the hole
713	 * (distance between the end of our previous node and the start of
714	 * or next), without poisoning the link so that we can restore it
715	 * later in drm_mm_scan_remove_block().
716	 */
717	hole = list_prev_entry(node, node_list);
718	DRM_MM_BUG_ON(list_next_entry(hole, node_list) != node);
719	__list_del_entry(&node->node_list);
720
721	hole_start = __drm_mm_hole_node_start(hole);
722	hole_end = __drm_mm_hole_node_end(hole);
723
724	col_start = hole_start;
725	col_end = hole_end;
726	if (mm->color_adjust)
727		mm->color_adjust(hole, scan->color, &col_start, &col_end);
728
729	adj_start = max(col_start, scan->range_start);
730	adj_end = min(col_end, scan->range_end);
731	if (adj_end <= adj_start || adj_end - adj_start < scan->size)
732		return false;
733
734	if (scan->mode == DRM_MM_INSERT_HIGH)
735		adj_start = adj_end - scan->size;
736
737	if (scan->alignment) {
738		u64 rem;
739
740		if (likely(scan->remainder_mask))
741			rem = adj_start & scan->remainder_mask;
742		else
743			div64_u64_rem(adj_start, scan->alignment, &rem);
744		if (rem) {
745			adj_start -= rem;
746			if (scan->mode != DRM_MM_INSERT_HIGH)
747				adj_start += scan->alignment;
748			if (adj_start < max(col_start, scan->range_start) ||
749			    min(col_end, scan->range_end) - adj_start < scan->size)
750				return false;
751
752			if (adj_end <= adj_start ||
753			    adj_end - adj_start < scan->size)
754				return false;
755		}
756	}
757
758	scan->hit_start = adj_start;
759	scan->hit_end = adj_start + scan->size;
760
761	DRM_MM_BUG_ON(scan->hit_start >= scan->hit_end);
762	DRM_MM_BUG_ON(scan->hit_start < hole_start);
763	DRM_MM_BUG_ON(scan->hit_end > hole_end);
764
765	return true;
766}
767EXPORT_SYMBOL(drm_mm_scan_add_block);
768
769/**
770 * drm_mm_scan_remove_block - remove a node from the scan list
771 * @scan: the active drm_mm scanner
772 * @node: drm_mm_node to remove
773 *
774 * Nodes **must** be removed in exactly the reverse order from the scan list as
775 * they have been added (e.g. using list_add() as they are added and then
776 * list_for_each() over that eviction list to remove), otherwise the internal
777 * state of the memory manager will be corrupted.
778 *
779 * When the scan list is empty, the selected memory nodes can be freed. An
780 * immediately following drm_mm_insert_node_in_range_generic() or one of the
781 * simpler versions of that function with !DRM_MM_SEARCH_BEST will then return
782 * the just freed block (because its at the top of the free_stack list).
783 *
784 * Returns:
785 * True if this block should be evicted, false otherwise. Will always
786 * return false when no hole has been found.
787 */
788bool drm_mm_scan_remove_block(struct drm_mm_scan *scan,
789			      struct drm_mm_node *node)
790{
791	struct drm_mm_node *prev_node;
792
793	DRM_MM_BUG_ON(node->mm != scan->mm);
794	DRM_MM_BUG_ON(!node->scanned_block);
795	node->scanned_block = false;
796
797	DRM_MM_BUG_ON(!node->mm->scan_active);
798	node->mm->scan_active--;
799
800	/* During drm_mm_scan_add_block() we decoupled this node leaving
801	 * its pointers intact. Now that the caller is walking back along
802	 * the eviction list we can restore this block into its rightful
803	 * place on the full node_list. To confirm that the caller is walking
804	 * backwards correctly we check that prev_node->next == node->next,
805	 * i.e. both believe the same node should be on the other side of the
806	 * hole.
807	 */
808	prev_node = list_prev_entry(node, node_list);
809	DRM_MM_BUG_ON(list_next_entry(prev_node, node_list) !=
810		      list_next_entry(node, node_list));
811	list_add(&node->node_list, &prev_node->node_list);
812
813	return (node->start + node->size > scan->hit_start &&
814		node->start < scan->hit_end);
815}
816EXPORT_SYMBOL(drm_mm_scan_remove_block);
817
818/**
819 * drm_mm_scan_color_evict - evict overlapping nodes on either side of hole
820 * @scan: drm_mm scan with target hole
821 *
822 * After completing an eviction scan and removing the selected nodes, we may
823 * need to remove a few more nodes from either side of the target hole if
824 * mm.color_adjust is being used.
825 *
826 * Returns:
827 * A node to evict, or NULL if there are no overlapping nodes.
828 */
829struct drm_mm_node *drm_mm_scan_color_evict(struct drm_mm_scan *scan)
830{
831	struct drm_mm *mm = scan->mm;
832	struct drm_mm_node *hole;
833	u64 hole_start, hole_end;
834
835	DRM_MM_BUG_ON(list_empty(&mm->hole_stack));
836
837	if (!mm->color_adjust)
838		return NULL;
839
840	/*
841	 * The hole found during scanning should ideally be the first element
842	 * in the hole_stack list, but due to side-effects in the driver it
843	 * may not be.
844	 */
845	list_for_each_entry(hole, &mm->hole_stack, hole_stack) {
846		hole_start = __drm_mm_hole_node_start(hole);
847		hole_end = hole_start + hole->hole_size;
848
849		if (hole_start <= scan->hit_start &&
850		    hole_end >= scan->hit_end)
851			break;
852	}
853
854	/* We should only be called after we found the hole previously */
855	DRM_MM_BUG_ON(&hole->hole_stack == &mm->hole_stack);
856	if (unlikely(&hole->hole_stack == &mm->hole_stack))
857		return NULL;
858
859	DRM_MM_BUG_ON(hole_start > scan->hit_start);
860	DRM_MM_BUG_ON(hole_end < scan->hit_end);
861
862	mm->color_adjust(hole, scan->color, &hole_start, &hole_end);
863	if (hole_start > scan->hit_start)
864		return hole;
865	if (hole_end < scan->hit_end)
866		return list_next_entry(hole, node_list);
867
868	return NULL;
869}
870EXPORT_SYMBOL(drm_mm_scan_color_evict);
871
872/**
873 * drm_mm_init - initialize a drm-mm allocator
874 * @mm: the drm_mm structure to initialize
875 * @start: start of the range managed by @mm
876 * @size: end of the range managed by @mm
877 *
878 * Note that @mm must be cleared to 0 before calling this function.
879 */
880void drm_mm_init(struct drm_mm *mm, u64 start, u64 size)
881{
882	DRM_MM_BUG_ON(start + size <= start);
883
884	mm->color_adjust = NULL;
885
886	INIT_LIST_HEAD(&mm->hole_stack);
887	mm->interval_tree = RB_ROOT_CACHED;
888	mm->holes_size = RB_ROOT;
889	mm->holes_addr = RB_ROOT;
890
891	/* Clever trick to avoid a special case in the free hole tracking. */
892	INIT_LIST_HEAD(&mm->head_node.node_list);
893	mm->head_node.allocated = false;
894	mm->head_node.mm = mm;
895	mm->head_node.start = start + size;
896	mm->head_node.size = -size;
897	add_hole(&mm->head_node);
898
899	mm->scan_active = 0;
900}
901EXPORT_SYMBOL(drm_mm_init);
902
903/**
904 * drm_mm_takedown - clean up a drm_mm allocator
905 * @mm: drm_mm allocator to clean up
906 *
907 * Note that it is a bug to call this function on an allocator which is not
908 * clean.
909 */
910void drm_mm_takedown(struct drm_mm *mm)
911{
912	if (WARN(!drm_mm_clean(mm),
913		 "Memory manager not clean during takedown.\n"))
914		show_leaks(mm);
915}
916EXPORT_SYMBOL(drm_mm_takedown);
917
918static u64 drm_mm_dump_hole(struct drm_printer *p, const struct drm_mm_node *entry)
919{
920	u64 start, size;
921
922	size = entry->hole_size;
923	if (size) {
924		start = drm_mm_hole_node_start(entry);
925		drm_printf(p, "%#018llx-%#018llx: %llu: free\n",
926			   start, start + size, size);
927	}
928
929	return size;
930}
931/**
932 * drm_mm_print - print allocator state
933 * @mm: drm_mm allocator to print
934 * @p: DRM printer to use
935 */
936void drm_mm_print(const struct drm_mm *mm, struct drm_printer *p)
937{
938	const struct drm_mm_node *entry;
939	u64 total_used = 0, total_free = 0, total = 0;
940
941	total_free += drm_mm_dump_hole(p, &mm->head_node);
942
943	drm_mm_for_each_node(entry, mm) {
944		drm_printf(p, "%#018llx-%#018llx: %llu: used\n", entry->start,
945			   entry->start + entry->size, entry->size);
946		total_used += entry->size;
947		total_free += drm_mm_dump_hole(p, entry);
948	}
949	total = total_free + total_used;
950
951	drm_printf(p, "total: %llu, used %llu free %llu\n", total,
952		   total_used, total_free);
953}
954EXPORT_SYMBOL(drm_mm_print);