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 (drm_mm_node_allocated(hole_node)) { 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 HOLE_SIZE(NODE) ((NODE)->hole_size) 216#define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE)) 217 218static u64 rb_to_hole_size(struct rb_node *rb) 219{ 220 return rb_entry(rb, struct drm_mm_node, rb_hole_size)->hole_size; 221} 222 223static void insert_hole_size(struct rb_root_cached *root, 224 struct drm_mm_node *node) 225{ 226 struct rb_node **link = &root->rb_root.rb_node, *rb = NULL; 227 u64 x = node->hole_size; 228 bool first = true; 229 230 while (*link) { 231 rb = *link; 232 if (x > rb_to_hole_size(rb)) { 233 link = &rb->rb_left; 234 } else { 235 link = &rb->rb_right; 236 first = false; 237 } 238 } 239 240 rb_link_node(&node->rb_hole_size, rb, link); 241 rb_insert_color_cached(&node->rb_hole_size, root, first); 242} 243 244RB_DECLARE_CALLBACKS_MAX(static, augment_callbacks, 245 struct drm_mm_node, rb_hole_addr, 246 u64, subtree_max_hole, HOLE_SIZE) 247 248static void insert_hole_addr(struct rb_root *root, struct drm_mm_node *node) 249{ 250 struct rb_node **link = &root->rb_node, *rb_parent = NULL; 251 u64 start = HOLE_ADDR(node), subtree_max_hole = node->subtree_max_hole; 252 struct drm_mm_node *parent; 253 254 while (*link) { 255 rb_parent = *link; 256 parent = rb_entry(rb_parent, struct drm_mm_node, rb_hole_addr); 257 if (parent->subtree_max_hole < subtree_max_hole) 258 parent->subtree_max_hole = subtree_max_hole; 259 if (start < HOLE_ADDR(parent)) 260 link = &parent->rb_hole_addr.rb_left; 261 else 262 link = &parent->rb_hole_addr.rb_right; 263 } 264 265 rb_link_node(&node->rb_hole_addr, rb_parent, link); 266 rb_insert_augmented(&node->rb_hole_addr, root, &augment_callbacks); 267} 268 269static void add_hole(struct drm_mm_node *node) 270{ 271 struct drm_mm *mm = node->mm; 272 273 node->hole_size = 274 __drm_mm_hole_node_end(node) - __drm_mm_hole_node_start(node); 275 node->subtree_max_hole = node->hole_size; 276 DRM_MM_BUG_ON(!drm_mm_hole_follows(node)); 277 278 insert_hole_size(&mm->holes_size, node); 279 insert_hole_addr(&mm->holes_addr, node); 280 281 list_add(&node->hole_stack, &mm->hole_stack); 282} 283 284static void rm_hole(struct drm_mm_node *node) 285{ 286 DRM_MM_BUG_ON(!drm_mm_hole_follows(node)); 287 288 list_del(&node->hole_stack); 289 rb_erase_cached(&node->rb_hole_size, &node->mm->holes_size); 290 rb_erase_augmented(&node->rb_hole_addr, &node->mm->holes_addr, 291 &augment_callbacks); 292 node->hole_size = 0; 293 node->subtree_max_hole = 0; 294 295 DRM_MM_BUG_ON(drm_mm_hole_follows(node)); 296} 297 298static inline struct drm_mm_node *rb_hole_size_to_node(struct rb_node *rb) 299{ 300 return rb_entry_safe(rb, struct drm_mm_node, rb_hole_size); 301} 302 303static inline struct drm_mm_node *rb_hole_addr_to_node(struct rb_node *rb) 304{ 305 return rb_entry_safe(rb, struct drm_mm_node, rb_hole_addr); 306} 307 308static struct drm_mm_node *best_hole(struct drm_mm *mm, u64 size) 309{ 310 struct rb_node *rb = mm->holes_size.rb_root.rb_node; 311 struct drm_mm_node *best = NULL; 312 313 do { 314 struct drm_mm_node *node = 315 rb_entry(rb, struct drm_mm_node, rb_hole_size); 316 317 if (size <= node->hole_size) { 318 best = node; 319 rb = rb->rb_right; 320 } else { 321 rb = rb->rb_left; 322 } 323 } while (rb); 324 325 return best; 326} 327 328static bool usable_hole_addr(struct rb_node *rb, u64 size) 329{ 330 return rb && rb_hole_addr_to_node(rb)->subtree_max_hole >= size; 331} 332 333static struct drm_mm_node *find_hole_addr(struct drm_mm *mm, u64 addr, u64 size) 334{ 335 struct rb_node *rb = mm->holes_addr.rb_node; 336 struct drm_mm_node *node = NULL; 337 338 while (rb) { 339 u64 hole_start; 340 341 if (!usable_hole_addr(rb, size)) 342 break; 343 344 node = rb_hole_addr_to_node(rb); 345 hole_start = __drm_mm_hole_node_start(node); 346 347 if (addr < hole_start) 348 rb = node->rb_hole_addr.rb_left; 349 else if (addr > hole_start + node->hole_size) 350 rb = node->rb_hole_addr.rb_right; 351 else 352 break; 353 } 354 355 return node; 356} 357 358static struct drm_mm_node * 359first_hole(struct drm_mm *mm, 360 u64 start, u64 end, u64 size, 361 enum drm_mm_insert_mode mode) 362{ 363 switch (mode) { 364 default: 365 case DRM_MM_INSERT_BEST: 366 return best_hole(mm, size); 367 368 case DRM_MM_INSERT_LOW: 369 return find_hole_addr(mm, start, size); 370 371 case DRM_MM_INSERT_HIGH: 372 return find_hole_addr(mm, end, size); 373 374 case DRM_MM_INSERT_EVICT: 375 return list_first_entry_or_null(&mm->hole_stack, 376 struct drm_mm_node, 377 hole_stack); 378 } 379} 380 381/** 382 * DECLARE_NEXT_HOLE_ADDR - macro to declare next hole functions 383 * @name: name of function to declare 384 * @first: first rb member to traverse (either rb_left or rb_right). 385 * @last: last rb member to traverse (either rb_right or rb_left). 386 * 387 * This macro declares a function to return the next hole of the addr rb tree. 388 * While traversing the tree we take the searched size into account and only 389 * visit branches with potential big enough holes. 390 */ 391 392#define DECLARE_NEXT_HOLE_ADDR(name, first, last) \ 393static struct drm_mm_node *name(struct drm_mm_node *entry, u64 size) \ 394{ \ 395 struct rb_node *parent, *node = &entry->rb_hole_addr; \ 396 \ 397 if (!entry || RB_EMPTY_NODE(node)) \ 398 return NULL; \ 399 \ 400 if (usable_hole_addr(node->first, size)) { \ 401 node = node->first; \ 402 while (usable_hole_addr(node->last, size)) \ 403 node = node->last; \ 404 return rb_hole_addr_to_node(node); \ 405 } \ 406 \ 407 while ((parent = rb_parent(node)) && node == parent->first) \ 408 node = parent; \ 409 \ 410 return rb_hole_addr_to_node(parent); \ 411} 412 413DECLARE_NEXT_HOLE_ADDR(next_hole_high_addr, rb_left, rb_right) 414DECLARE_NEXT_HOLE_ADDR(next_hole_low_addr, rb_right, rb_left) 415 416static struct drm_mm_node * 417next_hole(struct drm_mm *mm, 418 struct drm_mm_node *node, 419 u64 size, 420 enum drm_mm_insert_mode mode) 421{ 422 switch (mode) { 423 default: 424 case DRM_MM_INSERT_BEST: 425 return rb_hole_size_to_node(rb_prev(&node->rb_hole_size)); 426 427 case DRM_MM_INSERT_LOW: 428 return next_hole_low_addr(node, size); 429 430 case DRM_MM_INSERT_HIGH: 431 return next_hole_high_addr(node, size); 432 433 case DRM_MM_INSERT_EVICT: 434 node = list_next_entry(node, hole_stack); 435 return &node->hole_stack == &mm->hole_stack ? NULL : node; 436 } 437} 438 439/** 440 * drm_mm_reserve_node - insert an pre-initialized node 441 * @mm: drm_mm allocator to insert @node into 442 * @node: drm_mm_node to insert 443 * 444 * This functions inserts an already set-up &drm_mm_node into the allocator, 445 * meaning that start, size and color must be set by the caller. All other 446 * fields must be cleared to 0. This is useful to initialize the allocator with 447 * preallocated objects which must be set-up before the range allocator can be 448 * set-up, e.g. when taking over a firmware framebuffer. 449 * 450 * Returns: 451 * 0 on success, -ENOSPC if there's no hole where @node is. 452 */ 453int drm_mm_reserve_node(struct drm_mm *mm, struct drm_mm_node *node) 454{ 455 struct drm_mm_node *hole; 456 u64 hole_start, hole_end; 457 u64 adj_start, adj_end; 458 u64 end; 459 460 end = node->start + node->size; 461 if (unlikely(end <= node->start)) 462 return -ENOSPC; 463 464 /* Find the relevant hole to add our node to */ 465 hole = find_hole_addr(mm, node->start, 0); 466 if (!hole) 467 return -ENOSPC; 468 469 adj_start = hole_start = __drm_mm_hole_node_start(hole); 470 adj_end = hole_end = hole_start + hole->hole_size; 471 472 if (mm->color_adjust) 473 mm->color_adjust(hole, node->color, &adj_start, &adj_end); 474 475 if (adj_start > node->start || adj_end < end) 476 return -ENOSPC; 477 478 node->mm = mm; 479 480 __set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags); 481 list_add(&node->node_list, &hole->node_list); 482 drm_mm_interval_tree_add_node(hole, node); 483 node->hole_size = 0; 484 485 rm_hole(hole); 486 if (node->start > hole_start) 487 add_hole(hole); 488 if (end < hole_end) 489 add_hole(node); 490 491 save_stack(node); 492 return 0; 493} 494EXPORT_SYMBOL(drm_mm_reserve_node); 495 496static u64 rb_to_hole_size_or_zero(struct rb_node *rb) 497{ 498 return rb ? rb_to_hole_size(rb) : 0; 499} 500 501/** 502 * drm_mm_insert_node_in_range - ranged search for space and insert @node 503 * @mm: drm_mm to allocate from 504 * @node: preallocate node to insert 505 * @size: size of the allocation 506 * @alignment: alignment of the allocation 507 * @color: opaque tag value to use for this node 508 * @range_start: start of the allowed range for this node 509 * @range_end: end of the allowed range for this node 510 * @mode: fine-tune the allocation search and placement 511 * 512 * The preallocated @node must be cleared to 0. 513 * 514 * Returns: 515 * 0 on success, -ENOSPC if there's no suitable hole. 516 */ 517int drm_mm_insert_node_in_range(struct drm_mm * const mm, 518 struct drm_mm_node * const node, 519 u64 size, u64 alignment, 520 unsigned long color, 521 u64 range_start, u64 range_end, 522 enum drm_mm_insert_mode mode) 523{ 524 struct drm_mm_node *hole; 525 u64 remainder_mask; 526 bool once; 527 528 DRM_MM_BUG_ON(range_start > range_end); 529 530 if (unlikely(size == 0 || range_end - range_start < size)) 531 return -ENOSPC; 532 533 if (rb_to_hole_size_or_zero(rb_first_cached(&mm->holes_size)) < size) 534 return -ENOSPC; 535 536 if (alignment <= 1) 537 alignment = 0; 538 539 once = mode & DRM_MM_INSERT_ONCE; 540 mode &= ~DRM_MM_INSERT_ONCE; 541 542 remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0; 543 for (hole = first_hole(mm, range_start, range_end, size, mode); 544 hole; 545 hole = once ? NULL : next_hole(mm, hole, size, mode)) { 546 u64 hole_start = __drm_mm_hole_node_start(hole); 547 u64 hole_end = hole_start + hole->hole_size; 548 u64 adj_start, adj_end; 549 u64 col_start, col_end; 550 551 if (mode == DRM_MM_INSERT_LOW && hole_start >= range_end) 552 break; 553 554 if (mode == DRM_MM_INSERT_HIGH && hole_end <= range_start) 555 break; 556 557 col_start = hole_start; 558 col_end = hole_end; 559 if (mm->color_adjust) 560 mm->color_adjust(hole, color, &col_start, &col_end); 561 562 adj_start = max(col_start, range_start); 563 adj_end = min(col_end, range_end); 564 565 if (adj_end <= adj_start || adj_end - adj_start < size) 566 continue; 567 568 if (mode == DRM_MM_INSERT_HIGH) 569 adj_start = adj_end - size; 570 571 if (alignment) { 572 u64 rem; 573 574 if (likely(remainder_mask)) 575 rem = adj_start & remainder_mask; 576 else 577 div64_u64_rem(adj_start, alignment, &rem); 578 if (rem) { 579 adj_start -= rem; 580 if (mode != DRM_MM_INSERT_HIGH) 581 adj_start += alignment; 582 583 if (adj_start < max(col_start, range_start) || 584 min(col_end, range_end) - adj_start < size) 585 continue; 586 587 if (adj_end <= adj_start || 588 adj_end - adj_start < size) 589 continue; 590 } 591 } 592 593 node->mm = mm; 594 node->size = size; 595 node->start = adj_start; 596 node->color = color; 597 node->hole_size = 0; 598 599 __set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags); 600 list_add(&node->node_list, &hole->node_list); 601 drm_mm_interval_tree_add_node(hole, node); 602 603 rm_hole(hole); 604 if (adj_start > hole_start) 605 add_hole(hole); 606 if (adj_start + size < hole_end) 607 add_hole(node); 608 609 save_stack(node); 610 return 0; 611 } 612 613 return -ENOSPC; 614} 615EXPORT_SYMBOL(drm_mm_insert_node_in_range); 616 617static inline bool drm_mm_node_scanned_block(const struct drm_mm_node *node) 618{ 619 return test_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags); 620} 621 622/** 623 * drm_mm_remove_node - Remove a memory node from the allocator. 624 * @node: drm_mm_node to remove 625 * 626 * This just removes a node from its drm_mm allocator. The node does not need to 627 * be cleared again before it can be re-inserted into this or any other drm_mm 628 * allocator. It is a bug to call this function on a unallocated node. 629 */ 630void drm_mm_remove_node(struct drm_mm_node *node) 631{ 632 struct drm_mm *mm = node->mm; 633 struct drm_mm_node *prev_node; 634 635 DRM_MM_BUG_ON(!drm_mm_node_allocated(node)); 636 DRM_MM_BUG_ON(drm_mm_node_scanned_block(node)); 637 638 prev_node = list_prev_entry(node, node_list); 639 640 if (drm_mm_hole_follows(node)) 641 rm_hole(node); 642 643 drm_mm_interval_tree_remove(node, &mm->interval_tree); 644 list_del(&node->node_list); 645 646 if (drm_mm_hole_follows(prev_node)) 647 rm_hole(prev_node); 648 add_hole(prev_node); 649 650 clear_bit_unlock(DRM_MM_NODE_ALLOCATED_BIT, &node->flags); 651} 652EXPORT_SYMBOL(drm_mm_remove_node); 653 654/** 655 * drm_mm_replace_node - move an allocation from @old to @new 656 * @old: drm_mm_node to remove from the allocator 657 * @new: drm_mm_node which should inherit @old's allocation 658 * 659 * This is useful for when drivers embed the drm_mm_node structure and hence 660 * can't move allocations by reassigning pointers. It's a combination of remove 661 * and insert with the guarantee that the allocation start will match. 662 */ 663void drm_mm_replace_node(struct drm_mm_node *old, struct drm_mm_node *new) 664{ 665 struct drm_mm *mm = old->mm; 666 667 DRM_MM_BUG_ON(!drm_mm_node_allocated(old)); 668 669 *new = *old; 670 671 __set_bit(DRM_MM_NODE_ALLOCATED_BIT, &new->flags); 672 list_replace(&old->node_list, &new->node_list); 673 rb_replace_node_cached(&old->rb, &new->rb, &mm->interval_tree); 674 675 if (drm_mm_hole_follows(old)) { 676 list_replace(&old->hole_stack, &new->hole_stack); 677 rb_replace_node_cached(&old->rb_hole_size, 678 &new->rb_hole_size, 679 &mm->holes_size); 680 rb_replace_node(&old->rb_hole_addr, 681 &new->rb_hole_addr, 682 &mm->holes_addr); 683 } 684 685 clear_bit_unlock(DRM_MM_NODE_ALLOCATED_BIT, &old->flags); 686} 687EXPORT_SYMBOL(drm_mm_replace_node); 688 689/** 690 * DOC: lru scan roster 691 * 692 * Very often GPUs need to have continuous allocations for a given object. When 693 * evicting objects to make space for a new one it is therefore not most 694 * efficient when we simply start to select all objects from the tail of an LRU 695 * until there's a suitable hole: Especially for big objects or nodes that 696 * otherwise have special allocation constraints there's a good chance we evict 697 * lots of (smaller) objects unnecessarily. 698 * 699 * The DRM range allocator supports this use-case through the scanning 700 * interfaces. First a scan operation needs to be initialized with 701 * drm_mm_scan_init() or drm_mm_scan_init_with_range(). The driver adds 702 * objects to the roster, probably by walking an LRU list, but this can be 703 * freely implemented. Eviction candiates are added using 704 * drm_mm_scan_add_block() until a suitable hole is found or there are no 705 * further evictable objects. Eviction roster metadata is tracked in &struct 706 * drm_mm_scan. 707 * 708 * The driver must walk through all objects again in exactly the reverse 709 * order to restore the allocator state. Note that while the allocator is used 710 * in the scan mode no other operation is allowed. 711 * 712 * Finally the driver evicts all objects selected (drm_mm_scan_remove_block() 713 * reported true) in the scan, and any overlapping nodes after color adjustment 714 * (drm_mm_scan_color_evict()). Adding and removing an object is O(1), and 715 * since freeing a node is also O(1) the overall complexity is 716 * O(scanned_objects). So like the free stack which needs to be walked before a 717 * scan operation even begins this is linear in the number of objects. It 718 * doesn't seem to hurt too badly. 719 */ 720 721/** 722 * drm_mm_scan_init_with_range - initialize range-restricted lru scanning 723 * @scan: scan state 724 * @mm: drm_mm to scan 725 * @size: size of the allocation 726 * @alignment: alignment of the allocation 727 * @color: opaque tag value to use for the allocation 728 * @start: start of the allowed range for the allocation 729 * @end: end of the allowed range for the allocation 730 * @mode: fine-tune the allocation search and placement 731 * 732 * This simply sets up the scanning routines with the parameters for the desired 733 * hole. 734 * 735 * Warning: 736 * As long as the scan list is non-empty, no other operations than 737 * adding/removing nodes to/from the scan list are allowed. 738 */ 739void drm_mm_scan_init_with_range(struct drm_mm_scan *scan, 740 struct drm_mm *mm, 741 u64 size, 742 u64 alignment, 743 unsigned long color, 744 u64 start, 745 u64 end, 746 enum drm_mm_insert_mode mode) 747{ 748 DRM_MM_BUG_ON(start >= end); 749 DRM_MM_BUG_ON(!size || size > end - start); 750 DRM_MM_BUG_ON(mm->scan_active); 751 752 scan->mm = mm; 753 754 if (alignment <= 1) 755 alignment = 0; 756 757 scan->color = color; 758 scan->alignment = alignment; 759 scan->remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0; 760 scan->size = size; 761 scan->mode = mode; 762 763 DRM_MM_BUG_ON(end <= start); 764 scan->range_start = start; 765 scan->range_end = end; 766 767 scan->hit_start = U64_MAX; 768 scan->hit_end = 0; 769} 770EXPORT_SYMBOL(drm_mm_scan_init_with_range); 771 772/** 773 * drm_mm_scan_add_block - add a node to the scan list 774 * @scan: the active drm_mm scanner 775 * @node: drm_mm_node to add 776 * 777 * Add a node to the scan list that might be freed to make space for the desired 778 * hole. 779 * 780 * Returns: 781 * True if a hole has been found, false otherwise. 782 */ 783bool drm_mm_scan_add_block(struct drm_mm_scan *scan, 784 struct drm_mm_node *node) 785{ 786 struct drm_mm *mm = scan->mm; 787 struct drm_mm_node *hole; 788 u64 hole_start, hole_end; 789 u64 col_start, col_end; 790 u64 adj_start, adj_end; 791 792 DRM_MM_BUG_ON(node->mm != mm); 793 DRM_MM_BUG_ON(!drm_mm_node_allocated(node)); 794 DRM_MM_BUG_ON(drm_mm_node_scanned_block(node)); 795 __set_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags); 796 mm->scan_active++; 797 798 /* Remove this block from the node_list so that we enlarge the hole 799 * (distance between the end of our previous node and the start of 800 * or next), without poisoning the link so that we can restore it 801 * later in drm_mm_scan_remove_block(). 802 */ 803 hole = list_prev_entry(node, node_list); 804 DRM_MM_BUG_ON(list_next_entry(hole, node_list) != node); 805 __list_del_entry(&node->node_list); 806 807 hole_start = __drm_mm_hole_node_start(hole); 808 hole_end = __drm_mm_hole_node_end(hole); 809 810 col_start = hole_start; 811 col_end = hole_end; 812 if (mm->color_adjust) 813 mm->color_adjust(hole, scan->color, &col_start, &col_end); 814 815 adj_start = max(col_start, scan->range_start); 816 adj_end = min(col_end, scan->range_end); 817 if (adj_end <= adj_start || adj_end - adj_start < scan->size) 818 return false; 819 820 if (scan->mode == DRM_MM_INSERT_HIGH) 821 adj_start = adj_end - scan->size; 822 823 if (scan->alignment) { 824 u64 rem; 825 826 if (likely(scan->remainder_mask)) 827 rem = adj_start & scan->remainder_mask; 828 else 829 div64_u64_rem(adj_start, scan->alignment, &rem); 830 if (rem) { 831 adj_start -= rem; 832 if (scan->mode != DRM_MM_INSERT_HIGH) 833 adj_start += scan->alignment; 834 if (adj_start < max(col_start, scan->range_start) || 835 min(col_end, scan->range_end) - adj_start < scan->size) 836 return false; 837 838 if (adj_end <= adj_start || 839 adj_end - adj_start < scan->size) 840 return false; 841 } 842 } 843 844 scan->hit_start = adj_start; 845 scan->hit_end = adj_start + scan->size; 846 847 DRM_MM_BUG_ON(scan->hit_start >= scan->hit_end); 848 DRM_MM_BUG_ON(scan->hit_start < hole_start); 849 DRM_MM_BUG_ON(scan->hit_end > hole_end); 850 851 return true; 852} 853EXPORT_SYMBOL(drm_mm_scan_add_block); 854 855/** 856 * drm_mm_scan_remove_block - remove a node from the scan list 857 * @scan: the active drm_mm scanner 858 * @node: drm_mm_node to remove 859 * 860 * Nodes **must** be removed in exactly the reverse order from the scan list as 861 * they have been added (e.g. using list_add() as they are added and then 862 * list_for_each() over that eviction list to remove), otherwise the internal 863 * state of the memory manager will be corrupted. 864 * 865 * When the scan list is empty, the selected memory nodes can be freed. An 866 * immediately following drm_mm_insert_node_in_range_generic() or one of the 867 * simpler versions of that function with !DRM_MM_SEARCH_BEST will then return 868 * the just freed block (because it's at the top of the free_stack list). 869 * 870 * Returns: 871 * True if this block should be evicted, false otherwise. Will always 872 * return false when no hole has been found. 873 */ 874bool drm_mm_scan_remove_block(struct drm_mm_scan *scan, 875 struct drm_mm_node *node) 876{ 877 struct drm_mm_node *prev_node; 878 879 DRM_MM_BUG_ON(node->mm != scan->mm); 880 DRM_MM_BUG_ON(!drm_mm_node_scanned_block(node)); 881 __clear_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags); 882 883 DRM_MM_BUG_ON(!node->mm->scan_active); 884 node->mm->scan_active--; 885 886 /* During drm_mm_scan_add_block() we decoupled this node leaving 887 * its pointers intact. Now that the caller is walking back along 888 * the eviction list we can restore this block into its rightful 889 * place on the full node_list. To confirm that the caller is walking 890 * backwards correctly we check that prev_node->next == node->next, 891 * i.e. both believe the same node should be on the other side of the 892 * hole. 893 */ 894 prev_node = list_prev_entry(node, node_list); 895 DRM_MM_BUG_ON(list_next_entry(prev_node, node_list) != 896 list_next_entry(node, node_list)); 897 list_add(&node->node_list, &prev_node->node_list); 898 899 return (node->start + node->size > scan->hit_start && 900 node->start < scan->hit_end); 901} 902EXPORT_SYMBOL(drm_mm_scan_remove_block); 903 904/** 905 * drm_mm_scan_color_evict - evict overlapping nodes on either side of hole 906 * @scan: drm_mm scan with target hole 907 * 908 * After completing an eviction scan and removing the selected nodes, we may 909 * need to remove a few more nodes from either side of the target hole if 910 * mm.color_adjust is being used. 911 * 912 * Returns: 913 * A node to evict, or NULL if there are no overlapping nodes. 914 */ 915struct drm_mm_node *drm_mm_scan_color_evict(struct drm_mm_scan *scan) 916{ 917 struct drm_mm *mm = scan->mm; 918 struct drm_mm_node *hole; 919 u64 hole_start, hole_end; 920 921 DRM_MM_BUG_ON(list_empty(&mm->hole_stack)); 922 923 if (!mm->color_adjust) 924 return NULL; 925 926 /* 927 * The hole found during scanning should ideally be the first element 928 * in the hole_stack list, but due to side-effects in the driver it 929 * may not be. 930 */ 931 list_for_each_entry(hole, &mm->hole_stack, hole_stack) { 932 hole_start = __drm_mm_hole_node_start(hole); 933 hole_end = hole_start + hole->hole_size; 934 935 if (hole_start <= scan->hit_start && 936 hole_end >= scan->hit_end) 937 break; 938 } 939 940 /* We should only be called after we found the hole previously */ 941 DRM_MM_BUG_ON(&hole->hole_stack == &mm->hole_stack); 942 if (unlikely(&hole->hole_stack == &mm->hole_stack)) 943 return NULL; 944 945 DRM_MM_BUG_ON(hole_start > scan->hit_start); 946 DRM_MM_BUG_ON(hole_end < scan->hit_end); 947 948 mm->color_adjust(hole, scan->color, &hole_start, &hole_end); 949 if (hole_start > scan->hit_start) 950 return hole; 951 if (hole_end < scan->hit_end) 952 return list_next_entry(hole, node_list); 953 954 return NULL; 955} 956EXPORT_SYMBOL(drm_mm_scan_color_evict); 957 958/** 959 * drm_mm_init - initialize a drm-mm allocator 960 * @mm: the drm_mm structure to initialize 961 * @start: start of the range managed by @mm 962 * @size: end of the range managed by @mm 963 * 964 * Note that @mm must be cleared to 0 before calling this function. 965 */ 966void drm_mm_init(struct drm_mm *mm, u64 start, u64 size) 967{ 968 DRM_MM_BUG_ON(start + size <= start); 969 970 mm->color_adjust = NULL; 971 972 INIT_LIST_HEAD(&mm->hole_stack); 973 mm->interval_tree = RB_ROOT_CACHED; 974 mm->holes_size = RB_ROOT_CACHED; 975 mm->holes_addr = RB_ROOT; 976 977 /* Clever trick to avoid a special case in the free hole tracking. */ 978 INIT_LIST_HEAD(&mm->head_node.node_list); 979 mm->head_node.flags = 0; 980 mm->head_node.mm = mm; 981 mm->head_node.start = start + size; 982 mm->head_node.size = -size; 983 add_hole(&mm->head_node); 984 985 mm->scan_active = 0; 986} 987EXPORT_SYMBOL(drm_mm_init); 988 989/** 990 * drm_mm_takedown - clean up a drm_mm allocator 991 * @mm: drm_mm allocator to clean up 992 * 993 * Note that it is a bug to call this function on an allocator which is not 994 * clean. 995 */ 996void drm_mm_takedown(struct drm_mm *mm) 997{ 998 if (WARN(!drm_mm_clean(mm), 999 "Memory manager not clean during takedown.\n")) 1000 show_leaks(mm); 1001} 1002EXPORT_SYMBOL(drm_mm_takedown); 1003 1004static u64 drm_mm_dump_hole(struct drm_printer *p, const struct drm_mm_node *entry) 1005{ 1006 u64 start, size; 1007 1008 size = entry->hole_size; 1009 if (size) { 1010 start = drm_mm_hole_node_start(entry); 1011 drm_printf(p, "%#018llx-%#018llx: %llu: free\n", 1012 start, start + size, size); 1013 } 1014 1015 return size; 1016} 1017/** 1018 * drm_mm_print - print allocator state 1019 * @mm: drm_mm allocator to print 1020 * @p: DRM printer to use 1021 */ 1022void drm_mm_print(const struct drm_mm *mm, struct drm_printer *p) 1023{ 1024 const struct drm_mm_node *entry; 1025 u64 total_used = 0, total_free = 0, total = 0; 1026 1027 total_free += drm_mm_dump_hole(p, &mm->head_node); 1028 1029 drm_mm_for_each_node(entry, mm) { 1030 drm_printf(p, "%#018llx-%#018llx: %llu: used\n", entry->start, 1031 entry->start + entry->size, entry->size); 1032 total_used += entry->size; 1033 total_free += drm_mm_dump_hole(p, entry); 1034 } 1035 total = total_free + total_used; 1036 1037 drm_printf(p, "total: %llu, used %llu free %llu\n", total, 1038 total_used, total_free); 1039} 1040EXPORT_SYMBOL(drm_mm_print); 1041