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->error = BLK_STS_OK;
 575                 cmd->nq = nq;
 576                 if (nq->dev->irqmode == NULL_IRQ_TIMER) {
 577                         hrtimer_init(&cmd->timer, CLOCK_MONOTONIC,
 578                                      HRTIMER_MODE_REL);
 579                         cmd->timer.function = null_cmd_timer_expired;
 580                 }
 581                 return cmd;
 582         }
 583
 584         return NULL;
 585 }
 586
 587 static struct nullb_cmd *alloc_cmd(struct nullb_queue *nq, int can_wait)
 588 {
 589         struct nullb_cmd *cmd;
 590         DEFINE_WAIT(wait);
 591
 592         cmd = __alloc_cmd(nq);
 593         if (cmd || !can_wait)
 594                 return cmd;
 595
 596         do {
 597                 prepare_to_wait(&nq->wait, &wait, TASK_UNINTERRUPTIBLE);
 598                 cmd = __alloc_cmd(nq);
 599                 if (cmd)
 600                         break;
 601
 602                 io_schedule();
 603         } while (1);
 604
 605         finish_wait(&nq->wait, &wait);
 606         return cmd;
 607 }
 608
 609 static void end_cmd(struct nullb_cmd *cmd)
 610 {
 611         struct request_queue *q = NULL;
 612         int queue_mode = cmd->nq->dev->queue_mode;
 613
 614         if (cmd->rq)
 615                 q = cmd->rq->q;
 616
 617         switch (queue_mode)  {
 618         case NULL_Q_MQ:
 619                 blk_mq_end_request(cmd->rq, cmd->error);
 620                 return;
 621         case NULL_Q_RQ:
 622                 INIT_LIST_HEAD(&cmd->rq->queuelist);
 623                 blk_end_request_all(cmd->rq, cmd->error);
 624                 break;
 625         case NULL_Q_BIO:
 626                 cmd->bio->bi_status = cmd->error;
 627                 bio_endio(cmd->bio);
 628                 break;
 629         }
 630
 631         free_cmd(cmd);
 632
 633         /* Restart queue if needed, as we are freeing a tag */
 634         if (queue_mode == NULL_Q_RQ && blk_queue_stopped(q)) {
 635                 unsigned long flags;
 636
 637                 spin_lock_irqsave(q->queue_lock, flags);
 638                 blk_start_queue_async(q);
 639                 spin_unlock_irqrestore(q->queue_lock, flags);
 640         }
 641 }
 642
 643 static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer)
 644 {
 645         end_cmd(container_of(timer, struct nullb_cmd, timer));
 646
 647         return HRTIMER_NORESTART;
 648 }
 649
 650 static void null_cmd_end_timer(struct nullb_cmd *cmd)
 651 {
 652         ktime_t kt = cmd->nq->dev->completion_nsec;
 653
 654         hrtimer_start(&cmd->timer, kt, HRTIMER_MODE_REL);
 655 }
 656
 657 static void null_softirq_done_fn(struct request *rq)
 658 {
 659         struct nullb *nullb = rq->q->queuedata;
 660
 661         if (nullb->dev->queue_mode == NULL_Q_MQ)
 662                 end_cmd(blk_mq_rq_to_pdu(rq));
 663         else
 664                 end_cmd(rq->special);
 665 }
 666
 667 static struct nullb_page *null_alloc_page(gfp_t gfp_flags)
 668 {
 669         struct nullb_page *t_page;
 670
 671         t_page = kmalloc(sizeof(struct nullb_page), gfp_flags);
 672         if (!t_page)
 673                 goto out;
 674
 675         t_page->page = alloc_pages(gfp_flags, 0);
 676         if (!t_page->page)
 677                 goto out_freepage;
 678
 679         memset(t_page->bitmap, 0, sizeof(t_page->bitmap));
 680         return t_page;
 681 out_freepage:
 682         kfree(t_page);
 683 out:
 684         return NULL;
 685 }
 686
 687 static void null_free_page(struct nullb_page *t_page)
 688 {
 689         __set_bit(NULLB_PAGE_FREE, t_page->bitmap);
 690         if (test_bit(NULLB_PAGE_LOCK, t_page->bitmap))
 691                 return;
 692         __free_page(t_page->page);
 693         kfree(t_page);
 694 }
 695
 696 static bool null_page_empty(struct nullb_page *page)
 697 {
 698         int size = MAP_SZ - 2;
 699
 700         return find_first_bit(page->bitmap, size) == size;
 701 }
 702
 703 static void null_free_sector(struct nullb *nullb, sector_t sector,
 704         bool is_cache)
 705 {
 706         unsigned int sector_bit;
 707         u64 idx;
 708         struct nullb_page *t_page, *ret;
 709         struct radix_tree_root *root;
 710
 711         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 712         idx = sector >> PAGE_SECTORS_SHIFT;
 713         sector_bit = (sector & SECTOR_MASK);
 714
 715         t_page = radix_tree_lookup(root, idx);
 716         if (t_page) {
 717                 __clear_bit(sector_bit, t_page->bitmap);
 718
 719                 if (null_page_empty(t_page)) {
 720                         ret = radix_tree_delete_item(root, idx, t_page);
 721                         WARN_ON(ret != t_page);
 722                         null_free_page(ret);
 723                         if (is_cache)
 724                                 nullb->dev->curr_cache -= PAGE_SIZE;
 725                 }
 726         }
 727 }
 728
 729 static struct nullb_page *null_radix_tree_insert(struct nullb *nullb, u64 idx,
 730         struct nullb_page *t_page, bool is_cache)
 731 {
 732         struct radix_tree_root *root;
 733
 734         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 735
 736         if (radix_tree_insert(root, idx, t_page)) {
 737                 null_free_page(t_page);
 738                 t_page = radix_tree_lookup(root, idx);
 739                 WARN_ON(!t_page || t_page->page->index != idx);
 740         } else if (is_cache)
 741                 nullb->dev->curr_cache += PAGE_SIZE;
 742
 743         return t_page;
 744 }
 745
 746 static void null_free_device_storage(struct nullb_device *dev, bool is_cache)
 747 {
 748         unsigned long pos = 0;
 749         int nr_pages;
 750         struct nullb_page *ret, *t_pages[FREE_BATCH];
 751         struct radix_tree_root *root;
 752
 753         root = is_cache ? &dev->cache : &dev->data;
 754
 755         do {
 756                 int i;
 757
 758                 nr_pages = radix_tree_gang_lookup(root,
 759                                 (void **)t_pages, pos, FREE_BATCH);
 760
 761                 for (i = 0; i < nr_pages; i++) {
 762                         pos = t_pages[i]->page->index;
 763                         ret = radix_tree_delete_item(root, pos, t_pages[i]);
 764                         WARN_ON(ret != t_pages[i]);
 765                         null_free_page(ret);
 766                 }
 767
 768                 pos++;
 769         } while (nr_pages == FREE_BATCH);
 770
 771         if (is_cache)
 772                 dev->curr_cache = 0;
 773 }
 774
 775 static struct nullb_page *__null_lookup_page(struct nullb *nullb,
 776         sector_t sector, bool for_write, bool is_cache)
 777 {
 778         unsigned int sector_bit;
 779         u64 idx;
 780         struct nullb_page *t_page;
 781         struct radix_tree_root *root;
 782
 783         idx = sector >> PAGE_SECTORS_SHIFT;
 784         sector_bit = (sector & SECTOR_MASK);
 785
 786         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 787         t_page = radix_tree_lookup(root, idx);
 788         WARN_ON(t_page && t_page->page->index != idx);
 789
 790         if (t_page && (for_write || test_bit(sector_bit, t_page->bitmap)))
 791                 return t_page;
 792
 793         return NULL;
 794 }
 795
 796 static struct nullb_page *null_lookup_page(struct nullb *nullb,
 797         sector_t sector, bool for_write, bool ignore_cache)
 798 {
 799         struct nullb_page *page = NULL;
 800
 801         if (!ignore_cache)
 802                 page = __null_lookup_page(nullb, sector, for_write, true);
 803         if (page)
 804                 return page;
 805         return __null_lookup_page(nullb, sector, for_write, false);
 806 }
 807
 808 static struct nullb_page *null_insert_page(struct nullb *nullb,
 809                                            sector_t sector, bool ignore_cache)
 810         __releases(&nullb->lock)
 811         __acquires(&nullb->lock)
 812 {
 813         u64 idx;
 814         struct nullb_page *t_page;
 815
 816         t_page = null_lookup_page(nullb, sector, true, ignore_cache);
 817         if (t_page)
 818                 return t_page;
 819
 820         spin_unlock_irq(&nullb->lock);
 821
 822         t_page = null_alloc_page(GFP_NOIO);
 823         if (!t_page)
 824                 goto out_lock;
 825
 826         if (radix_tree_preload(GFP_NOIO))
 827                 goto out_freepage;
 828
 829         spin_lock_irq(&nullb->lock);
 830         idx = sector >> PAGE_SECTORS_SHIFT;
 831         t_page->page->index = idx;
 832         t_page = null_radix_tree_insert(nullb, idx, t_page, !ignore_cache);
 833         radix_tree_preload_end();
 834
 835         return t_page;
 836 out_freepage:
 837         null_free_page(t_page);
 838 out_lock:
 839         spin_lock_irq(&nullb->lock);
 840         return null_lookup_page(nullb, sector, true, ignore_cache);
 841 }
 842
 843 static int null_flush_cache_page(struct nullb *nullb, struct nullb_page *c_page)
 844 {
 845         int i;
 846         unsigned int offset;
 847         u64 idx;
 848         struct nullb_page *t_page, *ret;
 849         void *dst, *src;
 850
 851         idx = c_page->page->index;
 852
 853         t_page = null_insert_page(nullb, idx << PAGE_SECTORS_SHIFT, true);
 854
 855         __clear_bit(NULLB_PAGE_LOCK, c_page->bitmap);
 856         if (test_bit(NULLB_PAGE_FREE, c_page->bitmap)) {
 857                 null_free_page(c_page);
 858                 if (t_page && null_page_empty(t_page)) {
 859                         ret = radix_tree_delete_item(&nullb->dev->data,
 860                                 idx, t_page);
 861                         null_free_page(t_page);
 862                 }
 863                 return 0;
 864         }
 865
 866         if (!t_page)
 867                 return -ENOMEM;
 868
 869         src = kmap_atomic(c_page->page);
 870         dst = kmap_atomic(t_page->page);
 871
 872         for (i = 0; i < PAGE_SECTORS;
 873                         i += (nullb->dev->blocksize >> SECTOR_SHIFT)) {
 874                 if (test_bit(i, c_page->bitmap)) {
 875                         offset = (i << SECTOR_SHIFT);
 876                         memcpy(dst + offset, src + offset,
 877                                 nullb->dev->blocksize);
 878                         __set_bit(i, t_page->bitmap);
 879                 }
 880         }
 881
 882         kunmap_atomic(dst);
 883         kunmap_atomic(src);
 884
 885         ret = radix_tree_delete_item(&nullb->dev->cache, idx, c_page);
 886         null_free_page(ret);
 887         nullb->dev->curr_cache -= PAGE_SIZE;
 888
 889         return 0;
 890 }
 891
 892 static int null_make_cache_space(struct nullb *nullb, unsigned long n)
 893 {
 894         int i, err, nr_pages;
 895         struct nullb_page *c_pages[FREE_BATCH];
 896         unsigned long flushed = 0, one_round;
 897
 898 again:
 899         if ((nullb->dev->cache_size * 1024 * 1024) >
 900              nullb->dev->curr_cache + n || nullb->dev->curr_cache == 0)
 901                 return 0;
 902
 903         nr_pages = radix_tree_gang_lookup(&nullb->dev->cache,
 904                         (void **)c_pages, nullb->cache_flush_pos, FREE_BATCH);
 905         /*
 906          * nullb_flush_cache_page could unlock before using the c_pages. To
 907          * avoid race, we don't allow page free
 908          */
 909         for (i = 0; i < nr_pages; i++) {
 910                 nullb->cache_flush_pos = c_pages[i]->page->index;
 911                 /*
 912                  * We found the page which is being flushed to disk by other
 913                  * threads
 914                  */
 915                 if (test_bit(NULLB_PAGE_LOCK, c_pages[i]->bitmap))
 916                         c_pages[i] = NULL;
 917                 else
 918                         __set_bit(NULLB_PAGE_LOCK, c_pages[i]->bitmap);
 919         }
 920
 921         one_round = 0;
 922         for (i = 0; i < nr_pages; i++) {
 923                 if (c_pages[i] == NULL)
 924                         continue;
 925                 err = null_flush_cache_page(nullb, c_pages[i]);
 926                 if (err)
 927                         return err;
 928                 one_round++;
 929         }
 930         flushed += one_round << PAGE_SHIFT;
 931
 932         if (n > flushed) {
 933                 if (nr_pages == 0)
 934                         nullb->cache_flush_pos = 0;
 935                 if (one_round == 0) {
 936                         /* give other threads a chance */
 937                         spin_unlock_irq(&nullb->lock);
 938                         spin_lock_irq(&nullb->lock);
 939                 }
 940                 goto again;
 941         }
 942         return 0;
 943 }
 944
 945 static int copy_to_nullb(struct nullb *nullb, struct page *source,
 946         unsigned int off, sector_t sector, size_t n, bool is_fua)
 947 {
 948         size_t temp, count = 0;
 949         unsigned int offset;
 950         struct nullb_page *t_page;
 951         void *dst, *src;
 952
 953         while (count < n) {
 954                 temp = min_t(size_t, nullb->dev->blocksize, n - count);
 955
 956                 if (null_cache_active(nullb) && !is_fua)
 957                         null_make_cache_space(nullb, PAGE_SIZE);
 958
 959                 offset = (sector & SECTOR_MASK) << SECTOR_SHIFT;
 960                 t_page = null_insert_page(nullb, sector,
 961                         !null_cache_active(nullb) || is_fua);
 962                 if (!t_page)
 963                         return -ENOSPC;
 964
 965                 src = kmap_atomic(source);
 966                 dst = kmap_atomic(t_page->page);
 967                 memcpy(dst + offset, src + off + count, temp);
 968                 kunmap_atomic(dst);
 969                 kunmap_atomic(src);
 970
 971                 __set_bit(sector & SECTOR_MASK, t_page->bitmap);
 972
 973                 if (is_fua)
 974                         null_free_sector(nullb, sector, true);
 975
 976                 count += temp;
 977                 sector += temp >> SECTOR_SHIFT;
 978         }
 979         return 0;
 980 }
 981
 982 static int copy_from_nullb(struct nullb *nullb, struct page *dest,
 983         unsigned int off, sector_t sector, size_t n)
 984 {
 985         size_t temp, count = 0;
 986         unsigned int offset;
 987         struct nullb_page *t_page;
 988         void *dst, *src;
 989
 990         while (count < n) {
 991                 temp = min_t(size_t, nullb->dev->blocksize, n - count);
 992
 993                 offset = (sector & SECTOR_MASK) << SECTOR_SHIFT;
 994                 t_page = null_lookup_page(nullb, sector, false,
 995                         !null_cache_active(nullb));
 996
 997                 dst = kmap_atomic(dest);
 998                 if (!t_page) {
 999                         memset(dst + off + count, 0, temp);
1000                         goto next;
1001                 }
1002                 src = kmap_atomic(t_page->page);
1003                 memcpy(dst + off + count, src + offset, temp);
1004                 kunmap_atomic(src);
1005 next:
1006                 kunmap_atomic(dst);
1007
1008                 count += temp;
1009                 sector += temp >> SECTOR_SHIFT;
1010         }
1011         return 0;
1012 }
1013
1014 static void null_handle_discard(struct nullb *nullb, sector_t sector, size_t n)
1015 {
1016         size_t temp;
1017
1018         spin_lock_irq(&nullb->lock);
1019         while (n > 0) {
1020                 temp = min_t(size_t, n, nullb->dev->blocksize);
1021                 null_free_sector(nullb, sector, false);
1022                 if (null_cache_active(nullb))
1023                         null_free_sector(nullb, sector, true);
1024                 sector += temp >> SECTOR_SHIFT;
1025                 n -= temp;
1026         }
1027         spin_unlock_irq(&nullb->lock);
1028 }
1029
1030 static int null_handle_flush(struct nullb *nullb)
1031 {
1032         int err;
1033
1034         if (!null_cache_active(nullb))
1035                 return 0;
1036
1037         spin_lock_irq(&nullb->lock);
1038         while (true) {
1039                 err = null_make_cache_space(nullb,
1040                         nullb->dev->cache_size * 1024 * 1024);
1041                 if (err || nullb->dev->curr_cache == 0)
1042                         break;
1043         }
1044
1045         WARN_ON(!radix_tree_empty(&nullb->dev->cache));
1046         spin_unlock_irq(&nullb->lock);
1047         return err;
1048 }
1049
1050 static int null_transfer(struct nullb *nullb, struct page *page,
1051         unsigned int len, unsigned int off, bool is_write, sector_t sector,
1052         bool is_fua)
1053 {
1054         int err = 0;
1055
1056         if (!is_write) {
1057                 err = copy_from_nullb(nullb, page, off, sector, len);
1058                 flush_dcache_page(page);
1059         } else {
1060                 flush_dcache_page(page);
1061                 err = copy_to_nullb(nullb, page, off, sector, len, is_fua);
1062         }
1063
1064         return err;
1065 }
1066
1067 static int null_handle_rq(struct nullb_cmd *cmd)
1068 {
1069         struct request *rq = cmd->rq;
1070         struct nullb *nullb = cmd->nq->dev->nullb;
1071         int err;
1072         unsigned int len;
1073         sector_t sector;
1074         struct req_iterator iter;
1075         struct bio_vec bvec;
1076
1077         sector = blk_rq_pos(rq);
1078
1079         if (req_op(rq) == REQ_OP_DISCARD) {
1080                 null_handle_discard(nullb, sector, blk_rq_bytes(rq));
1081                 return 0;
1082         }
1083
1084         spin_lock_irq(&nullb->lock);
1085         rq_for_each_segment(bvec, rq, iter) {
1086                 len = bvec.bv_len;
1087                 err = null_transfer(nullb, bvec.bv_page, len, bvec.bv_offset,
1088                                      op_is_write(req_op(rq)), sector,
1089                                      rq->cmd_flags & REQ_FUA);
1090                 if (err) {
1091                         spin_unlock_irq(&nullb->lock);
1092                         return err;
1093                 }
1094                 sector += len >> SECTOR_SHIFT;
1095         }
1096         spin_unlock_irq(&nullb->lock);
1097
1098         return 0;
1099 }
1100
1101 static int null_handle_bio(struct nullb_cmd *cmd)
1102 {
1103         struct bio *bio = cmd->bio;
1104         struct nullb *nullb = cmd->nq->dev->nullb;
1105         int err;
1106         unsigned int len;
1107         sector_t sector;
1108         struct bio_vec bvec;
1109         struct bvec_iter iter;
1110
1111         sector = bio->bi_iter.bi_sector;
1112
1113         if (bio_op(bio) == REQ_OP_DISCARD) {
1114                 null_handle_discard(nullb, sector,
1115                         bio_sectors(bio) << SECTOR_SHIFT);
1116                 return 0;
1117         }
1118
1119         spin_lock_irq(&nullb->lock);
1120         bio_for_each_segment(bvec, bio, iter) {
1121                 len = bvec.bv_len;
1122                 err = null_transfer(nullb, bvec.bv_page, len, bvec.bv_offset,
1123                                      op_is_write(bio_op(bio)), sector,
1124                                      bio_op(bio) & REQ_FUA);
1125                 if (err) {
1126                         spin_unlock_irq(&nullb->lock);
1127                         return err;
1128                 }
1129                 sector += len >> SECTOR_SHIFT;
1130         }
1131         spin_unlock_irq(&nullb->lock);
1132         return 0;
1133 }
1134
1135 static void null_stop_queue(struct nullb *nullb)
1136 {
1137         struct request_queue *q = nullb->q;
1138
1139         if (nullb->dev->queue_mode == NULL_Q_MQ)
1140                 blk_mq_stop_hw_queues(q);
1141         else {
1142                 spin_lock_irq(q->queue_lock);
1143                 blk_stop_queue(q);
1144                 spin_unlock_irq(q->queue_lock);
1145         }
1146 }
1147
1148 static void null_restart_queue_async(struct nullb *nullb)
1149 {
1150         struct request_queue *q = nullb->q;
1151         unsigned long flags;
1152
1153         if (nullb->dev->queue_mode == NULL_Q_MQ)
1154                 blk_mq_start_stopped_hw_queues(q, true);
1155         else {
1156                 spin_lock_irqsave(q->queue_lock, flags);
1157                 blk_start_queue_async(q);
1158                 spin_unlock_irqrestore(q->queue_lock, flags);
1159         }
1160 }
1161
1162 static bool cmd_report_zone(struct nullb *nullb, struct nullb_cmd *cmd)
1163 {
1164         struct nullb_device *dev = cmd->nq->dev;
1165
1166         if (dev->queue_mode == NULL_Q_BIO) {
1167                 if (bio_op(cmd->bio) == REQ_OP_ZONE_REPORT) {
1168                         cmd->error = null_zone_report(nullb, cmd->bio);
1169                         return true;
1170                 }
1171         } else {
1172                 if (req_op(cmd->rq) == REQ_OP_ZONE_REPORT) {
1173                         cmd->error = null_zone_report(nullb, cmd->rq->bio);
1174                         return true;
1175                 }
1176         }
1177
1178         return false;
1179 }
1180
1181 static blk_status_t null_handle_cmd(struct nullb_cmd *cmd)
1182 {
1183         struct nullb_device *dev = cmd->nq->dev;
1184         struct nullb *nullb = dev->nullb;
1185         int err = 0;
1186
1187         if (cmd_report_zone(nullb, cmd))
1188                 goto out;
1189
1190         if (test_bit(NULLB_DEV_FL_THROTTLED, &dev->flags)) {
1191                 struct request *rq = cmd->rq;
1192
1193                 if (!hrtimer_active(&nullb->bw_timer))
1194                         hrtimer_restart(&nullb->bw_timer);
1195
1196                 if (atomic_long_sub_return(blk_rq_bytes(rq),
1197                                 &nullb->cur_bytes) < 0) {
1198                         null_stop_queue(nullb);
1199                         /* race with timer */
1200                         if (atomic_long_read(&nullb->cur_bytes) > 0)
1201                                 null_restart_queue_async(nullb);
1202                         if (dev->queue_mode == NULL_Q_RQ) {
1203                                 struct request_queue *q = nullb->q;
1204
1205                                 spin_lock_irq(q->queue_lock);
1206                                 rq->rq_flags |= RQF_DONTPREP;
1207                                 blk_requeue_request(q, rq);
1208                                 spin_unlock_irq(q->queue_lock);
1209                                 return BLK_STS_OK;
1210                         } else
1211                                 /* requeue request */
1212                                 return BLK_STS_DEV_RESOURCE;
1213                 }
1214         }
1215
1216         if (nullb->dev->badblocks.shift != -1) {
1217                 int bad_sectors;
1218                 sector_t sector, size, first_bad;
1219                 bool is_flush = true;
1220
1221                 if (dev->queue_mode == NULL_Q_BIO &&
1222                                 bio_op(cmd->bio) != REQ_OP_FLUSH) {
1223                         is_flush = false;
1224                         sector = cmd->bio->bi_iter.bi_sector;
1225                         size = bio_sectors(cmd->bio);
1226                 }
1227                 if (dev->queue_mode != NULL_Q_BIO &&
1228                                 req_op(cmd->rq) != REQ_OP_FLUSH) {
1229                         is_flush = false;
1230                         sector = blk_rq_pos(cmd->rq);
1231                         size = blk_rq_sectors(cmd->rq);
1232                 }
1233                 if (!is_flush && badblocks_check(&nullb->dev->badblocks, sector,
1234                                 size, &first_bad, &bad_sectors)) {
1235                         cmd->error = BLK_STS_IOERR;
1236                         goto out;
1237                 }
1238         }
1239
1240         if (dev->memory_backed) {
1241                 if (dev->queue_mode == NULL_Q_BIO) {
1242                         if (bio_op(cmd->bio) == REQ_OP_FLUSH)
1243                                 err = null_handle_flush(nullb);
1244                         else
1245                                 err = null_handle_bio(cmd);
1246                 } else {
1247                         if (req_op(cmd->rq) == REQ_OP_FLUSH)
1248                                 err = null_handle_flush(nullb);
1249                         else
1250                                 err = null_handle_rq(cmd);
1251                 }
1252         }
1253         cmd->error = errno_to_blk_status(err);
1254
1255         if (!cmd->error && dev->zoned) {
1256                 sector_t sector;
1257                 unsigned int nr_sectors;
1258                 int op;
1259
1260                 if (dev->queue_mode == NULL_Q_BIO) {
1261                         op = bio_op(cmd->bio);
1262                         sector = cmd->bio->bi_iter.bi_sector;
1263                         nr_sectors = cmd->bio->bi_iter.bi_size >> 9;
1264                 } else {
1265                         op = req_op(cmd->rq);
1266                         sector = blk_rq_pos(cmd->rq);
1267                         nr_sectors = blk_rq_sectors(cmd->rq);
1268                 }
1269
1270                 if (op == REQ_OP_WRITE)
1271                         null_zone_write(cmd, sector, nr_sectors);
1272                 else if (op == REQ_OP_ZONE_RESET)
1273                         null_zone_reset(cmd, sector);
1274         }
1275 out:
1276         /* Complete IO by inline, softirq or timer */
1277         switch (dev->irqmode) {
1278         case NULL_IRQ_SOFTIRQ:
1279                 switch (dev->queue_mode)  {
1280                 case NULL_Q_MQ:
1281                         blk_mq_complete_request(cmd->rq);
1282                         break;
1283                 case NULL_Q_RQ:
1284                         blk_complete_request(cmd->rq);
1285                         break;
1286                 case NULL_Q_BIO:
1287                         /*
1288                          * XXX: no proper submitting cpu information available.
1289                          */
1290                         end_cmd(cmd);
1291                         break;
1292                 }
1293                 break;
1294         case NULL_IRQ_NONE:
1295                 end_cmd(cmd);
1296                 break;
1297         case NULL_IRQ_TIMER:
1298                 null_cmd_end_timer(cmd);
1299                 break;
1300         }
1301         return BLK_STS_OK;
1302 }
1303
1304 static enum hrtimer_restart nullb_bwtimer_fn(struct hrtimer *timer)
1305 {
1306         struct nullb *nullb = container_of(timer, struct nullb, bw_timer);
1307         ktime_t timer_interval = ktime_set(0, TIMER_INTERVAL);
1308         unsigned int mbps = nullb->dev->mbps;
1309
1310         if (atomic_long_read(&nullb->cur_bytes) == mb_per_tick(mbps))
1311                 return HRTIMER_NORESTART;
1312
1313         atomic_long_set(&nullb->cur_bytes, mb_per_tick(mbps));
1314         null_restart_queue_async(nullb);
1315
1316         hrtimer_forward_now(&nullb->bw_timer, timer_interval);
1317
1318         return HRTIMER_RESTART;
1319 }
1320
1321 static void nullb_setup_bwtimer(struct nullb *nullb)
1322 {
1323         ktime_t timer_interval = ktime_set(0, TIMER_INTERVAL);
1324
1325         hrtimer_init(&nullb->bw_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1326         nullb->bw_timer.function = nullb_bwtimer_fn;
1327         atomic_long_set(&nullb->cur_bytes, mb_per_tick(nullb->dev->mbps));
1328         hrtimer_start(&nullb->bw_timer, timer_interval, HRTIMER_MODE_REL);
1329 }
1330
1331 static struct nullb_queue *nullb_to_queue(struct nullb *nullb)
1332 {
1333         int index = 0;
1334
1335         if (nullb->nr_queues != 1)
1336                 index = raw_smp_processor_id() / ((nr_cpu_ids + nullb->nr_queues - 1) / nullb->nr_queues);
1337
1338         return &nullb->queues[index];
1339 }
1340
1341 static blk_qc_t null_queue_bio(struct request_queue *q, struct bio *bio)
1342 {
1343         struct nullb *nullb = q->queuedata;
1344         struct nullb_queue *nq = nullb_to_queue(nullb);
1345         struct nullb_cmd *cmd;
1346
1347         cmd = alloc_cmd(nq, 1);
1348         cmd->bio = bio;
1349
1350         null_handle_cmd(cmd);
1351         return BLK_QC_T_NONE;
1352 }
1353
1354 static enum blk_eh_timer_return null_rq_timed_out_fn(struct request *rq)
1355 {
1356         pr_info("null: rq %p timed out\n", rq);
1357         __blk_complete_request(rq);
1358         return BLK_EH_DONE;
1359 }
1360
1361 static int null_rq_prep_fn(struct request_queue *q, struct request *req)
1362 {
1363         struct nullb *nullb = q->queuedata;
1364         struct nullb_queue *nq = nullb_to_queue(nullb);
1365         struct nullb_cmd *cmd;
1366
1367         cmd = alloc_cmd(nq, 0);
1368         if (cmd) {
1369                 cmd->rq = req;
1370                 req->special = cmd;
1371                 return BLKPREP_OK;
1372         }
1373         blk_stop_queue(q);
1374
1375         return BLKPREP_DEFER;
1376 }
1377
1378 static bool should_timeout_request(struct request *rq)
1379 {
1380 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1381         if (g_timeout_str[0])
1382                 return should_fail(&null_timeout_attr, 1);
1383 #endif
1384         return false;
1385 }
1386
1387 static bool should_requeue_request(struct request *rq)
1388 {
1389 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1390         if (g_requeue_str[0])
1391                 return should_fail(&null_requeue_attr, 1);
1392 #endif
1393         return false;
1394 }
1395
1396 static void null_request_fn(struct request_queue *q)
1397 {
1398         struct request *rq;
1399
1400         while ((rq = blk_fetch_request(q)) != NULL) {
1401                 struct nullb_cmd *cmd = rq->special;
1402
1403                 /* just ignore the request */
1404                 if (should_timeout_request(rq))
1405                         continue;
1406                 if (should_requeue_request(rq)) {
1407                         blk_requeue_request(q, rq);
1408                         continue;
1409                 }
1410
1411                 spin_unlock_irq(q->queue_lock);
1412                 null_handle_cmd(cmd);
1413                 spin_lock_irq(q->queue_lock);
1414         }
1415 }
1416
1417 static enum blk_eh_timer_return null_timeout_rq(struct request *rq, bool res)
1418 {
1419         pr_info("null: rq %p timed out\n", rq);
1420         blk_mq_complete_request(rq);
1421         return BLK_EH_DONE;
1422 }
1423
1424 static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx,
1425                          const struct blk_mq_queue_data *bd)
1426 {
1427         struct nullb_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1428         struct nullb_queue *nq = hctx->driver_data;
1429
1430         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1431
1432         if (nq->dev->irqmode == NULL_IRQ_TIMER) {
1433                 hrtimer_init(&cmd->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1434                 cmd->timer.function = null_cmd_timer_expired;
1435         }
1436         cmd->rq = bd->rq;
1437         cmd->error = BLK_STS_OK;
1438         cmd->nq = nq;
1439
1440         blk_mq_start_request(bd->rq);
1441
1442         if (should_requeue_request(bd->rq)) {
1443                 /*
1444                  * Alternate between hitting the core BUSY path, and the
1445                  * driver driven requeue path
1446                  */
1447                 nq->requeue_selection++;
1448                 if (nq->requeue_selection & 1)
1449                         return BLK_STS_RESOURCE;
1450                 else {
1451                         blk_mq_requeue_request(bd->rq, true);
1452                         return BLK_STS_OK;
1453                 }
1454         }
1455         if (should_timeout_request(bd->rq))
1456                 return BLK_STS_OK;
1457
1458         return null_handle_cmd(cmd);
1459 }
1460
1461 static const struct blk_mq_ops null_mq_ops = {
1462         .queue_rq       = null_queue_rq,
1463         .complete       = null_softirq_done_fn,
1464         .timeout        = null_timeout_rq,
1465 };
1466
1467 static void cleanup_queue(struct nullb_queue *nq)
1468 {
1469         kfree(nq->tag_map);
1470         kfree(nq->cmds);
1471 }
1472
1473 static void cleanup_queues(struct nullb *nullb)
1474 {
1475         int i;
1476
1477         for (i = 0; i < nullb->nr_queues; i++)
1478                 cleanup_queue(&nullb->queues[i]);
1479
1480         kfree(nullb->queues);
1481 }
1482
1483 static void null_del_dev(struct nullb *nullb)
1484 {
1485         struct nullb_device *dev;
1486
1487         if (!nullb)
1488                 return;
1489
1490         dev = nullb->dev;
1491
1492         ida_simple_remove(&nullb_indexes, nullb->index);
1493
1494         list_del_init(&nullb->list);
1495
1496         del_gendisk(nullb->disk);
1497
1498         if (test_bit(NULLB_DEV_FL_THROTTLED, &nullb->dev->flags)) {
1499                 hrtimer_cancel(&nullb->bw_timer);
1500                 atomic_long_set(&nullb->cur_bytes, LONG_MAX);
1501                 null_restart_queue_async(nullb);
1502         }
1503
1504         blk_cleanup_queue(nullb->q);
1505         if (dev->queue_mode == NULL_Q_MQ &&
1506             nullb->tag_set == &nullb->__tag_set)
1507                 blk_mq_free_tag_set(nullb->tag_set);
1508         put_disk(nullb->disk);
1509         cleanup_queues(nullb);
1510         if (null_cache_active(nullb))
1511                 null_free_device_storage(nullb->dev, true);
1512         kfree(nullb);
1513         dev->nullb = NULL;
1514 }
1515
1516 static void null_config_discard(struct nullb *nullb)
1517 {
1518         if (nullb->dev->discard == false)
1519                 return;
1520         nullb->q->limits.discard_granularity = nullb->dev->blocksize;
1521         nullb->q->limits.discard_alignment = nullb->dev->blocksize;
1522         blk_queue_max_discard_sectors(nullb->q, UINT_MAX >> 9);
1523         blk_queue_flag_set(QUEUE_FLAG_DISCARD, nullb->q);
1524 }
1525
1526 static int null_open(struct block_device *bdev, fmode_t mode)
1527 {
1528         return 0;
1529 }
1530
1531 static void null_release(struct gendisk *disk, fmode_t mode)
1532 {
1533 }
1534
1535 static const struct block_device_operations null_fops = {
1536         .owner =        THIS_MODULE,
1537         .open =         null_open,
1538         .release =      null_release,
1539 };
1540
1541 static void null_init_queue(struct nullb *nullb, struct nullb_queue *nq)
1542 {
1543         BUG_ON(!nullb);
1544         BUG_ON(!nq);
1545
1546         init_waitqueue_head(&nq->wait);
1547         nq->queue_depth = nullb->queue_depth;
1548         nq->dev = nullb->dev;
1549 }
1550
1551 static void null_init_queues(struct nullb *nullb)
1552 {
1553         struct request_queue *q = nullb->q;
1554         struct blk_mq_hw_ctx *hctx;
1555         struct nullb_queue *nq;
1556         int i;
1557
1558         queue_for_each_hw_ctx(q, hctx, i) {
1559                 if (!hctx->nr_ctx || !hctx->tags)
1560                         continue;
1561                 nq = &nullb->queues[i];
1562                 hctx->driver_data = nq;
1563                 null_init_queue(nullb, nq);
1564                 nullb->nr_queues++;
1565         }
1566 }
1567
1568 static int setup_commands(struct nullb_queue *nq)
1569 {
1570         struct nullb_cmd *cmd;
1571         int i, tag_size;
1572
1573         nq->cmds = kcalloc(nq->queue_depth, sizeof(*cmd), GFP_KERNEL);
1574         if (!nq->cmds)
1575                 return -ENOMEM;
1576
1577         tag_size = ALIGN(nq->queue_depth, BITS_PER_LONG) / BITS_PER_LONG;
1578         nq->tag_map = kcalloc(tag_size, sizeof(unsigned long), GFP_KERNEL);
1579         if (!nq->tag_map) {
1580                 kfree(nq->cmds);
1581                 return -ENOMEM;
1582         }
1583
1584         for (i = 0; i < nq->queue_depth; i++) {
1585                 cmd = &nq->cmds[i];
1586                 INIT_LIST_HEAD(&cmd->list);
1587                 cmd->ll_list.next = NULL;
1588                 cmd->tag = -1U;
1589         }
1590
1591         return 0;
1592 }
1593
1594 static int setup_queues(struct nullb *nullb)
1595 {
1596         nullb->queues = kcalloc(nullb->dev->submit_queues,
1597                                 sizeof(struct nullb_queue),
1598                                 GFP_KERNEL);
1599         if (!nullb->queues)
1600                 return -ENOMEM;
1601
1602         nullb->nr_queues = 0;
1603         nullb->queue_depth = nullb->dev->hw_queue_depth;
1604
1605         return 0;
1606 }
1607
1608 static int init_driver_queues(struct nullb *nullb)
1609 {
1610         struct nullb_queue *nq;
1611         int i, ret = 0;
1612
1613         for (i = 0; i < nullb->dev->submit_queues; i++) {
1614                 nq = &nullb->queues[i];
1615
1616                 null_init_queue(nullb, nq);
1617
1618                 ret = setup_commands(nq);
1619                 if (ret)
1620                         return ret;
1621                 nullb->nr_queues++;
1622         }
1623         return 0;
1624 }
1625
1626 static int null_gendisk_register(struct nullb *nullb)
1627 {
1628         struct gendisk *disk;
1629         sector_t size;
1630
1631         disk = nullb->disk = alloc_disk_node(1, nullb->dev->home_node);
1632         if (!disk)
1633                 return -ENOMEM;
1634         size = (sector_t)nullb->dev->size * 1024 * 1024ULL;
1635         set_capacity(disk, size >> 9);
1636
1637         disk->flags |= GENHD_FL_EXT_DEVT | GENHD_FL_SUPPRESS_PARTITION_INFO;
1638         disk->major             = null_major;
1639         disk->first_minor       = nullb->index;
1640         disk->fops              = &null_fops;
1641         disk->private_data      = nullb;
1642         disk->queue             = nullb->q;
1643         strncpy(disk->disk_name, nullb->disk_name, DISK_NAME_LEN);
1644
1645         add_disk(disk);
1646         return 0;
1647 }
1648
1649 static int null_init_tag_set(struct nullb *nullb, struct blk_mq_tag_set *set)
1650 {
1651         set->ops = &null_mq_ops;
1652         set->nr_hw_queues = nullb ? nullb->dev->submit_queues :
1653                                                 g_submit_queues;
1654         set->queue_depth = nullb ? nullb->dev->hw_queue_depth :
1655                                                 g_hw_queue_depth;
1656         set->numa_node = nullb ? nullb->dev->home_node : g_home_node;
1657         set->cmd_size   = sizeof(struct nullb_cmd);
1658         set->flags = BLK_MQ_F_SHOULD_MERGE;
1659         if (g_no_sched)
1660                 set->flags |= BLK_MQ_F_NO_SCHED;
1661         set->driver_data = NULL;
1662
1663         if ((nullb && nullb->dev->blocking) || g_blocking)
1664                 set->flags |= BLK_MQ_F_BLOCKING;
1665
1666         return blk_mq_alloc_tag_set(set);
1667 }
1668
1669 static void null_validate_conf(struct nullb_device *dev)
1670 {
1671         dev->blocksize = round_down(dev->blocksize, 512);
1672         dev->blocksize = clamp_t(unsigned int, dev->blocksize, 512, 4096);
1673
1674         if (dev->queue_mode == NULL_Q_MQ && dev->use_per_node_hctx) {
1675                 if (dev->submit_queues != nr_online_nodes)
1676                         dev->submit_queues = nr_online_nodes;
1677         } else if (dev->submit_queues > nr_cpu_ids)
1678                 dev->submit_queues = nr_cpu_ids;
1679         else if (dev->submit_queues == 0)
1680                 dev->submit_queues = 1;
1681
1682         dev->queue_mode = min_t(unsigned int, dev->queue_mode, NULL_Q_MQ);
1683         dev->irqmode = min_t(unsigned int, dev->irqmode, NULL_IRQ_TIMER);
1684
1685         /* Do memory allocation, so set blocking */
1686         if (dev->memory_backed)
1687                 dev->blocking = true;
1688         else /* cache is meaningless */
1689                 dev->cache_size = 0;
1690         dev->cache_size = min_t(unsigned long, ULONG_MAX / 1024 / 1024,
1691                                                 dev->cache_size);
1692         dev->mbps = min_t(unsigned int, 1024 * 40, dev->mbps);
1693         /* can not stop a queue */
1694         if (dev->queue_mode == NULL_Q_BIO)
1695                 dev->mbps = 0;
1696 }
1697
1698 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1699 static bool __null_setup_fault(struct fault_attr *attr, char *str)
1700 {
1701         if (!str[0])
1702                 return true;
1703
1704         if (!setup_fault_attr(attr, str))
1705                 return false;
1706
1707         attr->verbose = 0;
1708         return true;
1709 }
1710 #endif
1711
1712 static bool null_setup_fault(void)
1713 {
1714 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1715         if (!__null_setup_fault(&null_timeout_attr, g_timeout_str))
1716                 return false;
1717         if (!__null_setup_fault(&null_requeue_attr, g_requeue_str))
1718                 return false;
1719 #endif
1720         return true;
1721 }
1722
1723 static int null_add_dev(struct nullb_device *dev)
1724 {
1725         struct nullb *nullb;
1726         int rv;
1727
1728         null_validate_conf(dev);
1729
1730         nullb = kzalloc_node(sizeof(*nullb), GFP_KERNEL, dev->home_node);
1731         if (!nullb) {
1732                 rv = -ENOMEM;
1733                 goto out;
1734         }
1735         nullb->dev = dev;
1736         dev->nullb = nullb;
1737
1738         spin_lock_init(&nullb->lock);
1739
1740         rv = setup_queues(nullb);
1741         if (rv)
1742                 goto out_free_nullb;
1743
1744         if (dev->queue_mode == NULL_Q_MQ) {
1745                 if (shared_tags) {
1746                         nullb->tag_set = &tag_set;
1747                         rv = 0;
1748                 } else {
1749                         nullb->tag_set = &nullb->__tag_set;
1750                         rv = null_init_tag_set(nullb, nullb->tag_set);
1751                 }
1752
1753                 if (rv)
1754                         goto out_cleanup_queues;
1755
1756                 if (!null_setup_fault())
1757                         goto out_cleanup_queues;
1758
1759                 nullb->tag_set->timeout = 5 * HZ;
1760                 nullb->q = blk_mq_init_queue(nullb->tag_set);
1761                 if (IS_ERR(nullb->q)) {
1762                         rv = -ENOMEM;
1763                         goto out_cleanup_tags;
1764                 }
1765                 null_init_queues(nullb);
1766         } else if (dev->queue_mode == NULL_Q_BIO) {
1767                 nullb->q = blk_alloc_queue_node(GFP_KERNEL, dev->home_node,
1768                                                 NULL);
1769                 if (!nullb->q) {
1770                         rv = -ENOMEM;
1771                         goto out_cleanup_queues;
1772                 }
1773                 blk_queue_make_request(nullb->q, null_queue_bio);
1774                 rv = init_driver_queues(nullb);
1775                 if (rv)
1776                         goto out_cleanup_blk_queue;
1777         } else {
1778                 nullb->q = blk_init_queue_node(null_request_fn, &nullb->lock,
1779                                                 dev->home_node);
1780                 if (!nullb->q) {
1781                         rv = -ENOMEM;
1782                         goto out_cleanup_queues;
1783                 }
1784
1785                 if (!null_setup_fault())
1786                         goto out_cleanup_blk_queue;
1787
1788                 blk_queue_prep_rq(nullb->q, null_rq_prep_fn);
1789                 blk_queue_softirq_done(nullb->q, null_softirq_done_fn);
1790                 blk_queue_rq_timed_out(nullb->q, null_rq_timed_out_fn);
1791                 nullb->q->rq_timeout = 5 * HZ;
1792                 rv = init_driver_queues(nullb);
1793                 if (rv)
1794                         goto out_cleanup_blk_queue;
1795         }
1796
1797         if (dev->mbps) {
1798                 set_bit(NULLB_DEV_FL_THROTTLED, &dev->flags);
1799                 nullb_setup_bwtimer(nullb);
1800         }
1801
1802         if (dev->cache_size > 0) {
1803                 set_bit(NULLB_DEV_FL_CACHE, &nullb->dev->flags);
1804                 blk_queue_write_cache(nullb->q, true, true);
1805                 blk_queue_flush_queueable(nullb->q, true);
1806         }
1807
1808         if (dev->zoned) {
1809                 rv = null_zone_init(dev);
1810                 if (rv)
1811                         goto out_cleanup_blk_queue;
1812
1813                 blk_queue_chunk_sectors(nullb->q, dev->zone_size_sects);
1814                 nullb->q->limits.zoned = BLK_ZONED_HM;
1815         }
1816
1817         nullb->q->queuedata = nullb;
1818         blk_queue_flag_set(QUEUE_FLAG_NONROT, nullb->q);
1819         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, nullb->q);
1820
1821         mutex_lock(&lock);
1822         rv = ida_simple_get(&nullb_indexes, 0, 0, GFP_KERNEL);
1823         if (rv < 0) {
1824                 mutex_unlock(&lock);
1825                 goto out_cleanup_zone;
1826         }
1827         nullb->index = rv;
1828         dev->index = rv;
1829         mutex_unlock(&lock);
1830
1831         blk_queue_logical_block_size(nullb->q, dev->blocksize);
1832         blk_queue_physical_block_size(nullb->q, dev->blocksize);
1833
1834         null_config_discard(nullb);
1835
1836         sprintf(nullb->disk_name, "nullb%d", nullb->index);
1837
1838         rv = null_gendisk_register(nullb);
1839         if (rv)
1840                 goto out_ida_free;
1841
1842         mutex_lock(&lock);
1843         list_add_tail(&nullb->list, &nullb_list);
1844         mutex_unlock(&lock);
1845
1846         return 0;
1847
1848 out_ida_free:
1849         ida_free(&nullb_indexes, nullb->index);
1850 out_cleanup_zone:
1851         if (dev->zoned)
1852                 null_zone_exit(dev);
1853 out_cleanup_blk_queue:
1854         blk_cleanup_queue(nullb->q);
1855 out_cleanup_tags:
1856         if (dev->queue_mode == NULL_Q_MQ && nullb->tag_set == &nullb->__tag_set)
1857                 blk_mq_free_tag_set(nullb->tag_set);
1858 out_cleanup_queues:
1859         cleanup_queues(nullb);
1860 out_free_nullb:
1861         kfree(nullb);
1862         dev->nullb = NULL;
1863 out:
1864         return rv;
1865 }
1866
1867 static int __init null_init(void)
1868 {
1869         int ret = 0;
1870         unsigned int i;
1871         struct nullb *nullb;
1872         struct nullb_device *dev;
1873
1874         if (g_bs > PAGE_SIZE) {
1875                 pr_warn("null_blk: invalid block size\n");
1876                 pr_warn("null_blk: defaults block size to %lu\n", PAGE_SIZE);
1877                 g_bs = PAGE_SIZE;
1878         }
1879
1880         if (!is_power_of_2(g_zone_size)) {
1881                 pr_err("null_blk: zone_size must be power-of-two\n");
1882                 return -EINVAL;
1883         }
1884
1885         if (g_queue_mode == NULL_Q_MQ && g_use_per_node_hctx) {
1886                 if (g_submit_queues != nr_online_nodes) {
1887                         pr_warn("null_blk: submit_queues param is set to %u.\n",
1888                                                         nr_online_nodes);
1889                         g_submit_queues = nr_online_nodes;
1890                 }
1891         } else if (g_submit_queues > nr_cpu_ids)
1892                 g_submit_queues = nr_cpu_ids;
1893         else if (g_submit_queues <= 0)
1894                 g_submit_queues = 1;
1895
1896         if (g_queue_mode == NULL_Q_MQ && shared_tags) {
1897                 ret = null_init_tag_set(NULL, &tag_set);
1898                 if (ret)
1899                         return ret;
1900         }
1901
1902         config_group_init(&nullb_subsys.su_group);
1903         mutex_init(&nullb_subsys.su_mutex);
1904
1905         ret = configfs_register_subsystem(&nullb_subsys);
1906         if (ret)
1907                 goto err_tagset;
1908
1909         mutex_init(&lock);
1910
1911         null_major = register_blkdev(0, "nullb");
1912         if (null_major < 0) {
1913                 ret = null_major;
1914                 goto err_conf;
1915         }
1916
1917         for (i = 0; i < nr_devices; i++) {
1918                 dev = null_alloc_dev();
1919                 if (!dev) {
1920                         ret = -ENOMEM;
1921                         goto err_dev;
1922                 }
1923                 ret = null_add_dev(dev);
1924                 if (ret) {
1925                         null_free_dev(dev);
1926                         goto err_dev;
1927                 }
1928         }
1929
1930         pr_info("null: module loaded\n");
1931         return 0;
1932
1933 err_dev:
1934         while (!list_empty(&nullb_list)) {
1935                 nullb = list_entry(nullb_list.next, struct nullb, list);
1936                 dev = nullb->dev;
1937                 null_del_dev(nullb);
1938                 null_free_dev(dev);
1939         }
1940         unregister_blkdev(null_major, "nullb");
1941 err_conf:
1942         configfs_unregister_subsystem(&nullb_subsys);
1943 err_tagset:
1944         if (g_queue_mode == NULL_Q_MQ && shared_tags)
1945                 blk_mq_free_tag_set(&tag_set);
1946         return ret;
1947 }
1948
1949 static void __exit null_exit(void)
1950 {
1951         struct nullb *nullb;
1952
1953         configfs_unregister_subsystem(&nullb_subsys);
1954
1955         unregister_blkdev(null_major, "nullb");
1956
1957         mutex_lock(&lock);
1958         while (!list_empty(&nullb_list)) {
1959                 struct nullb_device *dev;
1960
1961                 nullb = list_entry(nullb_list.next, struct nullb, list);
1962                 dev = nullb->dev;
1963                 null_del_dev(nullb);
1964                 null_free_dev(dev);
1965         }
1966         mutex_unlock(&lock);
1967
1968         if (g_queue_mode == NULL_Q_MQ && shared_tags)
1969                 blk_mq_free_tag_set(&tag_set);
1970 }
1971
1972 module_init(null_init);
1973 module_exit(null_exit);
1974
1975 MODULE_AUTHOR("Jens Axboe <axboe@kernel.dk>");
1976 MODULE_LICENSE("GPL");