drivers/block/rbd.c

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