block/blk-mq-tag.c

   1 /*
   2  * Fast and scalable bitmap tagging variant. Uses sparser bitmaps spread
   3  * over multiple cachelines to avoid ping-pong between multiple submitters
   4  * or submitter and completer. Uses rolling wakeups to avoid falling of
   5  * the scaling cliff when we run out of tags and have to start putting
   6  * submitters to sleep.
   7  *
   8  * Uses active queue tracking to support fairer distribution of tags
   9  * between multiple submitters when a shared tag map is used.
  10  *
  11  * Copyright (C) 2013-2014 Jens Axboe
  12  */
  13 #include <linux/kernel.h>
  14 #include <linux/module.h>
  15 #include <linux/random.h>
  16
  17 #include <linux/blk-mq.h>
  18 #include "blk.h"
  19 #include "blk-mq.h"
  20 #include "blk-mq-tag.h"
  21
  22 static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
  23 {
  24         int i;
  25
  26         for (i = 0; i < bt->map_nr; i++) {
  27                 struct blk_align_bitmap *bm = &bt->map[i];
  28                 int ret;
  29
  30                 ret = find_first_zero_bit(&bm->word, bm->depth);
  31                 if (ret < bm->depth)
  32                         return true;
  33         }
  34
  35         return false;
  36 }
  37
  38 bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
  39 {
  40         if (!tags)
  41                 return true;
  42
  43         return bt_has_free_tags(&tags->bitmap_tags);
  44 }
  45
  46 static inline int bt_index_inc(int index)
  47 {
  48         return (index + 1) & (BT_WAIT_QUEUES - 1);
  49 }
  50
  51 static inline void bt_index_atomic_inc(atomic_t *index)
  52 {
  53         int old = atomic_read(index);
  54         int new = bt_index_inc(old);
  55         atomic_cmpxchg(index, old, new);
  56 }
  57
  58 /*
  59  * If a previously inactive queue goes active, bump the active user count.
  60  */
  61 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
  62 {
  63         if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
  64             !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
  65                 atomic_inc(&hctx->tags->active_queues);
  66
  67         return true;
  68 }
  69
  70 /*
  71  * Wakeup all potentially sleeping on normal (non-reserved) tags
  72  */
  73 static void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags)
  74 {
  75         struct blk_mq_bitmap_tags *bt;
  76         int i, wake_index;
  77
  78         bt = &tags->bitmap_tags;
  79         wake_index = atomic_read(&bt->wake_index);
  80         for (i = 0; i < BT_WAIT_QUEUES; i++) {
  81                 struct bt_wait_state *bs = &bt->bs[wake_index];
  82
  83                 if (waitqueue_active(&bs->wait))
  84                         wake_up(&bs->wait);
  85
  86                 wake_index = bt_index_inc(wake_index);
  87         }
  88 }
  89
  90 /*
  91  * If a previously busy queue goes inactive, potential waiters could now
  92  * be allowed to queue. Wake them up and check.
  93  */
  94 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
  95 {
  96         struct blk_mq_tags *tags = hctx->tags;
  97
  98         if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
  99                 return;
 100
 101         atomic_dec(&tags->active_queues);
 102
 103         blk_mq_tag_wakeup_all(tags);
 104 }
 105
 106 /*
 107  * For shared tag users, we track the number of currently active users
 108  * and attempt to provide a fair share of the tag depth for each of them.
 109  */
 110 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
 111                                   struct blk_mq_bitmap_tags *bt)
 112 {
 113         unsigned int depth, users;
 114
 115         if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
 116                 return true;
 117         if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
 118                 return true;
 119
 120         /*
 121          * Don't try dividing an ant
 122          */
 123         if (bt->depth == 1)
 124                 return true;
 125
 126         users = atomic_read(&hctx->tags->active_queues);
 127         if (!users)
 128                 return true;
 129
 130         /*
 131          * Allow at least some tags
 132          */
 133         depth = max((bt->depth + users - 1) / users, 4U);
 134         return atomic_read(&hctx->nr_active) < depth;
 135 }
 136
 137 static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag)
 138 {
 139         int tag, org_last_tag, end;
 140
 141         org_last_tag = last_tag;
 142         end = bm->depth;
 143         do {
 144 restart:
 145                 tag = find_next_zero_bit(&bm->word, end, last_tag);
 146                 if (unlikely(tag >= end)) {
 147                         /*
 148                          * We started with an offset, start from 0 to
 149                          * exhaust the map.
 150                          */
 151                         if (org_last_tag && last_tag) {
 152                                 end = last_tag;
 153                                 last_tag = 0;
 154                                 goto restart;
 155                         }
 156                         return -1;
 157                 }
 158                 last_tag = tag + 1;
 159         } while (test_and_set_bit_lock(tag, &bm->word));
 160
 161         return tag;
 162 }
 163
 164 /*
 165  * Straight forward bitmap tag implementation, where each bit is a tag
 166  * (cleared == free, and set == busy). The small twist is using per-cpu
 167  * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
 168  * contexts. This enables us to drastically limit the space searched,
 169  * without dirtying an extra shared cacheline like we would if we stored
 170  * the cache value inside the shared blk_mq_bitmap_tags structure. On top
 171  * of that, each word of tags is in a separate cacheline. This means that
 172  * multiple users will tend to stick to different cachelines, at least
 173  * until the map is exhausted.
 174  */
 175 static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
 176                     unsigned int *tag_cache)
 177 {
 178         unsigned int last_tag, org_last_tag;
 179         int index, i, tag;
 180
 181         if (!hctx_may_queue(hctx, bt))
 182                 return -1;
 183
 184         last_tag = org_last_tag = *tag_cache;
 185         index = TAG_TO_INDEX(bt, last_tag);
 186
 187         for (i = 0; i < bt->map_nr; i++) {
 188                 tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag));
 189                 if (tag != -1) {
 190                         tag += (index << bt->bits_per_word);
 191                         goto done;
 192                 }
 193
 194                 last_tag = 0;
 195                 if (++index >= bt->map_nr)
 196                         index = 0;
 197         }
 198
 199         *tag_cache = 0;
 200         return -1;
 201
 202         /*
 203          * Only update the cache from the allocation path, if we ended
 204          * up using the specific cached tag.
 205          */
 206 done:
 207         if (tag == org_last_tag) {
 208                 last_tag = tag + 1;
 209                 if (last_tag >= bt->depth - 1)
 210                         last_tag = 0;
 211
 212                 *tag_cache = last_tag;
 213         }
 214
 215         return tag;
 216 }
 217
 218 static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
 219                                          struct blk_mq_hw_ctx *hctx)
 220 {
 221         struct bt_wait_state *bs;
 222         int wait_index;
 223
 224         if (!hctx)
 225                 return &bt->bs[0];
 226
 227         wait_index = atomic_read(&hctx->wait_index);
 228         bs = &bt->bs[wait_index];
 229         bt_index_atomic_inc(&hctx->wait_index);
 230         return bs;
 231 }
 232
 233 static int bt_get(struct blk_mq_alloc_data *data,
 234                 struct blk_mq_bitmap_tags *bt,
 235                 struct blk_mq_hw_ctx *hctx,
 236                 unsigned int *last_tag)
 237 {
 238         struct bt_wait_state *bs;
 239         DEFINE_WAIT(wait);
 240         int tag;
 241
 242         tag = __bt_get(hctx, bt, last_tag);
 243         if (tag != -1)
 244                 return tag;
 245
 246         if (!(data->gfp & __GFP_WAIT))
 247                 return -1;
 248
 249         bs = bt_wait_ptr(bt, hctx);
 250         do {
 251                 prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
 252
 253                 tag = __bt_get(hctx, bt, last_tag);
 254                 if (tag != -1)
 255                         break;
 256
 257                 blk_mq_put_ctx(data->ctx);
 258
 259                 io_schedule();
 260
 261                 data->ctx = blk_mq_get_ctx(data->q);
 262                 data->hctx = data->q->mq_ops->map_queue(data->q,
 263                                 data->ctx->cpu);
 264                 if (data->reserved) {
 265                         bt = &data->hctx->tags->breserved_tags;
 266                 } else {
 267                         last_tag = &data->ctx->last_tag;
 268                         hctx = data->hctx;
 269                         bt = &hctx->tags->bitmap_tags;
 270                 }
 271                 finish_wait(&bs->wait, &wait);
 272                 bs = bt_wait_ptr(bt, hctx);
 273         } while (1);
 274
 275         finish_wait(&bs->wait, &wait);
 276         return tag;
 277 }
 278
 279 static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
 280 {
 281         int tag;
 282
 283         tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
 284                         &data->ctx->last_tag);
 285         if (tag >= 0)
 286                 return tag + data->hctx->tags->nr_reserved_tags;
 287
 288         return BLK_MQ_TAG_FAIL;
 289 }
 290
 291 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
 292 {
 293         int tag, zero = 0;
 294
 295         if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
 296                 WARN_ON_ONCE(1);
 297                 return BLK_MQ_TAG_FAIL;
 298         }
 299
 300         tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero);
 301         if (tag < 0)
 302                 return BLK_MQ_TAG_FAIL;
 303
 304         return tag;
 305 }
 306
 307 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
 308 {
 309         if (!data->reserved)
 310                 return __blk_mq_get_tag(data);
 311
 312         return __blk_mq_get_reserved_tag(data);
 313 }
 314
 315 static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
 316 {
 317         int i, wake_index;
 318
 319         wake_index = atomic_read(&bt->wake_index);
 320         for (i = 0; i < BT_WAIT_QUEUES; i++) {
 321                 struct bt_wait_state *bs = &bt->bs[wake_index];
 322
 323                 if (waitqueue_active(&bs->wait)) {
 324                         int o = atomic_read(&bt->wake_index);
 325                         if (wake_index != o)
 326                                 atomic_cmpxchg(&bt->wake_index, o, wake_index);
 327
 328                         return bs;
 329                 }
 330
 331                 wake_index = bt_index_inc(wake_index);
 332         }
 333
 334         return NULL;
 335 }
 336
 337 static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
 338 {
 339         const int index = TAG_TO_INDEX(bt, tag);
 340         struct bt_wait_state *bs;
 341         int wait_cnt;
 342
 343         /*
 344          * The unlock memory barrier need to order access to req in free
 345          * path and clearing tag bit
 346          */
 347         clear_bit_unlock(TAG_TO_BIT(bt, tag), &bt->map[index].word);
 348
 349         bs = bt_wake_ptr(bt);
 350         if (!bs)
 351                 return;
 352
 353         wait_cnt = atomic_dec_return(&bs->wait_cnt);
 354         if (wait_cnt == 0) {
 355 wake:
 356                 atomic_add(bt->wake_cnt, &bs->wait_cnt);
 357                 bt_index_atomic_inc(&bt->wake_index);
 358                 wake_up(&bs->wait);
 359         } else if (wait_cnt < 0) {
 360                 wait_cnt = atomic_inc_return(&bs->wait_cnt);
 361                 if (!wait_cnt)
 362                         goto wake;
 363         }
 364 }
 365
 366 static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
 367 {
 368         BUG_ON(tag >= tags->nr_tags);
 369
 370         bt_clear_tag(&tags->bitmap_tags, tag);
 371 }
 372
 373 static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
 374                                       unsigned int tag)
 375 {
 376         BUG_ON(tag >= tags->nr_reserved_tags);
 377
 378         bt_clear_tag(&tags->breserved_tags, tag);
 379 }
 380
 381 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
 382                     unsigned int *last_tag)
 383 {
 384         struct blk_mq_tags *tags = hctx->tags;
 385
 386         if (tag >= tags->nr_reserved_tags) {
 387                 const int real_tag = tag - tags->nr_reserved_tags;
 388
 389                 __blk_mq_put_tag(tags, real_tag);
 390                 *last_tag = real_tag;
 391         } else
 392                 __blk_mq_put_reserved_tag(tags, tag);
 393 }
 394
 395 static void bt_for_each_free(struct blk_mq_bitmap_tags *bt,
 396                              unsigned long *free_map, unsigned int off)
 397 {
 398         int i;
 399
 400         for (i = 0; i < bt->map_nr; i++) {
 401                 struct blk_align_bitmap *bm = &bt->map[i];
 402                 int bit = 0;
 403
 404                 do {
 405                         bit = find_next_zero_bit(&bm->word, bm->depth, bit);
 406                         if (bit >= bm->depth)
 407                                 break;
 408
 409                         __set_bit(bit + off, free_map);
 410                         bit++;
 411                 } while (1);
 412
 413                 off += (1 << bt->bits_per_word);
 414         }
 415 }
 416
 417 void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
 418                           void (*fn)(void *, unsigned long *), void *data)
 419 {
 420         unsigned long *tag_map;
 421         size_t map_size;
 422
 423         map_size = ALIGN(tags->nr_tags, BITS_PER_LONG) / BITS_PER_LONG;
 424         tag_map = kzalloc(map_size * sizeof(unsigned long), GFP_ATOMIC);
 425         if (!tag_map)
 426                 return;
 427
 428         bt_for_each_free(&tags->bitmap_tags, tag_map, tags->nr_reserved_tags);
 429         if (tags->nr_reserved_tags)
 430                 bt_for_each_free(&tags->breserved_tags, tag_map, 0);
 431
 432         fn(data, tag_map);
 433         kfree(tag_map);
 434 }
 435 EXPORT_SYMBOL(blk_mq_tag_busy_iter);
 436
 437 static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
 438 {
 439         unsigned int i, used;
 440
 441         for (i = 0, used = 0; i < bt->map_nr; i++) {
 442                 struct blk_align_bitmap *bm = &bt->map[i];
 443
 444                 used += bitmap_weight(&bm->word, bm->depth);
 445         }
 446
 447         return bt->depth - used;
 448 }
 449
 450 static void bt_update_count(struct blk_mq_bitmap_tags *bt,
 451                             unsigned int depth)
 452 {
 453         unsigned int tags_per_word = 1U << bt->bits_per_word;
 454         unsigned int map_depth = depth;
 455
 456         if (depth) {
 457                 int i;
 458
 459                 for (i = 0; i < bt->map_nr; i++) {
 460                         bt->map[i].depth = min(map_depth, tags_per_word);
 461                         map_depth -= bt->map[i].depth;
 462                 }
 463         }
 464
 465         bt->wake_cnt = BT_WAIT_BATCH;
 466         if (bt->wake_cnt > depth / 4)
 467                 bt->wake_cnt = max(1U, depth / 4);
 468
 469         bt->depth = depth;
 470 }
 471
 472 static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
 473                         int node, bool reserved)
 474 {
 475         int i;
 476
 477         bt->bits_per_word = ilog2(BITS_PER_LONG);
 478
 479         /*
 480          * Depth can be zero for reserved tags, that's not a failure
 481          * condition.
 482          */
 483         if (depth) {
 484                 unsigned int nr, tags_per_word;
 485
 486                 tags_per_word = (1 << bt->bits_per_word);
 487
 488                 /*
 489                  * If the tag space is small, shrink the number of tags
 490                  * per word so we spread over a few cachelines, at least.
 491                  * If less than 4 tags, just forget about it, it's not
 492                  * going to work optimally anyway.
 493                  */
 494                 if (depth >= 4) {
 495                         while (tags_per_word * 4 > depth) {
 496                                 bt->bits_per_word--;
 497                                 tags_per_word = (1 << bt->bits_per_word);
 498                         }
 499                 }
 500
 501                 nr = ALIGN(depth, tags_per_word) / tags_per_word;
 502                 bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
 503                                                 GFP_KERNEL, node);
 504                 if (!bt->map)
 505                         return -ENOMEM;
 506
 507                 bt->map_nr = nr;
 508         }
 509
 510         bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
 511         if (!bt->bs) {
 512                 kfree(bt->map);
 513                 return -ENOMEM;
 514         }
 515
 516         bt_update_count(bt, depth);
 517
 518         for (i = 0; i < BT_WAIT_QUEUES; i++) {
 519                 init_waitqueue_head(&bt->bs[i].wait);
 520                 atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt);
 521         }
 522
 523         return 0;
 524 }
 525
 526 static void bt_free(struct blk_mq_bitmap_tags *bt)
 527 {
 528         kfree(bt->map);
 529         kfree(bt->bs);
 530 }
 531
 532 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
 533                                                    int node)
 534 {
 535         unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
 536
 537         if (bt_alloc(&tags->bitmap_tags, depth, node, false))
 538                 goto enomem;
 539         if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
 540                 goto enomem;
 541
 542         return tags;
 543 enomem:
 544         bt_free(&tags->bitmap_tags);
 545         kfree(tags);
 546         return NULL;
 547 }
 548
 549 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
 550                                      unsigned int reserved_tags, int node)
 551 {
 552         struct blk_mq_tags *tags;
 553
 554         if (total_tags > BLK_MQ_TAG_MAX) {
 555                 pr_err("blk-mq: tag depth too large\n");
 556                 return NULL;
 557         }
 558
 559         tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
 560         if (!tags)
 561                 return NULL;
 562
 563         tags->nr_tags = total_tags;
 564         tags->nr_reserved_tags = reserved_tags;
 565
 566         return blk_mq_init_bitmap_tags(tags, node);
 567 }
 568
 569 void blk_mq_free_tags(struct blk_mq_tags *tags)
 570 {
 571         bt_free(&tags->bitmap_tags);
 572         bt_free(&tags->breserved_tags);
 573         kfree(tags);
 574 }
 575
 576 void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
 577 {
 578         unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
 579
 580         *tag = prandom_u32() % depth;
 581 }
 582
 583 int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
 584 {
 585         tdepth -= tags->nr_reserved_tags;
 586         if (tdepth > tags->nr_tags)
 587                 return -EINVAL;
 588
 589         /*
 590          * Don't need (or can't) update reserved tags here, they remain
 591          * static and should never need resizing.
 592          */
 593         bt_update_count(&tags->bitmap_tags, tdepth);
 594         blk_mq_tag_wakeup_all(tags);
 595         return 0;
 596 }
 597
 598 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
 599 {
 600         char *orig_page = page;
 601         unsigned int free, res;
 602
 603         if (!tags)
 604                 return 0;
 605
 606         page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
 607                         "bits_per_word=%u\n",
 608                         tags->nr_tags, tags->nr_reserved_tags,
 609                         tags->bitmap_tags.bits_per_word);
 610
 611         free = bt_unused_tags(&tags->bitmap_tags);
 612         res = bt_unused_tags(&tags->breserved_tags);
 613
 614         page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
 615         page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
 616
 617         return page - orig_page;
 618 }