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