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/decode.h>
36 #include <linux/parser.h>
37 #include <linux/bsearch.h>
39 #include <linux/kernel.h>
40 #include <linux/device.h>
41 #include <linux/module.h>
42 #include <linux/blk-mq.h>
44 #include <linux/blkdev.h>
45 #include <linux/slab.h>
46 #include <linux/idr.h>
47 #include <linux/workqueue.h>
49 #include "rbd_types.h"
51 #define RBD_DEBUG /* Activate rbd_assert() calls */
54 * The basic unit of block I/O is a sector. It is interpreted in a
55 * number of contexts in Linux (blk, bio, genhd), but the default is
56 * universally 512 bytes. These symbols are just slightly more
57 * meaningful than the bare numbers they represent.
59 #define SECTOR_SHIFT 9
60 #define SECTOR_SIZE (1ULL << SECTOR_SHIFT)
63 * Increment the given counter and return its updated value.
64 * If the counter is already 0 it will not be incremented.
65 * If the counter is already at its maximum value returns
66 * -EINVAL without updating it.
68 static int atomic_inc_return_safe(atomic_t *v)
72 counter = (unsigned int)__atomic_add_unless(v, 1, 0);
73 if (counter <= (unsigned int)INT_MAX)
81 /* Decrement the counter. Return the resulting value, or -EINVAL */
82 static int atomic_dec_return_safe(atomic_t *v)
86 counter = atomic_dec_return(v);
95 #define RBD_DRV_NAME "rbd"
97 #define RBD_MINORS_PER_MAJOR 256
98 #define RBD_SINGLE_MAJOR_PART_SHIFT 4
100 #define RBD_MAX_PARENT_CHAIN_LEN 16
102 #define RBD_SNAP_DEV_NAME_PREFIX "snap_"
103 #define RBD_MAX_SNAP_NAME_LEN \
104 (NAME_MAX - (sizeof (RBD_SNAP_DEV_NAME_PREFIX) - 1))
106 #define RBD_MAX_SNAP_COUNT 510 /* allows max snapc to fit in 4KB */
108 #define RBD_SNAP_HEAD_NAME "-"
110 #define BAD_SNAP_INDEX U32_MAX /* invalid index into snap array */
112 /* This allows a single page to hold an image name sent by OSD */
113 #define RBD_IMAGE_NAME_LEN_MAX (PAGE_SIZE - sizeof (__le32) - 1)
114 #define RBD_IMAGE_ID_LEN_MAX 64
116 #define RBD_OBJ_PREFIX_LEN_MAX 64
118 #define RBD_NOTIFY_TIMEOUT 5 /* seconds */
119 #define RBD_RETRY_DELAY msecs_to_jiffies(1000)
123 #define RBD_FEATURE_LAYERING (1<<0)
124 #define RBD_FEATURE_STRIPINGV2 (1<<1)
125 #define RBD_FEATURE_EXCLUSIVE_LOCK (1<<2)
126 #define RBD_FEATURES_ALL (RBD_FEATURE_LAYERING | \
127 RBD_FEATURE_STRIPINGV2 | \
128 RBD_FEATURE_EXCLUSIVE_LOCK)
130 /* Features supported by this (client software) implementation. */
132 #define RBD_FEATURES_SUPPORTED (RBD_FEATURES_ALL)
135 * An RBD device name will be "rbd#", where the "rbd" comes from
136 * RBD_DRV_NAME above, and # is a unique integer identifier.
138 #define DEV_NAME_LEN 32
141 * block device image metadata (in-memory version)
143 struct rbd_image_header {
144 /* These six fields never change for a given rbd image */
151 u64 features; /* Might be changeable someday? */
153 /* The remaining fields need to be updated occasionally */
155 struct ceph_snap_context *snapc;
156 char *snap_names; /* format 1 only */
157 u64 *snap_sizes; /* format 1 only */
161 * An rbd image specification.
163 * The tuple (pool_id, image_id, snap_id) is sufficient to uniquely
164 * identify an image. Each rbd_dev structure includes a pointer to
165 * an rbd_spec structure that encapsulates this identity.
167 * Each of the id's in an rbd_spec has an associated name. For a
168 * user-mapped image, the names are supplied and the id's associated
169 * with them are looked up. For a layered image, a parent image is
170 * defined by the tuple, and the names are looked up.
172 * An rbd_dev structure contains a parent_spec pointer which is
173 * non-null if the image it represents is a child in a layered
174 * image. This pointer will refer to the rbd_spec structure used
175 * by the parent rbd_dev for its own identity (i.e., the structure
176 * is shared between the parent and child).
178 * Since these structures are populated once, during the discovery
179 * phase of image construction, they are effectively immutable so
180 * we make no effort to synchronize access to them.
182 * Note that code herein does not assume the image name is known (it
183 * could be a null pointer).
187 const char *pool_name;
189 const char *image_id;
190 const char *image_name;
193 const char *snap_name;
199 * an instance of the client. multiple devices may share an rbd client.
202 struct ceph_client *client;
204 struct list_head node;
207 struct rbd_img_request;
208 typedef void (*rbd_img_callback_t)(struct rbd_img_request *);
210 #define BAD_WHICH U32_MAX /* Good which or bad which, which? */
212 struct rbd_obj_request;
213 typedef void (*rbd_obj_callback_t)(struct rbd_obj_request *);
215 enum obj_request_type {
216 OBJ_REQUEST_NODATA, OBJ_REQUEST_BIO, OBJ_REQUEST_PAGES
219 enum obj_operation_type {
226 OBJ_REQ_DONE, /* completion flag: not done = 0, done = 1 */
227 OBJ_REQ_IMG_DATA, /* object usage: standalone = 0, image = 1 */
228 OBJ_REQ_KNOWN, /* EXISTS flag valid: no = 0, yes = 1 */
229 OBJ_REQ_EXISTS, /* target exists: no = 0, yes = 1 */
232 struct rbd_obj_request {
233 const char *object_name;
234 u64 offset; /* object start byte */
235 u64 length; /* bytes from offset */
239 * An object request associated with an image will have its
240 * img_data flag set; a standalone object request will not.
242 * A standalone object request will have which == BAD_WHICH
243 * and a null obj_request pointer.
245 * An object request initiated in support of a layered image
246 * object (to check for its existence before a write) will
247 * have which == BAD_WHICH and a non-null obj_request pointer.
249 * Finally, an object request for rbd image data will have
250 * which != BAD_WHICH, and will have a non-null img_request
251 * pointer. The value of which will be in the range
252 * 0..(img_request->obj_request_count-1).
255 struct rbd_obj_request *obj_request; /* STAT op */
257 struct rbd_img_request *img_request;
259 /* links for img_request->obj_requests list */
260 struct list_head links;
263 u32 which; /* posn image request list */
265 enum obj_request_type type;
267 struct bio *bio_list;
273 struct page **copyup_pages;
274 u32 copyup_page_count;
276 struct ceph_osd_request *osd_req;
278 u64 xferred; /* bytes transferred */
281 rbd_obj_callback_t callback;
282 struct completion completion;
288 IMG_REQ_WRITE, /* I/O direction: read = 0, write = 1 */
289 IMG_REQ_CHILD, /* initiator: block = 0, child image = 1 */
290 IMG_REQ_LAYERED, /* ENOENT handling: normal = 0, layered = 1 */
291 IMG_REQ_DISCARD, /* discard: normal = 0, discard request = 1 */
294 struct rbd_img_request {
295 struct rbd_device *rbd_dev;
296 u64 offset; /* starting image byte offset */
297 u64 length; /* byte count from offset */
300 u64 snap_id; /* for reads */
301 struct ceph_snap_context *snapc; /* for writes */
304 struct request *rq; /* block request */
305 struct rbd_obj_request *obj_request; /* obj req initiator */
307 struct page **copyup_pages;
308 u32 copyup_page_count;
309 spinlock_t completion_lock;/* protects next_completion */
311 rbd_img_callback_t callback;
312 u64 xferred;/* aggregate bytes transferred */
313 int result; /* first nonzero obj_request result */
315 u32 obj_request_count;
316 struct list_head obj_requests; /* rbd_obj_request structs */
321 #define for_each_obj_request(ireq, oreq) \
322 list_for_each_entry(oreq, &(ireq)->obj_requests, links)
323 #define for_each_obj_request_from(ireq, oreq) \
324 list_for_each_entry_from(oreq, &(ireq)->obj_requests, links)
325 #define for_each_obj_request_safe(ireq, oreq, n) \
326 list_for_each_entry_safe_reverse(oreq, n, &(ireq)->obj_requests, links)
328 enum rbd_watch_state {
329 RBD_WATCH_STATE_UNREGISTERED,
330 RBD_WATCH_STATE_REGISTERED,
331 RBD_WATCH_STATE_ERROR,
334 enum rbd_lock_state {
335 RBD_LOCK_STATE_UNLOCKED,
336 RBD_LOCK_STATE_LOCKED,
337 RBD_LOCK_STATE_RELEASING,
340 /* WatchNotify::ClientId */
341 struct rbd_client_id {
356 int dev_id; /* blkdev unique id */
358 int major; /* blkdev assigned major */
360 struct gendisk *disk; /* blkdev's gendisk and rq */
362 u32 image_format; /* Either 1 or 2 */
363 struct rbd_client *rbd_client;
365 char name[DEV_NAME_LEN]; /* blkdev name, e.g. rbd3 */
367 spinlock_t lock; /* queue, flags, open_count */
369 struct rbd_image_header header;
370 unsigned long flags; /* possibly lock protected */
371 struct rbd_spec *spec;
372 struct rbd_options *opts;
373 char *config_info; /* add{,_single_major} string */
375 struct ceph_object_id header_oid;
376 struct ceph_object_locator header_oloc;
378 struct ceph_file_layout layout; /* used for all rbd requests */
380 struct mutex watch_mutex;
381 enum rbd_watch_state watch_state;
382 struct ceph_osd_linger_request *watch_handle;
384 struct delayed_work watch_dwork;
386 struct rw_semaphore lock_rwsem;
387 enum rbd_lock_state lock_state;
388 struct rbd_client_id owner_cid;
389 struct work_struct acquired_lock_work;
390 struct work_struct released_lock_work;
391 struct delayed_work lock_dwork;
392 struct work_struct unlock_work;
393 wait_queue_head_t lock_waitq;
395 struct workqueue_struct *task_wq;
397 struct rbd_spec *parent_spec;
400 struct rbd_device *parent;
402 /* Block layer tags. */
403 struct blk_mq_tag_set tag_set;
405 /* protects updating the header */
406 struct rw_semaphore header_rwsem;
408 struct rbd_mapping mapping;
410 struct list_head node;
414 unsigned long open_count; /* protected by lock */
418 * Flag bits for rbd_dev->flags. If atomicity is required,
419 * rbd_dev->lock is used to protect access.
421 * Currently, only the "removing" flag (which is coupled with the
422 * "open_count" field) requires atomic access.
425 RBD_DEV_FLAG_EXISTS, /* mapped snapshot has not been deleted */
426 RBD_DEV_FLAG_REMOVING, /* this mapping is being removed */
429 static DEFINE_MUTEX(client_mutex); /* Serialize client creation */
431 static LIST_HEAD(rbd_dev_list); /* devices */
432 static DEFINE_SPINLOCK(rbd_dev_list_lock);
434 static LIST_HEAD(rbd_client_list); /* clients */
435 static DEFINE_SPINLOCK(rbd_client_list_lock);
437 /* Slab caches for frequently-allocated structures */
439 static struct kmem_cache *rbd_img_request_cache;
440 static struct kmem_cache *rbd_obj_request_cache;
441 static struct kmem_cache *rbd_segment_name_cache;
443 static int rbd_major;
444 static DEFINE_IDA(rbd_dev_id_ida);
446 static struct workqueue_struct *rbd_wq;
449 * Default to false for now, as single-major requires >= 0.75 version of
450 * userspace rbd utility.
452 static bool single_major = false;
453 module_param(single_major, bool, S_IRUGO);
454 MODULE_PARM_DESC(single_major, "Use a single major number for all rbd devices (default: false)");
456 static int rbd_img_request_submit(struct rbd_img_request *img_request);
458 static ssize_t rbd_add(struct bus_type *bus, const char *buf,
460 static ssize_t rbd_remove(struct bus_type *bus, const char *buf,
462 static ssize_t rbd_add_single_major(struct bus_type *bus, const char *buf,
464 static ssize_t rbd_remove_single_major(struct bus_type *bus, const char *buf,
466 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth);
467 static void rbd_spec_put(struct rbd_spec *spec);
469 static int rbd_dev_id_to_minor(int dev_id)
471 return dev_id << RBD_SINGLE_MAJOR_PART_SHIFT;
474 static int minor_to_rbd_dev_id(int minor)
476 return minor >> RBD_SINGLE_MAJOR_PART_SHIFT;
479 static bool rbd_is_lock_supported(struct rbd_device *rbd_dev)
481 return (rbd_dev->header.features & RBD_FEATURE_EXCLUSIVE_LOCK) &&
482 rbd_dev->spec->snap_id == CEPH_NOSNAP &&
483 !rbd_dev->mapping.read_only;
486 static bool __rbd_is_lock_owner(struct rbd_device *rbd_dev)
488 return rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED ||
489 rbd_dev->lock_state == RBD_LOCK_STATE_RELEASING;
492 static bool rbd_is_lock_owner(struct rbd_device *rbd_dev)
496 down_read(&rbd_dev->lock_rwsem);
497 is_lock_owner = __rbd_is_lock_owner(rbd_dev);
498 up_read(&rbd_dev->lock_rwsem);
499 return is_lock_owner;
502 static BUS_ATTR(add, S_IWUSR, NULL, rbd_add);
503 static BUS_ATTR(remove, S_IWUSR, NULL, rbd_remove);
504 static BUS_ATTR(add_single_major, S_IWUSR, NULL, rbd_add_single_major);
505 static BUS_ATTR(remove_single_major, S_IWUSR, NULL, rbd_remove_single_major);
507 static struct attribute *rbd_bus_attrs[] = {
509 &bus_attr_remove.attr,
510 &bus_attr_add_single_major.attr,
511 &bus_attr_remove_single_major.attr,
515 static umode_t rbd_bus_is_visible(struct kobject *kobj,
516 struct attribute *attr, int index)
519 (attr == &bus_attr_add_single_major.attr ||
520 attr == &bus_attr_remove_single_major.attr))
526 static const struct attribute_group rbd_bus_group = {
527 .attrs = rbd_bus_attrs,
528 .is_visible = rbd_bus_is_visible,
530 __ATTRIBUTE_GROUPS(rbd_bus);
532 static struct bus_type rbd_bus_type = {
534 .bus_groups = rbd_bus_groups,
537 static void rbd_root_dev_release(struct device *dev)
541 static struct device rbd_root_dev = {
543 .release = rbd_root_dev_release,
546 static __printf(2, 3)
547 void rbd_warn(struct rbd_device *rbd_dev, const char *fmt, ...)
549 struct va_format vaf;
557 printk(KERN_WARNING "%s: %pV\n", RBD_DRV_NAME, &vaf);
558 else if (rbd_dev->disk)
559 printk(KERN_WARNING "%s: %s: %pV\n",
560 RBD_DRV_NAME, rbd_dev->disk->disk_name, &vaf);
561 else if (rbd_dev->spec && rbd_dev->spec->image_name)
562 printk(KERN_WARNING "%s: image %s: %pV\n",
563 RBD_DRV_NAME, rbd_dev->spec->image_name, &vaf);
564 else if (rbd_dev->spec && rbd_dev->spec->image_id)
565 printk(KERN_WARNING "%s: id %s: %pV\n",
566 RBD_DRV_NAME, rbd_dev->spec->image_id, &vaf);
568 printk(KERN_WARNING "%s: rbd_dev %p: %pV\n",
569 RBD_DRV_NAME, rbd_dev, &vaf);
574 #define rbd_assert(expr) \
575 if (unlikely(!(expr))) { \
576 printk(KERN_ERR "\nAssertion failure in %s() " \
578 "\trbd_assert(%s);\n\n", \
579 __func__, __LINE__, #expr); \
582 #else /* !RBD_DEBUG */
583 # define rbd_assert(expr) ((void) 0)
584 #endif /* !RBD_DEBUG */
586 static void rbd_osd_copyup_callback(struct rbd_obj_request *obj_request);
587 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request);
588 static void rbd_img_parent_read(struct rbd_obj_request *obj_request);
589 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev);
591 static int rbd_dev_refresh(struct rbd_device *rbd_dev);
592 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev);
593 static int rbd_dev_header_info(struct rbd_device *rbd_dev);
594 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev);
595 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
597 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
598 u8 *order, u64 *snap_size);
599 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
602 static int rbd_open(struct block_device *bdev, fmode_t mode)
604 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
605 bool removing = false;
607 if ((mode & FMODE_WRITE) && rbd_dev->mapping.read_only)
610 spin_lock_irq(&rbd_dev->lock);
611 if (test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags))
614 rbd_dev->open_count++;
615 spin_unlock_irq(&rbd_dev->lock);
619 (void) get_device(&rbd_dev->dev);
624 static void rbd_release(struct gendisk *disk, fmode_t mode)
626 struct rbd_device *rbd_dev = disk->private_data;
627 unsigned long open_count_before;
629 spin_lock_irq(&rbd_dev->lock);
630 open_count_before = rbd_dev->open_count--;
631 spin_unlock_irq(&rbd_dev->lock);
632 rbd_assert(open_count_before > 0);
634 put_device(&rbd_dev->dev);
637 static int rbd_ioctl_set_ro(struct rbd_device *rbd_dev, unsigned long arg)
642 bool ro_changed = false;
644 /* get_user() may sleep, so call it before taking rbd_dev->lock */
645 if (get_user(val, (int __user *)(arg)))
648 ro = val ? true : false;
649 /* Snapshot doesn't allow to write*/
650 if (rbd_dev->spec->snap_id != CEPH_NOSNAP && !ro)
653 spin_lock_irq(&rbd_dev->lock);
654 /* prevent others open this device */
655 if (rbd_dev->open_count > 1) {
660 if (rbd_dev->mapping.read_only != ro) {
661 rbd_dev->mapping.read_only = ro;
666 spin_unlock_irq(&rbd_dev->lock);
667 /* set_disk_ro() may sleep, so call it after releasing rbd_dev->lock */
668 if (ret == 0 && ro_changed)
669 set_disk_ro(rbd_dev->disk, ro ? 1 : 0);
674 static int rbd_ioctl(struct block_device *bdev, fmode_t mode,
675 unsigned int cmd, unsigned long arg)
677 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
682 ret = rbd_ioctl_set_ro(rbd_dev, arg);
692 static int rbd_compat_ioctl(struct block_device *bdev, fmode_t mode,
693 unsigned int cmd, unsigned long arg)
695 return rbd_ioctl(bdev, mode, cmd, arg);
697 #endif /* CONFIG_COMPAT */
699 static const struct block_device_operations rbd_bd_ops = {
700 .owner = THIS_MODULE,
702 .release = rbd_release,
705 .compat_ioctl = rbd_compat_ioctl,
710 * Initialize an rbd client instance. Success or not, this function
711 * consumes ceph_opts. Caller holds client_mutex.
713 static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts)
715 struct rbd_client *rbdc;
718 dout("%s:\n", __func__);
719 rbdc = kmalloc(sizeof(struct rbd_client), GFP_KERNEL);
723 kref_init(&rbdc->kref);
724 INIT_LIST_HEAD(&rbdc->node);
726 rbdc->client = ceph_create_client(ceph_opts, rbdc, 0, 0);
727 if (IS_ERR(rbdc->client))
729 ceph_opts = NULL; /* Now rbdc->client is responsible for ceph_opts */
731 ret = ceph_open_session(rbdc->client);
735 spin_lock(&rbd_client_list_lock);
736 list_add_tail(&rbdc->node, &rbd_client_list);
737 spin_unlock(&rbd_client_list_lock);
739 dout("%s: rbdc %p\n", __func__, rbdc);
743 ceph_destroy_client(rbdc->client);
748 ceph_destroy_options(ceph_opts);
749 dout("%s: error %d\n", __func__, ret);
754 static struct rbd_client *__rbd_get_client(struct rbd_client *rbdc)
756 kref_get(&rbdc->kref);
762 * Find a ceph client with specific addr and configuration. If
763 * found, bump its reference count.
765 static struct rbd_client *rbd_client_find(struct ceph_options *ceph_opts)
767 struct rbd_client *client_node;
770 if (ceph_opts->flags & CEPH_OPT_NOSHARE)
773 spin_lock(&rbd_client_list_lock);
774 list_for_each_entry(client_node, &rbd_client_list, node) {
775 if (!ceph_compare_options(ceph_opts, client_node->client)) {
776 __rbd_get_client(client_node);
782 spin_unlock(&rbd_client_list_lock);
784 return found ? client_node : NULL;
788 * (Per device) rbd map options
795 /* string args above */
802 static match_table_t rbd_opts_tokens = {
803 {Opt_queue_depth, "queue_depth=%d"},
805 /* string args above */
806 {Opt_read_only, "read_only"},
807 {Opt_read_only, "ro"}, /* Alternate spelling */
808 {Opt_read_write, "read_write"},
809 {Opt_read_write, "rw"}, /* Alternate spelling */
810 {Opt_lock_on_read, "lock_on_read"},
820 #define RBD_QUEUE_DEPTH_DEFAULT BLKDEV_MAX_RQ
821 #define RBD_READ_ONLY_DEFAULT false
822 #define RBD_LOCK_ON_READ_DEFAULT false
824 static int parse_rbd_opts_token(char *c, void *private)
826 struct rbd_options *rbd_opts = private;
827 substring_t argstr[MAX_OPT_ARGS];
828 int token, intval, ret;
830 token = match_token(c, rbd_opts_tokens, argstr);
831 if (token < Opt_last_int) {
832 ret = match_int(&argstr[0], &intval);
834 pr_err("bad mount option arg (not int) at '%s'\n", c);
837 dout("got int token %d val %d\n", token, intval);
838 } else if (token > Opt_last_int && token < Opt_last_string) {
839 dout("got string token %d val %s\n", token, argstr[0].from);
841 dout("got token %d\n", token);
845 case Opt_queue_depth:
847 pr_err("queue_depth out of range\n");
850 rbd_opts->queue_depth = intval;
853 rbd_opts->read_only = true;
856 rbd_opts->read_only = false;
858 case Opt_lock_on_read:
859 rbd_opts->lock_on_read = true;
862 /* libceph prints "bad option" msg */
869 static char* obj_op_name(enum obj_operation_type op_type)
884 * Get a ceph client with specific addr and configuration, if one does
885 * not exist create it. Either way, ceph_opts is consumed by this
888 static struct rbd_client *rbd_get_client(struct ceph_options *ceph_opts)
890 struct rbd_client *rbdc;
892 mutex_lock_nested(&client_mutex, SINGLE_DEPTH_NESTING);
893 rbdc = rbd_client_find(ceph_opts);
894 if (rbdc) /* using an existing client */
895 ceph_destroy_options(ceph_opts);
897 rbdc = rbd_client_create(ceph_opts);
898 mutex_unlock(&client_mutex);
904 * Destroy ceph client
906 * Caller must hold rbd_client_list_lock.
908 static void rbd_client_release(struct kref *kref)
910 struct rbd_client *rbdc = container_of(kref, struct rbd_client, kref);
912 dout("%s: rbdc %p\n", __func__, rbdc);
913 spin_lock(&rbd_client_list_lock);
914 list_del(&rbdc->node);
915 spin_unlock(&rbd_client_list_lock);
917 ceph_destroy_client(rbdc->client);
922 * Drop reference to ceph client node. If it's not referenced anymore, release
925 static void rbd_put_client(struct rbd_client *rbdc)
928 kref_put(&rbdc->kref, rbd_client_release);
931 static bool rbd_image_format_valid(u32 image_format)
933 return image_format == 1 || image_format == 2;
936 static bool rbd_dev_ondisk_valid(struct rbd_image_header_ondisk *ondisk)
941 /* The header has to start with the magic rbd header text */
942 if (memcmp(&ondisk->text, RBD_HEADER_TEXT, sizeof (RBD_HEADER_TEXT)))
945 /* The bio layer requires at least sector-sized I/O */
947 if (ondisk->options.order < SECTOR_SHIFT)
950 /* If we use u64 in a few spots we may be able to loosen this */
952 if (ondisk->options.order > 8 * sizeof (int) - 1)
956 * The size of a snapshot header has to fit in a size_t, and
957 * that limits the number of snapshots.
959 snap_count = le32_to_cpu(ondisk->snap_count);
960 size = SIZE_MAX - sizeof (struct ceph_snap_context);
961 if (snap_count > size / sizeof (__le64))
965 * Not only that, but the size of the entire the snapshot
966 * header must also be representable in a size_t.
968 size -= snap_count * sizeof (__le64);
969 if ((u64) size < le64_to_cpu(ondisk->snap_names_len))
976 * Fill an rbd image header with information from the given format 1
979 static int rbd_header_from_disk(struct rbd_device *rbd_dev,
980 struct rbd_image_header_ondisk *ondisk)
982 struct rbd_image_header *header = &rbd_dev->header;
983 bool first_time = header->object_prefix == NULL;
984 struct ceph_snap_context *snapc;
985 char *object_prefix = NULL;
986 char *snap_names = NULL;
987 u64 *snap_sizes = NULL;
993 /* Allocate this now to avoid having to handle failure below */
998 len = strnlen(ondisk->object_prefix,
999 sizeof (ondisk->object_prefix));
1000 object_prefix = kmalloc(len + 1, GFP_KERNEL);
1003 memcpy(object_prefix, ondisk->object_prefix, len);
1004 object_prefix[len] = '\0';
1007 /* Allocate the snapshot context and fill it in */
1009 snap_count = le32_to_cpu(ondisk->snap_count);
1010 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
1013 snapc->seq = le64_to_cpu(ondisk->snap_seq);
1015 struct rbd_image_snap_ondisk *snaps;
1016 u64 snap_names_len = le64_to_cpu(ondisk->snap_names_len);
1018 /* We'll keep a copy of the snapshot names... */
1020 if (snap_names_len > (u64)SIZE_MAX)
1022 snap_names = kmalloc(snap_names_len, GFP_KERNEL);
1026 /* ...as well as the array of their sizes. */
1028 size = snap_count * sizeof (*header->snap_sizes);
1029 snap_sizes = kmalloc(size, GFP_KERNEL);
1034 * Copy the names, and fill in each snapshot's id
1037 * Note that rbd_dev_v1_header_info() guarantees the
1038 * ondisk buffer we're working with has
1039 * snap_names_len bytes beyond the end of the
1040 * snapshot id array, this memcpy() is safe.
1042 memcpy(snap_names, &ondisk->snaps[snap_count], snap_names_len);
1043 snaps = ondisk->snaps;
1044 for (i = 0; i < snap_count; i++) {
1045 snapc->snaps[i] = le64_to_cpu(snaps[i].id);
1046 snap_sizes[i] = le64_to_cpu(snaps[i].image_size);
1050 /* We won't fail any more, fill in the header */
1053 header->object_prefix = object_prefix;
1054 header->obj_order = ondisk->options.order;
1055 header->crypt_type = ondisk->options.crypt_type;
1056 header->comp_type = ondisk->options.comp_type;
1057 /* The rest aren't used for format 1 images */
1058 header->stripe_unit = 0;
1059 header->stripe_count = 0;
1060 header->features = 0;
1062 ceph_put_snap_context(header->snapc);
1063 kfree(header->snap_names);
1064 kfree(header->snap_sizes);
1067 /* The remaining fields always get updated (when we refresh) */
1069 header->image_size = le64_to_cpu(ondisk->image_size);
1070 header->snapc = snapc;
1071 header->snap_names = snap_names;
1072 header->snap_sizes = snap_sizes;
1080 ceph_put_snap_context(snapc);
1081 kfree(object_prefix);
1086 static const char *_rbd_dev_v1_snap_name(struct rbd_device *rbd_dev, u32 which)
1088 const char *snap_name;
1090 rbd_assert(which < rbd_dev->header.snapc->num_snaps);
1092 /* Skip over names until we find the one we are looking for */
1094 snap_name = rbd_dev->header.snap_names;
1096 snap_name += strlen(snap_name) + 1;
1098 return kstrdup(snap_name, GFP_KERNEL);
1102 * Snapshot id comparison function for use with qsort()/bsearch().
1103 * Note that result is for snapshots in *descending* order.
1105 static int snapid_compare_reverse(const void *s1, const void *s2)
1107 u64 snap_id1 = *(u64 *)s1;
1108 u64 snap_id2 = *(u64 *)s2;
1110 if (snap_id1 < snap_id2)
1112 return snap_id1 == snap_id2 ? 0 : -1;
1116 * Search a snapshot context to see if the given snapshot id is
1119 * Returns the position of the snapshot id in the array if it's found,
1120 * or BAD_SNAP_INDEX otherwise.
1122 * Note: The snapshot array is in kept sorted (by the osd) in
1123 * reverse order, highest snapshot id first.
1125 static u32 rbd_dev_snap_index(struct rbd_device *rbd_dev, u64 snap_id)
1127 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
1130 found = bsearch(&snap_id, &snapc->snaps, snapc->num_snaps,
1131 sizeof (snap_id), snapid_compare_reverse);
1133 return found ? (u32)(found - &snapc->snaps[0]) : BAD_SNAP_INDEX;
1136 static const char *rbd_dev_v1_snap_name(struct rbd_device *rbd_dev,
1140 const char *snap_name;
1142 which = rbd_dev_snap_index(rbd_dev, snap_id);
1143 if (which == BAD_SNAP_INDEX)
1144 return ERR_PTR(-ENOENT);
1146 snap_name = _rbd_dev_v1_snap_name(rbd_dev, which);
1147 return snap_name ? snap_name : ERR_PTR(-ENOMEM);
1150 static const char *rbd_snap_name(struct rbd_device *rbd_dev, u64 snap_id)
1152 if (snap_id == CEPH_NOSNAP)
1153 return RBD_SNAP_HEAD_NAME;
1155 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1156 if (rbd_dev->image_format == 1)
1157 return rbd_dev_v1_snap_name(rbd_dev, snap_id);
1159 return rbd_dev_v2_snap_name(rbd_dev, snap_id);
1162 static int rbd_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
1165 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1166 if (snap_id == CEPH_NOSNAP) {
1167 *snap_size = rbd_dev->header.image_size;
1168 } else if (rbd_dev->image_format == 1) {
1171 which = rbd_dev_snap_index(rbd_dev, snap_id);
1172 if (which == BAD_SNAP_INDEX)
1175 *snap_size = rbd_dev->header.snap_sizes[which];
1180 ret = _rbd_dev_v2_snap_size(rbd_dev, snap_id, NULL, &size);
1189 static int rbd_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
1192 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1193 if (snap_id == CEPH_NOSNAP) {
1194 *snap_features = rbd_dev->header.features;
1195 } else if (rbd_dev->image_format == 1) {
1196 *snap_features = 0; /* No features for format 1 */
1201 ret = _rbd_dev_v2_snap_features(rbd_dev, snap_id, &features);
1205 *snap_features = features;
1210 static int rbd_dev_mapping_set(struct rbd_device *rbd_dev)
1212 u64 snap_id = rbd_dev->spec->snap_id;
1217 ret = rbd_snap_size(rbd_dev, snap_id, &size);
1220 ret = rbd_snap_features(rbd_dev, snap_id, &features);
1224 rbd_dev->mapping.size = size;
1225 rbd_dev->mapping.features = features;
1230 static void rbd_dev_mapping_clear(struct rbd_device *rbd_dev)
1232 rbd_dev->mapping.size = 0;
1233 rbd_dev->mapping.features = 0;
1236 static void rbd_segment_name_free(const char *name)
1238 /* The explicit cast here is needed to drop the const qualifier */
1240 kmem_cache_free(rbd_segment_name_cache, (void *)name);
1243 static const char *rbd_segment_name(struct rbd_device *rbd_dev, u64 offset)
1250 name = kmem_cache_alloc(rbd_segment_name_cache, GFP_NOIO);
1253 segment = offset >> rbd_dev->header.obj_order;
1254 name_format = "%s.%012llx";
1255 if (rbd_dev->image_format == 2)
1256 name_format = "%s.%016llx";
1257 ret = snprintf(name, CEPH_MAX_OID_NAME_LEN + 1, name_format,
1258 rbd_dev->header.object_prefix, segment);
1259 if (ret < 0 || ret > CEPH_MAX_OID_NAME_LEN) {
1260 pr_err("error formatting segment name for #%llu (%d)\n",
1262 rbd_segment_name_free(name);
1269 static u64 rbd_segment_offset(struct rbd_device *rbd_dev, u64 offset)
1271 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1273 return offset & (segment_size - 1);
1276 static u64 rbd_segment_length(struct rbd_device *rbd_dev,
1277 u64 offset, u64 length)
1279 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1281 offset &= segment_size - 1;
1283 rbd_assert(length <= U64_MAX - offset);
1284 if (offset + length > segment_size)
1285 length = segment_size - offset;
1291 * returns the size of an object in the image
1293 static u64 rbd_obj_bytes(struct rbd_image_header *header)
1295 return 1 << header->obj_order;
1302 static void bio_chain_put(struct bio *chain)
1308 chain = chain->bi_next;
1314 * zeros a bio chain, starting at specific offset
1316 static void zero_bio_chain(struct bio *chain, int start_ofs)
1319 struct bvec_iter iter;
1320 unsigned long flags;
1325 bio_for_each_segment(bv, chain, iter) {
1326 if (pos + bv.bv_len > start_ofs) {
1327 int remainder = max(start_ofs - pos, 0);
1328 buf = bvec_kmap_irq(&bv, &flags);
1329 memset(buf + remainder, 0,
1330 bv.bv_len - remainder);
1331 flush_dcache_page(bv.bv_page);
1332 bvec_kunmap_irq(buf, &flags);
1337 chain = chain->bi_next;
1342 * similar to zero_bio_chain(), zeros data defined by a page array,
1343 * starting at the given byte offset from the start of the array and
1344 * continuing up to the given end offset. The pages array is
1345 * assumed to be big enough to hold all bytes up to the end.
1347 static void zero_pages(struct page **pages, u64 offset, u64 end)
1349 struct page **page = &pages[offset >> PAGE_SHIFT];
1351 rbd_assert(end > offset);
1352 rbd_assert(end - offset <= (u64)SIZE_MAX);
1353 while (offset < end) {
1356 unsigned long flags;
1359 page_offset = offset & ~PAGE_MASK;
1360 length = min_t(size_t, PAGE_SIZE - page_offset, end - offset);
1361 local_irq_save(flags);
1362 kaddr = kmap_atomic(*page);
1363 memset(kaddr + page_offset, 0, length);
1364 flush_dcache_page(*page);
1365 kunmap_atomic(kaddr);
1366 local_irq_restore(flags);
1374 * Clone a portion of a bio, starting at the given byte offset
1375 * and continuing for the number of bytes indicated.
1377 static struct bio *bio_clone_range(struct bio *bio_src,
1378 unsigned int offset,
1384 bio = bio_clone(bio_src, gfpmask);
1386 return NULL; /* ENOMEM */
1388 bio_advance(bio, offset);
1389 bio->bi_iter.bi_size = len;
1395 * Clone a portion of a bio chain, starting at the given byte offset
1396 * into the first bio in the source chain and continuing for the
1397 * number of bytes indicated. The result is another bio chain of
1398 * exactly the given length, or a null pointer on error.
1400 * The bio_src and offset parameters are both in-out. On entry they
1401 * refer to the first source bio and the offset into that bio where
1402 * the start of data to be cloned is located.
1404 * On return, bio_src is updated to refer to the bio in the source
1405 * chain that contains first un-cloned byte, and *offset will
1406 * contain the offset of that byte within that bio.
1408 static struct bio *bio_chain_clone_range(struct bio **bio_src,
1409 unsigned int *offset,
1413 struct bio *bi = *bio_src;
1414 unsigned int off = *offset;
1415 struct bio *chain = NULL;
1418 /* Build up a chain of clone bios up to the limit */
1420 if (!bi || off >= bi->bi_iter.bi_size || !len)
1421 return NULL; /* Nothing to clone */
1425 unsigned int bi_size;
1429 rbd_warn(NULL, "bio_chain exhausted with %u left", len);
1430 goto out_err; /* EINVAL; ran out of bio's */
1432 bi_size = min_t(unsigned int, bi->bi_iter.bi_size - off, len);
1433 bio = bio_clone_range(bi, off, bi_size, gfpmask);
1435 goto out_err; /* ENOMEM */
1438 end = &bio->bi_next;
1441 if (off == bi->bi_iter.bi_size) {
1452 bio_chain_put(chain);
1458 * The default/initial value for all object request flags is 0. For
1459 * each flag, once its value is set to 1 it is never reset to 0
1462 static void obj_request_img_data_set(struct rbd_obj_request *obj_request)
1464 if (test_and_set_bit(OBJ_REQ_IMG_DATA, &obj_request->flags)) {
1465 struct rbd_device *rbd_dev;
1467 rbd_dev = obj_request->img_request->rbd_dev;
1468 rbd_warn(rbd_dev, "obj_request %p already marked img_data",
1473 static bool obj_request_img_data_test(struct rbd_obj_request *obj_request)
1476 return test_bit(OBJ_REQ_IMG_DATA, &obj_request->flags) != 0;
1479 static void obj_request_done_set(struct rbd_obj_request *obj_request)
1481 if (test_and_set_bit(OBJ_REQ_DONE, &obj_request->flags)) {
1482 struct rbd_device *rbd_dev = NULL;
1484 if (obj_request_img_data_test(obj_request))
1485 rbd_dev = obj_request->img_request->rbd_dev;
1486 rbd_warn(rbd_dev, "obj_request %p already marked done",
1491 static bool obj_request_done_test(struct rbd_obj_request *obj_request)
1494 return test_bit(OBJ_REQ_DONE, &obj_request->flags) != 0;
1498 * This sets the KNOWN flag after (possibly) setting the EXISTS
1499 * flag. The latter is set based on the "exists" value provided.
1501 * Note that for our purposes once an object exists it never goes
1502 * away again. It's possible that the response from two existence
1503 * checks are separated by the creation of the target object, and
1504 * the first ("doesn't exist") response arrives *after* the second
1505 * ("does exist"). In that case we ignore the second one.
1507 static void obj_request_existence_set(struct rbd_obj_request *obj_request,
1511 set_bit(OBJ_REQ_EXISTS, &obj_request->flags);
1512 set_bit(OBJ_REQ_KNOWN, &obj_request->flags);
1516 static bool obj_request_known_test(struct rbd_obj_request *obj_request)
1519 return test_bit(OBJ_REQ_KNOWN, &obj_request->flags) != 0;
1522 static bool obj_request_exists_test(struct rbd_obj_request *obj_request)
1525 return test_bit(OBJ_REQ_EXISTS, &obj_request->flags) != 0;
1528 static bool obj_request_overlaps_parent(struct rbd_obj_request *obj_request)
1530 struct rbd_device *rbd_dev = obj_request->img_request->rbd_dev;
1532 return obj_request->img_offset <
1533 round_up(rbd_dev->parent_overlap, rbd_obj_bytes(&rbd_dev->header));
1536 static void rbd_obj_request_get(struct rbd_obj_request *obj_request)
1538 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1539 atomic_read(&obj_request->kref.refcount));
1540 kref_get(&obj_request->kref);
1543 static void rbd_obj_request_destroy(struct kref *kref);
1544 static void rbd_obj_request_put(struct rbd_obj_request *obj_request)
1546 rbd_assert(obj_request != NULL);
1547 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1548 atomic_read(&obj_request->kref.refcount));
1549 kref_put(&obj_request->kref, rbd_obj_request_destroy);
1552 static void rbd_img_request_get(struct rbd_img_request *img_request)
1554 dout("%s: img %p (was %d)\n", __func__, img_request,
1555 atomic_read(&img_request->kref.refcount));
1556 kref_get(&img_request->kref);
1559 static bool img_request_child_test(struct rbd_img_request *img_request);
1560 static void rbd_parent_request_destroy(struct kref *kref);
1561 static void rbd_img_request_destroy(struct kref *kref);
1562 static void rbd_img_request_put(struct rbd_img_request *img_request)
1564 rbd_assert(img_request != NULL);
1565 dout("%s: img %p (was %d)\n", __func__, img_request,
1566 atomic_read(&img_request->kref.refcount));
1567 if (img_request_child_test(img_request))
1568 kref_put(&img_request->kref, rbd_parent_request_destroy);
1570 kref_put(&img_request->kref, rbd_img_request_destroy);
1573 static inline void rbd_img_obj_request_add(struct rbd_img_request *img_request,
1574 struct rbd_obj_request *obj_request)
1576 rbd_assert(obj_request->img_request == NULL);
1578 /* Image request now owns object's original reference */
1579 obj_request->img_request = img_request;
1580 obj_request->which = img_request->obj_request_count;
1581 rbd_assert(!obj_request_img_data_test(obj_request));
1582 obj_request_img_data_set(obj_request);
1583 rbd_assert(obj_request->which != BAD_WHICH);
1584 img_request->obj_request_count++;
1585 list_add_tail(&obj_request->links, &img_request->obj_requests);
1586 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1587 obj_request->which);
1590 static inline void rbd_img_obj_request_del(struct rbd_img_request *img_request,
1591 struct rbd_obj_request *obj_request)
1593 rbd_assert(obj_request->which != BAD_WHICH);
1595 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1596 obj_request->which);
1597 list_del(&obj_request->links);
1598 rbd_assert(img_request->obj_request_count > 0);
1599 img_request->obj_request_count--;
1600 rbd_assert(obj_request->which == img_request->obj_request_count);
1601 obj_request->which = BAD_WHICH;
1602 rbd_assert(obj_request_img_data_test(obj_request));
1603 rbd_assert(obj_request->img_request == img_request);
1604 obj_request->img_request = NULL;
1605 obj_request->callback = NULL;
1606 rbd_obj_request_put(obj_request);
1609 static bool obj_request_type_valid(enum obj_request_type type)
1612 case OBJ_REQUEST_NODATA:
1613 case OBJ_REQUEST_BIO:
1614 case OBJ_REQUEST_PAGES:
1621 static void rbd_obj_request_submit(struct rbd_obj_request *obj_request)
1623 struct ceph_osd_request *osd_req = obj_request->osd_req;
1625 dout("%s %p osd_req %p\n", __func__, obj_request, osd_req);
1626 ceph_osdc_start_request(osd_req->r_osdc, osd_req, false);
1629 static void rbd_obj_request_end(struct rbd_obj_request *obj_request)
1631 dout("%s %p\n", __func__, obj_request);
1632 ceph_osdc_cancel_request(obj_request->osd_req);
1636 * Wait for an object request to complete. If interrupted, cancel the
1637 * underlying osd request.
1639 * @timeout: in jiffies, 0 means "wait forever"
1641 static int __rbd_obj_request_wait(struct rbd_obj_request *obj_request,
1642 unsigned long timeout)
1646 dout("%s %p\n", __func__, obj_request);
1647 ret = wait_for_completion_interruptible_timeout(
1648 &obj_request->completion,
1649 ceph_timeout_jiffies(timeout));
1653 rbd_obj_request_end(obj_request);
1658 dout("%s %p ret %d\n", __func__, obj_request, (int)ret);
1662 static int rbd_obj_request_wait(struct rbd_obj_request *obj_request)
1664 return __rbd_obj_request_wait(obj_request, 0);
1667 static void rbd_img_request_complete(struct rbd_img_request *img_request)
1670 dout("%s: img %p\n", __func__, img_request);
1673 * If no error occurred, compute the aggregate transfer
1674 * count for the image request. We could instead use
1675 * atomic64_cmpxchg() to update it as each object request
1676 * completes; not clear which way is better off hand.
1678 if (!img_request->result) {
1679 struct rbd_obj_request *obj_request;
1682 for_each_obj_request(img_request, obj_request)
1683 xferred += obj_request->xferred;
1684 img_request->xferred = xferred;
1687 if (img_request->callback)
1688 img_request->callback(img_request);
1690 rbd_img_request_put(img_request);
1694 * The default/initial value for all image request flags is 0. Each
1695 * is conditionally set to 1 at image request initialization time
1696 * and currently never change thereafter.
1698 static void img_request_write_set(struct rbd_img_request *img_request)
1700 set_bit(IMG_REQ_WRITE, &img_request->flags);
1704 static bool img_request_write_test(struct rbd_img_request *img_request)
1707 return test_bit(IMG_REQ_WRITE, &img_request->flags) != 0;
1711 * Set the discard flag when the img_request is an discard request
1713 static void img_request_discard_set(struct rbd_img_request *img_request)
1715 set_bit(IMG_REQ_DISCARD, &img_request->flags);
1719 static bool img_request_discard_test(struct rbd_img_request *img_request)
1722 return test_bit(IMG_REQ_DISCARD, &img_request->flags) != 0;
1725 static void img_request_child_set(struct rbd_img_request *img_request)
1727 set_bit(IMG_REQ_CHILD, &img_request->flags);
1731 static void img_request_child_clear(struct rbd_img_request *img_request)
1733 clear_bit(IMG_REQ_CHILD, &img_request->flags);
1737 static bool img_request_child_test(struct rbd_img_request *img_request)
1740 return test_bit(IMG_REQ_CHILD, &img_request->flags) != 0;
1743 static void img_request_layered_set(struct rbd_img_request *img_request)
1745 set_bit(IMG_REQ_LAYERED, &img_request->flags);
1749 static void img_request_layered_clear(struct rbd_img_request *img_request)
1751 clear_bit(IMG_REQ_LAYERED, &img_request->flags);
1755 static bool img_request_layered_test(struct rbd_img_request *img_request)
1758 return test_bit(IMG_REQ_LAYERED, &img_request->flags) != 0;
1761 static enum obj_operation_type
1762 rbd_img_request_op_type(struct rbd_img_request *img_request)
1764 if (img_request_write_test(img_request))
1765 return OBJ_OP_WRITE;
1766 else if (img_request_discard_test(img_request))
1767 return OBJ_OP_DISCARD;
1773 rbd_img_obj_request_read_callback(struct rbd_obj_request *obj_request)
1775 u64 xferred = obj_request->xferred;
1776 u64 length = obj_request->length;
1778 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1779 obj_request, obj_request->img_request, obj_request->result,
1782 * ENOENT means a hole in the image. We zero-fill the entire
1783 * length of the request. A short read also implies zero-fill
1784 * to the end of the request. An error requires the whole
1785 * length of the request to be reported finished with an error
1786 * to the block layer. In each case we update the xferred
1787 * count to indicate the whole request was satisfied.
1789 rbd_assert(obj_request->type != OBJ_REQUEST_NODATA);
1790 if (obj_request->result == -ENOENT) {
1791 if (obj_request->type == OBJ_REQUEST_BIO)
1792 zero_bio_chain(obj_request->bio_list, 0);
1794 zero_pages(obj_request->pages, 0, length);
1795 obj_request->result = 0;
1796 } else if (xferred < length && !obj_request->result) {
1797 if (obj_request->type == OBJ_REQUEST_BIO)
1798 zero_bio_chain(obj_request->bio_list, xferred);
1800 zero_pages(obj_request->pages, xferred, length);
1802 obj_request->xferred = length;
1803 obj_request_done_set(obj_request);
1806 static void rbd_obj_request_complete(struct rbd_obj_request *obj_request)
1808 dout("%s: obj %p cb %p\n", __func__, obj_request,
1809 obj_request->callback);
1810 if (obj_request->callback)
1811 obj_request->callback(obj_request);
1813 complete_all(&obj_request->completion);
1816 static void rbd_osd_read_callback(struct rbd_obj_request *obj_request)
1818 struct rbd_img_request *img_request = NULL;
1819 struct rbd_device *rbd_dev = NULL;
1820 bool layered = false;
1822 if (obj_request_img_data_test(obj_request)) {
1823 img_request = obj_request->img_request;
1824 layered = img_request && img_request_layered_test(img_request);
1825 rbd_dev = img_request->rbd_dev;
1828 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1829 obj_request, img_request, obj_request->result,
1830 obj_request->xferred, obj_request->length);
1831 if (layered && obj_request->result == -ENOENT &&
1832 obj_request->img_offset < rbd_dev->parent_overlap)
1833 rbd_img_parent_read(obj_request);
1834 else if (img_request)
1835 rbd_img_obj_request_read_callback(obj_request);
1837 obj_request_done_set(obj_request);
1840 static void rbd_osd_write_callback(struct rbd_obj_request *obj_request)
1842 dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1843 obj_request->result, obj_request->length);
1845 * There is no such thing as a successful short write. Set
1846 * it to our originally-requested length.
1848 obj_request->xferred = obj_request->length;
1849 obj_request_done_set(obj_request);
1852 static void rbd_osd_discard_callback(struct rbd_obj_request *obj_request)
1854 dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1855 obj_request->result, obj_request->length);
1857 * There is no such thing as a successful short discard. Set
1858 * it to our originally-requested length.
1860 obj_request->xferred = obj_request->length;
1861 /* discarding a non-existent object is not a problem */
1862 if (obj_request->result == -ENOENT)
1863 obj_request->result = 0;
1864 obj_request_done_set(obj_request);
1868 * For a simple stat call there's nothing to do. We'll do more if
1869 * this is part of a write sequence for a layered image.
1871 static void rbd_osd_stat_callback(struct rbd_obj_request *obj_request)
1873 dout("%s: obj %p\n", __func__, obj_request);
1874 obj_request_done_set(obj_request);
1877 static void rbd_osd_call_callback(struct rbd_obj_request *obj_request)
1879 dout("%s: obj %p\n", __func__, obj_request);
1881 if (obj_request_img_data_test(obj_request))
1882 rbd_osd_copyup_callback(obj_request);
1884 obj_request_done_set(obj_request);
1887 static void rbd_osd_req_callback(struct ceph_osd_request *osd_req)
1889 struct rbd_obj_request *obj_request = osd_req->r_priv;
1892 dout("%s: osd_req %p\n", __func__, osd_req);
1893 rbd_assert(osd_req == obj_request->osd_req);
1894 if (obj_request_img_data_test(obj_request)) {
1895 rbd_assert(obj_request->img_request);
1896 rbd_assert(obj_request->which != BAD_WHICH);
1898 rbd_assert(obj_request->which == BAD_WHICH);
1901 if (osd_req->r_result < 0)
1902 obj_request->result = osd_req->r_result;
1905 * We support a 64-bit length, but ultimately it has to be
1906 * passed to the block layer, which just supports a 32-bit
1909 obj_request->xferred = osd_req->r_ops[0].outdata_len;
1910 rbd_assert(obj_request->xferred < (u64)UINT_MAX);
1912 opcode = osd_req->r_ops[0].op;
1914 case CEPH_OSD_OP_READ:
1915 rbd_osd_read_callback(obj_request);
1917 case CEPH_OSD_OP_SETALLOCHINT:
1918 rbd_assert(osd_req->r_ops[1].op == CEPH_OSD_OP_WRITE ||
1919 osd_req->r_ops[1].op == CEPH_OSD_OP_WRITEFULL);
1921 case CEPH_OSD_OP_WRITE:
1922 case CEPH_OSD_OP_WRITEFULL:
1923 rbd_osd_write_callback(obj_request);
1925 case CEPH_OSD_OP_STAT:
1926 rbd_osd_stat_callback(obj_request);
1928 case CEPH_OSD_OP_DELETE:
1929 case CEPH_OSD_OP_TRUNCATE:
1930 case CEPH_OSD_OP_ZERO:
1931 rbd_osd_discard_callback(obj_request);
1933 case CEPH_OSD_OP_CALL:
1934 rbd_osd_call_callback(obj_request);
1937 rbd_warn(NULL, "%s: unsupported op %hu",
1938 obj_request->object_name, (unsigned short) opcode);
1942 if (obj_request_done_test(obj_request))
1943 rbd_obj_request_complete(obj_request);
1946 static void rbd_osd_req_format_read(struct rbd_obj_request *obj_request)
1948 struct rbd_img_request *img_request = obj_request->img_request;
1949 struct ceph_osd_request *osd_req = obj_request->osd_req;
1952 osd_req->r_snapid = img_request->snap_id;
1955 static void rbd_osd_req_format_write(struct rbd_obj_request *obj_request)
1957 struct ceph_osd_request *osd_req = obj_request->osd_req;
1959 osd_req->r_mtime = CURRENT_TIME;
1960 osd_req->r_data_offset = obj_request->offset;
1964 * Create an osd request. A read request has one osd op (read).
1965 * A write request has either one (watch) or two (hint+write) osd ops.
1966 * (All rbd data writes are prefixed with an allocation hint op, but
1967 * technically osd watch is a write request, hence this distinction.)
1969 static struct ceph_osd_request *rbd_osd_req_create(
1970 struct rbd_device *rbd_dev,
1971 enum obj_operation_type op_type,
1972 unsigned int num_ops,
1973 struct rbd_obj_request *obj_request)
1975 struct ceph_snap_context *snapc = NULL;
1976 struct ceph_osd_client *osdc;
1977 struct ceph_osd_request *osd_req;
1979 if (obj_request_img_data_test(obj_request) &&
1980 (op_type == OBJ_OP_DISCARD || op_type == OBJ_OP_WRITE)) {
1981 struct rbd_img_request *img_request = obj_request->img_request;
1982 if (op_type == OBJ_OP_WRITE) {
1983 rbd_assert(img_request_write_test(img_request));
1985 rbd_assert(img_request_discard_test(img_request));
1987 snapc = img_request->snapc;
1990 rbd_assert(num_ops == 1 || ((op_type == OBJ_OP_WRITE) && num_ops == 2));
1992 /* Allocate and initialize the request, for the num_ops ops */
1994 osdc = &rbd_dev->rbd_client->client->osdc;
1995 osd_req = ceph_osdc_alloc_request(osdc, snapc, num_ops, false,
2000 if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
2001 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
2003 osd_req->r_flags = CEPH_OSD_FLAG_READ;
2005 osd_req->r_callback = rbd_osd_req_callback;
2006 osd_req->r_priv = obj_request;
2008 osd_req->r_base_oloc.pool = rbd_dev->layout.pool_id;
2009 if (ceph_oid_aprintf(&osd_req->r_base_oid, GFP_NOIO, "%s",
2010 obj_request->object_name))
2013 if (ceph_osdc_alloc_messages(osd_req, GFP_NOIO))
2019 ceph_osdc_put_request(osd_req);
2024 * Create a copyup osd request based on the information in the object
2025 * request supplied. A copyup request has two or three osd ops, a
2026 * copyup method call, potentially a hint op, and a write or truncate
2029 static struct ceph_osd_request *
2030 rbd_osd_req_create_copyup(struct rbd_obj_request *obj_request)
2032 struct rbd_img_request *img_request;
2033 struct ceph_snap_context *snapc;
2034 struct rbd_device *rbd_dev;
2035 struct ceph_osd_client *osdc;
2036 struct ceph_osd_request *osd_req;
2037 int num_osd_ops = 3;
2039 rbd_assert(obj_request_img_data_test(obj_request));
2040 img_request = obj_request->img_request;
2041 rbd_assert(img_request);
2042 rbd_assert(img_request_write_test(img_request) ||
2043 img_request_discard_test(img_request));
2045 if (img_request_discard_test(img_request))
2048 /* Allocate and initialize the request, for all the ops */
2050 snapc = img_request->snapc;
2051 rbd_dev = img_request->rbd_dev;
2052 osdc = &rbd_dev->rbd_client->client->osdc;
2053 osd_req = ceph_osdc_alloc_request(osdc, snapc, num_osd_ops,
2058 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
2059 osd_req->r_callback = rbd_osd_req_callback;
2060 osd_req->r_priv = obj_request;
2062 osd_req->r_base_oloc.pool = rbd_dev->layout.pool_id;
2063 if (ceph_oid_aprintf(&osd_req->r_base_oid, GFP_NOIO, "%s",
2064 obj_request->object_name))
2067 if (ceph_osdc_alloc_messages(osd_req, GFP_NOIO))
2073 ceph_osdc_put_request(osd_req);
2078 static void rbd_osd_req_destroy(struct ceph_osd_request *osd_req)
2080 ceph_osdc_put_request(osd_req);
2083 /* object_name is assumed to be a non-null pointer and NUL-terminated */
2085 static struct rbd_obj_request *rbd_obj_request_create(const char *object_name,
2086 u64 offset, u64 length,
2087 enum obj_request_type type)
2089 struct rbd_obj_request *obj_request;
2093 rbd_assert(obj_request_type_valid(type));
2095 size = strlen(object_name) + 1;
2096 name = kmalloc(size, GFP_NOIO);
2100 obj_request = kmem_cache_zalloc(rbd_obj_request_cache, GFP_NOIO);
2106 obj_request->object_name = memcpy(name, object_name, size);
2107 obj_request->offset = offset;
2108 obj_request->length = length;
2109 obj_request->flags = 0;
2110 obj_request->which = BAD_WHICH;
2111 obj_request->type = type;
2112 INIT_LIST_HEAD(&obj_request->links);
2113 init_completion(&obj_request->completion);
2114 kref_init(&obj_request->kref);
2116 dout("%s: \"%s\" %llu/%llu %d -> obj %p\n", __func__, object_name,
2117 offset, length, (int)type, obj_request);
2122 static void rbd_obj_request_destroy(struct kref *kref)
2124 struct rbd_obj_request *obj_request;
2126 obj_request = container_of(kref, struct rbd_obj_request, kref);
2128 dout("%s: obj %p\n", __func__, obj_request);
2130 rbd_assert(obj_request->img_request == NULL);
2131 rbd_assert(obj_request->which == BAD_WHICH);
2133 if (obj_request->osd_req)
2134 rbd_osd_req_destroy(obj_request->osd_req);
2136 rbd_assert(obj_request_type_valid(obj_request->type));
2137 switch (obj_request->type) {
2138 case OBJ_REQUEST_NODATA:
2139 break; /* Nothing to do */
2140 case OBJ_REQUEST_BIO:
2141 if (obj_request->bio_list)
2142 bio_chain_put(obj_request->bio_list);
2144 case OBJ_REQUEST_PAGES:
2145 if (obj_request->pages)
2146 ceph_release_page_vector(obj_request->pages,
2147 obj_request->page_count);
2151 kfree(obj_request->object_name);
2152 obj_request->object_name = NULL;
2153 kmem_cache_free(rbd_obj_request_cache, obj_request);
2156 /* It's OK to call this for a device with no parent */
2158 static void rbd_spec_put(struct rbd_spec *spec);
2159 static void rbd_dev_unparent(struct rbd_device *rbd_dev)
2161 rbd_dev_remove_parent(rbd_dev);
2162 rbd_spec_put(rbd_dev->parent_spec);
2163 rbd_dev->parent_spec = NULL;
2164 rbd_dev->parent_overlap = 0;
2168 * Parent image reference counting is used to determine when an
2169 * image's parent fields can be safely torn down--after there are no
2170 * more in-flight requests to the parent image. When the last
2171 * reference is dropped, cleaning them up is safe.
2173 static void rbd_dev_parent_put(struct rbd_device *rbd_dev)
2177 if (!rbd_dev->parent_spec)
2180 counter = atomic_dec_return_safe(&rbd_dev->parent_ref);
2184 /* Last reference; clean up parent data structures */
2187 rbd_dev_unparent(rbd_dev);
2189 rbd_warn(rbd_dev, "parent reference underflow");
2193 * If an image has a non-zero parent overlap, get a reference to its
2196 * Returns true if the rbd device has a parent with a non-zero
2197 * overlap and a reference for it was successfully taken, or
2200 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
2204 if (!rbd_dev->parent_spec)
2207 down_read(&rbd_dev->header_rwsem);
2208 if (rbd_dev->parent_overlap)
2209 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
2210 up_read(&rbd_dev->header_rwsem);
2213 rbd_warn(rbd_dev, "parent reference overflow");
2219 * Caller is responsible for filling in the list of object requests
2220 * that comprises the image request, and the Linux request pointer
2221 * (if there is one).
2223 static struct rbd_img_request *rbd_img_request_create(
2224 struct rbd_device *rbd_dev,
2225 u64 offset, u64 length,
2226 enum obj_operation_type op_type,
2227 struct ceph_snap_context *snapc)
2229 struct rbd_img_request *img_request;
2231 img_request = kmem_cache_alloc(rbd_img_request_cache, GFP_NOIO);
2235 img_request->rq = NULL;
2236 img_request->rbd_dev = rbd_dev;
2237 img_request->offset = offset;
2238 img_request->length = length;
2239 img_request->flags = 0;
2240 if (op_type == OBJ_OP_DISCARD) {
2241 img_request_discard_set(img_request);
2242 img_request->snapc = snapc;
2243 } else if (op_type == OBJ_OP_WRITE) {
2244 img_request_write_set(img_request);
2245 img_request->snapc = snapc;
2247 img_request->snap_id = rbd_dev->spec->snap_id;
2249 if (rbd_dev_parent_get(rbd_dev))
2250 img_request_layered_set(img_request);
2251 spin_lock_init(&img_request->completion_lock);
2252 img_request->next_completion = 0;
2253 img_request->callback = NULL;
2254 img_request->result = 0;
2255 img_request->obj_request_count = 0;
2256 INIT_LIST_HEAD(&img_request->obj_requests);
2257 kref_init(&img_request->kref);
2259 dout("%s: rbd_dev %p %s %llu/%llu -> img %p\n", __func__, rbd_dev,
2260 obj_op_name(op_type), offset, length, img_request);
2265 static void rbd_img_request_destroy(struct kref *kref)
2267 struct rbd_img_request *img_request;
2268 struct rbd_obj_request *obj_request;
2269 struct rbd_obj_request *next_obj_request;
2271 img_request = container_of(kref, struct rbd_img_request, kref);
2273 dout("%s: img %p\n", __func__, img_request);
2275 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2276 rbd_img_obj_request_del(img_request, obj_request);
2277 rbd_assert(img_request->obj_request_count == 0);
2279 if (img_request_layered_test(img_request)) {
2280 img_request_layered_clear(img_request);
2281 rbd_dev_parent_put(img_request->rbd_dev);
2284 if (img_request_write_test(img_request) ||
2285 img_request_discard_test(img_request))
2286 ceph_put_snap_context(img_request->snapc);
2288 kmem_cache_free(rbd_img_request_cache, img_request);
2291 static struct rbd_img_request *rbd_parent_request_create(
2292 struct rbd_obj_request *obj_request,
2293 u64 img_offset, u64 length)
2295 struct rbd_img_request *parent_request;
2296 struct rbd_device *rbd_dev;
2298 rbd_assert(obj_request->img_request);
2299 rbd_dev = obj_request->img_request->rbd_dev;
2301 parent_request = rbd_img_request_create(rbd_dev->parent, img_offset,
2302 length, OBJ_OP_READ, NULL);
2303 if (!parent_request)
2306 img_request_child_set(parent_request);
2307 rbd_obj_request_get(obj_request);
2308 parent_request->obj_request = obj_request;
2310 return parent_request;
2313 static void rbd_parent_request_destroy(struct kref *kref)
2315 struct rbd_img_request *parent_request;
2316 struct rbd_obj_request *orig_request;
2318 parent_request = container_of(kref, struct rbd_img_request, kref);
2319 orig_request = parent_request->obj_request;
2321 parent_request->obj_request = NULL;
2322 rbd_obj_request_put(orig_request);
2323 img_request_child_clear(parent_request);
2325 rbd_img_request_destroy(kref);
2328 static bool rbd_img_obj_end_request(struct rbd_obj_request *obj_request)
2330 struct rbd_img_request *img_request;
2331 unsigned int xferred;
2335 rbd_assert(obj_request_img_data_test(obj_request));
2336 img_request = obj_request->img_request;
2338 rbd_assert(obj_request->xferred <= (u64)UINT_MAX);
2339 xferred = (unsigned int)obj_request->xferred;
2340 result = obj_request->result;
2342 struct rbd_device *rbd_dev = img_request->rbd_dev;
2343 enum obj_operation_type op_type;
2345 if (img_request_discard_test(img_request))
2346 op_type = OBJ_OP_DISCARD;
2347 else if (img_request_write_test(img_request))
2348 op_type = OBJ_OP_WRITE;
2350 op_type = OBJ_OP_READ;
2352 rbd_warn(rbd_dev, "%s %llx at %llx (%llx)",
2353 obj_op_name(op_type), obj_request->length,
2354 obj_request->img_offset, obj_request->offset);
2355 rbd_warn(rbd_dev, " result %d xferred %x",
2357 if (!img_request->result)
2358 img_request->result = result;
2360 * Need to end I/O on the entire obj_request worth of
2361 * bytes in case of error.
2363 xferred = obj_request->length;
2366 /* Image object requests don't own their page array */
2368 if (obj_request->type == OBJ_REQUEST_PAGES) {
2369 obj_request->pages = NULL;
2370 obj_request->page_count = 0;
2373 if (img_request_child_test(img_request)) {
2374 rbd_assert(img_request->obj_request != NULL);
2375 more = obj_request->which < img_request->obj_request_count - 1;
2377 rbd_assert(img_request->rq != NULL);
2379 more = blk_update_request(img_request->rq, result, xferred);
2381 __blk_mq_end_request(img_request->rq, result);
2387 static void rbd_img_obj_callback(struct rbd_obj_request *obj_request)
2389 struct rbd_img_request *img_request;
2390 u32 which = obj_request->which;
2393 rbd_assert(obj_request_img_data_test(obj_request));
2394 img_request = obj_request->img_request;
2396 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
2397 rbd_assert(img_request != NULL);
2398 rbd_assert(img_request->obj_request_count > 0);
2399 rbd_assert(which != BAD_WHICH);
2400 rbd_assert(which < img_request->obj_request_count);
2402 spin_lock_irq(&img_request->completion_lock);
2403 if (which != img_request->next_completion)
2406 for_each_obj_request_from(img_request, obj_request) {
2408 rbd_assert(which < img_request->obj_request_count);
2410 if (!obj_request_done_test(obj_request))
2412 more = rbd_img_obj_end_request(obj_request);
2416 rbd_assert(more ^ (which == img_request->obj_request_count));
2417 img_request->next_completion = which;
2419 spin_unlock_irq(&img_request->completion_lock);
2420 rbd_img_request_put(img_request);
2423 rbd_img_request_complete(img_request);
2427 * Add individual osd ops to the given ceph_osd_request and prepare
2428 * them for submission. num_ops is the current number of
2429 * osd operations already to the object request.
2431 static void rbd_img_obj_request_fill(struct rbd_obj_request *obj_request,
2432 struct ceph_osd_request *osd_request,
2433 enum obj_operation_type op_type,
2434 unsigned int num_ops)
2436 struct rbd_img_request *img_request = obj_request->img_request;
2437 struct rbd_device *rbd_dev = img_request->rbd_dev;
2438 u64 object_size = rbd_obj_bytes(&rbd_dev->header);
2439 u64 offset = obj_request->offset;
2440 u64 length = obj_request->length;
2444 if (op_type == OBJ_OP_DISCARD) {
2445 if (!offset && length == object_size &&
2446 (!img_request_layered_test(img_request) ||
2447 !obj_request_overlaps_parent(obj_request))) {
2448 opcode = CEPH_OSD_OP_DELETE;
2449 } else if ((offset + length == object_size)) {
2450 opcode = CEPH_OSD_OP_TRUNCATE;
2452 down_read(&rbd_dev->header_rwsem);
2453 img_end = rbd_dev->header.image_size;
2454 up_read(&rbd_dev->header_rwsem);
2456 if (obj_request->img_offset + length == img_end)
2457 opcode = CEPH_OSD_OP_TRUNCATE;
2459 opcode = CEPH_OSD_OP_ZERO;
2461 } else if (op_type == OBJ_OP_WRITE) {
2462 if (!offset && length == object_size)
2463 opcode = CEPH_OSD_OP_WRITEFULL;
2465 opcode = CEPH_OSD_OP_WRITE;
2466 osd_req_op_alloc_hint_init(osd_request, num_ops,
2467 object_size, object_size);
2470 opcode = CEPH_OSD_OP_READ;
2473 if (opcode == CEPH_OSD_OP_DELETE)
2474 osd_req_op_init(osd_request, num_ops, opcode, 0);
2476 osd_req_op_extent_init(osd_request, num_ops, opcode,
2477 offset, length, 0, 0);
2479 if (obj_request->type == OBJ_REQUEST_BIO)
2480 osd_req_op_extent_osd_data_bio(osd_request, num_ops,
2481 obj_request->bio_list, length);
2482 else if (obj_request->type == OBJ_REQUEST_PAGES)
2483 osd_req_op_extent_osd_data_pages(osd_request, num_ops,
2484 obj_request->pages, length,
2485 offset & ~PAGE_MASK, false, false);
2487 /* Discards are also writes */
2488 if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
2489 rbd_osd_req_format_write(obj_request);
2491 rbd_osd_req_format_read(obj_request);
2495 * Split up an image request into one or more object requests, each
2496 * to a different object. The "type" parameter indicates whether
2497 * "data_desc" is the pointer to the head of a list of bio
2498 * structures, or the base of a page array. In either case this
2499 * function assumes data_desc describes memory sufficient to hold
2500 * all data described by the image request.
2502 static int rbd_img_request_fill(struct rbd_img_request *img_request,
2503 enum obj_request_type type,
2506 struct rbd_device *rbd_dev = img_request->rbd_dev;
2507 struct rbd_obj_request *obj_request = NULL;
2508 struct rbd_obj_request *next_obj_request;
2509 struct bio *bio_list = NULL;
2510 unsigned int bio_offset = 0;
2511 struct page **pages = NULL;
2512 enum obj_operation_type op_type;
2516 dout("%s: img %p type %d data_desc %p\n", __func__, img_request,
2517 (int)type, data_desc);
2519 img_offset = img_request->offset;
2520 resid = img_request->length;
2521 rbd_assert(resid > 0);
2522 op_type = rbd_img_request_op_type(img_request);
2524 if (type == OBJ_REQUEST_BIO) {
2525 bio_list = data_desc;
2526 rbd_assert(img_offset ==
2527 bio_list->bi_iter.bi_sector << SECTOR_SHIFT);
2528 } else if (type == OBJ_REQUEST_PAGES) {
2533 struct ceph_osd_request *osd_req;
2534 const char *object_name;
2538 object_name = rbd_segment_name(rbd_dev, img_offset);
2541 offset = rbd_segment_offset(rbd_dev, img_offset);
2542 length = rbd_segment_length(rbd_dev, img_offset, resid);
2543 obj_request = rbd_obj_request_create(object_name,
2544 offset, length, type);
2545 /* object request has its own copy of the object name */
2546 rbd_segment_name_free(object_name);
2551 * set obj_request->img_request before creating the
2552 * osd_request so that it gets the right snapc
2554 rbd_img_obj_request_add(img_request, obj_request);
2556 if (type == OBJ_REQUEST_BIO) {
2557 unsigned int clone_size;
2559 rbd_assert(length <= (u64)UINT_MAX);
2560 clone_size = (unsigned int)length;
2561 obj_request->bio_list =
2562 bio_chain_clone_range(&bio_list,
2566 if (!obj_request->bio_list)
2568 } else if (type == OBJ_REQUEST_PAGES) {
2569 unsigned int page_count;
2571 obj_request->pages = pages;
2572 page_count = (u32)calc_pages_for(offset, length);
2573 obj_request->page_count = page_count;
2574 if ((offset + length) & ~PAGE_MASK)
2575 page_count--; /* more on last page */
2576 pages += page_count;
2579 osd_req = rbd_osd_req_create(rbd_dev, op_type,
2580 (op_type == OBJ_OP_WRITE) ? 2 : 1,
2585 obj_request->osd_req = osd_req;
2586 obj_request->callback = rbd_img_obj_callback;
2587 obj_request->img_offset = img_offset;
2589 rbd_img_obj_request_fill(obj_request, osd_req, op_type, 0);
2591 rbd_img_request_get(img_request);
2593 img_offset += length;
2600 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2601 rbd_img_obj_request_del(img_request, obj_request);
2607 rbd_osd_copyup_callback(struct rbd_obj_request *obj_request)
2609 struct rbd_img_request *img_request;
2610 struct rbd_device *rbd_dev;
2611 struct page **pages;
2614 dout("%s: obj %p\n", __func__, obj_request);
2616 rbd_assert(obj_request->type == OBJ_REQUEST_BIO ||
2617 obj_request->type == OBJ_REQUEST_NODATA);
2618 rbd_assert(obj_request_img_data_test(obj_request));
2619 img_request = obj_request->img_request;
2620 rbd_assert(img_request);
2622 rbd_dev = img_request->rbd_dev;
2623 rbd_assert(rbd_dev);
2625 pages = obj_request->copyup_pages;
2626 rbd_assert(pages != NULL);
2627 obj_request->copyup_pages = NULL;
2628 page_count = obj_request->copyup_page_count;
2629 rbd_assert(page_count);
2630 obj_request->copyup_page_count = 0;
2631 ceph_release_page_vector(pages, page_count);
2634 * We want the transfer count to reflect the size of the
2635 * original write request. There is no such thing as a
2636 * successful short write, so if the request was successful
2637 * we can just set it to the originally-requested length.
2639 if (!obj_request->result)
2640 obj_request->xferred = obj_request->length;
2642 obj_request_done_set(obj_request);
2646 rbd_img_obj_parent_read_full_callback(struct rbd_img_request *img_request)
2648 struct rbd_obj_request *orig_request;
2649 struct ceph_osd_request *osd_req;
2650 struct rbd_device *rbd_dev;
2651 struct page **pages;
2652 enum obj_operation_type op_type;
2657 rbd_assert(img_request_child_test(img_request));
2659 /* First get what we need from the image request */
2661 pages = img_request->copyup_pages;
2662 rbd_assert(pages != NULL);
2663 img_request->copyup_pages = NULL;
2664 page_count = img_request->copyup_page_count;
2665 rbd_assert(page_count);
2666 img_request->copyup_page_count = 0;
2668 orig_request = img_request->obj_request;
2669 rbd_assert(orig_request != NULL);
2670 rbd_assert(obj_request_type_valid(orig_request->type));
2671 img_result = img_request->result;
2672 parent_length = img_request->length;
2673 rbd_assert(parent_length == img_request->xferred);
2674 rbd_img_request_put(img_request);
2676 rbd_assert(orig_request->img_request);
2677 rbd_dev = orig_request->img_request->rbd_dev;
2678 rbd_assert(rbd_dev);
2681 * If the overlap has become 0 (most likely because the
2682 * image has been flattened) we need to free the pages
2683 * and re-submit the original write request.
2685 if (!rbd_dev->parent_overlap) {
2686 ceph_release_page_vector(pages, page_count);
2687 rbd_obj_request_submit(orig_request);
2695 * The original osd request is of no use to use any more.
2696 * We need a new one that can hold the three ops in a copyup
2697 * request. Allocate the new copyup osd request for the
2698 * original request, and release the old one.
2700 img_result = -ENOMEM;
2701 osd_req = rbd_osd_req_create_copyup(orig_request);
2704 rbd_osd_req_destroy(orig_request->osd_req);
2705 orig_request->osd_req = osd_req;
2706 orig_request->copyup_pages = pages;
2707 orig_request->copyup_page_count = page_count;
2709 /* Initialize the copyup op */
2711 osd_req_op_cls_init(osd_req, 0, CEPH_OSD_OP_CALL, "rbd", "copyup");
2712 osd_req_op_cls_request_data_pages(osd_req, 0, pages, parent_length, 0,
2715 /* Add the other op(s) */
2717 op_type = rbd_img_request_op_type(orig_request->img_request);
2718 rbd_img_obj_request_fill(orig_request, osd_req, op_type, 1);
2720 /* All set, send it off. */
2722 rbd_obj_request_submit(orig_request);
2726 /* Record the error code and complete the request */
2728 orig_request->result = img_result;
2729 orig_request->xferred = 0;
2730 obj_request_done_set(orig_request);
2731 rbd_obj_request_complete(orig_request);
2735 * Read from the parent image the range of data that covers the
2736 * entire target of the given object request. This is used for
2737 * satisfying a layered image write request when the target of an
2738 * object request from the image request does not exist.
2740 * A page array big enough to hold the returned data is allocated
2741 * and supplied to rbd_img_request_fill() as the "data descriptor."
2742 * When the read completes, this page array will be transferred to
2743 * the original object request for the copyup operation.
2745 * If an error occurs, record it as the result of the original
2746 * object request and mark it done so it gets completed.
2748 static int rbd_img_obj_parent_read_full(struct rbd_obj_request *obj_request)
2750 struct rbd_device *rbd_dev = obj_request->img_request->rbd_dev;
2751 struct rbd_img_request *parent_request = NULL;
2754 struct page **pages = NULL;
2758 rbd_assert(rbd_dev->parent != NULL);
2761 * Determine the byte range covered by the object in the
2762 * child image to which the original request was to be sent.
2764 img_offset = obj_request->img_offset - obj_request->offset;
2765 length = (u64)1 << rbd_dev->header.obj_order;
2768 * There is no defined parent data beyond the parent
2769 * overlap, so limit what we read at that boundary if
2772 if (img_offset + length > rbd_dev->parent_overlap) {
2773 rbd_assert(img_offset < rbd_dev->parent_overlap);
2774 length = rbd_dev->parent_overlap - img_offset;
2778 * Allocate a page array big enough to receive the data read
2781 page_count = (u32)calc_pages_for(0, length);
2782 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2783 if (IS_ERR(pages)) {
2784 result = PTR_ERR(pages);
2790 parent_request = rbd_parent_request_create(obj_request,
2791 img_offset, length);
2792 if (!parent_request)
2795 result = rbd_img_request_fill(parent_request, OBJ_REQUEST_PAGES, pages);
2799 parent_request->copyup_pages = pages;
2800 parent_request->copyup_page_count = page_count;
2801 parent_request->callback = rbd_img_obj_parent_read_full_callback;
2803 result = rbd_img_request_submit(parent_request);
2807 parent_request->copyup_pages = NULL;
2808 parent_request->copyup_page_count = 0;
2809 parent_request->obj_request = NULL;
2810 rbd_obj_request_put(obj_request);
2813 ceph_release_page_vector(pages, page_count);
2815 rbd_img_request_put(parent_request);
2816 obj_request->result = result;
2817 obj_request->xferred = 0;
2818 obj_request_done_set(obj_request);
2823 static void rbd_img_obj_exists_callback(struct rbd_obj_request *obj_request)
2825 struct rbd_obj_request *orig_request;
2826 struct rbd_device *rbd_dev;
2829 rbd_assert(!obj_request_img_data_test(obj_request));
2832 * All we need from the object request is the original
2833 * request and the result of the STAT op. Grab those, then
2834 * we're done with the request.
2836 orig_request = obj_request->obj_request;
2837 obj_request->obj_request = NULL;
2838 rbd_obj_request_put(orig_request);
2839 rbd_assert(orig_request);
2840 rbd_assert(orig_request->img_request);
2842 result = obj_request->result;
2843 obj_request->result = 0;
2845 dout("%s: obj %p for obj %p result %d %llu/%llu\n", __func__,
2846 obj_request, orig_request, result,
2847 obj_request->xferred, obj_request->length);
2848 rbd_obj_request_put(obj_request);
2851 * If the overlap has become 0 (most likely because the
2852 * image has been flattened) we need to re-submit the
2855 rbd_dev = orig_request->img_request->rbd_dev;
2856 if (!rbd_dev->parent_overlap) {
2857 rbd_obj_request_submit(orig_request);
2862 * Our only purpose here is to determine whether the object
2863 * exists, and we don't want to treat the non-existence as
2864 * an error. If something else comes back, transfer the
2865 * error to the original request and complete it now.
2868 obj_request_existence_set(orig_request, true);
2869 } else if (result == -ENOENT) {
2870 obj_request_existence_set(orig_request, false);
2871 } else if (result) {
2872 orig_request->result = result;
2877 * Resubmit the original request now that we have recorded
2878 * whether the target object exists.
2880 orig_request->result = rbd_img_obj_request_submit(orig_request);
2882 if (orig_request->result)
2883 rbd_obj_request_complete(orig_request);
2886 static int rbd_img_obj_exists_submit(struct rbd_obj_request *obj_request)
2888 struct rbd_device *rbd_dev = obj_request->img_request->rbd_dev;
2889 struct rbd_obj_request *stat_request;
2890 struct page **pages = NULL;
2896 * The response data for a STAT call consists of:
2903 size = sizeof (__le64) + sizeof (__le32) + sizeof (__le32);
2904 page_count = (u32)calc_pages_for(0, size);
2905 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2907 return PTR_ERR(pages);
2910 stat_request = rbd_obj_request_create(obj_request->object_name, 0, 0,
2915 rbd_obj_request_get(obj_request);
2916 stat_request->obj_request = obj_request;
2917 stat_request->pages = pages;
2918 stat_request->page_count = page_count;
2920 stat_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
2922 if (!stat_request->osd_req)
2924 stat_request->callback = rbd_img_obj_exists_callback;
2926 osd_req_op_init(stat_request->osd_req, 0, CEPH_OSD_OP_STAT, 0);
2927 osd_req_op_raw_data_in_pages(stat_request->osd_req, 0, pages, size, 0,
2929 rbd_osd_req_format_read(stat_request);
2931 rbd_obj_request_submit(stat_request);
2936 rbd_obj_request_put(obj_request);
2941 static bool img_obj_request_simple(struct rbd_obj_request *obj_request)
2943 struct rbd_img_request *img_request = obj_request->img_request;
2944 struct rbd_device *rbd_dev = img_request->rbd_dev;
2947 if (!img_request_write_test(img_request) &&
2948 !img_request_discard_test(img_request))
2951 /* Non-layered writes */
2952 if (!img_request_layered_test(img_request))
2956 * Layered writes outside of the parent overlap range don't
2957 * share any data with the parent.
2959 if (!obj_request_overlaps_parent(obj_request))
2963 * Entire-object layered writes - we will overwrite whatever
2964 * parent data there is anyway.
2966 if (!obj_request->offset &&
2967 obj_request->length == rbd_obj_bytes(&rbd_dev->header))
2971 * If the object is known to already exist, its parent data has
2972 * already been copied.
2974 if (obj_request_known_test(obj_request) &&
2975 obj_request_exists_test(obj_request))
2981 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request)
2983 rbd_assert(obj_request_img_data_test(obj_request));
2984 rbd_assert(obj_request_type_valid(obj_request->type));
2985 rbd_assert(obj_request->img_request);
2987 if (img_obj_request_simple(obj_request)) {
2988 rbd_obj_request_submit(obj_request);
2993 * It's a layered write. The target object might exist but
2994 * we may not know that yet. If we know it doesn't exist,
2995 * start by reading the data for the full target object from
2996 * the parent so we can use it for a copyup to the target.
2998 if (obj_request_known_test(obj_request))
2999 return rbd_img_obj_parent_read_full(obj_request);
3001 /* We don't know whether the target exists. Go find out. */
3003 return rbd_img_obj_exists_submit(obj_request);
3006 static int rbd_img_request_submit(struct rbd_img_request *img_request)
3008 struct rbd_obj_request *obj_request;
3009 struct rbd_obj_request *next_obj_request;
3012 dout("%s: img %p\n", __func__, img_request);
3014 rbd_img_request_get(img_request);
3015 for_each_obj_request_safe(img_request, obj_request, next_obj_request) {
3016 ret = rbd_img_obj_request_submit(obj_request);
3022 rbd_img_request_put(img_request);
3026 static void rbd_img_parent_read_callback(struct rbd_img_request *img_request)
3028 struct rbd_obj_request *obj_request;
3029 struct rbd_device *rbd_dev;
3034 rbd_assert(img_request_child_test(img_request));
3036 /* First get what we need from the image request and release it */
3038 obj_request = img_request->obj_request;
3039 img_xferred = img_request->xferred;
3040 img_result = img_request->result;
3041 rbd_img_request_put(img_request);
3044 * If the overlap has become 0 (most likely because the
3045 * image has been flattened) we need to re-submit the
3048 rbd_assert(obj_request);
3049 rbd_assert(obj_request->img_request);
3050 rbd_dev = obj_request->img_request->rbd_dev;
3051 if (!rbd_dev->parent_overlap) {
3052 rbd_obj_request_submit(obj_request);
3056 obj_request->result = img_result;
3057 if (obj_request->result)
3061 * We need to zero anything beyond the parent overlap
3062 * boundary. Since rbd_img_obj_request_read_callback()
3063 * will zero anything beyond the end of a short read, an
3064 * easy way to do this is to pretend the data from the
3065 * parent came up short--ending at the overlap boundary.
3067 rbd_assert(obj_request->img_offset < U64_MAX - obj_request->length);
3068 obj_end = obj_request->img_offset + obj_request->length;
3069 if (obj_end > rbd_dev->parent_overlap) {
3072 if (obj_request->img_offset < rbd_dev->parent_overlap)
3073 xferred = rbd_dev->parent_overlap -
3074 obj_request->img_offset;
3076 obj_request->xferred = min(img_xferred, xferred);
3078 obj_request->xferred = img_xferred;
3081 rbd_img_obj_request_read_callback(obj_request);
3082 rbd_obj_request_complete(obj_request);
3085 static void rbd_img_parent_read(struct rbd_obj_request *obj_request)
3087 struct rbd_img_request *img_request;
3090 rbd_assert(obj_request_img_data_test(obj_request));
3091 rbd_assert(obj_request->img_request != NULL);
3092 rbd_assert(obj_request->result == (s32) -ENOENT);
3093 rbd_assert(obj_request_type_valid(obj_request->type));
3095 /* rbd_read_finish(obj_request, obj_request->length); */
3096 img_request = rbd_parent_request_create(obj_request,
3097 obj_request->img_offset,
3098 obj_request->length);
3103 if (obj_request->type == OBJ_REQUEST_BIO)
3104 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3105 obj_request->bio_list);
3107 result = rbd_img_request_fill(img_request, OBJ_REQUEST_PAGES,
3108 obj_request->pages);
3112 img_request->callback = rbd_img_parent_read_callback;
3113 result = rbd_img_request_submit(img_request);
3120 rbd_img_request_put(img_request);
3121 obj_request->result = result;
3122 obj_request->xferred = 0;
3123 obj_request_done_set(obj_request);
3126 static const struct rbd_client_id rbd_empty_cid;
3128 static bool rbd_cid_equal(const struct rbd_client_id *lhs,
3129 const struct rbd_client_id *rhs)
3131 return lhs->gid == rhs->gid && lhs->handle == rhs->handle;
3134 static struct rbd_client_id rbd_get_cid(struct rbd_device *rbd_dev)
3136 struct rbd_client_id cid;
3138 mutex_lock(&rbd_dev->watch_mutex);
3139 cid.gid = ceph_client_gid(rbd_dev->rbd_client->client);
3140 cid.handle = rbd_dev->watch_cookie;
3141 mutex_unlock(&rbd_dev->watch_mutex);
3146 * lock_rwsem must be held for write
3148 static void rbd_set_owner_cid(struct rbd_device *rbd_dev,
3149 const struct rbd_client_id *cid)
3151 dout("%s rbd_dev %p %llu-%llu -> %llu-%llu\n", __func__, rbd_dev,
3152 rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle,
3153 cid->gid, cid->handle);
3154 rbd_dev->owner_cid = *cid; /* struct */
3157 static void format_lock_cookie(struct rbd_device *rbd_dev, char *buf)
3159 mutex_lock(&rbd_dev->watch_mutex);
3160 sprintf(buf, "%s %llu", RBD_LOCK_COOKIE_PREFIX, rbd_dev->watch_cookie);
3161 mutex_unlock(&rbd_dev->watch_mutex);
3165 * lock_rwsem must be held for write
3167 static int rbd_lock(struct rbd_device *rbd_dev)
3169 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3170 struct rbd_client_id cid = rbd_get_cid(rbd_dev);
3174 WARN_ON(__rbd_is_lock_owner(rbd_dev));
3176 format_lock_cookie(rbd_dev, cookie);
3177 ret = ceph_cls_lock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
3178 RBD_LOCK_NAME, CEPH_CLS_LOCK_EXCLUSIVE, cookie,
3179 RBD_LOCK_TAG, "", 0);
3183 rbd_dev->lock_state = RBD_LOCK_STATE_LOCKED;
3184 rbd_set_owner_cid(rbd_dev, &cid);
3185 queue_work(rbd_dev->task_wq, &rbd_dev->acquired_lock_work);
3190 * lock_rwsem must be held for write
3192 static int rbd_unlock(struct rbd_device *rbd_dev)
3194 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3198 WARN_ON(!__rbd_is_lock_owner(rbd_dev));
3200 rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
3202 format_lock_cookie(rbd_dev, cookie);
3203 ret = ceph_cls_unlock(osdc, &rbd_dev->header_oid, &rbd_dev->header_oloc,
3204 RBD_LOCK_NAME, cookie);
3205 if (ret && ret != -ENOENT) {
3206 rbd_warn(rbd_dev, "cls_unlock failed: %d", ret);
3210 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3211 queue_work(rbd_dev->task_wq, &rbd_dev->released_lock_work);
3215 static int __rbd_notify_op_lock(struct rbd_device *rbd_dev,
3216 enum rbd_notify_op notify_op,
3217 struct page ***preply_pages,
3220 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3221 struct rbd_client_id cid = rbd_get_cid(rbd_dev);
3222 int buf_size = 4 + 8 + 8 + CEPH_ENCODING_START_BLK_LEN;
3226 dout("%s rbd_dev %p notify_op %d\n", __func__, rbd_dev, notify_op);
3228 /* encode *LockPayload NotifyMessage (op + ClientId) */
3229 ceph_start_encoding(&p, 2, 1, buf_size - CEPH_ENCODING_START_BLK_LEN);
3230 ceph_encode_32(&p, notify_op);
3231 ceph_encode_64(&p, cid.gid);
3232 ceph_encode_64(&p, cid.handle);
3234 return ceph_osdc_notify(osdc, &rbd_dev->header_oid,
3235 &rbd_dev->header_oloc, buf, buf_size,
3236 RBD_NOTIFY_TIMEOUT, preply_pages, preply_len);
3239 static void rbd_notify_op_lock(struct rbd_device *rbd_dev,
3240 enum rbd_notify_op notify_op)
3242 struct page **reply_pages;
3245 __rbd_notify_op_lock(rbd_dev, notify_op, &reply_pages, &reply_len);
3246 ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len));
3249 static void rbd_notify_acquired_lock(struct work_struct *work)
3251 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3252 acquired_lock_work);
3254 rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_ACQUIRED_LOCK);
3257 static void rbd_notify_released_lock(struct work_struct *work)
3259 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3260 released_lock_work);
3262 rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_RELEASED_LOCK);
3265 static int rbd_request_lock(struct rbd_device *rbd_dev)
3267 struct page **reply_pages;
3269 bool lock_owner_responded = false;
3272 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3274 ret = __rbd_notify_op_lock(rbd_dev, RBD_NOTIFY_OP_REQUEST_LOCK,
3275 &reply_pages, &reply_len);
3276 if (ret && ret != -ETIMEDOUT) {
3277 rbd_warn(rbd_dev, "failed to request lock: %d", ret);
3281 if (reply_len > 0 && reply_len <= PAGE_SIZE) {
3282 void *p = page_address(reply_pages[0]);
3283 void *const end = p + reply_len;
3286 ceph_decode_32_safe(&p, end, n, e_inval); /* num_acks */
3291 ceph_decode_need(&p, end, 8 + 8, e_inval);
3292 p += 8 + 8; /* skip gid and cookie */
3294 ceph_decode_32_safe(&p, end, len, e_inval);
3298 if (lock_owner_responded) {
3300 "duplicate lock owners detected");
3305 lock_owner_responded = true;
3306 ret = ceph_start_decoding(&p, end, 1, "ResponseMessage",
3310 "failed to decode ResponseMessage: %d",
3315 ret = ceph_decode_32(&p);
3319 if (!lock_owner_responded) {
3320 rbd_warn(rbd_dev, "no lock owners detected");
3325 ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len));
3333 static void wake_requests(struct rbd_device *rbd_dev, bool wake_all)
3335 dout("%s rbd_dev %p wake_all %d\n", __func__, rbd_dev, wake_all);
3337 cancel_delayed_work(&rbd_dev->lock_dwork);
3339 wake_up_all(&rbd_dev->lock_waitq);
3341 wake_up(&rbd_dev->lock_waitq);
3344 static int get_lock_owner_info(struct rbd_device *rbd_dev,
3345 struct ceph_locker **lockers, u32 *num_lockers)
3347 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3352 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3354 ret = ceph_cls_lock_info(osdc, &rbd_dev->header_oid,
3355 &rbd_dev->header_oloc, RBD_LOCK_NAME,
3356 &lock_type, &lock_tag, lockers, num_lockers);
3360 if (*num_lockers == 0) {
3361 dout("%s rbd_dev %p no lockers detected\n", __func__, rbd_dev);
3365 if (strcmp(lock_tag, RBD_LOCK_TAG)) {
3366 rbd_warn(rbd_dev, "locked by external mechanism, tag %s",
3372 if (lock_type == CEPH_CLS_LOCK_SHARED) {
3373 rbd_warn(rbd_dev, "shared lock type detected");
3378 if (strncmp((*lockers)[0].id.cookie, RBD_LOCK_COOKIE_PREFIX,
3379 strlen(RBD_LOCK_COOKIE_PREFIX))) {
3380 rbd_warn(rbd_dev, "locked by external mechanism, cookie %s",
3381 (*lockers)[0].id.cookie);
3391 static int find_watcher(struct rbd_device *rbd_dev,
3392 const struct ceph_locker *locker)
3394 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3395 struct ceph_watch_item *watchers;
3401 ret = ceph_osdc_list_watchers(osdc, &rbd_dev->header_oid,
3402 &rbd_dev->header_oloc, &watchers,
3407 sscanf(locker->id.cookie, RBD_LOCK_COOKIE_PREFIX " %llu", &cookie);
3408 for (i = 0; i < num_watchers; i++) {
3409 if (!memcmp(&watchers[i].addr, &locker->info.addr,
3410 sizeof(locker->info.addr)) &&
3411 watchers[i].cookie == cookie) {
3412 struct rbd_client_id cid = {
3413 .gid = le64_to_cpu(watchers[i].name.num),
3417 dout("%s rbd_dev %p found cid %llu-%llu\n", __func__,
3418 rbd_dev, cid.gid, cid.handle);
3419 rbd_set_owner_cid(rbd_dev, &cid);
3425 dout("%s rbd_dev %p no watchers\n", __func__, rbd_dev);
3433 * lock_rwsem must be held for write
3435 static int rbd_try_lock(struct rbd_device *rbd_dev)
3437 struct ceph_client *client = rbd_dev->rbd_client->client;
3438 struct ceph_locker *lockers;
3443 ret = rbd_lock(rbd_dev);
3447 /* determine if the current lock holder is still alive */
3448 ret = get_lock_owner_info(rbd_dev, &lockers, &num_lockers);
3452 if (num_lockers == 0)
3455 ret = find_watcher(rbd_dev, lockers);
3458 ret = 0; /* have to request lock */
3462 rbd_warn(rbd_dev, "%s%llu seems dead, breaking lock",
3463 ENTITY_NAME(lockers[0].id.name));
3465 ret = ceph_monc_blacklist_add(&client->monc,
3466 &lockers[0].info.addr);
3468 rbd_warn(rbd_dev, "blacklist of %s%llu failed: %d",
3469 ENTITY_NAME(lockers[0].id.name), ret);
3473 ret = ceph_cls_break_lock(&client->osdc, &rbd_dev->header_oid,
3474 &rbd_dev->header_oloc, RBD_LOCK_NAME,
3475 lockers[0].id.cookie,
3476 &lockers[0].id.name);
3477 if (ret && ret != -ENOENT)
3481 ceph_free_lockers(lockers, num_lockers);
3485 ceph_free_lockers(lockers, num_lockers);
3490 * ret is set only if lock_state is RBD_LOCK_STATE_UNLOCKED
3492 static enum rbd_lock_state rbd_try_acquire_lock(struct rbd_device *rbd_dev,
3495 enum rbd_lock_state lock_state;
3497 down_read(&rbd_dev->lock_rwsem);
3498 dout("%s rbd_dev %p read lock_state %d\n", __func__, rbd_dev,
3499 rbd_dev->lock_state);
3500 if (__rbd_is_lock_owner(rbd_dev)) {
3501 lock_state = rbd_dev->lock_state;
3502 up_read(&rbd_dev->lock_rwsem);
3506 up_read(&rbd_dev->lock_rwsem);
3507 down_write(&rbd_dev->lock_rwsem);
3508 dout("%s rbd_dev %p write lock_state %d\n", __func__, rbd_dev,
3509 rbd_dev->lock_state);
3510 if (!__rbd_is_lock_owner(rbd_dev)) {
3511 *pret = rbd_try_lock(rbd_dev);
3513 rbd_warn(rbd_dev, "failed to acquire lock: %d", *pret);
3516 lock_state = rbd_dev->lock_state;
3517 up_write(&rbd_dev->lock_rwsem);
3521 static void rbd_acquire_lock(struct work_struct *work)
3523 struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
3524 struct rbd_device, lock_dwork);
3525 enum rbd_lock_state lock_state;
3528 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3530 lock_state = rbd_try_acquire_lock(rbd_dev, &ret);
3531 if (lock_state != RBD_LOCK_STATE_UNLOCKED || ret == -EBLACKLISTED) {
3532 if (lock_state == RBD_LOCK_STATE_LOCKED)
3533 wake_requests(rbd_dev, true);
3534 dout("%s rbd_dev %p lock_state %d ret %d - done\n", __func__,
3535 rbd_dev, lock_state, ret);
3539 ret = rbd_request_lock(rbd_dev);
3540 if (ret == -ETIMEDOUT) {
3541 goto again; /* treat this as a dead client */
3542 } else if (ret < 0) {
3543 rbd_warn(rbd_dev, "error requesting lock: %d", ret);
3544 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
3548 * lock owner acked, but resend if we don't see them
3551 dout("%s rbd_dev %p requeueing lock_dwork\n", __func__,
3553 mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork,
3554 msecs_to_jiffies(2 * RBD_NOTIFY_TIMEOUT * MSEC_PER_SEC));
3559 * lock_rwsem must be held for write
3561 static bool rbd_release_lock(struct rbd_device *rbd_dev)
3563 dout("%s rbd_dev %p read lock_state %d\n", __func__, rbd_dev,
3564 rbd_dev->lock_state);
3565 if (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED)
3568 rbd_dev->lock_state = RBD_LOCK_STATE_RELEASING;
3569 downgrade_write(&rbd_dev->lock_rwsem);
3571 * Ensure that all in-flight IO is flushed.
3573 * FIXME: ceph_osdc_sync() flushes the entire OSD client, which
3574 * may be shared with other devices.
3576 ceph_osdc_sync(&rbd_dev->rbd_client->client->osdc);
3577 up_read(&rbd_dev->lock_rwsem);
3579 down_write(&rbd_dev->lock_rwsem);
3580 dout("%s rbd_dev %p write lock_state %d\n", __func__, rbd_dev,
3581 rbd_dev->lock_state);
3582 if (rbd_dev->lock_state != RBD_LOCK_STATE_RELEASING)
3585 if (!rbd_unlock(rbd_dev))
3587 * Give others a chance to grab the lock - we would re-acquire
3588 * almost immediately if we got new IO during ceph_osdc_sync()
3589 * otherwise. We need to ack our own notifications, so this
3590 * lock_dwork will be requeued from rbd_wait_state_locked()
3591 * after wake_requests() in rbd_handle_released_lock().
3593 cancel_delayed_work(&rbd_dev->lock_dwork);
3598 static void rbd_release_lock_work(struct work_struct *work)
3600 struct rbd_device *rbd_dev = container_of(work, struct rbd_device,
3603 down_write(&rbd_dev->lock_rwsem);
3604 rbd_release_lock(rbd_dev);
3605 up_write(&rbd_dev->lock_rwsem);
3608 static void rbd_handle_acquired_lock(struct rbd_device *rbd_dev, u8 struct_v,
3611 struct rbd_client_id cid = { 0 };
3613 if (struct_v >= 2) {
3614 cid.gid = ceph_decode_64(p);
3615 cid.handle = ceph_decode_64(p);
3618 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3620 if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
3621 down_write(&rbd_dev->lock_rwsem);
3622 if (rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
3624 * we already know that the remote client is
3627 up_write(&rbd_dev->lock_rwsem);
3631 rbd_set_owner_cid(rbd_dev, &cid);
3632 downgrade_write(&rbd_dev->lock_rwsem);
3634 down_read(&rbd_dev->lock_rwsem);
3637 if (!__rbd_is_lock_owner(rbd_dev))
3638 wake_requests(rbd_dev, false);
3639 up_read(&rbd_dev->lock_rwsem);
3642 static void rbd_handle_released_lock(struct rbd_device *rbd_dev, u8 struct_v,
3645 struct rbd_client_id cid = { 0 };
3647 if (struct_v >= 2) {
3648 cid.gid = ceph_decode_64(p);
3649 cid.handle = ceph_decode_64(p);
3652 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3654 if (!rbd_cid_equal(&cid, &rbd_empty_cid)) {
3655 down_write(&rbd_dev->lock_rwsem);
3656 if (!rbd_cid_equal(&cid, &rbd_dev->owner_cid)) {
3657 dout("%s rbd_dev %p unexpected owner, cid %llu-%llu != owner_cid %llu-%llu\n",
3658 __func__, rbd_dev, cid.gid, cid.handle,
3659 rbd_dev->owner_cid.gid, rbd_dev->owner_cid.handle);
3660 up_write(&rbd_dev->lock_rwsem);
3664 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3665 downgrade_write(&rbd_dev->lock_rwsem);
3667 down_read(&rbd_dev->lock_rwsem);
3670 if (!__rbd_is_lock_owner(rbd_dev))
3671 wake_requests(rbd_dev, false);
3672 up_read(&rbd_dev->lock_rwsem);
3675 static bool rbd_handle_request_lock(struct rbd_device *rbd_dev, u8 struct_v,
3678 struct rbd_client_id my_cid = rbd_get_cid(rbd_dev);
3679 struct rbd_client_id cid = { 0 };
3682 if (struct_v >= 2) {
3683 cid.gid = ceph_decode_64(p);
3684 cid.handle = ceph_decode_64(p);
3687 dout("%s rbd_dev %p cid %llu-%llu\n", __func__, rbd_dev, cid.gid,
3689 if (rbd_cid_equal(&cid, &my_cid))
3692 down_read(&rbd_dev->lock_rwsem);
3693 need_to_send = __rbd_is_lock_owner(rbd_dev);
3694 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED) {
3695 if (!rbd_cid_equal(&rbd_dev->owner_cid, &rbd_empty_cid)) {
3696 dout("%s rbd_dev %p queueing unlock_work\n", __func__,
3698 queue_work(rbd_dev->task_wq, &rbd_dev->unlock_work);
3701 up_read(&rbd_dev->lock_rwsem);
3702 return need_to_send;
3705 static void __rbd_acknowledge_notify(struct rbd_device *rbd_dev,
3706 u64 notify_id, u64 cookie, s32 *result)
3708 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3709 int buf_size = 4 + CEPH_ENCODING_START_BLK_LEN;
3716 /* encode ResponseMessage */
3717 ceph_start_encoding(&p, 1, 1,
3718 buf_size - CEPH_ENCODING_START_BLK_LEN);
3719 ceph_encode_32(&p, *result);
3724 ret = ceph_osdc_notify_ack(osdc, &rbd_dev->header_oid,
3725 &rbd_dev->header_oloc, notify_id, cookie,
3728 rbd_warn(rbd_dev, "acknowledge_notify failed: %d", ret);
3731 static void rbd_acknowledge_notify(struct rbd_device *rbd_dev, u64 notify_id,
3734 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3735 __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, NULL);
3738 static void rbd_acknowledge_notify_result(struct rbd_device *rbd_dev,
3739 u64 notify_id, u64 cookie, s32 result)
3741 dout("%s rbd_dev %p result %d\n", __func__, rbd_dev, result);
3742 __rbd_acknowledge_notify(rbd_dev, notify_id, cookie, &result);
3745 static void rbd_watch_cb(void *arg, u64 notify_id, u64 cookie,
3746 u64 notifier_id, void *data, size_t data_len)
3748 struct rbd_device *rbd_dev = arg;
3750 void *const end = p + data_len;
3756 dout("%s rbd_dev %p cookie %llu notify_id %llu data_len %zu\n",
3757 __func__, rbd_dev, cookie, notify_id, data_len);
3759 ret = ceph_start_decoding(&p, end, 1, "NotifyMessage",
3762 rbd_warn(rbd_dev, "failed to decode NotifyMessage: %d",
3767 notify_op = ceph_decode_32(&p);
3769 /* legacy notification for header updates */
3770 notify_op = RBD_NOTIFY_OP_HEADER_UPDATE;
3774 dout("%s rbd_dev %p notify_op %u\n", __func__, rbd_dev, notify_op);
3775 switch (notify_op) {
3776 case RBD_NOTIFY_OP_ACQUIRED_LOCK:
3777 rbd_handle_acquired_lock(rbd_dev, struct_v, &p);
3778 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3780 case RBD_NOTIFY_OP_RELEASED_LOCK:
3781 rbd_handle_released_lock(rbd_dev, struct_v, &p);
3782 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3784 case RBD_NOTIFY_OP_REQUEST_LOCK:
3785 if (rbd_handle_request_lock(rbd_dev, struct_v, &p))
3787 * send ResponseMessage(0) back so the client
3788 * can detect a missing owner
3790 rbd_acknowledge_notify_result(rbd_dev, notify_id,
3793 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3795 case RBD_NOTIFY_OP_HEADER_UPDATE:
3796 ret = rbd_dev_refresh(rbd_dev);
3798 rbd_warn(rbd_dev, "refresh failed: %d", ret);
3800 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3803 if (rbd_is_lock_owner(rbd_dev))
3804 rbd_acknowledge_notify_result(rbd_dev, notify_id,
3805 cookie, -EOPNOTSUPP);
3807 rbd_acknowledge_notify(rbd_dev, notify_id, cookie);
3812 static void __rbd_unregister_watch(struct rbd_device *rbd_dev);
3814 static void rbd_watch_errcb(void *arg, u64 cookie, int err)
3816 struct rbd_device *rbd_dev = arg;
3818 rbd_warn(rbd_dev, "encountered watch error: %d", err);
3820 down_write(&rbd_dev->lock_rwsem);
3821 rbd_set_owner_cid(rbd_dev, &rbd_empty_cid);
3822 up_write(&rbd_dev->lock_rwsem);
3824 mutex_lock(&rbd_dev->watch_mutex);
3825 if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED) {
3826 __rbd_unregister_watch(rbd_dev);
3827 rbd_dev->watch_state = RBD_WATCH_STATE_ERROR;
3829 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->watch_dwork, 0);
3831 mutex_unlock(&rbd_dev->watch_mutex);
3835 * watch_mutex must be locked
3837 static int __rbd_register_watch(struct rbd_device *rbd_dev)
3839 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3840 struct ceph_osd_linger_request *handle;
3842 rbd_assert(!rbd_dev->watch_handle);
3843 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3845 handle = ceph_osdc_watch(osdc, &rbd_dev->header_oid,
3846 &rbd_dev->header_oloc, rbd_watch_cb,
3847 rbd_watch_errcb, rbd_dev);
3849 return PTR_ERR(handle);
3851 rbd_dev->watch_handle = handle;
3856 * watch_mutex must be locked
3858 static void __rbd_unregister_watch(struct rbd_device *rbd_dev)
3860 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3863 rbd_assert(rbd_dev->watch_handle);
3864 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3866 ret = ceph_osdc_unwatch(osdc, rbd_dev->watch_handle);
3868 rbd_warn(rbd_dev, "failed to unwatch: %d", ret);
3870 rbd_dev->watch_handle = NULL;
3873 static int rbd_register_watch(struct rbd_device *rbd_dev)
3877 mutex_lock(&rbd_dev->watch_mutex);
3878 rbd_assert(rbd_dev->watch_state == RBD_WATCH_STATE_UNREGISTERED);
3879 ret = __rbd_register_watch(rbd_dev);
3883 rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
3884 rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
3887 mutex_unlock(&rbd_dev->watch_mutex);
3891 static void cancel_tasks_sync(struct rbd_device *rbd_dev)
3893 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3895 cancel_delayed_work_sync(&rbd_dev->watch_dwork);
3896 cancel_work_sync(&rbd_dev->acquired_lock_work);
3897 cancel_work_sync(&rbd_dev->released_lock_work);
3898 cancel_delayed_work_sync(&rbd_dev->lock_dwork);
3899 cancel_work_sync(&rbd_dev->unlock_work);
3902 static void rbd_unregister_watch(struct rbd_device *rbd_dev)
3904 WARN_ON(waitqueue_active(&rbd_dev->lock_waitq));
3905 cancel_tasks_sync(rbd_dev);
3907 mutex_lock(&rbd_dev->watch_mutex);
3908 if (rbd_dev->watch_state == RBD_WATCH_STATE_REGISTERED)
3909 __rbd_unregister_watch(rbd_dev);
3910 rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
3911 mutex_unlock(&rbd_dev->watch_mutex);
3913 ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
3916 static void rbd_reregister_watch(struct work_struct *work)
3918 struct rbd_device *rbd_dev = container_of(to_delayed_work(work),
3919 struct rbd_device, watch_dwork);
3920 bool was_lock_owner = false;
3923 dout("%s rbd_dev %p\n", __func__, rbd_dev);
3925 down_write(&rbd_dev->lock_rwsem);
3926 if (rbd_dev->lock_state == RBD_LOCK_STATE_LOCKED)
3927 was_lock_owner = rbd_release_lock(rbd_dev);
3929 mutex_lock(&rbd_dev->watch_mutex);
3930 if (rbd_dev->watch_state != RBD_WATCH_STATE_ERROR)
3933 ret = __rbd_register_watch(rbd_dev);
3935 rbd_warn(rbd_dev, "failed to reregister watch: %d", ret);
3936 if (ret != -EBLACKLISTED)
3937 queue_delayed_work(rbd_dev->task_wq,
3938 &rbd_dev->watch_dwork,
3943 rbd_dev->watch_state = RBD_WATCH_STATE_REGISTERED;
3944 rbd_dev->watch_cookie = rbd_dev->watch_handle->linger_id;
3945 mutex_unlock(&rbd_dev->watch_mutex);
3947 ret = rbd_dev_refresh(rbd_dev);
3949 rbd_warn(rbd_dev, "reregisteration refresh failed: %d", ret);
3951 if (was_lock_owner) {
3952 ret = rbd_try_lock(rbd_dev);
3954 rbd_warn(rbd_dev, "reregisteration lock failed: %d",
3958 up_write(&rbd_dev->lock_rwsem);
3959 wake_requests(rbd_dev, true);
3963 mutex_unlock(&rbd_dev->watch_mutex);
3964 up_write(&rbd_dev->lock_rwsem);
3968 * Synchronous osd object method call. Returns the number of bytes
3969 * returned in the outbound buffer, or a negative error code.
3971 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
3972 const char *object_name,
3973 const char *class_name,
3974 const char *method_name,
3975 const void *outbound,
3976 size_t outbound_size,
3978 size_t inbound_size)
3980 struct rbd_obj_request *obj_request;
3981 struct page **pages;
3986 * Method calls are ultimately read operations. The result
3987 * should placed into the inbound buffer provided. They
3988 * also supply outbound data--parameters for the object
3989 * method. Currently if this is present it will be a
3992 page_count = (u32)calc_pages_for(0, inbound_size);
3993 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
3995 return PTR_ERR(pages);
3998 obj_request = rbd_obj_request_create(object_name, 0, inbound_size,
4003 obj_request->pages = pages;
4004 obj_request->page_count = page_count;
4006 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
4008 if (!obj_request->osd_req)
4011 osd_req_op_cls_init(obj_request->osd_req, 0, CEPH_OSD_OP_CALL,
4012 class_name, method_name);
4013 if (outbound_size) {
4014 struct ceph_pagelist *pagelist;
4016 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
4020 ceph_pagelist_init(pagelist);
4021 ceph_pagelist_append(pagelist, outbound, outbound_size);
4022 osd_req_op_cls_request_data_pagelist(obj_request->osd_req, 0,
4025 osd_req_op_cls_response_data_pages(obj_request->osd_req, 0,
4026 obj_request->pages, inbound_size,
4028 rbd_osd_req_format_read(obj_request);
4030 rbd_obj_request_submit(obj_request);
4031 ret = rbd_obj_request_wait(obj_request);
4035 ret = obj_request->result;
4039 rbd_assert(obj_request->xferred < (u64)INT_MAX);
4040 ret = (int)obj_request->xferred;
4041 ceph_copy_from_page_vector(pages, inbound, 0, obj_request->xferred);
4044 rbd_obj_request_put(obj_request);
4046 ceph_release_page_vector(pages, page_count);
4052 * lock_rwsem must be held for read
4054 static void rbd_wait_state_locked(struct rbd_device *rbd_dev)
4060 * Note the use of mod_delayed_work() in rbd_acquire_lock()
4061 * and cancel_delayed_work() in wake_requests().
4063 dout("%s rbd_dev %p queueing lock_dwork\n", __func__, rbd_dev);
4064 queue_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork, 0);
4065 prepare_to_wait_exclusive(&rbd_dev->lock_waitq, &wait,
4066 TASK_UNINTERRUPTIBLE);
4067 up_read(&rbd_dev->lock_rwsem);
4069 down_read(&rbd_dev->lock_rwsem);
4070 } while (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED);
4071 finish_wait(&rbd_dev->lock_waitq, &wait);
4074 static void rbd_queue_workfn(struct work_struct *work)
4076 struct request *rq = blk_mq_rq_from_pdu(work);
4077 struct rbd_device *rbd_dev = rq->q->queuedata;
4078 struct rbd_img_request *img_request;
4079 struct ceph_snap_context *snapc = NULL;
4080 u64 offset = (u64)blk_rq_pos(rq) << SECTOR_SHIFT;
4081 u64 length = blk_rq_bytes(rq);
4082 enum obj_operation_type op_type;
4084 bool must_be_locked;
4087 if (rq->cmd_type != REQ_TYPE_FS) {
4088 dout("%s: non-fs request type %d\n", __func__,
4089 (int) rq->cmd_type);
4094 if (req_op(rq) == REQ_OP_DISCARD)
4095 op_type = OBJ_OP_DISCARD;
4096 else if (req_op(rq) == REQ_OP_WRITE)
4097 op_type = OBJ_OP_WRITE;
4099 op_type = OBJ_OP_READ;
4101 /* Ignore/skip any zero-length requests */
4104 dout("%s: zero-length request\n", __func__);
4109 /* Only reads are allowed to a read-only device */
4111 if (op_type != OBJ_OP_READ) {
4112 if (rbd_dev->mapping.read_only) {
4116 rbd_assert(rbd_dev->spec->snap_id == CEPH_NOSNAP);
4120 * Quit early if the mapped snapshot no longer exists. It's
4121 * still possible the snapshot will have disappeared by the
4122 * time our request arrives at the osd, but there's no sense in
4123 * sending it if we already know.
4125 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
4126 dout("request for non-existent snapshot");
4127 rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
4132 if (offset && length > U64_MAX - offset + 1) {
4133 rbd_warn(rbd_dev, "bad request range (%llu~%llu)", offset,
4136 goto err_rq; /* Shouldn't happen */
4139 blk_mq_start_request(rq);
4141 down_read(&rbd_dev->header_rwsem);
4142 mapping_size = rbd_dev->mapping.size;
4143 if (op_type != OBJ_OP_READ) {
4144 snapc = rbd_dev->header.snapc;
4145 ceph_get_snap_context(snapc);
4146 must_be_locked = rbd_is_lock_supported(rbd_dev);
4148 must_be_locked = rbd_dev->opts->lock_on_read &&
4149 rbd_is_lock_supported(rbd_dev);
4151 up_read(&rbd_dev->header_rwsem);
4153 if (offset + length > mapping_size) {
4154 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)", offset,
4155 length, mapping_size);
4160 if (must_be_locked) {
4161 down_read(&rbd_dev->lock_rwsem);
4162 if (rbd_dev->lock_state != RBD_LOCK_STATE_LOCKED)
4163 rbd_wait_state_locked(rbd_dev);
4166 img_request = rbd_img_request_create(rbd_dev, offset, length, op_type,
4172 img_request->rq = rq;
4173 snapc = NULL; /* img_request consumes a ref */
4175 if (op_type == OBJ_OP_DISCARD)
4176 result = rbd_img_request_fill(img_request, OBJ_REQUEST_NODATA,
4179 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
4182 goto err_img_request;
4184 result = rbd_img_request_submit(img_request);
4186 goto err_img_request;
4189 up_read(&rbd_dev->lock_rwsem);
4193 rbd_img_request_put(img_request);
4196 up_read(&rbd_dev->lock_rwsem);
4199 rbd_warn(rbd_dev, "%s %llx at %llx result %d",
4200 obj_op_name(op_type), length, offset, result);
4201 ceph_put_snap_context(snapc);
4203 blk_mq_end_request(rq, result);
4206 static int rbd_queue_rq(struct blk_mq_hw_ctx *hctx,
4207 const struct blk_mq_queue_data *bd)
4209 struct request *rq = bd->rq;
4210 struct work_struct *work = blk_mq_rq_to_pdu(rq);
4212 queue_work(rbd_wq, work);
4213 return BLK_MQ_RQ_QUEUE_OK;
4216 static void rbd_free_disk(struct rbd_device *rbd_dev)
4218 struct gendisk *disk = rbd_dev->disk;
4223 rbd_dev->disk = NULL;
4224 if (disk->flags & GENHD_FL_UP) {
4227 blk_cleanup_queue(disk->queue);
4228 blk_mq_free_tag_set(&rbd_dev->tag_set);
4233 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
4234 const char *object_name,
4235 u64 offset, u64 length, void *buf)
4238 struct rbd_obj_request *obj_request;
4239 struct page **pages = NULL;
4244 page_count = (u32) calc_pages_for(offset, length);
4245 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
4247 return PTR_ERR(pages);
4250 obj_request = rbd_obj_request_create(object_name, offset, length,
4255 obj_request->pages = pages;
4256 obj_request->page_count = page_count;
4258 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
4260 if (!obj_request->osd_req)
4263 osd_req_op_extent_init(obj_request->osd_req, 0, CEPH_OSD_OP_READ,
4264 offset, length, 0, 0);
4265 osd_req_op_extent_osd_data_pages(obj_request->osd_req, 0,
4267 obj_request->length,
4268 obj_request->offset & ~PAGE_MASK,
4270 rbd_osd_req_format_read(obj_request);
4272 rbd_obj_request_submit(obj_request);
4273 ret = rbd_obj_request_wait(obj_request);
4277 ret = obj_request->result;
4281 rbd_assert(obj_request->xferred <= (u64) SIZE_MAX);
4282 size = (size_t) obj_request->xferred;
4283 ceph_copy_from_page_vector(pages, buf, 0, size);
4284 rbd_assert(size <= (size_t)INT_MAX);
4288 rbd_obj_request_put(obj_request);
4290 ceph_release_page_vector(pages, page_count);
4296 * Read the complete header for the given rbd device. On successful
4297 * return, the rbd_dev->header field will contain up-to-date
4298 * information about the image.
4300 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev)
4302 struct rbd_image_header_ondisk *ondisk = NULL;
4309 * The complete header will include an array of its 64-bit
4310 * snapshot ids, followed by the names of those snapshots as
4311 * a contiguous block of NUL-terminated strings. Note that
4312 * the number of snapshots could change by the time we read
4313 * it in, in which case we re-read it.
4320 size = sizeof (*ondisk);
4321 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
4323 ondisk = kmalloc(size, GFP_KERNEL);
4327 ret = rbd_obj_read_sync(rbd_dev, rbd_dev->header_oid.name,
4331 if ((size_t)ret < size) {
4333 rbd_warn(rbd_dev, "short header read (want %zd got %d)",
4337 if (!rbd_dev_ondisk_valid(ondisk)) {
4339 rbd_warn(rbd_dev, "invalid header");
4343 names_size = le64_to_cpu(ondisk->snap_names_len);
4344 want_count = snap_count;
4345 snap_count = le32_to_cpu(ondisk->snap_count);
4346 } while (snap_count != want_count);
4348 ret = rbd_header_from_disk(rbd_dev, ondisk);
4356 * Clear the rbd device's EXISTS flag if the snapshot it's mapped to
4357 * has disappeared from the (just updated) snapshot context.
4359 static void rbd_exists_validate(struct rbd_device *rbd_dev)
4363 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags))
4366 snap_id = rbd_dev->spec->snap_id;
4367 if (snap_id == CEPH_NOSNAP)
4370 if (rbd_dev_snap_index(rbd_dev, snap_id) == BAD_SNAP_INDEX)
4371 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
4374 static void rbd_dev_update_size(struct rbd_device *rbd_dev)
4379 * If EXISTS is not set, rbd_dev->disk may be NULL, so don't
4380 * try to update its size. If REMOVING is set, updating size
4381 * is just useless work since the device can't be opened.
4383 if (test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags) &&
4384 !test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags)) {
4385 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
4386 dout("setting size to %llu sectors", (unsigned long long)size);
4387 set_capacity(rbd_dev->disk, size);
4388 revalidate_disk(rbd_dev->disk);
4392 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
4397 down_write(&rbd_dev->header_rwsem);
4398 mapping_size = rbd_dev->mapping.size;
4400 ret = rbd_dev_header_info(rbd_dev);
4405 * If there is a parent, see if it has disappeared due to the
4406 * mapped image getting flattened.
4408 if (rbd_dev->parent) {
4409 ret = rbd_dev_v2_parent_info(rbd_dev);
4414 if (rbd_dev->spec->snap_id == CEPH_NOSNAP) {
4415 rbd_dev->mapping.size = rbd_dev->header.image_size;
4417 /* validate mapped snapshot's EXISTS flag */
4418 rbd_exists_validate(rbd_dev);
4422 up_write(&rbd_dev->header_rwsem);
4423 if (!ret && mapping_size != rbd_dev->mapping.size)
4424 rbd_dev_update_size(rbd_dev);
4429 static int rbd_init_request(void *data, struct request *rq,
4430 unsigned int hctx_idx, unsigned int request_idx,
4431 unsigned int numa_node)
4433 struct work_struct *work = blk_mq_rq_to_pdu(rq);
4435 INIT_WORK(work, rbd_queue_workfn);
4439 static struct blk_mq_ops rbd_mq_ops = {
4440 .queue_rq = rbd_queue_rq,
4441 .map_queue = blk_mq_map_queue,
4442 .init_request = rbd_init_request,
4445 static int rbd_init_disk(struct rbd_device *rbd_dev)
4447 struct gendisk *disk;
4448 struct request_queue *q;
4452 /* create gendisk info */
4453 disk = alloc_disk(single_major ?
4454 (1 << RBD_SINGLE_MAJOR_PART_SHIFT) :
4455 RBD_MINORS_PER_MAJOR);
4459 snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
4461 disk->major = rbd_dev->major;
4462 disk->first_minor = rbd_dev->minor;
4464 disk->flags |= GENHD_FL_EXT_DEVT;
4465 disk->fops = &rbd_bd_ops;
4466 disk->private_data = rbd_dev;
4468 memset(&rbd_dev->tag_set, 0, sizeof(rbd_dev->tag_set));
4469 rbd_dev->tag_set.ops = &rbd_mq_ops;
4470 rbd_dev->tag_set.queue_depth = rbd_dev->opts->queue_depth;
4471 rbd_dev->tag_set.numa_node = NUMA_NO_NODE;
4472 rbd_dev->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
4473 rbd_dev->tag_set.nr_hw_queues = 1;
4474 rbd_dev->tag_set.cmd_size = sizeof(struct work_struct);
4476 err = blk_mq_alloc_tag_set(&rbd_dev->tag_set);
4480 q = blk_mq_init_queue(&rbd_dev->tag_set);
4486 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
4487 /* QUEUE_FLAG_ADD_RANDOM is off by default for blk-mq */
4489 /* set io sizes to object size */
4490 segment_size = rbd_obj_bytes(&rbd_dev->header);
4491 blk_queue_max_hw_sectors(q, segment_size / SECTOR_SIZE);
4492 q->limits.max_sectors = queue_max_hw_sectors(q);
4493 blk_queue_max_segments(q, segment_size / SECTOR_SIZE);
4494 blk_queue_max_segment_size(q, segment_size);
4495 blk_queue_io_min(q, segment_size);
4496 blk_queue_io_opt(q, segment_size);
4498 /* enable the discard support */
4499 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
4500 q->limits.discard_granularity = segment_size;
4501 q->limits.discard_alignment = segment_size;
4502 blk_queue_max_discard_sectors(q, segment_size / SECTOR_SIZE);
4503 q->limits.discard_zeroes_data = 1;
4505 if (!ceph_test_opt(rbd_dev->rbd_client->client, NOCRC))
4506 q->backing_dev_info.capabilities |= BDI_CAP_STABLE_WRITES;
4510 q->queuedata = rbd_dev;
4512 rbd_dev->disk = disk;
4516 blk_mq_free_tag_set(&rbd_dev->tag_set);
4526 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
4528 return container_of(dev, struct rbd_device, dev);
4531 static ssize_t rbd_size_show(struct device *dev,
4532 struct device_attribute *attr, char *buf)
4534 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4536 return sprintf(buf, "%llu\n",
4537 (unsigned long long)rbd_dev->mapping.size);
4541 * Note this shows the features for whatever's mapped, which is not
4542 * necessarily the base image.
4544 static ssize_t rbd_features_show(struct device *dev,
4545 struct device_attribute *attr, char *buf)
4547 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4549 return sprintf(buf, "0x%016llx\n",
4550 (unsigned long long)rbd_dev->mapping.features);
4553 static ssize_t rbd_major_show(struct device *dev,
4554 struct device_attribute *attr, char *buf)
4556 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4559 return sprintf(buf, "%d\n", rbd_dev->major);
4561 return sprintf(buf, "(none)\n");
4564 static ssize_t rbd_minor_show(struct device *dev,
4565 struct device_attribute *attr, char *buf)
4567 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4569 return sprintf(buf, "%d\n", rbd_dev->minor);
4572 static ssize_t rbd_client_addr_show(struct device *dev,
4573 struct device_attribute *attr, char *buf)
4575 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4576 struct ceph_entity_addr *client_addr =
4577 ceph_client_addr(rbd_dev->rbd_client->client);
4579 return sprintf(buf, "%pISpc/%u\n", &client_addr->in_addr,
4580 le32_to_cpu(client_addr->nonce));
4583 static ssize_t rbd_client_id_show(struct device *dev,
4584 struct device_attribute *attr, char *buf)
4586 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4588 return sprintf(buf, "client%lld\n",
4589 ceph_client_gid(rbd_dev->rbd_client->client));
4592 static ssize_t rbd_cluster_fsid_show(struct device *dev,
4593 struct device_attribute *attr, char *buf)
4595 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4597 return sprintf(buf, "%pU\n", &rbd_dev->rbd_client->client->fsid);
4600 static ssize_t rbd_config_info_show(struct device *dev,
4601 struct device_attribute *attr, char *buf)
4603 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4605 return sprintf(buf, "%s\n", rbd_dev->config_info);
4608 static ssize_t rbd_pool_show(struct device *dev,
4609 struct device_attribute *attr, char *buf)
4611 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4613 return sprintf(buf, "%s\n", rbd_dev->spec->pool_name);
4616 static ssize_t rbd_pool_id_show(struct device *dev,
4617 struct device_attribute *attr, char *buf)
4619 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4621 return sprintf(buf, "%llu\n",
4622 (unsigned long long) rbd_dev->spec->pool_id);
4625 static ssize_t rbd_name_show(struct device *dev,
4626 struct device_attribute *attr, char *buf)
4628 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4630 if (rbd_dev->spec->image_name)
4631 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
4633 return sprintf(buf, "(unknown)\n");
4636 static ssize_t rbd_image_id_show(struct device *dev,
4637 struct device_attribute *attr, char *buf)
4639 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4641 return sprintf(buf, "%s\n", rbd_dev->spec->image_id);
4645 * Shows the name of the currently-mapped snapshot (or
4646 * RBD_SNAP_HEAD_NAME for the base image).
4648 static ssize_t rbd_snap_show(struct device *dev,
4649 struct device_attribute *attr,
4652 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4654 return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
4657 static ssize_t rbd_snap_id_show(struct device *dev,
4658 struct device_attribute *attr, char *buf)
4660 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4662 return sprintf(buf, "%llu\n", rbd_dev->spec->snap_id);
4666 * For a v2 image, shows the chain of parent images, separated by empty
4667 * lines. For v1 images or if there is no parent, shows "(no parent
4670 static ssize_t rbd_parent_show(struct device *dev,
4671 struct device_attribute *attr,
4674 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4677 if (!rbd_dev->parent)
4678 return sprintf(buf, "(no parent image)\n");
4680 for ( ; rbd_dev->parent; rbd_dev = rbd_dev->parent) {
4681 struct rbd_spec *spec = rbd_dev->parent_spec;
4683 count += sprintf(&buf[count], "%s"
4684 "pool_id %llu\npool_name %s\n"
4685 "image_id %s\nimage_name %s\n"
4686 "snap_id %llu\nsnap_name %s\n"
4688 !count ? "" : "\n", /* first? */
4689 spec->pool_id, spec->pool_name,
4690 spec->image_id, spec->image_name ?: "(unknown)",
4691 spec->snap_id, spec->snap_name,
4692 rbd_dev->parent_overlap);
4698 static ssize_t rbd_image_refresh(struct device *dev,
4699 struct device_attribute *attr,
4703 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4706 ret = rbd_dev_refresh(rbd_dev);
4713 static DEVICE_ATTR(size, S_IRUGO, rbd_size_show, NULL);
4714 static DEVICE_ATTR(features, S_IRUGO, rbd_features_show, NULL);
4715 static DEVICE_ATTR(major, S_IRUGO, rbd_major_show, NULL);
4716 static DEVICE_ATTR(minor, S_IRUGO, rbd_minor_show, NULL);
4717 static DEVICE_ATTR(client_addr, S_IRUGO, rbd_client_addr_show, NULL);
4718 static DEVICE_ATTR(client_id, S_IRUGO, rbd_client_id_show, NULL);
4719 static DEVICE_ATTR(cluster_fsid, S_IRUGO, rbd_cluster_fsid_show, NULL);
4720 static DEVICE_ATTR(config_info, S_IRUSR, rbd_config_info_show, NULL);
4721 static DEVICE_ATTR(pool, S_IRUGO, rbd_pool_show, NULL);
4722 static DEVICE_ATTR(pool_id, S_IRUGO, rbd_pool_id_show, NULL);
4723 static DEVICE_ATTR(name, S_IRUGO, rbd_name_show, NULL);
4724 static DEVICE_ATTR(image_id, S_IRUGO, rbd_image_id_show, NULL);
4725 static DEVICE_ATTR(refresh, S_IWUSR, NULL, rbd_image_refresh);
4726 static DEVICE_ATTR(current_snap, S_IRUGO, rbd_snap_show, NULL);
4727 static DEVICE_ATTR(snap_id, S_IRUGO, rbd_snap_id_show, NULL);
4728 static DEVICE_ATTR(parent, S_IRUGO, rbd_parent_show, NULL);
4730 static struct attribute *rbd_attrs[] = {
4731 &dev_attr_size.attr,
4732 &dev_attr_features.attr,
4733 &dev_attr_major.attr,
4734 &dev_attr_minor.attr,
4735 &dev_attr_client_addr.attr,
4736 &dev_attr_client_id.attr,
4737 &dev_attr_cluster_fsid.attr,
4738 &dev_attr_config_info.attr,
4739 &dev_attr_pool.attr,
4740 &dev_attr_pool_id.attr,
4741 &dev_attr_name.attr,
4742 &dev_attr_image_id.attr,
4743 &dev_attr_current_snap.attr,
4744 &dev_attr_snap_id.attr,
4745 &dev_attr_parent.attr,
4746 &dev_attr_refresh.attr,
4750 static struct attribute_group rbd_attr_group = {
4754 static const struct attribute_group *rbd_attr_groups[] = {
4759 static void rbd_dev_release(struct device *dev);
4761 static struct device_type rbd_device_type = {
4763 .groups = rbd_attr_groups,
4764 .release = rbd_dev_release,
4767 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
4769 kref_get(&spec->kref);
4774 static void rbd_spec_free(struct kref *kref);
4775 static void rbd_spec_put(struct rbd_spec *spec)
4778 kref_put(&spec->kref, rbd_spec_free);
4781 static struct rbd_spec *rbd_spec_alloc(void)
4783 struct rbd_spec *spec;
4785 spec = kzalloc(sizeof (*spec), GFP_KERNEL);
4789 spec->pool_id = CEPH_NOPOOL;
4790 spec->snap_id = CEPH_NOSNAP;
4791 kref_init(&spec->kref);
4796 static void rbd_spec_free(struct kref *kref)
4798 struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
4800 kfree(spec->pool_name);
4801 kfree(spec->image_id);
4802 kfree(spec->image_name);
4803 kfree(spec->snap_name);
4807 static void rbd_dev_free(struct rbd_device *rbd_dev)
4809 WARN_ON(rbd_dev->watch_state != RBD_WATCH_STATE_UNREGISTERED);
4810 WARN_ON(rbd_dev->lock_state != RBD_LOCK_STATE_UNLOCKED);
4812 ceph_oid_destroy(&rbd_dev->header_oid);
4813 ceph_oloc_destroy(&rbd_dev->header_oloc);
4814 kfree(rbd_dev->config_info);
4816 rbd_put_client(rbd_dev->rbd_client);
4817 rbd_spec_put(rbd_dev->spec);
4818 kfree(rbd_dev->opts);
4822 static void rbd_dev_release(struct device *dev)
4824 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4825 bool need_put = !!rbd_dev->opts;
4828 destroy_workqueue(rbd_dev->task_wq);
4829 ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4832 rbd_dev_free(rbd_dev);
4835 * This is racy, but way better than putting module outside of
4836 * the release callback. The race window is pretty small, so
4837 * doing something similar to dm (dm-builtin.c) is overkill.
4840 module_put(THIS_MODULE);
4843 static struct rbd_device *__rbd_dev_create(struct rbd_client *rbdc,
4844 struct rbd_spec *spec)
4846 struct rbd_device *rbd_dev;
4848 rbd_dev = kzalloc(sizeof(*rbd_dev), GFP_KERNEL);
4852 spin_lock_init(&rbd_dev->lock);
4853 INIT_LIST_HEAD(&rbd_dev->node);
4854 init_rwsem(&rbd_dev->header_rwsem);
4856 ceph_oid_init(&rbd_dev->header_oid);
4857 ceph_oloc_init(&rbd_dev->header_oloc);
4859 mutex_init(&rbd_dev->watch_mutex);
4860 rbd_dev->watch_state = RBD_WATCH_STATE_UNREGISTERED;
4861 INIT_DELAYED_WORK(&rbd_dev->watch_dwork, rbd_reregister_watch);
4863 init_rwsem(&rbd_dev->lock_rwsem);
4864 rbd_dev->lock_state = RBD_LOCK_STATE_UNLOCKED;
4865 INIT_WORK(&rbd_dev->acquired_lock_work, rbd_notify_acquired_lock);
4866 INIT_WORK(&rbd_dev->released_lock_work, rbd_notify_released_lock);
4867 INIT_DELAYED_WORK(&rbd_dev->lock_dwork, rbd_acquire_lock);
4868 INIT_WORK(&rbd_dev->unlock_work, rbd_release_lock_work);
4869 init_waitqueue_head(&rbd_dev->lock_waitq);
4871 rbd_dev->dev.bus = &rbd_bus_type;
4872 rbd_dev->dev.type = &rbd_device_type;
4873 rbd_dev->dev.parent = &rbd_root_dev;
4874 device_initialize(&rbd_dev->dev);
4876 rbd_dev->rbd_client = rbdc;
4877 rbd_dev->spec = spec;
4879 rbd_dev->layout.stripe_unit = 1 << RBD_MAX_OBJ_ORDER;
4880 rbd_dev->layout.stripe_count = 1;
4881 rbd_dev->layout.object_size = 1 << RBD_MAX_OBJ_ORDER;
4882 rbd_dev->layout.pool_id = spec->pool_id;
4883 RCU_INIT_POINTER(rbd_dev->layout.pool_ns, NULL);
4889 * Create a mapping rbd_dev.
4891 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
4892 struct rbd_spec *spec,
4893 struct rbd_options *opts)
4895 struct rbd_device *rbd_dev;
4897 rbd_dev = __rbd_dev_create(rbdc, spec);
4901 rbd_dev->opts = opts;
4903 /* get an id and fill in device name */
4904 rbd_dev->dev_id = ida_simple_get(&rbd_dev_id_ida, 0,
4905 minor_to_rbd_dev_id(1 << MINORBITS),
4907 if (rbd_dev->dev_id < 0)
4910 sprintf(rbd_dev->name, RBD_DRV_NAME "%d", rbd_dev->dev_id);
4911 rbd_dev->task_wq = alloc_ordered_workqueue("%s-tasks", WQ_MEM_RECLAIM,
4913 if (!rbd_dev->task_wq)
4916 /* we have a ref from do_rbd_add() */
4917 __module_get(THIS_MODULE);
4919 dout("%s rbd_dev %p dev_id %d\n", __func__, rbd_dev, rbd_dev->dev_id);
4923 ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4925 rbd_dev_free(rbd_dev);
4929 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
4932 put_device(&rbd_dev->dev);
4936 * Get the size and object order for an image snapshot, or if
4937 * snap_id is CEPH_NOSNAP, gets this information for the base
4940 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
4941 u8 *order, u64 *snap_size)
4943 __le64 snapid = cpu_to_le64(snap_id);
4948 } __attribute__ ((packed)) size_buf = { 0 };
4950 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
4952 &snapid, sizeof (snapid),
4953 &size_buf, sizeof (size_buf));
4954 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4957 if (ret < sizeof (size_buf))
4961 *order = size_buf.order;
4962 dout(" order %u", (unsigned int)*order);
4964 *snap_size = le64_to_cpu(size_buf.size);
4966 dout(" snap_id 0x%016llx snap_size = %llu\n",
4967 (unsigned long long)snap_id,
4968 (unsigned long long)*snap_size);
4973 static int rbd_dev_v2_image_size(struct rbd_device *rbd_dev)
4975 return _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
4976 &rbd_dev->header.obj_order,
4977 &rbd_dev->header.image_size);
4980 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev)
4986 reply_buf = kzalloc(RBD_OBJ_PREFIX_LEN_MAX, GFP_KERNEL);
4990 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
4991 "rbd", "get_object_prefix", NULL, 0,
4992 reply_buf, RBD_OBJ_PREFIX_LEN_MAX);
4993 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4998 rbd_dev->header.object_prefix = ceph_extract_encoded_string(&p,
4999 p + ret, NULL, GFP_NOIO);
5002 if (IS_ERR(rbd_dev->header.object_prefix)) {
5003 ret = PTR_ERR(rbd_dev->header.object_prefix);
5004 rbd_dev->header.object_prefix = NULL;
5006 dout(" object_prefix = %s\n", rbd_dev->header.object_prefix);
5014 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
5017 __le64 snapid = cpu_to_le64(snap_id);
5021 } __attribute__ ((packed)) features_buf = { 0 };
5025 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5026 "rbd", "get_features",
5027 &snapid, sizeof (snapid),
5028 &features_buf, sizeof (features_buf));
5029 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5032 if (ret < sizeof (features_buf))
5035 unsup = le64_to_cpu(features_buf.incompat) & ~RBD_FEATURES_SUPPORTED;
5037 rbd_warn(rbd_dev, "image uses unsupported features: 0x%llx",
5042 *snap_features = le64_to_cpu(features_buf.features);
5044 dout(" snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
5045 (unsigned long long)snap_id,
5046 (unsigned long long)*snap_features,
5047 (unsigned long long)le64_to_cpu(features_buf.incompat));
5052 static int rbd_dev_v2_features(struct rbd_device *rbd_dev)
5054 return _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
5055 &rbd_dev->header.features);
5058 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev)
5060 struct rbd_spec *parent_spec;
5062 void *reply_buf = NULL;
5072 parent_spec = rbd_spec_alloc();
5076 size = sizeof (__le64) + /* pool_id */
5077 sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX + /* image_id */
5078 sizeof (__le64) + /* snap_id */
5079 sizeof (__le64); /* overlap */
5080 reply_buf = kmalloc(size, GFP_KERNEL);
5086 snapid = cpu_to_le64(rbd_dev->spec->snap_id);
5087 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5088 "rbd", "get_parent",
5089 &snapid, sizeof (snapid),
5091 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5096 end = reply_buf + ret;
5098 ceph_decode_64_safe(&p, end, pool_id, out_err);
5099 if (pool_id == CEPH_NOPOOL) {
5101 * Either the parent never existed, or we have
5102 * record of it but the image got flattened so it no
5103 * longer has a parent. When the parent of a
5104 * layered image disappears we immediately set the
5105 * overlap to 0. The effect of this is that all new
5106 * requests will be treated as if the image had no
5109 if (rbd_dev->parent_overlap) {
5110 rbd_dev->parent_overlap = 0;
5111 rbd_dev_parent_put(rbd_dev);
5112 pr_info("%s: clone image has been flattened\n",
5113 rbd_dev->disk->disk_name);
5116 goto out; /* No parent? No problem. */
5119 /* The ceph file layout needs to fit pool id in 32 bits */
5122 if (pool_id > (u64)U32_MAX) {
5123 rbd_warn(NULL, "parent pool id too large (%llu > %u)",
5124 (unsigned long long)pool_id, U32_MAX);
5128 image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
5129 if (IS_ERR(image_id)) {
5130 ret = PTR_ERR(image_id);
5133 ceph_decode_64_safe(&p, end, snap_id, out_err);
5134 ceph_decode_64_safe(&p, end, overlap, out_err);
5137 * The parent won't change (except when the clone is
5138 * flattened, already handled that). So we only need to
5139 * record the parent spec we have not already done so.
5141 if (!rbd_dev->parent_spec) {
5142 parent_spec->pool_id = pool_id;
5143 parent_spec->image_id = image_id;
5144 parent_spec->snap_id = snap_id;
5145 rbd_dev->parent_spec = parent_spec;
5146 parent_spec = NULL; /* rbd_dev now owns this */
5152 * We always update the parent overlap. If it's zero we issue
5153 * a warning, as we will proceed as if there was no parent.
5157 /* refresh, careful to warn just once */
5158 if (rbd_dev->parent_overlap)
5160 "clone now standalone (overlap became 0)");
5163 rbd_warn(rbd_dev, "clone is standalone (overlap 0)");
5166 rbd_dev->parent_overlap = overlap;
5172 rbd_spec_put(parent_spec);
5177 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev)
5181 __le64 stripe_count;
5182 } __attribute__ ((packed)) striping_info_buf = { 0 };
5183 size_t size = sizeof (striping_info_buf);
5190 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5191 "rbd", "get_stripe_unit_count", NULL, 0,
5192 (char *)&striping_info_buf, size);
5193 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5200 * We don't actually support the "fancy striping" feature
5201 * (STRIPINGV2) yet, but if the striping sizes are the
5202 * defaults the behavior is the same as before. So find
5203 * out, and only fail if the image has non-default values.
5206 obj_size = (u64)1 << rbd_dev->header.obj_order;
5207 p = &striping_info_buf;
5208 stripe_unit = ceph_decode_64(&p);
5209 if (stripe_unit != obj_size) {
5210 rbd_warn(rbd_dev, "unsupported stripe unit "
5211 "(got %llu want %llu)",
5212 stripe_unit, obj_size);
5215 stripe_count = ceph_decode_64(&p);
5216 if (stripe_count != 1) {
5217 rbd_warn(rbd_dev, "unsupported stripe count "
5218 "(got %llu want 1)", stripe_count);
5221 rbd_dev->header.stripe_unit = stripe_unit;
5222 rbd_dev->header.stripe_count = stripe_count;
5227 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
5229 size_t image_id_size;
5234 void *reply_buf = NULL;
5236 char *image_name = NULL;
5239 rbd_assert(!rbd_dev->spec->image_name);
5241 len = strlen(rbd_dev->spec->image_id);
5242 image_id_size = sizeof (__le32) + len;
5243 image_id = kmalloc(image_id_size, GFP_KERNEL);
5248 end = image_id + image_id_size;
5249 ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
5251 size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
5252 reply_buf = kmalloc(size, GFP_KERNEL);
5256 ret = rbd_obj_method_sync(rbd_dev, RBD_DIRECTORY,
5257 "rbd", "dir_get_name",
5258 image_id, image_id_size,
5263 end = reply_buf + ret;
5265 image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
5266 if (IS_ERR(image_name))
5269 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
5277 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5279 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
5280 const char *snap_name;
5283 /* Skip over names until we find the one we are looking for */
5285 snap_name = rbd_dev->header.snap_names;
5286 while (which < snapc->num_snaps) {
5287 if (!strcmp(name, snap_name))
5288 return snapc->snaps[which];
5289 snap_name += strlen(snap_name) + 1;
5295 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5297 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
5302 for (which = 0; !found && which < snapc->num_snaps; which++) {
5303 const char *snap_name;
5305 snap_id = snapc->snaps[which];
5306 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
5307 if (IS_ERR(snap_name)) {
5308 /* ignore no-longer existing snapshots */
5309 if (PTR_ERR(snap_name) == -ENOENT)
5314 found = !strcmp(name, snap_name);
5317 return found ? snap_id : CEPH_NOSNAP;
5321 * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
5322 * no snapshot by that name is found, or if an error occurs.
5324 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
5326 if (rbd_dev->image_format == 1)
5327 return rbd_v1_snap_id_by_name(rbd_dev, name);
5329 return rbd_v2_snap_id_by_name(rbd_dev, name);
5333 * An image being mapped will have everything but the snap id.
5335 static int rbd_spec_fill_snap_id(struct rbd_device *rbd_dev)
5337 struct rbd_spec *spec = rbd_dev->spec;
5339 rbd_assert(spec->pool_id != CEPH_NOPOOL && spec->pool_name);
5340 rbd_assert(spec->image_id && spec->image_name);
5341 rbd_assert(spec->snap_name);
5343 if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
5346 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
5347 if (snap_id == CEPH_NOSNAP)
5350 spec->snap_id = snap_id;
5352 spec->snap_id = CEPH_NOSNAP;
5359 * A parent image will have all ids but none of the names.
5361 * All names in an rbd spec are dynamically allocated. It's OK if we
5362 * can't figure out the name for an image id.
5364 static int rbd_spec_fill_names(struct rbd_device *rbd_dev)
5366 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
5367 struct rbd_spec *spec = rbd_dev->spec;
5368 const char *pool_name;
5369 const char *image_name;
5370 const char *snap_name;
5373 rbd_assert(spec->pool_id != CEPH_NOPOOL);
5374 rbd_assert(spec->image_id);
5375 rbd_assert(spec->snap_id != CEPH_NOSNAP);
5377 /* Get the pool name; we have to make our own copy of this */
5379 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
5381 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
5384 pool_name = kstrdup(pool_name, GFP_KERNEL);
5388 /* Fetch the image name; tolerate failure here */
5390 image_name = rbd_dev_image_name(rbd_dev);
5392 rbd_warn(rbd_dev, "unable to get image name");
5394 /* Fetch the snapshot name */
5396 snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
5397 if (IS_ERR(snap_name)) {
5398 ret = PTR_ERR(snap_name);
5402 spec->pool_name = pool_name;
5403 spec->image_name = image_name;
5404 spec->snap_name = snap_name;
5414 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev)
5423 struct ceph_snap_context *snapc;
5427 * We'll need room for the seq value (maximum snapshot id),
5428 * snapshot count, and array of that many snapshot ids.
5429 * For now we have a fixed upper limit on the number we're
5430 * prepared to receive.
5432 size = sizeof (__le64) + sizeof (__le32) +
5433 RBD_MAX_SNAP_COUNT * sizeof (__le64);
5434 reply_buf = kzalloc(size, GFP_KERNEL);
5438 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5439 "rbd", "get_snapcontext", NULL, 0,
5441 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5446 end = reply_buf + ret;
5448 ceph_decode_64_safe(&p, end, seq, out);
5449 ceph_decode_32_safe(&p, end, snap_count, out);
5452 * Make sure the reported number of snapshot ids wouldn't go
5453 * beyond the end of our buffer. But before checking that,
5454 * make sure the computed size of the snapshot context we
5455 * allocate is representable in a size_t.
5457 if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
5462 if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
5466 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
5472 for (i = 0; i < snap_count; i++)
5473 snapc->snaps[i] = ceph_decode_64(&p);
5475 ceph_put_snap_context(rbd_dev->header.snapc);
5476 rbd_dev->header.snapc = snapc;
5478 dout(" snap context seq = %llu, snap_count = %u\n",
5479 (unsigned long long)seq, (unsigned int)snap_count);
5486 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
5497 size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
5498 reply_buf = kmalloc(size, GFP_KERNEL);
5500 return ERR_PTR(-ENOMEM);
5502 snapid = cpu_to_le64(snap_id);
5503 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
5504 "rbd", "get_snapshot_name",
5505 &snapid, sizeof (snapid),
5507 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5509 snap_name = ERR_PTR(ret);
5514 end = reply_buf + ret;
5515 snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
5516 if (IS_ERR(snap_name))
5519 dout(" snap_id 0x%016llx snap_name = %s\n",
5520 (unsigned long long)snap_id, snap_name);
5527 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev)
5529 bool first_time = rbd_dev->header.object_prefix == NULL;
5532 ret = rbd_dev_v2_image_size(rbd_dev);
5537 ret = rbd_dev_v2_header_onetime(rbd_dev);
5542 ret = rbd_dev_v2_snap_context(rbd_dev);
5543 if (ret && first_time) {
5544 kfree(rbd_dev->header.object_prefix);
5545 rbd_dev->header.object_prefix = NULL;
5551 static int rbd_dev_header_info(struct rbd_device *rbd_dev)
5553 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
5555 if (rbd_dev->image_format == 1)
5556 return rbd_dev_v1_header_info(rbd_dev);
5558 return rbd_dev_v2_header_info(rbd_dev);
5562 * Skips over white space at *buf, and updates *buf to point to the
5563 * first found non-space character (if any). Returns the length of
5564 * the token (string of non-white space characters) found. Note
5565 * that *buf must be terminated with '\0'.
5567 static inline size_t next_token(const char **buf)
5570 * These are the characters that produce nonzero for
5571 * isspace() in the "C" and "POSIX" locales.
5573 const char *spaces = " \f\n\r\t\v";
5575 *buf += strspn(*buf, spaces); /* Find start of token */
5577 return strcspn(*buf, spaces); /* Return token length */
5581 * Finds the next token in *buf, dynamically allocates a buffer big
5582 * enough to hold a copy of it, and copies the token into the new
5583 * buffer. The copy is guaranteed to be terminated with '\0'. Note
5584 * that a duplicate buffer is created even for a zero-length token.
5586 * Returns a pointer to the newly-allocated duplicate, or a null
5587 * pointer if memory for the duplicate was not available. If
5588 * the lenp argument is a non-null pointer, the length of the token
5589 * (not including the '\0') is returned in *lenp.
5591 * If successful, the *buf pointer will be updated to point beyond
5592 * the end of the found token.
5594 * Note: uses GFP_KERNEL for allocation.
5596 static inline char *dup_token(const char **buf, size_t *lenp)
5601 len = next_token(buf);
5602 dup = kmemdup(*buf, len + 1, GFP_KERNEL);
5605 *(dup + len) = '\0';
5615 * Parse the options provided for an "rbd add" (i.e., rbd image
5616 * mapping) request. These arrive via a write to /sys/bus/rbd/add,
5617 * and the data written is passed here via a NUL-terminated buffer.
5618 * Returns 0 if successful or an error code otherwise.
5620 * The information extracted from these options is recorded in
5621 * the other parameters which return dynamically-allocated
5624 * The address of a pointer that will refer to a ceph options
5625 * structure. Caller must release the returned pointer using
5626 * ceph_destroy_options() when it is no longer needed.
5628 * Address of an rbd options pointer. Fully initialized by
5629 * this function; caller must release with kfree().
5631 * Address of an rbd image specification pointer. Fully
5632 * initialized by this function based on parsed options.
5633 * Caller must release with rbd_spec_put().
5635 * The options passed take this form:
5636 * <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
5639 * A comma-separated list of one or more monitor addresses.
5640 * A monitor address is an ip address, optionally followed
5641 * by a port number (separated by a colon).
5642 * I.e.: ip1[:port1][,ip2[:port2]...]
5644 * A comma-separated list of ceph and/or rbd options.
5646 * The name of the rados pool containing the rbd image.
5648 * The name of the image in that pool to map.
5650 * An optional snapshot id. If provided, the mapping will
5651 * present data from the image at the time that snapshot was
5652 * created. The image head is used if no snapshot id is
5653 * provided. Snapshot mappings are always read-only.
5655 static int rbd_add_parse_args(const char *buf,
5656 struct ceph_options **ceph_opts,
5657 struct rbd_options **opts,
5658 struct rbd_spec **rbd_spec)
5662 const char *mon_addrs;
5664 size_t mon_addrs_size;
5665 struct rbd_spec *spec = NULL;
5666 struct rbd_options *rbd_opts = NULL;
5667 struct ceph_options *copts;
5670 /* The first four tokens are required */
5672 len = next_token(&buf);
5674 rbd_warn(NULL, "no monitor address(es) provided");
5678 mon_addrs_size = len + 1;
5682 options = dup_token(&buf, NULL);
5686 rbd_warn(NULL, "no options provided");
5690 spec = rbd_spec_alloc();
5694 spec->pool_name = dup_token(&buf, NULL);
5695 if (!spec->pool_name)
5697 if (!*spec->pool_name) {
5698 rbd_warn(NULL, "no pool name provided");
5702 spec->image_name = dup_token(&buf, NULL);
5703 if (!spec->image_name)
5705 if (!*spec->image_name) {
5706 rbd_warn(NULL, "no image name provided");
5711 * Snapshot name is optional; default is to use "-"
5712 * (indicating the head/no snapshot).
5714 len = next_token(&buf);
5716 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
5717 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
5718 } else if (len > RBD_MAX_SNAP_NAME_LEN) {
5719 ret = -ENAMETOOLONG;
5722 snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
5725 *(snap_name + len) = '\0';
5726 spec->snap_name = snap_name;
5728 /* Initialize all rbd options to the defaults */
5730 rbd_opts = kzalloc(sizeof (*rbd_opts), GFP_KERNEL);
5734 rbd_opts->read_only = RBD_READ_ONLY_DEFAULT;
5735 rbd_opts->queue_depth = RBD_QUEUE_DEPTH_DEFAULT;
5736 rbd_opts->lock_on_read = RBD_LOCK_ON_READ_DEFAULT;
5738 copts = ceph_parse_options(options, mon_addrs,
5739 mon_addrs + mon_addrs_size - 1,
5740 parse_rbd_opts_token, rbd_opts);
5741 if (IS_ERR(copts)) {
5742 ret = PTR_ERR(copts);
5763 * Return pool id (>= 0) or a negative error code.
5765 static int rbd_add_get_pool_id(struct rbd_client *rbdc, const char *pool_name)
5767 struct ceph_options *opts = rbdc->client->options;
5773 ret = ceph_pg_poolid_by_name(rbdc->client->osdc.osdmap, pool_name);
5774 if (ret == -ENOENT && tries++ < 1) {
5775 ret = ceph_monc_get_version(&rbdc->client->monc, "osdmap",
5780 if (rbdc->client->osdc.osdmap->epoch < newest_epoch) {
5781 ceph_osdc_maybe_request_map(&rbdc->client->osdc);
5782 (void) ceph_monc_wait_osdmap(&rbdc->client->monc,
5784 opts->mount_timeout);
5787 /* the osdmap we have is new enough */
5796 * An rbd format 2 image has a unique identifier, distinct from the
5797 * name given to it by the user. Internally, that identifier is
5798 * what's used to specify the names of objects related to the image.
5800 * A special "rbd id" object is used to map an rbd image name to its
5801 * id. If that object doesn't exist, then there is no v2 rbd image
5802 * with the supplied name.
5804 * This function will record the given rbd_dev's image_id field if
5805 * it can be determined, and in that case will return 0. If any
5806 * errors occur a negative errno will be returned and the rbd_dev's
5807 * image_id field will be unchanged (and should be NULL).
5809 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
5818 * When probing a parent image, the image id is already
5819 * known (and the image name likely is not). There's no
5820 * need to fetch the image id again in this case. We
5821 * do still need to set the image format though.
5823 if (rbd_dev->spec->image_id) {
5824 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
5830 * First, see if the format 2 image id file exists, and if
5831 * so, get the image's persistent id from it.
5833 size = sizeof (RBD_ID_PREFIX) + strlen(rbd_dev->spec->image_name);
5834 object_name = kmalloc(size, GFP_NOIO);
5837 sprintf(object_name, "%s%s", RBD_ID_PREFIX, rbd_dev->spec->image_name);
5838 dout("rbd id object name is %s\n", object_name);
5840 /* Response will be an encoded string, which includes a length */
5842 size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
5843 response = kzalloc(size, GFP_NOIO);
5849 /* If it doesn't exist we'll assume it's a format 1 image */
5851 ret = rbd_obj_method_sync(rbd_dev, object_name,
5852 "rbd", "get_id", NULL, 0,
5853 response, RBD_IMAGE_ID_LEN_MAX);
5854 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5855 if (ret == -ENOENT) {
5856 image_id = kstrdup("", GFP_KERNEL);
5857 ret = image_id ? 0 : -ENOMEM;
5859 rbd_dev->image_format = 1;
5860 } else if (ret >= 0) {
5863 image_id = ceph_extract_encoded_string(&p, p + ret,
5865 ret = PTR_ERR_OR_ZERO(image_id);
5867 rbd_dev->image_format = 2;
5871 rbd_dev->spec->image_id = image_id;
5872 dout("image_id is %s\n", image_id);
5882 * Undo whatever state changes are made by v1 or v2 header info
5885 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
5887 struct rbd_image_header *header;
5889 rbd_dev_parent_put(rbd_dev);
5891 /* Free dynamic fields from the header, then zero it out */
5893 header = &rbd_dev->header;
5894 ceph_put_snap_context(header->snapc);
5895 kfree(header->snap_sizes);
5896 kfree(header->snap_names);
5897 kfree(header->object_prefix);
5898 memset(header, 0, sizeof (*header));
5901 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev)
5905 ret = rbd_dev_v2_object_prefix(rbd_dev);
5910 * Get the and check features for the image. Currently the
5911 * features are assumed to never change.
5913 ret = rbd_dev_v2_features(rbd_dev);
5917 /* If the image supports fancy striping, get its parameters */
5919 if (rbd_dev->header.features & RBD_FEATURE_STRIPINGV2) {
5920 ret = rbd_dev_v2_striping_info(rbd_dev);
5924 /* No support for crypto and compression type format 2 images */
5928 rbd_dev->header.features = 0;
5929 kfree(rbd_dev->header.object_prefix);
5930 rbd_dev->header.object_prefix = NULL;
5936 * @depth is rbd_dev_image_probe() -> rbd_dev_probe_parent() ->
5937 * rbd_dev_image_probe() recursion depth, which means it's also the
5938 * length of the already discovered part of the parent chain.
5940 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev, int depth)
5942 struct rbd_device *parent = NULL;
5945 if (!rbd_dev->parent_spec)
5948 if (++depth > RBD_MAX_PARENT_CHAIN_LEN) {
5949 pr_info("parent chain is too long (%d)\n", depth);
5954 parent = __rbd_dev_create(rbd_dev->rbd_client, rbd_dev->parent_spec);
5961 * Images related by parent/child relationships always share
5962 * rbd_client and spec/parent_spec, so bump their refcounts.
5964 __rbd_get_client(rbd_dev->rbd_client);
5965 rbd_spec_get(rbd_dev->parent_spec);
5967 ret = rbd_dev_image_probe(parent, depth);
5971 rbd_dev->parent = parent;
5972 atomic_set(&rbd_dev->parent_ref, 1);
5976 rbd_dev_unparent(rbd_dev);
5977 rbd_dev_destroy(parent);
5982 * rbd_dev->header_rwsem must be locked for write and will be unlocked
5985 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
5989 /* Record our major and minor device numbers. */
5991 if (!single_major) {
5992 ret = register_blkdev(0, rbd_dev->name);
5994 goto err_out_unlock;
5996 rbd_dev->major = ret;
5999 rbd_dev->major = rbd_major;
6000 rbd_dev->minor = rbd_dev_id_to_minor(rbd_dev->dev_id);
6003 /* Set up the blkdev mapping. */
6005 ret = rbd_init_disk(rbd_dev);
6007 goto err_out_blkdev;
6009 ret = rbd_dev_mapping_set(rbd_dev);
6013 set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
6014 set_disk_ro(rbd_dev->disk, rbd_dev->mapping.read_only);
6016 dev_set_name(&rbd_dev->dev, "%d", rbd_dev->dev_id);
6017 ret = device_add(&rbd_dev->dev);
6019 goto err_out_mapping;
6021 /* Everything's ready. Announce the disk to the world. */
6023 set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
6024 up_write(&rbd_dev->header_rwsem);
6026 spin_lock(&rbd_dev_list_lock);
6027 list_add_tail(&rbd_dev->node, &rbd_dev_list);
6028 spin_unlock(&rbd_dev_list_lock);
6030 add_disk(rbd_dev->disk);
6031 pr_info("%s: capacity %llu features 0x%llx\n", rbd_dev->disk->disk_name,
6032 (unsigned long long)get_capacity(rbd_dev->disk) << SECTOR_SHIFT,
6033 rbd_dev->header.features);
6038 rbd_dev_mapping_clear(rbd_dev);
6040 rbd_free_disk(rbd_dev);
6043 unregister_blkdev(rbd_dev->major, rbd_dev->name);
6045 up_write(&rbd_dev->header_rwsem);
6049 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
6051 struct rbd_spec *spec = rbd_dev->spec;
6054 /* Record the header object name for this rbd image. */
6056 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
6058 rbd_dev->header_oloc.pool = rbd_dev->layout.pool_id;
6059 if (rbd_dev->image_format == 1)
6060 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
6061 spec->image_name, RBD_SUFFIX);
6063 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
6064 RBD_HEADER_PREFIX, spec->image_id);
6069 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
6071 rbd_dev_unprobe(rbd_dev);
6072 rbd_dev->image_format = 0;
6073 kfree(rbd_dev->spec->image_id);
6074 rbd_dev->spec->image_id = NULL;
6076 rbd_dev_destroy(rbd_dev);
6080 * Probe for the existence of the header object for the given rbd
6081 * device. If this image is the one being mapped (i.e., not a
6082 * parent), initiate a watch on its header object before using that
6083 * object to get detailed information about the rbd image.
6085 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth)
6090 * Get the id from the image id object. Unless there's an
6091 * error, rbd_dev->spec->image_id will be filled in with
6092 * a dynamically-allocated string, and rbd_dev->image_format
6093 * will be set to either 1 or 2.
6095 ret = rbd_dev_image_id(rbd_dev);
6099 ret = rbd_dev_header_name(rbd_dev);
6101 goto err_out_format;
6104 ret = rbd_register_watch(rbd_dev);
6107 pr_info("image %s/%s does not exist\n",
6108 rbd_dev->spec->pool_name,
6109 rbd_dev->spec->image_name);
6110 goto err_out_format;
6114 ret = rbd_dev_header_info(rbd_dev);
6119 * If this image is the one being mapped, we have pool name and
6120 * id, image name and id, and snap name - need to fill snap id.
6121 * Otherwise this is a parent image, identified by pool, image
6122 * and snap ids - need to fill in names for those ids.
6125 ret = rbd_spec_fill_snap_id(rbd_dev);
6127 ret = rbd_spec_fill_names(rbd_dev);
6130 pr_info("snap %s/%s@%s does not exist\n",
6131 rbd_dev->spec->pool_name,
6132 rbd_dev->spec->image_name,
6133 rbd_dev->spec->snap_name);
6137 if (rbd_dev->header.features & RBD_FEATURE_LAYERING) {
6138 ret = rbd_dev_v2_parent_info(rbd_dev);
6143 * Need to warn users if this image is the one being
6144 * mapped and has a parent.
6146 if (!depth && rbd_dev->parent_spec)
6148 "WARNING: kernel layering is EXPERIMENTAL!");
6151 ret = rbd_dev_probe_parent(rbd_dev, depth);
6155 dout("discovered format %u image, header name is %s\n",
6156 rbd_dev->image_format, rbd_dev->header_oid.name);
6160 rbd_dev_unprobe(rbd_dev);
6163 rbd_unregister_watch(rbd_dev);
6165 rbd_dev->image_format = 0;
6166 kfree(rbd_dev->spec->image_id);
6167 rbd_dev->spec->image_id = NULL;
6171 static ssize_t do_rbd_add(struct bus_type *bus,
6175 struct rbd_device *rbd_dev = NULL;
6176 struct ceph_options *ceph_opts = NULL;
6177 struct rbd_options *rbd_opts = NULL;
6178 struct rbd_spec *spec = NULL;
6179 struct rbd_client *rbdc;
6183 if (!try_module_get(THIS_MODULE))
6186 /* parse add command */
6187 rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
6191 rbdc = rbd_get_client(ceph_opts);
6198 rc = rbd_add_get_pool_id(rbdc, spec->pool_name);
6201 pr_info("pool %s does not exist\n", spec->pool_name);
6202 goto err_out_client;
6204 spec->pool_id = (u64)rc;
6206 rbd_dev = rbd_dev_create(rbdc, spec, rbd_opts);
6209 goto err_out_client;
6211 rbdc = NULL; /* rbd_dev now owns this */
6212 spec = NULL; /* rbd_dev now owns this */
6213 rbd_opts = NULL; /* rbd_dev now owns this */
6215 rbd_dev->config_info = kstrdup(buf, GFP_KERNEL);
6216 if (!rbd_dev->config_info) {
6218 goto err_out_rbd_dev;
6221 down_write(&rbd_dev->header_rwsem);
6222 rc = rbd_dev_image_probe(rbd_dev, 0);
6224 up_write(&rbd_dev->header_rwsem);
6225 goto err_out_rbd_dev;
6228 /* If we are mapping a snapshot it must be marked read-only */
6230 read_only = rbd_dev->opts->read_only;
6231 if (rbd_dev->spec->snap_id != CEPH_NOSNAP)
6233 rbd_dev->mapping.read_only = read_only;
6235 rc = rbd_dev_device_setup(rbd_dev);
6238 * rbd_unregister_watch() can't be moved into
6239 * rbd_dev_image_release() without refactoring, see
6240 * commit 1f3ef78861ac.
6242 rbd_unregister_watch(rbd_dev);
6243 rbd_dev_image_release(rbd_dev);
6249 module_put(THIS_MODULE);
6253 rbd_dev_destroy(rbd_dev);
6255 rbd_put_client(rbdc);
6262 static ssize_t rbd_add(struct bus_type *bus,
6269 return do_rbd_add(bus, buf, count);
6272 static ssize_t rbd_add_single_major(struct bus_type *bus,
6276 return do_rbd_add(bus, buf, count);
6279 static void rbd_dev_device_release(struct rbd_device *rbd_dev)
6281 rbd_free_disk(rbd_dev);
6283 spin_lock(&rbd_dev_list_lock);
6284 list_del_init(&rbd_dev->node);
6285 spin_unlock(&rbd_dev_list_lock);
6287 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
6288 device_del(&rbd_dev->dev);
6289 rbd_dev_mapping_clear(rbd_dev);
6291 unregister_blkdev(rbd_dev->major, rbd_dev->name);
6294 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
6296 while (rbd_dev->parent) {
6297 struct rbd_device *first = rbd_dev;
6298 struct rbd_device *second = first->parent;
6299 struct rbd_device *third;
6302 * Follow to the parent with no grandparent and
6305 while (second && (third = second->parent)) {
6310 rbd_dev_image_release(second);
6311 first->parent = NULL;
6312 first->parent_overlap = 0;
6314 rbd_assert(first->parent_spec);
6315 rbd_spec_put(first->parent_spec);
6316 first->parent_spec = NULL;
6320 static ssize_t do_rbd_remove(struct bus_type *bus,
6324 struct rbd_device *rbd_dev = NULL;
6325 struct list_head *tmp;
6328 bool already = false;
6334 sscanf(buf, "%d %5s", &dev_id, opt_buf);
6336 pr_err("dev_id out of range\n");
6339 if (opt_buf[0] != '\0') {
6340 if (!strcmp(opt_buf, "force")) {
6343 pr_err("bad remove option at '%s'\n", opt_buf);
6349 spin_lock(&rbd_dev_list_lock);
6350 list_for_each(tmp, &rbd_dev_list) {
6351 rbd_dev = list_entry(tmp, struct rbd_device, node);
6352 if (rbd_dev->dev_id == dev_id) {
6358 spin_lock_irq(&rbd_dev->lock);
6359 if (rbd_dev->open_count && !force)
6362 already = test_and_set_bit(RBD_DEV_FLAG_REMOVING,
6364 spin_unlock_irq(&rbd_dev->lock);
6366 spin_unlock(&rbd_dev_list_lock);
6367 if (ret < 0 || already)
6372 * Prevent new IO from being queued and wait for existing
6373 * IO to complete/fail.
6375 blk_mq_freeze_queue(rbd_dev->disk->queue);
6376 blk_set_queue_dying(rbd_dev->disk->queue);
6379 down_write(&rbd_dev->lock_rwsem);
6380 if (__rbd_is_lock_owner(rbd_dev))
6381 rbd_unlock(rbd_dev);
6382 up_write(&rbd_dev->lock_rwsem);
6383 rbd_unregister_watch(rbd_dev);
6386 * Don't free anything from rbd_dev->disk until after all
6387 * notifies are completely processed. Otherwise
6388 * rbd_bus_del_dev() will race with rbd_watch_cb(), resulting
6389 * in a potential use after free of rbd_dev->disk or rbd_dev.
6391 rbd_dev_device_release(rbd_dev);
6392 rbd_dev_image_release(rbd_dev);
6397 static ssize_t rbd_remove(struct bus_type *bus,
6404 return do_rbd_remove(bus, buf, count);
6407 static ssize_t rbd_remove_single_major(struct bus_type *bus,
6411 return do_rbd_remove(bus, buf, count);
6415 * create control files in sysfs
6418 static int rbd_sysfs_init(void)
6422 ret = device_register(&rbd_root_dev);
6426 ret = bus_register(&rbd_bus_type);
6428 device_unregister(&rbd_root_dev);
6433 static void rbd_sysfs_cleanup(void)
6435 bus_unregister(&rbd_bus_type);
6436 device_unregister(&rbd_root_dev);
6439 static int rbd_slab_init(void)
6441 rbd_assert(!rbd_img_request_cache);
6442 rbd_img_request_cache = KMEM_CACHE(rbd_img_request, 0);
6443 if (!rbd_img_request_cache)
6446 rbd_assert(!rbd_obj_request_cache);
6447 rbd_obj_request_cache = KMEM_CACHE(rbd_obj_request, 0);
6448 if (!rbd_obj_request_cache)
6451 rbd_assert(!rbd_segment_name_cache);
6452 rbd_segment_name_cache = kmem_cache_create("rbd_segment_name",
6453 CEPH_MAX_OID_NAME_LEN + 1, 1, 0, NULL);
6454 if (rbd_segment_name_cache)
6457 kmem_cache_destroy(rbd_obj_request_cache);
6458 rbd_obj_request_cache = NULL;
6460 kmem_cache_destroy(rbd_img_request_cache);
6461 rbd_img_request_cache = NULL;
6466 static void rbd_slab_exit(void)
6468 rbd_assert(rbd_segment_name_cache);
6469 kmem_cache_destroy(rbd_segment_name_cache);
6470 rbd_segment_name_cache = NULL;
6472 rbd_assert(rbd_obj_request_cache);
6473 kmem_cache_destroy(rbd_obj_request_cache);
6474 rbd_obj_request_cache = NULL;
6476 rbd_assert(rbd_img_request_cache);
6477 kmem_cache_destroy(rbd_img_request_cache);
6478 rbd_img_request_cache = NULL;
6481 static int __init rbd_init(void)
6485 if (!libceph_compatible(NULL)) {
6486 rbd_warn(NULL, "libceph incompatibility (quitting)");
6490 rc = rbd_slab_init();
6495 * The number of active work items is limited by the number of
6496 * rbd devices * queue depth, so leave @max_active at default.
6498 rbd_wq = alloc_workqueue(RBD_DRV_NAME, WQ_MEM_RECLAIM, 0);
6505 rbd_major = register_blkdev(0, RBD_DRV_NAME);
6506 if (rbd_major < 0) {
6512 rc = rbd_sysfs_init();
6514 goto err_out_blkdev;
6517 pr_info("loaded (major %d)\n", rbd_major);
6519 pr_info("loaded\n");
6525 unregister_blkdev(rbd_major, RBD_DRV_NAME);
6527 destroy_workqueue(rbd_wq);
6533 static void __exit rbd_exit(void)
6535 ida_destroy(&rbd_dev_id_ida);
6536 rbd_sysfs_cleanup();
6538 unregister_blkdev(rbd_major, RBD_DRV_NAME);
6539 destroy_workqueue(rbd_wq);
6543 module_init(rbd_init);
6544 module_exit(rbd_exit);
6546 MODULE_AUTHOR("Alex Elder <elder@inktank.com>");
6547 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
6548 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
6549 /* following authorship retained from original osdblk.c */
6550 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
6552 MODULE_DESCRIPTION("RADOS Block Device (RBD) driver");
6553 MODULE_LICENSE("GPL");