1 /* SPDX-License-Identifier: GPL-2.0-only */ 2 /* 3 * Fast and scalable bitmaps. 4 * 5 * Copyright (C) 2016 Facebook 6 * Copyright (C) 2013-2014 Jens Axboe 7 */ 8 9 #ifndef __LINUX_SCALE_BITMAP_H 10 #define __LINUX_SCALE_BITMAP_H 11 12 #include <linux/kernel.h> 13 #include <linux/slab.h> 14 15 struct seq_file; 16 17 /** 18 * struct sbitmap_word - Word in a &struct sbitmap. 19 */ 20 struct sbitmap_word { 21 /** 22 * @depth: Number of bits being used in @word/@cleared 23 */ 24 unsigned long depth; 25 26 /** 27 * @word: word holding free bits 28 */ 29 unsigned long word ____cacheline_aligned_in_smp; 30 31 /** 32 * @cleared: word holding cleared bits 33 */ 34 unsigned long cleared ____cacheline_aligned_in_smp; 35 } ____cacheline_aligned_in_smp; 36 37 /** 38 * struct sbitmap - Scalable bitmap. 39 * 40 * A &struct sbitmap is spread over multiple cachelines to avoid ping-pong. This 41 * trades off higher memory usage for better scalability. 42 */ 43 struct sbitmap { 44 /** 45 * @depth: Number of bits used in the whole bitmap. 46 */ 47 unsigned int depth; 48 49 /** 50 * @shift: log2(number of bits used per word) 51 */ 52 unsigned int shift; 53 54 /** 55 * @map_nr: Number of words (cachelines) being used for the bitmap. 56 */ 57 unsigned int map_nr; 58 59 /** 60 * @round_robin: Allocate bits in strict round-robin order. 61 */ 62 bool round_robin; 63 64 /** 65 * @map: Allocated bitmap. 66 */ 67 struct sbitmap_word *map; 68 69 /* 70 * @alloc_hint: Cache of last successfully allocated or freed bit. 71 * 72 * This is per-cpu, which allows multiple users to stick to different 73 * cachelines until the map is exhausted. 74 */ 75 unsigned int __percpu *alloc_hint; 76 }; 77 78 #define SBQ_WAIT_QUEUES 8 79 #define SBQ_WAKE_BATCH 8 80 81 /** 82 * struct sbq_wait_state - Wait queue in a &struct sbitmap_queue. 83 */ 84 struct sbq_wait_state { 85 /** 86 * @wait_cnt: Number of frees remaining before we wake up. 87 */ 88 atomic_t wait_cnt; 89 90 /** 91 * @wait: Wait queue. 92 */ 93 wait_queue_head_t wait; 94 } ____cacheline_aligned_in_smp; 95 96 /** 97 * struct sbitmap_queue - Scalable bitmap with the added ability to wait on free 98 * bits. 99 * 100 * A &struct sbitmap_queue uses multiple wait queues and rolling wakeups to 101 * avoid contention on the wait queue spinlock. This ensures that we don't hit a 102 * scalability wall when we run out of free bits and have to start putting tasks 103 * to sleep. 104 */ 105 struct sbitmap_queue { 106 /** 107 * @sb: Scalable bitmap. 108 */ 109 struct sbitmap sb; 110 111 /** 112 * @wake_batch: Number of bits which must be freed before we wake up any 113 * waiters. 114 */ 115 unsigned int wake_batch; 116 117 /** 118 * @wake_index: Next wait queue in @ws to wake up. 119 */ 120 atomic_t wake_index; 121 122 /** 123 * @ws: Wait queues. 124 */ 125 struct sbq_wait_state *ws; 126 127 /* 128 * @ws_active: count of currently active ws waitqueues 129 */ 130 atomic_t ws_active; 131 132 /** 133 * @min_shallow_depth: The minimum shallow depth which may be passed to 134 * sbitmap_queue_get_shallow() or __sbitmap_queue_get_shallow(). 135 */ 136 unsigned int min_shallow_depth; 137 }; 138 139 /** 140 * sbitmap_init_node() - Initialize a &struct sbitmap on a specific memory node. 141 * @sb: Bitmap to initialize. 142 * @depth: Number of bits to allocate. 143 * @shift: Use 2^@shift bits per word in the bitmap; if a negative number if 144 * given, a good default is chosen. 145 * @flags: Allocation flags. 146 * @node: Memory node to allocate on. 147 * @round_robin: If true, be stricter about allocation order; always allocate 148 * starting from the last allocated bit. This is less efficient 149 * than the default behavior (false). 150 * @alloc_hint: If true, apply percpu hint for where to start searching for 151 * a free bit. 152 * 153 * Return: Zero on success or negative errno on failure. 154 */ 155 int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift, 156 gfp_t flags, int node, bool round_robin, bool alloc_hint); 157 158 /** 159 * sbitmap_free() - Free memory used by a &struct sbitmap. 160 * @sb: Bitmap to free. 161 */ 162 static inline void sbitmap_free(struct sbitmap *sb) 163 { 164 free_percpu(sb->alloc_hint); 165 kfree(sb->map); 166 sb->map = NULL; 167 } 168 169 /** 170 * sbitmap_resize() - Resize a &struct sbitmap. 171 * @sb: Bitmap to resize. 172 * @depth: New number of bits to resize to. 173 * 174 * Doesn't reallocate anything. It's up to the caller to ensure that the new 175 * depth doesn't exceed the depth that the sb was initialized with. 176 */ 177 void sbitmap_resize(struct sbitmap *sb, unsigned int depth); 178 179 /** 180 * sbitmap_get() - Try to allocate a free bit from a &struct sbitmap. 181 * @sb: Bitmap to allocate from. 182 * 183 * This operation provides acquire barrier semantics if it succeeds. 184 * 185 * Return: Non-negative allocated bit number if successful, -1 otherwise. 186 */ 187 int sbitmap_get(struct sbitmap *sb); 188 189 /** 190 * sbitmap_get_shallow() - Try to allocate a free bit from a &struct sbitmap, 191 * limiting the depth used from each word. 192 * @sb: Bitmap to allocate from. 193 * @shallow_depth: The maximum number of bits to allocate from a single word. 194 * 195 * This rather specific operation allows for having multiple users with 196 * different allocation limits. E.g., there can be a high-priority class that 197 * uses sbitmap_get() and a low-priority class that uses sbitmap_get_shallow() 198 * with a @shallow_depth of (1 << (@sb->shift - 1)). Then, the low-priority 199 * class can only allocate half of the total bits in the bitmap, preventing it 200 * from starving out the high-priority class. 201 * 202 * Return: Non-negative allocated bit number if successful, -1 otherwise. 203 */ 204 int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth); 205 206 /** 207 * sbitmap_any_bit_set() - Check for a set bit in a &struct sbitmap. 208 * @sb: Bitmap to check. 209 * 210 * Return: true if any bit in the bitmap is set, false otherwise. 211 */ 212 bool sbitmap_any_bit_set(const struct sbitmap *sb); 213 214 #define SB_NR_TO_INDEX(sb, bitnr) ((bitnr) >> (sb)->shift) 215 #define SB_NR_TO_BIT(sb, bitnr) ((bitnr) & ((1U << (sb)->shift) - 1U)) 216 217 typedef bool (*sb_for_each_fn)(struct sbitmap *, unsigned int, void *); 218 219 /** 220 * __sbitmap_for_each_set() - Iterate over each set bit in a &struct sbitmap. 221 * @start: Where to start the iteration. 222 * @sb: Bitmap to iterate over. 223 * @fn: Callback. Should return true to continue or false to break early. 224 * @data: Pointer to pass to callback. 225 * 226 * This is inline even though it's non-trivial so that the function calls to the 227 * callback will hopefully get optimized away. 228 */ 229 static inline void __sbitmap_for_each_set(struct sbitmap *sb, 230 unsigned int start, 231 sb_for_each_fn fn, void *data) 232 { 233 unsigned int index; 234 unsigned int nr; 235 unsigned int scanned = 0; 236 237 if (start >= sb->depth) 238 start = 0; 239 index = SB_NR_TO_INDEX(sb, start); 240 nr = SB_NR_TO_BIT(sb, start); 241 242 while (scanned < sb->depth) { 243 unsigned long word; 244 unsigned int depth = min_t(unsigned int, 245 sb->map[index].depth - nr, 246 sb->depth - scanned); 247 248 scanned += depth; 249 word = sb->map[index].word & ~sb->map[index].cleared; 250 if (!word) 251 goto next; 252 253 /* 254 * On the first iteration of the outer loop, we need to add the 255 * bit offset back to the size of the word for find_next_bit(). 256 * On all other iterations, nr is zero, so this is a noop. 257 */ 258 depth += nr; 259 while (1) { 260 nr = find_next_bit(&word, depth, nr); 261 if (nr >= depth) 262 break; 263 if (!fn(sb, (index << sb->shift) + nr, data)) 264 return; 265 266 nr++; 267 } 268 next: 269 nr = 0; 270 if (++index >= sb->map_nr) 271 index = 0; 272 } 273 } 274 275 /** 276 * sbitmap_for_each_set() - Iterate over each set bit in a &struct sbitmap. 277 * @sb: Bitmap to iterate over. 278 * @fn: Callback. Should return true to continue or false to break early. 279 * @data: Pointer to pass to callback. 280 */ 281 static inline void sbitmap_for_each_set(struct sbitmap *sb, sb_for_each_fn fn, 282 void *data) 283 { 284 __sbitmap_for_each_set(sb, 0, fn, data); 285 } 286 287 static inline unsigned long *__sbitmap_word(struct sbitmap *sb, 288 unsigned int bitnr) 289 { 290 return &sb->map[SB_NR_TO_INDEX(sb, bitnr)].word; 291 } 292 293 /* Helpers equivalent to the operations in asm/bitops.h and linux/bitmap.h */ 294 295 static inline void sbitmap_set_bit(struct sbitmap *sb, unsigned int bitnr) 296 { 297 set_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr)); 298 } 299 300 static inline void sbitmap_clear_bit(struct sbitmap *sb, unsigned int bitnr) 301 { 302 clear_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr)); 303 } 304 305 /* 306 * This one is special, since it doesn't actually clear the bit, rather it 307 * sets the corresponding bit in the ->cleared mask instead. Paired with 308 * the caller doing sbitmap_deferred_clear() if a given index is full, which 309 * will clear the previously freed entries in the corresponding ->word. 310 */ 311 static inline void sbitmap_deferred_clear_bit(struct sbitmap *sb, unsigned int bitnr) 312 { 313 unsigned long *addr = &sb->map[SB_NR_TO_INDEX(sb, bitnr)].cleared; 314 315 set_bit(SB_NR_TO_BIT(sb, bitnr), addr); 316 } 317 318 /* 319 * Pair of sbitmap_get, and this one applies both cleared bit and 320 * allocation hint. 321 */ 322 static inline void sbitmap_put(struct sbitmap *sb, unsigned int bitnr) 323 { 324 sbitmap_deferred_clear_bit(sb, bitnr); 325 326 if (likely(sb->alloc_hint && !sb->round_robin && bitnr < sb->depth)) 327 *raw_cpu_ptr(sb->alloc_hint) = bitnr; 328 } 329 330 static inline int sbitmap_test_bit(struct sbitmap *sb, unsigned int bitnr) 331 { 332 return test_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr)); 333 } 334 335 static inline int sbitmap_calculate_shift(unsigned int depth) 336 { 337 int shift = ilog2(BITS_PER_LONG); 338 339 /* 340 * If the bitmap is small, shrink the number of bits per word so 341 * we spread over a few cachelines, at least. If less than 4 342 * bits, just forget about it, it's not going to work optimally 343 * anyway. 344 */ 345 if (depth >= 4) { 346 while ((4U << shift) > depth) 347 shift--; 348 } 349 350 return shift; 351 } 352 353 /** 354 * sbitmap_show() - Dump &struct sbitmap information to a &struct seq_file. 355 * @sb: Bitmap to show. 356 * @m: struct seq_file to write to. 357 * 358 * This is intended for debugging. The format may change at any time. 359 */ 360 void sbitmap_show(struct sbitmap *sb, struct seq_file *m); 361 362 363 /** 364 * sbitmap_weight() - Return how many set and not cleared bits in a &struct 365 * sbitmap. 366 * @sb: Bitmap to check. 367 * 368 * Return: How many set and not cleared bits set 369 */ 370 unsigned int sbitmap_weight(const struct sbitmap *sb); 371 372 /** 373 * sbitmap_bitmap_show() - Write a hex dump of a &struct sbitmap to a &struct 374 * seq_file. 375 * @sb: Bitmap to show. 376 * @m: struct seq_file to write to. 377 * 378 * This is intended for debugging. The output isn't guaranteed to be internally 379 * consistent. 380 */ 381 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m); 382 383 /** 384 * sbitmap_queue_init_node() - Initialize a &struct sbitmap_queue on a specific 385 * memory node. 386 * @sbq: Bitmap queue to initialize. 387 * @depth: See sbitmap_init_node(). 388 * @shift: See sbitmap_init_node(). 389 * @round_robin: See sbitmap_get(). 390 * @flags: Allocation flags. 391 * @node: Memory node to allocate on. 392 * 393 * Return: Zero on success or negative errno on failure. 394 */ 395 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth, 396 int shift, bool round_robin, gfp_t flags, int node); 397 398 /** 399 * sbitmap_queue_free() - Free memory used by a &struct sbitmap_queue. 400 * 401 * @sbq: Bitmap queue to free. 402 */ 403 static inline void sbitmap_queue_free(struct sbitmap_queue *sbq) 404 { 405 kfree(sbq->ws); 406 sbitmap_free(&sbq->sb); 407 } 408 409 /** 410 * sbitmap_queue_resize() - Resize a &struct sbitmap_queue. 411 * @sbq: Bitmap queue to resize. 412 * @depth: New number of bits to resize to. 413 * 414 * Like sbitmap_resize(), this doesn't reallocate anything. It has to do 415 * some extra work on the &struct sbitmap_queue, so it's not safe to just 416 * resize the underlying &struct sbitmap. 417 */ 418 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth); 419 420 /** 421 * __sbitmap_queue_get() - Try to allocate a free bit from a &struct 422 * sbitmap_queue with preemption already disabled. 423 * @sbq: Bitmap queue to allocate from. 424 * 425 * Return: Non-negative allocated bit number if successful, -1 otherwise. 426 */ 427 int __sbitmap_queue_get(struct sbitmap_queue *sbq); 428 429 /** 430 * __sbitmap_queue_get_shallow() - Try to allocate a free bit from a &struct 431 * sbitmap_queue, limiting the depth used from each word, with preemption 432 * already disabled. 433 * @sbq: Bitmap queue to allocate from. 434 * @shallow_depth: The maximum number of bits to allocate from a single word. 435 * See sbitmap_get_shallow(). 436 * 437 * If you call this, make sure to call sbitmap_queue_min_shallow_depth() after 438 * initializing @sbq. 439 * 440 * Return: Non-negative allocated bit number if successful, -1 otherwise. 441 */ 442 int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq, 443 unsigned int shallow_depth); 444 445 /** 446 * sbitmap_queue_get() - Try to allocate a free bit from a &struct 447 * sbitmap_queue. 448 * @sbq: Bitmap queue to allocate from. 449 * @cpu: Output parameter; will contain the CPU we ran on (e.g., to be passed to 450 * sbitmap_queue_clear()). 451 * 452 * Return: Non-negative allocated bit number if successful, -1 otherwise. 453 */ 454 static inline int sbitmap_queue_get(struct sbitmap_queue *sbq, 455 unsigned int *cpu) 456 { 457 int nr; 458 459 *cpu = get_cpu(); 460 nr = __sbitmap_queue_get(sbq); 461 put_cpu(); 462 return nr; 463 } 464 465 /** 466 * sbitmap_queue_get_shallow() - Try to allocate a free bit from a &struct 467 * sbitmap_queue, limiting the depth used from each word. 468 * @sbq: Bitmap queue to allocate from. 469 * @cpu: Output parameter; will contain the CPU we ran on (e.g., to be passed to 470 * sbitmap_queue_clear()). 471 * @shallow_depth: The maximum number of bits to allocate from a single word. 472 * See sbitmap_get_shallow(). 473 * 474 * If you call this, make sure to call sbitmap_queue_min_shallow_depth() after 475 * initializing @sbq. 476 * 477 * Return: Non-negative allocated bit number if successful, -1 otherwise. 478 */ 479 static inline int sbitmap_queue_get_shallow(struct sbitmap_queue *sbq, 480 unsigned int *cpu, 481 unsigned int shallow_depth) 482 { 483 int nr; 484 485 *cpu = get_cpu(); 486 nr = __sbitmap_queue_get_shallow(sbq, shallow_depth); 487 put_cpu(); 488 return nr; 489 } 490 491 /** 492 * sbitmap_queue_min_shallow_depth() - Inform a &struct sbitmap_queue of the 493 * minimum shallow depth that will be used. 494 * @sbq: Bitmap queue in question. 495 * @min_shallow_depth: The minimum shallow depth that will be passed to 496 * sbitmap_queue_get_shallow() or __sbitmap_queue_get_shallow(). 497 * 498 * sbitmap_queue_clear() batches wakeups as an optimization. The batch size 499 * depends on the depth of the bitmap. Since the shallow allocation functions 500 * effectively operate with a different depth, the shallow depth must be taken 501 * into account when calculating the batch size. This function must be called 502 * with the minimum shallow depth that will be used. Failure to do so can result 503 * in missed wakeups. 504 */ 505 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq, 506 unsigned int min_shallow_depth); 507 508 /** 509 * sbitmap_queue_clear() - Free an allocated bit and wake up waiters on a 510 * &struct sbitmap_queue. 511 * @sbq: Bitmap to free from. 512 * @nr: Bit number to free. 513 * @cpu: CPU the bit was allocated on. 514 */ 515 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr, 516 unsigned int cpu); 517 518 static inline int sbq_index_inc(int index) 519 { 520 return (index + 1) & (SBQ_WAIT_QUEUES - 1); 521 } 522 523 static inline void sbq_index_atomic_inc(atomic_t *index) 524 { 525 int old = atomic_read(index); 526 int new = sbq_index_inc(old); 527 atomic_cmpxchg(index, old, new); 528 } 529 530 /** 531 * sbq_wait_ptr() - Get the next wait queue to use for a &struct 532 * sbitmap_queue. 533 * @sbq: Bitmap queue to wait on. 534 * @wait_index: A counter per "user" of @sbq. 535 */ 536 static inline struct sbq_wait_state *sbq_wait_ptr(struct sbitmap_queue *sbq, 537 atomic_t *wait_index) 538 { 539 struct sbq_wait_state *ws; 540 541 ws = &sbq->ws[atomic_read(wait_index)]; 542 sbq_index_atomic_inc(wait_index); 543 return ws; 544 } 545 546 /** 547 * sbitmap_queue_wake_all() - Wake up everything waiting on a &struct 548 * sbitmap_queue. 549 * @sbq: Bitmap queue to wake up. 550 */ 551 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq); 552 553 /** 554 * sbitmap_queue_wake_up() - Wake up some of waiters in one waitqueue 555 * on a &struct sbitmap_queue. 556 * @sbq: Bitmap queue to wake up. 557 */ 558 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq); 559 560 /** 561 * sbitmap_queue_show() - Dump &struct sbitmap_queue information to a &struct 562 * seq_file. 563 * @sbq: Bitmap queue to show. 564 * @m: struct seq_file to write to. 565 * 566 * This is intended for debugging. The format may change at any time. 567 */ 568 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m); 569 570 struct sbq_wait { 571 struct sbitmap_queue *sbq; /* if set, sbq_wait is accounted */ 572 struct wait_queue_entry wait; 573 }; 574 575 #define DEFINE_SBQ_WAIT(name) \ 576 struct sbq_wait name = { \ 577 .sbq = NULL, \ 578 .wait = { \ 579 .private = current, \ 580 .func = autoremove_wake_function, \ 581 .entry = LIST_HEAD_INIT((name).wait.entry), \ 582 } \ 583 } 584 585 /* 586 * Wrapper around prepare_to_wait_exclusive(), which maintains some extra 587 * internal state. 588 */ 589 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq, 590 struct sbq_wait_state *ws, 591 struct sbq_wait *sbq_wait, int state); 592 593 /* 594 * Must be paired with sbitmap_prepare_to_wait(). 595 */ 596 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws, 597 struct sbq_wait *sbq_wait); 598 599 /* 600 * Wrapper around add_wait_queue(), which maintains some extra internal state 601 */ 602 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq, 603 struct sbq_wait_state *ws, 604 struct sbq_wait *sbq_wait); 605 606 /* 607 * Must be paired with sbitmap_add_wait_queue() 608 */ 609 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait); 610 611 #endif /* __LINUX_SCALE_BITMAP_H */ 612