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 int i;
300 int num = realm->num_prior_parent_snaps + realm->num_snaps;
301
302 /*
303 * build parent context, if it hasn't been built.
304 * conservatively estimate that all parent snaps might be
305 * included by us.
306 */
307 if (parent) {
308 if (!parent->cached_context) {
309 err = build_snap_context(parent);
310 if (err)
311 goto fail;
312 }
313 num += parent->cached_context->num_snaps;
314 }
315
316 /* do i actually need to update? not if my context seq
317 matches realm seq, and my parents' does to. (this works
318 because we rebuild_snap_realms() works _downward_ in
319 hierarchy after each update.) */
320 if (realm->cached_context &&
321 realm->cached_context->seq == realm->seq &&
322 (!parent ||
323 realm->cached_context->seq >= parent->cached_context->seq)) {
324 dout("build_snap_context %llx %p: %p seq %lld (%d snaps)"
325 " (unchanged)\n",
326 realm->ino, realm, realm->cached_context,
327 realm->cached_context->seq,
328 realm->cached_context->num_snaps);
329 return 0;
330 }
331
332 /* alloc new snap context */
333 err = -ENOMEM;
334 if (num > ULONG_MAX / sizeof(u64) - sizeof(*snapc))
335 goto fail;
336 snapc = kzalloc(sizeof(*snapc) + num*sizeof(u64), GFP_NOFS);
337 if (!snapc)
338 goto fail;
339 atomic_set(&snapc->nref, 1);
340
341 /* build (reverse sorted) snap vector */
342 num = 0;
343 snapc->seq = realm->seq;
344 if (parent) {
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 (%d snaps)\n",
365 realm->ino, realm, snapc, snapc->seq, snapc->num_snaps);
366
367 if (realm->cached_context)
368 ceph_put_snap_context(realm->cached_context);
369 realm->cached_context = snapc;
370 return 0;
371
372fail:
373 /*
374 * if we fail, clear old (incorrect) cached_context... hopefully
375 * we'll have better luck building it later
376 */
377 if (realm->cached_context) {
378 ceph_put_snap_context(realm->cached_context);
379 realm->cached_context = NULL;
380 }
381 pr_err("build_snap_context %llx %p fail %d\n", realm->ino,
382 realm, err);
383 return err;
384}
385
386/*
387 * rebuild snap context for the given realm and all of its children.
388 */
389static void rebuild_snap_realms(struct ceph_snap_realm *realm)
390{
391 struct ceph_snap_realm *child;
392
393 dout("rebuild_snap_realms %llx %p\n", realm->ino, realm);
394 build_snap_context(realm);
395
396 list_for_each_entry(child, &realm->children, child_item)
397 rebuild_snap_realms(child);
398}
399
400
401/*
402 * helper to allocate and decode an array of snapids. free prior
403 * instance, if any.
404 */
405static int dup_array(u64 **dst, __le64 *src, int num)
406{
407 int i;
408
409 kfree(*dst);
410 if (num) {
411 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
412 if (!*dst)
413 return -ENOMEM;
414 for (i = 0; i < num; i++)
415 (*dst)[i] = get_unaligned_le64(src + i);
416 } else {
417 *dst = NULL;
418 }
419 return 0;
420}
421
422
423/*
424 * When a snapshot is applied, the size/mtime inode metadata is queued
425 * in a ceph_cap_snap (one for each snapshot) until writeback
426 * completes and the metadata can be flushed back to the MDS.
427 *
428 * However, if a (sync) write is currently in-progress when we apply
429 * the snapshot, we have to wait until the write succeeds or fails
430 * (and a final size/mtime is known). In this case the
431 * cap_snap->writing = 1, and is said to be "pending." When the write
432 * finishes, we __ceph_finish_cap_snap().
433 *
434 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
435 * change).
436 */
437void ceph_queue_cap_snap(struct ceph_inode_info *ci)
438{
439 struct inode *inode = &ci->vfs_inode;
440 struct ceph_cap_snap *capsnap;
441 int used, dirty;
442
443 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
444 if (!capsnap) {
445 pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode);
446 return;
447 }
448
449 spin_lock(&inode->i_lock);
450 used = __ceph_caps_used(ci);
451 dirty = __ceph_caps_dirty(ci);
452
453 /*
454 * If there is a write in progress, treat that as a dirty Fw,
455 * even though it hasn't completed yet; by the time we finish
456 * up this capsnap it will be.
457 */
458 if (used & CEPH_CAP_FILE_WR)
459 dirty |= CEPH_CAP_FILE_WR;
460
461 if (__ceph_have_pending_cap_snap(ci)) {
462 /* there is no point in queuing multiple "pending" cap_snaps,
463 as no new writes are allowed to start when pending, so any
464 writes in progress now were started before the previous
465 cap_snap. lucky us. */
466 dout("queue_cap_snap %p already pending\n", inode);
467 kfree(capsnap);
468 } else if (dirty & (CEPH_CAP_AUTH_EXCL|CEPH_CAP_XATTR_EXCL|
469 CEPH_CAP_FILE_EXCL|CEPH_CAP_FILE_WR)) {
470 struct ceph_snap_context *snapc = ci->i_head_snapc;
471
472 /*
473 * if we are a sync write, we may need to go to the snaprealm
474 * to get the current snapc.
475 */
476 if (!snapc)
477 snapc = ci->i_snap_realm->cached_context;
478
479 dout("queue_cap_snap %p cap_snap %p queuing under %p %s\n",
480 inode, capsnap, snapc, ceph_cap_string(dirty));
481 ihold(inode);
482
483 atomic_set(&capsnap->nref, 1);
484 capsnap->ci = ci;
485 INIT_LIST_HEAD(&capsnap->ci_item);
486 INIT_LIST_HEAD(&capsnap->flushing_item);
487
488 capsnap->follows = snapc->seq;
489 capsnap->issued = __ceph_caps_issued(ci, NULL);
490 capsnap->dirty = dirty;
491
492 capsnap->mode = inode->i_mode;
493 capsnap->uid = inode->i_uid;
494 capsnap->gid = inode->i_gid;
495
496 if (dirty & CEPH_CAP_XATTR_EXCL) {
497 __ceph_build_xattrs_blob(ci);
498 capsnap->xattr_blob =
499 ceph_buffer_get(ci->i_xattrs.blob);
500 capsnap->xattr_version = ci->i_xattrs.version;
501 } else {
502 capsnap->xattr_blob = NULL;
503 capsnap->xattr_version = 0;
504 }
505
506 /* dirty page count moved from _head to this cap_snap;
507 all subsequent writes page dirties occur _after_ this
508 snapshot. */
509 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
510 ci->i_wrbuffer_ref_head = 0;
511 capsnap->context = snapc;
512 ci->i_head_snapc =
513 ceph_get_snap_context(ci->i_snap_realm->cached_context);
514 dout(" new snapc is %p\n", ci->i_head_snapc);
515 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
516
517 if (used & CEPH_CAP_FILE_WR) {
518 dout("queue_cap_snap %p cap_snap %p snapc %p"
519 " seq %llu used WR, now pending\n", inode,
520 capsnap, snapc, snapc->seq);
521 capsnap->writing = 1;
522 } else {
523 /* note mtime, size NOW. */
524 __ceph_finish_cap_snap(ci, capsnap);
525 }
526 } else {
527 dout("queue_cap_snap %p nothing dirty|writing\n", inode);
528 kfree(capsnap);
529 }
530
531 spin_unlock(&inode->i_lock);
532}
533
534/*
535 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
536 * to be used for the snapshot, to be flushed back to the mds.
537 *
538 * If capsnap can now be flushed, add to snap_flush list, and return 1.
539 *
540 * Caller must hold i_lock.
541 */
542int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
543 struct ceph_cap_snap *capsnap)
544{
545 struct inode *inode = &ci->vfs_inode;
546 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
547
548 BUG_ON(capsnap->writing);
549 capsnap->size = inode->i_size;
550 capsnap->mtime = inode->i_mtime;
551 capsnap->atime = inode->i_atime;
552 capsnap->ctime = inode->i_ctime;
553 capsnap->time_warp_seq = ci->i_time_warp_seq;
554 if (capsnap->dirty_pages) {
555 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu "
556 "still has %d dirty pages\n", inode, capsnap,
557 capsnap->context, capsnap->context->seq,
558 ceph_cap_string(capsnap->dirty), capsnap->size,
559 capsnap->dirty_pages);
560 return 0;
561 }
562 dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n",
563 inode, capsnap, capsnap->context,
564 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
565 capsnap->size);
566
567 spin_lock(&mdsc->snap_flush_lock);
568 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
569 spin_unlock(&mdsc->snap_flush_lock);
570 return 1; /* caller may want to ceph_flush_snaps */
571}
572
573/*
574 * Queue cap_snaps for snap writeback for this realm and its children.
575 * Called under snap_rwsem, so realm topology won't change.
576 */
577static void queue_realm_cap_snaps(struct ceph_snap_realm *realm)
578{
579 struct ceph_inode_info *ci;
580 struct inode *lastinode = NULL;
581 struct ceph_snap_realm *child;
582
583 dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino);
584
585 spin_lock(&realm->inodes_with_caps_lock);
586 list_for_each_entry(ci, &realm->inodes_with_caps,
587 i_snap_realm_item) {
588 struct inode *inode = igrab(&ci->vfs_inode);
589 if (!inode)
590 continue;
591 spin_unlock(&realm->inodes_with_caps_lock);
592 if (lastinode)
593 iput(lastinode);
594 lastinode = inode;
595 ceph_queue_cap_snap(ci);
596 spin_lock(&realm->inodes_with_caps_lock);
597 }
598 spin_unlock(&realm->inodes_with_caps_lock);
599 if (lastinode)
600 iput(lastinode);
601
602 list_for_each_entry(child, &realm->children, child_item) {
603 dout("queue_realm_cap_snaps %p %llx queue child %p %llx\n",
604 realm, realm->ino, child, child->ino);
605 list_del_init(&child->dirty_item);
606 list_add(&child->dirty_item, &realm->dirty_item);
607 }
608
609 list_del_init(&realm->dirty_item);
610 dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino);
611}
612
613/*
614 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
615 * the snap realm parameters from a given realm and all of its ancestors,
616 * up to the root.
617 *
618 * Caller must hold snap_rwsem for write.
619 */
620int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
621 void *p, void *e, bool deletion)
622{
623 struct ceph_mds_snap_realm *ri; /* encoded */
624 __le64 *snaps; /* encoded */
625 __le64 *prior_parent_snaps; /* encoded */
626 struct ceph_snap_realm *realm;
627 int invalidate = 0;
628 int err = -ENOMEM;
629 LIST_HEAD(dirty_realms);
630
631 dout("update_snap_trace deletion=%d\n", deletion);
632more:
633 ceph_decode_need(&p, e, sizeof(*ri), bad);
634 ri = p;
635 p += sizeof(*ri);
636 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
637 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
638 snaps = p;
639 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
640 prior_parent_snaps = p;
641 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
642
643 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
644 if (!realm) {
645 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
646 if (IS_ERR(realm)) {
647 err = PTR_ERR(realm);
648 goto fail;
649 }
650 }
651
652 /* ensure the parent is correct */
653 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
654 if (err < 0)
655 goto fail;
656 invalidate += err;
657
658 if (le64_to_cpu(ri->seq) > realm->seq) {
659 dout("update_snap_trace updating %llx %p %lld -> %lld\n",
660 realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
661 /* update realm parameters, snap lists */
662 realm->seq = le64_to_cpu(ri->seq);
663 realm->created = le64_to_cpu(ri->created);
664 realm->parent_since = le64_to_cpu(ri->parent_since);
665
666 realm->num_snaps = le32_to_cpu(ri->num_snaps);
667 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
668 if (err < 0)
669 goto fail;
670
671 realm->num_prior_parent_snaps =
672 le32_to_cpu(ri->num_prior_parent_snaps);
673 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
674 realm->num_prior_parent_snaps);
675 if (err < 0)
676 goto fail;
677
678 /* queue realm for cap_snap creation */
679 list_add(&realm->dirty_item, &dirty_realms);
680
681 invalidate = 1;
682 } else if (!realm->cached_context) {
683 dout("update_snap_trace %llx %p seq %lld new\n",
684 realm->ino, realm, realm->seq);
685 invalidate = 1;
686 } else {
687 dout("update_snap_trace %llx %p seq %lld unchanged\n",
688 realm->ino, realm, realm->seq);
689 }
690
691 dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
692 realm, invalidate, p, e);
693
694 if (p < e)
695 goto more;
696
697 /* invalidate when we reach the _end_ (root) of the trace */
698 if (invalidate)
699 rebuild_snap_realms(realm);
700
701 /*
702 * queue cap snaps _after_ we've built the new snap contexts,
703 * so that i_head_snapc can be set appropriately.
704 */
705 while (!list_empty(&dirty_realms)) {
706 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
707 dirty_item);
708 queue_realm_cap_snaps(realm);
709 }
710
711 __cleanup_empty_realms(mdsc);
712 return 0;
713
714bad:
715 err = -EINVAL;
716fail:
717 pr_err("update_snap_trace error %d\n", err);
718 return err;
719}
720
721
722/*
723 * Send any cap_snaps that are queued for flush. Try to carry
724 * s_mutex across multiple snap flushes to avoid locking overhead.
725 *
726 * Caller holds no locks.
727 */
728static void flush_snaps(struct ceph_mds_client *mdsc)
729{
730 struct ceph_inode_info *ci;
731 struct inode *inode;
732 struct ceph_mds_session *session = NULL;
733
734 dout("flush_snaps\n");
735 spin_lock(&mdsc->snap_flush_lock);
736 while (!list_empty(&mdsc->snap_flush_list)) {
737 ci = list_first_entry(&mdsc->snap_flush_list,
738 struct ceph_inode_info, i_snap_flush_item);
739 inode = &ci->vfs_inode;
740 ihold(inode);
741 spin_unlock(&mdsc->snap_flush_lock);
742 spin_lock(&inode->i_lock);
743 __ceph_flush_snaps(ci, &session, 0);
744 spin_unlock(&inode->i_lock);
745 iput(inode);
746 spin_lock(&mdsc->snap_flush_lock);
747 }
748 spin_unlock(&mdsc->snap_flush_lock);
749
750 if (session) {
751 mutex_unlock(&session->s_mutex);
752 ceph_put_mds_session(session);
753 }
754 dout("flush_snaps done\n");
755}
756
757
758/*
759 * Handle a snap notification from the MDS.
760 *
761 * This can take two basic forms: the simplest is just a snap creation
762 * or deletion notification on an existing realm. This should update the
763 * realm and its children.
764 *
765 * The more difficult case is realm creation, due to snap creation at a
766 * new point in the file hierarchy, or due to a rename that moves a file or
767 * directory into another realm.
768 */
769void ceph_handle_snap(struct ceph_mds_client *mdsc,
770 struct ceph_mds_session *session,
771 struct ceph_msg *msg)
772{
773 struct super_block *sb = mdsc->fsc->sb;
774 int mds = session->s_mds;
775 u64 split;
776 int op;
777 int trace_len;
778 struct ceph_snap_realm *realm = NULL;
779 void *p = msg->front.iov_base;
780 void *e = p + msg->front.iov_len;
781 struct ceph_mds_snap_head *h;
782 int num_split_inos, num_split_realms;
783 __le64 *split_inos = NULL, *split_realms = NULL;
784 int i;
785 int locked_rwsem = 0;
786
787 /* decode */
788 if (msg->front.iov_len < sizeof(*h))
789 goto bad;
790 h = p;
791 op = le32_to_cpu(h->op);
792 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
793 * existing realm */
794 num_split_inos = le32_to_cpu(h->num_split_inos);
795 num_split_realms = le32_to_cpu(h->num_split_realms);
796 trace_len = le32_to_cpu(h->trace_len);
797 p += sizeof(*h);
798
799 dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds,
800 ceph_snap_op_name(op), split, trace_len);
801
802 mutex_lock(&session->s_mutex);
803 session->s_seq++;
804 mutex_unlock(&session->s_mutex);
805
806 down_write(&mdsc->snap_rwsem);
807 locked_rwsem = 1;
808
809 if (op == CEPH_SNAP_OP_SPLIT) {
810 struct ceph_mds_snap_realm *ri;
811
812 /*
813 * A "split" breaks part of an existing realm off into
814 * a new realm. The MDS provides a list of inodes
815 * (with caps) and child realms that belong to the new
816 * child.
817 */
818 split_inos = p;
819 p += sizeof(u64) * num_split_inos;
820 split_realms = p;
821 p += sizeof(u64) * num_split_realms;
822 ceph_decode_need(&p, e, sizeof(*ri), bad);
823 /* we will peek at realm info here, but will _not_
824 * advance p, as the realm update will occur below in
825 * ceph_update_snap_trace. */
826 ri = p;
827
828 realm = ceph_lookup_snap_realm(mdsc, split);
829 if (!realm) {
830 realm = ceph_create_snap_realm(mdsc, split);
831 if (IS_ERR(realm))
832 goto out;
833 }
834 ceph_get_snap_realm(mdsc, realm);
835
836 dout("splitting snap_realm %llx %p\n", realm->ino, realm);
837 for (i = 0; i < num_split_inos; i++) {
838 struct ceph_vino vino = {
839 .ino = le64_to_cpu(split_inos[i]),
840 .snap = CEPH_NOSNAP,
841 };
842 struct inode *inode = ceph_find_inode(sb, vino);
843 struct ceph_inode_info *ci;
844 struct ceph_snap_realm *oldrealm;
845
846 if (!inode)
847 continue;
848 ci = ceph_inode(inode);
849
850 spin_lock(&inode->i_lock);
851 if (!ci->i_snap_realm)
852 goto skip_inode;
853 /*
854 * If this inode belongs to a realm that was
855 * created after our new realm, we experienced
856 * a race (due to another split notifications
857 * arriving from a different MDS). So skip
858 * this inode.
859 */
860 if (ci->i_snap_realm->created >
861 le64_to_cpu(ri->created)) {
862 dout(" leaving %p in newer realm %llx %p\n",
863 inode, ci->i_snap_realm->ino,
864 ci->i_snap_realm);
865 goto skip_inode;
866 }
867 dout(" will move %p to split realm %llx %p\n",
868 inode, realm->ino, realm);
869 /*
870 * Move the inode to the new realm
871 */
872 spin_lock(&realm->inodes_with_caps_lock);
873 list_del_init(&ci->i_snap_realm_item);
874 list_add(&ci->i_snap_realm_item,
875 &realm->inodes_with_caps);
876 oldrealm = ci->i_snap_realm;
877 ci->i_snap_realm = realm;
878 spin_unlock(&realm->inodes_with_caps_lock);
879 spin_unlock(&inode->i_lock);
880
881 ceph_get_snap_realm(mdsc, realm);
882 ceph_put_snap_realm(mdsc, oldrealm);
883
884 iput(inode);
885 continue;
886
887skip_inode:
888 spin_unlock(&inode->i_lock);
889 iput(inode);
890 }
891
892 /* we may have taken some of the old realm's children. */
893 for (i = 0; i < num_split_realms; i++) {
894 struct ceph_snap_realm *child =
895 ceph_lookup_snap_realm(mdsc,
896 le64_to_cpu(split_realms[i]));
897 if (!child)
898 continue;
899 adjust_snap_realm_parent(mdsc, child, realm->ino);
900 }
901 }
902
903 /*
904 * update using the provided snap trace. if we are deleting a
905 * snap, we can avoid queueing cap_snaps.
906 */
907 ceph_update_snap_trace(mdsc, p, e,
908 op == CEPH_SNAP_OP_DESTROY);
909
910 if (op == CEPH_SNAP_OP_SPLIT)
911 /* we took a reference when we created the realm, above */
912 ceph_put_snap_realm(mdsc, realm);
913
914 __cleanup_empty_realms(mdsc);
915
916 up_write(&mdsc->snap_rwsem);
917
918 flush_snaps(mdsc);
919 return;
920
921bad:
922 pr_err("corrupt snap message from mds%d\n", mds);
923 ceph_msg_dump(msg);
924out:
925 if (locked_rwsem)
926 up_write(&mdsc->snap_rwsem);
927 return;
928}
929
930
931
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 doutc(mdsc->fsc->client, "%llx %p\n", 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 ceph_client *cl = mdsc->fsc->client;
154 struct rb_node *n = mdsc->snap_realms.rb_node;
155 struct ceph_snap_realm *r;
156
157 lockdep_assert_held(&mdsc->snap_rwsem);
158
159 while (n) {
160 r = rb_entry(n, struct ceph_snap_realm, node);
161 if (ino < r->ino)
162 n = n->rb_left;
163 else if (ino > r->ino)
164 n = n->rb_right;
165 else {
166 doutc(cl, "%llx %p\n", r->ino, r);
167 return r;
168 }
169 }
170 return NULL;
171}
172
173struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
174 u64 ino)
175{
176 struct ceph_snap_realm *r;
177 r = __lookup_snap_realm(mdsc, ino);
178 if (r)
179 ceph_get_snap_realm(mdsc, r);
180 return r;
181}
182
183static void __put_snap_realm(struct ceph_mds_client *mdsc,
184 struct ceph_snap_realm *realm);
185
186/*
187 * called with snap_rwsem (write)
188 */
189static void __destroy_snap_realm(struct ceph_mds_client *mdsc,
190 struct ceph_snap_realm *realm)
191{
192 struct ceph_client *cl = mdsc->fsc->client;
193 lockdep_assert_held_write(&mdsc->snap_rwsem);
194
195 doutc(cl, "%p %llx\n", realm, realm->ino);
196
197 rb_erase(&realm->node, &mdsc->snap_realms);
198 mdsc->num_snap_realms--;
199
200 if (realm->parent) {
201 list_del_init(&realm->child_item);
202 __put_snap_realm(mdsc, realm->parent);
203 }
204
205 kfree(realm->prior_parent_snaps);
206 kfree(realm->snaps);
207 ceph_put_snap_context(realm->cached_context);
208 kfree(realm);
209}
210
211/*
212 * caller holds snap_rwsem (write)
213 */
214static void __put_snap_realm(struct ceph_mds_client *mdsc,
215 struct ceph_snap_realm *realm)
216{
217 lockdep_assert_held_write(&mdsc->snap_rwsem);
218
219 /*
220 * We do not require the snap_empty_lock here, as any caller that
221 * increments the value must hold the snap_rwsem.
222 */
223 if (atomic_dec_and_test(&realm->nref))
224 __destroy_snap_realm(mdsc, realm);
225}
226
227/*
228 * See comments in ceph_get_snap_realm. Caller needn't hold any locks.
229 */
230void ceph_put_snap_realm(struct ceph_mds_client *mdsc,
231 struct ceph_snap_realm *realm)
232{
233 if (!atomic_dec_and_lock(&realm->nref, &mdsc->snap_empty_lock))
234 return;
235
236 if (down_write_trylock(&mdsc->snap_rwsem)) {
237 spin_unlock(&mdsc->snap_empty_lock);
238 __destroy_snap_realm(mdsc, realm);
239 up_write(&mdsc->snap_rwsem);
240 } else {
241 list_add(&realm->empty_item, &mdsc->snap_empty);
242 spin_unlock(&mdsc->snap_empty_lock);
243 }
244}
245
246/*
247 * Clean up any realms whose ref counts have dropped to zero. Note
248 * that this does not include realms who were created but not yet
249 * used.
250 *
251 * Called under snap_rwsem (write)
252 */
253static void __cleanup_empty_realms(struct ceph_mds_client *mdsc)
254{
255 struct ceph_snap_realm *realm;
256
257 lockdep_assert_held_write(&mdsc->snap_rwsem);
258
259 spin_lock(&mdsc->snap_empty_lock);
260 while (!list_empty(&mdsc->snap_empty)) {
261 realm = list_first_entry(&mdsc->snap_empty,
262 struct ceph_snap_realm, empty_item);
263 list_del(&realm->empty_item);
264 spin_unlock(&mdsc->snap_empty_lock);
265 __destroy_snap_realm(mdsc, realm);
266 spin_lock(&mdsc->snap_empty_lock);
267 }
268 spin_unlock(&mdsc->snap_empty_lock);
269}
270
271void ceph_cleanup_global_and_empty_realms(struct ceph_mds_client *mdsc)
272{
273 struct ceph_snap_realm *global_realm;
274
275 down_write(&mdsc->snap_rwsem);
276 global_realm = __lookup_snap_realm(mdsc, CEPH_INO_GLOBAL_SNAPREALM);
277 if (global_realm)
278 ceph_put_snap_realm(mdsc, global_realm);
279 __cleanup_empty_realms(mdsc);
280 up_write(&mdsc->snap_rwsem);
281}
282
283/*
284 * adjust the parent realm of a given @realm. adjust child list, and parent
285 * pointers, and ref counts appropriately.
286 *
287 * return true if parent was changed, 0 if unchanged, <0 on error.
288 *
289 * caller must hold snap_rwsem for write.
290 */
291static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc,
292 struct ceph_snap_realm *realm,
293 u64 parentino)
294{
295 struct ceph_client *cl = mdsc->fsc->client;
296 struct ceph_snap_realm *parent;
297
298 lockdep_assert_held_write(&mdsc->snap_rwsem);
299
300 if (realm->parent_ino == parentino)
301 return 0;
302
303 parent = ceph_lookup_snap_realm(mdsc, parentino);
304 if (!parent) {
305 parent = ceph_create_snap_realm(mdsc, parentino);
306 if (IS_ERR(parent))
307 return PTR_ERR(parent);
308 }
309 doutc(cl, "%llx %p: %llx %p -> %llx %p\n", realm->ino, realm,
310 realm->parent_ino, realm->parent, parentino, parent);
311 if (realm->parent) {
312 list_del_init(&realm->child_item);
313 ceph_put_snap_realm(mdsc, realm->parent);
314 }
315 realm->parent_ino = parentino;
316 realm->parent = parent;
317 list_add(&realm->child_item, &parent->children);
318 return 1;
319}
320
321
322static int cmpu64_rev(const void *a, const void *b)
323{
324 if (*(u64 *)a < *(u64 *)b)
325 return 1;
326 if (*(u64 *)a > *(u64 *)b)
327 return -1;
328 return 0;
329}
330
331
332/*
333 * build the snap context for a given realm.
334 */
335static int build_snap_context(struct ceph_mds_client *mdsc,
336 struct ceph_snap_realm *realm,
337 struct list_head *realm_queue,
338 struct list_head *dirty_realms)
339{
340 struct ceph_client *cl = mdsc->fsc->client;
341 struct ceph_snap_realm *parent = realm->parent;
342 struct ceph_snap_context *snapc;
343 int err = 0;
344 u32 num = realm->num_prior_parent_snaps + realm->num_snaps;
345
346 /*
347 * build parent context, if it hasn't been built.
348 * conservatively estimate that all parent snaps might be
349 * included by us.
350 */
351 if (parent) {
352 if (!parent->cached_context) {
353 /* add to the queue head */
354 list_add(&parent->rebuild_item, realm_queue);
355 return 1;
356 }
357 num += parent->cached_context->num_snaps;
358 }
359
360 /* do i actually need to update? not if my context seq
361 matches realm seq, and my parents' does to. (this works
362 because we rebuild_snap_realms() works _downward_ in
363 hierarchy after each update.) */
364 if (realm->cached_context &&
365 realm->cached_context->seq == realm->seq &&
366 (!parent ||
367 realm->cached_context->seq >= parent->cached_context->seq)) {
368 doutc(cl, "%llx %p: %p seq %lld (%u snaps) (unchanged)\n",
369 realm->ino, realm, realm->cached_context,
370 realm->cached_context->seq,
371 (unsigned int)realm->cached_context->num_snaps);
372 return 0;
373 }
374
375 /* alloc new snap context */
376 err = -ENOMEM;
377 if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
378 goto fail;
379 snapc = ceph_create_snap_context(num, GFP_NOFS);
380 if (!snapc)
381 goto fail;
382
383 /* build (reverse sorted) snap vector */
384 num = 0;
385 snapc->seq = realm->seq;
386 if (parent) {
387 u32 i;
388
389 /* include any of parent's snaps occurring _after_ my
390 parent became my parent */
391 for (i = 0; i < parent->cached_context->num_snaps; i++)
392 if (parent->cached_context->snaps[i] >=
393 realm->parent_since)
394 snapc->snaps[num++] =
395 parent->cached_context->snaps[i];
396 if (parent->cached_context->seq > snapc->seq)
397 snapc->seq = parent->cached_context->seq;
398 }
399 memcpy(snapc->snaps + num, realm->snaps,
400 sizeof(u64)*realm->num_snaps);
401 num += realm->num_snaps;
402 memcpy(snapc->snaps + num, realm->prior_parent_snaps,
403 sizeof(u64)*realm->num_prior_parent_snaps);
404 num += realm->num_prior_parent_snaps;
405
406 sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL);
407 snapc->num_snaps = num;
408 doutc(cl, "%llx %p: %p seq %lld (%u snaps)\n", realm->ino, realm,
409 snapc, snapc->seq, (unsigned int) snapc->num_snaps);
410
411 ceph_put_snap_context(realm->cached_context);
412 realm->cached_context = snapc;
413 /* queue realm for cap_snap creation */
414 list_add_tail(&realm->dirty_item, dirty_realms);
415 return 0;
416
417fail:
418 /*
419 * if we fail, clear old (incorrect) cached_context... hopefully
420 * we'll have better luck building it later
421 */
422 if (realm->cached_context) {
423 ceph_put_snap_context(realm->cached_context);
424 realm->cached_context = NULL;
425 }
426 pr_err_client(cl, "%llx %p fail %d\n", realm->ino, realm, err);
427 return err;
428}
429
430/*
431 * rebuild snap context for the given realm and all of its children.
432 */
433static void rebuild_snap_realms(struct ceph_mds_client *mdsc,
434 struct ceph_snap_realm *realm,
435 struct list_head *dirty_realms)
436{
437 struct ceph_client *cl = mdsc->fsc->client;
438 LIST_HEAD(realm_queue);
439 int last = 0;
440 bool skip = false;
441
442 list_add_tail(&realm->rebuild_item, &realm_queue);
443
444 while (!list_empty(&realm_queue)) {
445 struct ceph_snap_realm *_realm, *child;
446
447 _realm = list_first_entry(&realm_queue,
448 struct ceph_snap_realm,
449 rebuild_item);
450
451 /*
452 * If the last building failed dues to memory
453 * issue, just empty the realm_queue and return
454 * to avoid infinite loop.
455 */
456 if (last < 0) {
457 list_del_init(&_realm->rebuild_item);
458 continue;
459 }
460
461 last = build_snap_context(mdsc, _realm, &realm_queue,
462 dirty_realms);
463 doutc(cl, "%llx %p, %s\n", realm->ino, realm,
464 last > 0 ? "is deferred" : !last ? "succeeded" : "failed");
465
466 /* is any child in the list ? */
467 list_for_each_entry(child, &_realm->children, child_item) {
468 if (!list_empty(&child->rebuild_item)) {
469 skip = true;
470 break;
471 }
472 }
473
474 if (!skip) {
475 list_for_each_entry(child, &_realm->children, child_item)
476 list_add_tail(&child->rebuild_item, &realm_queue);
477 }
478
479 /* last == 1 means need to build parent first */
480 if (last <= 0)
481 list_del_init(&_realm->rebuild_item);
482 }
483}
484
485
486/*
487 * helper to allocate and decode an array of snapids. free prior
488 * instance, if any.
489 */
490static int dup_array(u64 **dst, __le64 *src, u32 num)
491{
492 u32 i;
493
494 kfree(*dst);
495 if (num) {
496 *dst = kcalloc(num, sizeof(u64), GFP_NOFS);
497 if (!*dst)
498 return -ENOMEM;
499 for (i = 0; i < num; i++)
500 (*dst)[i] = get_unaligned_le64(src + i);
501 } else {
502 *dst = NULL;
503 }
504 return 0;
505}
506
507static bool has_new_snaps(struct ceph_snap_context *o,
508 struct ceph_snap_context *n)
509{
510 if (n->num_snaps == 0)
511 return false;
512 /* snaps are in descending order */
513 return n->snaps[0] > o->seq;
514}
515
516/*
517 * When a snapshot is applied, the size/mtime inode metadata is queued
518 * in a ceph_cap_snap (one for each snapshot) until writeback
519 * completes and the metadata can be flushed back to the MDS.
520 *
521 * However, if a (sync) write is currently in-progress when we apply
522 * the snapshot, we have to wait until the write succeeds or fails
523 * (and a final size/mtime is known). In this case the
524 * cap_snap->writing = 1, and is said to be "pending." When the write
525 * finishes, we __ceph_finish_cap_snap().
526 *
527 * Caller must hold snap_rwsem for read (i.e., the realm topology won't
528 * change).
529 */
530static void ceph_queue_cap_snap(struct ceph_inode_info *ci,
531 struct ceph_cap_snap **pcapsnap)
532{
533 struct inode *inode = &ci->netfs.inode;
534 struct ceph_client *cl = ceph_inode_to_client(inode);
535 struct ceph_snap_context *old_snapc, *new_snapc;
536 struct ceph_cap_snap *capsnap = *pcapsnap;
537 struct ceph_buffer *old_blob = NULL;
538 int used, dirty;
539
540 spin_lock(&ci->i_ceph_lock);
541 used = __ceph_caps_used(ci);
542 dirty = __ceph_caps_dirty(ci);
543
544 old_snapc = ci->i_head_snapc;
545 new_snapc = ci->i_snap_realm->cached_context;
546
547 /*
548 * If there is a write in progress, treat that as a dirty Fw,
549 * even though it hasn't completed yet; by the time we finish
550 * up this capsnap it will be.
551 */
552 if (used & CEPH_CAP_FILE_WR)
553 dirty |= CEPH_CAP_FILE_WR;
554
555 if (__ceph_have_pending_cap_snap(ci)) {
556 /* there is no point in queuing multiple "pending" cap_snaps,
557 as no new writes are allowed to start when pending, so any
558 writes in progress now were started before the previous
559 cap_snap. lucky us. */
560 doutc(cl, "%p %llx.%llx already pending\n", inode,
561 ceph_vinop(inode));
562 goto update_snapc;
563 }
564 if (ci->i_wrbuffer_ref_head == 0 &&
565 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
566 doutc(cl, "%p %llx.%llx nothing dirty|writing\n", inode,
567 ceph_vinop(inode));
568 goto update_snapc;
569 }
570
571 BUG_ON(!old_snapc);
572
573 /*
574 * There is no need to send FLUSHSNAP message to MDS if there is
575 * no new snapshot. But when there is dirty pages or on-going
576 * writes, we still need to create cap_snap. cap_snap is needed
577 * by the write path and page writeback path.
578 *
579 * also see ceph_try_drop_cap_snap()
580 */
581 if (has_new_snaps(old_snapc, new_snapc)) {
582 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
583 capsnap->need_flush = true;
584 } else {
585 if (!(used & CEPH_CAP_FILE_WR) &&
586 ci->i_wrbuffer_ref_head == 0) {
587 doutc(cl, "%p %llx.%llx no new_snap|dirty_page|writing\n",
588 inode, ceph_vinop(inode));
589 goto update_snapc;
590 }
591 }
592
593 doutc(cl, "%p %llx.%llx cap_snap %p queuing under %p %s %s\n",
594 inode, ceph_vinop(inode), capsnap, old_snapc,
595 ceph_cap_string(dirty), capsnap->need_flush ? "" : "no_flush");
596 ihold(inode);
597
598 capsnap->follows = old_snapc->seq;
599 capsnap->issued = __ceph_caps_issued(ci, NULL);
600 capsnap->dirty = dirty;
601
602 capsnap->mode = inode->i_mode;
603 capsnap->uid = inode->i_uid;
604 capsnap->gid = inode->i_gid;
605
606 if (dirty & CEPH_CAP_XATTR_EXCL) {
607 old_blob = __ceph_build_xattrs_blob(ci);
608 capsnap->xattr_blob =
609 ceph_buffer_get(ci->i_xattrs.blob);
610 capsnap->xattr_version = ci->i_xattrs.version;
611 } else {
612 capsnap->xattr_blob = NULL;
613 capsnap->xattr_version = 0;
614 }
615
616 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
617
618 /* dirty page count moved from _head to this cap_snap;
619 all subsequent writes page dirties occur _after_ this
620 snapshot. */
621 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
622 ci->i_wrbuffer_ref_head = 0;
623 capsnap->context = old_snapc;
624 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
625
626 if (used & CEPH_CAP_FILE_WR) {
627 doutc(cl, "%p %llx.%llx cap_snap %p snapc %p seq %llu used WR,"
628 " now pending\n", inode, ceph_vinop(inode), capsnap,
629 old_snapc, old_snapc->seq);
630 capsnap->writing = 1;
631 } else {
632 /* note mtime, size NOW. */
633 __ceph_finish_cap_snap(ci, capsnap);
634 }
635 *pcapsnap = NULL;
636 old_snapc = NULL;
637
638update_snapc:
639 if (ci->i_wrbuffer_ref_head == 0 &&
640 ci->i_wr_ref == 0 &&
641 ci->i_dirty_caps == 0 &&
642 ci->i_flushing_caps == 0) {
643 ci->i_head_snapc = NULL;
644 } else {
645 ci->i_head_snapc = ceph_get_snap_context(new_snapc);
646 doutc(cl, " new snapc is %p\n", new_snapc);
647 }
648 spin_unlock(&ci->i_ceph_lock);
649
650 ceph_buffer_put(old_blob);
651 ceph_put_snap_context(old_snapc);
652}
653
654/*
655 * Finalize the size, mtime for a cap_snap.. that is, settle on final values
656 * to be used for the snapshot, to be flushed back to the mds.
657 *
658 * If capsnap can now be flushed, add to snap_flush list, and return 1.
659 *
660 * Caller must hold i_ceph_lock.
661 */
662int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
663 struct ceph_cap_snap *capsnap)
664{
665 struct inode *inode = &ci->netfs.inode;
666 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
667 struct ceph_client *cl = mdsc->fsc->client;
668
669 BUG_ON(capsnap->writing);
670 capsnap->size = i_size_read(inode);
671 capsnap->mtime = inode_get_mtime(inode);
672 capsnap->atime = inode_get_atime(inode);
673 capsnap->ctime = inode_get_ctime(inode);
674 capsnap->btime = ci->i_btime;
675 capsnap->change_attr = inode_peek_iversion_raw(inode);
676 capsnap->time_warp_seq = ci->i_time_warp_seq;
677 capsnap->truncate_size = ci->i_truncate_size;
678 capsnap->truncate_seq = ci->i_truncate_seq;
679 if (capsnap->dirty_pages) {
680 doutc(cl, "%p %llx.%llx cap_snap %p snapc %p %llu %s "
681 "s=%llu still has %d dirty pages\n", inode,
682 ceph_vinop(inode), capsnap, capsnap->context,
683 capsnap->context->seq,
684 ceph_cap_string(capsnap->dirty),
685 capsnap->size, capsnap->dirty_pages);
686 return 0;
687 }
688
689 /*
690 * Defer flushing the capsnap if the dirty buffer not flushed yet.
691 * And trigger to flush the buffer immediately.
692 */
693 if (ci->i_wrbuffer_ref) {
694 doutc(cl, "%p %llx.%llx cap_snap %p snapc %p %llu %s "
695 "s=%llu used WRBUFFER, delaying\n", inode,
696 ceph_vinop(inode), capsnap, capsnap->context,
697 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
698 capsnap->size);
699 ceph_queue_writeback(inode);
700 return 0;
701 }
702
703 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
704 doutc(cl, "%p %llx.%llx cap_snap %p snapc %p %llu %s s=%llu\n",
705 inode, ceph_vinop(inode), capsnap, capsnap->context,
706 capsnap->context->seq, ceph_cap_string(capsnap->dirty),
707 capsnap->size);
708
709 spin_lock(&mdsc->snap_flush_lock);
710 if (list_empty(&ci->i_snap_flush_item)) {
711 ihold(inode);
712 list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
713 }
714 spin_unlock(&mdsc->snap_flush_lock);
715 return 1; /* caller may want to ceph_flush_snaps */
716}
717
718/*
719 * Queue cap_snaps for snap writeback for this realm and its children.
720 * Called under snap_rwsem, so realm topology won't change.
721 */
722static void queue_realm_cap_snaps(struct ceph_mds_client *mdsc,
723 struct ceph_snap_realm *realm)
724{
725 struct ceph_client *cl = mdsc->fsc->client;
726 struct ceph_inode_info *ci;
727 struct inode *lastinode = NULL;
728 struct ceph_cap_snap *capsnap = NULL;
729
730 doutc(cl, "%p %llx inode\n", realm, realm->ino);
731
732 spin_lock(&realm->inodes_with_caps_lock);
733 list_for_each_entry(ci, &realm->inodes_with_caps, i_snap_realm_item) {
734 struct inode *inode = igrab(&ci->netfs.inode);
735 if (!inode)
736 continue;
737 spin_unlock(&realm->inodes_with_caps_lock);
738 iput(lastinode);
739 lastinode = inode;
740
741 /*
742 * Allocate the capsnap memory outside of ceph_queue_cap_snap()
743 * to reduce very possible but unnecessary frequently memory
744 * allocate/free in this loop.
745 */
746 if (!capsnap) {
747 capsnap = kmem_cache_zalloc(ceph_cap_snap_cachep, GFP_NOFS);
748 if (!capsnap) {
749 pr_err_client(cl,
750 "ENOMEM allocating ceph_cap_snap on %p\n",
751 inode);
752 return;
753 }
754 }
755 capsnap->cap_flush.is_capsnap = true;
756 refcount_set(&capsnap->nref, 1);
757 INIT_LIST_HEAD(&capsnap->cap_flush.i_list);
758 INIT_LIST_HEAD(&capsnap->cap_flush.g_list);
759 INIT_LIST_HEAD(&capsnap->ci_item);
760
761 ceph_queue_cap_snap(ci, &capsnap);
762 spin_lock(&realm->inodes_with_caps_lock);
763 }
764 spin_unlock(&realm->inodes_with_caps_lock);
765 iput(lastinode);
766
767 if (capsnap)
768 kmem_cache_free(ceph_cap_snap_cachep, capsnap);
769 doutc(cl, "%p %llx done\n", realm, realm->ino);
770}
771
772/*
773 * Parse and apply a snapblob "snap trace" from the MDS. This specifies
774 * the snap realm parameters from a given realm and all of its ancestors,
775 * up to the root.
776 *
777 * Caller must hold snap_rwsem for write.
778 */
779int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
780 void *p, void *e, bool deletion,
781 struct ceph_snap_realm **realm_ret)
782{
783 struct ceph_client *cl = mdsc->fsc->client;
784 struct ceph_mds_snap_realm *ri; /* encoded */
785 __le64 *snaps; /* encoded */
786 __le64 *prior_parent_snaps; /* encoded */
787 struct ceph_snap_realm *realm;
788 struct ceph_snap_realm *first_realm = NULL;
789 struct ceph_snap_realm *realm_to_rebuild = NULL;
790 struct ceph_client *client = mdsc->fsc->client;
791 int rebuild_snapcs;
792 int err = -ENOMEM;
793 int ret;
794 LIST_HEAD(dirty_realms);
795
796 lockdep_assert_held_write(&mdsc->snap_rwsem);
797
798 doutc(cl, "deletion=%d\n", deletion);
799more:
800 realm = NULL;
801 rebuild_snapcs = 0;
802 ceph_decode_need(&p, e, sizeof(*ri), bad);
803 ri = p;
804 p += sizeof(*ri);
805 ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) +
806 le32_to_cpu(ri->num_prior_parent_snaps)), bad);
807 snaps = p;
808 p += sizeof(u64) * le32_to_cpu(ri->num_snaps);
809 prior_parent_snaps = p;
810 p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps);
811
812 realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino));
813 if (!realm) {
814 realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino));
815 if (IS_ERR(realm)) {
816 err = PTR_ERR(realm);
817 goto fail;
818 }
819 }
820
821 /* ensure the parent is correct */
822 err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
823 if (err < 0)
824 goto fail;
825 rebuild_snapcs += err;
826
827 if (le64_to_cpu(ri->seq) > realm->seq) {
828 doutc(cl, "updating %llx %p %lld -> %lld\n", realm->ino,
829 realm, realm->seq, le64_to_cpu(ri->seq));
830 /* update realm parameters, snap lists */
831 realm->seq = le64_to_cpu(ri->seq);
832 realm->created = le64_to_cpu(ri->created);
833 realm->parent_since = le64_to_cpu(ri->parent_since);
834
835 realm->num_snaps = le32_to_cpu(ri->num_snaps);
836 err = dup_array(&realm->snaps, snaps, realm->num_snaps);
837 if (err < 0)
838 goto fail;
839
840 realm->num_prior_parent_snaps =
841 le32_to_cpu(ri->num_prior_parent_snaps);
842 err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps,
843 realm->num_prior_parent_snaps);
844 if (err < 0)
845 goto fail;
846
847 if (realm->seq > mdsc->last_snap_seq)
848 mdsc->last_snap_seq = realm->seq;
849
850 rebuild_snapcs = 1;
851 } else if (!realm->cached_context) {
852 doutc(cl, "%llx %p seq %lld new\n", realm->ino, realm,
853 realm->seq);
854 rebuild_snapcs = 1;
855 } else {
856 doutc(cl, "%llx %p seq %lld unchanged\n", realm->ino, realm,
857 realm->seq);
858 }
859
860 doutc(cl, "done with %llx %p, rebuild_snapcs=%d, %p %p\n", realm->ino,
861 realm, rebuild_snapcs, p, e);
862
863 /*
864 * this will always track the uppest parent realm from which
865 * we need to rebuild the snapshot contexts _downward_ in
866 * hierarchy.
867 */
868 if (rebuild_snapcs)
869 realm_to_rebuild = realm;
870
871 /* rebuild_snapcs when we reach the _end_ (root) of the trace */
872 if (realm_to_rebuild && p >= e)
873 rebuild_snap_realms(mdsc, realm_to_rebuild, &dirty_realms);
874
875 if (!first_realm)
876 first_realm = realm;
877 else
878 ceph_put_snap_realm(mdsc, realm);
879
880 if (p < e)
881 goto more;
882
883 /*
884 * queue cap snaps _after_ we've built the new snap contexts,
885 * so that i_head_snapc can be set appropriately.
886 */
887 while (!list_empty(&dirty_realms)) {
888 realm = list_first_entry(&dirty_realms, struct ceph_snap_realm,
889 dirty_item);
890 list_del_init(&realm->dirty_item);
891 queue_realm_cap_snaps(mdsc, realm);
892 }
893
894 if (realm_ret)
895 *realm_ret = first_realm;
896 else
897 ceph_put_snap_realm(mdsc, first_realm);
898
899 __cleanup_empty_realms(mdsc);
900 return 0;
901
902bad:
903 err = -EIO;
904fail:
905 if (realm && !IS_ERR(realm))
906 ceph_put_snap_realm(mdsc, realm);
907 if (first_realm)
908 ceph_put_snap_realm(mdsc, first_realm);
909 pr_err_client(cl, "error %d\n", err);
910
911 /*
912 * When receiving a corrupted snap trace we don't know what
913 * exactly has happened in MDS side. And we shouldn't continue
914 * writing to OSD, which may corrupt the snapshot contents.
915 *
916 * Just try to blocklist this kclient and then this kclient
917 * must be remounted to continue after the corrupted metadata
918 * fixed in the MDS side.
919 */
920 WRITE_ONCE(mdsc->fsc->mount_state, CEPH_MOUNT_FENCE_IO);
921 ret = ceph_monc_blocklist_add(&client->monc, &client->msgr.inst.addr);
922 if (ret)
923 pr_err_client(cl, "failed to blocklist %s: %d\n",
924 ceph_pr_addr(&client->msgr.inst.addr), ret);
925
926 WARN(1, "[client.%lld] %s %s%sdo remount to continue%s",
927 client->monc.auth->global_id, __func__,
928 ret ? "" : ceph_pr_addr(&client->msgr.inst.addr),
929 ret ? "" : " was blocklisted, ",
930 err == -EIO ? " after corrupted snaptrace is fixed" : "");
931
932 return err;
933}
934
935
936/*
937 * Send any cap_snaps that are queued for flush. Try to carry
938 * s_mutex across multiple snap flushes to avoid locking overhead.
939 *
940 * Caller holds no locks.
941 */
942static void flush_snaps(struct ceph_mds_client *mdsc)
943{
944 struct ceph_client *cl = mdsc->fsc->client;
945 struct ceph_inode_info *ci;
946 struct inode *inode;
947 struct ceph_mds_session *session = NULL;
948
949 doutc(cl, "begin\n");
950 spin_lock(&mdsc->snap_flush_lock);
951 while (!list_empty(&mdsc->snap_flush_list)) {
952 ci = list_first_entry(&mdsc->snap_flush_list,
953 struct ceph_inode_info, i_snap_flush_item);
954 inode = &ci->netfs.inode;
955 ihold(inode);
956 spin_unlock(&mdsc->snap_flush_lock);
957 ceph_flush_snaps(ci, &session);
958 iput(inode);
959 spin_lock(&mdsc->snap_flush_lock);
960 }
961 spin_unlock(&mdsc->snap_flush_lock);
962
963 ceph_put_mds_session(session);
964 doutc(cl, "done\n");
965}
966
967/**
968 * ceph_change_snap_realm - change the snap_realm for an inode
969 * @inode: inode to move to new snap realm
970 * @realm: new realm to move inode into (may be NULL)
971 *
972 * Detach an inode from its old snaprealm (if any) and attach it to
973 * the new snaprealm (if any). The old snap realm reference held by
974 * the inode is put. If realm is non-NULL, then the caller's reference
975 * to it is taken over by the inode.
976 */
977void ceph_change_snap_realm(struct inode *inode, struct ceph_snap_realm *realm)
978{
979 struct ceph_inode_info *ci = ceph_inode(inode);
980 struct ceph_mds_client *mdsc = ceph_inode_to_fs_client(inode)->mdsc;
981 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
982
983 lockdep_assert_held(&ci->i_ceph_lock);
984
985 if (oldrealm) {
986 spin_lock(&oldrealm->inodes_with_caps_lock);
987 list_del_init(&ci->i_snap_realm_item);
988 if (oldrealm->ino == ci->i_vino.ino)
989 oldrealm->inode = NULL;
990 spin_unlock(&oldrealm->inodes_with_caps_lock);
991 ceph_put_snap_realm(mdsc, oldrealm);
992 }
993
994 ci->i_snap_realm = realm;
995
996 if (realm) {
997 spin_lock(&realm->inodes_with_caps_lock);
998 list_add(&ci->i_snap_realm_item, &realm->inodes_with_caps);
999 if (realm->ino == ci->i_vino.ino)
1000 realm->inode = inode;
1001 spin_unlock(&realm->inodes_with_caps_lock);
1002 }
1003}
1004
1005/*
1006 * Handle a snap notification from the MDS.
1007 *
1008 * This can take two basic forms: the simplest is just a snap creation
1009 * or deletion notification on an existing realm. This should update the
1010 * realm and its children.
1011 *
1012 * The more difficult case is realm creation, due to snap creation at a
1013 * new point in the file hierarchy, or due to a rename that moves a file or
1014 * directory into another realm.
1015 */
1016void ceph_handle_snap(struct ceph_mds_client *mdsc,
1017 struct ceph_mds_session *session,
1018 struct ceph_msg *msg)
1019{
1020 struct ceph_client *cl = mdsc->fsc->client;
1021 struct super_block *sb = mdsc->fsc->sb;
1022 int mds = session->s_mds;
1023 u64 split;
1024 int op;
1025 int trace_len;
1026 struct ceph_snap_realm *realm = NULL;
1027 void *p = msg->front.iov_base;
1028 void *e = p + msg->front.iov_len;
1029 struct ceph_mds_snap_head *h;
1030 int num_split_inos, num_split_realms;
1031 __le64 *split_inos = NULL, *split_realms = NULL;
1032 int i;
1033 int locked_rwsem = 0;
1034 bool close_sessions = false;
1035
1036 if (!ceph_inc_mds_stopping_blocker(mdsc, session))
1037 return;
1038
1039 /* decode */
1040 if (msg->front.iov_len < sizeof(*h))
1041 goto bad;
1042 h = p;
1043 op = le32_to_cpu(h->op);
1044 split = le64_to_cpu(h->split); /* non-zero if we are splitting an
1045 * existing realm */
1046 num_split_inos = le32_to_cpu(h->num_split_inos);
1047 num_split_realms = le32_to_cpu(h->num_split_realms);
1048 trace_len = le32_to_cpu(h->trace_len);
1049 p += sizeof(*h);
1050
1051 doutc(cl, "from mds%d op %s split %llx tracelen %d\n", mds,
1052 ceph_snap_op_name(op), split, trace_len);
1053
1054 down_write(&mdsc->snap_rwsem);
1055 locked_rwsem = 1;
1056
1057 if (op == CEPH_SNAP_OP_SPLIT) {
1058 struct ceph_mds_snap_realm *ri;
1059
1060 /*
1061 * A "split" breaks part of an existing realm off into
1062 * a new realm. The MDS provides a list of inodes
1063 * (with caps) and child realms that belong to the new
1064 * child.
1065 */
1066 split_inos = p;
1067 p += sizeof(u64) * num_split_inos;
1068 split_realms = p;
1069 p += sizeof(u64) * num_split_realms;
1070 ceph_decode_need(&p, e, sizeof(*ri), bad);
1071 /* we will peek at realm info here, but will _not_
1072 * advance p, as the realm update will occur below in
1073 * ceph_update_snap_trace. */
1074 ri = p;
1075
1076 realm = ceph_lookup_snap_realm(mdsc, split);
1077 if (!realm) {
1078 realm = ceph_create_snap_realm(mdsc, split);
1079 if (IS_ERR(realm))
1080 goto out;
1081 }
1082
1083 doutc(cl, "splitting snap_realm %llx %p\n", realm->ino, realm);
1084 for (i = 0; i < num_split_inos; i++) {
1085 struct ceph_vino vino = {
1086 .ino = le64_to_cpu(split_inos[i]),
1087 .snap = CEPH_NOSNAP,
1088 };
1089 struct inode *inode = ceph_find_inode(sb, vino);
1090 struct ceph_inode_info *ci;
1091
1092 if (!inode)
1093 continue;
1094 ci = ceph_inode(inode);
1095
1096 spin_lock(&ci->i_ceph_lock);
1097 if (!ci->i_snap_realm)
1098 goto skip_inode;
1099 /*
1100 * If this inode belongs to a realm that was
1101 * created after our new realm, we experienced
1102 * a race (due to another split notifications
1103 * arriving from a different MDS). So skip
1104 * this inode.
1105 */
1106 if (ci->i_snap_realm->created >
1107 le64_to_cpu(ri->created)) {
1108 doutc(cl, " leaving %p %llx.%llx in newer realm %llx %p\n",
1109 inode, ceph_vinop(inode), ci->i_snap_realm->ino,
1110 ci->i_snap_realm);
1111 goto skip_inode;
1112 }
1113 doutc(cl, " will move %p %llx.%llx to split realm %llx %p\n",
1114 inode, ceph_vinop(inode), realm->ino, realm);
1115
1116 ceph_get_snap_realm(mdsc, realm);
1117 ceph_change_snap_realm(inode, realm);
1118 spin_unlock(&ci->i_ceph_lock);
1119 iput(inode);
1120 continue;
1121
1122skip_inode:
1123 spin_unlock(&ci->i_ceph_lock);
1124 iput(inode);
1125 }
1126
1127 /* we may have taken some of the old realm's children. */
1128 for (i = 0; i < num_split_realms; i++) {
1129 struct ceph_snap_realm *child =
1130 __lookup_snap_realm(mdsc,
1131 le64_to_cpu(split_realms[i]));
1132 if (!child)
1133 continue;
1134 adjust_snap_realm_parent(mdsc, child, realm->ino);
1135 }
1136 } else {
1137 /*
1138 * In the non-split case both 'num_split_inos' and
1139 * 'num_split_realms' should be 0, making this a no-op.
1140 * However the MDS happens to populate 'split_realms' list
1141 * in one of the UPDATE op cases by mistake.
1142 *
1143 * Skip both lists just in case to ensure that 'p' is
1144 * positioned at the start of realm info, as expected by
1145 * ceph_update_snap_trace().
1146 */
1147 p += sizeof(u64) * num_split_inos;
1148 p += sizeof(u64) * num_split_realms;
1149 }
1150
1151 /*
1152 * update using the provided snap trace. if we are deleting a
1153 * snap, we can avoid queueing cap_snaps.
1154 */
1155 if (ceph_update_snap_trace(mdsc, p, e,
1156 op == CEPH_SNAP_OP_DESTROY,
1157 NULL)) {
1158 close_sessions = true;
1159 goto bad;
1160 }
1161
1162 if (op == CEPH_SNAP_OP_SPLIT)
1163 /* we took a reference when we created the realm, above */
1164 ceph_put_snap_realm(mdsc, realm);
1165
1166 __cleanup_empty_realms(mdsc);
1167
1168 up_write(&mdsc->snap_rwsem);
1169
1170 flush_snaps(mdsc);
1171 ceph_dec_mds_stopping_blocker(mdsc);
1172 return;
1173
1174bad:
1175 pr_err_client(cl, "corrupt snap message from mds%d\n", mds);
1176 ceph_msg_dump(msg);
1177out:
1178 if (locked_rwsem)
1179 up_write(&mdsc->snap_rwsem);
1180
1181 ceph_dec_mds_stopping_blocker(mdsc);
1182
1183 if (close_sessions)
1184 ceph_mdsc_close_sessions(mdsc);
1185 return;
1186}
1187
1188struct ceph_snapid_map* ceph_get_snapid_map(struct ceph_mds_client *mdsc,
1189 u64 snap)
1190{
1191 struct ceph_client *cl = mdsc->fsc->client;
1192 struct ceph_snapid_map *sm, *exist;
1193 struct rb_node **p, *parent;
1194 int ret;
1195
1196 exist = NULL;
1197 spin_lock(&mdsc->snapid_map_lock);
1198 p = &mdsc->snapid_map_tree.rb_node;
1199 while (*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 if (atomic_inc_return(&exist->ref) == 1)
1207 list_del_init(&exist->lru);
1208 break;
1209 }
1210 exist = NULL;
1211 }
1212 spin_unlock(&mdsc->snapid_map_lock);
1213 if (exist) {
1214 doutc(cl, "found snapid map %llx -> %x\n", exist->snap,
1215 exist->dev);
1216 return exist;
1217 }
1218
1219 sm = kmalloc(sizeof(*sm), GFP_NOFS);
1220 if (!sm)
1221 return NULL;
1222
1223 ret = get_anon_bdev(&sm->dev);
1224 if (ret < 0) {
1225 kfree(sm);
1226 return NULL;
1227 }
1228
1229 INIT_LIST_HEAD(&sm->lru);
1230 atomic_set(&sm->ref, 1);
1231 sm->snap = snap;
1232
1233 exist = NULL;
1234 parent = NULL;
1235 p = &mdsc->snapid_map_tree.rb_node;
1236 spin_lock(&mdsc->snapid_map_lock);
1237 while (*p) {
1238 parent = *p;
1239 exist = rb_entry(*p, struct ceph_snapid_map, node);
1240 if (snap > exist->snap)
1241 p = &(*p)->rb_left;
1242 else if (snap < exist->snap)
1243 p = &(*p)->rb_right;
1244 else
1245 break;
1246 exist = NULL;
1247 }
1248 if (exist) {
1249 if (atomic_inc_return(&exist->ref) == 1)
1250 list_del_init(&exist->lru);
1251 } else {
1252 rb_link_node(&sm->node, parent, p);
1253 rb_insert_color(&sm->node, &mdsc->snapid_map_tree);
1254 }
1255 spin_unlock(&mdsc->snapid_map_lock);
1256 if (exist) {
1257 free_anon_bdev(sm->dev);
1258 kfree(sm);
1259 doutc(cl, "found snapid map %llx -> %x\n", exist->snap,
1260 exist->dev);
1261 return exist;
1262 }
1263
1264 doutc(cl, "create snapid map %llx -> %x\n", sm->snap, sm->dev);
1265 return sm;
1266}
1267
1268void ceph_put_snapid_map(struct ceph_mds_client* mdsc,
1269 struct ceph_snapid_map *sm)
1270{
1271 if (!sm)
1272 return;
1273 if (atomic_dec_and_lock(&sm->ref, &mdsc->snapid_map_lock)) {
1274 if (!RB_EMPTY_NODE(&sm->node)) {
1275 sm->last_used = jiffies;
1276 list_add_tail(&sm->lru, &mdsc->snapid_map_lru);
1277 spin_unlock(&mdsc->snapid_map_lock);
1278 } else {
1279 /* already cleaned up by
1280 * ceph_cleanup_snapid_map() */
1281 spin_unlock(&mdsc->snapid_map_lock);
1282 kfree(sm);
1283 }
1284 }
1285}
1286
1287void ceph_trim_snapid_map(struct ceph_mds_client *mdsc)
1288{
1289 struct ceph_client *cl = mdsc->fsc->client;
1290 struct ceph_snapid_map *sm;
1291 unsigned long now;
1292 LIST_HEAD(to_free);
1293
1294 spin_lock(&mdsc->snapid_map_lock);
1295 now = jiffies;
1296
1297 while (!list_empty(&mdsc->snapid_map_lru)) {
1298 sm = list_first_entry(&mdsc->snapid_map_lru,
1299 struct ceph_snapid_map, lru);
1300 if (time_after(sm->last_used + CEPH_SNAPID_MAP_TIMEOUT, now))
1301 break;
1302
1303 rb_erase(&sm->node, &mdsc->snapid_map_tree);
1304 list_move(&sm->lru, &to_free);
1305 }
1306 spin_unlock(&mdsc->snapid_map_lock);
1307
1308 while (!list_empty(&to_free)) {
1309 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1310 list_del(&sm->lru);
1311 doutc(cl, "trim snapid map %llx -> %x\n", sm->snap, sm->dev);
1312 free_anon_bdev(sm->dev);
1313 kfree(sm);
1314 }
1315}
1316
1317void ceph_cleanup_snapid_map(struct ceph_mds_client *mdsc)
1318{
1319 struct ceph_client *cl = mdsc->fsc->client;
1320 struct ceph_snapid_map *sm;
1321 struct rb_node *p;
1322 LIST_HEAD(to_free);
1323
1324 spin_lock(&mdsc->snapid_map_lock);
1325 while ((p = rb_first(&mdsc->snapid_map_tree))) {
1326 sm = rb_entry(p, struct ceph_snapid_map, node);
1327 rb_erase(p, &mdsc->snapid_map_tree);
1328 RB_CLEAR_NODE(p);
1329 list_move(&sm->lru, &to_free);
1330 }
1331 spin_unlock(&mdsc->snapid_map_lock);
1332
1333 while (!list_empty(&to_free)) {
1334 sm = list_first_entry(&to_free, struct ceph_snapid_map, lru);
1335 list_del(&sm->lru);
1336 free_anon_bdev(sm->dev);
1337 if (WARN_ON_ONCE(atomic_read(&sm->ref))) {
1338 pr_err_client(cl, "snapid map %llx -> %x still in use\n",
1339 sm->snap, sm->dev);
1340 }
1341 kfree(sm);
1342 }
1343}