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_fetch_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;
184 const char *pool_ns; /* NULL if default, never "" */
186 const char *image_id;
187 const char *image_name;
190 const char *snap_name;
196 * an instance of the client. multiple devices may share an rbd client.
199 struct ceph_client *client;
201 struct list_head node;
204 struct rbd_img_request;
206 enum obj_request_type {
207 OBJ_REQUEST_NODATA = 1,
208 OBJ_REQUEST_BIO, /* pointer into provided bio (list) */
209 OBJ_REQUEST_BVECS, /* pointer into provided bio_vec array */
210 OBJ_REQUEST_OWN_BVECS, /* private bio_vec array, doesn't own pages */
213 enum obj_operation_type {
220 * Writes go through the following state machine to deal with
224 * RBD_OBJ_WRITE_GUARD ---------------> RBD_OBJ_WRITE_COPYUP
226 * v \------------------------------/
232 * Writes start in RBD_OBJ_WRITE_GUARD or _FLAT, depending on whether
233 * there is a parent or not.
235 enum rbd_obj_write_state {
236 RBD_OBJ_WRITE_FLAT = 1,
238 RBD_OBJ_WRITE_COPYUP,
241 struct rbd_obj_request {
242 struct ceph_object_extent ex;
244 bool tried_parent; /* for reads */
245 enum rbd_obj_write_state write_state; /* for writes */
248 struct rbd_img_request *img_request;
249 struct ceph_file_extent *img_extents;
253 struct ceph_bio_iter bio_pos;
255 struct ceph_bvec_iter bvec_pos;
260 struct bio_vec *copyup_bvecs;
261 u32 copyup_bvec_count;
263 struct ceph_osd_request *osd_req;
265 u64 xferred; /* bytes transferred */
272 IMG_REQ_CHILD, /* initiator: block = 0, child image = 1 */
273 IMG_REQ_LAYERED, /* ENOENT handling: normal = 0, layered = 1 */
276 struct rbd_img_request {
277 struct rbd_device *rbd_dev;
278 enum obj_operation_type op_type;
279 enum obj_request_type data_type;
282 u64 snap_id; /* for reads */
283 struct ceph_snap_context *snapc; /* for writes */
286 struct request *rq; /* block request */
287 struct rbd_obj_request *obj_request; /* obj req initiator */
289 spinlock_t completion_lock;
290 u64 xferred;/* aggregate bytes transferred */
291 int result; /* first nonzero obj_request result */
293 struct list_head object_extents; /* obj_req.ex structs */
294 u32 obj_request_count;
300 #define for_each_obj_request(ireq, oreq) \
301 list_for_each_entry(oreq, &(ireq)->object_extents, ex.oe_item)
302 #define for_each_obj_request_safe(ireq, oreq, n) \
303 list_for_each_entry_safe(oreq, n, &(ireq)->object_extents, ex.oe_item)
305 enum rbd_watch_state {
306 RBD_WATCH_STATE_UNREGISTERED,
307 RBD_WATCH_STATE_REGISTERED,
308 RBD_WATCH_STATE_ERROR,
311 enum rbd_lock_state {
312 RBD_LOCK_STATE_UNLOCKED,
313 RBD_LOCK_STATE_LOCKED,
314 RBD_LOCK_STATE_RELEASING,
317 /* WatchNotify::ClientId */
318 struct rbd_client_id {
332 int dev_id; /* blkdev unique id */
334 int major; /* blkdev assigned major */
336 struct gendisk *disk; /* blkdev's gendisk and rq */
338 u32 image_format; /* Either 1 or 2 */
339 struct rbd_client *rbd_client;
341 char name[DEV_NAME_LEN]; /* blkdev name, e.g. rbd3 */
343 spinlock_t lock; /* queue, flags, open_count */
345 struct rbd_image_header header;
346 unsigned long flags; /* possibly lock protected */
347 struct rbd_spec *spec;
348 struct rbd_options *opts;
349 char *config_info; /* add{,_single_major} string */
351 struct ceph_object_id header_oid;
352 struct ceph_object_locator header_oloc;
354 struct ceph_file_layout layout; /* used for all rbd requests */
356 struct mutex watch_mutex;
357 enum rbd_watch_state watch_state;
358 struct ceph_osd_linger_request *watch_handle;
360 struct delayed_work watch_dwork;
362 struct rw_semaphore lock_rwsem;
363 enum rbd_lock_state lock_state;
364 char lock_cookie[32];
365 struct rbd_client_id owner_cid;
366 struct work_struct acquired_lock_work;
367 struct work_struct released_lock_work;
368 struct delayed_work lock_dwork;
369 struct work_struct unlock_work;
370 wait_queue_head_t lock_waitq;
372 struct workqueue_struct *task_wq;
374 struct rbd_spec *parent_spec;
377 struct rbd_device *parent;
379 /* Block layer tags. */
380 struct blk_mq_tag_set tag_set;
382 /* protects updating the header */
383 struct rw_semaphore header_rwsem;
385 struct rbd_mapping mapping;
387 struct list_head node;
391 unsigned long open_count; /* protected by lock */
395 * Flag bits for rbd_dev->flags:
396 * - REMOVING (which is coupled with rbd_dev->open_count) is protected
398 * - BLACKLISTED is protected by rbd_dev->lock_rwsem
401 RBD_DEV_FLAG_EXISTS, /* mapped snapshot has not been deleted */
402 RBD_DEV_FLAG_REMOVING, /* this mapping is being removed */
403 RBD_DEV_FLAG_BLACKLISTED, /* our ceph_client is blacklisted */
406 static DEFINE_MUTEX(client_mutex); /* Serialize client creation */
408 static LIST_HEAD(rbd_dev_list); /* devices */
409 static DEFINE_SPINLOCK(rbd_dev_list_lock);
411 static LIST_HEAD(rbd_client_list); /* clients */
412 static DEFINE_SPINLOCK(rbd_client_list_lock);
414 /* Slab caches for frequently-allocated structures */
416 static struct kmem_cache *rbd_img_request_cache;
417 static struct kmem_cache *rbd_obj_request_cache;
419 static int rbd_major;
420 static DEFINE_IDA(rbd_dev_id_ida);
422 static struct workqueue_struct *rbd_wq;
425 * single-major requires >= 0.75 version of userspace rbd utility.
427 static bool single_major = true;
428 module_param(single_major, bool, 0444);
429 MODULE_PARM_DESC(single_major, "Use a single major number for all rbd devices (default: true)");
431 static ssize_t rbd_add(struct bus_type *bus, const char *buf,
433 static ssize_t rbd_remove(struct bus_type *bus, const char *buf,
435 static ssize_t rbd_add_single_major(struct bus_type *bus, const char *buf,
437 static ssize_t rbd_remove_single_major(struct bus_type *bus, const char *buf,
439 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth);
441 static int rbd_dev_id_to_minor(int dev_id)
443 return dev_id << RBD_SINGLE_MAJOR_PART_SHIFT;
446 static int minor_to_rbd_dev_id(int minor)
448 return minor >> RBD_SINGLE_MAJOR_PART_SHIFT;
451 static bool __rbd_is_lock_owner(struct rbd_device *rbd_dev)
453 return rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED ||
454 rbd_dev->lock_state == RBD_LOCK_STATE_RELEASING;
457 static bool rbd_is_lock_owner(struct rbd_device *rbd_dev)
461 down_read(&rbd_dev->lock_rwsem);
462 is_lock_owner = __rbd_is_lock_owner(rbd_dev);
463 up_read(&rbd_dev->lock_rwsem);
464 return is_lock_owner;
467 static ssize_t rbd_supported_features_show(struct bus_type *bus, char *buf)
469 return sprintf(buf, "0x%llx\n", RBD_FEATURES_SUPPORTED);
472 static BUS_ATTR(add, 0200, NULL, rbd_add);
473 static BUS_ATTR(remove, 0200, NULL, rbd_remove);
474 static BUS_ATTR(add_single_major, 0200, NULL, rbd_add_single_major);
475 static BUS_ATTR(remove_single_major, 0200, NULL, rbd_remove_single_major);
476 static BUS_ATTR(supported_features, 0444, rbd_supported_features_show, NULL);
478 static struct attribute *rbd_bus_attrs[] = {
480 &bus_attr_remove.attr,
481 &bus_attr_add_single_major.attr,
482 &bus_attr_remove_single_major.attr,
483 &bus_attr_supported_features.attr,
487 static umode_t rbd_bus_is_visible(struct kobject *kobj,
488 struct attribute *attr, int index)
491 (attr == &bus_attr_add_single_major.attr ||
492 attr == &bus_attr_remove_single_major.attr))
498 static const struct attribute_group rbd_bus_group = {
499 .attrs = rbd_bus_attrs,
500 .is_visible = rbd_bus_is_visible,
502 __ATTRIBUTE_GROUPS(rbd_bus);
504 static struct bus_type rbd_bus_type = {
506 .bus_groups = rbd_bus_groups,
509 static void rbd_root_dev_release(struct device *dev)
513 static struct device rbd_root_dev = {
515 .release = rbd_root_dev_release,
518 static __printf(2, 3)
519 void rbd_warn(struct rbd_device *rbd_dev, const char *fmt, ...)
521 struct va_format vaf;
529 printk(KERN_WARNING "%s: %pV\n", RBD_DRV_NAME, &vaf);
530 else if (rbd_dev->disk)
531 printk(KERN_WARNING "%s: %s: %pV\n",
532 RBD_DRV_NAME, rbd_dev->disk->disk_name, &vaf);
533 else if (rbd_dev->spec && rbd_dev->spec->image_name)
534 printk(KERN_WARNING "%s: image %s: %pV\n",
535 RBD_DRV_NAME, rbd_dev->spec->image_name, &vaf);
536 else if (rbd_dev->spec && rbd_dev->spec->image_id)
537 printk(KERN_WARNING "%s: id %s: %pV\n",
538 RBD_DRV_NAME, rbd_dev->spec->image_id, &vaf);
540 printk(KERN_WARNING "%s: rbd_dev %p: %pV\n",
541 RBD_DRV_NAME, rbd_dev, &vaf);
546 #define rbd_assert(expr) \
547 if (unlikely(!(expr))) { \
548 printk(KERN_ERR "\nAssertion failure in %s() " \
550 "\trbd_assert(%s);\n\n", \
551 __func__, __LINE__, #expr); \
554 #else /* !RBD_DEBUG */
555 # define rbd_assert(expr) ((void) 0)
556 #endif /* !RBD_DEBUG */
558 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev);
560 static int rbd_dev_refresh(struct rbd_device *rbd_dev);
561 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev);
562 static int rbd_dev_header_info(struct rbd_device *rbd_dev);
563 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev);
564 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
566 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
567 u8 *order, u64 *snap_size);
568 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
571 static int rbd_open(struct block_device *bdev, fmode_t mode)
573 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
574 bool removing = false;
576 spin_lock_irq(&rbd_dev->lock);
577 if (test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags))
580 rbd_dev->open_count++;
581 spin_unlock_irq(&rbd_dev->lock);
585 (void) get_device(&rbd_dev->dev);
590 static void rbd_release(struct gendisk *disk, fmode_t mode)
592 struct rbd_device *rbd_dev = disk->private_data;
593 unsigned long open_count_before;
595 spin_lock_irq(&rbd_dev->lock);
596 open_count_before = rbd_dev->open_count--;
597 spin_unlock_irq(&rbd_dev->lock);
598 rbd_assert(open_count_before > 0);
600 put_device(&rbd_dev->dev);
603 static int rbd_ioctl_set_ro(struct rbd_device *rbd_dev, unsigned long arg)
607 if (get_user(ro, (int __user *)arg))
610 /* Snapshots can't be marked read-write */
611 if (rbd_dev->spec->snap_id != CEPH_NOSNAP && !ro)
614 /* Let blkdev_roset() handle it */
618 static int rbd_ioctl(struct block_device *bdev, fmode_t mode,
619 unsigned int cmd, unsigned long arg)
621 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
626 ret = rbd_ioctl_set_ro(rbd_dev, arg);
636 static int rbd_compat_ioctl(struct block_device *bdev, fmode_t mode,
637 unsigned int cmd, unsigned long arg)
639 return rbd_ioctl(bdev, mode, cmd, arg);
641 #endif /* CONFIG_COMPAT */
643 static const struct block_device_operations rbd_bd_ops = {
644 .owner = THIS_MODULE,
646 .release = rbd_release,
649 .compat_ioctl = rbd_compat_ioctl,
654 * Initialize an rbd client instance. Success or not, this function
655 * consumes ceph_opts. Caller holds client_mutex.
657 static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts)
659 struct rbd_client *rbdc;
662 dout("%s:\n", __func__);
663 rbdc = kmalloc(sizeof(struct rbd_client), GFP_KERNEL);
667 kref_init(&rbdc->kref);
668 INIT_LIST_HEAD(&rbdc->node);
670 rbdc->client = ceph_create_client(ceph_opts, rbdc);
671 if (IS_ERR(rbdc->client))
673 ceph_opts = NULL; /* Now rbdc->client is responsible for ceph_opts */
675 ret = ceph_open_session(rbdc->client);
679 spin_lock(&rbd_client_list_lock);
680 list_add_tail(&rbdc->node, &rbd_client_list);
681 spin_unlock(&rbd_client_list_lock);
683 dout("%s: rbdc %p\n", __func__, rbdc);
687 ceph_destroy_client(rbdc->client);
692 ceph_destroy_options(ceph_opts);
693 dout("%s: error %d\n", __func__, ret);
698 static struct rbd_client *__rbd_get_client(struct rbd_client *rbdc)
700 kref_get(&rbdc->kref);
706 * Find a ceph client with specific addr and configuration. If
707 * found, bump its reference count.
709 static struct rbd_client *rbd_client_find(struct ceph_options *ceph_opts)
711 struct rbd_client *client_node;
714 if (ceph_opts->flags & CEPH_OPT_NOSHARE)
717 spin_lock(&rbd_client_list_lock);
718 list_for_each_entry(client_node, &rbd_client_list, node) {
719 if (!ceph_compare_options(ceph_opts, client_node->client)) {
720 __rbd_get_client(client_node);
726 spin_unlock(&rbd_client_list_lock);
728 return found ? client_node : NULL;
732 * (Per device) rbd map options
741 /* string args above */
750 static match_table_t rbd_opts_tokens = {
751 {Opt_queue_depth, "queue_depth=%d"},
752 {Opt_lock_timeout, "lock_timeout=%d"},
754 {Opt_pool_ns, "_pool_ns=%s"},
755 /* string args above */
756 {Opt_read_only, "read_only"},
757 {Opt_read_only, "ro"}, /* Alternate spelling */
758 {Opt_read_write, "read_write"},
759 {Opt_read_write, "rw"}, /* Alternate spelling */
760 {Opt_lock_on_read, "lock_on_read"},
761 {Opt_exclusive, "exclusive"},
762 {Opt_notrim, "notrim"},
768 unsigned long lock_timeout;
775 #define RBD_QUEUE_DEPTH_DEFAULT BLKDEV_MAX_RQ
776 #define RBD_LOCK_TIMEOUT_DEFAULT 0 /* no timeout */
777 #define RBD_READ_ONLY_DEFAULT false
778 #define RBD_LOCK_ON_READ_DEFAULT false
779 #define RBD_EXCLUSIVE_DEFAULT false
780 #define RBD_TRIM_DEFAULT true
782 struct parse_rbd_opts_ctx {
783 struct rbd_spec *spec;
784 struct rbd_options *opts;
787 static int parse_rbd_opts_token(char *c, void *private)
789 struct parse_rbd_opts_ctx *pctx = private;
790 substring_t argstr[MAX_OPT_ARGS];
791 int token, intval, ret;
793 token = match_token(c, rbd_opts_tokens, argstr);
794 if (token < Opt_last_int) {
795 ret = match_int(&argstr[0], &intval);
797 pr_err("bad option arg (not int) at '%s'\n", c);
800 dout("got int token %d val %d\n", token, intval);
801 } else if (token > Opt_last_int && token < Opt_last_string) {
802 dout("got string token %d val %s\n", token, argstr[0].from);
804 dout("got token %d\n", token);
808 case Opt_queue_depth:
810 pr_err("queue_depth out of range\n");
813 pctx->opts->queue_depth = intval;
815 case Opt_lock_timeout:
816 /* 0 is "wait forever" (i.e. infinite timeout) */
817 if (intval < 0 || intval > INT_MAX / 1000) {
818 pr_err("lock_timeout out of range\n");
821 pctx->opts->lock_timeout = msecs_to_jiffies(intval * 1000);
824 kfree(pctx->spec->pool_ns);
825 pctx->spec->pool_ns = match_strdup(argstr);
826 if (!pctx->spec->pool_ns)
830 pctx->opts->read_only = true;
833 pctx->opts->read_only = false;
835 case Opt_lock_on_read:
836 pctx->opts->lock_on_read = true;
839 pctx->opts->exclusive = true;
842 pctx->opts->trim = false;
845 /* libceph prints "bad option" msg */
852 static char* obj_op_name(enum obj_operation_type op_type)
867 * Destroy ceph client
869 * Caller must hold rbd_client_list_lock.
871 static void rbd_client_release(struct kref *kref)
873 struct rbd_client *rbdc = container_of(kref, struct rbd_client, kref);
875 dout("%s: rbdc %p\n", __func__, rbdc);
876 spin_lock(&rbd_client_list_lock);
877 list_del(&rbdc->node);
878 spin_unlock(&rbd_client_list_lock);
880 ceph_destroy_client(rbdc->client);
885 * Drop reference to ceph client node. If it's not referenced anymore, release
888 static void rbd_put_client(struct rbd_client *rbdc)
891 kref_put(&rbdc->kref, rbd_client_release);
894 static int wait_for_latest_osdmap(struct ceph_client *client)
899 ret = ceph_monc_get_version(&client->monc, "osdmap", &newest_epoch);
903 if (client->osdc.osdmap->epoch >= newest_epoch)
906 ceph_osdc_maybe_request_map(&client->osdc);
907 return ceph_monc_wait_osdmap(&client->monc, newest_epoch,
908 client->options->mount_timeout);
912 * Get a ceph client with specific addr and configuration, if one does
913 * not exist create it. Either way, ceph_opts is consumed by this
916 static struct rbd_client *rbd_get_client(struct ceph_options *ceph_opts)
918 struct rbd_client *rbdc;
921 mutex_lock_nested(&client_mutex, SINGLE_DEPTH_NESTING);
922 rbdc = rbd_client_find(ceph_opts);
924 ceph_destroy_options(ceph_opts);
927 * Using an existing client. Make sure ->pg_pools is up to
928 * date before we look up the pool id in do_rbd_add().
930 ret = wait_for_latest_osdmap(rbdc->client);
932 rbd_warn(NULL, "failed to get latest osdmap: %d", ret);
933 rbd_put_client(rbdc);
937 rbdc = rbd_client_create(ceph_opts);
939 mutex_unlock(&client_mutex);
944 static bool rbd_image_format_valid(u32 image_format)
946 return image_format == 1 || image_format == 2;
949 static bool rbd_dev_ondisk_valid(struct rbd_image_header_ondisk *ondisk)
954 /* The header has to start with the magic rbd header text */
955 if (memcmp(&ondisk->text, RBD_HEADER_TEXT, sizeof (RBD_HEADER_TEXT)))
958 /* The bio layer requires at least sector-sized I/O */
960 if (ondisk->options.order < SECTOR_SHIFT)
963 /* If we use u64 in a few spots we may be able to loosen this */
965 if (ondisk->options.order > 8 * sizeof (int) - 1)
969 * The size of a snapshot header has to fit in a size_t, and
970 * that limits the number of snapshots.
972 snap_count = le32_to_cpu(ondisk->snap_count);
973 size = SIZE_MAX - sizeof (struct ceph_snap_context);
974 if (snap_count > size / sizeof (__le64))
978 * Not only that, but the size of the entire the snapshot
979 * header must also be representable in a size_t.
981 size -= snap_count * sizeof (__le64);
982 if ((u64) size < le64_to_cpu(ondisk->snap_names_len))
989 * returns the size of an object in the image
991 static u32 rbd_obj_bytes(struct rbd_image_header *header)
993 return 1U << header->obj_order;
996 static void rbd_init_layout(struct rbd_device *rbd_dev)
998 if (rbd_dev->header.stripe_unit == 0 ||
999 rbd_dev->header.stripe_count == 0) {
1000 rbd_dev->header.stripe_unit = rbd_obj_bytes(&rbd_dev->header);
1001 rbd_dev->header.stripe_count = 1;
1004 rbd_dev->layout.stripe_unit = rbd_dev->header.stripe_unit;
1005 rbd_dev->layout.stripe_count = rbd_dev->header.stripe_count;
1006 rbd_dev->layout.object_size = rbd_obj_bytes(&rbd_dev->header);
1007 rbd_dev->layout.pool_id = rbd_dev->header.data_pool_id == CEPH_NOPOOL ?
1008 rbd_dev->spec->pool_id : rbd_dev->header.data_pool_id;
1009 RCU_INIT_POINTER(rbd_dev->layout.pool_ns, NULL);
1013 * Fill an rbd image header with information from the given format 1
1016 static int rbd_header_from_disk(struct rbd_device *rbd_dev,
1017 struct rbd_image_header_ondisk *ondisk)
1019 struct rbd_image_header *header = &rbd_dev->header;
1020 bool first_time = header->object_prefix == NULL;
1021 struct ceph_snap_context *snapc;
1022 char *object_prefix = NULL;
1023 char *snap_names = NULL;
1024 u64 *snap_sizes = NULL;
1029 /* Allocate this now to avoid having to handle failure below */
1032 object_prefix = kstrndup(ondisk->object_prefix,
1033 sizeof(ondisk->object_prefix),
1039 /* Allocate the snapshot context and fill it in */
1041 snap_count = le32_to_cpu(ondisk->snap_count);
1042 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
1045 snapc->seq = le64_to_cpu(ondisk->snap_seq);
1047 struct rbd_image_snap_ondisk *snaps;
1048 u64 snap_names_len = le64_to_cpu(ondisk->snap_names_len);
1050 /* We'll keep a copy of the snapshot names... */
1052 if (snap_names_len > (u64)SIZE_MAX)
1054 snap_names = kmalloc(snap_names_len, GFP_KERNEL);
1058 /* ...as well as the array of their sizes. */
1059 snap_sizes = kmalloc_array(snap_count,
1060 sizeof(*header->snap_sizes),
1066 * Copy the names, and fill in each snapshot's id
1069 * Note that rbd_dev_v1_header_info() guarantees the
1070 * ondisk buffer we're working with has
1071 * snap_names_len bytes beyond the end of the
1072 * snapshot id array, this memcpy() is safe.
1074 memcpy(snap_names, &ondisk->snaps[snap_count], snap_names_len);
1075 snaps = ondisk->snaps;
1076 for (i = 0; i < snap_count; i++) {
1077 snapc->snaps[i] = le64_to_cpu(snaps[i].id);
1078 snap_sizes[i] = le64_to_cpu(snaps[i].image_size);
1082 /* We won't fail any more, fill in the header */
1085 header->object_prefix = object_prefix;
1086 header->obj_order = ondisk->options.order;
1087 rbd_init_layout(rbd_dev);
1089 ceph_put_snap_context(header->snapc);
1090 kfree(header->snap_names);
1091 kfree(header->snap_sizes);
1094 /* The remaining fields always get updated (when we refresh) */
1096 header->image_size = le64_to_cpu(ondisk->image_size);
1097 header->snapc = snapc;
1098 header->snap_names = snap_names;
1099 header->snap_sizes = snap_sizes;
1107 ceph_put_snap_context(snapc);
1108 kfree(object_prefix);
1113 static const char *_rbd_dev_v1_snap_name(struct rbd_device *rbd_dev, u32 which)
1115 const char *snap_name;
1117 rbd_assert(which < rbd_dev->header.snapc->num_snaps);
1119 /* Skip over names until we find the one we are looking for */
1121 snap_name = rbd_dev->header.snap_names;
1123 snap_name += strlen(snap_name) + 1;
1125 return kstrdup(snap_name, GFP_KERNEL);
1129 * Snapshot id comparison function for use with qsort()/bsearch().
1130 * Note that result is for snapshots in *descending* order.
1132 static int snapid_compare_reverse(const void *s1, const void *s2)
1134 u64 snap_id1 = *(u64 *)s1;
1135 u64 snap_id2 = *(u64 *)s2;
1137 if (snap_id1 < snap_id2)
1139 return snap_id1 == snap_id2 ? 0 : -1;
1143 * Search a snapshot context to see if the given snapshot id is
1146 * Returns the position of the snapshot id in the array if it's found,
1147 * or BAD_SNAP_INDEX otherwise.
1149 * Note: The snapshot array is in kept sorted (by the osd) in
1150 * reverse order, highest snapshot id first.
1152 static u32 rbd_dev_snap_index(struct rbd_device *rbd_dev, u64 snap_id)
1154 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
1157 found = bsearch(&snap_id, &snapc->snaps, snapc->num_snaps,
1158 sizeof (snap_id), snapid_compare_reverse);
1160 return found ? (u32)(found - &snapc->snaps[0]) : BAD_SNAP_INDEX;
1163 static const char *rbd_dev_v1_snap_name(struct rbd_device *rbd_dev,
1167 const char *snap_name;
1169 which = rbd_dev_snap_index(rbd_dev, snap_id);
1170 if (which == BAD_SNAP_INDEX)
1171 return ERR_PTR(-ENOENT);
1173 snap_name = _rbd_dev_v1_snap_name(rbd_dev, which);
1174 return snap_name ? snap_name : ERR_PTR(-ENOMEM);
1177 static const char *rbd_snap_name(struct rbd_device *rbd_dev, u64 snap_id)
1179 if (snap_id == CEPH_NOSNAP)
1180 return RBD_SNAP_HEAD_NAME;
1182 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1183 if (rbd_dev->image_format == 1)
1184 return rbd_dev_v1_snap_name(rbd_dev, snap_id);
1186 return rbd_dev_v2_snap_name(rbd_dev, snap_id);
1189 static int rbd_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
1192 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1193 if (snap_id == CEPH_NOSNAP) {
1194 *snap_size = rbd_dev->header.image_size;
1195 } else if (rbd_dev->image_format == 1) {
1198 which = rbd_dev_snap_index(rbd_dev, snap_id);
1199 if (which == BAD_SNAP_INDEX)
1202 *snap_size = rbd_dev->header.snap_sizes[which];
1207 ret = _rbd_dev_v2_snap_size(rbd_dev, snap_id, NULL, &size);
1216 static int rbd_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
1219 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1220 if (snap_id == CEPH_NOSNAP) {
1221 *snap_features = rbd_dev->header.features;
1222 } else if (rbd_dev->image_format == 1) {
1223 *snap_features = 0; /* No features for format 1 */
1228 ret = _rbd_dev_v2_snap_features(rbd_dev, snap_id, &features);
1232 *snap_features = features;
1237 static int rbd_dev_mapping_set(struct rbd_device *rbd_dev)
1239 u64 snap_id = rbd_dev->spec->snap_id;
1244 ret = rbd_snap_size(rbd_dev, snap_id, &size);
1247 ret = rbd_snap_features(rbd_dev, snap_id, &features);
1251 rbd_dev->mapping.size = size;
1252 rbd_dev->mapping.features = features;
1257 static void rbd_dev_mapping_clear(struct rbd_device *rbd_dev)
1259 rbd_dev->mapping.size = 0;
1260 rbd_dev->mapping.features = 0;
1263 static void zero_bvec(struct bio_vec *bv)
1266 unsigned long flags;
1268 buf = bvec_kmap_irq(bv, &flags);
1269 memset(buf, 0, bv->bv_len);
1270 flush_dcache_page(bv->bv_page);
1271 bvec_kunmap_irq(buf, &flags);
1274 static void zero_bios(struct ceph_bio_iter *bio_pos, u32 off, u32 bytes)
1276 struct ceph_bio_iter it = *bio_pos;
1278 ceph_bio_iter_advance(&it, off);
1279 ceph_bio_iter_advance_step(&it, bytes, ({
1284 static void zero_bvecs(struct ceph_bvec_iter *bvec_pos, u32 off, u32 bytes)
1286 struct ceph_bvec_iter it = *bvec_pos;
1288 ceph_bvec_iter_advance(&it, off);
1289 ceph_bvec_iter_advance_step(&it, bytes, ({
1295 * Zero a range in @obj_req data buffer defined by a bio (list) or
1296 * (private) bio_vec array.
1298 * @off is relative to the start of the data buffer.
1300 static void rbd_obj_zero_range(struct rbd_obj_request *obj_req, u32 off,
1303 switch (obj_req->img_request->data_type) {
1304 case OBJ_REQUEST_BIO:
1305 zero_bios(&obj_req->bio_pos, off, bytes);
1307 case OBJ_REQUEST_BVECS:
1308 case OBJ_REQUEST_OWN_BVECS:
1309 zero_bvecs(&obj_req->bvec_pos, off, bytes);
1316 static void rbd_obj_request_destroy(struct kref *kref);
1317 static void rbd_obj_request_put(struct rbd_obj_request *obj_request)
1319 rbd_assert(obj_request != NULL);
1320 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1321 kref_read(&obj_request->kref));
1322 kref_put(&obj_request->kref, rbd_obj_request_destroy);
1325 static void rbd_img_request_get(struct rbd_img_request *img_request)
1327 dout("%s: img %p (was %d)\n", __func__, img_request,
1328 kref_read(&img_request->kref));
1329 kref_get(&img_request->kref);
1332 static void rbd_img_request_destroy(struct kref *kref);
1333 static void rbd_img_request_put(struct rbd_img_request *img_request)
1335 rbd_assert(img_request != NULL);
1336 dout("%s: img %p (was %d)\n", __func__, img_request,
1337 kref_read(&img_request->kref));
1338 kref_put(&img_request->kref, rbd_img_request_destroy);
1341 static inline void rbd_img_obj_request_add(struct rbd_img_request *img_request,
1342 struct rbd_obj_request *obj_request)
1344 rbd_assert(obj_request->img_request == NULL);
1346 /* Image request now owns object's original reference */
1347 obj_request->img_request = img_request;
1348 img_request->obj_request_count++;
1349 img_request->pending_count++;
1350 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
1353 static inline void rbd_img_obj_request_del(struct rbd_img_request *img_request,
1354 struct rbd_obj_request *obj_request)
1356 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
1357 list_del(&obj_request->ex.oe_item);
1358 rbd_assert(img_request->obj_request_count > 0);
1359 img_request->obj_request_count--;
1360 rbd_assert(obj_request->img_request == img_request);
1361 rbd_obj_request_put(obj_request);
1364 static void rbd_obj_request_submit(struct rbd_obj_request *obj_request)
1366 struct ceph_osd_request *osd_req = obj_request->osd_req;
1368 dout("%s %p object_no %016llx %llu~%llu osd_req %p\n", __func__,
1369 obj_request, obj_request->ex.oe_objno, obj_request->ex.oe_off,
1370 obj_request->ex.oe_len, osd_req);
1371 ceph_osdc_start_request(osd_req->r_osdc, osd_req, false);
1375 * The default/initial value for all image request flags is 0. Each
1376 * is conditionally set to 1 at image request initialization time
1377 * and currently never change thereafter.
1379 static void img_request_layered_set(struct rbd_img_request *img_request)
1381 set_bit(IMG_REQ_LAYERED, &img_request->flags);
1385 static void img_request_layered_clear(struct rbd_img_request *img_request)
1387 clear_bit(IMG_REQ_LAYERED, &img_request->flags);
1391 static bool img_request_layered_test(struct rbd_img_request *img_request)
1394 return test_bit(IMG_REQ_LAYERED, &img_request->flags) != 0;
1397 static bool rbd_obj_is_entire(struct rbd_obj_request *obj_req)
1399 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1401 return !obj_req->ex.oe_off &&
1402 obj_req->ex.oe_len == rbd_dev->layout.object_size;
1405 static bool rbd_obj_is_tail(struct rbd_obj_request *obj_req)
1407 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1409 return obj_req->ex.oe_off + obj_req->ex.oe_len ==
1410 rbd_dev->layout.object_size;
1413 static u64 rbd_obj_img_extents_bytes(struct rbd_obj_request *obj_req)
1415 return ceph_file_extents_bytes(obj_req->img_extents,
1416 obj_req->num_img_extents);
1419 static bool rbd_img_is_write(struct rbd_img_request *img_req)
1421 switch (img_req->op_type) {
1425 case OBJ_OP_DISCARD:
1432 static void rbd_obj_handle_request(struct rbd_obj_request *obj_req);
1434 static void rbd_osd_req_callback(struct ceph_osd_request *osd_req)
1436 struct rbd_obj_request *obj_req = osd_req->r_priv;
1438 dout("%s osd_req %p result %d for obj_req %p\n", __func__, osd_req,
1439 osd_req->r_result, obj_req);
1440 rbd_assert(osd_req == obj_req->osd_req);
1442 obj_req->result = osd_req->r_result < 0 ? osd_req->r_result : 0;
1443 if (!obj_req->result && !rbd_img_is_write(obj_req->img_request))
1444 obj_req->xferred = osd_req->r_result;
1447 * Writes aren't allowed to return a data payload. In some
1448 * guarded write cases (e.g. stat + zero on an empty object)
1449 * a stat response makes it through, but we don't care.
1451 obj_req->xferred = 0;
1453 rbd_obj_handle_request(obj_req);
1456 static void rbd_osd_req_format_read(struct rbd_obj_request *obj_request)
1458 struct ceph_osd_request *osd_req = obj_request->osd_req;
1460 osd_req->r_flags = CEPH_OSD_FLAG_READ;
1461 osd_req->r_snapid = obj_request->img_request->snap_id;
1464 static void rbd_osd_req_format_write(struct rbd_obj_request *obj_request)
1466 struct ceph_osd_request *osd_req = obj_request->osd_req;
1468 osd_req->r_flags = CEPH_OSD_FLAG_WRITE;
1469 ktime_get_real_ts64(&osd_req->r_mtime);
1470 osd_req->r_data_offset = obj_request->ex.oe_off;
1473 static struct ceph_osd_request *
1474 rbd_osd_req_create(struct rbd_obj_request *obj_req, unsigned int num_ops)
1476 struct rbd_img_request *img_req = obj_req->img_request;
1477 struct rbd_device *rbd_dev = img_req->rbd_dev;
1478 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
1479 struct ceph_osd_request *req;
1480 const char *name_format = rbd_dev->image_format == 1 ?
1481 RBD_V1_DATA_FORMAT : RBD_V2_DATA_FORMAT;
1483 req = ceph_osdc_alloc_request(osdc,
1484 (rbd_img_is_write(img_req) ? img_req->snapc : NULL),
1485 num_ops, false, GFP_NOIO);
1489 req->r_callback = rbd_osd_req_callback;
1490 req->r_priv = obj_req;
1493 * Data objects may be stored in a separate pool, but always in
1494 * the same namespace in that pool as the header in its pool.
1496 ceph_oloc_copy(&req->r_base_oloc, &rbd_dev->header_oloc);
1497 req->r_base_oloc.pool = rbd_dev->layout.pool_id;
1499 if (ceph_oid_aprintf(&req->r_base_oid, GFP_NOIO, name_format,
1500 rbd_dev->header.object_prefix, obj_req->ex.oe_objno))
1503 if (ceph_osdc_alloc_messages(req, GFP_NOIO))
1509 ceph_osdc_put_request(req);
1513 static void rbd_osd_req_destroy(struct ceph_osd_request *osd_req)
1515 ceph_osdc_put_request(osd_req);
1518 static struct rbd_obj_request *rbd_obj_request_create(void)
1520 struct rbd_obj_request *obj_request;
1522 obj_request = kmem_cache_zalloc(rbd_obj_request_cache, GFP_NOIO);
1526 ceph_object_extent_init(&obj_request->ex);
1527 kref_init(&obj_request->kref);
1529 dout("%s %p\n", __func__, obj_request);
1533 static void rbd_obj_request_destroy(struct kref *kref)
1535 struct rbd_obj_request *obj_request;
1538 obj_request = container_of(kref, struct rbd_obj_request, kref);
1540 dout("%s: obj %p\n", __func__, obj_request);
1542 if (obj_request->osd_req)
1543 rbd_osd_req_destroy(obj_request->osd_req);
1545 switch (obj_request->img_request->data_type) {
1546 case OBJ_REQUEST_NODATA:
1547 case OBJ_REQUEST_BIO:
1548 case OBJ_REQUEST_BVECS:
1549 break; /* Nothing to do */
1550 case OBJ_REQUEST_OWN_BVECS:
1551 kfree(obj_request->bvec_pos.bvecs);
1557 kfree(obj_request->img_extents);
1558 if (obj_request->copyup_bvecs) {
1559 for (i = 0; i < obj_request->copyup_bvec_count; i++) {
1560 if (obj_request->copyup_bvecs[i].bv_page)
1561 __free_page(obj_request->copyup_bvecs[i].bv_page);
1563 kfree(obj_request->copyup_bvecs);
1566 kmem_cache_free(rbd_obj_request_cache, obj_request);
1569 /* It's OK to call this for a device with no parent */
1571 static void rbd_spec_put(struct rbd_spec *spec);
1572 static void rbd_dev_unparent(struct rbd_device *rbd_dev)
1574 rbd_dev_remove_parent(rbd_dev);
1575 rbd_spec_put(rbd_dev->parent_spec);
1576 rbd_dev->parent_spec = NULL;
1577 rbd_dev->parent_overlap = 0;
1581 * Parent image reference counting is used to determine when an
1582 * image's parent fields can be safely torn down--after there are no
1583 * more in-flight requests to the parent image. When the last
1584 * reference is dropped, cleaning them up is safe.
1586 static void rbd_dev_parent_put(struct rbd_device *rbd_dev)
1590 if (!rbd_dev->parent_spec)
1593 counter = atomic_dec_return_safe(&rbd_dev->parent_ref);
1597 /* Last reference; clean up parent data structures */
1600 rbd_dev_unparent(rbd_dev);
1602 rbd_warn(rbd_dev, "parent reference underflow");
1606 * If an image has a non-zero parent overlap, get a reference to its
1609 * Returns true if the rbd device has a parent with a non-zero
1610 * overlap and a reference for it was successfully taken, or
1613 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
1617 if (!rbd_dev->parent_spec)
1620 down_read(&rbd_dev->header_rwsem);
1621 if (rbd_dev->parent_overlap)
1622 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
1623 up_read(&rbd_dev->header_rwsem);
1626 rbd_warn(rbd_dev, "parent reference overflow");
1632 * Caller is responsible for filling in the list of object requests
1633 * that comprises the image request, and the Linux request pointer
1634 * (if there is one).
1636 static struct rbd_img_request *rbd_img_request_create(
1637 struct rbd_device *rbd_dev,
1638 enum obj_operation_type op_type,
1639 struct ceph_snap_context *snapc)
1641 struct rbd_img_request *img_request;
1643 img_request = kmem_cache_zalloc(rbd_img_request_cache, GFP_NOIO);
1647 img_request->rbd_dev = rbd_dev;
1648 img_request->op_type = op_type;
1649 if (!rbd_img_is_write(img_request))
1650 img_request->snap_id = rbd_dev->spec->snap_id;
1652 img_request->snapc = snapc;
1654 if (rbd_dev_parent_get(rbd_dev))
1655 img_request_layered_set(img_request);
1657 spin_lock_init(&img_request->completion_lock);
1658 INIT_LIST_HEAD(&img_request->object_extents);
1659 kref_init(&img_request->kref);
1661 dout("%s: rbd_dev %p %s -> img %p\n", __func__, rbd_dev,
1662 obj_op_name(op_type), img_request);
1666 static void rbd_img_request_destroy(struct kref *kref)
1668 struct rbd_img_request *img_request;
1669 struct rbd_obj_request *obj_request;
1670 struct rbd_obj_request *next_obj_request;
1672 img_request = container_of(kref, struct rbd_img_request, kref);
1674 dout("%s: img %p\n", __func__, img_request);
1676 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
1677 rbd_img_obj_request_del(img_request, obj_request);
1678 rbd_assert(img_request->obj_request_count == 0);
1680 if (img_request_layered_test(img_request)) {
1681 img_request_layered_clear(img_request);
1682 rbd_dev_parent_put(img_request->rbd_dev);
1685 if (rbd_img_is_write(img_request))
1686 ceph_put_snap_context(img_request->snapc);
1688 kmem_cache_free(rbd_img_request_cache, img_request);
1691 static void prune_extents(struct ceph_file_extent *img_extents,
1692 u32 *num_img_extents, u64 overlap)
1694 u32 cnt = *num_img_extents;
1696 /* drop extents completely beyond the overlap */
1697 while (cnt && img_extents[cnt - 1].fe_off >= overlap)
1701 struct ceph_file_extent *ex = &img_extents[cnt - 1];
1703 /* trim final overlapping extent */
1704 if (ex->fe_off + ex->fe_len > overlap)
1705 ex->fe_len = overlap - ex->fe_off;
1708 *num_img_extents = cnt;
1712 * Determine the byte range(s) covered by either just the object extent
1713 * or the entire object in the parent image.
1715 static int rbd_obj_calc_img_extents(struct rbd_obj_request *obj_req,
1718 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1721 if (!rbd_dev->parent_overlap)
1724 ret = ceph_extent_to_file(&rbd_dev->layout, obj_req->ex.oe_objno,
1725 entire ? 0 : obj_req->ex.oe_off,
1726 entire ? rbd_dev->layout.object_size :
1728 &obj_req->img_extents,
1729 &obj_req->num_img_extents);
1733 prune_extents(obj_req->img_extents, &obj_req->num_img_extents,
1734 rbd_dev->parent_overlap);
1738 static void rbd_osd_req_setup_data(struct rbd_obj_request *obj_req, u32 which)
1740 switch (obj_req->img_request->data_type) {
1741 case OBJ_REQUEST_BIO:
1742 osd_req_op_extent_osd_data_bio(obj_req->osd_req, which,
1744 obj_req->ex.oe_len);
1746 case OBJ_REQUEST_BVECS:
1747 case OBJ_REQUEST_OWN_BVECS:
1748 rbd_assert(obj_req->bvec_pos.iter.bi_size ==
1749 obj_req->ex.oe_len);
1750 rbd_assert(obj_req->bvec_idx == obj_req->bvec_count);
1751 osd_req_op_extent_osd_data_bvec_pos(obj_req->osd_req, which,
1752 &obj_req->bvec_pos);
1759 static int rbd_obj_setup_read(struct rbd_obj_request *obj_req)
1761 obj_req->osd_req = rbd_osd_req_create(obj_req, 1);
1762 if (!obj_req->osd_req)
1765 osd_req_op_extent_init(obj_req->osd_req, 0, CEPH_OSD_OP_READ,
1766 obj_req->ex.oe_off, obj_req->ex.oe_len, 0, 0);
1767 rbd_osd_req_setup_data(obj_req, 0);
1769 rbd_osd_req_format_read(obj_req);
1773 static int __rbd_obj_setup_stat(struct rbd_obj_request *obj_req,
1776 struct page **pages;
1779 * The response data for a STAT call consists of:
1786 pages = ceph_alloc_page_vector(1, GFP_NOIO);
1788 return PTR_ERR(pages);
1790 osd_req_op_init(obj_req->osd_req, which, CEPH_OSD_OP_STAT, 0);
1791 osd_req_op_raw_data_in_pages(obj_req->osd_req, which, pages,
1792 8 + sizeof(struct ceph_timespec),
1797 static void __rbd_obj_setup_write(struct rbd_obj_request *obj_req,
1800 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1803 osd_req_op_alloc_hint_init(obj_req->osd_req, which++,
1804 rbd_dev->layout.object_size,
1805 rbd_dev->layout.object_size);
1807 if (rbd_obj_is_entire(obj_req))
1808 opcode = CEPH_OSD_OP_WRITEFULL;
1810 opcode = CEPH_OSD_OP_WRITE;
1812 osd_req_op_extent_init(obj_req->osd_req, which, opcode,
1813 obj_req->ex.oe_off, obj_req->ex.oe_len, 0, 0);
1814 rbd_osd_req_setup_data(obj_req, which++);
1816 rbd_assert(which == obj_req->osd_req->r_num_ops);
1817 rbd_osd_req_format_write(obj_req);
1820 static int rbd_obj_setup_write(struct rbd_obj_request *obj_req)
1822 unsigned int num_osd_ops, which = 0;
1825 /* reverse map the entire object onto the parent */
1826 ret = rbd_obj_calc_img_extents(obj_req, true);
1830 if (obj_req->num_img_extents) {
1831 obj_req->write_state = RBD_OBJ_WRITE_GUARD;
1832 num_osd_ops = 3; /* stat + setallochint + write/writefull */
1834 obj_req->write_state = RBD_OBJ_WRITE_FLAT;
1835 num_osd_ops = 2; /* setallochint + write/writefull */
1838 obj_req->osd_req = rbd_osd_req_create(obj_req, num_osd_ops);
1839 if (!obj_req->osd_req)
1842 if (obj_req->num_img_extents) {
1843 ret = __rbd_obj_setup_stat(obj_req, which++);
1848 __rbd_obj_setup_write(obj_req, which);
1852 static void __rbd_obj_setup_discard(struct rbd_obj_request *obj_req,
1857 if (rbd_obj_is_entire(obj_req)) {
1858 if (obj_req->num_img_extents) {
1859 osd_req_op_init(obj_req->osd_req, which++,
1860 CEPH_OSD_OP_CREATE, 0);
1861 opcode = CEPH_OSD_OP_TRUNCATE;
1863 osd_req_op_init(obj_req->osd_req, which++,
1864 CEPH_OSD_OP_DELETE, 0);
1867 } else if (rbd_obj_is_tail(obj_req)) {
1868 opcode = CEPH_OSD_OP_TRUNCATE;
1870 opcode = CEPH_OSD_OP_ZERO;
1874 osd_req_op_extent_init(obj_req->osd_req, which++, opcode,
1875 obj_req->ex.oe_off, obj_req->ex.oe_len,
1878 rbd_assert(which == obj_req->osd_req->r_num_ops);
1879 rbd_osd_req_format_write(obj_req);
1882 static int rbd_obj_setup_discard(struct rbd_obj_request *obj_req)
1884 unsigned int num_osd_ops, which = 0;
1887 /* reverse map the entire object onto the parent */
1888 ret = rbd_obj_calc_img_extents(obj_req, true);
1892 if (rbd_obj_is_entire(obj_req)) {
1893 obj_req->write_state = RBD_OBJ_WRITE_FLAT;
1894 if (obj_req->num_img_extents)
1895 num_osd_ops = 2; /* create + truncate */
1897 num_osd_ops = 1; /* delete */
1899 if (obj_req->num_img_extents) {
1900 obj_req->write_state = RBD_OBJ_WRITE_GUARD;
1901 num_osd_ops = 2; /* stat + truncate/zero */
1903 obj_req->write_state = RBD_OBJ_WRITE_FLAT;
1904 num_osd_ops = 1; /* truncate/zero */
1908 obj_req->osd_req = rbd_osd_req_create(obj_req, num_osd_ops);
1909 if (!obj_req->osd_req)
1912 if (!rbd_obj_is_entire(obj_req) && obj_req->num_img_extents) {
1913 ret = __rbd_obj_setup_stat(obj_req, which++);
1918 __rbd_obj_setup_discard(obj_req, which);
1923 * For each object request in @img_req, allocate an OSD request, add
1924 * individual OSD ops and prepare them for submission. The number of
1925 * OSD ops depends on op_type and the overlap point (if any).
1927 static int __rbd_img_fill_request(struct rbd_img_request *img_req)
1929 struct rbd_obj_request *obj_req;
1932 for_each_obj_request(img_req, obj_req) {
1933 switch (img_req->op_type) {
1935 ret = rbd_obj_setup_read(obj_req);
1938 ret = rbd_obj_setup_write(obj_req);
1940 case OBJ_OP_DISCARD:
1941 ret = rbd_obj_setup_discard(obj_req);
1953 union rbd_img_fill_iter {
1954 struct ceph_bio_iter bio_iter;
1955 struct ceph_bvec_iter bvec_iter;
1958 struct rbd_img_fill_ctx {
1959 enum obj_request_type pos_type;
1960 union rbd_img_fill_iter *pos;
1961 union rbd_img_fill_iter iter;
1962 ceph_object_extent_fn_t set_pos_fn;
1963 ceph_object_extent_fn_t count_fn;
1964 ceph_object_extent_fn_t copy_fn;
1967 static struct ceph_object_extent *alloc_object_extent(void *arg)
1969 struct rbd_img_request *img_req = arg;
1970 struct rbd_obj_request *obj_req;
1972 obj_req = rbd_obj_request_create();
1976 rbd_img_obj_request_add(img_req, obj_req);
1977 return &obj_req->ex;
1981 * While su != os && sc == 1 is technically not fancy (it's the same
1982 * layout as su == os && sc == 1), we can't use the nocopy path for it
1983 * because ->set_pos_fn() should be called only once per object.
1984 * ceph_file_to_extents() invokes action_fn once per stripe unit, so
1985 * treat su != os && sc == 1 as fancy.
1987 static bool rbd_layout_is_fancy(struct ceph_file_layout *l)
1989 return l->stripe_unit != l->object_size;
1992 static int rbd_img_fill_request_nocopy(struct rbd_img_request *img_req,
1993 struct ceph_file_extent *img_extents,
1994 u32 num_img_extents,
1995 struct rbd_img_fill_ctx *fctx)
2000 img_req->data_type = fctx->pos_type;
2003 * Create object requests and set each object request's starting
2004 * position in the provided bio (list) or bio_vec array.
2006 fctx->iter = *fctx->pos;
2007 for (i = 0; i < num_img_extents; i++) {
2008 ret = ceph_file_to_extents(&img_req->rbd_dev->layout,
2009 img_extents[i].fe_off,
2010 img_extents[i].fe_len,
2011 &img_req->object_extents,
2012 alloc_object_extent, img_req,
2013 fctx->set_pos_fn, &fctx->iter);
2018 return __rbd_img_fill_request(img_req);
2022 * Map a list of image extents to a list of object extents, create the
2023 * corresponding object requests (normally each to a different object,
2024 * but not always) and add them to @img_req. For each object request,
2025 * set up its data descriptor to point to the corresponding chunk(s) of
2026 * @fctx->pos data buffer.
2028 * Because ceph_file_to_extents() will merge adjacent object extents
2029 * together, each object request's data descriptor may point to multiple
2030 * different chunks of @fctx->pos data buffer.
2032 * @fctx->pos data buffer is assumed to be large enough.
2034 static int rbd_img_fill_request(struct rbd_img_request *img_req,
2035 struct ceph_file_extent *img_extents,
2036 u32 num_img_extents,
2037 struct rbd_img_fill_ctx *fctx)
2039 struct rbd_device *rbd_dev = img_req->rbd_dev;
2040 struct rbd_obj_request *obj_req;
2044 if (fctx->pos_type == OBJ_REQUEST_NODATA ||
2045 !rbd_layout_is_fancy(&rbd_dev->layout))
2046 return rbd_img_fill_request_nocopy(img_req, img_extents,
2047 num_img_extents, fctx);
2049 img_req->data_type = OBJ_REQUEST_OWN_BVECS;
2052 * Create object requests and determine ->bvec_count for each object
2053 * request. Note that ->bvec_count sum over all object requests may
2054 * be greater than the number of bio_vecs in the provided bio (list)
2055 * or bio_vec array because when mapped, those bio_vecs can straddle
2056 * stripe unit boundaries.
2058 fctx->iter = *fctx->pos;
2059 for (i = 0; i < num_img_extents; i++) {
2060 ret = ceph_file_to_extents(&rbd_dev->layout,
2061 img_extents[i].fe_off,
2062 img_extents[i].fe_len,
2063 &img_req->object_extents,
2064 alloc_object_extent, img_req,
2065 fctx->count_fn, &fctx->iter);
2070 for_each_obj_request(img_req, obj_req) {
2071 obj_req->bvec_pos.bvecs = kmalloc_array(obj_req->bvec_count,
2072 sizeof(*obj_req->bvec_pos.bvecs),
2074 if (!obj_req->bvec_pos.bvecs)
2079 * Fill in each object request's private bio_vec array, splitting and
2080 * rearranging the provided bio_vecs in stripe unit chunks as needed.
2082 fctx->iter = *fctx->pos;
2083 for (i = 0; i < num_img_extents; i++) {
2084 ret = ceph_iterate_extents(&rbd_dev->layout,
2085 img_extents[i].fe_off,
2086 img_extents[i].fe_len,
2087 &img_req->object_extents,
2088 fctx->copy_fn, &fctx->iter);
2093 return __rbd_img_fill_request(img_req);
2096 static int rbd_img_fill_nodata(struct rbd_img_request *img_req,
2099 struct ceph_file_extent ex = { off, len };
2100 union rbd_img_fill_iter dummy = {};
2101 struct rbd_img_fill_ctx fctx = {
2102 .pos_type = OBJ_REQUEST_NODATA,
2106 return rbd_img_fill_request(img_req, &ex, 1, &fctx);
2109 static void set_bio_pos(struct ceph_object_extent *ex, u32 bytes, void *arg)
2111 struct rbd_obj_request *obj_req =
2112 container_of(ex, struct rbd_obj_request, ex);
2113 struct ceph_bio_iter *it = arg;
2115 dout("%s objno %llu bytes %u\n", __func__, ex->oe_objno, bytes);
2116 obj_req->bio_pos = *it;
2117 ceph_bio_iter_advance(it, bytes);
2120 static void count_bio_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2122 struct rbd_obj_request *obj_req =
2123 container_of(ex, struct rbd_obj_request, ex);
2124 struct ceph_bio_iter *it = arg;
2126 dout("%s objno %llu bytes %u\n", __func__, ex->oe_objno, bytes);
2127 ceph_bio_iter_advance_step(it, bytes, ({
2128 obj_req->bvec_count++;
2133 static void copy_bio_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2135 struct rbd_obj_request *obj_req =
2136 container_of(ex, struct rbd_obj_request, ex);
2137 struct ceph_bio_iter *it = arg;
2139 dout("%s objno %llu bytes %u\n", __func__, ex->oe_objno, bytes);
2140 ceph_bio_iter_advance_step(it, bytes, ({
2141 obj_req->bvec_pos.bvecs[obj_req->bvec_idx++] = bv;
2142 obj_req->bvec_pos.iter.bi_size += bv.bv_len;
2146 static int __rbd_img_fill_from_bio(struct rbd_img_request *img_req,
2147 struct ceph_file_extent *img_extents,
2148 u32 num_img_extents,
2149 struct ceph_bio_iter *bio_pos)
2151 struct rbd_img_fill_ctx fctx = {
2152 .pos_type = OBJ_REQUEST_BIO,
2153 .pos = (union rbd_img_fill_iter *)bio_pos,
2154 .set_pos_fn = set_bio_pos,
2155 .count_fn = count_bio_bvecs,
2156 .copy_fn = copy_bio_bvecs,
2159 return rbd_img_fill_request(img_req, img_extents, num_img_extents,
2163 static int rbd_img_fill_from_bio(struct rbd_img_request *img_req,
2164 u64 off, u64 len, struct bio *bio)
2166 struct ceph_file_extent ex = { off, len };
2167 struct ceph_bio_iter it = { .bio = bio, .iter = bio->bi_iter };
2169 return __rbd_img_fill_from_bio(img_req, &ex, 1, &it);
2172 static void set_bvec_pos(struct ceph_object_extent *ex, u32 bytes, void *arg)
2174 struct rbd_obj_request *obj_req =
2175 container_of(ex, struct rbd_obj_request, ex);
2176 struct ceph_bvec_iter *it = arg;
2178 obj_req->bvec_pos = *it;
2179 ceph_bvec_iter_shorten(&obj_req->bvec_pos, bytes);
2180 ceph_bvec_iter_advance(it, bytes);
2183 static void count_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2185 struct rbd_obj_request *obj_req =
2186 container_of(ex, struct rbd_obj_request, ex);
2187 struct ceph_bvec_iter *it = arg;
2189 ceph_bvec_iter_advance_step(it, bytes, ({
2190 obj_req->bvec_count++;
2194 static void copy_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2196 struct rbd_obj_request *obj_req =
2197 container_of(ex, struct rbd_obj_request, ex);
2198 struct ceph_bvec_iter *it = arg;
2200 ceph_bvec_iter_advance_step(it, bytes, ({
2201 obj_req->bvec_pos.bvecs[obj_req->bvec_idx++] = bv;
2202 obj_req->bvec_pos.iter.bi_size += bv.bv_len;
2206 static int __rbd_img_fill_from_bvecs(struct rbd_img_request *img_req,
2207 struct ceph_file_extent *img_extents,
2208 u32 num_img_extents,
2209 struct ceph_bvec_iter *bvec_pos)
2211 struct rbd_img_fill_ctx fctx = {
2212 .pos_type = OBJ_REQUEST_BVECS,
2213 .pos = (union rbd_img_fill_iter *)bvec_pos,
2214 .set_pos_fn = set_bvec_pos,
2215 .count_fn = count_bvecs,
2216 .copy_fn = copy_bvecs,
2219 return rbd_img_fill_request(img_req, img_extents, num_img_extents,
2223 static int rbd_img_fill_from_bvecs(struct rbd_img_request *img_req,
2224 struct ceph_file_extent *img_extents,
2225 u32 num_img_extents,
2226 struct bio_vec *bvecs)
2228 struct ceph_bvec_iter it = {
2230 .iter = { .bi_size = ceph_file_extents_bytes(img_extents,
2234 return __rbd_img_fill_from_bvecs(img_req, img_extents, num_img_extents,
2238 static void rbd_img_request_submit(struct rbd_img_request *img_request)
2240 struct rbd_obj_request *obj_request;
2242 dout("%s: img %p\n", __func__, img_request);
2244 rbd_img_request_get(img_request);
2245 for_each_obj_request(img_request, obj_request)
2246 rbd_obj_request_submit(obj_request);
2248 rbd_img_request_put(img_request);
2251 static int rbd_obj_read_from_parent(struct rbd_obj_request *obj_req)
2253 struct rbd_img_request *img_req = obj_req->img_request;
2254 struct rbd_img_request *child_img_req;
2257 child_img_req = rbd_img_request_create(img_req->rbd_dev->parent,
2262 __set_bit(IMG_REQ_CHILD, &child_img_req->flags);
2263 child_img_req->obj_request = obj_req;
2265 if (!rbd_img_is_write(img_req)) {
2266 switch (img_req->data_type) {
2267 case OBJ_REQUEST_BIO:
2268 ret = __rbd_img_fill_from_bio(child_img_req,
2269 obj_req->img_extents,
2270 obj_req->num_img_extents,
2273 case OBJ_REQUEST_BVECS:
2274 case OBJ_REQUEST_OWN_BVECS:
2275 ret = __rbd_img_fill_from_bvecs(child_img_req,
2276 obj_req->img_extents,
2277 obj_req->num_img_extents,
2278 &obj_req->bvec_pos);
2284 ret = rbd_img_fill_from_bvecs(child_img_req,
2285 obj_req->img_extents,
2286 obj_req->num_img_extents,
2287 obj_req->copyup_bvecs);
2290 rbd_img_request_put(child_img_req);
2294 rbd_img_request_submit(child_img_req);
2298 static bool rbd_obj_handle_read(struct rbd_obj_request *obj_req)
2300 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2303 if (obj_req->result == -ENOENT &&
2304 rbd_dev->parent_overlap && !obj_req->tried_parent) {
2305 /* reverse map this object extent onto the parent */
2306 ret = rbd_obj_calc_img_extents(obj_req, false);
2308 obj_req->result = ret;
2312 if (obj_req->num_img_extents) {
2313 obj_req->tried_parent = true;
2314 ret = rbd_obj_read_from_parent(obj_req);
2316 obj_req->result = ret;
2324 * -ENOENT means a hole in the image -- zero-fill the entire
2325 * length of the request. A short read also implies zero-fill
2326 * to the end of the request. In both cases we update xferred
2327 * count to indicate the whole request was satisfied.
2329 if (obj_req->result == -ENOENT ||
2330 (!obj_req->result && obj_req->xferred < obj_req->ex.oe_len)) {
2331 rbd_assert(!obj_req->xferred || !obj_req->result);
2332 rbd_obj_zero_range(obj_req, obj_req->xferred,
2333 obj_req->ex.oe_len - obj_req->xferred);
2334 obj_req->result = 0;
2335 obj_req->xferred = obj_req->ex.oe_len;
2342 * copyup_bvecs pages are never highmem pages
2344 static bool is_zero_bvecs(struct bio_vec *bvecs, u32 bytes)
2346 struct ceph_bvec_iter it = {
2348 .iter = { .bi_size = bytes },
2351 ceph_bvec_iter_advance_step(&it, bytes, ({
2352 if (memchr_inv(page_address(bv.bv_page) + bv.bv_offset, 0,
2359 static int rbd_obj_issue_copyup(struct rbd_obj_request *obj_req, u32 bytes)
2361 unsigned int num_osd_ops = obj_req->osd_req->r_num_ops;
2364 dout("%s obj_req %p bytes %u\n", __func__, obj_req, bytes);
2365 rbd_assert(obj_req->osd_req->r_ops[0].op == CEPH_OSD_OP_STAT);
2366 rbd_osd_req_destroy(obj_req->osd_req);
2369 * Create a copyup request with the same number of OSD ops as
2370 * the original request. The original request was stat + op(s),
2371 * the new copyup request will be copyup + the same op(s).
2373 obj_req->osd_req = rbd_osd_req_create(obj_req, num_osd_ops);
2374 if (!obj_req->osd_req)
2377 ret = osd_req_op_cls_init(obj_req->osd_req, 0, CEPH_OSD_OP_CALL, "rbd",
2383 * Only send non-zero copyup data to save some I/O and network
2384 * bandwidth -- zero copyup data is equivalent to the object not
2387 if (is_zero_bvecs(obj_req->copyup_bvecs, bytes)) {
2388 dout("%s obj_req %p detected zeroes\n", __func__, obj_req);
2391 osd_req_op_cls_request_data_bvecs(obj_req->osd_req, 0,
2392 obj_req->copyup_bvecs,
2393 obj_req->copyup_bvec_count,
2396 switch (obj_req->img_request->op_type) {
2398 __rbd_obj_setup_write(obj_req, 1);
2400 case OBJ_OP_DISCARD:
2401 rbd_assert(!rbd_obj_is_entire(obj_req));
2402 __rbd_obj_setup_discard(obj_req, 1);
2408 rbd_obj_request_submit(obj_req);
2412 static int setup_copyup_bvecs(struct rbd_obj_request *obj_req, u64 obj_overlap)
2416 rbd_assert(!obj_req->copyup_bvecs);
2417 obj_req->copyup_bvec_count = calc_pages_for(0, obj_overlap);
2418 obj_req->copyup_bvecs = kcalloc(obj_req->copyup_bvec_count,
2419 sizeof(*obj_req->copyup_bvecs),
2421 if (!obj_req->copyup_bvecs)
2424 for (i = 0; i < obj_req->copyup_bvec_count; i++) {
2425 unsigned int len = min(obj_overlap, (u64)PAGE_SIZE);
2427 obj_req->copyup_bvecs[i].bv_page = alloc_page(GFP_NOIO);
2428 if (!obj_req->copyup_bvecs[i].bv_page)
2431 obj_req->copyup_bvecs[i].bv_offset = 0;
2432 obj_req->copyup_bvecs[i].bv_len = len;
2436 rbd_assert(!obj_overlap);
2440 static int rbd_obj_handle_write_guard(struct rbd_obj_request *obj_req)
2442 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2445 rbd_assert(obj_req->num_img_extents);
2446 prune_extents(obj_req->img_extents, &obj_req->num_img_extents,
2447 rbd_dev->parent_overlap);
2448 if (!obj_req->num_img_extents) {
2450 * The overlap has become 0 (most likely because the
2451 * image has been flattened). Use rbd_obj_issue_copyup()
2452 * to re-submit the original write request -- the copyup
2453 * operation itself will be a no-op, since someone must
2454 * have populated the child object while we weren't
2455 * looking. Move to WRITE_FLAT state as we'll be done
2456 * with the operation once the null copyup completes.
2458 obj_req->write_state = RBD_OBJ_WRITE_FLAT;
2459 return rbd_obj_issue_copyup(obj_req, 0);
2462 ret = setup_copyup_bvecs(obj_req, rbd_obj_img_extents_bytes(obj_req));
2466 obj_req->write_state = RBD_OBJ_WRITE_COPYUP;
2467 return rbd_obj_read_from_parent(obj_req);
2470 static bool rbd_obj_handle_write(struct rbd_obj_request *obj_req)
2475 switch (obj_req->write_state) {
2476 case RBD_OBJ_WRITE_GUARD:
2477 rbd_assert(!obj_req->xferred);
2478 if (obj_req->result == -ENOENT) {
2480 * The target object doesn't exist. Read the data for
2481 * the entire target object up to the overlap point (if
2482 * any) from the parent, so we can use it for a copyup.
2484 ret = rbd_obj_handle_write_guard(obj_req);
2486 obj_req->result = ret;
2492 case RBD_OBJ_WRITE_FLAT:
2493 if (!obj_req->result)
2495 * There is no such thing as a successful short
2496 * write -- indicate the whole request was satisfied.
2498 obj_req->xferred = obj_req->ex.oe_len;
2500 case RBD_OBJ_WRITE_COPYUP:
2501 obj_req->write_state = RBD_OBJ_WRITE_GUARD;
2502 if (obj_req->result)
2505 rbd_assert(obj_req->xferred);
2506 ret = rbd_obj_issue_copyup(obj_req, obj_req->xferred);
2508 obj_req->result = ret;
2509 obj_req->xferred = 0;
2519 * Returns true if @obj_req is completed, or false otherwise.
2521 static bool __rbd_obj_handle_request(struct rbd_obj_request *obj_req)
2523 switch (obj_req->img_request->op_type) {
2525 return rbd_obj_handle_read(obj_req);
2527 return rbd_obj_handle_write(obj_req);
2528 case OBJ_OP_DISCARD:
2529 if (rbd_obj_handle_write(obj_req)) {
2531 * Hide -ENOENT from delete/truncate/zero -- discarding
2532 * a non-existent object is not a problem.
2534 if (obj_req->result == -ENOENT) {
2535 obj_req->result = 0;
2536 obj_req->xferred = obj_req->ex.oe_len;
2546 static void rbd_obj_end_request(struct rbd_obj_request *obj_req)
2548 struct rbd_img_request *img_req = obj_req->img_request;
2550 rbd_assert((!obj_req->result &&
2551 obj_req->xferred == obj_req->ex.oe_len) ||
2552 (obj_req->result < 0 && !obj_req->xferred));
2553 if (!obj_req->result) {
2554 img_req->xferred += obj_req->xferred;
2558 rbd_warn(img_req->rbd_dev,
2559 "%s at objno %llu %llu~%llu result %d xferred %llu",
2560 obj_op_name(img_req->op_type), obj_req->ex.oe_objno,
2561 obj_req->ex.oe_off, obj_req->ex.oe_len, obj_req->result,
2563 if (!img_req->result) {
2564 img_req->result = obj_req->result;
2565 img_req->xferred = 0;
2569 static void rbd_img_end_child_request(struct rbd_img_request *img_req)
2571 struct rbd_obj_request *obj_req = img_req->obj_request;
2573 rbd_assert(test_bit(IMG_REQ_CHILD, &img_req->flags));
2574 rbd_assert((!img_req->result &&
2575 img_req->xferred == rbd_obj_img_extents_bytes(obj_req)) ||
2576 (img_req->result < 0 && !img_req->xferred));
2578 obj_req->result = img_req->result;
2579 obj_req->xferred = img_req->xferred;
2580 rbd_img_request_put(img_req);
2583 static void rbd_img_end_request(struct rbd_img_request *img_req)
2585 rbd_assert(!test_bit(IMG_REQ_CHILD, &img_req->flags));
2586 rbd_assert((!img_req->result &&
2587 img_req->xferred == blk_rq_bytes(img_req->rq)) ||
2588 (img_req->result < 0 && !img_req->xferred));
2590 blk_mq_end_request(img_req->rq,
2591 errno_to_blk_status(img_req->result));
2592 rbd_img_request_put(img_req);
2595 static void rbd_obj_handle_request(struct rbd_obj_request *obj_req)
2597 struct rbd_img_request *img_req;
2600 if (!__rbd_obj_handle_request(obj_req))
2603 img_req = obj_req->img_request;
2604 spin_lock(&img_req->completion_lock);
2605 rbd_obj_end_request(obj_req);
2606 rbd_assert(img_req->pending_count);
2607 if (--img_req->pending_count) {
2608 spin_unlock(&img_req->completion_lock);
2612 spin_unlock(&img_req->completion_lock);
2613 if (test_bit(IMG_REQ_CHILD, &img_req->flags)) {
2614 obj_req = img_req->obj_request;
2615 rbd_img_end_child_request(img_req);
2618 rbd_img_end_request(img_req);
2621 static const struct rbd_client_id rbd_empty_cid;
2623 static bool rbd_cid_equal(const struct rbd_client_id *lhs,
2624 const struct rbd_client_id *rhs)
2626 return lhs->gid == rhs->gid && lhs->handle == rhs->handle;
2629 static struct rbd_client_id rbd_get_cid(struct rbd_device *rbd_dev)
2631 struct rbd_client_id cid;
2633 mutex_lock(&rbd_dev->watch_mutex);
2634 cid.gid = ceph_client_gid(rbd_dev->rbd_client->client);
2635 cid.handle = rbd_dev->watch_cookie;
2636 mutex_unlock(&rbd_dev->watch_mutex);
2641 * lock_rwsem must be held for write
2643 static void rbd_set_owner_cid(struct rbd_device *rbd_dev,
2644 const struct rbd_client_id *cid)
2646 dout("%s rbd_dev %p %llu-%llu -> %llu-%llu\n", __func__, rbd_dev,
2647 rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle,
2648 cid->gid, cid->handle);
2649 rbd_dev->owner_cid = *cid; /* struct */
2652 static void format_lock_cookie(struct rbd_device *rbd_dev, char *buf)
2654 mutex_lock(&rbd_dev->watch_mutex);
2655 sprintf(buf, "%s %llu", RBD_LOCK_COOKIE_PREFIX, rbd_dev->watch_cookie);
2656 mutex_unlock(&rbd_dev->watch_mutex);
2659 static void __rbd_lock(struct rbd_device *rbd_dev, const char *cookie)
2661 struct rbd_client_id cid = rbd_get_cid(rbd_dev);
2663 strcpy(rbd_dev->lock_cookie, cookie);
2664 rbd_set_owner_cid(rbd_dev, &cid);
2665 queue_work(rbd_dev->task_wq, &rbd_dev->acquired_lock_work);
2669 * lock_rwsem must be held for write
2671 static int rbd_lock(struct rbd_device *rbd_dev)
2673 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2677 WARN_ON(__rbd_is_lock_owner(rbd_dev) ||
2678 rbd_dev->lock_cookie[0] != '\0');
2680 format_lock_cookie(rbd_dev, cookie);
2681 ret = ceph_cls_lock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
2682 RBD_LOCK_NAME, CEPH_CLS_LOCK_EXCLUSIVE, cookie,
2683 RBD_LOCK_TAG, "", 0);
2687 rbd_dev->lock_state = RBD_LOCK_STATE_LOCKED;
2688 __rbd_lock(rbd_dev, cookie);
2693 * lock_rwsem must be held for write
2695 static void rbd_unlock(struct rbd_device *rbd_dev)
2697 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2700 WARN_ON(!__rbd_is_lock_owner(rbd_dev) ||
2701 rbd_dev->lock_cookie[0] == '\0');
2703 ret = ceph_cls_unlock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
2704 RBD_LOCK_NAME, rbd_dev->lock_cookie);
2705 if (ret && ret != -ENOENT)
2706 rbd_warn(rbd_dev, "failed to unlock: %d", ret);
2708 /* treat errors as the image is unlocked */
2709 rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
2710 rbd_dev->lock_cookie[0] = '\0';
2711 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
2712 queue_work(rbd_dev->task_wq, &rbd_dev->released_lock_work);
2715 static int __rbd_notify_op_lock(struct rbd_device *rbd_dev,
2716 enum rbd_notify_op notify_op,
2717 struct page ***preply_pages,
2720 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2721 struct rbd_client_id cid = rbd_get_cid(rbd_dev);
2722 char buf[4 + 8 + 8 + CEPH_ENCODING_START_BLK_LEN];
2723 int buf_size = sizeof(buf);
2726 dout("%s rbd_dev %p notify_op %d\n", __func__, rbd_dev, notify_op);
2728 /* encode *LockPayload NotifyMessage (op + ClientId) */
2729 ceph_start_encoding(&p, 2, 1, buf_size - CEPH_ENCODING_START_BLK_LEN);
2730 ceph_encode_32(&p, notify_op);
2731 ceph_encode_64(&p, cid.gid);
2732 ceph_encode_64(&p, cid.handle);
2734 return ceph_osdc_notify(osdc, &rbd_dev->header_oid,
2735 &rbd_dev->header_oloc, buf, buf_size,
2736 RBD_NOTIFY_TIMEOUT, preply_pages, preply_len);
2739 static void rbd_notify_op_lock(struct rbd_device *rbd_dev,
2740 enum rbd_notify_op notify_op)
2742 struct page **reply_pages;
2745 __rbd_notify_op_lock(rbd_dev, notify_op, &reply_pages, &reply_len);
2746 ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len));
2749 static void rbd_notify_acquired_lock(struct work_struct *work)
2751 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
2752 acquired_lock_work);
2754 rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_ACQUIRED_LOCK);
2757 static void rbd_notify_released_lock(struct work_struct *work)
2759 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
2760 released_lock_work);
2762 rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_RELEASED_LOCK);
2765 static int rbd_request_lock(struct rbd_device *rbd_dev)
2767 struct page **reply_pages;
2769 bool lock_owner_responded = false;
2772 dout("%s rbd_dev %p\n", __func__, rbd_dev);
2774 ret = __rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_REQUEST_LOCK,
2775 &reply_pages, &reply_len);
2776 if (ret && ret != -ETIMEDOUT) {
2777 rbd_warn(rbd_dev, "failed to request lock: %d", ret);
2781 if (reply_len > 0 && reply_len <= PAGE_SIZE) {
2782 void *p = page_address(reply_pages[0]);
2783 void *const end = p + reply_len;
2786 ceph_decode_32_safe(&p, end, n, e_inval); /* num_acks */
2791 ceph_decode_need(&p, end, 8 + 8, e_inval);
2792 p += 8 + 8; /* skip gid and cookie */
2794 ceph_decode_32_safe(&p, end, len, e_inval);
2798 if (lock_owner_responded) {
2800 "duplicate lock owners detected");
2805 lock_owner_responded = true;
2806 ret = ceph_start_decoding(&p, end, 1, "ResponseMessage",
2810 "failed to decode ResponseMessage: %d",
2815 ret = ceph_decode_32(&p);
2819 if (!lock_owner_responded) {
2820 rbd_warn(rbd_dev, "no lock owners detected");
2825 ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len));
2833 static void wake_requests(struct rbd_device *rbd_dev, bool wake_all)
2835 dout("%s rbd_dev %p wake_all %d\n", __func__, rbd_dev, wake_all);
2837 cancel_delayed_work(&rbd_dev->lock_dwork);
2839 wake_up_all(&rbd_dev->lock_waitq);
2841 wake_up(&rbd_dev->lock_waitq);
2844 static int get_lock_owner_info(struct rbd_device *rbd_dev,
2845 struct ceph_locker **lockers, u32 *num_lockers)
2847 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2852 dout("%s rbd_dev %p\n", __func__, rbd_dev);
2854 ret = ceph_cls_lock_info(osdc, &rbd_dev->header_oid,
2855 &rbd_dev->header_oloc, RBD_LOCK_NAME,
2856 &lock_type, &lock_tag, lockers, num_lockers);
2860 if (*num_lockers == 0) {
2861 dout("%s rbd_dev %p no lockers detected\n", __func__, rbd_dev);
2865 if (strcmp(lock_tag, RBD_LOCK_TAG)) {
2866 rbd_warn(rbd_dev, "locked by external mechanism, tag %s",
2872 if (lock_type == CEPH_CLS_LOCK_SHARED) {
2873 rbd_warn(rbd_dev, "shared lock type detected");
2878 if (strncmp((*lockers)[0].id.cookie, RBD_LOCK_COOKIE_PREFIX,
2879 strlen(RBD_LOCK_COOKIE_PREFIX))) {
2880 rbd_warn(rbd_dev, "locked by external mechanism, cookie %s",
2881 (*lockers)[0].id.cookie);
2891 static int find_watcher(struct rbd_device *rbd_dev,
2892 const struct ceph_locker *locker)
2894 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2895 struct ceph_watch_item *watchers;
2901 ret = ceph_osdc_list_watchers(osdc, &rbd_dev->header_oid,
2902 &rbd_dev->header_oloc, &watchers,
2907 sscanf(locker->id.cookie, RBD_LOCK_COOKIE_PREFIX " %llu", &cookie);
2908 for (i = 0; i < num_watchers; i++) {
2909 if (!memcmp(&watchers[i].addr, &locker->info.addr,
2910 sizeof(locker->info.addr)) &&
2911 watchers[i].cookie == cookie) {
2912 struct rbd_client_id cid = {
2913 .gid = le64_to_cpu(watchers[i].name.num),
2917 dout("%s rbd_dev %p found cid %llu-%llu\n", __func__,
2918 rbd_dev, cid.gid, cid.handle);
2919 rbd_set_owner_cid(rbd_dev, &cid);
2925 dout("%s rbd_dev %p no watchers\n", __func__, rbd_dev);
2933 * lock_rwsem must be held for write
2935 static int rbd_try_lock(struct rbd_device *rbd_dev)
2937 struct ceph_client *client = rbd_dev->rbd_client->client;
2938 struct ceph_locker *lockers;
2943 ret = rbd_lock(rbd_dev);
2947 /* determine if the current lock holder is still alive */
2948 ret = get_lock_owner_info(rbd_dev, &lockers, &num_lockers);
2952 if (num_lockers == 0)
2955 ret = find_watcher(rbd_dev, lockers);
2958 ret = 0; /* have to request lock */
2962 rbd_warn(rbd_dev, "%s%llu seems dead, breaking lock",
2963 ENTITY_NAME(lockers[0].id.name));
2965 ret = ceph_monc_blacklist_add(&client->monc,
2966 &lockers[0].info.addr);
2968 rbd_warn(rbd_dev, "blacklist of %s%llu failed: %d",
2969 ENTITY_NAME(lockers[0].id.name), ret);
2973 ret = ceph_cls_break_lock(&client->osdc, &rbd_dev->header_oid,
2974 &rbd_dev->header_oloc, RBD_LOCK_NAME,
2975 lockers[0].id.cookie,
2976 &lockers[0].id.name);
2977 if (ret && ret != -ENOENT)
2981 ceph_free_lockers(lockers, num_lockers);
2985 ceph_free_lockers(lockers, num_lockers);
2990 * ret is set only if lock_state is RBD_LOCK_STATE_UNLOCKED
2992 static enum rbd_lock_state rbd_try_acquire_lock(struct rbd_device *rbd_dev,
2995 enum rbd_lock_state lock_state;
2997 down_read(&rbd_dev->lock_rwsem);
2998 dout("%s rbd_dev %p read lock_state %d\n", __func__, rbd_dev,
2999 rbd_dev->lock_state);
3000 if (__rbd_is_lock_owner(rbd_dev)) {
3001 lock_state = rbd_dev->lock_state;
3002 up_read(&rbd_dev->lock_rwsem);
3006 up_read(&rbd_dev->lock_rwsem);
3007 down_write(&rbd_dev->lock_rwsem);
3008 dout("%s rbd_dev %p write lock_state %d\n", __func__, rbd_dev,
3009 rbd_dev->lock_state);
3010 if (!__rbd_is_lock_owner(rbd_dev)) {
3011 *pret = rbd_try_lock(rbd_dev);
3013 rbd_warn(rbd_dev, "failed to acquire lock: %d", *pret);
3016 lock_state = rbd_dev->lock_state;
3017 up_write(&rbd_dev->lock_rwsem);
3021 static void rbd_acquire_lock(struct work_struct *work)
3023 struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
3024 struct rbd_device, lock_dwork);
3025 enum rbd_lock_state lock_state;
3028 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3030 lock_state = rbd_try_acquire_lock(rbd_dev, &ret);
3031 if (lock_state != RBD_LOCK_STATE_UNLOCKED || ret == -EBLACKLISTED) {
3032 if (lock_state == RBD_LOCK_STATE_LOCKED)
3033 wake_requests(rbd_dev, true);
3034 dout("%s rbd_dev %p lock_state %d ret %d - done\n", __func__,
3035 rbd_dev, lock_state, ret);
3039 ret = rbd_request_lock(rbd_dev);
3040 if (ret == -ETIMEDOUT) {
3041 goto again; /* treat this as a dead client */
3042 } else if (ret == -EROFS) {
3043 rbd_warn(rbd_dev, "peer will not release lock");
3045 * If this is rbd_add_acquire_lock(), we want to fail
3046 * immediately -- reuse BLACKLISTED flag. Otherwise we
3049 if (!(rbd_dev->disk->flags & GENHD_FL_UP)) {
3050 set_bit(RBD_DEV_FLAG_BLACKLISTED, &rbd_dev->flags);
3051 /* wake "rbd map --exclusive" process */
3052 wake_requests(rbd_dev, false);
3054 } else if (ret < 0) {
3055 rbd_warn(rbd_dev, "error requesting lock: %d", ret);
3056 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
3060 * lock owner acked, but resend if we don't see them
3063 dout("%s rbd_dev %p requeueing lock_dwork\n", __func__,
3065 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
3066 msecs_to_jiffies(2 * RBD_NOTIFY_TIMEOUT * MSEC_PER_SEC));
3071 * lock_rwsem must be held for write
3073 static bool rbd_release_lock(struct rbd_device *rbd_dev)
3075 dout("%s rbd_dev %p read lock_state %d\n", __func__, rbd_dev,
3076 rbd_dev->lock_state);
3077 if (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED)
3080 rbd_dev->lock_state = RBD_LOCK_STATE_RELEASING;
3081 downgrade_write(&rbd_dev->lock_rwsem);
3083 * Ensure that all in-flight IO is flushed.
3085 * FIXME: ceph_osdc_sync() flushes the entire OSD client, which
3086 * may be shared with other devices.
3088 ceph_osdc_sync(&rbd_dev->rbd_client->client->osdc);
3089 up_read(&rbd_dev->lock_rwsem);
3091 down_write(&rbd_dev->lock_rwsem);
3092 dout("%s rbd_dev %p write lock_state %d\n", __func__, rbd_dev,
3093 rbd_dev->lock_state);
3094 if (rbd_dev->lock_state != RBD_LOCK_STATE_RELEASING)
3097 rbd_unlock(rbd_dev);
3099 * Give others a chance to grab the lock - we would re-acquire
3100 * almost immediately if we got new IO during ceph_osdc_sync()
3101 * otherwise. We need to ack our own notifications, so this
3102 * lock_dwork will be requeued from rbd_wait_state_locked()
3103 * after wake_requests() in rbd_handle_released_lock().
3105 cancel_delayed_work(&rbd_dev->lock_dwork);
3109 static void rbd_release_lock_work(struct work_struct *work)
3111 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3114 down_write(&rbd_dev->lock_rwsem);
3115 rbd_release_lock(rbd_dev);
3116 up_write(&rbd_dev->lock_rwsem);
3119 static void rbd_handle_acquired_lock(struct rbd_device *rbd_dev, u8 struct_v,
3122 struct rbd_client_id cid = { 0 };
3124 if (struct_v >= 2) {
3125 cid.gid = ceph_decode_64(p);
3126 cid.handle = ceph_decode_64(p);
3129 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3131 if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
3132 down_write(&rbd_dev->lock_rwsem);
3133 if (rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
3135 * we already know that the remote client is
3138 up_write(&rbd_dev->lock_rwsem);
3142 rbd_set_owner_cid(rbd_dev, &cid);
3143 downgrade_write(&rbd_dev->lock_rwsem);
3145 down_read(&rbd_dev->lock_rwsem);
3148 if (!__rbd_is_lock_owner(rbd_dev))
3149 wake_requests(rbd_dev, false);
3150 up_read(&rbd_dev->lock_rwsem);
3153 static void rbd_handle_released_lock(struct rbd_device *rbd_dev, u8 struct_v,
3156 struct rbd_client_id cid = { 0 };
3158 if (struct_v >= 2) {
3159 cid.gid = ceph_decode_64(p);
3160 cid.handle = ceph_decode_64(p);
3163 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3165 if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
3166 down_write(&rbd_dev->lock_rwsem);
3167 if (!rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
3168 dout("%s rbd_dev %p unexpected owner, cid %llu-%llu != owner_cid %llu-%llu\n",
3169 __func__, rbd_dev, cid.gid, cid.handle,
3170 rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle);
3171 up_write(&rbd_dev->lock_rwsem);
3175 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3176 downgrade_write(&rbd_dev->lock_rwsem);
3178 down_read(&rbd_dev->lock_rwsem);
3181 if (!__rbd_is_lock_owner(rbd_dev))
3182 wake_requests(rbd_dev, false);
3183 up_read(&rbd_dev->lock_rwsem);
3187 * Returns result for ResponseMessage to be encoded (<= 0), or 1 if no
3188 * ResponseMessage is needed.
3190 static int rbd_handle_request_lock(struct rbd_device *rbd_dev, u8 struct_v,
3193 struct rbd_client_id my_cid = rbd_get_cid(rbd_dev);
3194 struct rbd_client_id cid = { 0 };
3197 if (struct_v >= 2) {
3198 cid.gid = ceph_decode_64(p);
3199 cid.handle = ceph_decode_64(p);
3202 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3204 if (rbd_cid_equal(&cid, &my_cid))
3207 down_read(&rbd_dev->lock_rwsem);
3208 if (__rbd_is_lock_owner(rbd_dev)) {
3209 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED &&
3210 rbd_cid_equal(&rbd_dev->owner_cid, &rbd_empty_cid))
3214 * encode ResponseMessage(0) so the peer can detect
3219 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED) {
3220 if (!rbd_dev->opts->exclusive) {
3221 dout("%s rbd_dev %p queueing unlock_work\n",
3223 queue_work(rbd_dev->task_wq,
3224 &rbd_dev->unlock_work);
3226 /* refuse to release the lock */
3233 up_read(&rbd_dev->lock_rwsem);
3237 static void __rbd_acknowledge_notify(struct rbd_device *rbd_dev,
3238 u64 notify_id, u64 cookie, s32 *result)
3240 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3241 char buf[4 + CEPH_ENCODING_START_BLK_LEN];
3242 int buf_size = sizeof(buf);
3248 /* encode ResponseMessage */
3249 ceph_start_encoding(&p, 1, 1,
3250 buf_size - CEPH_ENCODING_START_BLK_LEN);
3251 ceph_encode_32(&p, *result);
3256 ret = ceph_osdc_notify_ack(osdc, &rbd_dev->header_oid,
3257 &rbd_dev->header_oloc, notify_id, cookie,
3260 rbd_warn(rbd_dev, "acknowledge_notify failed: %d", ret);
3263 static void rbd_acknowledge_notify(struct rbd_device *rbd_dev, u64 notify_id,
3266 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3267 __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, NULL);
3270 static void rbd_acknowledge_notify_result(struct rbd_device *rbd_dev,
3271 u64 notify_id, u64 cookie, s32 result)
3273 dout("%s rbd_dev %p result %d\n", __func__, rbd_dev, result);
3274 __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, &result);
3277 static void rbd_watch_cb(void *arg, u64 notify_id, u64 cookie,
3278 u64 notifier_id, void *data, size_t data_len)
3280 struct rbd_device *rbd_dev = arg;
3282 void *const end = p + data_len;
3288 dout("%s rbd_dev %p cookie %llu notify_id %llu data_len %zu\n",
3289 __func__, rbd_dev, cookie, notify_id, data_len);
3291 ret = ceph_start_decoding(&p, end, 1, "NotifyMessage",
3294 rbd_warn(rbd_dev, "failed to decode NotifyMessage: %d",
3299 notify_op = ceph_decode_32(&p);
3301 /* legacy notification for header updates */
3302 notify_op = RBD_NOTIFY_OP_HEADER_UPDATE;
3306 dout("%s rbd_dev %p notify_op %u\n", __func__, rbd_dev, notify_op);
3307 switch (notify_op) {
3308 case RBD_NOTIFY_OP_ACQUIRED_LOCK:
3309 rbd_handle_acquired_lock(rbd_dev, struct_v, &p);
3310 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3312 case RBD_NOTIFY_OP_RELEASED_LOCK:
3313 rbd_handle_released_lock(rbd_dev, struct_v, &p);
3314 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3316 case RBD_NOTIFY_OP_REQUEST_LOCK:
3317 ret = rbd_handle_request_lock(rbd_dev, struct_v, &p);
3319 rbd_acknowledge_notify_result(rbd_dev, notify_id,
3322 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3324 case RBD_NOTIFY_OP_HEADER_UPDATE:
3325 ret = rbd_dev_refresh(rbd_dev);
3327 rbd_warn(rbd_dev, "refresh failed: %d", ret);
3329 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3332 if (rbd_is_lock_owner(rbd_dev))
3333 rbd_acknowledge_notify_result(rbd_dev, notify_id,
3334 cookie, -EOPNOTSUPP);
3336 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3341 static void __rbd_unregister_watch(struct rbd_device *rbd_dev);
3343 static void rbd_watch_errcb(void *arg, u64 cookie, int err)
3345 struct rbd_device *rbd_dev = arg;
3347 rbd_warn(rbd_dev, "encountered watch error: %d", err);
3349 down_write(&rbd_dev->lock_rwsem);
3350 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3351 up_write(&rbd_dev->lock_rwsem);
3353 mutex_lock(&rbd_dev->watch_mutex);
3354 if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED) {
3355 __rbd_unregister_watch(rbd_dev);
3356 rbd_dev->watch_state = RBD_WATCH_STATE_ERROR;
3358 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->watch_dwork, 0);
3360 mutex_unlock(&rbd_dev->watch_mutex);
3364 * watch_mutex must be locked
3366 static int __rbd_register_watch(struct rbd_device *rbd_dev)
3368 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3369 struct ceph_osd_linger_request *handle;
3371 rbd_assert(!rbd_dev->watch_handle);
3372 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3374 handle = ceph_osdc_watch(osdc, &rbd_dev->header_oid,
3375 &rbd_dev->header_oloc, rbd_watch_cb,
3376 rbd_watch_errcb, rbd_dev);
3378 return PTR_ERR(handle);
3380 rbd_dev->watch_handle = handle;
3385 * watch_mutex must be locked
3387 static void __rbd_unregister_watch(struct rbd_device *rbd_dev)
3389 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3392 rbd_assert(rbd_dev->watch_handle);
3393 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3395 ret = ceph_osdc_unwatch(osdc, rbd_dev->watch_handle);
3397 rbd_warn(rbd_dev, "failed to unwatch: %d", ret);
3399 rbd_dev->watch_handle = NULL;
3402 static int rbd_register_watch(struct rbd_device *rbd_dev)
3406 mutex_lock(&rbd_dev->watch_mutex);
3407 rbd_assert(rbd_dev->watch_state == RBD_WATCH_STATE_UNREGISTERED);
3408 ret = __rbd_register_watch(rbd_dev);
3412 rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
3413 rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
3416 mutex_unlock(&rbd_dev->watch_mutex);
3420 static void cancel_tasks_sync(struct rbd_device *rbd_dev)
3422 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3424 cancel_work_sync(&rbd_dev->acquired_lock_work);
3425 cancel_work_sync(&rbd_dev->released_lock_work);
3426 cancel_delayed_work_sync(&rbd_dev->lock_dwork);
3427 cancel_work_sync(&rbd_dev->unlock_work);
3431 * header_rwsem must not be held to avoid a deadlock with
3432 * rbd_dev_refresh() when flushing notifies.
3434 static void rbd_unregister_watch(struct rbd_device *rbd_dev)
3436 WARN_ON(waitqueue_active(&rbd_dev->lock_waitq));
3437 cancel_tasks_sync(rbd_dev);
3439 mutex_lock(&rbd_dev->watch_mutex);
3440 if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED)
3441 __rbd_unregister_watch(rbd_dev);
3442 rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
3443 mutex_unlock(&rbd_dev->watch_mutex);
3445 cancel_delayed_work_sync(&rbd_dev->watch_dwork);
3446 ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
3450 * lock_rwsem must be held for write
3452 static void rbd_reacquire_lock(struct rbd_device *rbd_dev)
3454 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3458 WARN_ON(rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED);
3460 format_lock_cookie(rbd_dev, cookie);
3461 ret = ceph_cls_set_cookie(osdc, &rbd_dev->header_oid,
3462 &rbd_dev->header_oloc, RBD_LOCK_NAME,
3463 CEPH_CLS_LOCK_EXCLUSIVE, rbd_dev->lock_cookie,
3464 RBD_LOCK_TAG, cookie);
3466 if (ret != -EOPNOTSUPP)
3467 rbd_warn(rbd_dev, "failed to update lock cookie: %d",
3471 * Lock cookie cannot be updated on older OSDs, so do
3472 * a manual release and queue an acquire.
3474 if (rbd_release_lock(rbd_dev))
3475 queue_delayed_work(rbd_dev->task_wq,
3476 &rbd_dev->lock_dwork, 0);
3478 __rbd_lock(rbd_dev, cookie);
3482 static void rbd_reregister_watch(struct work_struct *work)
3484 struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
3485 struct rbd_device, watch_dwork);
3488 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3490 mutex_lock(&rbd_dev->watch_mutex);
3491 if (rbd_dev->watch_state != RBD_WATCH_STATE_ERROR) {
3492 mutex_unlock(&rbd_dev->watch_mutex);
3496 ret = __rbd_register_watch(rbd_dev);
3498 rbd_warn(rbd_dev, "failed to reregister watch: %d", ret);
3499 if (ret == -EBLACKLISTED || ret == -ENOENT) {
3500 set_bit(RBD_DEV_FLAG_BLACKLISTED, &rbd_dev->flags);
3501 wake_requests(rbd_dev, true);
3503 queue_delayed_work(rbd_dev->task_wq,
3504 &rbd_dev->watch_dwork,
3507 mutex_unlock(&rbd_dev->watch_mutex);
3511 rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
3512 rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
3513 mutex_unlock(&rbd_dev->watch_mutex);
3515 down_write(&rbd_dev->lock_rwsem);
3516 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED)
3517 rbd_reacquire_lock(rbd_dev);
3518 up_write(&rbd_dev->lock_rwsem);
3520 ret = rbd_dev_refresh(rbd_dev);
3522 rbd_warn(rbd_dev, "reregistration refresh failed: %d", ret);
3526 * Synchronous osd object method call. Returns the number of bytes
3527 * returned in the outbound buffer, or a negative error code.
3529 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
3530 struct ceph_object_id *oid,
3531 struct ceph_object_locator *oloc,
3532 const char *method_name,
3533 const void *outbound,
3534 size_t outbound_size,
3536 size_t inbound_size)
3538 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3539 struct page *req_page = NULL;
3540 struct page *reply_page;
3544 * Method calls are ultimately read operations. The result
3545 * should placed into the inbound buffer provided. They
3546 * also supply outbound data--parameters for the object
3547 * method. Currently if this is present it will be a
3551 if (outbound_size > PAGE_SIZE)
3554 req_page = alloc_page(GFP_KERNEL);
3558 memcpy(page_address(req_page), outbound, outbound_size);
3561 reply_page = alloc_page(GFP_KERNEL);
3564 __free_page(req_page);
3568 ret = ceph_osdc_call(osdc, oid, oloc, RBD_DRV_NAME, method_name,
3569 CEPH_OSD_FLAG_READ, req_page, outbound_size,
3570 reply_page, &inbound_size);
3572 memcpy(inbound, page_address(reply_page), inbound_size);
3577 __free_page(req_page);
3578 __free_page(reply_page);
3583 * lock_rwsem must be held for read
3585 static int rbd_wait_state_locked(struct rbd_device *rbd_dev, bool may_acquire)
3588 unsigned long timeout;
3591 if (test_bit(RBD_DEV_FLAG_BLACKLISTED, &rbd_dev->flags))
3592 return -EBLACKLISTED;
3594 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED)
3598 rbd_warn(rbd_dev, "exclusive lock required");
3604 * Note the use of mod_delayed_work() in rbd_acquire_lock()
3605 * and cancel_delayed_work() in wake_requests().
3607 dout("%s rbd_dev %p queueing lock_dwork\n", __func__, rbd_dev);
3608 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork, 0);
3609 prepare_to_wait_exclusive(&rbd_dev->lock_waitq, &wait,
3610 TASK_UNINTERRUPTIBLE);
3611 up_read(&rbd_dev->lock_rwsem);
3612 timeout = schedule_timeout(ceph_timeout_jiffies(
3613 rbd_dev->opts->lock_timeout));
3614 down_read(&rbd_dev->lock_rwsem);
3615 if (test_bit(RBD_DEV_FLAG_BLACKLISTED, &rbd_dev->flags)) {
3616 ret = -EBLACKLISTED;
3620 rbd_warn(rbd_dev, "timed out waiting for lock");
3624 } while (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED);
3626 finish_wait(&rbd_dev->lock_waitq, &wait);
3630 static void rbd_queue_workfn(struct work_struct *work)
3632 struct request *rq = blk_mq_rq_from_pdu(work);
3633 struct rbd_device *rbd_dev = rq->q->queuedata;
3634 struct rbd_img_request *img_request;
3635 struct ceph_snap_context *snapc = NULL;
3636 u64 offset = (u64)blk_rq_pos(rq) << SECTOR_SHIFT;
3637 u64 length = blk_rq_bytes(rq);
3638 enum obj_operation_type op_type;
3640 bool must_be_locked;
3643 switch (req_op(rq)) {
3644 case REQ_OP_DISCARD:
3645 case REQ_OP_WRITE_ZEROES:
3646 op_type = OBJ_OP_DISCARD;
3649 op_type = OBJ_OP_WRITE;
3652 op_type = OBJ_OP_READ;
3655 dout("%s: non-fs request type %d\n", __func__, req_op(rq));
3660 /* Ignore/skip any zero-length requests */
3663 dout("%s: zero-length request\n", __func__);
3668 rbd_assert(op_type == OBJ_OP_READ ||
3669 rbd_dev->spec->snap_id == CEPH_NOSNAP);
3672 * Quit early if the mapped snapshot no longer exists. It's
3673 * still possible the snapshot will have disappeared by the
3674 * time our request arrives at the osd, but there's no sense in
3675 * sending it if we already know.
3677 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
3678 dout("request for non-existent snapshot");
3679 rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
3684 if (offset && length > U64_MAX - offset + 1) {
3685 rbd_warn(rbd_dev, "bad request range (%llu~%llu)", offset,
3688 goto err_rq; /* Shouldn't happen */
3691 blk_mq_start_request(rq);
3693 down_read(&rbd_dev->header_rwsem);
3694 mapping_size = rbd_dev->mapping.size;
3695 if (op_type != OBJ_OP_READ) {
3696 snapc = rbd_dev->header.snapc;
3697 ceph_get_snap_context(snapc);
3699 up_read(&rbd_dev->header_rwsem);
3701 if (offset + length > mapping_size) {
3702 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)", offset,
3703 length, mapping_size);
3709 (rbd_dev->header.features & RBD_FEATURE_EXCLUSIVE_LOCK) &&
3710 (op_type != OBJ_OP_READ || rbd_dev->opts->lock_on_read);
3711 if (must_be_locked) {
3712 down_read(&rbd_dev->lock_rwsem);
3713 result = rbd_wait_state_locked(rbd_dev,
3714 !rbd_dev->opts->exclusive);
3719 img_request = rbd_img_request_create(rbd_dev, op_type, snapc);
3724 img_request->rq = rq;
3725 snapc = NULL; /* img_request consumes a ref */
3727 if (op_type == OBJ_OP_DISCARD)
3728 result = rbd_img_fill_nodata(img_request, offset, length);
3730 result = rbd_img_fill_from_bio(img_request, offset, length,
3733 goto err_img_request;
3735 rbd_img_request_submit(img_request);
3737 up_read(&rbd_dev->lock_rwsem);
3741 rbd_img_request_put(img_request);
3744 up_read(&rbd_dev->lock_rwsem);
3747 rbd_warn(rbd_dev, "%s %llx at %llx result %d",
3748 obj_op_name(op_type), length, offset, result);
3749 ceph_put_snap_context(snapc);
3751 blk_mq_end_request(rq, errno_to_blk_status(result));
3754 static blk_status_t rbd_queue_rq(struct blk_mq_hw_ctx *hctx,
3755 const struct blk_mq_queue_data *bd)
3757 struct request *rq = bd->rq;
3758 struct work_struct *work = blk_mq_rq_to_pdu(rq);
3760 queue_work(rbd_wq, work);
3764 static void rbd_free_disk(struct rbd_device *rbd_dev)
3766 blk_cleanup_queue(rbd_dev->disk->queue);
3767 blk_mq_free_tag_set(&rbd_dev->tag_set);
3768 put_disk(rbd_dev->disk);
3769 rbd_dev->disk = NULL;
3772 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
3773 struct ceph_object_id *oid,
3774 struct ceph_object_locator *oloc,
3775 void *buf, int buf_len)
3778 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3779 struct ceph_osd_request *req;
3780 struct page **pages;
3781 int num_pages = calc_pages_for(0, buf_len);
3784 req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
3788 ceph_oid_copy(&req->r_base_oid, oid);
3789 ceph_oloc_copy(&req->r_base_oloc, oloc);
3790 req->r_flags = CEPH_OSD_FLAG_READ;
3792 ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
3796 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
3797 if (IS_ERR(pages)) {
3798 ret = PTR_ERR(pages);
3802 osd_req_op_extent_init(req, 0, CEPH_OSD_OP_READ, 0, buf_len, 0, 0);
3803 osd_req_op_extent_osd_data_pages(req, 0, pages, buf_len, 0, false,
3806 ceph_osdc_start_request(osdc, req, false);
3807 ret = ceph_osdc_wait_request(osdc, req);
3809 ceph_copy_from_page_vector(pages, buf, 0, ret);
3812 ceph_osdc_put_request(req);
3817 * Read the complete header for the given rbd device. On successful
3818 * return, the rbd_dev->header field will contain up-to-date
3819 * information about the image.
3821 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev)
3823 struct rbd_image_header_ondisk *ondisk = NULL;
3830 * The complete header will include an array of its 64-bit
3831 * snapshot ids, followed by the names of those snapshots as
3832 * a contiguous block of NUL-terminated strings. Note that
3833 * the number of snapshots could change by the time we read
3834 * it in, in which case we re-read it.
3841 size = sizeof (*ondisk);
3842 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
3844 ondisk = kmalloc(size, GFP_KERNEL);
3848 ret = rbd_obj_read_sync(rbd_dev, &rbd_dev->header_oid,
3849 &rbd_dev->header_oloc, ondisk, size);
3852 if ((size_t)ret < size) {
3854 rbd_warn(rbd_dev, "short header read (want %zd got %d)",
3858 if (!rbd_dev_ondisk_valid(ondisk)) {
3860 rbd_warn(rbd_dev, "invalid header");
3864 names_size = le64_to_cpu(ondisk->snap_names_len);
3865 want_count = snap_count;
3866 snap_count = le32_to_cpu(ondisk->snap_count);
3867 } while (snap_count != want_count);
3869 ret = rbd_header_from_disk(rbd_dev, ondisk);
3877 * Clear the rbd device's EXISTS flag if the snapshot it's mapped to
3878 * has disappeared from the (just updated) snapshot context.
3880 static void rbd_exists_validate(struct rbd_device *rbd_dev)
3884 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags))
3887 snap_id = rbd_dev->spec->snap_id;
3888 if (snap_id == CEPH_NOSNAP)
3891 if (rbd_dev_snap_index(rbd_dev, snap_id) == BAD_SNAP_INDEX)
3892 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
3895 static void rbd_dev_update_size(struct rbd_device *rbd_dev)
3900 * If EXISTS is not set, rbd_dev->disk may be NULL, so don't
3901 * try to update its size. If REMOVING is set, updating size
3902 * is just useless work since the device can't be opened.
3904 if (test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags) &&
3905 !test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags)) {
3906 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
3907 dout("setting size to %llu sectors", (unsigned long long)size);
3908 set_capacity(rbd_dev->disk, size);
3909 revalidate_disk(rbd_dev->disk);
3913 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
3918 down_write(&rbd_dev->header_rwsem);
3919 mapping_size = rbd_dev->mapping.size;
3921 ret = rbd_dev_header_info(rbd_dev);
3926 * If there is a parent, see if it has disappeared due to the
3927 * mapped image getting flattened.
3929 if (rbd_dev->parent) {
3930 ret = rbd_dev_v2_parent_info(rbd_dev);
3935 if (rbd_dev->spec->snap_id == CEPH_NOSNAP) {
3936 rbd_dev->mapping.size = rbd_dev->header.image_size;
3938 /* validate mapped snapshot's EXISTS flag */
3939 rbd_exists_validate(rbd_dev);
3943 up_write(&rbd_dev->header_rwsem);
3944 if (!ret && mapping_size != rbd_dev->mapping.size)
3945 rbd_dev_update_size(rbd_dev);
3950 static int rbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
3951 unsigned int hctx_idx, unsigned int numa_node)
3953 struct work_struct *work = blk_mq_rq_to_pdu(rq);
3955 INIT_WORK(work, rbd_queue_workfn);
3959 static const struct blk_mq_ops rbd_mq_ops = {
3960 .queue_rq = rbd_queue_rq,
3961 .init_request = rbd_init_request,
3964 static int rbd_init_disk(struct rbd_device *rbd_dev)
3966 struct gendisk *disk;
3967 struct request_queue *q;
3968 unsigned int objset_bytes =
3969 rbd_dev->layout.object_size * rbd_dev->layout.stripe_count;
3972 /* create gendisk info */
3973 disk = alloc_disk(single_major ?
3974 (1 << RBD_SINGLE_MAJOR_PART_SHIFT) :
3975 RBD_MINORS_PER_MAJOR);
3979 snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
3981 disk->major = rbd_dev->major;
3982 disk->first_minor = rbd_dev->minor;
3984 disk->flags |= GENHD_FL_EXT_DEVT;
3985 disk->fops = &rbd_bd_ops;
3986 disk->private_data = rbd_dev;
3988 memset(&rbd_dev->tag_set, 0, sizeof(rbd_dev->tag_set));
3989 rbd_dev->tag_set.ops = &rbd_mq_ops;
3990 rbd_dev->tag_set.queue_depth = rbd_dev->opts->queue_depth;
3991 rbd_dev->tag_set.numa_node = NUMA_NO_NODE;
3992 rbd_dev->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
3993 rbd_dev->tag_set.nr_hw_queues = 1;
3994 rbd_dev->tag_set.cmd_size = sizeof(struct work_struct);
3996 err = blk_mq_alloc_tag_set(&rbd_dev->tag_set);
4000 q = blk_mq_init_queue(&rbd_dev->tag_set);
4006 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
4007 /* QUEUE_FLAG_ADD_RANDOM is off by default for blk-mq */
4009 blk_queue_max_hw_sectors(q, objset_bytes >> SECTOR_SHIFT);
4010 q->limits.max_sectors = queue_max_hw_sectors(q);
4011 blk_queue_max_segments(q, USHRT_MAX);
4012 blk_queue_max_segment_size(q, UINT_MAX);
4013 blk_queue_io_min(q, objset_bytes);
4014 blk_queue_io_opt(q, objset_bytes);
4016 if (rbd_dev->opts->trim) {
4017 blk_queue_flag_set(QUEUE_FLAG_DISCARD, q);
4018 q->limits.discard_granularity = objset_bytes;
4019 blk_queue_max_discard_sectors(q, objset_bytes >> SECTOR_SHIFT);
4020 blk_queue_max_write_zeroes_sectors(q, objset_bytes >> SECTOR_SHIFT);
4023 if (!ceph_test_opt(rbd_dev->rbd_client->client, NOCRC))
4024 q->backing_dev_info->capabilities |= BDI_CAP_STABLE_WRITES;
4027 * disk_release() expects a queue ref from add_disk() and will
4028 * put it. Hold an extra ref until add_disk() is called.
4030 WARN_ON(!blk_get_queue(q));
4032 q->queuedata = rbd_dev;
4034 rbd_dev->disk = disk;
4038 blk_mq_free_tag_set(&rbd_dev->tag_set);
4048 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
4050 return container_of(dev, struct rbd_device, dev);
4053 static ssize_t rbd_size_show(struct device *dev,
4054 struct device_attribute *attr, char *buf)
4056 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4058 return sprintf(buf, "%llu\n",
4059 (unsigned long long)rbd_dev->mapping.size);
4063 * Note this shows the features for whatever's mapped, which is not
4064 * necessarily the base image.
4066 static ssize_t rbd_features_show(struct device *dev,
4067 struct device_attribute *attr, char *buf)
4069 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4071 return sprintf(buf, "0x%016llx\n",
4072 (unsigned long long)rbd_dev->mapping.features);
4075 static ssize_t rbd_major_show(struct device *dev,
4076 struct device_attribute *attr, char *buf)
4078 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4081 return sprintf(buf, "%d\n", rbd_dev->major);
4083 return sprintf(buf, "(none)\n");
4086 static ssize_t rbd_minor_show(struct device *dev,
4087 struct device_attribute *attr, char *buf)
4089 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4091 return sprintf(buf, "%d\n", rbd_dev->minor);
4094 static ssize_t rbd_client_addr_show(struct device *dev,
4095 struct device_attribute *attr, char *buf)
4097 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4098 struct ceph_entity_addr *client_addr =
4099 ceph_client_addr(rbd_dev->rbd_client->client);
4101 return sprintf(buf, "%pISpc/%u\n", &client_addr->in_addr,
4102 le32_to_cpu(client_addr->nonce));
4105 static ssize_t rbd_client_id_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, "client%lld\n",
4111 ceph_client_gid(rbd_dev->rbd_client->client));
4114 static ssize_t rbd_cluster_fsid_show(struct device *dev,
4115 struct device_attribute *attr, char *buf)
4117 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4119 return sprintf(buf, "%pU\n", &rbd_dev->rbd_client->client->fsid);
4122 static ssize_t rbd_config_info_show(struct device *dev,
4123 struct device_attribute *attr, char *buf)
4125 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4127 if (!capable(CAP_SYS_ADMIN))
4130 return sprintf(buf, "%s\n", rbd_dev->config_info);
4133 static ssize_t rbd_pool_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->pool_name);
4141 static ssize_t rbd_pool_id_show(struct device *dev,
4142 struct device_attribute *attr, char *buf)
4144 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4146 return sprintf(buf, "%llu\n",
4147 (unsigned long long) rbd_dev->spec->pool_id);
4150 static ssize_t rbd_pool_ns_show(struct device *dev,
4151 struct device_attribute *attr, char *buf)
4153 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4155 return sprintf(buf, "%s\n", rbd_dev->spec->pool_ns ?: "");
4158 static ssize_t rbd_name_show(struct device *dev,
4159 struct device_attribute *attr, char *buf)
4161 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4163 if (rbd_dev->spec->image_name)
4164 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
4166 return sprintf(buf, "(unknown)\n");
4169 static ssize_t rbd_image_id_show(struct device *dev,
4170 struct device_attribute *attr, char *buf)
4172 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4174 return sprintf(buf, "%s\n", rbd_dev->spec->image_id);
4178 * Shows the name of the currently-mapped snapshot (or
4179 * RBD_SNAP_HEAD_NAME for the base image).
4181 static ssize_t rbd_snap_show(struct device *dev,
4182 struct device_attribute *attr,
4185 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4187 return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
4190 static ssize_t rbd_snap_id_show(struct device *dev,
4191 struct device_attribute *attr, char *buf)
4193 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4195 return sprintf(buf, "%llu\n", rbd_dev->spec->snap_id);
4199 * For a v2 image, shows the chain of parent images, separated by empty
4200 * lines. For v1 images or if there is no parent, shows "(no parent
4203 static ssize_t rbd_parent_show(struct device *dev,
4204 struct device_attribute *attr,
4207 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4210 if (!rbd_dev->parent)
4211 return sprintf(buf, "(no parent image)\n");
4213 for ( ; rbd_dev->parent; rbd_dev = rbd_dev->parent) {
4214 struct rbd_spec *spec = rbd_dev->parent_spec;
4216 count += sprintf(&buf[count], "%s"
4217 "pool_id %llu\npool_name %s\n"
4219 "image_id %s\nimage_name %s\n"
4220 "snap_id %llu\nsnap_name %s\n"
4222 !count ? "" : "\n", /* first? */
4223 spec->pool_id, spec->pool_name,
4224 spec->pool_ns ?: "",
4225 spec->image_id, spec->image_name ?: "(unknown)",
4226 spec->snap_id, spec->snap_name,
4227 rbd_dev->parent_overlap);
4233 static ssize_t rbd_image_refresh(struct device *dev,
4234 struct device_attribute *attr,
4238 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4241 if (!capable(CAP_SYS_ADMIN))
4244 ret = rbd_dev_refresh(rbd_dev);
4251 static DEVICE_ATTR(size, 0444, rbd_size_show, NULL);
4252 static DEVICE_ATTR(features, 0444, rbd_features_show, NULL);
4253 static DEVICE_ATTR(major, 0444, rbd_major_show, NULL);
4254 static DEVICE_ATTR(minor, 0444, rbd_minor_show, NULL);
4255 static DEVICE_ATTR(client_addr, 0444, rbd_client_addr_show, NULL);
4256 static DEVICE_ATTR(client_id, 0444, rbd_client_id_show, NULL);
4257 static DEVICE_ATTR(cluster_fsid, 0444, rbd_cluster_fsid_show, NULL);
4258 static DEVICE_ATTR(config_info, 0400, rbd_config_info_show, NULL);
4259 static DEVICE_ATTR(pool, 0444, rbd_pool_show, NULL);
4260 static DEVICE_ATTR(pool_id, 0444, rbd_pool_id_show, NULL);
4261 static DEVICE_ATTR(pool_ns, 0444, rbd_pool_ns_show, NULL);
4262 static DEVICE_ATTR(name, 0444, rbd_name_show, NULL);
4263 static DEVICE_ATTR(image_id, 0444, rbd_image_id_show, NULL);
4264 static DEVICE_ATTR(refresh, 0200, NULL, rbd_image_refresh);
4265 static DEVICE_ATTR(current_snap, 0444, rbd_snap_show, NULL);
4266 static DEVICE_ATTR(snap_id, 0444, rbd_snap_id_show, NULL);
4267 static DEVICE_ATTR(parent, 0444, rbd_parent_show, NULL);
4269 static struct attribute *rbd_attrs[] = {
4270 &dev_attr_size.attr,
4271 &dev_attr_features.attr,
4272 &dev_attr_major.attr,
4273 &dev_attr_minor.attr,
4274 &dev_attr_client_addr.attr,
4275 &dev_attr_client_id.attr,
4276 &dev_attr_cluster_fsid.attr,
4277 &dev_attr_config_info.attr,
4278 &dev_attr_pool.attr,
4279 &dev_attr_pool_id.attr,
4280 &dev_attr_pool_ns.attr,
4281 &dev_attr_name.attr,
4282 &dev_attr_image_id.attr,
4283 &dev_attr_current_snap.attr,
4284 &dev_attr_snap_id.attr,
4285 &dev_attr_parent.attr,
4286 &dev_attr_refresh.attr,
4290 static struct attribute_group rbd_attr_group = {
4294 static const struct attribute_group *rbd_attr_groups[] = {
4299 static void rbd_dev_release(struct device *dev);
4301 static const struct device_type rbd_device_type = {
4303 .groups = rbd_attr_groups,
4304 .release = rbd_dev_release,
4307 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
4309 kref_get(&spec->kref);
4314 static void rbd_spec_free(struct kref *kref);
4315 static void rbd_spec_put(struct rbd_spec *spec)
4318 kref_put(&spec->kref, rbd_spec_free);
4321 static struct rbd_spec *rbd_spec_alloc(void)
4323 struct rbd_spec *spec;
4325 spec = kzalloc(sizeof (*spec), GFP_KERNEL);
4329 spec->pool_id = CEPH_NOPOOL;
4330 spec->snap_id = CEPH_NOSNAP;
4331 kref_init(&spec->kref);
4336 static void rbd_spec_free(struct kref *kref)
4338 struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
4340 kfree(spec->pool_name);
4341 kfree(spec->pool_ns);
4342 kfree(spec->image_id);
4343 kfree(spec->image_name);
4344 kfree(spec->snap_name);
4348 static void rbd_dev_free(struct rbd_device *rbd_dev)
4350 WARN_ON(rbd_dev->watch_state != RBD_WATCH_STATE_UNREGISTERED);
4351 WARN_ON(rbd_dev->lock_state != RBD_LOCK_STATE_UNLOCKED);
4353 ceph_oid_destroy(&rbd_dev->header_oid);
4354 ceph_oloc_destroy(&rbd_dev->header_oloc);
4355 kfree(rbd_dev->config_info);
4357 rbd_put_client(rbd_dev->rbd_client);
4358 rbd_spec_put(rbd_dev->spec);
4359 kfree(rbd_dev->opts);
4363 static void rbd_dev_release(struct device *dev)
4365 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4366 bool need_put = !!rbd_dev->opts;
4369 destroy_workqueue(rbd_dev->task_wq);
4370 ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4373 rbd_dev_free(rbd_dev);
4376 * This is racy, but way better than putting module outside of
4377 * the release callback. The race window is pretty small, so
4378 * doing something similar to dm (dm-builtin.c) is overkill.
4381 module_put(THIS_MODULE);
4384 static struct rbd_device *__rbd_dev_create(struct rbd_client *rbdc,
4385 struct rbd_spec *spec)
4387 struct rbd_device *rbd_dev;
4389 rbd_dev = kzalloc(sizeof(*rbd_dev), GFP_KERNEL);
4393 spin_lock_init(&rbd_dev->lock);
4394 INIT_LIST_HEAD(&rbd_dev->node);
4395 init_rwsem(&rbd_dev->header_rwsem);
4397 rbd_dev->header.data_pool_id = CEPH_NOPOOL;
4398 ceph_oid_init(&rbd_dev->header_oid);
4399 rbd_dev->header_oloc.pool = spec->pool_id;
4400 if (spec->pool_ns) {
4401 WARN_ON(!*spec->pool_ns);
4402 rbd_dev->header_oloc.pool_ns =
4403 ceph_find_or_create_string(spec->pool_ns,
4404 strlen(spec->pool_ns));
4407 mutex_init(&rbd_dev->watch_mutex);
4408 rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
4409 INIT_DELAYED_WORK(&rbd_dev->watch_dwork, rbd_reregister_watch);
4411 init_rwsem(&rbd_dev->lock_rwsem);
4412 rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
4413 INIT_WORK(&rbd_dev->acquired_lock_work, rbd_notify_acquired_lock);
4414 INIT_WORK(&rbd_dev->released_lock_work, rbd_notify_released_lock);
4415 INIT_DELAYED_WORK(&rbd_dev->lock_dwork, rbd_acquire_lock);
4416 INIT_WORK(&rbd_dev->unlock_work, rbd_release_lock_work);
4417 init_waitqueue_head(&rbd_dev->lock_waitq);
4419 rbd_dev->dev.bus = &rbd_bus_type;
4420 rbd_dev->dev.type = &rbd_device_type;
4421 rbd_dev->dev.parent = &rbd_root_dev;
4422 device_initialize(&rbd_dev->dev);
4424 rbd_dev->rbd_client = rbdc;
4425 rbd_dev->spec = spec;
4431 * Create a mapping rbd_dev.
4433 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
4434 struct rbd_spec *spec,
4435 struct rbd_options *opts)
4437 struct rbd_device *rbd_dev;
4439 rbd_dev = __rbd_dev_create(rbdc, spec);
4443 rbd_dev->opts = opts;
4445 /* get an id and fill in device name */
4446 rbd_dev->dev_id = ida_simple_get(&rbd_dev_id_ida, 0,
4447 minor_to_rbd_dev_id(1 << MINORBITS),
4449 if (rbd_dev->dev_id < 0)
4452 sprintf(rbd_dev->name, RBD_DRV_NAME "%d", rbd_dev->dev_id);
4453 rbd_dev->task_wq = alloc_ordered_workqueue("%s-tasks", WQ_MEM_RECLAIM,
4455 if (!rbd_dev->task_wq)
4458 /* we have a ref from do_rbd_add() */
4459 __module_get(THIS_MODULE);
4461 dout("%s rbd_dev %p dev_id %d\n", __func__, rbd_dev, rbd_dev->dev_id);
4465 ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4467 rbd_dev_free(rbd_dev);
4471 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
4474 put_device(&rbd_dev->dev);
4478 * Get the size and object order for an image snapshot, or if
4479 * snap_id is CEPH_NOSNAP, gets this information for the base
4482 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
4483 u8 *order, u64 *snap_size)
4485 __le64 snapid = cpu_to_le64(snap_id);
4490 } __attribute__ ((packed)) size_buf = { 0 };
4492 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4493 &rbd_dev->header_oloc, "get_size",
4494 &snapid, sizeof(snapid),
4495 &size_buf, sizeof(size_buf));
4496 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4499 if (ret < sizeof (size_buf))
4503 *order = size_buf.order;
4504 dout(" order %u", (unsigned int)*order);
4506 *snap_size = le64_to_cpu(size_buf.size);
4508 dout(" snap_id 0x%016llx snap_size = %llu\n",
4509 (unsigned long long)snap_id,
4510 (unsigned long long)*snap_size);
4515 static int rbd_dev_v2_image_size(struct rbd_device *rbd_dev)
4517 return _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
4518 &rbd_dev->header.obj_order,
4519 &rbd_dev->header.image_size);
4522 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev)
4528 reply_buf = kzalloc(RBD_OBJ_PREFIX_LEN_MAX, GFP_KERNEL);
4532 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4533 &rbd_dev->header_oloc, "get_object_prefix",
4534 NULL, 0, reply_buf, RBD_OBJ_PREFIX_LEN_MAX);
4535 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4540 rbd_dev->header.object_prefix = ceph_extract_encoded_string(&p,
4541 p + ret, NULL, GFP_NOIO);
4544 if (IS_ERR(rbd_dev->header.object_prefix)) {
4545 ret = PTR_ERR(rbd_dev->header.object_prefix);
4546 rbd_dev->header.object_prefix = NULL;
4548 dout(" object_prefix = %s\n", rbd_dev->header.object_prefix);
4556 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
4559 __le64 snapid = cpu_to_le64(snap_id);
4563 } __attribute__ ((packed)) features_buf = { 0 };
4567 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4568 &rbd_dev->header_oloc, "get_features",
4569 &snapid, sizeof(snapid),
4570 &features_buf, sizeof(features_buf));
4571 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4574 if (ret < sizeof (features_buf))
4577 unsup = le64_to_cpu(features_buf.incompat) & ~RBD_FEATURES_SUPPORTED;
4579 rbd_warn(rbd_dev, "image uses unsupported features: 0x%llx",
4584 *snap_features = le64_to_cpu(features_buf.features);
4586 dout(" snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
4587 (unsigned long long)snap_id,
4588 (unsigned long long)*snap_features,
4589 (unsigned long long)le64_to_cpu(features_buf.incompat));
4594 static int rbd_dev_v2_features(struct rbd_device *rbd_dev)
4596 return _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
4597 &rbd_dev->header.features);
4600 struct parent_image_info {
4602 const char *pool_ns;
4603 const char *image_id;
4611 * The caller is responsible for @pii.
4613 static int decode_parent_image_spec(void **p, void *end,
4614 struct parent_image_info *pii)
4620 ret = ceph_start_decoding(p, end, 1, "ParentImageSpec",
4621 &struct_v, &struct_len);
4625 ceph_decode_64_safe(p, end, pii->pool_id, e_inval);
4626 pii->pool_ns = ceph_extract_encoded_string(p, end, NULL, GFP_KERNEL);
4627 if (IS_ERR(pii->pool_ns)) {
4628 ret = PTR_ERR(pii->pool_ns);
4629 pii->pool_ns = NULL;
4632 pii->image_id = ceph_extract_encoded_string(p, end, NULL, GFP_KERNEL);
4633 if (IS_ERR(pii->image_id)) {
4634 ret = PTR_ERR(pii->image_id);
4635 pii->image_id = NULL;
4638 ceph_decode_64_safe(p, end, pii->snap_id, e_inval);
4645 static int __get_parent_info(struct rbd_device *rbd_dev,
4646 struct page *req_page,
4647 struct page *reply_page,
4648 struct parent_image_info *pii)
4650 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4651 size_t reply_len = PAGE_SIZE;
4655 ret = ceph_osdc_call(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
4656 "rbd", "parent_get", CEPH_OSD_FLAG_READ,
4657 req_page, sizeof(u64), reply_page, &reply_len);
4659 return ret == -EOPNOTSUPP ? 1 : ret;
4661 p = page_address(reply_page);
4662 end = p + reply_len;
4663 ret = decode_parent_image_spec(&p, end, pii);
4667 ret = ceph_osdc_call(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
4668 "rbd", "parent_overlap_get", CEPH_OSD_FLAG_READ,
4669 req_page, sizeof(u64), reply_page, &reply_len);
4673 p = page_address(reply_page);
4674 end = p + reply_len;
4675 ceph_decode_8_safe(&p, end, pii->has_overlap, e_inval);
4676 if (pii->has_overlap)
4677 ceph_decode_64_safe(&p, end, pii->overlap, e_inval);
4686 * The caller is responsible for @pii.
4688 static int __get_parent_info_legacy(struct rbd_device *rbd_dev,
4689 struct page *req_page,
4690 struct page *reply_page,
4691 struct parent_image_info *pii)
4693 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4694 size_t reply_len = PAGE_SIZE;
4698 ret = ceph_osdc_call(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
4699 "rbd", "get_parent", CEPH_OSD_FLAG_READ,
4700 req_page, sizeof(u64), reply_page, &reply_len);
4704 p = page_address(reply_page);
4705 end = p + reply_len;
4706 ceph_decode_64_safe(&p, end, pii->pool_id, e_inval);
4707 pii->image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
4708 if (IS_ERR(pii->image_id)) {
4709 ret = PTR_ERR(pii->image_id);
4710 pii->image_id = NULL;
4713 ceph_decode_64_safe(&p, end, pii->snap_id, e_inval);
4714 pii->has_overlap = true;
4715 ceph_decode_64_safe(&p, end, pii->overlap, e_inval);
4723 static int get_parent_info(struct rbd_device *rbd_dev,
4724 struct parent_image_info *pii)
4726 struct page *req_page, *reply_page;
4730 req_page = alloc_page(GFP_KERNEL);
4734 reply_page = alloc_page(GFP_KERNEL);
4736 __free_page(req_page);
4740 p = page_address(req_page);
4741 ceph_encode_64(&p, rbd_dev->spec->snap_id);
4742 ret = __get_parent_info(rbd_dev, req_page, reply_page, pii);
4744 ret = __get_parent_info_legacy(rbd_dev, req_page, reply_page,
4747 __free_page(req_page);
4748 __free_page(reply_page);
4752 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev)
4754 struct rbd_spec *parent_spec;
4755 struct parent_image_info pii = { 0 };
4758 parent_spec = rbd_spec_alloc();
4762 ret = get_parent_info(rbd_dev, &pii);
4766 dout("%s pool_id %llu pool_ns %s image_id %s snap_id %llu has_overlap %d overlap %llu\n",
4767 __func__, pii.pool_id, pii.pool_ns, pii.image_id, pii.snap_id,
4768 pii.has_overlap, pii.overlap);
4770 if (pii.pool_id == CEPH_NOPOOL || !pii.has_overlap) {
4772 * Either the parent never existed, or we have
4773 * record of it but the image got flattened so it no
4774 * longer has a parent. When the parent of a
4775 * layered image disappears we immediately set the
4776 * overlap to 0. The effect of this is that all new
4777 * requests will be treated as if the image had no
4780 * If !pii.has_overlap, the parent image spec is not
4781 * applicable. It's there to avoid duplication in each
4784 if (rbd_dev->parent_overlap) {
4785 rbd_dev->parent_overlap = 0;
4786 rbd_dev_parent_put(rbd_dev);
4787 pr_info("%s: clone image has been flattened\n",
4788 rbd_dev->disk->disk_name);
4791 goto out; /* No parent? No problem. */
4794 /* The ceph file layout needs to fit pool id in 32 bits */
4797 if (pii.pool_id > (u64)U32_MAX) {
4798 rbd_warn(NULL, "parent pool id too large (%llu > %u)",
4799 (unsigned long long)pii.pool_id, U32_MAX);
4804 * The parent won't change (except when the clone is
4805 * flattened, already handled that). So we only need to
4806 * record the parent spec we have not already done so.
4808 if (!rbd_dev->parent_spec) {
4809 parent_spec->pool_id = pii.pool_id;
4810 if (pii.pool_ns && *pii.pool_ns) {
4811 parent_spec->pool_ns = pii.pool_ns;
4814 parent_spec->image_id = pii.image_id;
4815 pii.image_id = NULL;
4816 parent_spec->snap_id = pii.snap_id;
4818 rbd_dev->parent_spec = parent_spec;
4819 parent_spec = NULL; /* rbd_dev now owns this */
4823 * We always update the parent overlap. If it's zero we issue
4824 * a warning, as we will proceed as if there was no parent.
4828 /* refresh, careful to warn just once */
4829 if (rbd_dev->parent_overlap)
4831 "clone now standalone (overlap became 0)");
4834 rbd_warn(rbd_dev, "clone is standalone (overlap 0)");
4837 rbd_dev->parent_overlap = pii.overlap;
4843 kfree(pii.image_id);
4844 rbd_spec_put(parent_spec);
4848 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev)
4852 __le64 stripe_count;
4853 } __attribute__ ((packed)) striping_info_buf = { 0 };
4854 size_t size = sizeof (striping_info_buf);
4858 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4859 &rbd_dev->header_oloc, "get_stripe_unit_count",
4860 NULL, 0, &striping_info_buf, size);
4861 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4867 p = &striping_info_buf;
4868 rbd_dev->header.stripe_unit = ceph_decode_64(&p);
4869 rbd_dev->header.stripe_count = ceph_decode_64(&p);
4873 static int rbd_dev_v2_data_pool(struct rbd_device *rbd_dev)
4875 __le64 data_pool_id;
4878 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
4879 &rbd_dev->header_oloc, "get_data_pool",
4880 NULL, 0, &data_pool_id, sizeof(data_pool_id));
4883 if (ret < sizeof(data_pool_id))
4886 rbd_dev->header.data_pool_id = le64_to_cpu(data_pool_id);
4887 WARN_ON(rbd_dev->header.data_pool_id == CEPH_NOPOOL);
4891 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
4893 CEPH_DEFINE_OID_ONSTACK(oid);
4894 size_t image_id_size;
4899 void *reply_buf = NULL;
4901 char *image_name = NULL;
4904 rbd_assert(!rbd_dev->spec->image_name);
4906 len = strlen(rbd_dev->spec->image_id);
4907 image_id_size = sizeof (__le32) + len;
4908 image_id = kmalloc(image_id_size, GFP_KERNEL);
4913 end = image_id + image_id_size;
4914 ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
4916 size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
4917 reply_buf = kmalloc(size, GFP_KERNEL);
4921 ceph_oid_printf(&oid, "%s", RBD_DIRECTORY);
4922 ret = rbd_obj_method_sync(rbd_dev, &oid, &rbd_dev->header_oloc,
4923 "dir_get_name", image_id, image_id_size,
4928 end = reply_buf + ret;
4930 image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
4931 if (IS_ERR(image_name))
4934 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
4942 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4944 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4945 const char *snap_name;
4948 /* Skip over names until we find the one we are looking for */
4950 snap_name = rbd_dev->header.snap_names;
4951 while (which < snapc->num_snaps) {
4952 if (!strcmp(name, snap_name))
4953 return snapc->snaps[which];
4954 snap_name += strlen(snap_name) + 1;
4960 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4962 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4967 for (which = 0; !found && which < snapc->num_snaps; which++) {
4968 const char *snap_name;
4970 snap_id = snapc->snaps[which];
4971 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
4972 if (IS_ERR(snap_name)) {
4973 /* ignore no-longer existing snapshots */
4974 if (PTR_ERR(snap_name) == -ENOENT)
4979 found = !strcmp(name, snap_name);
4982 return found ? snap_id : CEPH_NOSNAP;
4986 * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
4987 * no snapshot by that name is found, or if an error occurs.
4989 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4991 if (rbd_dev->image_format == 1)
4992 return rbd_v1_snap_id_by_name(rbd_dev, name);
4994 return rbd_v2_snap_id_by_name(rbd_dev, name);
4998 * An image being mapped will have everything but the snap id.
5000 static int rbd_spec_fill_snap_id(struct rbd_device *rbd_dev)
5002 struct rbd_spec *spec = rbd_dev->spec;
5004 rbd_assert(spec->pool_id != CEPH_NOPOOL && spec->pool_name);
5005 rbd_assert(spec->image_id && spec->image_name);
5006 rbd_assert(spec->snap_name);
5008 if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
5011 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
5012 if (snap_id == CEPH_NOSNAP)
5015 spec->snap_id = snap_id;
5017 spec->snap_id = CEPH_NOSNAP;
5024 * A parent image will have all ids but none of the names.
5026 * All names in an rbd spec are dynamically allocated. It's OK if we
5027 * can't figure out the name for an image id.
5029 static int rbd_spec_fill_names(struct rbd_device *rbd_dev)
5031 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
5032 struct rbd_spec *spec = rbd_dev->spec;
5033 const char *pool_name;
5034 const char *image_name;
5035 const char *snap_name;
5038 rbd_assert(spec->pool_id != CEPH_NOPOOL);
5039 rbd_assert(spec->image_id);
5040 rbd_assert(spec->snap_id != CEPH_NOSNAP);
5042 /* Get the pool name; we have to make our own copy of this */
5044 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
5046 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
5049 pool_name = kstrdup(pool_name, GFP_KERNEL);
5053 /* Fetch the image name; tolerate failure here */
5055 image_name = rbd_dev_image_name(rbd_dev);
5057 rbd_warn(rbd_dev, "unable to get image name");
5059 /* Fetch the snapshot name */
5061 snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
5062 if (IS_ERR(snap_name)) {
5063 ret = PTR_ERR(snap_name);
5067 spec->pool_name = pool_name;
5068 spec->image_name = image_name;
5069 spec->snap_name = snap_name;
5079 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev)
5088 struct ceph_snap_context *snapc;
5092 * We'll need room for the seq value (maximum snapshot id),
5093 * snapshot count, and array of that many snapshot ids.
5094 * For now we have a fixed upper limit on the number we're
5095 * prepared to receive.
5097 size = sizeof (__le64) + sizeof (__le32) +
5098 RBD_MAX_SNAP_COUNT * sizeof (__le64);
5099 reply_buf = kzalloc(size, GFP_KERNEL);
5103 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
5104 &rbd_dev->header_oloc, "get_snapcontext",
5105 NULL, 0, reply_buf, size);
5106 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5111 end = reply_buf + ret;
5113 ceph_decode_64_safe(&p, end, seq, out);
5114 ceph_decode_32_safe(&p, end, snap_count, out);
5117 * Make sure the reported number of snapshot ids wouldn't go
5118 * beyond the end of our buffer. But before checking that,
5119 * make sure the computed size of the snapshot context we
5120 * allocate is representable in a size_t.
5122 if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
5127 if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
5131 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
5137 for (i = 0; i < snap_count; i++)
5138 snapc->snaps[i] = ceph_decode_64(&p);
5140 ceph_put_snap_context(rbd_dev->header.snapc);
5141 rbd_dev->header.snapc = snapc;
5143 dout(" snap context seq = %llu, snap_count = %u\n",
5144 (unsigned long long)seq, (unsigned int)snap_count);
5151 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
5162 size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
5163 reply_buf = kmalloc(size, GFP_KERNEL);
5165 return ERR_PTR(-ENOMEM);
5167 snapid = cpu_to_le64(snap_id);
5168 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
5169 &rbd_dev->header_oloc, "get_snapshot_name",
5170 &snapid, sizeof(snapid), reply_buf, size);
5171 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5173 snap_name = ERR_PTR(ret);
5178 end = reply_buf + ret;
5179 snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
5180 if (IS_ERR(snap_name))
5183 dout(" snap_id 0x%016llx snap_name = %s\n",
5184 (unsigned long long)snap_id, snap_name);
5191 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev)
5193 bool first_time = rbd_dev->header.object_prefix == NULL;
5196 ret = rbd_dev_v2_image_size(rbd_dev);
5201 ret = rbd_dev_v2_header_onetime(rbd_dev);
5206 ret = rbd_dev_v2_snap_context(rbd_dev);
5207 if (ret && first_time) {
5208 kfree(rbd_dev->header.object_prefix);
5209 rbd_dev->header.object_prefix = NULL;
5215 static int rbd_dev_header_info(struct rbd_device *rbd_dev)
5217 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
5219 if (rbd_dev->image_format == 1)
5220 return rbd_dev_v1_header_info(rbd_dev);
5222 return rbd_dev_v2_header_info(rbd_dev);
5226 * Skips over white space at *buf, and updates *buf to point to the
5227 * first found non-space character (if any). Returns the length of
5228 * the token (string of non-white space characters) found. Note
5229 * that *buf must be terminated with '\0'.
5231 static inline size_t next_token(const char **buf)
5234 * These are the characters that produce nonzero for
5235 * isspace() in the "C" and "POSIX" locales.
5237 const char *spaces = " \f\n\r\t\v";
5239 *buf += strspn(*buf, spaces); /* Find start of token */
5241 return strcspn(*buf, spaces); /* Return token length */
5245 * Finds the next token in *buf, dynamically allocates a buffer big
5246 * enough to hold a copy of it, and copies the token into the new
5247 * buffer. The copy is guaranteed to be terminated with '\0'. Note
5248 * that a duplicate buffer is created even for a zero-length token.
5250 * Returns a pointer to the newly-allocated duplicate, or a null
5251 * pointer if memory for the duplicate was not available. If
5252 * the lenp argument is a non-null pointer, the length of the token
5253 * (not including the '\0') is returned in *lenp.
5255 * If successful, the *buf pointer will be updated to point beyond
5256 * the end of the found token.
5258 * Note: uses GFP_KERNEL for allocation.
5260 static inline char *dup_token(const char **buf, size_t *lenp)
5265 len = next_token(buf);
5266 dup = kmemdup(*buf, len + 1, GFP_KERNEL);
5269 *(dup + len) = '\0';
5279 * Parse the options provided for an "rbd add" (i.e., rbd image
5280 * mapping) request. These arrive via a write to /sys/bus/rbd/add,
5281 * and the data written is passed here via a NUL-terminated buffer.
5282 * Returns 0 if successful or an error code otherwise.
5284 * The information extracted from these options is recorded in
5285 * the other parameters which return dynamically-allocated
5288 * The address of a pointer that will refer to a ceph options
5289 * structure. Caller must release the returned pointer using
5290 * ceph_destroy_options() when it is no longer needed.
5292 * Address of an rbd options pointer. Fully initialized by
5293 * this function; caller must release with kfree().
5295 * Address of an rbd image specification pointer. Fully
5296 * initialized by this function based on parsed options.
5297 * Caller must release with rbd_spec_put().
5299 * The options passed take this form:
5300 * <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
5303 * A comma-separated list of one or more monitor addresses.
5304 * A monitor address is an ip address, optionally followed
5305 * by a port number (separated by a colon).
5306 * I.e.: ip1[:port1][,ip2[:port2]...]
5308 * A comma-separated list of ceph and/or rbd options.
5310 * The name of the rados pool containing the rbd image.
5312 * The name of the image in that pool to map.
5314 * An optional snapshot id. If provided, the mapping will
5315 * present data from the image at the time that snapshot was
5316 * created. The image head is used if no snapshot id is
5317 * provided. Snapshot mappings are always read-only.
5319 static int rbd_add_parse_args(const char *buf,
5320 struct ceph_options **ceph_opts,
5321 struct rbd_options **opts,
5322 struct rbd_spec **rbd_spec)
5326 const char *mon_addrs;
5328 size_t mon_addrs_size;
5329 struct parse_rbd_opts_ctx pctx = { 0 };
5330 struct ceph_options *copts;
5333 /* The first four tokens are required */
5335 len = next_token(&buf);
5337 rbd_warn(NULL, "no monitor address(es) provided");
5341 mon_addrs_size = len + 1;
5345 options = dup_token(&buf, NULL);
5349 rbd_warn(NULL, "no options provided");
5353 pctx.spec = rbd_spec_alloc();
5357 pctx.spec->pool_name = dup_token(&buf, NULL);
5358 if (!pctx.spec->pool_name)
5360 if (!*pctx.spec->pool_name) {
5361 rbd_warn(NULL, "no pool name provided");
5365 pctx.spec->image_name = dup_token(&buf, NULL);
5366 if (!pctx.spec->image_name)
5368 if (!*pctx.spec->image_name) {
5369 rbd_warn(NULL, "no image name provided");
5374 * Snapshot name is optional; default is to use "-"
5375 * (indicating the head/no snapshot).
5377 len = next_token(&buf);
5379 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
5380 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
5381 } else if (len > RBD_MAX_SNAP_NAME_LEN) {
5382 ret = -ENAMETOOLONG;
5385 snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
5388 *(snap_name + len) = '\0';
5389 pctx.spec->snap_name = snap_name;
5391 /* Initialize all rbd options to the defaults */
5393 pctx.opts = kzalloc(sizeof(*pctx.opts), GFP_KERNEL);
5397 pctx.opts->read_only = RBD_READ_ONLY_DEFAULT;
5398 pctx.opts->queue_depth = RBD_QUEUE_DEPTH_DEFAULT;
5399 pctx.opts->lock_timeout = RBD_LOCK_TIMEOUT_DEFAULT;
5400 pctx.opts->lock_on_read = RBD_LOCK_ON_READ_DEFAULT;
5401 pctx.opts->exclusive = RBD_EXCLUSIVE_DEFAULT;
5402 pctx.opts->trim = RBD_TRIM_DEFAULT;
5404 copts = ceph_parse_options(options, mon_addrs,
5405 mon_addrs + mon_addrs_size - 1,
5406 parse_rbd_opts_token, &pctx);
5407 if (IS_ERR(copts)) {
5408 ret = PTR_ERR(copts);
5415 *rbd_spec = pctx.spec;
5422 rbd_spec_put(pctx.spec);
5428 static void rbd_dev_image_unlock(struct rbd_device *rbd_dev)
5430 down_write(&rbd_dev->lock_rwsem);
5431 if (__rbd_is_lock_owner(rbd_dev))
5432 rbd_unlock(rbd_dev);
5433 up_write(&rbd_dev->lock_rwsem);
5436 static int rbd_add_acquire_lock(struct rbd_device *rbd_dev)
5440 if (!(rbd_dev->header.features & RBD_FEATURE_EXCLUSIVE_LOCK)) {
5441 rbd_warn(rbd_dev, "exclusive-lock feature is not enabled");
5445 /* FIXME: "rbd map --exclusive" should be in interruptible */
5446 down_read(&rbd_dev->lock_rwsem);
5447 ret = rbd_wait_state_locked(rbd_dev, true);
5448 up_read(&rbd_dev->lock_rwsem);
5450 rbd_warn(rbd_dev, "failed to acquire exclusive lock");
5458 * An rbd format 2 image has a unique identifier, distinct from the
5459 * name given to it by the user. Internally, that identifier is
5460 * what's used to specify the names of objects related to the image.
5462 * A special "rbd id" object is used to map an rbd image name to its
5463 * id. If that object doesn't exist, then there is no v2 rbd image
5464 * with the supplied name.
5466 * This function will record the given rbd_dev's image_id field if
5467 * it can be determined, and in that case will return 0. If any
5468 * errors occur a negative errno will be returned and the rbd_dev's
5469 * image_id field will be unchanged (and should be NULL).
5471 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
5475 CEPH_DEFINE_OID_ONSTACK(oid);
5480 * When probing a parent image, the image id is already
5481 * known (and the image name likely is not). There's no
5482 * need to fetch the image id again in this case. We
5483 * do still need to set the image format though.
5485 if (rbd_dev->spec->image_id) {
5486 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
5492 * First, see if the format 2 image id file exists, and if
5493 * so, get the image's persistent id from it.
5495 ret = ceph_oid_aprintf(&oid, GFP_KERNEL, "%s%s", RBD_ID_PREFIX,
5496 rbd_dev->spec->image_name);
5500 dout("rbd id object name is %s\n", oid.name);
5502 /* Response will be an encoded string, which includes a length */
5504 size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
5505 response = kzalloc(size, GFP_NOIO);
5511 /* If it doesn't exist we'll assume it's a format 1 image */
5513 ret = rbd_obj_method_sync(rbd_dev, &oid, &rbd_dev->header_oloc,
5515 response, RBD_IMAGE_ID_LEN_MAX);
5516 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5517 if (ret == -ENOENT) {
5518 image_id = kstrdup("", GFP_KERNEL);
5519 ret = image_id ? 0 : -ENOMEM;
5521 rbd_dev->image_format = 1;
5522 } else if (ret >= 0) {
5525 image_id = ceph_extract_encoded_string(&p, p + ret,
5527 ret = PTR_ERR_OR_ZERO(image_id);
5529 rbd_dev->image_format = 2;
5533 rbd_dev->spec->image_id = image_id;
5534 dout("image_id is %s\n", image_id);
5538 ceph_oid_destroy(&oid);
5543 * Undo whatever state changes are made by v1 or v2 header info
5546 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
5548 struct rbd_image_header *header;
5550 rbd_dev_parent_put(rbd_dev);
5552 /* Free dynamic fields from the header, then zero it out */
5554 header = &rbd_dev->header;
5555 ceph_put_snap_context(header->snapc);
5556 kfree(header->snap_sizes);
5557 kfree(header->snap_names);
5558 kfree(header->object_prefix);
5559 memset(header, 0, sizeof (*header));
5562 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev)
5566 ret = rbd_dev_v2_object_prefix(rbd_dev);
5571 * Get the and check features for the image. Currently the
5572 * features are assumed to never change.
5574 ret = rbd_dev_v2_features(rbd_dev);
5578 /* If the image supports fancy striping, get its parameters */
5580 if (rbd_dev->header.features & RBD_FEATURE_STRIPINGV2) {
5581 ret = rbd_dev_v2_striping_info(rbd_dev);
5586 if (rbd_dev->header.features & RBD_FEATURE_DATA_POOL) {
5587 ret = rbd_dev_v2_data_pool(rbd_dev);
5592 rbd_init_layout(rbd_dev);
5596 rbd_dev->header.features = 0;
5597 kfree(rbd_dev->header.object_prefix);
5598 rbd_dev->header.object_prefix = NULL;
5603 * @depth is rbd_dev_image_probe() -> rbd_dev_probe_parent() ->
5604 * rbd_dev_image_probe() recursion depth, which means it's also the
5605 * length of the already discovered part of the parent chain.
5607 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev, int depth)
5609 struct rbd_device *parent = NULL;
5612 if (!rbd_dev->parent_spec)
5615 if (++depth > RBD_MAX_PARENT_CHAIN_LEN) {
5616 pr_info("parent chain is too long (%d)\n", depth);
5621 parent = __rbd_dev_create(rbd_dev->rbd_client, rbd_dev->parent_spec);
5628 * Images related by parent/child relationships always share
5629 * rbd_client and spec/parent_spec, so bump their refcounts.
5631 __rbd_get_client(rbd_dev->rbd_client);
5632 rbd_spec_get(rbd_dev->parent_spec);
5634 ret = rbd_dev_image_probe(parent, depth);
5638 rbd_dev->parent = parent;
5639 atomic_set(&rbd_dev->parent_ref, 1);
5643 rbd_dev_unparent(rbd_dev);
5644 rbd_dev_destroy(parent);
5648 static void rbd_dev_device_release(struct rbd_device *rbd_dev)
5650 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5651 rbd_dev_mapping_clear(rbd_dev);
5652 rbd_free_disk(rbd_dev);
5654 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5658 * rbd_dev->header_rwsem must be locked for write and will be unlocked
5661 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
5665 /* Record our major and minor device numbers. */
5667 if (!single_major) {
5668 ret = register_blkdev(0, rbd_dev->name);
5670 goto err_out_unlock;
5672 rbd_dev->major = ret;
5675 rbd_dev->major = rbd_major;
5676 rbd_dev->minor = rbd_dev_id_to_minor(rbd_dev->dev_id);
5679 /* Set up the blkdev mapping. */
5681 ret = rbd_init_disk(rbd_dev);
5683 goto err_out_blkdev;
5685 ret = rbd_dev_mapping_set(rbd_dev);
5689 set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
5690 set_disk_ro(rbd_dev->disk, rbd_dev->opts->read_only);
5692 ret = dev_set_name(&rbd_dev->dev, "%d", rbd_dev->dev_id);
5694 goto err_out_mapping;
5696 set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5697 up_write(&rbd_dev->header_rwsem);
5701 rbd_dev_mapping_clear(rbd_dev);
5703 rbd_free_disk(rbd_dev);
5706 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5708 up_write(&rbd_dev->header_rwsem);
5712 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
5714 struct rbd_spec *spec = rbd_dev->spec;
5717 /* Record the header object name for this rbd image. */
5719 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
5720 if (rbd_dev->image_format == 1)
5721 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
5722 spec->image_name, RBD_SUFFIX);
5724 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
5725 RBD_HEADER_PREFIX, spec->image_id);
5730 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
5733 rbd_unregister_watch(rbd_dev);
5735 rbd_dev_unprobe(rbd_dev);
5736 rbd_dev->image_format = 0;
5737 kfree(rbd_dev->spec->image_id);
5738 rbd_dev->spec->image_id = NULL;
5742 * Probe for the existence of the header object for the given rbd
5743 * device. If this image is the one being mapped (i.e., not a
5744 * parent), initiate a watch on its header object before using that
5745 * object to get detailed information about the rbd image.
5747 * On success, returns with header_rwsem held for write if called
5750 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth)
5755 * Get the id from the image id object. Unless there's an
5756 * error, rbd_dev->spec->image_id will be filled in with
5757 * a dynamically-allocated string, and rbd_dev->image_format
5758 * will be set to either 1 or 2.
5760 ret = rbd_dev_image_id(rbd_dev);
5764 ret = rbd_dev_header_name(rbd_dev);
5766 goto err_out_format;
5769 ret = rbd_register_watch(rbd_dev);
5772 pr_info("image %s/%s%s%s does not exist\n",
5773 rbd_dev->spec->pool_name,
5774 rbd_dev->spec->pool_ns ?: "",
5775 rbd_dev->spec->pool_ns ? "/" : "",
5776 rbd_dev->spec->image_name);
5777 goto err_out_format;
5782 down_write(&rbd_dev->header_rwsem);
5784 ret = rbd_dev_header_info(rbd_dev);
5789 * If this image is the one being mapped, we have pool name and
5790 * id, image name and id, and snap name - need to fill snap id.
5791 * Otherwise this is a parent image, identified by pool, image
5792 * and snap ids - need to fill in names for those ids.
5795 ret = rbd_spec_fill_snap_id(rbd_dev);
5797 ret = rbd_spec_fill_names(rbd_dev);
5800 pr_info("snap %s/%s%s%s@%s does not exist\n",
5801 rbd_dev->spec->pool_name,
5802 rbd_dev->spec->pool_ns ?: "",
5803 rbd_dev->spec->pool_ns ? "/" : "",
5804 rbd_dev->spec->image_name,
5805 rbd_dev->spec->snap_name);
5809 if (rbd_dev->header.features & RBD_FEATURE_LAYERING) {
5810 ret = rbd_dev_v2_parent_info(rbd_dev);
5815 * Need to warn users if this image is the one being
5816 * mapped and has a parent.
5818 if (!depth && rbd_dev->parent_spec)
5820 "WARNING: kernel layering is EXPERIMENTAL!");
5823 ret = rbd_dev_probe_parent(rbd_dev, depth);
5827 dout("discovered format %u image, header name is %s\n",
5828 rbd_dev->image_format, rbd_dev->header_oid.name);
5833 up_write(&rbd_dev->header_rwsem);
5835 rbd_unregister_watch(rbd_dev);
5836 rbd_dev_unprobe(rbd_dev);
5838 rbd_dev->image_format = 0;
5839 kfree(rbd_dev->spec->image_id);
5840 rbd_dev->spec->image_id = NULL;
5844 static ssize_t do_rbd_add(struct bus_type *bus,
5848 struct rbd_device *rbd_dev = NULL;
5849 struct ceph_options *ceph_opts = NULL;
5850 struct rbd_options *rbd_opts = NULL;
5851 struct rbd_spec *spec = NULL;
5852 struct rbd_client *rbdc;
5855 if (!capable(CAP_SYS_ADMIN))
5858 if (!try_module_get(THIS_MODULE))
5861 /* parse add command */
5862 rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
5866 rbdc = rbd_get_client(ceph_opts);
5873 rc = ceph_pg_poolid_by_name(rbdc->client->osdc.osdmap, spec->pool_name);
5876 pr_info("pool %s does not exist\n", spec->pool_name);
5877 goto err_out_client;
5879 spec->pool_id = (u64)rc;
5881 rbd_dev = rbd_dev_create(rbdc, spec, rbd_opts);
5884 goto err_out_client;
5886 rbdc = NULL; /* rbd_dev now owns this */
5887 spec = NULL; /* rbd_dev now owns this */
5888 rbd_opts = NULL; /* rbd_dev now owns this */
5890 rbd_dev->config_info = kstrdup(buf, GFP_KERNEL);
5891 if (!rbd_dev->config_info) {
5893 goto err_out_rbd_dev;
5896 rc = rbd_dev_image_probe(rbd_dev, 0);
5898 goto err_out_rbd_dev;
5900 /* If we are mapping a snapshot it must be marked read-only */
5901 if (rbd_dev->spec->snap_id != CEPH_NOSNAP)
5902 rbd_dev->opts->read_only = true;
5904 rc = rbd_dev_device_setup(rbd_dev);
5906 goto err_out_image_probe;
5908 if (rbd_dev->opts->exclusive) {
5909 rc = rbd_add_acquire_lock(rbd_dev);
5911 goto err_out_device_setup;
5914 /* Everything's ready. Announce the disk to the world. */
5916 rc = device_add(&rbd_dev->dev);
5918 goto err_out_image_lock;
5920 add_disk(rbd_dev->disk);
5921 /* see rbd_init_disk() */
5922 blk_put_queue(rbd_dev->disk->queue);
5924 spin_lock(&rbd_dev_list_lock);
5925 list_add_tail(&rbd_dev->node, &rbd_dev_list);
5926 spin_unlock(&rbd_dev_list_lock);
5928 pr_info("%s: capacity %llu features 0x%llx\n", rbd_dev->disk->disk_name,
5929 (unsigned long long)get_capacity(rbd_dev->disk) << SECTOR_SHIFT,
5930 rbd_dev->header.features);
5933 module_put(THIS_MODULE);
5937 rbd_dev_image_unlock(rbd_dev);
5938 err_out_device_setup:
5939 rbd_dev_device_release(rbd_dev);
5940 err_out_image_probe:
5941 rbd_dev_image_release(rbd_dev);
5943 rbd_dev_destroy(rbd_dev);
5945 rbd_put_client(rbdc);
5952 static ssize_t rbd_add(struct bus_type *bus,
5959 return do_rbd_add(bus, buf, count);
5962 static ssize_t rbd_add_single_major(struct bus_type *bus,
5966 return do_rbd_add(bus, buf, count);
5969 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
5971 while (rbd_dev->parent) {
5972 struct rbd_device *first = rbd_dev;
5973 struct rbd_device *second = first->parent;
5974 struct rbd_device *third;
5977 * Follow to the parent with no grandparent and
5980 while (second && (third = second->parent)) {
5985 rbd_dev_image_release(second);
5986 rbd_dev_destroy(second);
5987 first->parent = NULL;
5988 first->parent_overlap = 0;
5990 rbd_assert(first->parent_spec);
5991 rbd_spec_put(first->parent_spec);
5992 first->parent_spec = NULL;
5996 static ssize_t do_rbd_remove(struct bus_type *bus,
6000 struct rbd_device *rbd_dev = NULL;
6001 struct list_head *tmp;
6007 if (!capable(CAP_SYS_ADMIN))
6012 sscanf(buf, "%d %5s", &dev_id, opt_buf);
6014 pr_err("dev_id out of range\n");
6017 if (opt_buf[0] != '\0') {
6018 if (!strcmp(opt_buf, "force")) {
6021 pr_err("bad remove option at '%s'\n", opt_buf);
6027 spin_lock(&rbd_dev_list_lock);
6028 list_for_each(tmp, &rbd_dev_list) {
6029 rbd_dev = list_entry(tmp, struct rbd_device, node);
6030 if (rbd_dev->dev_id == dev_id) {
6036 spin_lock_irq(&rbd_dev->lock);
6037 if (rbd_dev->open_count && !force)
6039 else if (test_and_set_bit(RBD_DEV_FLAG_REMOVING,
6042 spin_unlock_irq(&rbd_dev->lock);
6044 spin_unlock(&rbd_dev_list_lock);
6050 * Prevent new IO from being queued and wait for existing
6051 * IO to complete/fail.
6053 blk_mq_freeze_queue(rbd_dev->disk->queue);
6054 blk_set_queue_dying(rbd_dev->disk->queue);
6057 del_gendisk(rbd_dev->disk);
6058 spin_lock(&rbd_dev_list_lock);
6059 list_del_init(&rbd_dev->node);
6060 spin_unlock(&rbd_dev_list_lock);
6061 device_del(&rbd_dev->dev);
6063 rbd_dev_image_unlock(rbd_dev);
6064 rbd_dev_device_release(rbd_dev);
6065 rbd_dev_image_release(rbd_dev);
6066 rbd_dev_destroy(rbd_dev);
6070 static ssize_t rbd_remove(struct bus_type *bus,
6077 return do_rbd_remove(bus, buf, count);
6080 static ssize_t rbd_remove_single_major(struct bus_type *bus,
6084 return do_rbd_remove(bus, buf, count);
6088 * create control files in sysfs
6091 static int rbd_sysfs_init(void)
6095 ret = device_register(&rbd_root_dev);
6099 ret = bus_register(&rbd_bus_type);
6101 device_unregister(&rbd_root_dev);
6106 static void rbd_sysfs_cleanup(void)
6108 bus_unregister(&rbd_bus_type);
6109 device_unregister(&rbd_root_dev);
6112 static int rbd_slab_init(void)
6114 rbd_assert(!rbd_img_request_cache);
6115 rbd_img_request_cache = KMEM_CACHE(rbd_img_request, 0);
6116 if (!rbd_img_request_cache)
6119 rbd_assert(!rbd_obj_request_cache);
6120 rbd_obj_request_cache = KMEM_CACHE(rbd_obj_request, 0);
6121 if (!rbd_obj_request_cache)
6127 kmem_cache_destroy(rbd_img_request_cache);
6128 rbd_img_request_cache = NULL;
6132 static void rbd_slab_exit(void)
6134 rbd_assert(rbd_obj_request_cache);
6135 kmem_cache_destroy(rbd_obj_request_cache);
6136 rbd_obj_request_cache = NULL;
6138 rbd_assert(rbd_img_request_cache);
6139 kmem_cache_destroy(rbd_img_request_cache);
6140 rbd_img_request_cache = NULL;
6143 static int __init rbd_init(void)
6147 if (!libceph_compatible(NULL)) {
6148 rbd_warn(NULL, "libceph incompatibility (quitting)");
6152 rc = rbd_slab_init();
6157 * The number of active work items is limited by the number of
6158 * rbd devices * queue depth, so leave @max_active at default.
6160 rbd_wq = alloc_workqueue(RBD_DRV_NAME, WQ_MEM_RECLAIM, 0);
6167 rbd_major = register_blkdev(0, RBD_DRV_NAME);
6168 if (rbd_major < 0) {
6174 rc = rbd_sysfs_init();
6176 goto err_out_blkdev;
6179 pr_info("loaded (major %d)\n", rbd_major);
6181 pr_info("loaded\n");
6187 unregister_blkdev(rbd_major, RBD_DRV_NAME);
6189 destroy_workqueue(rbd_wq);
6195 static void __exit rbd_exit(void)
6197 ida_destroy(&rbd_dev_id_ida);
6198 rbd_sysfs_cleanup();
6200 unregister_blkdev(rbd_major, RBD_DRV_NAME);
6201 destroy_workqueue(rbd_wq);
6205 module_init(rbd_init);
6206 module_exit(rbd_exit);
6208 MODULE_AUTHOR("Alex Elder <elder@inktank.com>");
6209 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
6210 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
6211 /* following authorship retained from original osdblk.c */
6212 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
6214 MODULE_DESCRIPTION("RADOS Block Device (RBD) driver");
6215 MODULE_LICENSE("GPL");