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/cls_lock_client.h>
35 #include <linux/ceph/striper.h>
36 #include <linux/ceph/decode.h>
37 #include <linux/parser.h>
38 #include <linux/bsearch.h>
40 #include <linux/kernel.h>
41 #include <linux/device.h>
42 #include <linux/module.h>
43 #include <linux/blk-mq.h>
45 #include <linux/blkdev.h>
46 #include <linux/slab.h>
47 #include <linux/idr.h>
48 #include <linux/workqueue.h>
50 #include "rbd_types.h"
52 #define RBD_DEBUG /* Activate rbd_assert() calls */
55 * Increment the given counter and return its updated value.
56 * If the counter is already 0 it will not be incremented.
57 * If the counter is already at its maximum value returns
58 * -EINVAL without updating it.
60 static int atomic_inc_return_safe(atomic_t *v)
64 counter = (unsigned int)__atomic_add_unless(v, 1, 0);
65 if (counter <= (unsigned int)INT_MAX)
73 /* Decrement the counter. Return the resulting value, or -EINVAL */
74 static int atomic_dec_return_safe(atomic_t *v)
78 counter = atomic_dec_return(v);
87 #define RBD_DRV_NAME "rbd"
89 #define RBD_MINORS_PER_MAJOR 256
90 #define RBD_SINGLE_MAJOR_PART_SHIFT 4
92 #define RBD_MAX_PARENT_CHAIN_LEN 16
94 #define RBD_SNAP_DEV_NAME_PREFIX "snap_"
95 #define RBD_MAX_SNAP_NAME_LEN \
96 (NAME_MAX - (sizeof (RBD_SNAP_DEV_NAME_PREFIX) - 1))
98 #define RBD_MAX_SNAP_COUNT 510 /* allows max snapc to fit in 4KB */
100 #define RBD_SNAP_HEAD_NAME "-"
102 #define BAD_SNAP_INDEX U32_MAX /* invalid index into snap array */
104 /* This allows a single page to hold an image name sent by OSD */
105 #define RBD_IMAGE_NAME_LEN_MAX (PAGE_SIZE - sizeof (__le32) - 1)
106 #define RBD_IMAGE_ID_LEN_MAX 64
108 #define RBD_OBJ_PREFIX_LEN_MAX 64
110 #define RBD_NOTIFY_TIMEOUT 5 /* seconds */
111 #define RBD_RETRY_DELAY msecs_to_jiffies(1000)
115 #define RBD_FEATURE_LAYERING (1ULL<<0)
116 #define RBD_FEATURE_STRIPINGV2 (1ULL<<1)
117 #define RBD_FEATURE_EXCLUSIVE_LOCK (1ULL<<2)
118 #define RBD_FEATURE_DATA_POOL (1ULL<<7)
119 #define RBD_FEATURE_OPERATIONS (1ULL<<8)
121 #define RBD_FEATURES_ALL (RBD_FEATURE_LAYERING | \
122 RBD_FEATURE_STRIPINGV2 | \
123 RBD_FEATURE_EXCLUSIVE_LOCK | \
124 RBD_FEATURE_DATA_POOL | \
125 RBD_FEATURE_OPERATIONS)
127 /* Features supported by this (client software) implementation. */
129 #define RBD_FEATURES_SUPPORTED (RBD_FEATURES_ALL)
132 * An RBD device name will be "rbd#", where the "rbd" comes from
133 * RBD_DRV_NAME above, and # is a unique integer identifier.
135 #define DEV_NAME_LEN 32
138 * block device image metadata (in-memory version)
140 struct rbd_image_header {
141 /* These six fields never change for a given rbd image */
147 u64 features; /* Might be changeable someday? */
149 /* The remaining fields need to be updated occasionally */
151 struct ceph_snap_context *snapc;
152 char *snap_names; /* format 1 only */
153 u64 *snap_sizes; /* format 1 only */
157 * An rbd image specification.
159 * The tuple (pool_id, image_id, snap_id) is sufficient to uniquely
160 * identify an image. Each rbd_dev structure includes a pointer to
161 * an rbd_spec structure that encapsulates this identity.
163 * Each of the id's in an rbd_spec has an associated name. For a
164 * user-mapped image, the names are supplied and the id's associated
165 * with them are looked up. For a layered image, a parent image is
166 * defined by the tuple, and the names are looked up.
168 * An rbd_dev structure contains a parent_spec pointer which is
169 * non-null if the image it represents is a child in a layered
170 * image. This pointer will refer to the rbd_spec structure used
171 * by the parent rbd_dev for its own identity (i.e., the structure
172 * is shared between the parent and child).
174 * Since these structures are populated once, during the discovery
175 * phase of image construction, they are effectively immutable so
176 * we make no effort to synchronize access to them.
178 * Note that code herein does not assume the image name is known (it
179 * could be a null pointer).
183 const char *pool_name;
185 const char *image_id;
186 const char *image_name;
189 const char *snap_name;
195 * an instance of the client. multiple devices may share an rbd client.
198 struct ceph_client *client;
200 struct list_head node;
203 struct rbd_img_request;
205 enum obj_request_type {
206 OBJ_REQUEST_NODATA = 1,
207 OBJ_REQUEST_BIO, /* pointer into provided bio (list) */
208 OBJ_REQUEST_BVECS, /* pointer into provided bio_vec array */
209 OBJ_REQUEST_OWN_BVECS, /* private bio_vec array, doesn't own pages */
212 enum obj_operation_type {
219 * Writes go through the following state machine to deal with
223 * RBD_OBJ_WRITE_GUARD ---------------> RBD_OBJ_WRITE_COPYUP
225 * v \------------------------------/
231 * Writes start in RBD_OBJ_WRITE_GUARD or _FLAT, depending on whether
232 * there is a parent or not.
234 enum rbd_obj_write_state {
235 RBD_OBJ_WRITE_FLAT = 1,
237 RBD_OBJ_WRITE_COPYUP,
240 struct rbd_obj_request {
241 struct ceph_object_extent ex;
243 bool tried_parent; /* for reads */
244 enum rbd_obj_write_state write_state; /* for writes */
247 struct rbd_img_request *img_request;
248 struct ceph_file_extent *img_extents;
252 struct ceph_bio_iter bio_pos;
254 struct ceph_bvec_iter bvec_pos;
259 struct bio_vec *copyup_bvecs;
260 u32 copyup_bvec_count;
262 struct ceph_osd_request *osd_req;
264 u64 xferred; /* bytes transferred */
271 IMG_REQ_CHILD, /* initiator: block = 0, child image = 1 */
272 IMG_REQ_LAYERED, /* ENOENT handling: normal = 0, layered = 1 */
275 struct rbd_img_request {
276 struct rbd_device *rbd_dev;
277 enum obj_operation_type op_type;
278 enum obj_request_type data_type;
281 u64 snap_id; /* for reads */
282 struct ceph_snap_context *snapc; /* for writes */
285 struct request *rq; /* block request */
286 struct rbd_obj_request *obj_request; /* obj req initiator */
288 spinlock_t completion_lock;
289 u64 xferred;/* aggregate bytes transferred */
290 int result; /* first nonzero obj_request result */
292 struct list_head object_extents; /* obj_req.ex structs */
293 u32 obj_request_count;
299 #define for_each_obj_request(ireq, oreq) \
300 list_for_each_entry(oreq, &(ireq)->object_extents, ex.oe_item)
301 #define for_each_obj_request_safe(ireq, oreq, n) \
302 list_for_each_entry_safe(oreq, n, &(ireq)->object_extents, ex.oe_item)
304 enum rbd_watch_state {
305 RBD_WATCH_STATE_UNREGISTERED,
306 RBD_WATCH_STATE_REGISTERED,
307 RBD_WATCH_STATE_ERROR,
310 enum rbd_lock_state {
311 RBD_LOCK_STATE_UNLOCKED,
312 RBD_LOCK_STATE_LOCKED,
313 RBD_LOCK_STATE_RELEASING,
316 /* WatchNotify::ClientId */
317 struct rbd_client_id {
331 int dev_id; /* blkdev unique id */
333 int major; /* blkdev assigned major */
335 struct gendisk *disk; /* blkdev's gendisk and rq */
337 u32 image_format; /* Either 1 or 2 */
338 struct rbd_client *rbd_client;
340 char name[DEV_NAME_LEN]; /* blkdev name, e.g. rbd3 */
342 spinlock_t lock; /* queue, flags, open_count */
344 struct rbd_image_header header;
345 unsigned long flags; /* possibly lock protected */
346 struct rbd_spec *spec;
347 struct rbd_options *opts;
348 char *config_info; /* add{,_single_major} string */
350 struct ceph_object_id header_oid;
351 struct ceph_object_locator header_oloc;
353 struct ceph_file_layout layout; /* used for all rbd requests */
355 struct mutex watch_mutex;
356 enum rbd_watch_state watch_state;
357 struct ceph_osd_linger_request *watch_handle;
359 struct delayed_work watch_dwork;
361 struct rw_semaphore lock_rwsem;
362 enum rbd_lock_state lock_state;
363 char lock_cookie[32];
364 struct rbd_client_id owner_cid;
365 struct work_struct acquired_lock_work;
366 struct work_struct released_lock_work;
367 struct delayed_work lock_dwork;
368 struct work_struct unlock_work;
369 wait_queue_head_t lock_waitq;
371 struct workqueue_struct *task_wq;
373 struct rbd_spec *parent_spec;
376 struct rbd_device *parent;
378 /* Block layer tags. */
379 struct blk_mq_tag_set tag_set;
381 /* protects updating the header */
382 struct rw_semaphore header_rwsem;
384 struct rbd_mapping mapping;
386 struct list_head node;
390 unsigned long open_count; /* protected by lock */
394 * Flag bits for rbd_dev->flags:
395 * - REMOVING (which is coupled with rbd_dev->open_count) is protected
397 * - BLACKLISTED is protected by rbd_dev->lock_rwsem
400 RBD_DEV_FLAG_EXISTS, /* mapped snapshot has not been deleted */
401 RBD_DEV_FLAG_REMOVING, /* this mapping is being removed */
402 RBD_DEV_FLAG_BLACKLISTED, /* our ceph_client is blacklisted */
405 static DEFINE_MUTEX(client_mutex); /* Serialize client creation */
407 static LIST_HEAD(rbd_dev_list); /* devices */
408 static DEFINE_SPINLOCK(rbd_dev_list_lock);
410 static LIST_HEAD(rbd_client_list); /* clients */
411 static DEFINE_SPINLOCK(rbd_client_list_lock);
413 /* Slab caches for frequently-allocated structures */
415 static struct kmem_cache *rbd_img_request_cache;
416 static struct kmem_cache *rbd_obj_request_cache;
418 static int rbd_major;
419 static DEFINE_IDA(rbd_dev_id_ida);
421 static struct workqueue_struct *rbd_wq;
424 * single-major requires >= 0.75 version of userspace rbd utility.
426 static bool single_major = true;
427 module_param(single_major, bool, 0444);
428 MODULE_PARM_DESC(single_major, "Use a single major number for all rbd devices (default: true)");
430 static ssize_t rbd_add(struct bus_type *bus, const char *buf,
432 static ssize_t rbd_remove(struct bus_type *bus, const char *buf,
434 static ssize_t rbd_add_single_major(struct bus_type *bus, const char *buf,
436 static ssize_t rbd_remove_single_major(struct bus_type *bus, const char *buf,
438 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth);
440 static int rbd_dev_id_to_minor(int dev_id)
442 return dev_id << RBD_SINGLE_MAJOR_PART_SHIFT;
445 static int minor_to_rbd_dev_id(int minor)
447 return minor >> RBD_SINGLE_MAJOR_PART_SHIFT;
450 static bool __rbd_is_lock_owner(struct rbd_device *rbd_dev)
452 return rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED ||
453 rbd_dev->lock_state == RBD_LOCK_STATE_RELEASING;
456 static bool rbd_is_lock_owner(struct rbd_device *rbd_dev)
460 down_read(&rbd_dev->lock_rwsem);
461 is_lock_owner = __rbd_is_lock_owner(rbd_dev);
462 up_read(&rbd_dev->lock_rwsem);
463 return is_lock_owner;
466 static ssize_t rbd_supported_features_show(struct bus_type *bus, char *buf)
468 return sprintf(buf, "0x%llx\n", RBD_FEATURES_SUPPORTED);
471 static BUS_ATTR(add, 0200, NULL, rbd_add);
472 static BUS_ATTR(remove, 0200, NULL, rbd_remove);
473 static BUS_ATTR(add_single_major, 0200, NULL, rbd_add_single_major);
474 static BUS_ATTR(remove_single_major, 0200, NULL, rbd_remove_single_major);
475 static BUS_ATTR(supported_features, 0444, rbd_supported_features_show, NULL);
477 static struct attribute *rbd_bus_attrs[] = {
479 &bus_attr_remove.attr,
480 &bus_attr_add_single_major.attr,
481 &bus_attr_remove_single_major.attr,
482 &bus_attr_supported_features.attr,
486 static umode_t rbd_bus_is_visible(struct kobject *kobj,
487 struct attribute *attr, int index)
490 (attr == &bus_attr_add_single_major.attr ||
491 attr == &bus_attr_remove_single_major.attr))
497 static const struct attribute_group rbd_bus_group = {
498 .attrs = rbd_bus_attrs,
499 .is_visible = rbd_bus_is_visible,
501 __ATTRIBUTE_GROUPS(rbd_bus);
503 static struct bus_type rbd_bus_type = {
505 .bus_groups = rbd_bus_groups,
508 static void rbd_root_dev_release(struct device *dev)
512 static struct device rbd_root_dev = {
514 .release = rbd_root_dev_release,
517 static __printf(2, 3)
518 void rbd_warn(struct rbd_device *rbd_dev, const char *fmt, ...)
520 struct va_format vaf;
528 printk(KERN_WARNING "%s: %pV\n", RBD_DRV_NAME, &vaf);
529 else if (rbd_dev->disk)
530 printk(KERN_WARNING "%s: %s: %pV\n",
531 RBD_DRV_NAME, rbd_dev->disk->disk_name, &vaf);
532 else if (rbd_dev->spec && rbd_dev->spec->image_name)
533 printk(KERN_WARNING "%s: image %s: %pV\n",
534 RBD_DRV_NAME, rbd_dev->spec->image_name, &vaf);
535 else if (rbd_dev->spec && rbd_dev->spec->image_id)
536 printk(KERN_WARNING "%s: id %s: %pV\n",
537 RBD_DRV_NAME, rbd_dev->spec->image_id, &vaf);
539 printk(KERN_WARNING "%s: rbd_dev %p: %pV\n",
540 RBD_DRV_NAME, rbd_dev, &vaf);
545 #define rbd_assert(expr) \
546 if (unlikely(!(expr))) { \
547 printk(KERN_ERR "\nAssertion failure in %s() " \
549 "\trbd_assert(%s);\n\n", \
550 __func__, __LINE__, #expr); \
553 #else /* !RBD_DEBUG */
554 # define rbd_assert(expr) ((void) 0)
555 #endif /* !RBD_DEBUG */
557 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev);
559 static int rbd_dev_refresh(struct rbd_device *rbd_dev);
560 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev);
561 static int rbd_dev_header_info(struct rbd_device *rbd_dev);
562 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev);
563 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
565 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
566 u8 *order, u64 *snap_size);
567 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
570 static int rbd_open(struct block_device *bdev, fmode_t mode)
572 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
573 bool removing = false;
575 spin_lock_irq(&rbd_dev->lock);
576 if (test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags))
579 rbd_dev->open_count++;
580 spin_unlock_irq(&rbd_dev->lock);
584 (void) get_device(&rbd_dev->dev);
589 static void rbd_release(struct gendisk *disk, fmode_t mode)
591 struct rbd_device *rbd_dev = disk->private_data;
592 unsigned long open_count_before;
594 spin_lock_irq(&rbd_dev->lock);
595 open_count_before = rbd_dev->open_count--;
596 spin_unlock_irq(&rbd_dev->lock);
597 rbd_assert(open_count_before > 0);
599 put_device(&rbd_dev->dev);
602 static int rbd_ioctl_set_ro(struct rbd_device *rbd_dev, unsigned long arg)
606 if (get_user(ro, (int __user *)arg))
609 /* Snapshots can't be marked read-write */
610 if (rbd_dev->spec->snap_id != CEPH_NOSNAP && !ro)
613 /* Let blkdev_roset() handle it */
617 static int rbd_ioctl(struct block_device *bdev, fmode_t mode,
618 unsigned int cmd, unsigned long arg)
620 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
625 ret = rbd_ioctl_set_ro(rbd_dev, arg);
635 static int rbd_compat_ioctl(struct block_device *bdev, fmode_t mode,
636 unsigned int cmd, unsigned long arg)
638 return rbd_ioctl(bdev, mode, cmd, arg);
640 #endif /* CONFIG_COMPAT */
642 static const struct block_device_operations rbd_bd_ops = {
643 .owner = THIS_MODULE,
645 .release = rbd_release,
648 .compat_ioctl = rbd_compat_ioctl,
653 * Initialize an rbd client instance. Success or not, this function
654 * consumes ceph_opts. Caller holds client_mutex.
656 static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts)
658 struct rbd_client *rbdc;
661 dout("%s:\n", __func__);
662 rbdc = kmalloc(sizeof(struct rbd_client), GFP_KERNEL);
666 kref_init(&rbdc->kref);
667 INIT_LIST_HEAD(&rbdc->node);
669 rbdc->client = ceph_create_client(ceph_opts, rbdc);
670 if (IS_ERR(rbdc->client))
672 ceph_opts = NULL; /* Now rbdc->client is responsible for ceph_opts */
674 ret = ceph_open_session(rbdc->client);
678 spin_lock(&rbd_client_list_lock);
679 list_add_tail(&rbdc->node, &rbd_client_list);
680 spin_unlock(&rbd_client_list_lock);
682 dout("%s: rbdc %p\n", __func__, rbdc);
686 ceph_destroy_client(rbdc->client);
691 ceph_destroy_options(ceph_opts);
692 dout("%s: error %d\n", __func__, ret);
697 static struct rbd_client *__rbd_get_client(struct rbd_client *rbdc)
699 kref_get(&rbdc->kref);
705 * Find a ceph client with specific addr and configuration. If
706 * found, bump its reference count.
708 static struct rbd_client *rbd_client_find(struct ceph_options *ceph_opts)
710 struct rbd_client *client_node;
713 if (ceph_opts->flags & CEPH_OPT_NOSHARE)
716 spin_lock(&rbd_client_list_lock);
717 list_for_each_entry(client_node, &rbd_client_list, node) {
718 if (!ceph_compare_options(ceph_opts, client_node->client)) {
719 __rbd_get_client(client_node);
725 spin_unlock(&rbd_client_list_lock);
727 return found ? client_node : NULL;
731 * (Per device) rbd map options
739 /* string args above */
748 static match_table_t rbd_opts_tokens = {
749 {Opt_queue_depth, "queue_depth=%d"},
750 {Opt_lock_timeout, "lock_timeout=%d"},
752 /* string args above */
753 {Opt_read_only, "read_only"},
754 {Opt_read_only, "ro"}, /* Alternate spelling */
755 {Opt_read_write, "read_write"},
756 {Opt_read_write, "rw"}, /* Alternate spelling */
757 {Opt_lock_on_read, "lock_on_read"},
758 {Opt_exclusive, "exclusive"},
759 {Opt_notrim, "notrim"},
765 unsigned long lock_timeout;
772 #define RBD_QUEUE_DEPTH_DEFAULT BLKDEV_MAX_RQ
773 #define RBD_LOCK_TIMEOUT_DEFAULT 0 /* no timeout */
774 #define RBD_READ_ONLY_DEFAULT false
775 #define RBD_LOCK_ON_READ_DEFAULT false
776 #define RBD_EXCLUSIVE_DEFAULT false
777 #define RBD_TRIM_DEFAULT true
779 static int parse_rbd_opts_token(char *c, void *private)
781 struct rbd_options *rbd_opts = private;
782 substring_t argstr[MAX_OPT_ARGS];
783 int token, intval, ret;
785 token = match_token(c, rbd_opts_tokens, argstr);
786 if (token < Opt_last_int) {
787 ret = match_int(&argstr[0], &intval);
789 pr_err("bad mount option arg (not int) at '%s'\n", c);
792 dout("got int token %d val %d\n", token, intval);
793 } else if (token > Opt_last_int && token < Opt_last_string) {
794 dout("got string token %d val %s\n", token, argstr[0].from);
796 dout("got token %d\n", token);
800 case Opt_queue_depth:
802 pr_err("queue_depth out of range\n");
805 rbd_opts->queue_depth = intval;
807 case Opt_lock_timeout:
808 /* 0 is "wait forever" (i.e. infinite timeout) */
809 if (intval < 0 || intval > INT_MAX / 1000) {
810 pr_err("lock_timeout out of range\n");
813 rbd_opts->lock_timeout = msecs_to_jiffies(intval * 1000);
816 rbd_opts->read_only = true;
819 rbd_opts->read_only = false;
821 case Opt_lock_on_read:
822 rbd_opts->lock_on_read = true;
825 rbd_opts->exclusive = true;
828 rbd_opts->trim = false;
831 /* libceph prints "bad option" msg */
838 static char* obj_op_name(enum obj_operation_type op_type)
853 * Destroy ceph client
855 * Caller must hold rbd_client_list_lock.
857 static void rbd_client_release(struct kref *kref)
859 struct rbd_client *rbdc = container_of(kref, struct rbd_client, kref);
861 dout("%s: rbdc %p\n", __func__, rbdc);
862 spin_lock(&rbd_client_list_lock);
863 list_del(&rbdc->node);
864 spin_unlock(&rbd_client_list_lock);
866 ceph_destroy_client(rbdc->client);
871 * Drop reference to ceph client node. If it's not referenced anymore, release
874 static void rbd_put_client(struct rbd_client *rbdc)
877 kref_put(&rbdc->kref, rbd_client_release);
880 static int wait_for_latest_osdmap(struct ceph_client *client)
885 ret = ceph_monc_get_version(&client->monc, "osdmap", &newest_epoch);
889 if (client->osdc.osdmap->epoch >= newest_epoch)
892 ceph_osdc_maybe_request_map(&client->osdc);
893 return ceph_monc_wait_osdmap(&client->monc, newest_epoch,
894 client->options->mount_timeout);
898 * Get a ceph client with specific addr and configuration, if one does
899 * not exist create it. Either way, ceph_opts is consumed by this
902 static struct rbd_client *rbd_get_client(struct ceph_options *ceph_opts)
904 struct rbd_client *rbdc;
907 mutex_lock_nested(&client_mutex, SINGLE_DEPTH_NESTING);
908 rbdc = rbd_client_find(ceph_opts);
910 ceph_destroy_options(ceph_opts);
913 * Using an existing client. Make sure ->pg_pools is up to
914 * date before we look up the pool id in do_rbd_add().
916 ret = wait_for_latest_osdmap(rbdc->client);
918 rbd_warn(NULL, "failed to get latest osdmap: %d", ret);
919 rbd_put_client(rbdc);
923 rbdc = rbd_client_create(ceph_opts);
925 mutex_unlock(&client_mutex);
930 static bool rbd_image_format_valid(u32 image_format)
932 return image_format == 1 || image_format == 2;
935 static bool rbd_dev_ondisk_valid(struct rbd_image_header_ondisk *ondisk)
940 /* The header has to start with the magic rbd header text */
941 if (memcmp(&ondisk->text, RBD_HEADER_TEXT, sizeof (RBD_HEADER_TEXT)))
944 /* The bio layer requires at least sector-sized I/O */
946 if (ondisk->options.order < SECTOR_SHIFT)
949 /* If we use u64 in a few spots we may be able to loosen this */
951 if (ondisk->options.order > 8 * sizeof (int) - 1)
955 * The size of a snapshot header has to fit in a size_t, and
956 * that limits the number of snapshots.
958 snap_count = le32_to_cpu(ondisk->snap_count);
959 size = SIZE_MAX - sizeof (struct ceph_snap_context);
960 if (snap_count > size / sizeof (__le64))
964 * Not only that, but the size of the entire the snapshot
965 * header must also be representable in a size_t.
967 size -= snap_count * sizeof (__le64);
968 if ((u64) size < le64_to_cpu(ondisk->snap_names_len))
975 * returns the size of an object in the image
977 static u32 rbd_obj_bytes(struct rbd_image_header *header)
979 return 1U << header->obj_order;
982 static void rbd_init_layout(struct rbd_device *rbd_dev)
984 if (rbd_dev->header.stripe_unit == 0 ||
985 rbd_dev->header.stripe_count == 0) {
986 rbd_dev->header.stripe_unit = rbd_obj_bytes(&rbd_dev->header);
987 rbd_dev->header.stripe_count = 1;
990 rbd_dev->layout.stripe_unit = rbd_dev->header.stripe_unit;
991 rbd_dev->layout.stripe_count = rbd_dev->header.stripe_count;
992 rbd_dev->layout.object_size = rbd_obj_bytes(&rbd_dev->header);
993 rbd_dev->layout.pool_id = rbd_dev->header.data_pool_id == CEPH_NOPOOL ?
994 rbd_dev->spec->pool_id : rbd_dev->header.data_pool_id;
995 RCU_INIT_POINTER(rbd_dev->layout.pool_ns, NULL);
999 * Fill an rbd image header with information from the given format 1
1002 static int rbd_header_from_disk(struct rbd_device *rbd_dev,
1003 struct rbd_image_header_ondisk *ondisk)
1005 struct rbd_image_header *header = &rbd_dev->header;
1006 bool first_time = header->object_prefix == NULL;
1007 struct ceph_snap_context *snapc;
1008 char *object_prefix = NULL;
1009 char *snap_names = NULL;
1010 u64 *snap_sizes = NULL;
1015 /* Allocate this now to avoid having to handle failure below */
1018 object_prefix = kstrndup(ondisk->object_prefix,
1019 sizeof(ondisk->object_prefix),
1025 /* Allocate the snapshot context and fill it in */
1027 snap_count = le32_to_cpu(ondisk->snap_count);
1028 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
1031 snapc->seq = le64_to_cpu(ondisk->snap_seq);
1033 struct rbd_image_snap_ondisk *snaps;
1034 u64 snap_names_len = le64_to_cpu(ondisk->snap_names_len);
1036 /* We'll keep a copy of the snapshot names... */
1038 if (snap_names_len > (u64)SIZE_MAX)
1040 snap_names = kmalloc(snap_names_len, GFP_KERNEL);
1044 /* ...as well as the array of their sizes. */
1045 snap_sizes = kmalloc_array(snap_count,
1046 sizeof(*header->snap_sizes),
1052 * Copy the names, and fill in each snapshot's id
1055 * Note that rbd_dev_v1_header_info() guarantees the
1056 * ondisk buffer we're working with has
1057 * snap_names_len bytes beyond the end of the
1058 * snapshot id array, this memcpy() is safe.
1060 memcpy(snap_names, &ondisk->snaps[snap_count], snap_names_len);
1061 snaps = ondisk->snaps;
1062 for (i = 0; i < snap_count; i++) {
1063 snapc->snaps[i] = le64_to_cpu(snaps[i].id);
1064 snap_sizes[i] = le64_to_cpu(snaps[i].image_size);
1068 /* We won't fail any more, fill in the header */
1071 header->object_prefix = object_prefix;
1072 header->obj_order = ondisk->options.order;
1073 rbd_init_layout(rbd_dev);
1075 ceph_put_snap_context(header->snapc);
1076 kfree(header->snap_names);
1077 kfree(header->snap_sizes);
1080 /* The remaining fields always get updated (when we refresh) */
1082 header->image_size = le64_to_cpu(ondisk->image_size);
1083 header->snapc = snapc;
1084 header->snap_names = snap_names;
1085 header->snap_sizes = snap_sizes;
1093 ceph_put_snap_context(snapc);
1094 kfree(object_prefix);
1099 static const char *_rbd_dev_v1_snap_name(struct rbd_device *rbd_dev, u32 which)
1101 const char *snap_name;
1103 rbd_assert(which < rbd_dev->header.snapc->num_snaps);
1105 /* Skip over names until we find the one we are looking for */
1107 snap_name = rbd_dev->header.snap_names;
1109 snap_name += strlen(snap_name) + 1;
1111 return kstrdup(snap_name, GFP_KERNEL);
1115 * Snapshot id comparison function for use with qsort()/bsearch().
1116 * Note that result is for snapshots in *descending* order.
1118 static int snapid_compare_reverse(const void *s1, const void *s2)
1120 u64 snap_id1 = *(u64 *)s1;
1121 u64 snap_id2 = *(u64 *)s2;
1123 if (snap_id1 < snap_id2)
1125 return snap_id1 == snap_id2 ? 0 : -1;
1129 * Search a snapshot context to see if the given snapshot id is
1132 * Returns the position of the snapshot id in the array if it's found,
1133 * or BAD_SNAP_INDEX otherwise.
1135 * Note: The snapshot array is in kept sorted (by the osd) in
1136 * reverse order, highest snapshot id first.
1138 static u32 rbd_dev_snap_index(struct rbd_device *rbd_dev, u64 snap_id)
1140 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
1143 found = bsearch(&snap_id, &snapc->snaps, snapc->num_snaps,
1144 sizeof (snap_id), snapid_compare_reverse);
1146 return found ? (u32)(found - &snapc->snaps[0]) : BAD_SNAP_INDEX;
1149 static const char *rbd_dev_v1_snap_name(struct rbd_device *rbd_dev,
1153 const char *snap_name;
1155 which = rbd_dev_snap_index(rbd_dev, snap_id);
1156 if (which == BAD_SNAP_INDEX)
1157 return ERR_PTR(-ENOENT);
1159 snap_name = _rbd_dev_v1_snap_name(rbd_dev, which);
1160 return snap_name ? snap_name : ERR_PTR(-ENOMEM);
1163 static const char *rbd_snap_name(struct rbd_device *rbd_dev, u64 snap_id)
1165 if (snap_id == CEPH_NOSNAP)
1166 return RBD_SNAP_HEAD_NAME;
1168 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1169 if (rbd_dev->image_format == 1)
1170 return rbd_dev_v1_snap_name(rbd_dev, snap_id);
1172 return rbd_dev_v2_snap_name(rbd_dev, snap_id);
1175 static int rbd_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
1178 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1179 if (snap_id == CEPH_NOSNAP) {
1180 *snap_size = rbd_dev->header.image_size;
1181 } else if (rbd_dev->image_format == 1) {
1184 which = rbd_dev_snap_index(rbd_dev, snap_id);
1185 if (which == BAD_SNAP_INDEX)
1188 *snap_size = rbd_dev->header.snap_sizes[which];
1193 ret = _rbd_dev_v2_snap_size(rbd_dev, snap_id, NULL, &size);
1202 static int rbd_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
1205 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1206 if (snap_id == CEPH_NOSNAP) {
1207 *snap_features = rbd_dev->header.features;
1208 } else if (rbd_dev->image_format == 1) {
1209 *snap_features = 0; /* No features for format 1 */
1214 ret = _rbd_dev_v2_snap_features(rbd_dev, snap_id, &features);
1218 *snap_features = features;
1223 static int rbd_dev_mapping_set(struct rbd_device *rbd_dev)
1225 u64 snap_id = rbd_dev->spec->snap_id;
1230 ret = rbd_snap_size(rbd_dev, snap_id, &size);
1233 ret = rbd_snap_features(rbd_dev, snap_id, &features);
1237 rbd_dev->mapping.size = size;
1238 rbd_dev->mapping.features = features;
1243 static void rbd_dev_mapping_clear(struct rbd_device *rbd_dev)
1245 rbd_dev->mapping.size = 0;
1246 rbd_dev->mapping.features = 0;
1249 static void zero_bvec(struct bio_vec *bv)
1252 unsigned long flags;
1254 buf = bvec_kmap_irq(bv, &flags);
1255 memset(buf, 0, bv->bv_len);
1256 flush_dcache_page(bv->bv_page);
1257 bvec_kunmap_irq(buf, &flags);
1260 static void zero_bios(struct ceph_bio_iter *bio_pos, u32 off, u32 bytes)
1262 struct ceph_bio_iter it = *bio_pos;
1264 ceph_bio_iter_advance(&it, off);
1265 ceph_bio_iter_advance_step(&it, bytes, ({
1270 static void zero_bvecs(struct ceph_bvec_iter *bvec_pos, u32 off, u32 bytes)
1272 struct ceph_bvec_iter it = *bvec_pos;
1274 ceph_bvec_iter_advance(&it, off);
1275 ceph_bvec_iter_advance_step(&it, bytes, ({
1281 * Zero a range in @obj_req data buffer defined by a bio (list) or
1282 * (private) bio_vec array.
1284 * @off is relative to the start of the data buffer.
1286 static void rbd_obj_zero_range(struct rbd_obj_request *obj_req, u32 off,
1289 switch (obj_req->img_request->data_type) {
1290 case OBJ_REQUEST_BIO:
1291 zero_bios(&obj_req->bio_pos, off, bytes);
1293 case OBJ_REQUEST_BVECS:
1294 case OBJ_REQUEST_OWN_BVECS:
1295 zero_bvecs(&obj_req->bvec_pos, off, bytes);
1302 static void rbd_obj_request_destroy(struct kref *kref);
1303 static void rbd_obj_request_put(struct rbd_obj_request *obj_request)
1305 rbd_assert(obj_request != NULL);
1306 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1307 kref_read(&obj_request->kref));
1308 kref_put(&obj_request->kref, rbd_obj_request_destroy);
1311 static void rbd_img_request_get(struct rbd_img_request *img_request)
1313 dout("%s: img %p (was %d)\n", __func__, img_request,
1314 kref_read(&img_request->kref));
1315 kref_get(&img_request->kref);
1318 static void rbd_img_request_destroy(struct kref *kref);
1319 static void rbd_img_request_put(struct rbd_img_request *img_request)
1321 rbd_assert(img_request != NULL);
1322 dout("%s: img %p (was %d)\n", __func__, img_request,
1323 kref_read(&img_request->kref));
1324 kref_put(&img_request->kref, rbd_img_request_destroy);
1327 static inline void rbd_img_obj_request_add(struct rbd_img_request *img_request,
1328 struct rbd_obj_request *obj_request)
1330 rbd_assert(obj_request->img_request == NULL);
1332 /* Image request now owns object's original reference */
1333 obj_request->img_request = img_request;
1334 img_request->obj_request_count++;
1335 img_request->pending_count++;
1336 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
1339 static inline void rbd_img_obj_request_del(struct rbd_img_request *img_request,
1340 struct rbd_obj_request *obj_request)
1342 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
1343 list_del(&obj_request->ex.oe_item);
1344 rbd_assert(img_request->obj_request_count > 0);
1345 img_request->obj_request_count--;
1346 rbd_assert(obj_request->img_request == img_request);
1347 rbd_obj_request_put(obj_request);
1350 static void rbd_obj_request_submit(struct rbd_obj_request *obj_request)
1352 struct ceph_osd_request *osd_req = obj_request->osd_req;
1354 dout("%s %p object_no %016llx %llu~%llu osd_req %p\n", __func__,
1355 obj_request, obj_request->ex.oe_objno, obj_request->ex.oe_off,
1356 obj_request->ex.oe_len, osd_req);
1357 ceph_osdc_start_request(osd_req->r_osdc, osd_req, false);
1361 * The default/initial value for all image request flags is 0. Each
1362 * is conditionally set to 1 at image request initialization time
1363 * and currently never change thereafter.
1365 static void img_request_layered_set(struct rbd_img_request *img_request)
1367 set_bit(IMG_REQ_LAYERED, &img_request->flags);
1371 static void img_request_layered_clear(struct rbd_img_request *img_request)
1373 clear_bit(IMG_REQ_LAYERED, &img_request->flags);
1377 static bool img_request_layered_test(struct rbd_img_request *img_request)
1380 return test_bit(IMG_REQ_LAYERED, &img_request->flags) != 0;
1383 static bool rbd_obj_is_entire(struct rbd_obj_request *obj_req)
1385 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1387 return !obj_req->ex.oe_off &&
1388 obj_req->ex.oe_len == rbd_dev->layout.object_size;
1391 static bool rbd_obj_is_tail(struct rbd_obj_request *obj_req)
1393 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1395 return obj_req->ex.oe_off + obj_req->ex.oe_len ==
1396 rbd_dev->layout.object_size;
1399 static u64 rbd_obj_img_extents_bytes(struct rbd_obj_request *obj_req)
1401 return ceph_file_extents_bytes(obj_req->img_extents,
1402 obj_req->num_img_extents);
1405 static bool rbd_img_is_write(struct rbd_img_request *img_req)
1407 switch (img_req->op_type) {
1411 case OBJ_OP_DISCARD:
1418 static void rbd_obj_handle_request(struct rbd_obj_request *obj_req);
1420 static void rbd_osd_req_callback(struct ceph_osd_request *osd_req)
1422 struct rbd_obj_request *obj_req = osd_req->r_priv;
1424 dout("%s osd_req %p result %d for obj_req %p\n", __func__, osd_req,
1425 osd_req->r_result, obj_req);
1426 rbd_assert(osd_req == obj_req->osd_req);
1428 obj_req->result = osd_req->r_result < 0 ? osd_req->r_result : 0;
1429 if (!obj_req->result && !rbd_img_is_write(obj_req->img_request))
1430 obj_req->xferred = osd_req->r_result;
1433 * Writes aren't allowed to return a data payload. In some
1434 * guarded write cases (e.g. stat + zero on an empty object)
1435 * a stat response makes it through, but we don't care.
1437 obj_req->xferred = 0;
1439 rbd_obj_handle_request(obj_req);
1442 static void rbd_osd_req_format_read(struct rbd_obj_request *obj_request)
1444 struct ceph_osd_request *osd_req = obj_request->osd_req;
1446 osd_req->r_flags = CEPH_OSD_FLAG_READ;
1447 osd_req->r_snapid = obj_request->img_request->snap_id;
1450 static void rbd_osd_req_format_write(struct rbd_obj_request *obj_request)
1452 struct ceph_osd_request *osd_req = obj_request->osd_req;
1454 osd_req->r_flags = CEPH_OSD_FLAG_WRITE;
1455 ktime_get_real_ts(&osd_req->r_mtime);
1456 osd_req->r_data_offset = obj_request->ex.oe_off;
1459 static struct ceph_osd_request *
1460 rbd_osd_req_create(struct rbd_obj_request *obj_req, unsigned int num_ops)
1462 struct rbd_img_request *img_req = obj_req->img_request;
1463 struct rbd_device *rbd_dev = img_req->rbd_dev;
1464 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
1465 struct ceph_osd_request *req;
1466 const char *name_format = rbd_dev->image_format == 1 ?
1467 RBD_V1_DATA_FORMAT : RBD_V2_DATA_FORMAT;
1469 req = ceph_osdc_alloc_request(osdc,
1470 (rbd_img_is_write(img_req) ? img_req->snapc : NULL),
1471 num_ops, false, GFP_NOIO);
1475 req->r_callback = rbd_osd_req_callback;
1476 req->r_priv = obj_req;
1478 req->r_base_oloc.pool = rbd_dev->layout.pool_id;
1479 if (ceph_oid_aprintf(&req->r_base_oid, GFP_NOIO, name_format,
1480 rbd_dev->header.object_prefix, obj_req->ex.oe_objno))
1483 if (ceph_osdc_alloc_messages(req, GFP_NOIO))
1489 ceph_osdc_put_request(req);
1493 static void rbd_osd_req_destroy(struct ceph_osd_request *osd_req)
1495 ceph_osdc_put_request(osd_req);
1498 static struct rbd_obj_request *rbd_obj_request_create(void)
1500 struct rbd_obj_request *obj_request;
1502 obj_request = kmem_cache_zalloc(rbd_obj_request_cache, GFP_NOIO);
1506 ceph_object_extent_init(&obj_request->ex);
1507 kref_init(&obj_request->kref);
1509 dout("%s %p\n", __func__, obj_request);
1513 static void rbd_obj_request_destroy(struct kref *kref)
1515 struct rbd_obj_request *obj_request;
1518 obj_request = container_of(kref, struct rbd_obj_request, kref);
1520 dout("%s: obj %p\n", __func__, obj_request);
1522 if (obj_request->osd_req)
1523 rbd_osd_req_destroy(obj_request->osd_req);
1525 switch (obj_request->img_request->data_type) {
1526 case OBJ_REQUEST_NODATA:
1527 case OBJ_REQUEST_BIO:
1528 case OBJ_REQUEST_BVECS:
1529 break; /* Nothing to do */
1530 case OBJ_REQUEST_OWN_BVECS:
1531 kfree(obj_request->bvec_pos.bvecs);
1537 kfree(obj_request->img_extents);
1538 if (obj_request->copyup_bvecs) {
1539 for (i = 0; i < obj_request->copyup_bvec_count; i++) {
1540 if (obj_request->copyup_bvecs[i].bv_page)
1541 __free_page(obj_request->copyup_bvecs[i].bv_page);
1543 kfree(obj_request->copyup_bvecs);
1546 kmem_cache_free(rbd_obj_request_cache, obj_request);
1549 /* It's OK to call this for a device with no parent */
1551 static void rbd_spec_put(struct rbd_spec *spec);
1552 static void rbd_dev_unparent(struct rbd_device *rbd_dev)
1554 rbd_dev_remove_parent(rbd_dev);
1555 rbd_spec_put(rbd_dev->parent_spec);
1556 rbd_dev->parent_spec = NULL;
1557 rbd_dev->parent_overlap = 0;
1561 * Parent image reference counting is used to determine when an
1562 * image's parent fields can be safely torn down--after there are no
1563 * more in-flight requests to the parent image. When the last
1564 * reference is dropped, cleaning them up is safe.
1566 static void rbd_dev_parent_put(struct rbd_device *rbd_dev)
1570 if (!rbd_dev->parent_spec)
1573 counter = atomic_dec_return_safe(&rbd_dev->parent_ref);
1577 /* Last reference; clean up parent data structures */
1580 rbd_dev_unparent(rbd_dev);
1582 rbd_warn(rbd_dev, "parent reference underflow");
1586 * If an image has a non-zero parent overlap, get a reference to its
1589 * Returns true if the rbd device has a parent with a non-zero
1590 * overlap and a reference for it was successfully taken, or
1593 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
1597 if (!rbd_dev->parent_spec)
1600 down_read(&rbd_dev->header_rwsem);
1601 if (rbd_dev->parent_overlap)
1602 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
1603 up_read(&rbd_dev->header_rwsem);
1606 rbd_warn(rbd_dev, "parent reference overflow");
1612 * Caller is responsible for filling in the list of object requests
1613 * that comprises the image request, and the Linux request pointer
1614 * (if there is one).
1616 static struct rbd_img_request *rbd_img_request_create(
1617 struct rbd_device *rbd_dev,
1618 enum obj_operation_type op_type,
1619 struct ceph_snap_context *snapc)
1621 struct rbd_img_request *img_request;
1623 img_request = kmem_cache_zalloc(rbd_img_request_cache, GFP_NOIO);
1627 img_request->rbd_dev = rbd_dev;
1628 img_request->op_type = op_type;
1629 if (!rbd_img_is_write(img_request))
1630 img_request->snap_id = rbd_dev->spec->snap_id;
1632 img_request->snapc = snapc;
1634 if (rbd_dev_parent_get(rbd_dev))
1635 img_request_layered_set(img_request);
1637 spin_lock_init(&img_request->completion_lock);
1638 INIT_LIST_HEAD(&img_request->object_extents);
1639 kref_init(&img_request->kref);
1641 dout("%s: rbd_dev %p %s -> img %p\n", __func__, rbd_dev,
1642 obj_op_name(op_type), img_request);
1646 static void rbd_img_request_destroy(struct kref *kref)
1648 struct rbd_img_request *img_request;
1649 struct rbd_obj_request *obj_request;
1650 struct rbd_obj_request *next_obj_request;
1652 img_request = container_of(kref, struct rbd_img_request, kref);
1654 dout("%s: img %p\n", __func__, img_request);
1656 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
1657 rbd_img_obj_request_del(img_request, obj_request);
1658 rbd_assert(img_request->obj_request_count == 0);
1660 if (img_request_layered_test(img_request)) {
1661 img_request_layered_clear(img_request);
1662 rbd_dev_parent_put(img_request->rbd_dev);
1665 if (rbd_img_is_write(img_request))
1666 ceph_put_snap_context(img_request->snapc);
1668 kmem_cache_free(rbd_img_request_cache, img_request);
1671 static void prune_extents(struct ceph_file_extent *img_extents,
1672 u32 *num_img_extents, u64 overlap)
1674 u32 cnt = *num_img_extents;
1676 /* drop extents completely beyond the overlap */
1677 while (cnt && img_extents[cnt - 1].fe_off >= overlap)
1681 struct ceph_file_extent *ex = &img_extents[cnt - 1];
1683 /* trim final overlapping extent */
1684 if (ex->fe_off + ex->fe_len > overlap)
1685 ex->fe_len = overlap - ex->fe_off;
1688 *num_img_extents = cnt;
1692 * Determine the byte range(s) covered by either just the object extent
1693 * or the entire object in the parent image.
1695 static int rbd_obj_calc_img_extents(struct rbd_obj_request *obj_req,
1698 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1701 if (!rbd_dev->parent_overlap)
1704 ret = ceph_extent_to_file(&rbd_dev->layout, obj_req->ex.oe_objno,
1705 entire ? 0 : obj_req->ex.oe_off,
1706 entire ? rbd_dev->layout.object_size :
1708 &obj_req->img_extents,
1709 &obj_req->num_img_extents);
1713 prune_extents(obj_req->img_extents, &obj_req->num_img_extents,
1714 rbd_dev->parent_overlap);
1718 static void rbd_osd_req_setup_data(struct rbd_obj_request *obj_req, u32 which)
1720 switch (obj_req->img_request->data_type) {
1721 case OBJ_REQUEST_BIO:
1722 osd_req_op_extent_osd_data_bio(obj_req->osd_req, which,
1724 obj_req->ex.oe_len);
1726 case OBJ_REQUEST_BVECS:
1727 case OBJ_REQUEST_OWN_BVECS:
1728 rbd_assert(obj_req->bvec_pos.iter.bi_size ==
1729 obj_req->ex.oe_len);
1730 rbd_assert(obj_req->bvec_idx == obj_req->bvec_count);
1731 osd_req_op_extent_osd_data_bvec_pos(obj_req->osd_req, which,
1732 &obj_req->bvec_pos);
1739 static int rbd_obj_setup_read(struct rbd_obj_request *obj_req)
1741 obj_req->osd_req = rbd_osd_req_create(obj_req, 1);
1742 if (!obj_req->osd_req)
1745 osd_req_op_extent_init(obj_req->osd_req, 0, CEPH_OSD_OP_READ,
1746 obj_req->ex.oe_off, obj_req->ex.oe_len, 0, 0);
1747 rbd_osd_req_setup_data(obj_req, 0);
1749 rbd_osd_req_format_read(obj_req);
1753 static int __rbd_obj_setup_stat(struct rbd_obj_request *obj_req,
1756 struct page **pages;
1759 * The response data for a STAT call consists of:
1766 pages = ceph_alloc_page_vector(1, GFP_NOIO);
1768 return PTR_ERR(pages);
1770 osd_req_op_init(obj_req->osd_req, which, CEPH_OSD_OP_STAT, 0);
1771 osd_req_op_raw_data_in_pages(obj_req->osd_req, which, pages,
1772 8 + sizeof(struct ceph_timespec),
1777 static void __rbd_obj_setup_write(struct rbd_obj_request *obj_req,
1780 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1783 osd_req_op_alloc_hint_init(obj_req->osd_req, which++,
1784 rbd_dev->layout.object_size,
1785 rbd_dev->layout.object_size);
1787 if (rbd_obj_is_entire(obj_req))
1788 opcode = CEPH_OSD_OP_WRITEFULL;
1790 opcode = CEPH_OSD_OP_WRITE;
1792 osd_req_op_extent_init(obj_req->osd_req, which, opcode,
1793 obj_req->ex.oe_off, obj_req->ex.oe_len, 0, 0);
1794 rbd_osd_req_setup_data(obj_req, which++);
1796 rbd_assert(which == obj_req->osd_req->r_num_ops);
1797 rbd_osd_req_format_write(obj_req);
1800 static int rbd_obj_setup_write(struct rbd_obj_request *obj_req)
1802 unsigned int num_osd_ops, which = 0;
1805 /* reverse map the entire object onto the parent */
1806 ret = rbd_obj_calc_img_extents(obj_req, true);
1810 if (obj_req->num_img_extents) {
1811 obj_req->write_state = RBD_OBJ_WRITE_GUARD;
1812 num_osd_ops = 3; /* stat + setallochint + write/writefull */
1814 obj_req->write_state = RBD_OBJ_WRITE_FLAT;
1815 num_osd_ops = 2; /* setallochint + write/writefull */
1818 obj_req->osd_req = rbd_osd_req_create(obj_req, num_osd_ops);
1819 if (!obj_req->osd_req)
1822 if (obj_req->num_img_extents) {
1823 ret = __rbd_obj_setup_stat(obj_req, which++);
1828 __rbd_obj_setup_write(obj_req, which);
1832 static void __rbd_obj_setup_discard(struct rbd_obj_request *obj_req,
1837 if (rbd_obj_is_entire(obj_req)) {
1838 if (obj_req->num_img_extents) {
1839 osd_req_op_init(obj_req->osd_req, which++,
1840 CEPH_OSD_OP_CREATE, 0);
1841 opcode = CEPH_OSD_OP_TRUNCATE;
1843 osd_req_op_init(obj_req->osd_req, which++,
1844 CEPH_OSD_OP_DELETE, 0);
1847 } else if (rbd_obj_is_tail(obj_req)) {
1848 opcode = CEPH_OSD_OP_TRUNCATE;
1850 opcode = CEPH_OSD_OP_ZERO;
1854 osd_req_op_extent_init(obj_req->osd_req, which++, opcode,
1855 obj_req->ex.oe_off, obj_req->ex.oe_len,
1858 rbd_assert(which == obj_req->osd_req->r_num_ops);
1859 rbd_osd_req_format_write(obj_req);
1862 static int rbd_obj_setup_discard(struct rbd_obj_request *obj_req)
1864 unsigned int num_osd_ops, which = 0;
1867 /* reverse map the entire object onto the parent */
1868 ret = rbd_obj_calc_img_extents(obj_req, true);
1872 if (rbd_obj_is_entire(obj_req)) {
1873 obj_req->write_state = RBD_OBJ_WRITE_FLAT;
1874 if (obj_req->num_img_extents)
1875 num_osd_ops = 2; /* create + truncate */
1877 num_osd_ops = 1; /* delete */
1879 if (obj_req->num_img_extents) {
1880 obj_req->write_state = RBD_OBJ_WRITE_GUARD;
1881 num_osd_ops = 2; /* stat + truncate/zero */
1883 obj_req->write_state = RBD_OBJ_WRITE_FLAT;
1884 num_osd_ops = 1; /* truncate/zero */
1888 obj_req->osd_req = rbd_osd_req_create(obj_req, num_osd_ops);
1889 if (!obj_req->osd_req)
1892 if (!rbd_obj_is_entire(obj_req) && obj_req->num_img_extents) {
1893 ret = __rbd_obj_setup_stat(obj_req, which++);
1898 __rbd_obj_setup_discard(obj_req, which);
1903 * For each object request in @img_req, allocate an OSD request, add
1904 * individual OSD ops and prepare them for submission. The number of
1905 * OSD ops depends on op_type and the overlap point (if any).
1907 static int __rbd_img_fill_request(struct rbd_img_request *img_req)
1909 struct rbd_obj_request *obj_req;
1912 for_each_obj_request(img_req, obj_req) {
1913 switch (img_req->op_type) {
1915 ret = rbd_obj_setup_read(obj_req);
1918 ret = rbd_obj_setup_write(obj_req);
1920 case OBJ_OP_DISCARD:
1921 ret = rbd_obj_setup_discard(obj_req);
1933 union rbd_img_fill_iter {
1934 struct ceph_bio_iter bio_iter;
1935 struct ceph_bvec_iter bvec_iter;
1938 struct rbd_img_fill_ctx {
1939 enum obj_request_type pos_type;
1940 union rbd_img_fill_iter *pos;
1941 union rbd_img_fill_iter iter;
1942 ceph_object_extent_fn_t set_pos_fn;
1943 ceph_object_extent_fn_t count_fn;
1944 ceph_object_extent_fn_t copy_fn;
1947 static struct ceph_object_extent *alloc_object_extent(void *arg)
1949 struct rbd_img_request *img_req = arg;
1950 struct rbd_obj_request *obj_req;
1952 obj_req = rbd_obj_request_create();
1956 rbd_img_obj_request_add(img_req, obj_req);
1957 return &obj_req->ex;
1961 * While su != os && sc == 1 is technically not fancy (it's the same
1962 * layout as su == os && sc == 1), we can't use the nocopy path for it
1963 * because ->set_pos_fn() should be called only once per object.
1964 * ceph_file_to_extents() invokes action_fn once per stripe unit, so
1965 * treat su != os && sc == 1 as fancy.
1967 static bool rbd_layout_is_fancy(struct ceph_file_layout *l)
1969 return l->stripe_unit != l->object_size;
1972 static int rbd_img_fill_request_nocopy(struct rbd_img_request *img_req,
1973 struct ceph_file_extent *img_extents,
1974 u32 num_img_extents,
1975 struct rbd_img_fill_ctx *fctx)
1980 img_req->data_type = fctx->pos_type;
1983 * Create object requests and set each object request's starting
1984 * position in the provided bio (list) or bio_vec array.
1986 fctx->iter = *fctx->pos;
1987 for (i = 0; i < num_img_extents; i++) {
1988 ret = ceph_file_to_extents(&img_req->rbd_dev->layout,
1989 img_extents[i].fe_off,
1990 img_extents[i].fe_len,
1991 &img_req->object_extents,
1992 alloc_object_extent, img_req,
1993 fctx->set_pos_fn, &fctx->iter);
1998 return __rbd_img_fill_request(img_req);
2002 * Map a list of image extents to a list of object extents, create the
2003 * corresponding object requests (normally each to a different object,
2004 * but not always) and add them to @img_req. For each object request,
2005 * set up its data descriptor to point to the corresponding chunk(s) of
2006 * @fctx->pos data buffer.
2008 * Because ceph_file_to_extents() will merge adjacent object extents
2009 * together, each object request's data descriptor may point to multiple
2010 * different chunks of @fctx->pos data buffer.
2012 * @fctx->pos data buffer is assumed to be large enough.
2014 static int rbd_img_fill_request(struct rbd_img_request *img_req,
2015 struct ceph_file_extent *img_extents,
2016 u32 num_img_extents,
2017 struct rbd_img_fill_ctx *fctx)
2019 struct rbd_device *rbd_dev = img_req->rbd_dev;
2020 struct rbd_obj_request *obj_req;
2024 if (fctx->pos_type == OBJ_REQUEST_NODATA ||
2025 !rbd_layout_is_fancy(&rbd_dev->layout))
2026 return rbd_img_fill_request_nocopy(img_req, img_extents,
2027 num_img_extents, fctx);
2029 img_req->data_type = OBJ_REQUEST_OWN_BVECS;
2032 * Create object requests and determine ->bvec_count for each object
2033 * request. Note that ->bvec_count sum over all object requests may
2034 * be greater than the number of bio_vecs in the provided bio (list)
2035 * or bio_vec array because when mapped, those bio_vecs can straddle
2036 * stripe unit boundaries.
2038 fctx->iter = *fctx->pos;
2039 for (i = 0; i < num_img_extents; i++) {
2040 ret = ceph_file_to_extents(&rbd_dev->layout,
2041 img_extents[i].fe_off,
2042 img_extents[i].fe_len,
2043 &img_req->object_extents,
2044 alloc_object_extent, img_req,
2045 fctx->count_fn, &fctx->iter);
2050 for_each_obj_request(img_req, obj_req) {
2051 obj_req->bvec_pos.bvecs = kmalloc_array(obj_req->bvec_count,
2052 sizeof(*obj_req->bvec_pos.bvecs),
2054 if (!obj_req->bvec_pos.bvecs)
2059 * Fill in each object request's private bio_vec array, splitting and
2060 * rearranging the provided bio_vecs in stripe unit chunks as needed.
2062 fctx->iter = *fctx->pos;
2063 for (i = 0; i < num_img_extents; i++) {
2064 ret = ceph_iterate_extents(&rbd_dev->layout,
2065 img_extents[i].fe_off,
2066 img_extents[i].fe_len,
2067 &img_req->object_extents,
2068 fctx->copy_fn, &fctx->iter);
2073 return __rbd_img_fill_request(img_req);
2076 static int rbd_img_fill_nodata(struct rbd_img_request *img_req,
2079 struct ceph_file_extent ex = { off, len };
2080 union rbd_img_fill_iter dummy;
2081 struct rbd_img_fill_ctx fctx = {
2082 .pos_type = OBJ_REQUEST_NODATA,
2086 return rbd_img_fill_request(img_req, &ex, 1, &fctx);
2089 static void set_bio_pos(struct ceph_object_extent *ex, u32 bytes, void *arg)
2091 struct rbd_obj_request *obj_req =
2092 container_of(ex, struct rbd_obj_request, ex);
2093 struct ceph_bio_iter *it = arg;
2095 dout("%s objno %llu bytes %u\n", __func__, ex->oe_objno, bytes);
2096 obj_req->bio_pos = *it;
2097 ceph_bio_iter_advance(it, bytes);
2100 static void count_bio_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2102 struct rbd_obj_request *obj_req =
2103 container_of(ex, struct rbd_obj_request, ex);
2104 struct ceph_bio_iter *it = arg;
2106 dout("%s objno %llu bytes %u\n", __func__, ex->oe_objno, bytes);
2107 ceph_bio_iter_advance_step(it, bytes, ({
2108 obj_req->bvec_count++;
2113 static void copy_bio_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2115 struct rbd_obj_request *obj_req =
2116 container_of(ex, struct rbd_obj_request, ex);
2117 struct ceph_bio_iter *it = arg;
2119 dout("%s objno %llu bytes %u\n", __func__, ex->oe_objno, bytes);
2120 ceph_bio_iter_advance_step(it, bytes, ({
2121 obj_req->bvec_pos.bvecs[obj_req->bvec_idx++] = bv;
2122 obj_req->bvec_pos.iter.bi_size += bv.bv_len;
2126 static int __rbd_img_fill_from_bio(struct rbd_img_request *img_req,
2127 struct ceph_file_extent *img_extents,
2128 u32 num_img_extents,
2129 struct ceph_bio_iter *bio_pos)
2131 struct rbd_img_fill_ctx fctx = {
2132 .pos_type = OBJ_REQUEST_BIO,
2133 .pos = (union rbd_img_fill_iter *)bio_pos,
2134 .set_pos_fn = set_bio_pos,
2135 .count_fn = count_bio_bvecs,
2136 .copy_fn = copy_bio_bvecs,
2139 return rbd_img_fill_request(img_req, img_extents, num_img_extents,
2143 static int rbd_img_fill_from_bio(struct rbd_img_request *img_req,
2144 u64 off, u64 len, struct bio *bio)
2146 struct ceph_file_extent ex = { off, len };
2147 struct ceph_bio_iter it = { .bio = bio, .iter = bio->bi_iter };
2149 return __rbd_img_fill_from_bio(img_req, &ex, 1, &it);
2152 static void set_bvec_pos(struct ceph_object_extent *ex, u32 bytes, void *arg)
2154 struct rbd_obj_request *obj_req =
2155 container_of(ex, struct rbd_obj_request, ex);
2156 struct ceph_bvec_iter *it = arg;
2158 obj_req->bvec_pos = *it;
2159 ceph_bvec_iter_shorten(&obj_req->bvec_pos, bytes);
2160 ceph_bvec_iter_advance(it, bytes);
2163 static void count_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2165 struct rbd_obj_request *obj_req =
2166 container_of(ex, struct rbd_obj_request, ex);
2167 struct ceph_bvec_iter *it = arg;
2169 ceph_bvec_iter_advance_step(it, bytes, ({
2170 obj_req->bvec_count++;
2174 static void copy_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2176 struct rbd_obj_request *obj_req =
2177 container_of(ex, struct rbd_obj_request, ex);
2178 struct ceph_bvec_iter *it = arg;
2180 ceph_bvec_iter_advance_step(it, bytes, ({
2181 obj_req->bvec_pos.bvecs[obj_req->bvec_idx++] = bv;
2182 obj_req->bvec_pos.iter.bi_size += bv.bv_len;
2186 static int __rbd_img_fill_from_bvecs(struct rbd_img_request *img_req,
2187 struct ceph_file_extent *img_extents,
2188 u32 num_img_extents,
2189 struct ceph_bvec_iter *bvec_pos)
2191 struct rbd_img_fill_ctx fctx = {
2192 .pos_type = OBJ_REQUEST_BVECS,
2193 .pos = (union rbd_img_fill_iter *)bvec_pos,
2194 .set_pos_fn = set_bvec_pos,
2195 .count_fn = count_bvecs,
2196 .copy_fn = copy_bvecs,
2199 return rbd_img_fill_request(img_req, img_extents, num_img_extents,
2203 static int rbd_img_fill_from_bvecs(struct rbd_img_request *img_req,
2204 struct ceph_file_extent *img_extents,
2205 u32 num_img_extents,
2206 struct bio_vec *bvecs)
2208 struct ceph_bvec_iter it = {
2210 .iter = { .bi_size = ceph_file_extents_bytes(img_extents,
2214 return __rbd_img_fill_from_bvecs(img_req, img_extents, num_img_extents,
2218 static void rbd_img_request_submit(struct rbd_img_request *img_request)
2220 struct rbd_obj_request *obj_request;
2222 dout("%s: img %p\n", __func__, img_request);
2224 rbd_img_request_get(img_request);
2225 for_each_obj_request(img_request, obj_request)
2226 rbd_obj_request_submit(obj_request);
2228 rbd_img_request_put(img_request);
2231 static int rbd_obj_read_from_parent(struct rbd_obj_request *obj_req)
2233 struct rbd_img_request *img_req = obj_req->img_request;
2234 struct rbd_img_request *child_img_req;
2237 child_img_req = rbd_img_request_create(img_req->rbd_dev->parent,
2242 __set_bit(IMG_REQ_CHILD, &child_img_req->flags);
2243 child_img_req->obj_request = obj_req;
2245 if (!rbd_img_is_write(img_req)) {
2246 switch (img_req->data_type) {
2247 case OBJ_REQUEST_BIO:
2248 ret = __rbd_img_fill_from_bio(child_img_req,
2249 obj_req->img_extents,
2250 obj_req->num_img_extents,
2253 case OBJ_REQUEST_BVECS:
2254 case OBJ_REQUEST_OWN_BVECS:
2255 ret = __rbd_img_fill_from_bvecs(child_img_req,
2256 obj_req->img_extents,
2257 obj_req->num_img_extents,
2258 &obj_req->bvec_pos);
2264 ret = rbd_img_fill_from_bvecs(child_img_req,
2265 obj_req->img_extents,
2266 obj_req->num_img_extents,
2267 obj_req->copyup_bvecs);
2270 rbd_img_request_put(child_img_req);
2274 rbd_img_request_submit(child_img_req);
2278 static bool rbd_obj_handle_read(struct rbd_obj_request *obj_req)
2280 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2283 if (obj_req->result == -ENOENT &&
2284 rbd_dev->parent_overlap && !obj_req->tried_parent) {
2285 /* reverse map this object extent onto the parent */
2286 ret = rbd_obj_calc_img_extents(obj_req, false);
2288 obj_req->result = ret;
2292 if (obj_req->num_img_extents) {
2293 obj_req->tried_parent = true;
2294 ret = rbd_obj_read_from_parent(obj_req);
2296 obj_req->result = ret;
2304 * -ENOENT means a hole in the image -- zero-fill the entire
2305 * length of the request. A short read also implies zero-fill
2306 * to the end of the request. In both cases we update xferred
2307 * count to indicate the whole request was satisfied.
2309 if (obj_req->result == -ENOENT ||
2310 (!obj_req->result && obj_req->xferred < obj_req->ex.oe_len)) {
2311 rbd_assert(!obj_req->xferred || !obj_req->result);
2312 rbd_obj_zero_range(obj_req, obj_req->xferred,
2313 obj_req->ex.oe_len - obj_req->xferred);
2314 obj_req->result = 0;
2315 obj_req->xferred = obj_req->ex.oe_len;
2322 * copyup_bvecs pages are never highmem pages
2324 static bool is_zero_bvecs(struct bio_vec *bvecs, u32 bytes)
2326 struct ceph_bvec_iter it = {
2328 .iter = { .bi_size = bytes },
2331 ceph_bvec_iter_advance_step(&it, bytes, ({
2332 if (memchr_inv(page_address(bv.bv_page) + bv.bv_offset, 0,
2339 static int rbd_obj_issue_copyup(struct rbd_obj_request *obj_req, u32 bytes)
2341 unsigned int num_osd_ops = obj_req->osd_req->r_num_ops;
2344 dout("%s obj_req %p bytes %u\n", __func__, obj_req, bytes);
2345 rbd_assert(obj_req->osd_req->r_ops[0].op == CEPH_OSD_OP_STAT);
2346 rbd_osd_req_destroy(obj_req->osd_req);
2349 * Create a copyup request with the same number of OSD ops as
2350 * the original request. The original request was stat + op(s),
2351 * the new copyup request will be copyup + the same op(s).
2353 obj_req->osd_req = rbd_osd_req_create(obj_req, num_osd_ops);
2354 if (!obj_req->osd_req)
2357 ret = osd_req_op_cls_init(obj_req->osd_req, 0, CEPH_OSD_OP_CALL, "rbd",
2363 * Only send non-zero copyup data to save some I/O and network
2364 * bandwidth -- zero copyup data is equivalent to the object not
2367 if (is_zero_bvecs(obj_req->copyup_bvecs, bytes)) {
2368 dout("%s obj_req %p detected zeroes\n", __func__, obj_req);
2371 osd_req_op_cls_request_data_bvecs(obj_req->osd_req, 0,
2372 obj_req->copyup_bvecs,
2373 obj_req->copyup_bvec_count,
2376 switch (obj_req->img_request->op_type) {
2378 __rbd_obj_setup_write(obj_req, 1);
2380 case OBJ_OP_DISCARD:
2381 rbd_assert(!rbd_obj_is_entire(obj_req));
2382 __rbd_obj_setup_discard(obj_req, 1);
2388 rbd_obj_request_submit(obj_req);
2392 static int setup_copyup_bvecs(struct rbd_obj_request *obj_req, u64 obj_overlap)
2396 rbd_assert(!obj_req->copyup_bvecs);
2397 obj_req->copyup_bvec_count = calc_pages_for(0, obj_overlap);
2398 obj_req->copyup_bvecs = kcalloc(obj_req->copyup_bvec_count,
2399 sizeof(*obj_req->copyup_bvecs),
2401 if (!obj_req->copyup_bvecs)
2404 for (i = 0; i < obj_req->copyup_bvec_count; i++) {
2405 unsigned int len = min(obj_overlap, (u64)PAGE_SIZE);
2407 obj_req->copyup_bvecs[i].bv_page = alloc_page(GFP_NOIO);
2408 if (!obj_req->copyup_bvecs[i].bv_page)
2411 obj_req->copyup_bvecs[i].bv_offset = 0;
2412 obj_req->copyup_bvecs[i].bv_len = len;
2416 rbd_assert(!obj_overlap);
2420 static int rbd_obj_handle_write_guard(struct rbd_obj_request *obj_req)
2422 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2425 rbd_assert(obj_req->num_img_extents);
2426 prune_extents(obj_req->img_extents, &obj_req->num_img_extents,
2427 rbd_dev->parent_overlap);
2428 if (!obj_req->num_img_extents) {
2430 * The overlap has become 0 (most likely because the
2431 * image has been flattened). Use rbd_obj_issue_copyup()
2432 * to re-submit the original write request -- the copyup
2433 * operation itself will be a no-op, since someone must
2434 * have populated the child object while we weren't
2435 * looking. Move to WRITE_FLAT state as we'll be done
2436 * with the operation once the null copyup completes.
2438 obj_req->write_state = RBD_OBJ_WRITE_FLAT;
2439 return rbd_obj_issue_copyup(obj_req, 0);
2442 ret = setup_copyup_bvecs(obj_req, rbd_obj_img_extents_bytes(obj_req));
2446 obj_req->write_state = RBD_OBJ_WRITE_COPYUP;
2447 return rbd_obj_read_from_parent(obj_req);
2450 static bool rbd_obj_handle_write(struct rbd_obj_request *obj_req)
2455 switch (obj_req->write_state) {
2456 case RBD_OBJ_WRITE_GUARD:
2457 rbd_assert(!obj_req->xferred);
2458 if (obj_req->result == -ENOENT) {
2460 * The target object doesn't exist. Read the data for
2461 * the entire target object up to the overlap point (if
2462 * any) from the parent, so we can use it for a copyup.
2464 ret = rbd_obj_handle_write_guard(obj_req);
2466 obj_req->result = ret;
2472 case RBD_OBJ_WRITE_FLAT:
2473 if (!obj_req->result)
2475 * There is no such thing as a successful short
2476 * write -- indicate the whole request was satisfied.
2478 obj_req->xferred = obj_req->ex.oe_len;
2480 case RBD_OBJ_WRITE_COPYUP:
2481 obj_req->write_state = RBD_OBJ_WRITE_GUARD;
2482 if (obj_req->result)
2485 rbd_assert(obj_req->xferred);
2486 ret = rbd_obj_issue_copyup(obj_req, obj_req->xferred);
2488 obj_req->result = ret;
2498 * Returns true if @obj_req is completed, or false otherwise.
2500 static bool __rbd_obj_handle_request(struct rbd_obj_request *obj_req)
2502 switch (obj_req->img_request->op_type) {
2504 return rbd_obj_handle_read(obj_req);
2506 return rbd_obj_handle_write(obj_req);
2507 case OBJ_OP_DISCARD:
2508 if (rbd_obj_handle_write(obj_req)) {
2510 * Hide -ENOENT from delete/truncate/zero -- discarding
2511 * a non-existent object is not a problem.
2513 if (obj_req->result == -ENOENT) {
2514 obj_req->result = 0;
2515 obj_req->xferred = obj_req->ex.oe_len;
2525 static void rbd_obj_end_request(struct rbd_obj_request *obj_req)
2527 struct rbd_img_request *img_req = obj_req->img_request;
2529 rbd_assert((!obj_req->result &&
2530 obj_req->xferred == obj_req->ex.oe_len) ||
2531 (obj_req->result < 0 && !obj_req->xferred));
2532 if (!obj_req->result) {
2533 img_req->xferred += obj_req->xferred;
2537 rbd_warn(img_req->rbd_dev,
2538 "%s at objno %llu %llu~%llu result %d xferred %llu",
2539 obj_op_name(img_req->op_type), obj_req->ex.oe_objno,
2540 obj_req->ex.oe_off, obj_req->ex.oe_len, obj_req->result,
2542 if (!img_req->result) {
2543 img_req->result = obj_req->result;
2544 img_req->xferred = 0;
2548 static void rbd_img_end_child_request(struct rbd_img_request *img_req)
2550 struct rbd_obj_request *obj_req = img_req->obj_request;
2552 rbd_assert(test_bit(IMG_REQ_CHILD, &img_req->flags));
2553 rbd_assert((!img_req->result &&
2554 img_req->xferred == rbd_obj_img_extents_bytes(obj_req)) ||
2555 (img_req->result < 0 && !img_req->xferred));
2557 obj_req->result = img_req->result;
2558 obj_req->xferred = img_req->xferred;
2559 rbd_img_request_put(img_req);
2562 static void rbd_img_end_request(struct rbd_img_request *img_req)
2564 rbd_assert(!test_bit(IMG_REQ_CHILD, &img_req->flags));
2565 rbd_assert((!img_req->result &&
2566 img_req->xferred == blk_rq_bytes(img_req->rq)) ||
2567 (img_req->result < 0 && !img_req->xferred));
2569 blk_mq_end_request(img_req->rq,
2570 errno_to_blk_status(img_req->result));
2571 rbd_img_request_put(img_req);
2574 static void rbd_obj_handle_request(struct rbd_obj_request *obj_req)
2576 struct rbd_img_request *img_req;
2579 if (!__rbd_obj_handle_request(obj_req))
2582 img_req = obj_req->img_request;
2583 spin_lock(&img_req->completion_lock);
2584 rbd_obj_end_request(obj_req);
2585 rbd_assert(img_req->pending_count);
2586 if (--img_req->pending_count) {
2587 spin_unlock(&img_req->completion_lock);
2591 spin_unlock(&img_req->completion_lock);
2592 if (test_bit(IMG_REQ_CHILD, &img_req->flags)) {
2593 obj_req = img_req->obj_request;
2594 rbd_img_end_child_request(img_req);
2597 rbd_img_end_request(img_req);
2600 static const struct rbd_client_id rbd_empty_cid;
2602 static bool rbd_cid_equal(const struct rbd_client_id *lhs,
2603 const struct rbd_client_id *rhs)
2605 return lhs->gid == rhs->gid && lhs->handle == rhs->handle;
2608 static struct rbd_client_id rbd_get_cid(struct rbd_device *rbd_dev)
2610 struct rbd_client_id cid;
2612 mutex_lock(&rbd_dev->watch_mutex);
2613 cid.gid = ceph_client_gid(rbd_dev->rbd_client->client);
2614 cid.handle = rbd_dev->watch_cookie;
2615 mutex_unlock(&rbd_dev->watch_mutex);
2620 * lock_rwsem must be held for write
2622 static void rbd_set_owner_cid(struct rbd_device *rbd_dev,
2623 const struct rbd_client_id *cid)
2625 dout("%s rbd_dev %p %llu-%llu -> %llu-%llu\n", __func__, rbd_dev,
2626 rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle,
2627 cid->gid, cid->handle);
2628 rbd_dev->owner_cid = *cid; /* struct */
2631 static void format_lock_cookie(struct rbd_device *rbd_dev, char *buf)
2633 mutex_lock(&rbd_dev->watch_mutex);
2634 sprintf(buf, "%s %llu", RBD_LOCK_COOKIE_PREFIX, rbd_dev->watch_cookie);
2635 mutex_unlock(&rbd_dev->watch_mutex);
2638 static void __rbd_lock(struct rbd_device *rbd_dev, const char *cookie)
2640 struct rbd_client_id cid = rbd_get_cid(rbd_dev);
2642 strcpy(rbd_dev->lock_cookie, cookie);
2643 rbd_set_owner_cid(rbd_dev, &cid);
2644 queue_work(rbd_dev->task_wq, &rbd_dev->acquired_lock_work);
2648 * lock_rwsem must be held for write
2650 static int rbd_lock(struct rbd_device *rbd_dev)
2652 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2656 WARN_ON(__rbd_is_lock_owner(rbd_dev) ||
2657 rbd_dev->lock_cookie[0] != '\0');
2659 format_lock_cookie(rbd_dev, cookie);
2660 ret = ceph_cls_lock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
2661 RBD_LOCK_NAME, CEPH_CLS_LOCK_EXCLUSIVE, cookie,
2662 RBD_LOCK_TAG, "", 0);
2666 rbd_dev->lock_state = RBD_LOCK_STATE_LOCKED;
2667 __rbd_lock(rbd_dev, cookie);
2672 * lock_rwsem must be held for write
2674 static void rbd_unlock(struct rbd_device *rbd_dev)
2676 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2679 WARN_ON(!__rbd_is_lock_owner(rbd_dev) ||
2680 rbd_dev->lock_cookie[0] == '\0');
2682 ret = ceph_cls_unlock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
2683 RBD_LOCK_NAME, rbd_dev->lock_cookie);
2684 if (ret && ret != -ENOENT)
2685 rbd_warn(rbd_dev, "failed to unlock: %d", ret);
2687 /* treat errors as the image is unlocked */
2688 rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
2689 rbd_dev->lock_cookie[0] = '\0';
2690 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
2691 queue_work(rbd_dev->task_wq, &rbd_dev->released_lock_work);
2694 static int __rbd_notify_op_lock(struct rbd_device *rbd_dev,
2695 enum rbd_notify_op notify_op,
2696 struct page ***preply_pages,
2699 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2700 struct rbd_client_id cid = rbd_get_cid(rbd_dev);
2701 char buf[4 + 8 + 8 + CEPH_ENCODING_START_BLK_LEN];
2702 int buf_size = sizeof(buf);
2705 dout("%s rbd_dev %p notify_op %d\n", __func__, rbd_dev, notify_op);
2707 /* encode *LockPayload NotifyMessage (op + ClientId) */
2708 ceph_start_encoding(&p, 2, 1, buf_size - CEPH_ENCODING_START_BLK_LEN);
2709 ceph_encode_32(&p, notify_op);
2710 ceph_encode_64(&p, cid.gid);
2711 ceph_encode_64(&p, cid.handle);
2713 return ceph_osdc_notify(osdc, &rbd_dev->header_oid,
2714 &rbd_dev->header_oloc, buf, buf_size,
2715 RBD_NOTIFY_TIMEOUT, preply_pages, preply_len);
2718 static void rbd_notify_op_lock(struct rbd_device *rbd_dev,
2719 enum rbd_notify_op notify_op)
2721 struct page **reply_pages;
2724 __rbd_notify_op_lock(rbd_dev, notify_op, &reply_pages, &reply_len);
2725 ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len));
2728 static void rbd_notify_acquired_lock(struct work_struct *work)
2730 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
2731 acquired_lock_work);
2733 rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_ACQUIRED_LOCK);
2736 static void rbd_notify_released_lock(struct work_struct *work)
2738 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
2739 released_lock_work);
2741 rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_RELEASED_LOCK);
2744 static int rbd_request_lock(struct rbd_device *rbd_dev)
2746 struct page **reply_pages;
2748 bool lock_owner_responded = false;
2751 dout("%s rbd_dev %p\n", __func__, rbd_dev);
2753 ret = __rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_REQUEST_LOCK,
2754 &reply_pages, &reply_len);
2755 if (ret && ret != -ETIMEDOUT) {
2756 rbd_warn(rbd_dev, "failed to request lock: %d", ret);
2760 if (reply_len > 0 && reply_len <= PAGE_SIZE) {
2761 void *p = page_address(reply_pages[0]);
2762 void *const end = p + reply_len;
2765 ceph_decode_32_safe(&p, end, n, e_inval); /* num_acks */
2770 ceph_decode_need(&p, end, 8 + 8, e_inval);
2771 p += 8 + 8; /* skip gid and cookie */
2773 ceph_decode_32_safe(&p, end, len, e_inval);
2777 if (lock_owner_responded) {
2779 "duplicate lock owners detected");
2784 lock_owner_responded = true;
2785 ret = ceph_start_decoding(&p, end, 1, "ResponseMessage",
2789 "failed to decode ResponseMessage: %d",
2794 ret = ceph_decode_32(&p);
2798 if (!lock_owner_responded) {
2799 rbd_warn(rbd_dev, "no lock owners detected");
2804 ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len));
2812 static void wake_requests(struct rbd_device *rbd_dev, bool wake_all)
2814 dout("%s rbd_dev %p wake_all %d\n", __func__, rbd_dev, wake_all);
2816 cancel_delayed_work(&rbd_dev->lock_dwork);
2818 wake_up_all(&rbd_dev->lock_waitq);
2820 wake_up(&rbd_dev->lock_waitq);
2823 static int get_lock_owner_info(struct rbd_device *rbd_dev,
2824 struct ceph_locker **lockers, u32 *num_lockers)
2826 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2831 dout("%s rbd_dev %p\n", __func__, rbd_dev);
2833 ret = ceph_cls_lock_info(osdc, &rbd_dev->header_oid,
2834 &rbd_dev->header_oloc, RBD_LOCK_NAME,
2835 &lock_type, &lock_tag, lockers, num_lockers);
2839 if (*num_lockers == 0) {
2840 dout("%s rbd_dev %p no lockers detected\n", __func__, rbd_dev);
2844 if (strcmp(lock_tag, RBD_LOCK_TAG)) {
2845 rbd_warn(rbd_dev, "locked by external mechanism, tag %s",
2851 if (lock_type == CEPH_CLS_LOCK_SHARED) {
2852 rbd_warn(rbd_dev, "shared lock type detected");
2857 if (strncmp((*lockers)[0].id.cookie, RBD_LOCK_COOKIE_PREFIX,
2858 strlen(RBD_LOCK_COOKIE_PREFIX))) {
2859 rbd_warn(rbd_dev, "locked by external mechanism, cookie %s",
2860 (*lockers)[0].id.cookie);
2870 static int find_watcher(struct rbd_device *rbd_dev,
2871 const struct ceph_locker *locker)
2873 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2874 struct ceph_watch_item *watchers;
2880 ret = ceph_osdc_list_watchers(osdc, &rbd_dev->header_oid,
2881 &rbd_dev->header_oloc, &watchers,
2886 sscanf(locker->id.cookie, RBD_LOCK_COOKIE_PREFIX " %llu", &cookie);
2887 for (i = 0; i < num_watchers; i++) {
2888 if (!memcmp(&watchers[i].addr, &locker->info.addr,
2889 sizeof(locker->info.addr)) &&
2890 watchers[i].cookie == cookie) {
2891 struct rbd_client_id cid = {
2892 .gid = le64_to_cpu(watchers[i].name.num),
2896 dout("%s rbd_dev %p found cid %llu-%llu\n", __func__,
2897 rbd_dev, cid.gid, cid.handle);
2898 rbd_set_owner_cid(rbd_dev, &cid);
2904 dout("%s rbd_dev %p no watchers\n", __func__, rbd_dev);
2912 * lock_rwsem must be held for write
2914 static int rbd_try_lock(struct rbd_device *rbd_dev)
2916 struct ceph_client *client = rbd_dev->rbd_client->client;
2917 struct ceph_locker *lockers;
2922 ret = rbd_lock(rbd_dev);
2926 /* determine if the current lock holder is still alive */
2927 ret = get_lock_owner_info(rbd_dev, &lockers, &num_lockers);
2931 if (num_lockers == 0)
2934 ret = find_watcher(rbd_dev, lockers);
2937 ret = 0; /* have to request lock */
2941 rbd_warn(rbd_dev, "%s%llu seems dead, breaking lock",
2942 ENTITY_NAME(lockers[0].id.name));
2944 ret = ceph_monc_blacklist_add(&client->monc,
2945 &lockers[0].info.addr);
2947 rbd_warn(rbd_dev, "blacklist of %s%llu failed: %d",
2948 ENTITY_NAME(lockers[0].id.name), ret);
2952 ret = ceph_cls_break_lock(&client->osdc, &rbd_dev->header_oid,
2953 &rbd_dev->header_oloc, RBD_LOCK_NAME,
2954 lockers[0].id.cookie,
2955 &lockers[0].id.name);
2956 if (ret && ret != -ENOENT)
2960 ceph_free_lockers(lockers, num_lockers);
2964 ceph_free_lockers(lockers, num_lockers);
2969 * ret is set only if lock_state is RBD_LOCK_STATE_UNLOCKED
2971 static enum rbd_lock_state rbd_try_acquire_lock(struct rbd_device *rbd_dev,
2974 enum rbd_lock_state lock_state;
2976 down_read(&rbd_dev->lock_rwsem);
2977 dout("%s rbd_dev %p read lock_state %d\n", __func__, rbd_dev,
2978 rbd_dev->lock_state);
2979 if (__rbd_is_lock_owner(rbd_dev)) {
2980 lock_state = rbd_dev->lock_state;
2981 up_read(&rbd_dev->lock_rwsem);
2985 up_read(&rbd_dev->lock_rwsem);
2986 down_write(&rbd_dev->lock_rwsem);
2987 dout("%s rbd_dev %p write lock_state %d\n", __func__, rbd_dev,
2988 rbd_dev->lock_state);
2989 if (!__rbd_is_lock_owner(rbd_dev)) {
2990 *pret = rbd_try_lock(rbd_dev);
2992 rbd_warn(rbd_dev, "failed to acquire lock: %d", *pret);
2995 lock_state = rbd_dev->lock_state;
2996 up_write(&rbd_dev->lock_rwsem);
3000 static void rbd_acquire_lock(struct work_struct *work)
3002 struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
3003 struct rbd_device, lock_dwork);
3004 enum rbd_lock_state lock_state;
3007 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3009 lock_state = rbd_try_acquire_lock(rbd_dev, &ret);
3010 if (lock_state != RBD_LOCK_STATE_UNLOCKED || ret == -EBLACKLISTED) {
3011 if (lock_state == RBD_LOCK_STATE_LOCKED)
3012 wake_requests(rbd_dev, true);
3013 dout("%s rbd_dev %p lock_state %d ret %d - done\n", __func__,
3014 rbd_dev, lock_state, ret);
3018 ret = rbd_request_lock(rbd_dev);
3019 if (ret == -ETIMEDOUT) {
3020 goto again; /* treat this as a dead client */
3021 } else if (ret == -EROFS) {
3022 rbd_warn(rbd_dev, "peer will not release lock");
3024 * If this is rbd_add_acquire_lock(), we want to fail
3025 * immediately -- reuse BLACKLISTED flag. Otherwise we
3028 if (!(rbd_dev->disk->flags & GENHD_FL_UP)) {
3029 set_bit(RBD_DEV_FLAG_BLACKLISTED, &rbd_dev->flags);
3030 /* wake "rbd map --exclusive" process */
3031 wake_requests(rbd_dev, false);
3033 } else if (ret < 0) {
3034 rbd_warn(rbd_dev, "error requesting lock: %d", ret);
3035 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
3039 * lock owner acked, but resend if we don't see them
3042 dout("%s rbd_dev %p requeueing lock_dwork\n", __func__,
3044 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
3045 msecs_to_jiffies(2 * RBD_NOTIFY_TIMEOUT * MSEC_PER_SEC));
3050 * lock_rwsem must be held for write
3052 static bool rbd_release_lock(struct rbd_device *rbd_dev)
3054 dout("%s rbd_dev %p read lock_state %d\n", __func__, rbd_dev,
3055 rbd_dev->lock_state);
3056 if (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED)
3059 rbd_dev->lock_state = RBD_LOCK_STATE_RELEASING;
3060 downgrade_write(&rbd_dev->lock_rwsem);
3062 * Ensure that all in-flight IO is flushed.
3064 * FIXME: ceph_osdc_sync() flushes the entire OSD client, which
3065 * may be shared with other devices.
3067 ceph_osdc_sync(&rbd_dev->rbd_client->client->osdc);
3068 up_read(&rbd_dev->lock_rwsem);
3070 down_write(&rbd_dev->lock_rwsem);
3071 dout("%s rbd_dev %p write lock_state %d\n", __func__, rbd_dev,
3072 rbd_dev->lock_state);
3073 if (rbd_dev->lock_state != RBD_LOCK_STATE_RELEASING)
3076 rbd_unlock(rbd_dev);
3078 * Give others a chance to grab the lock - we would re-acquire
3079 * almost immediately if we got new IO during ceph_osdc_sync()
3080 * otherwise. We need to ack our own notifications, so this
3081 * lock_dwork will be requeued from rbd_wait_state_locked()
3082 * after wake_requests() in rbd_handle_released_lock().
3084 cancel_delayed_work(&rbd_dev->lock_dwork);
3088 static void rbd_release_lock_work(struct work_struct *work)
3090 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3093 down_write(&rbd_dev->lock_rwsem);
3094 rbd_release_lock(rbd_dev);
3095 up_write(&rbd_dev->lock_rwsem);
3098 static void rbd_handle_acquired_lock(struct rbd_device *rbd_dev, u8 struct_v,
3101 struct rbd_client_id cid = { 0 };
3103 if (struct_v >= 2) {
3104 cid.gid = ceph_decode_64(p);
3105 cid.handle = ceph_decode_64(p);
3108 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3110 if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
3111 down_write(&rbd_dev->lock_rwsem);
3112 if (rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
3114 * we already know that the remote client is
3117 up_write(&rbd_dev->lock_rwsem);
3121 rbd_set_owner_cid(rbd_dev, &cid);
3122 downgrade_write(&rbd_dev->lock_rwsem);
3124 down_read(&rbd_dev->lock_rwsem);
3127 if (!__rbd_is_lock_owner(rbd_dev))
3128 wake_requests(rbd_dev, false);
3129 up_read(&rbd_dev->lock_rwsem);
3132 static void rbd_handle_released_lock(struct rbd_device *rbd_dev, u8 struct_v,
3135 struct rbd_client_id cid = { 0 };
3137 if (struct_v >= 2) {
3138 cid.gid = ceph_decode_64(p);
3139 cid.handle = ceph_decode_64(p);
3142 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3144 if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
3145 down_write(&rbd_dev->lock_rwsem);
3146 if (!rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
3147 dout("%s rbd_dev %p unexpected owner, cid %llu-%llu != owner_cid %llu-%llu\n",
3148 __func__, rbd_dev, cid.gid, cid.handle,
3149 rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle);
3150 up_write(&rbd_dev->lock_rwsem);
3154 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3155 downgrade_write(&rbd_dev->lock_rwsem);
3157 down_read(&rbd_dev->lock_rwsem);
3160 if (!__rbd_is_lock_owner(rbd_dev))
3161 wake_requests(rbd_dev, false);
3162 up_read(&rbd_dev->lock_rwsem);
3166 * Returns result for ResponseMessage to be encoded (<= 0), or 1 if no
3167 * ResponseMessage is needed.
3169 static int rbd_handle_request_lock(struct rbd_device *rbd_dev, u8 struct_v,
3172 struct rbd_client_id my_cid = rbd_get_cid(rbd_dev);
3173 struct rbd_client_id cid = { 0 };
3176 if (struct_v >= 2) {
3177 cid.gid = ceph_decode_64(p);
3178 cid.handle = ceph_decode_64(p);
3181 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3183 if (rbd_cid_equal(&cid, &my_cid))
3186 down_read(&rbd_dev->lock_rwsem);
3187 if (__rbd_is_lock_owner(rbd_dev)) {
3188 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED &&
3189 rbd_cid_equal(&rbd_dev->owner_cid, &rbd_empty_cid))
3193 * encode ResponseMessage(0) so the peer can detect
3198 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED) {
3199 if (!rbd_dev->opts->exclusive) {
3200 dout("%s rbd_dev %p queueing unlock_work\n",
3202 queue_work(rbd_dev->task_wq,
3203 &rbd_dev->unlock_work);
3205 /* refuse to release the lock */
3212 up_read(&rbd_dev->lock_rwsem);
3216 static void __rbd_acknowledge_notify(struct rbd_device *rbd_dev,
3217 u64 notify_id, u64 cookie, s32 *result)
3219 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3220 char buf[4 + CEPH_ENCODING_START_BLK_LEN];
3221 int buf_size = sizeof(buf);
3227 /* encode ResponseMessage */
3228 ceph_start_encoding(&p, 1, 1,
3229 buf_size - CEPH_ENCODING_START_BLK_LEN);
3230 ceph_encode_32(&p, *result);
3235 ret = ceph_osdc_notify_ack(osdc, &rbd_dev->header_oid,
3236 &rbd_dev->header_oloc, notify_id, cookie,
3239 rbd_warn(rbd_dev, "acknowledge_notify failed: %d", ret);
3242 static void rbd_acknowledge_notify(struct rbd_device *rbd_dev, u64 notify_id,
3245 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3246 __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, NULL);
3249 static void rbd_acknowledge_notify_result(struct rbd_device *rbd_dev,
3250 u64 notify_id, u64 cookie, s32 result)
3252 dout("%s rbd_dev %p result %d\n", __func__, rbd_dev, result);
3253 __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, &result);
3256 static void rbd_watch_cb(void *arg, u64 notify_id, u64 cookie,
3257 u64 notifier_id, void *data, size_t data_len)
3259 struct rbd_device *rbd_dev = arg;
3261 void *const end = p + data_len;
3267 dout("%s rbd_dev %p cookie %llu notify_id %llu data_len %zu\n",
3268 __func__, rbd_dev, cookie, notify_id, data_len);
3270 ret = ceph_start_decoding(&p, end, 1, "NotifyMessage",
3273 rbd_warn(rbd_dev, "failed to decode NotifyMessage: %d",
3278 notify_op = ceph_decode_32(&p);
3280 /* legacy notification for header updates */
3281 notify_op = RBD_NOTIFY_OP_HEADER_UPDATE;
3285 dout("%s rbd_dev %p notify_op %u\n", __func__, rbd_dev, notify_op);
3286 switch (notify_op) {
3287 case RBD_NOTIFY_OP_ACQUIRED_LOCK:
3288 rbd_handle_acquired_lock(rbd_dev, struct_v, &p);
3289 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3291 case RBD_NOTIFY_OP_RELEASED_LOCK:
3292 rbd_handle_released_lock(rbd_dev, struct_v, &p);
3293 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3295 case RBD_NOTIFY_OP_REQUEST_LOCK:
3296 ret = rbd_handle_request_lock(rbd_dev, struct_v, &p);
3298 rbd_acknowledge_notify_result(rbd_dev, notify_id,
3301 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3303 case RBD_NOTIFY_OP_HEADER_UPDATE:
3304 ret = rbd_dev_refresh(rbd_dev);
3306 rbd_warn(rbd_dev, "refresh failed: %d", ret);
3308 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3311 if (rbd_is_lock_owner(rbd_dev))
3312 rbd_acknowledge_notify_result(rbd_dev, notify_id,
3313 cookie, -EOPNOTSUPP);
3315 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3320 static void __rbd_unregister_watch(struct rbd_device *rbd_dev);
3322 static void rbd_watch_errcb(void *arg, u64 cookie, int err)
3324 struct rbd_device *rbd_dev = arg;
3326 rbd_warn(rbd_dev, "encountered watch error: %d", err);
3328 down_write(&rbd_dev->lock_rwsem);
3329 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3330 up_write(&rbd_dev->lock_rwsem);
3332 mutex_lock(&rbd_dev->watch_mutex);
3333 if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED) {
3334 __rbd_unregister_watch(rbd_dev);
3335 rbd_dev->watch_state = RBD_WATCH_STATE_ERROR;
3337 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->watch_dwork, 0);
3339 mutex_unlock(&rbd_dev->watch_mutex);
3343 * watch_mutex must be locked
3345 static int __rbd_register_watch(struct rbd_device *rbd_dev)
3347 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3348 struct ceph_osd_linger_request *handle;
3350 rbd_assert(!rbd_dev->watch_handle);
3351 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3353 handle = ceph_osdc_watch(osdc, &rbd_dev->header_oid,
3354 &rbd_dev->header_oloc, rbd_watch_cb,
3355 rbd_watch_errcb, rbd_dev);
3357 return PTR_ERR(handle);
3359 rbd_dev->watch_handle = handle;
3364 * watch_mutex must be locked
3366 static void __rbd_unregister_watch(struct rbd_device *rbd_dev)
3368 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3371 rbd_assert(rbd_dev->watch_handle);
3372 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3374 ret = ceph_osdc_unwatch(osdc, rbd_dev->watch_handle);
3376 rbd_warn(rbd_dev, "failed to unwatch: %d", ret);
3378 rbd_dev->watch_handle = NULL;
3381 static int rbd_register_watch(struct rbd_device *rbd_dev)
3385 mutex_lock(&rbd_dev->watch_mutex);
3386 rbd_assert(rbd_dev->watch_state == RBD_WATCH_STATE_UNREGISTERED);
3387 ret = __rbd_register_watch(rbd_dev);
3391 rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
3392 rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
3395 mutex_unlock(&rbd_dev->watch_mutex);
3399 static void cancel_tasks_sync(struct rbd_device *rbd_dev)
3401 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3403 cancel_work_sync(&rbd_dev->acquired_lock_work);
3404 cancel_work_sync(&rbd_dev->released_lock_work);
3405 cancel_delayed_work_sync(&rbd_dev->lock_dwork);
3406 cancel_work_sync(&rbd_dev->unlock_work);
3409 static void rbd_unregister_watch(struct rbd_device *rbd_dev)
3411 WARN_ON(waitqueue_active(&rbd_dev->lock_waitq));
3412 cancel_tasks_sync(rbd_dev);
3414 mutex_lock(&rbd_dev->watch_mutex);
3415 if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED)
3416 __rbd_unregister_watch(rbd_dev);
3417 rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
3418 mutex_unlock(&rbd_dev->watch_mutex);
3420 cancel_delayed_work_sync(&rbd_dev->watch_dwork);
3421 ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
3425 * lock_rwsem must be held for write
3427 static void rbd_reacquire_lock(struct rbd_device *rbd_dev)
3429 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3433 WARN_ON(rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED);
3435 format_lock_cookie(rbd_dev, cookie);
3436 ret = ceph_cls_set_cookie(osdc, &rbd_dev->header_oid,
3437 &rbd_dev->header_oloc, RBD_LOCK_NAME,
3438 CEPH_CLS_LOCK_EXCLUSIVE, rbd_dev->lock_cookie,
3439 RBD_LOCK_TAG, cookie);
3441 if (ret != -EOPNOTSUPP)
3442 rbd_warn(rbd_dev, "failed to update lock cookie: %d",
3446 * Lock cookie cannot be updated on older OSDs, so do
3447 * a manual release and queue an acquire.
3449 if (rbd_release_lock(rbd_dev))
3450 queue_delayed_work(rbd_dev->task_wq,
3451 &rbd_dev->lock_dwork, 0);
3453 __rbd_lock(rbd_dev, cookie);
3457 static void rbd_reregister_watch(struct work_struct *work)
3459 struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
3460 struct rbd_device, watch_dwork);
3463 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3465 mutex_lock(&rbd_dev->watch_mutex);
3466 if (rbd_dev->watch_state != RBD_WATCH_STATE_ERROR) {
3467 mutex_unlock(&rbd_dev->watch_mutex);
3471 ret = __rbd_register_watch(rbd_dev);
3473 rbd_warn(rbd_dev, "failed to reregister watch: %d", ret);
3474 if (ret == -EBLACKLISTED || ret == -ENOENT) {
3475 set_bit(RBD_DEV_FLAG_BLACKLISTED, &rbd_dev->flags);
3476 wake_requests(rbd_dev, true);
3478 queue_delayed_work(rbd_dev->task_wq,
3479 &rbd_dev->watch_dwork,
3482 mutex_unlock(&rbd_dev->watch_mutex);
3486 rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
3487 rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
3488 mutex_unlock(&rbd_dev->watch_mutex);
3490 down_write(&rbd_dev->lock_rwsem);
3491 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED)
3492 rbd_reacquire_lock(rbd_dev);
3493 up_write(&rbd_dev->lock_rwsem);
3495 ret = rbd_dev_refresh(rbd_dev);
3497 rbd_warn(rbd_dev, "reregistration refresh failed: %d", ret);
3501 * Synchronous osd object method call. Returns the number of bytes
3502 * returned in the outbound buffer, or a negative error code.
3504 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
3505 struct ceph_object_id *oid,
3506 struct ceph_object_locator *oloc,
3507 const char *method_name,
3508 const void *outbound,
3509 size_t outbound_size,
3511 size_t inbound_size)
3513 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3514 struct page *req_page = NULL;
3515 struct page *reply_page;
3519 * Method calls are ultimately read operations. The result
3520 * should placed into the inbound buffer provided. They
3521 * also supply outbound data--parameters for the object
3522 * method. Currently if this is present it will be a
3526 if (outbound_size > PAGE_SIZE)
3529 req_page = alloc_page(GFP_KERNEL);
3533 memcpy(page_address(req_page), outbound, outbound_size);
3536 reply_page = alloc_page(GFP_KERNEL);
3539 __free_page(req_page);
3543 ret = ceph_osdc_call(osdc, oid, oloc, RBD_DRV_NAME, method_name,
3544 CEPH_OSD_FLAG_READ, req_page, outbound_size,
3545 reply_page, &inbound_size);
3547 memcpy(inbound, page_address(reply_page), inbound_size);
3552 __free_page(req_page);
3553 __free_page(reply_page);
3558 * lock_rwsem must be held for read
3560 static int rbd_wait_state_locked(struct rbd_device *rbd_dev, bool may_acquire)
3563 unsigned long timeout;
3566 if (test_bit(RBD_DEV_FLAG_BLACKLISTED, &rbd_dev->flags))
3567 return -EBLACKLISTED;
3569 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED)
3573 rbd_warn(rbd_dev, "exclusive lock required");
3579 * Note the use of mod_delayed_work() in rbd_acquire_lock()
3580 * and cancel_delayed_work() in wake_requests().
3582 dout("%s rbd_dev %p queueing lock_dwork\n", __func__, rbd_dev);
3583 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork, 0);
3584 prepare_to_wait_exclusive(&rbd_dev->lock_waitq, &wait,
3585 TASK_UNINTERRUPTIBLE);
3586 up_read(&rbd_dev->lock_rwsem);
3587 timeout = schedule_timeout(ceph_timeout_jiffies(
3588 rbd_dev->opts->lock_timeout));
3589 down_read(&rbd_dev->lock_rwsem);
3590 if (test_bit(RBD_DEV_FLAG_BLACKLISTED, &rbd_dev->flags)) {
3591 ret = -EBLACKLISTED;
3595 rbd_warn(rbd_dev, "timed out waiting for lock");
3599 } while (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED);
3601 finish_wait(&rbd_dev->lock_waitq, &wait);
3605 static void rbd_queue_workfn(struct work_struct *work)
3607 struct request *rq = blk_mq_rq_from_pdu(work);
3608 struct rbd_device *rbd_dev = rq->q->queuedata;
3609 struct rbd_img_request *img_request;
3610 struct ceph_snap_context *snapc = NULL;
3611 u64 offset = (u64)blk_rq_pos(rq) << SECTOR_SHIFT;
3612 u64 length = blk_rq_bytes(rq);
3613 enum obj_operation_type op_type;
3615 bool must_be_locked;
3618 switch (req_op(rq)) {
3619 case REQ_OP_DISCARD:
3620 case REQ_OP_WRITE_ZEROES:
3621 op_type = OBJ_OP_DISCARD;
3624 op_type = OBJ_OP_WRITE;
3627 op_type = OBJ_OP_READ;
3630 dout("%s: non-fs request type %d\n", __func__, req_op(rq));
3635 /* Ignore/skip any zero-length requests */
3638 dout("%s: zero-length request\n", __func__);
3643 rbd_assert(op_type == OBJ_OP_READ ||
3644 rbd_dev->spec->snap_id == CEPH_NOSNAP);
3647 * Quit early if the mapped snapshot no longer exists. It's
3648 * still possible the snapshot will have disappeared by the
3649 * time our request arrives at the osd, but there's no sense in
3650 * sending it if we already know.
3652 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
3653 dout("request for non-existent snapshot");
3654 rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
3659 if (offset && length > U64_MAX - offset + 1) {
3660 rbd_warn(rbd_dev, "bad request range (%llu~%llu)", offset,
3663 goto err_rq; /* Shouldn't happen */
3666 blk_mq_start_request(rq);
3668 down_read(&rbd_dev->header_rwsem);
3669 mapping_size = rbd_dev->mapping.size;
3670 if (op_type != OBJ_OP_READ) {
3671 snapc = rbd_dev->header.snapc;
3672 ceph_get_snap_context(snapc);
3674 up_read(&rbd_dev->header_rwsem);
3676 if (offset + length > mapping_size) {
3677 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)", offset,
3678 length, mapping_size);
3684 (rbd_dev->header.features & RBD_FEATURE_EXCLUSIVE_LOCK) &&
3685 (op_type != OBJ_OP_READ || rbd_dev->opts->lock_on_read);
3686 if (must_be_locked) {
3687 down_read(&rbd_dev->lock_rwsem);
3688 result = rbd_wait_state_locked(rbd_dev,
3689 !rbd_dev->opts->exclusive);
3694 img_request = rbd_img_request_create(rbd_dev, op_type, snapc);
3699 img_request->rq = rq;
3700 snapc = NULL; /* img_request consumes a ref */
3702 if (op_type == OBJ_OP_DISCARD)
3703 result = rbd_img_fill_nodata(img_request, offset, length);
3705 result = rbd_img_fill_from_bio(img_request, offset, length,
3708 goto err_img_request;
3710 rbd_img_request_submit(img_request);
3712 up_read(&rbd_dev->lock_rwsem);
3716 rbd_img_request_put(img_request);
3719 up_read(&rbd_dev->lock_rwsem);
3722 rbd_warn(rbd_dev, "%s %llx at %llx result %d",
3723 obj_op_name(op_type), length, offset, result);
3724 ceph_put_snap_context(snapc);
3726 blk_mq_end_request(rq, errno_to_blk_status(result));
3729 static blk_status_t rbd_queue_rq(struct blk_mq_hw_ctx *hctx,
3730 const struct blk_mq_queue_data *bd)
3732 struct request *rq = bd->rq;
3733 struct work_struct *work = blk_mq_rq_to_pdu(rq);
3735 queue_work(rbd_wq, work);
3739 static void rbd_free_disk(struct rbd_device *rbd_dev)
3741 blk_cleanup_queue(rbd_dev->disk->queue);
3742 blk_mq_free_tag_set(&rbd_dev->tag_set);
3743 put_disk(rbd_dev->disk);
3744 rbd_dev->disk = NULL;
3747 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
3748 struct ceph_object_id *oid,
3749 struct ceph_object_locator *oloc,
3750 void *buf, int buf_len)
3753 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3754 struct ceph_osd_request *req;
3755 struct page **pages;
3756 int num_pages = calc_pages_for(0, buf_len);
3759 req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
3763 ceph_oid_copy(&req->r_base_oid, oid);
3764 ceph_oloc_copy(&req->r_base_oloc, oloc);
3765 req->r_flags = CEPH_OSD_FLAG_READ;
3767 ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
3771 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
3772 if (IS_ERR(pages)) {
3773 ret = PTR_ERR(pages);
3777 osd_req_op_extent_init(req, 0, CEPH_OSD_OP_READ, 0, buf_len, 0, 0);
3778 osd_req_op_extent_osd_data_pages(req, 0, pages, buf_len, 0, false,
3781 ceph_osdc_start_request(osdc, req, false);
3782 ret = ceph_osdc_wait_request(osdc, req);
3784 ceph_copy_from_page_vector(pages, buf, 0, ret);
3787 ceph_osdc_put_request(req);
3792 * Read the complete header for the given rbd device. On successful
3793 * return, the rbd_dev->header field will contain up-to-date
3794 * information about the image.
3796 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev)
3798 struct rbd_image_header_ondisk *ondisk = NULL;
3805 * The complete header will include an array of its 64-bit
3806 * snapshot ids, followed by the names of those snapshots as
3807 * a contiguous block of NUL-terminated strings. Note that
3808 * the number of snapshots could change by the time we read
3809 * it in, in which case we re-read it.
3816 size = sizeof (*ondisk);
3817 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
3819 ondisk = kmalloc(size, GFP_KERNEL);
3823 ret = rbd_obj_read_sync(rbd_dev, &rbd_dev->header_oid,
3824 &rbd_dev->header_oloc, ondisk, size);
3827 if ((size_t)ret < size) {
3829 rbd_warn(rbd_dev, "short header read (want %zd got %d)",
3833 if (!rbd_dev_ondisk_valid(ondisk)) {
3835 rbd_warn(rbd_dev, "invalid header");
3839 names_size = le64_to_cpu(ondisk->snap_names_len);
3840 want_count = snap_count;
3841 snap_count = le32_to_cpu(ondisk->snap_count);
3842 } while (snap_count != want_count);
3844 ret = rbd_header_from_disk(rbd_dev, ondisk);
3852 * Clear the rbd device's EXISTS flag if the snapshot it's mapped to
3853 * has disappeared from the (just updated) snapshot context.
3855 static void rbd_exists_validate(struct rbd_device *rbd_dev)
3859 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags))
3862 snap_id = rbd_dev->spec->snap_id;
3863 if (snap_id == CEPH_NOSNAP)
3866 if (rbd_dev_snap_index(rbd_dev, snap_id) == BAD_SNAP_INDEX)
3867 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
3870 static void rbd_dev_update_size(struct rbd_device *rbd_dev)
3875 * If EXISTS is not set, rbd_dev->disk may be NULL, so don't
3876 * try to update its size. If REMOVING is set, updating size
3877 * is just useless work since the device can't be opened.
3879 if (test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags) &&
3880 !test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags)) {
3881 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
3882 dout("setting size to %llu sectors", (unsigned long long)size);
3883 set_capacity(rbd_dev->disk, size);
3884 revalidate_disk(rbd_dev->disk);
3888 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
3893 down_write(&rbd_dev->header_rwsem);
3894 mapping_size = rbd_dev->mapping.size;
3896 ret = rbd_dev_header_info(rbd_dev);
3901 * If there is a parent, see if it has disappeared due to the
3902 * mapped image getting flattened.
3904 if (rbd_dev->parent) {
3905 ret = rbd_dev_v2_parent_info(rbd_dev);
3910 if (rbd_dev->spec->snap_id == CEPH_NOSNAP) {
3911 rbd_dev->mapping.size = rbd_dev->header.image_size;
3913 /* validate mapped snapshot's EXISTS flag */
3914 rbd_exists_validate(rbd_dev);
3918 up_write(&rbd_dev->header_rwsem);
3919 if (!ret && mapping_size != rbd_dev->mapping.size)
3920 rbd_dev_update_size(rbd_dev);
3925 static int rbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
3926 unsigned int hctx_idx, unsigned int numa_node)
3928 struct work_struct *work = blk_mq_rq_to_pdu(rq);
3930 INIT_WORK(work, rbd_queue_workfn);
3934 static const struct blk_mq_ops rbd_mq_ops = {
3935 .queue_rq = rbd_queue_rq,
3936 .init_request = rbd_init_request,
3939 static int rbd_init_disk(struct rbd_device *rbd_dev)
3941 struct gendisk *disk;
3942 struct request_queue *q;
3943 unsigned int objset_bytes =
3944 rbd_dev->layout.object_size * rbd_dev->layout.stripe_count;
3947 /* create gendisk info */
3948 disk = alloc_disk(single_major ?
3949 (1 << RBD_SINGLE_MAJOR_PART_SHIFT) :
3950 RBD_MINORS_PER_MAJOR);
3954 snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
3956 disk->major = rbd_dev->major;
3957 disk->first_minor = rbd_dev->minor;
3959 disk->flags |= GENHD_FL_EXT_DEVT;
3960 disk->fops = &rbd_bd_ops;
3961 disk->private_data = rbd_dev;
3963 memset(&rbd_dev->tag_set, 0, sizeof(rbd_dev->tag_set));
3964 rbd_dev->tag_set.ops = &rbd_mq_ops;
3965 rbd_dev->tag_set.queue_depth = rbd_dev->opts->queue_depth;
3966 rbd_dev->tag_set.numa_node = NUMA_NO_NODE;
3967 rbd_dev->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
3968 rbd_dev->tag_set.nr_hw_queues = 1;
3969 rbd_dev->tag_set.cmd_size = sizeof(struct work_struct);
3971 err = blk_mq_alloc_tag_set(&rbd_dev->tag_set);
3975 q = blk_mq_init_queue(&rbd_dev->tag_set);
3981 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
3982 /* QUEUE_FLAG_ADD_RANDOM is off by default for blk-mq */
3984 blk_queue_max_hw_sectors(q, objset_bytes >> SECTOR_SHIFT);
3985 q->limits.max_sectors = queue_max_hw_sectors(q);
3986 blk_queue_max_segments(q, USHRT_MAX);
3987 blk_queue_max_segment_size(q, UINT_MAX);
3988 blk_queue_io_min(q, objset_bytes);
3989 blk_queue_io_opt(q, objset_bytes);
3991 if (rbd_dev->opts->trim) {
3992 blk_queue_flag_set(QUEUE_FLAG_DISCARD, q);
3993 q->limits.discard_granularity = objset_bytes;
3994 blk_queue_max_discard_sectors(q, objset_bytes >> SECTOR_SHIFT);
3995 blk_queue_max_write_zeroes_sectors(q, objset_bytes >> SECTOR_SHIFT);
3998 if (!ceph_test_opt(rbd_dev->rbd_client->client, NOCRC))
3999 q->backing_dev_info->capabilities |= BDI_CAP_STABLE_WRITES;
4002 * disk_release() expects a queue ref from add_disk() and will
4003 * put it. Hold an extra ref until add_disk() is called.
4005 WARN_ON(!blk_get_queue(q));
4007 q->queuedata = rbd_dev;
4009 rbd_dev->disk = disk;
4013 blk_mq_free_tag_set(&rbd_dev->tag_set);
4023 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
4025 return container_of(dev, struct rbd_device, dev);
4028 static ssize_t rbd_size_show(struct device *dev,
4029 struct device_attribute *attr, char *buf)
4031 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4033 return sprintf(buf, "%llu\n",
4034 (unsigned long long)rbd_dev->mapping.size);
4038 * Note this shows the features for whatever's mapped, which is not
4039 * necessarily the base image.
4041 static ssize_t rbd_features_show(struct device *dev,
4042 struct device_attribute *attr, char *buf)
4044 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4046 return sprintf(buf, "0x%016llx\n",
4047 (unsigned long long)rbd_dev->mapping.features);
4050 static ssize_t rbd_major_show(struct device *dev,
4051 struct device_attribute *attr, char *buf)
4053 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4056 return sprintf(buf, "%d\n", rbd_dev->major);
4058 return sprintf(buf, "(none)\n");
4061 static ssize_t rbd_minor_show(struct device *dev,
4062 struct device_attribute *attr, char *buf)
4064 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4066 return sprintf(buf, "%d\n", rbd_dev->minor);
4069 static ssize_t rbd_client_addr_show(struct device *dev,
4070 struct device_attribute *attr, char *buf)
4072 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4073 struct ceph_entity_addr *client_addr =
4074 ceph_client_addr(rbd_dev->rbd_client->client);
4076 return sprintf(buf, "%pISpc/%u\n", &client_addr->in_addr,
4077 le32_to_cpu(client_addr->nonce));
4080 static ssize_t rbd_client_id_show(struct device *dev,
4081 struct device_attribute *attr, char *buf)
4083 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4085 return sprintf(buf, "client%lld\n",
4086 ceph_client_gid(rbd_dev->rbd_client->client));
4089 static ssize_t rbd_cluster_fsid_show(struct device *dev,
4090 struct device_attribute *attr, char *buf)
4092 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4094 return sprintf(buf, "%pU\n", &rbd_dev->rbd_client->client->fsid);
4097 static ssize_t rbd_config_info_show(struct device *dev,
4098 struct device_attribute *attr, char *buf)
4100 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4102 return sprintf(buf, "%s\n", rbd_dev->config_info);
4105 static ssize_t rbd_pool_show(struct device *dev,
4106 struct device_attribute *attr, char *buf)
4108 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4110 return sprintf(buf, "%s\n", rbd_dev->spec->pool_name);
4113 static ssize_t rbd_pool_id_show(struct device *dev,
4114 struct device_attribute *attr, char *buf)
4116 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4118 return sprintf(buf, "%llu\n",
4119 (unsigned long long) rbd_dev->spec->pool_id);
4122 static ssize_t rbd_name_show(struct device *dev,
4123 struct device_attribute *attr, char *buf)
4125 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4127 if (rbd_dev->spec->image_name)
4128 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
4130 return sprintf(buf, "(unknown)\n");
4133 static ssize_t rbd_image_id_show(struct device *dev,
4134 struct device_attribute *attr, char *buf)
4136 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4138 return sprintf(buf, "%s\n", rbd_dev->spec->image_id);
4142 * Shows the name of the currently-mapped snapshot (or
4143 * RBD_SNAP_HEAD_NAME for the base image).
4145 static ssize_t rbd_snap_show(struct device *dev,
4146 struct device_attribute *attr,
4149 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4151 return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
4154 static ssize_t rbd_snap_id_show(struct device *dev,
4155 struct device_attribute *attr, char *buf)
4157 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4159 return sprintf(buf, "%llu\n", rbd_dev->spec->snap_id);
4163 * For a v2 image, shows the chain of parent images, separated by empty
4164 * lines. For v1 images or if there is no parent, shows "(no parent
4167 static ssize_t rbd_parent_show(struct device *dev,
4168 struct device_attribute *attr,
4171 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4174 if (!rbd_dev->parent)
4175 return sprintf(buf, "(no parent image)\n");
4177 for ( ; rbd_dev->parent; rbd_dev = rbd_dev->parent) {
4178 struct rbd_spec *spec = rbd_dev->parent_spec;
4180 count += sprintf(&buf[count], "%s"
4181 "pool_id %llu\npool_name %s\n"
4182 "image_id %s\nimage_name %s\n"
4183 "snap_id %llu\nsnap_name %s\n"
4185 !count ? "" : "\n", /* first? */
4186 spec->pool_id, spec->pool_name,
4187 spec->image_id, spec->image_name ?: "(unknown)",
4188 spec->snap_id, spec->snap_name,
4189 rbd_dev->parent_overlap);
4195 static ssize_t rbd_image_refresh(struct device *dev,
4196 struct device_attribute *attr,
4200 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4203 ret = rbd_dev_refresh(rbd_dev);
4210 static DEVICE_ATTR(size, 0444, rbd_size_show, NULL);
4211 static DEVICE_ATTR(features, 0444, rbd_features_show, NULL);
4212 static DEVICE_ATTR(major, 0444, rbd_major_show, NULL);
4213 static DEVICE_ATTR(minor, 0444, rbd_minor_show, NULL);
4214 static DEVICE_ATTR(client_addr, 0444, rbd_client_addr_show, NULL);
4215 static DEVICE_ATTR(client_id, 0444, rbd_client_id_show, NULL);
4216 static DEVICE_ATTR(cluster_fsid, 0444, rbd_cluster_fsid_show, NULL);
4217 static DEVICE_ATTR(config_info, 0400, rbd_config_info_show, NULL);
4218 static DEVICE_ATTR(pool, 0444, rbd_pool_show, NULL);
4219 static DEVICE_ATTR(pool_id, 0444, rbd_pool_id_show, NULL);
4220 static DEVICE_ATTR(name, 0444, rbd_name_show, NULL);
4221 static DEVICE_ATTR(image_id, 0444, rbd_image_id_show, NULL);
4222 static DEVICE_ATTR(refresh, 0200, NULL, rbd_image_refresh);
4223 static DEVICE_ATTR(current_snap, 0444, rbd_snap_show, NULL);
4224 static DEVICE_ATTR(snap_id, 0444, rbd_snap_id_show, NULL);
4225 static DEVICE_ATTR(parent, 0444, rbd_parent_show, NULL);
4227 static struct attribute *rbd_attrs[] = {
4228 &dev_attr_size.attr,
4229 &dev_attr_features.attr,
4230 &dev_attr_major.attr,
4231 &dev_attr_minor.attr,
4232 &dev_attr_client_addr.attr,
4233 &dev_attr_client_id.attr,
4234 &dev_attr_cluster_fsid.attr,
4235 &dev_attr_config_info.attr,
4236 &dev_attr_pool.attr,
4237 &dev_attr_pool_id.attr,
4238 &dev_attr_name.attr,
4239 &dev_attr_image_id.attr,
4240 &dev_attr_current_snap.attr,
4241 &dev_attr_snap_id.attr,
4242 &dev_attr_parent.attr,
4243 &dev_attr_refresh.attr,
4247 static struct attribute_group rbd_attr_group = {
4251 static const struct attribute_group *rbd_attr_groups[] = {
4256 static void rbd_dev_release(struct device *dev);
4258 static const struct device_type rbd_device_type = {
4260 .groups = rbd_attr_groups,
4261 .release = rbd_dev_release,
4264 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
4266 kref_get(&spec->kref);
4271 static void rbd_spec_free(struct kref *kref);
4272 static void rbd_spec_put(struct rbd_spec *spec)
4275 kref_put(&spec->kref, rbd_spec_free);
4278 static struct rbd_spec *rbd_spec_alloc(void)
4280 struct rbd_spec *spec;
4282 spec = kzalloc(sizeof (*spec), GFP_KERNEL);
4286 spec->pool_id = CEPH_NOPOOL;
4287 spec->snap_id = CEPH_NOSNAP;
4288 kref_init(&spec->kref);
4293 static void rbd_spec_free(struct kref *kref)
4295 struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
4297 kfree(spec->pool_name);
4298 kfree(spec->image_id);
4299 kfree(spec->image_name);
4300 kfree(spec->snap_name);
4304 static void rbd_dev_free(struct rbd_device *rbd_dev)
4306 WARN_ON(rbd_dev->watch_state != RBD_WATCH_STATE_UNREGISTERED);
4307 WARN_ON(rbd_dev->lock_state != RBD_LOCK_STATE_UNLOCKED);
4309 ceph_oid_destroy(&rbd_dev->header_oid);
4310 ceph_oloc_destroy(&rbd_dev->header_oloc);
4311 kfree(rbd_dev->config_info);
4313 rbd_put_client(rbd_dev->rbd_client);
4314 rbd_spec_put(rbd_dev->spec);
4315 kfree(rbd_dev->opts);
4319 static void rbd_dev_release(struct device *dev)
4321 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4322 bool need_put = !!rbd_dev->opts;
4325 destroy_workqueue(rbd_dev->task_wq);
4326 ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4329 rbd_dev_free(rbd_dev);
4332 * This is racy, but way better than putting module outside of
4333 * the release callback. The race window is pretty small, so
4334 * doing something similar to dm (dm-builtin.c) is overkill.
4337 module_put(THIS_MODULE);
4340 static struct rbd_device *__rbd_dev_create(struct rbd_client *rbdc,
4341 struct rbd_spec *spec)
4343 struct rbd_device *rbd_dev;
4345 rbd_dev = kzalloc(sizeof(*rbd_dev), GFP_KERNEL);
4349 spin_lock_init(&rbd_dev->lock);
4350 INIT_LIST_HEAD(&rbd_dev->node);
4351 init_rwsem(&rbd_dev->header_rwsem);
4353 rbd_dev->header.data_pool_id = CEPH_NOPOOL;
4354 ceph_oid_init(&rbd_dev->header_oid);
4355 rbd_dev->header_oloc.pool = spec->pool_id;
4357 mutex_init(&rbd_dev->watch_mutex);
4358 rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
4359 INIT_DELAYED_WORK(&rbd_dev->watch_dwork, rbd_reregister_watch);
4361 init_rwsem(&rbd_dev->lock_rwsem);
4362 rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
4363 INIT_WORK(&rbd_dev->acquired_lock_work, rbd_notify_acquired_lock);
4364 INIT_WORK(&rbd_dev->released_lock_work, rbd_notify_released_lock);
4365 INIT_DELAYED_WORK(&rbd_dev->lock_dwork, rbd_acquire_lock);
4366 INIT_WORK(&rbd_dev->unlock_work, rbd_release_lock_work);
4367 init_waitqueue_head(&rbd_dev->lock_waitq);
4369 rbd_dev->dev.bus = &rbd_bus_type;
4370 rbd_dev->dev.type = &rbd_device_type;
4371 rbd_dev->dev.parent = &rbd_root_dev;
4372 device_initialize(&rbd_dev->dev);
4374 rbd_dev->rbd_client = rbdc;
4375 rbd_dev->spec = spec;
4381 * Create a mapping rbd_dev.
4383 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
4384 struct rbd_spec *spec,
4385 struct rbd_options *opts)
4387 struct rbd_device *rbd_dev;
4389 rbd_dev = __rbd_dev_create(rbdc, spec);
4393 rbd_dev->opts = opts;
4395 /* get an id and fill in device name */
4396 rbd_dev->dev_id = ida_simple_get(&rbd_dev_id_ida, 0,
4397 minor_to_rbd_dev_id(1 << MINORBITS),
4399 if (rbd_dev->dev_id < 0)
4402 sprintf(rbd_dev->name, RBD_DRV_NAME "%d", rbd_dev->dev_id);
4403 rbd_dev->task_wq = alloc_ordered_workqueue("%s-tasks", WQ_MEM_RECLAIM,
4405 if (!rbd_dev->task_wq)
4408 /* we have a ref from do_rbd_add() */
4409 __module_get(THIS_MODULE);
4411 dout("%s rbd_dev %p dev_id %d\n", __func__, rbd_dev, rbd_dev->dev_id);
4415 ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4417 rbd_dev_free(rbd_dev);
4421 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
4424 put_device(&rbd_dev->dev);
4428 * Get the size and object order for an image snapshot, or if
4429 * snap_id is CEPH_NOSNAP, gets this information for the base
4432 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
4433 u8 *order, u64 *snap_size)
4435 __le64 snapid = cpu_to_le64(snap_id);
4440 } __attribute__ ((packed)) size_buf = { 0 };
4442 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4443 &rbd_dev->header_oloc, "get_size",
4444 &snapid, sizeof(snapid),
4445 &size_buf, sizeof(size_buf));
4446 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4449 if (ret < sizeof (size_buf))
4453 *order = size_buf.order;
4454 dout(" order %u", (unsigned int)*order);
4456 *snap_size = le64_to_cpu(size_buf.size);
4458 dout(" snap_id 0x%016llx snap_size = %llu\n",
4459 (unsigned long long)snap_id,
4460 (unsigned long long)*snap_size);
4465 static int rbd_dev_v2_image_size(struct rbd_device *rbd_dev)
4467 return _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
4468 &rbd_dev->header.obj_order,
4469 &rbd_dev->header.image_size);
4472 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev)
4478 reply_buf = kzalloc(RBD_OBJ_PREFIX_LEN_MAX, GFP_KERNEL);
4482 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4483 &rbd_dev->header_oloc, "get_object_prefix",
4484 NULL, 0, reply_buf, RBD_OBJ_PREFIX_LEN_MAX);
4485 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4490 rbd_dev->header.object_prefix = ceph_extract_encoded_string(&p,
4491 p + ret, NULL, GFP_NOIO);
4494 if (IS_ERR(rbd_dev->header.object_prefix)) {
4495 ret = PTR_ERR(rbd_dev->header.object_prefix);
4496 rbd_dev->header.object_prefix = NULL;
4498 dout(" object_prefix = %s\n", rbd_dev->header.object_prefix);
4506 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
4509 __le64 snapid = cpu_to_le64(snap_id);
4513 } __attribute__ ((packed)) features_buf = { 0 };
4517 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4518 &rbd_dev->header_oloc, "get_features",
4519 &snapid, sizeof(snapid),
4520 &features_buf, sizeof(features_buf));
4521 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4524 if (ret < sizeof (features_buf))
4527 unsup = le64_to_cpu(features_buf.incompat) & ~RBD_FEATURES_SUPPORTED;
4529 rbd_warn(rbd_dev, "image uses unsupported features: 0x%llx",
4534 *snap_features = le64_to_cpu(features_buf.features);
4536 dout(" snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
4537 (unsigned long long)snap_id,
4538 (unsigned long long)*snap_features,
4539 (unsigned long long)le64_to_cpu(features_buf.incompat));
4544 static int rbd_dev_v2_features(struct rbd_device *rbd_dev)
4546 return _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
4547 &rbd_dev->header.features);
4550 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev)
4552 struct rbd_spec *parent_spec;
4554 void *reply_buf = NULL;
4564 parent_spec = rbd_spec_alloc();
4568 size = sizeof (__le64) + /* pool_id */
4569 sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX + /* image_id */
4570 sizeof (__le64) + /* snap_id */
4571 sizeof (__le64); /* overlap */
4572 reply_buf = kmalloc(size, GFP_KERNEL);
4578 snapid = cpu_to_le64(rbd_dev->spec->snap_id);
4579 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4580 &rbd_dev->header_oloc, "get_parent",
4581 &snapid, sizeof(snapid), reply_buf, size);
4582 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4587 end = reply_buf + ret;
4589 ceph_decode_64_safe(&p, end, pool_id, out_err);
4590 if (pool_id == CEPH_NOPOOL) {
4592 * Either the parent never existed, or we have
4593 * record of it but the image got flattened so it no
4594 * longer has a parent. When the parent of a
4595 * layered image disappears we immediately set the
4596 * overlap to 0. The effect of this is that all new
4597 * requests will be treated as if the image had no
4600 if (rbd_dev->parent_overlap) {
4601 rbd_dev->parent_overlap = 0;
4602 rbd_dev_parent_put(rbd_dev);
4603 pr_info("%s: clone image has been flattened\n",
4604 rbd_dev->disk->disk_name);
4607 goto out; /* No parent? No problem. */
4610 /* The ceph file layout needs to fit pool id in 32 bits */
4613 if (pool_id > (u64)U32_MAX) {
4614 rbd_warn(NULL, "parent pool id too large (%llu > %u)",
4615 (unsigned long long)pool_id, U32_MAX);
4619 image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
4620 if (IS_ERR(image_id)) {
4621 ret = PTR_ERR(image_id);
4624 ceph_decode_64_safe(&p, end, snap_id, out_err);
4625 ceph_decode_64_safe(&p, end, overlap, out_err);
4628 * The parent won't change (except when the clone is
4629 * flattened, already handled that). So we only need to
4630 * record the parent spec we have not already done so.
4632 if (!rbd_dev->parent_spec) {
4633 parent_spec->pool_id = pool_id;
4634 parent_spec->image_id = image_id;
4635 parent_spec->snap_id = snap_id;
4636 rbd_dev->parent_spec = parent_spec;
4637 parent_spec = NULL; /* rbd_dev now owns this */
4643 * We always update the parent overlap. If it's zero we issue
4644 * a warning, as we will proceed as if there was no parent.
4648 /* refresh, careful to warn just once */
4649 if (rbd_dev->parent_overlap)
4651 "clone now standalone (overlap became 0)");
4654 rbd_warn(rbd_dev, "clone is standalone (overlap 0)");
4657 rbd_dev->parent_overlap = overlap;
4663 rbd_spec_put(parent_spec);
4668 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev)
4672 __le64 stripe_count;
4673 } __attribute__ ((packed)) striping_info_buf = { 0 };
4674 size_t size = sizeof (striping_info_buf);
4678 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4679 &rbd_dev->header_oloc, "get_stripe_unit_count",
4680 NULL, 0, &striping_info_buf, size);
4681 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4687 p = &striping_info_buf;
4688 rbd_dev->header.stripe_unit = ceph_decode_64(&p);
4689 rbd_dev->header.stripe_count = ceph_decode_64(&p);
4693 static int rbd_dev_v2_data_pool(struct rbd_device *rbd_dev)
4695 __le64 data_pool_id;
4698 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4699 &rbd_dev->header_oloc, "get_data_pool",
4700 NULL, 0, &data_pool_id, sizeof(data_pool_id));
4703 if (ret < sizeof(data_pool_id))
4706 rbd_dev->header.data_pool_id = le64_to_cpu(data_pool_id);
4707 WARN_ON(rbd_dev->header.data_pool_id == CEPH_NOPOOL);
4711 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
4713 CEPH_DEFINE_OID_ONSTACK(oid);
4714 size_t image_id_size;
4719 void *reply_buf = NULL;
4721 char *image_name = NULL;
4724 rbd_assert(!rbd_dev->spec->image_name);
4726 len = strlen(rbd_dev->spec->image_id);
4727 image_id_size = sizeof (__le32) + len;
4728 image_id = kmalloc(image_id_size, GFP_KERNEL);
4733 end = image_id + image_id_size;
4734 ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
4736 size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
4737 reply_buf = kmalloc(size, GFP_KERNEL);
4741 ceph_oid_printf(&oid, "%s", RBD_DIRECTORY);
4742 ret = rbd_obj_method_sync(rbd_dev, &oid, &rbd_dev->header_oloc,
4743 "dir_get_name", image_id, image_id_size,
4748 end = reply_buf + ret;
4750 image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
4751 if (IS_ERR(image_name))
4754 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
4762 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4764 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4765 const char *snap_name;
4768 /* Skip over names until we find the one we are looking for */
4770 snap_name = rbd_dev->header.snap_names;
4771 while (which < snapc->num_snaps) {
4772 if (!strcmp(name, snap_name))
4773 return snapc->snaps[which];
4774 snap_name += strlen(snap_name) + 1;
4780 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4782 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4787 for (which = 0; !found && which < snapc->num_snaps; which++) {
4788 const char *snap_name;
4790 snap_id = snapc->snaps[which];
4791 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
4792 if (IS_ERR(snap_name)) {
4793 /* ignore no-longer existing snapshots */
4794 if (PTR_ERR(snap_name) == -ENOENT)
4799 found = !strcmp(name, snap_name);
4802 return found ? snap_id : CEPH_NOSNAP;
4806 * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
4807 * no snapshot by that name is found, or if an error occurs.
4809 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4811 if (rbd_dev->image_format == 1)
4812 return rbd_v1_snap_id_by_name(rbd_dev, name);
4814 return rbd_v2_snap_id_by_name(rbd_dev, name);
4818 * An image being mapped will have everything but the snap id.
4820 static int rbd_spec_fill_snap_id(struct rbd_device *rbd_dev)
4822 struct rbd_spec *spec = rbd_dev->spec;
4824 rbd_assert(spec->pool_id != CEPH_NOPOOL && spec->pool_name);
4825 rbd_assert(spec->image_id && spec->image_name);
4826 rbd_assert(spec->snap_name);
4828 if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
4831 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
4832 if (snap_id == CEPH_NOSNAP)
4835 spec->snap_id = snap_id;
4837 spec->snap_id = CEPH_NOSNAP;
4844 * A parent image will have all ids but none of the names.
4846 * All names in an rbd spec are dynamically allocated. It's OK if we
4847 * can't figure out the name for an image id.
4849 static int rbd_spec_fill_names(struct rbd_device *rbd_dev)
4851 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4852 struct rbd_spec *spec = rbd_dev->spec;
4853 const char *pool_name;
4854 const char *image_name;
4855 const char *snap_name;
4858 rbd_assert(spec->pool_id != CEPH_NOPOOL);
4859 rbd_assert(spec->image_id);
4860 rbd_assert(spec->snap_id != CEPH_NOSNAP);
4862 /* Get the pool name; we have to make our own copy of this */
4864 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
4866 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
4869 pool_name = kstrdup(pool_name, GFP_KERNEL);
4873 /* Fetch the image name; tolerate failure here */
4875 image_name = rbd_dev_image_name(rbd_dev);
4877 rbd_warn(rbd_dev, "unable to get image name");
4879 /* Fetch the snapshot name */
4881 snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
4882 if (IS_ERR(snap_name)) {
4883 ret = PTR_ERR(snap_name);
4887 spec->pool_name = pool_name;
4888 spec->image_name = image_name;
4889 spec->snap_name = snap_name;
4899 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev)
4908 struct ceph_snap_context *snapc;
4912 * We'll need room for the seq value (maximum snapshot id),
4913 * snapshot count, and array of that many snapshot ids.
4914 * For now we have a fixed upper limit on the number we're
4915 * prepared to receive.
4917 size = sizeof (__le64) + sizeof (__le32) +
4918 RBD_MAX_SNAP_COUNT * sizeof (__le64);
4919 reply_buf = kzalloc(size, GFP_KERNEL);
4923 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4924 &rbd_dev->header_oloc, "get_snapcontext",
4925 NULL, 0, reply_buf, size);
4926 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4931 end = reply_buf + ret;
4933 ceph_decode_64_safe(&p, end, seq, out);
4934 ceph_decode_32_safe(&p, end, snap_count, out);
4937 * Make sure the reported number of snapshot ids wouldn't go
4938 * beyond the end of our buffer. But before checking that,
4939 * make sure the computed size of the snapshot context we
4940 * allocate is representable in a size_t.
4942 if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
4947 if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
4951 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
4957 for (i = 0; i < snap_count; i++)
4958 snapc->snaps[i] = ceph_decode_64(&p);
4960 ceph_put_snap_context(rbd_dev->header.snapc);
4961 rbd_dev->header.snapc = snapc;
4963 dout(" snap context seq = %llu, snap_count = %u\n",
4964 (unsigned long long)seq, (unsigned int)snap_count);
4971 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
4982 size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
4983 reply_buf = kmalloc(size, GFP_KERNEL);
4985 return ERR_PTR(-ENOMEM);
4987 snapid = cpu_to_le64(snap_id);
4988 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4989 &rbd_dev->header_oloc, "get_snapshot_name",
4990 &snapid, sizeof(snapid), reply_buf, size);
4991 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4993 snap_name = ERR_PTR(ret);
4998 end = reply_buf + ret;
4999 snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
5000 if (IS_ERR(snap_name))
5003 dout(" snap_id 0x%016llx snap_name = %s\n",
5004 (unsigned long long)snap_id, snap_name);
5011 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev)
5013 bool first_time = rbd_dev->header.object_prefix == NULL;
5016 ret = rbd_dev_v2_image_size(rbd_dev);
5021 ret = rbd_dev_v2_header_onetime(rbd_dev);
5026 ret = rbd_dev_v2_snap_context(rbd_dev);
5027 if (ret && first_time) {
5028 kfree(rbd_dev->header.object_prefix);
5029 rbd_dev->header.object_prefix = NULL;
5035 static int rbd_dev_header_info(struct rbd_device *rbd_dev)
5037 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
5039 if (rbd_dev->image_format == 1)
5040 return rbd_dev_v1_header_info(rbd_dev);
5042 return rbd_dev_v2_header_info(rbd_dev);
5046 * Skips over white space at *buf, and updates *buf to point to the
5047 * first found non-space character (if any). Returns the length of
5048 * the token (string of non-white space characters) found. Note
5049 * that *buf must be terminated with '\0'.
5051 static inline size_t next_token(const char **buf)
5054 * These are the characters that produce nonzero for
5055 * isspace() in the "C" and "POSIX" locales.
5057 const char *spaces = " \f\n\r\t\v";
5059 *buf += strspn(*buf, spaces); /* Find start of token */
5061 return strcspn(*buf, spaces); /* Return token length */
5065 * Finds the next token in *buf, dynamically allocates a buffer big
5066 * enough to hold a copy of it, and copies the token into the new
5067 * buffer. The copy is guaranteed to be terminated with '\0'. Note
5068 * that a duplicate buffer is created even for a zero-length token.
5070 * Returns a pointer to the newly-allocated duplicate, or a null
5071 * pointer if memory for the duplicate was not available. If
5072 * the lenp argument is a non-null pointer, the length of the token
5073 * (not including the '\0') is returned in *lenp.
5075 * If successful, the *buf pointer will be updated to point beyond
5076 * the end of the found token.
5078 * Note: uses GFP_KERNEL for allocation.
5080 static inline char *dup_token(const char **buf, size_t *lenp)
5085 len = next_token(buf);
5086 dup = kmemdup(*buf, len + 1, GFP_KERNEL);
5089 *(dup + len) = '\0';
5099 * Parse the options provided for an "rbd add" (i.e., rbd image
5100 * mapping) request. These arrive via a write to /sys/bus/rbd/add,
5101 * and the data written is passed here via a NUL-terminated buffer.
5102 * Returns 0 if successful or an error code otherwise.
5104 * The information extracted from these options is recorded in
5105 * the other parameters which return dynamically-allocated
5108 * The address of a pointer that will refer to a ceph options
5109 * structure. Caller must release the returned pointer using
5110 * ceph_destroy_options() when it is no longer needed.
5112 * Address of an rbd options pointer. Fully initialized by
5113 * this function; caller must release with kfree().
5115 * Address of an rbd image specification pointer. Fully
5116 * initialized by this function based on parsed options.
5117 * Caller must release with rbd_spec_put().
5119 * The options passed take this form:
5120 * <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
5123 * A comma-separated list of one or more monitor addresses.
5124 * A monitor address is an ip address, optionally followed
5125 * by a port number (separated by a colon).
5126 * I.e.: ip1[:port1][,ip2[:port2]...]
5128 * A comma-separated list of ceph and/or rbd options.
5130 * The name of the rados pool containing the rbd image.
5132 * The name of the image in that pool to map.
5134 * An optional snapshot id. If provided, the mapping will
5135 * present data from the image at the time that snapshot was
5136 * created. The image head is used if no snapshot id is
5137 * provided. Snapshot mappings are always read-only.
5139 static int rbd_add_parse_args(const char *buf,
5140 struct ceph_options **ceph_opts,
5141 struct rbd_options **opts,
5142 struct rbd_spec **rbd_spec)
5146 const char *mon_addrs;
5148 size_t mon_addrs_size;
5149 struct rbd_spec *spec = NULL;
5150 struct rbd_options *rbd_opts = NULL;
5151 struct ceph_options *copts;
5154 /* The first four tokens are required */
5156 len = next_token(&buf);
5158 rbd_warn(NULL, "no monitor address(es) provided");
5162 mon_addrs_size = len + 1;
5166 options = dup_token(&buf, NULL);
5170 rbd_warn(NULL, "no options provided");
5174 spec = rbd_spec_alloc();
5178 spec->pool_name = dup_token(&buf, NULL);
5179 if (!spec->pool_name)
5181 if (!*spec->pool_name) {
5182 rbd_warn(NULL, "no pool name provided");
5186 spec->image_name = dup_token(&buf, NULL);
5187 if (!spec->image_name)
5189 if (!*spec->image_name) {
5190 rbd_warn(NULL, "no image name provided");
5195 * Snapshot name is optional; default is to use "-"
5196 * (indicating the head/no snapshot).
5198 len = next_token(&buf);
5200 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
5201 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
5202 } else if (len > RBD_MAX_SNAP_NAME_LEN) {
5203 ret = -ENAMETOOLONG;
5206 snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
5209 *(snap_name + len) = '\0';
5210 spec->snap_name = snap_name;
5212 /* Initialize all rbd options to the defaults */
5214 rbd_opts = kzalloc(sizeof (*rbd_opts), GFP_KERNEL);
5218 rbd_opts->read_only = RBD_READ_ONLY_DEFAULT;
5219 rbd_opts->queue_depth = RBD_QUEUE_DEPTH_DEFAULT;
5220 rbd_opts->lock_timeout = RBD_LOCK_TIMEOUT_DEFAULT;
5221 rbd_opts->lock_on_read = RBD_LOCK_ON_READ_DEFAULT;
5222 rbd_opts->exclusive = RBD_EXCLUSIVE_DEFAULT;
5223 rbd_opts->trim = RBD_TRIM_DEFAULT;
5225 copts = ceph_parse_options(options, mon_addrs,
5226 mon_addrs + mon_addrs_size - 1,
5227 parse_rbd_opts_token, rbd_opts);
5228 if (IS_ERR(copts)) {
5229 ret = PTR_ERR(copts);
5249 static void rbd_dev_image_unlock(struct rbd_device *rbd_dev)
5251 down_write(&rbd_dev->lock_rwsem);
5252 if (__rbd_is_lock_owner(rbd_dev))
5253 rbd_unlock(rbd_dev);
5254 up_write(&rbd_dev->lock_rwsem);
5257 static int rbd_add_acquire_lock(struct rbd_device *rbd_dev)
5261 if (!(rbd_dev->header.features & RBD_FEATURE_EXCLUSIVE_LOCK)) {
5262 rbd_warn(rbd_dev, "exclusive-lock feature is not enabled");
5266 /* FIXME: "rbd map --exclusive" should be in interruptible */
5267 down_read(&rbd_dev->lock_rwsem);
5268 ret = rbd_wait_state_locked(rbd_dev, true);
5269 up_read(&rbd_dev->lock_rwsem);
5271 rbd_warn(rbd_dev, "failed to acquire exclusive lock");
5279 * An rbd format 2 image has a unique identifier, distinct from the
5280 * name given to it by the user. Internally, that identifier is
5281 * what's used to specify the names of objects related to the image.
5283 * A special "rbd id" object is used to map an rbd image name to its
5284 * id. If that object doesn't exist, then there is no v2 rbd image
5285 * with the supplied name.
5287 * This function will record the given rbd_dev's image_id field if
5288 * it can be determined, and in that case will return 0. If any
5289 * errors occur a negative errno will be returned and the rbd_dev's
5290 * image_id field will be unchanged (and should be NULL).
5292 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
5296 CEPH_DEFINE_OID_ONSTACK(oid);
5301 * When probing a parent image, the image id is already
5302 * known (and the image name likely is not). There's no
5303 * need to fetch the image id again in this case. We
5304 * do still need to set the image format though.
5306 if (rbd_dev->spec->image_id) {
5307 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
5313 * First, see if the format 2 image id file exists, and if
5314 * so, get the image's persistent id from it.
5316 ret = ceph_oid_aprintf(&oid, GFP_KERNEL, "%s%s", RBD_ID_PREFIX,
5317 rbd_dev->spec->image_name);
5321 dout("rbd id object name is %s\n", oid.name);
5323 /* Response will be an encoded string, which includes a length */
5325 size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
5326 response = kzalloc(size, GFP_NOIO);
5332 /* If it doesn't exist we'll assume it's a format 1 image */
5334 ret = rbd_obj_method_sync(rbd_dev, &oid, &rbd_dev->header_oloc,
5336 response, RBD_IMAGE_ID_LEN_MAX);
5337 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5338 if (ret == -ENOENT) {
5339 image_id = kstrdup("", GFP_KERNEL);
5340 ret = image_id ? 0 : -ENOMEM;
5342 rbd_dev->image_format = 1;
5343 } else if (ret >= 0) {
5346 image_id = ceph_extract_encoded_string(&p, p + ret,
5348 ret = PTR_ERR_OR_ZERO(image_id);
5350 rbd_dev->image_format = 2;
5354 rbd_dev->spec->image_id = image_id;
5355 dout("image_id is %s\n", image_id);
5359 ceph_oid_destroy(&oid);
5364 * Undo whatever state changes are made by v1 or v2 header info
5367 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
5369 struct rbd_image_header *header;
5371 rbd_dev_parent_put(rbd_dev);
5373 /* Free dynamic fields from the header, then zero it out */
5375 header = &rbd_dev->header;
5376 ceph_put_snap_context(header->snapc);
5377 kfree(header->snap_sizes);
5378 kfree(header->snap_names);
5379 kfree(header->object_prefix);
5380 memset(header, 0, sizeof (*header));
5383 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev)
5387 ret = rbd_dev_v2_object_prefix(rbd_dev);
5392 * Get the and check features for the image. Currently the
5393 * features are assumed to never change.
5395 ret = rbd_dev_v2_features(rbd_dev);
5399 /* If the image supports fancy striping, get its parameters */
5401 if (rbd_dev->header.features & RBD_FEATURE_STRIPINGV2) {
5402 ret = rbd_dev_v2_striping_info(rbd_dev);
5407 if (rbd_dev->header.features & RBD_FEATURE_DATA_POOL) {
5408 ret = rbd_dev_v2_data_pool(rbd_dev);
5413 rbd_init_layout(rbd_dev);
5417 rbd_dev->header.features = 0;
5418 kfree(rbd_dev->header.object_prefix);
5419 rbd_dev->header.object_prefix = NULL;
5424 * @depth is rbd_dev_image_probe() -> rbd_dev_probe_parent() ->
5425 * rbd_dev_image_probe() recursion depth, which means it's also the
5426 * length of the already discovered part of the parent chain.
5428 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev, int depth)
5430 struct rbd_device *parent = NULL;
5433 if (!rbd_dev->parent_spec)
5436 if (++depth > RBD_MAX_PARENT_CHAIN_LEN) {
5437 pr_info("parent chain is too long (%d)\n", depth);
5442 parent = __rbd_dev_create(rbd_dev->rbd_client, rbd_dev->parent_spec);
5449 * Images related by parent/child relationships always share
5450 * rbd_client and spec/parent_spec, so bump their refcounts.
5452 __rbd_get_client(rbd_dev->rbd_client);
5453 rbd_spec_get(rbd_dev->parent_spec);
5455 ret = rbd_dev_image_probe(parent, depth);
5459 rbd_dev->parent = parent;
5460 atomic_set(&rbd_dev->parent_ref, 1);
5464 rbd_dev_unparent(rbd_dev);
5465 rbd_dev_destroy(parent);
5469 static void rbd_dev_device_release(struct rbd_device *rbd_dev)
5471 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5472 rbd_dev_mapping_clear(rbd_dev);
5473 rbd_free_disk(rbd_dev);
5475 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5479 * rbd_dev->header_rwsem must be locked for write and will be unlocked
5482 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
5486 /* Record our major and minor device numbers. */
5488 if (!single_major) {
5489 ret = register_blkdev(0, rbd_dev->name);
5491 goto err_out_unlock;
5493 rbd_dev->major = ret;
5496 rbd_dev->major = rbd_major;
5497 rbd_dev->minor = rbd_dev_id_to_minor(rbd_dev->dev_id);
5500 /* Set up the blkdev mapping. */
5502 ret = rbd_init_disk(rbd_dev);
5504 goto err_out_blkdev;
5506 ret = rbd_dev_mapping_set(rbd_dev);
5510 set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
5511 set_disk_ro(rbd_dev->disk, rbd_dev->opts->read_only);
5513 ret = dev_set_name(&rbd_dev->dev, "%d", rbd_dev->dev_id);
5515 goto err_out_mapping;
5517 set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5518 up_write(&rbd_dev->header_rwsem);
5522 rbd_dev_mapping_clear(rbd_dev);
5524 rbd_free_disk(rbd_dev);
5527 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5529 up_write(&rbd_dev->header_rwsem);
5533 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
5535 struct rbd_spec *spec = rbd_dev->spec;
5538 /* Record the header object name for this rbd image. */
5540 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
5541 if (rbd_dev->image_format == 1)
5542 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
5543 spec->image_name, RBD_SUFFIX);
5545 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
5546 RBD_HEADER_PREFIX, spec->image_id);
5551 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
5553 rbd_dev_unprobe(rbd_dev);
5555 rbd_unregister_watch(rbd_dev);
5556 rbd_dev->image_format = 0;
5557 kfree(rbd_dev->spec->image_id);
5558 rbd_dev->spec->image_id = NULL;
5562 * Probe for the existence of the header object for the given rbd
5563 * device. If this image is the one being mapped (i.e., not a
5564 * parent), initiate a watch on its header object before using that
5565 * object to get detailed information about the rbd image.
5567 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth)
5572 * Get the id from the image id object. Unless there's an
5573 * error, rbd_dev->spec->image_id will be filled in with
5574 * a dynamically-allocated string, and rbd_dev->image_format
5575 * will be set to either 1 or 2.
5577 ret = rbd_dev_image_id(rbd_dev);
5581 ret = rbd_dev_header_name(rbd_dev);
5583 goto err_out_format;
5586 ret = rbd_register_watch(rbd_dev);
5589 pr_info("image %s/%s does not exist\n",
5590 rbd_dev->spec->pool_name,
5591 rbd_dev->spec->image_name);
5592 goto err_out_format;
5596 ret = rbd_dev_header_info(rbd_dev);
5601 * If this image is the one being mapped, we have pool name and
5602 * id, image name and id, and snap name - need to fill snap id.
5603 * Otherwise this is a parent image, identified by pool, image
5604 * and snap ids - need to fill in names for those ids.
5607 ret = rbd_spec_fill_snap_id(rbd_dev);
5609 ret = rbd_spec_fill_names(rbd_dev);
5612 pr_info("snap %s/%s@%s does not exist\n",
5613 rbd_dev->spec->pool_name,
5614 rbd_dev->spec->image_name,
5615 rbd_dev->spec->snap_name);
5619 if (rbd_dev->header.features & RBD_FEATURE_LAYERING) {
5620 ret = rbd_dev_v2_parent_info(rbd_dev);
5625 * Need to warn users if this image is the one being
5626 * mapped and has a parent.
5628 if (!depth && rbd_dev->parent_spec)
5630 "WARNING: kernel layering is EXPERIMENTAL!");
5633 ret = rbd_dev_probe_parent(rbd_dev, depth);
5637 dout("discovered format %u image, header name is %s\n",
5638 rbd_dev->image_format, rbd_dev->header_oid.name);
5642 rbd_dev_unprobe(rbd_dev);
5645 rbd_unregister_watch(rbd_dev);
5647 rbd_dev->image_format = 0;
5648 kfree(rbd_dev->spec->image_id);
5649 rbd_dev->spec->image_id = NULL;
5653 static ssize_t do_rbd_add(struct bus_type *bus,
5657 struct rbd_device *rbd_dev = NULL;
5658 struct ceph_options *ceph_opts = NULL;
5659 struct rbd_options *rbd_opts = NULL;
5660 struct rbd_spec *spec = NULL;
5661 struct rbd_client *rbdc;
5664 if (!try_module_get(THIS_MODULE))
5667 /* parse add command */
5668 rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
5672 rbdc = rbd_get_client(ceph_opts);
5679 rc = ceph_pg_poolid_by_name(rbdc->client->osdc.osdmap, spec->pool_name);
5682 pr_info("pool %s does not exist\n", spec->pool_name);
5683 goto err_out_client;
5685 spec->pool_id = (u64)rc;
5687 rbd_dev = rbd_dev_create(rbdc, spec, rbd_opts);
5690 goto err_out_client;
5692 rbdc = NULL; /* rbd_dev now owns this */
5693 spec = NULL; /* rbd_dev now owns this */
5694 rbd_opts = NULL; /* rbd_dev now owns this */
5696 rbd_dev->config_info = kstrdup(buf, GFP_KERNEL);
5697 if (!rbd_dev->config_info) {
5699 goto err_out_rbd_dev;
5702 down_write(&rbd_dev->header_rwsem);
5703 rc = rbd_dev_image_probe(rbd_dev, 0);
5705 up_write(&rbd_dev->header_rwsem);
5706 goto err_out_rbd_dev;
5709 /* If we are mapping a snapshot it must be marked read-only */
5710 if (rbd_dev->spec->snap_id != CEPH_NOSNAP)
5711 rbd_dev->opts->read_only = true;
5713 rc = rbd_dev_device_setup(rbd_dev);
5715 goto err_out_image_probe;
5717 if (rbd_dev->opts->exclusive) {
5718 rc = rbd_add_acquire_lock(rbd_dev);
5720 goto err_out_device_setup;
5723 /* Everything's ready. Announce the disk to the world. */
5725 rc = device_add(&rbd_dev->dev);
5727 goto err_out_image_lock;
5729 add_disk(rbd_dev->disk);
5730 /* see rbd_init_disk() */
5731 blk_put_queue(rbd_dev->disk->queue);
5733 spin_lock(&rbd_dev_list_lock);
5734 list_add_tail(&rbd_dev->node, &rbd_dev_list);
5735 spin_unlock(&rbd_dev_list_lock);
5737 pr_info("%s: capacity %llu features 0x%llx\n", rbd_dev->disk->disk_name,
5738 (unsigned long long)get_capacity(rbd_dev->disk) << SECTOR_SHIFT,
5739 rbd_dev->header.features);
5742 module_put(THIS_MODULE);
5746 rbd_dev_image_unlock(rbd_dev);
5747 err_out_device_setup:
5748 rbd_dev_device_release(rbd_dev);
5749 err_out_image_probe:
5750 rbd_dev_image_release(rbd_dev);
5752 rbd_dev_destroy(rbd_dev);
5754 rbd_put_client(rbdc);
5761 static ssize_t rbd_add(struct bus_type *bus,
5768 return do_rbd_add(bus, buf, count);
5771 static ssize_t rbd_add_single_major(struct bus_type *bus,
5775 return do_rbd_add(bus, buf, count);
5778 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
5780 while (rbd_dev->parent) {
5781 struct rbd_device *first = rbd_dev;
5782 struct rbd_device *second = first->parent;
5783 struct rbd_device *third;
5786 * Follow to the parent with no grandparent and
5789 while (second && (third = second->parent)) {
5794 rbd_dev_image_release(second);
5795 rbd_dev_destroy(second);
5796 first->parent = NULL;
5797 first->parent_overlap = 0;
5799 rbd_assert(first->parent_spec);
5800 rbd_spec_put(first->parent_spec);
5801 first->parent_spec = NULL;
5805 static ssize_t do_rbd_remove(struct bus_type *bus,
5809 struct rbd_device *rbd_dev = NULL;
5810 struct list_head *tmp;
5813 bool already = false;
5819 sscanf(buf, "%d %5s", &dev_id, opt_buf);
5821 pr_err("dev_id out of range\n");
5824 if (opt_buf[0] != '\0') {
5825 if (!strcmp(opt_buf, "force")) {
5828 pr_err("bad remove option at '%s'\n", opt_buf);
5834 spin_lock(&rbd_dev_list_lock);
5835 list_for_each(tmp, &rbd_dev_list) {
5836 rbd_dev = list_entry(tmp, struct rbd_device, node);
5837 if (rbd_dev->dev_id == dev_id) {
5843 spin_lock_irq(&rbd_dev->lock);
5844 if (rbd_dev->open_count && !force)
5847 already = test_and_set_bit(RBD_DEV_FLAG_REMOVING,
5849 spin_unlock_irq(&rbd_dev->lock);
5851 spin_unlock(&rbd_dev_list_lock);
5852 if (ret < 0 || already)
5857 * Prevent new IO from being queued and wait for existing
5858 * IO to complete/fail.
5860 blk_mq_freeze_queue(rbd_dev->disk->queue);
5861 blk_set_queue_dying(rbd_dev->disk->queue);
5864 del_gendisk(rbd_dev->disk);
5865 spin_lock(&rbd_dev_list_lock);
5866 list_del_init(&rbd_dev->node);
5867 spin_unlock(&rbd_dev_list_lock);
5868 device_del(&rbd_dev->dev);
5870 rbd_dev_image_unlock(rbd_dev);
5871 rbd_dev_device_release(rbd_dev);
5872 rbd_dev_image_release(rbd_dev);
5873 rbd_dev_destroy(rbd_dev);
5877 static ssize_t rbd_remove(struct bus_type *bus,
5884 return do_rbd_remove(bus, buf, count);
5887 static ssize_t rbd_remove_single_major(struct bus_type *bus,
5891 return do_rbd_remove(bus, buf, count);
5895 * create control files in sysfs
5898 static int rbd_sysfs_init(void)
5902 ret = device_register(&rbd_root_dev);
5906 ret = bus_register(&rbd_bus_type);
5908 device_unregister(&rbd_root_dev);
5913 static void rbd_sysfs_cleanup(void)
5915 bus_unregister(&rbd_bus_type);
5916 device_unregister(&rbd_root_dev);
5919 static int rbd_slab_init(void)
5921 rbd_assert(!rbd_img_request_cache);
5922 rbd_img_request_cache = KMEM_CACHE(rbd_img_request, 0);
5923 if (!rbd_img_request_cache)
5926 rbd_assert(!rbd_obj_request_cache);
5927 rbd_obj_request_cache = KMEM_CACHE(rbd_obj_request, 0);
5928 if (!rbd_obj_request_cache)
5934 kmem_cache_destroy(rbd_img_request_cache);
5935 rbd_img_request_cache = NULL;
5939 static void rbd_slab_exit(void)
5941 rbd_assert(rbd_obj_request_cache);
5942 kmem_cache_destroy(rbd_obj_request_cache);
5943 rbd_obj_request_cache = NULL;
5945 rbd_assert(rbd_img_request_cache);
5946 kmem_cache_destroy(rbd_img_request_cache);
5947 rbd_img_request_cache = NULL;
5950 static int __init rbd_init(void)
5954 if (!libceph_compatible(NULL)) {
5955 rbd_warn(NULL, "libceph incompatibility (quitting)");
5959 rc = rbd_slab_init();
5964 * The number of active work items is limited by the number of
5965 * rbd devices * queue depth, so leave @max_active at default.
5967 rbd_wq = alloc_workqueue(RBD_DRV_NAME, WQ_MEM_RECLAIM, 0);
5974 rbd_major = register_blkdev(0, RBD_DRV_NAME);
5975 if (rbd_major < 0) {
5981 rc = rbd_sysfs_init();
5983 goto err_out_blkdev;
5986 pr_info("loaded (major %d)\n", rbd_major);
5988 pr_info("loaded\n");
5994 unregister_blkdev(rbd_major, RBD_DRV_NAME);
5996 destroy_workqueue(rbd_wq);
6002 static void __exit rbd_exit(void)
6004 ida_destroy(&rbd_dev_id_ida);
6005 rbd_sysfs_cleanup();
6007 unregister_blkdev(rbd_major, RBD_DRV_NAME);
6008 destroy_workqueue(rbd_wq);
6012 module_init(rbd_init);
6013 module_exit(rbd_exit);
6015 MODULE_AUTHOR("Alex Elder <elder@inktank.com>");
6016 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
6017 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
6018 /* following authorship retained from original osdblk.c */
6019 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
6021 MODULE_DESCRIPTION("RADOS Block Device (RBD) driver");
6022 MODULE_LICENSE("GPL");