drivers/block/null_blk.c

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