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