drivers/block/null_blk_main.c

   1 /*
   2  * Add configfs and memory store: Kyungchan Koh <kkc6196@fb.com> and
   3  * Shaohua Li <shli@fb.com>
   4  */
   5 #include <linux/module.h>
   6
   7 #include <linux/moduleparam.h>
   8 #include <linux/sched.h>
   9 #include <linux/fs.h>
  10 #include <linux/init.h>
  11 #include "null_blk.h"
  12
  13 #define PAGE_SECTORS_SHIFT      (PAGE_SHIFT - SECTOR_SHIFT)
  14 #define PAGE_SECTORS            (1 << PAGE_SECTORS_SHIFT)
  15 #define SECTOR_MASK             (PAGE_SECTORS - 1)
  16
  17 #define FREE_BATCH              16
  18
  19 #define TICKS_PER_SEC           50ULL
  20 #define TIMER_INTERVAL          (NSEC_PER_SEC / TICKS_PER_SEC)
  21
  22 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
  23 static DECLARE_FAULT_ATTR(null_timeout_attr);
  24 static DECLARE_FAULT_ATTR(null_requeue_attr);
  25 #endif
  26
  27 static inline u64 mb_per_tick(int mbps)
  28 {
  29         return (1 << 20) / TICKS_PER_SEC * ((u64) mbps);
  30 }
  31
  32 /*
  33  * Status flags for nullb_device.
  34  *
  35  * CONFIGURED:  Device has been configured and turned on. Cannot reconfigure.
  36  * UP:          Device is currently on and visible in userspace.
  37  * THROTTLED:   Device is being throttled.
  38  * CACHE:       Device is using a write-back cache.
  39  */
  40 enum nullb_device_flags {
  41         NULLB_DEV_FL_CONFIGURED = 0,
  42         NULLB_DEV_FL_UP         = 1,
  43         NULLB_DEV_FL_THROTTLED  = 2,
  44         NULLB_DEV_FL_CACHE      = 3,
  45 };
  46
  47 #define MAP_SZ          ((PAGE_SIZE >> SECTOR_SHIFT) + 2)
  48 /*
  49  * nullb_page is a page in memory for nullb devices.
  50  *
  51  * @page:       The page holding the data.
  52  * @bitmap:     The bitmap represents which sector in the page has data.
  53  *              Each bit represents one block size. For example, sector 8
  54  *              will use the 7th bit
  55  * The highest 2 bits of bitmap are for special purpose. LOCK means the cache
  56  * page is being flushing to storage. FREE means the cache page is freed and
  57  * should be skipped from flushing to storage. Please see
  58  * null_make_cache_space
  59  */
  60 struct nullb_page {
  61         struct page *page;
  62         DECLARE_BITMAP(bitmap, MAP_SZ);
  63 };
  64 #define NULLB_PAGE_LOCK (MAP_SZ - 1)
  65 #define NULLB_PAGE_FREE (MAP_SZ - 2)
  66
  67 static LIST_HEAD(nullb_list);
  68 static struct mutex lock;
  69 static int null_major;
  70 static DEFINE_IDA(nullb_indexes);
  71 static struct blk_mq_tag_set tag_set;
  72
  73 enum {
  74         NULL_IRQ_NONE           = 0,
  75         NULL_IRQ_SOFTIRQ        = 1,
  76         NULL_IRQ_TIMER          = 2,
  77 };
  78
  79 enum {
  80         NULL_Q_BIO              = 0,
  81         NULL_Q_RQ               = 1,
  82         NULL_Q_MQ               = 2,
  83 };
  84
  85 static int g_no_sched;
  86 module_param_named(no_sched, g_no_sched, int, 0444);
  87 MODULE_PARM_DESC(no_sched, "No io scheduler");
  88
  89 static int g_submit_queues = 1;
  90 module_param_named(submit_queues, g_submit_queues, int, 0444);
  91 MODULE_PARM_DESC(submit_queues, "Number of submission queues");
  92
  93 static int g_home_node = NUMA_NO_NODE;
  94 module_param_named(home_node, g_home_node, int, 0444);
  95 MODULE_PARM_DESC(home_node, "Home node for the device");
  96
  97 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
  98 static char g_timeout_str[80];
  99 module_param_string(timeout, g_timeout_str, sizeof(g_timeout_str), 0444);
 100
 101 static char g_requeue_str[80];
 102 module_param_string(requeue, g_requeue_str, sizeof(g_requeue_str), 0444);
 103 #endif
 104
 105 static int g_queue_mode = NULL_Q_MQ;
 106
 107 static int null_param_store_val(const char *str, int *val, int min, int max)
 108 {
 109         int ret, new_val;
 110
 111         ret = kstrtoint(str, 10, &new_val);
 112         if (ret)
 113                 return -EINVAL;
 114
 115         if (new_val < min || new_val > max)
 116                 return -EINVAL;
 117
 118         *val = new_val;
 119         return 0;
 120 }
 121
 122 static int null_set_queue_mode(const char *str, const struct kernel_param *kp)
 123 {
 124         return null_param_store_val(str, &g_queue_mode, NULL_Q_BIO, NULL_Q_MQ);
 125 }
 126
 127 static const struct kernel_param_ops null_queue_mode_param_ops = {
 128         .set    = null_set_queue_mode,
 129         .get    = param_get_int,
 130 };
 131
 132 device_param_cb(queue_mode, &null_queue_mode_param_ops, &g_queue_mode, 0444);
 133 MODULE_PARM_DESC(queue_mode, "Block interface to use (0=bio,1=rq,2=multiqueue)");
 134
 135 static int g_gb = 250;
 136 module_param_named(gb, g_gb, int, 0444);
 137 MODULE_PARM_DESC(gb, "Size in GB");
 138
 139 static int g_bs = 512;
 140 module_param_named(bs, g_bs, int, 0444);
 141 MODULE_PARM_DESC(bs, "Block size (in bytes)");
 142
 143 static int nr_devices = 1;
 144 module_param(nr_devices, int, 0444);
 145 MODULE_PARM_DESC(nr_devices, "Number of devices to register");
 146
 147 static bool g_blocking;
 148 module_param_named(blocking, g_blocking, bool, 0444);
 149 MODULE_PARM_DESC(blocking, "Register as a blocking blk-mq driver device");
 150
 151 static bool shared_tags;
 152 module_param(shared_tags, bool, 0444);
 153 MODULE_PARM_DESC(shared_tags, "Share tag set between devices for blk-mq");
 154
 155 static int g_irqmode = NULL_IRQ_SOFTIRQ;
 156
 157 static int null_set_irqmode(const char *str, const struct kernel_param *kp)
 158 {
 159         return null_param_store_val(str, &g_irqmode, NULL_IRQ_NONE,
 160                                         NULL_IRQ_TIMER);
 161 }
 162
 163 static const struct kernel_param_ops null_irqmode_param_ops = {
 164         .set    = null_set_irqmode,
 165         .get    = param_get_int,
 166 };
 167
 168 device_param_cb(irqmode, &null_irqmode_param_ops, &g_irqmode, 0444);
 169 MODULE_PARM_DESC(irqmode, "IRQ completion handler. 0-none, 1-softirq, 2-timer");
 170
 171 static unsigned long g_completion_nsec = 10000;
 172 module_param_named(completion_nsec, g_completion_nsec, ulong, 0444);
 173 MODULE_PARM_DESC(completion_nsec, "Time in ns to complete a request in hardware. Default: 10,000ns");
 174
 175 static int g_hw_queue_depth = 64;
 176 module_param_named(hw_queue_depth, g_hw_queue_depth, int, 0444);
 177 MODULE_PARM_DESC(hw_queue_depth, "Queue depth for each hardware queue. Default: 64");
 178
 179 static bool g_use_per_node_hctx;
 180 module_param_named(use_per_node_hctx, g_use_per_node_hctx, bool, 0444);
 181 MODULE_PARM_DESC(use_per_node_hctx, "Use per-node allocation for hardware context queues. Default: false");
 182
 183 static bool g_zoned;
 184 module_param_named(zoned, g_zoned, bool, S_IRUGO);
 185 MODULE_PARM_DESC(zoned, "Make device as a host-managed zoned block device. Default: false");
 186
 187 static unsigned long g_zone_size = 256;
 188 module_param_named(zone_size, g_zone_size, ulong, S_IRUGO);
 189 MODULE_PARM_DESC(zone_size, "Zone size in MB when block device is zoned. Must be power-of-two: Default: 256");
 190
 191 static struct nullb_device *null_alloc_dev(void);
 192 static void null_free_dev(struct nullb_device *dev);
 193 static void null_del_dev(struct nullb *nullb);
 194 static int null_add_dev(struct nullb_device *dev);
 195 static void null_free_device_storage(struct nullb_device *dev, bool is_cache);
 196
 197 static inline struct nullb_device *to_nullb_device(struct config_item *item)
 198 {
 199         return item ? container_of(item, struct nullb_device, item) : NULL;
 200 }
 201
 202 static inline ssize_t nullb_device_uint_attr_show(unsigned int val, char *page)
 203 {
 204         return snprintf(page, PAGE_SIZE, "%u\n", val);
 205 }
 206
 207 static inline ssize_t nullb_device_ulong_attr_show(unsigned long val,
 208         char *page)
 209 {
 210         return snprintf(page, PAGE_SIZE, "%lu\n", val);
 211 }
 212
 213 static inline ssize_t nullb_device_bool_attr_show(bool val, char *page)
 214 {
 215         return snprintf(page, PAGE_SIZE, "%u\n", val);
 216 }
 217
 218 static ssize_t nullb_device_uint_attr_store(unsigned int *val,
 219         const char *page, size_t count)
 220 {
 221         unsigned int tmp;
 222         int result;
 223
 224         result = kstrtouint(page, 0, &tmp);
 225         if (result)
 226                 return result;
 227
 228         *val = tmp;
 229         return count;
 230 }
 231
 232 static ssize_t nullb_device_ulong_attr_store(unsigned long *val,
 233         const char *page, size_t count)
 234 {
 235         int result;
 236         unsigned long tmp;
 237
 238         result = kstrtoul(page, 0, &tmp);
 239         if (result)
 240                 return result;
 241
 242         *val = tmp;
 243         return count;
 244 }
 245
 246 static ssize_t nullb_device_bool_attr_store(bool *val, const char *page,
 247         size_t count)
 248 {
 249         bool tmp;
 250         int result;
 251
 252         result = kstrtobool(page,  &tmp);
 253         if (result)
 254                 return result;
 255
 256         *val = tmp;
 257         return count;
 258 }
 259
 260 /* The following macro should only be used with TYPE = {uint, ulong, bool}. */
 261 #define NULLB_DEVICE_ATTR(NAME, TYPE)                                           \
 262 static ssize_t                                                                  \
 263 nullb_device_##NAME##_show(struct config_item *item, char *page)                \
 264 {                                                                               \
 265         return nullb_device_##TYPE##_attr_show(                                 \
 266                                 to_nullb_device(item)->NAME, page);             \
 267 }                                                                               \
 268 static ssize_t                                                                  \
 269 nullb_device_##NAME##_store(struct config_item *item, const char *page,         \
 270                             size_t count)                                       \
 271 {                                                                               \
 272         if (test_bit(NULLB_DEV_FL_CONFIGURED, &to_nullb_device(item)->flags))   \
 273                 return -EBUSY;                                                  \
 274         return nullb_device_##TYPE##_attr_store(                                \
 275                         &to_nullb_device(item)->NAME, page, count);             \
 276 }                                                                               \
 277 CONFIGFS_ATTR(nullb_device_, NAME);
 278
 279 NULLB_DEVICE_ATTR(size, ulong);
 280 NULLB_DEVICE_ATTR(completion_nsec, ulong);
 281 NULLB_DEVICE_ATTR(submit_queues, uint);
 282 NULLB_DEVICE_ATTR(home_node, uint);
 283 NULLB_DEVICE_ATTR(queue_mode, uint);
 284 NULLB_DEVICE_ATTR(blocksize, uint);
 285 NULLB_DEVICE_ATTR(irqmode, uint);
 286 NULLB_DEVICE_ATTR(hw_queue_depth, uint);
 287 NULLB_DEVICE_ATTR(index, uint);
 288 NULLB_DEVICE_ATTR(blocking, bool);
 289 NULLB_DEVICE_ATTR(use_per_node_hctx, bool);
 290 NULLB_DEVICE_ATTR(memory_backed, bool);
 291 NULLB_DEVICE_ATTR(discard, bool);
 292 NULLB_DEVICE_ATTR(mbps, uint);
 293 NULLB_DEVICE_ATTR(cache_size, ulong);
 294 NULLB_DEVICE_ATTR(zoned, bool);
 295 NULLB_DEVICE_ATTR(zone_size, ulong);
 296
 297 static ssize_t nullb_device_power_show(struct config_item *item, char *page)
 298 {
 299         return nullb_device_bool_attr_show(to_nullb_device(item)->power, page);
 300 }
 301
 302 static ssize_t nullb_device_power_store(struct config_item *item,
 303                                      const char *page, size_t count)
 304 {
 305         struct nullb_device *dev = to_nullb_device(item);
 306         bool newp = false;
 307         ssize_t ret;
 308
 309         ret = nullb_device_bool_attr_store(&newp, page, count);
 310         if (ret < 0)
 311                 return ret;
 312
 313         if (!dev->power && newp) {
 314                 if (test_and_set_bit(NULLB_DEV_FL_UP, &dev->flags))
 315                         return count;
 316                 if (null_add_dev(dev)) {
 317                         clear_bit(NULLB_DEV_FL_UP, &dev->flags);
 318                         return -ENOMEM;
 319                 }
 320
 321                 set_bit(NULLB_DEV_FL_CONFIGURED, &dev->flags);
 322                 dev->power = newp;
 323         } else if (dev->power && !newp) {
 324                 if (test_and_clear_bit(NULLB_DEV_FL_UP, &dev->flags)) {
 325                         mutex_lock(&lock);
 326                         dev->power = newp;
 327                         null_del_dev(dev->nullb);
 328                         mutex_unlock(&lock);
 329                 }
 330                 clear_bit(NULLB_DEV_FL_CONFIGURED, &dev->flags);
 331         }
 332
 333         return count;
 334 }
 335
 336 CONFIGFS_ATTR(nullb_device_, power);
 337
 338 static ssize_t nullb_device_badblocks_show(struct config_item *item, char *page)
 339 {
 340         struct nullb_device *t_dev = to_nullb_device(item);
 341
 342         return badblocks_show(&t_dev->badblocks, page, 0);
 343 }
 344
 345 static ssize_t nullb_device_badblocks_store(struct config_item *item,
 346                                      const char *page, size_t count)
 347 {
 348         struct nullb_device *t_dev = to_nullb_device(item);
 349         char *orig, *buf, *tmp;
 350         u64 start, end;
 351         int ret;
 352
 353         orig = kstrndup(page, count, GFP_KERNEL);
 354         if (!orig)
 355                 return -ENOMEM;
 356
 357         buf = strstrip(orig);
 358
 359         ret = -EINVAL;
 360         if (buf[0] != '+' && buf[0] != '-')
 361                 goto out;
 362         tmp = strchr(&buf[1], '-');
 363         if (!tmp)
 364                 goto out;
 365         *tmp = '\0';
 366         ret = kstrtoull(buf + 1, 0, &start);
 367         if (ret)
 368                 goto out;
 369         ret = kstrtoull(tmp + 1, 0, &end);
 370         if (ret)
 371                 goto out;
 372         ret = -EINVAL;
 373         if (start > end)
 374                 goto out;
 375         /* enable badblocks */
 376         cmpxchg(&t_dev->badblocks.shift, -1, 0);
 377         if (buf[0] == '+')
 378                 ret = badblocks_set(&t_dev->badblocks, start,
 379                         end - start + 1, 1);
 380         else
 381                 ret = badblocks_clear(&t_dev->badblocks, start,
 382                         end - start + 1);
 383         if (ret == 0)
 384                 ret = count;
 385 out:
 386         kfree(orig);
 387         return ret;
 388 }
 389 CONFIGFS_ATTR(nullb_device_, badblocks);
 390
 391 static struct configfs_attribute *nullb_device_attrs[] = {
 392         &nullb_device_attr_size,
 393         &nullb_device_attr_completion_nsec,
 394         &nullb_device_attr_submit_queues,
 395         &nullb_device_attr_home_node,
 396         &nullb_device_attr_queue_mode,
 397         &nullb_device_attr_blocksize,
 398         &nullb_device_attr_irqmode,
 399         &nullb_device_attr_hw_queue_depth,
 400         &nullb_device_attr_index,
 401         &nullb_device_attr_blocking,
 402         &nullb_device_attr_use_per_node_hctx,
 403         &nullb_device_attr_power,
 404         &nullb_device_attr_memory_backed,
 405         &nullb_device_attr_discard,
 406         &nullb_device_attr_mbps,
 407         &nullb_device_attr_cache_size,
 408         &nullb_device_attr_badblocks,
 409         &nullb_device_attr_zoned,
 410         &nullb_device_attr_zone_size,
 411         NULL,
 412 };
 413
 414 static void nullb_device_release(struct config_item *item)
 415 {
 416         struct nullb_device *dev = to_nullb_device(item);
 417
 418         null_free_device_storage(dev, false);
 419         null_free_dev(dev);
 420 }
 421
 422 static struct configfs_item_operations nullb_device_ops = {
 423         .release        = nullb_device_release,
 424 };
 425
 426 static const struct config_item_type nullb_device_type = {
 427         .ct_item_ops    = &nullb_device_ops,
 428         .ct_attrs       = nullb_device_attrs,
 429         .ct_owner       = THIS_MODULE,
 430 };
 431
 432 static struct
 433 config_item *nullb_group_make_item(struct config_group *group, const char *name)
 434 {
 435         struct nullb_device *dev;
 436
 437         dev = null_alloc_dev();
 438         if (!dev)
 439                 return ERR_PTR(-ENOMEM);
 440
 441         config_item_init_type_name(&dev->item, name, &nullb_device_type);
 442
 443         return &dev->item;
 444 }
 445
 446 static void
 447 nullb_group_drop_item(struct config_group *group, struct config_item *item)
 448 {
 449         struct nullb_device *dev = to_nullb_device(item);
 450
 451         if (test_and_clear_bit(NULLB_DEV_FL_UP, &dev->flags)) {
 452                 mutex_lock(&lock);
 453                 dev->power = false;
 454                 null_del_dev(dev->nullb);
 455                 mutex_unlock(&lock);
 456         }
 457
 458         config_item_put(item);
 459 }
 460
 461 static ssize_t memb_group_features_show(struct config_item *item, char *page)
 462 {
 463         return snprintf(page, PAGE_SIZE, "memory_backed,discard,bandwidth,cache,badblocks,zoned,zone_size\n");
 464 }
 465
 466 CONFIGFS_ATTR_RO(memb_group_, features);
 467
 468 static struct configfs_attribute *nullb_group_attrs[] = {
 469         &memb_group_attr_features,
 470         NULL,
 471 };
 472
 473 static struct configfs_group_operations nullb_group_ops = {
 474         .make_item      = nullb_group_make_item,
 475         .drop_item      = nullb_group_drop_item,
 476 };
 477
 478 static const struct config_item_type nullb_group_type = {
 479         .ct_group_ops   = &nullb_group_ops,
 480         .ct_attrs       = nullb_group_attrs,
 481         .ct_owner       = THIS_MODULE,
 482 };
 483
 484 static struct configfs_subsystem nullb_subsys = {
 485         .su_group = {
 486                 .cg_item = {
 487                         .ci_namebuf = "nullb",
 488                         .ci_type = &nullb_group_type,
 489                 },
 490         },
 491 };
 492
 493 static inline int null_cache_active(struct nullb *nullb)
 494 {
 495         return test_bit(NULLB_DEV_FL_CACHE, &nullb->dev->flags);
 496 }
 497
 498 static struct nullb_device *null_alloc_dev(void)
 499 {
 500         struct nullb_device *dev;
 501
 502         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
 503         if (!dev)
 504                 return NULL;
 505         INIT_RADIX_TREE(&dev->data, GFP_ATOMIC);
 506         INIT_RADIX_TREE(&dev->cache, GFP_ATOMIC);
 507         if (badblocks_init(&dev->badblocks, 0)) {
 508                 kfree(dev);
 509                 return NULL;
 510         }
 511
 512         dev->size = g_gb * 1024;
 513         dev->completion_nsec = g_completion_nsec;
 514         dev->submit_queues = g_submit_queues;
 515         dev->home_node = g_home_node;
 516         dev->queue_mode = g_queue_mode;
 517         dev->blocksize = g_bs;
 518         dev->irqmode = g_irqmode;
 519         dev->hw_queue_depth = g_hw_queue_depth;
 520         dev->blocking = g_blocking;
 521         dev->use_per_node_hctx = g_use_per_node_hctx;
 522         dev->zoned = g_zoned;
 523         dev->zone_size = g_zone_size;
 524         return dev;
 525 }
 526
 527 static void null_free_dev(struct nullb_device *dev)
 528 {
 529         if (!dev)
 530                 return;
 531
 532         null_zone_exit(dev);
 533         badblocks_exit(&dev->badblocks);
 534         kfree(dev);
 535 }
 536
 537 static void put_tag(struct nullb_queue *nq, unsigned int tag)
 538 {
 539         clear_bit_unlock(tag, nq->tag_map);
 540
 541         if (waitqueue_active(&nq->wait))
 542                 wake_up(&nq->wait);
 543 }
 544
 545 static unsigned int get_tag(struct nullb_queue *nq)
 546 {
 547         unsigned int tag;
 548
 549         do {
 550                 tag = find_first_zero_bit(nq->tag_map, nq->queue_depth);
 551                 if (tag >= nq->queue_depth)
 552                         return -1U;
 553         } while (test_and_set_bit_lock(tag, nq->tag_map));
 554
 555         return tag;
 556 }
 557
 558 static void free_cmd(struct nullb_cmd *cmd)
 559 {
 560         put_tag(cmd->nq, cmd->tag);
 561 }
 562
 563 static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer);
 564
 565 static struct nullb_cmd *__alloc_cmd(struct nullb_queue *nq)
 566 {
 567         struct nullb_cmd *cmd;
 568         unsigned int tag;
 569
 570         tag = get_tag(nq);
 571         if (tag != -1U) {
 572                 cmd = &nq->cmds[tag];
 573                 cmd->tag = tag;
 574                 cmd->nq = nq;
 575                 if (nq->dev->irqmode == NULL_IRQ_TIMER) {
 576                         hrtimer_init(&cmd->timer, CLOCK_MONOTONIC,
 577                                      HRTIMER_MODE_REL);
 578                         cmd->timer.function = null_cmd_timer_expired;
 579                 }
 580                 return cmd;
 581         }
 582
 583         return NULL;
 584 }
 585
 586 static struct nullb_cmd *alloc_cmd(struct nullb_queue *nq, int can_wait)
 587 {
 588         struct nullb_cmd *cmd;
 589         DEFINE_WAIT(wait);
 590
 591         cmd = __alloc_cmd(nq);
 592         if (cmd || !can_wait)
 593                 return cmd;
 594
 595         do {
 596                 prepare_to_wait(&nq->wait, &wait, TASK_UNINTERRUPTIBLE);
 597                 cmd = __alloc_cmd(nq);
 598                 if (cmd)
 599                         break;
 600
 601                 io_schedule();
 602         } while (1);
 603
 604         finish_wait(&nq->wait, &wait);
 605         return cmd;
 606 }
 607
 608 static void end_cmd(struct nullb_cmd *cmd)
 609 {
 610         struct request_queue *q = NULL;
 611         int queue_mode = cmd->nq->dev->queue_mode;
 612
 613         if (cmd->rq)
 614                 q = cmd->rq->q;
 615
 616         switch (queue_mode)  {
 617         case NULL_Q_MQ:
 618                 blk_mq_end_request(cmd->rq, cmd->error);
 619                 return;
 620         case NULL_Q_RQ:
 621                 INIT_LIST_HEAD(&cmd->rq->queuelist);
 622                 blk_end_request_all(cmd->rq, cmd->error);
 623                 break;
 624         case NULL_Q_BIO:
 625                 cmd->bio->bi_status = cmd->error;
 626                 bio_endio(cmd->bio);
 627                 break;
 628         }
 629
 630         free_cmd(cmd);
 631
 632         /* Restart queue if needed, as we are freeing a tag */
 633         if (queue_mode == NULL_Q_RQ && blk_queue_stopped(q)) {
 634                 unsigned long flags;
 635
 636                 spin_lock_irqsave(q->queue_lock, flags);
 637                 blk_start_queue_async(q);
 638                 spin_unlock_irqrestore(q->queue_lock, flags);
 639         }
 640 }
 641
 642 static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer)
 643 {
 644         end_cmd(container_of(timer, struct nullb_cmd, timer));
 645
 646         return HRTIMER_NORESTART;
 647 }
 648
 649 static void null_cmd_end_timer(struct nullb_cmd *cmd)
 650 {
 651         ktime_t kt = cmd->nq->dev->completion_nsec;
 652
 653         hrtimer_start(&cmd->timer, kt, HRTIMER_MODE_REL);
 654 }
 655
 656 static void null_softirq_done_fn(struct request *rq)
 657 {
 658         struct nullb *nullb = rq->q->queuedata;
 659
 660         if (nullb->dev->queue_mode == NULL_Q_MQ)
 661                 end_cmd(blk_mq_rq_to_pdu(rq));
 662         else
 663                 end_cmd(rq->special);
 664 }
 665
 666 static struct nullb_page *null_alloc_page(gfp_t gfp_flags)
 667 {
 668         struct nullb_page *t_page;
 669
 670         t_page = kmalloc(sizeof(struct nullb_page), gfp_flags);
 671         if (!t_page)
 672                 goto out;
 673
 674         t_page->page = alloc_pages(gfp_flags, 0);
 675         if (!t_page->page)
 676                 goto out_freepage;
 677
 678         memset(t_page->bitmap, 0, sizeof(t_page->bitmap));
 679         return t_page;
 680 out_freepage:
 681         kfree(t_page);
 682 out:
 683         return NULL;
 684 }
 685
 686 static void null_free_page(struct nullb_page *t_page)
 687 {
 688         __set_bit(NULLB_PAGE_FREE, t_page->bitmap);
 689         if (test_bit(NULLB_PAGE_LOCK, t_page->bitmap))
 690                 return;
 691         __free_page(t_page->page);
 692         kfree(t_page);
 693 }
 694
 695 static bool null_page_empty(struct nullb_page *page)
 696 {
 697         int size = MAP_SZ - 2;
 698
 699         return find_first_bit(page->bitmap, size) == size;
 700 }
 701
 702 static void null_free_sector(struct nullb *nullb, sector_t sector,
 703         bool is_cache)
 704 {
 705         unsigned int sector_bit;
 706         u64 idx;
 707         struct nullb_page *t_page, *ret;
 708         struct radix_tree_root *root;
 709
 710         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 711         idx = sector >> PAGE_SECTORS_SHIFT;
 712         sector_bit = (sector & SECTOR_MASK);
 713
 714         t_page = radix_tree_lookup(root, idx);
 715         if (t_page) {
 716                 __clear_bit(sector_bit, t_page->bitmap);
 717
 718                 if (null_page_empty(t_page)) {
 719                         ret = radix_tree_delete_item(root, idx, t_page);
 720                         WARN_ON(ret != t_page);
 721                         null_free_page(ret);
 722                         if (is_cache)
 723                                 nullb->dev->curr_cache -= PAGE_SIZE;
 724                 }
 725         }
 726 }
 727
 728 static struct nullb_page *null_radix_tree_insert(struct nullb *nullb, u64 idx,
 729         struct nullb_page *t_page, bool is_cache)
 730 {
 731         struct radix_tree_root *root;
 732
 733         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 734
 735         if (radix_tree_insert(root, idx, t_page)) {
 736                 null_free_page(t_page);
 737                 t_page = radix_tree_lookup(root, idx);
 738                 WARN_ON(!t_page || t_page->page->index != idx);
 739         } else if (is_cache)
 740                 nullb->dev->curr_cache += PAGE_SIZE;
 741
 742         return t_page;
 743 }
 744
 745 static void null_free_device_storage(struct nullb_device *dev, bool is_cache)
 746 {
 747         unsigned long pos = 0;
 748         int nr_pages;
 749         struct nullb_page *ret, *t_pages[FREE_BATCH];
 750         struct radix_tree_root *root;
 751
 752         root = is_cache ? &dev->cache : &dev->data;
 753
 754         do {
 755                 int i;
 756
 757                 nr_pages = radix_tree_gang_lookup(root,
 758                                 (void **)t_pages, pos, FREE_BATCH);
 759
 760                 for (i = 0; i < nr_pages; i++) {
 761                         pos = t_pages[i]->page->index;
 762                         ret = radix_tree_delete_item(root, pos, t_pages[i]);
 763                         WARN_ON(ret != t_pages[i]);
 764                         null_free_page(ret);
 765                 }
 766
 767                 pos++;
 768         } while (nr_pages == FREE_BATCH);
 769
 770         if (is_cache)
 771                 dev->curr_cache = 0;
 772 }
 773
 774 static struct nullb_page *__null_lookup_page(struct nullb *nullb,
 775         sector_t sector, bool for_write, bool is_cache)
 776 {
 777         unsigned int sector_bit;
 778         u64 idx;
 779         struct nullb_page *t_page;
 780         struct radix_tree_root *root;
 781
 782         idx = sector >> PAGE_SECTORS_SHIFT;
 783         sector_bit = (sector & SECTOR_MASK);
 784
 785         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 786         t_page = radix_tree_lookup(root, idx);
 787         WARN_ON(t_page && t_page->page->index != idx);
 788
 789         if (t_page && (for_write || test_bit(sector_bit, t_page->bitmap)))
 790                 return t_page;
 791
 792         return NULL;
 793 }
 794
 795 static struct nullb_page *null_lookup_page(struct nullb *nullb,
 796         sector_t sector, bool for_write, bool ignore_cache)
 797 {
 798         struct nullb_page *page = NULL;
 799
 800         if (!ignore_cache)
 801                 page = __null_lookup_page(nullb, sector, for_write, true);
 802         if (page)
 803                 return page;
 804         return __null_lookup_page(nullb, sector, for_write, false);
 805 }
 806
 807 static struct nullb_page *null_insert_page(struct nullb *nullb,
 808                                            sector_t sector, bool ignore_cache)
 809         __releases(&nullb->lock)
 810         __acquires(&nullb->lock)
 811 {
 812         u64 idx;
 813         struct nullb_page *t_page;
 814
 815         t_page = null_lookup_page(nullb, sector, true, ignore_cache);
 816         if (t_page)
 817                 return t_page;
 818
 819         spin_unlock_irq(&nullb->lock);
 820
 821         t_page = null_alloc_page(GFP_NOIO);
 822         if (!t_page)
 823                 goto out_lock;
 824
 825         if (radix_tree_preload(GFP_NOIO))
 826                 goto out_freepage;
 827
 828         spin_lock_irq(&nullb->lock);
 829         idx = sector >> PAGE_SECTORS_SHIFT;
 830         t_page->page->index = idx;
 831         t_page = null_radix_tree_insert(nullb, idx, t_page, !ignore_cache);
 832         radix_tree_preload_end();
 833
 834         return t_page;
 835 out_freepage:
 836         null_free_page(t_page);
 837 out_lock:
 838         spin_lock_irq(&nullb->lock);
 839         return null_lookup_page(nullb, sector, true, ignore_cache);
 840 }
 841
 842 static int null_flush_cache_page(struct nullb *nullb, struct nullb_page *c_page)
 843 {
 844         int i;
 845         unsigned int offset;
 846         u64 idx;
 847         struct nullb_page *t_page, *ret;
 848         void *dst, *src;
 849
 850         idx = c_page->page->index;
 851
 852         t_page = null_insert_page(nullb, idx << PAGE_SECTORS_SHIFT, true);
 853
 854         __clear_bit(NULLB_PAGE_LOCK, c_page->bitmap);
 855         if (test_bit(NULLB_PAGE_FREE, c_page->bitmap)) {
 856                 null_free_page(c_page);
 857                 if (t_page && null_page_empty(t_page)) {
 858                         ret = radix_tree_delete_item(&nullb->dev->data,
 859                                 idx, t_page);
 860                         null_free_page(t_page);
 861                 }
 862                 return 0;
 863         }
 864
 865         if (!t_page)
 866                 return -ENOMEM;
 867
 868         src = kmap_atomic(c_page->page);
 869         dst = kmap_atomic(t_page->page);
 870
 871         for (i = 0; i < PAGE_SECTORS;
 872                         i += (nullb->dev->blocksize >> SECTOR_SHIFT)) {
 873                 if (test_bit(i, c_page->bitmap)) {
 874                         offset = (i << SECTOR_SHIFT);
 875                         memcpy(dst + offset, src + offset,
 876                                 nullb->dev->blocksize);
 877                         __set_bit(i, t_page->bitmap);
 878                 }
 879         }
 880
 881         kunmap_atomic(dst);
 882         kunmap_atomic(src);
 883
 884         ret = radix_tree_delete_item(&nullb->dev->cache, idx, c_page);
 885         null_free_page(ret);
 886         nullb->dev->curr_cache -= PAGE_SIZE;
 887
 888         return 0;
 889 }
 890
 891 static int null_make_cache_space(struct nullb *nullb, unsigned long n)
 892 {
 893         int i, err, nr_pages;
 894         struct nullb_page *c_pages[FREE_BATCH];
 895         unsigned long flushed = 0, one_round;
 896
 897 again:
 898         if ((nullb->dev->cache_size * 1024 * 1024) >
 899              nullb->dev->curr_cache + n || nullb->dev->curr_cache == 0)
 900                 return 0;
 901
 902         nr_pages = radix_tree_gang_lookup(&nullb->dev->cache,
 903                         (void **)c_pages, nullb->cache_flush_pos, FREE_BATCH);
 904         /*
 905          * nullb_flush_cache_page could unlock before using the c_pages. To
 906          * avoid race, we don't allow page free
 907          */
 908         for (i = 0; i < nr_pages; i++) {
 909                 nullb->cache_flush_pos = c_pages[i]->page->index;
 910                 /*
 911                  * We found the page which is being flushed to disk by other
 912                  * threads
 913                  */
 914                 if (test_bit(NULLB_PAGE_LOCK, c_pages[i]->bitmap))
 915                         c_pages[i] = NULL;
 916                 else
 917                         __set_bit(NULLB_PAGE_LOCK, c_pages[i]->bitmap);
 918         }
 919
 920         one_round = 0;
 921         for (i = 0; i < nr_pages; i++) {
 922                 if (c_pages[i] == NULL)
 923                         continue;
 924                 err = null_flush_cache_page(nullb, c_pages[i]);
 925                 if (err)
 926                         return err;
 927                 one_round++;
 928         }
 929         flushed += one_round << PAGE_SHIFT;
 930
 931         if (n > flushed) {
 932                 if (nr_pages == 0)
 933                         nullb->cache_flush_pos = 0;
 934                 if (one_round == 0) {
 935                         /* give other threads a chance */
 936                         spin_unlock_irq(&nullb->lock);
 937                         spin_lock_irq(&nullb->lock);
 938                 }
 939                 goto again;
 940         }
 941         return 0;
 942 }
 943
 944 static int copy_to_nullb(struct nullb *nullb, struct page *source,
 945         unsigned int off, sector_t sector, size_t n, bool is_fua)
 946 {
 947         size_t temp, count = 0;
 948         unsigned int offset;
 949         struct nullb_page *t_page;
 950         void *dst, *src;
 951
 952         while (count < n) {
 953                 temp = min_t(size_t, nullb->dev->blocksize, n - count);
 954
 955                 if (null_cache_active(nullb) && !is_fua)
 956                         null_make_cache_space(nullb, PAGE_SIZE);
 957
 958                 offset = (sector & SECTOR_MASK) << SECTOR_SHIFT;
 959                 t_page = null_insert_page(nullb, sector,
 960                         !null_cache_active(nullb) || is_fua);
 961                 if (!t_page)
 962                         return -ENOSPC;
 963
 964                 src = kmap_atomic(source);
 965                 dst = kmap_atomic(t_page->page);
 966                 memcpy(dst + offset, src + off + count, temp);
 967                 kunmap_atomic(dst);
 968                 kunmap_atomic(src);
 969
 970                 __set_bit(sector & SECTOR_MASK, t_page->bitmap);
 971
 972                 if (is_fua)
 973                         null_free_sector(nullb, sector, true);
 974
 975                 count += temp;
 976                 sector += temp >> SECTOR_SHIFT;
 977         }
 978         return 0;
 979 }
 980
 981 static int copy_from_nullb(struct nullb *nullb, struct page *dest,
 982         unsigned int off, sector_t sector, size_t n)
 983 {
 984         size_t temp, count = 0;
 985         unsigned int offset;
 986         struct nullb_page *t_page;
 987         void *dst, *src;
 988
 989         while (count < n) {
 990                 temp = min_t(size_t, nullb->dev->blocksize, n - count);
 991
 992                 offset = (sector & SECTOR_MASK) << SECTOR_SHIFT;
 993                 t_page = null_lookup_page(nullb, sector, false,
 994                         !null_cache_active(nullb));
 995
 996                 dst = kmap_atomic(dest);
 997                 if (!t_page) {
 998                         memset(dst + off + count, 0, temp);
 999                         goto next;
1000                 }
1001                 src = kmap_atomic(t_page->page);
1002                 memcpy(dst + off + count, src + offset, temp);
1003                 kunmap_atomic(src);
1004 next:
1005                 kunmap_atomic(dst);
1006
1007                 count += temp;
1008                 sector += temp >> SECTOR_SHIFT;
1009         }
1010         return 0;
1011 }
1012
1013 static void null_handle_discard(struct nullb *nullb, sector_t sector, size_t n)
1014 {
1015         size_t temp;
1016
1017         spin_lock_irq(&nullb->lock);
1018         while (n > 0) {
1019                 temp = min_t(size_t, n, nullb->dev->blocksize);
1020                 null_free_sector(nullb, sector, false);
1021                 if (null_cache_active(nullb))
1022                         null_free_sector(nullb, sector, true);
1023                 sector += temp >> SECTOR_SHIFT;
1024                 n -= temp;
1025         }
1026         spin_unlock_irq(&nullb->lock);
1027 }
1028
1029 static int null_handle_flush(struct nullb *nullb)
1030 {
1031         int err;
1032
1033         if (!null_cache_active(nullb))
1034                 return 0;
1035
1036         spin_lock_irq(&nullb->lock);
1037         while (true) {
1038                 err = null_make_cache_space(nullb,
1039                         nullb->dev->cache_size * 1024 * 1024);
1040                 if (err || nullb->dev->curr_cache == 0)
1041                         break;
1042         }
1043
1044         WARN_ON(!radix_tree_empty(&nullb->dev->cache));
1045         spin_unlock_irq(&nullb->lock);
1046         return err;
1047 }
1048
1049 static int null_transfer(struct nullb *nullb, struct page *page,
1050         unsigned int len, unsigned int off, bool is_write, sector_t sector,
1051         bool is_fua)
1052 {
1053         int err = 0;
1054
1055         if (!is_write) {
1056                 err = copy_from_nullb(nullb, page, off, sector, len);
1057                 flush_dcache_page(page);
1058         } else {
1059                 flush_dcache_page(page);
1060                 err = copy_to_nullb(nullb, page, off, sector, len, is_fua);
1061         }
1062
1063         return err;
1064 }
1065
1066 static int null_handle_rq(struct nullb_cmd *cmd)
1067 {
1068         struct request *rq = cmd->rq;
1069         struct nullb *nullb = cmd->nq->dev->nullb;
1070         int err;
1071         unsigned int len;
1072         sector_t sector;
1073         struct req_iterator iter;
1074         struct bio_vec bvec;
1075
1076         sector = blk_rq_pos(rq);
1077
1078         if (req_op(rq) == REQ_OP_DISCARD) {
1079                 null_handle_discard(nullb, sector, blk_rq_bytes(rq));
1080                 return 0;
1081         }
1082
1083         spin_lock_irq(&nullb->lock);
1084         rq_for_each_segment(bvec, rq, iter) {
1085                 len = bvec.bv_len;
1086                 err = null_transfer(nullb, bvec.bv_page, len, bvec.bv_offset,
1087                                      op_is_write(req_op(rq)), sector,
1088                                      req_op(rq) & REQ_FUA);
1089                 if (err) {
1090                         spin_unlock_irq(&nullb->lock);
1091                         return err;
1092                 }
1093                 sector += len >> SECTOR_SHIFT;
1094         }
1095         spin_unlock_irq(&nullb->lock);
1096
1097         return 0;
1098 }
1099
1100 static int null_handle_bio(struct nullb_cmd *cmd)
1101 {
1102         struct bio *bio = cmd->bio;
1103         struct nullb *nullb = cmd->nq->dev->nullb;
1104         int err;
1105         unsigned int len;
1106         sector_t sector;
1107         struct bio_vec bvec;
1108         struct bvec_iter iter;
1109
1110         sector = bio->bi_iter.bi_sector;
1111
1112         if (bio_op(bio) == REQ_OP_DISCARD) {
1113                 null_handle_discard(nullb, sector,
1114                         bio_sectors(bio) << SECTOR_SHIFT);
1115                 return 0;
1116         }
1117
1118         spin_lock_irq(&nullb->lock);
1119         bio_for_each_segment(bvec, bio, iter) {
1120                 len = bvec.bv_len;
1121                 err = null_transfer(nullb, bvec.bv_page, len, bvec.bv_offset,
1122                                      op_is_write(bio_op(bio)), sector,
1123                                      bio_op(bio) & REQ_FUA);
1124                 if (err) {
1125                         spin_unlock_irq(&nullb->lock);
1126                         return err;
1127                 }
1128                 sector += len >> SECTOR_SHIFT;
1129         }
1130         spin_unlock_irq(&nullb->lock);
1131         return 0;
1132 }
1133
1134 static void null_stop_queue(struct nullb *nullb)
1135 {
1136         struct request_queue *q = nullb->q;
1137
1138         if (nullb->dev->queue_mode == NULL_Q_MQ)
1139                 blk_mq_stop_hw_queues(q);
1140         else {
1141                 spin_lock_irq(q->queue_lock);
1142                 blk_stop_queue(q);
1143                 spin_unlock_irq(q->queue_lock);
1144         }
1145 }
1146
1147 static void null_restart_queue_async(struct nullb *nullb)
1148 {
1149         struct request_queue *q = nullb->q;
1150         unsigned long flags;
1151
1152         if (nullb->dev->queue_mode == NULL_Q_MQ)
1153                 blk_mq_start_stopped_hw_queues(q, true);
1154         else {
1155                 spin_lock_irqsave(q->queue_lock, flags);
1156                 blk_start_queue_async(q);
1157                 spin_unlock_irqrestore(q->queue_lock, flags);
1158         }
1159 }
1160
1161 static bool cmd_report_zone(struct nullb *nullb, struct nullb_cmd *cmd)
1162 {
1163         struct nullb_device *dev = cmd->nq->dev;
1164
1165         if (dev->queue_mode == NULL_Q_BIO) {
1166                 if (bio_op(cmd->bio) == REQ_OP_ZONE_REPORT) {
1167                         cmd->error = null_zone_report(nullb, cmd->bio);
1168                         return true;
1169                 }
1170         } else {
1171                 if (req_op(cmd->rq) == REQ_OP_ZONE_REPORT) {
1172                         cmd->error = null_zone_report(nullb, cmd->rq->bio);
1173                         return true;
1174                 }
1175         }
1176
1177         return false;
1178 }
1179
1180 static blk_status_t null_handle_cmd(struct nullb_cmd *cmd)
1181 {
1182         struct nullb_device *dev = cmd->nq->dev;
1183         struct nullb *nullb = dev->nullb;
1184         int err = 0;
1185
1186         if (cmd_report_zone(nullb, cmd))
1187                 goto out;
1188
1189         if (test_bit(NULLB_DEV_FL_THROTTLED, &dev->flags)) {
1190                 struct request *rq = cmd->rq;
1191
1192                 if (!hrtimer_active(&nullb->bw_timer))
1193                         hrtimer_restart(&nullb->bw_timer);
1194
1195                 if (atomic_long_sub_return(blk_rq_bytes(rq),
1196                                 &nullb->cur_bytes) < 0) {
1197                         null_stop_queue(nullb);
1198                         /* race with timer */
1199                         if (atomic_long_read(&nullb->cur_bytes) > 0)
1200                                 null_restart_queue_async(nullb);
1201                         if (dev->queue_mode == NULL_Q_RQ) {
1202                                 struct request_queue *q = nullb->q;
1203
1204                                 spin_lock_irq(q->queue_lock);
1205                                 rq->rq_flags |= RQF_DONTPREP;
1206                                 blk_requeue_request(q, rq);
1207                                 spin_unlock_irq(q->queue_lock);
1208                                 return BLK_STS_OK;
1209                         } else
1210                                 /* requeue request */
1211                                 return BLK_STS_DEV_RESOURCE;
1212                 }
1213         }
1214
1215         if (nullb->dev->badblocks.shift != -1) {
1216                 int bad_sectors;
1217                 sector_t sector, size, first_bad;
1218                 bool is_flush = true;
1219
1220                 if (dev->queue_mode == NULL_Q_BIO &&
1221                                 bio_op(cmd->bio) != REQ_OP_FLUSH) {
1222                         is_flush = false;
1223                         sector = cmd->bio->bi_iter.bi_sector;
1224                         size = bio_sectors(cmd->bio);
1225                 }
1226                 if (dev->queue_mode != NULL_Q_BIO &&
1227                                 req_op(cmd->rq) != REQ_OP_FLUSH) {
1228                         is_flush = false;
1229                         sector = blk_rq_pos(cmd->rq);
1230                         size = blk_rq_sectors(cmd->rq);
1231                 }
1232                 if (!is_flush && badblocks_check(&nullb->dev->badblocks, sector,
1233                                 size, &first_bad, &bad_sectors)) {
1234                         cmd->error = BLK_STS_IOERR;
1235                         goto out;
1236                 }
1237         }
1238
1239         if (dev->memory_backed) {
1240                 if (dev->queue_mode == NULL_Q_BIO) {
1241                         if (bio_op(cmd->bio) == REQ_OP_FLUSH)
1242                                 err = null_handle_flush(nullb);
1243                         else
1244                                 err = null_handle_bio(cmd);
1245                 } else {
1246                         if (req_op(cmd->rq) == REQ_OP_FLUSH)
1247                                 err = null_handle_flush(nullb);
1248                         else
1249                                 err = null_handle_rq(cmd);
1250                 }
1251         }
1252         cmd->error = errno_to_blk_status(err);
1253
1254         if (!cmd->error && dev->zoned) {
1255                 sector_t sector;
1256                 unsigned int nr_sectors;
1257                 int op;
1258
1259                 if (dev->queue_mode == NULL_Q_BIO) {
1260                         op = bio_op(cmd->bio);
1261                         sector = cmd->bio->bi_iter.bi_sector;
1262                         nr_sectors = cmd->bio->bi_iter.bi_size >> 9;
1263                 } else {
1264                         op = req_op(cmd->rq);
1265                         sector = blk_rq_pos(cmd->rq);
1266                         nr_sectors = blk_rq_sectors(cmd->rq);
1267                 }
1268
1269                 if (op == REQ_OP_WRITE)
1270                         null_zone_write(cmd, sector, nr_sectors);
1271                 else if (op == REQ_OP_ZONE_RESET)
1272                         null_zone_reset(cmd, sector);
1273         }
1274 out:
1275         /* Complete IO by inline, softirq or timer */
1276         switch (dev->irqmode) {
1277         case NULL_IRQ_SOFTIRQ:
1278                 switch (dev->queue_mode)  {
1279                 case NULL_Q_MQ:
1280                         blk_mq_complete_request(cmd->rq);
1281                         break;
1282                 case NULL_Q_RQ:
1283                         blk_complete_request(cmd->rq);
1284                         break;
1285                 case NULL_Q_BIO:
1286                         /*
1287                          * XXX: no proper submitting cpu information available.
1288                          */
1289                         end_cmd(cmd);
1290                         break;
1291                 }
1292                 break;
1293         case NULL_IRQ_NONE:
1294                 end_cmd(cmd);
1295                 break;
1296         case NULL_IRQ_TIMER:
1297                 null_cmd_end_timer(cmd);
1298                 break;
1299         }
1300         return BLK_STS_OK;
1301 }
1302
1303 static enum hrtimer_restart nullb_bwtimer_fn(struct hrtimer *timer)
1304 {
1305         struct nullb *nullb = container_of(timer, struct nullb, bw_timer);
1306         ktime_t timer_interval = ktime_set(0, TIMER_INTERVAL);
1307         unsigned int mbps = nullb->dev->mbps;
1308
1309         if (atomic_long_read(&nullb->cur_bytes) == mb_per_tick(mbps))
1310                 return HRTIMER_NORESTART;
1311
1312         atomic_long_set(&nullb->cur_bytes, mb_per_tick(mbps));
1313         null_restart_queue_async(nullb);
1314
1315         hrtimer_forward_now(&nullb->bw_timer, timer_interval);
1316
1317         return HRTIMER_RESTART;
1318 }
1319
1320 static void nullb_setup_bwtimer(struct nullb *nullb)
1321 {
1322         ktime_t timer_interval = ktime_set(0, TIMER_INTERVAL);
1323
1324         hrtimer_init(&nullb->bw_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1325         nullb->bw_timer.function = nullb_bwtimer_fn;
1326         atomic_long_set(&nullb->cur_bytes, mb_per_tick(nullb->dev->mbps));
1327         hrtimer_start(&nullb->bw_timer, timer_interval, HRTIMER_MODE_REL);
1328 }
1329
1330 static struct nullb_queue *nullb_to_queue(struct nullb *nullb)
1331 {
1332         int index = 0;
1333
1334         if (nullb->nr_queues != 1)
1335                 index = raw_smp_processor_id() / ((nr_cpu_ids + nullb->nr_queues - 1) / nullb->nr_queues);
1336
1337         return &nullb->queues[index];
1338 }
1339
1340 static blk_qc_t null_queue_bio(struct request_queue *q, struct bio *bio)
1341 {
1342         struct nullb *nullb = q->queuedata;
1343         struct nullb_queue *nq = nullb_to_queue(nullb);
1344         struct nullb_cmd *cmd;
1345
1346         cmd = alloc_cmd(nq, 1);
1347         cmd->bio = bio;
1348
1349         null_handle_cmd(cmd);
1350         return BLK_QC_T_NONE;
1351 }
1352
1353 static enum blk_eh_timer_return null_rq_timed_out_fn(struct request *rq)
1354 {
1355         pr_info("null: rq %p timed out\n", rq);
1356         __blk_complete_request(rq);
1357         return BLK_EH_DONE;
1358 }
1359
1360 static int null_rq_prep_fn(struct request_queue *q, struct request *req)
1361 {
1362         struct nullb *nullb = q->queuedata;
1363         struct nullb_queue *nq = nullb_to_queue(nullb);
1364         struct nullb_cmd *cmd;
1365
1366         cmd = alloc_cmd(nq, 0);
1367         if (cmd) {
1368                 cmd->rq = req;
1369                 req->special = cmd;
1370                 return BLKPREP_OK;
1371         }
1372         blk_stop_queue(q);
1373
1374         return BLKPREP_DEFER;
1375 }
1376
1377 static bool should_timeout_request(struct request *rq)
1378 {
1379 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1380         if (g_timeout_str[0])
1381                 return should_fail(&null_timeout_attr, 1);
1382 #endif
1383         return false;
1384 }
1385
1386 static bool should_requeue_request(struct request *rq)
1387 {
1388 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1389         if (g_requeue_str[0])
1390                 return should_fail(&null_requeue_attr, 1);
1391 #endif
1392         return false;
1393 }
1394
1395 static void null_request_fn(struct request_queue *q)
1396 {
1397         struct request *rq;
1398
1399         while ((rq = blk_fetch_request(q)) != NULL) {
1400                 struct nullb_cmd *cmd = rq->special;
1401
1402                 /* just ignore the request */
1403                 if (should_timeout_request(rq))
1404                         continue;
1405                 if (should_requeue_request(rq)) {
1406                         blk_requeue_request(q, rq);
1407                         continue;
1408                 }
1409
1410                 spin_unlock_irq(q->queue_lock);
1411                 null_handle_cmd(cmd);
1412                 spin_lock_irq(q->queue_lock);
1413         }
1414 }
1415
1416 static enum blk_eh_timer_return null_timeout_rq(struct request *rq, bool res)
1417 {
1418         pr_info("null: rq %p timed out\n", rq);
1419         blk_mq_complete_request(rq);
1420         return BLK_EH_DONE;
1421 }
1422
1423 static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx,
1424                          const struct blk_mq_queue_data *bd)
1425 {
1426         struct nullb_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1427         struct nullb_queue *nq = hctx->driver_data;
1428
1429         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1430
1431         if (nq->dev->irqmode == NULL_IRQ_TIMER) {
1432                 hrtimer_init(&cmd->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1433                 cmd->timer.function = null_cmd_timer_expired;
1434         }
1435         cmd->rq = bd->rq;
1436         cmd->nq = nq;
1437
1438         blk_mq_start_request(bd->rq);
1439
1440         if (should_requeue_request(bd->rq)) {
1441                 /*
1442                  * Alternate between hitting the core BUSY path, and the
1443                  * driver driven requeue path
1444                  */
1445                 nq->requeue_selection++;
1446                 if (nq->requeue_selection & 1)
1447                         return BLK_STS_RESOURCE;
1448                 else {
1449                         blk_mq_requeue_request(bd->rq, true);
1450                         return BLK_STS_OK;
1451                 }
1452         }
1453         if (should_timeout_request(bd->rq))
1454                 return BLK_STS_OK;
1455
1456         return null_handle_cmd(cmd);
1457 }
1458
1459 static const struct blk_mq_ops null_mq_ops = {
1460         .queue_rq       = null_queue_rq,
1461         .complete       = null_softirq_done_fn,
1462         .timeout        = null_timeout_rq,
1463 };
1464
1465 static void cleanup_queue(struct nullb_queue *nq)
1466 {
1467         kfree(nq->tag_map);
1468         kfree(nq->cmds);
1469 }
1470
1471 static void cleanup_queues(struct nullb *nullb)
1472 {
1473         int i;
1474
1475         for (i = 0; i < nullb->nr_queues; i++)
1476                 cleanup_queue(&nullb->queues[i]);
1477
1478         kfree(nullb->queues);
1479 }
1480
1481 static void null_del_dev(struct nullb *nullb)
1482 {
1483         struct nullb_device *dev = nullb->dev;
1484
1485         ida_simple_remove(&nullb_indexes, nullb->index);
1486
1487         list_del_init(&nullb->list);
1488
1489         del_gendisk(nullb->disk);
1490
1491         if (test_bit(NULLB_DEV_FL_THROTTLED, &nullb->dev->flags)) {
1492                 hrtimer_cancel(&nullb->bw_timer);
1493                 atomic_long_set(&nullb->cur_bytes, LONG_MAX);
1494                 null_restart_queue_async(nullb);
1495         }
1496
1497         blk_cleanup_queue(nullb->q);
1498         if (dev->queue_mode == NULL_Q_MQ &&
1499             nullb->tag_set == &nullb->__tag_set)
1500                 blk_mq_free_tag_set(nullb->tag_set);
1501         put_disk(nullb->disk);
1502         cleanup_queues(nullb);
1503         if (null_cache_active(nullb))
1504                 null_free_device_storage(nullb->dev, true);
1505         kfree(nullb);
1506         dev->nullb = NULL;
1507 }
1508
1509 static void null_config_discard(struct nullb *nullb)
1510 {
1511         if (nullb->dev->discard == false)
1512                 return;
1513         nullb->q->limits.discard_granularity = nullb->dev->blocksize;
1514         nullb->q->limits.discard_alignment = nullb->dev->blocksize;
1515         blk_queue_max_discard_sectors(nullb->q, UINT_MAX >> 9);
1516         blk_queue_flag_set(QUEUE_FLAG_DISCARD, nullb->q);
1517 }
1518
1519 static int null_open(struct block_device *bdev, fmode_t mode)
1520 {
1521         return 0;
1522 }
1523
1524 static void null_release(struct gendisk *disk, fmode_t mode)
1525 {
1526 }
1527
1528 static const struct block_device_operations null_fops = {
1529         .owner =        THIS_MODULE,
1530         .open =         null_open,
1531         .release =      null_release,
1532 };
1533
1534 static void null_init_queue(struct nullb *nullb, struct nullb_queue *nq)
1535 {
1536         BUG_ON(!nullb);
1537         BUG_ON(!nq);
1538
1539         init_waitqueue_head(&nq->wait);
1540         nq->queue_depth = nullb->queue_depth;
1541         nq->dev = nullb->dev;
1542 }
1543
1544 static void null_init_queues(struct nullb *nullb)
1545 {
1546         struct request_queue *q = nullb->q;
1547         struct blk_mq_hw_ctx *hctx;
1548         struct nullb_queue *nq;
1549         int i;
1550
1551         queue_for_each_hw_ctx(q, hctx, i) {
1552                 if (!hctx->nr_ctx || !hctx->tags)
1553                         continue;
1554                 nq = &nullb->queues[i];
1555                 hctx->driver_data = nq;
1556                 null_init_queue(nullb, nq);
1557                 nullb->nr_queues++;
1558         }
1559 }
1560
1561 static int setup_commands(struct nullb_queue *nq)
1562 {
1563         struct nullb_cmd *cmd;
1564         int i, tag_size;
1565
1566         nq->cmds = kcalloc(nq->queue_depth, sizeof(*cmd), GFP_KERNEL);
1567         if (!nq->cmds)
1568                 return -ENOMEM;
1569
1570         tag_size = ALIGN(nq->queue_depth, BITS_PER_LONG) / BITS_PER_LONG;
1571         nq->tag_map = kcalloc(tag_size, sizeof(unsigned long), GFP_KERNEL);
1572         if (!nq->tag_map) {
1573                 kfree(nq->cmds);
1574                 return -ENOMEM;
1575         }
1576
1577         for (i = 0; i < nq->queue_depth; i++) {
1578                 cmd = &nq->cmds[i];
1579                 INIT_LIST_HEAD(&cmd->list);
1580                 cmd->ll_list.next = NULL;
1581                 cmd->tag = -1U;
1582         }
1583
1584         return 0;
1585 }
1586
1587 static int setup_queues(struct nullb *nullb)
1588 {
1589         nullb->queues = kcalloc(nullb->dev->submit_queues,
1590                                 sizeof(struct nullb_queue),
1591                                 GFP_KERNEL);
1592         if (!nullb->queues)
1593                 return -ENOMEM;
1594
1595         nullb->nr_queues = 0;
1596         nullb->queue_depth = nullb->dev->hw_queue_depth;
1597
1598         return 0;
1599 }
1600
1601 static int init_driver_queues(struct nullb *nullb)
1602 {
1603         struct nullb_queue *nq;
1604         int i, ret = 0;
1605
1606         for (i = 0; i < nullb->dev->submit_queues; i++) {
1607                 nq = &nullb->queues[i];
1608
1609                 null_init_queue(nullb, nq);
1610
1611                 ret = setup_commands(nq);
1612                 if (ret)
1613                         return ret;
1614                 nullb->nr_queues++;
1615         }
1616         return 0;
1617 }
1618
1619 static int null_gendisk_register(struct nullb *nullb)
1620 {
1621         struct gendisk *disk;
1622         sector_t size;
1623
1624         disk = nullb->disk = alloc_disk_node(1, nullb->dev->home_node);
1625         if (!disk)
1626                 return -ENOMEM;
1627         size = (sector_t)nullb->dev->size * 1024 * 1024ULL;
1628         set_capacity(disk, size >> 9);
1629
1630         disk->flags |= GENHD_FL_EXT_DEVT | GENHD_FL_SUPPRESS_PARTITION_INFO;
1631         disk->major             = null_major;
1632         disk->first_minor       = nullb->index;
1633         disk->fops              = &null_fops;
1634         disk->private_data      = nullb;
1635         disk->queue             = nullb->q;
1636         strncpy(disk->disk_name, nullb->disk_name, DISK_NAME_LEN);
1637
1638         add_disk(disk);
1639         return 0;
1640 }
1641
1642 static int null_init_tag_set(struct nullb *nullb, struct blk_mq_tag_set *set)
1643 {
1644         set->ops = &null_mq_ops;
1645         set->nr_hw_queues = nullb ? nullb->dev->submit_queues :
1646                                                 g_submit_queues;
1647         set->queue_depth = nullb ? nullb->dev->hw_queue_depth :
1648                                                 g_hw_queue_depth;
1649         set->numa_node = nullb ? nullb->dev->home_node : g_home_node;
1650         set->cmd_size   = sizeof(struct nullb_cmd);
1651         set->flags = BLK_MQ_F_SHOULD_MERGE;
1652         if (g_no_sched)
1653                 set->flags |= BLK_MQ_F_NO_SCHED;
1654         set->driver_data = NULL;
1655
1656         if ((nullb && nullb->dev->blocking) || g_blocking)
1657                 set->flags |= BLK_MQ_F_BLOCKING;
1658
1659         return blk_mq_alloc_tag_set(set);
1660 }
1661
1662 static void null_validate_conf(struct nullb_device *dev)
1663 {
1664         dev->blocksize = round_down(dev->blocksize, 512);
1665         dev->blocksize = clamp_t(unsigned int, dev->blocksize, 512, 4096);
1666
1667         if (dev->queue_mode == NULL_Q_MQ && dev->use_per_node_hctx) {
1668                 if (dev->submit_queues != nr_online_nodes)
1669                         dev->submit_queues = nr_online_nodes;
1670         } else if (dev->submit_queues > nr_cpu_ids)
1671                 dev->submit_queues = nr_cpu_ids;
1672         else if (dev->submit_queues == 0)
1673                 dev->submit_queues = 1;
1674
1675         dev->queue_mode = min_t(unsigned int, dev->queue_mode, NULL_Q_MQ);
1676         dev->irqmode = min_t(unsigned int, dev->irqmode, NULL_IRQ_TIMER);
1677
1678         /* Do memory allocation, so set blocking */
1679         if (dev->memory_backed)
1680                 dev->blocking = true;
1681         else /* cache is meaningless */
1682                 dev->cache_size = 0;
1683         dev->cache_size = min_t(unsigned long, ULONG_MAX / 1024 / 1024,
1684                                                 dev->cache_size);
1685         dev->mbps = min_t(unsigned int, 1024 * 40, dev->mbps);
1686         /* can not stop a queue */
1687         if (dev->queue_mode == NULL_Q_BIO)
1688                 dev->mbps = 0;
1689 }
1690
1691 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1692 static bool __null_setup_fault(struct fault_attr *attr, char *str)
1693 {
1694         if (!str[0])
1695                 return true;
1696
1697         if (!setup_fault_attr(attr, str))
1698                 return false;
1699
1700         attr->verbose = 0;
1701         return true;
1702 }
1703 #endif
1704
1705 static bool null_setup_fault(void)
1706 {
1707 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1708         if (!__null_setup_fault(&null_timeout_attr, g_timeout_str))
1709                 return false;
1710         if (!__null_setup_fault(&null_requeue_attr, g_requeue_str))
1711                 return false;
1712 #endif
1713         return true;
1714 }
1715
1716 static int null_add_dev(struct nullb_device *dev)
1717 {
1718         struct nullb *nullb;
1719         int rv;
1720
1721         null_validate_conf(dev);
1722
1723         nullb = kzalloc_node(sizeof(*nullb), GFP_KERNEL, dev->home_node);
1724         if (!nullb) {
1725                 rv = -ENOMEM;
1726                 goto out;
1727         }
1728         nullb->dev = dev;
1729         dev->nullb = nullb;
1730
1731         spin_lock_init(&nullb->lock);
1732
1733         rv = setup_queues(nullb);
1734         if (rv)
1735                 goto out_free_nullb;
1736
1737         if (dev->queue_mode == NULL_Q_MQ) {
1738                 if (shared_tags) {
1739                         nullb->tag_set = &tag_set;
1740                         rv = 0;
1741                 } else {
1742                         nullb->tag_set = &nullb->__tag_set;
1743                         rv = null_init_tag_set(nullb, nullb->tag_set);
1744                 }
1745
1746                 if (rv)
1747                         goto out_cleanup_queues;
1748
1749                 if (!null_setup_fault())
1750                         goto out_cleanup_queues;
1751
1752                 nullb->tag_set->timeout = 5 * HZ;
1753                 nullb->q = blk_mq_init_queue(nullb->tag_set);
1754                 if (IS_ERR(nullb->q)) {
1755                         rv = -ENOMEM;
1756                         goto out_cleanup_tags;
1757                 }
1758                 null_init_queues(nullb);
1759         } else if (dev->queue_mode == NULL_Q_BIO) {
1760                 nullb->q = blk_alloc_queue_node(GFP_KERNEL, dev->home_node,
1761                                                 NULL);
1762                 if (!nullb->q) {
1763                         rv = -ENOMEM;
1764                         goto out_cleanup_queues;
1765                 }
1766                 blk_queue_make_request(nullb->q, null_queue_bio);
1767                 rv = init_driver_queues(nullb);
1768                 if (rv)
1769                         goto out_cleanup_blk_queue;
1770         } else {
1771                 nullb->q = blk_init_queue_node(null_request_fn, &nullb->lock,
1772                                                 dev->home_node);
1773                 if (!nullb->q) {
1774                         rv = -ENOMEM;
1775                         goto out_cleanup_queues;
1776                 }
1777
1778                 if (!null_setup_fault())
1779                         goto out_cleanup_blk_queue;
1780
1781                 blk_queue_prep_rq(nullb->q, null_rq_prep_fn);
1782                 blk_queue_softirq_done(nullb->q, null_softirq_done_fn);
1783                 blk_queue_rq_timed_out(nullb->q, null_rq_timed_out_fn);
1784                 nullb->q->rq_timeout = 5 * HZ;
1785                 rv = init_driver_queues(nullb);
1786                 if (rv)
1787                         goto out_cleanup_blk_queue;
1788         }
1789
1790         if (dev->mbps) {
1791                 set_bit(NULLB_DEV_FL_THROTTLED, &dev->flags);
1792                 nullb_setup_bwtimer(nullb);
1793         }
1794
1795         if (dev->cache_size > 0) {
1796                 set_bit(NULLB_DEV_FL_CACHE, &nullb->dev->flags);
1797                 blk_queue_write_cache(nullb->q, true, true);
1798                 blk_queue_flush_queueable(nullb->q, true);
1799         }
1800
1801         if (dev->zoned) {
1802                 rv = null_zone_init(dev);
1803                 if (rv)
1804                         goto out_cleanup_blk_queue;
1805
1806                 blk_queue_chunk_sectors(nullb->q, dev->zone_size_sects);
1807                 nullb->q->limits.zoned = BLK_ZONED_HM;
1808         }
1809
1810         nullb->q->queuedata = nullb;
1811         blk_queue_flag_set(QUEUE_FLAG_NONROT, nullb->q);
1812         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, nullb->q);
1813
1814         mutex_lock(&lock);
1815         nullb->index = ida_simple_get(&nullb_indexes, 0, 0, GFP_KERNEL);
1816         dev->index = nullb->index;
1817         mutex_unlock(&lock);
1818
1819         blk_queue_logical_block_size(nullb->q, dev->blocksize);
1820         blk_queue_physical_block_size(nullb->q, dev->blocksize);
1821
1822         null_config_discard(nullb);
1823
1824         sprintf(nullb->disk_name, "nullb%d", nullb->index);
1825
1826         rv = null_gendisk_register(nullb);
1827         if (rv)
1828                 goto out_cleanup_zone;
1829
1830         mutex_lock(&lock);
1831         list_add_tail(&nullb->list, &nullb_list);
1832         mutex_unlock(&lock);
1833
1834         return 0;
1835 out_cleanup_zone:
1836         if (dev->zoned)
1837                 null_zone_exit(dev);
1838 out_cleanup_blk_queue:
1839         blk_cleanup_queue(nullb->q);
1840 out_cleanup_tags:
1841         if (dev->queue_mode == NULL_Q_MQ && nullb->tag_set == &nullb->__tag_set)
1842                 blk_mq_free_tag_set(nullb->tag_set);
1843 out_cleanup_queues:
1844         cleanup_queues(nullb);
1845 out_free_nullb:
1846         kfree(nullb);
1847 out:
1848         return rv;
1849 }
1850
1851 static int __init null_init(void)
1852 {
1853         int ret = 0;
1854         unsigned int i;
1855         struct nullb *nullb;
1856         struct nullb_device *dev;
1857
1858         if (g_bs > PAGE_SIZE) {
1859                 pr_warn("null_blk: invalid block size\n");
1860                 pr_warn("null_blk: defaults block size to %lu\n", PAGE_SIZE);
1861                 g_bs = PAGE_SIZE;
1862         }
1863
1864         if (!is_power_of_2(g_zone_size)) {
1865                 pr_err("null_blk: zone_size must be power-of-two\n");
1866                 return -EINVAL;
1867         }
1868
1869         if (g_queue_mode == NULL_Q_MQ && g_use_per_node_hctx) {
1870                 if (g_submit_queues != nr_online_nodes) {
1871                         pr_warn("null_blk: submit_queues param is set to %u.\n",
1872                                                         nr_online_nodes);
1873                         g_submit_queues = nr_online_nodes;
1874                 }
1875         } else if (g_submit_queues > nr_cpu_ids)
1876                 g_submit_queues = nr_cpu_ids;
1877         else if (g_submit_queues <= 0)
1878                 g_submit_queues = 1;
1879
1880         if (g_queue_mode == NULL_Q_MQ && shared_tags) {
1881                 ret = null_init_tag_set(NULL, &tag_set);
1882                 if (ret)
1883                         return ret;
1884         }
1885
1886         config_group_init(&nullb_subsys.su_group);
1887         mutex_init(&nullb_subsys.su_mutex);
1888
1889         ret = configfs_register_subsystem(&nullb_subsys);
1890         if (ret)
1891                 goto err_tagset;
1892
1893         mutex_init(&lock);
1894
1895         null_major = register_blkdev(0, "nullb");
1896         if (null_major < 0) {
1897                 ret = null_major;
1898                 goto err_conf;
1899         }
1900
1901         for (i = 0; i < nr_devices; i++) {
1902                 dev = null_alloc_dev();
1903                 if (!dev) {
1904                         ret = -ENOMEM;
1905                         goto err_dev;
1906                 }
1907                 ret = null_add_dev(dev);
1908                 if (ret) {
1909                         null_free_dev(dev);
1910                         goto err_dev;
1911                 }
1912         }
1913
1914         pr_info("null: module loaded\n");
1915         return 0;
1916
1917 err_dev:
1918         while (!list_empty(&nullb_list)) {
1919                 nullb = list_entry(nullb_list.next, struct nullb, list);
1920                 dev = nullb->dev;
1921                 null_del_dev(nullb);
1922                 null_free_dev(dev);
1923         }
1924         unregister_blkdev(null_major, "nullb");
1925 err_conf:
1926         configfs_unregister_subsystem(&nullb_subsys);
1927 err_tagset:
1928         if (g_queue_mode == NULL_Q_MQ && shared_tags)
1929                 blk_mq_free_tag_set(&tag_set);
1930         return ret;
1931 }
1932
1933 static void __exit null_exit(void)
1934 {
1935         struct nullb *nullb;
1936
1937         configfs_unregister_subsystem(&nullb_subsys);
1938
1939         unregister_blkdev(null_major, "nullb");
1940
1941         mutex_lock(&lock);
1942         while (!list_empty(&nullb_list)) {
1943                 struct nullb_device *dev;
1944
1945                 nullb = list_entry(nullb_list.next, struct nullb, list);
1946                 dev = nullb->dev;
1947                 null_del_dev(nullb);
1948                 null_free_dev(dev);
1949         }
1950         mutex_unlock(&lock);
1951
1952         if (g_queue_mode == NULL_Q_MQ && shared_tags)
1953                 blk_mq_free_tag_set(&tag_set);
1954 }
1955
1956 module_init(null_init);
1957 module_exit(null_exit);
1958
1959 MODULE_AUTHOR("Jens Axboe <axboe@kernel.dk>");
1960 MODULE_LICENSE("GPL");