blob: 49d0c4c59d81b6cdc919b1b088dd9e8db92c355d [file] [log] [blame]
Sage Weil963b61e2009-10-06 11:31:12 -07001#include "ceph_debug.h"
2
Sage Weil963b61e2009-10-06 11:31:12 -07003#include <linux/sort.h>
4
5#include "super.h"
6#include "decode.h"
7
8/*
9 * Snapshots in ceph are driven in large part by cooperation from the
10 * client. In contrast to local file systems or file servers that
11 * implement snapshots at a single point in the system, ceph's
12 * distributed access to storage requires clients to help decide
13 * whether a write logically occurs before or after a recently created
14 * snapshot.
15 *
16 * This provides a perfect instantanous client-wide snapshot. Between
17 * clients, however, snapshots may appear to be applied at slightly
18 * different points in time, depending on delays in delivering the
19 * snapshot notification.
20 *
21 * Snapshots are _not_ file system-wide. Instead, each snapshot
22 * applies to the subdirectory nested beneath some directory. This
23 * effectively divides the hierarchy into multiple "realms," where all
24 * of the files contained by each realm share the same set of
25 * snapshots. An individual realm's snap set contains snapshots
26 * explicitly created on that realm, as well as any snaps in its
27 * parent's snap set _after_ the point at which the parent became it's
28 * parent (due to, say, a rename). Similarly, snaps from prior parents
29 * during the time intervals during which they were the parent are included.
30 *
31 * The client is spared most of this detail, fortunately... it must only
32 * maintains a hierarchy of realms reflecting the current parent/child
33 * realm relationship, and for each realm has an explicit list of snaps
34 * inherited from prior parents.
35 *
36 * A snap_realm struct is maintained for realms containing every inode
37 * with an open cap in the system. (The needed snap realm information is
38 * provided by the MDS whenever a cap is issued, i.e., on open.) A 'seq'
39 * version number is used to ensure that as realm parameters change (new
40 * snapshot, new parent, etc.) the client's realm hierarchy is updated.
41 *
42 * The realm hierarchy drives the generation of a 'snap context' for each
43 * realm, which simply lists the resulting set of snaps for the realm. This
44 * is attached to any writes sent to OSDs.
45 */
46/*
47 * Unfortunately error handling is a bit mixed here. If we get a snap
48 * update, but don't have enough memory to update our realm hierarchy,
49 * it's not clear what we can do about it (besides complaining to the
50 * console).
51 */
52
53
54/*
55 * increase ref count for the realm
56 *
57 * caller must hold snap_rwsem for write.
58 */
59void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
60 struct ceph_snap_realm *realm)
61{
62 dout("get_realm %p %d -> %d\n", realm,
63 atomic_read(&realm->nref), atomic_read(&realm->nref)+1);
64 /*
65 * since we _only_ increment realm refs or empty the empty
66 * list with snap_rwsem held, adjusting the empty list here is
67 * safe. we do need to protect against concurrent empty list
68 * additions, however.
69 */
70 if (atomic_read(&realm->nref) == 0) {
71 spin_lock(&mdsc->snap_empty_lock);
72 list_del_init(&realm->empty_item);
73 spin_unlock(&mdsc->snap_empty_lock);
74 }
75
76 atomic_inc(&realm->nref);
77}
78
Sage Weila105f002010-02-15 14:37:55 -080079static void __insert_snap_realm(struct rb_root *root,
80 struct ceph_snap_realm *new)
81{
82 struct rb_node **p = &root->rb_node;
83 struct rb_node *parent = NULL;
84 struct ceph_snap_realm *r = NULL;
85
86 while (*p) {
87 parent = *p;
88 r = rb_entry(parent, struct ceph_snap_realm, node);
89 if (new->ino < r->ino)
90 p = &(*p)->rb_left;
91 else if (new->ino > r->ino)
92 p = &(*p)->rb_right;
93 else
94 BUG();
95 }
96
97 rb_link_node(&new->node, parent, p);
98 rb_insert_color(&new->node, root);
99}
100
Sage Weil963b61e2009-10-06 11:31:12 -0700101/*
102 * create and get the realm rooted at @ino and bump its ref count.
103 *
104 * caller must hold snap_rwsem for write.
105 */
106static struct ceph_snap_realm *ceph_create_snap_realm(
107 struct ceph_mds_client *mdsc,
108 u64 ino)
109{
110 struct ceph_snap_realm *realm;
111
112 realm = kzalloc(sizeof(*realm), GFP_NOFS);
113 if (!realm)
114 return ERR_PTR(-ENOMEM);
115
Sage Weil963b61e2009-10-06 11:31:12 -0700116 atomic_set(&realm->nref, 0); /* tree does not take a ref */
117 realm->ino = ino;
118 INIT_LIST_HEAD(&realm->children);
119 INIT_LIST_HEAD(&realm->child_item);
120 INIT_LIST_HEAD(&realm->empty_item);
121 INIT_LIST_HEAD(&realm->inodes_with_caps);
122 spin_lock_init(&realm->inodes_with_caps_lock);
Sage Weila105f002010-02-15 14:37:55 -0800123 __insert_snap_realm(&mdsc->snap_realms, realm);
Sage Weil963b61e2009-10-06 11:31:12 -0700124 dout("create_snap_realm %llx %p\n", realm->ino, realm);
125 return realm;
126}
127
128/*
Sage Weila105f002010-02-15 14:37:55 -0800129 * lookup the realm rooted at @ino.
Sage Weil963b61e2009-10-06 11:31:12 -0700130 *
131 * caller must hold snap_rwsem for write.
132 */
133struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
134 u64 ino)
135{
Sage Weila105f002010-02-15 14:37:55 -0800136 struct rb_node *n = mdsc->snap_realms.rb_node;
137 struct ceph_snap_realm *r;
Sage Weil963b61e2009-10-06 11:31:12 -0700138
Sage Weila105f002010-02-15 14:37:55 -0800139 while (n) {
140 r = rb_entry(n, struct ceph_snap_realm, node);
141 if (ino < r->ino)
142 n = n->rb_left;
143 else if (ino > r->ino)
144 n = n->rb_right;
145 else {
146 dout("lookup_snap_realm %llx %p\n", r->ino, r);
147 return r;
148 }
149 }
150 return NULL;
Sage Weil963b61e2009-10-06 11:31:12 -0700151}
152
153static void __put_snap_realm(struct ceph_mds_client *mdsc,
154 struct ceph_snap_realm *realm);
155
156/*
157 * called with snap_rwsem (write)
158 */
159static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
160 struct ceph_snap_realm *realm)
161{
162 dout("__destroy_snap_realm %p %llx\n", realm, realm->ino);
163
Sage Weila105f002010-02-15 14:37:55 -0800164 rb_erase(&realm->node, &mdsc->snap_realms);
Sage Weil963b61e2009-10-06 11:31:12 -0700165
166 if (realm->parent) {
167 list_del_init(&realm->child_item);
168 __put_snap_realm(mdsc, realm->parent);
169 }
170
171 kfree(realm->prior_parent_snaps);
172 kfree(realm->snaps);
173 ceph_put_snap_context(realm->cached_context);
174 kfree(realm);
175}
176
177/*
178 * caller holds snap_rwsem (write)
179 */
180static void __put_snap_realm(struct ceph_mds_client *mdsc,
181 struct ceph_snap_realm *realm)
182{
183 dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
184 atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
185 if (atomic_dec_and_test(&realm->nref))
186 __destroy_snap_realm(mdsc, realm);
187}
188
189/*
190 * caller needn't hold any locks
191 */
192void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
193 struct ceph_snap_realm *realm)
194{
195 dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm,
196 atomic_read(&realm->nref), atomic_read(&realm->nref)-1);
197 if (!atomic_dec_and_test(&realm->nref))
198 return;
199
200 if (down_write_trylock(&mdsc->snap_rwsem)) {
201 __destroy_snap_realm(mdsc, realm);
202 up_write(&mdsc->snap_rwsem);
203 } else {
204 spin_lock(&mdsc->snap_empty_lock);
205 list_add(&mdsc->snap_empty, &realm->empty_item);
206 spin_unlock(&mdsc->snap_empty_lock);
207 }
208}
209
210/*
211 * Clean up any realms whose ref counts have dropped to zero. Note
212 * that this does not include realms who were created but not yet
213 * used.
214 *
215 * Called under snap_rwsem (write)
216 */
217static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
218{
219 struct ceph_snap_realm *realm;
220
221 spin_lock(&mdsc->snap_empty_lock);
222 while (!list_empty(&mdsc->snap_empty)) {
223 realm = list_first_entry(&mdsc->snap_empty,
224 struct ceph_snap_realm, empty_item);
225 list_del(&realm->empty_item);
226 spin_unlock(&mdsc->snap_empty_lock);
227 __destroy_snap_realm(mdsc, realm);
228 spin_lock(&mdsc->snap_empty_lock);
229 }
230 spin_unlock(&mdsc->snap_empty_lock);
231}
232
233void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc)
234{
235 down_write(&mdsc->snap_rwsem);
236 __cleanup_empty_realms(mdsc);
237 up_write(&mdsc->snap_rwsem);
238}
239
240/*
241 * adjust the parent realm of a given @realm. adjust child list, and parent
242 * pointers, and ref counts appropriately.
243 *
244 * return true if parent was changed, 0 if unchanged, <0 on error.
245 *
246 * caller must hold snap_rwsem for write.
247 */
248static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
249 struct ceph_snap_realm *realm,
250 u64 parentino)
251{
252 struct ceph_snap_realm *parent;
253
254 if (realm->parent_ino == parentino)
255 return 0;
256
257 parent = ceph_lookup_snap_realm(mdsc, parentino);
Sage Weil963b61e2009-10-06 11:31:12 -0700258 if (!parent) {
259 parent = ceph_create_snap_realm(mdsc, parentino);
260 if (IS_ERR(parent))
261 return PTR_ERR(parent);
262 }
263 dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n",
264 realm->ino, realm, realm->parent_ino, realm->parent,
265 parentino, parent);
266 if (realm->parent) {
267 list_del_init(&realm->child_item);
268 ceph_put_snap_realm(mdsc, realm->parent);
269 }
270 realm->parent_ino = parentino;
271 realm->parent = parent;
272 ceph_get_snap_realm(mdsc, parent);
273 list_add(&realm->child_item, &parent->children);
274 return 1;
275}
276
277
278static int cmpu64_rev(const void *a, const void *b)
279{
280 if (*(u64 *)a < *(u64 *)b)
281 return 1;
282 if (*(u64 *)a > *(u64 *)b)
283 return -1;
284 return 0;
285}
286
287/*
288 * build the snap context for a given realm.
289 */
290static int build_snap_context(struct ceph_snap_realm *realm)
291{
292 struct ceph_snap_realm *parent = realm->parent;
293 struct ceph_snap_context *snapc;
294 int err = 0;
295 int i;
296 int num = realm->num_prior_parent_snaps + realm->num_snaps;
297
298 /*
299 * build parent context, if it hasn't been built.
300 * conservatively estimate that all parent snaps might be
301 * included by us.
302 */
303 if (parent) {
304 if (!parent->cached_context) {
305 err = build_snap_context(parent);
306 if (err)
307 goto fail;
308 }
309 num += parent->cached_context->num_snaps;
310 }
311
312 /* do i actually need to update? not if my context seq
313 matches realm seq, and my parents' does to. (this works
314 because we rebuild_snap_realms() works _downward_ in
315 hierarchy after each update.) */
316 if (realm->cached_context &&
317 realm->cached_context->seq <= realm->seq &&
318 (!parent ||
319 realm->cached_context->seq <= parent->cached_context->seq)) {
320 dout("build_snap_context %llx %p: %p seq %lld (%d snaps)"
321 " (unchanged)\n",
322 realm->ino, realm, realm->cached_context,
323 realm->cached_context->seq,
324 realm->cached_context->num_snaps);
325 return 0;
326 }
327
328 /* alloc new snap context */
329 err = -ENOMEM;
330 if (num > ULONG_MAX / sizeof(u64) - sizeof(*snapc))
331 goto fail;
332 snapc = kzalloc(sizeof(*snapc) + num*sizeof(u64), GFP_NOFS);
333 if (!snapc)
334 goto fail;
335 atomic_set(&snapc->nref, 1);
336
337 /* build (reverse sorted) snap vector */
338 num = 0;
339 snapc->seq = realm->seq;
340 if (parent) {
341 /* include any of parent's snaps occuring _after_ my
342 parent became my parent */
343 for (i = 0; i < parent->cached_context->num_snaps; i++)
344 if (parent->cached_context->snaps[i] >=
345 realm->parent_since)
346 snapc->snaps[num++] =
347 parent->cached_context->snaps[i];
348 if (parent->cached_context->seq > snapc->seq)
349 snapc->seq = parent->cached_context->seq;
350 }
351 memcpy(snapc->snaps + num, realm->snaps,
352 sizeof(u64)*realm->num_snaps);
353 num += realm->num_snaps;
354 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
355 sizeof(u64)*realm->num_prior_parent_snaps);
356 num += realm->num_prior_parent_snaps;
357
358 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
359 snapc->num_snaps = num;
360 dout("build_snap_context %llx %p: %p seq %lld (%d snaps)\n",
361 realm->ino, realm, snapc, snapc->seq, snapc->num_snaps);
362
363 if (realm->cached_context)
364 ceph_put_snap_context(realm->cached_context);
365 realm->cached_context = snapc;
366 return 0;
367
368fail:
369 /*
370 * if we fail, clear old (incorrect) cached_context... hopefully
371 * we'll have better luck building it later
372 */
373 if (realm->cached_context) {
374 ceph_put_snap_context(realm->cached_context);
375 realm->cached_context = NULL;
376 }
377 pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
378 realm, err);
379 return err;
380}
381
382/*
383 * rebuild snap context for the given realm and all of its children.
384 */
385static void rebuild_snap_realms(struct ceph_snap_realm *realm)
386{
387 struct ceph_snap_realm *child;
388
389 dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
390 build_snap_context(realm);
391
392 list_for_each_entry(child, &realm->children, child_item)
393 rebuild_snap_realms(child);
394}
395
396
397/*
398 * helper to allocate and decode an array of snapids. free prior
399 * instance, if any.
400 */
401static int dup_array(u64 **dst, __le64 *src, int num)
402{
403 int i;
404
405 kfree(*dst);
406 if (num) {
407 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
408 if (!*dst)
409 return -ENOMEM;
410 for (i = 0; i < num; i++)
411 (*dst)[i] = get_unaligned_le64(src + i);
412 } else {
413 *dst = NULL;
414 }
415 return 0;
416}
417
418
419/*
420 * When a snapshot is applied, the size/mtime inode metadata is queued
421 * in a ceph_cap_snap (one for each snapshot) until writeback
422 * completes and the metadata can be flushed back to the MDS.
423 *
424 * However, if a (sync) write is currently in-progress when we apply
425 * the snapshot, we have to wait until the write succeeds or fails
426 * (and a final size/mtime is known). In this case the
427 * cap_snap->writing = 1, and is said to be "pending." When the write
428 * finishes, we __ceph_finish_cap_snap().
429 *
430 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
431 * change).
432 */
433void ceph_queue_cap_snap(struct ceph_inode_info *ci,
434 struct ceph_snap_context *snapc)
435{
436 struct inode *inode = &ci->vfs_inode;
437 struct ceph_cap_snap *capsnap;
438 int used;
439
440 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
441 if (!capsnap) {
442 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
443 return;
444 }
445
446 spin_lock(&inode->i_lock);
447 used = __ceph_caps_used(ci);
448 if (__ceph_have_pending_cap_snap(ci)) {
449 /* there is no point in queuing multiple "pending" cap_snaps,
450 as no new writes are allowed to start when pending, so any
451 writes in progress now were started before the previous
452 cap_snap. lucky us. */
453 dout("queue_cap_snap %p snapc %p seq %llu used %d"
454 " already pending\n", inode, snapc, snapc->seq, used);
455 kfree(capsnap);
456 } else if (ci->i_wrbuffer_ref_head || (used & CEPH_CAP_FILE_WR)) {
457 igrab(inode);
458
459 atomic_set(&capsnap->nref, 1);
460 capsnap->ci = ci;
461 INIT_LIST_HEAD(&capsnap->ci_item);
462 INIT_LIST_HEAD(&capsnap->flushing_item);
463
464 capsnap->follows = snapc->seq - 1;
465 capsnap->context = ceph_get_snap_context(snapc);
466 capsnap->issued = __ceph_caps_issued(ci, NULL);
467 capsnap->dirty = __ceph_caps_dirty(ci);
468
469 capsnap->mode = inode->i_mode;
470 capsnap->uid = inode->i_uid;
471 capsnap->gid = inode->i_gid;
472
473 /* fixme? */
474 capsnap->xattr_blob = NULL;
475 capsnap->xattr_len = 0;
476
477 /* dirty page count moved from _head to this cap_snap;
478 all subsequent writes page dirties occur _after_ this
479 snapshot. */
480 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
481 ci->i_wrbuffer_ref_head = 0;
482 ceph_put_snap_context(ci->i_head_snapc);
483 ci->i_head_snapc = NULL;
484 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
485
486 if (used & CEPH_CAP_FILE_WR) {
487 dout("queue_cap_snap %p cap_snap %p snapc %p"
488 " seq %llu used WR, now pending\n", inode,
489 capsnap, snapc, snapc->seq);
490 capsnap->writing = 1;
491 } else {
492 /* note mtime, size NOW. */
493 __ceph_finish_cap_snap(ci, capsnap);
494 }
495 } else {
496 dout("queue_cap_snap %p nothing dirty|writing\n", inode);
497 kfree(capsnap);
498 }
499
500 spin_unlock(&inode->i_lock);
501}
502
503/*
504 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
505 * to be used for the snapshot, to be flushed back to the mds.
506 *
507 * If capsnap can now be flushed, add to snap_flush list, and return 1.
508 *
509 * Caller must hold i_lock.
510 */
511int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
512 struct ceph_cap_snap *capsnap)
513{
514 struct inode *inode = &ci->vfs_inode;
515 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
516
517 BUG_ON(capsnap->writing);
518 capsnap->size = inode->i_size;
519 capsnap->mtime = inode->i_mtime;
520 capsnap->atime = inode->i_atime;
521 capsnap->ctime = inode->i_ctime;
522 capsnap->time_warp_seq = ci->i_time_warp_seq;
523 if (capsnap->dirty_pages) {
524 dout("finish_cap_snap %p cap_snap %p snapc %p %llu s=%llu "
525 "still has %d dirty pages\n", inode, capsnap,
526 capsnap->context, capsnap->context->seq,
527 capsnap->size, capsnap->dirty_pages);
528 return 0;
529 }
530 dout("finish_cap_snap %p cap_snap %p snapc %p %llu s=%llu clean\n",
531 inode, capsnap, capsnap->context,
532 capsnap->context->seq, capsnap->size);
533
534 spin_lock(&mdsc->snap_flush_lock);
535 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
536 spin_unlock(&mdsc->snap_flush_lock);
537 return 1; /* caller may want to ceph_flush_snaps */
538}
539
540
541/*
542 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
543 * the snap realm parameters from a given realm and all of its ancestors,
544 * up to the root.
545 *
546 * Caller must hold snap_rwsem for write.
547 */
548int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
549 void *p, void *e, bool deletion)
550{
551 struct ceph_mds_snap_realm *ri; /* encoded */
552 __le64 *snaps; /* encoded */
553 __le64 *prior_parent_snaps; /* encoded */
554 struct ceph_snap_realm *realm;
555 int invalidate = 0;
556 int err = -ENOMEM;
557
558 dout("update_snap_trace deletion=%d\n", deletion);
559more:
560 ceph_decode_need(&p, e, sizeof(*ri), bad);
561 ri = p;
562 p += sizeof(*ri);
563 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
564 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
565 snaps = p;
566 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
567 prior_parent_snaps = p;
568 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
569
570 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
Sage Weil963b61e2009-10-06 11:31:12 -0700571 if (!realm) {
572 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
573 if (IS_ERR(realm)) {
574 err = PTR_ERR(realm);
575 goto fail;
576 }
577 }
578
579 if (le64_to_cpu(ri->seq) > realm->seq) {
580 dout("update_snap_trace updating %llx %p %lld -> %lld\n",
581 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
582 /*
583 * if the realm seq has changed, queue a cap_snap for every
584 * inode with open caps. we do this _before_ we update
585 * the realm info so that we prepare for writeback under the
586 * _previous_ snap context.
587 *
588 * ...unless it's a snap deletion!
589 */
590 if (!deletion) {
591 struct ceph_inode_info *ci;
592 struct inode *lastinode = NULL;
593
594 spin_lock(&realm->inodes_with_caps_lock);
595 list_for_each_entry(ci, &realm->inodes_with_caps,
596 i_snap_realm_item) {
597 struct inode *inode = igrab(&ci->vfs_inode);
598 if (!inode)
599 continue;
600 spin_unlock(&realm->inodes_with_caps_lock);
601 if (lastinode)
602 iput(lastinode);
603 lastinode = inode;
604 ceph_queue_cap_snap(ci, realm->cached_context);
605 spin_lock(&realm->inodes_with_caps_lock);
606 }
607 spin_unlock(&realm->inodes_with_caps_lock);
608 if (lastinode)
609 iput(lastinode);
610 dout("update_snap_trace cap_snaps queued\n");
611 }
612
613 } else {
614 dout("update_snap_trace %llx %p seq %lld unchanged\n",
615 realm->ino, realm, realm->seq);
616 }
617
618 /* ensure the parent is correct */
619 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
620 if (err < 0)
621 goto fail;
622 invalidate += err;
623
624 if (le64_to_cpu(ri->seq) > realm->seq) {
625 /* update realm parameters, snap lists */
626 realm->seq = le64_to_cpu(ri->seq);
627 realm->created = le64_to_cpu(ri->created);
628 realm->parent_since = le64_to_cpu(ri->parent_since);
629
630 realm->num_snaps = le32_to_cpu(ri->num_snaps);
631 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
632 if (err < 0)
633 goto fail;
634
635 realm->num_prior_parent_snaps =
636 le32_to_cpu(ri->num_prior_parent_snaps);
637 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
638 realm->num_prior_parent_snaps);
639 if (err < 0)
640 goto fail;
641
642 invalidate = 1;
643 } else if (!realm->cached_context) {
644 invalidate = 1;
645 }
646
647 dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
648 realm, invalidate, p, e);
649
650 if (p < e)
651 goto more;
652
653 /* invalidate when we reach the _end_ (root) of the trace */
654 if (invalidate)
655 rebuild_snap_realms(realm);
656
657 __cleanup_empty_realms(mdsc);
658 return 0;
659
660bad:
661 err = -EINVAL;
662fail:
663 pr_err("update_snap_trace error %d\n", err);
664 return err;
665}
666
667
668/*
669 * Send any cap_snaps that are queued for flush. Try to carry
670 * s_mutex across multiple snap flushes to avoid locking overhead.
671 *
672 * Caller holds no locks.
673 */
674static void flush_snaps(struct ceph_mds_client *mdsc)
675{
676 struct ceph_inode_info *ci;
677 struct inode *inode;
678 struct ceph_mds_session *session = NULL;
679
680 dout("flush_snaps\n");
681 spin_lock(&mdsc->snap_flush_lock);
682 while (!list_empty(&mdsc->snap_flush_list)) {
683 ci = list_first_entry(&mdsc->snap_flush_list,
684 struct ceph_inode_info, i_snap_flush_item);
685 inode = &ci->vfs_inode;
686 igrab(inode);
687 spin_unlock(&mdsc->snap_flush_lock);
688 spin_lock(&inode->i_lock);
689 __ceph_flush_snaps(ci, &session);
690 spin_unlock(&inode->i_lock);
691 iput(inode);
692 spin_lock(&mdsc->snap_flush_lock);
693 }
694 spin_unlock(&mdsc->snap_flush_lock);
695
696 if (session) {
697 mutex_unlock(&session->s_mutex);
698 ceph_put_mds_session(session);
699 }
700 dout("flush_snaps done\n");
701}
702
703
704/*
705 * Handle a snap notification from the MDS.
706 *
707 * This can take two basic forms: the simplest is just a snap creation
708 * or deletion notification on an existing realm. This should update the
709 * realm and its children.
710 *
711 * The more difficult case is realm creation, due to snap creation at a
712 * new point in the file hierarchy, or due to a rename that moves a file or
713 * directory into another realm.
714 */
715void ceph_handle_snap(struct ceph_mds_client *mdsc,
716 struct ceph_msg *msg)
717{
718 struct super_block *sb = mdsc->client->sb;
719 struct ceph_mds_session *session;
720 int mds;
721 u64 split;
722 int op;
723 int trace_len;
724 struct ceph_snap_realm *realm = NULL;
725 void *p = msg->front.iov_base;
726 void *e = p + msg->front.iov_len;
727 struct ceph_mds_snap_head *h;
728 int num_split_inos, num_split_realms;
729 __le64 *split_inos = NULL, *split_realms = NULL;
730 int i;
731 int locked_rwsem = 0;
732
733 if (msg->hdr.src.name.type != CEPH_ENTITY_TYPE_MDS)
734 return;
735 mds = le64_to_cpu(msg->hdr.src.name.num);
736
737 /* decode */
738 if (msg->front.iov_len < sizeof(*h))
739 goto bad;
740 h = p;
741 op = le32_to_cpu(h->op);
742 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
743 * existing realm */
744 num_split_inos = le32_to_cpu(h->num_split_inos);
745 num_split_realms = le32_to_cpu(h->num_split_realms);
746 trace_len = le32_to_cpu(h->trace_len);
747 p += sizeof(*h);
748
749 dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
750 ceph_snap_op_name(op), split, trace_len);
751
752 /* find session */
753 mutex_lock(&mdsc->mutex);
754 session = __ceph_lookup_mds_session(mdsc, mds);
755 mutex_unlock(&mdsc->mutex);
756 if (!session) {
757 dout("WTF, got snap but no session for mds%d\n", mds);
758 return;
759 }
760
761 mutex_lock(&session->s_mutex);
762 session->s_seq++;
763 mutex_unlock(&session->s_mutex);
764
765 down_write(&mdsc->snap_rwsem);
766 locked_rwsem = 1;
767
768 if (op == CEPH_SNAP_OP_SPLIT) {
769 struct ceph_mds_snap_realm *ri;
770
771 /*
772 * A "split" breaks part of an existing realm off into
773 * a new realm. The MDS provides a list of inodes
774 * (with caps) and child realms that belong to the new
775 * child.
776 */
777 split_inos = p;
778 p += sizeof(u64) * num_split_inos;
779 split_realms = p;
780 p += sizeof(u64) * num_split_realms;
781 ceph_decode_need(&p, e, sizeof(*ri), bad);
782 /* we will peek at realm info here, but will _not_
783 * advance p, as the realm update will occur below in
784 * ceph_update_snap_trace. */
785 ri = p;
786
787 realm = ceph_lookup_snap_realm(mdsc, split);
Sage Weil963b61e2009-10-06 11:31:12 -0700788 if (!realm) {
789 realm = ceph_create_snap_realm(mdsc, split);
790 if (IS_ERR(realm))
791 goto out;
792 }
793 ceph_get_snap_realm(mdsc, realm);
794
795 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
796 for (i = 0; i < num_split_inos; i++) {
797 struct ceph_vino vino = {
798 .ino = le64_to_cpu(split_inos[i]),
799 .snap = CEPH_NOSNAP,
800 };
801 struct inode *inode = ceph_find_inode(sb, vino);
802 struct ceph_inode_info *ci;
803
804 if (!inode)
805 continue;
806 ci = ceph_inode(inode);
807
808 spin_lock(&inode->i_lock);
809 if (!ci->i_snap_realm)
810 goto skip_inode;
811 /*
812 * If this inode belongs to a realm that was
813 * created after our new realm, we experienced
814 * a race (due to another split notifications
815 * arriving from a different MDS). So skip
816 * this inode.
817 */
818 if (ci->i_snap_realm->created >
819 le64_to_cpu(ri->created)) {
820 dout(" leaving %p in newer realm %llx %p\n",
821 inode, ci->i_snap_realm->ino,
822 ci->i_snap_realm);
823 goto skip_inode;
824 }
825 dout(" will move %p to split realm %llx %p\n",
826 inode, realm->ino, realm);
827 /*
828 * Remove the inode from the realm's inode
829 * list, but don't add it to the new realm
830 * yet. We don't want the cap_snap to be
831 * queued (again) by ceph_update_snap_trace()
832 * below. Queue it _now_, under the old context.
833 */
834 list_del_init(&ci->i_snap_realm_item);
835 spin_unlock(&inode->i_lock);
836
837 ceph_queue_cap_snap(ci,
838 ci->i_snap_realm->cached_context);
839
840 iput(inode);
841 continue;
842
843skip_inode:
844 spin_unlock(&inode->i_lock);
845 iput(inode);
846 }
847
848 /* we may have taken some of the old realm's children. */
849 for (i = 0; i < num_split_realms; i++) {
850 struct ceph_snap_realm *child =
851 ceph_lookup_snap_realm(mdsc,
852 le64_to_cpu(split_realms[i]));
Sage Weil963b61e2009-10-06 11:31:12 -0700853 if (!child)
854 continue;
855 adjust_snap_realm_parent(mdsc, child, realm->ino);
856 }
857 }
858
859 /*
860 * update using the provided snap trace. if we are deleting a
861 * snap, we can avoid queueing cap_snaps.
862 */
863 ceph_update_snap_trace(mdsc, p, e,
864 op == CEPH_SNAP_OP_DESTROY);
865
866 if (op == CEPH_SNAP_OP_SPLIT) {
867 /*
868 * ok, _now_ add the inodes into the new realm.
869 */
870 for (i = 0; i < num_split_inos; i++) {
871 struct ceph_vino vino = {
872 .ino = le64_to_cpu(split_inos[i]),
873 .snap = CEPH_NOSNAP,
874 };
875 struct inode *inode = ceph_find_inode(sb, vino);
876 struct ceph_inode_info *ci;
877
878 if (!inode)
879 continue;
880 ci = ceph_inode(inode);
881 spin_lock(&inode->i_lock);
882 if (!ci->i_snap_realm)
883 goto split_skip_inode;
884 ceph_put_snap_realm(mdsc, ci->i_snap_realm);
885 spin_lock(&realm->inodes_with_caps_lock);
886 list_add(&ci->i_snap_realm_item,
887 &realm->inodes_with_caps);
888 ci->i_snap_realm = realm;
889 spin_unlock(&realm->inodes_with_caps_lock);
890 ceph_get_snap_realm(mdsc, realm);
891split_skip_inode:
892 spin_unlock(&inode->i_lock);
893 iput(inode);
894 }
895
896 /* we took a reference when we created the realm, above */
897 ceph_put_snap_realm(mdsc, realm);
898 }
899
900 __cleanup_empty_realms(mdsc);
901
902 up_write(&mdsc->snap_rwsem);
903
904 flush_snaps(mdsc);
905 return;
906
907bad:
908 pr_err("corrupt snap message from mds%d\n", mds);
Sage Weil9ec7cab2009-12-14 15:13:47 -0800909 ceph_msg_dump(msg);
Sage Weil963b61e2009-10-06 11:31:12 -0700910out:
911 if (locked_rwsem)
912 up_write(&mdsc->snap_rwsem);
913 return;
914}
915
916
917