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