3 rbd.c -- Export ceph rados objects as a Linux block device
6 based on drivers/block/osdblk.c:
8 Copyright 2009 Red Hat, Inc.
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.
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.
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.
25 For usage instructions, please refer to:
27 Documentation/ABI/testing/sysfs-bus-rbd
31 #include <linux/ceph/libceph.h>
32 #include <linux/ceph/osd_client.h>
33 #include <linux/ceph/mon_client.h>
34 #include <linux/ceph/decode.h>
35 #include <linux/parser.h>
36 #include <linux/bsearch.h>
38 #include <linux/kernel.h>
39 #include <linux/device.h>
40 #include <linux/module.h>
42 #include <linux/blkdev.h>
43 #include <linux/slab.h>
45 #include "rbd_types.h"
47 #define RBD_DEBUG /* Activate rbd_assert() calls */
50 * The basic unit of block I/O is a sector. It is interpreted in a
51 * number of contexts in Linux (blk, bio, genhd), but the default is
52 * universally 512 bytes. These symbols are just slightly more
53 * meaningful than the bare numbers they represent.
55 #define SECTOR_SHIFT 9
56 #define SECTOR_SIZE (1ULL << SECTOR_SHIFT)
59 * Increment the given counter and return its updated value.
60 * If the counter is already 0 it will not be incremented.
61 * If the counter is already at its maximum value returns
62 * -EINVAL without updating it.
64 static int atomic_inc_return_safe(atomic_t *v)
68 counter = (unsigned int)__atomic_add_unless(v, 1, 0);
69 if (counter <= (unsigned int)INT_MAX)
77 /* Decrement the counter. Return the resulting value, or -EINVAL */
78 static int atomic_dec_return_safe(atomic_t *v)
82 counter = atomic_dec_return(v);
91 #define RBD_DRV_NAME "rbd"
92 #define RBD_DRV_NAME_LONG "rbd (rados block device)"
94 #define RBD_MINORS_PER_MAJOR 256 /* max minors per blkdev */
96 #define RBD_SNAP_DEV_NAME_PREFIX "snap_"
97 #define RBD_MAX_SNAP_NAME_LEN \
98 (NAME_MAX - (sizeof (RBD_SNAP_DEV_NAME_PREFIX) - 1))
100 #define RBD_MAX_SNAP_COUNT 510 /* allows max snapc to fit in 4KB */
102 #define RBD_SNAP_HEAD_NAME "-"
104 #define BAD_SNAP_INDEX U32_MAX /* invalid index into snap array */
106 /* This allows a single page to hold an image name sent by OSD */
107 #define RBD_IMAGE_NAME_LEN_MAX (PAGE_SIZE - sizeof (__le32) - 1)
108 #define RBD_IMAGE_ID_LEN_MAX 64
110 #define RBD_OBJ_PREFIX_LEN_MAX 64
114 #define RBD_FEATURE_LAYERING (1<<0)
115 #define RBD_FEATURE_STRIPINGV2 (1<<1)
116 #define RBD_FEATURES_ALL \
117 (RBD_FEATURE_LAYERING | RBD_FEATURE_STRIPINGV2)
119 /* Features supported by this (client software) implementation. */
121 #define RBD_FEATURES_SUPPORTED (RBD_FEATURES_ALL)
124 * An RBD device name will be "rbd#", where the "rbd" comes from
125 * RBD_DRV_NAME above, and # is a unique integer identifier.
126 * MAX_INT_FORMAT_WIDTH is used in ensuring DEV_NAME_LEN is big
127 * enough to hold all possible device names.
129 #define DEV_NAME_LEN 32
130 #define MAX_INT_FORMAT_WIDTH ((5 * sizeof (int)) / 2 + 1)
133 * block device image metadata (in-memory version)
135 struct rbd_image_header {
136 /* These six fields never change for a given rbd image */
143 u64 features; /* Might be changeable someday? */
145 /* The remaining fields need to be updated occasionally */
147 struct ceph_snap_context *snapc;
148 char *snap_names; /* format 1 only */
149 u64 *snap_sizes; /* format 1 only */
153 * An rbd image specification.
155 * The tuple (pool_id, image_id, snap_id) is sufficient to uniquely
156 * identify an image. Each rbd_dev structure includes a pointer to
157 * an rbd_spec structure that encapsulates this identity.
159 * Each of the id's in an rbd_spec has an associated name. For a
160 * user-mapped image, the names are supplied and the id's associated
161 * with them are looked up. For a layered image, a parent image is
162 * defined by the tuple, and the names are looked up.
164 * An rbd_dev structure contains a parent_spec pointer which is
165 * non-null if the image it represents is a child in a layered
166 * image. This pointer will refer to the rbd_spec structure used
167 * by the parent rbd_dev for its own identity (i.e., the structure
168 * is shared between the parent and child).
170 * Since these structures are populated once, during the discovery
171 * phase of image construction, they are effectively immutable so
172 * we make no effort to synchronize access to them.
174 * Note that code herein does not assume the image name is known (it
175 * could be a null pointer).
179 const char *pool_name;
181 const char *image_id;
182 const char *image_name;
185 const char *snap_name;
191 * an instance of the client. multiple devices may share an rbd client.
194 struct ceph_client *client;
196 struct list_head node;
199 struct rbd_img_request;
200 typedef void (*rbd_img_callback_t)(struct rbd_img_request *);
202 #define BAD_WHICH U32_MAX /* Good which or bad which, which? */
204 struct rbd_obj_request;
205 typedef void (*rbd_obj_callback_t)(struct rbd_obj_request *);
207 enum obj_request_type {
208 OBJ_REQUEST_NODATA, OBJ_REQUEST_BIO, OBJ_REQUEST_PAGES
212 OBJ_REQ_DONE, /* completion flag: not done = 0, done = 1 */
213 OBJ_REQ_IMG_DATA, /* object usage: standalone = 0, image = 1 */
214 OBJ_REQ_KNOWN, /* EXISTS flag valid: no = 0, yes = 1 */
215 OBJ_REQ_EXISTS, /* target exists: no = 0, yes = 1 */
218 struct rbd_obj_request {
219 const char *object_name;
220 u64 offset; /* object start byte */
221 u64 length; /* bytes from offset */
225 * An object request associated with an image will have its
226 * img_data flag set; a standalone object request will not.
228 * A standalone object request will have which == BAD_WHICH
229 * and a null obj_request pointer.
231 * An object request initiated in support of a layered image
232 * object (to check for its existence before a write) will
233 * have which == BAD_WHICH and a non-null obj_request pointer.
235 * Finally, an object request for rbd image data will have
236 * which != BAD_WHICH, and will have a non-null img_request
237 * pointer. The value of which will be in the range
238 * 0..(img_request->obj_request_count-1).
241 struct rbd_obj_request *obj_request; /* STAT op */
243 struct rbd_img_request *img_request;
245 /* links for img_request->obj_requests list */
246 struct list_head links;
249 u32 which; /* posn image request list */
251 enum obj_request_type type;
253 struct bio *bio_list;
259 struct page **copyup_pages;
260 u32 copyup_page_count;
262 struct ceph_osd_request *osd_req;
264 u64 xferred; /* bytes transferred */
267 rbd_obj_callback_t callback;
268 struct completion completion;
274 IMG_REQ_WRITE, /* I/O direction: read = 0, write = 1 */
275 IMG_REQ_CHILD, /* initiator: block = 0, child image = 1 */
276 IMG_REQ_LAYERED, /* ENOENT handling: normal = 0, layered = 1 */
279 struct rbd_img_request {
280 struct rbd_device *rbd_dev;
281 u64 offset; /* starting image byte offset */
282 u64 length; /* byte count from offset */
285 u64 snap_id; /* for reads */
286 struct ceph_snap_context *snapc; /* for writes */
289 struct request *rq; /* block request */
290 struct rbd_obj_request *obj_request; /* obj req initiator */
292 struct page **copyup_pages;
293 u32 copyup_page_count;
294 spinlock_t completion_lock;/* protects next_completion */
296 rbd_img_callback_t callback;
297 u64 xferred;/* aggregate bytes transferred */
298 int result; /* first nonzero obj_request result */
300 u32 obj_request_count;
301 struct list_head obj_requests; /* rbd_obj_request structs */
306 #define for_each_obj_request(ireq, oreq) \
307 list_for_each_entry(oreq, &(ireq)->obj_requests, links)
308 #define for_each_obj_request_from(ireq, oreq) \
309 list_for_each_entry_from(oreq, &(ireq)->obj_requests, links)
310 #define for_each_obj_request_safe(ireq, oreq, n) \
311 list_for_each_entry_safe_reverse(oreq, n, &(ireq)->obj_requests, links)
323 int dev_id; /* blkdev unique id */
325 int major; /* blkdev assigned major */
326 struct gendisk *disk; /* blkdev's gendisk and rq */
328 u32 image_format; /* Either 1 or 2 */
329 struct rbd_client *rbd_client;
331 char name[DEV_NAME_LEN]; /* blkdev name, e.g. rbd3 */
333 spinlock_t lock; /* queue, flags, open_count */
335 struct rbd_image_header header;
336 unsigned long flags; /* possibly lock protected */
337 struct rbd_spec *spec;
341 struct ceph_file_layout layout;
343 struct ceph_osd_event *watch_event;
344 struct rbd_obj_request *watch_request;
346 struct rbd_spec *parent_spec;
349 struct rbd_device *parent;
351 /* protects updating the header */
352 struct rw_semaphore header_rwsem;
354 struct rbd_mapping mapping;
356 struct list_head node;
360 unsigned long open_count; /* protected by lock */
364 * Flag bits for rbd_dev->flags. If atomicity is required,
365 * rbd_dev->lock is used to protect access.
367 * Currently, only the "removing" flag (which is coupled with the
368 * "open_count" field) requires atomic access.
371 RBD_DEV_FLAG_EXISTS, /* mapped snapshot has not been deleted */
372 RBD_DEV_FLAG_REMOVING, /* this mapping is being removed */
375 static DEFINE_MUTEX(ctl_mutex); /* Serialize open/close/setup/teardown */
377 static LIST_HEAD(rbd_dev_list); /* devices */
378 static DEFINE_SPINLOCK(rbd_dev_list_lock);
380 static LIST_HEAD(rbd_client_list); /* clients */
381 static DEFINE_SPINLOCK(rbd_client_list_lock);
383 /* Slab caches for frequently-allocated structures */
385 static struct kmem_cache *rbd_img_request_cache;
386 static struct kmem_cache *rbd_obj_request_cache;
387 static struct kmem_cache *rbd_segment_name_cache;
389 static int rbd_img_request_submit(struct rbd_img_request *img_request);
391 static void rbd_dev_device_release(struct device *dev);
393 static ssize_t rbd_add(struct bus_type *bus, const char *buf,
395 static ssize_t rbd_remove(struct bus_type *bus, const char *buf,
397 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, bool mapping);
398 static void rbd_spec_put(struct rbd_spec *spec);
400 static struct bus_attribute rbd_bus_attrs[] = {
401 __ATTR(add, S_IWUSR, NULL, rbd_add),
402 __ATTR(remove, S_IWUSR, NULL, rbd_remove),
406 static struct bus_type rbd_bus_type = {
408 .bus_attrs = rbd_bus_attrs,
411 static void rbd_root_dev_release(struct device *dev)
415 static struct device rbd_root_dev = {
417 .release = rbd_root_dev_release,
420 static __printf(2, 3)
421 void rbd_warn(struct rbd_device *rbd_dev, const char *fmt, ...)
423 struct va_format vaf;
431 printk(KERN_WARNING "%s: %pV\n", RBD_DRV_NAME, &vaf);
432 else if (rbd_dev->disk)
433 printk(KERN_WARNING "%s: %s: %pV\n",
434 RBD_DRV_NAME, rbd_dev->disk->disk_name, &vaf);
435 else if (rbd_dev->spec && rbd_dev->spec->image_name)
436 printk(KERN_WARNING "%s: image %s: %pV\n",
437 RBD_DRV_NAME, rbd_dev->spec->image_name, &vaf);
438 else if (rbd_dev->spec && rbd_dev->spec->image_id)
439 printk(KERN_WARNING "%s: id %s: %pV\n",
440 RBD_DRV_NAME, rbd_dev->spec->image_id, &vaf);
442 printk(KERN_WARNING "%s: rbd_dev %p: %pV\n",
443 RBD_DRV_NAME, rbd_dev, &vaf);
448 #define rbd_assert(expr) \
449 if (unlikely(!(expr))) { \
450 printk(KERN_ERR "\nAssertion failure in %s() " \
452 "\trbd_assert(%s);\n\n", \
453 __func__, __LINE__, #expr); \
456 #else /* !RBD_DEBUG */
457 # define rbd_assert(expr) ((void) 0)
458 #endif /* !RBD_DEBUG */
460 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request);
461 static void rbd_img_parent_read(struct rbd_obj_request *obj_request);
462 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev);
464 static int rbd_dev_refresh(struct rbd_device *rbd_dev);
465 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev);
466 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev);
467 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
469 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
470 u8 *order, u64 *snap_size);
471 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
473 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name);
475 static int rbd_open(struct block_device *bdev, fmode_t mode)
477 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
478 bool removing = false;
480 if ((mode & FMODE_WRITE) && rbd_dev->mapping.read_only)
483 spin_lock_irq(&rbd_dev->lock);
484 if (test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags))
487 rbd_dev->open_count++;
488 spin_unlock_irq(&rbd_dev->lock);
492 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
493 (void) get_device(&rbd_dev->dev);
494 set_device_ro(bdev, rbd_dev->mapping.read_only);
495 mutex_unlock(&ctl_mutex);
500 static void rbd_release(struct gendisk *disk, fmode_t mode)
502 struct rbd_device *rbd_dev = disk->private_data;
503 unsigned long open_count_before;
505 spin_lock_irq(&rbd_dev->lock);
506 open_count_before = rbd_dev->open_count--;
507 spin_unlock_irq(&rbd_dev->lock);
508 rbd_assert(open_count_before > 0);
510 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
511 put_device(&rbd_dev->dev);
512 mutex_unlock(&ctl_mutex);
515 static const struct block_device_operations rbd_bd_ops = {
516 .owner = THIS_MODULE,
518 .release = rbd_release,
522 * Initialize an rbd client instance. Success or not, this function
523 * consumes ceph_opts.
525 static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts)
527 struct rbd_client *rbdc;
530 dout("%s:\n", __func__);
531 rbdc = kmalloc(sizeof(struct rbd_client), GFP_KERNEL);
535 kref_init(&rbdc->kref);
536 INIT_LIST_HEAD(&rbdc->node);
538 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
540 rbdc->client = ceph_create_client(ceph_opts, rbdc, 0, 0);
541 if (IS_ERR(rbdc->client))
543 ceph_opts = NULL; /* Now rbdc->client is responsible for ceph_opts */
545 ret = ceph_open_session(rbdc->client);
549 spin_lock(&rbd_client_list_lock);
550 list_add_tail(&rbdc->node, &rbd_client_list);
551 spin_unlock(&rbd_client_list_lock);
553 mutex_unlock(&ctl_mutex);
554 dout("%s: rbdc %p\n", __func__, rbdc);
559 ceph_destroy_client(rbdc->client);
561 mutex_unlock(&ctl_mutex);
565 ceph_destroy_options(ceph_opts);
566 dout("%s: error %d\n", __func__, ret);
571 static struct rbd_client *__rbd_get_client(struct rbd_client *rbdc)
573 kref_get(&rbdc->kref);
579 * Find a ceph client with specific addr and configuration. If
580 * found, bump its reference count.
582 static struct rbd_client *rbd_client_find(struct ceph_options *ceph_opts)
584 struct rbd_client *client_node;
587 if (ceph_opts->flags & CEPH_OPT_NOSHARE)
590 spin_lock(&rbd_client_list_lock);
591 list_for_each_entry(client_node, &rbd_client_list, node) {
592 if (!ceph_compare_options(ceph_opts, client_node->client)) {
593 __rbd_get_client(client_node);
599 spin_unlock(&rbd_client_list_lock);
601 return found ? client_node : NULL;
611 /* string args above */
614 /* Boolean args above */
618 static match_table_t rbd_opts_tokens = {
620 /* string args above */
621 {Opt_read_only, "read_only"},
622 {Opt_read_only, "ro"}, /* Alternate spelling */
623 {Opt_read_write, "read_write"},
624 {Opt_read_write, "rw"}, /* Alternate spelling */
625 /* Boolean args above */
633 #define RBD_READ_ONLY_DEFAULT false
635 static int parse_rbd_opts_token(char *c, void *private)
637 struct rbd_options *rbd_opts = private;
638 substring_t argstr[MAX_OPT_ARGS];
639 int token, intval, ret;
641 token = match_token(c, rbd_opts_tokens, argstr);
645 if (token < Opt_last_int) {
646 ret = match_int(&argstr[0], &intval);
648 pr_err("bad mount option arg (not int) "
652 dout("got int token %d val %d\n", token, intval);
653 } else if (token > Opt_last_int && token < Opt_last_string) {
654 dout("got string token %d val %s\n", token,
656 } else if (token > Opt_last_string && token < Opt_last_bool) {
657 dout("got Boolean token %d\n", token);
659 dout("got token %d\n", token);
664 rbd_opts->read_only = true;
667 rbd_opts->read_only = false;
677 * Get a ceph client with specific addr and configuration, if one does
678 * not exist create it. Either way, ceph_opts is consumed by this
681 static struct rbd_client *rbd_get_client(struct ceph_options *ceph_opts)
683 struct rbd_client *rbdc;
685 rbdc = rbd_client_find(ceph_opts);
686 if (rbdc) /* using an existing client */
687 ceph_destroy_options(ceph_opts);
689 rbdc = rbd_client_create(ceph_opts);
695 * Destroy ceph client
697 * Caller must hold rbd_client_list_lock.
699 static void rbd_client_release(struct kref *kref)
701 struct rbd_client *rbdc = container_of(kref, struct rbd_client, kref);
703 dout("%s: rbdc %p\n", __func__, rbdc);
704 spin_lock(&rbd_client_list_lock);
705 list_del(&rbdc->node);
706 spin_unlock(&rbd_client_list_lock);
708 ceph_destroy_client(rbdc->client);
713 * Drop reference to ceph client node. If it's not referenced anymore, release
716 static void rbd_put_client(struct rbd_client *rbdc)
719 kref_put(&rbdc->kref, rbd_client_release);
722 static bool rbd_image_format_valid(u32 image_format)
724 return image_format == 1 || image_format == 2;
727 static bool rbd_dev_ondisk_valid(struct rbd_image_header_ondisk *ondisk)
732 /* The header has to start with the magic rbd header text */
733 if (memcmp(&ondisk->text, RBD_HEADER_TEXT, sizeof (RBD_HEADER_TEXT)))
736 /* The bio layer requires at least sector-sized I/O */
738 if (ondisk->options.order < SECTOR_SHIFT)
741 /* If we use u64 in a few spots we may be able to loosen this */
743 if (ondisk->options.order > 8 * sizeof (int) - 1)
747 * The size of a snapshot header has to fit in a size_t, and
748 * that limits the number of snapshots.
750 snap_count = le32_to_cpu(ondisk->snap_count);
751 size = SIZE_MAX - sizeof (struct ceph_snap_context);
752 if (snap_count > size / sizeof (__le64))
756 * Not only that, but the size of the entire the snapshot
757 * header must also be representable in a size_t.
759 size -= snap_count * sizeof (__le64);
760 if ((u64) size < le64_to_cpu(ondisk->snap_names_len))
767 * Fill an rbd image header with information from the given format 1
770 static int rbd_header_from_disk(struct rbd_device *rbd_dev,
771 struct rbd_image_header_ondisk *ondisk)
773 struct rbd_image_header *header = &rbd_dev->header;
774 bool first_time = header->object_prefix == NULL;
775 struct ceph_snap_context *snapc;
776 char *object_prefix = NULL;
777 char *snap_names = NULL;
778 u64 *snap_sizes = NULL;
784 /* Allocate this now to avoid having to handle failure below */
789 len = strnlen(ondisk->object_prefix,
790 sizeof (ondisk->object_prefix));
791 object_prefix = kmalloc(len + 1, GFP_KERNEL);
794 memcpy(object_prefix, ondisk->object_prefix, len);
795 object_prefix[len] = '\0';
798 /* Allocate the snapshot context and fill it in */
800 snap_count = le32_to_cpu(ondisk->snap_count);
801 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
804 snapc->seq = le64_to_cpu(ondisk->snap_seq);
806 struct rbd_image_snap_ondisk *snaps;
807 u64 snap_names_len = le64_to_cpu(ondisk->snap_names_len);
809 /* We'll keep a copy of the snapshot names... */
811 if (snap_names_len > (u64)SIZE_MAX)
813 snap_names = kmalloc(snap_names_len, GFP_KERNEL);
817 /* ...as well as the array of their sizes. */
819 size = snap_count * sizeof (*header->snap_sizes);
820 snap_sizes = kmalloc(size, GFP_KERNEL);
825 * Copy the names, and fill in each snapshot's id
828 * Note that rbd_dev_v1_header_info() guarantees the
829 * ondisk buffer we're working with has
830 * snap_names_len bytes beyond the end of the
831 * snapshot id array, this memcpy() is safe.
833 memcpy(snap_names, &ondisk->snaps[snap_count], snap_names_len);
834 snaps = ondisk->snaps;
835 for (i = 0; i < snap_count; i++) {
836 snapc->snaps[i] = le64_to_cpu(snaps[i].id);
837 snap_sizes[i] = le64_to_cpu(snaps[i].image_size);
841 /* We won't fail any more, fill in the header */
843 down_write(&rbd_dev->header_rwsem);
845 header->object_prefix = object_prefix;
846 header->obj_order = ondisk->options.order;
847 header->crypt_type = ondisk->options.crypt_type;
848 header->comp_type = ondisk->options.comp_type;
849 /* The rest aren't used for format 1 images */
850 header->stripe_unit = 0;
851 header->stripe_count = 0;
852 header->features = 0;
854 ceph_put_snap_context(header->snapc);
855 kfree(header->snap_names);
856 kfree(header->snap_sizes);
859 /* The remaining fields always get updated (when we refresh) */
861 header->image_size = le64_to_cpu(ondisk->image_size);
862 header->snapc = snapc;
863 header->snap_names = snap_names;
864 header->snap_sizes = snap_sizes;
866 /* Make sure mapping size is consistent with header info */
868 if (rbd_dev->spec->snap_id == CEPH_NOSNAP || first_time)
869 if (rbd_dev->mapping.size != header->image_size)
870 rbd_dev->mapping.size = header->image_size;
872 up_write(&rbd_dev->header_rwsem);
880 ceph_put_snap_context(snapc);
881 kfree(object_prefix);
886 static const char *_rbd_dev_v1_snap_name(struct rbd_device *rbd_dev, u32 which)
888 const char *snap_name;
890 rbd_assert(which < rbd_dev->header.snapc->num_snaps);
892 /* Skip over names until we find the one we are looking for */
894 snap_name = rbd_dev->header.snap_names;
896 snap_name += strlen(snap_name) + 1;
898 return kstrdup(snap_name, GFP_KERNEL);
902 * Snapshot id comparison function for use with qsort()/bsearch().
903 * Note that result is for snapshots in *descending* order.
905 static int snapid_compare_reverse(const void *s1, const void *s2)
907 u64 snap_id1 = *(u64 *)s1;
908 u64 snap_id2 = *(u64 *)s2;
910 if (snap_id1 < snap_id2)
912 return snap_id1 == snap_id2 ? 0 : -1;
916 * Search a snapshot context to see if the given snapshot id is
919 * Returns the position of the snapshot id in the array if it's found,
920 * or BAD_SNAP_INDEX otherwise.
922 * Note: The snapshot array is in kept sorted (by the osd) in
923 * reverse order, highest snapshot id first.
925 static u32 rbd_dev_snap_index(struct rbd_device *rbd_dev, u64 snap_id)
927 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
930 found = bsearch(&snap_id, &snapc->snaps, snapc->num_snaps,
931 sizeof (snap_id), snapid_compare_reverse);
933 return found ? (u32)(found - &snapc->snaps[0]) : BAD_SNAP_INDEX;
936 static const char *rbd_dev_v1_snap_name(struct rbd_device *rbd_dev,
940 const char *snap_name;
942 which = rbd_dev_snap_index(rbd_dev, snap_id);
943 if (which == BAD_SNAP_INDEX)
944 return ERR_PTR(-ENOENT);
946 snap_name = _rbd_dev_v1_snap_name(rbd_dev, which);
947 return snap_name ? snap_name : ERR_PTR(-ENOMEM);
950 static const char *rbd_snap_name(struct rbd_device *rbd_dev, u64 snap_id)
952 if (snap_id == CEPH_NOSNAP)
953 return RBD_SNAP_HEAD_NAME;
955 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
956 if (rbd_dev->image_format == 1)
957 return rbd_dev_v1_snap_name(rbd_dev, snap_id);
959 return rbd_dev_v2_snap_name(rbd_dev, snap_id);
962 static int rbd_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
965 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
966 if (snap_id == CEPH_NOSNAP) {
967 *snap_size = rbd_dev->header.image_size;
968 } else if (rbd_dev->image_format == 1) {
971 which = rbd_dev_snap_index(rbd_dev, snap_id);
972 if (which == BAD_SNAP_INDEX)
975 *snap_size = rbd_dev->header.snap_sizes[which];
980 ret = _rbd_dev_v2_snap_size(rbd_dev, snap_id, NULL, &size);
989 static int rbd_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
992 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
993 if (snap_id == CEPH_NOSNAP) {
994 *snap_features = rbd_dev->header.features;
995 } else if (rbd_dev->image_format == 1) {
996 *snap_features = 0; /* No features for format 1 */
1001 ret = _rbd_dev_v2_snap_features(rbd_dev, snap_id, &features);
1005 *snap_features = features;
1010 static int rbd_dev_mapping_set(struct rbd_device *rbd_dev)
1012 u64 snap_id = rbd_dev->spec->snap_id;
1017 ret = rbd_snap_size(rbd_dev, snap_id, &size);
1020 ret = rbd_snap_features(rbd_dev, snap_id, &features);
1024 rbd_dev->mapping.size = size;
1025 rbd_dev->mapping.features = features;
1030 static void rbd_dev_mapping_clear(struct rbd_device *rbd_dev)
1032 rbd_dev->mapping.size = 0;
1033 rbd_dev->mapping.features = 0;
1036 static const char *rbd_segment_name(struct rbd_device *rbd_dev, u64 offset)
1043 name = kmem_cache_alloc(rbd_segment_name_cache, GFP_NOIO);
1046 segment = offset >> rbd_dev->header.obj_order;
1047 name_format = "%s.%012llx";
1048 if (rbd_dev->image_format == 2)
1049 name_format = "%s.%016llx";
1050 ret = snprintf(name, MAX_OBJ_NAME_SIZE + 1, name_format,
1051 rbd_dev->header.object_prefix, segment);
1052 if (ret < 0 || ret > MAX_OBJ_NAME_SIZE) {
1053 pr_err("error formatting segment name for #%llu (%d)\n",
1062 static void rbd_segment_name_free(const char *name)
1064 /* The explicit cast here is needed to drop the const qualifier */
1066 kmem_cache_free(rbd_segment_name_cache, (void *)name);
1069 static u64 rbd_segment_offset(struct rbd_device *rbd_dev, u64 offset)
1071 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1073 return offset & (segment_size - 1);
1076 static u64 rbd_segment_length(struct rbd_device *rbd_dev,
1077 u64 offset, u64 length)
1079 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1081 offset &= segment_size - 1;
1083 rbd_assert(length <= U64_MAX - offset);
1084 if (offset + length > segment_size)
1085 length = segment_size - offset;
1091 * returns the size of an object in the image
1093 static u64 rbd_obj_bytes(struct rbd_image_header *header)
1095 return 1 << header->obj_order;
1102 static void bio_chain_put(struct bio *chain)
1108 chain = chain->bi_next;
1114 * zeros a bio chain, starting at specific offset
1116 static void zero_bio_chain(struct bio *chain, int start_ofs)
1119 unsigned long flags;
1125 bio_for_each_segment(bv, chain, i) {
1126 if (pos + bv->bv_len > start_ofs) {
1127 int remainder = max(start_ofs - pos, 0);
1128 buf = bvec_kmap_irq(bv, &flags);
1129 memset(buf + remainder, 0,
1130 bv->bv_len - remainder);
1131 flush_dcache_page(bv->bv_page);
1132 bvec_kunmap_irq(buf, &flags);
1137 chain = chain->bi_next;
1142 * similar to zero_bio_chain(), zeros data defined by a page array,
1143 * starting at the given byte offset from the start of the array and
1144 * continuing up to the given end offset. The pages array is
1145 * assumed to be big enough to hold all bytes up to the end.
1147 static void zero_pages(struct page **pages, u64 offset, u64 end)
1149 struct page **page = &pages[offset >> PAGE_SHIFT];
1151 rbd_assert(end > offset);
1152 rbd_assert(end - offset <= (u64)SIZE_MAX);
1153 while (offset < end) {
1156 unsigned long flags;
1159 page_offset = (size_t)(offset & ~PAGE_MASK);
1160 length = min(PAGE_SIZE - page_offset, (size_t)(end - offset));
1161 local_irq_save(flags);
1162 kaddr = kmap_atomic(*page);
1163 memset(kaddr + page_offset, 0, length);
1164 flush_dcache_page(*page);
1165 kunmap_atomic(kaddr);
1166 local_irq_restore(flags);
1174 * Clone a portion of a bio, starting at the given byte offset
1175 * and continuing for the number of bytes indicated.
1177 static struct bio *bio_clone_range(struct bio *bio_src,
1178 unsigned int offset,
1186 unsigned short end_idx;
1187 unsigned short vcnt;
1190 /* Handle the easy case for the caller */
1192 if (!offset && len == bio_src->bi_size)
1193 return bio_clone(bio_src, gfpmask);
1195 if (WARN_ON_ONCE(!len))
1197 if (WARN_ON_ONCE(len > bio_src->bi_size))
1199 if (WARN_ON_ONCE(offset > bio_src->bi_size - len))
1202 /* Find first affected segment... */
1205 bio_for_each_segment(bv, bio_src, idx) {
1206 if (resid < bv->bv_len)
1208 resid -= bv->bv_len;
1212 /* ...and the last affected segment */
1215 __bio_for_each_segment(bv, bio_src, end_idx, idx) {
1216 if (resid <= bv->bv_len)
1218 resid -= bv->bv_len;
1220 vcnt = end_idx - idx + 1;
1222 /* Build the clone */
1224 bio = bio_alloc(gfpmask, (unsigned int) vcnt);
1226 return NULL; /* ENOMEM */
1228 bio->bi_bdev = bio_src->bi_bdev;
1229 bio->bi_sector = bio_src->bi_sector + (offset >> SECTOR_SHIFT);
1230 bio->bi_rw = bio_src->bi_rw;
1231 bio->bi_flags |= 1 << BIO_CLONED;
1234 * Copy over our part of the bio_vec, then update the first
1235 * and last (or only) entries.
1237 memcpy(&bio->bi_io_vec[0], &bio_src->bi_io_vec[idx],
1238 vcnt * sizeof (struct bio_vec));
1239 bio->bi_io_vec[0].bv_offset += voff;
1241 bio->bi_io_vec[0].bv_len -= voff;
1242 bio->bi_io_vec[vcnt - 1].bv_len = resid;
1244 bio->bi_io_vec[0].bv_len = len;
1247 bio->bi_vcnt = vcnt;
1255 * Clone a portion of a bio chain, starting at the given byte offset
1256 * into the first bio in the source chain and continuing for the
1257 * number of bytes indicated. The result is another bio chain of
1258 * exactly the given length, or a null pointer on error.
1260 * The bio_src and offset parameters are both in-out. On entry they
1261 * refer to the first source bio and the offset into that bio where
1262 * the start of data to be cloned is located.
1264 * On return, bio_src is updated to refer to the bio in the source
1265 * chain that contains first un-cloned byte, and *offset will
1266 * contain the offset of that byte within that bio.
1268 static struct bio *bio_chain_clone_range(struct bio **bio_src,
1269 unsigned int *offset,
1273 struct bio *bi = *bio_src;
1274 unsigned int off = *offset;
1275 struct bio *chain = NULL;
1278 /* Build up a chain of clone bios up to the limit */
1280 if (!bi || off >= bi->bi_size || !len)
1281 return NULL; /* Nothing to clone */
1285 unsigned int bi_size;
1289 rbd_warn(NULL, "bio_chain exhausted with %u left", len);
1290 goto out_err; /* EINVAL; ran out of bio's */
1292 bi_size = min_t(unsigned int, bi->bi_size - off, len);
1293 bio = bio_clone_range(bi, off, bi_size, gfpmask);
1295 goto out_err; /* ENOMEM */
1298 end = &bio->bi_next;
1301 if (off == bi->bi_size) {
1312 bio_chain_put(chain);
1318 * The default/initial value for all object request flags is 0. For
1319 * each flag, once its value is set to 1 it is never reset to 0
1322 static void obj_request_img_data_set(struct rbd_obj_request *obj_request)
1324 if (test_and_set_bit(OBJ_REQ_IMG_DATA, &obj_request->flags)) {
1325 struct rbd_device *rbd_dev;
1327 rbd_dev = obj_request->img_request->rbd_dev;
1328 rbd_warn(rbd_dev, "obj_request %p already marked img_data\n",
1333 static bool obj_request_img_data_test(struct rbd_obj_request *obj_request)
1336 return test_bit(OBJ_REQ_IMG_DATA, &obj_request->flags) != 0;
1339 static void obj_request_done_set(struct rbd_obj_request *obj_request)
1341 if (test_and_set_bit(OBJ_REQ_DONE, &obj_request->flags)) {
1342 struct rbd_device *rbd_dev = NULL;
1344 if (obj_request_img_data_test(obj_request))
1345 rbd_dev = obj_request->img_request->rbd_dev;
1346 rbd_warn(rbd_dev, "obj_request %p already marked done\n",
1351 static bool obj_request_done_test(struct rbd_obj_request *obj_request)
1354 return test_bit(OBJ_REQ_DONE, &obj_request->flags) != 0;
1358 * This sets the KNOWN flag after (possibly) setting the EXISTS
1359 * flag. The latter is set based on the "exists" value provided.
1361 * Note that for our purposes once an object exists it never goes
1362 * away again. It's possible that the response from two existence
1363 * checks are separated by the creation of the target object, and
1364 * the first ("doesn't exist") response arrives *after* the second
1365 * ("does exist"). In that case we ignore the second one.
1367 static void obj_request_existence_set(struct rbd_obj_request *obj_request,
1371 set_bit(OBJ_REQ_EXISTS, &obj_request->flags);
1372 set_bit(OBJ_REQ_KNOWN, &obj_request->flags);
1376 static bool obj_request_known_test(struct rbd_obj_request *obj_request)
1379 return test_bit(OBJ_REQ_KNOWN, &obj_request->flags) != 0;
1382 static bool obj_request_exists_test(struct rbd_obj_request *obj_request)
1385 return test_bit(OBJ_REQ_EXISTS, &obj_request->flags) != 0;
1388 static void rbd_obj_request_get(struct rbd_obj_request *obj_request)
1390 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1391 atomic_read(&obj_request->kref.refcount));
1392 kref_get(&obj_request->kref);
1395 static void rbd_obj_request_destroy(struct kref *kref);
1396 static void rbd_obj_request_put(struct rbd_obj_request *obj_request)
1398 rbd_assert(obj_request != NULL);
1399 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1400 atomic_read(&obj_request->kref.refcount));
1401 kref_put(&obj_request->kref, rbd_obj_request_destroy);
1404 static bool img_request_child_test(struct rbd_img_request *img_request);
1405 static void rbd_parent_request_destroy(struct kref *kref);
1406 static void rbd_img_request_destroy(struct kref *kref);
1407 static void rbd_img_request_put(struct rbd_img_request *img_request)
1409 rbd_assert(img_request != NULL);
1410 dout("%s: img %p (was %d)\n", __func__, img_request,
1411 atomic_read(&img_request->kref.refcount));
1412 if (img_request_child_test(img_request))
1413 kref_put(&img_request->kref, rbd_parent_request_destroy);
1415 kref_put(&img_request->kref, rbd_img_request_destroy);
1418 static inline void rbd_img_obj_request_add(struct rbd_img_request *img_request,
1419 struct rbd_obj_request *obj_request)
1421 rbd_assert(obj_request->img_request == NULL);
1423 /* Image request now owns object's original reference */
1424 obj_request->img_request = img_request;
1425 obj_request->which = img_request->obj_request_count;
1426 rbd_assert(!obj_request_img_data_test(obj_request));
1427 obj_request_img_data_set(obj_request);
1428 rbd_assert(obj_request->which != BAD_WHICH);
1429 img_request->obj_request_count++;
1430 list_add_tail(&obj_request->links, &img_request->obj_requests);
1431 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1432 obj_request->which);
1435 static inline void rbd_img_obj_request_del(struct rbd_img_request *img_request,
1436 struct rbd_obj_request *obj_request)
1438 rbd_assert(obj_request->which != BAD_WHICH);
1440 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1441 obj_request->which);
1442 list_del(&obj_request->links);
1443 rbd_assert(img_request->obj_request_count > 0);
1444 img_request->obj_request_count--;
1445 rbd_assert(obj_request->which == img_request->obj_request_count);
1446 obj_request->which = BAD_WHICH;
1447 rbd_assert(obj_request_img_data_test(obj_request));
1448 rbd_assert(obj_request->img_request == img_request);
1449 obj_request->img_request = NULL;
1450 obj_request->callback = NULL;
1451 rbd_obj_request_put(obj_request);
1454 static bool obj_request_type_valid(enum obj_request_type type)
1457 case OBJ_REQUEST_NODATA:
1458 case OBJ_REQUEST_BIO:
1459 case OBJ_REQUEST_PAGES:
1466 static int rbd_obj_request_submit(struct ceph_osd_client *osdc,
1467 struct rbd_obj_request *obj_request)
1469 dout("%s: osdc %p obj %p\n", __func__, osdc, obj_request);
1471 return ceph_osdc_start_request(osdc, obj_request->osd_req, false);
1474 static void rbd_img_request_complete(struct rbd_img_request *img_request)
1477 dout("%s: img %p\n", __func__, img_request);
1480 * If no error occurred, compute the aggregate transfer
1481 * count for the image request. We could instead use
1482 * atomic64_cmpxchg() to update it as each object request
1483 * completes; not clear which way is better off hand.
1485 if (!img_request->result) {
1486 struct rbd_obj_request *obj_request;
1489 for_each_obj_request(img_request, obj_request)
1490 xferred += obj_request->xferred;
1491 img_request->xferred = xferred;
1494 if (img_request->callback)
1495 img_request->callback(img_request);
1497 rbd_img_request_put(img_request);
1500 /* Caller is responsible for rbd_obj_request_destroy(obj_request) */
1502 static int rbd_obj_request_wait(struct rbd_obj_request *obj_request)
1504 dout("%s: obj %p\n", __func__, obj_request);
1506 return wait_for_completion_interruptible(&obj_request->completion);
1510 * The default/initial value for all image request flags is 0. Each
1511 * is conditionally set to 1 at image request initialization time
1512 * and currently never change thereafter.
1514 static void img_request_write_set(struct rbd_img_request *img_request)
1516 set_bit(IMG_REQ_WRITE, &img_request->flags);
1520 static bool img_request_write_test(struct rbd_img_request *img_request)
1523 return test_bit(IMG_REQ_WRITE, &img_request->flags) != 0;
1526 static void img_request_child_set(struct rbd_img_request *img_request)
1528 set_bit(IMG_REQ_CHILD, &img_request->flags);
1532 static void img_request_child_clear(struct rbd_img_request *img_request)
1534 clear_bit(IMG_REQ_CHILD, &img_request->flags);
1538 static bool img_request_child_test(struct rbd_img_request *img_request)
1541 return test_bit(IMG_REQ_CHILD, &img_request->flags) != 0;
1544 static void img_request_layered_set(struct rbd_img_request *img_request)
1546 set_bit(IMG_REQ_LAYERED, &img_request->flags);
1550 static void img_request_layered_clear(struct rbd_img_request *img_request)
1552 clear_bit(IMG_REQ_LAYERED, &img_request->flags);
1556 static bool img_request_layered_test(struct rbd_img_request *img_request)
1559 return test_bit(IMG_REQ_LAYERED, &img_request->flags) != 0;
1563 rbd_img_obj_request_read_callback(struct rbd_obj_request *obj_request)
1565 u64 xferred = obj_request->xferred;
1566 u64 length = obj_request->length;
1568 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1569 obj_request, obj_request->img_request, obj_request->result,
1572 * ENOENT means a hole in the image. We zero-fill the entire
1573 * length of the request. A short read also implies zero-fill
1574 * to the end of the request. An error requires the whole
1575 * length of the request to be reported finished with an error
1576 * to the block layer. In each case we update the xferred
1577 * count to indicate the whole request was satisfied.
1579 rbd_assert(obj_request->type != OBJ_REQUEST_NODATA);
1580 if (obj_request->result == -ENOENT) {
1581 if (obj_request->type == OBJ_REQUEST_BIO)
1582 zero_bio_chain(obj_request->bio_list, 0);
1584 zero_pages(obj_request->pages, 0, length);
1585 obj_request->result = 0;
1586 } else if (xferred < length && !obj_request->result) {
1587 if (obj_request->type == OBJ_REQUEST_BIO)
1588 zero_bio_chain(obj_request->bio_list, xferred);
1590 zero_pages(obj_request->pages, xferred, length);
1592 obj_request->xferred = length;
1593 obj_request_done_set(obj_request);
1596 static void rbd_obj_request_complete(struct rbd_obj_request *obj_request)
1598 dout("%s: obj %p cb %p\n", __func__, obj_request,
1599 obj_request->callback);
1600 if (obj_request->callback)
1601 obj_request->callback(obj_request);
1603 complete_all(&obj_request->completion);
1606 static void rbd_osd_trivial_callback(struct rbd_obj_request *obj_request)
1608 dout("%s: obj %p\n", __func__, obj_request);
1609 obj_request_done_set(obj_request);
1612 static void rbd_osd_read_callback(struct rbd_obj_request *obj_request)
1614 struct rbd_img_request *img_request = NULL;
1615 struct rbd_device *rbd_dev = NULL;
1616 bool layered = false;
1618 if (obj_request_img_data_test(obj_request)) {
1619 img_request = obj_request->img_request;
1620 layered = img_request && img_request_layered_test(img_request);
1621 rbd_dev = img_request->rbd_dev;
1624 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1625 obj_request, img_request, obj_request->result,
1626 obj_request->xferred, obj_request->length);
1627 if (layered && obj_request->result == -ENOENT &&
1628 obj_request->img_offset < rbd_dev->parent_overlap)
1629 rbd_img_parent_read(obj_request);
1630 else if (img_request)
1631 rbd_img_obj_request_read_callback(obj_request);
1633 obj_request_done_set(obj_request);
1636 static void rbd_osd_write_callback(struct rbd_obj_request *obj_request)
1638 dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1639 obj_request->result, obj_request->length);
1641 * There is no such thing as a successful short write. Set
1642 * it to our originally-requested length.
1644 obj_request->xferred = obj_request->length;
1645 obj_request_done_set(obj_request);
1649 * For a simple stat call there's nothing to do. We'll do more if
1650 * this is part of a write sequence for a layered image.
1652 static void rbd_osd_stat_callback(struct rbd_obj_request *obj_request)
1654 dout("%s: obj %p\n", __func__, obj_request);
1655 obj_request_done_set(obj_request);
1658 static void rbd_osd_req_callback(struct ceph_osd_request *osd_req,
1659 struct ceph_msg *msg)
1661 struct rbd_obj_request *obj_request = osd_req->r_priv;
1664 dout("%s: osd_req %p msg %p\n", __func__, osd_req, msg);
1665 rbd_assert(osd_req == obj_request->osd_req);
1666 if (obj_request_img_data_test(obj_request)) {
1667 rbd_assert(obj_request->img_request);
1668 rbd_assert(obj_request->which != BAD_WHICH);
1670 rbd_assert(obj_request->which == BAD_WHICH);
1673 if (osd_req->r_result < 0)
1674 obj_request->result = osd_req->r_result;
1676 BUG_ON(osd_req->r_num_ops > 2);
1679 * We support a 64-bit length, but ultimately it has to be
1680 * passed to blk_end_request(), which takes an unsigned int.
1682 obj_request->xferred = osd_req->r_reply_op_len[0];
1683 rbd_assert(obj_request->xferred < (u64)UINT_MAX);
1684 opcode = osd_req->r_ops[0].op;
1686 case CEPH_OSD_OP_READ:
1687 rbd_osd_read_callback(obj_request);
1689 case CEPH_OSD_OP_WRITE:
1690 rbd_osd_write_callback(obj_request);
1692 case CEPH_OSD_OP_STAT:
1693 rbd_osd_stat_callback(obj_request);
1695 case CEPH_OSD_OP_CALL:
1696 case CEPH_OSD_OP_NOTIFY_ACK:
1697 case CEPH_OSD_OP_WATCH:
1698 rbd_osd_trivial_callback(obj_request);
1701 rbd_warn(NULL, "%s: unsupported op %hu\n",
1702 obj_request->object_name, (unsigned short) opcode);
1706 if (obj_request_done_test(obj_request))
1707 rbd_obj_request_complete(obj_request);
1710 static void rbd_osd_req_format_read(struct rbd_obj_request *obj_request)
1712 struct rbd_img_request *img_request = obj_request->img_request;
1713 struct ceph_osd_request *osd_req = obj_request->osd_req;
1716 rbd_assert(osd_req != NULL);
1718 snap_id = img_request ? img_request->snap_id : CEPH_NOSNAP;
1719 ceph_osdc_build_request(osd_req, obj_request->offset,
1720 NULL, snap_id, NULL);
1723 static void rbd_osd_req_format_write(struct rbd_obj_request *obj_request)
1725 struct rbd_img_request *img_request = obj_request->img_request;
1726 struct ceph_osd_request *osd_req = obj_request->osd_req;
1727 struct ceph_snap_context *snapc;
1728 struct timespec mtime = CURRENT_TIME;
1730 rbd_assert(osd_req != NULL);
1732 snapc = img_request ? img_request->snapc : NULL;
1733 ceph_osdc_build_request(osd_req, obj_request->offset,
1734 snapc, CEPH_NOSNAP, &mtime);
1737 static struct ceph_osd_request *rbd_osd_req_create(
1738 struct rbd_device *rbd_dev,
1740 struct rbd_obj_request *obj_request)
1742 struct ceph_snap_context *snapc = NULL;
1743 struct ceph_osd_client *osdc;
1744 struct ceph_osd_request *osd_req;
1746 if (obj_request_img_data_test(obj_request)) {
1747 struct rbd_img_request *img_request = obj_request->img_request;
1749 rbd_assert(write_request ==
1750 img_request_write_test(img_request));
1752 snapc = img_request->snapc;
1755 /* Allocate and initialize the request, for the single op */
1757 osdc = &rbd_dev->rbd_client->client->osdc;
1758 osd_req = ceph_osdc_alloc_request(osdc, snapc, 1, false, GFP_ATOMIC);
1760 return NULL; /* ENOMEM */
1763 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
1765 osd_req->r_flags = CEPH_OSD_FLAG_READ;
1767 osd_req->r_callback = rbd_osd_req_callback;
1768 osd_req->r_priv = obj_request;
1770 osd_req->r_oid_len = strlen(obj_request->object_name);
1771 rbd_assert(osd_req->r_oid_len < sizeof (osd_req->r_oid));
1772 memcpy(osd_req->r_oid, obj_request->object_name, osd_req->r_oid_len);
1774 osd_req->r_file_layout = rbd_dev->layout; /* struct */
1780 * Create a copyup osd request based on the information in the
1781 * object request supplied. A copyup request has two osd ops,
1782 * a copyup method call, and a "normal" write request.
1784 static struct ceph_osd_request *
1785 rbd_osd_req_create_copyup(struct rbd_obj_request *obj_request)
1787 struct rbd_img_request *img_request;
1788 struct ceph_snap_context *snapc;
1789 struct rbd_device *rbd_dev;
1790 struct ceph_osd_client *osdc;
1791 struct ceph_osd_request *osd_req;
1793 rbd_assert(obj_request_img_data_test(obj_request));
1794 img_request = obj_request->img_request;
1795 rbd_assert(img_request);
1796 rbd_assert(img_request_write_test(img_request));
1798 /* Allocate and initialize the request, for the two ops */
1800 snapc = img_request->snapc;
1801 rbd_dev = img_request->rbd_dev;
1802 osdc = &rbd_dev->rbd_client->client->osdc;
1803 osd_req = ceph_osdc_alloc_request(osdc, snapc, 2, false, GFP_ATOMIC);
1805 return NULL; /* ENOMEM */
1807 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
1808 osd_req->r_callback = rbd_osd_req_callback;
1809 osd_req->r_priv = obj_request;
1811 osd_req->r_oid_len = strlen(obj_request->object_name);
1812 rbd_assert(osd_req->r_oid_len < sizeof (osd_req->r_oid));
1813 memcpy(osd_req->r_oid, obj_request->object_name, osd_req->r_oid_len);
1815 osd_req->r_file_layout = rbd_dev->layout; /* struct */
1821 static void rbd_osd_req_destroy(struct ceph_osd_request *osd_req)
1823 ceph_osdc_put_request(osd_req);
1826 /* object_name is assumed to be a non-null pointer and NUL-terminated */
1828 static struct rbd_obj_request *rbd_obj_request_create(const char *object_name,
1829 u64 offset, u64 length,
1830 enum obj_request_type type)
1832 struct rbd_obj_request *obj_request;
1836 rbd_assert(obj_request_type_valid(type));
1838 size = strlen(object_name) + 1;
1839 name = kmalloc(size, GFP_KERNEL);
1843 obj_request = kmem_cache_zalloc(rbd_obj_request_cache, GFP_KERNEL);
1849 obj_request->object_name = memcpy(name, object_name, size);
1850 obj_request->offset = offset;
1851 obj_request->length = length;
1852 obj_request->flags = 0;
1853 obj_request->which = BAD_WHICH;
1854 obj_request->type = type;
1855 INIT_LIST_HEAD(&obj_request->links);
1856 init_completion(&obj_request->completion);
1857 kref_init(&obj_request->kref);
1859 dout("%s: \"%s\" %llu/%llu %d -> obj %p\n", __func__, object_name,
1860 offset, length, (int)type, obj_request);
1865 static void rbd_obj_request_destroy(struct kref *kref)
1867 struct rbd_obj_request *obj_request;
1869 obj_request = container_of(kref, struct rbd_obj_request, kref);
1871 dout("%s: obj %p\n", __func__, obj_request);
1873 rbd_assert(obj_request->img_request == NULL);
1874 rbd_assert(obj_request->which == BAD_WHICH);
1876 if (obj_request->osd_req)
1877 rbd_osd_req_destroy(obj_request->osd_req);
1879 rbd_assert(obj_request_type_valid(obj_request->type));
1880 switch (obj_request->type) {
1881 case OBJ_REQUEST_NODATA:
1882 break; /* Nothing to do */
1883 case OBJ_REQUEST_BIO:
1884 if (obj_request->bio_list)
1885 bio_chain_put(obj_request->bio_list);
1887 case OBJ_REQUEST_PAGES:
1888 if (obj_request->pages)
1889 ceph_release_page_vector(obj_request->pages,
1890 obj_request->page_count);
1894 kfree(obj_request->object_name);
1895 obj_request->object_name = NULL;
1896 kmem_cache_free(rbd_obj_request_cache, obj_request);
1899 /* It's OK to call this for a device with no parent */
1901 static void rbd_spec_put(struct rbd_spec *spec);
1902 static void rbd_dev_unparent(struct rbd_device *rbd_dev)
1904 rbd_dev_remove_parent(rbd_dev);
1905 rbd_spec_put(rbd_dev->parent_spec);
1906 rbd_dev->parent_spec = NULL;
1907 rbd_dev->parent_overlap = 0;
1911 * Parent image reference counting is used to determine when an
1912 * image's parent fields can be safely torn down--after there are no
1913 * more in-flight requests to the parent image. When the last
1914 * reference is dropped, cleaning them up is safe.
1916 static void rbd_dev_parent_put(struct rbd_device *rbd_dev)
1920 if (!rbd_dev->parent_spec)
1923 counter = atomic_dec_return_safe(&rbd_dev->parent_ref);
1927 /* Last reference; clean up parent data structures */
1930 rbd_dev_unparent(rbd_dev);
1932 rbd_warn(rbd_dev, "parent reference underflow\n");
1936 * If an image has a non-zero parent overlap, get a reference to its
1939 * We must get the reference before checking for the overlap to
1940 * coordinate properly with zeroing the parent overlap in
1941 * rbd_dev_v2_parent_info() when an image gets flattened. We
1942 * drop it again if there is no overlap.
1944 * Returns true if the rbd device has a parent with a non-zero
1945 * overlap and a reference for it was successfully taken, or
1948 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
1952 if (!rbd_dev->parent_spec)
1955 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
1956 if (counter > 0 && rbd_dev->parent_overlap)
1959 /* Image was flattened, but parent is not yet torn down */
1962 rbd_warn(rbd_dev, "parent reference overflow\n");
1968 * Caller is responsible for filling in the list of object requests
1969 * that comprises the image request, and the Linux request pointer
1970 * (if there is one).
1972 static struct rbd_img_request *rbd_img_request_create(
1973 struct rbd_device *rbd_dev,
1974 u64 offset, u64 length,
1977 struct rbd_img_request *img_request;
1979 img_request = kmem_cache_alloc(rbd_img_request_cache, GFP_ATOMIC);
1983 if (write_request) {
1984 down_read(&rbd_dev->header_rwsem);
1985 ceph_get_snap_context(rbd_dev->header.snapc);
1986 up_read(&rbd_dev->header_rwsem);
1989 img_request->rq = NULL;
1990 img_request->rbd_dev = rbd_dev;
1991 img_request->offset = offset;
1992 img_request->length = length;
1993 img_request->flags = 0;
1994 if (write_request) {
1995 img_request_write_set(img_request);
1996 img_request->snapc = rbd_dev->header.snapc;
1998 img_request->snap_id = rbd_dev->spec->snap_id;
2000 if (rbd_dev_parent_get(rbd_dev))
2001 img_request_layered_set(img_request);
2002 spin_lock_init(&img_request->completion_lock);
2003 img_request->next_completion = 0;
2004 img_request->callback = NULL;
2005 img_request->result = 0;
2006 img_request->obj_request_count = 0;
2007 INIT_LIST_HEAD(&img_request->obj_requests);
2008 kref_init(&img_request->kref);
2010 dout("%s: rbd_dev %p %s %llu/%llu -> img %p\n", __func__, rbd_dev,
2011 write_request ? "write" : "read", offset, length,
2017 static void rbd_img_request_destroy(struct kref *kref)
2019 struct rbd_img_request *img_request;
2020 struct rbd_obj_request *obj_request;
2021 struct rbd_obj_request *next_obj_request;
2023 img_request = container_of(kref, struct rbd_img_request, kref);
2025 dout("%s: img %p\n", __func__, img_request);
2027 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2028 rbd_img_obj_request_del(img_request, obj_request);
2029 rbd_assert(img_request->obj_request_count == 0);
2031 if (img_request_layered_test(img_request)) {
2032 img_request_layered_clear(img_request);
2033 rbd_dev_parent_put(img_request->rbd_dev);
2036 if (img_request_write_test(img_request))
2037 ceph_put_snap_context(img_request->snapc);
2039 kmem_cache_free(rbd_img_request_cache, img_request);
2042 static struct rbd_img_request *rbd_parent_request_create(
2043 struct rbd_obj_request *obj_request,
2044 u64 img_offset, u64 length)
2046 struct rbd_img_request *parent_request;
2047 struct rbd_device *rbd_dev;
2049 rbd_assert(obj_request->img_request);
2050 rbd_dev = obj_request->img_request->rbd_dev;
2052 parent_request = rbd_img_request_create(rbd_dev->parent,
2053 img_offset, length, false);
2054 if (!parent_request)
2057 img_request_child_set(parent_request);
2058 rbd_obj_request_get(obj_request);
2059 parent_request->obj_request = obj_request;
2061 return parent_request;
2064 static void rbd_parent_request_destroy(struct kref *kref)
2066 struct rbd_img_request *parent_request;
2067 struct rbd_obj_request *orig_request;
2069 parent_request = container_of(kref, struct rbd_img_request, kref);
2070 orig_request = parent_request->obj_request;
2072 parent_request->obj_request = NULL;
2073 rbd_obj_request_put(orig_request);
2074 img_request_child_clear(parent_request);
2076 rbd_img_request_destroy(kref);
2079 static bool rbd_img_obj_end_request(struct rbd_obj_request *obj_request)
2081 struct rbd_img_request *img_request;
2082 unsigned int xferred;
2086 rbd_assert(obj_request_img_data_test(obj_request));
2087 img_request = obj_request->img_request;
2089 rbd_assert(obj_request->xferred <= (u64)UINT_MAX);
2090 xferred = (unsigned int)obj_request->xferred;
2091 result = obj_request->result;
2093 struct rbd_device *rbd_dev = img_request->rbd_dev;
2095 rbd_warn(rbd_dev, "%s %llx at %llx (%llx)\n",
2096 img_request_write_test(img_request) ? "write" : "read",
2097 obj_request->length, obj_request->img_offset,
2098 obj_request->offset);
2099 rbd_warn(rbd_dev, " result %d xferred %x\n",
2101 if (!img_request->result)
2102 img_request->result = result;
2105 /* Image object requests don't own their page array */
2107 if (obj_request->type == OBJ_REQUEST_PAGES) {
2108 obj_request->pages = NULL;
2109 obj_request->page_count = 0;
2112 if (img_request_child_test(img_request)) {
2113 rbd_assert(img_request->obj_request != NULL);
2114 more = obj_request->which < img_request->obj_request_count - 1;
2116 rbd_assert(img_request->rq != NULL);
2117 more = blk_end_request(img_request->rq, result, xferred);
2123 static void rbd_img_obj_callback(struct rbd_obj_request *obj_request)
2125 struct rbd_img_request *img_request;
2126 u32 which = obj_request->which;
2129 rbd_assert(obj_request_img_data_test(obj_request));
2130 img_request = obj_request->img_request;
2132 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
2133 rbd_assert(img_request != NULL);
2134 rbd_assert(img_request->obj_request_count > 0);
2135 rbd_assert(which != BAD_WHICH);
2136 rbd_assert(which < img_request->obj_request_count);
2137 rbd_assert(which >= img_request->next_completion);
2139 spin_lock_irq(&img_request->completion_lock);
2140 if (which != img_request->next_completion)
2143 for_each_obj_request_from(img_request, obj_request) {
2145 rbd_assert(which < img_request->obj_request_count);
2147 if (!obj_request_done_test(obj_request))
2149 more = rbd_img_obj_end_request(obj_request);
2153 rbd_assert(more ^ (which == img_request->obj_request_count));
2154 img_request->next_completion = which;
2156 spin_unlock_irq(&img_request->completion_lock);
2159 rbd_img_request_complete(img_request);
2163 * Split up an image request into one or more object requests, each
2164 * to a different object. The "type" parameter indicates whether
2165 * "data_desc" is the pointer to the head of a list of bio
2166 * structures, or the base of a page array. In either case this
2167 * function assumes data_desc describes memory sufficient to hold
2168 * all data described by the image request.
2170 static int rbd_img_request_fill(struct rbd_img_request *img_request,
2171 enum obj_request_type type,
2174 struct rbd_device *rbd_dev = img_request->rbd_dev;
2175 struct rbd_obj_request *obj_request = NULL;
2176 struct rbd_obj_request *next_obj_request;
2177 bool write_request = img_request_write_test(img_request);
2178 struct bio *bio_list = 0;
2179 unsigned int bio_offset = 0;
2180 struct page **pages = 0;
2185 dout("%s: img %p type %d data_desc %p\n", __func__, img_request,
2186 (int)type, data_desc);
2188 opcode = write_request ? CEPH_OSD_OP_WRITE : CEPH_OSD_OP_READ;
2189 img_offset = img_request->offset;
2190 resid = img_request->length;
2191 rbd_assert(resid > 0);
2193 if (type == OBJ_REQUEST_BIO) {
2194 bio_list = data_desc;
2195 rbd_assert(img_offset == bio_list->bi_sector << SECTOR_SHIFT);
2197 rbd_assert(type == OBJ_REQUEST_PAGES);
2202 struct ceph_osd_request *osd_req;
2203 const char *object_name;
2207 object_name = rbd_segment_name(rbd_dev, img_offset);
2210 offset = rbd_segment_offset(rbd_dev, img_offset);
2211 length = rbd_segment_length(rbd_dev, img_offset, resid);
2212 obj_request = rbd_obj_request_create(object_name,
2213 offset, length, type);
2214 /* object request has its own copy of the object name */
2215 rbd_segment_name_free(object_name);
2219 * set obj_request->img_request before creating the
2220 * osd_request so that it gets the right snapc
2222 rbd_img_obj_request_add(img_request, obj_request);
2224 if (type == OBJ_REQUEST_BIO) {
2225 unsigned int clone_size;
2227 rbd_assert(length <= (u64)UINT_MAX);
2228 clone_size = (unsigned int)length;
2229 obj_request->bio_list =
2230 bio_chain_clone_range(&bio_list,
2234 if (!obj_request->bio_list)
2237 unsigned int page_count;
2239 obj_request->pages = pages;
2240 page_count = (u32)calc_pages_for(offset, length);
2241 obj_request->page_count = page_count;
2242 if ((offset + length) & ~PAGE_MASK)
2243 page_count--; /* more on last page */
2244 pages += page_count;
2247 osd_req = rbd_osd_req_create(rbd_dev, write_request,
2251 obj_request->osd_req = osd_req;
2252 obj_request->callback = rbd_img_obj_callback;
2254 osd_req_op_extent_init(osd_req, 0, opcode, offset, length,
2256 if (type == OBJ_REQUEST_BIO)
2257 osd_req_op_extent_osd_data_bio(osd_req, 0,
2258 obj_request->bio_list, length);
2260 osd_req_op_extent_osd_data_pages(osd_req, 0,
2261 obj_request->pages, length,
2262 offset & ~PAGE_MASK, false, false);
2265 rbd_osd_req_format_write(obj_request);
2267 rbd_osd_req_format_read(obj_request);
2269 obj_request->img_offset = img_offset;
2271 img_offset += length;
2278 rbd_obj_request_put(obj_request);
2280 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2281 rbd_obj_request_put(obj_request);
2287 rbd_img_obj_copyup_callback(struct rbd_obj_request *obj_request)
2289 struct rbd_img_request *img_request;
2290 struct rbd_device *rbd_dev;
2291 struct page **pages;
2294 rbd_assert(obj_request->type == OBJ_REQUEST_BIO);
2295 rbd_assert(obj_request_img_data_test(obj_request));
2296 img_request = obj_request->img_request;
2297 rbd_assert(img_request);
2299 rbd_dev = img_request->rbd_dev;
2300 rbd_assert(rbd_dev);
2302 pages = obj_request->copyup_pages;
2303 rbd_assert(pages != NULL);
2304 obj_request->copyup_pages = NULL;
2305 page_count = obj_request->copyup_page_count;
2306 rbd_assert(page_count);
2307 obj_request->copyup_page_count = 0;
2308 ceph_release_page_vector(pages, page_count);
2311 * We want the transfer count to reflect the size of the
2312 * original write request. There is no such thing as a
2313 * successful short write, so if the request was successful
2314 * we can just set it to the originally-requested length.
2316 if (!obj_request->result)
2317 obj_request->xferred = obj_request->length;
2319 /* Finish up with the normal image object callback */
2321 rbd_img_obj_callback(obj_request);
2325 rbd_img_obj_parent_read_full_callback(struct rbd_img_request *img_request)
2327 struct rbd_obj_request *orig_request;
2328 struct ceph_osd_request *osd_req;
2329 struct ceph_osd_client *osdc;
2330 struct rbd_device *rbd_dev;
2331 struct page **pages;
2338 rbd_assert(img_request_child_test(img_request));
2340 /* First get what we need from the image request */
2342 pages = img_request->copyup_pages;
2343 rbd_assert(pages != NULL);
2344 img_request->copyup_pages = NULL;
2345 page_count = img_request->copyup_page_count;
2346 rbd_assert(page_count);
2347 img_request->copyup_page_count = 0;
2349 orig_request = img_request->obj_request;
2350 rbd_assert(orig_request != NULL);
2351 rbd_assert(obj_request_type_valid(orig_request->type));
2352 img_result = img_request->result;
2353 parent_length = img_request->length;
2354 rbd_assert(parent_length == img_request->xferred);
2355 rbd_img_request_put(img_request);
2357 rbd_assert(orig_request->img_request);
2358 rbd_dev = orig_request->img_request->rbd_dev;
2359 rbd_assert(rbd_dev);
2362 * If the overlap has become 0 (most likely because the
2363 * image has been flattened) we need to free the pages
2364 * and re-submit the original write request.
2366 if (!rbd_dev->parent_overlap) {
2367 struct ceph_osd_client *osdc;
2369 ceph_release_page_vector(pages, page_count);
2370 osdc = &rbd_dev->rbd_client->client->osdc;
2371 img_result = rbd_obj_request_submit(osdc, orig_request);
2380 * The original osd request is of no use to use any more.
2381 * We need a new one that can hold the two ops in a copyup
2382 * request. Allocate the new copyup osd request for the
2383 * original request, and release the old one.
2385 img_result = -ENOMEM;
2386 osd_req = rbd_osd_req_create_copyup(orig_request);
2389 rbd_osd_req_destroy(orig_request->osd_req);
2390 orig_request->osd_req = osd_req;
2391 orig_request->copyup_pages = pages;
2392 orig_request->copyup_page_count = page_count;
2394 /* Initialize the copyup op */
2396 osd_req_op_cls_init(osd_req, 0, CEPH_OSD_OP_CALL, "rbd", "copyup");
2397 osd_req_op_cls_request_data_pages(osd_req, 0, pages, parent_length, 0,
2400 /* Then the original write request op */
2402 offset = orig_request->offset;
2403 length = orig_request->length;
2404 osd_req_op_extent_init(osd_req, 1, CEPH_OSD_OP_WRITE,
2405 offset, length, 0, 0);
2406 if (orig_request->type == OBJ_REQUEST_BIO)
2407 osd_req_op_extent_osd_data_bio(osd_req, 1,
2408 orig_request->bio_list, length);
2410 osd_req_op_extent_osd_data_pages(osd_req, 1,
2411 orig_request->pages, length,
2412 offset & ~PAGE_MASK, false, false);
2414 rbd_osd_req_format_write(orig_request);
2416 /* All set, send it off. */
2418 orig_request->callback = rbd_img_obj_copyup_callback;
2419 osdc = &rbd_dev->rbd_client->client->osdc;
2420 img_result = rbd_obj_request_submit(osdc, orig_request);
2424 /* Record the error code and complete the request */
2426 orig_request->result = img_result;
2427 orig_request->xferred = 0;
2428 obj_request_done_set(orig_request);
2429 rbd_obj_request_complete(orig_request);
2433 * Read from the parent image the range of data that covers the
2434 * entire target of the given object request. This is used for
2435 * satisfying a layered image write request when the target of an
2436 * object request from the image request does not exist.
2438 * A page array big enough to hold the returned data is allocated
2439 * and supplied to rbd_img_request_fill() as the "data descriptor."
2440 * When the read completes, this page array will be transferred to
2441 * the original object request for the copyup operation.
2443 * If an error occurs, record it as the result of the original
2444 * object request and mark it done so it gets completed.
2446 static int rbd_img_obj_parent_read_full(struct rbd_obj_request *obj_request)
2448 struct rbd_img_request *img_request = NULL;
2449 struct rbd_img_request *parent_request = NULL;
2450 struct rbd_device *rbd_dev;
2453 struct page **pages = NULL;
2457 rbd_assert(obj_request_img_data_test(obj_request));
2458 rbd_assert(obj_request_type_valid(obj_request->type));
2460 img_request = obj_request->img_request;
2461 rbd_assert(img_request != NULL);
2462 rbd_dev = img_request->rbd_dev;
2463 rbd_assert(rbd_dev->parent != NULL);
2466 * Determine the byte range covered by the object in the
2467 * child image to which the original request was to be sent.
2469 img_offset = obj_request->img_offset - obj_request->offset;
2470 length = (u64)1 << rbd_dev->header.obj_order;
2473 * There is no defined parent data beyond the parent
2474 * overlap, so limit what we read at that boundary if
2477 if (img_offset + length > rbd_dev->parent_overlap) {
2478 rbd_assert(img_offset < rbd_dev->parent_overlap);
2479 length = rbd_dev->parent_overlap - img_offset;
2483 * Allocate a page array big enough to receive the data read
2486 page_count = (u32)calc_pages_for(0, length);
2487 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2488 if (IS_ERR(pages)) {
2489 result = PTR_ERR(pages);
2495 parent_request = rbd_parent_request_create(obj_request,
2496 img_offset, length);
2497 if (!parent_request)
2500 result = rbd_img_request_fill(parent_request, OBJ_REQUEST_PAGES, pages);
2503 parent_request->copyup_pages = pages;
2504 parent_request->copyup_page_count = page_count;
2506 parent_request->callback = rbd_img_obj_parent_read_full_callback;
2507 result = rbd_img_request_submit(parent_request);
2511 parent_request->copyup_pages = NULL;
2512 parent_request->copyup_page_count = 0;
2513 parent_request->obj_request = NULL;
2514 rbd_obj_request_put(obj_request);
2517 ceph_release_page_vector(pages, page_count);
2519 rbd_img_request_put(parent_request);
2520 obj_request->result = result;
2521 obj_request->xferred = 0;
2522 obj_request_done_set(obj_request);
2527 static void rbd_img_obj_exists_callback(struct rbd_obj_request *obj_request)
2529 struct rbd_obj_request *orig_request;
2530 struct rbd_device *rbd_dev;
2533 rbd_assert(!obj_request_img_data_test(obj_request));
2536 * All we need from the object request is the original
2537 * request and the result of the STAT op. Grab those, then
2538 * we're done with the request.
2540 orig_request = obj_request->obj_request;
2541 obj_request->obj_request = NULL;
2542 rbd_assert(orig_request);
2543 rbd_assert(orig_request->img_request);
2545 result = obj_request->result;
2546 obj_request->result = 0;
2548 dout("%s: obj %p for obj %p result %d %llu/%llu\n", __func__,
2549 obj_request, orig_request, result,
2550 obj_request->xferred, obj_request->length);
2551 rbd_obj_request_put(obj_request);
2554 * If the overlap has become 0 (most likely because the
2555 * image has been flattened) we need to free the pages
2556 * and re-submit the original write request.
2558 rbd_dev = orig_request->img_request->rbd_dev;
2559 if (!rbd_dev->parent_overlap) {
2560 struct ceph_osd_client *osdc;
2562 rbd_obj_request_put(orig_request);
2563 osdc = &rbd_dev->rbd_client->client->osdc;
2564 result = rbd_obj_request_submit(osdc, orig_request);
2570 * Our only purpose here is to determine whether the object
2571 * exists, and we don't want to treat the non-existence as
2572 * an error. If something else comes back, transfer the
2573 * error to the original request and complete it now.
2576 obj_request_existence_set(orig_request, true);
2577 } else if (result == -ENOENT) {
2578 obj_request_existence_set(orig_request, false);
2579 } else if (result) {
2580 orig_request->result = result;
2585 * Resubmit the original request now that we have recorded
2586 * whether the target object exists.
2588 orig_request->result = rbd_img_obj_request_submit(orig_request);
2590 if (orig_request->result)
2591 rbd_obj_request_complete(orig_request);
2592 rbd_obj_request_put(orig_request);
2595 static int rbd_img_obj_exists_submit(struct rbd_obj_request *obj_request)
2597 struct rbd_obj_request *stat_request;
2598 struct rbd_device *rbd_dev;
2599 struct ceph_osd_client *osdc;
2600 struct page **pages = NULL;
2606 * The response data for a STAT call consists of:
2613 size = sizeof (__le64) + sizeof (__le32) + sizeof (__le32);
2614 page_count = (u32)calc_pages_for(0, size);
2615 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2617 return PTR_ERR(pages);
2620 stat_request = rbd_obj_request_create(obj_request->object_name, 0, 0,
2625 rbd_obj_request_get(obj_request);
2626 stat_request->obj_request = obj_request;
2627 stat_request->pages = pages;
2628 stat_request->page_count = page_count;
2630 rbd_assert(obj_request->img_request);
2631 rbd_dev = obj_request->img_request->rbd_dev;
2632 stat_request->osd_req = rbd_osd_req_create(rbd_dev, false,
2634 if (!stat_request->osd_req)
2636 stat_request->callback = rbd_img_obj_exists_callback;
2638 osd_req_op_init(stat_request->osd_req, 0, CEPH_OSD_OP_STAT);
2639 osd_req_op_raw_data_in_pages(stat_request->osd_req, 0, pages, size, 0,
2641 rbd_osd_req_format_read(stat_request);
2643 osdc = &rbd_dev->rbd_client->client->osdc;
2644 ret = rbd_obj_request_submit(osdc, stat_request);
2647 rbd_obj_request_put(obj_request);
2652 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request)
2654 struct rbd_img_request *img_request;
2655 struct rbd_device *rbd_dev;
2658 rbd_assert(obj_request_img_data_test(obj_request));
2660 img_request = obj_request->img_request;
2661 rbd_assert(img_request);
2662 rbd_dev = img_request->rbd_dev;
2665 * Only writes to layered images need special handling.
2666 * Reads and non-layered writes are simple object requests.
2667 * Layered writes that start beyond the end of the overlap
2668 * with the parent have no parent data, so they too are
2669 * simple object requests. Finally, if the target object is
2670 * known to already exist, its parent data has already been
2671 * copied, so a write to the object can also be handled as a
2672 * simple object request.
2674 if (!img_request_write_test(img_request) ||
2675 !img_request_layered_test(img_request) ||
2676 rbd_dev->parent_overlap <= obj_request->img_offset ||
2677 ((known = obj_request_known_test(obj_request)) &&
2678 obj_request_exists_test(obj_request))) {
2680 struct rbd_device *rbd_dev;
2681 struct ceph_osd_client *osdc;
2683 rbd_dev = obj_request->img_request->rbd_dev;
2684 osdc = &rbd_dev->rbd_client->client->osdc;
2686 return rbd_obj_request_submit(osdc, obj_request);
2690 * It's a layered write. The target object might exist but
2691 * we may not know that yet. If we know it doesn't exist,
2692 * start by reading the data for the full target object from
2693 * the parent so we can use it for a copyup to the target.
2696 return rbd_img_obj_parent_read_full(obj_request);
2698 /* We don't know whether the target exists. Go find out. */
2700 return rbd_img_obj_exists_submit(obj_request);
2703 static int rbd_img_request_submit(struct rbd_img_request *img_request)
2705 struct rbd_obj_request *obj_request;
2706 struct rbd_obj_request *next_obj_request;
2708 dout("%s: img %p\n", __func__, img_request);
2709 for_each_obj_request_safe(img_request, obj_request, next_obj_request) {
2712 ret = rbd_img_obj_request_submit(obj_request);
2720 static void rbd_img_parent_read_callback(struct rbd_img_request *img_request)
2722 struct rbd_obj_request *obj_request;
2723 struct rbd_device *rbd_dev;
2728 rbd_assert(img_request_child_test(img_request));
2730 /* First get what we need from the image request and release it */
2732 obj_request = img_request->obj_request;
2733 img_xferred = img_request->xferred;
2734 img_result = img_request->result;
2735 rbd_img_request_put(img_request);
2738 * If the overlap has become 0 (most likely because the
2739 * image has been flattened) we need to re-submit the
2742 rbd_assert(obj_request);
2743 rbd_assert(obj_request->img_request);
2744 rbd_dev = obj_request->img_request->rbd_dev;
2745 if (!rbd_dev->parent_overlap) {
2746 struct ceph_osd_client *osdc;
2748 osdc = &rbd_dev->rbd_client->client->osdc;
2749 img_result = rbd_obj_request_submit(osdc, obj_request);
2754 obj_request->result = img_result;
2755 if (obj_request->result)
2759 * We need to zero anything beyond the parent overlap
2760 * boundary. Since rbd_img_obj_request_read_callback()
2761 * will zero anything beyond the end of a short read, an
2762 * easy way to do this is to pretend the data from the
2763 * parent came up short--ending at the overlap boundary.
2765 rbd_assert(obj_request->img_offset < U64_MAX - obj_request->length);
2766 obj_end = obj_request->img_offset + obj_request->length;
2767 if (obj_end > rbd_dev->parent_overlap) {
2770 if (obj_request->img_offset < rbd_dev->parent_overlap)
2771 xferred = rbd_dev->parent_overlap -
2772 obj_request->img_offset;
2774 obj_request->xferred = min(img_xferred, xferred);
2776 obj_request->xferred = img_xferred;
2779 rbd_img_obj_request_read_callback(obj_request);
2780 rbd_obj_request_complete(obj_request);
2783 static void rbd_img_parent_read(struct rbd_obj_request *obj_request)
2785 struct rbd_img_request *img_request;
2788 rbd_assert(obj_request_img_data_test(obj_request));
2789 rbd_assert(obj_request->img_request != NULL);
2790 rbd_assert(obj_request->result == (s32) -ENOENT);
2791 rbd_assert(obj_request_type_valid(obj_request->type));
2793 /* rbd_read_finish(obj_request, obj_request->length); */
2794 img_request = rbd_parent_request_create(obj_request,
2795 obj_request->img_offset,
2796 obj_request->length);
2801 if (obj_request->type == OBJ_REQUEST_BIO)
2802 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
2803 obj_request->bio_list);
2805 result = rbd_img_request_fill(img_request, OBJ_REQUEST_PAGES,
2806 obj_request->pages);
2810 img_request->callback = rbd_img_parent_read_callback;
2811 result = rbd_img_request_submit(img_request);
2818 rbd_img_request_put(img_request);
2819 obj_request->result = result;
2820 obj_request->xferred = 0;
2821 obj_request_done_set(obj_request);
2824 static int rbd_obj_notify_ack_sync(struct rbd_device *rbd_dev, u64 notify_id)
2826 struct rbd_obj_request *obj_request;
2827 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2830 obj_request = rbd_obj_request_create(rbd_dev->header_name, 0, 0,
2831 OBJ_REQUEST_NODATA);
2836 obj_request->osd_req = rbd_osd_req_create(rbd_dev, false, obj_request);
2837 if (!obj_request->osd_req)
2840 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_NOTIFY_ACK,
2842 rbd_osd_req_format_read(obj_request);
2844 ret = rbd_obj_request_submit(osdc, obj_request);
2847 ret = rbd_obj_request_wait(obj_request);
2849 rbd_obj_request_put(obj_request);
2854 static void rbd_watch_cb(u64 ver, u64 notify_id, u8 opcode, void *data)
2856 struct rbd_device *rbd_dev = (struct rbd_device *)data;
2862 dout("%s: \"%s\" notify_id %llu opcode %u\n", __func__,
2863 rbd_dev->header_name, (unsigned long long)notify_id,
2864 (unsigned int)opcode);
2865 ret = rbd_dev_refresh(rbd_dev);
2867 rbd_warn(rbd_dev, ": header refresh error (%d)\n", ret);
2869 rbd_obj_notify_ack_sync(rbd_dev, notify_id);
2873 * Request sync osd watch/unwatch. The value of "start" determines
2874 * whether a watch request is being initiated or torn down.
2876 static int rbd_dev_header_watch_sync(struct rbd_device *rbd_dev, bool start)
2878 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2879 struct rbd_obj_request *obj_request;
2882 rbd_assert(start ^ !!rbd_dev->watch_event);
2883 rbd_assert(start ^ !!rbd_dev->watch_request);
2886 ret = ceph_osdc_create_event(osdc, rbd_watch_cb, rbd_dev,
2887 &rbd_dev->watch_event);
2890 rbd_assert(rbd_dev->watch_event != NULL);
2894 obj_request = rbd_obj_request_create(rbd_dev->header_name, 0, 0,
2895 OBJ_REQUEST_NODATA);
2899 obj_request->osd_req = rbd_osd_req_create(rbd_dev, true, obj_request);
2900 if (!obj_request->osd_req)
2904 ceph_osdc_set_request_linger(osdc, obj_request->osd_req);
2906 ceph_osdc_unregister_linger_request(osdc,
2907 rbd_dev->watch_request->osd_req);
2909 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_WATCH,
2910 rbd_dev->watch_event->cookie, 0, start ? 1 : 0);
2911 rbd_osd_req_format_write(obj_request);
2913 ret = rbd_obj_request_submit(osdc, obj_request);
2916 ret = rbd_obj_request_wait(obj_request);
2919 ret = obj_request->result;
2924 * A watch request is set to linger, so the underlying osd
2925 * request won't go away until we unregister it. We retain
2926 * a pointer to the object request during that time (in
2927 * rbd_dev->watch_request), so we'll keep a reference to
2928 * it. We'll drop that reference (below) after we've
2932 rbd_dev->watch_request = obj_request;
2937 /* We have successfully torn down the watch request */
2939 rbd_obj_request_put(rbd_dev->watch_request);
2940 rbd_dev->watch_request = NULL;
2942 /* Cancel the event if we're tearing down, or on error */
2943 ceph_osdc_cancel_event(rbd_dev->watch_event);
2944 rbd_dev->watch_event = NULL;
2946 rbd_obj_request_put(obj_request);
2952 * Synchronous osd object method call. Returns the number of bytes
2953 * returned in the outbound buffer, or a negative error code.
2955 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
2956 const char *object_name,
2957 const char *class_name,
2958 const char *method_name,
2959 const void *outbound,
2960 size_t outbound_size,
2962 size_t inbound_size)
2964 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2965 struct rbd_obj_request *obj_request;
2966 struct page **pages;
2971 * Method calls are ultimately read operations. The result
2972 * should placed into the inbound buffer provided. They
2973 * also supply outbound data--parameters for the object
2974 * method. Currently if this is present it will be a
2977 page_count = (u32)calc_pages_for(0, inbound_size);
2978 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2980 return PTR_ERR(pages);
2983 obj_request = rbd_obj_request_create(object_name, 0, inbound_size,
2988 obj_request->pages = pages;
2989 obj_request->page_count = page_count;
2991 obj_request->osd_req = rbd_osd_req_create(rbd_dev, false, obj_request);
2992 if (!obj_request->osd_req)
2995 osd_req_op_cls_init(obj_request->osd_req, 0, CEPH_OSD_OP_CALL,
2996 class_name, method_name);
2997 if (outbound_size) {
2998 struct ceph_pagelist *pagelist;
3000 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
3004 ceph_pagelist_init(pagelist);
3005 ceph_pagelist_append(pagelist, outbound, outbound_size);
3006 osd_req_op_cls_request_data_pagelist(obj_request->osd_req, 0,
3009 osd_req_op_cls_response_data_pages(obj_request->osd_req, 0,
3010 obj_request->pages, inbound_size,
3012 rbd_osd_req_format_read(obj_request);
3014 ret = rbd_obj_request_submit(osdc, obj_request);
3017 ret = rbd_obj_request_wait(obj_request);
3021 ret = obj_request->result;
3025 rbd_assert(obj_request->xferred < (u64)INT_MAX);
3026 ret = (int)obj_request->xferred;
3027 ceph_copy_from_page_vector(pages, inbound, 0, obj_request->xferred);
3030 rbd_obj_request_put(obj_request);
3032 ceph_release_page_vector(pages, page_count);
3037 static void rbd_request_fn(struct request_queue *q)
3038 __releases(q->queue_lock) __acquires(q->queue_lock)
3040 struct rbd_device *rbd_dev = q->queuedata;
3041 bool read_only = rbd_dev->mapping.read_only;
3045 while ((rq = blk_fetch_request(q))) {
3046 bool write_request = rq_data_dir(rq) == WRITE;
3047 struct rbd_img_request *img_request;
3051 /* Ignore any non-FS requests that filter through. */
3053 if (rq->cmd_type != REQ_TYPE_FS) {
3054 dout("%s: non-fs request type %d\n", __func__,
3055 (int) rq->cmd_type);
3056 __blk_end_request_all(rq, 0);
3060 /* Ignore/skip any zero-length requests */
3062 offset = (u64) blk_rq_pos(rq) << SECTOR_SHIFT;
3063 length = (u64) blk_rq_bytes(rq);
3066 dout("%s: zero-length request\n", __func__);
3067 __blk_end_request_all(rq, 0);
3071 spin_unlock_irq(q->queue_lock);
3073 /* Disallow writes to a read-only device */
3075 if (write_request) {
3079 rbd_assert(rbd_dev->spec->snap_id == CEPH_NOSNAP);
3083 * Quit early if the mapped snapshot no longer
3084 * exists. It's still possible the snapshot will
3085 * have disappeared by the time our request arrives
3086 * at the osd, but there's no sense in sending it if
3089 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
3090 dout("request for non-existent snapshot");
3091 rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
3097 if (offset && length > U64_MAX - offset + 1) {
3098 rbd_warn(rbd_dev, "bad request range (%llu~%llu)\n",
3100 goto end_request; /* Shouldn't happen */
3104 if (offset + length > rbd_dev->mapping.size) {
3105 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)\n",
3106 offset, length, rbd_dev->mapping.size);
3111 img_request = rbd_img_request_create(rbd_dev, offset, length,
3116 img_request->rq = rq;
3118 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3121 result = rbd_img_request_submit(img_request);
3123 rbd_img_request_put(img_request);
3125 spin_lock_irq(q->queue_lock);
3127 rbd_warn(rbd_dev, "%s %llx at %llx result %d\n",
3128 write_request ? "write" : "read",
3129 length, offset, result);
3131 __blk_end_request_all(rq, result);
3137 * a queue callback. Makes sure that we don't create a bio that spans across
3138 * multiple osd objects. One exception would be with a single page bios,
3139 * which we handle later at bio_chain_clone_range()
3141 static int rbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bmd,
3142 struct bio_vec *bvec)
3144 struct rbd_device *rbd_dev = q->queuedata;
3145 sector_t sector_offset;
3146 sector_t sectors_per_obj;
3147 sector_t obj_sector_offset;
3151 * Find how far into its rbd object the partition-relative
3152 * bio start sector is to offset relative to the enclosing
3155 sector_offset = get_start_sect(bmd->bi_bdev) + bmd->bi_sector;
3156 sectors_per_obj = 1 << (rbd_dev->header.obj_order - SECTOR_SHIFT);
3157 obj_sector_offset = sector_offset & (sectors_per_obj - 1);
3160 * Compute the number of bytes from that offset to the end
3161 * of the object. Account for what's already used by the bio.
3163 ret = (int) (sectors_per_obj - obj_sector_offset) << SECTOR_SHIFT;
3164 if (ret > bmd->bi_size)
3165 ret -= bmd->bi_size;
3170 * Don't send back more than was asked for. And if the bio
3171 * was empty, let the whole thing through because: "Note
3172 * that a block device *must* allow a single page to be
3173 * added to an empty bio."
3175 rbd_assert(bvec->bv_len <= PAGE_SIZE);
3176 if (ret > (int) bvec->bv_len || !bmd->bi_size)
3177 ret = (int) bvec->bv_len;
3182 static void rbd_free_disk(struct rbd_device *rbd_dev)
3184 struct gendisk *disk = rbd_dev->disk;
3189 rbd_dev->disk = NULL;
3190 if (disk->flags & GENHD_FL_UP) {
3193 blk_cleanup_queue(disk->queue);
3198 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
3199 const char *object_name,
3200 u64 offset, u64 length, void *buf)
3203 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3204 struct rbd_obj_request *obj_request;
3205 struct page **pages = NULL;
3210 page_count = (u32) calc_pages_for(offset, length);
3211 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
3213 ret = PTR_ERR(pages);
3216 obj_request = rbd_obj_request_create(object_name, offset, length,
3221 obj_request->pages = pages;
3222 obj_request->page_count = page_count;
3224 obj_request->osd_req = rbd_osd_req_create(rbd_dev, false, obj_request);
3225 if (!obj_request->osd_req)
3228 osd_req_op_extent_init(obj_request->osd_req, 0, CEPH_OSD_OP_READ,
3229 offset, length, 0, 0);
3230 osd_req_op_extent_osd_data_pages(obj_request->osd_req, 0,
3232 obj_request->length,
3233 obj_request->offset & ~PAGE_MASK,
3235 rbd_osd_req_format_read(obj_request);
3237 ret = rbd_obj_request_submit(osdc, obj_request);
3240 ret = rbd_obj_request_wait(obj_request);
3244 ret = obj_request->result;
3248 rbd_assert(obj_request->xferred <= (u64) SIZE_MAX);
3249 size = (size_t) obj_request->xferred;
3250 ceph_copy_from_page_vector(pages, buf, 0, size);
3251 rbd_assert(size <= (size_t)INT_MAX);
3255 rbd_obj_request_put(obj_request);
3257 ceph_release_page_vector(pages, page_count);
3263 * Read the complete header for the given rbd device. On successful
3264 * return, the rbd_dev->header field will contain up-to-date
3265 * information about the image.
3267 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev)
3269 struct rbd_image_header_ondisk *ondisk = NULL;
3276 * The complete header will include an array of its 64-bit
3277 * snapshot ids, followed by the names of those snapshots as
3278 * a contiguous block of NUL-terminated strings. Note that
3279 * the number of snapshots could change by the time we read
3280 * it in, in which case we re-read it.
3287 size = sizeof (*ondisk);
3288 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
3290 ondisk = kmalloc(size, GFP_KERNEL);
3294 ret = rbd_obj_read_sync(rbd_dev, rbd_dev->header_name,
3298 if ((size_t)ret < size) {
3300 rbd_warn(rbd_dev, "short header read (want %zd got %d)",
3304 if (!rbd_dev_ondisk_valid(ondisk)) {
3306 rbd_warn(rbd_dev, "invalid header");
3310 names_size = le64_to_cpu(ondisk->snap_names_len);
3311 want_count = snap_count;
3312 snap_count = le32_to_cpu(ondisk->snap_count);
3313 } while (snap_count != want_count);
3315 ret = rbd_header_from_disk(rbd_dev, ondisk);
3323 * Clear the rbd device's EXISTS flag if the snapshot it's mapped to
3324 * has disappeared from the (just updated) snapshot context.
3326 static void rbd_exists_validate(struct rbd_device *rbd_dev)
3330 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags))
3333 snap_id = rbd_dev->spec->snap_id;
3334 if (snap_id == CEPH_NOSNAP)
3337 if (rbd_dev_snap_index(rbd_dev, snap_id) == BAD_SNAP_INDEX)
3338 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
3341 static void rbd_dev_update_size(struct rbd_device *rbd_dev)
3347 * Don't hold the lock while doing disk operations,
3348 * or lock ordering will conflict with the bdev mutex via:
3349 * rbd_add() -> blkdev_get() -> rbd_open()
3351 spin_lock_irq(&rbd_dev->lock);
3352 removing = test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags);
3353 spin_unlock_irq(&rbd_dev->lock);
3355 * If the device is being removed, rbd_dev->disk has
3356 * been destroyed, so don't try to update its size
3359 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
3360 dout("setting size to %llu sectors", (unsigned long long)size);
3361 set_capacity(rbd_dev->disk, size);
3362 revalidate_disk(rbd_dev->disk);
3366 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
3371 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
3372 mapping_size = rbd_dev->mapping.size;
3373 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
3374 if (rbd_dev->image_format == 1)
3375 ret = rbd_dev_v1_header_info(rbd_dev);
3377 ret = rbd_dev_v2_header_info(rbd_dev);
3379 /* If it's a mapped snapshot, validate its EXISTS flag */
3381 rbd_exists_validate(rbd_dev);
3382 mutex_unlock(&ctl_mutex);
3383 if (mapping_size != rbd_dev->mapping.size) {
3384 rbd_dev_update_size(rbd_dev);
3390 static int rbd_init_disk(struct rbd_device *rbd_dev)
3392 struct gendisk *disk;
3393 struct request_queue *q;
3396 /* create gendisk info */
3397 disk = alloc_disk(RBD_MINORS_PER_MAJOR);
3401 snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
3403 disk->major = rbd_dev->major;
3404 disk->first_minor = 0;
3405 disk->fops = &rbd_bd_ops;
3406 disk->private_data = rbd_dev;
3408 q = blk_init_queue(rbd_request_fn, &rbd_dev->lock);
3412 /* We use the default size, but let's be explicit about it. */
3413 blk_queue_physical_block_size(q, SECTOR_SIZE);
3415 /* set io sizes to object size */
3416 segment_size = rbd_obj_bytes(&rbd_dev->header);
3417 blk_queue_max_hw_sectors(q, segment_size / SECTOR_SIZE);
3418 blk_queue_max_segment_size(q, segment_size);
3419 blk_queue_io_min(q, segment_size);
3420 blk_queue_io_opt(q, segment_size);
3422 blk_queue_merge_bvec(q, rbd_merge_bvec);
3425 q->queuedata = rbd_dev;
3427 rbd_dev->disk = disk;
3440 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
3442 return container_of(dev, struct rbd_device, dev);
3445 static ssize_t rbd_size_show(struct device *dev,
3446 struct device_attribute *attr, char *buf)
3448 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3450 return sprintf(buf, "%llu\n",
3451 (unsigned long long)rbd_dev->mapping.size);
3455 * Note this shows the features for whatever's mapped, which is not
3456 * necessarily the base image.
3458 static ssize_t rbd_features_show(struct device *dev,
3459 struct device_attribute *attr, char *buf)
3461 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3463 return sprintf(buf, "0x%016llx\n",
3464 (unsigned long long)rbd_dev->mapping.features);
3467 static ssize_t rbd_major_show(struct device *dev,
3468 struct device_attribute *attr, char *buf)
3470 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3473 return sprintf(buf, "%d\n", rbd_dev->major);
3475 return sprintf(buf, "(none)\n");
3479 static ssize_t rbd_client_id_show(struct device *dev,
3480 struct device_attribute *attr, char *buf)
3482 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3484 return sprintf(buf, "client%lld\n",
3485 ceph_client_id(rbd_dev->rbd_client->client));
3488 static ssize_t rbd_pool_show(struct device *dev,
3489 struct device_attribute *attr, char *buf)
3491 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3493 return sprintf(buf, "%s\n", rbd_dev->spec->pool_name);
3496 static ssize_t rbd_pool_id_show(struct device *dev,
3497 struct device_attribute *attr, char *buf)
3499 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3501 return sprintf(buf, "%llu\n",
3502 (unsigned long long) rbd_dev->spec->pool_id);
3505 static ssize_t rbd_name_show(struct device *dev,
3506 struct device_attribute *attr, char *buf)
3508 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3510 if (rbd_dev->spec->image_name)
3511 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
3513 return sprintf(buf, "(unknown)\n");
3516 static ssize_t rbd_image_id_show(struct device *dev,
3517 struct device_attribute *attr, char *buf)
3519 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3521 return sprintf(buf, "%s\n", rbd_dev->spec->image_id);
3525 * Shows the name of the currently-mapped snapshot (or
3526 * RBD_SNAP_HEAD_NAME for the base image).
3528 static ssize_t rbd_snap_show(struct device *dev,
3529 struct device_attribute *attr,
3532 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3534 return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
3538 * For an rbd v2 image, shows the pool id, image id, and snapshot id
3539 * for the parent image. If there is no parent, simply shows
3540 * "(no parent image)".
3542 static ssize_t rbd_parent_show(struct device *dev,
3543 struct device_attribute *attr,
3546 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3547 struct rbd_spec *spec = rbd_dev->parent_spec;
3552 return sprintf(buf, "(no parent image)\n");
3554 count = sprintf(bufp, "pool_id %llu\npool_name %s\n",
3555 (unsigned long long) spec->pool_id, spec->pool_name);
3560 count = sprintf(bufp, "image_id %s\nimage_name %s\n", spec->image_id,
3561 spec->image_name ? spec->image_name : "(unknown)");
3566 count = sprintf(bufp, "snap_id %llu\nsnap_name %s\n",
3567 (unsigned long long) spec->snap_id, spec->snap_name);
3572 count = sprintf(bufp, "overlap %llu\n", rbd_dev->parent_overlap);
3577 return (ssize_t) (bufp - buf);
3580 static ssize_t rbd_image_refresh(struct device *dev,
3581 struct device_attribute *attr,
3585 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3588 ret = rbd_dev_refresh(rbd_dev);
3590 rbd_warn(rbd_dev, ": manual header refresh error (%d)\n", ret);
3592 return ret < 0 ? ret : size;
3595 static DEVICE_ATTR(size, S_IRUGO, rbd_size_show, NULL);
3596 static DEVICE_ATTR(features, S_IRUGO, rbd_features_show, NULL);
3597 static DEVICE_ATTR(major, S_IRUGO, rbd_major_show, NULL);
3598 static DEVICE_ATTR(client_id, S_IRUGO, rbd_client_id_show, NULL);
3599 static DEVICE_ATTR(pool, S_IRUGO, rbd_pool_show, NULL);
3600 static DEVICE_ATTR(pool_id, S_IRUGO, rbd_pool_id_show, NULL);
3601 static DEVICE_ATTR(name, S_IRUGO, rbd_name_show, NULL);
3602 static DEVICE_ATTR(image_id, S_IRUGO, rbd_image_id_show, NULL);
3603 static DEVICE_ATTR(refresh, S_IWUSR, NULL, rbd_image_refresh);
3604 static DEVICE_ATTR(current_snap, S_IRUGO, rbd_snap_show, NULL);
3605 static DEVICE_ATTR(parent, S_IRUGO, rbd_parent_show, NULL);
3607 static struct attribute *rbd_attrs[] = {
3608 &dev_attr_size.attr,
3609 &dev_attr_features.attr,
3610 &dev_attr_major.attr,
3611 &dev_attr_client_id.attr,
3612 &dev_attr_pool.attr,
3613 &dev_attr_pool_id.attr,
3614 &dev_attr_name.attr,
3615 &dev_attr_image_id.attr,
3616 &dev_attr_current_snap.attr,
3617 &dev_attr_parent.attr,
3618 &dev_attr_refresh.attr,
3622 static struct attribute_group rbd_attr_group = {
3626 static const struct attribute_group *rbd_attr_groups[] = {
3631 static void rbd_sysfs_dev_release(struct device *dev)
3635 static struct device_type rbd_device_type = {
3637 .groups = rbd_attr_groups,
3638 .release = rbd_sysfs_dev_release,
3641 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
3643 kref_get(&spec->kref);
3648 static void rbd_spec_free(struct kref *kref);
3649 static void rbd_spec_put(struct rbd_spec *spec)
3652 kref_put(&spec->kref, rbd_spec_free);
3655 static struct rbd_spec *rbd_spec_alloc(void)
3657 struct rbd_spec *spec;
3659 spec = kzalloc(sizeof (*spec), GFP_KERNEL);
3662 kref_init(&spec->kref);
3667 static void rbd_spec_free(struct kref *kref)
3669 struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
3671 kfree(spec->pool_name);
3672 kfree(spec->image_id);
3673 kfree(spec->image_name);
3674 kfree(spec->snap_name);
3678 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
3679 struct rbd_spec *spec)
3681 struct rbd_device *rbd_dev;
3683 rbd_dev = kzalloc(sizeof (*rbd_dev), GFP_KERNEL);
3687 spin_lock_init(&rbd_dev->lock);
3689 atomic_set(&rbd_dev->parent_ref, 0);
3690 INIT_LIST_HEAD(&rbd_dev->node);
3691 init_rwsem(&rbd_dev->header_rwsem);
3693 rbd_dev->spec = spec;
3694 rbd_dev->rbd_client = rbdc;
3696 /* Initialize the layout used for all rbd requests */
3698 rbd_dev->layout.fl_stripe_unit = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
3699 rbd_dev->layout.fl_stripe_count = cpu_to_le32(1);
3700 rbd_dev->layout.fl_object_size = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
3701 rbd_dev->layout.fl_pg_pool = cpu_to_le32((u32) spec->pool_id);
3706 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
3708 rbd_put_client(rbd_dev->rbd_client);
3709 rbd_spec_put(rbd_dev->spec);
3714 * Get the size and object order for an image snapshot, or if
3715 * snap_id is CEPH_NOSNAP, gets this information for the base
3718 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
3719 u8 *order, u64 *snap_size)
3721 __le64 snapid = cpu_to_le64(snap_id);
3726 } __attribute__ ((packed)) size_buf = { 0 };
3728 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3730 &snapid, sizeof (snapid),
3731 &size_buf, sizeof (size_buf));
3732 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3735 if (ret < sizeof (size_buf))
3739 *order = size_buf.order;
3740 dout(" order %u", (unsigned int)*order);
3742 *snap_size = le64_to_cpu(size_buf.size);
3744 dout(" snap_id 0x%016llx snap_size = %llu\n",
3745 (unsigned long long)snap_id,
3746 (unsigned long long)*snap_size);
3751 static int rbd_dev_v2_image_size(struct rbd_device *rbd_dev)
3753 return _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
3754 &rbd_dev->header.obj_order,
3755 &rbd_dev->header.image_size);
3758 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev)
3764 reply_buf = kzalloc(RBD_OBJ_PREFIX_LEN_MAX, GFP_KERNEL);
3768 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3769 "rbd", "get_object_prefix", NULL, 0,
3770 reply_buf, RBD_OBJ_PREFIX_LEN_MAX);
3771 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3776 rbd_dev->header.object_prefix = ceph_extract_encoded_string(&p,
3777 p + ret, NULL, GFP_NOIO);
3780 if (IS_ERR(rbd_dev->header.object_prefix)) {
3781 ret = PTR_ERR(rbd_dev->header.object_prefix);
3782 rbd_dev->header.object_prefix = NULL;
3784 dout(" object_prefix = %s\n", rbd_dev->header.object_prefix);
3792 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
3795 __le64 snapid = cpu_to_le64(snap_id);
3799 } __attribute__ ((packed)) features_buf = { 0 };
3803 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3804 "rbd", "get_features",
3805 &snapid, sizeof (snapid),
3806 &features_buf, sizeof (features_buf));
3807 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3810 if (ret < sizeof (features_buf))
3813 incompat = le64_to_cpu(features_buf.incompat);
3814 if (incompat & ~RBD_FEATURES_SUPPORTED)
3817 *snap_features = le64_to_cpu(features_buf.features);
3819 dout(" snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
3820 (unsigned long long)snap_id,
3821 (unsigned long long)*snap_features,
3822 (unsigned long long)le64_to_cpu(features_buf.incompat));
3827 static int rbd_dev_v2_features(struct rbd_device *rbd_dev)
3829 return _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
3830 &rbd_dev->header.features);
3833 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev)
3835 struct rbd_spec *parent_spec;
3837 void *reply_buf = NULL;
3846 parent_spec = rbd_spec_alloc();
3850 size = sizeof (__le64) + /* pool_id */
3851 sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX + /* image_id */
3852 sizeof (__le64) + /* snap_id */
3853 sizeof (__le64); /* overlap */
3854 reply_buf = kmalloc(size, GFP_KERNEL);
3860 snapid = cpu_to_le64(CEPH_NOSNAP);
3861 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3862 "rbd", "get_parent",
3863 &snapid, sizeof (snapid),
3865 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3870 end = reply_buf + ret;
3872 ceph_decode_64_safe(&p, end, pool_id, out_err);
3873 if (pool_id == CEPH_NOPOOL) {
3875 * Either the parent never existed, or we have
3876 * record of it but the image got flattened so it no
3877 * longer has a parent. When the parent of a
3878 * layered image disappears we immediately set the
3879 * overlap to 0. The effect of this is that all new
3880 * requests will be treated as if the image had no
3883 if (rbd_dev->parent_overlap) {
3884 rbd_dev->parent_overlap = 0;
3886 rbd_dev_parent_put(rbd_dev);
3887 pr_info("%s: clone image has been flattened\n",
3888 rbd_dev->disk->disk_name);
3891 goto out; /* No parent? No problem. */
3894 /* The ceph file layout needs to fit pool id in 32 bits */
3897 if (pool_id > (u64)U32_MAX) {
3898 rbd_warn(NULL, "parent pool id too large (%llu > %u)\n",
3899 (unsigned long long)pool_id, U32_MAX);
3902 parent_spec->pool_id = pool_id;
3904 image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
3905 if (IS_ERR(image_id)) {
3906 ret = PTR_ERR(image_id);
3909 parent_spec->image_id = image_id;
3910 ceph_decode_64_safe(&p, end, parent_spec->snap_id, out_err);
3911 ceph_decode_64_safe(&p, end, overlap, out_err);
3914 rbd_spec_put(rbd_dev->parent_spec);
3915 rbd_dev->parent_spec = parent_spec;
3916 parent_spec = NULL; /* rbd_dev now owns this */
3917 rbd_dev->parent_overlap = overlap;
3919 rbd_warn(rbd_dev, "ignoring parent of clone with overlap 0\n");
3925 rbd_spec_put(parent_spec);
3930 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev)
3934 __le64 stripe_count;
3935 } __attribute__ ((packed)) striping_info_buf = { 0 };
3936 size_t size = sizeof (striping_info_buf);
3943 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3944 "rbd", "get_stripe_unit_count", NULL, 0,
3945 (char *)&striping_info_buf, size);
3946 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3953 * We don't actually support the "fancy striping" feature
3954 * (STRIPINGV2) yet, but if the striping sizes are the
3955 * defaults the behavior is the same as before. So find
3956 * out, and only fail if the image has non-default values.
3959 obj_size = (u64)1 << rbd_dev->header.obj_order;
3960 p = &striping_info_buf;
3961 stripe_unit = ceph_decode_64(&p);
3962 if (stripe_unit != obj_size) {
3963 rbd_warn(rbd_dev, "unsupported stripe unit "
3964 "(got %llu want %llu)",
3965 stripe_unit, obj_size);
3968 stripe_count = ceph_decode_64(&p);
3969 if (stripe_count != 1) {
3970 rbd_warn(rbd_dev, "unsupported stripe count "
3971 "(got %llu want 1)", stripe_count);
3974 rbd_dev->header.stripe_unit = stripe_unit;
3975 rbd_dev->header.stripe_count = stripe_count;
3980 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
3982 size_t image_id_size;
3987 void *reply_buf = NULL;
3989 char *image_name = NULL;
3992 rbd_assert(!rbd_dev->spec->image_name);
3994 len = strlen(rbd_dev->spec->image_id);
3995 image_id_size = sizeof (__le32) + len;
3996 image_id = kmalloc(image_id_size, GFP_KERNEL);
4001 end = image_id + image_id_size;
4002 ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
4004 size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
4005 reply_buf = kmalloc(size, GFP_KERNEL);
4009 ret = rbd_obj_method_sync(rbd_dev, RBD_DIRECTORY,
4010 "rbd", "dir_get_name",
4011 image_id, image_id_size,
4016 end = reply_buf + ret;
4018 image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
4019 if (IS_ERR(image_name))
4022 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
4030 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4032 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4033 const char *snap_name;
4036 /* Skip over names until we find the one we are looking for */
4038 snap_name = rbd_dev->header.snap_names;
4039 while (which < snapc->num_snaps) {
4040 if (!strcmp(name, snap_name))
4041 return snapc->snaps[which];
4042 snap_name += strlen(snap_name) + 1;
4048 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4050 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4055 for (which = 0; !found && which < snapc->num_snaps; which++) {
4056 const char *snap_name;
4058 snap_id = snapc->snaps[which];
4059 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
4060 if (IS_ERR(snap_name)) {
4061 /* ignore no-longer existing snapshots */
4062 if (PTR_ERR(snap_name) == -ENOENT)
4067 found = !strcmp(name, snap_name);
4070 return found ? snap_id : CEPH_NOSNAP;
4074 * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
4075 * no snapshot by that name is found, or if an error occurs.
4077 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4079 if (rbd_dev->image_format == 1)
4080 return rbd_v1_snap_id_by_name(rbd_dev, name);
4082 return rbd_v2_snap_id_by_name(rbd_dev, name);
4086 * When an rbd image has a parent image, it is identified by the
4087 * pool, image, and snapshot ids (not names). This function fills
4088 * in the names for those ids. (It's OK if we can't figure out the
4089 * name for an image id, but the pool and snapshot ids should always
4090 * exist and have names.) All names in an rbd spec are dynamically
4093 * When an image being mapped (not a parent) is probed, we have the
4094 * pool name and pool id, image name and image id, and the snapshot
4095 * name. The only thing we're missing is the snapshot id.
4097 static int rbd_dev_spec_update(struct rbd_device *rbd_dev)
4099 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4100 struct rbd_spec *spec = rbd_dev->spec;
4101 const char *pool_name;
4102 const char *image_name;
4103 const char *snap_name;
4107 * An image being mapped will have the pool name (etc.), but
4108 * we need to look up the snapshot id.
4110 if (spec->pool_name) {
4111 if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
4114 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
4115 if (snap_id == CEPH_NOSNAP)
4117 spec->snap_id = snap_id;
4119 spec->snap_id = CEPH_NOSNAP;
4125 /* Get the pool name; we have to make our own copy of this */
4127 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
4129 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
4132 pool_name = kstrdup(pool_name, GFP_KERNEL);
4136 /* Fetch the image name; tolerate failure here */
4138 image_name = rbd_dev_image_name(rbd_dev);
4140 rbd_warn(rbd_dev, "unable to get image name");
4142 /* Look up the snapshot name, and make a copy */
4144 snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
4145 if (IS_ERR(snap_name)) {
4146 ret = PTR_ERR(snap_name);
4150 spec->pool_name = pool_name;
4151 spec->image_name = image_name;
4152 spec->snap_name = snap_name;
4162 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev)
4171 struct ceph_snap_context *snapc;
4175 * We'll need room for the seq value (maximum snapshot id),
4176 * snapshot count, and array of that many snapshot ids.
4177 * For now we have a fixed upper limit on the number we're
4178 * prepared to receive.
4180 size = sizeof (__le64) + sizeof (__le32) +
4181 RBD_MAX_SNAP_COUNT * sizeof (__le64);
4182 reply_buf = kzalloc(size, GFP_KERNEL);
4186 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4187 "rbd", "get_snapcontext", NULL, 0,
4189 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4194 end = reply_buf + ret;
4196 ceph_decode_64_safe(&p, end, seq, out);
4197 ceph_decode_32_safe(&p, end, snap_count, out);
4200 * Make sure the reported number of snapshot ids wouldn't go
4201 * beyond the end of our buffer. But before checking that,
4202 * make sure the computed size of the snapshot context we
4203 * allocate is representable in a size_t.
4205 if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
4210 if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
4214 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
4220 for (i = 0; i < snap_count; i++)
4221 snapc->snaps[i] = ceph_decode_64(&p);
4223 ceph_put_snap_context(rbd_dev->header.snapc);
4224 rbd_dev->header.snapc = snapc;
4226 dout(" snap context seq = %llu, snap_count = %u\n",
4227 (unsigned long long)seq, (unsigned int)snap_count);
4234 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
4245 size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
4246 reply_buf = kmalloc(size, GFP_KERNEL);
4248 return ERR_PTR(-ENOMEM);
4250 snapid = cpu_to_le64(snap_id);
4251 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4252 "rbd", "get_snapshot_name",
4253 &snapid, sizeof (snapid),
4255 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4257 snap_name = ERR_PTR(ret);
4262 end = reply_buf + ret;
4263 snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
4264 if (IS_ERR(snap_name))
4267 dout(" snap_id 0x%016llx snap_name = %s\n",
4268 (unsigned long long)snap_id, snap_name);
4275 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev)
4277 bool first_time = rbd_dev->header.object_prefix == NULL;
4280 down_write(&rbd_dev->header_rwsem);
4282 ret = rbd_dev_v2_image_size(rbd_dev);
4287 ret = rbd_dev_v2_header_onetime(rbd_dev);
4293 * If the image supports layering, get the parent info. We
4294 * need to probe the first time regardless. Thereafter we
4295 * only need to if there's a parent, to see if it has
4296 * disappeared due to the mapped image getting flattened.
4298 if (rbd_dev->header.features & RBD_FEATURE_LAYERING &&
4299 (first_time || rbd_dev->parent_spec)) {
4302 ret = rbd_dev_v2_parent_info(rbd_dev);
4307 * Print a warning if this is the initial probe and
4308 * the image has a parent. Don't print it if the
4309 * image now being probed is itself a parent. We
4310 * can tell at this point because we won't know its
4311 * pool name yet (just its pool id).
4313 warn = rbd_dev->parent_spec && rbd_dev->spec->pool_name;
4314 if (first_time && warn)
4315 rbd_warn(rbd_dev, "WARNING: kernel layering "
4316 "is EXPERIMENTAL!");
4319 if (rbd_dev->spec->snap_id == CEPH_NOSNAP)
4320 if (rbd_dev->mapping.size != rbd_dev->header.image_size)
4321 rbd_dev->mapping.size = rbd_dev->header.image_size;
4323 ret = rbd_dev_v2_snap_context(rbd_dev);
4324 dout("rbd_dev_v2_snap_context returned %d\n", ret);
4326 up_write(&rbd_dev->header_rwsem);
4331 static int rbd_bus_add_dev(struct rbd_device *rbd_dev)
4336 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
4338 dev = &rbd_dev->dev;
4339 dev->bus = &rbd_bus_type;
4340 dev->type = &rbd_device_type;
4341 dev->parent = &rbd_root_dev;
4342 dev->release = rbd_dev_device_release;
4343 dev_set_name(dev, "%d", rbd_dev->dev_id);
4344 ret = device_register(dev);
4346 mutex_unlock(&ctl_mutex);
4351 static void rbd_bus_del_dev(struct rbd_device *rbd_dev)
4353 device_unregister(&rbd_dev->dev);
4356 static atomic64_t rbd_dev_id_max = ATOMIC64_INIT(0);
4359 * Get a unique rbd identifier for the given new rbd_dev, and add
4360 * the rbd_dev to the global list. The minimum rbd id is 1.
4362 static void rbd_dev_id_get(struct rbd_device *rbd_dev)
4364 rbd_dev->dev_id = atomic64_inc_return(&rbd_dev_id_max);
4366 spin_lock(&rbd_dev_list_lock);
4367 list_add_tail(&rbd_dev->node, &rbd_dev_list);
4368 spin_unlock(&rbd_dev_list_lock);
4369 dout("rbd_dev %p given dev id %llu\n", rbd_dev,
4370 (unsigned long long) rbd_dev->dev_id);
4374 * Remove an rbd_dev from the global list, and record that its
4375 * identifier is no longer in use.
4377 static void rbd_dev_id_put(struct rbd_device *rbd_dev)
4379 struct list_head *tmp;
4380 int rbd_id = rbd_dev->dev_id;
4383 rbd_assert(rbd_id > 0);
4385 dout("rbd_dev %p released dev id %llu\n", rbd_dev,
4386 (unsigned long long) rbd_dev->dev_id);
4387 spin_lock(&rbd_dev_list_lock);
4388 list_del_init(&rbd_dev->node);
4391 * If the id being "put" is not the current maximum, there
4392 * is nothing special we need to do.
4394 if (rbd_id != atomic64_read(&rbd_dev_id_max)) {
4395 spin_unlock(&rbd_dev_list_lock);
4400 * We need to update the current maximum id. Search the
4401 * list to find out what it is. We're more likely to find
4402 * the maximum at the end, so search the list backward.
4405 list_for_each_prev(tmp, &rbd_dev_list) {
4406 struct rbd_device *rbd_dev;
4408 rbd_dev = list_entry(tmp, struct rbd_device, node);
4409 if (rbd_dev->dev_id > max_id)
4410 max_id = rbd_dev->dev_id;
4412 spin_unlock(&rbd_dev_list_lock);
4415 * The max id could have been updated by rbd_dev_id_get(), in
4416 * which case it now accurately reflects the new maximum.
4417 * Be careful not to overwrite the maximum value in that
4420 atomic64_cmpxchg(&rbd_dev_id_max, rbd_id, max_id);
4421 dout(" max dev id has been reset\n");
4425 * Skips over white space at *buf, and updates *buf to point to the
4426 * first found non-space character (if any). Returns the length of
4427 * the token (string of non-white space characters) found. Note
4428 * that *buf must be terminated with '\0'.
4430 static inline size_t next_token(const char **buf)
4433 * These are the characters that produce nonzero for
4434 * isspace() in the "C" and "POSIX" locales.
4436 const char *spaces = " \f\n\r\t\v";
4438 *buf += strspn(*buf, spaces); /* Find start of token */
4440 return strcspn(*buf, spaces); /* Return token length */
4444 * Finds the next token in *buf, and if the provided token buffer is
4445 * big enough, copies the found token into it. The result, if
4446 * copied, is guaranteed to be terminated with '\0'. Note that *buf
4447 * must be terminated with '\0' on entry.
4449 * Returns the length of the token found (not including the '\0').
4450 * Return value will be 0 if no token is found, and it will be >=
4451 * token_size if the token would not fit.
4453 * The *buf pointer will be updated to point beyond the end of the
4454 * found token. Note that this occurs even if the token buffer is
4455 * too small to hold it.
4457 static inline size_t copy_token(const char **buf,
4463 len = next_token(buf);
4464 if (len < token_size) {
4465 memcpy(token, *buf, len);
4466 *(token + len) = '\0';
4474 * Finds the next token in *buf, dynamically allocates a buffer big
4475 * enough to hold a copy of it, and copies the token into the new
4476 * buffer. The copy is guaranteed to be terminated with '\0'. Note
4477 * that a duplicate buffer is created even for a zero-length token.
4479 * Returns a pointer to the newly-allocated duplicate, or a null
4480 * pointer if memory for the duplicate was not available. If
4481 * the lenp argument is a non-null pointer, the length of the token
4482 * (not including the '\0') is returned in *lenp.
4484 * If successful, the *buf pointer will be updated to point beyond
4485 * the end of the found token.
4487 * Note: uses GFP_KERNEL for allocation.
4489 static inline char *dup_token(const char **buf, size_t *lenp)
4494 len = next_token(buf);
4495 dup = kmemdup(*buf, len + 1, GFP_KERNEL);
4498 *(dup + len) = '\0';
4508 * Parse the options provided for an "rbd add" (i.e., rbd image
4509 * mapping) request. These arrive via a write to /sys/bus/rbd/add,
4510 * and the data written is passed here via a NUL-terminated buffer.
4511 * Returns 0 if successful or an error code otherwise.
4513 * The information extracted from these options is recorded in
4514 * the other parameters which return dynamically-allocated
4517 * The address of a pointer that will refer to a ceph options
4518 * structure. Caller must release the returned pointer using
4519 * ceph_destroy_options() when it is no longer needed.
4521 * Address of an rbd options pointer. Fully initialized by
4522 * this function; caller must release with kfree().
4524 * Address of an rbd image specification pointer. Fully
4525 * initialized by this function based on parsed options.
4526 * Caller must release with rbd_spec_put().
4528 * The options passed take this form:
4529 * <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
4532 * A comma-separated list of one or more monitor addresses.
4533 * A monitor address is an ip address, optionally followed
4534 * by a port number (separated by a colon).
4535 * I.e.: ip1[:port1][,ip2[:port2]...]
4537 * A comma-separated list of ceph and/or rbd options.
4539 * The name of the rados pool containing the rbd image.
4541 * The name of the image in that pool to map.
4543 * An optional snapshot id. If provided, the mapping will
4544 * present data from the image at the time that snapshot was
4545 * created. The image head is used if no snapshot id is
4546 * provided. Snapshot mappings are always read-only.
4548 static int rbd_add_parse_args(const char *buf,
4549 struct ceph_options **ceph_opts,
4550 struct rbd_options **opts,
4551 struct rbd_spec **rbd_spec)
4555 const char *mon_addrs;
4557 size_t mon_addrs_size;
4558 struct rbd_spec *spec = NULL;
4559 struct rbd_options *rbd_opts = NULL;
4560 struct ceph_options *copts;
4563 /* The first four tokens are required */
4565 len = next_token(&buf);
4567 rbd_warn(NULL, "no monitor address(es) provided");
4571 mon_addrs_size = len + 1;
4575 options = dup_token(&buf, NULL);
4579 rbd_warn(NULL, "no options provided");
4583 spec = rbd_spec_alloc();
4587 spec->pool_name = dup_token(&buf, NULL);
4588 if (!spec->pool_name)
4590 if (!*spec->pool_name) {
4591 rbd_warn(NULL, "no pool name provided");
4595 spec->image_name = dup_token(&buf, NULL);
4596 if (!spec->image_name)
4598 if (!*spec->image_name) {
4599 rbd_warn(NULL, "no image name provided");
4604 * Snapshot name is optional; default is to use "-"
4605 * (indicating the head/no snapshot).
4607 len = next_token(&buf);
4609 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
4610 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
4611 } else if (len > RBD_MAX_SNAP_NAME_LEN) {
4612 ret = -ENAMETOOLONG;
4615 snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
4618 *(snap_name + len) = '\0';
4619 spec->snap_name = snap_name;
4621 /* Initialize all rbd options to the defaults */
4623 rbd_opts = kzalloc(sizeof (*rbd_opts), GFP_KERNEL);
4627 rbd_opts->read_only = RBD_READ_ONLY_DEFAULT;
4629 copts = ceph_parse_options(options, mon_addrs,
4630 mon_addrs + mon_addrs_size - 1,
4631 parse_rbd_opts_token, rbd_opts);
4632 if (IS_ERR(copts)) {
4633 ret = PTR_ERR(copts);
4654 * An rbd format 2 image has a unique identifier, distinct from the
4655 * name given to it by the user. Internally, that identifier is
4656 * what's used to specify the names of objects related to the image.
4658 * A special "rbd id" object is used to map an rbd image name to its
4659 * id. If that object doesn't exist, then there is no v2 rbd image
4660 * with the supplied name.
4662 * This function will record the given rbd_dev's image_id field if
4663 * it can be determined, and in that case will return 0. If any
4664 * errors occur a negative errno will be returned and the rbd_dev's
4665 * image_id field will be unchanged (and should be NULL).
4667 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
4676 * When probing a parent image, the image id is already
4677 * known (and the image name likely is not). There's no
4678 * need to fetch the image id again in this case. We
4679 * do still need to set the image format though.
4681 if (rbd_dev->spec->image_id) {
4682 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
4688 * First, see if the format 2 image id file exists, and if
4689 * so, get the image's persistent id from it.
4691 size = sizeof (RBD_ID_PREFIX) + strlen(rbd_dev->spec->image_name);
4692 object_name = kmalloc(size, GFP_NOIO);
4695 sprintf(object_name, "%s%s", RBD_ID_PREFIX, rbd_dev->spec->image_name);
4696 dout("rbd id object name is %s\n", object_name);
4698 /* Response will be an encoded string, which includes a length */
4700 size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
4701 response = kzalloc(size, GFP_NOIO);
4707 /* If it doesn't exist we'll assume it's a format 1 image */
4709 ret = rbd_obj_method_sync(rbd_dev, object_name,
4710 "rbd", "get_id", NULL, 0,
4711 response, RBD_IMAGE_ID_LEN_MAX);
4712 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4713 if (ret == -ENOENT) {
4714 image_id = kstrdup("", GFP_KERNEL);
4715 ret = image_id ? 0 : -ENOMEM;
4717 rbd_dev->image_format = 1;
4718 } else if (ret > sizeof (__le32)) {
4721 image_id = ceph_extract_encoded_string(&p, p + ret,
4723 ret = IS_ERR(image_id) ? PTR_ERR(image_id) : 0;
4725 rbd_dev->image_format = 2;
4731 rbd_dev->spec->image_id = image_id;
4732 dout("image_id is %s\n", image_id);
4742 * Undo whatever state changes are made by v1 or v2 header info
4745 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
4747 struct rbd_image_header *header;
4749 /* Drop parent reference unless it's already been done (or none) */
4751 if (rbd_dev->parent_overlap)
4752 rbd_dev_parent_put(rbd_dev);
4754 /* Free dynamic fields from the header, then zero it out */
4756 header = &rbd_dev->header;
4757 ceph_put_snap_context(header->snapc);
4758 kfree(header->snap_sizes);
4759 kfree(header->snap_names);
4760 kfree(header->object_prefix);
4761 memset(header, 0, sizeof (*header));
4764 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev)
4768 ret = rbd_dev_v2_object_prefix(rbd_dev);
4773 * Get the and check features for the image. Currently the
4774 * features are assumed to never change.
4776 ret = rbd_dev_v2_features(rbd_dev);
4780 /* If the image supports fancy striping, get its parameters */
4782 if (rbd_dev->header.features & RBD_FEATURE_STRIPINGV2) {
4783 ret = rbd_dev_v2_striping_info(rbd_dev);
4787 /* No support for crypto and compression type format 2 images */
4791 rbd_dev->header.features = 0;
4792 kfree(rbd_dev->header.object_prefix);
4793 rbd_dev->header.object_prefix = NULL;
4798 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev)
4800 struct rbd_device *parent = NULL;
4801 struct rbd_spec *parent_spec;
4802 struct rbd_client *rbdc;
4805 if (!rbd_dev->parent_spec)
4808 * We need to pass a reference to the client and the parent
4809 * spec when creating the parent rbd_dev. Images related by
4810 * parent/child relationships always share both.
4812 parent_spec = rbd_spec_get(rbd_dev->parent_spec);
4813 rbdc = __rbd_get_client(rbd_dev->rbd_client);
4816 parent = rbd_dev_create(rbdc, parent_spec);
4820 ret = rbd_dev_image_probe(parent, false);
4823 rbd_dev->parent = parent;
4824 atomic_set(&rbd_dev->parent_ref, 1);
4829 rbd_dev_unparent(rbd_dev);
4830 kfree(rbd_dev->header_name);
4831 rbd_dev_destroy(parent);
4833 rbd_put_client(rbdc);
4834 rbd_spec_put(parent_spec);
4840 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
4844 /* generate unique id: find highest unique id, add one */
4845 rbd_dev_id_get(rbd_dev);
4847 /* Fill in the device name, now that we have its id. */
4848 BUILD_BUG_ON(DEV_NAME_LEN
4849 < sizeof (RBD_DRV_NAME) + MAX_INT_FORMAT_WIDTH);
4850 sprintf(rbd_dev->name, "%s%d", RBD_DRV_NAME, rbd_dev->dev_id);
4852 /* Get our block major device number. */
4854 ret = register_blkdev(0, rbd_dev->name);
4857 rbd_dev->major = ret;
4859 /* Set up the blkdev mapping. */
4861 ret = rbd_init_disk(rbd_dev);
4863 goto err_out_blkdev;
4865 ret = rbd_dev_mapping_set(rbd_dev);
4868 set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
4870 ret = rbd_bus_add_dev(rbd_dev);
4872 goto err_out_mapping;
4874 /* Everything's ready. Announce the disk to the world. */
4876 set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
4877 add_disk(rbd_dev->disk);
4879 pr_info("%s: added with size 0x%llx\n", rbd_dev->disk->disk_name,
4880 (unsigned long long) rbd_dev->mapping.size);
4885 rbd_dev_mapping_clear(rbd_dev);
4887 rbd_free_disk(rbd_dev);
4889 unregister_blkdev(rbd_dev->major, rbd_dev->name);
4891 rbd_dev_id_put(rbd_dev);
4892 rbd_dev_mapping_clear(rbd_dev);
4897 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
4899 struct rbd_spec *spec = rbd_dev->spec;
4902 /* Record the header object name for this rbd image. */
4904 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
4906 if (rbd_dev->image_format == 1)
4907 size = strlen(spec->image_name) + sizeof (RBD_SUFFIX);
4909 size = sizeof (RBD_HEADER_PREFIX) + strlen(spec->image_id);
4911 rbd_dev->header_name = kmalloc(size, GFP_KERNEL);
4912 if (!rbd_dev->header_name)
4915 if (rbd_dev->image_format == 1)
4916 sprintf(rbd_dev->header_name, "%s%s",
4917 spec->image_name, RBD_SUFFIX);
4919 sprintf(rbd_dev->header_name, "%s%s",
4920 RBD_HEADER_PREFIX, spec->image_id);
4924 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
4926 rbd_dev_unprobe(rbd_dev);
4927 kfree(rbd_dev->header_name);
4928 rbd_dev->header_name = NULL;
4929 rbd_dev->image_format = 0;
4930 kfree(rbd_dev->spec->image_id);
4931 rbd_dev->spec->image_id = NULL;
4933 rbd_dev_destroy(rbd_dev);
4937 * Probe for the existence of the header object for the given rbd
4938 * device. If this image is the one being mapped (i.e., not a
4939 * parent), initiate a watch on its header object before using that
4940 * object to get detailed information about the rbd image.
4942 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, bool mapping)
4948 * Get the id from the image id object. Unless there's an
4949 * error, rbd_dev->spec->image_id will be filled in with
4950 * a dynamically-allocated string, and rbd_dev->image_format
4951 * will be set to either 1 or 2.
4953 ret = rbd_dev_image_id(rbd_dev);
4956 rbd_assert(rbd_dev->spec->image_id);
4957 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
4959 ret = rbd_dev_header_name(rbd_dev);
4961 goto err_out_format;
4964 ret = rbd_dev_header_watch_sync(rbd_dev, true);
4966 goto out_header_name;
4969 if (rbd_dev->image_format == 1)
4970 ret = rbd_dev_v1_header_info(rbd_dev);
4972 ret = rbd_dev_v2_header_info(rbd_dev);
4976 ret = rbd_dev_spec_update(rbd_dev);
4980 ret = rbd_dev_probe_parent(rbd_dev);
4984 dout("discovered format %u image, header name is %s\n",
4985 rbd_dev->image_format, rbd_dev->header_name);
4989 rbd_dev_unprobe(rbd_dev);
4992 tmp = rbd_dev_header_watch_sync(rbd_dev, false);
4994 rbd_warn(rbd_dev, "unable to tear down "
4995 "watch request (%d)\n", tmp);
4998 kfree(rbd_dev->header_name);
4999 rbd_dev->header_name = NULL;
5001 rbd_dev->image_format = 0;
5002 kfree(rbd_dev->spec->image_id);
5003 rbd_dev->spec->image_id = NULL;
5005 dout("probe failed, returning %d\n", ret);
5010 static ssize_t rbd_add(struct bus_type *bus,
5014 struct rbd_device *rbd_dev = NULL;
5015 struct ceph_options *ceph_opts = NULL;
5016 struct rbd_options *rbd_opts = NULL;
5017 struct rbd_spec *spec = NULL;
5018 struct rbd_client *rbdc;
5019 struct ceph_osd_client *osdc;
5023 if (!try_module_get(THIS_MODULE))
5026 /* parse add command */
5027 rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
5029 goto err_out_module;
5030 read_only = rbd_opts->read_only;
5032 rbd_opts = NULL; /* done with this */
5034 rbdc = rbd_get_client(ceph_opts);
5041 osdc = &rbdc->client->osdc;
5042 rc = ceph_pg_poolid_by_name(osdc->osdmap, spec->pool_name);
5044 goto err_out_client;
5045 spec->pool_id = (u64)rc;
5047 /* The ceph file layout needs to fit pool id in 32 bits */
5049 if (spec->pool_id > (u64)U32_MAX) {
5050 rbd_warn(NULL, "pool id too large (%llu > %u)\n",
5051 (unsigned long long)spec->pool_id, U32_MAX);
5053 goto err_out_client;
5056 rbd_dev = rbd_dev_create(rbdc, spec);
5058 goto err_out_client;
5059 rbdc = NULL; /* rbd_dev now owns this */
5060 spec = NULL; /* rbd_dev now owns this */
5062 rc = rbd_dev_image_probe(rbd_dev, true);
5064 goto err_out_rbd_dev;
5066 /* If we are mapping a snapshot it must be marked read-only */
5068 if (rbd_dev->spec->snap_id != CEPH_NOSNAP)
5070 rbd_dev->mapping.read_only = read_only;
5072 rc = rbd_dev_device_setup(rbd_dev);
5074 rbd_dev_image_release(rbd_dev);
5075 goto err_out_module;
5081 rbd_dev_destroy(rbd_dev);
5083 rbd_put_client(rbdc);
5087 module_put(THIS_MODULE);
5089 dout("Error adding device %s\n", buf);
5094 static void rbd_dev_device_release(struct device *dev)
5096 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5098 rbd_free_disk(rbd_dev);
5099 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5100 rbd_dev_mapping_clear(rbd_dev);
5101 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5103 rbd_dev_id_put(rbd_dev);
5104 rbd_dev_mapping_clear(rbd_dev);
5107 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
5109 while (rbd_dev->parent) {
5110 struct rbd_device *first = rbd_dev;
5111 struct rbd_device *second = first->parent;
5112 struct rbd_device *third;
5115 * Follow to the parent with no grandparent and
5118 while (second && (third = second->parent)) {
5123 rbd_dev_image_release(second);
5124 first->parent = NULL;
5125 first->parent_overlap = 0;
5127 rbd_assert(first->parent_spec);
5128 rbd_spec_put(first->parent_spec);
5129 first->parent_spec = NULL;
5133 static ssize_t rbd_remove(struct bus_type *bus,
5137 struct rbd_device *rbd_dev = NULL;
5138 struct list_head *tmp;
5141 bool already = false;
5144 ret = strict_strtoul(buf, 10, &ul);
5148 /* convert to int; abort if we lost anything in the conversion */
5153 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
5156 spin_lock(&rbd_dev_list_lock);
5157 list_for_each(tmp, &rbd_dev_list) {
5158 rbd_dev = list_entry(tmp, struct rbd_device, node);
5159 if (rbd_dev->dev_id == dev_id) {
5165 spin_lock_irq(&rbd_dev->lock);
5166 if (rbd_dev->open_count)
5169 already = test_and_set_bit(RBD_DEV_FLAG_REMOVING,
5171 spin_unlock_irq(&rbd_dev->lock);
5173 spin_unlock(&rbd_dev_list_lock);
5174 if (ret < 0 || already)
5177 ret = rbd_dev_header_watch_sync(rbd_dev, false);
5179 rbd_warn(rbd_dev, "failed to cancel watch event (%d)\n", ret);
5182 * flush remaining watch callbacks - these must be complete
5183 * before the osd_client is shutdown
5185 dout("%s: flushing notifies", __func__);
5186 ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
5188 * Don't free anything from rbd_dev->disk until after all
5189 * notifies are completely processed. Otherwise
5190 * rbd_bus_del_dev() will race with rbd_watch_cb(), resulting
5191 * in a potential use after free of rbd_dev->disk or rbd_dev.
5193 rbd_bus_del_dev(rbd_dev);
5194 rbd_dev_image_release(rbd_dev);
5195 module_put(THIS_MODULE);
5198 mutex_unlock(&ctl_mutex);
5204 * create control files in sysfs
5207 static int rbd_sysfs_init(void)
5211 ret = device_register(&rbd_root_dev);
5215 ret = bus_register(&rbd_bus_type);
5217 device_unregister(&rbd_root_dev);
5222 static void rbd_sysfs_cleanup(void)
5224 bus_unregister(&rbd_bus_type);
5225 device_unregister(&rbd_root_dev);
5228 static int rbd_slab_init(void)
5230 rbd_assert(!rbd_img_request_cache);
5231 rbd_img_request_cache = kmem_cache_create("rbd_img_request",
5232 sizeof (struct rbd_img_request),
5233 __alignof__(struct rbd_img_request),
5235 if (!rbd_img_request_cache)
5238 rbd_assert(!rbd_obj_request_cache);
5239 rbd_obj_request_cache = kmem_cache_create("rbd_obj_request",
5240 sizeof (struct rbd_obj_request),
5241 __alignof__(struct rbd_obj_request),
5243 if (!rbd_obj_request_cache)
5246 rbd_assert(!rbd_segment_name_cache);
5247 rbd_segment_name_cache = kmem_cache_create("rbd_segment_name",
5248 MAX_OBJ_NAME_SIZE + 1, 1, 0, NULL);
5249 if (rbd_segment_name_cache)
5252 if (rbd_obj_request_cache) {
5253 kmem_cache_destroy(rbd_obj_request_cache);
5254 rbd_obj_request_cache = NULL;
5257 kmem_cache_destroy(rbd_img_request_cache);
5258 rbd_img_request_cache = NULL;
5263 static void rbd_slab_exit(void)
5265 rbd_assert(rbd_segment_name_cache);
5266 kmem_cache_destroy(rbd_segment_name_cache);
5267 rbd_segment_name_cache = NULL;
5269 rbd_assert(rbd_obj_request_cache);
5270 kmem_cache_destroy(rbd_obj_request_cache);
5271 rbd_obj_request_cache = NULL;
5273 rbd_assert(rbd_img_request_cache);
5274 kmem_cache_destroy(rbd_img_request_cache);
5275 rbd_img_request_cache = NULL;
5278 static int __init rbd_init(void)
5282 if (!libceph_compatible(NULL)) {
5283 rbd_warn(NULL, "libceph incompatibility (quitting)");
5287 rc = rbd_slab_init();
5290 rc = rbd_sysfs_init();
5294 pr_info("loaded " RBD_DRV_NAME_LONG "\n");
5299 static void __exit rbd_exit(void)
5301 rbd_sysfs_cleanup();
5305 module_init(rbd_init);
5306 module_exit(rbd_exit);
5308 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
5309 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
5310 MODULE_DESCRIPTION("rados block device");
5312 /* following authorship retained from original osdblk.c */
5313 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
5315 MODULE_LICENSE("GPL");