1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/sort.h>
6 #include <linux/slab.h>
7 #include <linux/iversion.h>
9 #include "mds_client.h"
10 #include <linux/ceph/decode.h>
12 /* unused map expires after 5 minutes */
13 #define CEPH_SNAPID_MAP_TIMEOUT (5 * 60 * HZ)
16 * Snapshots in ceph are driven in large part by cooperation from the
17 * client. In contrast to local file systems or file servers that
18 * implement snapshots at a single point in the system, ceph's
19 * distributed access to storage requires clients to help decide
20 * whether a write logically occurs before or after a recently created
23 * This provides a perfect instantanous client-wide snapshot. Between
24 * clients, however, snapshots may appear to be applied at slightly
25 * different points in time, depending on delays in delivering the
26 * snapshot notification.
28 * Snapshots are _not_ file system-wide. Instead, each snapshot
29 * applies to the subdirectory nested beneath some directory. This
30 * effectively divides the hierarchy into multiple "realms," where all
31 * of the files contained by each realm share the same set of
32 * snapshots. An individual realm's snap set contains snapshots
33 * explicitly created on that realm, as well as any snaps in its
34 * parent's snap set _after_ the point at which the parent became it's
35 * parent (due to, say, a rename). Similarly, snaps from prior parents
36 * during the time intervals during which they were the parent are included.
38 * The client is spared most of this detail, fortunately... it must only
39 * maintains a hierarchy of realms reflecting the current parent/child
40 * realm relationship, and for each realm has an explicit list of snaps
41 * inherited from prior parents.
43 * A snap_realm struct is maintained for realms containing every inode
44 * with an open cap in the system. (The needed snap realm information is
45 * provided by the MDS whenever a cap is issued, i.e., on open.) A 'seq'
46 * version number is used to ensure that as realm parameters change (new
47 * snapshot, new parent, etc.) the client's realm hierarchy is updated.
49 * The realm hierarchy drives the generation of a 'snap context' for each
50 * realm, which simply lists the resulting set of snaps for the realm. This
51 * is attached to any writes sent to OSDs.
54 * Unfortunately error handling is a bit mixed here. If we get a snap
55 * update, but don't have enough memory to update our realm hierarchy,
56 * it's not clear what we can do about it (besides complaining to the
62 * increase ref count for the realm
64 * caller must hold snap_rwsem.
66 void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
67 struct ceph_snap_realm *realm)
69 lockdep_assert_held(&mdsc->snap_rwsem);
72 * The 0->1 and 1->0 transitions must take the snap_empty_lock
73 * atomically with the refcount change. Go ahead and bump the
74 * nref here, unless it's 0, in which case we take the spinlock
75 * and then do the increment and remove it from the list.
77 if (atomic_inc_not_zero(&realm->nref))
80 spin_lock(&mdsc->snap_empty_lock);
81 if (atomic_inc_return(&realm->nref) == 1)
82 list_del_init(&realm->empty_item);
83 spin_unlock(&mdsc->snap_empty_lock);
86 static void __insert_snap_realm(struct rb_root *root,
87 struct ceph_snap_realm *new)
89 struct rb_node **p = &root->rb_node;
90 struct rb_node *parent = NULL;
91 struct ceph_snap_realm *r = NULL;
95 r = rb_entry(parent, struct ceph_snap_realm, node);
96 if (new->ino < r->ino)
98 else if (new->ino > r->ino)
104 rb_link_node(&new->node, parent, p);
105 rb_insert_color(&new->node, root);
109 * create and get the realm rooted at @ino and bump its ref count.
111 * caller must hold snap_rwsem for write.
113 static struct ceph_snap_realm *ceph_create_snap_realm(
114 struct ceph_mds_client *mdsc,
117 struct ceph_snap_realm *realm;
119 lockdep_assert_held_write(&mdsc->snap_rwsem);
121 realm = kzalloc(sizeof(*realm), GFP_NOFS);
123 return ERR_PTR(-ENOMEM);
125 /* Do not release the global dummy snaprealm until unmouting */
126 if (ino == CEPH_INO_GLOBAL_SNAPREALM)
127 atomic_set(&realm->nref, 2);
129 atomic_set(&realm->nref, 1);
131 INIT_LIST_HEAD(&realm->children);
132 INIT_LIST_HEAD(&realm->child_item);
133 INIT_LIST_HEAD(&realm->empty_item);
134 INIT_LIST_HEAD(&realm->dirty_item);
135 INIT_LIST_HEAD(&realm->rebuild_item);
136 INIT_LIST_HEAD(&realm->inodes_with_caps);
137 spin_lock_init(&realm->inodes_with_caps_lock);
138 __insert_snap_realm(&mdsc->snap_realms, realm);
139 mdsc->num_snap_realms++;
141 doutc(mdsc->fsc->client, "%llx %p\n", realm->ino, realm);
146 * lookup the realm rooted at @ino.
148 * caller must hold snap_rwsem.
150 static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
153 struct ceph_client *cl = mdsc->fsc->client;
154 struct rb_node *n = mdsc->snap_realms.rb_node;
155 struct ceph_snap_realm *r;
157 lockdep_assert_held(&mdsc->snap_rwsem);
160 r = rb_entry(n, struct ceph_snap_realm, node);
163 else if (ino > r->ino)
166 doutc(cl, "%llx %p\n", r->ino, r);
173 struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
176 struct ceph_snap_realm *r;
177 r = __lookup_snap_realm(mdsc, ino);
179 ceph_get_snap_realm(mdsc, r);
183 static void __put_snap_realm(struct ceph_mds_client *mdsc,
184 struct ceph_snap_realm *realm);
187 * called with snap_rwsem (write)
189 static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
190 struct ceph_snap_realm *realm)
192 struct ceph_client *cl = mdsc->fsc->client;
193 lockdep_assert_held_write(&mdsc->snap_rwsem);
195 doutc(cl, "%p %llx\n", realm, realm->ino);
197 rb_erase(&realm->node, &mdsc->snap_realms);
198 mdsc->num_snap_realms--;
201 list_del_init(&realm->child_item);
202 __put_snap_realm(mdsc, realm->parent);
205 kfree(realm->prior_parent_snaps);
207 ceph_put_snap_context(realm->cached_context);
212 * caller holds snap_rwsem (write)
214 static void __put_snap_realm(struct ceph_mds_client *mdsc,
215 struct ceph_snap_realm *realm)
217 lockdep_assert_held_write(&mdsc->snap_rwsem);
220 * We do not require the snap_empty_lock here, as any caller that
221 * increments the value must hold the snap_rwsem.
223 if (atomic_dec_and_test(&realm->nref))
224 __destroy_snap_realm(mdsc, realm);
228 * See comments in ceph_get_snap_realm. Caller needn't hold any locks.
230 void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
231 struct ceph_snap_realm *realm)
233 if (!atomic_dec_and_lock(&realm->nref, &mdsc->snap_empty_lock))
236 if (down_write_trylock(&mdsc->snap_rwsem)) {
237 spin_unlock(&mdsc->snap_empty_lock);
238 __destroy_snap_realm(mdsc, realm);
239 up_write(&mdsc->snap_rwsem);
241 list_add(&realm->empty_item, &mdsc->snap_empty);
242 spin_unlock(&mdsc->snap_empty_lock);
247 * Clean up any realms whose ref counts have dropped to zero. Note
248 * that this does not include realms who were created but not yet
251 * Called under snap_rwsem (write)
253 static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
255 struct ceph_snap_realm *realm;
257 lockdep_assert_held_write(&mdsc->snap_rwsem);
259 spin_lock(&mdsc->snap_empty_lock);
260 while (!list_empty(&mdsc->snap_empty)) {
261 realm = list_first_entry(&mdsc->snap_empty,
262 struct ceph_snap_realm, empty_item);
263 list_del(&realm->empty_item);
264 spin_unlock(&mdsc->snap_empty_lock);
265 __destroy_snap_realm(mdsc, realm);
266 spin_lock(&mdsc->snap_empty_lock);
268 spin_unlock(&mdsc->snap_empty_lock);
271 void ceph_cleanup_global_and_empty_realms(struct ceph_mds_client *mdsc)
273 struct ceph_snap_realm *global_realm;
275 down_write(&mdsc->snap_rwsem);
276 global_realm = __lookup_snap_realm(mdsc, CEPH_INO_GLOBAL_SNAPREALM);
278 ceph_put_snap_realm(mdsc, global_realm);
279 __cleanup_empty_realms(mdsc);
280 up_write(&mdsc->snap_rwsem);
284 * adjust the parent realm of a given @realm. adjust child list, and parent
285 * pointers, and ref counts appropriately.
287 * return true if parent was changed, 0 if unchanged, <0 on error.
289 * caller must hold snap_rwsem for write.
291 static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
292 struct ceph_snap_realm *realm,
295 struct ceph_client *cl = mdsc->fsc->client;
296 struct ceph_snap_realm *parent;
298 lockdep_assert_held_write(&mdsc->snap_rwsem);
300 if (realm->parent_ino == parentino)
303 parent = ceph_lookup_snap_realm(mdsc, parentino);
305 parent = ceph_create_snap_realm(mdsc, parentino);
307 return PTR_ERR(parent);
309 doutc(cl, "%llx %p: %llx %p -> %llx %p\n", realm->ino, realm,
310 realm->parent_ino, realm->parent, parentino, parent);
312 list_del_init(&realm->child_item);
313 ceph_put_snap_realm(mdsc, realm->parent);
315 realm->parent_ino = parentino;
316 realm->parent = parent;
317 list_add(&realm->child_item, &parent->children);
322 static int cmpu64_rev(const void *a, const void *b)
324 if (*(u64 *)a < *(u64 *)b)
326 if (*(u64 *)a > *(u64 *)b)
333 * build the snap context for a given realm.
335 static int build_snap_context(struct ceph_mds_client *mdsc,
336 struct ceph_snap_realm *realm,
337 struct list_head *realm_queue,
338 struct list_head *dirty_realms)
340 struct ceph_client *cl = mdsc->fsc->client;
341 struct ceph_snap_realm *parent = realm->parent;
342 struct ceph_snap_context *snapc;
344 u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
347 * build parent context, if it hasn't been built.
348 * conservatively estimate that all parent snaps might be
352 if (!parent->cached_context) {
353 /* add to the queue head */
354 list_add(&parent->rebuild_item, realm_queue);
357 num += parent->cached_context->num_snaps;
360 /* do i actually need to update? not if my context seq
361 matches realm seq, and my parents' does to. (this works
362 because we rebuild_snap_realms() works _downward_ in
363 hierarchy after each update.) */
364 if (realm->cached_context &&
365 realm->cached_context->seq == realm->seq &&
367 realm->cached_context->seq >= parent->cached_context->seq)) {
368 doutc(cl, "%llx %p: %p seq %lld (%u snaps) (unchanged)\n",
369 realm->ino, realm, realm->cached_context,
370 realm->cached_context->seq,
371 (unsigned int)realm->cached_context->num_snaps);
375 /* alloc new snap context */
377 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
379 snapc = ceph_create_snap_context(num, GFP_NOFS);
383 /* build (reverse sorted) snap vector */
385 snapc->seq = realm->seq;
389 /* include any of parent's snaps occurring _after_ my
390 parent became my parent */
391 for (i = 0; i < parent->cached_context->num_snaps; i++)
392 if (parent->cached_context->snaps[i] >=
394 snapc->snaps[num++] =
395 parent->cached_context->snaps[i];
396 if (parent->cached_context->seq > snapc->seq)
397 snapc->seq = parent->cached_context->seq;
399 memcpy(snapc->snaps + num, realm->snaps,
400 sizeof(u64)*realm->num_snaps);
401 num += realm->num_snaps;
402 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
403 sizeof(u64)*realm->num_prior_parent_snaps);
404 num += realm->num_prior_parent_snaps;
406 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
407 snapc->num_snaps = num;
408 doutc(cl, "%llx %p: %p seq %lld (%u snaps)\n", realm->ino, realm,
409 snapc, snapc->seq, (unsigned int) snapc->num_snaps);
411 ceph_put_snap_context(realm->cached_context);
412 realm->cached_context = snapc;
413 /* queue realm for cap_snap creation */
414 list_add_tail(&realm->dirty_item, dirty_realms);
419 * if we fail, clear old (incorrect) cached_context... hopefully
420 * we'll have better luck building it later
422 if (realm->cached_context) {
423 ceph_put_snap_context(realm->cached_context);
424 realm->cached_context = NULL;
426 pr_err_client(cl, "%llx %p fail %d\n", realm->ino, realm, err);
431 * rebuild snap context for the given realm and all of its children.
433 static void rebuild_snap_realms(struct ceph_mds_client *mdsc,
434 struct ceph_snap_realm *realm,
435 struct list_head *dirty_realms)
437 struct ceph_client *cl = mdsc->fsc->client;
438 LIST_HEAD(realm_queue);
442 list_add_tail(&realm->rebuild_item, &realm_queue);
444 while (!list_empty(&realm_queue)) {
445 struct ceph_snap_realm *_realm, *child;
447 _realm = list_first_entry(&realm_queue,
448 struct ceph_snap_realm,
452 * If the last building failed dues to memory
453 * issue, just empty the realm_queue and return
454 * to avoid infinite loop.
457 list_del_init(&_realm->rebuild_item);
461 last = build_snap_context(mdsc, _realm, &realm_queue,
463 doutc(cl, "%llx %p, %s\n", realm->ino, realm,
464 last > 0 ? "is deferred" : !last ? "succeeded" : "failed");
466 /* is any child in the list ? */
467 list_for_each_entry(child, &_realm->children, child_item) {
468 if (!list_empty(&child->rebuild_item)) {
475 list_for_each_entry(child, &_realm->children, child_item)
476 list_add_tail(&child->rebuild_item, &realm_queue);
479 /* last == 1 means need to build parent first */
481 list_del_init(&_realm->rebuild_item);
487 * helper to allocate and decode an array of snapids. free prior
490 static int dup_array(u64 **dst, __le64 *src, u32 num)
496 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
499 for (i = 0; i < num; i++)
500 (*dst)[i] = get_unaligned_le64(src + i);
507 static bool has_new_snaps(struct ceph_snap_context *o,
508 struct ceph_snap_context *n)
510 if (n->num_snaps == 0)
512 /* snaps are in descending order */
513 return n->snaps[0] > o->seq;
517 * When a snapshot is applied, the size/mtime inode metadata is queued
518 * in a ceph_cap_snap (one for each snapshot) until writeback
519 * completes and the metadata can be flushed back to the MDS.
521 * However, if a (sync) write is currently in-progress when we apply
522 * the snapshot, we have to wait until the write succeeds or fails
523 * (and a final size/mtime is known). In this case the
524 * cap_snap->writing = 1, and is said to be "pending." When the write
525 * finishes, we __ceph_finish_cap_snap().
527 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
530 static void ceph_queue_cap_snap(struct ceph_inode_info *ci,
531 struct ceph_cap_snap **pcapsnap)
533 struct inode *inode = &ci->netfs.inode;
534 struct ceph_client *cl = ceph_inode_to_client(inode);
535 struct ceph_snap_context *old_snapc, *new_snapc;
536 struct ceph_cap_snap *capsnap = *pcapsnap;
537 struct ceph_buffer *old_blob = NULL;
540 spin_lock(&ci->i_ceph_lock);
541 used = __ceph_caps_used(ci);
542 dirty = __ceph_caps_dirty(ci);
544 old_snapc = ci->i_head_snapc;
545 new_snapc = ci->i_snap_realm->cached_context;
548 * If there is a write in progress, treat that as a dirty Fw,
549 * even though it hasn't completed yet; by the time we finish
550 * up this capsnap it will be.
552 if (used & CEPH_CAP_FILE_WR)
553 dirty |= CEPH_CAP_FILE_WR;
555 if (__ceph_have_pending_cap_snap(ci)) {
556 /* there is no point in queuing multiple "pending" cap_snaps,
557 as no new writes are allowed to start when pending, so any
558 writes in progress now were started before the previous
559 cap_snap. lucky us. */
560 doutc(cl, "%p %llx.%llx already pending\n", inode,
564 if (ci->i_wrbuffer_ref_head == 0 &&
565 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
566 doutc(cl, "%p %llx.%llx nothing dirty|writing\n", inode,
574 * There is no need to send FLUSHSNAP message to MDS if there is
575 * no new snapshot. But when there is dirty pages or on-going
576 * writes, we still need to create cap_snap. cap_snap is needed
577 * by the write path and page writeback path.
579 * also see ceph_try_drop_cap_snap()
581 if (has_new_snaps(old_snapc, new_snapc)) {
582 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
583 capsnap->need_flush = true;
585 if (!(used & CEPH_CAP_FILE_WR) &&
586 ci->i_wrbuffer_ref_head == 0) {
587 doutc(cl, "%p %llx.%llx no new_snap|dirty_page|writing\n",
588 inode, ceph_vinop(inode));
593 doutc(cl, "%p %llx.%llx cap_snap %p queuing under %p %s %s\n",
594 inode, ceph_vinop(inode), capsnap, old_snapc,
595 ceph_cap_string(dirty), capsnap->need_flush ? "" : "no_flush");
598 capsnap->follows = old_snapc->seq;
599 capsnap->issued = __ceph_caps_issued(ci, NULL);
600 capsnap->dirty = dirty;
602 capsnap->mode = inode->i_mode;
603 capsnap->uid = inode->i_uid;
604 capsnap->gid = inode->i_gid;
606 if (dirty & CEPH_CAP_XATTR_EXCL) {
607 old_blob = __ceph_build_xattrs_blob(ci);
608 capsnap->xattr_blob =
609 ceph_buffer_get(ci->i_xattrs.blob);
610 capsnap->xattr_version = ci->i_xattrs.version;
612 capsnap->xattr_blob = NULL;
613 capsnap->xattr_version = 0;
616 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
618 /* dirty page count moved from _head to this cap_snap;
619 all subsequent writes page dirties occur _after_ this
621 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
622 ci->i_wrbuffer_ref_head = 0;
623 capsnap->context = old_snapc;
624 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
626 if (used & CEPH_CAP_FILE_WR) {
627 doutc(cl, "%p %llx.%llx cap_snap %p snapc %p seq %llu used WR,"
628 " now pending\n", inode, ceph_vinop(inode), capsnap,
629 old_snapc, old_snapc->seq);
630 capsnap->writing = 1;
632 /* note mtime, size NOW. */
633 __ceph_finish_cap_snap(ci, capsnap);
639 if (ci->i_wrbuffer_ref_head == 0 &&
641 ci->i_dirty_caps == 0 &&
642 ci->i_flushing_caps == 0) {
643 ci->i_head_snapc = NULL;
645 ci->i_head_snapc = ceph_get_snap_context(new_snapc);
646 doutc(cl, " new snapc is %p\n", new_snapc);
648 spin_unlock(&ci->i_ceph_lock);
650 ceph_buffer_put(old_blob);
651 ceph_put_snap_context(old_snapc);
655 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
656 * to be used for the snapshot, to be flushed back to the mds.
658 * If capsnap can now be flushed, add to snap_flush list, and return 1.
660 * Caller must hold i_ceph_lock.
662 int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
663 struct ceph_cap_snap *capsnap)
665 struct inode *inode = &ci->netfs.inode;
666 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
667 struct ceph_client *cl = mdsc->fsc->client;
669 BUG_ON(capsnap->writing);
670 capsnap->size = i_size_read(inode);
671 capsnap->mtime = inode_get_mtime(inode);
672 capsnap->atime = inode_get_atime(inode);
673 capsnap->ctime = inode_get_ctime(inode);
674 capsnap->btime = ci->i_btime;
675 capsnap->change_attr = inode_peek_iversion_raw(inode);
676 capsnap->time_warp_seq = ci->i_time_warp_seq;
677 capsnap->truncate_size = ci->i_truncate_size;
678 capsnap->truncate_seq = ci->i_truncate_seq;
679 if (capsnap->dirty_pages) {
680 doutc(cl, "%p %llx.%llx cap_snap %p snapc %p %llu %s "
681 "s=%llu still has %d dirty pages\n", inode,
682 ceph_vinop(inode), capsnap, capsnap->context,
683 capsnap->context->seq,
684 ceph_cap_string(capsnap->dirty),
685 capsnap->size, capsnap->dirty_pages);
690 * Defer flushing the capsnap if the dirty buffer not flushed yet.
691 * And trigger to flush the buffer immediately.
693 if (ci->i_wrbuffer_ref) {
694 doutc(cl, "%p %llx.%llx cap_snap %p snapc %p %llu %s "
695 "s=%llu used WRBUFFER, delaying\n", inode,
696 ceph_vinop(inode), capsnap, capsnap->context,
697 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
699 ceph_queue_writeback(inode);
703 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
704 doutc(cl, "%p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu\n",
705 inode, ceph_vinop(inode), capsnap, capsnap->context,
706 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
709 spin_lock(&mdsc->snap_flush_lock);
710 if (list_empty(&ci->i_snap_flush_item)) {
712 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
714 spin_unlock(&mdsc->snap_flush_lock);
715 return 1; /* caller may want to ceph_flush_snaps */
719 * Queue cap_snaps for snap writeback for this realm and its children.
720 * Called under snap_rwsem, so realm topology won't change.
722 static void queue_realm_cap_snaps(struct ceph_mds_client *mdsc,
723 struct ceph_snap_realm *realm)
725 struct ceph_client *cl = mdsc->fsc->client;
726 struct ceph_inode_info *ci;
727 struct inode *lastinode = NULL;
728 struct ceph_cap_snap *capsnap = NULL;
730 doutc(cl, "%p %llx inode\n", realm, realm->ino);
732 spin_lock(&realm->inodes_with_caps_lock);
733 list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
734 struct inode *inode = igrab(&ci->netfs.inode);
737 spin_unlock(&realm->inodes_with_caps_lock);
742 * Allocate the capsnap memory outside of ceph_queue_cap_snap()
743 * to reduce very possible but unnecessary frequently memory
744 * allocate/free in this loop.
747 capsnap = kmem_cache_zalloc(ceph_cap_snap_cachep, GFP_NOFS);
750 "ENOMEM allocating ceph_cap_snap on %p\n",
755 capsnap->cap_flush.is_capsnap = true;
756 refcount_set(&capsnap->nref, 1);
757 INIT_LIST_HEAD(&capsnap->cap_flush.i_list);
758 INIT_LIST_HEAD(&capsnap->cap_flush.g_list);
759 INIT_LIST_HEAD(&capsnap->ci_item);
761 ceph_queue_cap_snap(ci, &capsnap);
762 spin_lock(&realm->inodes_with_caps_lock);
764 spin_unlock(&realm->inodes_with_caps_lock);
768 kmem_cache_free(ceph_cap_snap_cachep, capsnap);
769 doutc(cl, "%p %llx done\n", realm, realm->ino);
773 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
774 * the snap realm parameters from a given realm and all of its ancestors,
777 * Caller must hold snap_rwsem for write.
779 int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
780 void *p, void *e, bool deletion,
781 struct ceph_snap_realm **realm_ret)
783 struct ceph_client *cl = mdsc->fsc->client;
784 struct ceph_mds_snap_realm *ri; /* encoded */
785 __le64 *snaps; /* encoded */
786 __le64 *prior_parent_snaps; /* encoded */
787 struct ceph_snap_realm *realm;
788 struct ceph_snap_realm *first_realm = NULL;
789 struct ceph_snap_realm *realm_to_rebuild = NULL;
790 struct ceph_client *client = mdsc->fsc->client;
794 LIST_HEAD(dirty_realms);
796 lockdep_assert_held_write(&mdsc->snap_rwsem);
798 doutc(cl, "deletion=%d\n", deletion);
802 ceph_decode_need(&p, e, sizeof(*ri), bad);
805 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
806 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
808 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
809 prior_parent_snaps = p;
810 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
812 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
814 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
816 err = PTR_ERR(realm);
821 /* ensure the parent is correct */
822 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
825 rebuild_snapcs += err;
827 if (le64_to_cpu(ri->seq) > realm->seq) {
828 doutc(cl, "updating %llx %p %lld -> %lld\n", realm->ino,
829 realm, realm->seq, le64_to_cpu(ri->seq));
830 /* update realm parameters, snap lists */
831 realm->seq = le64_to_cpu(ri->seq);
832 realm->created = le64_to_cpu(ri->created);
833 realm->parent_since = le64_to_cpu(ri->parent_since);
835 realm->num_snaps = le32_to_cpu(ri->num_snaps);
836 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
840 realm->num_prior_parent_snaps =
841 le32_to_cpu(ri->num_prior_parent_snaps);
842 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
843 realm->num_prior_parent_snaps);
847 if (realm->seq > mdsc->last_snap_seq)
848 mdsc->last_snap_seq = realm->seq;
851 } else if (!realm->cached_context) {
852 doutc(cl, "%llx %p seq %lld new\n", realm->ino, realm,
856 doutc(cl, "%llx %p seq %lld unchanged\n", realm->ino, realm,
860 doutc(cl, "done with %llx %p, rebuild_snapcs=%d, %p %p\n", realm->ino,
861 realm, rebuild_snapcs, p, e);
864 * this will always track the uppest parent realm from which
865 * we need to rebuild the snapshot contexts _downward_ in
869 realm_to_rebuild = realm;
871 /* rebuild_snapcs when we reach the _end_ (root) of the trace */
872 if (realm_to_rebuild && p >= e)
873 rebuild_snap_realms(mdsc, realm_to_rebuild, &dirty_realms);
878 ceph_put_snap_realm(mdsc, realm);
884 * queue cap snaps _after_ we've built the new snap contexts,
885 * so that i_head_snapc can be set appropriately.
887 while (!list_empty(&dirty_realms)) {
888 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
890 list_del_init(&realm->dirty_item);
891 queue_realm_cap_snaps(mdsc, realm);
895 *realm_ret = first_realm;
897 ceph_put_snap_realm(mdsc, first_realm);
899 __cleanup_empty_realms(mdsc);
905 if (realm && !IS_ERR(realm))
906 ceph_put_snap_realm(mdsc, realm);
908 ceph_put_snap_realm(mdsc, first_realm);
909 pr_err_client(cl, "error %d\n", err);
912 * When receiving a corrupted snap trace we don't know what
913 * exactly has happened in MDS side. And we shouldn't continue
914 * writing to OSD, which may corrupt the snapshot contents.
916 * Just try to blocklist this kclient and then this kclient
917 * must be remounted to continue after the corrupted metadata
918 * fixed in the MDS side.
920 WRITE_ONCE(mdsc->fsc->mount_state, CEPH_MOUNT_FENCE_IO);
921 ret = ceph_monc_blocklist_add(&client->monc, &client->msgr.inst.addr);
923 pr_err_client(cl, "failed to blocklist %s: %d\n",
924 ceph_pr_addr(&client->msgr.inst.addr), ret);
926 WARN(1, "[client.%lld] %s %s%sdo remount to continue%s",
927 client->monc.auth->global_id, __func__,
928 ret ? "" : ceph_pr_addr(&client->msgr.inst.addr),
929 ret ? "" : " was blocklisted, ",
930 err == -EIO ? " after corrupted snaptrace is fixed" : "");
937 * Send any cap_snaps that are queued for flush. Try to carry
938 * s_mutex across multiple snap flushes to avoid locking overhead.
940 * Caller holds no locks.
942 static void flush_snaps(struct ceph_mds_client *mdsc)
944 struct ceph_client *cl = mdsc->fsc->client;
945 struct ceph_inode_info *ci;
947 struct ceph_mds_session *session = NULL;
949 doutc(cl, "begin\n");
950 spin_lock(&mdsc->snap_flush_lock);
951 while (!list_empty(&mdsc->snap_flush_list)) {
952 ci = list_first_entry(&mdsc->snap_flush_list,
953 struct ceph_inode_info, i_snap_flush_item);
954 inode = &ci->netfs.inode;
956 spin_unlock(&mdsc->snap_flush_lock);
957 ceph_flush_snaps(ci, &session);
959 spin_lock(&mdsc->snap_flush_lock);
961 spin_unlock(&mdsc->snap_flush_lock);
963 ceph_put_mds_session(session);
968 * ceph_change_snap_realm - change the snap_realm for an inode
969 * @inode: inode to move to new snap realm
970 * @realm: new realm to move inode into (may be NULL)
972 * Detach an inode from its old snaprealm (if any) and attach it to
973 * the new snaprealm (if any). The old snap realm reference held by
974 * the inode is put. If realm is non-NULL, then the caller's reference
975 * to it is taken over by the inode.
977 void ceph_change_snap_realm(struct inode *inode, struct ceph_snap_realm *realm)
979 struct ceph_inode_info *ci = ceph_inode(inode);
980 struct ceph_mds_client *mdsc = ceph_inode_to_fs_client(inode)->mdsc;
981 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
983 lockdep_assert_held(&ci->i_ceph_lock);
986 spin_lock(&oldrealm->inodes_with_caps_lock);
987 list_del_init(&ci->i_snap_realm_item);
988 if (oldrealm->ino == ci->i_vino.ino)
989 oldrealm->inode = NULL;
990 spin_unlock(&oldrealm->inodes_with_caps_lock);
991 ceph_put_snap_realm(mdsc, oldrealm);
994 ci->i_snap_realm = realm;
997 spin_lock(&realm->inodes_with_caps_lock);
998 list_add(&ci->i_snap_realm_item, &realm->inodes_with_caps);
999 if (realm->ino == ci->i_vino.ino)
1000 realm->inode = inode;
1001 spin_unlock(&realm->inodes_with_caps_lock);
1006 * Handle a snap notification from the MDS.
1008 * This can take two basic forms: the simplest is just a snap creation
1009 * or deletion notification on an existing realm. This should update the
1010 * realm and its children.
1012 * The more difficult case is realm creation, due to snap creation at a
1013 * new point in the file hierarchy, or due to a rename that moves a file or
1014 * directory into another realm.
1016 void ceph_handle_snap(struct ceph_mds_client *mdsc,
1017 struct ceph_mds_session *session,
1018 struct ceph_msg *msg)
1020 struct ceph_client *cl = mdsc->fsc->client;
1021 struct super_block *sb = mdsc->fsc->sb;
1022 int mds = session->s_mds;
1026 struct ceph_snap_realm *realm = NULL;
1027 void *p = msg->front.iov_base;
1028 void *e = p + msg->front.iov_len;
1029 struct ceph_mds_snap_head *h;
1030 int num_split_inos, num_split_realms;
1031 __le64 *split_inos = NULL, *split_realms = NULL;
1033 int locked_rwsem = 0;
1034 bool close_sessions = false;
1036 if (!ceph_inc_mds_stopping_blocker(mdsc, session))
1040 if (msg->front.iov_len < sizeof(*h))
1043 op = le32_to_cpu(h->op);
1044 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
1046 num_split_inos = le32_to_cpu(h->num_split_inos);
1047 num_split_realms = le32_to_cpu(h->num_split_realms);
1048 trace_len = le32_to_cpu(h->trace_len);
1051 doutc(cl, "from mds%d op %s split %llx tracelen %d\n", mds,
1052 ceph_snap_op_name(op), split, trace_len);
1054 down_write(&mdsc->snap_rwsem);
1057 if (op == CEPH_SNAP_OP_SPLIT) {
1058 struct ceph_mds_snap_realm *ri;
1061 * A "split" breaks part of an existing realm off into
1062 * a new realm. The MDS provides a list of inodes
1063 * (with caps) and child realms that belong to the new
1067 p += sizeof(u64) * num_split_inos;
1069 p += sizeof(u64) * num_split_realms;
1070 ceph_decode_need(&p, e, sizeof(*ri), bad);
1071 /* we will peek at realm info here, but will _not_
1072 * advance p, as the realm update will occur below in
1073 * ceph_update_snap_trace. */
1076 realm = ceph_lookup_snap_realm(mdsc, split);
1078 realm = ceph_create_snap_realm(mdsc, split);
1083 doutc(cl, "splitting snap_realm %llx %p\n", realm->ino, realm);
1084 for (i = 0; i < num_split_inos; i++) {
1085 struct ceph_vino vino = {
1086 .ino = le64_to_cpu(split_inos[i]),
1087 .snap = CEPH_NOSNAP,
1089 struct inode *inode = ceph_find_inode(sb, vino);
1090 struct ceph_inode_info *ci;
1094 ci = ceph_inode(inode);
1096 spin_lock(&ci->i_ceph_lock);
1097 if (!ci->i_snap_realm)
1100 * If this inode belongs to a realm that was
1101 * created after our new realm, we experienced
1102 * a race (due to another split notifications
1103 * arriving from a different MDS). So skip
1106 if (ci->i_snap_realm->created >
1107 le64_to_cpu(ri->created)) {
1108 doutc(cl, " leaving %p %llx.%llx in newer realm %llx %p\n",
1109 inode, ceph_vinop(inode), ci->i_snap_realm->ino,
1113 doutc(cl, " will move %p %llx.%llx to split realm %llx %p\n",
1114 inode, ceph_vinop(inode), realm->ino, realm);
1116 ceph_get_snap_realm(mdsc, realm);
1117 ceph_change_snap_realm(inode, realm);
1118 spin_unlock(&ci->i_ceph_lock);
1123 spin_unlock(&ci->i_ceph_lock);
1127 /* we may have taken some of the old realm's children. */
1128 for (i = 0; i < num_split_realms; i++) {
1129 struct ceph_snap_realm *child =
1130 __lookup_snap_realm(mdsc,
1131 le64_to_cpu(split_realms[i]));
1134 adjust_snap_realm_parent(mdsc, child, realm->ino);
1138 * In the non-split case both 'num_split_inos' and
1139 * 'num_split_realms' should be 0, making this a no-op.
1140 * However the MDS happens to populate 'split_realms' list
1141 * in one of the UPDATE op cases by mistake.
1143 * Skip both lists just in case to ensure that 'p' is
1144 * positioned at the start of realm info, as expected by
1145 * ceph_update_snap_trace().
1147 p += sizeof(u64) * num_split_inos;
1148 p += sizeof(u64) * num_split_realms;
1152 * update using the provided snap trace. if we are deleting a
1153 * snap, we can avoid queueing cap_snaps.
1155 if (ceph_update_snap_trace(mdsc, p, e,
1156 op == CEPH_SNAP_OP_DESTROY,
1158 close_sessions = true;
1162 if (op == CEPH_SNAP_OP_SPLIT)
1163 /* we took a reference when we created the realm, above */
1164 ceph_put_snap_realm(mdsc, realm);
1166 __cleanup_empty_realms(mdsc);
1168 up_write(&mdsc->snap_rwsem);
1171 ceph_dec_mds_stopping_blocker(mdsc);
1175 pr_err_client(cl, "corrupt snap message from mds%d\n", mds);
1179 up_write(&mdsc->snap_rwsem);
1181 ceph_dec_mds_stopping_blocker(mdsc);
1184 ceph_mdsc_close_sessions(mdsc);
1188 struct ceph_snapid_map* ceph_get_snapid_map(struct ceph_mds_client *mdsc,
1191 struct ceph_client *cl = mdsc->fsc->client;
1192 struct ceph_snapid_map *sm, *exist;
1193 struct rb_node **p, *parent;
1197 spin_lock(&mdsc->snapid_map_lock);
1198 p = &mdsc->snapid_map_tree.rb_node;
1200 exist = rb_entry(*p, struct ceph_snapid_map, node);
1201 if (snap > exist->snap) {
1203 } else if (snap < exist->snap) {
1204 p = &(*p)->rb_right;
1206 if (atomic_inc_return(&exist->ref) == 1)
1207 list_del_init(&exist->lru);
1212 spin_unlock(&mdsc->snapid_map_lock);
1214 doutc(cl, "found snapid map %llx -> %x\n", exist->snap,
1219 sm = kmalloc(sizeof(*sm), GFP_NOFS);
1223 ret = get_anon_bdev(&sm->dev);
1229 INIT_LIST_HEAD(&sm->lru);
1230 atomic_set(&sm->ref, 1);
1235 p = &mdsc->snapid_map_tree.rb_node;
1236 spin_lock(&mdsc->snapid_map_lock);
1239 exist = rb_entry(*p, struct ceph_snapid_map, node);
1240 if (snap > exist->snap)
1242 else if (snap < exist->snap)
1243 p = &(*p)->rb_right;
1249 if (atomic_inc_return(&exist->ref) == 1)
1250 list_del_init(&exist->lru);
1252 rb_link_node(&sm->node, parent, p);
1253 rb_insert_color(&sm->node, &mdsc->snapid_map_tree);
1255 spin_unlock(&mdsc->snapid_map_lock);
1257 free_anon_bdev(sm->dev);
1259 doutc(cl, "found snapid map %llx -> %x\n", exist->snap,
1264 doutc(cl, "create snapid map %llx -> %x\n", sm->snap, sm->dev);
1268 void ceph_put_snapid_map(struct ceph_mds_client* mdsc,
1269 struct ceph_snapid_map *sm)
1273 if (atomic_dec_and_lock(&sm->ref, &mdsc->snapid_map_lock)) {
1274 if (!RB_EMPTY_NODE(&sm->node)) {
1275 sm->last_used = jiffies;
1276 list_add_tail(&sm->lru, &mdsc->snapid_map_lru);
1277 spin_unlock(&mdsc->snapid_map_lock);
1279 /* already cleaned up by
1280 * ceph_cleanup_snapid_map() */
1281 spin_unlock(&mdsc->snapid_map_lock);
1287 void ceph_trim_snapid_map(struct ceph_mds_client *mdsc)
1289 struct ceph_client *cl = mdsc->fsc->client;
1290 struct ceph_snapid_map *sm;
1294 spin_lock(&mdsc->snapid_map_lock);
1297 while (!list_empty(&mdsc->snapid_map_lru)) {
1298 sm = list_first_entry(&mdsc->snapid_map_lru,
1299 struct ceph_snapid_map, lru);
1300 if (time_after(sm->last_used + CEPH_SNAPID_MAP_TIMEOUT, now))
1303 rb_erase(&sm->node, &mdsc->snapid_map_tree);
1304 list_move(&sm->lru, &to_free);
1306 spin_unlock(&mdsc->snapid_map_lock);
1308 while (!list_empty(&to_free)) {
1309 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1311 doutc(cl, "trim snapid map %llx -> %x\n", sm->snap, sm->dev);
1312 free_anon_bdev(sm->dev);
1317 void ceph_cleanup_snapid_map(struct ceph_mds_client *mdsc)
1319 struct ceph_client *cl = mdsc->fsc->client;
1320 struct ceph_snapid_map *sm;
1324 spin_lock(&mdsc->snapid_map_lock);
1325 while ((p = rb_first(&mdsc->snapid_map_tree))) {
1326 sm = rb_entry(p, struct ceph_snapid_map, node);
1327 rb_erase(p, &mdsc->snapid_map_tree);
1329 list_move(&sm->lru, &to_free);
1331 spin_unlock(&mdsc->snapid_map_lock);
1333 while (!list_empty(&to_free)) {
1334 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1336 free_anon_bdev(sm->dev);
1337 if (WARN_ON_ONCE(atomic_read(&sm->ref))) {
1338 pr_err_client(cl, "snapid map %llx -> %x still in use\n",