drm_mm.c

   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#include <drm/drm_print.h>
  53
  54/**
  55 * DOC: Overview
  56 *
  57 * drm_mm provides a simple range allocator. The drivers are free to use the
  58 * resource allocator from the linux core if it suits them, the upside of drm_mm
  59 * is that it's in the DRM core. Which means that it's easier to extend for
  60 * some of the crazier special purpose needs of gpus.
  61 *
  62 * The main data struct is &drm_mm, allocations are tracked in &drm_mm_node.
  63 * Drivers are free to embed either of them into their own suitable
  64 * datastructures. drm_mm itself will not do any memory allocations of its own,
  65 * so if drivers choose not to embed nodes they need to still allocate them
  66 * themselves.
  67 *
  68 * The range allocator also supports reservation of preallocated blocks. This is
  69 * useful for taking over initial mode setting configurations from the firmware,
  70 * where an object needs to be created which exactly matches the firmware's
  71 * scanout target. As long as the range is still free it can be inserted anytime
  72 * after the allocator is initialized, which helps with avoiding looped
  73 * dependencies in the driver load sequence.
  74 *
  75 * drm_mm maintains a stack of most recently freed holes, which of all
  76 * simplistic datastructures seems to be a fairly decent approach to clustering
  77 * allocations and avoiding too much fragmentation. This means free space
  78 * searches are O(num_holes). Given that all the fancy features drm_mm supports
  79 * something better would be fairly complex and since gfx thrashing is a fairly
  80 * steep cliff not a real concern. Removing a node again is O(1).
  81 *
  82 * drm_mm supports a few features: Alignment and range restrictions can be
  83 * supplied. Furthermore every &drm_mm_node has a color value (which is just an
  84 * opaque unsigned long) which in conjunction with a driver callback can be used
  85 * to implement sophisticated placement restrictions. The i915 DRM driver uses
  86 * this to implement guard pages between incompatible caching domains in the
  87 * graphics TT.
  88 *
  89 * Two behaviors are supported for searching and allocating: bottom-up and
  90 * top-down. The default is bottom-up. Top-down allocation can be used if the
  91 * memory area has different restrictions, or just to reduce fragmentation.
  92 *
  93 * Finally iteration helpers to walk all nodes and all holes are provided as are
  94 * some basic allocator dumpers for debugging.
  95 *
  96 * Note that this range allocator is not thread-safe, drivers need to protect
  97 * modifications with their own locking. The idea behind this is that for a full
  98 * memory manager additional data needs to be protected anyway, hence internal
  99 * locking would be fully redundant.
 100 */
 101
 102#ifdef CONFIG_DRM_DEBUG_MM
 103#include <linux/stackdepot.h>
 104
 105#define STACKDEPTH 32
 106#define BUFSZ 4096
 107
 108static noinline void save_stack(struct drm_mm_node *node)
 109{
 110	unsigned long entries[STACKDEPTH];
 111	unsigned int n;
 112
 113	n = stack_trace_save(entries, ARRAY_SIZE(entries), 1);
 114
 115	/* May be called under spinlock, so avoid sleeping */
 116	node->stack = stack_depot_save(entries, n, GFP_NOWAIT);
 117}
 118
 119static void show_leaks(struct drm_mm *mm)
 120{
 121	struct drm_mm_node *node;
 122	char *buf;
 123
 124	buf = kmalloc(BUFSZ, GFP_KERNEL);
 125	if (!buf)
 126		return;
 127
 128	list_for_each_entry(node, drm_mm_nodes(mm), node_list) {
 129		if (!node->stack) {
 130			DRM_ERROR("node [%08llx + %08llx]: unknown owner\n",
 131				  node->start, node->size);
 132			continue;
 133		}
 134
 135		stack_depot_snprint(node->stack, buf, BUFSZ, 0);
 136		DRM_ERROR("node [%08llx + %08llx]: inserted at\n%s",
 137			  node->start, node->size, buf);
 138	}
 139
 140	kfree(buf);
 141}
 142
 143#undef STACKDEPTH
 144#undef BUFSZ
 145#else
 146static void save_stack(struct drm_mm_node *node) { }
 147static void show_leaks(struct drm_mm *mm) { }
 148#endif
 149
 150#define START(node) ((node)->start)
 151#define LAST(node)  ((node)->start + (node)->size - 1)
 152
 153INTERVAL_TREE_DEFINE(struct drm_mm_node, rb,
 154		     u64, __subtree_last,
 155		     START, LAST, static inline __maybe_unused, drm_mm_interval_tree)
 156
 157struct drm_mm_node *
 158__drm_mm_interval_first(const struct drm_mm *mm, u64 start, u64 last)
 159{
 160	return drm_mm_interval_tree_iter_first((struct rb_root_cached *)&mm->interval_tree,
 161					       start, last) ?: (struct drm_mm_node *)&mm->head_node;
 162}
 163EXPORT_SYMBOL(__drm_mm_interval_first);
 164
 165static void drm_mm_interval_tree_add_node(struct drm_mm_node *hole_node,
 166					  struct drm_mm_node *node)
 167{
 168	struct drm_mm *mm = hole_node->mm;
 169	struct rb_node **link, *rb;
 170	struct drm_mm_node *parent;
 171	bool leftmost;
 172
 173	node->__subtree_last = LAST(node);
 174
 175	if (drm_mm_node_allocated(hole_node)) {
 176		rb = &hole_node->rb;
 177		while (rb) {
 178			parent = rb_entry(rb, struct drm_mm_node, rb);
 179			if (parent->__subtree_last >= node->__subtree_last)
 180				break;
 181
 182			parent->__subtree_last = node->__subtree_last;
 183			rb = rb_parent(rb);
 184		}
 185
 186		rb = &hole_node->rb;
 187		link = &hole_node->rb.rb_right;
 188		leftmost = false;
 189	} else {
 190		rb = NULL;
 191		link = &mm->interval_tree.rb_root.rb_node;
 192		leftmost = true;
 193	}
 194
 195	while (*link) {
 196		rb = *link;
 197		parent = rb_entry(rb, struct drm_mm_node, rb);
 198		if (parent->__subtree_last < node->__subtree_last)
 199			parent->__subtree_last = node->__subtree_last;
 200		if (node->start < parent->start) {
 201			link = &parent->rb.rb_left;
 202		} else {
 203			link = &parent->rb.rb_right;
 204			leftmost = false;
 205		}
 206	}
 207
 208	rb_link_node(&node->rb, rb, link);
 209	rb_insert_augmented_cached(&node->rb, &mm->interval_tree, leftmost,
 210				   &drm_mm_interval_tree_augment);
 211}
 212
 213#define HOLE_SIZE(NODE) ((NODE)->hole_size)
 214#define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE))
 215
 216static u64 rb_to_hole_size(struct rb_node *rb)
 217{
 218	return rb_entry(rb, struct drm_mm_node, rb_hole_size)->hole_size;
 219}
 220
 221static void insert_hole_size(struct rb_root_cached *root,
 222			     struct drm_mm_node *node)
 223{
 224	struct rb_node **link = &root->rb_root.rb_node, *rb = NULL;
 225	u64 x = node->hole_size;
 226	bool first = true;
 227
 228	while (*link) {
 229		rb = *link;
 230		if (x > rb_to_hole_size(rb)) {
 231			link = &rb->rb_left;
 232		} else {
 233			link = &rb->rb_right;
 234			first = false;
 235		}
 236	}
 237
 238	rb_link_node(&node->rb_hole_size, rb, link);
 239	rb_insert_color_cached(&node->rb_hole_size, root, first);
 240}
 241
 242RB_DECLARE_CALLBACKS_MAX(static, augment_callbacks,
 243			 struct drm_mm_node, rb_hole_addr,
 244			 u64, subtree_max_hole, HOLE_SIZE)
 245
 246static void insert_hole_addr(struct rb_root *root, struct drm_mm_node *node)
 247{
 248	struct rb_node **link = &root->rb_node, *rb_parent = NULL;
 249	u64 start = HOLE_ADDR(node), subtree_max_hole = node->subtree_max_hole;
 250	struct drm_mm_node *parent;
 251
 252	while (*link) {
 253		rb_parent = *link;
 254		parent = rb_entry(rb_parent, struct drm_mm_node, rb_hole_addr);
 255		if (parent->subtree_max_hole < subtree_max_hole)
 256			parent->subtree_max_hole = subtree_max_hole;
 257		if (start < HOLE_ADDR(parent))
 258			link = &parent->rb_hole_addr.rb_left;
 259		else
 260			link = &parent->rb_hole_addr.rb_right;
 261	}
 262
 263	rb_link_node(&node->rb_hole_addr, rb_parent, link);
 264	rb_insert_augmented(&node->rb_hole_addr, root, &augment_callbacks);
 265}
 266
 267static void add_hole(struct drm_mm_node *node)
 268{
 269	struct drm_mm *mm = node->mm;
 270
 271	node->hole_size =
 272		__drm_mm_hole_node_end(node) - __drm_mm_hole_node_start(node);
 273	node->subtree_max_hole = node->hole_size;
 274	DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
 275
 276	insert_hole_size(&mm->holes_size, node);
 277	insert_hole_addr(&mm->holes_addr, node);
 278
 279	list_add(&node->hole_stack, &mm->hole_stack);
 280}
 281
 282static void rm_hole(struct drm_mm_node *node)
 283{
 284	DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
 285
 286	list_del(&node->hole_stack);
 287	rb_erase_cached(&node->rb_hole_size, &node->mm->holes_size);
 288	rb_erase_augmented(&node->rb_hole_addr, &node->mm->holes_addr,
 289			   &augment_callbacks);
 290	node->hole_size = 0;
 291	node->subtree_max_hole = 0;
 292
 293	DRM_MM_BUG_ON(drm_mm_hole_follows(node));
 294}
 295
 296static inline struct drm_mm_node *rb_hole_size_to_node(struct rb_node *rb)
 297{
 298	return rb_entry_safe(rb, struct drm_mm_node, rb_hole_size);
 299}
 300
 301static inline struct drm_mm_node *rb_hole_addr_to_node(struct rb_node *rb)
 302{
 303	return rb_entry_safe(rb, struct drm_mm_node, rb_hole_addr);
 304}
 305
 306static struct drm_mm_node *best_hole(struct drm_mm *mm, u64 size)
 307{
 308	struct rb_node *rb = mm->holes_size.rb_root.rb_node;
 309	struct drm_mm_node *best = NULL;
 310
 311	do {
 312		struct drm_mm_node *node =
 313			rb_entry(rb, struct drm_mm_node, rb_hole_size);
 314
 315		if (size <= node->hole_size) {
 316			best = node;
 317			rb = rb->rb_right;
 318		} else {
 319			rb = rb->rb_left;
 320		}
 321	} while (rb);
 322
 323	return best;
 324}
 325
 326static bool usable_hole_addr(struct rb_node *rb, u64 size)
 327{
 328	return rb && rb_hole_addr_to_node(rb)->subtree_max_hole >= size;
 329}
 330
 331static struct drm_mm_node *find_hole_addr(struct drm_mm *mm, u64 addr, u64 size)
 332{
 333	struct rb_node *rb = mm->holes_addr.rb_node;
 334	struct drm_mm_node *node = NULL;
 335
 336	while (rb) {
 337		u64 hole_start;
 338
 339		if (!usable_hole_addr(rb, size))
 340			break;
 341
 342		node = rb_hole_addr_to_node(rb);
 343		hole_start = __drm_mm_hole_node_start(node);
 344
 345		if (addr < hole_start)
 346			rb = node->rb_hole_addr.rb_left;
 347		else if (addr > hole_start + node->hole_size)
 348			rb = node->rb_hole_addr.rb_right;
 349		else
 350			break;
 351	}
 352
 353	return node;
 354}
 355
 356static struct drm_mm_node *
 357first_hole(struct drm_mm *mm,
 358	   u64 start, u64 end, u64 size,
 359	   enum drm_mm_insert_mode mode)
 360{
 361	switch (mode) {
 362	default:
 363	case DRM_MM_INSERT_BEST:
 364		return best_hole(mm, size);
 365
 366	case DRM_MM_INSERT_LOW:
 367		return find_hole_addr(mm, start, size);
 368
 369	case DRM_MM_INSERT_HIGH:
 370		return find_hole_addr(mm, end, size);
 371
 372	case DRM_MM_INSERT_EVICT:
 373		return list_first_entry_or_null(&mm->hole_stack,
 374						struct drm_mm_node,
 375						hole_stack);
 376	}
 377}
 378
 379/**
 380 * DECLARE_NEXT_HOLE_ADDR - macro to declare next hole functions
 381 * @name: name of function to declare
 382 * @first: first rb member to traverse (either rb_left or rb_right).
 383 * @last: last rb member to traverse (either rb_right or rb_left).
 384 *
 385 * This macro declares a function to return the next hole of the addr rb tree.
 386 * While traversing the tree we take the searched size into account and only
 387 * visit branches with potential big enough holes.
 388 */
 389
 390#define DECLARE_NEXT_HOLE_ADDR(name, first, last)			\
 391static struct drm_mm_node *name(struct drm_mm_node *entry, u64 size)	\
 392{									\
 393	struct rb_node *parent, *node = &entry->rb_hole_addr;		\
 394									\
 395	if (!entry || RB_EMPTY_NODE(node))				\
 396		return NULL;						\
 397									\
 398	if (usable_hole_addr(node->first, size)) {			\
 399		node = node->first;					\
 400		while (usable_hole_addr(node->last, size))		\
 401			node = node->last;				\
 402		return rb_hole_addr_to_node(node);			\
 403	}								\
 404									\
 405	while ((parent = rb_parent(node)) && node == parent->first)	\
 406		node = parent;						\
 407									\
 408	return rb_hole_addr_to_node(parent);				\
 409}
 410
 411DECLARE_NEXT_HOLE_ADDR(next_hole_high_addr, rb_left, rb_right)
 412DECLARE_NEXT_HOLE_ADDR(next_hole_low_addr, rb_right, rb_left)
 413
 414static struct drm_mm_node *
 415next_hole(struct drm_mm *mm,
 416	  struct drm_mm_node *node,
 417	  u64 size,
 418	  enum drm_mm_insert_mode mode)
 419{
 420	switch (mode) {
 421	default:
 422	case DRM_MM_INSERT_BEST:
 423		return rb_hole_size_to_node(rb_prev(&node->rb_hole_size));
 424
 425	case DRM_MM_INSERT_LOW:
 426		return next_hole_low_addr(node, size);
 427
 428	case DRM_MM_INSERT_HIGH:
 429		return next_hole_high_addr(node, size);
 430
 431	case DRM_MM_INSERT_EVICT:
 432		node = list_next_entry(node, hole_stack);
 433		return &node->hole_stack == &mm->hole_stack ? NULL : node;
 434	}
 435}
 436
 437/**
 438 * drm_mm_reserve_node - insert an pre-initialized node
 439 * @mm: drm_mm allocator to insert @node into
 440 * @node: drm_mm_node to insert
 441 *
 442 * This functions inserts an already set-up &drm_mm_node into the allocator,
 443 * meaning that start, size and color must be set by the caller. All other
 444 * fields must be cleared to 0. This is useful to initialize the allocator with
 445 * preallocated objects which must be set-up before the range allocator can be
 446 * set-up, e.g. when taking over a firmware framebuffer.
 447 *
 448 * Returns:
 449 * 0 on success, -ENOSPC if there's no hole where @node is.
 450 */
 451int drm_mm_reserve_node(struct drm_mm *mm, struct drm_mm_node *node)
 452{
 453	struct drm_mm_node *hole;
 454	u64 hole_start, hole_end;
 455	u64 adj_start, adj_end;
 456	u64 end;
 457
 458	end = node->start + node->size;
 459	if (unlikely(end <= node->start))
 460		return -ENOSPC;
 461
 462	/* Find the relevant hole to add our node to */
 463	hole = find_hole_addr(mm, node->start, 0);
 464	if (!hole)
 465		return -ENOSPC;
 466
 467	adj_start = hole_start = __drm_mm_hole_node_start(hole);
 468	adj_end = hole_end = hole_start + hole->hole_size;
 469
 470	if (mm->color_adjust)
 471		mm->color_adjust(hole, node->color, &adj_start, &adj_end);
 472
 473	if (adj_start > node->start || adj_end < end)
 474		return -ENOSPC;
 475
 476	node->mm = mm;
 477
 478	__set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
 479	list_add(&node->node_list, &hole->node_list);
 480	drm_mm_interval_tree_add_node(hole, node);
 481	node->hole_size = 0;
 482
 483	rm_hole(hole);
 484	if (node->start > hole_start)
 485		add_hole(hole);
 486	if (end < hole_end)
 487		add_hole(node);
 488
 489	save_stack(node);
 490	return 0;
 491}
 492EXPORT_SYMBOL(drm_mm_reserve_node);
 493
 494static u64 rb_to_hole_size_or_zero(struct rb_node *rb)
 495{
 496	return rb ? rb_to_hole_size(rb) : 0;
 497}
 498
 499/**
 500 * drm_mm_insert_node_in_range - ranged search for space and insert @node
 501 * @mm: drm_mm to allocate from
 502 * @node: preallocate node to insert
 503 * @size: size of the allocation
 504 * @alignment: alignment of the allocation
 505 * @color: opaque tag value to use for this node
 506 * @range_start: start of the allowed range for this node
 507 * @range_end: end of the allowed range for this node
 508 * @mode: fine-tune the allocation search and placement
 509 *
 510 * The preallocated @node must be cleared to 0.
 511 *
 512 * Returns:
 513 * 0 on success, -ENOSPC if there's no suitable hole.
 514 */
 515int drm_mm_insert_node_in_range(struct drm_mm * const mm,
 516				struct drm_mm_node * const node,
 517				u64 size, u64 alignment,
 518				unsigned long color,
 519				u64 range_start, u64 range_end,
 520				enum drm_mm_insert_mode mode)
 521{
 522	struct drm_mm_node *hole;
 523	u64 remainder_mask;
 524	bool once;
 525
 526	DRM_MM_BUG_ON(range_start > range_end);
 527
 528	if (unlikely(size == 0 || range_end - range_start < size))
 529		return -ENOSPC;
 530
 531	if (rb_to_hole_size_or_zero(rb_first_cached(&mm->holes_size)) < size)
 532		return -ENOSPC;
 533
 534	if (alignment <= 1)
 535		alignment = 0;
 536
 537	once = mode & DRM_MM_INSERT_ONCE;
 538	mode &= ~DRM_MM_INSERT_ONCE;
 539
 540	remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
 541	for (hole = first_hole(mm, range_start, range_end, size, mode);
 542	     hole;
 543	     hole = once ? NULL : next_hole(mm, hole, size, mode)) {
 544		u64 hole_start = __drm_mm_hole_node_start(hole);
 545		u64 hole_end = hole_start + hole->hole_size;
 546		u64 adj_start, adj_end;
 547		u64 col_start, col_end;
 548
 549		if (mode == DRM_MM_INSERT_LOW && hole_start >= range_end)
 550			break;
 551
 552		if (mode == DRM_MM_INSERT_HIGH && hole_end <= range_start)
 553			break;
 554
 555		col_start = hole_start;
 556		col_end = hole_end;
 557		if (mm->color_adjust)
 558			mm->color_adjust(hole, color, &col_start, &col_end);
 559
 560		adj_start = max(col_start, range_start);
 561		adj_end = min(col_end, range_end);
 562
 563		if (adj_end <= adj_start || adj_end - adj_start < size)
 564			continue;
 565
 566		if (mode == DRM_MM_INSERT_HIGH)
 567			adj_start = adj_end - size;
 568
 569		if (alignment) {
 570			u64 rem;
 571
 572			if (likely(remainder_mask))
 573				rem = adj_start & remainder_mask;
 574			else
 575				div64_u64_rem(adj_start, alignment, &rem);
 576			if (rem) {
 577				adj_start -= rem;
 578				if (mode != DRM_MM_INSERT_HIGH)
 579					adj_start += alignment;
 580
 581				if (adj_start < max(col_start, range_start) ||
 582				    min(col_end, range_end) - adj_start < size)
 583					continue;
 584
 585				if (adj_end <= adj_start ||
 586				    adj_end - adj_start < size)
 587					continue;
 588			}
 589		}
 590
 591		node->mm = mm;
 592		node->size = size;
 593		node->start = adj_start;
 594		node->color = color;
 595		node->hole_size = 0;
 596
 597		__set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
 598		list_add(&node->node_list, &hole->node_list);
 599		drm_mm_interval_tree_add_node(hole, node);
 600
 601		rm_hole(hole);
 602		if (adj_start > hole_start)
 603			add_hole(hole);
 604		if (adj_start + size < hole_end)
 605			add_hole(node);
 606
 607		save_stack(node);
 608		return 0;
 609	}
 610
 611	return -ENOSPC;
 612}
 613EXPORT_SYMBOL(drm_mm_insert_node_in_range);
 614
 615static inline __maybe_unused bool drm_mm_node_scanned_block(const struct drm_mm_node *node)
 616{
 617	return test_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
 618}
 619
 620/**
 621 * drm_mm_remove_node - Remove a memory node from the allocator.
 622 * @node: drm_mm_node to remove
 623 *
 624 * This just removes a node from its drm_mm allocator. The node does not need to
 625 * be cleared again before it can be re-inserted into this or any other drm_mm
 626 * allocator. It is a bug to call this function on a unallocated node.
 627 */
 628void drm_mm_remove_node(struct drm_mm_node *node)
 629{
 630	struct drm_mm *mm = node->mm;
 631	struct drm_mm_node *prev_node;
 632
 633	DRM_MM_BUG_ON(!drm_mm_node_allocated(node));
 634	DRM_MM_BUG_ON(drm_mm_node_scanned_block(node));
 635
 636	prev_node = list_prev_entry(node, node_list);
 637
 638	if (drm_mm_hole_follows(node))
 639		rm_hole(node);
 640
 641	drm_mm_interval_tree_remove(node, &mm->interval_tree);
 642	list_del(&node->node_list);
 643
 644	if (drm_mm_hole_follows(prev_node))
 645		rm_hole(prev_node);
 646	add_hole(prev_node);
 647
 648	clear_bit_unlock(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
 649}
 650EXPORT_SYMBOL(drm_mm_remove_node);
 651
 652/**
 653 * DOC: lru scan roster
 654 *
 655 * Very often GPUs need to have continuous allocations for a given object. When
 656 * evicting objects to make space for a new one it is therefore not most
 657 * efficient when we simply start to select all objects from the tail of an LRU
 658 * until there's a suitable hole: Especially for big objects or nodes that
 659 * otherwise have special allocation constraints there's a good chance we evict
 660 * lots of (smaller) objects unnecessarily.
 661 *
 662 * The DRM range allocator supports this use-case through the scanning
 663 * interfaces. First a scan operation needs to be initialized with
 664 * drm_mm_scan_init() or drm_mm_scan_init_with_range(). The driver adds
 665 * objects to the roster, probably by walking an LRU list, but this can be
 666 * freely implemented. Eviction candidates are added using
 667 * drm_mm_scan_add_block() until a suitable hole is found or there are no
 668 * further evictable objects. Eviction roster metadata is tracked in &struct
 669 * drm_mm_scan.
 670 *
 671 * The driver must walk through all objects again in exactly the reverse
 672 * order to restore the allocator state. Note that while the allocator is used
 673 * in the scan mode no other operation is allowed.
 674 *
 675 * Finally the driver evicts all objects selected (drm_mm_scan_remove_block()
 676 * reported true) in the scan, and any overlapping nodes after color adjustment
 677 * (drm_mm_scan_color_evict()). Adding and removing an object is O(1), and
 678 * since freeing a node is also O(1) the overall complexity is
 679 * O(scanned_objects). So like the free stack which needs to be walked before a
 680 * scan operation even begins this is linear in the number of objects. It
 681 * doesn't seem to hurt too badly.
 682 */
 683
 684/**
 685 * drm_mm_scan_init_with_range - initialize range-restricted lru scanning
 686 * @scan: scan state
 687 * @mm: drm_mm to scan
 688 * @size: size of the allocation
 689 * @alignment: alignment of the allocation
 690 * @color: opaque tag value to use for the allocation
 691 * @start: start of the allowed range for the allocation
 692 * @end: end of the allowed range for the allocation
 693 * @mode: fine-tune the allocation search and placement
 694 *
 695 * This simply sets up the scanning routines with the parameters for the desired
 696 * hole.
 697 *
 698 * Warning:
 699 * As long as the scan list is non-empty, no other operations than
 700 * adding/removing nodes to/from the scan list are allowed.
 701 */
 702void drm_mm_scan_init_with_range(struct drm_mm_scan *scan,
 703				 struct drm_mm *mm,
 704				 u64 size,
 705				 u64 alignment,
 706				 unsigned long color,
 707				 u64 start,
 708				 u64 end,
 709				 enum drm_mm_insert_mode mode)
 710{
 711	DRM_MM_BUG_ON(start >= end);
 712	DRM_MM_BUG_ON(!size || size > end - start);
 713	DRM_MM_BUG_ON(mm->scan_active);
 714
 715	scan->mm = mm;
 716
 717	if (alignment <= 1)
 718		alignment = 0;
 719
 720	scan->color = color;
 721	scan->alignment = alignment;
 722	scan->remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
 723	scan->size = size;
 724	scan->mode = mode;
 725
 726	DRM_MM_BUG_ON(end <= start);
 727	scan->range_start = start;
 728	scan->range_end = end;
 729
 730	scan->hit_start = U64_MAX;
 731	scan->hit_end = 0;
 732}
 733EXPORT_SYMBOL(drm_mm_scan_init_with_range);
 734
 735/**
 736 * drm_mm_scan_add_block - add a node to the scan list
 737 * @scan: the active drm_mm scanner
 738 * @node: drm_mm_node to add
 739 *
 740 * Add a node to the scan list that might be freed to make space for the desired
 741 * hole.
 742 *
 743 * Returns:
 744 * True if a hole has been found, false otherwise.
 745 */
 746bool drm_mm_scan_add_block(struct drm_mm_scan *scan,
 747			   struct drm_mm_node *node)
 748{
 749	struct drm_mm *mm = scan->mm;
 750	struct drm_mm_node *hole;
 751	u64 hole_start, hole_end;
 752	u64 col_start, col_end;
 753	u64 adj_start, adj_end;
 754
 755	DRM_MM_BUG_ON(node->mm != mm);
 756	DRM_MM_BUG_ON(!drm_mm_node_allocated(node));
 757	DRM_MM_BUG_ON(drm_mm_node_scanned_block(node));
 758	__set_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
 759	mm->scan_active++;
 760
 761	/* Remove this block from the node_list so that we enlarge the hole
 762	 * (distance between the end of our previous node and the start of
 763	 * or next), without poisoning the link so that we can restore it
 764	 * later in drm_mm_scan_remove_block().
 765	 */
 766	hole = list_prev_entry(node, node_list);
 767	DRM_MM_BUG_ON(list_next_entry(hole, node_list) != node);
 768	__list_del_entry(&node->node_list);
 769
 770	hole_start = __drm_mm_hole_node_start(hole);
 771	hole_end = __drm_mm_hole_node_end(hole);
 772
 773	col_start = hole_start;
 774	col_end = hole_end;
 775	if (mm->color_adjust)
 776		mm->color_adjust(hole, scan->color, &col_start, &col_end);
 777
 778	adj_start = max(col_start, scan->range_start);
 779	adj_end = min(col_end, scan->range_end);
 780	if (adj_end <= adj_start || adj_end - adj_start < scan->size)
 781		return false;
 782
 783	if (scan->mode == DRM_MM_INSERT_HIGH)
 784		adj_start = adj_end - scan->size;
 785
 786	if (scan->alignment) {
 787		u64 rem;
 788
 789		if (likely(scan->remainder_mask))
 790			rem = adj_start & scan->remainder_mask;
 791		else
 792			div64_u64_rem(adj_start, scan->alignment, &rem);
 793		if (rem) {
 794			adj_start -= rem;
 795			if (scan->mode != DRM_MM_INSERT_HIGH)
 796				adj_start += scan->alignment;
 797			if (adj_start < max(col_start, scan->range_start) ||
 798			    min(col_end, scan->range_end) - adj_start < scan->size)
 799				return false;
 800
 801			if (adj_end <= adj_start ||
 802			    adj_end - adj_start < scan->size)
 803				return false;
 804		}
 805	}
 806
 807	scan->hit_start = adj_start;
 808	scan->hit_end = adj_start + scan->size;
 809
 810	DRM_MM_BUG_ON(scan->hit_start >= scan->hit_end);
 811	DRM_MM_BUG_ON(scan->hit_start < hole_start);
 812	DRM_MM_BUG_ON(scan->hit_end > hole_end);
 813
 814	return true;
 815}
 816EXPORT_SYMBOL(drm_mm_scan_add_block);
 817
 818/**
 819 * drm_mm_scan_remove_block - remove a node from the scan list
 820 * @scan: the active drm_mm scanner
 821 * @node: drm_mm_node to remove
 822 *
 823 * Nodes **must** be removed in exactly the reverse order from the scan list as
 824 * they have been added (e.g. using list_add() as they are added and then
 825 * list_for_each() over that eviction list to remove), otherwise the internal
 826 * state of the memory manager will be corrupted.
 827 *
 828 * When the scan list is empty, the selected memory nodes can be freed. An
 829 * immediately following drm_mm_insert_node_in_range_generic() or one of the
 830 * simpler versions of that function with !DRM_MM_SEARCH_BEST will then return
 831 * the just freed block (because it's at the top of the free_stack list).
 832 *
 833 * Returns:
 834 * True if this block should be evicted, false otherwise. Will always
 835 * return false when no hole has been found.
 836 */
 837bool drm_mm_scan_remove_block(struct drm_mm_scan *scan,
 838			      struct drm_mm_node *node)
 839{
 840	struct drm_mm_node *prev_node;
 841
 842	DRM_MM_BUG_ON(node->mm != scan->mm);
 843	DRM_MM_BUG_ON(!drm_mm_node_scanned_block(node));
 844	__clear_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
 845
 846	DRM_MM_BUG_ON(!node->mm->scan_active);
 847	node->mm->scan_active--;
 848
 849	/* During drm_mm_scan_add_block() we decoupled this node leaving
 850	 * its pointers intact. Now that the caller is walking back along
 851	 * the eviction list we can restore this block into its rightful
 852	 * place on the full node_list. To confirm that the caller is walking
 853	 * backwards correctly we check that prev_node->next == node->next,
 854	 * i.e. both believe the same node should be on the other side of the
 855	 * hole.
 856	 */
 857	prev_node = list_prev_entry(node, node_list);
 858	DRM_MM_BUG_ON(list_next_entry(prev_node, node_list) !=
 859		      list_next_entry(node, node_list));
 860	list_add(&node->node_list, &prev_node->node_list);
 861
 862	return (node->start + node->size > scan->hit_start &&
 863		node->start < scan->hit_end);
 864}
 865EXPORT_SYMBOL(drm_mm_scan_remove_block);
 866
 867/**
 868 * drm_mm_scan_color_evict - evict overlapping nodes on either side of hole
 869 * @scan: drm_mm scan with target hole
 870 *
 871 * After completing an eviction scan and removing the selected nodes, we may
 872 * need to remove a few more nodes from either side of the target hole if
 873 * mm.color_adjust is being used.
 874 *
 875 * Returns:
 876 * A node to evict, or NULL if there are no overlapping nodes.
 877 */
 878struct drm_mm_node *drm_mm_scan_color_evict(struct drm_mm_scan *scan)
 879{
 880	struct drm_mm *mm = scan->mm;
 881	struct drm_mm_node *hole;
 882	u64 hole_start, hole_end;
 883
 884	DRM_MM_BUG_ON(list_empty(&mm->hole_stack));
 885
 886	if (!mm->color_adjust)
 887		return NULL;
 888
 889	/*
 890	 * The hole found during scanning should ideally be the first element
 891	 * in the hole_stack list, but due to side-effects in the driver it
 892	 * may not be.
 893	 */
 894	list_for_each_entry(hole, &mm->hole_stack, hole_stack) {
 895		hole_start = __drm_mm_hole_node_start(hole);
 896		hole_end = hole_start + hole->hole_size;
 897
 898		if (hole_start <= scan->hit_start &&
 899		    hole_end >= scan->hit_end)
 900			break;
 901	}
 902
 903	/* We should only be called after we found the hole previously */
 904	DRM_MM_BUG_ON(&hole->hole_stack == &mm->hole_stack);
 905	if (unlikely(&hole->hole_stack == &mm->hole_stack))
 906		return NULL;
 907
 908	DRM_MM_BUG_ON(hole_start > scan->hit_start);
 909	DRM_MM_BUG_ON(hole_end < scan->hit_end);
 910
 911	mm->color_adjust(hole, scan->color, &hole_start, &hole_end);
 912	if (hole_start > scan->hit_start)
 913		return hole;
 914	if (hole_end < scan->hit_end)
 915		return list_next_entry(hole, node_list);
 916
 917	return NULL;
 918}
 919EXPORT_SYMBOL(drm_mm_scan_color_evict);
 920
 921/**
 922 * drm_mm_init - initialize a drm-mm allocator
 923 * @mm: the drm_mm structure to initialize
 924 * @start: start of the range managed by @mm
 925 * @size: end of the range managed by @mm
 926 *
 927 * Note that @mm must be cleared to 0 before calling this function.
 928 */
 929void drm_mm_init(struct drm_mm *mm, u64 start, u64 size)
 930{
 931	DRM_MM_BUG_ON(start + size <= start);
 932
 933	mm->color_adjust = NULL;
 934
 935	INIT_LIST_HEAD(&mm->hole_stack);
 936	mm->interval_tree = RB_ROOT_CACHED;
 937	mm->holes_size = RB_ROOT_CACHED;
 938	mm->holes_addr = RB_ROOT;
 939
 940	/* Clever trick to avoid a special case in the free hole tracking. */
 941	INIT_LIST_HEAD(&mm->head_node.node_list);
 942	mm->head_node.flags = 0;
 943	mm->head_node.mm = mm;
 944	mm->head_node.start = start + size;
 945	mm->head_node.size = -size;
 946	add_hole(&mm->head_node);
 947
 948	mm->scan_active = 0;
 949
 950#ifdef CONFIG_DRM_DEBUG_MM
 951	stack_depot_init();
 952#endif
 953}
 954EXPORT_SYMBOL(drm_mm_init);
 955
 956/**
 957 * drm_mm_takedown - clean up a drm_mm allocator
 958 * @mm: drm_mm allocator to clean up
 959 *
 960 * Note that it is a bug to call this function on an allocator which is not
 961 * clean.
 962 */
 963void drm_mm_takedown(struct drm_mm *mm)
 964{
 965	if (WARN(!drm_mm_clean(mm),
 966		 "Memory manager not clean during takedown.\n"))
 967		show_leaks(mm);
 968}
 969EXPORT_SYMBOL(drm_mm_takedown);
 970
 971static u64 drm_mm_dump_hole(struct drm_printer *p, const struct drm_mm_node *entry)
 972{
 973	u64 start, size;
 974
 975	size = entry->hole_size;
 976	if (size) {
 977		start = drm_mm_hole_node_start(entry);
 978		drm_printf(p, "%#018llx-%#018llx: %llu: free\n",
 979			   start, start + size, size);
 980	}
 981
 982	return size;
 983}
 984/**
 985 * drm_mm_print - print allocator state
 986 * @mm: drm_mm allocator to print
 987 * @p: DRM printer to use
 988 */
 989void drm_mm_print(const struct drm_mm *mm, struct drm_printer *p)
 990{
 991	const struct drm_mm_node *entry;
 992	u64 total_used = 0, total_free = 0, total = 0;
 993
 994	total_free += drm_mm_dump_hole(p, &mm->head_node);
 995
 996	drm_mm_for_each_node(entry, mm) {
 997		drm_printf(p, "%#018llx-%#018llx: %llu: used\n", entry->start,
 998			   entry->start + entry->size, entry->size);
 999		total_used += entry->size;
1000		total_free += drm_mm_dump_hole(p, entry);
1001	}
1002	total = total_free + total_used;
1003
1004	drm_printf(p, "total: %llu, used %llu free %llu\n", total,
1005		   total_used, total_free);
1006}
1007EXPORT_SYMBOL(drm_mm_print);