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/fs_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_OBJECT_MAP (1ULL<<3)
119 #define RBD_FEATURE_FAST_DIFF (1ULL<<4)
120 #define RBD_FEATURE_DEEP_FLATTEN (1ULL<<5)
121 #define RBD_FEATURE_DATA_POOL (1ULL<<7)
122 #define RBD_FEATURE_OPERATIONS (1ULL<<8)
124 #define RBD_FEATURES_ALL (RBD_FEATURE_LAYERING | \
125 RBD_FEATURE_STRIPINGV2 | \
126 RBD_FEATURE_EXCLUSIVE_LOCK | \
127 RBD_FEATURE_OBJECT_MAP | \
128 RBD_FEATURE_FAST_DIFF | \
129 RBD_FEATURE_DEEP_FLATTEN | \
130 RBD_FEATURE_DATA_POOL | \
131 RBD_FEATURE_OPERATIONS)
133 /* Features supported by this (client software) implementation. */
135 #define RBD_FEATURES_SUPPORTED (RBD_FEATURES_ALL)
138 * An RBD device name will be "rbd#", where the "rbd" comes from
139 * RBD_DRV_NAME above, and # is a unique integer identifier.
141 #define DEV_NAME_LEN 32
144 * block device image metadata (in-memory version)
146 struct rbd_image_header {
147 /* These six fields never change for a given rbd image */
153 u64 features; /* Might be changeable someday? */
155 /* The remaining fields need to be updated occasionally */
157 struct ceph_snap_context *snapc;
158 char *snap_names; /* format 1 only */
159 u64 *snap_sizes; /* format 1 only */
163 * An rbd image specification.
165 * The tuple (pool_id, image_id, snap_id) is sufficient to uniquely
166 * identify an image. Each rbd_dev structure includes a pointer to
167 * an rbd_spec structure that encapsulates this identity.
169 * Each of the id's in an rbd_spec has an associated name. For a
170 * user-mapped image, the names are supplied and the id's associated
171 * with them are looked up. For a layered image, a parent image is
172 * defined by the tuple, and the names are looked up.
174 * An rbd_dev structure contains a parent_spec pointer which is
175 * non-null if the image it represents is a child in a layered
176 * image. This pointer will refer to the rbd_spec structure used
177 * by the parent rbd_dev for its own identity (i.e., the structure
178 * is shared between the parent and child).
180 * Since these structures are populated once, during the discovery
181 * phase of image construction, they are effectively immutable so
182 * we make no effort to synchronize access to them.
184 * Note that code herein does not assume the image name is known (it
185 * could be a null pointer).
189 const char *pool_name;
190 const char *pool_ns; /* NULL if default, never "" */
192 const char *image_id;
193 const char *image_name;
196 const char *snap_name;
202 * an instance of the client. multiple devices may share an rbd client.
205 struct ceph_client *client;
207 struct list_head node;
210 struct pending_result {
211 int result; /* first nonzero result */
215 struct rbd_img_request;
217 enum obj_request_type {
218 OBJ_REQUEST_NODATA = 1,
219 OBJ_REQUEST_BIO, /* pointer into provided bio (list) */
220 OBJ_REQUEST_BVECS, /* pointer into provided bio_vec array */
221 OBJ_REQUEST_OWN_BVECS, /* private bio_vec array, doesn't own pages */
224 enum obj_operation_type {
231 #define RBD_OBJ_FLAG_DELETION (1U << 0)
232 #define RBD_OBJ_FLAG_COPYUP_ENABLED (1U << 1)
233 #define RBD_OBJ_FLAG_COPYUP_ZEROS (1U << 2)
234 #define RBD_OBJ_FLAG_MAY_EXIST (1U << 3)
235 #define RBD_OBJ_FLAG_NOOP_FOR_NONEXISTENT (1U << 4)
237 enum rbd_obj_read_state {
238 RBD_OBJ_READ_START = 1,
244 * Writes go through the following state machine to deal with
247 * . . . . . RBD_OBJ_WRITE_GUARD. . . . . . . . . . . . . .
250 * . RBD_OBJ_WRITE_READ_FROM_PARENT. . . .
252 * . v v (deep-copyup .
253 * (image . RBD_OBJ_WRITE_COPYUP_EMPTY_SNAPC . not needed) .
256 * . . . .RBD_OBJ_WRITE_COPYUP_OPS. . . . . (copyup .
259 * done . . . . . . . . . . . . . . . . . .
264 * Writes start in RBD_OBJ_WRITE_GUARD or _FLAT, depending on whether
265 * assert_exists guard is needed or not (in some cases it's not needed
266 * even if there is a parent).
268 enum rbd_obj_write_state {
269 RBD_OBJ_WRITE_START = 1,
270 RBD_OBJ_WRITE_PRE_OBJECT_MAP,
271 RBD_OBJ_WRITE_OBJECT,
272 __RBD_OBJ_WRITE_COPYUP,
273 RBD_OBJ_WRITE_COPYUP,
274 RBD_OBJ_WRITE_POST_OBJECT_MAP,
277 enum rbd_obj_copyup_state {
278 RBD_OBJ_COPYUP_START = 1,
279 RBD_OBJ_COPYUP_READ_PARENT,
280 __RBD_OBJ_COPYUP_OBJECT_MAPS,
281 RBD_OBJ_COPYUP_OBJECT_MAPS,
282 __RBD_OBJ_COPYUP_WRITE_OBJECT,
283 RBD_OBJ_COPYUP_WRITE_OBJECT,
286 struct rbd_obj_request {
287 struct ceph_object_extent ex;
288 unsigned int flags; /* RBD_OBJ_FLAG_* */
290 enum rbd_obj_read_state read_state; /* for reads */
291 enum rbd_obj_write_state write_state; /* for writes */
294 struct rbd_img_request *img_request;
295 struct ceph_file_extent *img_extents;
299 struct ceph_bio_iter bio_pos;
301 struct ceph_bvec_iter bvec_pos;
307 enum rbd_obj_copyup_state copyup_state;
308 struct bio_vec *copyup_bvecs;
309 u32 copyup_bvec_count;
311 struct list_head osd_reqs; /* w/ r_private_item */
313 struct mutex state_mutex;
314 struct pending_result pending;
319 IMG_REQ_CHILD, /* initiator: block = 0, child image = 1 */
320 IMG_REQ_LAYERED, /* ENOENT handling: normal = 0, layered = 1 */
325 RBD_IMG_EXCLUSIVE_LOCK,
326 __RBD_IMG_OBJECT_REQUESTS,
327 RBD_IMG_OBJECT_REQUESTS,
330 struct rbd_img_request {
331 struct rbd_device *rbd_dev;
332 enum obj_operation_type op_type;
333 enum obj_request_type data_type;
335 enum rbd_img_state state;
337 u64 snap_id; /* for reads */
338 struct ceph_snap_context *snapc; /* for writes */
340 struct rbd_obj_request *obj_request; /* obj req initiator */
342 struct list_head lock_item;
343 struct list_head object_extents; /* obj_req.ex structs */
345 struct mutex state_mutex;
346 struct pending_result pending;
347 struct work_struct work;
351 #define for_each_obj_request(ireq, oreq) \
352 list_for_each_entry(oreq, &(ireq)->object_extents, ex.oe_item)
353 #define for_each_obj_request_safe(ireq, oreq, n) \
354 list_for_each_entry_safe(oreq, n, &(ireq)->object_extents, ex.oe_item)
356 enum rbd_watch_state {
357 RBD_WATCH_STATE_UNREGISTERED,
358 RBD_WATCH_STATE_REGISTERED,
359 RBD_WATCH_STATE_ERROR,
362 enum rbd_lock_state {
363 RBD_LOCK_STATE_UNLOCKED,
364 RBD_LOCK_STATE_LOCKED,
365 RBD_LOCK_STATE_RELEASING,
368 /* WatchNotify::ClientId */
369 struct rbd_client_id {
382 int dev_id; /* blkdev unique id */
384 int major; /* blkdev assigned major */
386 struct gendisk *disk; /* blkdev's gendisk and rq */
388 u32 image_format; /* Either 1 or 2 */
389 struct rbd_client *rbd_client;
391 char name[DEV_NAME_LEN]; /* blkdev name, e.g. rbd3 */
393 spinlock_t lock; /* queue, flags, open_count */
395 struct rbd_image_header header;
396 unsigned long flags; /* possibly lock protected */
397 struct rbd_spec *spec;
398 struct rbd_options *opts;
399 char *config_info; /* add{,_single_major} string */
401 struct ceph_object_id header_oid;
402 struct ceph_object_locator header_oloc;
404 struct ceph_file_layout layout; /* used for all rbd requests */
406 struct mutex watch_mutex;
407 enum rbd_watch_state watch_state;
408 struct ceph_osd_linger_request *watch_handle;
410 struct delayed_work watch_dwork;
412 struct rw_semaphore lock_rwsem;
413 enum rbd_lock_state lock_state;
414 char lock_cookie[32];
415 struct rbd_client_id owner_cid;
416 struct work_struct acquired_lock_work;
417 struct work_struct released_lock_work;
418 struct delayed_work lock_dwork;
419 struct work_struct unlock_work;
420 spinlock_t lock_lists_lock;
421 struct list_head acquiring_list;
422 struct list_head running_list;
423 struct completion acquire_wait;
425 struct completion releasing_wait;
427 spinlock_t object_map_lock;
429 u64 object_map_size; /* in objects */
430 u64 object_map_flags;
432 struct workqueue_struct *task_wq;
434 struct rbd_spec *parent_spec;
437 struct rbd_device *parent;
439 /* Block layer tags. */
440 struct blk_mq_tag_set tag_set;
442 /* protects updating the header */
443 struct rw_semaphore header_rwsem;
445 struct rbd_mapping mapping;
447 struct list_head node;
451 unsigned long open_count; /* protected by lock */
455 * Flag bits for rbd_dev->flags:
456 * - REMOVING (which is coupled with rbd_dev->open_count) is protected
460 RBD_DEV_FLAG_EXISTS, /* rbd_dev_device_setup() ran */
461 RBD_DEV_FLAG_REMOVING, /* this mapping is being removed */
462 RBD_DEV_FLAG_READONLY, /* -o ro or snapshot */
465 static DEFINE_MUTEX(client_mutex); /* Serialize client creation */
467 static LIST_HEAD(rbd_dev_list); /* devices */
468 static DEFINE_SPINLOCK(rbd_dev_list_lock);
470 static LIST_HEAD(rbd_client_list); /* clients */
471 static DEFINE_SPINLOCK(rbd_client_list_lock);
473 /* Slab caches for frequently-allocated structures */
475 static struct kmem_cache *rbd_img_request_cache;
476 static struct kmem_cache *rbd_obj_request_cache;
478 static int rbd_major;
479 static DEFINE_IDA(rbd_dev_id_ida);
481 static struct workqueue_struct *rbd_wq;
483 static struct ceph_snap_context rbd_empty_snapc = {
484 .nref = REFCOUNT_INIT(1),
488 * single-major requires >= 0.75 version of userspace rbd utility.
490 static bool single_major = true;
491 module_param(single_major, bool, 0444);
492 MODULE_PARM_DESC(single_major, "Use a single major number for all rbd devices (default: true)");
494 static ssize_t add_store(const struct bus_type *bus, const char *buf, size_t count);
495 static ssize_t remove_store(const struct bus_type *bus, const char *buf,
497 static ssize_t add_single_major_store(const struct bus_type *bus, const char *buf,
499 static ssize_t remove_single_major_store(const struct bus_type *bus, const char *buf,
501 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth);
503 static int rbd_dev_id_to_minor(int dev_id)
505 return dev_id << RBD_SINGLE_MAJOR_PART_SHIFT;
508 static int minor_to_rbd_dev_id(int minor)
510 return minor >> RBD_SINGLE_MAJOR_PART_SHIFT;
513 static bool rbd_is_ro(struct rbd_device *rbd_dev)
515 return test_bit(RBD_DEV_FLAG_READONLY, &rbd_dev->flags);
518 static bool rbd_is_snap(struct rbd_device *rbd_dev)
520 return rbd_dev->spec->snap_id != CEPH_NOSNAP;
523 static bool __rbd_is_lock_owner(struct rbd_device *rbd_dev)
525 lockdep_assert_held(&rbd_dev->lock_rwsem);
527 return rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED ||
528 rbd_dev->lock_state == RBD_LOCK_STATE_RELEASING;
531 static bool rbd_is_lock_owner(struct rbd_device *rbd_dev)
535 down_read(&rbd_dev->lock_rwsem);
536 is_lock_owner = __rbd_is_lock_owner(rbd_dev);
537 up_read(&rbd_dev->lock_rwsem);
538 return is_lock_owner;
541 static ssize_t supported_features_show(const struct bus_type *bus, char *buf)
543 return sprintf(buf, "0x%llx\n", RBD_FEATURES_SUPPORTED);
546 static BUS_ATTR_WO(add);
547 static BUS_ATTR_WO(remove);
548 static BUS_ATTR_WO(add_single_major);
549 static BUS_ATTR_WO(remove_single_major);
550 static BUS_ATTR_RO(supported_features);
552 static struct attribute *rbd_bus_attrs[] = {
554 &bus_attr_remove.attr,
555 &bus_attr_add_single_major.attr,
556 &bus_attr_remove_single_major.attr,
557 &bus_attr_supported_features.attr,
561 static umode_t rbd_bus_is_visible(struct kobject *kobj,
562 struct attribute *attr, int index)
565 (attr == &bus_attr_add_single_major.attr ||
566 attr == &bus_attr_remove_single_major.attr))
572 static const struct attribute_group rbd_bus_group = {
573 .attrs = rbd_bus_attrs,
574 .is_visible = rbd_bus_is_visible,
576 __ATTRIBUTE_GROUPS(rbd_bus);
578 static const struct bus_type rbd_bus_type = {
580 .bus_groups = rbd_bus_groups,
583 static void rbd_root_dev_release(struct device *dev)
587 static struct device rbd_root_dev = {
589 .release = rbd_root_dev_release,
592 static __printf(2, 3)
593 void rbd_warn(struct rbd_device *rbd_dev, const char *fmt, ...)
595 struct va_format vaf;
603 printk(KERN_WARNING "%s: %pV\n", RBD_DRV_NAME, &vaf);
604 else if (rbd_dev->disk)
605 printk(KERN_WARNING "%s: %s: %pV\n",
606 RBD_DRV_NAME, rbd_dev->disk->disk_name, &vaf);
607 else if (rbd_dev->spec && rbd_dev->spec->image_name)
608 printk(KERN_WARNING "%s: image %s: %pV\n",
609 RBD_DRV_NAME, rbd_dev->spec->image_name, &vaf);
610 else if (rbd_dev->spec && rbd_dev->spec->image_id)
611 printk(KERN_WARNING "%s: id %s: %pV\n",
612 RBD_DRV_NAME, rbd_dev->spec->image_id, &vaf);
614 printk(KERN_WARNING "%s: rbd_dev %p: %pV\n",
615 RBD_DRV_NAME, rbd_dev, &vaf);
620 #define rbd_assert(expr) \
621 if (unlikely(!(expr))) { \
622 printk(KERN_ERR "\nAssertion failure in %s() " \
624 "\trbd_assert(%s);\n\n", \
625 __func__, __LINE__, #expr); \
628 #else /* !RBD_DEBUG */
629 # define rbd_assert(expr) ((void) 0)
630 #endif /* !RBD_DEBUG */
632 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev);
634 static int rbd_dev_refresh(struct rbd_device *rbd_dev);
635 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev,
636 struct rbd_image_header *header);
637 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
639 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
640 u8 *order, u64 *snap_size);
641 static int rbd_dev_v2_get_flags(struct rbd_device *rbd_dev);
643 static void rbd_obj_handle_request(struct rbd_obj_request *obj_req, int result);
644 static void rbd_img_handle_request(struct rbd_img_request *img_req, int result);
647 * Return true if nothing else is pending.
649 static bool pending_result_dec(struct pending_result *pending, int *result)
651 rbd_assert(pending->num_pending > 0);
653 if (*result && !pending->result)
654 pending->result = *result;
655 if (--pending->num_pending)
658 *result = pending->result;
662 static int rbd_open(struct gendisk *disk, blk_mode_t mode)
664 struct rbd_device *rbd_dev = disk->private_data;
665 bool removing = false;
667 spin_lock_irq(&rbd_dev->lock);
668 if (test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags))
671 rbd_dev->open_count++;
672 spin_unlock_irq(&rbd_dev->lock);
676 (void) get_device(&rbd_dev->dev);
681 static void rbd_release(struct gendisk *disk)
683 struct rbd_device *rbd_dev = disk->private_data;
684 unsigned long open_count_before;
686 spin_lock_irq(&rbd_dev->lock);
687 open_count_before = rbd_dev->open_count--;
688 spin_unlock_irq(&rbd_dev->lock);
689 rbd_assert(open_count_before > 0);
691 put_device(&rbd_dev->dev);
694 static const struct block_device_operations rbd_bd_ops = {
695 .owner = THIS_MODULE,
697 .release = rbd_release,
701 * Initialize an rbd client instance. Success or not, this function
702 * consumes ceph_opts. Caller holds client_mutex.
704 static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts)
706 struct rbd_client *rbdc;
709 dout("%s:\n", __func__);
710 rbdc = kmalloc(sizeof(struct rbd_client), GFP_KERNEL);
714 kref_init(&rbdc->kref);
715 INIT_LIST_HEAD(&rbdc->node);
717 rbdc->client = ceph_create_client(ceph_opts, rbdc);
718 if (IS_ERR(rbdc->client))
720 ceph_opts = NULL; /* Now rbdc->client is responsible for ceph_opts */
722 ret = ceph_open_session(rbdc->client);
726 spin_lock(&rbd_client_list_lock);
727 list_add_tail(&rbdc->node, &rbd_client_list);
728 spin_unlock(&rbd_client_list_lock);
730 dout("%s: rbdc %p\n", __func__, rbdc);
734 ceph_destroy_client(rbdc->client);
739 ceph_destroy_options(ceph_opts);
740 dout("%s: error %d\n", __func__, ret);
745 static struct rbd_client *__rbd_get_client(struct rbd_client *rbdc)
747 kref_get(&rbdc->kref);
753 * Find a ceph client with specific addr and configuration. If
754 * found, bump its reference count.
756 static struct rbd_client *rbd_client_find(struct ceph_options *ceph_opts)
758 struct rbd_client *rbdc = NULL, *iter;
760 if (ceph_opts->flags & CEPH_OPT_NOSHARE)
763 spin_lock(&rbd_client_list_lock);
764 list_for_each_entry(iter, &rbd_client_list, node) {
765 if (!ceph_compare_options(ceph_opts, iter->client)) {
766 __rbd_get_client(iter);
772 spin_unlock(&rbd_client_list_lock);
778 * (Per device) rbd map options
786 Opt_compression_hint,
787 /* string args above */
796 Opt_compression_hint_none,
797 Opt_compression_hint_compressible,
798 Opt_compression_hint_incompressible,
801 static const struct constant_table rbd_param_compression_hint[] = {
802 {"none", Opt_compression_hint_none},
803 {"compressible", Opt_compression_hint_compressible},
804 {"incompressible", Opt_compression_hint_incompressible},
808 static const struct fs_parameter_spec rbd_parameters[] = {
809 fsparam_u32 ("alloc_size", Opt_alloc_size),
810 fsparam_enum ("compression_hint", Opt_compression_hint,
811 rbd_param_compression_hint),
812 fsparam_flag ("exclusive", Opt_exclusive),
813 fsparam_flag ("lock_on_read", Opt_lock_on_read),
814 fsparam_u32 ("lock_timeout", Opt_lock_timeout),
815 fsparam_flag ("notrim", Opt_notrim),
816 fsparam_string ("_pool_ns", Opt_pool_ns),
817 fsparam_u32 ("queue_depth", Opt_queue_depth),
818 fsparam_flag ("read_only", Opt_read_only),
819 fsparam_flag ("read_write", Opt_read_write),
820 fsparam_flag ("ro", Opt_read_only),
821 fsparam_flag ("rw", Opt_read_write),
828 unsigned long lock_timeout;
834 u32 alloc_hint_flags; /* CEPH_OSD_OP_ALLOC_HINT_FLAG_* */
837 #define RBD_QUEUE_DEPTH_DEFAULT BLKDEV_DEFAULT_RQ
838 #define RBD_ALLOC_SIZE_DEFAULT (64 * 1024)
839 #define RBD_LOCK_TIMEOUT_DEFAULT 0 /* no timeout */
840 #define RBD_READ_ONLY_DEFAULT false
841 #define RBD_LOCK_ON_READ_DEFAULT false
842 #define RBD_EXCLUSIVE_DEFAULT false
843 #define RBD_TRIM_DEFAULT true
845 struct rbd_parse_opts_ctx {
846 struct rbd_spec *spec;
847 struct ceph_options *copts;
848 struct rbd_options *opts;
851 static char* obj_op_name(enum obj_operation_type op_type)
868 * Destroy ceph client
870 * Caller must hold rbd_client_list_lock.
872 static void rbd_client_release(struct kref *kref)
874 struct rbd_client *rbdc = container_of(kref, struct rbd_client, kref);
876 dout("%s: rbdc %p\n", __func__, rbdc);
877 spin_lock(&rbd_client_list_lock);
878 list_del(&rbdc->node);
879 spin_unlock(&rbd_client_list_lock);
881 ceph_destroy_client(rbdc->client);
886 * Drop reference to ceph client node. If it's not referenced anymore, release
889 static void rbd_put_client(struct rbd_client *rbdc)
892 kref_put(&rbdc->kref, rbd_client_release);
896 * Get a ceph client with specific addr and configuration, if one does
897 * not exist create it. Either way, ceph_opts is consumed by this
900 static struct rbd_client *rbd_get_client(struct ceph_options *ceph_opts)
902 struct rbd_client *rbdc;
905 mutex_lock(&client_mutex);
906 rbdc = rbd_client_find(ceph_opts);
908 ceph_destroy_options(ceph_opts);
911 * Using an existing client. Make sure ->pg_pools is up to
912 * date before we look up the pool id in do_rbd_add().
914 ret = ceph_wait_for_latest_osdmap(rbdc->client,
915 rbdc->client->options->mount_timeout);
917 rbd_warn(NULL, "failed to get latest osdmap: %d", ret);
918 rbd_put_client(rbdc);
922 rbdc = rbd_client_create(ceph_opts);
924 mutex_unlock(&client_mutex);
929 static bool rbd_image_format_valid(u32 image_format)
931 return image_format == 1 || image_format == 2;
934 static bool rbd_dev_ondisk_valid(struct rbd_image_header_ondisk *ondisk)
939 /* The header has to start with the magic rbd header text */
940 if (memcmp(&ondisk->text, RBD_HEADER_TEXT, sizeof (RBD_HEADER_TEXT)))
943 /* The bio layer requires at least sector-sized I/O */
945 if (ondisk->options.order < SECTOR_SHIFT)
948 /* If we use u64 in a few spots we may be able to loosen this */
950 if (ondisk->options.order > 8 * sizeof (int) - 1)
954 * The size of a snapshot header has to fit in a size_t, and
955 * that limits the number of snapshots.
957 snap_count = le32_to_cpu(ondisk->snap_count);
958 size = SIZE_MAX - sizeof (struct ceph_snap_context);
959 if (snap_count > size / sizeof (__le64))
963 * Not only that, but the size of the entire the snapshot
964 * header must also be representable in a size_t.
966 size -= snap_count * sizeof (__le64);
967 if ((u64) size < le64_to_cpu(ondisk->snap_names_len))
974 * returns the size of an object in the image
976 static u32 rbd_obj_bytes(struct rbd_image_header *header)
978 return 1U << header->obj_order;
981 static void rbd_init_layout(struct rbd_device *rbd_dev)
983 if (rbd_dev->header.stripe_unit == 0 ||
984 rbd_dev->header.stripe_count == 0) {
985 rbd_dev->header.stripe_unit = rbd_obj_bytes(&rbd_dev->header);
986 rbd_dev->header.stripe_count = 1;
989 rbd_dev->layout.stripe_unit = rbd_dev->header.stripe_unit;
990 rbd_dev->layout.stripe_count = rbd_dev->header.stripe_count;
991 rbd_dev->layout.object_size = rbd_obj_bytes(&rbd_dev->header);
992 rbd_dev->layout.pool_id = rbd_dev->header.data_pool_id == CEPH_NOPOOL ?
993 rbd_dev->spec->pool_id : rbd_dev->header.data_pool_id;
994 RCU_INIT_POINTER(rbd_dev->layout.pool_ns, NULL);
997 static void rbd_image_header_cleanup(struct rbd_image_header *header)
999 kfree(header->object_prefix);
1000 ceph_put_snap_context(header->snapc);
1001 kfree(header->snap_sizes);
1002 kfree(header->snap_names);
1004 memset(header, 0, sizeof(*header));
1008 * Fill an rbd image header with information from the given format 1
1011 static int rbd_header_from_disk(struct rbd_image_header *header,
1012 struct rbd_image_header_ondisk *ondisk,
1015 struct ceph_snap_context *snapc;
1016 char *object_prefix = NULL;
1017 char *snap_names = NULL;
1018 u64 *snap_sizes = NULL;
1023 /* Allocate this now to avoid having to handle failure below */
1026 object_prefix = kstrndup(ondisk->object_prefix,
1027 sizeof(ondisk->object_prefix),
1033 /* Allocate the snapshot context and fill it in */
1035 snap_count = le32_to_cpu(ondisk->snap_count);
1036 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
1039 snapc->seq = le64_to_cpu(ondisk->snap_seq);
1041 struct rbd_image_snap_ondisk *snaps;
1042 u64 snap_names_len = le64_to_cpu(ondisk->snap_names_len);
1044 /* We'll keep a copy of the snapshot names... */
1046 if (snap_names_len > (u64)SIZE_MAX)
1048 snap_names = kmalloc(snap_names_len, GFP_KERNEL);
1052 /* ...as well as the array of their sizes. */
1053 snap_sizes = kmalloc_array(snap_count,
1054 sizeof(*header->snap_sizes),
1060 * Copy the names, and fill in each snapshot's id
1063 * Note that rbd_dev_v1_header_info() guarantees the
1064 * ondisk buffer we're working with has
1065 * snap_names_len bytes beyond the end of the
1066 * snapshot id array, this memcpy() is safe.
1068 memcpy(snap_names, &ondisk->snaps[snap_count], snap_names_len);
1069 snaps = ondisk->snaps;
1070 for (i = 0; i < snap_count; i++) {
1071 snapc->snaps[i] = le64_to_cpu(snaps[i].id);
1072 snap_sizes[i] = le64_to_cpu(snaps[i].image_size);
1076 /* We won't fail any more, fill in the header */
1079 header->object_prefix = object_prefix;
1080 header->obj_order = ondisk->options.order;
1083 /* The remaining fields always get updated (when we refresh) */
1085 header->image_size = le64_to_cpu(ondisk->image_size);
1086 header->snapc = snapc;
1087 header->snap_names = snap_names;
1088 header->snap_sizes = snap_sizes;
1096 ceph_put_snap_context(snapc);
1097 kfree(object_prefix);
1102 static const char *_rbd_dev_v1_snap_name(struct rbd_device *rbd_dev, u32 which)
1104 const char *snap_name;
1106 rbd_assert(which < rbd_dev->header.snapc->num_snaps);
1108 /* Skip over names until we find the one we are looking for */
1110 snap_name = rbd_dev->header.snap_names;
1112 snap_name += strlen(snap_name) + 1;
1114 return kstrdup(snap_name, GFP_KERNEL);
1118 * Snapshot id comparison function for use with qsort()/bsearch().
1119 * Note that result is for snapshots in *descending* order.
1121 static int snapid_compare_reverse(const void *s1, const void *s2)
1123 u64 snap_id1 = *(u64 *)s1;
1124 u64 snap_id2 = *(u64 *)s2;
1126 if (snap_id1 < snap_id2)
1128 return snap_id1 == snap_id2 ? 0 : -1;
1132 * Search a snapshot context to see if the given snapshot id is
1135 * Returns the position of the snapshot id in the array if it's found,
1136 * or BAD_SNAP_INDEX otherwise.
1138 * Note: The snapshot array is in kept sorted (by the osd) in
1139 * reverse order, highest snapshot id first.
1141 static u32 rbd_dev_snap_index(struct rbd_device *rbd_dev, u64 snap_id)
1143 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
1146 found = bsearch(&snap_id, &snapc->snaps, snapc->num_snaps,
1147 sizeof (snap_id), snapid_compare_reverse);
1149 return found ? (u32)(found - &snapc->snaps[0]) : BAD_SNAP_INDEX;
1152 static const char *rbd_dev_v1_snap_name(struct rbd_device *rbd_dev,
1156 const char *snap_name;
1158 which = rbd_dev_snap_index(rbd_dev, snap_id);
1159 if (which == BAD_SNAP_INDEX)
1160 return ERR_PTR(-ENOENT);
1162 snap_name = _rbd_dev_v1_snap_name(rbd_dev, which);
1163 return snap_name ? snap_name : ERR_PTR(-ENOMEM);
1166 static const char *rbd_snap_name(struct rbd_device *rbd_dev, u64 snap_id)
1168 if (snap_id == CEPH_NOSNAP)
1169 return RBD_SNAP_HEAD_NAME;
1171 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1172 if (rbd_dev->image_format == 1)
1173 return rbd_dev_v1_snap_name(rbd_dev, snap_id);
1175 return rbd_dev_v2_snap_name(rbd_dev, snap_id);
1178 static int rbd_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
1181 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1182 if (snap_id == CEPH_NOSNAP) {
1183 *snap_size = rbd_dev->header.image_size;
1184 } else if (rbd_dev->image_format == 1) {
1187 which = rbd_dev_snap_index(rbd_dev, snap_id);
1188 if (which == BAD_SNAP_INDEX)
1191 *snap_size = rbd_dev->header.snap_sizes[which];
1196 ret = _rbd_dev_v2_snap_size(rbd_dev, snap_id, NULL, &size);
1205 static int rbd_dev_mapping_set(struct rbd_device *rbd_dev)
1207 u64 snap_id = rbd_dev->spec->snap_id;
1211 ret = rbd_snap_size(rbd_dev, snap_id, &size);
1215 rbd_dev->mapping.size = size;
1219 static void rbd_dev_mapping_clear(struct rbd_device *rbd_dev)
1221 rbd_dev->mapping.size = 0;
1224 static void zero_bios(struct ceph_bio_iter *bio_pos, u32 off, u32 bytes)
1226 struct ceph_bio_iter it = *bio_pos;
1228 ceph_bio_iter_advance(&it, off);
1229 ceph_bio_iter_advance_step(&it, bytes, ({
1234 static void zero_bvecs(struct ceph_bvec_iter *bvec_pos, u32 off, u32 bytes)
1236 struct ceph_bvec_iter it = *bvec_pos;
1238 ceph_bvec_iter_advance(&it, off);
1239 ceph_bvec_iter_advance_step(&it, bytes, ({
1245 * Zero a range in @obj_req data buffer defined by a bio (list) or
1246 * (private) bio_vec array.
1248 * @off is relative to the start of the data buffer.
1250 static void rbd_obj_zero_range(struct rbd_obj_request *obj_req, u32 off,
1253 dout("%s %p data buf %u~%u\n", __func__, obj_req, off, bytes);
1255 switch (obj_req->img_request->data_type) {
1256 case OBJ_REQUEST_BIO:
1257 zero_bios(&obj_req->bio_pos, off, bytes);
1259 case OBJ_REQUEST_BVECS:
1260 case OBJ_REQUEST_OWN_BVECS:
1261 zero_bvecs(&obj_req->bvec_pos, off, bytes);
1268 static void rbd_obj_request_destroy(struct kref *kref);
1269 static void rbd_obj_request_put(struct rbd_obj_request *obj_request)
1271 rbd_assert(obj_request != NULL);
1272 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1273 kref_read(&obj_request->kref));
1274 kref_put(&obj_request->kref, rbd_obj_request_destroy);
1277 static inline void rbd_img_obj_request_add(struct rbd_img_request *img_request,
1278 struct rbd_obj_request *obj_request)
1280 rbd_assert(obj_request->img_request == NULL);
1282 /* Image request now owns object's original reference */
1283 obj_request->img_request = img_request;
1284 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
1287 static inline void rbd_img_obj_request_del(struct rbd_img_request *img_request,
1288 struct rbd_obj_request *obj_request)
1290 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
1291 list_del(&obj_request->ex.oe_item);
1292 rbd_assert(obj_request->img_request == img_request);
1293 rbd_obj_request_put(obj_request);
1296 static void rbd_osd_submit(struct ceph_osd_request *osd_req)
1298 struct rbd_obj_request *obj_req = osd_req->r_priv;
1300 dout("%s osd_req %p for obj_req %p objno %llu %llu~%llu\n",
1301 __func__, osd_req, obj_req, obj_req->ex.oe_objno,
1302 obj_req->ex.oe_off, obj_req->ex.oe_len);
1303 ceph_osdc_start_request(osd_req->r_osdc, osd_req);
1307 * The default/initial value for all image request flags is 0. Each
1308 * is conditionally set to 1 at image request initialization time
1309 * and currently never change thereafter.
1311 static void img_request_layered_set(struct rbd_img_request *img_request)
1313 set_bit(IMG_REQ_LAYERED, &img_request->flags);
1316 static bool img_request_layered_test(struct rbd_img_request *img_request)
1318 return test_bit(IMG_REQ_LAYERED, &img_request->flags) != 0;
1321 static bool rbd_obj_is_entire(struct rbd_obj_request *obj_req)
1323 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1325 return !obj_req->ex.oe_off &&
1326 obj_req->ex.oe_len == rbd_dev->layout.object_size;
1329 static bool rbd_obj_is_tail(struct rbd_obj_request *obj_req)
1331 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1333 return obj_req->ex.oe_off + obj_req->ex.oe_len ==
1334 rbd_dev->layout.object_size;
1338 * Must be called after rbd_obj_calc_img_extents().
1340 static void rbd_obj_set_copyup_enabled(struct rbd_obj_request *obj_req)
1342 rbd_assert(obj_req->img_request->snapc);
1344 if (obj_req->img_request->op_type == OBJ_OP_DISCARD) {
1345 dout("%s %p objno %llu discard\n", __func__, obj_req,
1346 obj_req->ex.oe_objno);
1350 if (!obj_req->num_img_extents) {
1351 dout("%s %p objno %llu not overlapping\n", __func__, obj_req,
1352 obj_req->ex.oe_objno);
1356 if (rbd_obj_is_entire(obj_req) &&
1357 !obj_req->img_request->snapc->num_snaps) {
1358 dout("%s %p objno %llu entire\n", __func__, obj_req,
1359 obj_req->ex.oe_objno);
1363 obj_req->flags |= RBD_OBJ_FLAG_COPYUP_ENABLED;
1366 static u64 rbd_obj_img_extents_bytes(struct rbd_obj_request *obj_req)
1368 return ceph_file_extents_bytes(obj_req->img_extents,
1369 obj_req->num_img_extents);
1372 static bool rbd_img_is_write(struct rbd_img_request *img_req)
1374 switch (img_req->op_type) {
1378 case OBJ_OP_DISCARD:
1379 case OBJ_OP_ZEROOUT:
1386 static void rbd_osd_req_callback(struct ceph_osd_request *osd_req)
1388 struct rbd_obj_request *obj_req = osd_req->r_priv;
1391 dout("%s osd_req %p result %d for obj_req %p\n", __func__, osd_req,
1392 osd_req->r_result, obj_req);
1395 * Writes aren't allowed to return a data payload. In some
1396 * guarded write cases (e.g. stat + zero on an empty object)
1397 * a stat response makes it through, but we don't care.
1399 if (osd_req->r_result > 0 && rbd_img_is_write(obj_req->img_request))
1402 result = osd_req->r_result;
1404 rbd_obj_handle_request(obj_req, result);
1407 static void rbd_osd_format_read(struct ceph_osd_request *osd_req)
1409 struct rbd_obj_request *obj_request = osd_req->r_priv;
1410 struct rbd_device *rbd_dev = obj_request->img_request->rbd_dev;
1411 struct ceph_options *opt = rbd_dev->rbd_client->client->options;
1413 osd_req->r_flags = CEPH_OSD_FLAG_READ | opt->read_from_replica;
1414 osd_req->r_snapid = obj_request->img_request->snap_id;
1417 static void rbd_osd_format_write(struct ceph_osd_request *osd_req)
1419 struct rbd_obj_request *obj_request = osd_req->r_priv;
1421 osd_req->r_flags = CEPH_OSD_FLAG_WRITE;
1422 ktime_get_real_ts64(&osd_req->r_mtime);
1423 osd_req->r_data_offset = obj_request->ex.oe_off;
1426 static struct ceph_osd_request *
1427 __rbd_obj_add_osd_request(struct rbd_obj_request *obj_req,
1428 struct ceph_snap_context *snapc, int num_ops)
1430 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1431 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
1432 struct ceph_osd_request *req;
1433 const char *name_format = rbd_dev->image_format == 1 ?
1434 RBD_V1_DATA_FORMAT : RBD_V2_DATA_FORMAT;
1437 req = ceph_osdc_alloc_request(osdc, snapc, num_ops, false, GFP_NOIO);
1439 return ERR_PTR(-ENOMEM);
1441 list_add_tail(&req->r_private_item, &obj_req->osd_reqs);
1442 req->r_callback = rbd_osd_req_callback;
1443 req->r_priv = obj_req;
1446 * Data objects may be stored in a separate pool, but always in
1447 * the same namespace in that pool as the header in its pool.
1449 ceph_oloc_copy(&req->r_base_oloc, &rbd_dev->header_oloc);
1450 req->r_base_oloc.pool = rbd_dev->layout.pool_id;
1452 ret = ceph_oid_aprintf(&req->r_base_oid, GFP_NOIO, name_format,
1453 rbd_dev->header.object_prefix,
1454 obj_req->ex.oe_objno);
1456 return ERR_PTR(ret);
1461 static struct ceph_osd_request *
1462 rbd_obj_add_osd_request(struct rbd_obj_request *obj_req, int num_ops)
1464 rbd_assert(obj_req->img_request->snapc);
1465 return __rbd_obj_add_osd_request(obj_req, obj_req->img_request->snapc,
1469 static struct rbd_obj_request *rbd_obj_request_create(void)
1471 struct rbd_obj_request *obj_request;
1473 obj_request = kmem_cache_zalloc(rbd_obj_request_cache, GFP_NOIO);
1477 ceph_object_extent_init(&obj_request->ex);
1478 INIT_LIST_HEAD(&obj_request->osd_reqs);
1479 mutex_init(&obj_request->state_mutex);
1480 kref_init(&obj_request->kref);
1482 dout("%s %p\n", __func__, obj_request);
1486 static void rbd_obj_request_destroy(struct kref *kref)
1488 struct rbd_obj_request *obj_request;
1489 struct ceph_osd_request *osd_req;
1492 obj_request = container_of(kref, struct rbd_obj_request, kref);
1494 dout("%s: obj %p\n", __func__, obj_request);
1496 while (!list_empty(&obj_request->osd_reqs)) {
1497 osd_req = list_first_entry(&obj_request->osd_reqs,
1498 struct ceph_osd_request, r_private_item);
1499 list_del_init(&osd_req->r_private_item);
1500 ceph_osdc_put_request(osd_req);
1503 switch (obj_request->img_request->data_type) {
1504 case OBJ_REQUEST_NODATA:
1505 case OBJ_REQUEST_BIO:
1506 case OBJ_REQUEST_BVECS:
1507 break; /* Nothing to do */
1508 case OBJ_REQUEST_OWN_BVECS:
1509 kfree(obj_request->bvec_pos.bvecs);
1515 kfree(obj_request->img_extents);
1516 if (obj_request->copyup_bvecs) {
1517 for (i = 0; i < obj_request->copyup_bvec_count; i++) {
1518 if (obj_request->copyup_bvecs[i].bv_page)
1519 __free_page(obj_request->copyup_bvecs[i].bv_page);
1521 kfree(obj_request->copyup_bvecs);
1524 kmem_cache_free(rbd_obj_request_cache, obj_request);
1527 /* It's OK to call this for a device with no parent */
1529 static void rbd_spec_put(struct rbd_spec *spec);
1530 static void rbd_dev_unparent(struct rbd_device *rbd_dev)
1532 rbd_dev_remove_parent(rbd_dev);
1533 rbd_spec_put(rbd_dev->parent_spec);
1534 rbd_dev->parent_spec = NULL;
1535 rbd_dev->parent_overlap = 0;
1539 * Parent image reference counting is used to determine when an
1540 * image's parent fields can be safely torn down--after there are no
1541 * more in-flight requests to the parent image. When the last
1542 * reference is dropped, cleaning them up is safe.
1544 static void rbd_dev_parent_put(struct rbd_device *rbd_dev)
1548 if (!rbd_dev->parent_spec)
1551 counter = atomic_dec_return_safe(&rbd_dev->parent_ref);
1555 /* Last reference; clean up parent data structures */
1558 rbd_dev_unparent(rbd_dev);
1560 rbd_warn(rbd_dev, "parent reference underflow");
1564 * If an image has a non-zero parent overlap, get a reference to its
1567 * Returns true if the rbd device has a parent with a non-zero
1568 * overlap and a reference for it was successfully taken, or
1571 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
1575 if (!rbd_dev->parent_spec)
1578 if (rbd_dev->parent_overlap)
1579 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
1582 rbd_warn(rbd_dev, "parent reference overflow");
1587 static void rbd_img_request_init(struct rbd_img_request *img_request,
1588 struct rbd_device *rbd_dev,
1589 enum obj_operation_type op_type)
1591 memset(img_request, 0, sizeof(*img_request));
1593 img_request->rbd_dev = rbd_dev;
1594 img_request->op_type = op_type;
1596 INIT_LIST_HEAD(&img_request->lock_item);
1597 INIT_LIST_HEAD(&img_request->object_extents);
1598 mutex_init(&img_request->state_mutex);
1602 * Only snap_id is captured here, for reads. For writes, snapshot
1603 * context is captured in rbd_img_object_requests() after exclusive
1604 * lock is ensured to be held.
1606 static void rbd_img_capture_header(struct rbd_img_request *img_req)
1608 struct rbd_device *rbd_dev = img_req->rbd_dev;
1610 lockdep_assert_held(&rbd_dev->header_rwsem);
1612 if (!rbd_img_is_write(img_req))
1613 img_req->snap_id = rbd_dev->spec->snap_id;
1615 if (rbd_dev_parent_get(rbd_dev))
1616 img_request_layered_set(img_req);
1619 static void rbd_img_request_destroy(struct rbd_img_request *img_request)
1621 struct rbd_obj_request *obj_request;
1622 struct rbd_obj_request *next_obj_request;
1624 dout("%s: img %p\n", __func__, img_request);
1626 WARN_ON(!list_empty(&img_request->lock_item));
1627 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
1628 rbd_img_obj_request_del(img_request, obj_request);
1630 if (img_request_layered_test(img_request))
1631 rbd_dev_parent_put(img_request->rbd_dev);
1633 if (rbd_img_is_write(img_request))
1634 ceph_put_snap_context(img_request->snapc);
1636 if (test_bit(IMG_REQ_CHILD, &img_request->flags))
1637 kmem_cache_free(rbd_img_request_cache, img_request);
1640 #define BITS_PER_OBJ 2
1641 #define OBJS_PER_BYTE (BITS_PER_BYTE / BITS_PER_OBJ)
1642 #define OBJ_MASK ((1 << BITS_PER_OBJ) - 1)
1644 static void __rbd_object_map_index(struct rbd_device *rbd_dev, u64 objno,
1645 u64 *index, u8 *shift)
1649 rbd_assert(objno < rbd_dev->object_map_size);
1650 *index = div_u64_rem(objno, OBJS_PER_BYTE, &off);
1651 *shift = (OBJS_PER_BYTE - off - 1) * BITS_PER_OBJ;
1654 static u8 __rbd_object_map_get(struct rbd_device *rbd_dev, u64 objno)
1659 lockdep_assert_held(&rbd_dev->object_map_lock);
1660 __rbd_object_map_index(rbd_dev, objno, &index, &shift);
1661 return (rbd_dev->object_map[index] >> shift) & OBJ_MASK;
1664 static void __rbd_object_map_set(struct rbd_device *rbd_dev, u64 objno, u8 val)
1670 lockdep_assert_held(&rbd_dev->object_map_lock);
1671 rbd_assert(!(val & ~OBJ_MASK));
1673 __rbd_object_map_index(rbd_dev, objno, &index, &shift);
1674 p = &rbd_dev->object_map[index];
1675 *p = (*p & ~(OBJ_MASK << shift)) | (val << shift);
1678 static u8 rbd_object_map_get(struct rbd_device *rbd_dev, u64 objno)
1682 spin_lock(&rbd_dev->object_map_lock);
1683 state = __rbd_object_map_get(rbd_dev, objno);
1684 spin_unlock(&rbd_dev->object_map_lock);
1688 static bool use_object_map(struct rbd_device *rbd_dev)
1691 * An image mapped read-only can't use the object map -- it isn't
1692 * loaded because the header lock isn't acquired. Someone else can
1693 * write to the image and update the object map behind our back.
1695 * A snapshot can't be written to, so using the object map is always
1698 if (!rbd_is_snap(rbd_dev) && rbd_is_ro(rbd_dev))
1701 return ((rbd_dev->header.features & RBD_FEATURE_OBJECT_MAP) &&
1702 !(rbd_dev->object_map_flags & RBD_FLAG_OBJECT_MAP_INVALID));
1705 static bool rbd_object_map_may_exist(struct rbd_device *rbd_dev, u64 objno)
1709 /* fall back to default logic if object map is disabled or invalid */
1710 if (!use_object_map(rbd_dev))
1713 state = rbd_object_map_get(rbd_dev, objno);
1714 return state != OBJECT_NONEXISTENT;
1717 static void rbd_object_map_name(struct rbd_device *rbd_dev, u64 snap_id,
1718 struct ceph_object_id *oid)
1720 if (snap_id == CEPH_NOSNAP)
1721 ceph_oid_printf(oid, "%s%s", RBD_OBJECT_MAP_PREFIX,
1722 rbd_dev->spec->image_id);
1724 ceph_oid_printf(oid, "%s%s.%016llx", RBD_OBJECT_MAP_PREFIX,
1725 rbd_dev->spec->image_id, snap_id);
1728 static int rbd_object_map_lock(struct rbd_device *rbd_dev)
1730 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
1731 CEPH_DEFINE_OID_ONSTACK(oid);
1734 struct ceph_locker *lockers;
1736 bool broke_lock = false;
1739 rbd_object_map_name(rbd_dev, CEPH_NOSNAP, &oid);
1742 ret = ceph_cls_lock(osdc, &oid, &rbd_dev->header_oloc, RBD_LOCK_NAME,
1743 CEPH_CLS_LOCK_EXCLUSIVE, "", "", "", 0);
1744 if (ret != -EBUSY || broke_lock) {
1746 ret = 0; /* already locked by myself */
1748 rbd_warn(rbd_dev, "failed to lock object map: %d", ret);
1752 ret = ceph_cls_lock_info(osdc, &oid, &rbd_dev->header_oloc,
1753 RBD_LOCK_NAME, &lock_type, &lock_tag,
1754 &lockers, &num_lockers);
1759 rbd_warn(rbd_dev, "failed to get object map lockers: %d", ret);
1764 if (num_lockers == 0)
1767 rbd_warn(rbd_dev, "breaking object map lock owned by %s%llu",
1768 ENTITY_NAME(lockers[0].id.name));
1770 ret = ceph_cls_break_lock(osdc, &oid, &rbd_dev->header_oloc,
1771 RBD_LOCK_NAME, lockers[0].id.cookie,
1772 &lockers[0].id.name);
1773 ceph_free_lockers(lockers, num_lockers);
1778 rbd_warn(rbd_dev, "failed to break object map lock: %d", ret);
1786 static void rbd_object_map_unlock(struct rbd_device *rbd_dev)
1788 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
1789 CEPH_DEFINE_OID_ONSTACK(oid);
1792 rbd_object_map_name(rbd_dev, CEPH_NOSNAP, &oid);
1794 ret = ceph_cls_unlock(osdc, &oid, &rbd_dev->header_oloc, RBD_LOCK_NAME,
1796 if (ret && ret != -ENOENT)
1797 rbd_warn(rbd_dev, "failed to unlock object map: %d", ret);
1800 static int decode_object_map_header(void **p, void *end, u64 *object_map_size)
1808 ceph_decode_32_safe(p, end, header_len, e_inval);
1809 header_end = *p + header_len;
1811 ret = ceph_start_decoding(p, end, 1, "BitVector header", &struct_v,
1816 ceph_decode_64_safe(p, end, *object_map_size, e_inval);
1825 static int __rbd_object_map_load(struct rbd_device *rbd_dev)
1827 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
1828 CEPH_DEFINE_OID_ONSTACK(oid);
1829 struct page **pages;
1833 u64 object_map_bytes;
1834 u64 object_map_size;
1838 rbd_assert(!rbd_dev->object_map && !rbd_dev->object_map_size);
1840 num_objects = ceph_get_num_objects(&rbd_dev->layout,
1841 rbd_dev->mapping.size);
1842 object_map_bytes = DIV_ROUND_UP_ULL(num_objects * BITS_PER_OBJ,
1844 num_pages = calc_pages_for(0, object_map_bytes) + 1;
1845 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1847 return PTR_ERR(pages);
1849 reply_len = num_pages * PAGE_SIZE;
1850 rbd_object_map_name(rbd_dev, rbd_dev->spec->snap_id, &oid);
1851 ret = ceph_osdc_call(osdc, &oid, &rbd_dev->header_oloc,
1852 "rbd", "object_map_load", CEPH_OSD_FLAG_READ,
1853 NULL, 0, pages, &reply_len);
1857 p = page_address(pages[0]);
1858 end = p + min(reply_len, (size_t)PAGE_SIZE);
1859 ret = decode_object_map_header(&p, end, &object_map_size);
1863 if (object_map_size != num_objects) {
1864 rbd_warn(rbd_dev, "object map size mismatch: %llu vs %llu",
1865 object_map_size, num_objects);
1870 if (offset_in_page(p) + object_map_bytes > reply_len) {
1875 rbd_dev->object_map = kvmalloc(object_map_bytes, GFP_KERNEL);
1876 if (!rbd_dev->object_map) {
1881 rbd_dev->object_map_size = object_map_size;
1882 ceph_copy_from_page_vector(pages, rbd_dev->object_map,
1883 offset_in_page(p), object_map_bytes);
1886 ceph_release_page_vector(pages, num_pages);
1890 static void rbd_object_map_free(struct rbd_device *rbd_dev)
1892 kvfree(rbd_dev->object_map);
1893 rbd_dev->object_map = NULL;
1894 rbd_dev->object_map_size = 0;
1897 static int rbd_object_map_load(struct rbd_device *rbd_dev)
1901 ret = __rbd_object_map_load(rbd_dev);
1905 ret = rbd_dev_v2_get_flags(rbd_dev);
1907 rbd_object_map_free(rbd_dev);
1911 if (rbd_dev->object_map_flags & RBD_FLAG_OBJECT_MAP_INVALID)
1912 rbd_warn(rbd_dev, "object map is invalid");
1917 static int rbd_object_map_open(struct rbd_device *rbd_dev)
1921 ret = rbd_object_map_lock(rbd_dev);
1925 ret = rbd_object_map_load(rbd_dev);
1927 rbd_object_map_unlock(rbd_dev);
1934 static void rbd_object_map_close(struct rbd_device *rbd_dev)
1936 rbd_object_map_free(rbd_dev);
1937 rbd_object_map_unlock(rbd_dev);
1941 * This function needs snap_id (or more precisely just something to
1942 * distinguish between HEAD and snapshot object maps), new_state and
1943 * current_state that were passed to rbd_object_map_update().
1945 * To avoid allocating and stashing a context we piggyback on the OSD
1946 * request. A HEAD update has two ops (assert_locked). For new_state
1947 * and current_state we decode our own object_map_update op, encoded in
1948 * rbd_cls_object_map_update().
1950 static int rbd_object_map_update_finish(struct rbd_obj_request *obj_req,
1951 struct ceph_osd_request *osd_req)
1953 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
1954 struct ceph_osd_data *osd_data;
1956 u8 state, new_state, current_state;
1957 bool has_current_state;
1960 if (osd_req->r_result)
1961 return osd_req->r_result;
1964 * Nothing to do for a snapshot object map.
1966 if (osd_req->r_num_ops == 1)
1970 * Update in-memory HEAD object map.
1972 rbd_assert(osd_req->r_num_ops == 2);
1973 osd_data = osd_req_op_data(osd_req, 1, cls, request_data);
1974 rbd_assert(osd_data->type == CEPH_OSD_DATA_TYPE_PAGES);
1976 p = page_address(osd_data->pages[0]);
1977 objno = ceph_decode_64(&p);
1978 rbd_assert(objno == obj_req->ex.oe_objno);
1979 rbd_assert(ceph_decode_64(&p) == objno + 1);
1980 new_state = ceph_decode_8(&p);
1981 has_current_state = ceph_decode_8(&p);
1982 if (has_current_state)
1983 current_state = ceph_decode_8(&p);
1985 spin_lock(&rbd_dev->object_map_lock);
1986 state = __rbd_object_map_get(rbd_dev, objno);
1987 if (!has_current_state || current_state == state ||
1988 (current_state == OBJECT_EXISTS && state == OBJECT_EXISTS_CLEAN))
1989 __rbd_object_map_set(rbd_dev, objno, new_state);
1990 spin_unlock(&rbd_dev->object_map_lock);
1995 static void rbd_object_map_callback(struct ceph_osd_request *osd_req)
1997 struct rbd_obj_request *obj_req = osd_req->r_priv;
2000 dout("%s osd_req %p result %d for obj_req %p\n", __func__, osd_req,
2001 osd_req->r_result, obj_req);
2003 result = rbd_object_map_update_finish(obj_req, osd_req);
2004 rbd_obj_handle_request(obj_req, result);
2007 static bool update_needed(struct rbd_device *rbd_dev, u64 objno, u8 new_state)
2009 u8 state = rbd_object_map_get(rbd_dev, objno);
2011 if (state == new_state ||
2012 (new_state == OBJECT_PENDING && state == OBJECT_NONEXISTENT) ||
2013 (new_state == OBJECT_NONEXISTENT && state != OBJECT_PENDING))
2019 static int rbd_cls_object_map_update(struct ceph_osd_request *req,
2020 int which, u64 objno, u8 new_state,
2021 const u8 *current_state)
2023 struct page **pages;
2027 ret = osd_req_op_cls_init(req, which, "rbd", "object_map_update");
2031 pages = ceph_alloc_page_vector(1, GFP_NOIO);
2033 return PTR_ERR(pages);
2035 p = start = page_address(pages[0]);
2036 ceph_encode_64(&p, objno);
2037 ceph_encode_64(&p, objno + 1);
2038 ceph_encode_8(&p, new_state);
2039 if (current_state) {
2040 ceph_encode_8(&p, 1);
2041 ceph_encode_8(&p, *current_state);
2043 ceph_encode_8(&p, 0);
2046 osd_req_op_cls_request_data_pages(req, which, pages, p - start, 0,
2053 * 0 - object map update sent
2054 * 1 - object map update isn't needed
2057 static int rbd_object_map_update(struct rbd_obj_request *obj_req, u64 snap_id,
2058 u8 new_state, const u8 *current_state)
2060 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2061 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
2062 struct ceph_osd_request *req;
2067 if (snap_id == CEPH_NOSNAP) {
2068 if (!update_needed(rbd_dev, obj_req->ex.oe_objno, new_state))
2071 num_ops++; /* assert_locked */
2074 req = ceph_osdc_alloc_request(osdc, NULL, num_ops, false, GFP_NOIO);
2078 list_add_tail(&req->r_private_item, &obj_req->osd_reqs);
2079 req->r_callback = rbd_object_map_callback;
2080 req->r_priv = obj_req;
2082 rbd_object_map_name(rbd_dev, snap_id, &req->r_base_oid);
2083 ceph_oloc_copy(&req->r_base_oloc, &rbd_dev->header_oloc);
2084 req->r_flags = CEPH_OSD_FLAG_WRITE;
2085 ktime_get_real_ts64(&req->r_mtime);
2087 if (snap_id == CEPH_NOSNAP) {
2089 * Protect against possible race conditions during lock
2090 * ownership transitions.
2092 ret = ceph_cls_assert_locked(req, which++, RBD_LOCK_NAME,
2093 CEPH_CLS_LOCK_EXCLUSIVE, "", "");
2098 ret = rbd_cls_object_map_update(req, which, obj_req->ex.oe_objno,
2099 new_state, current_state);
2103 ret = ceph_osdc_alloc_messages(req, GFP_NOIO);
2107 ceph_osdc_start_request(osdc, req);
2111 static void prune_extents(struct ceph_file_extent *img_extents,
2112 u32 *num_img_extents, u64 overlap)
2114 u32 cnt = *num_img_extents;
2116 /* drop extents completely beyond the overlap */
2117 while (cnt && img_extents[cnt - 1].fe_off >= overlap)
2121 struct ceph_file_extent *ex = &img_extents[cnt - 1];
2123 /* trim final overlapping extent */
2124 if (ex->fe_off + ex->fe_len > overlap)
2125 ex->fe_len = overlap - ex->fe_off;
2128 *num_img_extents = cnt;
2132 * Determine the byte range(s) covered by either just the object extent
2133 * or the entire object in the parent image.
2135 static int rbd_obj_calc_img_extents(struct rbd_obj_request *obj_req,
2138 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2141 if (!rbd_dev->parent_overlap)
2144 ret = ceph_extent_to_file(&rbd_dev->layout, obj_req->ex.oe_objno,
2145 entire ? 0 : obj_req->ex.oe_off,
2146 entire ? rbd_dev->layout.object_size :
2148 &obj_req->img_extents,
2149 &obj_req->num_img_extents);
2153 prune_extents(obj_req->img_extents, &obj_req->num_img_extents,
2154 rbd_dev->parent_overlap);
2158 static void rbd_osd_setup_data(struct ceph_osd_request *osd_req, int which)
2160 struct rbd_obj_request *obj_req = osd_req->r_priv;
2162 switch (obj_req->img_request->data_type) {
2163 case OBJ_REQUEST_BIO:
2164 osd_req_op_extent_osd_data_bio(osd_req, which,
2166 obj_req->ex.oe_len);
2168 case OBJ_REQUEST_BVECS:
2169 case OBJ_REQUEST_OWN_BVECS:
2170 rbd_assert(obj_req->bvec_pos.iter.bi_size ==
2171 obj_req->ex.oe_len);
2172 rbd_assert(obj_req->bvec_idx == obj_req->bvec_count);
2173 osd_req_op_extent_osd_data_bvec_pos(osd_req, which,
2174 &obj_req->bvec_pos);
2181 static int rbd_osd_setup_stat(struct ceph_osd_request *osd_req, int which)
2183 struct page **pages;
2186 * The response data for a STAT call consists of:
2193 pages = ceph_alloc_page_vector(1, GFP_NOIO);
2195 return PTR_ERR(pages);
2197 osd_req_op_init(osd_req, which, CEPH_OSD_OP_STAT, 0);
2198 osd_req_op_raw_data_in_pages(osd_req, which, pages,
2199 8 + sizeof(struct ceph_timespec),
2204 static int rbd_osd_setup_copyup(struct ceph_osd_request *osd_req, int which,
2207 struct rbd_obj_request *obj_req = osd_req->r_priv;
2210 ret = osd_req_op_cls_init(osd_req, which, "rbd", "copyup");
2214 osd_req_op_cls_request_data_bvecs(osd_req, which, obj_req->copyup_bvecs,
2215 obj_req->copyup_bvec_count, bytes);
2219 static int rbd_obj_init_read(struct rbd_obj_request *obj_req)
2221 obj_req->read_state = RBD_OBJ_READ_START;
2225 static void __rbd_osd_setup_write_ops(struct ceph_osd_request *osd_req,
2228 struct rbd_obj_request *obj_req = osd_req->r_priv;
2229 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2232 if (!use_object_map(rbd_dev) ||
2233 !(obj_req->flags & RBD_OBJ_FLAG_MAY_EXIST)) {
2234 osd_req_op_alloc_hint_init(osd_req, which++,
2235 rbd_dev->layout.object_size,
2236 rbd_dev->layout.object_size,
2237 rbd_dev->opts->alloc_hint_flags);
2240 if (rbd_obj_is_entire(obj_req))
2241 opcode = CEPH_OSD_OP_WRITEFULL;
2243 opcode = CEPH_OSD_OP_WRITE;
2245 osd_req_op_extent_init(osd_req, which, opcode,
2246 obj_req->ex.oe_off, obj_req->ex.oe_len, 0, 0);
2247 rbd_osd_setup_data(osd_req, which);
2250 static int rbd_obj_init_write(struct rbd_obj_request *obj_req)
2254 /* reverse map the entire object onto the parent */
2255 ret = rbd_obj_calc_img_extents(obj_req, true);
2259 obj_req->write_state = RBD_OBJ_WRITE_START;
2263 static u16 truncate_or_zero_opcode(struct rbd_obj_request *obj_req)
2265 return rbd_obj_is_tail(obj_req) ? CEPH_OSD_OP_TRUNCATE :
2269 static void __rbd_osd_setup_discard_ops(struct ceph_osd_request *osd_req,
2272 struct rbd_obj_request *obj_req = osd_req->r_priv;
2274 if (rbd_obj_is_entire(obj_req) && !obj_req->num_img_extents) {
2275 rbd_assert(obj_req->flags & RBD_OBJ_FLAG_DELETION);
2276 osd_req_op_init(osd_req, which, CEPH_OSD_OP_DELETE, 0);
2278 osd_req_op_extent_init(osd_req, which,
2279 truncate_or_zero_opcode(obj_req),
2280 obj_req->ex.oe_off, obj_req->ex.oe_len,
2285 static int rbd_obj_init_discard(struct rbd_obj_request *obj_req)
2287 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2292 * Align the range to alloc_size boundary and punt on discards
2293 * that are too small to free up any space.
2295 * alloc_size == object_size && is_tail() is a special case for
2296 * filestore with filestore_punch_hole = false, needed to allow
2297 * truncate (in addition to delete).
2299 if (rbd_dev->opts->alloc_size != rbd_dev->layout.object_size ||
2300 !rbd_obj_is_tail(obj_req)) {
2301 off = round_up(obj_req->ex.oe_off, rbd_dev->opts->alloc_size);
2302 next_off = round_down(obj_req->ex.oe_off + obj_req->ex.oe_len,
2303 rbd_dev->opts->alloc_size);
2304 if (off >= next_off)
2307 dout("%s %p %llu~%llu -> %llu~%llu\n", __func__,
2308 obj_req, obj_req->ex.oe_off, obj_req->ex.oe_len,
2309 off, next_off - off);
2310 obj_req->ex.oe_off = off;
2311 obj_req->ex.oe_len = next_off - off;
2314 /* reverse map the entire object onto the parent */
2315 ret = rbd_obj_calc_img_extents(obj_req, true);
2319 obj_req->flags |= RBD_OBJ_FLAG_NOOP_FOR_NONEXISTENT;
2320 if (rbd_obj_is_entire(obj_req) && !obj_req->num_img_extents)
2321 obj_req->flags |= RBD_OBJ_FLAG_DELETION;
2323 obj_req->write_state = RBD_OBJ_WRITE_START;
2327 static void __rbd_osd_setup_zeroout_ops(struct ceph_osd_request *osd_req,
2330 struct rbd_obj_request *obj_req = osd_req->r_priv;
2333 if (rbd_obj_is_entire(obj_req)) {
2334 if (obj_req->num_img_extents) {
2335 if (!(obj_req->flags & RBD_OBJ_FLAG_COPYUP_ENABLED))
2336 osd_req_op_init(osd_req, which++,
2337 CEPH_OSD_OP_CREATE, 0);
2338 opcode = CEPH_OSD_OP_TRUNCATE;
2340 rbd_assert(obj_req->flags & RBD_OBJ_FLAG_DELETION);
2341 osd_req_op_init(osd_req, which++,
2342 CEPH_OSD_OP_DELETE, 0);
2346 opcode = truncate_or_zero_opcode(obj_req);
2350 osd_req_op_extent_init(osd_req, which, opcode,
2351 obj_req->ex.oe_off, obj_req->ex.oe_len,
2355 static int rbd_obj_init_zeroout(struct rbd_obj_request *obj_req)
2359 /* reverse map the entire object onto the parent */
2360 ret = rbd_obj_calc_img_extents(obj_req, true);
2364 if (!obj_req->num_img_extents) {
2365 obj_req->flags |= RBD_OBJ_FLAG_NOOP_FOR_NONEXISTENT;
2366 if (rbd_obj_is_entire(obj_req))
2367 obj_req->flags |= RBD_OBJ_FLAG_DELETION;
2370 obj_req->write_state = RBD_OBJ_WRITE_START;
2374 static int count_write_ops(struct rbd_obj_request *obj_req)
2376 struct rbd_img_request *img_req = obj_req->img_request;
2378 switch (img_req->op_type) {
2380 if (!use_object_map(img_req->rbd_dev) ||
2381 !(obj_req->flags & RBD_OBJ_FLAG_MAY_EXIST))
2382 return 2; /* setallochint + write/writefull */
2384 return 1; /* write/writefull */
2385 case OBJ_OP_DISCARD:
2386 return 1; /* delete/truncate/zero */
2387 case OBJ_OP_ZEROOUT:
2388 if (rbd_obj_is_entire(obj_req) && obj_req->num_img_extents &&
2389 !(obj_req->flags & RBD_OBJ_FLAG_COPYUP_ENABLED))
2390 return 2; /* create + truncate */
2392 return 1; /* delete/truncate/zero */
2398 static void rbd_osd_setup_write_ops(struct ceph_osd_request *osd_req,
2401 struct rbd_obj_request *obj_req = osd_req->r_priv;
2403 switch (obj_req->img_request->op_type) {
2405 __rbd_osd_setup_write_ops(osd_req, which);
2407 case OBJ_OP_DISCARD:
2408 __rbd_osd_setup_discard_ops(osd_req, which);
2410 case OBJ_OP_ZEROOUT:
2411 __rbd_osd_setup_zeroout_ops(osd_req, which);
2419 * Prune the list of object requests (adjust offset and/or length, drop
2420 * redundant requests). Prepare object request state machines and image
2421 * request state machine for execution.
2423 static int __rbd_img_fill_request(struct rbd_img_request *img_req)
2425 struct rbd_obj_request *obj_req, *next_obj_req;
2428 for_each_obj_request_safe(img_req, obj_req, next_obj_req) {
2429 switch (img_req->op_type) {
2431 ret = rbd_obj_init_read(obj_req);
2434 ret = rbd_obj_init_write(obj_req);
2436 case OBJ_OP_DISCARD:
2437 ret = rbd_obj_init_discard(obj_req);
2439 case OBJ_OP_ZEROOUT:
2440 ret = rbd_obj_init_zeroout(obj_req);
2448 rbd_img_obj_request_del(img_req, obj_req);
2453 img_req->state = RBD_IMG_START;
2457 union rbd_img_fill_iter {
2458 struct ceph_bio_iter bio_iter;
2459 struct ceph_bvec_iter bvec_iter;
2462 struct rbd_img_fill_ctx {
2463 enum obj_request_type pos_type;
2464 union rbd_img_fill_iter *pos;
2465 union rbd_img_fill_iter iter;
2466 ceph_object_extent_fn_t set_pos_fn;
2467 ceph_object_extent_fn_t count_fn;
2468 ceph_object_extent_fn_t copy_fn;
2471 static struct ceph_object_extent *alloc_object_extent(void *arg)
2473 struct rbd_img_request *img_req = arg;
2474 struct rbd_obj_request *obj_req;
2476 obj_req = rbd_obj_request_create();
2480 rbd_img_obj_request_add(img_req, obj_req);
2481 return &obj_req->ex;
2485 * While su != os && sc == 1 is technically not fancy (it's the same
2486 * layout as su == os && sc == 1), we can't use the nocopy path for it
2487 * because ->set_pos_fn() should be called only once per object.
2488 * ceph_file_to_extents() invokes action_fn once per stripe unit, so
2489 * treat su != os && sc == 1 as fancy.
2491 static bool rbd_layout_is_fancy(struct ceph_file_layout *l)
2493 return l->stripe_unit != l->object_size;
2496 static int rbd_img_fill_request_nocopy(struct rbd_img_request *img_req,
2497 struct ceph_file_extent *img_extents,
2498 u32 num_img_extents,
2499 struct rbd_img_fill_ctx *fctx)
2504 img_req->data_type = fctx->pos_type;
2507 * Create object requests and set each object request's starting
2508 * position in the provided bio (list) or bio_vec array.
2510 fctx->iter = *fctx->pos;
2511 for (i = 0; i < num_img_extents; i++) {
2512 ret = ceph_file_to_extents(&img_req->rbd_dev->layout,
2513 img_extents[i].fe_off,
2514 img_extents[i].fe_len,
2515 &img_req->object_extents,
2516 alloc_object_extent, img_req,
2517 fctx->set_pos_fn, &fctx->iter);
2522 return __rbd_img_fill_request(img_req);
2526 * Map a list of image extents to a list of object extents, create the
2527 * corresponding object requests (normally each to a different object,
2528 * but not always) and add them to @img_req. For each object request,
2529 * set up its data descriptor to point to the corresponding chunk(s) of
2530 * @fctx->pos data buffer.
2532 * Because ceph_file_to_extents() will merge adjacent object extents
2533 * together, each object request's data descriptor may point to multiple
2534 * different chunks of @fctx->pos data buffer.
2536 * @fctx->pos data buffer is assumed to be large enough.
2538 static int rbd_img_fill_request(struct rbd_img_request *img_req,
2539 struct ceph_file_extent *img_extents,
2540 u32 num_img_extents,
2541 struct rbd_img_fill_ctx *fctx)
2543 struct rbd_device *rbd_dev = img_req->rbd_dev;
2544 struct rbd_obj_request *obj_req;
2548 if (fctx->pos_type == OBJ_REQUEST_NODATA ||
2549 !rbd_layout_is_fancy(&rbd_dev->layout))
2550 return rbd_img_fill_request_nocopy(img_req, img_extents,
2551 num_img_extents, fctx);
2553 img_req->data_type = OBJ_REQUEST_OWN_BVECS;
2556 * Create object requests and determine ->bvec_count for each object
2557 * request. Note that ->bvec_count sum over all object requests may
2558 * be greater than the number of bio_vecs in the provided bio (list)
2559 * or bio_vec array because when mapped, those bio_vecs can straddle
2560 * stripe unit boundaries.
2562 fctx->iter = *fctx->pos;
2563 for (i = 0; i < num_img_extents; i++) {
2564 ret = ceph_file_to_extents(&rbd_dev->layout,
2565 img_extents[i].fe_off,
2566 img_extents[i].fe_len,
2567 &img_req->object_extents,
2568 alloc_object_extent, img_req,
2569 fctx->count_fn, &fctx->iter);
2574 for_each_obj_request(img_req, obj_req) {
2575 obj_req->bvec_pos.bvecs = kmalloc_array(obj_req->bvec_count,
2576 sizeof(*obj_req->bvec_pos.bvecs),
2578 if (!obj_req->bvec_pos.bvecs)
2583 * Fill in each object request's private bio_vec array, splitting and
2584 * rearranging the provided bio_vecs in stripe unit chunks as needed.
2586 fctx->iter = *fctx->pos;
2587 for (i = 0; i < num_img_extents; i++) {
2588 ret = ceph_iterate_extents(&rbd_dev->layout,
2589 img_extents[i].fe_off,
2590 img_extents[i].fe_len,
2591 &img_req->object_extents,
2592 fctx->copy_fn, &fctx->iter);
2597 return __rbd_img_fill_request(img_req);
2600 static int rbd_img_fill_nodata(struct rbd_img_request *img_req,
2603 struct ceph_file_extent ex = { off, len };
2604 union rbd_img_fill_iter dummy = {};
2605 struct rbd_img_fill_ctx fctx = {
2606 .pos_type = OBJ_REQUEST_NODATA,
2610 return rbd_img_fill_request(img_req, &ex, 1, &fctx);
2613 static void set_bio_pos(struct ceph_object_extent *ex, u32 bytes, void *arg)
2615 struct rbd_obj_request *obj_req =
2616 container_of(ex, struct rbd_obj_request, ex);
2617 struct ceph_bio_iter *it = arg;
2619 dout("%s objno %llu bytes %u\n", __func__, ex->oe_objno, bytes);
2620 obj_req->bio_pos = *it;
2621 ceph_bio_iter_advance(it, bytes);
2624 static void count_bio_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2626 struct rbd_obj_request *obj_req =
2627 container_of(ex, struct rbd_obj_request, ex);
2628 struct ceph_bio_iter *it = arg;
2630 dout("%s objno %llu bytes %u\n", __func__, ex->oe_objno, bytes);
2631 ceph_bio_iter_advance_step(it, bytes, ({
2632 obj_req->bvec_count++;
2637 static void copy_bio_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2639 struct rbd_obj_request *obj_req =
2640 container_of(ex, struct rbd_obj_request, ex);
2641 struct ceph_bio_iter *it = arg;
2643 dout("%s objno %llu bytes %u\n", __func__, ex->oe_objno, bytes);
2644 ceph_bio_iter_advance_step(it, bytes, ({
2645 obj_req->bvec_pos.bvecs[obj_req->bvec_idx++] = bv;
2646 obj_req->bvec_pos.iter.bi_size += bv.bv_len;
2650 static int __rbd_img_fill_from_bio(struct rbd_img_request *img_req,
2651 struct ceph_file_extent *img_extents,
2652 u32 num_img_extents,
2653 struct ceph_bio_iter *bio_pos)
2655 struct rbd_img_fill_ctx fctx = {
2656 .pos_type = OBJ_REQUEST_BIO,
2657 .pos = (union rbd_img_fill_iter *)bio_pos,
2658 .set_pos_fn = set_bio_pos,
2659 .count_fn = count_bio_bvecs,
2660 .copy_fn = copy_bio_bvecs,
2663 return rbd_img_fill_request(img_req, img_extents, num_img_extents,
2667 static int rbd_img_fill_from_bio(struct rbd_img_request *img_req,
2668 u64 off, u64 len, struct bio *bio)
2670 struct ceph_file_extent ex = { off, len };
2671 struct ceph_bio_iter it = { .bio = bio, .iter = bio->bi_iter };
2673 return __rbd_img_fill_from_bio(img_req, &ex, 1, &it);
2676 static void set_bvec_pos(struct ceph_object_extent *ex, u32 bytes, void *arg)
2678 struct rbd_obj_request *obj_req =
2679 container_of(ex, struct rbd_obj_request, ex);
2680 struct ceph_bvec_iter *it = arg;
2682 obj_req->bvec_pos = *it;
2683 ceph_bvec_iter_shorten(&obj_req->bvec_pos, bytes);
2684 ceph_bvec_iter_advance(it, bytes);
2687 static void count_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2689 struct rbd_obj_request *obj_req =
2690 container_of(ex, struct rbd_obj_request, ex);
2691 struct ceph_bvec_iter *it = arg;
2693 ceph_bvec_iter_advance_step(it, bytes, ({
2694 obj_req->bvec_count++;
2698 static void copy_bvecs(struct ceph_object_extent *ex, u32 bytes, void *arg)
2700 struct rbd_obj_request *obj_req =
2701 container_of(ex, struct rbd_obj_request, ex);
2702 struct ceph_bvec_iter *it = arg;
2704 ceph_bvec_iter_advance_step(it, bytes, ({
2705 obj_req->bvec_pos.bvecs[obj_req->bvec_idx++] = bv;
2706 obj_req->bvec_pos.iter.bi_size += bv.bv_len;
2710 static int __rbd_img_fill_from_bvecs(struct rbd_img_request *img_req,
2711 struct ceph_file_extent *img_extents,
2712 u32 num_img_extents,
2713 struct ceph_bvec_iter *bvec_pos)
2715 struct rbd_img_fill_ctx fctx = {
2716 .pos_type = OBJ_REQUEST_BVECS,
2717 .pos = (union rbd_img_fill_iter *)bvec_pos,
2718 .set_pos_fn = set_bvec_pos,
2719 .count_fn = count_bvecs,
2720 .copy_fn = copy_bvecs,
2723 return rbd_img_fill_request(img_req, img_extents, num_img_extents,
2727 static int rbd_img_fill_from_bvecs(struct rbd_img_request *img_req,
2728 struct ceph_file_extent *img_extents,
2729 u32 num_img_extents,
2730 struct bio_vec *bvecs)
2732 struct ceph_bvec_iter it = {
2734 .iter = { .bi_size = ceph_file_extents_bytes(img_extents,
2738 return __rbd_img_fill_from_bvecs(img_req, img_extents, num_img_extents,
2742 static void rbd_img_handle_request_work(struct work_struct *work)
2744 struct rbd_img_request *img_req =
2745 container_of(work, struct rbd_img_request, work);
2747 rbd_img_handle_request(img_req, img_req->work_result);
2750 static void rbd_img_schedule(struct rbd_img_request *img_req, int result)
2752 INIT_WORK(&img_req->work, rbd_img_handle_request_work);
2753 img_req->work_result = result;
2754 queue_work(rbd_wq, &img_req->work);
2757 static bool rbd_obj_may_exist(struct rbd_obj_request *obj_req)
2759 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2761 if (rbd_object_map_may_exist(rbd_dev, obj_req->ex.oe_objno)) {
2762 obj_req->flags |= RBD_OBJ_FLAG_MAY_EXIST;
2766 dout("%s %p objno %llu assuming dne\n", __func__, obj_req,
2767 obj_req->ex.oe_objno);
2771 static int rbd_obj_read_object(struct rbd_obj_request *obj_req)
2773 struct ceph_osd_request *osd_req;
2776 osd_req = __rbd_obj_add_osd_request(obj_req, NULL, 1);
2777 if (IS_ERR(osd_req))
2778 return PTR_ERR(osd_req);
2780 osd_req_op_extent_init(osd_req, 0, CEPH_OSD_OP_READ,
2781 obj_req->ex.oe_off, obj_req->ex.oe_len, 0, 0);
2782 rbd_osd_setup_data(osd_req, 0);
2783 rbd_osd_format_read(osd_req);
2785 ret = ceph_osdc_alloc_messages(osd_req, GFP_NOIO);
2789 rbd_osd_submit(osd_req);
2793 static int rbd_obj_read_from_parent(struct rbd_obj_request *obj_req)
2795 struct rbd_img_request *img_req = obj_req->img_request;
2796 struct rbd_device *parent = img_req->rbd_dev->parent;
2797 struct rbd_img_request *child_img_req;
2800 child_img_req = kmem_cache_alloc(rbd_img_request_cache, GFP_NOIO);
2804 rbd_img_request_init(child_img_req, parent, OBJ_OP_READ);
2805 __set_bit(IMG_REQ_CHILD, &child_img_req->flags);
2806 child_img_req->obj_request = obj_req;
2808 down_read(&parent->header_rwsem);
2809 rbd_img_capture_header(child_img_req);
2810 up_read(&parent->header_rwsem);
2812 dout("%s child_img_req %p for obj_req %p\n", __func__, child_img_req,
2815 if (!rbd_img_is_write(img_req)) {
2816 switch (img_req->data_type) {
2817 case OBJ_REQUEST_BIO:
2818 ret = __rbd_img_fill_from_bio(child_img_req,
2819 obj_req->img_extents,
2820 obj_req->num_img_extents,
2823 case OBJ_REQUEST_BVECS:
2824 case OBJ_REQUEST_OWN_BVECS:
2825 ret = __rbd_img_fill_from_bvecs(child_img_req,
2826 obj_req->img_extents,
2827 obj_req->num_img_extents,
2828 &obj_req->bvec_pos);
2834 ret = rbd_img_fill_from_bvecs(child_img_req,
2835 obj_req->img_extents,
2836 obj_req->num_img_extents,
2837 obj_req->copyup_bvecs);
2840 rbd_img_request_destroy(child_img_req);
2844 /* avoid parent chain recursion */
2845 rbd_img_schedule(child_img_req, 0);
2849 static bool rbd_obj_advance_read(struct rbd_obj_request *obj_req, int *result)
2851 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2855 switch (obj_req->read_state) {
2856 case RBD_OBJ_READ_START:
2857 rbd_assert(!*result);
2859 if (!rbd_obj_may_exist(obj_req)) {
2861 obj_req->read_state = RBD_OBJ_READ_OBJECT;
2865 ret = rbd_obj_read_object(obj_req);
2870 obj_req->read_state = RBD_OBJ_READ_OBJECT;
2872 case RBD_OBJ_READ_OBJECT:
2873 if (*result == -ENOENT && rbd_dev->parent_overlap) {
2874 /* reverse map this object extent onto the parent */
2875 ret = rbd_obj_calc_img_extents(obj_req, false);
2880 if (obj_req->num_img_extents) {
2881 ret = rbd_obj_read_from_parent(obj_req);
2886 obj_req->read_state = RBD_OBJ_READ_PARENT;
2892 * -ENOENT means a hole in the image -- zero-fill the entire
2893 * length of the request. A short read also implies zero-fill
2894 * to the end of the request.
2896 if (*result == -ENOENT) {
2897 rbd_obj_zero_range(obj_req, 0, obj_req->ex.oe_len);
2899 } else if (*result >= 0) {
2900 if (*result < obj_req->ex.oe_len)
2901 rbd_obj_zero_range(obj_req, *result,
2902 obj_req->ex.oe_len - *result);
2904 rbd_assert(*result == obj_req->ex.oe_len);
2908 case RBD_OBJ_READ_PARENT:
2910 * The parent image is read only up to the overlap -- zero-fill
2911 * from the overlap to the end of the request.
2914 u32 obj_overlap = rbd_obj_img_extents_bytes(obj_req);
2916 if (obj_overlap < obj_req->ex.oe_len)
2917 rbd_obj_zero_range(obj_req, obj_overlap,
2918 obj_req->ex.oe_len - obj_overlap);
2926 static bool rbd_obj_write_is_noop(struct rbd_obj_request *obj_req)
2928 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2930 if (rbd_object_map_may_exist(rbd_dev, obj_req->ex.oe_objno))
2931 obj_req->flags |= RBD_OBJ_FLAG_MAY_EXIST;
2933 if (!(obj_req->flags & RBD_OBJ_FLAG_MAY_EXIST) &&
2934 (obj_req->flags & RBD_OBJ_FLAG_NOOP_FOR_NONEXISTENT)) {
2935 dout("%s %p noop for nonexistent\n", __func__, obj_req);
2944 * 0 - object map update sent
2945 * 1 - object map update isn't needed
2948 static int rbd_obj_write_pre_object_map(struct rbd_obj_request *obj_req)
2950 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
2953 if (!(rbd_dev->header.features & RBD_FEATURE_OBJECT_MAP))
2956 if (obj_req->flags & RBD_OBJ_FLAG_DELETION)
2957 new_state = OBJECT_PENDING;
2959 new_state = OBJECT_EXISTS;
2961 return rbd_object_map_update(obj_req, CEPH_NOSNAP, new_state, NULL);
2964 static int rbd_obj_write_object(struct rbd_obj_request *obj_req)
2966 struct ceph_osd_request *osd_req;
2967 int num_ops = count_write_ops(obj_req);
2971 if (obj_req->flags & RBD_OBJ_FLAG_COPYUP_ENABLED)
2972 num_ops++; /* stat */
2974 osd_req = rbd_obj_add_osd_request(obj_req, num_ops);
2975 if (IS_ERR(osd_req))
2976 return PTR_ERR(osd_req);
2978 if (obj_req->flags & RBD_OBJ_FLAG_COPYUP_ENABLED) {
2979 ret = rbd_osd_setup_stat(osd_req, which++);
2984 rbd_osd_setup_write_ops(osd_req, which);
2985 rbd_osd_format_write(osd_req);
2987 ret = ceph_osdc_alloc_messages(osd_req, GFP_NOIO);
2991 rbd_osd_submit(osd_req);
2996 * copyup_bvecs pages are never highmem pages
2998 static bool is_zero_bvecs(struct bio_vec *bvecs, u32 bytes)
3000 struct ceph_bvec_iter it = {
3002 .iter = { .bi_size = bytes },
3005 ceph_bvec_iter_advance_step(&it, bytes, ({
3006 if (memchr_inv(bvec_virt(&bv), 0, bv.bv_len))
3012 #define MODS_ONLY U32_MAX
3014 static int rbd_obj_copyup_empty_snapc(struct rbd_obj_request *obj_req,
3017 struct ceph_osd_request *osd_req;
3020 dout("%s obj_req %p bytes %u\n", __func__, obj_req, bytes);
3021 rbd_assert(bytes > 0 && bytes != MODS_ONLY);
3023 osd_req = __rbd_obj_add_osd_request(obj_req, &rbd_empty_snapc, 1);
3024 if (IS_ERR(osd_req))
3025 return PTR_ERR(osd_req);
3027 ret = rbd_osd_setup_copyup(osd_req, 0, bytes);
3031 rbd_osd_format_write(osd_req);
3033 ret = ceph_osdc_alloc_messages(osd_req, GFP_NOIO);
3037 rbd_osd_submit(osd_req);
3041 static int rbd_obj_copyup_current_snapc(struct rbd_obj_request *obj_req,
3044 struct ceph_osd_request *osd_req;
3045 int num_ops = count_write_ops(obj_req);
3049 dout("%s obj_req %p bytes %u\n", __func__, obj_req, bytes);
3051 if (bytes != MODS_ONLY)
3052 num_ops++; /* copyup */
3054 osd_req = rbd_obj_add_osd_request(obj_req, num_ops);
3055 if (IS_ERR(osd_req))
3056 return PTR_ERR(osd_req);
3058 if (bytes != MODS_ONLY) {
3059 ret = rbd_osd_setup_copyup(osd_req, which++, bytes);
3064 rbd_osd_setup_write_ops(osd_req, which);
3065 rbd_osd_format_write(osd_req);
3067 ret = ceph_osdc_alloc_messages(osd_req, GFP_NOIO);
3071 rbd_osd_submit(osd_req);
3075 static int setup_copyup_bvecs(struct rbd_obj_request *obj_req, u64 obj_overlap)
3079 rbd_assert(!obj_req->copyup_bvecs);
3080 obj_req->copyup_bvec_count = calc_pages_for(0, obj_overlap);
3081 obj_req->copyup_bvecs = kcalloc(obj_req->copyup_bvec_count,
3082 sizeof(*obj_req->copyup_bvecs),
3084 if (!obj_req->copyup_bvecs)
3087 for (i = 0; i < obj_req->copyup_bvec_count; i++) {
3088 unsigned int len = min(obj_overlap, (u64)PAGE_SIZE);
3089 struct page *page = alloc_page(GFP_NOIO);
3094 bvec_set_page(&obj_req->copyup_bvecs[i], page, len, 0);
3098 rbd_assert(!obj_overlap);
3103 * The target object doesn't exist. Read the data for the entire
3104 * target object up to the overlap point (if any) from the parent,
3105 * so we can use it for a copyup.
3107 static int rbd_obj_copyup_read_parent(struct rbd_obj_request *obj_req)
3109 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
3112 rbd_assert(obj_req->num_img_extents);
3113 prune_extents(obj_req->img_extents, &obj_req->num_img_extents,
3114 rbd_dev->parent_overlap);
3115 if (!obj_req->num_img_extents) {
3117 * The overlap has become 0 (most likely because the
3118 * image has been flattened). Re-submit the original write
3119 * request -- pass MODS_ONLY since the copyup isn't needed
3122 return rbd_obj_copyup_current_snapc(obj_req, MODS_ONLY);
3125 ret = setup_copyup_bvecs(obj_req, rbd_obj_img_extents_bytes(obj_req));
3129 return rbd_obj_read_from_parent(obj_req);
3132 static void rbd_obj_copyup_object_maps(struct rbd_obj_request *obj_req)
3134 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
3135 struct ceph_snap_context *snapc = obj_req->img_request->snapc;
3140 rbd_assert(!obj_req->pending.result && !obj_req->pending.num_pending);
3142 if (!(rbd_dev->header.features & RBD_FEATURE_OBJECT_MAP))
3145 if (obj_req->flags & RBD_OBJ_FLAG_COPYUP_ZEROS)
3148 for (i = 0; i < snapc->num_snaps; i++) {
3149 if ((rbd_dev->header.features & RBD_FEATURE_FAST_DIFF) &&
3150 i + 1 < snapc->num_snaps)
3151 new_state = OBJECT_EXISTS_CLEAN;
3153 new_state = OBJECT_EXISTS;
3155 ret = rbd_object_map_update(obj_req, snapc->snaps[i],
3158 obj_req->pending.result = ret;
3163 obj_req->pending.num_pending++;
3167 static void rbd_obj_copyup_write_object(struct rbd_obj_request *obj_req)
3169 u32 bytes = rbd_obj_img_extents_bytes(obj_req);
3172 rbd_assert(!obj_req->pending.result && !obj_req->pending.num_pending);
3175 * Only send non-zero copyup data to save some I/O and network
3176 * bandwidth -- zero copyup data is equivalent to the object not
3179 if (obj_req->flags & RBD_OBJ_FLAG_COPYUP_ZEROS)
3182 if (obj_req->img_request->snapc->num_snaps && bytes > 0) {
3184 * Send a copyup request with an empty snapshot context to
3185 * deep-copyup the object through all existing snapshots.
3186 * A second request with the current snapshot context will be
3187 * sent for the actual modification.
3189 ret = rbd_obj_copyup_empty_snapc(obj_req, bytes);
3191 obj_req->pending.result = ret;
3195 obj_req->pending.num_pending++;
3199 ret = rbd_obj_copyup_current_snapc(obj_req, bytes);
3201 obj_req->pending.result = ret;
3205 obj_req->pending.num_pending++;
3208 static bool rbd_obj_advance_copyup(struct rbd_obj_request *obj_req, int *result)
3210 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
3214 switch (obj_req->copyup_state) {
3215 case RBD_OBJ_COPYUP_START:
3216 rbd_assert(!*result);
3218 ret = rbd_obj_copyup_read_parent(obj_req);
3223 if (obj_req->num_img_extents)
3224 obj_req->copyup_state = RBD_OBJ_COPYUP_READ_PARENT;
3226 obj_req->copyup_state = RBD_OBJ_COPYUP_WRITE_OBJECT;
3228 case RBD_OBJ_COPYUP_READ_PARENT:
3232 if (is_zero_bvecs(obj_req->copyup_bvecs,
3233 rbd_obj_img_extents_bytes(obj_req))) {
3234 dout("%s %p detected zeros\n", __func__, obj_req);
3235 obj_req->flags |= RBD_OBJ_FLAG_COPYUP_ZEROS;
3238 rbd_obj_copyup_object_maps(obj_req);
3239 if (!obj_req->pending.num_pending) {
3240 *result = obj_req->pending.result;
3241 obj_req->copyup_state = RBD_OBJ_COPYUP_OBJECT_MAPS;
3244 obj_req->copyup_state = __RBD_OBJ_COPYUP_OBJECT_MAPS;
3246 case __RBD_OBJ_COPYUP_OBJECT_MAPS:
3247 if (!pending_result_dec(&obj_req->pending, result))
3250 case RBD_OBJ_COPYUP_OBJECT_MAPS:
3252 rbd_warn(rbd_dev, "snap object map update failed: %d",
3257 rbd_obj_copyup_write_object(obj_req);
3258 if (!obj_req->pending.num_pending) {
3259 *result = obj_req->pending.result;
3260 obj_req->copyup_state = RBD_OBJ_COPYUP_WRITE_OBJECT;
3263 obj_req->copyup_state = __RBD_OBJ_COPYUP_WRITE_OBJECT;
3265 case __RBD_OBJ_COPYUP_WRITE_OBJECT:
3266 if (!pending_result_dec(&obj_req->pending, result))
3269 case RBD_OBJ_COPYUP_WRITE_OBJECT:
3278 * 0 - object map update sent
3279 * 1 - object map update isn't needed
3282 static int rbd_obj_write_post_object_map(struct rbd_obj_request *obj_req)
3284 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
3285 u8 current_state = OBJECT_PENDING;
3287 if (!(rbd_dev->header.features & RBD_FEATURE_OBJECT_MAP))
3290 if (!(obj_req->flags & RBD_OBJ_FLAG_DELETION))
3293 return rbd_object_map_update(obj_req, CEPH_NOSNAP, OBJECT_NONEXISTENT,
3297 static bool rbd_obj_advance_write(struct rbd_obj_request *obj_req, int *result)
3299 struct rbd_device *rbd_dev = obj_req->img_request->rbd_dev;
3303 switch (obj_req->write_state) {
3304 case RBD_OBJ_WRITE_START:
3305 rbd_assert(!*result);
3307 rbd_obj_set_copyup_enabled(obj_req);
3308 if (rbd_obj_write_is_noop(obj_req))
3311 ret = rbd_obj_write_pre_object_map(obj_req);
3316 obj_req->write_state = RBD_OBJ_WRITE_PRE_OBJECT_MAP;
3320 case RBD_OBJ_WRITE_PRE_OBJECT_MAP:
3322 rbd_warn(rbd_dev, "pre object map update failed: %d",
3326 ret = rbd_obj_write_object(obj_req);
3331 obj_req->write_state = RBD_OBJ_WRITE_OBJECT;
3333 case RBD_OBJ_WRITE_OBJECT:
3334 if (*result == -ENOENT) {
3335 if (obj_req->flags & RBD_OBJ_FLAG_COPYUP_ENABLED) {
3337 obj_req->copyup_state = RBD_OBJ_COPYUP_START;
3338 obj_req->write_state = __RBD_OBJ_WRITE_COPYUP;
3342 * On a non-existent object:
3343 * delete - -ENOENT, truncate/zero - 0
3345 if (obj_req->flags & RBD_OBJ_FLAG_DELETION)
3351 obj_req->write_state = RBD_OBJ_WRITE_COPYUP;
3353 case __RBD_OBJ_WRITE_COPYUP:
3354 if (!rbd_obj_advance_copyup(obj_req, result))
3357 case RBD_OBJ_WRITE_COPYUP:
3359 rbd_warn(rbd_dev, "copyup failed: %d", *result);
3362 ret = rbd_obj_write_post_object_map(obj_req);
3367 obj_req->write_state = RBD_OBJ_WRITE_POST_OBJECT_MAP;
3371 case RBD_OBJ_WRITE_POST_OBJECT_MAP:
3373 rbd_warn(rbd_dev, "post object map update failed: %d",
3382 * Return true if @obj_req is completed.
3384 static bool __rbd_obj_handle_request(struct rbd_obj_request *obj_req,
3387 struct rbd_img_request *img_req = obj_req->img_request;
3388 struct rbd_device *rbd_dev = img_req->rbd_dev;
3391 mutex_lock(&obj_req->state_mutex);
3392 if (!rbd_img_is_write(img_req))
3393 done = rbd_obj_advance_read(obj_req, result);
3395 done = rbd_obj_advance_write(obj_req, result);
3396 mutex_unlock(&obj_req->state_mutex);
3398 if (done && *result) {
3399 rbd_assert(*result < 0);
3400 rbd_warn(rbd_dev, "%s at objno %llu %llu~%llu result %d",
3401 obj_op_name(img_req->op_type), obj_req->ex.oe_objno,
3402 obj_req->ex.oe_off, obj_req->ex.oe_len, *result);
3408 * This is open-coded in rbd_img_handle_request() to avoid parent chain
3411 static void rbd_obj_handle_request(struct rbd_obj_request *obj_req, int result)
3413 if (__rbd_obj_handle_request(obj_req, &result))
3414 rbd_img_handle_request(obj_req->img_request, result);
3417 static bool need_exclusive_lock(struct rbd_img_request *img_req)
3419 struct rbd_device *rbd_dev = img_req->rbd_dev;
3421 if (!(rbd_dev->header.features & RBD_FEATURE_EXCLUSIVE_LOCK))
3424 if (rbd_is_ro(rbd_dev))
3427 rbd_assert(!test_bit(IMG_REQ_CHILD, &img_req->flags));
3428 if (rbd_dev->opts->lock_on_read ||
3429 (rbd_dev->header.features & RBD_FEATURE_OBJECT_MAP))
3432 return rbd_img_is_write(img_req);
3435 static bool rbd_lock_add_request(struct rbd_img_request *img_req)
3437 struct rbd_device *rbd_dev = img_req->rbd_dev;
3440 lockdep_assert_held(&rbd_dev->lock_rwsem);
3441 locked = rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED;
3442 spin_lock(&rbd_dev->lock_lists_lock);
3443 rbd_assert(list_empty(&img_req->lock_item));
3445 list_add_tail(&img_req->lock_item, &rbd_dev->acquiring_list);
3447 list_add_tail(&img_req->lock_item, &rbd_dev->running_list);
3448 spin_unlock(&rbd_dev->lock_lists_lock);
3452 static void rbd_lock_del_request(struct rbd_img_request *img_req)
3454 struct rbd_device *rbd_dev = img_req->rbd_dev;
3455 bool need_wakeup = false;
3457 lockdep_assert_held(&rbd_dev->lock_rwsem);
3458 spin_lock(&rbd_dev->lock_lists_lock);
3459 if (!list_empty(&img_req->lock_item)) {
3460 list_del_init(&img_req->lock_item);
3461 need_wakeup = (rbd_dev->lock_state == RBD_LOCK_STATE_RELEASING &&
3462 list_empty(&rbd_dev->running_list));
3464 spin_unlock(&rbd_dev->lock_lists_lock);
3466 complete(&rbd_dev->releasing_wait);
3469 static int rbd_img_exclusive_lock(struct rbd_img_request *img_req)
3471 struct rbd_device *rbd_dev = img_req->rbd_dev;
3473 if (!need_exclusive_lock(img_req))
3476 if (rbd_lock_add_request(img_req))
3479 if (rbd_dev->opts->exclusive) {
3480 WARN_ON(1); /* lock got released? */
3485 * Note the use of mod_delayed_work() in rbd_acquire_lock()
3486 * and cancel_delayed_work() in wake_lock_waiters().
3488 dout("%s rbd_dev %p queueing lock_dwork\n", __func__, rbd_dev);
3489 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork, 0);
3493 static void rbd_img_object_requests(struct rbd_img_request *img_req)
3495 struct rbd_device *rbd_dev = img_req->rbd_dev;
3496 struct rbd_obj_request *obj_req;
3498 rbd_assert(!img_req->pending.result && !img_req->pending.num_pending);
3499 rbd_assert(!need_exclusive_lock(img_req) ||
3500 __rbd_is_lock_owner(rbd_dev));
3502 if (rbd_img_is_write(img_req)) {
3503 rbd_assert(!img_req->snapc);
3504 down_read(&rbd_dev->header_rwsem);
3505 img_req->snapc = ceph_get_snap_context(rbd_dev->header.snapc);
3506 up_read(&rbd_dev->header_rwsem);
3509 for_each_obj_request(img_req, obj_req) {
3512 if (__rbd_obj_handle_request(obj_req, &result)) {
3514 img_req->pending.result = result;
3518 img_req->pending.num_pending++;
3523 static bool rbd_img_advance(struct rbd_img_request *img_req, int *result)
3528 switch (img_req->state) {
3530 rbd_assert(!*result);
3532 ret = rbd_img_exclusive_lock(img_req);
3537 img_req->state = RBD_IMG_EXCLUSIVE_LOCK;
3541 case RBD_IMG_EXCLUSIVE_LOCK:
3545 rbd_img_object_requests(img_req);
3546 if (!img_req->pending.num_pending) {
3547 *result = img_req->pending.result;
3548 img_req->state = RBD_IMG_OBJECT_REQUESTS;
3551 img_req->state = __RBD_IMG_OBJECT_REQUESTS;
3553 case __RBD_IMG_OBJECT_REQUESTS:
3554 if (!pending_result_dec(&img_req->pending, result))
3557 case RBD_IMG_OBJECT_REQUESTS:
3565 * Return true if @img_req is completed.
3567 static bool __rbd_img_handle_request(struct rbd_img_request *img_req,
3570 struct rbd_device *rbd_dev = img_req->rbd_dev;
3573 if (need_exclusive_lock(img_req)) {
3574 down_read(&rbd_dev->lock_rwsem);
3575 mutex_lock(&img_req->state_mutex);
3576 done = rbd_img_advance(img_req, result);
3578 rbd_lock_del_request(img_req);
3579 mutex_unlock(&img_req->state_mutex);
3580 up_read(&rbd_dev->lock_rwsem);
3582 mutex_lock(&img_req->state_mutex);
3583 done = rbd_img_advance(img_req, result);
3584 mutex_unlock(&img_req->state_mutex);
3587 if (done && *result) {
3588 rbd_assert(*result < 0);
3589 rbd_warn(rbd_dev, "%s%s result %d",
3590 test_bit(IMG_REQ_CHILD, &img_req->flags) ? "child " : "",
3591 obj_op_name(img_req->op_type), *result);
3596 static void rbd_img_handle_request(struct rbd_img_request *img_req, int result)
3599 if (!__rbd_img_handle_request(img_req, &result))
3602 if (test_bit(IMG_REQ_CHILD, &img_req->flags)) {
3603 struct rbd_obj_request *obj_req = img_req->obj_request;
3605 rbd_img_request_destroy(img_req);
3606 if (__rbd_obj_handle_request(obj_req, &result)) {
3607 img_req = obj_req->img_request;
3611 struct request *rq = blk_mq_rq_from_pdu(img_req);
3613 rbd_img_request_destroy(img_req);
3614 blk_mq_end_request(rq, errno_to_blk_status(result));
3618 static const struct rbd_client_id rbd_empty_cid;
3620 static bool rbd_cid_equal(const struct rbd_client_id *lhs,
3621 const struct rbd_client_id *rhs)
3623 return lhs->gid == rhs->gid && lhs->handle == rhs->handle;
3626 static struct rbd_client_id rbd_get_cid(struct rbd_device *rbd_dev)
3628 struct rbd_client_id cid;
3630 mutex_lock(&rbd_dev->watch_mutex);
3631 cid.gid = ceph_client_gid(rbd_dev->rbd_client->client);
3632 cid.handle = rbd_dev->watch_cookie;
3633 mutex_unlock(&rbd_dev->watch_mutex);
3638 * lock_rwsem must be held for write
3640 static void rbd_set_owner_cid(struct rbd_device *rbd_dev,
3641 const struct rbd_client_id *cid)
3643 dout("%s rbd_dev %p %llu-%llu -> %llu-%llu\n", __func__, rbd_dev,
3644 rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle,
3645 cid->gid, cid->handle);
3646 rbd_dev->owner_cid = *cid; /* struct */
3649 static void format_lock_cookie(struct rbd_device *rbd_dev, char *buf)
3651 mutex_lock(&rbd_dev->watch_mutex);
3652 sprintf(buf, "%s %llu", RBD_LOCK_COOKIE_PREFIX, rbd_dev->watch_cookie);
3653 mutex_unlock(&rbd_dev->watch_mutex);
3656 static void __rbd_lock(struct rbd_device *rbd_dev, const char *cookie)
3658 struct rbd_client_id cid = rbd_get_cid(rbd_dev);
3660 rbd_dev->lock_state = RBD_LOCK_STATE_LOCKED;
3661 strcpy(rbd_dev->lock_cookie, cookie);
3662 rbd_set_owner_cid(rbd_dev, &cid);
3663 queue_work(rbd_dev->task_wq, &rbd_dev->acquired_lock_work);
3667 * lock_rwsem must be held for write
3669 static int rbd_lock(struct rbd_device *rbd_dev)
3671 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3675 WARN_ON(__rbd_is_lock_owner(rbd_dev) ||
3676 rbd_dev->lock_cookie[0] != '\0');
3678 format_lock_cookie(rbd_dev, cookie);
3679 ret = ceph_cls_lock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
3680 RBD_LOCK_NAME, CEPH_CLS_LOCK_EXCLUSIVE, cookie,
3681 RBD_LOCK_TAG, "", 0);
3682 if (ret && ret != -EEXIST)
3685 __rbd_lock(rbd_dev, cookie);
3690 * lock_rwsem must be held for write
3692 static void rbd_unlock(struct rbd_device *rbd_dev)
3694 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3697 WARN_ON(!__rbd_is_lock_owner(rbd_dev) ||
3698 rbd_dev->lock_cookie[0] == '\0');
3700 ret = ceph_cls_unlock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
3701 RBD_LOCK_NAME, rbd_dev->lock_cookie);
3702 if (ret && ret != -ENOENT)
3703 rbd_warn(rbd_dev, "failed to unlock header: %d", ret);
3705 /* treat errors as the image is unlocked */
3706 rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
3707 rbd_dev->lock_cookie[0] = '\0';
3708 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3709 queue_work(rbd_dev->task_wq, &rbd_dev->released_lock_work);
3712 static int __rbd_notify_op_lock(struct rbd_device *rbd_dev,
3713 enum rbd_notify_op notify_op,
3714 struct page ***preply_pages,
3717 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3718 struct rbd_client_id cid = rbd_get_cid(rbd_dev);
3719 char buf[4 + 8 + 8 + CEPH_ENCODING_START_BLK_LEN];
3720 int buf_size = sizeof(buf);
3723 dout("%s rbd_dev %p notify_op %d\n", __func__, rbd_dev, notify_op);
3725 /* encode *LockPayload NotifyMessage (op + ClientId) */
3726 ceph_start_encoding(&p, 2, 1, buf_size - CEPH_ENCODING_START_BLK_LEN);
3727 ceph_encode_32(&p, notify_op);
3728 ceph_encode_64(&p, cid.gid);
3729 ceph_encode_64(&p, cid.handle);
3731 return ceph_osdc_notify(osdc, &rbd_dev->header_oid,
3732 &rbd_dev->header_oloc, buf, buf_size,
3733 RBD_NOTIFY_TIMEOUT, preply_pages, preply_len);
3736 static void rbd_notify_op_lock(struct rbd_device *rbd_dev,
3737 enum rbd_notify_op notify_op)
3739 __rbd_notify_op_lock(rbd_dev, notify_op, NULL, NULL);
3742 static void rbd_notify_acquired_lock(struct work_struct *work)
3744 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3745 acquired_lock_work);
3747 rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_ACQUIRED_LOCK);
3750 static void rbd_notify_released_lock(struct work_struct *work)
3752 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3753 released_lock_work);
3755 rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_RELEASED_LOCK);
3758 static int rbd_request_lock(struct rbd_device *rbd_dev)
3760 struct page **reply_pages;
3762 bool lock_owner_responded = false;
3765 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3767 ret = __rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_REQUEST_LOCK,
3768 &reply_pages, &reply_len);
3769 if (ret && ret != -ETIMEDOUT) {
3770 rbd_warn(rbd_dev, "failed to request lock: %d", ret);
3774 if (reply_len > 0 && reply_len <= PAGE_SIZE) {
3775 void *p = page_address(reply_pages[0]);
3776 void *const end = p + reply_len;
3779 ceph_decode_32_safe(&p, end, n, e_inval); /* num_acks */
3784 ceph_decode_need(&p, end, 8 + 8, e_inval);
3785 p += 8 + 8; /* skip gid and cookie */
3787 ceph_decode_32_safe(&p, end, len, e_inval);
3791 if (lock_owner_responded) {
3793 "duplicate lock owners detected");
3798 lock_owner_responded = true;
3799 ret = ceph_start_decoding(&p, end, 1, "ResponseMessage",
3803 "failed to decode ResponseMessage: %d",
3808 ret = ceph_decode_32(&p);
3812 if (!lock_owner_responded) {
3813 rbd_warn(rbd_dev, "no lock owners detected");
3818 ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len));
3827 * Either image request state machine(s) or rbd_add_acquire_lock()
3830 static void wake_lock_waiters(struct rbd_device *rbd_dev, int result)
3832 struct rbd_img_request *img_req;
3834 dout("%s rbd_dev %p result %d\n", __func__, rbd_dev, result);
3835 lockdep_assert_held_write(&rbd_dev->lock_rwsem);
3837 cancel_delayed_work(&rbd_dev->lock_dwork);
3838 if (!completion_done(&rbd_dev->acquire_wait)) {
3839 rbd_assert(list_empty(&rbd_dev->acquiring_list) &&
3840 list_empty(&rbd_dev->running_list));
3841 rbd_dev->acquire_err = result;
3842 complete_all(&rbd_dev->acquire_wait);
3846 while (!list_empty(&rbd_dev->acquiring_list)) {
3847 img_req = list_first_entry(&rbd_dev->acquiring_list,
3848 struct rbd_img_request, lock_item);
3849 mutex_lock(&img_req->state_mutex);
3850 rbd_assert(img_req->state == RBD_IMG_EXCLUSIVE_LOCK);
3852 list_move_tail(&img_req->lock_item,
3853 &rbd_dev->running_list);
3855 list_del_init(&img_req->lock_item);
3856 rbd_img_schedule(img_req, result);
3857 mutex_unlock(&img_req->state_mutex);
3861 static bool locker_equal(const struct ceph_locker *lhs,
3862 const struct ceph_locker *rhs)
3864 return lhs->id.name.type == rhs->id.name.type &&
3865 lhs->id.name.num == rhs->id.name.num &&
3866 !strcmp(lhs->id.cookie, rhs->id.cookie) &&
3867 ceph_addr_equal_no_type(&lhs->info.addr, &rhs->info.addr);
3870 static void free_locker(struct ceph_locker *locker)
3873 ceph_free_lockers(locker, 1);
3876 static struct ceph_locker *get_lock_owner_info(struct rbd_device *rbd_dev)
3878 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3879 struct ceph_locker *lockers;
3886 ret = ceph_cls_lock_info(osdc, &rbd_dev->header_oid,
3887 &rbd_dev->header_oloc, RBD_LOCK_NAME,
3888 &lock_type, &lock_tag, &lockers, &num_lockers);
3890 rbd_warn(rbd_dev, "failed to get header lockers: %d", ret);
3891 return ERR_PTR(ret);
3894 if (num_lockers == 0) {
3895 dout("%s rbd_dev %p no lockers detected\n", __func__, rbd_dev);
3900 if (strcmp(lock_tag, RBD_LOCK_TAG)) {
3901 rbd_warn(rbd_dev, "locked by external mechanism, tag %s",
3906 if (lock_type != CEPH_CLS_LOCK_EXCLUSIVE) {
3907 rbd_warn(rbd_dev, "incompatible lock type detected");
3911 WARN_ON(num_lockers != 1);
3912 ret = sscanf(lockers[0].id.cookie, RBD_LOCK_COOKIE_PREFIX " %llu",
3915 rbd_warn(rbd_dev, "locked by external mechanism, cookie %s",
3916 lockers[0].id.cookie);
3919 if (ceph_addr_is_blank(&lockers[0].info.addr)) {
3920 rbd_warn(rbd_dev, "locker has a blank address");
3924 dout("%s rbd_dev %p got locker %s%llu@%pISpc/%u handle %llu\n",
3925 __func__, rbd_dev, ENTITY_NAME(lockers[0].id.name),
3926 &lockers[0].info.addr.in_addr,
3927 le32_to_cpu(lockers[0].info.addr.nonce), handle);
3935 ceph_free_lockers(lockers, num_lockers);
3936 return ERR_PTR(-EBUSY);
3939 static int find_watcher(struct rbd_device *rbd_dev,
3940 const struct ceph_locker *locker)
3942 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3943 struct ceph_watch_item *watchers;
3949 ret = ceph_osdc_list_watchers(osdc, &rbd_dev->header_oid,
3950 &rbd_dev->header_oloc, &watchers,
3953 rbd_warn(rbd_dev, "failed to get watchers: %d", ret);
3957 sscanf(locker->id.cookie, RBD_LOCK_COOKIE_PREFIX " %llu", &cookie);
3958 for (i = 0; i < num_watchers; i++) {
3960 * Ignore addr->type while comparing. This mimics
3961 * entity_addr_t::get_legacy_str() + strcmp().
3963 if (ceph_addr_equal_no_type(&watchers[i].addr,
3964 &locker->info.addr) &&
3965 watchers[i].cookie == cookie) {
3966 struct rbd_client_id cid = {
3967 .gid = le64_to_cpu(watchers[i].name.num),
3971 dout("%s rbd_dev %p found cid %llu-%llu\n", __func__,
3972 rbd_dev, cid.gid, cid.handle);
3973 rbd_set_owner_cid(rbd_dev, &cid);
3979 dout("%s rbd_dev %p no watchers\n", __func__, rbd_dev);
3987 * lock_rwsem must be held for write
3989 static int rbd_try_lock(struct rbd_device *rbd_dev)
3991 struct ceph_client *client = rbd_dev->rbd_client->client;
3992 struct ceph_locker *locker, *refreshed_locker;
3996 locker = refreshed_locker = NULL;
3998 ret = rbd_lock(rbd_dev);
4001 if (ret != -EBUSY) {
4002 rbd_warn(rbd_dev, "failed to lock header: %d", ret);
4006 /* determine if the current lock holder is still alive */
4007 locker = get_lock_owner_info(rbd_dev);
4008 if (IS_ERR(locker)) {
4009 ret = PTR_ERR(locker);
4016 ret = find_watcher(rbd_dev, locker);
4018 goto out; /* request lock or error */
4020 refreshed_locker = get_lock_owner_info(rbd_dev);
4021 if (IS_ERR(refreshed_locker)) {
4022 ret = PTR_ERR(refreshed_locker);
4023 refreshed_locker = NULL;
4026 if (!refreshed_locker ||
4027 !locker_equal(locker, refreshed_locker))
4030 rbd_warn(rbd_dev, "breaking header lock owned by %s%llu",
4031 ENTITY_NAME(locker->id.name));
4033 ret = ceph_monc_blocklist_add(&client->monc,
4034 &locker->info.addr);
4036 rbd_warn(rbd_dev, "failed to blocklist %s%llu: %d",
4037 ENTITY_NAME(locker->id.name), ret);
4041 ret = ceph_cls_break_lock(&client->osdc, &rbd_dev->header_oid,
4042 &rbd_dev->header_oloc, RBD_LOCK_NAME,
4043 locker->id.cookie, &locker->id.name);
4044 if (ret && ret != -ENOENT) {
4045 rbd_warn(rbd_dev, "failed to break header lock: %d",
4051 free_locker(refreshed_locker);
4052 free_locker(locker);
4056 free_locker(refreshed_locker);
4057 free_locker(locker);
4061 static int rbd_post_acquire_action(struct rbd_device *rbd_dev)
4065 ret = rbd_dev_refresh(rbd_dev);
4069 if (rbd_dev->header.features & RBD_FEATURE_OBJECT_MAP) {
4070 ret = rbd_object_map_open(rbd_dev);
4081 * 1 - caller should call rbd_request_lock()
4084 static int rbd_try_acquire_lock(struct rbd_device *rbd_dev)
4088 down_read(&rbd_dev->lock_rwsem);
4089 dout("%s rbd_dev %p read lock_state %d\n", __func__, rbd_dev,
4090 rbd_dev->lock_state);
4091 if (__rbd_is_lock_owner(rbd_dev)) {
4092 up_read(&rbd_dev->lock_rwsem);
4096 up_read(&rbd_dev->lock_rwsem);
4097 down_write(&rbd_dev->lock_rwsem);
4098 dout("%s rbd_dev %p write lock_state %d\n", __func__, rbd_dev,
4099 rbd_dev->lock_state);
4100 if (__rbd_is_lock_owner(rbd_dev)) {
4101 up_write(&rbd_dev->lock_rwsem);
4105 ret = rbd_try_lock(rbd_dev);
4107 rbd_warn(rbd_dev, "failed to acquire lock: %d", ret);
4111 up_write(&rbd_dev->lock_rwsem);
4115 rbd_assert(rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED);
4116 rbd_assert(list_empty(&rbd_dev->running_list));
4118 ret = rbd_post_acquire_action(rbd_dev);
4120 rbd_warn(rbd_dev, "post-acquire action failed: %d", ret);
4122 * Can't stay in RBD_LOCK_STATE_LOCKED because
4123 * rbd_lock_add_request() would let the request through,
4124 * assuming that e.g. object map is locked and loaded.
4126 rbd_unlock(rbd_dev);
4130 wake_lock_waiters(rbd_dev, ret);
4131 up_write(&rbd_dev->lock_rwsem);
4135 static void rbd_acquire_lock(struct work_struct *work)
4137 struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
4138 struct rbd_device, lock_dwork);
4141 dout("%s rbd_dev %p\n", __func__, rbd_dev);
4143 ret = rbd_try_acquire_lock(rbd_dev);
4145 dout("%s rbd_dev %p ret %d - done\n", __func__, rbd_dev, ret);
4149 ret = rbd_request_lock(rbd_dev);
4150 if (ret == -ETIMEDOUT) {
4151 goto again; /* treat this as a dead client */
4152 } else if (ret == -EROFS) {
4153 rbd_warn(rbd_dev, "peer will not release lock");
4154 down_write(&rbd_dev->lock_rwsem);
4155 wake_lock_waiters(rbd_dev, ret);
4156 up_write(&rbd_dev->lock_rwsem);
4157 } else if (ret < 0) {
4158 rbd_warn(rbd_dev, "error requesting lock: %d", ret);
4159 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
4163 * lock owner acked, but resend if we don't see them
4166 dout("%s rbd_dev %p requeuing lock_dwork\n", __func__,
4168 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
4169 msecs_to_jiffies(2 * RBD_NOTIFY_TIMEOUT * MSEC_PER_SEC));
4173 static bool rbd_quiesce_lock(struct rbd_device *rbd_dev)
4175 dout("%s rbd_dev %p\n", __func__, rbd_dev);
4176 lockdep_assert_held_write(&rbd_dev->lock_rwsem);
4178 if (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED)
4182 * Ensure that all in-flight IO is flushed.
4184 rbd_dev->lock_state = RBD_LOCK_STATE_RELEASING;
4185 rbd_assert(!completion_done(&rbd_dev->releasing_wait));
4186 if (list_empty(&rbd_dev->running_list))
4189 up_write(&rbd_dev->lock_rwsem);
4190 wait_for_completion(&rbd_dev->releasing_wait);
4192 down_write(&rbd_dev->lock_rwsem);
4193 if (rbd_dev->lock_state != RBD_LOCK_STATE_RELEASING)
4196 rbd_assert(list_empty(&rbd_dev->running_list));
4200 static void rbd_pre_release_action(struct rbd_device *rbd_dev)
4202 if (rbd_dev->header.features & RBD_FEATURE_OBJECT_MAP)
4203 rbd_object_map_close(rbd_dev);
4206 static void __rbd_release_lock(struct rbd_device *rbd_dev)
4208 rbd_assert(list_empty(&rbd_dev->running_list));
4210 rbd_pre_release_action(rbd_dev);
4211 rbd_unlock(rbd_dev);
4215 * lock_rwsem must be held for write
4217 static void rbd_release_lock(struct rbd_device *rbd_dev)
4219 if (!rbd_quiesce_lock(rbd_dev))
4222 __rbd_release_lock(rbd_dev);
4225 * Give others a chance to grab the lock - we would re-acquire
4226 * almost immediately if we got new IO while draining the running
4227 * list otherwise. We need to ack our own notifications, so this
4228 * lock_dwork will be requeued from rbd_handle_released_lock() by
4229 * way of maybe_kick_acquire().
4231 cancel_delayed_work(&rbd_dev->lock_dwork);
4234 static void rbd_release_lock_work(struct work_struct *work)
4236 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
4239 down_write(&rbd_dev->lock_rwsem);
4240 rbd_release_lock(rbd_dev);
4241 up_write(&rbd_dev->lock_rwsem);
4244 static void maybe_kick_acquire(struct rbd_device *rbd_dev)
4248 dout("%s rbd_dev %p\n", __func__, rbd_dev);
4249 if (__rbd_is_lock_owner(rbd_dev))
4252 spin_lock(&rbd_dev->lock_lists_lock);
4253 have_requests = !list_empty(&rbd_dev->acquiring_list);
4254 spin_unlock(&rbd_dev->lock_lists_lock);
4255 if (have_requests || delayed_work_pending(&rbd_dev->lock_dwork)) {
4256 dout("%s rbd_dev %p kicking lock_dwork\n", __func__, rbd_dev);
4257 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork, 0);
4261 static void rbd_handle_acquired_lock(struct rbd_device *rbd_dev, u8 struct_v,
4264 struct rbd_client_id cid = { 0 };
4266 if (struct_v >= 2) {
4267 cid.gid = ceph_decode_64(p);
4268 cid.handle = ceph_decode_64(p);
4271 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
4273 if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
4274 down_write(&rbd_dev->lock_rwsem);
4275 if (rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
4276 dout("%s rbd_dev %p cid %llu-%llu == owner_cid\n",
4277 __func__, rbd_dev, cid.gid, cid.handle);
4279 rbd_set_owner_cid(rbd_dev, &cid);
4281 downgrade_write(&rbd_dev->lock_rwsem);
4283 down_read(&rbd_dev->lock_rwsem);
4286 maybe_kick_acquire(rbd_dev);
4287 up_read(&rbd_dev->lock_rwsem);
4290 static void rbd_handle_released_lock(struct rbd_device *rbd_dev, u8 struct_v,
4293 struct rbd_client_id cid = { 0 };
4295 if (struct_v >= 2) {
4296 cid.gid = ceph_decode_64(p);
4297 cid.handle = ceph_decode_64(p);
4300 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
4302 if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
4303 down_write(&rbd_dev->lock_rwsem);
4304 if (!rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
4305 dout("%s rbd_dev %p cid %llu-%llu != owner_cid %llu-%llu\n",
4306 __func__, rbd_dev, cid.gid, cid.handle,
4307 rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle);
4309 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
4311 downgrade_write(&rbd_dev->lock_rwsem);
4313 down_read(&rbd_dev->lock_rwsem);
4316 maybe_kick_acquire(rbd_dev);
4317 up_read(&rbd_dev->lock_rwsem);
4321 * Returns result for ResponseMessage to be encoded (<= 0), or 1 if no
4322 * ResponseMessage is needed.
4324 static int rbd_handle_request_lock(struct rbd_device *rbd_dev, u8 struct_v,
4327 struct rbd_client_id my_cid = rbd_get_cid(rbd_dev);
4328 struct rbd_client_id cid = { 0 };
4331 if (struct_v >= 2) {
4332 cid.gid = ceph_decode_64(p);
4333 cid.handle = ceph_decode_64(p);
4336 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
4338 if (rbd_cid_equal(&cid, &my_cid))
4341 down_read(&rbd_dev->lock_rwsem);
4342 if (__rbd_is_lock_owner(rbd_dev)) {
4343 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED &&
4344 rbd_cid_equal(&rbd_dev->owner_cid, &rbd_empty_cid))
4348 * encode ResponseMessage(0) so the peer can detect
4353 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED) {
4354 if (!rbd_dev->opts->exclusive) {
4355 dout("%s rbd_dev %p queueing unlock_work\n",
4357 queue_work(rbd_dev->task_wq,
4358 &rbd_dev->unlock_work);
4360 /* refuse to release the lock */
4367 up_read(&rbd_dev->lock_rwsem);
4371 static void __rbd_acknowledge_notify(struct rbd_device *rbd_dev,
4372 u64 notify_id, u64 cookie, s32 *result)
4374 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4375 char buf[4 + CEPH_ENCODING_START_BLK_LEN];
4376 int buf_size = sizeof(buf);
4382 /* encode ResponseMessage */
4383 ceph_start_encoding(&p, 1, 1,
4384 buf_size - CEPH_ENCODING_START_BLK_LEN);
4385 ceph_encode_32(&p, *result);
4390 ret = ceph_osdc_notify_ack(osdc, &rbd_dev->header_oid,
4391 &rbd_dev->header_oloc, notify_id, cookie,
4394 rbd_warn(rbd_dev, "acknowledge_notify failed: %d", ret);
4397 static void rbd_acknowledge_notify(struct rbd_device *rbd_dev, u64 notify_id,
4400 dout("%s rbd_dev %p\n", __func__, rbd_dev);
4401 __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, NULL);
4404 static void rbd_acknowledge_notify_result(struct rbd_device *rbd_dev,
4405 u64 notify_id, u64 cookie, s32 result)
4407 dout("%s rbd_dev %p result %d\n", __func__, rbd_dev, result);
4408 __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, &result);
4411 static void rbd_watch_cb(void *arg, u64 notify_id, u64 cookie,
4412 u64 notifier_id, void *data, size_t data_len)
4414 struct rbd_device *rbd_dev = arg;
4416 void *const end = p + data_len;
4422 dout("%s rbd_dev %p cookie %llu notify_id %llu data_len %zu\n",
4423 __func__, rbd_dev, cookie, notify_id, data_len);
4425 ret = ceph_start_decoding(&p, end, 1, "NotifyMessage",
4428 rbd_warn(rbd_dev, "failed to decode NotifyMessage: %d",
4433 notify_op = ceph_decode_32(&p);
4435 /* legacy notification for header updates */
4436 notify_op = RBD_NOTIFY_OP_HEADER_UPDATE;
4440 dout("%s rbd_dev %p notify_op %u\n", __func__, rbd_dev, notify_op);
4441 switch (notify_op) {
4442 case RBD_NOTIFY_OP_ACQUIRED_LOCK:
4443 rbd_handle_acquired_lock(rbd_dev, struct_v, &p);
4444 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
4446 case RBD_NOTIFY_OP_RELEASED_LOCK:
4447 rbd_handle_released_lock(rbd_dev, struct_v, &p);
4448 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
4450 case RBD_NOTIFY_OP_REQUEST_LOCK:
4451 ret = rbd_handle_request_lock(rbd_dev, struct_v, &p);
4453 rbd_acknowledge_notify_result(rbd_dev, notify_id,
4456 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
4458 case RBD_NOTIFY_OP_HEADER_UPDATE:
4459 ret = rbd_dev_refresh(rbd_dev);
4461 rbd_warn(rbd_dev, "refresh failed: %d", ret);
4463 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
4466 if (rbd_is_lock_owner(rbd_dev))
4467 rbd_acknowledge_notify_result(rbd_dev, notify_id,
4468 cookie, -EOPNOTSUPP);
4470 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
4475 static void __rbd_unregister_watch(struct rbd_device *rbd_dev);
4477 static void rbd_watch_errcb(void *arg, u64 cookie, int err)
4479 struct rbd_device *rbd_dev = arg;
4481 rbd_warn(rbd_dev, "encountered watch error: %d", err);
4483 down_write(&rbd_dev->lock_rwsem);
4484 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
4485 up_write(&rbd_dev->lock_rwsem);
4487 mutex_lock(&rbd_dev->watch_mutex);
4488 if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED) {
4489 __rbd_unregister_watch(rbd_dev);
4490 rbd_dev->watch_state = RBD_WATCH_STATE_ERROR;
4492 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->watch_dwork, 0);
4494 mutex_unlock(&rbd_dev->watch_mutex);
4498 * watch_mutex must be locked
4500 static int __rbd_register_watch(struct rbd_device *rbd_dev)
4502 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4503 struct ceph_osd_linger_request *handle;
4505 rbd_assert(!rbd_dev->watch_handle);
4506 dout("%s rbd_dev %p\n", __func__, rbd_dev);
4508 handle = ceph_osdc_watch(osdc, &rbd_dev->header_oid,
4509 &rbd_dev->header_oloc, rbd_watch_cb,
4510 rbd_watch_errcb, rbd_dev);
4512 return PTR_ERR(handle);
4514 rbd_dev->watch_handle = handle;
4519 * watch_mutex must be locked
4521 static void __rbd_unregister_watch(struct rbd_device *rbd_dev)
4523 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4526 rbd_assert(rbd_dev->watch_handle);
4527 dout("%s rbd_dev %p\n", __func__, rbd_dev);
4529 ret = ceph_osdc_unwatch(osdc, rbd_dev->watch_handle);
4531 rbd_warn(rbd_dev, "failed to unwatch: %d", ret);
4533 rbd_dev->watch_handle = NULL;
4536 static int rbd_register_watch(struct rbd_device *rbd_dev)
4540 mutex_lock(&rbd_dev->watch_mutex);
4541 rbd_assert(rbd_dev->watch_state == RBD_WATCH_STATE_UNREGISTERED);
4542 ret = __rbd_register_watch(rbd_dev);
4546 rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
4547 rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
4550 mutex_unlock(&rbd_dev->watch_mutex);
4554 static void cancel_tasks_sync(struct rbd_device *rbd_dev)
4556 dout("%s rbd_dev %p\n", __func__, rbd_dev);
4558 cancel_work_sync(&rbd_dev->acquired_lock_work);
4559 cancel_work_sync(&rbd_dev->released_lock_work);
4560 cancel_delayed_work_sync(&rbd_dev->lock_dwork);
4561 cancel_work_sync(&rbd_dev->unlock_work);
4565 * header_rwsem must not be held to avoid a deadlock with
4566 * rbd_dev_refresh() when flushing notifies.
4568 static void rbd_unregister_watch(struct rbd_device *rbd_dev)
4570 cancel_tasks_sync(rbd_dev);
4572 mutex_lock(&rbd_dev->watch_mutex);
4573 if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED)
4574 __rbd_unregister_watch(rbd_dev);
4575 rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
4576 mutex_unlock(&rbd_dev->watch_mutex);
4578 cancel_delayed_work_sync(&rbd_dev->watch_dwork);
4579 ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
4583 * lock_rwsem must be held for write
4585 static void rbd_reacquire_lock(struct rbd_device *rbd_dev)
4587 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4591 if (!rbd_quiesce_lock(rbd_dev))
4594 format_lock_cookie(rbd_dev, cookie);
4595 ret = ceph_cls_set_cookie(osdc, &rbd_dev->header_oid,
4596 &rbd_dev->header_oloc, RBD_LOCK_NAME,
4597 CEPH_CLS_LOCK_EXCLUSIVE, rbd_dev->lock_cookie,
4598 RBD_LOCK_TAG, cookie);
4600 if (ret != -EOPNOTSUPP)
4601 rbd_warn(rbd_dev, "failed to update lock cookie: %d",
4605 * Lock cookie cannot be updated on older OSDs, so do
4606 * a manual release and queue an acquire.
4608 __rbd_release_lock(rbd_dev);
4609 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork, 0);
4611 __rbd_lock(rbd_dev, cookie);
4612 wake_lock_waiters(rbd_dev, 0);
4616 static void rbd_reregister_watch(struct work_struct *work)
4618 struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
4619 struct rbd_device, watch_dwork);
4622 dout("%s rbd_dev %p\n", __func__, rbd_dev);
4624 mutex_lock(&rbd_dev->watch_mutex);
4625 if (rbd_dev->watch_state != RBD_WATCH_STATE_ERROR) {
4626 mutex_unlock(&rbd_dev->watch_mutex);
4630 ret = __rbd_register_watch(rbd_dev);
4632 rbd_warn(rbd_dev, "failed to reregister watch: %d", ret);
4633 if (ret != -EBLOCKLISTED && ret != -ENOENT) {
4634 queue_delayed_work(rbd_dev->task_wq,
4635 &rbd_dev->watch_dwork,
4637 mutex_unlock(&rbd_dev->watch_mutex);
4641 mutex_unlock(&rbd_dev->watch_mutex);
4642 down_write(&rbd_dev->lock_rwsem);
4643 wake_lock_waiters(rbd_dev, ret);
4644 up_write(&rbd_dev->lock_rwsem);
4648 rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
4649 rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
4650 mutex_unlock(&rbd_dev->watch_mutex);
4652 down_write(&rbd_dev->lock_rwsem);
4653 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED)
4654 rbd_reacquire_lock(rbd_dev);
4655 up_write(&rbd_dev->lock_rwsem);
4657 ret = rbd_dev_refresh(rbd_dev);
4659 rbd_warn(rbd_dev, "reregistration refresh failed: %d", ret);
4663 * Synchronous osd object method call. Returns the number of bytes
4664 * returned in the outbound buffer, or a negative error code.
4666 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
4667 struct ceph_object_id *oid,
4668 struct ceph_object_locator *oloc,
4669 const char *method_name,
4670 const void *outbound,
4671 size_t outbound_size,
4673 size_t inbound_size)
4675 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4676 struct page *req_page = NULL;
4677 struct page *reply_page;
4681 * Method calls are ultimately read operations. The result
4682 * should placed into the inbound buffer provided. They
4683 * also supply outbound data--parameters for the object
4684 * method. Currently if this is present it will be a
4688 if (outbound_size > PAGE_SIZE)
4691 req_page = alloc_page(GFP_KERNEL);
4695 memcpy(page_address(req_page), outbound, outbound_size);
4698 reply_page = alloc_page(GFP_KERNEL);
4701 __free_page(req_page);
4705 ret = ceph_osdc_call(osdc, oid, oloc, RBD_DRV_NAME, method_name,
4706 CEPH_OSD_FLAG_READ, req_page, outbound_size,
4707 &reply_page, &inbound_size);
4709 memcpy(inbound, page_address(reply_page), inbound_size);
4714 __free_page(req_page);
4715 __free_page(reply_page);
4719 static void rbd_queue_workfn(struct work_struct *work)
4721 struct rbd_img_request *img_request =
4722 container_of(work, struct rbd_img_request, work);
4723 struct rbd_device *rbd_dev = img_request->rbd_dev;
4724 enum obj_operation_type op_type = img_request->op_type;
4725 struct request *rq = blk_mq_rq_from_pdu(img_request);
4726 u64 offset = (u64)blk_rq_pos(rq) << SECTOR_SHIFT;
4727 u64 length = blk_rq_bytes(rq);
4731 /* Ignore/skip any zero-length requests */
4733 dout("%s: zero-length request\n", __func__);
4735 goto err_img_request;
4738 blk_mq_start_request(rq);
4740 down_read(&rbd_dev->header_rwsem);
4741 mapping_size = rbd_dev->mapping.size;
4742 rbd_img_capture_header(img_request);
4743 up_read(&rbd_dev->header_rwsem);
4745 if (offset + length > mapping_size) {
4746 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)", offset,
4747 length, mapping_size);
4749 goto err_img_request;
4752 dout("%s rbd_dev %p img_req %p %s %llu~%llu\n", __func__, rbd_dev,
4753 img_request, obj_op_name(op_type), offset, length);
4755 if (op_type == OBJ_OP_DISCARD || op_type == OBJ_OP_ZEROOUT)
4756 result = rbd_img_fill_nodata(img_request, offset, length);
4758 result = rbd_img_fill_from_bio(img_request, offset, length,
4761 goto err_img_request;
4763 rbd_img_handle_request(img_request, 0);
4767 rbd_img_request_destroy(img_request);
4769 rbd_warn(rbd_dev, "%s %llx at %llx result %d",
4770 obj_op_name(op_type), length, offset, result);
4771 blk_mq_end_request(rq, errno_to_blk_status(result));
4774 static blk_status_t rbd_queue_rq(struct blk_mq_hw_ctx *hctx,
4775 const struct blk_mq_queue_data *bd)
4777 struct rbd_device *rbd_dev = hctx->queue->queuedata;
4778 struct rbd_img_request *img_req = blk_mq_rq_to_pdu(bd->rq);
4779 enum obj_operation_type op_type;
4781 switch (req_op(bd->rq)) {
4782 case REQ_OP_DISCARD:
4783 op_type = OBJ_OP_DISCARD;
4785 case REQ_OP_WRITE_ZEROES:
4786 op_type = OBJ_OP_ZEROOUT;
4789 op_type = OBJ_OP_WRITE;
4792 op_type = OBJ_OP_READ;
4795 rbd_warn(rbd_dev, "unknown req_op %d", req_op(bd->rq));
4796 return BLK_STS_IOERR;
4799 rbd_img_request_init(img_req, rbd_dev, op_type);
4801 if (rbd_img_is_write(img_req)) {
4802 if (rbd_is_ro(rbd_dev)) {
4803 rbd_warn(rbd_dev, "%s on read-only mapping",
4804 obj_op_name(img_req->op_type));
4805 return BLK_STS_IOERR;
4807 rbd_assert(!rbd_is_snap(rbd_dev));
4810 INIT_WORK(&img_req->work, rbd_queue_workfn);
4811 queue_work(rbd_wq, &img_req->work);
4815 static void rbd_free_disk(struct rbd_device *rbd_dev)
4817 put_disk(rbd_dev->disk);
4818 blk_mq_free_tag_set(&rbd_dev->tag_set);
4819 rbd_dev->disk = NULL;
4822 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
4823 struct ceph_object_id *oid,
4824 struct ceph_object_locator *oloc,
4825 void *buf, int buf_len)
4828 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4829 struct ceph_osd_request *req;
4830 struct page **pages;
4831 int num_pages = calc_pages_for(0, buf_len);
4834 req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
4838 ceph_oid_copy(&req->r_base_oid, oid);
4839 ceph_oloc_copy(&req->r_base_oloc, oloc);
4840 req->r_flags = CEPH_OSD_FLAG_READ;
4842 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
4843 if (IS_ERR(pages)) {
4844 ret = PTR_ERR(pages);
4848 osd_req_op_extent_init(req, 0, CEPH_OSD_OP_READ, 0, buf_len, 0, 0);
4849 osd_req_op_extent_osd_data_pages(req, 0, pages, buf_len, 0, false,
4852 ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
4856 ceph_osdc_start_request(osdc, req);
4857 ret = ceph_osdc_wait_request(osdc, req);
4859 ceph_copy_from_page_vector(pages, buf, 0, ret);
4862 ceph_osdc_put_request(req);
4867 * Read the complete header for the given rbd device. On successful
4868 * return, the rbd_dev->header field will contain up-to-date
4869 * information about the image.
4871 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev,
4872 struct rbd_image_header *header,
4875 struct rbd_image_header_ondisk *ondisk = NULL;
4882 * The complete header will include an array of its 64-bit
4883 * snapshot ids, followed by the names of those snapshots as
4884 * a contiguous block of NUL-terminated strings. Note that
4885 * the number of snapshots could change by the time we read
4886 * it in, in which case we re-read it.
4893 size = sizeof (*ondisk);
4894 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
4896 ondisk = kmalloc(size, GFP_KERNEL);
4900 ret = rbd_obj_read_sync(rbd_dev, &rbd_dev->header_oid,
4901 &rbd_dev->header_oloc, ondisk, size);
4904 if ((size_t)ret < size) {
4906 rbd_warn(rbd_dev, "short header read (want %zd got %d)",
4910 if (!rbd_dev_ondisk_valid(ondisk)) {
4912 rbd_warn(rbd_dev, "invalid header");
4916 names_size = le64_to_cpu(ondisk->snap_names_len);
4917 want_count = snap_count;
4918 snap_count = le32_to_cpu(ondisk->snap_count);
4919 } while (snap_count != want_count);
4921 ret = rbd_header_from_disk(header, ondisk, first_time);
4928 static void rbd_dev_update_size(struct rbd_device *rbd_dev)
4933 * If EXISTS is not set, rbd_dev->disk may be NULL, so don't
4934 * try to update its size. If REMOVING is set, updating size
4935 * is just useless work since the device can't be opened.
4937 if (test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags) &&
4938 !test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags)) {
4939 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
4940 dout("setting size to %llu sectors", (unsigned long long)size);
4941 set_capacity_and_notify(rbd_dev->disk, size);
4945 static const struct blk_mq_ops rbd_mq_ops = {
4946 .queue_rq = rbd_queue_rq,
4949 static int rbd_init_disk(struct rbd_device *rbd_dev)
4951 struct gendisk *disk;
4952 struct request_queue *q;
4953 unsigned int objset_bytes =
4954 rbd_dev->layout.object_size * rbd_dev->layout.stripe_count;
4955 struct queue_limits lim = {
4956 .max_hw_sectors = objset_bytes >> SECTOR_SHIFT,
4957 .max_user_sectors = objset_bytes >> SECTOR_SHIFT,
4958 .io_min = rbd_dev->opts->alloc_size,
4959 .io_opt = rbd_dev->opts->alloc_size,
4960 .max_segments = USHRT_MAX,
4961 .max_segment_size = UINT_MAX,
4965 memset(&rbd_dev->tag_set, 0, sizeof(rbd_dev->tag_set));
4966 rbd_dev->tag_set.ops = &rbd_mq_ops;
4967 rbd_dev->tag_set.queue_depth = rbd_dev->opts->queue_depth;
4968 rbd_dev->tag_set.numa_node = NUMA_NO_NODE;
4969 rbd_dev->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
4970 rbd_dev->tag_set.nr_hw_queues = num_present_cpus();
4971 rbd_dev->tag_set.cmd_size = sizeof(struct rbd_img_request);
4973 err = blk_mq_alloc_tag_set(&rbd_dev->tag_set);
4977 if (rbd_dev->opts->trim) {
4978 lim.discard_granularity = rbd_dev->opts->alloc_size;
4979 lim.max_hw_discard_sectors = objset_bytes >> SECTOR_SHIFT;
4980 lim.max_write_zeroes_sectors = objset_bytes >> SECTOR_SHIFT;
4983 disk = blk_mq_alloc_disk(&rbd_dev->tag_set, &lim, rbd_dev);
4985 err = PTR_ERR(disk);
4990 snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
4992 disk->major = rbd_dev->major;
4993 disk->first_minor = rbd_dev->minor;
4995 disk->minors = (1 << RBD_SINGLE_MAJOR_PART_SHIFT);
4997 disk->minors = RBD_MINORS_PER_MAJOR;
4998 disk->fops = &rbd_bd_ops;
4999 disk->private_data = rbd_dev;
5001 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
5002 /* QUEUE_FLAG_ADD_RANDOM is off by default for blk-mq */
5004 if (!ceph_test_opt(rbd_dev->rbd_client->client, NOCRC))
5005 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, q);
5007 rbd_dev->disk = disk;
5011 blk_mq_free_tag_set(&rbd_dev->tag_set);
5019 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
5021 return container_of(dev, struct rbd_device, dev);
5024 static ssize_t rbd_size_show(struct device *dev,
5025 struct device_attribute *attr, char *buf)
5027 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5029 return sprintf(buf, "%llu\n",
5030 (unsigned long long)rbd_dev->mapping.size);
5033 static ssize_t rbd_features_show(struct device *dev,
5034 struct device_attribute *attr, char *buf)
5036 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5038 return sprintf(buf, "0x%016llx\n", rbd_dev->header.features);
5041 static ssize_t rbd_major_show(struct device *dev,
5042 struct device_attribute *attr, char *buf)
5044 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5047 return sprintf(buf, "%d\n", rbd_dev->major);
5049 return sprintf(buf, "(none)\n");
5052 static ssize_t rbd_minor_show(struct device *dev,
5053 struct device_attribute *attr, char *buf)
5055 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5057 return sprintf(buf, "%d\n", rbd_dev->minor);
5060 static ssize_t rbd_client_addr_show(struct device *dev,
5061 struct device_attribute *attr, char *buf)
5063 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5064 struct ceph_entity_addr *client_addr =
5065 ceph_client_addr(rbd_dev->rbd_client->client);
5067 return sprintf(buf, "%pISpc/%u\n", &client_addr->in_addr,
5068 le32_to_cpu(client_addr->nonce));
5071 static ssize_t rbd_client_id_show(struct device *dev,
5072 struct device_attribute *attr, char *buf)
5074 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5076 return sprintf(buf, "client%lld\n",
5077 ceph_client_gid(rbd_dev->rbd_client->client));
5080 static ssize_t rbd_cluster_fsid_show(struct device *dev,
5081 struct device_attribute *attr, char *buf)
5083 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5085 return sprintf(buf, "%pU\n", &rbd_dev->rbd_client->client->fsid);
5088 static ssize_t rbd_config_info_show(struct device *dev,
5089 struct device_attribute *attr, char *buf)
5091 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5093 if (!capable(CAP_SYS_ADMIN))
5096 return sprintf(buf, "%s\n", rbd_dev->config_info);
5099 static ssize_t rbd_pool_show(struct device *dev,
5100 struct device_attribute *attr, char *buf)
5102 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5104 return sprintf(buf, "%s\n", rbd_dev->spec->pool_name);
5107 static ssize_t rbd_pool_id_show(struct device *dev,
5108 struct device_attribute *attr, char *buf)
5110 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5112 return sprintf(buf, "%llu\n",
5113 (unsigned long long) rbd_dev->spec->pool_id);
5116 static ssize_t rbd_pool_ns_show(struct device *dev,
5117 struct device_attribute *attr, char *buf)
5119 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5121 return sprintf(buf, "%s\n", rbd_dev->spec->pool_ns ?: "");
5124 static ssize_t rbd_name_show(struct device *dev,
5125 struct device_attribute *attr, char *buf)
5127 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5129 if (rbd_dev->spec->image_name)
5130 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
5132 return sprintf(buf, "(unknown)\n");
5135 static ssize_t rbd_image_id_show(struct device *dev,
5136 struct device_attribute *attr, char *buf)
5138 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5140 return sprintf(buf, "%s\n", rbd_dev->spec->image_id);
5144 * Shows the name of the currently-mapped snapshot (or
5145 * RBD_SNAP_HEAD_NAME for the base image).
5147 static ssize_t rbd_snap_show(struct device *dev,
5148 struct device_attribute *attr,
5151 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5153 return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
5156 static ssize_t rbd_snap_id_show(struct device *dev,
5157 struct device_attribute *attr, char *buf)
5159 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5161 return sprintf(buf, "%llu\n", rbd_dev->spec->snap_id);
5165 * For a v2 image, shows the chain of parent images, separated by empty
5166 * lines. For v1 images or if there is no parent, shows "(no parent
5169 static ssize_t rbd_parent_show(struct device *dev,
5170 struct device_attribute *attr,
5173 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5176 if (!rbd_dev->parent)
5177 return sprintf(buf, "(no parent image)\n");
5179 for ( ; rbd_dev->parent; rbd_dev = rbd_dev->parent) {
5180 struct rbd_spec *spec = rbd_dev->parent_spec;
5182 count += sprintf(&buf[count], "%s"
5183 "pool_id %llu\npool_name %s\n"
5185 "image_id %s\nimage_name %s\n"
5186 "snap_id %llu\nsnap_name %s\n"
5188 !count ? "" : "\n", /* first? */
5189 spec->pool_id, spec->pool_name,
5190 spec->pool_ns ?: "",
5191 spec->image_id, spec->image_name ?: "(unknown)",
5192 spec->snap_id, spec->snap_name,
5193 rbd_dev->parent_overlap);
5199 static ssize_t rbd_image_refresh(struct device *dev,
5200 struct device_attribute *attr,
5204 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5207 if (!capable(CAP_SYS_ADMIN))
5210 ret = rbd_dev_refresh(rbd_dev);
5217 static DEVICE_ATTR(size, 0444, rbd_size_show, NULL);
5218 static DEVICE_ATTR(features, 0444, rbd_features_show, NULL);
5219 static DEVICE_ATTR(major, 0444, rbd_major_show, NULL);
5220 static DEVICE_ATTR(minor, 0444, rbd_minor_show, NULL);
5221 static DEVICE_ATTR(client_addr, 0444, rbd_client_addr_show, NULL);
5222 static DEVICE_ATTR(client_id, 0444, rbd_client_id_show, NULL);
5223 static DEVICE_ATTR(cluster_fsid, 0444, rbd_cluster_fsid_show, NULL);
5224 static DEVICE_ATTR(config_info, 0400, rbd_config_info_show, NULL);
5225 static DEVICE_ATTR(pool, 0444, rbd_pool_show, NULL);
5226 static DEVICE_ATTR(pool_id, 0444, rbd_pool_id_show, NULL);
5227 static DEVICE_ATTR(pool_ns, 0444, rbd_pool_ns_show, NULL);
5228 static DEVICE_ATTR(name, 0444, rbd_name_show, NULL);
5229 static DEVICE_ATTR(image_id, 0444, rbd_image_id_show, NULL);
5230 static DEVICE_ATTR(refresh, 0200, NULL, rbd_image_refresh);
5231 static DEVICE_ATTR(current_snap, 0444, rbd_snap_show, NULL);
5232 static DEVICE_ATTR(snap_id, 0444, rbd_snap_id_show, NULL);
5233 static DEVICE_ATTR(parent, 0444, rbd_parent_show, NULL);
5235 static struct attribute *rbd_attrs[] = {
5236 &dev_attr_size.attr,
5237 &dev_attr_features.attr,
5238 &dev_attr_major.attr,
5239 &dev_attr_minor.attr,
5240 &dev_attr_client_addr.attr,
5241 &dev_attr_client_id.attr,
5242 &dev_attr_cluster_fsid.attr,
5243 &dev_attr_config_info.attr,
5244 &dev_attr_pool.attr,
5245 &dev_attr_pool_id.attr,
5246 &dev_attr_pool_ns.attr,
5247 &dev_attr_name.attr,
5248 &dev_attr_image_id.attr,
5249 &dev_attr_current_snap.attr,
5250 &dev_attr_snap_id.attr,
5251 &dev_attr_parent.attr,
5252 &dev_attr_refresh.attr,
5256 static struct attribute_group rbd_attr_group = {
5260 static const struct attribute_group *rbd_attr_groups[] = {
5265 static void rbd_dev_release(struct device *dev);
5267 static const struct device_type rbd_device_type = {
5269 .groups = rbd_attr_groups,
5270 .release = rbd_dev_release,
5273 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
5275 kref_get(&spec->kref);
5280 static void rbd_spec_free(struct kref *kref);
5281 static void rbd_spec_put(struct rbd_spec *spec)
5284 kref_put(&spec->kref, rbd_spec_free);
5287 static struct rbd_spec *rbd_spec_alloc(void)
5289 struct rbd_spec *spec;
5291 spec = kzalloc(sizeof (*spec), GFP_KERNEL);
5295 spec->pool_id = CEPH_NOPOOL;
5296 spec->snap_id = CEPH_NOSNAP;
5297 kref_init(&spec->kref);
5302 static void rbd_spec_free(struct kref *kref)
5304 struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
5306 kfree(spec->pool_name);
5307 kfree(spec->pool_ns);
5308 kfree(spec->image_id);
5309 kfree(spec->image_name);
5310 kfree(spec->snap_name);
5314 static void rbd_dev_free(struct rbd_device *rbd_dev)
5316 WARN_ON(rbd_dev->watch_state != RBD_WATCH_STATE_UNREGISTERED);
5317 WARN_ON(rbd_dev->lock_state != RBD_LOCK_STATE_UNLOCKED);
5319 ceph_oid_destroy(&rbd_dev->header_oid);
5320 ceph_oloc_destroy(&rbd_dev->header_oloc);
5321 kfree(rbd_dev->config_info);
5323 rbd_put_client(rbd_dev->rbd_client);
5324 rbd_spec_put(rbd_dev->spec);
5325 kfree(rbd_dev->opts);
5329 static void rbd_dev_release(struct device *dev)
5331 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5332 bool need_put = !!rbd_dev->opts;
5335 destroy_workqueue(rbd_dev->task_wq);
5336 ida_free(&rbd_dev_id_ida, rbd_dev->dev_id);
5339 rbd_dev_free(rbd_dev);
5342 * This is racy, but way better than putting module outside of
5343 * the release callback. The race window is pretty small, so
5344 * doing something similar to dm (dm-builtin.c) is overkill.
5347 module_put(THIS_MODULE);
5350 static struct rbd_device *__rbd_dev_create(struct rbd_spec *spec)
5352 struct rbd_device *rbd_dev;
5354 rbd_dev = kzalloc(sizeof(*rbd_dev), GFP_KERNEL);
5358 spin_lock_init(&rbd_dev->lock);
5359 INIT_LIST_HEAD(&rbd_dev->node);
5360 init_rwsem(&rbd_dev->header_rwsem);
5362 rbd_dev->header.data_pool_id = CEPH_NOPOOL;
5363 ceph_oid_init(&rbd_dev->header_oid);
5364 rbd_dev->header_oloc.pool = spec->pool_id;
5365 if (spec->pool_ns) {
5366 WARN_ON(!*spec->pool_ns);
5367 rbd_dev->header_oloc.pool_ns =
5368 ceph_find_or_create_string(spec->pool_ns,
5369 strlen(spec->pool_ns));
5372 mutex_init(&rbd_dev->watch_mutex);
5373 rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
5374 INIT_DELAYED_WORK(&rbd_dev->watch_dwork, rbd_reregister_watch);
5376 init_rwsem(&rbd_dev->lock_rwsem);
5377 rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
5378 INIT_WORK(&rbd_dev->acquired_lock_work, rbd_notify_acquired_lock);
5379 INIT_WORK(&rbd_dev->released_lock_work, rbd_notify_released_lock);
5380 INIT_DELAYED_WORK(&rbd_dev->lock_dwork, rbd_acquire_lock);
5381 INIT_WORK(&rbd_dev->unlock_work, rbd_release_lock_work);
5382 spin_lock_init(&rbd_dev->lock_lists_lock);
5383 INIT_LIST_HEAD(&rbd_dev->acquiring_list);
5384 INIT_LIST_HEAD(&rbd_dev->running_list);
5385 init_completion(&rbd_dev->acquire_wait);
5386 init_completion(&rbd_dev->releasing_wait);
5388 spin_lock_init(&rbd_dev->object_map_lock);
5390 rbd_dev->dev.bus = &rbd_bus_type;
5391 rbd_dev->dev.type = &rbd_device_type;
5392 rbd_dev->dev.parent = &rbd_root_dev;
5393 device_initialize(&rbd_dev->dev);
5399 * Create a mapping rbd_dev.
5401 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
5402 struct rbd_spec *spec,
5403 struct rbd_options *opts)
5405 struct rbd_device *rbd_dev;
5407 rbd_dev = __rbd_dev_create(spec);
5411 /* get an id and fill in device name */
5412 rbd_dev->dev_id = ida_alloc_max(&rbd_dev_id_ida,
5413 minor_to_rbd_dev_id(1 << MINORBITS) - 1,
5415 if (rbd_dev->dev_id < 0)
5418 sprintf(rbd_dev->name, RBD_DRV_NAME "%d", rbd_dev->dev_id);
5419 rbd_dev->task_wq = alloc_ordered_workqueue("%s-tasks", WQ_MEM_RECLAIM,
5421 if (!rbd_dev->task_wq)
5424 /* we have a ref from do_rbd_add() */
5425 __module_get(THIS_MODULE);
5427 rbd_dev->rbd_client = rbdc;
5428 rbd_dev->spec = spec;
5429 rbd_dev->opts = opts;
5431 dout("%s rbd_dev %p dev_id %d\n", __func__, rbd_dev, rbd_dev->dev_id);
5435 ida_free(&rbd_dev_id_ida, rbd_dev->dev_id);
5437 rbd_dev_free(rbd_dev);
5441 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
5444 put_device(&rbd_dev->dev);
5448 * Get the size and object order for an image snapshot, or if
5449 * snap_id is CEPH_NOSNAP, gets this information for the base
5452 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
5453 u8 *order, u64 *snap_size)
5455 __le64 snapid = cpu_to_le64(snap_id);
5460 } __attribute__ ((packed)) size_buf = { 0 };
5462 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
5463 &rbd_dev->header_oloc, "get_size",
5464 &snapid, sizeof(snapid),
5465 &size_buf, sizeof(size_buf));
5466 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5469 if (ret < sizeof (size_buf))
5473 *order = size_buf.order;
5474 dout(" order %u", (unsigned int)*order);
5476 *snap_size = le64_to_cpu(size_buf.size);
5478 dout(" snap_id 0x%016llx snap_size = %llu\n",
5479 (unsigned long long)snap_id,
5480 (unsigned long long)*snap_size);
5485 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev,
5486 char **pobject_prefix)
5490 char *object_prefix;
5494 /* Response will be an encoded string, which includes a length */
5495 size = sizeof(__le32) + RBD_OBJ_PREFIX_LEN_MAX;
5496 reply_buf = kzalloc(size, GFP_KERNEL);
5500 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
5501 &rbd_dev->header_oloc, "get_object_prefix",
5502 NULL, 0, reply_buf, size);
5503 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5508 object_prefix = ceph_extract_encoded_string(&p, p + ret, NULL,
5510 if (IS_ERR(object_prefix)) {
5511 ret = PTR_ERR(object_prefix);
5516 *pobject_prefix = object_prefix;
5517 dout(" object_prefix = %s\n", object_prefix);
5524 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
5525 bool read_only, u64 *snap_features)
5534 } __attribute__ ((packed)) features_buf = { 0 };
5538 features_in.snap_id = cpu_to_le64(snap_id);
5539 features_in.read_only = read_only;
5541 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
5542 &rbd_dev->header_oloc, "get_features",
5543 &features_in, sizeof(features_in),
5544 &features_buf, sizeof(features_buf));
5545 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5548 if (ret < sizeof (features_buf))
5551 unsup = le64_to_cpu(features_buf.incompat) & ~RBD_FEATURES_SUPPORTED;
5553 rbd_warn(rbd_dev, "image uses unsupported features: 0x%llx",
5558 *snap_features = le64_to_cpu(features_buf.features);
5560 dout(" snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
5561 (unsigned long long)snap_id,
5562 (unsigned long long)*snap_features,
5563 (unsigned long long)le64_to_cpu(features_buf.incompat));
5569 * These are generic image flags, but since they are used only for
5570 * object map, store them in rbd_dev->object_map_flags.
5572 * For the same reason, this function is called only on object map
5573 * (re)load and not on header refresh.
5575 static int rbd_dev_v2_get_flags(struct rbd_device *rbd_dev)
5577 __le64 snapid = cpu_to_le64(rbd_dev->spec->snap_id);
5581 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
5582 &rbd_dev->header_oloc, "get_flags",
5583 &snapid, sizeof(snapid),
5584 &flags, sizeof(flags));
5587 if (ret < sizeof(flags))
5590 rbd_dev->object_map_flags = le64_to_cpu(flags);
5594 struct parent_image_info {
5596 const char *pool_ns;
5597 const char *image_id;
5604 static void rbd_parent_info_cleanup(struct parent_image_info *pii)
5606 kfree(pii->pool_ns);
5607 kfree(pii->image_id);
5609 memset(pii, 0, sizeof(*pii));
5613 * The caller is responsible for @pii.
5615 static int decode_parent_image_spec(void **p, void *end,
5616 struct parent_image_info *pii)
5622 ret = ceph_start_decoding(p, end, 1, "ParentImageSpec",
5623 &struct_v, &struct_len);
5627 ceph_decode_64_safe(p, end, pii->pool_id, e_inval);
5628 pii->pool_ns = ceph_extract_encoded_string(p, end, NULL, GFP_KERNEL);
5629 if (IS_ERR(pii->pool_ns)) {
5630 ret = PTR_ERR(pii->pool_ns);
5631 pii->pool_ns = NULL;
5634 pii->image_id = ceph_extract_encoded_string(p, end, NULL, GFP_KERNEL);
5635 if (IS_ERR(pii->image_id)) {
5636 ret = PTR_ERR(pii->image_id);
5637 pii->image_id = NULL;
5640 ceph_decode_64_safe(p, end, pii->snap_id, e_inval);
5647 static int __get_parent_info(struct rbd_device *rbd_dev,
5648 struct page *req_page,
5649 struct page *reply_page,
5650 struct parent_image_info *pii)
5652 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
5653 size_t reply_len = PAGE_SIZE;
5657 ret = ceph_osdc_call(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
5658 "rbd", "parent_get", CEPH_OSD_FLAG_READ,
5659 req_page, sizeof(u64), &reply_page, &reply_len);
5661 return ret == -EOPNOTSUPP ? 1 : ret;
5663 p = page_address(reply_page);
5664 end = p + reply_len;
5665 ret = decode_parent_image_spec(&p, end, pii);
5669 ret = ceph_osdc_call(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
5670 "rbd", "parent_overlap_get", CEPH_OSD_FLAG_READ,
5671 req_page, sizeof(u64), &reply_page, &reply_len);
5675 p = page_address(reply_page);
5676 end = p + reply_len;
5677 ceph_decode_8_safe(&p, end, pii->has_overlap, e_inval);
5678 if (pii->has_overlap)
5679 ceph_decode_64_safe(&p, end, pii->overlap, e_inval);
5681 dout("%s pool_id %llu pool_ns %s image_id %s snap_id %llu has_overlap %d overlap %llu\n",
5682 __func__, pii->pool_id, pii->pool_ns, pii->image_id, pii->snap_id,
5683 pii->has_overlap, pii->overlap);
5691 * The caller is responsible for @pii.
5693 static int __get_parent_info_legacy(struct rbd_device *rbd_dev,
5694 struct page *req_page,
5695 struct page *reply_page,
5696 struct parent_image_info *pii)
5698 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
5699 size_t reply_len = PAGE_SIZE;
5703 ret = ceph_osdc_call(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
5704 "rbd", "get_parent", CEPH_OSD_FLAG_READ,
5705 req_page, sizeof(u64), &reply_page, &reply_len);
5709 p = page_address(reply_page);
5710 end = p + reply_len;
5711 ceph_decode_64_safe(&p, end, pii->pool_id, e_inval);
5712 pii->image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
5713 if (IS_ERR(pii->image_id)) {
5714 ret = PTR_ERR(pii->image_id);
5715 pii->image_id = NULL;
5718 ceph_decode_64_safe(&p, end, pii->snap_id, e_inval);
5719 pii->has_overlap = true;
5720 ceph_decode_64_safe(&p, end, pii->overlap, e_inval);
5722 dout("%s pool_id %llu pool_ns %s image_id %s snap_id %llu has_overlap %d overlap %llu\n",
5723 __func__, pii->pool_id, pii->pool_ns, pii->image_id, pii->snap_id,
5724 pii->has_overlap, pii->overlap);
5731 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev,
5732 struct parent_image_info *pii)
5734 struct page *req_page, *reply_page;
5738 req_page = alloc_page(GFP_KERNEL);
5742 reply_page = alloc_page(GFP_KERNEL);
5744 __free_page(req_page);
5748 p = page_address(req_page);
5749 ceph_encode_64(&p, rbd_dev->spec->snap_id);
5750 ret = __get_parent_info(rbd_dev, req_page, reply_page, pii);
5752 ret = __get_parent_info_legacy(rbd_dev, req_page, reply_page,
5755 __free_page(req_page);
5756 __free_page(reply_page);
5760 static int rbd_dev_setup_parent(struct rbd_device *rbd_dev)
5762 struct rbd_spec *parent_spec;
5763 struct parent_image_info pii = { 0 };
5766 parent_spec = rbd_spec_alloc();
5770 ret = rbd_dev_v2_parent_info(rbd_dev, &pii);
5774 if (pii.pool_id == CEPH_NOPOOL || !pii.has_overlap)
5775 goto out; /* No parent? No problem. */
5777 /* The ceph file layout needs to fit pool id in 32 bits */
5780 if (pii.pool_id > (u64)U32_MAX) {
5781 rbd_warn(NULL, "parent pool id too large (%llu > %u)",
5782 (unsigned long long)pii.pool_id, U32_MAX);
5787 * The parent won't change except when the clone is flattened,
5788 * so we only need to record the parent image spec once.
5790 parent_spec->pool_id = pii.pool_id;
5791 if (pii.pool_ns && *pii.pool_ns) {
5792 parent_spec->pool_ns = pii.pool_ns;
5795 parent_spec->image_id = pii.image_id;
5796 pii.image_id = NULL;
5797 parent_spec->snap_id = pii.snap_id;
5799 rbd_assert(!rbd_dev->parent_spec);
5800 rbd_dev->parent_spec = parent_spec;
5801 parent_spec = NULL; /* rbd_dev now owns this */
5804 * Record the parent overlap. If it's zero, issue a warning as
5805 * we will proceed as if there is no parent.
5808 rbd_warn(rbd_dev, "clone is standalone (overlap 0)");
5809 rbd_dev->parent_overlap = pii.overlap;
5814 rbd_parent_info_cleanup(&pii);
5815 rbd_spec_put(parent_spec);
5819 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev,
5820 u64 *stripe_unit, u64 *stripe_count)
5824 __le64 stripe_count;
5825 } __attribute__ ((packed)) striping_info_buf = { 0 };
5826 size_t size = sizeof (striping_info_buf);
5829 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
5830 &rbd_dev->header_oloc, "get_stripe_unit_count",
5831 NULL, 0, &striping_info_buf, size);
5832 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5838 *stripe_unit = le64_to_cpu(striping_info_buf.stripe_unit);
5839 *stripe_count = le64_to_cpu(striping_info_buf.stripe_count);
5840 dout(" stripe_unit = %llu stripe_count = %llu\n", *stripe_unit,
5846 static int rbd_dev_v2_data_pool(struct rbd_device *rbd_dev, s64 *data_pool_id)
5848 __le64 data_pool_buf;
5851 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
5852 &rbd_dev->header_oloc, "get_data_pool",
5853 NULL, 0, &data_pool_buf,
5854 sizeof(data_pool_buf));
5855 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5858 if (ret < sizeof(data_pool_buf))
5861 *data_pool_id = le64_to_cpu(data_pool_buf);
5862 dout(" data_pool_id = %lld\n", *data_pool_id);
5863 WARN_ON(*data_pool_id == CEPH_NOPOOL);
5868 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
5870 CEPH_DEFINE_OID_ONSTACK(oid);
5871 size_t image_id_size;
5876 void *reply_buf = NULL;
5878 char *image_name = NULL;
5881 rbd_assert(!rbd_dev->spec->image_name);
5883 len = strlen(rbd_dev->spec->image_id);
5884 image_id_size = sizeof (__le32) + len;
5885 image_id = kmalloc(image_id_size, GFP_KERNEL);
5890 end = image_id + image_id_size;
5891 ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
5893 size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
5894 reply_buf = kmalloc(size, GFP_KERNEL);
5898 ceph_oid_printf(&oid, "%s", RBD_DIRECTORY);
5899 ret = rbd_obj_method_sync(rbd_dev, &oid, &rbd_dev->header_oloc,
5900 "dir_get_name", image_id, image_id_size,
5905 end = reply_buf + ret;
5907 image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
5908 if (IS_ERR(image_name))
5911 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
5919 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5921 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
5922 const char *snap_name;
5925 /* Skip over names until we find the one we are looking for */
5927 snap_name = rbd_dev->header.snap_names;
5928 while (which < snapc->num_snaps) {
5929 if (!strcmp(name, snap_name))
5930 return snapc->snaps[which];
5931 snap_name += strlen(snap_name) + 1;
5937 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5939 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
5944 for (which = 0; !found && which < snapc->num_snaps; which++) {
5945 const char *snap_name;
5947 snap_id = snapc->snaps[which];
5948 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
5949 if (IS_ERR(snap_name)) {
5950 /* ignore no-longer existing snapshots */
5951 if (PTR_ERR(snap_name) == -ENOENT)
5956 found = !strcmp(name, snap_name);
5959 return found ? snap_id : CEPH_NOSNAP;
5963 * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
5964 * no snapshot by that name is found, or if an error occurs.
5966 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5968 if (rbd_dev->image_format == 1)
5969 return rbd_v1_snap_id_by_name(rbd_dev, name);
5971 return rbd_v2_snap_id_by_name(rbd_dev, name);
5975 * An image being mapped will have everything but the snap id.
5977 static int rbd_spec_fill_snap_id(struct rbd_device *rbd_dev)
5979 struct rbd_spec *spec = rbd_dev->spec;
5981 rbd_assert(spec->pool_id != CEPH_NOPOOL && spec->pool_name);
5982 rbd_assert(spec->image_id && spec->image_name);
5983 rbd_assert(spec->snap_name);
5985 if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
5988 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
5989 if (snap_id == CEPH_NOSNAP)
5992 spec->snap_id = snap_id;
5994 spec->snap_id = CEPH_NOSNAP;
6001 * A parent image will have all ids but none of the names.
6003 * All names in an rbd spec are dynamically allocated. It's OK if we
6004 * can't figure out the name for an image id.
6006 static int rbd_spec_fill_names(struct rbd_device *rbd_dev)
6008 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
6009 struct rbd_spec *spec = rbd_dev->spec;
6010 const char *pool_name;
6011 const char *image_name;
6012 const char *snap_name;
6015 rbd_assert(spec->pool_id != CEPH_NOPOOL);
6016 rbd_assert(spec->image_id);
6017 rbd_assert(spec->snap_id != CEPH_NOSNAP);
6019 /* Get the pool name; we have to make our own copy of this */
6021 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
6023 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
6026 pool_name = kstrdup(pool_name, GFP_KERNEL);
6030 /* Fetch the image name; tolerate failure here */
6032 image_name = rbd_dev_image_name(rbd_dev);
6034 rbd_warn(rbd_dev, "unable to get image name");
6036 /* Fetch the snapshot name */
6038 snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
6039 if (IS_ERR(snap_name)) {
6040 ret = PTR_ERR(snap_name);
6044 spec->pool_name = pool_name;
6045 spec->image_name = image_name;
6046 spec->snap_name = snap_name;
6056 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev,
6057 struct ceph_snap_context **psnapc)
6066 struct ceph_snap_context *snapc;
6070 * We'll need room for the seq value (maximum snapshot id),
6071 * snapshot count, and array of that many snapshot ids.
6072 * For now we have a fixed upper limit on the number we're
6073 * prepared to receive.
6075 size = sizeof (__le64) + sizeof (__le32) +
6076 RBD_MAX_SNAP_COUNT * sizeof (__le64);
6077 reply_buf = kzalloc(size, GFP_KERNEL);
6081 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
6082 &rbd_dev->header_oloc, "get_snapcontext",
6083 NULL, 0, reply_buf, size);
6084 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
6089 end = reply_buf + ret;
6091 ceph_decode_64_safe(&p, end, seq, out);
6092 ceph_decode_32_safe(&p, end, snap_count, out);
6095 * Make sure the reported number of snapshot ids wouldn't go
6096 * beyond the end of our buffer. But before checking that,
6097 * make sure the computed size of the snapshot context we
6098 * allocate is representable in a size_t.
6100 if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
6105 if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
6109 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
6115 for (i = 0; i < snap_count; i++)
6116 snapc->snaps[i] = ceph_decode_64(&p);
6119 dout(" snap context seq = %llu, snap_count = %u\n",
6120 (unsigned long long)seq, (unsigned int)snap_count);
6127 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
6138 size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
6139 reply_buf = kmalloc(size, GFP_KERNEL);
6141 return ERR_PTR(-ENOMEM);
6143 snapid = cpu_to_le64(snap_id);
6144 ret = rbd_obj_method_sync(rbd_dev, &rbd_dev->header_oid,
6145 &rbd_dev->header_oloc, "get_snapshot_name",
6146 &snapid, sizeof(snapid), reply_buf, size);
6147 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
6149 snap_name = ERR_PTR(ret);
6154 end = reply_buf + ret;
6155 snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
6156 if (IS_ERR(snap_name))
6159 dout(" snap_id 0x%016llx snap_name = %s\n",
6160 (unsigned long long)snap_id, snap_name);
6167 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev,
6168 struct rbd_image_header *header,
6173 ret = _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
6174 first_time ? &header->obj_order : NULL,
6175 &header->image_size);
6180 ret = rbd_dev_v2_header_onetime(rbd_dev, header);
6185 ret = rbd_dev_v2_snap_context(rbd_dev, &header->snapc);
6192 static int rbd_dev_header_info(struct rbd_device *rbd_dev,
6193 struct rbd_image_header *header,
6196 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
6197 rbd_assert(!header->object_prefix && !header->snapc);
6199 if (rbd_dev->image_format == 1)
6200 return rbd_dev_v1_header_info(rbd_dev, header, first_time);
6202 return rbd_dev_v2_header_info(rbd_dev, header, first_time);
6206 * Skips over white space at *buf, and updates *buf to point to the
6207 * first found non-space character (if any). Returns the length of
6208 * the token (string of non-white space characters) found. Note
6209 * that *buf must be terminated with '\0'.
6211 static inline size_t next_token(const char **buf)
6214 * These are the characters that produce nonzero for
6215 * isspace() in the "C" and "POSIX" locales.
6217 static const char spaces[] = " \f\n\r\t\v";
6219 *buf += strspn(*buf, spaces); /* Find start of token */
6221 return strcspn(*buf, spaces); /* Return token length */
6225 * Finds the next token in *buf, dynamically allocates a buffer big
6226 * enough to hold a copy of it, and copies the token into the new
6227 * buffer. The copy is guaranteed to be terminated with '\0'. Note
6228 * that a duplicate buffer is created even for a zero-length token.
6230 * Returns a pointer to the newly-allocated duplicate, or a null
6231 * pointer if memory for the duplicate was not available. If
6232 * the lenp argument is a non-null pointer, the length of the token
6233 * (not including the '\0') is returned in *lenp.
6235 * If successful, the *buf pointer will be updated to point beyond
6236 * the end of the found token.
6238 * Note: uses GFP_KERNEL for allocation.
6240 static inline char *dup_token(const char **buf, size_t *lenp)
6245 len = next_token(buf);
6246 dup = kmemdup(*buf, len + 1, GFP_KERNEL);
6249 *(dup + len) = '\0';
6258 static int rbd_parse_param(struct fs_parameter *param,
6259 struct rbd_parse_opts_ctx *pctx)
6261 struct rbd_options *opt = pctx->opts;
6262 struct fs_parse_result result;
6263 struct p_log log = {.prefix = "rbd"};
6266 ret = ceph_parse_param(param, pctx->copts, NULL);
6267 if (ret != -ENOPARAM)
6270 token = __fs_parse(&log, rbd_parameters, param, &result);
6271 dout("%s fs_parse '%s' token %d\n", __func__, param->key, token);
6273 if (token == -ENOPARAM)
6274 return inval_plog(&log, "Unknown parameter '%s'",
6280 case Opt_queue_depth:
6281 if (result.uint_32 < 1)
6283 opt->queue_depth = result.uint_32;
6285 case Opt_alloc_size:
6286 if (result.uint_32 < SECTOR_SIZE)
6288 if (!is_power_of_2(result.uint_32))
6289 return inval_plog(&log, "alloc_size must be a power of 2");
6290 opt->alloc_size = result.uint_32;
6292 case Opt_lock_timeout:
6293 /* 0 is "wait forever" (i.e. infinite timeout) */
6294 if (result.uint_32 > INT_MAX / 1000)
6296 opt->lock_timeout = msecs_to_jiffies(result.uint_32 * 1000);
6299 kfree(pctx->spec->pool_ns);
6300 pctx->spec->pool_ns = param->string;
6301 param->string = NULL;
6303 case Opt_compression_hint:
6304 switch (result.uint_32) {
6305 case Opt_compression_hint_none:
6306 opt->alloc_hint_flags &=
6307 ~(CEPH_OSD_ALLOC_HINT_FLAG_COMPRESSIBLE |
6308 CEPH_OSD_ALLOC_HINT_FLAG_INCOMPRESSIBLE);
6310 case Opt_compression_hint_compressible:
6311 opt->alloc_hint_flags |=
6312 CEPH_OSD_ALLOC_HINT_FLAG_COMPRESSIBLE;
6313 opt->alloc_hint_flags &=
6314 ~CEPH_OSD_ALLOC_HINT_FLAG_INCOMPRESSIBLE;
6316 case Opt_compression_hint_incompressible:
6317 opt->alloc_hint_flags |=
6318 CEPH_OSD_ALLOC_HINT_FLAG_INCOMPRESSIBLE;
6319 opt->alloc_hint_flags &=
6320 ~CEPH_OSD_ALLOC_HINT_FLAG_COMPRESSIBLE;
6327 opt->read_only = true;
6329 case Opt_read_write:
6330 opt->read_only = false;
6332 case Opt_lock_on_read:
6333 opt->lock_on_read = true;
6336 opt->exclusive = true;
6348 return inval_plog(&log, "%s out of range", param->key);
6352 * This duplicates most of generic_parse_monolithic(), untying it from
6353 * fs_context and skipping standard superblock and security options.
6355 static int rbd_parse_options(char *options, struct rbd_parse_opts_ctx *pctx)
6360 dout("%s '%s'\n", __func__, options);
6361 while ((key = strsep(&options, ",")) != NULL) {
6363 struct fs_parameter param = {
6365 .type = fs_value_is_flag,
6367 char *value = strchr(key, '=');
6374 v_len = strlen(value);
6375 param.string = kmemdup_nul(value, v_len,
6379 param.type = fs_value_is_string;
6383 ret = rbd_parse_param(¶m, pctx);
6384 kfree(param.string);
6394 * Parse the options provided for an "rbd add" (i.e., rbd image
6395 * mapping) request. These arrive via a write to /sys/bus/rbd/add,
6396 * and the data written is passed here via a NUL-terminated buffer.
6397 * Returns 0 if successful or an error code otherwise.
6399 * The information extracted from these options is recorded in
6400 * the other parameters which return dynamically-allocated
6403 * The address of a pointer that will refer to a ceph options
6404 * structure. Caller must release the returned pointer using
6405 * ceph_destroy_options() when it is no longer needed.
6407 * Address of an rbd options pointer. Fully initialized by
6408 * this function; caller must release with kfree().
6410 * Address of an rbd image specification pointer. Fully
6411 * initialized by this function based on parsed options.
6412 * Caller must release with rbd_spec_put().
6414 * The options passed take this form:
6415 * <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
6418 * A comma-separated list of one or more monitor addresses.
6419 * A monitor address is an ip address, optionally followed
6420 * by a port number (separated by a colon).
6421 * I.e.: ip1[:port1][,ip2[:port2]...]
6423 * A comma-separated list of ceph and/or rbd options.
6425 * The name of the rados pool containing the rbd image.
6427 * The name of the image in that pool to map.
6429 * An optional snapshot id. If provided, the mapping will
6430 * present data from the image at the time that snapshot was
6431 * created. The image head is used if no snapshot id is
6432 * provided. Snapshot mappings are always read-only.
6434 static int rbd_add_parse_args(const char *buf,
6435 struct ceph_options **ceph_opts,
6436 struct rbd_options **opts,
6437 struct rbd_spec **rbd_spec)
6441 const char *mon_addrs;
6443 size_t mon_addrs_size;
6444 struct rbd_parse_opts_ctx pctx = { 0 };
6447 /* The first four tokens are required */
6449 len = next_token(&buf);
6451 rbd_warn(NULL, "no monitor address(es) provided");
6455 mon_addrs_size = len;
6459 options = dup_token(&buf, NULL);
6463 rbd_warn(NULL, "no options provided");
6467 pctx.spec = rbd_spec_alloc();
6471 pctx.spec->pool_name = dup_token(&buf, NULL);
6472 if (!pctx.spec->pool_name)
6474 if (!*pctx.spec->pool_name) {
6475 rbd_warn(NULL, "no pool name provided");
6479 pctx.spec->image_name = dup_token(&buf, NULL);
6480 if (!pctx.spec->image_name)
6482 if (!*pctx.spec->image_name) {
6483 rbd_warn(NULL, "no image name provided");
6488 * Snapshot name is optional; default is to use "-"
6489 * (indicating the head/no snapshot).
6491 len = next_token(&buf);
6493 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
6494 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
6495 } else if (len > RBD_MAX_SNAP_NAME_LEN) {
6496 ret = -ENAMETOOLONG;
6499 snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
6502 *(snap_name + len) = '\0';
6503 pctx.spec->snap_name = snap_name;
6505 pctx.copts = ceph_alloc_options();
6509 /* Initialize all rbd options to the defaults */
6511 pctx.opts = kzalloc(sizeof(*pctx.opts), GFP_KERNEL);
6515 pctx.opts->read_only = RBD_READ_ONLY_DEFAULT;
6516 pctx.opts->queue_depth = RBD_QUEUE_DEPTH_DEFAULT;
6517 pctx.opts->alloc_size = RBD_ALLOC_SIZE_DEFAULT;
6518 pctx.opts->lock_timeout = RBD_LOCK_TIMEOUT_DEFAULT;
6519 pctx.opts->lock_on_read = RBD_LOCK_ON_READ_DEFAULT;
6520 pctx.opts->exclusive = RBD_EXCLUSIVE_DEFAULT;
6521 pctx.opts->trim = RBD_TRIM_DEFAULT;
6523 ret = ceph_parse_mon_ips(mon_addrs, mon_addrs_size, pctx.copts, NULL,
6528 ret = rbd_parse_options(options, &pctx);
6532 *ceph_opts = pctx.copts;
6534 *rbd_spec = pctx.spec;
6542 ceph_destroy_options(pctx.copts);
6543 rbd_spec_put(pctx.spec);
6548 static void rbd_dev_image_unlock(struct rbd_device *rbd_dev)
6550 down_write(&rbd_dev->lock_rwsem);
6551 if (__rbd_is_lock_owner(rbd_dev))
6552 __rbd_release_lock(rbd_dev);
6553 up_write(&rbd_dev->lock_rwsem);
6557 * If the wait is interrupted, an error is returned even if the lock
6558 * was successfully acquired. rbd_dev_image_unlock() will release it
6561 static int rbd_add_acquire_lock(struct rbd_device *rbd_dev)
6565 if (!(rbd_dev->header.features & RBD_FEATURE_EXCLUSIVE_LOCK)) {
6566 if (!rbd_dev->opts->exclusive && !rbd_dev->opts->lock_on_read)
6569 rbd_warn(rbd_dev, "exclusive-lock feature is not enabled");
6573 if (rbd_is_ro(rbd_dev))
6576 rbd_assert(!rbd_is_lock_owner(rbd_dev));
6577 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork, 0);
6578 ret = wait_for_completion_killable_timeout(&rbd_dev->acquire_wait,
6579 ceph_timeout_jiffies(rbd_dev->opts->lock_timeout));
6581 ret = rbd_dev->acquire_err;
6583 cancel_delayed_work_sync(&rbd_dev->lock_dwork);
6587 rbd_warn(rbd_dev, "failed to acquire lock: %ld", ret);
6593 * The lock may have been released by now, unless automatic lock
6594 * transitions are disabled.
6596 rbd_assert(!rbd_dev->opts->exclusive || rbd_is_lock_owner(rbd_dev));
6601 * An rbd format 2 image has a unique identifier, distinct from the
6602 * name given to it by the user. Internally, that identifier is
6603 * what's used to specify the names of objects related to the image.
6605 * A special "rbd id" object is used to map an rbd image name to its
6606 * id. If that object doesn't exist, then there is no v2 rbd image
6607 * with the supplied name.
6609 * This function will record the given rbd_dev's image_id field if
6610 * it can be determined, and in that case will return 0. If any
6611 * errors occur a negative errno will be returned and the rbd_dev's
6612 * image_id field will be unchanged (and should be NULL).
6614 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
6618 CEPH_DEFINE_OID_ONSTACK(oid);
6623 * When probing a parent image, the image id is already
6624 * known (and the image name likely is not). There's no
6625 * need to fetch the image id again in this case. We
6626 * do still need to set the image format though.
6628 if (rbd_dev->spec->image_id) {
6629 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
6635 * First, see if the format 2 image id file exists, and if
6636 * so, get the image's persistent id from it.
6638 ret = ceph_oid_aprintf(&oid, GFP_KERNEL, "%s%s", RBD_ID_PREFIX,
6639 rbd_dev->spec->image_name);
6643 dout("rbd id object name is %s\n", oid.name);
6645 /* Response will be an encoded string, which includes a length */
6646 size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
6647 response = kzalloc(size, GFP_NOIO);
6653 /* If it doesn't exist we'll assume it's a format 1 image */
6655 ret = rbd_obj_method_sync(rbd_dev, &oid, &rbd_dev->header_oloc,
6658 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
6659 if (ret == -ENOENT) {
6660 image_id = kstrdup("", GFP_KERNEL);
6661 ret = image_id ? 0 : -ENOMEM;
6663 rbd_dev->image_format = 1;
6664 } else if (ret >= 0) {
6667 image_id = ceph_extract_encoded_string(&p, p + ret,
6669 ret = PTR_ERR_OR_ZERO(image_id);
6671 rbd_dev->image_format = 2;
6675 rbd_dev->spec->image_id = image_id;
6676 dout("image_id is %s\n", image_id);
6680 ceph_oid_destroy(&oid);
6685 * Undo whatever state changes are made by v1 or v2 header info
6688 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
6690 rbd_dev_parent_put(rbd_dev);
6691 rbd_object_map_free(rbd_dev);
6692 rbd_dev_mapping_clear(rbd_dev);
6694 /* Free dynamic fields from the header, then zero it out */
6696 rbd_image_header_cleanup(&rbd_dev->header);
6699 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev,
6700 struct rbd_image_header *header)
6704 ret = rbd_dev_v2_object_prefix(rbd_dev, &header->object_prefix);
6709 * Get the and check features for the image. Currently the
6710 * features are assumed to never change.
6712 ret = _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
6713 rbd_is_ro(rbd_dev), &header->features);
6717 /* If the image supports fancy striping, get its parameters */
6719 if (header->features & RBD_FEATURE_STRIPINGV2) {
6720 ret = rbd_dev_v2_striping_info(rbd_dev, &header->stripe_unit,
6721 &header->stripe_count);
6726 if (header->features & RBD_FEATURE_DATA_POOL) {
6727 ret = rbd_dev_v2_data_pool(rbd_dev, &header->data_pool_id);
6736 * @depth is rbd_dev_image_probe() -> rbd_dev_probe_parent() ->
6737 * rbd_dev_image_probe() recursion depth, which means it's also the
6738 * length of the already discovered part of the parent chain.
6740 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev, int depth)
6742 struct rbd_device *parent = NULL;
6745 if (!rbd_dev->parent_spec)
6748 if (++depth > RBD_MAX_PARENT_CHAIN_LEN) {
6749 pr_info("parent chain is too long (%d)\n", depth);
6754 parent = __rbd_dev_create(rbd_dev->parent_spec);
6761 * Images related by parent/child relationships always share
6762 * rbd_client and spec/parent_spec, so bump their refcounts.
6764 parent->rbd_client = __rbd_get_client(rbd_dev->rbd_client);
6765 parent->spec = rbd_spec_get(rbd_dev->parent_spec);
6767 __set_bit(RBD_DEV_FLAG_READONLY, &parent->flags);
6769 ret = rbd_dev_image_probe(parent, depth);
6773 rbd_dev->parent = parent;
6774 atomic_set(&rbd_dev->parent_ref, 1);
6778 rbd_dev_unparent(rbd_dev);
6779 rbd_dev_destroy(parent);
6783 static void rbd_dev_device_release(struct rbd_device *rbd_dev)
6785 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
6786 rbd_free_disk(rbd_dev);
6788 unregister_blkdev(rbd_dev->major, rbd_dev->name);
6792 * rbd_dev->header_rwsem must be locked for write and will be unlocked
6795 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
6799 /* Record our major and minor device numbers. */
6801 if (!single_major) {
6802 ret = register_blkdev(0, rbd_dev->name);
6804 goto err_out_unlock;
6806 rbd_dev->major = ret;
6809 rbd_dev->major = rbd_major;
6810 rbd_dev->minor = rbd_dev_id_to_minor(rbd_dev->dev_id);
6813 /* Set up the blkdev mapping. */
6815 ret = rbd_init_disk(rbd_dev);
6817 goto err_out_blkdev;
6819 set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
6820 set_disk_ro(rbd_dev->disk, rbd_is_ro(rbd_dev));
6822 ret = dev_set_name(&rbd_dev->dev, "%d", rbd_dev->dev_id);
6826 set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
6827 up_write(&rbd_dev->header_rwsem);
6831 rbd_free_disk(rbd_dev);
6834 unregister_blkdev(rbd_dev->major, rbd_dev->name);
6836 up_write(&rbd_dev->header_rwsem);
6840 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
6842 struct rbd_spec *spec = rbd_dev->spec;
6845 /* Record the header object name for this rbd image. */
6847 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
6848 if (rbd_dev->image_format == 1)
6849 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
6850 spec->image_name, RBD_SUFFIX);
6852 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
6853 RBD_HEADER_PREFIX, spec->image_id);
6858 static void rbd_print_dne(struct rbd_device *rbd_dev, bool is_snap)
6861 pr_info("image %s/%s%s%s does not exist\n",
6862 rbd_dev->spec->pool_name,
6863 rbd_dev->spec->pool_ns ?: "",
6864 rbd_dev->spec->pool_ns ? "/" : "",
6865 rbd_dev->spec->image_name);
6867 pr_info("snap %s/%s%s%s@%s does not exist\n",
6868 rbd_dev->spec->pool_name,
6869 rbd_dev->spec->pool_ns ?: "",
6870 rbd_dev->spec->pool_ns ? "/" : "",
6871 rbd_dev->spec->image_name,
6872 rbd_dev->spec->snap_name);
6876 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
6878 if (!rbd_is_ro(rbd_dev))
6879 rbd_unregister_watch(rbd_dev);
6881 rbd_dev_unprobe(rbd_dev);
6882 rbd_dev->image_format = 0;
6883 kfree(rbd_dev->spec->image_id);
6884 rbd_dev->spec->image_id = NULL;
6888 * Probe for the existence of the header object for the given rbd
6889 * device. If this image is the one being mapped (i.e., not a
6890 * parent), initiate a watch on its header object before using that
6891 * object to get detailed information about the rbd image.
6893 * On success, returns with header_rwsem held for write if called
6896 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth)
6898 bool need_watch = !rbd_is_ro(rbd_dev);
6902 * Get the id from the image id object. Unless there's an
6903 * error, rbd_dev->spec->image_id will be filled in with
6904 * a dynamically-allocated string, and rbd_dev->image_format
6905 * will be set to either 1 or 2.
6907 ret = rbd_dev_image_id(rbd_dev);
6911 ret = rbd_dev_header_name(rbd_dev);
6913 goto err_out_format;
6916 ret = rbd_register_watch(rbd_dev);
6919 rbd_print_dne(rbd_dev, false);
6920 goto err_out_format;
6925 down_write(&rbd_dev->header_rwsem);
6927 ret = rbd_dev_header_info(rbd_dev, &rbd_dev->header, true);
6929 if (ret == -ENOENT && !need_watch)
6930 rbd_print_dne(rbd_dev, false);
6934 rbd_init_layout(rbd_dev);
6937 * If this image is the one being mapped, we have pool name and
6938 * id, image name and id, and snap name - need to fill snap id.
6939 * Otherwise this is a parent image, identified by pool, image
6940 * and snap ids - need to fill in names for those ids.
6943 ret = rbd_spec_fill_snap_id(rbd_dev);
6945 ret = rbd_spec_fill_names(rbd_dev);
6948 rbd_print_dne(rbd_dev, true);
6952 ret = rbd_dev_mapping_set(rbd_dev);
6956 if (rbd_is_snap(rbd_dev) &&
6957 (rbd_dev->header.features & RBD_FEATURE_OBJECT_MAP)) {
6958 ret = rbd_object_map_load(rbd_dev);
6963 if (rbd_dev->header.features & RBD_FEATURE_LAYERING) {
6964 ret = rbd_dev_setup_parent(rbd_dev);
6969 ret = rbd_dev_probe_parent(rbd_dev, depth);
6973 dout("discovered format %u image, header name is %s\n",
6974 rbd_dev->image_format, rbd_dev->header_oid.name);
6979 up_write(&rbd_dev->header_rwsem);
6981 rbd_unregister_watch(rbd_dev);
6982 rbd_dev_unprobe(rbd_dev);
6984 rbd_dev->image_format = 0;
6985 kfree(rbd_dev->spec->image_id);
6986 rbd_dev->spec->image_id = NULL;
6990 static void rbd_dev_update_header(struct rbd_device *rbd_dev,
6991 struct rbd_image_header *header)
6993 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
6994 rbd_assert(rbd_dev->header.object_prefix); /* !first_time */
6996 if (rbd_dev->header.image_size != header->image_size) {
6997 rbd_dev->header.image_size = header->image_size;
6999 if (!rbd_is_snap(rbd_dev)) {
7000 rbd_dev->mapping.size = header->image_size;
7001 rbd_dev_update_size(rbd_dev);
7005 ceph_put_snap_context(rbd_dev->header.snapc);
7006 rbd_dev->header.snapc = header->snapc;
7007 header->snapc = NULL;
7009 if (rbd_dev->image_format == 1) {
7010 kfree(rbd_dev->header.snap_names);
7011 rbd_dev->header.snap_names = header->snap_names;
7012 header->snap_names = NULL;
7014 kfree(rbd_dev->header.snap_sizes);
7015 rbd_dev->header.snap_sizes = header->snap_sizes;
7016 header->snap_sizes = NULL;
7020 static void rbd_dev_update_parent(struct rbd_device *rbd_dev,
7021 struct parent_image_info *pii)
7023 if (pii->pool_id == CEPH_NOPOOL || !pii->has_overlap) {
7025 * Either the parent never existed, or we have
7026 * record of it but the image got flattened so it no
7027 * longer has a parent. When the parent of a
7028 * layered image disappears we immediately set the
7029 * overlap to 0. The effect of this is that all new
7030 * requests will be treated as if the image had no
7033 * If !pii.has_overlap, the parent image spec is not
7034 * applicable. It's there to avoid duplication in each
7037 if (rbd_dev->parent_overlap) {
7038 rbd_dev->parent_overlap = 0;
7039 rbd_dev_parent_put(rbd_dev);
7040 pr_info("%s: clone has been flattened\n",
7041 rbd_dev->disk->disk_name);
7044 rbd_assert(rbd_dev->parent_spec);
7047 * Update the parent overlap. If it became zero, issue
7048 * a warning as we will proceed as if there is no parent.
7050 if (!pii->overlap && rbd_dev->parent_overlap)
7052 "clone has become standalone (overlap 0)");
7053 rbd_dev->parent_overlap = pii->overlap;
7057 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
7059 struct rbd_image_header header = { 0 };
7060 struct parent_image_info pii = { 0 };
7063 dout("%s rbd_dev %p\n", __func__, rbd_dev);
7065 ret = rbd_dev_header_info(rbd_dev, &header, false);
7070 * If there is a parent, see if it has disappeared due to the
7071 * mapped image getting flattened.
7073 if (rbd_dev->parent) {
7074 ret = rbd_dev_v2_parent_info(rbd_dev, &pii);
7079 down_write(&rbd_dev->header_rwsem);
7080 rbd_dev_update_header(rbd_dev, &header);
7081 if (rbd_dev->parent)
7082 rbd_dev_update_parent(rbd_dev, &pii);
7083 up_write(&rbd_dev->header_rwsem);
7086 rbd_parent_info_cleanup(&pii);
7087 rbd_image_header_cleanup(&header);
7091 static ssize_t do_rbd_add(const char *buf, size_t count)
7093 struct rbd_device *rbd_dev = NULL;
7094 struct ceph_options *ceph_opts = NULL;
7095 struct rbd_options *rbd_opts = NULL;
7096 struct rbd_spec *spec = NULL;
7097 struct rbd_client *rbdc;
7100 if (!capable(CAP_SYS_ADMIN))
7103 if (!try_module_get(THIS_MODULE))
7106 /* parse add command */
7107 rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
7111 rbdc = rbd_get_client(ceph_opts);
7118 rc = ceph_pg_poolid_by_name(rbdc->client->osdc.osdmap, spec->pool_name);
7121 pr_info("pool %s does not exist\n", spec->pool_name);
7122 goto err_out_client;
7124 spec->pool_id = (u64)rc;
7126 rbd_dev = rbd_dev_create(rbdc, spec, rbd_opts);
7129 goto err_out_client;
7131 rbdc = NULL; /* rbd_dev now owns this */
7132 spec = NULL; /* rbd_dev now owns this */
7133 rbd_opts = NULL; /* rbd_dev now owns this */
7135 /* if we are mapping a snapshot it will be a read-only mapping */
7136 if (rbd_dev->opts->read_only ||
7137 strcmp(rbd_dev->spec->snap_name, RBD_SNAP_HEAD_NAME))
7138 __set_bit(RBD_DEV_FLAG_READONLY, &rbd_dev->flags);
7140 rbd_dev->config_info = kstrdup(buf, GFP_KERNEL);
7141 if (!rbd_dev->config_info) {
7143 goto err_out_rbd_dev;
7146 rc = rbd_dev_image_probe(rbd_dev, 0);
7148 goto err_out_rbd_dev;
7150 if (rbd_dev->opts->alloc_size > rbd_dev->layout.object_size) {
7151 rbd_warn(rbd_dev, "alloc_size adjusted to %u",
7152 rbd_dev->layout.object_size);
7153 rbd_dev->opts->alloc_size = rbd_dev->layout.object_size;
7156 rc = rbd_dev_device_setup(rbd_dev);
7158 goto err_out_image_probe;
7160 rc = rbd_add_acquire_lock(rbd_dev);
7162 goto err_out_image_lock;
7164 /* Everything's ready. Announce the disk to the world. */
7166 rc = device_add(&rbd_dev->dev);
7168 goto err_out_image_lock;
7170 rc = device_add_disk(&rbd_dev->dev, rbd_dev->disk, NULL);
7172 goto err_out_cleanup_disk;
7174 spin_lock(&rbd_dev_list_lock);
7175 list_add_tail(&rbd_dev->node, &rbd_dev_list);
7176 spin_unlock(&rbd_dev_list_lock);
7178 pr_info("%s: capacity %llu features 0x%llx\n", rbd_dev->disk->disk_name,
7179 (unsigned long long)get_capacity(rbd_dev->disk) << SECTOR_SHIFT,
7180 rbd_dev->header.features);
7183 module_put(THIS_MODULE);
7186 err_out_cleanup_disk:
7187 rbd_free_disk(rbd_dev);
7189 rbd_dev_image_unlock(rbd_dev);
7190 rbd_dev_device_release(rbd_dev);
7191 err_out_image_probe:
7192 rbd_dev_image_release(rbd_dev);
7194 rbd_dev_destroy(rbd_dev);
7196 rbd_put_client(rbdc);
7203 static ssize_t add_store(const struct bus_type *bus, const char *buf, size_t count)
7208 return do_rbd_add(buf, count);
7211 static ssize_t add_single_major_store(const struct bus_type *bus, const char *buf,
7214 return do_rbd_add(buf, count);
7217 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
7219 while (rbd_dev->parent) {
7220 struct rbd_device *first = rbd_dev;
7221 struct rbd_device *second = first->parent;
7222 struct rbd_device *third;
7225 * Follow to the parent with no grandparent and
7228 while (second && (third = second->parent)) {
7233 rbd_dev_image_release(second);
7234 rbd_dev_destroy(second);
7235 first->parent = NULL;
7236 first->parent_overlap = 0;
7238 rbd_assert(first->parent_spec);
7239 rbd_spec_put(first->parent_spec);
7240 first->parent_spec = NULL;
7244 static ssize_t do_rbd_remove(const char *buf, size_t count)
7246 struct rbd_device *rbd_dev = NULL;
7252 if (!capable(CAP_SYS_ADMIN))
7257 sscanf(buf, "%d %5s", &dev_id, opt_buf);
7259 pr_err("dev_id out of range\n");
7262 if (opt_buf[0] != '\0') {
7263 if (!strcmp(opt_buf, "force")) {
7266 pr_err("bad remove option at '%s'\n", opt_buf);
7272 spin_lock(&rbd_dev_list_lock);
7273 list_for_each_entry(rbd_dev, &rbd_dev_list, node) {
7274 if (rbd_dev->dev_id == dev_id) {
7280 spin_lock_irq(&rbd_dev->lock);
7281 if (rbd_dev->open_count && !force)
7283 else if (test_and_set_bit(RBD_DEV_FLAG_REMOVING,
7286 spin_unlock_irq(&rbd_dev->lock);
7288 spin_unlock(&rbd_dev_list_lock);
7294 * Prevent new IO from being queued and wait for existing
7295 * IO to complete/fail.
7297 blk_mq_freeze_queue(rbd_dev->disk->queue);
7298 blk_mark_disk_dead(rbd_dev->disk);
7301 del_gendisk(rbd_dev->disk);
7302 spin_lock(&rbd_dev_list_lock);
7303 list_del_init(&rbd_dev->node);
7304 spin_unlock(&rbd_dev_list_lock);
7305 device_del(&rbd_dev->dev);
7307 rbd_dev_image_unlock(rbd_dev);
7308 rbd_dev_device_release(rbd_dev);
7309 rbd_dev_image_release(rbd_dev);
7310 rbd_dev_destroy(rbd_dev);
7314 static ssize_t remove_store(const struct bus_type *bus, const char *buf, size_t count)
7319 return do_rbd_remove(buf, count);
7322 static ssize_t remove_single_major_store(const struct bus_type *bus, const char *buf,
7325 return do_rbd_remove(buf, count);
7329 * create control files in sysfs
7332 static int __init rbd_sysfs_init(void)
7336 ret = device_register(&rbd_root_dev);
7338 put_device(&rbd_root_dev);
7342 ret = bus_register(&rbd_bus_type);
7344 device_unregister(&rbd_root_dev);
7349 static void __exit rbd_sysfs_cleanup(void)
7351 bus_unregister(&rbd_bus_type);
7352 device_unregister(&rbd_root_dev);
7355 static int __init rbd_slab_init(void)
7357 rbd_assert(!rbd_img_request_cache);
7358 rbd_img_request_cache = KMEM_CACHE(rbd_img_request, 0);
7359 if (!rbd_img_request_cache)
7362 rbd_assert(!rbd_obj_request_cache);
7363 rbd_obj_request_cache = KMEM_CACHE(rbd_obj_request, 0);
7364 if (!rbd_obj_request_cache)
7370 kmem_cache_destroy(rbd_img_request_cache);
7371 rbd_img_request_cache = NULL;
7375 static void rbd_slab_exit(void)
7377 rbd_assert(rbd_obj_request_cache);
7378 kmem_cache_destroy(rbd_obj_request_cache);
7379 rbd_obj_request_cache = NULL;
7381 rbd_assert(rbd_img_request_cache);
7382 kmem_cache_destroy(rbd_img_request_cache);
7383 rbd_img_request_cache = NULL;
7386 static int __init rbd_init(void)
7390 if (!libceph_compatible(NULL)) {
7391 rbd_warn(NULL, "libceph incompatibility (quitting)");
7395 rc = rbd_slab_init();
7400 * The number of active work items is limited by the number of
7401 * rbd devices * queue depth, so leave @max_active at default.
7403 rbd_wq = alloc_workqueue(RBD_DRV_NAME, WQ_MEM_RECLAIM, 0);
7410 rbd_major = register_blkdev(0, RBD_DRV_NAME);
7411 if (rbd_major < 0) {
7417 rc = rbd_sysfs_init();
7419 goto err_out_blkdev;
7422 pr_info("loaded (major %d)\n", rbd_major);
7424 pr_info("loaded\n");
7430 unregister_blkdev(rbd_major, RBD_DRV_NAME);
7432 destroy_workqueue(rbd_wq);
7438 static void __exit rbd_exit(void)
7440 ida_destroy(&rbd_dev_id_ida);
7441 rbd_sysfs_cleanup();
7443 unregister_blkdev(rbd_major, RBD_DRV_NAME);
7444 destroy_workqueue(rbd_wq);
7448 module_init(rbd_init);
7449 module_exit(rbd_exit);
7451 MODULE_AUTHOR("Alex Elder <elder@inktank.com>");
7452 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
7453 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
7454 /* following authorship retained from original osdblk.c */
7455 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
7457 MODULE_DESCRIPTION("RADOS Block Device (RBD) driver");
7458 MODULE_LICENSE("GPL");