]>
Commit | Line | Data |
---|---|---|
1 | #include <linux/ceph/ceph_debug.h> | |
2 | ||
3 | #include <linux/fs.h> | |
4 | #include <linux/kernel.h> | |
5 | #include <linux/sched.h> | |
6 | #include <linux/slab.h> | |
7 | #include <linux/vmalloc.h> | |
8 | #include <linux/wait.h> | |
9 | #include <linux/writeback.h> | |
10 | ||
11 | #include "super.h" | |
12 | #include "mds_client.h" | |
13 | #include "cache.h" | |
14 | #include <linux/ceph/decode.h> | |
15 | #include <linux/ceph/messenger.h> | |
16 | ||
17 | /* | |
18 | * Capability management | |
19 | * | |
20 | * The Ceph metadata servers control client access to inode metadata | |
21 | * and file data by issuing capabilities, granting clients permission | |
22 | * to read and/or write both inode field and file data to OSDs | |
23 | * (storage nodes). Each capability consists of a set of bits | |
24 | * indicating which operations are allowed. | |
25 | * | |
26 | * If the client holds a *_SHARED cap, the client has a coherent value | |
27 | * that can be safely read from the cached inode. | |
28 | * | |
29 | * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the | |
30 | * client is allowed to change inode attributes (e.g., file size, | |
31 | * mtime), note its dirty state in the ceph_cap, and asynchronously | |
32 | * flush that metadata change to the MDS. | |
33 | * | |
34 | * In the event of a conflicting operation (perhaps by another | |
35 | * client), the MDS will revoke the conflicting client capabilities. | |
36 | * | |
37 | * In order for a client to cache an inode, it must hold a capability | |
38 | * with at least one MDS server. When inodes are released, release | |
39 | * notifications are batched and periodically sent en masse to the MDS | |
40 | * cluster to release server state. | |
41 | */ | |
42 | ||
43 | static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc); | |
44 | static void __kick_flushing_caps(struct ceph_mds_client *mdsc, | |
45 | struct ceph_mds_session *session, | |
46 | struct ceph_inode_info *ci, | |
47 | u64 oldest_flush_tid); | |
48 | ||
49 | /* | |
50 | * Generate readable cap strings for debugging output. | |
51 | */ | |
52 | #define MAX_CAP_STR 20 | |
53 | static char cap_str[MAX_CAP_STR][40]; | |
54 | static DEFINE_SPINLOCK(cap_str_lock); | |
55 | static int last_cap_str; | |
56 | ||
57 | static char *gcap_string(char *s, int c) | |
58 | { | |
59 | if (c & CEPH_CAP_GSHARED) | |
60 | *s++ = 's'; | |
61 | if (c & CEPH_CAP_GEXCL) | |
62 | *s++ = 'x'; | |
63 | if (c & CEPH_CAP_GCACHE) | |
64 | *s++ = 'c'; | |
65 | if (c & CEPH_CAP_GRD) | |
66 | *s++ = 'r'; | |
67 | if (c & CEPH_CAP_GWR) | |
68 | *s++ = 'w'; | |
69 | if (c & CEPH_CAP_GBUFFER) | |
70 | *s++ = 'b'; | |
71 | if (c & CEPH_CAP_GLAZYIO) | |
72 | *s++ = 'l'; | |
73 | return s; | |
74 | } | |
75 | ||
76 | const char *ceph_cap_string(int caps) | |
77 | { | |
78 | int i; | |
79 | char *s; | |
80 | int c; | |
81 | ||
82 | spin_lock(&cap_str_lock); | |
83 | i = last_cap_str++; | |
84 | if (last_cap_str == MAX_CAP_STR) | |
85 | last_cap_str = 0; | |
86 | spin_unlock(&cap_str_lock); | |
87 | ||
88 | s = cap_str[i]; | |
89 | ||
90 | if (caps & CEPH_CAP_PIN) | |
91 | *s++ = 'p'; | |
92 | ||
93 | c = (caps >> CEPH_CAP_SAUTH) & 3; | |
94 | if (c) { | |
95 | *s++ = 'A'; | |
96 | s = gcap_string(s, c); | |
97 | } | |
98 | ||
99 | c = (caps >> CEPH_CAP_SLINK) & 3; | |
100 | if (c) { | |
101 | *s++ = 'L'; | |
102 | s = gcap_string(s, c); | |
103 | } | |
104 | ||
105 | c = (caps >> CEPH_CAP_SXATTR) & 3; | |
106 | if (c) { | |
107 | *s++ = 'X'; | |
108 | s = gcap_string(s, c); | |
109 | } | |
110 | ||
111 | c = caps >> CEPH_CAP_SFILE; | |
112 | if (c) { | |
113 | *s++ = 'F'; | |
114 | s = gcap_string(s, c); | |
115 | } | |
116 | ||
117 | if (s == cap_str[i]) | |
118 | *s++ = '-'; | |
119 | *s = 0; | |
120 | return cap_str[i]; | |
121 | } | |
122 | ||
123 | void ceph_caps_init(struct ceph_mds_client *mdsc) | |
124 | { | |
125 | INIT_LIST_HEAD(&mdsc->caps_list); | |
126 | spin_lock_init(&mdsc->caps_list_lock); | |
127 | } | |
128 | ||
129 | void ceph_caps_finalize(struct ceph_mds_client *mdsc) | |
130 | { | |
131 | struct ceph_cap *cap; | |
132 | ||
133 | spin_lock(&mdsc->caps_list_lock); | |
134 | while (!list_empty(&mdsc->caps_list)) { | |
135 | cap = list_first_entry(&mdsc->caps_list, | |
136 | struct ceph_cap, caps_item); | |
137 | list_del(&cap->caps_item); | |
138 | kmem_cache_free(ceph_cap_cachep, cap); | |
139 | } | |
140 | mdsc->caps_total_count = 0; | |
141 | mdsc->caps_avail_count = 0; | |
142 | mdsc->caps_use_count = 0; | |
143 | mdsc->caps_reserve_count = 0; | |
144 | mdsc->caps_min_count = 0; | |
145 | spin_unlock(&mdsc->caps_list_lock); | |
146 | } | |
147 | ||
148 | void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta) | |
149 | { | |
150 | spin_lock(&mdsc->caps_list_lock); | |
151 | mdsc->caps_min_count += delta; | |
152 | BUG_ON(mdsc->caps_min_count < 0); | |
153 | spin_unlock(&mdsc->caps_list_lock); | |
154 | } | |
155 | ||
156 | void ceph_reserve_caps(struct ceph_mds_client *mdsc, | |
157 | struct ceph_cap_reservation *ctx, int need) | |
158 | { | |
159 | int i; | |
160 | struct ceph_cap *cap; | |
161 | int have; | |
162 | int alloc = 0; | |
163 | LIST_HEAD(newcaps); | |
164 | ||
165 | dout("reserve caps ctx=%p need=%d\n", ctx, need); | |
166 | ||
167 | /* first reserve any caps that are already allocated */ | |
168 | spin_lock(&mdsc->caps_list_lock); | |
169 | if (mdsc->caps_avail_count >= need) | |
170 | have = need; | |
171 | else | |
172 | have = mdsc->caps_avail_count; | |
173 | mdsc->caps_avail_count -= have; | |
174 | mdsc->caps_reserve_count += have; | |
175 | BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + | |
176 | mdsc->caps_reserve_count + | |
177 | mdsc->caps_avail_count); | |
178 | spin_unlock(&mdsc->caps_list_lock); | |
179 | ||
180 | for (i = have; i < need; i++) { | |
181 | cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS); | |
182 | if (!cap) | |
183 | break; | |
184 | list_add(&cap->caps_item, &newcaps); | |
185 | alloc++; | |
186 | } | |
187 | /* we didn't manage to reserve as much as we needed */ | |
188 | if (have + alloc != need) | |
189 | pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n", | |
190 | ctx, need, have + alloc); | |
191 | ||
192 | spin_lock(&mdsc->caps_list_lock); | |
193 | mdsc->caps_total_count += alloc; | |
194 | mdsc->caps_reserve_count += alloc; | |
195 | list_splice(&newcaps, &mdsc->caps_list); | |
196 | ||
197 | BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + | |
198 | mdsc->caps_reserve_count + | |
199 | mdsc->caps_avail_count); | |
200 | spin_unlock(&mdsc->caps_list_lock); | |
201 | ||
202 | ctx->count = need; | |
203 | dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n", | |
204 | ctx, mdsc->caps_total_count, mdsc->caps_use_count, | |
205 | mdsc->caps_reserve_count, mdsc->caps_avail_count); | |
206 | } | |
207 | ||
208 | int ceph_unreserve_caps(struct ceph_mds_client *mdsc, | |
209 | struct ceph_cap_reservation *ctx) | |
210 | { | |
211 | dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count); | |
212 | if (ctx->count) { | |
213 | spin_lock(&mdsc->caps_list_lock); | |
214 | BUG_ON(mdsc->caps_reserve_count < ctx->count); | |
215 | mdsc->caps_reserve_count -= ctx->count; | |
216 | mdsc->caps_avail_count += ctx->count; | |
217 | ctx->count = 0; | |
218 | dout("unreserve caps %d = %d used + %d resv + %d avail\n", | |
219 | mdsc->caps_total_count, mdsc->caps_use_count, | |
220 | mdsc->caps_reserve_count, mdsc->caps_avail_count); | |
221 | BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + | |
222 | mdsc->caps_reserve_count + | |
223 | mdsc->caps_avail_count); | |
224 | spin_unlock(&mdsc->caps_list_lock); | |
225 | } | |
226 | return 0; | |
227 | } | |
228 | ||
229 | struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc, | |
230 | struct ceph_cap_reservation *ctx) | |
231 | { | |
232 | struct ceph_cap *cap = NULL; | |
233 | ||
234 | /* temporary, until we do something about cap import/export */ | |
235 | if (!ctx) { | |
236 | cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS); | |
237 | if (cap) { | |
238 | spin_lock(&mdsc->caps_list_lock); | |
239 | mdsc->caps_use_count++; | |
240 | mdsc->caps_total_count++; | |
241 | spin_unlock(&mdsc->caps_list_lock); | |
242 | } | |
243 | return cap; | |
244 | } | |
245 | ||
246 | spin_lock(&mdsc->caps_list_lock); | |
247 | dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n", | |
248 | ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count, | |
249 | mdsc->caps_reserve_count, mdsc->caps_avail_count); | |
250 | BUG_ON(!ctx->count); | |
251 | BUG_ON(ctx->count > mdsc->caps_reserve_count); | |
252 | BUG_ON(list_empty(&mdsc->caps_list)); | |
253 | ||
254 | ctx->count--; | |
255 | mdsc->caps_reserve_count--; | |
256 | mdsc->caps_use_count++; | |
257 | ||
258 | cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item); | |
259 | list_del(&cap->caps_item); | |
260 | ||
261 | BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + | |
262 | mdsc->caps_reserve_count + mdsc->caps_avail_count); | |
263 | spin_unlock(&mdsc->caps_list_lock); | |
264 | return cap; | |
265 | } | |
266 | ||
267 | void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap) | |
268 | { | |
269 | spin_lock(&mdsc->caps_list_lock); | |
270 | dout("put_cap %p %d = %d used + %d resv + %d avail\n", | |
271 | cap, mdsc->caps_total_count, mdsc->caps_use_count, | |
272 | mdsc->caps_reserve_count, mdsc->caps_avail_count); | |
273 | mdsc->caps_use_count--; | |
274 | /* | |
275 | * Keep some preallocated caps around (ceph_min_count), to | |
276 | * avoid lots of free/alloc churn. | |
277 | */ | |
278 | if (mdsc->caps_avail_count >= mdsc->caps_reserve_count + | |
279 | mdsc->caps_min_count) { | |
280 | mdsc->caps_total_count--; | |
281 | kmem_cache_free(ceph_cap_cachep, cap); | |
282 | } else { | |
283 | mdsc->caps_avail_count++; | |
284 | list_add(&cap->caps_item, &mdsc->caps_list); | |
285 | } | |
286 | ||
287 | BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count + | |
288 | mdsc->caps_reserve_count + mdsc->caps_avail_count); | |
289 | spin_unlock(&mdsc->caps_list_lock); | |
290 | } | |
291 | ||
292 | void ceph_reservation_status(struct ceph_fs_client *fsc, | |
293 | int *total, int *avail, int *used, int *reserved, | |
294 | int *min) | |
295 | { | |
296 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
297 | ||
298 | if (total) | |
299 | *total = mdsc->caps_total_count; | |
300 | if (avail) | |
301 | *avail = mdsc->caps_avail_count; | |
302 | if (used) | |
303 | *used = mdsc->caps_use_count; | |
304 | if (reserved) | |
305 | *reserved = mdsc->caps_reserve_count; | |
306 | if (min) | |
307 | *min = mdsc->caps_min_count; | |
308 | } | |
309 | ||
310 | /* | |
311 | * Find ceph_cap for given mds, if any. | |
312 | * | |
313 | * Called with i_ceph_lock held. | |
314 | */ | |
315 | static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds) | |
316 | { | |
317 | struct ceph_cap *cap; | |
318 | struct rb_node *n = ci->i_caps.rb_node; | |
319 | ||
320 | while (n) { | |
321 | cap = rb_entry(n, struct ceph_cap, ci_node); | |
322 | if (mds < cap->mds) | |
323 | n = n->rb_left; | |
324 | else if (mds > cap->mds) | |
325 | n = n->rb_right; | |
326 | else | |
327 | return cap; | |
328 | } | |
329 | return NULL; | |
330 | } | |
331 | ||
332 | struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds) | |
333 | { | |
334 | struct ceph_cap *cap; | |
335 | ||
336 | spin_lock(&ci->i_ceph_lock); | |
337 | cap = __get_cap_for_mds(ci, mds); | |
338 | spin_unlock(&ci->i_ceph_lock); | |
339 | return cap; | |
340 | } | |
341 | ||
342 | /* | |
343 | * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1. | |
344 | */ | |
345 | static int __ceph_get_cap_mds(struct ceph_inode_info *ci) | |
346 | { | |
347 | struct ceph_cap *cap; | |
348 | int mds = -1; | |
349 | struct rb_node *p; | |
350 | ||
351 | /* prefer mds with WR|BUFFER|EXCL caps */ | |
352 | for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { | |
353 | cap = rb_entry(p, struct ceph_cap, ci_node); | |
354 | mds = cap->mds; | |
355 | if (cap->issued & (CEPH_CAP_FILE_WR | | |
356 | CEPH_CAP_FILE_BUFFER | | |
357 | CEPH_CAP_FILE_EXCL)) | |
358 | break; | |
359 | } | |
360 | return mds; | |
361 | } | |
362 | ||
363 | int ceph_get_cap_mds(struct inode *inode) | |
364 | { | |
365 | struct ceph_inode_info *ci = ceph_inode(inode); | |
366 | int mds; | |
367 | spin_lock(&ci->i_ceph_lock); | |
368 | mds = __ceph_get_cap_mds(ceph_inode(inode)); | |
369 | spin_unlock(&ci->i_ceph_lock); | |
370 | return mds; | |
371 | } | |
372 | ||
373 | /* | |
374 | * Called under i_ceph_lock. | |
375 | */ | |
376 | static void __insert_cap_node(struct ceph_inode_info *ci, | |
377 | struct ceph_cap *new) | |
378 | { | |
379 | struct rb_node **p = &ci->i_caps.rb_node; | |
380 | struct rb_node *parent = NULL; | |
381 | struct ceph_cap *cap = NULL; | |
382 | ||
383 | while (*p) { | |
384 | parent = *p; | |
385 | cap = rb_entry(parent, struct ceph_cap, ci_node); | |
386 | if (new->mds < cap->mds) | |
387 | p = &(*p)->rb_left; | |
388 | else if (new->mds > cap->mds) | |
389 | p = &(*p)->rb_right; | |
390 | else | |
391 | BUG(); | |
392 | } | |
393 | ||
394 | rb_link_node(&new->ci_node, parent, p); | |
395 | rb_insert_color(&new->ci_node, &ci->i_caps); | |
396 | } | |
397 | ||
398 | /* | |
399 | * (re)set cap hold timeouts, which control the delayed release | |
400 | * of unused caps back to the MDS. Should be called on cap use. | |
401 | */ | |
402 | static void __cap_set_timeouts(struct ceph_mds_client *mdsc, | |
403 | struct ceph_inode_info *ci) | |
404 | { | |
405 | struct ceph_mount_options *ma = mdsc->fsc->mount_options; | |
406 | ||
407 | ci->i_hold_caps_min = round_jiffies(jiffies + | |
408 | ma->caps_wanted_delay_min * HZ); | |
409 | ci->i_hold_caps_max = round_jiffies(jiffies + | |
410 | ma->caps_wanted_delay_max * HZ); | |
411 | dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode, | |
412 | ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies); | |
413 | } | |
414 | ||
415 | /* | |
416 | * (Re)queue cap at the end of the delayed cap release list. | |
417 | * | |
418 | * If I_FLUSH is set, leave the inode at the front of the list. | |
419 | * | |
420 | * Caller holds i_ceph_lock | |
421 | * -> we take mdsc->cap_delay_lock | |
422 | */ | |
423 | static void __cap_delay_requeue(struct ceph_mds_client *mdsc, | |
424 | struct ceph_inode_info *ci) | |
425 | { | |
426 | __cap_set_timeouts(mdsc, ci); | |
427 | dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode, | |
428 | ci->i_ceph_flags, ci->i_hold_caps_max); | |
429 | if (!mdsc->stopping) { | |
430 | spin_lock(&mdsc->cap_delay_lock); | |
431 | if (!list_empty(&ci->i_cap_delay_list)) { | |
432 | if (ci->i_ceph_flags & CEPH_I_FLUSH) | |
433 | goto no_change; | |
434 | list_del_init(&ci->i_cap_delay_list); | |
435 | } | |
436 | list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list); | |
437 | no_change: | |
438 | spin_unlock(&mdsc->cap_delay_lock); | |
439 | } | |
440 | } | |
441 | ||
442 | /* | |
443 | * Queue an inode for immediate writeback. Mark inode with I_FLUSH, | |
444 | * indicating we should send a cap message to flush dirty metadata | |
445 | * asap, and move to the front of the delayed cap list. | |
446 | */ | |
447 | static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc, | |
448 | struct ceph_inode_info *ci) | |
449 | { | |
450 | dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode); | |
451 | spin_lock(&mdsc->cap_delay_lock); | |
452 | ci->i_ceph_flags |= CEPH_I_FLUSH; | |
453 | if (!list_empty(&ci->i_cap_delay_list)) | |
454 | list_del_init(&ci->i_cap_delay_list); | |
455 | list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list); | |
456 | spin_unlock(&mdsc->cap_delay_lock); | |
457 | } | |
458 | ||
459 | /* | |
460 | * Cancel delayed work on cap. | |
461 | * | |
462 | * Caller must hold i_ceph_lock. | |
463 | */ | |
464 | static void __cap_delay_cancel(struct ceph_mds_client *mdsc, | |
465 | struct ceph_inode_info *ci) | |
466 | { | |
467 | dout("__cap_delay_cancel %p\n", &ci->vfs_inode); | |
468 | if (list_empty(&ci->i_cap_delay_list)) | |
469 | return; | |
470 | spin_lock(&mdsc->cap_delay_lock); | |
471 | list_del_init(&ci->i_cap_delay_list); | |
472 | spin_unlock(&mdsc->cap_delay_lock); | |
473 | } | |
474 | ||
475 | /* | |
476 | * Common issue checks for add_cap, handle_cap_grant. | |
477 | */ | |
478 | static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap, | |
479 | unsigned issued) | |
480 | { | |
481 | unsigned had = __ceph_caps_issued(ci, NULL); | |
482 | ||
483 | /* | |
484 | * Each time we receive FILE_CACHE anew, we increment | |
485 | * i_rdcache_gen. | |
486 | */ | |
487 | if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) && | |
488 | (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) { | |
489 | ci->i_rdcache_gen++; | |
490 | } | |
491 | ||
492 | /* | |
493 | * if we are newly issued FILE_SHARED, mark dir not complete; we | |
494 | * don't know what happened to this directory while we didn't | |
495 | * have the cap. | |
496 | */ | |
497 | if ((issued & CEPH_CAP_FILE_SHARED) && | |
498 | (had & CEPH_CAP_FILE_SHARED) == 0) { | |
499 | ci->i_shared_gen++; | |
500 | if (S_ISDIR(ci->vfs_inode.i_mode)) { | |
501 | dout(" marking %p NOT complete\n", &ci->vfs_inode); | |
502 | __ceph_dir_clear_complete(ci); | |
503 | } | |
504 | } | |
505 | } | |
506 | ||
507 | /* | |
508 | * Add a capability under the given MDS session. | |
509 | * | |
510 | * Caller should hold session snap_rwsem (read) and s_mutex. | |
511 | * | |
512 | * @fmode is the open file mode, if we are opening a file, otherwise | |
513 | * it is < 0. (This is so we can atomically add the cap and add an | |
514 | * open file reference to it.) | |
515 | */ | |
516 | void ceph_add_cap(struct inode *inode, | |
517 | struct ceph_mds_session *session, u64 cap_id, | |
518 | int fmode, unsigned issued, unsigned wanted, | |
519 | unsigned seq, unsigned mseq, u64 realmino, int flags, | |
520 | struct ceph_cap **new_cap) | |
521 | { | |
522 | struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; | |
523 | struct ceph_inode_info *ci = ceph_inode(inode); | |
524 | struct ceph_cap *cap; | |
525 | int mds = session->s_mds; | |
526 | int actual_wanted; | |
527 | ||
528 | dout("add_cap %p mds%d cap %llx %s seq %d\n", inode, | |
529 | session->s_mds, cap_id, ceph_cap_string(issued), seq); | |
530 | ||
531 | /* | |
532 | * If we are opening the file, include file mode wanted bits | |
533 | * in wanted. | |
534 | */ | |
535 | if (fmode >= 0) | |
536 | wanted |= ceph_caps_for_mode(fmode); | |
537 | ||
538 | cap = __get_cap_for_mds(ci, mds); | |
539 | if (!cap) { | |
540 | cap = *new_cap; | |
541 | *new_cap = NULL; | |
542 | ||
543 | cap->issued = 0; | |
544 | cap->implemented = 0; | |
545 | cap->mds = mds; | |
546 | cap->mds_wanted = 0; | |
547 | cap->mseq = 0; | |
548 | ||
549 | cap->ci = ci; | |
550 | __insert_cap_node(ci, cap); | |
551 | ||
552 | /* add to session cap list */ | |
553 | cap->session = session; | |
554 | spin_lock(&session->s_cap_lock); | |
555 | list_add_tail(&cap->session_caps, &session->s_caps); | |
556 | session->s_nr_caps++; | |
557 | spin_unlock(&session->s_cap_lock); | |
558 | } else { | |
559 | /* | |
560 | * auth mds of the inode changed. we received the cap export | |
561 | * message, but still haven't received the cap import message. | |
562 | * handle_cap_export() updated the new auth MDS' cap. | |
563 | * | |
564 | * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing | |
565 | * a message that was send before the cap import message. So | |
566 | * don't remove caps. | |
567 | */ | |
568 | if (ceph_seq_cmp(seq, cap->seq) <= 0) { | |
569 | WARN_ON(cap != ci->i_auth_cap); | |
570 | WARN_ON(cap->cap_id != cap_id); | |
571 | seq = cap->seq; | |
572 | mseq = cap->mseq; | |
573 | issued |= cap->issued; | |
574 | flags |= CEPH_CAP_FLAG_AUTH; | |
575 | } | |
576 | } | |
577 | ||
578 | if (!ci->i_snap_realm) { | |
579 | /* | |
580 | * add this inode to the appropriate snap realm | |
581 | */ | |
582 | struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc, | |
583 | realmino); | |
584 | if (realm) { | |
585 | spin_lock(&realm->inodes_with_caps_lock); | |
586 | ci->i_snap_realm = realm; | |
587 | list_add(&ci->i_snap_realm_item, | |
588 | &realm->inodes_with_caps); | |
589 | spin_unlock(&realm->inodes_with_caps_lock); | |
590 | } else { | |
591 | pr_err("ceph_add_cap: couldn't find snap realm %llx\n", | |
592 | realmino); | |
593 | WARN_ON(!realm); | |
594 | } | |
595 | } | |
596 | ||
597 | __check_cap_issue(ci, cap, issued); | |
598 | ||
599 | /* | |
600 | * If we are issued caps we don't want, or the mds' wanted | |
601 | * value appears to be off, queue a check so we'll release | |
602 | * later and/or update the mds wanted value. | |
603 | */ | |
604 | actual_wanted = __ceph_caps_wanted(ci); | |
605 | if ((wanted & ~actual_wanted) || | |
606 | (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) { | |
607 | dout(" issued %s, mds wanted %s, actual %s, queueing\n", | |
608 | ceph_cap_string(issued), ceph_cap_string(wanted), | |
609 | ceph_cap_string(actual_wanted)); | |
610 | __cap_delay_requeue(mdsc, ci); | |
611 | } | |
612 | ||
613 | if (flags & CEPH_CAP_FLAG_AUTH) { | |
614 | if (ci->i_auth_cap == NULL || | |
615 | ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) { | |
616 | ci->i_auth_cap = cap; | |
617 | cap->mds_wanted = wanted; | |
618 | } | |
619 | } else { | |
620 | WARN_ON(ci->i_auth_cap == cap); | |
621 | } | |
622 | ||
623 | dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n", | |
624 | inode, ceph_vinop(inode), cap, ceph_cap_string(issued), | |
625 | ceph_cap_string(issued|cap->issued), seq, mds); | |
626 | cap->cap_id = cap_id; | |
627 | cap->issued = issued; | |
628 | cap->implemented |= issued; | |
629 | if (ceph_seq_cmp(mseq, cap->mseq) > 0) | |
630 | cap->mds_wanted = wanted; | |
631 | else | |
632 | cap->mds_wanted |= wanted; | |
633 | cap->seq = seq; | |
634 | cap->issue_seq = seq; | |
635 | cap->mseq = mseq; | |
636 | cap->cap_gen = session->s_cap_gen; | |
637 | ||
638 | if (fmode >= 0) | |
639 | __ceph_get_fmode(ci, fmode); | |
640 | } | |
641 | ||
642 | /* | |
643 | * Return true if cap has not timed out and belongs to the current | |
644 | * generation of the MDS session (i.e. has not gone 'stale' due to | |
645 | * us losing touch with the mds). | |
646 | */ | |
647 | static int __cap_is_valid(struct ceph_cap *cap) | |
648 | { | |
649 | unsigned long ttl; | |
650 | u32 gen; | |
651 | ||
652 | spin_lock(&cap->session->s_gen_ttl_lock); | |
653 | gen = cap->session->s_cap_gen; | |
654 | ttl = cap->session->s_cap_ttl; | |
655 | spin_unlock(&cap->session->s_gen_ttl_lock); | |
656 | ||
657 | if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) { | |
658 | dout("__cap_is_valid %p cap %p issued %s " | |
659 | "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode, | |
660 | cap, ceph_cap_string(cap->issued), cap->cap_gen, gen); | |
661 | return 0; | |
662 | } | |
663 | ||
664 | return 1; | |
665 | } | |
666 | ||
667 | /* | |
668 | * Return set of valid cap bits issued to us. Note that caps time | |
669 | * out, and may be invalidated in bulk if the client session times out | |
670 | * and session->s_cap_gen is bumped. | |
671 | */ | |
672 | int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented) | |
673 | { | |
674 | int have = ci->i_snap_caps; | |
675 | struct ceph_cap *cap; | |
676 | struct rb_node *p; | |
677 | ||
678 | if (implemented) | |
679 | *implemented = 0; | |
680 | for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { | |
681 | cap = rb_entry(p, struct ceph_cap, ci_node); | |
682 | if (!__cap_is_valid(cap)) | |
683 | continue; | |
684 | dout("__ceph_caps_issued %p cap %p issued %s\n", | |
685 | &ci->vfs_inode, cap, ceph_cap_string(cap->issued)); | |
686 | have |= cap->issued; | |
687 | if (implemented) | |
688 | *implemented |= cap->implemented; | |
689 | } | |
690 | /* | |
691 | * exclude caps issued by non-auth MDS, but are been revoking | |
692 | * by the auth MDS. The non-auth MDS should be revoking/exporting | |
693 | * these caps, but the message is delayed. | |
694 | */ | |
695 | if (ci->i_auth_cap) { | |
696 | cap = ci->i_auth_cap; | |
697 | have &= ~cap->implemented | cap->issued; | |
698 | } | |
699 | return have; | |
700 | } | |
701 | ||
702 | /* | |
703 | * Get cap bits issued by caps other than @ocap | |
704 | */ | |
705 | int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap) | |
706 | { | |
707 | int have = ci->i_snap_caps; | |
708 | struct ceph_cap *cap; | |
709 | struct rb_node *p; | |
710 | ||
711 | for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { | |
712 | cap = rb_entry(p, struct ceph_cap, ci_node); | |
713 | if (cap == ocap) | |
714 | continue; | |
715 | if (!__cap_is_valid(cap)) | |
716 | continue; | |
717 | have |= cap->issued; | |
718 | } | |
719 | return have; | |
720 | } | |
721 | ||
722 | /* | |
723 | * Move a cap to the end of the LRU (oldest caps at list head, newest | |
724 | * at list tail). | |
725 | */ | |
726 | static void __touch_cap(struct ceph_cap *cap) | |
727 | { | |
728 | struct ceph_mds_session *s = cap->session; | |
729 | ||
730 | spin_lock(&s->s_cap_lock); | |
731 | if (s->s_cap_iterator == NULL) { | |
732 | dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap, | |
733 | s->s_mds); | |
734 | list_move_tail(&cap->session_caps, &s->s_caps); | |
735 | } else { | |
736 | dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n", | |
737 | &cap->ci->vfs_inode, cap, s->s_mds); | |
738 | } | |
739 | spin_unlock(&s->s_cap_lock); | |
740 | } | |
741 | ||
742 | /* | |
743 | * Check if we hold the given mask. If so, move the cap(s) to the | |
744 | * front of their respective LRUs. (This is the preferred way for | |
745 | * callers to check for caps they want.) | |
746 | */ | |
747 | int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch) | |
748 | { | |
749 | struct ceph_cap *cap; | |
750 | struct rb_node *p; | |
751 | int have = ci->i_snap_caps; | |
752 | ||
753 | if ((have & mask) == mask) { | |
754 | dout("__ceph_caps_issued_mask %p snap issued %s" | |
755 | " (mask %s)\n", &ci->vfs_inode, | |
756 | ceph_cap_string(have), | |
757 | ceph_cap_string(mask)); | |
758 | return 1; | |
759 | } | |
760 | ||
761 | for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { | |
762 | cap = rb_entry(p, struct ceph_cap, ci_node); | |
763 | if (!__cap_is_valid(cap)) | |
764 | continue; | |
765 | if ((cap->issued & mask) == mask) { | |
766 | dout("__ceph_caps_issued_mask %p cap %p issued %s" | |
767 | " (mask %s)\n", &ci->vfs_inode, cap, | |
768 | ceph_cap_string(cap->issued), | |
769 | ceph_cap_string(mask)); | |
770 | if (touch) | |
771 | __touch_cap(cap); | |
772 | return 1; | |
773 | } | |
774 | ||
775 | /* does a combination of caps satisfy mask? */ | |
776 | have |= cap->issued; | |
777 | if ((have & mask) == mask) { | |
778 | dout("__ceph_caps_issued_mask %p combo issued %s" | |
779 | " (mask %s)\n", &ci->vfs_inode, | |
780 | ceph_cap_string(cap->issued), | |
781 | ceph_cap_string(mask)); | |
782 | if (touch) { | |
783 | struct rb_node *q; | |
784 | ||
785 | /* touch this + preceding caps */ | |
786 | __touch_cap(cap); | |
787 | for (q = rb_first(&ci->i_caps); q != p; | |
788 | q = rb_next(q)) { | |
789 | cap = rb_entry(q, struct ceph_cap, | |
790 | ci_node); | |
791 | if (!__cap_is_valid(cap)) | |
792 | continue; | |
793 | __touch_cap(cap); | |
794 | } | |
795 | } | |
796 | return 1; | |
797 | } | |
798 | } | |
799 | ||
800 | return 0; | |
801 | } | |
802 | ||
803 | /* | |
804 | * Return true if mask caps are currently being revoked by an MDS. | |
805 | */ | |
806 | int __ceph_caps_revoking_other(struct ceph_inode_info *ci, | |
807 | struct ceph_cap *ocap, int mask) | |
808 | { | |
809 | struct ceph_cap *cap; | |
810 | struct rb_node *p; | |
811 | ||
812 | for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { | |
813 | cap = rb_entry(p, struct ceph_cap, ci_node); | |
814 | if (cap != ocap && | |
815 | (cap->implemented & ~cap->issued & mask)) | |
816 | return 1; | |
817 | } | |
818 | return 0; | |
819 | } | |
820 | ||
821 | int ceph_caps_revoking(struct ceph_inode_info *ci, int mask) | |
822 | { | |
823 | struct inode *inode = &ci->vfs_inode; | |
824 | int ret; | |
825 | ||
826 | spin_lock(&ci->i_ceph_lock); | |
827 | ret = __ceph_caps_revoking_other(ci, NULL, mask); | |
828 | spin_unlock(&ci->i_ceph_lock); | |
829 | dout("ceph_caps_revoking %p %s = %d\n", inode, | |
830 | ceph_cap_string(mask), ret); | |
831 | return ret; | |
832 | } | |
833 | ||
834 | int __ceph_caps_used(struct ceph_inode_info *ci) | |
835 | { | |
836 | int used = 0; | |
837 | if (ci->i_pin_ref) | |
838 | used |= CEPH_CAP_PIN; | |
839 | if (ci->i_rd_ref) | |
840 | used |= CEPH_CAP_FILE_RD; | |
841 | if (ci->i_rdcache_ref || | |
842 | (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */ | |
843 | ci->vfs_inode.i_data.nrpages)) | |
844 | used |= CEPH_CAP_FILE_CACHE; | |
845 | if (ci->i_wr_ref) | |
846 | used |= CEPH_CAP_FILE_WR; | |
847 | if (ci->i_wb_ref || ci->i_wrbuffer_ref) | |
848 | used |= CEPH_CAP_FILE_BUFFER; | |
849 | return used; | |
850 | } | |
851 | ||
852 | /* | |
853 | * wanted, by virtue of open file modes | |
854 | */ | |
855 | int __ceph_caps_file_wanted(struct ceph_inode_info *ci) | |
856 | { | |
857 | int i, bits = 0; | |
858 | for (i = 0; i < CEPH_FILE_MODE_BITS; i++) { | |
859 | if (ci->i_nr_by_mode[i]) | |
860 | bits |= 1 << i; | |
861 | } | |
862 | if (bits == 0) | |
863 | return 0; | |
864 | return ceph_caps_for_mode(bits >> 1); | |
865 | } | |
866 | ||
867 | /* | |
868 | * Return caps we have registered with the MDS(s) as 'wanted'. | |
869 | */ | |
870 | int __ceph_caps_mds_wanted(struct ceph_inode_info *ci) | |
871 | { | |
872 | struct ceph_cap *cap; | |
873 | struct rb_node *p; | |
874 | int mds_wanted = 0; | |
875 | ||
876 | for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { | |
877 | cap = rb_entry(p, struct ceph_cap, ci_node); | |
878 | if (!__cap_is_valid(cap)) | |
879 | continue; | |
880 | if (cap == ci->i_auth_cap) | |
881 | mds_wanted |= cap->mds_wanted; | |
882 | else | |
883 | mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR); | |
884 | } | |
885 | return mds_wanted; | |
886 | } | |
887 | ||
888 | /* | |
889 | * called under i_ceph_lock | |
890 | */ | |
891 | static int __ceph_is_any_caps(struct ceph_inode_info *ci) | |
892 | { | |
893 | return !RB_EMPTY_ROOT(&ci->i_caps); | |
894 | } | |
895 | ||
896 | int ceph_is_any_caps(struct inode *inode) | |
897 | { | |
898 | struct ceph_inode_info *ci = ceph_inode(inode); | |
899 | int ret; | |
900 | ||
901 | spin_lock(&ci->i_ceph_lock); | |
902 | ret = __ceph_is_any_caps(ci); | |
903 | spin_unlock(&ci->i_ceph_lock); | |
904 | ||
905 | return ret; | |
906 | } | |
907 | ||
908 | static void drop_inode_snap_realm(struct ceph_inode_info *ci) | |
909 | { | |
910 | struct ceph_snap_realm *realm = ci->i_snap_realm; | |
911 | spin_lock(&realm->inodes_with_caps_lock); | |
912 | list_del_init(&ci->i_snap_realm_item); | |
913 | ci->i_snap_realm_counter++; | |
914 | ci->i_snap_realm = NULL; | |
915 | spin_unlock(&realm->inodes_with_caps_lock); | |
916 | ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc, | |
917 | realm); | |
918 | } | |
919 | ||
920 | /* | |
921 | * Remove a cap. Take steps to deal with a racing iterate_session_caps. | |
922 | * | |
923 | * caller should hold i_ceph_lock. | |
924 | * caller will not hold session s_mutex if called from destroy_inode. | |
925 | */ | |
926 | void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release) | |
927 | { | |
928 | struct ceph_mds_session *session = cap->session; | |
929 | struct ceph_inode_info *ci = cap->ci; | |
930 | struct ceph_mds_client *mdsc = | |
931 | ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc; | |
932 | int removed = 0; | |
933 | ||
934 | dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode); | |
935 | ||
936 | /* remove from session list */ | |
937 | spin_lock(&session->s_cap_lock); | |
938 | if (session->s_cap_iterator == cap) { | |
939 | /* not yet, we are iterating over this very cap */ | |
940 | dout("__ceph_remove_cap delaying %p removal from session %p\n", | |
941 | cap, cap->session); | |
942 | } else { | |
943 | list_del_init(&cap->session_caps); | |
944 | session->s_nr_caps--; | |
945 | cap->session = NULL; | |
946 | removed = 1; | |
947 | } | |
948 | /* protect backpointer with s_cap_lock: see iterate_session_caps */ | |
949 | cap->ci = NULL; | |
950 | ||
951 | /* | |
952 | * s_cap_reconnect is protected by s_cap_lock. no one changes | |
953 | * s_cap_gen while session is in the reconnect state. | |
954 | */ | |
955 | if (queue_release && | |
956 | (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) { | |
957 | cap->queue_release = 1; | |
958 | if (removed) { | |
959 | list_add_tail(&cap->session_caps, | |
960 | &session->s_cap_releases); | |
961 | session->s_num_cap_releases++; | |
962 | removed = 0; | |
963 | } | |
964 | } else { | |
965 | cap->queue_release = 0; | |
966 | } | |
967 | cap->cap_ino = ci->i_vino.ino; | |
968 | ||
969 | spin_unlock(&session->s_cap_lock); | |
970 | ||
971 | /* remove from inode list */ | |
972 | rb_erase(&cap->ci_node, &ci->i_caps); | |
973 | if (ci->i_auth_cap == cap) | |
974 | ci->i_auth_cap = NULL; | |
975 | ||
976 | if (removed) | |
977 | ceph_put_cap(mdsc, cap); | |
978 | ||
979 | /* when reconnect denied, we remove session caps forcibly, | |
980 | * i_wr_ref can be non-zero. If there are ongoing write, | |
981 | * keep i_snap_realm. | |
982 | */ | |
983 | if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm) | |
984 | drop_inode_snap_realm(ci); | |
985 | ||
986 | if (!__ceph_is_any_real_caps(ci)) | |
987 | __cap_delay_cancel(mdsc, ci); | |
988 | } | |
989 | ||
990 | /* | |
991 | * Build and send a cap message to the given MDS. | |
992 | * | |
993 | * Caller should be holding s_mutex. | |
994 | */ | |
995 | static int send_cap_msg(struct ceph_mds_session *session, | |
996 | u64 ino, u64 cid, int op, | |
997 | int caps, int wanted, int dirty, | |
998 | u32 seq, u64 flush_tid, u64 oldest_flush_tid, | |
999 | u32 issue_seq, u32 mseq, u64 size, u64 max_size, | |
1000 | struct timespec *mtime, struct timespec *atime, | |
1001 | struct timespec *ctime, u32 time_warp_seq, | |
1002 | kuid_t uid, kgid_t gid, umode_t mode, | |
1003 | u64 xattr_version, | |
1004 | struct ceph_buffer *xattrs_buf, | |
1005 | u64 follows, bool inline_data) | |
1006 | { | |
1007 | struct ceph_mds_caps *fc; | |
1008 | struct ceph_msg *msg; | |
1009 | void *p; | |
1010 | size_t extra_len; | |
1011 | ||
1012 | dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s" | |
1013 | " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu" | |
1014 | " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op), | |
1015 | cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted), | |
1016 | ceph_cap_string(dirty), | |
1017 | seq, issue_seq, flush_tid, oldest_flush_tid, | |
1018 | mseq, follows, size, max_size, | |
1019 | xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0); | |
1020 | ||
1021 | /* flock buffer size + inline version + inline data size + | |
1022 | * osd_epoch_barrier + oldest_flush_tid */ | |
1023 | extra_len = 4 + 8 + 4 + 4 + 8; | |
1024 | msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len, | |
1025 | GFP_NOFS, false); | |
1026 | if (!msg) | |
1027 | return -ENOMEM; | |
1028 | ||
1029 | msg->hdr.version = cpu_to_le16(6); | |
1030 | msg->hdr.tid = cpu_to_le64(flush_tid); | |
1031 | ||
1032 | fc = msg->front.iov_base; | |
1033 | memset(fc, 0, sizeof(*fc)); | |
1034 | ||
1035 | fc->cap_id = cpu_to_le64(cid); | |
1036 | fc->op = cpu_to_le32(op); | |
1037 | fc->seq = cpu_to_le32(seq); | |
1038 | fc->issue_seq = cpu_to_le32(issue_seq); | |
1039 | fc->migrate_seq = cpu_to_le32(mseq); | |
1040 | fc->caps = cpu_to_le32(caps); | |
1041 | fc->wanted = cpu_to_le32(wanted); | |
1042 | fc->dirty = cpu_to_le32(dirty); | |
1043 | fc->ino = cpu_to_le64(ino); | |
1044 | fc->snap_follows = cpu_to_le64(follows); | |
1045 | ||
1046 | fc->size = cpu_to_le64(size); | |
1047 | fc->max_size = cpu_to_le64(max_size); | |
1048 | if (mtime) | |
1049 | ceph_encode_timespec(&fc->mtime, mtime); | |
1050 | if (atime) | |
1051 | ceph_encode_timespec(&fc->atime, atime); | |
1052 | if (ctime) | |
1053 | ceph_encode_timespec(&fc->ctime, ctime); | |
1054 | fc->time_warp_seq = cpu_to_le32(time_warp_seq); | |
1055 | ||
1056 | fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid)); | |
1057 | fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid)); | |
1058 | fc->mode = cpu_to_le32(mode); | |
1059 | ||
1060 | p = fc + 1; | |
1061 | /* flock buffer size */ | |
1062 | ceph_encode_32(&p, 0); | |
1063 | /* inline version */ | |
1064 | ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE); | |
1065 | /* inline data size */ | |
1066 | ceph_encode_32(&p, 0); | |
1067 | /* osd_epoch_barrier */ | |
1068 | ceph_encode_32(&p, 0); | |
1069 | /* oldest_flush_tid */ | |
1070 | ceph_encode_64(&p, oldest_flush_tid); | |
1071 | ||
1072 | fc->xattr_version = cpu_to_le64(xattr_version); | |
1073 | if (xattrs_buf) { | |
1074 | msg->middle = ceph_buffer_get(xattrs_buf); | |
1075 | fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len); | |
1076 | msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len); | |
1077 | } | |
1078 | ||
1079 | ceph_con_send(&session->s_con, msg); | |
1080 | return 0; | |
1081 | } | |
1082 | ||
1083 | /* | |
1084 | * Queue cap releases when an inode is dropped from our cache. Since | |
1085 | * inode is about to be destroyed, there is no need for i_ceph_lock. | |
1086 | */ | |
1087 | void ceph_queue_caps_release(struct inode *inode) | |
1088 | { | |
1089 | struct ceph_inode_info *ci = ceph_inode(inode); | |
1090 | struct rb_node *p; | |
1091 | ||
1092 | p = rb_first(&ci->i_caps); | |
1093 | while (p) { | |
1094 | struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node); | |
1095 | p = rb_next(p); | |
1096 | __ceph_remove_cap(cap, true); | |
1097 | } | |
1098 | } | |
1099 | ||
1100 | /* | |
1101 | * Send a cap msg on the given inode. Update our caps state, then | |
1102 | * drop i_ceph_lock and send the message. | |
1103 | * | |
1104 | * Make note of max_size reported/requested from mds, revoked caps | |
1105 | * that have now been implemented. | |
1106 | * | |
1107 | * Make half-hearted attempt ot to invalidate page cache if we are | |
1108 | * dropping RDCACHE. Note that this will leave behind locked pages | |
1109 | * that we'll then need to deal with elsewhere. | |
1110 | * | |
1111 | * Return non-zero if delayed release, or we experienced an error | |
1112 | * such that the caller should requeue + retry later. | |
1113 | * | |
1114 | * called with i_ceph_lock, then drops it. | |
1115 | * caller should hold snap_rwsem (read), s_mutex. | |
1116 | */ | |
1117 | static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap, | |
1118 | int op, int used, int want, int retain, int flushing, | |
1119 | u64 flush_tid, u64 oldest_flush_tid) | |
1120 | __releases(cap->ci->i_ceph_lock) | |
1121 | { | |
1122 | struct ceph_inode_info *ci = cap->ci; | |
1123 | struct inode *inode = &ci->vfs_inode; | |
1124 | u64 cap_id = cap->cap_id; | |
1125 | int held, revoking, dropping, keep; | |
1126 | u64 follows, size, max_size; | |
1127 | u32 seq, issue_seq, mseq, time_warp_seq; | |
1128 | struct timespec mtime, atime, ctime; | |
1129 | int wake = 0; | |
1130 | umode_t mode; | |
1131 | kuid_t uid; | |
1132 | kgid_t gid; | |
1133 | struct ceph_mds_session *session; | |
1134 | u64 xattr_version = 0; | |
1135 | struct ceph_buffer *xattr_blob = NULL; | |
1136 | int delayed = 0; | |
1137 | int ret; | |
1138 | bool inline_data; | |
1139 | ||
1140 | held = cap->issued | cap->implemented; | |
1141 | revoking = cap->implemented & ~cap->issued; | |
1142 | retain &= ~revoking; | |
1143 | dropping = cap->issued & ~retain; | |
1144 | ||
1145 | dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n", | |
1146 | inode, cap, cap->session, | |
1147 | ceph_cap_string(held), ceph_cap_string(held & retain), | |
1148 | ceph_cap_string(revoking)); | |
1149 | BUG_ON((retain & CEPH_CAP_PIN) == 0); | |
1150 | ||
1151 | session = cap->session; | |
1152 | ||
1153 | /* don't release wanted unless we've waited a bit. */ | |
1154 | if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 && | |
1155 | time_before(jiffies, ci->i_hold_caps_min)) { | |
1156 | dout(" delaying issued %s -> %s, wanted %s -> %s on send\n", | |
1157 | ceph_cap_string(cap->issued), | |
1158 | ceph_cap_string(cap->issued & retain), | |
1159 | ceph_cap_string(cap->mds_wanted), | |
1160 | ceph_cap_string(want)); | |
1161 | want |= cap->mds_wanted; | |
1162 | retain |= cap->issued; | |
1163 | delayed = 1; | |
1164 | } | |
1165 | ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH); | |
1166 | ||
1167 | cap->issued &= retain; /* drop bits we don't want */ | |
1168 | if (cap->implemented & ~cap->issued) { | |
1169 | /* | |
1170 | * Wake up any waiters on wanted -> needed transition. | |
1171 | * This is due to the weird transition from buffered | |
1172 | * to sync IO... we need to flush dirty pages _before_ | |
1173 | * allowing sync writes to avoid reordering. | |
1174 | */ | |
1175 | wake = 1; | |
1176 | } | |
1177 | cap->implemented &= cap->issued | used; | |
1178 | cap->mds_wanted = want; | |
1179 | ||
1180 | follows = flushing ? ci->i_head_snapc->seq : 0; | |
1181 | ||
1182 | keep = cap->implemented; | |
1183 | seq = cap->seq; | |
1184 | issue_seq = cap->issue_seq; | |
1185 | mseq = cap->mseq; | |
1186 | size = inode->i_size; | |
1187 | ci->i_reported_size = size; | |
1188 | max_size = ci->i_wanted_max_size; | |
1189 | ci->i_requested_max_size = max_size; | |
1190 | mtime = inode->i_mtime; | |
1191 | atime = inode->i_atime; | |
1192 | ctime = inode->i_ctime; | |
1193 | time_warp_seq = ci->i_time_warp_seq; | |
1194 | uid = inode->i_uid; | |
1195 | gid = inode->i_gid; | |
1196 | mode = inode->i_mode; | |
1197 | ||
1198 | if (flushing & CEPH_CAP_XATTR_EXCL) { | |
1199 | __ceph_build_xattrs_blob(ci); | |
1200 | xattr_blob = ci->i_xattrs.blob; | |
1201 | xattr_version = ci->i_xattrs.version; | |
1202 | } | |
1203 | ||
1204 | inline_data = ci->i_inline_version != CEPH_INLINE_NONE; | |
1205 | ||
1206 | spin_unlock(&ci->i_ceph_lock); | |
1207 | ||
1208 | ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id, | |
1209 | op, keep, want, flushing, seq, | |
1210 | flush_tid, oldest_flush_tid, issue_seq, mseq, | |
1211 | size, max_size, &mtime, &atime, &ctime, time_warp_seq, | |
1212 | uid, gid, mode, xattr_version, xattr_blob, | |
1213 | follows, inline_data); | |
1214 | if (ret < 0) { | |
1215 | dout("error sending cap msg, must requeue %p\n", inode); | |
1216 | delayed = 1; | |
1217 | } | |
1218 | ||
1219 | if (wake) | |
1220 | wake_up_all(&ci->i_cap_wq); | |
1221 | ||
1222 | return delayed; | |
1223 | } | |
1224 | ||
1225 | static inline int __send_flush_snap(struct inode *inode, | |
1226 | struct ceph_mds_session *session, | |
1227 | struct ceph_cap_snap *capsnap, | |
1228 | u32 mseq, u64 oldest_flush_tid) | |
1229 | { | |
1230 | return send_cap_msg(session, ceph_vino(inode).ino, 0, | |
1231 | CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0, | |
1232 | capsnap->dirty, 0, capsnap->cap_flush.tid, | |
1233 | oldest_flush_tid, 0, mseq, capsnap->size, 0, | |
1234 | &capsnap->mtime, &capsnap->atime, | |
1235 | &capsnap->ctime, capsnap->time_warp_seq, | |
1236 | capsnap->uid, capsnap->gid, capsnap->mode, | |
1237 | capsnap->xattr_version, capsnap->xattr_blob, | |
1238 | capsnap->follows, capsnap->inline_data); | |
1239 | } | |
1240 | ||
1241 | /* | |
1242 | * When a snapshot is taken, clients accumulate dirty metadata on | |
1243 | * inodes with capabilities in ceph_cap_snaps to describe the file | |
1244 | * state at the time the snapshot was taken. This must be flushed | |
1245 | * asynchronously back to the MDS once sync writes complete and dirty | |
1246 | * data is written out. | |
1247 | * | |
1248 | * Called under i_ceph_lock. Takes s_mutex as needed. | |
1249 | */ | |
1250 | static void __ceph_flush_snaps(struct ceph_inode_info *ci, | |
1251 | struct ceph_mds_session *session) | |
1252 | __releases(ci->i_ceph_lock) | |
1253 | __acquires(ci->i_ceph_lock) | |
1254 | { | |
1255 | struct inode *inode = &ci->vfs_inode; | |
1256 | struct ceph_mds_client *mdsc = session->s_mdsc; | |
1257 | struct ceph_cap_snap *capsnap; | |
1258 | u64 oldest_flush_tid = 0; | |
1259 | u64 first_tid = 1, last_tid = 0; | |
1260 | ||
1261 | dout("__flush_snaps %p session %p\n", inode, session); | |
1262 | ||
1263 | list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { | |
1264 | /* | |
1265 | * we need to wait for sync writes to complete and for dirty | |
1266 | * pages to be written out. | |
1267 | */ | |
1268 | if (capsnap->dirty_pages || capsnap->writing) | |
1269 | break; | |
1270 | ||
1271 | /* should be removed by ceph_try_drop_cap_snap() */ | |
1272 | BUG_ON(!capsnap->need_flush); | |
1273 | ||
1274 | /* only flush each capsnap once */ | |
1275 | if (capsnap->cap_flush.tid > 0) { | |
1276 | dout(" already flushed %p, skipping\n", capsnap); | |
1277 | continue; | |
1278 | } | |
1279 | ||
1280 | spin_lock(&mdsc->cap_dirty_lock); | |
1281 | capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid; | |
1282 | list_add_tail(&capsnap->cap_flush.g_list, | |
1283 | &mdsc->cap_flush_list); | |
1284 | if (oldest_flush_tid == 0) | |
1285 | oldest_flush_tid = __get_oldest_flush_tid(mdsc); | |
1286 | if (list_empty(&ci->i_flushing_item)) { | |
1287 | list_add_tail(&ci->i_flushing_item, | |
1288 | &session->s_cap_flushing); | |
1289 | } | |
1290 | spin_unlock(&mdsc->cap_dirty_lock); | |
1291 | ||
1292 | list_add_tail(&capsnap->cap_flush.i_list, | |
1293 | &ci->i_cap_flush_list); | |
1294 | ||
1295 | if (first_tid == 1) | |
1296 | first_tid = capsnap->cap_flush.tid; | |
1297 | last_tid = capsnap->cap_flush.tid; | |
1298 | } | |
1299 | ||
1300 | ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS; | |
1301 | ||
1302 | while (first_tid <= last_tid) { | |
1303 | struct ceph_cap *cap = ci->i_auth_cap; | |
1304 | struct ceph_cap_flush *cf; | |
1305 | int ret; | |
1306 | ||
1307 | if (!(cap && cap->session == session)) { | |
1308 | dout("__flush_snaps %p auth cap %p not mds%d, " | |
1309 | "stop\n", inode, cap, session->s_mds); | |
1310 | break; | |
1311 | } | |
1312 | ||
1313 | ret = -ENOENT; | |
1314 | list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) { | |
1315 | if (cf->tid >= first_tid) { | |
1316 | ret = 0; | |
1317 | break; | |
1318 | } | |
1319 | } | |
1320 | if (ret < 0) | |
1321 | break; | |
1322 | ||
1323 | first_tid = cf->tid + 1; | |
1324 | ||
1325 | capsnap = container_of(cf, struct ceph_cap_snap, cap_flush); | |
1326 | atomic_inc(&capsnap->nref); | |
1327 | spin_unlock(&ci->i_ceph_lock); | |
1328 | ||
1329 | dout("__flush_snaps %p capsnap %p tid %llu %s\n", | |
1330 | inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty)); | |
1331 | ||
1332 | ret = __send_flush_snap(inode, session, capsnap, cap->mseq, | |
1333 | oldest_flush_tid); | |
1334 | if (ret < 0) { | |
1335 | pr_err("__flush_snaps: error sending cap flushsnap, " | |
1336 | "ino (%llx.%llx) tid %llu follows %llu\n", | |
1337 | ceph_vinop(inode), cf->tid, capsnap->follows); | |
1338 | } | |
1339 | ||
1340 | ceph_put_cap_snap(capsnap); | |
1341 | spin_lock(&ci->i_ceph_lock); | |
1342 | } | |
1343 | } | |
1344 | ||
1345 | void ceph_flush_snaps(struct ceph_inode_info *ci, | |
1346 | struct ceph_mds_session **psession) | |
1347 | { | |
1348 | struct inode *inode = &ci->vfs_inode; | |
1349 | struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; | |
1350 | struct ceph_mds_session *session = *psession; | |
1351 | int mds; | |
1352 | dout("ceph_flush_snaps %p\n", inode); | |
1353 | retry: | |
1354 | spin_lock(&ci->i_ceph_lock); | |
1355 | if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) { | |
1356 | dout(" no capsnap needs flush, doing nothing\n"); | |
1357 | goto out; | |
1358 | } | |
1359 | if (!ci->i_auth_cap) { | |
1360 | dout(" no auth cap (migrating?), doing nothing\n"); | |
1361 | goto out; | |
1362 | } | |
1363 | ||
1364 | mds = ci->i_auth_cap->session->s_mds; | |
1365 | if (session && session->s_mds != mds) { | |
1366 | dout(" oops, wrong session %p mutex\n", session); | |
1367 | mutex_unlock(&session->s_mutex); | |
1368 | ceph_put_mds_session(session); | |
1369 | session = NULL; | |
1370 | } | |
1371 | if (!session) { | |
1372 | spin_unlock(&ci->i_ceph_lock); | |
1373 | mutex_lock(&mdsc->mutex); | |
1374 | session = __ceph_lookup_mds_session(mdsc, mds); | |
1375 | mutex_unlock(&mdsc->mutex); | |
1376 | if (session) { | |
1377 | dout(" inverting session/ino locks on %p\n", session); | |
1378 | mutex_lock(&session->s_mutex); | |
1379 | } | |
1380 | goto retry; | |
1381 | } | |
1382 | ||
1383 | __ceph_flush_snaps(ci, session); | |
1384 | out: | |
1385 | spin_unlock(&ci->i_ceph_lock); | |
1386 | ||
1387 | if (psession) { | |
1388 | *psession = session; | |
1389 | } else { | |
1390 | mutex_unlock(&session->s_mutex); | |
1391 | ceph_put_mds_session(session); | |
1392 | } | |
1393 | /* we flushed them all; remove this inode from the queue */ | |
1394 | spin_lock(&mdsc->snap_flush_lock); | |
1395 | list_del_init(&ci->i_snap_flush_item); | |
1396 | spin_unlock(&mdsc->snap_flush_lock); | |
1397 | } | |
1398 | ||
1399 | /* | |
1400 | * Mark caps dirty. If inode is newly dirty, return the dirty flags. | |
1401 | * Caller is then responsible for calling __mark_inode_dirty with the | |
1402 | * returned flags value. | |
1403 | */ | |
1404 | int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask, | |
1405 | struct ceph_cap_flush **pcf) | |
1406 | { | |
1407 | struct ceph_mds_client *mdsc = | |
1408 | ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc; | |
1409 | struct inode *inode = &ci->vfs_inode; | |
1410 | int was = ci->i_dirty_caps; | |
1411 | int dirty = 0; | |
1412 | ||
1413 | if (!ci->i_auth_cap) { | |
1414 | pr_warn("__mark_dirty_caps %p %llx mask %s, " | |
1415 | "but no auth cap (session was closed?)\n", | |
1416 | inode, ceph_ino(inode), ceph_cap_string(mask)); | |
1417 | return 0; | |
1418 | } | |
1419 | ||
1420 | dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode, | |
1421 | ceph_cap_string(mask), ceph_cap_string(was), | |
1422 | ceph_cap_string(was | mask)); | |
1423 | ci->i_dirty_caps |= mask; | |
1424 | if (was == 0) { | |
1425 | WARN_ON_ONCE(ci->i_prealloc_cap_flush); | |
1426 | swap(ci->i_prealloc_cap_flush, *pcf); | |
1427 | ||
1428 | if (!ci->i_head_snapc) { | |
1429 | WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem)); | |
1430 | ci->i_head_snapc = ceph_get_snap_context( | |
1431 | ci->i_snap_realm->cached_context); | |
1432 | } | |
1433 | dout(" inode %p now dirty snapc %p auth cap %p\n", | |
1434 | &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap); | |
1435 | BUG_ON(!list_empty(&ci->i_dirty_item)); | |
1436 | spin_lock(&mdsc->cap_dirty_lock); | |
1437 | list_add(&ci->i_dirty_item, &mdsc->cap_dirty); | |
1438 | spin_unlock(&mdsc->cap_dirty_lock); | |
1439 | if (ci->i_flushing_caps == 0) { | |
1440 | ihold(inode); | |
1441 | dirty |= I_DIRTY_SYNC; | |
1442 | } | |
1443 | } else { | |
1444 | WARN_ON_ONCE(!ci->i_prealloc_cap_flush); | |
1445 | } | |
1446 | BUG_ON(list_empty(&ci->i_dirty_item)); | |
1447 | if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) && | |
1448 | (mask & CEPH_CAP_FILE_BUFFER)) | |
1449 | dirty |= I_DIRTY_DATASYNC; | |
1450 | __cap_delay_requeue(mdsc, ci); | |
1451 | return dirty; | |
1452 | } | |
1453 | ||
1454 | struct ceph_cap_flush *ceph_alloc_cap_flush(void) | |
1455 | { | |
1456 | return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL); | |
1457 | } | |
1458 | ||
1459 | void ceph_free_cap_flush(struct ceph_cap_flush *cf) | |
1460 | { | |
1461 | if (cf) | |
1462 | kmem_cache_free(ceph_cap_flush_cachep, cf); | |
1463 | } | |
1464 | ||
1465 | static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc) | |
1466 | { | |
1467 | if (!list_empty(&mdsc->cap_flush_list)) { | |
1468 | struct ceph_cap_flush *cf = | |
1469 | list_first_entry(&mdsc->cap_flush_list, | |
1470 | struct ceph_cap_flush, g_list); | |
1471 | return cf->tid; | |
1472 | } | |
1473 | return 0; | |
1474 | } | |
1475 | ||
1476 | /* | |
1477 | * Add dirty inode to the flushing list. Assigned a seq number so we | |
1478 | * can wait for caps to flush without starving. | |
1479 | * | |
1480 | * Called under i_ceph_lock. | |
1481 | */ | |
1482 | static int __mark_caps_flushing(struct inode *inode, | |
1483 | struct ceph_mds_session *session, | |
1484 | u64 *flush_tid, u64 *oldest_flush_tid) | |
1485 | { | |
1486 | struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; | |
1487 | struct ceph_inode_info *ci = ceph_inode(inode); | |
1488 | struct ceph_cap_flush *cf = NULL; | |
1489 | int flushing; | |
1490 | ||
1491 | BUG_ON(ci->i_dirty_caps == 0); | |
1492 | BUG_ON(list_empty(&ci->i_dirty_item)); | |
1493 | BUG_ON(!ci->i_prealloc_cap_flush); | |
1494 | ||
1495 | flushing = ci->i_dirty_caps; | |
1496 | dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n", | |
1497 | ceph_cap_string(flushing), | |
1498 | ceph_cap_string(ci->i_flushing_caps), | |
1499 | ceph_cap_string(ci->i_flushing_caps | flushing)); | |
1500 | ci->i_flushing_caps |= flushing; | |
1501 | ci->i_dirty_caps = 0; | |
1502 | dout(" inode %p now !dirty\n", inode); | |
1503 | ||
1504 | swap(cf, ci->i_prealloc_cap_flush); | |
1505 | cf->caps = flushing; | |
1506 | ||
1507 | spin_lock(&mdsc->cap_dirty_lock); | |
1508 | list_del_init(&ci->i_dirty_item); | |
1509 | ||
1510 | cf->tid = ++mdsc->last_cap_flush_tid; | |
1511 | list_add_tail(&cf->g_list, &mdsc->cap_flush_list); | |
1512 | *oldest_flush_tid = __get_oldest_flush_tid(mdsc); | |
1513 | ||
1514 | if (list_empty(&ci->i_flushing_item)) { | |
1515 | list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing); | |
1516 | mdsc->num_cap_flushing++; | |
1517 | } | |
1518 | spin_unlock(&mdsc->cap_dirty_lock); | |
1519 | ||
1520 | list_add_tail(&cf->i_list, &ci->i_cap_flush_list); | |
1521 | ||
1522 | *flush_tid = cf->tid; | |
1523 | return flushing; | |
1524 | } | |
1525 | ||
1526 | /* | |
1527 | * try to invalidate mapping pages without blocking. | |
1528 | */ | |
1529 | static int try_nonblocking_invalidate(struct inode *inode) | |
1530 | { | |
1531 | struct ceph_inode_info *ci = ceph_inode(inode); | |
1532 | u32 invalidating_gen = ci->i_rdcache_gen; | |
1533 | ||
1534 | spin_unlock(&ci->i_ceph_lock); | |
1535 | invalidate_mapping_pages(&inode->i_data, 0, -1); | |
1536 | spin_lock(&ci->i_ceph_lock); | |
1537 | ||
1538 | if (inode->i_data.nrpages == 0 && | |
1539 | invalidating_gen == ci->i_rdcache_gen) { | |
1540 | /* success. */ | |
1541 | dout("try_nonblocking_invalidate %p success\n", inode); | |
1542 | /* save any racing async invalidate some trouble */ | |
1543 | ci->i_rdcache_revoking = ci->i_rdcache_gen - 1; | |
1544 | return 0; | |
1545 | } | |
1546 | dout("try_nonblocking_invalidate %p failed\n", inode); | |
1547 | return -1; | |
1548 | } | |
1549 | ||
1550 | /* | |
1551 | * Swiss army knife function to examine currently used and wanted | |
1552 | * versus held caps. Release, flush, ack revoked caps to mds as | |
1553 | * appropriate. | |
1554 | * | |
1555 | * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay | |
1556 | * cap release further. | |
1557 | * CHECK_CAPS_AUTHONLY - we should only check the auth cap | |
1558 | * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without | |
1559 | * further delay. | |
1560 | */ | |
1561 | void ceph_check_caps(struct ceph_inode_info *ci, int flags, | |
1562 | struct ceph_mds_session *session) | |
1563 | { | |
1564 | struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode); | |
1565 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
1566 | struct inode *inode = &ci->vfs_inode; | |
1567 | struct ceph_cap *cap; | |
1568 | u64 flush_tid, oldest_flush_tid; | |
1569 | int file_wanted, used, cap_used; | |
1570 | int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */ | |
1571 | int issued, implemented, want, retain, revoking, flushing = 0; | |
1572 | int mds = -1; /* keep track of how far we've gone through i_caps list | |
1573 | to avoid an infinite loop on retry */ | |
1574 | struct rb_node *p; | |
1575 | int delayed = 0, sent = 0, num; | |
1576 | bool is_delayed = flags & CHECK_CAPS_NODELAY; | |
1577 | bool queue_invalidate = false; | |
1578 | bool force_requeue = false; | |
1579 | bool tried_invalidate = false; | |
1580 | ||
1581 | /* if we are unmounting, flush any unused caps immediately. */ | |
1582 | if (mdsc->stopping) | |
1583 | is_delayed = 1; | |
1584 | ||
1585 | spin_lock(&ci->i_ceph_lock); | |
1586 | ||
1587 | if (ci->i_ceph_flags & CEPH_I_FLUSH) | |
1588 | flags |= CHECK_CAPS_FLUSH; | |
1589 | ||
1590 | goto retry_locked; | |
1591 | retry: | |
1592 | spin_lock(&ci->i_ceph_lock); | |
1593 | retry_locked: | |
1594 | file_wanted = __ceph_caps_file_wanted(ci); | |
1595 | used = __ceph_caps_used(ci); | |
1596 | issued = __ceph_caps_issued(ci, &implemented); | |
1597 | revoking = implemented & ~issued; | |
1598 | ||
1599 | want = file_wanted; | |
1600 | retain = file_wanted | used | CEPH_CAP_PIN; | |
1601 | if (!mdsc->stopping && inode->i_nlink > 0) { | |
1602 | if (file_wanted) { | |
1603 | retain |= CEPH_CAP_ANY; /* be greedy */ | |
1604 | } else if (S_ISDIR(inode->i_mode) && | |
1605 | (issued & CEPH_CAP_FILE_SHARED) && | |
1606 | __ceph_dir_is_complete(ci)) { | |
1607 | /* | |
1608 | * If a directory is complete, we want to keep | |
1609 | * the exclusive cap. So that MDS does not end up | |
1610 | * revoking the shared cap on every create/unlink | |
1611 | * operation. | |
1612 | */ | |
1613 | want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL; | |
1614 | retain |= want; | |
1615 | } else { | |
1616 | ||
1617 | retain |= CEPH_CAP_ANY_SHARED; | |
1618 | /* | |
1619 | * keep RD only if we didn't have the file open RW, | |
1620 | * because then the mds would revoke it anyway to | |
1621 | * journal max_size=0. | |
1622 | */ | |
1623 | if (ci->i_max_size == 0) | |
1624 | retain |= CEPH_CAP_ANY_RD; | |
1625 | } | |
1626 | } | |
1627 | ||
1628 | dout("check_caps %p file_want %s used %s dirty %s flushing %s" | |
1629 | " issued %s revoking %s retain %s %s%s%s\n", inode, | |
1630 | ceph_cap_string(file_wanted), | |
1631 | ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps), | |
1632 | ceph_cap_string(ci->i_flushing_caps), | |
1633 | ceph_cap_string(issued), ceph_cap_string(revoking), | |
1634 | ceph_cap_string(retain), | |
1635 | (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "", | |
1636 | (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "", | |
1637 | (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : ""); | |
1638 | ||
1639 | /* | |
1640 | * If we no longer need to hold onto old our caps, and we may | |
1641 | * have cached pages, but don't want them, then try to invalidate. | |
1642 | * If we fail, it's because pages are locked.... try again later. | |
1643 | */ | |
1644 | if ((!is_delayed || mdsc->stopping) && | |
1645 | !S_ISDIR(inode->i_mode) && /* ignore readdir cache */ | |
1646 | !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */ | |
1647 | inode->i_data.nrpages && /* have cached pages */ | |
1648 | (revoking & (CEPH_CAP_FILE_CACHE| | |
1649 | CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */ | |
1650 | !tried_invalidate) { | |
1651 | dout("check_caps trying to invalidate on %p\n", inode); | |
1652 | if (try_nonblocking_invalidate(inode) < 0) { | |
1653 | if (revoking & (CEPH_CAP_FILE_CACHE| | |
1654 | CEPH_CAP_FILE_LAZYIO)) { | |
1655 | dout("check_caps queuing invalidate\n"); | |
1656 | queue_invalidate = true; | |
1657 | ci->i_rdcache_revoking = ci->i_rdcache_gen; | |
1658 | } else { | |
1659 | dout("check_caps failed to invalidate pages\n"); | |
1660 | /* we failed to invalidate pages. check these | |
1661 | caps again later. */ | |
1662 | force_requeue = true; | |
1663 | __cap_set_timeouts(mdsc, ci); | |
1664 | } | |
1665 | } | |
1666 | tried_invalidate = true; | |
1667 | goto retry_locked; | |
1668 | } | |
1669 | ||
1670 | num = 0; | |
1671 | for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) { | |
1672 | cap = rb_entry(p, struct ceph_cap, ci_node); | |
1673 | num++; | |
1674 | ||
1675 | /* avoid looping forever */ | |
1676 | if (mds >= cap->mds || | |
1677 | ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap)) | |
1678 | continue; | |
1679 | ||
1680 | /* NOTE: no side-effects allowed, until we take s_mutex */ | |
1681 | ||
1682 | cap_used = used; | |
1683 | if (ci->i_auth_cap && cap != ci->i_auth_cap) | |
1684 | cap_used &= ~ci->i_auth_cap->issued; | |
1685 | ||
1686 | revoking = cap->implemented & ~cap->issued; | |
1687 | dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n", | |
1688 | cap->mds, cap, ceph_cap_string(cap_used), | |
1689 | ceph_cap_string(cap->issued), | |
1690 | ceph_cap_string(cap->implemented), | |
1691 | ceph_cap_string(revoking)); | |
1692 | ||
1693 | if (cap == ci->i_auth_cap && | |
1694 | (cap->issued & CEPH_CAP_FILE_WR)) { | |
1695 | /* request larger max_size from MDS? */ | |
1696 | if (ci->i_wanted_max_size > ci->i_max_size && | |
1697 | ci->i_wanted_max_size > ci->i_requested_max_size) { | |
1698 | dout("requesting new max_size\n"); | |
1699 | goto ack; | |
1700 | } | |
1701 | ||
1702 | /* approaching file_max? */ | |
1703 | if ((inode->i_size << 1) >= ci->i_max_size && | |
1704 | (ci->i_reported_size << 1) < ci->i_max_size) { | |
1705 | dout("i_size approaching max_size\n"); | |
1706 | goto ack; | |
1707 | } | |
1708 | } | |
1709 | /* flush anything dirty? */ | |
1710 | if (cap == ci->i_auth_cap) { | |
1711 | if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) { | |
1712 | dout("flushing dirty caps\n"); | |
1713 | goto ack; | |
1714 | } | |
1715 | if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) { | |
1716 | dout("flushing snap caps\n"); | |
1717 | goto ack; | |
1718 | } | |
1719 | } | |
1720 | ||
1721 | /* completed revocation? going down and there are no caps? */ | |
1722 | if (revoking && (revoking & cap_used) == 0) { | |
1723 | dout("completed revocation of %s\n", | |
1724 | ceph_cap_string(cap->implemented & ~cap->issued)); | |
1725 | goto ack; | |
1726 | } | |
1727 | ||
1728 | /* want more caps from mds? */ | |
1729 | if (want & ~(cap->mds_wanted | cap->issued)) | |
1730 | goto ack; | |
1731 | ||
1732 | /* things we might delay */ | |
1733 | if ((cap->issued & ~retain) == 0 && | |
1734 | cap->mds_wanted == want) | |
1735 | continue; /* nope, all good */ | |
1736 | ||
1737 | if (is_delayed) | |
1738 | goto ack; | |
1739 | ||
1740 | /* delay? */ | |
1741 | if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 && | |
1742 | time_before(jiffies, ci->i_hold_caps_max)) { | |
1743 | dout(" delaying issued %s -> %s, wanted %s -> %s\n", | |
1744 | ceph_cap_string(cap->issued), | |
1745 | ceph_cap_string(cap->issued & retain), | |
1746 | ceph_cap_string(cap->mds_wanted), | |
1747 | ceph_cap_string(want)); | |
1748 | delayed++; | |
1749 | continue; | |
1750 | } | |
1751 | ||
1752 | ack: | |
1753 | if (ci->i_ceph_flags & CEPH_I_NOFLUSH) { | |
1754 | dout(" skipping %p I_NOFLUSH set\n", inode); | |
1755 | continue; | |
1756 | } | |
1757 | ||
1758 | if (session && session != cap->session) { | |
1759 | dout("oops, wrong session %p mutex\n", session); | |
1760 | mutex_unlock(&session->s_mutex); | |
1761 | session = NULL; | |
1762 | } | |
1763 | if (!session) { | |
1764 | session = cap->session; | |
1765 | if (mutex_trylock(&session->s_mutex) == 0) { | |
1766 | dout("inverting session/ino locks on %p\n", | |
1767 | session); | |
1768 | spin_unlock(&ci->i_ceph_lock); | |
1769 | if (took_snap_rwsem) { | |
1770 | up_read(&mdsc->snap_rwsem); | |
1771 | took_snap_rwsem = 0; | |
1772 | } | |
1773 | mutex_lock(&session->s_mutex); | |
1774 | goto retry; | |
1775 | } | |
1776 | } | |
1777 | ||
1778 | /* kick flushing and flush snaps before sending normal | |
1779 | * cap message */ | |
1780 | if (cap == ci->i_auth_cap && | |
1781 | (ci->i_ceph_flags & | |
1782 | (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) { | |
1783 | if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) { | |
1784 | spin_lock(&mdsc->cap_dirty_lock); | |
1785 | oldest_flush_tid = __get_oldest_flush_tid(mdsc); | |
1786 | spin_unlock(&mdsc->cap_dirty_lock); | |
1787 | __kick_flushing_caps(mdsc, session, ci, | |
1788 | oldest_flush_tid); | |
1789 | ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; | |
1790 | } | |
1791 | if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) | |
1792 | __ceph_flush_snaps(ci, session); | |
1793 | ||
1794 | goto retry_locked; | |
1795 | } | |
1796 | ||
1797 | /* take snap_rwsem after session mutex */ | |
1798 | if (!took_snap_rwsem) { | |
1799 | if (down_read_trylock(&mdsc->snap_rwsem) == 0) { | |
1800 | dout("inverting snap/in locks on %p\n", | |
1801 | inode); | |
1802 | spin_unlock(&ci->i_ceph_lock); | |
1803 | down_read(&mdsc->snap_rwsem); | |
1804 | took_snap_rwsem = 1; | |
1805 | goto retry; | |
1806 | } | |
1807 | took_snap_rwsem = 1; | |
1808 | } | |
1809 | ||
1810 | if (cap == ci->i_auth_cap && ci->i_dirty_caps) { | |
1811 | flushing = __mark_caps_flushing(inode, session, | |
1812 | &flush_tid, | |
1813 | &oldest_flush_tid); | |
1814 | } else { | |
1815 | flushing = 0; | |
1816 | flush_tid = 0; | |
1817 | spin_lock(&mdsc->cap_dirty_lock); | |
1818 | oldest_flush_tid = __get_oldest_flush_tid(mdsc); | |
1819 | spin_unlock(&mdsc->cap_dirty_lock); | |
1820 | } | |
1821 | ||
1822 | mds = cap->mds; /* remember mds, so we don't repeat */ | |
1823 | sent++; | |
1824 | ||
1825 | /* __send_cap drops i_ceph_lock */ | |
1826 | delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used, | |
1827 | want, retain, flushing, | |
1828 | flush_tid, oldest_flush_tid); | |
1829 | goto retry; /* retake i_ceph_lock and restart our cap scan. */ | |
1830 | } | |
1831 | ||
1832 | /* | |
1833 | * Reschedule delayed caps release if we delayed anything, | |
1834 | * otherwise cancel. | |
1835 | */ | |
1836 | if (delayed && is_delayed) | |
1837 | force_requeue = true; /* __send_cap delayed release; requeue */ | |
1838 | if (!delayed && !is_delayed) | |
1839 | __cap_delay_cancel(mdsc, ci); | |
1840 | else if (!is_delayed || force_requeue) | |
1841 | __cap_delay_requeue(mdsc, ci); | |
1842 | ||
1843 | spin_unlock(&ci->i_ceph_lock); | |
1844 | ||
1845 | if (queue_invalidate) | |
1846 | ceph_queue_invalidate(inode); | |
1847 | ||
1848 | if (session) | |
1849 | mutex_unlock(&session->s_mutex); | |
1850 | if (took_snap_rwsem) | |
1851 | up_read(&mdsc->snap_rwsem); | |
1852 | } | |
1853 | ||
1854 | /* | |
1855 | * Try to flush dirty caps back to the auth mds. | |
1856 | */ | |
1857 | static int try_flush_caps(struct inode *inode, u64 *ptid) | |
1858 | { | |
1859 | struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; | |
1860 | struct ceph_inode_info *ci = ceph_inode(inode); | |
1861 | struct ceph_mds_session *session = NULL; | |
1862 | int flushing = 0; | |
1863 | u64 flush_tid = 0, oldest_flush_tid = 0; | |
1864 | ||
1865 | retry: | |
1866 | spin_lock(&ci->i_ceph_lock); | |
1867 | if (ci->i_ceph_flags & CEPH_I_NOFLUSH) { | |
1868 | dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode); | |
1869 | goto out; | |
1870 | } | |
1871 | if (ci->i_dirty_caps && ci->i_auth_cap) { | |
1872 | struct ceph_cap *cap = ci->i_auth_cap; | |
1873 | int used = __ceph_caps_used(ci); | |
1874 | int want = __ceph_caps_wanted(ci); | |
1875 | int delayed; | |
1876 | ||
1877 | if (!session || session != cap->session) { | |
1878 | spin_unlock(&ci->i_ceph_lock); | |
1879 | if (session) | |
1880 | mutex_unlock(&session->s_mutex); | |
1881 | session = cap->session; | |
1882 | mutex_lock(&session->s_mutex); | |
1883 | goto retry; | |
1884 | } | |
1885 | if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) | |
1886 | goto out; | |
1887 | ||
1888 | flushing = __mark_caps_flushing(inode, session, &flush_tid, | |
1889 | &oldest_flush_tid); | |
1890 | ||
1891 | /* __send_cap drops i_ceph_lock */ | |
1892 | delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want, | |
1893 | (cap->issued | cap->implemented), | |
1894 | flushing, flush_tid, oldest_flush_tid); | |
1895 | ||
1896 | if (delayed) { | |
1897 | spin_lock(&ci->i_ceph_lock); | |
1898 | __cap_delay_requeue(mdsc, ci); | |
1899 | spin_unlock(&ci->i_ceph_lock); | |
1900 | } | |
1901 | } else { | |
1902 | if (!list_empty(&ci->i_cap_flush_list)) { | |
1903 | struct ceph_cap_flush *cf = | |
1904 | list_last_entry(&ci->i_cap_flush_list, | |
1905 | struct ceph_cap_flush, i_list); | |
1906 | flush_tid = cf->tid; | |
1907 | } | |
1908 | flushing = ci->i_flushing_caps; | |
1909 | spin_unlock(&ci->i_ceph_lock); | |
1910 | } | |
1911 | out: | |
1912 | if (session) | |
1913 | mutex_unlock(&session->s_mutex); | |
1914 | ||
1915 | *ptid = flush_tid; | |
1916 | return flushing; | |
1917 | } | |
1918 | ||
1919 | /* | |
1920 | * Return true if we've flushed caps through the given flush_tid. | |
1921 | */ | |
1922 | static int caps_are_flushed(struct inode *inode, u64 flush_tid) | |
1923 | { | |
1924 | struct ceph_inode_info *ci = ceph_inode(inode); | |
1925 | int ret = 1; | |
1926 | ||
1927 | spin_lock(&ci->i_ceph_lock); | |
1928 | if (!list_empty(&ci->i_cap_flush_list)) { | |
1929 | struct ceph_cap_flush * cf = | |
1930 | list_first_entry(&ci->i_cap_flush_list, | |
1931 | struct ceph_cap_flush, i_list); | |
1932 | if (cf->tid <= flush_tid) | |
1933 | ret = 0; | |
1934 | } | |
1935 | spin_unlock(&ci->i_ceph_lock); | |
1936 | return ret; | |
1937 | } | |
1938 | ||
1939 | /* | |
1940 | * wait for any unsafe requests to complete. | |
1941 | */ | |
1942 | static int unsafe_request_wait(struct inode *inode) | |
1943 | { | |
1944 | struct ceph_inode_info *ci = ceph_inode(inode); | |
1945 | struct ceph_mds_request *req1 = NULL, *req2 = NULL; | |
1946 | int ret, err = 0; | |
1947 | ||
1948 | spin_lock(&ci->i_unsafe_lock); | |
1949 | if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) { | |
1950 | req1 = list_last_entry(&ci->i_unsafe_dirops, | |
1951 | struct ceph_mds_request, | |
1952 | r_unsafe_dir_item); | |
1953 | ceph_mdsc_get_request(req1); | |
1954 | } | |
1955 | if (!list_empty(&ci->i_unsafe_iops)) { | |
1956 | req2 = list_last_entry(&ci->i_unsafe_iops, | |
1957 | struct ceph_mds_request, | |
1958 | r_unsafe_target_item); | |
1959 | ceph_mdsc_get_request(req2); | |
1960 | } | |
1961 | spin_unlock(&ci->i_unsafe_lock); | |
1962 | ||
1963 | dout("unsafe_requeset_wait %p wait on tid %llu %llu\n", | |
1964 | inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL); | |
1965 | if (req1) { | |
1966 | ret = !wait_for_completion_timeout(&req1->r_safe_completion, | |
1967 | ceph_timeout_jiffies(req1->r_timeout)); | |
1968 | if (ret) | |
1969 | err = -EIO; | |
1970 | ceph_mdsc_put_request(req1); | |
1971 | } | |
1972 | if (req2) { | |
1973 | ret = !wait_for_completion_timeout(&req2->r_safe_completion, | |
1974 | ceph_timeout_jiffies(req2->r_timeout)); | |
1975 | if (ret) | |
1976 | err = -EIO; | |
1977 | ceph_mdsc_put_request(req2); | |
1978 | } | |
1979 | return err; | |
1980 | } | |
1981 | ||
1982 | int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync) | |
1983 | { | |
1984 | struct inode *inode = file->f_mapping->host; | |
1985 | struct ceph_inode_info *ci = ceph_inode(inode); | |
1986 | u64 flush_tid; | |
1987 | int ret; | |
1988 | int dirty; | |
1989 | ||
1990 | dout("fsync %p%s\n", inode, datasync ? " datasync" : ""); | |
1991 | ||
1992 | ceph_sync_write_wait(inode); | |
1993 | ||
1994 | ret = filemap_write_and_wait_range(inode->i_mapping, start, end); | |
1995 | if (ret < 0) | |
1996 | goto out; | |
1997 | ||
1998 | if (datasync) | |
1999 | goto out; | |
2000 | ||
2001 | inode_lock(inode); | |
2002 | ||
2003 | dirty = try_flush_caps(inode, &flush_tid); | |
2004 | dout("fsync dirty caps are %s\n", ceph_cap_string(dirty)); | |
2005 | ||
2006 | ret = unsafe_request_wait(inode); | |
2007 | ||
2008 | /* | |
2009 | * only wait on non-file metadata writeback (the mds | |
2010 | * can recover size and mtime, so we don't need to | |
2011 | * wait for that) | |
2012 | */ | |
2013 | if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) { | |
2014 | ret = wait_event_interruptible(ci->i_cap_wq, | |
2015 | caps_are_flushed(inode, flush_tid)); | |
2016 | } | |
2017 | inode_unlock(inode); | |
2018 | out: | |
2019 | dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret); | |
2020 | return ret; | |
2021 | } | |
2022 | ||
2023 | /* | |
2024 | * Flush any dirty caps back to the mds. If we aren't asked to wait, | |
2025 | * queue inode for flush but don't do so immediately, because we can | |
2026 | * get by with fewer MDS messages if we wait for data writeback to | |
2027 | * complete first. | |
2028 | */ | |
2029 | int ceph_write_inode(struct inode *inode, struct writeback_control *wbc) | |
2030 | { | |
2031 | struct ceph_inode_info *ci = ceph_inode(inode); | |
2032 | u64 flush_tid; | |
2033 | int err = 0; | |
2034 | int dirty; | |
2035 | int wait = wbc->sync_mode == WB_SYNC_ALL; | |
2036 | ||
2037 | dout("write_inode %p wait=%d\n", inode, wait); | |
2038 | if (wait) { | |
2039 | dirty = try_flush_caps(inode, &flush_tid); | |
2040 | if (dirty) | |
2041 | err = wait_event_interruptible(ci->i_cap_wq, | |
2042 | caps_are_flushed(inode, flush_tid)); | |
2043 | } else { | |
2044 | struct ceph_mds_client *mdsc = | |
2045 | ceph_sb_to_client(inode->i_sb)->mdsc; | |
2046 | ||
2047 | spin_lock(&ci->i_ceph_lock); | |
2048 | if (__ceph_caps_dirty(ci)) | |
2049 | __cap_delay_requeue_front(mdsc, ci); | |
2050 | spin_unlock(&ci->i_ceph_lock); | |
2051 | } | |
2052 | return err; | |
2053 | } | |
2054 | ||
2055 | static void __kick_flushing_caps(struct ceph_mds_client *mdsc, | |
2056 | struct ceph_mds_session *session, | |
2057 | struct ceph_inode_info *ci, | |
2058 | u64 oldest_flush_tid) | |
2059 | __releases(ci->i_ceph_lock) | |
2060 | __acquires(ci->i_ceph_lock) | |
2061 | { | |
2062 | struct inode *inode = &ci->vfs_inode; | |
2063 | struct ceph_cap *cap; | |
2064 | struct ceph_cap_flush *cf; | |
2065 | int ret; | |
2066 | u64 first_tid = 0; | |
2067 | ||
2068 | list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) { | |
2069 | if (cf->tid < first_tid) | |
2070 | continue; | |
2071 | ||
2072 | cap = ci->i_auth_cap; | |
2073 | if (!(cap && cap->session == session)) { | |
2074 | pr_err("%p auth cap %p not mds%d ???\n", | |
2075 | inode, cap, session->s_mds); | |
2076 | break; | |
2077 | } | |
2078 | ||
2079 | first_tid = cf->tid + 1; | |
2080 | ||
2081 | if (cf->caps) { | |
2082 | dout("kick_flushing_caps %p cap %p tid %llu %s\n", | |
2083 | inode, cap, cf->tid, ceph_cap_string(cf->caps)); | |
2084 | ci->i_ceph_flags |= CEPH_I_NODELAY; | |
2085 | ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, | |
2086 | __ceph_caps_used(ci), | |
2087 | __ceph_caps_wanted(ci), | |
2088 | cap->issued | cap->implemented, | |
2089 | cf->caps, cf->tid, oldest_flush_tid); | |
2090 | if (ret) { | |
2091 | pr_err("kick_flushing_caps: error sending " | |
2092 | "cap flush, ino (%llx.%llx) " | |
2093 | "tid %llu flushing %s\n", | |
2094 | ceph_vinop(inode), cf->tid, | |
2095 | ceph_cap_string(cf->caps)); | |
2096 | } | |
2097 | } else { | |
2098 | struct ceph_cap_snap *capsnap = | |
2099 | container_of(cf, struct ceph_cap_snap, | |
2100 | cap_flush); | |
2101 | dout("kick_flushing_caps %p capsnap %p tid %llu %s\n", | |
2102 | inode, capsnap, cf->tid, | |
2103 | ceph_cap_string(capsnap->dirty)); | |
2104 | ||
2105 | atomic_inc(&capsnap->nref); | |
2106 | spin_unlock(&ci->i_ceph_lock); | |
2107 | ||
2108 | ret = __send_flush_snap(inode, session, capsnap, cap->mseq, | |
2109 | oldest_flush_tid); | |
2110 | if (ret < 0) { | |
2111 | pr_err("kick_flushing_caps: error sending " | |
2112 | "cap flushsnap, ino (%llx.%llx) " | |
2113 | "tid %llu follows %llu\n", | |
2114 | ceph_vinop(inode), cf->tid, | |
2115 | capsnap->follows); | |
2116 | } | |
2117 | ||
2118 | ceph_put_cap_snap(capsnap); | |
2119 | } | |
2120 | ||
2121 | spin_lock(&ci->i_ceph_lock); | |
2122 | } | |
2123 | } | |
2124 | ||
2125 | void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc, | |
2126 | struct ceph_mds_session *session) | |
2127 | { | |
2128 | struct ceph_inode_info *ci; | |
2129 | struct ceph_cap *cap; | |
2130 | u64 oldest_flush_tid; | |
2131 | ||
2132 | dout("early_kick_flushing_caps mds%d\n", session->s_mds); | |
2133 | ||
2134 | spin_lock(&mdsc->cap_dirty_lock); | |
2135 | oldest_flush_tid = __get_oldest_flush_tid(mdsc); | |
2136 | spin_unlock(&mdsc->cap_dirty_lock); | |
2137 | ||
2138 | list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { | |
2139 | spin_lock(&ci->i_ceph_lock); | |
2140 | cap = ci->i_auth_cap; | |
2141 | if (!(cap && cap->session == session)) { | |
2142 | pr_err("%p auth cap %p not mds%d ???\n", | |
2143 | &ci->vfs_inode, cap, session->s_mds); | |
2144 | spin_unlock(&ci->i_ceph_lock); | |
2145 | continue; | |
2146 | } | |
2147 | ||
2148 | ||
2149 | /* | |
2150 | * if flushing caps were revoked, we re-send the cap flush | |
2151 | * in client reconnect stage. This guarantees MDS * processes | |
2152 | * the cap flush message before issuing the flushing caps to | |
2153 | * other client. | |
2154 | */ | |
2155 | if ((cap->issued & ci->i_flushing_caps) != | |
2156 | ci->i_flushing_caps) { | |
2157 | ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; | |
2158 | __kick_flushing_caps(mdsc, session, ci, | |
2159 | oldest_flush_tid); | |
2160 | } else { | |
2161 | ci->i_ceph_flags |= CEPH_I_KICK_FLUSH; | |
2162 | } | |
2163 | ||
2164 | spin_unlock(&ci->i_ceph_lock); | |
2165 | } | |
2166 | } | |
2167 | ||
2168 | void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc, | |
2169 | struct ceph_mds_session *session) | |
2170 | { | |
2171 | struct ceph_inode_info *ci; | |
2172 | struct ceph_cap *cap; | |
2173 | u64 oldest_flush_tid; | |
2174 | ||
2175 | dout("kick_flushing_caps mds%d\n", session->s_mds); | |
2176 | ||
2177 | spin_lock(&mdsc->cap_dirty_lock); | |
2178 | oldest_flush_tid = __get_oldest_flush_tid(mdsc); | |
2179 | spin_unlock(&mdsc->cap_dirty_lock); | |
2180 | ||
2181 | list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { | |
2182 | spin_lock(&ci->i_ceph_lock); | |
2183 | cap = ci->i_auth_cap; | |
2184 | if (!(cap && cap->session == session)) { | |
2185 | pr_err("%p auth cap %p not mds%d ???\n", | |
2186 | &ci->vfs_inode, cap, session->s_mds); | |
2187 | spin_unlock(&ci->i_ceph_lock); | |
2188 | continue; | |
2189 | } | |
2190 | if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) { | |
2191 | ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; | |
2192 | __kick_flushing_caps(mdsc, session, ci, | |
2193 | oldest_flush_tid); | |
2194 | } | |
2195 | spin_unlock(&ci->i_ceph_lock); | |
2196 | } | |
2197 | } | |
2198 | ||
2199 | static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc, | |
2200 | struct ceph_mds_session *session, | |
2201 | struct inode *inode) | |
2202 | __releases(ci->i_ceph_lock) | |
2203 | { | |
2204 | struct ceph_inode_info *ci = ceph_inode(inode); | |
2205 | struct ceph_cap *cap; | |
2206 | ||
2207 | cap = ci->i_auth_cap; | |
2208 | dout("kick_flushing_inode_caps %p flushing %s\n", inode, | |
2209 | ceph_cap_string(ci->i_flushing_caps)); | |
2210 | ||
2211 | if (!list_empty(&ci->i_cap_flush_list)) { | |
2212 | u64 oldest_flush_tid; | |
2213 | spin_lock(&mdsc->cap_dirty_lock); | |
2214 | list_move_tail(&ci->i_flushing_item, | |
2215 | &cap->session->s_cap_flushing); | |
2216 | oldest_flush_tid = __get_oldest_flush_tid(mdsc); | |
2217 | spin_unlock(&mdsc->cap_dirty_lock); | |
2218 | ||
2219 | ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH; | |
2220 | __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid); | |
2221 | spin_unlock(&ci->i_ceph_lock); | |
2222 | } else { | |
2223 | spin_unlock(&ci->i_ceph_lock); | |
2224 | } | |
2225 | } | |
2226 | ||
2227 | ||
2228 | /* | |
2229 | * Take references to capabilities we hold, so that we don't release | |
2230 | * them to the MDS prematurely. | |
2231 | * | |
2232 | * Protected by i_ceph_lock. | |
2233 | */ | |
2234 | static void __take_cap_refs(struct ceph_inode_info *ci, int got, | |
2235 | bool snap_rwsem_locked) | |
2236 | { | |
2237 | if (got & CEPH_CAP_PIN) | |
2238 | ci->i_pin_ref++; | |
2239 | if (got & CEPH_CAP_FILE_RD) | |
2240 | ci->i_rd_ref++; | |
2241 | if (got & CEPH_CAP_FILE_CACHE) | |
2242 | ci->i_rdcache_ref++; | |
2243 | if (got & CEPH_CAP_FILE_WR) { | |
2244 | if (ci->i_wr_ref == 0 && !ci->i_head_snapc) { | |
2245 | BUG_ON(!snap_rwsem_locked); | |
2246 | ci->i_head_snapc = ceph_get_snap_context( | |
2247 | ci->i_snap_realm->cached_context); | |
2248 | } | |
2249 | ci->i_wr_ref++; | |
2250 | } | |
2251 | if (got & CEPH_CAP_FILE_BUFFER) { | |
2252 | if (ci->i_wb_ref == 0) | |
2253 | ihold(&ci->vfs_inode); | |
2254 | ci->i_wb_ref++; | |
2255 | dout("__take_cap_refs %p wb %d -> %d (?)\n", | |
2256 | &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref); | |
2257 | } | |
2258 | } | |
2259 | ||
2260 | /* | |
2261 | * Try to grab cap references. Specify those refs we @want, and the | |
2262 | * minimal set we @need. Also include the larger offset we are writing | |
2263 | * to (when applicable), and check against max_size here as well. | |
2264 | * Note that caller is responsible for ensuring max_size increases are | |
2265 | * requested from the MDS. | |
2266 | */ | |
2267 | static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want, | |
2268 | loff_t endoff, bool nonblock, int *got, int *err) | |
2269 | { | |
2270 | struct inode *inode = &ci->vfs_inode; | |
2271 | struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; | |
2272 | int ret = 0; | |
2273 | int have, implemented; | |
2274 | int file_wanted; | |
2275 | bool snap_rwsem_locked = false; | |
2276 | ||
2277 | dout("get_cap_refs %p need %s want %s\n", inode, | |
2278 | ceph_cap_string(need), ceph_cap_string(want)); | |
2279 | ||
2280 | again: | |
2281 | spin_lock(&ci->i_ceph_lock); | |
2282 | ||
2283 | /* make sure file is actually open */ | |
2284 | file_wanted = __ceph_caps_file_wanted(ci); | |
2285 | if ((file_wanted & need) != need) { | |
2286 | dout("try_get_cap_refs need %s file_wanted %s, EBADF\n", | |
2287 | ceph_cap_string(need), ceph_cap_string(file_wanted)); | |
2288 | *err = -EBADF; | |
2289 | ret = 1; | |
2290 | goto out_unlock; | |
2291 | } | |
2292 | ||
2293 | /* finish pending truncate */ | |
2294 | while (ci->i_truncate_pending) { | |
2295 | spin_unlock(&ci->i_ceph_lock); | |
2296 | if (snap_rwsem_locked) { | |
2297 | up_read(&mdsc->snap_rwsem); | |
2298 | snap_rwsem_locked = false; | |
2299 | } | |
2300 | __ceph_do_pending_vmtruncate(inode); | |
2301 | spin_lock(&ci->i_ceph_lock); | |
2302 | } | |
2303 | ||
2304 | have = __ceph_caps_issued(ci, &implemented); | |
2305 | ||
2306 | if (have & need & CEPH_CAP_FILE_WR) { | |
2307 | if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) { | |
2308 | dout("get_cap_refs %p endoff %llu > maxsize %llu\n", | |
2309 | inode, endoff, ci->i_max_size); | |
2310 | if (endoff > ci->i_requested_max_size) { | |
2311 | *err = -EAGAIN; | |
2312 | ret = 1; | |
2313 | } | |
2314 | goto out_unlock; | |
2315 | } | |
2316 | /* | |
2317 | * If a sync write is in progress, we must wait, so that we | |
2318 | * can get a final snapshot value for size+mtime. | |
2319 | */ | |
2320 | if (__ceph_have_pending_cap_snap(ci)) { | |
2321 | dout("get_cap_refs %p cap_snap_pending\n", inode); | |
2322 | goto out_unlock; | |
2323 | } | |
2324 | } | |
2325 | ||
2326 | if ((have & need) == need) { | |
2327 | /* | |
2328 | * Look at (implemented & ~have & not) so that we keep waiting | |
2329 | * on transition from wanted -> needed caps. This is needed | |
2330 | * for WRBUFFER|WR -> WR to avoid a new WR sync write from | |
2331 | * going before a prior buffered writeback happens. | |
2332 | */ | |
2333 | int not = want & ~(have & need); | |
2334 | int revoking = implemented & ~have; | |
2335 | dout("get_cap_refs %p have %s but not %s (revoking %s)\n", | |
2336 | inode, ceph_cap_string(have), ceph_cap_string(not), | |
2337 | ceph_cap_string(revoking)); | |
2338 | if ((revoking & not) == 0) { | |
2339 | if (!snap_rwsem_locked && | |
2340 | !ci->i_head_snapc && | |
2341 | (need & CEPH_CAP_FILE_WR)) { | |
2342 | if (!down_read_trylock(&mdsc->snap_rwsem)) { | |
2343 | /* | |
2344 | * we can not call down_read() when | |
2345 | * task isn't in TASK_RUNNING state | |
2346 | */ | |
2347 | if (nonblock) { | |
2348 | *err = -EAGAIN; | |
2349 | ret = 1; | |
2350 | goto out_unlock; | |
2351 | } | |
2352 | ||
2353 | spin_unlock(&ci->i_ceph_lock); | |
2354 | down_read(&mdsc->snap_rwsem); | |
2355 | snap_rwsem_locked = true; | |
2356 | goto again; | |
2357 | } | |
2358 | snap_rwsem_locked = true; | |
2359 | } | |
2360 | *got = need | (have & want); | |
2361 | if ((need & CEPH_CAP_FILE_RD) && | |
2362 | !(*got & CEPH_CAP_FILE_CACHE)) | |
2363 | ceph_disable_fscache_readpage(ci); | |
2364 | __take_cap_refs(ci, *got, true); | |
2365 | ret = 1; | |
2366 | } | |
2367 | } else { | |
2368 | int session_readonly = false; | |
2369 | if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) { | |
2370 | struct ceph_mds_session *s = ci->i_auth_cap->session; | |
2371 | spin_lock(&s->s_cap_lock); | |
2372 | session_readonly = s->s_readonly; | |
2373 | spin_unlock(&s->s_cap_lock); | |
2374 | } | |
2375 | if (session_readonly) { | |
2376 | dout("get_cap_refs %p needed %s but mds%d readonly\n", | |
2377 | inode, ceph_cap_string(need), ci->i_auth_cap->mds); | |
2378 | *err = -EROFS; | |
2379 | ret = 1; | |
2380 | goto out_unlock; | |
2381 | } | |
2382 | ||
2383 | if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) { | |
2384 | int mds_wanted; | |
2385 | if (ACCESS_ONCE(mdsc->fsc->mount_state) == | |
2386 | CEPH_MOUNT_SHUTDOWN) { | |
2387 | dout("get_cap_refs %p forced umount\n", inode); | |
2388 | *err = -EIO; | |
2389 | ret = 1; | |
2390 | goto out_unlock; | |
2391 | } | |
2392 | mds_wanted = __ceph_caps_mds_wanted(ci); | |
2393 | if ((mds_wanted & need) != need) { | |
2394 | dout("get_cap_refs %p caps were dropped" | |
2395 | " (session killed?)\n", inode); | |
2396 | *err = -ESTALE; | |
2397 | ret = 1; | |
2398 | goto out_unlock; | |
2399 | } | |
2400 | if ((mds_wanted & file_wanted) == | |
2401 | (file_wanted & (CEPH_CAP_FILE_RD|CEPH_CAP_FILE_WR))) | |
2402 | ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED; | |
2403 | } | |
2404 | ||
2405 | dout("get_cap_refs %p have %s needed %s\n", inode, | |
2406 | ceph_cap_string(have), ceph_cap_string(need)); | |
2407 | } | |
2408 | out_unlock: | |
2409 | spin_unlock(&ci->i_ceph_lock); | |
2410 | if (snap_rwsem_locked) | |
2411 | up_read(&mdsc->snap_rwsem); | |
2412 | ||
2413 | dout("get_cap_refs %p ret %d got %s\n", inode, | |
2414 | ret, ceph_cap_string(*got)); | |
2415 | return ret; | |
2416 | } | |
2417 | ||
2418 | /* | |
2419 | * Check the offset we are writing up to against our current | |
2420 | * max_size. If necessary, tell the MDS we want to write to | |
2421 | * a larger offset. | |
2422 | */ | |
2423 | static void check_max_size(struct inode *inode, loff_t endoff) | |
2424 | { | |
2425 | struct ceph_inode_info *ci = ceph_inode(inode); | |
2426 | int check = 0; | |
2427 | ||
2428 | /* do we need to explicitly request a larger max_size? */ | |
2429 | spin_lock(&ci->i_ceph_lock); | |
2430 | if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) { | |
2431 | dout("write %p at large endoff %llu, req max_size\n", | |
2432 | inode, endoff); | |
2433 | ci->i_wanted_max_size = endoff; | |
2434 | } | |
2435 | /* duplicate ceph_check_caps()'s logic */ | |
2436 | if (ci->i_auth_cap && | |
2437 | (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) && | |
2438 | ci->i_wanted_max_size > ci->i_max_size && | |
2439 | ci->i_wanted_max_size > ci->i_requested_max_size) | |
2440 | check = 1; | |
2441 | spin_unlock(&ci->i_ceph_lock); | |
2442 | if (check) | |
2443 | ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); | |
2444 | } | |
2445 | ||
2446 | /* | |
2447 | * Wait for caps, and take cap references. If we can't get a WR cap | |
2448 | * due to a small max_size, make sure we check_max_size (and possibly | |
2449 | * ask the mds) so we don't get hung up indefinitely. | |
2450 | */ | |
2451 | int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, | |
2452 | loff_t endoff, int *got, struct page **pinned_page) | |
2453 | { | |
2454 | int _got, ret, err = 0; | |
2455 | ||
2456 | ret = ceph_pool_perm_check(ci, need); | |
2457 | if (ret < 0) | |
2458 | return ret; | |
2459 | ||
2460 | while (true) { | |
2461 | if (endoff > 0) | |
2462 | check_max_size(&ci->vfs_inode, endoff); | |
2463 | ||
2464 | err = 0; | |
2465 | _got = 0; | |
2466 | ret = try_get_cap_refs(ci, need, want, endoff, | |
2467 | false, &_got, &err); | |
2468 | if (ret) { | |
2469 | if (err == -EAGAIN) | |
2470 | continue; | |
2471 | if (err < 0) | |
2472 | ret = err; | |
2473 | } else { | |
2474 | ret = wait_event_interruptible(ci->i_cap_wq, | |
2475 | try_get_cap_refs(ci, need, want, endoff, | |
2476 | true, &_got, &err)); | |
2477 | if (err == -EAGAIN) | |
2478 | continue; | |
2479 | if (err < 0) | |
2480 | ret = err; | |
2481 | } | |
2482 | if (ret < 0) { | |
2483 | if (err == -ESTALE) { | |
2484 | /* session was killed, try renew caps */ | |
2485 | ret = ceph_renew_caps(&ci->vfs_inode); | |
2486 | if (ret == 0) | |
2487 | continue; | |
2488 | } | |
2489 | return ret; | |
2490 | } | |
2491 | ||
2492 | if (ci->i_inline_version != CEPH_INLINE_NONE && | |
2493 | (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) && | |
2494 | i_size_read(&ci->vfs_inode) > 0) { | |
2495 | struct page *page = | |
2496 | find_get_page(ci->vfs_inode.i_mapping, 0); | |
2497 | if (page) { | |
2498 | if (PageUptodate(page)) { | |
2499 | *pinned_page = page; | |
2500 | break; | |
2501 | } | |
2502 | put_page(page); | |
2503 | } | |
2504 | /* | |
2505 | * drop cap refs first because getattr while | |
2506 | * holding * caps refs can cause deadlock. | |
2507 | */ | |
2508 | ceph_put_cap_refs(ci, _got); | |
2509 | _got = 0; | |
2510 | ||
2511 | /* | |
2512 | * getattr request will bring inline data into | |
2513 | * page cache | |
2514 | */ | |
2515 | ret = __ceph_do_getattr(&ci->vfs_inode, NULL, | |
2516 | CEPH_STAT_CAP_INLINE_DATA, | |
2517 | true); | |
2518 | if (ret < 0) | |
2519 | return ret; | |
2520 | continue; | |
2521 | } | |
2522 | break; | |
2523 | } | |
2524 | ||
2525 | if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE)) | |
2526 | ceph_fscache_revalidate_cookie(ci); | |
2527 | ||
2528 | *got = _got; | |
2529 | return 0; | |
2530 | } | |
2531 | ||
2532 | /* | |
2533 | * Take cap refs. Caller must already know we hold at least one ref | |
2534 | * on the caps in question or we don't know this is safe. | |
2535 | */ | |
2536 | void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps) | |
2537 | { | |
2538 | spin_lock(&ci->i_ceph_lock); | |
2539 | __take_cap_refs(ci, caps, false); | |
2540 | spin_unlock(&ci->i_ceph_lock); | |
2541 | } | |
2542 | ||
2543 | ||
2544 | /* | |
2545 | * drop cap_snap that is not associated with any snapshot. | |
2546 | * we don't need to send FLUSHSNAP message for it. | |
2547 | */ | |
2548 | static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci, | |
2549 | struct ceph_cap_snap *capsnap) | |
2550 | { | |
2551 | if (!capsnap->need_flush && | |
2552 | !capsnap->writing && !capsnap->dirty_pages) { | |
2553 | dout("dropping cap_snap %p follows %llu\n", | |
2554 | capsnap, capsnap->follows); | |
2555 | BUG_ON(capsnap->cap_flush.tid > 0); | |
2556 | ceph_put_snap_context(capsnap->context); | |
2557 | if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps)) | |
2558 | ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS; | |
2559 | ||
2560 | list_del(&capsnap->ci_item); | |
2561 | ceph_put_cap_snap(capsnap); | |
2562 | return 1; | |
2563 | } | |
2564 | return 0; | |
2565 | } | |
2566 | ||
2567 | /* | |
2568 | * Release cap refs. | |
2569 | * | |
2570 | * If we released the last ref on any given cap, call ceph_check_caps | |
2571 | * to release (or schedule a release). | |
2572 | * | |
2573 | * If we are releasing a WR cap (from a sync write), finalize any affected | |
2574 | * cap_snap, and wake up any waiters. | |
2575 | */ | |
2576 | void ceph_put_cap_refs(struct ceph_inode_info *ci, int had) | |
2577 | { | |
2578 | struct inode *inode = &ci->vfs_inode; | |
2579 | int last = 0, put = 0, flushsnaps = 0, wake = 0; | |
2580 | ||
2581 | spin_lock(&ci->i_ceph_lock); | |
2582 | if (had & CEPH_CAP_PIN) | |
2583 | --ci->i_pin_ref; | |
2584 | if (had & CEPH_CAP_FILE_RD) | |
2585 | if (--ci->i_rd_ref == 0) | |
2586 | last++; | |
2587 | if (had & CEPH_CAP_FILE_CACHE) | |
2588 | if (--ci->i_rdcache_ref == 0) | |
2589 | last++; | |
2590 | if (had & CEPH_CAP_FILE_BUFFER) { | |
2591 | if (--ci->i_wb_ref == 0) { | |
2592 | last++; | |
2593 | put++; | |
2594 | } | |
2595 | dout("put_cap_refs %p wb %d -> %d (?)\n", | |
2596 | inode, ci->i_wb_ref+1, ci->i_wb_ref); | |
2597 | } | |
2598 | if (had & CEPH_CAP_FILE_WR) | |
2599 | if (--ci->i_wr_ref == 0) { | |
2600 | last++; | |
2601 | if (__ceph_have_pending_cap_snap(ci)) { | |
2602 | struct ceph_cap_snap *capsnap = | |
2603 | list_last_entry(&ci->i_cap_snaps, | |
2604 | struct ceph_cap_snap, | |
2605 | ci_item); | |
2606 | capsnap->writing = 0; | |
2607 | if (ceph_try_drop_cap_snap(ci, capsnap)) | |
2608 | put++; | |
2609 | else if (__ceph_finish_cap_snap(ci, capsnap)) | |
2610 | flushsnaps = 1; | |
2611 | wake = 1; | |
2612 | } | |
2613 | if (ci->i_wrbuffer_ref_head == 0 && | |
2614 | ci->i_dirty_caps == 0 && | |
2615 | ci->i_flushing_caps == 0) { | |
2616 | BUG_ON(!ci->i_head_snapc); | |
2617 | ceph_put_snap_context(ci->i_head_snapc); | |
2618 | ci->i_head_snapc = NULL; | |
2619 | } | |
2620 | /* see comment in __ceph_remove_cap() */ | |
2621 | if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) | |
2622 | drop_inode_snap_realm(ci); | |
2623 | } | |
2624 | spin_unlock(&ci->i_ceph_lock); | |
2625 | ||
2626 | dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had), | |
2627 | last ? " last" : "", put ? " put" : ""); | |
2628 | ||
2629 | if (last && !flushsnaps) | |
2630 | ceph_check_caps(ci, 0, NULL); | |
2631 | else if (flushsnaps) | |
2632 | ceph_flush_snaps(ci, NULL); | |
2633 | if (wake) | |
2634 | wake_up_all(&ci->i_cap_wq); | |
2635 | while (put-- > 0) | |
2636 | iput(inode); | |
2637 | } | |
2638 | ||
2639 | /* | |
2640 | * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap | |
2641 | * context. Adjust per-snap dirty page accounting as appropriate. | |
2642 | * Once all dirty data for a cap_snap is flushed, flush snapped file | |
2643 | * metadata back to the MDS. If we dropped the last ref, call | |
2644 | * ceph_check_caps. | |
2645 | */ | |
2646 | void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr, | |
2647 | struct ceph_snap_context *snapc) | |
2648 | { | |
2649 | struct inode *inode = &ci->vfs_inode; | |
2650 | struct ceph_cap_snap *capsnap = NULL; | |
2651 | int put = 0; | |
2652 | bool last = false; | |
2653 | bool found = false; | |
2654 | bool flush_snaps = false; | |
2655 | bool complete_capsnap = false; | |
2656 | ||
2657 | spin_lock(&ci->i_ceph_lock); | |
2658 | ci->i_wrbuffer_ref -= nr; | |
2659 | if (ci->i_wrbuffer_ref == 0) { | |
2660 | last = true; | |
2661 | put++; | |
2662 | } | |
2663 | ||
2664 | if (ci->i_head_snapc == snapc) { | |
2665 | ci->i_wrbuffer_ref_head -= nr; | |
2666 | if (ci->i_wrbuffer_ref_head == 0 && | |
2667 | ci->i_wr_ref == 0 && | |
2668 | ci->i_dirty_caps == 0 && | |
2669 | ci->i_flushing_caps == 0) { | |
2670 | BUG_ON(!ci->i_head_snapc); | |
2671 | ceph_put_snap_context(ci->i_head_snapc); | |
2672 | ci->i_head_snapc = NULL; | |
2673 | } | |
2674 | dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n", | |
2675 | inode, | |
2676 | ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr, | |
2677 | ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head, | |
2678 | last ? " LAST" : ""); | |
2679 | } else { | |
2680 | list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { | |
2681 | if (capsnap->context == snapc) { | |
2682 | found = true; | |
2683 | break; | |
2684 | } | |
2685 | } | |
2686 | BUG_ON(!found); | |
2687 | capsnap->dirty_pages -= nr; | |
2688 | if (capsnap->dirty_pages == 0) { | |
2689 | complete_capsnap = true; | |
2690 | if (!capsnap->writing) { | |
2691 | if (ceph_try_drop_cap_snap(ci, capsnap)) { | |
2692 | put++; | |
2693 | } else { | |
2694 | ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS; | |
2695 | flush_snaps = true; | |
2696 | } | |
2697 | } | |
2698 | } | |
2699 | dout("put_wrbuffer_cap_refs on %p cap_snap %p " | |
2700 | " snap %lld %d/%d -> %d/%d %s%s\n", | |
2701 | inode, capsnap, capsnap->context->seq, | |
2702 | ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr, | |
2703 | ci->i_wrbuffer_ref, capsnap->dirty_pages, | |
2704 | last ? " (wrbuffer last)" : "", | |
2705 | complete_capsnap ? " (complete capsnap)" : ""); | |
2706 | } | |
2707 | ||
2708 | spin_unlock(&ci->i_ceph_lock); | |
2709 | ||
2710 | if (last) { | |
2711 | ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL); | |
2712 | } else if (flush_snaps) { | |
2713 | ceph_flush_snaps(ci, NULL); | |
2714 | } | |
2715 | if (complete_capsnap) | |
2716 | wake_up_all(&ci->i_cap_wq); | |
2717 | while (put-- > 0) | |
2718 | iput(inode); | |
2719 | } | |
2720 | ||
2721 | /* | |
2722 | * Invalidate unlinked inode's aliases, so we can drop the inode ASAP. | |
2723 | */ | |
2724 | static void invalidate_aliases(struct inode *inode) | |
2725 | { | |
2726 | struct dentry *dn, *prev = NULL; | |
2727 | ||
2728 | dout("invalidate_aliases inode %p\n", inode); | |
2729 | d_prune_aliases(inode); | |
2730 | /* | |
2731 | * For non-directory inode, d_find_alias() only returns | |
2732 | * hashed dentry. After calling d_invalidate(), the | |
2733 | * dentry becomes unhashed. | |
2734 | * | |
2735 | * For directory inode, d_find_alias() can return | |
2736 | * unhashed dentry. But directory inode should have | |
2737 | * one alias at most. | |
2738 | */ | |
2739 | while ((dn = d_find_alias(inode))) { | |
2740 | if (dn == prev) { | |
2741 | dput(dn); | |
2742 | break; | |
2743 | } | |
2744 | d_invalidate(dn); | |
2745 | if (prev) | |
2746 | dput(prev); | |
2747 | prev = dn; | |
2748 | } | |
2749 | if (prev) | |
2750 | dput(prev); | |
2751 | } | |
2752 | ||
2753 | /* | |
2754 | * Handle a cap GRANT message from the MDS. (Note that a GRANT may | |
2755 | * actually be a revocation if it specifies a smaller cap set.) | |
2756 | * | |
2757 | * caller holds s_mutex and i_ceph_lock, we drop both. | |
2758 | */ | |
2759 | static void handle_cap_grant(struct ceph_mds_client *mdsc, | |
2760 | struct inode *inode, struct ceph_mds_caps *grant, | |
2761 | struct ceph_string **pns, u64 inline_version, | |
2762 | void *inline_data, u32 inline_len, | |
2763 | struct ceph_buffer *xattr_buf, | |
2764 | struct ceph_mds_session *session, | |
2765 | struct ceph_cap *cap, int issued) | |
2766 | __releases(ci->i_ceph_lock) | |
2767 | __releases(mdsc->snap_rwsem) | |
2768 | { | |
2769 | struct ceph_inode_info *ci = ceph_inode(inode); | |
2770 | int mds = session->s_mds; | |
2771 | int seq = le32_to_cpu(grant->seq); | |
2772 | int newcaps = le32_to_cpu(grant->caps); | |
2773 | int used, wanted, dirty; | |
2774 | u64 size = le64_to_cpu(grant->size); | |
2775 | u64 max_size = le64_to_cpu(grant->max_size); | |
2776 | struct timespec mtime, atime, ctime; | |
2777 | int check_caps = 0; | |
2778 | bool wake = false; | |
2779 | bool writeback = false; | |
2780 | bool queue_trunc = false; | |
2781 | bool queue_invalidate = false; | |
2782 | bool deleted_inode = false; | |
2783 | bool fill_inline = false; | |
2784 | ||
2785 | dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n", | |
2786 | inode, cap, mds, seq, ceph_cap_string(newcaps)); | |
2787 | dout(" size %llu max_size %llu, i_size %llu\n", size, max_size, | |
2788 | inode->i_size); | |
2789 | ||
2790 | ||
2791 | /* | |
2792 | * auth mds of the inode changed. we received the cap export message, | |
2793 | * but still haven't received the cap import message. handle_cap_export | |
2794 | * updated the new auth MDS' cap. | |
2795 | * | |
2796 | * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message | |
2797 | * that was sent before the cap import message. So don't remove caps. | |
2798 | */ | |
2799 | if (ceph_seq_cmp(seq, cap->seq) <= 0) { | |
2800 | WARN_ON(cap != ci->i_auth_cap); | |
2801 | WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id)); | |
2802 | seq = cap->seq; | |
2803 | newcaps |= cap->issued; | |
2804 | } | |
2805 | ||
2806 | /* | |
2807 | * If CACHE is being revoked, and we have no dirty buffers, | |
2808 | * try to invalidate (once). (If there are dirty buffers, we | |
2809 | * will invalidate _after_ writeback.) | |
2810 | */ | |
2811 | if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */ | |
2812 | ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) && | |
2813 | (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 && | |
2814 | !(ci->i_wrbuffer_ref || ci->i_wb_ref)) { | |
2815 | if (try_nonblocking_invalidate(inode)) { | |
2816 | /* there were locked pages.. invalidate later | |
2817 | in a separate thread. */ | |
2818 | if (ci->i_rdcache_revoking != ci->i_rdcache_gen) { | |
2819 | queue_invalidate = true; | |
2820 | ci->i_rdcache_revoking = ci->i_rdcache_gen; | |
2821 | } | |
2822 | } | |
2823 | } | |
2824 | ||
2825 | /* side effects now are allowed */ | |
2826 | cap->cap_gen = session->s_cap_gen; | |
2827 | cap->seq = seq; | |
2828 | ||
2829 | __check_cap_issue(ci, cap, newcaps); | |
2830 | ||
2831 | if ((newcaps & CEPH_CAP_AUTH_SHARED) && | |
2832 | (issued & CEPH_CAP_AUTH_EXCL) == 0) { | |
2833 | inode->i_mode = le32_to_cpu(grant->mode); | |
2834 | inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid)); | |
2835 | inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid)); | |
2836 | dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode, | |
2837 | from_kuid(&init_user_ns, inode->i_uid), | |
2838 | from_kgid(&init_user_ns, inode->i_gid)); | |
2839 | } | |
2840 | ||
2841 | if ((newcaps & CEPH_CAP_AUTH_SHARED) && | |
2842 | (issued & CEPH_CAP_LINK_EXCL) == 0) { | |
2843 | set_nlink(inode, le32_to_cpu(grant->nlink)); | |
2844 | if (inode->i_nlink == 0 && | |
2845 | (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL))) | |
2846 | deleted_inode = true; | |
2847 | } | |
2848 | ||
2849 | if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) { | |
2850 | int len = le32_to_cpu(grant->xattr_len); | |
2851 | u64 version = le64_to_cpu(grant->xattr_version); | |
2852 | ||
2853 | if (version > ci->i_xattrs.version) { | |
2854 | dout(" got new xattrs v%llu on %p len %d\n", | |
2855 | version, inode, len); | |
2856 | if (ci->i_xattrs.blob) | |
2857 | ceph_buffer_put(ci->i_xattrs.blob); | |
2858 | ci->i_xattrs.blob = ceph_buffer_get(xattr_buf); | |
2859 | ci->i_xattrs.version = version; | |
2860 | ceph_forget_all_cached_acls(inode); | |
2861 | } | |
2862 | } | |
2863 | ||
2864 | if (newcaps & CEPH_CAP_ANY_RD) { | |
2865 | /* ctime/mtime/atime? */ | |
2866 | ceph_decode_timespec(&mtime, &grant->mtime); | |
2867 | ceph_decode_timespec(&atime, &grant->atime); | |
2868 | ceph_decode_timespec(&ctime, &grant->ctime); | |
2869 | ceph_fill_file_time(inode, issued, | |
2870 | le32_to_cpu(grant->time_warp_seq), | |
2871 | &ctime, &mtime, &atime); | |
2872 | } | |
2873 | ||
2874 | if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) { | |
2875 | /* file layout may have changed */ | |
2876 | s64 old_pool = ci->i_layout.pool_id; | |
2877 | struct ceph_string *old_ns; | |
2878 | ||
2879 | ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout); | |
2880 | old_ns = rcu_dereference_protected(ci->i_layout.pool_ns, | |
2881 | lockdep_is_held(&ci->i_ceph_lock)); | |
2882 | rcu_assign_pointer(ci->i_layout.pool_ns, *pns); | |
2883 | ||
2884 | if (ci->i_layout.pool_id != old_pool || *pns != old_ns) | |
2885 | ci->i_ceph_flags &= ~CEPH_I_POOL_PERM; | |
2886 | ||
2887 | *pns = old_ns; | |
2888 | ||
2889 | /* size/truncate_seq? */ | |
2890 | queue_trunc = ceph_fill_file_size(inode, issued, | |
2891 | le32_to_cpu(grant->truncate_seq), | |
2892 | le64_to_cpu(grant->truncate_size), | |
2893 | size); | |
2894 | /* max size increase? */ | |
2895 | if (ci->i_auth_cap == cap && max_size != ci->i_max_size) { | |
2896 | dout("max_size %lld -> %llu\n", | |
2897 | ci->i_max_size, max_size); | |
2898 | ci->i_max_size = max_size; | |
2899 | if (max_size >= ci->i_wanted_max_size) { | |
2900 | ci->i_wanted_max_size = 0; /* reset */ | |
2901 | ci->i_requested_max_size = 0; | |
2902 | } | |
2903 | wake = true; | |
2904 | } | |
2905 | } | |
2906 | ||
2907 | /* check cap bits */ | |
2908 | wanted = __ceph_caps_wanted(ci); | |
2909 | used = __ceph_caps_used(ci); | |
2910 | dirty = __ceph_caps_dirty(ci); | |
2911 | dout(" my wanted = %s, used = %s, dirty %s\n", | |
2912 | ceph_cap_string(wanted), | |
2913 | ceph_cap_string(used), | |
2914 | ceph_cap_string(dirty)); | |
2915 | if (wanted != le32_to_cpu(grant->wanted)) { | |
2916 | dout("mds wanted %s -> %s\n", | |
2917 | ceph_cap_string(le32_to_cpu(grant->wanted)), | |
2918 | ceph_cap_string(wanted)); | |
2919 | /* imported cap may not have correct mds_wanted */ | |
2920 | if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) | |
2921 | check_caps = 1; | |
2922 | } | |
2923 | ||
2924 | /* revocation, grant, or no-op? */ | |
2925 | if (cap->issued & ~newcaps) { | |
2926 | int revoking = cap->issued & ~newcaps; | |
2927 | ||
2928 | dout("revocation: %s -> %s (revoking %s)\n", | |
2929 | ceph_cap_string(cap->issued), | |
2930 | ceph_cap_string(newcaps), | |
2931 | ceph_cap_string(revoking)); | |
2932 | if (revoking & used & CEPH_CAP_FILE_BUFFER) | |
2933 | writeback = true; /* initiate writeback; will delay ack */ | |
2934 | else if (revoking == CEPH_CAP_FILE_CACHE && | |
2935 | (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 && | |
2936 | queue_invalidate) | |
2937 | ; /* do nothing yet, invalidation will be queued */ | |
2938 | else if (cap == ci->i_auth_cap) | |
2939 | check_caps = 1; /* check auth cap only */ | |
2940 | else | |
2941 | check_caps = 2; /* check all caps */ | |
2942 | cap->issued = newcaps; | |
2943 | cap->implemented |= newcaps; | |
2944 | } else if (cap->issued == newcaps) { | |
2945 | dout("caps unchanged: %s -> %s\n", | |
2946 | ceph_cap_string(cap->issued), ceph_cap_string(newcaps)); | |
2947 | } else { | |
2948 | dout("grant: %s -> %s\n", ceph_cap_string(cap->issued), | |
2949 | ceph_cap_string(newcaps)); | |
2950 | /* non-auth MDS is revoking the newly grant caps ? */ | |
2951 | if (cap == ci->i_auth_cap && | |
2952 | __ceph_caps_revoking_other(ci, cap, newcaps)) | |
2953 | check_caps = 2; | |
2954 | ||
2955 | cap->issued = newcaps; | |
2956 | cap->implemented |= newcaps; /* add bits only, to | |
2957 | * avoid stepping on a | |
2958 | * pending revocation */ | |
2959 | wake = true; | |
2960 | } | |
2961 | BUG_ON(cap->issued & ~cap->implemented); | |
2962 | ||
2963 | if (inline_version > 0 && inline_version >= ci->i_inline_version) { | |
2964 | ci->i_inline_version = inline_version; | |
2965 | if (ci->i_inline_version != CEPH_INLINE_NONE && | |
2966 | (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO))) | |
2967 | fill_inline = true; | |
2968 | } | |
2969 | ||
2970 | if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) { | |
2971 | if (newcaps & ~issued) | |
2972 | wake = true; | |
2973 | kick_flushing_inode_caps(mdsc, session, inode); | |
2974 | up_read(&mdsc->snap_rwsem); | |
2975 | } else { | |
2976 | spin_unlock(&ci->i_ceph_lock); | |
2977 | } | |
2978 | ||
2979 | if (fill_inline) | |
2980 | ceph_fill_inline_data(inode, NULL, inline_data, inline_len); | |
2981 | ||
2982 | if (queue_trunc) | |
2983 | ceph_queue_vmtruncate(inode); | |
2984 | ||
2985 | if (writeback) | |
2986 | /* | |
2987 | * queue inode for writeback: we can't actually call | |
2988 | * filemap_write_and_wait, etc. from message handler | |
2989 | * context. | |
2990 | */ | |
2991 | ceph_queue_writeback(inode); | |
2992 | if (queue_invalidate) | |
2993 | ceph_queue_invalidate(inode); | |
2994 | if (deleted_inode) | |
2995 | invalidate_aliases(inode); | |
2996 | if (wake) | |
2997 | wake_up_all(&ci->i_cap_wq); | |
2998 | ||
2999 | if (check_caps == 1) | |
3000 | ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY, | |
3001 | session); | |
3002 | else if (check_caps == 2) | |
3003 | ceph_check_caps(ci, CHECK_CAPS_NODELAY, session); | |
3004 | else | |
3005 | mutex_unlock(&session->s_mutex); | |
3006 | } | |
3007 | ||
3008 | /* | |
3009 | * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the | |
3010 | * MDS has been safely committed. | |
3011 | */ | |
3012 | static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid, | |
3013 | struct ceph_mds_caps *m, | |
3014 | struct ceph_mds_session *session, | |
3015 | struct ceph_cap *cap) | |
3016 | __releases(ci->i_ceph_lock) | |
3017 | { | |
3018 | struct ceph_inode_info *ci = ceph_inode(inode); | |
3019 | struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; | |
3020 | struct ceph_cap_flush *cf, *tmp_cf; | |
3021 | LIST_HEAD(to_remove); | |
3022 | unsigned seq = le32_to_cpu(m->seq); | |
3023 | int dirty = le32_to_cpu(m->dirty); | |
3024 | int cleaned = 0; | |
3025 | int drop = 0; | |
3026 | ||
3027 | list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) { | |
3028 | if (cf->tid == flush_tid) | |
3029 | cleaned = cf->caps; | |
3030 | if (cf->caps == 0) /* capsnap */ | |
3031 | continue; | |
3032 | if (cf->tid <= flush_tid) { | |
3033 | list_del(&cf->i_list); | |
3034 | list_add_tail(&cf->i_list, &to_remove); | |
3035 | } else { | |
3036 | cleaned &= ~cf->caps; | |
3037 | if (!cleaned) | |
3038 | break; | |
3039 | } | |
3040 | } | |
3041 | ||
3042 | dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s," | |
3043 | " flushing %s -> %s\n", | |
3044 | inode, session->s_mds, seq, ceph_cap_string(dirty), | |
3045 | ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps), | |
3046 | ceph_cap_string(ci->i_flushing_caps & ~cleaned)); | |
3047 | ||
3048 | if (list_empty(&to_remove) && !cleaned) | |
3049 | goto out; | |
3050 | ||
3051 | ci->i_flushing_caps &= ~cleaned; | |
3052 | ||
3053 | spin_lock(&mdsc->cap_dirty_lock); | |
3054 | ||
3055 | if (!list_empty(&to_remove)) { | |
3056 | u64 oldest_flush_tid; | |
3057 | list_for_each_entry(cf, &to_remove, i_list) | |
3058 | list_del(&cf->g_list); | |
3059 | ||
3060 | oldest_flush_tid = __get_oldest_flush_tid(mdsc); | |
3061 | if (oldest_flush_tid == 0 || oldest_flush_tid > flush_tid) | |
3062 | wake_up_all(&mdsc->cap_flushing_wq); | |
3063 | } | |
3064 | ||
3065 | if (ci->i_flushing_caps == 0) { | |
3066 | if (list_empty(&ci->i_cap_flush_list)) { | |
3067 | list_del_init(&ci->i_flushing_item); | |
3068 | if (!list_empty(&session->s_cap_flushing)) { | |
3069 | dout(" mds%d still flushing cap on %p\n", | |
3070 | session->s_mds, | |
3071 | &list_first_entry(&session->s_cap_flushing, | |
3072 | struct ceph_inode_info, | |
3073 | i_flushing_item)->vfs_inode); | |
3074 | } | |
3075 | } | |
3076 | mdsc->num_cap_flushing--; | |
3077 | dout(" inode %p now !flushing\n", inode); | |
3078 | ||
3079 | if (ci->i_dirty_caps == 0) { | |
3080 | dout(" inode %p now clean\n", inode); | |
3081 | BUG_ON(!list_empty(&ci->i_dirty_item)); | |
3082 | drop = 1; | |
3083 | if (ci->i_wr_ref == 0 && | |
3084 | ci->i_wrbuffer_ref_head == 0) { | |
3085 | BUG_ON(!ci->i_head_snapc); | |
3086 | ceph_put_snap_context(ci->i_head_snapc); | |
3087 | ci->i_head_snapc = NULL; | |
3088 | } | |
3089 | } else { | |
3090 | BUG_ON(list_empty(&ci->i_dirty_item)); | |
3091 | } | |
3092 | } | |
3093 | spin_unlock(&mdsc->cap_dirty_lock); | |
3094 | wake_up_all(&ci->i_cap_wq); | |
3095 | ||
3096 | out: | |
3097 | spin_unlock(&ci->i_ceph_lock); | |
3098 | ||
3099 | while (!list_empty(&to_remove)) { | |
3100 | cf = list_first_entry(&to_remove, | |
3101 | struct ceph_cap_flush, i_list); | |
3102 | list_del(&cf->i_list); | |
3103 | ceph_free_cap_flush(cf); | |
3104 | } | |
3105 | if (drop) | |
3106 | iput(inode); | |
3107 | } | |
3108 | ||
3109 | /* | |
3110 | * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can | |
3111 | * throw away our cap_snap. | |
3112 | * | |
3113 | * Caller hold s_mutex. | |
3114 | */ | |
3115 | static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid, | |
3116 | struct ceph_mds_caps *m, | |
3117 | struct ceph_mds_session *session) | |
3118 | { | |
3119 | struct ceph_inode_info *ci = ceph_inode(inode); | |
3120 | struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; | |
3121 | u64 follows = le64_to_cpu(m->snap_follows); | |
3122 | struct ceph_cap_snap *capsnap; | |
3123 | int flushed = 0; | |
3124 | ||
3125 | dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n", | |
3126 | inode, ci, session->s_mds, follows); | |
3127 | ||
3128 | spin_lock(&ci->i_ceph_lock); | |
3129 | list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) { | |
3130 | if (capsnap->follows == follows) { | |
3131 | if (capsnap->cap_flush.tid != flush_tid) { | |
3132 | dout(" cap_snap %p follows %lld tid %lld !=" | |
3133 | " %lld\n", capsnap, follows, | |
3134 | flush_tid, capsnap->cap_flush.tid); | |
3135 | break; | |
3136 | } | |
3137 | flushed = 1; | |
3138 | break; | |
3139 | } else { | |
3140 | dout(" skipping cap_snap %p follows %lld\n", | |
3141 | capsnap, capsnap->follows); | |
3142 | } | |
3143 | } | |
3144 | if (flushed) { | |
3145 | u64 oldest_flush_tid; | |
3146 | WARN_ON(capsnap->dirty_pages || capsnap->writing); | |
3147 | dout(" removing %p cap_snap %p follows %lld\n", | |
3148 | inode, capsnap, follows); | |
3149 | list_del(&capsnap->ci_item); | |
3150 | list_del(&capsnap->cap_flush.i_list); | |
3151 | ||
3152 | spin_lock(&mdsc->cap_dirty_lock); | |
3153 | ||
3154 | if (list_empty(&ci->i_cap_flush_list)) | |
3155 | list_del_init(&ci->i_flushing_item); | |
3156 | ||
3157 | list_del(&capsnap->cap_flush.g_list); | |
3158 | ||
3159 | oldest_flush_tid = __get_oldest_flush_tid(mdsc); | |
3160 | if (oldest_flush_tid == 0 || oldest_flush_tid > flush_tid) | |
3161 | wake_up_all(&mdsc->cap_flushing_wq); | |
3162 | ||
3163 | spin_unlock(&mdsc->cap_dirty_lock); | |
3164 | wake_up_all(&ci->i_cap_wq); | |
3165 | } | |
3166 | spin_unlock(&ci->i_ceph_lock); | |
3167 | if (flushed) { | |
3168 | ceph_put_snap_context(capsnap->context); | |
3169 | ceph_put_cap_snap(capsnap); | |
3170 | iput(inode); | |
3171 | } | |
3172 | } | |
3173 | ||
3174 | /* | |
3175 | * Handle TRUNC from MDS, indicating file truncation. | |
3176 | * | |
3177 | * caller hold s_mutex. | |
3178 | */ | |
3179 | static void handle_cap_trunc(struct inode *inode, | |
3180 | struct ceph_mds_caps *trunc, | |
3181 | struct ceph_mds_session *session) | |
3182 | __releases(ci->i_ceph_lock) | |
3183 | { | |
3184 | struct ceph_inode_info *ci = ceph_inode(inode); | |
3185 | int mds = session->s_mds; | |
3186 | int seq = le32_to_cpu(trunc->seq); | |
3187 | u32 truncate_seq = le32_to_cpu(trunc->truncate_seq); | |
3188 | u64 truncate_size = le64_to_cpu(trunc->truncate_size); | |
3189 | u64 size = le64_to_cpu(trunc->size); | |
3190 | int implemented = 0; | |
3191 | int dirty = __ceph_caps_dirty(ci); | |
3192 | int issued = __ceph_caps_issued(ceph_inode(inode), &implemented); | |
3193 | int queue_trunc = 0; | |
3194 | ||
3195 | issued |= implemented | dirty; | |
3196 | ||
3197 | dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n", | |
3198 | inode, mds, seq, truncate_size, truncate_seq); | |
3199 | queue_trunc = ceph_fill_file_size(inode, issued, | |
3200 | truncate_seq, truncate_size, size); | |
3201 | spin_unlock(&ci->i_ceph_lock); | |
3202 | ||
3203 | if (queue_trunc) | |
3204 | ceph_queue_vmtruncate(inode); | |
3205 | } | |
3206 | ||
3207 | /* | |
3208 | * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a | |
3209 | * different one. If we are the most recent migration we've seen (as | |
3210 | * indicated by mseq), make note of the migrating cap bits for the | |
3211 | * duration (until we see the corresponding IMPORT). | |
3212 | * | |
3213 | * caller holds s_mutex | |
3214 | */ | |
3215 | static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex, | |
3216 | struct ceph_mds_cap_peer *ph, | |
3217 | struct ceph_mds_session *session) | |
3218 | { | |
3219 | struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; | |
3220 | struct ceph_mds_session *tsession = NULL; | |
3221 | struct ceph_cap *cap, *tcap, *new_cap = NULL; | |
3222 | struct ceph_inode_info *ci = ceph_inode(inode); | |
3223 | u64 t_cap_id; | |
3224 | unsigned mseq = le32_to_cpu(ex->migrate_seq); | |
3225 | unsigned t_seq, t_mseq; | |
3226 | int target, issued; | |
3227 | int mds = session->s_mds; | |
3228 | ||
3229 | if (ph) { | |
3230 | t_cap_id = le64_to_cpu(ph->cap_id); | |
3231 | t_seq = le32_to_cpu(ph->seq); | |
3232 | t_mseq = le32_to_cpu(ph->mseq); | |
3233 | target = le32_to_cpu(ph->mds); | |
3234 | } else { | |
3235 | t_cap_id = t_seq = t_mseq = 0; | |
3236 | target = -1; | |
3237 | } | |
3238 | ||
3239 | dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n", | |
3240 | inode, ci, mds, mseq, target); | |
3241 | retry: | |
3242 | spin_lock(&ci->i_ceph_lock); | |
3243 | cap = __get_cap_for_mds(ci, mds); | |
3244 | if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id)) | |
3245 | goto out_unlock; | |
3246 | ||
3247 | if (target < 0) { | |
3248 | __ceph_remove_cap(cap, false); | |
3249 | if (!ci->i_auth_cap) | |
3250 | ci->i_ceph_flags |= CEPH_I_CAP_DROPPED; | |
3251 | goto out_unlock; | |
3252 | } | |
3253 | ||
3254 | /* | |
3255 | * now we know we haven't received the cap import message yet | |
3256 | * because the exported cap still exist. | |
3257 | */ | |
3258 | ||
3259 | issued = cap->issued; | |
3260 | WARN_ON(issued != cap->implemented); | |
3261 | ||
3262 | tcap = __get_cap_for_mds(ci, target); | |
3263 | if (tcap) { | |
3264 | /* already have caps from the target */ | |
3265 | if (tcap->cap_id != t_cap_id || | |
3266 | ceph_seq_cmp(tcap->seq, t_seq) < 0) { | |
3267 | dout(" updating import cap %p mds%d\n", tcap, target); | |
3268 | tcap->cap_id = t_cap_id; | |
3269 | tcap->seq = t_seq - 1; | |
3270 | tcap->issue_seq = t_seq - 1; | |
3271 | tcap->mseq = t_mseq; | |
3272 | tcap->issued |= issued; | |
3273 | tcap->implemented |= issued; | |
3274 | if (cap == ci->i_auth_cap) | |
3275 | ci->i_auth_cap = tcap; | |
3276 | if (!list_empty(&ci->i_cap_flush_list) && | |
3277 | ci->i_auth_cap == tcap) { | |
3278 | spin_lock(&mdsc->cap_dirty_lock); | |
3279 | list_move_tail(&ci->i_flushing_item, | |
3280 | &tcap->session->s_cap_flushing); | |
3281 | spin_unlock(&mdsc->cap_dirty_lock); | |
3282 | } | |
3283 | } | |
3284 | __ceph_remove_cap(cap, false); | |
3285 | goto out_unlock; | |
3286 | } else if (tsession) { | |
3287 | /* add placeholder for the export tagert */ | |
3288 | int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0; | |
3289 | ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0, | |
3290 | t_seq - 1, t_mseq, (u64)-1, flag, &new_cap); | |
3291 | ||
3292 | __ceph_remove_cap(cap, false); | |
3293 | goto out_unlock; | |
3294 | } | |
3295 | ||
3296 | spin_unlock(&ci->i_ceph_lock); | |
3297 | mutex_unlock(&session->s_mutex); | |
3298 | ||
3299 | /* open target session */ | |
3300 | tsession = ceph_mdsc_open_export_target_session(mdsc, target); | |
3301 | if (!IS_ERR(tsession)) { | |
3302 | if (mds > target) { | |
3303 | mutex_lock(&session->s_mutex); | |
3304 | mutex_lock_nested(&tsession->s_mutex, | |
3305 | SINGLE_DEPTH_NESTING); | |
3306 | } else { | |
3307 | mutex_lock(&tsession->s_mutex); | |
3308 | mutex_lock_nested(&session->s_mutex, | |
3309 | SINGLE_DEPTH_NESTING); | |
3310 | } | |
3311 | new_cap = ceph_get_cap(mdsc, NULL); | |
3312 | } else { | |
3313 | WARN_ON(1); | |
3314 | tsession = NULL; | |
3315 | target = -1; | |
3316 | } | |
3317 | goto retry; | |
3318 | ||
3319 | out_unlock: | |
3320 | spin_unlock(&ci->i_ceph_lock); | |
3321 | mutex_unlock(&session->s_mutex); | |
3322 | if (tsession) { | |
3323 | mutex_unlock(&tsession->s_mutex); | |
3324 | ceph_put_mds_session(tsession); | |
3325 | } | |
3326 | if (new_cap) | |
3327 | ceph_put_cap(mdsc, new_cap); | |
3328 | } | |
3329 | ||
3330 | /* | |
3331 | * Handle cap IMPORT. | |
3332 | * | |
3333 | * caller holds s_mutex. acquires i_ceph_lock | |
3334 | */ | |
3335 | static void handle_cap_import(struct ceph_mds_client *mdsc, | |
3336 | struct inode *inode, struct ceph_mds_caps *im, | |
3337 | struct ceph_mds_cap_peer *ph, | |
3338 | struct ceph_mds_session *session, | |
3339 | struct ceph_cap **target_cap, int *old_issued) | |
3340 | __acquires(ci->i_ceph_lock) | |
3341 | { | |
3342 | struct ceph_inode_info *ci = ceph_inode(inode); | |
3343 | struct ceph_cap *cap, *ocap, *new_cap = NULL; | |
3344 | int mds = session->s_mds; | |
3345 | int issued; | |
3346 | unsigned caps = le32_to_cpu(im->caps); | |
3347 | unsigned wanted = le32_to_cpu(im->wanted); | |
3348 | unsigned seq = le32_to_cpu(im->seq); | |
3349 | unsigned mseq = le32_to_cpu(im->migrate_seq); | |
3350 | u64 realmino = le64_to_cpu(im->realm); | |
3351 | u64 cap_id = le64_to_cpu(im->cap_id); | |
3352 | u64 p_cap_id; | |
3353 | int peer; | |
3354 | ||
3355 | if (ph) { | |
3356 | p_cap_id = le64_to_cpu(ph->cap_id); | |
3357 | peer = le32_to_cpu(ph->mds); | |
3358 | } else { | |
3359 | p_cap_id = 0; | |
3360 | peer = -1; | |
3361 | } | |
3362 | ||
3363 | dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n", | |
3364 | inode, ci, mds, mseq, peer); | |
3365 | ||
3366 | retry: | |
3367 | spin_lock(&ci->i_ceph_lock); | |
3368 | cap = __get_cap_for_mds(ci, mds); | |
3369 | if (!cap) { | |
3370 | if (!new_cap) { | |
3371 | spin_unlock(&ci->i_ceph_lock); | |
3372 | new_cap = ceph_get_cap(mdsc, NULL); | |
3373 | goto retry; | |
3374 | } | |
3375 | cap = new_cap; | |
3376 | } else { | |
3377 | if (new_cap) { | |
3378 | ceph_put_cap(mdsc, new_cap); | |
3379 | new_cap = NULL; | |
3380 | } | |
3381 | } | |
3382 | ||
3383 | __ceph_caps_issued(ci, &issued); | |
3384 | issued |= __ceph_caps_dirty(ci); | |
3385 | ||
3386 | ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq, | |
3387 | realmino, CEPH_CAP_FLAG_AUTH, &new_cap); | |
3388 | ||
3389 | ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL; | |
3390 | if (ocap && ocap->cap_id == p_cap_id) { | |
3391 | dout(" remove export cap %p mds%d flags %d\n", | |
3392 | ocap, peer, ph->flags); | |
3393 | if ((ph->flags & CEPH_CAP_FLAG_AUTH) && | |
3394 | (ocap->seq != le32_to_cpu(ph->seq) || | |
3395 | ocap->mseq != le32_to_cpu(ph->mseq))) { | |
3396 | pr_err("handle_cap_import: mismatched seq/mseq: " | |
3397 | "ino (%llx.%llx) mds%d seq %d mseq %d " | |
3398 | "importer mds%d has peer seq %d mseq %d\n", | |
3399 | ceph_vinop(inode), peer, ocap->seq, | |
3400 | ocap->mseq, mds, le32_to_cpu(ph->seq), | |
3401 | le32_to_cpu(ph->mseq)); | |
3402 | } | |
3403 | __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE)); | |
3404 | } | |
3405 | ||
3406 | /* make sure we re-request max_size, if necessary */ | |
3407 | ci->i_wanted_max_size = 0; | |
3408 | ci->i_requested_max_size = 0; | |
3409 | ||
3410 | *old_issued = issued; | |
3411 | *target_cap = cap; | |
3412 | } | |
3413 | ||
3414 | /* | |
3415 | * Handle a caps message from the MDS. | |
3416 | * | |
3417 | * Identify the appropriate session, inode, and call the right handler | |
3418 | * based on the cap op. | |
3419 | */ | |
3420 | void ceph_handle_caps(struct ceph_mds_session *session, | |
3421 | struct ceph_msg *msg) | |
3422 | { | |
3423 | struct ceph_mds_client *mdsc = session->s_mdsc; | |
3424 | struct super_block *sb = mdsc->fsc->sb; | |
3425 | struct inode *inode; | |
3426 | struct ceph_inode_info *ci; | |
3427 | struct ceph_cap *cap; | |
3428 | struct ceph_mds_caps *h; | |
3429 | struct ceph_mds_cap_peer *peer = NULL; | |
3430 | struct ceph_snap_realm *realm = NULL; | |
3431 | struct ceph_string *pool_ns = NULL; | |
3432 | int mds = session->s_mds; | |
3433 | int op, issued; | |
3434 | u32 seq, mseq; | |
3435 | struct ceph_vino vino; | |
3436 | u64 tid; | |
3437 | u64 inline_version = 0; | |
3438 | void *inline_data = NULL; | |
3439 | u32 inline_len = 0; | |
3440 | void *snaptrace; | |
3441 | size_t snaptrace_len; | |
3442 | void *p, *end; | |
3443 | ||
3444 | dout("handle_caps from mds%d\n", mds); | |
3445 | ||
3446 | /* decode */ | |
3447 | end = msg->front.iov_base + msg->front.iov_len; | |
3448 | tid = le64_to_cpu(msg->hdr.tid); | |
3449 | if (msg->front.iov_len < sizeof(*h)) | |
3450 | goto bad; | |
3451 | h = msg->front.iov_base; | |
3452 | op = le32_to_cpu(h->op); | |
3453 | vino.ino = le64_to_cpu(h->ino); | |
3454 | vino.snap = CEPH_NOSNAP; | |
3455 | seq = le32_to_cpu(h->seq); | |
3456 | mseq = le32_to_cpu(h->migrate_seq); | |
3457 | ||
3458 | snaptrace = h + 1; | |
3459 | snaptrace_len = le32_to_cpu(h->snap_trace_len); | |
3460 | p = snaptrace + snaptrace_len; | |
3461 | ||
3462 | if (le16_to_cpu(msg->hdr.version) >= 2) { | |
3463 | u32 flock_len; | |
3464 | ceph_decode_32_safe(&p, end, flock_len, bad); | |
3465 | if (p + flock_len > end) | |
3466 | goto bad; | |
3467 | p += flock_len; | |
3468 | } | |
3469 | ||
3470 | if (le16_to_cpu(msg->hdr.version) >= 3) { | |
3471 | if (op == CEPH_CAP_OP_IMPORT) { | |
3472 | if (p + sizeof(*peer) > end) | |
3473 | goto bad; | |
3474 | peer = p; | |
3475 | p += sizeof(*peer); | |
3476 | } else if (op == CEPH_CAP_OP_EXPORT) { | |
3477 | /* recorded in unused fields */ | |
3478 | peer = (void *)&h->size; | |
3479 | } | |
3480 | } | |
3481 | ||
3482 | if (le16_to_cpu(msg->hdr.version) >= 4) { | |
3483 | ceph_decode_64_safe(&p, end, inline_version, bad); | |
3484 | ceph_decode_32_safe(&p, end, inline_len, bad); | |
3485 | if (p + inline_len > end) | |
3486 | goto bad; | |
3487 | inline_data = p; | |
3488 | p += inline_len; | |
3489 | } | |
3490 | ||
3491 | if (le16_to_cpu(msg->hdr.version) >= 8) { | |
3492 | u64 flush_tid; | |
3493 | u32 caller_uid, caller_gid; | |
3494 | u32 osd_epoch_barrier; | |
3495 | u32 pool_ns_len; | |
3496 | /* version >= 5 */ | |
3497 | ceph_decode_32_safe(&p, end, osd_epoch_barrier, bad); | |
3498 | /* version >= 6 */ | |
3499 | ceph_decode_64_safe(&p, end, flush_tid, bad); | |
3500 | /* version >= 7 */ | |
3501 | ceph_decode_32_safe(&p, end, caller_uid, bad); | |
3502 | ceph_decode_32_safe(&p, end, caller_gid, bad); | |
3503 | /* version >= 8 */ | |
3504 | ceph_decode_32_safe(&p, end, pool_ns_len, bad); | |
3505 | if (pool_ns_len > 0) { | |
3506 | ceph_decode_need(&p, end, pool_ns_len, bad); | |
3507 | pool_ns = ceph_find_or_create_string(p, pool_ns_len); | |
3508 | p += pool_ns_len; | |
3509 | } | |
3510 | } | |
3511 | ||
3512 | /* lookup ino */ | |
3513 | inode = ceph_find_inode(sb, vino); | |
3514 | ci = ceph_inode(inode); | |
3515 | dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino, | |
3516 | vino.snap, inode); | |
3517 | ||
3518 | mutex_lock(&session->s_mutex); | |
3519 | session->s_seq++; | |
3520 | dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq, | |
3521 | (unsigned)seq); | |
3522 | ||
3523 | if (!inode) { | |
3524 | dout(" i don't have ino %llx\n", vino.ino); | |
3525 | ||
3526 | if (op == CEPH_CAP_OP_IMPORT) { | |
3527 | cap = ceph_get_cap(mdsc, NULL); | |
3528 | cap->cap_ino = vino.ino; | |
3529 | cap->queue_release = 1; | |
3530 | cap->cap_id = le64_to_cpu(h->cap_id); | |
3531 | cap->mseq = mseq; | |
3532 | cap->seq = seq; | |
3533 | spin_lock(&session->s_cap_lock); | |
3534 | list_add_tail(&cap->session_caps, | |
3535 | &session->s_cap_releases); | |
3536 | session->s_num_cap_releases++; | |
3537 | spin_unlock(&session->s_cap_lock); | |
3538 | } | |
3539 | goto flush_cap_releases; | |
3540 | } | |
3541 | ||
3542 | /* these will work even if we don't have a cap yet */ | |
3543 | switch (op) { | |
3544 | case CEPH_CAP_OP_FLUSHSNAP_ACK: | |
3545 | handle_cap_flushsnap_ack(inode, tid, h, session); | |
3546 | goto done; | |
3547 | ||
3548 | case CEPH_CAP_OP_EXPORT: | |
3549 | handle_cap_export(inode, h, peer, session); | |
3550 | goto done_unlocked; | |
3551 | ||
3552 | case CEPH_CAP_OP_IMPORT: | |
3553 | realm = NULL; | |
3554 | if (snaptrace_len) { | |
3555 | down_write(&mdsc->snap_rwsem); | |
3556 | ceph_update_snap_trace(mdsc, snaptrace, | |
3557 | snaptrace + snaptrace_len, | |
3558 | false, &realm); | |
3559 | downgrade_write(&mdsc->snap_rwsem); | |
3560 | } else { | |
3561 | down_read(&mdsc->snap_rwsem); | |
3562 | } | |
3563 | handle_cap_import(mdsc, inode, h, peer, session, | |
3564 | &cap, &issued); | |
3565 | handle_cap_grant(mdsc, inode, h, &pool_ns, | |
3566 | inline_version, inline_data, inline_len, | |
3567 | msg->middle, session, cap, issued); | |
3568 | if (realm) | |
3569 | ceph_put_snap_realm(mdsc, realm); | |
3570 | goto done_unlocked; | |
3571 | } | |
3572 | ||
3573 | /* the rest require a cap */ | |
3574 | spin_lock(&ci->i_ceph_lock); | |
3575 | cap = __get_cap_for_mds(ceph_inode(inode), mds); | |
3576 | if (!cap) { | |
3577 | dout(" no cap on %p ino %llx.%llx from mds%d\n", | |
3578 | inode, ceph_ino(inode), ceph_snap(inode), mds); | |
3579 | spin_unlock(&ci->i_ceph_lock); | |
3580 | goto flush_cap_releases; | |
3581 | } | |
3582 | ||
3583 | /* note that each of these drops i_ceph_lock for us */ | |
3584 | switch (op) { | |
3585 | case CEPH_CAP_OP_REVOKE: | |
3586 | case CEPH_CAP_OP_GRANT: | |
3587 | __ceph_caps_issued(ci, &issued); | |
3588 | issued |= __ceph_caps_dirty(ci); | |
3589 | handle_cap_grant(mdsc, inode, h, &pool_ns, | |
3590 | inline_version, inline_data, inline_len, | |
3591 | msg->middle, session, cap, issued); | |
3592 | goto done_unlocked; | |
3593 | ||
3594 | case CEPH_CAP_OP_FLUSH_ACK: | |
3595 | handle_cap_flush_ack(inode, tid, h, session, cap); | |
3596 | break; | |
3597 | ||
3598 | case CEPH_CAP_OP_TRUNC: | |
3599 | handle_cap_trunc(inode, h, session); | |
3600 | break; | |
3601 | ||
3602 | default: | |
3603 | spin_unlock(&ci->i_ceph_lock); | |
3604 | pr_err("ceph_handle_caps: unknown cap op %d %s\n", op, | |
3605 | ceph_cap_op_name(op)); | |
3606 | } | |
3607 | ||
3608 | goto done; | |
3609 | ||
3610 | flush_cap_releases: | |
3611 | /* | |
3612 | * send any cap release message to try to move things | |
3613 | * along for the mds (who clearly thinks we still have this | |
3614 | * cap). | |
3615 | */ | |
3616 | ceph_send_cap_releases(mdsc, session); | |
3617 | ||
3618 | done: | |
3619 | mutex_unlock(&session->s_mutex); | |
3620 | done_unlocked: | |
3621 | iput(inode); | |
3622 | ceph_put_string(pool_ns); | |
3623 | return; | |
3624 | ||
3625 | bad: | |
3626 | pr_err("ceph_handle_caps: corrupt message\n"); | |
3627 | ceph_msg_dump(msg); | |
3628 | return; | |
3629 | } | |
3630 | ||
3631 | /* | |
3632 | * Delayed work handler to process end of delayed cap release LRU list. | |
3633 | */ | |
3634 | void ceph_check_delayed_caps(struct ceph_mds_client *mdsc) | |
3635 | { | |
3636 | struct ceph_inode_info *ci; | |
3637 | int flags = CHECK_CAPS_NODELAY; | |
3638 | ||
3639 | dout("check_delayed_caps\n"); | |
3640 | while (1) { | |
3641 | spin_lock(&mdsc->cap_delay_lock); | |
3642 | if (list_empty(&mdsc->cap_delay_list)) | |
3643 | break; | |
3644 | ci = list_first_entry(&mdsc->cap_delay_list, | |
3645 | struct ceph_inode_info, | |
3646 | i_cap_delay_list); | |
3647 | if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 && | |
3648 | time_before(jiffies, ci->i_hold_caps_max)) | |
3649 | break; | |
3650 | list_del_init(&ci->i_cap_delay_list); | |
3651 | spin_unlock(&mdsc->cap_delay_lock); | |
3652 | dout("check_delayed_caps on %p\n", &ci->vfs_inode); | |
3653 | ceph_check_caps(ci, flags, NULL); | |
3654 | } | |
3655 | spin_unlock(&mdsc->cap_delay_lock); | |
3656 | } | |
3657 | ||
3658 | /* | |
3659 | * Flush all dirty caps to the mds | |
3660 | */ | |
3661 | void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc) | |
3662 | { | |
3663 | struct ceph_inode_info *ci; | |
3664 | struct inode *inode; | |
3665 | ||
3666 | dout("flush_dirty_caps\n"); | |
3667 | spin_lock(&mdsc->cap_dirty_lock); | |
3668 | while (!list_empty(&mdsc->cap_dirty)) { | |
3669 | ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info, | |
3670 | i_dirty_item); | |
3671 | inode = &ci->vfs_inode; | |
3672 | ihold(inode); | |
3673 | dout("flush_dirty_caps %p\n", inode); | |
3674 | spin_unlock(&mdsc->cap_dirty_lock); | |
3675 | ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL); | |
3676 | iput(inode); | |
3677 | spin_lock(&mdsc->cap_dirty_lock); | |
3678 | } | |
3679 | spin_unlock(&mdsc->cap_dirty_lock); | |
3680 | dout("flush_dirty_caps done\n"); | |
3681 | } | |
3682 | ||
3683 | void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode) | |
3684 | { | |
3685 | int i; | |
3686 | int bits = (fmode << 1) | 1; | |
3687 | for (i = 0; i < CEPH_FILE_MODE_BITS; i++) { | |
3688 | if (bits & (1 << i)) | |
3689 | ci->i_nr_by_mode[i]++; | |
3690 | } | |
3691 | } | |
3692 | ||
3693 | /* | |
3694 | * Drop open file reference. If we were the last open file, | |
3695 | * we may need to release capabilities to the MDS (or schedule | |
3696 | * their delayed release). | |
3697 | */ | |
3698 | void ceph_put_fmode(struct ceph_inode_info *ci, int fmode) | |
3699 | { | |
3700 | int i, last = 0; | |
3701 | int bits = (fmode << 1) | 1; | |
3702 | spin_lock(&ci->i_ceph_lock); | |
3703 | for (i = 0; i < CEPH_FILE_MODE_BITS; i++) { | |
3704 | if (bits & (1 << i)) { | |
3705 | BUG_ON(ci->i_nr_by_mode[i] == 0); | |
3706 | if (--ci->i_nr_by_mode[i] == 0) | |
3707 | last++; | |
3708 | } | |
3709 | } | |
3710 | dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n", | |
3711 | &ci->vfs_inode, fmode, | |
3712 | ci->i_nr_by_mode[0], ci->i_nr_by_mode[1], | |
3713 | ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]); | |
3714 | spin_unlock(&ci->i_ceph_lock); | |
3715 | ||
3716 | if (last && ci->i_vino.snap == CEPH_NOSNAP) | |
3717 | ceph_check_caps(ci, 0, NULL); | |
3718 | } | |
3719 | ||
3720 | /* | |
3721 | * Helpers for embedding cap and dentry lease releases into mds | |
3722 | * requests. | |
3723 | * | |
3724 | * @force is used by dentry_release (below) to force inclusion of a | |
3725 | * record for the directory inode, even when there aren't any caps to | |
3726 | * drop. | |
3727 | */ | |
3728 | int ceph_encode_inode_release(void **p, struct inode *inode, | |
3729 | int mds, int drop, int unless, int force) | |
3730 | { | |
3731 | struct ceph_inode_info *ci = ceph_inode(inode); | |
3732 | struct ceph_cap *cap; | |
3733 | struct ceph_mds_request_release *rel = *p; | |
3734 | int used, dirty; | |
3735 | int ret = 0; | |
3736 | ||
3737 | spin_lock(&ci->i_ceph_lock); | |
3738 | used = __ceph_caps_used(ci); | |
3739 | dirty = __ceph_caps_dirty(ci); | |
3740 | ||
3741 | dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n", | |
3742 | inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop), | |
3743 | ceph_cap_string(unless)); | |
3744 | ||
3745 | /* only drop unused, clean caps */ | |
3746 | drop &= ~(used | dirty); | |
3747 | ||
3748 | cap = __get_cap_for_mds(ci, mds); | |
3749 | if (cap && __cap_is_valid(cap)) { | |
3750 | if (force || | |
3751 | ((cap->issued & drop) && | |
3752 | (cap->issued & unless) == 0)) { | |
3753 | if ((cap->issued & drop) && | |
3754 | (cap->issued & unless) == 0) { | |
3755 | int wanted = __ceph_caps_wanted(ci); | |
3756 | if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0) | |
3757 | wanted |= cap->mds_wanted; | |
3758 | dout("encode_inode_release %p cap %p " | |
3759 | "%s -> %s, wanted %s -> %s\n", inode, cap, | |
3760 | ceph_cap_string(cap->issued), | |
3761 | ceph_cap_string(cap->issued & ~drop), | |
3762 | ceph_cap_string(cap->mds_wanted), | |
3763 | ceph_cap_string(wanted)); | |
3764 | ||
3765 | cap->issued &= ~drop; | |
3766 | cap->implemented &= ~drop; | |
3767 | cap->mds_wanted = wanted; | |
3768 | } else { | |
3769 | dout("encode_inode_release %p cap %p %s" | |
3770 | " (force)\n", inode, cap, | |
3771 | ceph_cap_string(cap->issued)); | |
3772 | } | |
3773 | ||
3774 | rel->ino = cpu_to_le64(ceph_ino(inode)); | |
3775 | rel->cap_id = cpu_to_le64(cap->cap_id); | |
3776 | rel->seq = cpu_to_le32(cap->seq); | |
3777 | rel->issue_seq = cpu_to_le32(cap->issue_seq); | |
3778 | rel->mseq = cpu_to_le32(cap->mseq); | |
3779 | rel->caps = cpu_to_le32(cap->implemented); | |
3780 | rel->wanted = cpu_to_le32(cap->mds_wanted); | |
3781 | rel->dname_len = 0; | |
3782 | rel->dname_seq = 0; | |
3783 | *p += sizeof(*rel); | |
3784 | ret = 1; | |
3785 | } else { | |
3786 | dout("encode_inode_release %p cap %p %s\n", | |
3787 | inode, cap, ceph_cap_string(cap->issued)); | |
3788 | } | |
3789 | } | |
3790 | spin_unlock(&ci->i_ceph_lock); | |
3791 | return ret; | |
3792 | } | |
3793 | ||
3794 | int ceph_encode_dentry_release(void **p, struct dentry *dentry, | |
3795 | int mds, int drop, int unless) | |
3796 | { | |
3797 | struct inode *dir = d_inode(dentry->d_parent); | |
3798 | struct ceph_mds_request_release *rel = *p; | |
3799 | struct ceph_dentry_info *di = ceph_dentry(dentry); | |
3800 | int force = 0; | |
3801 | int ret; | |
3802 | ||
3803 | /* | |
3804 | * force an record for the directory caps if we have a dentry lease. | |
3805 | * this is racy (can't take i_ceph_lock and d_lock together), but it | |
3806 | * doesn't have to be perfect; the mds will revoke anything we don't | |
3807 | * release. | |
3808 | */ | |
3809 | spin_lock(&dentry->d_lock); | |
3810 | if (di->lease_session && di->lease_session->s_mds == mds) | |
3811 | force = 1; | |
3812 | spin_unlock(&dentry->d_lock); | |
3813 | ||
3814 | ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force); | |
3815 | ||
3816 | spin_lock(&dentry->d_lock); | |
3817 | if (ret && di->lease_session && di->lease_session->s_mds == mds) { | |
3818 | dout("encode_dentry_release %p mds%d seq %d\n", | |
3819 | dentry, mds, (int)di->lease_seq); | |
3820 | rel->dname_len = cpu_to_le32(dentry->d_name.len); | |
3821 | memcpy(*p, dentry->d_name.name, dentry->d_name.len); | |
3822 | *p += dentry->d_name.len; | |
3823 | rel->dname_seq = cpu_to_le32(di->lease_seq); | |
3824 | __ceph_mdsc_drop_dentry_lease(dentry); | |
3825 | } | |
3826 | spin_unlock(&dentry->d_lock); | |
3827 | return ret; | |
3828 | } |