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 * Returns true if the rbd device has a parent with a non-zero
1933 * overlap and a reference for it was successfully taken, or
1936 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
1940 if (!rbd_dev->parent_spec)
1943 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
1944 if (counter > 0 && rbd_dev->parent_overlap)
1947 /* Image was flattened, but parent is not yet torn down */
1950 rbd_warn(rbd_dev, "parent reference overflow\n");
1956 * Caller is responsible for filling in the list of object requests
1957 * that comprises the image request, and the Linux request pointer
1958 * (if there is one).
1960 static struct rbd_img_request *rbd_img_request_create(
1961 struct rbd_device *rbd_dev,
1962 u64 offset, u64 length,
1965 struct rbd_img_request *img_request;
1967 img_request = kmem_cache_alloc(rbd_img_request_cache, GFP_ATOMIC);
1971 if (write_request) {
1972 down_read(&rbd_dev->header_rwsem);
1973 ceph_get_snap_context(rbd_dev->header.snapc);
1974 up_read(&rbd_dev->header_rwsem);
1977 img_request->rq = NULL;
1978 img_request->rbd_dev = rbd_dev;
1979 img_request->offset = offset;
1980 img_request->length = length;
1981 img_request->flags = 0;
1982 if (write_request) {
1983 img_request_write_set(img_request);
1984 img_request->snapc = rbd_dev->header.snapc;
1986 img_request->snap_id = rbd_dev->spec->snap_id;
1988 if (rbd_dev_parent_get(rbd_dev))
1989 img_request_layered_set(img_request);
1990 spin_lock_init(&img_request->completion_lock);
1991 img_request->next_completion = 0;
1992 img_request->callback = NULL;
1993 img_request->result = 0;
1994 img_request->obj_request_count = 0;
1995 INIT_LIST_HEAD(&img_request->obj_requests);
1996 kref_init(&img_request->kref);
1998 dout("%s: rbd_dev %p %s %llu/%llu -> img %p\n", __func__, rbd_dev,
1999 write_request ? "write" : "read", offset, length,
2005 static void rbd_img_request_destroy(struct kref *kref)
2007 struct rbd_img_request *img_request;
2008 struct rbd_obj_request *obj_request;
2009 struct rbd_obj_request *next_obj_request;
2011 img_request = container_of(kref, struct rbd_img_request, kref);
2013 dout("%s: img %p\n", __func__, img_request);
2015 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2016 rbd_img_obj_request_del(img_request, obj_request);
2017 rbd_assert(img_request->obj_request_count == 0);
2019 if (img_request_layered_test(img_request)) {
2020 img_request_layered_clear(img_request);
2021 rbd_dev_parent_put(img_request->rbd_dev);
2024 if (img_request_write_test(img_request))
2025 ceph_put_snap_context(img_request->snapc);
2027 kmem_cache_free(rbd_img_request_cache, img_request);
2030 static struct rbd_img_request *rbd_parent_request_create(
2031 struct rbd_obj_request *obj_request,
2032 u64 img_offset, u64 length)
2034 struct rbd_img_request *parent_request;
2035 struct rbd_device *rbd_dev;
2037 rbd_assert(obj_request->img_request);
2038 rbd_dev = obj_request->img_request->rbd_dev;
2040 parent_request = rbd_img_request_create(rbd_dev->parent,
2041 img_offset, length, false);
2042 if (!parent_request)
2045 img_request_child_set(parent_request);
2046 rbd_obj_request_get(obj_request);
2047 parent_request->obj_request = obj_request;
2049 return parent_request;
2052 static void rbd_parent_request_destroy(struct kref *kref)
2054 struct rbd_img_request *parent_request;
2055 struct rbd_obj_request *orig_request;
2057 parent_request = container_of(kref, struct rbd_img_request, kref);
2058 orig_request = parent_request->obj_request;
2060 parent_request->obj_request = NULL;
2061 rbd_obj_request_put(orig_request);
2062 img_request_child_clear(parent_request);
2064 rbd_img_request_destroy(kref);
2067 static bool rbd_img_obj_end_request(struct rbd_obj_request *obj_request)
2069 struct rbd_img_request *img_request;
2070 unsigned int xferred;
2074 rbd_assert(obj_request_img_data_test(obj_request));
2075 img_request = obj_request->img_request;
2077 rbd_assert(obj_request->xferred <= (u64)UINT_MAX);
2078 xferred = (unsigned int)obj_request->xferred;
2079 result = obj_request->result;
2081 struct rbd_device *rbd_dev = img_request->rbd_dev;
2083 rbd_warn(rbd_dev, "%s %llx at %llx (%llx)\n",
2084 img_request_write_test(img_request) ? "write" : "read",
2085 obj_request->length, obj_request->img_offset,
2086 obj_request->offset);
2087 rbd_warn(rbd_dev, " result %d xferred %x\n",
2089 if (!img_request->result)
2090 img_request->result = result;
2093 /* Image object requests don't own their page array */
2095 if (obj_request->type == OBJ_REQUEST_PAGES) {
2096 obj_request->pages = NULL;
2097 obj_request->page_count = 0;
2100 if (img_request_child_test(img_request)) {
2101 rbd_assert(img_request->obj_request != NULL);
2102 more = obj_request->which < img_request->obj_request_count - 1;
2104 rbd_assert(img_request->rq != NULL);
2105 more = blk_end_request(img_request->rq, result, xferred);
2111 static void rbd_img_obj_callback(struct rbd_obj_request *obj_request)
2113 struct rbd_img_request *img_request;
2114 u32 which = obj_request->which;
2117 rbd_assert(obj_request_img_data_test(obj_request));
2118 img_request = obj_request->img_request;
2120 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
2121 rbd_assert(img_request != NULL);
2122 rbd_assert(img_request->obj_request_count > 0);
2123 rbd_assert(which != BAD_WHICH);
2124 rbd_assert(which < img_request->obj_request_count);
2125 rbd_assert(which >= img_request->next_completion);
2127 spin_lock_irq(&img_request->completion_lock);
2128 if (which != img_request->next_completion)
2131 for_each_obj_request_from(img_request, obj_request) {
2133 rbd_assert(which < img_request->obj_request_count);
2135 if (!obj_request_done_test(obj_request))
2137 more = rbd_img_obj_end_request(obj_request);
2141 rbd_assert(more ^ (which == img_request->obj_request_count));
2142 img_request->next_completion = which;
2144 spin_unlock_irq(&img_request->completion_lock);
2147 rbd_img_request_complete(img_request);
2151 * Split up an image request into one or more object requests, each
2152 * to a different object. The "type" parameter indicates whether
2153 * "data_desc" is the pointer to the head of a list of bio
2154 * structures, or the base of a page array. In either case this
2155 * function assumes data_desc describes memory sufficient to hold
2156 * all data described by the image request.
2158 static int rbd_img_request_fill(struct rbd_img_request *img_request,
2159 enum obj_request_type type,
2162 struct rbd_device *rbd_dev = img_request->rbd_dev;
2163 struct rbd_obj_request *obj_request = NULL;
2164 struct rbd_obj_request *next_obj_request;
2165 bool write_request = img_request_write_test(img_request);
2166 struct bio *bio_list;
2167 unsigned int bio_offset = 0;
2168 struct page **pages;
2173 dout("%s: img %p type %d data_desc %p\n", __func__, img_request,
2174 (int)type, data_desc);
2176 opcode = write_request ? CEPH_OSD_OP_WRITE : CEPH_OSD_OP_READ;
2177 img_offset = img_request->offset;
2178 resid = img_request->length;
2179 rbd_assert(resid > 0);
2181 if (type == OBJ_REQUEST_BIO) {
2182 bio_list = data_desc;
2183 rbd_assert(img_offset == bio_list->bi_sector << SECTOR_SHIFT);
2185 rbd_assert(type == OBJ_REQUEST_PAGES);
2190 struct ceph_osd_request *osd_req;
2191 const char *object_name;
2195 object_name = rbd_segment_name(rbd_dev, img_offset);
2198 offset = rbd_segment_offset(rbd_dev, img_offset);
2199 length = rbd_segment_length(rbd_dev, img_offset, resid);
2200 obj_request = rbd_obj_request_create(object_name,
2201 offset, length, type);
2202 /* object request has its own copy of the object name */
2203 rbd_segment_name_free(object_name);
2207 if (type == OBJ_REQUEST_BIO) {
2208 unsigned int clone_size;
2210 rbd_assert(length <= (u64)UINT_MAX);
2211 clone_size = (unsigned int)length;
2212 obj_request->bio_list =
2213 bio_chain_clone_range(&bio_list,
2217 if (!obj_request->bio_list)
2220 unsigned int page_count;
2222 obj_request->pages = pages;
2223 page_count = (u32)calc_pages_for(offset, length);
2224 obj_request->page_count = page_count;
2225 if ((offset + length) & ~PAGE_MASK)
2226 page_count--; /* more on last page */
2227 pages += page_count;
2230 osd_req = rbd_osd_req_create(rbd_dev, write_request,
2234 obj_request->osd_req = osd_req;
2235 obj_request->callback = rbd_img_obj_callback;
2237 osd_req_op_extent_init(osd_req, 0, opcode, offset, length,
2239 if (type == OBJ_REQUEST_BIO)
2240 osd_req_op_extent_osd_data_bio(osd_req, 0,
2241 obj_request->bio_list, length);
2243 osd_req_op_extent_osd_data_pages(osd_req, 0,
2244 obj_request->pages, length,
2245 offset & ~PAGE_MASK, false, false);
2248 rbd_osd_req_format_write(obj_request);
2250 rbd_osd_req_format_read(obj_request);
2252 obj_request->img_offset = img_offset;
2253 rbd_img_obj_request_add(img_request, obj_request);
2255 img_offset += length;
2262 rbd_obj_request_put(obj_request);
2264 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2265 rbd_obj_request_put(obj_request);
2271 rbd_img_obj_copyup_callback(struct rbd_obj_request *obj_request)
2273 struct rbd_img_request *img_request;
2274 struct rbd_device *rbd_dev;
2275 struct page **pages;
2278 rbd_assert(obj_request->type == OBJ_REQUEST_BIO);
2279 rbd_assert(obj_request_img_data_test(obj_request));
2280 img_request = obj_request->img_request;
2281 rbd_assert(img_request);
2283 rbd_dev = img_request->rbd_dev;
2284 rbd_assert(rbd_dev);
2286 pages = obj_request->copyup_pages;
2287 rbd_assert(pages != NULL);
2288 obj_request->copyup_pages = NULL;
2289 page_count = obj_request->copyup_page_count;
2290 rbd_assert(page_count);
2291 obj_request->copyup_page_count = 0;
2292 ceph_release_page_vector(pages, page_count);
2295 * We want the transfer count to reflect the size of the
2296 * original write request. There is no such thing as a
2297 * successful short write, so if the request was successful
2298 * we can just set it to the originally-requested length.
2300 if (!obj_request->result)
2301 obj_request->xferred = obj_request->length;
2303 /* Finish up with the normal image object callback */
2305 rbd_img_obj_callback(obj_request);
2309 rbd_img_obj_parent_read_full_callback(struct rbd_img_request *img_request)
2311 struct rbd_obj_request *orig_request;
2312 struct ceph_osd_request *osd_req;
2313 struct ceph_osd_client *osdc;
2314 struct rbd_device *rbd_dev;
2315 struct page **pages;
2322 rbd_assert(img_request_child_test(img_request));
2324 /* First get what we need from the image request */
2326 pages = img_request->copyup_pages;
2327 rbd_assert(pages != NULL);
2328 img_request->copyup_pages = NULL;
2329 page_count = img_request->copyup_page_count;
2330 rbd_assert(page_count);
2331 img_request->copyup_page_count = 0;
2333 orig_request = img_request->obj_request;
2334 rbd_assert(orig_request != NULL);
2335 rbd_assert(obj_request_type_valid(orig_request->type));
2336 result = img_request->result;
2337 parent_length = img_request->length;
2338 rbd_assert(parent_length == img_request->xferred);
2339 rbd_img_request_put(img_request);
2341 rbd_assert(orig_request->img_request);
2342 rbd_dev = orig_request->img_request->rbd_dev;
2343 rbd_assert(rbd_dev);
2349 * The original osd request is of no use to use any more.
2350 * We need a new one that can hold the two ops in a copyup
2351 * request. Allocate the new copyup osd request for the
2352 * original request, and release the old one.
2355 osd_req = rbd_osd_req_create_copyup(orig_request);
2358 rbd_osd_req_destroy(orig_request->osd_req);
2359 orig_request->osd_req = osd_req;
2360 orig_request->copyup_pages = pages;
2361 orig_request->copyup_page_count = page_count;
2363 /* Initialize the copyup op */
2365 osd_req_op_cls_init(osd_req, 0, CEPH_OSD_OP_CALL, "rbd", "copyup");
2366 osd_req_op_cls_request_data_pages(osd_req, 0, pages, parent_length, 0,
2369 /* Then the original write request op */
2371 offset = orig_request->offset;
2372 length = orig_request->length;
2373 osd_req_op_extent_init(osd_req, 1, CEPH_OSD_OP_WRITE,
2374 offset, length, 0, 0);
2375 if (orig_request->type == OBJ_REQUEST_BIO)
2376 osd_req_op_extent_osd_data_bio(osd_req, 1,
2377 orig_request->bio_list, length);
2379 osd_req_op_extent_osd_data_pages(osd_req, 1,
2380 orig_request->pages, length,
2381 offset & ~PAGE_MASK, false, false);
2383 rbd_osd_req_format_write(orig_request);
2385 /* All set, send it off. */
2387 orig_request->callback = rbd_img_obj_copyup_callback;
2388 osdc = &rbd_dev->rbd_client->client->osdc;
2389 result = rbd_obj_request_submit(osdc, orig_request);
2393 /* Record the error code and complete the request */
2395 orig_request->result = result;
2396 orig_request->xferred = 0;
2397 obj_request_done_set(orig_request);
2398 rbd_obj_request_complete(orig_request);
2402 * Read from the parent image the range of data that covers the
2403 * entire target of the given object request. This is used for
2404 * satisfying a layered image write request when the target of an
2405 * object request from the image request does not exist.
2407 * A page array big enough to hold the returned data is allocated
2408 * and supplied to rbd_img_request_fill() as the "data descriptor."
2409 * When the read completes, this page array will be transferred to
2410 * the original object request for the copyup operation.
2412 * If an error occurs, record it as the result of the original
2413 * object request and mark it done so it gets completed.
2415 static int rbd_img_obj_parent_read_full(struct rbd_obj_request *obj_request)
2417 struct rbd_img_request *img_request = NULL;
2418 struct rbd_img_request *parent_request = NULL;
2419 struct rbd_device *rbd_dev;
2422 struct page **pages = NULL;
2426 rbd_assert(obj_request_img_data_test(obj_request));
2427 rbd_assert(obj_request_type_valid(obj_request->type));
2429 img_request = obj_request->img_request;
2430 rbd_assert(img_request != NULL);
2431 rbd_dev = img_request->rbd_dev;
2432 rbd_assert(rbd_dev->parent != NULL);
2435 * Determine the byte range covered by the object in the
2436 * child image to which the original request was to be sent.
2438 img_offset = obj_request->img_offset - obj_request->offset;
2439 length = (u64)1 << rbd_dev->header.obj_order;
2442 * There is no defined parent data beyond the parent
2443 * overlap, so limit what we read at that boundary if
2446 if (img_offset + length > rbd_dev->parent_overlap) {
2447 rbd_assert(img_offset < rbd_dev->parent_overlap);
2448 length = rbd_dev->parent_overlap - img_offset;
2452 * Allocate a page array big enough to receive the data read
2455 page_count = (u32)calc_pages_for(0, length);
2456 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2457 if (IS_ERR(pages)) {
2458 result = PTR_ERR(pages);
2464 parent_request = rbd_parent_request_create(obj_request,
2465 img_offset, length);
2466 if (!parent_request)
2469 result = rbd_img_request_fill(parent_request, OBJ_REQUEST_PAGES, pages);
2472 parent_request->copyup_pages = pages;
2473 parent_request->copyup_page_count = page_count;
2475 parent_request->callback = rbd_img_obj_parent_read_full_callback;
2476 result = rbd_img_request_submit(parent_request);
2480 parent_request->copyup_pages = NULL;
2481 parent_request->copyup_page_count = 0;
2482 parent_request->obj_request = NULL;
2483 rbd_obj_request_put(obj_request);
2486 ceph_release_page_vector(pages, page_count);
2488 rbd_img_request_put(parent_request);
2489 obj_request->result = result;
2490 obj_request->xferred = 0;
2491 obj_request_done_set(obj_request);
2496 static void rbd_img_obj_exists_callback(struct rbd_obj_request *obj_request)
2498 struct rbd_obj_request *orig_request;
2501 rbd_assert(!obj_request_img_data_test(obj_request));
2504 * All we need from the object request is the original
2505 * request and the result of the STAT op. Grab those, then
2506 * we're done with the request.
2508 orig_request = obj_request->obj_request;
2509 obj_request->obj_request = NULL;
2510 rbd_assert(orig_request);
2511 rbd_assert(orig_request->img_request);
2513 result = obj_request->result;
2514 obj_request->result = 0;
2516 dout("%s: obj %p for obj %p result %d %llu/%llu\n", __func__,
2517 obj_request, orig_request, result,
2518 obj_request->xferred, obj_request->length);
2519 rbd_obj_request_put(obj_request);
2521 rbd_assert(orig_request);
2522 rbd_assert(orig_request->img_request);
2525 * Our only purpose here is to determine whether the object
2526 * exists, and we don't want to treat the non-existence as
2527 * an error. If something else comes back, transfer the
2528 * error to the original request and complete it now.
2531 obj_request_existence_set(orig_request, true);
2532 } else if (result == -ENOENT) {
2533 obj_request_existence_set(orig_request, false);
2534 } else if (result) {
2535 orig_request->result = result;
2540 * Resubmit the original request now that we have recorded
2541 * whether the target object exists.
2543 orig_request->result = rbd_img_obj_request_submit(orig_request);
2545 if (orig_request->result)
2546 rbd_obj_request_complete(orig_request);
2547 rbd_obj_request_put(orig_request);
2550 static int rbd_img_obj_exists_submit(struct rbd_obj_request *obj_request)
2552 struct rbd_obj_request *stat_request;
2553 struct rbd_device *rbd_dev;
2554 struct ceph_osd_client *osdc;
2555 struct page **pages = NULL;
2561 * The response data for a STAT call consists of:
2568 size = sizeof (__le64) + sizeof (__le32) + sizeof (__le32);
2569 page_count = (u32)calc_pages_for(0, size);
2570 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2572 return PTR_ERR(pages);
2575 stat_request = rbd_obj_request_create(obj_request->object_name, 0, 0,
2580 rbd_obj_request_get(obj_request);
2581 stat_request->obj_request = obj_request;
2582 stat_request->pages = pages;
2583 stat_request->page_count = page_count;
2585 rbd_assert(obj_request->img_request);
2586 rbd_dev = obj_request->img_request->rbd_dev;
2587 stat_request->osd_req = rbd_osd_req_create(rbd_dev, false,
2589 if (!stat_request->osd_req)
2591 stat_request->callback = rbd_img_obj_exists_callback;
2593 osd_req_op_init(stat_request->osd_req, 0, CEPH_OSD_OP_STAT);
2594 osd_req_op_raw_data_in_pages(stat_request->osd_req, 0, pages, size, 0,
2596 rbd_osd_req_format_read(stat_request);
2598 osdc = &rbd_dev->rbd_client->client->osdc;
2599 ret = rbd_obj_request_submit(osdc, stat_request);
2602 rbd_obj_request_put(obj_request);
2607 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request)
2609 struct rbd_img_request *img_request;
2610 struct rbd_device *rbd_dev;
2613 rbd_assert(obj_request_img_data_test(obj_request));
2615 img_request = obj_request->img_request;
2616 rbd_assert(img_request);
2617 rbd_dev = img_request->rbd_dev;
2620 * Only writes to layered images need special handling.
2621 * Reads and non-layered writes are simple object requests.
2622 * Layered writes that start beyond the end of the overlap
2623 * with the parent have no parent data, so they too are
2624 * simple object requests. Finally, if the target object is
2625 * known to already exist, its parent data has already been
2626 * copied, so a write to the object can also be handled as a
2627 * simple object request.
2629 if (!img_request_write_test(img_request) ||
2630 !img_request_layered_test(img_request) ||
2631 rbd_dev->parent_overlap <= obj_request->img_offset ||
2632 ((known = obj_request_known_test(obj_request)) &&
2633 obj_request_exists_test(obj_request))) {
2635 struct rbd_device *rbd_dev;
2636 struct ceph_osd_client *osdc;
2638 rbd_dev = obj_request->img_request->rbd_dev;
2639 osdc = &rbd_dev->rbd_client->client->osdc;
2641 return rbd_obj_request_submit(osdc, obj_request);
2645 * It's a layered write. The target object might exist but
2646 * we may not know that yet. If we know it doesn't exist,
2647 * start by reading the data for the full target object from
2648 * the parent so we can use it for a copyup to the target.
2651 return rbd_img_obj_parent_read_full(obj_request);
2653 /* We don't know whether the target exists. Go find out. */
2655 return rbd_img_obj_exists_submit(obj_request);
2658 static int rbd_img_request_submit(struct rbd_img_request *img_request)
2660 struct rbd_obj_request *obj_request;
2661 struct rbd_obj_request *next_obj_request;
2663 dout("%s: img %p\n", __func__, img_request);
2664 for_each_obj_request_safe(img_request, obj_request, next_obj_request) {
2667 ret = rbd_img_obj_request_submit(obj_request);
2675 static void rbd_img_parent_read_callback(struct rbd_img_request *img_request)
2677 struct rbd_obj_request *obj_request;
2678 struct rbd_device *rbd_dev;
2681 rbd_assert(img_request_child_test(img_request));
2683 obj_request = img_request->obj_request;
2684 rbd_assert(obj_request);
2685 rbd_assert(obj_request->img_request);
2687 obj_request->result = img_request->result;
2688 if (obj_request->result)
2692 * We need to zero anything beyond the parent overlap
2693 * boundary. Since rbd_img_obj_request_read_callback()
2694 * will zero anything beyond the end of a short read, an
2695 * easy way to do this is to pretend the data from the
2696 * parent came up short--ending at the overlap boundary.
2698 rbd_assert(obj_request->img_offset < U64_MAX - obj_request->length);
2699 obj_end = obj_request->img_offset + obj_request->length;
2700 rbd_dev = obj_request->img_request->rbd_dev;
2701 if (obj_end > rbd_dev->parent_overlap) {
2704 if (obj_request->img_offset < rbd_dev->parent_overlap)
2705 xferred = rbd_dev->parent_overlap -
2706 obj_request->img_offset;
2708 obj_request->xferred = min(img_request->xferred, xferred);
2710 obj_request->xferred = img_request->xferred;
2713 rbd_img_request_put(img_request);
2714 rbd_img_obj_request_read_callback(obj_request);
2715 rbd_obj_request_complete(obj_request);
2718 static void rbd_img_parent_read(struct rbd_obj_request *obj_request)
2720 struct rbd_img_request *img_request;
2723 rbd_assert(obj_request_img_data_test(obj_request));
2724 rbd_assert(obj_request->img_request != NULL);
2725 rbd_assert(obj_request->result == (s32) -ENOENT);
2726 rbd_assert(obj_request_type_valid(obj_request->type));
2728 /* rbd_read_finish(obj_request, obj_request->length); */
2729 img_request = rbd_parent_request_create(obj_request,
2730 obj_request->img_offset,
2731 obj_request->length);
2736 if (obj_request->type == OBJ_REQUEST_BIO)
2737 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
2738 obj_request->bio_list);
2740 result = rbd_img_request_fill(img_request, OBJ_REQUEST_PAGES,
2741 obj_request->pages);
2745 img_request->callback = rbd_img_parent_read_callback;
2746 result = rbd_img_request_submit(img_request);
2753 rbd_img_request_put(img_request);
2754 obj_request->result = result;
2755 obj_request->xferred = 0;
2756 obj_request_done_set(obj_request);
2759 static int rbd_obj_notify_ack(struct rbd_device *rbd_dev, u64 notify_id)
2761 struct rbd_obj_request *obj_request;
2762 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2765 obj_request = rbd_obj_request_create(rbd_dev->header_name, 0, 0,
2766 OBJ_REQUEST_NODATA);
2771 obj_request->osd_req = rbd_osd_req_create(rbd_dev, false, obj_request);
2772 if (!obj_request->osd_req)
2774 obj_request->callback = rbd_obj_request_put;
2776 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_NOTIFY_ACK,
2778 rbd_osd_req_format_read(obj_request);
2780 ret = rbd_obj_request_submit(osdc, obj_request);
2783 rbd_obj_request_put(obj_request);
2788 static void rbd_watch_cb(u64 ver, u64 notify_id, u8 opcode, void *data)
2790 struct rbd_device *rbd_dev = (struct rbd_device *)data;
2796 dout("%s: \"%s\" notify_id %llu opcode %u\n", __func__,
2797 rbd_dev->header_name, (unsigned long long)notify_id,
2798 (unsigned int)opcode);
2799 ret = rbd_dev_refresh(rbd_dev);
2801 rbd_warn(rbd_dev, ": header refresh error (%d)\n", ret);
2803 rbd_obj_notify_ack(rbd_dev, notify_id);
2807 * Request sync osd watch/unwatch. The value of "start" determines
2808 * whether a watch request is being initiated or torn down.
2810 static int rbd_dev_header_watch_sync(struct rbd_device *rbd_dev, bool start)
2812 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2813 struct rbd_obj_request *obj_request;
2816 rbd_assert(start ^ !!rbd_dev->watch_event);
2817 rbd_assert(start ^ !!rbd_dev->watch_request);
2820 ret = ceph_osdc_create_event(osdc, rbd_watch_cb, rbd_dev,
2821 &rbd_dev->watch_event);
2824 rbd_assert(rbd_dev->watch_event != NULL);
2828 obj_request = rbd_obj_request_create(rbd_dev->header_name, 0, 0,
2829 OBJ_REQUEST_NODATA);
2833 obj_request->osd_req = rbd_osd_req_create(rbd_dev, true, obj_request);
2834 if (!obj_request->osd_req)
2838 ceph_osdc_set_request_linger(osdc, obj_request->osd_req);
2840 ceph_osdc_unregister_linger_request(osdc,
2841 rbd_dev->watch_request->osd_req);
2843 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_WATCH,
2844 rbd_dev->watch_event->cookie, 0, start ? 1 : 0);
2845 rbd_osd_req_format_write(obj_request);
2847 ret = rbd_obj_request_submit(osdc, obj_request);
2850 ret = rbd_obj_request_wait(obj_request);
2853 ret = obj_request->result;
2858 * A watch request is set to linger, so the underlying osd
2859 * request won't go away until we unregister it. We retain
2860 * a pointer to the object request during that time (in
2861 * rbd_dev->watch_request), so we'll keep a reference to
2862 * it. We'll drop that reference (below) after we've
2866 rbd_dev->watch_request = obj_request;
2871 /* We have successfully torn down the watch request */
2873 rbd_obj_request_put(rbd_dev->watch_request);
2874 rbd_dev->watch_request = NULL;
2876 /* Cancel the event if we're tearing down, or on error */
2877 ceph_osdc_cancel_event(rbd_dev->watch_event);
2878 rbd_dev->watch_event = NULL;
2880 rbd_obj_request_put(obj_request);
2886 * Synchronous osd object method call. Returns the number of bytes
2887 * returned in the outbound buffer, or a negative error code.
2889 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
2890 const char *object_name,
2891 const char *class_name,
2892 const char *method_name,
2893 const void *outbound,
2894 size_t outbound_size,
2896 size_t inbound_size)
2898 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2899 struct rbd_obj_request *obj_request;
2900 struct page **pages;
2905 * Method calls are ultimately read operations. The result
2906 * should placed into the inbound buffer provided. They
2907 * also supply outbound data--parameters for the object
2908 * method. Currently if this is present it will be a
2911 page_count = (u32)calc_pages_for(0, inbound_size);
2912 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2914 return PTR_ERR(pages);
2917 obj_request = rbd_obj_request_create(object_name, 0, inbound_size,
2922 obj_request->pages = pages;
2923 obj_request->page_count = page_count;
2925 obj_request->osd_req = rbd_osd_req_create(rbd_dev, false, obj_request);
2926 if (!obj_request->osd_req)
2929 osd_req_op_cls_init(obj_request->osd_req, 0, CEPH_OSD_OP_CALL,
2930 class_name, method_name);
2931 if (outbound_size) {
2932 struct ceph_pagelist *pagelist;
2934 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
2938 ceph_pagelist_init(pagelist);
2939 ceph_pagelist_append(pagelist, outbound, outbound_size);
2940 osd_req_op_cls_request_data_pagelist(obj_request->osd_req, 0,
2943 osd_req_op_cls_response_data_pages(obj_request->osd_req, 0,
2944 obj_request->pages, inbound_size,
2946 rbd_osd_req_format_read(obj_request);
2948 ret = rbd_obj_request_submit(osdc, obj_request);
2951 ret = rbd_obj_request_wait(obj_request);
2955 ret = obj_request->result;
2959 rbd_assert(obj_request->xferred < (u64)INT_MAX);
2960 ret = (int)obj_request->xferred;
2961 ceph_copy_from_page_vector(pages, inbound, 0, obj_request->xferred);
2964 rbd_obj_request_put(obj_request);
2966 ceph_release_page_vector(pages, page_count);
2971 static void rbd_request_fn(struct request_queue *q)
2972 __releases(q->queue_lock) __acquires(q->queue_lock)
2974 struct rbd_device *rbd_dev = q->queuedata;
2975 bool read_only = rbd_dev->mapping.read_only;
2979 while ((rq = blk_fetch_request(q))) {
2980 bool write_request = rq_data_dir(rq) == WRITE;
2981 struct rbd_img_request *img_request;
2985 /* Ignore any non-FS requests that filter through. */
2987 if (rq->cmd_type != REQ_TYPE_FS) {
2988 dout("%s: non-fs request type %d\n", __func__,
2989 (int) rq->cmd_type);
2990 __blk_end_request_all(rq, 0);
2994 /* Ignore/skip any zero-length requests */
2996 offset = (u64) blk_rq_pos(rq) << SECTOR_SHIFT;
2997 length = (u64) blk_rq_bytes(rq);
3000 dout("%s: zero-length request\n", __func__);
3001 __blk_end_request_all(rq, 0);
3005 spin_unlock_irq(q->queue_lock);
3007 /* Disallow writes to a read-only device */
3009 if (write_request) {
3013 rbd_assert(rbd_dev->spec->snap_id == CEPH_NOSNAP);
3017 * Quit early if the mapped snapshot no longer
3018 * exists. It's still possible the snapshot will
3019 * have disappeared by the time our request arrives
3020 * at the osd, but there's no sense in sending it if
3023 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
3024 dout("request for non-existent snapshot");
3025 rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
3031 if (offset && length > U64_MAX - offset + 1) {
3032 rbd_warn(rbd_dev, "bad request range (%llu~%llu)\n",
3034 goto end_request; /* Shouldn't happen */
3038 if (offset + length > rbd_dev->mapping.size) {
3039 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)\n",
3040 offset, length, rbd_dev->mapping.size);
3045 img_request = rbd_img_request_create(rbd_dev, offset, length,
3050 img_request->rq = rq;
3052 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3055 result = rbd_img_request_submit(img_request);
3057 rbd_img_request_put(img_request);
3059 spin_lock_irq(q->queue_lock);
3061 rbd_warn(rbd_dev, "%s %llx at %llx result %d\n",
3062 write_request ? "write" : "read",
3063 length, offset, result);
3065 __blk_end_request_all(rq, result);
3071 * a queue callback. Makes sure that we don't create a bio that spans across
3072 * multiple osd objects. One exception would be with a single page bios,
3073 * which we handle later at bio_chain_clone_range()
3075 static int rbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bmd,
3076 struct bio_vec *bvec)
3078 struct rbd_device *rbd_dev = q->queuedata;
3079 sector_t sector_offset;
3080 sector_t sectors_per_obj;
3081 sector_t obj_sector_offset;
3085 * Find how far into its rbd object the partition-relative
3086 * bio start sector is to offset relative to the enclosing
3089 sector_offset = get_start_sect(bmd->bi_bdev) + bmd->bi_sector;
3090 sectors_per_obj = 1 << (rbd_dev->header.obj_order - SECTOR_SHIFT);
3091 obj_sector_offset = sector_offset & (sectors_per_obj - 1);
3094 * Compute the number of bytes from that offset to the end
3095 * of the object. Account for what's already used by the bio.
3097 ret = (int) (sectors_per_obj - obj_sector_offset) << SECTOR_SHIFT;
3098 if (ret > bmd->bi_size)
3099 ret -= bmd->bi_size;
3104 * Don't send back more than was asked for. And if the bio
3105 * was empty, let the whole thing through because: "Note
3106 * that a block device *must* allow a single page to be
3107 * added to an empty bio."
3109 rbd_assert(bvec->bv_len <= PAGE_SIZE);
3110 if (ret > (int) bvec->bv_len || !bmd->bi_size)
3111 ret = (int) bvec->bv_len;
3116 static void rbd_free_disk(struct rbd_device *rbd_dev)
3118 struct gendisk *disk = rbd_dev->disk;
3123 rbd_dev->disk = NULL;
3124 if (disk->flags & GENHD_FL_UP) {
3127 blk_cleanup_queue(disk->queue);
3132 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
3133 const char *object_name,
3134 u64 offset, u64 length, void *buf)
3137 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3138 struct rbd_obj_request *obj_request;
3139 struct page **pages = NULL;
3144 page_count = (u32) calc_pages_for(offset, length);
3145 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
3147 ret = PTR_ERR(pages);
3150 obj_request = rbd_obj_request_create(object_name, offset, length,
3155 obj_request->pages = pages;
3156 obj_request->page_count = page_count;
3158 obj_request->osd_req = rbd_osd_req_create(rbd_dev, false, obj_request);
3159 if (!obj_request->osd_req)
3162 osd_req_op_extent_init(obj_request->osd_req, 0, CEPH_OSD_OP_READ,
3163 offset, length, 0, 0);
3164 osd_req_op_extent_osd_data_pages(obj_request->osd_req, 0,
3166 obj_request->length,
3167 obj_request->offset & ~PAGE_MASK,
3169 rbd_osd_req_format_read(obj_request);
3171 ret = rbd_obj_request_submit(osdc, obj_request);
3174 ret = rbd_obj_request_wait(obj_request);
3178 ret = obj_request->result;
3182 rbd_assert(obj_request->xferred <= (u64) SIZE_MAX);
3183 size = (size_t) obj_request->xferred;
3184 ceph_copy_from_page_vector(pages, buf, 0, size);
3185 rbd_assert(size <= (size_t)INT_MAX);
3189 rbd_obj_request_put(obj_request);
3191 ceph_release_page_vector(pages, page_count);
3197 * Read the complete header for the given rbd device. On successful
3198 * return, the rbd_dev->header field will contain up-to-date
3199 * information about the image.
3201 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev)
3203 struct rbd_image_header_ondisk *ondisk = NULL;
3210 * The complete header will include an array of its 64-bit
3211 * snapshot ids, followed by the names of those snapshots as
3212 * a contiguous block of NUL-terminated strings. Note that
3213 * the number of snapshots could change by the time we read
3214 * it in, in which case we re-read it.
3221 size = sizeof (*ondisk);
3222 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
3224 ondisk = kmalloc(size, GFP_KERNEL);
3228 ret = rbd_obj_read_sync(rbd_dev, rbd_dev->header_name,
3232 if ((size_t)ret < size) {
3234 rbd_warn(rbd_dev, "short header read (want %zd got %d)",
3238 if (!rbd_dev_ondisk_valid(ondisk)) {
3240 rbd_warn(rbd_dev, "invalid header");
3244 names_size = le64_to_cpu(ondisk->snap_names_len);
3245 want_count = snap_count;
3246 snap_count = le32_to_cpu(ondisk->snap_count);
3247 } while (snap_count != want_count);
3249 ret = rbd_header_from_disk(rbd_dev, ondisk);
3257 * Clear the rbd device's EXISTS flag if the snapshot it's mapped to
3258 * has disappeared from the (just updated) snapshot context.
3260 static void rbd_exists_validate(struct rbd_device *rbd_dev)
3264 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags))
3267 snap_id = rbd_dev->spec->snap_id;
3268 if (snap_id == CEPH_NOSNAP)
3271 if (rbd_dev_snap_index(rbd_dev, snap_id) == BAD_SNAP_INDEX)
3272 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
3275 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
3280 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
3281 mapping_size = rbd_dev->mapping.size;
3282 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
3283 if (rbd_dev->image_format == 1)
3284 ret = rbd_dev_v1_header_info(rbd_dev);
3286 ret = rbd_dev_v2_header_info(rbd_dev);
3288 /* If it's a mapped snapshot, validate its EXISTS flag */
3290 rbd_exists_validate(rbd_dev);
3291 mutex_unlock(&ctl_mutex);
3292 if (mapping_size != rbd_dev->mapping.size) {
3295 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
3296 dout("setting size to %llu sectors", (unsigned long long)size);
3297 set_capacity(rbd_dev->disk, size);
3298 revalidate_disk(rbd_dev->disk);
3304 static int rbd_init_disk(struct rbd_device *rbd_dev)
3306 struct gendisk *disk;
3307 struct request_queue *q;
3310 /* create gendisk info */
3311 disk = alloc_disk(RBD_MINORS_PER_MAJOR);
3315 snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
3317 disk->major = rbd_dev->major;
3318 disk->first_minor = 0;
3319 disk->fops = &rbd_bd_ops;
3320 disk->private_data = rbd_dev;
3322 q = blk_init_queue(rbd_request_fn, &rbd_dev->lock);
3326 /* We use the default size, but let's be explicit about it. */
3327 blk_queue_physical_block_size(q, SECTOR_SIZE);
3329 /* set io sizes to object size */
3330 segment_size = rbd_obj_bytes(&rbd_dev->header);
3331 blk_queue_max_hw_sectors(q, segment_size / SECTOR_SIZE);
3332 blk_queue_max_segment_size(q, segment_size);
3333 blk_queue_io_min(q, segment_size);
3334 blk_queue_io_opt(q, segment_size);
3336 blk_queue_merge_bvec(q, rbd_merge_bvec);
3339 q->queuedata = rbd_dev;
3341 rbd_dev->disk = disk;
3354 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
3356 return container_of(dev, struct rbd_device, dev);
3359 static ssize_t rbd_size_show(struct device *dev,
3360 struct device_attribute *attr, char *buf)
3362 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3364 return sprintf(buf, "%llu\n",
3365 (unsigned long long)rbd_dev->mapping.size);
3369 * Note this shows the features for whatever's mapped, which is not
3370 * necessarily the base image.
3372 static ssize_t rbd_features_show(struct device *dev,
3373 struct device_attribute *attr, char *buf)
3375 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3377 return sprintf(buf, "0x%016llx\n",
3378 (unsigned long long)rbd_dev->mapping.features);
3381 static ssize_t rbd_major_show(struct device *dev,
3382 struct device_attribute *attr, char *buf)
3384 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3387 return sprintf(buf, "%d\n", rbd_dev->major);
3389 return sprintf(buf, "(none)\n");
3393 static ssize_t rbd_client_id_show(struct device *dev,
3394 struct device_attribute *attr, char *buf)
3396 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3398 return sprintf(buf, "client%lld\n",
3399 ceph_client_id(rbd_dev->rbd_client->client));
3402 static ssize_t rbd_pool_show(struct device *dev,
3403 struct device_attribute *attr, char *buf)
3405 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3407 return sprintf(buf, "%s\n", rbd_dev->spec->pool_name);
3410 static ssize_t rbd_pool_id_show(struct device *dev,
3411 struct device_attribute *attr, char *buf)
3413 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3415 return sprintf(buf, "%llu\n",
3416 (unsigned long long) rbd_dev->spec->pool_id);
3419 static ssize_t rbd_name_show(struct device *dev,
3420 struct device_attribute *attr, char *buf)
3422 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3424 if (rbd_dev->spec->image_name)
3425 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
3427 return sprintf(buf, "(unknown)\n");
3430 static ssize_t rbd_image_id_show(struct device *dev,
3431 struct device_attribute *attr, char *buf)
3433 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3435 return sprintf(buf, "%s\n", rbd_dev->spec->image_id);
3439 * Shows the name of the currently-mapped snapshot (or
3440 * RBD_SNAP_HEAD_NAME for the base image).
3442 static ssize_t rbd_snap_show(struct device *dev,
3443 struct device_attribute *attr,
3446 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3448 return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
3452 * For an rbd v2 image, shows the pool id, image id, and snapshot id
3453 * for the parent image. If there is no parent, simply shows
3454 * "(no parent image)".
3456 static ssize_t rbd_parent_show(struct device *dev,
3457 struct device_attribute *attr,
3460 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3461 struct rbd_spec *spec = rbd_dev->parent_spec;
3466 return sprintf(buf, "(no parent image)\n");
3468 count = sprintf(bufp, "pool_id %llu\npool_name %s\n",
3469 (unsigned long long) spec->pool_id, spec->pool_name);
3474 count = sprintf(bufp, "image_id %s\nimage_name %s\n", spec->image_id,
3475 spec->image_name ? spec->image_name : "(unknown)");
3480 count = sprintf(bufp, "snap_id %llu\nsnap_name %s\n",
3481 (unsigned long long) spec->snap_id, spec->snap_name);
3486 count = sprintf(bufp, "overlap %llu\n", rbd_dev->parent_overlap);
3491 return (ssize_t) (bufp - buf);
3494 static ssize_t rbd_image_refresh(struct device *dev,
3495 struct device_attribute *attr,
3499 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3502 ret = rbd_dev_refresh(rbd_dev);
3504 rbd_warn(rbd_dev, ": manual header refresh error (%d)\n", ret);
3506 return ret < 0 ? ret : size;
3509 static DEVICE_ATTR(size, S_IRUGO, rbd_size_show, NULL);
3510 static DEVICE_ATTR(features, S_IRUGO, rbd_features_show, NULL);
3511 static DEVICE_ATTR(major, S_IRUGO, rbd_major_show, NULL);
3512 static DEVICE_ATTR(client_id, S_IRUGO, rbd_client_id_show, NULL);
3513 static DEVICE_ATTR(pool, S_IRUGO, rbd_pool_show, NULL);
3514 static DEVICE_ATTR(pool_id, S_IRUGO, rbd_pool_id_show, NULL);
3515 static DEVICE_ATTR(name, S_IRUGO, rbd_name_show, NULL);
3516 static DEVICE_ATTR(image_id, S_IRUGO, rbd_image_id_show, NULL);
3517 static DEVICE_ATTR(refresh, S_IWUSR, NULL, rbd_image_refresh);
3518 static DEVICE_ATTR(current_snap, S_IRUGO, rbd_snap_show, NULL);
3519 static DEVICE_ATTR(parent, S_IRUGO, rbd_parent_show, NULL);
3521 static struct attribute *rbd_attrs[] = {
3522 &dev_attr_size.attr,
3523 &dev_attr_features.attr,
3524 &dev_attr_major.attr,
3525 &dev_attr_client_id.attr,
3526 &dev_attr_pool.attr,
3527 &dev_attr_pool_id.attr,
3528 &dev_attr_name.attr,
3529 &dev_attr_image_id.attr,
3530 &dev_attr_current_snap.attr,
3531 &dev_attr_parent.attr,
3532 &dev_attr_refresh.attr,
3536 static struct attribute_group rbd_attr_group = {
3540 static const struct attribute_group *rbd_attr_groups[] = {
3545 static void rbd_sysfs_dev_release(struct device *dev)
3549 static struct device_type rbd_device_type = {
3551 .groups = rbd_attr_groups,
3552 .release = rbd_sysfs_dev_release,
3555 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
3557 kref_get(&spec->kref);
3562 static void rbd_spec_free(struct kref *kref);
3563 static void rbd_spec_put(struct rbd_spec *spec)
3566 kref_put(&spec->kref, rbd_spec_free);
3569 static struct rbd_spec *rbd_spec_alloc(void)
3571 struct rbd_spec *spec;
3573 spec = kzalloc(sizeof (*spec), GFP_KERNEL);
3576 kref_init(&spec->kref);
3581 static void rbd_spec_free(struct kref *kref)
3583 struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
3585 kfree(spec->pool_name);
3586 kfree(spec->image_id);
3587 kfree(spec->image_name);
3588 kfree(spec->snap_name);
3592 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
3593 struct rbd_spec *spec)
3595 struct rbd_device *rbd_dev;
3597 rbd_dev = kzalloc(sizeof (*rbd_dev), GFP_KERNEL);
3601 spin_lock_init(&rbd_dev->lock);
3603 atomic_set(&rbd_dev->parent_ref, 0);
3604 INIT_LIST_HEAD(&rbd_dev->node);
3605 init_rwsem(&rbd_dev->header_rwsem);
3607 rbd_dev->spec = spec;
3608 rbd_dev->rbd_client = rbdc;
3610 /* Initialize the layout used for all rbd requests */
3612 rbd_dev->layout.fl_stripe_unit = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
3613 rbd_dev->layout.fl_stripe_count = cpu_to_le32(1);
3614 rbd_dev->layout.fl_object_size = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
3615 rbd_dev->layout.fl_pg_pool = cpu_to_le32((u32) spec->pool_id);
3620 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
3622 rbd_put_client(rbd_dev->rbd_client);
3623 rbd_spec_put(rbd_dev->spec);
3628 * Get the size and object order for an image snapshot, or if
3629 * snap_id is CEPH_NOSNAP, gets this information for the base
3632 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
3633 u8 *order, u64 *snap_size)
3635 __le64 snapid = cpu_to_le64(snap_id);
3640 } __attribute__ ((packed)) size_buf = { 0 };
3642 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3644 &snapid, sizeof (snapid),
3645 &size_buf, sizeof (size_buf));
3646 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3649 if (ret < sizeof (size_buf))
3653 *order = size_buf.order;
3654 *snap_size = le64_to_cpu(size_buf.size);
3656 dout(" snap_id 0x%016llx order = %u, snap_size = %llu\n",
3657 (unsigned long long)snap_id, (unsigned int)*order,
3658 (unsigned long long)*snap_size);
3663 static int rbd_dev_v2_image_size(struct rbd_device *rbd_dev)
3665 return _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
3666 &rbd_dev->header.obj_order,
3667 &rbd_dev->header.image_size);
3670 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev)
3676 reply_buf = kzalloc(RBD_OBJ_PREFIX_LEN_MAX, GFP_KERNEL);
3680 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3681 "rbd", "get_object_prefix", NULL, 0,
3682 reply_buf, RBD_OBJ_PREFIX_LEN_MAX);
3683 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3688 rbd_dev->header.object_prefix = ceph_extract_encoded_string(&p,
3689 p + ret, NULL, GFP_NOIO);
3692 if (IS_ERR(rbd_dev->header.object_prefix)) {
3693 ret = PTR_ERR(rbd_dev->header.object_prefix);
3694 rbd_dev->header.object_prefix = NULL;
3696 dout(" object_prefix = %s\n", rbd_dev->header.object_prefix);
3704 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
3707 __le64 snapid = cpu_to_le64(snap_id);
3711 } __attribute__ ((packed)) features_buf = { 0 };
3715 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3716 "rbd", "get_features",
3717 &snapid, sizeof (snapid),
3718 &features_buf, sizeof (features_buf));
3719 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3722 if (ret < sizeof (features_buf))
3725 incompat = le64_to_cpu(features_buf.incompat);
3726 if (incompat & ~RBD_FEATURES_SUPPORTED)
3729 *snap_features = le64_to_cpu(features_buf.features);
3731 dout(" snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
3732 (unsigned long long)snap_id,
3733 (unsigned long long)*snap_features,
3734 (unsigned long long)le64_to_cpu(features_buf.incompat));
3739 static int rbd_dev_v2_features(struct rbd_device *rbd_dev)
3741 return _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
3742 &rbd_dev->header.features);
3745 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev)
3747 struct rbd_spec *parent_spec;
3749 void *reply_buf = NULL;
3758 parent_spec = rbd_spec_alloc();
3762 size = sizeof (__le64) + /* pool_id */
3763 sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX + /* image_id */
3764 sizeof (__le64) + /* snap_id */
3765 sizeof (__le64); /* overlap */
3766 reply_buf = kmalloc(size, GFP_KERNEL);
3772 snapid = cpu_to_le64(CEPH_NOSNAP);
3773 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3774 "rbd", "get_parent",
3775 &snapid, sizeof (snapid),
3777 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3782 end = reply_buf + ret;
3784 ceph_decode_64_safe(&p, end, pool_id, out_err);
3785 if (pool_id == CEPH_NOPOOL)
3786 goto out; /* No parent? No problem. */
3788 /* The ceph file layout needs to fit pool id in 32 bits */
3791 if (pool_id > (u64)U32_MAX) {
3792 rbd_warn(NULL, "parent pool id too large (%llu > %u)\n",
3793 (unsigned long long)pool_id, U32_MAX);
3796 parent_spec->pool_id = pool_id;
3798 image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
3799 if (IS_ERR(image_id)) {
3800 ret = PTR_ERR(image_id);
3803 parent_spec->image_id = image_id;
3804 ceph_decode_64_safe(&p, end, parent_spec->snap_id, out_err);
3805 ceph_decode_64_safe(&p, end, overlap, out_err);
3808 rbd_spec_put(rbd_dev->parent_spec);
3809 rbd_dev->parent_spec = parent_spec;
3810 parent_spec = NULL; /* rbd_dev now owns this */
3811 rbd_dev->parent_overlap = overlap;
3813 rbd_warn(rbd_dev, "ignoring parent of clone with overlap 0\n");
3819 rbd_spec_put(parent_spec);
3824 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev)
3828 __le64 stripe_count;
3829 } __attribute__ ((packed)) striping_info_buf = { 0 };
3830 size_t size = sizeof (striping_info_buf);
3837 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
3838 "rbd", "get_stripe_unit_count", NULL, 0,
3839 (char *)&striping_info_buf, size);
3840 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
3847 * We don't actually support the "fancy striping" feature
3848 * (STRIPINGV2) yet, but if the striping sizes are the
3849 * defaults the behavior is the same as before. So find
3850 * out, and only fail if the image has non-default values.
3853 obj_size = (u64)1 << rbd_dev->header.obj_order;
3854 p = &striping_info_buf;
3855 stripe_unit = ceph_decode_64(&p);
3856 if (stripe_unit != obj_size) {
3857 rbd_warn(rbd_dev, "unsupported stripe unit "
3858 "(got %llu want %llu)",
3859 stripe_unit, obj_size);
3862 stripe_count = ceph_decode_64(&p);
3863 if (stripe_count != 1) {
3864 rbd_warn(rbd_dev, "unsupported stripe count "
3865 "(got %llu want 1)", stripe_count);
3868 rbd_dev->header.stripe_unit = stripe_unit;
3869 rbd_dev->header.stripe_count = stripe_count;
3874 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
3876 size_t image_id_size;
3881 void *reply_buf = NULL;
3883 char *image_name = NULL;
3886 rbd_assert(!rbd_dev->spec->image_name);
3888 len = strlen(rbd_dev->spec->image_id);
3889 image_id_size = sizeof (__le32) + len;
3890 image_id = kmalloc(image_id_size, GFP_KERNEL);
3895 end = image_id + image_id_size;
3896 ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
3898 size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
3899 reply_buf = kmalloc(size, GFP_KERNEL);
3903 ret = rbd_obj_method_sync(rbd_dev, RBD_DIRECTORY,
3904 "rbd", "dir_get_name",
3905 image_id, image_id_size,
3910 end = reply_buf + ret;
3912 image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
3913 if (IS_ERR(image_name))
3916 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
3924 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
3926 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
3927 const char *snap_name;
3930 /* Skip over names until we find the one we are looking for */
3932 snap_name = rbd_dev->header.snap_names;
3933 while (which < snapc->num_snaps) {
3934 if (!strcmp(name, snap_name))
3935 return snapc->snaps[which];
3936 snap_name += strlen(snap_name) + 1;
3942 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
3944 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
3949 for (which = 0; !found && which < snapc->num_snaps; which++) {
3950 const char *snap_name;
3952 snap_id = snapc->snaps[which];
3953 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
3954 if (IS_ERR(snap_name))
3956 found = !strcmp(name, snap_name);
3959 return found ? snap_id : CEPH_NOSNAP;
3963 * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
3964 * no snapshot by that name is found, or if an error occurs.
3966 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
3968 if (rbd_dev->image_format == 1)
3969 return rbd_v1_snap_id_by_name(rbd_dev, name);
3971 return rbd_v2_snap_id_by_name(rbd_dev, name);
3975 * When an rbd image has a parent image, it is identified by the
3976 * pool, image, and snapshot ids (not names). This function fills
3977 * in the names for those ids. (It's OK if we can't figure out the
3978 * name for an image id, but the pool and snapshot ids should always
3979 * exist and have names.) All names in an rbd spec are dynamically
3982 * When an image being mapped (not a parent) is probed, we have the
3983 * pool name and pool id, image name and image id, and the snapshot
3984 * name. The only thing we're missing is the snapshot id.
3986 static int rbd_dev_spec_update(struct rbd_device *rbd_dev)
3988 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3989 struct rbd_spec *spec = rbd_dev->spec;
3990 const char *pool_name;
3991 const char *image_name;
3992 const char *snap_name;
3996 * An image being mapped will have the pool name (etc.), but
3997 * we need to look up the snapshot id.
3999 if (spec->pool_name) {
4000 if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
4003 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
4004 if (snap_id == CEPH_NOSNAP)
4006 spec->snap_id = snap_id;
4008 spec->snap_id = CEPH_NOSNAP;
4014 /* Get the pool name; we have to make our own copy of this */
4016 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
4018 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
4021 pool_name = kstrdup(pool_name, GFP_KERNEL);
4025 /* Fetch the image name; tolerate failure here */
4027 image_name = rbd_dev_image_name(rbd_dev);
4029 rbd_warn(rbd_dev, "unable to get image name");
4031 /* Look up the snapshot name, and make a copy */
4033 snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
4039 spec->pool_name = pool_name;
4040 spec->image_name = image_name;
4041 spec->snap_name = snap_name;
4051 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev)
4060 struct ceph_snap_context *snapc;
4064 * We'll need room for the seq value (maximum snapshot id),
4065 * snapshot count, and array of that many snapshot ids.
4066 * For now we have a fixed upper limit on the number we're
4067 * prepared to receive.
4069 size = sizeof (__le64) + sizeof (__le32) +
4070 RBD_MAX_SNAP_COUNT * sizeof (__le64);
4071 reply_buf = kzalloc(size, GFP_KERNEL);
4075 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4076 "rbd", "get_snapcontext", NULL, 0,
4078 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4083 end = reply_buf + ret;
4085 ceph_decode_64_safe(&p, end, seq, out);
4086 ceph_decode_32_safe(&p, end, snap_count, out);
4089 * Make sure the reported number of snapshot ids wouldn't go
4090 * beyond the end of our buffer. But before checking that,
4091 * make sure the computed size of the snapshot context we
4092 * allocate is representable in a size_t.
4094 if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
4099 if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
4103 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
4109 for (i = 0; i < snap_count; i++)
4110 snapc->snaps[i] = ceph_decode_64(&p);
4112 ceph_put_snap_context(rbd_dev->header.snapc);
4113 rbd_dev->header.snapc = snapc;
4115 dout(" snap context seq = %llu, snap_count = %u\n",
4116 (unsigned long long)seq, (unsigned int)snap_count);
4123 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
4134 size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
4135 reply_buf = kmalloc(size, GFP_KERNEL);
4137 return ERR_PTR(-ENOMEM);
4139 snapid = cpu_to_le64(snap_id);
4140 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4141 "rbd", "get_snapshot_name",
4142 &snapid, sizeof (snapid),
4144 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4146 snap_name = ERR_PTR(ret);
4151 end = reply_buf + ret;
4152 snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
4153 if (IS_ERR(snap_name))
4156 dout(" snap_id 0x%016llx snap_name = %s\n",
4157 (unsigned long long)snap_id, snap_name);
4164 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev)
4166 bool first_time = rbd_dev->header.object_prefix == NULL;
4169 down_write(&rbd_dev->header_rwsem);
4172 ret = rbd_dev_v2_header_onetime(rbd_dev);
4178 * If the image supports layering, get the parent info. We
4179 * need to probe the first time regardless. Thereafter we
4180 * only need to if there's a parent, to see if it has
4181 * disappeared due to the mapped image getting flattened.
4183 if (rbd_dev->header.features & RBD_FEATURE_LAYERING &&
4184 (first_time || rbd_dev->parent_spec)) {
4187 ret = rbd_dev_v2_parent_info(rbd_dev);
4192 * Print a warning if this is the initial probe and
4193 * the image has a parent. Don't print it if the
4194 * image now being probed is itself a parent. We
4195 * can tell at this point because we won't know its
4196 * pool name yet (just its pool id).
4198 warn = rbd_dev->parent_spec && rbd_dev->spec->pool_name;
4199 if (first_time && warn)
4200 rbd_warn(rbd_dev, "WARNING: kernel layering "
4201 "is EXPERIMENTAL!");
4204 ret = rbd_dev_v2_image_size(rbd_dev);
4208 if (rbd_dev->spec->snap_id == CEPH_NOSNAP)
4209 if (rbd_dev->mapping.size != rbd_dev->header.image_size)
4210 rbd_dev->mapping.size = rbd_dev->header.image_size;
4212 ret = rbd_dev_v2_snap_context(rbd_dev);
4213 dout("rbd_dev_v2_snap_context returned %d\n", ret);
4215 up_write(&rbd_dev->header_rwsem);
4220 static int rbd_bus_add_dev(struct rbd_device *rbd_dev)
4225 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
4227 dev = &rbd_dev->dev;
4228 dev->bus = &rbd_bus_type;
4229 dev->type = &rbd_device_type;
4230 dev->parent = &rbd_root_dev;
4231 dev->release = rbd_dev_device_release;
4232 dev_set_name(dev, "%d", rbd_dev->dev_id);
4233 ret = device_register(dev);
4235 mutex_unlock(&ctl_mutex);
4240 static void rbd_bus_del_dev(struct rbd_device *rbd_dev)
4242 device_unregister(&rbd_dev->dev);
4245 static atomic64_t rbd_dev_id_max = ATOMIC64_INIT(0);
4248 * Get a unique rbd identifier for the given new rbd_dev, and add
4249 * the rbd_dev to the global list. The minimum rbd id is 1.
4251 static void rbd_dev_id_get(struct rbd_device *rbd_dev)
4253 rbd_dev->dev_id = atomic64_inc_return(&rbd_dev_id_max);
4255 spin_lock(&rbd_dev_list_lock);
4256 list_add_tail(&rbd_dev->node, &rbd_dev_list);
4257 spin_unlock(&rbd_dev_list_lock);
4258 dout("rbd_dev %p given dev id %llu\n", rbd_dev,
4259 (unsigned long long) rbd_dev->dev_id);
4263 * Remove an rbd_dev from the global list, and record that its
4264 * identifier is no longer in use.
4266 static void rbd_dev_id_put(struct rbd_device *rbd_dev)
4268 struct list_head *tmp;
4269 int rbd_id = rbd_dev->dev_id;
4272 rbd_assert(rbd_id > 0);
4274 dout("rbd_dev %p released dev id %llu\n", rbd_dev,
4275 (unsigned long long) rbd_dev->dev_id);
4276 spin_lock(&rbd_dev_list_lock);
4277 list_del_init(&rbd_dev->node);
4280 * If the id being "put" is not the current maximum, there
4281 * is nothing special we need to do.
4283 if (rbd_id != atomic64_read(&rbd_dev_id_max)) {
4284 spin_unlock(&rbd_dev_list_lock);
4289 * We need to update the current maximum id. Search the
4290 * list to find out what it is. We're more likely to find
4291 * the maximum at the end, so search the list backward.
4294 list_for_each_prev(tmp, &rbd_dev_list) {
4295 struct rbd_device *rbd_dev;
4297 rbd_dev = list_entry(tmp, struct rbd_device, node);
4298 if (rbd_dev->dev_id > max_id)
4299 max_id = rbd_dev->dev_id;
4301 spin_unlock(&rbd_dev_list_lock);
4304 * The max id could have been updated by rbd_dev_id_get(), in
4305 * which case it now accurately reflects the new maximum.
4306 * Be careful not to overwrite the maximum value in that
4309 atomic64_cmpxchg(&rbd_dev_id_max, rbd_id, max_id);
4310 dout(" max dev id has been reset\n");
4314 * Skips over white space at *buf, and updates *buf to point to the
4315 * first found non-space character (if any). Returns the length of
4316 * the token (string of non-white space characters) found. Note
4317 * that *buf must be terminated with '\0'.
4319 static inline size_t next_token(const char **buf)
4322 * These are the characters that produce nonzero for
4323 * isspace() in the "C" and "POSIX" locales.
4325 const char *spaces = " \f\n\r\t\v";
4327 *buf += strspn(*buf, spaces); /* Find start of token */
4329 return strcspn(*buf, spaces); /* Return token length */
4333 * Finds the next token in *buf, and if the provided token buffer is
4334 * big enough, copies the found token into it. The result, if
4335 * copied, is guaranteed to be terminated with '\0'. Note that *buf
4336 * must be terminated with '\0' on entry.
4338 * Returns the length of the token found (not including the '\0').
4339 * Return value will be 0 if no token is found, and it will be >=
4340 * token_size if the token would not fit.
4342 * The *buf pointer will be updated to point beyond the end of the
4343 * found token. Note that this occurs even if the token buffer is
4344 * too small to hold it.
4346 static inline size_t copy_token(const char **buf,
4352 len = next_token(buf);
4353 if (len < token_size) {
4354 memcpy(token, *buf, len);
4355 *(token + len) = '\0';
4363 * Finds the next token in *buf, dynamically allocates a buffer big
4364 * enough to hold a copy of it, and copies the token into the new
4365 * buffer. The copy is guaranteed to be terminated with '\0'. Note
4366 * that a duplicate buffer is created even for a zero-length token.
4368 * Returns a pointer to the newly-allocated duplicate, or a null
4369 * pointer if memory for the duplicate was not available. If
4370 * the lenp argument is a non-null pointer, the length of the token
4371 * (not including the '\0') is returned in *lenp.
4373 * If successful, the *buf pointer will be updated to point beyond
4374 * the end of the found token.
4376 * Note: uses GFP_KERNEL for allocation.
4378 static inline char *dup_token(const char **buf, size_t *lenp)
4383 len = next_token(buf);
4384 dup = kmemdup(*buf, len + 1, GFP_KERNEL);
4387 *(dup + len) = '\0';
4397 * Parse the options provided for an "rbd add" (i.e., rbd image
4398 * mapping) request. These arrive via a write to /sys/bus/rbd/add,
4399 * and the data written is passed here via a NUL-terminated buffer.
4400 * Returns 0 if successful or an error code otherwise.
4402 * The information extracted from these options is recorded in
4403 * the other parameters which return dynamically-allocated
4406 * The address of a pointer that will refer to a ceph options
4407 * structure. Caller must release the returned pointer using
4408 * ceph_destroy_options() when it is no longer needed.
4410 * Address of an rbd options pointer. Fully initialized by
4411 * this function; caller must release with kfree().
4413 * Address of an rbd image specification pointer. Fully
4414 * initialized by this function based on parsed options.
4415 * Caller must release with rbd_spec_put().
4417 * The options passed take this form:
4418 * <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
4421 * A comma-separated list of one or more monitor addresses.
4422 * A monitor address is an ip address, optionally followed
4423 * by a port number (separated by a colon).
4424 * I.e.: ip1[:port1][,ip2[:port2]...]
4426 * A comma-separated list of ceph and/or rbd options.
4428 * The name of the rados pool containing the rbd image.
4430 * The name of the image in that pool to map.
4432 * An optional snapshot id. If provided, the mapping will
4433 * present data from the image at the time that snapshot was
4434 * created. The image head is used if no snapshot id is
4435 * provided. Snapshot mappings are always read-only.
4437 static int rbd_add_parse_args(const char *buf,
4438 struct ceph_options **ceph_opts,
4439 struct rbd_options **opts,
4440 struct rbd_spec **rbd_spec)
4444 const char *mon_addrs;
4446 size_t mon_addrs_size;
4447 struct rbd_spec *spec = NULL;
4448 struct rbd_options *rbd_opts = NULL;
4449 struct ceph_options *copts;
4452 /* The first four tokens are required */
4454 len = next_token(&buf);
4456 rbd_warn(NULL, "no monitor address(es) provided");
4460 mon_addrs_size = len + 1;
4464 options = dup_token(&buf, NULL);
4468 rbd_warn(NULL, "no options provided");
4472 spec = rbd_spec_alloc();
4476 spec->pool_name = dup_token(&buf, NULL);
4477 if (!spec->pool_name)
4479 if (!*spec->pool_name) {
4480 rbd_warn(NULL, "no pool name provided");
4484 spec->image_name = dup_token(&buf, NULL);
4485 if (!spec->image_name)
4487 if (!*spec->image_name) {
4488 rbd_warn(NULL, "no image name provided");
4493 * Snapshot name is optional; default is to use "-"
4494 * (indicating the head/no snapshot).
4496 len = next_token(&buf);
4498 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
4499 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
4500 } else if (len > RBD_MAX_SNAP_NAME_LEN) {
4501 ret = -ENAMETOOLONG;
4504 snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
4507 *(snap_name + len) = '\0';
4508 spec->snap_name = snap_name;
4510 /* Initialize all rbd options to the defaults */
4512 rbd_opts = kzalloc(sizeof (*rbd_opts), GFP_KERNEL);
4516 rbd_opts->read_only = RBD_READ_ONLY_DEFAULT;
4518 copts = ceph_parse_options(options, mon_addrs,
4519 mon_addrs + mon_addrs_size - 1,
4520 parse_rbd_opts_token, rbd_opts);
4521 if (IS_ERR(copts)) {
4522 ret = PTR_ERR(copts);
4543 * An rbd format 2 image has a unique identifier, distinct from the
4544 * name given to it by the user. Internally, that identifier is
4545 * what's used to specify the names of objects related to the image.
4547 * A special "rbd id" object is used to map an rbd image name to its
4548 * id. If that object doesn't exist, then there is no v2 rbd image
4549 * with the supplied name.
4551 * This function will record the given rbd_dev's image_id field if
4552 * it can be determined, and in that case will return 0. If any
4553 * errors occur a negative errno will be returned and the rbd_dev's
4554 * image_id field will be unchanged (and should be NULL).
4556 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
4565 * When probing a parent image, the image id is already
4566 * known (and the image name likely is not). There's no
4567 * need to fetch the image id again in this case. We
4568 * do still need to set the image format though.
4570 if (rbd_dev->spec->image_id) {
4571 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
4577 * First, see if the format 2 image id file exists, and if
4578 * so, get the image's persistent id from it.
4580 size = sizeof (RBD_ID_PREFIX) + strlen(rbd_dev->spec->image_name);
4581 object_name = kmalloc(size, GFP_NOIO);
4584 sprintf(object_name, "%s%s", RBD_ID_PREFIX, rbd_dev->spec->image_name);
4585 dout("rbd id object name is %s\n", object_name);
4587 /* Response will be an encoded string, which includes a length */
4589 size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
4590 response = kzalloc(size, GFP_NOIO);
4596 /* If it doesn't exist we'll assume it's a format 1 image */
4598 ret = rbd_obj_method_sync(rbd_dev, object_name,
4599 "rbd", "get_id", NULL, 0,
4600 response, RBD_IMAGE_ID_LEN_MAX);
4601 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4602 if (ret == -ENOENT) {
4603 image_id = kstrdup("", GFP_KERNEL);
4604 ret = image_id ? 0 : -ENOMEM;
4606 rbd_dev->image_format = 1;
4607 } else if (ret > sizeof (__le32)) {
4610 image_id = ceph_extract_encoded_string(&p, p + ret,
4612 ret = IS_ERR(image_id) ? PTR_ERR(image_id) : 0;
4614 rbd_dev->image_format = 2;
4620 rbd_dev->spec->image_id = image_id;
4621 dout("image_id is %s\n", image_id);
4630 /* Undo whatever state changes are made by v1 or v2 image probe */
4632 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
4634 struct rbd_image_header *header;
4636 rbd_dev_parent_put(rbd_dev);
4638 /* Free dynamic fields from the header, then zero it out */
4640 header = &rbd_dev->header;
4641 ceph_put_snap_context(header->snapc);
4642 kfree(header->snap_sizes);
4643 kfree(header->snap_names);
4644 kfree(header->object_prefix);
4645 memset(header, 0, sizeof (*header));
4648 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev)
4652 ret = rbd_dev_v2_object_prefix(rbd_dev);
4657 * Get the and check features for the image. Currently the
4658 * features are assumed to never change.
4660 ret = rbd_dev_v2_features(rbd_dev);
4664 /* If the image supports fancy striping, get its parameters */
4666 if (rbd_dev->header.features & RBD_FEATURE_STRIPINGV2) {
4667 ret = rbd_dev_v2_striping_info(rbd_dev);
4671 /* No support for crypto and compression type format 2 images */
4675 rbd_dev->header.features = 0;
4676 kfree(rbd_dev->header.object_prefix);
4677 rbd_dev->header.object_prefix = NULL;
4682 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev)
4684 struct rbd_device *parent = NULL;
4685 struct rbd_spec *parent_spec;
4686 struct rbd_client *rbdc;
4689 if (!rbd_dev->parent_spec)
4692 * We need to pass a reference to the client and the parent
4693 * spec when creating the parent rbd_dev. Images related by
4694 * parent/child relationships always share both.
4696 parent_spec = rbd_spec_get(rbd_dev->parent_spec);
4697 rbdc = __rbd_get_client(rbd_dev->rbd_client);
4700 parent = rbd_dev_create(rbdc, parent_spec);
4704 ret = rbd_dev_image_probe(parent, false);
4707 rbd_dev->parent = parent;
4708 atomic_set(&rbd_dev->parent_ref, 1);
4713 rbd_dev_unparent(rbd_dev);
4714 kfree(rbd_dev->header_name);
4715 rbd_dev_destroy(parent);
4717 rbd_put_client(rbdc);
4718 rbd_spec_put(parent_spec);
4724 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
4728 /* generate unique id: find highest unique id, add one */
4729 rbd_dev_id_get(rbd_dev);
4731 /* Fill in the device name, now that we have its id. */
4732 BUILD_BUG_ON(DEV_NAME_LEN
4733 < sizeof (RBD_DRV_NAME) + MAX_INT_FORMAT_WIDTH);
4734 sprintf(rbd_dev->name, "%s%d", RBD_DRV_NAME, rbd_dev->dev_id);
4736 /* Get our block major device number. */
4738 ret = register_blkdev(0, rbd_dev->name);
4741 rbd_dev->major = ret;
4743 /* Set up the blkdev mapping. */
4745 ret = rbd_init_disk(rbd_dev);
4747 goto err_out_blkdev;
4749 ret = rbd_dev_mapping_set(rbd_dev);
4752 set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
4754 ret = rbd_bus_add_dev(rbd_dev);
4756 goto err_out_mapping;
4758 /* Everything's ready. Announce the disk to the world. */
4760 set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
4761 add_disk(rbd_dev->disk);
4763 pr_info("%s: added with size 0x%llx\n", rbd_dev->disk->disk_name,
4764 (unsigned long long) rbd_dev->mapping.size);
4769 rbd_dev_mapping_clear(rbd_dev);
4771 rbd_free_disk(rbd_dev);
4773 unregister_blkdev(rbd_dev->major, rbd_dev->name);
4775 rbd_dev_id_put(rbd_dev);
4776 rbd_dev_mapping_clear(rbd_dev);
4781 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
4783 struct rbd_spec *spec = rbd_dev->spec;
4786 /* Record the header object name for this rbd image. */
4788 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
4790 if (rbd_dev->image_format == 1)
4791 size = strlen(spec->image_name) + sizeof (RBD_SUFFIX);
4793 size = sizeof (RBD_HEADER_PREFIX) + strlen(spec->image_id);
4795 rbd_dev->header_name = kmalloc(size, GFP_KERNEL);
4796 if (!rbd_dev->header_name)
4799 if (rbd_dev->image_format == 1)
4800 sprintf(rbd_dev->header_name, "%s%s",
4801 spec->image_name, RBD_SUFFIX);
4803 sprintf(rbd_dev->header_name, "%s%s",
4804 RBD_HEADER_PREFIX, spec->image_id);
4808 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
4810 rbd_dev_unprobe(rbd_dev);
4811 kfree(rbd_dev->header_name);
4812 rbd_dev->header_name = NULL;
4813 rbd_dev->image_format = 0;
4814 kfree(rbd_dev->spec->image_id);
4815 rbd_dev->spec->image_id = NULL;
4817 rbd_dev_destroy(rbd_dev);
4821 * Probe for the existence of the header object for the given rbd
4822 * device. If this image is the one being mapped (i.e., not a
4823 * parent), initiate a watch on its header object before using that
4824 * object to get detailed information about the rbd image.
4826 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, bool mapping)
4832 * Get the id from the image id object. If it's not a
4833 * format 2 image, we'll get ENOENT back, and we'll assume
4834 * it's a format 1 image.
4836 ret = rbd_dev_image_id(rbd_dev);
4839 rbd_assert(rbd_dev->spec->image_id);
4840 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
4842 ret = rbd_dev_header_name(rbd_dev);
4844 goto err_out_format;
4847 ret = rbd_dev_header_watch_sync(rbd_dev, true);
4849 goto out_header_name;
4852 if (rbd_dev->image_format == 1)
4853 ret = rbd_dev_v1_header_info(rbd_dev);
4855 ret = rbd_dev_v2_header_info(rbd_dev);
4859 ret = rbd_dev_spec_update(rbd_dev);
4863 ret = rbd_dev_probe_parent(rbd_dev);
4867 dout("discovered format %u image, header name is %s\n",
4868 rbd_dev->image_format, rbd_dev->header_name);
4872 rbd_dev_unprobe(rbd_dev);
4875 tmp = rbd_dev_header_watch_sync(rbd_dev, false);
4877 rbd_warn(rbd_dev, "unable to tear down "
4878 "watch request (%d)\n", tmp);
4881 kfree(rbd_dev->header_name);
4882 rbd_dev->header_name = NULL;
4884 rbd_dev->image_format = 0;
4885 kfree(rbd_dev->spec->image_id);
4886 rbd_dev->spec->image_id = NULL;
4888 dout("probe failed, returning %d\n", ret);
4893 static ssize_t rbd_add(struct bus_type *bus,
4897 struct rbd_device *rbd_dev = NULL;
4898 struct ceph_options *ceph_opts = NULL;
4899 struct rbd_options *rbd_opts = NULL;
4900 struct rbd_spec *spec = NULL;
4901 struct rbd_client *rbdc;
4902 struct ceph_osd_client *osdc;
4906 if (!try_module_get(THIS_MODULE))
4909 /* parse add command */
4910 rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
4912 goto err_out_module;
4913 read_only = rbd_opts->read_only;
4915 rbd_opts = NULL; /* done with this */
4917 rbdc = rbd_get_client(ceph_opts);
4922 ceph_opts = NULL; /* rbd_dev client now owns this */
4925 osdc = &rbdc->client->osdc;
4926 rc = ceph_pg_poolid_by_name(osdc->osdmap, spec->pool_name);
4928 goto err_out_client;
4929 spec->pool_id = (u64)rc;
4931 /* The ceph file layout needs to fit pool id in 32 bits */
4933 if (spec->pool_id > (u64)U32_MAX) {
4934 rbd_warn(NULL, "pool id too large (%llu > %u)\n",
4935 (unsigned long long)spec->pool_id, U32_MAX);
4937 goto err_out_client;
4940 rbd_dev = rbd_dev_create(rbdc, spec);
4942 goto err_out_client;
4943 rbdc = NULL; /* rbd_dev now owns this */
4944 spec = NULL; /* rbd_dev now owns this */
4946 rc = rbd_dev_image_probe(rbd_dev, true);
4948 goto err_out_rbd_dev;
4950 /* If we are mapping a snapshot it must be marked read-only */
4952 if (rbd_dev->spec->snap_id != CEPH_NOSNAP)
4954 rbd_dev->mapping.read_only = read_only;
4956 rc = rbd_dev_device_setup(rbd_dev);
4960 rbd_dev_image_release(rbd_dev);
4962 rbd_dev_destroy(rbd_dev);
4964 rbd_put_client(rbdc);
4967 ceph_destroy_options(ceph_opts);
4971 module_put(THIS_MODULE);
4973 dout("Error adding device %s\n", buf);
4978 static struct rbd_device *__rbd_get_dev(unsigned long dev_id)
4980 struct list_head *tmp;
4981 struct rbd_device *rbd_dev;
4983 spin_lock(&rbd_dev_list_lock);
4984 list_for_each(tmp, &rbd_dev_list) {
4985 rbd_dev = list_entry(tmp, struct rbd_device, node);
4986 if (rbd_dev->dev_id == dev_id) {
4987 spin_unlock(&rbd_dev_list_lock);
4991 spin_unlock(&rbd_dev_list_lock);
4995 static void rbd_dev_device_release(struct device *dev)
4997 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4999 rbd_free_disk(rbd_dev);
5000 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5001 rbd_dev_mapping_clear(rbd_dev);
5002 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5004 rbd_dev_id_put(rbd_dev);
5005 rbd_dev_mapping_clear(rbd_dev);
5008 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
5010 while (rbd_dev->parent) {
5011 struct rbd_device *first = rbd_dev;
5012 struct rbd_device *second = first->parent;
5013 struct rbd_device *third;
5016 * Follow to the parent with no grandparent and
5019 while (second && (third = second->parent)) {
5024 rbd_dev_image_release(second);
5025 first->parent = NULL;
5026 first->parent_overlap = 0;
5028 rbd_assert(first->parent_spec);
5029 rbd_spec_put(first->parent_spec);
5030 first->parent_spec = NULL;
5034 static ssize_t rbd_remove(struct bus_type *bus,
5038 struct rbd_device *rbd_dev = NULL;
5043 ret = strict_strtoul(buf, 10, &ul);
5047 /* convert to int; abort if we lost anything in the conversion */
5048 target_id = (int) ul;
5049 if (target_id != ul)
5052 mutex_lock_nested(&ctl_mutex, SINGLE_DEPTH_NESTING);
5054 rbd_dev = __rbd_get_dev(target_id);
5060 spin_lock_irq(&rbd_dev->lock);
5061 if (rbd_dev->open_count)
5064 set_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags);
5065 spin_unlock_irq(&rbd_dev->lock);
5068 rbd_bus_del_dev(rbd_dev);
5069 ret = rbd_dev_header_watch_sync(rbd_dev, false);
5071 rbd_warn(rbd_dev, "failed to cancel watch event (%d)\n", ret);
5072 rbd_dev_image_release(rbd_dev);
5073 module_put(THIS_MODULE);
5076 mutex_unlock(&ctl_mutex);
5082 * create control files in sysfs
5085 static int rbd_sysfs_init(void)
5089 ret = device_register(&rbd_root_dev);
5093 ret = bus_register(&rbd_bus_type);
5095 device_unregister(&rbd_root_dev);
5100 static void rbd_sysfs_cleanup(void)
5102 bus_unregister(&rbd_bus_type);
5103 device_unregister(&rbd_root_dev);
5106 static int rbd_slab_init(void)
5108 rbd_assert(!rbd_img_request_cache);
5109 rbd_img_request_cache = kmem_cache_create("rbd_img_request",
5110 sizeof (struct rbd_img_request),
5111 __alignof__(struct rbd_img_request),
5113 if (!rbd_img_request_cache)
5116 rbd_assert(!rbd_obj_request_cache);
5117 rbd_obj_request_cache = kmem_cache_create("rbd_obj_request",
5118 sizeof (struct rbd_obj_request),
5119 __alignof__(struct rbd_obj_request),
5121 if (!rbd_obj_request_cache)
5124 rbd_assert(!rbd_segment_name_cache);
5125 rbd_segment_name_cache = kmem_cache_create("rbd_segment_name",
5126 MAX_OBJ_NAME_SIZE + 1, 1, 0, NULL);
5127 if (rbd_segment_name_cache)
5130 if (rbd_obj_request_cache) {
5131 kmem_cache_destroy(rbd_obj_request_cache);
5132 rbd_obj_request_cache = NULL;
5135 kmem_cache_destroy(rbd_img_request_cache);
5136 rbd_img_request_cache = NULL;
5141 static void rbd_slab_exit(void)
5143 rbd_assert(rbd_segment_name_cache);
5144 kmem_cache_destroy(rbd_segment_name_cache);
5145 rbd_segment_name_cache = NULL;
5147 rbd_assert(rbd_obj_request_cache);
5148 kmem_cache_destroy(rbd_obj_request_cache);
5149 rbd_obj_request_cache = NULL;
5151 rbd_assert(rbd_img_request_cache);
5152 kmem_cache_destroy(rbd_img_request_cache);
5153 rbd_img_request_cache = NULL;
5156 static int __init rbd_init(void)
5160 if (!libceph_compatible(NULL)) {
5161 rbd_warn(NULL, "libceph incompatibility (quitting)");
5165 rc = rbd_slab_init();
5168 rc = rbd_sysfs_init();
5172 pr_info("loaded " RBD_DRV_NAME_LONG "\n");
5177 static void __exit rbd_exit(void)
5179 rbd_sysfs_cleanup();
5183 module_init(rbd_init);
5184 module_exit(rbd_exit);
5186 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
5187 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
5188 MODULE_DESCRIPTION("rados block device");
5190 /* following authorship retained from original osdblk.c */
5191 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
5193 MODULE_LICENSE("GPL");