Loading...
1#include <linux/ceph/ceph_debug.h>
2
3#include <linux/sort.h>
4#include <linux/slab.h>
5
6#include "super.h"
7#include "mds_client.h"
8
9#include <linux/ceph/decode.h>
10
11/*
12 * Snapshots in ceph are driven in large part by cooperation from the
13 * client. In contrast to local file systems or file servers that
14 * implement snapshots at a single point in the system, ceph's
15 * distributed access to storage requires clients to help decide
16 * whether a write logically occurs before or after a recently created
17 * snapshot.
18 *
19 * This provides a perfect instantanous client-wide snapshot. Between
20 * clients, however, snapshots may appear to be applied at slightly
21 * different points in time, depending on delays in delivering the
22 * snapshot notification.
23 *
24 * Snapshots are _not_ file system-wide. Instead, each snapshot
25 * applies to the subdirectory nested beneath some directory. This
26 * effectively divides the hierarchy into multiple "realms," where all
27 * of the files contained by each realm share the same set of
28 * snapshots. An individual realm's snap set contains snapshots
29 * explicitly created on that realm, as well as any snaps in its
30 * parent's snap set _after_ the point at which the parent became it's
31 * parent (due to, say, a rename). Similarly, snaps from prior parents
32 * during the time intervals during which they were the parent are included.
33 *
34 * The client is spared most of this detail, fortunately... it must only
35 * maintains a hierarchy of realms reflecting the current parent/child
36 * realm relationship, and for each realm has an explicit list of snaps
37 * inherited from prior parents.
38 *
39 * A snap_realm struct is maintained for realms containing every inode
40 * with an open cap in the system. (The needed snap realm information is
41 * provided by the MDS whenever a cap is issued, i.e., on open.) A 'seq'
42 * version number is used to ensure that as realm parameters change (new
43 * snapshot, new parent, etc.) the client's realm hierarchy is updated.
44 *
45 * The realm hierarchy drives the generation of a 'snap context' for each
46 * realm, which simply lists the resulting set of snaps for the realm. This
47 * is attached to any writes sent to OSDs.
48 */
49/*
50 * Unfortunately error handling is a bit mixed here. If we get a snap
51 * update, but don't have enough memory to update our realm hierarchy,
52 * it's not clear what we can do about it (besides complaining to the
53 * console).
54 */
55
56
57/*
58 * increase ref count for the realm
59 *
60 * caller must hold snap_rwsem for write.
61 */
62void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
63 struct ceph_snap_realm *realm)
64{
65 dout("get_realm %p %d -> %d\n", realm,
66 atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
67 /*
68 * since we _only_ increment realm refs or empty the empty
69 * list with snap_rwsem held, adjusting the empty list here is
70 * safe. we do need to protect against concurrent empty list
71 * additions, however.
72 */
73 if (atomic_read(&realm->nref) == 0) {
74 spin_lock(&mdsc->snap_empty_lock);
75 list_del_init(&realm->empty_item);
76 spin_unlock(&mdsc->snap_empty_lock);
77 }
78
79 atomic_inc(&realm->nref);
80}
81
82static void __insert_snap_realm(struct rb_root *root,
83 struct ceph_snap_realm *new)
84{
85 struct rb_node **p = &root->rb_node;
86 struct rb_node *parent = NULL;
87 struct ceph_snap_realm *r = NULL;
88
89 while (*p) {
90 parent = *p;
91 r = rb_entry(parent, struct ceph_snap_realm, node);
92 if (new->ino < r->ino)
93 p = &(*p)->rb_left;
94 else if (new->ino > r->ino)
95 p = &(*p)->rb_right;
96 else
97 BUG();
98 }
99
100 rb_link_node(&new->node, parent, p);
101 rb_insert_color(&new->node, root);
102}
103
104/*
105 * create and get the realm rooted at @ino and bump its ref count.
106 *
107 * caller must hold snap_rwsem for write.
108 */
109static struct ceph_snap_realm *ceph_create_snap_realm(
110 struct ceph_mds_client *mdsc,
111 u64 ino)
112{
113 struct ceph_snap_realm *realm;
114
115 realm = kzalloc(sizeof(*realm), GFP_NOFS);
116 if (!realm)
117 return ERR_PTR(-ENOMEM);
118
119 atomic_set(&realm->nref, 0); /* tree does not take a ref */
120 realm->ino = ino;
121 INIT_LIST_HEAD(&realm->children);
122 INIT_LIST_HEAD(&realm->child_item);
123 INIT_LIST_HEAD(&realm->empty_item);
124 INIT_LIST_HEAD(&realm->dirty_item);
125 INIT_LIST_HEAD(&realm->inodes_with_caps);
126 spin_lock_init(&realm->inodes_with_caps_lock);
127 __insert_snap_realm(&mdsc->snap_realms, realm);
128 dout("create_snap_realm %llx %p\n", realm->ino, realm);
129 return realm;
130}
131
132/*
133 * lookup the realm rooted at @ino.
134 *
135 * caller must hold snap_rwsem for write.
136 */
137struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
138 u64 ino)
139{
140 struct rb_node *n = mdsc->snap_realms.rb_node;
141 struct ceph_snap_realm *r;
142
143 while (n) {
144 r = rb_entry(n, struct ceph_snap_realm, node);
145 if (ino < r->ino)
146 n = n->rb_left;
147 else if (ino > r->ino)
148 n = n->rb_right;
149 else {
150 dout("lookup_snap_realm %llx %p\n", r->ino, r);
151 return r;
152 }
153 }
154 return NULL;
155}
156
157static void __put_snap_realm(struct ceph_mds_client *mdsc,
158 struct ceph_snap_realm *realm);
159
160/*
161 * called with snap_rwsem (write)
162 */
163static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
164 struct ceph_snap_realm *realm)
165{
166 dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
167
168 rb_erase(&realm->node, &mdsc->snap_realms);
169
170 if (realm->parent) {
171 list_del_init(&realm->child_item);
172 __put_snap_realm(mdsc, realm->parent);
173 }
174
175 kfree(realm->prior_parent_snaps);
176 kfree(realm->snaps);
177 ceph_put_snap_context(realm->cached_context);
178 kfree(realm);
179}
180
181/*
182 * caller holds snap_rwsem (write)
183 */
184static void __put_snap_realm(struct ceph_mds_client *mdsc,
185 struct ceph_snap_realm *realm)
186{
187 dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
188 atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
189 if (atomic_dec_and_test(&realm->nref))
190 __destroy_snap_realm(mdsc, realm);
191}
192
193/*
194 * caller needn't hold any locks
195 */
196void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
197 struct ceph_snap_realm *realm)
198{
199 dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
200 atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
201 if (!atomic_dec_and_test(&realm->nref))
202 return;
203
204 if (down_write_trylock(&mdsc->snap_rwsem)) {
205 __destroy_snap_realm(mdsc, realm);
206 up_write(&mdsc->snap_rwsem);
207 } else {
208 spin_lock(&mdsc->snap_empty_lock);
209 list_add(&realm->empty_item, &mdsc->snap_empty);
210 spin_unlock(&mdsc->snap_empty_lock);
211 }
212}
213
214/*
215 * Clean up any realms whose ref counts have dropped to zero. Note
216 * that this does not include realms who were created but not yet
217 * used.
218 *
219 * Called under snap_rwsem (write)
220 */
221static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
222{
223 struct ceph_snap_realm *realm;
224
225 spin_lock(&mdsc->snap_empty_lock);
226 while (!list_empty(&mdsc->snap_empty)) {
227 realm = list_first_entry(&mdsc->snap_empty,
228 struct ceph_snap_realm, empty_item);
229 list_del(&realm->empty_item);
230 spin_unlock(&mdsc->snap_empty_lock);
231 __destroy_snap_realm(mdsc, realm);
232 spin_lock(&mdsc->snap_empty_lock);
233 }
234 spin_unlock(&mdsc->snap_empty_lock);
235}
236
237void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
238{
239 down_write(&mdsc->snap_rwsem);
240 __cleanup_empty_realms(mdsc);
241 up_write(&mdsc->snap_rwsem);
242}
243
244/*
245 * adjust the parent realm of a given @realm. adjust child list, and parent
246 * pointers, and ref counts appropriately.
247 *
248 * return true if parent was changed, 0 if unchanged, <0 on error.
249 *
250 * caller must hold snap_rwsem for write.
251 */
252static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
253 struct ceph_snap_realm *realm,
254 u64 parentino)
255{
256 struct ceph_snap_realm *parent;
257
258 if (realm->parent_ino == parentino)
259 return 0;
260
261 parent = ceph_lookup_snap_realm(mdsc, parentino);
262 if (!parent) {
263 parent = ceph_create_snap_realm(mdsc, parentino);
264 if (IS_ERR(parent))
265 return PTR_ERR(parent);
266 }
267 dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
268 realm->ino, realm, realm->parent_ino, realm->parent,
269 parentino, parent);
270 if (realm->parent) {
271 list_del_init(&realm->child_item);
272 ceph_put_snap_realm(mdsc, realm->parent);
273 }
274 realm->parent_ino = parentino;
275 realm->parent = parent;
276 ceph_get_snap_realm(mdsc, parent);
277 list_add(&realm->child_item, &parent->children);
278 return 1;
279}
280
281
282static int cmpu64_rev(const void *a, const void *b)
283{
284 if (*(u64 *)a < *(u64 *)b)
285 return 1;
286 if (*(u64 *)a > *(u64 *)b)
287 return -1;
288 return 0;
289}
290
291/*
292 * build the snap context for a given realm.
293 */
294static int build_snap_context(struct ceph_snap_realm *realm)
295{
296 struct ceph_snap_realm *parent = realm->parent;
297 struct ceph_snap_context *snapc;
298 int err = 0;
299 u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
300
301 /*
302 * build parent context, if it hasn't been built.
303 * conservatively estimate that all parent snaps might be
304 * included by us.
305 */
306 if (parent) {
307 if (!parent->cached_context) {
308 err = build_snap_context(parent);
309 if (err)
310 goto fail;
311 }
312 num += parent->cached_context->num_snaps;
313 }
314
315 /* do i actually need to update? not if my context seq
316 matches realm seq, and my parents' does to. (this works
317 because we rebuild_snap_realms() works _downward_ in
318 hierarchy after each update.) */
319 if (realm->cached_context &&
320 realm->cached_context->seq == realm->seq &&
321 (!parent ||
322 realm->cached_context->seq >= parent->cached_context->seq)) {
323 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)"
324 " (unchanged)\n",
325 realm->ino, realm, realm->cached_context,
326 realm->cached_context->seq,
327 (unsigned int) realm->cached_context->num_snaps);
328 return 0;
329 }
330
331 /* alloc new snap context */
332 err = -ENOMEM;
333 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
334 goto fail;
335 snapc = ceph_create_snap_context(num, GFP_NOFS);
336 if (!snapc)
337 goto fail;
338
339 /* build (reverse sorted) snap vector */
340 num = 0;
341 snapc->seq = realm->seq;
342 if (parent) {
343 u32 i;
344
345 /* include any of parent's snaps occurring _after_ my
346 parent became my parent */
347 for (i = 0; i < parent->cached_context->num_snaps; i++)
348 if (parent->cached_context->snaps[i] >=
349 realm->parent_since)
350 snapc->snaps[num++] =
351 parent->cached_context->snaps[i];
352 if (parent->cached_context->seq > snapc->seq)
353 snapc->seq = parent->cached_context->seq;
354 }
355 memcpy(snapc->snaps + num, realm->snaps,
356 sizeof(u64)*realm->num_snaps);
357 num += realm->num_snaps;
358 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
359 sizeof(u64)*realm->num_prior_parent_snaps);
360 num += realm->num_prior_parent_snaps;
361
362 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
363 snapc->num_snaps = num;
364 dout("build_snap_context %llx %p: %p seq %lld (%u snaps)\n",
365 realm->ino, realm, snapc, snapc->seq,
366 (unsigned int) snapc->num_snaps);
367
368 if (realm->cached_context)
369 ceph_put_snap_context(realm->cached_context);
370 realm->cached_context = snapc;
371 return 0;
372
373fail:
374 /*
375 * if we fail, clear old (incorrect) cached_context... hopefully
376 * we'll have better luck building it later
377 */
378 if (realm->cached_context) {
379 ceph_put_snap_context(realm->cached_context);
380 realm->cached_context = NULL;
381 }
382 pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
383 realm, err);
384 return err;
385}
386
387/*
388 * rebuild snap context for the given realm and all of its children.
389 */
390static void rebuild_snap_realms(struct ceph_snap_realm *realm)
391{
392 struct ceph_snap_realm *child;
393
394 dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
395 build_snap_context(realm);
396
397 list_for_each_entry(child, &realm->children, child_item)
398 rebuild_snap_realms(child);
399}
400
401
402/*
403 * helper to allocate and decode an array of snapids. free prior
404 * instance, if any.
405 */
406static int dup_array(u64 **dst, __le64 *src, u32 num)
407{
408 u32 i;
409
410 kfree(*dst);
411 if (num) {
412 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
413 if (!*dst)
414 return -ENOMEM;
415 for (i = 0; i < num; i++)
416 (*dst)[i] = get_unaligned_le64(src + i);
417 } else {
418 *dst = NULL;
419 }
420 return 0;
421}
422
423
424/*
425 * When a snapshot is applied, the size/mtime inode metadata is queued
426 * in a ceph_cap_snap (one for each snapshot) until writeback
427 * completes and the metadata can be flushed back to the MDS.
428 *
429 * However, if a (sync) write is currently in-progress when we apply
430 * the snapshot, we have to wait until the write succeeds or fails
431 * (and a final size/mtime is known). In this case the
432 * cap_snap->writing = 1, and is said to be "pending." When the write
433 * finishes, we __ceph_finish_cap_snap().
434 *
435 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
436 * change).
437 */
438void ceph_queue_cap_snap(struct ceph_inode_info *ci)
439{
440 struct inode *inode = &ci->vfs_inode;
441 struct ceph_cap_snap *capsnap;
442 int used, dirty;
443
444 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
445 if (!capsnap) {
446 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
447 return;
448 }
449
450 spin_lock(&ci->i_ceph_lock);
451 used = __ceph_caps_used(ci);
452 dirty = __ceph_caps_dirty(ci);
453
454 /*
455 * If there is a write in progress, treat that as a dirty Fw,
456 * even though it hasn't completed yet; by the time we finish
457 * up this capsnap it will be.
458 */
459 if (used & CEPH_CAP_FILE_WR)
460 dirty |= CEPH_CAP_FILE_WR;
461
462 if (__ceph_have_pending_cap_snap(ci)) {
463 /* there is no point in queuing multiple "pending" cap_snaps,
464 as no new writes are allowed to start when pending, so any
465 writes in progress now were started before the previous
466 cap_snap. lucky us. */
467 dout("queue_cap_snap %p already pending\n", inode);
468 kfree(capsnap);
469 } else if (dirty & (CEPH_CAP_AUTH_EXCL|CEPH_CAP_XATTR_EXCL|
470 CEPH_CAP_FILE_EXCL|CEPH_CAP_FILE_WR)) {
471 struct ceph_snap_context *snapc = ci->i_head_snapc;
472
473 /*
474 * if we are a sync write, we may need to go to the snaprealm
475 * to get the current snapc.
476 */
477 if (!snapc)
478 snapc = ci->i_snap_realm->cached_context;
479
480 dout("queue_cap_snap %p cap_snap %p queuing under %p %s\n",
481 inode, capsnap, snapc, ceph_cap_string(dirty));
482 ihold(inode);
483
484 atomic_set(&capsnap->nref, 1);
485 capsnap->ci = ci;
486 INIT_LIST_HEAD(&capsnap->ci_item);
487 INIT_LIST_HEAD(&capsnap->flushing_item);
488
489 capsnap->follows = snapc->seq;
490 capsnap->issued = __ceph_caps_issued(ci, NULL);
491 capsnap->dirty = dirty;
492
493 capsnap->mode = inode->i_mode;
494 capsnap->uid = inode->i_uid;
495 capsnap->gid = inode->i_gid;
496
497 if (dirty & CEPH_CAP_XATTR_EXCL) {
498 __ceph_build_xattrs_blob(ci);
499 capsnap->xattr_blob =
500 ceph_buffer_get(ci->i_xattrs.blob);
501 capsnap->xattr_version = ci->i_xattrs.version;
502 } else {
503 capsnap->xattr_blob = NULL;
504 capsnap->xattr_version = 0;
505 }
506
507 /* dirty page count moved from _head to this cap_snap;
508 all subsequent writes page dirties occur _after_ this
509 snapshot. */
510 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
511 ci->i_wrbuffer_ref_head = 0;
512 capsnap->context = snapc;
513 ci->i_head_snapc =
514 ceph_get_snap_context(ci->i_snap_realm->cached_context);
515 dout(" new snapc is %p\n", ci->i_head_snapc);
516 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
517
518 if (used & CEPH_CAP_FILE_WR) {
519 dout("queue_cap_snap %p cap_snap %p snapc %p"
520 " seq %llu used WR, now pending\n", inode,
521 capsnap, snapc, snapc->seq);
522 capsnap->writing = 1;
523 } else {
524 /* note mtime, size NOW. */
525 __ceph_finish_cap_snap(ci, capsnap);
526 }
527 } else {
528 dout("queue_cap_snap %p nothing dirty|writing\n", inode);
529 kfree(capsnap);
530 }
531
532 spin_unlock(&ci->i_ceph_lock);
533}
534
535/*
536 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
537 * to be used for the snapshot, to be flushed back to the mds.
538 *
539 * If capsnap can now be flushed, add to snap_flush list, and return 1.
540 *
541 * Caller must hold i_ceph_lock.
542 */
543int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
544 struct ceph_cap_snap *capsnap)
545{
546 struct inode *inode = &ci->vfs_inode;
547 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
548
549 BUG_ON(capsnap->writing);
550 capsnap->size = inode->i_size;
551 capsnap->mtime = inode->i_mtime;
552 capsnap->atime = inode->i_atime;
553 capsnap->ctime = inode->i_ctime;
554 capsnap->time_warp_seq = ci->i_time_warp_seq;
555 if (capsnap->dirty_pages) {
556 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
557 "still has %d dirty pages\n", inode, capsnap,
558 capsnap->context, capsnap->context->seq,
559 ceph_cap_string(capsnap->dirty), capsnap->size,
560 capsnap->dirty_pages);
561 return 0;
562 }
563 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
564 inode, capsnap, capsnap->context,
565 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
566 capsnap->size);
567
568 spin_lock(&mdsc->snap_flush_lock);
569 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
570 spin_unlock(&mdsc->snap_flush_lock);
571 return 1; /* caller may want to ceph_flush_snaps */
572}
573
574/*
575 * Queue cap_snaps for snap writeback for this realm and its children.
576 * Called under snap_rwsem, so realm topology won't change.
577 */
578static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
579{
580 struct ceph_inode_info *ci;
581 struct inode *lastinode = NULL;
582 struct ceph_snap_realm *child;
583
584 dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
585
586 spin_lock(&realm->inodes_with_caps_lock);
587 list_for_each_entry(ci, &realm->inodes_with_caps,
588 i_snap_realm_item) {
589 struct inode *inode = igrab(&ci->vfs_inode);
590 if (!inode)
591 continue;
592 spin_unlock(&realm->inodes_with_caps_lock);
593 if (lastinode)
594 iput(lastinode);
595 lastinode = inode;
596 ceph_queue_cap_snap(ci);
597 spin_lock(&realm->inodes_with_caps_lock);
598 }
599 spin_unlock(&realm->inodes_with_caps_lock);
600 if (lastinode)
601 iput(lastinode);
602
603 list_for_each_entry(child, &realm->children, child_item) {
604 dout("queue_realm_cap_snaps %p %llx queue child %p %llx\n",
605 realm, realm->ino, child, child->ino);
606 list_del_init(&child->dirty_item);
607 list_add(&child->dirty_item, &realm->dirty_item);
608 }
609
610 list_del_init(&realm->dirty_item);
611 dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
612}
613
614/*
615 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
616 * the snap realm parameters from a given realm and all of its ancestors,
617 * up to the root.
618 *
619 * Caller must hold snap_rwsem for write.
620 */
621int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
622 void *p, void *e, bool deletion)
623{
624 struct ceph_mds_snap_realm *ri; /* encoded */
625 __le64 *snaps; /* encoded */
626 __le64 *prior_parent_snaps; /* encoded */
627 struct ceph_snap_realm *realm;
628 int invalidate = 0;
629 int err = -ENOMEM;
630 LIST_HEAD(dirty_realms);
631
632 dout("update_snap_trace deletion=%d\n", deletion);
633more:
634 ceph_decode_need(&p, e, sizeof(*ri), bad);
635 ri = p;
636 p += sizeof(*ri);
637 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
638 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
639 snaps = p;
640 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
641 prior_parent_snaps = p;
642 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
643
644 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
645 if (!realm) {
646 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
647 if (IS_ERR(realm)) {
648 err = PTR_ERR(realm);
649 goto fail;
650 }
651 }
652
653 /* ensure the parent is correct */
654 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
655 if (err < 0)
656 goto fail;
657 invalidate += err;
658
659 if (le64_to_cpu(ri->seq) > realm->seq) {
660 dout("update_snap_trace updating %llx %p %lld -> %lld\n",
661 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
662 /* update realm parameters, snap lists */
663 realm->seq = le64_to_cpu(ri->seq);
664 realm->created = le64_to_cpu(ri->created);
665 realm->parent_since = le64_to_cpu(ri->parent_since);
666
667 realm->num_snaps = le32_to_cpu(ri->num_snaps);
668 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
669 if (err < 0)
670 goto fail;
671
672 realm->num_prior_parent_snaps =
673 le32_to_cpu(ri->num_prior_parent_snaps);
674 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
675 realm->num_prior_parent_snaps);
676 if (err < 0)
677 goto fail;
678
679 /* queue realm for cap_snap creation */
680 list_add(&realm->dirty_item, &dirty_realms);
681
682 invalidate = 1;
683 } else if (!realm->cached_context) {
684 dout("update_snap_trace %llx %p seq %lld new\n",
685 realm->ino, realm, realm->seq);
686 invalidate = 1;
687 } else {
688 dout("update_snap_trace %llx %p seq %lld unchanged\n",
689 realm->ino, realm, realm->seq);
690 }
691
692 dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
693 realm, invalidate, p, e);
694
695 if (p < e)
696 goto more;
697
698 /* invalidate when we reach the _end_ (root) of the trace */
699 if (invalidate)
700 rebuild_snap_realms(realm);
701
702 /*
703 * queue cap snaps _after_ we've built the new snap contexts,
704 * so that i_head_snapc can be set appropriately.
705 */
706 while (!list_empty(&dirty_realms)) {
707 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
708 dirty_item);
709 queue_realm_cap_snaps(realm);
710 }
711
712 __cleanup_empty_realms(mdsc);
713 return 0;
714
715bad:
716 err = -EINVAL;
717fail:
718 pr_err("update_snap_trace error %d\n", err);
719 return err;
720}
721
722
723/*
724 * Send any cap_snaps that are queued for flush. Try to carry
725 * s_mutex across multiple snap flushes to avoid locking overhead.
726 *
727 * Caller holds no locks.
728 */
729static void flush_snaps(struct ceph_mds_client *mdsc)
730{
731 struct ceph_inode_info *ci;
732 struct inode *inode;
733 struct ceph_mds_session *session = NULL;
734
735 dout("flush_snaps\n");
736 spin_lock(&mdsc->snap_flush_lock);
737 while (!list_empty(&mdsc->snap_flush_list)) {
738 ci = list_first_entry(&mdsc->snap_flush_list,
739 struct ceph_inode_info, i_snap_flush_item);
740 inode = &ci->vfs_inode;
741 ihold(inode);
742 spin_unlock(&mdsc->snap_flush_lock);
743 spin_lock(&ci->i_ceph_lock);
744 __ceph_flush_snaps(ci, &session, 0);
745 spin_unlock(&ci->i_ceph_lock);
746 iput(inode);
747 spin_lock(&mdsc->snap_flush_lock);
748 }
749 spin_unlock(&mdsc->snap_flush_lock);
750
751 if (session) {
752 mutex_unlock(&session->s_mutex);
753 ceph_put_mds_session(session);
754 }
755 dout("flush_snaps done\n");
756}
757
758
759/*
760 * Handle a snap notification from the MDS.
761 *
762 * This can take two basic forms: the simplest is just a snap creation
763 * or deletion notification on an existing realm. This should update the
764 * realm and its children.
765 *
766 * The more difficult case is realm creation, due to snap creation at a
767 * new point in the file hierarchy, or due to a rename that moves a file or
768 * directory into another realm.
769 */
770void ceph_handle_snap(struct ceph_mds_client *mdsc,
771 struct ceph_mds_session *session,
772 struct ceph_msg *msg)
773{
774 struct super_block *sb = mdsc->fsc->sb;
775 int mds = session->s_mds;
776 u64 split;
777 int op;
778 int trace_len;
779 struct ceph_snap_realm *realm = NULL;
780 void *p = msg->front.iov_base;
781 void *e = p + msg->front.iov_len;
782 struct ceph_mds_snap_head *h;
783 int num_split_inos, num_split_realms;
784 __le64 *split_inos = NULL, *split_realms = NULL;
785 int i;
786 int locked_rwsem = 0;
787
788 /* decode */
789 if (msg->front.iov_len < sizeof(*h))
790 goto bad;
791 h = p;
792 op = le32_to_cpu(h->op);
793 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
794 * existing realm */
795 num_split_inos = le32_to_cpu(h->num_split_inos);
796 num_split_realms = le32_to_cpu(h->num_split_realms);
797 trace_len = le32_to_cpu(h->trace_len);
798 p += sizeof(*h);
799
800 dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
801 ceph_snap_op_name(op), split, trace_len);
802
803 mutex_lock(&session->s_mutex);
804 session->s_seq++;
805 mutex_unlock(&session->s_mutex);
806
807 down_write(&mdsc->snap_rwsem);
808 locked_rwsem = 1;
809
810 if (op == CEPH_SNAP_OP_SPLIT) {
811 struct ceph_mds_snap_realm *ri;
812
813 /*
814 * A "split" breaks part of an existing realm off into
815 * a new realm. The MDS provides a list of inodes
816 * (with caps) and child realms that belong to the new
817 * child.
818 */
819 split_inos = p;
820 p += sizeof(u64) * num_split_inos;
821 split_realms = p;
822 p += sizeof(u64) * num_split_realms;
823 ceph_decode_need(&p, e, sizeof(*ri), bad);
824 /* we will peek at realm info here, but will _not_
825 * advance p, as the realm update will occur below in
826 * ceph_update_snap_trace. */
827 ri = p;
828
829 realm = ceph_lookup_snap_realm(mdsc, split);
830 if (!realm) {
831 realm = ceph_create_snap_realm(mdsc, split);
832 if (IS_ERR(realm))
833 goto out;
834 }
835 ceph_get_snap_realm(mdsc, realm);
836
837 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
838 for (i = 0; i < num_split_inos; i++) {
839 struct ceph_vino vino = {
840 .ino = le64_to_cpu(split_inos[i]),
841 .snap = CEPH_NOSNAP,
842 };
843 struct inode *inode = ceph_find_inode(sb, vino);
844 struct ceph_inode_info *ci;
845 struct ceph_snap_realm *oldrealm;
846
847 if (!inode)
848 continue;
849 ci = ceph_inode(inode);
850
851 spin_lock(&ci->i_ceph_lock);
852 if (!ci->i_snap_realm)
853 goto skip_inode;
854 /*
855 * If this inode belongs to a realm that was
856 * created after our new realm, we experienced
857 * a race (due to another split notifications
858 * arriving from a different MDS). So skip
859 * this inode.
860 */
861 if (ci->i_snap_realm->created >
862 le64_to_cpu(ri->created)) {
863 dout(" leaving %p in newer realm %llx %p\n",
864 inode, ci->i_snap_realm->ino,
865 ci->i_snap_realm);
866 goto skip_inode;
867 }
868 dout(" will move %p to split realm %llx %p\n",
869 inode, realm->ino, realm);
870 /*
871 * Move the inode to the new realm
872 */
873 spin_lock(&realm->inodes_with_caps_lock);
874 list_del_init(&ci->i_snap_realm_item);
875 list_add(&ci->i_snap_realm_item,
876 &realm->inodes_with_caps);
877 oldrealm = ci->i_snap_realm;
878 ci->i_snap_realm = realm;
879 spin_unlock(&realm->inodes_with_caps_lock);
880 spin_unlock(&ci->i_ceph_lock);
881
882 ceph_get_snap_realm(mdsc, realm);
883 ceph_put_snap_realm(mdsc, oldrealm);
884
885 iput(inode);
886 continue;
887
888skip_inode:
889 spin_unlock(&ci->i_ceph_lock);
890 iput(inode);
891 }
892
893 /* we may have taken some of the old realm's children. */
894 for (i = 0; i < num_split_realms; i++) {
895 struct ceph_snap_realm *child =
896 ceph_lookup_snap_realm(mdsc,
897 le64_to_cpu(split_realms[i]));
898 if (!child)
899 continue;
900 adjust_snap_realm_parent(mdsc, child, realm->ino);
901 }
902 }
903
904 /*
905 * update using the provided snap trace. if we are deleting a
906 * snap, we can avoid queueing cap_snaps.
907 */
908 ceph_update_snap_trace(mdsc, p, e,
909 op == CEPH_SNAP_OP_DESTROY);
910
911 if (op == CEPH_SNAP_OP_SPLIT)
912 /* we took a reference when we created the realm, above */
913 ceph_put_snap_realm(mdsc, realm);
914
915 __cleanup_empty_realms(mdsc);
916
917 up_write(&mdsc->snap_rwsem);
918
919 flush_snaps(mdsc);
920 return;
921
922bad:
923 pr_err("corrupt snap message from mds%d\n", mds);
924 ceph_msg_dump(msg);
925out:
926 if (locked_rwsem)
927 up_write(&mdsc->snap_rwsem);
928 return;
929}
930
931
932
1// SPDX-License-Identifier: GPL-2.0
2#include <linux/ceph/ceph_debug.h>
3
4#include <linux/fs.h>
5#include <linux/sort.h>
6#include <linux/slab.h>
7#include <linux/iversion.h>
8#include "super.h"
9#include "mds_client.h"
10#include <linux/ceph/decode.h>
11
12/* unused map expires after 5 minutes */
13#define CEPH_SNAPID_MAP_TIMEOUT (5 * 60 * HZ)
14
15/*
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
21 * snapshot.
22 *
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.
27 *
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.
37 *
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.
42 *
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.
48 *
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.
52 */
53/*
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
57 * console).
58 */
59
60
61/*
62 * increase ref count for the realm
63 *
64 * caller must hold snap_rwsem.
65 */
66void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
67 struct ceph_snap_realm *realm)
68{
69 lockdep_assert_held(&mdsc->snap_rwsem);
70
71 /*
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.
76 */
77 if (atomic_inc_not_zero(&realm->nref))
78 return;
79
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);
84}
85
86static void __insert_snap_realm(struct rb_root *root,
87 struct ceph_snap_realm *new)
88{
89 struct rb_node **p = &root->rb_node;
90 struct rb_node *parent = NULL;
91 struct ceph_snap_realm *r = NULL;
92
93 while (*p) {
94 parent = *p;
95 r = rb_entry(parent, struct ceph_snap_realm, node);
96 if (new->ino < r->ino)
97 p = &(*p)->rb_left;
98 else if (new->ino > r->ino)
99 p = &(*p)->rb_right;
100 else
101 BUG();
102 }
103
104 rb_link_node(&new->node, parent, p);
105 rb_insert_color(&new->node, root);
106}
107
108/*
109 * create and get the realm rooted at @ino and bump its ref count.
110 *
111 * caller must hold snap_rwsem for write.
112 */
113static struct ceph_snap_realm *ceph_create_snap_realm(
114 struct ceph_mds_client *mdsc,
115 u64 ino)
116{
117 struct ceph_snap_realm *realm;
118
119 lockdep_assert_held_write(&mdsc->snap_rwsem);
120
121 realm = kzalloc(sizeof(*realm), GFP_NOFS);
122 if (!realm)
123 return ERR_PTR(-ENOMEM);
124
125 /* Do not release the global dummy snaprealm until unmouting */
126 if (ino == CEPH_INO_GLOBAL_SNAPREALM)
127 atomic_set(&realm->nref, 2);
128 else
129 atomic_set(&realm->nref, 1);
130 realm->ino = ino;
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++;
140
141 dout("%s %llx %p\n", __func__, realm->ino, realm);
142 return realm;
143}
144
145/*
146 * lookup the realm rooted at @ino.
147 *
148 * caller must hold snap_rwsem.
149 */
150static struct ceph_snap_realm *__lookup_snap_realm(struct ceph_mds_client *mdsc,
151 u64 ino)
152{
153 struct rb_node *n = mdsc->snap_realms.rb_node;
154 struct ceph_snap_realm *r;
155
156 lockdep_assert_held(&mdsc->snap_rwsem);
157
158 while (n) {
159 r = rb_entry(n, struct ceph_snap_realm, node);
160 if (ino < r->ino)
161 n = n->rb_left;
162 else if (ino > r->ino)
163 n = n->rb_right;
164 else {
165 dout("%s %llx %p\n", __func__, r->ino, r);
166 return r;
167 }
168 }
169 return NULL;
170}
171
172struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
173 u64 ino)
174{
175 struct ceph_snap_realm *r;
176 r = __lookup_snap_realm(mdsc, ino);
177 if (r)
178 ceph_get_snap_realm(mdsc, r);
179 return r;
180}
181
182static void __put_snap_realm(struct ceph_mds_client *mdsc,
183 struct ceph_snap_realm *realm);
184
185/*
186 * called with snap_rwsem (write)
187 */
188static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
189 struct ceph_snap_realm *realm)
190{
191 lockdep_assert_held_write(&mdsc->snap_rwsem);
192
193 dout("%s %p %llx\n", __func__, realm, realm->ino);
194
195 rb_erase(&realm->node, &mdsc->snap_realms);
196 mdsc->num_snap_realms--;
197
198 if (realm->parent) {
199 list_del_init(&realm->child_item);
200 __put_snap_realm(mdsc, realm->parent);
201 }
202
203 kfree(realm->prior_parent_snaps);
204 kfree(realm->snaps);
205 ceph_put_snap_context(realm->cached_context);
206 kfree(realm);
207}
208
209/*
210 * caller holds snap_rwsem (write)
211 */
212static void __put_snap_realm(struct ceph_mds_client *mdsc,
213 struct ceph_snap_realm *realm)
214{
215 lockdep_assert_held_write(&mdsc->snap_rwsem);
216
217 /*
218 * We do not require the snap_empty_lock here, as any caller that
219 * increments the value must hold the snap_rwsem.
220 */
221 if (atomic_dec_and_test(&realm->nref))
222 __destroy_snap_realm(mdsc, realm);
223}
224
225/*
226 * See comments in ceph_get_snap_realm. Caller needn't hold any locks.
227 */
228void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
229 struct ceph_snap_realm *realm)
230{
231 if (!atomic_dec_and_lock(&realm->nref, &mdsc->snap_empty_lock))
232 return;
233
234 if (down_write_trylock(&mdsc->snap_rwsem)) {
235 spin_unlock(&mdsc->snap_empty_lock);
236 __destroy_snap_realm(mdsc, realm);
237 up_write(&mdsc->snap_rwsem);
238 } else {
239 list_add(&realm->empty_item, &mdsc->snap_empty);
240 spin_unlock(&mdsc->snap_empty_lock);
241 }
242}
243
244/*
245 * Clean up any realms whose ref counts have dropped to zero. Note
246 * that this does not include realms who were created but not yet
247 * used.
248 *
249 * Called under snap_rwsem (write)
250 */
251static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
252{
253 struct ceph_snap_realm *realm;
254
255 lockdep_assert_held_write(&mdsc->snap_rwsem);
256
257 spin_lock(&mdsc->snap_empty_lock);
258 while (!list_empty(&mdsc->snap_empty)) {
259 realm = list_first_entry(&mdsc->snap_empty,
260 struct ceph_snap_realm, empty_item);
261 list_del(&realm->empty_item);
262 spin_unlock(&mdsc->snap_empty_lock);
263 __destroy_snap_realm(mdsc, realm);
264 spin_lock(&mdsc->snap_empty_lock);
265 }
266 spin_unlock(&mdsc->snap_empty_lock);
267}
268
269void ceph_cleanup_global_and_empty_realms(struct ceph_mds_client *mdsc)
270{
271 struct ceph_snap_realm *global_realm;
272
273 down_write(&mdsc->snap_rwsem);
274 global_realm = __lookup_snap_realm(mdsc, CEPH_INO_GLOBAL_SNAPREALM);
275 if (global_realm)
276 ceph_put_snap_realm(mdsc, global_realm);
277 __cleanup_empty_realms(mdsc);
278 up_write(&mdsc->snap_rwsem);
279}
280
281/*
282 * adjust the parent realm of a given @realm. adjust child list, and parent
283 * pointers, and ref counts appropriately.
284 *
285 * return true if parent was changed, 0 if unchanged, <0 on error.
286 *
287 * caller must hold snap_rwsem for write.
288 */
289static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
290 struct ceph_snap_realm *realm,
291 u64 parentino)
292{
293 struct ceph_snap_realm *parent;
294
295 lockdep_assert_held_write(&mdsc->snap_rwsem);
296
297 if (realm->parent_ino == parentino)
298 return 0;
299
300 parent = ceph_lookup_snap_realm(mdsc, parentino);
301 if (!parent) {
302 parent = ceph_create_snap_realm(mdsc, parentino);
303 if (IS_ERR(parent))
304 return PTR_ERR(parent);
305 }
306 dout("%s %llx %p: %llx %p -> %llx %p\n", __func__, realm->ino,
307 realm, realm->parent_ino, realm->parent, parentino, parent);
308 if (realm->parent) {
309 list_del_init(&realm->child_item);
310 ceph_put_snap_realm(mdsc, realm->parent);
311 }
312 realm->parent_ino = parentino;
313 realm->parent = parent;
314 list_add(&realm->child_item, &parent->children);
315 return 1;
316}
317
318
319static int cmpu64_rev(const void *a, const void *b)
320{
321 if (*(u64 *)a < *(u64 *)b)
322 return 1;
323 if (*(u64 *)a > *(u64 *)b)
324 return -1;
325 return 0;
326}
327
328
329/*
330 * build the snap context for a given realm.
331 */
332static int build_snap_context(struct ceph_snap_realm *realm,
333 struct list_head *realm_queue,
334 struct list_head *dirty_realms)
335{
336 struct ceph_snap_realm *parent = realm->parent;
337 struct ceph_snap_context *snapc;
338 int err = 0;
339 u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
340
341 /*
342 * build parent context, if it hasn't been built.
343 * conservatively estimate that all parent snaps might be
344 * included by us.
345 */
346 if (parent) {
347 if (!parent->cached_context) {
348 /* add to the queue head */
349 list_add(&parent->rebuild_item, realm_queue);
350 return 1;
351 }
352 num += parent->cached_context->num_snaps;
353 }
354
355 /* do i actually need to update? not if my context seq
356 matches realm seq, and my parents' does to. (this works
357 because we rebuild_snap_realms() works _downward_ in
358 hierarchy after each update.) */
359 if (realm->cached_context &&
360 realm->cached_context->seq == realm->seq &&
361 (!parent ||
362 realm->cached_context->seq >= parent->cached_context->seq)) {
363 dout("%s %llx %p: %p seq %lld (%u snaps) (unchanged)\n",
364 __func__, realm->ino, realm, realm->cached_context,
365 realm->cached_context->seq,
366 (unsigned int)realm->cached_context->num_snaps);
367 return 0;
368 }
369
370 /* alloc new snap context */
371 err = -ENOMEM;
372 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
373 goto fail;
374 snapc = ceph_create_snap_context(num, GFP_NOFS);
375 if (!snapc)
376 goto fail;
377
378 /* build (reverse sorted) snap vector */
379 num = 0;
380 snapc->seq = realm->seq;
381 if (parent) {
382 u32 i;
383
384 /* include any of parent's snaps occurring _after_ my
385 parent became my parent */
386 for (i = 0; i < parent->cached_context->num_snaps; i++)
387 if (parent->cached_context->snaps[i] >=
388 realm->parent_since)
389 snapc->snaps[num++] =
390 parent->cached_context->snaps[i];
391 if (parent->cached_context->seq > snapc->seq)
392 snapc->seq = parent->cached_context->seq;
393 }
394 memcpy(snapc->snaps + num, realm->snaps,
395 sizeof(u64)*realm->num_snaps);
396 num += realm->num_snaps;
397 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
398 sizeof(u64)*realm->num_prior_parent_snaps);
399 num += realm->num_prior_parent_snaps;
400
401 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
402 snapc->num_snaps = num;
403 dout("%s %llx %p: %p seq %lld (%u snaps)\n", __func__, realm->ino,
404 realm, snapc, snapc->seq, (unsigned int) snapc->num_snaps);
405
406 ceph_put_snap_context(realm->cached_context);
407 realm->cached_context = snapc;
408 /* queue realm for cap_snap creation */
409 list_add_tail(&realm->dirty_item, dirty_realms);
410 return 0;
411
412fail:
413 /*
414 * if we fail, clear old (incorrect) cached_context... hopefully
415 * we'll have better luck building it later
416 */
417 if (realm->cached_context) {
418 ceph_put_snap_context(realm->cached_context);
419 realm->cached_context = NULL;
420 }
421 pr_err("%s %llx %p fail %d\n", __func__, realm->ino, realm, err);
422 return err;
423}
424
425/*
426 * rebuild snap context for the given realm and all of its children.
427 */
428static void rebuild_snap_realms(struct ceph_snap_realm *realm,
429 struct list_head *dirty_realms)
430{
431 LIST_HEAD(realm_queue);
432 int last = 0;
433 bool skip = false;
434
435 list_add_tail(&realm->rebuild_item, &realm_queue);
436
437 while (!list_empty(&realm_queue)) {
438 struct ceph_snap_realm *_realm, *child;
439
440 _realm = list_first_entry(&realm_queue,
441 struct ceph_snap_realm,
442 rebuild_item);
443
444 /*
445 * If the last building failed dues to memory
446 * issue, just empty the realm_queue and return
447 * to avoid infinite loop.
448 */
449 if (last < 0) {
450 list_del_init(&_realm->rebuild_item);
451 continue;
452 }
453
454 last = build_snap_context(_realm, &realm_queue, dirty_realms);
455 dout("%s %llx %p, %s\n", __func__, _realm->ino, _realm,
456 last > 0 ? "is deferred" : !last ? "succeeded" : "failed");
457
458 /* is any child in the list ? */
459 list_for_each_entry(child, &_realm->children, child_item) {
460 if (!list_empty(&child->rebuild_item)) {
461 skip = true;
462 break;
463 }
464 }
465
466 if (!skip) {
467 list_for_each_entry(child, &_realm->children, child_item)
468 list_add_tail(&child->rebuild_item, &realm_queue);
469 }
470
471 /* last == 1 means need to build parent first */
472 if (last <= 0)
473 list_del_init(&_realm->rebuild_item);
474 }
475}
476
477
478/*
479 * helper to allocate and decode an array of snapids. free prior
480 * instance, if any.
481 */
482static int dup_array(u64 **dst, __le64 *src, u32 num)
483{
484 u32 i;
485
486 kfree(*dst);
487 if (num) {
488 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
489 if (!*dst)
490 return -ENOMEM;
491 for (i = 0; i < num; i++)
492 (*dst)[i] = get_unaligned_le64(src + i);
493 } else {
494 *dst = NULL;
495 }
496 return 0;
497}
498
499static bool has_new_snaps(struct ceph_snap_context *o,
500 struct ceph_snap_context *n)
501{
502 if (n->num_snaps == 0)
503 return false;
504 /* snaps are in descending order */
505 return n->snaps[0] > o->seq;
506}
507
508/*
509 * When a snapshot is applied, the size/mtime inode metadata is queued
510 * in a ceph_cap_snap (one for each snapshot) until writeback
511 * completes and the metadata can be flushed back to the MDS.
512 *
513 * However, if a (sync) write is currently in-progress when we apply
514 * the snapshot, we have to wait until the write succeeds or fails
515 * (and a final size/mtime is known). In this case the
516 * cap_snap->writing = 1, and is said to be "pending." When the write
517 * finishes, we __ceph_finish_cap_snap().
518 *
519 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
520 * change).
521 */
522static void ceph_queue_cap_snap(struct ceph_inode_info *ci,
523 struct ceph_cap_snap **pcapsnap)
524{
525 struct inode *inode = &ci->netfs.inode;
526 struct ceph_snap_context *old_snapc, *new_snapc;
527 struct ceph_cap_snap *capsnap = *pcapsnap;
528 struct ceph_buffer *old_blob = NULL;
529 int used, dirty;
530
531 spin_lock(&ci->i_ceph_lock);
532 used = __ceph_caps_used(ci);
533 dirty = __ceph_caps_dirty(ci);
534
535 old_snapc = ci->i_head_snapc;
536 new_snapc = ci->i_snap_realm->cached_context;
537
538 /*
539 * If there is a write in progress, treat that as a dirty Fw,
540 * even though it hasn't completed yet; by the time we finish
541 * up this capsnap it will be.
542 */
543 if (used & CEPH_CAP_FILE_WR)
544 dirty |= CEPH_CAP_FILE_WR;
545
546 if (__ceph_have_pending_cap_snap(ci)) {
547 /* there is no point in queuing multiple "pending" cap_snaps,
548 as no new writes are allowed to start when pending, so any
549 writes in progress now were started before the previous
550 cap_snap. lucky us. */
551 dout("%s %p %llx.%llx already pending\n",
552 __func__, inode, ceph_vinop(inode));
553 goto update_snapc;
554 }
555 if (ci->i_wrbuffer_ref_head == 0 &&
556 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
557 dout("%s %p %llx.%llx nothing dirty|writing\n",
558 __func__, inode, ceph_vinop(inode));
559 goto update_snapc;
560 }
561
562 BUG_ON(!old_snapc);
563
564 /*
565 * There is no need to send FLUSHSNAP message to MDS if there is
566 * no new snapshot. But when there is dirty pages or on-going
567 * writes, we still need to create cap_snap. cap_snap is needed
568 * by the write path and page writeback path.
569 *
570 * also see ceph_try_drop_cap_snap()
571 */
572 if (has_new_snaps(old_snapc, new_snapc)) {
573 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
574 capsnap->need_flush = true;
575 } else {
576 if (!(used & CEPH_CAP_FILE_WR) &&
577 ci->i_wrbuffer_ref_head == 0) {
578 dout("%s %p %llx.%llx no new_snap|dirty_page|writing\n",
579 __func__, inode, ceph_vinop(inode));
580 goto update_snapc;
581 }
582 }
583
584 dout("%s %p %llx.%llx cap_snap %p queuing under %p %s %s\n",
585 __func__, inode, ceph_vinop(inode), capsnap, old_snapc,
586 ceph_cap_string(dirty), capsnap->need_flush ? "" : "no_flush");
587 ihold(inode);
588
589 capsnap->follows = old_snapc->seq;
590 capsnap->issued = __ceph_caps_issued(ci, NULL);
591 capsnap->dirty = dirty;
592
593 capsnap->mode = inode->i_mode;
594 capsnap->uid = inode->i_uid;
595 capsnap->gid = inode->i_gid;
596
597 if (dirty & CEPH_CAP_XATTR_EXCL) {
598 old_blob = __ceph_build_xattrs_blob(ci);
599 capsnap->xattr_blob =
600 ceph_buffer_get(ci->i_xattrs.blob);
601 capsnap->xattr_version = ci->i_xattrs.version;
602 } else {
603 capsnap->xattr_blob = NULL;
604 capsnap->xattr_version = 0;
605 }
606
607 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
608
609 /* dirty page count moved from _head to this cap_snap;
610 all subsequent writes page dirties occur _after_ this
611 snapshot. */
612 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
613 ci->i_wrbuffer_ref_head = 0;
614 capsnap->context = old_snapc;
615 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
616
617 if (used & CEPH_CAP_FILE_WR) {
618 dout("%s %p %llx.%llx cap_snap %p snapc %p seq %llu used WR,"
619 " now pending\n", __func__, inode, ceph_vinop(inode),
620 capsnap, old_snapc, old_snapc->seq);
621 capsnap->writing = 1;
622 } else {
623 /* note mtime, size NOW. */
624 __ceph_finish_cap_snap(ci, capsnap);
625 }
626 *pcapsnap = NULL;
627 old_snapc = NULL;
628
629update_snapc:
630 if (ci->i_wrbuffer_ref_head == 0 &&
631 ci->i_wr_ref == 0 &&
632 ci->i_dirty_caps == 0 &&
633 ci->i_flushing_caps == 0) {
634 ci->i_head_snapc = NULL;
635 } else {
636 ci->i_head_snapc = ceph_get_snap_context(new_snapc);
637 dout(" new snapc is %p\n", new_snapc);
638 }
639 spin_unlock(&ci->i_ceph_lock);
640
641 ceph_buffer_put(old_blob);
642 ceph_put_snap_context(old_snapc);
643}
644
645/*
646 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
647 * to be used for the snapshot, to be flushed back to the mds.
648 *
649 * If capsnap can now be flushed, add to snap_flush list, and return 1.
650 *
651 * Caller must hold i_ceph_lock.
652 */
653int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
654 struct ceph_cap_snap *capsnap)
655{
656 struct inode *inode = &ci->netfs.inode;
657 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
658
659 BUG_ON(capsnap->writing);
660 capsnap->size = i_size_read(inode);
661 capsnap->mtime = inode->i_mtime;
662 capsnap->atime = inode->i_atime;
663 capsnap->ctime = inode->i_ctime;
664 capsnap->btime = ci->i_btime;
665 capsnap->change_attr = inode_peek_iversion_raw(inode);
666 capsnap->time_warp_seq = ci->i_time_warp_seq;
667 capsnap->truncate_size = ci->i_truncate_size;
668 capsnap->truncate_seq = ci->i_truncate_seq;
669 if (capsnap->dirty_pages) {
670 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu "
671 "still has %d dirty pages\n", __func__, inode,
672 ceph_vinop(inode), capsnap, capsnap->context,
673 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
674 capsnap->size, capsnap->dirty_pages);
675 return 0;
676 }
677
678 /* Fb cap still in use, delay it */
679 if (ci->i_wb_ref) {
680 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu "
681 "used WRBUFFER, delaying\n", __func__, inode,
682 ceph_vinop(inode), capsnap, capsnap->context,
683 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
684 capsnap->size);
685 capsnap->writing = 1;
686 return 0;
687 }
688
689 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
690 dout("%s %p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu\n",
691 __func__, inode, ceph_vinop(inode), capsnap, capsnap->context,
692 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
693 capsnap->size);
694
695 spin_lock(&mdsc->snap_flush_lock);
696 if (list_empty(&ci->i_snap_flush_item))
697 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
698 spin_unlock(&mdsc->snap_flush_lock);
699 return 1; /* caller may want to ceph_flush_snaps */
700}
701
702/*
703 * Queue cap_snaps for snap writeback for this realm and its children.
704 * Called under snap_rwsem, so realm topology won't change.
705 */
706static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
707{
708 struct ceph_inode_info *ci;
709 struct inode *lastinode = NULL;
710 struct ceph_cap_snap *capsnap = NULL;
711
712 dout("%s %p %llx inode\n", __func__, realm, realm->ino);
713
714 spin_lock(&realm->inodes_with_caps_lock);
715 list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
716 struct inode *inode = igrab(&ci->netfs.inode);
717 if (!inode)
718 continue;
719 spin_unlock(&realm->inodes_with_caps_lock);
720 iput(lastinode);
721 lastinode = inode;
722
723 /*
724 * Allocate the capsnap memory outside of ceph_queue_cap_snap()
725 * to reduce very possible but unnecessary frequently memory
726 * allocate/free in this loop.
727 */
728 if (!capsnap) {
729 capsnap = kmem_cache_zalloc(ceph_cap_snap_cachep, GFP_NOFS);
730 if (!capsnap) {
731 pr_err("ENOMEM allocating ceph_cap_snap on %p\n",
732 inode);
733 return;
734 }
735 }
736 capsnap->cap_flush.is_capsnap = true;
737 refcount_set(&capsnap->nref, 1);
738 INIT_LIST_HEAD(&capsnap->cap_flush.i_list);
739 INIT_LIST_HEAD(&capsnap->cap_flush.g_list);
740 INIT_LIST_HEAD(&capsnap->ci_item);
741
742 ceph_queue_cap_snap(ci, &capsnap);
743 spin_lock(&realm->inodes_with_caps_lock);
744 }
745 spin_unlock(&realm->inodes_with_caps_lock);
746 iput(lastinode);
747
748 if (capsnap)
749 kmem_cache_free(ceph_cap_snap_cachep, capsnap);
750 dout("%s %p %llx done\n", __func__, realm, realm->ino);
751}
752
753/*
754 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
755 * the snap realm parameters from a given realm and all of its ancestors,
756 * up to the root.
757 *
758 * Caller must hold snap_rwsem for write.
759 */
760int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
761 void *p, void *e, bool deletion,
762 struct ceph_snap_realm **realm_ret)
763{
764 struct ceph_mds_snap_realm *ri; /* encoded */
765 __le64 *snaps; /* encoded */
766 __le64 *prior_parent_snaps; /* encoded */
767 struct ceph_snap_realm *realm;
768 struct ceph_snap_realm *first_realm = NULL;
769 struct ceph_snap_realm *realm_to_rebuild = NULL;
770 struct ceph_client *client = mdsc->fsc->client;
771 int rebuild_snapcs;
772 int err = -ENOMEM;
773 int ret;
774 LIST_HEAD(dirty_realms);
775
776 lockdep_assert_held_write(&mdsc->snap_rwsem);
777
778 dout("%s deletion=%d\n", __func__, deletion);
779more:
780 realm = NULL;
781 rebuild_snapcs = 0;
782 ceph_decode_need(&p, e, sizeof(*ri), bad);
783 ri = p;
784 p += sizeof(*ri);
785 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
786 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
787 snaps = p;
788 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
789 prior_parent_snaps = p;
790 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
791
792 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
793 if (!realm) {
794 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
795 if (IS_ERR(realm)) {
796 err = PTR_ERR(realm);
797 goto fail;
798 }
799 }
800
801 /* ensure the parent is correct */
802 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
803 if (err < 0)
804 goto fail;
805 rebuild_snapcs += err;
806
807 if (le64_to_cpu(ri->seq) > realm->seq) {
808 dout("%s updating %llx %p %lld -> %lld\n", __func__,
809 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
810 /* update realm parameters, snap lists */
811 realm->seq = le64_to_cpu(ri->seq);
812 realm->created = le64_to_cpu(ri->created);
813 realm->parent_since = le64_to_cpu(ri->parent_since);
814
815 realm->num_snaps = le32_to_cpu(ri->num_snaps);
816 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
817 if (err < 0)
818 goto fail;
819
820 realm->num_prior_parent_snaps =
821 le32_to_cpu(ri->num_prior_parent_snaps);
822 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
823 realm->num_prior_parent_snaps);
824 if (err < 0)
825 goto fail;
826
827 if (realm->seq > mdsc->last_snap_seq)
828 mdsc->last_snap_seq = realm->seq;
829
830 rebuild_snapcs = 1;
831 } else if (!realm->cached_context) {
832 dout("%s %llx %p seq %lld new\n", __func__,
833 realm->ino, realm, realm->seq);
834 rebuild_snapcs = 1;
835 } else {
836 dout("%s %llx %p seq %lld unchanged\n", __func__,
837 realm->ino, realm, realm->seq);
838 }
839
840 dout("done with %llx %p, rebuild_snapcs=%d, %p %p\n", realm->ino,
841 realm, rebuild_snapcs, p, e);
842
843 /*
844 * this will always track the uppest parent realm from which
845 * we need to rebuild the snapshot contexts _downward_ in
846 * hierarchy.
847 */
848 if (rebuild_snapcs)
849 realm_to_rebuild = realm;
850
851 /* rebuild_snapcs when we reach the _end_ (root) of the trace */
852 if (realm_to_rebuild && p >= e)
853 rebuild_snap_realms(realm_to_rebuild, &dirty_realms);
854
855 if (!first_realm)
856 first_realm = realm;
857 else
858 ceph_put_snap_realm(mdsc, realm);
859
860 if (p < e)
861 goto more;
862
863 /*
864 * queue cap snaps _after_ we've built the new snap contexts,
865 * so that i_head_snapc can be set appropriately.
866 */
867 while (!list_empty(&dirty_realms)) {
868 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
869 dirty_item);
870 list_del_init(&realm->dirty_item);
871 queue_realm_cap_snaps(realm);
872 }
873
874 if (realm_ret)
875 *realm_ret = first_realm;
876 else
877 ceph_put_snap_realm(mdsc, first_realm);
878
879 __cleanup_empty_realms(mdsc);
880 return 0;
881
882bad:
883 err = -EIO;
884fail:
885 if (realm && !IS_ERR(realm))
886 ceph_put_snap_realm(mdsc, realm);
887 if (first_realm)
888 ceph_put_snap_realm(mdsc, first_realm);
889 pr_err("%s error %d\n", __func__, err);
890
891 /*
892 * When receiving a corrupted snap trace we don't know what
893 * exactly has happened in MDS side. And we shouldn't continue
894 * writing to OSD, which may corrupt the snapshot contents.
895 *
896 * Just try to blocklist this kclient and then this kclient
897 * must be remounted to continue after the corrupted metadata
898 * fixed in the MDS side.
899 */
900 WRITE_ONCE(mdsc->fsc->mount_state, CEPH_MOUNT_FENCE_IO);
901 ret = ceph_monc_blocklist_add(&client->monc, &client->msgr.inst.addr);
902 if (ret)
903 pr_err("%s failed to blocklist %s: %d\n", __func__,
904 ceph_pr_addr(&client->msgr.inst.addr), ret);
905
906 WARN(1, "%s: %s%sdo remount to continue%s",
907 __func__, ret ? "" : ceph_pr_addr(&client->msgr.inst.addr),
908 ret ? "" : " was blocklisted, ",
909 err == -EIO ? " after corrupted snaptrace is fixed" : "");
910
911 return err;
912}
913
914
915/*
916 * Send any cap_snaps that are queued for flush. Try to carry
917 * s_mutex across multiple snap flushes to avoid locking overhead.
918 *
919 * Caller holds no locks.
920 */
921static void flush_snaps(struct ceph_mds_client *mdsc)
922{
923 struct ceph_inode_info *ci;
924 struct inode *inode;
925 struct ceph_mds_session *session = NULL;
926
927 dout("%s\n", __func__);
928 spin_lock(&mdsc->snap_flush_lock);
929 while (!list_empty(&mdsc->snap_flush_list)) {
930 ci = list_first_entry(&mdsc->snap_flush_list,
931 struct ceph_inode_info, i_snap_flush_item);
932 inode = &ci->netfs.inode;
933 ihold(inode);
934 spin_unlock(&mdsc->snap_flush_lock);
935 ceph_flush_snaps(ci, &session);
936 iput(inode);
937 spin_lock(&mdsc->snap_flush_lock);
938 }
939 spin_unlock(&mdsc->snap_flush_lock);
940
941 ceph_put_mds_session(session);
942 dout("%s done\n", __func__);
943}
944
945/**
946 * ceph_change_snap_realm - change the snap_realm for an inode
947 * @inode: inode to move to new snap realm
948 * @realm: new realm to move inode into (may be NULL)
949 *
950 * Detach an inode from its old snaprealm (if any) and attach it to
951 * the new snaprealm (if any). The old snap realm reference held by
952 * the inode is put. If realm is non-NULL, then the caller's reference
953 * to it is taken over by the inode.
954 */
955void ceph_change_snap_realm(struct inode *inode, struct ceph_snap_realm *realm)
956{
957 struct ceph_inode_info *ci = ceph_inode(inode);
958 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
959 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
960
961 lockdep_assert_held(&ci->i_ceph_lock);
962
963 if (oldrealm) {
964 spin_lock(&oldrealm->inodes_with_caps_lock);
965 list_del_init(&ci->i_snap_realm_item);
966 if (oldrealm->ino == ci->i_vino.ino)
967 oldrealm->inode = NULL;
968 spin_unlock(&oldrealm->inodes_with_caps_lock);
969 ceph_put_snap_realm(mdsc, oldrealm);
970 }
971
972 ci->i_snap_realm = realm;
973
974 if (realm) {
975 spin_lock(&realm->inodes_with_caps_lock);
976 list_add(&ci->i_snap_realm_item, &realm->inodes_with_caps);
977 if (realm->ino == ci->i_vino.ino)
978 realm->inode = inode;
979 spin_unlock(&realm->inodes_with_caps_lock);
980 }
981}
982
983/*
984 * Handle a snap notification from the MDS.
985 *
986 * This can take two basic forms: the simplest is just a snap creation
987 * or deletion notification on an existing realm. This should update the
988 * realm and its children.
989 *
990 * The more difficult case is realm creation, due to snap creation at a
991 * new point in the file hierarchy, or due to a rename that moves a file or
992 * directory into another realm.
993 */
994void ceph_handle_snap(struct ceph_mds_client *mdsc,
995 struct ceph_mds_session *session,
996 struct ceph_msg *msg)
997{
998 struct super_block *sb = mdsc->fsc->sb;
999 int mds = session->s_mds;
1000 u64 split;
1001 int op;
1002 int trace_len;
1003 struct ceph_snap_realm *realm = NULL;
1004 void *p = msg->front.iov_base;
1005 void *e = p + msg->front.iov_len;
1006 struct ceph_mds_snap_head *h;
1007 int num_split_inos, num_split_realms;
1008 __le64 *split_inos = NULL, *split_realms = NULL;
1009 int i;
1010 int locked_rwsem = 0;
1011 bool close_sessions = false;
1012
1013 /* decode */
1014 if (msg->front.iov_len < sizeof(*h))
1015 goto bad;
1016 h = p;
1017 op = le32_to_cpu(h->op);
1018 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
1019 * existing realm */
1020 num_split_inos = le32_to_cpu(h->num_split_inos);
1021 num_split_realms = le32_to_cpu(h->num_split_realms);
1022 trace_len = le32_to_cpu(h->trace_len);
1023 p += sizeof(*h);
1024
1025 dout("%s from mds%d op %s split %llx tracelen %d\n", __func__,
1026 mds, ceph_snap_op_name(op), split, trace_len);
1027
1028 mutex_lock(&session->s_mutex);
1029 inc_session_sequence(session);
1030 mutex_unlock(&session->s_mutex);
1031
1032 down_write(&mdsc->snap_rwsem);
1033 locked_rwsem = 1;
1034
1035 if (op == CEPH_SNAP_OP_SPLIT) {
1036 struct ceph_mds_snap_realm *ri;
1037
1038 /*
1039 * A "split" breaks part of an existing realm off into
1040 * a new realm. The MDS provides a list of inodes
1041 * (with caps) and child realms that belong to the new
1042 * child.
1043 */
1044 split_inos = p;
1045 p += sizeof(u64) * num_split_inos;
1046 split_realms = p;
1047 p += sizeof(u64) * num_split_realms;
1048 ceph_decode_need(&p, e, sizeof(*ri), bad);
1049 /* we will peek at realm info here, but will _not_
1050 * advance p, as the realm update will occur below in
1051 * ceph_update_snap_trace. */
1052 ri = p;
1053
1054 realm = ceph_lookup_snap_realm(mdsc, split);
1055 if (!realm) {
1056 realm = ceph_create_snap_realm(mdsc, split);
1057 if (IS_ERR(realm))
1058 goto out;
1059 }
1060
1061 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
1062 for (i = 0; i < num_split_inos; i++) {
1063 struct ceph_vino vino = {
1064 .ino = le64_to_cpu(split_inos[i]),
1065 .snap = CEPH_NOSNAP,
1066 };
1067 struct inode *inode = ceph_find_inode(sb, vino);
1068 struct ceph_inode_info *ci;
1069
1070 if (!inode)
1071 continue;
1072 ci = ceph_inode(inode);
1073
1074 spin_lock(&ci->i_ceph_lock);
1075 if (!ci->i_snap_realm)
1076 goto skip_inode;
1077 /*
1078 * If this inode belongs to a realm that was
1079 * created after our new realm, we experienced
1080 * a race (due to another split notifications
1081 * arriving from a different MDS). So skip
1082 * this inode.
1083 */
1084 if (ci->i_snap_realm->created >
1085 le64_to_cpu(ri->created)) {
1086 dout(" leaving %p %llx.%llx in newer realm %llx %p\n",
1087 inode, ceph_vinop(inode), ci->i_snap_realm->ino,
1088 ci->i_snap_realm);
1089 goto skip_inode;
1090 }
1091 dout(" will move %p %llx.%llx to split realm %llx %p\n",
1092 inode, ceph_vinop(inode), realm->ino, realm);
1093
1094 ceph_get_snap_realm(mdsc, realm);
1095 ceph_change_snap_realm(inode, realm);
1096 spin_unlock(&ci->i_ceph_lock);
1097 iput(inode);
1098 continue;
1099
1100skip_inode:
1101 spin_unlock(&ci->i_ceph_lock);
1102 iput(inode);
1103 }
1104
1105 /* we may have taken some of the old realm's children. */
1106 for (i = 0; i < num_split_realms; i++) {
1107 struct ceph_snap_realm *child =
1108 __lookup_snap_realm(mdsc,
1109 le64_to_cpu(split_realms[i]));
1110 if (!child)
1111 continue;
1112 adjust_snap_realm_parent(mdsc, child, realm->ino);
1113 }
1114 }
1115
1116 /*
1117 * update using the provided snap trace. if we are deleting a
1118 * snap, we can avoid queueing cap_snaps.
1119 */
1120 if (ceph_update_snap_trace(mdsc, p, e,
1121 op == CEPH_SNAP_OP_DESTROY,
1122 NULL)) {
1123 close_sessions = true;
1124 goto bad;
1125 }
1126
1127 if (op == CEPH_SNAP_OP_SPLIT)
1128 /* we took a reference when we created the realm, above */
1129 ceph_put_snap_realm(mdsc, realm);
1130
1131 __cleanup_empty_realms(mdsc);
1132
1133 up_write(&mdsc->snap_rwsem);
1134
1135 flush_snaps(mdsc);
1136 return;
1137
1138bad:
1139 pr_err("%s corrupt snap message from mds%d\n", __func__, mds);
1140 ceph_msg_dump(msg);
1141out:
1142 if (locked_rwsem)
1143 up_write(&mdsc->snap_rwsem);
1144
1145 if (close_sessions)
1146 ceph_mdsc_close_sessions(mdsc);
1147 return;
1148}
1149
1150struct ceph_snapid_map* ceph_get_snapid_map(struct ceph_mds_client *mdsc,
1151 u64 snap)
1152{
1153 struct ceph_snapid_map *sm, *exist;
1154 struct rb_node **p, *parent;
1155 int ret;
1156
1157 exist = NULL;
1158 spin_lock(&mdsc->snapid_map_lock);
1159 p = &mdsc->snapid_map_tree.rb_node;
1160 while (*p) {
1161 exist = rb_entry(*p, struct ceph_snapid_map, node);
1162 if (snap > exist->snap) {
1163 p = &(*p)->rb_left;
1164 } else if (snap < exist->snap) {
1165 p = &(*p)->rb_right;
1166 } else {
1167 if (atomic_inc_return(&exist->ref) == 1)
1168 list_del_init(&exist->lru);
1169 break;
1170 }
1171 exist = NULL;
1172 }
1173 spin_unlock(&mdsc->snapid_map_lock);
1174 if (exist) {
1175 dout("%s found snapid map %llx -> %x\n", __func__,
1176 exist->snap, exist->dev);
1177 return exist;
1178 }
1179
1180 sm = kmalloc(sizeof(*sm), GFP_NOFS);
1181 if (!sm)
1182 return NULL;
1183
1184 ret = get_anon_bdev(&sm->dev);
1185 if (ret < 0) {
1186 kfree(sm);
1187 return NULL;
1188 }
1189
1190 INIT_LIST_HEAD(&sm->lru);
1191 atomic_set(&sm->ref, 1);
1192 sm->snap = snap;
1193
1194 exist = NULL;
1195 parent = NULL;
1196 p = &mdsc->snapid_map_tree.rb_node;
1197 spin_lock(&mdsc->snapid_map_lock);
1198 while (*p) {
1199 parent = *p;
1200 exist = rb_entry(*p, struct ceph_snapid_map, node);
1201 if (snap > exist->snap)
1202 p = &(*p)->rb_left;
1203 else if (snap < exist->snap)
1204 p = &(*p)->rb_right;
1205 else
1206 break;
1207 exist = NULL;
1208 }
1209 if (exist) {
1210 if (atomic_inc_return(&exist->ref) == 1)
1211 list_del_init(&exist->lru);
1212 } else {
1213 rb_link_node(&sm->node, parent, p);
1214 rb_insert_color(&sm->node, &mdsc->snapid_map_tree);
1215 }
1216 spin_unlock(&mdsc->snapid_map_lock);
1217 if (exist) {
1218 free_anon_bdev(sm->dev);
1219 kfree(sm);
1220 dout("%s found snapid map %llx -> %x\n", __func__,
1221 exist->snap, exist->dev);
1222 return exist;
1223 }
1224
1225 dout("%s create snapid map %llx -> %x\n", __func__,
1226 sm->snap, sm->dev);
1227 return sm;
1228}
1229
1230void ceph_put_snapid_map(struct ceph_mds_client* mdsc,
1231 struct ceph_snapid_map *sm)
1232{
1233 if (!sm)
1234 return;
1235 if (atomic_dec_and_lock(&sm->ref, &mdsc->snapid_map_lock)) {
1236 if (!RB_EMPTY_NODE(&sm->node)) {
1237 sm->last_used = jiffies;
1238 list_add_tail(&sm->lru, &mdsc->snapid_map_lru);
1239 spin_unlock(&mdsc->snapid_map_lock);
1240 } else {
1241 /* already cleaned up by
1242 * ceph_cleanup_snapid_map() */
1243 spin_unlock(&mdsc->snapid_map_lock);
1244 kfree(sm);
1245 }
1246 }
1247}
1248
1249void ceph_trim_snapid_map(struct ceph_mds_client *mdsc)
1250{
1251 struct ceph_snapid_map *sm;
1252 unsigned long now;
1253 LIST_HEAD(to_free);
1254
1255 spin_lock(&mdsc->snapid_map_lock);
1256 now = jiffies;
1257
1258 while (!list_empty(&mdsc->snapid_map_lru)) {
1259 sm = list_first_entry(&mdsc->snapid_map_lru,
1260 struct ceph_snapid_map, lru);
1261 if (time_after(sm->last_used + CEPH_SNAPID_MAP_TIMEOUT, now))
1262 break;
1263
1264 rb_erase(&sm->node, &mdsc->snapid_map_tree);
1265 list_move(&sm->lru, &to_free);
1266 }
1267 spin_unlock(&mdsc->snapid_map_lock);
1268
1269 while (!list_empty(&to_free)) {
1270 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1271 list_del(&sm->lru);
1272 dout("trim snapid map %llx -> %x\n", sm->snap, sm->dev);
1273 free_anon_bdev(sm->dev);
1274 kfree(sm);
1275 }
1276}
1277
1278void ceph_cleanup_snapid_map(struct ceph_mds_client *mdsc)
1279{
1280 struct ceph_snapid_map *sm;
1281 struct rb_node *p;
1282 LIST_HEAD(to_free);
1283
1284 spin_lock(&mdsc->snapid_map_lock);
1285 while ((p = rb_first(&mdsc->snapid_map_tree))) {
1286 sm = rb_entry(p, struct ceph_snapid_map, node);
1287 rb_erase(p, &mdsc->snapid_map_tree);
1288 RB_CLEAR_NODE(p);
1289 list_move(&sm->lru, &to_free);
1290 }
1291 spin_unlock(&mdsc->snapid_map_lock);
1292
1293 while (!list_empty(&to_free)) {
1294 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1295 list_del(&sm->lru);
1296 free_anon_bdev(sm->dev);
1297 if (WARN_ON_ONCE(atomic_read(&sm->ref))) {
1298 pr_err("snapid map %llx -> %x still in use\n",
1299 sm->snap, sm->dev);
1300 }
1301 kfree(sm);
1302 }
1303}