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 void rbd_osd_copyup_callback(struct rbd_obj_request *obj_request);
527 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request);
528 static void rbd_img_parent_read(struct rbd_obj_request *obj_request);
529 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev);
531 static int rbd_dev_refresh(struct rbd_device *rbd_dev);
532 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev);
533 static int rbd_dev_header_info(struct rbd_device *rbd_dev);
534 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev);
535 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
537 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
538 u8 *order, u64 *snap_size);
539 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
541 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name);
543 static int rbd_open(struct block_device *bdev, fmode_t mode)
545 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
546 bool removing = false;
548 if ((mode & FMODE_WRITE) && rbd_dev->mapping.read_only)
551 spin_lock_irq(&rbd_dev->lock);
552 if (test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags))
555 rbd_dev->open_count++;
556 spin_unlock_irq(&rbd_dev->lock);
560 (void) get_device(&rbd_dev->dev);
565 static void rbd_release(struct gendisk *disk, fmode_t mode)
567 struct rbd_device *rbd_dev = disk->private_data;
568 unsigned long open_count_before;
570 spin_lock_irq(&rbd_dev->lock);
571 open_count_before = rbd_dev->open_count--;
572 spin_unlock_irq(&rbd_dev->lock);
573 rbd_assert(open_count_before > 0);
575 put_device(&rbd_dev->dev);
578 static int rbd_ioctl_set_ro(struct rbd_device *rbd_dev, unsigned long arg)
583 bool ro_changed = false;
585 /* get_user() may sleep, so call it before taking rbd_dev->lock */
586 if (get_user(val, (int __user *)(arg)))
589 ro = val ? true : false;
590 /* Snapshot doesn't allow to write*/
591 if (rbd_dev->spec->snap_id != CEPH_NOSNAP && !ro)
594 spin_lock_irq(&rbd_dev->lock);
595 /* prevent others open this device */
596 if (rbd_dev->open_count > 1) {
601 if (rbd_dev->mapping.read_only != ro) {
602 rbd_dev->mapping.read_only = ro;
607 spin_unlock_irq(&rbd_dev->lock);
608 /* set_disk_ro() may sleep, so call it after releasing rbd_dev->lock */
609 if (ret == 0 && ro_changed)
610 set_disk_ro(rbd_dev->disk, ro ? 1 : 0);
615 static int rbd_ioctl(struct block_device *bdev, fmode_t mode,
616 unsigned int cmd, unsigned long arg)
618 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
623 ret = rbd_ioctl_set_ro(rbd_dev, arg);
633 static int rbd_compat_ioctl(struct block_device *bdev, fmode_t mode,
634 unsigned int cmd, unsigned long arg)
636 return rbd_ioctl(bdev, mode, cmd, arg);
638 #endif /* CONFIG_COMPAT */
640 static const struct block_device_operations rbd_bd_ops = {
641 .owner = THIS_MODULE,
643 .release = rbd_release,
646 .compat_ioctl = rbd_compat_ioctl,
651 * Initialize an rbd client instance. Success or not, this function
652 * consumes ceph_opts. Caller holds client_mutex.
654 static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts)
656 struct rbd_client *rbdc;
659 dout("%s:\n", __func__);
660 rbdc = kmalloc(sizeof(struct rbd_client), GFP_KERNEL);
664 kref_init(&rbdc->kref);
665 INIT_LIST_HEAD(&rbdc->node);
667 rbdc->client = ceph_create_client(ceph_opts, rbdc, 0, 0);
668 if (IS_ERR(rbdc->client))
670 ceph_opts = NULL; /* Now rbdc->client is responsible for ceph_opts */
672 ret = ceph_open_session(rbdc->client);
676 spin_lock(&rbd_client_list_lock);
677 list_add_tail(&rbdc->node, &rbd_client_list);
678 spin_unlock(&rbd_client_list_lock);
680 dout("%s: rbdc %p\n", __func__, rbdc);
684 ceph_destroy_client(rbdc->client);
689 ceph_destroy_options(ceph_opts);
690 dout("%s: error %d\n", __func__, ret);
695 static struct rbd_client *__rbd_get_client(struct rbd_client *rbdc)
697 kref_get(&rbdc->kref);
703 * Find a ceph client with specific addr and configuration. If
704 * found, bump its reference count.
706 static struct rbd_client *rbd_client_find(struct ceph_options *ceph_opts)
708 struct rbd_client *client_node;
711 if (ceph_opts->flags & CEPH_OPT_NOSHARE)
714 spin_lock(&rbd_client_list_lock);
715 list_for_each_entry(client_node, &rbd_client_list, node) {
716 if (!ceph_compare_options(ceph_opts, client_node->client)) {
717 __rbd_get_client(client_node);
723 spin_unlock(&rbd_client_list_lock);
725 return found ? client_node : NULL;
729 * (Per device) rbd map options
736 /* string args above */
742 static match_table_t rbd_opts_tokens = {
743 {Opt_queue_depth, "queue_depth=%d"},
745 /* string args above */
746 {Opt_read_only, "read_only"},
747 {Opt_read_only, "ro"}, /* Alternate spelling */
748 {Opt_read_write, "read_write"},
749 {Opt_read_write, "rw"}, /* Alternate spelling */
758 #define RBD_QUEUE_DEPTH_DEFAULT BLKDEV_MAX_RQ
759 #define RBD_READ_ONLY_DEFAULT false
761 static int parse_rbd_opts_token(char *c, void *private)
763 struct rbd_options *rbd_opts = private;
764 substring_t argstr[MAX_OPT_ARGS];
765 int token, intval, ret;
767 token = match_token(c, rbd_opts_tokens, argstr);
768 if (token < Opt_last_int) {
769 ret = match_int(&argstr[0], &intval);
771 pr_err("bad mount option arg (not int) at '%s'\n", c);
774 dout("got int token %d val %d\n", token, intval);
775 } else if (token > Opt_last_int && token < Opt_last_string) {
776 dout("got string token %d val %s\n", token, argstr[0].from);
778 dout("got token %d\n", token);
782 case Opt_queue_depth:
784 pr_err("queue_depth out of range\n");
787 rbd_opts->queue_depth = intval;
790 rbd_opts->read_only = true;
793 rbd_opts->read_only = false;
796 /* libceph prints "bad option" msg */
803 static char* obj_op_name(enum obj_operation_type op_type)
818 * Get a ceph client with specific addr and configuration, if one does
819 * not exist create it. Either way, ceph_opts is consumed by this
822 static struct rbd_client *rbd_get_client(struct ceph_options *ceph_opts)
824 struct rbd_client *rbdc;
826 mutex_lock_nested(&client_mutex, SINGLE_DEPTH_NESTING);
827 rbdc = rbd_client_find(ceph_opts);
828 if (rbdc) /* using an existing client */
829 ceph_destroy_options(ceph_opts);
831 rbdc = rbd_client_create(ceph_opts);
832 mutex_unlock(&client_mutex);
838 * Destroy ceph client
840 * Caller must hold rbd_client_list_lock.
842 static void rbd_client_release(struct kref *kref)
844 struct rbd_client *rbdc = container_of(kref, struct rbd_client, kref);
846 dout("%s: rbdc %p\n", __func__, rbdc);
847 spin_lock(&rbd_client_list_lock);
848 list_del(&rbdc->node);
849 spin_unlock(&rbd_client_list_lock);
851 ceph_destroy_client(rbdc->client);
856 * Drop reference to ceph client node. If it's not referenced anymore, release
859 static void rbd_put_client(struct rbd_client *rbdc)
862 kref_put(&rbdc->kref, rbd_client_release);
865 static bool rbd_image_format_valid(u32 image_format)
867 return image_format == 1 || image_format == 2;
870 static bool rbd_dev_ondisk_valid(struct rbd_image_header_ondisk *ondisk)
875 /* The header has to start with the magic rbd header text */
876 if (memcmp(&ondisk->text, RBD_HEADER_TEXT, sizeof (RBD_HEADER_TEXT)))
879 /* The bio layer requires at least sector-sized I/O */
881 if (ondisk->options.order < SECTOR_SHIFT)
884 /* If we use u64 in a few spots we may be able to loosen this */
886 if (ondisk->options.order > 8 * sizeof (int) - 1)
890 * The size of a snapshot header has to fit in a size_t, and
891 * that limits the number of snapshots.
893 snap_count = le32_to_cpu(ondisk->snap_count);
894 size = SIZE_MAX - sizeof (struct ceph_snap_context);
895 if (snap_count > size / sizeof (__le64))
899 * Not only that, but the size of the entire the snapshot
900 * header must also be representable in a size_t.
902 size -= snap_count * sizeof (__le64);
903 if ((u64) size < le64_to_cpu(ondisk->snap_names_len))
910 * Fill an rbd image header with information from the given format 1
913 static int rbd_header_from_disk(struct rbd_device *rbd_dev,
914 struct rbd_image_header_ondisk *ondisk)
916 struct rbd_image_header *header = &rbd_dev->header;
917 bool first_time = header->object_prefix == NULL;
918 struct ceph_snap_context *snapc;
919 char *object_prefix = NULL;
920 char *snap_names = NULL;
921 u64 *snap_sizes = NULL;
927 /* Allocate this now to avoid having to handle failure below */
932 len = strnlen(ondisk->object_prefix,
933 sizeof (ondisk->object_prefix));
934 object_prefix = kmalloc(len + 1, GFP_KERNEL);
937 memcpy(object_prefix, ondisk->object_prefix, len);
938 object_prefix[len] = '\0';
941 /* Allocate the snapshot context and fill it in */
943 snap_count = le32_to_cpu(ondisk->snap_count);
944 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
947 snapc->seq = le64_to_cpu(ondisk->snap_seq);
949 struct rbd_image_snap_ondisk *snaps;
950 u64 snap_names_len = le64_to_cpu(ondisk->snap_names_len);
952 /* We'll keep a copy of the snapshot names... */
954 if (snap_names_len > (u64)SIZE_MAX)
956 snap_names = kmalloc(snap_names_len, GFP_KERNEL);
960 /* ...as well as the array of their sizes. */
962 size = snap_count * sizeof (*header->snap_sizes);
963 snap_sizes = kmalloc(size, GFP_KERNEL);
968 * Copy the names, and fill in each snapshot's id
971 * Note that rbd_dev_v1_header_info() guarantees the
972 * ondisk buffer we're working with has
973 * snap_names_len bytes beyond the end of the
974 * snapshot id array, this memcpy() is safe.
976 memcpy(snap_names, &ondisk->snaps[snap_count], snap_names_len);
977 snaps = ondisk->snaps;
978 for (i = 0; i < snap_count; i++) {
979 snapc->snaps[i] = le64_to_cpu(snaps[i].id);
980 snap_sizes[i] = le64_to_cpu(snaps[i].image_size);
984 /* We won't fail any more, fill in the header */
987 header->object_prefix = object_prefix;
988 header->obj_order = ondisk->options.order;
989 header->crypt_type = ondisk->options.crypt_type;
990 header->comp_type = ondisk->options.comp_type;
991 /* The rest aren't used for format 1 images */
992 header->stripe_unit = 0;
993 header->stripe_count = 0;
994 header->features = 0;
996 ceph_put_snap_context(header->snapc);
997 kfree(header->snap_names);
998 kfree(header->snap_sizes);
1001 /* The remaining fields always get updated (when we refresh) */
1003 header->image_size = le64_to_cpu(ondisk->image_size);
1004 header->snapc = snapc;
1005 header->snap_names = snap_names;
1006 header->snap_sizes = snap_sizes;
1014 ceph_put_snap_context(snapc);
1015 kfree(object_prefix);
1020 static const char *_rbd_dev_v1_snap_name(struct rbd_device *rbd_dev, u32 which)
1022 const char *snap_name;
1024 rbd_assert(which < rbd_dev->header.snapc->num_snaps);
1026 /* Skip over names until we find the one we are looking for */
1028 snap_name = rbd_dev->header.snap_names;
1030 snap_name += strlen(snap_name) + 1;
1032 return kstrdup(snap_name, GFP_KERNEL);
1036 * Snapshot id comparison function for use with qsort()/bsearch().
1037 * Note that result is for snapshots in *descending* order.
1039 static int snapid_compare_reverse(const void *s1, const void *s2)
1041 u64 snap_id1 = *(u64 *)s1;
1042 u64 snap_id2 = *(u64 *)s2;
1044 if (snap_id1 < snap_id2)
1046 return snap_id1 == snap_id2 ? 0 : -1;
1050 * Search a snapshot context to see if the given snapshot id is
1053 * Returns the position of the snapshot id in the array if it's found,
1054 * or BAD_SNAP_INDEX otherwise.
1056 * Note: The snapshot array is in kept sorted (by the osd) in
1057 * reverse order, highest snapshot id first.
1059 static u32 rbd_dev_snap_index(struct rbd_device *rbd_dev, u64 snap_id)
1061 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
1064 found = bsearch(&snap_id, &snapc->snaps, snapc->num_snaps,
1065 sizeof (snap_id), snapid_compare_reverse);
1067 return found ? (u32)(found - &snapc->snaps[0]) : BAD_SNAP_INDEX;
1070 static const char *rbd_dev_v1_snap_name(struct rbd_device *rbd_dev,
1074 const char *snap_name;
1076 which = rbd_dev_snap_index(rbd_dev, snap_id);
1077 if (which == BAD_SNAP_INDEX)
1078 return ERR_PTR(-ENOENT);
1080 snap_name = _rbd_dev_v1_snap_name(rbd_dev, which);
1081 return snap_name ? snap_name : ERR_PTR(-ENOMEM);
1084 static const char *rbd_snap_name(struct rbd_device *rbd_dev, u64 snap_id)
1086 if (snap_id == CEPH_NOSNAP)
1087 return RBD_SNAP_HEAD_NAME;
1089 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1090 if (rbd_dev->image_format == 1)
1091 return rbd_dev_v1_snap_name(rbd_dev, snap_id);
1093 return rbd_dev_v2_snap_name(rbd_dev, snap_id);
1096 static int rbd_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
1099 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1100 if (snap_id == CEPH_NOSNAP) {
1101 *snap_size = rbd_dev->header.image_size;
1102 } else if (rbd_dev->image_format == 1) {
1105 which = rbd_dev_snap_index(rbd_dev, snap_id);
1106 if (which == BAD_SNAP_INDEX)
1109 *snap_size = rbd_dev->header.snap_sizes[which];
1114 ret = _rbd_dev_v2_snap_size(rbd_dev, snap_id, NULL, &size);
1123 static int rbd_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
1126 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1127 if (snap_id == CEPH_NOSNAP) {
1128 *snap_features = rbd_dev->header.features;
1129 } else if (rbd_dev->image_format == 1) {
1130 *snap_features = 0; /* No features for format 1 */
1135 ret = _rbd_dev_v2_snap_features(rbd_dev, snap_id, &features);
1139 *snap_features = features;
1144 static int rbd_dev_mapping_set(struct rbd_device *rbd_dev)
1146 u64 snap_id = rbd_dev->spec->snap_id;
1151 ret = rbd_snap_size(rbd_dev, snap_id, &size);
1154 ret = rbd_snap_features(rbd_dev, snap_id, &features);
1158 rbd_dev->mapping.size = size;
1159 rbd_dev->mapping.features = features;
1164 static void rbd_dev_mapping_clear(struct rbd_device *rbd_dev)
1166 rbd_dev->mapping.size = 0;
1167 rbd_dev->mapping.features = 0;
1170 static void rbd_segment_name_free(const char *name)
1172 /* The explicit cast here is needed to drop the const qualifier */
1174 kmem_cache_free(rbd_segment_name_cache, (void *)name);
1177 static const char *rbd_segment_name(struct rbd_device *rbd_dev, u64 offset)
1184 name = kmem_cache_alloc(rbd_segment_name_cache, GFP_NOIO);
1187 segment = offset >> rbd_dev->header.obj_order;
1188 name_format = "%s.%012llx";
1189 if (rbd_dev->image_format == 2)
1190 name_format = "%s.%016llx";
1191 ret = snprintf(name, CEPH_MAX_OID_NAME_LEN + 1, name_format,
1192 rbd_dev->header.object_prefix, segment);
1193 if (ret < 0 || ret > CEPH_MAX_OID_NAME_LEN) {
1194 pr_err("error formatting segment name for #%llu (%d)\n",
1196 rbd_segment_name_free(name);
1203 static u64 rbd_segment_offset(struct rbd_device *rbd_dev, u64 offset)
1205 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1207 return offset & (segment_size - 1);
1210 static u64 rbd_segment_length(struct rbd_device *rbd_dev,
1211 u64 offset, u64 length)
1213 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1215 offset &= segment_size - 1;
1217 rbd_assert(length <= U64_MAX - offset);
1218 if (offset + length > segment_size)
1219 length = segment_size - offset;
1225 * returns the size of an object in the image
1227 static u64 rbd_obj_bytes(struct rbd_image_header *header)
1229 return 1 << header->obj_order;
1236 static void bio_chain_put(struct bio *chain)
1242 chain = chain->bi_next;
1248 * zeros a bio chain, starting at specific offset
1250 static void zero_bio_chain(struct bio *chain, int start_ofs)
1253 struct bvec_iter iter;
1254 unsigned long flags;
1259 bio_for_each_segment(bv, chain, iter) {
1260 if (pos + bv.bv_len > start_ofs) {
1261 int remainder = max(start_ofs - pos, 0);
1262 buf = bvec_kmap_irq(&bv, &flags);
1263 memset(buf + remainder, 0,
1264 bv.bv_len - remainder);
1265 flush_dcache_page(bv.bv_page);
1266 bvec_kunmap_irq(buf, &flags);
1271 chain = chain->bi_next;
1276 * similar to zero_bio_chain(), zeros data defined by a page array,
1277 * starting at the given byte offset from the start of the array and
1278 * continuing up to the given end offset. The pages array is
1279 * assumed to be big enough to hold all bytes up to the end.
1281 static void zero_pages(struct page **pages, u64 offset, u64 end)
1283 struct page **page = &pages[offset >> PAGE_SHIFT];
1285 rbd_assert(end > offset);
1286 rbd_assert(end - offset <= (u64)SIZE_MAX);
1287 while (offset < end) {
1290 unsigned long flags;
1293 page_offset = offset & ~PAGE_MASK;
1294 length = min_t(size_t, PAGE_SIZE - page_offset, end - offset);
1295 local_irq_save(flags);
1296 kaddr = kmap_atomic(*page);
1297 memset(kaddr + page_offset, 0, length);
1298 flush_dcache_page(*page);
1299 kunmap_atomic(kaddr);
1300 local_irq_restore(flags);
1308 * Clone a portion of a bio, starting at the given byte offset
1309 * and continuing for the number of bytes indicated.
1311 static struct bio *bio_clone_range(struct bio *bio_src,
1312 unsigned int offset,
1318 bio = bio_clone(bio_src, gfpmask);
1320 return NULL; /* ENOMEM */
1322 bio_advance(bio, offset);
1323 bio->bi_iter.bi_size = len;
1329 * Clone a portion of a bio chain, starting at the given byte offset
1330 * into the first bio in the source chain and continuing for the
1331 * number of bytes indicated. The result is another bio chain of
1332 * exactly the given length, or a null pointer on error.
1334 * The bio_src and offset parameters are both in-out. On entry they
1335 * refer to the first source bio and the offset into that bio where
1336 * the start of data to be cloned is located.
1338 * On return, bio_src is updated to refer to the bio in the source
1339 * chain that contains first un-cloned byte, and *offset will
1340 * contain the offset of that byte within that bio.
1342 static struct bio *bio_chain_clone_range(struct bio **bio_src,
1343 unsigned int *offset,
1347 struct bio *bi = *bio_src;
1348 unsigned int off = *offset;
1349 struct bio *chain = NULL;
1352 /* Build up a chain of clone bios up to the limit */
1354 if (!bi || off >= bi->bi_iter.bi_size || !len)
1355 return NULL; /* Nothing to clone */
1359 unsigned int bi_size;
1363 rbd_warn(NULL, "bio_chain exhausted with %u left", len);
1364 goto out_err; /* EINVAL; ran out of bio's */
1366 bi_size = min_t(unsigned int, bi->bi_iter.bi_size - off, len);
1367 bio = bio_clone_range(bi, off, bi_size, gfpmask);
1369 goto out_err; /* ENOMEM */
1372 end = &bio->bi_next;
1375 if (off == bi->bi_iter.bi_size) {
1386 bio_chain_put(chain);
1392 * The default/initial value for all object request flags is 0. For
1393 * each flag, once its value is set to 1 it is never reset to 0
1396 static void obj_request_img_data_set(struct rbd_obj_request *obj_request)
1398 if (test_and_set_bit(OBJ_REQ_IMG_DATA, &obj_request->flags)) {
1399 struct rbd_device *rbd_dev;
1401 rbd_dev = obj_request->img_request->rbd_dev;
1402 rbd_warn(rbd_dev, "obj_request %p already marked img_data",
1407 static bool obj_request_img_data_test(struct rbd_obj_request *obj_request)
1410 return test_bit(OBJ_REQ_IMG_DATA, &obj_request->flags) != 0;
1413 static void obj_request_done_set(struct rbd_obj_request *obj_request)
1415 if (test_and_set_bit(OBJ_REQ_DONE, &obj_request->flags)) {
1416 struct rbd_device *rbd_dev = NULL;
1418 if (obj_request_img_data_test(obj_request))
1419 rbd_dev = obj_request->img_request->rbd_dev;
1420 rbd_warn(rbd_dev, "obj_request %p already marked done",
1425 static bool obj_request_done_test(struct rbd_obj_request *obj_request)
1428 return test_bit(OBJ_REQ_DONE, &obj_request->flags) != 0;
1432 * This sets the KNOWN flag after (possibly) setting the EXISTS
1433 * flag. The latter is set based on the "exists" value provided.
1435 * Note that for our purposes once an object exists it never goes
1436 * away again. It's possible that the response from two existence
1437 * checks are separated by the creation of the target object, and
1438 * the first ("doesn't exist") response arrives *after* the second
1439 * ("does exist"). In that case we ignore the second one.
1441 static void obj_request_existence_set(struct rbd_obj_request *obj_request,
1445 set_bit(OBJ_REQ_EXISTS, &obj_request->flags);
1446 set_bit(OBJ_REQ_KNOWN, &obj_request->flags);
1450 static bool obj_request_known_test(struct rbd_obj_request *obj_request)
1453 return test_bit(OBJ_REQ_KNOWN, &obj_request->flags) != 0;
1456 static bool obj_request_exists_test(struct rbd_obj_request *obj_request)
1459 return test_bit(OBJ_REQ_EXISTS, &obj_request->flags) != 0;
1462 static bool obj_request_overlaps_parent(struct rbd_obj_request *obj_request)
1464 struct rbd_device *rbd_dev = obj_request->img_request->rbd_dev;
1466 return obj_request->img_offset <
1467 round_up(rbd_dev->parent_overlap, rbd_obj_bytes(&rbd_dev->header));
1470 static void rbd_obj_request_get(struct rbd_obj_request *obj_request)
1472 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1473 atomic_read(&obj_request->kref.refcount));
1474 kref_get(&obj_request->kref);
1477 static void rbd_obj_request_destroy(struct kref *kref);
1478 static void rbd_obj_request_put(struct rbd_obj_request *obj_request)
1480 rbd_assert(obj_request != NULL);
1481 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1482 atomic_read(&obj_request->kref.refcount));
1483 kref_put(&obj_request->kref, rbd_obj_request_destroy);
1486 static void rbd_img_request_get(struct rbd_img_request *img_request)
1488 dout("%s: img %p (was %d)\n", __func__, img_request,
1489 atomic_read(&img_request->kref.refcount));
1490 kref_get(&img_request->kref);
1493 static bool img_request_child_test(struct rbd_img_request *img_request);
1494 static void rbd_parent_request_destroy(struct kref *kref);
1495 static void rbd_img_request_destroy(struct kref *kref);
1496 static void rbd_img_request_put(struct rbd_img_request *img_request)
1498 rbd_assert(img_request != NULL);
1499 dout("%s: img %p (was %d)\n", __func__, img_request,
1500 atomic_read(&img_request->kref.refcount));
1501 if (img_request_child_test(img_request))
1502 kref_put(&img_request->kref, rbd_parent_request_destroy);
1504 kref_put(&img_request->kref, rbd_img_request_destroy);
1507 static inline void rbd_img_obj_request_add(struct rbd_img_request *img_request,
1508 struct rbd_obj_request *obj_request)
1510 rbd_assert(obj_request->img_request == NULL);
1512 /* Image request now owns object's original reference */
1513 obj_request->img_request = img_request;
1514 obj_request->which = img_request->obj_request_count;
1515 rbd_assert(!obj_request_img_data_test(obj_request));
1516 obj_request_img_data_set(obj_request);
1517 rbd_assert(obj_request->which != BAD_WHICH);
1518 img_request->obj_request_count++;
1519 list_add_tail(&obj_request->links, &img_request->obj_requests);
1520 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1521 obj_request->which);
1524 static inline void rbd_img_obj_request_del(struct rbd_img_request *img_request,
1525 struct rbd_obj_request *obj_request)
1527 rbd_assert(obj_request->which != BAD_WHICH);
1529 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1530 obj_request->which);
1531 list_del(&obj_request->links);
1532 rbd_assert(img_request->obj_request_count > 0);
1533 img_request->obj_request_count--;
1534 rbd_assert(obj_request->which == img_request->obj_request_count);
1535 obj_request->which = BAD_WHICH;
1536 rbd_assert(obj_request_img_data_test(obj_request));
1537 rbd_assert(obj_request->img_request == img_request);
1538 obj_request->img_request = NULL;
1539 obj_request->callback = NULL;
1540 rbd_obj_request_put(obj_request);
1543 static bool obj_request_type_valid(enum obj_request_type type)
1546 case OBJ_REQUEST_NODATA:
1547 case OBJ_REQUEST_BIO:
1548 case OBJ_REQUEST_PAGES:
1555 static int rbd_obj_request_submit(struct ceph_osd_client *osdc,
1556 struct rbd_obj_request *obj_request)
1558 dout("%s %p\n", __func__, obj_request);
1559 return ceph_osdc_start_request(osdc, obj_request->osd_req, false);
1562 static void rbd_obj_request_end(struct rbd_obj_request *obj_request)
1564 dout("%s %p\n", __func__, obj_request);
1565 ceph_osdc_cancel_request(obj_request->osd_req);
1569 * Wait for an object request to complete. If interrupted, cancel the
1570 * underlying osd request.
1572 * @timeout: in jiffies, 0 means "wait forever"
1574 static int __rbd_obj_request_wait(struct rbd_obj_request *obj_request,
1575 unsigned long timeout)
1579 dout("%s %p\n", __func__, obj_request);
1580 ret = wait_for_completion_interruptible_timeout(
1581 &obj_request->completion,
1582 ceph_timeout_jiffies(timeout));
1586 rbd_obj_request_end(obj_request);
1591 dout("%s %p ret %d\n", __func__, obj_request, (int)ret);
1595 static int rbd_obj_request_wait(struct rbd_obj_request *obj_request)
1597 return __rbd_obj_request_wait(obj_request, 0);
1600 static int rbd_obj_request_wait_timeout(struct rbd_obj_request *obj_request,
1601 unsigned long timeout)
1603 return __rbd_obj_request_wait(obj_request, timeout);
1606 static void rbd_img_request_complete(struct rbd_img_request *img_request)
1609 dout("%s: img %p\n", __func__, img_request);
1612 * If no error occurred, compute the aggregate transfer
1613 * count for the image request. We could instead use
1614 * atomic64_cmpxchg() to update it as each object request
1615 * completes; not clear which way is better off hand.
1617 if (!img_request->result) {
1618 struct rbd_obj_request *obj_request;
1621 for_each_obj_request(img_request, obj_request)
1622 xferred += obj_request->xferred;
1623 img_request->xferred = xferred;
1626 if (img_request->callback)
1627 img_request->callback(img_request);
1629 rbd_img_request_put(img_request);
1633 * The default/initial value for all image request flags is 0. Each
1634 * is conditionally set to 1 at image request initialization time
1635 * and currently never change thereafter.
1637 static void img_request_write_set(struct rbd_img_request *img_request)
1639 set_bit(IMG_REQ_WRITE, &img_request->flags);
1643 static bool img_request_write_test(struct rbd_img_request *img_request)
1646 return test_bit(IMG_REQ_WRITE, &img_request->flags) != 0;
1650 * Set the discard flag when the img_request is an discard request
1652 static void img_request_discard_set(struct rbd_img_request *img_request)
1654 set_bit(IMG_REQ_DISCARD, &img_request->flags);
1658 static bool img_request_discard_test(struct rbd_img_request *img_request)
1661 return test_bit(IMG_REQ_DISCARD, &img_request->flags) != 0;
1664 static void img_request_child_set(struct rbd_img_request *img_request)
1666 set_bit(IMG_REQ_CHILD, &img_request->flags);
1670 static void img_request_child_clear(struct rbd_img_request *img_request)
1672 clear_bit(IMG_REQ_CHILD, &img_request->flags);
1676 static bool img_request_child_test(struct rbd_img_request *img_request)
1679 return test_bit(IMG_REQ_CHILD, &img_request->flags) != 0;
1682 static void img_request_layered_set(struct rbd_img_request *img_request)
1684 set_bit(IMG_REQ_LAYERED, &img_request->flags);
1688 static void img_request_layered_clear(struct rbd_img_request *img_request)
1690 clear_bit(IMG_REQ_LAYERED, &img_request->flags);
1694 static bool img_request_layered_test(struct rbd_img_request *img_request)
1697 return test_bit(IMG_REQ_LAYERED, &img_request->flags) != 0;
1700 static enum obj_operation_type
1701 rbd_img_request_op_type(struct rbd_img_request *img_request)
1703 if (img_request_write_test(img_request))
1704 return OBJ_OP_WRITE;
1705 else if (img_request_discard_test(img_request))
1706 return OBJ_OP_DISCARD;
1712 rbd_img_obj_request_read_callback(struct rbd_obj_request *obj_request)
1714 u64 xferred = obj_request->xferred;
1715 u64 length = obj_request->length;
1717 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1718 obj_request, obj_request->img_request, obj_request->result,
1721 * ENOENT means a hole in the image. We zero-fill the entire
1722 * length of the request. A short read also implies zero-fill
1723 * to the end of the request. An error requires the whole
1724 * length of the request to be reported finished with an error
1725 * to the block layer. In each case we update the xferred
1726 * count to indicate the whole request was satisfied.
1728 rbd_assert(obj_request->type != OBJ_REQUEST_NODATA);
1729 if (obj_request->result == -ENOENT) {
1730 if (obj_request->type == OBJ_REQUEST_BIO)
1731 zero_bio_chain(obj_request->bio_list, 0);
1733 zero_pages(obj_request->pages, 0, length);
1734 obj_request->result = 0;
1735 } else if (xferred < length && !obj_request->result) {
1736 if (obj_request->type == OBJ_REQUEST_BIO)
1737 zero_bio_chain(obj_request->bio_list, xferred);
1739 zero_pages(obj_request->pages, xferred, length);
1741 obj_request->xferred = length;
1742 obj_request_done_set(obj_request);
1745 static void rbd_obj_request_complete(struct rbd_obj_request *obj_request)
1747 dout("%s: obj %p cb %p\n", __func__, obj_request,
1748 obj_request->callback);
1749 if (obj_request->callback)
1750 obj_request->callback(obj_request);
1752 complete_all(&obj_request->completion);
1755 static void rbd_osd_trivial_callback(struct rbd_obj_request *obj_request)
1757 dout("%s: obj %p\n", __func__, obj_request);
1758 obj_request_done_set(obj_request);
1761 static void rbd_osd_read_callback(struct rbd_obj_request *obj_request)
1763 struct rbd_img_request *img_request = NULL;
1764 struct rbd_device *rbd_dev = NULL;
1765 bool layered = false;
1767 if (obj_request_img_data_test(obj_request)) {
1768 img_request = obj_request->img_request;
1769 layered = img_request && img_request_layered_test(img_request);
1770 rbd_dev = img_request->rbd_dev;
1773 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1774 obj_request, img_request, obj_request->result,
1775 obj_request->xferred, obj_request->length);
1776 if (layered && obj_request->result == -ENOENT &&
1777 obj_request->img_offset < rbd_dev->parent_overlap)
1778 rbd_img_parent_read(obj_request);
1779 else if (img_request)
1780 rbd_img_obj_request_read_callback(obj_request);
1782 obj_request_done_set(obj_request);
1785 static void rbd_osd_write_callback(struct rbd_obj_request *obj_request)
1787 dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1788 obj_request->result, obj_request->length);
1790 * There is no such thing as a successful short write. Set
1791 * it to our originally-requested length.
1793 obj_request->xferred = obj_request->length;
1794 obj_request_done_set(obj_request);
1797 static void rbd_osd_discard_callback(struct rbd_obj_request *obj_request)
1799 dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1800 obj_request->result, obj_request->length);
1802 * There is no such thing as a successful short discard. Set
1803 * it to our originally-requested length.
1805 obj_request->xferred = obj_request->length;
1806 /* discarding a non-existent object is not a problem */
1807 if (obj_request->result == -ENOENT)
1808 obj_request->result = 0;
1809 obj_request_done_set(obj_request);
1813 * For a simple stat call there's nothing to do. We'll do more if
1814 * this is part of a write sequence for a layered image.
1816 static void rbd_osd_stat_callback(struct rbd_obj_request *obj_request)
1818 dout("%s: obj %p\n", __func__, obj_request);
1819 obj_request_done_set(obj_request);
1822 static void rbd_osd_call_callback(struct rbd_obj_request *obj_request)
1824 dout("%s: obj %p\n", __func__, obj_request);
1826 if (obj_request_img_data_test(obj_request))
1827 rbd_osd_copyup_callback(obj_request);
1829 obj_request_done_set(obj_request);
1832 static void rbd_osd_req_callback(struct ceph_osd_request *osd_req,
1833 struct ceph_msg *msg)
1835 struct rbd_obj_request *obj_request = osd_req->r_priv;
1838 dout("%s: osd_req %p msg %p\n", __func__, osd_req, msg);
1839 rbd_assert(osd_req == obj_request->osd_req);
1840 if (obj_request_img_data_test(obj_request)) {
1841 rbd_assert(obj_request->img_request);
1842 rbd_assert(obj_request->which != BAD_WHICH);
1844 rbd_assert(obj_request->which == BAD_WHICH);
1847 if (osd_req->r_result < 0)
1848 obj_request->result = osd_req->r_result;
1850 rbd_assert(osd_req->r_num_ops <= CEPH_OSD_MAX_OP);
1853 * We support a 64-bit length, but ultimately it has to be
1854 * passed to the block layer, which just supports a 32-bit
1857 obj_request->xferred = osd_req->r_reply_op_len[0];
1858 rbd_assert(obj_request->xferred < (u64)UINT_MAX);
1860 opcode = osd_req->r_ops[0].op;
1862 case CEPH_OSD_OP_READ:
1863 rbd_osd_read_callback(obj_request);
1865 case CEPH_OSD_OP_SETALLOCHINT:
1866 rbd_assert(osd_req->r_ops[1].op == CEPH_OSD_OP_WRITE);
1868 case CEPH_OSD_OP_WRITE:
1869 rbd_osd_write_callback(obj_request);
1871 case CEPH_OSD_OP_STAT:
1872 rbd_osd_stat_callback(obj_request);
1874 case CEPH_OSD_OP_DELETE:
1875 case CEPH_OSD_OP_TRUNCATE:
1876 case CEPH_OSD_OP_ZERO:
1877 rbd_osd_discard_callback(obj_request);
1879 case CEPH_OSD_OP_CALL:
1880 rbd_osd_call_callback(obj_request);
1882 case CEPH_OSD_OP_NOTIFY_ACK:
1883 case CEPH_OSD_OP_WATCH:
1884 rbd_osd_trivial_callback(obj_request);
1887 rbd_warn(NULL, "%s: unsupported op %hu",
1888 obj_request->object_name, (unsigned short) opcode);
1892 if (obj_request_done_test(obj_request))
1893 rbd_obj_request_complete(obj_request);
1896 static void rbd_osd_req_format_read(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;
1902 rbd_assert(osd_req != NULL);
1904 snap_id = img_request ? img_request->snap_id : CEPH_NOSNAP;
1905 ceph_osdc_build_request(osd_req, obj_request->offset,
1906 NULL, snap_id, NULL);
1909 static void rbd_osd_req_format_write(struct rbd_obj_request *obj_request)
1911 struct rbd_img_request *img_request = obj_request->img_request;
1912 struct ceph_osd_request *osd_req = obj_request->osd_req;
1913 struct ceph_snap_context *snapc;
1914 struct timespec mtime = CURRENT_TIME;
1916 rbd_assert(osd_req != NULL);
1918 snapc = img_request ? img_request->snapc : NULL;
1919 ceph_osdc_build_request(osd_req, obj_request->offset,
1920 snapc, CEPH_NOSNAP, &mtime);
1924 * Create an osd request. A read request has one osd op (read).
1925 * A write request has either one (watch) or two (hint+write) osd ops.
1926 * (All rbd data writes are prefixed with an allocation hint op, but
1927 * technically osd watch is a write request, hence this distinction.)
1929 static struct ceph_osd_request *rbd_osd_req_create(
1930 struct rbd_device *rbd_dev,
1931 enum obj_operation_type op_type,
1932 unsigned int num_ops,
1933 struct rbd_obj_request *obj_request)
1935 struct ceph_snap_context *snapc = NULL;
1936 struct ceph_osd_client *osdc;
1937 struct ceph_osd_request *osd_req;
1939 if (obj_request_img_data_test(obj_request) &&
1940 (op_type == OBJ_OP_DISCARD || op_type == OBJ_OP_WRITE)) {
1941 struct rbd_img_request *img_request = obj_request->img_request;
1942 if (op_type == OBJ_OP_WRITE) {
1943 rbd_assert(img_request_write_test(img_request));
1945 rbd_assert(img_request_discard_test(img_request));
1947 snapc = img_request->snapc;
1950 rbd_assert(num_ops == 1 || ((op_type == OBJ_OP_WRITE) && num_ops == 2));
1952 /* Allocate and initialize the request, for the num_ops ops */
1954 osdc = &rbd_dev->rbd_client->client->osdc;
1955 osd_req = ceph_osdc_alloc_request(osdc, snapc, num_ops, false,
1958 return NULL; /* ENOMEM */
1960 if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
1961 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
1963 osd_req->r_flags = CEPH_OSD_FLAG_READ;
1965 osd_req->r_callback = rbd_osd_req_callback;
1966 osd_req->r_priv = obj_request;
1968 osd_req->r_base_oloc.pool = ceph_file_layout_pg_pool(rbd_dev->layout);
1969 ceph_oid_set_name(&osd_req->r_base_oid, obj_request->object_name);
1975 * Create a copyup osd request based on the information in the object
1976 * request supplied. A copyup request has two or three osd ops, a
1977 * copyup method call, potentially a hint op, and a write or truncate
1980 static struct ceph_osd_request *
1981 rbd_osd_req_create_copyup(struct rbd_obj_request *obj_request)
1983 struct rbd_img_request *img_request;
1984 struct ceph_snap_context *snapc;
1985 struct rbd_device *rbd_dev;
1986 struct ceph_osd_client *osdc;
1987 struct ceph_osd_request *osd_req;
1988 int num_osd_ops = 3;
1990 rbd_assert(obj_request_img_data_test(obj_request));
1991 img_request = obj_request->img_request;
1992 rbd_assert(img_request);
1993 rbd_assert(img_request_write_test(img_request) ||
1994 img_request_discard_test(img_request));
1996 if (img_request_discard_test(img_request))
1999 /* Allocate and initialize the request, for all the ops */
2001 snapc = img_request->snapc;
2002 rbd_dev = img_request->rbd_dev;
2003 osdc = &rbd_dev->rbd_client->client->osdc;
2004 osd_req = ceph_osdc_alloc_request(osdc, snapc, num_osd_ops,
2007 return NULL; /* ENOMEM */
2009 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
2010 osd_req->r_callback = rbd_osd_req_callback;
2011 osd_req->r_priv = obj_request;
2013 osd_req->r_base_oloc.pool = ceph_file_layout_pg_pool(rbd_dev->layout);
2014 ceph_oid_set_name(&osd_req->r_base_oid, obj_request->object_name);
2020 static void rbd_osd_req_destroy(struct ceph_osd_request *osd_req)
2022 ceph_osdc_put_request(osd_req);
2025 /* object_name is assumed to be a non-null pointer and NUL-terminated */
2027 static struct rbd_obj_request *rbd_obj_request_create(const char *object_name,
2028 u64 offset, u64 length,
2029 enum obj_request_type type)
2031 struct rbd_obj_request *obj_request;
2035 rbd_assert(obj_request_type_valid(type));
2037 size = strlen(object_name) + 1;
2038 name = kmalloc(size, GFP_NOIO);
2042 obj_request = kmem_cache_zalloc(rbd_obj_request_cache, GFP_NOIO);
2048 obj_request->object_name = memcpy(name, object_name, size);
2049 obj_request->offset = offset;
2050 obj_request->length = length;
2051 obj_request->flags = 0;
2052 obj_request->which = BAD_WHICH;
2053 obj_request->type = type;
2054 INIT_LIST_HEAD(&obj_request->links);
2055 init_completion(&obj_request->completion);
2056 kref_init(&obj_request->kref);
2058 dout("%s: \"%s\" %llu/%llu %d -> obj %p\n", __func__, object_name,
2059 offset, length, (int)type, obj_request);
2064 static void rbd_obj_request_destroy(struct kref *kref)
2066 struct rbd_obj_request *obj_request;
2068 obj_request = container_of(kref, struct rbd_obj_request, kref);
2070 dout("%s: obj %p\n", __func__, obj_request);
2072 rbd_assert(obj_request->img_request == NULL);
2073 rbd_assert(obj_request->which == BAD_WHICH);
2075 if (obj_request->osd_req)
2076 rbd_osd_req_destroy(obj_request->osd_req);
2078 rbd_assert(obj_request_type_valid(obj_request->type));
2079 switch (obj_request->type) {
2080 case OBJ_REQUEST_NODATA:
2081 break; /* Nothing to do */
2082 case OBJ_REQUEST_BIO:
2083 if (obj_request->bio_list)
2084 bio_chain_put(obj_request->bio_list);
2086 case OBJ_REQUEST_PAGES:
2087 if (obj_request->pages)
2088 ceph_release_page_vector(obj_request->pages,
2089 obj_request->page_count);
2093 kfree(obj_request->object_name);
2094 obj_request->object_name = NULL;
2095 kmem_cache_free(rbd_obj_request_cache, obj_request);
2098 /* It's OK to call this for a device with no parent */
2100 static void rbd_spec_put(struct rbd_spec *spec);
2101 static void rbd_dev_unparent(struct rbd_device *rbd_dev)
2103 rbd_dev_remove_parent(rbd_dev);
2104 rbd_spec_put(rbd_dev->parent_spec);
2105 rbd_dev->parent_spec = NULL;
2106 rbd_dev->parent_overlap = 0;
2110 * Parent image reference counting is used to determine when an
2111 * image's parent fields can be safely torn down--after there are no
2112 * more in-flight requests to the parent image. When the last
2113 * reference is dropped, cleaning them up is safe.
2115 static void rbd_dev_parent_put(struct rbd_device *rbd_dev)
2119 if (!rbd_dev->parent_spec)
2122 counter = atomic_dec_return_safe(&rbd_dev->parent_ref);
2126 /* Last reference; clean up parent data structures */
2129 rbd_dev_unparent(rbd_dev);
2131 rbd_warn(rbd_dev, "parent reference underflow");
2135 * If an image has a non-zero parent overlap, get a reference to its
2138 * Returns true if the rbd device has a parent with a non-zero
2139 * overlap and a reference for it was successfully taken, or
2142 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
2146 if (!rbd_dev->parent_spec)
2149 down_read(&rbd_dev->header_rwsem);
2150 if (rbd_dev->parent_overlap)
2151 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
2152 up_read(&rbd_dev->header_rwsem);
2155 rbd_warn(rbd_dev, "parent reference overflow");
2161 * Caller is responsible for filling in the list of object requests
2162 * that comprises the image request, and the Linux request pointer
2163 * (if there is one).
2165 static struct rbd_img_request *rbd_img_request_create(
2166 struct rbd_device *rbd_dev,
2167 u64 offset, u64 length,
2168 enum obj_operation_type op_type,
2169 struct ceph_snap_context *snapc)
2171 struct rbd_img_request *img_request;
2173 img_request = kmem_cache_alloc(rbd_img_request_cache, GFP_NOIO);
2177 img_request->rq = NULL;
2178 img_request->rbd_dev = rbd_dev;
2179 img_request->offset = offset;
2180 img_request->length = length;
2181 img_request->flags = 0;
2182 if (op_type == OBJ_OP_DISCARD) {
2183 img_request_discard_set(img_request);
2184 img_request->snapc = snapc;
2185 } else if (op_type == OBJ_OP_WRITE) {
2186 img_request_write_set(img_request);
2187 img_request->snapc = snapc;
2189 img_request->snap_id = rbd_dev->spec->snap_id;
2191 if (rbd_dev_parent_get(rbd_dev))
2192 img_request_layered_set(img_request);
2193 spin_lock_init(&img_request->completion_lock);
2194 img_request->next_completion = 0;
2195 img_request->callback = NULL;
2196 img_request->result = 0;
2197 img_request->obj_request_count = 0;
2198 INIT_LIST_HEAD(&img_request->obj_requests);
2199 kref_init(&img_request->kref);
2201 dout("%s: rbd_dev %p %s %llu/%llu -> img %p\n", __func__, rbd_dev,
2202 obj_op_name(op_type), offset, length, img_request);
2207 static void rbd_img_request_destroy(struct kref *kref)
2209 struct rbd_img_request *img_request;
2210 struct rbd_obj_request *obj_request;
2211 struct rbd_obj_request *next_obj_request;
2213 img_request = container_of(kref, struct rbd_img_request, kref);
2215 dout("%s: img %p\n", __func__, img_request);
2217 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2218 rbd_img_obj_request_del(img_request, obj_request);
2219 rbd_assert(img_request->obj_request_count == 0);
2221 if (img_request_layered_test(img_request)) {
2222 img_request_layered_clear(img_request);
2223 rbd_dev_parent_put(img_request->rbd_dev);
2226 if (img_request_write_test(img_request) ||
2227 img_request_discard_test(img_request))
2228 ceph_put_snap_context(img_request->snapc);
2230 kmem_cache_free(rbd_img_request_cache, img_request);
2233 static struct rbd_img_request *rbd_parent_request_create(
2234 struct rbd_obj_request *obj_request,
2235 u64 img_offset, u64 length)
2237 struct rbd_img_request *parent_request;
2238 struct rbd_device *rbd_dev;
2240 rbd_assert(obj_request->img_request);
2241 rbd_dev = obj_request->img_request->rbd_dev;
2243 parent_request = rbd_img_request_create(rbd_dev->parent, img_offset,
2244 length, OBJ_OP_READ, NULL);
2245 if (!parent_request)
2248 img_request_child_set(parent_request);
2249 rbd_obj_request_get(obj_request);
2250 parent_request->obj_request = obj_request;
2252 return parent_request;
2255 static void rbd_parent_request_destroy(struct kref *kref)
2257 struct rbd_img_request *parent_request;
2258 struct rbd_obj_request *orig_request;
2260 parent_request = container_of(kref, struct rbd_img_request, kref);
2261 orig_request = parent_request->obj_request;
2263 parent_request->obj_request = NULL;
2264 rbd_obj_request_put(orig_request);
2265 img_request_child_clear(parent_request);
2267 rbd_img_request_destroy(kref);
2270 static bool rbd_img_obj_end_request(struct rbd_obj_request *obj_request)
2272 struct rbd_img_request *img_request;
2273 unsigned int xferred;
2277 rbd_assert(obj_request_img_data_test(obj_request));
2278 img_request = obj_request->img_request;
2280 rbd_assert(obj_request->xferred <= (u64)UINT_MAX);
2281 xferred = (unsigned int)obj_request->xferred;
2282 result = obj_request->result;
2284 struct rbd_device *rbd_dev = img_request->rbd_dev;
2285 enum obj_operation_type op_type;
2287 if (img_request_discard_test(img_request))
2288 op_type = OBJ_OP_DISCARD;
2289 else if (img_request_write_test(img_request))
2290 op_type = OBJ_OP_WRITE;
2292 op_type = OBJ_OP_READ;
2294 rbd_warn(rbd_dev, "%s %llx at %llx (%llx)",
2295 obj_op_name(op_type), obj_request->length,
2296 obj_request->img_offset, obj_request->offset);
2297 rbd_warn(rbd_dev, " result %d xferred %x",
2299 if (!img_request->result)
2300 img_request->result = result;
2302 * Need to end I/O on the entire obj_request worth of
2303 * bytes in case of error.
2305 xferred = obj_request->length;
2308 /* Image object requests don't own their page array */
2310 if (obj_request->type == OBJ_REQUEST_PAGES) {
2311 obj_request->pages = NULL;
2312 obj_request->page_count = 0;
2315 if (img_request_child_test(img_request)) {
2316 rbd_assert(img_request->obj_request != NULL);
2317 more = obj_request->which < img_request->obj_request_count - 1;
2319 rbd_assert(img_request->rq != NULL);
2321 more = blk_update_request(img_request->rq, result, xferred);
2323 __blk_mq_end_request(img_request->rq, result);
2329 static void rbd_img_obj_callback(struct rbd_obj_request *obj_request)
2331 struct rbd_img_request *img_request;
2332 u32 which = obj_request->which;
2335 rbd_assert(obj_request_img_data_test(obj_request));
2336 img_request = obj_request->img_request;
2338 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
2339 rbd_assert(img_request != NULL);
2340 rbd_assert(img_request->obj_request_count > 0);
2341 rbd_assert(which != BAD_WHICH);
2342 rbd_assert(which < img_request->obj_request_count);
2344 spin_lock_irq(&img_request->completion_lock);
2345 if (which != img_request->next_completion)
2348 for_each_obj_request_from(img_request, obj_request) {
2350 rbd_assert(which < img_request->obj_request_count);
2352 if (!obj_request_done_test(obj_request))
2354 more = rbd_img_obj_end_request(obj_request);
2358 rbd_assert(more ^ (which == img_request->obj_request_count));
2359 img_request->next_completion = which;
2361 spin_unlock_irq(&img_request->completion_lock);
2362 rbd_img_request_put(img_request);
2365 rbd_img_request_complete(img_request);
2369 * Add individual osd ops to the given ceph_osd_request and prepare
2370 * them for submission. num_ops is the current number of
2371 * osd operations already to the object request.
2373 static void rbd_img_obj_request_fill(struct rbd_obj_request *obj_request,
2374 struct ceph_osd_request *osd_request,
2375 enum obj_operation_type op_type,
2376 unsigned int num_ops)
2378 struct rbd_img_request *img_request = obj_request->img_request;
2379 struct rbd_device *rbd_dev = img_request->rbd_dev;
2380 u64 object_size = rbd_obj_bytes(&rbd_dev->header);
2381 u64 offset = obj_request->offset;
2382 u64 length = obj_request->length;
2386 if (op_type == OBJ_OP_DISCARD) {
2387 if (!offset && length == object_size &&
2388 (!img_request_layered_test(img_request) ||
2389 !obj_request_overlaps_parent(obj_request))) {
2390 opcode = CEPH_OSD_OP_DELETE;
2391 } else if ((offset + length == object_size)) {
2392 opcode = CEPH_OSD_OP_TRUNCATE;
2394 down_read(&rbd_dev->header_rwsem);
2395 img_end = rbd_dev->header.image_size;
2396 up_read(&rbd_dev->header_rwsem);
2398 if (obj_request->img_offset + length == img_end)
2399 opcode = CEPH_OSD_OP_TRUNCATE;
2401 opcode = CEPH_OSD_OP_ZERO;
2403 } else if (op_type == OBJ_OP_WRITE) {
2404 opcode = CEPH_OSD_OP_WRITE;
2405 osd_req_op_alloc_hint_init(osd_request, num_ops,
2406 object_size, object_size);
2409 opcode = CEPH_OSD_OP_READ;
2412 if (opcode == CEPH_OSD_OP_DELETE)
2413 osd_req_op_init(osd_request, num_ops, opcode, 0);
2415 osd_req_op_extent_init(osd_request, num_ops, opcode,
2416 offset, length, 0, 0);
2418 if (obj_request->type == OBJ_REQUEST_BIO)
2419 osd_req_op_extent_osd_data_bio(osd_request, num_ops,
2420 obj_request->bio_list, length);
2421 else if (obj_request->type == OBJ_REQUEST_PAGES)
2422 osd_req_op_extent_osd_data_pages(osd_request, num_ops,
2423 obj_request->pages, length,
2424 offset & ~PAGE_MASK, false, false);
2426 /* Discards are also writes */
2427 if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
2428 rbd_osd_req_format_write(obj_request);
2430 rbd_osd_req_format_read(obj_request);
2434 * Split up an image request into one or more object requests, each
2435 * to a different object. The "type" parameter indicates whether
2436 * "data_desc" is the pointer to the head of a list of bio
2437 * structures, or the base of a page array. In either case this
2438 * function assumes data_desc describes memory sufficient to hold
2439 * all data described by the image request.
2441 static int rbd_img_request_fill(struct rbd_img_request *img_request,
2442 enum obj_request_type type,
2445 struct rbd_device *rbd_dev = img_request->rbd_dev;
2446 struct rbd_obj_request *obj_request = NULL;
2447 struct rbd_obj_request *next_obj_request;
2448 struct bio *bio_list = NULL;
2449 unsigned int bio_offset = 0;
2450 struct page **pages = NULL;
2451 enum obj_operation_type op_type;
2455 dout("%s: img %p type %d data_desc %p\n", __func__, img_request,
2456 (int)type, data_desc);
2458 img_offset = img_request->offset;
2459 resid = img_request->length;
2460 rbd_assert(resid > 0);
2461 op_type = rbd_img_request_op_type(img_request);
2463 if (type == OBJ_REQUEST_BIO) {
2464 bio_list = data_desc;
2465 rbd_assert(img_offset ==
2466 bio_list->bi_iter.bi_sector << SECTOR_SHIFT);
2467 } else if (type == OBJ_REQUEST_PAGES) {
2472 struct ceph_osd_request *osd_req;
2473 const char *object_name;
2477 object_name = rbd_segment_name(rbd_dev, img_offset);
2480 offset = rbd_segment_offset(rbd_dev, img_offset);
2481 length = rbd_segment_length(rbd_dev, img_offset, resid);
2482 obj_request = rbd_obj_request_create(object_name,
2483 offset, length, type);
2484 /* object request has its own copy of the object name */
2485 rbd_segment_name_free(object_name);
2490 * set obj_request->img_request before creating the
2491 * osd_request so that it gets the right snapc
2493 rbd_img_obj_request_add(img_request, obj_request);
2495 if (type == OBJ_REQUEST_BIO) {
2496 unsigned int clone_size;
2498 rbd_assert(length <= (u64)UINT_MAX);
2499 clone_size = (unsigned int)length;
2500 obj_request->bio_list =
2501 bio_chain_clone_range(&bio_list,
2505 if (!obj_request->bio_list)
2507 } else if (type == OBJ_REQUEST_PAGES) {
2508 unsigned int page_count;
2510 obj_request->pages = pages;
2511 page_count = (u32)calc_pages_for(offset, length);
2512 obj_request->page_count = page_count;
2513 if ((offset + length) & ~PAGE_MASK)
2514 page_count--; /* more on last page */
2515 pages += page_count;
2518 osd_req = rbd_osd_req_create(rbd_dev, op_type,
2519 (op_type == OBJ_OP_WRITE) ? 2 : 1,
2524 obj_request->osd_req = osd_req;
2525 obj_request->callback = rbd_img_obj_callback;
2526 obj_request->img_offset = img_offset;
2528 rbd_img_obj_request_fill(obj_request, osd_req, op_type, 0);
2530 rbd_img_request_get(img_request);
2532 img_offset += length;
2539 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2540 rbd_img_obj_request_del(img_request, obj_request);
2546 rbd_osd_copyup_callback(struct rbd_obj_request *obj_request)
2548 struct rbd_img_request *img_request;
2549 struct rbd_device *rbd_dev;
2550 struct page **pages;
2553 dout("%s: obj %p\n", __func__, obj_request);
2555 rbd_assert(obj_request->type == OBJ_REQUEST_BIO ||
2556 obj_request->type == OBJ_REQUEST_NODATA);
2557 rbd_assert(obj_request_img_data_test(obj_request));
2558 img_request = obj_request->img_request;
2559 rbd_assert(img_request);
2561 rbd_dev = img_request->rbd_dev;
2562 rbd_assert(rbd_dev);
2564 pages = obj_request->copyup_pages;
2565 rbd_assert(pages != NULL);
2566 obj_request->copyup_pages = NULL;
2567 page_count = obj_request->copyup_page_count;
2568 rbd_assert(page_count);
2569 obj_request->copyup_page_count = 0;
2570 ceph_release_page_vector(pages, page_count);
2573 * We want the transfer count to reflect the size of the
2574 * original write request. There is no such thing as a
2575 * successful short write, so if the request was successful
2576 * we can just set it to the originally-requested length.
2578 if (!obj_request->result)
2579 obj_request->xferred = obj_request->length;
2581 obj_request_done_set(obj_request);
2585 rbd_img_obj_parent_read_full_callback(struct rbd_img_request *img_request)
2587 struct rbd_obj_request *orig_request;
2588 struct ceph_osd_request *osd_req;
2589 struct ceph_osd_client *osdc;
2590 struct rbd_device *rbd_dev;
2591 struct page **pages;
2592 enum obj_operation_type op_type;
2597 rbd_assert(img_request_child_test(img_request));
2599 /* First get what we need from the image request */
2601 pages = img_request->copyup_pages;
2602 rbd_assert(pages != NULL);
2603 img_request->copyup_pages = NULL;
2604 page_count = img_request->copyup_page_count;
2605 rbd_assert(page_count);
2606 img_request->copyup_page_count = 0;
2608 orig_request = img_request->obj_request;
2609 rbd_assert(orig_request != NULL);
2610 rbd_assert(obj_request_type_valid(orig_request->type));
2611 img_result = img_request->result;
2612 parent_length = img_request->length;
2613 rbd_assert(parent_length == img_request->xferred);
2614 rbd_img_request_put(img_request);
2616 rbd_assert(orig_request->img_request);
2617 rbd_dev = orig_request->img_request->rbd_dev;
2618 rbd_assert(rbd_dev);
2621 * If the overlap has become 0 (most likely because the
2622 * image has been flattened) we need to free the pages
2623 * and re-submit the original write request.
2625 if (!rbd_dev->parent_overlap) {
2626 struct ceph_osd_client *osdc;
2628 ceph_release_page_vector(pages, page_count);
2629 osdc = &rbd_dev->rbd_client->client->osdc;
2630 img_result = rbd_obj_request_submit(osdc, orig_request);
2639 * The original osd request is of no use to use any more.
2640 * We need a new one that can hold the three ops in a copyup
2641 * request. Allocate the new copyup osd request for the
2642 * original request, and release the old one.
2644 img_result = -ENOMEM;
2645 osd_req = rbd_osd_req_create_copyup(orig_request);
2648 rbd_osd_req_destroy(orig_request->osd_req);
2649 orig_request->osd_req = osd_req;
2650 orig_request->copyup_pages = pages;
2651 orig_request->copyup_page_count = page_count;
2653 /* Initialize the copyup op */
2655 osd_req_op_cls_init(osd_req, 0, CEPH_OSD_OP_CALL, "rbd", "copyup");
2656 osd_req_op_cls_request_data_pages(osd_req, 0, pages, parent_length, 0,
2659 /* Add the other op(s) */
2661 op_type = rbd_img_request_op_type(orig_request->img_request);
2662 rbd_img_obj_request_fill(orig_request, osd_req, op_type, 1);
2664 /* All set, send it off. */
2666 osdc = &rbd_dev->rbd_client->client->osdc;
2667 img_result = rbd_obj_request_submit(osdc, orig_request);
2671 /* Record the error code and complete the request */
2673 orig_request->result = img_result;
2674 orig_request->xferred = 0;
2675 obj_request_done_set(orig_request);
2676 rbd_obj_request_complete(orig_request);
2680 * Read from the parent image the range of data that covers the
2681 * entire target of the given object request. This is used for
2682 * satisfying a layered image write request when the target of an
2683 * object request from the image request does not exist.
2685 * A page array big enough to hold the returned data is allocated
2686 * and supplied to rbd_img_request_fill() as the "data descriptor."
2687 * When the read completes, this page array will be transferred to
2688 * the original object request for the copyup operation.
2690 * If an error occurs, record it as the result of the original
2691 * object request and mark it done so it gets completed.
2693 static int rbd_img_obj_parent_read_full(struct rbd_obj_request *obj_request)
2695 struct rbd_img_request *img_request = NULL;
2696 struct rbd_img_request *parent_request = NULL;
2697 struct rbd_device *rbd_dev;
2700 struct page **pages = NULL;
2704 rbd_assert(obj_request_img_data_test(obj_request));
2705 rbd_assert(obj_request_type_valid(obj_request->type));
2707 img_request = obj_request->img_request;
2708 rbd_assert(img_request != NULL);
2709 rbd_dev = img_request->rbd_dev;
2710 rbd_assert(rbd_dev->parent != NULL);
2713 * Determine the byte range covered by the object in the
2714 * child image to which the original request was to be sent.
2716 img_offset = obj_request->img_offset - obj_request->offset;
2717 length = (u64)1 << rbd_dev->header.obj_order;
2720 * There is no defined parent data beyond the parent
2721 * overlap, so limit what we read at that boundary if
2724 if (img_offset + length > rbd_dev->parent_overlap) {
2725 rbd_assert(img_offset < rbd_dev->parent_overlap);
2726 length = rbd_dev->parent_overlap - img_offset;
2730 * Allocate a page array big enough to receive the data read
2733 page_count = (u32)calc_pages_for(0, length);
2734 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2735 if (IS_ERR(pages)) {
2736 result = PTR_ERR(pages);
2742 parent_request = rbd_parent_request_create(obj_request,
2743 img_offset, length);
2744 if (!parent_request)
2747 result = rbd_img_request_fill(parent_request, OBJ_REQUEST_PAGES, pages);
2750 parent_request->copyup_pages = pages;
2751 parent_request->copyup_page_count = page_count;
2753 parent_request->callback = rbd_img_obj_parent_read_full_callback;
2754 result = rbd_img_request_submit(parent_request);
2758 parent_request->copyup_pages = NULL;
2759 parent_request->copyup_page_count = 0;
2760 parent_request->obj_request = NULL;
2761 rbd_obj_request_put(obj_request);
2764 ceph_release_page_vector(pages, page_count);
2766 rbd_img_request_put(parent_request);
2767 obj_request->result = result;
2768 obj_request->xferred = 0;
2769 obj_request_done_set(obj_request);
2774 static void rbd_img_obj_exists_callback(struct rbd_obj_request *obj_request)
2776 struct rbd_obj_request *orig_request;
2777 struct rbd_device *rbd_dev;
2780 rbd_assert(!obj_request_img_data_test(obj_request));
2783 * All we need from the object request is the original
2784 * request and the result of the STAT op. Grab those, then
2785 * we're done with the request.
2787 orig_request = obj_request->obj_request;
2788 obj_request->obj_request = NULL;
2789 rbd_obj_request_put(orig_request);
2790 rbd_assert(orig_request);
2791 rbd_assert(orig_request->img_request);
2793 result = obj_request->result;
2794 obj_request->result = 0;
2796 dout("%s: obj %p for obj %p result %d %llu/%llu\n", __func__,
2797 obj_request, orig_request, result,
2798 obj_request->xferred, obj_request->length);
2799 rbd_obj_request_put(obj_request);
2802 * If the overlap has become 0 (most likely because the
2803 * image has been flattened) we need to free the pages
2804 * and re-submit the original write request.
2806 rbd_dev = orig_request->img_request->rbd_dev;
2807 if (!rbd_dev->parent_overlap) {
2808 struct ceph_osd_client *osdc;
2810 osdc = &rbd_dev->rbd_client->client->osdc;
2811 result = rbd_obj_request_submit(osdc, orig_request);
2817 * Our only purpose here is to determine whether the object
2818 * exists, and we don't want to treat the non-existence as
2819 * an error. If something else comes back, transfer the
2820 * error to the original request and complete it now.
2823 obj_request_existence_set(orig_request, true);
2824 } else if (result == -ENOENT) {
2825 obj_request_existence_set(orig_request, false);
2826 } else if (result) {
2827 orig_request->result = result;
2832 * Resubmit the original request now that we have recorded
2833 * whether the target object exists.
2835 orig_request->result = rbd_img_obj_request_submit(orig_request);
2837 if (orig_request->result)
2838 rbd_obj_request_complete(orig_request);
2841 static int rbd_img_obj_exists_submit(struct rbd_obj_request *obj_request)
2843 struct rbd_obj_request *stat_request;
2844 struct rbd_device *rbd_dev;
2845 struct ceph_osd_client *osdc;
2846 struct page **pages = NULL;
2852 * The response data for a STAT call consists of:
2859 size = sizeof (__le64) + sizeof (__le32) + sizeof (__le32);
2860 page_count = (u32)calc_pages_for(0, size);
2861 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2863 return PTR_ERR(pages);
2866 stat_request = rbd_obj_request_create(obj_request->object_name, 0, 0,
2871 rbd_obj_request_get(obj_request);
2872 stat_request->obj_request = obj_request;
2873 stat_request->pages = pages;
2874 stat_request->page_count = page_count;
2876 rbd_assert(obj_request->img_request);
2877 rbd_dev = obj_request->img_request->rbd_dev;
2878 stat_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
2880 if (!stat_request->osd_req)
2882 stat_request->callback = rbd_img_obj_exists_callback;
2884 osd_req_op_init(stat_request->osd_req, 0, CEPH_OSD_OP_STAT, 0);
2885 osd_req_op_raw_data_in_pages(stat_request->osd_req, 0, pages, size, 0,
2887 rbd_osd_req_format_read(stat_request);
2889 osdc = &rbd_dev->rbd_client->client->osdc;
2890 ret = rbd_obj_request_submit(osdc, stat_request);
2893 rbd_obj_request_put(obj_request);
2898 static bool img_obj_request_simple(struct rbd_obj_request *obj_request)
2900 struct rbd_img_request *img_request;
2901 struct rbd_device *rbd_dev;
2903 rbd_assert(obj_request_img_data_test(obj_request));
2905 img_request = obj_request->img_request;
2906 rbd_assert(img_request);
2907 rbd_dev = img_request->rbd_dev;
2910 if (!img_request_write_test(img_request) &&
2911 !img_request_discard_test(img_request))
2914 /* Non-layered writes */
2915 if (!img_request_layered_test(img_request))
2919 * Layered writes outside of the parent overlap range don't
2920 * share any data with the parent.
2922 if (!obj_request_overlaps_parent(obj_request))
2926 * Entire-object layered writes - we will overwrite whatever
2927 * parent data there is anyway.
2929 if (!obj_request->offset &&
2930 obj_request->length == rbd_obj_bytes(&rbd_dev->header))
2934 * If the object is known to already exist, its parent data has
2935 * already been copied.
2937 if (obj_request_known_test(obj_request) &&
2938 obj_request_exists_test(obj_request))
2944 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request)
2946 if (img_obj_request_simple(obj_request)) {
2947 struct rbd_device *rbd_dev;
2948 struct ceph_osd_client *osdc;
2950 rbd_dev = obj_request->img_request->rbd_dev;
2951 osdc = &rbd_dev->rbd_client->client->osdc;
2953 return rbd_obj_request_submit(osdc, obj_request);
2957 * It's a layered write. The target object might exist but
2958 * we may not know that yet. If we know it doesn't exist,
2959 * start by reading the data for the full target object from
2960 * the parent so we can use it for a copyup to the target.
2962 if (obj_request_known_test(obj_request))
2963 return rbd_img_obj_parent_read_full(obj_request);
2965 /* We don't know whether the target exists. Go find out. */
2967 return rbd_img_obj_exists_submit(obj_request);
2970 static int rbd_img_request_submit(struct rbd_img_request *img_request)
2972 struct rbd_obj_request *obj_request;
2973 struct rbd_obj_request *next_obj_request;
2975 dout("%s: img %p\n", __func__, img_request);
2976 for_each_obj_request_safe(img_request, obj_request, next_obj_request) {
2979 ret = rbd_img_obj_request_submit(obj_request);
2987 static void rbd_img_parent_read_callback(struct rbd_img_request *img_request)
2989 struct rbd_obj_request *obj_request;
2990 struct rbd_device *rbd_dev;
2995 rbd_assert(img_request_child_test(img_request));
2997 /* First get what we need from the image request and release it */
2999 obj_request = img_request->obj_request;
3000 img_xferred = img_request->xferred;
3001 img_result = img_request->result;
3002 rbd_img_request_put(img_request);
3005 * If the overlap has become 0 (most likely because the
3006 * image has been flattened) we need to re-submit the
3009 rbd_assert(obj_request);
3010 rbd_assert(obj_request->img_request);
3011 rbd_dev = obj_request->img_request->rbd_dev;
3012 if (!rbd_dev->parent_overlap) {
3013 struct ceph_osd_client *osdc;
3015 osdc = &rbd_dev->rbd_client->client->osdc;
3016 img_result = rbd_obj_request_submit(osdc, obj_request);
3021 obj_request->result = img_result;
3022 if (obj_request->result)
3026 * We need to zero anything beyond the parent overlap
3027 * boundary. Since rbd_img_obj_request_read_callback()
3028 * will zero anything beyond the end of a short read, an
3029 * easy way to do this is to pretend the data from the
3030 * parent came up short--ending at the overlap boundary.
3032 rbd_assert(obj_request->img_offset < U64_MAX - obj_request->length);
3033 obj_end = obj_request->img_offset + obj_request->length;
3034 if (obj_end > rbd_dev->parent_overlap) {
3037 if (obj_request->img_offset < rbd_dev->parent_overlap)
3038 xferred = rbd_dev->parent_overlap -
3039 obj_request->img_offset;
3041 obj_request->xferred = min(img_xferred, xferred);
3043 obj_request->xferred = img_xferred;
3046 rbd_img_obj_request_read_callback(obj_request);
3047 rbd_obj_request_complete(obj_request);
3050 static void rbd_img_parent_read(struct rbd_obj_request *obj_request)
3052 struct rbd_img_request *img_request;
3055 rbd_assert(obj_request_img_data_test(obj_request));
3056 rbd_assert(obj_request->img_request != NULL);
3057 rbd_assert(obj_request->result == (s32) -ENOENT);
3058 rbd_assert(obj_request_type_valid(obj_request->type));
3060 /* rbd_read_finish(obj_request, obj_request->length); */
3061 img_request = rbd_parent_request_create(obj_request,
3062 obj_request->img_offset,
3063 obj_request->length);
3068 if (obj_request->type == OBJ_REQUEST_BIO)
3069 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3070 obj_request->bio_list);
3072 result = rbd_img_request_fill(img_request, OBJ_REQUEST_PAGES,
3073 obj_request->pages);
3077 img_request->callback = rbd_img_parent_read_callback;
3078 result = rbd_img_request_submit(img_request);
3085 rbd_img_request_put(img_request);
3086 obj_request->result = result;
3087 obj_request->xferred = 0;
3088 obj_request_done_set(obj_request);
3091 static int rbd_obj_notify_ack_sync(struct rbd_device *rbd_dev, u64 notify_id)
3093 struct rbd_obj_request *obj_request;
3094 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3097 obj_request = rbd_obj_request_create(rbd_dev->header_name, 0, 0,
3098 OBJ_REQUEST_NODATA);
3103 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
3105 if (!obj_request->osd_req)
3108 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_NOTIFY_ACK,
3110 rbd_osd_req_format_read(obj_request);
3112 ret = rbd_obj_request_submit(osdc, obj_request);
3115 ret = rbd_obj_request_wait(obj_request);
3117 rbd_obj_request_put(obj_request);
3122 static void rbd_watch_cb(u64 ver, u64 notify_id, u8 opcode, void *data)
3124 struct rbd_device *rbd_dev = (struct rbd_device *)data;
3130 dout("%s: \"%s\" notify_id %llu opcode %u\n", __func__,
3131 rbd_dev->header_name, (unsigned long long)notify_id,
3132 (unsigned int)opcode);
3135 * Until adequate refresh error handling is in place, there is
3136 * not much we can do here, except warn.
3138 * See http://tracker.ceph.com/issues/5040
3140 ret = rbd_dev_refresh(rbd_dev);
3142 rbd_warn(rbd_dev, "refresh failed: %d", ret);
3144 ret = rbd_obj_notify_ack_sync(rbd_dev, notify_id);
3146 rbd_warn(rbd_dev, "notify_ack ret %d", ret);
3150 * Send a (un)watch request and wait for the ack. Return a request
3151 * with a ref held on success or error.
3153 static struct rbd_obj_request *rbd_obj_watch_request_helper(
3154 struct rbd_device *rbd_dev,
3157 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3158 struct ceph_options *opts = osdc->client->options;
3159 struct rbd_obj_request *obj_request;
3162 obj_request = rbd_obj_request_create(rbd_dev->header_name, 0, 0,
3163 OBJ_REQUEST_NODATA);
3165 return ERR_PTR(-ENOMEM);
3167 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_WRITE, 1,
3169 if (!obj_request->osd_req) {
3174 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_WATCH,
3175 rbd_dev->watch_event->cookie, 0, watch);
3176 rbd_osd_req_format_write(obj_request);
3179 ceph_osdc_set_request_linger(osdc, obj_request->osd_req);
3181 ret = rbd_obj_request_submit(osdc, obj_request);
3185 ret = rbd_obj_request_wait_timeout(obj_request, opts->mount_timeout);
3189 ret = obj_request->result;
3192 rbd_obj_request_end(obj_request);
3199 rbd_obj_request_put(obj_request);
3200 return ERR_PTR(ret);
3204 * Initiate a watch request, synchronously.
3206 static int rbd_dev_header_watch_sync(struct rbd_device *rbd_dev)
3208 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3209 struct rbd_obj_request *obj_request;
3212 rbd_assert(!rbd_dev->watch_event);
3213 rbd_assert(!rbd_dev->watch_request);
3215 ret = ceph_osdc_create_event(osdc, rbd_watch_cb, rbd_dev,
3216 &rbd_dev->watch_event);
3220 obj_request = rbd_obj_watch_request_helper(rbd_dev, true);
3221 if (IS_ERR(obj_request)) {
3222 ceph_osdc_cancel_event(rbd_dev->watch_event);
3223 rbd_dev->watch_event = NULL;
3224 return PTR_ERR(obj_request);
3228 * A watch request is set to linger, so the underlying osd
3229 * request won't go away until we unregister it. We retain
3230 * a pointer to the object request during that time (in
3231 * rbd_dev->watch_request), so we'll keep a reference to it.
3232 * We'll drop that reference after we've unregistered it in
3233 * rbd_dev_header_unwatch_sync().
3235 rbd_dev->watch_request = obj_request;
3241 * Tear down a watch request, synchronously.
3243 static void rbd_dev_header_unwatch_sync(struct rbd_device *rbd_dev)
3245 struct rbd_obj_request *obj_request;
3247 rbd_assert(rbd_dev->watch_event);
3248 rbd_assert(rbd_dev->watch_request);
3250 rbd_obj_request_end(rbd_dev->watch_request);
3251 rbd_obj_request_put(rbd_dev->watch_request);
3252 rbd_dev->watch_request = NULL;
3254 obj_request = rbd_obj_watch_request_helper(rbd_dev, false);
3255 if (!IS_ERR(obj_request))
3256 rbd_obj_request_put(obj_request);
3258 rbd_warn(rbd_dev, "unable to tear down watch request (%ld)",
3259 PTR_ERR(obj_request));
3261 ceph_osdc_cancel_event(rbd_dev->watch_event);
3262 rbd_dev->watch_event = NULL;
3266 * Synchronous osd object method call. Returns the number of bytes
3267 * returned in the outbound buffer, or a negative error code.
3269 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
3270 const char *object_name,
3271 const char *class_name,
3272 const char *method_name,
3273 const void *outbound,
3274 size_t outbound_size,
3276 size_t inbound_size)
3278 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3279 struct rbd_obj_request *obj_request;
3280 struct page **pages;
3285 * Method calls are ultimately read operations. The result
3286 * should placed into the inbound buffer provided. They
3287 * also supply outbound data--parameters for the object
3288 * method. Currently if this is present it will be a
3291 page_count = (u32)calc_pages_for(0, inbound_size);
3292 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
3294 return PTR_ERR(pages);
3297 obj_request = rbd_obj_request_create(object_name, 0, inbound_size,
3302 obj_request->pages = pages;
3303 obj_request->page_count = page_count;
3305 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
3307 if (!obj_request->osd_req)
3310 osd_req_op_cls_init(obj_request->osd_req, 0, CEPH_OSD_OP_CALL,
3311 class_name, method_name);
3312 if (outbound_size) {
3313 struct ceph_pagelist *pagelist;
3315 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
3319 ceph_pagelist_init(pagelist);
3320 ceph_pagelist_append(pagelist, outbound, outbound_size);
3321 osd_req_op_cls_request_data_pagelist(obj_request->osd_req, 0,
3324 osd_req_op_cls_response_data_pages(obj_request->osd_req, 0,
3325 obj_request->pages, inbound_size,
3327 rbd_osd_req_format_read(obj_request);
3329 ret = rbd_obj_request_submit(osdc, obj_request);
3332 ret = rbd_obj_request_wait(obj_request);
3336 ret = obj_request->result;
3340 rbd_assert(obj_request->xferred < (u64)INT_MAX);
3341 ret = (int)obj_request->xferred;
3342 ceph_copy_from_page_vector(pages, inbound, 0, obj_request->xferred);
3345 rbd_obj_request_put(obj_request);
3347 ceph_release_page_vector(pages, page_count);
3352 static void rbd_queue_workfn(struct work_struct *work)
3354 struct request *rq = blk_mq_rq_from_pdu(work);
3355 struct rbd_device *rbd_dev = rq->q->queuedata;
3356 struct rbd_img_request *img_request;
3357 struct ceph_snap_context *snapc = NULL;
3358 u64 offset = (u64)blk_rq_pos(rq) << SECTOR_SHIFT;
3359 u64 length = blk_rq_bytes(rq);
3360 enum obj_operation_type op_type;
3364 if (rq->cmd_type != REQ_TYPE_FS) {
3365 dout("%s: non-fs request type %d\n", __func__,
3366 (int) rq->cmd_type);
3371 if (rq->cmd_flags & REQ_DISCARD)
3372 op_type = OBJ_OP_DISCARD;
3373 else if (rq->cmd_flags & REQ_WRITE)
3374 op_type = OBJ_OP_WRITE;
3376 op_type = OBJ_OP_READ;
3378 /* Ignore/skip any zero-length requests */
3381 dout("%s: zero-length request\n", __func__);
3386 /* Only reads are allowed to a read-only device */
3388 if (op_type != OBJ_OP_READ) {
3389 if (rbd_dev->mapping.read_only) {
3393 rbd_assert(rbd_dev->spec->snap_id == CEPH_NOSNAP);
3397 * Quit early if the mapped snapshot no longer exists. It's
3398 * still possible the snapshot will have disappeared by the
3399 * time our request arrives at the osd, but there's no sense in
3400 * sending it if we already know.
3402 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
3403 dout("request for non-existent snapshot");
3404 rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
3409 if (offset && length > U64_MAX - offset + 1) {
3410 rbd_warn(rbd_dev, "bad request range (%llu~%llu)", offset,
3413 goto err_rq; /* Shouldn't happen */
3416 blk_mq_start_request(rq);
3418 down_read(&rbd_dev->header_rwsem);
3419 mapping_size = rbd_dev->mapping.size;
3420 if (op_type != OBJ_OP_READ) {
3421 snapc = rbd_dev->header.snapc;
3422 ceph_get_snap_context(snapc);
3424 up_read(&rbd_dev->header_rwsem);
3426 if (offset + length > mapping_size) {
3427 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)", offset,
3428 length, mapping_size);
3433 img_request = rbd_img_request_create(rbd_dev, offset, length, op_type,
3439 img_request->rq = rq;
3441 if (op_type == OBJ_OP_DISCARD)
3442 result = rbd_img_request_fill(img_request, OBJ_REQUEST_NODATA,
3445 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3448 goto err_img_request;
3450 result = rbd_img_request_submit(img_request);
3452 goto err_img_request;
3457 rbd_img_request_put(img_request);
3460 rbd_warn(rbd_dev, "%s %llx at %llx result %d",
3461 obj_op_name(op_type), length, offset, result);
3462 ceph_put_snap_context(snapc);
3464 blk_mq_end_request(rq, result);
3467 static int rbd_queue_rq(struct blk_mq_hw_ctx *hctx,
3468 const struct blk_mq_queue_data *bd)
3470 struct request *rq = bd->rq;
3471 struct work_struct *work = blk_mq_rq_to_pdu(rq);
3473 queue_work(rbd_wq, work);
3474 return BLK_MQ_RQ_QUEUE_OK;
3478 * a queue callback. Makes sure that we don't create a bio that spans across
3479 * multiple osd objects. One exception would be with a single page bios,
3480 * which we handle later at bio_chain_clone_range()
3482 static int rbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bmd,
3483 struct bio_vec *bvec)
3485 struct rbd_device *rbd_dev = q->queuedata;
3486 sector_t sector_offset;
3487 sector_t sectors_per_obj;
3488 sector_t obj_sector_offset;
3492 * Find how far into its rbd object the partition-relative
3493 * bio start sector is to offset relative to the enclosing
3496 sector_offset = get_start_sect(bmd->bi_bdev) + bmd->bi_sector;
3497 sectors_per_obj = 1 << (rbd_dev->header.obj_order - SECTOR_SHIFT);
3498 obj_sector_offset = sector_offset & (sectors_per_obj - 1);
3501 * Compute the number of bytes from that offset to the end
3502 * of the object. Account for what's already used by the bio.
3504 ret = (int) (sectors_per_obj - obj_sector_offset) << SECTOR_SHIFT;
3505 if (ret > bmd->bi_size)
3506 ret -= bmd->bi_size;
3511 * Don't send back more than was asked for. And if the bio
3512 * was empty, let the whole thing through because: "Note
3513 * that a block device *must* allow a single page to be
3514 * added to an empty bio."
3516 rbd_assert(bvec->bv_len <= PAGE_SIZE);
3517 if (ret > (int) bvec->bv_len || !bmd->bi_size)
3518 ret = (int) bvec->bv_len;
3523 static void rbd_free_disk(struct rbd_device *rbd_dev)
3525 struct gendisk *disk = rbd_dev->disk;
3530 rbd_dev->disk = NULL;
3531 if (disk->flags & GENHD_FL_UP) {
3534 blk_cleanup_queue(disk->queue);
3535 blk_mq_free_tag_set(&rbd_dev->tag_set);
3540 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
3541 const char *object_name,
3542 u64 offset, u64 length, void *buf)
3545 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3546 struct rbd_obj_request *obj_request;
3547 struct page **pages = NULL;
3552 page_count = (u32) calc_pages_for(offset, length);
3553 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
3555 return PTR_ERR(pages);
3558 obj_request = rbd_obj_request_create(object_name, offset, length,
3563 obj_request->pages = pages;
3564 obj_request->page_count = page_count;
3566 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
3568 if (!obj_request->osd_req)
3571 osd_req_op_extent_init(obj_request->osd_req, 0, CEPH_OSD_OP_READ,
3572 offset, length, 0, 0);
3573 osd_req_op_extent_osd_data_pages(obj_request->osd_req, 0,
3575 obj_request->length,
3576 obj_request->offset & ~PAGE_MASK,
3578 rbd_osd_req_format_read(obj_request);
3580 ret = rbd_obj_request_submit(osdc, obj_request);
3583 ret = rbd_obj_request_wait(obj_request);
3587 ret = obj_request->result;
3591 rbd_assert(obj_request->xferred <= (u64) SIZE_MAX);
3592 size = (size_t) obj_request->xferred;
3593 ceph_copy_from_page_vector(pages, buf, 0, size);
3594 rbd_assert(size <= (size_t)INT_MAX);
3598 rbd_obj_request_put(obj_request);
3600 ceph_release_page_vector(pages, page_count);
3606 * Read the complete header for the given rbd device. On successful
3607 * return, the rbd_dev->header field will contain up-to-date
3608 * information about the image.
3610 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev)
3612 struct rbd_image_header_ondisk *ondisk = NULL;
3619 * The complete header will include an array of its 64-bit
3620 * snapshot ids, followed by the names of those snapshots as
3621 * a contiguous block of NUL-terminated strings. Note that
3622 * the number of snapshots could change by the time we read
3623 * it in, in which case we re-read it.
3630 size = sizeof (*ondisk);
3631 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
3633 ondisk = kmalloc(size, GFP_KERNEL);
3637 ret = rbd_obj_read_sync(rbd_dev, rbd_dev->header_name,
3641 if ((size_t)ret < size) {
3643 rbd_warn(rbd_dev, "short header read (want %zd got %d)",
3647 if (!rbd_dev_ondisk_valid(ondisk)) {
3649 rbd_warn(rbd_dev, "invalid header");
3653 names_size = le64_to_cpu(ondisk->snap_names_len);
3654 want_count = snap_count;
3655 snap_count = le32_to_cpu(ondisk->snap_count);
3656 } while (snap_count != want_count);
3658 ret = rbd_header_from_disk(rbd_dev, ondisk);
3666 * Clear the rbd device's EXISTS flag if the snapshot it's mapped to
3667 * has disappeared from the (just updated) snapshot context.
3669 static void rbd_exists_validate(struct rbd_device *rbd_dev)
3673 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags))
3676 snap_id = rbd_dev->spec->snap_id;
3677 if (snap_id == CEPH_NOSNAP)
3680 if (rbd_dev_snap_index(rbd_dev, snap_id) == BAD_SNAP_INDEX)
3681 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
3684 static void rbd_dev_update_size(struct rbd_device *rbd_dev)
3690 * Don't hold the lock while doing disk operations,
3691 * or lock ordering will conflict with the bdev mutex via:
3692 * rbd_add() -> blkdev_get() -> rbd_open()
3694 spin_lock_irq(&rbd_dev->lock);
3695 removing = test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags);
3696 spin_unlock_irq(&rbd_dev->lock);
3698 * If the device is being removed, rbd_dev->disk has
3699 * been destroyed, so don't try to update its size
3702 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
3703 dout("setting size to %llu sectors", (unsigned long long)size);
3704 set_capacity(rbd_dev->disk, size);
3705 revalidate_disk(rbd_dev->disk);
3709 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
3714 down_write(&rbd_dev->header_rwsem);
3715 mapping_size = rbd_dev->mapping.size;
3717 ret = rbd_dev_header_info(rbd_dev);
3722 * If there is a parent, see if it has disappeared due to the
3723 * mapped image getting flattened.
3725 if (rbd_dev->parent) {
3726 ret = rbd_dev_v2_parent_info(rbd_dev);
3731 if (rbd_dev->spec->snap_id == CEPH_NOSNAP) {
3732 rbd_dev->mapping.size = rbd_dev->header.image_size;
3734 /* validate mapped snapshot's EXISTS flag */
3735 rbd_exists_validate(rbd_dev);
3739 up_write(&rbd_dev->header_rwsem);
3740 if (!ret && mapping_size != rbd_dev->mapping.size)
3741 rbd_dev_update_size(rbd_dev);
3746 static int rbd_init_request(void *data, struct request *rq,
3747 unsigned int hctx_idx, unsigned int request_idx,
3748 unsigned int numa_node)
3750 struct work_struct *work = blk_mq_rq_to_pdu(rq);
3752 INIT_WORK(work, rbd_queue_workfn);
3756 static struct blk_mq_ops rbd_mq_ops = {
3757 .queue_rq = rbd_queue_rq,
3758 .map_queue = blk_mq_map_queue,
3759 .init_request = rbd_init_request,
3762 static int rbd_init_disk(struct rbd_device *rbd_dev)
3764 struct gendisk *disk;
3765 struct request_queue *q;
3769 /* create gendisk info */
3770 disk = alloc_disk(single_major ?
3771 (1 << RBD_SINGLE_MAJOR_PART_SHIFT) :
3772 RBD_MINORS_PER_MAJOR);
3776 snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
3778 disk->major = rbd_dev->major;
3779 disk->first_minor = rbd_dev->minor;
3781 disk->flags |= GENHD_FL_EXT_DEVT;
3782 disk->fops = &rbd_bd_ops;
3783 disk->private_data = rbd_dev;
3785 memset(&rbd_dev->tag_set, 0, sizeof(rbd_dev->tag_set));
3786 rbd_dev->tag_set.ops = &rbd_mq_ops;
3787 rbd_dev->tag_set.queue_depth = rbd_dev->opts->queue_depth;
3788 rbd_dev->tag_set.numa_node = NUMA_NO_NODE;
3789 rbd_dev->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
3790 rbd_dev->tag_set.nr_hw_queues = 1;
3791 rbd_dev->tag_set.cmd_size = sizeof(struct work_struct);
3793 err = blk_mq_alloc_tag_set(&rbd_dev->tag_set);
3797 q = blk_mq_init_queue(&rbd_dev->tag_set);
3803 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
3804 /* QUEUE_FLAG_ADD_RANDOM is off by default for blk-mq */
3806 /* set io sizes to object size */
3807 segment_size = rbd_obj_bytes(&rbd_dev->header);
3808 blk_queue_max_hw_sectors(q, segment_size / SECTOR_SIZE);
3809 blk_queue_max_segments(q, segment_size / SECTOR_SIZE);
3810 blk_queue_max_segment_size(q, segment_size);
3811 blk_queue_io_min(q, segment_size);
3812 blk_queue_io_opt(q, segment_size);
3814 /* enable the discard support */
3815 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
3816 q->limits.discard_granularity = segment_size;
3817 q->limits.discard_alignment = segment_size;
3818 q->limits.max_discard_sectors = segment_size / SECTOR_SIZE;
3819 q->limits.discard_zeroes_data = 1;
3821 blk_queue_merge_bvec(q, rbd_merge_bvec);
3824 q->queuedata = rbd_dev;
3826 rbd_dev->disk = disk;
3830 blk_mq_free_tag_set(&rbd_dev->tag_set);
3840 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
3842 return container_of(dev, struct rbd_device, dev);
3845 static ssize_t rbd_size_show(struct device *dev,
3846 struct device_attribute *attr, char *buf)
3848 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3850 return sprintf(buf, "%llu\n",
3851 (unsigned long long)rbd_dev->mapping.size);
3855 * Note this shows the features for whatever's mapped, which is not
3856 * necessarily the base image.
3858 static ssize_t rbd_features_show(struct device *dev,
3859 struct device_attribute *attr, char *buf)
3861 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3863 return sprintf(buf, "0x%016llx\n",
3864 (unsigned long long)rbd_dev->mapping.features);
3867 static ssize_t rbd_major_show(struct device *dev,
3868 struct device_attribute *attr, char *buf)
3870 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3873 return sprintf(buf, "%d\n", rbd_dev->major);
3875 return sprintf(buf, "(none)\n");
3878 static ssize_t rbd_minor_show(struct device *dev,
3879 struct device_attribute *attr, char *buf)
3881 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3883 return sprintf(buf, "%d\n", rbd_dev->minor);
3886 static ssize_t rbd_client_id_show(struct device *dev,
3887 struct device_attribute *attr, char *buf)
3889 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3891 return sprintf(buf, "client%lld\n",
3892 ceph_client_id(rbd_dev->rbd_client->client));
3895 static ssize_t rbd_pool_show(struct device *dev,
3896 struct device_attribute *attr, char *buf)
3898 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3900 return sprintf(buf, "%s\n", rbd_dev->spec->pool_name);
3903 static ssize_t rbd_pool_id_show(struct device *dev,
3904 struct device_attribute *attr, char *buf)
3906 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3908 return sprintf(buf, "%llu\n",
3909 (unsigned long long) rbd_dev->spec->pool_id);
3912 static ssize_t rbd_name_show(struct device *dev,
3913 struct device_attribute *attr, char *buf)
3915 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3917 if (rbd_dev->spec->image_name)
3918 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
3920 return sprintf(buf, "(unknown)\n");
3923 static ssize_t rbd_image_id_show(struct device *dev,
3924 struct device_attribute *attr, char *buf)
3926 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3928 return sprintf(buf, "%s\n", rbd_dev->spec->image_id);
3932 * Shows the name of the currently-mapped snapshot (or
3933 * RBD_SNAP_HEAD_NAME for the base image).
3935 static ssize_t rbd_snap_show(struct device *dev,
3936 struct device_attribute *attr,
3939 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3941 return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
3945 * For a v2 image, shows the chain of parent images, separated by empty
3946 * lines. For v1 images or if there is no parent, shows "(no parent
3949 static ssize_t rbd_parent_show(struct device *dev,
3950 struct device_attribute *attr,
3953 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3956 if (!rbd_dev->parent)
3957 return sprintf(buf, "(no parent image)\n");
3959 for ( ; rbd_dev->parent; rbd_dev = rbd_dev->parent) {
3960 struct rbd_spec *spec = rbd_dev->parent_spec;
3962 count += sprintf(&buf[count], "%s"
3963 "pool_id %llu\npool_name %s\n"
3964 "image_id %s\nimage_name %s\n"
3965 "snap_id %llu\nsnap_name %s\n"
3967 !count ? "" : "\n", /* first? */
3968 spec->pool_id, spec->pool_name,
3969 spec->image_id, spec->image_name ?: "(unknown)",
3970 spec->snap_id, spec->snap_name,
3971 rbd_dev->parent_overlap);
3977 static ssize_t rbd_image_refresh(struct device *dev,
3978 struct device_attribute *attr,
3982 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3985 ret = rbd_dev_refresh(rbd_dev);
3992 static DEVICE_ATTR(size, S_IRUGO, rbd_size_show, NULL);
3993 static DEVICE_ATTR(features, S_IRUGO, rbd_features_show, NULL);
3994 static DEVICE_ATTR(major, S_IRUGO, rbd_major_show, NULL);
3995 static DEVICE_ATTR(minor, S_IRUGO, rbd_minor_show, NULL);
3996 static DEVICE_ATTR(client_id, S_IRUGO, rbd_client_id_show, NULL);
3997 static DEVICE_ATTR(pool, S_IRUGO, rbd_pool_show, NULL);
3998 static DEVICE_ATTR(pool_id, S_IRUGO, rbd_pool_id_show, NULL);
3999 static DEVICE_ATTR(name, S_IRUGO, rbd_name_show, NULL);
4000 static DEVICE_ATTR(image_id, S_IRUGO, rbd_image_id_show, NULL);
4001 static DEVICE_ATTR(refresh, S_IWUSR, NULL, rbd_image_refresh);
4002 static DEVICE_ATTR(current_snap, S_IRUGO, rbd_snap_show, NULL);
4003 static DEVICE_ATTR(parent, S_IRUGO, rbd_parent_show, NULL);
4005 static struct attribute *rbd_attrs[] = {
4006 &dev_attr_size.attr,
4007 &dev_attr_features.attr,
4008 &dev_attr_major.attr,
4009 &dev_attr_minor.attr,
4010 &dev_attr_client_id.attr,
4011 &dev_attr_pool.attr,
4012 &dev_attr_pool_id.attr,
4013 &dev_attr_name.attr,
4014 &dev_attr_image_id.attr,
4015 &dev_attr_current_snap.attr,
4016 &dev_attr_parent.attr,
4017 &dev_attr_refresh.attr,
4021 static struct attribute_group rbd_attr_group = {
4025 static const struct attribute_group *rbd_attr_groups[] = {
4030 static void rbd_sysfs_dev_release(struct device *dev)
4034 static struct device_type rbd_device_type = {
4036 .groups = rbd_attr_groups,
4037 .release = rbd_sysfs_dev_release,
4040 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
4042 kref_get(&spec->kref);
4047 static void rbd_spec_free(struct kref *kref);
4048 static void rbd_spec_put(struct rbd_spec *spec)
4051 kref_put(&spec->kref, rbd_spec_free);
4054 static struct rbd_spec *rbd_spec_alloc(void)
4056 struct rbd_spec *spec;
4058 spec = kzalloc(sizeof (*spec), GFP_KERNEL);
4062 spec->pool_id = CEPH_NOPOOL;
4063 spec->snap_id = CEPH_NOSNAP;
4064 kref_init(&spec->kref);
4069 static void rbd_spec_free(struct kref *kref)
4071 struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
4073 kfree(spec->pool_name);
4074 kfree(spec->image_id);
4075 kfree(spec->image_name);
4076 kfree(spec->snap_name);
4080 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
4081 struct rbd_spec *spec,
4082 struct rbd_options *opts)
4084 struct rbd_device *rbd_dev;
4086 rbd_dev = kzalloc(sizeof (*rbd_dev), GFP_KERNEL);
4090 spin_lock_init(&rbd_dev->lock);
4092 atomic_set(&rbd_dev->parent_ref, 0);
4093 INIT_LIST_HEAD(&rbd_dev->node);
4094 init_rwsem(&rbd_dev->header_rwsem);
4096 rbd_dev->rbd_client = rbdc;
4097 rbd_dev->spec = spec;
4098 rbd_dev->opts = opts;
4100 /* Initialize the layout used for all rbd requests */
4102 rbd_dev->layout.fl_stripe_unit = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
4103 rbd_dev->layout.fl_stripe_count = cpu_to_le32(1);
4104 rbd_dev->layout.fl_object_size = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
4105 rbd_dev->layout.fl_pg_pool = cpu_to_le32((u32) spec->pool_id);
4110 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
4112 rbd_put_client(rbd_dev->rbd_client);
4113 rbd_spec_put(rbd_dev->spec);
4114 kfree(rbd_dev->opts);
4119 * Get the size and object order for an image snapshot, or if
4120 * snap_id is CEPH_NOSNAP, gets this information for the base
4123 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
4124 u8 *order, u64 *snap_size)
4126 __le64 snapid = cpu_to_le64(snap_id);
4131 } __attribute__ ((packed)) size_buf = { 0 };
4133 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4135 &snapid, sizeof (snapid),
4136 &size_buf, sizeof (size_buf));
4137 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4140 if (ret < sizeof (size_buf))
4144 *order = size_buf.order;
4145 dout(" order %u", (unsigned int)*order);
4147 *snap_size = le64_to_cpu(size_buf.size);
4149 dout(" snap_id 0x%016llx snap_size = %llu\n",
4150 (unsigned long long)snap_id,
4151 (unsigned long long)*snap_size);
4156 static int rbd_dev_v2_image_size(struct rbd_device *rbd_dev)
4158 return _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
4159 &rbd_dev->header.obj_order,
4160 &rbd_dev->header.image_size);
4163 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev)
4169 reply_buf = kzalloc(RBD_OBJ_PREFIX_LEN_MAX, GFP_KERNEL);
4173 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4174 "rbd", "get_object_prefix", NULL, 0,
4175 reply_buf, RBD_OBJ_PREFIX_LEN_MAX);
4176 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4181 rbd_dev->header.object_prefix = ceph_extract_encoded_string(&p,
4182 p + ret, NULL, GFP_NOIO);
4185 if (IS_ERR(rbd_dev->header.object_prefix)) {
4186 ret = PTR_ERR(rbd_dev->header.object_prefix);
4187 rbd_dev->header.object_prefix = NULL;
4189 dout(" object_prefix = %s\n", rbd_dev->header.object_prefix);
4197 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
4200 __le64 snapid = cpu_to_le64(snap_id);
4204 } __attribute__ ((packed)) features_buf = { 0 };
4208 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4209 "rbd", "get_features",
4210 &snapid, sizeof (snapid),
4211 &features_buf, sizeof (features_buf));
4212 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4215 if (ret < sizeof (features_buf))
4218 incompat = le64_to_cpu(features_buf.incompat);
4219 if (incompat & ~RBD_FEATURES_SUPPORTED)
4222 *snap_features = le64_to_cpu(features_buf.features);
4224 dout(" snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
4225 (unsigned long long)snap_id,
4226 (unsigned long long)*snap_features,
4227 (unsigned long long)le64_to_cpu(features_buf.incompat));
4232 static int rbd_dev_v2_features(struct rbd_device *rbd_dev)
4234 return _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
4235 &rbd_dev->header.features);
4238 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev)
4240 struct rbd_spec *parent_spec;
4242 void *reply_buf = NULL;
4252 parent_spec = rbd_spec_alloc();
4256 size = sizeof (__le64) + /* pool_id */
4257 sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX + /* image_id */
4258 sizeof (__le64) + /* snap_id */
4259 sizeof (__le64); /* overlap */
4260 reply_buf = kmalloc(size, GFP_KERNEL);
4266 snapid = cpu_to_le64(rbd_dev->spec->snap_id);
4267 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4268 "rbd", "get_parent",
4269 &snapid, sizeof (snapid),
4271 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4276 end = reply_buf + ret;
4278 ceph_decode_64_safe(&p, end, pool_id, out_err);
4279 if (pool_id == CEPH_NOPOOL) {
4281 * Either the parent never existed, or we have
4282 * record of it but the image got flattened so it no
4283 * longer has a parent. When the parent of a
4284 * layered image disappears we immediately set the
4285 * overlap to 0. The effect of this is that all new
4286 * requests will be treated as if the image had no
4289 if (rbd_dev->parent_overlap) {
4290 rbd_dev->parent_overlap = 0;
4291 rbd_dev_parent_put(rbd_dev);
4292 pr_info("%s: clone image has been flattened\n",
4293 rbd_dev->disk->disk_name);
4296 goto out; /* No parent? No problem. */
4299 /* The ceph file layout needs to fit pool id in 32 bits */
4302 if (pool_id > (u64)U32_MAX) {
4303 rbd_warn(NULL, "parent pool id too large (%llu > %u)",
4304 (unsigned long long)pool_id, U32_MAX);
4308 image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
4309 if (IS_ERR(image_id)) {
4310 ret = PTR_ERR(image_id);
4313 ceph_decode_64_safe(&p, end, snap_id, out_err);
4314 ceph_decode_64_safe(&p, end, overlap, out_err);
4317 * The parent won't change (except when the clone is
4318 * flattened, already handled that). So we only need to
4319 * record the parent spec we have not already done so.
4321 if (!rbd_dev->parent_spec) {
4322 parent_spec->pool_id = pool_id;
4323 parent_spec->image_id = image_id;
4324 parent_spec->snap_id = snap_id;
4325 rbd_dev->parent_spec = parent_spec;
4326 parent_spec = NULL; /* rbd_dev now owns this */
4332 * We always update the parent overlap. If it's zero we issue
4333 * a warning, as we will proceed as if there was no parent.
4337 /* refresh, careful to warn just once */
4338 if (rbd_dev->parent_overlap)
4340 "clone now standalone (overlap became 0)");
4343 rbd_warn(rbd_dev, "clone is standalone (overlap 0)");
4346 rbd_dev->parent_overlap = overlap;
4352 rbd_spec_put(parent_spec);
4357 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev)
4361 __le64 stripe_count;
4362 } __attribute__ ((packed)) striping_info_buf = { 0 };
4363 size_t size = sizeof (striping_info_buf);
4370 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4371 "rbd", "get_stripe_unit_count", NULL, 0,
4372 (char *)&striping_info_buf, size);
4373 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4380 * We don't actually support the "fancy striping" feature
4381 * (STRIPINGV2) yet, but if the striping sizes are the
4382 * defaults the behavior is the same as before. So find
4383 * out, and only fail if the image has non-default values.
4386 obj_size = (u64)1 << rbd_dev->header.obj_order;
4387 p = &striping_info_buf;
4388 stripe_unit = ceph_decode_64(&p);
4389 if (stripe_unit != obj_size) {
4390 rbd_warn(rbd_dev, "unsupported stripe unit "
4391 "(got %llu want %llu)",
4392 stripe_unit, obj_size);
4395 stripe_count = ceph_decode_64(&p);
4396 if (stripe_count != 1) {
4397 rbd_warn(rbd_dev, "unsupported stripe count "
4398 "(got %llu want 1)", stripe_count);
4401 rbd_dev->header.stripe_unit = stripe_unit;
4402 rbd_dev->header.stripe_count = stripe_count;
4407 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
4409 size_t image_id_size;
4414 void *reply_buf = NULL;
4416 char *image_name = NULL;
4419 rbd_assert(!rbd_dev->spec->image_name);
4421 len = strlen(rbd_dev->spec->image_id);
4422 image_id_size = sizeof (__le32) + len;
4423 image_id = kmalloc(image_id_size, GFP_KERNEL);
4428 end = image_id + image_id_size;
4429 ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
4431 size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
4432 reply_buf = kmalloc(size, GFP_KERNEL);
4436 ret = rbd_obj_method_sync(rbd_dev, RBD_DIRECTORY,
4437 "rbd", "dir_get_name",
4438 image_id, image_id_size,
4443 end = reply_buf + ret;
4445 image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
4446 if (IS_ERR(image_name))
4449 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
4457 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4459 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4460 const char *snap_name;
4463 /* Skip over names until we find the one we are looking for */
4465 snap_name = rbd_dev->header.snap_names;
4466 while (which < snapc->num_snaps) {
4467 if (!strcmp(name, snap_name))
4468 return snapc->snaps[which];
4469 snap_name += strlen(snap_name) + 1;
4475 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4477 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4482 for (which = 0; !found && which < snapc->num_snaps; which++) {
4483 const char *snap_name;
4485 snap_id = snapc->snaps[which];
4486 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
4487 if (IS_ERR(snap_name)) {
4488 /* ignore no-longer existing snapshots */
4489 if (PTR_ERR(snap_name) == -ENOENT)
4494 found = !strcmp(name, snap_name);
4497 return found ? snap_id : CEPH_NOSNAP;
4501 * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
4502 * no snapshot by that name is found, or if an error occurs.
4504 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4506 if (rbd_dev->image_format == 1)
4507 return rbd_v1_snap_id_by_name(rbd_dev, name);
4509 return rbd_v2_snap_id_by_name(rbd_dev, name);
4513 * An image being mapped will have everything but the snap id.
4515 static int rbd_spec_fill_snap_id(struct rbd_device *rbd_dev)
4517 struct rbd_spec *spec = rbd_dev->spec;
4519 rbd_assert(spec->pool_id != CEPH_NOPOOL && spec->pool_name);
4520 rbd_assert(spec->image_id && spec->image_name);
4521 rbd_assert(spec->snap_name);
4523 if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
4526 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
4527 if (snap_id == CEPH_NOSNAP)
4530 spec->snap_id = snap_id;
4532 spec->snap_id = CEPH_NOSNAP;
4539 * A parent image will have all ids but none of the names.
4541 * All names in an rbd spec are dynamically allocated. It's OK if we
4542 * can't figure out the name for an image id.
4544 static int rbd_spec_fill_names(struct rbd_device *rbd_dev)
4546 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4547 struct rbd_spec *spec = rbd_dev->spec;
4548 const char *pool_name;
4549 const char *image_name;
4550 const char *snap_name;
4553 rbd_assert(spec->pool_id != CEPH_NOPOOL);
4554 rbd_assert(spec->image_id);
4555 rbd_assert(spec->snap_id != CEPH_NOSNAP);
4557 /* Get the pool name; we have to make our own copy of this */
4559 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
4561 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
4564 pool_name = kstrdup(pool_name, GFP_KERNEL);
4568 /* Fetch the image name; tolerate failure here */
4570 image_name = rbd_dev_image_name(rbd_dev);
4572 rbd_warn(rbd_dev, "unable to get image name");
4574 /* Fetch the snapshot name */
4576 snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
4577 if (IS_ERR(snap_name)) {
4578 ret = PTR_ERR(snap_name);
4582 spec->pool_name = pool_name;
4583 spec->image_name = image_name;
4584 spec->snap_name = snap_name;
4594 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev)
4603 struct ceph_snap_context *snapc;
4607 * We'll need room for the seq value (maximum snapshot id),
4608 * snapshot count, and array of that many snapshot ids.
4609 * For now we have a fixed upper limit on the number we're
4610 * prepared to receive.
4612 size = sizeof (__le64) + sizeof (__le32) +
4613 RBD_MAX_SNAP_COUNT * sizeof (__le64);
4614 reply_buf = kzalloc(size, GFP_KERNEL);
4618 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4619 "rbd", "get_snapcontext", NULL, 0,
4621 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4626 end = reply_buf + ret;
4628 ceph_decode_64_safe(&p, end, seq, out);
4629 ceph_decode_32_safe(&p, end, snap_count, out);
4632 * Make sure the reported number of snapshot ids wouldn't go
4633 * beyond the end of our buffer. But before checking that,
4634 * make sure the computed size of the snapshot context we
4635 * allocate is representable in a size_t.
4637 if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
4642 if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
4646 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
4652 for (i = 0; i < snap_count; i++)
4653 snapc->snaps[i] = ceph_decode_64(&p);
4655 ceph_put_snap_context(rbd_dev->header.snapc);
4656 rbd_dev->header.snapc = snapc;
4658 dout(" snap context seq = %llu, snap_count = %u\n",
4659 (unsigned long long)seq, (unsigned int)snap_count);
4666 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
4677 size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
4678 reply_buf = kmalloc(size, GFP_KERNEL);
4680 return ERR_PTR(-ENOMEM);
4682 snapid = cpu_to_le64(snap_id);
4683 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_name,
4684 "rbd", "get_snapshot_name",
4685 &snapid, sizeof (snapid),
4687 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4689 snap_name = ERR_PTR(ret);
4694 end = reply_buf + ret;
4695 snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
4696 if (IS_ERR(snap_name))
4699 dout(" snap_id 0x%016llx snap_name = %s\n",
4700 (unsigned long long)snap_id, snap_name);
4707 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev)
4709 bool first_time = rbd_dev->header.object_prefix == NULL;
4712 ret = rbd_dev_v2_image_size(rbd_dev);
4717 ret = rbd_dev_v2_header_onetime(rbd_dev);
4722 ret = rbd_dev_v2_snap_context(rbd_dev);
4723 dout("rbd_dev_v2_snap_context returned %d\n", ret);
4728 static int rbd_dev_header_info(struct rbd_device *rbd_dev)
4730 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
4732 if (rbd_dev->image_format == 1)
4733 return rbd_dev_v1_header_info(rbd_dev);
4735 return rbd_dev_v2_header_info(rbd_dev);
4738 static int rbd_bus_add_dev(struct rbd_device *rbd_dev)
4743 dev = &rbd_dev->dev;
4744 dev->bus = &rbd_bus_type;
4745 dev->type = &rbd_device_type;
4746 dev->parent = &rbd_root_dev;
4747 dev->release = rbd_dev_device_release;
4748 dev_set_name(dev, "%d", rbd_dev->dev_id);
4749 ret = device_register(dev);
4754 static void rbd_bus_del_dev(struct rbd_device *rbd_dev)
4756 device_unregister(&rbd_dev->dev);
4760 * Get a unique rbd identifier for the given new rbd_dev, and add
4761 * the rbd_dev to the global list.
4763 static int rbd_dev_id_get(struct rbd_device *rbd_dev)
4767 new_dev_id = ida_simple_get(&rbd_dev_id_ida,
4768 0, minor_to_rbd_dev_id(1 << MINORBITS),
4773 rbd_dev->dev_id = new_dev_id;
4775 spin_lock(&rbd_dev_list_lock);
4776 list_add_tail(&rbd_dev->node, &rbd_dev_list);
4777 spin_unlock(&rbd_dev_list_lock);
4779 dout("rbd_dev %p given dev id %d\n", rbd_dev, rbd_dev->dev_id);
4785 * Remove an rbd_dev from the global list, and record that its
4786 * identifier is no longer in use.
4788 static void rbd_dev_id_put(struct rbd_device *rbd_dev)
4790 spin_lock(&rbd_dev_list_lock);
4791 list_del_init(&rbd_dev->node);
4792 spin_unlock(&rbd_dev_list_lock);
4794 ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4796 dout("rbd_dev %p released dev id %d\n", rbd_dev, rbd_dev->dev_id);
4800 * Skips over white space at *buf, and updates *buf to point to the
4801 * first found non-space character (if any). Returns the length of
4802 * the token (string of non-white space characters) found. Note
4803 * that *buf must be terminated with '\0'.
4805 static inline size_t next_token(const char **buf)
4808 * These are the characters that produce nonzero for
4809 * isspace() in the "C" and "POSIX" locales.
4811 const char *spaces = " \f\n\r\t\v";
4813 *buf += strspn(*buf, spaces); /* Find start of token */
4815 return strcspn(*buf, spaces); /* Return token length */
4819 * Finds the next token in *buf, dynamically allocates a buffer big
4820 * enough to hold a copy of it, and copies the token into the new
4821 * buffer. The copy is guaranteed to be terminated with '\0'. Note
4822 * that a duplicate buffer is created even for a zero-length token.
4824 * Returns a pointer to the newly-allocated duplicate, or a null
4825 * pointer if memory for the duplicate was not available. If
4826 * the lenp argument is a non-null pointer, the length of the token
4827 * (not including the '\0') is returned in *lenp.
4829 * If successful, the *buf pointer will be updated to point beyond
4830 * the end of the found token.
4832 * Note: uses GFP_KERNEL for allocation.
4834 static inline char *dup_token(const char **buf, size_t *lenp)
4839 len = next_token(buf);
4840 dup = kmemdup(*buf, len + 1, GFP_KERNEL);
4843 *(dup + len) = '\0';
4853 * Parse the options provided for an "rbd add" (i.e., rbd image
4854 * mapping) request. These arrive via a write to /sys/bus/rbd/add,
4855 * and the data written is passed here via a NUL-terminated buffer.
4856 * Returns 0 if successful or an error code otherwise.
4858 * The information extracted from these options is recorded in
4859 * the other parameters which return dynamically-allocated
4862 * The address of a pointer that will refer to a ceph options
4863 * structure. Caller must release the returned pointer using
4864 * ceph_destroy_options() when it is no longer needed.
4866 * Address of an rbd options pointer. Fully initialized by
4867 * this function; caller must release with kfree().
4869 * Address of an rbd image specification pointer. Fully
4870 * initialized by this function based on parsed options.
4871 * Caller must release with rbd_spec_put().
4873 * The options passed take this form:
4874 * <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
4877 * A comma-separated list of one or more monitor addresses.
4878 * A monitor address is an ip address, optionally followed
4879 * by a port number (separated by a colon).
4880 * I.e.: ip1[:port1][,ip2[:port2]...]
4882 * A comma-separated list of ceph and/or rbd options.
4884 * The name of the rados pool containing the rbd image.
4886 * The name of the image in that pool to map.
4888 * An optional snapshot id. If provided, the mapping will
4889 * present data from the image at the time that snapshot was
4890 * created. The image head is used if no snapshot id is
4891 * provided. Snapshot mappings are always read-only.
4893 static int rbd_add_parse_args(const char *buf,
4894 struct ceph_options **ceph_opts,
4895 struct rbd_options **opts,
4896 struct rbd_spec **rbd_spec)
4900 const char *mon_addrs;
4902 size_t mon_addrs_size;
4903 struct rbd_spec *spec = NULL;
4904 struct rbd_options *rbd_opts = NULL;
4905 struct ceph_options *copts;
4908 /* The first four tokens are required */
4910 len = next_token(&buf);
4912 rbd_warn(NULL, "no monitor address(es) provided");
4916 mon_addrs_size = len + 1;
4920 options = dup_token(&buf, NULL);
4924 rbd_warn(NULL, "no options provided");
4928 spec = rbd_spec_alloc();
4932 spec->pool_name = dup_token(&buf, NULL);
4933 if (!spec->pool_name)
4935 if (!*spec->pool_name) {
4936 rbd_warn(NULL, "no pool name provided");
4940 spec->image_name = dup_token(&buf, NULL);
4941 if (!spec->image_name)
4943 if (!*spec->image_name) {
4944 rbd_warn(NULL, "no image name provided");
4949 * Snapshot name is optional; default is to use "-"
4950 * (indicating the head/no snapshot).
4952 len = next_token(&buf);
4954 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
4955 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
4956 } else if (len > RBD_MAX_SNAP_NAME_LEN) {
4957 ret = -ENAMETOOLONG;
4960 snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
4963 *(snap_name + len) = '\0';
4964 spec->snap_name = snap_name;
4966 /* Initialize all rbd options to the defaults */
4968 rbd_opts = kzalloc(sizeof (*rbd_opts), GFP_KERNEL);
4972 rbd_opts->read_only = RBD_READ_ONLY_DEFAULT;
4973 rbd_opts->queue_depth = RBD_QUEUE_DEPTH_DEFAULT;
4975 copts = ceph_parse_options(options, mon_addrs,
4976 mon_addrs + mon_addrs_size - 1,
4977 parse_rbd_opts_token, rbd_opts);
4978 if (IS_ERR(copts)) {
4979 ret = PTR_ERR(copts);
5000 * Return pool id (>= 0) or a negative error code.
5002 static int rbd_add_get_pool_id(struct rbd_client *rbdc, const char *pool_name)
5004 struct ceph_options *opts = rbdc->client->options;
5010 ret = ceph_pg_poolid_by_name(rbdc->client->osdc.osdmap, pool_name);
5011 if (ret == -ENOENT && tries++ < 1) {
5012 ret = ceph_monc_do_get_version(&rbdc->client->monc, "osdmap",
5017 if (rbdc->client->osdc.osdmap->epoch < newest_epoch) {
5018 ceph_monc_request_next_osdmap(&rbdc->client->monc);
5019 (void) ceph_monc_wait_osdmap(&rbdc->client->monc,
5021 opts->mount_timeout);
5024 /* the osdmap we have is new enough */
5033 * An rbd format 2 image has a unique identifier, distinct from the
5034 * name given to it by the user. Internally, that identifier is
5035 * what's used to specify the names of objects related to the image.
5037 * A special "rbd id" object is used to map an rbd image name to its
5038 * id. If that object doesn't exist, then there is no v2 rbd image
5039 * with the supplied name.
5041 * This function will record the given rbd_dev's image_id field if
5042 * it can be determined, and in that case will return 0. If any
5043 * errors occur a negative errno will be returned and the rbd_dev's
5044 * image_id field will be unchanged (and should be NULL).
5046 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
5055 * When probing a parent image, the image id is already
5056 * known (and the image name likely is not). There's no
5057 * need to fetch the image id again in this case. We
5058 * do still need to set the image format though.
5060 if (rbd_dev->spec->image_id) {
5061 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
5067 * First, see if the format 2 image id file exists, and if
5068 * so, get the image's persistent id from it.
5070 size = sizeof (RBD_ID_PREFIX) + strlen(rbd_dev->spec->image_name);
5071 object_name = kmalloc(size, GFP_NOIO);
5074 sprintf(object_name, "%s%s", RBD_ID_PREFIX, rbd_dev->spec->image_name);
5075 dout("rbd id object name is %s\n", object_name);
5077 /* Response will be an encoded string, which includes a length */
5079 size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
5080 response = kzalloc(size, GFP_NOIO);
5086 /* If it doesn't exist we'll assume it's a format 1 image */
5088 ret = rbd_obj_method_sync(rbd_dev, object_name,
5089 "rbd", "get_id", NULL, 0,
5090 response, RBD_IMAGE_ID_LEN_MAX);
5091 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5092 if (ret == -ENOENT) {
5093 image_id = kstrdup("", GFP_KERNEL);
5094 ret = image_id ? 0 : -ENOMEM;
5096 rbd_dev->image_format = 1;
5097 } else if (ret >= 0) {
5100 image_id = ceph_extract_encoded_string(&p, p + ret,
5102 ret = PTR_ERR_OR_ZERO(image_id);
5104 rbd_dev->image_format = 2;
5108 rbd_dev->spec->image_id = image_id;
5109 dout("image_id is %s\n", image_id);
5119 * Undo whatever state changes are made by v1 or v2 header info
5122 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
5124 struct rbd_image_header *header;
5126 rbd_dev_parent_put(rbd_dev);
5128 /* Free dynamic fields from the header, then zero it out */
5130 header = &rbd_dev->header;
5131 ceph_put_snap_context(header->snapc);
5132 kfree(header->snap_sizes);
5133 kfree(header->snap_names);
5134 kfree(header->object_prefix);
5135 memset(header, 0, sizeof (*header));
5138 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev)
5142 ret = rbd_dev_v2_object_prefix(rbd_dev);
5147 * Get the and check features for the image. Currently the
5148 * features are assumed to never change.
5150 ret = rbd_dev_v2_features(rbd_dev);
5154 /* If the image supports fancy striping, get its parameters */
5156 if (rbd_dev->header.features & RBD_FEATURE_STRIPINGV2) {
5157 ret = rbd_dev_v2_striping_info(rbd_dev);
5161 /* No support for crypto and compression type format 2 images */
5165 rbd_dev->header.features = 0;
5166 kfree(rbd_dev->header.object_prefix);
5167 rbd_dev->header.object_prefix = NULL;
5172 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev)
5174 struct rbd_device *parent = NULL;
5175 struct rbd_spec *parent_spec;
5176 struct rbd_client *rbdc;
5179 if (!rbd_dev->parent_spec)
5182 * We need to pass a reference to the client and the parent
5183 * spec when creating the parent rbd_dev. Images related by
5184 * parent/child relationships always share both.
5186 parent_spec = rbd_spec_get(rbd_dev->parent_spec);
5187 rbdc = __rbd_get_client(rbd_dev->rbd_client);
5190 parent = rbd_dev_create(rbdc, parent_spec, NULL);
5194 ret = rbd_dev_image_probe(parent, false);
5197 rbd_dev->parent = parent;
5198 atomic_set(&rbd_dev->parent_ref, 1);
5203 rbd_dev_unparent(rbd_dev);
5204 kfree(rbd_dev->header_name);
5205 rbd_dev_destroy(parent);
5207 rbd_put_client(rbdc);
5208 rbd_spec_put(parent_spec);
5214 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
5218 /* Get an id and fill in device name. */
5220 ret = rbd_dev_id_get(rbd_dev);
5224 BUILD_BUG_ON(DEV_NAME_LEN
5225 < sizeof (RBD_DRV_NAME) + MAX_INT_FORMAT_WIDTH);
5226 sprintf(rbd_dev->name, "%s%d", RBD_DRV_NAME, rbd_dev->dev_id);
5228 /* Record our major and minor device numbers. */
5230 if (!single_major) {
5231 ret = register_blkdev(0, rbd_dev->name);
5235 rbd_dev->major = ret;
5238 rbd_dev->major = rbd_major;
5239 rbd_dev->minor = rbd_dev_id_to_minor(rbd_dev->dev_id);
5242 /* Set up the blkdev mapping. */
5244 ret = rbd_init_disk(rbd_dev);
5246 goto err_out_blkdev;
5248 ret = rbd_dev_mapping_set(rbd_dev);
5252 set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
5253 set_disk_ro(rbd_dev->disk, rbd_dev->mapping.read_only);
5255 ret = rbd_bus_add_dev(rbd_dev);
5257 goto err_out_mapping;
5259 /* Everything's ready. Announce the disk to the world. */
5261 set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5262 add_disk(rbd_dev->disk);
5264 pr_info("%s: added with size 0x%llx\n", rbd_dev->disk->disk_name,
5265 (unsigned long long) rbd_dev->mapping.size);
5270 rbd_dev_mapping_clear(rbd_dev);
5272 rbd_free_disk(rbd_dev);
5275 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5277 rbd_dev_id_put(rbd_dev);
5278 rbd_dev_mapping_clear(rbd_dev);
5283 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
5285 struct rbd_spec *spec = rbd_dev->spec;
5288 /* Record the header object name for this rbd image. */
5290 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
5292 if (rbd_dev->image_format == 1)
5293 size = strlen(spec->image_name) + sizeof (RBD_SUFFIX);
5295 size = sizeof (RBD_HEADER_PREFIX) + strlen(spec->image_id);
5297 rbd_dev->header_name = kmalloc(size, GFP_KERNEL);
5298 if (!rbd_dev->header_name)
5301 if (rbd_dev->image_format == 1)
5302 sprintf(rbd_dev->header_name, "%s%s",
5303 spec->image_name, RBD_SUFFIX);
5305 sprintf(rbd_dev->header_name, "%s%s",
5306 RBD_HEADER_PREFIX, spec->image_id);
5310 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
5312 rbd_dev_unprobe(rbd_dev);
5313 kfree(rbd_dev->header_name);
5314 rbd_dev->header_name = NULL;
5315 rbd_dev->image_format = 0;
5316 kfree(rbd_dev->spec->image_id);
5317 rbd_dev->spec->image_id = NULL;
5319 rbd_dev_destroy(rbd_dev);
5323 * Probe for the existence of the header object for the given rbd
5324 * device. If this image is the one being mapped (i.e., not a
5325 * parent), initiate a watch on its header object before using that
5326 * object to get detailed information about the rbd image.
5328 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, bool mapping)
5333 * Get the id from the image id object. Unless there's an
5334 * error, rbd_dev->spec->image_id will be filled in with
5335 * a dynamically-allocated string, and rbd_dev->image_format
5336 * will be set to either 1 or 2.
5338 ret = rbd_dev_image_id(rbd_dev);
5342 ret = rbd_dev_header_name(rbd_dev);
5344 goto err_out_format;
5347 ret = rbd_dev_header_watch_sync(rbd_dev);
5350 pr_info("image %s/%s does not exist\n",
5351 rbd_dev->spec->pool_name,
5352 rbd_dev->spec->image_name);
5353 goto out_header_name;
5357 ret = rbd_dev_header_info(rbd_dev);
5362 * If this image is the one being mapped, we have pool name and
5363 * id, image name and id, and snap name - need to fill snap id.
5364 * Otherwise this is a parent image, identified by pool, image
5365 * and snap ids - need to fill in names for those ids.
5368 ret = rbd_spec_fill_snap_id(rbd_dev);
5370 ret = rbd_spec_fill_names(rbd_dev);
5373 pr_info("snap %s/%s@%s does not exist\n",
5374 rbd_dev->spec->pool_name,
5375 rbd_dev->spec->image_name,
5376 rbd_dev->spec->snap_name);
5380 if (rbd_dev->header.features & RBD_FEATURE_LAYERING) {
5381 ret = rbd_dev_v2_parent_info(rbd_dev);
5386 * Need to warn users if this image is the one being
5387 * mapped and has a parent.
5389 if (mapping && rbd_dev->parent_spec)
5391 "WARNING: kernel layering is EXPERIMENTAL!");
5394 ret = rbd_dev_probe_parent(rbd_dev);
5398 dout("discovered format %u image, header name is %s\n",
5399 rbd_dev->image_format, rbd_dev->header_name);
5403 rbd_dev_unprobe(rbd_dev);
5406 rbd_dev_header_unwatch_sync(rbd_dev);
5408 kfree(rbd_dev->header_name);
5409 rbd_dev->header_name = NULL;
5411 rbd_dev->image_format = 0;
5412 kfree(rbd_dev->spec->image_id);
5413 rbd_dev->spec->image_id = NULL;
5417 static ssize_t do_rbd_add(struct bus_type *bus,
5421 struct rbd_device *rbd_dev = NULL;
5422 struct ceph_options *ceph_opts = NULL;
5423 struct rbd_options *rbd_opts = NULL;
5424 struct rbd_spec *spec = NULL;
5425 struct rbd_client *rbdc;
5429 if (!try_module_get(THIS_MODULE))
5432 /* parse add command */
5433 rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
5435 goto err_out_module;
5437 rbdc = rbd_get_client(ceph_opts);
5444 rc = rbd_add_get_pool_id(rbdc, spec->pool_name);
5447 pr_info("pool %s does not exist\n", spec->pool_name);
5448 goto err_out_client;
5450 spec->pool_id = (u64)rc;
5452 /* The ceph file layout needs to fit pool id in 32 bits */
5454 if (spec->pool_id > (u64)U32_MAX) {
5455 rbd_warn(NULL, "pool id too large (%llu > %u)",
5456 (unsigned long long)spec->pool_id, U32_MAX);
5458 goto err_out_client;
5461 rbd_dev = rbd_dev_create(rbdc, spec, rbd_opts);
5463 goto err_out_client;
5464 rbdc = NULL; /* rbd_dev now owns this */
5465 spec = NULL; /* rbd_dev now owns this */
5466 rbd_opts = NULL; /* rbd_dev now owns this */
5468 rc = rbd_dev_image_probe(rbd_dev, true);
5470 goto err_out_rbd_dev;
5472 /* If we are mapping a snapshot it must be marked read-only */
5474 read_only = rbd_dev->opts->read_only;
5475 if (rbd_dev->spec->snap_id != CEPH_NOSNAP)
5477 rbd_dev->mapping.read_only = read_only;
5479 rc = rbd_dev_device_setup(rbd_dev);
5482 * rbd_dev_header_unwatch_sync() can't be moved into
5483 * rbd_dev_image_release() without refactoring, see
5484 * commit 1f3ef78861ac.
5486 rbd_dev_header_unwatch_sync(rbd_dev);
5487 rbd_dev_image_release(rbd_dev);
5488 goto err_out_module;
5494 rbd_dev_destroy(rbd_dev);
5496 rbd_put_client(rbdc);
5501 module_put(THIS_MODULE);
5503 dout("Error adding device %s\n", buf);
5508 static ssize_t rbd_add(struct bus_type *bus,
5515 return do_rbd_add(bus, buf, count);
5518 static ssize_t rbd_add_single_major(struct bus_type *bus,
5522 return do_rbd_add(bus, buf, count);
5525 static void rbd_dev_device_release(struct device *dev)
5527 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
5529 rbd_free_disk(rbd_dev);
5530 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5531 rbd_dev_mapping_clear(rbd_dev);
5533 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5534 rbd_dev_id_put(rbd_dev);
5535 rbd_dev_mapping_clear(rbd_dev);
5538 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
5540 while (rbd_dev->parent) {
5541 struct rbd_device *first = rbd_dev;
5542 struct rbd_device *second = first->parent;
5543 struct rbd_device *third;
5546 * Follow to the parent with no grandparent and
5549 while (second && (third = second->parent)) {
5554 rbd_dev_image_release(second);
5555 first->parent = NULL;
5556 first->parent_overlap = 0;
5558 rbd_assert(first->parent_spec);
5559 rbd_spec_put(first->parent_spec);
5560 first->parent_spec = NULL;
5564 static ssize_t do_rbd_remove(struct bus_type *bus,
5568 struct rbd_device *rbd_dev = NULL;
5569 struct list_head *tmp;
5572 bool already = false;
5575 ret = kstrtoul(buf, 10, &ul);
5579 /* convert to int; abort if we lost anything in the conversion */
5585 spin_lock(&rbd_dev_list_lock);
5586 list_for_each(tmp, &rbd_dev_list) {
5587 rbd_dev = list_entry(tmp, struct rbd_device, node);
5588 if (rbd_dev->dev_id == dev_id) {
5594 spin_lock_irq(&rbd_dev->lock);
5595 if (rbd_dev->open_count)
5598 already = test_and_set_bit(RBD_DEV_FLAG_REMOVING,
5600 spin_unlock_irq(&rbd_dev->lock);
5602 spin_unlock(&rbd_dev_list_lock);
5603 if (ret < 0 || already)
5606 rbd_dev_header_unwatch_sync(rbd_dev);
5608 * flush remaining watch callbacks - these must be complete
5609 * before the osd_client is shutdown
5611 dout("%s: flushing notifies", __func__);
5612 ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
5615 * Don't free anything from rbd_dev->disk until after all
5616 * notifies are completely processed. Otherwise
5617 * rbd_bus_del_dev() will race with rbd_watch_cb(), resulting
5618 * in a potential use after free of rbd_dev->disk or rbd_dev.
5620 rbd_bus_del_dev(rbd_dev);
5621 rbd_dev_image_release(rbd_dev);
5622 module_put(THIS_MODULE);
5627 static ssize_t rbd_remove(struct bus_type *bus,
5634 return do_rbd_remove(bus, buf, count);
5637 static ssize_t rbd_remove_single_major(struct bus_type *bus,
5641 return do_rbd_remove(bus, buf, count);
5645 * create control files in sysfs
5648 static int rbd_sysfs_init(void)
5652 ret = device_register(&rbd_root_dev);
5656 ret = bus_register(&rbd_bus_type);
5658 device_unregister(&rbd_root_dev);
5663 static void rbd_sysfs_cleanup(void)
5665 bus_unregister(&rbd_bus_type);
5666 device_unregister(&rbd_root_dev);
5669 static int rbd_slab_init(void)
5671 rbd_assert(!rbd_img_request_cache);
5672 rbd_img_request_cache = kmem_cache_create("rbd_img_request",
5673 sizeof (struct rbd_img_request),
5674 __alignof__(struct rbd_img_request),
5676 if (!rbd_img_request_cache)
5679 rbd_assert(!rbd_obj_request_cache);
5680 rbd_obj_request_cache = kmem_cache_create("rbd_obj_request",
5681 sizeof (struct rbd_obj_request),
5682 __alignof__(struct rbd_obj_request),
5684 if (!rbd_obj_request_cache)
5687 rbd_assert(!rbd_segment_name_cache);
5688 rbd_segment_name_cache = kmem_cache_create("rbd_segment_name",
5689 CEPH_MAX_OID_NAME_LEN + 1, 1, 0, NULL);
5690 if (rbd_segment_name_cache)
5693 if (rbd_obj_request_cache) {
5694 kmem_cache_destroy(rbd_obj_request_cache);
5695 rbd_obj_request_cache = NULL;
5698 kmem_cache_destroy(rbd_img_request_cache);
5699 rbd_img_request_cache = NULL;
5704 static void rbd_slab_exit(void)
5706 rbd_assert(rbd_segment_name_cache);
5707 kmem_cache_destroy(rbd_segment_name_cache);
5708 rbd_segment_name_cache = NULL;
5710 rbd_assert(rbd_obj_request_cache);
5711 kmem_cache_destroy(rbd_obj_request_cache);
5712 rbd_obj_request_cache = NULL;
5714 rbd_assert(rbd_img_request_cache);
5715 kmem_cache_destroy(rbd_img_request_cache);
5716 rbd_img_request_cache = NULL;
5719 static int __init rbd_init(void)
5723 if (!libceph_compatible(NULL)) {
5724 rbd_warn(NULL, "libceph incompatibility (quitting)");
5728 rc = rbd_slab_init();
5733 * The number of active work items is limited by the number of
5734 * rbd devices * queue depth, so leave @max_active at default.
5736 rbd_wq = alloc_workqueue(RBD_DRV_NAME, WQ_MEM_RECLAIM, 0);
5743 rbd_major = register_blkdev(0, RBD_DRV_NAME);
5744 if (rbd_major < 0) {
5750 rc = rbd_sysfs_init();
5752 goto err_out_blkdev;
5755 pr_info("loaded (major %d)\n", rbd_major);
5757 pr_info("loaded\n");
5763 unregister_blkdev(rbd_major, RBD_DRV_NAME);
5765 destroy_workqueue(rbd_wq);
5771 static void __exit rbd_exit(void)
5773 ida_destroy(&rbd_dev_id_ida);
5774 rbd_sysfs_cleanup();
5776 unregister_blkdev(rbd_major, RBD_DRV_NAME);
5777 destroy_workqueue(rbd_wq);
5781 module_init(rbd_init);
5782 module_exit(rbd_exit);
5784 MODULE_AUTHOR("Alex Elder <elder@inktank.com>");
5785 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
5786 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
5787 /* following authorship retained from original osdblk.c */
5788 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
5790 MODULE_DESCRIPTION("RADOS Block Device (RBD) driver");
5791 MODULE_LICENSE("GPL");