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