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.
 140                  */
 141                 blk_mq_run_hw_queue(data->hctx, false);
 142
 143                 /*
 144                  * Retry tag allocation after running the hardware queue,
 145                  * as running the queue may also have found completions.
 146                  */
 147                 tag = __blk_mq_get_tag(data->hctx, bt);
 148                 if (tag != -1)
 149                         break;
 150
 151                 blk_mq_put_ctx(data->ctx);
 152
 153                 io_schedule();
 154
 155                 data->ctx = blk_mq_get_ctx(data->q);
 156                 data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu);
 157                 tags = blk_mq_tags_from_data(data);
 158                 if (data->flags & BLK_MQ_REQ_RESERVED)
 159                         bt = &tags->breserved_tags;
 160                 else
 161                         bt = &tags->bitmap_tags;
 162
 163                 finish_wait(&ws->wait, &wait);
 164                 ws = bt_wait_ptr(bt, data->hctx);
 165         } while (1);
 166
 167         finish_wait(&ws->wait, &wait);
 168
 169 found_tag:
 170         return tag + tag_offset;
 171 }
 172
 173 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags,
 174                     struct blk_mq_ctx *ctx, unsigned int tag)
 175 {
 176         if (tag >= tags->nr_reserved_tags) {
 177                 const int real_tag = tag - tags->nr_reserved_tags;
 178
 179                 BUG_ON(real_tag >= tags->nr_tags);
 180                 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
 181         } else {
 182                 BUG_ON(tag >= tags->nr_reserved_tags);
 183                 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
 184         }
 185 }
 186
 187 struct bt_iter_data {
 188         struct blk_mq_hw_ctx *hctx;
 189         busy_iter_fn *fn;
 190         void *data;
 191         bool reserved;
 192 };
 193
 194 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
 195 {
 196         struct bt_iter_data *iter_data = data;
 197         struct blk_mq_hw_ctx *hctx = iter_data->hctx;
 198         struct blk_mq_tags *tags = hctx->tags;
 199         bool reserved = iter_data->reserved;
 200         struct request *rq;
 201
 202         if (!reserved)
 203                 bitnr += tags->nr_reserved_tags;
 204         rq = tags->rqs[bitnr];
 205
 206         if (rq->q == hctx->queue)
 207                 iter_data->fn(hctx, rq, iter_data->data, reserved);
 208         return true;
 209 }
 210
 211 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
 212                         busy_iter_fn *fn, void *data, bool reserved)
 213 {
 214         struct bt_iter_data iter_data = {
 215                 .hctx = hctx,
 216                 .fn = fn,
 217                 .data = data,
 218                 .reserved = reserved,
 219         };
 220
 221         sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
 222 }
 223
 224 struct bt_tags_iter_data {
 225         struct blk_mq_tags *tags;
 226         busy_tag_iter_fn *fn;
 227         void *data;
 228         bool reserved;
 229 };
 230
 231 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
 232 {
 233         struct bt_tags_iter_data *iter_data = data;
 234         struct blk_mq_tags *tags = iter_data->tags;
 235         bool reserved = iter_data->reserved;
 236         struct request *rq;
 237
 238         if (!reserved)
 239                 bitnr += tags->nr_reserved_tags;
 240         rq = tags->rqs[bitnr];
 241
 242         iter_data->fn(rq, iter_data->data, reserved);
 243         return true;
 244 }
 245
 246 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
 247                              busy_tag_iter_fn *fn, void *data, bool reserved)
 248 {
 249         struct bt_tags_iter_data iter_data = {
 250                 .tags = tags,
 251                 .fn = fn,
 252                 .data = data,
 253                 .reserved = reserved,
 254         };
 255
 256         if (tags->rqs)
 257                 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
 258 }
 259
 260 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
 261                 busy_tag_iter_fn *fn, void *priv)
 262 {
 263         if (tags->nr_reserved_tags)
 264                 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
 265         bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
 266 }
 267
 268 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
 269                 busy_tag_iter_fn *fn, void *priv)
 270 {
 271         int i;
 272
 273         for (i = 0; i < tagset->nr_hw_queues; i++) {
 274                 if (tagset->tags && tagset->tags[i])
 275                         blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
 276         }
 277 }
 278 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
 279
 280 int blk_mq_reinit_tagset(struct blk_mq_tag_set *set)
 281 {
 282         int i, j, ret = 0;
 283
 284         if (!set->ops->reinit_request)
 285                 goto out;
 286
 287         for (i = 0; i < set->nr_hw_queues; i++) {
 288                 struct blk_mq_tags *tags = set->tags[i];
 289
 290                 for (j = 0; j < tags->nr_tags; j++) {
 291                         if (!tags->static_rqs[j])
 292                                 continue;
 293
 294                         ret = set->ops->reinit_request(set->driver_data,
 295                                                 tags->static_rqs[j]);
 296                         if (ret)
 297                                 goto out;
 298                 }
 299         }
 300
 301 out:
 302         return ret;
 303 }
 304 EXPORT_SYMBOL_GPL(blk_mq_reinit_tagset);
 305
 306 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
 307                 void *priv)
 308 {
 309         struct blk_mq_hw_ctx *hctx;
 310         int i;
 311
 312
 313         queue_for_each_hw_ctx(q, hctx, i) {
 314                 struct blk_mq_tags *tags = hctx->tags;
 315
 316                 /*
 317                  * If not software queues are currently mapped to this
 318                  * hardware queue, there's nothing to check
 319                  */
 320                 if (!blk_mq_hw_queue_mapped(hctx))
 321                         continue;
 322
 323                 if (tags->nr_reserved_tags)
 324                         bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
 325                 bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
 326         }
 327
 328 }
 329
 330 static unsigned int bt_unused_tags(const struct sbitmap_queue *bt)
 331 {
 332         return bt->sb.depth - sbitmap_weight(&bt->sb);
 333 }
 334
 335 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
 336                     bool round_robin, int node)
 337 {
 338         return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
 339                                        node);
 340 }
 341
 342 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
 343                                                    int node, int alloc_policy)
 344 {
 345         unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
 346         bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
 347
 348         if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
 349                 goto free_tags;
 350         if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
 351                      node))
 352                 goto free_bitmap_tags;
 353
 354         return tags;
 355 free_bitmap_tags:
 356         sbitmap_queue_free(&tags->bitmap_tags);
 357 free_tags:
 358         kfree(tags);
 359         return NULL;
 360 }
 361
 362 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
 363                                      unsigned int reserved_tags,
 364                                      int node, int alloc_policy)
 365 {
 366         struct blk_mq_tags *tags;
 367
 368         if (total_tags > BLK_MQ_TAG_MAX) {
 369                 pr_err("blk-mq: tag depth too large\n");
 370                 return NULL;
 371         }
 372
 373         tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
 374         if (!tags)
 375                 return NULL;
 376
 377         tags->nr_tags = total_tags;
 378         tags->nr_reserved_tags = reserved_tags;
 379
 380         return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
 381 }
 382
 383 void blk_mq_free_tags(struct blk_mq_tags *tags)
 384 {
 385         sbitmap_queue_free(&tags->bitmap_tags);
 386         sbitmap_queue_free(&tags->breserved_tags);
 387         kfree(tags);
 388 }
 389
 390 int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
 391 {
 392         tdepth -= tags->nr_reserved_tags;
 393         if (tdepth > tags->nr_tags)
 394                 return -EINVAL;
 395
 396         /*
 397          * Don't need (or can't) update reserved tags here, they remain
 398          * static and should never need resizing.
 399          */
 400         sbitmap_queue_resize(&tags->bitmap_tags, tdepth);
 401
 402         blk_mq_tag_wakeup_all(tags, false);
 403         return 0;
 404 }
 405
 406 /**
 407  * blk_mq_unique_tag() - return a tag that is unique queue-wide
 408  * @rq: request for which to compute a unique tag
 409  *
 410  * The tag field in struct request is unique per hardware queue but not over
 411  * all hardware queues. Hence this function that returns a tag with the
 412  * hardware context index in the upper bits and the per hardware queue tag in
 413  * the lower bits.
 414  *
 415  * Note: When called for a request that is queued on a non-multiqueue request
 416  * queue, the hardware context index is set to zero.
 417  */
 418 u32 blk_mq_unique_tag(struct request *rq)
 419 {
 420         struct request_queue *q = rq->q;
 421         struct blk_mq_hw_ctx *hctx;
 422         int hwq = 0;
 423
 424         if (q->mq_ops) {
 425                 hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
 426                 hwq = hctx->queue_num;
 427         }
 428
 429         return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
 430                 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
 431 }
 432 EXPORT_SYMBOL(blk_mq_unique_tag);
 433
 434 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
 435 {
 436         char *orig_page = page;
 437         unsigned int free, res;
 438
 439         if (!tags)
 440                 return 0;
 441
 442         page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
 443                         "bits_per_word=%u\n",
 444                         tags->nr_tags, tags->nr_reserved_tags,
 445                         1U << tags->bitmap_tags.sb.shift);
 446
 447         free = bt_unused_tags(&tags->bitmap_tags);
 448         res = bt_unused_tags(&tags->breserved_tags);
 449
 450         page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
 451         page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
 452
 453         return page - orig_page;
 454 }