drivers/block/null_blk_main.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Add configfs and memory store: Kyungchan Koh <kkc6196@fb.com> and
   4  * Shaohua Li <shli@fb.com>
   5  */
   6 #include <linux/module.h>
   7
   8 #include <linux/moduleparam.h>
   9 #include <linux/sched.h>
  10 #include <linux/fs.h>
  11 #include <linux/init.h>
  12 #include "null_blk.h"
  13
  14 #define PAGE_SECTORS_SHIFT      (PAGE_SHIFT - SECTOR_SHIFT)
  15 #define PAGE_SECTORS            (1 << PAGE_SECTORS_SHIFT)
  16 #define SECTOR_MASK             (PAGE_SECTORS - 1)
  17
  18 #define FREE_BATCH              16
  19
  20 #define TICKS_PER_SEC           50ULL
  21 #define TIMER_INTERVAL          (NSEC_PER_SEC / TICKS_PER_SEC)
  22
  23 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
  24 static DECLARE_FAULT_ATTR(null_timeout_attr);
  25 static DECLARE_FAULT_ATTR(null_requeue_attr);
  26 static DECLARE_FAULT_ATTR(null_init_hctx_attr);
  27 #endif
  28
  29 static inline u64 mb_per_tick(int mbps)
  30 {
  31         return (1 << 20) / TICKS_PER_SEC * ((u64) mbps);
  32 }
  33
  34 /*
  35  * Status flags for nullb_device.
  36  *
  37  * CONFIGURED:  Device has been configured and turned on. Cannot reconfigure.
  38  * UP:          Device is currently on and visible in userspace.
  39  * THROTTLED:   Device is being throttled.
  40  * CACHE:       Device is using a write-back cache.
  41  */
  42 enum nullb_device_flags {
  43         NULLB_DEV_FL_CONFIGURED = 0,
  44         NULLB_DEV_FL_UP         = 1,
  45         NULLB_DEV_FL_THROTTLED  = 2,
  46         NULLB_DEV_FL_CACHE      = 3,
  47 };
  48
  49 #define MAP_SZ          ((PAGE_SIZE >> SECTOR_SHIFT) + 2)
  50 /*
  51  * nullb_page is a page in memory for nullb devices.
  52  *
  53  * @page:       The page holding the data.
  54  * @bitmap:     The bitmap represents which sector in the page has data.
  55  *              Each bit represents one block size. For example, sector 8
  56  *              will use the 7th bit
  57  * The highest 2 bits of bitmap are for special purpose. LOCK means the cache
  58  * page is being flushing to storage. FREE means the cache page is freed and
  59  * should be skipped from flushing to storage. Please see
  60  * null_make_cache_space
  61  */
  62 struct nullb_page {
  63         struct page *page;
  64         DECLARE_BITMAP(bitmap, MAP_SZ);
  65 };
  66 #define NULLB_PAGE_LOCK (MAP_SZ - 1)
  67 #define NULLB_PAGE_FREE (MAP_SZ - 2)
  68
  69 static LIST_HEAD(nullb_list);
  70 static struct mutex lock;
  71 static int null_major;
  72 static DEFINE_IDA(nullb_indexes);
  73 static struct blk_mq_tag_set tag_set;
  74
  75 enum {
  76         NULL_IRQ_NONE           = 0,
  77         NULL_IRQ_SOFTIRQ        = 1,
  78         NULL_IRQ_TIMER          = 2,
  79 };
  80
  81 enum {
  82         NULL_Q_BIO              = 0,
  83         NULL_Q_RQ               = 1,
  84         NULL_Q_MQ               = 2,
  85 };
  86
  87 static int g_no_sched;
  88 module_param_named(no_sched, g_no_sched, int, 0444);
  89 MODULE_PARM_DESC(no_sched, "No io scheduler");
  90
  91 static int g_submit_queues = 1;
  92 module_param_named(submit_queues, g_submit_queues, int, 0444);
  93 MODULE_PARM_DESC(submit_queues, "Number of submission queues");
  94
  95 static int g_home_node = NUMA_NO_NODE;
  96 module_param_named(home_node, g_home_node, int, 0444);
  97 MODULE_PARM_DESC(home_node, "Home node for the device");
  98
  99 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
 100 static char g_timeout_str[80];
 101 module_param_string(timeout, g_timeout_str, sizeof(g_timeout_str), 0444);
 102
 103 static char g_requeue_str[80];
 104 module_param_string(requeue, g_requeue_str, sizeof(g_requeue_str), 0444);
 105
 106 static char g_init_hctx_str[80];
 107 module_param_string(init_hctx, g_init_hctx_str, sizeof(g_init_hctx_str), 0444);
 108 #endif
 109
 110 static int g_queue_mode = NULL_Q_MQ;
 111
 112 static int null_param_store_val(const char *str, int *val, int min, int max)
 113 {
 114         int ret, new_val;
 115
 116         ret = kstrtoint(str, 10, &new_val);
 117         if (ret)
 118                 return -EINVAL;
 119
 120         if (new_val < min || new_val > max)
 121                 return -EINVAL;
 122
 123         *val = new_val;
 124         return 0;
 125 }
 126
 127 static int null_set_queue_mode(const char *str, const struct kernel_param *kp)
 128 {
 129         return null_param_store_val(str, &g_queue_mode, NULL_Q_BIO, NULL_Q_MQ);
 130 }
 131
 132 static const struct kernel_param_ops null_queue_mode_param_ops = {
 133         .set    = null_set_queue_mode,
 134         .get    = param_get_int,
 135 };
 136
 137 device_param_cb(queue_mode, &null_queue_mode_param_ops, &g_queue_mode, 0444);
 138 MODULE_PARM_DESC(queue_mode, "Block interface to use (0=bio,1=rq,2=multiqueue)");
 139
 140 static int g_gb = 250;
 141 module_param_named(gb, g_gb, int, 0444);
 142 MODULE_PARM_DESC(gb, "Size in GB");
 143
 144 static int g_bs = 512;
 145 module_param_named(bs, g_bs, int, 0444);
 146 MODULE_PARM_DESC(bs, "Block size (in bytes)");
 147
 148 static unsigned int nr_devices = 1;
 149 module_param(nr_devices, uint, 0444);
 150 MODULE_PARM_DESC(nr_devices, "Number of devices to register");
 151
 152 static bool g_blocking;
 153 module_param_named(blocking, g_blocking, bool, 0444);
 154 MODULE_PARM_DESC(blocking, "Register as a blocking blk-mq driver device");
 155
 156 static bool shared_tags;
 157 module_param(shared_tags, bool, 0444);
 158 MODULE_PARM_DESC(shared_tags, "Share tag set between devices for blk-mq");
 159
 160 static int g_irqmode = NULL_IRQ_SOFTIRQ;
 161
 162 static int null_set_irqmode(const char *str, const struct kernel_param *kp)
 163 {
 164         return null_param_store_val(str, &g_irqmode, NULL_IRQ_NONE,
 165                                         NULL_IRQ_TIMER);
 166 }
 167
 168 static const struct kernel_param_ops null_irqmode_param_ops = {
 169         .set    = null_set_irqmode,
 170         .get    = param_get_int,
 171 };
 172
 173 device_param_cb(irqmode, &null_irqmode_param_ops, &g_irqmode, 0444);
 174 MODULE_PARM_DESC(irqmode, "IRQ completion handler. 0-none, 1-softirq, 2-timer");
 175
 176 static unsigned long g_completion_nsec = 10000;
 177 module_param_named(completion_nsec, g_completion_nsec, ulong, 0444);
 178 MODULE_PARM_DESC(completion_nsec, "Time in ns to complete a request in hardware. Default: 10,000ns");
 179
 180 static int g_hw_queue_depth = 64;
 181 module_param_named(hw_queue_depth, g_hw_queue_depth, int, 0444);
 182 MODULE_PARM_DESC(hw_queue_depth, "Queue depth for each hardware queue. Default: 64");
 183
 184 static bool g_use_per_node_hctx;
 185 module_param_named(use_per_node_hctx, g_use_per_node_hctx, bool, 0444);
 186 MODULE_PARM_DESC(use_per_node_hctx, "Use per-node allocation for hardware context queues. Default: false");
 187
 188 static bool g_zoned;
 189 module_param_named(zoned, g_zoned, bool, S_IRUGO);
 190 MODULE_PARM_DESC(zoned, "Make device as a host-managed zoned block device. Default: false");
 191
 192 static unsigned long g_zone_size = 256;
 193 module_param_named(zone_size, g_zone_size, ulong, S_IRUGO);
 194 MODULE_PARM_DESC(zone_size, "Zone size in MB when block device is zoned. Must be power-of-two: Default: 256");
 195
 196 static unsigned int g_zone_nr_conv;
 197 module_param_named(zone_nr_conv, g_zone_nr_conv, uint, 0444);
 198 MODULE_PARM_DESC(zone_nr_conv, "Number of conventional zones when block device is zoned. Default: 0");
 199
 200 static struct nullb_device *null_alloc_dev(void);
 201 static void null_free_dev(struct nullb_device *dev);
 202 static void null_del_dev(struct nullb *nullb);
 203 static int null_add_dev(struct nullb_device *dev);
 204 static void null_free_device_storage(struct nullb_device *dev, bool is_cache);
 205
 206 static inline struct nullb_device *to_nullb_device(struct config_item *item)
 207 {
 208         return item ? container_of(item, struct nullb_device, item) : NULL;
 209 }
 210
 211 static inline ssize_t nullb_device_uint_attr_show(unsigned int val, char *page)
 212 {
 213         return snprintf(page, PAGE_SIZE, "%u\n", val);
 214 }
 215
 216 static inline ssize_t nullb_device_ulong_attr_show(unsigned long val,
 217         char *page)
 218 {
 219         return snprintf(page, PAGE_SIZE, "%lu\n", val);
 220 }
 221
 222 static inline ssize_t nullb_device_bool_attr_show(bool val, char *page)
 223 {
 224         return snprintf(page, PAGE_SIZE, "%u\n", val);
 225 }
 226
 227 static ssize_t nullb_device_uint_attr_store(unsigned int *val,
 228         const char *page, size_t count)
 229 {
 230         unsigned int tmp;
 231         int result;
 232
 233         result = kstrtouint(page, 0, &tmp);
 234         if (result < 0)
 235                 return result;
 236
 237         *val = tmp;
 238         return count;
 239 }
 240
 241 static ssize_t nullb_device_ulong_attr_store(unsigned long *val,
 242         const char *page, size_t count)
 243 {
 244         int result;
 245         unsigned long tmp;
 246
 247         result = kstrtoul(page, 0, &tmp);
 248         if (result < 0)
 249                 return result;
 250
 251         *val = tmp;
 252         return count;
 253 }
 254
 255 static ssize_t nullb_device_bool_attr_store(bool *val, const char *page,
 256         size_t count)
 257 {
 258         bool tmp;
 259         int result;
 260
 261         result = kstrtobool(page,  &tmp);
 262         if (result < 0)
 263                 return result;
 264
 265         *val = tmp;
 266         return count;
 267 }
 268
 269 /* The following macro should only be used with TYPE = {uint, ulong, bool}. */
 270 #define NULLB_DEVICE_ATTR(NAME, TYPE, APPLY)                            \
 271 static ssize_t                                                          \
 272 nullb_device_##NAME##_show(struct config_item *item, char *page)        \
 273 {                                                                       \
 274         return nullb_device_##TYPE##_attr_show(                         \
 275                                 to_nullb_device(item)->NAME, page);     \
 276 }                                                                       \
 277 static ssize_t                                                          \
 278 nullb_device_##NAME##_store(struct config_item *item, const char *page, \
 279                             size_t count)                               \
 280 {                                                                       \
 281         int (*apply_fn)(struct nullb_device *dev, TYPE new_value) = APPLY;\
 282         struct nullb_device *dev = to_nullb_device(item);               \
 283         TYPE new_value = 0;                                             \
 284         int ret;                                                        \
 285                                                                         \
 286         ret = nullb_device_##TYPE##_attr_store(&new_value, page, count);\
 287         if (ret < 0)                                                    \
 288                 return ret;                                             \
 289         if (apply_fn)                                                   \
 290                 ret = apply_fn(dev, new_value);                         \
 291         else if (test_bit(NULLB_DEV_FL_CONFIGURED, &dev->flags))        \
 292                 ret = -EBUSY;                                           \
 293         if (ret < 0)                                                    \
 294                 return ret;                                             \
 295         dev->NAME = new_value;                                          \
 296         return count;                                                   \
 297 }                                                                       \
 298 CONFIGFS_ATTR(nullb_device_, NAME);
 299
 300 static int nullb_apply_submit_queues(struct nullb_device *dev,
 301                                      unsigned int submit_queues)
 302 {
 303         struct nullb *nullb = dev->nullb;
 304         struct blk_mq_tag_set *set;
 305
 306         if (!nullb)
 307                 return 0;
 308
 309         /*
 310          * Make sure that null_init_hctx() does not access nullb->queues[] past
 311          * the end of that array.
 312          */
 313         if (submit_queues > nr_cpu_ids)
 314                 return -EINVAL;
 315         set = nullb->tag_set;
 316         blk_mq_update_nr_hw_queues(set, submit_queues);
 317         return set->nr_hw_queues == submit_queues ? 0 : -ENOMEM;
 318 }
 319
 320 NULLB_DEVICE_ATTR(size, ulong, NULL);
 321 NULLB_DEVICE_ATTR(completion_nsec, ulong, NULL);
 322 NULLB_DEVICE_ATTR(submit_queues, uint, nullb_apply_submit_queues);
 323 NULLB_DEVICE_ATTR(home_node, uint, NULL);
 324 NULLB_DEVICE_ATTR(queue_mode, uint, NULL);
 325 NULLB_DEVICE_ATTR(blocksize, uint, NULL);
 326 NULLB_DEVICE_ATTR(irqmode, uint, NULL);
 327 NULLB_DEVICE_ATTR(hw_queue_depth, uint, NULL);
 328 NULLB_DEVICE_ATTR(index, uint, NULL);
 329 NULLB_DEVICE_ATTR(blocking, bool, NULL);
 330 NULLB_DEVICE_ATTR(use_per_node_hctx, bool, NULL);
 331 NULLB_DEVICE_ATTR(memory_backed, bool, NULL);
 332 NULLB_DEVICE_ATTR(discard, bool, NULL);
 333 NULLB_DEVICE_ATTR(mbps, uint, NULL);
 334 NULLB_DEVICE_ATTR(cache_size, ulong, NULL);
 335 NULLB_DEVICE_ATTR(zoned, bool, NULL);
 336 NULLB_DEVICE_ATTR(zone_size, ulong, NULL);
 337 NULLB_DEVICE_ATTR(zone_nr_conv, uint, NULL);
 338
 339 static ssize_t nullb_device_power_show(struct config_item *item, char *page)
 340 {
 341         return nullb_device_bool_attr_show(to_nullb_device(item)->power, page);
 342 }
 343
 344 static ssize_t nullb_device_power_store(struct config_item *item,
 345                                      const char *page, size_t count)
 346 {
 347         struct nullb_device *dev = to_nullb_device(item);
 348         bool newp = false;
 349         ssize_t ret;
 350
 351         ret = nullb_device_bool_attr_store(&newp, page, count);
 352         if (ret < 0)
 353                 return ret;
 354
 355         if (!dev->power && newp) {
 356                 if (test_and_set_bit(NULLB_DEV_FL_UP, &dev->flags))
 357                         return count;
 358                 if (null_add_dev(dev)) {
 359                         clear_bit(NULLB_DEV_FL_UP, &dev->flags);
 360                         return -ENOMEM;
 361                 }
 362
 363                 set_bit(NULLB_DEV_FL_CONFIGURED, &dev->flags);
 364                 dev->power = newp;
 365         } else if (dev->power && !newp) {
 366                 if (test_and_clear_bit(NULLB_DEV_FL_UP, &dev->flags)) {
 367                         mutex_lock(&lock);
 368                         dev->power = newp;
 369                         null_del_dev(dev->nullb);
 370                         mutex_unlock(&lock);
 371                 }
 372                 clear_bit(NULLB_DEV_FL_CONFIGURED, &dev->flags);
 373         }
 374
 375         return count;
 376 }
 377
 378 CONFIGFS_ATTR(nullb_device_, power);
 379
 380 static ssize_t nullb_device_badblocks_show(struct config_item *item, char *page)
 381 {
 382         struct nullb_device *t_dev = to_nullb_device(item);
 383
 384         return badblocks_show(&t_dev->badblocks, page, 0);
 385 }
 386
 387 static ssize_t nullb_device_badblocks_store(struct config_item *item,
 388                                      const char *page, size_t count)
 389 {
 390         struct nullb_device *t_dev = to_nullb_device(item);
 391         char *orig, *buf, *tmp;
 392         u64 start, end;
 393         int ret;
 394
 395         orig = kstrndup(page, count, GFP_KERNEL);
 396         if (!orig)
 397                 return -ENOMEM;
 398
 399         buf = strstrip(orig);
 400
 401         ret = -EINVAL;
 402         if (buf[0] != '+' && buf[0] != '-')
 403                 goto out;
 404         tmp = strchr(&buf[1], '-');
 405         if (!tmp)
 406                 goto out;
 407         *tmp = '\0';
 408         ret = kstrtoull(buf + 1, 0, &start);
 409         if (ret)
 410                 goto out;
 411         ret = kstrtoull(tmp + 1, 0, &end);
 412         if (ret)
 413                 goto out;
 414         ret = -EINVAL;
 415         if (start > end)
 416                 goto out;
 417         /* enable badblocks */
 418         cmpxchg(&t_dev->badblocks.shift, -1, 0);
 419         if (buf[0] == '+')
 420                 ret = badblocks_set(&t_dev->badblocks, start,
 421                         end - start + 1, 1);
 422         else
 423                 ret = badblocks_clear(&t_dev->badblocks, start,
 424                         end - start + 1);
 425         if (ret == 0)
 426                 ret = count;
 427 out:
 428         kfree(orig);
 429         return ret;
 430 }
 431 CONFIGFS_ATTR(nullb_device_, badblocks);
 432
 433 static struct configfs_attribute *nullb_device_attrs[] = {
 434         &nullb_device_attr_size,
 435         &nullb_device_attr_completion_nsec,
 436         &nullb_device_attr_submit_queues,
 437         &nullb_device_attr_home_node,
 438         &nullb_device_attr_queue_mode,
 439         &nullb_device_attr_blocksize,
 440         &nullb_device_attr_irqmode,
 441         &nullb_device_attr_hw_queue_depth,
 442         &nullb_device_attr_index,
 443         &nullb_device_attr_blocking,
 444         &nullb_device_attr_use_per_node_hctx,
 445         &nullb_device_attr_power,
 446         &nullb_device_attr_memory_backed,
 447         &nullb_device_attr_discard,
 448         &nullb_device_attr_mbps,
 449         &nullb_device_attr_cache_size,
 450         &nullb_device_attr_badblocks,
 451         &nullb_device_attr_zoned,
 452         &nullb_device_attr_zone_size,
 453         &nullb_device_attr_zone_nr_conv,
 454         NULL,
 455 };
 456
 457 static void nullb_device_release(struct config_item *item)
 458 {
 459         struct nullb_device *dev = to_nullb_device(item);
 460
 461         null_free_device_storage(dev, false);
 462         null_free_dev(dev);
 463 }
 464
 465 static struct configfs_item_operations nullb_device_ops = {
 466         .release        = nullb_device_release,
 467 };
 468
 469 static const struct config_item_type nullb_device_type = {
 470         .ct_item_ops    = &nullb_device_ops,
 471         .ct_attrs       = nullb_device_attrs,
 472         .ct_owner       = THIS_MODULE,
 473 };
 474
 475 static struct
 476 config_item *nullb_group_make_item(struct config_group *group, const char *name)
 477 {
 478         struct nullb_device *dev;
 479
 480         dev = null_alloc_dev();
 481         if (!dev)
 482                 return ERR_PTR(-ENOMEM);
 483
 484         config_item_init_type_name(&dev->item, name, &nullb_device_type);
 485
 486         return &dev->item;
 487 }
 488
 489 static void
 490 nullb_group_drop_item(struct config_group *group, struct config_item *item)
 491 {
 492         struct nullb_device *dev = to_nullb_device(item);
 493
 494         if (test_and_clear_bit(NULLB_DEV_FL_UP, &dev->flags)) {
 495                 mutex_lock(&lock);
 496                 dev->power = false;
 497                 null_del_dev(dev->nullb);
 498                 mutex_unlock(&lock);
 499         }
 500
 501         config_item_put(item);
 502 }
 503
 504 static ssize_t memb_group_features_show(struct config_item *item, char *page)
 505 {
 506         return snprintf(page, PAGE_SIZE, "memory_backed,discard,bandwidth,cache,badblocks,zoned,zone_size,zone_nr_conv\n");
 507 }
 508
 509 CONFIGFS_ATTR_RO(memb_group_, features);
 510
 511 static struct configfs_attribute *nullb_group_attrs[] = {
 512         &memb_group_attr_features,
 513         NULL,
 514 };
 515
 516 static struct configfs_group_operations nullb_group_ops = {
 517         .make_item      = nullb_group_make_item,
 518         .drop_item      = nullb_group_drop_item,
 519 };
 520
 521 static const struct config_item_type nullb_group_type = {
 522         .ct_group_ops   = &nullb_group_ops,
 523         .ct_attrs       = nullb_group_attrs,
 524         .ct_owner       = THIS_MODULE,
 525 };
 526
 527 static struct configfs_subsystem nullb_subsys = {
 528         .su_group = {
 529                 .cg_item = {
 530                         .ci_namebuf = "nullb",
 531                         .ci_type = &nullb_group_type,
 532                 },
 533         },
 534 };
 535
 536 static inline int null_cache_active(struct nullb *nullb)
 537 {
 538         return test_bit(NULLB_DEV_FL_CACHE, &nullb->dev->flags);
 539 }
 540
 541 static struct nullb_device *null_alloc_dev(void)
 542 {
 543         struct nullb_device *dev;
 544
 545         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
 546         if (!dev)
 547                 return NULL;
 548         INIT_RADIX_TREE(&dev->data, GFP_ATOMIC);
 549         INIT_RADIX_TREE(&dev->cache, GFP_ATOMIC);
 550         if (badblocks_init(&dev->badblocks, 0)) {
 551                 kfree(dev);
 552                 return NULL;
 553         }
 554
 555         dev->size = g_gb * 1024;
 556         dev->completion_nsec = g_completion_nsec;
 557         dev->submit_queues = g_submit_queues;
 558         dev->home_node = g_home_node;
 559         dev->queue_mode = g_queue_mode;
 560         dev->blocksize = g_bs;
 561         dev->irqmode = g_irqmode;
 562         dev->hw_queue_depth = g_hw_queue_depth;
 563         dev->blocking = g_blocking;
 564         dev->use_per_node_hctx = g_use_per_node_hctx;
 565         dev->zoned = g_zoned;
 566         dev->zone_size = g_zone_size;
 567         dev->zone_nr_conv = g_zone_nr_conv;
 568         return dev;
 569 }
 570
 571 static void null_free_dev(struct nullb_device *dev)
 572 {
 573         if (!dev)
 574                 return;
 575
 576         null_zone_exit(dev);
 577         badblocks_exit(&dev->badblocks);
 578         kfree(dev);
 579 }
 580
 581 static void put_tag(struct nullb_queue *nq, unsigned int tag)
 582 {
 583         clear_bit_unlock(tag, nq->tag_map);
 584
 585         if (waitqueue_active(&nq->wait))
 586                 wake_up(&nq->wait);
 587 }
 588
 589 static unsigned int get_tag(struct nullb_queue *nq)
 590 {
 591         unsigned int tag;
 592
 593         do {
 594                 tag = find_first_zero_bit(nq->tag_map, nq->queue_depth);
 595                 if (tag >= nq->queue_depth)
 596                         return -1U;
 597         } while (test_and_set_bit_lock(tag, nq->tag_map));
 598
 599         return tag;
 600 }
 601
 602 static void free_cmd(struct nullb_cmd *cmd)
 603 {
 604         put_tag(cmd->nq, cmd->tag);
 605 }
 606
 607 static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer);
 608
 609 static struct nullb_cmd *__alloc_cmd(struct nullb_queue *nq)
 610 {
 611         struct nullb_cmd *cmd;
 612         unsigned int tag;
 613
 614         tag = get_tag(nq);
 615         if (tag != -1U) {
 616                 cmd = &nq->cmds[tag];
 617                 cmd->tag = tag;
 618                 cmd->nq = nq;
 619                 if (nq->dev->irqmode == NULL_IRQ_TIMER) {
 620                         hrtimer_init(&cmd->timer, CLOCK_MONOTONIC,
 621                                      HRTIMER_MODE_REL);
 622                         cmd->timer.function = null_cmd_timer_expired;
 623                 }
 624                 return cmd;
 625         }
 626
 627         return NULL;
 628 }
 629
 630 static struct nullb_cmd *alloc_cmd(struct nullb_queue *nq, int can_wait)
 631 {
 632         struct nullb_cmd *cmd;
 633         DEFINE_WAIT(wait);
 634
 635         cmd = __alloc_cmd(nq);
 636         if (cmd || !can_wait)
 637                 return cmd;
 638
 639         do {
 640                 prepare_to_wait(&nq->wait, &wait, TASK_UNINTERRUPTIBLE);
 641                 cmd = __alloc_cmd(nq);
 642                 if (cmd)
 643                         break;
 644
 645                 io_schedule();
 646         } while (1);
 647
 648         finish_wait(&nq->wait, &wait);
 649         return cmd;
 650 }
 651
 652 static void end_cmd(struct nullb_cmd *cmd)
 653 {
 654         int queue_mode = cmd->nq->dev->queue_mode;
 655
 656         switch (queue_mode)  {
 657         case NULL_Q_MQ:
 658                 blk_mq_end_request(cmd->rq, cmd->error);
 659                 return;
 660         case NULL_Q_BIO:
 661                 cmd->bio->bi_status = cmd->error;
 662                 bio_endio(cmd->bio);
 663                 break;
 664         }
 665
 666         free_cmd(cmd);
 667 }
 668
 669 static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer)
 670 {
 671         end_cmd(container_of(timer, struct nullb_cmd, timer));
 672
 673         return HRTIMER_NORESTART;
 674 }
 675
 676 static void null_cmd_end_timer(struct nullb_cmd *cmd)
 677 {
 678         ktime_t kt = cmd->nq->dev->completion_nsec;
 679
 680         hrtimer_start(&cmd->timer, kt, HRTIMER_MODE_REL);
 681 }
 682
 683 static void null_complete_rq(struct request *rq)
 684 {
 685         end_cmd(blk_mq_rq_to_pdu(rq));
 686 }
 687
 688 static struct nullb_page *null_alloc_page(gfp_t gfp_flags)
 689 {
 690         struct nullb_page *t_page;
 691
 692         t_page = kmalloc(sizeof(struct nullb_page), gfp_flags);
 693         if (!t_page)
 694                 goto out;
 695
 696         t_page->page = alloc_pages(gfp_flags, 0);
 697         if (!t_page->page)
 698                 goto out_freepage;
 699
 700         memset(t_page->bitmap, 0, sizeof(t_page->bitmap));
 701         return t_page;
 702 out_freepage:
 703         kfree(t_page);
 704 out:
 705         return NULL;
 706 }
 707
 708 static void null_free_page(struct nullb_page *t_page)
 709 {
 710         __set_bit(NULLB_PAGE_FREE, t_page->bitmap);
 711         if (test_bit(NULLB_PAGE_LOCK, t_page->bitmap))
 712                 return;
 713         __free_page(t_page->page);
 714         kfree(t_page);
 715 }
 716
 717 static bool null_page_empty(struct nullb_page *page)
 718 {
 719         int size = MAP_SZ - 2;
 720
 721         return find_first_bit(page->bitmap, size) == size;
 722 }
 723
 724 static void null_free_sector(struct nullb *nullb, sector_t sector,
 725         bool is_cache)
 726 {
 727         unsigned int sector_bit;
 728         u64 idx;
 729         struct nullb_page *t_page, *ret;
 730         struct radix_tree_root *root;
 731
 732         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 733         idx = sector >> PAGE_SECTORS_SHIFT;
 734         sector_bit = (sector & SECTOR_MASK);
 735
 736         t_page = radix_tree_lookup(root, idx);
 737         if (t_page) {
 738                 __clear_bit(sector_bit, t_page->bitmap);
 739
 740                 if (null_page_empty(t_page)) {
 741                         ret = radix_tree_delete_item(root, idx, t_page);
 742                         WARN_ON(ret != t_page);
 743                         null_free_page(ret);
 744                         if (is_cache)
 745                                 nullb->dev->curr_cache -= PAGE_SIZE;
 746                 }
 747         }
 748 }
 749
 750 static struct nullb_page *null_radix_tree_insert(struct nullb *nullb, u64 idx,
 751         struct nullb_page *t_page, bool is_cache)
 752 {
 753         struct radix_tree_root *root;
 754
 755         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 756
 757         if (radix_tree_insert(root, idx, t_page)) {
 758                 null_free_page(t_page);
 759                 t_page = radix_tree_lookup(root, idx);
 760                 WARN_ON(!t_page || t_page->page->index != idx);
 761         } else if (is_cache)
 762                 nullb->dev->curr_cache += PAGE_SIZE;
 763
 764         return t_page;
 765 }
 766
 767 static void null_free_device_storage(struct nullb_device *dev, bool is_cache)
 768 {
 769         unsigned long pos = 0;
 770         int nr_pages;
 771         struct nullb_page *ret, *t_pages[FREE_BATCH];
 772         struct radix_tree_root *root;
 773
 774         root = is_cache ? &dev->cache : &dev->data;
 775
 776         do {
 777                 int i;
 778
 779                 nr_pages = radix_tree_gang_lookup(root,
 780                                 (void **)t_pages, pos, FREE_BATCH);
 781
 782                 for (i = 0; i < nr_pages; i++) {
 783                         pos = t_pages[i]->page->index;
 784                         ret = radix_tree_delete_item(root, pos, t_pages[i]);
 785                         WARN_ON(ret != t_pages[i]);
 786                         null_free_page(ret);
 787                 }
 788
 789                 pos++;
 790         } while (nr_pages == FREE_BATCH);
 791
 792         if (is_cache)
 793                 dev->curr_cache = 0;
 794 }
 795
 796 static struct nullb_page *__null_lookup_page(struct nullb *nullb,
 797         sector_t sector, bool for_write, bool is_cache)
 798 {
 799         unsigned int sector_bit;
 800         u64 idx;
 801         struct nullb_page *t_page;
 802         struct radix_tree_root *root;
 803
 804         idx = sector >> PAGE_SECTORS_SHIFT;
 805         sector_bit = (sector & SECTOR_MASK);
 806
 807         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 808         t_page = radix_tree_lookup(root, idx);
 809         WARN_ON(t_page && t_page->page->index != idx);
 810
 811         if (t_page && (for_write || test_bit(sector_bit, t_page->bitmap)))
 812                 return t_page;
 813
 814         return NULL;
 815 }
 816
 817 static struct nullb_page *null_lookup_page(struct nullb *nullb,
 818         sector_t sector, bool for_write, bool ignore_cache)
 819 {
 820         struct nullb_page *page = NULL;
 821
 822         if (!ignore_cache)
 823                 page = __null_lookup_page(nullb, sector, for_write, true);
 824         if (page)
 825                 return page;
 826         return __null_lookup_page(nullb, sector, for_write, false);
 827 }
 828
 829 static struct nullb_page *null_insert_page(struct nullb *nullb,
 830                                            sector_t sector, bool ignore_cache)
 831         __releases(&nullb->lock)
 832         __acquires(&nullb->lock)
 833 {
 834         u64 idx;
 835         struct nullb_page *t_page;
 836
 837         t_page = null_lookup_page(nullb, sector, true, ignore_cache);
 838         if (t_page)
 839                 return t_page;
 840
 841         spin_unlock_irq(&nullb->lock);
 842
 843         t_page = null_alloc_page(GFP_NOIO);
 844         if (!t_page)
 845                 goto out_lock;
 846
 847         if (radix_tree_preload(GFP_NOIO))
 848                 goto out_freepage;
 849
 850         spin_lock_irq(&nullb->lock);
 851         idx = sector >> PAGE_SECTORS_SHIFT;
 852         t_page->page->index = idx;
 853         t_page = null_radix_tree_insert(nullb, idx, t_page, !ignore_cache);
 854         radix_tree_preload_end();
 855
 856         return t_page;
 857 out_freepage:
 858         null_free_page(t_page);
 859 out_lock:
 860         spin_lock_irq(&nullb->lock);
 861         return null_lookup_page(nullb, sector, true, ignore_cache);
 862 }
 863
 864 static int null_flush_cache_page(struct nullb *nullb, struct nullb_page *c_page)
 865 {
 866         int i;
 867         unsigned int offset;
 868         u64 idx;
 869         struct nullb_page *t_page, *ret;
 870         void *dst, *src;
 871
 872         idx = c_page->page->index;
 873
 874         t_page = null_insert_page(nullb, idx << PAGE_SECTORS_SHIFT, true);
 875
 876         __clear_bit(NULLB_PAGE_LOCK, c_page->bitmap);
 877         if (test_bit(NULLB_PAGE_FREE, c_page->bitmap)) {
 878                 null_free_page(c_page);
 879                 if (t_page && null_page_empty(t_page)) {
 880                         ret = radix_tree_delete_item(&nullb->dev->data,
 881                                 idx, t_page);
 882                         null_free_page(t_page);
 883                 }
 884                 return 0;
 885         }
 886
 887         if (!t_page)
 888                 return -ENOMEM;
 889
 890         src = kmap_atomic(c_page->page);
 891         dst = kmap_atomic(t_page->page);
 892
 893         for (i = 0; i < PAGE_SECTORS;
 894                         i += (nullb->dev->blocksize >> SECTOR_SHIFT)) {
 895                 if (test_bit(i, c_page->bitmap)) {
 896                         offset = (i << SECTOR_SHIFT);
 897                         memcpy(dst + offset, src + offset,
 898                                 nullb->dev->blocksize);
 899                         __set_bit(i, t_page->bitmap);
 900                 }
 901         }
 902
 903         kunmap_atomic(dst);
 904         kunmap_atomic(src);
 905
 906         ret = radix_tree_delete_item(&nullb->dev->cache, idx, c_page);
 907         null_free_page(ret);
 908         nullb->dev->curr_cache -= PAGE_SIZE;
 909
 910         return 0;
 911 }
 912
 913 static int null_make_cache_space(struct nullb *nullb, unsigned long n)
 914 {
 915         int i, err, nr_pages;
 916         struct nullb_page *c_pages[FREE_BATCH];
 917         unsigned long flushed = 0, one_round;
 918
 919 again:
 920         if ((nullb->dev->cache_size * 1024 * 1024) >
 921              nullb->dev->curr_cache + n || nullb->dev->curr_cache == 0)
 922                 return 0;
 923
 924         nr_pages = radix_tree_gang_lookup(&nullb->dev->cache,
 925                         (void **)c_pages, nullb->cache_flush_pos, FREE_BATCH);
 926         /*
 927          * nullb_flush_cache_page could unlock before using the c_pages. To
 928          * avoid race, we don't allow page free
 929          */
 930         for (i = 0; i < nr_pages; i++) {
 931                 nullb->cache_flush_pos = c_pages[i]->page->index;
 932                 /*
 933                  * We found the page which is being flushed to disk by other
 934                  * threads
 935                  */
 936                 if (test_bit(NULLB_PAGE_LOCK, c_pages[i]->bitmap))
 937                         c_pages[i] = NULL;
 938                 else
 939                         __set_bit(NULLB_PAGE_LOCK, c_pages[i]->bitmap);
 940         }
 941
 942         one_round = 0;
 943         for (i = 0; i < nr_pages; i++) {
 944                 if (c_pages[i] == NULL)
 945                         continue;
 946                 err = null_flush_cache_page(nullb, c_pages[i]);
 947                 if (err)
 948                         return err;
 949                 one_round++;
 950         }
 951         flushed += one_round << PAGE_SHIFT;
 952
 953         if (n > flushed) {
 954                 if (nr_pages == 0)
 955                         nullb->cache_flush_pos = 0;
 956                 if (one_round == 0) {
 957                         /* give other threads a chance */
 958                         spin_unlock_irq(&nullb->lock);
 959                         spin_lock_irq(&nullb->lock);
 960                 }
 961                 goto again;
 962         }
 963         return 0;
 964 }
 965
 966 static int copy_to_nullb(struct nullb *nullb, struct page *source,
 967         unsigned int off, sector_t sector, size_t n, bool is_fua)
 968 {
 969         size_t temp, count = 0;
 970         unsigned int offset;
 971         struct nullb_page *t_page;
 972         void *dst, *src;
 973
 974         while (count < n) {
 975                 temp = min_t(size_t, nullb->dev->blocksize, n - count);
 976
 977                 if (null_cache_active(nullb) && !is_fua)
 978                         null_make_cache_space(nullb, PAGE_SIZE);
 979
 980                 offset = (sector & SECTOR_MASK) << SECTOR_SHIFT;
 981                 t_page = null_insert_page(nullb, sector,
 982                         !null_cache_active(nullb) || is_fua);
 983                 if (!t_page)
 984                         return -ENOSPC;
 985
 986                 src = kmap_atomic(source);
 987                 dst = kmap_atomic(t_page->page);
 988                 memcpy(dst + offset, src + off + count, temp);
 989                 kunmap_atomic(dst);
 990                 kunmap_atomic(src);
 991
 992                 __set_bit(sector & SECTOR_MASK, t_page->bitmap);
 993
 994                 if (is_fua)
 995                         null_free_sector(nullb, sector, true);
 996
 997                 count += temp;
 998                 sector += temp >> SECTOR_SHIFT;
 999         }
1000         return 0;
1001 }
1002
1003 static int copy_from_nullb(struct nullb *nullb, struct page *dest,
1004         unsigned int off, sector_t sector, size_t n)
1005 {
1006         size_t temp, count = 0;
1007         unsigned int offset;
1008         struct nullb_page *t_page;
1009         void *dst, *src;
1010
1011         while (count < n) {
1012                 temp = min_t(size_t, nullb->dev->blocksize, n - count);
1013
1014                 offset = (sector & SECTOR_MASK) << SECTOR_SHIFT;
1015                 t_page = null_lookup_page(nullb, sector, false,
1016                         !null_cache_active(nullb));
1017
1018                 dst = kmap_atomic(dest);
1019                 if (!t_page) {
1020                         memset(dst + off + count, 0, temp);
1021                         goto next;
1022                 }
1023                 src = kmap_atomic(t_page->page);
1024                 memcpy(dst + off + count, src + offset, temp);
1025                 kunmap_atomic(src);
1026 next:
1027                 kunmap_atomic(dst);
1028
1029                 count += temp;
1030                 sector += temp >> SECTOR_SHIFT;
1031         }
1032         return 0;
1033 }
1034
1035 static void nullb_fill_pattern(struct nullb *nullb, struct page *page,
1036                                unsigned int len, unsigned int off)
1037 {
1038         void *dst;
1039
1040         dst = kmap_atomic(page);
1041         memset(dst + off, 0xFF, len);
1042         kunmap_atomic(dst);
1043 }
1044
1045 static void null_handle_discard(struct nullb *nullb, sector_t sector, size_t n)
1046 {
1047         size_t temp;
1048
1049         spin_lock_irq(&nullb->lock);
1050         while (n > 0) {
1051                 temp = min_t(size_t, n, nullb->dev->blocksize);
1052                 null_free_sector(nullb, sector, false);
1053                 if (null_cache_active(nullb))
1054                         null_free_sector(nullb, sector, true);
1055                 sector += temp >> SECTOR_SHIFT;
1056                 n -= temp;
1057         }
1058         spin_unlock_irq(&nullb->lock);
1059 }
1060
1061 static int null_handle_flush(struct nullb *nullb)
1062 {
1063         int err;
1064
1065         if (!null_cache_active(nullb))
1066                 return 0;
1067
1068         spin_lock_irq(&nullb->lock);
1069         while (true) {
1070                 err = null_make_cache_space(nullb,
1071                         nullb->dev->cache_size * 1024 * 1024);
1072                 if (err || nullb->dev->curr_cache == 0)
1073                         break;
1074         }
1075
1076         WARN_ON(!radix_tree_empty(&nullb->dev->cache));
1077         spin_unlock_irq(&nullb->lock);
1078         return err;
1079 }
1080
1081 static int null_transfer(struct nullb *nullb, struct page *page,
1082         unsigned int len, unsigned int off, bool is_write, sector_t sector,
1083         bool is_fua)
1084 {
1085         struct nullb_device *dev = nullb->dev;
1086         unsigned int valid_len = len;
1087         int err = 0;
1088
1089         if (!is_write) {
1090                 if (dev->zoned)
1091                         valid_len = null_zone_valid_read_len(nullb,
1092                                 sector, len);
1093
1094                 if (valid_len) {
1095                         err = copy_from_nullb(nullb, page, off,
1096                                 sector, valid_len);
1097                         off += valid_len;
1098                         len -= valid_len;
1099                 }
1100
1101                 if (len)
1102                         nullb_fill_pattern(nullb, page, len, off);
1103                 flush_dcache_page(page);
1104         } else {
1105                 flush_dcache_page(page);
1106                 err = copy_to_nullb(nullb, page, off, sector, len, is_fua);
1107         }
1108
1109         return err;
1110 }
1111
1112 static int null_handle_rq(struct nullb_cmd *cmd)
1113 {
1114         struct request *rq = cmd->rq;
1115         struct nullb *nullb = cmd->nq->dev->nullb;
1116         int err;
1117         unsigned int len;
1118         sector_t sector;
1119         struct req_iterator iter;
1120         struct bio_vec bvec;
1121
1122         sector = blk_rq_pos(rq);
1123
1124         if (req_op(rq) == REQ_OP_DISCARD) {
1125                 null_handle_discard(nullb, sector, blk_rq_bytes(rq));
1126                 return 0;
1127         }
1128
1129         spin_lock_irq(&nullb->lock);
1130         rq_for_each_segment(bvec, rq, iter) {
1131                 len = bvec.bv_len;
1132                 err = null_transfer(nullb, bvec.bv_page, len, bvec.bv_offset,
1133                                      op_is_write(req_op(rq)), sector,
1134                                      req_op(rq) & REQ_FUA);
1135                 if (err) {
1136                         spin_unlock_irq(&nullb->lock);
1137                         return err;
1138                 }
1139                 sector += len >> SECTOR_SHIFT;
1140         }
1141         spin_unlock_irq(&nullb->lock);
1142
1143         return 0;
1144 }
1145
1146 static int null_handle_bio(struct nullb_cmd *cmd)
1147 {
1148         struct bio *bio = cmd->bio;
1149         struct nullb *nullb = cmd->nq->dev->nullb;
1150         int err;
1151         unsigned int len;
1152         sector_t sector;
1153         struct bio_vec bvec;
1154         struct bvec_iter iter;
1155
1156         sector = bio->bi_iter.bi_sector;
1157
1158         if (bio_op(bio) == REQ_OP_DISCARD) {
1159                 null_handle_discard(nullb, sector,
1160                         bio_sectors(bio) << SECTOR_SHIFT);
1161                 return 0;
1162         }
1163
1164         spin_lock_irq(&nullb->lock);
1165         bio_for_each_segment(bvec, bio, iter) {
1166                 len = bvec.bv_len;
1167                 err = null_transfer(nullb, bvec.bv_page, len, bvec.bv_offset,
1168                                      op_is_write(bio_op(bio)), sector,
1169                                      bio->bi_opf & REQ_FUA);
1170                 if (err) {
1171                         spin_unlock_irq(&nullb->lock);
1172                         return err;
1173                 }
1174                 sector += len >> SECTOR_SHIFT;
1175         }
1176         spin_unlock_irq(&nullb->lock);
1177         return 0;
1178 }
1179
1180 static void null_stop_queue(struct nullb *nullb)
1181 {
1182         struct request_queue *q = nullb->q;
1183
1184         if (nullb->dev->queue_mode == NULL_Q_MQ)
1185                 blk_mq_stop_hw_queues(q);
1186 }
1187
1188 static void null_restart_queue_async(struct nullb *nullb)
1189 {
1190         struct request_queue *q = nullb->q;
1191
1192         if (nullb->dev->queue_mode == NULL_Q_MQ)
1193                 blk_mq_start_stopped_hw_queues(q, true);
1194 }
1195
1196 static inline blk_status_t null_handle_throttled(struct nullb_cmd *cmd)
1197 {
1198         struct nullb_device *dev = cmd->nq->dev;
1199         struct nullb *nullb = dev->nullb;
1200         blk_status_t sts = BLK_STS_OK;
1201         struct request *rq = cmd->rq;
1202
1203         if (!hrtimer_active(&nullb->bw_timer))
1204                 hrtimer_restart(&nullb->bw_timer);
1205
1206         if (atomic_long_sub_return(blk_rq_bytes(rq), &nullb->cur_bytes) < 0) {
1207                 null_stop_queue(nullb);
1208                 /* race with timer */
1209                 if (atomic_long_read(&nullb->cur_bytes) > 0)
1210                         null_restart_queue_async(nullb);
1211                 /* requeue request */
1212                 sts = BLK_STS_DEV_RESOURCE;
1213         }
1214         return sts;
1215 }
1216
1217 static inline blk_status_t null_handle_badblocks(struct nullb_cmd *cmd,
1218                                                  sector_t sector,
1219                                                  sector_t nr_sectors)
1220 {
1221         struct badblocks *bb = &cmd->nq->dev->badblocks;
1222         sector_t first_bad;
1223         int bad_sectors;
1224
1225         if (badblocks_check(bb, sector, nr_sectors, &first_bad, &bad_sectors))
1226                 return BLK_STS_IOERR;
1227
1228         return BLK_STS_OK;
1229 }
1230
1231 static inline blk_status_t null_handle_memory_backed(struct nullb_cmd *cmd,
1232                                                      enum req_opf op)
1233 {
1234         struct nullb_device *dev = cmd->nq->dev;
1235         int err;
1236
1237         if (dev->queue_mode == NULL_Q_BIO)
1238                 err = null_handle_bio(cmd);
1239         else
1240                 err = null_handle_rq(cmd);
1241
1242         return errno_to_blk_status(err);
1243 }
1244
1245 static inline void nullb_complete_cmd(struct nullb_cmd *cmd)
1246 {
1247         /* Complete IO by inline, softirq or timer */
1248         switch (cmd->nq->dev->irqmode) {
1249         case NULL_IRQ_SOFTIRQ:
1250                 switch (cmd->nq->dev->queue_mode) {
1251                 case NULL_Q_MQ:
1252                         blk_mq_complete_request(cmd->rq);
1253                         break;
1254                 case NULL_Q_BIO:
1255                         /*
1256                          * XXX: no proper submitting cpu information available.
1257                          */
1258                         end_cmd(cmd);
1259                         break;
1260                 }
1261                 break;
1262         case NULL_IRQ_NONE:
1263                 end_cmd(cmd);
1264                 break;
1265         case NULL_IRQ_TIMER:
1266                 null_cmd_end_timer(cmd);
1267                 break;
1268         }
1269 }
1270
1271 static blk_status_t null_handle_cmd(struct nullb_cmd *cmd, sector_t sector,
1272                                     sector_t nr_sectors, enum req_opf op)
1273 {
1274         struct nullb_device *dev = cmd->nq->dev;
1275         struct nullb *nullb = dev->nullb;
1276         blk_status_t sts;
1277
1278         if (test_bit(NULLB_DEV_FL_THROTTLED, &dev->flags)) {
1279                 sts = null_handle_throttled(cmd);
1280                 if (sts != BLK_STS_OK)
1281                         return sts;
1282         }
1283
1284         if (op == REQ_OP_FLUSH) {
1285                 cmd->error = errno_to_blk_status(null_handle_flush(nullb));
1286                 goto out;
1287         }
1288
1289         if (nullb->dev->badblocks.shift != -1) {
1290                 cmd->error = null_handle_badblocks(cmd, sector, nr_sectors);
1291                 if (cmd->error != BLK_STS_OK)
1292                         goto out;
1293         }
1294
1295         if (dev->memory_backed)
1296                 cmd->error = null_handle_memory_backed(cmd, op);
1297
1298         if (!cmd->error && dev->zoned)
1299                 cmd->error = null_handle_zoned(cmd, op, sector, nr_sectors);
1300
1301 out:
1302         nullb_complete_cmd(cmd);
1303         return BLK_STS_OK;
1304 }
1305
1306 static enum hrtimer_restart nullb_bwtimer_fn(struct hrtimer *timer)
1307 {
1308         struct nullb *nullb = container_of(timer, struct nullb, bw_timer);
1309         ktime_t timer_interval = ktime_set(0, TIMER_INTERVAL);
1310         unsigned int mbps = nullb->dev->mbps;
1311
1312         if (atomic_long_read(&nullb->cur_bytes) == mb_per_tick(mbps))
1313                 return HRTIMER_NORESTART;
1314
1315         atomic_long_set(&nullb->cur_bytes, mb_per_tick(mbps));
1316         null_restart_queue_async(nullb);
1317
1318         hrtimer_forward_now(&nullb->bw_timer, timer_interval);
1319
1320         return HRTIMER_RESTART;
1321 }
1322
1323 static void nullb_setup_bwtimer(struct nullb *nullb)
1324 {
1325         ktime_t timer_interval = ktime_set(0, TIMER_INTERVAL);
1326
1327         hrtimer_init(&nullb->bw_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1328         nullb->bw_timer.function = nullb_bwtimer_fn;
1329         atomic_long_set(&nullb->cur_bytes, mb_per_tick(nullb->dev->mbps));
1330         hrtimer_start(&nullb->bw_timer, timer_interval, HRTIMER_MODE_REL);
1331 }
1332
1333 static struct nullb_queue *nullb_to_queue(struct nullb *nullb)
1334 {
1335         int index = 0;
1336
1337         if (nullb->nr_queues != 1)
1338                 index = raw_smp_processor_id() / ((nr_cpu_ids + nullb->nr_queues - 1) / nullb->nr_queues);
1339
1340         return &nullb->queues[index];
1341 }
1342
1343 static blk_qc_t null_queue_bio(struct request_queue *q, struct bio *bio)
1344 {
1345         sector_t sector = bio->bi_iter.bi_sector;
1346         sector_t nr_sectors = bio_sectors(bio);
1347         struct nullb *nullb = q->queuedata;
1348         struct nullb_queue *nq = nullb_to_queue(nullb);
1349         struct nullb_cmd *cmd;
1350
1351         cmd = alloc_cmd(nq, 1);
1352         cmd->bio = bio;
1353
1354         null_handle_cmd(cmd, sector, nr_sectors, bio_op(bio));
1355         return BLK_QC_T_NONE;
1356 }
1357
1358 static bool should_timeout_request(struct request *rq)
1359 {
1360 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1361         if (g_timeout_str[0])
1362                 return should_fail(&null_timeout_attr, 1);
1363 #endif
1364         return false;
1365 }
1366
1367 static bool should_requeue_request(struct request *rq)
1368 {
1369 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1370         if (g_requeue_str[0])
1371                 return should_fail(&null_requeue_attr, 1);
1372 #endif
1373         return false;
1374 }
1375
1376 static enum blk_eh_timer_return null_timeout_rq(struct request *rq, bool res)
1377 {
1378         pr_info("rq %p timed out\n", rq);
1379         blk_mq_complete_request(rq);
1380         return BLK_EH_DONE;
1381 }
1382
1383 static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx,
1384                          const struct blk_mq_queue_data *bd)
1385 {
1386         struct nullb_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1387         struct nullb_queue *nq = hctx->driver_data;
1388         sector_t nr_sectors = blk_rq_sectors(bd->rq);
1389         sector_t sector = blk_rq_pos(bd->rq);
1390
1391         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1392
1393         if (nq->dev->irqmode == NULL_IRQ_TIMER) {
1394                 hrtimer_init(&cmd->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1395                 cmd->timer.function = null_cmd_timer_expired;
1396         }
1397         cmd->rq = bd->rq;
1398         cmd->nq = nq;
1399
1400         blk_mq_start_request(bd->rq);
1401
1402         if (should_requeue_request(bd->rq)) {
1403                 /*
1404                  * Alternate between hitting the core BUSY path, and the
1405                  * driver driven requeue path
1406                  */
1407                 nq->requeue_selection++;
1408                 if (nq->requeue_selection & 1)
1409                         return BLK_STS_RESOURCE;
1410                 else {
1411                         blk_mq_requeue_request(bd->rq, true);
1412                         return BLK_STS_OK;
1413                 }
1414         }
1415         if (should_timeout_request(bd->rq))
1416                 return BLK_STS_OK;
1417
1418         return null_handle_cmd(cmd, sector, nr_sectors, req_op(bd->rq));
1419 }
1420
1421 static void cleanup_queue(struct nullb_queue *nq)
1422 {
1423         kfree(nq->tag_map);
1424         kfree(nq->cmds);
1425 }
1426
1427 static void cleanup_queues(struct nullb *nullb)
1428 {
1429         int i;
1430
1431         for (i = 0; i < nullb->nr_queues; i++)
1432                 cleanup_queue(&nullb->queues[i]);
1433
1434         kfree(nullb->queues);
1435 }
1436
1437 static void null_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1438 {
1439         struct nullb_queue *nq = hctx->driver_data;
1440         struct nullb *nullb = nq->dev->nullb;
1441
1442         nullb->nr_queues--;
1443 }
1444
1445 static void null_init_queue(struct nullb *nullb, struct nullb_queue *nq)
1446 {
1447         init_waitqueue_head(&nq->wait);
1448         nq->queue_depth = nullb->queue_depth;
1449         nq->dev = nullb->dev;
1450 }
1451
1452 static int null_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
1453                           unsigned int hctx_idx)
1454 {
1455         struct nullb *nullb = hctx->queue->queuedata;
1456         struct nullb_queue *nq;
1457
1458 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1459         if (g_init_hctx_str[0] && should_fail(&null_init_hctx_attr, 1))
1460                 return -EFAULT;
1461 #endif
1462
1463         nq = &nullb->queues[hctx_idx];
1464         hctx->driver_data = nq;
1465         null_init_queue(nullb, nq);
1466         nullb->nr_queues++;
1467
1468         return 0;
1469 }
1470
1471 static const struct blk_mq_ops null_mq_ops = {
1472         .queue_rq       = null_queue_rq,
1473         .complete       = null_complete_rq,
1474         .timeout        = null_timeout_rq,
1475         .init_hctx      = null_init_hctx,
1476         .exit_hctx      = null_exit_hctx,
1477 };
1478
1479 static void null_del_dev(struct nullb *nullb)
1480 {
1481         struct nullb_device *dev;
1482
1483         if (!nullb)
1484                 return;
1485
1486         dev = nullb->dev;
1487
1488         ida_simple_remove(&nullb_indexes, nullb->index);
1489
1490         list_del_init(&nullb->list);
1491
1492         del_gendisk(nullb->disk);
1493
1494         if (test_bit(NULLB_DEV_FL_THROTTLED, &nullb->dev->flags)) {
1495                 hrtimer_cancel(&nullb->bw_timer);
1496                 atomic_long_set(&nullb->cur_bytes, LONG_MAX);
1497                 null_restart_queue_async(nullb);
1498         }
1499
1500         blk_cleanup_queue(nullb->q);
1501         if (dev->queue_mode == NULL_Q_MQ &&
1502             nullb->tag_set == &nullb->__tag_set)
1503                 blk_mq_free_tag_set(nullb->tag_set);
1504         put_disk(nullb->disk);
1505         cleanup_queues(nullb);
1506         if (null_cache_active(nullb))
1507                 null_free_device_storage(nullb->dev, true);
1508         kfree(nullb);
1509         dev->nullb = NULL;
1510 }
1511
1512 static void null_config_discard(struct nullb *nullb)
1513 {
1514         if (nullb->dev->discard == false)
1515                 return;
1516         nullb->q->limits.discard_granularity = nullb->dev->blocksize;
1517         nullb->q->limits.discard_alignment = nullb->dev->blocksize;
1518         blk_queue_max_discard_sectors(nullb->q, UINT_MAX >> 9);
1519         blk_queue_flag_set(QUEUE_FLAG_DISCARD, nullb->q);
1520 }
1521
1522 static const struct block_device_operations null_ops = {
1523         .owner          = THIS_MODULE,
1524         .report_zones   = null_report_zones,
1525 };
1526
1527 static int setup_commands(struct nullb_queue *nq)
1528 {
1529         struct nullb_cmd *cmd;
1530         int i, tag_size;
1531
1532         nq->cmds = kcalloc(nq->queue_depth, sizeof(*cmd), GFP_KERNEL);
1533         if (!nq->cmds)
1534                 return -ENOMEM;
1535
1536         tag_size = ALIGN(nq->queue_depth, BITS_PER_LONG) / BITS_PER_LONG;
1537         nq->tag_map = kcalloc(tag_size, sizeof(unsigned long), GFP_KERNEL);
1538         if (!nq->tag_map) {
1539                 kfree(nq->cmds);
1540                 return -ENOMEM;
1541         }
1542
1543         for (i = 0; i < nq->queue_depth; i++) {
1544                 cmd = &nq->cmds[i];
1545                 cmd->tag = -1U;
1546         }
1547
1548         return 0;
1549 }
1550
1551 static int setup_queues(struct nullb *nullb)
1552 {
1553         nullb->queues = kcalloc(nr_cpu_ids, sizeof(struct nullb_queue),
1554                                 GFP_KERNEL);
1555         if (!nullb->queues)
1556                 return -ENOMEM;
1557
1558         nullb->queue_depth = nullb->dev->hw_queue_depth;
1559
1560         return 0;
1561 }
1562
1563 static int init_driver_queues(struct nullb *nullb)
1564 {
1565         struct nullb_queue *nq;
1566         int i, ret = 0;
1567
1568         for (i = 0; i < nullb->dev->submit_queues; i++) {
1569                 nq = &nullb->queues[i];
1570
1571                 null_init_queue(nullb, nq);
1572
1573                 ret = setup_commands(nq);
1574                 if (ret)
1575                         return ret;
1576                 nullb->nr_queues++;
1577         }
1578         return 0;
1579 }
1580
1581 static int null_gendisk_register(struct nullb *nullb)
1582 {
1583         sector_t size = ((sector_t)nullb->dev->size * SZ_1M) >> SECTOR_SHIFT;
1584         struct gendisk *disk;
1585
1586         disk = nullb->disk = alloc_disk_node(1, nullb->dev->home_node);
1587         if (!disk)
1588                 return -ENOMEM;
1589         set_capacity(disk, size);
1590
1591         disk->flags |= GENHD_FL_EXT_DEVT | GENHD_FL_SUPPRESS_PARTITION_INFO;
1592         disk->major             = null_major;
1593         disk->first_minor       = nullb->index;
1594         disk->fops              = &null_ops;
1595         disk->private_data      = nullb;
1596         disk->queue             = nullb->q;
1597         strncpy(disk->disk_name, nullb->disk_name, DISK_NAME_LEN);
1598
1599 #ifdef CONFIG_BLK_DEV_ZONED
1600         if (nullb->dev->zoned) {
1601                 if (queue_is_mq(nullb->q)) {
1602                         int ret = blk_revalidate_disk_zones(disk);
1603                         if (ret)
1604                                 return ret;
1605                 } else {
1606                         blk_queue_chunk_sectors(nullb->q,
1607                                         nullb->dev->zone_size_sects);
1608                         nullb->q->nr_zones = blkdev_nr_zones(disk);
1609                 }
1610         }
1611 #endif
1612
1613         add_disk(disk);
1614         return 0;
1615 }
1616
1617 static int null_init_tag_set(struct nullb *nullb, struct blk_mq_tag_set *set)
1618 {
1619         set->ops = &null_mq_ops;
1620         set->nr_hw_queues = nullb ? nullb->dev->submit_queues :
1621                                                 g_submit_queues;
1622         set->queue_depth = nullb ? nullb->dev->hw_queue_depth :
1623                                                 g_hw_queue_depth;
1624         set->numa_node = nullb ? nullb->dev->home_node : g_home_node;
1625         set->cmd_size   = sizeof(struct nullb_cmd);
1626         set->flags = BLK_MQ_F_SHOULD_MERGE;
1627         if (g_no_sched)
1628                 set->flags |= BLK_MQ_F_NO_SCHED;
1629         set->driver_data = NULL;
1630
1631         if ((nullb && nullb->dev->blocking) || g_blocking)
1632                 set->flags |= BLK_MQ_F_BLOCKING;
1633
1634         return blk_mq_alloc_tag_set(set);
1635 }
1636
1637 static int null_validate_conf(struct nullb_device *dev)
1638 {
1639         dev->blocksize = round_down(dev->blocksize, 512);
1640         dev->blocksize = clamp_t(unsigned int, dev->blocksize, 512, 4096);
1641
1642         if (dev->queue_mode == NULL_Q_MQ && dev->use_per_node_hctx) {
1643                 if (dev->submit_queues != nr_online_nodes)
1644                         dev->submit_queues = nr_online_nodes;
1645         } else if (dev->submit_queues > nr_cpu_ids)
1646                 dev->submit_queues = nr_cpu_ids;
1647         else if (dev->submit_queues == 0)
1648                 dev->submit_queues = 1;
1649
1650         dev->queue_mode = min_t(unsigned int, dev->queue_mode, NULL_Q_MQ);
1651         dev->irqmode = min_t(unsigned int, dev->irqmode, NULL_IRQ_TIMER);
1652
1653         /* Do memory allocation, so set blocking */
1654         if (dev->memory_backed)
1655                 dev->blocking = true;
1656         else /* cache is meaningless */
1657                 dev->cache_size = 0;
1658         dev->cache_size = min_t(unsigned long, ULONG_MAX / 1024 / 1024,
1659                                                 dev->cache_size);
1660         dev->mbps = min_t(unsigned int, 1024 * 40, dev->mbps);
1661         /* can not stop a queue */
1662         if (dev->queue_mode == NULL_Q_BIO)
1663                 dev->mbps = 0;
1664
1665         if (dev->zoned &&
1666             (!dev->zone_size || !is_power_of_2(dev->zone_size))) {
1667                 pr_err("zone_size must be power-of-two\n");
1668                 return -EINVAL;
1669         }
1670
1671         return 0;
1672 }
1673
1674 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1675 static bool __null_setup_fault(struct fault_attr *attr, char *str)
1676 {
1677         if (!str[0])
1678                 return true;
1679
1680         if (!setup_fault_attr(attr, str))
1681                 return false;
1682
1683         attr->verbose = 0;
1684         return true;
1685 }
1686 #endif
1687
1688 static bool null_setup_fault(void)
1689 {
1690 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1691         if (!__null_setup_fault(&null_timeout_attr, g_timeout_str))
1692                 return false;
1693         if (!__null_setup_fault(&null_requeue_attr, g_requeue_str))
1694                 return false;
1695         if (!__null_setup_fault(&null_init_hctx_attr, g_init_hctx_str))
1696                 return false;
1697 #endif
1698         return true;
1699 }
1700
1701 static int null_add_dev(struct nullb_device *dev)
1702 {
1703         struct nullb *nullb;
1704         int rv;
1705
1706         rv = null_validate_conf(dev);
1707         if (rv)
1708                 return rv;
1709
1710         nullb = kzalloc_node(sizeof(*nullb), GFP_KERNEL, dev->home_node);
1711         if (!nullb) {
1712                 rv = -ENOMEM;
1713                 goto out;
1714         }
1715         nullb->dev = dev;
1716         dev->nullb = nullb;
1717
1718         spin_lock_init(&nullb->lock);
1719
1720         rv = setup_queues(nullb);
1721         if (rv)
1722                 goto out_free_nullb;
1723
1724         if (dev->queue_mode == NULL_Q_MQ) {
1725                 if (shared_tags) {
1726                         nullb->tag_set = &tag_set;
1727                         rv = 0;
1728                 } else {
1729                         nullb->tag_set = &nullb->__tag_set;
1730                         rv = null_init_tag_set(nullb, nullb->tag_set);
1731                 }
1732
1733                 if (rv)
1734                         goto out_cleanup_queues;
1735
1736                 if (!null_setup_fault())
1737                         goto out_cleanup_queues;
1738
1739                 nullb->tag_set->timeout = 5 * HZ;
1740                 nullb->q = blk_mq_init_queue_data(nullb->tag_set, nullb);
1741                 if (IS_ERR(nullb->q)) {
1742                         rv = -ENOMEM;
1743                         goto out_cleanup_tags;
1744                 }
1745         } else if (dev->queue_mode == NULL_Q_BIO) {
1746                 nullb->q = blk_alloc_queue_node(GFP_KERNEL, dev->home_node);
1747                 if (!nullb->q) {
1748                         rv = -ENOMEM;
1749                         goto out_cleanup_queues;
1750                 }
1751                 blk_queue_make_request(nullb->q, null_queue_bio);
1752                 rv = init_driver_queues(nullb);
1753                 if (rv)
1754                         goto out_cleanup_blk_queue;
1755         }
1756
1757         if (dev->mbps) {
1758                 set_bit(NULLB_DEV_FL_THROTTLED, &dev->flags);
1759                 nullb_setup_bwtimer(nullb);
1760         }
1761
1762         if (dev->cache_size > 0) {
1763                 set_bit(NULLB_DEV_FL_CACHE, &nullb->dev->flags);
1764                 blk_queue_write_cache(nullb->q, true, true);
1765         }
1766
1767         if (dev->zoned) {
1768                 rv = null_zone_init(dev);
1769                 if (rv)
1770                         goto out_cleanup_blk_queue;
1771
1772                 nullb->q->limits.zoned = BLK_ZONED_HM;
1773                 blk_queue_flag_set(QUEUE_FLAG_ZONE_RESETALL, nullb->q);
1774                 blk_queue_required_elevator_features(nullb->q,
1775                                                 ELEVATOR_F_ZBD_SEQ_WRITE);
1776         }
1777
1778         nullb->q->queuedata = nullb;
1779         blk_queue_flag_set(QUEUE_FLAG_NONROT, nullb->q);
1780         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, nullb->q);
1781
1782         mutex_lock(&lock);
1783         nullb->index = ida_simple_get(&nullb_indexes, 0, 0, GFP_KERNEL);
1784         dev->index = nullb->index;
1785         mutex_unlock(&lock);
1786
1787         blk_queue_logical_block_size(nullb->q, dev->blocksize);
1788         blk_queue_physical_block_size(nullb->q, dev->blocksize);
1789
1790         null_config_discard(nullb);
1791
1792         sprintf(nullb->disk_name, "nullb%d", nullb->index);
1793
1794         rv = null_gendisk_register(nullb);
1795         if (rv)
1796                 goto out_cleanup_zone;
1797
1798         mutex_lock(&lock);
1799         list_add_tail(&nullb->list, &nullb_list);
1800         mutex_unlock(&lock);
1801
1802         return 0;
1803 out_cleanup_zone:
1804         if (dev->zoned)
1805                 null_zone_exit(dev);
1806 out_cleanup_blk_queue:
1807         blk_cleanup_queue(nullb->q);
1808 out_cleanup_tags:
1809         if (dev->queue_mode == NULL_Q_MQ && nullb->tag_set == &nullb->__tag_set)
1810                 blk_mq_free_tag_set(nullb->tag_set);
1811 out_cleanup_queues:
1812         cleanup_queues(nullb);
1813 out_free_nullb:
1814         kfree(nullb);
1815         dev->nullb = NULL;
1816 out:
1817         return rv;
1818 }
1819
1820 static int __init null_init(void)
1821 {
1822         int ret = 0;
1823         unsigned int i;
1824         struct nullb *nullb;
1825         struct nullb_device *dev;
1826
1827         if (g_bs > PAGE_SIZE) {
1828                 pr_warn("invalid block size\n");
1829                 pr_warn("defaults block size to %lu\n", PAGE_SIZE);
1830                 g_bs = PAGE_SIZE;
1831         }
1832
1833         if (g_home_node != NUMA_NO_NODE && g_home_node >= nr_online_nodes) {
1834                 pr_err("invalid home_node value\n");
1835                 g_home_node = NUMA_NO_NODE;
1836         }
1837
1838         if (g_queue_mode == NULL_Q_RQ) {
1839                 pr_err("legacy IO path no longer available\n");
1840                 return -EINVAL;
1841         }
1842         if (g_queue_mode == NULL_Q_MQ && g_use_per_node_hctx) {
1843                 if (g_submit_queues != nr_online_nodes) {
1844                         pr_warn("submit_queues param is set to %u.\n",
1845                                                         nr_online_nodes);
1846                         g_submit_queues = nr_online_nodes;
1847                 }
1848         } else if (g_submit_queues > nr_cpu_ids)
1849                 g_submit_queues = nr_cpu_ids;
1850         else if (g_submit_queues <= 0)
1851                 g_submit_queues = 1;
1852
1853         if (g_queue_mode == NULL_Q_MQ && shared_tags) {
1854                 ret = null_init_tag_set(NULL, &tag_set);
1855                 if (ret)
1856                         return ret;
1857         }
1858
1859         config_group_init(&nullb_subsys.su_group);
1860         mutex_init(&nullb_subsys.su_mutex);
1861
1862         ret = configfs_register_subsystem(&nullb_subsys);
1863         if (ret)
1864                 goto err_tagset;
1865
1866         mutex_init(&lock);
1867
1868         null_major = register_blkdev(0, "nullb");
1869         if (null_major < 0) {
1870                 ret = null_major;
1871                 goto err_conf;
1872         }
1873
1874         for (i = 0; i < nr_devices; i++) {
1875                 dev = null_alloc_dev();
1876                 if (!dev) {
1877                         ret = -ENOMEM;
1878                         goto err_dev;
1879                 }
1880                 ret = null_add_dev(dev);
1881                 if (ret) {
1882                         null_free_dev(dev);
1883                         goto err_dev;
1884                 }
1885         }
1886
1887         pr_info("module loaded\n");
1888         return 0;
1889
1890 err_dev:
1891         while (!list_empty(&nullb_list)) {
1892                 nullb = list_entry(nullb_list.next, struct nullb, list);
1893                 dev = nullb->dev;
1894                 null_del_dev(nullb);
1895                 null_free_dev(dev);
1896         }
1897         unregister_blkdev(null_major, "nullb");
1898 err_conf:
1899         configfs_unregister_subsystem(&nullb_subsys);
1900 err_tagset:
1901         if (g_queue_mode == NULL_Q_MQ && shared_tags)
1902                 blk_mq_free_tag_set(&tag_set);
1903         return ret;
1904 }
1905
1906 static void __exit null_exit(void)
1907 {
1908         struct nullb *nullb;
1909
1910         configfs_unregister_subsystem(&nullb_subsys);
1911
1912         unregister_blkdev(null_major, "nullb");
1913
1914         mutex_lock(&lock);
1915         while (!list_empty(&nullb_list)) {
1916                 struct nullb_device *dev;
1917
1918                 nullb = list_entry(nullb_list.next, struct nullb, list);
1919                 dev = nullb->dev;
1920                 null_del_dev(nullb);
1921                 null_free_dev(dev);
1922         }
1923         mutex_unlock(&lock);
1924
1925         if (g_queue_mode == NULL_Q_MQ && shared_tags)
1926                 blk_mq_free_tag_set(&tag_set);
1927 }
1928
1929 module_init(null_init);
1930 module_exit(null_exit);
1931
1932 MODULE_AUTHOR("Jens Axboe <axboe@kernel.dk>");
1933 MODULE_LICENSE("GPL");