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_MAX_PARENT_CHAIN_LEN 16
101 #define RBD_SNAP_DEV_NAME_PREFIX "snap_"
102 #define RBD_MAX_SNAP_NAME_LEN \
103 (NAME_MAX - (sizeof (RBD_SNAP_DEV_NAME_PREFIX) - 1))
105 #define RBD_MAX_SNAP_COUNT 510 /* allows max snapc to fit in 4KB */
107 #define RBD_SNAP_HEAD_NAME "-"
109 #define BAD_SNAP_INDEX U32_MAX /* invalid index into snap array */
111 /* This allows a single page to hold an image name sent by OSD */
112 #define RBD_IMAGE_NAME_LEN_MAX (PAGE_SIZE - sizeof (__le32) - 1)
113 #define RBD_IMAGE_ID_LEN_MAX 64
115 #define RBD_OBJ_PREFIX_LEN_MAX 64
119 #define RBD_FEATURE_LAYERING (1<<0)
120 #define RBD_FEATURE_STRIPINGV2 (1<<1)
121 #define RBD_FEATURES_ALL \
122 (RBD_FEATURE_LAYERING | RBD_FEATURE_STRIPINGV2)
124 /* Features supported by this (client software) implementation. */
126 #define RBD_FEATURES_SUPPORTED (RBD_FEATURES_ALL)
129 * An RBD device name will be "rbd#", where the "rbd" comes from
130 * RBD_DRV_NAME above, and # is a unique integer identifier.
131 * MAX_INT_FORMAT_WIDTH is used in ensuring DEV_NAME_LEN is big
132 * enough to hold all possible device names.
134 #define DEV_NAME_LEN 32
135 #define MAX_INT_FORMAT_WIDTH ((5 * sizeof (int)) / 2 + 1)
138 * block device image metadata (in-memory version)
140 struct rbd_image_header {
141 /* These six fields never change for a given rbd image */
148 u64 features; /* Might be changeable someday? */
150 /* The remaining fields need to be updated occasionally */
152 struct ceph_snap_context *snapc;
153 char *snap_names; /* format 1 only */
154 u64 *snap_sizes; /* format 1 only */
158 * An rbd image specification.
160 * The tuple (pool_id, image_id, snap_id) is sufficient to uniquely
161 * identify an image. Each rbd_dev structure includes a pointer to
162 * an rbd_spec structure that encapsulates this identity.
164 * Each of the id's in an rbd_spec has an associated name. For a
165 * user-mapped image, the names are supplied and the id's associated
166 * with them are looked up. For a layered image, a parent image is
167 * defined by the tuple, and the names are looked up.
169 * An rbd_dev structure contains a parent_spec pointer which is
170 * non-null if the image it represents is a child in a layered
171 * image. This pointer will refer to the rbd_spec structure used
172 * by the parent rbd_dev for its own identity (i.e., the structure
173 * is shared between the parent and child).
175 * Since these structures are populated once, during the discovery
176 * phase of image construction, they are effectively immutable so
177 * we make no effort to synchronize access to them.
179 * Note that code herein does not assume the image name is known (it
180 * could be a null pointer).
184 const char *pool_name;
186 const char *image_id;
187 const char *image_name;
190 const char *snap_name;
196 * an instance of the client. multiple devices may share an rbd client.
199 struct ceph_client *client;
201 struct list_head node;
204 struct rbd_img_request;
205 typedef void (*rbd_img_callback_t)(struct rbd_img_request *);
207 #define BAD_WHICH U32_MAX /* Good which or bad which, which? */
209 struct rbd_obj_request;
210 typedef void (*rbd_obj_callback_t)(struct rbd_obj_request *);
212 enum obj_request_type {
213 OBJ_REQUEST_NODATA, OBJ_REQUEST_BIO, OBJ_REQUEST_PAGES
216 enum obj_operation_type {
223 OBJ_REQ_DONE, /* completion flag: not done = 0, done = 1 */
224 OBJ_REQ_IMG_DATA, /* object usage: standalone = 0, image = 1 */
225 OBJ_REQ_KNOWN, /* EXISTS flag valid: no = 0, yes = 1 */
226 OBJ_REQ_EXISTS, /* target exists: no = 0, yes = 1 */
229 struct rbd_obj_request {
230 const char *object_name;
231 u64 offset; /* object start byte */
232 u64 length; /* bytes from offset */
236 * An object request associated with an image will have its
237 * img_data flag set; a standalone object request will not.
239 * A standalone object request will have which == BAD_WHICH
240 * and a null obj_request pointer.
242 * An object request initiated in support of a layered image
243 * object (to check for its existence before a write) will
244 * have which == BAD_WHICH and a non-null obj_request pointer.
246 * Finally, an object request for rbd image data will have
247 * which != BAD_WHICH, and will have a non-null img_request
248 * pointer. The value of which will be in the range
249 * 0..(img_request->obj_request_count-1).
252 struct rbd_obj_request *obj_request; /* STAT op */
254 struct rbd_img_request *img_request;
256 /* links for img_request->obj_requests list */
257 struct list_head links;
260 u32 which; /* posn image request list */
262 enum obj_request_type type;
264 struct bio *bio_list;
270 struct page **copyup_pages;
271 u32 copyup_page_count;
273 struct ceph_osd_request *osd_req;
275 u64 xferred; /* bytes transferred */
278 rbd_obj_callback_t callback;
279 struct completion completion;
285 IMG_REQ_WRITE, /* I/O direction: read = 0, write = 1 */
286 IMG_REQ_CHILD, /* initiator: block = 0, child image = 1 */
287 IMG_REQ_LAYERED, /* ENOENT handling: normal = 0, layered = 1 */
288 IMG_REQ_DISCARD, /* discard: normal = 0, discard request = 1 */
291 struct rbd_img_request {
292 struct rbd_device *rbd_dev;
293 u64 offset; /* starting image byte offset */
294 u64 length; /* byte count from offset */
297 u64 snap_id; /* for reads */
298 struct ceph_snap_context *snapc; /* for writes */
301 struct request *rq; /* block request */
302 struct rbd_obj_request *obj_request; /* obj req initiator */
304 struct page **copyup_pages;
305 u32 copyup_page_count;
306 spinlock_t completion_lock;/* protects next_completion */
308 rbd_img_callback_t callback;
309 u64 xferred;/* aggregate bytes transferred */
310 int result; /* first nonzero obj_request result */
312 u32 obj_request_count;
313 struct list_head obj_requests; /* rbd_obj_request structs */
318 #define for_each_obj_request(ireq, oreq) \
319 list_for_each_entry(oreq, &(ireq)->obj_requests, links)
320 #define for_each_obj_request_from(ireq, oreq) \
321 list_for_each_entry_from(oreq, &(ireq)->obj_requests, links)
322 #define for_each_obj_request_safe(ireq, oreq, n) \
323 list_for_each_entry_safe_reverse(oreq, n, &(ireq)->obj_requests, links)
335 int dev_id; /* blkdev unique id */
337 int major; /* blkdev assigned major */
339 struct gendisk *disk; /* blkdev's gendisk and rq */
341 u32 image_format; /* Either 1 or 2 */
342 struct rbd_client *rbd_client;
344 char name[DEV_NAME_LEN]; /* blkdev name, e.g. rbd3 */
346 spinlock_t lock; /* queue, flags, open_count */
348 struct rbd_image_header header;
349 unsigned long flags; /* possibly lock protected */
350 struct rbd_spec *spec;
351 struct rbd_options *opts;
353 struct ceph_object_id header_oid;
355 struct ceph_file_layout layout;
357 struct ceph_osd_event *watch_event;
358 struct rbd_obj_request *watch_request;
360 struct rbd_spec *parent_spec;
363 struct rbd_device *parent;
365 /* Block layer tags. */
366 struct blk_mq_tag_set tag_set;
368 /* protects updating the header */
369 struct rw_semaphore header_rwsem;
371 struct rbd_mapping mapping;
373 struct list_head node;
377 unsigned long open_count; /* protected by lock */
381 * Flag bits for rbd_dev->flags. If atomicity is required,
382 * rbd_dev->lock is used to protect access.
384 * Currently, only the "removing" flag (which is coupled with the
385 * "open_count" field) requires atomic access.
388 RBD_DEV_FLAG_EXISTS, /* mapped snapshot has not been deleted */
389 RBD_DEV_FLAG_REMOVING, /* this mapping is being removed */
392 static DEFINE_MUTEX(client_mutex); /* Serialize client creation */
394 static LIST_HEAD(rbd_dev_list); /* devices */
395 static DEFINE_SPINLOCK(rbd_dev_list_lock);
397 static LIST_HEAD(rbd_client_list); /* clients */
398 static DEFINE_SPINLOCK(rbd_client_list_lock);
400 /* Slab caches for frequently-allocated structures */
402 static struct kmem_cache *rbd_img_request_cache;
403 static struct kmem_cache *rbd_obj_request_cache;
404 static struct kmem_cache *rbd_segment_name_cache;
406 static int rbd_major;
407 static DEFINE_IDA(rbd_dev_id_ida);
409 static struct workqueue_struct *rbd_wq;
412 * Default to false for now, as single-major requires >= 0.75 version of
413 * userspace rbd utility.
415 static bool single_major = false;
416 module_param(single_major, bool, S_IRUGO);
417 MODULE_PARM_DESC(single_major, "Use a single major number for all rbd devices (default: false)");
419 static int rbd_img_request_submit(struct rbd_img_request *img_request);
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, int depth);
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,
542 static int rbd_open(struct block_device *bdev, fmode_t mode)
544 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
545 bool removing = false;
547 if ((mode & FMODE_WRITE) && rbd_dev->mapping.read_only)
550 spin_lock_irq(&rbd_dev->lock);
551 if (test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags))
554 rbd_dev->open_count++;
555 spin_unlock_irq(&rbd_dev->lock);
559 (void) get_device(&rbd_dev->dev);
564 static void rbd_release(struct gendisk *disk, fmode_t mode)
566 struct rbd_device *rbd_dev = disk->private_data;
567 unsigned long open_count_before;
569 spin_lock_irq(&rbd_dev->lock);
570 open_count_before = rbd_dev->open_count--;
571 spin_unlock_irq(&rbd_dev->lock);
572 rbd_assert(open_count_before > 0);
574 put_device(&rbd_dev->dev);
577 static int rbd_ioctl_set_ro(struct rbd_device *rbd_dev, unsigned long arg)
582 bool ro_changed = false;
584 /* get_user() may sleep, so call it before taking rbd_dev->lock */
585 if (get_user(val, (int __user *)(arg)))
588 ro = val ? true : false;
589 /* Snapshot doesn't allow to write*/
590 if (rbd_dev->spec->snap_id != CEPH_NOSNAP && !ro)
593 spin_lock_irq(&rbd_dev->lock);
594 /* prevent others open this device */
595 if (rbd_dev->open_count > 1) {
600 if (rbd_dev->mapping.read_only != ro) {
601 rbd_dev->mapping.read_only = ro;
606 spin_unlock_irq(&rbd_dev->lock);
607 /* set_disk_ro() may sleep, so call it after releasing rbd_dev->lock */
608 if (ret == 0 && ro_changed)
609 set_disk_ro(rbd_dev->disk, ro ? 1 : 0);
614 static int rbd_ioctl(struct block_device *bdev, fmode_t mode,
615 unsigned int cmd, unsigned long arg)
617 struct rbd_device *rbd_dev = bdev->bd_disk->private_data;
622 ret = rbd_ioctl_set_ro(rbd_dev, arg);
632 static int rbd_compat_ioctl(struct block_device *bdev, fmode_t mode,
633 unsigned int cmd, unsigned long arg)
635 return rbd_ioctl(bdev, mode, cmd, arg);
637 #endif /* CONFIG_COMPAT */
639 static const struct block_device_operations rbd_bd_ops = {
640 .owner = THIS_MODULE,
642 .release = rbd_release,
645 .compat_ioctl = rbd_compat_ioctl,
650 * Initialize an rbd client instance. Success or not, this function
651 * consumes ceph_opts. Caller holds client_mutex.
653 static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts)
655 struct rbd_client *rbdc;
658 dout("%s:\n", __func__);
659 rbdc = kmalloc(sizeof(struct rbd_client), GFP_KERNEL);
663 kref_init(&rbdc->kref);
664 INIT_LIST_HEAD(&rbdc->node);
666 rbdc->client = ceph_create_client(ceph_opts, rbdc, 0, 0);
667 if (IS_ERR(rbdc->client))
669 ceph_opts = NULL; /* Now rbdc->client is responsible for ceph_opts */
671 ret = ceph_open_session(rbdc->client);
675 spin_lock(&rbd_client_list_lock);
676 list_add_tail(&rbdc->node, &rbd_client_list);
677 spin_unlock(&rbd_client_list_lock);
679 dout("%s: rbdc %p\n", __func__, rbdc);
683 ceph_destroy_client(rbdc->client);
688 ceph_destroy_options(ceph_opts);
689 dout("%s: error %d\n", __func__, ret);
694 static struct rbd_client *__rbd_get_client(struct rbd_client *rbdc)
696 kref_get(&rbdc->kref);
702 * Find a ceph client with specific addr and configuration. If
703 * found, bump its reference count.
705 static struct rbd_client *rbd_client_find(struct ceph_options *ceph_opts)
707 struct rbd_client *client_node;
710 if (ceph_opts->flags & CEPH_OPT_NOSHARE)
713 spin_lock(&rbd_client_list_lock);
714 list_for_each_entry(client_node, &rbd_client_list, node) {
715 if (!ceph_compare_options(ceph_opts, client_node->client)) {
716 __rbd_get_client(client_node);
722 spin_unlock(&rbd_client_list_lock);
724 return found ? client_node : NULL;
728 * (Per device) rbd map options
735 /* string args above */
741 static match_table_t rbd_opts_tokens = {
742 {Opt_queue_depth, "queue_depth=%d"},
744 /* string args above */
745 {Opt_read_only, "read_only"},
746 {Opt_read_only, "ro"}, /* Alternate spelling */
747 {Opt_read_write, "read_write"},
748 {Opt_read_write, "rw"}, /* Alternate spelling */
757 #define RBD_QUEUE_DEPTH_DEFAULT BLKDEV_MAX_RQ
758 #define RBD_READ_ONLY_DEFAULT false
760 static int parse_rbd_opts_token(char *c, void *private)
762 struct rbd_options *rbd_opts = private;
763 substring_t argstr[MAX_OPT_ARGS];
764 int token, intval, ret;
766 token = match_token(c, rbd_opts_tokens, argstr);
767 if (token < Opt_last_int) {
768 ret = match_int(&argstr[0], &intval);
770 pr_err("bad mount option arg (not int) at '%s'\n", c);
773 dout("got int token %d val %d\n", token, intval);
774 } else if (token > Opt_last_int && token < Opt_last_string) {
775 dout("got string token %d val %s\n", token, argstr[0].from);
777 dout("got token %d\n", token);
781 case Opt_queue_depth:
783 pr_err("queue_depth out of range\n");
786 rbd_opts->queue_depth = intval;
789 rbd_opts->read_only = true;
792 rbd_opts->read_only = false;
795 /* libceph prints "bad option" msg */
802 static char* obj_op_name(enum obj_operation_type op_type)
817 * Get a ceph client with specific addr and configuration, if one does
818 * not exist create it. Either way, ceph_opts is consumed by this
821 static struct rbd_client *rbd_get_client(struct ceph_options *ceph_opts)
823 struct rbd_client *rbdc;
825 mutex_lock_nested(&client_mutex, SINGLE_DEPTH_NESTING);
826 rbdc = rbd_client_find(ceph_opts);
827 if (rbdc) /* using an existing client */
828 ceph_destroy_options(ceph_opts);
830 rbdc = rbd_client_create(ceph_opts);
831 mutex_unlock(&client_mutex);
837 * Destroy ceph client
839 * Caller must hold rbd_client_list_lock.
841 static void rbd_client_release(struct kref *kref)
843 struct rbd_client *rbdc = container_of(kref, struct rbd_client, kref);
845 dout("%s: rbdc %p\n", __func__, rbdc);
846 spin_lock(&rbd_client_list_lock);
847 list_del(&rbdc->node);
848 spin_unlock(&rbd_client_list_lock);
850 ceph_destroy_client(rbdc->client);
855 * Drop reference to ceph client node. If it's not referenced anymore, release
858 static void rbd_put_client(struct rbd_client *rbdc)
861 kref_put(&rbdc->kref, rbd_client_release);
864 static bool rbd_image_format_valid(u32 image_format)
866 return image_format == 1 || image_format == 2;
869 static bool rbd_dev_ondisk_valid(struct rbd_image_header_ondisk *ondisk)
874 /* The header has to start with the magic rbd header text */
875 if (memcmp(&ondisk->text, RBD_HEADER_TEXT, sizeof (RBD_HEADER_TEXT)))
878 /* The bio layer requires at least sector-sized I/O */
880 if (ondisk->options.order < SECTOR_SHIFT)
883 /* If we use u64 in a few spots we may be able to loosen this */
885 if (ondisk->options.order > 8 * sizeof (int) - 1)
889 * The size of a snapshot header has to fit in a size_t, and
890 * that limits the number of snapshots.
892 snap_count = le32_to_cpu(ondisk->snap_count);
893 size = SIZE_MAX - sizeof (struct ceph_snap_context);
894 if (snap_count > size / sizeof (__le64))
898 * Not only that, but the size of the entire the snapshot
899 * header must also be representable in a size_t.
901 size -= snap_count * sizeof (__le64);
902 if ((u64) size < le64_to_cpu(ondisk->snap_names_len))
909 * Fill an rbd image header with information from the given format 1
912 static int rbd_header_from_disk(struct rbd_device *rbd_dev,
913 struct rbd_image_header_ondisk *ondisk)
915 struct rbd_image_header *header = &rbd_dev->header;
916 bool first_time = header->object_prefix == NULL;
917 struct ceph_snap_context *snapc;
918 char *object_prefix = NULL;
919 char *snap_names = NULL;
920 u64 *snap_sizes = NULL;
926 /* Allocate this now to avoid having to handle failure below */
931 len = strnlen(ondisk->object_prefix,
932 sizeof (ondisk->object_prefix));
933 object_prefix = kmalloc(len + 1, GFP_KERNEL);
936 memcpy(object_prefix, ondisk->object_prefix, len);
937 object_prefix[len] = '\0';
940 /* Allocate the snapshot context and fill it in */
942 snap_count = le32_to_cpu(ondisk->snap_count);
943 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
946 snapc->seq = le64_to_cpu(ondisk->snap_seq);
948 struct rbd_image_snap_ondisk *snaps;
949 u64 snap_names_len = le64_to_cpu(ondisk->snap_names_len);
951 /* We'll keep a copy of the snapshot names... */
953 if (snap_names_len > (u64)SIZE_MAX)
955 snap_names = kmalloc(snap_names_len, GFP_KERNEL);
959 /* ...as well as the array of their sizes. */
961 size = snap_count * sizeof (*header->snap_sizes);
962 snap_sizes = kmalloc(size, GFP_KERNEL);
967 * Copy the names, and fill in each snapshot's id
970 * Note that rbd_dev_v1_header_info() guarantees the
971 * ondisk buffer we're working with has
972 * snap_names_len bytes beyond the end of the
973 * snapshot id array, this memcpy() is safe.
975 memcpy(snap_names, &ondisk->snaps[snap_count], snap_names_len);
976 snaps = ondisk->snaps;
977 for (i = 0; i < snap_count; i++) {
978 snapc->snaps[i] = le64_to_cpu(snaps[i].id);
979 snap_sizes[i] = le64_to_cpu(snaps[i].image_size);
983 /* We won't fail any more, fill in the header */
986 header->object_prefix = object_prefix;
987 header->obj_order = ondisk->options.order;
988 header->crypt_type = ondisk->options.crypt_type;
989 header->comp_type = ondisk->options.comp_type;
990 /* The rest aren't used for format 1 images */
991 header->stripe_unit = 0;
992 header->stripe_count = 0;
993 header->features = 0;
995 ceph_put_snap_context(header->snapc);
996 kfree(header->snap_names);
997 kfree(header->snap_sizes);
1000 /* The remaining fields always get updated (when we refresh) */
1002 header->image_size = le64_to_cpu(ondisk->image_size);
1003 header->snapc = snapc;
1004 header->snap_names = snap_names;
1005 header->snap_sizes = snap_sizes;
1013 ceph_put_snap_context(snapc);
1014 kfree(object_prefix);
1019 static const char *_rbd_dev_v1_snap_name(struct rbd_device *rbd_dev, u32 which)
1021 const char *snap_name;
1023 rbd_assert(which < rbd_dev->header.snapc->num_snaps);
1025 /* Skip over names until we find the one we are looking for */
1027 snap_name = rbd_dev->header.snap_names;
1029 snap_name += strlen(snap_name) + 1;
1031 return kstrdup(snap_name, GFP_KERNEL);
1035 * Snapshot id comparison function for use with qsort()/bsearch().
1036 * Note that result is for snapshots in *descending* order.
1038 static int snapid_compare_reverse(const void *s1, const void *s2)
1040 u64 snap_id1 = *(u64 *)s1;
1041 u64 snap_id2 = *(u64 *)s2;
1043 if (snap_id1 < snap_id2)
1045 return snap_id1 == snap_id2 ? 0 : -1;
1049 * Search a snapshot context to see if the given snapshot id is
1052 * Returns the position of the snapshot id in the array if it's found,
1053 * or BAD_SNAP_INDEX otherwise.
1055 * Note: The snapshot array is in kept sorted (by the osd) in
1056 * reverse order, highest snapshot id first.
1058 static u32 rbd_dev_snap_index(struct rbd_device *rbd_dev, u64 snap_id)
1060 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
1063 found = bsearch(&snap_id, &snapc->snaps, snapc->num_snaps,
1064 sizeof (snap_id), snapid_compare_reverse);
1066 return found ? (u32)(found - &snapc->snaps[0]) : BAD_SNAP_INDEX;
1069 static const char *rbd_dev_v1_snap_name(struct rbd_device *rbd_dev,
1073 const char *snap_name;
1075 which = rbd_dev_snap_index(rbd_dev, snap_id);
1076 if (which == BAD_SNAP_INDEX)
1077 return ERR_PTR(-ENOENT);
1079 snap_name = _rbd_dev_v1_snap_name(rbd_dev, which);
1080 return snap_name ? snap_name : ERR_PTR(-ENOMEM);
1083 static const char *rbd_snap_name(struct rbd_device *rbd_dev, u64 snap_id)
1085 if (snap_id == CEPH_NOSNAP)
1086 return RBD_SNAP_HEAD_NAME;
1088 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1089 if (rbd_dev->image_format == 1)
1090 return rbd_dev_v1_snap_name(rbd_dev, snap_id);
1092 return rbd_dev_v2_snap_name(rbd_dev, snap_id);
1095 static int rbd_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
1098 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1099 if (snap_id == CEPH_NOSNAP) {
1100 *snap_size = rbd_dev->header.image_size;
1101 } else if (rbd_dev->image_format == 1) {
1104 which = rbd_dev_snap_index(rbd_dev, snap_id);
1105 if (which == BAD_SNAP_INDEX)
1108 *snap_size = rbd_dev->header.snap_sizes[which];
1113 ret = _rbd_dev_v2_snap_size(rbd_dev, snap_id, NULL, &size);
1122 static int rbd_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
1125 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
1126 if (snap_id == CEPH_NOSNAP) {
1127 *snap_features = rbd_dev->header.features;
1128 } else if (rbd_dev->image_format == 1) {
1129 *snap_features = 0; /* No features for format 1 */
1134 ret = _rbd_dev_v2_snap_features(rbd_dev, snap_id, &features);
1138 *snap_features = features;
1143 static int rbd_dev_mapping_set(struct rbd_device *rbd_dev)
1145 u64 snap_id = rbd_dev->spec->snap_id;
1150 ret = rbd_snap_size(rbd_dev, snap_id, &size);
1153 ret = rbd_snap_features(rbd_dev, snap_id, &features);
1157 rbd_dev->mapping.size = size;
1158 rbd_dev->mapping.features = features;
1163 static void rbd_dev_mapping_clear(struct rbd_device *rbd_dev)
1165 rbd_dev->mapping.size = 0;
1166 rbd_dev->mapping.features = 0;
1169 static void rbd_segment_name_free(const char *name)
1171 /* The explicit cast here is needed to drop the const qualifier */
1173 kmem_cache_free(rbd_segment_name_cache, (void *)name);
1176 static const char *rbd_segment_name(struct rbd_device *rbd_dev, u64 offset)
1183 name = kmem_cache_alloc(rbd_segment_name_cache, GFP_NOIO);
1186 segment = offset >> rbd_dev->header.obj_order;
1187 name_format = "%s.%012llx";
1188 if (rbd_dev->image_format == 2)
1189 name_format = "%s.%016llx";
1190 ret = snprintf(name, CEPH_MAX_OID_NAME_LEN + 1, name_format,
1191 rbd_dev->header.object_prefix, segment);
1192 if (ret < 0 || ret > CEPH_MAX_OID_NAME_LEN) {
1193 pr_err("error formatting segment name for #%llu (%d)\n",
1195 rbd_segment_name_free(name);
1202 static u64 rbd_segment_offset(struct rbd_device *rbd_dev, u64 offset)
1204 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1206 return offset & (segment_size - 1);
1209 static u64 rbd_segment_length(struct rbd_device *rbd_dev,
1210 u64 offset, u64 length)
1212 u64 segment_size = (u64) 1 << rbd_dev->header.obj_order;
1214 offset &= segment_size - 1;
1216 rbd_assert(length <= U64_MAX - offset);
1217 if (offset + length > segment_size)
1218 length = segment_size - offset;
1224 * returns the size of an object in the image
1226 static u64 rbd_obj_bytes(struct rbd_image_header *header)
1228 return 1 << header->obj_order;
1235 static void bio_chain_put(struct bio *chain)
1241 chain = chain->bi_next;
1247 * zeros a bio chain, starting at specific offset
1249 static void zero_bio_chain(struct bio *chain, int start_ofs)
1252 struct bvec_iter iter;
1253 unsigned long flags;
1258 bio_for_each_segment(bv, chain, iter) {
1259 if (pos + bv.bv_len > start_ofs) {
1260 int remainder = max(start_ofs - pos, 0);
1261 buf = bvec_kmap_irq(&bv, &flags);
1262 memset(buf + remainder, 0,
1263 bv.bv_len - remainder);
1264 flush_dcache_page(bv.bv_page);
1265 bvec_kunmap_irq(buf, &flags);
1270 chain = chain->bi_next;
1275 * similar to zero_bio_chain(), zeros data defined by a page array,
1276 * starting at the given byte offset from the start of the array and
1277 * continuing up to the given end offset. The pages array is
1278 * assumed to be big enough to hold all bytes up to the end.
1280 static void zero_pages(struct page **pages, u64 offset, u64 end)
1282 struct page **page = &pages[offset >> PAGE_SHIFT];
1284 rbd_assert(end > offset);
1285 rbd_assert(end - offset <= (u64)SIZE_MAX);
1286 while (offset < end) {
1289 unsigned long flags;
1292 page_offset = offset & ~PAGE_MASK;
1293 length = min_t(size_t, PAGE_SIZE - page_offset, end - offset);
1294 local_irq_save(flags);
1295 kaddr = kmap_atomic(*page);
1296 memset(kaddr + page_offset, 0, length);
1297 flush_dcache_page(*page);
1298 kunmap_atomic(kaddr);
1299 local_irq_restore(flags);
1307 * Clone a portion of a bio, starting at the given byte offset
1308 * and continuing for the number of bytes indicated.
1310 static struct bio *bio_clone_range(struct bio *bio_src,
1311 unsigned int offset,
1317 bio = bio_clone(bio_src, gfpmask);
1319 return NULL; /* ENOMEM */
1321 bio_advance(bio, offset);
1322 bio->bi_iter.bi_size = len;
1328 * Clone a portion of a bio chain, starting at the given byte offset
1329 * into the first bio in the source chain and continuing for the
1330 * number of bytes indicated. The result is another bio chain of
1331 * exactly the given length, or a null pointer on error.
1333 * The bio_src and offset parameters are both in-out. On entry they
1334 * refer to the first source bio and the offset into that bio where
1335 * the start of data to be cloned is located.
1337 * On return, bio_src is updated to refer to the bio in the source
1338 * chain that contains first un-cloned byte, and *offset will
1339 * contain the offset of that byte within that bio.
1341 static struct bio *bio_chain_clone_range(struct bio **bio_src,
1342 unsigned int *offset,
1346 struct bio *bi = *bio_src;
1347 unsigned int off = *offset;
1348 struct bio *chain = NULL;
1351 /* Build up a chain of clone bios up to the limit */
1353 if (!bi || off >= bi->bi_iter.bi_size || !len)
1354 return NULL; /* Nothing to clone */
1358 unsigned int bi_size;
1362 rbd_warn(NULL, "bio_chain exhausted with %u left", len);
1363 goto out_err; /* EINVAL; ran out of bio's */
1365 bi_size = min_t(unsigned int, bi->bi_iter.bi_size - off, len);
1366 bio = bio_clone_range(bi, off, bi_size, gfpmask);
1368 goto out_err; /* ENOMEM */
1371 end = &bio->bi_next;
1374 if (off == bi->bi_iter.bi_size) {
1385 bio_chain_put(chain);
1391 * The default/initial value for all object request flags is 0. For
1392 * each flag, once its value is set to 1 it is never reset to 0
1395 static void obj_request_img_data_set(struct rbd_obj_request *obj_request)
1397 if (test_and_set_bit(OBJ_REQ_IMG_DATA, &obj_request->flags)) {
1398 struct rbd_device *rbd_dev;
1400 rbd_dev = obj_request->img_request->rbd_dev;
1401 rbd_warn(rbd_dev, "obj_request %p already marked img_data",
1406 static bool obj_request_img_data_test(struct rbd_obj_request *obj_request)
1409 return test_bit(OBJ_REQ_IMG_DATA, &obj_request->flags) != 0;
1412 static void obj_request_done_set(struct rbd_obj_request *obj_request)
1414 if (test_and_set_bit(OBJ_REQ_DONE, &obj_request->flags)) {
1415 struct rbd_device *rbd_dev = NULL;
1417 if (obj_request_img_data_test(obj_request))
1418 rbd_dev = obj_request->img_request->rbd_dev;
1419 rbd_warn(rbd_dev, "obj_request %p already marked done",
1424 static bool obj_request_done_test(struct rbd_obj_request *obj_request)
1427 return test_bit(OBJ_REQ_DONE, &obj_request->flags) != 0;
1431 * This sets the KNOWN flag after (possibly) setting the EXISTS
1432 * flag. The latter is set based on the "exists" value provided.
1434 * Note that for our purposes once an object exists it never goes
1435 * away again. It's possible that the response from two existence
1436 * checks are separated by the creation of the target object, and
1437 * the first ("doesn't exist") response arrives *after* the second
1438 * ("does exist"). In that case we ignore the second one.
1440 static void obj_request_existence_set(struct rbd_obj_request *obj_request,
1444 set_bit(OBJ_REQ_EXISTS, &obj_request->flags);
1445 set_bit(OBJ_REQ_KNOWN, &obj_request->flags);
1449 static bool obj_request_known_test(struct rbd_obj_request *obj_request)
1452 return test_bit(OBJ_REQ_KNOWN, &obj_request->flags) != 0;
1455 static bool obj_request_exists_test(struct rbd_obj_request *obj_request)
1458 return test_bit(OBJ_REQ_EXISTS, &obj_request->flags) != 0;
1461 static bool obj_request_overlaps_parent(struct rbd_obj_request *obj_request)
1463 struct rbd_device *rbd_dev = obj_request->img_request->rbd_dev;
1465 return obj_request->img_offset <
1466 round_up(rbd_dev->parent_overlap, rbd_obj_bytes(&rbd_dev->header));
1469 static void rbd_obj_request_get(struct rbd_obj_request *obj_request)
1471 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1472 atomic_read(&obj_request->kref.refcount));
1473 kref_get(&obj_request->kref);
1476 static void rbd_obj_request_destroy(struct kref *kref);
1477 static void rbd_obj_request_put(struct rbd_obj_request *obj_request)
1479 rbd_assert(obj_request != NULL);
1480 dout("%s: obj %p (was %d)\n", __func__, obj_request,
1481 atomic_read(&obj_request->kref.refcount));
1482 kref_put(&obj_request->kref, rbd_obj_request_destroy);
1485 static void rbd_img_request_get(struct rbd_img_request *img_request)
1487 dout("%s: img %p (was %d)\n", __func__, img_request,
1488 atomic_read(&img_request->kref.refcount));
1489 kref_get(&img_request->kref);
1492 static bool img_request_child_test(struct rbd_img_request *img_request);
1493 static void rbd_parent_request_destroy(struct kref *kref);
1494 static void rbd_img_request_destroy(struct kref *kref);
1495 static void rbd_img_request_put(struct rbd_img_request *img_request)
1497 rbd_assert(img_request != NULL);
1498 dout("%s: img %p (was %d)\n", __func__, img_request,
1499 atomic_read(&img_request->kref.refcount));
1500 if (img_request_child_test(img_request))
1501 kref_put(&img_request->kref, rbd_parent_request_destroy);
1503 kref_put(&img_request->kref, rbd_img_request_destroy);
1506 static inline void rbd_img_obj_request_add(struct rbd_img_request *img_request,
1507 struct rbd_obj_request *obj_request)
1509 rbd_assert(obj_request->img_request == NULL);
1511 /* Image request now owns object's original reference */
1512 obj_request->img_request = img_request;
1513 obj_request->which = img_request->obj_request_count;
1514 rbd_assert(!obj_request_img_data_test(obj_request));
1515 obj_request_img_data_set(obj_request);
1516 rbd_assert(obj_request->which != BAD_WHICH);
1517 img_request->obj_request_count++;
1518 list_add_tail(&obj_request->links, &img_request->obj_requests);
1519 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1520 obj_request->which);
1523 static inline void rbd_img_obj_request_del(struct rbd_img_request *img_request,
1524 struct rbd_obj_request *obj_request)
1526 rbd_assert(obj_request->which != BAD_WHICH);
1528 dout("%s: img %p obj %p w=%u\n", __func__, img_request, obj_request,
1529 obj_request->which);
1530 list_del(&obj_request->links);
1531 rbd_assert(img_request->obj_request_count > 0);
1532 img_request->obj_request_count--;
1533 rbd_assert(obj_request->which == img_request->obj_request_count);
1534 obj_request->which = BAD_WHICH;
1535 rbd_assert(obj_request_img_data_test(obj_request));
1536 rbd_assert(obj_request->img_request == img_request);
1537 obj_request->img_request = NULL;
1538 obj_request->callback = NULL;
1539 rbd_obj_request_put(obj_request);
1542 static bool obj_request_type_valid(enum obj_request_type type)
1545 case OBJ_REQUEST_NODATA:
1546 case OBJ_REQUEST_BIO:
1547 case OBJ_REQUEST_PAGES:
1554 static int rbd_obj_request_submit(struct ceph_osd_client *osdc,
1555 struct rbd_obj_request *obj_request)
1557 dout("%s %p\n", __func__, obj_request);
1558 return ceph_osdc_start_request(osdc, obj_request->osd_req, false);
1561 static void rbd_obj_request_end(struct rbd_obj_request *obj_request)
1563 dout("%s %p\n", __func__, obj_request);
1564 ceph_osdc_cancel_request(obj_request->osd_req);
1568 * Wait for an object request to complete. If interrupted, cancel the
1569 * underlying osd request.
1571 * @timeout: in jiffies, 0 means "wait forever"
1573 static int __rbd_obj_request_wait(struct rbd_obj_request *obj_request,
1574 unsigned long timeout)
1578 dout("%s %p\n", __func__, obj_request);
1579 ret = wait_for_completion_interruptible_timeout(
1580 &obj_request->completion,
1581 ceph_timeout_jiffies(timeout));
1585 rbd_obj_request_end(obj_request);
1590 dout("%s %p ret %d\n", __func__, obj_request, (int)ret);
1594 static int rbd_obj_request_wait(struct rbd_obj_request *obj_request)
1596 return __rbd_obj_request_wait(obj_request, 0);
1599 static int rbd_obj_request_wait_timeout(struct rbd_obj_request *obj_request,
1600 unsigned long timeout)
1602 return __rbd_obj_request_wait(obj_request, timeout);
1605 static void rbd_img_request_complete(struct rbd_img_request *img_request)
1608 dout("%s: img %p\n", __func__, img_request);
1611 * If no error occurred, compute the aggregate transfer
1612 * count for the image request. We could instead use
1613 * atomic64_cmpxchg() to update it as each object request
1614 * completes; not clear which way is better off hand.
1616 if (!img_request->result) {
1617 struct rbd_obj_request *obj_request;
1620 for_each_obj_request(img_request, obj_request)
1621 xferred += obj_request->xferred;
1622 img_request->xferred = xferred;
1625 if (img_request->callback)
1626 img_request->callback(img_request);
1628 rbd_img_request_put(img_request);
1632 * The default/initial value for all image request flags is 0. Each
1633 * is conditionally set to 1 at image request initialization time
1634 * and currently never change thereafter.
1636 static void img_request_write_set(struct rbd_img_request *img_request)
1638 set_bit(IMG_REQ_WRITE, &img_request->flags);
1642 static bool img_request_write_test(struct rbd_img_request *img_request)
1645 return test_bit(IMG_REQ_WRITE, &img_request->flags) != 0;
1649 * Set the discard flag when the img_request is an discard request
1651 static void img_request_discard_set(struct rbd_img_request *img_request)
1653 set_bit(IMG_REQ_DISCARD, &img_request->flags);
1657 static bool img_request_discard_test(struct rbd_img_request *img_request)
1660 return test_bit(IMG_REQ_DISCARD, &img_request->flags) != 0;
1663 static void img_request_child_set(struct rbd_img_request *img_request)
1665 set_bit(IMG_REQ_CHILD, &img_request->flags);
1669 static void img_request_child_clear(struct rbd_img_request *img_request)
1671 clear_bit(IMG_REQ_CHILD, &img_request->flags);
1675 static bool img_request_child_test(struct rbd_img_request *img_request)
1678 return test_bit(IMG_REQ_CHILD, &img_request->flags) != 0;
1681 static void img_request_layered_set(struct rbd_img_request *img_request)
1683 set_bit(IMG_REQ_LAYERED, &img_request->flags);
1687 static void img_request_layered_clear(struct rbd_img_request *img_request)
1689 clear_bit(IMG_REQ_LAYERED, &img_request->flags);
1693 static bool img_request_layered_test(struct rbd_img_request *img_request)
1696 return test_bit(IMG_REQ_LAYERED, &img_request->flags) != 0;
1699 static enum obj_operation_type
1700 rbd_img_request_op_type(struct rbd_img_request *img_request)
1702 if (img_request_write_test(img_request))
1703 return OBJ_OP_WRITE;
1704 else if (img_request_discard_test(img_request))
1705 return OBJ_OP_DISCARD;
1711 rbd_img_obj_request_read_callback(struct rbd_obj_request *obj_request)
1713 u64 xferred = obj_request->xferred;
1714 u64 length = obj_request->length;
1716 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1717 obj_request, obj_request->img_request, obj_request->result,
1720 * ENOENT means a hole in the image. We zero-fill the entire
1721 * length of the request. A short read also implies zero-fill
1722 * to the end of the request. An error requires the whole
1723 * length of the request to be reported finished with an error
1724 * to the block layer. In each case we update the xferred
1725 * count to indicate the whole request was satisfied.
1727 rbd_assert(obj_request->type != OBJ_REQUEST_NODATA);
1728 if (obj_request->result == -ENOENT) {
1729 if (obj_request->type == OBJ_REQUEST_BIO)
1730 zero_bio_chain(obj_request->bio_list, 0);
1732 zero_pages(obj_request->pages, 0, length);
1733 obj_request->result = 0;
1734 } else if (xferred < length && !obj_request->result) {
1735 if (obj_request->type == OBJ_REQUEST_BIO)
1736 zero_bio_chain(obj_request->bio_list, xferred);
1738 zero_pages(obj_request->pages, xferred, length);
1740 obj_request->xferred = length;
1741 obj_request_done_set(obj_request);
1744 static void rbd_obj_request_complete(struct rbd_obj_request *obj_request)
1746 dout("%s: obj %p cb %p\n", __func__, obj_request,
1747 obj_request->callback);
1748 if (obj_request->callback)
1749 obj_request->callback(obj_request);
1751 complete_all(&obj_request->completion);
1754 static void rbd_osd_trivial_callback(struct rbd_obj_request *obj_request)
1756 dout("%s: obj %p\n", __func__, obj_request);
1757 obj_request_done_set(obj_request);
1760 static void rbd_osd_read_callback(struct rbd_obj_request *obj_request)
1762 struct rbd_img_request *img_request = NULL;
1763 struct rbd_device *rbd_dev = NULL;
1764 bool layered = false;
1766 if (obj_request_img_data_test(obj_request)) {
1767 img_request = obj_request->img_request;
1768 layered = img_request && img_request_layered_test(img_request);
1769 rbd_dev = img_request->rbd_dev;
1772 dout("%s: obj %p img %p result %d %llu/%llu\n", __func__,
1773 obj_request, img_request, obj_request->result,
1774 obj_request->xferred, obj_request->length);
1775 if (layered && obj_request->result == -ENOENT &&
1776 obj_request->img_offset < rbd_dev->parent_overlap)
1777 rbd_img_parent_read(obj_request);
1778 else if (img_request)
1779 rbd_img_obj_request_read_callback(obj_request);
1781 obj_request_done_set(obj_request);
1784 static void rbd_osd_write_callback(struct rbd_obj_request *obj_request)
1786 dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1787 obj_request->result, obj_request->length);
1789 * There is no such thing as a successful short write. Set
1790 * it to our originally-requested length.
1792 obj_request->xferred = obj_request->length;
1793 obj_request_done_set(obj_request);
1796 static void rbd_osd_discard_callback(struct rbd_obj_request *obj_request)
1798 dout("%s: obj %p result %d %llu\n", __func__, obj_request,
1799 obj_request->result, obj_request->length);
1801 * There is no such thing as a successful short discard. Set
1802 * it to our originally-requested length.
1804 obj_request->xferred = obj_request->length;
1805 /* discarding a non-existent object is not a problem */
1806 if (obj_request->result == -ENOENT)
1807 obj_request->result = 0;
1808 obj_request_done_set(obj_request);
1812 * For a simple stat call there's nothing to do. We'll do more if
1813 * this is part of a write sequence for a layered image.
1815 static void rbd_osd_stat_callback(struct rbd_obj_request *obj_request)
1817 dout("%s: obj %p\n", __func__, obj_request);
1818 obj_request_done_set(obj_request);
1821 static void rbd_osd_call_callback(struct rbd_obj_request *obj_request)
1823 dout("%s: obj %p\n", __func__, obj_request);
1825 if (obj_request_img_data_test(obj_request))
1826 rbd_osd_copyup_callback(obj_request);
1828 obj_request_done_set(obj_request);
1831 static void rbd_osd_req_callback(struct ceph_osd_request *osd_req)
1833 struct rbd_obj_request *obj_request = osd_req->r_priv;
1836 dout("%s: osd_req %p\n", __func__, osd_req);
1837 rbd_assert(osd_req == obj_request->osd_req);
1838 if (obj_request_img_data_test(obj_request)) {
1839 rbd_assert(obj_request->img_request);
1840 rbd_assert(obj_request->which != BAD_WHICH);
1842 rbd_assert(obj_request->which == BAD_WHICH);
1845 if (osd_req->r_result < 0)
1846 obj_request->result = osd_req->r_result;
1849 * We support a 64-bit length, but ultimately it has to be
1850 * passed to the block layer, which just supports a 32-bit
1853 obj_request->xferred = osd_req->r_ops[0].outdata_len;
1854 rbd_assert(obj_request->xferred < (u64)UINT_MAX);
1856 opcode = osd_req->r_ops[0].op;
1858 case CEPH_OSD_OP_READ:
1859 rbd_osd_read_callback(obj_request);
1861 case CEPH_OSD_OP_SETALLOCHINT:
1862 rbd_assert(osd_req->r_ops[1].op == CEPH_OSD_OP_WRITE ||
1863 osd_req->r_ops[1].op == CEPH_OSD_OP_WRITEFULL);
1865 case CEPH_OSD_OP_WRITE:
1866 case CEPH_OSD_OP_WRITEFULL:
1867 rbd_osd_write_callback(obj_request);
1869 case CEPH_OSD_OP_STAT:
1870 rbd_osd_stat_callback(obj_request);
1872 case CEPH_OSD_OP_DELETE:
1873 case CEPH_OSD_OP_TRUNCATE:
1874 case CEPH_OSD_OP_ZERO:
1875 rbd_osd_discard_callback(obj_request);
1877 case CEPH_OSD_OP_CALL:
1878 rbd_osd_call_callback(obj_request);
1880 case CEPH_OSD_OP_NOTIFY_ACK:
1881 case CEPH_OSD_OP_WATCH:
1882 rbd_osd_trivial_callback(obj_request);
1885 rbd_warn(NULL, "%s: unsupported op %hu",
1886 obj_request->object_name, (unsigned short) opcode);
1890 if (obj_request_done_test(obj_request))
1891 rbd_obj_request_complete(obj_request);
1894 static void rbd_osd_req_format_read(struct rbd_obj_request *obj_request)
1896 struct rbd_img_request *img_request = obj_request->img_request;
1897 struct ceph_osd_request *osd_req = obj_request->osd_req;
1900 osd_req->r_snapid = img_request->snap_id;
1903 static void rbd_osd_req_format_write(struct rbd_obj_request *obj_request)
1905 struct ceph_osd_request *osd_req = obj_request->osd_req;
1907 osd_req->r_mtime = CURRENT_TIME;
1908 osd_req->r_data_offset = obj_request->offset;
1912 * Create an osd request. A read request has one osd op (read).
1913 * A write request has either one (watch) or two (hint+write) osd ops.
1914 * (All rbd data writes are prefixed with an allocation hint op, but
1915 * technically osd watch is a write request, hence this distinction.)
1917 static struct ceph_osd_request *rbd_osd_req_create(
1918 struct rbd_device *rbd_dev,
1919 enum obj_operation_type op_type,
1920 unsigned int num_ops,
1921 struct rbd_obj_request *obj_request)
1923 struct ceph_snap_context *snapc = NULL;
1924 struct ceph_osd_client *osdc;
1925 struct ceph_osd_request *osd_req;
1927 if (obj_request_img_data_test(obj_request) &&
1928 (op_type == OBJ_OP_DISCARD || op_type == OBJ_OP_WRITE)) {
1929 struct rbd_img_request *img_request = obj_request->img_request;
1930 if (op_type == OBJ_OP_WRITE) {
1931 rbd_assert(img_request_write_test(img_request));
1933 rbd_assert(img_request_discard_test(img_request));
1935 snapc = img_request->snapc;
1938 rbd_assert(num_ops == 1 || ((op_type == OBJ_OP_WRITE) && num_ops == 2));
1940 /* Allocate and initialize the request, for the num_ops ops */
1942 osdc = &rbd_dev->rbd_client->client->osdc;
1943 osd_req = ceph_osdc_alloc_request(osdc, snapc, num_ops, false,
1948 if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
1949 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
1951 osd_req->r_flags = CEPH_OSD_FLAG_READ;
1953 osd_req->r_callback = rbd_osd_req_callback;
1954 osd_req->r_priv = obj_request;
1956 osd_req->r_base_oloc.pool = ceph_file_layout_pg_pool(rbd_dev->layout);
1957 if (ceph_oid_aprintf(&osd_req->r_base_oid, GFP_NOIO, "%s",
1958 obj_request->object_name))
1961 if (ceph_osdc_alloc_messages(osd_req, GFP_NOIO))
1967 ceph_osdc_put_request(osd_req);
1972 * Create a copyup osd request based on the information in the object
1973 * request supplied. A copyup request has two or three osd ops, a
1974 * copyup method call, potentially a hint op, and a write or truncate
1977 static struct ceph_osd_request *
1978 rbd_osd_req_create_copyup(struct rbd_obj_request *obj_request)
1980 struct rbd_img_request *img_request;
1981 struct ceph_snap_context *snapc;
1982 struct rbd_device *rbd_dev;
1983 struct ceph_osd_client *osdc;
1984 struct ceph_osd_request *osd_req;
1985 int num_osd_ops = 3;
1987 rbd_assert(obj_request_img_data_test(obj_request));
1988 img_request = obj_request->img_request;
1989 rbd_assert(img_request);
1990 rbd_assert(img_request_write_test(img_request) ||
1991 img_request_discard_test(img_request));
1993 if (img_request_discard_test(img_request))
1996 /* Allocate and initialize the request, for all the ops */
1998 snapc = img_request->snapc;
1999 rbd_dev = img_request->rbd_dev;
2000 osdc = &rbd_dev->rbd_client->client->osdc;
2001 osd_req = ceph_osdc_alloc_request(osdc, snapc, num_osd_ops,
2006 osd_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ONDISK;
2007 osd_req->r_callback = rbd_osd_req_callback;
2008 osd_req->r_priv = obj_request;
2010 osd_req->r_base_oloc.pool = ceph_file_layout_pg_pool(rbd_dev->layout);
2011 if (ceph_oid_aprintf(&osd_req->r_base_oid, GFP_NOIO, "%s",
2012 obj_request->object_name))
2015 if (ceph_osdc_alloc_messages(osd_req, GFP_NOIO))
2021 ceph_osdc_put_request(osd_req);
2026 static void rbd_osd_req_destroy(struct ceph_osd_request *osd_req)
2028 ceph_osdc_put_request(osd_req);
2031 /* object_name is assumed to be a non-null pointer and NUL-terminated */
2033 static struct rbd_obj_request *rbd_obj_request_create(const char *object_name,
2034 u64 offset, u64 length,
2035 enum obj_request_type type)
2037 struct rbd_obj_request *obj_request;
2041 rbd_assert(obj_request_type_valid(type));
2043 size = strlen(object_name) + 1;
2044 name = kmalloc(size, GFP_NOIO);
2048 obj_request = kmem_cache_zalloc(rbd_obj_request_cache, GFP_NOIO);
2054 obj_request->object_name = memcpy(name, object_name, size);
2055 obj_request->offset = offset;
2056 obj_request->length = length;
2057 obj_request->flags = 0;
2058 obj_request->which = BAD_WHICH;
2059 obj_request->type = type;
2060 INIT_LIST_HEAD(&obj_request->links);
2061 init_completion(&obj_request->completion);
2062 kref_init(&obj_request->kref);
2064 dout("%s: \"%s\" %llu/%llu %d -> obj %p\n", __func__, object_name,
2065 offset, length, (int)type, obj_request);
2070 static void rbd_obj_request_destroy(struct kref *kref)
2072 struct rbd_obj_request *obj_request;
2074 obj_request = container_of(kref, struct rbd_obj_request, kref);
2076 dout("%s: obj %p\n", __func__, obj_request);
2078 rbd_assert(obj_request->img_request == NULL);
2079 rbd_assert(obj_request->which == BAD_WHICH);
2081 if (obj_request->osd_req)
2082 rbd_osd_req_destroy(obj_request->osd_req);
2084 rbd_assert(obj_request_type_valid(obj_request->type));
2085 switch (obj_request->type) {
2086 case OBJ_REQUEST_NODATA:
2087 break; /* Nothing to do */
2088 case OBJ_REQUEST_BIO:
2089 if (obj_request->bio_list)
2090 bio_chain_put(obj_request->bio_list);
2092 case OBJ_REQUEST_PAGES:
2093 if (obj_request->pages)
2094 ceph_release_page_vector(obj_request->pages,
2095 obj_request->page_count);
2099 kfree(obj_request->object_name);
2100 obj_request->object_name = NULL;
2101 kmem_cache_free(rbd_obj_request_cache, obj_request);
2104 /* It's OK to call this for a device with no parent */
2106 static void rbd_spec_put(struct rbd_spec *spec);
2107 static void rbd_dev_unparent(struct rbd_device *rbd_dev)
2109 rbd_dev_remove_parent(rbd_dev);
2110 rbd_spec_put(rbd_dev->parent_spec);
2111 rbd_dev->parent_spec = NULL;
2112 rbd_dev->parent_overlap = 0;
2116 * Parent image reference counting is used to determine when an
2117 * image's parent fields can be safely torn down--after there are no
2118 * more in-flight requests to the parent image. When the last
2119 * reference is dropped, cleaning them up is safe.
2121 static void rbd_dev_parent_put(struct rbd_device *rbd_dev)
2125 if (!rbd_dev->parent_spec)
2128 counter = atomic_dec_return_safe(&rbd_dev->parent_ref);
2132 /* Last reference; clean up parent data structures */
2135 rbd_dev_unparent(rbd_dev);
2137 rbd_warn(rbd_dev, "parent reference underflow");
2141 * If an image has a non-zero parent overlap, get a reference to its
2144 * Returns true if the rbd device has a parent with a non-zero
2145 * overlap and a reference for it was successfully taken, or
2148 static bool rbd_dev_parent_get(struct rbd_device *rbd_dev)
2152 if (!rbd_dev->parent_spec)
2155 down_read(&rbd_dev->header_rwsem);
2156 if (rbd_dev->parent_overlap)
2157 counter = atomic_inc_return_safe(&rbd_dev->parent_ref);
2158 up_read(&rbd_dev->header_rwsem);
2161 rbd_warn(rbd_dev, "parent reference overflow");
2167 * Caller is responsible for filling in the list of object requests
2168 * that comprises the image request, and the Linux request pointer
2169 * (if there is one).
2171 static struct rbd_img_request *rbd_img_request_create(
2172 struct rbd_device *rbd_dev,
2173 u64 offset, u64 length,
2174 enum obj_operation_type op_type,
2175 struct ceph_snap_context *snapc)
2177 struct rbd_img_request *img_request;
2179 img_request = kmem_cache_alloc(rbd_img_request_cache, GFP_NOIO);
2183 img_request->rq = NULL;
2184 img_request->rbd_dev = rbd_dev;
2185 img_request->offset = offset;
2186 img_request->length = length;
2187 img_request->flags = 0;
2188 if (op_type == OBJ_OP_DISCARD) {
2189 img_request_discard_set(img_request);
2190 img_request->snapc = snapc;
2191 } else if (op_type == OBJ_OP_WRITE) {
2192 img_request_write_set(img_request);
2193 img_request->snapc = snapc;
2195 img_request->snap_id = rbd_dev->spec->snap_id;
2197 if (rbd_dev_parent_get(rbd_dev))
2198 img_request_layered_set(img_request);
2199 spin_lock_init(&img_request->completion_lock);
2200 img_request->next_completion = 0;
2201 img_request->callback = NULL;
2202 img_request->result = 0;
2203 img_request->obj_request_count = 0;
2204 INIT_LIST_HEAD(&img_request->obj_requests);
2205 kref_init(&img_request->kref);
2207 dout("%s: rbd_dev %p %s %llu/%llu -> img %p\n", __func__, rbd_dev,
2208 obj_op_name(op_type), offset, length, img_request);
2213 static void rbd_img_request_destroy(struct kref *kref)
2215 struct rbd_img_request *img_request;
2216 struct rbd_obj_request *obj_request;
2217 struct rbd_obj_request *next_obj_request;
2219 img_request = container_of(kref, struct rbd_img_request, kref);
2221 dout("%s: img %p\n", __func__, img_request);
2223 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2224 rbd_img_obj_request_del(img_request, obj_request);
2225 rbd_assert(img_request->obj_request_count == 0);
2227 if (img_request_layered_test(img_request)) {
2228 img_request_layered_clear(img_request);
2229 rbd_dev_parent_put(img_request->rbd_dev);
2232 if (img_request_write_test(img_request) ||
2233 img_request_discard_test(img_request))
2234 ceph_put_snap_context(img_request->snapc);
2236 kmem_cache_free(rbd_img_request_cache, img_request);
2239 static struct rbd_img_request *rbd_parent_request_create(
2240 struct rbd_obj_request *obj_request,
2241 u64 img_offset, u64 length)
2243 struct rbd_img_request *parent_request;
2244 struct rbd_device *rbd_dev;
2246 rbd_assert(obj_request->img_request);
2247 rbd_dev = obj_request->img_request->rbd_dev;
2249 parent_request = rbd_img_request_create(rbd_dev->parent, img_offset,
2250 length, OBJ_OP_READ, NULL);
2251 if (!parent_request)
2254 img_request_child_set(parent_request);
2255 rbd_obj_request_get(obj_request);
2256 parent_request->obj_request = obj_request;
2258 return parent_request;
2261 static void rbd_parent_request_destroy(struct kref *kref)
2263 struct rbd_img_request *parent_request;
2264 struct rbd_obj_request *orig_request;
2266 parent_request = container_of(kref, struct rbd_img_request, kref);
2267 orig_request = parent_request->obj_request;
2269 parent_request->obj_request = NULL;
2270 rbd_obj_request_put(orig_request);
2271 img_request_child_clear(parent_request);
2273 rbd_img_request_destroy(kref);
2276 static bool rbd_img_obj_end_request(struct rbd_obj_request *obj_request)
2278 struct rbd_img_request *img_request;
2279 unsigned int xferred;
2283 rbd_assert(obj_request_img_data_test(obj_request));
2284 img_request = obj_request->img_request;
2286 rbd_assert(obj_request->xferred <= (u64)UINT_MAX);
2287 xferred = (unsigned int)obj_request->xferred;
2288 result = obj_request->result;
2290 struct rbd_device *rbd_dev = img_request->rbd_dev;
2291 enum obj_operation_type op_type;
2293 if (img_request_discard_test(img_request))
2294 op_type = OBJ_OP_DISCARD;
2295 else if (img_request_write_test(img_request))
2296 op_type = OBJ_OP_WRITE;
2298 op_type = OBJ_OP_READ;
2300 rbd_warn(rbd_dev, "%s %llx at %llx (%llx)",
2301 obj_op_name(op_type), obj_request->length,
2302 obj_request->img_offset, obj_request->offset);
2303 rbd_warn(rbd_dev, " result %d xferred %x",
2305 if (!img_request->result)
2306 img_request->result = result;
2308 * Need to end I/O on the entire obj_request worth of
2309 * bytes in case of error.
2311 xferred = obj_request->length;
2314 /* Image object requests don't own their page array */
2316 if (obj_request->type == OBJ_REQUEST_PAGES) {
2317 obj_request->pages = NULL;
2318 obj_request->page_count = 0;
2321 if (img_request_child_test(img_request)) {
2322 rbd_assert(img_request->obj_request != NULL);
2323 more = obj_request->which < img_request->obj_request_count - 1;
2325 rbd_assert(img_request->rq != NULL);
2327 more = blk_update_request(img_request->rq, result, xferred);
2329 __blk_mq_end_request(img_request->rq, result);
2335 static void rbd_img_obj_callback(struct rbd_obj_request *obj_request)
2337 struct rbd_img_request *img_request;
2338 u32 which = obj_request->which;
2341 rbd_assert(obj_request_img_data_test(obj_request));
2342 img_request = obj_request->img_request;
2344 dout("%s: img %p obj %p\n", __func__, img_request, obj_request);
2345 rbd_assert(img_request != NULL);
2346 rbd_assert(img_request->obj_request_count > 0);
2347 rbd_assert(which != BAD_WHICH);
2348 rbd_assert(which < img_request->obj_request_count);
2350 spin_lock_irq(&img_request->completion_lock);
2351 if (which != img_request->next_completion)
2354 for_each_obj_request_from(img_request, obj_request) {
2356 rbd_assert(which < img_request->obj_request_count);
2358 if (!obj_request_done_test(obj_request))
2360 more = rbd_img_obj_end_request(obj_request);
2364 rbd_assert(more ^ (which == img_request->obj_request_count));
2365 img_request->next_completion = which;
2367 spin_unlock_irq(&img_request->completion_lock);
2368 rbd_img_request_put(img_request);
2371 rbd_img_request_complete(img_request);
2375 * Add individual osd ops to the given ceph_osd_request and prepare
2376 * them for submission. num_ops is the current number of
2377 * osd operations already to the object request.
2379 static void rbd_img_obj_request_fill(struct rbd_obj_request *obj_request,
2380 struct ceph_osd_request *osd_request,
2381 enum obj_operation_type op_type,
2382 unsigned int num_ops)
2384 struct rbd_img_request *img_request = obj_request->img_request;
2385 struct rbd_device *rbd_dev = img_request->rbd_dev;
2386 u64 object_size = rbd_obj_bytes(&rbd_dev->header);
2387 u64 offset = obj_request->offset;
2388 u64 length = obj_request->length;
2392 if (op_type == OBJ_OP_DISCARD) {
2393 if (!offset && length == object_size &&
2394 (!img_request_layered_test(img_request) ||
2395 !obj_request_overlaps_parent(obj_request))) {
2396 opcode = CEPH_OSD_OP_DELETE;
2397 } else if ((offset + length == object_size)) {
2398 opcode = CEPH_OSD_OP_TRUNCATE;
2400 down_read(&rbd_dev->header_rwsem);
2401 img_end = rbd_dev->header.image_size;
2402 up_read(&rbd_dev->header_rwsem);
2404 if (obj_request->img_offset + length == img_end)
2405 opcode = CEPH_OSD_OP_TRUNCATE;
2407 opcode = CEPH_OSD_OP_ZERO;
2409 } else if (op_type == OBJ_OP_WRITE) {
2410 if (!offset && length == object_size)
2411 opcode = CEPH_OSD_OP_WRITEFULL;
2413 opcode = CEPH_OSD_OP_WRITE;
2414 osd_req_op_alloc_hint_init(osd_request, num_ops,
2415 object_size, object_size);
2418 opcode = CEPH_OSD_OP_READ;
2421 if (opcode == CEPH_OSD_OP_DELETE)
2422 osd_req_op_init(osd_request, num_ops, opcode, 0);
2424 osd_req_op_extent_init(osd_request, num_ops, opcode,
2425 offset, length, 0, 0);
2427 if (obj_request->type == OBJ_REQUEST_BIO)
2428 osd_req_op_extent_osd_data_bio(osd_request, num_ops,
2429 obj_request->bio_list, length);
2430 else if (obj_request->type == OBJ_REQUEST_PAGES)
2431 osd_req_op_extent_osd_data_pages(osd_request, num_ops,
2432 obj_request->pages, length,
2433 offset & ~PAGE_MASK, false, false);
2435 /* Discards are also writes */
2436 if (op_type == OBJ_OP_WRITE || op_type == OBJ_OP_DISCARD)
2437 rbd_osd_req_format_write(obj_request);
2439 rbd_osd_req_format_read(obj_request);
2443 * Split up an image request into one or more object requests, each
2444 * to a different object. The "type" parameter indicates whether
2445 * "data_desc" is the pointer to the head of a list of bio
2446 * structures, or the base of a page array. In either case this
2447 * function assumes data_desc describes memory sufficient to hold
2448 * all data described by the image request.
2450 static int rbd_img_request_fill(struct rbd_img_request *img_request,
2451 enum obj_request_type type,
2454 struct rbd_device *rbd_dev = img_request->rbd_dev;
2455 struct rbd_obj_request *obj_request = NULL;
2456 struct rbd_obj_request *next_obj_request;
2457 struct bio *bio_list = NULL;
2458 unsigned int bio_offset = 0;
2459 struct page **pages = NULL;
2460 enum obj_operation_type op_type;
2464 dout("%s: img %p type %d data_desc %p\n", __func__, img_request,
2465 (int)type, data_desc);
2467 img_offset = img_request->offset;
2468 resid = img_request->length;
2469 rbd_assert(resid > 0);
2470 op_type = rbd_img_request_op_type(img_request);
2472 if (type == OBJ_REQUEST_BIO) {
2473 bio_list = data_desc;
2474 rbd_assert(img_offset ==
2475 bio_list->bi_iter.bi_sector << SECTOR_SHIFT);
2476 } else if (type == OBJ_REQUEST_PAGES) {
2481 struct ceph_osd_request *osd_req;
2482 const char *object_name;
2486 object_name = rbd_segment_name(rbd_dev, img_offset);
2489 offset = rbd_segment_offset(rbd_dev, img_offset);
2490 length = rbd_segment_length(rbd_dev, img_offset, resid);
2491 obj_request = rbd_obj_request_create(object_name,
2492 offset, length, type);
2493 /* object request has its own copy of the object name */
2494 rbd_segment_name_free(object_name);
2499 * set obj_request->img_request before creating the
2500 * osd_request so that it gets the right snapc
2502 rbd_img_obj_request_add(img_request, obj_request);
2504 if (type == OBJ_REQUEST_BIO) {
2505 unsigned int clone_size;
2507 rbd_assert(length <= (u64)UINT_MAX);
2508 clone_size = (unsigned int)length;
2509 obj_request->bio_list =
2510 bio_chain_clone_range(&bio_list,
2514 if (!obj_request->bio_list)
2516 } else if (type == OBJ_REQUEST_PAGES) {
2517 unsigned int page_count;
2519 obj_request->pages = pages;
2520 page_count = (u32)calc_pages_for(offset, length);
2521 obj_request->page_count = page_count;
2522 if ((offset + length) & ~PAGE_MASK)
2523 page_count--; /* more on last page */
2524 pages += page_count;
2527 osd_req = rbd_osd_req_create(rbd_dev, op_type,
2528 (op_type == OBJ_OP_WRITE) ? 2 : 1,
2533 obj_request->osd_req = osd_req;
2534 obj_request->callback = rbd_img_obj_callback;
2535 obj_request->img_offset = img_offset;
2537 rbd_img_obj_request_fill(obj_request, osd_req, op_type, 0);
2539 rbd_img_request_get(img_request);
2541 img_offset += length;
2548 for_each_obj_request_safe(img_request, obj_request, next_obj_request)
2549 rbd_img_obj_request_del(img_request, obj_request);
2555 rbd_osd_copyup_callback(struct rbd_obj_request *obj_request)
2557 struct rbd_img_request *img_request;
2558 struct rbd_device *rbd_dev;
2559 struct page **pages;
2562 dout("%s: obj %p\n", __func__, obj_request);
2564 rbd_assert(obj_request->type == OBJ_REQUEST_BIO ||
2565 obj_request->type == OBJ_REQUEST_NODATA);
2566 rbd_assert(obj_request_img_data_test(obj_request));
2567 img_request = obj_request->img_request;
2568 rbd_assert(img_request);
2570 rbd_dev = img_request->rbd_dev;
2571 rbd_assert(rbd_dev);
2573 pages = obj_request->copyup_pages;
2574 rbd_assert(pages != NULL);
2575 obj_request->copyup_pages = NULL;
2576 page_count = obj_request->copyup_page_count;
2577 rbd_assert(page_count);
2578 obj_request->copyup_page_count = 0;
2579 ceph_release_page_vector(pages, page_count);
2582 * We want the transfer count to reflect the size of the
2583 * original write request. There is no such thing as a
2584 * successful short write, so if the request was successful
2585 * we can just set it to the originally-requested length.
2587 if (!obj_request->result)
2588 obj_request->xferred = obj_request->length;
2590 obj_request_done_set(obj_request);
2594 rbd_img_obj_parent_read_full_callback(struct rbd_img_request *img_request)
2596 struct rbd_obj_request *orig_request;
2597 struct ceph_osd_request *osd_req;
2598 struct ceph_osd_client *osdc;
2599 struct rbd_device *rbd_dev;
2600 struct page **pages;
2601 enum obj_operation_type op_type;
2606 rbd_assert(img_request_child_test(img_request));
2608 /* First get what we need from the image request */
2610 pages = img_request->copyup_pages;
2611 rbd_assert(pages != NULL);
2612 img_request->copyup_pages = NULL;
2613 page_count = img_request->copyup_page_count;
2614 rbd_assert(page_count);
2615 img_request->copyup_page_count = 0;
2617 orig_request = img_request->obj_request;
2618 rbd_assert(orig_request != NULL);
2619 rbd_assert(obj_request_type_valid(orig_request->type));
2620 img_result = img_request->result;
2621 parent_length = img_request->length;
2622 rbd_assert(parent_length == img_request->xferred);
2623 rbd_img_request_put(img_request);
2625 rbd_assert(orig_request->img_request);
2626 rbd_dev = orig_request->img_request->rbd_dev;
2627 rbd_assert(rbd_dev);
2630 * If the overlap has become 0 (most likely because the
2631 * image has been flattened) we need to free the pages
2632 * and re-submit the original write request.
2634 if (!rbd_dev->parent_overlap) {
2635 struct ceph_osd_client *osdc;
2637 ceph_release_page_vector(pages, page_count);
2638 osdc = &rbd_dev->rbd_client->client->osdc;
2639 img_result = rbd_obj_request_submit(osdc, orig_request);
2648 * The original osd request is of no use to use any more.
2649 * We need a new one that can hold the three ops in a copyup
2650 * request. Allocate the new copyup osd request for the
2651 * original request, and release the old one.
2653 img_result = -ENOMEM;
2654 osd_req = rbd_osd_req_create_copyup(orig_request);
2657 rbd_osd_req_destroy(orig_request->osd_req);
2658 orig_request->osd_req = osd_req;
2659 orig_request->copyup_pages = pages;
2660 orig_request->copyup_page_count = page_count;
2662 /* Initialize the copyup op */
2664 osd_req_op_cls_init(osd_req, 0, CEPH_OSD_OP_CALL, "rbd", "copyup");
2665 osd_req_op_cls_request_data_pages(osd_req, 0, pages, parent_length, 0,
2668 /* Add the other op(s) */
2670 op_type = rbd_img_request_op_type(orig_request->img_request);
2671 rbd_img_obj_request_fill(orig_request, osd_req, op_type, 1);
2673 /* All set, send it off. */
2675 osdc = &rbd_dev->rbd_client->client->osdc;
2676 img_result = rbd_obj_request_submit(osdc, orig_request);
2680 /* Record the error code and complete the request */
2682 orig_request->result = img_result;
2683 orig_request->xferred = 0;
2684 obj_request_done_set(orig_request);
2685 rbd_obj_request_complete(orig_request);
2689 * Read from the parent image the range of data that covers the
2690 * entire target of the given object request. This is used for
2691 * satisfying a layered image write request when the target of an
2692 * object request from the image request does not exist.
2694 * A page array big enough to hold the returned data is allocated
2695 * and supplied to rbd_img_request_fill() as the "data descriptor."
2696 * When the read completes, this page array will be transferred to
2697 * the original object request for the copyup operation.
2699 * If an error occurs, record it as the result of the original
2700 * object request and mark it done so it gets completed.
2702 static int rbd_img_obj_parent_read_full(struct rbd_obj_request *obj_request)
2704 struct rbd_img_request *img_request = NULL;
2705 struct rbd_img_request *parent_request = NULL;
2706 struct rbd_device *rbd_dev;
2709 struct page **pages = NULL;
2713 rbd_assert(obj_request_img_data_test(obj_request));
2714 rbd_assert(obj_request_type_valid(obj_request->type));
2716 img_request = obj_request->img_request;
2717 rbd_assert(img_request != NULL);
2718 rbd_dev = img_request->rbd_dev;
2719 rbd_assert(rbd_dev->parent != NULL);
2722 * Determine the byte range covered by the object in the
2723 * child image to which the original request was to be sent.
2725 img_offset = obj_request->img_offset - obj_request->offset;
2726 length = (u64)1 << rbd_dev->header.obj_order;
2729 * There is no defined parent data beyond the parent
2730 * overlap, so limit what we read at that boundary if
2733 if (img_offset + length > rbd_dev->parent_overlap) {
2734 rbd_assert(img_offset < rbd_dev->parent_overlap);
2735 length = rbd_dev->parent_overlap - img_offset;
2739 * Allocate a page array big enough to receive the data read
2742 page_count = (u32)calc_pages_for(0, length);
2743 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2744 if (IS_ERR(pages)) {
2745 result = PTR_ERR(pages);
2751 parent_request = rbd_parent_request_create(obj_request,
2752 img_offset, length);
2753 if (!parent_request)
2756 result = rbd_img_request_fill(parent_request, OBJ_REQUEST_PAGES, pages);
2759 parent_request->copyup_pages = pages;
2760 parent_request->copyup_page_count = page_count;
2762 parent_request->callback = rbd_img_obj_parent_read_full_callback;
2763 result = rbd_img_request_submit(parent_request);
2767 parent_request->copyup_pages = NULL;
2768 parent_request->copyup_page_count = 0;
2769 parent_request->obj_request = NULL;
2770 rbd_obj_request_put(obj_request);
2773 ceph_release_page_vector(pages, page_count);
2775 rbd_img_request_put(parent_request);
2776 obj_request->result = result;
2777 obj_request->xferred = 0;
2778 obj_request_done_set(obj_request);
2783 static void rbd_img_obj_exists_callback(struct rbd_obj_request *obj_request)
2785 struct rbd_obj_request *orig_request;
2786 struct rbd_device *rbd_dev;
2789 rbd_assert(!obj_request_img_data_test(obj_request));
2792 * All we need from the object request is the original
2793 * request and the result of the STAT op. Grab those, then
2794 * we're done with the request.
2796 orig_request = obj_request->obj_request;
2797 obj_request->obj_request = NULL;
2798 rbd_obj_request_put(orig_request);
2799 rbd_assert(orig_request);
2800 rbd_assert(orig_request->img_request);
2802 result = obj_request->result;
2803 obj_request->result = 0;
2805 dout("%s: obj %p for obj %p result %d %llu/%llu\n", __func__,
2806 obj_request, orig_request, result,
2807 obj_request->xferred, obj_request->length);
2808 rbd_obj_request_put(obj_request);
2811 * If the overlap has become 0 (most likely because the
2812 * image has been flattened) we need to free the pages
2813 * and re-submit the original write request.
2815 rbd_dev = orig_request->img_request->rbd_dev;
2816 if (!rbd_dev->parent_overlap) {
2817 struct ceph_osd_client *osdc;
2819 osdc = &rbd_dev->rbd_client->client->osdc;
2820 result = rbd_obj_request_submit(osdc, orig_request);
2826 * Our only purpose here is to determine whether the object
2827 * exists, and we don't want to treat the non-existence as
2828 * an error. If something else comes back, transfer the
2829 * error to the original request and complete it now.
2832 obj_request_existence_set(orig_request, true);
2833 } else if (result == -ENOENT) {
2834 obj_request_existence_set(orig_request, false);
2835 } else if (result) {
2836 orig_request->result = result;
2841 * Resubmit the original request now that we have recorded
2842 * whether the target object exists.
2844 orig_request->result = rbd_img_obj_request_submit(orig_request);
2846 if (orig_request->result)
2847 rbd_obj_request_complete(orig_request);
2850 static int rbd_img_obj_exists_submit(struct rbd_obj_request *obj_request)
2852 struct rbd_obj_request *stat_request;
2853 struct rbd_device *rbd_dev;
2854 struct ceph_osd_client *osdc;
2855 struct page **pages = NULL;
2861 * The response data for a STAT call consists of:
2868 size = sizeof (__le64) + sizeof (__le32) + sizeof (__le32);
2869 page_count = (u32)calc_pages_for(0, size);
2870 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
2872 return PTR_ERR(pages);
2875 stat_request = rbd_obj_request_create(obj_request->object_name, 0, 0,
2880 rbd_obj_request_get(obj_request);
2881 stat_request->obj_request = obj_request;
2882 stat_request->pages = pages;
2883 stat_request->page_count = page_count;
2885 rbd_assert(obj_request->img_request);
2886 rbd_dev = obj_request->img_request->rbd_dev;
2887 stat_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
2889 if (!stat_request->osd_req)
2891 stat_request->callback = rbd_img_obj_exists_callback;
2893 osd_req_op_init(stat_request->osd_req, 0, CEPH_OSD_OP_STAT, 0);
2894 osd_req_op_raw_data_in_pages(stat_request->osd_req, 0, pages, size, 0,
2896 rbd_osd_req_format_read(stat_request);
2898 osdc = &rbd_dev->rbd_client->client->osdc;
2899 ret = rbd_obj_request_submit(osdc, stat_request);
2902 rbd_obj_request_put(obj_request);
2907 static bool img_obj_request_simple(struct rbd_obj_request *obj_request)
2909 struct rbd_img_request *img_request;
2910 struct rbd_device *rbd_dev;
2912 rbd_assert(obj_request_img_data_test(obj_request));
2914 img_request = obj_request->img_request;
2915 rbd_assert(img_request);
2916 rbd_dev = img_request->rbd_dev;
2919 if (!img_request_write_test(img_request) &&
2920 !img_request_discard_test(img_request))
2923 /* Non-layered writes */
2924 if (!img_request_layered_test(img_request))
2928 * Layered writes outside of the parent overlap range don't
2929 * share any data with the parent.
2931 if (!obj_request_overlaps_parent(obj_request))
2935 * Entire-object layered writes - we will overwrite whatever
2936 * parent data there is anyway.
2938 if (!obj_request->offset &&
2939 obj_request->length == rbd_obj_bytes(&rbd_dev->header))
2943 * If the object is known to already exist, its parent data has
2944 * already been copied.
2946 if (obj_request_known_test(obj_request) &&
2947 obj_request_exists_test(obj_request))
2953 static int rbd_img_obj_request_submit(struct rbd_obj_request *obj_request)
2955 if (img_obj_request_simple(obj_request)) {
2956 struct rbd_device *rbd_dev;
2957 struct ceph_osd_client *osdc;
2959 rbd_dev = obj_request->img_request->rbd_dev;
2960 osdc = &rbd_dev->rbd_client->client->osdc;
2962 return rbd_obj_request_submit(osdc, obj_request);
2966 * It's a layered write. The target object might exist but
2967 * we may not know that yet. If we know it doesn't exist,
2968 * start by reading the data for the full target object from
2969 * the parent so we can use it for a copyup to the target.
2971 if (obj_request_known_test(obj_request))
2972 return rbd_img_obj_parent_read_full(obj_request);
2974 /* We don't know whether the target exists. Go find out. */
2976 return rbd_img_obj_exists_submit(obj_request);
2979 static int rbd_img_request_submit(struct rbd_img_request *img_request)
2981 struct rbd_obj_request *obj_request;
2982 struct rbd_obj_request *next_obj_request;
2985 dout("%s: img %p\n", __func__, img_request);
2987 rbd_img_request_get(img_request);
2988 for_each_obj_request_safe(img_request, obj_request, next_obj_request) {
2989 ret = rbd_img_obj_request_submit(obj_request);
2995 rbd_img_request_put(img_request);
2999 static void rbd_img_parent_read_callback(struct rbd_img_request *img_request)
3001 struct rbd_obj_request *obj_request;
3002 struct rbd_device *rbd_dev;
3007 rbd_assert(img_request_child_test(img_request));
3009 /* First get what we need from the image request and release it */
3011 obj_request = img_request->obj_request;
3012 img_xferred = img_request->xferred;
3013 img_result = img_request->result;
3014 rbd_img_request_put(img_request);
3017 * If the overlap has become 0 (most likely because the
3018 * image has been flattened) we need to re-submit the
3021 rbd_assert(obj_request);
3022 rbd_assert(obj_request->img_request);
3023 rbd_dev = obj_request->img_request->rbd_dev;
3024 if (!rbd_dev->parent_overlap) {
3025 struct ceph_osd_client *osdc;
3027 osdc = &rbd_dev->rbd_client->client->osdc;
3028 img_result = rbd_obj_request_submit(osdc, obj_request);
3033 obj_request->result = img_result;
3034 if (obj_request->result)
3038 * We need to zero anything beyond the parent overlap
3039 * boundary. Since rbd_img_obj_request_read_callback()
3040 * will zero anything beyond the end of a short read, an
3041 * easy way to do this is to pretend the data from the
3042 * parent came up short--ending at the overlap boundary.
3044 rbd_assert(obj_request->img_offset < U64_MAX - obj_request->length);
3045 obj_end = obj_request->img_offset + obj_request->length;
3046 if (obj_end > rbd_dev->parent_overlap) {
3049 if (obj_request->img_offset < rbd_dev->parent_overlap)
3050 xferred = rbd_dev->parent_overlap -
3051 obj_request->img_offset;
3053 obj_request->xferred = min(img_xferred, xferred);
3055 obj_request->xferred = img_xferred;
3058 rbd_img_obj_request_read_callback(obj_request);
3059 rbd_obj_request_complete(obj_request);
3062 static void rbd_img_parent_read(struct rbd_obj_request *obj_request)
3064 struct rbd_img_request *img_request;
3067 rbd_assert(obj_request_img_data_test(obj_request));
3068 rbd_assert(obj_request->img_request != NULL);
3069 rbd_assert(obj_request->result == (s32) -ENOENT);
3070 rbd_assert(obj_request_type_valid(obj_request->type));
3072 /* rbd_read_finish(obj_request, obj_request->length); */
3073 img_request = rbd_parent_request_create(obj_request,
3074 obj_request->img_offset,
3075 obj_request->length);
3080 if (obj_request->type == OBJ_REQUEST_BIO)
3081 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3082 obj_request->bio_list);
3084 result = rbd_img_request_fill(img_request, OBJ_REQUEST_PAGES,
3085 obj_request->pages);
3089 img_request->callback = rbd_img_parent_read_callback;
3090 result = rbd_img_request_submit(img_request);
3097 rbd_img_request_put(img_request);
3098 obj_request->result = result;
3099 obj_request->xferred = 0;
3100 obj_request_done_set(obj_request);
3103 static int rbd_obj_notify_ack_sync(struct rbd_device *rbd_dev, u64 notify_id)
3105 struct rbd_obj_request *obj_request;
3106 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3109 obj_request = rbd_obj_request_create(rbd_dev->header_oid.name, 0, 0,
3110 OBJ_REQUEST_NODATA);
3115 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
3117 if (!obj_request->osd_req)
3120 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_NOTIFY_ACK,
3122 rbd_osd_req_format_read(obj_request);
3124 ret = rbd_obj_request_submit(osdc, obj_request);
3127 ret = rbd_obj_request_wait(obj_request);
3129 rbd_obj_request_put(obj_request);
3134 static void rbd_watch_cb(u64 ver, u64 notify_id, u8 opcode, void *data)
3136 struct rbd_device *rbd_dev = (struct rbd_device *)data;
3139 dout("%s: \"%s\" notify_id %llu opcode %u\n", __func__,
3140 rbd_dev->header_oid.name, (unsigned long long)notify_id,
3141 (unsigned int)opcode);
3144 * Until adequate refresh error handling is in place, there is
3145 * not much we can do here, except warn.
3147 * See http://tracker.ceph.com/issues/5040
3149 ret = rbd_dev_refresh(rbd_dev);
3151 rbd_warn(rbd_dev, "refresh failed: %d", ret);
3153 ret = rbd_obj_notify_ack_sync(rbd_dev, notify_id);
3155 rbd_warn(rbd_dev, "notify_ack ret %d", ret);
3159 * Send a (un)watch request and wait for the ack. Return a request
3160 * with a ref held on success or error.
3162 static struct rbd_obj_request *rbd_obj_watch_request_helper(
3163 struct rbd_device *rbd_dev,
3166 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3167 struct ceph_options *opts = osdc->client->options;
3168 struct rbd_obj_request *obj_request;
3171 obj_request = rbd_obj_request_create(rbd_dev->header_oid.name, 0, 0,
3172 OBJ_REQUEST_NODATA);
3174 return ERR_PTR(-ENOMEM);
3176 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_WRITE, 1,
3178 if (!obj_request->osd_req) {
3183 osd_req_op_watch_init(obj_request->osd_req, 0, CEPH_OSD_OP_WATCH,
3184 rbd_dev->watch_event->cookie, 0, watch);
3185 rbd_osd_req_format_write(obj_request);
3188 ceph_osdc_set_request_linger(osdc, obj_request->osd_req);
3190 ret = rbd_obj_request_submit(osdc, obj_request);
3194 ret = rbd_obj_request_wait_timeout(obj_request, opts->mount_timeout);
3198 ret = obj_request->result;
3201 rbd_obj_request_end(obj_request);
3208 rbd_obj_request_put(obj_request);
3209 return ERR_PTR(ret);
3213 * Initiate a watch request, synchronously.
3215 static int rbd_dev_header_watch_sync(struct rbd_device *rbd_dev)
3217 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3218 struct rbd_obj_request *obj_request;
3221 rbd_assert(!rbd_dev->watch_event);
3222 rbd_assert(!rbd_dev->watch_request);
3224 ret = ceph_osdc_create_event(osdc, rbd_watch_cb, rbd_dev,
3225 &rbd_dev->watch_event);
3229 obj_request = rbd_obj_watch_request_helper(rbd_dev, true);
3230 if (IS_ERR(obj_request)) {
3231 ceph_osdc_cancel_event(rbd_dev->watch_event);
3232 rbd_dev->watch_event = NULL;
3233 return PTR_ERR(obj_request);
3237 * A watch request is set to linger, so the underlying osd
3238 * request won't go away until we unregister it. We retain
3239 * a pointer to the object request during that time (in
3240 * rbd_dev->watch_request), so we'll keep a reference to it.
3241 * We'll drop that reference after we've unregistered it in
3242 * rbd_dev_header_unwatch_sync().
3244 rbd_dev->watch_request = obj_request;
3250 * Tear down a watch request, synchronously.
3252 static void rbd_dev_header_unwatch_sync(struct rbd_device *rbd_dev)
3254 struct rbd_obj_request *obj_request;
3256 rbd_assert(rbd_dev->watch_event);
3257 rbd_assert(rbd_dev->watch_request);
3259 rbd_obj_request_end(rbd_dev->watch_request);
3260 rbd_obj_request_put(rbd_dev->watch_request);
3261 rbd_dev->watch_request = NULL;
3263 obj_request = rbd_obj_watch_request_helper(rbd_dev, false);
3264 if (!IS_ERR(obj_request))
3265 rbd_obj_request_put(obj_request);
3267 rbd_warn(rbd_dev, "unable to tear down watch request (%ld)",
3268 PTR_ERR(obj_request));
3270 ceph_osdc_cancel_event(rbd_dev->watch_event);
3271 rbd_dev->watch_event = NULL;
3273 dout("%s flushing notifies\n", __func__);
3274 ceph_osdc_flush_notifies(&rbd_dev->rbd_client->client->osdc);
3278 * Synchronous osd object method call. Returns the number of bytes
3279 * returned in the outbound buffer, or a negative error code.
3281 static int rbd_obj_method_sync(struct rbd_device *rbd_dev,
3282 const char *object_name,
3283 const char *class_name,
3284 const char *method_name,
3285 const void *outbound,
3286 size_t outbound_size,
3288 size_t inbound_size)
3290 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3291 struct rbd_obj_request *obj_request;
3292 struct page **pages;
3297 * Method calls are ultimately read operations. The result
3298 * should placed into the inbound buffer provided. They
3299 * also supply outbound data--parameters for the object
3300 * method. Currently if this is present it will be a
3303 page_count = (u32)calc_pages_for(0, inbound_size);
3304 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
3306 return PTR_ERR(pages);
3309 obj_request = rbd_obj_request_create(object_name, 0, inbound_size,
3314 obj_request->pages = pages;
3315 obj_request->page_count = page_count;
3317 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
3319 if (!obj_request->osd_req)
3322 osd_req_op_cls_init(obj_request->osd_req, 0, CEPH_OSD_OP_CALL,
3323 class_name, method_name);
3324 if (outbound_size) {
3325 struct ceph_pagelist *pagelist;
3327 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
3331 ceph_pagelist_init(pagelist);
3332 ceph_pagelist_append(pagelist, outbound, outbound_size);
3333 osd_req_op_cls_request_data_pagelist(obj_request->osd_req, 0,
3336 osd_req_op_cls_response_data_pages(obj_request->osd_req, 0,
3337 obj_request->pages, inbound_size,
3339 rbd_osd_req_format_read(obj_request);
3341 ret = rbd_obj_request_submit(osdc, obj_request);
3344 ret = rbd_obj_request_wait(obj_request);
3348 ret = obj_request->result;
3352 rbd_assert(obj_request->xferred < (u64)INT_MAX);
3353 ret = (int)obj_request->xferred;
3354 ceph_copy_from_page_vector(pages, inbound, 0, obj_request->xferred);
3357 rbd_obj_request_put(obj_request);
3359 ceph_release_page_vector(pages, page_count);
3364 static void rbd_queue_workfn(struct work_struct *work)
3366 struct request *rq = blk_mq_rq_from_pdu(work);
3367 struct rbd_device *rbd_dev = rq->q->queuedata;
3368 struct rbd_img_request *img_request;
3369 struct ceph_snap_context *snapc = NULL;
3370 u64 offset = (u64)blk_rq_pos(rq) << SECTOR_SHIFT;
3371 u64 length = blk_rq_bytes(rq);
3372 enum obj_operation_type op_type;
3376 if (rq->cmd_type != REQ_TYPE_FS) {
3377 dout("%s: non-fs request type %d\n", __func__,
3378 (int) rq->cmd_type);
3383 if (rq->cmd_flags & REQ_DISCARD)
3384 op_type = OBJ_OP_DISCARD;
3385 else if (rq->cmd_flags & REQ_WRITE)
3386 op_type = OBJ_OP_WRITE;
3388 op_type = OBJ_OP_READ;
3390 /* Ignore/skip any zero-length requests */
3393 dout("%s: zero-length request\n", __func__);
3398 /* Only reads are allowed to a read-only device */
3400 if (op_type != OBJ_OP_READ) {
3401 if (rbd_dev->mapping.read_only) {
3405 rbd_assert(rbd_dev->spec->snap_id == CEPH_NOSNAP);
3409 * Quit early if the mapped snapshot no longer exists. It's
3410 * still possible the snapshot will have disappeared by the
3411 * time our request arrives at the osd, but there's no sense in
3412 * sending it if we already know.
3414 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
3415 dout("request for non-existent snapshot");
3416 rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
3421 if (offset && length > U64_MAX - offset + 1) {
3422 rbd_warn(rbd_dev, "bad request range (%llu~%llu)", offset,
3425 goto err_rq; /* Shouldn't happen */
3428 blk_mq_start_request(rq);
3430 down_read(&rbd_dev->header_rwsem);
3431 mapping_size = rbd_dev->mapping.size;
3432 if (op_type != OBJ_OP_READ) {
3433 snapc = rbd_dev->header.snapc;
3434 ceph_get_snap_context(snapc);
3436 up_read(&rbd_dev->header_rwsem);
3438 if (offset + length > mapping_size) {
3439 rbd_warn(rbd_dev, "beyond EOD (%llu~%llu > %llu)", offset,
3440 length, mapping_size);
3445 img_request = rbd_img_request_create(rbd_dev, offset, length, op_type,
3451 img_request->rq = rq;
3452 snapc = NULL; /* img_request consumes a ref */
3454 if (op_type == OBJ_OP_DISCARD)
3455 result = rbd_img_request_fill(img_request, OBJ_REQUEST_NODATA,
3458 result = rbd_img_request_fill(img_request, OBJ_REQUEST_BIO,
3461 goto err_img_request;
3463 result = rbd_img_request_submit(img_request);
3465 goto err_img_request;
3470 rbd_img_request_put(img_request);
3473 rbd_warn(rbd_dev, "%s %llx at %llx result %d",
3474 obj_op_name(op_type), length, offset, result);
3475 ceph_put_snap_context(snapc);
3477 blk_mq_end_request(rq, result);
3480 static int rbd_queue_rq(struct blk_mq_hw_ctx *hctx,
3481 const struct blk_mq_queue_data *bd)
3483 struct request *rq = bd->rq;
3484 struct work_struct *work = blk_mq_rq_to_pdu(rq);
3486 queue_work(rbd_wq, work);
3487 return BLK_MQ_RQ_QUEUE_OK;
3490 static void rbd_free_disk(struct rbd_device *rbd_dev)
3492 struct gendisk *disk = rbd_dev->disk;
3497 rbd_dev->disk = NULL;
3498 if (disk->flags & GENHD_FL_UP) {
3501 blk_cleanup_queue(disk->queue);
3502 blk_mq_free_tag_set(&rbd_dev->tag_set);
3507 static int rbd_obj_read_sync(struct rbd_device *rbd_dev,
3508 const char *object_name,
3509 u64 offset, u64 length, void *buf)
3512 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
3513 struct rbd_obj_request *obj_request;
3514 struct page **pages = NULL;
3519 page_count = (u32) calc_pages_for(offset, length);
3520 pages = ceph_alloc_page_vector(page_count, GFP_KERNEL);
3522 return PTR_ERR(pages);
3525 obj_request = rbd_obj_request_create(object_name, offset, length,
3530 obj_request->pages = pages;
3531 obj_request->page_count = page_count;
3533 obj_request->osd_req = rbd_osd_req_create(rbd_dev, OBJ_OP_READ, 1,
3535 if (!obj_request->osd_req)
3538 osd_req_op_extent_init(obj_request->osd_req, 0, CEPH_OSD_OP_READ,
3539 offset, length, 0, 0);
3540 osd_req_op_extent_osd_data_pages(obj_request->osd_req, 0,
3542 obj_request->length,
3543 obj_request->offset & ~PAGE_MASK,
3545 rbd_osd_req_format_read(obj_request);
3547 ret = rbd_obj_request_submit(osdc, obj_request);
3550 ret = rbd_obj_request_wait(obj_request);
3554 ret = obj_request->result;
3558 rbd_assert(obj_request->xferred <= (u64) SIZE_MAX);
3559 size = (size_t) obj_request->xferred;
3560 ceph_copy_from_page_vector(pages, buf, 0, size);
3561 rbd_assert(size <= (size_t)INT_MAX);
3565 rbd_obj_request_put(obj_request);
3567 ceph_release_page_vector(pages, page_count);
3573 * Read the complete header for the given rbd device. On successful
3574 * return, the rbd_dev->header field will contain up-to-date
3575 * information about the image.
3577 static int rbd_dev_v1_header_info(struct rbd_device *rbd_dev)
3579 struct rbd_image_header_ondisk *ondisk = NULL;
3586 * The complete header will include an array of its 64-bit
3587 * snapshot ids, followed by the names of those snapshots as
3588 * a contiguous block of NUL-terminated strings. Note that
3589 * the number of snapshots could change by the time we read
3590 * it in, in which case we re-read it.
3597 size = sizeof (*ondisk);
3598 size += snap_count * sizeof (struct rbd_image_snap_ondisk);
3600 ondisk = kmalloc(size, GFP_KERNEL);
3604 ret = rbd_obj_read_sync(rbd_dev, rbd_dev->header_oid.name,
3608 if ((size_t)ret < size) {
3610 rbd_warn(rbd_dev, "short header read (want %zd got %d)",
3614 if (!rbd_dev_ondisk_valid(ondisk)) {
3616 rbd_warn(rbd_dev, "invalid header");
3620 names_size = le64_to_cpu(ondisk->snap_names_len);
3621 want_count = snap_count;
3622 snap_count = le32_to_cpu(ondisk->snap_count);
3623 } while (snap_count != want_count);
3625 ret = rbd_header_from_disk(rbd_dev, ondisk);
3633 * Clear the rbd device's EXISTS flag if the snapshot it's mapped to
3634 * has disappeared from the (just updated) snapshot context.
3636 static void rbd_exists_validate(struct rbd_device *rbd_dev)
3640 if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags))
3643 snap_id = rbd_dev->spec->snap_id;
3644 if (snap_id == CEPH_NOSNAP)
3647 if (rbd_dev_snap_index(rbd_dev, snap_id) == BAD_SNAP_INDEX)
3648 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
3651 static void rbd_dev_update_size(struct rbd_device *rbd_dev)
3656 * If EXISTS is not set, rbd_dev->disk may be NULL, so don't
3657 * try to update its size. If REMOVING is set, updating size
3658 * is just useless work since the device can't be opened.
3660 if (test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags) &&
3661 !test_bit(RBD_DEV_FLAG_REMOVING, &rbd_dev->flags)) {
3662 size = (sector_t)rbd_dev->mapping.size / SECTOR_SIZE;
3663 dout("setting size to %llu sectors", (unsigned long long)size);
3664 set_capacity(rbd_dev->disk, size);
3665 revalidate_disk(rbd_dev->disk);
3669 static int rbd_dev_refresh(struct rbd_device *rbd_dev)
3674 down_write(&rbd_dev->header_rwsem);
3675 mapping_size = rbd_dev->mapping.size;
3677 ret = rbd_dev_header_info(rbd_dev);
3682 * If there is a parent, see if it has disappeared due to the
3683 * mapped image getting flattened.
3685 if (rbd_dev->parent) {
3686 ret = rbd_dev_v2_parent_info(rbd_dev);
3691 if (rbd_dev->spec->snap_id == CEPH_NOSNAP) {
3692 rbd_dev->mapping.size = rbd_dev->header.image_size;
3694 /* validate mapped snapshot's EXISTS flag */
3695 rbd_exists_validate(rbd_dev);
3699 up_write(&rbd_dev->header_rwsem);
3700 if (!ret && mapping_size != rbd_dev->mapping.size)
3701 rbd_dev_update_size(rbd_dev);
3706 static int rbd_init_request(void *data, struct request *rq,
3707 unsigned int hctx_idx, unsigned int request_idx,
3708 unsigned int numa_node)
3710 struct work_struct *work = blk_mq_rq_to_pdu(rq);
3712 INIT_WORK(work, rbd_queue_workfn);
3716 static struct blk_mq_ops rbd_mq_ops = {
3717 .queue_rq = rbd_queue_rq,
3718 .map_queue = blk_mq_map_queue,
3719 .init_request = rbd_init_request,
3722 static int rbd_init_disk(struct rbd_device *rbd_dev)
3724 struct gendisk *disk;
3725 struct request_queue *q;
3729 /* create gendisk info */
3730 disk = alloc_disk(single_major ?
3731 (1 << RBD_SINGLE_MAJOR_PART_SHIFT) :
3732 RBD_MINORS_PER_MAJOR);
3736 snprintf(disk->disk_name, sizeof(disk->disk_name), RBD_DRV_NAME "%d",
3738 disk->major = rbd_dev->major;
3739 disk->first_minor = rbd_dev->minor;
3741 disk->flags |= GENHD_FL_EXT_DEVT;
3742 disk->fops = &rbd_bd_ops;
3743 disk->private_data = rbd_dev;
3745 memset(&rbd_dev->tag_set, 0, sizeof(rbd_dev->tag_set));
3746 rbd_dev->tag_set.ops = &rbd_mq_ops;
3747 rbd_dev->tag_set.queue_depth = rbd_dev->opts->queue_depth;
3748 rbd_dev->tag_set.numa_node = NUMA_NO_NODE;
3749 rbd_dev->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
3750 rbd_dev->tag_set.nr_hw_queues = 1;
3751 rbd_dev->tag_set.cmd_size = sizeof(struct work_struct);
3753 err = blk_mq_alloc_tag_set(&rbd_dev->tag_set);
3757 q = blk_mq_init_queue(&rbd_dev->tag_set);
3763 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
3764 /* QUEUE_FLAG_ADD_RANDOM is off by default for blk-mq */
3766 /* set io sizes to object size */
3767 segment_size = rbd_obj_bytes(&rbd_dev->header);
3768 blk_queue_max_hw_sectors(q, segment_size / SECTOR_SIZE);
3769 q->limits.max_sectors = queue_max_hw_sectors(q);
3770 blk_queue_max_segments(q, segment_size / SECTOR_SIZE);
3771 blk_queue_max_segment_size(q, segment_size);
3772 blk_queue_io_min(q, segment_size);
3773 blk_queue_io_opt(q, segment_size);
3775 /* enable the discard support */
3776 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
3777 q->limits.discard_granularity = segment_size;
3778 q->limits.discard_alignment = segment_size;
3779 blk_queue_max_discard_sectors(q, segment_size / SECTOR_SIZE);
3780 q->limits.discard_zeroes_data = 1;
3782 if (!ceph_test_opt(rbd_dev->rbd_client->client, NOCRC))
3783 q->backing_dev_info.capabilities |= BDI_CAP_STABLE_WRITES;
3787 q->queuedata = rbd_dev;
3789 rbd_dev->disk = disk;
3793 blk_mq_free_tag_set(&rbd_dev->tag_set);
3803 static struct rbd_device *dev_to_rbd_dev(struct device *dev)
3805 return container_of(dev, struct rbd_device, dev);
3808 static ssize_t rbd_size_show(struct device *dev,
3809 struct device_attribute *attr, char *buf)
3811 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3813 return sprintf(buf, "%llu\n",
3814 (unsigned long long)rbd_dev->mapping.size);
3818 * Note this shows the features for whatever's mapped, which is not
3819 * necessarily the base image.
3821 static ssize_t rbd_features_show(struct device *dev,
3822 struct device_attribute *attr, char *buf)
3824 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3826 return sprintf(buf, "0x%016llx\n",
3827 (unsigned long long)rbd_dev->mapping.features);
3830 static ssize_t rbd_major_show(struct device *dev,
3831 struct device_attribute *attr, char *buf)
3833 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3836 return sprintf(buf, "%d\n", rbd_dev->major);
3838 return sprintf(buf, "(none)\n");
3841 static ssize_t rbd_minor_show(struct device *dev,
3842 struct device_attribute *attr, char *buf)
3844 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3846 return sprintf(buf, "%d\n", rbd_dev->minor);
3849 static ssize_t rbd_client_id_show(struct device *dev,
3850 struct device_attribute *attr, char *buf)
3852 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3854 return sprintf(buf, "client%lld\n",
3855 ceph_client_id(rbd_dev->rbd_client->client));
3858 static ssize_t rbd_pool_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, "%s\n", rbd_dev->spec->pool_name);
3866 static ssize_t rbd_pool_id_show(struct device *dev,
3867 struct device_attribute *attr, char *buf)
3869 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3871 return sprintf(buf, "%llu\n",
3872 (unsigned long long) rbd_dev->spec->pool_id);
3875 static ssize_t rbd_name_show(struct device *dev,
3876 struct device_attribute *attr, char *buf)
3878 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3880 if (rbd_dev->spec->image_name)
3881 return sprintf(buf, "%s\n", rbd_dev->spec->image_name);
3883 return sprintf(buf, "(unknown)\n");
3886 static ssize_t rbd_image_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, "%s\n", rbd_dev->spec->image_id);
3895 * Shows the name of the currently-mapped snapshot (or
3896 * RBD_SNAP_HEAD_NAME for the base image).
3898 static ssize_t rbd_snap_show(struct device *dev,
3899 struct device_attribute *attr,
3902 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3904 return sprintf(buf, "%s\n", rbd_dev->spec->snap_name);
3908 * For a v2 image, shows the chain of parent images, separated by empty
3909 * lines. For v1 images or if there is no parent, shows "(no parent
3912 static ssize_t rbd_parent_show(struct device *dev,
3913 struct device_attribute *attr,
3916 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3919 if (!rbd_dev->parent)
3920 return sprintf(buf, "(no parent image)\n");
3922 for ( ; rbd_dev->parent; rbd_dev = rbd_dev->parent) {
3923 struct rbd_spec *spec = rbd_dev->parent_spec;
3925 count += sprintf(&buf[count], "%s"
3926 "pool_id %llu\npool_name %s\n"
3927 "image_id %s\nimage_name %s\n"
3928 "snap_id %llu\nsnap_name %s\n"
3930 !count ? "" : "\n", /* first? */
3931 spec->pool_id, spec->pool_name,
3932 spec->image_id, spec->image_name ?: "(unknown)",
3933 spec->snap_id, spec->snap_name,
3934 rbd_dev->parent_overlap);
3940 static ssize_t rbd_image_refresh(struct device *dev,
3941 struct device_attribute *attr,
3945 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
3948 ret = rbd_dev_refresh(rbd_dev);
3955 static DEVICE_ATTR(size, S_IRUGO, rbd_size_show, NULL);
3956 static DEVICE_ATTR(features, S_IRUGO, rbd_features_show, NULL);
3957 static DEVICE_ATTR(major, S_IRUGO, rbd_major_show, NULL);
3958 static DEVICE_ATTR(minor, S_IRUGO, rbd_minor_show, NULL);
3959 static DEVICE_ATTR(client_id, S_IRUGO, rbd_client_id_show, NULL);
3960 static DEVICE_ATTR(pool, S_IRUGO, rbd_pool_show, NULL);
3961 static DEVICE_ATTR(pool_id, S_IRUGO, rbd_pool_id_show, NULL);
3962 static DEVICE_ATTR(name, S_IRUGO, rbd_name_show, NULL);
3963 static DEVICE_ATTR(image_id, S_IRUGO, rbd_image_id_show, NULL);
3964 static DEVICE_ATTR(refresh, S_IWUSR, NULL, rbd_image_refresh);
3965 static DEVICE_ATTR(current_snap, S_IRUGO, rbd_snap_show, NULL);
3966 static DEVICE_ATTR(parent, S_IRUGO, rbd_parent_show, NULL);
3968 static struct attribute *rbd_attrs[] = {
3969 &dev_attr_size.attr,
3970 &dev_attr_features.attr,
3971 &dev_attr_major.attr,
3972 &dev_attr_minor.attr,
3973 &dev_attr_client_id.attr,
3974 &dev_attr_pool.attr,
3975 &dev_attr_pool_id.attr,
3976 &dev_attr_name.attr,
3977 &dev_attr_image_id.attr,
3978 &dev_attr_current_snap.attr,
3979 &dev_attr_parent.attr,
3980 &dev_attr_refresh.attr,
3984 static struct attribute_group rbd_attr_group = {
3988 static const struct attribute_group *rbd_attr_groups[] = {
3993 static void rbd_dev_release(struct device *dev);
3995 static struct device_type rbd_device_type = {
3997 .groups = rbd_attr_groups,
3998 .release = rbd_dev_release,
4001 static struct rbd_spec *rbd_spec_get(struct rbd_spec *spec)
4003 kref_get(&spec->kref);
4008 static void rbd_spec_free(struct kref *kref);
4009 static void rbd_spec_put(struct rbd_spec *spec)
4012 kref_put(&spec->kref, rbd_spec_free);
4015 static struct rbd_spec *rbd_spec_alloc(void)
4017 struct rbd_spec *spec;
4019 spec = kzalloc(sizeof (*spec), GFP_KERNEL);
4023 spec->pool_id = CEPH_NOPOOL;
4024 spec->snap_id = CEPH_NOSNAP;
4025 kref_init(&spec->kref);
4030 static void rbd_spec_free(struct kref *kref)
4032 struct rbd_spec *spec = container_of(kref, struct rbd_spec, kref);
4034 kfree(spec->pool_name);
4035 kfree(spec->image_id);
4036 kfree(spec->image_name);
4037 kfree(spec->snap_name);
4041 static void rbd_dev_release(struct device *dev)
4043 struct rbd_device *rbd_dev = dev_to_rbd_dev(dev);
4044 bool need_put = !!rbd_dev->opts;
4046 ceph_oid_destroy(&rbd_dev->header_oid);
4048 rbd_put_client(rbd_dev->rbd_client);
4049 rbd_spec_put(rbd_dev->spec);
4050 kfree(rbd_dev->opts);
4054 * This is racy, but way better than putting module outside of
4055 * the release callback. The race window is pretty small, so
4056 * doing something similar to dm (dm-builtin.c) is overkill.
4059 module_put(THIS_MODULE);
4062 static struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
4063 struct rbd_spec *spec,
4064 struct rbd_options *opts)
4066 struct rbd_device *rbd_dev;
4068 rbd_dev = kzalloc(sizeof (*rbd_dev), GFP_KERNEL);
4072 spin_lock_init(&rbd_dev->lock);
4074 atomic_set(&rbd_dev->parent_ref, 0);
4075 INIT_LIST_HEAD(&rbd_dev->node);
4076 init_rwsem(&rbd_dev->header_rwsem);
4078 ceph_oid_init(&rbd_dev->header_oid);
4080 rbd_dev->dev.bus = &rbd_bus_type;
4081 rbd_dev->dev.type = &rbd_device_type;
4082 rbd_dev->dev.parent = &rbd_root_dev;
4083 device_initialize(&rbd_dev->dev);
4085 rbd_dev->rbd_client = rbdc;
4086 rbd_dev->spec = spec;
4087 rbd_dev->opts = opts;
4089 /* Initialize the layout used for all rbd requests */
4091 rbd_dev->layout.fl_stripe_unit = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
4092 rbd_dev->layout.fl_stripe_count = cpu_to_le32(1);
4093 rbd_dev->layout.fl_object_size = cpu_to_le32(1 << RBD_MAX_OBJ_ORDER);
4094 rbd_dev->layout.fl_pg_pool = cpu_to_le32((u32) spec->pool_id);
4097 * If this is a mapping rbd_dev (as opposed to a parent one),
4098 * pin our module. We have a ref from do_rbd_add(), so use
4102 __module_get(THIS_MODULE);
4107 static void rbd_dev_destroy(struct rbd_device *rbd_dev)
4110 put_device(&rbd_dev->dev);
4114 * Get the size and object order for an image snapshot, or if
4115 * snap_id is CEPH_NOSNAP, gets this information for the base
4118 static int _rbd_dev_v2_snap_size(struct rbd_device *rbd_dev, u64 snap_id,
4119 u8 *order, u64 *snap_size)
4121 __le64 snapid = cpu_to_le64(snap_id);
4126 } __attribute__ ((packed)) size_buf = { 0 };
4128 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
4130 &snapid, sizeof (snapid),
4131 &size_buf, sizeof (size_buf));
4132 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4135 if (ret < sizeof (size_buf))
4139 *order = size_buf.order;
4140 dout(" order %u", (unsigned int)*order);
4142 *snap_size = le64_to_cpu(size_buf.size);
4144 dout(" snap_id 0x%016llx snap_size = %llu\n",
4145 (unsigned long long)snap_id,
4146 (unsigned long long)*snap_size);
4151 static int rbd_dev_v2_image_size(struct rbd_device *rbd_dev)
4153 return _rbd_dev_v2_snap_size(rbd_dev, CEPH_NOSNAP,
4154 &rbd_dev->header.obj_order,
4155 &rbd_dev->header.image_size);
4158 static int rbd_dev_v2_object_prefix(struct rbd_device *rbd_dev)
4164 reply_buf = kzalloc(RBD_OBJ_PREFIX_LEN_MAX, GFP_KERNEL);
4168 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
4169 "rbd", "get_object_prefix", NULL, 0,
4170 reply_buf, RBD_OBJ_PREFIX_LEN_MAX);
4171 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4176 rbd_dev->header.object_prefix = ceph_extract_encoded_string(&p,
4177 p + ret, NULL, GFP_NOIO);
4180 if (IS_ERR(rbd_dev->header.object_prefix)) {
4181 ret = PTR_ERR(rbd_dev->header.object_prefix);
4182 rbd_dev->header.object_prefix = NULL;
4184 dout(" object_prefix = %s\n", rbd_dev->header.object_prefix);
4192 static int _rbd_dev_v2_snap_features(struct rbd_device *rbd_dev, u64 snap_id,
4195 __le64 snapid = cpu_to_le64(snap_id);
4199 } __attribute__ ((packed)) features_buf = { 0 };
4203 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
4204 "rbd", "get_features",
4205 &snapid, sizeof (snapid),
4206 &features_buf, sizeof (features_buf));
4207 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4210 if (ret < sizeof (features_buf))
4213 unsup = le64_to_cpu(features_buf.incompat) & ~RBD_FEATURES_SUPPORTED;
4215 rbd_warn(rbd_dev, "image uses unsupported features: 0x%llx",
4220 *snap_features = le64_to_cpu(features_buf.features);
4222 dout(" snap_id 0x%016llx features = 0x%016llx incompat = 0x%016llx\n",
4223 (unsigned long long)snap_id,
4224 (unsigned long long)*snap_features,
4225 (unsigned long long)le64_to_cpu(features_buf.incompat));
4230 static int rbd_dev_v2_features(struct rbd_device *rbd_dev)
4232 return _rbd_dev_v2_snap_features(rbd_dev, CEPH_NOSNAP,
4233 &rbd_dev->header.features);
4236 static int rbd_dev_v2_parent_info(struct rbd_device *rbd_dev)
4238 struct rbd_spec *parent_spec;
4240 void *reply_buf = NULL;
4250 parent_spec = rbd_spec_alloc();
4254 size = sizeof (__le64) + /* pool_id */
4255 sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX + /* image_id */
4256 sizeof (__le64) + /* snap_id */
4257 sizeof (__le64); /* overlap */
4258 reply_buf = kmalloc(size, GFP_KERNEL);
4264 snapid = cpu_to_le64(rbd_dev->spec->snap_id);
4265 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
4266 "rbd", "get_parent",
4267 &snapid, sizeof (snapid),
4269 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4274 end = reply_buf + ret;
4276 ceph_decode_64_safe(&p, end, pool_id, out_err);
4277 if (pool_id == CEPH_NOPOOL) {
4279 * Either the parent never existed, or we have
4280 * record of it but the image got flattened so it no
4281 * longer has a parent. When the parent of a
4282 * layered image disappears we immediately set the
4283 * overlap to 0. The effect of this is that all new
4284 * requests will be treated as if the image had no
4287 if (rbd_dev->parent_overlap) {
4288 rbd_dev->parent_overlap = 0;
4289 rbd_dev_parent_put(rbd_dev);
4290 pr_info("%s: clone image has been flattened\n",
4291 rbd_dev->disk->disk_name);
4294 goto out; /* No parent? No problem. */
4297 /* The ceph file layout needs to fit pool id in 32 bits */
4300 if (pool_id > (u64)U32_MAX) {
4301 rbd_warn(NULL, "parent pool id too large (%llu > %u)",
4302 (unsigned long long)pool_id, U32_MAX);
4306 image_id = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
4307 if (IS_ERR(image_id)) {
4308 ret = PTR_ERR(image_id);
4311 ceph_decode_64_safe(&p, end, snap_id, out_err);
4312 ceph_decode_64_safe(&p, end, overlap, out_err);
4315 * The parent won't change (except when the clone is
4316 * flattened, already handled that). So we only need to
4317 * record the parent spec we have not already done so.
4319 if (!rbd_dev->parent_spec) {
4320 parent_spec->pool_id = pool_id;
4321 parent_spec->image_id = image_id;
4322 parent_spec->snap_id = snap_id;
4323 rbd_dev->parent_spec = parent_spec;
4324 parent_spec = NULL; /* rbd_dev now owns this */
4330 * We always update the parent overlap. If it's zero we issue
4331 * a warning, as we will proceed as if there was no parent.
4335 /* refresh, careful to warn just once */
4336 if (rbd_dev->parent_overlap)
4338 "clone now standalone (overlap became 0)");
4341 rbd_warn(rbd_dev, "clone is standalone (overlap 0)");
4344 rbd_dev->parent_overlap = overlap;
4350 rbd_spec_put(parent_spec);
4355 static int rbd_dev_v2_striping_info(struct rbd_device *rbd_dev)
4359 __le64 stripe_count;
4360 } __attribute__ ((packed)) striping_info_buf = { 0 };
4361 size_t size = sizeof (striping_info_buf);
4368 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
4369 "rbd", "get_stripe_unit_count", NULL, 0,
4370 (char *)&striping_info_buf, size);
4371 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4378 * We don't actually support the "fancy striping" feature
4379 * (STRIPINGV2) yet, but if the striping sizes are the
4380 * defaults the behavior is the same as before. So find
4381 * out, and only fail if the image has non-default values.
4384 obj_size = (u64)1 << rbd_dev->header.obj_order;
4385 p = &striping_info_buf;
4386 stripe_unit = ceph_decode_64(&p);
4387 if (stripe_unit != obj_size) {
4388 rbd_warn(rbd_dev, "unsupported stripe unit "
4389 "(got %llu want %llu)",
4390 stripe_unit, obj_size);
4393 stripe_count = ceph_decode_64(&p);
4394 if (stripe_count != 1) {
4395 rbd_warn(rbd_dev, "unsupported stripe count "
4396 "(got %llu want 1)", stripe_count);
4399 rbd_dev->header.stripe_unit = stripe_unit;
4400 rbd_dev->header.stripe_count = stripe_count;
4405 static char *rbd_dev_image_name(struct rbd_device *rbd_dev)
4407 size_t image_id_size;
4412 void *reply_buf = NULL;
4414 char *image_name = NULL;
4417 rbd_assert(!rbd_dev->spec->image_name);
4419 len = strlen(rbd_dev->spec->image_id);
4420 image_id_size = sizeof (__le32) + len;
4421 image_id = kmalloc(image_id_size, GFP_KERNEL);
4426 end = image_id + image_id_size;
4427 ceph_encode_string(&p, end, rbd_dev->spec->image_id, (u32)len);
4429 size = sizeof (__le32) + RBD_IMAGE_NAME_LEN_MAX;
4430 reply_buf = kmalloc(size, GFP_KERNEL);
4434 ret = rbd_obj_method_sync(rbd_dev, RBD_DIRECTORY,
4435 "rbd", "dir_get_name",
4436 image_id, image_id_size,
4441 end = reply_buf + ret;
4443 image_name = ceph_extract_encoded_string(&p, end, &len, GFP_KERNEL);
4444 if (IS_ERR(image_name))
4447 dout("%s: name is %s len is %zd\n", __func__, image_name, len);
4455 static u64 rbd_v1_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4457 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4458 const char *snap_name;
4461 /* Skip over names until we find the one we are looking for */
4463 snap_name = rbd_dev->header.snap_names;
4464 while (which < snapc->num_snaps) {
4465 if (!strcmp(name, snap_name))
4466 return snapc->snaps[which];
4467 snap_name += strlen(snap_name) + 1;
4473 static u64 rbd_v2_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4475 struct ceph_snap_context *snapc = rbd_dev->header.snapc;
4480 for (which = 0; !found && which < snapc->num_snaps; which++) {
4481 const char *snap_name;
4483 snap_id = snapc->snaps[which];
4484 snap_name = rbd_dev_v2_snap_name(rbd_dev, snap_id);
4485 if (IS_ERR(snap_name)) {
4486 /* ignore no-longer existing snapshots */
4487 if (PTR_ERR(snap_name) == -ENOENT)
4492 found = !strcmp(name, snap_name);
4495 return found ? snap_id : CEPH_NOSNAP;
4499 * Assumes name is never RBD_SNAP_HEAD_NAME; returns CEPH_NOSNAP if
4500 * no snapshot by that name is found, or if an error occurs.
4502 static u64 rbd_snap_id_by_name(struct rbd_device *rbd_dev, const char *name)
4504 if (rbd_dev->image_format == 1)
4505 return rbd_v1_snap_id_by_name(rbd_dev, name);
4507 return rbd_v2_snap_id_by_name(rbd_dev, name);
4511 * An image being mapped will have everything but the snap id.
4513 static int rbd_spec_fill_snap_id(struct rbd_device *rbd_dev)
4515 struct rbd_spec *spec = rbd_dev->spec;
4517 rbd_assert(spec->pool_id != CEPH_NOPOOL && spec->pool_name);
4518 rbd_assert(spec->image_id && spec->image_name);
4519 rbd_assert(spec->snap_name);
4521 if (strcmp(spec->snap_name, RBD_SNAP_HEAD_NAME)) {
4524 snap_id = rbd_snap_id_by_name(rbd_dev, spec->snap_name);
4525 if (snap_id == CEPH_NOSNAP)
4528 spec->snap_id = snap_id;
4530 spec->snap_id = CEPH_NOSNAP;
4537 * A parent image will have all ids but none of the names.
4539 * All names in an rbd spec are dynamically allocated. It's OK if we
4540 * can't figure out the name for an image id.
4542 static int rbd_spec_fill_names(struct rbd_device *rbd_dev)
4544 struct ceph_osd_client *osdc = &rbd_dev->rbd_client->client->osdc;
4545 struct rbd_spec *spec = rbd_dev->spec;
4546 const char *pool_name;
4547 const char *image_name;
4548 const char *snap_name;
4551 rbd_assert(spec->pool_id != CEPH_NOPOOL);
4552 rbd_assert(spec->image_id);
4553 rbd_assert(spec->snap_id != CEPH_NOSNAP);
4555 /* Get the pool name; we have to make our own copy of this */
4557 pool_name = ceph_pg_pool_name_by_id(osdc->osdmap, spec->pool_id);
4559 rbd_warn(rbd_dev, "no pool with id %llu", spec->pool_id);
4562 pool_name = kstrdup(pool_name, GFP_KERNEL);
4566 /* Fetch the image name; tolerate failure here */
4568 image_name = rbd_dev_image_name(rbd_dev);
4570 rbd_warn(rbd_dev, "unable to get image name");
4572 /* Fetch the snapshot name */
4574 snap_name = rbd_snap_name(rbd_dev, spec->snap_id);
4575 if (IS_ERR(snap_name)) {
4576 ret = PTR_ERR(snap_name);
4580 spec->pool_name = pool_name;
4581 spec->image_name = image_name;
4582 spec->snap_name = snap_name;
4592 static int rbd_dev_v2_snap_context(struct rbd_device *rbd_dev)
4601 struct ceph_snap_context *snapc;
4605 * We'll need room for the seq value (maximum snapshot id),
4606 * snapshot count, and array of that many snapshot ids.
4607 * For now we have a fixed upper limit on the number we're
4608 * prepared to receive.
4610 size = sizeof (__le64) + sizeof (__le32) +
4611 RBD_MAX_SNAP_COUNT * sizeof (__le64);
4612 reply_buf = kzalloc(size, GFP_KERNEL);
4616 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
4617 "rbd", "get_snapcontext", NULL, 0,
4619 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4624 end = reply_buf + ret;
4626 ceph_decode_64_safe(&p, end, seq, out);
4627 ceph_decode_32_safe(&p, end, snap_count, out);
4630 * Make sure the reported number of snapshot ids wouldn't go
4631 * beyond the end of our buffer. But before checking that,
4632 * make sure the computed size of the snapshot context we
4633 * allocate is representable in a size_t.
4635 if (snap_count > (SIZE_MAX - sizeof (struct ceph_snap_context))
4640 if (!ceph_has_room(&p, end, snap_count * sizeof (__le64)))
4644 snapc = ceph_create_snap_context(snap_count, GFP_KERNEL);
4650 for (i = 0; i < snap_count; i++)
4651 snapc->snaps[i] = ceph_decode_64(&p);
4653 ceph_put_snap_context(rbd_dev->header.snapc);
4654 rbd_dev->header.snapc = snapc;
4656 dout(" snap context seq = %llu, snap_count = %u\n",
4657 (unsigned long long)seq, (unsigned int)snap_count);
4664 static const char *rbd_dev_v2_snap_name(struct rbd_device *rbd_dev,
4675 size = sizeof (__le32) + RBD_MAX_SNAP_NAME_LEN;
4676 reply_buf = kmalloc(size, GFP_KERNEL);
4678 return ERR_PTR(-ENOMEM);
4680 snapid = cpu_to_le64(snap_id);
4681 ret = rbd_obj_method_sync(rbd_dev, rbd_dev->header_oid.name,
4682 "rbd", "get_snapshot_name",
4683 &snapid, sizeof (snapid),
4685 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
4687 snap_name = ERR_PTR(ret);
4692 end = reply_buf + ret;
4693 snap_name = ceph_extract_encoded_string(&p, end, NULL, GFP_KERNEL);
4694 if (IS_ERR(snap_name))
4697 dout(" snap_id 0x%016llx snap_name = %s\n",
4698 (unsigned long long)snap_id, snap_name);
4705 static int rbd_dev_v2_header_info(struct rbd_device *rbd_dev)
4707 bool first_time = rbd_dev->header.object_prefix == NULL;
4710 ret = rbd_dev_v2_image_size(rbd_dev);
4715 ret = rbd_dev_v2_header_onetime(rbd_dev);
4720 ret = rbd_dev_v2_snap_context(rbd_dev);
4721 if (ret && first_time) {
4722 kfree(rbd_dev->header.object_prefix);
4723 rbd_dev->header.object_prefix = NULL;
4729 static int rbd_dev_header_info(struct rbd_device *rbd_dev)
4731 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
4733 if (rbd_dev->image_format == 1)
4734 return rbd_dev_v1_header_info(rbd_dev);
4736 return rbd_dev_v2_header_info(rbd_dev);
4740 * Get a unique rbd identifier for the given new rbd_dev, and add
4741 * the rbd_dev to the global list.
4743 static int rbd_dev_id_get(struct rbd_device *rbd_dev)
4747 new_dev_id = ida_simple_get(&rbd_dev_id_ida,
4748 0, minor_to_rbd_dev_id(1 << MINORBITS),
4753 rbd_dev->dev_id = new_dev_id;
4755 spin_lock(&rbd_dev_list_lock);
4756 list_add_tail(&rbd_dev->node, &rbd_dev_list);
4757 spin_unlock(&rbd_dev_list_lock);
4759 dout("rbd_dev %p given dev id %d\n", rbd_dev, rbd_dev->dev_id);
4765 * Remove an rbd_dev from the global list, and record that its
4766 * identifier is no longer in use.
4768 static void rbd_dev_id_put(struct rbd_device *rbd_dev)
4770 spin_lock(&rbd_dev_list_lock);
4771 list_del_init(&rbd_dev->node);
4772 spin_unlock(&rbd_dev_list_lock);
4774 ida_simple_remove(&rbd_dev_id_ida, rbd_dev->dev_id);
4776 dout("rbd_dev %p released dev id %d\n", rbd_dev, rbd_dev->dev_id);
4780 * Skips over white space at *buf, and updates *buf to point to the
4781 * first found non-space character (if any). Returns the length of
4782 * the token (string of non-white space characters) found. Note
4783 * that *buf must be terminated with '\0'.
4785 static inline size_t next_token(const char **buf)
4788 * These are the characters that produce nonzero for
4789 * isspace() in the "C" and "POSIX" locales.
4791 const char *spaces = " \f\n\r\t\v";
4793 *buf += strspn(*buf, spaces); /* Find start of token */
4795 return strcspn(*buf, spaces); /* Return token length */
4799 * Finds the next token in *buf, dynamically allocates a buffer big
4800 * enough to hold a copy of it, and copies the token into the new
4801 * buffer. The copy is guaranteed to be terminated with '\0'. Note
4802 * that a duplicate buffer is created even for a zero-length token.
4804 * Returns a pointer to the newly-allocated duplicate, or a null
4805 * pointer if memory for the duplicate was not available. If
4806 * the lenp argument is a non-null pointer, the length of the token
4807 * (not including the '\0') is returned in *lenp.
4809 * If successful, the *buf pointer will be updated to point beyond
4810 * the end of the found token.
4812 * Note: uses GFP_KERNEL for allocation.
4814 static inline char *dup_token(const char **buf, size_t *lenp)
4819 len = next_token(buf);
4820 dup = kmemdup(*buf, len + 1, GFP_KERNEL);
4823 *(dup + len) = '\0';
4833 * Parse the options provided for an "rbd add" (i.e., rbd image
4834 * mapping) request. These arrive via a write to /sys/bus/rbd/add,
4835 * and the data written is passed here via a NUL-terminated buffer.
4836 * Returns 0 if successful or an error code otherwise.
4838 * The information extracted from these options is recorded in
4839 * the other parameters which return dynamically-allocated
4842 * The address of a pointer that will refer to a ceph options
4843 * structure. Caller must release the returned pointer using
4844 * ceph_destroy_options() when it is no longer needed.
4846 * Address of an rbd options pointer. Fully initialized by
4847 * this function; caller must release with kfree().
4849 * Address of an rbd image specification pointer. Fully
4850 * initialized by this function based on parsed options.
4851 * Caller must release with rbd_spec_put().
4853 * The options passed take this form:
4854 * <mon_addrs> <options> <pool_name> <image_name> [<snap_id>]
4857 * A comma-separated list of one or more monitor addresses.
4858 * A monitor address is an ip address, optionally followed
4859 * by a port number (separated by a colon).
4860 * I.e.: ip1[:port1][,ip2[:port2]...]
4862 * A comma-separated list of ceph and/or rbd options.
4864 * The name of the rados pool containing the rbd image.
4866 * The name of the image in that pool to map.
4868 * An optional snapshot id. If provided, the mapping will
4869 * present data from the image at the time that snapshot was
4870 * created. The image head is used if no snapshot id is
4871 * provided. Snapshot mappings are always read-only.
4873 static int rbd_add_parse_args(const char *buf,
4874 struct ceph_options **ceph_opts,
4875 struct rbd_options **opts,
4876 struct rbd_spec **rbd_spec)
4880 const char *mon_addrs;
4882 size_t mon_addrs_size;
4883 struct rbd_spec *spec = NULL;
4884 struct rbd_options *rbd_opts = NULL;
4885 struct ceph_options *copts;
4888 /* The first four tokens are required */
4890 len = next_token(&buf);
4892 rbd_warn(NULL, "no monitor address(es) provided");
4896 mon_addrs_size = len + 1;
4900 options = dup_token(&buf, NULL);
4904 rbd_warn(NULL, "no options provided");
4908 spec = rbd_spec_alloc();
4912 spec->pool_name = dup_token(&buf, NULL);
4913 if (!spec->pool_name)
4915 if (!*spec->pool_name) {
4916 rbd_warn(NULL, "no pool name provided");
4920 spec->image_name = dup_token(&buf, NULL);
4921 if (!spec->image_name)
4923 if (!*spec->image_name) {
4924 rbd_warn(NULL, "no image name provided");
4929 * Snapshot name is optional; default is to use "-"
4930 * (indicating the head/no snapshot).
4932 len = next_token(&buf);
4934 buf = RBD_SNAP_HEAD_NAME; /* No snapshot supplied */
4935 len = sizeof (RBD_SNAP_HEAD_NAME) - 1;
4936 } else if (len > RBD_MAX_SNAP_NAME_LEN) {
4937 ret = -ENAMETOOLONG;
4940 snap_name = kmemdup(buf, len + 1, GFP_KERNEL);
4943 *(snap_name + len) = '\0';
4944 spec->snap_name = snap_name;
4946 /* Initialize all rbd options to the defaults */
4948 rbd_opts = kzalloc(sizeof (*rbd_opts), GFP_KERNEL);
4952 rbd_opts->read_only = RBD_READ_ONLY_DEFAULT;
4953 rbd_opts->queue_depth = RBD_QUEUE_DEPTH_DEFAULT;
4955 copts = ceph_parse_options(options, mon_addrs,
4956 mon_addrs + mon_addrs_size - 1,
4957 parse_rbd_opts_token, rbd_opts);
4958 if (IS_ERR(copts)) {
4959 ret = PTR_ERR(copts);
4980 * Return pool id (>= 0) or a negative error code.
4982 static int rbd_add_get_pool_id(struct rbd_client *rbdc, const char *pool_name)
4984 struct ceph_options *opts = rbdc->client->options;
4990 ret = ceph_pg_poolid_by_name(rbdc->client->osdc.osdmap, pool_name);
4991 if (ret == -ENOENT && tries++ < 1) {
4992 ret = ceph_monc_do_get_version(&rbdc->client->monc, "osdmap",
4997 if (rbdc->client->osdc.osdmap->epoch < newest_epoch) {
4998 ceph_monc_request_next_osdmap(&rbdc->client->monc);
4999 (void) ceph_monc_wait_osdmap(&rbdc->client->monc,
5001 opts->mount_timeout);
5004 /* the osdmap we have is new enough */
5013 * An rbd format 2 image has a unique identifier, distinct from the
5014 * name given to it by the user. Internally, that identifier is
5015 * what's used to specify the names of objects related to the image.
5017 * A special "rbd id" object is used to map an rbd image name to its
5018 * id. If that object doesn't exist, then there is no v2 rbd image
5019 * with the supplied name.
5021 * This function will record the given rbd_dev's image_id field if
5022 * it can be determined, and in that case will return 0. If any
5023 * errors occur a negative errno will be returned and the rbd_dev's
5024 * image_id field will be unchanged (and should be NULL).
5026 static int rbd_dev_image_id(struct rbd_device *rbd_dev)
5035 * When probing a parent image, the image id is already
5036 * known (and the image name likely is not). There's no
5037 * need to fetch the image id again in this case. We
5038 * do still need to set the image format though.
5040 if (rbd_dev->spec->image_id) {
5041 rbd_dev->image_format = *rbd_dev->spec->image_id ? 2 : 1;
5047 * First, see if the format 2 image id file exists, and if
5048 * so, get the image's persistent id from it.
5050 size = sizeof (RBD_ID_PREFIX) + strlen(rbd_dev->spec->image_name);
5051 object_name = kmalloc(size, GFP_NOIO);
5054 sprintf(object_name, "%s%s", RBD_ID_PREFIX, rbd_dev->spec->image_name);
5055 dout("rbd id object name is %s\n", object_name);
5057 /* Response will be an encoded string, which includes a length */
5059 size = sizeof (__le32) + RBD_IMAGE_ID_LEN_MAX;
5060 response = kzalloc(size, GFP_NOIO);
5066 /* If it doesn't exist we'll assume it's a format 1 image */
5068 ret = rbd_obj_method_sync(rbd_dev, object_name,
5069 "rbd", "get_id", NULL, 0,
5070 response, RBD_IMAGE_ID_LEN_MAX);
5071 dout("%s: rbd_obj_method_sync returned %d\n", __func__, ret);
5072 if (ret == -ENOENT) {
5073 image_id = kstrdup("", GFP_KERNEL);
5074 ret = image_id ? 0 : -ENOMEM;
5076 rbd_dev->image_format = 1;
5077 } else if (ret >= 0) {
5080 image_id = ceph_extract_encoded_string(&p, p + ret,
5082 ret = PTR_ERR_OR_ZERO(image_id);
5084 rbd_dev->image_format = 2;
5088 rbd_dev->spec->image_id = image_id;
5089 dout("image_id is %s\n", image_id);
5099 * Undo whatever state changes are made by v1 or v2 header info
5102 static void rbd_dev_unprobe(struct rbd_device *rbd_dev)
5104 struct rbd_image_header *header;
5106 rbd_dev_parent_put(rbd_dev);
5108 /* Free dynamic fields from the header, then zero it out */
5110 header = &rbd_dev->header;
5111 ceph_put_snap_context(header->snapc);
5112 kfree(header->snap_sizes);
5113 kfree(header->snap_names);
5114 kfree(header->object_prefix);
5115 memset(header, 0, sizeof (*header));
5118 static int rbd_dev_v2_header_onetime(struct rbd_device *rbd_dev)
5122 ret = rbd_dev_v2_object_prefix(rbd_dev);
5127 * Get the and check features for the image. Currently the
5128 * features are assumed to never change.
5130 ret = rbd_dev_v2_features(rbd_dev);
5134 /* If the image supports fancy striping, get its parameters */
5136 if (rbd_dev->header.features & RBD_FEATURE_STRIPINGV2) {
5137 ret = rbd_dev_v2_striping_info(rbd_dev);
5141 /* No support for crypto and compression type format 2 images */
5145 rbd_dev->header.features = 0;
5146 kfree(rbd_dev->header.object_prefix);
5147 rbd_dev->header.object_prefix = NULL;
5153 * @depth is rbd_dev_image_probe() -> rbd_dev_probe_parent() ->
5154 * rbd_dev_image_probe() recursion depth, which means it's also the
5155 * length of the already discovered part of the parent chain.
5157 static int rbd_dev_probe_parent(struct rbd_device *rbd_dev, int depth)
5159 struct rbd_device *parent = NULL;
5162 if (!rbd_dev->parent_spec)
5165 if (++depth > RBD_MAX_PARENT_CHAIN_LEN) {
5166 pr_info("parent chain is too long (%d)\n", depth);
5171 parent = rbd_dev_create(rbd_dev->rbd_client, rbd_dev->parent_spec,
5179 * Images related by parent/child relationships always share
5180 * rbd_client and spec/parent_spec, so bump their refcounts.
5182 __rbd_get_client(rbd_dev->rbd_client);
5183 rbd_spec_get(rbd_dev->parent_spec);
5185 ret = rbd_dev_image_probe(parent, depth);
5189 rbd_dev->parent = parent;
5190 atomic_set(&rbd_dev->parent_ref, 1);
5194 rbd_dev_unparent(rbd_dev);
5195 rbd_dev_destroy(parent);
5200 * rbd_dev->header_rwsem must be locked for write and will be unlocked
5203 static int rbd_dev_device_setup(struct rbd_device *rbd_dev)
5207 /* Get an id and fill in device name. */
5209 ret = rbd_dev_id_get(rbd_dev);
5211 goto err_out_unlock;
5213 BUILD_BUG_ON(DEV_NAME_LEN
5214 < sizeof (RBD_DRV_NAME) + MAX_INT_FORMAT_WIDTH);
5215 sprintf(rbd_dev->name, "%s%d", RBD_DRV_NAME, rbd_dev->dev_id);
5217 /* Record our major and minor device numbers. */
5219 if (!single_major) {
5220 ret = register_blkdev(0, rbd_dev->name);
5224 rbd_dev->major = ret;
5227 rbd_dev->major = rbd_major;
5228 rbd_dev->minor = rbd_dev_id_to_minor(rbd_dev->dev_id);
5231 /* Set up the blkdev mapping. */
5233 ret = rbd_init_disk(rbd_dev);
5235 goto err_out_blkdev;
5237 ret = rbd_dev_mapping_set(rbd_dev);
5241 set_capacity(rbd_dev->disk, rbd_dev->mapping.size / SECTOR_SIZE);
5242 set_disk_ro(rbd_dev->disk, rbd_dev->mapping.read_only);
5244 dev_set_name(&rbd_dev->dev, "%d", rbd_dev->dev_id);
5245 ret = device_add(&rbd_dev->dev);
5247 goto err_out_mapping;
5249 /* Everything's ready. Announce the disk to the world. */
5251 set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5252 up_write(&rbd_dev->header_rwsem);
5254 add_disk(rbd_dev->disk);
5255 pr_info("%s: added with size 0x%llx\n", rbd_dev->disk->disk_name,
5256 (unsigned long long) rbd_dev->mapping.size);
5261 rbd_dev_mapping_clear(rbd_dev);
5263 rbd_free_disk(rbd_dev);
5266 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5268 rbd_dev_id_put(rbd_dev);
5270 up_write(&rbd_dev->header_rwsem);
5274 static int rbd_dev_header_name(struct rbd_device *rbd_dev)
5276 struct rbd_spec *spec = rbd_dev->spec;
5279 /* Record the header object name for this rbd image. */
5281 rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
5283 if (rbd_dev->image_format == 1)
5284 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
5285 spec->image_name, RBD_SUFFIX);
5287 ret = ceph_oid_aprintf(&rbd_dev->header_oid, GFP_KERNEL, "%s%s",
5288 RBD_HEADER_PREFIX, spec->image_id);
5293 static void rbd_dev_image_release(struct rbd_device *rbd_dev)
5295 rbd_dev_unprobe(rbd_dev);
5296 rbd_dev->image_format = 0;
5297 kfree(rbd_dev->spec->image_id);
5298 rbd_dev->spec->image_id = NULL;
5300 rbd_dev_destroy(rbd_dev);
5304 * Probe for the existence of the header object for the given rbd
5305 * device. If this image is the one being mapped (i.e., not a
5306 * parent), initiate a watch on its header object before using that
5307 * object to get detailed information about the rbd image.
5309 static int rbd_dev_image_probe(struct rbd_device *rbd_dev, int depth)
5314 * Get the id from the image id object. Unless there's an
5315 * error, rbd_dev->spec->image_id will be filled in with
5316 * a dynamically-allocated string, and rbd_dev->image_format
5317 * will be set to either 1 or 2.
5319 ret = rbd_dev_image_id(rbd_dev);
5323 ret = rbd_dev_header_name(rbd_dev);
5325 goto err_out_format;
5328 ret = rbd_dev_header_watch_sync(rbd_dev);
5331 pr_info("image %s/%s does not exist\n",
5332 rbd_dev->spec->pool_name,
5333 rbd_dev->spec->image_name);
5334 goto err_out_format;
5338 ret = rbd_dev_header_info(rbd_dev);
5343 * If this image is the one being mapped, we have pool name and
5344 * id, image name and id, and snap name - need to fill snap id.
5345 * Otherwise this is a parent image, identified by pool, image
5346 * and snap ids - need to fill in names for those ids.
5349 ret = rbd_spec_fill_snap_id(rbd_dev);
5351 ret = rbd_spec_fill_names(rbd_dev);
5354 pr_info("snap %s/%s@%s does not exist\n",
5355 rbd_dev->spec->pool_name,
5356 rbd_dev->spec->image_name,
5357 rbd_dev->spec->snap_name);
5361 if (rbd_dev->header.features & RBD_FEATURE_LAYERING) {
5362 ret = rbd_dev_v2_parent_info(rbd_dev);
5367 * Need to warn users if this image is the one being
5368 * mapped and has a parent.
5370 if (!depth && rbd_dev->parent_spec)
5372 "WARNING: kernel layering is EXPERIMENTAL!");
5375 ret = rbd_dev_probe_parent(rbd_dev, depth);
5379 dout("discovered format %u image, header name is %s\n",
5380 rbd_dev->image_format, rbd_dev->header_oid.name);
5384 rbd_dev_unprobe(rbd_dev);
5387 rbd_dev_header_unwatch_sync(rbd_dev);
5389 rbd_dev->image_format = 0;
5390 kfree(rbd_dev->spec->image_id);
5391 rbd_dev->spec->image_id = NULL;
5395 static ssize_t do_rbd_add(struct bus_type *bus,
5399 struct rbd_device *rbd_dev = NULL;
5400 struct ceph_options *ceph_opts = NULL;
5401 struct rbd_options *rbd_opts = NULL;
5402 struct rbd_spec *spec = NULL;
5403 struct rbd_client *rbdc;
5407 if (!try_module_get(THIS_MODULE))
5410 /* parse add command */
5411 rc = rbd_add_parse_args(buf, &ceph_opts, &rbd_opts, &spec);
5415 rbdc = rbd_get_client(ceph_opts);
5422 rc = rbd_add_get_pool_id(rbdc, spec->pool_name);
5425 pr_info("pool %s does not exist\n", spec->pool_name);
5426 goto err_out_client;
5428 spec->pool_id = (u64)rc;
5430 /* The ceph file layout needs to fit pool id in 32 bits */
5432 if (spec->pool_id > (u64)U32_MAX) {
5433 rbd_warn(NULL, "pool id too large (%llu > %u)",
5434 (unsigned long long)spec->pool_id, U32_MAX);
5436 goto err_out_client;
5439 rbd_dev = rbd_dev_create(rbdc, spec, rbd_opts);
5442 goto err_out_client;
5444 rbdc = NULL; /* rbd_dev now owns this */
5445 spec = NULL; /* rbd_dev now owns this */
5446 rbd_opts = NULL; /* rbd_dev now owns this */
5448 down_write(&rbd_dev->header_rwsem);
5449 rc = rbd_dev_image_probe(rbd_dev, 0);
5451 goto err_out_rbd_dev;
5453 /* If we are mapping a snapshot it must be marked read-only */
5455 read_only = rbd_dev->opts->read_only;
5456 if (rbd_dev->spec->snap_id != CEPH_NOSNAP)
5458 rbd_dev->mapping.read_only = read_only;
5460 rc = rbd_dev_device_setup(rbd_dev);
5463 * rbd_dev_header_unwatch_sync() can't be moved into
5464 * rbd_dev_image_release() without refactoring, see
5465 * commit 1f3ef78861ac.
5467 rbd_dev_header_unwatch_sync(rbd_dev);
5468 rbd_dev_image_release(rbd_dev);
5474 module_put(THIS_MODULE);
5478 up_write(&rbd_dev->header_rwsem);
5479 rbd_dev_destroy(rbd_dev);
5481 rbd_put_client(rbdc);
5488 static ssize_t rbd_add(struct bus_type *bus,
5495 return do_rbd_add(bus, buf, count);
5498 static ssize_t rbd_add_single_major(struct bus_type *bus,
5502 return do_rbd_add(bus, buf, count);
5505 static void rbd_dev_device_release(struct rbd_device *rbd_dev)
5507 rbd_free_disk(rbd_dev);
5508 clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
5509 device_del(&rbd_dev->dev);
5510 rbd_dev_mapping_clear(rbd_dev);
5512 unregister_blkdev(rbd_dev->major, rbd_dev->name);
5513 rbd_dev_id_put(rbd_dev);
5516 static void rbd_dev_remove_parent(struct rbd_device *rbd_dev)
5518 while (rbd_dev->parent) {
5519 struct rbd_device *first = rbd_dev;
5520 struct rbd_device *second = first->parent;
5521 struct rbd_device *third;
5524 * Follow to the parent with no grandparent and
5527 while (second && (third = second->parent)) {
5532 rbd_dev_image_release(second);
5533 first->parent = NULL;
5534 first->parent_overlap = 0;
5536 rbd_assert(first->parent_spec);
5537 rbd_spec_put(first->parent_spec);
5538 first->parent_spec = NULL;
5542 static ssize_t do_rbd_remove(struct bus_type *bus,
5546 struct rbd_device *rbd_dev = NULL;
5547 struct list_head *tmp;
5550 bool already = false;
5553 ret = kstrtoul(buf, 10, &ul);
5557 /* convert to int; abort if we lost anything in the conversion */
5563 spin_lock(&rbd_dev_list_lock);
5564 list_for_each(tmp, &rbd_dev_list) {
5565 rbd_dev = list_entry(tmp, struct rbd_device, node);
5566 if (rbd_dev->dev_id == dev_id) {
5572 spin_lock_irq(&rbd_dev->lock);
5573 if (rbd_dev->open_count)
5576 already = test_and_set_bit(RBD_DEV_FLAG_REMOVING,
5578 spin_unlock_irq(&rbd_dev->lock);
5580 spin_unlock(&rbd_dev_list_lock);
5581 if (ret < 0 || already)
5584 rbd_dev_header_unwatch_sync(rbd_dev);
5587 * Don't free anything from rbd_dev->disk until after all
5588 * notifies are completely processed. Otherwise
5589 * rbd_bus_del_dev() will race with rbd_watch_cb(), resulting
5590 * in a potential use after free of rbd_dev->disk or rbd_dev.
5592 rbd_dev_device_release(rbd_dev);
5593 rbd_dev_image_release(rbd_dev);
5598 static ssize_t rbd_remove(struct bus_type *bus,
5605 return do_rbd_remove(bus, buf, count);
5608 static ssize_t rbd_remove_single_major(struct bus_type *bus,
5612 return do_rbd_remove(bus, buf, count);
5616 * create control files in sysfs
5619 static int rbd_sysfs_init(void)
5623 ret = device_register(&rbd_root_dev);
5627 ret = bus_register(&rbd_bus_type);
5629 device_unregister(&rbd_root_dev);
5634 static void rbd_sysfs_cleanup(void)
5636 bus_unregister(&rbd_bus_type);
5637 device_unregister(&rbd_root_dev);
5640 static int rbd_slab_init(void)
5642 rbd_assert(!rbd_img_request_cache);
5643 rbd_img_request_cache = KMEM_CACHE(rbd_img_request, 0);
5644 if (!rbd_img_request_cache)
5647 rbd_assert(!rbd_obj_request_cache);
5648 rbd_obj_request_cache = KMEM_CACHE(rbd_obj_request, 0);
5649 if (!rbd_obj_request_cache)
5652 rbd_assert(!rbd_segment_name_cache);
5653 rbd_segment_name_cache = kmem_cache_create("rbd_segment_name",
5654 CEPH_MAX_OID_NAME_LEN + 1, 1, 0, NULL);
5655 if (rbd_segment_name_cache)
5658 kmem_cache_destroy(rbd_obj_request_cache);
5659 rbd_obj_request_cache = NULL;
5661 kmem_cache_destroy(rbd_img_request_cache);
5662 rbd_img_request_cache = NULL;
5667 static void rbd_slab_exit(void)
5669 rbd_assert(rbd_segment_name_cache);
5670 kmem_cache_destroy(rbd_segment_name_cache);
5671 rbd_segment_name_cache = NULL;
5673 rbd_assert(rbd_obj_request_cache);
5674 kmem_cache_destroy(rbd_obj_request_cache);
5675 rbd_obj_request_cache = NULL;
5677 rbd_assert(rbd_img_request_cache);
5678 kmem_cache_destroy(rbd_img_request_cache);
5679 rbd_img_request_cache = NULL;
5682 static int __init rbd_init(void)
5686 if (!libceph_compatible(NULL)) {
5687 rbd_warn(NULL, "libceph incompatibility (quitting)");
5691 rc = rbd_slab_init();
5696 * The number of active work items is limited by the number of
5697 * rbd devices * queue depth, so leave @max_active at default.
5699 rbd_wq = alloc_workqueue(RBD_DRV_NAME, WQ_MEM_RECLAIM, 0);
5706 rbd_major = register_blkdev(0, RBD_DRV_NAME);
5707 if (rbd_major < 0) {
5713 rc = rbd_sysfs_init();
5715 goto err_out_blkdev;
5718 pr_info("loaded (major %d)\n", rbd_major);
5720 pr_info("loaded\n");
5726 unregister_blkdev(rbd_major, RBD_DRV_NAME);
5728 destroy_workqueue(rbd_wq);
5734 static void __exit rbd_exit(void)
5736 ida_destroy(&rbd_dev_id_ida);
5737 rbd_sysfs_cleanup();
5739 unregister_blkdev(rbd_major, RBD_DRV_NAME);
5740 destroy_workqueue(rbd_wq);
5744 module_init(rbd_init);
5745 module_exit(rbd_exit);
5747 MODULE_AUTHOR("Alex Elder <elder@inktank.com>");
5748 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
5749 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
5750 /* following authorship retained from original osdblk.c */
5751 MODULE_AUTHOR("Jeff Garzik <jeff@garzik.org>");
5753 MODULE_DESCRIPTION("RADOS Block Device (RBD) driver");
5754 MODULE_LICENSE("GPL");