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