xref: /openbmc/linux/fs/ceph/caps.c (revision b34e08d5)
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 
44 /*
45  * Generate readable cap strings for debugging output.
46  */
47 #define MAX_CAP_STR 20
48 static char cap_str[MAX_CAP_STR][40];
49 static DEFINE_SPINLOCK(cap_str_lock);
50 static int last_cap_str;
51 
52 static char *gcap_string(char *s, int c)
53 {
54 	if (c & CEPH_CAP_GSHARED)
55 		*s++ = 's';
56 	if (c & CEPH_CAP_GEXCL)
57 		*s++ = 'x';
58 	if (c & CEPH_CAP_GCACHE)
59 		*s++ = 'c';
60 	if (c & CEPH_CAP_GRD)
61 		*s++ = 'r';
62 	if (c & CEPH_CAP_GWR)
63 		*s++ = 'w';
64 	if (c & CEPH_CAP_GBUFFER)
65 		*s++ = 'b';
66 	if (c & CEPH_CAP_GLAZYIO)
67 		*s++ = 'l';
68 	return s;
69 }
70 
71 const char *ceph_cap_string(int caps)
72 {
73 	int i;
74 	char *s;
75 	int c;
76 
77 	spin_lock(&cap_str_lock);
78 	i = last_cap_str++;
79 	if (last_cap_str == MAX_CAP_STR)
80 		last_cap_str = 0;
81 	spin_unlock(&cap_str_lock);
82 
83 	s = cap_str[i];
84 
85 	if (caps & CEPH_CAP_PIN)
86 		*s++ = 'p';
87 
88 	c = (caps >> CEPH_CAP_SAUTH) & 3;
89 	if (c) {
90 		*s++ = 'A';
91 		s = gcap_string(s, c);
92 	}
93 
94 	c = (caps >> CEPH_CAP_SLINK) & 3;
95 	if (c) {
96 		*s++ = 'L';
97 		s = gcap_string(s, c);
98 	}
99 
100 	c = (caps >> CEPH_CAP_SXATTR) & 3;
101 	if (c) {
102 		*s++ = 'X';
103 		s = gcap_string(s, c);
104 	}
105 
106 	c = caps >> CEPH_CAP_SFILE;
107 	if (c) {
108 		*s++ = 'F';
109 		s = gcap_string(s, c);
110 	}
111 
112 	if (s == cap_str[i])
113 		*s++ = '-';
114 	*s = 0;
115 	return cap_str[i];
116 }
117 
118 void ceph_caps_init(struct ceph_mds_client *mdsc)
119 {
120 	INIT_LIST_HEAD(&mdsc->caps_list);
121 	spin_lock_init(&mdsc->caps_list_lock);
122 }
123 
124 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
125 {
126 	struct ceph_cap *cap;
127 
128 	spin_lock(&mdsc->caps_list_lock);
129 	while (!list_empty(&mdsc->caps_list)) {
130 		cap = list_first_entry(&mdsc->caps_list,
131 				       struct ceph_cap, caps_item);
132 		list_del(&cap->caps_item);
133 		kmem_cache_free(ceph_cap_cachep, cap);
134 	}
135 	mdsc->caps_total_count = 0;
136 	mdsc->caps_avail_count = 0;
137 	mdsc->caps_use_count = 0;
138 	mdsc->caps_reserve_count = 0;
139 	mdsc->caps_min_count = 0;
140 	spin_unlock(&mdsc->caps_list_lock);
141 }
142 
143 void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
144 {
145 	spin_lock(&mdsc->caps_list_lock);
146 	mdsc->caps_min_count += delta;
147 	BUG_ON(mdsc->caps_min_count < 0);
148 	spin_unlock(&mdsc->caps_list_lock);
149 }
150 
151 void ceph_reserve_caps(struct ceph_mds_client *mdsc,
152 		      struct ceph_cap_reservation *ctx, int need)
153 {
154 	int i;
155 	struct ceph_cap *cap;
156 	int have;
157 	int alloc = 0;
158 	LIST_HEAD(newcaps);
159 
160 	dout("reserve caps ctx=%p need=%d\n", ctx, need);
161 
162 	/* first reserve any caps that are already allocated */
163 	spin_lock(&mdsc->caps_list_lock);
164 	if (mdsc->caps_avail_count >= need)
165 		have = need;
166 	else
167 		have = mdsc->caps_avail_count;
168 	mdsc->caps_avail_count -= have;
169 	mdsc->caps_reserve_count += have;
170 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
171 					 mdsc->caps_reserve_count +
172 					 mdsc->caps_avail_count);
173 	spin_unlock(&mdsc->caps_list_lock);
174 
175 	for (i = have; i < need; i++) {
176 		cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
177 		if (!cap)
178 			break;
179 		list_add(&cap->caps_item, &newcaps);
180 		alloc++;
181 	}
182 	/* we didn't manage to reserve as much as we needed */
183 	if (have + alloc != need)
184 		pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
185 			ctx, need, have + alloc);
186 
187 	spin_lock(&mdsc->caps_list_lock);
188 	mdsc->caps_total_count += alloc;
189 	mdsc->caps_reserve_count += alloc;
190 	list_splice(&newcaps, &mdsc->caps_list);
191 
192 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
193 					 mdsc->caps_reserve_count +
194 					 mdsc->caps_avail_count);
195 	spin_unlock(&mdsc->caps_list_lock);
196 
197 	ctx->count = need;
198 	dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
199 	     ctx, mdsc->caps_total_count, mdsc->caps_use_count,
200 	     mdsc->caps_reserve_count, mdsc->caps_avail_count);
201 }
202 
203 int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
204 			struct ceph_cap_reservation *ctx)
205 {
206 	dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
207 	if (ctx->count) {
208 		spin_lock(&mdsc->caps_list_lock);
209 		BUG_ON(mdsc->caps_reserve_count < ctx->count);
210 		mdsc->caps_reserve_count -= ctx->count;
211 		mdsc->caps_avail_count += ctx->count;
212 		ctx->count = 0;
213 		dout("unreserve caps %d = %d used + %d resv + %d avail\n",
214 		     mdsc->caps_total_count, mdsc->caps_use_count,
215 		     mdsc->caps_reserve_count, mdsc->caps_avail_count);
216 		BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
217 						 mdsc->caps_reserve_count +
218 						 mdsc->caps_avail_count);
219 		spin_unlock(&mdsc->caps_list_lock);
220 	}
221 	return 0;
222 }
223 
224 static struct ceph_cap *get_cap(struct ceph_mds_client *mdsc,
225 				struct ceph_cap_reservation *ctx)
226 {
227 	struct ceph_cap *cap = NULL;
228 
229 	/* temporary, until we do something about cap import/export */
230 	if (!ctx) {
231 		cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
232 		if (cap) {
233 			spin_lock(&mdsc->caps_list_lock);
234 			mdsc->caps_use_count++;
235 			mdsc->caps_total_count++;
236 			spin_unlock(&mdsc->caps_list_lock);
237 		}
238 		return cap;
239 	}
240 
241 	spin_lock(&mdsc->caps_list_lock);
242 	dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
243 	     ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
244 	     mdsc->caps_reserve_count, mdsc->caps_avail_count);
245 	BUG_ON(!ctx->count);
246 	BUG_ON(ctx->count > mdsc->caps_reserve_count);
247 	BUG_ON(list_empty(&mdsc->caps_list));
248 
249 	ctx->count--;
250 	mdsc->caps_reserve_count--;
251 	mdsc->caps_use_count++;
252 
253 	cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
254 	list_del(&cap->caps_item);
255 
256 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
257 	       mdsc->caps_reserve_count + mdsc->caps_avail_count);
258 	spin_unlock(&mdsc->caps_list_lock);
259 	return cap;
260 }
261 
262 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
263 {
264 	spin_lock(&mdsc->caps_list_lock);
265 	dout("put_cap %p %d = %d used + %d resv + %d avail\n",
266 	     cap, mdsc->caps_total_count, mdsc->caps_use_count,
267 	     mdsc->caps_reserve_count, mdsc->caps_avail_count);
268 	mdsc->caps_use_count--;
269 	/*
270 	 * Keep some preallocated caps around (ceph_min_count), to
271 	 * avoid lots of free/alloc churn.
272 	 */
273 	if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
274 				      mdsc->caps_min_count) {
275 		mdsc->caps_total_count--;
276 		kmem_cache_free(ceph_cap_cachep, cap);
277 	} else {
278 		mdsc->caps_avail_count++;
279 		list_add(&cap->caps_item, &mdsc->caps_list);
280 	}
281 
282 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
283 	       mdsc->caps_reserve_count + mdsc->caps_avail_count);
284 	spin_unlock(&mdsc->caps_list_lock);
285 }
286 
287 void ceph_reservation_status(struct ceph_fs_client *fsc,
288 			     int *total, int *avail, int *used, int *reserved,
289 			     int *min)
290 {
291 	struct ceph_mds_client *mdsc = fsc->mdsc;
292 
293 	if (total)
294 		*total = mdsc->caps_total_count;
295 	if (avail)
296 		*avail = mdsc->caps_avail_count;
297 	if (used)
298 		*used = mdsc->caps_use_count;
299 	if (reserved)
300 		*reserved = mdsc->caps_reserve_count;
301 	if (min)
302 		*min = mdsc->caps_min_count;
303 }
304 
305 /*
306  * Find ceph_cap for given mds, if any.
307  *
308  * Called with i_ceph_lock held.
309  */
310 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
311 {
312 	struct ceph_cap *cap;
313 	struct rb_node *n = ci->i_caps.rb_node;
314 
315 	while (n) {
316 		cap = rb_entry(n, struct ceph_cap, ci_node);
317 		if (mds < cap->mds)
318 			n = n->rb_left;
319 		else if (mds > cap->mds)
320 			n = n->rb_right;
321 		else
322 			return cap;
323 	}
324 	return NULL;
325 }
326 
327 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
328 {
329 	struct ceph_cap *cap;
330 
331 	spin_lock(&ci->i_ceph_lock);
332 	cap = __get_cap_for_mds(ci, mds);
333 	spin_unlock(&ci->i_ceph_lock);
334 	return cap;
335 }
336 
337 /*
338  * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
339  */
340 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
341 {
342 	struct ceph_cap *cap;
343 	int mds = -1;
344 	struct rb_node *p;
345 
346 	/* prefer mds with WR|BUFFER|EXCL caps */
347 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
348 		cap = rb_entry(p, struct ceph_cap, ci_node);
349 		mds = cap->mds;
350 		if (cap->issued & (CEPH_CAP_FILE_WR |
351 				   CEPH_CAP_FILE_BUFFER |
352 				   CEPH_CAP_FILE_EXCL))
353 			break;
354 	}
355 	return mds;
356 }
357 
358 int ceph_get_cap_mds(struct inode *inode)
359 {
360 	struct ceph_inode_info *ci = ceph_inode(inode);
361 	int mds;
362 	spin_lock(&ci->i_ceph_lock);
363 	mds = __ceph_get_cap_mds(ceph_inode(inode));
364 	spin_unlock(&ci->i_ceph_lock);
365 	return mds;
366 }
367 
368 /*
369  * Called under i_ceph_lock.
370  */
371 static void __insert_cap_node(struct ceph_inode_info *ci,
372 			      struct ceph_cap *new)
373 {
374 	struct rb_node **p = &ci->i_caps.rb_node;
375 	struct rb_node *parent = NULL;
376 	struct ceph_cap *cap = NULL;
377 
378 	while (*p) {
379 		parent = *p;
380 		cap = rb_entry(parent, struct ceph_cap, ci_node);
381 		if (new->mds < cap->mds)
382 			p = &(*p)->rb_left;
383 		else if (new->mds > cap->mds)
384 			p = &(*p)->rb_right;
385 		else
386 			BUG();
387 	}
388 
389 	rb_link_node(&new->ci_node, parent, p);
390 	rb_insert_color(&new->ci_node, &ci->i_caps);
391 }
392 
393 /*
394  * (re)set cap hold timeouts, which control the delayed release
395  * of unused caps back to the MDS.  Should be called on cap use.
396  */
397 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
398 			       struct ceph_inode_info *ci)
399 {
400 	struct ceph_mount_options *ma = mdsc->fsc->mount_options;
401 
402 	ci->i_hold_caps_min = round_jiffies(jiffies +
403 					    ma->caps_wanted_delay_min * HZ);
404 	ci->i_hold_caps_max = round_jiffies(jiffies +
405 					    ma->caps_wanted_delay_max * HZ);
406 	dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
407 	     ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
408 }
409 
410 /*
411  * (Re)queue cap at the end of the delayed cap release list.
412  *
413  * If I_FLUSH is set, leave the inode at the front of the list.
414  *
415  * Caller holds i_ceph_lock
416  *    -> we take mdsc->cap_delay_lock
417  */
418 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
419 				struct ceph_inode_info *ci)
420 {
421 	__cap_set_timeouts(mdsc, ci);
422 	dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
423 	     ci->i_ceph_flags, ci->i_hold_caps_max);
424 	if (!mdsc->stopping) {
425 		spin_lock(&mdsc->cap_delay_lock);
426 		if (!list_empty(&ci->i_cap_delay_list)) {
427 			if (ci->i_ceph_flags & CEPH_I_FLUSH)
428 				goto no_change;
429 			list_del_init(&ci->i_cap_delay_list);
430 		}
431 		list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
432 no_change:
433 		spin_unlock(&mdsc->cap_delay_lock);
434 	}
435 }
436 
437 /*
438  * Queue an inode for immediate writeback.  Mark inode with I_FLUSH,
439  * indicating we should send a cap message to flush dirty metadata
440  * asap, and move to the front of the delayed cap list.
441  */
442 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
443 				      struct ceph_inode_info *ci)
444 {
445 	dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
446 	spin_lock(&mdsc->cap_delay_lock);
447 	ci->i_ceph_flags |= CEPH_I_FLUSH;
448 	if (!list_empty(&ci->i_cap_delay_list))
449 		list_del_init(&ci->i_cap_delay_list);
450 	list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
451 	spin_unlock(&mdsc->cap_delay_lock);
452 }
453 
454 /*
455  * Cancel delayed work on cap.
456  *
457  * Caller must hold i_ceph_lock.
458  */
459 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
460 			       struct ceph_inode_info *ci)
461 {
462 	dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
463 	if (list_empty(&ci->i_cap_delay_list))
464 		return;
465 	spin_lock(&mdsc->cap_delay_lock);
466 	list_del_init(&ci->i_cap_delay_list);
467 	spin_unlock(&mdsc->cap_delay_lock);
468 }
469 
470 /*
471  * Common issue checks for add_cap, handle_cap_grant.
472  */
473 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
474 			      unsigned issued)
475 {
476 	unsigned had = __ceph_caps_issued(ci, NULL);
477 
478 	/*
479 	 * Each time we receive FILE_CACHE anew, we increment
480 	 * i_rdcache_gen.
481 	 */
482 	if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
483 	    (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
484 		ci->i_rdcache_gen++;
485 	}
486 
487 	/*
488 	 * if we are newly issued FILE_SHARED, mark dir not complete; we
489 	 * don't know what happened to this directory while we didn't
490 	 * have the cap.
491 	 */
492 	if ((issued & CEPH_CAP_FILE_SHARED) &&
493 	    (had & CEPH_CAP_FILE_SHARED) == 0) {
494 		ci->i_shared_gen++;
495 		if (S_ISDIR(ci->vfs_inode.i_mode)) {
496 			dout(" marking %p NOT complete\n", &ci->vfs_inode);
497 			__ceph_dir_clear_complete(ci);
498 		}
499 	}
500 }
501 
502 /*
503  * Add a capability under the given MDS session.
504  *
505  * Caller should hold session snap_rwsem (read) and s_mutex.
506  *
507  * @fmode is the open file mode, if we are opening a file, otherwise
508  * it is < 0.  (This is so we can atomically add the cap and add an
509  * open file reference to it.)
510  */
511 int ceph_add_cap(struct inode *inode,
512 		 struct ceph_mds_session *session, u64 cap_id,
513 		 int fmode, unsigned issued, unsigned wanted,
514 		 unsigned seq, unsigned mseq, u64 realmino, int flags,
515 		 struct ceph_cap_reservation *caps_reservation)
516 {
517 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
518 	struct ceph_inode_info *ci = ceph_inode(inode);
519 	struct ceph_cap *new_cap = NULL;
520 	struct ceph_cap *cap;
521 	int mds = session->s_mds;
522 	int actual_wanted;
523 
524 	dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
525 	     session->s_mds, cap_id, ceph_cap_string(issued), seq);
526 
527 	/*
528 	 * If we are opening the file, include file mode wanted bits
529 	 * in wanted.
530 	 */
531 	if (fmode >= 0)
532 		wanted |= ceph_caps_for_mode(fmode);
533 
534 retry:
535 	spin_lock(&ci->i_ceph_lock);
536 	cap = __get_cap_for_mds(ci, mds);
537 	if (!cap) {
538 		if (new_cap) {
539 			cap = new_cap;
540 			new_cap = NULL;
541 		} else {
542 			spin_unlock(&ci->i_ceph_lock);
543 			new_cap = get_cap(mdsc, caps_reservation);
544 			if (new_cap == NULL)
545 				return -ENOMEM;
546 			goto retry;
547 		}
548 
549 		cap->issued = 0;
550 		cap->implemented = 0;
551 		cap->mds = mds;
552 		cap->mds_wanted = 0;
553 		cap->mseq = 0;
554 
555 		cap->ci = ci;
556 		__insert_cap_node(ci, cap);
557 
558 		/* add to session cap list */
559 		cap->session = session;
560 		spin_lock(&session->s_cap_lock);
561 		list_add_tail(&cap->session_caps, &session->s_caps);
562 		session->s_nr_caps++;
563 		spin_unlock(&session->s_cap_lock);
564 	} else {
565 		if (new_cap)
566 			ceph_put_cap(mdsc, new_cap);
567 
568 		/*
569 		 * auth mds of the inode changed. we received the cap export
570 		 * message, but still haven't received the cap import message.
571 		 * handle_cap_export() updated the new auth MDS' cap.
572 		 *
573 		 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
574 		 * a message that was send before the cap import message. So
575 		 * don't remove caps.
576 		 */
577 		if (ceph_seq_cmp(seq, cap->seq) <= 0) {
578 			WARN_ON(cap != ci->i_auth_cap);
579 			WARN_ON(cap->cap_id != cap_id);
580 			seq = cap->seq;
581 			mseq = cap->mseq;
582 			issued |= cap->issued;
583 			flags |= CEPH_CAP_FLAG_AUTH;
584 		}
585 	}
586 
587 	if (!ci->i_snap_realm) {
588 		/*
589 		 * add this inode to the appropriate snap realm
590 		 */
591 		struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
592 							       realmino);
593 		if (realm) {
594 			ceph_get_snap_realm(mdsc, realm);
595 			spin_lock(&realm->inodes_with_caps_lock);
596 			ci->i_snap_realm = realm;
597 			list_add(&ci->i_snap_realm_item,
598 				 &realm->inodes_with_caps);
599 			spin_unlock(&realm->inodes_with_caps_lock);
600 		} else {
601 			pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
602 			       realmino);
603 			WARN_ON(!realm);
604 		}
605 	}
606 
607 	__check_cap_issue(ci, cap, issued);
608 
609 	/*
610 	 * If we are issued caps we don't want, or the mds' wanted
611 	 * value appears to be off, queue a check so we'll release
612 	 * later and/or update the mds wanted value.
613 	 */
614 	actual_wanted = __ceph_caps_wanted(ci);
615 	if ((wanted & ~actual_wanted) ||
616 	    (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
617 		dout(" issued %s, mds wanted %s, actual %s, queueing\n",
618 		     ceph_cap_string(issued), ceph_cap_string(wanted),
619 		     ceph_cap_string(actual_wanted));
620 		__cap_delay_requeue(mdsc, ci);
621 	}
622 
623 	if (flags & CEPH_CAP_FLAG_AUTH) {
624 		if (ci->i_auth_cap == NULL ||
625 		    ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
626 			ci->i_auth_cap = cap;
627 			cap->mds_wanted = wanted;
628 		}
629 		ci->i_cap_exporting_issued = 0;
630 	} else {
631 		WARN_ON(ci->i_auth_cap == cap);
632 	}
633 
634 	dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
635 	     inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
636 	     ceph_cap_string(issued|cap->issued), seq, mds);
637 	cap->cap_id = cap_id;
638 	cap->issued = issued;
639 	cap->implemented |= issued;
640 	if (ceph_seq_cmp(mseq, cap->mseq) > 0)
641 		cap->mds_wanted = wanted;
642 	else
643 		cap->mds_wanted |= wanted;
644 	cap->seq = seq;
645 	cap->issue_seq = seq;
646 	cap->mseq = mseq;
647 	cap->cap_gen = session->s_cap_gen;
648 
649 	if (fmode >= 0)
650 		__ceph_get_fmode(ci, fmode);
651 	spin_unlock(&ci->i_ceph_lock);
652 	wake_up_all(&ci->i_cap_wq);
653 	return 0;
654 }
655 
656 /*
657  * Return true if cap has not timed out and belongs to the current
658  * generation of the MDS session (i.e. has not gone 'stale' due to
659  * us losing touch with the mds).
660  */
661 static int __cap_is_valid(struct ceph_cap *cap)
662 {
663 	unsigned long ttl;
664 	u32 gen;
665 
666 	spin_lock(&cap->session->s_gen_ttl_lock);
667 	gen = cap->session->s_cap_gen;
668 	ttl = cap->session->s_cap_ttl;
669 	spin_unlock(&cap->session->s_gen_ttl_lock);
670 
671 	if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
672 		dout("__cap_is_valid %p cap %p issued %s "
673 		     "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
674 		     cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
675 		return 0;
676 	}
677 
678 	return 1;
679 }
680 
681 /*
682  * Return set of valid cap bits issued to us.  Note that caps time
683  * out, and may be invalidated in bulk if the client session times out
684  * and session->s_cap_gen is bumped.
685  */
686 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
687 {
688 	int have = ci->i_snap_caps | ci->i_cap_exporting_issued;
689 	struct ceph_cap *cap;
690 	struct rb_node *p;
691 
692 	if (implemented)
693 		*implemented = 0;
694 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
695 		cap = rb_entry(p, struct ceph_cap, ci_node);
696 		if (!__cap_is_valid(cap))
697 			continue;
698 		dout("__ceph_caps_issued %p cap %p issued %s\n",
699 		     &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
700 		have |= cap->issued;
701 		if (implemented)
702 			*implemented |= cap->implemented;
703 	}
704 	/*
705 	 * exclude caps issued by non-auth MDS, but are been revoking
706 	 * by the auth MDS. The non-auth MDS should be revoking/exporting
707 	 * these caps, but the message is delayed.
708 	 */
709 	if (ci->i_auth_cap) {
710 		cap = ci->i_auth_cap;
711 		have &= ~cap->implemented | cap->issued;
712 	}
713 	return have;
714 }
715 
716 /*
717  * Get cap bits issued by caps other than @ocap
718  */
719 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
720 {
721 	int have = ci->i_snap_caps;
722 	struct ceph_cap *cap;
723 	struct rb_node *p;
724 
725 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
726 		cap = rb_entry(p, struct ceph_cap, ci_node);
727 		if (cap == ocap)
728 			continue;
729 		if (!__cap_is_valid(cap))
730 			continue;
731 		have |= cap->issued;
732 	}
733 	return have;
734 }
735 
736 /*
737  * Move a cap to the end of the LRU (oldest caps at list head, newest
738  * at list tail).
739  */
740 static void __touch_cap(struct ceph_cap *cap)
741 {
742 	struct ceph_mds_session *s = cap->session;
743 
744 	spin_lock(&s->s_cap_lock);
745 	if (s->s_cap_iterator == NULL) {
746 		dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
747 		     s->s_mds);
748 		list_move_tail(&cap->session_caps, &s->s_caps);
749 	} else {
750 		dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
751 		     &cap->ci->vfs_inode, cap, s->s_mds);
752 	}
753 	spin_unlock(&s->s_cap_lock);
754 }
755 
756 /*
757  * Check if we hold the given mask.  If so, move the cap(s) to the
758  * front of their respective LRUs.  (This is the preferred way for
759  * callers to check for caps they want.)
760  */
761 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
762 {
763 	struct ceph_cap *cap;
764 	struct rb_node *p;
765 	int have = ci->i_snap_caps;
766 
767 	if ((have & mask) == mask) {
768 		dout("__ceph_caps_issued_mask %p snap issued %s"
769 		     " (mask %s)\n", &ci->vfs_inode,
770 		     ceph_cap_string(have),
771 		     ceph_cap_string(mask));
772 		return 1;
773 	}
774 
775 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
776 		cap = rb_entry(p, struct ceph_cap, ci_node);
777 		if (!__cap_is_valid(cap))
778 			continue;
779 		if ((cap->issued & mask) == mask) {
780 			dout("__ceph_caps_issued_mask %p cap %p issued %s"
781 			     " (mask %s)\n", &ci->vfs_inode, cap,
782 			     ceph_cap_string(cap->issued),
783 			     ceph_cap_string(mask));
784 			if (touch)
785 				__touch_cap(cap);
786 			return 1;
787 		}
788 
789 		/* does a combination of caps satisfy mask? */
790 		have |= cap->issued;
791 		if ((have & mask) == mask) {
792 			dout("__ceph_caps_issued_mask %p combo issued %s"
793 			     " (mask %s)\n", &ci->vfs_inode,
794 			     ceph_cap_string(cap->issued),
795 			     ceph_cap_string(mask));
796 			if (touch) {
797 				struct rb_node *q;
798 
799 				/* touch this + preceding caps */
800 				__touch_cap(cap);
801 				for (q = rb_first(&ci->i_caps); q != p;
802 				     q = rb_next(q)) {
803 					cap = rb_entry(q, struct ceph_cap,
804 						       ci_node);
805 					if (!__cap_is_valid(cap))
806 						continue;
807 					__touch_cap(cap);
808 				}
809 			}
810 			return 1;
811 		}
812 	}
813 
814 	return 0;
815 }
816 
817 /*
818  * Return true if mask caps are currently being revoked by an MDS.
819  */
820 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
821 			       struct ceph_cap *ocap, int mask)
822 {
823 	struct ceph_cap *cap;
824 	struct rb_node *p;
825 
826 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
827 		cap = rb_entry(p, struct ceph_cap, ci_node);
828 		if (cap != ocap &&
829 		    (cap->implemented & ~cap->issued & mask))
830 			return 1;
831 	}
832 	return 0;
833 }
834 
835 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
836 {
837 	struct inode *inode = &ci->vfs_inode;
838 	int ret;
839 
840 	spin_lock(&ci->i_ceph_lock);
841 	ret = __ceph_caps_revoking_other(ci, NULL, mask);
842 	spin_unlock(&ci->i_ceph_lock);
843 	dout("ceph_caps_revoking %p %s = %d\n", inode,
844 	     ceph_cap_string(mask), ret);
845 	return ret;
846 }
847 
848 int __ceph_caps_used(struct ceph_inode_info *ci)
849 {
850 	int used = 0;
851 	if (ci->i_pin_ref)
852 		used |= CEPH_CAP_PIN;
853 	if (ci->i_rd_ref)
854 		used |= CEPH_CAP_FILE_RD;
855 	if (ci->i_rdcache_ref || ci->vfs_inode.i_data.nrpages)
856 		used |= CEPH_CAP_FILE_CACHE;
857 	if (ci->i_wr_ref)
858 		used |= CEPH_CAP_FILE_WR;
859 	if (ci->i_wb_ref || ci->i_wrbuffer_ref)
860 		used |= CEPH_CAP_FILE_BUFFER;
861 	return used;
862 }
863 
864 /*
865  * wanted, by virtue of open file modes
866  */
867 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
868 {
869 	int want = 0;
870 	int mode;
871 	for (mode = 0; mode < CEPH_FILE_MODE_NUM; mode++)
872 		if (ci->i_nr_by_mode[mode])
873 			want |= ceph_caps_for_mode(mode);
874 	return want;
875 }
876 
877 /*
878  * Return caps we have registered with the MDS(s) as 'wanted'.
879  */
880 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
881 {
882 	struct ceph_cap *cap;
883 	struct rb_node *p;
884 	int mds_wanted = 0;
885 
886 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
887 		cap = rb_entry(p, struct ceph_cap, ci_node);
888 		if (!__cap_is_valid(cap))
889 			continue;
890 		if (cap == ci->i_auth_cap)
891 			mds_wanted |= cap->mds_wanted;
892 		else
893 			mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
894 	}
895 	return mds_wanted;
896 }
897 
898 /*
899  * called under i_ceph_lock
900  */
901 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
902 {
903 	return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_issued;
904 }
905 
906 int ceph_is_any_caps(struct inode *inode)
907 {
908 	struct ceph_inode_info *ci = ceph_inode(inode);
909 	int ret;
910 
911 	spin_lock(&ci->i_ceph_lock);
912 	ret = __ceph_is_any_caps(ci);
913 	spin_unlock(&ci->i_ceph_lock);
914 
915 	return ret;
916 }
917 
918 /*
919  * Remove a cap.  Take steps to deal with a racing iterate_session_caps.
920  *
921  * caller should hold i_ceph_lock.
922  * caller will not hold session s_mutex if called from destroy_inode.
923  */
924 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
925 {
926 	struct ceph_mds_session *session = cap->session;
927 	struct ceph_inode_info *ci = cap->ci;
928 	struct ceph_mds_client *mdsc =
929 		ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
930 	int removed = 0;
931 
932 	dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
933 
934 	/* remove from session list */
935 	spin_lock(&session->s_cap_lock);
936 	/*
937 	 * s_cap_reconnect is protected by s_cap_lock. no one changes
938 	 * s_cap_gen while session is in the reconnect state.
939 	 */
940 	if (queue_release &&
941 	    (!session->s_cap_reconnect ||
942 	     cap->cap_gen == session->s_cap_gen))
943 		__queue_cap_release(session, ci->i_vino.ino, cap->cap_id,
944 				    cap->mseq, cap->issue_seq);
945 
946 	if (session->s_cap_iterator == cap) {
947 		/* not yet, we are iterating over this very cap */
948 		dout("__ceph_remove_cap  delaying %p removal from session %p\n",
949 		     cap, cap->session);
950 	} else {
951 		list_del_init(&cap->session_caps);
952 		session->s_nr_caps--;
953 		cap->session = NULL;
954 		removed = 1;
955 	}
956 	/* protect backpointer with s_cap_lock: see iterate_session_caps */
957 	cap->ci = NULL;
958 	spin_unlock(&session->s_cap_lock);
959 
960 	/* remove from inode list */
961 	rb_erase(&cap->ci_node, &ci->i_caps);
962 	if (ci->i_auth_cap == cap)
963 		ci->i_auth_cap = NULL;
964 
965 	if (removed)
966 		ceph_put_cap(mdsc, cap);
967 
968 	if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
969 		struct ceph_snap_realm *realm = ci->i_snap_realm;
970 		spin_lock(&realm->inodes_with_caps_lock);
971 		list_del_init(&ci->i_snap_realm_item);
972 		ci->i_snap_realm_counter++;
973 		ci->i_snap_realm = NULL;
974 		spin_unlock(&realm->inodes_with_caps_lock);
975 		ceph_put_snap_realm(mdsc, realm);
976 	}
977 	if (!__ceph_is_any_real_caps(ci))
978 		__cap_delay_cancel(mdsc, ci);
979 }
980 
981 /*
982  * Build and send a cap message to the given MDS.
983  *
984  * Caller should be holding s_mutex.
985  */
986 static int send_cap_msg(struct ceph_mds_session *session,
987 			u64 ino, u64 cid, int op,
988 			int caps, int wanted, int dirty,
989 			u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
990 			u64 size, u64 max_size,
991 			struct timespec *mtime, struct timespec *atime,
992 			u64 time_warp_seq,
993 			kuid_t uid, kgid_t gid, umode_t mode,
994 			u64 xattr_version,
995 			struct ceph_buffer *xattrs_buf,
996 			u64 follows)
997 {
998 	struct ceph_mds_caps *fc;
999 	struct ceph_msg *msg;
1000 
1001 	dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1002 	     " seq %u/%u mseq %u follows %lld size %llu/%llu"
1003 	     " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
1004 	     cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
1005 	     ceph_cap_string(dirty),
1006 	     seq, issue_seq, mseq, follows, size, max_size,
1007 	     xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
1008 
1009 	msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), GFP_NOFS, false);
1010 	if (!msg)
1011 		return -ENOMEM;
1012 
1013 	msg->hdr.tid = cpu_to_le64(flush_tid);
1014 
1015 	fc = msg->front.iov_base;
1016 	memset(fc, 0, sizeof(*fc));
1017 
1018 	fc->cap_id = cpu_to_le64(cid);
1019 	fc->op = cpu_to_le32(op);
1020 	fc->seq = cpu_to_le32(seq);
1021 	fc->issue_seq = cpu_to_le32(issue_seq);
1022 	fc->migrate_seq = cpu_to_le32(mseq);
1023 	fc->caps = cpu_to_le32(caps);
1024 	fc->wanted = cpu_to_le32(wanted);
1025 	fc->dirty = cpu_to_le32(dirty);
1026 	fc->ino = cpu_to_le64(ino);
1027 	fc->snap_follows = cpu_to_le64(follows);
1028 
1029 	fc->size = cpu_to_le64(size);
1030 	fc->max_size = cpu_to_le64(max_size);
1031 	if (mtime)
1032 		ceph_encode_timespec(&fc->mtime, mtime);
1033 	if (atime)
1034 		ceph_encode_timespec(&fc->atime, atime);
1035 	fc->time_warp_seq = cpu_to_le32(time_warp_seq);
1036 
1037 	fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid));
1038 	fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid));
1039 	fc->mode = cpu_to_le32(mode);
1040 
1041 	fc->xattr_version = cpu_to_le64(xattr_version);
1042 	if (xattrs_buf) {
1043 		msg->middle = ceph_buffer_get(xattrs_buf);
1044 		fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1045 		msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1046 	}
1047 
1048 	ceph_con_send(&session->s_con, msg);
1049 	return 0;
1050 }
1051 
1052 void __queue_cap_release(struct ceph_mds_session *session,
1053 			 u64 ino, u64 cap_id, u32 migrate_seq,
1054 			 u32 issue_seq)
1055 {
1056 	struct ceph_msg *msg;
1057 	struct ceph_mds_cap_release *head;
1058 	struct ceph_mds_cap_item *item;
1059 
1060 	BUG_ON(!session->s_num_cap_releases);
1061 	msg = list_first_entry(&session->s_cap_releases,
1062 			       struct ceph_msg, list_head);
1063 
1064 	dout(" adding %llx release to mds%d msg %p (%d left)\n",
1065 	     ino, session->s_mds, msg, session->s_num_cap_releases);
1066 
1067 	BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1068 	head = msg->front.iov_base;
1069 	le32_add_cpu(&head->num, 1);
1070 	item = msg->front.iov_base + msg->front.iov_len;
1071 	item->ino = cpu_to_le64(ino);
1072 	item->cap_id = cpu_to_le64(cap_id);
1073 	item->migrate_seq = cpu_to_le32(migrate_seq);
1074 	item->seq = cpu_to_le32(issue_seq);
1075 
1076 	session->s_num_cap_releases--;
1077 
1078 	msg->front.iov_len += sizeof(*item);
1079 	if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1080 		dout(" release msg %p full\n", msg);
1081 		list_move_tail(&msg->list_head, &session->s_cap_releases_done);
1082 	} else {
1083 		dout(" release msg %p at %d/%d (%d)\n", msg,
1084 		     (int)le32_to_cpu(head->num),
1085 		     (int)CEPH_CAPS_PER_RELEASE,
1086 		     (int)msg->front.iov_len);
1087 	}
1088 }
1089 
1090 /*
1091  * Queue cap releases when an inode is dropped from our cache.  Since
1092  * inode is about to be destroyed, there is no need for i_ceph_lock.
1093  */
1094 void ceph_queue_caps_release(struct inode *inode)
1095 {
1096 	struct ceph_inode_info *ci = ceph_inode(inode);
1097 	struct rb_node *p;
1098 
1099 	p = rb_first(&ci->i_caps);
1100 	while (p) {
1101 		struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1102 		p = rb_next(p);
1103 		__ceph_remove_cap(cap, true);
1104 	}
1105 }
1106 
1107 /*
1108  * Send a cap msg on the given inode.  Update our caps state, then
1109  * drop i_ceph_lock and send the message.
1110  *
1111  * Make note of max_size reported/requested from mds, revoked caps
1112  * that have now been implemented.
1113  *
1114  * Make half-hearted attempt ot to invalidate page cache if we are
1115  * dropping RDCACHE.  Note that this will leave behind locked pages
1116  * that we'll then need to deal with elsewhere.
1117  *
1118  * Return non-zero if delayed release, or we experienced an error
1119  * such that the caller should requeue + retry later.
1120  *
1121  * called with i_ceph_lock, then drops it.
1122  * caller should hold snap_rwsem (read), s_mutex.
1123  */
1124 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1125 		      int op, int used, int want, int retain, int flushing,
1126 		      unsigned *pflush_tid)
1127 	__releases(cap->ci->i_ceph_lock)
1128 {
1129 	struct ceph_inode_info *ci = cap->ci;
1130 	struct inode *inode = &ci->vfs_inode;
1131 	u64 cap_id = cap->cap_id;
1132 	int held, revoking, dropping, keep;
1133 	u64 seq, issue_seq, mseq, time_warp_seq, follows;
1134 	u64 size, max_size;
1135 	struct timespec mtime, atime;
1136 	int wake = 0;
1137 	umode_t mode;
1138 	kuid_t uid;
1139 	kgid_t gid;
1140 	struct ceph_mds_session *session;
1141 	u64 xattr_version = 0;
1142 	struct ceph_buffer *xattr_blob = NULL;
1143 	int delayed = 0;
1144 	u64 flush_tid = 0;
1145 	int i;
1146 	int ret;
1147 
1148 	held = cap->issued | cap->implemented;
1149 	revoking = cap->implemented & ~cap->issued;
1150 	retain &= ~revoking;
1151 	dropping = cap->issued & ~retain;
1152 
1153 	dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1154 	     inode, cap, cap->session,
1155 	     ceph_cap_string(held), ceph_cap_string(held & retain),
1156 	     ceph_cap_string(revoking));
1157 	BUG_ON((retain & CEPH_CAP_PIN) == 0);
1158 
1159 	session = cap->session;
1160 
1161 	/* don't release wanted unless we've waited a bit. */
1162 	if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1163 	    time_before(jiffies, ci->i_hold_caps_min)) {
1164 		dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1165 		     ceph_cap_string(cap->issued),
1166 		     ceph_cap_string(cap->issued & retain),
1167 		     ceph_cap_string(cap->mds_wanted),
1168 		     ceph_cap_string(want));
1169 		want |= cap->mds_wanted;
1170 		retain |= cap->issued;
1171 		delayed = 1;
1172 	}
1173 	ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1174 
1175 	cap->issued &= retain;  /* drop bits we don't want */
1176 	if (cap->implemented & ~cap->issued) {
1177 		/*
1178 		 * Wake up any waiters on wanted -> needed transition.
1179 		 * This is due to the weird transition from buffered
1180 		 * to sync IO... we need to flush dirty pages _before_
1181 		 * allowing sync writes to avoid reordering.
1182 		 */
1183 		wake = 1;
1184 	}
1185 	cap->implemented &= cap->issued | used;
1186 	cap->mds_wanted = want;
1187 
1188 	if (flushing) {
1189 		/*
1190 		 * assign a tid for flush operations so we can avoid
1191 		 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1192 		 * clean type races.  track latest tid for every bit
1193 		 * so we can handle flush AxFw, flush Fw, and have the
1194 		 * first ack clean Ax.
1195 		 */
1196 		flush_tid = ++ci->i_cap_flush_last_tid;
1197 		if (pflush_tid)
1198 			*pflush_tid = flush_tid;
1199 		dout(" cap_flush_tid %d\n", (int)flush_tid);
1200 		for (i = 0; i < CEPH_CAP_BITS; i++)
1201 			if (flushing & (1 << i))
1202 				ci->i_cap_flush_tid[i] = flush_tid;
1203 
1204 		follows = ci->i_head_snapc->seq;
1205 	} else {
1206 		follows = 0;
1207 	}
1208 
1209 	keep = cap->implemented;
1210 	seq = cap->seq;
1211 	issue_seq = cap->issue_seq;
1212 	mseq = cap->mseq;
1213 	size = inode->i_size;
1214 	ci->i_reported_size = size;
1215 	max_size = ci->i_wanted_max_size;
1216 	ci->i_requested_max_size = max_size;
1217 	mtime = inode->i_mtime;
1218 	atime = inode->i_atime;
1219 	time_warp_seq = ci->i_time_warp_seq;
1220 	uid = inode->i_uid;
1221 	gid = inode->i_gid;
1222 	mode = inode->i_mode;
1223 
1224 	if (flushing & CEPH_CAP_XATTR_EXCL) {
1225 		__ceph_build_xattrs_blob(ci);
1226 		xattr_blob = ci->i_xattrs.blob;
1227 		xattr_version = ci->i_xattrs.version;
1228 	}
1229 
1230 	spin_unlock(&ci->i_ceph_lock);
1231 
1232 	ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1233 		op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1234 		size, max_size, &mtime, &atime, time_warp_seq,
1235 		uid, gid, mode, xattr_version, xattr_blob,
1236 		follows);
1237 	if (ret < 0) {
1238 		dout("error sending cap msg, must requeue %p\n", inode);
1239 		delayed = 1;
1240 	}
1241 
1242 	if (wake)
1243 		wake_up_all(&ci->i_cap_wq);
1244 
1245 	return delayed;
1246 }
1247 
1248 /*
1249  * When a snapshot is taken, clients accumulate dirty metadata on
1250  * inodes with capabilities in ceph_cap_snaps to describe the file
1251  * state at the time the snapshot was taken.  This must be flushed
1252  * asynchronously back to the MDS once sync writes complete and dirty
1253  * data is written out.
1254  *
1255  * Unless @again is true, skip cap_snaps that were already sent to
1256  * the MDS (i.e., during this session).
1257  *
1258  * Called under i_ceph_lock.  Takes s_mutex as needed.
1259  */
1260 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1261 			struct ceph_mds_session **psession,
1262 			int again)
1263 		__releases(ci->i_ceph_lock)
1264 		__acquires(ci->i_ceph_lock)
1265 {
1266 	struct inode *inode = &ci->vfs_inode;
1267 	int mds;
1268 	struct ceph_cap_snap *capsnap;
1269 	u32 mseq;
1270 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1271 	struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1272 						    session->s_mutex */
1273 	u64 next_follows = 0;  /* keep track of how far we've gotten through the
1274 			     i_cap_snaps list, and skip these entries next time
1275 			     around to avoid an infinite loop */
1276 
1277 	if (psession)
1278 		session = *psession;
1279 
1280 	dout("__flush_snaps %p\n", inode);
1281 retry:
1282 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1283 		/* avoid an infiniute loop after retry */
1284 		if (capsnap->follows < next_follows)
1285 			continue;
1286 		/*
1287 		 * we need to wait for sync writes to complete and for dirty
1288 		 * pages to be written out.
1289 		 */
1290 		if (capsnap->dirty_pages || capsnap->writing)
1291 			break;
1292 
1293 		/*
1294 		 * if cap writeback already occurred, we should have dropped
1295 		 * the capsnap in ceph_put_wrbuffer_cap_refs.
1296 		 */
1297 		BUG_ON(capsnap->dirty == 0);
1298 
1299 		/* pick mds, take s_mutex */
1300 		if (ci->i_auth_cap == NULL) {
1301 			dout("no auth cap (migrating?), doing nothing\n");
1302 			goto out;
1303 		}
1304 
1305 		/* only flush each capsnap once */
1306 		if (!again && !list_empty(&capsnap->flushing_item)) {
1307 			dout("already flushed %p, skipping\n", capsnap);
1308 			continue;
1309 		}
1310 
1311 		mds = ci->i_auth_cap->session->s_mds;
1312 		mseq = ci->i_auth_cap->mseq;
1313 
1314 		if (session && session->s_mds != mds) {
1315 			dout("oops, wrong session %p mutex\n", session);
1316 			mutex_unlock(&session->s_mutex);
1317 			ceph_put_mds_session(session);
1318 			session = NULL;
1319 		}
1320 		if (!session) {
1321 			spin_unlock(&ci->i_ceph_lock);
1322 			mutex_lock(&mdsc->mutex);
1323 			session = __ceph_lookup_mds_session(mdsc, mds);
1324 			mutex_unlock(&mdsc->mutex);
1325 			if (session) {
1326 				dout("inverting session/ino locks on %p\n",
1327 				     session);
1328 				mutex_lock(&session->s_mutex);
1329 			}
1330 			/*
1331 			 * if session == NULL, we raced against a cap
1332 			 * deletion or migration.  retry, and we'll
1333 			 * get a better @mds value next time.
1334 			 */
1335 			spin_lock(&ci->i_ceph_lock);
1336 			goto retry;
1337 		}
1338 
1339 		capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1340 		atomic_inc(&capsnap->nref);
1341 		if (!list_empty(&capsnap->flushing_item))
1342 			list_del_init(&capsnap->flushing_item);
1343 		list_add_tail(&capsnap->flushing_item,
1344 			      &session->s_cap_snaps_flushing);
1345 		spin_unlock(&ci->i_ceph_lock);
1346 
1347 		dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1348 		     inode, capsnap, capsnap->follows, capsnap->flush_tid);
1349 		send_cap_msg(session, ceph_vino(inode).ino, 0,
1350 			     CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1351 			     capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1352 			     capsnap->size, 0,
1353 			     &capsnap->mtime, &capsnap->atime,
1354 			     capsnap->time_warp_seq,
1355 			     capsnap->uid, capsnap->gid, capsnap->mode,
1356 			     capsnap->xattr_version, capsnap->xattr_blob,
1357 			     capsnap->follows);
1358 
1359 		next_follows = capsnap->follows + 1;
1360 		ceph_put_cap_snap(capsnap);
1361 
1362 		spin_lock(&ci->i_ceph_lock);
1363 		goto retry;
1364 	}
1365 
1366 	/* we flushed them all; remove this inode from the queue */
1367 	spin_lock(&mdsc->snap_flush_lock);
1368 	list_del_init(&ci->i_snap_flush_item);
1369 	spin_unlock(&mdsc->snap_flush_lock);
1370 
1371 out:
1372 	if (psession)
1373 		*psession = session;
1374 	else if (session) {
1375 		mutex_unlock(&session->s_mutex);
1376 		ceph_put_mds_session(session);
1377 	}
1378 }
1379 
1380 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1381 {
1382 	spin_lock(&ci->i_ceph_lock);
1383 	__ceph_flush_snaps(ci, NULL, 0);
1384 	spin_unlock(&ci->i_ceph_lock);
1385 }
1386 
1387 /*
1388  * Mark caps dirty.  If inode is newly dirty, return the dirty flags.
1389  * Caller is then responsible for calling __mark_inode_dirty with the
1390  * returned flags value.
1391  */
1392 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1393 {
1394 	struct ceph_mds_client *mdsc =
1395 		ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1396 	struct inode *inode = &ci->vfs_inode;
1397 	int was = ci->i_dirty_caps;
1398 	int dirty = 0;
1399 
1400 	dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1401 	     ceph_cap_string(mask), ceph_cap_string(was),
1402 	     ceph_cap_string(was | mask));
1403 	ci->i_dirty_caps |= mask;
1404 	if (was == 0) {
1405 		if (!ci->i_head_snapc)
1406 			ci->i_head_snapc = ceph_get_snap_context(
1407 				ci->i_snap_realm->cached_context);
1408 		dout(" inode %p now dirty snapc %p auth cap %p\n",
1409 		     &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1410 		WARN_ON(!ci->i_auth_cap);
1411 		BUG_ON(!list_empty(&ci->i_dirty_item));
1412 		spin_lock(&mdsc->cap_dirty_lock);
1413 		list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1414 		spin_unlock(&mdsc->cap_dirty_lock);
1415 		if (ci->i_flushing_caps == 0) {
1416 			ihold(inode);
1417 			dirty |= I_DIRTY_SYNC;
1418 		}
1419 	}
1420 	BUG_ON(list_empty(&ci->i_dirty_item));
1421 	if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1422 	    (mask & CEPH_CAP_FILE_BUFFER))
1423 		dirty |= I_DIRTY_DATASYNC;
1424 	__cap_delay_requeue(mdsc, ci);
1425 	return dirty;
1426 }
1427 
1428 /*
1429  * Add dirty inode to the flushing list.  Assigned a seq number so we
1430  * can wait for caps to flush without starving.
1431  *
1432  * Called under i_ceph_lock.
1433  */
1434 static int __mark_caps_flushing(struct inode *inode,
1435 				 struct ceph_mds_session *session)
1436 {
1437 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1438 	struct ceph_inode_info *ci = ceph_inode(inode);
1439 	int flushing;
1440 
1441 	BUG_ON(ci->i_dirty_caps == 0);
1442 	BUG_ON(list_empty(&ci->i_dirty_item));
1443 
1444 	flushing = ci->i_dirty_caps;
1445 	dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1446 	     ceph_cap_string(flushing),
1447 	     ceph_cap_string(ci->i_flushing_caps),
1448 	     ceph_cap_string(ci->i_flushing_caps | flushing));
1449 	ci->i_flushing_caps |= flushing;
1450 	ci->i_dirty_caps = 0;
1451 	dout(" inode %p now !dirty\n", inode);
1452 
1453 	spin_lock(&mdsc->cap_dirty_lock);
1454 	list_del_init(&ci->i_dirty_item);
1455 
1456 	ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1457 	if (list_empty(&ci->i_flushing_item)) {
1458 		list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1459 		mdsc->num_cap_flushing++;
1460 		dout(" inode %p now flushing seq %lld\n", inode,
1461 		     ci->i_cap_flush_seq);
1462 	} else {
1463 		list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1464 		dout(" inode %p now flushing (more) seq %lld\n", inode,
1465 		     ci->i_cap_flush_seq);
1466 	}
1467 	spin_unlock(&mdsc->cap_dirty_lock);
1468 
1469 	return flushing;
1470 }
1471 
1472 /*
1473  * try to invalidate mapping pages without blocking.
1474  */
1475 static int try_nonblocking_invalidate(struct inode *inode)
1476 {
1477 	struct ceph_inode_info *ci = ceph_inode(inode);
1478 	u32 invalidating_gen = ci->i_rdcache_gen;
1479 
1480 	spin_unlock(&ci->i_ceph_lock);
1481 	invalidate_mapping_pages(&inode->i_data, 0, -1);
1482 	spin_lock(&ci->i_ceph_lock);
1483 
1484 	if (inode->i_data.nrpages == 0 &&
1485 	    invalidating_gen == ci->i_rdcache_gen) {
1486 		/* success. */
1487 		dout("try_nonblocking_invalidate %p success\n", inode);
1488 		/* save any racing async invalidate some trouble */
1489 		ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1490 		return 0;
1491 	}
1492 	dout("try_nonblocking_invalidate %p failed\n", inode);
1493 	return -1;
1494 }
1495 
1496 /*
1497  * Swiss army knife function to examine currently used and wanted
1498  * versus held caps.  Release, flush, ack revoked caps to mds as
1499  * appropriate.
1500  *
1501  *  CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1502  *    cap release further.
1503  *  CHECK_CAPS_AUTHONLY - we should only check the auth cap
1504  *  CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1505  *    further delay.
1506  */
1507 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1508 		     struct ceph_mds_session *session)
1509 {
1510 	struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1511 	struct ceph_mds_client *mdsc = fsc->mdsc;
1512 	struct inode *inode = &ci->vfs_inode;
1513 	struct ceph_cap *cap;
1514 	int file_wanted, used, cap_used;
1515 	int took_snap_rwsem = 0;             /* true if mdsc->snap_rwsem held */
1516 	int issued, implemented, want, retain, revoking, flushing = 0;
1517 	int mds = -1;   /* keep track of how far we've gone through i_caps list
1518 			   to avoid an infinite loop on retry */
1519 	struct rb_node *p;
1520 	int tried_invalidate = 0;
1521 	int delayed = 0, sent = 0, force_requeue = 0, num;
1522 	int queue_invalidate = 0;
1523 	int is_delayed = flags & CHECK_CAPS_NODELAY;
1524 
1525 	/* if we are unmounting, flush any unused caps immediately. */
1526 	if (mdsc->stopping)
1527 		is_delayed = 1;
1528 
1529 	spin_lock(&ci->i_ceph_lock);
1530 
1531 	if (ci->i_ceph_flags & CEPH_I_FLUSH)
1532 		flags |= CHECK_CAPS_FLUSH;
1533 
1534 	/* flush snaps first time around only */
1535 	if (!list_empty(&ci->i_cap_snaps))
1536 		__ceph_flush_snaps(ci, &session, 0);
1537 	goto retry_locked;
1538 retry:
1539 	spin_lock(&ci->i_ceph_lock);
1540 retry_locked:
1541 	file_wanted = __ceph_caps_file_wanted(ci);
1542 	used = __ceph_caps_used(ci);
1543 	want = file_wanted | used;
1544 	issued = __ceph_caps_issued(ci, &implemented);
1545 	revoking = implemented & ~issued;
1546 
1547 	retain = want | CEPH_CAP_PIN;
1548 	if (!mdsc->stopping && inode->i_nlink > 0) {
1549 		if (want) {
1550 			retain |= CEPH_CAP_ANY;       /* be greedy */
1551 		} else {
1552 			retain |= CEPH_CAP_ANY_SHARED;
1553 			/*
1554 			 * keep RD only if we didn't have the file open RW,
1555 			 * because then the mds would revoke it anyway to
1556 			 * journal max_size=0.
1557 			 */
1558 			if (ci->i_max_size == 0)
1559 				retain |= CEPH_CAP_ANY_RD;
1560 		}
1561 	}
1562 
1563 	dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1564 	     " issued %s revoking %s retain %s %s%s%s\n", inode,
1565 	     ceph_cap_string(file_wanted),
1566 	     ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1567 	     ceph_cap_string(ci->i_flushing_caps),
1568 	     ceph_cap_string(issued), ceph_cap_string(revoking),
1569 	     ceph_cap_string(retain),
1570 	     (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1571 	     (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1572 	     (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1573 
1574 	/*
1575 	 * If we no longer need to hold onto old our caps, and we may
1576 	 * have cached pages, but don't want them, then try to invalidate.
1577 	 * If we fail, it's because pages are locked.... try again later.
1578 	 */
1579 	if ((!is_delayed || mdsc->stopping) &&
1580 	    ci->i_wrbuffer_ref == 0 &&               /* no dirty pages... */
1581 	    inode->i_data.nrpages &&                 /* have cached pages */
1582 	    (file_wanted == 0 ||                     /* no open files */
1583 	     (revoking & (CEPH_CAP_FILE_CACHE|
1584 			  CEPH_CAP_FILE_LAZYIO))) && /*  or revoking cache */
1585 	    !tried_invalidate) {
1586 		dout("check_caps trying to invalidate on %p\n", inode);
1587 		if (try_nonblocking_invalidate(inode) < 0) {
1588 			if (revoking & (CEPH_CAP_FILE_CACHE|
1589 					CEPH_CAP_FILE_LAZYIO)) {
1590 				dout("check_caps queuing invalidate\n");
1591 				queue_invalidate = 1;
1592 				ci->i_rdcache_revoking = ci->i_rdcache_gen;
1593 			} else {
1594 				dout("check_caps failed to invalidate pages\n");
1595 				/* we failed to invalidate pages.  check these
1596 				   caps again later. */
1597 				force_requeue = 1;
1598 				__cap_set_timeouts(mdsc, ci);
1599 			}
1600 		}
1601 		tried_invalidate = 1;
1602 		goto retry_locked;
1603 	}
1604 
1605 	num = 0;
1606 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1607 		cap = rb_entry(p, struct ceph_cap, ci_node);
1608 		num++;
1609 
1610 		/* avoid looping forever */
1611 		if (mds >= cap->mds ||
1612 		    ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1613 			continue;
1614 
1615 		/* NOTE: no side-effects allowed, until we take s_mutex */
1616 
1617 		cap_used = used;
1618 		if (ci->i_auth_cap && cap != ci->i_auth_cap)
1619 			cap_used &= ~ci->i_auth_cap->issued;
1620 
1621 		revoking = cap->implemented & ~cap->issued;
1622 		dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1623 		     cap->mds, cap, ceph_cap_string(cap->issued),
1624 		     ceph_cap_string(cap_used),
1625 		     ceph_cap_string(cap->implemented),
1626 		     ceph_cap_string(revoking));
1627 
1628 		if (cap == ci->i_auth_cap &&
1629 		    (cap->issued & CEPH_CAP_FILE_WR)) {
1630 			/* request larger max_size from MDS? */
1631 			if (ci->i_wanted_max_size > ci->i_max_size &&
1632 			    ci->i_wanted_max_size > ci->i_requested_max_size) {
1633 				dout("requesting new max_size\n");
1634 				goto ack;
1635 			}
1636 
1637 			/* approaching file_max? */
1638 			if ((inode->i_size << 1) >= ci->i_max_size &&
1639 			    (ci->i_reported_size << 1) < ci->i_max_size) {
1640 				dout("i_size approaching max_size\n");
1641 				goto ack;
1642 			}
1643 		}
1644 		/* flush anything dirty? */
1645 		if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1646 		    ci->i_dirty_caps) {
1647 			dout("flushing dirty caps\n");
1648 			goto ack;
1649 		}
1650 
1651 		/* completed revocation? going down and there are no caps? */
1652 		if (revoking && (revoking & cap_used) == 0) {
1653 			dout("completed revocation of %s\n",
1654 			     ceph_cap_string(cap->implemented & ~cap->issued));
1655 			goto ack;
1656 		}
1657 
1658 		/* want more caps from mds? */
1659 		if (want & ~(cap->mds_wanted | cap->issued))
1660 			goto ack;
1661 
1662 		/* things we might delay */
1663 		if ((cap->issued & ~retain) == 0 &&
1664 		    cap->mds_wanted == want)
1665 			continue;     /* nope, all good */
1666 
1667 		if (is_delayed)
1668 			goto ack;
1669 
1670 		/* delay? */
1671 		if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1672 		    time_before(jiffies, ci->i_hold_caps_max)) {
1673 			dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1674 			     ceph_cap_string(cap->issued),
1675 			     ceph_cap_string(cap->issued & retain),
1676 			     ceph_cap_string(cap->mds_wanted),
1677 			     ceph_cap_string(want));
1678 			delayed++;
1679 			continue;
1680 		}
1681 
1682 ack:
1683 		if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1684 			dout(" skipping %p I_NOFLUSH set\n", inode);
1685 			continue;
1686 		}
1687 
1688 		if (session && session != cap->session) {
1689 			dout("oops, wrong session %p mutex\n", session);
1690 			mutex_unlock(&session->s_mutex);
1691 			session = NULL;
1692 		}
1693 		if (!session) {
1694 			session = cap->session;
1695 			if (mutex_trylock(&session->s_mutex) == 0) {
1696 				dout("inverting session/ino locks on %p\n",
1697 				     session);
1698 				spin_unlock(&ci->i_ceph_lock);
1699 				if (took_snap_rwsem) {
1700 					up_read(&mdsc->snap_rwsem);
1701 					took_snap_rwsem = 0;
1702 				}
1703 				mutex_lock(&session->s_mutex);
1704 				goto retry;
1705 			}
1706 		}
1707 		/* take snap_rwsem after session mutex */
1708 		if (!took_snap_rwsem) {
1709 			if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1710 				dout("inverting snap/in locks on %p\n",
1711 				     inode);
1712 				spin_unlock(&ci->i_ceph_lock);
1713 				down_read(&mdsc->snap_rwsem);
1714 				took_snap_rwsem = 1;
1715 				goto retry;
1716 			}
1717 			took_snap_rwsem = 1;
1718 		}
1719 
1720 		if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1721 			flushing = __mark_caps_flushing(inode, session);
1722 		else
1723 			flushing = 0;
1724 
1725 		mds = cap->mds;  /* remember mds, so we don't repeat */
1726 		sent++;
1727 
1728 		/* __send_cap drops i_ceph_lock */
1729 		delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1730 				      want, retain, flushing, NULL);
1731 		goto retry; /* retake i_ceph_lock and restart our cap scan. */
1732 	}
1733 
1734 	/*
1735 	 * Reschedule delayed caps release if we delayed anything,
1736 	 * otherwise cancel.
1737 	 */
1738 	if (delayed && is_delayed)
1739 		force_requeue = 1;   /* __send_cap delayed release; requeue */
1740 	if (!delayed && !is_delayed)
1741 		__cap_delay_cancel(mdsc, ci);
1742 	else if (!is_delayed || force_requeue)
1743 		__cap_delay_requeue(mdsc, ci);
1744 
1745 	spin_unlock(&ci->i_ceph_lock);
1746 
1747 	if (queue_invalidate)
1748 		ceph_queue_invalidate(inode);
1749 
1750 	if (session)
1751 		mutex_unlock(&session->s_mutex);
1752 	if (took_snap_rwsem)
1753 		up_read(&mdsc->snap_rwsem);
1754 }
1755 
1756 /*
1757  * Try to flush dirty caps back to the auth mds.
1758  */
1759 static int try_flush_caps(struct inode *inode, unsigned *flush_tid)
1760 {
1761 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1762 	struct ceph_inode_info *ci = ceph_inode(inode);
1763 	int flushing = 0;
1764 	struct ceph_mds_session *session = NULL;
1765 
1766 retry:
1767 	spin_lock(&ci->i_ceph_lock);
1768 	if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1769 		dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1770 		goto out;
1771 	}
1772 	if (ci->i_dirty_caps && ci->i_auth_cap) {
1773 		struct ceph_cap *cap = ci->i_auth_cap;
1774 		int used = __ceph_caps_used(ci);
1775 		int want = __ceph_caps_wanted(ci);
1776 		int delayed;
1777 
1778 		if (!session || session != cap->session) {
1779 			spin_unlock(&ci->i_ceph_lock);
1780 			if (session)
1781 				mutex_unlock(&session->s_mutex);
1782 			session = cap->session;
1783 			mutex_lock(&session->s_mutex);
1784 			goto retry;
1785 		}
1786 		if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1787 			goto out;
1788 
1789 		flushing = __mark_caps_flushing(inode, session);
1790 
1791 		/* __send_cap drops i_ceph_lock */
1792 		delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1793 				     cap->issued | cap->implemented, flushing,
1794 				     flush_tid);
1795 		if (!delayed)
1796 			goto out_unlocked;
1797 
1798 		spin_lock(&ci->i_ceph_lock);
1799 		__cap_delay_requeue(mdsc, ci);
1800 	}
1801 out:
1802 	spin_unlock(&ci->i_ceph_lock);
1803 out_unlocked:
1804 	if (session)
1805 		mutex_unlock(&session->s_mutex);
1806 	return flushing;
1807 }
1808 
1809 /*
1810  * Return true if we've flushed caps through the given flush_tid.
1811  */
1812 static int caps_are_flushed(struct inode *inode, unsigned tid)
1813 {
1814 	struct ceph_inode_info *ci = ceph_inode(inode);
1815 	int i, ret = 1;
1816 
1817 	spin_lock(&ci->i_ceph_lock);
1818 	for (i = 0; i < CEPH_CAP_BITS; i++)
1819 		if ((ci->i_flushing_caps & (1 << i)) &&
1820 		    ci->i_cap_flush_tid[i] <= tid) {
1821 			/* still flushing this bit */
1822 			ret = 0;
1823 			break;
1824 		}
1825 	spin_unlock(&ci->i_ceph_lock);
1826 	return ret;
1827 }
1828 
1829 /*
1830  * Wait on any unsafe replies for the given inode.  First wait on the
1831  * newest request, and make that the upper bound.  Then, if there are
1832  * more requests, keep waiting on the oldest as long as it is still older
1833  * than the original request.
1834  */
1835 static void sync_write_wait(struct inode *inode)
1836 {
1837 	struct ceph_inode_info *ci = ceph_inode(inode);
1838 	struct list_head *head = &ci->i_unsafe_writes;
1839 	struct ceph_osd_request *req;
1840 	u64 last_tid;
1841 
1842 	spin_lock(&ci->i_unsafe_lock);
1843 	if (list_empty(head))
1844 		goto out;
1845 
1846 	/* set upper bound as _last_ entry in chain */
1847 	req = list_entry(head->prev, struct ceph_osd_request,
1848 			 r_unsafe_item);
1849 	last_tid = req->r_tid;
1850 
1851 	do {
1852 		ceph_osdc_get_request(req);
1853 		spin_unlock(&ci->i_unsafe_lock);
1854 		dout("sync_write_wait on tid %llu (until %llu)\n",
1855 		     req->r_tid, last_tid);
1856 		wait_for_completion(&req->r_safe_completion);
1857 		spin_lock(&ci->i_unsafe_lock);
1858 		ceph_osdc_put_request(req);
1859 
1860 		/*
1861 		 * from here on look at first entry in chain, since we
1862 		 * only want to wait for anything older than last_tid
1863 		 */
1864 		if (list_empty(head))
1865 			break;
1866 		req = list_entry(head->next, struct ceph_osd_request,
1867 				 r_unsafe_item);
1868 	} while (req->r_tid < last_tid);
1869 out:
1870 	spin_unlock(&ci->i_unsafe_lock);
1871 }
1872 
1873 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1874 {
1875 	struct inode *inode = file->f_mapping->host;
1876 	struct ceph_inode_info *ci = ceph_inode(inode);
1877 	unsigned flush_tid;
1878 	int ret;
1879 	int dirty;
1880 
1881 	dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1882 	sync_write_wait(inode);
1883 
1884 	ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1885 	if (ret < 0)
1886 		return ret;
1887 	mutex_lock(&inode->i_mutex);
1888 
1889 	dirty = try_flush_caps(inode, &flush_tid);
1890 	dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1891 
1892 	/*
1893 	 * only wait on non-file metadata writeback (the mds
1894 	 * can recover size and mtime, so we don't need to
1895 	 * wait for that)
1896 	 */
1897 	if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1898 		dout("fsync waiting for flush_tid %u\n", flush_tid);
1899 		ret = wait_event_interruptible(ci->i_cap_wq,
1900 				       caps_are_flushed(inode, flush_tid));
1901 	}
1902 
1903 	dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1904 	mutex_unlock(&inode->i_mutex);
1905 	return ret;
1906 }
1907 
1908 /*
1909  * Flush any dirty caps back to the mds.  If we aren't asked to wait,
1910  * queue inode for flush but don't do so immediately, because we can
1911  * get by with fewer MDS messages if we wait for data writeback to
1912  * complete first.
1913  */
1914 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1915 {
1916 	struct ceph_inode_info *ci = ceph_inode(inode);
1917 	unsigned flush_tid;
1918 	int err = 0;
1919 	int dirty;
1920 	int wait = wbc->sync_mode == WB_SYNC_ALL;
1921 
1922 	dout("write_inode %p wait=%d\n", inode, wait);
1923 	if (wait) {
1924 		dirty = try_flush_caps(inode, &flush_tid);
1925 		if (dirty)
1926 			err = wait_event_interruptible(ci->i_cap_wq,
1927 				       caps_are_flushed(inode, flush_tid));
1928 	} else {
1929 		struct ceph_mds_client *mdsc =
1930 			ceph_sb_to_client(inode->i_sb)->mdsc;
1931 
1932 		spin_lock(&ci->i_ceph_lock);
1933 		if (__ceph_caps_dirty(ci))
1934 			__cap_delay_requeue_front(mdsc, ci);
1935 		spin_unlock(&ci->i_ceph_lock);
1936 	}
1937 	return err;
1938 }
1939 
1940 /*
1941  * After a recovering MDS goes active, we need to resend any caps
1942  * we were flushing.
1943  *
1944  * Caller holds session->s_mutex.
1945  */
1946 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1947 				   struct ceph_mds_session *session)
1948 {
1949 	struct ceph_cap_snap *capsnap;
1950 
1951 	dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1952 	list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1953 			    flushing_item) {
1954 		struct ceph_inode_info *ci = capsnap->ci;
1955 		struct inode *inode = &ci->vfs_inode;
1956 		struct ceph_cap *cap;
1957 
1958 		spin_lock(&ci->i_ceph_lock);
1959 		cap = ci->i_auth_cap;
1960 		if (cap && cap->session == session) {
1961 			dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1962 			     cap, capsnap);
1963 			__ceph_flush_snaps(ci, &session, 1);
1964 		} else {
1965 			pr_err("%p auth cap %p not mds%d ???\n", inode,
1966 			       cap, session->s_mds);
1967 		}
1968 		spin_unlock(&ci->i_ceph_lock);
1969 	}
1970 }
1971 
1972 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1973 			     struct ceph_mds_session *session)
1974 {
1975 	struct ceph_inode_info *ci;
1976 
1977 	kick_flushing_capsnaps(mdsc, session);
1978 
1979 	dout("kick_flushing_caps mds%d\n", session->s_mds);
1980 	list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1981 		struct inode *inode = &ci->vfs_inode;
1982 		struct ceph_cap *cap;
1983 		int delayed = 0;
1984 
1985 		spin_lock(&ci->i_ceph_lock);
1986 		cap = ci->i_auth_cap;
1987 		if (cap && cap->session == session) {
1988 			dout("kick_flushing_caps %p cap %p %s\n", inode,
1989 			     cap, ceph_cap_string(ci->i_flushing_caps));
1990 			delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1991 					     __ceph_caps_used(ci),
1992 					     __ceph_caps_wanted(ci),
1993 					     cap->issued | cap->implemented,
1994 					     ci->i_flushing_caps, NULL);
1995 			if (delayed) {
1996 				spin_lock(&ci->i_ceph_lock);
1997 				__cap_delay_requeue(mdsc, ci);
1998 				spin_unlock(&ci->i_ceph_lock);
1999 			}
2000 		} else {
2001 			pr_err("%p auth cap %p not mds%d ???\n", inode,
2002 			       cap, session->s_mds);
2003 			spin_unlock(&ci->i_ceph_lock);
2004 		}
2005 	}
2006 }
2007 
2008 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2009 				     struct ceph_mds_session *session,
2010 				     struct inode *inode)
2011 {
2012 	struct ceph_inode_info *ci = ceph_inode(inode);
2013 	struct ceph_cap *cap;
2014 	int delayed = 0;
2015 
2016 	spin_lock(&ci->i_ceph_lock);
2017 	cap = ci->i_auth_cap;
2018 	dout("kick_flushing_inode_caps %p flushing %s flush_seq %lld\n", inode,
2019 	     ceph_cap_string(ci->i_flushing_caps), ci->i_cap_flush_seq);
2020 
2021 	__ceph_flush_snaps(ci, &session, 1);
2022 
2023 	if (ci->i_flushing_caps) {
2024 		spin_lock(&mdsc->cap_dirty_lock);
2025 		list_move_tail(&ci->i_flushing_item,
2026 			       &cap->session->s_cap_flushing);
2027 		spin_unlock(&mdsc->cap_dirty_lock);
2028 
2029 		delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2030 				     __ceph_caps_used(ci),
2031 				     __ceph_caps_wanted(ci),
2032 				     cap->issued | cap->implemented,
2033 				     ci->i_flushing_caps, NULL);
2034 		if (delayed) {
2035 			spin_lock(&ci->i_ceph_lock);
2036 			__cap_delay_requeue(mdsc, ci);
2037 			spin_unlock(&ci->i_ceph_lock);
2038 		}
2039 	} else {
2040 		spin_unlock(&ci->i_ceph_lock);
2041 	}
2042 }
2043 
2044 
2045 /*
2046  * Take references to capabilities we hold, so that we don't release
2047  * them to the MDS prematurely.
2048  *
2049  * Protected by i_ceph_lock.
2050  */
2051 static void __take_cap_refs(struct ceph_inode_info *ci, int got)
2052 {
2053 	if (got & CEPH_CAP_PIN)
2054 		ci->i_pin_ref++;
2055 	if (got & CEPH_CAP_FILE_RD)
2056 		ci->i_rd_ref++;
2057 	if (got & CEPH_CAP_FILE_CACHE)
2058 		ci->i_rdcache_ref++;
2059 	if (got & CEPH_CAP_FILE_WR)
2060 		ci->i_wr_ref++;
2061 	if (got & CEPH_CAP_FILE_BUFFER) {
2062 		if (ci->i_wb_ref == 0)
2063 			ihold(&ci->vfs_inode);
2064 		ci->i_wb_ref++;
2065 		dout("__take_cap_refs %p wb %d -> %d (?)\n",
2066 		     &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2067 	}
2068 }
2069 
2070 /*
2071  * Try to grab cap references.  Specify those refs we @want, and the
2072  * minimal set we @need.  Also include the larger offset we are writing
2073  * to (when applicable), and check against max_size here as well.
2074  * Note that caller is responsible for ensuring max_size increases are
2075  * requested from the MDS.
2076  */
2077 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2078 			    int *got, loff_t endoff, int *check_max, int *err)
2079 {
2080 	struct inode *inode = &ci->vfs_inode;
2081 	int ret = 0;
2082 	int have, implemented;
2083 	int file_wanted;
2084 
2085 	dout("get_cap_refs %p need %s want %s\n", inode,
2086 	     ceph_cap_string(need), ceph_cap_string(want));
2087 	spin_lock(&ci->i_ceph_lock);
2088 
2089 	/* make sure file is actually open */
2090 	file_wanted = __ceph_caps_file_wanted(ci);
2091 	if ((file_wanted & need) == 0) {
2092 		dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2093 		     ceph_cap_string(need), ceph_cap_string(file_wanted));
2094 		*err = -EBADF;
2095 		ret = 1;
2096 		goto out;
2097 	}
2098 
2099 	/* finish pending truncate */
2100 	while (ci->i_truncate_pending) {
2101 		spin_unlock(&ci->i_ceph_lock);
2102 		__ceph_do_pending_vmtruncate(inode);
2103 		spin_lock(&ci->i_ceph_lock);
2104 	}
2105 
2106 	have = __ceph_caps_issued(ci, &implemented);
2107 
2108 	if (have & need & CEPH_CAP_FILE_WR) {
2109 		if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2110 			dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2111 			     inode, endoff, ci->i_max_size);
2112 			if (endoff > ci->i_requested_max_size) {
2113 				*check_max = 1;
2114 				ret = 1;
2115 			}
2116 			goto out;
2117 		}
2118 		/*
2119 		 * If a sync write is in progress, we must wait, so that we
2120 		 * can get a final snapshot value for size+mtime.
2121 		 */
2122 		if (__ceph_have_pending_cap_snap(ci)) {
2123 			dout("get_cap_refs %p cap_snap_pending\n", inode);
2124 			goto out;
2125 		}
2126 	}
2127 
2128 	if ((have & need) == need) {
2129 		/*
2130 		 * Look at (implemented & ~have & not) so that we keep waiting
2131 		 * on transition from wanted -> needed caps.  This is needed
2132 		 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2133 		 * going before a prior buffered writeback happens.
2134 		 */
2135 		int not = want & ~(have & need);
2136 		int revoking = implemented & ~have;
2137 		dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2138 		     inode, ceph_cap_string(have), ceph_cap_string(not),
2139 		     ceph_cap_string(revoking));
2140 		if ((revoking & not) == 0) {
2141 			*got = need | (have & want);
2142 			__take_cap_refs(ci, *got);
2143 			ret = 1;
2144 		}
2145 	} else {
2146 		dout("get_cap_refs %p have %s needed %s\n", inode,
2147 		     ceph_cap_string(have), ceph_cap_string(need));
2148 	}
2149 out:
2150 	spin_unlock(&ci->i_ceph_lock);
2151 	dout("get_cap_refs %p ret %d got %s\n", inode,
2152 	     ret, ceph_cap_string(*got));
2153 	return ret;
2154 }
2155 
2156 /*
2157  * Check the offset we are writing up to against our current
2158  * max_size.  If necessary, tell the MDS we want to write to
2159  * a larger offset.
2160  */
2161 static void check_max_size(struct inode *inode, loff_t endoff)
2162 {
2163 	struct ceph_inode_info *ci = ceph_inode(inode);
2164 	int check = 0;
2165 
2166 	/* do we need to explicitly request a larger max_size? */
2167 	spin_lock(&ci->i_ceph_lock);
2168 	if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2169 		dout("write %p at large endoff %llu, req max_size\n",
2170 		     inode, endoff);
2171 		ci->i_wanted_max_size = endoff;
2172 	}
2173 	/* duplicate ceph_check_caps()'s logic */
2174 	if (ci->i_auth_cap &&
2175 	    (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2176 	    ci->i_wanted_max_size > ci->i_max_size &&
2177 	    ci->i_wanted_max_size > ci->i_requested_max_size)
2178 		check = 1;
2179 	spin_unlock(&ci->i_ceph_lock);
2180 	if (check)
2181 		ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2182 }
2183 
2184 /*
2185  * Wait for caps, and take cap references.  If we can't get a WR cap
2186  * due to a small max_size, make sure we check_max_size (and possibly
2187  * ask the mds) so we don't get hung up indefinitely.
2188  */
2189 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got,
2190 		  loff_t endoff)
2191 {
2192 	int check_max, ret, err;
2193 
2194 retry:
2195 	if (endoff > 0)
2196 		check_max_size(&ci->vfs_inode, endoff);
2197 	check_max = 0;
2198 	err = 0;
2199 	ret = wait_event_interruptible(ci->i_cap_wq,
2200 				       try_get_cap_refs(ci, need, want,
2201 							got, endoff,
2202 							&check_max, &err));
2203 	if (err)
2204 		ret = err;
2205 	if (check_max)
2206 		goto retry;
2207 	return ret;
2208 }
2209 
2210 /*
2211  * Take cap refs.  Caller must already know we hold at least one ref
2212  * on the caps in question or we don't know this is safe.
2213  */
2214 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2215 {
2216 	spin_lock(&ci->i_ceph_lock);
2217 	__take_cap_refs(ci, caps);
2218 	spin_unlock(&ci->i_ceph_lock);
2219 }
2220 
2221 /*
2222  * Release cap refs.
2223  *
2224  * If we released the last ref on any given cap, call ceph_check_caps
2225  * to release (or schedule a release).
2226  *
2227  * If we are releasing a WR cap (from a sync write), finalize any affected
2228  * cap_snap, and wake up any waiters.
2229  */
2230 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2231 {
2232 	struct inode *inode = &ci->vfs_inode;
2233 	int last = 0, put = 0, flushsnaps = 0, wake = 0;
2234 	struct ceph_cap_snap *capsnap;
2235 
2236 	spin_lock(&ci->i_ceph_lock);
2237 	if (had & CEPH_CAP_PIN)
2238 		--ci->i_pin_ref;
2239 	if (had & CEPH_CAP_FILE_RD)
2240 		if (--ci->i_rd_ref == 0)
2241 			last++;
2242 	if (had & CEPH_CAP_FILE_CACHE)
2243 		if (--ci->i_rdcache_ref == 0)
2244 			last++;
2245 	if (had & CEPH_CAP_FILE_BUFFER) {
2246 		if (--ci->i_wb_ref == 0) {
2247 			last++;
2248 			put++;
2249 		}
2250 		dout("put_cap_refs %p wb %d -> %d (?)\n",
2251 		     inode, ci->i_wb_ref+1, ci->i_wb_ref);
2252 	}
2253 	if (had & CEPH_CAP_FILE_WR)
2254 		if (--ci->i_wr_ref == 0) {
2255 			last++;
2256 			if (!list_empty(&ci->i_cap_snaps)) {
2257 				capsnap = list_first_entry(&ci->i_cap_snaps,
2258 						     struct ceph_cap_snap,
2259 						     ci_item);
2260 				if (capsnap->writing) {
2261 					capsnap->writing = 0;
2262 					flushsnaps =
2263 						__ceph_finish_cap_snap(ci,
2264 								       capsnap);
2265 					wake = 1;
2266 				}
2267 			}
2268 		}
2269 	spin_unlock(&ci->i_ceph_lock);
2270 
2271 	dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2272 	     last ? " last" : "", put ? " put" : "");
2273 
2274 	if (last && !flushsnaps)
2275 		ceph_check_caps(ci, 0, NULL);
2276 	else if (flushsnaps)
2277 		ceph_flush_snaps(ci);
2278 	if (wake)
2279 		wake_up_all(&ci->i_cap_wq);
2280 	if (put)
2281 		iput(inode);
2282 }
2283 
2284 /*
2285  * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2286  * context.  Adjust per-snap dirty page accounting as appropriate.
2287  * Once all dirty data for a cap_snap is flushed, flush snapped file
2288  * metadata back to the MDS.  If we dropped the last ref, call
2289  * ceph_check_caps.
2290  */
2291 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2292 				struct ceph_snap_context *snapc)
2293 {
2294 	struct inode *inode = &ci->vfs_inode;
2295 	int last = 0;
2296 	int complete_capsnap = 0;
2297 	int drop_capsnap = 0;
2298 	int found = 0;
2299 	struct ceph_cap_snap *capsnap = NULL;
2300 
2301 	spin_lock(&ci->i_ceph_lock);
2302 	ci->i_wrbuffer_ref -= nr;
2303 	last = !ci->i_wrbuffer_ref;
2304 
2305 	if (ci->i_head_snapc == snapc) {
2306 		ci->i_wrbuffer_ref_head -= nr;
2307 		if (ci->i_wrbuffer_ref_head == 0 &&
2308 		    ci->i_dirty_caps == 0 && ci->i_flushing_caps == 0) {
2309 			BUG_ON(!ci->i_head_snapc);
2310 			ceph_put_snap_context(ci->i_head_snapc);
2311 			ci->i_head_snapc = NULL;
2312 		}
2313 		dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2314 		     inode,
2315 		     ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2316 		     ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2317 		     last ? " LAST" : "");
2318 	} else {
2319 		list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2320 			if (capsnap->context == snapc) {
2321 				found = 1;
2322 				break;
2323 			}
2324 		}
2325 		BUG_ON(!found);
2326 		capsnap->dirty_pages -= nr;
2327 		if (capsnap->dirty_pages == 0) {
2328 			complete_capsnap = 1;
2329 			if (capsnap->dirty == 0)
2330 				/* cap writeback completed before we created
2331 				 * the cap_snap; no FLUSHSNAP is needed */
2332 				drop_capsnap = 1;
2333 		}
2334 		dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2335 		     " snap %lld %d/%d -> %d/%d %s%s%s\n",
2336 		     inode, capsnap, capsnap->context->seq,
2337 		     ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2338 		     ci->i_wrbuffer_ref, capsnap->dirty_pages,
2339 		     last ? " (wrbuffer last)" : "",
2340 		     complete_capsnap ? " (complete capsnap)" : "",
2341 		     drop_capsnap ? " (drop capsnap)" : "");
2342 		if (drop_capsnap) {
2343 			ceph_put_snap_context(capsnap->context);
2344 			list_del(&capsnap->ci_item);
2345 			list_del(&capsnap->flushing_item);
2346 			ceph_put_cap_snap(capsnap);
2347 		}
2348 	}
2349 
2350 	spin_unlock(&ci->i_ceph_lock);
2351 
2352 	if (last) {
2353 		ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2354 		iput(inode);
2355 	} else if (complete_capsnap) {
2356 		ceph_flush_snaps(ci);
2357 		wake_up_all(&ci->i_cap_wq);
2358 	}
2359 	if (drop_capsnap)
2360 		iput(inode);
2361 }
2362 
2363 /*
2364  * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2365  */
2366 static void invalidate_aliases(struct inode *inode)
2367 {
2368 	struct dentry *dn, *prev = NULL;
2369 
2370 	dout("invalidate_aliases inode %p\n", inode);
2371 	d_prune_aliases(inode);
2372 	/*
2373 	 * For non-directory inode, d_find_alias() only returns
2374 	 * hashed dentry. After calling d_invalidate(), the
2375 	 * dentry becomes unhashed.
2376 	 *
2377 	 * For directory inode, d_find_alias() can return
2378 	 * unhashed dentry. But directory inode should have
2379 	 * one alias at most.
2380 	 */
2381 	while ((dn = d_find_alias(inode))) {
2382 		if (dn == prev) {
2383 			dput(dn);
2384 			break;
2385 		}
2386 		d_invalidate(dn);
2387 		if (prev)
2388 			dput(prev);
2389 		prev = dn;
2390 	}
2391 	if (prev)
2392 		dput(prev);
2393 }
2394 
2395 /*
2396  * Handle a cap GRANT message from the MDS.  (Note that a GRANT may
2397  * actually be a revocation if it specifies a smaller cap set.)
2398  *
2399  * caller holds s_mutex and i_ceph_lock, we drop both.
2400  *
2401  * return value:
2402  *  0 - ok
2403  *  1 - check_caps on auth cap only (writeback)
2404  *  2 - check_caps (ack revoke)
2405  */
2406 static void handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant,
2407 			     struct ceph_mds_session *session,
2408 			     struct ceph_cap *cap,
2409 			     struct ceph_buffer *xattr_buf)
2410 		__releases(ci->i_ceph_lock)
2411 {
2412 	struct ceph_inode_info *ci = ceph_inode(inode);
2413 	int mds = session->s_mds;
2414 	int seq = le32_to_cpu(grant->seq);
2415 	int newcaps = le32_to_cpu(grant->caps);
2416 	int issued, implemented, used, wanted, dirty;
2417 	u64 size = le64_to_cpu(grant->size);
2418 	u64 max_size = le64_to_cpu(grant->max_size);
2419 	struct timespec mtime, atime, ctime;
2420 	int check_caps = 0;
2421 	int wake = 0;
2422 	int writeback = 0;
2423 	int queue_invalidate = 0;
2424 	int deleted_inode = 0;
2425 	int queue_revalidate = 0;
2426 
2427 	dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2428 	     inode, cap, mds, seq, ceph_cap_string(newcaps));
2429 	dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2430 		inode->i_size);
2431 
2432 
2433 	/*
2434 	 * auth mds of the inode changed. we received the cap export message,
2435 	 * but still haven't received the cap import message. handle_cap_export
2436 	 * updated the new auth MDS' cap.
2437 	 *
2438 	 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2439 	 * that was sent before the cap import message. So don't remove caps.
2440 	 */
2441 	if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2442 		WARN_ON(cap != ci->i_auth_cap);
2443 		WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2444 		seq = cap->seq;
2445 		newcaps |= cap->issued;
2446 	}
2447 
2448 	/*
2449 	 * If CACHE is being revoked, and we have no dirty buffers,
2450 	 * try to invalidate (once).  (If there are dirty buffers, we
2451 	 * will invalidate _after_ writeback.)
2452 	 */
2453 	if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2454 	    (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2455 	    !ci->i_wrbuffer_ref) {
2456 		if (try_nonblocking_invalidate(inode)) {
2457 			/* there were locked pages.. invalidate later
2458 			   in a separate thread. */
2459 			if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2460 				queue_invalidate = 1;
2461 				ci->i_rdcache_revoking = ci->i_rdcache_gen;
2462 			}
2463 		}
2464 
2465 		ceph_fscache_invalidate(inode);
2466 	}
2467 
2468 	/* side effects now are allowed */
2469 
2470 	issued = __ceph_caps_issued(ci, &implemented);
2471 	issued |= implemented | __ceph_caps_dirty(ci);
2472 
2473 	cap->cap_gen = session->s_cap_gen;
2474 	cap->seq = seq;
2475 
2476 	__check_cap_issue(ci, cap, newcaps);
2477 
2478 	if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
2479 		inode->i_mode = le32_to_cpu(grant->mode);
2480 		inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2481 		inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2482 		dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2483 		     from_kuid(&init_user_ns, inode->i_uid),
2484 		     from_kgid(&init_user_ns, inode->i_gid));
2485 	}
2486 
2487 	if ((issued & CEPH_CAP_LINK_EXCL) == 0) {
2488 		set_nlink(inode, le32_to_cpu(grant->nlink));
2489 		if (inode->i_nlink == 0 &&
2490 		    (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2491 			deleted_inode = 1;
2492 	}
2493 
2494 	if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2495 		int len = le32_to_cpu(grant->xattr_len);
2496 		u64 version = le64_to_cpu(grant->xattr_version);
2497 
2498 		if (version > ci->i_xattrs.version) {
2499 			dout(" got new xattrs v%llu on %p len %d\n",
2500 			     version, inode, len);
2501 			if (ci->i_xattrs.blob)
2502 				ceph_buffer_put(ci->i_xattrs.blob);
2503 			ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2504 			ci->i_xattrs.version = version;
2505 			ceph_forget_all_cached_acls(inode);
2506 		}
2507 	}
2508 
2509 	/* Do we need to revalidate our fscache cookie. Don't bother on the
2510 	 * first cache cap as we already validate at cookie creation time. */
2511 	if ((issued & CEPH_CAP_FILE_CACHE) && ci->i_rdcache_gen > 1)
2512 		queue_revalidate = 1;
2513 
2514 	/* size/ctime/mtime/atime? */
2515 	ceph_fill_file_size(inode, issued,
2516 			    le32_to_cpu(grant->truncate_seq),
2517 			    le64_to_cpu(grant->truncate_size), size);
2518 	ceph_decode_timespec(&mtime, &grant->mtime);
2519 	ceph_decode_timespec(&atime, &grant->atime);
2520 	ceph_decode_timespec(&ctime, &grant->ctime);
2521 	ceph_fill_file_time(inode, issued,
2522 			    le32_to_cpu(grant->time_warp_seq), &ctime, &mtime,
2523 			    &atime);
2524 
2525 
2526 	/* file layout may have changed */
2527 	ci->i_layout = grant->layout;
2528 
2529 	/* max size increase? */
2530 	if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
2531 		dout("max_size %lld -> %llu\n", ci->i_max_size, max_size);
2532 		ci->i_max_size = max_size;
2533 		if (max_size >= ci->i_wanted_max_size) {
2534 			ci->i_wanted_max_size = 0;  /* reset */
2535 			ci->i_requested_max_size = 0;
2536 		}
2537 		wake = 1;
2538 	}
2539 
2540 	/* check cap bits */
2541 	wanted = __ceph_caps_wanted(ci);
2542 	used = __ceph_caps_used(ci);
2543 	dirty = __ceph_caps_dirty(ci);
2544 	dout(" my wanted = %s, used = %s, dirty %s\n",
2545 	     ceph_cap_string(wanted),
2546 	     ceph_cap_string(used),
2547 	     ceph_cap_string(dirty));
2548 	if (wanted != le32_to_cpu(grant->wanted)) {
2549 		dout("mds wanted %s -> %s\n",
2550 		     ceph_cap_string(le32_to_cpu(grant->wanted)),
2551 		     ceph_cap_string(wanted));
2552 		/* imported cap may not have correct mds_wanted */
2553 		if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
2554 			check_caps = 1;
2555 	}
2556 
2557 	/* revocation, grant, or no-op? */
2558 	if (cap->issued & ~newcaps) {
2559 		int revoking = cap->issued & ~newcaps;
2560 
2561 		dout("revocation: %s -> %s (revoking %s)\n",
2562 		     ceph_cap_string(cap->issued),
2563 		     ceph_cap_string(newcaps),
2564 		     ceph_cap_string(revoking));
2565 		if (revoking & used & CEPH_CAP_FILE_BUFFER)
2566 			writeback = 1;  /* initiate writeback; will delay ack */
2567 		else if (revoking == CEPH_CAP_FILE_CACHE &&
2568 			 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2569 			 queue_invalidate)
2570 			; /* do nothing yet, invalidation will be queued */
2571 		else if (cap == ci->i_auth_cap)
2572 			check_caps = 1; /* check auth cap only */
2573 		else
2574 			check_caps = 2; /* check all caps */
2575 		cap->issued = newcaps;
2576 		cap->implemented |= newcaps;
2577 	} else if (cap->issued == newcaps) {
2578 		dout("caps unchanged: %s -> %s\n",
2579 		     ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2580 	} else {
2581 		dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2582 		     ceph_cap_string(newcaps));
2583 		/* non-auth MDS is revoking the newly grant caps ? */
2584 		if (cap == ci->i_auth_cap &&
2585 		    __ceph_caps_revoking_other(ci, cap, newcaps))
2586 		    check_caps = 2;
2587 
2588 		cap->issued = newcaps;
2589 		cap->implemented |= newcaps; /* add bits only, to
2590 					      * avoid stepping on a
2591 					      * pending revocation */
2592 		wake = 1;
2593 	}
2594 	BUG_ON(cap->issued & ~cap->implemented);
2595 
2596 	spin_unlock(&ci->i_ceph_lock);
2597 
2598 	if (writeback)
2599 		/*
2600 		 * queue inode for writeback: we can't actually call
2601 		 * filemap_write_and_wait, etc. from message handler
2602 		 * context.
2603 		 */
2604 		ceph_queue_writeback(inode);
2605 	if (queue_invalidate)
2606 		ceph_queue_invalidate(inode);
2607 	if (deleted_inode)
2608 		invalidate_aliases(inode);
2609 	if (queue_revalidate)
2610 		ceph_queue_revalidate(inode);
2611 	if (wake)
2612 		wake_up_all(&ci->i_cap_wq);
2613 
2614 	if (check_caps == 1)
2615 		ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2616 				session);
2617 	else if (check_caps == 2)
2618 		ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2619 	else
2620 		mutex_unlock(&session->s_mutex);
2621 }
2622 
2623 /*
2624  * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2625  * MDS has been safely committed.
2626  */
2627 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2628 				 struct ceph_mds_caps *m,
2629 				 struct ceph_mds_session *session,
2630 				 struct ceph_cap *cap)
2631 	__releases(ci->i_ceph_lock)
2632 {
2633 	struct ceph_inode_info *ci = ceph_inode(inode);
2634 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2635 	unsigned seq = le32_to_cpu(m->seq);
2636 	int dirty = le32_to_cpu(m->dirty);
2637 	int cleaned = 0;
2638 	int drop = 0;
2639 	int i;
2640 
2641 	for (i = 0; i < CEPH_CAP_BITS; i++)
2642 		if ((dirty & (1 << i)) &&
2643 		    flush_tid == ci->i_cap_flush_tid[i])
2644 			cleaned |= 1 << i;
2645 
2646 	dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2647 	     " flushing %s -> %s\n",
2648 	     inode, session->s_mds, seq, ceph_cap_string(dirty),
2649 	     ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2650 	     ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2651 
2652 	if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2653 		goto out;
2654 
2655 	ci->i_flushing_caps &= ~cleaned;
2656 
2657 	spin_lock(&mdsc->cap_dirty_lock);
2658 	if (ci->i_flushing_caps == 0) {
2659 		list_del_init(&ci->i_flushing_item);
2660 		if (!list_empty(&session->s_cap_flushing))
2661 			dout(" mds%d still flushing cap on %p\n",
2662 			     session->s_mds,
2663 			     &list_entry(session->s_cap_flushing.next,
2664 					 struct ceph_inode_info,
2665 					 i_flushing_item)->vfs_inode);
2666 		mdsc->num_cap_flushing--;
2667 		wake_up_all(&mdsc->cap_flushing_wq);
2668 		dout(" inode %p now !flushing\n", inode);
2669 
2670 		if (ci->i_dirty_caps == 0) {
2671 			dout(" inode %p now clean\n", inode);
2672 			BUG_ON(!list_empty(&ci->i_dirty_item));
2673 			drop = 1;
2674 			if (ci->i_wrbuffer_ref_head == 0) {
2675 				BUG_ON(!ci->i_head_snapc);
2676 				ceph_put_snap_context(ci->i_head_snapc);
2677 				ci->i_head_snapc = NULL;
2678 			}
2679 		} else {
2680 			BUG_ON(list_empty(&ci->i_dirty_item));
2681 		}
2682 	}
2683 	spin_unlock(&mdsc->cap_dirty_lock);
2684 	wake_up_all(&ci->i_cap_wq);
2685 
2686 out:
2687 	spin_unlock(&ci->i_ceph_lock);
2688 	if (drop)
2689 		iput(inode);
2690 }
2691 
2692 /*
2693  * Handle FLUSHSNAP_ACK.  MDS has flushed snap data to disk and we can
2694  * throw away our cap_snap.
2695  *
2696  * Caller hold s_mutex.
2697  */
2698 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2699 				     struct ceph_mds_caps *m,
2700 				     struct ceph_mds_session *session)
2701 {
2702 	struct ceph_inode_info *ci = ceph_inode(inode);
2703 	u64 follows = le64_to_cpu(m->snap_follows);
2704 	struct ceph_cap_snap *capsnap;
2705 	int drop = 0;
2706 
2707 	dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2708 	     inode, ci, session->s_mds, follows);
2709 
2710 	spin_lock(&ci->i_ceph_lock);
2711 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2712 		if (capsnap->follows == follows) {
2713 			if (capsnap->flush_tid != flush_tid) {
2714 				dout(" cap_snap %p follows %lld tid %lld !="
2715 				     " %lld\n", capsnap, follows,
2716 				     flush_tid, capsnap->flush_tid);
2717 				break;
2718 			}
2719 			WARN_ON(capsnap->dirty_pages || capsnap->writing);
2720 			dout(" removing %p cap_snap %p follows %lld\n",
2721 			     inode, capsnap, follows);
2722 			ceph_put_snap_context(capsnap->context);
2723 			list_del(&capsnap->ci_item);
2724 			list_del(&capsnap->flushing_item);
2725 			ceph_put_cap_snap(capsnap);
2726 			drop = 1;
2727 			break;
2728 		} else {
2729 			dout(" skipping cap_snap %p follows %lld\n",
2730 			     capsnap, capsnap->follows);
2731 		}
2732 	}
2733 	spin_unlock(&ci->i_ceph_lock);
2734 	if (drop)
2735 		iput(inode);
2736 }
2737 
2738 /*
2739  * Handle TRUNC from MDS, indicating file truncation.
2740  *
2741  * caller hold s_mutex.
2742  */
2743 static void handle_cap_trunc(struct inode *inode,
2744 			     struct ceph_mds_caps *trunc,
2745 			     struct ceph_mds_session *session)
2746 	__releases(ci->i_ceph_lock)
2747 {
2748 	struct ceph_inode_info *ci = ceph_inode(inode);
2749 	int mds = session->s_mds;
2750 	int seq = le32_to_cpu(trunc->seq);
2751 	u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2752 	u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2753 	u64 size = le64_to_cpu(trunc->size);
2754 	int implemented = 0;
2755 	int dirty = __ceph_caps_dirty(ci);
2756 	int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2757 	int queue_trunc = 0;
2758 
2759 	issued |= implemented | dirty;
2760 
2761 	dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2762 	     inode, mds, seq, truncate_size, truncate_seq);
2763 	queue_trunc = ceph_fill_file_size(inode, issued,
2764 					  truncate_seq, truncate_size, size);
2765 	spin_unlock(&ci->i_ceph_lock);
2766 
2767 	if (queue_trunc) {
2768 		ceph_queue_vmtruncate(inode);
2769 		ceph_fscache_invalidate(inode);
2770 	}
2771 }
2772 
2773 /*
2774  * Handle EXPORT from MDS.  Cap is being migrated _from_ this mds to a
2775  * different one.  If we are the most recent migration we've seen (as
2776  * indicated by mseq), make note of the migrating cap bits for the
2777  * duration (until we see the corresponding IMPORT).
2778  *
2779  * caller holds s_mutex
2780  */
2781 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2782 			      struct ceph_mds_cap_peer *ph,
2783 			      struct ceph_mds_session *session)
2784 {
2785 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2786 	struct ceph_mds_session *tsession = NULL;
2787 	struct ceph_cap *cap, *tcap;
2788 	struct ceph_inode_info *ci = ceph_inode(inode);
2789 	u64 t_cap_id;
2790 	unsigned mseq = le32_to_cpu(ex->migrate_seq);
2791 	unsigned t_seq, t_mseq;
2792 	int target, issued;
2793 	int mds = session->s_mds;
2794 
2795 	if (ph) {
2796 		t_cap_id = le64_to_cpu(ph->cap_id);
2797 		t_seq = le32_to_cpu(ph->seq);
2798 		t_mseq = le32_to_cpu(ph->mseq);
2799 		target = le32_to_cpu(ph->mds);
2800 	} else {
2801 		t_cap_id = t_seq = t_mseq = 0;
2802 		target = -1;
2803 	}
2804 
2805 	dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
2806 	     inode, ci, mds, mseq, target);
2807 retry:
2808 	spin_lock(&ci->i_ceph_lock);
2809 	cap = __get_cap_for_mds(ci, mds);
2810 	if (!cap)
2811 		goto out_unlock;
2812 
2813 	if (target < 0) {
2814 		__ceph_remove_cap(cap, false);
2815 		goto out_unlock;
2816 	}
2817 
2818 	/*
2819 	 * now we know we haven't received the cap import message yet
2820 	 * because the exported cap still exist.
2821 	 */
2822 
2823 	issued = cap->issued;
2824 	WARN_ON(issued != cap->implemented);
2825 
2826 	tcap = __get_cap_for_mds(ci, target);
2827 	if (tcap) {
2828 		/* already have caps from the target */
2829 		if (tcap->cap_id != t_cap_id ||
2830 		    ceph_seq_cmp(tcap->seq, t_seq) < 0) {
2831 			dout(" updating import cap %p mds%d\n", tcap, target);
2832 			tcap->cap_id = t_cap_id;
2833 			tcap->seq = t_seq - 1;
2834 			tcap->issue_seq = t_seq - 1;
2835 			tcap->mseq = t_mseq;
2836 			tcap->issued |= issued;
2837 			tcap->implemented |= issued;
2838 			if (cap == ci->i_auth_cap)
2839 				ci->i_auth_cap = tcap;
2840 			if (ci->i_flushing_caps && ci->i_auth_cap == tcap) {
2841 				spin_lock(&mdsc->cap_dirty_lock);
2842 				list_move_tail(&ci->i_flushing_item,
2843 					       &tcap->session->s_cap_flushing);
2844 				spin_unlock(&mdsc->cap_dirty_lock);
2845 			}
2846 		}
2847 		__ceph_remove_cap(cap, false);
2848 		goto out_unlock;
2849 	}
2850 
2851 	if (tsession) {
2852 		int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
2853 		spin_unlock(&ci->i_ceph_lock);
2854 		/* add placeholder for the export tagert */
2855 		ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
2856 			     t_seq - 1, t_mseq, (u64)-1, flag, NULL);
2857 		goto retry;
2858 	}
2859 
2860 	spin_unlock(&ci->i_ceph_lock);
2861 	mutex_unlock(&session->s_mutex);
2862 
2863 	/* open target session */
2864 	tsession = ceph_mdsc_open_export_target_session(mdsc, target);
2865 	if (!IS_ERR(tsession)) {
2866 		if (mds > target) {
2867 			mutex_lock(&session->s_mutex);
2868 			mutex_lock_nested(&tsession->s_mutex,
2869 					  SINGLE_DEPTH_NESTING);
2870 		} else {
2871 			mutex_lock(&tsession->s_mutex);
2872 			mutex_lock_nested(&session->s_mutex,
2873 					  SINGLE_DEPTH_NESTING);
2874 		}
2875 		ceph_add_cap_releases(mdsc, tsession);
2876 	} else {
2877 		WARN_ON(1);
2878 		tsession = NULL;
2879 		target = -1;
2880 	}
2881 	goto retry;
2882 
2883 out_unlock:
2884 	spin_unlock(&ci->i_ceph_lock);
2885 	mutex_unlock(&session->s_mutex);
2886 	if (tsession) {
2887 		mutex_unlock(&tsession->s_mutex);
2888 		ceph_put_mds_session(tsession);
2889 	}
2890 }
2891 
2892 /*
2893  * Handle cap IMPORT.  If there are temp bits from an older EXPORT,
2894  * clean them up.
2895  *
2896  * caller holds s_mutex.
2897  */
2898 static void handle_cap_import(struct ceph_mds_client *mdsc,
2899 			      struct inode *inode, struct ceph_mds_caps *im,
2900 			      struct ceph_mds_cap_peer *ph,
2901 			      struct ceph_mds_session *session,
2902 			      void *snaptrace, int snaptrace_len)
2903 {
2904 	struct ceph_inode_info *ci = ceph_inode(inode);
2905 	struct ceph_cap *cap;
2906 	int mds = session->s_mds;
2907 	unsigned issued = le32_to_cpu(im->caps);
2908 	unsigned wanted = le32_to_cpu(im->wanted);
2909 	unsigned seq = le32_to_cpu(im->seq);
2910 	unsigned mseq = le32_to_cpu(im->migrate_seq);
2911 	u64 realmino = le64_to_cpu(im->realm);
2912 	u64 cap_id = le64_to_cpu(im->cap_id);
2913 	u64 p_cap_id;
2914 	int peer;
2915 
2916 	if (ph) {
2917 		p_cap_id = le64_to_cpu(ph->cap_id);
2918 		peer = le32_to_cpu(ph->mds);
2919 	} else {
2920 		p_cap_id = 0;
2921 		peer = -1;
2922 	}
2923 
2924 	dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
2925 	     inode, ci, mds, mseq, peer);
2926 
2927 	spin_lock(&ci->i_ceph_lock);
2928 	cap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
2929 	if (cap && cap->cap_id == p_cap_id) {
2930 		dout(" remove export cap %p mds%d flags %d\n",
2931 		     cap, peer, ph->flags);
2932 		if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
2933 		    (cap->seq != le32_to_cpu(ph->seq) ||
2934 		     cap->mseq != le32_to_cpu(ph->mseq))) {
2935 			pr_err("handle_cap_import: mismatched seq/mseq: "
2936 			       "ino (%llx.%llx) mds%d seq %d mseq %d "
2937 			       "importer mds%d has peer seq %d mseq %d\n",
2938 			       ceph_vinop(inode), peer, cap->seq,
2939 			       cap->mseq, mds, le32_to_cpu(ph->seq),
2940 			       le32_to_cpu(ph->mseq));
2941 		}
2942 		ci->i_cap_exporting_issued = cap->issued;
2943 		__ceph_remove_cap(cap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
2944 	}
2945 
2946 	/* make sure we re-request max_size, if necessary */
2947 	ci->i_wanted_max_size = 0;
2948 	ci->i_requested_max_size = 0;
2949 	spin_unlock(&ci->i_ceph_lock);
2950 
2951 	down_write(&mdsc->snap_rwsem);
2952 	ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len,
2953 			       false);
2954 	downgrade_write(&mdsc->snap_rwsem);
2955 	ceph_add_cap(inode, session, cap_id, -1,
2956 		     issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH,
2957 		     NULL /* no caps context */);
2958 	kick_flushing_inode_caps(mdsc, session, inode);
2959 	up_read(&mdsc->snap_rwsem);
2960 
2961 }
2962 
2963 /*
2964  * Handle a caps message from the MDS.
2965  *
2966  * Identify the appropriate session, inode, and call the right handler
2967  * based on the cap op.
2968  */
2969 void ceph_handle_caps(struct ceph_mds_session *session,
2970 		      struct ceph_msg *msg)
2971 {
2972 	struct ceph_mds_client *mdsc = session->s_mdsc;
2973 	struct super_block *sb = mdsc->fsc->sb;
2974 	struct inode *inode;
2975 	struct ceph_inode_info *ci;
2976 	struct ceph_cap *cap;
2977 	struct ceph_mds_caps *h;
2978 	struct ceph_mds_cap_peer *peer = NULL;
2979 	int mds = session->s_mds;
2980 	int op;
2981 	u32 seq, mseq;
2982 	struct ceph_vino vino;
2983 	u64 cap_id;
2984 	u64 size, max_size;
2985 	u64 tid;
2986 	void *snaptrace;
2987 	size_t snaptrace_len;
2988 	void *flock;
2989 	void *end;
2990 	u32 flock_len;
2991 
2992 	dout("handle_caps from mds%d\n", mds);
2993 
2994 	/* decode */
2995 	end = msg->front.iov_base + msg->front.iov_len;
2996 	tid = le64_to_cpu(msg->hdr.tid);
2997 	if (msg->front.iov_len < sizeof(*h))
2998 		goto bad;
2999 	h = msg->front.iov_base;
3000 	op = le32_to_cpu(h->op);
3001 	vino.ino = le64_to_cpu(h->ino);
3002 	vino.snap = CEPH_NOSNAP;
3003 	cap_id = le64_to_cpu(h->cap_id);
3004 	seq = le32_to_cpu(h->seq);
3005 	mseq = le32_to_cpu(h->migrate_seq);
3006 	size = le64_to_cpu(h->size);
3007 	max_size = le64_to_cpu(h->max_size);
3008 
3009 	snaptrace = h + 1;
3010 	snaptrace_len = le32_to_cpu(h->snap_trace_len);
3011 
3012 	if (le16_to_cpu(msg->hdr.version) >= 2) {
3013 		void *p = snaptrace + snaptrace_len;
3014 		ceph_decode_32_safe(&p, end, flock_len, bad);
3015 		if (p + flock_len > end)
3016 			goto bad;
3017 		flock = p;
3018 	} else {
3019 		flock = NULL;
3020 		flock_len = 0;
3021 	}
3022 
3023 	if (le16_to_cpu(msg->hdr.version) >= 3) {
3024 		if (op == CEPH_CAP_OP_IMPORT) {
3025 			void *p = flock + flock_len;
3026 			if (p + sizeof(*peer) > end)
3027 				goto bad;
3028 			peer = p;
3029 		} else if (op == CEPH_CAP_OP_EXPORT) {
3030 			/* recorded in unused fields */
3031 			peer = (void *)&h->size;
3032 		}
3033 	}
3034 
3035 	mutex_lock(&session->s_mutex);
3036 	session->s_seq++;
3037 	dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3038 	     (unsigned)seq);
3039 
3040 	if (op == CEPH_CAP_OP_IMPORT)
3041 		ceph_add_cap_releases(mdsc, session);
3042 
3043 	/* lookup ino */
3044 	inode = ceph_find_inode(sb, vino);
3045 	ci = ceph_inode(inode);
3046 	dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3047 	     vino.snap, inode);
3048 	if (!inode) {
3049 		dout(" i don't have ino %llx\n", vino.ino);
3050 
3051 		if (op == CEPH_CAP_OP_IMPORT) {
3052 			spin_lock(&session->s_cap_lock);
3053 			__queue_cap_release(session, vino.ino, cap_id,
3054 					    mseq, seq);
3055 			spin_unlock(&session->s_cap_lock);
3056 		}
3057 		goto flush_cap_releases;
3058 	}
3059 
3060 	/* these will work even if we don't have a cap yet */
3061 	switch (op) {
3062 	case CEPH_CAP_OP_FLUSHSNAP_ACK:
3063 		handle_cap_flushsnap_ack(inode, tid, h, session);
3064 		goto done;
3065 
3066 	case CEPH_CAP_OP_EXPORT:
3067 		handle_cap_export(inode, h, peer, session);
3068 		goto done_unlocked;
3069 
3070 	case CEPH_CAP_OP_IMPORT:
3071 		handle_cap_import(mdsc, inode, h, peer, session,
3072 				  snaptrace, snaptrace_len);
3073 	}
3074 
3075 	/* the rest require a cap */
3076 	spin_lock(&ci->i_ceph_lock);
3077 	cap = __get_cap_for_mds(ceph_inode(inode), mds);
3078 	if (!cap) {
3079 		dout(" no cap on %p ino %llx.%llx from mds%d\n",
3080 		     inode, ceph_ino(inode), ceph_snap(inode), mds);
3081 		spin_unlock(&ci->i_ceph_lock);
3082 		goto flush_cap_releases;
3083 	}
3084 
3085 	/* note that each of these drops i_ceph_lock for us */
3086 	switch (op) {
3087 	case CEPH_CAP_OP_REVOKE:
3088 	case CEPH_CAP_OP_GRANT:
3089 	case CEPH_CAP_OP_IMPORT:
3090 		handle_cap_grant(inode, h, session, cap, msg->middle);
3091 		goto done_unlocked;
3092 
3093 	case CEPH_CAP_OP_FLUSH_ACK:
3094 		handle_cap_flush_ack(inode, tid, h, session, cap);
3095 		break;
3096 
3097 	case CEPH_CAP_OP_TRUNC:
3098 		handle_cap_trunc(inode, h, session);
3099 		break;
3100 
3101 	default:
3102 		spin_unlock(&ci->i_ceph_lock);
3103 		pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3104 		       ceph_cap_op_name(op));
3105 	}
3106 
3107 	goto done;
3108 
3109 flush_cap_releases:
3110 	/*
3111 	 * send any full release message to try to move things
3112 	 * along for the mds (who clearly thinks we still have this
3113 	 * cap).
3114 	 */
3115 	ceph_add_cap_releases(mdsc, session);
3116 	ceph_send_cap_releases(mdsc, session);
3117 
3118 done:
3119 	mutex_unlock(&session->s_mutex);
3120 done_unlocked:
3121 	if (inode)
3122 		iput(inode);
3123 	return;
3124 
3125 bad:
3126 	pr_err("ceph_handle_caps: corrupt message\n");
3127 	ceph_msg_dump(msg);
3128 	return;
3129 }
3130 
3131 /*
3132  * Delayed work handler to process end of delayed cap release LRU list.
3133  */
3134 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3135 {
3136 	struct ceph_inode_info *ci;
3137 	int flags = CHECK_CAPS_NODELAY;
3138 
3139 	dout("check_delayed_caps\n");
3140 	while (1) {
3141 		spin_lock(&mdsc->cap_delay_lock);
3142 		if (list_empty(&mdsc->cap_delay_list))
3143 			break;
3144 		ci = list_first_entry(&mdsc->cap_delay_list,
3145 				      struct ceph_inode_info,
3146 				      i_cap_delay_list);
3147 		if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3148 		    time_before(jiffies, ci->i_hold_caps_max))
3149 			break;
3150 		list_del_init(&ci->i_cap_delay_list);
3151 		spin_unlock(&mdsc->cap_delay_lock);
3152 		dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3153 		ceph_check_caps(ci, flags, NULL);
3154 	}
3155 	spin_unlock(&mdsc->cap_delay_lock);
3156 }
3157 
3158 /*
3159  * Flush all dirty caps to the mds
3160  */
3161 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3162 {
3163 	struct ceph_inode_info *ci;
3164 	struct inode *inode;
3165 
3166 	dout("flush_dirty_caps\n");
3167 	spin_lock(&mdsc->cap_dirty_lock);
3168 	while (!list_empty(&mdsc->cap_dirty)) {
3169 		ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3170 				      i_dirty_item);
3171 		inode = &ci->vfs_inode;
3172 		ihold(inode);
3173 		dout("flush_dirty_caps %p\n", inode);
3174 		spin_unlock(&mdsc->cap_dirty_lock);
3175 		ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3176 		iput(inode);
3177 		spin_lock(&mdsc->cap_dirty_lock);
3178 	}
3179 	spin_unlock(&mdsc->cap_dirty_lock);
3180 	dout("flush_dirty_caps done\n");
3181 }
3182 
3183 /*
3184  * Drop open file reference.  If we were the last open file,
3185  * we may need to release capabilities to the MDS (or schedule
3186  * their delayed release).
3187  */
3188 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3189 {
3190 	struct inode *inode = &ci->vfs_inode;
3191 	int last = 0;
3192 
3193 	spin_lock(&ci->i_ceph_lock);
3194 	dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
3195 	     ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
3196 	BUG_ON(ci->i_nr_by_mode[fmode] == 0);
3197 	if (--ci->i_nr_by_mode[fmode] == 0)
3198 		last++;
3199 	spin_unlock(&ci->i_ceph_lock);
3200 
3201 	if (last && ci->i_vino.snap == CEPH_NOSNAP)
3202 		ceph_check_caps(ci, 0, NULL);
3203 }
3204 
3205 /*
3206  * Helpers for embedding cap and dentry lease releases into mds
3207  * requests.
3208  *
3209  * @force is used by dentry_release (below) to force inclusion of a
3210  * record for the directory inode, even when there aren't any caps to
3211  * drop.
3212  */
3213 int ceph_encode_inode_release(void **p, struct inode *inode,
3214 			      int mds, int drop, int unless, int force)
3215 {
3216 	struct ceph_inode_info *ci = ceph_inode(inode);
3217 	struct ceph_cap *cap;
3218 	struct ceph_mds_request_release *rel = *p;
3219 	int used, dirty;
3220 	int ret = 0;
3221 
3222 	spin_lock(&ci->i_ceph_lock);
3223 	used = __ceph_caps_used(ci);
3224 	dirty = __ceph_caps_dirty(ci);
3225 
3226 	dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3227 	     inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3228 	     ceph_cap_string(unless));
3229 
3230 	/* only drop unused, clean caps */
3231 	drop &= ~(used | dirty);
3232 
3233 	cap = __get_cap_for_mds(ci, mds);
3234 	if (cap && __cap_is_valid(cap)) {
3235 		if (force ||
3236 		    ((cap->issued & drop) &&
3237 		     (cap->issued & unless) == 0)) {
3238 			if ((cap->issued & drop) &&
3239 			    (cap->issued & unless) == 0) {
3240 				int wanted = __ceph_caps_wanted(ci);
3241 				if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3242 					wanted |= cap->mds_wanted;
3243 				dout("encode_inode_release %p cap %p "
3244 				     "%s -> %s, wanted %s -> %s\n", inode, cap,
3245 				     ceph_cap_string(cap->issued),
3246 				     ceph_cap_string(cap->issued & ~drop),
3247 				     ceph_cap_string(cap->mds_wanted),
3248 				     ceph_cap_string(wanted));
3249 
3250 				cap->issued &= ~drop;
3251 				cap->implemented &= ~drop;
3252 				cap->mds_wanted = wanted;
3253 			} else {
3254 				dout("encode_inode_release %p cap %p %s"
3255 				     " (force)\n", inode, cap,
3256 				     ceph_cap_string(cap->issued));
3257 			}
3258 
3259 			rel->ino = cpu_to_le64(ceph_ino(inode));
3260 			rel->cap_id = cpu_to_le64(cap->cap_id);
3261 			rel->seq = cpu_to_le32(cap->seq);
3262 			rel->issue_seq = cpu_to_le32(cap->issue_seq),
3263 			rel->mseq = cpu_to_le32(cap->mseq);
3264 			rel->caps = cpu_to_le32(cap->issued);
3265 			rel->wanted = cpu_to_le32(cap->mds_wanted);
3266 			rel->dname_len = 0;
3267 			rel->dname_seq = 0;
3268 			*p += sizeof(*rel);
3269 			ret = 1;
3270 		} else {
3271 			dout("encode_inode_release %p cap %p %s\n",
3272 			     inode, cap, ceph_cap_string(cap->issued));
3273 		}
3274 	}
3275 	spin_unlock(&ci->i_ceph_lock);
3276 	return ret;
3277 }
3278 
3279 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3280 			       int mds, int drop, int unless)
3281 {
3282 	struct inode *dir = dentry->d_parent->d_inode;
3283 	struct ceph_mds_request_release *rel = *p;
3284 	struct ceph_dentry_info *di = ceph_dentry(dentry);
3285 	int force = 0;
3286 	int ret;
3287 
3288 	/*
3289 	 * force an record for the directory caps if we have a dentry lease.
3290 	 * this is racy (can't take i_ceph_lock and d_lock together), but it
3291 	 * doesn't have to be perfect; the mds will revoke anything we don't
3292 	 * release.
3293 	 */
3294 	spin_lock(&dentry->d_lock);
3295 	if (di->lease_session && di->lease_session->s_mds == mds)
3296 		force = 1;
3297 	spin_unlock(&dentry->d_lock);
3298 
3299 	ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3300 
3301 	spin_lock(&dentry->d_lock);
3302 	if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3303 		dout("encode_dentry_release %p mds%d seq %d\n",
3304 		     dentry, mds, (int)di->lease_seq);
3305 		rel->dname_len = cpu_to_le32(dentry->d_name.len);
3306 		memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3307 		*p += dentry->d_name.len;
3308 		rel->dname_seq = cpu_to_le32(di->lease_seq);
3309 		__ceph_mdsc_drop_dentry_lease(dentry);
3310 	}
3311 	spin_unlock(&dentry->d_lock);
3312 	return ret;
3313 }
3314