drivers/block/rbd.c

   1
   2 /*
   3    rbd.c -- Export ceph rados objects as a Linux block device
   4
   5
   6    based on drivers/block/osdblk.c:
   7
   8    Copyright 2009 Red Hat, Inc.
   9
  10    This program is free software; you can redistribute it and/or modify
  11    it under the terms of the GNU General Public License as published by
  12    the Free Software Foundation.
  13
  14    This program is distributed in the hope that it will be useful,
  15    but WITHOUT ANY WARRANTY; without even the implied warranty of
  16    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17    GNU General Public License for more details.
  18
  19    You should have received a copy of the GNU General Public License
  20    along with this program; see the file COPYING.  If not, write to
  21    the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
  22
  23
  24
  25    For usage instructions, please refer to:
  26
  27                  Documentation/ABI/testing/sysfs-bus-rbd
  28
  29  */
  30
  31 #include <linux/ceph/libceph.h>
  32 #include <linux/ceph/osd_client.h>
  33 #include <linux/ceph/mon_client.h>
  34 #include <linux/ceph/cls_lock_client.h>
  35 #include <linux/ceph/decode.h>
  36 #include <linux/parser.h>
  37 #include <linux/bsearch.h>
  38
  39 #include <linux/kernel.h>
  40 #include <linux/device.h>
  41 #include <linux/module.h>
  42 #include <linux/blk-mq.h>
  43 #include <linux/fs.h>
  44 #include <linux/blkdev.h>
  45 #include <linux/slab.h>
  46 #include <linux/idr.h>
  47 #include <linux/workqueue.h>
  48
  49 #include "rbd_types.h"
  50
  51 #define RBD_DEBUG       /* Activate rbd_assert() calls */
  52
  53 /*
  54  * The basic unit of block I/O is a sector.  It is interpreted in a
  55  * number of contexts in Linux (blk, bio, genhd), but the default is
  56  * universally 512 bytes.  These symbols are just slightly more
  57  * meaningful than the bare numbers they represent.
  58  */
  59 #define SECTOR_SHIFT    9
  60 #define SECTOR_SIZE     (1ULL << SECTOR_SHIFT)
  61
  62 /*
  63  * Increment the given counter and return its updated value.
  64  * If the counter is already 0 it will not be incremented.
  65  * If the counter is already at its maximum value returns
  66  * -EINVAL without updating it.
  67  */
  68 static int atomic_inc_return_safe(atomic_t *v)
  69 {
  70         unsigned int counter;
  71
  72         counter = (unsigned int)__atomic_add_unless(v, 1, 0);
  73         if (counter <= (unsigned int)INT_MAX)
  74                 return (int)counter;
  75
  76         atomic_dec(v);
  77
  78         return -EINVAL;
  79 }
  80
  81 /* Decrement the counter.  Return the resulting value, or -EINVAL */
  82 static int atomic_dec_return_safe(atomic_t *v)
  83 {
  84         int counter;
  85
  86         counter = atomic_dec_return(v);
  87         if (counter >= 0)
  88                 return counter;
  89
  90         atomic_inc(v);
  91
  92         return -EINVAL;
  93 }
  94
  95 #define RBD_DRV_NAME "rbd"
  96
  97 #define RBD_MINORS_PER_MAJOR            256
  98 #define RBD_SINGLE_MAJOR_PART_SHIFT     4
  99
 100 #define RBD_MAX_PARENT_CHAIN_LEN        16
 101
 102 #define RBD_SNAP_DEV_NAME_PREFIX        "snap_"
 103 #define RBD_MAX_SNAP_NAME_LEN   \
 104                         (NAME_MAX - (sizeof (RBD_SNAP_DEV_NAME_PREFIX) - 1))
 105
 106 #define RBD_MAX_SNAP_COUNT      510     /* allows max snapc to fit in 4KB */
 107
 108 #define RBD_SNAP_HEAD_NAME      "-"
 109
 110 #define BAD_SNAP_INDEX  U32_MAX         /* invalid index into snap array */
 111
 112 /* This allows a single page to hold an image name sent by OSD */
 113 #define RBD_IMAGE_NAME_LEN_MAX  (PAGE_SIZE - sizeof (__le32) - 1)
 114 #define RBD_IMAGE_ID_LEN_MAX    64
 115
 116 #define RBD_OBJ_PREFIX_LEN_MAX  64
 117
 118 #define RBD_NOTIFY_TIMEOUT      5       /* seconds */
 119 #define RBD_RETRY_DELAY         msecs_to_jiffies(1000)
 120
 121 /* Feature bits */
 122
 123 #define RBD_FEATURE_LAYERING    (1<<0)
 124 #define RBD_FEATURE_STRIPINGV2  (1<<1)
 125 #define RBD_FEATURE_EXCLUSIVE_LOCK (1<<2)
 126 #define RBD_FEATURES_ALL        (RBD_FEATURE_LAYERING |         \
 127                                  RBD_FEATURE_STRIPINGV2 |       \
 128                                  RBD_FEATURE_EXCLUSIVE_LOCK)
 129
 130 /* Features supported by this (client software) implementation. */
 131
 132 #define RBD_FEATURES_SUPPORTED  (RBD_FEATURES_ALL)
 133
 134 /*
 135  * An RBD device name will be "rbd#", where the "rbd" comes from
 136  * RBD_DRV_NAME above, and # is a unique integer identifier.
 137  */
 138 #define DEV_NAME_LEN            32
 139
 140 /*
 141  * block device image metadata (in-memory version)
 142  */
 143 struct rbd_image_header {
 144         /* These six fields never change for a given rbd image */
 145         char *object_prefix;
 146         __u8 obj_order;
 147         __u8 crypt_type;
 148         __u8 comp_type;
 149         u64 stripe_unit;
 150         u64 stripe_count;
 151         u64 features;           /* Might be changeable someday? */
 152
 153         /* The remaining fields need to be updated occasionally */
 154         u64 image_size;
 155         struct ceph_snap_context *snapc;
 156         char *snap_names;       /* format 1 only */
 157         u64 *snap_sizes;        /* format 1 only */
 158 };
 159
 160 /*
 161  * An rbd image specification.
 162  *
 163  * The tuple (pool_id, image_id, snap_id) is sufficient to uniquely
 164  * identify an image.  Each rbd_dev structure includes a pointer to
 165  * an rbd_spec structure that encapsulates this identity.
 166  *
 167  * Each of the id's in an rbd_spec has an associated name.  For a
 168  * user-mapped image, the names are supplied and the id's associated
 169  * with them are looked up.  For a layered image, a parent image is
 170  * defined by the tuple, and the names are looked up.
 171  *
 172  * An rbd_dev structure contains a parent_spec pointer which is
 173  * non-null if the image it represents is a child in a layered
 174  * image.  This pointer will refer to the rbd_spec structure used
 175  * by the parent rbd_dev for its own identity (i.e., the structure
 176  * is shared between the parent and child).
 177  *
 178  * Since these structures are populated once, during the discovery
 179  * phase of image construction, they are effectively immutable so
 180  * we make no effort to synchronize access to them.
 181  *
 182  * Note that code herein does not assume the image name is known (it
 183  * could be a null pointer).
 184  */
 185 struct rbd_spec {
 186         u64             pool_id;
 187         const char      *pool_name;
 188
 189         const char      *image_id;
 190         const char      *image_name;
 191
 192         u64             snap_id;
 193         const char      *snap_name;
 194
 195         struct kref     kref;
 196 };
 197
 198 /*
 199  * an instance of the client.  multiple devices may share an rbd client.
 200  */
 201 struct rbd_client {
 202         struct ceph_client      *client;
 203         struct kref             kref;
 204         struct list_head        node;
 205 };
 206
 207 struct rbd_img_request;
 208 typedef void (*rbd_img_callback_t)(struct rbd_img_request *);
 209
 210 #define BAD_WHICH       U32_MAX         /* Good which or bad which, which? */
 211
 212 struct rbd_obj_request;
 213 typedef void (*rbd_obj_callback_t)(struct rbd_obj_request *);
 214
 215 enum obj_request_type {
 216         OBJ_REQUEST_NODATA, OBJ_REQUEST_BIO, OBJ_REQUEST_PAGES
 217 };
 218
 219 enum obj_operation_type {
 220         OBJ_OP_WRITE,
 221         OBJ_OP_READ,
 222         OBJ_OP_DISCARD,
 223 };
 224
 225 enum obj_req_flags {
 226         OBJ_REQ_DONE,           /* completion flag: not done = 0, done = 1 */
 227         OBJ_REQ_IMG_DATA,       /* object usage: standalone = 0, image = 1 */
 228         OBJ_REQ_KNOWN,          /* EXISTS flag valid: no = 0, yes = 1 */
 229         OBJ_REQ_EXISTS,         /* target exists: no = 0, yes = 1 */
 230 };
 231
 232 struct rbd_obj_request {
 233         const char              *object_name;
 234         u64                     offset;         /* object start byte */
 235         u64                     length;         /* bytes from offset */
 236         unsigned long           flags;
 237
 238         /*
 239          * An object request associated with an image will have its
 240          * img_data flag set; a standalone object request will not.
 241          *
 242          * A standalone object request will have which == BAD_WHICH
 243          * and a null obj_request pointer.
 244          *
 245          * An object request initiated in support of a layered image
 246          * object (to check for its existence before a write) will
 247          * have which == BAD_WHICH and a non-null obj_request pointer.
 248          *
 249          * Finally, an object request for rbd image data will have
 250          * which != BAD_WHICH, and will have a non-null img_request
 251          * pointer.  The value of which will be in the range
 252          * 0..(img_request->obj_request_count-1).
 253          */
 254         union {
 255                 struct rbd_obj_request  *obj_request;   /* STAT op */
 256                 struct {
 257                         struct rbd_img_request  *img_request;
 258                         u64                     img_offset;
 259                         /* links for img_request->obj_requests list */
 260                         struct list_head        links;
 261                 };
 262         };
 263         u32                     which;          /* posn image request list */
 264
 265         enum obj_request_type   type;
 266         union {
 267                 struct bio      *bio_list;
 268                 struct {
 269                         struct page     **pages;
 270                         u32             page_count;
 271                 };
 272         };
 273         struct page             **copyup_pages;
 274         u32                     copyup_page_count;
 275
 276         struct ceph_osd_request *osd_req;
 277
 278         u64                     xferred;        /* bytes transferred */
 279         int                     result;
 280
 281         rbd_obj_callback_t      callback;
 282         struct completion       completion;
 283
 284         struct kref             kref;
 285 };
 286
 287 enum img_req_flags {
 288         IMG_REQ_WRITE,          /* I/O direction: read = 0, write = 1 */
 289         IMG_REQ_CHILD,          /* initiator: block = 0, child image = 1 */
 290         IMG_REQ_LAYERED,        /* ENOENT handling: normal = 0, layered = 1 */
 291         IMG_REQ_DISCARD,        /* discard: normal = 0, discard request = 1 */
 292 };
 293
 294 struct rbd_img_request {
 295         struct rbd_device       *rbd_dev;
 296         u64                     offset; /* starting image byte offset */
 297         u64                     length; /* byte count from offset */
 298         unsigned long           flags;
 299         union {
 300                 u64                     snap_id;        /* for reads */
 301                 struct ceph_snap_context *snapc;        /* for writes */
 302         };
 303         union {
 304                 struct request          *rq;            /* block request */
 305                 struct rbd_obj_request  *obj_request;   /* obj req initiator */
 306         };
 307         struct page             **copyup_pages;
 308         u32                     copyup_page_count;
 309         spinlock_t              completion_lock;/* protects next_completion */
 310         u32                     next_completion;
 311         rbd_img_callback_t      callback;
 312         u64                     xferred;/* aggregate bytes transferred */
 313         int                     result; /* first nonzero obj_request result */
 314
 315         u32                     obj_request_count;
 316         struct list_head        obj_requests;   /* rbd_obj_request structs */
 317
 318         struct kref             kref;
 319 };
 320
 321 #define for_each_obj_request(ireq, oreq) \
 322         list_for_each_entry(oreq, &(ireq)->obj_requests, links)
 323 #define for_each_obj_request_from(ireq, oreq) \
 324         list_for_each_entry_from(oreq, &(ireq)->obj_requests, links)
 325 #define for_each_obj_request_safe(ireq, oreq, n) \
 326         list_for_each_entry_safe_reverse(oreq, n, &(ireq)->obj_requests, links)
 327
 328 enum rbd_watch_state {
 329         RBD_WATCH_STATE_UNREGISTERED,
 330         RBD_WATCH_STATE_REGISTERED,
 331         RBD_WATCH_STATE_ERROR,
 332 };
 333
 334 enum rbd_lock_state {
 335         RBD_LOCK_STATE_UNLOCKED,
 336         RBD_LOCK_STATE_LOCKED,
 337         RBD_LOCK_STATE_RELEASING,
 338 };
 339
 340 /* WatchNotify::ClientId */
 341 struct rbd_client_id {
 342         u64 gid;
 343         u64 handle;
 344 };
 345
 346 struct rbd_mapping {
 347         u64                     size;
 348         u64                     features;
 349         bool                    read_only;
 350 };
 351
 352 /*
 353  * a single device
 354  */
 355 struct rbd_device {
 356         int                     dev_id;         /* blkdev unique id */
 357
 358         int                     major;          /* blkdev assigned major */
 359         int                     minor;
 360         struct gendisk          *disk;          /* blkdev's gendisk and rq */
 361
 362         u32                     image_format;   /* Either 1 or 2 */
 363         struct rbd_client       *rbd_client;
 364
 365         char                    name[DEV_NAME_LEN]; /* blkdev name, e.g. rbd3 */
 366
 367         spinlock_t              lock;           /* queue, flags, open_count */
 368
 369         struct rbd_image_header header;
 370         unsigned long           flags;          /* possibly lock protected */
 371         struct rbd_spec         *spec;
 372         struct rbd_options      *opts;
 373         char                    *config_info;   /* add{,_single_major} string */
 374
 375         struct ceph_object_id   header_oid;
 376         struct ceph_object_locator header_oloc;
 377
 378         struct ceph_file_layout layout;         /* used for all rbd requests */
 379
 380         struct mutex            watch_mutex;
 381         enum rbd_watch_state    watch_state;
 382         struct ceph_osd_linger_request *watch_handle;
 383         u64                     watch_cookie;
 384         struct delayed_work     watch_dwork;
 385
 386         struct rw_semaphore     lock_rwsem;
 387         enum rbd_lock_state     lock_state;
 388         struct rbd_client_id    owner_cid;
 389         struct work_struct      acquired_lock_work;
 390         struct work_struct      released_lock_work;
 391         struct delayed_work     lock_dwork;
 392         struct work_struct      unlock_work;
 393         wait_queue_head_t       lock_waitq;
 394
 395         struct workqueue_struct *task_wq;
 396
 397         struct rbd_spec         *parent_spec;
 398         u64                     parent_overlap;
 399         atomic_t                parent_ref;
 400         struct rbd_device       *parent;
 401
 402         /* Block layer tags. */
 403         struct blk_mq_tag_set   tag_set;
 404
 405         /* protects updating the header */
 406         struct rw_semaphore     header_rwsem;
 407
 408         struct rbd_mapping      mapping;
 409
 410         struct list_head        node;
 411
 412         /* sysfs related */
 413         struct device           dev;
 414         unsigned long           open_count;     /* protected by lock */
 415 };
 416
 417 /*
 418  * Flag bits for rbd_dev->flags.  If atomicity is required,
 419  * rbd_dev->lock is used to protect access.
 420  *
 421  * Currently, only the "removing" flag (which is coupled with the
 422  * "open_count" field) requires atomic access.
 423  */
 424 enum rbd_dev_flags {
 425         RBD_DEV_FLAG_EXISTS,    /* mapped snapshot has not been deleted */
 426         RBD_DEV_FLAG_REMOVING,  /* this mapping is being removed */
 427 };
 428
 429 static DEFINE_MUTEX(client_mutex);      /* Serialize client creation */
 430
 431 static LIST_HEAD(rbd_dev_list);    /* devices */
 432 static DEFINE_SPINLOCK(rbd_dev_list_lock);
 433
 434 static LIST_HEAD(rbd_client_list);              /* clients */
 435 static DEFINE_SPINLOCK(rbd_client_list_lock);
 436
 437 /* Slab caches for frequently-allocated structures */
 438
 439 static struct kmem_cache        *rbd_img_request_cache;
 440 static struct kmem_cache        *rbd_obj_request_cache;
 441 static struct kmem_cache        *rbd_segment_name_cache;
 442
 443 static int rbd_major;
 444 static DEFINE_IDA(rbd_dev_id_ida);
 445
 446 static struct workqueue_struct *rbd_wq;
 447
 448 /*
 449  * Default to false for now, as single-major requires >= 0.75 version of
 450  * userspace rbd utility.
 451  */
 452 static bool single_major = false;
 453 module_param(single_major, bool, S_IRUGO);
 454 MODULE_PARM_DESC(single_major, "Use a single major number for all rbd devices (default: false)");
 455
 456 static int rbd_img_request_submit(struct rbd_img_request *img_request);
 457
 458 static ssize_t rbd_add(struct bus_type *bus, const char *buf,
 459                        size_t count);
 460 static ssize_t rbd_remove(struct bus_type *bus, const char *buf,
 461                           size_t count);
 462 static ssize_t rbd_add_single_major(struct bus_type *bus, const char *buf,
 463                                     size_t count);
 464 static ssize_t rbd_remove_single_major(struct bus_type *bus, const char *buf,
 465                                        size_t count);
 466 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth);
 467 static void rbd_spec_put(struct rbd_spec *spec);
 468
 469 static int rbd_dev_id_to_minor(int dev_id)
 470 {
 471         return dev_id << RBD_SINGLE_MAJOR_PART_SHIFT;
 472 }
 473
 474 static int minor_to_rbd_dev_id(int minor)
 475 {
 476         return minor >> RBD_SINGLE_MAJOR_PART_SHIFT;
 477 }
 478
 479 static bool rbd_is_lock_supported(struct rbd_device *rbd_dev)
 480 {
 481         return (rbd_dev->header.features & RBD_FEATURE_EXCLUSIVE_LOCK) &&
 482                rbd_dev->spec->snap_id == CEPH_NOSNAP &&
 483                !rbd_dev->mapping.read_only;
 484 }
 485
 486 static bool __rbd_is_lock_owner(struct rbd_device *rbd_dev)
 487 {
 488         return rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED ||
 489                rbd_dev->lock_state == RBD_LOCK_STATE_RELEASING;
 490 }
 491
 492 static bool rbd_is_lock_owner(struct rbd_device *rbd_dev)
 493 {
 494         bool is_lock_owner;
 495
 496         down_read(&rbd_dev->lock_rwsem);
 497         is_lock_owner = __rbd_is_lock_owner(rbd_dev);
 498         up_read(&rbd_dev->lock_rwsem);
 499         return is_lock_owner;
 500 }
 501
 502 static BUS_ATTR(add, S_IWUSR, NULL, rbd_add);
 503 static BUS_ATTR(remove, S_IWUSR, NULL, rbd_remove);
 504 static BUS_ATTR(add_single_major, S_IWUSR, NULL, rbd_add_single_major);
 505 static BUS_ATTR(remove_single_major, S_IWUSR, NULL, rbd_remove_single_major);
 506
 507 static struct attribute *rbd_bus_attrs[] = {
 508         &bus_attr_add.attr,
 509         &bus_attr_remove.attr,
 510         &bus_attr_add_single_major.attr,
 511         &bus_attr_remove_single_major.attr,
 512         NULL,
 513 };
 514
 515 static umode_t rbd_bus_is_visible(struct kobject *kobj,
 516                                   struct attribute *attr, int index)
 517 {
 518         if (!single_major &&
 519             (attr == &bus_attr_add_single_major.attr ||
 520              attr == &bus_attr_remove_single_major.attr))
 521                 return 0;
 522
 523         return attr->mode;
 524 }
 525
 526 static const struct attribute_group rbd_bus_group = {
 527         .attrs = rbd_bus_attrs,
 528         .is_visible = rbd_bus_is_visible,
 529 };
 530 __ATTRIBUTE_GROUPS(rbd_bus);
 531
 532 static struct bus_type rbd_bus_type = {
 533         .name           = "rbd",
 534         .bus_groups     = rbd_bus_groups,
 535 };
 536
 537 static void rbd_root_dev_release(struct device *dev)
 538 {
 539 }
 540
 541 static struct device rbd_root_dev = {
 542         .init_name =    "rbd",
 543         .release =      rbd_root_dev_release,
 544 };
 545
 546 static __printf(2, 3)
 547 void rbd_warn(struct rbd_device *rbd_dev, const char *fmt, ...)
 548 {
 549         struct va_format vaf;
 550         va_list args;
 551
 552         va_start(args, fmt);
 553         vaf.fmt = fmt;
 554         vaf.va = &args;
 555
 556         if (!rbd_dev)
 557                 printk(KERN_WARNING "%s: %pV\n", RBD_DRV_NAME, &vaf);
 558         else if (rbd_dev->disk)
 559                 printk(KERN_WARNING "%s: %s: %pV\n",
 560                         RBD_DRV_NAME, rbd_dev->disk->disk_name, &vaf);
 561         else if (rbd_dev->spec && rbd_dev->spec->image_name)
 562                 printk(KERN_WARNING "%s: image %s: %pV\n",
 563                         RBD_DRV_NAME, rbd_dev->spec->image_name, &vaf);
 564         else if (rbd_dev->spec && rbd_dev->spec->image_id)
 565                 printk(KERN_WARNING "%s: id %s: %pV\n",
 566                         RBD_DRV_NAME, rbd_dev->spec->image_id, &vaf);
 567         else    /* punt */
 568                 printk(KERN_WARNING "%s: rbd_dev %p: %pV\n",
 569                         RBD_DRV_NAME, rbd_dev, &vaf);
 570         va_end(args);
 571 }
 572
 573 #ifdef RBD_DEBUG
 574 #define rbd_assert(expr)                                                \
 575                 if (unlikely(!(expr))) {                                \
 576                         printk(KERN_ERR "\nAssertion failure in %s() "  \
 577                                                 "at line %d:\n\n"       \
 578                                         "\trbd_assert(%s);\n\n",        \
 579                                         __func__, __LINE__, #expr);     \
 580                         BUG();                                          \
 581                 }
 582 #else /* !RBD_DEBUG */
 583 #  define rbd_assert(expr)      ((void) 0)
 584 #endif /* !RBD_DEBUG */
 585
 586 static void rbd_osd_copyup_callback(struct rbd_obj_request *obj_request);
 587 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request);
 588 static void rbd_img_parent_read(struct rbd_obj_request *obj_request);
 589 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev);
 590
 591 static int rbd_dev_refresh(struct rbd_device *rbd_dev);
 592 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev);
 593 static int rbd_dev_header_info(struct rbd_device *rbd_dev);
 594 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev);
 595 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
 596                                         u64 snap_id);
 597 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
 598                                 u8 *order, u64 *snap_size);
 599 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
 600                 u64 *snap_features);
 601
 602 static int rbd_open(struct block_device *bdev, fmode_t mode)
 603 {
 604         struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
 605         bool removing = false;
 606
 607         if ((mode & FMODE_WRITE) && rbd_dev->mapping.read_only)
 608                 return -EROFS;
 609
 610         spin_lock_irq(&rbd_dev->lock);
 611         if (test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags))
 612                 removing = true;
 613         else
 614                 rbd_dev->open_count++;
 615         spin_unlock_irq(&rbd_dev->lock);
 616         if (removing)
 617                 return -ENOENT;
 618
 619         (void) get_device(&rbd_dev->dev);
 620
 621         return 0;
 622 }
 623
 624 static void rbd_release(struct gendisk *disk, fmode_t mode)
 625 {
 626         struct rbd_device *rbd_dev = disk->private_data;
 627         unsigned long open_count_before;
 628
 629         spin_lock_irq(&rbd_dev->lock);
 630         open_count_before = rbd_dev->open_count--;
 631         spin_unlock_irq(&rbd_dev->lock);
 632         rbd_assert(open_count_before > 0);
 633
 634         put_device(&rbd_dev->dev);
 635 }
 636
 637 static int rbd_ioctl_set_ro(struct rbd_device *rbd_dev, unsigned long arg)
 638 {
 639         int ret = 0;
 640         int val;
 641         bool ro;
 642         bool ro_changed = false;
 643
 644         /* get_user() may sleep, so call it before taking rbd_dev->lock */
 645         if (get_user(val, (int __user *)(arg)))
 646                 return -EFAULT;
 647
 648         ro = val ? true : false;
 649         /* Snapshot doesn't allow to write*/
 650         if (rbd_dev->spec->snap_id != CEPH_NOSNAP && !ro)
 651                 return -EROFS;
 652
 653         spin_lock_irq(&rbd_dev->lock);
 654         /* prevent others open this device */
 655         if (rbd_dev->open_count > 1) {
 656                 ret = -EBUSY;
 657                 goto out;
 658         }
 659
 660         if (rbd_dev->mapping.read_only != ro) {
 661                 rbd_dev->mapping.read_only = ro;
 662                 ro_changed = true;
 663         }
 664
 665 out:
 666         spin_unlock_irq(&rbd_dev->lock);
 667         /* set_disk_ro() may sleep, so call it after releasing rbd_dev->lock */
 668         if (ret == 0 && ro_changed)
 669                 set_disk_ro(rbd_dev->disk, ro ? 1 : 0);
 670
 671         return ret;
 672 }
 673
 674 static int rbd_ioctl(struct block_device *bdev, fmode_t mode,
 675                         unsigned int cmd, unsigned long arg)
 676 {
 677         struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
 678         int ret = 0;
 679
 680         switch (cmd) {
 681         case BLKROSET:
 682                 ret = rbd_ioctl_set_ro(rbd_dev, arg);
 683                 break;
 684         default:
 685                 ret = -ENOTTY;
 686         }
 687
 688         return ret;
 689 }
 690
 691 #ifdef CONFIG_COMPAT
 692 static int rbd_compat_ioctl(struct block_device *bdev, fmode_t mode,
 693                                 unsigned int cmd, unsigned long arg)
 694 {
 695         return rbd_ioctl(bdev, mode, cmd, arg);
 696 }
 697 #endif /* CONFIG_COMPAT */
 698
 699 static const struct block_device_operations rbd_bd_ops = {
 700         .owner                  = THIS_MODULE,
 701         .open                   = rbd_open,
 702         .release                = rbd_release,
 703         .ioctl                  = rbd_ioctl,
 704 #ifdef CONFIG_COMPAT
 705         .compat_ioctl           = rbd_compat_ioctl,
 706 #endif
 707 };
 708
 709 /*
 710  * Initialize an rbd client instance.  Success or not, this function
 711  * consumes ceph_opts.  Caller holds client_mutex.
 712  */
 713 static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts)
 714 {
 715         struct rbd_client *rbdc;
 716         int ret = -ENOMEM;
 717
 718         dout("%s:\n", __func__);
 719         rbdc = kmalloc(sizeof(struct rbd_client), GFP_KERNEL);
 720         if (!rbdc)
 721                 goto out_opt;
 722
 723         kref_init(&rbdc->kref);
 724         INIT_LIST_HEAD(&rbdc->node);
 725
 726         rbdc->client = ceph_create_client(ceph_opts, rbdc, 0, 0);
 727         if (IS_ERR(rbdc->client))
 728                 goto out_rbdc;
 729         ceph_opts = NULL; /* Now rbdc->client is responsible for ceph_opts */
 730
 731         ret = ceph_open_session(rbdc->client);
 732         if (ret < 0)
 733                 goto out_client;
 734
 735         spin_lock(&rbd_client_list_lock);
 736         list_add_tail(&rbdc->node, &rbd_client_list);
 737         spin_unlock(&rbd_client_list_lock);
 738
 739         dout("%s: rbdc %p\n", __func__, rbdc);
 740
 741         return rbdc;
 742 out_client:
 743         ceph_destroy_client(rbdc->client);
 744 out_rbdc:
 745         kfree(rbdc);
 746 out_opt:
 747         if (ceph_opts)
 748                 ceph_destroy_options(ceph_opts);
 749         dout("%s: error %d\n", __func__, ret);
 750
 751         return ERR_PTR(ret);
 752 }
 753
 754 static struct rbd_client *__rbd_get_client(struct rbd_client *rbdc)
 755 {
 756         kref_get(&rbdc->kref);
 757
 758         return rbdc;
 759 }
 760
 761 /*
 762  * Find a ceph client with specific addr and configuration.  If
 763  * found, bump its reference count.
 764  */
 765 static struct rbd_client *rbd_client_find(struct ceph_options *ceph_opts)
 766 {
 767         struct rbd_client *client_node;
 768         bool found = false;
 769
 770         if (ceph_opts->flags & CEPH_OPT_NOSHARE)
 771                 return NULL;
 772
 773         spin_lock(&rbd_client_list_lock);
 774         list_for_each_entry(client_node, &rbd_client_list, node) {
 775                 if (!ceph_compare_options(ceph_opts, client_node->client)) {
 776                         __rbd_get_client(client_node);
 777
 778                         found = true;
 779                         break;
 780                 }
 781         }
 782         spin_unlock(&rbd_client_list_lock);
 783
 784         return found ? client_node : NULL;
 785 }
 786
 787 /*
 788  * (Per device) rbd map options
 789  */
 790 enum {
 791         Opt_queue_depth,
 792         Opt_last_int,
 793         /* int args above */
 794         Opt_last_string,
 795         /* string args above */
 796         Opt_read_only,
 797         Opt_read_write,
 798         Opt_lock_on_read,
 799         Opt_err
 800 };
 801
 802 static match_table_t rbd_opts_tokens = {
 803         {Opt_queue_depth, "queue_depth=%d"},
 804         /* int args above */
 805         /* string args above */
 806         {Opt_read_only, "read_only"},
 807         {Opt_read_only, "ro"},          /* Alternate spelling */
 808         {Opt_read_write, "read_write"},
 809         {Opt_read_write, "rw"},         /* Alternate spelling */
 810         {Opt_lock_on_read, "lock_on_read"},
 811         {Opt_err, NULL}
 812 };
 813
 814 struct rbd_options {
 815         int     queue_depth;
 816         bool    read_only;
 817         bool    lock_on_read;
 818 };
 819
 820 #define RBD_QUEUE_DEPTH_DEFAULT BLKDEV_MAX_RQ
 821 #define RBD_READ_ONLY_DEFAULT   false
 822 #define RBD_LOCK_ON_READ_DEFAULT false
 823
 824 static int parse_rbd_opts_token(char *c, void *private)
 825 {
 826         struct rbd_options *rbd_opts = private;
 827         substring_t argstr[MAX_OPT_ARGS];
 828         int token, intval, ret;
 829
 830         token = match_token(c, rbd_opts_tokens, argstr);
 831         if (token < Opt_last_int) {
 832                 ret = match_int(&argstr[0], &intval);
 833                 if (ret < 0) {
 834                         pr_err("bad mount option arg (not int) at '%s'\n", c);
 835                         return ret;
 836                 }
 837                 dout("got int token %d val %d\n", token, intval);
 838         } else if (token > Opt_last_int && token < Opt_last_string) {
 839                 dout("got string token %d val %s\n", token, argstr[0].from);
 840         } else {
 841                 dout("got token %d\n", token);
 842         }
 843
 844         switch (token) {
 845         case Opt_queue_depth:
 846                 if (intval < 1) {
 847                         pr_err("queue_depth out of range\n");
 848                         return -EINVAL;
 849                 }
 850                 rbd_opts->queue_depth = intval;
 851                 break;
 852         case Opt_read_only:
 853                 rbd_opts->read_only = true;
 854                 break;
 855         case Opt_read_write:
 856                 rbd_opts->read_only = false;
 857                 break;
 858         case Opt_lock_on_read:
 859                 rbd_opts->lock_on_read = true;
 860                 break;
 861         default:
 862                 /* libceph prints "bad option" msg */
 863                 return -EINVAL;
 864         }
 865
 866         return 0;
 867 }
 868
 869 static char* obj_op_name(enum obj_operation_type op_type)
 870 {
 871         switch (op_type) {
 872         case OBJ_OP_READ:
 873                 return "read";
 874         case OBJ_OP_WRITE:
 875                 return "write";
 876         case OBJ_OP_DISCARD:
 877                 return "discard";
 878         default:
 879                 return "???";
 880         }
 881 }
 882
 883 /*
 884  * Get a ceph client with specific addr and configuration, if one does
 885  * not exist create it.  Either way, ceph_opts is consumed by this
 886  * function.
 887  */
 888 static struct rbd_client *rbd_get_client(struct ceph_options *ceph_opts)
 889 {
 890         struct rbd_client *rbdc;
 891
 892         mutex_lock_nested(&client_mutex, SINGLE_DEPTH_NESTING);
 893         rbdc = rbd_client_find(ceph_opts);
 894         if (rbdc)       /* using an existing client */
 895                 ceph_destroy_options(ceph_opts);
 896         else
 897                 rbdc = rbd_client_create(ceph_opts);
 898         mutex_unlock(&client_mutex);
 899
 900         return rbdc;
 901 }
 902
 903 /*
 904  * Destroy ceph client
 905  *
 906  * Caller must hold rbd_client_list_lock.
 907  */
 908 static void rbd_client_release(struct kref *kref)
 909 {
 910         struct rbd_client *rbdc = container_of(kref, struct rbd_client, kref);
 911
 912         dout("%s: rbdc %p\n", __func__, rbdc);
 913         spin_lock(&rbd_client_list_lock);
 914         list_del(&rbdc->node);
 915         spin_unlock(&rbd_client_list_lock);
 916
 917         ceph_destroy_client(rbdc->client);
 918         kfree(rbdc);
 919 }
 920
 921 /*
 922  * Drop reference to ceph client node. If it's not referenced anymore, release
 923  * it.
 924  */
 925 static void rbd_put_client(struct rbd_client *rbdc)
 926 {
 927         if (rbdc)
 928                 kref_put(&rbdc->kref, rbd_client_release);
 929 }
 930
 931 static bool rbd_image_format_valid(u32 image_format)
 932 {
 933         return image_format == 1 || image_format == 2;
 934 }
 935
 936 static bool rbd_dev_ondisk_valid(struct rbd_image_header_ondisk *ondisk)
 937 {
 938         size_t size;
 939         u32 snap_count;
 940
 941         /* The header has to start with the magic rbd header text */
 942         if (memcmp(&ondisk->text, RBD_HEADER_TEXT, sizeof (RBD_HEADER_TEXT)))
 943                 return false;
 944
 945         /* The bio layer requires at least sector-sized I/O */
 946
 947         if (ondisk->options.order < SECTOR_SHIFT)
 948                 return false;
 949
 950         /* If we use u64 in a few spots we may be able to loosen this */
 951
 952         if (ondisk->options.order > 8 * sizeof (int) - 1)
 953                 return false;
 954
 955         /*
 956          * The size of a snapshot header has to fit in a size_t, and
 957          * that limits the number of snapshots.
 958          */
 959         snap_count = le32_to_cpu(ondisk->snap_count);
 960         size = SIZE_MAX - sizeof (struct ceph_snap_context);
 961         if (snap_count > size / sizeof (__le64))
 962                 return false;
 963
 964         /*
 965          * Not only that, but the size of the entire the snapshot
 966          * header must also be representable in a size_t.
 967          */
 968         size -= snap_count * sizeof (__le64);
 969         if ((u64) size < le64_to_cpu(ondisk->snap_names_len))
 970                 return false;
 971
 972         return true;
 973 }
 974
 975 /*
 976  * Fill an rbd image header with information from the given format 1
 977  * on-disk header.
 978  */
 979 static int rbd_header_from_disk(struct rbd_device *rbd_dev,
 980                                  struct rbd_image_header_ondisk *ondisk)
 981 {
 982         struct rbd_image_header *header = &rbd_dev->header;
 983         bool first_time = header->object_prefix == NULL;
 984         struct ceph_snap_context *snapc;
 985         char *object_prefix = NULL;
 986         char *snap_names = NULL;
 987         u64 *snap_sizes = NULL;
 988         u32 snap_count;
 989         size_t size;
 990         int ret = -ENOMEM;
 991         u32 i;
 992
 993         /* Allocate this now to avoid having to handle failure below */
 994
 995         if (first_time) {
 996                 size_t len;
 997
 998                 len = strnlen(ondisk->object_prefix,
 999                                 sizeof (ondisk->object_prefix));
1000                 object_prefix = kmalloc(len + 1, GFP_KERNEL);
1001                 if (!object_prefix)
1002                         return -ENOMEM;
1003                 memcpy(object_prefix, ondisk->object_prefix, len);
1004                 object_prefix[len] = '\0';
1005         }
1006
1007         /* Allocate the snapshot context and fill it in */
1008
1009         snap_count = le32_to_cpu(ondisk->snap_count);
1010         snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
1011         if (!snapc)
1012                 goto out_err;
1013         snapc->seq = le64_to_cpu(ondisk->snap_seq);
1014         if (snap_count) {
1015                 struct rbd_image_snap_ondisk *snaps;
1016                 u64 snap_names_len = le64_to_cpu(ondisk->snap_names_len);
1017
1018                 /* We'll keep a copy of the snapshot names... */
1019
1020                 if (snap_names_len > (u64)SIZE_MAX)
1021                         goto out_2big;
1022                 snap_names = kmalloc(snap_names_len, GFP_KERNEL);
1023                 if (!snap_names)
1024                         goto out_err;
1025
1026                 /* ...as well as the array of their sizes. */
1027
1028                 size = snap_count * sizeof (*header->snap_sizes);
1029                 snap_sizes = kmalloc(size, GFP_KERNEL);
1030                 if (!snap_sizes)
1031                         goto out_err;
1032
1033                 /*
1034                  * Copy the names, and fill in each snapshot's id
1035                  * and size.
1036                  *
1037                  * Note that rbd_dev_v1_header_info() guarantees the
1038                  * ondisk buffer we're working with has
1039                  * snap_names_len bytes beyond the end of the
1040                  * snapshot id array, this memcpy() is safe.
1041                  */
1042                 memcpy(snap_names, &ondisk->snaps[snap_count], snap_names_len);
1043                 snaps = ondisk->snaps;
1044                 for (i = 0; i < snap_count; i++) {
1045                         snapc->snaps[i] = le64_to_cpu(snaps[i].id);
1046                         snap_sizes[i] = le64_to_cpu(snaps[i].image_size);
1047                 }
1048         }
1049
1050         /* We won't fail any more, fill in the header */
1051
1052         if (first_time) {
1053                 header->object_prefix = object_prefix;
1054                 header->obj_order = ondisk->options.order;
1055                 header->crypt_type = ondisk->options.crypt_type;
1056                 header->comp_type = ondisk->options.comp_type;
1057                 /* The rest aren't used for format 1 images */
1058                 header->stripe_unit = 0;
1059                 header->stripe_count = 0;
1060                 header->features = 0;
1061         } else {
1062                 ceph_put_snap_context(header->snapc);
1063                 kfree(header->snap_names);
1064                 kfree(header->snap_sizes);
1065         }
1066
1067         /* The remaining fields always get updated (when we refresh) */
1068
1069         header->image_size = le64_to_cpu(ondisk->image_size);
1070         header->snapc = snapc;
1071         header->snap_names = snap_names;
1072         header->snap_sizes = snap_sizes;
1073
1074         return 0;
1075 out_2big:
1076         ret = -EIO;
1077 out_err:
1078         kfree(snap_sizes);
1079         kfree(snap_names);
1080         ceph_put_snap_context(snapc);
1081         kfree(object_prefix);
1082
1083         return ret;
1084 }
1085
1086 static const char *_rbd_dev_v1_snap_name(struct rbd_device *rbd_dev, u32 which)
1087 {
1088         const char *snap_name;
1089
1090         rbd_assert(which < rbd_dev->header.snapc->num_snaps);
1091
1092         /* Skip over names until we find the one we are looking for */
1093
1094         snap_name = rbd_dev->header.snap_names;
1095         while (which--)
1096                 snap_name += strlen(snap_name) + 1;
1097
1098         return kstrdup(snap_name, GFP_KERNEL);
1099 }
1100
1101 /*
1102  * Snapshot id comparison function for use with qsort()/bsearch().
1103  * Note that result is for snapshots in *descending* order.
1104  */
1105 static int snapid_compare_reverse(const void *s1, const void *s2)
1106 {
1107         u64 snap_id1 = *(u64 *)s1;
1108         u64 snap_id2 = *(u64 *)s2;
1109
1110         if (snap_id1 < snap_id2)
1111                 return 1;
1112         return snap_id1 == snap_id2 ? 0 : -1;
1113 }
1114
1115 /*
1116  * Search a snapshot context to see if the given snapshot id is
1117  * present.
1118  *
1119  * Returns the position of the snapshot id in the array if it's found,
1120  * or BAD_SNAP_INDEX otherwise.
1121  *
1122  * Note: The snapshot array is in kept sorted (by the osd) in
1123  * reverse order, highest snapshot id first.
1124  */
1125 static u32 rbd_dev_snap_index(struct rbd_device *rbd_dev, u64 snap_id)
1126 {
1127         struct ceph_snap_context *snapc = rbd_dev->header.snapc;
1128         u64 *found;
1129
1130         found = bsearch(&snap_id, &snapc->snaps, snapc->num_snaps,
1131                                 sizeof (snap_id), snapid_compare_reverse);
1132
1133         return found ? (u32)(found - &snapc->snaps[0]) : BAD_SNAP_INDEX;
1134 }
1135
1136 static const char *rbd_dev_v1_snap_name(struct rbd_device *rbd_dev,
1137                                         u64 snap_id)
1138 {
1139         u32 which;
1140         const char *snap_name;
1141
1142         which = rbd_dev_snap_index(rbd_dev, snap_id);
1143         if (which == BAD_SNAP_INDEX)
1144                 return ERR_PTR(-ENOENT);
1145
1146         snap_name = _rbd_dev_v1_snap_name(rbd_dev, which);
1147         return snap_name ? snap_name : ERR_PTR(-ENOMEM);
1148 }
1149
1150 static const char *rbd_snap_name(struct rbd_device *rbd_dev, u64 snap_id)
1151 {
1152         if (snap_id == CEPH_NOSNAP)
1153                 return RBD_SNAP_HEAD_NAME;
1154
1155         rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1156         if (rbd_dev->image_format == 1)
1157                 return rbd_dev_v1_snap_name(rbd_dev, snap_id);
1158
1159         return rbd_dev_v2_snap_name(rbd_dev, snap_id);
1160 }
1161
1162 static int rbd_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
1163                                 u64 *snap_size)
1164 {
1165         rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1166         if (snap_id == CEPH_NOSNAP) {
1167                 *snap_size = rbd_dev->header.image_size;
1168         } else if (rbd_dev->image_format == 1) {
1169                 u32 which;
1170
1171                 which = rbd_dev_snap_index(rbd_dev, snap_id);
1172                 if (which == BAD_SNAP_INDEX)
1173                         return -ENOENT;
1174
1175                 *snap_size = rbd_dev->header.snap_sizes[which];
1176         } else {
1177                 u64 size = 0;
1178                 int ret;
1179
1180                 ret = _rbd_dev_v2_snap_size(rbd_dev, snap_id, NULL, &size);
1181                 if (ret)
1182                         return ret;
1183
1184                 *snap_size = size;
1185         }
1186         return 0;
1187 }
1188
1189 static int rbd_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
1190                         u64 *snap_features)
1191 {
1192         rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1193         if (snap_id == CEPH_NOSNAP) {
1194                 *snap_features = rbd_dev->header.features;
1195         } else if (rbd_dev->image_format == 1) {
1196                 *snap_features = 0;     /* No features for format 1 */
1197         } else {
1198                 u64 features = 0;
1199                 int ret;
1200
1201                 ret = _rbd_dev_v2_snap_features(rbd_dev, snap_id, &features);
1202                 if (ret)
1203                         return ret;
1204
1205                 *snap_features = features;
1206         }
1207         return 0;
1208 }
1209
1210 static int rbd_dev_mapping_set(struct rbd_device *rbd_dev)
1211 {
1212         u64 snap_id = rbd_dev->spec->snap_id;
1213         u64 size = 0;
1214         u64 features = 0;
1215         int ret;
1216
1217         ret = rbd_snap_size(rbd_dev, snap_id, &size);
1218         if (ret)
1219                 return ret;
1220         ret = rbd_snap_features(rbd_dev, snap_id, &features);
1221         if (ret)
1222                 return ret;
1223
1224         rbd_dev->mapping.size = size;
1225         rbd_dev->mapping.features = features;
1226
1227         return 0;
1228 }
1229
1230 static void rbd_dev_mapping_clear(struct rbd_device *rbd_dev)
1231 {
1232         rbd_dev->mapping.size = 0;
1233         rbd_dev->mapping.features = 0;
1234 }
1235
1236 static void rbd_segment_name_free(const char *name)
1237 {
1238         /* The explicit cast here is needed to drop the const qualifier */
1239
1240         kmem_cache_free(rbd_segment_name_cache, (void *)name);
1241 }
1242
1243 static const char *rbd_segment_name(struct rbd_device *rbd_dev, u64 offset)
1244 {
1245         char *name;
1246         u64 segment;
1247         int ret;
1248         char *name_format;
1249
1250         name = kmem_cache_alloc(rbd_segment_name_cache, GFP_NOIO);
1251         if (!name)
1252                 return NULL;
1253         segment = offset >> rbd_dev->header.obj_order;
1254         name_format = "%s.%012llx";
1255         if (rbd_dev->image_format == 2)
1256                 name_format = "%s.%016llx";
1257         ret = snprintf(name, CEPH_MAX_OID_NAME_LEN + 1, name_format,
1258                         rbd_dev->header.object_prefix, segment);
1259         if (ret < 0 || ret > CEPH_MAX_OID_NAME_LEN) {
1260                 pr_err("error formatting segment name for #%llu (%d)\n",
1261                         segment, ret);
1262                 rbd_segment_name_free(name);
1263                 name = NULL;
1264         }
1265
1266         return name;
1267 }
1268
1269 static u64 rbd_segment_offset(struct rbd_device *rbd_dev, u64 offset)
1270 {
1271         u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1272
1273         return offset & (segment_size - 1);
1274 }
1275
1276 static u64 rbd_segment_length(struct rbd_device *rbd_dev,
1277                                 u64 offset, u64 length)
1278 {
1279         u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1280
1281         offset &= segment_size - 1;
1282
1283         rbd_assert(length <= U64_MAX - offset);
1284         if (offset + length > segment_size)
1285                 length = segment_size - offset;
1286
1287         return length;
1288 }
1289
1290 /*
1291  * returns the size of an object in the image
1292  */
1293 static u64 rbd_obj_bytes(struct rbd_image_header *header)
1294 {
1295         return 1 << header->obj_order;
1296 }
1297
1298 /*
1299  * bio helpers
1300  */
1301
1302 static void bio_chain_put(struct bio *chain)
1303 {
1304         struct bio *tmp;
1305
1306         while (chain) {
1307                 tmp = chain;
1308                 chain = chain->bi_next;
1309                 bio_put(tmp);
1310         }
1311 }
1312
1313 /*
1314  * zeros a bio chain, starting at specific offset
1315  */
1316 static void zero_bio_chain(struct bio *chain, int start_ofs)
1317 {
1318         struct bio_vec bv;
1319         struct bvec_iter iter;
1320         unsigned long flags;
1321         void *buf;
1322         int pos = 0;
1323
1324         while (chain) {
1325                 bio_for_each_segment(bv, chain, iter) {
1326                         if (pos + bv.bv_len > start_ofs) {
1327                                 int remainder = max(start_ofs - pos, 0);
1328                                 buf = bvec_kmap_irq(&bv, &flags);
1329                                 memset(buf + remainder, 0,
1330                                        bv.bv_len - remainder);
1331                                 flush_dcache_page(bv.bv_page);
1332                                 bvec_kunmap_irq(buf, &flags);
1333                         }
1334                         pos += bv.bv_len;
1335                 }
1336
1337                 chain = chain->bi_next;
1338         }
1339 }
1340
1341 /*
1342  * similar to zero_bio_chain(), zeros data defined by a page array,
1343  * starting at the given byte offset from the start of the array and
1344  * continuing up to the given end offset.  The pages array is
1345  * assumed to be big enough to hold all bytes up to the end.
1346  */
1347 static void zero_pages(struct page **pages, u64 offset, u64 end)
1348 {
1349         struct page **page = &pages[offset >> PAGE_SHIFT];
1350
1351         rbd_assert(end > offset);
1352         rbd_assert(end - offset <= (u64)SIZE_MAX);
1353         while (offset < end) {
1354                 size_t page_offset;
1355                 size_t length;
1356                 unsigned long flags;
1357                 void *kaddr;
1358
1359                 page_offset = offset & ~PAGE_MASK;
1360                 length = min_t(size_t, PAGE_SIZE - page_offset, end - offset);
1361                 local_irq_save(flags);
1362                 kaddr = kmap_atomic(*page);
1363                 memset(kaddr + page_offset, 0, length);
1364                 flush_dcache_page(*page);
1365                 kunmap_atomic(kaddr);
1366                 local_irq_restore(flags);
1367
1368                 offset += length;
1369                 page++;
1370         }
1371 }
1372
1373 /*
1374  * Clone a portion of a bio, starting at the given byte offset
1375  * and continuing for the number of bytes indicated.
1376  */
1377 static struct bio *bio_clone_range(struct bio *bio_src,
1378                                         unsigned int offset,
1379                                         unsigned int len,
1380                                         gfp_t gfpmask)
1381 {
1382         struct bio *bio;
1383
1384         bio = bio_clone(bio_src, gfpmask);
1385         if (!bio)
1386                 return NULL;    /* ENOMEM */
1387
1388         bio_advance(bio, offset);
1389         bio->bi_iter.bi_size = len;
1390
1391         return bio;
1392 }
1393
1394 /*
1395  * Clone a portion of a bio chain, starting at the given byte offset
1396  * into the first bio in the source chain and continuing for the
1397  * number of bytes indicated.  The result is another bio chain of
1398  * exactly the given length, or a null pointer on error.
1399  *
1400  * The bio_src and offset parameters are both in-out.  On entry they
1401  * refer to the first source bio and the offset into that bio where
1402  * the start of data to be cloned is located.
1403  *
1404  * On return, bio_src is updated to refer to the bio in the source
1405  * chain that contains first un-cloned byte, and *offset will
1406  * contain the offset of that byte within that bio.
1407  */
1408 static struct bio *bio_chain_clone_range(struct bio **bio_src,
1409                                         unsigned int *offset,
1410                                         unsigned int len,
1411                                         gfp_t gfpmask)
1412 {
1413         struct bio *bi = *bio_src;
1414         unsigned int off = *offset;
1415         struct bio *chain = NULL;
1416         struct bio **end;
1417
1418         /* Build up a chain of clone bios up to the limit */
1419
1420         if (!bi || off >= bi->bi_iter.bi_size || !len)
1421                 return NULL;            /* Nothing to clone */
1422
1423         end = &chain;
1424         while (len) {
1425                 unsigned int bi_size;
1426                 struct bio *bio;
1427
1428                 if (!bi) {
1429                         rbd_warn(NULL, "bio_chain exhausted with %u left", len);
1430                         goto out_err;   /* EINVAL; ran out of bio's */
1431                 }
1432                 bi_size = min_t(unsigned int, bi->bi_iter.bi_size - off, len);
1433                 bio = bio_clone_range(bi, off, bi_size, gfpmask);
1434                 if (!bio)
1435                         goto out_err;   /* ENOMEM */
1436
1437                 *end = bio;
1438                 end = &bio->bi_next;
1439
1440                 off += bi_size;
1441                 if (off == bi->bi_iter.bi_size) {
1442                         bi = bi->bi_next;
1443                         off = 0;
1444                 }
1445                 len -= bi_size;
1446         }
1447         *bio_src = bi;
1448         *offset = off;
1449
1450         return chain;
1451 out_err:
1452         bio_chain_put(chain);
1453
1454         return NULL;
1455 }
1456
1457 /*
1458  * The default/initial value for all object request flags is 0.  For
1459  * each flag, once its value is set to 1 it is never reset to 0
1460  * again.
1461  */
1462 static void obj_request_img_data_set(struct rbd_obj_request *obj_request)
1463 {
1464         if (test_and_set_bit(OBJ_REQ_IMG_DATA, &obj_request->flags)) {
1465                 struct rbd_device *rbd_dev;
1466
1467                 rbd_dev = obj_request->img_request->rbd_dev;
1468                 rbd_warn(rbd_dev, "obj_request %p already marked img_data",
1469                         obj_request);
1470         }
1471 }
1472
1473 static bool obj_request_img_data_test(struct rbd_obj_request *obj_request)
1474 {
1475         smp_mb();
1476         return test_bit(OBJ_REQ_IMG_DATA, &obj_request->flags) != 0;
1477 }
1478
1479 static void obj_request_done_set(struct rbd_obj_request *obj_request)
1480 {
1481         if (test_and_set_bit(OBJ_REQ_DONE, &obj_request->flags)) {
1482                 struct rbd_device *rbd_dev = NULL;
1483
1484                 if (obj_request_img_data_test(obj_request))
1485                         rbd_dev = obj_request->img_request->rbd_dev;
1486                 rbd_warn(rbd_dev, "obj_request %p already marked done",
1487                         obj_request);
1488         }
1489 }
1490
1491 static bool obj_request_done_test(struct rbd_obj_request *obj_request)
1492 {
1493         smp_mb();
1494         return test_bit(OBJ_REQ_DONE, &obj_request->flags) != 0;
1495 }
1496
1497 /*
1498  * This sets the KNOWN flag after (possibly) setting the EXISTS
1499  * flag.  The latter is set based on the "exists" value provided.
1500  *
1501  * Note that for our purposes once an object exists it never goes
1502  * away again.  It's possible that the response from two existence
1503  * checks are separated by the creation of the target object, and
1504  * the first ("doesn't exist") response arrives *after* the second
1505  * ("does exist").  In that case we ignore the second one.
1506  */
1507 static void obj_request_existence_set(struct rbd_obj_request *obj_request,
1508                                 bool exists)
1509 {
1510         if (exists)
1511                 set_bit(OBJ_REQ_EXISTS, &obj_request->flags);
1512         set_bit(OBJ_REQ_KNOWN, &obj_request->flags);
1513         smp_mb();
1514 }
1515
1516 static bool obj_request_known_test(struct rbd_obj_request *obj_request)
1517 {
1518         smp_mb();
1519         return test_bit(OBJ_REQ_KNOWN, &obj_request->flags) != 0;
1520 }
1521
1522 static bool obj_request_exists_test(struct rbd_obj_request *obj_request)
1523 {
1524         smp_mb();
1525         return test_bit(OBJ_REQ_EXISTS, &obj_request->flags) != 0;
1526 }
1527
1528 static bool obj_request_overlaps_parent(struct rbd_obj_request *obj_request)
1529 {
1530         struct rbd_device *rbd_dev = obj_request->img_request->rbd_dev;
1531
1532         return obj_request->img_offset <
1533             round_up(rbd_dev->parent_overlap, rbd_obj_bytes(&rbd_dev->header));
1534 }
1535
1536 static void rbd_obj_request_get(struct rbd_obj_request *obj_request)
1537 {
1538         dout("%s: obj %p (was %d)\n", __func__, obj_request,
1539                 atomic_read(&obj_request->kref.refcount));
1540         kref_get(&obj_request->kref);
1541 }
1542
1543 static void rbd_obj_request_destroy(struct kref *kref);
1544 static void rbd_obj_request_put(struct rbd_obj_request *obj_request)
1545 {
1546         rbd_assert(obj_request != NULL);
1547         dout("%s: obj %p (was %d)\n", __func__, obj_request,
1548                 atomic_read(&obj_request->kref.refcount));
1549         kref_put(&obj_request->kref, rbd_obj_request_destroy);
1550 }
1551
1552 static void rbd_img_request_get(struct rbd_img_request *img_request)
1553 {
1554         dout("%s: img %p (was %d)\n", __func__, img_request,
1555              atomic_read(&img_request->kref.refcount));
1556         kref_get(&img_request->kref);
1557 }
1558
1559 static bool img_request_child_test(struct rbd_img_request *img_request);
1560 static void rbd_parent_request_destroy(struct kref *kref);
1561 static void rbd_img_request_destroy(struct kref *kref);
1562 static void rbd_img_request_put(struct rbd_img_request *img_request)
1563 {
1564         rbd_assert(img_request != NULL);
1565         dout("%s: img %p (was %d)\n", __func__, img_request,
1566                 atomic_read(&img_request->kref.refcount));
1567         if (img_request_child_test(img_request))
1568                 kref_put(&img_request->kref, rbd_parent_request_destroy);
1569         else
1570                 kref_put(&img_request->kref, rbd_img_request_destroy);
1571 }
1572
1573 static inline void rbd_img_obj_request_add(struct rbd_img_request *img_request,
1574                                         struct rbd_obj_request *obj_request)
1575 {
1576         rbd_assert(obj_request->img_request == NULL);
1577
1578         /* Image request now owns object's original reference */
1579         obj_request->img_request = img_request;
1580         obj_request->which = img_request->obj_request_count;
1581         rbd_assert(!obj_request_img_data_test(obj_request));
1582         obj_request_img_data_set(obj_request);
1583         rbd_assert(obj_request->which != BAD_WHICH);
1584         img_request->obj_request_count++;
1585         list_add_tail(&obj_request->links, &img_request->obj_requests);
1586         dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1587                 obj_request->which);
1588 }
1589
1590 static inline void rbd_img_obj_request_del(struct rbd_img_request *img_request,
1591                                         struct rbd_obj_request *obj_request)
1592 {
1593         rbd_assert(obj_request->which != BAD_WHICH);
1594
1595         dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1596                 obj_request->which);
1597         list_del(&obj_request->links);
1598         rbd_assert(img_request->obj_request_count > 0);
1599         img_request->obj_request_count--;
1600         rbd_assert(obj_request->which == img_request->obj_request_count);
1601         obj_request->which = BAD_WHICH;
1602         rbd_assert(obj_request_img_data_test(obj_request));
1603         rbd_assert(obj_request->img_request == img_request);
1604         obj_request->img_request = NULL;
1605         obj_request->callback = NULL;
1606         rbd_obj_request_put(obj_request);
1607 }
1608
1609 static bool obj_request_type_valid(enum obj_request_type type)
1610 {
1611         switch (type) {
1612         case OBJ_REQUEST_NODATA:
1613         case OBJ_REQUEST_BIO:
1614         case OBJ_REQUEST_PAGES:
1615                 return true;
1616         default:
1617                 return false;
1618         }
1619 }
1620
1621 static void rbd_img_obj_callback(struct rbd_obj_request *obj_request);
1622
1623 static void rbd_obj_request_submit(struct rbd_obj_request *obj_request)
1624 {
1625         struct ceph_osd_request *osd_req = obj_request->osd_req;
1626
1627         dout("%s %p osd_req %p\n", __func__, obj_request, osd_req);
1628         if (obj_request_img_data_test(obj_request)) {
1629                 WARN_ON(obj_request->callback != rbd_img_obj_callback);
1630                 rbd_img_request_get(obj_request->img_request);
1631         }
1632         ceph_osdc_start_request(osd_req->r_osdc, osd_req, false);
1633 }
1634
1635 static void rbd_obj_request_end(struct rbd_obj_request *obj_request)
1636 {
1637         dout("%s %p\n", __func__, obj_request);
1638         ceph_osdc_cancel_request(obj_request->osd_req);
1639 }
1640
1641 /*
1642  * Wait for an object request to complete.  If interrupted, cancel the
1643  * underlying osd request.
1644  *
1645  * @timeout: in jiffies, 0 means "wait forever"
1646  */
1647 static int __rbd_obj_request_wait(struct rbd_obj_request *obj_request,
1648                                   unsigned long timeout)
1649 {
1650         long ret;
1651
1652         dout("%s %p\n", __func__, obj_request);
1653         ret = wait_for_completion_interruptible_timeout(
1654                                         &obj_request->completion,
1655                                         ceph_timeout_jiffies(timeout));
1656         if (ret <= 0) {
1657                 if (ret == 0)
1658                         ret = -ETIMEDOUT;
1659                 rbd_obj_request_end(obj_request);
1660         } else {
1661                 ret = 0;
1662         }
1663
1664         dout("%s %p ret %d\n", __func__, obj_request, (int)ret);
1665         return ret;
1666 }
1667
1668 static int rbd_obj_request_wait(struct rbd_obj_request *obj_request)
1669 {
1670         return __rbd_obj_request_wait(obj_request, 0);
1671 }
1672
1673 static void rbd_img_request_complete(struct rbd_img_request *img_request)
1674 {
1675
1676         dout("%s: img %p\n", __func__, img_request);
1677
1678         /*
1679          * If no error occurred, compute the aggregate transfer
1680          * count for the image request.  We could instead use
1681          * atomic64_cmpxchg() to update it as each object request
1682          * completes; not clear which way is better off hand.
1683          */
1684         if (!img_request->result) {
1685                 struct rbd_obj_request *obj_request;
1686                 u64 xferred = 0;
1687
1688                 for_each_obj_request(img_request, obj_request)
1689                         xferred += obj_request->xferred;
1690                 img_request->xferred = xferred;
1691         }
1692
1693         if (img_request->callback)
1694                 img_request->callback(img_request);
1695         else
1696                 rbd_img_request_put(img_request);
1697 }
1698
1699 /*
1700  * The default/initial value for all image request flags is 0.  Each
1701  * is conditionally set to 1 at image request initialization time
1702  * and currently never change thereafter.
1703  */
1704 static void img_request_write_set(struct rbd_img_request *img_request)
1705 {
1706         set_bit(IMG_REQ_WRITE, &img_request->flags);
1707         smp_mb();
1708 }
1709
1710 static bool img_request_write_test(struct rbd_img_request *img_request)
1711 {
1712         smp_mb();
1713         return test_bit(IMG_REQ_WRITE, &img_request->flags) != 0;
1714 }
1715
1716 /*
1717  * Set the discard flag when the img_request is an discard request
1718  */
1719 static void img_request_discard_set(struct rbd_img_request *img_request)
1720 {
1721         set_bit(IMG_REQ_DISCARD, &img_request->flags);
1722         smp_mb();
1723 }
1724
1725 static bool img_request_discard_test(struct rbd_img_request *img_request)
1726 {
1727         smp_mb();
1728         return test_bit(IMG_REQ_DISCARD, &img_request->flags) != 0;
1729 }
1730
1731 static void img_request_child_set(struct rbd_img_request *img_request)
1732 {
1733         set_bit(IMG_REQ_CHILD, &img_request->flags);
1734         smp_mb();
1735 }
1736
1737 static void img_request_child_clear(struct rbd_img_request *img_request)
1738 {
1739         clear_bit(IMG_REQ_CHILD, &img_request->flags);
1740         smp_mb();
1741 }
1742
1743 static bool img_request_child_test(struct rbd_img_request *img_request)
1744 {
1745         smp_mb();
1746         return test_bit(IMG_REQ_CHILD, &img_request->flags) != 0;
1747 }
1748
1749 static void img_request_layered_set(struct rbd_img_request *img_request)
1750 {
1751         set_bit(IMG_REQ_LAYERED, &img_request->flags);
1752         smp_mb();
1753 }
1754
1755 static void img_request_layered_clear(struct rbd_img_request *img_request)
1756 {
1757         clear_bit(IMG_REQ_LAYERED, &img_request->flags);
1758         smp_mb();
1759 }
1760
1761 static bool img_request_layered_test(struct rbd_img_request *img_request)
1762 {
1763         smp_mb();
1764         return test_bit(IMG_REQ_LAYERED, &img_request->flags) != 0;
1765 }
1766
1767 static enum obj_operation_type
1768 rbd_img_request_op_type(struct rbd_img_request *img_request)
1769 {
1770         if (img_request_write_test(img_request))
1771                 return OBJ_OP_WRITE;
1772         else if (img_request_discard_test(img_request))
1773                 return OBJ_OP_DISCARD;
1774         else
1775                 return OBJ_OP_READ;
1776 }
1777
1778 static void
1779 rbd_img_obj_request_read_callback(struct rbd_obj_request *obj_request)
1780 {
1781         u64 xferred = obj_request->xferred;
1782         u64 length = obj_request->length;
1783
1784         dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1785                 obj_request, obj_request->img_request, obj_request->result,
1786                 xferred, length);
1787         /*
1788          * ENOENT means a hole in the image.  We zero-fill the entire
1789          * length of the request.  A short read also implies zero-fill
1790          * to the end of the request.  An error requires the whole
1791          * length of the request to be reported finished with an error
1792          * to the block layer.  In each case we update the xferred
1793          * count to indicate the whole request was satisfied.
1794          */
1795         rbd_assert(obj_request->type != OBJ_REQUEST_NODATA);
1796         if (obj_request->result == -ENOENT) {
1797                 if (obj_request->type == OBJ_REQUEST_BIO)
1798                         zero_bio_chain(obj_request->bio_list, 0);
1799                 else
1800                         zero_pages(obj_request->pages, 0, length);
1801                 obj_request->result = 0;
1802         } else if (xferred < length && !obj_request->result) {
1803                 if (obj_request->type == OBJ_REQUEST_BIO)
1804                         zero_bio_chain(obj_request->bio_list, xferred);
1805                 else
1806                         zero_pages(obj_request->pages, xferred, length);
1807         }
1808         obj_request->xferred = length;
1809         obj_request_done_set(obj_request);
1810 }
1811
1812 static void rbd_obj_request_complete(struct rbd_obj_request *obj_request)
1813 {
1814         dout("%s: obj %p cb %p\n", __func__, obj_request,
1815                 obj_request->callback);
1816         if (obj_request->callback)
1817                 obj_request->callback(obj_request);
1818         else
1819                 complete_all(&obj_request->completion);
1820 }
1821
1822 static void rbd_osd_read_callback(struct rbd_obj_request *obj_request)
1823 {
1824         struct rbd_img_request *img_request = NULL;
1825         struct rbd_device *rbd_dev = NULL;
1826         bool layered = false;
1827
1828         if (obj_request_img_data_test(obj_request)) {
1829                 img_request = obj_request->img_request;
1830                 layered = img_request && img_request_layered_test(img_request);
1831                 rbd_dev = img_request->rbd_dev;
1832         }
1833
1834         dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1835                 obj_request, img_request, obj_request->result,
1836                 obj_request->xferred, obj_request->length);
1837         if (layered && obj_request->result == -ENOENT &&
1838                         obj_request->img_offset < rbd_dev->parent_overlap)
1839                 rbd_img_parent_read(obj_request);
1840         else if (img_request)
1841                 rbd_img_obj_request_read_callback(obj_request);
1842         else
1843                 obj_request_done_set(obj_request);
1844 }
1845
1846 static void rbd_osd_write_callback(struct rbd_obj_request *obj_request)
1847 {
1848         dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1849                 obj_request->result, obj_request->length);
1850         /*
1851          * There is no such thing as a successful short write.  Set
1852          * it to our originally-requested length.
1853          */
1854         obj_request->xferred = obj_request->length;
1855         obj_request_done_set(obj_request);
1856 }
1857
1858 static void rbd_osd_discard_callback(struct rbd_obj_request *obj_request)
1859 {
1860         dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1861                 obj_request->result, obj_request->length);
1862         /*
1863          * There is no such thing as a successful short discard.  Set
1864          * it to our originally-requested length.
1865          */
1866         obj_request->xferred = obj_request->length;
1867         /* discarding a non-existent object is not a problem */
1868         if (obj_request->result == -ENOENT)
1869                 obj_request->result = 0;
1870         obj_request_done_set(obj_request);
1871 }
1872
1873 /*
1874  * For a simple stat call there's nothing to do.  We'll do more if
1875  * this is part of a write sequence for a layered image.
1876  */
1877 static void rbd_osd_stat_callback(struct rbd_obj_request *obj_request)
1878 {
1879         dout("%s: obj %p\n", __func__, obj_request);
1880         obj_request_done_set(obj_request);
1881 }
1882
1883 static void rbd_osd_call_callback(struct rbd_obj_request *obj_request)
1884 {
1885         dout("%s: obj %p\n", __func__, obj_request);
1886
1887         if (obj_request_img_data_test(obj_request))
1888                 rbd_osd_copyup_callback(obj_request);
1889         else
1890                 obj_request_done_set(obj_request);
1891 }
1892
1893 static void rbd_osd_req_callback(struct ceph_osd_request *osd_req)
1894 {
1895         struct rbd_obj_request *obj_request = osd_req->r_priv;
1896         u16 opcode;
1897
1898         dout("%s: osd_req %p\n", __func__, osd_req);
1899         rbd_assert(osd_req == obj_request->osd_req);
1900         if (obj_request_img_data_test(obj_request)) {
1901                 rbd_assert(obj_request->img_request);
1902                 rbd_assert(obj_request->which != BAD_WHICH);
1903         } else {
1904                 rbd_assert(obj_request->which == BAD_WHICH);
1905         }
1906
1907         if (osd_req->r_result < 0)
1908                 obj_request->result = osd_req->r_result;
1909
1910         /*
1911          * We support a 64-bit length, but ultimately it has to be
1912          * passed to the block layer, which just supports a 32-bit
1913          * length field.
1914          */
1915         obj_request->xferred = osd_req->r_ops[0].outdata_len;
1916         rbd_assert(obj_request->xferred < (u64)UINT_MAX);
1917
1918         opcode = osd_req->r_ops[0].op;
1919         switch (opcode) {
1920         case CEPH_OSD_OP_READ:
1921                 rbd_osd_read_callback(obj_request);
1922                 break;
1923         case CEPH_OSD_OP_SETALLOCHINT:
1924                 rbd_assert(osd_req->r_ops[1].op == CEPH_OSD_OP_WRITE ||
1925                            osd_req->r_ops[1].op == CEPH_OSD_OP_WRITEFULL);
1926                 /* fall through */
1927         case CEPH_OSD_OP_WRITE:
1928         case CEPH_OSD_OP_WRITEFULL:
1929                 rbd_osd_write_callback(obj_request);
1930                 break;
1931         case CEPH_OSD_OP_STAT:
1932                 rbd_osd_stat_callback(obj_request);
1933                 break;
1934         case CEPH_OSD_OP_DELETE:
1935         case CEPH_OSD_OP_TRUNCATE:
1936         case CEPH_OSD_OP_ZERO:
1937                 rbd_osd_discard_callback(obj_request);
1938                 break;
1939         case CEPH_OSD_OP_CALL:
1940                 rbd_osd_call_callback(obj_request);
1941                 break;
1942         default:
1943                 rbd_warn(NULL, "%s: unsupported op %hu",
1944                         obj_request->object_name, (unsigned short) opcode);
1945                 break;
1946         }
1947
1948         if (obj_request_done_test(obj_request))
1949                 rbd_obj_request_complete(obj_request);
1950 }
1951
1952 static void rbd_osd_req_format_read(struct rbd_obj_request *obj_request)
1953 {
1954         struct rbd_img_request *img_request = obj_request->img_request;
1955         struct ceph_osd_request *osd_req = obj_request->osd_req;
1956
1957         if (img_request)
1958                 osd_req->r_snapid = img_request->snap_id;
1959 }
1960
1961 static void rbd_osd_req_format_write(struct rbd_obj_request *obj_request)
1962 {
1963         struct ceph_osd_request *osd_req = obj_request->osd_req;
1964
1965         osd_req->r_mtime = CURRENT_TIME;
1966         osd_req->r_data_offset = obj_request->offset;
1967 }
1968
1969 /*
1970  * Create an osd request.  A read request has one osd op (read).
1971  * A write request has either one (watch) or two (hint+write) osd ops.
1972  * (All rbd data writes are prefixed with an allocation hint op, but
1973  * technically osd watch is a write request, hence this distinction.)
1974  */
1975 static struct ceph_osd_request *rbd_osd_req_create(
1976                                         struct rbd_device *rbd_dev,
1977                                         enum obj_operation_type op_type,
1978                                         unsigned int num_ops,
1979                                         struct rbd_obj_request *obj_request)
1980 {
1981         struct ceph_snap_context *snapc = NULL;
1982         struct ceph_osd_client *osdc;
1983         struct ceph_osd_request *osd_req;
1984
1985         if (obj_request_img_data_test(obj_request) &&
1986                 (op_type == OBJ_OP_DISCARD || op_type == OBJ_OP_WRITE)) {
1987                 struct rbd_img_request *img_request = obj_request->img_request;
1988                 if (op_type == OBJ_OP_WRITE) {
1989                         rbd_assert(img_request_write_test(img_request));
1990                 } else {
1991                         rbd_assert(img_request_discard_test(img_request));
1992                 }
1993                 snapc = img_request->snapc;
1994         }
1995
1996         rbd_assert(num_ops == 1 || ((op_type == OBJ_OP_WRITE) && num_ops == 2));
1997
1998         /* Allocate and initialize the request, for the num_ops ops */
1999
2000         osdc = &rbd_dev->rbd_client->client->osdc;
2001         osd_req = ceph_osdc_alloc_request(osdc, snapc, num_ops, false,
2002                                           GFP_NOIO);
2003         if (!osd_req)
2004                 goto fail;
2005
2006         if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
2007                 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
2008         else
2009                 osd_req->r_flags = CEPH_OSD_FLAG_READ;
2010
2011         osd_req->r_callback = rbd_osd_req_callback;
2012         osd_req->r_priv = obj_request;
2013
2014         osd_req->r_base_oloc.pool = rbd_dev->layout.pool_id;
2015         if (ceph_oid_aprintf(&osd_req->r_base_oid, GFP_NOIO, "%s",
2016                              obj_request->object_name))
2017                 goto fail;
2018
2019         if (ceph_osdc_alloc_messages(osd_req, GFP_NOIO))
2020                 goto fail;
2021
2022         return osd_req;
2023
2024 fail:
2025         ceph_osdc_put_request(osd_req);
2026         return NULL;
2027 }
2028
2029 /*
2030  * Create a copyup osd request based on the information in the object
2031  * request supplied.  A copyup request has two or three osd ops, a
2032  * copyup method call, potentially a hint op, and a write or truncate
2033  * or zero op.
2034  */
2035 static struct ceph_osd_request *
2036 rbd_osd_req_create_copyup(struct rbd_obj_request *obj_request)
2037 {
2038         struct rbd_img_request *img_request;
2039         struct ceph_snap_context *snapc;
2040         struct rbd_device *rbd_dev;
2041         struct ceph_osd_client *osdc;
2042         struct ceph_osd_request *osd_req;
2043         int num_osd_ops = 3;
2044
2045         rbd_assert(obj_request_img_data_test(obj_request));
2046         img_request = obj_request->img_request;
2047         rbd_assert(img_request);
2048         rbd_assert(img_request_write_test(img_request) ||
2049                         img_request_discard_test(img_request));
2050
2051         if (img_request_discard_test(img_request))
2052                 num_osd_ops = 2;
2053
2054         /* Allocate and initialize the request, for all the ops */
2055
2056         snapc = img_request->snapc;
2057         rbd_dev = img_request->rbd_dev;
2058         osdc = &rbd_dev->rbd_client->client->osdc;
2059         osd_req = ceph_osdc_alloc_request(osdc, snapc, num_osd_ops,
2060                                                 false, GFP_NOIO);
2061         if (!osd_req)
2062                 goto fail;
2063
2064         osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
2065         osd_req->r_callback = rbd_osd_req_callback;
2066         osd_req->r_priv = obj_request;
2067
2068         osd_req->r_base_oloc.pool = rbd_dev->layout.pool_id;
2069         if (ceph_oid_aprintf(&osd_req->r_base_oid, GFP_NOIO, "%s",
2070                              obj_request->object_name))
2071                 goto fail;
2072
2073         if (ceph_osdc_alloc_messages(osd_req, GFP_NOIO))
2074                 goto fail;
2075
2076         return osd_req;
2077
2078 fail:
2079         ceph_osdc_put_request(osd_req);
2080         return NULL;
2081 }
2082
2083
2084 static void rbd_osd_req_destroy(struct ceph_osd_request *osd_req)
2085 {
2086         ceph_osdc_put_request(osd_req);
2087 }
2088
2089 /* object_name is assumed to be a non-null pointer and NUL-terminated */
2090
2091 static struct rbd_obj_request *rbd_obj_request_create(const char *object_name,
2092                                                 u64 offset, u64 length,
2093                                                 enum obj_request_type type)
2094 {
2095         struct rbd_obj_request *obj_request;
2096         size_t size;
2097         char *name;
2098
2099         rbd_assert(obj_request_type_valid(type));
2100
2101         size = strlen(object_name) + 1;
2102         name = kmalloc(size, GFP_NOIO);
2103         if (!name)
2104                 return NULL;
2105
2106         obj_request = kmem_cache_zalloc(rbd_obj_request_cache, GFP_NOIO);
2107         if (!obj_request) {
2108                 kfree(name);
2109                 return NULL;
2110         }
2111
2112         obj_request->object_name = memcpy(name, object_name, size);
2113         obj_request->offset = offset;
2114         obj_request->length = length;
2115         obj_request->flags = 0;
2116         obj_request->which = BAD_WHICH;
2117         obj_request->type = type;
2118         INIT_LIST_HEAD(&obj_request->links);
2119         init_completion(&obj_request->completion);
2120         kref_init(&obj_request->kref);
2121
2122         dout("%s: \"%s\" %llu/%llu %d -> obj %p\n", __func__, object_name,
2123                 offset, length, (int)type, obj_request);
2124
2125         return obj_request;
2126 }
2127
2128 static void rbd_obj_request_destroy(struct kref *kref)
2129 {
2130         struct rbd_obj_request *obj_request;
2131
2132         obj_request = container_of(kref, struct rbd_obj_request, kref);
2133
2134         dout("%s: obj %p\n", __func__, obj_request);
2135
2136         rbd_assert(obj_request->img_request == NULL);
2137         rbd_assert(obj_request->which == BAD_WHICH);
2138
2139         if (obj_request->osd_req)
2140                 rbd_osd_req_destroy(obj_request->osd_req);
2141
2142         rbd_assert(obj_request_type_valid(obj_request->type));
2143         switch (obj_request->type) {
2144         case OBJ_REQUEST_NODATA:
2145                 break;          /* Nothing to do */
2146         case OBJ_REQUEST_BIO:
2147                 if (obj_request->bio_list)
2148                         bio_chain_put(obj_request->bio_list);
2149                 break;
2150         case OBJ_REQUEST_PAGES:
2151                 if (obj_request->pages)
2152                         ceph_release_page_vector(obj_request->pages,
2153                                                 obj_request->page_count);
2154                 break;
2155         }
2156
2157         kfree(obj_request->object_name);
2158         obj_request->object_name = NULL;
2159         kmem_cache_free(rbd_obj_request_cache, obj_request);
2160 }
2161
2162 /* It's OK to call this for a device with no parent */
2163
2164 static void rbd_spec_put(struct rbd_spec *spec);
2165 static void rbd_dev_unparent(struct rbd_device *rbd_dev)
2166 {
2167         rbd_dev_remove_parent(rbd_dev);
2168         rbd_spec_put(rbd_dev->parent_spec);
2169         rbd_dev->parent_spec = NULL;
2170         rbd_dev->parent_overlap = 0;
2171 }
2172
2173 /*
2174  * Parent image reference counting is used to determine when an
2175  * image's parent fields can be safely torn down--after there are no
2176  * more in-flight requests to the parent image.  When the last
2177  * reference is dropped, cleaning them up is safe.
2178  */
2179 static void rbd_dev_parent_put(struct rbd_device *rbd_dev)
2180 {
2181         int counter;
2182
2183         if (!rbd_dev->parent_spec)
2184                 return;
2185
2186         counter = atomic_dec_return_safe(&rbd_dev->parent_ref);
2187         if (counter > 0)
2188                 return;
2189
2190         /* Last reference; clean up parent data structures */
2191
2192         if (!counter)
2193                 rbd_dev_unparent(rbd_dev);
2194         else
2195                 rbd_warn(rbd_dev, "parent reference underflow");
2196 }
2197
2198 /*
2199  * If an image has a non-zero parent overlap, get a reference to its
2200  * parent.
2201  *
2202  * Returns true if the rbd device has a parent with a non-zero
2203  * overlap and a reference for it was successfully taken, or
2204  * false otherwise.
2205  */
2206 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
2207 {
2208         int counter = 0;
2209
2210         if (!rbd_dev->parent_spec)
2211                 return false;
2212
2213         down_read(&rbd_dev->header_rwsem);
2214         if (rbd_dev->parent_overlap)
2215                 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
2216         up_read(&rbd_dev->header_rwsem);
2217
2218         if (counter < 0)
2219                 rbd_warn(rbd_dev, "parent reference overflow");
2220
2221         return counter > 0;
2222 }
2223
2224 /*
2225  * Caller is responsible for filling in the list of object requests
2226  * that comprises the image request, and the Linux request pointer
2227  * (if there is one).
2228  */
2229 static struct rbd_img_request *rbd_img_request_create(
2230                                         struct rbd_device *rbd_dev,
2231                                         u64 offset, u64 length,
2232                                         enum obj_operation_type op_type,
2233                                         struct ceph_snap_context *snapc)
2234 {
2235         struct rbd_img_request *img_request;
2236
2237         img_request = kmem_cache_alloc(rbd_img_request_cache, GFP_NOIO);
2238         if (!img_request)
2239                 return NULL;
2240
2241         img_request->rq = NULL;
2242         img_request->rbd_dev = rbd_dev;
2243         img_request->offset = offset;
2244         img_request->length = length;
2245         img_request->flags = 0;
2246         if (op_type == OBJ_OP_DISCARD) {
2247                 img_request_discard_set(img_request);
2248                 img_request->snapc = snapc;
2249         } else if (op_type == OBJ_OP_WRITE) {
2250                 img_request_write_set(img_request);
2251                 img_request->snapc = snapc;
2252         } else {
2253                 img_request->snap_id = rbd_dev->spec->snap_id;
2254         }
2255         if (rbd_dev_parent_get(rbd_dev))
2256                 img_request_layered_set(img_request);
2257         spin_lock_init(&img_request->completion_lock);
2258         img_request->next_completion = 0;
2259         img_request->callback = NULL;
2260         img_request->result = 0;
2261         img_request->obj_request_count = 0;
2262         INIT_LIST_HEAD(&img_request->obj_requests);
2263         kref_init(&img_request->kref);
2264
2265         dout("%s: rbd_dev %p %s %llu/%llu -> img %p\n", __func__, rbd_dev,
2266                 obj_op_name(op_type), offset, length, img_request);
2267
2268         return img_request;
2269 }
2270
2271 static void rbd_img_request_destroy(struct kref *kref)
2272 {
2273         struct rbd_img_request *img_request;
2274         struct rbd_obj_request *obj_request;
2275         struct rbd_obj_request *next_obj_request;
2276
2277         img_request = container_of(kref, struct rbd_img_request, kref);
2278
2279         dout("%s: img %p\n", __func__, img_request);
2280
2281         for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2282                 rbd_img_obj_request_del(img_request, obj_request);
2283         rbd_assert(img_request->obj_request_count == 0);
2284
2285         if (img_request_layered_test(img_request)) {
2286                 img_request_layered_clear(img_request);
2287                 rbd_dev_parent_put(img_request->rbd_dev);
2288         }
2289
2290         if (img_request_write_test(img_request) ||
2291                 img_request_discard_test(img_request))
2292                 ceph_put_snap_context(img_request->snapc);
2293
2294         kmem_cache_free(rbd_img_request_cache, img_request);
2295 }
2296
2297 static struct rbd_img_request *rbd_parent_request_create(
2298                                         struct rbd_obj_request *obj_request,
2299                                         u64 img_offset, u64 length)
2300 {
2301         struct rbd_img_request *parent_request;
2302         struct rbd_device *rbd_dev;
2303
2304         rbd_assert(obj_request->img_request);
2305         rbd_dev = obj_request->img_request->rbd_dev;
2306
2307         parent_request = rbd_img_request_create(rbd_dev->parent, img_offset,
2308                                                 length, OBJ_OP_READ, NULL);
2309         if (!parent_request)
2310                 return NULL;
2311
2312         img_request_child_set(parent_request);
2313         rbd_obj_request_get(obj_request);
2314         parent_request->obj_request = obj_request;
2315
2316         return parent_request;
2317 }
2318
2319 static void rbd_parent_request_destroy(struct kref *kref)
2320 {
2321         struct rbd_img_request *parent_request;
2322         struct rbd_obj_request *orig_request;
2323
2324         parent_request = container_of(kref, struct rbd_img_request, kref);
2325         orig_request = parent_request->obj_request;
2326
2327         parent_request->obj_request = NULL;
2328         rbd_obj_request_put(orig_request);
2329         img_request_child_clear(parent_request);
2330
2331         rbd_img_request_destroy(kref);
2332 }
2333
2334 static bool rbd_img_obj_end_request(struct rbd_obj_request *obj_request)
2335 {
2336         struct rbd_img_request *img_request;
2337         unsigned int xferred;
2338         int result;
2339         bool more;
2340
2341         rbd_assert(obj_request_img_data_test(obj_request));
2342         img_request = obj_request->img_request;
2343
2344         rbd_assert(obj_request->xferred <= (u64)UINT_MAX);
2345         xferred = (unsigned int)obj_request->xferred;
2346         result = obj_request->result;
2347         if (result) {
2348                 struct rbd_device *rbd_dev = img_request->rbd_dev;
2349                 enum obj_operation_type op_type;
2350
2351                 if (img_request_discard_test(img_request))
2352                         op_type = OBJ_OP_DISCARD;
2353                 else if (img_request_write_test(img_request))
2354                         op_type = OBJ_OP_WRITE;
2355                 else
2356                         op_type = OBJ_OP_READ;
2357
2358                 rbd_warn(rbd_dev, "%s %llx at %llx (%llx)",
2359                         obj_op_name(op_type), obj_request->length,
2360                         obj_request->img_offset, obj_request->offset);
2361                 rbd_warn(rbd_dev, "  result %d xferred %x",
2362                         result, xferred);
2363                 if (!img_request->result)
2364                         img_request->result = result;
2365                 /*
2366                  * Need to end I/O on the entire obj_request worth of
2367                  * bytes in case of error.
2368                  */
2369                 xferred = obj_request->length;
2370         }
2371
2372         /* Image object requests don't own their page array */
2373
2374         if (obj_request->type == OBJ_REQUEST_PAGES) {
2375                 obj_request->pages = NULL;
2376                 obj_request->page_count = 0;
2377         }
2378
2379         if (img_request_child_test(img_request)) {
2380                 rbd_assert(img_request->obj_request != NULL);
2381                 more = obj_request->which < img_request->obj_request_count - 1;
2382         } else {
2383                 rbd_assert(img_request->rq != NULL);
2384
2385                 more = blk_update_request(img_request->rq, result, xferred);
2386                 if (!more)
2387                         __blk_mq_end_request(img_request->rq, result);
2388         }
2389
2390         return more;
2391 }
2392
2393 static void rbd_img_obj_callback(struct rbd_obj_request *obj_request)
2394 {
2395         struct rbd_img_request *img_request;
2396         u32 which = obj_request->which;
2397         bool more = true;
2398
2399         rbd_assert(obj_request_img_data_test(obj_request));
2400         img_request = obj_request->img_request;
2401
2402         dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
2403         rbd_assert(img_request != NULL);
2404         rbd_assert(img_request->obj_request_count > 0);
2405         rbd_assert(which != BAD_WHICH);
2406         rbd_assert(which < img_request->obj_request_count);
2407
2408         spin_lock_irq(&img_request->completion_lock);
2409         if (which != img_request->next_completion)
2410                 goto out;
2411
2412         for_each_obj_request_from(img_request, obj_request) {
2413                 rbd_assert(more);
2414                 rbd_assert(which < img_request->obj_request_count);
2415
2416                 if (!obj_request_done_test(obj_request))
2417                         break;
2418                 more = rbd_img_obj_end_request(obj_request);
2419                 which++;
2420         }
2421
2422         rbd_assert(more ^ (which == img_request->obj_request_count));
2423         img_request->next_completion = which;
2424 out:
2425         spin_unlock_irq(&img_request->completion_lock);
2426         rbd_img_request_put(img_request);
2427
2428         if (!more)
2429                 rbd_img_request_complete(img_request);
2430 }
2431
2432 /*
2433  * Add individual osd ops to the given ceph_osd_request and prepare
2434  * them for submission. num_ops is the current number of
2435  * osd operations already to the object request.
2436  */
2437 static void rbd_img_obj_request_fill(struct rbd_obj_request *obj_request,
2438                                 struct ceph_osd_request *osd_request,
2439                                 enum obj_operation_type op_type,
2440                                 unsigned int num_ops)
2441 {
2442         struct rbd_img_request *img_request = obj_request->img_request;
2443         struct rbd_device *rbd_dev = img_request->rbd_dev;
2444         u64 object_size = rbd_obj_bytes(&rbd_dev->header);
2445         u64 offset = obj_request->offset;
2446         u64 length = obj_request->length;
2447         u64 img_end;
2448         u16 opcode;
2449
2450         if (op_type == OBJ_OP_DISCARD) {
2451                 if (!offset && length == object_size &&
2452                     (!img_request_layered_test(img_request) ||
2453                      !obj_request_overlaps_parent(obj_request))) {
2454                         opcode = CEPH_OSD_OP_DELETE;
2455                 } else if ((offset + length == object_size)) {
2456                         opcode = CEPH_OSD_OP_TRUNCATE;
2457                 } else {
2458                         down_read(&rbd_dev->header_rwsem);
2459                         img_end = rbd_dev->header.image_size;
2460                         up_read(&rbd_dev->header_rwsem);
2461
2462                         if (obj_request->img_offset + length == img_end)
2463                                 opcode = CEPH_OSD_OP_TRUNCATE;
2464                         else
2465                                 opcode = CEPH_OSD_OP_ZERO;
2466                 }
2467         } else if (op_type == OBJ_OP_WRITE) {
2468                 if (!offset && length == object_size)
2469                         opcode = CEPH_OSD_OP_WRITEFULL;
2470                 else
2471                         opcode = CEPH_OSD_OP_WRITE;
2472                 osd_req_op_alloc_hint_init(osd_request, num_ops,
2473                                         object_size, object_size);
2474                 num_ops++;
2475         } else {
2476                 opcode = CEPH_OSD_OP_READ;
2477         }
2478
2479         if (opcode == CEPH_OSD_OP_DELETE)
2480                 osd_req_op_init(osd_request, num_ops, opcode, 0);
2481         else
2482                 osd_req_op_extent_init(osd_request, num_ops, opcode,
2483                                        offset, length, 0, 0);
2484
2485         if (obj_request->type == OBJ_REQUEST_BIO)
2486                 osd_req_op_extent_osd_data_bio(osd_request, num_ops,
2487                                         obj_request->bio_list, length);
2488         else if (obj_request->type == OBJ_REQUEST_PAGES)
2489                 osd_req_op_extent_osd_data_pages(osd_request, num_ops,
2490                                         obj_request->pages, length,
2491                                         offset & ~PAGE_MASK, false, false);
2492
2493         /* Discards are also writes */
2494         if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
2495                 rbd_osd_req_format_write(obj_request);
2496         else
2497                 rbd_osd_req_format_read(obj_request);
2498 }
2499
2500 /*
2501  * Split up an image request into one or more object requests, each
2502  * to a different object.  The "type" parameter indicates whether
2503  * "data_desc" is the pointer to the head of a list of bio
2504  * structures, or the base of a page array.  In either case this
2505  * function assumes data_desc describes memory sufficient to hold
2506  * all data described by the image request.
2507  */
2508 static int rbd_img_request_fill(struct rbd_img_request *img_request,
2509                                         enum obj_request_type type,
2510                                         void *data_desc)
2511 {
2512         struct rbd_device *rbd_dev = img_request->rbd_dev;
2513         struct rbd_obj_request *obj_request = NULL;
2514         struct rbd_obj_request *next_obj_request;
2515         struct bio *bio_list = NULL;
2516         unsigned int bio_offset = 0;
2517         struct page **pages = NULL;
2518         enum obj_operation_type op_type;
2519         u64 img_offset;
2520         u64 resid;
2521
2522         dout("%s: img %p type %d data_desc %p\n", __func__, img_request,
2523                 (int)type, data_desc);
2524
2525         img_offset = img_request->offset;
2526         resid = img_request->length;
2527         rbd_assert(resid > 0);
2528         op_type = rbd_img_request_op_type(img_request);
2529
2530         if (type == OBJ_REQUEST_BIO) {
2531                 bio_list = data_desc;
2532                 rbd_assert(img_offset ==
2533                            bio_list->bi_iter.bi_sector << SECTOR_SHIFT);
2534         } else if (type == OBJ_REQUEST_PAGES) {
2535                 pages = data_desc;
2536         }
2537
2538         while (resid) {
2539                 struct ceph_osd_request *osd_req;
2540                 const char *object_name;
2541                 u64 offset;
2542                 u64 length;
2543
2544                 object_name = rbd_segment_name(rbd_dev, img_offset);
2545                 if (!object_name)
2546                         goto out_unwind;
2547                 offset = rbd_segment_offset(rbd_dev, img_offset);
2548                 length = rbd_segment_length(rbd_dev, img_offset, resid);
2549                 obj_request = rbd_obj_request_create(object_name,
2550                                                 offset, length, type);
2551                 /* object request has its own copy of the object name */
2552                 rbd_segment_name_free(object_name);
2553                 if (!obj_request)
2554                         goto out_unwind;
2555
2556                 /*
2557                  * set obj_request->img_request before creating the
2558                  * osd_request so that it gets the right snapc
2559                  */
2560                 rbd_img_obj_request_add(img_request, obj_request);
2561
2562                 if (type == OBJ_REQUEST_BIO) {
2563                         unsigned int clone_size;
2564
2565                         rbd_assert(length <= (u64)UINT_MAX);
2566                         clone_size = (unsigned int)length;
2567                         obj_request->bio_list =
2568                                         bio_chain_clone_range(&bio_list,
2569                                                                 &bio_offset,
2570                                                                 clone_size,
2571                                                                 GFP_NOIO);
2572                         if (!obj_request->bio_list)
2573                                 goto out_unwind;
2574                 } else if (type == OBJ_REQUEST_PAGES) {
2575                         unsigned int page_count;
2576
2577                         obj_request->pages = pages;
2578                         page_count = (u32)calc_pages_for(offset, length);
2579                         obj_request->page_count = page_count;
2580                         if ((offset + length) & ~PAGE_MASK)
2581                                 page_count--;   /* more on last page */
2582                         pages += page_count;
2583                 }
2584
2585                 osd_req = rbd_osd_req_create(rbd_dev, op_type,
2586                                         (op_type == OBJ_OP_WRITE) ? 2 : 1,
2587                                         obj_request);
2588                 if (!osd_req)
2589                         goto out_unwind;
2590
2591                 obj_request->osd_req = osd_req;
2592                 obj_request->callback = rbd_img_obj_callback;
2593                 obj_request->img_offset = img_offset;
2594
2595                 rbd_img_obj_request_fill(obj_request, osd_req, op_type, 0);
2596
2597                 img_offset += length;
2598                 resid -= length;
2599         }
2600
2601         return 0;
2602
2603 out_unwind:
2604         for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2605                 rbd_img_obj_request_del(img_request, obj_request);
2606
2607         return -ENOMEM;
2608 }
2609
2610 static void
2611 rbd_osd_copyup_callback(struct rbd_obj_request *obj_request)
2612 {
2613         struct rbd_img_request *img_request;
2614         struct rbd_device *rbd_dev;
2615         struct page **pages;
2616         u32 page_count;
2617
2618         dout("%s: obj %p\n", __func__, obj_request);
2619
2620         rbd_assert(obj_request->type == OBJ_REQUEST_BIO ||
2621                 obj_request->type == OBJ_REQUEST_NODATA);
2622         rbd_assert(obj_request_img_data_test(obj_request));
2623         img_request = obj_request->img_request;
2624         rbd_assert(img_request);
2625
2626         rbd_dev = img_request->rbd_dev;
2627         rbd_assert(rbd_dev);
2628
2629         pages = obj_request->copyup_pages;
2630         rbd_assert(pages != NULL);
2631         obj_request->copyup_pages = NULL;
2632         page_count = obj_request->copyup_page_count;
2633         rbd_assert(page_count);
2634         obj_request->copyup_page_count = 0;
2635         ceph_release_page_vector(pages, page_count);
2636
2637         /*
2638          * We want the transfer count to reflect the size of the
2639          * original write request.  There is no such thing as a
2640          * successful short write, so if the request was successful
2641          * we can just set it to the originally-requested length.
2642          */
2643         if (!obj_request->result)
2644                 obj_request->xferred = obj_request->length;
2645
2646         obj_request_done_set(obj_request);
2647 }
2648
2649 static void
2650 rbd_img_obj_parent_read_full_callback(struct rbd_img_request *img_request)
2651 {
2652         struct rbd_obj_request *orig_request;
2653         struct ceph_osd_request *osd_req;
2654         struct rbd_device *rbd_dev;
2655         struct page **pages;
2656         enum obj_operation_type op_type;
2657         u32 page_count;
2658         int img_result;
2659         u64 parent_length;
2660
2661         rbd_assert(img_request_child_test(img_request));
2662
2663         /* First get what we need from the image request */
2664
2665         pages = img_request->copyup_pages;
2666         rbd_assert(pages != NULL);
2667         img_request->copyup_pages = NULL;
2668         page_count = img_request->copyup_page_count;
2669         rbd_assert(page_count);
2670         img_request->copyup_page_count = 0;
2671
2672         orig_request = img_request->obj_request;
2673         rbd_assert(orig_request != NULL);
2674         rbd_assert(obj_request_type_valid(orig_request->type));
2675         img_result = img_request->result;
2676         parent_length = img_request->length;
2677         rbd_assert(img_result || parent_length == img_request->xferred);
2678         rbd_img_request_put(img_request);
2679
2680         rbd_assert(orig_request->img_request);
2681         rbd_dev = orig_request->img_request->rbd_dev;
2682         rbd_assert(rbd_dev);
2683
2684         /*
2685          * If the overlap has become 0 (most likely because the
2686          * image has been flattened) we need to free the pages
2687          * and re-submit the original write request.
2688          */
2689         if (!rbd_dev->parent_overlap) {
2690                 ceph_release_page_vector(pages, page_count);
2691                 rbd_obj_request_submit(orig_request);
2692                 return;
2693         }
2694
2695         if (img_result)
2696                 goto out_err;
2697
2698         /*
2699          * The original osd request is of no use to use any more.
2700          * We need a new one that can hold the three ops in a copyup
2701          * request.  Allocate the new copyup osd request for the
2702          * original request, and release the old one.
2703          */
2704         img_result = -ENOMEM;
2705         osd_req = rbd_osd_req_create_copyup(orig_request);
2706         if (!osd_req)
2707                 goto out_err;
2708         rbd_osd_req_destroy(orig_request->osd_req);
2709         orig_request->osd_req = osd_req;
2710         orig_request->copyup_pages = pages;
2711         orig_request->copyup_page_count = page_count;
2712
2713         /* Initialize the copyup op */
2714
2715         osd_req_op_cls_init(osd_req, 0, CEPH_OSD_OP_CALL, "rbd", "copyup");
2716         osd_req_op_cls_request_data_pages(osd_req, 0, pages, parent_length, 0,
2717                                                 false, false);
2718
2719         /* Add the other op(s) */
2720
2721         op_type = rbd_img_request_op_type(orig_request->img_request);
2722         rbd_img_obj_request_fill(orig_request, osd_req, op_type, 1);
2723
2724         /* All set, send it off. */
2725
2726         rbd_obj_request_submit(orig_request);
2727         return;
2728
2729 out_err:
2730         ceph_release_page_vector(pages, page_count);
2731         orig_request->result = img_result;
2732         orig_request->xferred = 0;
2733         rbd_img_request_get(orig_request->img_request);
2734         obj_request_done_set(orig_request);
2735         rbd_obj_request_complete(orig_request);
2736 }
2737
2738 /*
2739  * Read from the parent image the range of data that covers the
2740  * entire target of the given object request.  This is used for
2741  * satisfying a layered image write request when the target of an
2742  * object request from the image request does not exist.
2743  *
2744  * A page array big enough to hold the returned data is allocated
2745  * and supplied to rbd_img_request_fill() as the "data descriptor."
2746  * When the read completes, this page array will be transferred to
2747  * the original object request for the copyup operation.
2748  *
2749  * If an error occurs, it is recorded as the result of the original
2750  * object request in rbd_img_obj_exists_callback().
2751  */
2752 static int rbd_img_obj_parent_read_full(struct rbd_obj_request *obj_request)
2753 {
2754         struct rbd_device *rbd_dev = obj_request->img_request->rbd_dev;
2755         struct rbd_img_request *parent_request = NULL;
2756         u64 img_offset;
2757         u64 length;
2758         struct page **pages = NULL;
2759         u32 page_count;
2760         int result;
2761
2762         rbd_assert(rbd_dev->parent != NULL);
2763
2764         /*
2765          * Determine the byte range covered by the object in the
2766          * child image to which the original request was to be sent.
2767          */
2768         img_offset = obj_request->img_offset - obj_request->offset;
2769         length = (u64)1 << rbd_dev->header.obj_order;
2770
2771         /*
2772          * There is no defined parent data beyond the parent
2773          * overlap, so limit what we read at that boundary if
2774          * necessary.
2775          */
2776         if (img_offset + length > rbd_dev->parent_overlap) {
2777                 rbd_assert(img_offset < rbd_dev->parent_overlap);
2778                 length = rbd_dev->parent_overlap - img_offset;
2779         }
2780
2781         /*
2782          * Allocate a page array big enough to receive the data read
2783          * from the parent.
2784          */
2785         page_count = (u32)calc_pages_for(0, length);
2786         pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2787         if (IS_ERR(pages)) {
2788                 result = PTR_ERR(pages);
2789                 pages = NULL;
2790                 goto out_err;
2791         }
2792
2793         result = -ENOMEM;
2794         parent_request = rbd_parent_request_create(obj_request,
2795                                                 img_offset, length);
2796         if (!parent_request)
2797                 goto out_err;
2798
2799         result = rbd_img_request_fill(parent_request, OBJ_REQUEST_PAGES, pages);
2800         if (result)
2801                 goto out_err;
2802
2803         parent_request->copyup_pages = pages;
2804         parent_request->copyup_page_count = page_count;
2805         parent_request->callback = rbd_img_obj_parent_read_full_callback;
2806
2807         result = rbd_img_request_submit(parent_request);
2808         if (!result)
2809                 return 0;
2810
2811         parent_request->copyup_pages = NULL;
2812         parent_request->copyup_page_count = 0;
2813         parent_request->obj_request = NULL;
2814         rbd_obj_request_put(obj_request);
2815 out_err:
2816         if (pages)
2817                 ceph_release_page_vector(pages, page_count);
2818         if (parent_request)
2819                 rbd_img_request_put(parent_request);
2820         return result;
2821 }
2822
2823 static void rbd_img_obj_exists_callback(struct rbd_obj_request *obj_request)
2824 {
2825         struct rbd_obj_request *orig_request;
2826         struct rbd_device *rbd_dev;
2827         int result;
2828
2829         rbd_assert(!obj_request_img_data_test(obj_request));
2830
2831         /*
2832          * All we need from the object request is the original
2833          * request and the result of the STAT op.  Grab those, then
2834          * we're done with the request.
2835          */
2836         orig_request = obj_request->obj_request;
2837         obj_request->obj_request = NULL;
2838         rbd_obj_request_put(orig_request);
2839         rbd_assert(orig_request);
2840         rbd_assert(orig_request->img_request);
2841
2842         result = obj_request->result;
2843         obj_request->result = 0;
2844
2845         dout("%s: obj %p for obj %p result %d %llu/%llu\n", __func__,
2846                 obj_request, orig_request, result,
2847                 obj_request->xferred, obj_request->length);
2848         rbd_obj_request_put(obj_request);
2849
2850         /*
2851          * If the overlap has become 0 (most likely because the
2852          * image has been flattened) we need to re-submit the
2853          * original request.
2854          */
2855         rbd_dev = orig_request->img_request->rbd_dev;
2856         if (!rbd_dev->parent_overlap) {
2857                 rbd_obj_request_submit(orig_request);
2858                 return;
2859         }
2860
2861         /*
2862          * Our only purpose here is to determine whether the object
2863          * exists, and we don't want to treat the non-existence as
2864          * an error.  If something else comes back, transfer the
2865          * error to the original request and complete it now.
2866          */
2867         if (!result) {
2868                 obj_request_existence_set(orig_request, true);
2869         } else if (result == -ENOENT) {
2870                 obj_request_existence_set(orig_request, false);
2871         } else {
2872                 goto fail_orig_request;
2873         }
2874
2875         /*
2876          * Resubmit the original request now that we have recorded
2877          * whether the target object exists.
2878          */
2879         result = rbd_img_obj_request_submit(orig_request);
2880         if (result)
2881                 goto fail_orig_request;
2882
2883         return;
2884
2885 fail_orig_request:
2886         orig_request->result = result;
2887         orig_request->xferred = 0;
2888         rbd_img_request_get(orig_request->img_request);
2889         obj_request_done_set(orig_request);
2890         rbd_obj_request_complete(orig_request);
2891 }
2892
2893 static int rbd_img_obj_exists_submit(struct rbd_obj_request *obj_request)
2894 {
2895         struct rbd_device *rbd_dev = obj_request->img_request->rbd_dev;
2896         struct rbd_obj_request *stat_request;
2897         struct page **pages;
2898         u32 page_count;
2899         size_t size;
2900         int ret;
2901
2902         stat_request = rbd_obj_request_create(obj_request->object_name, 0, 0,
2903                                               OBJ_REQUEST_PAGES);
2904         if (!stat_request)
2905                 return -ENOMEM;
2906
2907         stat_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
2908                                                    stat_request);
2909         if (!stat_request->osd_req) {
2910                 ret = -ENOMEM;
2911                 goto fail_stat_request;
2912         }
2913
2914         /*
2915          * The response data for a STAT call consists of:
2916          *     le64 length;
2917          *     struct {
2918          *         le32 tv_sec;
2919          *         le32 tv_nsec;
2920          *     } mtime;
2921          */
2922         size = sizeof (__le64) + sizeof (__le32) + sizeof (__le32);
2923         page_count = (u32)calc_pages_for(0, size);
2924         pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2925         if (IS_ERR(pages)) {
2926                 ret = PTR_ERR(pages);
2927                 goto fail_stat_request;
2928         }
2929
2930         osd_req_op_init(stat_request->osd_req, 0, CEPH_OSD_OP_STAT, 0);
2931         osd_req_op_raw_data_in_pages(stat_request->osd_req, 0, pages, size, 0,
2932                                      false, false);
2933
2934         rbd_obj_request_get(obj_request);
2935         stat_request->obj_request = obj_request;
2936         stat_request->pages = pages;
2937         stat_request->page_count = page_count;
2938         stat_request->callback = rbd_img_obj_exists_callback;
2939
2940         rbd_osd_req_format_read(stat_request);
2941
2942         rbd_obj_request_submit(stat_request);
2943         return 0;
2944
2945 fail_stat_request:
2946         rbd_obj_request_put(stat_request);
2947         return ret;
2948 }
2949
2950 static bool img_obj_request_simple(struct rbd_obj_request *obj_request)
2951 {
2952         struct rbd_img_request *img_request = obj_request->img_request;
2953         struct rbd_device *rbd_dev = img_request->rbd_dev;
2954
2955         /* Reads */
2956         if (!img_request_write_test(img_request) &&
2957             !img_request_discard_test(img_request))
2958                 return true;
2959
2960         /* Non-layered writes */
2961         if (!img_request_layered_test(img_request))
2962                 return true;
2963
2964         /*
2965          * Layered writes outside of the parent overlap range don't
2966          * share any data with the parent.
2967          */
2968         if (!obj_request_overlaps_parent(obj_request))
2969                 return true;
2970
2971         /*
2972          * Entire-object layered writes - we will overwrite whatever
2973          * parent data there is anyway.
2974          */
2975         if (!obj_request->offset &&
2976             obj_request->length == rbd_obj_bytes(&rbd_dev->header))
2977                 return true;
2978
2979         /*
2980          * If the object is known to already exist, its parent data has
2981          * already been copied.
2982          */
2983         if (obj_request_known_test(obj_request) &&
2984             obj_request_exists_test(obj_request))
2985                 return true;
2986
2987         return false;
2988 }
2989
2990 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request)
2991 {
2992         rbd_assert(obj_request_img_data_test(obj_request));
2993         rbd_assert(obj_request_type_valid(obj_request->type));
2994         rbd_assert(obj_request->img_request);
2995
2996         if (img_obj_request_simple(obj_request)) {
2997                 rbd_obj_request_submit(obj_request);
2998                 return 0;
2999         }
3000
3001         /*
3002          * It's a layered write.  The target object might exist but
3003          * we may not know that yet.  If we know it doesn't exist,
3004          * start by reading the data for the full target object from
3005          * the parent so we can use it for a copyup to the target.
3006          */
3007         if (obj_request_known_test(obj_request))
3008                 return rbd_img_obj_parent_read_full(obj_request);
3009
3010         /* We don't know whether the target exists.  Go find out. */
3011
3012         return rbd_img_obj_exists_submit(obj_request);
3013 }
3014
3015 static int rbd_img_request_submit(struct rbd_img_request *img_request)
3016 {
3017         struct rbd_obj_request *obj_request;
3018         struct rbd_obj_request *next_obj_request;
3019         int ret = 0;
3020
3021         dout("%s: img %p\n", __func__, img_request);
3022
3023         rbd_img_request_get(img_request);
3024         for_each_obj_request_safe(img_request, obj_request, next_obj_request) {
3025                 ret = rbd_img_obj_request_submit(obj_request);
3026                 if (ret)
3027                         goto out_put_ireq;
3028         }
3029
3030 out_put_ireq:
3031         rbd_img_request_put(img_request);
3032         return ret;
3033 }
3034
3035 static void rbd_img_parent_read_callback(struct rbd_img_request *img_request)
3036 {
3037         struct rbd_obj_request *obj_request;
3038         struct rbd_device *rbd_dev;
3039         u64 obj_end;
3040         u64 img_xferred;
3041         int img_result;
3042
3043         rbd_assert(img_request_child_test(img_request));
3044
3045         /* First get what we need from the image request and release it */
3046
3047         obj_request = img_request->obj_request;
3048         img_xferred = img_request->xferred;
3049         img_result = img_request->result;
3050         rbd_img_request_put(img_request);
3051
3052         /*
3053          * If the overlap has become 0 (most likely because the
3054          * image has been flattened) we need to re-submit the
3055          * original request.
3056          */
3057         rbd_assert(obj_request);
3058         rbd_assert(obj_request->img_request);
3059         rbd_dev = obj_request->img_request->rbd_dev;
3060         if (!rbd_dev->parent_overlap) {
3061                 rbd_obj_request_submit(obj_request);
3062                 return;
3063         }
3064
3065         obj_request->result = img_result;
3066         if (obj_request->result)
3067                 goto out;
3068
3069         /*
3070          * We need to zero anything beyond the parent overlap
3071          * boundary.  Since rbd_img_obj_request_read_callback()
3072          * will zero anything beyond the end of a short read, an
3073          * easy way to do this is to pretend the data from the
3074          * parent came up short--ending at the overlap boundary.
3075          */
3076         rbd_assert(obj_request->img_offset < U64_MAX - obj_request->length);
3077         obj_end = obj_request->img_offset + obj_request->length;
3078         if (obj_end > rbd_dev->parent_overlap) {
3079                 u64 xferred = 0;
3080
3081                 if (obj_request->img_offset < rbd_dev->parent_overlap)
3082                         xferred = rbd_dev->parent_overlap -
3083                                         obj_request->img_offset;
3084
3085                 obj_request->xferred = min(img_xferred, xferred);
3086         } else {
3087                 obj_request->xferred = img_xferred;
3088         }
3089 out:
3090         rbd_img_obj_request_read_callback(obj_request);
3091         rbd_obj_request_complete(obj_request);
3092 }
3093
3094 static void rbd_img_parent_read(struct rbd_obj_request *obj_request)
3095 {
3096         struct rbd_img_request *img_request;
3097         int result;
3098
3099         rbd_assert(obj_request_img_data_test(obj_request));
3100         rbd_assert(obj_request->img_request != NULL);
3101         rbd_assert(obj_request->result == (s32) -ENOENT);
3102         rbd_assert(obj_request_type_valid(obj_request->type));
3103
3104         /* rbd_read_finish(obj_request, obj_request->length); */
3105         img_request = rbd_parent_request_create(obj_request,
3106                                                 obj_request->img_offset,
3107                                                 obj_request->length);
3108         result = -ENOMEM;
3109         if (!img_request)
3110                 goto out_err;
3111
3112         if (obj_request->type == OBJ_REQUEST_BIO)
3113                 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3114                                                 obj_request->bio_list);
3115         else
3116                 result = rbd_img_request_fill(img_request, OBJ_REQUEST_PAGES,
3117                                                 obj_request->pages);
3118         if (result)
3119                 goto out_err;
3120
3121         img_request->callback = rbd_img_parent_read_callback;
3122         result = rbd_img_request_submit(img_request);
3123         if (result)
3124                 goto out_err;
3125
3126         return;
3127 out_err:
3128         if (img_request)
3129                 rbd_img_request_put(img_request);
3130         obj_request->result = result;
3131         obj_request->xferred = 0;
3132         obj_request_done_set(obj_request);
3133 }
3134
3135 static const struct rbd_client_id rbd_empty_cid;
3136
3137 static bool rbd_cid_equal(const struct rbd_client_id *lhs,
3138                           const struct rbd_client_id *rhs)
3139 {
3140         return lhs->gid == rhs->gid && lhs->handle == rhs->handle;
3141 }
3142
3143 static struct rbd_client_id rbd_get_cid(struct rbd_device *rbd_dev)
3144 {
3145         struct rbd_client_id cid;
3146
3147         mutex_lock(&rbd_dev->watch_mutex);
3148         cid.gid = ceph_client_gid(rbd_dev->rbd_client->client);
3149         cid.handle = rbd_dev->watch_cookie;
3150         mutex_unlock(&rbd_dev->watch_mutex);
3151         return cid;
3152 }
3153
3154 /*
3155  * lock_rwsem must be held for write
3156  */
3157 static void rbd_set_owner_cid(struct rbd_device *rbd_dev,
3158                               const struct rbd_client_id *cid)
3159 {
3160         dout("%s rbd_dev %p %llu-%llu -> %llu-%llu\n", __func__, rbd_dev,
3161              rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle,
3162              cid->gid, cid->handle);
3163         rbd_dev->owner_cid = *cid; /* struct */
3164 }
3165
3166 static void format_lock_cookie(struct rbd_device *rbd_dev, char *buf)
3167 {
3168         mutex_lock(&rbd_dev->watch_mutex);
3169         sprintf(buf, "%s %llu", RBD_LOCK_COOKIE_PREFIX, rbd_dev->watch_cookie);
3170         mutex_unlock(&rbd_dev->watch_mutex);
3171 }
3172
3173 /*
3174  * lock_rwsem must be held for write
3175  */
3176 static int rbd_lock(struct rbd_device *rbd_dev)
3177 {
3178         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3179         struct rbd_client_id cid = rbd_get_cid(rbd_dev);
3180         char cookie[32];
3181         int ret;
3182
3183         WARN_ON(__rbd_is_lock_owner(rbd_dev));
3184
3185         format_lock_cookie(rbd_dev, cookie);
3186         ret = ceph_cls_lock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
3187                             RBD_LOCK_NAME, CEPH_CLS_LOCK_EXCLUSIVE, cookie,
3188                             RBD_LOCK_TAG, "", 0);
3189         if (ret)
3190                 return ret;
3191
3192         rbd_dev->lock_state = RBD_LOCK_STATE_LOCKED;
3193         rbd_set_owner_cid(rbd_dev, &cid);
3194         queue_work(rbd_dev->task_wq, &rbd_dev->acquired_lock_work);
3195         return 0;
3196 }
3197
3198 /*
3199  * lock_rwsem must be held for write
3200  */
3201 static int rbd_unlock(struct rbd_device *rbd_dev)
3202 {
3203         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3204         char cookie[32];
3205         int ret;
3206
3207         WARN_ON(!__rbd_is_lock_owner(rbd_dev));
3208
3209         rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
3210
3211         format_lock_cookie(rbd_dev, cookie);
3212         ret = ceph_cls_unlock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
3213                               RBD_LOCK_NAME, cookie);
3214         if (ret && ret != -ENOENT) {
3215                 rbd_warn(rbd_dev, "cls_unlock failed: %d", ret);
3216                 return ret;
3217         }
3218
3219         rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3220         queue_work(rbd_dev->task_wq, &rbd_dev->released_lock_work);
3221         return 0;
3222 }
3223
3224 static int __rbd_notify_op_lock(struct rbd_device *rbd_dev,
3225                                 enum rbd_notify_op notify_op,
3226                                 struct page ***preply_pages,
3227                                 size_t *preply_len)
3228 {
3229         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3230         struct rbd_client_id cid = rbd_get_cid(rbd_dev);
3231         int buf_size = 4 + 8 + 8 + CEPH_ENCODING_START_BLK_LEN;
3232         char buf[buf_size];
3233         void *p = buf;
3234
3235         dout("%s rbd_dev %p notify_op %d\n", __func__, rbd_dev, notify_op);
3236
3237         /* encode *LockPayload NotifyMessage (op + ClientId) */
3238         ceph_start_encoding(&p, 2, 1, buf_size - CEPH_ENCODING_START_BLK_LEN);
3239         ceph_encode_32(&p, notify_op);
3240         ceph_encode_64(&p, cid.gid);
3241         ceph_encode_64(&p, cid.handle);
3242
3243         return ceph_osdc_notify(osdc, &rbd_dev->header_oid,
3244                                 &rbd_dev->header_oloc, buf, buf_size,
3245                                 RBD_NOTIFY_TIMEOUT, preply_pages, preply_len);
3246 }
3247
3248 static void rbd_notify_op_lock(struct rbd_device *rbd_dev,
3249                                enum rbd_notify_op notify_op)
3250 {
3251         struct page **reply_pages;
3252         size_t reply_len;
3253
3254         __rbd_notify_op_lock(rbd_dev, notify_op, &reply_pages, &reply_len);
3255         ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len));
3256 }
3257
3258 static void rbd_notify_acquired_lock(struct work_struct *work)
3259 {
3260         struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3261                                                   acquired_lock_work);
3262
3263         rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_ACQUIRED_LOCK);
3264 }
3265
3266 static void rbd_notify_released_lock(struct work_struct *work)
3267 {
3268         struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3269                                                   released_lock_work);
3270
3271         rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_RELEASED_LOCK);
3272 }
3273
3274 static int rbd_request_lock(struct rbd_device *rbd_dev)
3275 {
3276         struct page **reply_pages;
3277         size_t reply_len;
3278         bool lock_owner_responded = false;
3279         int ret;
3280
3281         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3282
3283         ret = __rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_REQUEST_LOCK,
3284                                    &reply_pages, &reply_len);
3285         if (ret && ret != -ETIMEDOUT) {
3286                 rbd_warn(rbd_dev, "failed to request lock: %d", ret);
3287                 goto out;
3288         }
3289
3290         if (reply_len > 0 && reply_len <= PAGE_SIZE) {
3291                 void *p = page_address(reply_pages[0]);
3292                 void *const end = p + reply_len;
3293                 u32 n;
3294
3295                 ceph_decode_32_safe(&p, end, n, e_inval); /* num_acks */
3296                 while (n--) {
3297                         u8 struct_v;
3298                         u32 len;
3299
3300                         ceph_decode_need(&p, end, 8 + 8, e_inval);
3301                         p += 8 + 8; /* skip gid and cookie */
3302
3303                         ceph_decode_32_safe(&p, end, len, e_inval);
3304                         if (!len)
3305                                 continue;
3306
3307                         if (lock_owner_responded) {
3308                                 rbd_warn(rbd_dev,
3309                                          "duplicate lock owners detected");
3310                                 ret = -EIO;
3311                                 goto out;
3312                         }
3313
3314                         lock_owner_responded = true;
3315                         ret = ceph_start_decoding(&p, end, 1, "ResponseMessage",
3316                                                   &struct_v, &len);
3317                         if (ret) {
3318                                 rbd_warn(rbd_dev,
3319                                          "failed to decode ResponseMessage: %d",
3320                                          ret);
3321                                 goto e_inval;
3322                         }
3323
3324                         ret = ceph_decode_32(&p);
3325                 }
3326         }
3327
3328         if (!lock_owner_responded) {
3329                 rbd_warn(rbd_dev, "no lock owners detected");
3330                 ret = -ETIMEDOUT;
3331         }
3332
3333 out:
3334         ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len));
3335         return ret;
3336
3337 e_inval:
3338         ret = -EINVAL;
3339         goto out;
3340 }
3341
3342 static void wake_requests(struct rbd_device *rbd_dev, bool wake_all)
3343 {
3344         dout("%s rbd_dev %p wake_all %d\n", __func__, rbd_dev, wake_all);
3345
3346         cancel_delayed_work(&rbd_dev->lock_dwork);
3347         if (wake_all)
3348                 wake_up_all(&rbd_dev->lock_waitq);
3349         else
3350                 wake_up(&rbd_dev->lock_waitq);
3351 }
3352
3353 static int get_lock_owner_info(struct rbd_device *rbd_dev,
3354                                struct ceph_locker **lockers, u32 *num_lockers)
3355 {
3356         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3357         u8 lock_type;
3358         char *lock_tag;
3359         int ret;
3360
3361         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3362
3363         ret = ceph_cls_lock_info(osdc, &rbd_dev->header_oid,
3364                                  &rbd_dev->header_oloc, RBD_LOCK_NAME,
3365                                  &lock_type, &lock_tag, lockers, num_lockers);
3366         if (ret)
3367                 return ret;
3368
3369         if (*num_lockers == 0) {
3370                 dout("%s rbd_dev %p no lockers detected\n", __func__, rbd_dev);
3371                 goto out;
3372         }
3373
3374         if (strcmp(lock_tag, RBD_LOCK_TAG)) {
3375                 rbd_warn(rbd_dev, "locked by external mechanism, tag %s",
3376                          lock_tag);
3377                 ret = -EBUSY;
3378                 goto out;
3379         }
3380
3381         if (lock_type == CEPH_CLS_LOCK_SHARED) {
3382                 rbd_warn(rbd_dev, "shared lock type detected");
3383                 ret = -EBUSY;
3384                 goto out;
3385         }
3386
3387         if (strncmp((*lockers)[0].id.cookie, RBD_LOCK_COOKIE_PREFIX,
3388                     strlen(RBD_LOCK_COOKIE_PREFIX))) {
3389                 rbd_warn(rbd_dev, "locked by external mechanism, cookie %s",
3390                          (*lockers)[0].id.cookie);
3391                 ret = -EBUSY;
3392                 goto out;
3393         }
3394
3395 out:
3396         kfree(lock_tag);
3397         return ret;
3398 }
3399
3400 static int find_watcher(struct rbd_device *rbd_dev,
3401                         const struct ceph_locker *locker)
3402 {
3403         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3404         struct ceph_watch_item *watchers;
3405         u32 num_watchers;
3406         u64 cookie;
3407         int i;
3408         int ret;
3409
3410         ret = ceph_osdc_list_watchers(osdc, &rbd_dev->header_oid,
3411                                       &rbd_dev->header_oloc, &watchers,
3412                                       &num_watchers);
3413         if (ret)
3414                 return ret;
3415
3416         sscanf(locker->id.cookie, RBD_LOCK_COOKIE_PREFIX " %llu", &cookie);
3417         for (i = 0; i < num_watchers; i++) {
3418                 if (!memcmp(&watchers[i].addr, &locker->info.addr,
3419                             sizeof(locker->info.addr)) &&
3420                     watchers[i].cookie == cookie) {
3421                         struct rbd_client_id cid = {
3422                                 .gid = le64_to_cpu(watchers[i].name.num),
3423                                 .handle = cookie,
3424                         };
3425
3426                         dout("%s rbd_dev %p found cid %llu-%llu\n", __func__,
3427                              rbd_dev, cid.gid, cid.handle);
3428                         rbd_set_owner_cid(rbd_dev, &cid);
3429                         ret = 1;
3430                         goto out;
3431                 }
3432         }
3433
3434         dout("%s rbd_dev %p no watchers\n", __func__, rbd_dev);
3435         ret = 0;
3436 out:
3437         kfree(watchers);
3438         return ret;
3439 }
3440
3441 /*
3442  * lock_rwsem must be held for write
3443  */
3444 static int rbd_try_lock(struct rbd_device *rbd_dev)
3445 {
3446         struct ceph_client *client = rbd_dev->rbd_client->client;
3447         struct ceph_locker *lockers;
3448         u32 num_lockers;
3449         int ret;
3450
3451         for (;;) {
3452                 ret = rbd_lock(rbd_dev);
3453                 if (ret != -EBUSY)
3454                         return ret;
3455
3456                 /* determine if the current lock holder is still alive */
3457                 ret = get_lock_owner_info(rbd_dev, &lockers, &num_lockers);
3458                 if (ret)
3459                         return ret;
3460
3461                 if (num_lockers == 0)
3462                         goto again;
3463
3464                 ret = find_watcher(rbd_dev, lockers);
3465                 if (ret) {
3466                         if (ret > 0)
3467                                 ret = 0; /* have to request lock */
3468                         goto out;
3469                 }
3470
3471                 rbd_warn(rbd_dev, "%s%llu seems dead, breaking lock",
3472                          ENTITY_NAME(lockers[0].id.name));
3473
3474                 ret = ceph_monc_blacklist_add(&client->monc,
3475                                               &lockers[0].info.addr);
3476                 if (ret) {
3477                         rbd_warn(rbd_dev, "blacklist of %s%llu failed: %d",
3478                                  ENTITY_NAME(lockers[0].id.name), ret);
3479                         goto out;
3480                 }
3481
3482                 ret = ceph_cls_break_lock(&client->osdc, &rbd_dev->header_oid,
3483                                           &rbd_dev->header_oloc, RBD_LOCK_NAME,
3484                                           lockers[0].id.cookie,
3485                                           &lockers[0].id.name);
3486                 if (ret && ret != -ENOENT)
3487                         goto out;
3488
3489 again:
3490                 ceph_free_lockers(lockers, num_lockers);
3491         }
3492
3493 out:
3494         ceph_free_lockers(lockers, num_lockers);
3495         return ret;
3496 }
3497
3498 /*
3499  * ret is set only if lock_state is RBD_LOCK_STATE_UNLOCKED
3500  */
3501 static enum rbd_lock_state rbd_try_acquire_lock(struct rbd_device *rbd_dev,
3502                                                 int *pret)
3503 {
3504         enum rbd_lock_state lock_state;
3505
3506         down_read(&rbd_dev->lock_rwsem);
3507         dout("%s rbd_dev %p read lock_state %d\n", __func__, rbd_dev,
3508              rbd_dev->lock_state);
3509         if (__rbd_is_lock_owner(rbd_dev)) {
3510                 lock_state = rbd_dev->lock_state;
3511                 up_read(&rbd_dev->lock_rwsem);
3512                 return lock_state;
3513         }
3514
3515         up_read(&rbd_dev->lock_rwsem);
3516         down_write(&rbd_dev->lock_rwsem);
3517         dout("%s rbd_dev %p write lock_state %d\n", __func__, rbd_dev,
3518              rbd_dev->lock_state);
3519         if (!__rbd_is_lock_owner(rbd_dev)) {
3520                 *pret = rbd_try_lock(rbd_dev);
3521                 if (*pret)
3522                         rbd_warn(rbd_dev, "failed to acquire lock: %d", *pret);
3523         }
3524
3525         lock_state = rbd_dev->lock_state;
3526         up_write(&rbd_dev->lock_rwsem);
3527         return lock_state;
3528 }
3529
3530 static void rbd_acquire_lock(struct work_struct *work)
3531 {
3532         struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
3533                                             struct rbd_device, lock_dwork);
3534         enum rbd_lock_state lock_state;
3535         int ret;
3536
3537         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3538 again:
3539         lock_state = rbd_try_acquire_lock(rbd_dev, &ret);
3540         if (lock_state != RBD_LOCK_STATE_UNLOCKED || ret == -EBLACKLISTED) {
3541                 if (lock_state == RBD_LOCK_STATE_LOCKED)
3542                         wake_requests(rbd_dev, true);
3543                 dout("%s rbd_dev %p lock_state %d ret %d - done\n", __func__,
3544                      rbd_dev, lock_state, ret);
3545                 return;
3546         }
3547
3548         ret = rbd_request_lock(rbd_dev);
3549         if (ret == -ETIMEDOUT) {
3550                 goto again; /* treat this as a dead client */
3551         } else if (ret < 0) {
3552                 rbd_warn(rbd_dev, "error requesting lock: %d", ret);
3553                 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
3554                                  RBD_RETRY_DELAY);
3555         } else {
3556                 /*
3557                  * lock owner acked, but resend if we don't see them
3558                  * release the lock
3559                  */
3560                 dout("%s rbd_dev %p requeueing lock_dwork\n", __func__,
3561                      rbd_dev);
3562                 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
3563                     msecs_to_jiffies(2 * RBD_NOTIFY_TIMEOUT * MSEC_PER_SEC));
3564         }
3565 }
3566
3567 /*
3568  * lock_rwsem must be held for write
3569  */
3570 static bool rbd_release_lock(struct rbd_device *rbd_dev)
3571 {
3572         dout("%s rbd_dev %p read lock_state %d\n", __func__, rbd_dev,
3573              rbd_dev->lock_state);
3574         if (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED)
3575                 return false;
3576
3577         rbd_dev->lock_state = RBD_LOCK_STATE_RELEASING;
3578         downgrade_write(&rbd_dev->lock_rwsem);
3579         /*
3580          * Ensure that all in-flight IO is flushed.
3581          *
3582          * FIXME: ceph_osdc_sync() flushes the entire OSD client, which
3583          * may be shared with other devices.
3584          */
3585         ceph_osdc_sync(&rbd_dev->rbd_client->client->osdc);
3586         up_read(&rbd_dev->lock_rwsem);
3587
3588         down_write(&rbd_dev->lock_rwsem);
3589         dout("%s rbd_dev %p write lock_state %d\n", __func__, rbd_dev,
3590              rbd_dev->lock_state);
3591         if (rbd_dev->lock_state != RBD_LOCK_STATE_RELEASING)
3592                 return false;
3593
3594         if (!rbd_unlock(rbd_dev))
3595                 /*
3596                  * Give others a chance to grab the lock - we would re-acquire
3597                  * almost immediately if we got new IO during ceph_osdc_sync()
3598                  * otherwise.  We need to ack our own notifications, so this
3599                  * lock_dwork will be requeued from rbd_wait_state_locked()
3600                  * after wake_requests() in rbd_handle_released_lock().
3601                  */
3602                 cancel_delayed_work(&rbd_dev->lock_dwork);
3603
3604         return true;
3605 }
3606
3607 static void rbd_release_lock_work(struct work_struct *work)
3608 {
3609         struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3610                                                   unlock_work);
3611
3612         down_write(&rbd_dev->lock_rwsem);
3613         rbd_release_lock(rbd_dev);
3614         up_write(&rbd_dev->lock_rwsem);
3615 }
3616
3617 static void rbd_handle_acquired_lock(struct rbd_device *rbd_dev, u8 struct_v,
3618                                      void **p)
3619 {
3620         struct rbd_client_id cid = { 0 };
3621
3622         if (struct_v >= 2) {
3623                 cid.gid = ceph_decode_64(p);
3624                 cid.handle = ceph_decode_64(p);
3625         }
3626
3627         dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3628              cid.handle);
3629         if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
3630                 down_write(&rbd_dev->lock_rwsem);
3631                 if (rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
3632                         /*
3633                          * we already know that the remote client is
3634                          * the owner
3635                          */
3636                         up_write(&rbd_dev->lock_rwsem);
3637                         return;
3638                 }
3639
3640                 rbd_set_owner_cid(rbd_dev, &cid);
3641                 downgrade_write(&rbd_dev->lock_rwsem);
3642         } else {
3643                 down_read(&rbd_dev->lock_rwsem);
3644         }
3645
3646         if (!__rbd_is_lock_owner(rbd_dev))
3647                 wake_requests(rbd_dev, false);
3648         up_read(&rbd_dev->lock_rwsem);
3649 }
3650
3651 static void rbd_handle_released_lock(struct rbd_device *rbd_dev, u8 struct_v,
3652                                      void **p)
3653 {
3654         struct rbd_client_id cid = { 0 };
3655
3656         if (struct_v >= 2) {
3657                 cid.gid = ceph_decode_64(p);
3658                 cid.handle = ceph_decode_64(p);
3659         }
3660
3661         dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3662              cid.handle);
3663         if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
3664                 down_write(&rbd_dev->lock_rwsem);
3665                 if (!rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
3666                         dout("%s rbd_dev %p unexpected owner, cid %llu-%llu != owner_cid %llu-%llu\n",
3667                              __func__, rbd_dev, cid.gid, cid.handle,
3668                              rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle);
3669                         up_write(&rbd_dev->lock_rwsem);
3670                         return;
3671                 }
3672
3673                 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3674                 downgrade_write(&rbd_dev->lock_rwsem);
3675         } else {
3676                 down_read(&rbd_dev->lock_rwsem);
3677         }
3678
3679         if (!__rbd_is_lock_owner(rbd_dev))
3680                 wake_requests(rbd_dev, false);
3681         up_read(&rbd_dev->lock_rwsem);
3682 }
3683
3684 static bool rbd_handle_request_lock(struct rbd_device *rbd_dev, u8 struct_v,
3685                                     void **p)
3686 {
3687         struct rbd_client_id my_cid = rbd_get_cid(rbd_dev);
3688         struct rbd_client_id cid = { 0 };
3689         bool need_to_send;
3690
3691         if (struct_v >= 2) {
3692                 cid.gid = ceph_decode_64(p);
3693                 cid.handle = ceph_decode_64(p);
3694         }
3695
3696         dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3697              cid.handle);
3698         if (rbd_cid_equal(&cid, &my_cid))
3699                 return false;
3700
3701         down_read(&rbd_dev->lock_rwsem);
3702         need_to_send = __rbd_is_lock_owner(rbd_dev);
3703         if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED) {
3704                 if (!rbd_cid_equal(&rbd_dev->owner_cid, &rbd_empty_cid)) {
3705                         dout("%s rbd_dev %p queueing unlock_work\n", __func__,
3706                              rbd_dev);
3707                         queue_work(rbd_dev->task_wq, &rbd_dev->unlock_work);
3708                 }
3709         }
3710         up_read(&rbd_dev->lock_rwsem);
3711         return need_to_send;
3712 }
3713
3714 static void __rbd_acknowledge_notify(struct rbd_device *rbd_dev,
3715                                      u64 notify_id, u64 cookie, s32 *result)
3716 {
3717         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3718         int buf_size = 4 + CEPH_ENCODING_START_BLK_LEN;
3719         char buf[buf_size];
3720         int ret;
3721
3722         if (result) {
3723                 void *p = buf;
3724
3725                 /* encode ResponseMessage */
3726                 ceph_start_encoding(&p, 1, 1,
3727                                     buf_size - CEPH_ENCODING_START_BLK_LEN);
3728                 ceph_encode_32(&p, *result);
3729         } else {
3730                 buf_size = 0;
3731         }
3732
3733         ret = ceph_osdc_notify_ack(osdc, &rbd_dev->header_oid,
3734                                    &rbd_dev->header_oloc, notify_id, cookie,
3735                                    buf, buf_size);
3736         if (ret)
3737                 rbd_warn(rbd_dev, "acknowledge_notify failed: %d", ret);
3738 }
3739
3740 static void rbd_acknowledge_notify(struct rbd_device *rbd_dev, u64 notify_id,
3741                                    u64 cookie)
3742 {
3743         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3744         __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, NULL);
3745 }
3746
3747 static void rbd_acknowledge_notify_result(struct rbd_device *rbd_dev,
3748                                           u64 notify_id, u64 cookie, s32 result)
3749 {
3750         dout("%s rbd_dev %p result %d\n", __func__, rbd_dev, result);
3751         __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, &result);
3752 }
3753
3754 static void rbd_watch_cb(void *arg, u64 notify_id, u64 cookie,
3755                          u64 notifier_id, void *data, size_t data_len)
3756 {
3757         struct rbd_device *rbd_dev = arg;
3758         void *p = data;
3759         void *const end = p + data_len;
3760         u8 struct_v;
3761         u32 len;
3762         u32 notify_op;
3763         int ret;
3764
3765         dout("%s rbd_dev %p cookie %llu notify_id %llu data_len %zu\n",
3766              __func__, rbd_dev, cookie, notify_id, data_len);
3767         if (data_len) {
3768                 ret = ceph_start_decoding(&p, end, 1, "NotifyMessage",
3769                                           &struct_v, &len);
3770                 if (ret) {
3771                         rbd_warn(rbd_dev, "failed to decode NotifyMessage: %d",
3772                                  ret);
3773                         return;
3774                 }
3775
3776                 notify_op = ceph_decode_32(&p);
3777         } else {
3778                 /* legacy notification for header updates */
3779                 notify_op = RBD_NOTIFY_OP_HEADER_UPDATE;
3780                 len = 0;
3781         }
3782
3783         dout("%s rbd_dev %p notify_op %u\n", __func__, rbd_dev, notify_op);
3784         switch (notify_op) {
3785         case RBD_NOTIFY_OP_ACQUIRED_LOCK:
3786                 rbd_handle_acquired_lock(rbd_dev, struct_v, &p);
3787                 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3788                 break;
3789         case RBD_NOTIFY_OP_RELEASED_LOCK:
3790                 rbd_handle_released_lock(rbd_dev, struct_v, &p);
3791                 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3792                 break;
3793         case RBD_NOTIFY_OP_REQUEST_LOCK:
3794                 if (rbd_handle_request_lock(rbd_dev, struct_v, &p))
3795                         /*
3796                          * send ResponseMessage(0) back so the client
3797                          * can detect a missing owner
3798                          */
3799                         rbd_acknowledge_notify_result(rbd_dev, notify_id,
3800                                                       cookie, 0);
3801                 else
3802                         rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3803                 break;
3804         case RBD_NOTIFY_OP_HEADER_UPDATE:
3805                 ret = rbd_dev_refresh(rbd_dev);
3806                 if (ret)
3807                         rbd_warn(rbd_dev, "refresh failed: %d", ret);
3808
3809                 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3810                 break;
3811         default:
3812                 if (rbd_is_lock_owner(rbd_dev))
3813                         rbd_acknowledge_notify_result(rbd_dev, notify_id,
3814                                                       cookie, -EOPNOTSUPP);
3815                 else
3816                         rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3817                 break;
3818         }
3819 }
3820
3821 static void __rbd_unregister_watch(struct rbd_device *rbd_dev);
3822
3823 static void rbd_watch_errcb(void *arg, u64 cookie, int err)
3824 {
3825         struct rbd_device *rbd_dev = arg;
3826
3827         rbd_warn(rbd_dev, "encountered watch error: %d", err);
3828
3829         down_write(&rbd_dev->lock_rwsem);
3830         rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3831         up_write(&rbd_dev->lock_rwsem);
3832
3833         mutex_lock(&rbd_dev->watch_mutex);
3834         if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED) {
3835                 __rbd_unregister_watch(rbd_dev);
3836                 rbd_dev->watch_state = RBD_WATCH_STATE_ERROR;
3837
3838                 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->watch_dwork, 0);
3839         }
3840         mutex_unlock(&rbd_dev->watch_mutex);
3841 }
3842
3843 /*
3844  * watch_mutex must be locked
3845  */
3846 static int __rbd_register_watch(struct rbd_device *rbd_dev)
3847 {
3848         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3849         struct ceph_osd_linger_request *handle;
3850
3851         rbd_assert(!rbd_dev->watch_handle);
3852         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3853
3854         handle = ceph_osdc_watch(osdc, &rbd_dev->header_oid,
3855                                  &rbd_dev->header_oloc, rbd_watch_cb,
3856                                  rbd_watch_errcb, rbd_dev);
3857         if (IS_ERR(handle))
3858                 return PTR_ERR(handle);
3859
3860         rbd_dev->watch_handle = handle;
3861         return 0;
3862 }
3863
3864 /*
3865  * watch_mutex must be locked
3866  */
3867 static void __rbd_unregister_watch(struct rbd_device *rbd_dev)
3868 {
3869         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3870         int ret;
3871
3872         rbd_assert(rbd_dev->watch_handle);
3873         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3874
3875         ret = ceph_osdc_unwatch(osdc, rbd_dev->watch_handle);
3876         if (ret)
3877                 rbd_warn(rbd_dev, "failed to unwatch: %d", ret);
3878
3879         rbd_dev->watch_handle = NULL;
3880 }
3881
3882 static int rbd_register_watch(struct rbd_device *rbd_dev)
3883 {
3884         int ret;
3885
3886         mutex_lock(&rbd_dev->watch_mutex);
3887         rbd_assert(rbd_dev->watch_state == RBD_WATCH_STATE_UNREGISTERED);
3888         ret = __rbd_register_watch(rbd_dev);
3889         if (ret)
3890                 goto out;
3891
3892         rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
3893         rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
3894
3895 out:
3896         mutex_unlock(&rbd_dev->watch_mutex);
3897         return ret;
3898 }
3899
3900 static void cancel_tasks_sync(struct rbd_device *rbd_dev)
3901 {
3902         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3903
3904         cancel_delayed_work_sync(&rbd_dev->watch_dwork);
3905         cancel_work_sync(&rbd_dev->acquired_lock_work);
3906         cancel_work_sync(&rbd_dev->released_lock_work);
3907         cancel_delayed_work_sync(&rbd_dev->lock_dwork);
3908         cancel_work_sync(&rbd_dev->unlock_work);
3909 }
3910
3911 static void rbd_unregister_watch(struct rbd_device *rbd_dev)
3912 {
3913         WARN_ON(waitqueue_active(&rbd_dev->lock_waitq));
3914         cancel_tasks_sync(rbd_dev);
3915
3916         mutex_lock(&rbd_dev->watch_mutex);
3917         if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED)
3918                 __rbd_unregister_watch(rbd_dev);
3919         rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
3920         mutex_unlock(&rbd_dev->watch_mutex);
3921
3922         ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
3923 }
3924
3925 static void rbd_reregister_watch(struct work_struct *work)
3926 {
3927         struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
3928                                             struct rbd_device, watch_dwork);
3929         bool was_lock_owner = false;
3930         int ret;
3931
3932         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3933
3934         down_write(&rbd_dev->lock_rwsem);
3935         if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED)
3936                 was_lock_owner = rbd_release_lock(rbd_dev);
3937
3938         mutex_lock(&rbd_dev->watch_mutex);
3939         if (rbd_dev->watch_state != RBD_WATCH_STATE_ERROR)
3940                 goto fail_unlock;
3941
3942         ret = __rbd_register_watch(rbd_dev);
3943         if (ret) {
3944                 rbd_warn(rbd_dev, "failed to reregister watch: %d", ret);
3945                 if (ret != -EBLACKLISTED)
3946                         queue_delayed_work(rbd_dev->task_wq,
3947                                            &rbd_dev->watch_dwork,
3948                                            RBD_RETRY_DELAY);
3949                 goto fail_unlock;
3950         }
3951
3952         rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
3953         rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
3954         mutex_unlock(&rbd_dev->watch_mutex);
3955
3956         ret = rbd_dev_refresh(rbd_dev);
3957         if (ret)
3958                 rbd_warn(rbd_dev, "reregisteration refresh failed: %d", ret);
3959
3960         if (was_lock_owner) {
3961                 ret = rbd_try_lock(rbd_dev);
3962                 if (ret)
3963                         rbd_warn(rbd_dev, "reregisteration lock failed: %d",
3964                                  ret);
3965         }
3966
3967         up_write(&rbd_dev->lock_rwsem);
3968         wake_requests(rbd_dev, true);
3969         return;
3970
3971 fail_unlock:
3972         mutex_unlock(&rbd_dev->watch_mutex);
3973         up_write(&rbd_dev->lock_rwsem);
3974 }
3975
3976 /*
3977  * Synchronous osd object method call.  Returns the number of bytes
3978  * returned in the outbound buffer, or a negative error code.
3979  */
3980 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
3981                              const char *object_name,
3982                              const char *class_name,
3983                              const char *method_name,
3984                              const void *outbound,
3985                              size_t outbound_size,
3986                              void *inbound,
3987                              size_t inbound_size)
3988 {
3989         struct rbd_obj_request *obj_request;
3990         struct page **pages;
3991         u32 page_count;
3992         int ret;
3993
3994         /*
3995          * Method calls are ultimately read operations.  The result
3996          * should placed into the inbound buffer provided.  They
3997          * also supply outbound data--parameters for the object
3998          * method.  Currently if this is present it will be a
3999          * snapshot id.
4000          */
4001         page_count = (u32)calc_pages_for(0, inbound_size);
4002         pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
4003         if (IS_ERR(pages))
4004                 return PTR_ERR(pages);
4005
4006         ret = -ENOMEM;
4007         obj_request = rbd_obj_request_create(object_name, 0, inbound_size,
4008                                                         OBJ_REQUEST_PAGES);
4009         if (!obj_request)
4010                 goto out;
4011
4012         obj_request->pages = pages;
4013         obj_request->page_count = page_count;
4014
4015         obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
4016                                                   obj_request);
4017         if (!obj_request->osd_req)
4018                 goto out;
4019
4020         osd_req_op_cls_init(obj_request->osd_req, 0, CEPH_OSD_OP_CALL,
4021                                         class_name, method_name);
4022         if (outbound_size) {
4023                 struct ceph_pagelist *pagelist;
4024
4025                 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
4026                 if (!pagelist)
4027                         goto out;
4028
4029                 ceph_pagelist_init(pagelist);
4030                 ceph_pagelist_append(pagelist, outbound, outbound_size);
4031                 osd_req_op_cls_request_data_pagelist(obj_request->osd_req, 0,
4032                                                 pagelist);
4033         }
4034         osd_req_op_cls_response_data_pages(obj_request->osd_req, 0,
4035                                         obj_request->pages, inbound_size,
4036                                         0, false, false);
4037         rbd_osd_req_format_read(obj_request);
4038
4039         rbd_obj_request_submit(obj_request);
4040         ret = rbd_obj_request_wait(obj_request);
4041         if (ret)
4042                 goto out;
4043
4044         ret = obj_request->result;
4045         if (ret < 0)
4046                 goto out;
4047
4048         rbd_assert(obj_request->xferred < (u64)INT_MAX);
4049         ret = (int)obj_request->xferred;
4050         ceph_copy_from_page_vector(pages, inbound, 0, obj_request->xferred);
4051 out:
4052         if (obj_request)
4053                 rbd_obj_request_put(obj_request);
4054         else
4055                 ceph_release_page_vector(pages, page_count);
4056
4057         return ret;
4058 }
4059
4060 /*
4061  * lock_rwsem must be held for read
4062  */
4063 static void rbd_wait_state_locked(struct rbd_device *rbd_dev)
4064 {
4065         DEFINE_WAIT(wait);
4066
4067         do {
4068                 /*
4069                  * Note the use of mod_delayed_work() in rbd_acquire_lock()
4070                  * and cancel_delayed_work() in wake_requests().
4071                  */
4072                 dout("%s rbd_dev %p queueing lock_dwork\n", __func__, rbd_dev);
4073                 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork, 0);
4074                 prepare_to_wait_exclusive(&rbd_dev->lock_waitq, &wait,
4075                                           TASK_UNINTERRUPTIBLE);
4076                 up_read(&rbd_dev->lock_rwsem);
4077                 schedule();
4078                 down_read(&rbd_dev->lock_rwsem);
4079         } while (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED);
4080         finish_wait(&rbd_dev->lock_waitq, &wait);
4081 }
4082
4083 static void rbd_queue_workfn(struct work_struct *work)
4084 {
4085         struct request *rq = blk_mq_rq_from_pdu(work);
4086         struct rbd_device *rbd_dev = rq->q->queuedata;
4087         struct rbd_img_request *img_request;
4088         struct ceph_snap_context *snapc = NULL;
4089         u64 offset = (u64)blk_rq_pos(rq) << SECTOR_SHIFT;
4090         u64 length = blk_rq_bytes(rq);
4091         enum obj_operation_type op_type;
4092         u64 mapping_size;
4093         bool must_be_locked;
4094         int result;
4095
4096         if (rq->cmd_type != REQ_TYPE_FS) {
4097                 dout("%s: non-fs request type %d\n", __func__,
4098                         (int) rq->cmd_type);
4099                 result = -EIO;
4100                 goto err;
4101         }
4102
4103         if (req_op(rq) == REQ_OP_DISCARD)
4104                 op_type = OBJ_OP_DISCARD;
4105         else if (req_op(rq) == REQ_OP_WRITE)
4106                 op_type = OBJ_OP_WRITE;
4107         else
4108                 op_type = OBJ_OP_READ;
4109
4110         /* Ignore/skip any zero-length requests */
4111
4112         if (!length) {
4113                 dout("%s: zero-length request\n", __func__);
4114                 result = 0;
4115                 goto err_rq;
4116         }
4117
4118         /* Only reads are allowed to a read-only device */
4119
4120         if (op_type != OBJ_OP_READ) {
4121                 if (rbd_dev->mapping.read_only) {
4122                         result = -EROFS;
4123                         goto err_rq;
4124                 }
4125                 rbd_assert(rbd_dev->spec->snap_id == CEPH_NOSNAP);
4126         }
4127
4128         /*
4129          * Quit early if the mapped snapshot no longer exists.  It's
4130          * still possible the snapshot will have disappeared by the
4131          * time our request arrives at the osd, but there's no sense in
4132          * sending it if we already know.
4133          */
4134         if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
4135                 dout("request for non-existent snapshot");
4136                 rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
4137                 result = -ENXIO;
4138                 goto err_rq;
4139         }
4140
4141         if (offset && length > U64_MAX - offset + 1) {
4142                 rbd_warn(rbd_dev, "bad request range (%llu~%llu)", offset,
4143                          length);
4144                 result = -EINVAL;
4145                 goto err_rq;    /* Shouldn't happen */
4146         }
4147
4148         blk_mq_start_request(rq);
4149
4150         down_read(&rbd_dev->header_rwsem);
4151         mapping_size = rbd_dev->mapping.size;
4152         if (op_type != OBJ_OP_READ) {
4153                 snapc = rbd_dev->header.snapc;
4154                 ceph_get_snap_context(snapc);
4155                 must_be_locked = rbd_is_lock_supported(rbd_dev);
4156         } else {
4157                 must_be_locked = rbd_dev->opts->lock_on_read &&
4158                                         rbd_is_lock_supported(rbd_dev);
4159         }
4160         up_read(&rbd_dev->header_rwsem);
4161
4162         if (offset + length > mapping_size) {
4163                 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)", offset,
4164                          length, mapping_size);
4165                 result = -EIO;
4166                 goto err_rq;
4167         }
4168
4169         if (must_be_locked) {
4170                 down_read(&rbd_dev->lock_rwsem);
4171                 if (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED)
4172                         rbd_wait_state_locked(rbd_dev);
4173         }
4174
4175         img_request = rbd_img_request_create(rbd_dev, offset, length, op_type,
4176                                              snapc);
4177         if (!img_request) {
4178                 result = -ENOMEM;
4179                 goto err_unlock;
4180         }
4181         img_request->rq = rq;
4182         snapc = NULL; /* img_request consumes a ref */
4183
4184         if (op_type == OBJ_OP_DISCARD)
4185                 result = rbd_img_request_fill(img_request, OBJ_REQUEST_NODATA,
4186                                               NULL);
4187         else
4188                 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
4189                                               rq->bio);
4190         if (result)
4191                 goto err_img_request;
4192
4193         result = rbd_img_request_submit(img_request);
4194         if (result)
4195                 goto err_img_request;
4196
4197         if (must_be_locked)
4198                 up_read(&rbd_dev->lock_rwsem);
4199         return;
4200
4201 err_img_request:
4202         rbd_img_request_put(img_request);
4203 err_unlock:
4204         if (must_be_locked)
4205                 up_read(&rbd_dev->lock_rwsem);
4206 err_rq:
4207         if (result)
4208                 rbd_warn(rbd_dev, "%s %llx at %llx result %d",
4209                          obj_op_name(op_type), length, offset, result);
4210         ceph_put_snap_context(snapc);
4211 err:
4212         blk_mq_end_request(rq, result);
4213 }
4214
4215 static int rbd_queue_rq(struct blk_mq_hw_ctx *hctx,
4216                 const struct blk_mq_queue_data *bd)
4217 {
4218         struct request *rq = bd->rq;
4219         struct work_struct *work = blk_mq_rq_to_pdu(rq);
4220
4221         queue_work(rbd_wq, work);
4222         return BLK_MQ_RQ_QUEUE_OK;
4223 }
4224
4225 static void rbd_free_disk(struct rbd_device *rbd_dev)
4226 {
4227         struct gendisk *disk = rbd_dev->disk;
4228
4229         if (!disk)
4230                 return;
4231
4232         rbd_dev->disk = NULL;
4233         if (disk->flags & GENHD_FL_UP) {
4234                 del_gendisk(disk);
4235                 if (disk->queue)
4236                         blk_cleanup_queue(disk->queue);
4237                 blk_mq_free_tag_set(&rbd_dev->tag_set);
4238         }
4239         put_disk(disk);
4240 }
4241
4242 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
4243                                 const char *object_name,
4244                                 u64 offset, u64 length, void *buf)
4245
4246 {
4247         struct rbd_obj_request *obj_request;
4248         struct page **pages = NULL;
4249         u32 page_count;
4250         size_t size;
4251         int ret;
4252
4253         page_count = (u32) calc_pages_for(offset, length);
4254         pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
4255         if (IS_ERR(pages))
4256                 return PTR_ERR(pages);
4257
4258         ret = -ENOMEM;
4259         obj_request = rbd_obj_request_create(object_name, offset, length,
4260                                                         OBJ_REQUEST_PAGES);
4261         if (!obj_request)
4262                 goto out;
4263
4264         obj_request->pages = pages;
4265         obj_request->page_count = page_count;
4266
4267         obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
4268                                                   obj_request);
4269         if (!obj_request->osd_req)
4270                 goto out;
4271
4272         osd_req_op_extent_init(obj_request->osd_req, 0, CEPH_OSD_OP_READ,
4273                                         offset, length, 0, 0);
4274         osd_req_op_extent_osd_data_pages(obj_request->osd_req, 0,
4275                                         obj_request->pages,
4276                                         obj_request->length,
4277                                         obj_request->offset & ~PAGE_MASK,
4278                                         false, false);
4279         rbd_osd_req_format_read(obj_request);
4280
4281         rbd_obj_request_submit(obj_request);
4282         ret = rbd_obj_request_wait(obj_request);
4283         if (ret)
4284                 goto out;
4285
4286         ret = obj_request->result;
4287         if (ret < 0)
4288                 goto out;
4289
4290         rbd_assert(obj_request->xferred <= (u64) SIZE_MAX);
4291         size = (size_t) obj_request->xferred;
4292         ceph_copy_from_page_vector(pages, buf, 0, size);
4293         rbd_assert(size <= (size_t)INT_MAX);
4294         ret = (int)size;
4295 out:
4296         if (obj_request)
4297                 rbd_obj_request_put(obj_request);
4298         else
4299                 ceph_release_page_vector(pages, page_count);
4300
4301         return ret;
4302 }
4303
4304 /*
4305  * Read the complete header for the given rbd device.  On successful
4306  * return, the rbd_dev->header field will contain up-to-date
4307  * information about the image.
4308  */
4309 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev)
4310 {
4311         struct rbd_image_header_ondisk *ondisk = NULL;
4312         u32 snap_count = 0;
4313         u64 names_size = 0;
4314         u32 want_count;
4315         int ret;
4316
4317         /*
4318          * The complete header will include an array of its 64-bit
4319          * snapshot ids, followed by the names of those snapshots as
4320          * a contiguous block of NUL-terminated strings.  Note that
4321          * the number of snapshots could change by the time we read
4322          * it in, in which case we re-read it.
4323          */
4324         do {
4325                 size_t size;
4326
4327                 kfree(ondisk);
4328
4329                 size = sizeof (*ondisk);
4330                 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
4331                 size += names_size;
4332                 ondisk = kmalloc(size, GFP_KERNEL);
4333                 if (!ondisk)
4334                         return -ENOMEM;
4335
4336                 ret = rbd_obj_read_sync(rbd_dev, rbd_dev->header_oid.name,
4337                                        0, size, ondisk);
4338                 if (ret < 0)
4339                         goto out;
4340                 if ((size_t)ret < size) {
4341                         ret = -ENXIO;
4342                         rbd_warn(rbd_dev, "short header read (want %zd got %d)",
4343                                 size, ret);
4344                         goto out;
4345                 }
4346                 if (!rbd_dev_ondisk_valid(ondisk)) {
4347                         ret = -ENXIO;
4348                         rbd_warn(rbd_dev, "invalid header");
4349                         goto out;
4350                 }
4351
4352                 names_size = le64_to_cpu(ondisk->snap_names_len);
4353                 want_count = snap_count;
4354                 snap_count = le32_to_cpu(ondisk->snap_count);
4355         } while (snap_count != want_count);
4356
4357         ret = rbd_header_from_disk(rbd_dev, ondisk);
4358 out:
4359         kfree(ondisk);
4360
4361         return ret;
4362 }
4363
4364 /*
4365  * Clear the rbd device's EXISTS flag if the snapshot it's mapped to
4366  * has disappeared from the (just updated) snapshot context.
4367  */
4368 static void rbd_exists_validate(struct rbd_device *rbd_dev)
4369 {
4370         u64 snap_id;
4371
4372         if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags))
4373                 return;
4374
4375         snap_id = rbd_dev->spec->snap_id;
4376         if (snap_id == CEPH_NOSNAP)
4377                 return;
4378
4379         if (rbd_dev_snap_index(rbd_dev, snap_id) == BAD_SNAP_INDEX)
4380                 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
4381 }
4382
4383 static void rbd_dev_update_size(struct rbd_device *rbd_dev)
4384 {
4385         sector_t size;
4386
4387         /*
4388          * If EXISTS is not set, rbd_dev->disk may be NULL, so don't
4389          * try to update its size.  If REMOVING is set, updating size
4390          * is just useless work since the device can't be opened.
4391          */
4392         if (test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags) &&
4393             !test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags)) {
4394                 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
4395                 dout("setting size to %llu sectors", (unsigned long long)size);
4396                 set_capacity(rbd_dev->disk, size);
4397                 revalidate_disk(rbd_dev->disk);
4398         }
4399 }
4400
4401 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
4402 {
4403         u64 mapping_size;
4404         int ret;
4405
4406         down_write(&rbd_dev->header_rwsem);
4407         mapping_size = rbd_dev->mapping.size;
4408
4409         ret = rbd_dev_header_info(rbd_dev);
4410         if (ret)
4411                 goto out;
4412
4413         /*
4414          * If there is a parent, see if it has disappeared due to the
4415          * mapped image getting flattened.
4416          */
4417         if (rbd_dev->parent) {
4418                 ret = rbd_dev_v2_parent_info(rbd_dev);
4419                 if (ret)
4420                         goto out;
4421         }
4422
4423         if (rbd_dev->spec->snap_id == CEPH_NOSNAP) {
4424                 rbd_dev->mapping.size = rbd_dev->header.image_size;
4425         } else {
4426                 /* validate mapped snapshot's EXISTS flag */
4427                 rbd_exists_validate(rbd_dev);
4428         }
4429
4430 out:
4431         up_write(&rbd_dev->header_rwsem);
4432         if (!ret && mapping_size != rbd_dev->mapping.size)
4433                 rbd_dev_update_size(rbd_dev);
4434
4435         return ret;
4436 }
4437
4438 static int rbd_init_request(void *data, struct request *rq,
4439                 unsigned int hctx_idx, unsigned int request_idx,
4440                 unsigned int numa_node)
4441 {
4442         struct work_struct *work = blk_mq_rq_to_pdu(rq);
4443
4444         INIT_WORK(work, rbd_queue_workfn);
4445         return 0;
4446 }
4447
4448 static struct blk_mq_ops rbd_mq_ops = {
4449         .queue_rq       = rbd_queue_rq,
4450         .map_queue      = blk_mq_map_queue,
4451         .init_request   = rbd_init_request,
4452 };
4453
4454 static int rbd_init_disk(struct rbd_device *rbd_dev)
4455 {
4456         struct gendisk *disk;
4457         struct request_queue *q;
4458         u64 segment_size;
4459         int err;
4460
4461         /* create gendisk info */
4462         disk = alloc_disk(single_major ?
4463                           (1 << RBD_SINGLE_MAJOR_PART_SHIFT) :
4464                           RBD_MINORS_PER_MAJOR);
4465         if (!disk)
4466                 return -ENOMEM;
4467
4468         snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
4469                  rbd_dev->dev_id);
4470         disk->major = rbd_dev->major;
4471         disk->first_minor = rbd_dev->minor;
4472         if (single_major)
4473                 disk->flags |= GENHD_FL_EXT_DEVT;
4474         disk->fops = &rbd_bd_ops;
4475         disk->private_data = rbd_dev;
4476
4477         memset(&rbd_dev->tag_set, 0, sizeof(rbd_dev->tag_set));
4478         rbd_dev->tag_set.ops = &rbd_mq_ops;
4479         rbd_dev->tag_set.queue_depth = rbd_dev->opts->queue_depth;
4480         rbd_dev->tag_set.numa_node = NUMA_NO_NODE;
4481         rbd_dev->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
4482         rbd_dev->tag_set.nr_hw_queues = 1;
4483         rbd_dev->tag_set.cmd_size = sizeof(struct work_struct);
4484
4485         err = blk_mq_alloc_tag_set(&rbd_dev->tag_set);
4486         if (err)
4487                 goto out_disk;
4488
4489         q = blk_mq_init_queue(&rbd_dev->tag_set);
4490         if (IS_ERR(q)) {
4491                 err = PTR_ERR(q);
4492                 goto out_tag_set;
4493         }
4494
4495         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
4496         /* QUEUE_FLAG_ADD_RANDOM is off by default for blk-mq */
4497
4498         /* set io sizes to object size */
4499         segment_size = rbd_obj_bytes(&rbd_dev->header);
4500         blk_queue_max_hw_sectors(q, segment_size / SECTOR_SIZE);
4501         q->limits.max_sectors = queue_max_hw_sectors(q);
4502         blk_queue_max_segments(q, segment_size / SECTOR_SIZE);
4503         blk_queue_max_segment_size(q, segment_size);
4504         blk_queue_io_min(q, segment_size);
4505         blk_queue_io_opt(q, segment_size);
4506
4507         /* enable the discard support */
4508         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
4509         q->limits.discard_granularity = segment_size;
4510         q->limits.discard_alignment = segment_size;
4511         blk_queue_max_discard_sectors(q, segment_size / SECTOR_SIZE);
4512         q->limits.discard_zeroes_data = 1;
4513
4514         if (!ceph_test_opt(rbd_dev->rbd_client->client, NOCRC))
4515                 q->backing_dev_info.capabilities |= BDI_CAP_STABLE_WRITES;
4516
4517         disk->queue = q;
4518
4519         q->queuedata = rbd_dev;
4520
4521         rbd_dev->disk = disk;
4522
4523         return 0;
4524 out_tag_set:
4525         blk_mq_free_tag_set(&rbd_dev->tag_set);
4526 out_disk:
4527         put_disk(disk);
4528         return err;
4529 }
4530
4531 /*
4532   sysfs
4533 */
4534
4535 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
4536 {
4537         return container_of(dev, struct rbd_device, dev);
4538 }
4539
4540 static ssize_t rbd_size_show(struct device *dev,
4541                              struct device_attribute *attr, char *buf)
4542 {
4543         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4544
4545         return sprintf(buf, "%llu\n",
4546                 (unsigned long long)rbd_dev->mapping.size);
4547 }
4548
4549 /*
4550  * Note this shows the features for whatever's mapped, which is not
4551  * necessarily the base image.
4552  */
4553 static ssize_t rbd_features_show(struct device *dev,
4554                              struct device_attribute *attr, char *buf)
4555 {
4556         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4557
4558         return sprintf(buf, "0x%016llx\n",
4559                         (unsigned long long)rbd_dev->mapping.features);
4560 }
4561
4562 static ssize_t rbd_major_show(struct device *dev,
4563                               struct device_attribute *attr, char *buf)
4564 {
4565         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4566
4567         if (rbd_dev->major)
4568                 return sprintf(buf, "%d\n", rbd_dev->major);
4569
4570         return sprintf(buf, "(none)\n");
4571 }
4572
4573 static ssize_t rbd_minor_show(struct device *dev,
4574                               struct device_attribute *attr, char *buf)
4575 {
4576         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4577
4578         return sprintf(buf, "%d\n", rbd_dev->minor);
4579 }
4580
4581 static ssize_t rbd_client_addr_show(struct device *dev,
4582                                     struct device_attribute *attr, char *buf)
4583 {
4584         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4585         struct ceph_entity_addr *client_addr =
4586             ceph_client_addr(rbd_dev->rbd_client->client);
4587
4588         return sprintf(buf, "%pISpc/%u\n", &client_addr->in_addr,
4589                        le32_to_cpu(client_addr->nonce));
4590 }
4591
4592 static ssize_t rbd_client_id_show(struct device *dev,
4593                                   struct device_attribute *attr, char *buf)
4594 {
4595         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4596
4597         return sprintf(buf, "client%lld\n",
4598                        ceph_client_gid(rbd_dev->rbd_client->client));
4599 }
4600
4601 static ssize_t rbd_cluster_fsid_show(struct device *dev,
4602                                      struct device_attribute *attr, char *buf)
4603 {
4604         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4605
4606         return sprintf(buf, "%pU\n", &rbd_dev->rbd_client->client->fsid);
4607 }
4608
4609 static ssize_t rbd_config_info_show(struct device *dev,
4610                                     struct device_attribute *attr, char *buf)
4611 {
4612         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4613
4614         return sprintf(buf, "%s\n", rbd_dev->config_info);
4615 }
4616
4617 static ssize_t rbd_pool_show(struct device *dev,
4618                              struct device_attribute *attr, char *buf)
4619 {
4620         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4621
4622         return sprintf(buf, "%s\n", rbd_dev->spec->pool_name);
4623 }
4624
4625 static ssize_t rbd_pool_id_show(struct device *dev,
4626                              struct device_attribute *attr, char *buf)
4627 {
4628         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4629
4630         return sprintf(buf, "%llu\n",
4631                         (unsigned long long) rbd_dev->spec->pool_id);
4632 }
4633
4634 static ssize_t rbd_name_show(struct device *dev,
4635                              struct device_attribute *attr, char *buf)
4636 {
4637         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4638
4639         if (rbd_dev->spec->image_name)
4640                 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
4641
4642         return sprintf(buf, "(unknown)\n");
4643 }
4644
4645 static ssize_t rbd_image_id_show(struct device *dev,
4646                              struct device_attribute *attr, char *buf)
4647 {
4648         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4649
4650         return sprintf(buf, "%s\n", rbd_dev->spec->image_id);
4651 }
4652
4653 /*
4654  * Shows the name of the currently-mapped snapshot (or
4655  * RBD_SNAP_HEAD_NAME for the base image).
4656  */
4657 static ssize_t rbd_snap_show(struct device *dev,
4658                              struct device_attribute *attr,
4659                              char *buf)
4660 {
4661         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4662
4663         return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
4664 }
4665
4666 static ssize_t rbd_snap_id_show(struct device *dev,
4667                                 struct device_attribute *attr, char *buf)
4668 {
4669         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4670
4671         return sprintf(buf, "%llu\n", rbd_dev->spec->snap_id);
4672 }
4673
4674 /*
4675  * For a v2 image, shows the chain of parent images, separated by empty
4676  * lines.  For v1 images or if there is no parent, shows "(no parent
4677  * image)".
4678  */
4679 static ssize_t rbd_parent_show(struct device *dev,
4680                                struct device_attribute *attr,
4681                                char *buf)
4682 {
4683         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4684         ssize_t count = 0;
4685
4686         if (!rbd_dev->parent)
4687                 return sprintf(buf, "(no parent image)\n");
4688
4689         for ( ; rbd_dev->parent; rbd_dev = rbd_dev->parent) {
4690                 struct rbd_spec *spec = rbd_dev->parent_spec;
4691
4692                 count += sprintf(&buf[count], "%s"
4693                             "pool_id %llu\npool_name %s\n"
4694                             "image_id %s\nimage_name %s\n"
4695                             "snap_id %llu\nsnap_name %s\n"
4696                             "overlap %llu\n",
4697                             !count ? "" : "\n", /* first? */
4698                             spec->pool_id, spec->pool_name,
4699                             spec->image_id, spec->image_name ?: "(unknown)",
4700                             spec->snap_id, spec->snap_name,
4701                             rbd_dev->parent_overlap);
4702         }
4703
4704         return count;
4705 }
4706
4707 static ssize_t rbd_image_refresh(struct device *dev,
4708                                  struct device_attribute *attr,
4709                                  const char *buf,
4710                                  size_t size)
4711 {
4712         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4713         int ret;
4714
4715         ret = rbd_dev_refresh(rbd_dev);
4716         if (ret)
4717                 return ret;
4718
4719         return size;
4720 }
4721
4722 static DEVICE_ATTR(size, S_IRUGO, rbd_size_show, NULL);
4723 static DEVICE_ATTR(features, S_IRUGO, rbd_features_show, NULL);
4724 static DEVICE_ATTR(major, S_IRUGO, rbd_major_show, NULL);
4725 static DEVICE_ATTR(minor, S_IRUGO, rbd_minor_show, NULL);
4726 static DEVICE_ATTR(client_addr, S_IRUGO, rbd_client_addr_show, NULL);
4727 static DEVICE_ATTR(client_id, S_IRUGO, rbd_client_id_show, NULL);
4728 static DEVICE_ATTR(cluster_fsid, S_IRUGO, rbd_cluster_fsid_show, NULL);
4729 static DEVICE_ATTR(config_info, S_IRUSR, rbd_config_info_show, NULL);
4730 static DEVICE_ATTR(pool, S_IRUGO, rbd_pool_show, NULL);
4731 static DEVICE_ATTR(pool_id, S_IRUGO, rbd_pool_id_show, NULL);
4732 static DEVICE_ATTR(name, S_IRUGO, rbd_name_show, NULL);
4733 static DEVICE_ATTR(image_id, S_IRUGO, rbd_image_id_show, NULL);
4734 static DEVICE_ATTR(refresh, S_IWUSR, NULL, rbd_image_refresh);
4735 static DEVICE_ATTR(current_snap, S_IRUGO, rbd_snap_show, NULL);
4736 static DEVICE_ATTR(snap_id, S_IRUGO, rbd_snap_id_show, NULL);
4737 static DEVICE_ATTR(parent, S_IRUGO, rbd_parent_show, NULL);
4738
4739 static struct attribute *rbd_attrs[] = {
4740         &dev_attr_size.attr,
4741         &dev_attr_features.attr,
4742         &dev_attr_major.attr,
4743         &dev_attr_minor.attr,
4744         &dev_attr_client_addr.attr,
4745         &dev_attr_client_id.attr,
4746         &dev_attr_cluster_fsid.attr,
4747         &dev_attr_config_info.attr,
4748         &dev_attr_pool.attr,
4749         &dev_attr_pool_id.attr,
4750         &dev_attr_name.attr,
4751         &dev_attr_image_id.attr,
4752         &dev_attr_current_snap.attr,
4753         &dev_attr_snap_id.attr,
4754         &dev_attr_parent.attr,
4755         &dev_attr_refresh.attr,
4756         NULL
4757 };
4758
4759 static struct attribute_group rbd_attr_group = {
4760         .attrs = rbd_attrs,
4761 };
4762
4763 static const struct attribute_group *rbd_attr_groups[] = {
4764         &rbd_attr_group,
4765         NULL
4766 };
4767
4768 static void rbd_dev_release(struct device *dev);
4769
4770 static struct device_type rbd_device_type = {
4771         .name           = "rbd",
4772         .groups         = rbd_attr_groups,
4773         .release        = rbd_dev_release,
4774 };
4775
4776 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
4777 {
4778         kref_get(&spec->kref);
4779
4780         return spec;
4781 }
4782
4783 static void rbd_spec_free(struct kref *kref);
4784 static void rbd_spec_put(struct rbd_spec *spec)
4785 {
4786         if (spec)
4787                 kref_put(&spec->kref, rbd_spec_free);
4788 }
4789
4790 static struct rbd_spec *rbd_spec_alloc(void)
4791 {
4792         struct rbd_spec *spec;
4793
4794         spec = kzalloc(sizeof (*spec), GFP_KERNEL);
4795         if (!spec)
4796                 return NULL;
4797
4798         spec->pool_id = CEPH_NOPOOL;
4799         spec->snap_id = CEPH_NOSNAP;
4800         kref_init(&spec->kref);
4801
4802         return spec;
4803 }
4804
4805 static void rbd_spec_free(struct kref *kref)
4806 {
4807         struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
4808
4809         kfree(spec->pool_name);
4810         kfree(spec->image_id);
4811         kfree(spec->image_name);
4812         kfree(spec->snap_name);
4813         kfree(spec);
4814 }
4815
4816 static void rbd_dev_free(struct rbd_device *rbd_dev)
4817 {
4818         WARN_ON(rbd_dev->watch_state != RBD_WATCH_STATE_UNREGISTERED);
4819         WARN_ON(rbd_dev->lock_state != RBD_LOCK_STATE_UNLOCKED);
4820
4821         ceph_oid_destroy(&rbd_dev->header_oid);
4822         ceph_oloc_destroy(&rbd_dev->header_oloc);
4823         kfree(rbd_dev->config_info);
4824
4825         rbd_put_client(rbd_dev->rbd_client);
4826         rbd_spec_put(rbd_dev->spec);
4827         kfree(rbd_dev->opts);
4828         kfree(rbd_dev);
4829 }
4830
4831 static void rbd_dev_release(struct device *dev)
4832 {
4833         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4834         bool need_put = !!rbd_dev->opts;
4835
4836         if (need_put) {
4837                 destroy_workqueue(rbd_dev->task_wq);
4838                 ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4839         }
4840
4841         rbd_dev_free(rbd_dev);
4842
4843         /*
4844          * This is racy, but way better than putting module outside of
4845          * the release callback.  The race window is pretty small, so
4846          * doing something similar to dm (dm-builtin.c) is overkill.
4847          */
4848         if (need_put)
4849                 module_put(THIS_MODULE);
4850 }
4851
4852 static struct rbd_device *__rbd_dev_create(struct rbd_client *rbdc,
4853                                            struct rbd_spec *spec)
4854 {
4855         struct rbd_device *rbd_dev;
4856
4857         rbd_dev = kzalloc(sizeof(*rbd_dev), GFP_KERNEL);
4858         if (!rbd_dev)
4859                 return NULL;
4860
4861         spin_lock_init(&rbd_dev->lock);
4862         INIT_LIST_HEAD(&rbd_dev->node);
4863         init_rwsem(&rbd_dev->header_rwsem);
4864
4865         ceph_oid_init(&rbd_dev->header_oid);
4866         ceph_oloc_init(&rbd_dev->header_oloc);
4867
4868         mutex_init(&rbd_dev->watch_mutex);
4869         rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
4870         INIT_DELAYED_WORK(&rbd_dev->watch_dwork, rbd_reregister_watch);
4871
4872         init_rwsem(&rbd_dev->lock_rwsem);
4873         rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
4874         INIT_WORK(&rbd_dev->acquired_lock_work, rbd_notify_acquired_lock);
4875         INIT_WORK(&rbd_dev->released_lock_work, rbd_notify_released_lock);
4876         INIT_DELAYED_WORK(&rbd_dev->lock_dwork, rbd_acquire_lock);
4877         INIT_WORK(&rbd_dev->unlock_work, rbd_release_lock_work);
4878         init_waitqueue_head(&rbd_dev->lock_waitq);
4879
4880         rbd_dev->dev.bus = &rbd_bus_type;
4881         rbd_dev->dev.type = &rbd_device_type;
4882         rbd_dev->dev.parent = &rbd_root_dev;
4883         device_initialize(&rbd_dev->dev);
4884
4885         rbd_dev->rbd_client = rbdc;
4886         rbd_dev->spec = spec;
4887
4888         rbd_dev->layout.stripe_unit = 1 << RBD_MAX_OBJ_ORDER;
4889         rbd_dev->layout.stripe_count = 1;
4890         rbd_dev->layout.object_size = 1 << RBD_MAX_OBJ_ORDER;
4891         rbd_dev->layout.pool_id = spec->pool_id;
4892         RCU_INIT_POINTER(rbd_dev->layout.pool_ns, NULL);
4893
4894         return rbd_dev;
4895 }
4896
4897 /*
4898  * Create a mapping rbd_dev.
4899  */
4900 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
4901                                          struct rbd_spec *spec,
4902                                          struct rbd_options *opts)
4903 {
4904         struct rbd_device *rbd_dev;
4905
4906         rbd_dev = __rbd_dev_create(rbdc, spec);
4907         if (!rbd_dev)
4908                 return NULL;
4909
4910         rbd_dev->opts = opts;
4911
4912         /* get an id and fill in device name */
4913         rbd_dev->dev_id = ida_simple_get(&rbd_dev_id_ida, 0,
4914                                          minor_to_rbd_dev_id(1 << MINORBITS),
4915                                          GFP_KERNEL);
4916         if (rbd_dev->dev_id < 0)
4917                 goto fail_rbd_dev;
4918
4919         sprintf(rbd_dev->name, RBD_DRV_NAME "%d", rbd_dev->dev_id);
4920         rbd_dev->task_wq = alloc_ordered_workqueue("%s-tasks", WQ_MEM_RECLAIM,
4921                                                    rbd_dev->name);
4922         if (!rbd_dev->task_wq)
4923                 goto fail_dev_id;
4924
4925         /* we have a ref from do_rbd_add() */
4926         __module_get(THIS_MODULE);
4927
4928         dout("%s rbd_dev %p dev_id %d\n", __func__, rbd_dev, rbd_dev->dev_id);
4929         return rbd_dev;
4930
4931 fail_dev_id:
4932         ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4933 fail_rbd_dev:
4934         rbd_dev_free(rbd_dev);
4935         return NULL;
4936 }
4937
4938 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
4939 {
4940         if (rbd_dev)
4941                 put_device(&rbd_dev->dev);
4942 }
4943
4944 /*
4945  * Get the size and object order for an image snapshot, or if
4946  * snap_id is CEPH_NOSNAP, gets this information for the base
4947  * image.
4948  */
4949 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
4950                                 u8 *order, u64 *snap_size)
4951 {
4952         __le64 snapid = cpu_to_le64(snap_id);
4953         int ret;
4954         struct {
4955                 u8 order;
4956                 __le64 size;
4957         } __attribute__ ((packed)) size_buf = { 0 };
4958
4959         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
4960                                 "rbd", "get_size",
4961                                 &snapid, sizeof (snapid),
4962                                 &size_buf, sizeof (size_buf));
4963         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4964         if (ret < 0)
4965                 return ret;
4966         if (ret < sizeof (size_buf))
4967                 return -ERANGE;
4968
4969         if (order) {
4970                 *order = size_buf.order;
4971                 dout("  order %u", (unsigned int)*order);
4972         }
4973         *snap_size = le64_to_cpu(size_buf.size);
4974
4975         dout("  snap_id 0x%016llx snap_size = %llu\n",
4976                 (unsigned long long)snap_id,
4977                 (unsigned long long)*snap_size);
4978
4979         return 0;
4980 }
4981
4982 static int rbd_dev_v2_image_size(struct rbd_device *rbd_dev)
4983 {
4984         return _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
4985                                         &rbd_dev->header.obj_order,
4986                                         &rbd_dev->header.image_size);
4987 }
4988
4989 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev)
4990 {
4991         void *reply_buf;
4992         int ret;
4993         void *p;
4994
4995         reply_buf = kzalloc(RBD_OBJ_PREFIX_LEN_MAX, GFP_KERNEL);
4996         if (!reply_buf)
4997                 return -ENOMEM;
4998
4999         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5000                                 "rbd", "get_object_prefix", NULL, 0,
5001                                 reply_buf, RBD_OBJ_PREFIX_LEN_MAX);
5002         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5003         if (ret < 0)
5004                 goto out;
5005
5006         p = reply_buf;
5007         rbd_dev->header.object_prefix = ceph_extract_encoded_string(&p,
5008                                                 p + ret, NULL, GFP_NOIO);
5009         ret = 0;
5010
5011         if (IS_ERR(rbd_dev->header.object_prefix)) {
5012                 ret = PTR_ERR(rbd_dev->header.object_prefix);
5013                 rbd_dev->header.object_prefix = NULL;
5014         } else {
5015                 dout("  object_prefix = %s\n", rbd_dev->header.object_prefix);
5016         }
5017 out:
5018         kfree(reply_buf);
5019
5020         return ret;
5021 }
5022
5023 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
5024                 u64 *snap_features)
5025 {
5026         __le64 snapid = cpu_to_le64(snap_id);
5027         struct {
5028                 __le64 features;
5029                 __le64 incompat;
5030         } __attribute__ ((packed)) features_buf = { 0 };
5031         u64 unsup;
5032         int ret;
5033
5034         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5035                                 "rbd", "get_features",
5036                                 &snapid, sizeof (snapid),
5037                                 &features_buf, sizeof (features_buf));
5038         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5039         if (ret < 0)
5040                 return ret;
5041         if (ret < sizeof (features_buf))
5042                 return -ERANGE;
5043
5044         unsup = le64_to_cpu(features_buf.incompat) & ~RBD_FEATURES_SUPPORTED;
5045         if (unsup) {
5046                 rbd_warn(rbd_dev, "image uses unsupported features: 0x%llx",
5047                          unsup);
5048                 return -ENXIO;
5049         }
5050
5051         *snap_features = le64_to_cpu(features_buf.features);
5052
5053         dout("  snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
5054                 (unsigned long long)snap_id,
5055                 (unsigned long long)*snap_features,
5056                 (unsigned long long)le64_to_cpu(features_buf.incompat));
5057
5058         return 0;
5059 }
5060
5061 static int rbd_dev_v2_features(struct rbd_device *rbd_dev)
5062 {
5063         return _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
5064                                                 &rbd_dev->header.features);
5065 }
5066
5067 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev)
5068 {
5069         struct rbd_spec *parent_spec;
5070         size_t size;
5071         void *reply_buf = NULL;
5072         __le64 snapid;
5073         void *p;
5074         void *end;
5075         u64 pool_id;
5076         char *image_id;
5077         u64 snap_id;
5078         u64 overlap;
5079         int ret;
5080
5081         parent_spec = rbd_spec_alloc();
5082         if (!parent_spec)
5083                 return -ENOMEM;
5084
5085         size = sizeof (__le64) +                                /* pool_id */
5086                 sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX +        /* image_id */
5087                 sizeof (__le64) +                               /* snap_id */
5088                 sizeof (__le64);                                /* overlap */
5089         reply_buf = kmalloc(size, GFP_KERNEL);
5090         if (!reply_buf) {
5091                 ret = -ENOMEM;
5092                 goto out_err;
5093         }
5094
5095         snapid = cpu_to_le64(rbd_dev->spec->snap_id);
5096         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5097                                 "rbd", "get_parent",
5098                                 &snapid, sizeof (snapid),
5099                                 reply_buf, size);
5100         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5101         if (ret < 0)
5102                 goto out_err;
5103
5104         p = reply_buf;
5105         end = reply_buf + ret;
5106         ret = -ERANGE;
5107         ceph_decode_64_safe(&p, end, pool_id, out_err);
5108         if (pool_id == CEPH_NOPOOL) {
5109                 /*
5110                  * Either the parent never existed, or we have
5111                  * record of it but the image got flattened so it no
5112                  * longer has a parent.  When the parent of a
5113                  * layered image disappears we immediately set the
5114                  * overlap to 0.  The effect of this is that all new
5115                  * requests will be treated as if the image had no
5116                  * parent.
5117                  */
5118                 if (rbd_dev->parent_overlap) {
5119                         rbd_dev->parent_overlap = 0;
5120                         rbd_dev_parent_put(rbd_dev);
5121                         pr_info("%s: clone image has been flattened\n",
5122                                 rbd_dev->disk->disk_name);
5123                 }
5124
5125                 goto out;       /* No parent?  No problem. */
5126         }
5127
5128         /* The ceph file layout needs to fit pool id in 32 bits */
5129
5130         ret = -EIO;
5131         if (pool_id > (u64)U32_MAX) {
5132                 rbd_warn(NULL, "parent pool id too large (%llu > %u)",
5133                         (unsigned long long)pool_id, U32_MAX);
5134                 goto out_err;
5135         }
5136
5137         image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
5138         if (IS_ERR(image_id)) {
5139                 ret = PTR_ERR(image_id);
5140                 goto out_err;
5141         }
5142         ceph_decode_64_safe(&p, end, snap_id, out_err);
5143         ceph_decode_64_safe(&p, end, overlap, out_err);
5144
5145         /*
5146          * The parent won't change (except when the clone is
5147          * flattened, already handled that).  So we only need to
5148          * record the parent spec we have not already done so.
5149          */
5150         if (!rbd_dev->parent_spec) {
5151                 parent_spec->pool_id = pool_id;
5152                 parent_spec->image_id = image_id;
5153                 parent_spec->snap_id = snap_id;
5154                 rbd_dev->parent_spec = parent_spec;
5155                 parent_spec = NULL;     /* rbd_dev now owns this */
5156         } else {
5157                 kfree(image_id);
5158         }
5159
5160         /*
5161          * We always update the parent overlap.  If it's zero we issue
5162          * a warning, as we will proceed as if there was no parent.
5163          */
5164         if (!overlap) {
5165                 if (parent_spec) {
5166                         /* refresh, careful to warn just once */
5167                         if (rbd_dev->parent_overlap)
5168                                 rbd_warn(rbd_dev,
5169                                     "clone now standalone (overlap became 0)");
5170                 } else {
5171                         /* initial probe */
5172                         rbd_warn(rbd_dev, "clone is standalone (overlap 0)");
5173                 }
5174         }
5175         rbd_dev->parent_overlap = overlap;
5176
5177 out:
5178         ret = 0;
5179 out_err:
5180         kfree(reply_buf);
5181         rbd_spec_put(parent_spec);
5182
5183         return ret;
5184 }
5185
5186 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev)
5187 {
5188         struct {
5189                 __le64 stripe_unit;
5190                 __le64 stripe_count;
5191         } __attribute__ ((packed)) striping_info_buf = { 0 };
5192         size_t size = sizeof (striping_info_buf);
5193         void *p;
5194         u64 obj_size;
5195         u64 stripe_unit;
5196         u64 stripe_count;
5197         int ret;
5198
5199         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5200                                 "rbd", "get_stripe_unit_count", NULL, 0,
5201                                 (char *)&striping_info_buf, size);
5202         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5203         if (ret < 0)
5204                 return ret;
5205         if (ret < size)
5206                 return -ERANGE;
5207
5208         /*
5209          * We don't actually support the "fancy striping" feature
5210          * (STRIPINGV2) yet, but if the striping sizes are the
5211          * defaults the behavior is the same as before.  So find
5212          * out, and only fail if the image has non-default values.
5213          */
5214         ret = -EINVAL;
5215         obj_size = (u64)1 << rbd_dev->header.obj_order;
5216         p = &striping_info_buf;
5217         stripe_unit = ceph_decode_64(&p);
5218         if (stripe_unit != obj_size) {
5219                 rbd_warn(rbd_dev, "unsupported stripe unit "
5220                                 "(got %llu want %llu)",
5221                                 stripe_unit, obj_size);
5222                 return -EINVAL;
5223         }
5224         stripe_count = ceph_decode_64(&p);
5225         if (stripe_count != 1) {
5226                 rbd_warn(rbd_dev, "unsupported stripe count "
5227                                 "(got %llu want 1)", stripe_count);
5228                 return -EINVAL;
5229         }
5230         rbd_dev->header.stripe_unit = stripe_unit;
5231         rbd_dev->header.stripe_count = stripe_count;
5232
5233         return 0;
5234 }
5235
5236 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
5237 {
5238         size_t image_id_size;
5239         char *image_id;
5240         void *p;
5241         void *end;
5242         size_t size;
5243         void *reply_buf = NULL;
5244         size_t len = 0;
5245         char *image_name = NULL;
5246         int ret;
5247
5248         rbd_assert(!rbd_dev->spec->image_name);
5249
5250         len = strlen(rbd_dev->spec->image_id);
5251         image_id_size = sizeof (__le32) + len;
5252         image_id = kmalloc(image_id_size, GFP_KERNEL);
5253         if (!image_id)
5254                 return NULL;
5255
5256         p = image_id;
5257         end = image_id + image_id_size;
5258         ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
5259
5260         size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
5261         reply_buf = kmalloc(size, GFP_KERNEL);
5262         if (!reply_buf)
5263                 goto out;
5264
5265         ret = rbd_obj_method_sync(rbd_dev, RBD_DIRECTORY,
5266                                 "rbd", "dir_get_name",
5267                                 image_id, image_id_size,
5268                                 reply_buf, size);
5269         if (ret < 0)
5270                 goto out;
5271         p = reply_buf;
5272         end = reply_buf + ret;
5273
5274         image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
5275         if (IS_ERR(image_name))
5276                 image_name = NULL;
5277         else
5278                 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
5279 out:
5280         kfree(reply_buf);
5281         kfree(image_id);
5282
5283         return image_name;
5284 }
5285
5286 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5287 {
5288         struct ceph_snap_context *snapc = rbd_dev->header.snapc;
5289         const char *snap_name;
5290         u32 which = 0;
5291
5292         /* Skip over names until we find the one we are looking for */
5293
5294         snap_name = rbd_dev->header.snap_names;
5295         while (which < snapc->num_snaps) {
5296                 if (!strcmp(name, snap_name))
5297                         return snapc->snaps[which];
5298                 snap_name += strlen(snap_name) + 1;
5299                 which++;
5300         }
5301         return CEPH_NOSNAP;
5302 }
5303
5304 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5305 {
5306         struct ceph_snap_context *snapc = rbd_dev->header.snapc;
5307         u32 which;
5308         bool found = false;
5309         u64 snap_id;
5310
5311         for (which = 0; !found && which < snapc->num_snaps; which++) {
5312                 const char *snap_name;
5313
5314                 snap_id = snapc->snaps[which];
5315                 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
5316                 if (IS_ERR(snap_name)) {
5317                         /* ignore no-longer existing snapshots */
5318                         if (PTR_ERR(snap_name) == -ENOENT)
5319                                 continue;
5320                         else
5321                                 break;
5322                 }
5323                 found = !strcmp(name, snap_name);
5324                 kfree(snap_name);
5325         }
5326         return found ? snap_id : CEPH_NOSNAP;
5327 }
5328
5329 /*
5330  * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
5331  * no snapshot by that name is found, or if an error occurs.
5332  */
5333 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5334 {
5335         if (rbd_dev->image_format == 1)
5336                 return rbd_v1_snap_id_by_name(rbd_dev, name);
5337
5338         return rbd_v2_snap_id_by_name(rbd_dev, name);
5339 }
5340
5341 /*
5342  * An image being mapped will have everything but the snap id.
5343  */
5344 static int rbd_spec_fill_snap_id(struct rbd_device *rbd_dev)
5345 {
5346         struct rbd_spec *spec = rbd_dev->spec;
5347
5348         rbd_assert(spec->pool_id != CEPH_NOPOOL && spec->pool_name);
5349         rbd_assert(spec->image_id && spec->image_name);
5350         rbd_assert(spec->snap_name);
5351
5352         if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
5353                 u64 snap_id;
5354
5355                 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
5356                 if (snap_id == CEPH_NOSNAP)
5357                         return -ENOENT;
5358
5359                 spec->snap_id = snap_id;
5360         } else {
5361                 spec->snap_id = CEPH_NOSNAP;
5362         }
5363
5364         return 0;
5365 }
5366
5367 /*
5368  * A parent image will have all ids but none of the names.
5369  *
5370  * All names in an rbd spec are dynamically allocated.  It's OK if we
5371  * can't figure out the name for an image id.
5372  */
5373 static int rbd_spec_fill_names(struct rbd_device *rbd_dev)
5374 {
5375         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
5376         struct rbd_spec *spec = rbd_dev->spec;
5377         const char *pool_name;
5378         const char *image_name;
5379         const char *snap_name;
5380         int ret;
5381
5382         rbd_assert(spec->pool_id != CEPH_NOPOOL);
5383         rbd_assert(spec->image_id);
5384         rbd_assert(spec->snap_id != CEPH_NOSNAP);
5385
5386         /* Get the pool name; we have to make our own copy of this */
5387
5388         pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
5389         if (!pool_name) {
5390                 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
5391                 return -EIO;
5392         }
5393         pool_name = kstrdup(pool_name, GFP_KERNEL);
5394         if (!pool_name)
5395                 return -ENOMEM;
5396
5397         /* Fetch the image name; tolerate failure here */
5398
5399         image_name = rbd_dev_image_name(rbd_dev);
5400         if (!image_name)
5401                 rbd_warn(rbd_dev, "unable to get image name");
5402
5403         /* Fetch the snapshot name */
5404
5405         snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
5406         if (IS_ERR(snap_name)) {
5407                 ret = PTR_ERR(snap_name);
5408                 goto out_err;
5409         }
5410
5411         spec->pool_name = pool_name;
5412         spec->image_name = image_name;
5413         spec->snap_name = snap_name;
5414
5415         return 0;
5416
5417 out_err:
5418         kfree(image_name);
5419         kfree(pool_name);
5420         return ret;
5421 }
5422
5423 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev)
5424 {
5425         size_t size;
5426         int ret;
5427         void *reply_buf;
5428         void *p;
5429         void *end;
5430         u64 seq;
5431         u32 snap_count;
5432         struct ceph_snap_context *snapc;
5433         u32 i;
5434
5435         /*
5436          * We'll need room for the seq value (maximum snapshot id),
5437          * snapshot count, and array of that many snapshot ids.
5438          * For now we have a fixed upper limit on the number we're
5439          * prepared to receive.
5440          */
5441         size = sizeof (__le64) + sizeof (__le32) +
5442                         RBD_MAX_SNAP_COUNT * sizeof (__le64);
5443         reply_buf = kzalloc(size, GFP_KERNEL);
5444         if (!reply_buf)
5445                 return -ENOMEM;
5446
5447         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5448                                 "rbd", "get_snapcontext", NULL, 0,
5449                                 reply_buf, size);
5450         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5451         if (ret < 0)
5452                 goto out;
5453
5454         p = reply_buf;
5455         end = reply_buf + ret;
5456         ret = -ERANGE;
5457         ceph_decode_64_safe(&p, end, seq, out);
5458         ceph_decode_32_safe(&p, end, snap_count, out);
5459
5460         /*
5461          * Make sure the reported number of snapshot ids wouldn't go
5462          * beyond the end of our buffer.  But before checking that,
5463          * make sure the computed size of the snapshot context we
5464          * allocate is representable in a size_t.
5465          */
5466         if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
5467                                  / sizeof (u64)) {
5468                 ret = -EINVAL;
5469                 goto out;
5470         }
5471         if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
5472                 goto out;
5473         ret = 0;
5474
5475         snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
5476         if (!snapc) {
5477                 ret = -ENOMEM;
5478                 goto out;
5479         }
5480         snapc->seq = seq;
5481         for (i = 0; i < snap_count; i++)
5482                 snapc->snaps[i] = ceph_decode_64(&p);
5483
5484         ceph_put_snap_context(rbd_dev->header.snapc);
5485         rbd_dev->header.snapc = snapc;
5486
5487         dout("  snap context seq = %llu, snap_count = %u\n",
5488                 (unsigned long long)seq, (unsigned int)snap_count);
5489 out:
5490         kfree(reply_buf);
5491
5492         return ret;
5493 }
5494
5495 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
5496                                         u64 snap_id)
5497 {
5498         size_t size;
5499         void *reply_buf;
5500         __le64 snapid;
5501         int ret;
5502         void *p;
5503         void *end;
5504         char *snap_name;
5505
5506         size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
5507         reply_buf = kmalloc(size, GFP_KERNEL);
5508         if (!reply_buf)
5509                 return ERR_PTR(-ENOMEM);
5510
5511         snapid = cpu_to_le64(snap_id);
5512         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5513                                 "rbd", "get_snapshot_name",
5514                                 &snapid, sizeof (snapid),
5515                                 reply_buf, size);
5516         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5517         if (ret < 0) {
5518                 snap_name = ERR_PTR(ret);
5519                 goto out;
5520         }
5521
5522         p = reply_buf;
5523         end = reply_buf + ret;
5524         snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
5525         if (IS_ERR(snap_name))
5526                 goto out;
5527
5528         dout("  snap_id 0x%016llx snap_name = %s\n",
5529                 (unsigned long long)snap_id, snap_name);
5530 out:
5531         kfree(reply_buf);
5532
5533         return snap_name;
5534 }
5535
5536 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev)
5537 {
5538         bool first_time = rbd_dev->header.object_prefix == NULL;
5539         int ret;
5540
5541         ret = rbd_dev_v2_image_size(rbd_dev);
5542         if (ret)
5543                 return ret;
5544
5545         if (first_time) {
5546                 ret = rbd_dev_v2_header_onetime(rbd_dev);
5547                 if (ret)
5548                         return ret;
5549         }
5550
5551         ret = rbd_dev_v2_snap_context(rbd_dev);
5552         if (ret && first_time) {
5553                 kfree(rbd_dev->header.object_prefix);
5554                 rbd_dev->header.object_prefix = NULL;
5555         }
5556
5557         return ret;
5558 }
5559
5560 static int rbd_dev_header_info(struct rbd_device *rbd_dev)
5561 {
5562         rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
5563
5564         if (rbd_dev->image_format == 1)
5565                 return rbd_dev_v1_header_info(rbd_dev);
5566
5567         return rbd_dev_v2_header_info(rbd_dev);
5568 }
5569
5570 /*
5571  * Skips over white space at *buf, and updates *buf to point to the
5572  * first found non-space character (if any). Returns the length of
5573  * the token (string of non-white space characters) found.  Note
5574  * that *buf must be terminated with '\0'.
5575  */
5576 static inline size_t next_token(const char **buf)
5577 {
5578         /*
5579         * These are the characters that produce nonzero for
5580         * isspace() in the "C" and "POSIX" locales.
5581         */
5582         const char *spaces = " \f\n\r\t\v";
5583
5584         *buf += strspn(*buf, spaces);   /* Find start of token */
5585
5586         return strcspn(*buf, spaces);   /* Return token length */
5587 }
5588
5589 /*
5590  * Finds the next token in *buf, dynamically allocates a buffer big
5591  * enough to hold a copy of it, and copies the token into the new
5592  * buffer.  The copy is guaranteed to be terminated with '\0'.  Note
5593  * that a duplicate buffer is created even for a zero-length token.
5594  *
5595  * Returns a pointer to the newly-allocated duplicate, or a null
5596  * pointer if memory for the duplicate was not available.  If
5597  * the lenp argument is a non-null pointer, the length of the token
5598  * (not including the '\0') is returned in *lenp.
5599  *
5600  * If successful, the *buf pointer will be updated to point beyond
5601  * the end of the found token.
5602  *
5603  * Note: uses GFP_KERNEL for allocation.
5604  */
5605 static inline char *dup_token(const char **buf, size_t *lenp)
5606 {
5607         char *dup;
5608         size_t len;
5609
5610         len = next_token(buf);
5611         dup = kmemdup(*buf, len + 1, GFP_KERNEL);
5612         if (!dup)
5613                 return NULL;
5614         *(dup + len) = '\0';
5615         *buf += len;
5616
5617         if (lenp)
5618                 *lenp = len;
5619
5620         return dup;
5621 }
5622
5623 /*
5624  * Parse the options provided for an "rbd add" (i.e., rbd image
5625  * mapping) request.  These arrive via a write to /sys/bus/rbd/add,
5626  * and the data written is passed here via a NUL-terminated buffer.
5627  * Returns 0 if successful or an error code otherwise.
5628  *
5629  * The information extracted from these options is recorded in
5630  * the other parameters which return dynamically-allocated
5631  * structures:
5632  *  ceph_opts
5633  *      The address of a pointer that will refer to a ceph options
5634  *      structure.  Caller must release the returned pointer using
5635  *      ceph_destroy_options() when it is no longer needed.
5636  *  rbd_opts
5637  *      Address of an rbd options pointer.  Fully initialized by
5638  *      this function; caller must release with kfree().
5639  *  spec
5640  *      Address of an rbd image specification pointer.  Fully
5641  *      initialized by this function based on parsed options.
5642  *      Caller must release with rbd_spec_put().
5643  *
5644  * The options passed take this form:
5645  *  <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
5646  * where:
5647  *  <mon_addrs>
5648  *      A comma-separated list of one or more monitor addresses.
5649  *      A monitor address is an ip address, optionally followed
5650  *      by a port number (separated by a colon).
5651  *        I.e.:  ip1[:port1][,ip2[:port2]...]
5652  *  <options>
5653  *      A comma-separated list of ceph and/or rbd options.
5654  *  <pool_name>
5655  *      The name of the rados pool containing the rbd image.
5656  *  <image_name>
5657  *      The name of the image in that pool to map.
5658  *  <snap_id>
5659  *      An optional snapshot id.  If provided, the mapping will
5660  *      present data from the image at the time that snapshot was
5661  *      created.  The image head is used if no snapshot id is
5662  *      provided.  Snapshot mappings are always read-only.
5663  */
5664 static int rbd_add_parse_args(const char *buf,
5665                                 struct ceph_options **ceph_opts,
5666                                 struct rbd_options **opts,
5667                                 struct rbd_spec **rbd_spec)
5668 {
5669         size_t len;
5670         char *options;
5671         const char *mon_addrs;
5672         char *snap_name;
5673         size_t mon_addrs_size;
5674         struct rbd_spec *spec = NULL;
5675         struct rbd_options *rbd_opts = NULL;
5676         struct ceph_options *copts;
5677         int ret;
5678
5679         /* The first four tokens are required */
5680
5681         len = next_token(&buf);
5682         if (!len) {
5683                 rbd_warn(NULL, "no monitor address(es) provided");
5684                 return -EINVAL;
5685         }
5686         mon_addrs = buf;
5687         mon_addrs_size = len + 1;
5688         buf += len;
5689
5690         ret = -EINVAL;
5691         options = dup_token(&buf, NULL);
5692         if (!options)
5693                 return -ENOMEM;
5694         if (!*options) {
5695                 rbd_warn(NULL, "no options provided");
5696                 goto out_err;
5697         }
5698
5699         spec = rbd_spec_alloc();
5700         if (!spec)
5701                 goto out_mem;
5702
5703         spec->pool_name = dup_token(&buf, NULL);
5704         if (!spec->pool_name)
5705                 goto out_mem;
5706         if (!*spec->pool_name) {
5707                 rbd_warn(NULL, "no pool name provided");
5708                 goto out_err;
5709         }
5710
5711         spec->image_name = dup_token(&buf, NULL);
5712         if (!spec->image_name)
5713                 goto out_mem;
5714         if (!*spec->image_name) {
5715                 rbd_warn(NULL, "no image name provided");
5716                 goto out_err;
5717         }
5718
5719         /*
5720          * Snapshot name is optional; default is to use "-"
5721          * (indicating the head/no snapshot).
5722          */
5723         len = next_token(&buf);
5724         if (!len) {
5725                 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
5726                 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
5727         } else if (len > RBD_MAX_SNAP_NAME_LEN) {
5728                 ret = -ENAMETOOLONG;
5729                 goto out_err;
5730         }
5731         snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
5732         if (!snap_name)
5733                 goto out_mem;
5734         *(snap_name + len) = '\0';
5735         spec->snap_name = snap_name;
5736
5737         /* Initialize all rbd options to the defaults */
5738
5739         rbd_opts = kzalloc(sizeof (*rbd_opts), GFP_KERNEL);
5740         if (!rbd_opts)
5741                 goto out_mem;
5742
5743         rbd_opts->read_only = RBD_READ_ONLY_DEFAULT;
5744         rbd_opts->queue_depth = RBD_QUEUE_DEPTH_DEFAULT;
5745         rbd_opts->lock_on_read = RBD_LOCK_ON_READ_DEFAULT;
5746
5747         copts = ceph_parse_options(options, mon_addrs,
5748                                         mon_addrs + mon_addrs_size - 1,
5749                                         parse_rbd_opts_token, rbd_opts);
5750         if (IS_ERR(copts)) {
5751                 ret = PTR_ERR(copts);
5752                 goto out_err;
5753         }
5754         kfree(options);
5755
5756         *ceph_opts = copts;
5757         *opts = rbd_opts;
5758         *rbd_spec = spec;
5759
5760         return 0;
5761 out_mem:
5762         ret = -ENOMEM;
5763 out_err:
5764         kfree(rbd_opts);
5765         rbd_spec_put(spec);
5766         kfree(options);
5767
5768         return ret;
5769 }
5770
5771 /*
5772  * Return pool id (>= 0) or a negative error code.
5773  */
5774 static int rbd_add_get_pool_id(struct rbd_client *rbdc, const char *pool_name)
5775 {
5776         struct ceph_options *opts = rbdc->client->options;
5777         u64 newest_epoch;
5778         int tries = 0;
5779         int ret;
5780
5781 again:
5782         ret = ceph_pg_poolid_by_name(rbdc->client->osdc.osdmap, pool_name);
5783         if (ret == -ENOENT && tries++ < 1) {
5784                 ret = ceph_monc_get_version(&rbdc->client->monc, "osdmap",
5785                                             &newest_epoch);
5786                 if (ret < 0)
5787                         return ret;
5788
5789                 if (rbdc->client->osdc.osdmap->epoch < newest_epoch) {
5790                         ceph_osdc_maybe_request_map(&rbdc->client->osdc);
5791                         (void) ceph_monc_wait_osdmap(&rbdc->client->monc,
5792                                                      newest_epoch,
5793                                                      opts->mount_timeout);
5794                         goto again;
5795                 } else {
5796                         /* the osdmap we have is new enough */
5797                         return -ENOENT;
5798                 }
5799         }
5800
5801         return ret;
5802 }
5803
5804 /*
5805  * An rbd format 2 image has a unique identifier, distinct from the
5806  * name given to it by the user.  Internally, that identifier is
5807  * what's used to specify the names of objects related to the image.
5808  *
5809  * A special "rbd id" object is used to map an rbd image name to its
5810  * id.  If that object doesn't exist, then there is no v2 rbd image
5811  * with the supplied name.
5812  *
5813  * This function will record the given rbd_dev's image_id field if
5814  * it can be determined, and in that case will return 0.  If any
5815  * errors occur a negative errno will be returned and the rbd_dev's
5816  * image_id field will be unchanged (and should be NULL).
5817  */
5818 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
5819 {
5820         int ret;
5821         size_t size;
5822         char *object_name;
5823         void *response;
5824         char *image_id;
5825
5826         /*
5827          * When probing a parent image, the image id is already
5828          * known (and the image name likely is not).  There's no
5829          * need to fetch the image id again in this case.  We
5830          * do still need to set the image format though.
5831          */
5832         if (rbd_dev->spec->image_id) {
5833                 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
5834
5835                 return 0;
5836         }
5837
5838         /*
5839          * First, see if the format 2 image id file exists, and if
5840          * so, get the image's persistent id from it.
5841          */
5842         size = sizeof (RBD_ID_PREFIX) + strlen(rbd_dev->spec->image_name);
5843         object_name = kmalloc(size, GFP_NOIO);
5844         if (!object_name)
5845                 return -ENOMEM;
5846         sprintf(object_name, "%s%s", RBD_ID_PREFIX, rbd_dev->spec->image_name);
5847         dout("rbd id object name is %s\n", object_name);
5848
5849         /* Response will be an encoded string, which includes a length */
5850
5851         size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
5852         response = kzalloc(size, GFP_NOIO);
5853         if (!response) {
5854                 ret = -ENOMEM;
5855                 goto out;
5856         }
5857
5858         /* If it doesn't exist we'll assume it's a format 1 image */
5859
5860         ret = rbd_obj_method_sync(rbd_dev, object_name,
5861                                 "rbd", "get_id", NULL, 0,
5862                                 response, RBD_IMAGE_ID_LEN_MAX);
5863         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5864         if (ret == -ENOENT) {
5865                 image_id = kstrdup("", GFP_KERNEL);
5866                 ret = image_id ? 0 : -ENOMEM;
5867                 if (!ret)
5868                         rbd_dev->image_format = 1;
5869         } else if (ret >= 0) {
5870                 void *p = response;
5871
5872                 image_id = ceph_extract_encoded_string(&p, p + ret,
5873                                                 NULL, GFP_NOIO);
5874                 ret = PTR_ERR_OR_ZERO(image_id);
5875                 if (!ret)
5876                         rbd_dev->image_format = 2;
5877         }
5878
5879         if (!ret) {
5880                 rbd_dev->spec->image_id = image_id;
5881                 dout("image_id is %s\n", image_id);
5882         }
5883 out:
5884         kfree(response);
5885         kfree(object_name);
5886
5887         return ret;
5888 }
5889
5890 /*
5891  * Undo whatever state changes are made by v1 or v2 header info
5892  * call.
5893  */
5894 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
5895 {
5896         struct rbd_image_header *header;
5897
5898         rbd_dev_parent_put(rbd_dev);
5899
5900         /* Free dynamic fields from the header, then zero it out */
5901
5902         header = &rbd_dev->header;
5903         ceph_put_snap_context(header->snapc);
5904         kfree(header->snap_sizes);
5905         kfree(header->snap_names);
5906         kfree(header->object_prefix);
5907         memset(header, 0, sizeof (*header));
5908 }
5909
5910 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev)
5911 {
5912         int ret;
5913
5914         ret = rbd_dev_v2_object_prefix(rbd_dev);
5915         if (ret)
5916                 goto out_err;
5917
5918         /*
5919          * Get the and check features for the image.  Currently the
5920          * features are assumed to never change.
5921          */
5922         ret = rbd_dev_v2_features(rbd_dev);
5923         if (ret)
5924                 goto out_err;
5925
5926         /* If the image supports fancy striping, get its parameters */
5927
5928         if (rbd_dev->header.features & RBD_FEATURE_STRIPINGV2) {
5929                 ret = rbd_dev_v2_striping_info(rbd_dev);
5930                 if (ret < 0)
5931                         goto out_err;
5932         }
5933         /* No support for crypto and compression type format 2 images */
5934
5935         return 0;
5936 out_err:
5937         rbd_dev->header.features = 0;
5938         kfree(rbd_dev->header.object_prefix);
5939         rbd_dev->header.object_prefix = NULL;
5940
5941         return ret;
5942 }
5943
5944 /*
5945  * @depth is rbd_dev_image_probe() -> rbd_dev_probe_parent() ->
5946  * rbd_dev_image_probe() recursion depth, which means it's also the
5947  * length of the already discovered part of the parent chain.
5948  */
5949 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev, int depth)
5950 {
5951         struct rbd_device *parent = NULL;
5952         int ret;
5953
5954         if (!rbd_dev->parent_spec)
5955                 return 0;
5956
5957         if (++depth > RBD_MAX_PARENT_CHAIN_LEN) {
5958                 pr_info("parent chain is too long (%d)\n", depth);
5959                 ret = -EINVAL;
5960                 goto out_err;
5961         }
5962
5963         parent = __rbd_dev_create(rbd_dev->rbd_client, rbd_dev->parent_spec);
5964         if (!parent) {
5965                 ret = -ENOMEM;
5966                 goto out_err;
5967         }
5968
5969         /*
5970          * Images related by parent/child relationships always share
5971          * rbd_client and spec/parent_spec, so bump their refcounts.
5972          */
5973         __rbd_get_client(rbd_dev->rbd_client);
5974         rbd_spec_get(rbd_dev->parent_spec);
5975
5976         ret = rbd_dev_image_probe(parent, depth);
5977         if (ret < 0)
5978                 goto out_err;
5979
5980         rbd_dev->parent = parent;
5981         atomic_set(&rbd_dev->parent_ref, 1);
5982         return 0;
5983
5984 out_err:
5985         rbd_dev_unparent(rbd_dev);
5986         rbd_dev_destroy(parent);
5987         return ret;
5988 }
5989
5990 /*
5991  * rbd_dev->header_rwsem must be locked for write and will be unlocked
5992  * upon return.
5993  */
5994 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
5995 {
5996         int ret;
5997
5998         /* Record our major and minor device numbers. */
5999
6000         if (!single_major) {
6001                 ret = register_blkdev(0, rbd_dev->name);
6002                 if (ret < 0)
6003                         goto err_out_unlock;
6004
6005                 rbd_dev->major = ret;
6006                 rbd_dev->minor = 0;
6007         } else {
6008                 rbd_dev->major = rbd_major;
6009                 rbd_dev->minor = rbd_dev_id_to_minor(rbd_dev->dev_id);
6010         }
6011
6012         /* Set up the blkdev mapping. */
6013
6014         ret = rbd_init_disk(rbd_dev);
6015         if (ret)
6016                 goto err_out_blkdev;
6017
6018         ret = rbd_dev_mapping_set(rbd_dev);
6019         if (ret)
6020                 goto err_out_disk;
6021
6022         set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
6023         set_disk_ro(rbd_dev->disk, rbd_dev->mapping.read_only);
6024
6025         dev_set_name(&rbd_dev->dev, "%d", rbd_dev->dev_id);
6026         ret = device_add(&rbd_dev->dev);
6027         if (ret)
6028                 goto err_out_mapping;
6029
6030         /* Everything's ready.  Announce the disk to the world. */
6031
6032         set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
6033         up_write(&rbd_dev->header_rwsem);
6034
6035         spin_lock(&rbd_dev_list_lock);
6036         list_add_tail(&rbd_dev->node, &rbd_dev_list);
6037         spin_unlock(&rbd_dev_list_lock);
6038
6039         add_disk(rbd_dev->disk);
6040         pr_info("%s: capacity %llu features 0x%llx\n", rbd_dev->disk->disk_name,
6041                 (unsigned long long)get_capacity(rbd_dev->disk) << SECTOR_SHIFT,
6042                 rbd_dev->header.features);
6043
6044         return ret;
6045
6046 err_out_mapping:
6047         rbd_dev_mapping_clear(rbd_dev);
6048 err_out_disk:
6049         rbd_free_disk(rbd_dev);
6050 err_out_blkdev:
6051         if (!single_major)
6052                 unregister_blkdev(rbd_dev->major, rbd_dev->name);
6053 err_out_unlock:
6054         up_write(&rbd_dev->header_rwsem);
6055         return ret;
6056 }
6057
6058 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
6059 {
6060         struct rbd_spec *spec = rbd_dev->spec;
6061         int ret;
6062
6063         /* Record the header object name for this rbd image. */
6064
6065         rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
6066
6067         rbd_dev->header_oloc.pool = rbd_dev->layout.pool_id;
6068         if (rbd_dev->image_format == 1)
6069                 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
6070                                        spec->image_name, RBD_SUFFIX);
6071         else
6072                 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
6073                                        RBD_HEADER_PREFIX, spec->image_id);
6074
6075         return ret;
6076 }
6077
6078 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
6079 {
6080         rbd_dev_unprobe(rbd_dev);
6081         rbd_dev->image_format = 0;
6082         kfree(rbd_dev->spec->image_id);
6083         rbd_dev->spec->image_id = NULL;
6084
6085         rbd_dev_destroy(rbd_dev);
6086 }
6087
6088 /*
6089  * Probe for the existence of the header object for the given rbd
6090  * device.  If this image is the one being mapped (i.e., not a
6091  * parent), initiate a watch on its header object before using that
6092  * object to get detailed information about the rbd image.
6093  */
6094 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth)
6095 {
6096         int ret;
6097
6098         /*
6099          * Get the id from the image id object.  Unless there's an
6100          * error, rbd_dev->spec->image_id will be filled in with
6101          * a dynamically-allocated string, and rbd_dev->image_format
6102          * will be set to either 1 or 2.
6103          */
6104         ret = rbd_dev_image_id(rbd_dev);
6105         if (ret)
6106                 return ret;
6107
6108         ret = rbd_dev_header_name(rbd_dev);
6109         if (ret)
6110                 goto err_out_format;
6111
6112         if (!depth) {
6113                 ret = rbd_register_watch(rbd_dev);
6114                 if (ret) {
6115                         if (ret == -ENOENT)
6116                                 pr_info("image %s/%s does not exist\n",
6117                                         rbd_dev->spec->pool_name,
6118                                         rbd_dev->spec->image_name);
6119                         goto err_out_format;
6120                 }
6121         }
6122
6123         ret = rbd_dev_header_info(rbd_dev);
6124         if (ret)
6125                 goto err_out_watch;
6126
6127         /*
6128          * If this image is the one being mapped, we have pool name and
6129          * id, image name and id, and snap name - need to fill snap id.
6130          * Otherwise this is a parent image, identified by pool, image
6131          * and snap ids - need to fill in names for those ids.
6132          */
6133         if (!depth)
6134                 ret = rbd_spec_fill_snap_id(rbd_dev);
6135         else
6136                 ret = rbd_spec_fill_names(rbd_dev);
6137         if (ret) {
6138                 if (ret == -ENOENT)
6139                         pr_info("snap %s/%s@%s does not exist\n",
6140                                 rbd_dev->spec->pool_name,
6141                                 rbd_dev->spec->image_name,
6142                                 rbd_dev->spec->snap_name);
6143                 goto err_out_probe;
6144         }
6145
6146         if (rbd_dev->header.features & RBD_FEATURE_LAYERING) {
6147                 ret = rbd_dev_v2_parent_info(rbd_dev);
6148                 if (ret)
6149                         goto err_out_probe;
6150
6151                 /*
6152                  * Need to warn users if this image is the one being
6153                  * mapped and has a parent.
6154                  */
6155                 if (!depth && rbd_dev->parent_spec)
6156                         rbd_warn(rbd_dev,
6157                                  "WARNING: kernel layering is EXPERIMENTAL!");
6158         }
6159
6160         ret = rbd_dev_probe_parent(rbd_dev, depth);
6161         if (ret)
6162                 goto err_out_probe;
6163
6164         dout("discovered format %u image, header name is %s\n",
6165                 rbd_dev->image_format, rbd_dev->header_oid.name);
6166         return 0;
6167
6168 err_out_probe:
6169         rbd_dev_unprobe(rbd_dev);
6170 err_out_watch:
6171         if (!depth)
6172                 rbd_unregister_watch(rbd_dev);
6173 err_out_format:
6174         rbd_dev->image_format = 0;
6175         kfree(rbd_dev->spec->image_id);
6176         rbd_dev->spec->image_id = NULL;
6177         return ret;
6178 }
6179
6180 static ssize_t do_rbd_add(struct bus_type *bus,
6181                           const char *buf,
6182                           size_t count)
6183 {
6184         struct rbd_device *rbd_dev = NULL;
6185         struct ceph_options *ceph_opts = NULL;
6186         struct rbd_options *rbd_opts = NULL;
6187         struct rbd_spec *spec = NULL;
6188         struct rbd_client *rbdc;
6189         bool read_only;
6190         int rc;
6191
6192         if (!try_module_get(THIS_MODULE))
6193                 return -ENODEV;
6194
6195         /* parse add command */
6196         rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
6197         if (rc < 0)
6198                 goto out;
6199
6200         rbdc = rbd_get_client(ceph_opts);
6201         if (IS_ERR(rbdc)) {
6202                 rc = PTR_ERR(rbdc);
6203                 goto err_out_args;
6204         }
6205
6206         /* pick the pool */
6207         rc = rbd_add_get_pool_id(rbdc, spec->pool_name);
6208         if (rc < 0) {
6209                 if (rc == -ENOENT)
6210                         pr_info("pool %s does not exist\n", spec->pool_name);
6211                 goto err_out_client;
6212         }
6213         spec->pool_id = (u64)rc;
6214
6215         rbd_dev = rbd_dev_create(rbdc, spec, rbd_opts);
6216         if (!rbd_dev) {
6217                 rc = -ENOMEM;
6218                 goto err_out_client;
6219         }
6220         rbdc = NULL;            /* rbd_dev now owns this */
6221         spec = NULL;            /* rbd_dev now owns this */
6222         rbd_opts = NULL;        /* rbd_dev now owns this */
6223
6224         rbd_dev->config_info = kstrdup(buf, GFP_KERNEL);
6225         if (!rbd_dev->config_info) {
6226                 rc = -ENOMEM;
6227                 goto err_out_rbd_dev;
6228         }
6229
6230         down_write(&rbd_dev->header_rwsem);
6231         rc = rbd_dev_image_probe(rbd_dev, 0);
6232         if (rc < 0) {
6233                 up_write(&rbd_dev->header_rwsem);
6234                 goto err_out_rbd_dev;
6235         }
6236
6237         /* If we are mapping a snapshot it must be marked read-only */
6238
6239         read_only = rbd_dev->opts->read_only;
6240         if (rbd_dev->spec->snap_id != CEPH_NOSNAP)
6241                 read_only = true;
6242         rbd_dev->mapping.read_only = read_only;
6243
6244         rc = rbd_dev_device_setup(rbd_dev);
6245         if (rc) {
6246                 /*
6247                  * rbd_unregister_watch() can't be moved into
6248                  * rbd_dev_image_release() without refactoring, see
6249                  * commit 1f3ef78861ac.
6250                  */
6251                 rbd_unregister_watch(rbd_dev);
6252                 rbd_dev_image_release(rbd_dev);
6253                 goto out;
6254         }
6255
6256         rc = count;
6257 out:
6258         module_put(THIS_MODULE);
6259         return rc;
6260
6261 err_out_rbd_dev:
6262         rbd_dev_destroy(rbd_dev);
6263 err_out_client:
6264         rbd_put_client(rbdc);
6265 err_out_args:
6266         rbd_spec_put(spec);
6267         kfree(rbd_opts);
6268         goto out;
6269 }
6270
6271 static ssize_t rbd_add(struct bus_type *bus,
6272                        const char *buf,
6273                        size_t count)
6274 {
6275         if (single_major)
6276                 return -EINVAL;
6277
6278         return do_rbd_add(bus, buf, count);
6279 }
6280
6281 static ssize_t rbd_add_single_major(struct bus_type *bus,
6282                                     const char *buf,
6283                                     size_t count)
6284 {
6285         return do_rbd_add(bus, buf, count);
6286 }
6287
6288 static void rbd_dev_device_release(struct rbd_device *rbd_dev)
6289 {
6290         rbd_free_disk(rbd_dev);
6291
6292         spin_lock(&rbd_dev_list_lock);
6293         list_del_init(&rbd_dev->node);
6294         spin_unlock(&rbd_dev_list_lock);
6295
6296         clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
6297         device_del(&rbd_dev->dev);
6298         rbd_dev_mapping_clear(rbd_dev);
6299         if (!single_major)
6300                 unregister_blkdev(rbd_dev->major, rbd_dev->name);
6301 }
6302
6303 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
6304 {
6305         while (rbd_dev->parent) {
6306                 struct rbd_device *first = rbd_dev;
6307                 struct rbd_device *second = first->parent;
6308                 struct rbd_device *third;
6309
6310                 /*
6311                  * Follow to the parent with no grandparent and
6312                  * remove it.
6313                  */
6314                 while (second && (third = second->parent)) {
6315                         first = second;
6316                         second = third;
6317                 }
6318                 rbd_assert(second);
6319                 rbd_dev_image_release(second);
6320                 first->parent = NULL;
6321                 first->parent_overlap = 0;
6322
6323                 rbd_assert(first->parent_spec);
6324                 rbd_spec_put(first->parent_spec);
6325                 first->parent_spec = NULL;
6326         }
6327 }
6328
6329 static ssize_t do_rbd_remove(struct bus_type *bus,
6330                              const char *buf,
6331                              size_t count)
6332 {
6333         struct rbd_device *rbd_dev = NULL;
6334         struct list_head *tmp;
6335         int dev_id;
6336         char opt_buf[6];
6337         bool already = false;
6338         bool force = false;
6339         int ret;
6340
6341         dev_id = -1;
6342         opt_buf[0] = '\0';
6343         sscanf(buf, "%d %5s", &dev_id, opt_buf);
6344         if (dev_id < 0) {
6345                 pr_err("dev_id out of range\n");
6346                 return -EINVAL;
6347         }
6348         if (opt_buf[0] != '\0') {
6349                 if (!strcmp(opt_buf, "force")) {
6350                         force = true;
6351                 } else {
6352                         pr_err("bad remove option at '%s'\n", opt_buf);
6353                         return -EINVAL;
6354                 }
6355         }
6356
6357         ret = -ENOENT;
6358         spin_lock(&rbd_dev_list_lock);
6359         list_for_each(tmp, &rbd_dev_list) {
6360                 rbd_dev = list_entry(tmp, struct rbd_device, node);
6361                 if (rbd_dev->dev_id == dev_id) {
6362                         ret = 0;
6363                         break;
6364                 }
6365         }
6366         if (!ret) {
6367                 spin_lock_irq(&rbd_dev->lock);
6368                 if (rbd_dev->open_count && !force)
6369                         ret = -EBUSY;
6370                 else
6371                         already = test_and_set_bit(RBD_DEV_FLAG_REMOVING,
6372                                                         &rbd_dev->flags);
6373                 spin_unlock_irq(&rbd_dev->lock);
6374         }
6375         spin_unlock(&rbd_dev_list_lock);
6376         if (ret < 0 || already)
6377                 return ret;
6378
6379         if (force) {
6380                 /*
6381                  * Prevent new IO from being queued and wait for existing
6382                  * IO to complete/fail.
6383                  */
6384                 blk_mq_freeze_queue(rbd_dev->disk->queue);
6385                 blk_set_queue_dying(rbd_dev->disk->queue);
6386         }
6387
6388         down_write(&rbd_dev->lock_rwsem);
6389         if (__rbd_is_lock_owner(rbd_dev))
6390                 rbd_unlock(rbd_dev);
6391         up_write(&rbd_dev->lock_rwsem);
6392         rbd_unregister_watch(rbd_dev);
6393
6394         /*
6395          * Don't free anything from rbd_dev->disk until after all
6396          * notifies are completely processed. Otherwise
6397          * rbd_bus_del_dev() will race with rbd_watch_cb(), resulting
6398          * in a potential use after free of rbd_dev->disk or rbd_dev.
6399          */
6400         rbd_dev_device_release(rbd_dev);
6401         rbd_dev_image_release(rbd_dev);
6402
6403         return count;
6404 }
6405
6406 static ssize_t rbd_remove(struct bus_type *bus,
6407                           const char *buf,
6408                           size_t count)
6409 {
6410         if (single_major)
6411                 return -EINVAL;
6412
6413         return do_rbd_remove(bus, buf, count);
6414 }
6415
6416 static ssize_t rbd_remove_single_major(struct bus_type *bus,
6417                                        const char *buf,
6418                                        size_t count)
6419 {
6420         return do_rbd_remove(bus, buf, count);
6421 }
6422
6423 /*
6424  * create control files in sysfs
6425  * /sys/bus/rbd/...
6426  */
6427 static int rbd_sysfs_init(void)
6428 {
6429         int ret;
6430
6431         ret = device_register(&rbd_root_dev);
6432         if (ret < 0)
6433                 return ret;
6434
6435         ret = bus_register(&rbd_bus_type);
6436         if (ret < 0)
6437                 device_unregister(&rbd_root_dev);
6438
6439         return ret;
6440 }
6441
6442 static void rbd_sysfs_cleanup(void)
6443 {
6444         bus_unregister(&rbd_bus_type);
6445         device_unregister(&rbd_root_dev);
6446 }
6447
6448 static int rbd_slab_init(void)
6449 {
6450         rbd_assert(!rbd_img_request_cache);
6451         rbd_img_request_cache = KMEM_CACHE(rbd_img_request, 0);
6452         if (!rbd_img_request_cache)
6453                 return -ENOMEM;
6454
6455         rbd_assert(!rbd_obj_request_cache);
6456         rbd_obj_request_cache = KMEM_CACHE(rbd_obj_request, 0);
6457         if (!rbd_obj_request_cache)
6458                 goto out_err;
6459
6460         rbd_assert(!rbd_segment_name_cache);
6461         rbd_segment_name_cache = kmem_cache_create("rbd_segment_name",
6462                                         CEPH_MAX_OID_NAME_LEN + 1, 1, 0, NULL);
6463         if (rbd_segment_name_cache)
6464                 return 0;
6465 out_err:
6466         kmem_cache_destroy(rbd_obj_request_cache);
6467         rbd_obj_request_cache = NULL;
6468
6469         kmem_cache_destroy(rbd_img_request_cache);
6470         rbd_img_request_cache = NULL;
6471
6472         return -ENOMEM;
6473 }
6474
6475 static void rbd_slab_exit(void)
6476 {
6477         rbd_assert(rbd_segment_name_cache);
6478         kmem_cache_destroy(rbd_segment_name_cache);
6479         rbd_segment_name_cache = NULL;
6480
6481         rbd_assert(rbd_obj_request_cache);
6482         kmem_cache_destroy(rbd_obj_request_cache);
6483         rbd_obj_request_cache = NULL;
6484
6485         rbd_assert(rbd_img_request_cache);
6486         kmem_cache_destroy(rbd_img_request_cache);
6487         rbd_img_request_cache = NULL;
6488 }
6489
6490 static int __init rbd_init(void)
6491 {
6492         int rc;
6493
6494         if (!libceph_compatible(NULL)) {
6495                 rbd_warn(NULL, "libceph incompatibility (quitting)");
6496                 return -EINVAL;
6497         }
6498
6499         rc = rbd_slab_init();
6500         if (rc)
6501                 return rc;
6502
6503         /*
6504          * The number of active work items is limited by the number of
6505          * rbd devices * queue depth, so leave @max_active at default.
6506          */
6507         rbd_wq = alloc_workqueue(RBD_DRV_NAME, WQ_MEM_RECLAIM, 0);
6508         if (!rbd_wq) {
6509                 rc = -ENOMEM;
6510                 goto err_out_slab;
6511         }
6512
6513         if (single_major) {
6514                 rbd_major = register_blkdev(0, RBD_DRV_NAME);
6515                 if (rbd_major < 0) {
6516                         rc = rbd_major;
6517                         goto err_out_wq;
6518                 }
6519         }
6520
6521         rc = rbd_sysfs_init();
6522         if (rc)
6523                 goto err_out_blkdev;
6524
6525         if (single_major)
6526                 pr_info("loaded (major %d)\n", rbd_major);
6527         else
6528                 pr_info("loaded\n");
6529
6530         return 0;
6531
6532 err_out_blkdev:
6533         if (single_major)
6534                 unregister_blkdev(rbd_major, RBD_DRV_NAME);
6535 err_out_wq:
6536         destroy_workqueue(rbd_wq);
6537 err_out_slab:
6538         rbd_slab_exit();
6539         return rc;
6540 }
6541
6542 static void __exit rbd_exit(void)
6543 {
6544         ida_destroy(&rbd_dev_id_ida);
6545         rbd_sysfs_cleanup();
6546         if (single_major)
6547                 unregister_blkdev(rbd_major, RBD_DRV_NAME);
6548         destroy_workqueue(rbd_wq);
6549         rbd_slab_exit();
6550 }
6551
6552 module_init(rbd_init);
6553 module_exit(rbd_exit);
6554
6555 MODULE_AUTHOR("Alex Elder <elder@inktank.com>");
6556 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
6557 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
6558 /* following authorship retained from original osdblk.c */
6559 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
6560
6561 MODULE_DESCRIPTION("RADOS Block Device (RBD) driver");
6562 MODULE_LICENSE("GPL");