block/blk-mq.c

   1 /*
   2  * Block multiqueue core code
   3  *
   4  * Copyright (C) 2013-2014 Jens Axboe
   5  * Copyright (C) 2013-2014 Christoph Hellwig
   6  */
   7 #include <linux/kernel.h>
   8 #include <linux/module.h>
   9 #include <linux/backing-dev.h>
  10 #include <linux/bio.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/kmemleak.h>
  13 #include <linux/mm.h>
  14 #include <linux/init.h>
  15 #include <linux/slab.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/smp.h>
  18 #include <linux/llist.h>
  19 #include <linux/list_sort.h>
  20 #include <linux/cpu.h>
  21 #include <linux/cache.h>
  22 #include <linux/sched/sysctl.h>
  23 #include <linux/sched/topology.h>
  24 #include <linux/sched/signal.h>
  25 #include <linux/delay.h>
  26 #include <linux/crash_dump.h>
  27 #include <linux/prefetch.h>
  28
  29 #include <trace/events/block.h>
  30
  31 #include <linux/blk-mq.h>
  32 #include "blk.h"
  33 #include "blk-mq.h"
  34 #include "blk-mq-debugfs.h"
  35 #include "blk-mq-tag.h"
  36 #include "blk-stat.h"
  37 #include "blk-wbt.h"
  38 #include "blk-mq-sched.h"
  39
  40 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie);
  41 static void blk_mq_poll_stats_start(struct request_queue *q);
  42 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
  43
  44 static int blk_mq_poll_stats_bkt(const struct request *rq)
  45 {
  46         int ddir, bytes, bucket;
  47
  48         ddir = rq_data_dir(rq);
  49         bytes = blk_rq_bytes(rq);
  50
  51         bucket = ddir + 2*(ilog2(bytes) - 9);
  52
  53         if (bucket < 0)
  54                 return -1;
  55         else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
  56                 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
  57
  58         return bucket;
  59 }
  60
  61 /*
  62  * Check if any of the ctx's have pending work in this hardware queue
  63  */
  64 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  65 {
  66         return !list_empty_careful(&hctx->dispatch) ||
  67                 sbitmap_any_bit_set(&hctx->ctx_map) ||
  68                         blk_mq_sched_has_work(hctx);
  69 }
  70
  71 /*
  72  * Mark this ctx as having pending work in this hardware queue
  73  */
  74 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  75                                      struct blk_mq_ctx *ctx)
  76 {
  77         if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
  78                 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
  79 }
  80
  81 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  82                                       struct blk_mq_ctx *ctx)
  83 {
  84         sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
  85 }
  86
  87 struct mq_inflight {
  88         struct hd_struct *part;
  89         unsigned int *inflight;
  90 };
  91
  92 static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
  93                                   struct request *rq, void *priv,
  94                                   bool reserved)
  95 {
  96         struct mq_inflight *mi = priv;
  97
  98         /*
  99          * index[0] counts the specific partition that was asked for. index[1]
 100          * counts the ones that are active on the whole device, so increment
 101          * that if mi->part is indeed a partition, and not a whole device.
 102          */
 103         if (rq->part == mi->part)
 104                 mi->inflight[0]++;
 105         if (mi->part->partno)
 106                 mi->inflight[1]++;
 107 }
 108
 109 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
 110                       unsigned int inflight[2])
 111 {
 112         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 113
 114         inflight[0] = inflight[1] = 0;
 115         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
 116 }
 117
 118 static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
 119                                      struct request *rq, void *priv,
 120                                      bool reserved)
 121 {
 122         struct mq_inflight *mi = priv;
 123
 124         if (rq->part == mi->part)
 125                 mi->inflight[rq_data_dir(rq)]++;
 126 }
 127
 128 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
 129                          unsigned int inflight[2])
 130 {
 131         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 132
 133         inflight[0] = inflight[1] = 0;
 134         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
 135 }
 136
 137 void blk_freeze_queue_start(struct request_queue *q)
 138 {
 139         int freeze_depth;
 140
 141         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
 142         if (freeze_depth == 1) {
 143                 percpu_ref_kill(&q->q_usage_counter);
 144                 if (q->mq_ops)
 145                         blk_mq_run_hw_queues(q, false);
 146         }
 147 }
 148 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
 149
 150 void blk_mq_freeze_queue_wait(struct request_queue *q)
 151 {
 152         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
 153 }
 154 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
 155
 156 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 157                                      unsigned long timeout)
 158 {
 159         return wait_event_timeout(q->mq_freeze_wq,
 160                                         percpu_ref_is_zero(&q->q_usage_counter),
 161                                         timeout);
 162 }
 163 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 164
 165 /*
 166  * Guarantee no request is in use, so we can change any data structure of
 167  * the queue afterward.
 168  */
 169 void blk_freeze_queue(struct request_queue *q)
 170 {
 171         /*
 172          * In the !blk_mq case we are only calling this to kill the
 173          * q_usage_counter, otherwise this increases the freeze depth
 174          * and waits for it to return to zero.  For this reason there is
 175          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 176          * exported to drivers as the only user for unfreeze is blk_mq.
 177          */
 178         blk_freeze_queue_start(q);
 179         if (!q->mq_ops)
 180                 blk_drain_queue(q);
 181         blk_mq_freeze_queue_wait(q);
 182 }
 183
 184 void blk_mq_freeze_queue(struct request_queue *q)
 185 {
 186         /*
 187          * ...just an alias to keep freeze and unfreeze actions balanced
 188          * in the blk_mq_* namespace
 189          */
 190         blk_freeze_queue(q);
 191 }
 192 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 193
 194 void blk_mq_unfreeze_queue(struct request_queue *q)
 195 {
 196         int freeze_depth;
 197
 198         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 199         WARN_ON_ONCE(freeze_depth < 0);
 200         if (!freeze_depth) {
 201                 percpu_ref_reinit(&q->q_usage_counter);
 202                 wake_up_all(&q->mq_freeze_wq);
 203         }
 204 }
 205 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 206
 207 /*
 208  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
 209  * mpt3sas driver such that this function can be removed.
 210  */
 211 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 212 {
 213         blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
 214 }
 215 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 216
 217 /**
 218  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
 219  * @q: request queue.
 220  *
 221  * Note: this function does not prevent that the struct request end_io()
 222  * callback function is invoked. Once this function is returned, we make
 223  * sure no dispatch can happen until the queue is unquiesced via
 224  * blk_mq_unquiesce_queue().
 225  */
 226 void blk_mq_quiesce_queue(struct request_queue *q)
 227 {
 228         struct blk_mq_hw_ctx *hctx;
 229         unsigned int i;
 230         bool rcu = false;
 231
 232         blk_mq_quiesce_queue_nowait(q);
 233
 234         queue_for_each_hw_ctx(q, hctx, i) {
 235                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 236                         synchronize_srcu(hctx->srcu);
 237                 else
 238                         rcu = true;
 239         }
 240         if (rcu)
 241                 synchronize_rcu();
 242 }
 243 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 244
 245 /*
 246  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
 247  * @q: request queue.
 248  *
 249  * This function recovers queue into the state before quiescing
 250  * which is done by blk_mq_quiesce_queue.
 251  */
 252 void blk_mq_unquiesce_queue(struct request_queue *q)
 253 {
 254         blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 255
 256         /* dispatch requests which are inserted during quiescing */
 257         blk_mq_run_hw_queues(q, true);
 258 }
 259 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
 260
 261 void blk_mq_wake_waiters(struct request_queue *q)
 262 {
 263         struct blk_mq_hw_ctx *hctx;
 264         unsigned int i;
 265
 266         queue_for_each_hw_ctx(q, hctx, i)
 267                 if (blk_mq_hw_queue_mapped(hctx))
 268                         blk_mq_tag_wakeup_all(hctx->tags, true);
 269 }
 270
 271 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 272 {
 273         return blk_mq_has_free_tags(hctx->tags);
 274 }
 275 EXPORT_SYMBOL(blk_mq_can_queue);
 276
 277 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 278                 unsigned int tag, unsigned int op)
 279 {
 280         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 281         struct request *rq = tags->static_rqs[tag];
 282         req_flags_t rq_flags = 0;
 283
 284         if (data->flags & BLK_MQ_REQ_INTERNAL) {
 285                 rq->tag = -1;
 286                 rq->internal_tag = tag;
 287         } else {
 288                 if (blk_mq_tag_busy(data->hctx)) {
 289                         rq_flags = RQF_MQ_INFLIGHT;
 290                         atomic_inc(&data->hctx->nr_active);
 291                 }
 292                 rq->tag = tag;
 293                 rq->internal_tag = -1;
 294                 data->hctx->tags->rqs[rq->tag] = rq;
 295         }
 296
 297         /* csd/requeue_work/fifo_time is initialized before use */
 298         rq->q = data->q;
 299         rq->mq_ctx = data->ctx;
 300         rq->rq_flags = rq_flags;
 301         rq->cpu = -1;
 302         rq->cmd_flags = op;
 303         if (data->flags & BLK_MQ_REQ_PREEMPT)
 304                 rq->rq_flags |= RQF_PREEMPT;
 305         if (blk_queue_io_stat(data->q))
 306                 rq->rq_flags |= RQF_IO_STAT;
 307         INIT_LIST_HEAD(&rq->queuelist);
 308         INIT_HLIST_NODE(&rq->hash);
 309         RB_CLEAR_NODE(&rq->rb_node);
 310         rq->rq_disk = NULL;
 311         rq->part = NULL;
 312         rq->start_time_ns = ktime_get_ns();
 313         rq->io_start_time_ns = 0;
 314         rq->nr_phys_segments = 0;
 315 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 316         rq->nr_integrity_segments = 0;
 317 #endif
 318         rq->special = NULL;
 319         /* tag was already set */
 320         rq->extra_len = 0;
 321         rq->__deadline = 0;
 322
 323         INIT_LIST_HEAD(&rq->timeout_list);
 324         rq->timeout = 0;
 325
 326         rq->end_io = NULL;
 327         rq->end_io_data = NULL;
 328         rq->next_rq = NULL;
 329
 330 #ifdef CONFIG_BLK_CGROUP
 331         rq->rl = NULL;
 332 #endif
 333
 334         data->ctx->rq_dispatched[op_is_sync(op)]++;
 335         refcount_set(&rq->ref, 1);
 336         return rq;
 337 }
 338
 339 static struct request *blk_mq_get_request(struct request_queue *q,
 340                 struct bio *bio, unsigned int op,
 341                 struct blk_mq_alloc_data *data)
 342 {
 343         struct elevator_queue *e = q->elevator;
 344         struct request *rq;
 345         unsigned int tag;
 346         bool put_ctx_on_error = false;
 347
 348         blk_queue_enter_live(q);
 349         data->q = q;
 350         if (likely(!data->ctx)) {
 351                 data->ctx = blk_mq_get_ctx(q);
 352                 put_ctx_on_error = true;
 353         }
 354         if (likely(!data->hctx))
 355                 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
 356         if (op & REQ_NOWAIT)
 357                 data->flags |= BLK_MQ_REQ_NOWAIT;
 358
 359         if (e) {
 360                 data->flags |= BLK_MQ_REQ_INTERNAL;
 361
 362                 /*
 363                  * Flush requests are special and go directly to the
 364                  * dispatch list. Don't include reserved tags in the
 365                  * limiting, as it isn't useful.
 366                  */
 367                 if (!op_is_flush(op) && e->type->ops.mq.limit_depth &&
 368                     !(data->flags & BLK_MQ_REQ_RESERVED))
 369                         e->type->ops.mq.limit_depth(op, data);
 370         }
 371
 372         tag = blk_mq_get_tag(data);
 373         if (tag == BLK_MQ_TAG_FAIL) {
 374                 if (put_ctx_on_error) {
 375                         blk_mq_put_ctx(data->ctx);
 376                         data->ctx = NULL;
 377                 }
 378                 blk_queue_exit(q);
 379                 return NULL;
 380         }
 381
 382         rq = blk_mq_rq_ctx_init(data, tag, op);
 383         if (!op_is_flush(op)) {
 384                 rq->elv.icq = NULL;
 385                 if (e && e->type->ops.mq.prepare_request) {
 386                         if (e->type->icq_cache && rq_ioc(bio))
 387                                 blk_mq_sched_assign_ioc(rq, bio);
 388
 389                         e->type->ops.mq.prepare_request(rq, bio);
 390                         rq->rq_flags |= RQF_ELVPRIV;
 391                 }
 392         }
 393         data->hctx->queued++;
 394         return rq;
 395 }
 396
 397 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 398                 blk_mq_req_flags_t flags)
 399 {
 400         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 401         struct request *rq;
 402         int ret;
 403
 404         ret = blk_queue_enter(q, flags);
 405         if (ret)
 406                 return ERR_PTR(ret);
 407
 408         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 409         blk_queue_exit(q);
 410
 411         if (!rq)
 412                 return ERR_PTR(-EWOULDBLOCK);
 413
 414         blk_mq_put_ctx(alloc_data.ctx);
 415
 416         rq->__data_len = 0;
 417         rq->__sector = (sector_t) -1;
 418         rq->bio = rq->biotail = NULL;
 419         return rq;
 420 }
 421 EXPORT_SYMBOL(blk_mq_alloc_request);
 422
 423 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 424         unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
 425 {
 426         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 427         struct request *rq;
 428         unsigned int cpu;
 429         int ret;
 430
 431         /*
 432          * If the tag allocator sleeps we could get an allocation for a
 433          * different hardware context.  No need to complicate the low level
 434          * allocator for this for the rare use case of a command tied to
 435          * a specific queue.
 436          */
 437         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 438                 return ERR_PTR(-EINVAL);
 439
 440         if (hctx_idx >= q->nr_hw_queues)
 441                 return ERR_PTR(-EIO);
 442
 443         ret = blk_queue_enter(q, flags);
 444         if (ret)
 445                 return ERR_PTR(ret);
 446
 447         /*
 448          * Check if the hardware context is actually mapped to anything.
 449          * If not tell the caller that it should skip this queue.
 450          */
 451         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 452         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 453                 blk_queue_exit(q);
 454                 return ERR_PTR(-EXDEV);
 455         }
 456         cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
 457         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 458
 459         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 460         blk_queue_exit(q);
 461
 462         if (!rq)
 463                 return ERR_PTR(-EWOULDBLOCK);
 464
 465         return rq;
 466 }
 467 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 468
 469 static void __blk_mq_free_request(struct request *rq)
 470 {
 471         struct request_queue *q = rq->q;
 472         struct blk_mq_ctx *ctx = rq->mq_ctx;
 473         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 474         const int sched_tag = rq->internal_tag;
 475
 476         if (rq->tag != -1)
 477                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 478         if (sched_tag != -1)
 479                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 480         blk_mq_sched_restart(hctx);
 481         blk_queue_exit(q);
 482 }
 483
 484 void blk_mq_free_request(struct request *rq)
 485 {
 486         struct request_queue *q = rq->q;
 487         struct elevator_queue *e = q->elevator;
 488         struct blk_mq_ctx *ctx = rq->mq_ctx;
 489         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 490
 491         if (rq->rq_flags & RQF_ELVPRIV) {
 492                 if (e && e->type->ops.mq.finish_request)
 493                         e->type->ops.mq.finish_request(rq);
 494                 if (rq->elv.icq) {
 495                         put_io_context(rq->elv.icq->ioc);
 496                         rq->elv.icq = NULL;
 497                 }
 498         }
 499
 500         ctx->rq_completed[rq_is_sync(rq)]++;
 501         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 502                 atomic_dec(&hctx->nr_active);
 503
 504         if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
 505                 laptop_io_completion(q->backing_dev_info);
 506
 507         wbt_done(q->rq_wb, rq);
 508
 509         if (blk_rq_rl(rq))
 510                 blk_put_rl(blk_rq_rl(rq));
 511
 512         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 513         if (refcount_dec_and_test(&rq->ref))
 514                 __blk_mq_free_request(rq);
 515 }
 516 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 517
 518 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 519 {
 520         u64 now = ktime_get_ns();
 521
 522         if (rq->rq_flags & RQF_STATS) {
 523                 blk_mq_poll_stats_start(rq->q);
 524                 blk_stat_add(rq, now);
 525         }
 526
 527         blk_account_io_done(rq, now);
 528
 529         if (rq->end_io) {
 530                 wbt_done(rq->q->rq_wb, rq);
 531                 rq->end_io(rq, error);
 532         } else {
 533                 if (unlikely(blk_bidi_rq(rq)))
 534                         blk_mq_free_request(rq->next_rq);
 535                 blk_mq_free_request(rq);
 536         }
 537 }
 538 EXPORT_SYMBOL(__blk_mq_end_request);
 539
 540 void blk_mq_end_request(struct request *rq, blk_status_t error)
 541 {
 542         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 543                 BUG();
 544         __blk_mq_end_request(rq, error);
 545 }
 546 EXPORT_SYMBOL(blk_mq_end_request);
 547
 548 static void __blk_mq_complete_request_remote(void *data)
 549 {
 550         struct request *rq = data;
 551
 552         rq->q->softirq_done_fn(rq);
 553 }
 554
 555 static void __blk_mq_complete_request(struct request *rq)
 556 {
 557         struct blk_mq_ctx *ctx = rq->mq_ctx;
 558         bool shared = false;
 559         int cpu;
 560
 561         if (!blk_mq_mark_complete(rq))
 562                 return;
 563         if (rq->internal_tag != -1)
 564                 blk_mq_sched_completed_request(rq);
 565
 566         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 567                 rq->q->softirq_done_fn(rq);
 568                 return;
 569         }
 570
 571         cpu = get_cpu();
 572         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 573                 shared = cpus_share_cache(cpu, ctx->cpu);
 574
 575         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 576                 rq->csd.func = __blk_mq_complete_request_remote;
 577                 rq->csd.info = rq;
 578                 rq->csd.flags = 0;
 579                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 580         } else {
 581                 rq->q->softirq_done_fn(rq);
 582         }
 583         put_cpu();
 584 }
 585
 586 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
 587         __releases(hctx->srcu)
 588 {
 589         if (!(hctx->flags & BLK_MQ_F_BLOCKING))
 590                 rcu_read_unlock();
 591         else
 592                 srcu_read_unlock(hctx->srcu, srcu_idx);
 593 }
 594
 595 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
 596         __acquires(hctx->srcu)
 597 {
 598         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
 599                 /* shut up gcc false positive */
 600                 *srcu_idx = 0;
 601                 rcu_read_lock();
 602         } else
 603                 *srcu_idx = srcu_read_lock(hctx->srcu);
 604 }
 605
 606 /**
 607  * blk_mq_complete_request - end I/O on a request
 608  * @rq:         the request being processed
 609  *
 610  * Description:
 611  *      Ends all I/O on a request. It does not handle partial completions.
 612  *      The actual completion happens out-of-order, through a IPI handler.
 613  **/
 614 void blk_mq_complete_request(struct request *rq)
 615 {
 616         if (unlikely(blk_should_fake_timeout(rq->q)))
 617                 return;
 618         __blk_mq_complete_request(rq);
 619 }
 620 EXPORT_SYMBOL(blk_mq_complete_request);
 621
 622 int blk_mq_request_started(struct request *rq)
 623 {
 624         return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
 625 }
 626 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 627
 628 void blk_mq_start_request(struct request *rq)
 629 {
 630         struct request_queue *q = rq->q;
 631
 632         blk_mq_sched_started_request(rq);
 633
 634         trace_block_rq_issue(q, rq);
 635
 636         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 637                 rq->io_start_time_ns = ktime_get_ns();
 638 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
 639                 rq->throtl_size = blk_rq_sectors(rq);
 640 #endif
 641                 rq->rq_flags |= RQF_STATS;
 642                 wbt_issue(q->rq_wb, rq);
 643         }
 644
 645         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
 646
 647         blk_add_timer(rq);
 648         WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
 649
 650         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 651                 /*
 652                  * Make sure space for the drain appears.  We know we can do
 653                  * this because max_hw_segments has been adjusted to be one
 654                  * fewer than the device can handle.
 655                  */
 656                 rq->nr_phys_segments++;
 657         }
 658 }
 659 EXPORT_SYMBOL(blk_mq_start_request);
 660
 661 static void __blk_mq_requeue_request(struct request *rq)
 662 {
 663         struct request_queue *q = rq->q;
 664
 665         blk_mq_put_driver_tag(rq);
 666
 667         trace_block_rq_requeue(q, rq);
 668         wbt_requeue(q->rq_wb, rq);
 669
 670         if (blk_mq_request_started(rq)) {
 671                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 672                 rq->rq_flags &= ~RQF_TIMED_OUT;
 673                 if (q->dma_drain_size && blk_rq_bytes(rq))
 674                         rq->nr_phys_segments--;
 675         }
 676 }
 677
 678 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 679 {
 680         __blk_mq_requeue_request(rq);
 681
 682         /* this request will be re-inserted to io scheduler queue */
 683         blk_mq_sched_requeue_request(rq);
 684
 685         BUG_ON(blk_queued_rq(rq));
 686         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 687 }
 688 EXPORT_SYMBOL(blk_mq_requeue_request);
 689
 690 static void blk_mq_requeue_work(struct work_struct *work)
 691 {
 692         struct request_queue *q =
 693                 container_of(work, struct request_queue, requeue_work.work);
 694         LIST_HEAD(rq_list);
 695         struct request *rq, *next;
 696
 697         spin_lock_irq(&q->requeue_lock);
 698         list_splice_init(&q->requeue_list, &rq_list);
 699         spin_unlock_irq(&q->requeue_lock);
 700
 701         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 702                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
 703                         continue;
 704
 705                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 706                 list_del_init(&rq->queuelist);
 707                 blk_mq_sched_insert_request(rq, true, false, false);
 708         }
 709
 710         while (!list_empty(&rq_list)) {
 711                 rq = list_entry(rq_list.next, struct request, queuelist);
 712                 list_del_init(&rq->queuelist);
 713                 blk_mq_sched_insert_request(rq, false, false, false);
 714         }
 715
 716         blk_mq_run_hw_queues(q, false);
 717 }
 718
 719 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 720                                 bool kick_requeue_list)
 721 {
 722         struct request_queue *q = rq->q;
 723         unsigned long flags;
 724
 725         /*
 726          * We abuse this flag that is otherwise used by the I/O scheduler to
 727          * request head insertion from the workqueue.
 728          */
 729         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 730
 731         spin_lock_irqsave(&q->requeue_lock, flags);
 732         if (at_head) {
 733                 rq->rq_flags |= RQF_SOFTBARRIER;
 734                 list_add(&rq->queuelist, &q->requeue_list);
 735         } else {
 736                 list_add_tail(&rq->queuelist, &q->requeue_list);
 737         }
 738         spin_unlock_irqrestore(&q->requeue_lock, flags);
 739
 740         if (kick_requeue_list)
 741                 blk_mq_kick_requeue_list(q);
 742 }
 743 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 744
 745 void blk_mq_kick_requeue_list(struct request_queue *q)
 746 {
 747         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
 748 }
 749 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 750
 751 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 752                                     unsigned long msecs)
 753 {
 754         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 755                                     msecs_to_jiffies(msecs));
 756 }
 757 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 758
 759 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 760 {
 761         if (tag < tags->nr_tags) {
 762                 prefetch(tags->rqs[tag]);
 763                 return tags->rqs[tag];
 764         }
 765
 766         return NULL;
 767 }
 768 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 769
 770 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 771 {
 772         req->rq_flags |= RQF_TIMED_OUT;
 773         if (req->q->mq_ops->timeout) {
 774                 enum blk_eh_timer_return ret;
 775
 776                 ret = req->q->mq_ops->timeout(req, reserved);
 777                 if (ret == BLK_EH_DONE)
 778                         return;
 779                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
 780         }
 781
 782         blk_add_timer(req);
 783 }
 784
 785 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
 786 {
 787         unsigned long deadline;
 788
 789         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
 790                 return false;
 791         if (rq->rq_flags & RQF_TIMED_OUT)
 792                 return false;
 793
 794         deadline = blk_rq_deadline(rq);
 795         if (time_after_eq(jiffies, deadline))
 796                 return true;
 797
 798         if (*next == 0)
 799                 *next = deadline;
 800         else if (time_after(*next, deadline))
 801                 *next = deadline;
 802         return false;
 803 }
 804
 805 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 806                 struct request *rq, void *priv, bool reserved)
 807 {
 808         unsigned long *next = priv;
 809
 810         /*
 811          * Just do a quick check if it is expired before locking the request in
 812          * so we're not unnecessarilly synchronizing across CPUs.
 813          */
 814         if (!blk_mq_req_expired(rq, next))
 815                 return;
 816
 817         /*
 818          * We have reason to believe the request may be expired. Take a
 819          * reference on the request to lock this request lifetime into its
 820          * currently allocated context to prevent it from being reallocated in
 821          * the event the completion by-passes this timeout handler.
 822          *
 823          * If the reference was already released, then the driver beat the
 824          * timeout handler to posting a natural completion.
 825          */
 826         if (!refcount_inc_not_zero(&rq->ref))
 827                 return;
 828
 829         /*
 830          * The request is now locked and cannot be reallocated underneath the
 831          * timeout handler's processing. Re-verify this exact request is truly
 832          * expired; if it is not expired, then the request was completed and
 833          * reallocated as a new request.
 834          */
 835         if (blk_mq_req_expired(rq, next))
 836                 blk_mq_rq_timed_out(rq, reserved);
 837         if (refcount_dec_and_test(&rq->ref))
 838                 __blk_mq_free_request(rq);
 839 }
 840
 841 static void blk_mq_timeout_work(struct work_struct *work)
 842 {
 843         struct request_queue *q =
 844                 container_of(work, struct request_queue, timeout_work);
 845         unsigned long next = 0;
 846         struct blk_mq_hw_ctx *hctx;
 847         int i;
 848
 849         /* A deadlock might occur if a request is stuck requiring a
 850          * timeout at the same time a queue freeze is waiting
 851          * completion, since the timeout code would not be able to
 852          * acquire the queue reference here.
 853          *
 854          * That's why we don't use blk_queue_enter here; instead, we use
 855          * percpu_ref_tryget directly, because we need to be able to
 856          * obtain a reference even in the short window between the queue
 857          * starting to freeze, by dropping the first reference in
 858          * blk_freeze_queue_start, and the moment the last request is
 859          * consumed, marked by the instant q_usage_counter reaches
 860          * zero.
 861          */
 862         if (!percpu_ref_tryget(&q->q_usage_counter))
 863                 return;
 864
 865         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
 866
 867         if (next != 0) {
 868                 mod_timer(&q->timeout, next);
 869         } else {
 870                 /*
 871                  * Request timeouts are handled as a forward rolling timer. If
 872                  * we end up here it means that no requests are pending and
 873                  * also that no request has been pending for a while. Mark
 874                  * each hctx as idle.
 875                  */
 876                 queue_for_each_hw_ctx(q, hctx, i) {
 877                         /* the hctx may be unmapped, so check it here */
 878                         if (blk_mq_hw_queue_mapped(hctx))
 879                                 blk_mq_tag_idle(hctx);
 880                 }
 881         }
 882         blk_queue_exit(q);
 883 }
 884
 885 struct flush_busy_ctx_data {
 886         struct blk_mq_hw_ctx *hctx;
 887         struct list_head *list;
 888 };
 889
 890 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 891 {
 892         struct flush_busy_ctx_data *flush_data = data;
 893         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 894         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 895
 896         spin_lock(&ctx->lock);
 897         list_splice_tail_init(&ctx->rq_list, flush_data->list);
 898         sbitmap_clear_bit(sb, bitnr);
 899         spin_unlock(&ctx->lock);
 900         return true;
 901 }
 902
 903 /*
 904  * Process software queues that have been marked busy, splicing them
 905  * to the for-dispatch
 906  */
 907 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 908 {
 909         struct flush_busy_ctx_data data = {
 910                 .hctx = hctx,
 911                 .list = list,
 912         };
 913
 914         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 915 }
 916 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 917
 918 struct dispatch_rq_data {
 919         struct blk_mq_hw_ctx *hctx;
 920         struct request *rq;
 921 };
 922
 923 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
 924                 void *data)
 925 {
 926         struct dispatch_rq_data *dispatch_data = data;
 927         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
 928         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 929
 930         spin_lock(&ctx->lock);
 931         if (!list_empty(&ctx->rq_list)) {
 932                 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
 933                 list_del_init(&dispatch_data->rq->queuelist);
 934                 if (list_empty(&ctx->rq_list))
 935                         sbitmap_clear_bit(sb, bitnr);
 936         }
 937         spin_unlock(&ctx->lock);
 938
 939         return !dispatch_data->rq;
 940 }
 941
 942 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
 943                                         struct blk_mq_ctx *start)
 944 {
 945         unsigned off = start ? start->index_hw : 0;
 946         struct dispatch_rq_data data = {
 947                 .hctx = hctx,
 948                 .rq   = NULL,
 949         };
 950
 951         __sbitmap_for_each_set(&hctx->ctx_map, off,
 952                                dispatch_rq_from_ctx, &data);
 953
 954         return data.rq;
 955 }
 956
 957 static inline unsigned int queued_to_index(unsigned int queued)
 958 {
 959         if (!queued)
 960                 return 0;
 961
 962         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
 963 }
 964
 965 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
 966                            bool wait)
 967 {
 968         struct blk_mq_alloc_data data = {
 969                 .q = rq->q,
 970                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
 971                 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
 972         };
 973
 974         might_sleep_if(wait);
 975
 976         if (rq->tag != -1)
 977                 goto done;
 978
 979         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
 980                 data.flags |= BLK_MQ_REQ_RESERVED;
 981
 982         rq->tag = blk_mq_get_tag(&data);
 983         if (rq->tag >= 0) {
 984                 if (blk_mq_tag_busy(data.hctx)) {
 985                         rq->rq_flags |= RQF_MQ_INFLIGHT;
 986                         atomic_inc(&data.hctx->nr_active);
 987                 }
 988                 data.hctx->tags->rqs[rq->tag] = rq;
 989         }
 990
 991 done:
 992         if (hctx)
 993                 *hctx = data.hctx;
 994         return rq->tag != -1;
 995 }
 996
 997 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
 998                                 int flags, void *key)
 999 {
1000         struct blk_mq_hw_ctx *hctx;
1001
1002         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1003
1004         list_del_init(&wait->entry);
1005         blk_mq_run_hw_queue(hctx, true);
1006         return 1;
1007 }
1008
1009 /*
1010  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1011  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1012  * restart. For both cases, take care to check the condition again after
1013  * marking us as waiting.
1014  */
1015 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1016                                  struct request *rq)
1017 {
1018         struct blk_mq_hw_ctx *this_hctx = *hctx;
1019         struct sbq_wait_state *ws;
1020         wait_queue_entry_t *wait;
1021         bool ret;
1022
1023         if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1024                 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1025                         set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1026
1027                 /*
1028                  * It's possible that a tag was freed in the window between the
1029                  * allocation failure and adding the hardware queue to the wait
1030                  * queue.
1031                  *
1032                  * Don't clear RESTART here, someone else could have set it.
1033                  * At most this will cost an extra queue run.
1034                  */
1035                 return blk_mq_get_driver_tag(rq, hctx, false);
1036         }
1037
1038         wait = &this_hctx->dispatch_wait;
1039         if (!list_empty_careful(&wait->entry))
1040                 return false;
1041
1042         spin_lock(&this_hctx->lock);
1043         if (!list_empty(&wait->entry)) {
1044                 spin_unlock(&this_hctx->lock);
1045                 return false;
1046         }
1047
1048         ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1049         add_wait_queue(&ws->wait, wait);
1050
1051         /*
1052          * It's possible that a tag was freed in the window between the
1053          * allocation failure and adding the hardware queue to the wait
1054          * queue.
1055          */
1056         ret = blk_mq_get_driver_tag(rq, hctx, false);
1057         if (!ret) {
1058                 spin_unlock(&this_hctx->lock);
1059                 return false;
1060         }
1061
1062         /*
1063          * We got a tag, remove ourselves from the wait queue to ensure
1064          * someone else gets the wakeup.
1065          */
1066         spin_lock_irq(&ws->wait.lock);
1067         list_del_init(&wait->entry);
1068         spin_unlock_irq(&ws->wait.lock);
1069         spin_unlock(&this_hctx->lock);
1070
1071         return true;
1072 }
1073
1074 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1075
1076 /*
1077  * Returns true if we did some work AND can potentially do more.
1078  */
1079 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1080                              bool got_budget)
1081 {
1082         struct blk_mq_hw_ctx *hctx;
1083         struct request *rq, *nxt;
1084         bool no_tag = false;
1085         int errors, queued;
1086         blk_status_t ret = BLK_STS_OK;
1087
1088         if (list_empty(list))
1089                 return false;
1090
1091         WARN_ON(!list_is_singular(list) && got_budget);
1092
1093         /*
1094          * Now process all the entries, sending them to the driver.
1095          */
1096         errors = queued = 0;
1097         do {
1098                 struct blk_mq_queue_data bd;
1099
1100                 rq = list_first_entry(list, struct request, queuelist);
1101
1102                 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1103                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1104                         break;
1105
1106                 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1107                         /*
1108                          * The initial allocation attempt failed, so we need to
1109                          * rerun the hardware queue when a tag is freed. The
1110                          * waitqueue takes care of that. If the queue is run
1111                          * before we add this entry back on the dispatch list,
1112                          * we'll re-run it below.
1113                          */
1114                         if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1115                                 blk_mq_put_dispatch_budget(hctx);
1116                                 /*
1117                                  * For non-shared tags, the RESTART check
1118                                  * will suffice.
1119                                  */
1120                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1121                                         no_tag = true;
1122                                 break;
1123                         }
1124                 }
1125
1126                 list_del_init(&rq->queuelist);
1127
1128                 bd.rq = rq;
1129
1130                 /*
1131                  * Flag last if we have no more requests, or if we have more
1132                  * but can't assign a driver tag to it.
1133                  */
1134                 if (list_empty(list))
1135                         bd.last = true;
1136                 else {
1137                         nxt = list_first_entry(list, struct request, queuelist);
1138                         bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1139                 }
1140
1141                 ret = q->mq_ops->queue_rq(hctx, &bd);
1142                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1143                         /*
1144                          * If an I/O scheduler has been configured and we got a
1145                          * driver tag for the next request already, free it
1146                          * again.
1147                          */
1148                         if (!list_empty(list)) {
1149                                 nxt = list_first_entry(list, struct request, queuelist);
1150                                 blk_mq_put_driver_tag(nxt);
1151                         }
1152                         list_add(&rq->queuelist, list);
1153                         __blk_mq_requeue_request(rq);
1154                         break;
1155                 }
1156
1157                 if (unlikely(ret != BLK_STS_OK)) {
1158                         errors++;
1159                         blk_mq_end_request(rq, BLK_STS_IOERR);
1160                         continue;
1161                 }
1162
1163                 queued++;
1164         } while (!list_empty(list));
1165
1166         hctx->dispatched[queued_to_index(queued)]++;
1167
1168         /*
1169          * Any items that need requeuing? Stuff them into hctx->dispatch,
1170          * that is where we will continue on next queue run.
1171          */
1172         if (!list_empty(list)) {
1173                 bool needs_restart;
1174
1175                 spin_lock(&hctx->lock);
1176                 list_splice_init(list, &hctx->dispatch);
1177                 spin_unlock(&hctx->lock);
1178
1179                 /*
1180                  * If SCHED_RESTART was set by the caller of this function and
1181                  * it is no longer set that means that it was cleared by another
1182                  * thread and hence that a queue rerun is needed.
1183                  *
1184                  * If 'no_tag' is set, that means that we failed getting
1185                  * a driver tag with an I/O scheduler attached. If our dispatch
1186                  * waitqueue is no longer active, ensure that we run the queue
1187                  * AFTER adding our entries back to the list.
1188                  *
1189                  * If no I/O scheduler has been configured it is possible that
1190                  * the hardware queue got stopped and restarted before requests
1191                  * were pushed back onto the dispatch list. Rerun the queue to
1192                  * avoid starvation. Notes:
1193                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1194                  *   been stopped before rerunning a queue.
1195                  * - Some but not all block drivers stop a queue before
1196                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1197                  *   and dm-rq.
1198                  *
1199                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1200                  * bit is set, run queue after a delay to avoid IO stalls
1201                  * that could otherwise occur if the queue is idle.
1202                  */
1203                 needs_restart = blk_mq_sched_needs_restart(hctx);
1204                 if (!needs_restart ||
1205                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1206                         blk_mq_run_hw_queue(hctx, true);
1207                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1208                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1209
1210                 return false;
1211         }
1212
1213         /*
1214          * If the host/device is unable to accept more work, inform the
1215          * caller of that.
1216          */
1217         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1218                 return false;
1219
1220         return (queued + errors) != 0;
1221 }
1222
1223 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1224 {
1225         int srcu_idx;
1226
1227         /*
1228          * We should be running this queue from one of the CPUs that
1229          * are mapped to it.
1230          *
1231          * There are at least two related races now between setting
1232          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1233          * __blk_mq_run_hw_queue():
1234          *
1235          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1236          *   but later it becomes online, then this warning is harmless
1237          *   at all
1238          *
1239          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1240          *   but later it becomes offline, then the warning can't be
1241          *   triggered, and we depend on blk-mq timeout handler to
1242          *   handle dispatched requests to this hctx
1243          */
1244         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1245                 cpu_online(hctx->next_cpu)) {
1246                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1247                         raw_smp_processor_id(),
1248                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1249                 dump_stack();
1250         }
1251
1252         /*
1253          * We can't run the queue inline with ints disabled. Ensure that
1254          * we catch bad users of this early.
1255          */
1256         WARN_ON_ONCE(in_interrupt());
1257
1258         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1259
1260         hctx_lock(hctx, &srcu_idx);
1261         blk_mq_sched_dispatch_requests(hctx);
1262         hctx_unlock(hctx, srcu_idx);
1263 }
1264
1265 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1266 {
1267         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1268
1269         if (cpu >= nr_cpu_ids)
1270                 cpu = cpumask_first(hctx->cpumask);
1271         return cpu;
1272 }
1273
1274 /*
1275  * It'd be great if the workqueue API had a way to pass
1276  * in a mask and had some smarts for more clever placement.
1277  * For now we just round-robin here, switching for every
1278  * BLK_MQ_CPU_WORK_BATCH queued items.
1279  */
1280 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1281 {
1282         bool tried = false;
1283         int next_cpu = hctx->next_cpu;
1284
1285         if (hctx->queue->nr_hw_queues == 1)
1286                 return WORK_CPU_UNBOUND;
1287
1288         if (--hctx->next_cpu_batch <= 0) {
1289 select_cpu:
1290                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1291                                 cpu_online_mask);
1292                 if (next_cpu >= nr_cpu_ids)
1293                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1294                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1295         }
1296
1297         /*
1298          * Do unbound schedule if we can't find a online CPU for this hctx,
1299          * and it should only happen in the path of handling CPU DEAD.
1300          */
1301         if (!cpu_online(next_cpu)) {
1302                 if (!tried) {
1303                         tried = true;
1304                         goto select_cpu;
1305                 }
1306
1307                 /*
1308                  * Make sure to re-select CPU next time once after CPUs
1309                  * in hctx->cpumask become online again.
1310                  */
1311                 hctx->next_cpu = next_cpu;
1312                 hctx->next_cpu_batch = 1;
1313                 return WORK_CPU_UNBOUND;
1314         }
1315
1316         hctx->next_cpu = next_cpu;
1317         return next_cpu;
1318 }
1319
1320 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1321                                         unsigned long msecs)
1322 {
1323         if (unlikely(blk_mq_hctx_stopped(hctx)))
1324                 return;
1325
1326         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1327                 int cpu = get_cpu();
1328                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1329                         __blk_mq_run_hw_queue(hctx);
1330                         put_cpu();
1331                         return;
1332                 }
1333
1334                 put_cpu();
1335         }
1336
1337         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1338                                     msecs_to_jiffies(msecs));
1339 }
1340
1341 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1342 {
1343         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1344 }
1345 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1346
1347 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1348 {
1349         int srcu_idx;
1350         bool need_run;
1351
1352         /*
1353          * When queue is quiesced, we may be switching io scheduler, or
1354          * updating nr_hw_queues, or other things, and we can't run queue
1355          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1356          *
1357          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1358          * quiesced.
1359          */
1360         hctx_lock(hctx, &srcu_idx);
1361         need_run = !blk_queue_quiesced(hctx->queue) &&
1362                 blk_mq_hctx_has_pending(hctx);
1363         hctx_unlock(hctx, srcu_idx);
1364
1365         if (need_run) {
1366                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1367                 return true;
1368         }
1369
1370         return false;
1371 }
1372 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1373
1374 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1375 {
1376         struct blk_mq_hw_ctx *hctx;
1377         int i;
1378
1379         queue_for_each_hw_ctx(q, hctx, i) {
1380                 if (blk_mq_hctx_stopped(hctx))
1381                         continue;
1382
1383                 blk_mq_run_hw_queue(hctx, async);
1384         }
1385 }
1386 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1387
1388 /**
1389  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1390  * @q: request queue.
1391  *
1392  * The caller is responsible for serializing this function against
1393  * blk_mq_{start,stop}_hw_queue().
1394  */
1395 bool blk_mq_queue_stopped(struct request_queue *q)
1396 {
1397         struct blk_mq_hw_ctx *hctx;
1398         int i;
1399
1400         queue_for_each_hw_ctx(q, hctx, i)
1401                 if (blk_mq_hctx_stopped(hctx))
1402                         return true;
1403
1404         return false;
1405 }
1406 EXPORT_SYMBOL(blk_mq_queue_stopped);
1407
1408 /*
1409  * This function is often used for pausing .queue_rq() by driver when
1410  * there isn't enough resource or some conditions aren't satisfied, and
1411  * BLK_STS_RESOURCE is usually returned.
1412  *
1413  * We do not guarantee that dispatch can be drained or blocked
1414  * after blk_mq_stop_hw_queue() returns. Please use
1415  * blk_mq_quiesce_queue() for that requirement.
1416  */
1417 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1418 {
1419         cancel_delayed_work(&hctx->run_work);
1420
1421         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1422 }
1423 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1424
1425 /*
1426  * This function is often used for pausing .queue_rq() by driver when
1427  * there isn't enough resource or some conditions aren't satisfied, and
1428  * BLK_STS_RESOURCE is usually returned.
1429  *
1430  * We do not guarantee that dispatch can be drained or blocked
1431  * after blk_mq_stop_hw_queues() returns. Please use
1432  * blk_mq_quiesce_queue() for that requirement.
1433  */
1434 void blk_mq_stop_hw_queues(struct request_queue *q)
1435 {
1436         struct blk_mq_hw_ctx *hctx;
1437         int i;
1438
1439         queue_for_each_hw_ctx(q, hctx, i)
1440                 blk_mq_stop_hw_queue(hctx);
1441 }
1442 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1443
1444 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1445 {
1446         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1447
1448         blk_mq_run_hw_queue(hctx, false);
1449 }
1450 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1451
1452 void blk_mq_start_hw_queues(struct request_queue *q)
1453 {
1454         struct blk_mq_hw_ctx *hctx;
1455         int i;
1456
1457         queue_for_each_hw_ctx(q, hctx, i)
1458                 blk_mq_start_hw_queue(hctx);
1459 }
1460 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1461
1462 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1463 {
1464         if (!blk_mq_hctx_stopped(hctx))
1465                 return;
1466
1467         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1468         blk_mq_run_hw_queue(hctx, async);
1469 }
1470 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1471
1472 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1473 {
1474         struct blk_mq_hw_ctx *hctx;
1475         int i;
1476
1477         queue_for_each_hw_ctx(q, hctx, i)
1478                 blk_mq_start_stopped_hw_queue(hctx, async);
1479 }
1480 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1481
1482 static void blk_mq_run_work_fn(struct work_struct *work)
1483 {
1484         struct blk_mq_hw_ctx *hctx;
1485
1486         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1487
1488         /*
1489          * If we are stopped, don't run the queue.
1490          */
1491         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1492                 return;
1493
1494         __blk_mq_run_hw_queue(hctx);
1495 }
1496
1497 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1498                                             struct request *rq,
1499                                             bool at_head)
1500 {
1501         struct blk_mq_ctx *ctx = rq->mq_ctx;
1502
1503         lockdep_assert_held(&ctx->lock);
1504
1505         trace_block_rq_insert(hctx->queue, rq);
1506
1507         if (at_head)
1508                 list_add(&rq->queuelist, &ctx->rq_list);
1509         else
1510                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1511 }
1512
1513 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1514                              bool at_head)
1515 {
1516         struct blk_mq_ctx *ctx = rq->mq_ctx;
1517
1518         lockdep_assert_held(&ctx->lock);
1519
1520         __blk_mq_insert_req_list(hctx, rq, at_head);
1521         blk_mq_hctx_mark_pending(hctx, ctx);
1522 }
1523
1524 /*
1525  * Should only be used carefully, when the caller knows we want to
1526  * bypass a potential IO scheduler on the target device.
1527  */
1528 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1529 {
1530         struct blk_mq_ctx *ctx = rq->mq_ctx;
1531         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1532
1533         spin_lock(&hctx->lock);
1534         list_add_tail(&rq->queuelist, &hctx->dispatch);
1535         spin_unlock(&hctx->lock);
1536
1537         if (run_queue)
1538                 blk_mq_run_hw_queue(hctx, false);
1539 }
1540
1541 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1542                             struct list_head *list)
1543
1544 {
1545         /*
1546          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1547          * offline now
1548          */
1549         spin_lock(&ctx->lock);
1550         while (!list_empty(list)) {
1551                 struct request *rq;
1552
1553                 rq = list_first_entry(list, struct request, queuelist);
1554                 BUG_ON(rq->mq_ctx != ctx);
1555                 list_del_init(&rq->queuelist);
1556                 __blk_mq_insert_req_list(hctx, rq, false);
1557         }
1558         blk_mq_hctx_mark_pending(hctx, ctx);
1559         spin_unlock(&ctx->lock);
1560 }
1561
1562 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1563 {
1564         struct request *rqa = container_of(a, struct request, queuelist);
1565         struct request *rqb = container_of(b, struct request, queuelist);
1566
1567         return !(rqa->mq_ctx < rqb->mq_ctx ||
1568                  (rqa->mq_ctx == rqb->mq_ctx &&
1569                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1570 }
1571
1572 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1573 {
1574         struct blk_mq_ctx *this_ctx;
1575         struct request_queue *this_q;
1576         struct request *rq;
1577         LIST_HEAD(list);
1578         LIST_HEAD(ctx_list);
1579         unsigned int depth;
1580
1581         list_splice_init(&plug->mq_list, &list);
1582
1583         list_sort(NULL, &list, plug_ctx_cmp);
1584
1585         this_q = NULL;
1586         this_ctx = NULL;
1587         depth = 0;
1588
1589         while (!list_empty(&list)) {
1590                 rq = list_entry_rq(list.next);
1591                 list_del_init(&rq->queuelist);
1592                 BUG_ON(!rq->q);
1593                 if (rq->mq_ctx != this_ctx) {
1594                         if (this_ctx) {
1595                                 trace_block_unplug(this_q, depth, from_schedule);
1596                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1597                                                                 &ctx_list,
1598                                                                 from_schedule);
1599                         }
1600
1601                         this_ctx = rq->mq_ctx;
1602                         this_q = rq->q;
1603                         depth = 0;
1604                 }
1605
1606                 depth++;
1607                 list_add_tail(&rq->queuelist, &ctx_list);
1608         }
1609
1610         /*
1611          * If 'this_ctx' is set, we know we have entries to complete
1612          * on 'ctx_list'. Do those.
1613          */
1614         if (this_ctx) {
1615                 trace_block_unplug(this_q, depth, from_schedule);
1616                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1617                                                 from_schedule);
1618         }
1619 }
1620
1621 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1622 {
1623         blk_init_request_from_bio(rq, bio);
1624
1625         blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1626
1627         blk_account_io_start(rq, true);
1628 }
1629
1630 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1631 {
1632         if (rq->tag != -1)
1633                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1634
1635         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1636 }
1637
1638 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1639                                             struct request *rq,
1640                                             blk_qc_t *cookie)
1641 {
1642         struct request_queue *q = rq->q;
1643         struct blk_mq_queue_data bd = {
1644                 .rq = rq,
1645                 .last = true,
1646         };
1647         blk_qc_t new_cookie;
1648         blk_status_t ret;
1649
1650         new_cookie = request_to_qc_t(hctx, rq);
1651
1652         /*
1653          * For OK queue, we are done. For error, caller may kill it.
1654          * Any other error (busy), just add it to our list as we
1655          * previously would have done.
1656          */
1657         ret = q->mq_ops->queue_rq(hctx, &bd);
1658         switch (ret) {
1659         case BLK_STS_OK:
1660                 *cookie = new_cookie;
1661                 break;
1662         case BLK_STS_RESOURCE:
1663         case BLK_STS_DEV_RESOURCE:
1664                 __blk_mq_requeue_request(rq);
1665                 break;
1666         default:
1667                 *cookie = BLK_QC_T_NONE;
1668                 break;
1669         }
1670
1671         return ret;
1672 }
1673
1674 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1675                                                 struct request *rq,
1676                                                 blk_qc_t *cookie,
1677                                                 bool bypass_insert)
1678 {
1679         struct request_queue *q = rq->q;
1680         bool run_queue = true;
1681
1682         /*
1683          * RCU or SRCU read lock is needed before checking quiesced flag.
1684          *
1685          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1686          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1687          * and avoid driver to try to dispatch again.
1688          */
1689         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1690                 run_queue = false;
1691                 bypass_insert = false;
1692                 goto insert;
1693         }
1694
1695         if (q->elevator && !bypass_insert)
1696                 goto insert;
1697
1698         if (!blk_mq_get_dispatch_budget(hctx))
1699                 goto insert;
1700
1701         if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1702                 blk_mq_put_dispatch_budget(hctx);
1703                 goto insert;
1704         }
1705
1706         return __blk_mq_issue_directly(hctx, rq, cookie);
1707 insert:
1708         if (bypass_insert)
1709                 return BLK_STS_RESOURCE;
1710
1711         blk_mq_sched_insert_request(rq, false, run_queue, false);
1712         return BLK_STS_OK;
1713 }
1714
1715 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1716                 struct request *rq, blk_qc_t *cookie)
1717 {
1718         blk_status_t ret;
1719         int srcu_idx;
1720
1721         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1722
1723         hctx_lock(hctx, &srcu_idx);
1724
1725         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1726         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1727                 blk_mq_sched_insert_request(rq, false, true, false);
1728         else if (ret != BLK_STS_OK)
1729                 blk_mq_end_request(rq, ret);
1730
1731         hctx_unlock(hctx, srcu_idx);
1732 }
1733
1734 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1735 {
1736         blk_status_t ret;
1737         int srcu_idx;
1738         blk_qc_t unused_cookie;
1739         struct blk_mq_ctx *ctx = rq->mq_ctx;
1740         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1741
1742         hctx_lock(hctx, &srcu_idx);
1743         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1744         hctx_unlock(hctx, srcu_idx);
1745
1746         return ret;
1747 }
1748
1749 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1750 {
1751         const int is_sync = op_is_sync(bio->bi_opf);
1752         const int is_flush_fua = op_is_flush(bio->bi_opf);
1753         struct blk_mq_alloc_data data = { .flags = 0 };
1754         struct request *rq;
1755         unsigned int request_count = 0;
1756         struct blk_plug *plug;
1757         struct request *same_queue_rq = NULL;
1758         blk_qc_t cookie;
1759         unsigned int wb_acct;
1760
1761         blk_queue_bounce(q, &bio);
1762
1763         blk_queue_split(q, &bio);
1764
1765         if (!bio_integrity_prep(bio))
1766                 return BLK_QC_T_NONE;
1767
1768         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1769             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1770                 return BLK_QC_T_NONE;
1771
1772         if (blk_mq_sched_bio_merge(q, bio))
1773                 return BLK_QC_T_NONE;
1774
1775         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1776
1777         trace_block_getrq(q, bio, bio->bi_opf);
1778
1779         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1780         if (unlikely(!rq)) {
1781                 __wbt_done(q->rq_wb, wb_acct);
1782                 if (bio->bi_opf & REQ_NOWAIT)
1783                         bio_wouldblock_error(bio);
1784                 return BLK_QC_T_NONE;
1785         }
1786
1787         wbt_track(rq, wb_acct);
1788
1789         cookie = request_to_qc_t(data.hctx, rq);
1790
1791         plug = current->plug;
1792         if (unlikely(is_flush_fua)) {
1793                 blk_mq_put_ctx(data.ctx);
1794                 blk_mq_bio_to_request(rq, bio);
1795
1796                 /* bypass scheduler for flush rq */
1797                 blk_insert_flush(rq);
1798                 blk_mq_run_hw_queue(data.hctx, true);
1799         } else if (plug && q->nr_hw_queues == 1) {
1800                 struct request *last = NULL;
1801
1802                 blk_mq_put_ctx(data.ctx);
1803                 blk_mq_bio_to_request(rq, bio);
1804
1805                 /*
1806                  * @request_count may become stale because of schedule
1807                  * out, so check the list again.
1808                  */
1809                 if (list_empty(&plug->mq_list))
1810                         request_count = 0;
1811                 else if (blk_queue_nomerges(q))
1812                         request_count = blk_plug_queued_count(q);
1813
1814                 if (!request_count)
1815                         trace_block_plug(q);
1816                 else
1817                         last = list_entry_rq(plug->mq_list.prev);
1818
1819                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1820                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1821                         blk_flush_plug_list(plug, false);
1822                         trace_block_plug(q);
1823                 }
1824
1825                 list_add_tail(&rq->queuelist, &plug->mq_list);
1826         } else if (plug && !blk_queue_nomerges(q)) {
1827                 blk_mq_bio_to_request(rq, bio);
1828
1829                 /*
1830                  * We do limited plugging. If the bio can be merged, do that.
1831                  * Otherwise the existing request in the plug list will be
1832                  * issued. So the plug list will have one request at most
1833                  * The plug list might get flushed before this. If that happens,
1834                  * the plug list is empty, and same_queue_rq is invalid.
1835                  */
1836                 if (list_empty(&plug->mq_list))
1837                         same_queue_rq = NULL;
1838                 if (same_queue_rq)
1839                         list_del_init(&same_queue_rq->queuelist);
1840                 list_add_tail(&rq->queuelist, &plug->mq_list);
1841
1842                 blk_mq_put_ctx(data.ctx);
1843
1844                 if (same_queue_rq) {
1845                         data.hctx = blk_mq_map_queue(q,
1846                                         same_queue_rq->mq_ctx->cpu);
1847                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1848                                         &cookie);
1849                 }
1850         } else if (q->nr_hw_queues > 1 && is_sync) {
1851                 blk_mq_put_ctx(data.ctx);
1852                 blk_mq_bio_to_request(rq, bio);
1853                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1854         } else {
1855                 blk_mq_put_ctx(data.ctx);
1856                 blk_mq_bio_to_request(rq, bio);
1857                 blk_mq_sched_insert_request(rq, false, true, true);
1858         }
1859
1860         return cookie;
1861 }
1862
1863 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1864                      unsigned int hctx_idx)
1865 {
1866         struct page *page;
1867
1868         if (tags->rqs && set->ops->exit_request) {
1869                 int i;
1870
1871                 for (i = 0; i < tags->nr_tags; i++) {
1872                         struct request *rq = tags->static_rqs[i];
1873
1874                         if (!rq)
1875                                 continue;
1876                         set->ops->exit_request(set, rq, hctx_idx);
1877                         tags->static_rqs[i] = NULL;
1878                 }
1879         }
1880
1881         while (!list_empty(&tags->page_list)) {
1882                 page = list_first_entry(&tags->page_list, struct page, lru);
1883                 list_del_init(&page->lru);
1884                 /*
1885                  * Remove kmemleak object previously allocated in
1886                  * blk_mq_init_rq_map().
1887                  */
1888                 kmemleak_free(page_address(page));
1889                 __free_pages(page, page->private);
1890         }
1891 }
1892
1893 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1894 {
1895         kfree(tags->rqs);
1896         tags->rqs = NULL;
1897         kfree(tags->static_rqs);
1898         tags->static_rqs = NULL;
1899
1900         blk_mq_free_tags(tags);
1901 }
1902
1903 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1904                                         unsigned int hctx_idx,
1905                                         unsigned int nr_tags,
1906                                         unsigned int reserved_tags)
1907 {
1908         struct blk_mq_tags *tags;
1909         int node;
1910
1911         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1912         if (node == NUMA_NO_NODE)
1913                 node = set->numa_node;
1914
1915         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1916                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1917         if (!tags)
1918                 return NULL;
1919
1920         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
1921                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1922                                  node);
1923         if (!tags->rqs) {
1924                 blk_mq_free_tags(tags);
1925                 return NULL;
1926         }
1927
1928         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
1929                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1930                                         node);
1931         if (!tags->static_rqs) {
1932                 kfree(tags->rqs);
1933                 blk_mq_free_tags(tags);
1934                 return NULL;
1935         }
1936
1937         return tags;
1938 }
1939
1940 static size_t order_to_size(unsigned int order)
1941 {
1942         return (size_t)PAGE_SIZE << order;
1943 }
1944
1945 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1946                                unsigned int hctx_idx, int node)
1947 {
1948         int ret;
1949
1950         if (set->ops->init_request) {
1951                 ret = set->ops->init_request(set, rq, hctx_idx, node);
1952                 if (ret)
1953                         return ret;
1954         }
1955
1956         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1957         return 0;
1958 }
1959
1960 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1961                      unsigned int hctx_idx, unsigned int depth)
1962 {
1963         unsigned int i, j, entries_per_page, max_order = 4;
1964         size_t rq_size, left;
1965         int node;
1966
1967         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1968         if (node == NUMA_NO_NODE)
1969                 node = set->numa_node;
1970
1971         INIT_LIST_HEAD(&tags->page_list);
1972
1973         /*
1974          * rq_size is the size of the request plus driver payload, rounded
1975          * to the cacheline size
1976          */
1977         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1978                                 cache_line_size());
1979         left = rq_size * depth;
1980
1981         for (i = 0; i < depth; ) {
1982                 int this_order = max_order;
1983                 struct page *page;
1984                 int to_do;
1985                 void *p;
1986
1987                 while (this_order && left < order_to_size(this_order - 1))
1988                         this_order--;
1989
1990                 do {
1991                         page = alloc_pages_node(node,
1992                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1993                                 this_order);
1994                         if (page)
1995                                 break;
1996                         if (!this_order--)
1997                                 break;
1998                         if (order_to_size(this_order) < rq_size)
1999                                 break;
2000                 } while (1);
2001
2002                 if (!page)
2003                         goto fail;
2004
2005                 page->private = this_order;
2006                 list_add_tail(&page->lru, &tags->page_list);
2007
2008                 p = page_address(page);
2009                 /*
2010                  * Allow kmemleak to scan these pages as they contain pointers
2011                  * to additional allocations like via ops->init_request().
2012                  */
2013                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2014                 entries_per_page = order_to_size(this_order) / rq_size;
2015                 to_do = min(entries_per_page, depth - i);
2016                 left -= to_do * rq_size;
2017                 for (j = 0; j < to_do; j++) {
2018                         struct request *rq = p;
2019
2020                         tags->static_rqs[i] = rq;
2021                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2022                                 tags->static_rqs[i] = NULL;
2023                                 goto fail;
2024                         }
2025
2026                         p += rq_size;
2027                         i++;
2028                 }
2029         }
2030         return 0;
2031
2032 fail:
2033         blk_mq_free_rqs(set, tags, hctx_idx);
2034         return -ENOMEM;
2035 }
2036
2037 /*
2038  * 'cpu' is going away. splice any existing rq_list entries from this
2039  * software queue to the hw queue dispatch list, and ensure that it
2040  * gets run.
2041  */
2042 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2043 {
2044         struct blk_mq_hw_ctx *hctx;
2045         struct blk_mq_ctx *ctx;
2046         LIST_HEAD(tmp);
2047
2048         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2049         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2050
2051         spin_lock(&ctx->lock);
2052         if (!list_empty(&ctx->rq_list)) {
2053                 list_splice_init(&ctx->rq_list, &tmp);
2054                 blk_mq_hctx_clear_pending(hctx, ctx);
2055         }
2056         spin_unlock(&ctx->lock);
2057
2058         if (list_empty(&tmp))
2059                 return 0;
2060
2061         spin_lock(&hctx->lock);
2062         list_splice_tail_init(&tmp, &hctx->dispatch);
2063         spin_unlock(&hctx->lock);
2064
2065         blk_mq_run_hw_queue(hctx, true);
2066         return 0;
2067 }
2068
2069 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2070 {
2071         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2072                                             &hctx->cpuhp_dead);
2073 }
2074
2075 /* hctx->ctxs will be freed in queue's release handler */
2076 static void blk_mq_exit_hctx(struct request_queue *q,
2077                 struct blk_mq_tag_set *set,
2078                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2079 {
2080         blk_mq_debugfs_unregister_hctx(hctx);
2081
2082         if (blk_mq_hw_queue_mapped(hctx))
2083                 blk_mq_tag_idle(hctx);
2084
2085         if (set->ops->exit_request)
2086                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2087
2088         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2089
2090         if (set->ops->exit_hctx)
2091                 set->ops->exit_hctx(hctx, hctx_idx);
2092
2093         if (hctx->flags & BLK_MQ_F_BLOCKING)
2094                 cleanup_srcu_struct(hctx->srcu);
2095
2096         blk_mq_remove_cpuhp(hctx);
2097         blk_free_flush_queue(hctx->fq);
2098         sbitmap_free(&hctx->ctx_map);
2099 }
2100
2101 static void blk_mq_exit_hw_queues(struct request_queue *q,
2102                 struct blk_mq_tag_set *set, int nr_queue)
2103 {
2104         struct blk_mq_hw_ctx *hctx;
2105         unsigned int i;
2106
2107         queue_for_each_hw_ctx(q, hctx, i) {
2108                 if (i == nr_queue)
2109                         break;
2110                 blk_mq_exit_hctx(q, set, hctx, i);
2111         }
2112 }
2113
2114 static int blk_mq_init_hctx(struct request_queue *q,
2115                 struct blk_mq_tag_set *set,
2116                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2117 {
2118         int node;
2119
2120         node = hctx->numa_node;
2121         if (node == NUMA_NO_NODE)
2122                 node = hctx->numa_node = set->numa_node;
2123
2124         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2125         spin_lock_init(&hctx->lock);
2126         INIT_LIST_HEAD(&hctx->dispatch);
2127         hctx->queue = q;
2128         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2129
2130         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2131
2132         hctx->tags = set->tags[hctx_idx];
2133
2134         /*
2135          * Allocate space for all possible cpus to avoid allocation at
2136          * runtime
2137          */
2138         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2139                                         GFP_KERNEL, node);
2140         if (!hctx->ctxs)
2141                 goto unregister_cpu_notifier;
2142
2143         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2144                               node))
2145                 goto free_ctxs;
2146
2147         hctx->nr_ctx = 0;
2148
2149         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2150         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2151
2152         if (set->ops->init_hctx &&
2153             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2154                 goto free_bitmap;
2155
2156         if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2157                 goto exit_hctx;
2158
2159         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2160         if (!hctx->fq)
2161                 goto sched_exit_hctx;
2162
2163         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2164                 goto free_fq;
2165
2166         if (hctx->flags & BLK_MQ_F_BLOCKING)
2167                 init_srcu_struct(hctx->srcu);
2168
2169         blk_mq_debugfs_register_hctx(q, hctx);
2170
2171         return 0;
2172
2173  free_fq:
2174         kfree(hctx->fq);
2175  sched_exit_hctx:
2176         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2177  exit_hctx:
2178         if (set->ops->exit_hctx)
2179                 set->ops->exit_hctx(hctx, hctx_idx);
2180  free_bitmap:
2181         sbitmap_free(&hctx->ctx_map);
2182  free_ctxs:
2183         kfree(hctx->ctxs);
2184  unregister_cpu_notifier:
2185         blk_mq_remove_cpuhp(hctx);
2186         return -1;
2187 }
2188
2189 static void blk_mq_init_cpu_queues(struct request_queue *q,
2190                                    unsigned int nr_hw_queues)
2191 {
2192         unsigned int i;
2193
2194         for_each_possible_cpu(i) {
2195                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2196                 struct blk_mq_hw_ctx *hctx;
2197
2198                 __ctx->cpu = i;
2199                 spin_lock_init(&__ctx->lock);
2200                 INIT_LIST_HEAD(&__ctx->rq_list);
2201                 __ctx->queue = q;
2202
2203                 /*
2204                  * Set local node, IFF we have more than one hw queue. If
2205                  * not, we remain on the home node of the device
2206                  */
2207                 hctx = blk_mq_map_queue(q, i);
2208                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2209                         hctx->numa_node = local_memory_node(cpu_to_node(i));
2210         }
2211 }
2212
2213 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2214 {
2215         int ret = 0;
2216
2217         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2218                                         set->queue_depth, set->reserved_tags);
2219         if (!set->tags[hctx_idx])
2220                 return false;
2221
2222         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2223                                 set->queue_depth);
2224         if (!ret)
2225                 return true;
2226
2227         blk_mq_free_rq_map(set->tags[hctx_idx]);
2228         set->tags[hctx_idx] = NULL;
2229         return false;
2230 }
2231
2232 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2233                                          unsigned int hctx_idx)
2234 {
2235         if (set->tags[hctx_idx]) {
2236                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2237                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2238                 set->tags[hctx_idx] = NULL;
2239         }
2240 }
2241
2242 static void blk_mq_map_swqueue(struct request_queue *q)
2243 {
2244         unsigned int i, hctx_idx;
2245         struct blk_mq_hw_ctx *hctx;
2246         struct blk_mq_ctx *ctx;
2247         struct blk_mq_tag_set *set = q->tag_set;
2248
2249         /*
2250          * Avoid others reading imcomplete hctx->cpumask through sysfs
2251          */
2252         mutex_lock(&q->sysfs_lock);
2253
2254         queue_for_each_hw_ctx(q, hctx, i) {
2255                 cpumask_clear(hctx->cpumask);
2256                 hctx->nr_ctx = 0;
2257                 hctx->dispatch_from = NULL;
2258         }
2259
2260         /*
2261          * Map software to hardware queues.
2262          *
2263          * If the cpu isn't present, the cpu is mapped to first hctx.
2264          */
2265         for_each_possible_cpu(i) {
2266                 hctx_idx = q->mq_map[i];
2267                 /* unmapped hw queue can be remapped after CPU topo changed */
2268                 if (!set->tags[hctx_idx] &&
2269                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2270                         /*
2271                          * If tags initialization fail for some hctx,
2272                          * that hctx won't be brought online.  In this
2273                          * case, remap the current ctx to hctx[0] which
2274                          * is guaranteed to always have tags allocated
2275                          */
2276                         q->mq_map[i] = 0;
2277                 }
2278
2279                 ctx = per_cpu_ptr(q->queue_ctx, i);
2280                 hctx = blk_mq_map_queue(q, i);
2281
2282                 cpumask_set_cpu(i, hctx->cpumask);
2283                 ctx->index_hw = hctx->nr_ctx;
2284                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2285         }
2286
2287         mutex_unlock(&q->sysfs_lock);
2288
2289         queue_for_each_hw_ctx(q, hctx, i) {
2290                 /*
2291                  * If no software queues are mapped to this hardware queue,
2292                  * disable it and free the request entries.
2293                  */
2294                 if (!hctx->nr_ctx) {
2295                         /* Never unmap queue 0.  We need it as a
2296                          * fallback in case of a new remap fails
2297                          * allocation
2298                          */
2299                         if (i && set->tags[i])
2300                                 blk_mq_free_map_and_requests(set, i);
2301
2302                         hctx->tags = NULL;
2303                         continue;
2304                 }
2305
2306                 hctx->tags = set->tags[i];
2307                 WARN_ON(!hctx->tags);
2308
2309                 /*
2310                  * Set the map size to the number of mapped software queues.
2311                  * This is more accurate and more efficient than looping
2312                  * over all possibly mapped software queues.
2313                  */
2314                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2315
2316                 /*
2317                  * Initialize batch roundrobin counts
2318                  */
2319                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2320                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2321         }
2322 }
2323
2324 /*
2325  * Caller needs to ensure that we're either frozen/quiesced, or that
2326  * the queue isn't live yet.
2327  */
2328 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2329 {
2330         struct blk_mq_hw_ctx *hctx;
2331         int i;
2332
2333         queue_for_each_hw_ctx(q, hctx, i) {
2334                 if (shared) {
2335                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2336                                 atomic_inc(&q->shared_hctx_restart);
2337                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2338                 } else {
2339                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2340                                 atomic_dec(&q->shared_hctx_restart);
2341                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2342                 }
2343         }
2344 }
2345
2346 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2347                                         bool shared)
2348 {
2349         struct request_queue *q;
2350
2351         lockdep_assert_held(&set->tag_list_lock);
2352
2353         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2354                 blk_mq_freeze_queue(q);
2355                 queue_set_hctx_shared(q, shared);
2356                 blk_mq_unfreeze_queue(q);
2357         }
2358 }
2359
2360 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2361 {
2362         struct blk_mq_tag_set *set = q->tag_set;
2363
2364         mutex_lock(&set->tag_list_lock);
2365         list_del_rcu(&q->tag_set_list);
2366         if (list_is_singular(&set->tag_list)) {
2367                 /* just transitioned to unshared */
2368                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2369                 /* update existing queue */
2370                 blk_mq_update_tag_set_depth(set, false);
2371         }
2372         mutex_unlock(&set->tag_list_lock);
2373         synchronize_rcu();
2374         INIT_LIST_HEAD(&q->tag_set_list);
2375 }
2376
2377 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2378                                      struct request_queue *q)
2379 {
2380         q->tag_set = set;
2381
2382         mutex_lock(&set->tag_list_lock);
2383
2384         /*
2385          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2386          */
2387         if (!list_empty(&set->tag_list) &&
2388             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2389                 set->flags |= BLK_MQ_F_TAG_SHARED;
2390                 /* update existing queue */
2391                 blk_mq_update_tag_set_depth(set, true);
2392         }
2393         if (set->flags & BLK_MQ_F_TAG_SHARED)
2394                 queue_set_hctx_shared(q, true);
2395         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2396
2397         mutex_unlock(&set->tag_list_lock);
2398 }
2399
2400 /*
2401  * It is the actual release handler for mq, but we do it from
2402  * request queue's release handler for avoiding use-after-free
2403  * and headache because q->mq_kobj shouldn't have been introduced,
2404  * but we can't group ctx/kctx kobj without it.
2405  */
2406 void blk_mq_release(struct request_queue *q)
2407 {
2408         struct blk_mq_hw_ctx *hctx;
2409         unsigned int i;
2410
2411         /* hctx kobj stays in hctx */
2412         queue_for_each_hw_ctx(q, hctx, i) {
2413                 if (!hctx)
2414                         continue;
2415                 kobject_put(&hctx->kobj);
2416         }
2417
2418         q->mq_map = NULL;
2419
2420         kfree(q->queue_hw_ctx);
2421
2422         /*
2423          * release .mq_kobj and sw queue's kobject now because
2424          * both share lifetime with request queue.
2425          */
2426         blk_mq_sysfs_deinit(q);
2427
2428         free_percpu(q->queue_ctx);
2429 }
2430
2431 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2432 {
2433         struct request_queue *uninit_q, *q;
2434
2435         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2436         if (!uninit_q)
2437                 return ERR_PTR(-ENOMEM);
2438
2439         q = blk_mq_init_allocated_queue(set, uninit_q);
2440         if (IS_ERR(q))
2441                 blk_cleanup_queue(uninit_q);
2442
2443         return q;
2444 }
2445 EXPORT_SYMBOL(blk_mq_init_queue);
2446
2447 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2448 {
2449         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2450
2451         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2452                            __alignof__(struct blk_mq_hw_ctx)) !=
2453                      sizeof(struct blk_mq_hw_ctx));
2454
2455         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2456                 hw_ctx_size += sizeof(struct srcu_struct);
2457
2458         return hw_ctx_size;
2459 }
2460
2461 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2462                                                 struct request_queue *q)
2463 {
2464         int i, j;
2465         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2466
2467         blk_mq_sysfs_unregister(q);
2468
2469         /* protect against switching io scheduler  */
2470         mutex_lock(&q->sysfs_lock);
2471         for (i = 0; i < set->nr_hw_queues; i++) {
2472                 int node;
2473
2474                 if (hctxs[i])
2475                         continue;
2476
2477                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2478                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2479                                         GFP_KERNEL, node);
2480                 if (!hctxs[i])
2481                         break;
2482
2483                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2484                                                 node)) {
2485                         kfree(hctxs[i]);
2486                         hctxs[i] = NULL;
2487                         break;
2488                 }
2489
2490                 atomic_set(&hctxs[i]->nr_active, 0);
2491                 hctxs[i]->numa_node = node;
2492                 hctxs[i]->queue_num = i;
2493
2494                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2495                         free_cpumask_var(hctxs[i]->cpumask);
2496                         kfree(hctxs[i]);
2497                         hctxs[i] = NULL;
2498                         break;
2499                 }
2500                 blk_mq_hctx_kobj_init(hctxs[i]);
2501         }
2502         for (j = i; j < q->nr_hw_queues; j++) {
2503                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2504
2505                 if (hctx) {
2506                         if (hctx->tags)
2507                                 blk_mq_free_map_and_requests(set, j);
2508                         blk_mq_exit_hctx(q, set, hctx, j);
2509                         kobject_put(&hctx->kobj);
2510                         hctxs[j] = NULL;
2511
2512                 }
2513         }
2514         q->nr_hw_queues = i;
2515         mutex_unlock(&q->sysfs_lock);
2516         blk_mq_sysfs_register(q);
2517 }
2518
2519 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2520                                                   struct request_queue *q)
2521 {
2522         /* mark the queue as mq asap */
2523         q->mq_ops = set->ops;
2524
2525         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2526                                              blk_mq_poll_stats_bkt,
2527                                              BLK_MQ_POLL_STATS_BKTS, q);
2528         if (!q->poll_cb)
2529                 goto err_exit;
2530
2531         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2532         if (!q->queue_ctx)
2533                 goto err_exit;
2534
2535         /* init q->mq_kobj and sw queues' kobjects */
2536         blk_mq_sysfs_init(q);
2537
2538         q->queue_hw_ctx = kcalloc_node(nr_cpu_ids, sizeof(*(q->queue_hw_ctx)),
2539                                                 GFP_KERNEL, set->numa_node);
2540         if (!q->queue_hw_ctx)
2541                 goto err_percpu;
2542
2543         q->mq_map = set->mq_map;
2544
2545         blk_mq_realloc_hw_ctxs(set, q);
2546         if (!q->nr_hw_queues)
2547                 goto err_hctxs;
2548
2549         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2550         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2551
2552         q->nr_queues = nr_cpu_ids;
2553
2554         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2555
2556         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2557                 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2558
2559         q->sg_reserved_size = INT_MAX;
2560
2561         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2562         INIT_LIST_HEAD(&q->requeue_list);
2563         spin_lock_init(&q->requeue_lock);
2564
2565         blk_queue_make_request(q, blk_mq_make_request);
2566         if (q->mq_ops->poll)
2567                 q->poll_fn = blk_mq_poll;
2568
2569         /*
2570          * Do this after blk_queue_make_request() overrides it...
2571          */
2572         q->nr_requests = set->queue_depth;
2573
2574         /*
2575          * Default to classic polling
2576          */
2577         q->poll_nsec = -1;
2578
2579         if (set->ops->complete)
2580                 blk_queue_softirq_done(q, set->ops->complete);
2581
2582         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2583         blk_mq_add_queue_tag_set(set, q);
2584         blk_mq_map_swqueue(q);
2585
2586         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2587                 int ret;
2588
2589                 ret = elevator_init_mq(q);
2590                 if (ret)
2591                         return ERR_PTR(ret);
2592         }
2593
2594         return q;
2595
2596 err_hctxs:
2597         kfree(q->queue_hw_ctx);
2598 err_percpu:
2599         free_percpu(q->queue_ctx);
2600 err_exit:
2601         q->mq_ops = NULL;
2602         return ERR_PTR(-ENOMEM);
2603 }
2604 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2605
2606 void blk_mq_free_queue(struct request_queue *q)
2607 {
2608         struct blk_mq_tag_set   *set = q->tag_set;
2609
2610         blk_mq_del_queue_tag_set(q);
2611         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2612 }
2613
2614 /* Basically redo blk_mq_init_queue with queue frozen */
2615 static void blk_mq_queue_reinit(struct request_queue *q)
2616 {
2617         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2618
2619         blk_mq_debugfs_unregister_hctxs(q);
2620         blk_mq_sysfs_unregister(q);
2621
2622         /*
2623          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2624          * we should change hctx numa_node according to the new topology (this
2625          * involves freeing and re-allocating memory, worth doing?)
2626          */
2627         blk_mq_map_swqueue(q);
2628
2629         blk_mq_sysfs_register(q);
2630         blk_mq_debugfs_register_hctxs(q);
2631 }
2632
2633 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2634 {
2635         int i;
2636
2637         for (i = 0; i < set->nr_hw_queues; i++)
2638                 if (!__blk_mq_alloc_rq_map(set, i))
2639                         goto out_unwind;
2640
2641         return 0;
2642
2643 out_unwind:
2644         while (--i >= 0)
2645                 blk_mq_free_rq_map(set->tags[i]);
2646
2647         return -ENOMEM;
2648 }
2649
2650 /*
2651  * Allocate the request maps associated with this tag_set. Note that this
2652  * may reduce the depth asked for, if memory is tight. set->queue_depth
2653  * will be updated to reflect the allocated depth.
2654  */
2655 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2656 {
2657         unsigned int depth;
2658         int err;
2659
2660         depth = set->queue_depth;
2661         do {
2662                 err = __blk_mq_alloc_rq_maps(set);
2663                 if (!err)
2664                         break;
2665
2666                 set->queue_depth >>= 1;
2667                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2668                         err = -ENOMEM;
2669                         break;
2670                 }
2671         } while (set->queue_depth);
2672
2673         if (!set->queue_depth || err) {
2674                 pr_err("blk-mq: failed to allocate request map\n");
2675                 return -ENOMEM;
2676         }
2677
2678         if (depth != set->queue_depth)
2679                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2680                                                 depth, set->queue_depth);
2681
2682         return 0;
2683 }
2684
2685 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2686 {
2687         if (set->ops->map_queues) {
2688                 int cpu;
2689                 /*
2690                  * transport .map_queues is usually done in the following
2691                  * way:
2692                  *
2693                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2694                  *      mask = get_cpu_mask(queue)
2695                  *      for_each_cpu(cpu, mask)
2696                  *              set->mq_map[cpu] = queue;
2697                  * }
2698                  *
2699                  * When we need to remap, the table has to be cleared for
2700                  * killing stale mapping since one CPU may not be mapped
2701                  * to any hw queue.
2702                  */
2703                 for_each_possible_cpu(cpu)
2704                         set->mq_map[cpu] = 0;
2705
2706                 return set->ops->map_queues(set);
2707         } else
2708                 return blk_mq_map_queues(set);
2709 }
2710
2711 /*
2712  * Alloc a tag set to be associated with one or more request queues.
2713  * May fail with EINVAL for various error conditions. May adjust the
2714  * requested depth down, if if it too large. In that case, the set
2715  * value will be stored in set->queue_depth.
2716  */
2717 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2718 {
2719         int ret;
2720
2721         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2722
2723         if (!set->nr_hw_queues)
2724                 return -EINVAL;
2725         if (!set->queue_depth)
2726                 return -EINVAL;
2727         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2728                 return -EINVAL;
2729
2730         if (!set->ops->queue_rq)
2731                 return -EINVAL;
2732
2733         if (!set->ops->get_budget ^ !set->ops->put_budget)
2734                 return -EINVAL;
2735
2736         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2737                 pr_info("blk-mq: reduced tag depth to %u\n",
2738                         BLK_MQ_MAX_DEPTH);
2739                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2740         }
2741
2742         /*
2743          * If a crashdump is active, then we are potentially in a very
2744          * memory constrained environment. Limit us to 1 queue and
2745          * 64 tags to prevent using too much memory.
2746          */
2747         if (is_kdump_kernel()) {
2748                 set->nr_hw_queues = 1;
2749                 set->queue_depth = min(64U, set->queue_depth);
2750         }
2751         /*
2752          * There is no use for more h/w queues than cpus.
2753          */
2754         if (set->nr_hw_queues > nr_cpu_ids)
2755                 set->nr_hw_queues = nr_cpu_ids;
2756
2757         set->tags = kcalloc_node(nr_cpu_ids, sizeof(struct blk_mq_tags *),
2758                                  GFP_KERNEL, set->numa_node);
2759         if (!set->tags)
2760                 return -ENOMEM;
2761
2762         ret = -ENOMEM;
2763         set->mq_map = kcalloc_node(nr_cpu_ids, sizeof(*set->mq_map),
2764                                    GFP_KERNEL, set->numa_node);
2765         if (!set->mq_map)
2766                 goto out_free_tags;
2767
2768         ret = blk_mq_update_queue_map(set);
2769         if (ret)
2770                 goto out_free_mq_map;
2771
2772         ret = blk_mq_alloc_rq_maps(set);
2773         if (ret)
2774                 goto out_free_mq_map;
2775
2776         mutex_init(&set->tag_list_lock);
2777         INIT_LIST_HEAD(&set->tag_list);
2778
2779         return 0;
2780
2781 out_free_mq_map:
2782         kfree(set->mq_map);
2783         set->mq_map = NULL;
2784 out_free_tags:
2785         kfree(set->tags);
2786         set->tags = NULL;
2787         return ret;
2788 }
2789 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2790
2791 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2792 {
2793         int i;
2794
2795         for (i = 0; i < nr_cpu_ids; i++)
2796                 blk_mq_free_map_and_requests(set, i);
2797
2798         kfree(set->mq_map);
2799         set->mq_map = NULL;
2800
2801         kfree(set->tags);
2802         set->tags = NULL;
2803 }
2804 EXPORT_SYMBOL(blk_mq_free_tag_set);
2805
2806 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2807 {
2808         struct blk_mq_tag_set *set = q->tag_set;
2809         struct blk_mq_hw_ctx *hctx;
2810         int i, ret;
2811
2812         if (!set)
2813                 return -EINVAL;
2814
2815         blk_mq_freeze_queue(q);
2816         blk_mq_quiesce_queue(q);
2817
2818         ret = 0;
2819         queue_for_each_hw_ctx(q, hctx, i) {
2820                 if (!hctx->tags)
2821                         continue;
2822                 /*
2823                  * If we're using an MQ scheduler, just update the scheduler
2824                  * queue depth. This is similar to what the old code would do.
2825                  */
2826                 if (!hctx->sched_tags) {
2827                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2828                                                         false);
2829                 } else {
2830                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2831                                                         nr, true);
2832                 }
2833                 if (ret)
2834                         break;
2835         }
2836
2837         if (!ret)
2838                 q->nr_requests = nr;
2839
2840         blk_mq_unquiesce_queue(q);
2841         blk_mq_unfreeze_queue(q);
2842
2843         return ret;
2844 }
2845
2846 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2847                                                         int nr_hw_queues)
2848 {
2849         struct request_queue *q;
2850
2851         lockdep_assert_held(&set->tag_list_lock);
2852
2853         if (nr_hw_queues > nr_cpu_ids)
2854                 nr_hw_queues = nr_cpu_ids;
2855         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2856                 return;
2857
2858         list_for_each_entry(q, &set->tag_list, tag_set_list)
2859                 blk_mq_freeze_queue(q);
2860
2861         set->nr_hw_queues = nr_hw_queues;
2862         blk_mq_update_queue_map(set);
2863         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2864                 blk_mq_realloc_hw_ctxs(set, q);
2865                 blk_mq_queue_reinit(q);
2866         }
2867
2868         list_for_each_entry(q, &set->tag_list, tag_set_list)
2869                 blk_mq_unfreeze_queue(q);
2870 }
2871
2872 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2873 {
2874         mutex_lock(&set->tag_list_lock);
2875         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2876         mutex_unlock(&set->tag_list_lock);
2877 }
2878 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2879
2880 /* Enable polling stats and return whether they were already enabled. */
2881 static bool blk_poll_stats_enable(struct request_queue *q)
2882 {
2883         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2884             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
2885                 return true;
2886         blk_stat_add_callback(q, q->poll_cb);
2887         return false;
2888 }
2889
2890 static void blk_mq_poll_stats_start(struct request_queue *q)
2891 {
2892         /*
2893          * We don't arm the callback if polling stats are not enabled or the
2894          * callback is already active.
2895          */
2896         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2897             blk_stat_is_active(q->poll_cb))
2898                 return;
2899
2900         blk_stat_activate_msecs(q->poll_cb, 100);
2901 }
2902
2903 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2904 {
2905         struct request_queue *q = cb->data;
2906         int bucket;
2907
2908         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2909                 if (cb->stat[bucket].nr_samples)
2910                         q->poll_stat[bucket] = cb->stat[bucket];
2911         }
2912 }
2913
2914 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2915                                        struct blk_mq_hw_ctx *hctx,
2916                                        struct request *rq)
2917 {
2918         unsigned long ret = 0;
2919         int bucket;
2920
2921         /*
2922          * If stats collection isn't on, don't sleep but turn it on for
2923          * future users
2924          */
2925         if (!blk_poll_stats_enable(q))
2926                 return 0;
2927
2928         /*
2929          * As an optimistic guess, use half of the mean service time
2930          * for this type of request. We can (and should) make this smarter.
2931          * For instance, if the completion latencies are tight, we can
2932          * get closer than just half the mean. This is especially
2933          * important on devices where the completion latencies are longer
2934          * than ~10 usec. We do use the stats for the relevant IO size
2935          * if available which does lead to better estimates.
2936          */
2937         bucket = blk_mq_poll_stats_bkt(rq);
2938         if (bucket < 0)
2939                 return ret;
2940
2941         if (q->poll_stat[bucket].nr_samples)
2942                 ret = (q->poll_stat[bucket].mean + 1) / 2;
2943
2944         return ret;
2945 }
2946
2947 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2948                                      struct blk_mq_hw_ctx *hctx,
2949                                      struct request *rq)
2950 {
2951         struct hrtimer_sleeper hs;
2952         enum hrtimer_mode mode;
2953         unsigned int nsecs;
2954         ktime_t kt;
2955
2956         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
2957                 return false;
2958
2959         /*
2960          * poll_nsec can be:
2961          *
2962          * -1:  don't ever hybrid sleep
2963          *  0:  use half of prev avg
2964          * >0:  use this specific value
2965          */
2966         if (q->poll_nsec == -1)
2967                 return false;
2968         else if (q->poll_nsec > 0)
2969                 nsecs = q->poll_nsec;
2970         else
2971                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2972
2973         if (!nsecs)
2974                 return false;
2975
2976         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
2977
2978         /*
2979          * This will be replaced with the stats tracking code, using
2980          * 'avg_completion_time / 2' as the pre-sleep target.
2981          */
2982         kt = nsecs;
2983
2984         mode = HRTIMER_MODE_REL;
2985         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2986         hrtimer_set_expires(&hs.timer, kt);
2987
2988         hrtimer_init_sleeper(&hs, current);
2989         do {
2990                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
2991                         break;
2992                 set_current_state(TASK_UNINTERRUPTIBLE);
2993                 hrtimer_start_expires(&hs.timer, mode);
2994                 if (hs.task)
2995                         io_schedule();
2996                 hrtimer_cancel(&hs.timer);
2997                 mode = HRTIMER_MODE_ABS;
2998         } while (hs.task && !signal_pending(current));
2999
3000         __set_current_state(TASK_RUNNING);
3001         destroy_hrtimer_on_stack(&hs.timer);
3002         return true;
3003 }
3004
3005 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3006 {
3007         struct request_queue *q = hctx->queue;
3008         long state;
3009
3010         /*
3011          * If we sleep, have the caller restart the poll loop to reset
3012          * the state. Like for the other success return cases, the
3013          * caller is responsible for checking if the IO completed. If
3014          * the IO isn't complete, we'll get called again and will go
3015          * straight to the busy poll loop.
3016          */
3017         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3018                 return true;
3019
3020         hctx->poll_considered++;
3021
3022         state = current->state;
3023         while (!need_resched()) {
3024                 int ret;
3025
3026                 hctx->poll_invoked++;
3027
3028                 ret = q->mq_ops->poll(hctx, rq->tag);
3029                 if (ret > 0) {
3030                         hctx->poll_success++;
3031                         set_current_state(TASK_RUNNING);
3032                         return true;
3033                 }
3034
3035                 if (signal_pending_state(state, current))
3036                         set_current_state(TASK_RUNNING);
3037
3038                 if (current->state == TASK_RUNNING)
3039                         return true;
3040                 if (ret < 0)
3041                         break;
3042                 cpu_relax();
3043         }
3044
3045         __set_current_state(TASK_RUNNING);
3046         return false;
3047 }
3048
3049 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3050 {
3051         struct blk_mq_hw_ctx *hctx;
3052         struct request *rq;
3053
3054         if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3055                 return false;
3056
3057         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3058         if (!blk_qc_t_is_internal(cookie))
3059                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3060         else {
3061                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3062                 /*
3063                  * With scheduling, if the request has completed, we'll
3064                  * get a NULL return here, as we clear the sched tag when
3065                  * that happens. The request still remains valid, like always,
3066                  * so we should be safe with just the NULL check.
3067                  */
3068                 if (!rq)
3069                         return false;
3070         }
3071
3072         return __blk_mq_poll(hctx, rq);
3073 }
3074
3075 static int __init blk_mq_init(void)
3076 {
3077         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3078                                 blk_mq_hctx_notify_dead);
3079         return 0;
3080 }
3081 subsys_initcall(blk_mq_init);