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 tags
  72  */
  73 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
  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         if (include_reserve) {
  90                 bt = &tags->breserved_tags;
  91                 if (waitqueue_active(&bt->bs[0].wait))
  92                         wake_up(&bt->bs[0].wait);
  93         }
  94 }
  95
  96 /*
  97  * If a previously busy queue goes inactive, potential waiters could now
  98  * be allowed to queue. Wake them up and check.
  99  */
 100 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
 101 {
 102         struct blk_mq_tags *tags = hctx->tags;
 103
 104         if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
 105                 return;
 106
 107         atomic_dec(&tags->active_queues);
 108
 109         blk_mq_tag_wakeup_all(tags, false);
 110 }
 111
 112 /*
 113  * For shared tag users, we track the number of currently active users
 114  * and attempt to provide a fair share of the tag depth for each of them.
 115  */
 116 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
 117                                   struct blk_mq_bitmap_tags *bt)
 118 {
 119         unsigned int depth, users;
 120
 121         if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
 122                 return true;
 123         if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
 124                 return true;
 125
 126         /*
 127          * Don't try dividing an ant
 128          */
 129         if (bt->depth == 1)
 130                 return true;
 131
 132         users = atomic_read(&hctx->tags->active_queues);
 133         if (!users)
 134                 return true;
 135
 136         /*
 137          * Allow at least some tags
 138          */
 139         depth = max((bt->depth + users - 1) / users, 4U);
 140         return atomic_read(&hctx->nr_active) < depth;
 141 }
 142
 143 static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag)
 144 {
 145         int tag, org_last_tag, end;
 146         bool wrap = last_tag != 0;
 147
 148         org_last_tag = last_tag;
 149         end = bm->depth;
 150         do {
 151 restart:
 152                 tag = find_next_zero_bit(&bm->word, end, last_tag);
 153                 if (unlikely(tag >= end)) {
 154                         /*
 155                          * We started with an offset, start from 0 to
 156                          * exhaust the map.
 157                          */
 158                         if (wrap) {
 159                                 wrap = false;
 160                                 end = org_last_tag;
 161                                 last_tag = 0;
 162                                 goto restart;
 163                         }
 164                         return -1;
 165                 }
 166                 last_tag = tag + 1;
 167         } while (test_and_set_bit(tag, &bm->word));
 168
 169         return tag;
 170 }
 171
 172 /*
 173  * Straight forward bitmap tag implementation, where each bit is a tag
 174  * (cleared == free, and set == busy). The small twist is using per-cpu
 175  * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
 176  * contexts. This enables us to drastically limit the space searched,
 177  * without dirtying an extra shared cacheline like we would if we stored
 178  * the cache value inside the shared blk_mq_bitmap_tags structure. On top
 179  * of that, each word of tags is in a separate cacheline. This means that
 180  * multiple users will tend to stick to different cachelines, at least
 181  * until the map is exhausted.
 182  */
 183 static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
 184                     unsigned int *tag_cache)
 185 {
 186         unsigned int last_tag, org_last_tag;
 187         int index, i, tag;
 188
 189         if (!hctx_may_queue(hctx, bt))
 190                 return -1;
 191
 192         last_tag = org_last_tag = *tag_cache;
 193         index = TAG_TO_INDEX(bt, last_tag);
 194
 195         for (i = 0; i < bt->map_nr; i++) {
 196                 tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag));
 197                 if (tag != -1) {
 198                         tag += (index << bt->bits_per_word);
 199                         goto done;
 200                 }
 201
 202                 last_tag = 0;
 203                 if (++index >= bt->map_nr)
 204                         index = 0;
 205         }
 206
 207         *tag_cache = 0;
 208         return -1;
 209
 210         /*
 211          * Only update the cache from the allocation path, if we ended
 212          * up using the specific cached tag.
 213          */
 214 done:
 215         if (tag == org_last_tag) {
 216                 last_tag = tag + 1;
 217                 if (last_tag >= bt->depth - 1)
 218                         last_tag = 0;
 219
 220                 *tag_cache = last_tag;
 221         }
 222
 223         return tag;
 224 }
 225
 226 static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
 227                                          struct blk_mq_hw_ctx *hctx)
 228 {
 229         struct bt_wait_state *bs;
 230         int wait_index;
 231
 232         if (!hctx)
 233                 return &bt->bs[0];
 234
 235         wait_index = atomic_read(&hctx->wait_index);
 236         bs = &bt->bs[wait_index];
 237         bt_index_atomic_inc(&hctx->wait_index);
 238         return bs;
 239 }
 240
 241 static int bt_get(struct blk_mq_alloc_data *data,
 242                 struct blk_mq_bitmap_tags *bt,
 243                 struct blk_mq_hw_ctx *hctx,
 244                 unsigned int *last_tag)
 245 {
 246         struct bt_wait_state *bs;
 247         DEFINE_WAIT(wait);
 248         int tag;
 249
 250         tag = __bt_get(hctx, bt, last_tag);
 251         if (tag != -1)
 252                 return tag;
 253
 254         if (!(data->gfp & __GFP_WAIT))
 255                 return -1;
 256
 257         bs = bt_wait_ptr(bt, hctx);
 258         do {
 259                 prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
 260
 261                 tag = __bt_get(hctx, bt, last_tag);
 262                 if (tag != -1)
 263                         break;
 264
 265                 /*
 266                  * We're out of tags on this hardware queue, kick any
 267                  * pending IO submits before going to sleep waiting for
 268                  * some to complete.
 269                  */
 270                 blk_mq_run_hw_queue(hctx, false);
 271
 272                 /*
 273                  * Retry tag allocation after running the hardware queue,
 274                  * as running the queue may also have found completions.
 275                  */
 276                 tag = __bt_get(hctx, bt, last_tag);
 277                 if (tag != -1)
 278                         break;
 279
 280                 blk_mq_put_ctx(data->ctx);
 281
 282                 io_schedule();
 283
 284                 data->ctx = blk_mq_get_ctx(data->q);
 285                 data->hctx = data->q->mq_ops->map_queue(data->q,
 286                                 data->ctx->cpu);
 287                 if (data->reserved) {
 288                         bt = &data->hctx->tags->breserved_tags;
 289                 } else {
 290                         last_tag = &data->ctx->last_tag;
 291                         hctx = data->hctx;
 292                         bt = &hctx->tags->bitmap_tags;
 293                 }
 294                 finish_wait(&bs->wait, &wait);
 295                 bs = bt_wait_ptr(bt, hctx);
 296         } while (1);
 297
 298         finish_wait(&bs->wait, &wait);
 299         return tag;
 300 }
 301
 302 static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
 303 {
 304         int tag;
 305
 306         tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
 307                         &data->ctx->last_tag);
 308         if (tag >= 0)
 309                 return tag + data->hctx->tags->nr_reserved_tags;
 310
 311         return BLK_MQ_TAG_FAIL;
 312 }
 313
 314 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
 315 {
 316         int tag, zero = 0;
 317
 318         if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
 319                 WARN_ON_ONCE(1);
 320                 return BLK_MQ_TAG_FAIL;
 321         }
 322
 323         tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero);
 324         if (tag < 0)
 325                 return BLK_MQ_TAG_FAIL;
 326
 327         return tag;
 328 }
 329
 330 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
 331 {
 332         if (!data->reserved)
 333                 return __blk_mq_get_tag(data);
 334
 335         return __blk_mq_get_reserved_tag(data);
 336 }
 337
 338 static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
 339 {
 340         int i, wake_index;
 341
 342         wake_index = atomic_read(&bt->wake_index);
 343         for (i = 0; i < BT_WAIT_QUEUES; i++) {
 344                 struct bt_wait_state *bs = &bt->bs[wake_index];
 345
 346                 if (waitqueue_active(&bs->wait)) {
 347                         int o = atomic_read(&bt->wake_index);
 348                         if (wake_index != o)
 349                                 atomic_cmpxchg(&bt->wake_index, o, wake_index);
 350
 351                         return bs;
 352                 }
 353
 354                 wake_index = bt_index_inc(wake_index);
 355         }
 356
 357         return NULL;
 358 }
 359
 360 static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
 361 {
 362         const int index = TAG_TO_INDEX(bt, tag);
 363         struct bt_wait_state *bs;
 364         int wait_cnt;
 365
 366         clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word);
 367
 368         /* Ensure that the wait list checks occur after clear_bit(). */
 369         smp_mb();
 370
 371         bs = bt_wake_ptr(bt);
 372         if (!bs)
 373                 return;
 374
 375         wait_cnt = atomic_dec_return(&bs->wait_cnt);
 376         if (unlikely(wait_cnt < 0))
 377                 wait_cnt = atomic_inc_return(&bs->wait_cnt);
 378         if (wait_cnt == 0) {
 379                 atomic_add(bt->wake_cnt, &bs->wait_cnt);
 380                 bt_index_atomic_inc(&bt->wake_index);
 381                 wake_up(&bs->wait);
 382         }
 383 }
 384
 385 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
 386                     unsigned int *last_tag)
 387 {
 388         struct blk_mq_tags *tags = hctx->tags;
 389
 390         if (tag >= tags->nr_reserved_tags) {
 391                 const int real_tag = tag - tags->nr_reserved_tags;
 392
 393                 BUG_ON(real_tag >= tags->nr_tags);
 394                 bt_clear_tag(&tags->bitmap_tags, real_tag);
 395                 *last_tag = real_tag;
 396         } else {
 397                 BUG_ON(tag >= tags->nr_reserved_tags);
 398                 bt_clear_tag(&tags->breserved_tags, tag);
 399         }
 400 }
 401
 402 static void bt_for_each(struct blk_mq_hw_ctx *hctx,
 403                 struct blk_mq_bitmap_tags *bt, unsigned int off,
 404                 busy_iter_fn *fn, void *data, bool reserved)
 405 {
 406         struct request *rq;
 407         int bit, i;
 408
 409         for (i = 0; i < bt->map_nr; i++) {
 410                 struct blk_align_bitmap *bm = &bt->map[i];
 411
 412                 for (bit = find_first_bit(&bm->word, bm->depth);
 413                      bit < bm->depth;
 414                      bit = find_next_bit(&bm->word, bm->depth, bit + 1)) {
 415                         rq = blk_mq_tag_to_rq(hctx->tags, off + bit);
 416                         if (rq->q == hctx->queue)
 417                                 fn(hctx, rq, data, reserved);
 418                 }
 419
 420                 off += (1 << bt->bits_per_word);
 421         }
 422 }
 423
 424 void blk_mq_tag_busy_iter(struct blk_mq_hw_ctx *hctx, busy_iter_fn *fn,
 425                 void *priv)
 426 {
 427         struct blk_mq_tags *tags = hctx->tags;
 428
 429         if (tags->nr_reserved_tags)
 430                 bt_for_each(hctx, &tags->breserved_tags, 0, fn, priv, true);
 431         bt_for_each(hctx, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv,
 432                         false);
 433 }
 434 EXPORT_SYMBOL(blk_mq_tag_busy_iter);
 435
 436 static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
 437 {
 438         unsigned int i, used;
 439
 440         for (i = 0, used = 0; i < bt->map_nr; i++) {
 441                 struct blk_align_bitmap *bm = &bt->map[i];
 442
 443                 used += bitmap_weight(&bm->word, bm->depth);
 444         }
 445
 446         return bt->depth - used;
 447 }
 448
 449 static void bt_update_count(struct blk_mq_bitmap_tags *bt,
 450                             unsigned int depth)
 451 {
 452         unsigned int tags_per_word = 1U << bt->bits_per_word;
 453         unsigned int map_depth = depth;
 454
 455         if (depth) {
 456                 int i;
 457
 458                 for (i = 0; i < bt->map_nr; i++) {
 459                         bt->map[i].depth = min(map_depth, tags_per_word);
 460                         map_depth -= bt->map[i].depth;
 461                 }
 462         }
 463
 464         bt->wake_cnt = BT_WAIT_BATCH;
 465         if (bt->wake_cnt > depth / BT_WAIT_QUEUES)
 466                 bt->wake_cnt = max(1U, depth / BT_WAIT_QUEUES);
 467
 468         bt->depth = depth;
 469 }
 470
 471 static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
 472                         int node, bool reserved)
 473 {
 474         int i;
 475
 476         bt->bits_per_word = ilog2(BITS_PER_LONG);
 477
 478         /*
 479          * Depth can be zero for reserved tags, that's not a failure
 480          * condition.
 481          */
 482         if (depth) {
 483                 unsigned int nr, tags_per_word;
 484
 485                 tags_per_word = (1 << bt->bits_per_word);
 486
 487                 /*
 488                  * If the tag space is small, shrink the number of tags
 489                  * per word so we spread over a few cachelines, at least.
 490                  * If less than 4 tags, just forget about it, it's not
 491                  * going to work optimally anyway.
 492                  */
 493                 if (depth >= 4) {
 494                         while (tags_per_word * 4 > depth) {
 495                                 bt->bits_per_word--;
 496                                 tags_per_word = (1 << bt->bits_per_word);
 497                         }
 498                 }
 499
 500                 nr = ALIGN(depth, tags_per_word) / tags_per_word;
 501                 bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
 502                                                 GFP_KERNEL, node);
 503                 if (!bt->map)
 504                         return -ENOMEM;
 505
 506                 bt->map_nr = nr;
 507         }
 508
 509         bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
 510         if (!bt->bs) {
 511                 kfree(bt->map);
 512                 return -ENOMEM;
 513         }
 514
 515         bt_update_count(bt, depth);
 516
 517         for (i = 0; i < BT_WAIT_QUEUES; i++) {
 518                 init_waitqueue_head(&bt->bs[i].wait);
 519                 atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt);
 520         }
 521
 522         return 0;
 523 }
 524
 525 static void bt_free(struct blk_mq_bitmap_tags *bt)
 526 {
 527         kfree(bt->map);
 528         kfree(bt->bs);
 529 }
 530
 531 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
 532                                                    int node)
 533 {
 534         unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
 535
 536         if (bt_alloc(&tags->bitmap_tags, depth, node, false))
 537                 goto enomem;
 538         if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
 539                 goto enomem;
 540
 541         return tags;
 542 enomem:
 543         bt_free(&tags->bitmap_tags);
 544         kfree(tags);
 545         return NULL;
 546 }
 547
 548 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
 549                                      unsigned int reserved_tags, int node)
 550 {
 551         struct blk_mq_tags *tags;
 552
 553         if (total_tags > BLK_MQ_TAG_MAX) {
 554                 pr_err("blk-mq: tag depth too large\n");
 555                 return NULL;
 556         }
 557
 558         tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
 559         if (!tags)
 560                 return NULL;
 561
 562         tags->nr_tags = total_tags;
 563         tags->nr_reserved_tags = reserved_tags;
 564
 565         return blk_mq_init_bitmap_tags(tags, node);
 566 }
 567
 568 void blk_mq_free_tags(struct blk_mq_tags *tags)
 569 {
 570         bt_free(&tags->bitmap_tags);
 571         bt_free(&tags->breserved_tags);
 572         kfree(tags);
 573 }
 574
 575 void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
 576 {
 577         unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
 578
 579         *tag = prandom_u32() % depth;
 580 }
 581
 582 int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
 583 {
 584         tdepth -= tags->nr_reserved_tags;
 585         if (tdepth > tags->nr_tags)
 586                 return -EINVAL;
 587
 588         /*
 589          * Don't need (or can't) update reserved tags here, they remain
 590          * static and should never need resizing.
 591          */
 592         bt_update_count(&tags->bitmap_tags, tdepth);
 593         blk_mq_tag_wakeup_all(tags, false);
 594         return 0;
 595 }
 596
 597 /**
 598  * blk_mq_unique_tag() - return a tag that is unique queue-wide
 599  * @rq: request for which to compute a unique tag
 600  *
 601  * The tag field in struct request is unique per hardware queue but not over
 602  * all hardware queues. Hence this function that returns a tag with the
 603  * hardware context index in the upper bits and the per hardware queue tag in
 604  * the lower bits.
 605  *
 606  * Note: When called for a request that is queued on a non-multiqueue request
 607  * queue, the hardware context index is set to zero.
 608  */
 609 u32 blk_mq_unique_tag(struct request *rq)
 610 {
 611         struct request_queue *q = rq->q;
 612         struct blk_mq_hw_ctx *hctx;
 613         int hwq = 0;
 614
 615         if (q->mq_ops) {
 616                 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
 617                 hwq = hctx->queue_num;
 618         }
 619
 620         return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
 621                 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
 622 }
 623 EXPORT_SYMBOL(blk_mq_unique_tag);
 624
 625 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
 626 {
 627         char *orig_page = page;
 628         unsigned int free, res;
 629
 630         if (!tags)
 631                 return 0;
 632
 633         page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
 634                         "bits_per_word=%u\n",
 635                         tags->nr_tags, tags->nr_reserved_tags,
 636                         tags->bitmap_tags.bits_per_word);
 637
 638         free = bt_unused_tags(&tags->bitmap_tags);
 639         res = bt_unused_tags(&tags->breserved_tags);
 640
 641         page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
 642         page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
 643
 644         return page - orig_page;
 645 }