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