blob: dcf18d92130a2c56dba4f1d92453abdcff828cd1 [file] [log] [blame]
Sage Weil963b61e2009-10-06 11:31:12 -07001#include "ceph_debug.h"
2
3#include <linux/radix-tree.h>
4#include <linux/sort.h>
5
6#include "super.h"
7#include "decode.h"
8
9/*
10 * Snapshots in ceph are driven in large part by cooperation from the
11 * client. In contrast to local file systems or file servers that
12 * implement snapshots at a single point in the system, ceph's
13 * distributed access to storage requires clients to help decide
14 * whether a write logically occurs before or after a recently created
15 * snapshot.
16 *
17 * This provides a perfect instantanous client-wide snapshot. Between
18 * clients, however, snapshots may appear to be applied at slightly
19 * different points in time, depending on delays in delivering the
20 * snapshot notification.
21 *
22 * Snapshots are _not_ file system-wide. Instead, each snapshot
23 * applies to the subdirectory nested beneath some directory. This
24 * effectively divides the hierarchy into multiple "realms," where all
25 * of the files contained by each realm share the same set of
26 * snapshots. An individual realm's snap set contains snapshots
27 * explicitly created on that realm, as well as any snaps in its
28 * parent's snap set _after_ the point at which the parent became it's
29 * parent (due to, say, a rename). Similarly, snaps from prior parents
30 * during the time intervals during which they were the parent are included.
31 *
32 * The client is spared most of this detail, fortunately... it must only
33 * maintains a hierarchy of realms reflecting the current parent/child
34 * realm relationship, and for each realm has an explicit list of snaps
35 * inherited from prior parents.
36 *
37 * A snap_realm struct is maintained for realms containing every inode
38 * with an open cap in the system. (The needed snap realm information is
39 * provided by the MDS whenever a cap is issued, i.e., on open.) A 'seq'
40 * version number is used to ensure that as realm parameters change (new
41 * snapshot, new parent, etc.) the client's realm hierarchy is updated.
42 *
43 * The realm hierarchy drives the generation of a 'snap context' for each
44 * realm, which simply lists the resulting set of snaps for the realm. This
45 * is attached to any writes sent to OSDs.
46 */
47/*
48 * Unfortunately error handling is a bit mixed here. If we get a snap
49 * update, but don't have enough memory to update our realm hierarchy,
50 * it's not clear what we can do about it (besides complaining to the
51 * console).
52 */
53
54
55/*
56 * increase ref count for the realm
57 *
58 * caller must hold snap_rwsem for write.
59 */
60void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
61 struct ceph_snap_realm *realm)
62{
63 dout("get_realm %p %d -> %d\n", realm,
64 atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
65 /*
66 * since we _only_ increment realm refs or empty the empty
67 * list with snap_rwsem held, adjusting the empty list here is
68 * safe. we do need to protect against concurrent empty list
69 * additions, however.
70 */
71 if (atomic_read(&realm->nref) == 0) {
72 spin_lock(&mdsc->snap_empty_lock);
73 list_del_init(&realm->empty_item);
74 spin_unlock(&mdsc->snap_empty_lock);
75 }
76
77 atomic_inc(&realm->nref);
78}
79
80/*
81 * create and get the realm rooted at @ino and bump its ref count.
82 *
83 * caller must hold snap_rwsem for write.
84 */
85static struct ceph_snap_realm *ceph_create_snap_realm(
86 struct ceph_mds_client *mdsc,
87 u64 ino)
88{
89 struct ceph_snap_realm *realm;
90
91 realm = kzalloc(sizeof(*realm), GFP_NOFS);
92 if (!realm)
93 return ERR_PTR(-ENOMEM);
94
95 radix_tree_insert(&mdsc->snap_realms, ino, realm);
96
97 atomic_set(&realm->nref, 0); /* tree does not take a ref */
98 realm->ino = ino;
99 INIT_LIST_HEAD(&realm->children);
100 INIT_LIST_HEAD(&realm->child_item);
101 INIT_LIST_HEAD(&realm->empty_item);
102 INIT_LIST_HEAD(&realm->inodes_with_caps);
103 spin_lock_init(&realm->inodes_with_caps_lock);
104 dout("create_snap_realm %llx %p\n", realm->ino, realm);
105 return realm;
106}
107
108/*
109 * find and get (if found) the realm rooted at @ino and bump its ref count.
110 *
111 * caller must hold snap_rwsem for write.
112 */
113struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
114 u64 ino)
115{
116 struct ceph_snap_realm *realm;
117
118 realm = radix_tree_lookup(&mdsc->snap_realms, ino);
119 if (realm)
120 dout("lookup_snap_realm %llx %p\n", realm->ino, realm);
121 return realm;
122}
123
124static void __put_snap_realm(struct ceph_mds_client *mdsc,
125 struct ceph_snap_realm *realm);
126
127/*
128 * called with snap_rwsem (write)
129 */
130static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
131 struct ceph_snap_realm *realm)
132{
133 dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
134
135 radix_tree_delete(&mdsc->snap_realms, realm->ino);
136
137 if (realm->parent) {
138 list_del_init(&realm->child_item);
139 __put_snap_realm(mdsc, realm->parent);
140 }
141
142 kfree(realm->prior_parent_snaps);
143 kfree(realm->snaps);
144 ceph_put_snap_context(realm->cached_context);
145 kfree(realm);
146}
147
148/*
149 * caller holds snap_rwsem (write)
150 */
151static void __put_snap_realm(struct ceph_mds_client *mdsc,
152 struct ceph_snap_realm *realm)
153{
154 dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
155 atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
156 if (atomic_dec_and_test(&realm->nref))
157 __destroy_snap_realm(mdsc, realm);
158}
159
160/*
161 * caller needn't hold any locks
162 */
163void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
164 struct ceph_snap_realm *realm)
165{
166 dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
167 atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
168 if (!atomic_dec_and_test(&realm->nref))
169 return;
170
171 if (down_write_trylock(&mdsc->snap_rwsem)) {
172 __destroy_snap_realm(mdsc, realm);
173 up_write(&mdsc->snap_rwsem);
174 } else {
175 spin_lock(&mdsc->snap_empty_lock);
176 list_add(&mdsc->snap_empty, &realm->empty_item);
177 spin_unlock(&mdsc->snap_empty_lock);
178 }
179}
180
181/*
182 * Clean up any realms whose ref counts have dropped to zero. Note
183 * that this does not include realms who were created but not yet
184 * used.
185 *
186 * Called under snap_rwsem (write)
187 */
188static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
189{
190 struct ceph_snap_realm *realm;
191
192 spin_lock(&mdsc->snap_empty_lock);
193 while (!list_empty(&mdsc->snap_empty)) {
194 realm = list_first_entry(&mdsc->snap_empty,
195 struct ceph_snap_realm, empty_item);
196 list_del(&realm->empty_item);
197 spin_unlock(&mdsc->snap_empty_lock);
198 __destroy_snap_realm(mdsc, realm);
199 spin_lock(&mdsc->snap_empty_lock);
200 }
201 spin_unlock(&mdsc->snap_empty_lock);
202}
203
204void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
205{
206 down_write(&mdsc->snap_rwsem);
207 __cleanup_empty_realms(mdsc);
208 up_write(&mdsc->snap_rwsem);
209}
210
211/*
212 * adjust the parent realm of a given @realm. adjust child list, and parent
213 * pointers, and ref counts appropriately.
214 *
215 * return true if parent was changed, 0 if unchanged, <0 on error.
216 *
217 * caller must hold snap_rwsem for write.
218 */
219static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
220 struct ceph_snap_realm *realm,
221 u64 parentino)
222{
223 struct ceph_snap_realm *parent;
224
225 if (realm->parent_ino == parentino)
226 return 0;
227
228 parent = ceph_lookup_snap_realm(mdsc, parentino);
Sage Weil963b61e2009-10-06 11:31:12 -0700229 if (!parent) {
230 parent = ceph_create_snap_realm(mdsc, parentino);
231 if (IS_ERR(parent))
232 return PTR_ERR(parent);
233 }
234 dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
235 realm->ino, realm, realm->parent_ino, realm->parent,
236 parentino, parent);
237 if (realm->parent) {
238 list_del_init(&realm->child_item);
239 ceph_put_snap_realm(mdsc, realm->parent);
240 }
241 realm->parent_ino = parentino;
242 realm->parent = parent;
243 ceph_get_snap_realm(mdsc, parent);
244 list_add(&realm->child_item, &parent->children);
245 return 1;
246}
247
248
249static int cmpu64_rev(const void *a, const void *b)
250{
251 if (*(u64 *)a < *(u64 *)b)
252 return 1;
253 if (*(u64 *)a > *(u64 *)b)
254 return -1;
255 return 0;
256}
257
258/*
259 * build the snap context for a given realm.
260 */
261static int build_snap_context(struct ceph_snap_realm *realm)
262{
263 struct ceph_snap_realm *parent = realm->parent;
264 struct ceph_snap_context *snapc;
265 int err = 0;
266 int i;
267 int num = realm->num_prior_parent_snaps + realm->num_snaps;
268
269 /*
270 * build parent context, if it hasn't been built.
271 * conservatively estimate that all parent snaps might be
272 * included by us.
273 */
274 if (parent) {
275 if (!parent->cached_context) {
276 err = build_snap_context(parent);
277 if (err)
278 goto fail;
279 }
280 num += parent->cached_context->num_snaps;
281 }
282
283 /* do i actually need to update? not if my context seq
284 matches realm seq, and my parents' does to. (this works
285 because we rebuild_snap_realms() works _downward_ in
286 hierarchy after each update.) */
287 if (realm->cached_context &&
288 realm->cached_context->seq <= realm->seq &&
289 (!parent ||
290 realm->cached_context->seq <= parent->cached_context->seq)) {
291 dout("build_snap_context %llx %p: %p seq %lld (%d snaps)"
292 " (unchanged)\n",
293 realm->ino, realm, realm->cached_context,
294 realm->cached_context->seq,
295 realm->cached_context->num_snaps);
296 return 0;
297 }
298
299 /* alloc new snap context */
300 err = -ENOMEM;
301 if (num > ULONG_MAX / sizeof(u64) - sizeof(*snapc))
302 goto fail;
303 snapc = kzalloc(sizeof(*snapc) + num*sizeof(u64), GFP_NOFS);
304 if (!snapc)
305 goto fail;
306 atomic_set(&snapc->nref, 1);
307
308 /* build (reverse sorted) snap vector */
309 num = 0;
310 snapc->seq = realm->seq;
311 if (parent) {
312 /* include any of parent's snaps occuring _after_ my
313 parent became my parent */
314 for (i = 0; i < parent->cached_context->num_snaps; i++)
315 if (parent->cached_context->snaps[i] >=
316 realm->parent_since)
317 snapc->snaps[num++] =
318 parent->cached_context->snaps[i];
319 if (parent->cached_context->seq > snapc->seq)
320 snapc->seq = parent->cached_context->seq;
321 }
322 memcpy(snapc->snaps + num, realm->snaps,
323 sizeof(u64)*realm->num_snaps);
324 num += realm->num_snaps;
325 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
326 sizeof(u64)*realm->num_prior_parent_snaps);
327 num += realm->num_prior_parent_snaps;
328
329 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
330 snapc->num_snaps = num;
331 dout("build_snap_context %llx %p: %p seq %lld (%d snaps)\n",
332 realm->ino, realm, snapc, snapc->seq, snapc->num_snaps);
333
334 if (realm->cached_context)
335 ceph_put_snap_context(realm->cached_context);
336 realm->cached_context = snapc;
337 return 0;
338
339fail:
340 /*
341 * if we fail, clear old (incorrect) cached_context... hopefully
342 * we'll have better luck building it later
343 */
344 if (realm->cached_context) {
345 ceph_put_snap_context(realm->cached_context);
346 realm->cached_context = NULL;
347 }
348 pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
349 realm, err);
350 return err;
351}
352
353/*
354 * rebuild snap context for the given realm and all of its children.
355 */
356static void rebuild_snap_realms(struct ceph_snap_realm *realm)
357{
358 struct ceph_snap_realm *child;
359
360 dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
361 build_snap_context(realm);
362
363 list_for_each_entry(child, &realm->children, child_item)
364 rebuild_snap_realms(child);
365}
366
367
368/*
369 * helper to allocate and decode an array of snapids. free prior
370 * instance, if any.
371 */
372static int dup_array(u64 **dst, __le64 *src, int num)
373{
374 int i;
375
376 kfree(*dst);
377 if (num) {
378 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
379 if (!*dst)
380 return -ENOMEM;
381 for (i = 0; i < num; i++)
382 (*dst)[i] = get_unaligned_le64(src + i);
383 } else {
384 *dst = NULL;
385 }
386 return 0;
387}
388
389
390/*
391 * When a snapshot is applied, the size/mtime inode metadata is queued
392 * in a ceph_cap_snap (one for each snapshot) until writeback
393 * completes and the metadata can be flushed back to the MDS.
394 *
395 * However, if a (sync) write is currently in-progress when we apply
396 * the snapshot, we have to wait until the write succeeds or fails
397 * (and a final size/mtime is known). In this case the
398 * cap_snap->writing = 1, and is said to be "pending." When the write
399 * finishes, we __ceph_finish_cap_snap().
400 *
401 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
402 * change).
403 */
404void ceph_queue_cap_snap(struct ceph_inode_info *ci,
405 struct ceph_snap_context *snapc)
406{
407 struct inode *inode = &ci->vfs_inode;
408 struct ceph_cap_snap *capsnap;
409 int used;
410
411 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
412 if (!capsnap) {
413 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
414 return;
415 }
416
417 spin_lock(&inode->i_lock);
418 used = __ceph_caps_used(ci);
419 if (__ceph_have_pending_cap_snap(ci)) {
420 /* there is no point in queuing multiple "pending" cap_snaps,
421 as no new writes are allowed to start when pending, so any
422 writes in progress now were started before the previous
423 cap_snap. lucky us. */
424 dout("queue_cap_snap %p snapc %p seq %llu used %d"
425 " already pending\n", inode, snapc, snapc->seq, used);
426 kfree(capsnap);
427 } else if (ci->i_wrbuffer_ref_head || (used & CEPH_CAP_FILE_WR)) {
428 igrab(inode);
429
430 atomic_set(&capsnap->nref, 1);
431 capsnap->ci = ci;
432 INIT_LIST_HEAD(&capsnap->ci_item);
433 INIT_LIST_HEAD(&capsnap->flushing_item);
434
435 capsnap->follows = snapc->seq - 1;
436 capsnap->context = ceph_get_snap_context(snapc);
437 capsnap->issued = __ceph_caps_issued(ci, NULL);
438 capsnap->dirty = __ceph_caps_dirty(ci);
439
440 capsnap->mode = inode->i_mode;
441 capsnap->uid = inode->i_uid;
442 capsnap->gid = inode->i_gid;
443
444 /* fixme? */
445 capsnap->xattr_blob = NULL;
446 capsnap->xattr_len = 0;
447
448 /* dirty page count moved from _head to this cap_snap;
449 all subsequent writes page dirties occur _after_ this
450 snapshot. */
451 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
452 ci->i_wrbuffer_ref_head = 0;
453 ceph_put_snap_context(ci->i_head_snapc);
454 ci->i_head_snapc = NULL;
455 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
456
457 if (used & CEPH_CAP_FILE_WR) {
458 dout("queue_cap_snap %p cap_snap %p snapc %p"
459 " seq %llu used WR, now pending\n", inode,
460 capsnap, snapc, snapc->seq);
461 capsnap->writing = 1;
462 } else {
463 /* note mtime, size NOW. */
464 __ceph_finish_cap_snap(ci, capsnap);
465 }
466 } else {
467 dout("queue_cap_snap %p nothing dirty|writing\n", inode);
468 kfree(capsnap);
469 }
470
471 spin_unlock(&inode->i_lock);
472}
473
474/*
475 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
476 * to be used for the snapshot, to be flushed back to the mds.
477 *
478 * If capsnap can now be flushed, add to snap_flush list, and return 1.
479 *
480 * Caller must hold i_lock.
481 */
482int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
483 struct ceph_cap_snap *capsnap)
484{
485 struct inode *inode = &ci->vfs_inode;
486 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
487
488 BUG_ON(capsnap->writing);
489 capsnap->size = inode->i_size;
490 capsnap->mtime = inode->i_mtime;
491 capsnap->atime = inode->i_atime;
492 capsnap->ctime = inode->i_ctime;
493 capsnap->time_warp_seq = ci->i_time_warp_seq;
494 if (capsnap->dirty_pages) {
495 dout("finish_cap_snap %p cap_snap %p snapc %p %llu s=%llu "
496 "still has %d dirty pages\n", inode, capsnap,
497 capsnap->context, capsnap->context->seq,
498 capsnap->size, capsnap->dirty_pages);
499 return 0;
500 }
501 dout("finish_cap_snap %p cap_snap %p snapc %p %llu s=%llu clean\n",
502 inode, capsnap, capsnap->context,
503 capsnap->context->seq, capsnap->size);
504
505 spin_lock(&mdsc->snap_flush_lock);
506 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
507 spin_unlock(&mdsc->snap_flush_lock);
508 return 1; /* caller may want to ceph_flush_snaps */
509}
510
511
512/*
513 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
514 * the snap realm parameters from a given realm and all of its ancestors,
515 * up to the root.
516 *
517 * Caller must hold snap_rwsem for write.
518 */
519int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
520 void *p, void *e, bool deletion)
521{
522 struct ceph_mds_snap_realm *ri; /* encoded */
523 __le64 *snaps; /* encoded */
524 __le64 *prior_parent_snaps; /* encoded */
525 struct ceph_snap_realm *realm;
526 int invalidate = 0;
527 int err = -ENOMEM;
528
529 dout("update_snap_trace deletion=%d\n", deletion);
530more:
531 ceph_decode_need(&p, e, sizeof(*ri), bad);
532 ri = p;
533 p += sizeof(*ri);
534 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
535 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
536 snaps = p;
537 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
538 prior_parent_snaps = p;
539 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
540
541 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
Sage Weil963b61e2009-10-06 11:31:12 -0700542 if (!realm) {
543 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
544 if (IS_ERR(realm)) {
545 err = PTR_ERR(realm);
546 goto fail;
547 }
548 }
549
550 if (le64_to_cpu(ri->seq) > realm->seq) {
551 dout("update_snap_trace updating %llx %p %lld -> %lld\n",
552 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
553 /*
554 * if the realm seq has changed, queue a cap_snap for every
555 * inode with open caps. we do this _before_ we update
556 * the realm info so that we prepare for writeback under the
557 * _previous_ snap context.
558 *
559 * ...unless it's a snap deletion!
560 */
561 if (!deletion) {
562 struct ceph_inode_info *ci;
563 struct inode *lastinode = NULL;
564
565 spin_lock(&realm->inodes_with_caps_lock);
566 list_for_each_entry(ci, &realm->inodes_with_caps,
567 i_snap_realm_item) {
568 struct inode *inode = igrab(&ci->vfs_inode);
569 if (!inode)
570 continue;
571 spin_unlock(&realm->inodes_with_caps_lock);
572 if (lastinode)
573 iput(lastinode);
574 lastinode = inode;
575 ceph_queue_cap_snap(ci, realm->cached_context);
576 spin_lock(&realm->inodes_with_caps_lock);
577 }
578 spin_unlock(&realm->inodes_with_caps_lock);
579 if (lastinode)
580 iput(lastinode);
581 dout("update_snap_trace cap_snaps queued\n");
582 }
583
584 } else {
585 dout("update_snap_trace %llx %p seq %lld unchanged\n",
586 realm->ino, realm, realm->seq);
587 }
588
589 /* ensure the parent is correct */
590 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
591 if (err < 0)
592 goto fail;
593 invalidate += err;
594
595 if (le64_to_cpu(ri->seq) > realm->seq) {
596 /* update realm parameters, snap lists */
597 realm->seq = le64_to_cpu(ri->seq);
598 realm->created = le64_to_cpu(ri->created);
599 realm->parent_since = le64_to_cpu(ri->parent_since);
600
601 realm->num_snaps = le32_to_cpu(ri->num_snaps);
602 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
603 if (err < 0)
604 goto fail;
605
606 realm->num_prior_parent_snaps =
607 le32_to_cpu(ri->num_prior_parent_snaps);
608 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
609 realm->num_prior_parent_snaps);
610 if (err < 0)
611 goto fail;
612
613 invalidate = 1;
614 } else if (!realm->cached_context) {
615 invalidate = 1;
616 }
617
618 dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
619 realm, invalidate, p, e);
620
621 if (p < e)
622 goto more;
623
624 /* invalidate when we reach the _end_ (root) of the trace */
625 if (invalidate)
626 rebuild_snap_realms(realm);
627
628 __cleanup_empty_realms(mdsc);
629 return 0;
630
631bad:
632 err = -EINVAL;
633fail:
634 pr_err("update_snap_trace error %d\n", err);
635 return err;
636}
637
638
639/*
640 * Send any cap_snaps that are queued for flush. Try to carry
641 * s_mutex across multiple snap flushes to avoid locking overhead.
642 *
643 * Caller holds no locks.
644 */
645static void flush_snaps(struct ceph_mds_client *mdsc)
646{
647 struct ceph_inode_info *ci;
648 struct inode *inode;
649 struct ceph_mds_session *session = NULL;
650
651 dout("flush_snaps\n");
652 spin_lock(&mdsc->snap_flush_lock);
653 while (!list_empty(&mdsc->snap_flush_list)) {
654 ci = list_first_entry(&mdsc->snap_flush_list,
655 struct ceph_inode_info, i_snap_flush_item);
656 inode = &ci->vfs_inode;
657 igrab(inode);
658 spin_unlock(&mdsc->snap_flush_lock);
659 spin_lock(&inode->i_lock);
660 __ceph_flush_snaps(ci, &session);
661 spin_unlock(&inode->i_lock);
662 iput(inode);
663 spin_lock(&mdsc->snap_flush_lock);
664 }
665 spin_unlock(&mdsc->snap_flush_lock);
666
667 if (session) {
668 mutex_unlock(&session->s_mutex);
669 ceph_put_mds_session(session);
670 }
671 dout("flush_snaps done\n");
672}
673
674
675/*
676 * Handle a snap notification from the MDS.
677 *
678 * This can take two basic forms: the simplest is just a snap creation
679 * or deletion notification on an existing realm. This should update the
680 * realm and its children.
681 *
682 * The more difficult case is realm creation, due to snap creation at a
683 * new point in the file hierarchy, or due to a rename that moves a file or
684 * directory into another realm.
685 */
686void ceph_handle_snap(struct ceph_mds_client *mdsc,
687 struct ceph_msg *msg)
688{
689 struct super_block *sb = mdsc->client->sb;
690 struct ceph_mds_session *session;
691 int mds;
692 u64 split;
693 int op;
694 int trace_len;
695 struct ceph_snap_realm *realm = NULL;
696 void *p = msg->front.iov_base;
697 void *e = p + msg->front.iov_len;
698 struct ceph_mds_snap_head *h;
699 int num_split_inos, num_split_realms;
700 __le64 *split_inos = NULL, *split_realms = NULL;
701 int i;
702 int locked_rwsem = 0;
703
704 if (msg->hdr.src.name.type != CEPH_ENTITY_TYPE_MDS)
705 return;
706 mds = le64_to_cpu(msg->hdr.src.name.num);
707
708 /* decode */
709 if (msg->front.iov_len < sizeof(*h))
710 goto bad;
711 h = p;
712 op = le32_to_cpu(h->op);
713 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
714 * existing realm */
715 num_split_inos = le32_to_cpu(h->num_split_inos);
716 num_split_realms = le32_to_cpu(h->num_split_realms);
717 trace_len = le32_to_cpu(h->trace_len);
718 p += sizeof(*h);
719
720 dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
721 ceph_snap_op_name(op), split, trace_len);
722
723 /* find session */
724 mutex_lock(&mdsc->mutex);
725 session = __ceph_lookup_mds_session(mdsc, mds);
726 mutex_unlock(&mdsc->mutex);
727 if (!session) {
728 dout("WTF, got snap but no session for mds%d\n", mds);
729 return;
730 }
731
732 mutex_lock(&session->s_mutex);
733 session->s_seq++;
734 mutex_unlock(&session->s_mutex);
735
736 down_write(&mdsc->snap_rwsem);
737 locked_rwsem = 1;
738
739 if (op == CEPH_SNAP_OP_SPLIT) {
740 struct ceph_mds_snap_realm *ri;
741
742 /*
743 * A "split" breaks part of an existing realm off into
744 * a new realm. The MDS provides a list of inodes
745 * (with caps) and child realms that belong to the new
746 * child.
747 */
748 split_inos = p;
749 p += sizeof(u64) * num_split_inos;
750 split_realms = p;
751 p += sizeof(u64) * num_split_realms;
752 ceph_decode_need(&p, e, sizeof(*ri), bad);
753 /* we will peek at realm info here, but will _not_
754 * advance p, as the realm update will occur below in
755 * ceph_update_snap_trace. */
756 ri = p;
757
758 realm = ceph_lookup_snap_realm(mdsc, split);
Sage Weil963b61e2009-10-06 11:31:12 -0700759 if (!realm) {
760 realm = ceph_create_snap_realm(mdsc, split);
761 if (IS_ERR(realm))
762 goto out;
763 }
764 ceph_get_snap_realm(mdsc, realm);
765
766 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
767 for (i = 0; i < num_split_inos; i++) {
768 struct ceph_vino vino = {
769 .ino = le64_to_cpu(split_inos[i]),
770 .snap = CEPH_NOSNAP,
771 };
772 struct inode *inode = ceph_find_inode(sb, vino);
773 struct ceph_inode_info *ci;
774
775 if (!inode)
776 continue;
777 ci = ceph_inode(inode);
778
779 spin_lock(&inode->i_lock);
780 if (!ci->i_snap_realm)
781 goto skip_inode;
782 /*
783 * If this inode belongs to a realm that was
784 * created after our new realm, we experienced
785 * a race (due to another split notifications
786 * arriving from a different MDS). So skip
787 * this inode.
788 */
789 if (ci->i_snap_realm->created >
790 le64_to_cpu(ri->created)) {
791 dout(" leaving %p in newer realm %llx %p\n",
792 inode, ci->i_snap_realm->ino,
793 ci->i_snap_realm);
794 goto skip_inode;
795 }
796 dout(" will move %p to split realm %llx %p\n",
797 inode, realm->ino, realm);
798 /*
799 * Remove the inode from the realm's inode
800 * list, but don't add it to the new realm
801 * yet. We don't want the cap_snap to be
802 * queued (again) by ceph_update_snap_trace()
803 * below. Queue it _now_, under the old context.
804 */
805 list_del_init(&ci->i_snap_realm_item);
806 spin_unlock(&inode->i_lock);
807
808 ceph_queue_cap_snap(ci,
809 ci->i_snap_realm->cached_context);
810
811 iput(inode);
812 continue;
813
814skip_inode:
815 spin_unlock(&inode->i_lock);
816 iput(inode);
817 }
818
819 /* we may have taken some of the old realm's children. */
820 for (i = 0; i < num_split_realms; i++) {
821 struct ceph_snap_realm *child =
822 ceph_lookup_snap_realm(mdsc,
823 le64_to_cpu(split_realms[i]));
Sage Weil963b61e2009-10-06 11:31:12 -0700824 if (!child)
825 continue;
826 adjust_snap_realm_parent(mdsc, child, realm->ino);
827 }
828 }
829
830 /*
831 * update using the provided snap trace. if we are deleting a
832 * snap, we can avoid queueing cap_snaps.
833 */
834 ceph_update_snap_trace(mdsc, p, e,
835 op == CEPH_SNAP_OP_DESTROY);
836
837 if (op == CEPH_SNAP_OP_SPLIT) {
838 /*
839 * ok, _now_ add the inodes into the new realm.
840 */
841 for (i = 0; i < num_split_inos; i++) {
842 struct ceph_vino vino = {
843 .ino = le64_to_cpu(split_inos[i]),
844 .snap = CEPH_NOSNAP,
845 };
846 struct inode *inode = ceph_find_inode(sb, vino);
847 struct ceph_inode_info *ci;
848
849 if (!inode)
850 continue;
851 ci = ceph_inode(inode);
852 spin_lock(&inode->i_lock);
853 if (!ci->i_snap_realm)
854 goto split_skip_inode;
855 ceph_put_snap_realm(mdsc, ci->i_snap_realm);
856 spin_lock(&realm->inodes_with_caps_lock);
857 list_add(&ci->i_snap_realm_item,
858 &realm->inodes_with_caps);
859 ci->i_snap_realm = realm;
860 spin_unlock(&realm->inodes_with_caps_lock);
861 ceph_get_snap_realm(mdsc, realm);
862split_skip_inode:
863 spin_unlock(&inode->i_lock);
864 iput(inode);
865 }
866
867 /* we took a reference when we created the realm, above */
868 ceph_put_snap_realm(mdsc, realm);
869 }
870
871 __cleanup_empty_realms(mdsc);
872
873 up_write(&mdsc->snap_rwsem);
874
875 flush_snaps(mdsc);
876 return;
877
878bad:
879 pr_err("corrupt snap message from mds%d\n", mds);
Sage Weil9ec7cab2009-12-14 15:13:47 -0800880 ceph_msg_dump(msg);
Sage Weil963b61e2009-10-06 11:31:12 -0700881out:
882 if (locked_rwsem)
883 up_write(&mdsc->snap_rwsem);
884 return;
885}
886
887
888