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