3 rbd.c -- Export ceph rados objects as a Linux block device
6 based on drivers/block/osdblk.c:
8 Copyright 2009 Red Hat, Inc.
10 This program is free software; you can redistribute it and/or modify
11 it under the terms of the GNU General Public License as published by
12 the Free Software Foundation.
14 This program is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 GNU General Public License for more details.
19 You should have received a copy of the GNU General Public License
20 along with this program; see the file COPYING. If not, write to
21 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
25 For usage instructions, please refer to:
27 Documentation/ABI/testing/sysfs-bus-rbd
31 #include <linux/ceph/libceph.h>
32 #include <linux/ceph/osd_client.h>
33 #include <linux/ceph/mon_client.h>
34 #include <linux/ceph/decode.h>
35 #include <linux/parser.h>
36 #include <linux/bsearch.h>
38 #include <linux/kernel.h>
39 #include <linux/device.h>
40 #include <linux/module.h>
41 #include <linux/blk-mq.h>
43 #include <linux/blkdev.h>
44 #include <linux/slab.h>
45 #include <linux/idr.h>
46 #include <linux/workqueue.h>
48 #include "rbd_types.h"
50 #define RBD_DEBUG /* Activate rbd_assert() calls */
53 * The basic unit of block I/O is a sector. It is interpreted in a
54 * number of contexts in Linux (blk, bio, genhd), but the default is
55 * universally 512 bytes. These symbols are just slightly more
56 * meaningful than the bare numbers they represent.
58 #define SECTOR_SHIFT 9
59 #define SECTOR_SIZE (1ULL << SECTOR_SHIFT)
62 * Increment the given counter and return its updated value.
63 * If the counter is already 0 it will not be incremented.
64 * If the counter is already at its maximum value returns
65 * -EINVAL without updating it.
67 static int atomic_inc_return_safe(atomic_t *v)
71 counter = (unsigned int)__atomic_add_unless(v, 1, 0);
72 if (counter <= (unsigned int)INT_MAX)
80 /* Decrement the counter. Return the resulting value, or -EINVAL */
81 static int atomic_dec_return_safe(atomic_t *v)
85 counter = atomic_dec_return(v);
94 #define RBD_DRV_NAME "rbd"
96 #define RBD_MINORS_PER_MAJOR 256
97 #define RBD_SINGLE_MAJOR_PART_SHIFT 4
99 #define RBD_SNAP_DEV_NAME_PREFIX "snap_"
100 #define RBD_MAX_SNAP_NAME_LEN \
101 (NAME_MAX - (sizeof (RBD_SNAP_DEV_NAME_PREFIX) - 1))
103 #define RBD_MAX_SNAP_COUNT 510 /* allows max snapc to fit in 4KB */
105 #define RBD_SNAP_HEAD_NAME "-"
107 #define BAD_SNAP_INDEX U32_MAX /* invalid index into snap array */
109 /* This allows a single page to hold an image name sent by OSD */
110 #define RBD_IMAGE_NAME_LEN_MAX (PAGE_SIZE - sizeof (__le32) - 1)
111 #define RBD_IMAGE_ID_LEN_MAX 64
113 #define RBD_OBJ_PREFIX_LEN_MAX 64
117 #define RBD_FEATURE_LAYERING (1<<0)
118 #define RBD_FEATURE_STRIPINGV2 (1<<1)
119 #define RBD_FEATURES_ALL \
120 (RBD_FEATURE_LAYERING | RBD_FEATURE_STRIPINGV2)
122 /* Features supported by this (client software) implementation. */
124 #define RBD_FEATURES_SUPPORTED (RBD_FEATURES_ALL)
127 * An RBD device name will be "rbd#", where the "rbd" comes from
128 * RBD_DRV_NAME above, and # is a unique integer identifier.
129 * MAX_INT_FORMAT_WIDTH is used in ensuring DEV_NAME_LEN is big
130 * enough to hold all possible device names.
132 #define DEV_NAME_LEN 32
133 #define MAX_INT_FORMAT_WIDTH ((5 * sizeof (int)) / 2 + 1)
136 * block device image metadata (in-memory version)
138 struct rbd_image_header {
139 /* These six fields never change for a given rbd image */
146 u64 features; /* Might be changeable someday? */
148 /* The remaining fields need to be updated occasionally */
150 struct ceph_snap_context *snapc;
151 char *snap_names; /* format 1 only */
152 u64 *snap_sizes; /* format 1 only */
156 * An rbd image specification.
158 * The tuple (pool_id, image_id, snap_id) is sufficient to uniquely
159 * identify an image. Each rbd_dev structure includes a pointer to
160 * an rbd_spec structure that encapsulates this identity.
162 * Each of the id's in an rbd_spec has an associated name. For a
163 * user-mapped image, the names are supplied and the id's associated
164 * with them are looked up. For a layered image, a parent image is
165 * defined by the tuple, and the names are looked up.
167 * An rbd_dev structure contains a parent_spec pointer which is
168 * non-null if the image it represents is a child in a layered
169 * image. This pointer will refer to the rbd_spec structure used
170 * by the parent rbd_dev for its own identity (i.e., the structure
171 * is shared between the parent and child).
173 * Since these structures are populated once, during the discovery
174 * phase of image construction, they are effectively immutable so
175 * we make no effort to synchronize access to them.
177 * Note that code herein does not assume the image name is known (it
178 * could be a null pointer).
182 const char *pool_name;
184 const char *image_id;
185 const char *image_name;
188 const char *snap_name;
194 * an instance of the client. multiple devices may share an rbd client.
197 struct ceph_client *client;
199 struct list_head node;
202 struct rbd_img_request;
203 typedef void (*rbd_img_callback_t)(struct rbd_img_request *);
205 #define BAD_WHICH U32_MAX /* Good which or bad which, which? */
207 struct rbd_obj_request;
208 typedef void (*rbd_obj_callback_t)(struct rbd_obj_request *);
210 enum obj_request_type {
211 OBJ_REQUEST_NODATA, OBJ_REQUEST_BIO, OBJ_REQUEST_PAGES
214 enum obj_operation_type {
221 OBJ_REQ_DONE, /* completion flag: not done = 0, done = 1 */
222 OBJ_REQ_IMG_DATA, /* object usage: standalone = 0, image = 1 */
223 OBJ_REQ_KNOWN, /* EXISTS flag valid: no = 0, yes = 1 */
224 OBJ_REQ_EXISTS, /* target exists: no = 0, yes = 1 */
227 struct rbd_obj_request {
228 const char *object_name;
229 u64 offset; /* object start byte */
230 u64 length; /* bytes from offset */
234 * An object request associated with an image will have its
235 * img_data flag set; a standalone object request will not.
237 * A standalone object request will have which == BAD_WHICH
238 * and a null obj_request pointer.
240 * An object request initiated in support of a layered image
241 * object (to check for its existence before a write) will
242 * have which == BAD_WHICH and a non-null obj_request pointer.
244 * Finally, an object request for rbd image data will have
245 * which != BAD_WHICH, and will have a non-null img_request
246 * pointer. The value of which will be in the range
247 * 0..(img_request->obj_request_count-1).
250 struct rbd_obj_request *obj_request; /* STAT op */
252 struct rbd_img_request *img_request;
254 /* links for img_request->obj_requests list */
255 struct list_head links;
258 u32 which; /* posn image request list */
260 enum obj_request_type type;
262 struct bio *bio_list;
268 struct page **copyup_pages;
269 u32 copyup_page_count;
271 struct ceph_osd_request *osd_req;
273 u64 xferred; /* bytes transferred */
276 rbd_obj_callback_t callback;
277 struct completion completion;
283 IMG_REQ_WRITE, /* I/O direction: read = 0, write = 1 */
284 IMG_REQ_CHILD, /* initiator: block = 0, child image = 1 */
285 IMG_REQ_LAYERED, /* ENOENT handling: normal = 0, layered = 1 */
286 IMG_REQ_DISCARD, /* discard: normal = 0, discard request = 1 */
289 struct rbd_img_request {
290 struct rbd_device *rbd_dev;
291 u64 offset; /* starting image byte offset */
292 u64 length; /* byte count from offset */
295 u64 snap_id; /* for reads */
296 struct ceph_snap_context *snapc; /* for writes */
299 struct request *rq; /* block request */
300 struct rbd_obj_request *obj_request; /* obj req initiator */
302 struct page **copyup_pages;
303 u32 copyup_page_count;
304 spinlock_t completion_lock;/* protects next_completion */
306 rbd_img_callback_t callback;
307 u64 xferred;/* aggregate bytes transferred */
308 int result; /* first nonzero obj_request result */
310 u32 obj_request_count;
311 struct list_head obj_requests; /* rbd_obj_request structs */
316 #define for_each_obj_request(ireq, oreq) \
317 list_for_each_entry(oreq, &(ireq)->obj_requests, links)
318 #define for_each_obj_request_from(ireq, oreq) \
319 list_for_each_entry_from(oreq, &(ireq)->obj_requests, links)
320 #define for_each_obj_request_safe(ireq, oreq, n) \
321 list_for_each_entry_safe_reverse(oreq, n, &(ireq)->obj_requests, links)
333 int dev_id; /* blkdev unique id */
335 int major; /* blkdev assigned major */
337 struct gendisk *disk; /* blkdev's gendisk and rq */
339 u32 image_format; /* Either 1 or 2 */
340 struct rbd_client *rbd_client;
342 char name[DEV_NAME_LEN]; /* blkdev name, e.g. rbd3 */
344 spinlock_t lock; /* queue, flags, open_count */
346 struct rbd_image_header header;
347 unsigned long flags; /* possibly lock protected */
348 struct rbd_spec *spec;
349 struct rbd_options *opts;
353 struct ceph_file_layout layout;
355 struct ceph_osd_event *watch_event;
356 struct rbd_obj_request *watch_request;
358 struct rbd_spec *parent_spec;
361 struct rbd_device *parent;
363 /* Block layer tags. */
364 struct blk_mq_tag_set tag_set;
366 /* protects updating the header */
367 struct rw_semaphore header_rwsem;
369 struct rbd_mapping mapping;
371 struct list_head node;
375 unsigned long open_count; /* protected by lock */
379 * Flag bits for rbd_dev->flags. If atomicity is required,
380 * rbd_dev->lock is used to protect access.
382 * Currently, only the "removing" flag (which is coupled with the
383 * "open_count" field) requires atomic access.
386 RBD_DEV_FLAG_EXISTS, /* mapped snapshot has not been deleted */
387 RBD_DEV_FLAG_REMOVING, /* this mapping is being removed */
390 static DEFINE_MUTEX(client_mutex); /* Serialize client creation */
392 static LIST_HEAD(rbd_dev_list); /* devices */
393 static DEFINE_SPINLOCK(rbd_dev_list_lock);
395 static LIST_HEAD(rbd_client_list); /* clients */
396 static DEFINE_SPINLOCK(rbd_client_list_lock);
398 /* Slab caches for frequently-allocated structures */
400 static struct kmem_cache *rbd_img_request_cache;
401 static struct kmem_cache *rbd_obj_request_cache;
402 static struct kmem_cache *rbd_segment_name_cache;
404 static int rbd_major;
405 static DEFINE_IDA(rbd_dev_id_ida);
407 static struct workqueue_struct *rbd_wq;
410 * Default to false for now, as single-major requires >= 0.75 version of
411 * userspace rbd utility.
413 static bool single_major = false;
414 module_param(single_major, bool, S_IRUGO);
415 MODULE_PARM_DESC(single_major, "Use a single major number for all rbd devices (default: false)");
417 static int rbd_img_request_submit(struct rbd_img_request *img_request);
419 static void rbd_dev_device_release(struct device *dev);
421 static ssize_t rbd_add(struct bus_type *bus, const char *buf,
423 static ssize_t rbd_remove(struct bus_type *bus, const char *buf,
425 static ssize_t rbd_add_single_major(struct bus_type *bus, const char *buf,
427 static ssize_t rbd_remove_single_major(struct bus_type *bus, const char *buf,
429 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, bool mapping);
430 static void rbd_spec_put(struct rbd_spec *spec);
432 static int rbd_dev_id_to_minor(int dev_id)
434 return dev_id << RBD_SINGLE_MAJOR_PART_SHIFT;
437 static int minor_to_rbd_dev_id(int minor)
439 return minor >> RBD_SINGLE_MAJOR_PART_SHIFT;
442 static BUS_ATTR(add, S_IWUSR, NULL, rbd_add);
443 static BUS_ATTR(remove, S_IWUSR, NULL, rbd_remove);
444 static BUS_ATTR(add_single_major, S_IWUSR, NULL, rbd_add_single_major);
445 static BUS_ATTR(remove_single_major, S_IWUSR, NULL, rbd_remove_single_major);
447 static struct attribute *rbd_bus_attrs[] = {
449 &bus_attr_remove.attr,
450 &bus_attr_add_single_major.attr,
451 &bus_attr_remove_single_major.attr,
455 static umode_t rbd_bus_is_visible(struct kobject *kobj,
456 struct attribute *attr, int index)
459 (attr == &bus_attr_add_single_major.attr ||
460 attr == &bus_attr_remove_single_major.attr))
466 static const struct attribute_group rbd_bus_group = {
467 .attrs = rbd_bus_attrs,
468 .is_visible = rbd_bus_is_visible,
470 __ATTRIBUTE_GROUPS(rbd_bus);
472 static struct bus_type rbd_bus_type = {
474 .bus_groups = rbd_bus_groups,
477 static void rbd_root_dev_release(struct device *dev)
481 static struct device rbd_root_dev = {
483 .release = rbd_root_dev_release,
486 static __printf(2, 3)
487 void rbd_warn(struct rbd_device *rbd_dev, const char *fmt, ...)
489 struct va_format vaf;
497 printk(KERN_WARNING "%s: %pV\n", RBD_DRV_NAME, &vaf);
498 else if (rbd_dev->disk)
499 printk(KERN_WARNING "%s: %s: %pV\n",
500 RBD_DRV_NAME, rbd_dev->disk->disk_name, &vaf);
501 else if (rbd_dev->spec && rbd_dev->spec->image_name)
502 printk(KERN_WARNING "%s: image %s: %pV\n",
503 RBD_DRV_NAME, rbd_dev->spec->image_name, &vaf);
504 else if (rbd_dev->spec && rbd_dev->spec->image_id)
505 printk(KERN_WARNING "%s: id %s: %pV\n",
506 RBD_DRV_NAME, rbd_dev->spec->image_id, &vaf);
508 printk(KERN_WARNING "%s: rbd_dev %p: %pV\n",
509 RBD_DRV_NAME, rbd_dev, &vaf);
514 #define rbd_assert(expr) \
515 if (unlikely(!(expr))) { \
516 printk(KERN_ERR "\nAssertion failure in %s() " \
518 "\trbd_assert(%s);\n\n", \
519 __func__, __LINE__, #expr); \
522 #else /* !RBD_DEBUG */
523 # define rbd_assert(expr) ((void) 0)
524 #endif /* !RBD_DEBUG */
526 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request);
527 static void rbd_img_parent_read(struct rbd_obj_request *obj_request);
528 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev);
530 static int rbd_dev_refresh(struct rbd_device *rbd_dev);
531 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev);
532 static int rbd_dev_header_info(struct rbd_device *rbd_dev);
533 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev);
534 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
536 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
537 u8 *order, u64 *snap_size);
538 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
540 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name);
542 static int rbd_open(struct block_device *bdev, fmode_t mode)
544 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
545 bool removing = false;
547 if ((mode & FMODE_WRITE) && rbd_dev->mapping.read_only)
550 spin_lock_irq(&rbd_dev->lock);
551 if (test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags))
554 rbd_dev->open_count++;
555 spin_unlock_irq(&rbd_dev->lock);
559 (void) get_device(&rbd_dev->dev);
564 static void rbd_release(struct gendisk *disk, fmode_t mode)
566 struct rbd_device *rbd_dev = disk->private_data;
567 unsigned long open_count_before;
569 spin_lock_irq(&rbd_dev->lock);
570 open_count_before = rbd_dev->open_count--;
571 spin_unlock_irq(&rbd_dev->lock);
572 rbd_assert(open_count_before > 0);
574 put_device(&rbd_dev->dev);
577 static int rbd_ioctl_set_ro(struct rbd_device *rbd_dev, unsigned long arg)
582 bool ro_changed = false;
584 /* get_user() may sleep, so call it before taking rbd_dev->lock */
585 if (get_user(val, (int __user *)(arg)))
588 ro = val ? true : false;
589 /* Snapshot doesn't allow to write*/
590 if (rbd_dev->spec->snap_id != CEPH_NOSNAP && !ro)
593 spin_lock_irq(&rbd_dev->lock);
594 /* prevent others open this device */
595 if (rbd_dev->open_count > 1) {
600 if (rbd_dev->mapping.read_only != ro) {
601 rbd_dev->mapping.read_only = ro;
606 spin_unlock_irq(&rbd_dev->lock);
607 /* set_disk_ro() may sleep, so call it after releasing rbd_dev->lock */
608 if (ret == 0 && ro_changed)
609 set_disk_ro(rbd_dev->disk, ro ? 1 : 0);
614 static int rbd_ioctl(struct block_device *bdev, fmode_t mode,
615 unsigned int cmd, unsigned long arg)
617 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
622 ret = rbd_ioctl_set_ro(rbd_dev, arg);
632 static int rbd_compat_ioctl(struct block_device *bdev, fmode_t mode,
633 unsigned int cmd, unsigned long arg)
635 return rbd_ioctl(bdev, mode, cmd, arg);
637 #endif /* CONFIG_COMPAT */
639 static const struct block_device_operations rbd_bd_ops = {
640 .owner = THIS_MODULE,
642 .release = rbd_release,
645 .compat_ioctl = rbd_compat_ioctl,
650 * Initialize an rbd client instance. Success or not, this function
651 * consumes ceph_opts. Caller holds client_mutex.
653 static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts)
655 struct rbd_client *rbdc;
658 dout("%s:\n", __func__);
659 rbdc = kmalloc(sizeof(struct rbd_client), GFP_KERNEL);
663 kref_init(&rbdc->kref);
664 INIT_LIST_HEAD(&rbdc->node);
666 rbdc->client = ceph_create_client(ceph_opts, rbdc, 0, 0);
667 if (IS_ERR(rbdc->client))
669 ceph_opts = NULL; /* Now rbdc->client is responsible for ceph_opts */
671 ret = ceph_open_session(rbdc->client);
675 spin_lock(&rbd_client_list_lock);
676 list_add_tail(&rbdc->node, &rbd_client_list);
677 spin_unlock(&rbd_client_list_lock);
679 dout("%s: rbdc %p\n", __func__, rbdc);
683 ceph_destroy_client(rbdc->client);
688 ceph_destroy_options(ceph_opts);
689 dout("%s: error %d\n", __func__, ret);
694 static struct rbd_client *__rbd_get_client(struct rbd_client *rbdc)
696 kref_get(&rbdc->kref);
702 * Find a ceph client with specific addr and configuration. If
703 * found, bump its reference count.
705 static struct rbd_client *rbd_client_find(struct ceph_options *ceph_opts)
707 struct rbd_client *client_node;
710 if (ceph_opts->flags & CEPH_OPT_NOSHARE)
713 spin_lock(&rbd_client_list_lock);
714 list_for_each_entry(client_node, &rbd_client_list, node) {
715 if (!ceph_compare_options(ceph_opts, client_node->client)) {
716 __rbd_get_client(client_node);
722 spin_unlock(&rbd_client_list_lock);
724 return found ? client_node : NULL;
728 * (Per device) rbd map options
735 /* string args above */
741 static match_table_t rbd_opts_tokens = {
742 {Opt_queue_depth, "queue_depth=%d"},
744 /* string args above */
745 {Opt_read_only, "read_only"},
746 {Opt_read_only, "ro"}, /* Alternate spelling */
747 {Opt_read_write, "read_write"},
748 {Opt_read_write, "rw"}, /* Alternate spelling */
757 #define RBD_QUEUE_DEPTH_DEFAULT BLKDEV_MAX_RQ
758 #define RBD_READ_ONLY_DEFAULT false
760 static int parse_rbd_opts_token(char *c, void *private)
762 struct rbd_options *rbd_opts = private;
763 substring_t argstr[MAX_OPT_ARGS];
764 int token, intval, ret;
766 token = match_token(c, rbd_opts_tokens, argstr);
767 if (token < Opt_last_int) {
768 ret = match_int(&argstr[0], &intval);
770 pr_err("bad mount option arg (not int) at '%s'\n", c);
773 dout("got int token %d val %d\n", token, intval);
774 } else if (token > Opt_last_int && token < Opt_last_string) {
775 dout("got string token %d val %s\n", token, argstr[0].from);
777 dout("got token %d\n", token);
781 case Opt_queue_depth:
783 pr_err("queue_depth out of range\n");
786 rbd_opts->queue_depth = intval;
789 rbd_opts->read_only = true;
792 rbd_opts->read_only = false;
795 /* libceph prints "bad option" msg */
802 static char* obj_op_name(enum obj_operation_type op_type)
817 * Get a ceph client with specific addr and configuration, if one does
818 * not exist create it. Either way, ceph_opts is consumed by this
821 static struct rbd_client *rbd_get_client(struct ceph_options *ceph_opts)
823 struct rbd_client *rbdc;
825 mutex_lock_nested(&client_mutex, SINGLE_DEPTH_NESTING);
826 rbdc = rbd_client_find(ceph_opts);
827 if (rbdc) /* using an existing client */
828 ceph_destroy_options(ceph_opts);
830 rbdc = rbd_client_create(ceph_opts);
831 mutex_unlock(&client_mutex);
837 * Destroy ceph client
839 * Caller must hold rbd_client_list_lock.
841 static void rbd_client_release(struct kref *kref)
843 struct rbd_client *rbdc = container_of(kref, struct rbd_client, kref);
845 dout("%s: rbdc %p\n", __func__, rbdc);
846 spin_lock(&rbd_client_list_lock);
847 list_del(&rbdc->node);
848 spin_unlock(&rbd_client_list_lock);
850 ceph_destroy_client(rbdc->client);
855 * Drop reference to ceph client node. If it's not referenced anymore, release
858 static void rbd_put_client(struct rbd_client *rbdc)
861 kref_put(&rbdc->kref, rbd_client_release);
864 static bool rbd_image_format_valid(u32 image_format)
866 return image_format == 1 || image_format == 2;
869 static bool rbd_dev_ondisk_valid(struct rbd_image_header_ondisk *ondisk)
874 /* The header has to start with the magic rbd header text */
875 if (memcmp(&ondisk->text, RBD_HEADER_TEXT, sizeof (RBD_HEADER_TEXT)))
878 /* The bio layer requires at least sector-sized I/O */
880 if (ondisk->options.order < SECTOR_SHIFT)
883 /* If we use u64 in a few spots we may be able to loosen this */
885 if (ondisk->options.order > 8 * sizeof (int) - 1)
889 * The size of a snapshot header has to fit in a size_t, and
890 * that limits the number of snapshots.
892 snap_count = le32_to_cpu(ondisk->snap_count);
893 size = SIZE_MAX - sizeof (struct ceph_snap_context);
894 if (snap_count > size / sizeof (__le64))
898 * Not only that, but the size of the entire the snapshot
899 * header must also be representable in a size_t.
901 size -= snap_count * sizeof (__le64);
902 if ((u64) size < le64_to_cpu(ondisk->snap_names_len))
909 * Fill an rbd image header with information from the given format 1
912 static int rbd_header_from_disk(struct rbd_device *rbd_dev,
913 struct rbd_image_header_ondisk *ondisk)
915 struct rbd_image_header *header = &rbd_dev->header;
916 bool first_time = header->object_prefix == NULL;
917 struct ceph_snap_context *snapc;
918 char *object_prefix = NULL;
919 char *snap_names = NULL;
920 u64 *snap_sizes = NULL;
926 /* Allocate this now to avoid having to handle failure below */
931 len = strnlen(ondisk->object_prefix,
932 sizeof (ondisk->object_prefix));
933 object_prefix = kmalloc(len + 1, GFP_KERNEL);
936 memcpy(object_prefix, ondisk->object_prefix, len);
937 object_prefix[len] = '\0';
940 /* Allocate the snapshot context and fill it in */
942 snap_count = le32_to_cpu(ondisk->snap_count);
943 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
946 snapc->seq = le64_to_cpu(ondisk->snap_seq);
948 struct rbd_image_snap_ondisk *snaps;
949 u64 snap_names_len = le64_to_cpu(ondisk->snap_names_len);
951 /* We'll keep a copy of the snapshot names... */
953 if (snap_names_len > (u64)SIZE_MAX)
955 snap_names = kmalloc(snap_names_len, GFP_KERNEL);
959 /* ...as well as the array of their sizes. */
961 size = snap_count * sizeof (*header->snap_sizes);
962 snap_sizes = kmalloc(size, GFP_KERNEL);
967 * Copy the names, and fill in each snapshot's id
970 * Note that rbd_dev_v1_header_info() guarantees the
971 * ondisk buffer we're working with has
972 * snap_names_len bytes beyond the end of the
973 * snapshot id array, this memcpy() is safe.
975 memcpy(snap_names, &ondisk->snaps[snap_count], snap_names_len);
976 snaps = ondisk->snaps;
977 for (i = 0; i < snap_count; i++) {
978 snapc->snaps[i] = le64_to_cpu(snaps[i].id);
979 snap_sizes[i] = le64_to_cpu(snaps[i].image_size);
983 /* We won't fail any more, fill in the header */
986 header->object_prefix = object_prefix;
987 header->obj_order = ondisk->options.order;
988 header->crypt_type = ondisk->options.crypt_type;
989 header->comp_type = ondisk->options.comp_type;
990 /* The rest aren't used for format 1 images */
991 header->stripe_unit = 0;
992 header->stripe_count = 0;
993 header->features = 0;
995 ceph_put_snap_context(header->snapc);
996 kfree(header->snap_names);
997 kfree(header->snap_sizes);
1000 /* The remaining fields always get updated (when we refresh) */
1002 header->image_size = le64_to_cpu(ondisk->image_size);
1003 header->snapc = snapc;
1004 header->snap_names = snap_names;
1005 header->snap_sizes = snap_sizes;
1013 ceph_put_snap_context(snapc);
1014 kfree(object_prefix);
1019 static const char *_rbd_dev_v1_snap_name(struct rbd_device *rbd_dev, u32 which)
1021 const char *snap_name;
1023 rbd_assert(which < rbd_dev->header.snapc->num_snaps);
1025 /* Skip over names until we find the one we are looking for */
1027 snap_name = rbd_dev->header.snap_names;
1029 snap_name += strlen(snap_name) + 1;
1031 return kstrdup(snap_name, GFP_KERNEL);
1035 * Snapshot id comparison function for use with qsort()/bsearch().
1036 * Note that result is for snapshots in *descending* order.
1038 static int snapid_compare_reverse(const void *s1, const void *s2)
1040 u64 snap_id1 = *(u64 *)s1;
1041 u64 snap_id2 = *(u64 *)s2;
1043 if (snap_id1 < snap_id2)
1045 return snap_id1 == snap_id2 ? 0 : -1;
1049 * Search a snapshot context to see if the given snapshot id is
1052 * Returns the position of the snapshot id in the array if it's found,
1053 * or BAD_SNAP_INDEX otherwise.
1055 * Note: The snapshot array is in kept sorted (by the osd) in
1056 * reverse order, highest snapshot id first.
1058 static u32 rbd_dev_snap_index(struct rbd_device *rbd_dev, u64 snap_id)
1060 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
1063 found = bsearch(&snap_id, &snapc->snaps, snapc->num_snaps,
1064 sizeof (snap_id), snapid_compare_reverse);
1066 return found ? (u32)(found - &snapc->snaps[0]) : BAD_SNAP_INDEX;
1069 static const char *rbd_dev_v1_snap_name(struct rbd_device *rbd_dev,
1073 const char *snap_name;
1075 which = rbd_dev_snap_index(rbd_dev, snap_id);
1076 if (which == BAD_SNAP_INDEX)
1077 return ERR_PTR(-ENOENT);
1079 snap_name = _rbd_dev_v1_snap_name(rbd_dev, which);
1080 return snap_name ? snap_name : ERR_PTR(-ENOMEM);
1083 static const char *rbd_snap_name(struct rbd_device *rbd_dev, u64 snap_id)
1085 if (snap_id == CEPH_NOSNAP)
1086 return RBD_SNAP_HEAD_NAME;
1088 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1089 if (rbd_dev->image_format == 1)
1090 return rbd_dev_v1_snap_name(rbd_dev, snap_id);
1092 return rbd_dev_v2_snap_name(rbd_dev, snap_id);
1095 static int rbd_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
1098 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1099 if (snap_id == CEPH_NOSNAP) {
1100 *snap_size = rbd_dev->header.image_size;
1101 } else if (rbd_dev->image_format == 1) {
1104 which = rbd_dev_snap_index(rbd_dev, snap_id);
1105 if (which == BAD_SNAP_INDEX)
1108 *snap_size = rbd_dev->header.snap_sizes[which];
1113 ret = _rbd_dev_v2_snap_size(rbd_dev, snap_id, NULL, &size);
1122 static int rbd_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
1125 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1126 if (snap_id == CEPH_NOSNAP) {
1127 *snap_features = rbd_dev->header.features;
1128 } else if (rbd_dev->image_format == 1) {
1129 *snap_features = 0; /* No features for format 1 */
1134 ret = _rbd_dev_v2_snap_features(rbd_dev, snap_id, &features);
1138 *snap_features = features;
1143 static int rbd_dev_mapping_set(struct rbd_device *rbd_dev)
1145 u64 snap_id = rbd_dev->spec->snap_id;
1150 ret = rbd_snap_size(rbd_dev, snap_id, &size);
1153 ret = rbd_snap_features(rbd_dev, snap_id, &features);
1157 rbd_dev->mapping.size = size;
1158 rbd_dev->mapping.features = features;
1163 static void rbd_dev_mapping_clear(struct rbd_device *rbd_dev)
1165 rbd_dev->mapping.size = 0;
1166 rbd_dev->mapping.features = 0;
1169 static void rbd_segment_name_free(const char *name)
1171 /* The explicit cast here is needed to drop the const qualifier */
1173 kmem_cache_free(rbd_segment_name_cache, (void *)name);
1176 static const char *rbd_segment_name(struct rbd_device *rbd_dev, u64 offset)
1183 name = kmem_cache_alloc(rbd_segment_name_cache, GFP_NOIO);
1186 segment = offset >> rbd_dev->header.obj_order;
1187 name_format = "%s.%012llx";
1188 if (rbd_dev->image_format == 2)
1189 name_format = "%s.%016llx";
1190 ret = snprintf(name, CEPH_MAX_OID_NAME_LEN + 1, name_format,
1191 rbd_dev->header.object_prefix, segment);
1192 if (ret < 0 || ret > CEPH_MAX_OID_NAME_LEN) {
1193 pr_err("error formatting segment name for #%llu (%d)\n",
1195 rbd_segment_name_free(name);
1202 static u64 rbd_segment_offset(struct rbd_device *rbd_dev, u64 offset)
1204 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1206 return offset & (segment_size - 1);
1209 static u64 rbd_segment_length(struct rbd_device *rbd_dev,
1210 u64 offset, u64 length)
1212 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1214 offset &= segment_size - 1;
1216 rbd_assert(length <= U64_MAX - offset);
1217 if (offset + length > segment_size)
1218 length = segment_size - offset;
1224 * returns the size of an object in the image
1226 static u64 rbd_obj_bytes(struct rbd_image_header *header)
1228 return 1 << header->obj_order;
1235 static void bio_chain_put(struct bio *chain)
1241 chain = chain->bi_next;
1247 * zeros a bio chain, starting at specific offset
1249 static void zero_bio_chain(struct bio *chain, int start_ofs)
1252 struct bvec_iter iter;
1253 unsigned long flags;
1258 bio_for_each_segment(bv, chain, iter) {
1259 if (pos + bv.bv_len > start_ofs) {
1260 int remainder = max(start_ofs - pos, 0);
1261 buf = bvec_kmap_irq(&bv, &flags);
1262 memset(buf + remainder, 0,
1263 bv.bv_len - remainder);
1264 flush_dcache_page(bv.bv_page);
1265 bvec_kunmap_irq(buf, &flags);
1270 chain = chain->bi_next;
1275 * similar to zero_bio_chain(), zeros data defined by a page array,
1276 * starting at the given byte offset from the start of the array and
1277 * continuing up to the given end offset. The pages array is
1278 * assumed to be big enough to hold all bytes up to the end.
1280 static void zero_pages(struct page **pages, u64 offset, u64 end)
1282 struct page **page = &pages[offset >> PAGE_SHIFT];
1284 rbd_assert(end > offset);
1285 rbd_assert(end - offset <= (u64)SIZE_MAX);
1286 while (offset < end) {
1289 unsigned long flags;
1292 page_offset = offset & ~PAGE_MASK;
1293 length = min_t(size_t, PAGE_SIZE - page_offset, end - offset);
1294 local_irq_save(flags);
1295 kaddr = kmap_atomic(*page);
1296 memset(kaddr + page_offset, 0, length);
1297 flush_dcache_page(*page);
1298 kunmap_atomic(kaddr);
1299 local_irq_restore(flags);
1307 * Clone a portion of a bio, starting at the given byte offset
1308 * and continuing for the number of bytes indicated.
1310 static struct bio *bio_clone_range(struct bio *bio_src,
1311 unsigned int offset,
1317 bio = bio_clone(bio_src, gfpmask);
1319 return NULL; /* ENOMEM */
1321 bio_advance(bio, offset);
1322 bio->bi_iter.bi_size = len;
1328 * Clone a portion of a bio chain, starting at the given byte offset
1329 * into the first bio in the source chain and continuing for the
1330 * number of bytes indicated. The result is another bio chain of
1331 * exactly the given length, or a null pointer on error.
1333 * The bio_src and offset parameters are both in-out. On entry they
1334 * refer to the first source bio and the offset into that bio where
1335 * the start of data to be cloned is located.
1337 * On return, bio_src is updated to refer to the bio in the source
1338 * chain that contains first un-cloned byte, and *offset will
1339 * contain the offset of that byte within that bio.
1341 static struct bio *bio_chain_clone_range(struct bio **bio_src,
1342 unsigned int *offset,
1346 struct bio *bi = *bio_src;
1347 unsigned int off = *offset;
1348 struct bio *chain = NULL;
1351 /* Build up a chain of clone bios up to the limit */
1353 if (!bi || off >= bi->bi_iter.bi_size || !len)
1354 return NULL; /* Nothing to clone */
1358 unsigned int bi_size;
1362 rbd_warn(NULL, "bio_chain exhausted with %u left", len);
1363 goto out_err; /* EINVAL; ran out of bio's */
1365 bi_size = min_t(unsigned int, bi->bi_iter.bi_size - off, len);
1366 bio = bio_clone_range(bi, off, bi_size, gfpmask);
1368 goto out_err; /* ENOMEM */
1371 end = &bio->bi_next;
1374 if (off == bi->bi_iter.bi_size) {
1385 bio_chain_put(chain);
1391 * The default/initial value for all object request flags is 0. For
1392 * each flag, once its value is set to 1 it is never reset to 0
1395 static void obj_request_img_data_set(struct rbd_obj_request *obj_request)
1397 if (test_and_set_bit(OBJ_REQ_IMG_DATA, &obj_request->flags)) {
1398 struct rbd_device *rbd_dev;
1400 rbd_dev = obj_request->img_request->rbd_dev;
1401 rbd_warn(rbd_dev, "obj_request %p already marked img_data",
1406 static bool obj_request_img_data_test(struct rbd_obj_request *obj_request)
1409 return test_bit(OBJ_REQ_IMG_DATA, &obj_request->flags) != 0;
1412 static void obj_request_done_set(struct rbd_obj_request *obj_request)
1414 if (test_and_set_bit(OBJ_REQ_DONE, &obj_request->flags)) {
1415 struct rbd_device *rbd_dev = NULL;
1417 if (obj_request_img_data_test(obj_request))
1418 rbd_dev = obj_request->img_request->rbd_dev;
1419 rbd_warn(rbd_dev, "obj_request %p already marked done",
1424 static bool obj_request_done_test(struct rbd_obj_request *obj_request)
1427 return test_bit(OBJ_REQ_DONE, &obj_request->flags) != 0;
1431 * This sets the KNOWN flag after (possibly) setting the EXISTS
1432 * flag. The latter is set based on the "exists" value provided.
1434 * Note that for our purposes once an object exists it never goes
1435 * away again. It's possible that the response from two existence
1436 * checks are separated by the creation of the target object, and
1437 * the first ("doesn't exist") response arrives *after* the second
1438 * ("does exist"). In that case we ignore the second one.
1440 static void obj_request_existence_set(struct rbd_obj_request *obj_request,
1444 set_bit(OBJ_REQ_EXISTS, &obj_request->flags);
1445 set_bit(OBJ_REQ_KNOWN, &obj_request->flags);
1449 static bool obj_request_known_test(struct rbd_obj_request *obj_request)
1452 return test_bit(OBJ_REQ_KNOWN, &obj_request->flags) != 0;
1455 static bool obj_request_exists_test(struct rbd_obj_request *obj_request)
1458 return test_bit(OBJ_REQ_EXISTS, &obj_request->flags) != 0;
1461 static bool obj_request_overlaps_parent(struct rbd_obj_request *obj_request)
1463 struct rbd_device *rbd_dev = obj_request->img_request->rbd_dev;
1465 return obj_request->img_offset <
1466 round_up(rbd_dev->parent_overlap, rbd_obj_bytes(&rbd_dev->header));
1469 static void rbd_obj_request_get(struct rbd_obj_request *obj_request)
1471 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1472 atomic_read(&obj_request->kref.refcount));
1473 kref_get(&obj_request->kref);
1476 static void rbd_obj_request_destroy(struct kref *kref);
1477 static void rbd_obj_request_put(struct rbd_obj_request *obj_request)
1479 rbd_assert(obj_request != NULL);
1480 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1481 atomic_read(&obj_request->kref.refcount));
1482 kref_put(&obj_request->kref, rbd_obj_request_destroy);
1485 static void rbd_img_request_get(struct rbd_img_request *img_request)
1487 dout("%s: img %p (was %d)\n", __func__, img_request,
1488 atomic_read(&img_request->kref.refcount));
1489 kref_get(&img_request->kref);
1492 static bool img_request_child_test(struct rbd_img_request *img_request);
1493 static void rbd_parent_request_destroy(struct kref *kref);
1494 static void rbd_img_request_destroy(struct kref *kref);
1495 static void rbd_img_request_put(struct rbd_img_request *img_request)
1497 rbd_assert(img_request != NULL);
1498 dout("%s: img %p (was %d)\n", __func__, img_request,
1499 atomic_read(&img_request->kref.refcount));
1500 if (img_request_child_test(img_request))
1501 kref_put(&img_request->kref, rbd_parent_request_destroy);
1503 kref_put(&img_request->kref, rbd_img_request_destroy);
1506 static inline void rbd_img_obj_request_add(struct rbd_img_request *img_request,
1507 struct rbd_obj_request *obj_request)
1509 rbd_assert(obj_request->img_request == NULL);
1511 /* Image request now owns object's original reference */
1512 obj_request->img_request = img_request;
1513 obj_request->which = img_request->obj_request_count;
1514 rbd_assert(!obj_request_img_data_test(obj_request));
1515 obj_request_img_data_set(obj_request);
1516 rbd_assert(obj_request->which != BAD_WHICH);
1517 img_request->obj_request_count++;
1518 list_add_tail(&obj_request->links, &img_request->obj_requests);
1519 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1520 obj_request->which);
1523 static inline void rbd_img_obj_request_del(struct rbd_img_request *img_request,
1524 struct rbd_obj_request *obj_request)
1526 rbd_assert(obj_request->which != BAD_WHICH);
1528 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1529 obj_request->which);
1530 list_del(&obj_request->links);
1531 rbd_assert(img_request->obj_request_count > 0);
1532 img_request->obj_request_count--;
1533 rbd_assert(obj_request->which == img_request->obj_request_count);
1534 obj_request->which = BAD_WHICH;
1535 rbd_assert(obj_request_img_data_test(obj_request));
1536 rbd_assert(obj_request->img_request == img_request);
1537 obj_request->img_request = NULL;
1538 obj_request->callback = NULL;
1539 rbd_obj_request_put(obj_request);
1542 static bool obj_request_type_valid(enum obj_request_type type)
1545 case OBJ_REQUEST_NODATA:
1546 case OBJ_REQUEST_BIO:
1547 case OBJ_REQUEST_PAGES:
1554 static int rbd_obj_request_submit(struct ceph_osd_client *osdc,
1555 struct rbd_obj_request *obj_request)
1557 dout("%s %p\n", __func__, obj_request);
1558 return ceph_osdc_start_request(osdc, obj_request->osd_req, false);
1561 static void rbd_obj_request_end(struct rbd_obj_request *obj_request)
1563 dout("%s %p\n", __func__, obj_request);
1564 ceph_osdc_cancel_request(obj_request->osd_req);
1568 * Wait for an object request to complete. If interrupted, cancel the
1569 * underlying osd request.
1571 * @timeout: in jiffies, 0 means "wait forever"
1573 static int __rbd_obj_request_wait(struct rbd_obj_request *obj_request,
1574 unsigned long timeout)
1578 dout("%s %p\n", __func__, obj_request);
1579 ret = wait_for_completion_interruptible_timeout(
1580 &obj_request->completion,
1581 ceph_timeout_jiffies(timeout));
1585 rbd_obj_request_end(obj_request);
1590 dout("%s %p ret %d\n", __func__, obj_request, (int)ret);
1594 static int rbd_obj_request_wait(struct rbd_obj_request *obj_request)
1596 return __rbd_obj_request_wait(obj_request, 0);
1599 static int rbd_obj_request_wait_timeout(struct rbd_obj_request *obj_request,
1600 unsigned long timeout)
1602 return __rbd_obj_request_wait(obj_request, timeout);
1605 static void rbd_img_request_complete(struct rbd_img_request *img_request)
1608 dout("%s: img %p\n", __func__, img_request);
1611 * If no error occurred, compute the aggregate transfer
1612 * count for the image request. We could instead use
1613 * atomic64_cmpxchg() to update it as each object request
1614 * completes; not clear which way is better off hand.
1616 if (!img_request->result) {
1617 struct rbd_obj_request *obj_request;
1620 for_each_obj_request(img_request, obj_request)
1621 xferred += obj_request->xferred;
1622 img_request->xferred = xferred;
1625 if (img_request->callback)
1626 img_request->callback(img_request);
1628 rbd_img_request_put(img_request);
1632 * The default/initial value for all image request flags is 0. Each
1633 * is conditionally set to 1 at image request initialization time
1634 * and currently never change thereafter.
1636 static void img_request_write_set(struct rbd_img_request *img_request)
1638 set_bit(IMG_REQ_WRITE, &img_request->flags);
1642 static bool img_request_write_test(struct rbd_img_request *img_request)
1645 return test_bit(IMG_REQ_WRITE, &img_request->flags) != 0;
1649 * Set the discard flag when the img_request is an discard request
1651 static void img_request_discard_set(struct rbd_img_request *img_request)
1653 set_bit(IMG_REQ_DISCARD, &img_request->flags);
1657 static bool img_request_discard_test(struct rbd_img_request *img_request)
1660 return test_bit(IMG_REQ_DISCARD, &img_request->flags) != 0;
1663 static void img_request_child_set(struct rbd_img_request *img_request)
1665 set_bit(IMG_REQ_CHILD, &img_request->flags);
1669 static void img_request_child_clear(struct rbd_img_request *img_request)
1671 clear_bit(IMG_REQ_CHILD, &img_request->flags);
1675 static bool img_request_child_test(struct rbd_img_request *img_request)
1678 return test_bit(IMG_REQ_CHILD, &img_request->flags) != 0;
1681 static void img_request_layered_set(struct rbd_img_request *img_request)
1683 set_bit(IMG_REQ_LAYERED, &img_request->flags);
1687 static void img_request_layered_clear(struct rbd_img_request *img_request)
1689 clear_bit(IMG_REQ_LAYERED, &img_request->flags);
1693 static bool img_request_layered_test(struct rbd_img_request *img_request)
1696 return test_bit(IMG_REQ_LAYERED, &img_request->flags) != 0;
1699 static enum obj_operation_type
1700 rbd_img_request_op_type(struct rbd_img_request *img_request)
1702 if (img_request_write_test(img_request))
1703 return OBJ_OP_WRITE;
1704 else if (img_request_discard_test(img_request))
1705 return OBJ_OP_DISCARD;
1711 rbd_img_obj_request_read_callback(struct rbd_obj_request *obj_request)
1713 u64 xferred = obj_request->xferred;
1714 u64 length = obj_request->length;
1716 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1717 obj_request, obj_request->img_request, obj_request->result,
1720 * ENOENT means a hole in the image. We zero-fill the entire
1721 * length of the request. A short read also implies zero-fill
1722 * to the end of the request. An error requires the whole
1723 * length of the request to be reported finished with an error
1724 * to the block layer. In each case we update the xferred
1725 * count to indicate the whole request was satisfied.
1727 rbd_assert(obj_request->type != OBJ_REQUEST_NODATA);
1728 if (obj_request->result == -ENOENT) {
1729 if (obj_request->type == OBJ_REQUEST_BIO)
1730 zero_bio_chain(obj_request->bio_list, 0);
1732 zero_pages(obj_request->pages, 0, length);
1733 obj_request->result = 0;
1734 } else if (xferred < length && !obj_request->result) {
1735 if (obj_request->type == OBJ_REQUEST_BIO)
1736 zero_bio_chain(obj_request->bio_list, xferred);
1738 zero_pages(obj_request->pages, xferred, length);
1740 obj_request->xferred = length;
1741 obj_request_done_set(obj_request);
1744 static void rbd_obj_request_complete(struct rbd_obj_request *obj_request)
1746 dout("%s: obj %p cb %p\n", __func__, obj_request,
1747 obj_request->callback);
1748 if (obj_request->callback)
1749 obj_request->callback(obj_request);
1751 complete_all(&obj_request->completion);
1754 static void rbd_osd_trivial_callback(struct rbd_obj_request *obj_request)
1756 dout("%s: obj %p\n", __func__, obj_request);
1757 obj_request_done_set(obj_request);
1760 static void rbd_osd_read_callback(struct rbd_obj_request *obj_request)
1762 struct rbd_img_request *img_request = NULL;
1763 struct rbd_device *rbd_dev = NULL;
1764 bool layered = false;
1766 if (obj_request_img_data_test(obj_request)) {
1767 img_request = obj_request->img_request;
1768 layered = img_request && img_request_layered_test(img_request);
1769 rbd_dev = img_request->rbd_dev;
1772 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1773 obj_request, img_request, obj_request->result,
1774 obj_request->xferred, obj_request->length);
1775 if (layered && obj_request->result == -ENOENT &&
1776 obj_request->img_offset < rbd_dev->parent_overlap)
1777 rbd_img_parent_read(obj_request);
1778 else if (img_request)
1779 rbd_img_obj_request_read_callback(obj_request);
1781 obj_request_done_set(obj_request);
1784 static void rbd_osd_write_callback(struct rbd_obj_request *obj_request)
1786 dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1787 obj_request->result, obj_request->length);
1789 * There is no such thing as a successful short write. Set
1790 * it to our originally-requested length.
1792 obj_request->xferred = obj_request->length;
1793 obj_request_done_set(obj_request);
1796 static void rbd_osd_discard_callback(struct rbd_obj_request *obj_request)
1798 dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1799 obj_request->result, obj_request->length);
1801 * There is no such thing as a successful short discard. Set
1802 * it to our originally-requested length.
1804 obj_request->xferred = obj_request->length;
1805 /* discarding a non-existent object is not a problem */
1806 if (obj_request->result == -ENOENT)
1807 obj_request->result = 0;
1808 obj_request_done_set(obj_request);
1812 * For a simple stat call there's nothing to do. We'll do more if
1813 * this is part of a write sequence for a layered image.
1815 static void rbd_osd_stat_callback(struct rbd_obj_request *obj_request)
1817 dout("%s: obj %p\n", __func__, obj_request);
1818 obj_request_done_set(obj_request);
1821 static void rbd_osd_req_callback(struct ceph_osd_request *osd_req,
1822 struct ceph_msg *msg)
1824 struct rbd_obj_request *obj_request = osd_req->r_priv;
1827 dout("%s: osd_req %p msg %p\n", __func__, osd_req, msg);
1828 rbd_assert(osd_req == obj_request->osd_req);
1829 if (obj_request_img_data_test(obj_request)) {
1830 rbd_assert(obj_request->img_request);
1831 rbd_assert(obj_request->which != BAD_WHICH);
1833 rbd_assert(obj_request->which == BAD_WHICH);
1836 if (osd_req->r_result < 0)
1837 obj_request->result = osd_req->r_result;
1839 rbd_assert(osd_req->r_num_ops <= CEPH_OSD_MAX_OP);
1842 * We support a 64-bit length, but ultimately it has to be
1843 * passed to the block layer, which just supports a 32-bit
1846 obj_request->xferred = osd_req->r_reply_op_len[0];
1847 rbd_assert(obj_request->xferred < (u64)UINT_MAX);
1849 opcode = osd_req->r_ops[0].op;
1851 case CEPH_OSD_OP_READ:
1852 rbd_osd_read_callback(obj_request);
1854 case CEPH_OSD_OP_SETALLOCHINT:
1855 rbd_assert(osd_req->r_ops[1].op == CEPH_OSD_OP_WRITE);
1857 case CEPH_OSD_OP_WRITE:
1858 rbd_osd_write_callback(obj_request);
1860 case CEPH_OSD_OP_STAT:
1861 rbd_osd_stat_callback(obj_request);
1863 case CEPH_OSD_OP_DELETE:
1864 case CEPH_OSD_OP_TRUNCATE:
1865 case CEPH_OSD_OP_ZERO:
1866 rbd_osd_discard_callback(obj_request);
1868 case CEPH_OSD_OP_CALL:
1869 case CEPH_OSD_OP_NOTIFY_ACK:
1870 case CEPH_OSD_OP_WATCH:
1871 rbd_osd_trivial_callback(obj_request);
1874 rbd_warn(NULL, "%s: unsupported op %hu",
1875 obj_request->object_name, (unsigned short) opcode);
1879 if (obj_request_done_test(obj_request))
1880 rbd_obj_request_complete(obj_request);
1883 static void rbd_osd_req_format_read(struct rbd_obj_request *obj_request)
1885 struct rbd_img_request *img_request = obj_request->img_request;
1886 struct ceph_osd_request *osd_req = obj_request->osd_req;
1889 rbd_assert(osd_req != NULL);
1891 snap_id = img_request ? img_request->snap_id : CEPH_NOSNAP;
1892 ceph_osdc_build_request(osd_req, obj_request->offset,
1893 NULL, snap_id, NULL);
1896 static void rbd_osd_req_format_write(struct rbd_obj_request *obj_request)
1898 struct rbd_img_request *img_request = obj_request->img_request;
1899 struct ceph_osd_request *osd_req = obj_request->osd_req;
1900 struct ceph_snap_context *snapc;
1901 struct timespec mtime = CURRENT_TIME;
1903 rbd_assert(osd_req != NULL);
1905 snapc = img_request ? img_request->snapc : NULL;
1906 ceph_osdc_build_request(osd_req, obj_request->offset,
1907 snapc, CEPH_NOSNAP, &mtime);
1911 * Create an osd request. A read request has one osd op (read).
1912 * A write request has either one (watch) or two (hint+write) osd ops.
1913 * (All rbd data writes are prefixed with an allocation hint op, but
1914 * technically osd watch is a write request, hence this distinction.)
1916 static struct ceph_osd_request *rbd_osd_req_create(
1917 struct rbd_device *rbd_dev,
1918 enum obj_operation_type op_type,
1919 unsigned int num_ops,
1920 struct rbd_obj_request *obj_request)
1922 struct ceph_snap_context *snapc = NULL;
1923 struct ceph_osd_client *osdc;
1924 struct ceph_osd_request *osd_req;
1926 if (obj_request_img_data_test(obj_request) &&
1927 (op_type == OBJ_OP_DISCARD || op_type == OBJ_OP_WRITE)) {
1928 struct rbd_img_request *img_request = obj_request->img_request;
1929 if (op_type == OBJ_OP_WRITE) {
1930 rbd_assert(img_request_write_test(img_request));
1932 rbd_assert(img_request_discard_test(img_request));
1934 snapc = img_request->snapc;
1937 rbd_assert(num_ops == 1 || ((op_type == OBJ_OP_WRITE) && num_ops == 2));
1939 /* Allocate and initialize the request, for the num_ops ops */
1941 osdc = &rbd_dev->rbd_client->client->osdc;
1942 osd_req = ceph_osdc_alloc_request(osdc, snapc, num_ops, false,
1945 return NULL; /* ENOMEM */
1947 if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
1948 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
1950 osd_req->r_flags = CEPH_OSD_FLAG_READ;
1952 osd_req->r_callback = rbd_osd_req_callback;
1953 osd_req->r_priv = obj_request;
1955 osd_req->r_base_oloc.pool = ceph_file_layout_pg_pool(rbd_dev->layout);
1956 ceph_oid_set_name(&osd_req->r_base_oid, obj_request->object_name);
1962 * Create a copyup osd request based on the information in the object
1963 * request supplied. A copyup request has two or three osd ops, a
1964 * copyup method call, potentially a hint op, and a write or truncate
1967 static struct ceph_osd_request *
1968 rbd_osd_req_create_copyup(struct rbd_obj_request *obj_request)
1970 struct rbd_img_request *img_request;
1971 struct ceph_snap_context *snapc;
1972 struct rbd_device *rbd_dev;
1973 struct ceph_osd_client *osdc;
1974 struct ceph_osd_request *osd_req;
1975 int num_osd_ops = 3;
1977 rbd_assert(obj_request_img_data_test(obj_request));
1978 img_request = obj_request->img_request;
1979 rbd_assert(img_request);
1980 rbd_assert(img_request_write_test(img_request) ||
1981 img_request_discard_test(img_request));
1983 if (img_request_discard_test(img_request))
1986 /* Allocate and initialize the request, for all the ops */
1988 snapc = img_request->snapc;
1989 rbd_dev = img_request->rbd_dev;
1990 osdc = &rbd_dev->rbd_client->client->osdc;
1991 osd_req = ceph_osdc_alloc_request(osdc, snapc, num_osd_ops,
1994 return NULL; /* ENOMEM */
1996 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
1997 osd_req->r_callback = rbd_osd_req_callback;
1998 osd_req->r_priv = obj_request;
2000 osd_req->r_base_oloc.pool = ceph_file_layout_pg_pool(rbd_dev->layout);
2001 ceph_oid_set_name(&osd_req->r_base_oid, obj_request->object_name);
2007 static void rbd_osd_req_destroy(struct ceph_osd_request *osd_req)
2009 ceph_osdc_put_request(osd_req);
2012 /* object_name is assumed to be a non-null pointer and NUL-terminated */
2014 static struct rbd_obj_request *rbd_obj_request_create(const char *object_name,
2015 u64 offset, u64 length,
2016 enum obj_request_type type)
2018 struct rbd_obj_request *obj_request;
2022 rbd_assert(obj_request_type_valid(type));
2024 size = strlen(object_name) + 1;
2025 name = kmalloc(size, GFP_NOIO);
2029 obj_request = kmem_cache_zalloc(rbd_obj_request_cache, GFP_NOIO);
2035 obj_request->object_name = memcpy(name, object_name, size);
2036 obj_request->offset = offset;
2037 obj_request->length = length;
2038 obj_request->flags = 0;
2039 obj_request->which = BAD_WHICH;
2040 obj_request->type = type;
2041 INIT_LIST_HEAD(&obj_request->links);
2042 init_completion(&obj_request->completion);
2043 kref_init(&obj_request->kref);
2045 dout("%s: \"%s\" %llu/%llu %d -> obj %p\n", __func__, object_name,
2046 offset, length, (int)type, obj_request);
2051 static void rbd_obj_request_destroy(struct kref *kref)
2053 struct rbd_obj_request *obj_request;
2055 obj_request = container_of(kref, struct rbd_obj_request, kref);
2057 dout("%s: obj %p\n", __func__, obj_request);
2059 rbd_assert(obj_request->img_request == NULL);
2060 rbd_assert(obj_request->which == BAD_WHICH);
2062 if (obj_request->osd_req)
2063 rbd_osd_req_destroy(obj_request->osd_req);
2065 rbd_assert(obj_request_type_valid(obj_request->type));
2066 switch (obj_request->type) {
2067 case OBJ_REQUEST_NODATA:
2068 break; /* Nothing to do */
2069 case OBJ_REQUEST_BIO:
2070 if (obj_request->bio_list)
2071 bio_chain_put(obj_request->bio_list);
2073 case OBJ_REQUEST_PAGES:
2074 if (obj_request->pages)
2075 ceph_release_page_vector(obj_request->pages,
2076 obj_request->page_count);
2080 kfree(obj_request->object_name);
2081 obj_request->object_name = NULL;
2082 kmem_cache_free(rbd_obj_request_cache, obj_request);
2085 /* It's OK to call this for a device with no parent */
2087 static void rbd_spec_put(struct rbd_spec *spec);
2088 static void rbd_dev_unparent(struct rbd_device *rbd_dev)
2090 rbd_dev_remove_parent(rbd_dev);
2091 rbd_spec_put(rbd_dev->parent_spec);
2092 rbd_dev->parent_spec = NULL;
2093 rbd_dev->parent_overlap = 0;
2097 * Parent image reference counting is used to determine when an
2098 * image's parent fields can be safely torn down--after there are no
2099 * more in-flight requests to the parent image. When the last
2100 * reference is dropped, cleaning them up is safe.
2102 static void rbd_dev_parent_put(struct rbd_device *rbd_dev)
2106 if (!rbd_dev->parent_spec)
2109 counter = atomic_dec_return_safe(&rbd_dev->parent_ref);
2113 /* Last reference; clean up parent data structures */
2116 rbd_dev_unparent(rbd_dev);
2118 rbd_warn(rbd_dev, "parent reference underflow");
2122 * If an image has a non-zero parent overlap, get a reference to its
2125 * Returns true if the rbd device has a parent with a non-zero
2126 * overlap and a reference for it was successfully taken, or
2129 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
2133 if (!rbd_dev->parent_spec)
2136 down_read(&rbd_dev->header_rwsem);
2137 if (rbd_dev->parent_overlap)
2138 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
2139 up_read(&rbd_dev->header_rwsem);
2142 rbd_warn(rbd_dev, "parent reference overflow");
2148 * Caller is responsible for filling in the list of object requests
2149 * that comprises the image request, and the Linux request pointer
2150 * (if there is one).
2152 static struct rbd_img_request *rbd_img_request_create(
2153 struct rbd_device *rbd_dev,
2154 u64 offset, u64 length,
2155 enum obj_operation_type op_type,
2156 struct ceph_snap_context *snapc)
2158 struct rbd_img_request *img_request;
2160 img_request = kmem_cache_alloc(rbd_img_request_cache, GFP_NOIO);
2164 img_request->rq = NULL;
2165 img_request->rbd_dev = rbd_dev;
2166 img_request->offset = offset;
2167 img_request->length = length;
2168 img_request->flags = 0;
2169 if (op_type == OBJ_OP_DISCARD) {
2170 img_request_discard_set(img_request);
2171 img_request->snapc = snapc;
2172 } else if (op_type == OBJ_OP_WRITE) {
2173 img_request_write_set(img_request);
2174 img_request->snapc = snapc;
2176 img_request->snap_id = rbd_dev->spec->snap_id;
2178 if (rbd_dev_parent_get(rbd_dev))
2179 img_request_layered_set(img_request);
2180 spin_lock_init(&img_request->completion_lock);
2181 img_request->next_completion = 0;
2182 img_request->callback = NULL;
2183 img_request->result = 0;
2184 img_request->obj_request_count = 0;
2185 INIT_LIST_HEAD(&img_request->obj_requests);
2186 kref_init(&img_request->kref);
2188 dout("%s: rbd_dev %p %s %llu/%llu -> img %p\n", __func__, rbd_dev,
2189 obj_op_name(op_type), offset, length, img_request);
2194 static void rbd_img_request_destroy(struct kref *kref)
2196 struct rbd_img_request *img_request;
2197 struct rbd_obj_request *obj_request;
2198 struct rbd_obj_request *next_obj_request;
2200 img_request = container_of(kref, struct rbd_img_request, kref);
2202 dout("%s: img %p\n", __func__, img_request);
2204 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2205 rbd_img_obj_request_del(img_request, obj_request);
2206 rbd_assert(img_request->obj_request_count == 0);
2208 if (img_request_layered_test(img_request)) {
2209 img_request_layered_clear(img_request);
2210 rbd_dev_parent_put(img_request->rbd_dev);
2213 if (img_request_write_test(img_request) ||
2214 img_request_discard_test(img_request))
2215 ceph_put_snap_context(img_request->snapc);
2217 kmem_cache_free(rbd_img_request_cache, img_request);
2220 static struct rbd_img_request *rbd_parent_request_create(
2221 struct rbd_obj_request *obj_request,
2222 u64 img_offset, u64 length)
2224 struct rbd_img_request *parent_request;
2225 struct rbd_device *rbd_dev;
2227 rbd_assert(obj_request->img_request);
2228 rbd_dev = obj_request->img_request->rbd_dev;
2230 parent_request = rbd_img_request_create(rbd_dev->parent, img_offset,
2231 length, OBJ_OP_READ, NULL);
2232 if (!parent_request)
2235 img_request_child_set(parent_request);
2236 rbd_obj_request_get(obj_request);
2237 parent_request->obj_request = obj_request;
2239 return parent_request;
2242 static void rbd_parent_request_destroy(struct kref *kref)
2244 struct rbd_img_request *parent_request;
2245 struct rbd_obj_request *orig_request;
2247 parent_request = container_of(kref, struct rbd_img_request, kref);
2248 orig_request = parent_request->obj_request;
2250 parent_request->obj_request = NULL;
2251 rbd_obj_request_put(orig_request);
2252 img_request_child_clear(parent_request);
2254 rbd_img_request_destroy(kref);
2257 static bool rbd_img_obj_end_request(struct rbd_obj_request *obj_request)
2259 struct rbd_img_request *img_request;
2260 unsigned int xferred;
2264 rbd_assert(obj_request_img_data_test(obj_request));
2265 img_request = obj_request->img_request;
2267 rbd_assert(obj_request->xferred <= (u64)UINT_MAX);
2268 xferred = (unsigned int)obj_request->xferred;
2269 result = obj_request->result;
2271 struct rbd_device *rbd_dev = img_request->rbd_dev;
2272 enum obj_operation_type op_type;
2274 if (img_request_discard_test(img_request))
2275 op_type = OBJ_OP_DISCARD;
2276 else if (img_request_write_test(img_request))
2277 op_type = OBJ_OP_WRITE;
2279 op_type = OBJ_OP_READ;
2281 rbd_warn(rbd_dev, "%s %llx at %llx (%llx)",
2282 obj_op_name(op_type), obj_request->length,
2283 obj_request->img_offset, obj_request->offset);
2284 rbd_warn(rbd_dev, " result %d xferred %x",
2286 if (!img_request->result)
2287 img_request->result = result;
2289 * Need to end I/O on the entire obj_request worth of
2290 * bytes in case of error.
2292 xferred = obj_request->length;
2295 /* Image object requests don't own their page array */
2297 if (obj_request->type == OBJ_REQUEST_PAGES) {
2298 obj_request->pages = NULL;
2299 obj_request->page_count = 0;
2302 if (img_request_child_test(img_request)) {
2303 rbd_assert(img_request->obj_request != NULL);
2304 more = obj_request->which < img_request->obj_request_count - 1;
2306 rbd_assert(img_request->rq != NULL);
2308 more = blk_update_request(img_request->rq, result, xferred);
2310 __blk_mq_end_request(img_request->rq, result);
2316 static void rbd_img_obj_callback(struct rbd_obj_request *obj_request)
2318 struct rbd_img_request *img_request;
2319 u32 which = obj_request->which;
2322 rbd_assert(obj_request_img_data_test(obj_request));
2323 img_request = obj_request->img_request;
2325 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
2326 rbd_assert(img_request != NULL);
2327 rbd_assert(img_request->obj_request_count > 0);
2328 rbd_assert(which != BAD_WHICH);
2329 rbd_assert(which < img_request->obj_request_count);
2331 spin_lock_irq(&img_request->completion_lock);
2332 if (which != img_request->next_completion)
2335 for_each_obj_request_from(img_request, obj_request) {
2337 rbd_assert(which < img_request->obj_request_count);
2339 if (!obj_request_done_test(obj_request))
2341 more = rbd_img_obj_end_request(obj_request);
2345 rbd_assert(more ^ (which == img_request->obj_request_count));
2346 img_request->next_completion = which;
2348 spin_unlock_irq(&img_request->completion_lock);
2349 rbd_img_request_put(img_request);
2352 rbd_img_request_complete(img_request);
2356 * Add individual osd ops to the given ceph_osd_request and prepare
2357 * them for submission. num_ops is the current number of
2358 * osd operations already to the object request.
2360 static void rbd_img_obj_request_fill(struct rbd_obj_request *obj_request,
2361 struct ceph_osd_request *osd_request,
2362 enum obj_operation_type op_type,
2363 unsigned int num_ops)
2365 struct rbd_img_request *img_request = obj_request->img_request;
2366 struct rbd_device *rbd_dev = img_request->rbd_dev;
2367 u64 object_size = rbd_obj_bytes(&rbd_dev->header);
2368 u64 offset = obj_request->offset;
2369 u64 length = obj_request->length;
2373 if (op_type == OBJ_OP_DISCARD) {
2374 if (!offset && length == object_size &&
2375 (!img_request_layered_test(img_request) ||
2376 !obj_request_overlaps_parent(obj_request))) {
2377 opcode = CEPH_OSD_OP_DELETE;
2378 } else if ((offset + length == object_size)) {
2379 opcode = CEPH_OSD_OP_TRUNCATE;
2381 down_read(&rbd_dev->header_rwsem);
2382 img_end = rbd_dev->header.image_size;
2383 up_read(&rbd_dev->header_rwsem);
2385 if (obj_request->img_offset + length == img_end)
2386 opcode = CEPH_OSD_OP_TRUNCATE;
2388 opcode = CEPH_OSD_OP_ZERO;
2390 } else if (op_type == OBJ_OP_WRITE) {
2391 opcode = CEPH_OSD_OP_WRITE;
2392 osd_req_op_alloc_hint_init(osd_request, num_ops,
2393 object_size, object_size);
2396 opcode = CEPH_OSD_OP_READ;
2399 if (opcode == CEPH_OSD_OP_DELETE)
2400 osd_req_op_init(osd_request, num_ops, opcode, 0);
2402 osd_req_op_extent_init(osd_request, num_ops, opcode,
2403 offset, length, 0, 0);
2405 if (obj_request->type == OBJ_REQUEST_BIO)
2406 osd_req_op_extent_osd_data_bio(osd_request, num_ops,
2407 obj_request->bio_list, length);
2408 else if (obj_request->type == OBJ_REQUEST_PAGES)
2409 osd_req_op_extent_osd_data_pages(osd_request, num_ops,
2410 obj_request->pages, length,
2411 offset & ~PAGE_MASK, false, false);
2413 /* Discards are also writes */
2414 if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
2415 rbd_osd_req_format_write(obj_request);
2417 rbd_osd_req_format_read(obj_request);
2421 * Split up an image request into one or more object requests, each
2422 * to a different object. The "type" parameter indicates whether
2423 * "data_desc" is the pointer to the head of a list of bio
2424 * structures, or the base of a page array. In either case this
2425 * function assumes data_desc describes memory sufficient to hold
2426 * all data described by the image request.
2428 static int rbd_img_request_fill(struct rbd_img_request *img_request,
2429 enum obj_request_type type,
2432 struct rbd_device *rbd_dev = img_request->rbd_dev;
2433 struct rbd_obj_request *obj_request = NULL;
2434 struct rbd_obj_request *next_obj_request;
2435 struct bio *bio_list = NULL;
2436 unsigned int bio_offset = 0;
2437 struct page **pages = NULL;
2438 enum obj_operation_type op_type;
2442 dout("%s: img %p type %d data_desc %p\n", __func__, img_request,
2443 (int)type, data_desc);
2445 img_offset = img_request->offset;
2446 resid = img_request->length;
2447 rbd_assert(resid > 0);
2448 op_type = rbd_img_request_op_type(img_request);
2450 if (type == OBJ_REQUEST_BIO) {
2451 bio_list = data_desc;
2452 rbd_assert(img_offset ==
2453 bio_list->bi_iter.bi_sector << SECTOR_SHIFT);
2454 } else if (type == OBJ_REQUEST_PAGES) {
2459 struct ceph_osd_request *osd_req;
2460 const char *object_name;
2464 object_name = rbd_segment_name(rbd_dev, img_offset);
2467 offset = rbd_segment_offset(rbd_dev, img_offset);
2468 length = rbd_segment_length(rbd_dev, img_offset, resid);
2469 obj_request = rbd_obj_request_create(object_name,
2470 offset, length, type);
2471 /* object request has its own copy of the object name */
2472 rbd_segment_name_free(object_name);
2477 * set obj_request->img_request before creating the
2478 * osd_request so that it gets the right snapc
2480 rbd_img_obj_request_add(img_request, obj_request);
2482 if (type == OBJ_REQUEST_BIO) {
2483 unsigned int clone_size;
2485 rbd_assert(length <= (u64)UINT_MAX);
2486 clone_size = (unsigned int)length;
2487 obj_request->bio_list =
2488 bio_chain_clone_range(&bio_list,
2492 if (!obj_request->bio_list)
2494 } else if (type == OBJ_REQUEST_PAGES) {
2495 unsigned int page_count;
2497 obj_request->pages = pages;
2498 page_count = (u32)calc_pages_for(offset, length);
2499 obj_request->page_count = page_count;
2500 if ((offset + length) & ~PAGE_MASK)
2501 page_count--; /* more on last page */
2502 pages += page_count;
2505 osd_req = rbd_osd_req_create(rbd_dev, op_type,
2506 (op_type == OBJ_OP_WRITE) ? 2 : 1,
2511 obj_request->osd_req = osd_req;
2512 obj_request->callback = rbd_img_obj_callback;
2513 obj_request->img_offset = img_offset;
2515 rbd_img_obj_request_fill(obj_request, osd_req, op_type, 0);
2517 rbd_img_request_get(img_request);
2519 img_offset += length;
2526 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2527 rbd_img_obj_request_del(img_request, obj_request);
2533 rbd_img_obj_copyup_callback(struct rbd_obj_request *obj_request)
2535 struct rbd_img_request *img_request;
2536 struct rbd_device *rbd_dev;
2537 struct page **pages;
2540 rbd_assert(obj_request->type == OBJ_REQUEST_BIO ||
2541 obj_request->type == OBJ_REQUEST_NODATA);
2542 rbd_assert(obj_request_img_data_test(obj_request));
2543 img_request = obj_request->img_request;
2544 rbd_assert(img_request);
2546 rbd_dev = img_request->rbd_dev;
2547 rbd_assert(rbd_dev);
2549 pages = obj_request->copyup_pages;
2550 rbd_assert(pages != NULL);
2551 obj_request->copyup_pages = NULL;
2552 page_count = obj_request->copyup_page_count;
2553 rbd_assert(page_count);
2554 obj_request->copyup_page_count = 0;
2555 ceph_release_page_vector(pages, page_count);
2558 * We want the transfer count to reflect the size of the
2559 * original write request. There is no such thing as a
2560 * successful short write, so if the request was successful
2561 * we can just set it to the originally-requested length.
2563 if (!obj_request->result)
2564 obj_request->xferred = obj_request->length;
2566 /* Finish up with the normal image object callback */
2568 rbd_img_obj_callback(obj_request);
2572 rbd_img_obj_parent_read_full_callback(struct rbd_img_request *img_request)
2574 struct rbd_obj_request *orig_request;
2575 struct ceph_osd_request *osd_req;
2576 struct ceph_osd_client *osdc;
2577 struct rbd_device *rbd_dev;
2578 struct page **pages;
2579 enum obj_operation_type op_type;
2584 rbd_assert(img_request_child_test(img_request));
2586 /* First get what we need from the image request */
2588 pages = img_request->copyup_pages;
2589 rbd_assert(pages != NULL);
2590 img_request->copyup_pages = NULL;
2591 page_count = img_request->copyup_page_count;
2592 rbd_assert(page_count);
2593 img_request->copyup_page_count = 0;
2595 orig_request = img_request->obj_request;
2596 rbd_assert(orig_request != NULL);
2597 rbd_assert(obj_request_type_valid(orig_request->type));
2598 img_result = img_request->result;
2599 parent_length = img_request->length;
2600 rbd_assert(parent_length == img_request->xferred);
2601 rbd_img_request_put(img_request);
2603 rbd_assert(orig_request->img_request);
2604 rbd_dev = orig_request->img_request->rbd_dev;
2605 rbd_assert(rbd_dev);
2608 * If the overlap has become 0 (most likely because the
2609 * image has been flattened) we need to free the pages
2610 * and re-submit the original write request.
2612 if (!rbd_dev->parent_overlap) {
2613 struct ceph_osd_client *osdc;
2615 ceph_release_page_vector(pages, page_count);
2616 osdc = &rbd_dev->rbd_client->client->osdc;
2617 img_result = rbd_obj_request_submit(osdc, orig_request);
2626 * The original osd request is of no use to use any more.
2627 * We need a new one that can hold the three ops in a copyup
2628 * request. Allocate the new copyup osd request for the
2629 * original request, and release the old one.
2631 img_result = -ENOMEM;
2632 osd_req = rbd_osd_req_create_copyup(orig_request);
2635 rbd_osd_req_destroy(orig_request->osd_req);
2636 orig_request->osd_req = osd_req;
2637 orig_request->copyup_pages = pages;
2638 orig_request->copyup_page_count = page_count;
2640 /* Initialize the copyup op */
2642 osd_req_op_cls_init(osd_req, 0, CEPH_OSD_OP_CALL, "rbd", "copyup");
2643 osd_req_op_cls_request_data_pages(osd_req, 0, pages, parent_length, 0,
2646 /* Add the other op(s) */
2648 op_type = rbd_img_request_op_type(orig_request->img_request);
2649 rbd_img_obj_request_fill(orig_request, osd_req, op_type, 1);
2651 /* All set, send it off. */
2653 orig_request->callback = rbd_img_obj_copyup_callback;
2654 osdc = &rbd_dev->rbd_client->client->osdc;
2655 img_result = rbd_obj_request_submit(osdc, orig_request);
2659 /* Record the error code and complete the request */
2661 orig_request->result = img_result;
2662 orig_request->xferred = 0;
2663 obj_request_done_set(orig_request);
2664 rbd_obj_request_complete(orig_request);
2668 * Read from the parent image the range of data that covers the
2669 * entire target of the given object request. This is used for
2670 * satisfying a layered image write request when the target of an
2671 * object request from the image request does not exist.
2673 * A page array big enough to hold the returned data is allocated
2674 * and supplied to rbd_img_request_fill() as the "data descriptor."
2675 * When the read completes, this page array will be transferred to
2676 * the original object request for the copyup operation.
2678 * If an error occurs, record it as the result of the original
2679 * object request and mark it done so it gets completed.
2681 static int rbd_img_obj_parent_read_full(struct rbd_obj_request *obj_request)
2683 struct rbd_img_request *img_request = NULL;
2684 struct rbd_img_request *parent_request = NULL;
2685 struct rbd_device *rbd_dev;
2688 struct page **pages = NULL;
2692 rbd_assert(obj_request_img_data_test(obj_request));
2693 rbd_assert(obj_request_type_valid(obj_request->type));
2695 img_request = obj_request->img_request;
2696 rbd_assert(img_request != NULL);
2697 rbd_dev = img_request->rbd_dev;
2698 rbd_assert(rbd_dev->parent != NULL);
2701 * Determine the byte range covered by the object in the
2702 * child image to which the original request was to be sent.
2704 img_offset = obj_request->img_offset - obj_request->offset;
2705 length = (u64)1 << rbd_dev->header.obj_order;
2708 * There is no defined parent data beyond the parent
2709 * overlap, so limit what we read at that boundary if
2712 if (img_offset + length > rbd_dev->parent_overlap) {
2713 rbd_assert(img_offset < rbd_dev->parent_overlap);
2714 length = rbd_dev->parent_overlap - img_offset;
2718 * Allocate a page array big enough to receive the data read
2721 page_count = (u32)calc_pages_for(0, length);
2722 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2723 if (IS_ERR(pages)) {
2724 result = PTR_ERR(pages);
2730 parent_request = rbd_parent_request_create(obj_request,
2731 img_offset, length);
2732 if (!parent_request)
2735 result = rbd_img_request_fill(parent_request, OBJ_REQUEST_PAGES, pages);
2738 parent_request->copyup_pages = pages;
2739 parent_request->copyup_page_count = page_count;
2741 parent_request->callback = rbd_img_obj_parent_read_full_callback;
2742 result = rbd_img_request_submit(parent_request);
2746 parent_request->copyup_pages = NULL;
2747 parent_request->copyup_page_count = 0;
2748 parent_request->obj_request = NULL;
2749 rbd_obj_request_put(obj_request);
2752 ceph_release_page_vector(pages, page_count);
2754 rbd_img_request_put(parent_request);
2755 obj_request->result = result;
2756 obj_request->xferred = 0;
2757 obj_request_done_set(obj_request);
2762 static void rbd_img_obj_exists_callback(struct rbd_obj_request *obj_request)
2764 struct rbd_obj_request *orig_request;
2765 struct rbd_device *rbd_dev;
2768 rbd_assert(!obj_request_img_data_test(obj_request));
2771 * All we need from the object request is the original
2772 * request and the result of the STAT op. Grab those, then
2773 * we're done with the request.
2775 orig_request = obj_request->obj_request;
2776 obj_request->obj_request = NULL;
2777 rbd_obj_request_put(orig_request);
2778 rbd_assert(orig_request);
2779 rbd_assert(orig_request->img_request);
2781 result = obj_request->result;
2782 obj_request->result = 0;
2784 dout("%s: obj %p for obj %p result %d %llu/%llu\n", __func__,
2785 obj_request, orig_request, result,
2786 obj_request->xferred, obj_request->length);
2787 rbd_obj_request_put(obj_request);
2790 * If the overlap has become 0 (most likely because the
2791 * image has been flattened) we need to free the pages
2792 * and re-submit the original write request.
2794 rbd_dev = orig_request->img_request->rbd_dev;
2795 if (!rbd_dev->parent_overlap) {
2796 struct ceph_osd_client *osdc;
2798 osdc = &rbd_dev->rbd_client->client->osdc;
2799 result = rbd_obj_request_submit(osdc, orig_request);
2805 * Our only purpose here is to determine whether the object
2806 * exists, and we don't want to treat the non-existence as
2807 * an error. If something else comes back, transfer the
2808 * error to the original request and complete it now.
2811 obj_request_existence_set(orig_request, true);
2812 } else if (result == -ENOENT) {
2813 obj_request_existence_set(orig_request, false);
2814 } else if (result) {
2815 orig_request->result = result;
2820 * Resubmit the original request now that we have recorded
2821 * whether the target object exists.
2823 orig_request->result = rbd_img_obj_request_submit(orig_request);
2825 if (orig_request->result)
2826 rbd_obj_request_complete(orig_request);
2829 static int rbd_img_obj_exists_submit(struct rbd_obj_request *obj_request)
2831 struct rbd_obj_request *stat_request;
2832 struct rbd_device *rbd_dev;
2833 struct ceph_osd_client *osdc;
2834 struct page **pages = NULL;
2840 * The response data for a STAT call consists of:
2847 size = sizeof (__le64) + sizeof (__le32) + sizeof (__le32);
2848 page_count = (u32)calc_pages_for(0, size);
2849 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2851 return PTR_ERR(pages);
2854 stat_request = rbd_obj_request_create(obj_request->object_name, 0, 0,
2859 rbd_obj_request_get(obj_request);
2860 stat_request->obj_request = obj_request;
2861 stat_request->pages = pages;
2862 stat_request->page_count = page_count;
2864 rbd_assert(obj_request->img_request);
2865 rbd_dev = obj_request->img_request->rbd_dev;
2866 stat_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
2868 if (!stat_request->osd_req)
2870 stat_request->callback = rbd_img_obj_exists_callback;
2872 osd_req_op_init(stat_request->osd_req, 0, CEPH_OSD_OP_STAT, 0);
2873 osd_req_op_raw_data_in_pages(stat_request->osd_req, 0, pages, size, 0,
2875 rbd_osd_req_format_read(stat_request);
2877 osdc = &rbd_dev->rbd_client->client->osdc;
2878 ret = rbd_obj_request_submit(osdc, stat_request);
2881 rbd_obj_request_put(obj_request);
2886 static bool img_obj_request_simple(struct rbd_obj_request *obj_request)
2888 struct rbd_img_request *img_request;
2889 struct rbd_device *rbd_dev;
2891 rbd_assert(obj_request_img_data_test(obj_request));
2893 img_request = obj_request->img_request;
2894 rbd_assert(img_request);
2895 rbd_dev = img_request->rbd_dev;
2898 if (!img_request_write_test(img_request) &&
2899 !img_request_discard_test(img_request))
2902 /* Non-layered writes */
2903 if (!img_request_layered_test(img_request))
2907 * Layered writes outside of the parent overlap range don't
2908 * share any data with the parent.
2910 if (!obj_request_overlaps_parent(obj_request))
2914 * Entire-object layered writes - we will overwrite whatever
2915 * parent data there is anyway.
2917 if (!obj_request->offset &&
2918 obj_request->length == rbd_obj_bytes(&rbd_dev->header))
2922 * If the object is known to already exist, its parent data has
2923 * already been copied.
2925 if (obj_request_known_test(obj_request) &&
2926 obj_request_exists_test(obj_request))
2932 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request)
2934 if (img_obj_request_simple(obj_request)) {
2935 struct rbd_device *rbd_dev;
2936 struct ceph_osd_client *osdc;
2938 rbd_dev = obj_request->img_request->rbd_dev;
2939 osdc = &rbd_dev->rbd_client->client->osdc;
2941 return rbd_obj_request_submit(osdc, obj_request);
2945 * It's a layered write. The target object might exist but
2946 * we may not know that yet. If we know it doesn't exist,
2947 * start by reading the data for the full target object from
2948 * the parent so we can use it for a copyup to the target.
2950 if (obj_request_known_test(obj_request))
2951 return rbd_img_obj_parent_read_full(obj_request);
2953 /* We don't know whether the target exists. Go find out. */
2955 return rbd_img_obj_exists_submit(obj_request);
2958 static int rbd_img_request_submit(struct rbd_img_request *img_request)
2960 struct rbd_obj_request *obj_request;
2961 struct rbd_obj_request *next_obj_request;
2963 dout("%s: img %p\n", __func__, img_request);
2964 for_each_obj_request_safe(img_request, obj_request, next_obj_request) {
2967 ret = rbd_img_obj_request_submit(obj_request);
2975 static void rbd_img_parent_read_callback(struct rbd_img_request *img_request)
2977 struct rbd_obj_request *obj_request;
2978 struct rbd_device *rbd_dev;
2983 rbd_assert(img_request_child_test(img_request));
2985 /* First get what we need from the image request and release it */
2987 obj_request = img_request->obj_request;
2988 img_xferred = img_request->xferred;
2989 img_result = img_request->result;
2990 rbd_img_request_put(img_request);
2993 * If the overlap has become 0 (most likely because the
2994 * image has been flattened) we need to re-submit the
2997 rbd_assert(obj_request);
2998 rbd_assert(obj_request->img_request);
2999 rbd_dev = obj_request->img_request->rbd_dev;
3000 if (!rbd_dev->parent_overlap) {
3001 struct ceph_osd_client *osdc;
3003 osdc = &rbd_dev->rbd_client->client->osdc;
3004 img_result = rbd_obj_request_submit(osdc, obj_request);
3009 obj_request->result = img_result;
3010 if (obj_request->result)
3014 * We need to zero anything beyond the parent overlap
3015 * boundary. Since rbd_img_obj_request_read_callback()
3016 * will zero anything beyond the end of a short read, an
3017 * easy way to do this is to pretend the data from the
3018 * parent came up short--ending at the overlap boundary.
3020 rbd_assert(obj_request->img_offset < U64_MAX - obj_request->length);
3021 obj_end = obj_request->img_offset + obj_request->length;
3022 if (obj_end > rbd_dev->parent_overlap) {
3025 if (obj_request->img_offset < rbd_dev->parent_overlap)
3026 xferred = rbd_dev->parent_overlap -
3027 obj_request->img_offset;
3029 obj_request->xferred = min(img_xferred, xferred);
3031 obj_request->xferred = img_xferred;
3034 rbd_img_obj_request_read_callback(obj_request);
3035 rbd_obj_request_complete(obj_request);
3038 static void rbd_img_parent_read(struct rbd_obj_request *obj_request)
3040 struct rbd_img_request *img_request;
3043 rbd_assert(obj_request_img_data_test(obj_request));
3044 rbd_assert(obj_request->img_request != NULL);
3045 rbd_assert(obj_request->result == (s32) -ENOENT);
3046 rbd_assert(obj_request_type_valid(obj_request->type));
3048 /* rbd_read_finish(obj_request, obj_request->length); */
3049 img_request = rbd_parent_request_create(obj_request,
3050 obj_request->img_offset,
3051 obj_request->length);
3056 if (obj_request->type == OBJ_REQUEST_BIO)
3057 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3058 obj_request->bio_list);
3060 result = rbd_img_request_fill(img_request, OBJ_REQUEST_PAGES,
3061 obj_request->pages);
3065 img_request->callback = rbd_img_parent_read_callback;
3066 result = rbd_img_request_submit(img_request);
3073 rbd_img_request_put(img_request);
3074 obj_request->result = result;
3075 obj_request->xferred = 0;
3076 obj_request_done_set(obj_request);
3079 static int rbd_obj_notify_ack_sync(struct rbd_device *rbd_dev, u64 notify_id)
3081 struct rbd_obj_request *obj_request;
3082 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3085 obj_request = rbd_obj_request_create(rbd_dev->header_name, 0, 0,
3086 OBJ_REQUEST_NODATA);
3091 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
3093 if (!obj_request->osd_req)
3096 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_NOTIFY_ACK,
3098 rbd_osd_req_format_read(obj_request);
3100 ret = rbd_obj_request_submit(osdc, obj_request);
3103 ret = rbd_obj_request_wait(obj_request);
3105 rbd_obj_request_put(obj_request);
3110 static void rbd_watch_cb(u64 ver, u64 notify_id, u8 opcode, void *data)
3112 struct rbd_device *rbd_dev = (struct rbd_device *)data;
3118 dout("%s: \"%s\" notify_id %llu opcode %u\n", __func__,
3119 rbd_dev->header_name, (unsigned long long)notify_id,
3120 (unsigned int)opcode);
3123 * Until adequate refresh error handling is in place, there is
3124 * not much we can do here, except warn.
3126 * See http://tracker.ceph.com/issues/5040
3128 ret = rbd_dev_refresh(rbd_dev);
3130 rbd_warn(rbd_dev, "refresh failed: %d", ret);
3132 ret = rbd_obj_notify_ack_sync(rbd_dev, notify_id);
3134 rbd_warn(rbd_dev, "notify_ack ret %d", ret);
3138 * Send a (un)watch request and wait for the ack. Return a request
3139 * with a ref held on success or error.
3141 static struct rbd_obj_request *rbd_obj_watch_request_helper(
3142 struct rbd_device *rbd_dev,
3145 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3146 struct ceph_options *opts = osdc->client->options;
3147 struct rbd_obj_request *obj_request;
3150 obj_request = rbd_obj_request_create(rbd_dev->header_name, 0, 0,
3151 OBJ_REQUEST_NODATA);
3153 return ERR_PTR(-ENOMEM);
3155 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_WRITE, 1,
3157 if (!obj_request->osd_req) {
3162 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_WATCH,
3163 rbd_dev->watch_event->cookie, 0, watch);
3164 rbd_osd_req_format_write(obj_request);
3167 ceph_osdc_set_request_linger(osdc, obj_request->osd_req);
3169 ret = rbd_obj_request_submit(osdc, obj_request);
3173 ret = rbd_obj_request_wait_timeout(obj_request, opts->mount_timeout);
3177 ret = obj_request->result;
3180 rbd_obj_request_end(obj_request);
3187 rbd_obj_request_put(obj_request);
3188 return ERR_PTR(ret);
3192 * Initiate a watch request, synchronously.
3194 static int rbd_dev_header_watch_sync(struct rbd_device *rbd_dev)
3196 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3197 struct rbd_obj_request *obj_request;
3200 rbd_assert(!rbd_dev->watch_event);
3201 rbd_assert(!rbd_dev->watch_request);
3203 ret = ceph_osdc_create_event(osdc, rbd_watch_cb, rbd_dev,
3204 &rbd_dev->watch_event);
3208 obj_request = rbd_obj_watch_request_helper(rbd_dev, true);
3209 if (IS_ERR(obj_request)) {
3210 ceph_osdc_cancel_event(rbd_dev->watch_event);
3211 rbd_dev->watch_event = NULL;
3212 return PTR_ERR(obj_request);
3216 * A watch request is set to linger, so the underlying osd
3217 * request won't go away until we unregister it. We retain
3218 * a pointer to the object request during that time (in
3219 * rbd_dev->watch_request), so we'll keep a reference to it.
3220 * We'll drop that reference after we've unregistered it in
3221 * rbd_dev_header_unwatch_sync().
3223 rbd_dev->watch_request = obj_request;
3229 * Tear down a watch request, synchronously.
3231 static void rbd_dev_header_unwatch_sync(struct rbd_device *rbd_dev)
3233 struct rbd_obj_request *obj_request;
3235 rbd_assert(rbd_dev->watch_event);
3236 rbd_assert(rbd_dev->watch_request);
3238 rbd_obj_request_end(rbd_dev->watch_request);
3239 rbd_obj_request_put(rbd_dev->watch_request);
3240 rbd_dev->watch_request = NULL;
3242 obj_request = rbd_obj_watch_request_helper(rbd_dev, false);
3243 if (!IS_ERR(obj_request))
3244 rbd_obj_request_put(obj_request);
3246 rbd_warn(rbd_dev, "unable to tear down watch request (%ld)",
3247 PTR_ERR(obj_request));
3249 ceph_osdc_cancel_event(rbd_dev->watch_event);
3250 rbd_dev->watch_event = NULL;
3254 * Synchronous osd object method call. Returns the number of bytes
3255 * returned in the outbound buffer, or a negative error code.
3257 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
3258 const char *object_name,
3259 const char *class_name,
3260 const char *method_name,
3261 const void *outbound,
3262 size_t outbound_size,
3264 size_t inbound_size)
3266 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3267 struct rbd_obj_request *obj_request;
3268 struct page **pages;
3273 * Method calls are ultimately read operations. The result
3274 * should placed into the inbound buffer provided. They
3275 * also supply outbound data--parameters for the object
3276 * method. Currently if this is present it will be a
3279 page_count = (u32)calc_pages_for(0, inbound_size);
3280 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
3282 return PTR_ERR(pages);
3285 obj_request = rbd_obj_request_create(object_name, 0, inbound_size,
3290 obj_request->pages = pages;
3291 obj_request->page_count = page_count;
3293 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
3295 if (!obj_request->osd_req)
3298 osd_req_op_cls_init(obj_request->osd_req, 0, CEPH_OSD_OP_CALL,
3299 class_name, method_name);
3300 if (outbound_size) {
3301 struct ceph_pagelist *pagelist;
3303 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
3307 ceph_pagelist_init(pagelist);
3308 ceph_pagelist_append(pagelist, outbound, outbound_size);
3309 osd_req_op_cls_request_data_pagelist(obj_request->osd_req, 0,
3312 osd_req_op_cls_response_data_pages(obj_request->osd_req, 0,
3313 obj_request->pages, inbound_size,
3315 rbd_osd_req_format_read(obj_request);
3317 ret = rbd_obj_request_submit(osdc, obj_request);
3320 ret = rbd_obj_request_wait(obj_request);
3324 ret = obj_request->result;
3328 rbd_assert(obj_request->xferred < (u64)INT_MAX);
3329 ret = (int)obj_request->xferred;
3330 ceph_copy_from_page_vector(pages, inbound, 0, obj_request->xferred);
3333 rbd_obj_request_put(obj_request);
3335 ceph_release_page_vector(pages, page_count);
3340 static void rbd_queue_workfn(struct work_struct *work)
3342 struct request *rq = blk_mq_rq_from_pdu(work);
3343 struct rbd_device *rbd_dev = rq->q->queuedata;
3344 struct rbd_img_request *img_request;
3345 struct ceph_snap_context *snapc = NULL;
3346 u64 offset = (u64)blk_rq_pos(rq) << SECTOR_SHIFT;
3347 u64 length = blk_rq_bytes(rq);
3348 enum obj_operation_type op_type;
3352 if (rq->cmd_type != REQ_TYPE_FS) {
3353 dout("%s: non-fs request type %d\n", __func__,
3354 (int) rq->cmd_type);
3359 if (rq->cmd_flags & REQ_DISCARD)
3360 op_type = OBJ_OP_DISCARD;
3361 else if (rq->cmd_flags & REQ_WRITE)
3362 op_type = OBJ_OP_WRITE;
3364 op_type = OBJ_OP_READ;
3366 /* Ignore/skip any zero-length requests */
3369 dout("%s: zero-length request\n", __func__);
3374 /* Only reads are allowed to a read-only device */
3376 if (op_type != OBJ_OP_READ) {
3377 if (rbd_dev->mapping.read_only) {
3381 rbd_assert(rbd_dev->spec->snap_id == CEPH_NOSNAP);
3385 * Quit early if the mapped snapshot no longer exists. It's
3386 * still possible the snapshot will have disappeared by the
3387 * time our request arrives at the osd, but there's no sense in
3388 * sending it if we already know.
3390 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
3391 dout("request for non-existent snapshot");
3392 rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
3397 if (offset && length > U64_MAX - offset + 1) {
3398 rbd_warn(rbd_dev, "bad request range (%llu~%llu)", offset,
3401 goto err_rq; /* Shouldn't happen */
3404 blk_mq_start_request(rq);
3406 down_read(&rbd_dev->header_rwsem);
3407 mapping_size = rbd_dev->mapping.size;
3408 if (op_type != OBJ_OP_READ) {
3409 snapc = rbd_dev->header.snapc;
3410 ceph_get_snap_context(snapc);
3412 up_read(&rbd_dev->header_rwsem);
3414 if (offset + length > mapping_size) {
3415 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)", offset,
3416 length, mapping_size);
3421 img_request = rbd_img_request_create(rbd_dev, offset, length, op_type,
3427 img_request->rq = rq;
3429 if (op_type == OBJ_OP_DISCARD)
3430 result = rbd_img_request_fill(img_request, OBJ_REQUEST_NODATA,
3433 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3436 goto err_img_request;
3438 result = rbd_img_request_submit(img_request);
3440 goto err_img_request;
3445 rbd_img_request_put(img_request);
3448 rbd_warn(rbd_dev, "%s %llx at %llx result %d",
3449 obj_op_name(op_type), length, offset, result);
3450 ceph_put_snap_context(snapc);
3452 blk_mq_end_request(rq, result);
3455 static int rbd_queue_rq(struct blk_mq_hw_ctx *hctx,
3456 const struct blk_mq_queue_data *bd)
3458 struct request *rq = bd->rq;
3459 struct work_struct *work = blk_mq_rq_to_pdu(rq);
3461 queue_work(rbd_wq, work);
3462 return BLK_MQ_RQ_QUEUE_OK;
3466 * a queue callback. Makes sure that we don't create a bio that spans across
3467 * multiple osd objects. One exception would be with a single page bios,
3468 * which we handle later at bio_chain_clone_range()
3470 static int rbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bmd,
3471 struct bio_vec *bvec)
3473 struct rbd_device *rbd_dev = q->queuedata;
3474 sector_t sector_offset;
3475 sector_t sectors_per_obj;
3476 sector_t obj_sector_offset;
3480 * Find how far into its rbd object the partition-relative
3481 * bio start sector is to offset relative to the enclosing
3484 sector_offset = get_start_sect(bmd->bi_bdev) + bmd->bi_sector;
3485 sectors_per_obj = 1 << (rbd_dev->header.obj_order - SECTOR_SHIFT);
3486 obj_sector_offset = sector_offset & (sectors_per_obj - 1);
3489 * Compute the number of bytes from that offset to the end
3490 * of the object. Account for what's already used by the bio.
3492 ret = (int) (sectors_per_obj - obj_sector_offset) << SECTOR_SHIFT;
3493 if (ret > bmd->bi_size)
3494 ret -= bmd->bi_size;
3499 * Don't send back more than was asked for. And if the bio
3500 * was empty, let the whole thing through because: "Note
3501 * that a block device *must* allow a single page to be
3502 * added to an empty bio."
3504 rbd_assert(bvec->bv_len <= PAGE_SIZE);
3505 if (ret > (int) bvec->bv_len || !bmd->bi_size)
3506 ret = (int) bvec->bv_len;
3511 static void rbd_free_disk(struct rbd_device *rbd_dev)
3513 struct gendisk *disk = rbd_dev->disk;
3518 rbd_dev->disk = NULL;
3519 if (disk->flags & GENHD_FL_UP) {
3522 blk_cleanup_queue(disk->queue);
3523 blk_mq_free_tag_set(&rbd_dev->tag_set);
3528 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
3529 const char *object_name,
3530 u64 offset, u64 length, void *buf)
3533 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3534 struct rbd_obj_request *obj_request;
3535 struct page **pages = NULL;
3540 page_count = (u32) calc_pages_for(offset, length);
3541 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
3543 return PTR_ERR(pages);
3546 obj_request = rbd_obj_request_create(object_name, offset, length,
3551 obj_request->pages = pages;
3552 obj_request->page_count = page_count;
3554 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
3556 if (!obj_request->osd_req)
3559 osd_req_op_extent_init(obj_request->osd_req, 0, CEPH_OSD_OP_READ,
3560 offset, length, 0, 0);
3561 osd_req_op_extent_osd_data_pages(obj_request->osd_req, 0,
3563 obj_request->length,
3564 obj_request->offset & ~PAGE_MASK,
3566 rbd_osd_req_format_read(obj_request);
3568 ret = rbd_obj_request_submit(osdc, obj_request);
3571 ret = rbd_obj_request_wait(obj_request);
3575 ret = obj_request->result;
3579 rbd_assert(obj_request->xferred <= (u64) SIZE_MAX);
3580 size = (size_t) obj_request->xferred;
3581 ceph_copy_from_page_vector(pages, buf, 0, size);
3582 rbd_assert(size <= (size_t)INT_MAX);
3586 rbd_obj_request_put(obj_request);
3588 ceph_release_page_vector(pages, page_count);
3594 * Read the complete header for the given rbd device. On successful
3595 * return, the rbd_dev->header field will contain up-to-date
3596 * information about the image.
3598 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev)
3600 struct rbd_image_header_ondisk *ondisk = NULL;
3607 * The complete header will include an array of its 64-bit
3608 * snapshot ids, followed by the names of those snapshots as
3609 * a contiguous block of NUL-terminated strings. Note that
3610 * the number of snapshots could change by the time we read
3611 * it in, in which case we re-read it.
3618 size = sizeof (*ondisk);
3619 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
3621 ondisk = kmalloc(size, GFP_KERNEL);
3625 ret = rbd_obj_read_sync(rbd_dev, rbd_dev->header_name,
3629 if ((size_t)ret < size) {
3631 rbd_warn(rbd_dev, "short header read (want %zd got %d)",
3635 if (!rbd_dev_ondisk_valid(ondisk)) {
3637 rbd_warn(rbd_dev, "invalid header");
3641 names_size = le64_to_cpu(ondisk->snap_names_len);
3642 want_count = snap_count;
3643 snap_count = le32_to_cpu(ondisk->snap_count);
3644 } while (snap_count != want_count);
3646 ret = rbd_header_from_disk(rbd_dev, ondisk);
3654 * Clear the rbd device's EXISTS flag if the snapshot it's mapped to
3655 * has disappeared from the (just updated) snapshot context.
3657 static void rbd_exists_validate(struct rbd_device *rbd_dev)
3661 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags))
3664 snap_id = rbd_dev->spec->snap_id;
3665 if (snap_id == CEPH_NOSNAP)
3668 if (rbd_dev_snap_index(rbd_dev, snap_id) == BAD_SNAP_INDEX)
3669 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
3672 static void rbd_dev_update_size(struct rbd_device *rbd_dev)
3678 * Don't hold the lock while doing disk operations,
3679 * or lock ordering will conflict with the bdev mutex via:
3680 * rbd_add() -> blkdev_get() -> rbd_open()
3682 spin_lock_irq(&rbd_dev->lock);
3683 removing = test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags);
3684 spin_unlock_irq(&rbd_dev->lock);
3686 * If the device is being removed, rbd_dev->disk has
3687 * been destroyed, so don't try to update its size
3690 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
3691 dout("setting size to %llu sectors", (unsigned long long)size);
3692 set_capacity(rbd_dev->disk, size);
3693 revalidate_disk(rbd_dev->disk);
3697 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
3702 down_write(&rbd_dev->header_rwsem);
3703 mapping_size = rbd_dev->mapping.size;
3705 ret = rbd_dev_header_info(rbd_dev);
3710 * If there is a parent, see if it has disappeared due to the
3711 * mapped image getting flattened.
3713 if (rbd_dev->parent) {
3714 ret = rbd_dev_v2_parent_info(rbd_dev);
3719 if (rbd_dev->spec->snap_id == CEPH_NOSNAP) {
3720 rbd_dev->mapping.size = rbd_dev->header.image_size;
3722 /* validate mapped snapshot's EXISTS flag */
3723 rbd_exists_validate(rbd_dev);
3727 up_write(&rbd_dev->header_rwsem);
3728 if (!ret && mapping_size != rbd_dev->mapping.size)
3729 rbd_dev_update_size(rbd_dev);
3734 static int rbd_init_request(void *data, struct request *rq,
3735 unsigned int hctx_idx, unsigned int request_idx,
3736 unsigned int numa_node)
3738 struct work_struct *work = blk_mq_rq_to_pdu(rq);
3740 INIT_WORK(work, rbd_queue_workfn);
3744 static struct blk_mq_ops rbd_mq_ops = {
3745 .queue_rq = rbd_queue_rq,
3746 .map_queue = blk_mq_map_queue,
3747 .init_request = rbd_init_request,
3750 static int rbd_init_disk(struct rbd_device *rbd_dev)
3752 struct gendisk *disk;
3753 struct request_queue *q;
3757 /* create gendisk info */
3758 disk = alloc_disk(single_major ?
3759 (1 << RBD_SINGLE_MAJOR_PART_SHIFT) :
3760 RBD_MINORS_PER_MAJOR);
3764 snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
3766 disk->major = rbd_dev->major;
3767 disk->first_minor = rbd_dev->minor;
3769 disk->flags |= GENHD_FL_EXT_DEVT;
3770 disk->fops = &rbd_bd_ops;
3771 disk->private_data = rbd_dev;
3773 memset(&rbd_dev->tag_set, 0, sizeof(rbd_dev->tag_set));
3774 rbd_dev->tag_set.ops = &rbd_mq_ops;
3775 rbd_dev->tag_set.queue_depth = rbd_dev->opts->queue_depth;
3776 rbd_dev->tag_set.numa_node = NUMA_NO_NODE;
3777 rbd_dev->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
3778 rbd_dev->tag_set.nr_hw_queues = 1;
3779 rbd_dev->tag_set.cmd_size = sizeof(struct work_struct);
3781 err = blk_mq_alloc_tag_set(&rbd_dev->tag_set);
3785 q = blk_mq_init_queue(&rbd_dev->tag_set);
3791 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
3792 /* QUEUE_FLAG_ADD_RANDOM is off by default for blk-mq */
3794 /* set io sizes to object size */
3795 segment_size = rbd_obj_bytes(&rbd_dev->header);
3796 blk_queue_max_hw_sectors(q, segment_size / SECTOR_SIZE);
3797 blk_queue_max_segments(q, segment_size / SECTOR_SIZE);
3798 blk_queue_max_segment_size(q, segment_size);
3799 blk_queue_io_min(q, segment_size);
3800 blk_queue_io_opt(q, segment_size);
3802 /* enable the discard support */
3803 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
3804 q->limits.discard_granularity = segment_size;
3805 q->limits.discard_alignment = segment_size;
3806 q->limits.max_discard_sectors = segment_size / SECTOR_SIZE;
3807 q->limits.discard_zeroes_data = 1;
3809 blk_queue_merge_bvec(q, rbd_merge_bvec);
3812 q->queuedata = rbd_dev;
3814 rbd_dev->disk = disk;
3818 blk_mq_free_tag_set(&rbd_dev->tag_set);
3828 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
3830 return container_of(dev, struct rbd_device, dev);
3833 static ssize_t rbd_size_show(struct device *dev,
3834 struct device_attribute *attr, char *buf)
3836 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3838 return sprintf(buf, "%llu\n",
3839 (unsigned long long)rbd_dev->mapping.size);
3843 * Note this shows the features for whatever's mapped, which is not
3844 * necessarily the base image.
3846 static ssize_t rbd_features_show(struct device *dev,
3847 struct device_attribute *attr, char *buf)
3849 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3851 return sprintf(buf, "0x%016llx\n",
3852 (unsigned long long)rbd_dev->mapping.features);
3855 static ssize_t rbd_major_show(struct device *dev,
3856 struct device_attribute *attr, char *buf)
3858 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3861 return sprintf(buf, "%d\n", rbd_dev->major);
3863 return sprintf(buf, "(none)\n");
3866 static ssize_t rbd_minor_show(struct device *dev,
3867 struct device_attribute *attr, char *buf)
3869 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3871 return sprintf(buf, "%d\n", rbd_dev->minor);
3874 static ssize_t rbd_client_id_show(struct device *dev,
3875 struct device_attribute *attr, char *buf)
3877 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3879 return sprintf(buf, "client%lld\n",
3880 ceph_client_id(rbd_dev->rbd_client->client));
3883 static ssize_t rbd_pool_show(struct device *dev,
3884 struct device_attribute *attr, char *buf)
3886 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3888 return sprintf(buf, "%s\n", rbd_dev->spec->pool_name);
3891 static ssize_t rbd_pool_id_show(struct device *dev,
3892 struct device_attribute *attr, char *buf)
3894 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3896 return sprintf(buf, "%llu\n",
3897 (unsigned long long) rbd_dev->spec->pool_id);
3900 static ssize_t rbd_name_show(struct device *dev,
3901 struct device_attribute *attr, char *buf)
3903 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3905 if (rbd_dev->spec->image_name)
3906 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
3908 return sprintf(buf, "(unknown)\n");
3911 static ssize_t rbd_image_id_show(struct device *dev,
3912 struct device_attribute *attr, char *buf)
3914 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3916 return sprintf(buf, "%s\n", rbd_dev->spec->image_id);
3920 * Shows the name of the currently-mapped snapshot (or
3921 * RBD_SNAP_HEAD_NAME for the base image).
3923 static ssize_t rbd_snap_show(struct device *dev,
3924 struct device_attribute *attr,
3927 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3929 return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
3933 * For a v2 image, shows the chain of parent images, separated by empty
3934 * lines. For v1 images or if there is no parent, shows "(no parent
3937 static ssize_t rbd_parent_show(struct device *dev,
3938 struct device_attribute *attr,
3941 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3944 if (!rbd_dev->parent)
3945 return sprintf(buf, "(no parent image)\n");
3947 for ( ; rbd_dev->parent; rbd_dev = rbd_dev->parent) {
3948 struct rbd_spec *spec = rbd_dev->parent_spec;
3950 count += sprintf(&buf[count], "%s"
3951 "pool_id %llu\npool_name %s\n"
3952 "image_id %s\nimage_name %s\n"
3953 "snap_id %llu\nsnap_name %s\n"
3955 !count ? "" : "\n", /* first? */
3956 spec->pool_id, spec->pool_name,
3957 spec->image_id, spec->image_name ?: "(unknown)",
3958 spec->snap_id, spec->snap_name,
3959 rbd_dev->parent_overlap);
3965 static ssize_t rbd_image_refresh(struct device *dev,
3966 struct device_attribute *attr,
3970 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3973 ret = rbd_dev_refresh(rbd_dev);
3980 static DEVICE_ATTR(size, S_IRUGO, rbd_size_show, NULL);
3981 static DEVICE_ATTR(features, S_IRUGO, rbd_features_show, NULL);
3982 static DEVICE_ATTR(major, S_IRUGO, rbd_major_show, NULL);
3983 static DEVICE_ATTR(minor, S_IRUGO, rbd_minor_show, NULL);
3984 static DEVICE_ATTR(client_id, S_IRUGO, rbd_client_id_show, NULL);
3985 static DEVICE_ATTR(pool, S_IRUGO, rbd_pool_show, NULL);
3986 static DEVICE_ATTR(pool_id, S_IRUGO, rbd_pool_id_show, NULL);
3987 static DEVICE_ATTR(name, S_IRUGO, rbd_name_show, NULL);
3988 static DEVICE_ATTR(image_id, S_IRUGO, rbd_image_id_show, NULL);
3989 static DEVICE_ATTR(refresh, S_IWUSR, NULL, rbd_image_refresh);
3990 static DEVICE_ATTR(current_snap, S_IRUGO, rbd_snap_show, NULL);
3991 static DEVICE_ATTR(parent, S_IRUGO, rbd_parent_show, NULL);
3993 static struct attribute *rbd_attrs[] = {
3994 &dev_attr_size.attr,
3995 &dev_attr_features.attr,
3996 &dev_attr_major.attr,
3997 &dev_attr_minor.attr,
3998 &dev_attr_client_id.attr,
3999 &dev_attr_pool.attr,
4000 &dev_attr_pool_id.attr,
4001 &dev_attr_name.attr,
4002 &dev_attr_image_id.attr,
4003 &dev_attr_current_snap.attr,
4004 &dev_attr_parent.attr,
4005 &dev_attr_refresh.attr,
4009 static struct attribute_group rbd_attr_group = {
4013 static const struct attribute_group *rbd_attr_groups[] = {
4018 static void rbd_sysfs_dev_release(struct device *dev)
4022 static struct device_type rbd_device_type = {
4024 .groups = rbd_attr_groups,
4025 .release = rbd_sysfs_dev_release,
4028 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
4030 kref_get(&spec->kref);
4035 static void rbd_spec_free(struct kref *kref);
4036 static void rbd_spec_put(struct rbd_spec *spec)
4039 kref_put(&spec->kref, rbd_spec_free);
4042 static struct rbd_spec *rbd_spec_alloc(void)
4044 struct rbd_spec *spec;
4046 spec = kzalloc(sizeof (*spec), GFP_KERNEL);
4050 spec->pool_id = CEPH_NOPOOL;
4051 spec->snap_id = CEPH_NOSNAP;
4052 kref_init(&spec->kref);
4057 static void rbd_spec_free(struct kref *kref)
4059 struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
4061 kfree(spec->pool_name);
4062 kfree(spec->image_id);
4063 kfree(spec->image_name);
4064 kfree(spec->snap_name);
4068 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
4069 struct rbd_spec *spec,
4070 struct rbd_options *opts)
4072 struct rbd_device *rbd_dev;
4074 rbd_dev = kzalloc(sizeof (*rbd_dev), GFP_KERNEL);
4078 spin_lock_init(&rbd_dev->lock);
4080 atomic_set(&rbd_dev->parent_ref, 0);
4081 INIT_LIST_HEAD(&rbd_dev->node);
4082 init_rwsem(&rbd_dev->header_rwsem);
4084 rbd_dev->rbd_client = rbdc;
4085 rbd_dev->spec = spec;
4086 rbd_dev->opts = opts;
4088 /* Initialize the layout used for all rbd requests */
4090 rbd_dev->layout.fl_stripe_unit = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
4091 rbd_dev->layout.fl_stripe_count = cpu_to_le32(1);
4092 rbd_dev->layout.fl_object_size = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
4093 rbd_dev->layout.fl_pg_pool = cpu_to_le32((u32) spec->pool_id);
4098 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
4100 rbd_put_client(rbd_dev->rbd_client);
4101 rbd_spec_put(rbd_dev->spec);
4102 kfree(rbd_dev->opts);
4107 * Get the size and object order for an image snapshot, or if
4108 * snap_id is CEPH_NOSNAP, gets this information for the base
4111 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
4112 u8 *order, u64 *snap_size)
4114 __le64 snapid = cpu_to_le64(snap_id);
4119 } __attribute__ ((packed)) size_buf = { 0 };
4121 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4123 &snapid, sizeof (snapid),
4124 &size_buf, sizeof (size_buf));
4125 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4128 if (ret < sizeof (size_buf))
4132 *order = size_buf.order;
4133 dout(" order %u", (unsigned int)*order);
4135 *snap_size = le64_to_cpu(size_buf.size);
4137 dout(" snap_id 0x%016llx snap_size = %llu\n",
4138 (unsigned long long)snap_id,
4139 (unsigned long long)*snap_size);
4144 static int rbd_dev_v2_image_size(struct rbd_device *rbd_dev)
4146 return _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
4147 &rbd_dev->header.obj_order,
4148 &rbd_dev->header.image_size);
4151 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev)
4157 reply_buf = kzalloc(RBD_OBJ_PREFIX_LEN_MAX, GFP_KERNEL);
4161 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4162 "rbd", "get_object_prefix", NULL, 0,
4163 reply_buf, RBD_OBJ_PREFIX_LEN_MAX);
4164 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4169 rbd_dev->header.object_prefix = ceph_extract_encoded_string(&p,
4170 p + ret, NULL, GFP_NOIO);
4173 if (IS_ERR(rbd_dev->header.object_prefix)) {
4174 ret = PTR_ERR(rbd_dev->header.object_prefix);
4175 rbd_dev->header.object_prefix = NULL;
4177 dout(" object_prefix = %s\n", rbd_dev->header.object_prefix);
4185 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
4188 __le64 snapid = cpu_to_le64(snap_id);
4192 } __attribute__ ((packed)) features_buf = { 0 };
4196 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4197 "rbd", "get_features",
4198 &snapid, sizeof (snapid),
4199 &features_buf, sizeof (features_buf));
4200 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4203 if (ret < sizeof (features_buf))
4206 incompat = le64_to_cpu(features_buf.incompat);
4207 if (incompat & ~RBD_FEATURES_SUPPORTED)
4210 *snap_features = le64_to_cpu(features_buf.features);
4212 dout(" snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
4213 (unsigned long long)snap_id,
4214 (unsigned long long)*snap_features,
4215 (unsigned long long)le64_to_cpu(features_buf.incompat));
4220 static int rbd_dev_v2_features(struct rbd_device *rbd_dev)
4222 return _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
4223 &rbd_dev->header.features);
4226 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev)
4228 struct rbd_spec *parent_spec;
4230 void *reply_buf = NULL;
4240 parent_spec = rbd_spec_alloc();
4244 size = sizeof (__le64) + /* pool_id */
4245 sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX + /* image_id */
4246 sizeof (__le64) + /* snap_id */
4247 sizeof (__le64); /* overlap */
4248 reply_buf = kmalloc(size, GFP_KERNEL);
4254 snapid = cpu_to_le64(rbd_dev->spec->snap_id);
4255 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4256 "rbd", "get_parent",
4257 &snapid, sizeof (snapid),
4259 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4264 end = reply_buf + ret;
4266 ceph_decode_64_safe(&p, end, pool_id, out_err);
4267 if (pool_id == CEPH_NOPOOL) {
4269 * Either the parent never existed, or we have
4270 * record of it but the image got flattened so it no
4271 * longer has a parent. When the parent of a
4272 * layered image disappears we immediately set the
4273 * overlap to 0. The effect of this is that all new
4274 * requests will be treated as if the image had no
4277 if (rbd_dev->parent_overlap) {
4278 rbd_dev->parent_overlap = 0;
4279 rbd_dev_parent_put(rbd_dev);
4280 pr_info("%s: clone image has been flattened\n",
4281 rbd_dev->disk->disk_name);
4284 goto out; /* No parent? No problem. */
4287 /* The ceph file layout needs to fit pool id in 32 bits */
4290 if (pool_id > (u64)U32_MAX) {
4291 rbd_warn(NULL, "parent pool id too large (%llu > %u)",
4292 (unsigned long long)pool_id, U32_MAX);
4296 image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
4297 if (IS_ERR(image_id)) {
4298 ret = PTR_ERR(image_id);
4301 ceph_decode_64_safe(&p, end, snap_id, out_err);
4302 ceph_decode_64_safe(&p, end, overlap, out_err);
4305 * The parent won't change (except when the clone is
4306 * flattened, already handled that). So we only need to
4307 * record the parent spec we have not already done so.
4309 if (!rbd_dev->parent_spec) {
4310 parent_spec->pool_id = pool_id;
4311 parent_spec->image_id = image_id;
4312 parent_spec->snap_id = snap_id;
4313 rbd_dev->parent_spec = parent_spec;
4314 parent_spec = NULL; /* rbd_dev now owns this */
4320 * We always update the parent overlap. If it's zero we issue
4321 * a warning, as we will proceed as if there was no parent.
4325 /* refresh, careful to warn just once */
4326 if (rbd_dev->parent_overlap)
4328 "clone now standalone (overlap became 0)");
4331 rbd_warn(rbd_dev, "clone is standalone (overlap 0)");
4334 rbd_dev->parent_overlap = overlap;
4340 rbd_spec_put(parent_spec);
4345 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev)
4349 __le64 stripe_count;
4350 } __attribute__ ((packed)) striping_info_buf = { 0 };
4351 size_t size = sizeof (striping_info_buf);
4358 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4359 "rbd", "get_stripe_unit_count", NULL, 0,
4360 (char *)&striping_info_buf, size);
4361 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4368 * We don't actually support the "fancy striping" feature
4369 * (STRIPINGV2) yet, but if the striping sizes are the
4370 * defaults the behavior is the same as before. So find
4371 * out, and only fail if the image has non-default values.
4374 obj_size = (u64)1 << rbd_dev->header.obj_order;
4375 p = &striping_info_buf;
4376 stripe_unit = ceph_decode_64(&p);
4377 if (stripe_unit != obj_size) {
4378 rbd_warn(rbd_dev, "unsupported stripe unit "
4379 "(got %llu want %llu)",
4380 stripe_unit, obj_size);
4383 stripe_count = ceph_decode_64(&p);
4384 if (stripe_count != 1) {
4385 rbd_warn(rbd_dev, "unsupported stripe count "
4386 "(got %llu want 1)", stripe_count);
4389 rbd_dev->header.stripe_unit = stripe_unit;
4390 rbd_dev->header.stripe_count = stripe_count;
4395 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
4397 size_t image_id_size;
4402 void *reply_buf = NULL;
4404 char *image_name = NULL;
4407 rbd_assert(!rbd_dev->spec->image_name);
4409 len = strlen(rbd_dev->spec->image_id);
4410 image_id_size = sizeof (__le32) + len;
4411 image_id = kmalloc(image_id_size, GFP_KERNEL);
4416 end = image_id + image_id_size;
4417 ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
4419 size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
4420 reply_buf = kmalloc(size, GFP_KERNEL);
4424 ret = rbd_obj_method_sync(rbd_dev, RBD_DIRECTORY,
4425 "rbd", "dir_get_name",
4426 image_id, image_id_size,
4431 end = reply_buf + ret;
4433 image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
4434 if (IS_ERR(image_name))
4437 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
4445 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4447 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4448 const char *snap_name;
4451 /* Skip over names until we find the one we are looking for */
4453 snap_name = rbd_dev->header.snap_names;
4454 while (which < snapc->num_snaps) {
4455 if (!strcmp(name, snap_name))
4456 return snapc->snaps[which];
4457 snap_name += strlen(snap_name) + 1;
4463 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4465 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4470 for (which = 0; !found && which < snapc->num_snaps; which++) {
4471 const char *snap_name;
4473 snap_id = snapc->snaps[which];
4474 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
4475 if (IS_ERR(snap_name)) {
4476 /* ignore no-longer existing snapshots */
4477 if (PTR_ERR(snap_name) == -ENOENT)
4482 found = !strcmp(name, snap_name);
4485 return found ? snap_id : CEPH_NOSNAP;
4489 * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
4490 * no snapshot by that name is found, or if an error occurs.
4492 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4494 if (rbd_dev->image_format == 1)
4495 return rbd_v1_snap_id_by_name(rbd_dev, name);
4497 return rbd_v2_snap_id_by_name(rbd_dev, name);
4501 * An image being mapped will have everything but the snap id.
4503 static int rbd_spec_fill_snap_id(struct rbd_device *rbd_dev)
4505 struct rbd_spec *spec = rbd_dev->spec;
4507 rbd_assert(spec->pool_id != CEPH_NOPOOL && spec->pool_name);
4508 rbd_assert(spec->image_id && spec->image_name);
4509 rbd_assert(spec->snap_name);
4511 if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
4514 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
4515 if (snap_id == CEPH_NOSNAP)
4518 spec->snap_id = snap_id;
4520 spec->snap_id = CEPH_NOSNAP;
4527 * A parent image will have all ids but none of the names.
4529 * All names in an rbd spec are dynamically allocated. It's OK if we
4530 * can't figure out the name for an image id.
4532 static int rbd_spec_fill_names(struct rbd_device *rbd_dev)
4534 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4535 struct rbd_spec *spec = rbd_dev->spec;
4536 const char *pool_name;
4537 const char *image_name;
4538 const char *snap_name;
4541 rbd_assert(spec->pool_id != CEPH_NOPOOL);
4542 rbd_assert(spec->image_id);
4543 rbd_assert(spec->snap_id != CEPH_NOSNAP);
4545 /* Get the pool name; we have to make our own copy of this */
4547 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
4549 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
4552 pool_name = kstrdup(pool_name, GFP_KERNEL);
4556 /* Fetch the image name; tolerate failure here */
4558 image_name = rbd_dev_image_name(rbd_dev);
4560 rbd_warn(rbd_dev, "unable to get image name");
4562 /* Fetch the snapshot name */
4564 snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
4565 if (IS_ERR(snap_name)) {
4566 ret = PTR_ERR(snap_name);
4570 spec->pool_name = pool_name;
4571 spec->image_name = image_name;
4572 spec->snap_name = snap_name;
4582 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev)
4591 struct ceph_snap_context *snapc;
4595 * We'll need room for the seq value (maximum snapshot id),
4596 * snapshot count, and array of that many snapshot ids.
4597 * For now we have a fixed upper limit on the number we're
4598 * prepared to receive.
4600 size = sizeof (__le64) + sizeof (__le32) +
4601 RBD_MAX_SNAP_COUNT * sizeof (__le64);
4602 reply_buf = kzalloc(size, GFP_KERNEL);
4606 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4607 "rbd", "get_snapcontext", NULL, 0,
4609 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4614 end = reply_buf + ret;
4616 ceph_decode_64_safe(&p, end, seq, out);
4617 ceph_decode_32_safe(&p, end, snap_count, out);
4620 * Make sure the reported number of snapshot ids wouldn't go
4621 * beyond the end of our buffer. But before checking that,
4622 * make sure the computed size of the snapshot context we
4623 * allocate is representable in a size_t.
4625 if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
4630 if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
4634 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
4640 for (i = 0; i < snap_count; i++)
4641 snapc->snaps[i] = ceph_decode_64(&p);
4643 ceph_put_snap_context(rbd_dev->header.snapc);
4644 rbd_dev->header.snapc = snapc;
4646 dout(" snap context seq = %llu, snap_count = %u\n",
4647 (unsigned long long)seq, (unsigned int)snap_count);
4654 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
4665 size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
4666 reply_buf = kmalloc(size, GFP_KERNEL);
4668 return ERR_PTR(-ENOMEM);
4670 snapid = cpu_to_le64(snap_id);
4671 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4672 "rbd", "get_snapshot_name",
4673 &snapid, sizeof (snapid),
4675 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4677 snap_name = ERR_PTR(ret);
4682 end = reply_buf + ret;
4683 snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
4684 if (IS_ERR(snap_name))
4687 dout(" snap_id 0x%016llx snap_name = %s\n",
4688 (unsigned long long)snap_id, snap_name);
4695 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev)
4697 bool first_time = rbd_dev->header.object_prefix == NULL;
4700 ret = rbd_dev_v2_image_size(rbd_dev);
4705 ret = rbd_dev_v2_header_onetime(rbd_dev);
4710 ret = rbd_dev_v2_snap_context(rbd_dev);
4711 dout("rbd_dev_v2_snap_context returned %d\n", ret);
4716 static int rbd_dev_header_info(struct rbd_device *rbd_dev)
4718 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
4720 if (rbd_dev->image_format == 1)
4721 return rbd_dev_v1_header_info(rbd_dev);
4723 return rbd_dev_v2_header_info(rbd_dev);
4726 static int rbd_bus_add_dev(struct rbd_device *rbd_dev)
4731 dev = &rbd_dev->dev;
4732 dev->bus = &rbd_bus_type;
4733 dev->type = &rbd_device_type;
4734 dev->parent = &rbd_root_dev;
4735 dev->release = rbd_dev_device_release;
4736 dev_set_name(dev, "%d", rbd_dev->dev_id);
4737 ret = device_register(dev);
4742 static void rbd_bus_del_dev(struct rbd_device *rbd_dev)
4744 device_unregister(&rbd_dev->dev);
4748 * Get a unique rbd identifier for the given new rbd_dev, and add
4749 * the rbd_dev to the global list.
4751 static int rbd_dev_id_get(struct rbd_device *rbd_dev)
4755 new_dev_id = ida_simple_get(&rbd_dev_id_ida,
4756 0, minor_to_rbd_dev_id(1 << MINORBITS),
4761 rbd_dev->dev_id = new_dev_id;
4763 spin_lock(&rbd_dev_list_lock);
4764 list_add_tail(&rbd_dev->node, &rbd_dev_list);
4765 spin_unlock(&rbd_dev_list_lock);
4767 dout("rbd_dev %p given dev id %d\n", rbd_dev, rbd_dev->dev_id);
4773 * Remove an rbd_dev from the global list, and record that its
4774 * identifier is no longer in use.
4776 static void rbd_dev_id_put(struct rbd_device *rbd_dev)
4778 spin_lock(&rbd_dev_list_lock);
4779 list_del_init(&rbd_dev->node);
4780 spin_unlock(&rbd_dev_list_lock);
4782 ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4784 dout("rbd_dev %p released dev id %d\n", rbd_dev, rbd_dev->dev_id);
4788 * Skips over white space at *buf, and updates *buf to point to the
4789 * first found non-space character (if any). Returns the length of
4790 * the token (string of non-white space characters) found. Note
4791 * that *buf must be terminated with '\0'.
4793 static inline size_t next_token(const char **buf)
4796 * These are the characters that produce nonzero for
4797 * isspace() in the "C" and "POSIX" locales.
4799 const char *spaces = " \f\n\r\t\v";
4801 *buf += strspn(*buf, spaces); /* Find start of token */
4803 return strcspn(*buf, spaces); /* Return token length */
4807 * Finds the next token in *buf, dynamically allocates a buffer big
4808 * enough to hold a copy of it, and copies the token into the new
4809 * buffer. The copy is guaranteed to be terminated with '\0'. Note
4810 * that a duplicate buffer is created even for a zero-length token.
4812 * Returns a pointer to the newly-allocated duplicate, or a null
4813 * pointer if memory for the duplicate was not available. If
4814 * the lenp argument is a non-null pointer, the length of the token
4815 * (not including the '\0') is returned in *lenp.
4817 * If successful, the *buf pointer will be updated to point beyond
4818 * the end of the found token.
4820 * Note: uses GFP_KERNEL for allocation.
4822 static inline char *dup_token(const char **buf, size_t *lenp)
4827 len = next_token(buf);
4828 dup = kmemdup(*buf, len + 1, GFP_KERNEL);
4831 *(dup + len) = '\0';
4841 * Parse the options provided for an "rbd add" (i.e., rbd image
4842 * mapping) request. These arrive via a write to /sys/bus/rbd/add,
4843 * and the data written is passed here via a NUL-terminated buffer.
4844 * Returns 0 if successful or an error code otherwise.
4846 * The information extracted from these options is recorded in
4847 * the other parameters which return dynamically-allocated
4850 * The address of a pointer that will refer to a ceph options
4851 * structure. Caller must release the returned pointer using
4852 * ceph_destroy_options() when it is no longer needed.
4854 * Address of an rbd options pointer. Fully initialized by
4855 * this function; caller must release with kfree().
4857 * Address of an rbd image specification pointer. Fully
4858 * initialized by this function based on parsed options.
4859 * Caller must release with rbd_spec_put().
4861 * The options passed take this form:
4862 * <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
4865 * A comma-separated list of one or more monitor addresses.
4866 * A monitor address is an ip address, optionally followed
4867 * by a port number (separated by a colon).
4868 * I.e.: ip1[:port1][,ip2[:port2]...]
4870 * A comma-separated list of ceph and/or rbd options.
4872 * The name of the rados pool containing the rbd image.
4874 * The name of the image in that pool to map.
4876 * An optional snapshot id. If provided, the mapping will
4877 * present data from the image at the time that snapshot was
4878 * created. The image head is used if no snapshot id is
4879 * provided. Snapshot mappings are always read-only.
4881 static int rbd_add_parse_args(const char *buf,
4882 struct ceph_options **ceph_opts,
4883 struct rbd_options **opts,
4884 struct rbd_spec **rbd_spec)
4888 const char *mon_addrs;
4890 size_t mon_addrs_size;
4891 struct rbd_spec *spec = NULL;
4892 struct rbd_options *rbd_opts = NULL;
4893 struct ceph_options *copts;
4896 /* The first four tokens are required */
4898 len = next_token(&buf);
4900 rbd_warn(NULL, "no monitor address(es) provided");
4904 mon_addrs_size = len + 1;
4908 options = dup_token(&buf, NULL);
4912 rbd_warn(NULL, "no options provided");
4916 spec = rbd_spec_alloc();
4920 spec->pool_name = dup_token(&buf, NULL);
4921 if (!spec->pool_name)
4923 if (!*spec->pool_name) {
4924 rbd_warn(NULL, "no pool name provided");
4928 spec->image_name = dup_token(&buf, NULL);
4929 if (!spec->image_name)
4931 if (!*spec->image_name) {
4932 rbd_warn(NULL, "no image name provided");
4937 * Snapshot name is optional; default is to use "-"
4938 * (indicating the head/no snapshot).
4940 len = next_token(&buf);
4942 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
4943 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
4944 } else if (len > RBD_MAX_SNAP_NAME_LEN) {
4945 ret = -ENAMETOOLONG;
4948 snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
4951 *(snap_name + len) = '\0';
4952 spec->snap_name = snap_name;
4954 /* Initialize all rbd options to the defaults */
4956 rbd_opts = kzalloc(sizeof (*rbd_opts), GFP_KERNEL);
4960 rbd_opts->read_only = RBD_READ_ONLY_DEFAULT;
4961 rbd_opts->queue_depth = RBD_QUEUE_DEPTH_DEFAULT;
4963 copts = ceph_parse_options(options, mon_addrs,
4964 mon_addrs + mon_addrs_size - 1,
4965 parse_rbd_opts_token, rbd_opts);
4966 if (IS_ERR(copts)) {
4967 ret = PTR_ERR(copts);
4988 * Return pool id (>= 0) or a negative error code.
4990 static int rbd_add_get_pool_id(struct rbd_client *rbdc, const char *pool_name)
4992 struct ceph_options *opts = rbdc->client->options;
4998 ret = ceph_pg_poolid_by_name(rbdc->client->osdc.osdmap, pool_name);
4999 if (ret == -ENOENT && tries++ < 1) {
5000 ret = ceph_monc_do_get_version(&rbdc->client->monc, "osdmap",
5005 if (rbdc->client->osdc.osdmap->epoch < newest_epoch) {
5006 ceph_monc_request_next_osdmap(&rbdc->client->monc);
5007 (void) ceph_monc_wait_osdmap(&rbdc->client->monc,
5009 opts->mount_timeout);
5012 /* the osdmap we have is new enough */
5021 * An rbd format 2 image has a unique identifier, distinct from the
5022 * name given to it by the user. Internally, that identifier is
5023 * what's used to specify the names of objects related to the image.
5025 * A special "rbd id" object is used to map an rbd image name to its
5026 * id. If that object doesn't exist, then there is no v2 rbd image
5027 * with the supplied name.
5029 * This function will record the given rbd_dev's image_id field if
5030 * it can be determined, and in that case will return 0. If any
5031 * errors occur a negative errno will be returned and the rbd_dev's
5032 * image_id field will be unchanged (and should be NULL).
5034 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
5043 * When probing a parent image, the image id is already
5044 * known (and the image name likely is not). There's no
5045 * need to fetch the image id again in this case. We
5046 * do still need to set the image format though.
5048 if (rbd_dev->spec->image_id) {
5049 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
5055 * First, see if the format 2 image id file exists, and if
5056 * so, get the image's persistent id from it.
5058 size = sizeof (RBD_ID_PREFIX) + strlen(rbd_dev->spec->image_name);
5059 object_name = kmalloc(size, GFP_NOIO);
5062 sprintf(object_name, "%s%s", RBD_ID_PREFIX, rbd_dev->spec->image_name);
5063 dout("rbd id object name is %s\n", object_name);
5065 /* Response will be an encoded string, which includes a length */
5067 size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
5068 response = kzalloc(size, GFP_NOIO);
5074 /* If it doesn't exist we'll assume it's a format 1 image */
5076 ret = rbd_obj_method_sync(rbd_dev, object_name,
5077 "rbd", "get_id", NULL, 0,
5078 response, RBD_IMAGE_ID_LEN_MAX);
5079 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5080 if (ret == -ENOENT) {
5081 image_id = kstrdup("", GFP_KERNEL);
5082 ret = image_id ? 0 : -ENOMEM;
5084 rbd_dev->image_format = 1;
5085 } else if (ret >= 0) {
5088 image_id = ceph_extract_encoded_string(&p, p + ret,
5090 ret = PTR_ERR_OR_ZERO(image_id);
5092 rbd_dev->image_format = 2;
5096 rbd_dev->spec->image_id = image_id;
5097 dout("image_id is %s\n", image_id);
5107 * Undo whatever state changes are made by v1 or v2 header info
5110 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
5112 struct rbd_image_header *header;
5114 rbd_dev_parent_put(rbd_dev);
5116 /* Free dynamic fields from the header, then zero it out */
5118 header = &rbd_dev->header;
5119 ceph_put_snap_context(header->snapc);
5120 kfree(header->snap_sizes);
5121 kfree(header->snap_names);
5122 kfree(header->object_prefix);
5123 memset(header, 0, sizeof (*header));
5126 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev)
5130 ret = rbd_dev_v2_object_prefix(rbd_dev);
5135 * Get the and check features for the image. Currently the
5136 * features are assumed to never change.
5138 ret = rbd_dev_v2_features(rbd_dev);
5142 /* If the image supports fancy striping, get its parameters */
5144 if (rbd_dev->header.features & RBD_FEATURE_STRIPINGV2) {
5145 ret = rbd_dev_v2_striping_info(rbd_dev);
5149 /* No support for crypto and compression type format 2 images */
5153 rbd_dev->header.features = 0;
5154 kfree(rbd_dev->header.object_prefix);
5155 rbd_dev->header.object_prefix = NULL;
5160 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev)
5162 struct rbd_device *parent = NULL;
5163 struct rbd_spec *parent_spec;
5164 struct rbd_client *rbdc;
5167 if (!rbd_dev->parent_spec)
5170 * We need to pass a reference to the client and the parent
5171 * spec when creating the parent rbd_dev. Images related by
5172 * parent/child relationships always share both.
5174 parent_spec = rbd_spec_get(rbd_dev->parent_spec);
5175 rbdc = __rbd_get_client(rbd_dev->rbd_client);
5178 parent = rbd_dev_create(rbdc, parent_spec, NULL);
5182 ret = rbd_dev_image_probe(parent, false);
5185 rbd_dev->parent = parent;
5186 atomic_set(&rbd_dev->parent_ref, 1);
5191 rbd_dev_unparent(rbd_dev);
5192 kfree(rbd_dev->header_name);
5193 rbd_dev_destroy(parent);
5195 rbd_put_client(rbdc);
5196 rbd_spec_put(parent_spec);
5202 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
5206 /* Get an id and fill in device name. */
5208 ret = rbd_dev_id_get(rbd_dev);
5212 BUILD_BUG_ON(DEV_NAME_LEN
5213 < sizeof (RBD_DRV_NAME) + MAX_INT_FORMAT_WIDTH);
5214 sprintf(rbd_dev->name, "%s%d", RBD_DRV_NAME, rbd_dev->dev_id);
5216 /* Record our major and minor device numbers. */
5218 if (!single_major) {
5219 ret = register_blkdev(0, rbd_dev->name);
5223 rbd_dev->major = ret;
5226 rbd_dev->major = rbd_major;
5227 rbd_dev->minor = rbd_dev_id_to_minor(rbd_dev->dev_id);
5230 /* Set up the blkdev mapping. */
5232 ret = rbd_init_disk(rbd_dev);
5234 goto err_out_blkdev;
5236 ret = rbd_dev_mapping_set(rbd_dev);
5240 set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
5241 set_disk_ro(rbd_dev->disk, rbd_dev->mapping.read_only);
5243 ret = rbd_bus_add_dev(rbd_dev);
5245 goto err_out_mapping;
5247 /* Everything's ready. Announce the disk to the world. */
5249 set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5250 add_disk(rbd_dev->disk);
5252 pr_info("%s: added with size 0x%llx\n", rbd_dev->disk->disk_name,
5253 (unsigned long long) rbd_dev->mapping.size);
5258 rbd_dev_mapping_clear(rbd_dev);
5260 rbd_free_disk(rbd_dev);
5263 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5265 rbd_dev_id_put(rbd_dev);
5266 rbd_dev_mapping_clear(rbd_dev);
5271 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
5273 struct rbd_spec *spec = rbd_dev->spec;
5276 /* Record the header object name for this rbd image. */
5278 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
5280 if (rbd_dev->image_format == 1)
5281 size = strlen(spec->image_name) + sizeof (RBD_SUFFIX);
5283 size = sizeof (RBD_HEADER_PREFIX) + strlen(spec->image_id);
5285 rbd_dev->header_name = kmalloc(size, GFP_KERNEL);
5286 if (!rbd_dev->header_name)
5289 if (rbd_dev->image_format == 1)
5290 sprintf(rbd_dev->header_name, "%s%s",
5291 spec->image_name, RBD_SUFFIX);
5293 sprintf(rbd_dev->header_name, "%s%s",
5294 RBD_HEADER_PREFIX, spec->image_id);
5298 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
5300 rbd_dev_unprobe(rbd_dev);
5301 kfree(rbd_dev->header_name);
5302 rbd_dev->header_name = NULL;
5303 rbd_dev->image_format = 0;
5304 kfree(rbd_dev->spec->image_id);
5305 rbd_dev->spec->image_id = NULL;
5307 rbd_dev_destroy(rbd_dev);
5311 * Probe for the existence of the header object for the given rbd
5312 * device. If this image is the one being mapped (i.e., not a
5313 * parent), initiate a watch on its header object before using that
5314 * object to get detailed information about the rbd image.
5316 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, bool mapping)
5321 * Get the id from the image id object. Unless there's an
5322 * error, rbd_dev->spec->image_id will be filled in with
5323 * a dynamically-allocated string, and rbd_dev->image_format
5324 * will be set to either 1 or 2.
5326 ret = rbd_dev_image_id(rbd_dev);
5330 ret = rbd_dev_header_name(rbd_dev);
5332 goto err_out_format;
5335 ret = rbd_dev_header_watch_sync(rbd_dev);
5338 pr_info("image %s/%s does not exist\n",
5339 rbd_dev->spec->pool_name,
5340 rbd_dev->spec->image_name);
5341 goto out_header_name;
5345 ret = rbd_dev_header_info(rbd_dev);
5350 * If this image is the one being mapped, we have pool name and
5351 * id, image name and id, and snap name - need to fill snap id.
5352 * Otherwise this is a parent image, identified by pool, image
5353 * and snap ids - need to fill in names for those ids.
5356 ret = rbd_spec_fill_snap_id(rbd_dev);
5358 ret = rbd_spec_fill_names(rbd_dev);
5361 pr_info("snap %s/%s@%s does not exist\n",
5362 rbd_dev->spec->pool_name,
5363 rbd_dev->spec->image_name,
5364 rbd_dev->spec->snap_name);
5368 if (rbd_dev->header.features & RBD_FEATURE_LAYERING) {
5369 ret = rbd_dev_v2_parent_info(rbd_dev);
5374 * Need to warn users if this image is the one being
5375 * mapped and has a parent.
5377 if (mapping && rbd_dev->parent_spec)
5379 "WARNING: kernel layering is EXPERIMENTAL!");
5382 ret = rbd_dev_probe_parent(rbd_dev);
5386 dout("discovered format %u image, header name is %s\n",
5387 rbd_dev->image_format, rbd_dev->header_name);
5391 rbd_dev_unprobe(rbd_dev);
5394 rbd_dev_header_unwatch_sync(rbd_dev);
5396 kfree(rbd_dev->header_name);
5397 rbd_dev->header_name = NULL;
5399 rbd_dev->image_format = 0;
5400 kfree(rbd_dev->spec->image_id);
5401 rbd_dev->spec->image_id = NULL;
5405 static ssize_t do_rbd_add(struct bus_type *bus,
5409 struct rbd_device *rbd_dev = NULL;
5410 struct ceph_options *ceph_opts = NULL;
5411 struct rbd_options *rbd_opts = NULL;
5412 struct rbd_spec *spec = NULL;
5413 struct rbd_client *rbdc;
5417 if (!try_module_get(THIS_MODULE))
5420 /* parse add command */
5421 rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
5423 goto err_out_module;
5425 rbdc = rbd_get_client(ceph_opts);
5432 rc = rbd_add_get_pool_id(rbdc, spec->pool_name);
5435 pr_info("pool %s does not exist\n", spec->pool_name);
5436 goto err_out_client;
5438 spec->pool_id = (u64)rc;
5440 /* The ceph file layout needs to fit pool id in 32 bits */
5442 if (spec->pool_id > (u64)U32_MAX) {
5443 rbd_warn(NULL, "pool id too large (%llu > %u)",
5444 (unsigned long long)spec->pool_id, U32_MAX);
5446 goto err_out_client;
5449 rbd_dev = rbd_dev_create(rbdc, spec, rbd_opts);
5451 goto err_out_client;
5452 rbdc = NULL; /* rbd_dev now owns this */
5453 spec = NULL; /* rbd_dev now owns this */
5454 rbd_opts = NULL; /* rbd_dev now owns this */
5456 rc = rbd_dev_image_probe(rbd_dev, true);
5458 goto err_out_rbd_dev;
5460 /* If we are mapping a snapshot it must be marked read-only */
5462 read_only = rbd_dev->opts->read_only;
5463 if (rbd_dev->spec->snap_id != CEPH_NOSNAP)
5465 rbd_dev->mapping.read_only = read_only;
5467 rc = rbd_dev_device_setup(rbd_dev);
5470 * rbd_dev_header_unwatch_sync() can't be moved into
5471 * rbd_dev_image_release() without refactoring, see
5472 * commit 1f3ef78861ac.
5474 rbd_dev_header_unwatch_sync(rbd_dev);
5475 rbd_dev_image_release(rbd_dev);
5476 goto err_out_module;
5482 rbd_dev_destroy(rbd_dev);
5484 rbd_put_client(rbdc);
5489 module_put(THIS_MODULE);
5491 dout("Error adding device %s\n", buf);
5496 static ssize_t rbd_add(struct bus_type *bus,
5503 return do_rbd_add(bus, buf, count);
5506 static ssize_t rbd_add_single_major(struct bus_type *bus,
5510 return do_rbd_add(bus, buf, count);
5513 static void rbd_dev_device_release(struct device *dev)
5515 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5517 rbd_free_disk(rbd_dev);
5518 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5519 rbd_dev_mapping_clear(rbd_dev);
5521 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5522 rbd_dev_id_put(rbd_dev);
5523 rbd_dev_mapping_clear(rbd_dev);
5526 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
5528 while (rbd_dev->parent) {
5529 struct rbd_device *first = rbd_dev;
5530 struct rbd_device *second = first->parent;
5531 struct rbd_device *third;
5534 * Follow to the parent with no grandparent and
5537 while (second && (third = second->parent)) {
5542 rbd_dev_image_release(second);
5543 first->parent = NULL;
5544 first->parent_overlap = 0;
5546 rbd_assert(first->parent_spec);
5547 rbd_spec_put(first->parent_spec);
5548 first->parent_spec = NULL;
5552 static ssize_t do_rbd_remove(struct bus_type *bus,
5556 struct rbd_device *rbd_dev = NULL;
5557 struct list_head *tmp;
5560 bool already = false;
5563 ret = kstrtoul(buf, 10, &ul);
5567 /* convert to int; abort if we lost anything in the conversion */
5573 spin_lock(&rbd_dev_list_lock);
5574 list_for_each(tmp, &rbd_dev_list) {
5575 rbd_dev = list_entry(tmp, struct rbd_device, node);
5576 if (rbd_dev->dev_id == dev_id) {
5582 spin_lock_irq(&rbd_dev->lock);
5583 if (rbd_dev->open_count)
5586 already = test_and_set_bit(RBD_DEV_FLAG_REMOVING,
5588 spin_unlock_irq(&rbd_dev->lock);
5590 spin_unlock(&rbd_dev_list_lock);
5591 if (ret < 0 || already)
5594 rbd_dev_header_unwatch_sync(rbd_dev);
5596 * flush remaining watch callbacks - these must be complete
5597 * before the osd_client is shutdown
5599 dout("%s: flushing notifies", __func__);
5600 ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
5603 * Don't free anything from rbd_dev->disk until after all
5604 * notifies are completely processed. Otherwise
5605 * rbd_bus_del_dev() will race with rbd_watch_cb(), resulting
5606 * in a potential use after free of rbd_dev->disk or rbd_dev.
5608 rbd_bus_del_dev(rbd_dev);
5609 rbd_dev_image_release(rbd_dev);
5610 module_put(THIS_MODULE);
5615 static ssize_t rbd_remove(struct bus_type *bus,
5622 return do_rbd_remove(bus, buf, count);
5625 static ssize_t rbd_remove_single_major(struct bus_type *bus,
5629 return do_rbd_remove(bus, buf, count);
5633 * create control files in sysfs
5636 static int rbd_sysfs_init(void)
5640 ret = device_register(&rbd_root_dev);
5644 ret = bus_register(&rbd_bus_type);
5646 device_unregister(&rbd_root_dev);
5651 static void rbd_sysfs_cleanup(void)
5653 bus_unregister(&rbd_bus_type);
5654 device_unregister(&rbd_root_dev);
5657 static int rbd_slab_init(void)
5659 rbd_assert(!rbd_img_request_cache);
5660 rbd_img_request_cache = kmem_cache_create("rbd_img_request",
5661 sizeof (struct rbd_img_request),
5662 __alignof__(struct rbd_img_request),
5664 if (!rbd_img_request_cache)
5667 rbd_assert(!rbd_obj_request_cache);
5668 rbd_obj_request_cache = kmem_cache_create("rbd_obj_request",
5669 sizeof (struct rbd_obj_request),
5670 __alignof__(struct rbd_obj_request),
5672 if (!rbd_obj_request_cache)
5675 rbd_assert(!rbd_segment_name_cache);
5676 rbd_segment_name_cache = kmem_cache_create("rbd_segment_name",
5677 CEPH_MAX_OID_NAME_LEN + 1, 1, 0, NULL);
5678 if (rbd_segment_name_cache)
5681 if (rbd_obj_request_cache) {
5682 kmem_cache_destroy(rbd_obj_request_cache);
5683 rbd_obj_request_cache = NULL;
5686 kmem_cache_destroy(rbd_img_request_cache);
5687 rbd_img_request_cache = NULL;
5692 static void rbd_slab_exit(void)
5694 rbd_assert(rbd_segment_name_cache);
5695 kmem_cache_destroy(rbd_segment_name_cache);
5696 rbd_segment_name_cache = NULL;
5698 rbd_assert(rbd_obj_request_cache);
5699 kmem_cache_destroy(rbd_obj_request_cache);
5700 rbd_obj_request_cache = NULL;
5702 rbd_assert(rbd_img_request_cache);
5703 kmem_cache_destroy(rbd_img_request_cache);
5704 rbd_img_request_cache = NULL;
5707 static int __init rbd_init(void)
5711 if (!libceph_compatible(NULL)) {
5712 rbd_warn(NULL, "libceph incompatibility (quitting)");
5716 rc = rbd_slab_init();
5721 * The number of active work items is limited by the number of
5722 * rbd devices * queue depth, so leave @max_active at default.
5724 rbd_wq = alloc_workqueue(RBD_DRV_NAME, WQ_MEM_RECLAIM, 0);
5731 rbd_major = register_blkdev(0, RBD_DRV_NAME);
5732 if (rbd_major < 0) {
5738 rc = rbd_sysfs_init();
5740 goto err_out_blkdev;
5743 pr_info("loaded (major %d)\n", rbd_major);
5745 pr_info("loaded\n");
5751 unregister_blkdev(rbd_major, RBD_DRV_NAME);
5753 destroy_workqueue(rbd_wq);
5759 static void __exit rbd_exit(void)
5761 ida_destroy(&rbd_dev_id_ida);
5762 rbd_sysfs_cleanup();
5764 unregister_blkdev(rbd_major, RBD_DRV_NAME);
5765 destroy_workqueue(rbd_wq);
5769 module_init(rbd_init);
5770 module_exit(rbd_exit);
5772 MODULE_AUTHOR("Alex Elder <elder@inktank.com>");
5773 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
5774 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
5775 /* following authorship retained from original osdblk.c */
5776 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
5778 MODULE_DESCRIPTION("RADOS Block Device (RBD) driver");
5779 MODULE_LICENSE("GPL");