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_obj_request_submit(struct rbd_obj_request *obj_request)
1622 {
1623         struct ceph_osd_request *osd_req = obj_request->osd_req;
1624
1625         dout("%s %p osd_req %p\n", __func__, obj_request, osd_req);
1626         ceph_osdc_start_request(osd_req->r_osdc, osd_req, false);
1627 }
1628
1629 static void rbd_obj_request_end(struct rbd_obj_request *obj_request)
1630 {
1631         dout("%s %p\n", __func__, obj_request);
1632         ceph_osdc_cancel_request(obj_request->osd_req);
1633 }
1634
1635 /*
1636  * Wait for an object request to complete.  If interrupted, cancel the
1637  * underlying osd request.
1638  *
1639  * @timeout: in jiffies, 0 means "wait forever"
1640  */
1641 static int __rbd_obj_request_wait(struct rbd_obj_request *obj_request,
1642                                   unsigned long timeout)
1643 {
1644         long ret;
1645
1646         dout("%s %p\n", __func__, obj_request);
1647         ret = wait_for_completion_interruptible_timeout(
1648                                         &obj_request->completion,
1649                                         ceph_timeout_jiffies(timeout));
1650         if (ret <= 0) {
1651                 if (ret == 0)
1652                         ret = -ETIMEDOUT;
1653                 rbd_obj_request_end(obj_request);
1654         } else {
1655                 ret = 0;
1656         }
1657
1658         dout("%s %p ret %d\n", __func__, obj_request, (int)ret);
1659         return ret;
1660 }
1661
1662 static int rbd_obj_request_wait(struct rbd_obj_request *obj_request)
1663 {
1664         return __rbd_obj_request_wait(obj_request, 0);
1665 }
1666
1667 static void rbd_img_request_complete(struct rbd_img_request *img_request)
1668 {
1669
1670         dout("%s: img %p\n", __func__, img_request);
1671
1672         /*
1673          * If no error occurred, compute the aggregate transfer
1674          * count for the image request.  We could instead use
1675          * atomic64_cmpxchg() to update it as each object request
1676          * completes; not clear which way is better off hand.
1677          */
1678         if (!img_request->result) {
1679                 struct rbd_obj_request *obj_request;
1680                 u64 xferred = 0;
1681
1682                 for_each_obj_request(img_request, obj_request)
1683                         xferred += obj_request->xferred;
1684                 img_request->xferred = xferred;
1685         }
1686
1687         if (img_request->callback)
1688                 img_request->callback(img_request);
1689         else
1690                 rbd_img_request_put(img_request);
1691 }
1692
1693 /*
1694  * The default/initial value for all image request flags is 0.  Each
1695  * is conditionally set to 1 at image request initialization time
1696  * and currently never change thereafter.
1697  */
1698 static void img_request_write_set(struct rbd_img_request *img_request)
1699 {
1700         set_bit(IMG_REQ_WRITE, &img_request->flags);
1701         smp_mb();
1702 }
1703
1704 static bool img_request_write_test(struct rbd_img_request *img_request)
1705 {
1706         smp_mb();
1707         return test_bit(IMG_REQ_WRITE, &img_request->flags) != 0;
1708 }
1709
1710 /*
1711  * Set the discard flag when the img_request is an discard request
1712  */
1713 static void img_request_discard_set(struct rbd_img_request *img_request)
1714 {
1715         set_bit(IMG_REQ_DISCARD, &img_request->flags);
1716         smp_mb();
1717 }
1718
1719 static bool img_request_discard_test(struct rbd_img_request *img_request)
1720 {
1721         smp_mb();
1722         return test_bit(IMG_REQ_DISCARD, &img_request->flags) != 0;
1723 }
1724
1725 static void img_request_child_set(struct rbd_img_request *img_request)
1726 {
1727         set_bit(IMG_REQ_CHILD, &img_request->flags);
1728         smp_mb();
1729 }
1730
1731 static void img_request_child_clear(struct rbd_img_request *img_request)
1732 {
1733         clear_bit(IMG_REQ_CHILD, &img_request->flags);
1734         smp_mb();
1735 }
1736
1737 static bool img_request_child_test(struct rbd_img_request *img_request)
1738 {
1739         smp_mb();
1740         return test_bit(IMG_REQ_CHILD, &img_request->flags) != 0;
1741 }
1742
1743 static void img_request_layered_set(struct rbd_img_request *img_request)
1744 {
1745         set_bit(IMG_REQ_LAYERED, &img_request->flags);
1746         smp_mb();
1747 }
1748
1749 static void img_request_layered_clear(struct rbd_img_request *img_request)
1750 {
1751         clear_bit(IMG_REQ_LAYERED, &img_request->flags);
1752         smp_mb();
1753 }
1754
1755 static bool img_request_layered_test(struct rbd_img_request *img_request)
1756 {
1757         smp_mb();
1758         return test_bit(IMG_REQ_LAYERED, &img_request->flags) != 0;
1759 }
1760
1761 static enum obj_operation_type
1762 rbd_img_request_op_type(struct rbd_img_request *img_request)
1763 {
1764         if (img_request_write_test(img_request))
1765                 return OBJ_OP_WRITE;
1766         else if (img_request_discard_test(img_request))
1767                 return OBJ_OP_DISCARD;
1768         else
1769                 return OBJ_OP_READ;
1770 }
1771
1772 static void
1773 rbd_img_obj_request_read_callback(struct rbd_obj_request *obj_request)
1774 {
1775         u64 xferred = obj_request->xferred;
1776         u64 length = obj_request->length;
1777
1778         dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1779                 obj_request, obj_request->img_request, obj_request->result,
1780                 xferred, length);
1781         /*
1782          * ENOENT means a hole in the image.  We zero-fill the entire
1783          * length of the request.  A short read also implies zero-fill
1784          * to the end of the request.  An error requires the whole
1785          * length of the request to be reported finished with an error
1786          * to the block layer.  In each case we update the xferred
1787          * count to indicate the whole request was satisfied.
1788          */
1789         rbd_assert(obj_request->type != OBJ_REQUEST_NODATA);
1790         if (obj_request->result == -ENOENT) {
1791                 if (obj_request->type == OBJ_REQUEST_BIO)
1792                         zero_bio_chain(obj_request->bio_list, 0);
1793                 else
1794                         zero_pages(obj_request->pages, 0, length);
1795                 obj_request->result = 0;
1796         } else if (xferred < length && !obj_request->result) {
1797                 if (obj_request->type == OBJ_REQUEST_BIO)
1798                         zero_bio_chain(obj_request->bio_list, xferred);
1799                 else
1800                         zero_pages(obj_request->pages, xferred, length);
1801         }
1802         obj_request->xferred = length;
1803         obj_request_done_set(obj_request);
1804 }
1805
1806 static void rbd_obj_request_complete(struct rbd_obj_request *obj_request)
1807 {
1808         dout("%s: obj %p cb %p\n", __func__, obj_request,
1809                 obj_request->callback);
1810         if (obj_request->callback)
1811                 obj_request->callback(obj_request);
1812         else
1813                 complete_all(&obj_request->completion);
1814 }
1815
1816 static void rbd_osd_read_callback(struct rbd_obj_request *obj_request)
1817 {
1818         struct rbd_img_request *img_request = NULL;
1819         struct rbd_device *rbd_dev = NULL;
1820         bool layered = false;
1821
1822         if (obj_request_img_data_test(obj_request)) {
1823                 img_request = obj_request->img_request;
1824                 layered = img_request && img_request_layered_test(img_request);
1825                 rbd_dev = img_request->rbd_dev;
1826         }
1827
1828         dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1829                 obj_request, img_request, obj_request->result,
1830                 obj_request->xferred, obj_request->length);
1831         if (layered && obj_request->result == -ENOENT &&
1832                         obj_request->img_offset < rbd_dev->parent_overlap)
1833                 rbd_img_parent_read(obj_request);
1834         else if (img_request)
1835                 rbd_img_obj_request_read_callback(obj_request);
1836         else
1837                 obj_request_done_set(obj_request);
1838 }
1839
1840 static void rbd_osd_write_callback(struct rbd_obj_request *obj_request)
1841 {
1842         dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1843                 obj_request->result, obj_request->length);
1844         /*
1845          * There is no such thing as a successful short write.  Set
1846          * it to our originally-requested length.
1847          */
1848         obj_request->xferred = obj_request->length;
1849         obj_request_done_set(obj_request);
1850 }
1851
1852 static void rbd_osd_discard_callback(struct rbd_obj_request *obj_request)
1853 {
1854         dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1855                 obj_request->result, obj_request->length);
1856         /*
1857          * There is no such thing as a successful short discard.  Set
1858          * it to our originally-requested length.
1859          */
1860         obj_request->xferred = obj_request->length;
1861         /* discarding a non-existent object is not a problem */
1862         if (obj_request->result == -ENOENT)
1863                 obj_request->result = 0;
1864         obj_request_done_set(obj_request);
1865 }
1866
1867 /*
1868  * For a simple stat call there's nothing to do.  We'll do more if
1869  * this is part of a write sequence for a layered image.
1870  */
1871 static void rbd_osd_stat_callback(struct rbd_obj_request *obj_request)
1872 {
1873         dout("%s: obj %p\n", __func__, obj_request);
1874         obj_request_done_set(obj_request);
1875 }
1876
1877 static void rbd_osd_call_callback(struct rbd_obj_request *obj_request)
1878 {
1879         dout("%s: obj %p\n", __func__, obj_request);
1880
1881         if (obj_request_img_data_test(obj_request))
1882                 rbd_osd_copyup_callback(obj_request);
1883         else
1884                 obj_request_done_set(obj_request);
1885 }
1886
1887 static void rbd_osd_req_callback(struct ceph_osd_request *osd_req)
1888 {
1889         struct rbd_obj_request *obj_request = osd_req->r_priv;
1890         u16 opcode;
1891
1892         dout("%s: osd_req %p\n", __func__, osd_req);
1893         rbd_assert(osd_req == obj_request->osd_req);
1894         if (obj_request_img_data_test(obj_request)) {
1895                 rbd_assert(obj_request->img_request);
1896                 rbd_assert(obj_request->which != BAD_WHICH);
1897         } else {
1898                 rbd_assert(obj_request->which == BAD_WHICH);
1899         }
1900
1901         if (osd_req->r_result < 0)
1902                 obj_request->result = osd_req->r_result;
1903
1904         /*
1905          * We support a 64-bit length, but ultimately it has to be
1906          * passed to the block layer, which just supports a 32-bit
1907          * length field.
1908          */
1909         obj_request->xferred = osd_req->r_ops[0].outdata_len;
1910         rbd_assert(obj_request->xferred < (u64)UINT_MAX);
1911
1912         opcode = osd_req->r_ops[0].op;
1913         switch (opcode) {
1914         case CEPH_OSD_OP_READ:
1915                 rbd_osd_read_callback(obj_request);
1916                 break;
1917         case CEPH_OSD_OP_SETALLOCHINT:
1918                 rbd_assert(osd_req->r_ops[1].op == CEPH_OSD_OP_WRITE ||
1919                            osd_req->r_ops[1].op == CEPH_OSD_OP_WRITEFULL);
1920                 /* fall through */
1921         case CEPH_OSD_OP_WRITE:
1922         case CEPH_OSD_OP_WRITEFULL:
1923                 rbd_osd_write_callback(obj_request);
1924                 break;
1925         case CEPH_OSD_OP_STAT:
1926                 rbd_osd_stat_callback(obj_request);
1927                 break;
1928         case CEPH_OSD_OP_DELETE:
1929         case CEPH_OSD_OP_TRUNCATE:
1930         case CEPH_OSD_OP_ZERO:
1931                 rbd_osd_discard_callback(obj_request);
1932                 break;
1933         case CEPH_OSD_OP_CALL:
1934                 rbd_osd_call_callback(obj_request);
1935                 break;
1936         default:
1937                 rbd_warn(NULL, "%s: unsupported op %hu",
1938                         obj_request->object_name, (unsigned short) opcode);
1939                 break;
1940         }
1941
1942         if (obj_request_done_test(obj_request))
1943                 rbd_obj_request_complete(obj_request);
1944 }
1945
1946 static void rbd_osd_req_format_read(struct rbd_obj_request *obj_request)
1947 {
1948         struct rbd_img_request *img_request = obj_request->img_request;
1949         struct ceph_osd_request *osd_req = obj_request->osd_req;
1950
1951         if (img_request)
1952                 osd_req->r_snapid = img_request->snap_id;
1953 }
1954
1955 static void rbd_osd_req_format_write(struct rbd_obj_request *obj_request)
1956 {
1957         struct ceph_osd_request *osd_req = obj_request->osd_req;
1958
1959         osd_req->r_mtime = CURRENT_TIME;
1960         osd_req->r_data_offset = obj_request->offset;
1961 }
1962
1963 /*
1964  * Create an osd request.  A read request has one osd op (read).
1965  * A write request has either one (watch) or two (hint+write) osd ops.
1966  * (All rbd data writes are prefixed with an allocation hint op, but
1967  * technically osd watch is a write request, hence this distinction.)
1968  */
1969 static struct ceph_osd_request *rbd_osd_req_create(
1970                                         struct rbd_device *rbd_dev,
1971                                         enum obj_operation_type op_type,
1972                                         unsigned int num_ops,
1973                                         struct rbd_obj_request *obj_request)
1974 {
1975         struct ceph_snap_context *snapc = NULL;
1976         struct ceph_osd_client *osdc;
1977         struct ceph_osd_request *osd_req;
1978
1979         if (obj_request_img_data_test(obj_request) &&
1980                 (op_type == OBJ_OP_DISCARD || op_type == OBJ_OP_WRITE)) {
1981                 struct rbd_img_request *img_request = obj_request->img_request;
1982                 if (op_type == OBJ_OP_WRITE) {
1983                         rbd_assert(img_request_write_test(img_request));
1984                 } else {
1985                         rbd_assert(img_request_discard_test(img_request));
1986                 }
1987                 snapc = img_request->snapc;
1988         }
1989
1990         rbd_assert(num_ops == 1 || ((op_type == OBJ_OP_WRITE) && num_ops == 2));
1991
1992         /* Allocate and initialize the request, for the num_ops ops */
1993
1994         osdc = &rbd_dev->rbd_client->client->osdc;
1995         osd_req = ceph_osdc_alloc_request(osdc, snapc, num_ops, false,
1996                                           GFP_NOIO);
1997         if (!osd_req)
1998                 goto fail;
1999
2000         if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
2001                 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
2002         else
2003                 osd_req->r_flags = CEPH_OSD_FLAG_READ;
2004
2005         osd_req->r_callback = rbd_osd_req_callback;
2006         osd_req->r_priv = obj_request;
2007
2008         osd_req->r_base_oloc.pool = rbd_dev->layout.pool_id;
2009         if (ceph_oid_aprintf(&osd_req->r_base_oid, GFP_NOIO, "%s",
2010                              obj_request->object_name))
2011                 goto fail;
2012
2013         if (ceph_osdc_alloc_messages(osd_req, GFP_NOIO))
2014                 goto fail;
2015
2016         return osd_req;
2017
2018 fail:
2019         ceph_osdc_put_request(osd_req);
2020         return NULL;
2021 }
2022
2023 /*
2024  * Create a copyup osd request based on the information in the object
2025  * request supplied.  A copyup request has two or three osd ops, a
2026  * copyup method call, potentially a hint op, and a write or truncate
2027  * or zero op.
2028  */
2029 static struct ceph_osd_request *
2030 rbd_osd_req_create_copyup(struct rbd_obj_request *obj_request)
2031 {
2032         struct rbd_img_request *img_request;
2033         struct ceph_snap_context *snapc;
2034         struct rbd_device *rbd_dev;
2035         struct ceph_osd_client *osdc;
2036         struct ceph_osd_request *osd_req;
2037         int num_osd_ops = 3;
2038
2039         rbd_assert(obj_request_img_data_test(obj_request));
2040         img_request = obj_request->img_request;
2041         rbd_assert(img_request);
2042         rbd_assert(img_request_write_test(img_request) ||
2043                         img_request_discard_test(img_request));
2044
2045         if (img_request_discard_test(img_request))
2046                 num_osd_ops = 2;
2047
2048         /* Allocate and initialize the request, for all the ops */
2049
2050         snapc = img_request->snapc;
2051         rbd_dev = img_request->rbd_dev;
2052         osdc = &rbd_dev->rbd_client->client->osdc;
2053         osd_req = ceph_osdc_alloc_request(osdc, snapc, num_osd_ops,
2054                                                 false, GFP_NOIO);
2055         if (!osd_req)
2056                 goto fail;
2057
2058         osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
2059         osd_req->r_callback = rbd_osd_req_callback;
2060         osd_req->r_priv = obj_request;
2061
2062         osd_req->r_base_oloc.pool = rbd_dev->layout.pool_id;
2063         if (ceph_oid_aprintf(&osd_req->r_base_oid, GFP_NOIO, "%s",
2064                              obj_request->object_name))
2065                 goto fail;
2066
2067         if (ceph_osdc_alloc_messages(osd_req, GFP_NOIO))
2068                 goto fail;
2069
2070         return osd_req;
2071
2072 fail:
2073         ceph_osdc_put_request(osd_req);
2074         return NULL;
2075 }
2076
2077
2078 static void rbd_osd_req_destroy(struct ceph_osd_request *osd_req)
2079 {
2080         ceph_osdc_put_request(osd_req);
2081 }
2082
2083 /* object_name is assumed to be a non-null pointer and NUL-terminated */
2084
2085 static struct rbd_obj_request *rbd_obj_request_create(const char *object_name,
2086                                                 u64 offset, u64 length,
2087                                                 enum obj_request_type type)
2088 {
2089         struct rbd_obj_request *obj_request;
2090         size_t size;
2091         char *name;
2092
2093         rbd_assert(obj_request_type_valid(type));
2094
2095         size = strlen(object_name) + 1;
2096         name = kmalloc(size, GFP_NOIO);
2097         if (!name)
2098                 return NULL;
2099
2100         obj_request = kmem_cache_zalloc(rbd_obj_request_cache, GFP_NOIO);
2101         if (!obj_request) {
2102                 kfree(name);
2103                 return NULL;
2104         }
2105
2106         obj_request->object_name = memcpy(name, object_name, size);
2107         obj_request->offset = offset;
2108         obj_request->length = length;
2109         obj_request->flags = 0;
2110         obj_request->which = BAD_WHICH;
2111         obj_request->type = type;
2112         INIT_LIST_HEAD(&obj_request->links);
2113         init_completion(&obj_request->completion);
2114         kref_init(&obj_request->kref);
2115
2116         dout("%s: \"%s\" %llu/%llu %d -> obj %p\n", __func__, object_name,
2117                 offset, length, (int)type, obj_request);
2118
2119         return obj_request;
2120 }
2121
2122 static void rbd_obj_request_destroy(struct kref *kref)
2123 {
2124         struct rbd_obj_request *obj_request;
2125
2126         obj_request = container_of(kref, struct rbd_obj_request, kref);
2127
2128         dout("%s: obj %p\n", __func__, obj_request);
2129
2130         rbd_assert(obj_request->img_request == NULL);
2131         rbd_assert(obj_request->which == BAD_WHICH);
2132
2133         if (obj_request->osd_req)
2134                 rbd_osd_req_destroy(obj_request->osd_req);
2135
2136         rbd_assert(obj_request_type_valid(obj_request->type));
2137         switch (obj_request->type) {
2138         case OBJ_REQUEST_NODATA:
2139                 break;          /* Nothing to do */
2140         case OBJ_REQUEST_BIO:
2141                 if (obj_request->bio_list)
2142                         bio_chain_put(obj_request->bio_list);
2143                 break;
2144         case OBJ_REQUEST_PAGES:
2145                 if (obj_request->pages)
2146                         ceph_release_page_vector(obj_request->pages,
2147                                                 obj_request->page_count);
2148                 break;
2149         }
2150
2151         kfree(obj_request->object_name);
2152         obj_request->object_name = NULL;
2153         kmem_cache_free(rbd_obj_request_cache, obj_request);
2154 }
2155
2156 /* It's OK to call this for a device with no parent */
2157
2158 static void rbd_spec_put(struct rbd_spec *spec);
2159 static void rbd_dev_unparent(struct rbd_device *rbd_dev)
2160 {
2161         rbd_dev_remove_parent(rbd_dev);
2162         rbd_spec_put(rbd_dev->parent_spec);
2163         rbd_dev->parent_spec = NULL;
2164         rbd_dev->parent_overlap = 0;
2165 }
2166
2167 /*
2168  * Parent image reference counting is used to determine when an
2169  * image's parent fields can be safely torn down--after there are no
2170  * more in-flight requests to the parent image.  When the last
2171  * reference is dropped, cleaning them up is safe.
2172  */
2173 static void rbd_dev_parent_put(struct rbd_device *rbd_dev)
2174 {
2175         int counter;
2176
2177         if (!rbd_dev->parent_spec)
2178                 return;
2179
2180         counter = atomic_dec_return_safe(&rbd_dev->parent_ref);
2181         if (counter > 0)
2182                 return;
2183
2184         /* Last reference; clean up parent data structures */
2185
2186         if (!counter)
2187                 rbd_dev_unparent(rbd_dev);
2188         else
2189                 rbd_warn(rbd_dev, "parent reference underflow");
2190 }
2191
2192 /*
2193  * If an image has a non-zero parent overlap, get a reference to its
2194  * parent.
2195  *
2196  * Returns true if the rbd device has a parent with a non-zero
2197  * overlap and a reference for it was successfully taken, or
2198  * false otherwise.
2199  */
2200 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
2201 {
2202         int counter = 0;
2203
2204         if (!rbd_dev->parent_spec)
2205                 return false;
2206
2207         down_read(&rbd_dev->header_rwsem);
2208         if (rbd_dev->parent_overlap)
2209                 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
2210         up_read(&rbd_dev->header_rwsem);
2211
2212         if (counter < 0)
2213                 rbd_warn(rbd_dev, "parent reference overflow");
2214
2215         return counter > 0;
2216 }
2217
2218 /*
2219  * Caller is responsible for filling in the list of object requests
2220  * that comprises the image request, and the Linux request pointer
2221  * (if there is one).
2222  */
2223 static struct rbd_img_request *rbd_img_request_create(
2224                                         struct rbd_device *rbd_dev,
2225                                         u64 offset, u64 length,
2226                                         enum obj_operation_type op_type,
2227                                         struct ceph_snap_context *snapc)
2228 {
2229         struct rbd_img_request *img_request;
2230
2231         img_request = kmem_cache_alloc(rbd_img_request_cache, GFP_NOIO);
2232         if (!img_request)
2233                 return NULL;
2234
2235         img_request->rq = NULL;
2236         img_request->rbd_dev = rbd_dev;
2237         img_request->offset = offset;
2238         img_request->length = length;
2239         img_request->flags = 0;
2240         if (op_type == OBJ_OP_DISCARD) {
2241                 img_request_discard_set(img_request);
2242                 img_request->snapc = snapc;
2243         } else if (op_type == OBJ_OP_WRITE) {
2244                 img_request_write_set(img_request);
2245                 img_request->snapc = snapc;
2246         } else {
2247                 img_request->snap_id = rbd_dev->spec->snap_id;
2248         }
2249         if (rbd_dev_parent_get(rbd_dev))
2250                 img_request_layered_set(img_request);
2251         spin_lock_init(&img_request->completion_lock);
2252         img_request->next_completion = 0;
2253         img_request->callback = NULL;
2254         img_request->result = 0;
2255         img_request->obj_request_count = 0;
2256         INIT_LIST_HEAD(&img_request->obj_requests);
2257         kref_init(&img_request->kref);
2258
2259         dout("%s: rbd_dev %p %s %llu/%llu -> img %p\n", __func__, rbd_dev,
2260                 obj_op_name(op_type), offset, length, img_request);
2261
2262         return img_request;
2263 }
2264
2265 static void rbd_img_request_destroy(struct kref *kref)
2266 {
2267         struct rbd_img_request *img_request;
2268         struct rbd_obj_request *obj_request;
2269         struct rbd_obj_request *next_obj_request;
2270
2271         img_request = container_of(kref, struct rbd_img_request, kref);
2272
2273         dout("%s: img %p\n", __func__, img_request);
2274
2275         for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2276                 rbd_img_obj_request_del(img_request, obj_request);
2277         rbd_assert(img_request->obj_request_count == 0);
2278
2279         if (img_request_layered_test(img_request)) {
2280                 img_request_layered_clear(img_request);
2281                 rbd_dev_parent_put(img_request->rbd_dev);
2282         }
2283
2284         if (img_request_write_test(img_request) ||
2285                 img_request_discard_test(img_request))
2286                 ceph_put_snap_context(img_request->snapc);
2287
2288         kmem_cache_free(rbd_img_request_cache, img_request);
2289 }
2290
2291 static struct rbd_img_request *rbd_parent_request_create(
2292                                         struct rbd_obj_request *obj_request,
2293                                         u64 img_offset, u64 length)
2294 {
2295         struct rbd_img_request *parent_request;
2296         struct rbd_device *rbd_dev;
2297
2298         rbd_assert(obj_request->img_request);
2299         rbd_dev = obj_request->img_request->rbd_dev;
2300
2301         parent_request = rbd_img_request_create(rbd_dev->parent, img_offset,
2302                                                 length, OBJ_OP_READ, NULL);
2303         if (!parent_request)
2304                 return NULL;
2305
2306         img_request_child_set(parent_request);
2307         rbd_obj_request_get(obj_request);
2308         parent_request->obj_request = obj_request;
2309
2310         return parent_request;
2311 }
2312
2313 static void rbd_parent_request_destroy(struct kref *kref)
2314 {
2315         struct rbd_img_request *parent_request;
2316         struct rbd_obj_request *orig_request;
2317
2318         parent_request = container_of(kref, struct rbd_img_request, kref);
2319         orig_request = parent_request->obj_request;
2320
2321         parent_request->obj_request = NULL;
2322         rbd_obj_request_put(orig_request);
2323         img_request_child_clear(parent_request);
2324
2325         rbd_img_request_destroy(kref);
2326 }
2327
2328 static bool rbd_img_obj_end_request(struct rbd_obj_request *obj_request)
2329 {
2330         struct rbd_img_request *img_request;
2331         unsigned int xferred;
2332         int result;
2333         bool more;
2334
2335         rbd_assert(obj_request_img_data_test(obj_request));
2336         img_request = obj_request->img_request;
2337
2338         rbd_assert(obj_request->xferred <= (u64)UINT_MAX);
2339         xferred = (unsigned int)obj_request->xferred;
2340         result = obj_request->result;
2341         if (result) {
2342                 struct rbd_device *rbd_dev = img_request->rbd_dev;
2343                 enum obj_operation_type op_type;
2344
2345                 if (img_request_discard_test(img_request))
2346                         op_type = OBJ_OP_DISCARD;
2347                 else if (img_request_write_test(img_request))
2348                         op_type = OBJ_OP_WRITE;
2349                 else
2350                         op_type = OBJ_OP_READ;
2351
2352                 rbd_warn(rbd_dev, "%s %llx at %llx (%llx)",
2353                         obj_op_name(op_type), obj_request->length,
2354                         obj_request->img_offset, obj_request->offset);
2355                 rbd_warn(rbd_dev, "  result %d xferred %x",
2356                         result, xferred);
2357                 if (!img_request->result)
2358                         img_request->result = result;
2359                 /*
2360                  * Need to end I/O on the entire obj_request worth of
2361                  * bytes in case of error.
2362                  */
2363                 xferred = obj_request->length;
2364         }
2365
2366         /* Image object requests don't own their page array */
2367
2368         if (obj_request->type == OBJ_REQUEST_PAGES) {
2369                 obj_request->pages = NULL;
2370                 obj_request->page_count = 0;
2371         }
2372
2373         if (img_request_child_test(img_request)) {
2374                 rbd_assert(img_request->obj_request != NULL);
2375                 more = obj_request->which < img_request->obj_request_count - 1;
2376         } else {
2377                 rbd_assert(img_request->rq != NULL);
2378
2379                 more = blk_update_request(img_request->rq, result, xferred);
2380                 if (!more)
2381                         __blk_mq_end_request(img_request->rq, result);
2382         }
2383
2384         return more;
2385 }
2386
2387 static void rbd_img_obj_callback(struct rbd_obj_request *obj_request)
2388 {
2389         struct rbd_img_request *img_request;
2390         u32 which = obj_request->which;
2391         bool more = true;
2392
2393         rbd_assert(obj_request_img_data_test(obj_request));
2394         img_request = obj_request->img_request;
2395
2396         dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
2397         rbd_assert(img_request != NULL);
2398         rbd_assert(img_request->obj_request_count > 0);
2399         rbd_assert(which != BAD_WHICH);
2400         rbd_assert(which < img_request->obj_request_count);
2401
2402         spin_lock_irq(&img_request->completion_lock);
2403         if (which != img_request->next_completion)
2404                 goto out;
2405
2406         for_each_obj_request_from(img_request, obj_request) {
2407                 rbd_assert(more);
2408                 rbd_assert(which < img_request->obj_request_count);
2409
2410                 if (!obj_request_done_test(obj_request))
2411                         break;
2412                 more = rbd_img_obj_end_request(obj_request);
2413                 which++;
2414         }
2415
2416         rbd_assert(more ^ (which == img_request->obj_request_count));
2417         img_request->next_completion = which;
2418 out:
2419         spin_unlock_irq(&img_request->completion_lock);
2420         rbd_img_request_put(img_request);
2421
2422         if (!more)
2423                 rbd_img_request_complete(img_request);
2424 }
2425
2426 /*
2427  * Add individual osd ops to the given ceph_osd_request and prepare
2428  * them for submission. num_ops is the current number of
2429  * osd operations already to the object request.
2430  */
2431 static void rbd_img_obj_request_fill(struct rbd_obj_request *obj_request,
2432                                 struct ceph_osd_request *osd_request,
2433                                 enum obj_operation_type op_type,
2434                                 unsigned int num_ops)
2435 {
2436         struct rbd_img_request *img_request = obj_request->img_request;
2437         struct rbd_device *rbd_dev = img_request->rbd_dev;
2438         u64 object_size = rbd_obj_bytes(&rbd_dev->header);
2439         u64 offset = obj_request->offset;
2440         u64 length = obj_request->length;
2441         u64 img_end;
2442         u16 opcode;
2443
2444         if (op_type == OBJ_OP_DISCARD) {
2445                 if (!offset && length == object_size &&
2446                     (!img_request_layered_test(img_request) ||
2447                      !obj_request_overlaps_parent(obj_request))) {
2448                         opcode = CEPH_OSD_OP_DELETE;
2449                 } else if ((offset + length == object_size)) {
2450                         opcode = CEPH_OSD_OP_TRUNCATE;
2451                 } else {
2452                         down_read(&rbd_dev->header_rwsem);
2453                         img_end = rbd_dev->header.image_size;
2454                         up_read(&rbd_dev->header_rwsem);
2455
2456                         if (obj_request->img_offset + length == img_end)
2457                                 opcode = CEPH_OSD_OP_TRUNCATE;
2458                         else
2459                                 opcode = CEPH_OSD_OP_ZERO;
2460                 }
2461         } else if (op_type == OBJ_OP_WRITE) {
2462                 if (!offset && length == object_size)
2463                         opcode = CEPH_OSD_OP_WRITEFULL;
2464                 else
2465                         opcode = CEPH_OSD_OP_WRITE;
2466                 osd_req_op_alloc_hint_init(osd_request, num_ops,
2467                                         object_size, object_size);
2468                 num_ops++;
2469         } else {
2470                 opcode = CEPH_OSD_OP_READ;
2471         }
2472
2473         if (opcode == CEPH_OSD_OP_DELETE)
2474                 osd_req_op_init(osd_request, num_ops, opcode, 0);
2475         else
2476                 osd_req_op_extent_init(osd_request, num_ops, opcode,
2477                                        offset, length, 0, 0);
2478
2479         if (obj_request->type == OBJ_REQUEST_BIO)
2480                 osd_req_op_extent_osd_data_bio(osd_request, num_ops,
2481                                         obj_request->bio_list, length);
2482         else if (obj_request->type == OBJ_REQUEST_PAGES)
2483                 osd_req_op_extent_osd_data_pages(osd_request, num_ops,
2484                                         obj_request->pages, length,
2485                                         offset & ~PAGE_MASK, false, false);
2486
2487         /* Discards are also writes */
2488         if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
2489                 rbd_osd_req_format_write(obj_request);
2490         else
2491                 rbd_osd_req_format_read(obj_request);
2492 }
2493
2494 /*
2495  * Split up an image request into one or more object requests, each
2496  * to a different object.  The "type" parameter indicates whether
2497  * "data_desc" is the pointer to the head of a list of bio
2498  * structures, or the base of a page array.  In either case this
2499  * function assumes data_desc describes memory sufficient to hold
2500  * all data described by the image request.
2501  */
2502 static int rbd_img_request_fill(struct rbd_img_request *img_request,
2503                                         enum obj_request_type type,
2504                                         void *data_desc)
2505 {
2506         struct rbd_device *rbd_dev = img_request->rbd_dev;
2507         struct rbd_obj_request *obj_request = NULL;
2508         struct rbd_obj_request *next_obj_request;
2509         struct bio *bio_list = NULL;
2510         unsigned int bio_offset = 0;
2511         struct page **pages = NULL;
2512         enum obj_operation_type op_type;
2513         u64 img_offset;
2514         u64 resid;
2515
2516         dout("%s: img %p type %d data_desc %p\n", __func__, img_request,
2517                 (int)type, data_desc);
2518
2519         img_offset = img_request->offset;
2520         resid = img_request->length;
2521         rbd_assert(resid > 0);
2522         op_type = rbd_img_request_op_type(img_request);
2523
2524         if (type == OBJ_REQUEST_BIO) {
2525                 bio_list = data_desc;
2526                 rbd_assert(img_offset ==
2527                            bio_list->bi_iter.bi_sector << SECTOR_SHIFT);
2528         } else if (type == OBJ_REQUEST_PAGES) {
2529                 pages = data_desc;
2530         }
2531
2532         while (resid) {
2533                 struct ceph_osd_request *osd_req;
2534                 const char *object_name;
2535                 u64 offset;
2536                 u64 length;
2537
2538                 object_name = rbd_segment_name(rbd_dev, img_offset);
2539                 if (!object_name)
2540                         goto out_unwind;
2541                 offset = rbd_segment_offset(rbd_dev, img_offset);
2542                 length = rbd_segment_length(rbd_dev, img_offset, resid);
2543                 obj_request = rbd_obj_request_create(object_name,
2544                                                 offset, length, type);
2545                 /* object request has its own copy of the object name */
2546                 rbd_segment_name_free(object_name);
2547                 if (!obj_request)
2548                         goto out_unwind;
2549
2550                 /*
2551                  * set obj_request->img_request before creating the
2552                  * osd_request so that it gets the right snapc
2553                  */
2554                 rbd_img_obj_request_add(img_request, obj_request);
2555
2556                 if (type == OBJ_REQUEST_BIO) {
2557                         unsigned int clone_size;
2558
2559                         rbd_assert(length <= (u64)UINT_MAX);
2560                         clone_size = (unsigned int)length;
2561                         obj_request->bio_list =
2562                                         bio_chain_clone_range(&bio_list,
2563                                                                 &bio_offset,
2564                                                                 clone_size,
2565                                                                 GFP_NOIO);
2566                         if (!obj_request->bio_list)
2567                                 goto out_unwind;
2568                 } else if (type == OBJ_REQUEST_PAGES) {
2569                         unsigned int page_count;
2570
2571                         obj_request->pages = pages;
2572                         page_count = (u32)calc_pages_for(offset, length);
2573                         obj_request->page_count = page_count;
2574                         if ((offset + length) & ~PAGE_MASK)
2575                                 page_count--;   /* more on last page */
2576                         pages += page_count;
2577                 }
2578
2579                 osd_req = rbd_osd_req_create(rbd_dev, op_type,
2580                                         (op_type == OBJ_OP_WRITE) ? 2 : 1,
2581                                         obj_request);
2582                 if (!osd_req)
2583                         goto out_unwind;
2584
2585                 obj_request->osd_req = osd_req;
2586                 obj_request->callback = rbd_img_obj_callback;
2587                 obj_request->img_offset = img_offset;
2588
2589                 rbd_img_obj_request_fill(obj_request, osd_req, op_type, 0);
2590
2591                 rbd_img_request_get(img_request);
2592
2593                 img_offset += length;
2594                 resid -= length;
2595         }
2596
2597         return 0;
2598
2599 out_unwind:
2600         for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2601                 rbd_img_obj_request_del(img_request, obj_request);
2602
2603         return -ENOMEM;
2604 }
2605
2606 static void
2607 rbd_osd_copyup_callback(struct rbd_obj_request *obj_request)
2608 {
2609         struct rbd_img_request *img_request;
2610         struct rbd_device *rbd_dev;
2611         struct page **pages;
2612         u32 page_count;
2613
2614         dout("%s: obj %p\n", __func__, obj_request);
2615
2616         rbd_assert(obj_request->type == OBJ_REQUEST_BIO ||
2617                 obj_request->type == OBJ_REQUEST_NODATA);
2618         rbd_assert(obj_request_img_data_test(obj_request));
2619         img_request = obj_request->img_request;
2620         rbd_assert(img_request);
2621
2622         rbd_dev = img_request->rbd_dev;
2623         rbd_assert(rbd_dev);
2624
2625         pages = obj_request->copyup_pages;
2626         rbd_assert(pages != NULL);
2627         obj_request->copyup_pages = NULL;
2628         page_count = obj_request->copyup_page_count;
2629         rbd_assert(page_count);
2630         obj_request->copyup_page_count = 0;
2631         ceph_release_page_vector(pages, page_count);
2632
2633         /*
2634          * We want the transfer count to reflect the size of the
2635          * original write request.  There is no such thing as a
2636          * successful short write, so if the request was successful
2637          * we can just set it to the originally-requested length.
2638          */
2639         if (!obj_request->result)
2640                 obj_request->xferred = obj_request->length;
2641
2642         obj_request_done_set(obj_request);
2643 }
2644
2645 static void
2646 rbd_img_obj_parent_read_full_callback(struct rbd_img_request *img_request)
2647 {
2648         struct rbd_obj_request *orig_request;
2649         struct ceph_osd_request *osd_req;
2650         struct rbd_device *rbd_dev;
2651         struct page **pages;
2652         enum obj_operation_type op_type;
2653         u32 page_count;
2654         int img_result;
2655         u64 parent_length;
2656
2657         rbd_assert(img_request_child_test(img_request));
2658
2659         /* First get what we need from the image request */
2660
2661         pages = img_request->copyup_pages;
2662         rbd_assert(pages != NULL);
2663         img_request->copyup_pages = NULL;
2664         page_count = img_request->copyup_page_count;
2665         rbd_assert(page_count);
2666         img_request->copyup_page_count = 0;
2667
2668         orig_request = img_request->obj_request;
2669         rbd_assert(orig_request != NULL);
2670         rbd_assert(obj_request_type_valid(orig_request->type));
2671         img_result = img_request->result;
2672         parent_length = img_request->length;
2673         rbd_assert(parent_length == img_request->xferred);
2674         rbd_img_request_put(img_request);
2675
2676         rbd_assert(orig_request->img_request);
2677         rbd_dev = orig_request->img_request->rbd_dev;
2678         rbd_assert(rbd_dev);
2679
2680         /*
2681          * If the overlap has become 0 (most likely because the
2682          * image has been flattened) we need to free the pages
2683          * and re-submit the original write request.
2684          */
2685         if (!rbd_dev->parent_overlap) {
2686                 ceph_release_page_vector(pages, page_count);
2687                 rbd_obj_request_submit(orig_request);
2688                 return;
2689         }
2690
2691         if (img_result)
2692                 goto out_err;
2693
2694         /*
2695          * The original osd request is of no use to use any more.
2696          * We need a new one that can hold the three ops in a copyup
2697          * request.  Allocate the new copyup osd request for the
2698          * original request, and release the old one.
2699          */
2700         img_result = -ENOMEM;
2701         osd_req = rbd_osd_req_create_copyup(orig_request);
2702         if (!osd_req)
2703                 goto out_err;
2704         rbd_osd_req_destroy(orig_request->osd_req);
2705         orig_request->osd_req = osd_req;
2706         orig_request->copyup_pages = pages;
2707         orig_request->copyup_page_count = page_count;
2708
2709         /* Initialize the copyup op */
2710
2711         osd_req_op_cls_init(osd_req, 0, CEPH_OSD_OP_CALL, "rbd", "copyup");
2712         osd_req_op_cls_request_data_pages(osd_req, 0, pages, parent_length, 0,
2713                                                 false, false);
2714
2715         /* Add the other op(s) */
2716
2717         op_type = rbd_img_request_op_type(orig_request->img_request);
2718         rbd_img_obj_request_fill(orig_request, osd_req, op_type, 1);
2719
2720         /* All set, send it off. */
2721
2722         rbd_obj_request_submit(orig_request);
2723         return;
2724
2725 out_err:
2726         /* Record the error code and complete the request */
2727
2728         orig_request->result = img_result;
2729         orig_request->xferred = 0;
2730         obj_request_done_set(orig_request);
2731         rbd_obj_request_complete(orig_request);
2732 }
2733
2734 /*
2735  * Read from the parent image the range of data that covers the
2736  * entire target of the given object request.  This is used for
2737  * satisfying a layered image write request when the target of an
2738  * object request from the image request does not exist.
2739  *
2740  * A page array big enough to hold the returned data is allocated
2741  * and supplied to rbd_img_request_fill() as the "data descriptor."
2742  * When the read completes, this page array will be transferred to
2743  * the original object request for the copyup operation.
2744  *
2745  * If an error occurs, record it as the result of the original
2746  * object request and mark it done so it gets completed.
2747  */
2748 static int rbd_img_obj_parent_read_full(struct rbd_obj_request *obj_request)
2749 {
2750         struct rbd_img_request *img_request = NULL;
2751         struct rbd_img_request *parent_request = NULL;
2752         struct rbd_device *rbd_dev;
2753         u64 img_offset;
2754         u64 length;
2755         struct page **pages = NULL;
2756         u32 page_count;
2757         int result;
2758
2759         rbd_assert(obj_request_img_data_test(obj_request));
2760         rbd_assert(obj_request_type_valid(obj_request->type));
2761
2762         img_request = obj_request->img_request;
2763         rbd_assert(img_request != NULL);
2764         rbd_dev = img_request->rbd_dev;
2765         rbd_assert(rbd_dev->parent != NULL);
2766
2767         /*
2768          * Determine the byte range covered by the object in the
2769          * child image to which the original request was to be sent.
2770          */
2771         img_offset = obj_request->img_offset - obj_request->offset;
2772         length = (u64)1 << rbd_dev->header.obj_order;
2773
2774         /*
2775          * There is no defined parent data beyond the parent
2776          * overlap, so limit what we read at that boundary if
2777          * necessary.
2778          */
2779         if (img_offset + length > rbd_dev->parent_overlap) {
2780                 rbd_assert(img_offset < rbd_dev->parent_overlap);
2781                 length = rbd_dev->parent_overlap - img_offset;
2782         }
2783
2784         /*
2785          * Allocate a page array big enough to receive the data read
2786          * from the parent.
2787          */
2788         page_count = (u32)calc_pages_for(0, length);
2789         pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2790         if (IS_ERR(pages)) {
2791                 result = PTR_ERR(pages);
2792                 pages = NULL;
2793                 goto out_err;
2794         }
2795
2796         result = -ENOMEM;
2797         parent_request = rbd_parent_request_create(obj_request,
2798                                                 img_offset, length);
2799         if (!parent_request)
2800                 goto out_err;
2801
2802         result = rbd_img_request_fill(parent_request, OBJ_REQUEST_PAGES, pages);
2803         if (result)
2804                 goto out_err;
2805         parent_request->copyup_pages = pages;
2806         parent_request->copyup_page_count = page_count;
2807
2808         parent_request->callback = rbd_img_obj_parent_read_full_callback;
2809         result = rbd_img_request_submit(parent_request);
2810         if (!result)
2811                 return 0;
2812
2813         parent_request->copyup_pages = NULL;
2814         parent_request->copyup_page_count = 0;
2815         parent_request->obj_request = NULL;
2816         rbd_obj_request_put(obj_request);
2817 out_err:
2818         if (pages)
2819                 ceph_release_page_vector(pages, page_count);
2820         if (parent_request)
2821                 rbd_img_request_put(parent_request);
2822         obj_request->result = result;
2823         obj_request->xferred = 0;
2824         obj_request_done_set(obj_request);
2825
2826         return result;
2827 }
2828
2829 static void rbd_img_obj_exists_callback(struct rbd_obj_request *obj_request)
2830 {
2831         struct rbd_obj_request *orig_request;
2832         struct rbd_device *rbd_dev;
2833         int result;
2834
2835         rbd_assert(!obj_request_img_data_test(obj_request));
2836
2837         /*
2838          * All we need from the object request is the original
2839          * request and the result of the STAT op.  Grab those, then
2840          * we're done with the request.
2841          */
2842         orig_request = obj_request->obj_request;
2843         obj_request->obj_request = NULL;
2844         rbd_obj_request_put(orig_request);
2845         rbd_assert(orig_request);
2846         rbd_assert(orig_request->img_request);
2847
2848         result = obj_request->result;
2849         obj_request->result = 0;
2850
2851         dout("%s: obj %p for obj %p result %d %llu/%llu\n", __func__,
2852                 obj_request, orig_request, result,
2853                 obj_request->xferred, obj_request->length);
2854         rbd_obj_request_put(obj_request);
2855
2856         /*
2857          * If the overlap has become 0 (most likely because the
2858          * image has been flattened) we need to re-submit the
2859          * original request.
2860          */
2861         rbd_dev = orig_request->img_request->rbd_dev;
2862         if (!rbd_dev->parent_overlap) {
2863                 rbd_obj_request_submit(orig_request);
2864                 return;
2865         }
2866
2867         /*
2868          * Our only purpose here is to determine whether the object
2869          * exists, and we don't want to treat the non-existence as
2870          * an error.  If something else comes back, transfer the
2871          * error to the original request and complete it now.
2872          */
2873         if (!result) {
2874                 obj_request_existence_set(orig_request, true);
2875         } else if (result == -ENOENT) {
2876                 obj_request_existence_set(orig_request, false);
2877         } else if (result) {
2878                 orig_request->result = result;
2879                 goto out;
2880         }
2881
2882         /*
2883          * Resubmit the original request now that we have recorded
2884          * whether the target object exists.
2885          */
2886         orig_request->result = rbd_img_obj_request_submit(orig_request);
2887 out:
2888         if (orig_request->result)
2889                 rbd_obj_request_complete(orig_request);
2890 }
2891
2892 static int rbd_img_obj_exists_submit(struct rbd_obj_request *obj_request)
2893 {
2894         struct rbd_obj_request *stat_request;
2895         struct rbd_device *rbd_dev;
2896         struct page **pages = NULL;
2897         u32 page_count;
2898         size_t size;
2899         int ret;
2900
2901         /*
2902          * The response data for a STAT call consists of:
2903          *     le64 length;
2904          *     struct {
2905          *         le32 tv_sec;
2906          *         le32 tv_nsec;
2907          *     } mtime;
2908          */
2909         size = sizeof (__le64) + sizeof (__le32) + sizeof (__le32);
2910         page_count = (u32)calc_pages_for(0, size);
2911         pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2912         if (IS_ERR(pages))
2913                 return PTR_ERR(pages);
2914
2915         ret = -ENOMEM;
2916         stat_request = rbd_obj_request_create(obj_request->object_name, 0, 0,
2917                                                         OBJ_REQUEST_PAGES);
2918         if (!stat_request)
2919                 goto out;
2920
2921         rbd_obj_request_get(obj_request);
2922         stat_request->obj_request = obj_request;
2923         stat_request->pages = pages;
2924         stat_request->page_count = page_count;
2925
2926         rbd_assert(obj_request->img_request);
2927         rbd_dev = obj_request->img_request->rbd_dev;
2928         stat_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
2929                                                    stat_request);
2930         if (!stat_request->osd_req)
2931                 goto out;
2932         stat_request->callback = rbd_img_obj_exists_callback;
2933
2934         osd_req_op_init(stat_request->osd_req, 0, CEPH_OSD_OP_STAT, 0);
2935         osd_req_op_raw_data_in_pages(stat_request->osd_req, 0, pages, size, 0,
2936                                         false, false);
2937         rbd_osd_req_format_read(stat_request);
2938
2939         rbd_obj_request_submit(stat_request);
2940         return 0;
2941
2942 out:
2943         if (ret)
2944                 rbd_obj_request_put(obj_request);
2945
2946         return ret;
2947 }
2948
2949 static bool img_obj_request_simple(struct rbd_obj_request *obj_request)
2950 {
2951         struct rbd_img_request *img_request;
2952         struct rbd_device *rbd_dev;
2953
2954         rbd_assert(obj_request_img_data_test(obj_request));
2955
2956         img_request = obj_request->img_request;
2957         rbd_assert(img_request);
2958         rbd_dev = img_request->rbd_dev;
2959
2960         /* Reads */
2961         if (!img_request_write_test(img_request) &&
2962             !img_request_discard_test(img_request))
2963                 return true;
2964
2965         /* Non-layered writes */
2966         if (!img_request_layered_test(img_request))
2967                 return true;
2968
2969         /*
2970          * Layered writes outside of the parent overlap range don't
2971          * share any data with the parent.
2972          */
2973         if (!obj_request_overlaps_parent(obj_request))
2974                 return true;
2975
2976         /*
2977          * Entire-object layered writes - we will overwrite whatever
2978          * parent data there is anyway.
2979          */
2980         if (!obj_request->offset &&
2981             obj_request->length == rbd_obj_bytes(&rbd_dev->header))
2982                 return true;
2983
2984         /*
2985          * If the object is known to already exist, its parent data has
2986          * already been copied.
2987          */
2988         if (obj_request_known_test(obj_request) &&
2989             obj_request_exists_test(obj_request))
2990                 return true;
2991
2992         return false;
2993 }
2994
2995 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request)
2996 {
2997         if (img_obj_request_simple(obj_request)) {
2998                 rbd_obj_request_submit(obj_request);
2999                 return 0;
3000         }
3001
3002         /*
3003          * It's a layered write.  The target object might exist but
3004          * we may not know that yet.  If we know it doesn't exist,
3005          * start by reading the data for the full target object from
3006          * the parent so we can use it for a copyup to the target.
3007          */
3008         if (obj_request_known_test(obj_request))
3009                 return rbd_img_obj_parent_read_full(obj_request);
3010
3011         /* We don't know whether the target exists.  Go find out. */
3012
3013         return rbd_img_obj_exists_submit(obj_request);
3014 }
3015
3016 static int rbd_img_request_submit(struct rbd_img_request *img_request)
3017 {
3018         struct rbd_obj_request *obj_request;
3019         struct rbd_obj_request *next_obj_request;
3020         int ret = 0;
3021
3022         dout("%s: img %p\n", __func__, img_request);
3023
3024         rbd_img_request_get(img_request);
3025         for_each_obj_request_safe(img_request, obj_request, next_obj_request) {
3026                 ret = rbd_img_obj_request_submit(obj_request);
3027                 if (ret)
3028                         goto out_put_ireq;
3029         }
3030
3031 out_put_ireq:
3032         rbd_img_request_put(img_request);
3033         return ret;
3034 }
3035
3036 static void rbd_img_parent_read_callback(struct rbd_img_request *img_request)
3037 {
3038         struct rbd_obj_request *obj_request;
3039         struct rbd_device *rbd_dev;
3040         u64 obj_end;
3041         u64 img_xferred;
3042         int img_result;
3043
3044         rbd_assert(img_request_child_test(img_request));
3045
3046         /* First get what we need from the image request and release it */
3047
3048         obj_request = img_request->obj_request;
3049         img_xferred = img_request->xferred;
3050         img_result = img_request->result;
3051         rbd_img_request_put(img_request);
3052
3053         /*
3054          * If the overlap has become 0 (most likely because the
3055          * image has been flattened) we need to re-submit the
3056          * original request.
3057          */
3058         rbd_assert(obj_request);
3059         rbd_assert(obj_request->img_request);
3060         rbd_dev = obj_request->img_request->rbd_dev;
3061         if (!rbd_dev->parent_overlap) {
3062                 rbd_obj_request_submit(obj_request);
3063                 return;
3064         }
3065
3066         obj_request->result = img_result;
3067         if (obj_request->result)
3068                 goto out;
3069
3070         /*
3071          * We need to zero anything beyond the parent overlap
3072          * boundary.  Since rbd_img_obj_request_read_callback()
3073          * will zero anything beyond the end of a short read, an
3074          * easy way to do this is to pretend the data from the
3075          * parent came up short--ending at the overlap boundary.
3076          */
3077         rbd_assert(obj_request->img_offset < U64_MAX - obj_request->length);
3078         obj_end = obj_request->img_offset + obj_request->length;
3079         if (obj_end > rbd_dev->parent_overlap) {
3080                 u64 xferred = 0;
3081
3082                 if (obj_request->img_offset < rbd_dev->parent_overlap)
3083                         xferred = rbd_dev->parent_overlap -
3084                                         obj_request->img_offset;
3085
3086                 obj_request->xferred = min(img_xferred, xferred);
3087         } else {
3088                 obj_request->xferred = img_xferred;
3089         }
3090 out:
3091         rbd_img_obj_request_read_callback(obj_request);
3092         rbd_obj_request_complete(obj_request);
3093 }
3094
3095 static void rbd_img_parent_read(struct rbd_obj_request *obj_request)
3096 {
3097         struct rbd_img_request *img_request;
3098         int result;
3099
3100         rbd_assert(obj_request_img_data_test(obj_request));
3101         rbd_assert(obj_request->img_request != NULL);
3102         rbd_assert(obj_request->result == (s32) -ENOENT);
3103         rbd_assert(obj_request_type_valid(obj_request->type));
3104
3105         /* rbd_read_finish(obj_request, obj_request->length); */
3106         img_request = rbd_parent_request_create(obj_request,
3107                                                 obj_request->img_offset,
3108                                                 obj_request->length);
3109         result = -ENOMEM;
3110         if (!img_request)
3111                 goto out_err;
3112
3113         if (obj_request->type == OBJ_REQUEST_BIO)
3114                 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3115                                                 obj_request->bio_list);
3116         else
3117                 result = rbd_img_request_fill(img_request, OBJ_REQUEST_PAGES,
3118                                                 obj_request->pages);
3119         if (result)
3120                 goto out_err;
3121
3122         img_request->callback = rbd_img_parent_read_callback;
3123         result = rbd_img_request_submit(img_request);
3124         if (result)
3125                 goto out_err;
3126
3127         return;
3128 out_err:
3129         if (img_request)
3130                 rbd_img_request_put(img_request);
3131         obj_request->result = result;
3132         obj_request->xferred = 0;
3133         obj_request_done_set(obj_request);
3134 }
3135
3136 static const struct rbd_client_id rbd_empty_cid;
3137
3138 static bool rbd_cid_equal(const struct rbd_client_id *lhs,
3139                           const struct rbd_client_id *rhs)
3140 {
3141         return lhs->gid == rhs->gid && lhs->handle == rhs->handle;
3142 }
3143
3144 static struct rbd_client_id rbd_get_cid(struct rbd_device *rbd_dev)
3145 {
3146         struct rbd_client_id cid;
3147
3148         mutex_lock(&rbd_dev->watch_mutex);
3149         cid.gid = ceph_client_gid(rbd_dev->rbd_client->client);
3150         cid.handle = rbd_dev->watch_cookie;
3151         mutex_unlock(&rbd_dev->watch_mutex);
3152         return cid;
3153 }
3154
3155 /*
3156  * lock_rwsem must be held for write
3157  */
3158 static void rbd_set_owner_cid(struct rbd_device *rbd_dev,
3159                               const struct rbd_client_id *cid)
3160 {
3161         dout("%s rbd_dev %p %llu-%llu -> %llu-%llu\n", __func__, rbd_dev,
3162              rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle,
3163              cid->gid, cid->handle);
3164         rbd_dev->owner_cid = *cid; /* struct */
3165 }
3166
3167 static void format_lock_cookie(struct rbd_device *rbd_dev, char *buf)
3168 {
3169         mutex_lock(&rbd_dev->watch_mutex);
3170         sprintf(buf, "%s %llu", RBD_LOCK_COOKIE_PREFIX, rbd_dev->watch_cookie);
3171         mutex_unlock(&rbd_dev->watch_mutex);
3172 }
3173
3174 /*
3175  * lock_rwsem must be held for write
3176  */
3177 static int rbd_lock(struct rbd_device *rbd_dev)
3178 {
3179         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3180         struct rbd_client_id cid = rbd_get_cid(rbd_dev);
3181         char cookie[32];
3182         int ret;
3183
3184         WARN_ON(__rbd_is_lock_owner(rbd_dev));
3185
3186         format_lock_cookie(rbd_dev, cookie);
3187         ret = ceph_cls_lock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
3188                             RBD_LOCK_NAME, CEPH_CLS_LOCK_EXCLUSIVE, cookie,
3189                             RBD_LOCK_TAG, "", 0);
3190         if (ret)
3191                 return ret;
3192
3193         rbd_dev->lock_state = RBD_LOCK_STATE_LOCKED;
3194         rbd_set_owner_cid(rbd_dev, &cid);
3195         queue_work(rbd_dev->task_wq, &rbd_dev->acquired_lock_work);
3196         return 0;
3197 }
3198
3199 /*
3200  * lock_rwsem must be held for write
3201  */
3202 static int rbd_unlock(struct rbd_device *rbd_dev)
3203 {
3204         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3205         char cookie[32];
3206         int ret;
3207
3208         WARN_ON(!__rbd_is_lock_owner(rbd_dev));
3209
3210         rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
3211
3212         format_lock_cookie(rbd_dev, cookie);
3213         ret = ceph_cls_unlock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
3214                               RBD_LOCK_NAME, cookie);
3215         if (ret && ret != -ENOENT) {
3216                 rbd_warn(rbd_dev, "cls_unlock failed: %d", ret);
3217                 return ret;
3218         }
3219
3220         rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3221         queue_work(rbd_dev->task_wq, &rbd_dev->released_lock_work);
3222         return 0;
3223 }
3224
3225 static int __rbd_notify_op_lock(struct rbd_device *rbd_dev,
3226                                 enum rbd_notify_op notify_op,
3227                                 struct page ***preply_pages,
3228                                 size_t *preply_len)
3229 {
3230         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3231         struct rbd_client_id cid = rbd_get_cid(rbd_dev);
3232         int buf_size = 4 + 8 + 8 + CEPH_ENCODING_START_BLK_LEN;
3233         char buf[buf_size];
3234         void *p = buf;
3235
3236         dout("%s rbd_dev %p notify_op %d\n", __func__, rbd_dev, notify_op);
3237
3238         /* encode *LockPayload NotifyMessage (op + ClientId) */
3239         ceph_start_encoding(&p, 2, 1, buf_size - CEPH_ENCODING_START_BLK_LEN);
3240         ceph_encode_32(&p, notify_op);
3241         ceph_encode_64(&p, cid.gid);
3242         ceph_encode_64(&p, cid.handle);
3243
3244         return ceph_osdc_notify(osdc, &rbd_dev->header_oid,
3245                                 &rbd_dev->header_oloc, buf, buf_size,
3246                                 RBD_NOTIFY_TIMEOUT, preply_pages, preply_len);
3247 }
3248
3249 static void rbd_notify_op_lock(struct rbd_device *rbd_dev,
3250                                enum rbd_notify_op notify_op)
3251 {
3252         struct page **reply_pages;
3253         size_t reply_len;
3254
3255         __rbd_notify_op_lock(rbd_dev, notify_op, &reply_pages, &reply_len);
3256         ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len));
3257 }
3258
3259 static void rbd_notify_acquired_lock(struct work_struct *work)
3260 {
3261         struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3262                                                   acquired_lock_work);
3263
3264         rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_ACQUIRED_LOCK);
3265 }
3266
3267 static void rbd_notify_released_lock(struct work_struct *work)
3268 {
3269         struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3270                                                   released_lock_work);
3271
3272         rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_RELEASED_LOCK);
3273 }
3274
3275 static int rbd_request_lock(struct rbd_device *rbd_dev)
3276 {
3277         struct page **reply_pages;
3278         size_t reply_len;
3279         bool lock_owner_responded = false;
3280         int ret;
3281
3282         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3283
3284         ret = __rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_REQUEST_LOCK,
3285                                    &reply_pages, &reply_len);
3286         if (ret && ret != -ETIMEDOUT) {
3287                 rbd_warn(rbd_dev, "failed to request lock: %d", ret);
3288                 goto out;
3289         }
3290
3291         if (reply_len > 0 && reply_len <= PAGE_SIZE) {
3292                 void *p = page_address(reply_pages[0]);
3293                 void *const end = p + reply_len;
3294                 u32 n;
3295
3296                 ceph_decode_32_safe(&p, end, n, e_inval); /* num_acks */
3297                 while (n--) {
3298                         u8 struct_v;
3299                         u32 len;
3300
3301                         ceph_decode_need(&p, end, 8 + 8, e_inval);
3302                         p += 8 + 8; /* skip gid and cookie */
3303
3304                         ceph_decode_32_safe(&p, end, len, e_inval);
3305                         if (!len)
3306                                 continue;
3307
3308                         if (lock_owner_responded) {
3309                                 rbd_warn(rbd_dev,
3310                                          "duplicate lock owners detected");
3311                                 ret = -EIO;
3312                                 goto out;
3313                         }
3314
3315                         lock_owner_responded = true;
3316                         ret = ceph_start_decoding(&p, end, 1, "ResponseMessage",
3317                                                   &struct_v, &len);
3318                         if (ret) {
3319                                 rbd_warn(rbd_dev,
3320                                          "failed to decode ResponseMessage: %d",
3321                                          ret);
3322                                 goto e_inval;
3323                         }
3324
3325                         ret = ceph_decode_32(&p);
3326                 }
3327         }
3328
3329         if (!lock_owner_responded) {
3330                 rbd_warn(rbd_dev, "no lock owners detected");
3331                 ret = -ETIMEDOUT;
3332         }
3333
3334 out:
3335         ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len));
3336         return ret;
3337
3338 e_inval:
3339         ret = -EINVAL;
3340         goto out;
3341 }
3342
3343 static void wake_requests(struct rbd_device *rbd_dev, bool wake_all)
3344 {
3345         dout("%s rbd_dev %p wake_all %d\n", __func__, rbd_dev, wake_all);
3346
3347         cancel_delayed_work(&rbd_dev->lock_dwork);
3348         if (wake_all)
3349                 wake_up_all(&rbd_dev->lock_waitq);
3350         else
3351                 wake_up(&rbd_dev->lock_waitq);
3352 }
3353
3354 static int get_lock_owner_info(struct rbd_device *rbd_dev,
3355                                struct ceph_locker **lockers, u32 *num_lockers)
3356 {
3357         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3358         u8 lock_type;
3359         char *lock_tag;
3360         int ret;
3361
3362         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3363
3364         ret = ceph_cls_lock_info(osdc, &rbd_dev->header_oid,
3365                                  &rbd_dev->header_oloc, RBD_LOCK_NAME,
3366                                  &lock_type, &lock_tag, lockers, num_lockers);
3367         if (ret)
3368                 return ret;
3369
3370         if (*num_lockers == 0) {
3371                 dout("%s rbd_dev %p no lockers detected\n", __func__, rbd_dev);
3372                 goto out;
3373         }
3374
3375         if (strcmp(lock_tag, RBD_LOCK_TAG)) {
3376                 rbd_warn(rbd_dev, "locked by external mechanism, tag %s",
3377                          lock_tag);
3378                 ret = -EBUSY;
3379                 goto out;
3380         }
3381
3382         if (lock_type == CEPH_CLS_LOCK_SHARED) {
3383                 rbd_warn(rbd_dev, "shared lock type detected");
3384                 ret = -EBUSY;
3385                 goto out;
3386         }
3387
3388         if (strncmp((*lockers)[0].id.cookie, RBD_LOCK_COOKIE_PREFIX,
3389                     strlen(RBD_LOCK_COOKIE_PREFIX))) {
3390                 rbd_warn(rbd_dev, "locked by external mechanism, cookie %s",
3391                          (*lockers)[0].id.cookie);
3392                 ret = -EBUSY;
3393                 goto out;
3394         }
3395
3396 out:
3397         kfree(lock_tag);
3398         return ret;
3399 }
3400
3401 static int find_watcher(struct rbd_device *rbd_dev,
3402                         const struct ceph_locker *locker)
3403 {
3404         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3405         struct ceph_watch_item *watchers;
3406         u32 num_watchers;
3407         u64 cookie;
3408         int i;
3409         int ret;
3410
3411         ret = ceph_osdc_list_watchers(osdc, &rbd_dev->header_oid,
3412                                       &rbd_dev->header_oloc, &watchers,
3413                                       &num_watchers);
3414         if (ret)
3415                 return ret;
3416
3417         sscanf(locker->id.cookie, RBD_LOCK_COOKIE_PREFIX " %llu", &cookie);
3418         for (i = 0; i < num_watchers; i++) {
3419                 if (!memcmp(&watchers[i].addr, &locker->info.addr,
3420                             sizeof(locker->info.addr)) &&
3421                     watchers[i].cookie == cookie) {
3422                         struct rbd_client_id cid = {
3423                                 .gid = le64_to_cpu(watchers[i].name.num),
3424                                 .handle = cookie,
3425                         };
3426
3427                         dout("%s rbd_dev %p found cid %llu-%llu\n", __func__,
3428                              rbd_dev, cid.gid, cid.handle);
3429                         rbd_set_owner_cid(rbd_dev, &cid);
3430                         ret = 1;
3431                         goto out;
3432                 }
3433         }
3434
3435         dout("%s rbd_dev %p no watchers\n", __func__, rbd_dev);
3436         ret = 0;
3437 out:
3438         kfree(watchers);
3439         return ret;
3440 }
3441
3442 /*
3443  * lock_rwsem must be held for write
3444  */
3445 static int rbd_try_lock(struct rbd_device *rbd_dev)
3446 {
3447         struct ceph_client *client = rbd_dev->rbd_client->client;
3448         struct ceph_locker *lockers;
3449         u32 num_lockers;
3450         int ret;
3451
3452         for (;;) {
3453                 ret = rbd_lock(rbd_dev);
3454                 if (ret != -EBUSY)
3455                         return ret;
3456
3457                 /* determine if the current lock holder is still alive */
3458                 ret = get_lock_owner_info(rbd_dev, &lockers, &num_lockers);
3459                 if (ret)
3460                         return ret;
3461
3462                 if (num_lockers == 0)
3463                         goto again;
3464
3465                 ret = find_watcher(rbd_dev, lockers);
3466                 if (ret) {
3467                         if (ret > 0)
3468                                 ret = 0; /* have to request lock */
3469                         goto out;
3470                 }
3471
3472                 rbd_warn(rbd_dev, "%s%llu seems dead, breaking lock",
3473                          ENTITY_NAME(lockers[0].id.name));
3474
3475                 ret = ceph_monc_blacklist_add(&client->monc,
3476                                               &lockers[0].info.addr);
3477                 if (ret) {
3478                         rbd_warn(rbd_dev, "blacklist of %s%llu failed: %d",
3479                                  ENTITY_NAME(lockers[0].id.name), ret);
3480                         goto out;
3481                 }
3482
3483                 ret = ceph_cls_break_lock(&client->osdc, &rbd_dev->header_oid,
3484                                           &rbd_dev->header_oloc, RBD_LOCK_NAME,
3485                                           lockers[0].id.cookie,
3486                                           &lockers[0].id.name);
3487                 if (ret && ret != -ENOENT)
3488                         goto out;
3489
3490 again:
3491                 ceph_free_lockers(lockers, num_lockers);
3492         }
3493
3494 out:
3495         ceph_free_lockers(lockers, num_lockers);
3496         return ret;
3497 }
3498
3499 /*
3500  * ret is set only if lock_state is RBD_LOCK_STATE_UNLOCKED
3501  */
3502 static enum rbd_lock_state rbd_try_acquire_lock(struct rbd_device *rbd_dev,
3503                                                 int *pret)
3504 {
3505         enum rbd_lock_state lock_state;
3506
3507         down_read(&rbd_dev->lock_rwsem);
3508         dout("%s rbd_dev %p read lock_state %d\n", __func__, rbd_dev,
3509              rbd_dev->lock_state);
3510         if (__rbd_is_lock_owner(rbd_dev)) {
3511                 lock_state = rbd_dev->lock_state;
3512                 up_read(&rbd_dev->lock_rwsem);
3513                 return lock_state;
3514         }
3515
3516         up_read(&rbd_dev->lock_rwsem);
3517         down_write(&rbd_dev->lock_rwsem);
3518         dout("%s rbd_dev %p write lock_state %d\n", __func__, rbd_dev,
3519              rbd_dev->lock_state);
3520         if (!__rbd_is_lock_owner(rbd_dev)) {
3521                 *pret = rbd_try_lock(rbd_dev);
3522                 if (*pret)
3523                         rbd_warn(rbd_dev, "failed to acquire lock: %d", *pret);
3524         }
3525
3526         lock_state = rbd_dev->lock_state;
3527         up_write(&rbd_dev->lock_rwsem);
3528         return lock_state;
3529 }
3530
3531 static void rbd_acquire_lock(struct work_struct *work)
3532 {
3533         struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
3534                                             struct rbd_device, lock_dwork);
3535         enum rbd_lock_state lock_state;
3536         int ret;
3537
3538         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3539 again:
3540         lock_state = rbd_try_acquire_lock(rbd_dev, &ret);
3541         if (lock_state != RBD_LOCK_STATE_UNLOCKED || ret == -EBLACKLISTED) {
3542                 if (lock_state == RBD_LOCK_STATE_LOCKED)
3543                         wake_requests(rbd_dev, true);
3544                 dout("%s rbd_dev %p lock_state %d ret %d - done\n", __func__,
3545                      rbd_dev, lock_state, ret);
3546                 return;
3547         }
3548
3549         ret = rbd_request_lock(rbd_dev);
3550         if (ret == -ETIMEDOUT) {
3551                 goto again; /* treat this as a dead client */
3552         } else if (ret < 0) {
3553                 rbd_warn(rbd_dev, "error requesting lock: %d", ret);
3554                 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
3555                                  RBD_RETRY_DELAY);
3556         } else {
3557                 /*
3558                  * lock owner acked, but resend if we don't see them
3559                  * release the lock
3560                  */
3561                 dout("%s rbd_dev %p requeueing lock_dwork\n", __func__,
3562                      rbd_dev);
3563                 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
3564                     msecs_to_jiffies(2 * RBD_NOTIFY_TIMEOUT * MSEC_PER_SEC));
3565         }
3566 }
3567
3568 /*
3569  * lock_rwsem must be held for write
3570  */
3571 static bool rbd_release_lock(struct rbd_device *rbd_dev)
3572 {
3573         dout("%s rbd_dev %p read lock_state %d\n", __func__, rbd_dev,
3574              rbd_dev->lock_state);
3575         if (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED)
3576                 return false;
3577
3578         rbd_dev->lock_state = RBD_LOCK_STATE_RELEASING;
3579         downgrade_write(&rbd_dev->lock_rwsem);
3580         /*
3581          * Ensure that all in-flight IO is flushed.
3582          *
3583          * FIXME: ceph_osdc_sync() flushes the entire OSD client, which
3584          * may be shared with other devices.
3585          */
3586         ceph_osdc_sync(&rbd_dev->rbd_client->client->osdc);
3587         up_read(&rbd_dev->lock_rwsem);
3588
3589         down_write(&rbd_dev->lock_rwsem);
3590         dout("%s rbd_dev %p write lock_state %d\n", __func__, rbd_dev,
3591              rbd_dev->lock_state);
3592         if (rbd_dev->lock_state != RBD_LOCK_STATE_RELEASING)
3593                 return false;
3594
3595         if (!rbd_unlock(rbd_dev))
3596                 /*
3597                  * Give others a chance to grab the lock - we would re-acquire
3598                  * almost immediately if we got new IO during ceph_osdc_sync()
3599                  * otherwise.  We need to ack our own notifications, so this
3600                  * lock_dwork will be requeued from rbd_wait_state_locked()
3601                  * after wake_requests() in rbd_handle_released_lock().
3602                  */
3603                 cancel_delayed_work(&rbd_dev->lock_dwork);
3604
3605         return true;
3606 }
3607
3608 static void rbd_release_lock_work(struct work_struct *work)
3609 {
3610         struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3611                                                   unlock_work);
3612
3613         down_write(&rbd_dev->lock_rwsem);
3614         rbd_release_lock(rbd_dev);
3615         up_write(&rbd_dev->lock_rwsem);
3616 }
3617
3618 static void rbd_handle_acquired_lock(struct rbd_device *rbd_dev, u8 struct_v,
3619                                      void **p)
3620 {
3621         struct rbd_client_id cid = { 0 };
3622
3623         if (struct_v >= 2) {
3624                 cid.gid = ceph_decode_64(p);
3625                 cid.handle = ceph_decode_64(p);
3626         }
3627
3628         dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3629              cid.handle);
3630         if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
3631                 down_write(&rbd_dev->lock_rwsem);
3632                 if (rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
3633                         /*
3634                          * we already know that the remote client is
3635                          * the owner
3636                          */
3637                         up_write(&rbd_dev->lock_rwsem);
3638                         return;
3639                 }
3640
3641                 rbd_set_owner_cid(rbd_dev, &cid);
3642                 downgrade_write(&rbd_dev->lock_rwsem);
3643         } else {
3644                 down_read(&rbd_dev->lock_rwsem);
3645         }
3646
3647         if (!__rbd_is_lock_owner(rbd_dev))
3648                 wake_requests(rbd_dev, false);
3649         up_read(&rbd_dev->lock_rwsem);
3650 }
3651
3652 static void rbd_handle_released_lock(struct rbd_device *rbd_dev, u8 struct_v,
3653                                      void **p)
3654 {
3655         struct rbd_client_id cid = { 0 };
3656
3657         if (struct_v >= 2) {
3658                 cid.gid = ceph_decode_64(p);
3659                 cid.handle = ceph_decode_64(p);
3660         }
3661
3662         dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3663              cid.handle);
3664         if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
3665                 down_write(&rbd_dev->lock_rwsem);
3666                 if (!rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
3667                         dout("%s rbd_dev %p unexpected owner, cid %llu-%llu != owner_cid %llu-%llu\n",
3668                              __func__, rbd_dev, cid.gid, cid.handle,
3669                              rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle);
3670                         up_write(&rbd_dev->lock_rwsem);
3671                         return;
3672                 }
3673
3674                 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3675                 downgrade_write(&rbd_dev->lock_rwsem);
3676         } else {
3677                 down_read(&rbd_dev->lock_rwsem);
3678         }
3679
3680         if (!__rbd_is_lock_owner(rbd_dev))
3681                 wake_requests(rbd_dev, false);
3682         up_read(&rbd_dev->lock_rwsem);
3683 }
3684
3685 static bool rbd_handle_request_lock(struct rbd_device *rbd_dev, u8 struct_v,
3686                                     void **p)
3687 {
3688         struct rbd_client_id my_cid = rbd_get_cid(rbd_dev);
3689         struct rbd_client_id cid = { 0 };
3690         bool need_to_send;
3691
3692         if (struct_v >= 2) {
3693                 cid.gid = ceph_decode_64(p);
3694                 cid.handle = ceph_decode_64(p);
3695         }
3696
3697         dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3698              cid.handle);
3699         if (rbd_cid_equal(&cid, &my_cid))
3700                 return false;
3701
3702         down_read(&rbd_dev->lock_rwsem);
3703         need_to_send = __rbd_is_lock_owner(rbd_dev);
3704         if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED) {
3705                 if (!rbd_cid_equal(&rbd_dev->owner_cid, &rbd_empty_cid)) {
3706                         dout("%s rbd_dev %p queueing unlock_work\n", __func__,
3707                              rbd_dev);
3708                         queue_work(rbd_dev->task_wq, &rbd_dev->unlock_work);
3709                 }
3710         }
3711         up_read(&rbd_dev->lock_rwsem);
3712         return need_to_send;
3713 }
3714
3715 static void __rbd_acknowledge_notify(struct rbd_device *rbd_dev,
3716                                      u64 notify_id, u64 cookie, s32 *result)
3717 {
3718         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3719         int buf_size = 4 + CEPH_ENCODING_START_BLK_LEN;
3720         char buf[buf_size];
3721         int ret;
3722
3723         if (result) {
3724                 void *p = buf;
3725
3726                 /* encode ResponseMessage */
3727                 ceph_start_encoding(&p, 1, 1,
3728                                     buf_size - CEPH_ENCODING_START_BLK_LEN);
3729                 ceph_encode_32(&p, *result);
3730         } else {
3731                 buf_size = 0;
3732         }
3733
3734         ret = ceph_osdc_notify_ack(osdc, &rbd_dev->header_oid,
3735                                    &rbd_dev->header_oloc, notify_id, cookie,
3736                                    buf, buf_size);
3737         if (ret)
3738                 rbd_warn(rbd_dev, "acknowledge_notify failed: %d", ret);
3739 }
3740
3741 static void rbd_acknowledge_notify(struct rbd_device *rbd_dev, u64 notify_id,
3742                                    u64 cookie)
3743 {
3744         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3745         __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, NULL);
3746 }
3747
3748 static void rbd_acknowledge_notify_result(struct rbd_device *rbd_dev,
3749                                           u64 notify_id, u64 cookie, s32 result)
3750 {
3751         dout("%s rbd_dev %p result %d\n", __func__, rbd_dev, result);
3752         __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, &result);
3753 }
3754
3755 static void rbd_watch_cb(void *arg, u64 notify_id, u64 cookie,
3756                          u64 notifier_id, void *data, size_t data_len)
3757 {
3758         struct rbd_device *rbd_dev = arg;
3759         void *p = data;
3760         void *const end = p + data_len;
3761         u8 struct_v;
3762         u32 len;
3763         u32 notify_op;
3764         int ret;
3765
3766         dout("%s rbd_dev %p cookie %llu notify_id %llu data_len %zu\n",
3767              __func__, rbd_dev, cookie, notify_id, data_len);
3768         if (data_len) {
3769                 ret = ceph_start_decoding(&p, end, 1, "NotifyMessage",
3770                                           &struct_v, &len);
3771                 if (ret) {
3772                         rbd_warn(rbd_dev, "failed to decode NotifyMessage: %d",
3773                                  ret);
3774                         return;
3775                 }
3776
3777                 notify_op = ceph_decode_32(&p);
3778         } else {
3779                 /* legacy notification for header updates */
3780                 notify_op = RBD_NOTIFY_OP_HEADER_UPDATE;
3781                 len = 0;
3782         }
3783
3784         dout("%s rbd_dev %p notify_op %u\n", __func__, rbd_dev, notify_op);
3785         switch (notify_op) {
3786         case RBD_NOTIFY_OP_ACQUIRED_LOCK:
3787                 rbd_handle_acquired_lock(rbd_dev, struct_v, &p);
3788                 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3789                 break;
3790         case RBD_NOTIFY_OP_RELEASED_LOCK:
3791                 rbd_handle_released_lock(rbd_dev, struct_v, &p);
3792                 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3793                 break;
3794         case RBD_NOTIFY_OP_REQUEST_LOCK:
3795                 if (rbd_handle_request_lock(rbd_dev, struct_v, &p))
3796                         /*
3797                          * send ResponseMessage(0) back so the client
3798                          * can detect a missing owner
3799                          */
3800                         rbd_acknowledge_notify_result(rbd_dev, notify_id,
3801                                                       cookie, 0);
3802                 else
3803                         rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3804                 break;
3805         case RBD_NOTIFY_OP_HEADER_UPDATE:
3806                 ret = rbd_dev_refresh(rbd_dev);
3807                 if (ret)
3808                         rbd_warn(rbd_dev, "refresh failed: %d", ret);
3809
3810                 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3811                 break;
3812         default:
3813                 if (rbd_is_lock_owner(rbd_dev))
3814                         rbd_acknowledge_notify_result(rbd_dev, notify_id,
3815                                                       cookie, -EOPNOTSUPP);
3816                 else
3817                         rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3818                 break;
3819         }
3820 }
3821
3822 static void __rbd_unregister_watch(struct rbd_device *rbd_dev);
3823
3824 static void rbd_watch_errcb(void *arg, u64 cookie, int err)
3825 {
3826         struct rbd_device *rbd_dev = arg;
3827
3828         rbd_warn(rbd_dev, "encountered watch error: %d", err);
3829
3830         down_write(&rbd_dev->lock_rwsem);
3831         rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3832         up_write(&rbd_dev->lock_rwsem);
3833
3834         mutex_lock(&rbd_dev->watch_mutex);
3835         if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED) {
3836                 __rbd_unregister_watch(rbd_dev);
3837                 rbd_dev->watch_state = RBD_WATCH_STATE_ERROR;
3838
3839                 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->watch_dwork, 0);
3840         }
3841         mutex_unlock(&rbd_dev->watch_mutex);
3842 }
3843
3844 /*
3845  * watch_mutex must be locked
3846  */
3847 static int __rbd_register_watch(struct rbd_device *rbd_dev)
3848 {
3849         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3850         struct ceph_osd_linger_request *handle;
3851
3852         rbd_assert(!rbd_dev->watch_handle);
3853         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3854
3855         handle = ceph_osdc_watch(osdc, &rbd_dev->header_oid,
3856                                  &rbd_dev->header_oloc, rbd_watch_cb,
3857                                  rbd_watch_errcb, rbd_dev);
3858         if (IS_ERR(handle))
3859                 return PTR_ERR(handle);
3860
3861         rbd_dev->watch_handle = handle;
3862         return 0;
3863 }
3864
3865 /*
3866  * watch_mutex must be locked
3867  */
3868 static void __rbd_unregister_watch(struct rbd_device *rbd_dev)
3869 {
3870         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3871         int ret;
3872
3873         rbd_assert(rbd_dev->watch_handle);
3874         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3875
3876         ret = ceph_osdc_unwatch(osdc, rbd_dev->watch_handle);
3877         if (ret)
3878                 rbd_warn(rbd_dev, "failed to unwatch: %d", ret);
3879
3880         rbd_dev->watch_handle = NULL;
3881 }
3882
3883 static int rbd_register_watch(struct rbd_device *rbd_dev)
3884 {
3885         int ret;
3886
3887         mutex_lock(&rbd_dev->watch_mutex);
3888         rbd_assert(rbd_dev->watch_state == RBD_WATCH_STATE_UNREGISTERED);
3889         ret = __rbd_register_watch(rbd_dev);
3890         if (ret)
3891                 goto out;
3892
3893         rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
3894         rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
3895
3896 out:
3897         mutex_unlock(&rbd_dev->watch_mutex);
3898         return ret;
3899 }
3900
3901 static void cancel_tasks_sync(struct rbd_device *rbd_dev)
3902 {
3903         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3904
3905         cancel_delayed_work_sync(&rbd_dev->watch_dwork);
3906         cancel_work_sync(&rbd_dev->acquired_lock_work);
3907         cancel_work_sync(&rbd_dev->released_lock_work);
3908         cancel_delayed_work_sync(&rbd_dev->lock_dwork);
3909         cancel_work_sync(&rbd_dev->unlock_work);
3910 }
3911
3912 static void rbd_unregister_watch(struct rbd_device *rbd_dev)
3913 {
3914         WARN_ON(waitqueue_active(&rbd_dev->lock_waitq));
3915         cancel_tasks_sync(rbd_dev);
3916
3917         mutex_lock(&rbd_dev->watch_mutex);
3918         if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED)
3919                 __rbd_unregister_watch(rbd_dev);
3920         rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
3921         mutex_unlock(&rbd_dev->watch_mutex);
3922
3923         ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
3924 }
3925
3926 static void rbd_reregister_watch(struct work_struct *work)
3927 {
3928         struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
3929                                             struct rbd_device, watch_dwork);
3930         bool was_lock_owner = false;
3931         int ret;
3932
3933         dout("%s rbd_dev %p\n", __func__, rbd_dev);
3934
3935         down_write(&rbd_dev->lock_rwsem);
3936         if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED)
3937                 was_lock_owner = rbd_release_lock(rbd_dev);
3938
3939         mutex_lock(&rbd_dev->watch_mutex);
3940         if (rbd_dev->watch_state != RBD_WATCH_STATE_ERROR)
3941                 goto fail_unlock;
3942
3943         ret = __rbd_register_watch(rbd_dev);
3944         if (ret) {
3945                 rbd_warn(rbd_dev, "failed to reregister watch: %d", ret);
3946                 if (ret != -EBLACKLISTED)
3947                         queue_delayed_work(rbd_dev->task_wq,
3948                                            &rbd_dev->watch_dwork,
3949                                            RBD_RETRY_DELAY);
3950                 goto fail_unlock;
3951         }
3952
3953         rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
3954         rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
3955         mutex_unlock(&rbd_dev->watch_mutex);
3956
3957         ret = rbd_dev_refresh(rbd_dev);
3958         if (ret)
3959                 rbd_warn(rbd_dev, "reregisteration refresh failed: %d", ret);
3960
3961         if (was_lock_owner) {
3962                 ret = rbd_try_lock(rbd_dev);
3963                 if (ret)
3964                         rbd_warn(rbd_dev, "reregisteration lock failed: %d",
3965                                  ret);
3966         }
3967
3968         up_write(&rbd_dev->lock_rwsem);
3969         wake_requests(rbd_dev, true);
3970         return;
3971
3972 fail_unlock:
3973         mutex_unlock(&rbd_dev->watch_mutex);
3974         up_write(&rbd_dev->lock_rwsem);
3975 }
3976
3977 /*
3978  * Synchronous osd object method call.  Returns the number of bytes
3979  * returned in the outbound buffer, or a negative error code.
3980  */
3981 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
3982                              const char *object_name,
3983                              const char *class_name,
3984                              const char *method_name,
3985                              const void *outbound,
3986                              size_t outbound_size,
3987                              void *inbound,
3988                              size_t inbound_size)
3989 {
3990         struct rbd_obj_request *obj_request;
3991         struct page **pages;
3992         u32 page_count;
3993         int ret;
3994
3995         /*
3996          * Method calls are ultimately read operations.  The result
3997          * should placed into the inbound buffer provided.  They
3998          * also supply outbound data--parameters for the object
3999          * method.  Currently if this is present it will be a
4000          * snapshot id.
4001          */
4002         page_count = (u32)calc_pages_for(0, inbound_size);
4003         pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
4004         if (IS_ERR(pages))
4005                 return PTR_ERR(pages);
4006
4007         ret = -ENOMEM;
4008         obj_request = rbd_obj_request_create(object_name, 0, inbound_size,
4009                                                         OBJ_REQUEST_PAGES);
4010         if (!obj_request)
4011                 goto out;
4012
4013         obj_request->pages = pages;
4014         obj_request->page_count = page_count;
4015
4016         obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
4017                                                   obj_request);
4018         if (!obj_request->osd_req)
4019                 goto out;
4020
4021         osd_req_op_cls_init(obj_request->osd_req, 0, CEPH_OSD_OP_CALL,
4022                                         class_name, method_name);
4023         if (outbound_size) {
4024                 struct ceph_pagelist *pagelist;
4025
4026                 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
4027                 if (!pagelist)
4028                         goto out;
4029
4030                 ceph_pagelist_init(pagelist);
4031                 ceph_pagelist_append(pagelist, outbound, outbound_size);
4032                 osd_req_op_cls_request_data_pagelist(obj_request->osd_req, 0,
4033                                                 pagelist);
4034         }
4035         osd_req_op_cls_response_data_pages(obj_request->osd_req, 0,
4036                                         obj_request->pages, inbound_size,
4037                                         0, false, false);
4038         rbd_osd_req_format_read(obj_request);
4039
4040         rbd_obj_request_submit(obj_request);
4041         ret = rbd_obj_request_wait(obj_request);
4042         if (ret)
4043                 goto out;
4044
4045         ret = obj_request->result;
4046         if (ret < 0)
4047                 goto out;
4048
4049         rbd_assert(obj_request->xferred < (u64)INT_MAX);
4050         ret = (int)obj_request->xferred;
4051         ceph_copy_from_page_vector(pages, inbound, 0, obj_request->xferred);
4052 out:
4053         if (obj_request)
4054                 rbd_obj_request_put(obj_request);
4055         else
4056                 ceph_release_page_vector(pages, page_count);
4057
4058         return ret;
4059 }
4060
4061 /*
4062  * lock_rwsem must be held for read
4063  */
4064 static void rbd_wait_state_locked(struct rbd_device *rbd_dev)
4065 {
4066         DEFINE_WAIT(wait);
4067
4068         do {
4069                 /*
4070                  * Note the use of mod_delayed_work() in rbd_acquire_lock()
4071                  * and cancel_delayed_work() in wake_requests().
4072                  */
4073                 dout("%s rbd_dev %p queueing lock_dwork\n", __func__, rbd_dev);
4074                 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork, 0);
4075                 prepare_to_wait_exclusive(&rbd_dev->lock_waitq, &wait,
4076                                           TASK_UNINTERRUPTIBLE);
4077                 up_read(&rbd_dev->lock_rwsem);
4078                 schedule();
4079                 down_read(&rbd_dev->lock_rwsem);
4080         } while (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED);
4081         finish_wait(&rbd_dev->lock_waitq, &wait);
4082 }
4083
4084 static void rbd_queue_workfn(struct work_struct *work)
4085 {
4086         struct request *rq = blk_mq_rq_from_pdu(work);
4087         struct rbd_device *rbd_dev = rq->q->queuedata;
4088         struct rbd_img_request *img_request;
4089         struct ceph_snap_context *snapc = NULL;
4090         u64 offset = (u64)blk_rq_pos(rq) << SECTOR_SHIFT;
4091         u64 length = blk_rq_bytes(rq);
4092         enum obj_operation_type op_type;
4093         u64 mapping_size;
4094         bool must_be_locked;
4095         int result;
4096
4097         if (rq->cmd_type != REQ_TYPE_FS) {
4098                 dout("%s: non-fs request type %d\n", __func__,
4099                         (int) rq->cmd_type);
4100                 result = -EIO;
4101                 goto err;
4102         }
4103
4104         if (req_op(rq) == REQ_OP_DISCARD)
4105                 op_type = OBJ_OP_DISCARD;
4106         else if (req_op(rq) == REQ_OP_WRITE)
4107                 op_type = OBJ_OP_WRITE;
4108         else
4109                 op_type = OBJ_OP_READ;
4110
4111         /* Ignore/skip any zero-length requests */
4112
4113         if (!length) {
4114                 dout("%s: zero-length request\n", __func__);
4115                 result = 0;
4116                 goto err_rq;
4117         }
4118
4119         /* Only reads are allowed to a read-only device */
4120
4121         if (op_type != OBJ_OP_READ) {
4122                 if (rbd_dev->mapping.read_only) {
4123                         result = -EROFS;
4124                         goto err_rq;
4125                 }
4126                 rbd_assert(rbd_dev->spec->snap_id == CEPH_NOSNAP);
4127         }
4128
4129         /*
4130          * Quit early if the mapped snapshot no longer exists.  It's
4131          * still possible the snapshot will have disappeared by the
4132          * time our request arrives at the osd, but there's no sense in
4133          * sending it if we already know.
4134          */
4135         if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
4136                 dout("request for non-existent snapshot");
4137                 rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
4138                 result = -ENXIO;
4139                 goto err_rq;
4140         }
4141
4142         if (offset && length > U64_MAX - offset + 1) {
4143                 rbd_warn(rbd_dev, "bad request range (%llu~%llu)", offset,
4144                          length);
4145                 result = -EINVAL;
4146                 goto err_rq;    /* Shouldn't happen */
4147         }
4148
4149         blk_mq_start_request(rq);
4150
4151         down_read(&rbd_dev->header_rwsem);
4152         mapping_size = rbd_dev->mapping.size;
4153         if (op_type != OBJ_OP_READ) {
4154                 snapc = rbd_dev->header.snapc;
4155                 ceph_get_snap_context(snapc);
4156                 must_be_locked = rbd_is_lock_supported(rbd_dev);
4157         } else {
4158                 must_be_locked = rbd_dev->opts->lock_on_read &&
4159                                         rbd_is_lock_supported(rbd_dev);
4160         }
4161         up_read(&rbd_dev->header_rwsem);
4162
4163         if (offset + length > mapping_size) {
4164                 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)", offset,
4165                          length, mapping_size);
4166                 result = -EIO;
4167                 goto err_rq;
4168         }
4169
4170         if (must_be_locked) {
4171                 down_read(&rbd_dev->lock_rwsem);
4172                 if (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED)
4173                         rbd_wait_state_locked(rbd_dev);
4174         }
4175
4176         img_request = rbd_img_request_create(rbd_dev, offset, length, op_type,
4177                                              snapc);
4178         if (!img_request) {
4179                 result = -ENOMEM;
4180                 goto err_unlock;
4181         }
4182         img_request->rq = rq;
4183         snapc = NULL; /* img_request consumes a ref */
4184
4185         if (op_type == OBJ_OP_DISCARD)
4186                 result = rbd_img_request_fill(img_request, OBJ_REQUEST_NODATA,
4187                                               NULL);
4188         else
4189                 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
4190                                               rq->bio);
4191         if (result)
4192                 goto err_img_request;
4193
4194         result = rbd_img_request_submit(img_request);
4195         if (result)
4196                 goto err_img_request;
4197
4198         if (must_be_locked)
4199                 up_read(&rbd_dev->lock_rwsem);
4200         return;
4201
4202 err_img_request:
4203         rbd_img_request_put(img_request);
4204 err_unlock:
4205         if (must_be_locked)
4206                 up_read(&rbd_dev->lock_rwsem);
4207 err_rq:
4208         if (result)
4209                 rbd_warn(rbd_dev, "%s %llx at %llx result %d",
4210                          obj_op_name(op_type), length, offset, result);
4211         ceph_put_snap_context(snapc);
4212 err:
4213         blk_mq_end_request(rq, result);
4214 }
4215
4216 static int rbd_queue_rq(struct blk_mq_hw_ctx *hctx,
4217                 const struct blk_mq_queue_data *bd)
4218 {
4219         struct request *rq = bd->rq;
4220         struct work_struct *work = blk_mq_rq_to_pdu(rq);
4221
4222         queue_work(rbd_wq, work);
4223         return BLK_MQ_RQ_QUEUE_OK;
4224 }
4225
4226 static void rbd_free_disk(struct rbd_device *rbd_dev)
4227 {
4228         struct gendisk *disk = rbd_dev->disk;
4229
4230         if (!disk)
4231                 return;
4232
4233         rbd_dev->disk = NULL;
4234         if (disk->flags & GENHD_FL_UP) {
4235                 del_gendisk(disk);
4236                 if (disk->queue)
4237                         blk_cleanup_queue(disk->queue);
4238                 blk_mq_free_tag_set(&rbd_dev->tag_set);
4239         }
4240         put_disk(disk);
4241 }
4242
4243 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
4244                                 const char *object_name,
4245                                 u64 offset, u64 length, void *buf)
4246
4247 {
4248         struct rbd_obj_request *obj_request;
4249         struct page **pages = NULL;
4250         u32 page_count;
4251         size_t size;
4252         int ret;
4253
4254         page_count = (u32) calc_pages_for(offset, length);
4255         pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
4256         if (IS_ERR(pages))
4257                 return PTR_ERR(pages);
4258
4259         ret = -ENOMEM;
4260         obj_request = rbd_obj_request_create(object_name, offset, length,
4261                                                         OBJ_REQUEST_PAGES);
4262         if (!obj_request)
4263                 goto out;
4264
4265         obj_request->pages = pages;
4266         obj_request->page_count = page_count;
4267
4268         obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
4269                                                   obj_request);
4270         if (!obj_request->osd_req)
4271                 goto out;
4272
4273         osd_req_op_extent_init(obj_request->osd_req, 0, CEPH_OSD_OP_READ,
4274                                         offset, length, 0, 0);
4275         osd_req_op_extent_osd_data_pages(obj_request->osd_req, 0,
4276                                         obj_request->pages,
4277                                         obj_request->length,
4278                                         obj_request->offset & ~PAGE_MASK,
4279                                         false, false);
4280         rbd_osd_req_format_read(obj_request);
4281
4282         rbd_obj_request_submit(obj_request);
4283         ret = rbd_obj_request_wait(obj_request);
4284         if (ret)
4285                 goto out;
4286
4287         ret = obj_request->result;
4288         if (ret < 0)
4289                 goto out;
4290
4291         rbd_assert(obj_request->xferred <= (u64) SIZE_MAX);
4292         size = (size_t) obj_request->xferred;
4293         ceph_copy_from_page_vector(pages, buf, 0, size);
4294         rbd_assert(size <= (size_t)INT_MAX);
4295         ret = (int)size;
4296 out:
4297         if (obj_request)
4298                 rbd_obj_request_put(obj_request);
4299         else
4300                 ceph_release_page_vector(pages, page_count);
4301
4302         return ret;
4303 }
4304
4305 /*
4306  * Read the complete header for the given rbd device.  On successful
4307  * return, the rbd_dev->header field will contain up-to-date
4308  * information about the image.
4309  */
4310 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev)
4311 {
4312         struct rbd_image_header_ondisk *ondisk = NULL;
4313         u32 snap_count = 0;
4314         u64 names_size = 0;
4315         u32 want_count;
4316         int ret;
4317
4318         /*
4319          * The complete header will include an array of its 64-bit
4320          * snapshot ids, followed by the names of those snapshots as
4321          * a contiguous block of NUL-terminated strings.  Note that
4322          * the number of snapshots could change by the time we read
4323          * it in, in which case we re-read it.
4324          */
4325         do {
4326                 size_t size;
4327
4328                 kfree(ondisk);
4329
4330                 size = sizeof (*ondisk);
4331                 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
4332                 size += names_size;
4333                 ondisk = kmalloc(size, GFP_KERNEL);
4334                 if (!ondisk)
4335                         return -ENOMEM;
4336
4337                 ret = rbd_obj_read_sync(rbd_dev, rbd_dev->header_oid.name,
4338                                        0, size, ondisk);
4339                 if (ret < 0)
4340                         goto out;
4341                 if ((size_t)ret < size) {
4342                         ret = -ENXIO;
4343                         rbd_warn(rbd_dev, "short header read (want %zd got %d)",
4344                                 size, ret);
4345                         goto out;
4346                 }
4347                 if (!rbd_dev_ondisk_valid(ondisk)) {
4348                         ret = -ENXIO;
4349                         rbd_warn(rbd_dev, "invalid header");
4350                         goto out;
4351                 }
4352
4353                 names_size = le64_to_cpu(ondisk->snap_names_len);
4354                 want_count = snap_count;
4355                 snap_count = le32_to_cpu(ondisk->snap_count);
4356         } while (snap_count != want_count);
4357
4358         ret = rbd_header_from_disk(rbd_dev, ondisk);
4359 out:
4360         kfree(ondisk);
4361
4362         return ret;
4363 }
4364
4365 /*
4366  * Clear the rbd device's EXISTS flag if the snapshot it's mapped to
4367  * has disappeared from the (just updated) snapshot context.
4368  */
4369 static void rbd_exists_validate(struct rbd_device *rbd_dev)
4370 {
4371         u64 snap_id;
4372
4373         if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags))
4374                 return;
4375
4376         snap_id = rbd_dev->spec->snap_id;
4377         if (snap_id == CEPH_NOSNAP)
4378                 return;
4379
4380         if (rbd_dev_snap_index(rbd_dev, snap_id) == BAD_SNAP_INDEX)
4381                 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
4382 }
4383
4384 static void rbd_dev_update_size(struct rbd_device *rbd_dev)
4385 {
4386         sector_t size;
4387
4388         /*
4389          * If EXISTS is not set, rbd_dev->disk may be NULL, so don't
4390          * try to update its size.  If REMOVING is set, updating size
4391          * is just useless work since the device can't be opened.
4392          */
4393         if (test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags) &&
4394             !test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags)) {
4395                 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
4396                 dout("setting size to %llu sectors", (unsigned long long)size);
4397                 set_capacity(rbd_dev->disk, size);
4398                 revalidate_disk(rbd_dev->disk);
4399         }
4400 }
4401
4402 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
4403 {
4404         u64 mapping_size;
4405         int ret;
4406
4407         down_write(&rbd_dev->header_rwsem);
4408         mapping_size = rbd_dev->mapping.size;
4409
4410         ret = rbd_dev_header_info(rbd_dev);
4411         if (ret)
4412                 goto out;
4413
4414         /*
4415          * If there is a parent, see if it has disappeared due to the
4416          * mapped image getting flattened.
4417          */
4418         if (rbd_dev->parent) {
4419                 ret = rbd_dev_v2_parent_info(rbd_dev);
4420                 if (ret)
4421                         goto out;
4422         }
4423
4424         if (rbd_dev->spec->snap_id == CEPH_NOSNAP) {
4425                 rbd_dev->mapping.size = rbd_dev->header.image_size;
4426         } else {
4427                 /* validate mapped snapshot's EXISTS flag */
4428                 rbd_exists_validate(rbd_dev);
4429         }
4430
4431 out:
4432         up_write(&rbd_dev->header_rwsem);
4433         if (!ret && mapping_size != rbd_dev->mapping.size)
4434                 rbd_dev_update_size(rbd_dev);
4435
4436         return ret;
4437 }
4438
4439 static int rbd_init_request(void *data, struct request *rq,
4440                 unsigned int hctx_idx, unsigned int request_idx,
4441                 unsigned int numa_node)
4442 {
4443         struct work_struct *work = blk_mq_rq_to_pdu(rq);
4444
4445         INIT_WORK(work, rbd_queue_workfn);
4446         return 0;
4447 }
4448
4449 static struct blk_mq_ops rbd_mq_ops = {
4450         .queue_rq       = rbd_queue_rq,
4451         .map_queue      = blk_mq_map_queue,
4452         .init_request   = rbd_init_request,
4453 };
4454
4455 static int rbd_init_disk(struct rbd_device *rbd_dev)
4456 {
4457         struct gendisk *disk;
4458         struct request_queue *q;
4459         u64 segment_size;
4460         int err;
4461
4462         /* create gendisk info */
4463         disk = alloc_disk(single_major ?
4464                           (1 << RBD_SINGLE_MAJOR_PART_SHIFT) :
4465                           RBD_MINORS_PER_MAJOR);
4466         if (!disk)
4467                 return -ENOMEM;
4468
4469         snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
4470                  rbd_dev->dev_id);
4471         disk->major = rbd_dev->major;
4472         disk->first_minor = rbd_dev->minor;
4473         if (single_major)
4474                 disk->flags |= GENHD_FL_EXT_DEVT;
4475         disk->fops = &rbd_bd_ops;
4476         disk->private_data = rbd_dev;
4477
4478         memset(&rbd_dev->tag_set, 0, sizeof(rbd_dev->tag_set));
4479         rbd_dev->tag_set.ops = &rbd_mq_ops;
4480         rbd_dev->tag_set.queue_depth = rbd_dev->opts->queue_depth;
4481         rbd_dev->tag_set.numa_node = NUMA_NO_NODE;
4482         rbd_dev->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
4483         rbd_dev->tag_set.nr_hw_queues = 1;
4484         rbd_dev->tag_set.cmd_size = sizeof(struct work_struct);
4485
4486         err = blk_mq_alloc_tag_set(&rbd_dev->tag_set);
4487         if (err)
4488                 goto out_disk;
4489
4490         q = blk_mq_init_queue(&rbd_dev->tag_set);
4491         if (IS_ERR(q)) {
4492                 err = PTR_ERR(q);
4493                 goto out_tag_set;
4494         }
4495
4496         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
4497         /* QUEUE_FLAG_ADD_RANDOM is off by default for blk-mq */
4498
4499         /* set io sizes to object size */
4500         segment_size = rbd_obj_bytes(&rbd_dev->header);
4501         blk_queue_max_hw_sectors(q, segment_size / SECTOR_SIZE);
4502         q->limits.max_sectors = queue_max_hw_sectors(q);
4503         blk_queue_max_segments(q, segment_size / SECTOR_SIZE);
4504         blk_queue_max_segment_size(q, segment_size);
4505         blk_queue_io_min(q, segment_size);
4506         blk_queue_io_opt(q, segment_size);
4507
4508         /* enable the discard support */
4509         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
4510         q->limits.discard_granularity = segment_size;
4511         q->limits.discard_alignment = segment_size;
4512         blk_queue_max_discard_sectors(q, segment_size / SECTOR_SIZE);
4513         q->limits.discard_zeroes_data = 1;
4514
4515         if (!ceph_test_opt(rbd_dev->rbd_client->client, NOCRC))
4516                 q->backing_dev_info.capabilities |= BDI_CAP_STABLE_WRITES;
4517
4518         disk->queue = q;
4519
4520         q->queuedata = rbd_dev;
4521
4522         rbd_dev->disk = disk;
4523
4524         return 0;
4525 out_tag_set:
4526         blk_mq_free_tag_set(&rbd_dev->tag_set);
4527 out_disk:
4528         put_disk(disk);
4529         return err;
4530 }
4531
4532 /*
4533   sysfs
4534 */
4535
4536 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
4537 {
4538         return container_of(dev, struct rbd_device, dev);
4539 }
4540
4541 static ssize_t rbd_size_show(struct device *dev,
4542                              struct device_attribute *attr, char *buf)
4543 {
4544         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4545
4546         return sprintf(buf, "%llu\n",
4547                 (unsigned long long)rbd_dev->mapping.size);
4548 }
4549
4550 /*
4551  * Note this shows the features for whatever's mapped, which is not
4552  * necessarily the base image.
4553  */
4554 static ssize_t rbd_features_show(struct device *dev,
4555                              struct device_attribute *attr, char *buf)
4556 {
4557         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4558
4559         return sprintf(buf, "0x%016llx\n",
4560                         (unsigned long long)rbd_dev->mapping.features);
4561 }
4562
4563 static ssize_t rbd_major_show(struct device *dev,
4564                               struct device_attribute *attr, char *buf)
4565 {
4566         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4567
4568         if (rbd_dev->major)
4569                 return sprintf(buf, "%d\n", rbd_dev->major);
4570
4571         return sprintf(buf, "(none)\n");
4572 }
4573
4574 static ssize_t rbd_minor_show(struct device *dev,
4575                               struct device_attribute *attr, char *buf)
4576 {
4577         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4578
4579         return sprintf(buf, "%d\n", rbd_dev->minor);
4580 }
4581
4582 static ssize_t rbd_client_addr_show(struct device *dev,
4583                                     struct device_attribute *attr, char *buf)
4584 {
4585         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4586         struct ceph_entity_addr *client_addr =
4587             ceph_client_addr(rbd_dev->rbd_client->client);
4588
4589         return sprintf(buf, "%pISpc/%u\n", &client_addr->in_addr,
4590                        le32_to_cpu(client_addr->nonce));
4591 }
4592
4593 static ssize_t rbd_client_id_show(struct device *dev,
4594                                   struct device_attribute *attr, char *buf)
4595 {
4596         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4597
4598         return sprintf(buf, "client%lld\n",
4599                        ceph_client_gid(rbd_dev->rbd_client->client));
4600 }
4601
4602 static ssize_t rbd_cluster_fsid_show(struct device *dev,
4603                                      struct device_attribute *attr, char *buf)
4604 {
4605         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4606
4607         return sprintf(buf, "%pU\n", &rbd_dev->rbd_client->client->fsid);
4608 }
4609
4610 static ssize_t rbd_config_info_show(struct device *dev,
4611                                     struct device_attribute *attr, char *buf)
4612 {
4613         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4614
4615         return sprintf(buf, "%s\n", rbd_dev->config_info);
4616 }
4617
4618 static ssize_t rbd_pool_show(struct device *dev,
4619                              struct device_attribute *attr, char *buf)
4620 {
4621         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4622
4623         return sprintf(buf, "%s\n", rbd_dev->spec->pool_name);
4624 }
4625
4626 static ssize_t rbd_pool_id_show(struct device *dev,
4627                              struct device_attribute *attr, char *buf)
4628 {
4629         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4630
4631         return sprintf(buf, "%llu\n",
4632                         (unsigned long long) rbd_dev->spec->pool_id);
4633 }
4634
4635 static ssize_t rbd_name_show(struct device *dev,
4636                              struct device_attribute *attr, char *buf)
4637 {
4638         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4639
4640         if (rbd_dev->spec->image_name)
4641                 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
4642
4643         return sprintf(buf, "(unknown)\n");
4644 }
4645
4646 static ssize_t rbd_image_id_show(struct device *dev,
4647                              struct device_attribute *attr, char *buf)
4648 {
4649         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4650
4651         return sprintf(buf, "%s\n", rbd_dev->spec->image_id);
4652 }
4653
4654 /*
4655  * Shows the name of the currently-mapped snapshot (or
4656  * RBD_SNAP_HEAD_NAME for the base image).
4657  */
4658 static ssize_t rbd_snap_show(struct device *dev,
4659                              struct device_attribute *attr,
4660                              char *buf)
4661 {
4662         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4663
4664         return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
4665 }
4666
4667 static ssize_t rbd_snap_id_show(struct device *dev,
4668                                 struct device_attribute *attr, char *buf)
4669 {
4670         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4671
4672         return sprintf(buf, "%llu\n", rbd_dev->spec->snap_id);
4673 }
4674
4675 /*
4676  * For a v2 image, shows the chain of parent images, separated by empty
4677  * lines.  For v1 images or if there is no parent, shows "(no parent
4678  * image)".
4679  */
4680 static ssize_t rbd_parent_show(struct device *dev,
4681                                struct device_attribute *attr,
4682                                char *buf)
4683 {
4684         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4685         ssize_t count = 0;
4686
4687         if (!rbd_dev->parent)
4688                 return sprintf(buf, "(no parent image)\n");
4689
4690         for ( ; rbd_dev->parent; rbd_dev = rbd_dev->parent) {
4691                 struct rbd_spec *spec = rbd_dev->parent_spec;
4692
4693                 count += sprintf(&buf[count], "%s"
4694                             "pool_id %llu\npool_name %s\n"
4695                             "image_id %s\nimage_name %s\n"
4696                             "snap_id %llu\nsnap_name %s\n"
4697                             "overlap %llu\n",
4698                             !count ? "" : "\n", /* first? */
4699                             spec->pool_id, spec->pool_name,
4700                             spec->image_id, spec->image_name ?: "(unknown)",
4701                             spec->snap_id, spec->snap_name,
4702                             rbd_dev->parent_overlap);
4703         }
4704
4705         return count;
4706 }
4707
4708 static ssize_t rbd_image_refresh(struct device *dev,
4709                                  struct device_attribute *attr,
4710                                  const char *buf,
4711                                  size_t size)
4712 {
4713         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4714         int ret;
4715
4716         ret = rbd_dev_refresh(rbd_dev);
4717         if (ret)
4718                 return ret;
4719
4720         return size;
4721 }
4722
4723 static DEVICE_ATTR(size, S_IRUGO, rbd_size_show, NULL);
4724 static DEVICE_ATTR(features, S_IRUGO, rbd_features_show, NULL);
4725 static DEVICE_ATTR(major, S_IRUGO, rbd_major_show, NULL);
4726 static DEVICE_ATTR(minor, S_IRUGO, rbd_minor_show, NULL);
4727 static DEVICE_ATTR(client_addr, S_IRUGO, rbd_client_addr_show, NULL);
4728 static DEVICE_ATTR(client_id, S_IRUGO, rbd_client_id_show, NULL);
4729 static DEVICE_ATTR(cluster_fsid, S_IRUGO, rbd_cluster_fsid_show, NULL);
4730 static DEVICE_ATTR(config_info, S_IRUSR, rbd_config_info_show, NULL);
4731 static DEVICE_ATTR(pool, S_IRUGO, rbd_pool_show, NULL);
4732 static DEVICE_ATTR(pool_id, S_IRUGO, rbd_pool_id_show, NULL);
4733 static DEVICE_ATTR(name, S_IRUGO, rbd_name_show, NULL);
4734 static DEVICE_ATTR(image_id, S_IRUGO, rbd_image_id_show, NULL);
4735 static DEVICE_ATTR(refresh, S_IWUSR, NULL, rbd_image_refresh);
4736 static DEVICE_ATTR(current_snap, S_IRUGO, rbd_snap_show, NULL);
4737 static DEVICE_ATTR(snap_id, S_IRUGO, rbd_snap_id_show, NULL);
4738 static DEVICE_ATTR(parent, S_IRUGO, rbd_parent_show, NULL);
4739
4740 static struct attribute *rbd_attrs[] = {
4741         &dev_attr_size.attr,
4742         &dev_attr_features.attr,
4743         &dev_attr_major.attr,
4744         &dev_attr_minor.attr,
4745         &dev_attr_client_addr.attr,
4746         &dev_attr_client_id.attr,
4747         &dev_attr_cluster_fsid.attr,
4748         &dev_attr_config_info.attr,
4749         &dev_attr_pool.attr,
4750         &dev_attr_pool_id.attr,
4751         &dev_attr_name.attr,
4752         &dev_attr_image_id.attr,
4753         &dev_attr_current_snap.attr,
4754         &dev_attr_snap_id.attr,
4755         &dev_attr_parent.attr,
4756         &dev_attr_refresh.attr,
4757         NULL
4758 };
4759
4760 static struct attribute_group rbd_attr_group = {
4761         .attrs = rbd_attrs,
4762 };
4763
4764 static const struct attribute_group *rbd_attr_groups[] = {
4765         &rbd_attr_group,
4766         NULL
4767 };
4768
4769 static void rbd_dev_release(struct device *dev);
4770
4771 static struct device_type rbd_device_type = {
4772         .name           = "rbd",
4773         .groups         = rbd_attr_groups,
4774         .release        = rbd_dev_release,
4775 };
4776
4777 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
4778 {
4779         kref_get(&spec->kref);
4780
4781         return spec;
4782 }
4783
4784 static void rbd_spec_free(struct kref *kref);
4785 static void rbd_spec_put(struct rbd_spec *spec)
4786 {
4787         if (spec)
4788                 kref_put(&spec->kref, rbd_spec_free);
4789 }
4790
4791 static struct rbd_spec *rbd_spec_alloc(void)
4792 {
4793         struct rbd_spec *spec;
4794
4795         spec = kzalloc(sizeof (*spec), GFP_KERNEL);
4796         if (!spec)
4797                 return NULL;
4798
4799         spec->pool_id = CEPH_NOPOOL;
4800         spec->snap_id = CEPH_NOSNAP;
4801         kref_init(&spec->kref);
4802
4803         return spec;
4804 }
4805
4806 static void rbd_spec_free(struct kref *kref)
4807 {
4808         struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
4809
4810         kfree(spec->pool_name);
4811         kfree(spec->image_id);
4812         kfree(spec->image_name);
4813         kfree(spec->snap_name);
4814         kfree(spec);
4815 }
4816
4817 static void rbd_dev_free(struct rbd_device *rbd_dev)
4818 {
4819         WARN_ON(rbd_dev->watch_state != RBD_WATCH_STATE_UNREGISTERED);
4820         WARN_ON(rbd_dev->lock_state != RBD_LOCK_STATE_UNLOCKED);
4821
4822         ceph_oid_destroy(&rbd_dev->header_oid);
4823         ceph_oloc_destroy(&rbd_dev->header_oloc);
4824         kfree(rbd_dev->config_info);
4825
4826         rbd_put_client(rbd_dev->rbd_client);
4827         rbd_spec_put(rbd_dev->spec);
4828         kfree(rbd_dev->opts);
4829         kfree(rbd_dev);
4830 }
4831
4832 static void rbd_dev_release(struct device *dev)
4833 {
4834         struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4835         bool need_put = !!rbd_dev->opts;
4836
4837         if (need_put) {
4838                 destroy_workqueue(rbd_dev->task_wq);
4839                 ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4840         }
4841
4842         rbd_dev_free(rbd_dev);
4843
4844         /*
4845          * This is racy, but way better than putting module outside of
4846          * the release callback.  The race window is pretty small, so
4847          * doing something similar to dm (dm-builtin.c) is overkill.
4848          */
4849         if (need_put)
4850                 module_put(THIS_MODULE);
4851 }
4852
4853 static struct rbd_device *__rbd_dev_create(struct rbd_client *rbdc,
4854                                            struct rbd_spec *spec)
4855 {
4856         struct rbd_device *rbd_dev;
4857
4858         rbd_dev = kzalloc(sizeof(*rbd_dev), GFP_KERNEL);
4859         if (!rbd_dev)
4860                 return NULL;
4861
4862         spin_lock_init(&rbd_dev->lock);
4863         INIT_LIST_HEAD(&rbd_dev->node);
4864         init_rwsem(&rbd_dev->header_rwsem);
4865
4866         ceph_oid_init(&rbd_dev->header_oid);
4867         ceph_oloc_init(&rbd_dev->header_oloc);
4868
4869         mutex_init(&rbd_dev->watch_mutex);
4870         rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
4871         INIT_DELAYED_WORK(&rbd_dev->watch_dwork, rbd_reregister_watch);
4872
4873         init_rwsem(&rbd_dev->lock_rwsem);
4874         rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
4875         INIT_WORK(&rbd_dev->acquired_lock_work, rbd_notify_acquired_lock);
4876         INIT_WORK(&rbd_dev->released_lock_work, rbd_notify_released_lock);
4877         INIT_DELAYED_WORK(&rbd_dev->lock_dwork, rbd_acquire_lock);
4878         INIT_WORK(&rbd_dev->unlock_work, rbd_release_lock_work);
4879         init_waitqueue_head(&rbd_dev->lock_waitq);
4880
4881         rbd_dev->dev.bus = &rbd_bus_type;
4882         rbd_dev->dev.type = &rbd_device_type;
4883         rbd_dev->dev.parent = &rbd_root_dev;
4884         device_initialize(&rbd_dev->dev);
4885
4886         rbd_dev->rbd_client = rbdc;
4887         rbd_dev->spec = spec;
4888
4889         rbd_dev->layout.stripe_unit = 1 << RBD_MAX_OBJ_ORDER;
4890         rbd_dev->layout.stripe_count = 1;
4891         rbd_dev->layout.object_size = 1 << RBD_MAX_OBJ_ORDER;
4892         rbd_dev->layout.pool_id = spec->pool_id;
4893         RCU_INIT_POINTER(rbd_dev->layout.pool_ns, NULL);
4894
4895         return rbd_dev;
4896 }
4897
4898 /*
4899  * Create a mapping rbd_dev.
4900  */
4901 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
4902                                          struct rbd_spec *spec,
4903                                          struct rbd_options *opts)
4904 {
4905         struct rbd_device *rbd_dev;
4906
4907         rbd_dev = __rbd_dev_create(rbdc, spec);
4908         if (!rbd_dev)
4909                 return NULL;
4910
4911         rbd_dev->opts = opts;
4912
4913         /* get an id and fill in device name */
4914         rbd_dev->dev_id = ida_simple_get(&rbd_dev_id_ida, 0,
4915                                          minor_to_rbd_dev_id(1 << MINORBITS),
4916                                          GFP_KERNEL);
4917         if (rbd_dev->dev_id < 0)
4918                 goto fail_rbd_dev;
4919
4920         sprintf(rbd_dev->name, RBD_DRV_NAME "%d", rbd_dev->dev_id);
4921         rbd_dev->task_wq = alloc_ordered_workqueue("%s-tasks", WQ_MEM_RECLAIM,
4922                                                    rbd_dev->name);
4923         if (!rbd_dev->task_wq)
4924                 goto fail_dev_id;
4925
4926         /* we have a ref from do_rbd_add() */
4927         __module_get(THIS_MODULE);
4928
4929         dout("%s rbd_dev %p dev_id %d\n", __func__, rbd_dev, rbd_dev->dev_id);
4930         return rbd_dev;
4931
4932 fail_dev_id:
4933         ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4934 fail_rbd_dev:
4935         rbd_dev_free(rbd_dev);
4936         return NULL;
4937 }
4938
4939 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
4940 {
4941         if (rbd_dev)
4942                 put_device(&rbd_dev->dev);
4943 }
4944
4945 /*
4946  * Get the size and object order for an image snapshot, or if
4947  * snap_id is CEPH_NOSNAP, gets this information for the base
4948  * image.
4949  */
4950 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
4951                                 u8 *order, u64 *snap_size)
4952 {
4953         __le64 snapid = cpu_to_le64(snap_id);
4954         int ret;
4955         struct {
4956                 u8 order;
4957                 __le64 size;
4958         } __attribute__ ((packed)) size_buf = { 0 };
4959
4960         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
4961                                 "rbd", "get_size",
4962                                 &snapid, sizeof (snapid),
4963                                 &size_buf, sizeof (size_buf));
4964         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4965         if (ret < 0)
4966                 return ret;
4967         if (ret < sizeof (size_buf))
4968                 return -ERANGE;
4969
4970         if (order) {
4971                 *order = size_buf.order;
4972                 dout("  order %u", (unsigned int)*order);
4973         }
4974         *snap_size = le64_to_cpu(size_buf.size);
4975
4976         dout("  snap_id 0x%016llx snap_size = %llu\n",
4977                 (unsigned long long)snap_id,
4978                 (unsigned long long)*snap_size);
4979
4980         return 0;
4981 }
4982
4983 static int rbd_dev_v2_image_size(struct rbd_device *rbd_dev)
4984 {
4985         return _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
4986                                         &rbd_dev->header.obj_order,
4987                                         &rbd_dev->header.image_size);
4988 }
4989
4990 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev)
4991 {
4992         void *reply_buf;
4993         int ret;
4994         void *p;
4995
4996         reply_buf = kzalloc(RBD_OBJ_PREFIX_LEN_MAX, GFP_KERNEL);
4997         if (!reply_buf)
4998                 return -ENOMEM;
4999
5000         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5001                                 "rbd", "get_object_prefix", NULL, 0,
5002                                 reply_buf, RBD_OBJ_PREFIX_LEN_MAX);
5003         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5004         if (ret < 0)
5005                 goto out;
5006
5007         p = reply_buf;
5008         rbd_dev->header.object_prefix = ceph_extract_encoded_string(&p,
5009                                                 p + ret, NULL, GFP_NOIO);
5010         ret = 0;
5011
5012         if (IS_ERR(rbd_dev->header.object_prefix)) {
5013                 ret = PTR_ERR(rbd_dev->header.object_prefix);
5014                 rbd_dev->header.object_prefix = NULL;
5015         } else {
5016                 dout("  object_prefix = %s\n", rbd_dev->header.object_prefix);
5017         }
5018 out:
5019         kfree(reply_buf);
5020
5021         return ret;
5022 }
5023
5024 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
5025                 u64 *snap_features)
5026 {
5027         __le64 snapid = cpu_to_le64(snap_id);
5028         struct {
5029                 __le64 features;
5030                 __le64 incompat;
5031         } __attribute__ ((packed)) features_buf = { 0 };
5032         u64 unsup;
5033         int ret;
5034
5035         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5036                                 "rbd", "get_features",
5037                                 &snapid, sizeof (snapid),
5038                                 &features_buf, sizeof (features_buf));
5039         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5040         if (ret < 0)
5041                 return ret;
5042         if (ret < sizeof (features_buf))
5043                 return -ERANGE;
5044
5045         unsup = le64_to_cpu(features_buf.incompat) & ~RBD_FEATURES_SUPPORTED;
5046         if (unsup) {
5047                 rbd_warn(rbd_dev, "image uses unsupported features: 0x%llx",
5048                          unsup);
5049                 return -ENXIO;
5050         }
5051
5052         *snap_features = le64_to_cpu(features_buf.features);
5053
5054         dout("  snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
5055                 (unsigned long long)snap_id,
5056                 (unsigned long long)*snap_features,
5057                 (unsigned long long)le64_to_cpu(features_buf.incompat));
5058
5059         return 0;
5060 }
5061
5062 static int rbd_dev_v2_features(struct rbd_device *rbd_dev)
5063 {
5064         return _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
5065                                                 &rbd_dev->header.features);
5066 }
5067
5068 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev)
5069 {
5070         struct rbd_spec *parent_spec;
5071         size_t size;
5072         void *reply_buf = NULL;
5073         __le64 snapid;
5074         void *p;
5075         void *end;
5076         u64 pool_id;
5077         char *image_id;
5078         u64 snap_id;
5079         u64 overlap;
5080         int ret;
5081
5082         parent_spec = rbd_spec_alloc();
5083         if (!parent_spec)
5084                 return -ENOMEM;
5085
5086         size = sizeof (__le64) +                                /* pool_id */
5087                 sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX +        /* image_id */
5088                 sizeof (__le64) +                               /* snap_id */
5089                 sizeof (__le64);                                /* overlap */
5090         reply_buf = kmalloc(size, GFP_KERNEL);
5091         if (!reply_buf) {
5092                 ret = -ENOMEM;
5093                 goto out_err;
5094         }
5095
5096         snapid = cpu_to_le64(rbd_dev->spec->snap_id);
5097         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5098                                 "rbd", "get_parent",
5099                                 &snapid, sizeof (snapid),
5100                                 reply_buf, size);
5101         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5102         if (ret < 0)
5103                 goto out_err;
5104
5105         p = reply_buf;
5106         end = reply_buf + ret;
5107         ret = -ERANGE;
5108         ceph_decode_64_safe(&p, end, pool_id, out_err);
5109         if (pool_id == CEPH_NOPOOL) {
5110                 /*
5111                  * Either the parent never existed, or we have
5112                  * record of it but the image got flattened so it no
5113                  * longer has a parent.  When the parent of a
5114                  * layered image disappears we immediately set the
5115                  * overlap to 0.  The effect of this is that all new
5116                  * requests will be treated as if the image had no
5117                  * parent.
5118                  */
5119                 if (rbd_dev->parent_overlap) {
5120                         rbd_dev->parent_overlap = 0;
5121                         rbd_dev_parent_put(rbd_dev);
5122                         pr_info("%s: clone image has been flattened\n",
5123                                 rbd_dev->disk->disk_name);
5124                 }
5125
5126                 goto out;       /* No parent?  No problem. */
5127         }
5128
5129         /* The ceph file layout needs to fit pool id in 32 bits */
5130
5131         ret = -EIO;
5132         if (pool_id > (u64)U32_MAX) {
5133                 rbd_warn(NULL, "parent pool id too large (%llu > %u)",
5134                         (unsigned long long)pool_id, U32_MAX);
5135                 goto out_err;
5136         }
5137
5138         image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
5139         if (IS_ERR(image_id)) {
5140                 ret = PTR_ERR(image_id);
5141                 goto out_err;
5142         }
5143         ceph_decode_64_safe(&p, end, snap_id, out_err);
5144         ceph_decode_64_safe(&p, end, overlap, out_err);
5145
5146         /*
5147          * The parent won't change (except when the clone is
5148          * flattened, already handled that).  So we only need to
5149          * record the parent spec we have not already done so.
5150          */
5151         if (!rbd_dev->parent_spec) {
5152                 parent_spec->pool_id = pool_id;
5153                 parent_spec->image_id = image_id;
5154                 parent_spec->snap_id = snap_id;
5155                 rbd_dev->parent_spec = parent_spec;
5156                 parent_spec = NULL;     /* rbd_dev now owns this */
5157         } else {
5158                 kfree(image_id);
5159         }
5160
5161         /*
5162          * We always update the parent overlap.  If it's zero we issue
5163          * a warning, as we will proceed as if there was no parent.
5164          */
5165         if (!overlap) {
5166                 if (parent_spec) {
5167                         /* refresh, careful to warn just once */
5168                         if (rbd_dev->parent_overlap)
5169                                 rbd_warn(rbd_dev,
5170                                     "clone now standalone (overlap became 0)");
5171                 } else {
5172                         /* initial probe */
5173                         rbd_warn(rbd_dev, "clone is standalone (overlap 0)");
5174                 }
5175         }
5176         rbd_dev->parent_overlap = overlap;
5177
5178 out:
5179         ret = 0;
5180 out_err:
5181         kfree(reply_buf);
5182         rbd_spec_put(parent_spec);
5183
5184         return ret;
5185 }
5186
5187 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev)
5188 {
5189         struct {
5190                 __le64 stripe_unit;
5191                 __le64 stripe_count;
5192         } __attribute__ ((packed)) striping_info_buf = { 0 };
5193         size_t size = sizeof (striping_info_buf);
5194         void *p;
5195         u64 obj_size;
5196         u64 stripe_unit;
5197         u64 stripe_count;
5198         int ret;
5199
5200         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5201                                 "rbd", "get_stripe_unit_count", NULL, 0,
5202                                 (char *)&striping_info_buf, size);
5203         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5204         if (ret < 0)
5205                 return ret;
5206         if (ret < size)
5207                 return -ERANGE;
5208
5209         /*
5210          * We don't actually support the "fancy striping" feature
5211          * (STRIPINGV2) yet, but if the striping sizes are the
5212          * defaults the behavior is the same as before.  So find
5213          * out, and only fail if the image has non-default values.
5214          */
5215         ret = -EINVAL;
5216         obj_size = (u64)1 << rbd_dev->header.obj_order;
5217         p = &striping_info_buf;
5218         stripe_unit = ceph_decode_64(&p);
5219         if (stripe_unit != obj_size) {
5220                 rbd_warn(rbd_dev, "unsupported stripe unit "
5221                                 "(got %llu want %llu)",
5222                                 stripe_unit, obj_size);
5223                 return -EINVAL;
5224         }
5225         stripe_count = ceph_decode_64(&p);
5226         if (stripe_count != 1) {
5227                 rbd_warn(rbd_dev, "unsupported stripe count "
5228                                 "(got %llu want 1)", stripe_count);
5229                 return -EINVAL;
5230         }
5231         rbd_dev->header.stripe_unit = stripe_unit;
5232         rbd_dev->header.stripe_count = stripe_count;
5233
5234         return 0;
5235 }
5236
5237 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
5238 {
5239         size_t image_id_size;
5240         char *image_id;
5241         void *p;
5242         void *end;
5243         size_t size;
5244         void *reply_buf = NULL;
5245         size_t len = 0;
5246         char *image_name = NULL;
5247         int ret;
5248
5249         rbd_assert(!rbd_dev->spec->image_name);
5250
5251         len = strlen(rbd_dev->spec->image_id);
5252         image_id_size = sizeof (__le32) + len;
5253         image_id = kmalloc(image_id_size, GFP_KERNEL);
5254         if (!image_id)
5255                 return NULL;
5256
5257         p = image_id;
5258         end = image_id + image_id_size;
5259         ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
5260
5261         size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
5262         reply_buf = kmalloc(size, GFP_KERNEL);
5263         if (!reply_buf)
5264                 goto out;
5265
5266         ret = rbd_obj_method_sync(rbd_dev, RBD_DIRECTORY,
5267                                 "rbd", "dir_get_name",
5268                                 image_id, image_id_size,
5269                                 reply_buf, size);
5270         if (ret < 0)
5271                 goto out;
5272         p = reply_buf;
5273         end = reply_buf + ret;
5274
5275         image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
5276         if (IS_ERR(image_name))
5277                 image_name = NULL;
5278         else
5279                 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
5280 out:
5281         kfree(reply_buf);
5282         kfree(image_id);
5283
5284         return image_name;
5285 }
5286
5287 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5288 {
5289         struct ceph_snap_context *snapc = rbd_dev->header.snapc;
5290         const char *snap_name;
5291         u32 which = 0;
5292
5293         /* Skip over names until we find the one we are looking for */
5294
5295         snap_name = rbd_dev->header.snap_names;
5296         while (which < snapc->num_snaps) {
5297                 if (!strcmp(name, snap_name))
5298                         return snapc->snaps[which];
5299                 snap_name += strlen(snap_name) + 1;
5300                 which++;
5301         }
5302         return CEPH_NOSNAP;
5303 }
5304
5305 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5306 {
5307         struct ceph_snap_context *snapc = rbd_dev->header.snapc;
5308         u32 which;
5309         bool found = false;
5310         u64 snap_id;
5311
5312         for (which = 0; !found && which < snapc->num_snaps; which++) {
5313                 const char *snap_name;
5314
5315                 snap_id = snapc->snaps[which];
5316                 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
5317                 if (IS_ERR(snap_name)) {
5318                         /* ignore no-longer existing snapshots */
5319                         if (PTR_ERR(snap_name) == -ENOENT)
5320                                 continue;
5321                         else
5322                                 break;
5323                 }
5324                 found = !strcmp(name, snap_name);
5325                 kfree(snap_name);
5326         }
5327         return found ? snap_id : CEPH_NOSNAP;
5328 }
5329
5330 /*
5331  * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
5332  * no snapshot by that name is found, or if an error occurs.
5333  */
5334 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5335 {
5336         if (rbd_dev->image_format == 1)
5337                 return rbd_v1_snap_id_by_name(rbd_dev, name);
5338
5339         return rbd_v2_snap_id_by_name(rbd_dev, name);
5340 }
5341
5342 /*
5343  * An image being mapped will have everything but the snap id.
5344  */
5345 static int rbd_spec_fill_snap_id(struct rbd_device *rbd_dev)
5346 {
5347         struct rbd_spec *spec = rbd_dev->spec;
5348
5349         rbd_assert(spec->pool_id != CEPH_NOPOOL && spec->pool_name);
5350         rbd_assert(spec->image_id && spec->image_name);
5351         rbd_assert(spec->snap_name);
5352
5353         if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
5354                 u64 snap_id;
5355
5356                 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
5357                 if (snap_id == CEPH_NOSNAP)
5358                         return -ENOENT;
5359
5360                 spec->snap_id = snap_id;
5361         } else {
5362                 spec->snap_id = CEPH_NOSNAP;
5363         }
5364
5365         return 0;
5366 }
5367
5368 /*
5369  * A parent image will have all ids but none of the names.
5370  *
5371  * All names in an rbd spec are dynamically allocated.  It's OK if we
5372  * can't figure out the name for an image id.
5373  */
5374 static int rbd_spec_fill_names(struct rbd_device *rbd_dev)
5375 {
5376         struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
5377         struct rbd_spec *spec = rbd_dev->spec;
5378         const char *pool_name;
5379         const char *image_name;
5380         const char *snap_name;
5381         int ret;
5382
5383         rbd_assert(spec->pool_id != CEPH_NOPOOL);
5384         rbd_assert(spec->image_id);
5385         rbd_assert(spec->snap_id != CEPH_NOSNAP);
5386
5387         /* Get the pool name; we have to make our own copy of this */
5388
5389         pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
5390         if (!pool_name) {
5391                 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
5392                 return -EIO;
5393         }
5394         pool_name = kstrdup(pool_name, GFP_KERNEL);
5395         if (!pool_name)
5396                 return -ENOMEM;
5397
5398         /* Fetch the image name; tolerate failure here */
5399
5400         image_name = rbd_dev_image_name(rbd_dev);
5401         if (!image_name)
5402                 rbd_warn(rbd_dev, "unable to get image name");
5403
5404         /* Fetch the snapshot name */
5405
5406         snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
5407         if (IS_ERR(snap_name)) {
5408                 ret = PTR_ERR(snap_name);
5409                 goto out_err;
5410         }
5411
5412         spec->pool_name = pool_name;
5413         spec->image_name = image_name;
5414         spec->snap_name = snap_name;
5415
5416         return 0;
5417
5418 out_err:
5419         kfree(image_name);
5420         kfree(pool_name);
5421         return ret;
5422 }
5423
5424 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev)
5425 {
5426         size_t size;
5427         int ret;
5428         void *reply_buf;
5429         void *p;
5430         void *end;
5431         u64 seq;
5432         u32 snap_count;
5433         struct ceph_snap_context *snapc;
5434         u32 i;
5435
5436         /*
5437          * We'll need room for the seq value (maximum snapshot id),
5438          * snapshot count, and array of that many snapshot ids.
5439          * For now we have a fixed upper limit on the number we're
5440          * prepared to receive.
5441          */
5442         size = sizeof (__le64) + sizeof (__le32) +
5443                         RBD_MAX_SNAP_COUNT * sizeof (__le64);
5444         reply_buf = kzalloc(size, GFP_KERNEL);
5445         if (!reply_buf)
5446                 return -ENOMEM;
5447
5448         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5449                                 "rbd", "get_snapcontext", NULL, 0,
5450                                 reply_buf, size);
5451         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5452         if (ret < 0)
5453                 goto out;
5454
5455         p = reply_buf;
5456         end = reply_buf + ret;
5457         ret = -ERANGE;
5458         ceph_decode_64_safe(&p, end, seq, out);
5459         ceph_decode_32_safe(&p, end, snap_count, out);
5460
5461         /*
5462          * Make sure the reported number of snapshot ids wouldn't go
5463          * beyond the end of our buffer.  But before checking that,
5464          * make sure the computed size of the snapshot context we
5465          * allocate is representable in a size_t.
5466          */
5467         if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
5468                                  / sizeof (u64)) {
5469                 ret = -EINVAL;
5470                 goto out;
5471         }
5472         if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
5473                 goto out;
5474         ret = 0;
5475
5476         snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
5477         if (!snapc) {
5478                 ret = -ENOMEM;
5479                 goto out;
5480         }
5481         snapc->seq = seq;
5482         for (i = 0; i < snap_count; i++)
5483                 snapc->snaps[i] = ceph_decode_64(&p);
5484
5485         ceph_put_snap_context(rbd_dev->header.snapc);
5486         rbd_dev->header.snapc = snapc;
5487
5488         dout("  snap context seq = %llu, snap_count = %u\n",
5489                 (unsigned long long)seq, (unsigned int)snap_count);
5490 out:
5491         kfree(reply_buf);
5492
5493         return ret;
5494 }
5495
5496 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
5497                                         u64 snap_id)
5498 {
5499         size_t size;
5500         void *reply_buf;
5501         __le64 snapid;
5502         int ret;
5503         void *p;
5504         void *end;
5505         char *snap_name;
5506
5507         size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
5508         reply_buf = kmalloc(size, GFP_KERNEL);
5509         if (!reply_buf)
5510                 return ERR_PTR(-ENOMEM);
5511
5512         snapid = cpu_to_le64(snap_id);
5513         ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5514                                 "rbd", "get_snapshot_name",
5515                                 &snapid, sizeof (snapid),
5516                                 reply_buf, size);
5517         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5518         if (ret < 0) {
5519                 snap_name = ERR_PTR(ret);
5520                 goto out;
5521         }
5522
5523         p = reply_buf;
5524         end = reply_buf + ret;
5525         snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
5526         if (IS_ERR(snap_name))
5527                 goto out;
5528
5529         dout("  snap_id 0x%016llx snap_name = %s\n",
5530                 (unsigned long long)snap_id, snap_name);
5531 out:
5532         kfree(reply_buf);
5533
5534         return snap_name;
5535 }
5536
5537 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev)
5538 {
5539         bool first_time = rbd_dev->header.object_prefix == NULL;
5540         int ret;
5541
5542         ret = rbd_dev_v2_image_size(rbd_dev);
5543         if (ret)
5544                 return ret;
5545
5546         if (first_time) {
5547                 ret = rbd_dev_v2_header_onetime(rbd_dev);
5548                 if (ret)
5549                         return ret;
5550         }
5551
5552         ret = rbd_dev_v2_snap_context(rbd_dev);
5553         if (ret && first_time) {
5554                 kfree(rbd_dev->header.object_prefix);
5555                 rbd_dev->header.object_prefix = NULL;
5556         }
5557
5558         return ret;
5559 }
5560
5561 static int rbd_dev_header_info(struct rbd_device *rbd_dev)
5562 {
5563         rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
5564
5565         if (rbd_dev->image_format == 1)
5566                 return rbd_dev_v1_header_info(rbd_dev);
5567
5568         return rbd_dev_v2_header_info(rbd_dev);
5569 }
5570
5571 /*
5572  * Skips over white space at *buf, and updates *buf to point to the
5573  * first found non-space character (if any). Returns the length of
5574  * the token (string of non-white space characters) found.  Note
5575  * that *buf must be terminated with '\0'.
5576  */
5577 static inline size_t next_token(const char **buf)
5578 {
5579         /*
5580         * These are the characters that produce nonzero for
5581         * isspace() in the "C" and "POSIX" locales.
5582         */
5583         const char *spaces = " \f\n\r\t\v";
5584
5585         *buf += strspn(*buf, spaces);   /* Find start of token */
5586
5587         return strcspn(*buf, spaces);   /* Return token length */
5588 }
5589
5590 /*
5591  * Finds the next token in *buf, dynamically allocates a buffer big
5592  * enough to hold a copy of it, and copies the token into the new
5593  * buffer.  The copy is guaranteed to be terminated with '\0'.  Note
5594  * that a duplicate buffer is created even for a zero-length token.
5595  *
5596  * Returns a pointer to the newly-allocated duplicate, or a null
5597  * pointer if memory for the duplicate was not available.  If
5598  * the lenp argument is a non-null pointer, the length of the token
5599  * (not including the '\0') is returned in *lenp.
5600  *
5601  * If successful, the *buf pointer will be updated to point beyond
5602  * the end of the found token.
5603  *
5604  * Note: uses GFP_KERNEL for allocation.
5605  */
5606 static inline char *dup_token(const char **buf, size_t *lenp)
5607 {
5608         char *dup;
5609         size_t len;
5610
5611         len = next_token(buf);
5612         dup = kmemdup(*buf, len + 1, GFP_KERNEL);
5613         if (!dup)
5614                 return NULL;
5615         *(dup + len) = '\0';
5616         *buf += len;
5617
5618         if (lenp)
5619                 *lenp = len;
5620
5621         return dup;
5622 }
5623
5624 /*
5625  * Parse the options provided for an "rbd add" (i.e., rbd image
5626  * mapping) request.  These arrive via a write to /sys/bus/rbd/add,
5627  * and the data written is passed here via a NUL-terminated buffer.
5628  * Returns 0 if successful or an error code otherwise.
5629  *
5630  * The information extracted from these options is recorded in
5631  * the other parameters which return dynamically-allocated
5632  * structures:
5633  *  ceph_opts
5634  *      The address of a pointer that will refer to a ceph options
5635  *      structure.  Caller must release the returned pointer using
5636  *      ceph_destroy_options() when it is no longer needed.
5637  *  rbd_opts
5638  *      Address of an rbd options pointer.  Fully initialized by
5639  *      this function; caller must release with kfree().
5640  *  spec
5641  *      Address of an rbd image specification pointer.  Fully
5642  *      initialized by this function based on parsed options.
5643  *      Caller must release with rbd_spec_put().
5644  *
5645  * The options passed take this form:
5646  *  <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
5647  * where:
5648  *  <mon_addrs>
5649  *      A comma-separated list of one or more monitor addresses.
5650  *      A monitor address is an ip address, optionally followed
5651  *      by a port number (separated by a colon).
5652  *        I.e.:  ip1[:port1][,ip2[:port2]...]
5653  *  <options>
5654  *      A comma-separated list of ceph and/or rbd options.
5655  *  <pool_name>
5656  *      The name of the rados pool containing the rbd image.
5657  *  <image_name>
5658  *      The name of the image in that pool to map.
5659  *  <snap_id>
5660  *      An optional snapshot id.  If provided, the mapping will
5661  *      present data from the image at the time that snapshot was
5662  *      created.  The image head is used if no snapshot id is
5663  *      provided.  Snapshot mappings are always read-only.
5664  */
5665 static int rbd_add_parse_args(const char *buf,
5666                                 struct ceph_options **ceph_opts,
5667                                 struct rbd_options **opts,
5668                                 struct rbd_spec **rbd_spec)
5669 {
5670         size_t len;
5671         char *options;
5672         const char *mon_addrs;
5673         char *snap_name;
5674         size_t mon_addrs_size;
5675         struct rbd_spec *spec = NULL;
5676         struct rbd_options *rbd_opts = NULL;
5677         struct ceph_options *copts;
5678         int ret;
5679
5680         /* The first four tokens are required */
5681
5682         len = next_token(&buf);
5683         if (!len) {
5684                 rbd_warn(NULL, "no monitor address(es) provided");
5685                 return -EINVAL;
5686         }
5687         mon_addrs = buf;
5688         mon_addrs_size = len + 1;
5689         buf += len;
5690
5691         ret = -EINVAL;
5692         options = dup_token(&buf, NULL);
5693         if (!options)
5694                 return -ENOMEM;
5695         if (!*options) {
5696                 rbd_warn(NULL, "no options provided");
5697                 goto out_err;
5698         }
5699
5700         spec = rbd_spec_alloc();
5701         if (!spec)
5702                 goto out_mem;
5703
5704         spec->pool_name = dup_token(&buf, NULL);
5705         if (!spec->pool_name)
5706                 goto out_mem;
5707         if (!*spec->pool_name) {
5708                 rbd_warn(NULL, "no pool name provided");
5709                 goto out_err;
5710         }
5711
5712         spec->image_name = dup_token(&buf, NULL);
5713         if (!spec->image_name)
5714                 goto out_mem;
5715         if (!*spec->image_name) {
5716                 rbd_warn(NULL, "no image name provided");
5717                 goto out_err;
5718         }
5719
5720         /*
5721          * Snapshot name is optional; default is to use "-"
5722          * (indicating the head/no snapshot).
5723          */
5724         len = next_token(&buf);
5725         if (!len) {
5726                 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
5727                 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
5728         } else if (len > RBD_MAX_SNAP_NAME_LEN) {
5729                 ret = -ENAMETOOLONG;
5730                 goto out_err;
5731         }
5732         snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
5733         if (!snap_name)
5734                 goto out_mem;
5735         *(snap_name + len) = '\0';
5736         spec->snap_name = snap_name;
5737
5738         /* Initialize all rbd options to the defaults */
5739
5740         rbd_opts = kzalloc(sizeof (*rbd_opts), GFP_KERNEL);
5741         if (!rbd_opts)
5742                 goto out_mem;
5743
5744         rbd_opts->read_only = RBD_READ_ONLY_DEFAULT;
5745         rbd_opts->queue_depth = RBD_QUEUE_DEPTH_DEFAULT;
5746         rbd_opts->lock_on_read = RBD_LOCK_ON_READ_DEFAULT;
5747
5748         copts = ceph_parse_options(options, mon_addrs,
5749                                         mon_addrs + mon_addrs_size - 1,
5750                                         parse_rbd_opts_token, rbd_opts);
5751         if (IS_ERR(copts)) {
5752                 ret = PTR_ERR(copts);
5753                 goto out_err;
5754         }
5755         kfree(options);
5756
5757         *ceph_opts = copts;
5758         *opts = rbd_opts;
5759         *rbd_spec = spec;
5760
5761         return 0;
5762 out_mem:
5763         ret = -ENOMEM;
5764 out_err:
5765         kfree(rbd_opts);
5766         rbd_spec_put(spec);
5767         kfree(options);
5768
5769         return ret;
5770 }
5771
5772 /*
5773  * Return pool id (>= 0) or a negative error code.
5774  */
5775 static int rbd_add_get_pool_id(struct rbd_client *rbdc, const char *pool_name)
5776 {
5777         struct ceph_options *opts = rbdc->client->options;
5778         u64 newest_epoch;
5779         int tries = 0;
5780         int ret;
5781
5782 again:
5783         ret = ceph_pg_poolid_by_name(rbdc->client->osdc.osdmap, pool_name);
5784         if (ret == -ENOENT && tries++ < 1) {
5785                 ret = ceph_monc_get_version(&rbdc->client->monc, "osdmap",
5786                                             &newest_epoch);
5787                 if (ret < 0)
5788                         return ret;
5789
5790                 if (rbdc->client->osdc.osdmap->epoch < newest_epoch) {
5791                         ceph_osdc_maybe_request_map(&rbdc->client->osdc);
5792                         (void) ceph_monc_wait_osdmap(&rbdc->client->monc,
5793                                                      newest_epoch,
5794                                                      opts->mount_timeout);
5795                         goto again;
5796                 } else {
5797                         /* the osdmap we have is new enough */
5798                         return -ENOENT;
5799                 }
5800         }
5801
5802         return ret;
5803 }
5804
5805 /*
5806  * An rbd format 2 image has a unique identifier, distinct from the
5807  * name given to it by the user.  Internally, that identifier is
5808  * what's used to specify the names of objects related to the image.
5809  *
5810  * A special "rbd id" object is used to map an rbd image name to its
5811  * id.  If that object doesn't exist, then there is no v2 rbd image
5812  * with the supplied name.
5813  *
5814  * This function will record the given rbd_dev's image_id field if
5815  * it can be determined, and in that case will return 0.  If any
5816  * errors occur a negative errno will be returned and the rbd_dev's
5817  * image_id field will be unchanged (and should be NULL).
5818  */
5819 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
5820 {
5821         int ret;
5822         size_t size;
5823         char *object_name;
5824         void *response;
5825         char *image_id;
5826
5827         /*
5828          * When probing a parent image, the image id is already
5829          * known (and the image name likely is not).  There's no
5830          * need to fetch the image id again in this case.  We
5831          * do still need to set the image format though.
5832          */
5833         if (rbd_dev->spec->image_id) {
5834                 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
5835
5836                 return 0;
5837         }
5838
5839         /*
5840          * First, see if the format 2 image id file exists, and if
5841          * so, get the image's persistent id from it.
5842          */
5843         size = sizeof (RBD_ID_PREFIX) + strlen(rbd_dev->spec->image_name);
5844         object_name = kmalloc(size, GFP_NOIO);
5845         if (!object_name)
5846                 return -ENOMEM;
5847         sprintf(object_name, "%s%s", RBD_ID_PREFIX, rbd_dev->spec->image_name);
5848         dout("rbd id object name is %s\n", object_name);
5849
5850         /* Response will be an encoded string, which includes a length */
5851
5852         size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
5853         response = kzalloc(size, GFP_NOIO);
5854         if (!response) {
5855                 ret = -ENOMEM;
5856                 goto out;
5857         }
5858
5859         /* If it doesn't exist we'll assume it's a format 1 image */
5860
5861         ret = rbd_obj_method_sync(rbd_dev, object_name,
5862                                 "rbd", "get_id", NULL, 0,
5863                                 response, RBD_IMAGE_ID_LEN_MAX);
5864         dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5865         if (ret == -ENOENT) {
5866                 image_id = kstrdup("", GFP_KERNEL);
5867                 ret = image_id ? 0 : -ENOMEM;
5868                 if (!ret)
5869                         rbd_dev->image_format = 1;
5870         } else if (ret >= 0) {
5871                 void *p = response;
5872
5873                 image_id = ceph_extract_encoded_string(&p, p + ret,
5874                                                 NULL, GFP_NOIO);
5875                 ret = PTR_ERR_OR_ZERO(image_id);
5876                 if (!ret)
5877                         rbd_dev->image_format = 2;
5878         }
5879
5880         if (!ret) {
5881                 rbd_dev->spec->image_id = image_id;
5882                 dout("image_id is %s\n", image_id);
5883         }
5884 out:
5885         kfree(response);
5886         kfree(object_name);
5887
5888         return ret;
5889 }
5890
5891 /*
5892  * Undo whatever state changes are made by v1 or v2 header info
5893  * call.
5894  */
5895 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
5896 {
5897         struct rbd_image_header *header;
5898
5899         rbd_dev_parent_put(rbd_dev);
5900
5901         /* Free dynamic fields from the header, then zero it out */
5902
5903         header = &rbd_dev->header;
5904         ceph_put_snap_context(header->snapc);
5905         kfree(header->snap_sizes);
5906         kfree(header->snap_names);
5907         kfree(header->object_prefix);
5908         memset(header, 0, sizeof (*header));
5909 }
5910
5911 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev)
5912 {
5913         int ret;
5914
5915         ret = rbd_dev_v2_object_prefix(rbd_dev);
5916         if (ret)
5917                 goto out_err;
5918
5919         /*
5920          * Get the and check features for the image.  Currently the
5921          * features are assumed to never change.
5922          */
5923         ret = rbd_dev_v2_features(rbd_dev);
5924         if (ret)
5925                 goto out_err;
5926
5927         /* If the image supports fancy striping, get its parameters */
5928
5929         if (rbd_dev->header.features & RBD_FEATURE_STRIPINGV2) {
5930                 ret = rbd_dev_v2_striping_info(rbd_dev);
5931                 if (ret < 0)
5932                         goto out_err;
5933         }
5934         /* No support for crypto and compression type format 2 images */
5935
5936         return 0;
5937 out_err:
5938         rbd_dev->header.features = 0;
5939         kfree(rbd_dev->header.object_prefix);
5940         rbd_dev->header.object_prefix = NULL;
5941
5942         return ret;
5943 }
5944
5945 /*
5946  * @depth is rbd_dev_image_probe() -> rbd_dev_probe_parent() ->
5947  * rbd_dev_image_probe() recursion depth, which means it's also the
5948  * length of the already discovered part of the parent chain.
5949  */
5950 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev, int depth)
5951 {
5952         struct rbd_device *parent = NULL;
5953         int ret;
5954
5955         if (!rbd_dev->parent_spec)
5956                 return 0;
5957
5958         if (++depth > RBD_MAX_PARENT_CHAIN_LEN) {
5959                 pr_info("parent chain is too long (%d)\n", depth);
5960                 ret = -EINVAL;
5961                 goto out_err;
5962         }
5963
5964         parent = __rbd_dev_create(rbd_dev->rbd_client, rbd_dev->parent_spec);
5965         if (!parent) {
5966                 ret = -ENOMEM;
5967                 goto out_err;
5968         }
5969
5970         /*
5971          * Images related by parent/child relationships always share
5972          * rbd_client and spec/parent_spec, so bump their refcounts.
5973          */
5974         __rbd_get_client(rbd_dev->rbd_client);
5975         rbd_spec_get(rbd_dev->parent_spec);
5976
5977         ret = rbd_dev_image_probe(parent, depth);
5978         if (ret < 0)
5979                 goto out_err;
5980
5981         rbd_dev->parent = parent;
5982         atomic_set(&rbd_dev->parent_ref, 1);
5983         return 0;
5984
5985 out_err:
5986         rbd_dev_unparent(rbd_dev);
5987         rbd_dev_destroy(parent);
5988         return ret;
5989 }
5990
5991 /*
5992  * rbd_dev->header_rwsem must be locked for write and will be unlocked
5993  * upon return.
5994  */
5995 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
5996 {
5997         int ret;
5998
5999         /* Record our major and minor device numbers. */
6000
6001         if (!single_major) {
6002                 ret = register_blkdev(0, rbd_dev->name);
6003                 if (ret < 0)
6004                         goto err_out_unlock;
6005
6006                 rbd_dev->major = ret;
6007                 rbd_dev->minor = 0;
6008         } else {
6009                 rbd_dev->major = rbd_major;
6010                 rbd_dev->minor = rbd_dev_id_to_minor(rbd_dev->dev_id);
6011         }
6012
6013         /* Set up the blkdev mapping. */
6014
6015         ret = rbd_init_disk(rbd_dev);
6016         if (ret)
6017                 goto err_out_blkdev;
6018
6019         ret = rbd_dev_mapping_set(rbd_dev);
6020         if (ret)
6021                 goto err_out_disk;
6022
6023         set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
6024         set_disk_ro(rbd_dev->disk, rbd_dev->mapping.read_only);
6025
6026         dev_set_name(&rbd_dev->dev, "%d", rbd_dev->dev_id);
6027         ret = device_add(&rbd_dev->dev);
6028         if (ret)
6029                 goto err_out_mapping;
6030
6031         /* Everything's ready.  Announce the disk to the world. */
6032
6033         set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
6034         up_write(&rbd_dev->header_rwsem);
6035
6036         spin_lock(&rbd_dev_list_lock);
6037         list_add_tail(&rbd_dev->node, &rbd_dev_list);
6038         spin_unlock(&rbd_dev_list_lock);
6039
6040         add_disk(rbd_dev->disk);
6041         pr_info("%s: capacity %llu features 0x%llx\n", rbd_dev->disk->disk_name,
6042                 (unsigned long long)get_capacity(rbd_dev->disk) << SECTOR_SHIFT,
6043                 rbd_dev->header.features);
6044
6045         return ret;
6046
6047 err_out_mapping:
6048         rbd_dev_mapping_clear(rbd_dev);
6049 err_out_disk:
6050         rbd_free_disk(rbd_dev);
6051 err_out_blkdev:
6052         if (!single_major)
6053                 unregister_blkdev(rbd_dev->major, rbd_dev->name);
6054 err_out_unlock:
6055         up_write(&rbd_dev->header_rwsem);
6056         return ret;
6057 }
6058
6059 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
6060 {
6061         struct rbd_spec *spec = rbd_dev->spec;
6062         int ret;
6063
6064         /* Record the header object name for this rbd image. */
6065
6066         rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
6067
6068         rbd_dev->header_oloc.pool = rbd_dev->layout.pool_id;
6069         if (rbd_dev->image_format == 1)
6070                 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
6071                                        spec->image_name, RBD_SUFFIX);
6072         else
6073                 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
6074                                        RBD_HEADER_PREFIX, spec->image_id);
6075
6076         return ret;
6077 }
6078
6079 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
6080 {
6081         rbd_dev_unprobe(rbd_dev);
6082         rbd_dev->image_format = 0;
6083         kfree(rbd_dev->spec->image_id);
6084         rbd_dev->spec->image_id = NULL;
6085
6086         rbd_dev_destroy(rbd_dev);
6087 }
6088
6089 /*
6090  * Probe for the existence of the header object for the given rbd
6091  * device.  If this image is the one being mapped (i.e., not a
6092  * parent), initiate a watch on its header object before using that
6093  * object to get detailed information about the rbd image.
6094  */
6095 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth)
6096 {
6097         int ret;
6098
6099         /*
6100          * Get the id from the image id object.  Unless there's an
6101          * error, rbd_dev->spec->image_id will be filled in with
6102          * a dynamically-allocated string, and rbd_dev->image_format
6103          * will be set to either 1 or 2.
6104          */
6105         ret = rbd_dev_image_id(rbd_dev);
6106         if (ret)
6107                 return ret;
6108
6109         ret = rbd_dev_header_name(rbd_dev);
6110         if (ret)
6111                 goto err_out_format;
6112
6113         if (!depth) {
6114                 ret = rbd_register_watch(rbd_dev);
6115                 if (ret) {
6116                         if (ret == -ENOENT)
6117                                 pr_info("image %s/%s does not exist\n",
6118                                         rbd_dev->spec->pool_name,
6119                                         rbd_dev->spec->image_name);
6120                         goto err_out_format;
6121                 }
6122         }
6123
6124         ret = rbd_dev_header_info(rbd_dev);
6125         if (ret)
6126                 goto err_out_watch;
6127
6128         /*
6129          * If this image is the one being mapped, we have pool name and
6130          * id, image name and id, and snap name - need to fill snap id.
6131          * Otherwise this is a parent image, identified by pool, image
6132          * and snap ids - need to fill in names for those ids.
6133          */
6134         if (!depth)
6135                 ret = rbd_spec_fill_snap_id(rbd_dev);
6136         else
6137                 ret = rbd_spec_fill_names(rbd_dev);
6138         if (ret) {
6139                 if (ret == -ENOENT)
6140                         pr_info("snap %s/%s@%s does not exist\n",
6141                                 rbd_dev->spec->pool_name,
6142                                 rbd_dev->spec->image_name,
6143                                 rbd_dev->spec->snap_name);
6144                 goto err_out_probe;
6145         }
6146
6147         if (rbd_dev->header.features & RBD_FEATURE_LAYERING) {
6148                 ret = rbd_dev_v2_parent_info(rbd_dev);
6149                 if (ret)
6150                         goto err_out_probe;
6151
6152                 /*
6153                  * Need to warn users if this image is the one being
6154                  * mapped and has a parent.
6155                  */
6156                 if (!depth && rbd_dev->parent_spec)
6157                         rbd_warn(rbd_dev,
6158                                  "WARNING: kernel layering is EXPERIMENTAL!");
6159         }
6160
6161         ret = rbd_dev_probe_parent(rbd_dev, depth);
6162         if (ret)
6163                 goto err_out_probe;
6164
6165         dout("discovered format %u image, header name is %s\n",
6166                 rbd_dev->image_format, rbd_dev->header_oid.name);
6167         return 0;
6168
6169 err_out_probe:
6170         rbd_dev_unprobe(rbd_dev);
6171 err_out_watch:
6172         if (!depth)
6173                 rbd_unregister_watch(rbd_dev);
6174 err_out_format:
6175         rbd_dev->image_format = 0;
6176         kfree(rbd_dev->spec->image_id);
6177         rbd_dev->spec->image_id = NULL;
6178         return ret;
6179 }
6180
6181 static ssize_t do_rbd_add(struct bus_type *bus,
6182                           const char *buf,
6183                           size_t count)
6184 {
6185         struct rbd_device *rbd_dev = NULL;
6186         struct ceph_options *ceph_opts = NULL;
6187         struct rbd_options *rbd_opts = NULL;
6188         struct rbd_spec *spec = NULL;
6189         struct rbd_client *rbdc;
6190         bool read_only;
6191         int rc;
6192
6193         if (!try_module_get(THIS_MODULE))
6194                 return -ENODEV;
6195
6196         /* parse add command */
6197         rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
6198         if (rc < 0)
6199                 goto out;
6200
6201         rbdc = rbd_get_client(ceph_opts);
6202         if (IS_ERR(rbdc)) {
6203                 rc = PTR_ERR(rbdc);
6204                 goto err_out_args;
6205         }
6206
6207         /* pick the pool */
6208         rc = rbd_add_get_pool_id(rbdc, spec->pool_name);
6209         if (rc < 0) {
6210                 if (rc == -ENOENT)
6211                         pr_info("pool %s does not exist\n", spec->pool_name);
6212                 goto err_out_client;
6213         }
6214         spec->pool_id = (u64)rc;
6215
6216         rbd_dev = rbd_dev_create(rbdc, spec, rbd_opts);
6217         if (!rbd_dev) {
6218                 rc = -ENOMEM;
6219                 goto err_out_client;
6220         }
6221         rbdc = NULL;            /* rbd_dev now owns this */
6222         spec = NULL;            /* rbd_dev now owns this */
6223         rbd_opts = NULL;        /* rbd_dev now owns this */
6224
6225         rbd_dev->config_info = kstrdup(buf, GFP_KERNEL);
6226         if (!rbd_dev->config_info) {
6227                 rc = -ENOMEM;
6228                 goto err_out_rbd_dev;
6229         }
6230
6231         down_write(&rbd_dev->header_rwsem);
6232         rc = rbd_dev_image_probe(rbd_dev, 0);
6233         if (rc < 0) {
6234                 up_write(&rbd_dev->header_rwsem);
6235                 goto err_out_rbd_dev;
6236         }
6237
6238         /* If we are mapping a snapshot it must be marked read-only */
6239
6240         read_only = rbd_dev->opts->read_only;
6241         if (rbd_dev->spec->snap_id != CEPH_NOSNAP)
6242                 read_only = true;
6243         rbd_dev->mapping.read_only = read_only;
6244
6245         rc = rbd_dev_device_setup(rbd_dev);
6246         if (rc) {
6247                 /*
6248                  * rbd_unregister_watch() can't be moved into
6249                  * rbd_dev_image_release() without refactoring, see
6250                  * commit 1f3ef78861ac.
6251                  */
6252                 rbd_unregister_watch(rbd_dev);
6253                 rbd_dev_image_release(rbd_dev);
6254                 goto out;
6255         }
6256
6257         rc = count;
6258 out:
6259         module_put(THIS_MODULE);
6260         return rc;
6261
6262 err_out_rbd_dev:
6263         rbd_dev_destroy(rbd_dev);
6264 err_out_client:
6265         rbd_put_client(rbdc);
6266 err_out_args:
6267         rbd_spec_put(spec);
6268         kfree(rbd_opts);
6269         goto out;
6270 }
6271
6272 static ssize_t rbd_add(struct bus_type *bus,
6273                        const char *buf,
6274                        size_t count)
6275 {
6276         if (single_major)
6277                 return -EINVAL;
6278
6279         return do_rbd_add(bus, buf, count);
6280 }
6281
6282 static ssize_t rbd_add_single_major(struct bus_type *bus,
6283                                     const char *buf,
6284                                     size_t count)
6285 {
6286         return do_rbd_add(bus, buf, count);
6287 }
6288
6289 static void rbd_dev_device_release(struct rbd_device *rbd_dev)
6290 {
6291         rbd_free_disk(rbd_dev);
6292
6293         spin_lock(&rbd_dev_list_lock);
6294         list_del_init(&rbd_dev->node);
6295         spin_unlock(&rbd_dev_list_lock);
6296
6297         clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
6298         device_del(&rbd_dev->dev);
6299         rbd_dev_mapping_clear(rbd_dev);
6300         if (!single_major)
6301                 unregister_blkdev(rbd_dev->major, rbd_dev->name);
6302 }
6303
6304 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
6305 {
6306         while (rbd_dev->parent) {
6307                 struct rbd_device *first = rbd_dev;
6308                 struct rbd_device *second = first->parent;
6309                 struct rbd_device *third;
6310
6311                 /*
6312                  * Follow to the parent with no grandparent and
6313                  * remove it.
6314                  */
6315                 while (second && (third = second->parent)) {
6316                         first = second;
6317                         second = third;
6318                 }
6319                 rbd_assert(second);
6320                 rbd_dev_image_release(second);
6321                 first->parent = NULL;
6322                 first->parent_overlap = 0;
6323
6324                 rbd_assert(first->parent_spec);
6325                 rbd_spec_put(first->parent_spec);
6326                 first->parent_spec = NULL;
6327         }
6328 }
6329
6330 static ssize_t do_rbd_remove(struct bus_type *bus,
6331                              const char *buf,
6332                              size_t count)
6333 {
6334         struct rbd_device *rbd_dev = NULL;
6335         struct list_head *tmp;
6336         int dev_id;
6337         char opt_buf[6];
6338         bool already = false;
6339         bool force = false;
6340         int ret;
6341
6342         dev_id = -1;
6343         opt_buf[0] = '\0';
6344         sscanf(buf, "%d %5s", &dev_id, opt_buf);
6345         if (dev_id < 0) {
6346                 pr_err("dev_id out of range\n");
6347                 return -EINVAL;
6348         }
6349         if (opt_buf[0] != '\0') {
6350                 if (!strcmp(opt_buf, "force")) {
6351                         force = true;
6352                 } else {
6353                         pr_err("bad remove option at '%s'\n", opt_buf);
6354                         return -EINVAL;
6355                 }
6356         }
6357
6358         ret = -ENOENT;
6359         spin_lock(&rbd_dev_list_lock);
6360         list_for_each(tmp, &rbd_dev_list) {
6361                 rbd_dev = list_entry(tmp, struct rbd_device, node);
6362                 if (rbd_dev->dev_id == dev_id) {
6363                         ret = 0;
6364                         break;
6365                 }
6366         }
6367         if (!ret) {
6368                 spin_lock_irq(&rbd_dev->lock);
6369                 if (rbd_dev->open_count && !force)
6370                         ret = -EBUSY;
6371                 else
6372                         already = test_and_set_bit(RBD_DEV_FLAG_REMOVING,
6373                                                         &rbd_dev->flags);
6374                 spin_unlock_irq(&rbd_dev->lock);
6375         }
6376         spin_unlock(&rbd_dev_list_lock);
6377         if (ret < 0 || already)
6378                 return ret;
6379
6380         if (force) {
6381                 /*
6382                  * Prevent new IO from being queued and wait for existing
6383                  * IO to complete/fail.
6384                  */
6385                 blk_mq_freeze_queue(rbd_dev->disk->queue);
6386                 blk_set_queue_dying(rbd_dev->disk->queue);
6387         }
6388
6389         down_write(&rbd_dev->lock_rwsem);
6390         if (__rbd_is_lock_owner(rbd_dev))
6391                 rbd_unlock(rbd_dev);
6392         up_write(&rbd_dev->lock_rwsem);
6393         rbd_unregister_watch(rbd_dev);
6394
6395         /*
6396          * Don't free anything from rbd_dev->disk until after all
6397          * notifies are completely processed. Otherwise
6398          * rbd_bus_del_dev() will race with rbd_watch_cb(), resulting
6399          * in a potential use after free of rbd_dev->disk or rbd_dev.
6400          */
6401         rbd_dev_device_release(rbd_dev);
6402         rbd_dev_image_release(rbd_dev);
6403
6404         return count;
6405 }
6406
6407 static ssize_t rbd_remove(struct bus_type *bus,
6408                           const char *buf,
6409                           size_t count)
6410 {
6411         if (single_major)
6412                 return -EINVAL;
6413
6414         return do_rbd_remove(bus, buf, count);
6415 }
6416
6417 static ssize_t rbd_remove_single_major(struct bus_type *bus,
6418                                        const char *buf,
6419                                        size_t count)
6420 {
6421         return do_rbd_remove(bus, buf, count);
6422 }
6423
6424 /*
6425  * create control files in sysfs
6426  * /sys/bus/rbd/...
6427  */
6428 static int rbd_sysfs_init(void)
6429 {
6430         int ret;
6431
6432         ret = device_register(&rbd_root_dev);
6433         if (ret < 0)
6434                 return ret;
6435
6436         ret = bus_register(&rbd_bus_type);
6437         if (ret < 0)
6438                 device_unregister(&rbd_root_dev);
6439
6440         return ret;
6441 }
6442
6443 static void rbd_sysfs_cleanup(void)
6444 {
6445         bus_unregister(&rbd_bus_type);
6446         device_unregister(&rbd_root_dev);
6447 }
6448
6449 static int rbd_slab_init(void)
6450 {
6451         rbd_assert(!rbd_img_request_cache);
6452         rbd_img_request_cache = KMEM_CACHE(rbd_img_request, 0);
6453         if (!rbd_img_request_cache)
6454                 return -ENOMEM;
6455
6456         rbd_assert(!rbd_obj_request_cache);
6457         rbd_obj_request_cache = KMEM_CACHE(rbd_obj_request, 0);
6458         if (!rbd_obj_request_cache)
6459                 goto out_err;
6460
6461         rbd_assert(!rbd_segment_name_cache);
6462         rbd_segment_name_cache = kmem_cache_create("rbd_segment_name",
6463                                         CEPH_MAX_OID_NAME_LEN + 1, 1, 0, NULL);
6464         if (rbd_segment_name_cache)
6465                 return 0;
6466 out_err:
6467         kmem_cache_destroy(rbd_obj_request_cache);
6468         rbd_obj_request_cache = NULL;
6469
6470         kmem_cache_destroy(rbd_img_request_cache);
6471         rbd_img_request_cache = NULL;
6472
6473         return -ENOMEM;
6474 }
6475
6476 static void rbd_slab_exit(void)
6477 {
6478         rbd_assert(rbd_segment_name_cache);
6479         kmem_cache_destroy(rbd_segment_name_cache);
6480         rbd_segment_name_cache = NULL;
6481
6482         rbd_assert(rbd_obj_request_cache);
6483         kmem_cache_destroy(rbd_obj_request_cache);
6484         rbd_obj_request_cache = NULL;
6485
6486         rbd_assert(rbd_img_request_cache);
6487         kmem_cache_destroy(rbd_img_request_cache);
6488         rbd_img_request_cache = NULL;
6489 }
6490
6491 static int __init rbd_init(void)
6492 {
6493         int rc;
6494
6495         if (!libceph_compatible(NULL)) {
6496                 rbd_warn(NULL, "libceph incompatibility (quitting)");
6497                 return -EINVAL;
6498         }
6499
6500         rc = rbd_slab_init();
6501         if (rc)
6502                 return rc;
6503
6504         /*
6505          * The number of active work items is limited by the number of
6506          * rbd devices * queue depth, so leave @max_active at default.
6507          */
6508         rbd_wq = alloc_workqueue(RBD_DRV_NAME, WQ_MEM_RECLAIM, 0);
6509         if (!rbd_wq) {
6510                 rc = -ENOMEM;
6511                 goto err_out_slab;
6512         }
6513
6514         if (single_major) {
6515                 rbd_major = register_blkdev(0, RBD_DRV_NAME);
6516                 if (rbd_major < 0) {
6517                         rc = rbd_major;
6518                         goto err_out_wq;
6519                 }
6520         }
6521
6522         rc = rbd_sysfs_init();
6523         if (rc)
6524                 goto err_out_blkdev;
6525
6526         if (single_major)
6527                 pr_info("loaded (major %d)\n", rbd_major);
6528         else
6529                 pr_info("loaded\n");
6530
6531         return 0;
6532
6533 err_out_blkdev:
6534         if (single_major)
6535                 unregister_blkdev(rbd_major, RBD_DRV_NAME);
6536 err_out_wq:
6537         destroy_workqueue(rbd_wq);
6538 err_out_slab:
6539         rbd_slab_exit();
6540         return rc;
6541 }
6542
6543 static void __exit rbd_exit(void)
6544 {
6545         ida_destroy(&rbd_dev_id_ida);
6546         rbd_sysfs_cleanup();
6547         if (single_major)
6548                 unregister_blkdev(rbd_major, RBD_DRV_NAME);
6549         destroy_workqueue(rbd_wq);
6550         rbd_slab_exit();
6551 }
6552
6553 module_init(rbd_init);
6554 module_exit(rbd_exit);
6555
6556 MODULE_AUTHOR("Alex Elder <elder@inktank.com>");
6557 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
6558 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
6559 /* following authorship retained from original osdblk.c */
6560 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
6561
6562 MODULE_DESCRIPTION("RADOS Block Device (RBD) driver");
6563 MODULE_LICENSE("GPL");