Linux Audio

Check our new training course

Loading...
v3.1
  1/**************************************************************************
  2 *
  3 * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA.
 
  4 * All Rights Reserved.
  5 *
  6 * Permission is hereby granted, free of charge, to any person obtaining a
  7 * copy of this software and associated documentation files (the
  8 * "Software"), to deal in the Software without restriction, including
  9 * without limitation the rights to use, copy, modify, merge, publish,
 10 * distribute, sub license, and/or sell copies of the Software, and to
 11 * permit persons to whom the Software is furnished to do so, subject to
 12 * the following conditions:
 13 *
 14 * The above copyright notice and this permission notice (including the
 15 * next paragraph) shall be included in all copies or substantial portions
 16 * of the Software.
 17 *
 18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
 21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
 22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 24 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 25 *
 26 *
 27 **************************************************************************/
 28
 29/*
 30 * Generic simple memory manager implementation. Intended to be used as a base
 31 * class implementation for more advanced memory managers.
 32 *
 33 * Note that the algorithm used is quite simple and there might be substantial
 34 * performance gains if a smarter free list is implemented. Currently it is just an
 35 * unordered stack of free regions. This could easily be improved if an RB-tree
 36 * is used instead. At least if we expect heavy fragmentation.
 37 *
 38 * Aligned allocations can also see improvement.
 39 *
 40 * Authors:
 41 * Thomas Hellström <thomas-at-tungstengraphics-dot-com>
 42 */
 43
 44#include "drmP.h"
 45#include "drm_mm.h"
 46#include <linux/slab.h>
 47#include <linux/seq_file.h>
 
 
 48
 49#define MM_UNUSED_TARGET 4
 50
 51static struct drm_mm_node *drm_mm_kmalloc(struct drm_mm *mm, int atomic)
 52{
 53	struct drm_mm_node *child;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 54
 55	if (atomic)
 56		child = kzalloc(sizeof(*child), GFP_ATOMIC);
 57	else
 58		child = kzalloc(sizeof(*child), GFP_KERNEL);
 59
 60	if (unlikely(child == NULL)) {
 61		spin_lock(&mm->unused_lock);
 62		if (list_empty(&mm->unused_nodes))
 63			child = NULL;
 64		else {
 65			child =
 66			    list_entry(mm->unused_nodes.next,
 67				       struct drm_mm_node, node_list);
 68			list_del(&child->node_list);
 69			--mm->num_unused;
 70		}
 71		spin_unlock(&mm->unused_lock);
 72	}
 73	return child;
 74}
 75
 76/* drm_mm_pre_get() - pre allocate drm_mm_node structure
 77 * drm_mm:	memory manager struct we are pre-allocating for
 78 *
 79 * Returns 0 on success or -ENOMEM if allocation fails.
 80 */
 81int drm_mm_pre_get(struct drm_mm *mm)
 82{
 83	struct drm_mm_node *node;
 
 
 
 
 
 
 84
 85	spin_lock(&mm->unused_lock);
 86	while (mm->num_unused < MM_UNUSED_TARGET) {
 87		spin_unlock(&mm->unused_lock);
 88		node = kzalloc(sizeof(*node), GFP_KERNEL);
 89		spin_lock(&mm->unused_lock);
 90
 91		if (unlikely(node == NULL)) {
 92			int ret = (mm->num_unused < 2) ? -ENOMEM : 0;
 93			spin_unlock(&mm->unused_lock);
 94			return ret;
 95		}
 96		++mm->num_unused;
 97		list_add_tail(&node->node_list, &mm->unused_nodes);
 
 
 
 98	}
 99	spin_unlock(&mm->unused_lock);
100	return 0;
101}
102EXPORT_SYMBOL(drm_mm_pre_get);
103
104static inline unsigned long drm_mm_hole_node_start(struct drm_mm_node *hole_node)
105{
106	return hole_node->start + hole_node->size;
107}
108
109static inline unsigned long drm_mm_hole_node_end(struct drm_mm_node *hole_node)
110{
111	struct drm_mm_node *next_node =
112		list_entry(hole_node->node_list.next, struct drm_mm_node,
113			   node_list);
 
 
 
 
114
115	return next_node->start;
 
 
 
 
 
 
 
 
116}
 
117
118static void drm_mm_insert_helper(struct drm_mm_node *hole_node,
119				 struct drm_mm_node *node,
120				 unsigned long size, unsigned alignment)
121{
122	struct drm_mm *mm = hole_node->mm;
123	unsigned long tmp = 0, wasted = 0;
124	unsigned long hole_start = drm_mm_hole_node_start(hole_node);
125	unsigned long hole_end = drm_mm_hole_node_end(hole_node);
126
127	BUG_ON(!hole_node->hole_follows || node->allocated);
128
129	if (alignment)
130		tmp = hole_start % alignment;
131
132	if (!tmp) {
133		hole_node->hole_follows = 0;
134		list_del_init(&hole_node->hole_stack);
135	} else
136		wasted = alignment - tmp;
137
138	node->start = hole_start + wasted;
139	node->size = size;
140	node->mm = mm;
141	node->allocated = 1;
142
143	INIT_LIST_HEAD(&node->hole_stack);
144	list_add(&node->node_list, &hole_node->node_list);
145
146	BUG_ON(node->start + node->size > hole_end);
 
 
147
148	if (node->start + node->size < hole_end) {
149		list_add(&node->hole_stack, &mm->hole_stack);
150		node->hole_follows = 1;
151	} else {
152		node->hole_follows = 0;
 
 
153	}
154}
155
156struct drm_mm_node *drm_mm_get_block_generic(struct drm_mm_node *hole_node,
157					     unsigned long size,
158					     unsigned alignment,
159					     int atomic)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
160{
161	struct drm_mm_node *node;
162
163	node = drm_mm_kmalloc(hole_node->mm, atomic);
164	if (unlikely(node == NULL))
165		return NULL;
166
167	drm_mm_insert_helper(hole_node, node, size, alignment);
 
168
169	return node;
170}
171EXPORT_SYMBOL(drm_mm_get_block_generic);
172
173/**
174 * Search for free space and insert a preallocated memory node. Returns
175 * -ENOSPC if no suitable free area is available. The preallocated memory node
176 * must be cleared.
177 */
178int drm_mm_insert_node(struct drm_mm *mm, struct drm_mm_node *node,
179		       unsigned long size, unsigned alignment)
180{
181	struct drm_mm_node *hole_node;
182
183	hole_node = drm_mm_search_free(mm, size, alignment, 0);
184	if (!hole_node)
185		return -ENOSPC;
 
186
187	drm_mm_insert_helper(hole_node, node, size, alignment);
 
188
189	return 0;
 
 
190}
191EXPORT_SYMBOL(drm_mm_insert_node);
192
193static void drm_mm_insert_helper_range(struct drm_mm_node *hole_node,
194				       struct drm_mm_node *node,
195				       unsigned long size, unsigned alignment,
196				       unsigned long start, unsigned long end)
197{
198	struct drm_mm *mm = hole_node->mm;
199	unsigned long tmp = 0, wasted = 0;
200	unsigned long hole_start = drm_mm_hole_node_start(hole_node);
201	unsigned long hole_end = drm_mm_hole_node_end(hole_node);
202
203	BUG_ON(!hole_node->hole_follows || node->allocated);
 
 
 
204
205	if (hole_start < start)
206		wasted += start - hole_start;
207	if (alignment)
208		tmp = (hole_start + wasted) % alignment;
209
210	if (tmp)
211		wasted += alignment - tmp;
212
213	if (!wasted) {
214		hole_node->hole_follows = 0;
215		list_del_init(&hole_node->hole_stack);
 
 
 
216	}
217
218	node->start = hole_start + wasted;
219	node->size = size;
220	node->mm = mm;
221	node->allocated = 1;
222
223	INIT_LIST_HEAD(&node->hole_stack);
224	list_add(&node->node_list, &hole_node->node_list);
 
 
225
226	BUG_ON(node->start + node->size > hole_end);
227	BUG_ON(node->start + node->size > end);
228
229	if (node->start + node->size < hole_end) {
230		list_add(&node->hole_stack, &mm->hole_stack);
231		node->hole_follows = 1;
232	} else {
233		node->hole_follows = 0;
 
 
 
 
234	}
 
 
235}
236
237struct drm_mm_node *drm_mm_get_block_range_generic(struct drm_mm_node *hole_node,
238						unsigned long size,
239						unsigned alignment,
240						unsigned long start,
241						unsigned long end,
242						int atomic)
243{
244	struct drm_mm_node *node;
245
246	node = drm_mm_kmalloc(hole_node->mm, atomic);
247	if (unlikely(node == NULL))
248		return NULL;
249
250	drm_mm_insert_helper_range(hole_node, node, size, alignment,
251				   start, end);
 
 
252
253	return node;
 
 
 
 
 
 
 
 
 
 
254}
255EXPORT_SYMBOL(drm_mm_get_block_range_generic);
256
257/**
258 * Search for free space and insert a preallocated memory node. Returns
259 * -ENOSPC if no suitable free area is available. This is for range
260 * restricted allocations. The preallocated memory node must be cleared.
261 */
262int drm_mm_insert_node_in_range(struct drm_mm *mm, struct drm_mm_node *node,
263				unsigned long size, unsigned alignment,
264				unsigned long start, unsigned long end)
265{
266	struct drm_mm_node *hole_node;
 
 
 
267
268	hole_node = drm_mm_search_free_in_range(mm, size, alignment,
269						start, end, 0);
270	if (!hole_node)
271		return -ENOSPC;
272
273	drm_mm_insert_helper_range(hole_node, node, size, alignment,
274				   start, end);
275
276	return 0;
 
 
 
277}
278EXPORT_SYMBOL(drm_mm_insert_node_in_range);
279
280/**
281 * Remove a memory node from the allocator.
282 */
283void drm_mm_remove_node(struct drm_mm_node *node)
284{
285	struct drm_mm *mm = node->mm;
286	struct drm_mm_node *prev_node;
 
 
 
 
 
 
 
 
 
 
 
 
 
287
288	BUG_ON(node->scanned_block || node->scanned_prev_free
289				   || node->scanned_next_free);
 
290
291	prev_node =
292	    list_entry(node->node_list.prev, struct drm_mm_node, node_list);
 
 
293
294	if (node->hole_follows) {
295		BUG_ON(drm_mm_hole_node_start(node)
296				== drm_mm_hole_node_end(node));
297		list_del(&node->hole_stack);
298	} else
299		BUG_ON(drm_mm_hole_node_start(node)
300				!= drm_mm_hole_node_end(node));
301
302	if (!prev_node->hole_follows) {
303		prev_node->hole_follows = 1;
304		list_add(&prev_node->hole_stack, &mm->hole_stack);
305	} else
306		list_move(&prev_node->hole_stack, &mm->hole_stack);
307
308	list_del(&node->node_list);
309	node->allocated = 0;
310}
311EXPORT_SYMBOL(drm_mm_remove_node);
312
313/*
314 * Remove a memory node from the allocator and free the allocated struct
315 * drm_mm_node. Only to be used on a struct drm_mm_node obtained by one of the
316 * drm_mm_get_block functions.
317 */
318void drm_mm_put_block(struct drm_mm_node *node)
319{
320
321	struct drm_mm *mm = node->mm;
322
323	drm_mm_remove_node(node);
 
 
 
 
 
 
 
 
 
324
325	spin_lock(&mm->unused_lock);
326	if (mm->num_unused < MM_UNUSED_TARGET) {
327		list_add(&node->node_list, &mm->unused_nodes);
328		++mm->num_unused;
329	} else
330		kfree(node);
331	spin_unlock(&mm->unused_lock);
332}
333EXPORT_SYMBOL(drm_mm_put_block);
334
335static int check_free_hole(unsigned long start, unsigned long end,
336			   unsigned long size, unsigned alignment)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
337{
338	unsigned wasted = 0;
 
339
340	if (end - start < size)
341		return 0;
342
343	if (alignment) {
344		unsigned tmp = start % alignment;
345		if (tmp)
346			wasted = alignment - tmp;
347	}
348
349	if (end >= start + size + wasted) {
350		return 1;
351	}
352
353	return 0;
354}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
355
356struct drm_mm_node *drm_mm_search_free(const struct drm_mm *mm,
357				       unsigned long size,
358				       unsigned alignment, int best_match)
359{
360	struct drm_mm_node *entry;
361	struct drm_mm_node *best;
362	unsigned long best_size;
363
364	BUG_ON(mm->scanned_blocks);
365
366	best = NULL;
367	best_size = ~0UL;
368
369	list_for_each_entry(entry, &mm->hole_stack, hole_stack) {
370		BUG_ON(!entry->hole_follows);
371		if (!check_free_hole(drm_mm_hole_node_start(entry),
372				     drm_mm_hole_node_end(entry),
373				     size, alignment))
374			continue;
375
376		if (!best_match)
377			return entry;
378
379		if (entry->size < best_size) {
380			best = entry;
381			best_size = entry->size;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
382		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
383	}
384
385	return best;
386}
387EXPORT_SYMBOL(drm_mm_search_free);
388
389struct drm_mm_node *drm_mm_search_free_in_range(const struct drm_mm *mm,
390						unsigned long size,
391						unsigned alignment,
392						unsigned long start,
393						unsigned long end,
394						int best_match)
395{
396	struct drm_mm_node *entry;
397	struct drm_mm_node *best;
398	unsigned long best_size;
399
400	BUG_ON(mm->scanned_blocks);
401
402	best = NULL;
403	best_size = ~0UL;
404
405	list_for_each_entry(entry, &mm->hole_stack, hole_stack) {
406		unsigned long adj_start = drm_mm_hole_node_start(entry) < start ?
407			start : drm_mm_hole_node_start(entry);
408		unsigned long adj_end = drm_mm_hole_node_end(entry) > end ?
409			end : drm_mm_hole_node_end(entry);
410
411		BUG_ON(!entry->hole_follows);
412		if (!check_free_hole(adj_start, adj_end, size, alignment))
413			continue;
414
415		if (!best_match)
416			return entry;
417
418		if (entry->size < best_size) {
419			best = entry;
420			best_size = entry->size;
421		}
422	}
423
424	return best;
 
 
 
 
 
 
425}
426EXPORT_SYMBOL(drm_mm_search_free_in_range);
427
428/**
429 * Moves an allocation. To be used with embedded struct drm_mm_node.
 
 
 
 
 
 
430 */
431void drm_mm_replace_node(struct drm_mm_node *old, struct drm_mm_node *new)
432{
 
 
 
 
 
 
433	list_replace(&old->node_list, &new->node_list);
434	list_replace(&old->hole_stack, &new->hole_stack);
435	new->hole_follows = old->hole_follows;
436	new->mm = old->mm;
437	new->start = old->start;
438	new->size = old->size;
 
 
 
 
 
 
439
440	old->allocated = 0;
441	new->allocated = 1;
442}
443EXPORT_SYMBOL(drm_mm_replace_node);
444
445/**
446 * Initializa lru scanning.
447 *
448 * This simply sets up the scanning routines with the parameters for the desired
449 * hole.
450 *
451 * Warning: As long as the scan list is non-empty, no other operations than
452 * adding/removing nodes to/from the scan list are allowed.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
453 */
454void drm_mm_init_scan(struct drm_mm *mm, unsigned long size,
455		      unsigned alignment)
456{
457	mm->scan_alignment = alignment;
458	mm->scan_size = size;
459	mm->scanned_blocks = 0;
460	mm->scan_hit_start = 0;
461	mm->scan_hit_size = 0;
462	mm->scan_check_range = 0;
463	mm->prev_scanned_node = NULL;
464}
465EXPORT_SYMBOL(drm_mm_init_scan);
466
467/**
468 * Initializa lru scanning.
 
 
 
 
 
 
 
 
469 *
470 * This simply sets up the scanning routines with the parameters for the desired
471 * hole. This version is for range-restricted scans.
472 *
473 * Warning: As long as the scan list is non-empty, no other operations than
 
474 * adding/removing nodes to/from the scan list are allowed.
475 */
476void drm_mm_init_scan_with_range(struct drm_mm *mm, unsigned long size,
477				 unsigned alignment,
478				 unsigned long start,
479				 unsigned long end)
480{
481	mm->scan_alignment = alignment;
482	mm->scan_size = size;
483	mm->scanned_blocks = 0;
484	mm->scan_hit_start = 0;
485	mm->scan_hit_size = 0;
486	mm->scan_start = start;
487	mm->scan_end = end;
488	mm->scan_check_range = 1;
489	mm->prev_scanned_node = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
490}
491EXPORT_SYMBOL(drm_mm_init_scan_with_range);
492
493/**
 
 
 
 
494 * Add a node to the scan list that might be freed to make space for the desired
495 * hole.
496 *
497 * Returns non-zero, if a hole has been found, zero otherwise.
 
498 */
499int drm_mm_scan_add_block(struct drm_mm_node *node)
 
500{
501	struct drm_mm *mm = node->mm;
502	struct drm_mm_node *prev_node;
503	unsigned long hole_start, hole_end;
504	unsigned long adj_start;
505	unsigned long adj_end;
506
507	mm->scanned_blocks++;
508
509	BUG_ON(node->scanned_block);
510	node->scanned_block = 1;
511
512	prev_node = list_entry(node->node_list.prev, struct drm_mm_node,
513			       node_list);
514
515	node->scanned_preceeds_hole = prev_node->hole_follows;
516	prev_node->hole_follows = 1;
517	list_del(&node->node_list);
518	node->node_list.prev = &prev_node->node_list;
519	node->node_list.next = &mm->prev_scanned_node->node_list;
520	mm->prev_scanned_node = node;
521
522	hole_start = drm_mm_hole_node_start(prev_node);
523	hole_end = drm_mm_hole_node_end(prev_node);
524	if (mm->scan_check_range) {
525		adj_start = hole_start < mm->scan_start ?
526			mm->scan_start : hole_start;
527		adj_end = hole_end > mm->scan_end ?
528			mm->scan_end : hole_end;
529	} else {
530		adj_start = hole_start;
531		adj_end = hole_end;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
532	}
533
534	if (check_free_hole(adj_start , adj_end,
535			    mm->scan_size, mm->scan_alignment)) {
536		mm->scan_hit_start = hole_start;
537		mm->scan_hit_size = hole_end;
538
539		return 1;
540	}
 
541
542	return 0;
543}
544EXPORT_SYMBOL(drm_mm_scan_add_block);
545
546/**
547 * Remove a node from the scan list.
548 *
549 * Nodes _must_ be removed in the exact same order from the scan list as they
550 * have been added, otherwise the internal state of the memory manager will be
551 * corrupted.
 
 
 
552 *
553 * When the scan list is empty, the selected memory nodes can be freed. An
554 * immediately following drm_mm_search_free with best_match = 0 will then return
 
555 * the just freed block (because its at the top of the free_stack list).
556 *
557 * Returns one if this block should be evicted, zero otherwise. Will always
558 * return zero when no hole has been found.
 
559 */
560int drm_mm_scan_remove_block(struct drm_mm_node *node)
 
561{
562	struct drm_mm *mm = node->mm;
563	struct drm_mm_node *prev_node;
564
565	mm->scanned_blocks--;
566
567	BUG_ON(!node->scanned_block);
568	node->scanned_block = 0;
569
570	prev_node = list_entry(node->node_list.prev, struct drm_mm_node,
571			       node_list);
572
573	prev_node->hole_follows = node->scanned_preceeds_hole;
574	INIT_LIST_HEAD(&node->node_list);
 
 
 
 
 
 
 
 
575	list_add(&node->node_list, &prev_node->node_list);
576
577	/* Only need to check for containement because start&size for the
578	 * complete resulting free block (not just the desired part) is
579	 * stored. */
580	if (node->start >= mm->scan_hit_start &&
581	    node->start + node->size
582	    		<= mm->scan_hit_start + mm->scan_hit_size) {
583		return 1;
584	}
585
586	return 0;
587}
588EXPORT_SYMBOL(drm_mm_scan_remove_block);
589
590int drm_mm_clean(struct drm_mm * mm)
 
 
 
 
 
 
 
 
 
 
 
591{
592	struct list_head *head = &mm->head_node.node_list;
 
 
 
 
593
594	return (head->next->next == head);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
595}
596EXPORT_SYMBOL(drm_mm_clean);
597
598int drm_mm_init(struct drm_mm * mm, unsigned long start, unsigned long size)
 
 
 
 
 
 
 
 
599{
 
 
 
 
600	INIT_LIST_HEAD(&mm->hole_stack);
601	INIT_LIST_HEAD(&mm->unused_nodes);
602	mm->num_unused = 0;
603	mm->scanned_blocks = 0;
604	spin_lock_init(&mm->unused_lock);
605
606	/* Clever trick to avoid a special case in the free hole tracking. */
607	INIT_LIST_HEAD(&mm->head_node.node_list);
608	INIT_LIST_HEAD(&mm->head_node.hole_stack);
609	mm->head_node.hole_follows = 1;
610	mm->head_node.scanned_block = 0;
611	mm->head_node.scanned_prev_free = 0;
612	mm->head_node.scanned_next_free = 0;
613	mm->head_node.mm = mm;
614	mm->head_node.start = start + size;
615	mm->head_node.size = start - mm->head_node.start;
616	list_add_tail(&mm->head_node.hole_stack, &mm->hole_stack);
617
618	return 0;
619}
620EXPORT_SYMBOL(drm_mm_init);
621
622void drm_mm_takedown(struct drm_mm * mm)
 
 
 
 
 
 
 
623{
624	struct drm_mm_node *entry, *next;
625
626	if (!list_empty(&mm->head_node.node_list)) {
627		DRM_ERROR("Memory manager not clean. Delaying takedown\n");
628		return;
629	}
630
631	spin_lock(&mm->unused_lock);
632	list_for_each_entry_safe(entry, next, &mm->unused_nodes, node_list) {
633		list_del(&entry->node_list);
634		kfree(entry);
635		--mm->num_unused;
636	}
637	spin_unlock(&mm->unused_lock);
638
639	BUG_ON(mm->num_unused != 0);
640}
641EXPORT_SYMBOL(drm_mm_takedown);
642
643void drm_mm_debug_table(struct drm_mm *mm, const char *prefix)
644{
645	struct drm_mm_node *entry;
646	unsigned long total_used = 0, total_free = 0, total = 0;
647	unsigned long hole_start, hole_end, hole_size;
648
649	hole_start = drm_mm_hole_node_start(&mm->head_node);
650	hole_end = drm_mm_hole_node_end(&mm->head_node);
651	hole_size = hole_end - hole_start;
652	if (hole_size)
653		printk(KERN_DEBUG "%s 0x%08lx-0x%08lx: %8lu: free\n",
654			prefix, hole_start, hole_end,
655			hole_size);
656	total_free += hole_size;
657
658	drm_mm_for_each_node(entry, mm) {
659		printk(KERN_DEBUG "%s 0x%08lx-0x%08lx: %8lu: used\n",
660			prefix, entry->start, entry->start + entry->size,
661			entry->size);
662		total_used += entry->size;
663
664		if (entry->hole_follows) {
665			hole_start = drm_mm_hole_node_start(entry);
666			hole_end = drm_mm_hole_node_end(entry);
667			hole_size = hole_end - hole_start;
668			printk(KERN_DEBUG "%s 0x%08lx-0x%08lx: %8lu: free\n",
669				prefix, hole_start, hole_end,
670				hole_size);
671			total_free += hole_size;
672		}
673	}
674	total = total_free + total_used;
675
676	printk(KERN_DEBUG "%s total: %lu, used %lu free %lu\n", prefix, total,
677		total_used, total_free);
678}
679EXPORT_SYMBOL(drm_mm_debug_table);
680
681#if defined(CONFIG_DEBUG_FS)
682int drm_mm_dump_table(struct seq_file *m, struct drm_mm *mm)
 
 
683{
684	struct drm_mm_node *entry;
685	unsigned long total_used = 0, total_free = 0, total = 0;
686	unsigned long hole_start, hole_end, hole_size;
687
688	hole_start = drm_mm_hole_node_start(&mm->head_node);
689	hole_end = drm_mm_hole_node_end(&mm->head_node);
690	hole_size = hole_end - hole_start;
691	if (hole_size)
692		seq_printf(m, "0x%08lx-0x%08lx: 0x%08lx: free\n",
693				hole_start, hole_end, hole_size);
694	total_free += hole_size;
695
696	drm_mm_for_each_node(entry, mm) {
697		seq_printf(m, "0x%08lx-0x%08lx: 0x%08lx: used\n",
698				entry->start, entry->start + entry->size,
699				entry->size);
700		total_used += entry->size;
701		if (entry->hole_follows) {
702			hole_start = drm_mm_hole_node_start(entry);
703			hole_end = drm_mm_hole_node_end(entry);
704			hole_size = hole_end - hole_start;
705			seq_printf(m, "0x%08lx-0x%08lx: 0x%08lx: free\n",
706					hole_start, hole_end, hole_size);
707			total_free += hole_size;
708		}
709	}
710	total = total_free + total_used;
711
712	seq_printf(m, "total: %lu, used %lu free %lu\n", total, total_used, total_free);
713	return 0;
714}
715EXPORT_SYMBOL(drm_mm_dump_table);
716#endif
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);