drivers/block/null_blk.c

   1 #include <linux/module.h>
   2
   3 #include <linux/moduleparam.h>
   4 #include <linux/sched.h>
   5 #include <linux/fs.h>
   6 #include <linux/blkdev.h>
   7 #include <linux/init.h>
   8 #include <linux/slab.h>
   9 #include <linux/blk-mq.h>
  10 #include <linux/hrtimer.h>
  11
  12 struct nullb_cmd {
  13         struct list_head list;
  14         struct llist_node ll_list;
  15         struct call_single_data csd;
  16         struct request *rq;
  17         struct bio *bio;
  18         unsigned int tag;
  19         struct nullb_queue *nq;
  20 };
  21
  22 struct nullb_queue {
  23         unsigned long *tag_map;
  24         wait_queue_head_t wait;
  25         unsigned int queue_depth;
  26
  27         struct nullb_cmd *cmds;
  28 };
  29
  30 struct nullb {
  31         struct list_head list;
  32         unsigned int index;
  33         struct request_queue *q;
  34         struct gendisk *disk;
  35         struct blk_mq_tag_set tag_set;
  36         struct hrtimer timer;
  37         unsigned int queue_depth;
  38         spinlock_t lock;
  39
  40         struct nullb_queue *queues;
  41         unsigned int nr_queues;
  42 };
  43
  44 static LIST_HEAD(nullb_list);
  45 static struct mutex lock;
  46 static int null_major;
  47 static int nullb_indexes;
  48
  49 struct completion_queue {
  50         struct llist_head list;
  51         struct hrtimer timer;
  52 };
  53
  54 /*
  55  * These are per-cpu for now, they will need to be configured by the
  56  * complete_queues parameter and appropriately mapped.
  57  */
  58 static DEFINE_PER_CPU(struct completion_queue, completion_queues);
  59
  60 enum {
  61         NULL_IRQ_NONE           = 0,
  62         NULL_IRQ_SOFTIRQ        = 1,
  63         NULL_IRQ_TIMER          = 2,
  64 };
  65
  66 enum {
  67         NULL_Q_BIO              = 0,
  68         NULL_Q_RQ               = 1,
  69         NULL_Q_MQ               = 2,
  70 };
  71
  72 static int submit_queues;
  73 module_param(submit_queues, int, S_IRUGO);
  74 MODULE_PARM_DESC(submit_queues, "Number of submission queues");
  75
  76 static int home_node = NUMA_NO_NODE;
  77 module_param(home_node, int, S_IRUGO);
  78 MODULE_PARM_DESC(home_node, "Home node for the device");
  79
  80 static int queue_mode = NULL_Q_MQ;
  81 module_param(queue_mode, int, S_IRUGO);
  82 MODULE_PARM_DESC(use_mq, "Use blk-mq interface (0=bio,1=rq,2=multiqueue)");
  83
  84 static int gb = 250;
  85 module_param(gb, int, S_IRUGO);
  86 MODULE_PARM_DESC(gb, "Size in GB");
  87
  88 static int bs = 512;
  89 module_param(bs, int, S_IRUGO);
  90 MODULE_PARM_DESC(bs, "Block size (in bytes)");
  91
  92 static int nr_devices = 2;
  93 module_param(nr_devices, int, S_IRUGO);
  94 MODULE_PARM_DESC(nr_devices, "Number of devices to register");
  95
  96 static int irqmode = NULL_IRQ_SOFTIRQ;
  97 module_param(irqmode, int, S_IRUGO);
  98 MODULE_PARM_DESC(irqmode, "IRQ completion handler. 0-none, 1-softirq, 2-timer");
  99
 100 static int completion_nsec = 10000;
 101 module_param(completion_nsec, int, S_IRUGO);
 102 MODULE_PARM_DESC(completion_nsec, "Time in ns to complete a request in hardware. Default: 10,000ns");
 103
 104 static int hw_queue_depth = 64;
 105 module_param(hw_queue_depth, int, S_IRUGO);
 106 MODULE_PARM_DESC(hw_queue_depth, "Queue depth for each hardware queue. Default: 64");
 107
 108 static bool use_per_node_hctx = false;
 109 module_param(use_per_node_hctx, bool, S_IRUGO);
 110 MODULE_PARM_DESC(use_per_node_hctx, "Use per-node allocation for hardware context queues. Default: false");
 111
 112 static void put_tag(struct nullb_queue *nq, unsigned int tag)
 113 {
 114         clear_bit_unlock(tag, nq->tag_map);
 115
 116         if (waitqueue_active(&nq->wait))
 117                 wake_up(&nq->wait);
 118 }
 119
 120 static unsigned int get_tag(struct nullb_queue *nq)
 121 {
 122         unsigned int tag;
 123
 124         do {
 125                 tag = find_first_zero_bit(nq->tag_map, nq->queue_depth);
 126                 if (tag >= nq->queue_depth)
 127                         return -1U;
 128         } while (test_and_set_bit_lock(tag, nq->tag_map));
 129
 130         return tag;
 131 }
 132
 133 static void free_cmd(struct nullb_cmd *cmd)
 134 {
 135         put_tag(cmd->nq, cmd->tag);
 136 }
 137
 138 static struct nullb_cmd *__alloc_cmd(struct nullb_queue *nq)
 139 {
 140         struct nullb_cmd *cmd;
 141         unsigned int tag;
 142
 143         tag = get_tag(nq);
 144         if (tag != -1U) {
 145                 cmd = &nq->cmds[tag];
 146                 cmd->tag = tag;
 147                 cmd->nq = nq;
 148                 return cmd;
 149         }
 150
 151         return NULL;
 152 }
 153
 154 static struct nullb_cmd *alloc_cmd(struct nullb_queue *nq, int can_wait)
 155 {
 156         struct nullb_cmd *cmd;
 157         DEFINE_WAIT(wait);
 158
 159         cmd = __alloc_cmd(nq);
 160         if (cmd || !can_wait)
 161                 return cmd;
 162
 163         do {
 164                 prepare_to_wait(&nq->wait, &wait, TASK_UNINTERRUPTIBLE);
 165                 cmd = __alloc_cmd(nq);
 166                 if (cmd)
 167                         break;
 168
 169                 io_schedule();
 170         } while (1);
 171
 172         finish_wait(&nq->wait, &wait);
 173         return cmd;
 174 }
 175
 176 static void end_cmd(struct nullb_cmd *cmd)
 177 {
 178         switch (queue_mode)  {
 179         case NULL_Q_MQ:
 180                 blk_mq_end_io(cmd->rq, 0);
 181                 return;
 182         case NULL_Q_RQ:
 183                 INIT_LIST_HEAD(&cmd->rq->queuelist);
 184                 blk_end_request_all(cmd->rq, 0);
 185                 break;
 186         case NULL_Q_BIO:
 187                 bio_endio(cmd->bio, 0);
 188                 break;
 189         }
 190
 191         free_cmd(cmd);
 192 }
 193
 194 static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer)
 195 {
 196         struct completion_queue *cq;
 197         struct llist_node *entry;
 198         struct nullb_cmd *cmd;
 199
 200         cq = &per_cpu(completion_queues, smp_processor_id());
 201
 202         while ((entry = llist_del_all(&cq->list)) != NULL) {
 203                 entry = llist_reverse_order(entry);
 204                 do {
 205                         cmd = container_of(entry, struct nullb_cmd, ll_list);
 206                         entry = entry->next;
 207                         end_cmd(cmd);
 208                 } while (entry);
 209         }
 210
 211         return HRTIMER_NORESTART;
 212 }
 213
 214 static void null_cmd_end_timer(struct nullb_cmd *cmd)
 215 {
 216         struct completion_queue *cq = &per_cpu(completion_queues, get_cpu());
 217
 218         cmd->ll_list.next = NULL;
 219         if (llist_add(&cmd->ll_list, &cq->list)) {
 220                 ktime_t kt = ktime_set(0, completion_nsec);
 221
 222                 hrtimer_start(&cq->timer, kt, HRTIMER_MODE_REL);
 223         }
 224
 225         put_cpu();
 226 }
 227
 228 static void null_softirq_done_fn(struct request *rq)
 229 {
 230         end_cmd(blk_mq_rq_to_pdu(rq));
 231 }
 232
 233 static inline void null_handle_cmd(struct nullb_cmd *cmd)
 234 {
 235         /* Complete IO by inline, softirq or timer */
 236         switch (irqmode) {
 237         case NULL_IRQ_SOFTIRQ:
 238                 switch (queue_mode)  {
 239                 case NULL_Q_MQ:
 240                         blk_mq_complete_request(cmd->rq);
 241                         break;
 242                 case NULL_Q_RQ:
 243                         blk_complete_request(cmd->rq);
 244                         break;
 245                 case NULL_Q_BIO:
 246                         /*
 247                          * XXX: no proper submitting cpu information available.
 248                          */
 249                         end_cmd(cmd);
 250                         break;
 251                 }
 252                 break;
 253         case NULL_IRQ_NONE:
 254                 end_cmd(cmd);
 255                 break;
 256         case NULL_IRQ_TIMER:
 257                 null_cmd_end_timer(cmd);
 258                 break;
 259         }
 260 }
 261
 262 static struct nullb_queue *nullb_to_queue(struct nullb *nullb)
 263 {
 264         int index = 0;
 265
 266         if (nullb->nr_queues != 1)
 267                 index = raw_smp_processor_id() / ((nr_cpu_ids + nullb->nr_queues - 1) / nullb->nr_queues);
 268
 269         return &nullb->queues[index];
 270 }
 271
 272 static void null_queue_bio(struct request_queue *q, struct bio *bio)
 273 {
 274         struct nullb *nullb = q->queuedata;
 275         struct nullb_queue *nq = nullb_to_queue(nullb);
 276         struct nullb_cmd *cmd;
 277
 278         cmd = alloc_cmd(nq, 1);
 279         cmd->bio = bio;
 280
 281         null_handle_cmd(cmd);
 282 }
 283
 284 static int null_rq_prep_fn(struct request_queue *q, struct request *req)
 285 {
 286         struct nullb *nullb = q->queuedata;
 287         struct nullb_queue *nq = nullb_to_queue(nullb);
 288         struct nullb_cmd *cmd;
 289
 290         cmd = alloc_cmd(nq, 0);
 291         if (cmd) {
 292                 cmd->rq = req;
 293                 req->special = cmd;
 294                 return BLKPREP_OK;
 295         }
 296
 297         return BLKPREP_DEFER;
 298 }
 299
 300 static void null_request_fn(struct request_queue *q)
 301 {
 302         struct request *rq;
 303
 304         while ((rq = blk_fetch_request(q)) != NULL) {
 305                 struct nullb_cmd *cmd = rq->special;
 306
 307                 spin_unlock_irq(q->queue_lock);
 308                 null_handle_cmd(cmd);
 309                 spin_lock_irq(q->queue_lock);
 310         }
 311 }
 312
 313 static int null_queue_rq(struct blk_mq_hw_ctx *hctx, struct request *rq)
 314 {
 315         struct nullb_cmd *cmd = blk_mq_rq_to_pdu(rq);
 316
 317         cmd->rq = rq;
 318         cmd->nq = hctx->driver_data;
 319
 320         null_handle_cmd(cmd);
 321         return BLK_MQ_RQ_QUEUE_OK;
 322 }
 323
 324 static void null_init_queue(struct nullb *nullb, struct nullb_queue *nq)
 325 {
 326         BUG_ON(!nullb);
 327         BUG_ON(!nq);
 328
 329         init_waitqueue_head(&nq->wait);
 330         nq->queue_depth = nullb->queue_depth;
 331 }
 332
 333 static int null_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 334                           unsigned int index)
 335 {
 336         struct nullb *nullb = data;
 337         struct nullb_queue *nq = &nullb->queues[index];
 338
 339         hctx->driver_data = nq;
 340         null_init_queue(nullb, nq);
 341         nullb->nr_queues++;
 342
 343         return 0;
 344 }
 345
 346 static struct blk_mq_ops null_mq_ops = {
 347         .queue_rq       = null_queue_rq,
 348         .map_queue      = blk_mq_map_queue,
 349         .init_hctx      = null_init_hctx,
 350         .complete       = null_softirq_done_fn,
 351 };
 352
 353 static void null_del_dev(struct nullb *nullb)
 354 {
 355         list_del_init(&nullb->list);
 356
 357         del_gendisk(nullb->disk);
 358         blk_cleanup_queue(nullb->q);
 359         if (queue_mode == NULL_Q_MQ)
 360                 blk_mq_free_tag_set(&nullb->tag_set);
 361         put_disk(nullb->disk);
 362         kfree(nullb);
 363 }
 364
 365 static int null_open(struct block_device *bdev, fmode_t mode)
 366 {
 367         return 0;
 368 }
 369
 370 static void null_release(struct gendisk *disk, fmode_t mode)
 371 {
 372 }
 373
 374 static const struct block_device_operations null_fops = {
 375         .owner =        THIS_MODULE,
 376         .open =         null_open,
 377         .release =      null_release,
 378 };
 379
 380 static int setup_commands(struct nullb_queue *nq)
 381 {
 382         struct nullb_cmd *cmd;
 383         int i, tag_size;
 384
 385         nq->cmds = kzalloc(nq->queue_depth * sizeof(*cmd), GFP_KERNEL);
 386         if (!nq->cmds)
 387                 return -ENOMEM;
 388
 389         tag_size = ALIGN(nq->queue_depth, BITS_PER_LONG) / BITS_PER_LONG;
 390         nq->tag_map = kzalloc(tag_size * sizeof(unsigned long), GFP_KERNEL);
 391         if (!nq->tag_map) {
 392                 kfree(nq->cmds);
 393                 return -ENOMEM;
 394         }
 395
 396         for (i = 0; i < nq->queue_depth; i++) {
 397                 cmd = &nq->cmds[i];
 398                 INIT_LIST_HEAD(&cmd->list);
 399                 cmd->ll_list.next = NULL;
 400                 cmd->tag = -1U;
 401         }
 402
 403         return 0;
 404 }
 405
 406 static void cleanup_queue(struct nullb_queue *nq)
 407 {
 408         kfree(nq->tag_map);
 409         kfree(nq->cmds);
 410 }
 411
 412 static void cleanup_queues(struct nullb *nullb)
 413 {
 414         int i;
 415
 416         for (i = 0; i < nullb->nr_queues; i++)
 417                 cleanup_queue(&nullb->queues[i]);
 418
 419         kfree(nullb->queues);
 420 }
 421
 422 static int setup_queues(struct nullb *nullb)
 423 {
 424         nullb->queues = kzalloc(submit_queues * sizeof(struct nullb_queue),
 425                                                                 GFP_KERNEL);
 426         if (!nullb->queues)
 427                 return -ENOMEM;
 428
 429         nullb->nr_queues = 0;
 430         nullb->queue_depth = hw_queue_depth;
 431
 432         return 0;
 433 }
 434
 435 static int init_driver_queues(struct nullb *nullb)
 436 {
 437         struct nullb_queue *nq;
 438         int i, ret = 0;
 439
 440         for (i = 0; i < submit_queues; i++) {
 441                 nq = &nullb->queues[i];
 442
 443                 null_init_queue(nullb, nq);
 444
 445                 ret = setup_commands(nq);
 446                 if (ret)
 447                         goto err_queue;
 448                 nullb->nr_queues++;
 449         }
 450
 451         return 0;
 452 err_queue:
 453         cleanup_queues(nullb);
 454         return ret;
 455 }
 456
 457 static int null_add_dev(void)
 458 {
 459         struct gendisk *disk;
 460         struct nullb *nullb;
 461         sector_t size;
 462
 463         nullb = kzalloc_node(sizeof(*nullb), GFP_KERNEL, home_node);
 464         if (!nullb)
 465                 goto out;
 466
 467         spin_lock_init(&nullb->lock);
 468
 469         if (queue_mode == NULL_Q_MQ && use_per_node_hctx)
 470                 submit_queues = nr_online_nodes;
 471
 472         if (setup_queues(nullb))
 473                 goto out_free_nullb;
 474
 475         if (queue_mode == NULL_Q_MQ) {
 476                 nullb->tag_set.ops = &null_mq_ops;
 477                 nullb->tag_set.nr_hw_queues = submit_queues;
 478                 nullb->tag_set.queue_depth = hw_queue_depth;
 479                 nullb->tag_set.numa_node = home_node;
 480                 nullb->tag_set.cmd_size = sizeof(struct nullb_cmd);
 481                 nullb->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 482                 nullb->tag_set.driver_data = nullb;
 483
 484                 if (blk_mq_alloc_tag_set(&nullb->tag_set))
 485                         goto out_cleanup_queues;
 486
 487                 nullb->q = blk_mq_init_queue(&nullb->tag_set);
 488                 if (!nullb->q)
 489                         goto out_cleanup_tags;
 490         } else if (queue_mode == NULL_Q_BIO) {
 491                 nullb->q = blk_alloc_queue_node(GFP_KERNEL, home_node);
 492                 if (!nullb->q)
 493                         goto out_cleanup_queues;
 494                 blk_queue_make_request(nullb->q, null_queue_bio);
 495                 init_driver_queues(nullb);
 496         } else {
 497                 nullb->q = blk_init_queue_node(null_request_fn, &nullb->lock, home_node);
 498                 if (!nullb->q)
 499                         goto out_cleanup_queues;
 500                 blk_queue_prep_rq(nullb->q, null_rq_prep_fn);
 501                 blk_queue_softirq_done(nullb->q, null_softirq_done_fn);
 502                 init_driver_queues(nullb);
 503         }
 504
 505         nullb->q->queuedata = nullb;
 506         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, nullb->q);
 507
 508         disk = nullb->disk = alloc_disk_node(1, home_node);
 509         if (!disk)
 510                 goto out_cleanup_blk_queue;
 511
 512         mutex_lock(&lock);
 513         list_add_tail(&nullb->list, &nullb_list);
 514         nullb->index = nullb_indexes++;
 515         mutex_unlock(&lock);
 516
 517         blk_queue_logical_block_size(nullb->q, bs);
 518         blk_queue_physical_block_size(nullb->q, bs);
 519
 520         size = gb * 1024 * 1024 * 1024ULL;
 521         sector_div(size, bs);
 522         set_capacity(disk, size);
 523
 524         disk->flags |= GENHD_FL_EXT_DEVT;
 525         disk->major             = null_major;
 526         disk->first_minor       = nullb->index;
 527         disk->fops              = &null_fops;
 528         disk->private_data      = nullb;
 529         disk->queue             = nullb->q;
 530         sprintf(disk->disk_name, "nullb%d", nullb->index);
 531         add_disk(disk);
 532         return 0;
 533
 534 out_cleanup_blk_queue:
 535         blk_cleanup_queue(nullb->q);
 536 out_cleanup_tags:
 537         if (queue_mode == NULL_Q_MQ)
 538                 blk_mq_free_tag_set(&nullb->tag_set);
 539 out_cleanup_queues:
 540         cleanup_queues(nullb);
 541 out_free_nullb:
 542         kfree(nullb);
 543 out:
 544         return -ENOMEM;
 545 }
 546
 547 static int __init null_init(void)
 548 {
 549         unsigned int i;
 550
 551         if (bs > PAGE_SIZE) {
 552                 pr_warn("null_blk: invalid block size\n");
 553                 pr_warn("null_blk: defaults block size to %lu\n", PAGE_SIZE);
 554                 bs = PAGE_SIZE;
 555         }
 556
 557         if (queue_mode == NULL_Q_MQ && use_per_node_hctx) {
 558                 if (submit_queues < nr_online_nodes) {
 559                         pr_warn("null_blk: submit_queues param is set to %u.",
 560                                                         nr_online_nodes);
 561                         submit_queues = nr_online_nodes;
 562                 }
 563         } else if (submit_queues > nr_cpu_ids)
 564                 submit_queues = nr_cpu_ids;
 565         else if (!submit_queues)
 566                 submit_queues = 1;
 567
 568         mutex_init(&lock);
 569
 570         /* Initialize a separate list for each CPU for issuing softirqs */
 571         for_each_possible_cpu(i) {
 572                 struct completion_queue *cq = &per_cpu(completion_queues, i);
 573
 574                 init_llist_head(&cq->list);
 575
 576                 if (irqmode != NULL_IRQ_TIMER)
 577                         continue;
 578
 579                 hrtimer_init(&cq->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 580                 cq->timer.function = null_cmd_timer_expired;
 581         }
 582
 583         null_major = register_blkdev(0, "nullb");
 584         if (null_major < 0)
 585                 return null_major;
 586
 587         for (i = 0; i < nr_devices; i++) {
 588                 if (null_add_dev()) {
 589                         unregister_blkdev(null_major, "nullb");
 590                         return -EINVAL;
 591                 }
 592         }
 593
 594         pr_info("null: module loaded\n");
 595         return 0;
 596 }
 597
 598 static void __exit null_exit(void)
 599 {
 600         struct nullb *nullb;
 601
 602         unregister_blkdev(null_major, "nullb");
 603
 604         mutex_lock(&lock);
 605         while (!list_empty(&nullb_list)) {
 606                 nullb = list_entry(nullb_list.next, struct nullb, list);
 607                 null_del_dev(nullb);
 608         }
 609         mutex_unlock(&lock);
 610 }
 611
 612 module_init(null_init);
 613 module_exit(null_exit);
 614
 615 MODULE_AUTHOR("Jens Axboe <jaxboe@fusionio.com>");
 616 MODULE_LICENSE("GPL");