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 /*
 101  * For more details about fault injection, please refer to
 102  * Documentation/fault-injection/fault-injection.rst.
 103  */
 104 static char g_timeout_str[80];
 105 module_param_string(timeout, g_timeout_str, sizeof(g_timeout_str), 0444);
 106 MODULE_PARM_DESC(timeout, "Fault injection. timeout=<interval>,<probability>,<space>,<times>");
 107
 108 static char g_requeue_str[80];
 109 module_param_string(requeue, g_requeue_str, sizeof(g_requeue_str), 0444);
 110 MODULE_PARM_DESC(requeue, "Fault injection. requeue=<interval>,<probability>,<space>,<times>");
 111
 112 static char g_init_hctx_str[80];
 113 module_param_string(init_hctx, g_init_hctx_str, sizeof(g_init_hctx_str), 0444);
 114 MODULE_PARM_DESC(init_hctx, "Fault injection to fail hctx init. init_hctx=<interval>,<probability>,<space>,<times>");
 115 #endif
 116
 117 static int g_queue_mode = NULL_Q_MQ;
 118
 119 static int null_param_store_val(const char *str, int *val, int min, int max)
 120 {
 121         int ret, new_val;
 122
 123         ret = kstrtoint(str, 10, &new_val);
 124         if (ret)
 125                 return -EINVAL;
 126
 127         if (new_val < min || new_val > max)
 128                 return -EINVAL;
 129
 130         *val = new_val;
 131         return 0;
 132 }
 133
 134 static int null_set_queue_mode(const char *str, const struct kernel_param *kp)
 135 {
 136         return null_param_store_val(str, &g_queue_mode, NULL_Q_BIO, NULL_Q_MQ);
 137 }
 138
 139 static const struct kernel_param_ops null_queue_mode_param_ops = {
 140         .set    = null_set_queue_mode,
 141         .get    = param_get_int,
 142 };
 143
 144 device_param_cb(queue_mode, &null_queue_mode_param_ops, &g_queue_mode, 0444);
 145 MODULE_PARM_DESC(queue_mode, "Block interface to use (0=bio,1=rq,2=multiqueue)");
 146
 147 static int g_gb = 250;
 148 module_param_named(gb, g_gb, int, 0444);
 149 MODULE_PARM_DESC(gb, "Size in GB");
 150
 151 static int g_bs = 512;
 152 module_param_named(bs, g_bs, int, 0444);
 153 MODULE_PARM_DESC(bs, "Block size (in bytes)");
 154
 155 static unsigned int nr_devices = 1;
 156 module_param(nr_devices, uint, 0444);
 157 MODULE_PARM_DESC(nr_devices, "Number of devices to register");
 158
 159 static bool g_blocking;
 160 module_param_named(blocking, g_blocking, bool, 0444);
 161 MODULE_PARM_DESC(blocking, "Register as a blocking blk-mq driver device");
 162
 163 static bool shared_tags;
 164 module_param(shared_tags, bool, 0444);
 165 MODULE_PARM_DESC(shared_tags, "Share tag set between devices for blk-mq");
 166
 167 static int g_irqmode = NULL_IRQ_SOFTIRQ;
 168
 169 static int null_set_irqmode(const char *str, const struct kernel_param *kp)
 170 {
 171         return null_param_store_val(str, &g_irqmode, NULL_IRQ_NONE,
 172                                         NULL_IRQ_TIMER);
 173 }
 174
 175 static const struct kernel_param_ops null_irqmode_param_ops = {
 176         .set    = null_set_irqmode,
 177         .get    = param_get_int,
 178 };
 179
 180 device_param_cb(irqmode, &null_irqmode_param_ops, &g_irqmode, 0444);
 181 MODULE_PARM_DESC(irqmode, "IRQ completion handler. 0-none, 1-softirq, 2-timer");
 182
 183 static unsigned long g_completion_nsec = 10000;
 184 module_param_named(completion_nsec, g_completion_nsec, ulong, 0444);
 185 MODULE_PARM_DESC(completion_nsec, "Time in ns to complete a request in hardware. Default: 10,000ns");
 186
 187 static int g_hw_queue_depth = 64;
 188 module_param_named(hw_queue_depth, g_hw_queue_depth, int, 0444);
 189 MODULE_PARM_DESC(hw_queue_depth, "Queue depth for each hardware queue. Default: 64");
 190
 191 static bool g_use_per_node_hctx;
 192 module_param_named(use_per_node_hctx, g_use_per_node_hctx, bool, 0444);
 193 MODULE_PARM_DESC(use_per_node_hctx, "Use per-node allocation for hardware context queues. Default: false");
 194
 195 static bool g_zoned;
 196 module_param_named(zoned, g_zoned, bool, S_IRUGO);
 197 MODULE_PARM_DESC(zoned, "Make device as a host-managed zoned block device. Default: false");
 198
 199 static unsigned long g_zone_size = 256;
 200 module_param_named(zone_size, g_zone_size, ulong, S_IRUGO);
 201 MODULE_PARM_DESC(zone_size, "Zone size in MB when block device is zoned. Must be power-of-two: Default: 256");
 202
 203 static unsigned int g_zone_nr_conv;
 204 module_param_named(zone_nr_conv, g_zone_nr_conv, uint, 0444);
 205 MODULE_PARM_DESC(zone_nr_conv, "Number of conventional zones when block device is zoned. Default: 0");
 206
 207 static struct nullb_device *null_alloc_dev(void);
 208 static void null_free_dev(struct nullb_device *dev);
 209 static void null_del_dev(struct nullb *nullb);
 210 static int null_add_dev(struct nullb_device *dev);
 211 static void null_free_device_storage(struct nullb_device *dev, bool is_cache);
 212
 213 static inline struct nullb_device *to_nullb_device(struct config_item *item)
 214 {
 215         return item ? container_of(item, struct nullb_device, item) : NULL;
 216 }
 217
 218 static inline ssize_t nullb_device_uint_attr_show(unsigned int val, char *page)
 219 {
 220         return snprintf(page, PAGE_SIZE, "%u\n", val);
 221 }
 222
 223 static inline ssize_t nullb_device_ulong_attr_show(unsigned long val,
 224         char *page)
 225 {
 226         return snprintf(page, PAGE_SIZE, "%lu\n", val);
 227 }
 228
 229 static inline ssize_t nullb_device_bool_attr_show(bool val, char *page)
 230 {
 231         return snprintf(page, PAGE_SIZE, "%u\n", val);
 232 }
 233
 234 static ssize_t nullb_device_uint_attr_store(unsigned int *val,
 235         const char *page, size_t count)
 236 {
 237         unsigned int tmp;
 238         int result;
 239
 240         result = kstrtouint(page, 0, &tmp);
 241         if (result < 0)
 242                 return result;
 243
 244         *val = tmp;
 245         return count;
 246 }
 247
 248 static ssize_t nullb_device_ulong_attr_store(unsigned long *val,
 249         const char *page, size_t count)
 250 {
 251         int result;
 252         unsigned long tmp;
 253
 254         result = kstrtoul(page, 0, &tmp);
 255         if (result < 0)
 256                 return result;
 257
 258         *val = tmp;
 259         return count;
 260 }
 261
 262 static ssize_t nullb_device_bool_attr_store(bool *val, const char *page,
 263         size_t count)
 264 {
 265         bool tmp;
 266         int result;
 267
 268         result = kstrtobool(page,  &tmp);
 269         if (result < 0)
 270                 return result;
 271
 272         *val = tmp;
 273         return count;
 274 }
 275
 276 /* The following macro should only be used with TYPE = {uint, ulong, bool}. */
 277 #define NULLB_DEVICE_ATTR(NAME, TYPE, APPLY)                            \
 278 static ssize_t                                                          \
 279 nullb_device_##NAME##_show(struct config_item *item, char *page)        \
 280 {                                                                       \
 281         return nullb_device_##TYPE##_attr_show(                         \
 282                                 to_nullb_device(item)->NAME, page);     \
 283 }                                                                       \
 284 static ssize_t                                                          \
 285 nullb_device_##NAME##_store(struct config_item *item, const char *page, \
 286                             size_t count)                               \
 287 {                                                                       \
 288         int (*apply_fn)(struct nullb_device *dev, TYPE new_value) = APPLY;\
 289         struct nullb_device *dev = to_nullb_device(item);               \
 290         TYPE new_value = 0;                                             \
 291         int ret;                                                        \
 292                                                                         \
 293         ret = nullb_device_##TYPE##_attr_store(&new_value, page, count);\
 294         if (ret < 0)                                                    \
 295                 return ret;                                             \
 296         if (apply_fn)                                                   \
 297                 ret = apply_fn(dev, new_value);                         \
 298         else if (test_bit(NULLB_DEV_FL_CONFIGURED, &dev->flags))        \
 299                 ret = -EBUSY;                                           \
 300         if (ret < 0)                                                    \
 301                 return ret;                                             \
 302         dev->NAME = new_value;                                          \
 303         return count;                                                   \
 304 }                                                                       \
 305 CONFIGFS_ATTR(nullb_device_, NAME);
 306
 307 static int nullb_apply_submit_queues(struct nullb_device *dev,
 308                                      unsigned int submit_queues)
 309 {
 310         struct nullb *nullb = dev->nullb;
 311         struct blk_mq_tag_set *set;
 312
 313         if (!nullb)
 314                 return 0;
 315
 316         /*
 317          * Make sure that null_init_hctx() does not access nullb->queues[] past
 318          * the end of that array.
 319          */
 320         if (submit_queues > nr_cpu_ids)
 321                 return -EINVAL;
 322         set = nullb->tag_set;
 323         blk_mq_update_nr_hw_queues(set, submit_queues);
 324         return set->nr_hw_queues == submit_queues ? 0 : -ENOMEM;
 325 }
 326
 327 NULLB_DEVICE_ATTR(size, ulong, NULL);
 328 NULLB_DEVICE_ATTR(completion_nsec, ulong, NULL);
 329 NULLB_DEVICE_ATTR(submit_queues, uint, nullb_apply_submit_queues);
 330 NULLB_DEVICE_ATTR(home_node, uint, NULL);
 331 NULLB_DEVICE_ATTR(queue_mode, uint, NULL);
 332 NULLB_DEVICE_ATTR(blocksize, uint, NULL);
 333 NULLB_DEVICE_ATTR(irqmode, uint, NULL);
 334 NULLB_DEVICE_ATTR(hw_queue_depth, uint, NULL);
 335 NULLB_DEVICE_ATTR(index, uint, NULL);
 336 NULLB_DEVICE_ATTR(blocking, bool, NULL);
 337 NULLB_DEVICE_ATTR(use_per_node_hctx, bool, NULL);
 338 NULLB_DEVICE_ATTR(memory_backed, bool, NULL);
 339 NULLB_DEVICE_ATTR(discard, bool, NULL);
 340 NULLB_DEVICE_ATTR(mbps, uint, NULL);
 341 NULLB_DEVICE_ATTR(cache_size, ulong, NULL);
 342 NULLB_DEVICE_ATTR(zoned, bool, NULL);
 343 NULLB_DEVICE_ATTR(zone_size, ulong, NULL);
 344 NULLB_DEVICE_ATTR(zone_nr_conv, uint, NULL);
 345
 346 static ssize_t nullb_device_power_show(struct config_item *item, char *page)
 347 {
 348         return nullb_device_bool_attr_show(to_nullb_device(item)->power, page);
 349 }
 350
 351 static ssize_t nullb_device_power_store(struct config_item *item,
 352                                      const char *page, size_t count)
 353 {
 354         struct nullb_device *dev = to_nullb_device(item);
 355         bool newp = false;
 356         ssize_t ret;
 357
 358         ret = nullb_device_bool_attr_store(&newp, page, count);
 359         if (ret < 0)
 360                 return ret;
 361
 362         if (!dev->power && newp) {
 363                 if (test_and_set_bit(NULLB_DEV_FL_UP, &dev->flags))
 364                         return count;
 365                 if (null_add_dev(dev)) {
 366                         clear_bit(NULLB_DEV_FL_UP, &dev->flags);
 367                         return -ENOMEM;
 368                 }
 369
 370                 set_bit(NULLB_DEV_FL_CONFIGURED, &dev->flags);
 371                 dev->power = newp;
 372         } else if (dev->power && !newp) {
 373                 if (test_and_clear_bit(NULLB_DEV_FL_UP, &dev->flags)) {
 374                         mutex_lock(&lock);
 375                         dev->power = newp;
 376                         null_del_dev(dev->nullb);
 377                         mutex_unlock(&lock);
 378                 }
 379                 clear_bit(NULLB_DEV_FL_CONFIGURED, &dev->flags);
 380         }
 381
 382         return count;
 383 }
 384
 385 CONFIGFS_ATTR(nullb_device_, power);
 386
 387 static ssize_t nullb_device_badblocks_show(struct config_item *item, char *page)
 388 {
 389         struct nullb_device *t_dev = to_nullb_device(item);
 390
 391         return badblocks_show(&t_dev->badblocks, page, 0);
 392 }
 393
 394 static ssize_t nullb_device_badblocks_store(struct config_item *item,
 395                                      const char *page, size_t count)
 396 {
 397         struct nullb_device *t_dev = to_nullb_device(item);
 398         char *orig, *buf, *tmp;
 399         u64 start, end;
 400         int ret;
 401
 402         orig = kstrndup(page, count, GFP_KERNEL);
 403         if (!orig)
 404                 return -ENOMEM;
 405
 406         buf = strstrip(orig);
 407
 408         ret = -EINVAL;
 409         if (buf[0] != '+' && buf[0] != '-')
 410                 goto out;
 411         tmp = strchr(&buf[1], '-');
 412         if (!tmp)
 413                 goto out;
 414         *tmp = '\0';
 415         ret = kstrtoull(buf + 1, 0, &start);
 416         if (ret)
 417                 goto out;
 418         ret = kstrtoull(tmp + 1, 0, &end);
 419         if (ret)
 420                 goto out;
 421         ret = -EINVAL;
 422         if (start > end)
 423                 goto out;
 424         /* enable badblocks */
 425         cmpxchg(&t_dev->badblocks.shift, -1, 0);
 426         if (buf[0] == '+')
 427                 ret = badblocks_set(&t_dev->badblocks, start,
 428                         end - start + 1, 1);
 429         else
 430                 ret = badblocks_clear(&t_dev->badblocks, start,
 431                         end - start + 1);
 432         if (ret == 0)
 433                 ret = count;
 434 out:
 435         kfree(orig);
 436         return ret;
 437 }
 438 CONFIGFS_ATTR(nullb_device_, badblocks);
 439
 440 static struct configfs_attribute *nullb_device_attrs[] = {
 441         &nullb_device_attr_size,
 442         &nullb_device_attr_completion_nsec,
 443         &nullb_device_attr_submit_queues,
 444         &nullb_device_attr_home_node,
 445         &nullb_device_attr_queue_mode,
 446         &nullb_device_attr_blocksize,
 447         &nullb_device_attr_irqmode,
 448         &nullb_device_attr_hw_queue_depth,
 449         &nullb_device_attr_index,
 450         &nullb_device_attr_blocking,
 451         &nullb_device_attr_use_per_node_hctx,
 452         &nullb_device_attr_power,
 453         &nullb_device_attr_memory_backed,
 454         &nullb_device_attr_discard,
 455         &nullb_device_attr_mbps,
 456         &nullb_device_attr_cache_size,
 457         &nullb_device_attr_badblocks,
 458         &nullb_device_attr_zoned,
 459         &nullb_device_attr_zone_size,
 460         &nullb_device_attr_zone_nr_conv,
 461         NULL,
 462 };
 463
 464 static void nullb_device_release(struct config_item *item)
 465 {
 466         struct nullb_device *dev = to_nullb_device(item);
 467
 468         null_free_device_storage(dev, false);
 469         null_free_dev(dev);
 470 }
 471
 472 static struct configfs_item_operations nullb_device_ops = {
 473         .release        = nullb_device_release,
 474 };
 475
 476 static const struct config_item_type nullb_device_type = {
 477         .ct_item_ops    = &nullb_device_ops,
 478         .ct_attrs       = nullb_device_attrs,
 479         .ct_owner       = THIS_MODULE,
 480 };
 481
 482 static struct
 483 config_item *nullb_group_make_item(struct config_group *group, const char *name)
 484 {
 485         struct nullb_device *dev;
 486
 487         dev = null_alloc_dev();
 488         if (!dev)
 489                 return ERR_PTR(-ENOMEM);
 490
 491         config_item_init_type_name(&dev->item, name, &nullb_device_type);
 492
 493         return &dev->item;
 494 }
 495
 496 static void
 497 nullb_group_drop_item(struct config_group *group, struct config_item *item)
 498 {
 499         struct nullb_device *dev = to_nullb_device(item);
 500
 501         if (test_and_clear_bit(NULLB_DEV_FL_UP, &dev->flags)) {
 502                 mutex_lock(&lock);
 503                 dev->power = false;
 504                 null_del_dev(dev->nullb);
 505                 mutex_unlock(&lock);
 506         }
 507
 508         config_item_put(item);
 509 }
 510
 511 static ssize_t memb_group_features_show(struct config_item *item, char *page)
 512 {
 513         return snprintf(page, PAGE_SIZE, "memory_backed,discard,bandwidth,cache,badblocks,zoned,zone_size,zone_nr_conv\n");
 514 }
 515
 516 CONFIGFS_ATTR_RO(memb_group_, features);
 517
 518 static struct configfs_attribute *nullb_group_attrs[] = {
 519         &memb_group_attr_features,
 520         NULL,
 521 };
 522
 523 static struct configfs_group_operations nullb_group_ops = {
 524         .make_item      = nullb_group_make_item,
 525         .drop_item      = nullb_group_drop_item,
 526 };
 527
 528 static const struct config_item_type nullb_group_type = {
 529         .ct_group_ops   = &nullb_group_ops,
 530         .ct_attrs       = nullb_group_attrs,
 531         .ct_owner       = THIS_MODULE,
 532 };
 533
 534 static struct configfs_subsystem nullb_subsys = {
 535         .su_group = {
 536                 .cg_item = {
 537                         .ci_namebuf = "nullb",
 538                         .ci_type = &nullb_group_type,
 539                 },
 540         },
 541 };
 542
 543 static inline int null_cache_active(struct nullb *nullb)
 544 {
 545         return test_bit(NULLB_DEV_FL_CACHE, &nullb->dev->flags);
 546 }
 547
 548 static struct nullb_device *null_alloc_dev(void)
 549 {
 550         struct nullb_device *dev;
 551
 552         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
 553         if (!dev)
 554                 return NULL;
 555         INIT_RADIX_TREE(&dev->data, GFP_ATOMIC);
 556         INIT_RADIX_TREE(&dev->cache, GFP_ATOMIC);
 557         if (badblocks_init(&dev->badblocks, 0)) {
 558                 kfree(dev);
 559                 return NULL;
 560         }
 561
 562         dev->size = g_gb * 1024;
 563         dev->completion_nsec = g_completion_nsec;
 564         dev->submit_queues = g_submit_queues;
 565         dev->home_node = g_home_node;
 566         dev->queue_mode = g_queue_mode;
 567         dev->blocksize = g_bs;
 568         dev->irqmode = g_irqmode;
 569         dev->hw_queue_depth = g_hw_queue_depth;
 570         dev->blocking = g_blocking;
 571         dev->use_per_node_hctx = g_use_per_node_hctx;
 572         dev->zoned = g_zoned;
 573         dev->zone_size = g_zone_size;
 574         dev->zone_nr_conv = g_zone_nr_conv;
 575         return dev;
 576 }
 577
 578 static void null_free_dev(struct nullb_device *dev)
 579 {
 580         if (!dev)
 581                 return;
 582
 583         null_free_zoned_dev(dev);
 584         badblocks_exit(&dev->badblocks);
 585         kfree(dev);
 586 }
 587
 588 static void put_tag(struct nullb_queue *nq, unsigned int tag)
 589 {
 590         clear_bit_unlock(tag, nq->tag_map);
 591
 592         if (waitqueue_active(&nq->wait))
 593                 wake_up(&nq->wait);
 594 }
 595
 596 static unsigned int get_tag(struct nullb_queue *nq)
 597 {
 598         unsigned int tag;
 599
 600         do {
 601                 tag = find_first_zero_bit(nq->tag_map, nq->queue_depth);
 602                 if (tag >= nq->queue_depth)
 603                         return -1U;
 604         } while (test_and_set_bit_lock(tag, nq->tag_map));
 605
 606         return tag;
 607 }
 608
 609 static void free_cmd(struct nullb_cmd *cmd)
 610 {
 611         put_tag(cmd->nq, cmd->tag);
 612 }
 613
 614 static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer);
 615
 616 static struct nullb_cmd *__alloc_cmd(struct nullb_queue *nq)
 617 {
 618         struct nullb_cmd *cmd;
 619         unsigned int tag;
 620
 621         tag = get_tag(nq);
 622         if (tag != -1U) {
 623                 cmd = &nq->cmds[tag];
 624                 cmd->tag = tag;
 625                 cmd->error = BLK_STS_OK;
 626                 cmd->nq = nq;
 627                 if (nq->dev->irqmode == NULL_IRQ_TIMER) {
 628                         hrtimer_init(&cmd->timer, CLOCK_MONOTONIC,
 629                                      HRTIMER_MODE_REL);
 630                         cmd->timer.function = null_cmd_timer_expired;
 631                 }
 632                 return cmd;
 633         }
 634
 635         return NULL;
 636 }
 637
 638 static struct nullb_cmd *alloc_cmd(struct nullb_queue *nq, int can_wait)
 639 {
 640         struct nullb_cmd *cmd;
 641         DEFINE_WAIT(wait);
 642
 643         cmd = __alloc_cmd(nq);
 644         if (cmd || !can_wait)
 645                 return cmd;
 646
 647         do {
 648                 prepare_to_wait(&nq->wait, &wait, TASK_UNINTERRUPTIBLE);
 649                 cmd = __alloc_cmd(nq);
 650                 if (cmd)
 651                         break;
 652
 653                 io_schedule();
 654         } while (1);
 655
 656         finish_wait(&nq->wait, &wait);
 657         return cmd;
 658 }
 659
 660 static void end_cmd(struct nullb_cmd *cmd)
 661 {
 662         int queue_mode = cmd->nq->dev->queue_mode;
 663
 664         switch (queue_mode)  {
 665         case NULL_Q_MQ:
 666                 blk_mq_end_request(cmd->rq, cmd->error);
 667                 return;
 668         case NULL_Q_BIO:
 669                 cmd->bio->bi_status = cmd->error;
 670                 bio_endio(cmd->bio);
 671                 break;
 672         }
 673
 674         free_cmd(cmd);
 675 }
 676
 677 static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer)
 678 {
 679         end_cmd(container_of(timer, struct nullb_cmd, timer));
 680
 681         return HRTIMER_NORESTART;
 682 }
 683
 684 static void null_cmd_end_timer(struct nullb_cmd *cmd)
 685 {
 686         ktime_t kt = cmd->nq->dev->completion_nsec;
 687
 688         hrtimer_start(&cmd->timer, kt, HRTIMER_MODE_REL);
 689 }
 690
 691 static void null_complete_rq(struct request *rq)
 692 {
 693         end_cmd(blk_mq_rq_to_pdu(rq));
 694 }
 695
 696 static struct nullb_page *null_alloc_page(gfp_t gfp_flags)
 697 {
 698         struct nullb_page *t_page;
 699
 700         t_page = kmalloc(sizeof(struct nullb_page), gfp_flags);
 701         if (!t_page)
 702                 goto out;
 703
 704         t_page->page = alloc_pages(gfp_flags, 0);
 705         if (!t_page->page)
 706                 goto out_freepage;
 707
 708         memset(t_page->bitmap, 0, sizeof(t_page->bitmap));
 709         return t_page;
 710 out_freepage:
 711         kfree(t_page);
 712 out:
 713         return NULL;
 714 }
 715
 716 static void null_free_page(struct nullb_page *t_page)
 717 {
 718         __set_bit(NULLB_PAGE_FREE, t_page->bitmap);
 719         if (test_bit(NULLB_PAGE_LOCK, t_page->bitmap))
 720                 return;
 721         __free_page(t_page->page);
 722         kfree(t_page);
 723 }
 724
 725 static bool null_page_empty(struct nullb_page *page)
 726 {
 727         int size = MAP_SZ - 2;
 728
 729         return find_first_bit(page->bitmap, size) == size;
 730 }
 731
 732 static void null_free_sector(struct nullb *nullb, sector_t sector,
 733         bool is_cache)
 734 {
 735         unsigned int sector_bit;
 736         u64 idx;
 737         struct nullb_page *t_page, *ret;
 738         struct radix_tree_root *root;
 739
 740         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 741         idx = sector >> PAGE_SECTORS_SHIFT;
 742         sector_bit = (sector & SECTOR_MASK);
 743
 744         t_page = radix_tree_lookup(root, idx);
 745         if (t_page) {
 746                 __clear_bit(sector_bit, t_page->bitmap);
 747
 748                 if (null_page_empty(t_page)) {
 749                         ret = radix_tree_delete_item(root, idx, t_page);
 750                         WARN_ON(ret != t_page);
 751                         null_free_page(ret);
 752                         if (is_cache)
 753                                 nullb->dev->curr_cache -= PAGE_SIZE;
 754                 }
 755         }
 756 }
 757
 758 static struct nullb_page *null_radix_tree_insert(struct nullb *nullb, u64 idx,
 759         struct nullb_page *t_page, bool is_cache)
 760 {
 761         struct radix_tree_root *root;
 762
 763         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 764
 765         if (radix_tree_insert(root, idx, t_page)) {
 766                 null_free_page(t_page);
 767                 t_page = radix_tree_lookup(root, idx);
 768                 WARN_ON(!t_page || t_page->page->index != idx);
 769         } else if (is_cache)
 770                 nullb->dev->curr_cache += PAGE_SIZE;
 771
 772         return t_page;
 773 }
 774
 775 static void null_free_device_storage(struct nullb_device *dev, bool is_cache)
 776 {
 777         unsigned long pos = 0;
 778         int nr_pages;
 779         struct nullb_page *ret, *t_pages[FREE_BATCH];
 780         struct radix_tree_root *root;
 781
 782         root = is_cache ? &dev->cache : &dev->data;
 783
 784         do {
 785                 int i;
 786
 787                 nr_pages = radix_tree_gang_lookup(root,
 788                                 (void **)t_pages, pos, FREE_BATCH);
 789
 790                 for (i = 0; i < nr_pages; i++) {
 791                         pos = t_pages[i]->page->index;
 792                         ret = radix_tree_delete_item(root, pos, t_pages[i]);
 793                         WARN_ON(ret != t_pages[i]);
 794                         null_free_page(ret);
 795                 }
 796
 797                 pos++;
 798         } while (nr_pages == FREE_BATCH);
 799
 800         if (is_cache)
 801                 dev->curr_cache = 0;
 802 }
 803
 804 static struct nullb_page *__null_lookup_page(struct nullb *nullb,
 805         sector_t sector, bool for_write, bool is_cache)
 806 {
 807         unsigned int sector_bit;
 808         u64 idx;
 809         struct nullb_page *t_page;
 810         struct radix_tree_root *root;
 811
 812         idx = sector >> PAGE_SECTORS_SHIFT;
 813         sector_bit = (sector & SECTOR_MASK);
 814
 815         root = is_cache ? &nullb->dev->cache : &nullb->dev->data;
 816         t_page = radix_tree_lookup(root, idx);
 817         WARN_ON(t_page && t_page->page->index != idx);
 818
 819         if (t_page && (for_write || test_bit(sector_bit, t_page->bitmap)))
 820                 return t_page;
 821
 822         return NULL;
 823 }
 824
 825 static struct nullb_page *null_lookup_page(struct nullb *nullb,
 826         sector_t sector, bool for_write, bool ignore_cache)
 827 {
 828         struct nullb_page *page = NULL;
 829
 830         if (!ignore_cache)
 831                 page = __null_lookup_page(nullb, sector, for_write, true);
 832         if (page)
 833                 return page;
 834         return __null_lookup_page(nullb, sector, for_write, false);
 835 }
 836
 837 static struct nullb_page *null_insert_page(struct nullb *nullb,
 838                                            sector_t sector, bool ignore_cache)
 839         __releases(&nullb->lock)
 840         __acquires(&nullb->lock)
 841 {
 842         u64 idx;
 843         struct nullb_page *t_page;
 844
 845         t_page = null_lookup_page(nullb, sector, true, ignore_cache);
 846         if (t_page)
 847                 return t_page;
 848
 849         spin_unlock_irq(&nullb->lock);
 850
 851         t_page = null_alloc_page(GFP_NOIO);
 852         if (!t_page)
 853                 goto out_lock;
 854
 855         if (radix_tree_preload(GFP_NOIO))
 856                 goto out_freepage;
 857
 858         spin_lock_irq(&nullb->lock);
 859         idx = sector >> PAGE_SECTORS_SHIFT;
 860         t_page->page->index = idx;
 861         t_page = null_radix_tree_insert(nullb, idx, t_page, !ignore_cache);
 862         radix_tree_preload_end();
 863
 864         return t_page;
 865 out_freepage:
 866         null_free_page(t_page);
 867 out_lock:
 868         spin_lock_irq(&nullb->lock);
 869         return null_lookup_page(nullb, sector, true, ignore_cache);
 870 }
 871
 872 static int null_flush_cache_page(struct nullb *nullb, struct nullb_page *c_page)
 873 {
 874         int i;
 875         unsigned int offset;
 876         u64 idx;
 877         struct nullb_page *t_page, *ret;
 878         void *dst, *src;
 879
 880         idx = c_page->page->index;
 881
 882         t_page = null_insert_page(nullb, idx << PAGE_SECTORS_SHIFT, true);
 883
 884         __clear_bit(NULLB_PAGE_LOCK, c_page->bitmap);
 885         if (test_bit(NULLB_PAGE_FREE, c_page->bitmap)) {
 886                 null_free_page(c_page);
 887                 if (t_page && null_page_empty(t_page)) {
 888                         ret = radix_tree_delete_item(&nullb->dev->data,
 889                                 idx, t_page);
 890                         null_free_page(t_page);
 891                 }
 892                 return 0;
 893         }
 894
 895         if (!t_page)
 896                 return -ENOMEM;
 897
 898         src = kmap_atomic(c_page->page);
 899         dst = kmap_atomic(t_page->page);
 900
 901         for (i = 0; i < PAGE_SECTORS;
 902                         i += (nullb->dev->blocksize >> SECTOR_SHIFT)) {
 903                 if (test_bit(i, c_page->bitmap)) {
 904                         offset = (i << SECTOR_SHIFT);
 905                         memcpy(dst + offset, src + offset,
 906                                 nullb->dev->blocksize);
 907                         __set_bit(i, t_page->bitmap);
 908                 }
 909         }
 910
 911         kunmap_atomic(dst);
 912         kunmap_atomic(src);
 913
 914         ret = radix_tree_delete_item(&nullb->dev->cache, idx, c_page);
 915         null_free_page(ret);
 916         nullb->dev->curr_cache -= PAGE_SIZE;
 917
 918         return 0;
 919 }
 920
 921 static int null_make_cache_space(struct nullb *nullb, unsigned long n)
 922 {
 923         int i, err, nr_pages;
 924         struct nullb_page *c_pages[FREE_BATCH];
 925         unsigned long flushed = 0, one_round;
 926
 927 again:
 928         if ((nullb->dev->cache_size * 1024 * 1024) >
 929              nullb->dev->curr_cache + n || nullb->dev->curr_cache == 0)
 930                 return 0;
 931
 932         nr_pages = radix_tree_gang_lookup(&nullb->dev->cache,
 933                         (void **)c_pages, nullb->cache_flush_pos, FREE_BATCH);
 934         /*
 935          * nullb_flush_cache_page could unlock before using the c_pages. To
 936          * avoid race, we don't allow page free
 937          */
 938         for (i = 0; i < nr_pages; i++) {
 939                 nullb->cache_flush_pos = c_pages[i]->page->index;
 940                 /*
 941                  * We found the page which is being flushed to disk by other
 942                  * threads
 943                  */
 944                 if (test_bit(NULLB_PAGE_LOCK, c_pages[i]->bitmap))
 945                         c_pages[i] = NULL;
 946                 else
 947                         __set_bit(NULLB_PAGE_LOCK, c_pages[i]->bitmap);
 948         }
 949
 950         one_round = 0;
 951         for (i = 0; i < nr_pages; i++) {
 952                 if (c_pages[i] == NULL)
 953                         continue;
 954                 err = null_flush_cache_page(nullb, c_pages[i]);
 955                 if (err)
 956                         return err;
 957                 one_round++;
 958         }
 959         flushed += one_round << PAGE_SHIFT;
 960
 961         if (n > flushed) {
 962                 if (nr_pages == 0)
 963                         nullb->cache_flush_pos = 0;
 964                 if (one_round == 0) {
 965                         /* give other threads a chance */
 966                         spin_unlock_irq(&nullb->lock);
 967                         spin_lock_irq(&nullb->lock);
 968                 }
 969                 goto again;
 970         }
 971         return 0;
 972 }
 973
 974 static int copy_to_nullb(struct nullb *nullb, struct page *source,
 975         unsigned int off, sector_t sector, size_t n, bool is_fua)
 976 {
 977         size_t temp, count = 0;
 978         unsigned int offset;
 979         struct nullb_page *t_page;
 980         void *dst, *src;
 981
 982         while (count < n) {
 983                 temp = min_t(size_t, nullb->dev->blocksize, n - count);
 984
 985                 if (null_cache_active(nullb) && !is_fua)
 986                         null_make_cache_space(nullb, PAGE_SIZE);
 987
 988                 offset = (sector & SECTOR_MASK) << SECTOR_SHIFT;
 989                 t_page = null_insert_page(nullb, sector,
 990                         !null_cache_active(nullb) || is_fua);
 991                 if (!t_page)
 992                         return -ENOSPC;
 993
 994                 src = kmap_atomic(source);
 995                 dst = kmap_atomic(t_page->page);
 996                 memcpy(dst + offset, src + off + count, temp);
 997                 kunmap_atomic(dst);
 998                 kunmap_atomic(src);
 999
1000                 __set_bit(sector & SECTOR_MASK, t_page->bitmap);
1001
1002                 if (is_fua)
1003                         null_free_sector(nullb, sector, true);
1004
1005                 count += temp;
1006                 sector += temp >> SECTOR_SHIFT;
1007         }
1008         return 0;
1009 }
1010
1011 static int copy_from_nullb(struct nullb *nullb, struct page *dest,
1012         unsigned int off, sector_t sector, size_t n)
1013 {
1014         size_t temp, count = 0;
1015         unsigned int offset;
1016         struct nullb_page *t_page;
1017         void *dst, *src;
1018
1019         while (count < n) {
1020                 temp = min_t(size_t, nullb->dev->blocksize, n - count);
1021
1022                 offset = (sector & SECTOR_MASK) << SECTOR_SHIFT;
1023                 t_page = null_lookup_page(nullb, sector, false,
1024                         !null_cache_active(nullb));
1025
1026                 dst = kmap_atomic(dest);
1027                 if (!t_page) {
1028                         memset(dst + off + count, 0, temp);
1029                         goto next;
1030                 }
1031                 src = kmap_atomic(t_page->page);
1032                 memcpy(dst + off + count, src + offset, temp);
1033                 kunmap_atomic(src);
1034 next:
1035                 kunmap_atomic(dst);
1036
1037                 count += temp;
1038                 sector += temp >> SECTOR_SHIFT;
1039         }
1040         return 0;
1041 }
1042
1043 static void nullb_fill_pattern(struct nullb *nullb, struct page *page,
1044                                unsigned int len, unsigned int off)
1045 {
1046         void *dst;
1047
1048         dst = kmap_atomic(page);
1049         memset(dst + off, 0xFF, len);
1050         kunmap_atomic(dst);
1051 }
1052
1053 static void null_handle_discard(struct nullb *nullb, sector_t sector, size_t n)
1054 {
1055         size_t temp;
1056
1057         spin_lock_irq(&nullb->lock);
1058         while (n > 0) {
1059                 temp = min_t(size_t, n, nullb->dev->blocksize);
1060                 null_free_sector(nullb, sector, false);
1061                 if (null_cache_active(nullb))
1062                         null_free_sector(nullb, sector, true);
1063                 sector += temp >> SECTOR_SHIFT;
1064                 n -= temp;
1065         }
1066         spin_unlock_irq(&nullb->lock);
1067 }
1068
1069 static int null_handle_flush(struct nullb *nullb)
1070 {
1071         int err;
1072
1073         if (!null_cache_active(nullb))
1074                 return 0;
1075
1076         spin_lock_irq(&nullb->lock);
1077         while (true) {
1078                 err = null_make_cache_space(nullb,
1079                         nullb->dev->cache_size * 1024 * 1024);
1080                 if (err || nullb->dev->curr_cache == 0)
1081                         break;
1082         }
1083
1084         WARN_ON(!radix_tree_empty(&nullb->dev->cache));
1085         spin_unlock_irq(&nullb->lock);
1086         return err;
1087 }
1088
1089 static int null_transfer(struct nullb *nullb, struct page *page,
1090         unsigned int len, unsigned int off, bool is_write, sector_t sector,
1091         bool is_fua)
1092 {
1093         struct nullb_device *dev = nullb->dev;
1094         unsigned int valid_len = len;
1095         int err = 0;
1096
1097         if (!is_write) {
1098                 if (dev->zoned)
1099                         valid_len = null_zone_valid_read_len(nullb,
1100                                 sector, len);
1101
1102                 if (valid_len) {
1103                         err = copy_from_nullb(nullb, page, off,
1104                                 sector, valid_len);
1105                         off += valid_len;
1106                         len -= valid_len;
1107                 }
1108
1109                 if (len)
1110                         nullb_fill_pattern(nullb, page, len, off);
1111                 flush_dcache_page(page);
1112         } else {
1113                 flush_dcache_page(page);
1114                 err = copy_to_nullb(nullb, page, off, sector, len, is_fua);
1115         }
1116
1117         return err;
1118 }
1119
1120 static int null_handle_rq(struct nullb_cmd *cmd)
1121 {
1122         struct request *rq = cmd->rq;
1123         struct nullb *nullb = cmd->nq->dev->nullb;
1124         int err;
1125         unsigned int len;
1126         sector_t sector;
1127         struct req_iterator iter;
1128         struct bio_vec bvec;
1129
1130         sector = blk_rq_pos(rq);
1131
1132         if (req_op(rq) == REQ_OP_DISCARD) {
1133                 null_handle_discard(nullb, sector, blk_rq_bytes(rq));
1134                 return 0;
1135         }
1136
1137         spin_lock_irq(&nullb->lock);
1138         rq_for_each_segment(bvec, rq, iter) {
1139                 len = bvec.bv_len;
1140                 err = null_transfer(nullb, bvec.bv_page, len, bvec.bv_offset,
1141                                      op_is_write(req_op(rq)), sector,
1142                                      req_op(rq) & REQ_FUA);
1143                 if (err) {
1144                         spin_unlock_irq(&nullb->lock);
1145                         return err;
1146                 }
1147                 sector += len >> SECTOR_SHIFT;
1148         }
1149         spin_unlock_irq(&nullb->lock);
1150
1151         return 0;
1152 }
1153
1154 static int null_handle_bio(struct nullb_cmd *cmd)
1155 {
1156         struct bio *bio = cmd->bio;
1157         struct nullb *nullb = cmd->nq->dev->nullb;
1158         int err;
1159         unsigned int len;
1160         sector_t sector;
1161         struct bio_vec bvec;
1162         struct bvec_iter iter;
1163
1164         sector = bio->bi_iter.bi_sector;
1165
1166         if (bio_op(bio) == REQ_OP_DISCARD) {
1167                 null_handle_discard(nullb, sector,
1168                         bio_sectors(bio) << SECTOR_SHIFT);
1169                 return 0;
1170         }
1171
1172         spin_lock_irq(&nullb->lock);
1173         bio_for_each_segment(bvec, bio, iter) {
1174                 len = bvec.bv_len;
1175                 err = null_transfer(nullb, bvec.bv_page, len, bvec.bv_offset,
1176                                      op_is_write(bio_op(bio)), sector,
1177                                      bio->bi_opf & REQ_FUA);
1178                 if (err) {
1179                         spin_unlock_irq(&nullb->lock);
1180                         return err;
1181                 }
1182                 sector += len >> SECTOR_SHIFT;
1183         }
1184         spin_unlock_irq(&nullb->lock);
1185         return 0;
1186 }
1187
1188 static void null_stop_queue(struct nullb *nullb)
1189 {
1190         struct request_queue *q = nullb->q;
1191
1192         if (nullb->dev->queue_mode == NULL_Q_MQ)
1193                 blk_mq_stop_hw_queues(q);
1194 }
1195
1196 static void null_restart_queue_async(struct nullb *nullb)
1197 {
1198         struct request_queue *q = nullb->q;
1199
1200         if (nullb->dev->queue_mode == NULL_Q_MQ)
1201                 blk_mq_start_stopped_hw_queues(q, true);
1202 }
1203
1204 static inline blk_status_t null_handle_throttled(struct nullb_cmd *cmd)
1205 {
1206         struct nullb_device *dev = cmd->nq->dev;
1207         struct nullb *nullb = dev->nullb;
1208         blk_status_t sts = BLK_STS_OK;
1209         struct request *rq = cmd->rq;
1210
1211         if (!hrtimer_active(&nullb->bw_timer))
1212                 hrtimer_restart(&nullb->bw_timer);
1213
1214         if (atomic_long_sub_return(blk_rq_bytes(rq), &nullb->cur_bytes) < 0) {
1215                 null_stop_queue(nullb);
1216                 /* race with timer */
1217                 if (atomic_long_read(&nullb->cur_bytes) > 0)
1218                         null_restart_queue_async(nullb);
1219                 /* requeue request */
1220                 sts = BLK_STS_DEV_RESOURCE;
1221         }
1222         return sts;
1223 }
1224
1225 static inline blk_status_t null_handle_badblocks(struct nullb_cmd *cmd,
1226                                                  sector_t sector,
1227                                                  sector_t nr_sectors)
1228 {
1229         struct badblocks *bb = &cmd->nq->dev->badblocks;
1230         sector_t first_bad;
1231         int bad_sectors;
1232
1233         if (badblocks_check(bb, sector, nr_sectors, &first_bad, &bad_sectors))
1234                 return BLK_STS_IOERR;
1235
1236         return BLK_STS_OK;
1237 }
1238
1239 static inline blk_status_t null_handle_memory_backed(struct nullb_cmd *cmd,
1240                                                      enum req_opf op)
1241 {
1242         struct nullb_device *dev = cmd->nq->dev;
1243         int err;
1244
1245         if (dev->queue_mode == NULL_Q_BIO)
1246                 err = null_handle_bio(cmd);
1247         else
1248                 err = null_handle_rq(cmd);
1249
1250         return errno_to_blk_status(err);
1251 }
1252
1253 static inline void nullb_complete_cmd(struct nullb_cmd *cmd)
1254 {
1255         /* Complete IO by inline, softirq or timer */
1256         switch (cmd->nq->dev->irqmode) {
1257         case NULL_IRQ_SOFTIRQ:
1258                 switch (cmd->nq->dev->queue_mode) {
1259                 case NULL_Q_MQ:
1260                         blk_mq_complete_request(cmd->rq);
1261                         break;
1262                 case NULL_Q_BIO:
1263                         /*
1264                          * XXX: no proper submitting cpu information available.
1265                          */
1266                         end_cmd(cmd);
1267                         break;
1268                 }
1269                 break;
1270         case NULL_IRQ_NONE:
1271                 end_cmd(cmd);
1272                 break;
1273         case NULL_IRQ_TIMER:
1274                 null_cmd_end_timer(cmd);
1275                 break;
1276         }
1277 }
1278
1279 blk_status_t null_process_cmd(struct nullb_cmd *cmd,
1280                               enum req_opf op, sector_t sector,
1281                               unsigned int nr_sectors)
1282 {
1283         struct nullb_device *dev = cmd->nq->dev;
1284         blk_status_t ret;
1285
1286         if (dev->badblocks.shift != -1) {
1287                 ret = null_handle_badblocks(cmd, sector, nr_sectors);
1288                 if (ret != BLK_STS_OK)
1289                         return ret;
1290         }
1291
1292         if (dev->memory_backed)
1293                 return null_handle_memory_backed(cmd, op);
1294
1295         return BLK_STS_OK;
1296 }
1297
1298 static blk_status_t null_handle_cmd(struct nullb_cmd *cmd, sector_t sector,
1299                                     sector_t nr_sectors, enum req_opf op)
1300 {
1301         struct nullb_device *dev = cmd->nq->dev;
1302         struct nullb *nullb = dev->nullb;
1303         blk_status_t sts;
1304
1305         if (test_bit(NULLB_DEV_FL_THROTTLED, &dev->flags)) {
1306                 sts = null_handle_throttled(cmd);
1307                 if (sts != BLK_STS_OK)
1308                         return sts;
1309         }
1310
1311         if (op == REQ_OP_FLUSH) {
1312                 cmd->error = errno_to_blk_status(null_handle_flush(nullb));
1313                 goto out;
1314         }
1315
1316         if (dev->zoned)
1317                 cmd->error = null_process_zoned_cmd(cmd, op,
1318                                                     sector, nr_sectors);
1319         else
1320                 cmd->error = null_process_cmd(cmd, op, sector, nr_sectors);
1321
1322 out:
1323         nullb_complete_cmd(cmd);
1324         return BLK_STS_OK;
1325 }
1326
1327 static enum hrtimer_restart nullb_bwtimer_fn(struct hrtimer *timer)
1328 {
1329         struct nullb *nullb = container_of(timer, struct nullb, bw_timer);
1330         ktime_t timer_interval = ktime_set(0, TIMER_INTERVAL);
1331         unsigned int mbps = nullb->dev->mbps;
1332
1333         if (atomic_long_read(&nullb->cur_bytes) == mb_per_tick(mbps))
1334                 return HRTIMER_NORESTART;
1335
1336         atomic_long_set(&nullb->cur_bytes, mb_per_tick(mbps));
1337         null_restart_queue_async(nullb);
1338
1339         hrtimer_forward_now(&nullb->bw_timer, timer_interval);
1340
1341         return HRTIMER_RESTART;
1342 }
1343
1344 static void nullb_setup_bwtimer(struct nullb *nullb)
1345 {
1346         ktime_t timer_interval = ktime_set(0, TIMER_INTERVAL);
1347
1348         hrtimer_init(&nullb->bw_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1349         nullb->bw_timer.function = nullb_bwtimer_fn;
1350         atomic_long_set(&nullb->cur_bytes, mb_per_tick(nullb->dev->mbps));
1351         hrtimer_start(&nullb->bw_timer, timer_interval, HRTIMER_MODE_REL);
1352 }
1353
1354 static struct nullb_queue *nullb_to_queue(struct nullb *nullb)
1355 {
1356         int index = 0;
1357
1358         if (nullb->nr_queues != 1)
1359                 index = raw_smp_processor_id() / ((nr_cpu_ids + nullb->nr_queues - 1) / nullb->nr_queues);
1360
1361         return &nullb->queues[index];
1362 }
1363
1364 static blk_qc_t null_queue_bio(struct request_queue *q, struct bio *bio)
1365 {
1366         sector_t sector = bio->bi_iter.bi_sector;
1367         sector_t nr_sectors = bio_sectors(bio);
1368         struct nullb *nullb = q->queuedata;
1369         struct nullb_queue *nq = nullb_to_queue(nullb);
1370         struct nullb_cmd *cmd;
1371
1372         cmd = alloc_cmd(nq, 1);
1373         cmd->bio = bio;
1374
1375         null_handle_cmd(cmd, sector, nr_sectors, bio_op(bio));
1376         return BLK_QC_T_NONE;
1377 }
1378
1379 static bool should_timeout_request(struct request *rq)
1380 {
1381 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1382         if (g_timeout_str[0])
1383                 return should_fail(&null_timeout_attr, 1);
1384 #endif
1385         return false;
1386 }
1387
1388 static bool should_requeue_request(struct request *rq)
1389 {
1390 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1391         if (g_requeue_str[0])
1392                 return should_fail(&null_requeue_attr, 1);
1393 #endif
1394         return false;
1395 }
1396
1397 static enum blk_eh_timer_return null_timeout_rq(struct request *rq, bool res)
1398 {
1399         pr_info("rq %p timed out\n", rq);
1400         blk_mq_complete_request(rq);
1401         return BLK_EH_DONE;
1402 }
1403
1404 static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx,
1405                          const struct blk_mq_queue_data *bd)
1406 {
1407         struct nullb_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1408         struct nullb_queue *nq = hctx->driver_data;
1409         sector_t nr_sectors = blk_rq_sectors(bd->rq);
1410         sector_t sector = blk_rq_pos(bd->rq);
1411
1412         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1413
1414         if (nq->dev->irqmode == NULL_IRQ_TIMER) {
1415                 hrtimer_init(&cmd->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1416                 cmd->timer.function = null_cmd_timer_expired;
1417         }
1418         cmd->rq = bd->rq;
1419         cmd->error = BLK_STS_OK;
1420         cmd->nq = nq;
1421
1422         blk_mq_start_request(bd->rq);
1423
1424         if (should_requeue_request(bd->rq)) {
1425                 /*
1426                  * Alternate between hitting the core BUSY path, and the
1427                  * driver driven requeue path
1428                  */
1429                 nq->requeue_selection++;
1430                 if (nq->requeue_selection & 1)
1431                         return BLK_STS_RESOURCE;
1432                 else {
1433                         blk_mq_requeue_request(bd->rq, true);
1434                         return BLK_STS_OK;
1435                 }
1436         }
1437         if (should_timeout_request(bd->rq))
1438                 return BLK_STS_OK;
1439
1440         return null_handle_cmd(cmd, sector, nr_sectors, req_op(bd->rq));
1441 }
1442
1443 static void cleanup_queue(struct nullb_queue *nq)
1444 {
1445         kfree(nq->tag_map);
1446         kfree(nq->cmds);
1447 }
1448
1449 static void cleanup_queues(struct nullb *nullb)
1450 {
1451         int i;
1452
1453         for (i = 0; i < nullb->nr_queues; i++)
1454                 cleanup_queue(&nullb->queues[i]);
1455
1456         kfree(nullb->queues);
1457 }
1458
1459 static void null_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1460 {
1461         struct nullb_queue *nq = hctx->driver_data;
1462         struct nullb *nullb = nq->dev->nullb;
1463
1464         nullb->nr_queues--;
1465 }
1466
1467 static void null_init_queue(struct nullb *nullb, struct nullb_queue *nq)
1468 {
1469         init_waitqueue_head(&nq->wait);
1470         nq->queue_depth = nullb->queue_depth;
1471         nq->dev = nullb->dev;
1472 }
1473
1474 static int null_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
1475                           unsigned int hctx_idx)
1476 {
1477         struct nullb *nullb = hctx->queue->queuedata;
1478         struct nullb_queue *nq;
1479
1480 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1481         if (g_init_hctx_str[0] && should_fail(&null_init_hctx_attr, 1))
1482                 return -EFAULT;
1483 #endif
1484
1485         nq = &nullb->queues[hctx_idx];
1486         hctx->driver_data = nq;
1487         null_init_queue(nullb, nq);
1488         nullb->nr_queues++;
1489
1490         return 0;
1491 }
1492
1493 static const struct blk_mq_ops null_mq_ops = {
1494         .queue_rq       = null_queue_rq,
1495         .complete       = null_complete_rq,
1496         .timeout        = null_timeout_rq,
1497         .init_hctx      = null_init_hctx,
1498         .exit_hctx      = null_exit_hctx,
1499 };
1500
1501 static void null_del_dev(struct nullb *nullb)
1502 {
1503         struct nullb_device *dev;
1504
1505         if (!nullb)
1506                 return;
1507
1508         dev = nullb->dev;
1509
1510         ida_simple_remove(&nullb_indexes, nullb->index);
1511
1512         list_del_init(&nullb->list);
1513
1514         del_gendisk(nullb->disk);
1515
1516         if (test_bit(NULLB_DEV_FL_THROTTLED, &nullb->dev->flags)) {
1517                 hrtimer_cancel(&nullb->bw_timer);
1518                 atomic_long_set(&nullb->cur_bytes, LONG_MAX);
1519                 null_restart_queue_async(nullb);
1520         }
1521
1522         blk_cleanup_queue(nullb->q);
1523         if (dev->queue_mode == NULL_Q_MQ &&
1524             nullb->tag_set == &nullb->__tag_set)
1525                 blk_mq_free_tag_set(nullb->tag_set);
1526         put_disk(nullb->disk);
1527         cleanup_queues(nullb);
1528         if (null_cache_active(nullb))
1529                 null_free_device_storage(nullb->dev, true);
1530         kfree(nullb);
1531         dev->nullb = NULL;
1532 }
1533
1534 static void null_config_discard(struct nullb *nullb)
1535 {
1536         if (nullb->dev->discard == false)
1537                 return;
1538
1539         if (nullb->dev->zoned) {
1540                 nullb->dev->discard = false;
1541                 pr_info("discard option is ignored in zoned mode\n");
1542                 return;
1543         }
1544
1545         nullb->q->limits.discard_granularity = nullb->dev->blocksize;
1546         nullb->q->limits.discard_alignment = nullb->dev->blocksize;
1547         blk_queue_max_discard_sectors(nullb->q, UINT_MAX >> 9);
1548         blk_queue_flag_set(QUEUE_FLAG_DISCARD, nullb->q);
1549 }
1550
1551 static const struct block_device_operations null_ops = {
1552         .owner          = THIS_MODULE,
1553         .report_zones   = null_report_zones,
1554 };
1555
1556 static int setup_commands(struct nullb_queue *nq)
1557 {
1558         struct nullb_cmd *cmd;
1559         int i, tag_size;
1560
1561         nq->cmds = kcalloc(nq->queue_depth, sizeof(*cmd), GFP_KERNEL);
1562         if (!nq->cmds)
1563                 return -ENOMEM;
1564
1565         tag_size = ALIGN(nq->queue_depth, BITS_PER_LONG) / BITS_PER_LONG;
1566         nq->tag_map = kcalloc(tag_size, sizeof(unsigned long), GFP_KERNEL);
1567         if (!nq->tag_map) {
1568                 kfree(nq->cmds);
1569                 return -ENOMEM;
1570         }
1571
1572         for (i = 0; i < nq->queue_depth; i++) {
1573                 cmd = &nq->cmds[i];
1574                 cmd->tag = -1U;
1575         }
1576
1577         return 0;
1578 }
1579
1580 static int setup_queues(struct nullb *nullb)
1581 {
1582         nullb->queues = kcalloc(nr_cpu_ids, sizeof(struct nullb_queue),
1583                                 GFP_KERNEL);
1584         if (!nullb->queues)
1585                 return -ENOMEM;
1586
1587         nullb->queue_depth = nullb->dev->hw_queue_depth;
1588
1589         return 0;
1590 }
1591
1592 static int init_driver_queues(struct nullb *nullb)
1593 {
1594         struct nullb_queue *nq;
1595         int i, ret = 0;
1596
1597         for (i = 0; i < nullb->dev->submit_queues; i++) {
1598                 nq = &nullb->queues[i];
1599
1600                 null_init_queue(nullb, nq);
1601
1602                 ret = setup_commands(nq);
1603                 if (ret)
1604                         return ret;
1605                 nullb->nr_queues++;
1606         }
1607         return 0;
1608 }
1609
1610 static int null_gendisk_register(struct nullb *nullb)
1611 {
1612         sector_t size = ((sector_t)nullb->dev->size * SZ_1M) >> SECTOR_SHIFT;
1613         struct gendisk *disk;
1614
1615         disk = nullb->disk = alloc_disk_node(1, nullb->dev->home_node);
1616         if (!disk)
1617                 return -ENOMEM;
1618         set_capacity(disk, size);
1619
1620         disk->flags |= GENHD_FL_EXT_DEVT | GENHD_FL_SUPPRESS_PARTITION_INFO;
1621         disk->major             = null_major;
1622         disk->first_minor       = nullb->index;
1623         disk->fops              = &null_ops;
1624         disk->private_data      = nullb;
1625         disk->queue             = nullb->q;
1626         strncpy(disk->disk_name, nullb->disk_name, DISK_NAME_LEN);
1627
1628         if (nullb->dev->zoned) {
1629                 int ret = null_register_zoned_dev(nullb);
1630
1631                 if (ret)
1632                         return ret;
1633         }
1634
1635         add_disk(disk);
1636         return 0;
1637 }
1638
1639 static int null_init_tag_set(struct nullb *nullb, struct blk_mq_tag_set *set)
1640 {
1641         set->ops = &null_mq_ops;
1642         set->nr_hw_queues = nullb ? nullb->dev->submit_queues :
1643                                                 g_submit_queues;
1644         set->queue_depth = nullb ? nullb->dev->hw_queue_depth :
1645                                                 g_hw_queue_depth;
1646         set->numa_node = nullb ? nullb->dev->home_node : g_home_node;
1647         set->cmd_size   = sizeof(struct nullb_cmd);
1648         set->flags = BLK_MQ_F_SHOULD_MERGE;
1649         if (g_no_sched)
1650                 set->flags |= BLK_MQ_F_NO_SCHED;
1651         set->driver_data = NULL;
1652
1653         if ((nullb && nullb->dev->blocking) || g_blocking)
1654                 set->flags |= BLK_MQ_F_BLOCKING;
1655
1656         return blk_mq_alloc_tag_set(set);
1657 }
1658
1659 static int null_validate_conf(struct nullb_device *dev)
1660 {
1661         dev->blocksize = round_down(dev->blocksize, 512);
1662         dev->blocksize = clamp_t(unsigned int, dev->blocksize, 512, 4096);
1663
1664         if (dev->queue_mode == NULL_Q_MQ && dev->use_per_node_hctx) {
1665                 if (dev->submit_queues != nr_online_nodes)
1666                         dev->submit_queues = nr_online_nodes;
1667         } else if (dev->submit_queues > nr_cpu_ids)
1668                 dev->submit_queues = nr_cpu_ids;
1669         else if (dev->submit_queues == 0)
1670                 dev->submit_queues = 1;
1671
1672         dev->queue_mode = min_t(unsigned int, dev->queue_mode, NULL_Q_MQ);
1673         dev->irqmode = min_t(unsigned int, dev->irqmode, NULL_IRQ_TIMER);
1674
1675         /* Do memory allocation, so set blocking */
1676         if (dev->memory_backed)
1677                 dev->blocking = true;
1678         else /* cache is meaningless */
1679                 dev->cache_size = 0;
1680         dev->cache_size = min_t(unsigned long, ULONG_MAX / 1024 / 1024,
1681                                                 dev->cache_size);
1682         dev->mbps = min_t(unsigned int, 1024 * 40, dev->mbps);
1683         /* can not stop a queue */
1684         if (dev->queue_mode == NULL_Q_BIO)
1685                 dev->mbps = 0;
1686
1687         if (dev->zoned &&
1688             (!dev->zone_size || !is_power_of_2(dev->zone_size))) {
1689                 pr_err("zone_size must be power-of-two\n");
1690                 return -EINVAL;
1691         }
1692
1693         return 0;
1694 }
1695
1696 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1697 static bool __null_setup_fault(struct fault_attr *attr, char *str)
1698 {
1699         if (!str[0])
1700                 return true;
1701
1702         if (!setup_fault_attr(attr, str))
1703                 return false;
1704
1705         attr->verbose = 0;
1706         return true;
1707 }
1708 #endif
1709
1710 static bool null_setup_fault(void)
1711 {
1712 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
1713         if (!__null_setup_fault(&null_timeout_attr, g_timeout_str))
1714                 return false;
1715         if (!__null_setup_fault(&null_requeue_attr, g_requeue_str))
1716                 return false;
1717         if (!__null_setup_fault(&null_init_hctx_attr, g_init_hctx_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         rv = null_validate_conf(dev);
1729         if (rv)
1730                 return rv;
1731
1732         nullb = kzalloc_node(sizeof(*nullb), GFP_KERNEL, dev->home_node);
1733         if (!nullb) {
1734                 rv = -ENOMEM;
1735                 goto out;
1736         }
1737         nullb->dev = dev;
1738         dev->nullb = nullb;
1739
1740         spin_lock_init(&nullb->lock);
1741
1742         rv = setup_queues(nullb);
1743         if (rv)
1744                 goto out_free_nullb;
1745
1746         if (dev->queue_mode == NULL_Q_MQ) {
1747                 if (shared_tags) {
1748                         nullb->tag_set = &tag_set;
1749                         rv = 0;
1750                 } else {
1751                         nullb->tag_set = &nullb->__tag_set;
1752                         rv = null_init_tag_set(nullb, nullb->tag_set);
1753                 }
1754
1755                 if (rv)
1756                         goto out_cleanup_queues;
1757
1758                 if (!null_setup_fault())
1759                         goto out_cleanup_queues;
1760
1761                 nullb->tag_set->timeout = 5 * HZ;
1762                 nullb->q = blk_mq_init_queue_data(nullb->tag_set, nullb);
1763                 if (IS_ERR(nullb->q)) {
1764                         rv = -ENOMEM;
1765                         goto out_cleanup_tags;
1766                 }
1767         } else if (dev->queue_mode == NULL_Q_BIO) {
1768                 nullb->q = blk_alloc_queue(null_queue_bio, dev->home_node);
1769                 if (!nullb->q) {
1770                         rv = -ENOMEM;
1771                         goto out_cleanup_queues;
1772                 }
1773                 rv = init_driver_queues(nullb);
1774                 if (rv)
1775                         goto out_cleanup_blk_queue;
1776         }
1777
1778         if (dev->mbps) {
1779                 set_bit(NULLB_DEV_FL_THROTTLED, &dev->flags);
1780                 nullb_setup_bwtimer(nullb);
1781         }
1782
1783         if (dev->cache_size > 0) {
1784                 set_bit(NULLB_DEV_FL_CACHE, &nullb->dev->flags);
1785                 blk_queue_write_cache(nullb->q, true, true);
1786         }
1787
1788         if (dev->zoned) {
1789                 rv = null_init_zoned_dev(dev, nullb->q);
1790                 if (rv)
1791                         goto out_cleanup_blk_queue;
1792         }
1793
1794         nullb->q->queuedata = nullb;
1795         blk_queue_flag_set(QUEUE_FLAG_NONROT, nullb->q);
1796         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, nullb->q);
1797
1798         mutex_lock(&lock);
1799         nullb->index = ida_simple_get(&nullb_indexes, 0, 0, GFP_KERNEL);
1800         dev->index = nullb->index;
1801         mutex_unlock(&lock);
1802
1803         blk_queue_logical_block_size(nullb->q, dev->blocksize);
1804         blk_queue_physical_block_size(nullb->q, dev->blocksize);
1805
1806         null_config_discard(nullb);
1807
1808         sprintf(nullb->disk_name, "nullb%d", nullb->index);
1809
1810         rv = null_gendisk_register(nullb);
1811         if (rv)
1812                 goto out_cleanup_zone;
1813
1814         mutex_lock(&lock);
1815         list_add_tail(&nullb->list, &nullb_list);
1816         mutex_unlock(&lock);
1817
1818         return 0;
1819 out_cleanup_zone:
1820         null_free_zoned_dev(dev);
1821 out_cleanup_blk_queue:
1822         blk_cleanup_queue(nullb->q);
1823 out_cleanup_tags:
1824         if (dev->queue_mode == NULL_Q_MQ && nullb->tag_set == &nullb->__tag_set)
1825                 blk_mq_free_tag_set(nullb->tag_set);
1826 out_cleanup_queues:
1827         cleanup_queues(nullb);
1828 out_free_nullb:
1829         kfree(nullb);
1830         dev->nullb = NULL;
1831 out:
1832         return rv;
1833 }
1834
1835 static int __init null_init(void)
1836 {
1837         int ret = 0;
1838         unsigned int i;
1839         struct nullb *nullb;
1840         struct nullb_device *dev;
1841
1842         if (g_bs > PAGE_SIZE) {
1843                 pr_warn("invalid block size\n");
1844                 pr_warn("defaults block size to %lu\n", PAGE_SIZE);
1845                 g_bs = PAGE_SIZE;
1846         }
1847
1848         if (g_home_node != NUMA_NO_NODE && g_home_node >= nr_online_nodes) {
1849                 pr_err("invalid home_node value\n");
1850                 g_home_node = NUMA_NO_NODE;
1851         }
1852
1853         if (g_queue_mode == NULL_Q_RQ) {
1854                 pr_err("legacy IO path no longer available\n");
1855                 return -EINVAL;
1856         }
1857         if (g_queue_mode == NULL_Q_MQ && g_use_per_node_hctx) {
1858                 if (g_submit_queues != nr_online_nodes) {
1859                         pr_warn("submit_queues param is set to %u.\n",
1860                                                         nr_online_nodes);
1861                         g_submit_queues = nr_online_nodes;
1862                 }
1863         } else if (g_submit_queues > nr_cpu_ids)
1864                 g_submit_queues = nr_cpu_ids;
1865         else if (g_submit_queues <= 0)
1866                 g_submit_queues = 1;
1867
1868         if (g_queue_mode == NULL_Q_MQ && shared_tags) {
1869                 ret = null_init_tag_set(NULL, &tag_set);
1870                 if (ret)
1871                         return ret;
1872         }
1873
1874         config_group_init(&nullb_subsys.su_group);
1875         mutex_init(&nullb_subsys.su_mutex);
1876
1877         ret = configfs_register_subsystem(&nullb_subsys);
1878         if (ret)
1879                 goto err_tagset;
1880
1881         mutex_init(&lock);
1882
1883         null_major = register_blkdev(0, "nullb");
1884         if (null_major < 0) {
1885                 ret = null_major;
1886                 goto err_conf;
1887         }
1888
1889         for (i = 0; i < nr_devices; i++) {
1890                 dev = null_alloc_dev();
1891                 if (!dev) {
1892                         ret = -ENOMEM;
1893                         goto err_dev;
1894                 }
1895                 ret = null_add_dev(dev);
1896                 if (ret) {
1897                         null_free_dev(dev);
1898                         goto err_dev;
1899                 }
1900         }
1901
1902         pr_info("module loaded\n");
1903         return 0;
1904
1905 err_dev:
1906         while (!list_empty(&nullb_list)) {
1907                 nullb = list_entry(nullb_list.next, struct nullb, list);
1908                 dev = nullb->dev;
1909                 null_del_dev(nullb);
1910                 null_free_dev(dev);
1911         }
1912         unregister_blkdev(null_major, "nullb");
1913 err_conf:
1914         configfs_unregister_subsystem(&nullb_subsys);
1915 err_tagset:
1916         if (g_queue_mode == NULL_Q_MQ && shared_tags)
1917                 blk_mq_free_tag_set(&tag_set);
1918         return ret;
1919 }
1920
1921 static void __exit null_exit(void)
1922 {
1923         struct nullb *nullb;
1924
1925         configfs_unregister_subsystem(&nullb_subsys);
1926
1927         unregister_blkdev(null_major, "nullb");
1928
1929         mutex_lock(&lock);
1930         while (!list_empty(&nullb_list)) {
1931                 struct nullb_device *dev;
1932
1933                 nullb = list_entry(nullb_list.next, struct nullb, list);
1934                 dev = nullb->dev;
1935                 null_del_dev(nullb);
1936                 null_free_dev(dev);
1937         }
1938         mutex_unlock(&lock);
1939
1940         if (g_queue_mode == NULL_Q_MQ && shared_tags)
1941                 blk_mq_free_tag_set(&tag_set);
1942 }
1943
1944 module_init(null_init);
1945 module_exit(null_exit);
1946
1947 MODULE_AUTHOR("Jens Axboe <axboe@kernel.dk>");
1948 MODULE_LICENSE("GPL");