block/blk-mq-tag.c

   1 /*
   2  * Tag allocation using scalable bitmaps. Uses active queue tracking to support
   3  * fairer distribution of tags between multiple submitters when a shared tag map
   4  * is used.
   5  *
   6  * Copyright (C) 2013-2014 Jens Axboe
   7  */
   8 #include <linux/kernel.h>
   9 #include <linux/module.h>
  10
  11 #include <linux/blk-mq.h>
  12 #include "blk.h"
  13 #include "blk-mq.h"
  14 #include "blk-mq-tag.h"
  15
  16 bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
  17 {
  18         if (!tags)
  19                 return true;
  20
  21         return sbitmap_any_bit_clear(&tags->bitmap_tags.sb);
  22 }
  23
  24 /*
  25  * If a previously inactive queue goes active, bump the active user count.
  26  * We need to do this before try to allocate driver tag, then even if fail
  27  * to get tag when first time, the other shared-tag users could reserve
  28  * budget for it.
  29  */
  30 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
  31 {
  32         if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
  33             !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
  34                 atomic_inc(&hctx->tags->active_queues);
  35
  36         return true;
  37 }
  38
  39 /*
  40  * Wakeup all potentially sleeping on tags
  41  */
  42 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
  43 {
  44         sbitmap_queue_wake_all(&tags->bitmap_tags);
  45         if (include_reserve)
  46                 sbitmap_queue_wake_all(&tags->breserved_tags);
  47 }
  48
  49 /*
  50  * If a previously busy queue goes inactive, potential waiters could now
  51  * be allowed to queue. Wake them up and check.
  52  */
  53 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
  54 {
  55         struct blk_mq_tags *tags = hctx->tags;
  56
  57         if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
  58                 return;
  59
  60         atomic_dec(&tags->active_queues);
  61
  62         blk_mq_tag_wakeup_all(tags, false);
  63 }
  64
  65 /*
  66  * For shared tag users, we track the number of currently active users
  67  * and attempt to provide a fair share of the tag depth for each of them.
  68  */
  69 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
  70                                   struct sbitmap_queue *bt)
  71 {
  72         unsigned int depth, users;
  73
  74         if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
  75                 return true;
  76         if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
  77                 return true;
  78
  79         /*
  80          * Don't try dividing an ant
  81          */
  82         if (bt->sb.depth == 1)
  83                 return true;
  84
  85         users = atomic_read(&hctx->tags->active_queues);
  86         if (!users)
  87                 return true;
  88
  89         /*
  90          * Allow at least some tags
  91          */
  92         depth = max((bt->sb.depth + users - 1) / users, 4U);
  93         return atomic_read(&hctx->nr_active) < depth;
  94 }
  95
  96 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
  97                             struct sbitmap_queue *bt)
  98 {
  99         if (!(data->flags & BLK_MQ_REQ_INTERNAL) &&
 100             !hctx_may_queue(data->hctx, bt))
 101                 return -1;
 102         if (data->shallow_depth)
 103                 return __sbitmap_queue_get_shallow(bt, data->shallow_depth);
 104         else
 105                 return __sbitmap_queue_get(bt);
 106 }
 107
 108 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
 109 {
 110         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 111         struct sbitmap_queue *bt;
 112         struct sbq_wait_state *ws;
 113         DEFINE_WAIT(wait);
 114         unsigned int tag_offset;
 115         bool drop_ctx;
 116         int tag;
 117
 118         if (data->flags & BLK_MQ_REQ_RESERVED) {
 119                 if (unlikely(!tags->nr_reserved_tags)) {
 120                         WARN_ON_ONCE(1);
 121                         return BLK_MQ_TAG_FAIL;
 122                 }
 123                 bt = &tags->breserved_tags;
 124                 tag_offset = 0;
 125         } else {
 126                 bt = &tags->bitmap_tags;
 127                 tag_offset = tags->nr_reserved_tags;
 128         }
 129
 130         tag = __blk_mq_get_tag(data, bt);
 131         if (tag != -1)
 132                 goto found_tag;
 133
 134         if (data->flags & BLK_MQ_REQ_NOWAIT)
 135                 return BLK_MQ_TAG_FAIL;
 136
 137         ws = bt_wait_ptr(bt, data->hctx);
 138         drop_ctx = data->ctx == NULL;
 139         do {
 140                 struct sbitmap_queue *bt_prev;
 141
 142                 /*
 143                  * We're out of tags on this hardware queue, kick any
 144                  * pending IO submits before going to sleep waiting for
 145                  * some to complete.
 146                  */
 147                 blk_mq_run_hw_queue(data->hctx, false);
 148
 149                 /*
 150                  * Retry tag allocation after running the hardware queue,
 151                  * as running the queue may also have found completions.
 152                  */
 153                 tag = __blk_mq_get_tag(data, bt);
 154                 if (tag != -1)
 155                         break;
 156
 157                 prepare_to_wait_exclusive(&ws->wait, &wait,
 158                                                 TASK_UNINTERRUPTIBLE);
 159
 160                 tag = __blk_mq_get_tag(data, bt);
 161                 if (tag != -1)
 162                         break;
 163
 164                 if (data->ctx)
 165                         blk_mq_put_ctx(data->ctx);
 166
 167                 bt_prev = bt;
 168                 io_schedule();
 169
 170                 data->ctx = blk_mq_get_ctx(data->q);
 171                 data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu);
 172                 tags = blk_mq_tags_from_data(data);
 173                 if (data->flags & BLK_MQ_REQ_RESERVED)
 174                         bt = &tags->breserved_tags;
 175                 else
 176                         bt = &tags->bitmap_tags;
 177
 178                 finish_wait(&ws->wait, &wait);
 179
 180                 /*
 181                  * If destination hw queue is changed, fake wake up on
 182                  * previous queue for compensating the wake up miss, so
 183                  * other allocations on previous queue won't be starved.
 184                  */
 185                 if (bt != bt_prev)
 186                         sbitmap_queue_wake_up(bt_prev);
 187
 188                 ws = bt_wait_ptr(bt, data->hctx);
 189         } while (1);
 190
 191         if (drop_ctx && data->ctx)
 192                 blk_mq_put_ctx(data->ctx);
 193
 194         finish_wait(&ws->wait, &wait);
 195
 196 found_tag:
 197         return tag + tag_offset;
 198 }
 199
 200 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags,
 201                     struct blk_mq_ctx *ctx, unsigned int tag)
 202 {
 203         if (!blk_mq_tag_is_reserved(tags, tag)) {
 204                 const int real_tag = tag - tags->nr_reserved_tags;
 205
 206                 BUG_ON(real_tag >= tags->nr_tags);
 207                 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
 208         } else {
 209                 BUG_ON(tag >= tags->nr_reserved_tags);
 210                 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
 211         }
 212 }
 213
 214 struct bt_iter_data {
 215         struct blk_mq_hw_ctx *hctx;
 216         busy_iter_fn *fn;
 217         void *data;
 218         bool reserved;
 219 };
 220
 221 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
 222 {
 223         struct bt_iter_data *iter_data = data;
 224         struct blk_mq_hw_ctx *hctx = iter_data->hctx;
 225         struct blk_mq_tags *tags = hctx->tags;
 226         bool reserved = iter_data->reserved;
 227         struct request *rq;
 228
 229         if (!reserved)
 230                 bitnr += tags->nr_reserved_tags;
 231         rq = tags->rqs[bitnr];
 232
 233         /*
 234          * We can hit rq == NULL here, because the tagging functions
 235          * test and set the bit before assining ->rqs[].
 236          */
 237         if (rq && rq->q == hctx->queue)
 238                 iter_data->fn(hctx, rq, iter_data->data, reserved);
 239         return true;
 240 }
 241
 242 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
 243                         busy_iter_fn *fn, void *data, bool reserved)
 244 {
 245         struct bt_iter_data iter_data = {
 246                 .hctx = hctx,
 247                 .fn = fn,
 248                 .data = data,
 249                 .reserved = reserved,
 250         };
 251
 252         sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
 253 }
 254
 255 struct bt_tags_iter_data {
 256         struct blk_mq_tags *tags;
 257         busy_tag_iter_fn *fn;
 258         void *data;
 259         bool reserved;
 260 };
 261
 262 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
 263 {
 264         struct bt_tags_iter_data *iter_data = data;
 265         struct blk_mq_tags *tags = iter_data->tags;
 266         bool reserved = iter_data->reserved;
 267         struct request *rq;
 268
 269         if (!reserved)
 270                 bitnr += tags->nr_reserved_tags;
 271
 272         /*
 273          * We can hit rq == NULL here, because the tagging functions
 274          * test and set the bit before assining ->rqs[].
 275          */
 276         rq = tags->rqs[bitnr];
 277         if (rq && blk_mq_request_started(rq))
 278                 iter_data->fn(rq, iter_data->data, reserved);
 279
 280         return true;
 281 }
 282
 283 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
 284                              busy_tag_iter_fn *fn, void *data, bool reserved)
 285 {
 286         struct bt_tags_iter_data iter_data = {
 287                 .tags = tags,
 288                 .fn = fn,
 289                 .data = data,
 290                 .reserved = reserved,
 291         };
 292
 293         if (tags->rqs)
 294                 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
 295 }
 296
 297 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
 298                 busy_tag_iter_fn *fn, void *priv)
 299 {
 300         if (tags->nr_reserved_tags)
 301                 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
 302         bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
 303 }
 304
 305 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
 306                 busy_tag_iter_fn *fn, void *priv)
 307 {
 308         int i;
 309
 310         for (i = 0; i < tagset->nr_hw_queues; i++) {
 311                 if (tagset->tags && tagset->tags[i])
 312                         blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
 313         }
 314 }
 315 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
 316
 317 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
 318                 void *priv)
 319 {
 320         struct blk_mq_hw_ctx *hctx;
 321         int i;
 322
 323         /*
 324          * __blk_mq_update_nr_hw_queues will update the nr_hw_queues and
 325          * queue_hw_ctx after freeze the queue, so we use q_usage_counter
 326          * to avoid race with it.
 327          */
 328         if (!percpu_ref_tryget(&q->q_usage_counter))
 329                 return;
 330
 331         queue_for_each_hw_ctx(q, hctx, i) {
 332                 struct blk_mq_tags *tags = hctx->tags;
 333
 334                 /*
 335                  * If not software queues are currently mapped to this
 336                  * hardware queue, there's nothing to check
 337                  */
 338                 if (!blk_mq_hw_queue_mapped(hctx))
 339                         continue;
 340
 341                 if (tags->nr_reserved_tags)
 342                         bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
 343                 bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
 344         }
 345         blk_queue_exit(q);
 346 }
 347
 348 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
 349                     bool round_robin, int node)
 350 {
 351         return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
 352                                        node);
 353 }
 354
 355 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
 356                                                    int node, int alloc_policy)
 357 {
 358         unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
 359         bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
 360
 361         if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
 362                 goto free_tags;
 363         if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
 364                      node))
 365                 goto free_bitmap_tags;
 366
 367         return tags;
 368 free_bitmap_tags:
 369         sbitmap_queue_free(&tags->bitmap_tags);
 370 free_tags:
 371         kfree(tags);
 372         return NULL;
 373 }
 374
 375 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
 376                                      unsigned int reserved_tags,
 377                                      int node, int alloc_policy)
 378 {
 379         struct blk_mq_tags *tags;
 380
 381         if (total_tags > BLK_MQ_TAG_MAX) {
 382                 pr_err("blk-mq: tag depth too large\n");
 383                 return NULL;
 384         }
 385
 386         tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
 387         if (!tags)
 388                 return NULL;
 389
 390         tags->nr_tags = total_tags;
 391         tags->nr_reserved_tags = reserved_tags;
 392
 393         return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
 394 }
 395
 396 void blk_mq_free_tags(struct blk_mq_tags *tags)
 397 {
 398         sbitmap_queue_free(&tags->bitmap_tags);
 399         sbitmap_queue_free(&tags->breserved_tags);
 400         kfree(tags);
 401 }
 402
 403 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
 404                             struct blk_mq_tags **tagsptr, unsigned int tdepth,
 405                             bool can_grow)
 406 {
 407         struct blk_mq_tags *tags = *tagsptr;
 408
 409         if (tdepth <= tags->nr_reserved_tags)
 410                 return -EINVAL;
 411
 412         /*
 413          * If we are allowed to grow beyond the original size, allocate
 414          * a new set of tags before freeing the old one.
 415          */
 416         if (tdepth > tags->nr_tags) {
 417                 struct blk_mq_tag_set *set = hctx->queue->tag_set;
 418                 struct blk_mq_tags *new;
 419                 bool ret;
 420
 421                 if (!can_grow)
 422                         return -EINVAL;
 423
 424                 /*
 425                  * We need some sort of upper limit, set it high enough that
 426                  * no valid use cases should require more.
 427                  */
 428                 if (tdepth > 16 * BLKDEV_MAX_RQ)
 429                         return -EINVAL;
 430
 431                 new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth,
 432                                 tags->nr_reserved_tags);
 433                 if (!new)
 434                         return -ENOMEM;
 435                 ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
 436                 if (ret) {
 437                         blk_mq_free_rq_map(new);
 438                         return -ENOMEM;
 439                 }
 440
 441                 blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
 442                 blk_mq_free_rq_map(*tagsptr);
 443                 *tagsptr = new;
 444         } else {
 445                 /*
 446                  * Don't need (or can't) update reserved tags here, they
 447                  * remain static and should never need resizing.
 448                  */
 449                 sbitmap_queue_resize(&tags->bitmap_tags,
 450                                 tdepth - tags->nr_reserved_tags);
 451         }
 452
 453         return 0;
 454 }
 455
 456 /**
 457  * blk_mq_unique_tag() - return a tag that is unique queue-wide
 458  * @rq: request for which to compute a unique tag
 459  *
 460  * The tag field in struct request is unique per hardware queue but not over
 461  * all hardware queues. Hence this function that returns a tag with the
 462  * hardware context index in the upper bits and the per hardware queue tag in
 463  * the lower bits.
 464  *
 465  * Note: When called for a request that is queued on a non-multiqueue request
 466  * queue, the hardware context index is set to zero.
 467  */
 468 u32 blk_mq_unique_tag(struct request *rq)
 469 {
 470         struct request_queue *q = rq->q;
 471         struct blk_mq_hw_ctx *hctx;
 472         int hwq = 0;
 473
 474         if (q->mq_ops) {
 475                 hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
 476                 hwq = hctx->queue_num;
 477         }
 478
 479         return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
 480                 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
 481 }
 482 EXPORT_SYMBOL(blk_mq_unique_tag);