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 hrtimer timer;
  36         unsigned int queue_depth;
  37         spinlock_t lock;
  38
  39         struct nullb_queue *queues;
  40         unsigned int nr_queues;
  41 };
  42
  43 static LIST_HEAD(nullb_list);
  44 static struct mutex lock;
  45 static int null_major;
  46 static int nullb_indexes;
  47
  48 struct completion_queue {
  49         struct llist_head list;
  50         struct hrtimer timer;
  51 };
  52
  53 /*
  54  * These are per-cpu for now, they will need to be configured by the
  55  * complete_queues parameter and appropriately mapped.
  56  */
  57 static DEFINE_PER_CPU(struct completion_queue, completion_queues);
  58
  59 enum {
  60         NULL_IRQ_NONE           = 0,
  61         NULL_IRQ_SOFTIRQ        = 1,
  62         NULL_IRQ_TIMER          = 2,
  63
  64         NULL_Q_BIO              = 0,
  65         NULL_Q_RQ               = 1,
  66         NULL_Q_MQ               = 2,
  67 };
  68
  69 static int submit_queues;
  70 module_param(submit_queues, int, S_IRUGO);
  71 MODULE_PARM_DESC(submit_queues, "Number of submission queues");
  72
  73 static int home_node = NUMA_NO_NODE;
  74 module_param(home_node, int, S_IRUGO);
  75 MODULE_PARM_DESC(home_node, "Home node for the device");
  76
  77 static int queue_mode = NULL_Q_MQ;
  78 module_param(queue_mode, int, S_IRUGO);
  79 MODULE_PARM_DESC(use_mq, "Use blk-mq interface (0=bio,1=rq,2=multiqueue)");
  80
  81 static int gb = 250;
  82 module_param(gb, int, S_IRUGO);
  83 MODULE_PARM_DESC(gb, "Size in GB");
  84
  85 static int bs = 512;
  86 module_param(bs, int, S_IRUGO);
  87 MODULE_PARM_DESC(bs, "Block size (in bytes)");
  88
  89 static int nr_devices = 2;
  90 module_param(nr_devices, int, S_IRUGO);
  91 MODULE_PARM_DESC(nr_devices, "Number of devices to register");
  92
  93 static int irqmode = NULL_IRQ_SOFTIRQ;
  94 module_param(irqmode, int, S_IRUGO);
  95 MODULE_PARM_DESC(irqmode, "IRQ completion handler. 0-none, 1-softirq, 2-timer");
  96
  97 static int completion_nsec = 10000;
  98 module_param(completion_nsec, int, S_IRUGO);
  99 MODULE_PARM_DESC(completion_nsec, "Time in ns to complete a request in hardware. Default: 10,000ns");
 100
 101 static int hw_queue_depth = 64;
 102 module_param(hw_queue_depth, int, S_IRUGO);
 103 MODULE_PARM_DESC(hw_queue_depth, "Queue depth for each hardware queue. Default: 64");
 104
 105 static bool use_per_node_hctx = false;
 106 module_param(use_per_node_hctx, bool, S_IRUGO);
 107 MODULE_PARM_DESC(use_per_node_hctx, "Use per-node allocation for hardware context queues. Default: false");
 108
 109 static void put_tag(struct nullb_queue *nq, unsigned int tag)
 110 {
 111         clear_bit_unlock(tag, nq->tag_map);
 112
 113         if (waitqueue_active(&nq->wait))
 114                 wake_up(&nq->wait);
 115 }
 116
 117 static unsigned int get_tag(struct nullb_queue *nq)
 118 {
 119         unsigned int tag;
 120
 121         do {
 122                 tag = find_first_zero_bit(nq->tag_map, nq->queue_depth);
 123                 if (tag >= nq->queue_depth)
 124                         return -1U;
 125         } while (test_and_set_bit_lock(tag, nq->tag_map));
 126
 127         return tag;
 128 }
 129
 130 static void free_cmd(struct nullb_cmd *cmd)
 131 {
 132         put_tag(cmd->nq, cmd->tag);
 133 }
 134
 135 static struct nullb_cmd *__alloc_cmd(struct nullb_queue *nq)
 136 {
 137         struct nullb_cmd *cmd;
 138         unsigned int tag;
 139
 140         tag = get_tag(nq);
 141         if (tag != -1U) {
 142                 cmd = &nq->cmds[tag];
 143                 cmd->tag = tag;
 144                 cmd->nq = nq;
 145                 return cmd;
 146         }
 147
 148         return NULL;
 149 }
 150
 151 static struct nullb_cmd *alloc_cmd(struct nullb_queue *nq, int can_wait)
 152 {
 153         struct nullb_cmd *cmd;
 154         DEFINE_WAIT(wait);
 155
 156         cmd = __alloc_cmd(nq);
 157         if (cmd || !can_wait)
 158                 return cmd;
 159
 160         do {
 161                 prepare_to_wait(&nq->wait, &wait, TASK_UNINTERRUPTIBLE);
 162                 cmd = __alloc_cmd(nq);
 163                 if (cmd)
 164                         break;
 165
 166                 io_schedule();
 167         } while (1);
 168
 169         finish_wait(&nq->wait, &wait);
 170         return cmd;
 171 }
 172
 173 static void end_cmd(struct nullb_cmd *cmd)
 174 {
 175         if (cmd->rq) {
 176                 if (queue_mode == NULL_Q_MQ)
 177                         blk_mq_end_io(cmd->rq, 0);
 178                 else {
 179                         INIT_LIST_HEAD(&cmd->rq->queuelist);
 180                         blk_end_request_all(cmd->rq, 0);
 181                 }
 182         } else if (cmd->bio)
 183                 bio_endio(cmd->bio, 0);
 184
 185         if (queue_mode != NULL_Q_MQ)
 186                 free_cmd(cmd);
 187 }
 188
 189 static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer)
 190 {
 191         struct completion_queue *cq;
 192         struct llist_node *entry;
 193         struct nullb_cmd *cmd;
 194
 195         cq = &per_cpu(completion_queues, smp_processor_id());
 196
 197         while ((entry = llist_del_all(&cq->list)) != NULL) {
 198                 entry = llist_reverse_order(entry);
 199                 do {
 200                         cmd = container_of(entry, struct nullb_cmd, ll_list);
 201                         end_cmd(cmd);
 202                         entry = entry->next;
 203                 } while (entry);
 204         }
 205
 206         return HRTIMER_NORESTART;
 207 }
 208
 209 static void null_cmd_end_timer(struct nullb_cmd *cmd)
 210 {
 211         struct completion_queue *cq = &per_cpu(completion_queues, get_cpu());
 212
 213         cmd->ll_list.next = NULL;
 214         if (llist_add(&cmd->ll_list, &cq->list)) {
 215                 ktime_t kt = ktime_set(0, completion_nsec);
 216
 217                 hrtimer_start(&cq->timer, kt, HRTIMER_MODE_REL);
 218         }
 219
 220         put_cpu();
 221 }
 222
 223 static void null_softirq_done_fn(struct request *rq)
 224 {
 225         blk_end_request_all(rq, 0);
 226 }
 227
 228 #ifdef CONFIG_SMP
 229
 230 static void null_ipi_cmd_end_io(void *data)
 231 {
 232         struct completion_queue *cq;
 233         struct llist_node *entry, *next;
 234         struct nullb_cmd *cmd;
 235
 236         cq = &per_cpu(completion_queues, smp_processor_id());
 237
 238         entry = llist_del_all(&cq->list);
 239         entry = llist_reverse_order(entry);
 240
 241         while (entry) {
 242                 next = entry->next;
 243                 cmd = llist_entry(entry, struct nullb_cmd, ll_list);
 244                 end_cmd(cmd);
 245                 entry = next;
 246         }
 247 }
 248
 249 static void null_cmd_end_ipi(struct nullb_cmd *cmd)
 250 {
 251         struct call_single_data *data = &cmd->csd;
 252         int cpu = get_cpu();
 253         struct completion_queue *cq = &per_cpu(completion_queues, cpu);
 254
 255         cmd->ll_list.next = NULL;
 256
 257         if (llist_add(&cmd->ll_list, &cq->list)) {
 258                 data->func = null_ipi_cmd_end_io;
 259                 data->flags = 0;
 260                 __smp_call_function_single(cpu, data, 0);
 261         }
 262
 263         put_cpu();
 264 }
 265
 266 #endif /* CONFIG_SMP */
 267
 268 static inline void null_handle_cmd(struct nullb_cmd *cmd)
 269 {
 270         /* Complete IO by inline, softirq or timer */
 271         switch (irqmode) {
 272         case NULL_IRQ_NONE:
 273                 end_cmd(cmd);
 274                 break;
 275         case NULL_IRQ_SOFTIRQ:
 276 #ifdef CONFIG_SMP
 277                 null_cmd_end_ipi(cmd);
 278 #else
 279                 end_cmd(cmd);
 280 #endif
 281                 break;
 282         case NULL_IRQ_TIMER:
 283                 null_cmd_end_timer(cmd);
 284                 break;
 285         }
 286 }
 287
 288 static struct nullb_queue *nullb_to_queue(struct nullb *nullb)
 289 {
 290         int index = 0;
 291
 292         if (nullb->nr_queues != 1)
 293                 index = raw_smp_processor_id() / ((nr_cpu_ids + nullb->nr_queues - 1) / nullb->nr_queues);
 294
 295         return &nullb->queues[index];
 296 }
 297
 298 static void null_queue_bio(struct request_queue *q, struct bio *bio)
 299 {
 300         struct nullb *nullb = q->queuedata;
 301         struct nullb_queue *nq = nullb_to_queue(nullb);
 302         struct nullb_cmd *cmd;
 303
 304         cmd = alloc_cmd(nq, 1);
 305         cmd->bio = bio;
 306
 307         null_handle_cmd(cmd);
 308 }
 309
 310 static int null_rq_prep_fn(struct request_queue *q, struct request *req)
 311 {
 312         struct nullb *nullb = q->queuedata;
 313         struct nullb_queue *nq = nullb_to_queue(nullb);
 314         struct nullb_cmd *cmd;
 315
 316         cmd = alloc_cmd(nq, 0);
 317         if (cmd) {
 318                 cmd->rq = req;
 319                 req->special = cmd;
 320                 return BLKPREP_OK;
 321         }
 322
 323         return BLKPREP_DEFER;
 324 }
 325
 326 static void null_request_fn(struct request_queue *q)
 327 {
 328         struct request *rq;
 329
 330         while ((rq = blk_fetch_request(q)) != NULL) {
 331                 struct nullb_cmd *cmd = rq->special;
 332
 333                 spin_unlock_irq(q->queue_lock);
 334                 null_handle_cmd(cmd);
 335                 spin_lock_irq(q->queue_lock);
 336         }
 337 }
 338
 339 static int null_queue_rq(struct blk_mq_hw_ctx *hctx, struct request *rq)
 340 {
 341         struct nullb_cmd *cmd = rq->special;
 342
 343         cmd->rq = rq;
 344         cmd->nq = hctx->driver_data;
 345
 346         null_handle_cmd(cmd);
 347         return BLK_MQ_RQ_QUEUE_OK;
 348 }
 349
 350 static struct blk_mq_hw_ctx *null_alloc_hctx(struct blk_mq_reg *reg, unsigned int hctx_index)
 351 {
 352         int b_size = DIV_ROUND_UP(reg->nr_hw_queues, nr_online_nodes);
 353         int tip = (reg->nr_hw_queues % nr_online_nodes);
 354         int node = 0, i, n;
 355
 356         /*
 357          * Split submit queues evenly wrt to the number of nodes. If uneven,
 358          * fill the first buckets with one extra, until the rest is filled with
 359          * no extra.
 360          */
 361         for (i = 0, n = 1; i < hctx_index; i++, n++) {
 362                 if (n % b_size == 0) {
 363                         n = 0;
 364                         node++;
 365
 366                         tip--;
 367                         if (!tip)
 368                                 b_size = reg->nr_hw_queues / nr_online_nodes;
 369                 }
 370         }
 371
 372         /*
 373          * A node might not be online, therefore map the relative node id to the
 374          * real node id.
 375          */
 376         for_each_online_node(n) {
 377                 if (!node)
 378                         break;
 379                 node--;
 380         }
 381
 382         return kzalloc_node(sizeof(struct blk_mq_hw_ctx), GFP_KERNEL, n);
 383 }
 384
 385 static void null_free_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_index)
 386 {
 387         kfree(hctx);
 388 }
 389
 390 static void null_init_queue(struct nullb *nullb, struct nullb_queue *nq)
 391 {
 392         BUG_ON(!nullb);
 393         BUG_ON(!nq);
 394
 395         init_waitqueue_head(&nq->wait);
 396         nq->queue_depth = nullb->queue_depth;
 397 }
 398
 399 static int null_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 400                           unsigned int index)
 401 {
 402         struct nullb *nullb = data;
 403         struct nullb_queue *nq = &nullb->queues[index];
 404
 405         hctx->driver_data = nq;
 406         null_init_queue(nullb, nq);
 407         nullb->nr_queues++;
 408
 409         return 0;
 410 }
 411
 412 static struct blk_mq_ops null_mq_ops = {
 413         .queue_rq       = null_queue_rq,
 414         .map_queue      = blk_mq_map_queue,
 415         .init_hctx      = null_init_hctx,
 416 };
 417
 418 static struct blk_mq_reg null_mq_reg = {
 419         .ops            = &null_mq_ops,
 420         .queue_depth    = 64,
 421         .cmd_size       = sizeof(struct nullb_cmd),
 422         .flags          = BLK_MQ_F_SHOULD_MERGE,
 423 };
 424
 425 static void null_del_dev(struct nullb *nullb)
 426 {
 427         list_del_init(&nullb->list);
 428
 429         del_gendisk(nullb->disk);
 430         blk_cleanup_queue(nullb->q);
 431         put_disk(nullb->disk);
 432         kfree(nullb);
 433 }
 434
 435 static int null_open(struct block_device *bdev, fmode_t mode)
 436 {
 437         return 0;
 438 }
 439
 440 static void null_release(struct gendisk *disk, fmode_t mode)
 441 {
 442 }
 443
 444 static const struct block_device_operations null_fops = {
 445         .owner =        THIS_MODULE,
 446         .open =         null_open,
 447         .release =      null_release,
 448 };
 449
 450 static int setup_commands(struct nullb_queue *nq)
 451 {
 452         struct nullb_cmd *cmd;
 453         int i, tag_size;
 454
 455         nq->cmds = kzalloc(nq->queue_depth * sizeof(*cmd), GFP_KERNEL);
 456         if (!nq->cmds)
 457                 return -ENOMEM;
 458
 459         tag_size = ALIGN(nq->queue_depth, BITS_PER_LONG) / BITS_PER_LONG;
 460         nq->tag_map = kzalloc(tag_size * sizeof(unsigned long), GFP_KERNEL);
 461         if (!nq->tag_map) {
 462                 kfree(nq->cmds);
 463                 return -ENOMEM;
 464         }
 465
 466         for (i = 0; i < nq->queue_depth; i++) {
 467                 cmd = &nq->cmds[i];
 468                 INIT_LIST_HEAD(&cmd->list);
 469                 cmd->ll_list.next = NULL;
 470                 cmd->tag = -1U;
 471         }
 472
 473         return 0;
 474 }
 475
 476 static void cleanup_queue(struct nullb_queue *nq)
 477 {
 478         kfree(nq->tag_map);
 479         kfree(nq->cmds);
 480 }
 481
 482 static void cleanup_queues(struct nullb *nullb)
 483 {
 484         int i;
 485
 486         for (i = 0; i < nullb->nr_queues; i++)
 487                 cleanup_queue(&nullb->queues[i]);
 488
 489         kfree(nullb->queues);
 490 }
 491
 492 static int setup_queues(struct nullb *nullb)
 493 {
 494         nullb->queues = kzalloc(submit_queues * sizeof(struct nullb_queue),
 495                                                                 GFP_KERNEL);
 496         if (!nullb->queues)
 497                 return -ENOMEM;
 498
 499         nullb->nr_queues = 0;
 500         nullb->queue_depth = hw_queue_depth;
 501
 502         return 0;
 503 }
 504
 505 static int init_driver_queues(struct nullb *nullb)
 506 {
 507         struct nullb_queue *nq;
 508         int i, ret = 0;
 509
 510         for (i = 0; i < submit_queues; i++) {
 511                 nq = &nullb->queues[i];
 512
 513                 null_init_queue(nullb, nq);
 514
 515                 ret = setup_commands(nq);
 516                 if (ret)
 517                         goto err_queue;
 518                 nullb->nr_queues++;
 519         }
 520
 521         return 0;
 522 err_queue:
 523         cleanup_queues(nullb);
 524         return ret;
 525 }
 526
 527 static int null_add_dev(void)
 528 {
 529         struct gendisk *disk;
 530         struct nullb *nullb;
 531         sector_t size;
 532
 533         nullb = kzalloc_node(sizeof(*nullb), GFP_KERNEL, home_node);
 534         if (!nullb)
 535                 return -ENOMEM;
 536
 537         spin_lock_init(&nullb->lock);
 538
 539         if (queue_mode == NULL_Q_MQ && use_per_node_hctx)
 540                 submit_queues = nr_online_nodes;
 541
 542         if (setup_queues(nullb))
 543                 goto err;
 544
 545         if (queue_mode == NULL_Q_MQ) {
 546                 null_mq_reg.numa_node = home_node;
 547                 null_mq_reg.queue_depth = hw_queue_depth;
 548                 null_mq_reg.nr_hw_queues = submit_queues;
 549
 550                 if (use_per_node_hctx) {
 551                         null_mq_reg.ops->alloc_hctx = null_alloc_hctx;
 552                         null_mq_reg.ops->free_hctx = null_free_hctx;
 553                 } else {
 554                         null_mq_reg.ops->alloc_hctx = blk_mq_alloc_single_hw_queue;
 555                         null_mq_reg.ops->free_hctx = blk_mq_free_single_hw_queue;
 556                 }
 557
 558                 nullb->q = blk_mq_init_queue(&null_mq_reg, nullb);
 559         } else if (queue_mode == NULL_Q_BIO) {
 560                 nullb->q = blk_alloc_queue_node(GFP_KERNEL, home_node);
 561                 blk_queue_make_request(nullb->q, null_queue_bio);
 562                 init_driver_queues(nullb);
 563         } else {
 564                 nullb->q = blk_init_queue_node(null_request_fn, &nullb->lock, home_node);
 565                 blk_queue_prep_rq(nullb->q, null_rq_prep_fn);
 566                 if (nullb->q)
 567                         blk_queue_softirq_done(nullb->q, null_softirq_done_fn);
 568                 init_driver_queues(nullb);
 569         }
 570
 571         if (!nullb->q)
 572                 goto queue_fail;
 573
 574         nullb->q->queuedata = nullb;
 575         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, nullb->q);
 576
 577         disk = nullb->disk = alloc_disk_node(1, home_node);
 578         if (!disk) {
 579 queue_fail:
 580                 blk_cleanup_queue(nullb->q);
 581                 cleanup_queues(nullb);
 582 err:
 583                 kfree(nullb);
 584                 return -ENOMEM;
 585         }
 586
 587         mutex_lock(&lock);
 588         list_add_tail(&nullb->list, &nullb_list);
 589         nullb->index = nullb_indexes++;
 590         mutex_unlock(&lock);
 591
 592         blk_queue_logical_block_size(nullb->q, bs);
 593         blk_queue_physical_block_size(nullb->q, bs);
 594
 595         size = gb * 1024 * 1024 * 1024ULL;
 596         sector_div(size, bs);
 597         set_capacity(disk, size);
 598
 599         disk->flags |= GENHD_FL_EXT_DEVT;
 600         disk->major             = null_major;
 601         disk->first_minor       = nullb->index;
 602         disk->fops              = &null_fops;
 603         disk->private_data      = nullb;
 604         disk->queue             = nullb->q;
 605         sprintf(disk->disk_name, "nullb%d", nullb->index);
 606         add_disk(disk);
 607         return 0;
 608 }
 609
 610 static int __init null_init(void)
 611 {
 612         unsigned int i;
 613
 614 #if !defined(CONFIG_SMP)
 615         if (irqmode == NULL_IRQ_SOFTIRQ) {
 616                 pr_warn("null_blk: softirq completions not available.\n");
 617                 pr_warn("null_blk: using direct completions.\n");
 618                 irqmode = NULL_IRQ_NONE;
 619         }
 620 #endif
 621         if (bs > PAGE_SIZE) {
 622                 pr_warn("null_blk: invalid block size\n");
 623                 pr_warn("null_blk: defaults block size to %lu\n", PAGE_SIZE);
 624                 bs = PAGE_SIZE;
 625         }
 626
 627         if (queue_mode == NULL_Q_MQ && use_per_node_hctx) {
 628                 if (submit_queues < nr_online_nodes) {
 629                         pr_warn("null_blk: submit_queues param is set to %u.",
 630                                                         nr_online_nodes);
 631                         submit_queues = nr_online_nodes;
 632                 }
 633         } else if (submit_queues > nr_cpu_ids)
 634                 submit_queues = nr_cpu_ids;
 635         else if (!submit_queues)
 636                 submit_queues = 1;
 637
 638         mutex_init(&lock);
 639
 640         /* Initialize a separate list for each CPU for issuing softirqs */
 641         for_each_possible_cpu(i) {
 642                 struct completion_queue *cq = &per_cpu(completion_queues, i);
 643
 644                 init_llist_head(&cq->list);
 645
 646                 if (irqmode != NULL_IRQ_TIMER)
 647                         continue;
 648
 649                 hrtimer_init(&cq->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 650                 cq->timer.function = null_cmd_timer_expired;
 651         }
 652
 653         null_major = register_blkdev(0, "nullb");
 654         if (null_major < 0)
 655                 return null_major;
 656
 657         for (i = 0; i < nr_devices; i++) {
 658                 if (null_add_dev()) {
 659                         unregister_blkdev(null_major, "nullb");
 660                         return -EINVAL;
 661                 }
 662         }
 663
 664         pr_info("null: module loaded\n");
 665         return 0;
 666 }
 667
 668 static void __exit null_exit(void)
 669 {
 670         struct nullb *nullb;
 671
 672         unregister_blkdev(null_major, "nullb");
 673
 674         mutex_lock(&lock);
 675         while (!list_empty(&nullb_list)) {
 676                 nullb = list_entry(nullb_list.next, struct nullb, list);
 677                 null_del_dev(nullb);
 678         }
 679         mutex_unlock(&lock);
 680 }
 681
 682 module_init(null_init);
 683 module_exit(null_exit);
 684
 685 MODULE_AUTHOR("Jens Axboe <jaxboe@fusionio.com>");
 686 MODULE_LICENSE("GPL");