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