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