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                          bool nowrap)
 145 {
 146         int tag, org_last_tag = last_tag;
 147
 148         while (1) {
 149                 tag = find_next_zero_bit(&bm->word, bm->depth, last_tag);
 150                 if (unlikely(tag >= bm->depth)) {
 151                         /*
 152                          * We started with an offset, and we didn't reset the
 153                          * offset to 0 in a failure case, so start from 0 to
 154                          * exhaust the map.
 155                          */
 156                         if (org_last_tag && last_tag && !nowrap) {
 157                                 last_tag = org_last_tag = 0;
 158                                 continue;
 159                         }
 160                         return -1;
 161                 }
 162
 163                 if (!test_and_set_bit(tag, &bm->word))
 164                         break;
 165
 166                 last_tag = tag + 1;
 167                 if (last_tag >= bm->depth - 1)
 168                         last_tag = 0;
 169         }
 170
 171         return tag;
 172 }
 173
 174 #define BT_ALLOC_RR(tags) (tags->alloc_policy == BLK_TAG_ALLOC_RR)
 175
 176 /*
 177  * Straight forward bitmap tag implementation, where each bit is a tag
 178  * (cleared == free, and set == busy). The small twist is using per-cpu
 179  * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
 180  * contexts. This enables us to drastically limit the space searched,
 181  * without dirtying an extra shared cacheline like we would if we stored
 182  * the cache value inside the shared blk_mq_bitmap_tags structure. On top
 183  * of that, each word of tags is in a separate cacheline. This means that
 184  * multiple users will tend to stick to different cachelines, at least
 185  * until the map is exhausted.
 186  */
 187 static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
 188                     unsigned int *tag_cache, struct blk_mq_tags *tags)
 189 {
 190         unsigned int last_tag, org_last_tag;
 191         int index, i, tag;
 192
 193         if (!hctx_may_queue(hctx, bt))
 194                 return -1;
 195
 196         last_tag = org_last_tag = *tag_cache;
 197         index = TAG_TO_INDEX(bt, last_tag);
 198
 199         for (i = 0; i < bt->map_nr; i++) {
 200                 tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag),
 201                                     BT_ALLOC_RR(tags));
 202                 if (tag != -1) {
 203                         tag += (index << bt->bits_per_word);
 204                         goto done;
 205                 }
 206
 207                 /*
 208                  * Jump to next index, and reset the last tag to be the
 209                  * first tag of that index
 210                  */
 211                 index++;
 212                 last_tag = (index << bt->bits_per_word);
 213
 214                 if (index >= bt->map_nr) {
 215                         index = 0;
 216                         last_tag = 0;
 217                 }
 218         }
 219
 220         *tag_cache = 0;
 221         return -1;
 222
 223         /*
 224          * Only update the cache from the allocation path, if we ended
 225          * up using the specific cached tag.
 226          */
 227 done:
 228         if (tag == org_last_tag || unlikely(BT_ALLOC_RR(tags))) {
 229                 last_tag = tag + 1;
 230                 if (last_tag >= bt->depth - 1)
 231                         last_tag = 0;
 232
 233                 *tag_cache = last_tag;
 234         }
 235
 236         return tag;
 237 }
 238
 239 static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
 240                                          struct blk_mq_hw_ctx *hctx)
 241 {
 242         struct bt_wait_state *bs;
 243         int wait_index;
 244
 245         if (!hctx)
 246                 return &bt->bs[0];
 247
 248         wait_index = atomic_read(&hctx->wait_index);
 249         bs = &bt->bs[wait_index];
 250         bt_index_atomic_inc(&hctx->wait_index);
 251         return bs;
 252 }
 253
 254 static int bt_get(struct blk_mq_alloc_data *data,
 255                 struct blk_mq_bitmap_tags *bt,
 256                 struct blk_mq_hw_ctx *hctx,
 257                 unsigned int *last_tag, struct blk_mq_tags *tags)
 258 {
 259         struct bt_wait_state *bs;
 260         DEFINE_WAIT(wait);
 261         int tag;
 262
 263         tag = __bt_get(hctx, bt, last_tag, tags);
 264         if (tag != -1)
 265                 return tag;
 266
 267         if (!(data->gfp & __GFP_WAIT))
 268                 return -1;
 269
 270         bs = bt_wait_ptr(bt, hctx);
 271         do {
 272                 prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
 273
 274                 tag = __bt_get(hctx, bt, last_tag, tags);
 275                 if (tag != -1)
 276                         break;
 277
 278                 /*
 279                  * We're out of tags on this hardware queue, kick any
 280                  * pending IO submits before going to sleep waiting for
 281                  * some to complete.
 282                  */
 283                 blk_mq_run_hw_queue(hctx, false);
 284
 285                 /*
 286                  * Retry tag allocation after running the hardware queue,
 287                  * as running the queue may also have found completions.
 288                  */
 289                 tag = __bt_get(hctx, bt, last_tag, tags);
 290                 if (tag != -1)
 291                         break;
 292
 293                 blk_mq_put_ctx(data->ctx);
 294
 295                 io_schedule();
 296
 297                 data->ctx = blk_mq_get_ctx(data->q);
 298                 data->hctx = data->q->mq_ops->map_queue(data->q,
 299                                 data->ctx->cpu);
 300                 if (data->reserved) {
 301                         bt = &data->hctx->tags->breserved_tags;
 302                 } else {
 303                         last_tag = &data->ctx->last_tag;
 304                         hctx = data->hctx;
 305                         bt = &hctx->tags->bitmap_tags;
 306                 }
 307                 finish_wait(&bs->wait, &wait);
 308                 bs = bt_wait_ptr(bt, hctx);
 309         } while (1);
 310
 311         finish_wait(&bs->wait, &wait);
 312         return tag;
 313 }
 314
 315 static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
 316 {
 317         int tag;
 318
 319         tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
 320                         &data->ctx->last_tag, data->hctx->tags);
 321         if (tag >= 0)
 322                 return tag + data->hctx->tags->nr_reserved_tags;
 323
 324         return BLK_MQ_TAG_FAIL;
 325 }
 326
 327 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
 328 {
 329         int tag, zero = 0;
 330
 331         if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
 332                 WARN_ON_ONCE(1);
 333                 return BLK_MQ_TAG_FAIL;
 334         }
 335
 336         tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero,
 337                 data->hctx->tags);
 338         if (tag < 0)
 339                 return BLK_MQ_TAG_FAIL;
 340
 341         return tag;
 342 }
 343
 344 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
 345 {
 346         if (!data->reserved)
 347                 return __blk_mq_get_tag(data);
 348
 349         return __blk_mq_get_reserved_tag(data);
 350 }
 351
 352 static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
 353 {
 354         int i, wake_index;
 355
 356         wake_index = atomic_read(&bt->wake_index);
 357         for (i = 0; i < BT_WAIT_QUEUES; i++) {
 358                 struct bt_wait_state *bs = &bt->bs[wake_index];
 359
 360                 if (waitqueue_active(&bs->wait)) {
 361                         int o = atomic_read(&bt->wake_index);
 362                         if (wake_index != o)
 363                                 atomic_cmpxchg(&bt->wake_index, o, wake_index);
 364
 365                         return bs;
 366                 }
 367
 368                 wake_index = bt_index_inc(wake_index);
 369         }
 370
 371         return NULL;
 372 }
 373
 374 static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
 375 {
 376         const int index = TAG_TO_INDEX(bt, tag);
 377         struct bt_wait_state *bs;
 378         int wait_cnt;
 379
 380         clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word);
 381
 382         /* Ensure that the wait list checks occur after clear_bit(). */
 383         smp_mb();
 384
 385         bs = bt_wake_ptr(bt);
 386         if (!bs)
 387                 return;
 388
 389         wait_cnt = atomic_dec_return(&bs->wait_cnt);
 390         if (unlikely(wait_cnt < 0))
 391                 wait_cnt = atomic_inc_return(&bs->wait_cnt);
 392         if (wait_cnt == 0) {
 393                 atomic_add(bt->wake_cnt, &bs->wait_cnt);
 394                 bt_index_atomic_inc(&bt->wake_index);
 395                 wake_up(&bs->wait);
 396         }
 397 }
 398
 399 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
 400                     unsigned int *last_tag)
 401 {
 402         struct blk_mq_tags *tags = hctx->tags;
 403
 404         if (tag >= tags->nr_reserved_tags) {
 405                 const int real_tag = tag - tags->nr_reserved_tags;
 406
 407                 BUG_ON(real_tag >= tags->nr_tags);
 408                 bt_clear_tag(&tags->bitmap_tags, real_tag);
 409                 if (likely(tags->alloc_policy == BLK_TAG_ALLOC_FIFO))
 410                         *last_tag = real_tag;
 411         } else {
 412                 BUG_ON(tag >= tags->nr_reserved_tags);
 413                 bt_clear_tag(&tags->breserved_tags, tag);
 414         }
 415 }
 416
 417 static void bt_for_each(struct blk_mq_hw_ctx *hctx,
 418                 struct blk_mq_bitmap_tags *bt, unsigned int off,
 419                 busy_iter_fn *fn, void *data, bool reserved)
 420 {
 421         struct request *rq;
 422         int bit, i;
 423
 424         for (i = 0; i < bt->map_nr; i++) {
 425                 struct blk_align_bitmap *bm = &bt->map[i];
 426
 427                 for (bit = find_first_bit(&bm->word, bm->depth);
 428                      bit < bm->depth;
 429                      bit = find_next_bit(&bm->word, bm->depth, bit + 1)) {
 430                         rq = blk_mq_tag_to_rq(hctx->tags, off + bit);
 431                         if (rq->q == hctx->queue)
 432                                 fn(hctx, rq, data, reserved);
 433                 }
 434
 435                 off += (1 << bt->bits_per_word);
 436         }
 437 }
 438
 439 void blk_mq_tag_busy_iter(struct blk_mq_hw_ctx *hctx, busy_iter_fn *fn,
 440                 void *priv)
 441 {
 442         struct blk_mq_tags *tags = hctx->tags;
 443
 444         if (tags->nr_reserved_tags)
 445                 bt_for_each(hctx, &tags->breserved_tags, 0, fn, priv, true);
 446         bt_for_each(hctx, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv,
 447                         false);
 448 }
 449 EXPORT_SYMBOL(blk_mq_tag_busy_iter);
 450
 451 static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
 452 {
 453         unsigned int i, used;
 454
 455         for (i = 0, used = 0; i < bt->map_nr; i++) {
 456                 struct blk_align_bitmap *bm = &bt->map[i];
 457
 458                 used += bitmap_weight(&bm->word, bm->depth);
 459         }
 460
 461         return bt->depth - used;
 462 }
 463
 464 static void bt_update_count(struct blk_mq_bitmap_tags *bt,
 465                             unsigned int depth)
 466 {
 467         unsigned int tags_per_word = 1U << bt->bits_per_word;
 468         unsigned int map_depth = depth;
 469
 470         if (depth) {
 471                 int i;
 472
 473                 for (i = 0; i < bt->map_nr; i++) {
 474                         bt->map[i].depth = min(map_depth, tags_per_word);
 475                         map_depth -= bt->map[i].depth;
 476                 }
 477         }
 478
 479         bt->wake_cnt = BT_WAIT_BATCH;
 480         if (bt->wake_cnt > depth / BT_WAIT_QUEUES)
 481                 bt->wake_cnt = max(1U, depth / BT_WAIT_QUEUES);
 482
 483         bt->depth = depth;
 484 }
 485
 486 static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
 487                         int node, bool reserved)
 488 {
 489         int i;
 490
 491         bt->bits_per_word = ilog2(BITS_PER_LONG);
 492
 493         /*
 494          * Depth can be zero for reserved tags, that's not a failure
 495          * condition.
 496          */
 497         if (depth) {
 498                 unsigned int nr, tags_per_word;
 499
 500                 tags_per_word = (1 << bt->bits_per_word);
 501
 502                 /*
 503                  * If the tag space is small, shrink the number of tags
 504                  * per word so we spread over a few cachelines, at least.
 505                  * If less than 4 tags, just forget about it, it's not
 506                  * going to work optimally anyway.
 507                  */
 508                 if (depth >= 4) {
 509                         while (tags_per_word * 4 > depth) {
 510                                 bt->bits_per_word--;
 511                                 tags_per_word = (1 << bt->bits_per_word);
 512                         }
 513                 }
 514
 515                 nr = ALIGN(depth, tags_per_word) / tags_per_word;
 516                 bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
 517                                                 GFP_KERNEL, node);
 518                 if (!bt->map)
 519                         return -ENOMEM;
 520
 521                 bt->map_nr = nr;
 522         }
 523
 524         bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
 525         if (!bt->bs) {
 526                 kfree(bt->map);
 527                 bt->map = NULL;
 528                 return -ENOMEM;
 529         }
 530
 531         bt_update_count(bt, depth);
 532
 533         for (i = 0; i < BT_WAIT_QUEUES; i++) {
 534                 init_waitqueue_head(&bt->bs[i].wait);
 535                 atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt);
 536         }
 537
 538         return 0;
 539 }
 540
 541 static void bt_free(struct blk_mq_bitmap_tags *bt)
 542 {
 543         kfree(bt->map);
 544         kfree(bt->bs);
 545 }
 546
 547 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
 548                                                    int node, int alloc_policy)
 549 {
 550         unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
 551
 552         tags->alloc_policy = alloc_policy;
 553
 554         if (bt_alloc(&tags->bitmap_tags, depth, node, false))
 555                 goto enomem;
 556         if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
 557                 goto enomem;
 558
 559         return tags;
 560 enomem:
 561         bt_free(&tags->bitmap_tags);
 562         kfree(tags);
 563         return NULL;
 564 }
 565
 566 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
 567                                      unsigned int reserved_tags,
 568                                      int node, int alloc_policy)
 569 {
 570         struct blk_mq_tags *tags;
 571
 572         if (total_tags > BLK_MQ_TAG_MAX) {
 573                 pr_err("blk-mq: tag depth too large\n");
 574                 return NULL;
 575         }
 576
 577         tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
 578         if (!tags)
 579                 return NULL;
 580
 581         tags->nr_tags = total_tags;
 582         tags->nr_reserved_tags = reserved_tags;
 583
 584         return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
 585 }
 586
 587 void blk_mq_free_tags(struct blk_mq_tags *tags)
 588 {
 589         bt_free(&tags->bitmap_tags);
 590         bt_free(&tags->breserved_tags);
 591         kfree(tags);
 592 }
 593
 594 void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
 595 {
 596         unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
 597
 598         *tag = prandom_u32() % depth;
 599 }
 600
 601 int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
 602 {
 603         tdepth -= tags->nr_reserved_tags;
 604         if (tdepth > tags->nr_tags)
 605                 return -EINVAL;
 606
 607         /*
 608          * Don't need (or can't) update reserved tags here, they remain
 609          * static and should never need resizing.
 610          */
 611         bt_update_count(&tags->bitmap_tags, tdepth);
 612         blk_mq_tag_wakeup_all(tags, false);
 613         return 0;
 614 }
 615
 616 /**
 617  * blk_mq_unique_tag() - return a tag that is unique queue-wide
 618  * @rq: request for which to compute a unique tag
 619  *
 620  * The tag field in struct request is unique per hardware queue but not over
 621  * all hardware queues. Hence this function that returns a tag with the
 622  * hardware context index in the upper bits and the per hardware queue tag in
 623  * the lower bits.
 624  *
 625  * Note: When called for a request that is queued on a non-multiqueue request
 626  * queue, the hardware context index is set to zero.
 627  */
 628 u32 blk_mq_unique_tag(struct request *rq)
 629 {
 630         struct request_queue *q = rq->q;
 631         struct blk_mq_hw_ctx *hctx;
 632         int hwq = 0;
 633
 634         if (q->mq_ops) {
 635                 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
 636                 hwq = hctx->queue_num;
 637         }
 638
 639         return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
 640                 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
 641 }
 642 EXPORT_SYMBOL(blk_mq_unique_tag);
 643
 644 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
 645 {
 646         char *orig_page = page;
 647         unsigned int free, res;
 648
 649         if (!tags)
 650                 return 0;
 651
 652         page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
 653                         "bits_per_word=%u\n",
 654                         tags->nr_tags, tags->nr_reserved_tags,
 655                         tags->bitmap_tags.bits_per_word);
 656
 657         free = bt_unused_tags(&tags->bitmap_tags);
 658         res = bt_unused_tags(&tags->breserved_tags);
 659
 660         page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
 661         page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
 662
 663         return page - orig_page;
 664 }