drivers/block/rbd.c

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