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/mm.h>
  13 #include <linux/init.h>
  14 #include <linux/slab.h>
  15 #include <linux/workqueue.h>
  16 #include <linux/smp.h>
  17 #include <linux/llist.h>
  18 #include <linux/list_sort.h>
  19 #include <linux/cpu.h>
  20 #include <linux/cache.h>
  21 #include <linux/sched/sysctl.h>
  22 #include <linux/delay.h>
  23 #include <linux/crash_dump.h>
  24
  25 #include <trace/events/block.h>
  26
  27 #include <linux/blk-mq.h>
  28 #include "blk.h"
  29 #include "blk-mq.h"
  30 #include "blk-mq-tag.h"
  31
  32 static DEFINE_MUTEX(all_q_mutex);
  33 static LIST_HEAD(all_q_list);
  34
  35 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
  36
  37 /*
  38  * Check if any of the ctx's have pending work in this hardware queue
  39  */
  40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  41 {
  42         unsigned int i;
  43
  44         for (i = 0; i < hctx->ctx_map.map_size; i++)
  45                 if (hctx->ctx_map.map[i].word)
  46                         return true;
  47
  48         return false;
  49 }
  50
  51 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
  52                                               struct blk_mq_ctx *ctx)
  53 {
  54         return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
  55 }
  56
  57 #define CTX_TO_BIT(hctx, ctx)   \
  58         ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
  59
  60 /*
  61  * Mark this ctx as having pending work in this hardware queue
  62  */
  63 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  64                                      struct blk_mq_ctx *ctx)
  65 {
  66         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  67
  68         if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
  69                 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  70 }
  71
  72 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  73                                       struct blk_mq_ctx *ctx)
  74 {
  75         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  76
  77         clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  78 }
  79
  80 static int blk_mq_queue_enter(struct request_queue *q)
  81 {
  82         while (true) {
  83                 int ret;
  84
  85                 if (percpu_ref_tryget_live(&q->mq_usage_counter))
  86                         return 0;
  87
  88                 ret = wait_event_interruptible(q->mq_freeze_wq,
  89                                 !q->mq_freeze_depth || blk_queue_dying(q));
  90                 if (blk_queue_dying(q))
  91                         return -ENODEV;
  92                 if (ret)
  93                         return ret;
  94         }
  95 }
  96
  97 static void blk_mq_queue_exit(struct request_queue *q)
  98 {
  99         percpu_ref_put(&q->mq_usage_counter);
 100 }
 101
 102 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
 103 {
 104         struct request_queue *q =
 105                 container_of(ref, struct request_queue, mq_usage_counter);
 106
 107         wake_up_all(&q->mq_freeze_wq);
 108 }
 109
 110 /*
 111  * Guarantee no request is in use, so we can change any data structure of
 112  * the queue afterward.
 113  */
 114 void blk_mq_freeze_queue(struct request_queue *q)
 115 {
 116         bool freeze;
 117
 118         spin_lock_irq(q->queue_lock);
 119         freeze = !q->mq_freeze_depth++;
 120         spin_unlock_irq(q->queue_lock);
 121
 122         if (freeze) {
 123                 percpu_ref_kill(&q->mq_usage_counter);
 124                 blk_mq_run_queues(q, false);
 125         }
 126         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
 127 }
 128
 129 static void blk_mq_unfreeze_queue(struct request_queue *q)
 130 {
 131         bool wake;
 132
 133         spin_lock_irq(q->queue_lock);
 134         wake = !--q->mq_freeze_depth;
 135         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 136         spin_unlock_irq(q->queue_lock);
 137         if (wake) {
 138                 percpu_ref_reinit(&q->mq_usage_counter);
 139                 wake_up_all(&q->mq_freeze_wq);
 140         }
 141 }
 142
 143 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 144 {
 145         return blk_mq_has_free_tags(hctx->tags);
 146 }
 147 EXPORT_SYMBOL(blk_mq_can_queue);
 148
 149 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 150                                struct request *rq, unsigned int rw_flags)
 151 {
 152         if (blk_queue_io_stat(q))
 153                 rw_flags |= REQ_IO_STAT;
 154
 155         INIT_LIST_HEAD(&rq->queuelist);
 156         /* csd/requeue_work/fifo_time is initialized before use */
 157         rq->q = q;
 158         rq->mq_ctx = ctx;
 159         rq->cmd_flags |= rw_flags;
 160         /* do not touch atomic flags, it needs atomic ops against the timer */
 161         rq->cpu = -1;
 162         INIT_HLIST_NODE(&rq->hash);
 163         RB_CLEAR_NODE(&rq->rb_node);
 164         rq->rq_disk = NULL;
 165         rq->part = NULL;
 166         rq->start_time = jiffies;
 167 #ifdef CONFIG_BLK_CGROUP
 168         rq->rl = NULL;
 169         set_start_time_ns(rq);
 170         rq->io_start_time_ns = 0;
 171 #endif
 172         rq->nr_phys_segments = 0;
 173 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 174         rq->nr_integrity_segments = 0;
 175 #endif
 176         rq->special = NULL;
 177         /* tag was already set */
 178         rq->errors = 0;
 179
 180         rq->cmd = rq->__cmd;
 181
 182         rq->extra_len = 0;
 183         rq->sense_len = 0;
 184         rq->resid_len = 0;
 185         rq->sense = NULL;
 186
 187         INIT_LIST_HEAD(&rq->timeout_list);
 188         rq->timeout = 0;
 189
 190         rq->end_io = NULL;
 191         rq->end_io_data = NULL;
 192         rq->next_rq = NULL;
 193
 194         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 195 }
 196
 197 static struct request *
 198 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 199 {
 200         struct request *rq;
 201         unsigned int tag;
 202
 203         tag = blk_mq_get_tag(data);
 204         if (tag != BLK_MQ_TAG_FAIL) {
 205                 rq = data->hctx->tags->rqs[tag];
 206
 207                 if (blk_mq_tag_busy(data->hctx)) {
 208                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 209                         atomic_inc(&data->hctx->nr_active);
 210                 }
 211
 212                 rq->tag = tag;
 213                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 214                 return rq;
 215         }
 216
 217         return NULL;
 218 }
 219
 220 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 221                 bool reserved)
 222 {
 223         struct blk_mq_ctx *ctx;
 224         struct blk_mq_hw_ctx *hctx;
 225         struct request *rq;
 226         struct blk_mq_alloc_data alloc_data;
 227         int ret;
 228
 229         ret = blk_mq_queue_enter(q);
 230         if (ret)
 231                 return ERR_PTR(ret);
 232
 233         ctx = blk_mq_get_ctx(q);
 234         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 235         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 236                         reserved, ctx, hctx);
 237
 238         rq = __blk_mq_alloc_request(&alloc_data, rw);
 239         if (!rq && (gfp & __GFP_WAIT)) {
 240                 __blk_mq_run_hw_queue(hctx);
 241                 blk_mq_put_ctx(ctx);
 242
 243                 ctx = blk_mq_get_ctx(q);
 244                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 245                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 246                                 hctx);
 247                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 248                 ctx = alloc_data.ctx;
 249         }
 250         blk_mq_put_ctx(ctx);
 251         if (!rq)
 252                 return ERR_PTR(-EWOULDBLOCK);
 253         return rq;
 254 }
 255 EXPORT_SYMBOL(blk_mq_alloc_request);
 256
 257 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 258                                   struct blk_mq_ctx *ctx, struct request *rq)
 259 {
 260         const int tag = rq->tag;
 261         struct request_queue *q = rq->q;
 262
 263         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 264                 atomic_dec(&hctx->nr_active);
 265         rq->cmd_flags = 0;
 266
 267         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 268         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 269         blk_mq_queue_exit(q);
 270 }
 271
 272 void blk_mq_free_request(struct request *rq)
 273 {
 274         struct blk_mq_ctx *ctx = rq->mq_ctx;
 275         struct blk_mq_hw_ctx *hctx;
 276         struct request_queue *q = rq->q;
 277
 278         ctx->rq_completed[rq_is_sync(rq)]++;
 279
 280         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 281         __blk_mq_free_request(hctx, ctx, rq);
 282 }
 283
 284 /*
 285  * Clone all relevant state from a request that has been put on hold in
 286  * the flush state machine into the preallocated flush request that hangs
 287  * off the request queue.
 288  *
 289  * For a driver the flush request should be invisible, that's why we are
 290  * impersonating the original request here.
 291  */
 292 void blk_mq_clone_flush_request(struct request *flush_rq,
 293                 struct request *orig_rq)
 294 {
 295         struct blk_mq_hw_ctx *hctx =
 296                 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
 297
 298         flush_rq->mq_ctx = orig_rq->mq_ctx;
 299         flush_rq->tag = orig_rq->tag;
 300         memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
 301                 hctx->cmd_size);
 302 }
 303
 304 inline void __blk_mq_end_request(struct request *rq, int error)
 305 {
 306         blk_account_io_done(rq);
 307
 308         if (rq->end_io) {
 309                 rq->end_io(rq, error);
 310         } else {
 311                 if (unlikely(blk_bidi_rq(rq)))
 312                         blk_mq_free_request(rq->next_rq);
 313                 blk_mq_free_request(rq);
 314         }
 315 }
 316 EXPORT_SYMBOL(__blk_mq_end_request);
 317
 318 void blk_mq_end_request(struct request *rq, int error)
 319 {
 320         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 321                 BUG();
 322         __blk_mq_end_request(rq, error);
 323 }
 324 EXPORT_SYMBOL(blk_mq_end_request);
 325
 326 static void __blk_mq_complete_request_remote(void *data)
 327 {
 328         struct request *rq = data;
 329
 330         rq->q->softirq_done_fn(rq);
 331 }
 332
 333 static void blk_mq_ipi_complete_request(struct request *rq)
 334 {
 335         struct blk_mq_ctx *ctx = rq->mq_ctx;
 336         bool shared = false;
 337         int cpu;
 338
 339         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 340                 rq->q->softirq_done_fn(rq);
 341                 return;
 342         }
 343
 344         cpu = get_cpu();
 345         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 346                 shared = cpus_share_cache(cpu, ctx->cpu);
 347
 348         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 349                 rq->csd.func = __blk_mq_complete_request_remote;
 350                 rq->csd.info = rq;
 351                 rq->csd.flags = 0;
 352                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 353         } else {
 354                 rq->q->softirq_done_fn(rq);
 355         }
 356         put_cpu();
 357 }
 358
 359 void __blk_mq_complete_request(struct request *rq)
 360 {
 361         struct request_queue *q = rq->q;
 362
 363         if (!q->softirq_done_fn)
 364                 blk_mq_end_request(rq, rq->errors);
 365         else
 366                 blk_mq_ipi_complete_request(rq);
 367 }
 368
 369 /**
 370  * blk_mq_complete_request - end I/O on a request
 371  * @rq:         the request being processed
 372  *
 373  * Description:
 374  *      Ends all I/O on a request. It does not handle partial completions.
 375  *      The actual completion happens out-of-order, through a IPI handler.
 376  **/
 377 void blk_mq_complete_request(struct request *rq)
 378 {
 379         struct request_queue *q = rq->q;
 380
 381         if (unlikely(blk_should_fake_timeout(q)))
 382                 return;
 383         if (!blk_mark_rq_complete(rq))
 384                 __blk_mq_complete_request(rq);
 385 }
 386 EXPORT_SYMBOL(blk_mq_complete_request);
 387
 388 void blk_mq_start_request(struct request *rq)
 389 {
 390         struct request_queue *q = rq->q;
 391
 392         trace_block_rq_issue(q, rq);
 393
 394         rq->resid_len = blk_rq_bytes(rq);
 395         if (unlikely(blk_bidi_rq(rq)))
 396                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 397
 398         blk_add_timer(rq);
 399
 400         /*
 401          * Ensure that ->deadline is visible before set the started
 402          * flag and clear the completed flag.
 403          */
 404         smp_mb__before_atomic();
 405
 406         /*
 407          * Mark us as started and clear complete. Complete might have been
 408          * set if requeue raced with timeout, which then marked it as
 409          * complete. So be sure to clear complete again when we start
 410          * the request, otherwise we'll ignore the completion event.
 411          */
 412         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 413                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 414         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 415                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 416
 417         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 418                 /*
 419                  * Make sure space for the drain appears.  We know we can do
 420                  * this because max_hw_segments has been adjusted to be one
 421                  * fewer than the device can handle.
 422                  */
 423                 rq->nr_phys_segments++;
 424         }
 425 }
 426 EXPORT_SYMBOL(blk_mq_start_request);
 427
 428 static void __blk_mq_requeue_request(struct request *rq)
 429 {
 430         struct request_queue *q = rq->q;
 431
 432         trace_block_rq_requeue(q, rq);
 433
 434         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 435                 if (q->dma_drain_size && blk_rq_bytes(rq))
 436                         rq->nr_phys_segments--;
 437         }
 438 }
 439
 440 void blk_mq_requeue_request(struct request *rq)
 441 {
 442         __blk_mq_requeue_request(rq);
 443
 444         BUG_ON(blk_queued_rq(rq));
 445         blk_mq_add_to_requeue_list(rq, true);
 446 }
 447 EXPORT_SYMBOL(blk_mq_requeue_request);
 448
 449 static void blk_mq_requeue_work(struct work_struct *work)
 450 {
 451         struct request_queue *q =
 452                 container_of(work, struct request_queue, requeue_work);
 453         LIST_HEAD(rq_list);
 454         struct request *rq, *next;
 455         unsigned long flags;
 456
 457         spin_lock_irqsave(&q->requeue_lock, flags);
 458         list_splice_init(&q->requeue_list, &rq_list);
 459         spin_unlock_irqrestore(&q->requeue_lock, flags);
 460
 461         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 462                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 463                         continue;
 464
 465                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 466                 list_del_init(&rq->queuelist);
 467                 blk_mq_insert_request(rq, true, false, false);
 468         }
 469
 470         while (!list_empty(&rq_list)) {
 471                 rq = list_entry(rq_list.next, struct request, queuelist);
 472                 list_del_init(&rq->queuelist);
 473                 blk_mq_insert_request(rq, false, false, false);
 474         }
 475
 476         /*
 477          * Use the start variant of queue running here, so that running
 478          * the requeue work will kick stopped queues.
 479          */
 480         blk_mq_start_hw_queues(q);
 481 }
 482
 483 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 484 {
 485         struct request_queue *q = rq->q;
 486         unsigned long flags;
 487
 488         /*
 489          * We abuse this flag that is otherwise used by the I/O scheduler to
 490          * request head insertation from the workqueue.
 491          */
 492         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 493
 494         spin_lock_irqsave(&q->requeue_lock, flags);
 495         if (at_head) {
 496                 rq->cmd_flags |= REQ_SOFTBARRIER;
 497                 list_add(&rq->queuelist, &q->requeue_list);
 498         } else {
 499                 list_add_tail(&rq->queuelist, &q->requeue_list);
 500         }
 501         spin_unlock_irqrestore(&q->requeue_lock, flags);
 502 }
 503 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 504
 505 void blk_mq_kick_requeue_list(struct request_queue *q)
 506 {
 507         kblockd_schedule_work(&q->requeue_work);
 508 }
 509 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 510
 511 static inline bool is_flush_request(struct request *rq, unsigned int tag)
 512 {
 513         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 514                         rq->q->flush_rq->tag == tag);
 515 }
 516
 517 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 518 {
 519         struct request *rq = tags->rqs[tag];
 520
 521         if (!is_flush_request(rq, tag))
 522                 return rq;
 523
 524         return rq->q->flush_rq;
 525 }
 526 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 527
 528 struct blk_mq_timeout_data {
 529         unsigned long next;
 530         unsigned int next_set;
 531 };
 532
 533 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 534 {
 535         struct blk_mq_ops *ops = req->q->mq_ops;
 536         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 537
 538         /*
 539          * We know that complete is set at this point. If STARTED isn't set
 540          * anymore, then the request isn't active and the "timeout" should
 541          * just be ignored. This can happen due to the bitflag ordering.
 542          * Timeout first checks if STARTED is set, and if it is, assumes
 543          * the request is active. But if we race with completion, then
 544          * we both flags will get cleared. So check here again, and ignore
 545          * a timeout event with a request that isn't active.
 546          */
 547         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 548                 return;
 549
 550         if (ops->timeout)
 551                 ret = ops->timeout(req, reserved);
 552
 553         switch (ret) {
 554         case BLK_EH_HANDLED:
 555                 __blk_mq_complete_request(req);
 556                 break;
 557         case BLK_EH_RESET_TIMER:
 558                 blk_add_timer(req);
 559                 blk_clear_rq_complete(req);
 560                 break;
 561         case BLK_EH_NOT_HANDLED:
 562                 break;
 563         default:
 564                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 565                 break;
 566         }
 567 }
 568
 569 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 570                 struct request *rq, void *priv, bool reserved)
 571 {
 572         struct blk_mq_timeout_data *data = priv;
 573
 574         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 575                 return;
 576
 577         if (time_after_eq(jiffies, rq->deadline)) {
 578                 if (!blk_mark_rq_complete(rq))
 579                         blk_mq_rq_timed_out(rq, reserved);
 580         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 581                 data->next = rq->deadline;
 582                 data->next_set = 1;
 583         }
 584 }
 585
 586 static void blk_mq_rq_timer(unsigned long priv)
 587 {
 588         struct request_queue *q = (struct request_queue *)priv;
 589         struct blk_mq_timeout_data data = {
 590                 .next           = 0,
 591                 .next_set       = 0,
 592         };
 593         struct blk_mq_hw_ctx *hctx;
 594         int i;
 595
 596         queue_for_each_hw_ctx(q, hctx, i) {
 597                 /*
 598                  * If not software queues are currently mapped to this
 599                  * hardware queue, there's nothing to check
 600                  */
 601                 if (!hctx->nr_ctx || !hctx->tags)
 602                         continue;
 603
 604                 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
 605         }
 606
 607         if (data.next_set) {
 608                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 609                 mod_timer(&q->timeout, data.next);
 610         } else {
 611                 queue_for_each_hw_ctx(q, hctx, i)
 612                         blk_mq_tag_idle(hctx);
 613         }
 614 }
 615
 616 /*
 617  * Reverse check our software queue for entries that we could potentially
 618  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 619  * too much time checking for merges.
 620  */
 621 static bool blk_mq_attempt_merge(struct request_queue *q,
 622                                  struct blk_mq_ctx *ctx, struct bio *bio)
 623 {
 624         struct request *rq;
 625         int checked = 8;
 626
 627         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 628                 int el_ret;
 629
 630                 if (!checked--)
 631                         break;
 632
 633                 if (!blk_rq_merge_ok(rq, bio))
 634                         continue;
 635
 636                 el_ret = blk_try_merge(rq, bio);
 637                 if (el_ret == ELEVATOR_BACK_MERGE) {
 638                         if (bio_attempt_back_merge(q, rq, bio)) {
 639                                 ctx->rq_merged++;
 640                                 return true;
 641                         }
 642                         break;
 643                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 644                         if (bio_attempt_front_merge(q, rq, bio)) {
 645                                 ctx->rq_merged++;
 646                                 return true;
 647                         }
 648                         break;
 649                 }
 650         }
 651
 652         return false;
 653 }
 654
 655 /*
 656  * Process software queues that have been marked busy, splicing them
 657  * to the for-dispatch
 658  */
 659 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 660 {
 661         struct blk_mq_ctx *ctx;
 662         int i;
 663
 664         for (i = 0; i < hctx->ctx_map.map_size; i++) {
 665                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 666                 unsigned int off, bit;
 667
 668                 if (!bm->word)
 669                         continue;
 670
 671                 bit = 0;
 672                 off = i * hctx->ctx_map.bits_per_word;
 673                 do {
 674                         bit = find_next_bit(&bm->word, bm->depth, bit);
 675                         if (bit >= bm->depth)
 676                                 break;
 677
 678                         ctx = hctx->ctxs[bit + off];
 679                         clear_bit(bit, &bm->word);
 680                         spin_lock(&ctx->lock);
 681                         list_splice_tail_init(&ctx->rq_list, list);
 682                         spin_unlock(&ctx->lock);
 683
 684                         bit++;
 685                 } while (1);
 686         }
 687 }
 688
 689 /*
 690  * Run this hardware queue, pulling any software queues mapped to it in.
 691  * Note that this function currently has various problems around ordering
 692  * of IO. In particular, we'd like FIFO behaviour on handling existing
 693  * items on the hctx->dispatch list. Ignore that for now.
 694  */
 695 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 696 {
 697         struct request_queue *q = hctx->queue;
 698         struct request *rq;
 699         LIST_HEAD(rq_list);
 700         int queued;
 701
 702         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 703
 704         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 705                 return;
 706
 707         hctx->run++;
 708
 709         /*
 710          * Touch any software queue that has pending entries.
 711          */
 712         flush_busy_ctxs(hctx, &rq_list);
 713
 714         /*
 715          * If we have previous entries on our dispatch list, grab them
 716          * and stuff them at the front for more fair dispatch.
 717          */
 718         if (!list_empty_careful(&hctx->dispatch)) {
 719                 spin_lock(&hctx->lock);
 720                 if (!list_empty(&hctx->dispatch))
 721                         list_splice_init(&hctx->dispatch, &rq_list);
 722                 spin_unlock(&hctx->lock);
 723         }
 724
 725         /*
 726          * Now process all the entries, sending them to the driver.
 727          */
 728         queued = 0;
 729         while (!list_empty(&rq_list)) {
 730                 int ret;
 731
 732                 rq = list_first_entry(&rq_list, struct request, queuelist);
 733                 list_del_init(&rq->queuelist);
 734
 735                 ret = q->mq_ops->queue_rq(hctx, rq, list_empty(&rq_list));
 736                 switch (ret) {
 737                 case BLK_MQ_RQ_QUEUE_OK:
 738                         queued++;
 739                         continue;
 740                 case BLK_MQ_RQ_QUEUE_BUSY:
 741                         list_add(&rq->queuelist, &rq_list);
 742                         __blk_mq_requeue_request(rq);
 743                         break;
 744                 default:
 745                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 746                 case BLK_MQ_RQ_QUEUE_ERROR:
 747                         rq->errors = -EIO;
 748                         blk_mq_end_request(rq, rq->errors);
 749                         break;
 750                 }
 751
 752                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 753                         break;
 754         }
 755
 756         if (!queued)
 757                 hctx->dispatched[0]++;
 758         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 759                 hctx->dispatched[ilog2(queued) + 1]++;
 760
 761         /*
 762          * Any items that need requeuing? Stuff them into hctx->dispatch,
 763          * that is where we will continue on next queue run.
 764          */
 765         if (!list_empty(&rq_list)) {
 766                 spin_lock(&hctx->lock);
 767                 list_splice(&rq_list, &hctx->dispatch);
 768                 spin_unlock(&hctx->lock);
 769         }
 770 }
 771
 772 /*
 773  * It'd be great if the workqueue API had a way to pass
 774  * in a mask and had some smarts for more clever placement.
 775  * For now we just round-robin here, switching for every
 776  * BLK_MQ_CPU_WORK_BATCH queued items.
 777  */
 778 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 779 {
 780         int cpu = hctx->next_cpu;
 781
 782         if (--hctx->next_cpu_batch <= 0) {
 783                 int next_cpu;
 784
 785                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 786                 if (next_cpu >= nr_cpu_ids)
 787                         next_cpu = cpumask_first(hctx->cpumask);
 788
 789                 hctx->next_cpu = next_cpu;
 790                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 791         }
 792
 793         return cpu;
 794 }
 795
 796 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 797 {
 798         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 799                 return;
 800
 801         if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
 802                 __blk_mq_run_hw_queue(hctx);
 803         else if (hctx->queue->nr_hw_queues == 1)
 804                 kblockd_schedule_delayed_work(&hctx->run_work, 0);
 805         else {
 806                 unsigned int cpu;
 807
 808                 cpu = blk_mq_hctx_next_cpu(hctx);
 809                 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
 810         }
 811 }
 812
 813 void blk_mq_run_queues(struct request_queue *q, bool async)
 814 {
 815         struct blk_mq_hw_ctx *hctx;
 816         int i;
 817
 818         queue_for_each_hw_ctx(q, hctx, i) {
 819                 if ((!blk_mq_hctx_has_pending(hctx) &&
 820                     list_empty_careful(&hctx->dispatch)) ||
 821                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 822                         continue;
 823
 824                 preempt_disable();
 825                 blk_mq_run_hw_queue(hctx, async);
 826                 preempt_enable();
 827         }
 828 }
 829 EXPORT_SYMBOL(blk_mq_run_queues);
 830
 831 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 832 {
 833         cancel_delayed_work(&hctx->run_work);
 834         cancel_delayed_work(&hctx->delay_work);
 835         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 836 }
 837 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 838
 839 void blk_mq_stop_hw_queues(struct request_queue *q)
 840 {
 841         struct blk_mq_hw_ctx *hctx;
 842         int i;
 843
 844         queue_for_each_hw_ctx(q, hctx, i)
 845                 blk_mq_stop_hw_queue(hctx);
 846 }
 847 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 848
 849 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 850 {
 851         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 852
 853         preempt_disable();
 854         blk_mq_run_hw_queue(hctx, false);
 855         preempt_enable();
 856 }
 857 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 858
 859 void blk_mq_start_hw_queues(struct request_queue *q)
 860 {
 861         struct blk_mq_hw_ctx *hctx;
 862         int i;
 863
 864         queue_for_each_hw_ctx(q, hctx, i)
 865                 blk_mq_start_hw_queue(hctx);
 866 }
 867 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 868
 869
 870 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 871 {
 872         struct blk_mq_hw_ctx *hctx;
 873         int i;
 874
 875         queue_for_each_hw_ctx(q, hctx, i) {
 876                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 877                         continue;
 878
 879                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 880                 preempt_disable();
 881                 blk_mq_run_hw_queue(hctx, async);
 882                 preempt_enable();
 883         }
 884 }
 885 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 886
 887 static void blk_mq_run_work_fn(struct work_struct *work)
 888 {
 889         struct blk_mq_hw_ctx *hctx;
 890
 891         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 892
 893         __blk_mq_run_hw_queue(hctx);
 894 }
 895
 896 static void blk_mq_delay_work_fn(struct work_struct *work)
 897 {
 898         struct blk_mq_hw_ctx *hctx;
 899
 900         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 901
 902         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 903                 __blk_mq_run_hw_queue(hctx);
 904 }
 905
 906 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 907 {
 908         unsigned long tmo = msecs_to_jiffies(msecs);
 909
 910         if (hctx->queue->nr_hw_queues == 1)
 911                 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
 912         else {
 913                 unsigned int cpu;
 914
 915                 cpu = blk_mq_hctx_next_cpu(hctx);
 916                 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
 917         }
 918 }
 919 EXPORT_SYMBOL(blk_mq_delay_queue);
 920
 921 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
 922                                     struct request *rq, bool at_head)
 923 {
 924         struct blk_mq_ctx *ctx = rq->mq_ctx;
 925
 926         trace_block_rq_insert(hctx->queue, rq);
 927
 928         if (at_head)
 929                 list_add(&rq->queuelist, &ctx->rq_list);
 930         else
 931                 list_add_tail(&rq->queuelist, &ctx->rq_list);
 932
 933         blk_mq_hctx_mark_pending(hctx, ctx);
 934 }
 935
 936 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
 937                 bool async)
 938 {
 939         struct request_queue *q = rq->q;
 940         struct blk_mq_hw_ctx *hctx;
 941         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
 942
 943         current_ctx = blk_mq_get_ctx(q);
 944         if (!cpu_online(ctx->cpu))
 945                 rq->mq_ctx = ctx = current_ctx;
 946
 947         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 948
 949         spin_lock(&ctx->lock);
 950         __blk_mq_insert_request(hctx, rq, at_head);
 951         spin_unlock(&ctx->lock);
 952
 953         if (run_queue)
 954                 blk_mq_run_hw_queue(hctx, async);
 955
 956         blk_mq_put_ctx(current_ctx);
 957 }
 958
 959 static void blk_mq_insert_requests(struct request_queue *q,
 960                                      struct blk_mq_ctx *ctx,
 961                                      struct list_head *list,
 962                                      int depth,
 963                                      bool from_schedule)
 964
 965 {
 966         struct blk_mq_hw_ctx *hctx;
 967         struct blk_mq_ctx *current_ctx;
 968
 969         trace_block_unplug(q, depth, !from_schedule);
 970
 971         current_ctx = blk_mq_get_ctx(q);
 972
 973         if (!cpu_online(ctx->cpu))
 974                 ctx = current_ctx;
 975         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 976
 977         /*
 978          * preemption doesn't flush plug list, so it's possible ctx->cpu is
 979          * offline now
 980          */
 981         spin_lock(&ctx->lock);
 982         while (!list_empty(list)) {
 983                 struct request *rq;
 984
 985                 rq = list_first_entry(list, struct request, queuelist);
 986                 list_del_init(&rq->queuelist);
 987                 rq->mq_ctx = ctx;
 988                 __blk_mq_insert_request(hctx, rq, false);
 989         }
 990         spin_unlock(&ctx->lock);
 991
 992         blk_mq_run_hw_queue(hctx, from_schedule);
 993         blk_mq_put_ctx(current_ctx);
 994 }
 995
 996 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
 997 {
 998         struct request *rqa = container_of(a, struct request, queuelist);
 999         struct request *rqb = container_of(b, struct request, queuelist);
1000
1001         return !(rqa->mq_ctx < rqb->mq_ctx ||
1002                  (rqa->mq_ctx == rqb->mq_ctx &&
1003                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1004 }
1005
1006 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1007 {
1008         struct blk_mq_ctx *this_ctx;
1009         struct request_queue *this_q;
1010         struct request *rq;
1011         LIST_HEAD(list);
1012         LIST_HEAD(ctx_list);
1013         unsigned int depth;
1014
1015         list_splice_init(&plug->mq_list, &list);
1016
1017         list_sort(NULL, &list, plug_ctx_cmp);
1018
1019         this_q = NULL;
1020         this_ctx = NULL;
1021         depth = 0;
1022
1023         while (!list_empty(&list)) {
1024                 rq = list_entry_rq(list.next);
1025                 list_del_init(&rq->queuelist);
1026                 BUG_ON(!rq->q);
1027                 if (rq->mq_ctx != this_ctx) {
1028                         if (this_ctx) {
1029                                 blk_mq_insert_requests(this_q, this_ctx,
1030                                                         &ctx_list, depth,
1031                                                         from_schedule);
1032                         }
1033
1034                         this_ctx = rq->mq_ctx;
1035                         this_q = rq->q;
1036                         depth = 0;
1037                 }
1038
1039                 depth++;
1040                 list_add_tail(&rq->queuelist, &ctx_list);
1041         }
1042
1043         /*
1044          * If 'this_ctx' is set, we know we have entries to complete
1045          * on 'ctx_list'. Do those.
1046          */
1047         if (this_ctx) {
1048                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1049                                        from_schedule);
1050         }
1051 }
1052
1053 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1054 {
1055         init_request_from_bio(rq, bio);
1056
1057         if (blk_do_io_stat(rq))
1058                 blk_account_io_start(rq, 1);
1059 }
1060
1061 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1062 {
1063         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1064                 !blk_queue_nomerges(hctx->queue);
1065 }
1066
1067 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1068                                          struct blk_mq_ctx *ctx,
1069                                          struct request *rq, struct bio *bio)
1070 {
1071         if (!hctx_allow_merges(hctx)) {
1072                 blk_mq_bio_to_request(rq, bio);
1073                 spin_lock(&ctx->lock);
1074 insert_rq:
1075                 __blk_mq_insert_request(hctx, rq, false);
1076                 spin_unlock(&ctx->lock);
1077                 return false;
1078         } else {
1079                 struct request_queue *q = hctx->queue;
1080
1081                 spin_lock(&ctx->lock);
1082                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1083                         blk_mq_bio_to_request(rq, bio);
1084                         goto insert_rq;
1085                 }
1086
1087                 spin_unlock(&ctx->lock);
1088                 __blk_mq_free_request(hctx, ctx, rq);
1089                 return true;
1090         }
1091 }
1092
1093 struct blk_map_ctx {
1094         struct blk_mq_hw_ctx *hctx;
1095         struct blk_mq_ctx *ctx;
1096 };
1097
1098 static struct request *blk_mq_map_request(struct request_queue *q,
1099                                           struct bio *bio,
1100                                           struct blk_map_ctx *data)
1101 {
1102         struct blk_mq_hw_ctx *hctx;
1103         struct blk_mq_ctx *ctx;
1104         struct request *rq;
1105         int rw = bio_data_dir(bio);
1106         struct blk_mq_alloc_data alloc_data;
1107
1108         if (unlikely(blk_mq_queue_enter(q))) {
1109                 bio_endio(bio, -EIO);
1110                 return NULL;
1111         }
1112
1113         ctx = blk_mq_get_ctx(q);
1114         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1115
1116         if (rw_is_sync(bio->bi_rw))
1117                 rw |= REQ_SYNC;
1118
1119         trace_block_getrq(q, bio, rw);
1120         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1121                         hctx);
1122         rq = __blk_mq_alloc_request(&alloc_data, rw);
1123         if (unlikely(!rq)) {
1124                 __blk_mq_run_hw_queue(hctx);
1125                 blk_mq_put_ctx(ctx);
1126                 trace_block_sleeprq(q, bio, rw);
1127
1128                 ctx = blk_mq_get_ctx(q);
1129                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1130                 blk_mq_set_alloc_data(&alloc_data, q,
1131                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1132                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1133                 ctx = alloc_data.ctx;
1134                 hctx = alloc_data.hctx;
1135         }
1136
1137         hctx->queued++;
1138         data->hctx = hctx;
1139         data->ctx = ctx;
1140         return rq;
1141 }
1142
1143 /*
1144  * Multiple hardware queue variant. This will not use per-process plugs,
1145  * but will attempt to bypass the hctx queueing if we can go straight to
1146  * hardware for SYNC IO.
1147  */
1148 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1149 {
1150         const int is_sync = rw_is_sync(bio->bi_rw);
1151         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1152         struct blk_map_ctx data;
1153         struct request *rq;
1154
1155         blk_queue_bounce(q, &bio);
1156
1157         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1158                 bio_endio(bio, -EIO);
1159                 return;
1160         }
1161
1162         rq = blk_mq_map_request(q, bio, &data);
1163         if (unlikely(!rq))
1164                 return;
1165
1166         if (unlikely(is_flush_fua)) {
1167                 blk_mq_bio_to_request(rq, bio);
1168                 blk_insert_flush(rq);
1169                 goto run_queue;
1170         }
1171
1172         if (is_sync) {
1173                 int ret;
1174
1175                 blk_mq_bio_to_request(rq, bio);
1176
1177                 /*
1178                  * For OK queue, we are done. For error, kill it. Any other
1179                  * error (busy), just add it to our list as we previously
1180                  * would have done
1181                  */
1182                 ret = q->mq_ops->queue_rq(data.hctx, rq, true);
1183                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1184                         goto done;
1185                 else {
1186                         __blk_mq_requeue_request(rq);
1187
1188                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1189                                 rq->errors = -EIO;
1190                                 blk_mq_end_request(rq, rq->errors);
1191                                 goto done;
1192                         }
1193                 }
1194         }
1195
1196         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1197                 /*
1198                  * For a SYNC request, send it to the hardware immediately. For
1199                  * an ASYNC request, just ensure that we run it later on. The
1200                  * latter allows for merging opportunities and more efficient
1201                  * dispatching.
1202                  */
1203 run_queue:
1204                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1205         }
1206 done:
1207         blk_mq_put_ctx(data.ctx);
1208 }
1209
1210 /*
1211  * Single hardware queue variant. This will attempt to use any per-process
1212  * plug for merging and IO deferral.
1213  */
1214 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1215 {
1216         const int is_sync = rw_is_sync(bio->bi_rw);
1217         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1218         unsigned int use_plug, request_count = 0;
1219         struct blk_map_ctx data;
1220         struct request *rq;
1221
1222         /*
1223          * If we have multiple hardware queues, just go directly to
1224          * one of those for sync IO.
1225          */
1226         use_plug = !is_flush_fua && !is_sync;
1227
1228         blk_queue_bounce(q, &bio);
1229
1230         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1231                 bio_endio(bio, -EIO);
1232                 return;
1233         }
1234
1235         if (use_plug && !blk_queue_nomerges(q) &&
1236             blk_attempt_plug_merge(q, bio, &request_count))
1237                 return;
1238
1239         rq = blk_mq_map_request(q, bio, &data);
1240         if (unlikely(!rq))
1241                 return;
1242
1243         if (unlikely(is_flush_fua)) {
1244                 blk_mq_bio_to_request(rq, bio);
1245                 blk_insert_flush(rq);
1246                 goto run_queue;
1247         }
1248
1249         /*
1250          * A task plug currently exists. Since this is completely lockless,
1251          * utilize that to temporarily store requests until the task is
1252          * either done or scheduled away.
1253          */
1254         if (use_plug) {
1255                 struct blk_plug *plug = current->plug;
1256
1257                 if (plug) {
1258                         blk_mq_bio_to_request(rq, bio);
1259                         if (list_empty(&plug->mq_list))
1260                                 trace_block_plug(q);
1261                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1262                                 blk_flush_plug_list(plug, false);
1263                                 trace_block_plug(q);
1264                         }
1265                         list_add_tail(&rq->queuelist, &plug->mq_list);
1266                         blk_mq_put_ctx(data.ctx);
1267                         return;
1268                 }
1269         }
1270
1271         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1272                 /*
1273                  * For a SYNC request, send it to the hardware immediately. For
1274                  * an ASYNC request, just ensure that we run it later on. The
1275                  * latter allows for merging opportunities and more efficient
1276                  * dispatching.
1277                  */
1278 run_queue:
1279                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1280         }
1281
1282         blk_mq_put_ctx(data.ctx);
1283 }
1284
1285 /*
1286  * Default mapping to a software queue, since we use one per CPU.
1287  */
1288 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1289 {
1290         return q->queue_hw_ctx[q->mq_map[cpu]];
1291 }
1292 EXPORT_SYMBOL(blk_mq_map_queue);
1293
1294 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1295                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1296 {
1297         struct page *page;
1298
1299         if (tags->rqs && set->ops->exit_request) {
1300                 int i;
1301
1302                 for (i = 0; i < tags->nr_tags; i++) {
1303                         if (!tags->rqs[i])
1304                                 continue;
1305                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1306                                                 hctx_idx, i);
1307                         tags->rqs[i] = NULL;
1308                 }
1309         }
1310
1311         while (!list_empty(&tags->page_list)) {
1312                 page = list_first_entry(&tags->page_list, struct page, lru);
1313                 list_del_init(&page->lru);
1314                 __free_pages(page, page->private);
1315         }
1316
1317         kfree(tags->rqs);
1318
1319         blk_mq_free_tags(tags);
1320 }
1321
1322 static size_t order_to_size(unsigned int order)
1323 {
1324         return (size_t)PAGE_SIZE << order;
1325 }
1326
1327 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1328                 unsigned int hctx_idx)
1329 {
1330         struct blk_mq_tags *tags;
1331         unsigned int i, j, entries_per_page, max_order = 4;
1332         size_t rq_size, left;
1333
1334         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1335                                 set->numa_node);
1336         if (!tags)
1337                 return NULL;
1338
1339         INIT_LIST_HEAD(&tags->page_list);
1340
1341         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1342                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1343                                  set->numa_node);
1344         if (!tags->rqs) {
1345                 blk_mq_free_tags(tags);
1346                 return NULL;
1347         }
1348
1349         /*
1350          * rq_size is the size of the request plus driver payload, rounded
1351          * to the cacheline size
1352          */
1353         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1354                                 cache_line_size());
1355         left = rq_size * set->queue_depth;
1356
1357         for (i = 0; i < set->queue_depth; ) {
1358                 int this_order = max_order;
1359                 struct page *page;
1360                 int to_do;
1361                 void *p;
1362
1363                 while (left < order_to_size(this_order - 1) && this_order)
1364                         this_order--;
1365
1366                 do {
1367                         page = alloc_pages_node(set->numa_node,
1368                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1369                                 this_order);
1370                         if (page)
1371                                 break;
1372                         if (!this_order--)
1373                                 break;
1374                         if (order_to_size(this_order) < rq_size)
1375                                 break;
1376                 } while (1);
1377
1378                 if (!page)
1379                         goto fail;
1380
1381                 page->private = this_order;
1382                 list_add_tail(&page->lru, &tags->page_list);
1383
1384                 p = page_address(page);
1385                 entries_per_page = order_to_size(this_order) / rq_size;
1386                 to_do = min(entries_per_page, set->queue_depth - i);
1387                 left -= to_do * rq_size;
1388                 for (j = 0; j < to_do; j++) {
1389                         tags->rqs[i] = p;
1390                         tags->rqs[i]->atomic_flags = 0;
1391                         tags->rqs[i]->cmd_flags = 0;
1392                         if (set->ops->init_request) {
1393                                 if (set->ops->init_request(set->driver_data,
1394                                                 tags->rqs[i], hctx_idx, i,
1395                                                 set->numa_node)) {
1396                                         tags->rqs[i] = NULL;
1397                                         goto fail;
1398                                 }
1399                         }
1400
1401                         p += rq_size;
1402                         i++;
1403                 }
1404         }
1405
1406         return tags;
1407
1408 fail:
1409         blk_mq_free_rq_map(set, tags, hctx_idx);
1410         return NULL;
1411 }
1412
1413 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1414 {
1415         kfree(bitmap->map);
1416 }
1417
1418 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1419 {
1420         unsigned int bpw = 8, total, num_maps, i;
1421
1422         bitmap->bits_per_word = bpw;
1423
1424         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1425         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1426                                         GFP_KERNEL, node);
1427         if (!bitmap->map)
1428                 return -ENOMEM;
1429
1430         bitmap->map_size = num_maps;
1431
1432         total = nr_cpu_ids;
1433         for (i = 0; i < num_maps; i++) {
1434                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1435                 total -= bitmap->map[i].depth;
1436         }
1437
1438         return 0;
1439 }
1440
1441 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1442 {
1443         struct request_queue *q = hctx->queue;
1444         struct blk_mq_ctx *ctx;
1445         LIST_HEAD(tmp);
1446
1447         /*
1448          * Move ctx entries to new CPU, if this one is going away.
1449          */
1450         ctx = __blk_mq_get_ctx(q, cpu);
1451
1452         spin_lock(&ctx->lock);
1453         if (!list_empty(&ctx->rq_list)) {
1454                 list_splice_init(&ctx->rq_list, &tmp);
1455                 blk_mq_hctx_clear_pending(hctx, ctx);
1456         }
1457         spin_unlock(&ctx->lock);
1458
1459         if (list_empty(&tmp))
1460                 return NOTIFY_OK;
1461
1462         ctx = blk_mq_get_ctx(q);
1463         spin_lock(&ctx->lock);
1464
1465         while (!list_empty(&tmp)) {
1466                 struct request *rq;
1467
1468                 rq = list_first_entry(&tmp, struct request, queuelist);
1469                 rq->mq_ctx = ctx;
1470                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1471         }
1472
1473         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1474         blk_mq_hctx_mark_pending(hctx, ctx);
1475
1476         spin_unlock(&ctx->lock);
1477
1478         blk_mq_run_hw_queue(hctx, true);
1479         blk_mq_put_ctx(ctx);
1480         return NOTIFY_OK;
1481 }
1482
1483 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1484 {
1485         struct request_queue *q = hctx->queue;
1486         struct blk_mq_tag_set *set = q->tag_set;
1487
1488         if (set->tags[hctx->queue_num])
1489                 return NOTIFY_OK;
1490
1491         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1492         if (!set->tags[hctx->queue_num])
1493                 return NOTIFY_STOP;
1494
1495         hctx->tags = set->tags[hctx->queue_num];
1496         return NOTIFY_OK;
1497 }
1498
1499 static int blk_mq_hctx_notify(void *data, unsigned long action,
1500                               unsigned int cpu)
1501 {
1502         struct blk_mq_hw_ctx *hctx = data;
1503
1504         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1505                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1506         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1507                 return blk_mq_hctx_cpu_online(hctx, cpu);
1508
1509         return NOTIFY_OK;
1510 }
1511
1512 static void blk_mq_exit_hctx(struct request_queue *q,
1513                 struct blk_mq_tag_set *set,
1514                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1515 {
1516         blk_mq_tag_idle(hctx);
1517
1518         if (set->ops->exit_hctx)
1519                 set->ops->exit_hctx(hctx, hctx_idx);
1520
1521         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1522         kfree(hctx->ctxs);
1523         blk_mq_free_bitmap(&hctx->ctx_map);
1524 }
1525
1526 static void blk_mq_exit_hw_queues(struct request_queue *q,
1527                 struct blk_mq_tag_set *set, int nr_queue)
1528 {
1529         struct blk_mq_hw_ctx *hctx;
1530         unsigned int i;
1531
1532         queue_for_each_hw_ctx(q, hctx, i) {
1533                 if (i == nr_queue)
1534                         break;
1535                 blk_mq_exit_hctx(q, set, hctx, i);
1536         }
1537 }
1538
1539 static void blk_mq_free_hw_queues(struct request_queue *q,
1540                 struct blk_mq_tag_set *set)
1541 {
1542         struct blk_mq_hw_ctx *hctx;
1543         unsigned int i;
1544
1545         queue_for_each_hw_ctx(q, hctx, i) {
1546                 free_cpumask_var(hctx->cpumask);
1547                 kfree(hctx);
1548         }
1549 }
1550
1551 static int blk_mq_init_hctx(struct request_queue *q,
1552                 struct blk_mq_tag_set *set,
1553                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1554 {
1555         int node;
1556
1557         node = hctx->numa_node;
1558         if (node == NUMA_NO_NODE)
1559                 node = hctx->numa_node = set->numa_node;
1560
1561         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1562         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1563         spin_lock_init(&hctx->lock);
1564         INIT_LIST_HEAD(&hctx->dispatch);
1565         hctx->queue = q;
1566         hctx->queue_num = hctx_idx;
1567         hctx->flags = set->flags;
1568         hctx->cmd_size = set->cmd_size;
1569
1570         blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1571                                         blk_mq_hctx_notify, hctx);
1572         blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1573
1574         hctx->tags = set->tags[hctx_idx];
1575
1576         /*
1577          * Allocate space for all possible cpus to avoid allocation at
1578          * runtime
1579          */
1580         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1581                                         GFP_KERNEL, node);
1582         if (!hctx->ctxs)
1583                 goto unregister_cpu_notifier;
1584
1585         if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1586                 goto free_ctxs;
1587
1588         hctx->nr_ctx = 0;
1589
1590         if (set->ops->init_hctx &&
1591             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1592                 goto free_bitmap;
1593
1594         return 0;
1595
1596  free_bitmap:
1597         blk_mq_free_bitmap(&hctx->ctx_map);
1598  free_ctxs:
1599         kfree(hctx->ctxs);
1600  unregister_cpu_notifier:
1601         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1602
1603         return -1;
1604 }
1605
1606 static int blk_mq_init_hw_queues(struct request_queue *q,
1607                 struct blk_mq_tag_set *set)
1608 {
1609         struct blk_mq_hw_ctx *hctx;
1610         unsigned int i;
1611
1612         /*
1613          * Initialize hardware queues
1614          */
1615         queue_for_each_hw_ctx(q, hctx, i) {
1616                 if (blk_mq_init_hctx(q, set, hctx, i))
1617                         break;
1618         }
1619
1620         if (i == q->nr_hw_queues)
1621                 return 0;
1622
1623         /*
1624          * Init failed
1625          */
1626         blk_mq_exit_hw_queues(q, set, i);
1627
1628         return 1;
1629 }
1630
1631 static void blk_mq_init_cpu_queues(struct request_queue *q,
1632                                    unsigned int nr_hw_queues)
1633 {
1634         unsigned int i;
1635
1636         for_each_possible_cpu(i) {
1637                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1638                 struct blk_mq_hw_ctx *hctx;
1639
1640                 memset(__ctx, 0, sizeof(*__ctx));
1641                 __ctx->cpu = i;
1642                 spin_lock_init(&__ctx->lock);
1643                 INIT_LIST_HEAD(&__ctx->rq_list);
1644                 __ctx->queue = q;
1645
1646                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1647                 if (!cpu_online(i))
1648                         continue;
1649
1650                 hctx = q->mq_ops->map_queue(q, i);
1651                 cpumask_set_cpu(i, hctx->cpumask);
1652                 hctx->nr_ctx++;
1653
1654                 /*
1655                  * Set local node, IFF we have more than one hw queue. If
1656                  * not, we remain on the home node of the device
1657                  */
1658                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1659                         hctx->numa_node = cpu_to_node(i);
1660         }
1661 }
1662
1663 static void blk_mq_map_swqueue(struct request_queue *q)
1664 {
1665         unsigned int i;
1666         struct blk_mq_hw_ctx *hctx;
1667         struct blk_mq_ctx *ctx;
1668
1669         queue_for_each_hw_ctx(q, hctx, i) {
1670                 cpumask_clear(hctx->cpumask);
1671                 hctx->nr_ctx = 0;
1672         }
1673
1674         /*
1675          * Map software to hardware queues
1676          */
1677         queue_for_each_ctx(q, ctx, i) {
1678                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1679                 if (!cpu_online(i))
1680                         continue;
1681
1682                 hctx = q->mq_ops->map_queue(q, i);
1683                 cpumask_set_cpu(i, hctx->cpumask);
1684                 ctx->index_hw = hctx->nr_ctx;
1685                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1686         }
1687
1688         queue_for_each_hw_ctx(q, hctx, i) {
1689                 /*
1690                  * If no software queues are mapped to this hardware queue,
1691                  * disable it and free the request entries.
1692                  */
1693                 if (!hctx->nr_ctx) {
1694                         struct blk_mq_tag_set *set = q->tag_set;
1695
1696                         if (set->tags[i]) {
1697                                 blk_mq_free_rq_map(set, set->tags[i], i);
1698                                 set->tags[i] = NULL;
1699                                 hctx->tags = NULL;
1700                         }
1701                         continue;
1702                 }
1703
1704                 /*
1705                  * Initialize batch roundrobin counts
1706                  */
1707                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1708                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1709         }
1710 }
1711
1712 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1713 {
1714         struct blk_mq_hw_ctx *hctx;
1715         struct request_queue *q;
1716         bool shared;
1717         int i;
1718
1719         if (set->tag_list.next == set->tag_list.prev)
1720                 shared = false;
1721         else
1722                 shared = true;
1723
1724         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1725                 blk_mq_freeze_queue(q);
1726
1727                 queue_for_each_hw_ctx(q, hctx, i) {
1728                         if (shared)
1729                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1730                         else
1731                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1732                 }
1733                 blk_mq_unfreeze_queue(q);
1734         }
1735 }
1736
1737 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1738 {
1739         struct blk_mq_tag_set *set = q->tag_set;
1740
1741         mutex_lock(&set->tag_list_lock);
1742         list_del_init(&q->tag_set_list);
1743         blk_mq_update_tag_set_depth(set);
1744         mutex_unlock(&set->tag_list_lock);
1745 }
1746
1747 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1748                                      struct request_queue *q)
1749 {
1750         q->tag_set = set;
1751
1752         mutex_lock(&set->tag_list_lock);
1753         list_add_tail(&q->tag_set_list, &set->tag_list);
1754         blk_mq_update_tag_set_depth(set);
1755         mutex_unlock(&set->tag_list_lock);
1756 }
1757
1758 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1759 {
1760         struct blk_mq_hw_ctx **hctxs;
1761         struct blk_mq_ctx __percpu *ctx;
1762         struct request_queue *q;
1763         unsigned int *map;
1764         int i;
1765
1766         ctx = alloc_percpu(struct blk_mq_ctx);
1767         if (!ctx)
1768                 return ERR_PTR(-ENOMEM);
1769
1770         /*
1771          * If a crashdump is active, then we are potentially in a very
1772          * memory constrained environment. Limit us to 1 queue and
1773          * 64 tags to prevent using too much memory.
1774          */
1775         if (is_kdump_kernel()) {
1776                 set->nr_hw_queues = 1;
1777                 set->queue_depth = min(64U, set->queue_depth);
1778         }
1779
1780         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1781                         set->numa_node);
1782
1783         if (!hctxs)
1784                 goto err_percpu;
1785
1786         map = blk_mq_make_queue_map(set);
1787         if (!map)
1788                 goto err_map;
1789
1790         for (i = 0; i < set->nr_hw_queues; i++) {
1791                 int node = blk_mq_hw_queue_to_node(map, i);
1792
1793                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1794                                         GFP_KERNEL, node);
1795                 if (!hctxs[i])
1796                         goto err_hctxs;
1797
1798                 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1799                         goto err_hctxs;
1800
1801                 atomic_set(&hctxs[i]->nr_active, 0);
1802                 hctxs[i]->numa_node = node;
1803                 hctxs[i]->queue_num = i;
1804         }
1805
1806         q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1807         if (!q)
1808                 goto err_hctxs;
1809
1810         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release))
1811                 goto err_map;
1812
1813         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1814         blk_queue_rq_timeout(q, 30000);
1815
1816         q->nr_queues = nr_cpu_ids;
1817         q->nr_hw_queues = set->nr_hw_queues;
1818         q->mq_map = map;
1819
1820         q->queue_ctx = ctx;
1821         q->queue_hw_ctx = hctxs;
1822
1823         q->mq_ops = set->ops;
1824         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1825
1826         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1827                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1828
1829         q->sg_reserved_size = INT_MAX;
1830
1831         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1832         INIT_LIST_HEAD(&q->requeue_list);
1833         spin_lock_init(&q->requeue_lock);
1834
1835         if (q->nr_hw_queues > 1)
1836                 blk_queue_make_request(q, blk_mq_make_request);
1837         else
1838                 blk_queue_make_request(q, blk_sq_make_request);
1839
1840         if (set->timeout)
1841                 blk_queue_rq_timeout(q, set->timeout);
1842
1843         /*
1844          * Do this after blk_queue_make_request() overrides it...
1845          */
1846         q->nr_requests = set->queue_depth;
1847
1848         if (set->ops->complete)
1849                 blk_queue_softirq_done(q, set->ops->complete);
1850
1851         blk_mq_init_flush(q);
1852         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1853
1854         q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1855                                 set->cmd_size, cache_line_size()),
1856                                 GFP_KERNEL);
1857         if (!q->flush_rq)
1858                 goto err_hw;
1859
1860         if (blk_mq_init_hw_queues(q, set))
1861                 goto err_flush_rq;
1862
1863         mutex_lock(&all_q_mutex);
1864         list_add_tail(&q->all_q_node, &all_q_list);
1865         mutex_unlock(&all_q_mutex);
1866
1867         blk_mq_add_queue_tag_set(set, q);
1868
1869         blk_mq_map_swqueue(q);
1870
1871         return q;
1872
1873 err_flush_rq:
1874         kfree(q->flush_rq);
1875 err_hw:
1876         blk_cleanup_queue(q);
1877 err_hctxs:
1878         kfree(map);
1879         for (i = 0; i < set->nr_hw_queues; i++) {
1880                 if (!hctxs[i])
1881                         break;
1882                 free_cpumask_var(hctxs[i]->cpumask);
1883                 kfree(hctxs[i]);
1884         }
1885 err_map:
1886         kfree(hctxs);
1887 err_percpu:
1888         free_percpu(ctx);
1889         return ERR_PTR(-ENOMEM);
1890 }
1891 EXPORT_SYMBOL(blk_mq_init_queue);
1892
1893 void blk_mq_free_queue(struct request_queue *q)
1894 {
1895         struct blk_mq_tag_set   *set = q->tag_set;
1896
1897         blk_mq_del_queue_tag_set(q);
1898
1899         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1900         blk_mq_free_hw_queues(q, set);
1901
1902         percpu_ref_exit(&q->mq_usage_counter);
1903
1904         free_percpu(q->queue_ctx);
1905         kfree(q->queue_hw_ctx);
1906         kfree(q->mq_map);
1907
1908         q->queue_ctx = NULL;
1909         q->queue_hw_ctx = NULL;
1910         q->mq_map = NULL;
1911
1912         mutex_lock(&all_q_mutex);
1913         list_del_init(&q->all_q_node);
1914         mutex_unlock(&all_q_mutex);
1915 }
1916
1917 /* Basically redo blk_mq_init_queue with queue frozen */
1918 static void blk_mq_queue_reinit(struct request_queue *q)
1919 {
1920         blk_mq_freeze_queue(q);
1921
1922         blk_mq_sysfs_unregister(q);
1923
1924         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1925
1926         /*
1927          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1928          * we should change hctx numa_node according to new topology (this
1929          * involves free and re-allocate memory, worthy doing?)
1930          */
1931
1932         blk_mq_map_swqueue(q);
1933
1934         blk_mq_sysfs_register(q);
1935
1936         blk_mq_unfreeze_queue(q);
1937 }
1938
1939 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1940                                       unsigned long action, void *hcpu)
1941 {
1942         struct request_queue *q;
1943
1944         /*
1945          * Before new mappings are established, hotadded cpu might already
1946          * start handling requests. This doesn't break anything as we map
1947          * offline CPUs to first hardware queue. We will re-init the queue
1948          * below to get optimal settings.
1949          */
1950         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1951             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1952                 return NOTIFY_OK;
1953
1954         mutex_lock(&all_q_mutex);
1955         list_for_each_entry(q, &all_q_list, all_q_node)
1956                 blk_mq_queue_reinit(q);
1957         mutex_unlock(&all_q_mutex);
1958         return NOTIFY_OK;
1959 }
1960
1961 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1962 {
1963         int i;
1964
1965         for (i = 0; i < set->nr_hw_queues; i++) {
1966                 set->tags[i] = blk_mq_init_rq_map(set, i);
1967                 if (!set->tags[i])
1968                         goto out_unwind;
1969         }
1970
1971         return 0;
1972
1973 out_unwind:
1974         while (--i >= 0)
1975                 blk_mq_free_rq_map(set, set->tags[i], i);
1976
1977         return -ENOMEM;
1978 }
1979
1980 /*
1981  * Allocate the request maps associated with this tag_set. Note that this
1982  * may reduce the depth asked for, if memory is tight. set->queue_depth
1983  * will be updated to reflect the allocated depth.
1984  */
1985 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1986 {
1987         unsigned int depth;
1988         int err;
1989
1990         depth = set->queue_depth;
1991         do {
1992                 err = __blk_mq_alloc_rq_maps(set);
1993                 if (!err)
1994                         break;
1995
1996                 set->queue_depth >>= 1;
1997                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
1998                         err = -ENOMEM;
1999                         break;
2000                 }
2001         } while (set->queue_depth);
2002
2003         if (!set->queue_depth || err) {
2004                 pr_err("blk-mq: failed to allocate request map\n");
2005                 return -ENOMEM;
2006         }
2007
2008         if (depth != set->queue_depth)
2009                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2010                                                 depth, set->queue_depth);
2011
2012         return 0;
2013 }
2014
2015 /*
2016  * Alloc a tag set to be associated with one or more request queues.
2017  * May fail with EINVAL for various error conditions. May adjust the
2018  * requested depth down, if if it too large. In that case, the set
2019  * value will be stored in set->queue_depth.
2020  */
2021 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2022 {
2023         if (!set->nr_hw_queues)
2024                 return -EINVAL;
2025         if (!set->queue_depth)
2026                 return -EINVAL;
2027         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2028                 return -EINVAL;
2029
2030         if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2031                 return -EINVAL;
2032
2033         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2034                 pr_info("blk-mq: reduced tag depth to %u\n",
2035                         BLK_MQ_MAX_DEPTH);
2036                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2037         }
2038
2039         set->tags = kmalloc_node(set->nr_hw_queues *
2040                                  sizeof(struct blk_mq_tags *),
2041                                  GFP_KERNEL, set->numa_node);
2042         if (!set->tags)
2043                 return -ENOMEM;
2044
2045         if (blk_mq_alloc_rq_maps(set))
2046                 goto enomem;
2047
2048         mutex_init(&set->tag_list_lock);
2049         INIT_LIST_HEAD(&set->tag_list);
2050
2051         return 0;
2052 enomem:
2053         kfree(set->tags);
2054         set->tags = NULL;
2055         return -ENOMEM;
2056 }
2057 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2058
2059 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2060 {
2061         int i;
2062
2063         for (i = 0; i < set->nr_hw_queues; i++) {
2064                 if (set->tags[i])
2065                         blk_mq_free_rq_map(set, set->tags[i], i);
2066         }
2067
2068         kfree(set->tags);
2069         set->tags = NULL;
2070 }
2071 EXPORT_SYMBOL(blk_mq_free_tag_set);
2072
2073 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2074 {
2075         struct blk_mq_tag_set *set = q->tag_set;
2076         struct blk_mq_hw_ctx *hctx;
2077         int i, ret;
2078
2079         if (!set || nr > set->queue_depth)
2080                 return -EINVAL;
2081
2082         ret = 0;
2083         queue_for_each_hw_ctx(q, hctx, i) {
2084                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2085                 if (ret)
2086                         break;
2087         }
2088
2089         if (!ret)
2090                 q->nr_requests = nr;
2091
2092         return ret;
2093 }
2094
2095 void blk_mq_disable_hotplug(void)
2096 {
2097         mutex_lock(&all_q_mutex);
2098 }
2099
2100 void blk_mq_enable_hotplug(void)
2101 {
2102         mutex_unlock(&all_q_mutex);
2103 }
2104
2105 static int __init blk_mq_init(void)
2106 {
2107         blk_mq_cpu_init();
2108
2109         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2110
2111         return 0;
2112 }
2113 subsys_initcall(blk_mq_init);