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 int 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);
517 static const struct block_device_operations rbd_bd_ops = {
518 .owner = THIS_MODULE,
520 .release = rbd_release,
524 * Initialize an rbd client instance.
527 static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts)
529 struct rbd_client *rbdc;
532 dout("%s:\n", __func__);
533 rbdc = kmalloc(sizeof(struct rbd_client), GFP_KERNEL);
537 kref_init(&rbdc->kref);
538 INIT_LIST_HEAD(&rbdc->node);
540 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
542 rbdc->client = ceph_create_client(ceph_opts, rbdc, 0, 0);
543 if (IS_ERR(rbdc->client))
545 ceph_opts = NULL; /* Now rbdc->client is responsible for ceph_opts */
547 ret = ceph_open_session(rbdc->client);
551 spin_lock(&rbd_client_list_lock);
552 list_add_tail(&rbdc->node, &rbd_client_list);
553 spin_unlock(&rbd_client_list_lock);
555 mutex_unlock(&ctl_mutex);
556 dout("%s: rbdc %p\n", __func__, rbdc);
561 ceph_destroy_client(rbdc->client);
563 mutex_unlock(&ctl_mutex);
567 ceph_destroy_options(ceph_opts);
568 dout("%s: error %d\n", __func__, ret);
573 static struct rbd_client *__rbd_get_client(struct rbd_client *rbdc)
575 kref_get(&rbdc->kref);
581 * Find a ceph client with specific addr and configuration. If
582 * found, bump its reference count.
584 static struct rbd_client *rbd_client_find(struct ceph_options *ceph_opts)
586 struct rbd_client *client_node;
589 if (ceph_opts->flags & CEPH_OPT_NOSHARE)
592 spin_lock(&rbd_client_list_lock);
593 list_for_each_entry(client_node, &rbd_client_list, node) {
594 if (!ceph_compare_options(ceph_opts, client_node->client)) {
595 __rbd_get_client(client_node);
601 spin_unlock(&rbd_client_list_lock);
603 return found ? client_node : NULL;
613 /* string args above */
616 /* Boolean args above */
620 static match_table_t rbd_opts_tokens = {
622 /* string args above */
623 {Opt_read_only, "read_only"},
624 {Opt_read_only, "ro"}, /* Alternate spelling */
625 {Opt_read_write, "read_write"},
626 {Opt_read_write, "rw"}, /* Alternate spelling */
627 /* Boolean args above */
635 #define RBD_READ_ONLY_DEFAULT false
637 static int parse_rbd_opts_token(char *c, void *private)
639 struct rbd_options *rbd_opts = private;
640 substring_t argstr[MAX_OPT_ARGS];
641 int token, intval, ret;
643 token = match_token(c, rbd_opts_tokens, argstr);
647 if (token < Opt_last_int) {
648 ret = match_int(&argstr[0], &intval);
650 pr_err("bad mount option arg (not int) "
654 dout("got int token %d val %d\n", token, intval);
655 } else if (token > Opt_last_int && token < Opt_last_string) {
656 dout("got string token %d val %s\n", token,
658 } else if (token > Opt_last_string && token < Opt_last_bool) {
659 dout("got Boolean token %d\n", token);
661 dout("got token %d\n", token);
666 rbd_opts->read_only = true;
669 rbd_opts->read_only = false;
679 * Get a ceph client with specific addr and configuration, if one does
680 * not exist create it.
682 static struct rbd_client *rbd_get_client(struct ceph_options *ceph_opts)
684 struct rbd_client *rbdc;
686 rbdc = rbd_client_find(ceph_opts);
687 if (rbdc) /* using an existing client */
688 ceph_destroy_options(ceph_opts);
690 rbdc = rbd_client_create(ceph_opts);
696 * Destroy ceph client
698 * Caller must hold rbd_client_list_lock.
700 static void rbd_client_release(struct kref *kref)
702 struct rbd_client *rbdc = container_of(kref, struct rbd_client, kref);
704 dout("%s: rbdc %p\n", __func__, rbdc);
705 spin_lock(&rbd_client_list_lock);
706 list_del(&rbdc->node);
707 spin_unlock(&rbd_client_list_lock);
709 ceph_destroy_client(rbdc->client);
714 * Drop reference to ceph client node. If it's not referenced anymore, release
717 static void rbd_put_client(struct rbd_client *rbdc)
720 kref_put(&rbdc->kref, rbd_client_release);
723 static bool rbd_image_format_valid(u32 image_format)
725 return image_format == 1 || image_format == 2;
728 static bool rbd_dev_ondisk_valid(struct rbd_image_header_ondisk *ondisk)
733 /* The header has to start with the magic rbd header text */
734 if (memcmp(&ondisk->text, RBD_HEADER_TEXT, sizeof (RBD_HEADER_TEXT)))
737 /* The bio layer requires at least sector-sized I/O */
739 if (ondisk->options.order < SECTOR_SHIFT)
742 /* If we use u64 in a few spots we may be able to loosen this */
744 if (ondisk->options.order > 8 * sizeof (int) - 1)
748 * The size of a snapshot header has to fit in a size_t, and
749 * that limits the number of snapshots.
751 snap_count = le32_to_cpu(ondisk->snap_count);
752 size = SIZE_MAX - sizeof (struct ceph_snap_context);
753 if (snap_count > size / sizeof (__le64))
757 * Not only that, but the size of the entire the snapshot
758 * header must also be representable in a size_t.
760 size -= snap_count * sizeof (__le64);
761 if ((u64) size < le64_to_cpu(ondisk->snap_names_len))
768 * Fill an rbd image header with information from the given format 1
771 static int rbd_header_from_disk(struct rbd_device *rbd_dev,
772 struct rbd_image_header_ondisk *ondisk)
774 struct rbd_image_header *header = &rbd_dev->header;
775 bool first_time = header->object_prefix == NULL;
776 struct ceph_snap_context *snapc;
777 char *object_prefix = NULL;
778 char *snap_names = NULL;
779 u64 *snap_sizes = NULL;
785 /* Allocate this now to avoid having to handle failure below */
790 len = strnlen(ondisk->object_prefix,
791 sizeof (ondisk->object_prefix));
792 object_prefix = kmalloc(len + 1, GFP_KERNEL);
795 memcpy(object_prefix, ondisk->object_prefix, len);
796 object_prefix[len] = '\0';
799 /* Allocate the snapshot context and fill it in */
801 snap_count = le32_to_cpu(ondisk->snap_count);
802 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
805 snapc->seq = le64_to_cpu(ondisk->snap_seq);
807 struct rbd_image_snap_ondisk *snaps;
808 u64 snap_names_len = le64_to_cpu(ondisk->snap_names_len);
810 /* We'll keep a copy of the snapshot names... */
812 if (snap_names_len > (u64)SIZE_MAX)
814 snap_names = kmalloc(snap_names_len, GFP_KERNEL);
818 /* ...as well as the array of their sizes. */
820 size = snap_count * sizeof (*header->snap_sizes);
821 snap_sizes = kmalloc(size, GFP_KERNEL);
826 * Copy the names, and fill in each snapshot's id
829 * Note that rbd_dev_v1_header_info() guarantees the
830 * ondisk buffer we're working with has
831 * snap_names_len bytes beyond the end of the
832 * snapshot id array, this memcpy() is safe.
834 memcpy(snap_names, &ondisk->snaps[snap_count], snap_names_len);
835 snaps = ondisk->snaps;
836 for (i = 0; i < snap_count; i++) {
837 snapc->snaps[i] = le64_to_cpu(snaps[i].id);
838 snap_sizes[i] = le64_to_cpu(snaps[i].image_size);
842 /* We won't fail any more, fill in the header */
844 down_write(&rbd_dev->header_rwsem);
846 header->object_prefix = object_prefix;
847 header->obj_order = ondisk->options.order;
848 header->crypt_type = ondisk->options.crypt_type;
849 header->comp_type = ondisk->options.comp_type;
850 /* The rest aren't used for format 1 images */
851 header->stripe_unit = 0;
852 header->stripe_count = 0;
853 header->features = 0;
855 ceph_put_snap_context(header->snapc);
856 kfree(header->snap_names);
857 kfree(header->snap_sizes);
860 /* The remaining fields always get updated (when we refresh) */
862 header->image_size = le64_to_cpu(ondisk->image_size);
863 header->snapc = snapc;
864 header->snap_names = snap_names;
865 header->snap_sizes = snap_sizes;
867 /* Make sure mapping size is consistent with header info */
869 if (rbd_dev->spec->snap_id == CEPH_NOSNAP || first_time)
870 if (rbd_dev->mapping.size != header->image_size)
871 rbd_dev->mapping.size = header->image_size;
873 up_write(&rbd_dev->header_rwsem);
881 ceph_put_snap_context(snapc);
882 kfree(object_prefix);
887 static const char *_rbd_dev_v1_snap_name(struct rbd_device *rbd_dev, u32 which)
889 const char *snap_name;
891 rbd_assert(which < rbd_dev->header.snapc->num_snaps);
893 /* Skip over names until we find the one we are looking for */
895 snap_name = rbd_dev->header.snap_names;
897 snap_name += strlen(snap_name) + 1;
899 return kstrdup(snap_name, GFP_KERNEL);
903 * Snapshot id comparison function for use with qsort()/bsearch().
904 * Note that result is for snapshots in *descending* order.
906 static int snapid_compare_reverse(const void *s1, const void *s2)
908 u64 snap_id1 = *(u64 *)s1;
909 u64 snap_id2 = *(u64 *)s2;
911 if (snap_id1 < snap_id2)
913 return snap_id1 == snap_id2 ? 0 : -1;
917 * Search a snapshot context to see if the given snapshot id is
920 * Returns the position of the snapshot id in the array if it's found,
921 * or BAD_SNAP_INDEX otherwise.
923 * Note: The snapshot array is in kept sorted (by the osd) in
924 * reverse order, highest snapshot id first.
926 static u32 rbd_dev_snap_index(struct rbd_device *rbd_dev, u64 snap_id)
928 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
931 found = bsearch(&snap_id, &snapc->snaps, snapc->num_snaps,
932 sizeof (snap_id), snapid_compare_reverse);
934 return found ? (u32)(found - &snapc->snaps[0]) : BAD_SNAP_INDEX;
937 static const char *rbd_dev_v1_snap_name(struct rbd_device *rbd_dev,
942 which = rbd_dev_snap_index(rbd_dev, snap_id);
943 if (which == BAD_SNAP_INDEX)
946 return _rbd_dev_v1_snap_name(rbd_dev, which);
949 static const char *rbd_snap_name(struct rbd_device *rbd_dev, u64 snap_id)
951 if (snap_id == CEPH_NOSNAP)
952 return RBD_SNAP_HEAD_NAME;
954 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
955 if (rbd_dev->image_format == 1)
956 return rbd_dev_v1_snap_name(rbd_dev, snap_id);
958 return rbd_dev_v2_snap_name(rbd_dev, snap_id);
961 static int rbd_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
964 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
965 if (snap_id == CEPH_NOSNAP) {
966 *snap_size = rbd_dev->header.image_size;
967 } else if (rbd_dev->image_format == 1) {
970 which = rbd_dev_snap_index(rbd_dev, snap_id);
971 if (which == BAD_SNAP_INDEX)
974 *snap_size = rbd_dev->header.snap_sizes[which];
979 ret = _rbd_dev_v2_snap_size(rbd_dev, snap_id, NULL, &size);
988 static int rbd_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
991 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
992 if (snap_id == CEPH_NOSNAP) {
993 *snap_features = rbd_dev->header.features;
994 } else if (rbd_dev->image_format == 1) {
995 *snap_features = 0; /* No features for format 1 */
1000 ret = _rbd_dev_v2_snap_features(rbd_dev, snap_id, &features);
1004 *snap_features = features;
1009 static int rbd_dev_mapping_set(struct rbd_device *rbd_dev)
1011 u64 snap_id = rbd_dev->spec->snap_id;
1016 ret = rbd_snap_size(rbd_dev, snap_id, &size);
1019 ret = rbd_snap_features(rbd_dev, snap_id, &features);
1023 rbd_dev->mapping.size = size;
1024 rbd_dev->mapping.features = features;
1029 static void rbd_dev_mapping_clear(struct rbd_device *rbd_dev)
1031 rbd_dev->mapping.size = 0;
1032 rbd_dev->mapping.features = 0;
1035 static const char *rbd_segment_name(struct rbd_device *rbd_dev, u64 offset)
1041 name = kmem_cache_alloc(rbd_segment_name_cache, GFP_NOIO);
1044 segment = offset >> rbd_dev->header.obj_order;
1045 ret = snprintf(name, MAX_OBJ_NAME_SIZE + 1, "%s.%012llx",
1046 rbd_dev->header.object_prefix, segment);
1047 if (ret < 0 || ret > MAX_OBJ_NAME_SIZE) {
1048 pr_err("error formatting segment name for #%llu (%d)\n",
1057 static void rbd_segment_name_free(const char *name)
1059 /* The explicit cast here is needed to drop the const qualifier */
1061 kmem_cache_free(rbd_segment_name_cache, (void *)name);
1064 static u64 rbd_segment_offset(struct rbd_device *rbd_dev, u64 offset)
1066 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1068 return offset & (segment_size - 1);
1071 static u64 rbd_segment_length(struct rbd_device *rbd_dev,
1072 u64 offset, u64 length)
1074 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1076 offset &= segment_size - 1;
1078 rbd_assert(length <= U64_MAX - offset);
1079 if (offset + length > segment_size)
1080 length = segment_size - offset;
1086 * returns the size of an object in the image
1088 static u64 rbd_obj_bytes(struct rbd_image_header *header)
1090 return 1 << header->obj_order;
1097 static void bio_chain_put(struct bio *chain)
1103 chain = chain->bi_next;
1109 * zeros a bio chain, starting at specific offset
1111 static void zero_bio_chain(struct bio *chain, int start_ofs)
1114 unsigned long flags;
1120 bio_for_each_segment(bv, chain, i) {
1121 if (pos + bv->bv_len > start_ofs) {
1122 int remainder = max(start_ofs - pos, 0);
1123 buf = bvec_kmap_irq(bv, &flags);
1124 memset(buf + remainder, 0,
1125 bv->bv_len - remainder);
1126 bvec_kunmap_irq(buf, &flags);
1131 chain = chain->bi_next;
1136 * similar to zero_bio_chain(), zeros data defined by a page array,
1137 * starting at the given byte offset from the start of the array and
1138 * continuing up to the given end offset. The pages array is
1139 * assumed to be big enough to hold all bytes up to the end.
1141 static void zero_pages(struct page **pages, u64 offset, u64 end)
1143 struct page **page = &pages[offset >> PAGE_SHIFT];
1145 rbd_assert(end > offset);
1146 rbd_assert(end - offset <= (u64)SIZE_MAX);
1147 while (offset < end) {
1150 unsigned long flags;
1153 page_offset = (size_t)(offset & ~PAGE_MASK);
1154 length = min(PAGE_SIZE - page_offset, (size_t)(end - offset));
1155 local_irq_save(flags);
1156 kaddr = kmap_atomic(*page);
1157 memset(kaddr + page_offset, 0, length);
1158 kunmap_atomic(kaddr);
1159 local_irq_restore(flags);
1167 * Clone a portion of a bio, starting at the given byte offset
1168 * and continuing for the number of bytes indicated.
1170 static struct bio *bio_clone_range(struct bio *bio_src,
1171 unsigned int offset,
1179 unsigned short end_idx;
1180 unsigned short vcnt;
1183 /* Handle the easy case for the caller */
1185 if (!offset && len == bio_src->bi_size)
1186 return bio_clone(bio_src, gfpmask);
1188 if (WARN_ON_ONCE(!len))
1190 if (WARN_ON_ONCE(len > bio_src->bi_size))
1192 if (WARN_ON_ONCE(offset > bio_src->bi_size - len))
1195 /* Find first affected segment... */
1198 __bio_for_each_segment(bv, bio_src, idx, 0) {
1199 if (resid < bv->bv_len)
1201 resid -= bv->bv_len;
1205 /* ...and the last affected segment */
1208 __bio_for_each_segment(bv, bio_src, end_idx, idx) {
1209 if (resid <= bv->bv_len)
1211 resid -= bv->bv_len;
1213 vcnt = end_idx - idx + 1;
1215 /* Build the clone */
1217 bio = bio_alloc(gfpmask, (unsigned int) vcnt);
1219 return NULL; /* ENOMEM */
1221 bio->bi_bdev = bio_src->bi_bdev;
1222 bio->bi_sector = bio_src->bi_sector + (offset >> SECTOR_SHIFT);
1223 bio->bi_rw = bio_src->bi_rw;
1224 bio->bi_flags |= 1 << BIO_CLONED;
1227 * Copy over our part of the bio_vec, then update the first
1228 * and last (or only) entries.
1230 memcpy(&bio->bi_io_vec[0], &bio_src->bi_io_vec[idx],
1231 vcnt * sizeof (struct bio_vec));
1232 bio->bi_io_vec[0].bv_offset += voff;
1234 bio->bi_io_vec[0].bv_len -= voff;
1235 bio->bi_io_vec[vcnt - 1].bv_len = resid;
1237 bio->bi_io_vec[0].bv_len = len;
1240 bio->bi_vcnt = vcnt;
1248 * Clone a portion of a bio chain, starting at the given byte offset
1249 * into the first bio in the source chain and continuing for the
1250 * number of bytes indicated. The result is another bio chain of
1251 * exactly the given length, or a null pointer on error.
1253 * The bio_src and offset parameters are both in-out. On entry they
1254 * refer to the first source bio and the offset into that bio where
1255 * the start of data to be cloned is located.
1257 * On return, bio_src is updated to refer to the bio in the source
1258 * chain that contains first un-cloned byte, and *offset will
1259 * contain the offset of that byte within that bio.
1261 static struct bio *bio_chain_clone_range(struct bio **bio_src,
1262 unsigned int *offset,
1266 struct bio *bi = *bio_src;
1267 unsigned int off = *offset;
1268 struct bio *chain = NULL;
1271 /* Build up a chain of clone bios up to the limit */
1273 if (!bi || off >= bi->bi_size || !len)
1274 return NULL; /* Nothing to clone */
1278 unsigned int bi_size;
1282 rbd_warn(NULL, "bio_chain exhausted with %u left", len);
1283 goto out_err; /* EINVAL; ran out of bio's */
1285 bi_size = min_t(unsigned int, bi->bi_size - off, len);
1286 bio = bio_clone_range(bi, off, bi_size, gfpmask);
1288 goto out_err; /* ENOMEM */
1291 end = &bio->bi_next;
1294 if (off == bi->bi_size) {
1305 bio_chain_put(chain);
1311 * The default/initial value for all object request flags is 0. For
1312 * each flag, once its value is set to 1 it is never reset to 0
1315 static void obj_request_img_data_set(struct rbd_obj_request *obj_request)
1317 if (test_and_set_bit(OBJ_REQ_IMG_DATA, &obj_request->flags)) {
1318 struct rbd_device *rbd_dev;
1320 rbd_dev = obj_request->img_request->rbd_dev;
1321 rbd_warn(rbd_dev, "obj_request %p already marked img_data\n",
1326 static bool obj_request_img_data_test(struct rbd_obj_request *obj_request)
1329 return test_bit(OBJ_REQ_IMG_DATA, &obj_request->flags) != 0;
1332 static void obj_request_done_set(struct rbd_obj_request *obj_request)
1334 if (test_and_set_bit(OBJ_REQ_DONE, &obj_request->flags)) {
1335 struct rbd_device *rbd_dev = NULL;
1337 if (obj_request_img_data_test(obj_request))
1338 rbd_dev = obj_request->img_request->rbd_dev;
1339 rbd_warn(rbd_dev, "obj_request %p already marked done\n",
1344 static bool obj_request_done_test(struct rbd_obj_request *obj_request)
1347 return test_bit(OBJ_REQ_DONE, &obj_request->flags) != 0;
1351 * This sets the KNOWN flag after (possibly) setting the EXISTS
1352 * flag. The latter is set based on the "exists" value provided.
1354 * Note that for our purposes once an object exists it never goes
1355 * away again. It's possible that the response from two existence
1356 * checks are separated by the creation of the target object, and
1357 * the first ("doesn't exist") response arrives *after* the second
1358 * ("does exist"). In that case we ignore the second one.
1360 static void obj_request_existence_set(struct rbd_obj_request *obj_request,
1364 set_bit(OBJ_REQ_EXISTS, &obj_request->flags);
1365 set_bit(OBJ_REQ_KNOWN, &obj_request->flags);
1369 static bool obj_request_known_test(struct rbd_obj_request *obj_request)
1372 return test_bit(OBJ_REQ_KNOWN, &obj_request->flags) != 0;
1375 static bool obj_request_exists_test(struct rbd_obj_request *obj_request)
1378 return test_bit(OBJ_REQ_EXISTS, &obj_request->flags) != 0;
1381 static void rbd_obj_request_get(struct rbd_obj_request *obj_request)
1383 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1384 atomic_read(&obj_request->kref.refcount));
1385 kref_get(&obj_request->kref);
1388 static void rbd_obj_request_destroy(struct kref *kref);
1389 static void rbd_obj_request_put(struct rbd_obj_request *obj_request)
1391 rbd_assert(obj_request != NULL);
1392 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1393 atomic_read(&obj_request->kref.refcount));
1394 kref_put(&obj_request->kref, rbd_obj_request_destroy);
1397 static bool img_request_child_test(struct rbd_img_request *img_request);
1398 static void rbd_parent_request_destroy(struct kref *kref);
1399 static void rbd_img_request_destroy(struct kref *kref);
1400 static void rbd_img_request_put(struct rbd_img_request *img_request)
1402 rbd_assert(img_request != NULL);
1403 dout("%s: img %p (was %d)\n", __func__, img_request,
1404 atomic_read(&img_request->kref.refcount));
1405 if (img_request_child_test(img_request))
1406 kref_put(&img_request->kref, rbd_parent_request_destroy);
1408 kref_put(&img_request->kref, rbd_img_request_destroy);
1411 static inline void rbd_img_obj_request_add(struct rbd_img_request *img_request,
1412 struct rbd_obj_request *obj_request)
1414 rbd_assert(obj_request->img_request == NULL);
1416 /* Image request now owns object's original reference */
1417 obj_request->img_request = img_request;
1418 obj_request->which = img_request->obj_request_count;
1419 rbd_assert(!obj_request_img_data_test(obj_request));
1420 obj_request_img_data_set(obj_request);
1421 rbd_assert(obj_request->which != BAD_WHICH);
1422 img_request->obj_request_count++;
1423 list_add_tail(&obj_request->links, &img_request->obj_requests);
1424 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1425 obj_request->which);
1428 static inline void rbd_img_obj_request_del(struct rbd_img_request *img_request,
1429 struct rbd_obj_request *obj_request)
1431 rbd_assert(obj_request->which != BAD_WHICH);
1433 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1434 obj_request->which);
1435 list_del(&obj_request->links);
1436 rbd_assert(img_request->obj_request_count > 0);
1437 img_request->obj_request_count--;
1438 rbd_assert(obj_request->which == img_request->obj_request_count);
1439 obj_request->which = BAD_WHICH;
1440 rbd_assert(obj_request_img_data_test(obj_request));
1441 rbd_assert(obj_request->img_request == img_request);
1442 obj_request->img_request = NULL;
1443 obj_request->callback = NULL;
1444 rbd_obj_request_put(obj_request);
1447 static bool obj_request_type_valid(enum obj_request_type type)
1450 case OBJ_REQUEST_NODATA:
1451 case OBJ_REQUEST_BIO:
1452 case OBJ_REQUEST_PAGES:
1459 static int rbd_obj_request_submit(struct ceph_osd_client *osdc,
1460 struct rbd_obj_request *obj_request)
1462 dout("%s: osdc %p obj %p\n", __func__, osdc, obj_request);
1464 return ceph_osdc_start_request(osdc, obj_request->osd_req, false);
1467 static void rbd_img_request_complete(struct rbd_img_request *img_request)
1470 dout("%s: img %p\n", __func__, img_request);
1473 * If no error occurred, compute the aggregate transfer
1474 * count for the image request. We could instead use
1475 * atomic64_cmpxchg() to update it as each object request
1476 * completes; not clear which way is better off hand.
1478 if (!img_request->result) {
1479 struct rbd_obj_request *obj_request;
1482 for_each_obj_request(img_request, obj_request)
1483 xferred += obj_request->xferred;
1484 img_request->xferred = xferred;
1487 if (img_request->callback)
1488 img_request->callback(img_request);
1490 rbd_img_request_put(img_request);
1493 /* Caller is responsible for rbd_obj_request_destroy(obj_request) */
1495 static int rbd_obj_request_wait(struct rbd_obj_request *obj_request)
1497 dout("%s: obj %p\n", __func__, obj_request);
1499 return wait_for_completion_interruptible(&obj_request->completion);
1503 * The default/initial value for all image request flags is 0. Each
1504 * is conditionally set to 1 at image request initialization time
1505 * and currently never change thereafter.
1507 static void img_request_write_set(struct rbd_img_request *img_request)
1509 set_bit(IMG_REQ_WRITE, &img_request->flags);
1513 static bool img_request_write_test(struct rbd_img_request *img_request)
1516 return test_bit(IMG_REQ_WRITE, &img_request->flags) != 0;
1519 static void img_request_child_set(struct rbd_img_request *img_request)
1521 set_bit(IMG_REQ_CHILD, &img_request->flags);
1525 static void img_request_child_clear(struct rbd_img_request *img_request)
1527 clear_bit(IMG_REQ_CHILD, &img_request->flags);
1531 static bool img_request_child_test(struct rbd_img_request *img_request)
1534 return test_bit(IMG_REQ_CHILD, &img_request->flags) != 0;
1537 static void img_request_layered_set(struct rbd_img_request *img_request)
1539 set_bit(IMG_REQ_LAYERED, &img_request->flags);
1543 static void img_request_layered_clear(struct rbd_img_request *img_request)
1545 clear_bit(IMG_REQ_LAYERED, &img_request->flags);
1549 static bool img_request_layered_test(struct rbd_img_request *img_request)
1552 return test_bit(IMG_REQ_LAYERED, &img_request->flags) != 0;
1556 rbd_img_obj_request_read_callback(struct rbd_obj_request *obj_request)
1558 u64 xferred = obj_request->xferred;
1559 u64 length = obj_request->length;
1561 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1562 obj_request, obj_request->img_request, obj_request->result,
1565 * ENOENT means a hole in the image. We zero-fill the
1566 * entire length of the request. A short read also implies
1567 * zero-fill to the end of the request. Either way we
1568 * update the xferred count to indicate the whole request
1571 rbd_assert(obj_request->type != OBJ_REQUEST_NODATA);
1572 if (obj_request->result == -ENOENT) {
1573 if (obj_request->type == OBJ_REQUEST_BIO)
1574 zero_bio_chain(obj_request->bio_list, 0);
1576 zero_pages(obj_request->pages, 0, length);
1577 obj_request->result = 0;
1578 obj_request->xferred = length;
1579 } else if (xferred < length && !obj_request->result) {
1580 if (obj_request->type == OBJ_REQUEST_BIO)
1581 zero_bio_chain(obj_request->bio_list, xferred);
1583 zero_pages(obj_request->pages, xferred, length);
1584 obj_request->xferred = length;
1586 obj_request_done_set(obj_request);
1589 static void rbd_obj_request_complete(struct rbd_obj_request *obj_request)
1591 dout("%s: obj %p cb %p\n", __func__, obj_request,
1592 obj_request->callback);
1593 if (obj_request->callback)
1594 obj_request->callback(obj_request);
1596 complete_all(&obj_request->completion);
1599 static void rbd_osd_trivial_callback(struct rbd_obj_request *obj_request)
1601 dout("%s: obj %p\n", __func__, obj_request);
1602 obj_request_done_set(obj_request);
1605 static void rbd_osd_read_callback(struct rbd_obj_request *obj_request)
1607 struct rbd_img_request *img_request = NULL;
1608 struct rbd_device *rbd_dev = NULL;
1609 bool layered = false;
1611 if (obj_request_img_data_test(obj_request)) {
1612 img_request = obj_request->img_request;
1613 layered = img_request && img_request_layered_test(img_request);
1614 rbd_dev = img_request->rbd_dev;
1617 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1618 obj_request, img_request, obj_request->result,
1619 obj_request->xferred, obj_request->length);
1620 if (layered && obj_request->result == -ENOENT &&
1621 obj_request->img_offset < rbd_dev->parent_overlap)
1622 rbd_img_parent_read(obj_request);
1623 else if (img_request)
1624 rbd_img_obj_request_read_callback(obj_request);
1626 obj_request_done_set(obj_request);
1629 static void rbd_osd_write_callback(struct rbd_obj_request *obj_request)
1631 dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1632 obj_request->result, obj_request->length);
1634 * There is no such thing as a successful short write. Set
1635 * it to our originally-requested length.
1637 obj_request->xferred = obj_request->length;
1638 obj_request_done_set(obj_request);
1642 * For a simple stat call there's nothing to do. We'll do more if
1643 * this is part of a write sequence for a layered image.
1645 static void rbd_osd_stat_callback(struct rbd_obj_request *obj_request)
1647 dout("%s: obj %p\n", __func__, obj_request);
1648 obj_request_done_set(obj_request);
1651 static void rbd_osd_req_callback(struct ceph_osd_request *osd_req,
1652 struct ceph_msg *msg)
1654 struct rbd_obj_request *obj_request = osd_req->r_priv;
1657 dout("%s: osd_req %p msg %p\n", __func__, osd_req, msg);
1658 rbd_assert(osd_req == obj_request->osd_req);
1659 if (obj_request_img_data_test(obj_request)) {
1660 rbd_assert(obj_request->img_request);
1661 rbd_assert(obj_request->which != BAD_WHICH);
1663 rbd_assert(obj_request->which == BAD_WHICH);
1666 if (osd_req->r_result < 0)
1667 obj_request->result = osd_req->r_result;
1669 BUG_ON(osd_req->r_num_ops > 2);
1672 * We support a 64-bit length, but ultimately it has to be
1673 * passed to blk_end_request(), which takes an unsigned int.
1675 obj_request->xferred = osd_req->r_reply_op_len[0];
1676 rbd_assert(obj_request->xferred < (u64)UINT_MAX);
1677 opcode = osd_req->r_ops[0].op;
1679 case CEPH_OSD_OP_READ:
1680 rbd_osd_read_callback(obj_request);
1682 case CEPH_OSD_OP_WRITE:
1683 rbd_osd_write_callback(obj_request);
1685 case CEPH_OSD_OP_STAT:
1686 rbd_osd_stat_callback(obj_request);
1688 case CEPH_OSD_OP_CALL:
1689 case CEPH_OSD_OP_NOTIFY_ACK:
1690 case CEPH_OSD_OP_WATCH:
1691 rbd_osd_trivial_callback(obj_request);
1694 rbd_warn(NULL, "%s: unsupported op %hu\n",
1695 obj_request->object_name, (unsigned short) opcode);
1699 if (obj_request_done_test(obj_request))
1700 rbd_obj_request_complete(obj_request);
1703 static void rbd_osd_req_format_read(struct rbd_obj_request *obj_request)
1705 struct rbd_img_request *img_request = obj_request->img_request;
1706 struct ceph_osd_request *osd_req = obj_request->osd_req;
1709 rbd_assert(osd_req != NULL);
1711 snap_id = img_request ? img_request->snap_id : CEPH_NOSNAP;
1712 ceph_osdc_build_request(osd_req, obj_request->offset,
1713 NULL, snap_id, NULL);
1716 static void rbd_osd_req_format_write(struct rbd_obj_request *obj_request)
1718 struct rbd_img_request *img_request = obj_request->img_request;
1719 struct ceph_osd_request *osd_req = obj_request->osd_req;
1720 struct ceph_snap_context *snapc;
1721 struct timespec mtime = CURRENT_TIME;
1723 rbd_assert(osd_req != NULL);
1725 snapc = img_request ? img_request->snapc : NULL;
1726 ceph_osdc_build_request(osd_req, obj_request->offset,
1727 snapc, CEPH_NOSNAP, &mtime);
1730 static struct ceph_osd_request *rbd_osd_req_create(
1731 struct rbd_device *rbd_dev,
1733 struct rbd_obj_request *obj_request)
1735 struct ceph_snap_context *snapc = NULL;
1736 struct ceph_osd_client *osdc;
1737 struct ceph_osd_request *osd_req;
1739 if (obj_request_img_data_test(obj_request)) {
1740 struct rbd_img_request *img_request = obj_request->img_request;
1742 rbd_assert(write_request ==
1743 img_request_write_test(img_request));
1745 snapc = img_request->snapc;
1748 /* Allocate and initialize the request, for the single op */
1750 osdc = &rbd_dev->rbd_client->client->osdc;
1751 osd_req = ceph_osdc_alloc_request(osdc, snapc, 1, false, GFP_ATOMIC);
1753 return NULL; /* ENOMEM */
1756 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
1758 osd_req->r_flags = CEPH_OSD_FLAG_READ;
1760 osd_req->r_callback = rbd_osd_req_callback;
1761 osd_req->r_priv = obj_request;
1763 osd_req->r_oid_len = strlen(obj_request->object_name);
1764 rbd_assert(osd_req->r_oid_len < sizeof (osd_req->r_oid));
1765 memcpy(osd_req->r_oid, obj_request->object_name, osd_req->r_oid_len);
1767 osd_req->r_file_layout = rbd_dev->layout; /* struct */
1773 * Create a copyup osd request based on the information in the
1774 * object request supplied. A copyup request has two osd ops,
1775 * a copyup method call, and a "normal" write request.
1777 static struct ceph_osd_request *
1778 rbd_osd_req_create_copyup(struct rbd_obj_request *obj_request)
1780 struct rbd_img_request *img_request;
1781 struct ceph_snap_context *snapc;
1782 struct rbd_device *rbd_dev;
1783 struct ceph_osd_client *osdc;
1784 struct ceph_osd_request *osd_req;
1786 rbd_assert(obj_request_img_data_test(obj_request));
1787 img_request = obj_request->img_request;
1788 rbd_assert(img_request);
1789 rbd_assert(img_request_write_test(img_request));
1791 /* Allocate and initialize the request, for the two ops */
1793 snapc = img_request->snapc;
1794 rbd_dev = img_request->rbd_dev;
1795 osdc = &rbd_dev->rbd_client->client->osdc;
1796 osd_req = ceph_osdc_alloc_request(osdc, snapc, 2, false, GFP_ATOMIC);
1798 return NULL; /* ENOMEM */
1800 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
1801 osd_req->r_callback = rbd_osd_req_callback;
1802 osd_req->r_priv = obj_request;
1804 osd_req->r_oid_len = strlen(obj_request->object_name);
1805 rbd_assert(osd_req->r_oid_len < sizeof (osd_req->r_oid));
1806 memcpy(osd_req->r_oid, obj_request->object_name, osd_req->r_oid_len);
1808 osd_req->r_file_layout = rbd_dev->layout; /* struct */
1814 static void rbd_osd_req_destroy(struct ceph_osd_request *osd_req)
1816 ceph_osdc_put_request(osd_req);
1819 /* object_name is assumed to be a non-null pointer and NUL-terminated */
1821 static struct rbd_obj_request *rbd_obj_request_create(const char *object_name,
1822 u64 offset, u64 length,
1823 enum obj_request_type type)
1825 struct rbd_obj_request *obj_request;
1829 rbd_assert(obj_request_type_valid(type));
1831 size = strlen(object_name) + 1;
1832 name = kmalloc(size, GFP_KERNEL);
1836 obj_request = kmem_cache_zalloc(rbd_obj_request_cache, GFP_KERNEL);
1842 obj_request->object_name = memcpy(name, object_name, size);
1843 obj_request->offset = offset;
1844 obj_request->length = length;
1845 obj_request->flags = 0;
1846 obj_request->which = BAD_WHICH;
1847 obj_request->type = type;
1848 INIT_LIST_HEAD(&obj_request->links);
1849 init_completion(&obj_request->completion);
1850 kref_init(&obj_request->kref);
1852 dout("%s: \"%s\" %llu/%llu %d -> obj %p\n", __func__, object_name,
1853 offset, length, (int)type, obj_request);
1858 static void rbd_obj_request_destroy(struct kref *kref)
1860 struct rbd_obj_request *obj_request;
1862 obj_request = container_of(kref, struct rbd_obj_request, kref);
1864 dout("%s: obj %p\n", __func__, obj_request);
1866 rbd_assert(obj_request->img_request == NULL);
1867 rbd_assert(obj_request->which == BAD_WHICH);
1869 if (obj_request->osd_req)
1870 rbd_osd_req_destroy(obj_request->osd_req);
1872 rbd_assert(obj_request_type_valid(obj_request->type));
1873 switch (obj_request->type) {
1874 case OBJ_REQUEST_NODATA:
1875 break; /* Nothing to do */
1876 case OBJ_REQUEST_BIO:
1877 if (obj_request->bio_list)
1878 bio_chain_put(obj_request->bio_list);
1880 case OBJ_REQUEST_PAGES:
1881 if (obj_request->pages)
1882 ceph_release_page_vector(obj_request->pages,
1883 obj_request->page_count);
1887 kfree(obj_request->object_name);
1888 obj_request->object_name = NULL;
1889 kmem_cache_free(rbd_obj_request_cache, obj_request);
1892 /* It's OK to call this for a device with no parent */
1894 static void rbd_spec_put(struct rbd_spec *spec);
1895 static void rbd_dev_unparent(struct rbd_device *rbd_dev)
1897 rbd_dev_remove_parent(rbd_dev);
1898 rbd_spec_put(rbd_dev->parent_spec);
1899 rbd_dev->parent_spec = NULL;
1900 rbd_dev->parent_overlap = 0;
1904 * Parent image reference counting is used to determine when an
1905 * image's parent fields can be safely torn down--after there are no
1906 * more in-flight requests to the parent image. When the last
1907 * reference is dropped, cleaning them up is safe.
1909 static void rbd_dev_parent_put(struct rbd_device *rbd_dev)
1913 if (!rbd_dev->parent_spec)
1916 counter = atomic_dec_return_safe(&rbd_dev->parent_ref);
1920 /* Last reference; clean up parent data structures */
1923 rbd_dev_unparent(rbd_dev);
1925 rbd_warn(rbd_dev, "parent reference underflow\n");
1929 * If an image has a non-zero parent overlap, get a reference to its
1932 * We must get the reference before checking for the overlap to
1933 * coordinate properly with zeroing the parent overlap in
1934 * rbd_dev_v2_parent_info() when an image gets flattened. We
1935 * drop it again if there is no overlap.
1937 * Returns true if the rbd device has a parent with a non-zero
1938 * overlap and a reference for it was successfully taken, or
1941 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
1945 if (!rbd_dev->parent_spec)
1948 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
1949 if (counter > 0 && rbd_dev->parent_overlap)
1952 /* Image was flattened, but parent is not yet torn down */
1955 rbd_warn(rbd_dev, "parent reference overflow\n");
1961 * Caller is responsible for filling in the list of object requests
1962 * that comprises the image request, and the Linux request pointer
1963 * (if there is one).
1965 static struct rbd_img_request *rbd_img_request_create(
1966 struct rbd_device *rbd_dev,
1967 u64 offset, u64 length,
1970 struct rbd_img_request *img_request;
1972 img_request = kmem_cache_alloc(rbd_img_request_cache, GFP_ATOMIC);
1976 if (write_request) {
1977 down_read(&rbd_dev->header_rwsem);
1978 ceph_get_snap_context(rbd_dev->header.snapc);
1979 up_read(&rbd_dev->header_rwsem);
1982 img_request->rq = NULL;
1983 img_request->rbd_dev = rbd_dev;
1984 img_request->offset = offset;
1985 img_request->length = length;
1986 img_request->flags = 0;
1987 if (write_request) {
1988 img_request_write_set(img_request);
1989 img_request->snapc = rbd_dev->header.snapc;
1991 img_request->snap_id = rbd_dev->spec->snap_id;
1993 if (rbd_dev_parent_get(rbd_dev))
1994 img_request_layered_set(img_request);
1995 spin_lock_init(&img_request->completion_lock);
1996 img_request->next_completion = 0;
1997 img_request->callback = NULL;
1998 img_request->result = 0;
1999 img_request->obj_request_count = 0;
2000 INIT_LIST_HEAD(&img_request->obj_requests);
2001 kref_init(&img_request->kref);
2003 dout("%s: rbd_dev %p %s %llu/%llu -> img %p\n", __func__, rbd_dev,
2004 write_request ? "write" : "read", offset, length,
2010 static void rbd_img_request_destroy(struct kref *kref)
2012 struct rbd_img_request *img_request;
2013 struct rbd_obj_request *obj_request;
2014 struct rbd_obj_request *next_obj_request;
2016 img_request = container_of(kref, struct rbd_img_request, kref);
2018 dout("%s: img %p\n", __func__, img_request);
2020 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2021 rbd_img_obj_request_del(img_request, obj_request);
2022 rbd_assert(img_request->obj_request_count == 0);
2024 if (img_request_layered_test(img_request)) {
2025 img_request_layered_clear(img_request);
2026 rbd_dev_parent_put(img_request->rbd_dev);
2029 if (img_request_write_test(img_request))
2030 ceph_put_snap_context(img_request->snapc);
2032 kmem_cache_free(rbd_img_request_cache, img_request);
2035 static struct rbd_img_request *rbd_parent_request_create(
2036 struct rbd_obj_request *obj_request,
2037 u64 img_offset, u64 length)
2039 struct rbd_img_request *parent_request;
2040 struct rbd_device *rbd_dev;
2042 rbd_assert(obj_request->img_request);
2043 rbd_dev = obj_request->img_request->rbd_dev;
2045 parent_request = rbd_img_request_create(rbd_dev->parent,
2046 img_offset, length, false);
2047 if (!parent_request)
2050 img_request_child_set(parent_request);
2051 rbd_obj_request_get(obj_request);
2052 parent_request->obj_request = obj_request;
2054 return parent_request;
2057 static void rbd_parent_request_destroy(struct kref *kref)
2059 struct rbd_img_request *parent_request;
2060 struct rbd_obj_request *orig_request;
2062 parent_request = container_of(kref, struct rbd_img_request, kref);
2063 orig_request = parent_request->obj_request;
2065 parent_request->obj_request = NULL;
2066 rbd_obj_request_put(orig_request);
2067 img_request_child_clear(parent_request);
2069 rbd_img_request_destroy(kref);
2072 static bool rbd_img_obj_end_request(struct rbd_obj_request *obj_request)
2074 struct rbd_img_request *img_request;
2075 unsigned int xferred;
2079 rbd_assert(obj_request_img_data_test(obj_request));
2080 img_request = obj_request->img_request;
2082 rbd_assert(obj_request->xferred <= (u64)UINT_MAX);
2083 xferred = (unsigned int)obj_request->xferred;
2084 result = obj_request->result;
2086 struct rbd_device *rbd_dev = img_request->rbd_dev;
2088 rbd_warn(rbd_dev, "%s %llx at %llx (%llx)\n",
2089 img_request_write_test(img_request) ? "write" : "read",
2090 obj_request->length, obj_request->img_offset,
2091 obj_request->offset);
2092 rbd_warn(rbd_dev, " result %d xferred %x\n",
2094 if (!img_request->result)
2095 img_request->result = result;
2098 /* Image object requests don't own their page array */
2100 if (obj_request->type == OBJ_REQUEST_PAGES) {
2101 obj_request->pages = NULL;
2102 obj_request->page_count = 0;
2105 if (img_request_child_test(img_request)) {
2106 rbd_assert(img_request->obj_request != NULL);
2107 more = obj_request->which < img_request->obj_request_count - 1;
2109 rbd_assert(img_request->rq != NULL);
2110 more = blk_end_request(img_request->rq, result, xferred);
2116 static void rbd_img_obj_callback(struct rbd_obj_request *obj_request)
2118 struct rbd_img_request *img_request;
2119 u32 which = obj_request->which;
2122 rbd_assert(obj_request_img_data_test(obj_request));
2123 img_request = obj_request->img_request;
2125 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
2126 rbd_assert(img_request != NULL);
2127 rbd_assert(img_request->obj_request_count > 0);
2128 rbd_assert(which != BAD_WHICH);
2129 rbd_assert(which < img_request->obj_request_count);
2130 rbd_assert(which >= img_request->next_completion);
2132 spin_lock_irq(&img_request->completion_lock);
2133 if (which != img_request->next_completion)
2136 for_each_obj_request_from(img_request, obj_request) {
2138 rbd_assert(which < img_request->obj_request_count);
2140 if (!obj_request_done_test(obj_request))
2142 more = rbd_img_obj_end_request(obj_request);
2146 rbd_assert(more ^ (which == img_request->obj_request_count));
2147 img_request->next_completion = which;
2149 spin_unlock_irq(&img_request->completion_lock);
2152 rbd_img_request_complete(img_request);
2156 * Split up an image request into one or more object requests, each
2157 * to a different object. The "type" parameter indicates whether
2158 * "data_desc" is the pointer to the head of a list of bio
2159 * structures, or the base of a page array. In either case this
2160 * function assumes data_desc describes memory sufficient to hold
2161 * all data described by the image request.
2163 static int rbd_img_request_fill(struct rbd_img_request *img_request,
2164 enum obj_request_type type,
2167 struct rbd_device *rbd_dev = img_request->rbd_dev;
2168 struct rbd_obj_request *obj_request = NULL;
2169 struct rbd_obj_request *next_obj_request;
2170 bool write_request = img_request_write_test(img_request);
2171 struct bio *bio_list;
2172 unsigned int bio_offset = 0;
2173 struct page **pages;
2178 dout("%s: img %p type %d data_desc %p\n", __func__, img_request,
2179 (int)type, data_desc);
2181 opcode = write_request ? CEPH_OSD_OP_WRITE : CEPH_OSD_OP_READ;
2182 img_offset = img_request->offset;
2183 resid = img_request->length;
2184 rbd_assert(resid > 0);
2186 if (type == OBJ_REQUEST_BIO) {
2187 bio_list = data_desc;
2188 rbd_assert(img_offset == bio_list->bi_sector << SECTOR_SHIFT);
2190 rbd_assert(type == OBJ_REQUEST_PAGES);
2195 struct ceph_osd_request *osd_req;
2196 const char *object_name;
2200 object_name = rbd_segment_name(rbd_dev, img_offset);
2203 offset = rbd_segment_offset(rbd_dev, img_offset);
2204 length = rbd_segment_length(rbd_dev, img_offset, resid);
2205 obj_request = rbd_obj_request_create(object_name,
2206 offset, length, type);
2207 /* object request has its own copy of the object name */
2208 rbd_segment_name_free(object_name);
2212 if (type == OBJ_REQUEST_BIO) {
2213 unsigned int clone_size;
2215 rbd_assert(length <= (u64)UINT_MAX);
2216 clone_size = (unsigned int)length;
2217 obj_request->bio_list =
2218 bio_chain_clone_range(&bio_list,
2222 if (!obj_request->bio_list)
2225 unsigned int page_count;
2227 obj_request->pages = pages;
2228 page_count = (u32)calc_pages_for(offset, length);
2229 obj_request->page_count = page_count;
2230 if ((offset + length) & ~PAGE_MASK)
2231 page_count--; /* more on last page */
2232 pages += page_count;
2235 osd_req = rbd_osd_req_create(rbd_dev, write_request,
2239 obj_request->osd_req = osd_req;
2240 obj_request->callback = rbd_img_obj_callback;
2242 osd_req_op_extent_init(osd_req, 0, opcode, offset, length,
2244 if (type == OBJ_REQUEST_BIO)
2245 osd_req_op_extent_osd_data_bio(osd_req, 0,
2246 obj_request->bio_list, length);
2248 osd_req_op_extent_osd_data_pages(osd_req, 0,
2249 obj_request->pages, length,
2250 offset & ~PAGE_MASK, false, false);
2253 rbd_osd_req_format_write(obj_request);
2255 rbd_osd_req_format_read(obj_request);
2257 obj_request->img_offset = img_offset;
2258 rbd_img_obj_request_add(img_request, obj_request);
2260 img_offset += length;
2267 rbd_obj_request_put(obj_request);
2269 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2270 rbd_obj_request_put(obj_request);
2276 rbd_img_obj_copyup_callback(struct rbd_obj_request *obj_request)
2278 struct rbd_img_request *img_request;
2279 struct rbd_device *rbd_dev;
2280 struct page **pages;
2283 rbd_assert(obj_request->type == OBJ_REQUEST_BIO);
2284 rbd_assert(obj_request_img_data_test(obj_request));
2285 img_request = obj_request->img_request;
2286 rbd_assert(img_request);
2288 rbd_dev = img_request->rbd_dev;
2289 rbd_assert(rbd_dev);
2291 pages = obj_request->copyup_pages;
2292 rbd_assert(pages != NULL);
2293 obj_request->copyup_pages = NULL;
2294 page_count = obj_request->copyup_page_count;
2295 rbd_assert(page_count);
2296 obj_request->copyup_page_count = 0;
2297 ceph_release_page_vector(pages, page_count);
2300 * We want the transfer count to reflect the size of the
2301 * original write request. There is no such thing as a
2302 * successful short write, so if the request was successful
2303 * we can just set it to the originally-requested length.
2305 if (!obj_request->result)
2306 obj_request->xferred = obj_request->length;
2308 /* Finish up with the normal image object callback */
2310 rbd_img_obj_callback(obj_request);
2314 rbd_img_obj_parent_read_full_callback(struct rbd_img_request *img_request)
2316 struct rbd_obj_request *orig_request;
2317 struct ceph_osd_request *osd_req;
2318 struct ceph_osd_client *osdc;
2319 struct rbd_device *rbd_dev;
2320 struct page **pages;
2327 rbd_assert(img_request_child_test(img_request));
2329 /* First get what we need from the image request */
2331 pages = img_request->copyup_pages;
2332 rbd_assert(pages != NULL);
2333 img_request->copyup_pages = NULL;
2334 page_count = img_request->copyup_page_count;
2335 rbd_assert(page_count);
2336 img_request->copyup_page_count = 0;
2338 orig_request = img_request->obj_request;
2339 rbd_assert(orig_request != NULL);
2340 rbd_assert(obj_request_type_valid(orig_request->type));
2341 result = img_request->result;
2342 parent_length = img_request->length;
2343 rbd_assert(parent_length == img_request->xferred);
2344 rbd_img_request_put(img_request);
2346 rbd_assert(orig_request->img_request);
2347 rbd_dev = orig_request->img_request->rbd_dev;
2348 rbd_assert(rbd_dev);
2354 * The original osd request is of no use to use any more.
2355 * We need a new one that can hold the two ops in a copyup
2356 * request. Allocate the new copyup osd request for the
2357 * original request, and release the old one.
2360 osd_req = rbd_osd_req_create_copyup(orig_request);
2363 rbd_osd_req_destroy(orig_request->osd_req);
2364 orig_request->osd_req = osd_req;
2365 orig_request->copyup_pages = pages;
2366 orig_request->copyup_page_count = page_count;
2368 /* Initialize the copyup op */
2370 osd_req_op_cls_init(osd_req, 0, CEPH_OSD_OP_CALL, "rbd", "copyup");
2371 osd_req_op_cls_request_data_pages(osd_req, 0, pages, parent_length, 0,
2374 /* Then the original write request op */
2376 offset = orig_request->offset;
2377 length = orig_request->length;
2378 osd_req_op_extent_init(osd_req, 1, CEPH_OSD_OP_WRITE,
2379 offset, length, 0, 0);
2380 if (orig_request->type == OBJ_REQUEST_BIO)
2381 osd_req_op_extent_osd_data_bio(osd_req, 1,
2382 orig_request->bio_list, length);
2384 osd_req_op_extent_osd_data_pages(osd_req, 1,
2385 orig_request->pages, length,
2386 offset & ~PAGE_MASK, false, false);
2388 rbd_osd_req_format_write(orig_request);
2390 /* All set, send it off. */
2392 orig_request->callback = rbd_img_obj_copyup_callback;
2393 osdc = &rbd_dev->rbd_client->client->osdc;
2394 result = rbd_obj_request_submit(osdc, orig_request);
2398 /* Record the error code and complete the request */
2400 orig_request->result = result;
2401 orig_request->xferred = 0;
2402 obj_request_done_set(orig_request);
2403 rbd_obj_request_complete(orig_request);
2407 * Read from the parent image the range of data that covers the
2408 * entire target of the given object request. This is used for
2409 * satisfying a layered image write request when the target of an
2410 * object request from the image request does not exist.
2412 * A page array big enough to hold the returned data is allocated
2413 * and supplied to rbd_img_request_fill() as the "data descriptor."
2414 * When the read completes, this page array will be transferred to
2415 * the original object request for the copyup operation.
2417 * If an error occurs, record it as the result of the original
2418 * object request and mark it done so it gets completed.
2420 static int rbd_img_obj_parent_read_full(struct rbd_obj_request *obj_request)
2422 struct rbd_img_request *img_request = NULL;
2423 struct rbd_img_request *parent_request = NULL;
2424 struct rbd_device *rbd_dev;
2427 struct page **pages = NULL;
2431 rbd_assert(obj_request_img_data_test(obj_request));
2432 rbd_assert(obj_request_type_valid(obj_request->type));
2434 img_request = obj_request->img_request;
2435 rbd_assert(img_request != NULL);
2436 rbd_dev = img_request->rbd_dev;
2437 rbd_assert(rbd_dev->parent != NULL);
2440 * Determine the byte range covered by the object in the
2441 * child image to which the original request was to be sent.
2443 img_offset = obj_request->img_offset - obj_request->offset;
2444 length = (u64)1 << rbd_dev->header.obj_order;
2447 * There is no defined parent data beyond the parent
2448 * overlap, so limit what we read at that boundary if
2451 if (img_offset + length > rbd_dev->parent_overlap) {
2452 rbd_assert(img_offset < rbd_dev->parent_overlap);
2453 length = rbd_dev->parent_overlap - img_offset;
2457 * Allocate a page array big enough to receive the data read
2460 page_count = (u32)calc_pages_for(0, length);
2461 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2462 if (IS_ERR(pages)) {
2463 result = PTR_ERR(pages);
2469 parent_request = rbd_parent_request_create(obj_request,
2470 img_offset, length);
2471 if (!parent_request)
2474 result = rbd_img_request_fill(parent_request, OBJ_REQUEST_PAGES, pages);
2477 parent_request->copyup_pages = pages;
2478 parent_request->copyup_page_count = page_count;
2480 parent_request->callback = rbd_img_obj_parent_read_full_callback;
2481 result = rbd_img_request_submit(parent_request);
2485 parent_request->copyup_pages = NULL;
2486 parent_request->copyup_page_count = 0;
2487 parent_request->obj_request = NULL;
2488 rbd_obj_request_put(obj_request);
2491 ceph_release_page_vector(pages, page_count);
2493 rbd_img_request_put(parent_request);
2494 obj_request->result = result;
2495 obj_request->xferred = 0;
2496 obj_request_done_set(obj_request);
2501 static void rbd_img_obj_exists_callback(struct rbd_obj_request *obj_request)
2503 struct rbd_obj_request *orig_request;
2506 rbd_assert(!obj_request_img_data_test(obj_request));
2509 * All we need from the object request is the original
2510 * request and the result of the STAT op. Grab those, then
2511 * we're done with the request.
2513 orig_request = obj_request->obj_request;
2514 obj_request->obj_request = NULL;
2515 rbd_assert(orig_request);
2516 rbd_assert(orig_request->img_request);
2518 result = obj_request->result;
2519 obj_request->result = 0;
2521 dout("%s: obj %p for obj %p result %d %llu/%llu\n", __func__,
2522 obj_request, orig_request, result,
2523 obj_request->xferred, obj_request->length);
2524 rbd_obj_request_put(obj_request);
2526 rbd_assert(orig_request);
2527 rbd_assert(orig_request->img_request);
2530 * Our only purpose here is to determine whether the object
2531 * exists, and we don't want to treat the non-existence as
2532 * an error. If something else comes back, transfer the
2533 * error to the original request and complete it now.
2536 obj_request_existence_set(orig_request, true);
2537 } else if (result == -ENOENT) {
2538 obj_request_existence_set(orig_request, false);
2539 } else if (result) {
2540 orig_request->result = result;
2545 * Resubmit the original request now that we have recorded
2546 * whether the target object exists.
2548 orig_request->result = rbd_img_obj_request_submit(orig_request);
2550 if (orig_request->result)
2551 rbd_obj_request_complete(orig_request);
2552 rbd_obj_request_put(orig_request);
2555 static int rbd_img_obj_exists_submit(struct rbd_obj_request *obj_request)
2557 struct rbd_obj_request *stat_request;
2558 struct rbd_device *rbd_dev;
2559 struct ceph_osd_client *osdc;
2560 struct page **pages = NULL;
2566 * The response data for a STAT call consists of:
2573 size = sizeof (__le64) + sizeof (__le32) + sizeof (__le32);
2574 page_count = (u32)calc_pages_for(0, size);
2575 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2577 return PTR_ERR(pages);
2580 stat_request = rbd_obj_request_create(obj_request->object_name, 0, 0,
2585 rbd_obj_request_get(obj_request);
2586 stat_request->obj_request = obj_request;
2587 stat_request->pages = pages;
2588 stat_request->page_count = page_count;
2590 rbd_assert(obj_request->img_request);
2591 rbd_dev = obj_request->img_request->rbd_dev;
2592 stat_request->osd_req = rbd_osd_req_create(rbd_dev, false,
2594 if (!stat_request->osd_req)
2596 stat_request->callback = rbd_img_obj_exists_callback;
2598 osd_req_op_init(stat_request->osd_req, 0, CEPH_OSD_OP_STAT);
2599 osd_req_op_raw_data_in_pages(stat_request->osd_req, 0, pages, size, 0,
2601 rbd_osd_req_format_read(stat_request);
2603 osdc = &rbd_dev->rbd_client->client->osdc;
2604 ret = rbd_obj_request_submit(osdc, stat_request);
2607 rbd_obj_request_put(obj_request);
2612 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request)
2614 struct rbd_img_request *img_request;
2615 struct rbd_device *rbd_dev;
2618 rbd_assert(obj_request_img_data_test(obj_request));
2620 img_request = obj_request->img_request;
2621 rbd_assert(img_request);
2622 rbd_dev = img_request->rbd_dev;
2625 * Only writes to layered images need special handling.
2626 * Reads and non-layered writes are simple object requests.
2627 * Layered writes that start beyond the end of the overlap
2628 * with the parent have no parent data, so they too are
2629 * simple object requests. Finally, if the target object is
2630 * known to already exist, its parent data has already been
2631 * copied, so a write to the object can also be handled as a
2632 * simple object request.
2634 if (!img_request_write_test(img_request) ||
2635 !img_request_layered_test(img_request) ||
2636 rbd_dev->parent_overlap <= obj_request->img_offset ||
2637 ((known = obj_request_known_test(obj_request)) &&
2638 obj_request_exists_test(obj_request))) {
2640 struct rbd_device *rbd_dev;
2641 struct ceph_osd_client *osdc;
2643 rbd_dev = obj_request->img_request->rbd_dev;
2644 osdc = &rbd_dev->rbd_client->client->osdc;
2646 return rbd_obj_request_submit(osdc, obj_request);
2650 * It's a layered write. The target object might exist but
2651 * we may not know that yet. If we know it doesn't exist,
2652 * start by reading the data for the full target object from
2653 * the parent so we can use it for a copyup to the target.
2656 return rbd_img_obj_parent_read_full(obj_request);
2658 /* We don't know whether the target exists. Go find out. */
2660 return rbd_img_obj_exists_submit(obj_request);
2663 static int rbd_img_request_submit(struct rbd_img_request *img_request)
2665 struct rbd_obj_request *obj_request;
2666 struct rbd_obj_request *next_obj_request;
2668 dout("%s: img %p\n", __func__, img_request);
2669 for_each_obj_request_safe(img_request, obj_request, next_obj_request) {
2672 ret = rbd_img_obj_request_submit(obj_request);
2680 static void rbd_img_parent_read_callback(struct rbd_img_request *img_request)
2682 struct rbd_obj_request *obj_request;
2683 struct rbd_device *rbd_dev;
2688 rbd_assert(img_request_child_test(img_request));
2690 /* First get what we need from the image request and release it */
2692 obj_request = img_request->obj_request;
2693 img_xferred = img_request->xferred;
2694 img_result = img_request->result;
2695 rbd_img_request_put(img_request);
2698 * If the overlap has become 0 (most likely because the
2699 * image has been flattened) we need to re-submit the
2702 rbd_assert(obj_request);
2703 rbd_assert(obj_request->img_request);
2704 rbd_dev = obj_request->img_request->rbd_dev;
2705 if (!rbd_dev->parent_overlap) {
2706 struct ceph_osd_client *osdc;
2708 osdc = &rbd_dev->rbd_client->client->osdc;
2709 img_result = rbd_obj_request_submit(osdc, obj_request);
2714 obj_request->result = img_result;
2715 if (obj_request->result)
2719 * We need to zero anything beyond the parent overlap
2720 * boundary. Since rbd_img_obj_request_read_callback()
2721 * will zero anything beyond the end of a short read, an
2722 * easy way to do this is to pretend the data from the
2723 * parent came up short--ending at the overlap boundary.
2725 rbd_assert(obj_request->img_offset < U64_MAX - obj_request->length);
2726 obj_end = obj_request->img_offset + obj_request->length;
2727 if (obj_end > rbd_dev->parent_overlap) {
2730 if (obj_request->img_offset < rbd_dev->parent_overlap)
2731 xferred = rbd_dev->parent_overlap -
2732 obj_request->img_offset;
2734 obj_request->xferred = min(img_xferred, xferred);
2736 obj_request->xferred = img_xferred;
2739 rbd_img_obj_request_read_callback(obj_request);
2740 rbd_obj_request_complete(obj_request);
2743 static void rbd_img_parent_read(struct rbd_obj_request *obj_request)
2745 struct rbd_img_request *img_request;
2748 rbd_assert(obj_request_img_data_test(obj_request));
2749 rbd_assert(obj_request->img_request != NULL);
2750 rbd_assert(obj_request->result == (s32) -ENOENT);
2751 rbd_assert(obj_request_type_valid(obj_request->type));
2753 /* rbd_read_finish(obj_request, obj_request->length); */
2754 img_request = rbd_parent_request_create(obj_request,
2755 obj_request->img_offset,
2756 obj_request->length);
2761 if (obj_request->type == OBJ_REQUEST_BIO)
2762 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
2763 obj_request->bio_list);
2765 result = rbd_img_request_fill(img_request, OBJ_REQUEST_PAGES,
2766 obj_request->pages);
2770 img_request->callback = rbd_img_parent_read_callback;
2771 result = rbd_img_request_submit(img_request);
2778 rbd_img_request_put(img_request);
2779 obj_request->result = result;
2780 obj_request->xferred = 0;
2781 obj_request_done_set(obj_request);
2784 static int rbd_obj_notify_ack(struct rbd_device *rbd_dev, u64 notify_id)
2786 struct rbd_obj_request *obj_request;
2787 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2790 obj_request = rbd_obj_request_create(rbd_dev->header_name, 0, 0,
2791 OBJ_REQUEST_NODATA);
2796 obj_request->osd_req = rbd_osd_req_create(rbd_dev, false, obj_request);
2797 if (!obj_request->osd_req)
2799 obj_request->callback = rbd_obj_request_put;
2801 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_NOTIFY_ACK,
2803 rbd_osd_req_format_read(obj_request);
2805 ret = rbd_obj_request_submit(osdc, obj_request);
2808 rbd_obj_request_put(obj_request);
2813 static void rbd_watch_cb(u64 ver, u64 notify_id, u8 opcode, void *data)
2815 struct rbd_device *rbd_dev = (struct rbd_device *)data;
2821 dout("%s: \"%s\" notify_id %llu opcode %u\n", __func__,
2822 rbd_dev->header_name, (unsigned long long)notify_id,
2823 (unsigned int)opcode);
2824 ret = rbd_dev_refresh(rbd_dev);
2826 rbd_warn(rbd_dev, ": header refresh error (%d)\n", ret);
2828 rbd_obj_notify_ack(rbd_dev, notify_id);
2832 * Request sync osd watch/unwatch. The value of "start" determines
2833 * whether a watch request is being initiated or torn down.
2835 static int rbd_dev_header_watch_sync(struct rbd_device *rbd_dev, bool start)
2837 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2838 struct rbd_obj_request *obj_request;
2841 rbd_assert(start ^ !!rbd_dev->watch_event);
2842 rbd_assert(start ^ !!rbd_dev->watch_request);
2845 ret = ceph_osdc_create_event(osdc, rbd_watch_cb, rbd_dev,
2846 &rbd_dev->watch_event);
2849 rbd_assert(rbd_dev->watch_event != NULL);
2853 obj_request = rbd_obj_request_create(rbd_dev->header_name, 0, 0,
2854 OBJ_REQUEST_NODATA);
2858 obj_request->osd_req = rbd_osd_req_create(rbd_dev, true, obj_request);
2859 if (!obj_request->osd_req)
2863 ceph_osdc_set_request_linger(osdc, obj_request->osd_req);
2865 ceph_osdc_unregister_linger_request(osdc,
2866 rbd_dev->watch_request->osd_req);
2868 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_WATCH,
2869 rbd_dev->watch_event->cookie, 0, start ? 1 : 0);
2870 rbd_osd_req_format_write(obj_request);
2872 ret = rbd_obj_request_submit(osdc, obj_request);
2875 ret = rbd_obj_request_wait(obj_request);
2878 ret = obj_request->result;
2883 * A watch request is set to linger, so the underlying osd
2884 * request won't go away until we unregister it. We retain
2885 * a pointer to the object request during that time (in
2886 * rbd_dev->watch_request), so we'll keep a reference to
2887 * it. We'll drop that reference (below) after we've
2891 rbd_dev->watch_request = obj_request;
2896 /* We have successfully torn down the watch request */
2898 rbd_obj_request_put(rbd_dev->watch_request);
2899 rbd_dev->watch_request = NULL;
2901 /* Cancel the event if we're tearing down, or on error */
2902 ceph_osdc_cancel_event(rbd_dev->watch_event);
2903 rbd_dev->watch_event = NULL;
2905 rbd_obj_request_put(obj_request);
2911 * Synchronous osd object method call. Returns the number of bytes
2912 * returned in the outbound buffer, or a negative error code.
2914 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
2915 const char *object_name,
2916 const char *class_name,
2917 const char *method_name,
2918 const void *outbound,
2919 size_t outbound_size,
2921 size_t inbound_size)
2923 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2924 struct rbd_obj_request *obj_request;
2925 struct page **pages;
2930 * Method calls are ultimately read operations. The result
2931 * should placed into the inbound buffer provided. They
2932 * also supply outbound data--parameters for the object
2933 * method. Currently if this is present it will be a
2936 page_count = (u32)calc_pages_for(0, inbound_size);
2937 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2939 return PTR_ERR(pages);
2942 obj_request = rbd_obj_request_create(object_name, 0, inbound_size,
2947 obj_request->pages = pages;
2948 obj_request->page_count = page_count;
2950 obj_request->osd_req = rbd_osd_req_create(rbd_dev, false, obj_request);
2951 if (!obj_request->osd_req)
2954 osd_req_op_cls_init(obj_request->osd_req, 0, CEPH_OSD_OP_CALL,
2955 class_name, method_name);
2956 if (outbound_size) {
2957 struct ceph_pagelist *pagelist;
2959 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
2963 ceph_pagelist_init(pagelist);
2964 ceph_pagelist_append(pagelist, outbound, outbound_size);
2965 osd_req_op_cls_request_data_pagelist(obj_request->osd_req, 0,
2968 osd_req_op_cls_response_data_pages(obj_request->osd_req, 0,
2969 obj_request->pages, inbound_size,
2971 rbd_osd_req_format_read(obj_request);
2973 ret = rbd_obj_request_submit(osdc, obj_request);
2976 ret = rbd_obj_request_wait(obj_request);
2980 ret = obj_request->result;
2984 rbd_assert(obj_request->xferred < (u64)INT_MAX);
2985 ret = (int)obj_request->xferred;
2986 ceph_copy_from_page_vector(pages, inbound, 0, obj_request->xferred);
2989 rbd_obj_request_put(obj_request);
2991 ceph_release_page_vector(pages, page_count);
2996 static void rbd_request_fn(struct request_queue *q)
2997 __releases(q->queue_lock) __acquires(q->queue_lock)
2999 struct rbd_device *rbd_dev = q->queuedata;
3000 bool read_only = rbd_dev->mapping.read_only;
3004 while ((rq = blk_fetch_request(q))) {
3005 bool write_request = rq_data_dir(rq) == WRITE;
3006 struct rbd_img_request *img_request;
3010 /* Ignore any non-FS requests that filter through. */
3012 if (rq->cmd_type != REQ_TYPE_FS) {
3013 dout("%s: non-fs request type %d\n", __func__,
3014 (int) rq->cmd_type);
3015 __blk_end_request_all(rq, 0);
3019 /* Ignore/skip any zero-length requests */
3021 offset = (u64) blk_rq_pos(rq) << SECTOR_SHIFT;
3022 length = (u64) blk_rq_bytes(rq);
3025 dout("%s: zero-length request\n", __func__);
3026 __blk_end_request_all(rq, 0);
3030 spin_unlock_irq(q->queue_lock);
3032 /* Disallow writes to a read-only device */
3034 if (write_request) {
3038 rbd_assert(rbd_dev->spec->snap_id == CEPH_NOSNAP);
3042 * Quit early if the mapped snapshot no longer
3043 * exists. It's still possible the snapshot will
3044 * have disappeared by the time our request arrives
3045 * at the osd, but there's no sense in sending it if
3048 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
3049 dout("request for non-existent snapshot");
3050 rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
3056 if (offset && length > U64_MAX - offset + 1) {
3057 rbd_warn(rbd_dev, "bad request range (%llu~%llu)\n",
3059 goto end_request; /* Shouldn't happen */
3063 if (offset + length > rbd_dev->mapping.size) {
3064 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)\n",
3065 offset, length, rbd_dev->mapping.size);
3070 img_request = rbd_img_request_create(rbd_dev, offset, length,
3075 img_request->rq = rq;
3077 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3080 result = rbd_img_request_submit(img_request);
3082 rbd_img_request_put(img_request);
3084 spin_lock_irq(q->queue_lock);
3086 rbd_warn(rbd_dev, "%s %llx at %llx result %d\n",
3087 write_request ? "write" : "read",
3088 length, offset, result);
3090 __blk_end_request_all(rq, result);
3096 * a queue callback. Makes sure that we don't create a bio that spans across
3097 * multiple osd objects. One exception would be with a single page bios,
3098 * which we handle later at bio_chain_clone_range()
3100 static int rbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bmd,
3101 struct bio_vec *bvec)
3103 struct rbd_device *rbd_dev = q->queuedata;
3104 sector_t sector_offset;
3105 sector_t sectors_per_obj;
3106 sector_t obj_sector_offset;
3110 * Find how far into its rbd object the partition-relative
3111 * bio start sector is to offset relative to the enclosing
3114 sector_offset = get_start_sect(bmd->bi_bdev) + bmd->bi_sector;
3115 sectors_per_obj = 1 << (rbd_dev->header.obj_order - SECTOR_SHIFT);
3116 obj_sector_offset = sector_offset & (sectors_per_obj - 1);
3119 * Compute the number of bytes from that offset to the end
3120 * of the object. Account for what's already used by the bio.
3122 ret = (int) (sectors_per_obj - obj_sector_offset) << SECTOR_SHIFT;
3123 if (ret > bmd->bi_size)
3124 ret -= bmd->bi_size;
3129 * Don't send back more than was asked for. And if the bio
3130 * was empty, let the whole thing through because: "Note
3131 * that a block device *must* allow a single page to be
3132 * added to an empty bio."
3134 rbd_assert(bvec->bv_len <= PAGE_SIZE);
3135 if (ret > (int) bvec->bv_len || !bmd->bi_size)
3136 ret = (int) bvec->bv_len;
3141 static void rbd_free_disk(struct rbd_device *rbd_dev)
3143 struct gendisk *disk = rbd_dev->disk;
3148 rbd_dev->disk = NULL;
3149 if (disk->flags & GENHD_FL_UP) {
3152 blk_cleanup_queue(disk->queue);
3157 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
3158 const char *object_name,
3159 u64 offset, u64 length, void *buf)
3162 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3163 struct rbd_obj_request *obj_request;
3164 struct page **pages = NULL;
3169 page_count = (u32) calc_pages_for(offset, length);
3170 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
3172 ret = PTR_ERR(pages);
3175 obj_request = rbd_obj_request_create(object_name, offset, length,
3180 obj_request->pages = pages;
3181 obj_request->page_count = page_count;
3183 obj_request->osd_req = rbd_osd_req_create(rbd_dev, false, obj_request);
3184 if (!obj_request->osd_req)
3187 osd_req_op_extent_init(obj_request->osd_req, 0, CEPH_OSD_OP_READ,
3188 offset, length, 0, 0);
3189 osd_req_op_extent_osd_data_pages(obj_request->osd_req, 0,
3191 obj_request->length,
3192 obj_request->offset & ~PAGE_MASK,
3194 rbd_osd_req_format_read(obj_request);
3196 ret = rbd_obj_request_submit(osdc, obj_request);
3199 ret = rbd_obj_request_wait(obj_request);
3203 ret = obj_request->result;
3207 rbd_assert(obj_request->xferred <= (u64) SIZE_MAX);
3208 size = (size_t) obj_request->xferred;
3209 ceph_copy_from_page_vector(pages, buf, 0, size);
3210 rbd_assert(size <= (size_t)INT_MAX);
3214 rbd_obj_request_put(obj_request);
3216 ceph_release_page_vector(pages, page_count);
3222 * Read the complete header for the given rbd device. On successful
3223 * return, the rbd_dev->header field will contain up-to-date
3224 * information about the image.
3226 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev)
3228 struct rbd_image_header_ondisk *ondisk = NULL;
3235 * The complete header will include an array of its 64-bit
3236 * snapshot ids, followed by the names of those snapshots as
3237 * a contiguous block of NUL-terminated strings. Note that
3238 * the number of snapshots could change by the time we read
3239 * it in, in which case we re-read it.
3246 size = sizeof (*ondisk);
3247 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
3249 ondisk = kmalloc(size, GFP_KERNEL);
3253 ret = rbd_obj_read_sync(rbd_dev, rbd_dev->header_name,
3257 if ((size_t)ret < size) {
3259 rbd_warn(rbd_dev, "short header read (want %zd got %d)",
3263 if (!rbd_dev_ondisk_valid(ondisk)) {
3265 rbd_warn(rbd_dev, "invalid header");
3269 names_size = le64_to_cpu(ondisk->snap_names_len);
3270 want_count = snap_count;
3271 snap_count = le32_to_cpu(ondisk->snap_count);
3272 } while (snap_count != want_count);
3274 ret = rbd_header_from_disk(rbd_dev, ondisk);
3282 * Clear the rbd device's EXISTS flag if the snapshot it's mapped to
3283 * has disappeared from the (just updated) snapshot context.
3285 static void rbd_exists_validate(struct rbd_device *rbd_dev)
3289 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags))
3292 snap_id = rbd_dev->spec->snap_id;
3293 if (snap_id == CEPH_NOSNAP)
3296 if (rbd_dev_snap_index(rbd_dev, snap_id) == BAD_SNAP_INDEX)
3297 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
3300 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
3305 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
3306 mapping_size = rbd_dev->mapping.size;
3307 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
3308 if (rbd_dev->image_format == 1)
3309 ret = rbd_dev_v1_header_info(rbd_dev);
3311 ret = rbd_dev_v2_header_info(rbd_dev);
3313 /* If it's a mapped snapshot, validate its EXISTS flag */
3315 rbd_exists_validate(rbd_dev);
3316 mutex_unlock(&ctl_mutex);
3317 if (mapping_size != rbd_dev->mapping.size) {
3320 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
3321 dout("setting size to %llu sectors", (unsigned long long)size);
3322 set_capacity(rbd_dev->disk, size);
3323 revalidate_disk(rbd_dev->disk);
3329 static int rbd_init_disk(struct rbd_device *rbd_dev)
3331 struct gendisk *disk;
3332 struct request_queue *q;
3335 /* create gendisk info */
3336 disk = alloc_disk(RBD_MINORS_PER_MAJOR);
3340 snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
3342 disk->major = rbd_dev->major;
3343 disk->first_minor = 0;
3344 disk->fops = &rbd_bd_ops;
3345 disk->private_data = rbd_dev;
3347 q = blk_init_queue(rbd_request_fn, &rbd_dev->lock);
3351 /* We use the default size, but let's be explicit about it. */
3352 blk_queue_physical_block_size(q, SECTOR_SIZE);
3354 /* set io sizes to object size */
3355 segment_size = rbd_obj_bytes(&rbd_dev->header);
3356 blk_queue_max_hw_sectors(q, segment_size / SECTOR_SIZE);
3357 blk_queue_max_segment_size(q, segment_size);
3358 blk_queue_io_min(q, segment_size);
3359 blk_queue_io_opt(q, segment_size);
3361 blk_queue_merge_bvec(q, rbd_merge_bvec);
3364 q->queuedata = rbd_dev;
3366 rbd_dev->disk = disk;
3379 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
3381 return container_of(dev, struct rbd_device, dev);
3384 static ssize_t rbd_size_show(struct device *dev,
3385 struct device_attribute *attr, char *buf)
3387 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3389 return sprintf(buf, "%llu\n",
3390 (unsigned long long)rbd_dev->mapping.size);
3394 * Note this shows the features for whatever's mapped, which is not
3395 * necessarily the base image.
3397 static ssize_t rbd_features_show(struct device *dev,
3398 struct device_attribute *attr, char *buf)
3400 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3402 return sprintf(buf, "0x%016llx\n",
3403 (unsigned long long)rbd_dev->mapping.features);
3406 static ssize_t rbd_major_show(struct device *dev,
3407 struct device_attribute *attr, char *buf)
3409 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3412 return sprintf(buf, "%d\n", rbd_dev->major);
3414 return sprintf(buf, "(none)\n");
3418 static ssize_t rbd_client_id_show(struct device *dev,
3419 struct device_attribute *attr, char *buf)
3421 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3423 return sprintf(buf, "client%lld\n",
3424 ceph_client_id(rbd_dev->rbd_client->client));
3427 static ssize_t rbd_pool_show(struct device *dev,
3428 struct device_attribute *attr, char *buf)
3430 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3432 return sprintf(buf, "%s\n", rbd_dev->spec->pool_name);
3435 static ssize_t rbd_pool_id_show(struct device *dev,
3436 struct device_attribute *attr, char *buf)
3438 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3440 return sprintf(buf, "%llu\n",
3441 (unsigned long long) rbd_dev->spec->pool_id);
3444 static ssize_t rbd_name_show(struct device *dev,
3445 struct device_attribute *attr, char *buf)
3447 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3449 if (rbd_dev->spec->image_name)
3450 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
3452 return sprintf(buf, "(unknown)\n");
3455 static ssize_t rbd_image_id_show(struct device *dev,
3456 struct device_attribute *attr, char *buf)
3458 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3460 return sprintf(buf, "%s\n", rbd_dev->spec->image_id);
3464 * Shows the name of the currently-mapped snapshot (or
3465 * RBD_SNAP_HEAD_NAME for the base image).
3467 static ssize_t rbd_snap_show(struct device *dev,
3468 struct device_attribute *attr,
3471 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3473 return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
3477 * For an rbd v2 image, shows the pool id, image id, and snapshot id
3478 * for the parent image. If there is no parent, simply shows
3479 * "(no parent image)".
3481 static ssize_t rbd_parent_show(struct device *dev,
3482 struct device_attribute *attr,
3485 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3486 struct rbd_spec *spec = rbd_dev->parent_spec;
3491 return sprintf(buf, "(no parent image)\n");
3493 count = sprintf(bufp, "pool_id %llu\npool_name %s\n",
3494 (unsigned long long) spec->pool_id, spec->pool_name);
3499 count = sprintf(bufp, "image_id %s\nimage_name %s\n", spec->image_id,
3500 spec->image_name ? spec->image_name : "(unknown)");
3505 count = sprintf(bufp, "snap_id %llu\nsnap_name %s\n",
3506 (unsigned long long) spec->snap_id, spec->snap_name);
3511 count = sprintf(bufp, "overlap %llu\n", rbd_dev->parent_overlap);
3516 return (ssize_t) (bufp - buf);
3519 static ssize_t rbd_image_refresh(struct device *dev,
3520 struct device_attribute *attr,
3524 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3527 ret = rbd_dev_refresh(rbd_dev);
3529 rbd_warn(rbd_dev, ": manual header refresh error (%d)\n", ret);
3531 return ret < 0 ? ret : size;
3534 static DEVICE_ATTR(size, S_IRUGO, rbd_size_show, NULL);
3535 static DEVICE_ATTR(features, S_IRUGO, rbd_features_show, NULL);
3536 static DEVICE_ATTR(major, S_IRUGO, rbd_major_show, NULL);
3537 static DEVICE_ATTR(client_id, S_IRUGO, rbd_client_id_show, NULL);
3538 static DEVICE_ATTR(pool, S_IRUGO, rbd_pool_show, NULL);
3539 static DEVICE_ATTR(pool_id, S_IRUGO, rbd_pool_id_show, NULL);
3540 static DEVICE_ATTR(name, S_IRUGO, rbd_name_show, NULL);
3541 static DEVICE_ATTR(image_id, S_IRUGO, rbd_image_id_show, NULL);
3542 static DEVICE_ATTR(refresh, S_IWUSR, NULL, rbd_image_refresh);
3543 static DEVICE_ATTR(current_snap, S_IRUGO, rbd_snap_show, NULL);
3544 static DEVICE_ATTR(parent, S_IRUGO, rbd_parent_show, NULL);
3546 static struct attribute *rbd_attrs[] = {
3547 &dev_attr_size.attr,
3548 &dev_attr_features.attr,
3549 &dev_attr_major.attr,
3550 &dev_attr_client_id.attr,
3551 &dev_attr_pool.attr,
3552 &dev_attr_pool_id.attr,
3553 &dev_attr_name.attr,
3554 &dev_attr_image_id.attr,
3555 &dev_attr_current_snap.attr,
3556 &dev_attr_parent.attr,
3557 &dev_attr_refresh.attr,
3561 static struct attribute_group rbd_attr_group = {
3565 static const struct attribute_group *rbd_attr_groups[] = {
3570 static void rbd_sysfs_dev_release(struct device *dev)
3574 static struct device_type rbd_device_type = {
3576 .groups = rbd_attr_groups,
3577 .release = rbd_sysfs_dev_release,
3580 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
3582 kref_get(&spec->kref);
3587 static void rbd_spec_free(struct kref *kref);
3588 static void rbd_spec_put(struct rbd_spec *spec)
3591 kref_put(&spec->kref, rbd_spec_free);
3594 static struct rbd_spec *rbd_spec_alloc(void)
3596 struct rbd_spec *spec;
3598 spec = kzalloc(sizeof (*spec), GFP_KERNEL);
3601 kref_init(&spec->kref);
3606 static void rbd_spec_free(struct kref *kref)
3608 struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
3610 kfree(spec->pool_name);
3611 kfree(spec->image_id);
3612 kfree(spec->image_name);
3613 kfree(spec->snap_name);
3617 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
3618 struct rbd_spec *spec)
3620 struct rbd_device *rbd_dev;
3622 rbd_dev = kzalloc(sizeof (*rbd_dev), GFP_KERNEL);
3626 spin_lock_init(&rbd_dev->lock);
3628 atomic_set(&rbd_dev->parent_ref, 0);
3629 INIT_LIST_HEAD(&rbd_dev->node);
3630 init_rwsem(&rbd_dev->header_rwsem);
3632 rbd_dev->spec = spec;
3633 rbd_dev->rbd_client = rbdc;
3635 /* Initialize the layout used for all rbd requests */
3637 rbd_dev->layout.fl_stripe_unit = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
3638 rbd_dev->layout.fl_stripe_count = cpu_to_le32(1);
3639 rbd_dev->layout.fl_object_size = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
3640 rbd_dev->layout.fl_pg_pool = cpu_to_le32((u32) spec->pool_id);
3645 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
3647 rbd_put_client(rbd_dev->rbd_client);
3648 rbd_spec_put(rbd_dev->spec);
3653 * Get the size and object order for an image snapshot, or if
3654 * snap_id is CEPH_NOSNAP, gets this information for the base
3657 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
3658 u8 *order, u64 *snap_size)
3660 __le64 snapid = cpu_to_le64(snap_id);
3665 } __attribute__ ((packed)) size_buf = { 0 };
3667 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3669 &snapid, sizeof (snapid),
3670 &size_buf, sizeof (size_buf));
3671 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3674 if (ret < sizeof (size_buf))
3678 *order = size_buf.order;
3679 *snap_size = le64_to_cpu(size_buf.size);
3681 dout(" snap_id 0x%016llx order = %u, snap_size = %llu\n",
3682 (unsigned long long)snap_id, (unsigned int)*order,
3683 (unsigned long long)*snap_size);
3688 static int rbd_dev_v2_image_size(struct rbd_device *rbd_dev)
3690 return _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
3691 &rbd_dev->header.obj_order,
3692 &rbd_dev->header.image_size);
3695 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev)
3701 reply_buf = kzalloc(RBD_OBJ_PREFIX_LEN_MAX, GFP_KERNEL);
3705 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3706 "rbd", "get_object_prefix", NULL, 0,
3707 reply_buf, RBD_OBJ_PREFIX_LEN_MAX);
3708 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3713 rbd_dev->header.object_prefix = ceph_extract_encoded_string(&p,
3714 p + ret, NULL, GFP_NOIO);
3717 if (IS_ERR(rbd_dev->header.object_prefix)) {
3718 ret = PTR_ERR(rbd_dev->header.object_prefix);
3719 rbd_dev->header.object_prefix = NULL;
3721 dout(" object_prefix = %s\n", rbd_dev->header.object_prefix);
3729 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
3732 __le64 snapid = cpu_to_le64(snap_id);
3736 } __attribute__ ((packed)) features_buf = { 0 };
3740 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3741 "rbd", "get_features",
3742 &snapid, sizeof (snapid),
3743 &features_buf, sizeof (features_buf));
3744 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3747 if (ret < sizeof (features_buf))
3750 incompat = le64_to_cpu(features_buf.incompat);
3751 if (incompat & ~RBD_FEATURES_SUPPORTED)
3754 *snap_features = le64_to_cpu(features_buf.features);
3756 dout(" snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
3757 (unsigned long long)snap_id,
3758 (unsigned long long)*snap_features,
3759 (unsigned long long)le64_to_cpu(features_buf.incompat));
3764 static int rbd_dev_v2_features(struct rbd_device *rbd_dev)
3766 return _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
3767 &rbd_dev->header.features);
3770 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev)
3772 struct rbd_spec *parent_spec;
3774 void *reply_buf = NULL;
3783 parent_spec = rbd_spec_alloc();
3787 size = sizeof (__le64) + /* pool_id */
3788 sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX + /* image_id */
3789 sizeof (__le64) + /* snap_id */
3790 sizeof (__le64); /* overlap */
3791 reply_buf = kmalloc(size, GFP_KERNEL);
3797 snapid = cpu_to_le64(CEPH_NOSNAP);
3798 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3799 "rbd", "get_parent",
3800 &snapid, sizeof (snapid),
3802 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3807 end = reply_buf + ret;
3809 ceph_decode_64_safe(&p, end, pool_id, out_err);
3810 if (pool_id == CEPH_NOPOOL) {
3812 * Either the parent never existed, or we have
3813 * record of it but the image got flattened so it no
3814 * longer has a parent. When the parent of a
3815 * layered image disappears we immediately set the
3816 * overlap to 0. The effect of this is that all new
3817 * requests will be treated as if the image had no
3820 if (rbd_dev->parent_overlap) {
3821 rbd_dev->parent_overlap = 0;
3823 rbd_dev_parent_put(rbd_dev);
3824 pr_info("%s: clone image has been flattened\n",
3825 rbd_dev->disk->disk_name);
3828 goto out; /* No parent? No problem. */
3831 /* The ceph file layout needs to fit pool id in 32 bits */
3834 if (pool_id > (u64)U32_MAX) {
3835 rbd_warn(NULL, "parent pool id too large (%llu > %u)\n",
3836 (unsigned long long)pool_id, U32_MAX);
3839 parent_spec->pool_id = pool_id;
3841 image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
3842 if (IS_ERR(image_id)) {
3843 ret = PTR_ERR(image_id);
3846 parent_spec->image_id = image_id;
3847 ceph_decode_64_safe(&p, end, parent_spec->snap_id, out_err);
3848 ceph_decode_64_safe(&p, end, overlap, out_err);
3851 rbd_spec_put(rbd_dev->parent_spec);
3852 rbd_dev->parent_spec = parent_spec;
3853 parent_spec = NULL; /* rbd_dev now owns this */
3854 rbd_dev->parent_overlap = overlap;
3856 rbd_warn(rbd_dev, "ignoring parent of clone with overlap 0\n");
3862 rbd_spec_put(parent_spec);
3867 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev)
3871 __le64 stripe_count;
3872 } __attribute__ ((packed)) striping_info_buf = { 0 };
3873 size_t size = sizeof (striping_info_buf);
3880 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3881 "rbd", "get_stripe_unit_count", NULL, 0,
3882 (char *)&striping_info_buf, size);
3883 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3890 * We don't actually support the "fancy striping" feature
3891 * (STRIPINGV2) yet, but if the striping sizes are the
3892 * defaults the behavior is the same as before. So find
3893 * out, and only fail if the image has non-default values.
3896 obj_size = (u64)1 << rbd_dev->header.obj_order;
3897 p = &striping_info_buf;
3898 stripe_unit = ceph_decode_64(&p);
3899 if (stripe_unit != obj_size) {
3900 rbd_warn(rbd_dev, "unsupported stripe unit "
3901 "(got %llu want %llu)",
3902 stripe_unit, obj_size);
3905 stripe_count = ceph_decode_64(&p);
3906 if (stripe_count != 1) {
3907 rbd_warn(rbd_dev, "unsupported stripe count "
3908 "(got %llu want 1)", stripe_count);
3911 rbd_dev->header.stripe_unit = stripe_unit;
3912 rbd_dev->header.stripe_count = stripe_count;
3917 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
3919 size_t image_id_size;
3924 void *reply_buf = NULL;
3926 char *image_name = NULL;
3929 rbd_assert(!rbd_dev->spec->image_name);
3931 len = strlen(rbd_dev->spec->image_id);
3932 image_id_size = sizeof (__le32) + len;
3933 image_id = kmalloc(image_id_size, GFP_KERNEL);
3938 end = image_id + image_id_size;
3939 ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
3941 size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
3942 reply_buf = kmalloc(size, GFP_KERNEL);
3946 ret = rbd_obj_method_sync(rbd_dev, RBD_DIRECTORY,
3947 "rbd", "dir_get_name",
3948 image_id, image_id_size,
3953 end = reply_buf + ret;
3955 image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
3956 if (IS_ERR(image_name))
3959 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
3967 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
3969 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
3970 const char *snap_name;
3973 /* Skip over names until we find the one we are looking for */
3975 snap_name = rbd_dev->header.snap_names;
3976 while (which < snapc->num_snaps) {
3977 if (!strcmp(name, snap_name))
3978 return snapc->snaps[which];
3979 snap_name += strlen(snap_name) + 1;
3985 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
3987 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
3992 for (which = 0; !found && which < snapc->num_snaps; which++) {
3993 const char *snap_name;
3995 snap_id = snapc->snaps[which];
3996 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
3997 if (IS_ERR(snap_name))
3999 found = !strcmp(name, snap_name);
4002 return found ? snap_id : CEPH_NOSNAP;
4006 * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
4007 * no snapshot by that name is found, or if an error occurs.
4009 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4011 if (rbd_dev->image_format == 1)
4012 return rbd_v1_snap_id_by_name(rbd_dev, name);
4014 return rbd_v2_snap_id_by_name(rbd_dev, name);
4018 * When an rbd image has a parent image, it is identified by the
4019 * pool, image, and snapshot ids (not names). This function fills
4020 * in the names for those ids. (It's OK if we can't figure out the
4021 * name for an image id, but the pool and snapshot ids should always
4022 * exist and have names.) All names in an rbd spec are dynamically
4025 * When an image being mapped (not a parent) is probed, we have the
4026 * pool name and pool id, image name and image id, and the snapshot
4027 * name. The only thing we're missing is the snapshot id.
4029 static int rbd_dev_spec_update(struct rbd_device *rbd_dev)
4031 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4032 struct rbd_spec *spec = rbd_dev->spec;
4033 const char *pool_name;
4034 const char *image_name;
4035 const char *snap_name;
4039 * An image being mapped will have the pool name (etc.), but
4040 * we need to look up the snapshot id.
4042 if (spec->pool_name) {
4043 if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
4046 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
4047 if (snap_id == CEPH_NOSNAP)
4049 spec->snap_id = snap_id;
4051 spec->snap_id = CEPH_NOSNAP;
4057 /* Get the pool name; we have to make our own copy of this */
4059 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
4061 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
4064 pool_name = kstrdup(pool_name, GFP_KERNEL);
4068 /* Fetch the image name; tolerate failure here */
4070 image_name = rbd_dev_image_name(rbd_dev);
4072 rbd_warn(rbd_dev, "unable to get image name");
4074 /* Look up the snapshot name, and make a copy */
4076 snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
4082 spec->pool_name = pool_name;
4083 spec->image_name = image_name;
4084 spec->snap_name = snap_name;
4094 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev)
4103 struct ceph_snap_context *snapc;
4107 * We'll need room for the seq value (maximum snapshot id),
4108 * snapshot count, and array of that many snapshot ids.
4109 * For now we have a fixed upper limit on the number we're
4110 * prepared to receive.
4112 size = sizeof (__le64) + sizeof (__le32) +
4113 RBD_MAX_SNAP_COUNT * sizeof (__le64);
4114 reply_buf = kzalloc(size, GFP_KERNEL);
4118 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4119 "rbd", "get_snapcontext", NULL, 0,
4121 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4126 end = reply_buf + ret;
4128 ceph_decode_64_safe(&p, end, seq, out);
4129 ceph_decode_32_safe(&p, end, snap_count, out);
4132 * Make sure the reported number of snapshot ids wouldn't go
4133 * beyond the end of our buffer. But before checking that,
4134 * make sure the computed size of the snapshot context we
4135 * allocate is representable in a size_t.
4137 if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
4142 if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
4146 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
4152 for (i = 0; i < snap_count; i++)
4153 snapc->snaps[i] = ceph_decode_64(&p);
4155 ceph_put_snap_context(rbd_dev->header.snapc);
4156 rbd_dev->header.snapc = snapc;
4158 dout(" snap context seq = %llu, snap_count = %u\n",
4159 (unsigned long long)seq, (unsigned int)snap_count);
4166 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
4177 size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
4178 reply_buf = kmalloc(size, GFP_KERNEL);
4180 return ERR_PTR(-ENOMEM);
4182 snapid = cpu_to_le64(snap_id);
4183 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4184 "rbd", "get_snapshot_name",
4185 &snapid, sizeof (snapid),
4187 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4189 snap_name = ERR_PTR(ret);
4194 end = reply_buf + ret;
4195 snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
4196 if (IS_ERR(snap_name))
4199 dout(" snap_id 0x%016llx snap_name = %s\n",
4200 (unsigned long long)snap_id, snap_name);
4207 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev)
4209 bool first_time = rbd_dev->header.object_prefix == NULL;
4212 down_write(&rbd_dev->header_rwsem);
4215 ret = rbd_dev_v2_header_onetime(rbd_dev);
4221 * If the image supports layering, get the parent info. We
4222 * need to probe the first time regardless. Thereafter we
4223 * only need to if there's a parent, to see if it has
4224 * disappeared due to the mapped image getting flattened.
4226 if (rbd_dev->header.features & RBD_FEATURE_LAYERING &&
4227 (first_time || rbd_dev->parent_spec)) {
4230 ret = rbd_dev_v2_parent_info(rbd_dev);
4235 * Print a warning if this is the initial probe and
4236 * the image has a parent. Don't print it if the
4237 * image now being probed is itself a parent. We
4238 * can tell at this point because we won't know its
4239 * pool name yet (just its pool id).
4241 warn = rbd_dev->parent_spec && rbd_dev->spec->pool_name;
4242 if (first_time && warn)
4243 rbd_warn(rbd_dev, "WARNING: kernel layering "
4244 "is EXPERIMENTAL!");
4247 ret = rbd_dev_v2_image_size(rbd_dev);
4251 if (rbd_dev->spec->snap_id == CEPH_NOSNAP)
4252 if (rbd_dev->mapping.size != rbd_dev->header.image_size)
4253 rbd_dev->mapping.size = rbd_dev->header.image_size;
4255 ret = rbd_dev_v2_snap_context(rbd_dev);
4256 dout("rbd_dev_v2_snap_context returned %d\n", ret);
4258 up_write(&rbd_dev->header_rwsem);
4263 static int rbd_bus_add_dev(struct rbd_device *rbd_dev)
4268 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
4270 dev = &rbd_dev->dev;
4271 dev->bus = &rbd_bus_type;
4272 dev->type = &rbd_device_type;
4273 dev->parent = &rbd_root_dev;
4274 dev->release = rbd_dev_device_release;
4275 dev_set_name(dev, "%d", rbd_dev->dev_id);
4276 ret = device_register(dev);
4278 mutex_unlock(&ctl_mutex);
4283 static void rbd_bus_del_dev(struct rbd_device *rbd_dev)
4285 device_unregister(&rbd_dev->dev);
4288 static atomic64_t rbd_dev_id_max = ATOMIC64_INIT(0);
4291 * Get a unique rbd identifier for the given new rbd_dev, and add
4292 * the rbd_dev to the global list. The minimum rbd id is 1.
4294 static void rbd_dev_id_get(struct rbd_device *rbd_dev)
4296 rbd_dev->dev_id = atomic64_inc_return(&rbd_dev_id_max);
4298 spin_lock(&rbd_dev_list_lock);
4299 list_add_tail(&rbd_dev->node, &rbd_dev_list);
4300 spin_unlock(&rbd_dev_list_lock);
4301 dout("rbd_dev %p given dev id %llu\n", rbd_dev,
4302 (unsigned long long) rbd_dev->dev_id);
4306 * Remove an rbd_dev from the global list, and record that its
4307 * identifier is no longer in use.
4309 static void rbd_dev_id_put(struct rbd_device *rbd_dev)
4311 struct list_head *tmp;
4312 int rbd_id = rbd_dev->dev_id;
4315 rbd_assert(rbd_id > 0);
4317 dout("rbd_dev %p released dev id %llu\n", rbd_dev,
4318 (unsigned long long) rbd_dev->dev_id);
4319 spin_lock(&rbd_dev_list_lock);
4320 list_del_init(&rbd_dev->node);
4323 * If the id being "put" is not the current maximum, there
4324 * is nothing special we need to do.
4326 if (rbd_id != atomic64_read(&rbd_dev_id_max)) {
4327 spin_unlock(&rbd_dev_list_lock);
4332 * We need to update the current maximum id. Search the
4333 * list to find out what it is. We're more likely to find
4334 * the maximum at the end, so search the list backward.
4337 list_for_each_prev(tmp, &rbd_dev_list) {
4338 struct rbd_device *rbd_dev;
4340 rbd_dev = list_entry(tmp, struct rbd_device, node);
4341 if (rbd_dev->dev_id > max_id)
4342 max_id = rbd_dev->dev_id;
4344 spin_unlock(&rbd_dev_list_lock);
4347 * The max id could have been updated by rbd_dev_id_get(), in
4348 * which case it now accurately reflects the new maximum.
4349 * Be careful not to overwrite the maximum value in that
4352 atomic64_cmpxchg(&rbd_dev_id_max, rbd_id, max_id);
4353 dout(" max dev id has been reset\n");
4357 * Skips over white space at *buf, and updates *buf to point to the
4358 * first found non-space character (if any). Returns the length of
4359 * the token (string of non-white space characters) found. Note
4360 * that *buf must be terminated with '\0'.
4362 static inline size_t next_token(const char **buf)
4365 * These are the characters that produce nonzero for
4366 * isspace() in the "C" and "POSIX" locales.
4368 const char *spaces = " \f\n\r\t\v";
4370 *buf += strspn(*buf, spaces); /* Find start of token */
4372 return strcspn(*buf, spaces); /* Return token length */
4376 * Finds the next token in *buf, and if the provided token buffer is
4377 * big enough, copies the found token into it. The result, if
4378 * copied, is guaranteed to be terminated with '\0'. Note that *buf
4379 * must be terminated with '\0' on entry.
4381 * Returns the length of the token found (not including the '\0').
4382 * Return value will be 0 if no token is found, and it will be >=
4383 * token_size if the token would not fit.
4385 * The *buf pointer will be updated to point beyond the end of the
4386 * found token. Note that this occurs even if the token buffer is
4387 * too small to hold it.
4389 static inline size_t copy_token(const char **buf,
4395 len = next_token(buf);
4396 if (len < token_size) {
4397 memcpy(token, *buf, len);
4398 *(token + len) = '\0';
4406 * Finds the next token in *buf, dynamically allocates a buffer big
4407 * enough to hold a copy of it, and copies the token into the new
4408 * buffer. The copy is guaranteed to be terminated with '\0'. Note
4409 * that a duplicate buffer is created even for a zero-length token.
4411 * Returns a pointer to the newly-allocated duplicate, or a null
4412 * pointer if memory for the duplicate was not available. If
4413 * the lenp argument is a non-null pointer, the length of the token
4414 * (not including the '\0') is returned in *lenp.
4416 * If successful, the *buf pointer will be updated to point beyond
4417 * the end of the found token.
4419 * Note: uses GFP_KERNEL for allocation.
4421 static inline char *dup_token(const char **buf, size_t *lenp)
4426 len = next_token(buf);
4427 dup = kmemdup(*buf, len + 1, GFP_KERNEL);
4430 *(dup + len) = '\0';
4440 * Parse the options provided for an "rbd add" (i.e., rbd image
4441 * mapping) request. These arrive via a write to /sys/bus/rbd/add,
4442 * and the data written is passed here via a NUL-terminated buffer.
4443 * Returns 0 if successful or an error code otherwise.
4445 * The information extracted from these options is recorded in
4446 * the other parameters which return dynamically-allocated
4449 * The address of a pointer that will refer to a ceph options
4450 * structure. Caller must release the returned pointer using
4451 * ceph_destroy_options() when it is no longer needed.
4453 * Address of an rbd options pointer. Fully initialized by
4454 * this function; caller must release with kfree().
4456 * Address of an rbd image specification pointer. Fully
4457 * initialized by this function based on parsed options.
4458 * Caller must release with rbd_spec_put().
4460 * The options passed take this form:
4461 * <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
4464 * A comma-separated list of one or more monitor addresses.
4465 * A monitor address is an ip address, optionally followed
4466 * by a port number (separated by a colon).
4467 * I.e.: ip1[:port1][,ip2[:port2]...]
4469 * A comma-separated list of ceph and/or rbd options.
4471 * The name of the rados pool containing the rbd image.
4473 * The name of the image in that pool to map.
4475 * An optional snapshot id. If provided, the mapping will
4476 * present data from the image at the time that snapshot was
4477 * created. The image head is used if no snapshot id is
4478 * provided. Snapshot mappings are always read-only.
4480 static int rbd_add_parse_args(const char *buf,
4481 struct ceph_options **ceph_opts,
4482 struct rbd_options **opts,
4483 struct rbd_spec **rbd_spec)
4487 const char *mon_addrs;
4489 size_t mon_addrs_size;
4490 struct rbd_spec *spec = NULL;
4491 struct rbd_options *rbd_opts = NULL;
4492 struct ceph_options *copts;
4495 /* The first four tokens are required */
4497 len = next_token(&buf);
4499 rbd_warn(NULL, "no monitor address(es) provided");
4503 mon_addrs_size = len + 1;
4507 options = dup_token(&buf, NULL);
4511 rbd_warn(NULL, "no options provided");
4515 spec = rbd_spec_alloc();
4519 spec->pool_name = dup_token(&buf, NULL);
4520 if (!spec->pool_name)
4522 if (!*spec->pool_name) {
4523 rbd_warn(NULL, "no pool name provided");
4527 spec->image_name = dup_token(&buf, NULL);
4528 if (!spec->image_name)
4530 if (!*spec->image_name) {
4531 rbd_warn(NULL, "no image name provided");
4536 * Snapshot name is optional; default is to use "-"
4537 * (indicating the head/no snapshot).
4539 len = next_token(&buf);
4541 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
4542 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
4543 } else if (len > RBD_MAX_SNAP_NAME_LEN) {
4544 ret = -ENAMETOOLONG;
4547 snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
4550 *(snap_name + len) = '\0';
4551 spec->snap_name = snap_name;
4553 /* Initialize all rbd options to the defaults */
4555 rbd_opts = kzalloc(sizeof (*rbd_opts), GFP_KERNEL);
4559 rbd_opts->read_only = RBD_READ_ONLY_DEFAULT;
4561 copts = ceph_parse_options(options, mon_addrs,
4562 mon_addrs + mon_addrs_size - 1,
4563 parse_rbd_opts_token, rbd_opts);
4564 if (IS_ERR(copts)) {
4565 ret = PTR_ERR(copts);
4586 * An rbd format 2 image has a unique identifier, distinct from the
4587 * name given to it by the user. Internally, that identifier is
4588 * what's used to specify the names of objects related to the image.
4590 * A special "rbd id" object is used to map an rbd image name to its
4591 * id. If that object doesn't exist, then there is no v2 rbd image
4592 * with the supplied name.
4594 * This function will record the given rbd_dev's image_id field if
4595 * it can be determined, and in that case will return 0. If any
4596 * errors occur a negative errno will be returned and the rbd_dev's
4597 * image_id field will be unchanged (and should be NULL).
4599 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
4608 * When probing a parent image, the image id is already
4609 * known (and the image name likely is not). There's no
4610 * need to fetch the image id again in this case. We
4611 * do still need to set the image format though.
4613 if (rbd_dev->spec->image_id) {
4614 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
4620 * First, see if the format 2 image id file exists, and if
4621 * so, get the image's persistent id from it.
4623 size = sizeof (RBD_ID_PREFIX) + strlen(rbd_dev->spec->image_name);
4624 object_name = kmalloc(size, GFP_NOIO);
4627 sprintf(object_name, "%s%s", RBD_ID_PREFIX, rbd_dev->spec->image_name);
4628 dout("rbd id object name is %s\n", object_name);
4630 /* Response will be an encoded string, which includes a length */
4632 size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
4633 response = kzalloc(size, GFP_NOIO);
4639 /* If it doesn't exist we'll assume it's a format 1 image */
4641 ret = rbd_obj_method_sync(rbd_dev, object_name,
4642 "rbd", "get_id", NULL, 0,
4643 response, RBD_IMAGE_ID_LEN_MAX);
4644 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4645 if (ret == -ENOENT) {
4646 image_id = kstrdup("", GFP_KERNEL);
4647 ret = image_id ? 0 : -ENOMEM;
4649 rbd_dev->image_format = 1;
4650 } else if (ret > sizeof (__le32)) {
4653 image_id = ceph_extract_encoded_string(&p, p + ret,
4655 ret = IS_ERR(image_id) ? PTR_ERR(image_id) : 0;
4657 rbd_dev->image_format = 2;
4663 rbd_dev->spec->image_id = image_id;
4664 dout("image_id is %s\n", image_id);
4673 /* Undo whatever state changes are made by v1 or v2 image probe */
4675 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
4677 struct rbd_image_header *header;
4679 /* Drop parent reference unless it's already been done (or none) */
4681 if (rbd_dev->parent_overlap)
4682 rbd_dev_parent_put(rbd_dev);
4684 /* Free dynamic fields from the header, then zero it out */
4686 header = &rbd_dev->header;
4687 ceph_put_snap_context(header->snapc);
4688 kfree(header->snap_sizes);
4689 kfree(header->snap_names);
4690 kfree(header->object_prefix);
4691 memset(header, 0, sizeof (*header));
4694 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev)
4698 ret = rbd_dev_v2_object_prefix(rbd_dev);
4703 * Get the and check features for the image. Currently the
4704 * features are assumed to never change.
4706 ret = rbd_dev_v2_features(rbd_dev);
4710 /* If the image supports fancy striping, get its parameters */
4712 if (rbd_dev->header.features & RBD_FEATURE_STRIPINGV2) {
4713 ret = rbd_dev_v2_striping_info(rbd_dev);
4717 /* No support for crypto and compression type format 2 images */
4721 rbd_dev->header.features = 0;
4722 kfree(rbd_dev->header.object_prefix);
4723 rbd_dev->header.object_prefix = NULL;
4728 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev)
4730 struct rbd_device *parent = NULL;
4731 struct rbd_spec *parent_spec;
4732 struct rbd_client *rbdc;
4735 if (!rbd_dev->parent_spec)
4738 * We need to pass a reference to the client and the parent
4739 * spec when creating the parent rbd_dev. Images related by
4740 * parent/child relationships always share both.
4742 parent_spec = rbd_spec_get(rbd_dev->parent_spec);
4743 rbdc = __rbd_get_client(rbd_dev->rbd_client);
4746 parent = rbd_dev_create(rbdc, parent_spec);
4750 ret = rbd_dev_image_probe(parent, false);
4753 rbd_dev->parent = parent;
4754 atomic_set(&rbd_dev->parent_ref, 1);
4759 rbd_dev_unparent(rbd_dev);
4760 kfree(rbd_dev->header_name);
4761 rbd_dev_destroy(parent);
4763 rbd_put_client(rbdc);
4764 rbd_spec_put(parent_spec);
4770 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
4774 /* generate unique id: find highest unique id, add one */
4775 rbd_dev_id_get(rbd_dev);
4777 /* Fill in the device name, now that we have its id. */
4778 BUILD_BUG_ON(DEV_NAME_LEN
4779 < sizeof (RBD_DRV_NAME) + MAX_INT_FORMAT_WIDTH);
4780 sprintf(rbd_dev->name, "%s%d", RBD_DRV_NAME, rbd_dev->dev_id);
4782 /* Get our block major device number. */
4784 ret = register_blkdev(0, rbd_dev->name);
4787 rbd_dev->major = ret;
4789 /* Set up the blkdev mapping. */
4791 ret = rbd_init_disk(rbd_dev);
4793 goto err_out_blkdev;
4795 ret = rbd_dev_mapping_set(rbd_dev);
4798 set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
4800 ret = rbd_bus_add_dev(rbd_dev);
4802 goto err_out_mapping;
4804 /* Everything's ready. Announce the disk to the world. */
4806 set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
4807 add_disk(rbd_dev->disk);
4809 pr_info("%s: added with size 0x%llx\n", rbd_dev->disk->disk_name,
4810 (unsigned long long) rbd_dev->mapping.size);
4815 rbd_dev_mapping_clear(rbd_dev);
4817 rbd_free_disk(rbd_dev);
4819 unregister_blkdev(rbd_dev->major, rbd_dev->name);
4821 rbd_dev_id_put(rbd_dev);
4822 rbd_dev_mapping_clear(rbd_dev);
4827 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
4829 struct rbd_spec *spec = rbd_dev->spec;
4832 /* Record the header object name for this rbd image. */
4834 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
4836 if (rbd_dev->image_format == 1)
4837 size = strlen(spec->image_name) + sizeof (RBD_SUFFIX);
4839 size = sizeof (RBD_HEADER_PREFIX) + strlen(spec->image_id);
4841 rbd_dev->header_name = kmalloc(size, GFP_KERNEL);
4842 if (!rbd_dev->header_name)
4845 if (rbd_dev->image_format == 1)
4846 sprintf(rbd_dev->header_name, "%s%s",
4847 spec->image_name, RBD_SUFFIX);
4849 sprintf(rbd_dev->header_name, "%s%s",
4850 RBD_HEADER_PREFIX, spec->image_id);
4854 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
4856 rbd_dev_unprobe(rbd_dev);
4857 kfree(rbd_dev->header_name);
4858 rbd_dev->header_name = NULL;
4859 rbd_dev->image_format = 0;
4860 kfree(rbd_dev->spec->image_id);
4861 rbd_dev->spec->image_id = NULL;
4863 rbd_dev_destroy(rbd_dev);
4867 * Probe for the existence of the header object for the given rbd
4868 * device. If this image is the one being mapped (i.e., not a
4869 * parent), initiate a watch on its header object before using that
4870 * object to get detailed information about the rbd image.
4872 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, bool mapping)
4878 * Get the id from the image id object. If it's not a
4879 * format 2 image, we'll get ENOENT back, and we'll assume
4880 * it's a format 1 image.
4882 ret = rbd_dev_image_id(rbd_dev);
4885 rbd_assert(rbd_dev->spec->image_id);
4886 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
4888 ret = rbd_dev_header_name(rbd_dev);
4890 goto err_out_format;
4893 ret = rbd_dev_header_watch_sync(rbd_dev, true);
4895 goto out_header_name;
4898 if (rbd_dev->image_format == 1)
4899 ret = rbd_dev_v1_header_info(rbd_dev);
4901 ret = rbd_dev_v2_header_info(rbd_dev);
4905 ret = rbd_dev_spec_update(rbd_dev);
4909 ret = rbd_dev_probe_parent(rbd_dev);
4913 dout("discovered format %u image, header name is %s\n",
4914 rbd_dev->image_format, rbd_dev->header_name);
4918 rbd_dev_unprobe(rbd_dev);
4921 tmp = rbd_dev_header_watch_sync(rbd_dev, false);
4923 rbd_warn(rbd_dev, "unable to tear down "
4924 "watch request (%d)\n", tmp);
4927 kfree(rbd_dev->header_name);
4928 rbd_dev->header_name = NULL;
4930 rbd_dev->image_format = 0;
4931 kfree(rbd_dev->spec->image_id);
4932 rbd_dev->spec->image_id = NULL;
4934 dout("probe failed, returning %d\n", ret);
4939 static ssize_t rbd_add(struct bus_type *bus,
4943 struct rbd_device *rbd_dev = NULL;
4944 struct ceph_options *ceph_opts = NULL;
4945 struct rbd_options *rbd_opts = NULL;
4946 struct rbd_spec *spec = NULL;
4947 struct rbd_client *rbdc;
4948 struct ceph_osd_client *osdc;
4952 if (!try_module_get(THIS_MODULE))
4955 /* parse add command */
4956 rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
4958 goto err_out_module;
4959 read_only = rbd_opts->read_only;
4961 rbd_opts = NULL; /* done with this */
4963 rbdc = rbd_get_client(ceph_opts);
4968 ceph_opts = NULL; /* rbd_dev client now owns this */
4971 osdc = &rbdc->client->osdc;
4972 rc = ceph_pg_poolid_by_name(osdc->osdmap, spec->pool_name);
4974 goto err_out_client;
4975 spec->pool_id = (u64)rc;
4977 /* The ceph file layout needs to fit pool id in 32 bits */
4979 if (spec->pool_id > (u64)U32_MAX) {
4980 rbd_warn(NULL, "pool id too large (%llu > %u)\n",
4981 (unsigned long long)spec->pool_id, U32_MAX);
4983 goto err_out_client;
4986 rbd_dev = rbd_dev_create(rbdc, spec);
4988 goto err_out_client;
4989 rbdc = NULL; /* rbd_dev now owns this */
4990 spec = NULL; /* rbd_dev now owns this */
4992 rc = rbd_dev_image_probe(rbd_dev, true);
4994 goto err_out_rbd_dev;
4996 /* If we are mapping a snapshot it must be marked read-only */
4998 if (rbd_dev->spec->snap_id != CEPH_NOSNAP)
5000 rbd_dev->mapping.read_only = read_only;
5002 rc = rbd_dev_device_setup(rbd_dev);
5006 rbd_dev_image_release(rbd_dev);
5008 rbd_dev_destroy(rbd_dev);
5010 rbd_put_client(rbdc);
5013 ceph_destroy_options(ceph_opts);
5017 module_put(THIS_MODULE);
5019 dout("Error adding device %s\n", buf);
5024 static struct rbd_device *__rbd_get_dev(unsigned long dev_id)
5026 struct list_head *tmp;
5027 struct rbd_device *rbd_dev;
5029 spin_lock(&rbd_dev_list_lock);
5030 list_for_each(tmp, &rbd_dev_list) {
5031 rbd_dev = list_entry(tmp, struct rbd_device, node);
5032 if (rbd_dev->dev_id == dev_id) {
5033 spin_unlock(&rbd_dev_list_lock);
5037 spin_unlock(&rbd_dev_list_lock);
5041 static void rbd_dev_device_release(struct device *dev)
5043 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5045 rbd_free_disk(rbd_dev);
5046 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5047 rbd_dev_mapping_clear(rbd_dev);
5048 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5050 rbd_dev_id_put(rbd_dev);
5051 rbd_dev_mapping_clear(rbd_dev);
5054 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
5056 while (rbd_dev->parent) {
5057 struct rbd_device *first = rbd_dev;
5058 struct rbd_device *second = first->parent;
5059 struct rbd_device *third;
5062 * Follow to the parent with no grandparent and
5065 while (second && (third = second->parent)) {
5070 rbd_dev_image_release(second);
5071 first->parent = NULL;
5072 first->parent_overlap = 0;
5074 rbd_assert(first->parent_spec);
5075 rbd_spec_put(first->parent_spec);
5076 first->parent_spec = NULL;
5080 static ssize_t rbd_remove(struct bus_type *bus,
5084 struct rbd_device *rbd_dev = NULL;
5089 ret = strict_strtoul(buf, 10, &ul);
5093 /* convert to int; abort if we lost anything in the conversion */
5094 target_id = (int) ul;
5095 if (target_id != ul)
5098 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
5100 rbd_dev = __rbd_get_dev(target_id);
5106 spin_lock_irq(&rbd_dev->lock);
5107 if (rbd_dev->open_count)
5110 set_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags);
5111 spin_unlock_irq(&rbd_dev->lock);
5114 rbd_bus_del_dev(rbd_dev);
5115 ret = rbd_dev_header_watch_sync(rbd_dev, false);
5117 rbd_warn(rbd_dev, "failed to cancel watch event (%d)\n", ret);
5118 rbd_dev_image_release(rbd_dev);
5119 module_put(THIS_MODULE);
5122 mutex_unlock(&ctl_mutex);
5128 * create control files in sysfs
5131 static int rbd_sysfs_init(void)
5135 ret = device_register(&rbd_root_dev);
5139 ret = bus_register(&rbd_bus_type);
5141 device_unregister(&rbd_root_dev);
5146 static void rbd_sysfs_cleanup(void)
5148 bus_unregister(&rbd_bus_type);
5149 device_unregister(&rbd_root_dev);
5152 static int rbd_slab_init(void)
5154 rbd_assert(!rbd_img_request_cache);
5155 rbd_img_request_cache = kmem_cache_create("rbd_img_request",
5156 sizeof (struct rbd_img_request),
5157 __alignof__(struct rbd_img_request),
5159 if (!rbd_img_request_cache)
5162 rbd_assert(!rbd_obj_request_cache);
5163 rbd_obj_request_cache = kmem_cache_create("rbd_obj_request",
5164 sizeof (struct rbd_obj_request),
5165 __alignof__(struct rbd_obj_request),
5167 if (!rbd_obj_request_cache)
5170 rbd_assert(!rbd_segment_name_cache);
5171 rbd_segment_name_cache = kmem_cache_create("rbd_segment_name",
5172 MAX_OBJ_NAME_SIZE + 1, 1, 0, NULL);
5173 if (rbd_segment_name_cache)
5176 if (rbd_obj_request_cache) {
5177 kmem_cache_destroy(rbd_obj_request_cache);
5178 rbd_obj_request_cache = NULL;
5181 kmem_cache_destroy(rbd_img_request_cache);
5182 rbd_img_request_cache = NULL;
5187 static void rbd_slab_exit(void)
5189 rbd_assert(rbd_segment_name_cache);
5190 kmem_cache_destroy(rbd_segment_name_cache);
5191 rbd_segment_name_cache = NULL;
5193 rbd_assert(rbd_obj_request_cache);
5194 kmem_cache_destroy(rbd_obj_request_cache);
5195 rbd_obj_request_cache = NULL;
5197 rbd_assert(rbd_img_request_cache);
5198 kmem_cache_destroy(rbd_img_request_cache);
5199 rbd_img_request_cache = NULL;
5202 static int __init rbd_init(void)
5206 if (!libceph_compatible(NULL)) {
5207 rbd_warn(NULL, "libceph incompatibility (quitting)");
5211 rc = rbd_slab_init();
5214 rc = rbd_sysfs_init();
5218 pr_info("loaded " RBD_DRV_NAME_LONG "\n");
5223 static void __exit rbd_exit(void)
5225 rbd_sysfs_cleanup();
5229 module_init(rbd_init);
5230 module_exit(rbd_exit);
5232 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
5233 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
5234 MODULE_DESCRIPTION("rados block device");
5236 /* following authorship retained from original osdblk.c */
5237 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
5239 MODULE_LICENSE("GPL");