Commit | Line | Data |
---|---|---|
963b61eb SW |
1 | #include "ceph_debug.h" |
2 | ||
963b61eb | 3 | #include <linux/sort.h> |
5a0e3ad6 | 4 | #include <linux/slab.h> |
963b61eb SW |
5 | |
6 | #include "super.h" | |
7 | #include "decode.h" | |
8 | ||
9 | /* | |
10 | * Snapshots in ceph are driven in large part by cooperation from the | |
11 | * client. In contrast to local file systems or file servers that | |
12 | * implement snapshots at a single point in the system, ceph's | |
13 | * distributed access to storage requires clients to help decide | |
14 | * whether a write logically occurs before or after a recently created | |
15 | * snapshot. | |
16 | * | |
17 | * This provides a perfect instantanous client-wide snapshot. Between | |
18 | * clients, however, snapshots may appear to be applied at slightly | |
19 | * different points in time, depending on delays in delivering the | |
20 | * snapshot notification. | |
21 | * | |
22 | * Snapshots are _not_ file system-wide. Instead, each snapshot | |
23 | * applies to the subdirectory nested beneath some directory. This | |
24 | * effectively divides the hierarchy into multiple "realms," where all | |
25 | * of the files contained by each realm share the same set of | |
26 | * snapshots. An individual realm's snap set contains snapshots | |
27 | * explicitly created on that realm, as well as any snaps in its | |
28 | * parent's snap set _after_ the point at which the parent became it's | |
29 | * parent (due to, say, a rename). Similarly, snaps from prior parents | |
30 | * during the time intervals during which they were the parent are included. | |
31 | * | |
32 | * The client is spared most of this detail, fortunately... it must only | |
33 | * maintains a hierarchy of realms reflecting the current parent/child | |
34 | * realm relationship, and for each realm has an explicit list of snaps | |
35 | * inherited from prior parents. | |
36 | * | |
37 | * A snap_realm struct is maintained for realms containing every inode | |
38 | * with an open cap in the system. (The needed snap realm information is | |
39 | * provided by the MDS whenever a cap is issued, i.e., on open.) A 'seq' | |
40 | * version number is used to ensure that as realm parameters change (new | |
41 | * snapshot, new parent, etc.) the client's realm hierarchy is updated. | |
42 | * | |
43 | * The realm hierarchy drives the generation of a 'snap context' for each | |
44 | * realm, which simply lists the resulting set of snaps for the realm. This | |
45 | * is attached to any writes sent to OSDs. | |
46 | */ | |
47 | /* | |
48 | * Unfortunately error handling is a bit mixed here. If we get a snap | |
49 | * update, but don't have enough memory to update our realm hierarchy, | |
50 | * it's not clear what we can do about it (besides complaining to the | |
51 | * console). | |
52 | */ | |
53 | ||
54 | ||
55 | /* | |
56 | * increase ref count for the realm | |
57 | * | |
58 | * caller must hold snap_rwsem for write. | |
59 | */ | |
60 | void ceph_get_snap_realm(struct ceph_mds_client *mdsc, | |
61 | struct ceph_snap_realm *realm) | |
62 | { | |
63 | dout("get_realm %p %d -> %d\n", realm, | |
64 | atomic_read(&realm->nref), atomic_read(&realm->nref)+1); | |
65 | /* | |
66 | * since we _only_ increment realm refs or empty the empty | |
67 | * list with snap_rwsem held, adjusting the empty list here is | |
68 | * safe. we do need to protect against concurrent empty list | |
69 | * additions, however. | |
70 | */ | |
71 | if (atomic_read(&realm->nref) == 0) { | |
72 | spin_lock(&mdsc->snap_empty_lock); | |
73 | list_del_init(&realm->empty_item); | |
74 | spin_unlock(&mdsc->snap_empty_lock); | |
75 | } | |
76 | ||
77 | atomic_inc(&realm->nref); | |
78 | } | |
79 | ||
a105f00c SW |
80 | static 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 | ||
963b61eb SW |
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 | */ | |
107 | static 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 | ||
963b61eb SW |
117 | atomic_set(&realm->nref, 0); /* tree does not take a ref */ |
118 | realm->ino = ino; | |
119 | INIT_LIST_HEAD(&realm->children); | |
120 | INIT_LIST_HEAD(&realm->child_item); | |
121 | INIT_LIST_HEAD(&realm->empty_item); | |
ae00d4f3 | 122 | INIT_LIST_HEAD(&realm->dirty_item); |
963b61eb SW |
123 | INIT_LIST_HEAD(&realm->inodes_with_caps); |
124 | spin_lock_init(&realm->inodes_with_caps_lock); | |
a105f00c | 125 | __insert_snap_realm(&mdsc->snap_realms, realm); |
963b61eb SW |
126 | dout("create_snap_realm %llx %p\n", realm->ino, realm); |
127 | return realm; | |
128 | } | |
129 | ||
130 | /* | |
a105f00c | 131 | * lookup the realm rooted at @ino. |
963b61eb SW |
132 | * |
133 | * caller must hold snap_rwsem for write. | |
134 | */ | |
135 | struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc, | |
136 | u64 ino) | |
137 | { | |
a105f00c SW |
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; | |
963b61eb SW |
153 | } |
154 | ||
155 | static void __put_snap_realm(struct ceph_mds_client *mdsc, | |
156 | struct ceph_snap_realm *realm); | |
157 | ||
158 | /* | |
159 | * called with snap_rwsem (write) | |
160 | */ | |
161 | static void __destroy_snap_realm(struct ceph_mds_client *mdsc, | |
162 | struct ceph_snap_realm *realm) | |
163 | { | |
164 | dout("__destroy_snap_realm %p %llx\n", realm, realm->ino); | |
165 | ||
a105f00c | 166 | rb_erase(&realm->node, &mdsc->snap_realms); |
963b61eb SW |
167 | |
168 | if (realm->parent) { | |
169 | list_del_init(&realm->child_item); | |
170 | __put_snap_realm(mdsc, realm->parent); | |
171 | } | |
172 | ||
173 | kfree(realm->prior_parent_snaps); | |
174 | kfree(realm->snaps); | |
175 | ceph_put_snap_context(realm->cached_context); | |
176 | kfree(realm); | |
177 | } | |
178 | ||
179 | /* | |
180 | * caller holds snap_rwsem (write) | |
181 | */ | |
182 | static void __put_snap_realm(struct ceph_mds_client *mdsc, | |
183 | struct ceph_snap_realm *realm) | |
184 | { | |
185 | dout("__put_snap_realm %llx %p %d -> %d\n", realm->ino, realm, | |
186 | atomic_read(&realm->nref), atomic_read(&realm->nref)-1); | |
187 | if (atomic_dec_and_test(&realm->nref)) | |
188 | __destroy_snap_realm(mdsc, realm); | |
189 | } | |
190 | ||
191 | /* | |
192 | * caller needn't hold any locks | |
193 | */ | |
194 | void ceph_put_snap_realm(struct ceph_mds_client *mdsc, | |
195 | struct ceph_snap_realm *realm) | |
196 | { | |
197 | dout("put_snap_realm %llx %p %d -> %d\n", realm->ino, realm, | |
198 | atomic_read(&realm->nref), atomic_read(&realm->nref)-1); | |
199 | if (!atomic_dec_and_test(&realm->nref)) | |
200 | return; | |
201 | ||
202 | if (down_write_trylock(&mdsc->snap_rwsem)) { | |
203 | __destroy_snap_realm(mdsc, realm); | |
204 | up_write(&mdsc->snap_rwsem); | |
205 | } else { | |
206 | spin_lock(&mdsc->snap_empty_lock); | |
207 | list_add(&mdsc->snap_empty, &realm->empty_item); | |
208 | spin_unlock(&mdsc->snap_empty_lock); | |
209 | } | |
210 | } | |
211 | ||
212 | /* | |
213 | * Clean up any realms whose ref counts have dropped to zero. Note | |
214 | * that this does not include realms who were created but not yet | |
215 | * used. | |
216 | * | |
217 | * Called under snap_rwsem (write) | |
218 | */ | |
219 | static void __cleanup_empty_realms(struct ceph_mds_client *mdsc) | |
220 | { | |
221 | struct ceph_snap_realm *realm; | |
222 | ||
223 | spin_lock(&mdsc->snap_empty_lock); | |
224 | while (!list_empty(&mdsc->snap_empty)) { | |
225 | realm = list_first_entry(&mdsc->snap_empty, | |
226 | struct ceph_snap_realm, empty_item); | |
227 | list_del(&realm->empty_item); | |
228 | spin_unlock(&mdsc->snap_empty_lock); | |
229 | __destroy_snap_realm(mdsc, realm); | |
230 | spin_lock(&mdsc->snap_empty_lock); | |
231 | } | |
232 | spin_unlock(&mdsc->snap_empty_lock); | |
233 | } | |
234 | ||
235 | void ceph_cleanup_empty_realms(struct ceph_mds_client *mdsc) | |
236 | { | |
237 | down_write(&mdsc->snap_rwsem); | |
238 | __cleanup_empty_realms(mdsc); | |
239 | up_write(&mdsc->snap_rwsem); | |
240 | } | |
241 | ||
242 | /* | |
243 | * adjust the parent realm of a given @realm. adjust child list, and parent | |
244 | * pointers, and ref counts appropriately. | |
245 | * | |
246 | * return true if parent was changed, 0 if unchanged, <0 on error. | |
247 | * | |
248 | * caller must hold snap_rwsem for write. | |
249 | */ | |
250 | static int adjust_snap_realm_parent(struct ceph_mds_client *mdsc, | |
251 | struct ceph_snap_realm *realm, | |
252 | u64 parentino) | |
253 | { | |
254 | struct ceph_snap_realm *parent; | |
255 | ||
256 | if (realm->parent_ino == parentino) | |
257 | return 0; | |
258 | ||
259 | parent = ceph_lookup_snap_realm(mdsc, parentino); | |
963b61eb SW |
260 | if (!parent) { |
261 | parent = ceph_create_snap_realm(mdsc, parentino); | |
262 | if (IS_ERR(parent)) | |
263 | return PTR_ERR(parent); | |
264 | } | |
265 | dout("adjust_snap_realm_parent %llx %p: %llx %p -> %llx %p\n", | |
266 | realm->ino, realm, realm->parent_ino, realm->parent, | |
267 | parentino, parent); | |
268 | if (realm->parent) { | |
269 | list_del_init(&realm->child_item); | |
270 | ceph_put_snap_realm(mdsc, realm->parent); | |
271 | } | |
272 | realm->parent_ino = parentino; | |
273 | realm->parent = parent; | |
274 | ceph_get_snap_realm(mdsc, parent); | |
275 | list_add(&realm->child_item, &parent->children); | |
276 | return 1; | |
277 | } | |
278 | ||
279 | ||
280 | static int cmpu64_rev(const void *a, const void *b) | |
281 | { | |
282 | if (*(u64 *)a < *(u64 *)b) | |
283 | return 1; | |
284 | if (*(u64 *)a > *(u64 *)b) | |
285 | return -1; | |
286 | return 0; | |
287 | } | |
288 | ||
289 | /* | |
290 | * build the snap context for a given realm. | |
291 | */ | |
292 | static int build_snap_context(struct ceph_snap_realm *realm) | |
293 | { | |
294 | struct ceph_snap_realm *parent = realm->parent; | |
295 | struct ceph_snap_context *snapc; | |
296 | int err = 0; | |
297 | int i; | |
298 | int num = realm->num_prior_parent_snaps + realm->num_snaps; | |
299 | ||
300 | /* | |
301 | * build parent context, if it hasn't been built. | |
302 | * conservatively estimate that all parent snaps might be | |
303 | * included by us. | |
304 | */ | |
305 | if (parent) { | |
306 | if (!parent->cached_context) { | |
307 | err = build_snap_context(parent); | |
308 | if (err) | |
309 | goto fail; | |
310 | } | |
311 | num += parent->cached_context->num_snaps; | |
312 | } | |
313 | ||
314 | /* do i actually need to update? not if my context seq | |
315 | matches realm seq, and my parents' does to. (this works | |
316 | because we rebuild_snap_realms() works _downward_ in | |
317 | hierarchy after each update.) */ | |
318 | if (realm->cached_context && | |
ec4318bc | 319 | realm->cached_context->seq == realm->seq && |
963b61eb | 320 | (!parent || |
ec4318bc | 321 | realm->cached_context->seq >= parent->cached_context->seq)) { |
963b61eb SW |
322 | dout("build_snap_context %llx %p: %p seq %lld (%d snaps)" |
323 | " (unchanged)\n", | |
324 | realm->ino, realm, realm->cached_context, | |
325 | realm->cached_context->seq, | |
326 | realm->cached_context->num_snaps); | |
327 | return 0; | |
328 | } | |
329 | ||
330 | /* alloc new snap context */ | |
331 | err = -ENOMEM; | |
332 | if (num > ULONG_MAX / sizeof(u64) - sizeof(*snapc)) | |
333 | goto fail; | |
334 | snapc = kzalloc(sizeof(*snapc) + num*sizeof(u64), GFP_NOFS); | |
335 | if (!snapc) | |
336 | goto fail; | |
337 | atomic_set(&snapc->nref, 1); | |
338 | ||
339 | /* build (reverse sorted) snap vector */ | |
340 | num = 0; | |
341 | snapc->seq = realm->seq; | |
342 | if (parent) { | |
343 | /* include any of parent's snaps occuring _after_ my | |
344 | parent became my parent */ | |
345 | for (i = 0; i < parent->cached_context->num_snaps; i++) | |
346 | if (parent->cached_context->snaps[i] >= | |
347 | realm->parent_since) | |
348 | snapc->snaps[num++] = | |
349 | parent->cached_context->snaps[i]; | |
350 | if (parent->cached_context->seq > snapc->seq) | |
351 | snapc->seq = parent->cached_context->seq; | |
352 | } | |
353 | memcpy(snapc->snaps + num, realm->snaps, | |
354 | sizeof(u64)*realm->num_snaps); | |
355 | num += realm->num_snaps; | |
356 | memcpy(snapc->snaps + num, realm->prior_parent_snaps, | |
357 | sizeof(u64)*realm->num_prior_parent_snaps); | |
358 | num += realm->num_prior_parent_snaps; | |
359 | ||
360 | sort(snapc->snaps, num, sizeof(u64), cmpu64_rev, NULL); | |
361 | snapc->num_snaps = num; | |
362 | dout("build_snap_context %llx %p: %p seq %lld (%d snaps)\n", | |
363 | realm->ino, realm, snapc, snapc->seq, snapc->num_snaps); | |
364 | ||
365 | if (realm->cached_context) | |
366 | ceph_put_snap_context(realm->cached_context); | |
367 | realm->cached_context = snapc; | |
368 | return 0; | |
369 | ||
370 | fail: | |
371 | /* | |
372 | * if we fail, clear old (incorrect) cached_context... hopefully | |
373 | * we'll have better luck building it later | |
374 | */ | |
375 | if (realm->cached_context) { | |
376 | ceph_put_snap_context(realm->cached_context); | |
377 | realm->cached_context = NULL; | |
378 | } | |
379 | pr_err("build_snap_context %llx %p fail %d\n", realm->ino, | |
380 | realm, err); | |
381 | return err; | |
382 | } | |
383 | ||
384 | /* | |
385 | * rebuild snap context for the given realm and all of its children. | |
386 | */ | |
387 | static void rebuild_snap_realms(struct ceph_snap_realm *realm) | |
388 | { | |
389 | struct ceph_snap_realm *child; | |
390 | ||
391 | dout("rebuild_snap_realms %llx %p\n", realm->ino, realm); | |
392 | build_snap_context(realm); | |
393 | ||
394 | list_for_each_entry(child, &realm->children, child_item) | |
395 | rebuild_snap_realms(child); | |
396 | } | |
397 | ||
398 | ||
399 | /* | |
400 | * helper to allocate and decode an array of snapids. free prior | |
401 | * instance, if any. | |
402 | */ | |
403 | static int dup_array(u64 **dst, __le64 *src, int num) | |
404 | { | |
405 | int i; | |
406 | ||
407 | kfree(*dst); | |
408 | if (num) { | |
409 | *dst = kcalloc(num, sizeof(u64), GFP_NOFS); | |
410 | if (!*dst) | |
411 | return -ENOMEM; | |
412 | for (i = 0; i < num; i++) | |
413 | (*dst)[i] = get_unaligned_le64(src + i); | |
414 | } else { | |
415 | *dst = NULL; | |
416 | } | |
417 | return 0; | |
418 | } | |
419 | ||
420 | ||
421 | /* | |
422 | * When a snapshot is applied, the size/mtime inode metadata is queued | |
423 | * in a ceph_cap_snap (one for each snapshot) until writeback | |
424 | * completes and the metadata can be flushed back to the MDS. | |
425 | * | |
426 | * However, if a (sync) write is currently in-progress when we apply | |
427 | * the snapshot, we have to wait until the write succeeds or fails | |
428 | * (and a final size/mtime is known). In this case the | |
429 | * cap_snap->writing = 1, and is said to be "pending." When the write | |
430 | * finishes, we __ceph_finish_cap_snap(). | |
431 | * | |
432 | * Caller must hold snap_rwsem for read (i.e., the realm topology won't | |
433 | * change). | |
434 | */ | |
fc837c8f | 435 | void ceph_queue_cap_snap(struct ceph_inode_info *ci) |
963b61eb SW |
436 | { |
437 | struct inode *inode = &ci->vfs_inode; | |
438 | struct ceph_cap_snap *capsnap; | |
4a625be4 | 439 | int used, dirty; |
963b61eb SW |
440 | |
441 | capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS); | |
442 | if (!capsnap) { | |
443 | pr_err("ENOMEM allocating ceph_cap_snap on %p\n", inode); | |
444 | return; | |
445 | } | |
446 | ||
447 | spin_lock(&inode->i_lock); | |
448 | used = __ceph_caps_used(ci); | |
4a625be4 | 449 | dirty = __ceph_caps_dirty(ci); |
963b61eb SW |
450 | if (__ceph_have_pending_cap_snap(ci)) { |
451 | /* there is no point in queuing multiple "pending" cap_snaps, | |
452 | as no new writes are allowed to start when pending, so any | |
453 | writes in progress now were started before the previous | |
454 | cap_snap. lucky us. */ | |
fc837c8f | 455 | dout("queue_cap_snap %p already pending\n", inode); |
963b61eb | 456 | kfree(capsnap); |
4a625be4 SW |
457 | } else if (ci->i_wrbuffer_ref_head || (used & CEPH_CAP_FILE_WR) || |
458 | (dirty & (CEPH_CAP_AUTH_EXCL|CEPH_CAP_XATTR_EXCL| | |
459 | CEPH_CAP_FILE_EXCL|CEPH_CAP_FILE_WR))) { | |
fc837c8f SW |
460 | struct ceph_snap_context *snapc = ci->i_head_snapc; |
461 | ||
7d8cb26d SW |
462 | dout("queue_cap_snap %p cap_snap %p queuing under %p\n", inode, |
463 | capsnap, snapc); | |
963b61eb | 464 | igrab(inode); |
4a625be4 | 465 | |
963b61eb SW |
466 | atomic_set(&capsnap->nref, 1); |
467 | capsnap->ci = ci; | |
468 | INIT_LIST_HEAD(&capsnap->ci_item); | |
469 | INIT_LIST_HEAD(&capsnap->flushing_item); | |
470 | ||
8bef9239 | 471 | capsnap->follows = snapc->seq; |
963b61eb | 472 | capsnap->issued = __ceph_caps_issued(ci, NULL); |
4a625be4 | 473 | capsnap->dirty = dirty; |
963b61eb SW |
474 | |
475 | capsnap->mode = inode->i_mode; | |
476 | capsnap->uid = inode->i_uid; | |
477 | capsnap->gid = inode->i_gid; | |
478 | ||
4a625be4 SW |
479 | if (dirty & CEPH_CAP_XATTR_EXCL) { |
480 | __ceph_build_xattrs_blob(ci); | |
481 | capsnap->xattr_blob = | |
482 | ceph_buffer_get(ci->i_xattrs.blob); | |
483 | capsnap->xattr_version = ci->i_xattrs.version; | |
484 | } else { | |
485 | capsnap->xattr_blob = NULL; | |
486 | capsnap->xattr_version = 0; | |
487 | } | |
963b61eb SW |
488 | |
489 | /* dirty page count moved from _head to this cap_snap; | |
490 | all subsequent writes page dirties occur _after_ this | |
491 | snapshot. */ | |
492 | capsnap->dirty_pages = ci->i_wrbuffer_ref_head; | |
493 | ci->i_wrbuffer_ref_head = 0; | |
fc837c8f | 494 | capsnap->context = snapc; |
7d8cb26d SW |
495 | ci->i_head_snapc = |
496 | ceph_get_snap_context(ci->i_snap_realm->cached_context); | |
497 | dout(" new snapc is %p\n", ci->i_head_snapc); | |
963b61eb SW |
498 | list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps); |
499 | ||
500 | if (used & CEPH_CAP_FILE_WR) { | |
501 | dout("queue_cap_snap %p cap_snap %p snapc %p" | |
502 | " seq %llu used WR, now pending\n", inode, | |
503 | capsnap, snapc, snapc->seq); | |
504 | capsnap->writing = 1; | |
505 | } else { | |
506 | /* note mtime, size NOW. */ | |
507 | __ceph_finish_cap_snap(ci, capsnap); | |
508 | } | |
509 | } else { | |
510 | dout("queue_cap_snap %p nothing dirty|writing\n", inode); | |
511 | kfree(capsnap); | |
512 | } | |
513 | ||
514 | spin_unlock(&inode->i_lock); | |
515 | } | |
516 | ||
517 | /* | |
518 | * Finalize the size, mtime for a cap_snap.. that is, settle on final values | |
519 | * to be used for the snapshot, to be flushed back to the mds. | |
520 | * | |
521 | * If capsnap can now be flushed, add to snap_flush list, and return 1. | |
522 | * | |
523 | * Caller must hold i_lock. | |
524 | */ | |
525 | int __ceph_finish_cap_snap(struct ceph_inode_info *ci, | |
526 | struct ceph_cap_snap *capsnap) | |
527 | { | |
528 | struct inode *inode = &ci->vfs_inode; | |
640ef79d | 529 | struct ceph_mds_client *mdsc = &ceph_sb_to_client(inode->i_sb)->mdsc; |
963b61eb SW |
530 | |
531 | BUG_ON(capsnap->writing); | |
532 | capsnap->size = inode->i_size; | |
533 | capsnap->mtime = inode->i_mtime; | |
534 | capsnap->atime = inode->i_atime; | |
535 | capsnap->ctime = inode->i_ctime; | |
536 | capsnap->time_warp_seq = ci->i_time_warp_seq; | |
537 | if (capsnap->dirty_pages) { | |
819ccbfa | 538 | dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu " |
963b61eb SW |
539 | "still has %d dirty pages\n", inode, capsnap, |
540 | capsnap->context, capsnap->context->seq, | |
819ccbfa SW |
541 | ceph_cap_string(capsnap->dirty), capsnap->size, |
542 | capsnap->dirty_pages); | |
963b61eb SW |
543 | return 0; |
544 | } | |
819ccbfa | 545 | dout("finish_cap_snap %p cap_snap %p snapc %p %llu %s s=%llu\n", |
963b61eb | 546 | inode, capsnap, capsnap->context, |
819ccbfa SW |
547 | capsnap->context->seq, ceph_cap_string(capsnap->dirty), |
548 | capsnap->size); | |
963b61eb SW |
549 | |
550 | spin_lock(&mdsc->snap_flush_lock); | |
551 | list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list); | |
552 | spin_unlock(&mdsc->snap_flush_lock); | |
553 | return 1; /* caller may want to ceph_flush_snaps */ | |
554 | } | |
555 | ||
ed326044 SW |
556 | /* |
557 | * Queue cap_snaps for snap writeback for this realm and its children. | |
558 | * Called under snap_rwsem, so realm topology won't change. | |
559 | */ | |
560 | static void queue_realm_cap_snaps(struct ceph_snap_realm *realm) | |
561 | { | |
562 | struct ceph_inode_info *ci; | |
563 | struct inode *lastinode = NULL; | |
564 | struct ceph_snap_realm *child; | |
565 | ||
566 | dout("queue_realm_cap_snaps %p %llx inodes\n", realm, realm->ino); | |
567 | ||
568 | spin_lock(&realm->inodes_with_caps_lock); | |
569 | list_for_each_entry(ci, &realm->inodes_with_caps, | |
570 | i_snap_realm_item) { | |
571 | struct inode *inode = igrab(&ci->vfs_inode); | |
572 | if (!inode) | |
573 | continue; | |
574 | spin_unlock(&realm->inodes_with_caps_lock); | |
575 | if (lastinode) | |
576 | iput(lastinode); | |
577 | lastinode = inode; | |
578 | ceph_queue_cap_snap(ci); | |
579 | spin_lock(&realm->inodes_with_caps_lock); | |
580 | } | |
581 | spin_unlock(&realm->inodes_with_caps_lock); | |
582 | if (lastinode) | |
583 | iput(lastinode); | |
584 | ||
585 | dout("queue_realm_cap_snaps %p %llx children\n", realm, realm->ino); | |
586 | list_for_each_entry(child, &realm->children, child_item) | |
587 | queue_realm_cap_snaps(child); | |
588 | ||
589 | dout("queue_realm_cap_snaps %p %llx done\n", realm, realm->ino); | |
590 | } | |
963b61eb SW |
591 | |
592 | /* | |
593 | * Parse and apply a snapblob "snap trace" from the MDS. This specifies | |
594 | * the snap realm parameters from a given realm and all of its ancestors, | |
595 | * up to the root. | |
596 | * | |
597 | * Caller must hold snap_rwsem for write. | |
598 | */ | |
599 | int ceph_update_snap_trace(struct ceph_mds_client *mdsc, | |
600 | void *p, void *e, bool deletion) | |
601 | { | |
602 | struct ceph_mds_snap_realm *ri; /* encoded */ | |
603 | __le64 *snaps; /* encoded */ | |
604 | __le64 *prior_parent_snaps; /* encoded */ | |
605 | struct ceph_snap_realm *realm; | |
606 | int invalidate = 0; | |
607 | int err = -ENOMEM; | |
ae00d4f3 | 608 | LIST_HEAD(dirty_realms); |
963b61eb SW |
609 | |
610 | dout("update_snap_trace deletion=%d\n", deletion); | |
611 | more: | |
612 | ceph_decode_need(&p, e, sizeof(*ri), bad); | |
613 | ri = p; | |
614 | p += sizeof(*ri); | |
615 | ceph_decode_need(&p, e, sizeof(u64)*(le32_to_cpu(ri->num_snaps) + | |
616 | le32_to_cpu(ri->num_prior_parent_snaps)), bad); | |
617 | snaps = p; | |
618 | p += sizeof(u64) * le32_to_cpu(ri->num_snaps); | |
619 | prior_parent_snaps = p; | |
620 | p += sizeof(u64) * le32_to_cpu(ri->num_prior_parent_snaps); | |
621 | ||
622 | realm = ceph_lookup_snap_realm(mdsc, le64_to_cpu(ri->ino)); | |
963b61eb SW |
623 | if (!realm) { |
624 | realm = ceph_create_snap_realm(mdsc, le64_to_cpu(ri->ino)); | |
625 | if (IS_ERR(realm)) { | |
626 | err = PTR_ERR(realm); | |
627 | goto fail; | |
628 | } | |
629 | } | |
630 | ||
963b61eb SW |
631 | /* ensure the parent is correct */ |
632 | err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent)); | |
633 | if (err < 0) | |
634 | goto fail; | |
635 | invalidate += err; | |
636 | ||
637 | if (le64_to_cpu(ri->seq) > realm->seq) { | |
ae00d4f3 SW |
638 | dout("update_snap_trace updating %llx %p %lld -> %lld\n", |
639 | realm->ino, realm, realm->seq, le64_to_cpu(ri->seq)); | |
963b61eb SW |
640 | /* update realm parameters, snap lists */ |
641 | realm->seq = le64_to_cpu(ri->seq); | |
642 | realm->created = le64_to_cpu(ri->created); | |
643 | realm->parent_since = le64_to_cpu(ri->parent_since); | |
644 | ||
645 | realm->num_snaps = le32_to_cpu(ri->num_snaps); | |
646 | err = dup_array(&realm->snaps, snaps, realm->num_snaps); | |
647 | if (err < 0) | |
648 | goto fail; | |
649 | ||
650 | realm->num_prior_parent_snaps = | |
651 | le32_to_cpu(ri->num_prior_parent_snaps); | |
652 | err = dup_array(&realm->prior_parent_snaps, prior_parent_snaps, | |
653 | realm->num_prior_parent_snaps); | |
654 | if (err < 0) | |
655 | goto fail; | |
656 | ||
ae00d4f3 SW |
657 | /* queue realm for cap_snap creation */ |
658 | list_add(&realm->dirty_item, &dirty_realms); | |
659 | ||
963b61eb SW |
660 | invalidate = 1; |
661 | } else if (!realm->cached_context) { | |
ae00d4f3 SW |
662 | dout("update_snap_trace %llx %p seq %lld new\n", |
663 | realm->ino, realm, realm->seq); | |
963b61eb | 664 | invalidate = 1; |
ae00d4f3 SW |
665 | } else { |
666 | dout("update_snap_trace %llx %p seq %lld unchanged\n", | |
667 | realm->ino, realm, realm->seq); | |
963b61eb SW |
668 | } |
669 | ||
670 | dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino, | |
671 | realm, invalidate, p, e); | |
672 | ||
673 | if (p < e) | |
674 | goto more; | |
675 | ||
676 | /* invalidate when we reach the _end_ (root) of the trace */ | |
677 | if (invalidate) | |
678 | rebuild_snap_realms(realm); | |
679 | ||
ae00d4f3 SW |
680 | /* |
681 | * queue cap snaps _after_ we've built the new snap contexts, | |
682 | * so that i_head_snapc can be set appropriately. | |
683 | */ | |
684 | list_for_each_entry(realm, &dirty_realms, dirty_item) { | |
685 | queue_realm_cap_snaps(realm); | |
686 | } | |
687 | ||
963b61eb SW |
688 | __cleanup_empty_realms(mdsc); |
689 | return 0; | |
690 | ||
691 | bad: | |
692 | err = -EINVAL; | |
693 | fail: | |
694 | pr_err("update_snap_trace error %d\n", err); | |
695 | return err; | |
696 | } | |
697 | ||
698 | ||
699 | /* | |
700 | * Send any cap_snaps that are queued for flush. Try to carry | |
701 | * s_mutex across multiple snap flushes to avoid locking overhead. | |
702 | * | |
703 | * Caller holds no locks. | |
704 | */ | |
705 | static void flush_snaps(struct ceph_mds_client *mdsc) | |
706 | { | |
707 | struct ceph_inode_info *ci; | |
708 | struct inode *inode; | |
709 | struct ceph_mds_session *session = NULL; | |
710 | ||
711 | dout("flush_snaps\n"); | |
712 | spin_lock(&mdsc->snap_flush_lock); | |
713 | while (!list_empty(&mdsc->snap_flush_list)) { | |
714 | ci = list_first_entry(&mdsc->snap_flush_list, | |
715 | struct ceph_inode_info, i_snap_flush_item); | |
716 | inode = &ci->vfs_inode; | |
717 | igrab(inode); | |
718 | spin_unlock(&mdsc->snap_flush_lock); | |
719 | spin_lock(&inode->i_lock); | |
720 | __ceph_flush_snaps(ci, &session); | |
721 | spin_unlock(&inode->i_lock); | |
722 | iput(inode); | |
723 | spin_lock(&mdsc->snap_flush_lock); | |
724 | } | |
725 | spin_unlock(&mdsc->snap_flush_lock); | |
726 | ||
727 | if (session) { | |
728 | mutex_unlock(&session->s_mutex); | |
729 | ceph_put_mds_session(session); | |
730 | } | |
731 | dout("flush_snaps done\n"); | |
732 | } | |
733 | ||
734 | ||
735 | /* | |
736 | * Handle a snap notification from the MDS. | |
737 | * | |
738 | * This can take two basic forms: the simplest is just a snap creation | |
739 | * or deletion notification on an existing realm. This should update the | |
740 | * realm and its children. | |
741 | * | |
742 | * The more difficult case is realm creation, due to snap creation at a | |
743 | * new point in the file hierarchy, or due to a rename that moves a file or | |
744 | * directory into another realm. | |
745 | */ | |
746 | void ceph_handle_snap(struct ceph_mds_client *mdsc, | |
2600d2dd | 747 | struct ceph_mds_session *session, |
963b61eb SW |
748 | struct ceph_msg *msg) |
749 | { | |
750 | struct super_block *sb = mdsc->client->sb; | |
2600d2dd | 751 | int mds = session->s_mds; |
963b61eb SW |
752 | u64 split; |
753 | int op; | |
754 | int trace_len; | |
755 | struct ceph_snap_realm *realm = NULL; | |
756 | void *p = msg->front.iov_base; | |
757 | void *e = p + msg->front.iov_len; | |
758 | struct ceph_mds_snap_head *h; | |
759 | int num_split_inos, num_split_realms; | |
760 | __le64 *split_inos = NULL, *split_realms = NULL; | |
761 | int i; | |
762 | int locked_rwsem = 0; | |
763 | ||
963b61eb SW |
764 | /* decode */ |
765 | if (msg->front.iov_len < sizeof(*h)) | |
766 | goto bad; | |
767 | h = p; | |
768 | op = le32_to_cpu(h->op); | |
769 | split = le64_to_cpu(h->split); /* non-zero if we are splitting an | |
770 | * existing realm */ | |
771 | num_split_inos = le32_to_cpu(h->num_split_inos); | |
772 | num_split_realms = le32_to_cpu(h->num_split_realms); | |
773 | trace_len = le32_to_cpu(h->trace_len); | |
774 | p += sizeof(*h); | |
775 | ||
776 | dout("handle_snap from mds%d op %s split %llx tracelen %d\n", mds, | |
777 | ceph_snap_op_name(op), split, trace_len); | |
778 | ||
963b61eb SW |
779 | mutex_lock(&session->s_mutex); |
780 | session->s_seq++; | |
781 | mutex_unlock(&session->s_mutex); | |
782 | ||
783 | down_write(&mdsc->snap_rwsem); | |
784 | locked_rwsem = 1; | |
785 | ||
786 | if (op == CEPH_SNAP_OP_SPLIT) { | |
787 | struct ceph_mds_snap_realm *ri; | |
788 | ||
789 | /* | |
790 | * A "split" breaks part of an existing realm off into | |
791 | * a new realm. The MDS provides a list of inodes | |
792 | * (with caps) and child realms that belong to the new | |
793 | * child. | |
794 | */ | |
795 | split_inos = p; | |
796 | p += sizeof(u64) * num_split_inos; | |
797 | split_realms = p; | |
798 | p += sizeof(u64) * num_split_realms; | |
799 | ceph_decode_need(&p, e, sizeof(*ri), bad); | |
800 | /* we will peek at realm info here, but will _not_ | |
801 | * advance p, as the realm update will occur below in | |
802 | * ceph_update_snap_trace. */ | |
803 | ri = p; | |
804 | ||
805 | realm = ceph_lookup_snap_realm(mdsc, split); | |
963b61eb SW |
806 | if (!realm) { |
807 | realm = ceph_create_snap_realm(mdsc, split); | |
808 | if (IS_ERR(realm)) | |
809 | goto out; | |
810 | } | |
811 | ceph_get_snap_realm(mdsc, realm); | |
812 | ||
813 | dout("splitting snap_realm %llx %p\n", realm->ino, realm); | |
814 | for (i = 0; i < num_split_inos; i++) { | |
815 | struct ceph_vino vino = { | |
816 | .ino = le64_to_cpu(split_inos[i]), | |
817 | .snap = CEPH_NOSNAP, | |
818 | }; | |
819 | struct inode *inode = ceph_find_inode(sb, vino); | |
820 | struct ceph_inode_info *ci; | |
ae00d4f3 | 821 | struct ceph_snap_realm *oldrealm; |
963b61eb SW |
822 | |
823 | if (!inode) | |
824 | continue; | |
825 | ci = ceph_inode(inode); | |
826 | ||
827 | spin_lock(&inode->i_lock); | |
828 | if (!ci->i_snap_realm) | |
829 | goto skip_inode; | |
830 | /* | |
831 | * If this inode belongs to a realm that was | |
832 | * created after our new realm, we experienced | |
833 | * a race (due to another split notifications | |
834 | * arriving from a different MDS). So skip | |
835 | * this inode. | |
836 | */ | |
837 | if (ci->i_snap_realm->created > | |
838 | le64_to_cpu(ri->created)) { | |
839 | dout(" leaving %p in newer realm %llx %p\n", | |
840 | inode, ci->i_snap_realm->ino, | |
841 | ci->i_snap_realm); | |
842 | goto skip_inode; | |
843 | } | |
844 | dout(" will move %p to split realm %llx %p\n", | |
845 | inode, realm->ino, realm); | |
846 | /* | |
ae00d4f3 | 847 | * Move the inode to the new realm |
963b61eb | 848 | */ |
052bb34a | 849 | spin_lock(&realm->inodes_with_caps_lock); |
963b61eb | 850 | list_del_init(&ci->i_snap_realm_item); |
ae00d4f3 SW |
851 | list_add(&ci->i_snap_realm_item, |
852 | &realm->inodes_with_caps); | |
853 | oldrealm = ci->i_snap_realm; | |
854 | ci->i_snap_realm = realm; | |
052bb34a | 855 | spin_unlock(&realm->inodes_with_caps_lock); |
963b61eb SW |
856 | spin_unlock(&inode->i_lock); |
857 | ||
ae00d4f3 SW |
858 | ceph_get_snap_realm(mdsc, realm); |
859 | ceph_put_snap_realm(mdsc, oldrealm); | |
963b61eb SW |
860 | |
861 | iput(inode); | |
862 | continue; | |
863 | ||
864 | skip_inode: | |
865 | spin_unlock(&inode->i_lock); | |
866 | iput(inode); | |
867 | } | |
868 | ||
869 | /* we may have taken some of the old realm's children. */ | |
870 | for (i = 0; i < num_split_realms; i++) { | |
871 | struct ceph_snap_realm *child = | |
872 | ceph_lookup_snap_realm(mdsc, | |
873 | le64_to_cpu(split_realms[i])); | |
963b61eb SW |
874 | if (!child) |
875 | continue; | |
876 | adjust_snap_realm_parent(mdsc, child, realm->ino); | |
877 | } | |
878 | } | |
879 | ||
880 | /* | |
881 | * update using the provided snap trace. if we are deleting a | |
882 | * snap, we can avoid queueing cap_snaps. | |
883 | */ | |
884 | ceph_update_snap_trace(mdsc, p, e, | |
885 | op == CEPH_SNAP_OP_DESTROY); | |
886 | ||
ae00d4f3 | 887 | if (op == CEPH_SNAP_OP_SPLIT) |
963b61eb SW |
888 | /* we took a reference when we created the realm, above */ |
889 | ceph_put_snap_realm(mdsc, realm); | |
963b61eb SW |
890 | |
891 | __cleanup_empty_realms(mdsc); | |
892 | ||
893 | up_write(&mdsc->snap_rwsem); | |
894 | ||
895 | flush_snaps(mdsc); | |
896 | return; | |
897 | ||
898 | bad: | |
899 | pr_err("corrupt snap message from mds%d\n", mds); | |
9ec7cab1 | 900 | ceph_msg_dump(msg); |
963b61eb SW |
901 | out: |
902 | if (locked_rwsem) | |
903 | up_write(&mdsc->snap_rwsem); | |
904 | return; | |
905 | } | |
906 | ||
907 | ||
908 |