1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2016 Facebook 4 * Copyright (C) 2013-2014 Jens Axboe 5 */ 6 7 #include <linux/sched.h> 8 #include <linux/random.h> 9 #include <linux/sbitmap.h> 10 #include <linux/seq_file.h> 11 12 /* 13 * See if we have deferred clears that we can batch move 14 */ 15 static inline bool sbitmap_deferred_clear(struct sbitmap_word *map) 16 { 17 unsigned long mask; 18 19 if (!READ_ONCE(map->cleared)) 20 return false; 21 22 /* 23 * First get a stable cleared mask, setting the old mask to 0. 24 */ 25 mask = xchg(&map->cleared, 0); 26 27 /* 28 * Now clear the masked bits in our free word 29 */ 30 atomic_long_andnot(mask, (atomic_long_t *)&map->word); 31 BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(map->word)); 32 return true; 33 } 34 35 int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift, 36 gfp_t flags, int node) 37 { 38 unsigned int bits_per_word; 39 unsigned int i; 40 41 if (shift < 0) { 42 shift = ilog2(BITS_PER_LONG); 43 /* 44 * If the bitmap is small, shrink the number of bits per word so 45 * we spread over a few cachelines, at least. If less than 4 46 * bits, just forget about it, it's not going to work optimally 47 * anyway. 48 */ 49 if (depth >= 4) { 50 while ((4U << shift) > depth) 51 shift--; 52 } 53 } 54 bits_per_word = 1U << shift; 55 if (bits_per_word > BITS_PER_LONG) 56 return -EINVAL; 57 58 sb->shift = shift; 59 sb->depth = depth; 60 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word); 61 62 if (depth == 0) { 63 sb->map = NULL; 64 return 0; 65 } 66 67 sb->map = kcalloc_node(sb->map_nr, sizeof(*sb->map), flags, node); 68 if (!sb->map) 69 return -ENOMEM; 70 71 for (i = 0; i < sb->map_nr; i++) { 72 sb->map[i].depth = min(depth, bits_per_word); 73 depth -= sb->map[i].depth; 74 } 75 return 0; 76 } 77 EXPORT_SYMBOL_GPL(sbitmap_init_node); 78 79 void sbitmap_resize(struct sbitmap *sb, unsigned int depth) 80 { 81 unsigned int bits_per_word = 1U << sb->shift; 82 unsigned int i; 83 84 for (i = 0; i < sb->map_nr; i++) 85 sbitmap_deferred_clear(&sb->map[i]); 86 87 sb->depth = depth; 88 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word); 89 90 for (i = 0; i < sb->map_nr; i++) { 91 sb->map[i].depth = min(depth, bits_per_word); 92 depth -= sb->map[i].depth; 93 } 94 } 95 EXPORT_SYMBOL_GPL(sbitmap_resize); 96 97 static int __sbitmap_get_word(unsigned long *word, unsigned long depth, 98 unsigned int hint, bool wrap) 99 { 100 int nr; 101 102 /* don't wrap if starting from 0 */ 103 wrap = wrap && hint; 104 105 while (1) { 106 nr = find_next_zero_bit(word, depth, hint); 107 if (unlikely(nr >= depth)) { 108 /* 109 * We started with an offset, and we didn't reset the 110 * offset to 0 in a failure case, so start from 0 to 111 * exhaust the map. 112 */ 113 if (hint && wrap) { 114 hint = 0; 115 continue; 116 } 117 return -1; 118 } 119 120 if (!test_and_set_bit_lock(nr, word)) 121 break; 122 123 hint = nr + 1; 124 if (hint >= depth - 1) 125 hint = 0; 126 } 127 128 return nr; 129 } 130 131 static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index, 132 unsigned int alloc_hint, bool round_robin) 133 { 134 struct sbitmap_word *map = &sb->map[index]; 135 int nr; 136 137 do { 138 nr = __sbitmap_get_word(&map->word, map->depth, alloc_hint, 139 !round_robin); 140 if (nr != -1) 141 break; 142 if (!sbitmap_deferred_clear(map)) 143 break; 144 } while (1); 145 146 return nr; 147 } 148 149 int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin) 150 { 151 unsigned int i, index; 152 int nr = -1; 153 154 index = SB_NR_TO_INDEX(sb, alloc_hint); 155 156 /* 157 * Unless we're doing round robin tag allocation, just use the 158 * alloc_hint to find the right word index. No point in looping 159 * twice in find_next_zero_bit() for that case. 160 */ 161 if (round_robin) 162 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint); 163 else 164 alloc_hint = 0; 165 166 for (i = 0; i < sb->map_nr; i++) { 167 nr = sbitmap_find_bit_in_index(sb, index, alloc_hint, 168 round_robin); 169 if (nr != -1) { 170 nr += index << sb->shift; 171 break; 172 } 173 174 /* Jump to next index. */ 175 alloc_hint = 0; 176 if (++index >= sb->map_nr) 177 index = 0; 178 } 179 180 return nr; 181 } 182 EXPORT_SYMBOL_GPL(sbitmap_get); 183 184 int sbitmap_get_shallow(struct sbitmap *sb, unsigned int alloc_hint, 185 unsigned long shallow_depth) 186 { 187 unsigned int i, index; 188 int nr = -1; 189 190 index = SB_NR_TO_INDEX(sb, alloc_hint); 191 192 for (i = 0; i < sb->map_nr; i++) { 193 again: 194 nr = __sbitmap_get_word(&sb->map[index].word, 195 min(sb->map[index].depth, shallow_depth), 196 SB_NR_TO_BIT(sb, alloc_hint), true); 197 if (nr != -1) { 198 nr += index << sb->shift; 199 break; 200 } 201 202 if (sbitmap_deferred_clear(&sb->map[index])) 203 goto again; 204 205 /* Jump to next index. */ 206 index++; 207 alloc_hint = index << sb->shift; 208 209 if (index >= sb->map_nr) { 210 index = 0; 211 alloc_hint = 0; 212 } 213 } 214 215 return nr; 216 } 217 EXPORT_SYMBOL_GPL(sbitmap_get_shallow); 218 219 bool sbitmap_any_bit_set(const struct sbitmap *sb) 220 { 221 unsigned int i; 222 223 for (i = 0; i < sb->map_nr; i++) { 224 if (sb->map[i].word & ~sb->map[i].cleared) 225 return true; 226 } 227 return false; 228 } 229 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set); 230 231 static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set) 232 { 233 unsigned int i, weight = 0; 234 235 for (i = 0; i < sb->map_nr; i++) { 236 const struct sbitmap_word *word = &sb->map[i]; 237 238 if (set) 239 weight += bitmap_weight(&word->word, word->depth); 240 else 241 weight += bitmap_weight(&word->cleared, word->depth); 242 } 243 return weight; 244 } 245 246 static unsigned int sbitmap_weight(const struct sbitmap *sb) 247 { 248 return __sbitmap_weight(sb, true); 249 } 250 251 static unsigned int sbitmap_cleared(const struct sbitmap *sb) 252 { 253 return __sbitmap_weight(sb, false); 254 } 255 256 void sbitmap_show(struct sbitmap *sb, struct seq_file *m) 257 { 258 seq_printf(m, "depth=%u\n", sb->depth); 259 seq_printf(m, "busy=%u\n", sbitmap_weight(sb) - sbitmap_cleared(sb)); 260 seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb)); 261 seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift); 262 seq_printf(m, "map_nr=%u\n", sb->map_nr); 263 } 264 EXPORT_SYMBOL_GPL(sbitmap_show); 265 266 static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte) 267 { 268 if ((offset & 0xf) == 0) { 269 if (offset != 0) 270 seq_putc(m, '\n'); 271 seq_printf(m, "%08x:", offset); 272 } 273 if ((offset & 0x1) == 0) 274 seq_putc(m, ' '); 275 seq_printf(m, "%02x", byte); 276 } 277 278 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m) 279 { 280 u8 byte = 0; 281 unsigned int byte_bits = 0; 282 unsigned int offset = 0; 283 int i; 284 285 for (i = 0; i < sb->map_nr; i++) { 286 unsigned long word = READ_ONCE(sb->map[i].word); 287 unsigned long cleared = READ_ONCE(sb->map[i].cleared); 288 unsigned int word_bits = READ_ONCE(sb->map[i].depth); 289 290 word &= ~cleared; 291 292 while (word_bits > 0) { 293 unsigned int bits = min(8 - byte_bits, word_bits); 294 295 byte |= (word & (BIT(bits) - 1)) << byte_bits; 296 byte_bits += bits; 297 if (byte_bits == 8) { 298 emit_byte(m, offset, byte); 299 byte = 0; 300 byte_bits = 0; 301 offset++; 302 } 303 word >>= bits; 304 word_bits -= bits; 305 } 306 } 307 if (byte_bits) { 308 emit_byte(m, offset, byte); 309 offset++; 310 } 311 if (offset) 312 seq_putc(m, '\n'); 313 } 314 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show); 315 316 static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq, 317 unsigned int depth) 318 { 319 unsigned int wake_batch; 320 unsigned int shallow_depth; 321 322 /* 323 * For each batch, we wake up one queue. We need to make sure that our 324 * batch size is small enough that the full depth of the bitmap, 325 * potentially limited by a shallow depth, is enough to wake up all of 326 * the queues. 327 * 328 * Each full word of the bitmap has bits_per_word bits, and there might 329 * be a partial word. There are depth / bits_per_word full words and 330 * depth % bits_per_word bits left over. In bitwise arithmetic: 331 * 332 * bits_per_word = 1 << shift 333 * depth / bits_per_word = depth >> shift 334 * depth % bits_per_word = depth & ((1 << shift) - 1) 335 * 336 * Each word can be limited to sbq->min_shallow_depth bits. 337 */ 338 shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth); 339 depth = ((depth >> sbq->sb.shift) * shallow_depth + 340 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth)); 341 wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1, 342 SBQ_WAKE_BATCH); 343 344 return wake_batch; 345 } 346 347 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth, 348 int shift, bool round_robin, gfp_t flags, int node) 349 { 350 int ret; 351 int i; 352 353 ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node); 354 if (ret) 355 return ret; 356 357 sbq->alloc_hint = alloc_percpu_gfp(unsigned int, flags); 358 if (!sbq->alloc_hint) { 359 sbitmap_free(&sbq->sb); 360 return -ENOMEM; 361 } 362 363 if (depth && !round_robin) { 364 for_each_possible_cpu(i) 365 *per_cpu_ptr(sbq->alloc_hint, i) = prandom_u32() % depth; 366 } 367 368 sbq->min_shallow_depth = UINT_MAX; 369 sbq->wake_batch = sbq_calc_wake_batch(sbq, depth); 370 atomic_set(&sbq->wake_index, 0); 371 atomic_set(&sbq->ws_active, 0); 372 373 sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node); 374 if (!sbq->ws) { 375 free_percpu(sbq->alloc_hint); 376 sbitmap_free(&sbq->sb); 377 return -ENOMEM; 378 } 379 380 for (i = 0; i < SBQ_WAIT_QUEUES; i++) { 381 init_waitqueue_head(&sbq->ws[i].wait); 382 atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch); 383 } 384 385 sbq->round_robin = round_robin; 386 return 0; 387 } 388 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node); 389 390 static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq, 391 unsigned int depth) 392 { 393 unsigned int wake_batch = sbq_calc_wake_batch(sbq, depth); 394 int i; 395 396 if (sbq->wake_batch != wake_batch) { 397 WRITE_ONCE(sbq->wake_batch, wake_batch); 398 /* 399 * Pairs with the memory barrier in sbitmap_queue_wake_up() 400 * to ensure that the batch size is updated before the wait 401 * counts. 402 */ 403 smp_mb(); 404 for (i = 0; i < SBQ_WAIT_QUEUES; i++) 405 atomic_set(&sbq->ws[i].wait_cnt, 1); 406 } 407 } 408 409 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth) 410 { 411 sbitmap_queue_update_wake_batch(sbq, depth); 412 sbitmap_resize(&sbq->sb, depth); 413 } 414 EXPORT_SYMBOL_GPL(sbitmap_queue_resize); 415 416 int __sbitmap_queue_get(struct sbitmap_queue *sbq) 417 { 418 unsigned int hint, depth; 419 int nr; 420 421 hint = this_cpu_read(*sbq->alloc_hint); 422 depth = READ_ONCE(sbq->sb.depth); 423 if (unlikely(hint >= depth)) { 424 hint = depth ? prandom_u32() % depth : 0; 425 this_cpu_write(*sbq->alloc_hint, hint); 426 } 427 nr = sbitmap_get(&sbq->sb, hint, sbq->round_robin); 428 429 if (nr == -1) { 430 /* If the map is full, a hint won't do us much good. */ 431 this_cpu_write(*sbq->alloc_hint, 0); 432 } else if (nr == hint || unlikely(sbq->round_robin)) { 433 /* Only update the hint if we used it. */ 434 hint = nr + 1; 435 if (hint >= depth - 1) 436 hint = 0; 437 this_cpu_write(*sbq->alloc_hint, hint); 438 } 439 440 return nr; 441 } 442 EXPORT_SYMBOL_GPL(__sbitmap_queue_get); 443 444 int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq, 445 unsigned int shallow_depth) 446 { 447 unsigned int hint, depth; 448 int nr; 449 450 WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth); 451 452 hint = this_cpu_read(*sbq->alloc_hint); 453 depth = READ_ONCE(sbq->sb.depth); 454 if (unlikely(hint >= depth)) { 455 hint = depth ? prandom_u32() % depth : 0; 456 this_cpu_write(*sbq->alloc_hint, hint); 457 } 458 nr = sbitmap_get_shallow(&sbq->sb, hint, shallow_depth); 459 460 if (nr == -1) { 461 /* If the map is full, a hint won't do us much good. */ 462 this_cpu_write(*sbq->alloc_hint, 0); 463 } else if (nr == hint || unlikely(sbq->round_robin)) { 464 /* Only update the hint if we used it. */ 465 hint = nr + 1; 466 if (hint >= depth - 1) 467 hint = 0; 468 this_cpu_write(*sbq->alloc_hint, hint); 469 } 470 471 return nr; 472 } 473 EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow); 474 475 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq, 476 unsigned int min_shallow_depth) 477 { 478 sbq->min_shallow_depth = min_shallow_depth; 479 sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth); 480 } 481 EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth); 482 483 static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq) 484 { 485 int i, wake_index; 486 487 if (!atomic_read(&sbq->ws_active)) 488 return NULL; 489 490 wake_index = atomic_read(&sbq->wake_index); 491 for (i = 0; i < SBQ_WAIT_QUEUES; i++) { 492 struct sbq_wait_state *ws = &sbq->ws[wake_index]; 493 494 if (waitqueue_active(&ws->wait)) { 495 if (wake_index != atomic_read(&sbq->wake_index)) 496 atomic_set(&sbq->wake_index, wake_index); 497 return ws; 498 } 499 500 wake_index = sbq_index_inc(wake_index); 501 } 502 503 return NULL; 504 } 505 506 static bool __sbq_wake_up(struct sbitmap_queue *sbq) 507 { 508 struct sbq_wait_state *ws; 509 unsigned int wake_batch; 510 int wait_cnt; 511 512 ws = sbq_wake_ptr(sbq); 513 if (!ws) 514 return false; 515 516 wait_cnt = atomic_dec_return(&ws->wait_cnt); 517 if (wait_cnt <= 0) { 518 int ret; 519 520 wake_batch = READ_ONCE(sbq->wake_batch); 521 522 /* 523 * Pairs with the memory barrier in sbitmap_queue_resize() to 524 * ensure that we see the batch size update before the wait 525 * count is reset. 526 */ 527 smp_mb__before_atomic(); 528 529 /* 530 * For concurrent callers of this, the one that failed the 531 * atomic_cmpxhcg() race should call this function again 532 * to wakeup a new batch on a different 'ws'. 533 */ 534 ret = atomic_cmpxchg(&ws->wait_cnt, wait_cnt, wake_batch); 535 if (ret == wait_cnt) { 536 sbq_index_atomic_inc(&sbq->wake_index); 537 wake_up_nr(&ws->wait, wake_batch); 538 return false; 539 } 540 541 return true; 542 } 543 544 return false; 545 } 546 547 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq) 548 { 549 while (__sbq_wake_up(sbq)) 550 ; 551 } 552 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up); 553 554 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr, 555 unsigned int cpu) 556 { 557 /* 558 * Once the clear bit is set, the bit may be allocated out. 559 * 560 * Orders READ/WRITE on the asssociated instance(such as request 561 * of blk_mq) by this bit for avoiding race with re-allocation, 562 * and its pair is the memory barrier implied in __sbitmap_get_word. 563 * 564 * One invariant is that the clear bit has to be zero when the bit 565 * is in use. 566 */ 567 smp_mb__before_atomic(); 568 sbitmap_deferred_clear_bit(&sbq->sb, nr); 569 570 /* 571 * Pairs with the memory barrier in set_current_state() to ensure the 572 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker 573 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the 574 * waiter. See the comment on waitqueue_active(). 575 */ 576 smp_mb__after_atomic(); 577 sbitmap_queue_wake_up(sbq); 578 579 if (likely(!sbq->round_robin && nr < sbq->sb.depth)) 580 *per_cpu_ptr(sbq->alloc_hint, cpu) = nr; 581 } 582 EXPORT_SYMBOL_GPL(sbitmap_queue_clear); 583 584 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq) 585 { 586 int i, wake_index; 587 588 /* 589 * Pairs with the memory barrier in set_current_state() like in 590 * sbitmap_queue_wake_up(). 591 */ 592 smp_mb(); 593 wake_index = atomic_read(&sbq->wake_index); 594 for (i = 0; i < SBQ_WAIT_QUEUES; i++) { 595 struct sbq_wait_state *ws = &sbq->ws[wake_index]; 596 597 if (waitqueue_active(&ws->wait)) 598 wake_up(&ws->wait); 599 600 wake_index = sbq_index_inc(wake_index); 601 } 602 } 603 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all); 604 605 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m) 606 { 607 bool first; 608 int i; 609 610 sbitmap_show(&sbq->sb, m); 611 612 seq_puts(m, "alloc_hint={"); 613 first = true; 614 for_each_possible_cpu(i) { 615 if (!first) 616 seq_puts(m, ", "); 617 first = false; 618 seq_printf(m, "%u", *per_cpu_ptr(sbq->alloc_hint, i)); 619 } 620 seq_puts(m, "}\n"); 621 622 seq_printf(m, "wake_batch=%u\n", sbq->wake_batch); 623 seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index)); 624 seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active)); 625 626 seq_puts(m, "ws={\n"); 627 for (i = 0; i < SBQ_WAIT_QUEUES; i++) { 628 struct sbq_wait_state *ws = &sbq->ws[i]; 629 630 seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n", 631 atomic_read(&ws->wait_cnt), 632 waitqueue_active(&ws->wait) ? "active" : "inactive"); 633 } 634 seq_puts(m, "}\n"); 635 636 seq_printf(m, "round_robin=%d\n", sbq->round_robin); 637 seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth); 638 } 639 EXPORT_SYMBOL_GPL(sbitmap_queue_show); 640 641 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq, 642 struct sbq_wait_state *ws, 643 struct sbq_wait *sbq_wait) 644 { 645 if (!sbq_wait->sbq) { 646 sbq_wait->sbq = sbq; 647 atomic_inc(&sbq->ws_active); 648 add_wait_queue(&ws->wait, &sbq_wait->wait); 649 } 650 } 651 EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue); 652 653 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait) 654 { 655 list_del_init(&sbq_wait->wait.entry); 656 if (sbq_wait->sbq) { 657 atomic_dec(&sbq_wait->sbq->ws_active); 658 sbq_wait->sbq = NULL; 659 } 660 } 661 EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue); 662 663 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq, 664 struct sbq_wait_state *ws, 665 struct sbq_wait *sbq_wait, int state) 666 { 667 if (!sbq_wait->sbq) { 668 atomic_inc(&sbq->ws_active); 669 sbq_wait->sbq = sbq; 670 } 671 prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state); 672 } 673 EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait); 674 675 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws, 676 struct sbq_wait *sbq_wait) 677 { 678 finish_wait(&ws->wait, &sbq_wait->wait); 679 if (sbq_wait->sbq) { 680 atomic_dec(&sbq->ws_active); 681 sbq_wait->sbq = NULL; 682 } 683 } 684 EXPORT_SYMBOL_GPL(sbitmap_finish_wait); 685