xref: /openbmc/linux/fs/ceph/caps.c (revision 87875c10)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
3 
4 #include <linux/fs.h>
5 #include <linux/kernel.h>
6 #include <linux/sched/signal.h>
7 #include <linux/slab.h>
8 #include <linux/vmalloc.h>
9 #include <linux/wait.h>
10 #include <linux/writeback.h>
11 #include <linux/iversion.h>
12 
13 #include "super.h"
14 #include "mds_client.h"
15 #include "cache.h"
16 #include <linux/ceph/decode.h>
17 #include <linux/ceph/messenger.h>
18 
19 /*
20  * Capability management
21  *
22  * The Ceph metadata servers control client access to inode metadata
23  * and file data by issuing capabilities, granting clients permission
24  * to read and/or write both inode field and file data to OSDs
25  * (storage nodes).  Each capability consists of a set of bits
26  * indicating which operations are allowed.
27  *
28  * If the client holds a *_SHARED cap, the client has a coherent value
29  * that can be safely read from the cached inode.
30  *
31  * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
32  * client is allowed to change inode attributes (e.g., file size,
33  * mtime), note its dirty state in the ceph_cap, and asynchronously
34  * flush that metadata change to the MDS.
35  *
36  * In the event of a conflicting operation (perhaps by another
37  * client), the MDS will revoke the conflicting client capabilities.
38  *
39  * In order for a client to cache an inode, it must hold a capability
40  * with at least one MDS server.  When inodes are released, release
41  * notifications are batched and periodically sent en masse to the MDS
42  * cluster to release server state.
43  */
44 
45 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
46 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
47 				 struct ceph_mds_session *session,
48 				 struct ceph_inode_info *ci,
49 				 u64 oldest_flush_tid);
50 
51 /*
52  * Generate readable cap strings for debugging output.
53  */
54 #define MAX_CAP_STR 20
55 static char cap_str[MAX_CAP_STR][40];
56 static DEFINE_SPINLOCK(cap_str_lock);
57 static int last_cap_str;
58 
59 static char *gcap_string(char *s, int c)
60 {
61 	if (c & CEPH_CAP_GSHARED)
62 		*s++ = 's';
63 	if (c & CEPH_CAP_GEXCL)
64 		*s++ = 'x';
65 	if (c & CEPH_CAP_GCACHE)
66 		*s++ = 'c';
67 	if (c & CEPH_CAP_GRD)
68 		*s++ = 'r';
69 	if (c & CEPH_CAP_GWR)
70 		*s++ = 'w';
71 	if (c & CEPH_CAP_GBUFFER)
72 		*s++ = 'b';
73 	if (c & CEPH_CAP_GWREXTEND)
74 		*s++ = 'a';
75 	if (c & CEPH_CAP_GLAZYIO)
76 		*s++ = 'l';
77 	return s;
78 }
79 
80 const char *ceph_cap_string(int caps)
81 {
82 	int i;
83 	char *s;
84 	int c;
85 
86 	spin_lock(&cap_str_lock);
87 	i = last_cap_str++;
88 	if (last_cap_str == MAX_CAP_STR)
89 		last_cap_str = 0;
90 	spin_unlock(&cap_str_lock);
91 
92 	s = cap_str[i];
93 
94 	if (caps & CEPH_CAP_PIN)
95 		*s++ = 'p';
96 
97 	c = (caps >> CEPH_CAP_SAUTH) & 3;
98 	if (c) {
99 		*s++ = 'A';
100 		s = gcap_string(s, c);
101 	}
102 
103 	c = (caps >> CEPH_CAP_SLINK) & 3;
104 	if (c) {
105 		*s++ = 'L';
106 		s = gcap_string(s, c);
107 	}
108 
109 	c = (caps >> CEPH_CAP_SXATTR) & 3;
110 	if (c) {
111 		*s++ = 'X';
112 		s = gcap_string(s, c);
113 	}
114 
115 	c = caps >> CEPH_CAP_SFILE;
116 	if (c) {
117 		*s++ = 'F';
118 		s = gcap_string(s, c);
119 	}
120 
121 	if (s == cap_str[i])
122 		*s++ = '-';
123 	*s = 0;
124 	return cap_str[i];
125 }
126 
127 void ceph_caps_init(struct ceph_mds_client *mdsc)
128 {
129 	INIT_LIST_HEAD(&mdsc->caps_list);
130 	spin_lock_init(&mdsc->caps_list_lock);
131 }
132 
133 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
134 {
135 	struct ceph_cap *cap;
136 
137 	spin_lock(&mdsc->caps_list_lock);
138 	while (!list_empty(&mdsc->caps_list)) {
139 		cap = list_first_entry(&mdsc->caps_list,
140 				       struct ceph_cap, caps_item);
141 		list_del(&cap->caps_item);
142 		kmem_cache_free(ceph_cap_cachep, cap);
143 	}
144 	mdsc->caps_total_count = 0;
145 	mdsc->caps_avail_count = 0;
146 	mdsc->caps_use_count = 0;
147 	mdsc->caps_reserve_count = 0;
148 	mdsc->caps_min_count = 0;
149 	spin_unlock(&mdsc->caps_list_lock);
150 }
151 
152 void ceph_adjust_caps_max_min(struct ceph_mds_client *mdsc,
153 			      struct ceph_mount_options *fsopt)
154 {
155 	spin_lock(&mdsc->caps_list_lock);
156 	mdsc->caps_min_count = fsopt->max_readdir;
157 	if (mdsc->caps_min_count < 1024)
158 		mdsc->caps_min_count = 1024;
159 	mdsc->caps_use_max = fsopt->caps_max;
160 	if (mdsc->caps_use_max > 0 &&
161 	    mdsc->caps_use_max < mdsc->caps_min_count)
162 		mdsc->caps_use_max = mdsc->caps_min_count;
163 	spin_unlock(&mdsc->caps_list_lock);
164 }
165 
166 static void __ceph_unreserve_caps(struct ceph_mds_client *mdsc, int nr_caps)
167 {
168 	struct ceph_cap *cap;
169 	int i;
170 
171 	if (nr_caps) {
172 		BUG_ON(mdsc->caps_reserve_count < nr_caps);
173 		mdsc->caps_reserve_count -= nr_caps;
174 		if (mdsc->caps_avail_count >=
175 		    mdsc->caps_reserve_count + mdsc->caps_min_count) {
176 			mdsc->caps_total_count -= nr_caps;
177 			for (i = 0; i < nr_caps; i++) {
178 				cap = list_first_entry(&mdsc->caps_list,
179 					struct ceph_cap, caps_item);
180 				list_del(&cap->caps_item);
181 				kmem_cache_free(ceph_cap_cachep, cap);
182 			}
183 		} else {
184 			mdsc->caps_avail_count += nr_caps;
185 		}
186 
187 		dout("%s: caps %d = %d used + %d resv + %d avail\n",
188 		     __func__,
189 		     mdsc->caps_total_count, mdsc->caps_use_count,
190 		     mdsc->caps_reserve_count, mdsc->caps_avail_count);
191 		BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
192 						 mdsc->caps_reserve_count +
193 						 mdsc->caps_avail_count);
194 	}
195 }
196 
197 /*
198  * Called under mdsc->mutex.
199  */
200 int ceph_reserve_caps(struct ceph_mds_client *mdsc,
201 		      struct ceph_cap_reservation *ctx, int need)
202 {
203 	int i, j;
204 	struct ceph_cap *cap;
205 	int have;
206 	int alloc = 0;
207 	int max_caps;
208 	int err = 0;
209 	bool trimmed = false;
210 	struct ceph_mds_session *s;
211 	LIST_HEAD(newcaps);
212 
213 	dout("reserve caps ctx=%p need=%d\n", ctx, need);
214 
215 	/* first reserve any caps that are already allocated */
216 	spin_lock(&mdsc->caps_list_lock);
217 	if (mdsc->caps_avail_count >= need)
218 		have = need;
219 	else
220 		have = mdsc->caps_avail_count;
221 	mdsc->caps_avail_count -= have;
222 	mdsc->caps_reserve_count += have;
223 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
224 					 mdsc->caps_reserve_count +
225 					 mdsc->caps_avail_count);
226 	spin_unlock(&mdsc->caps_list_lock);
227 
228 	for (i = have; i < need; ) {
229 		cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
230 		if (cap) {
231 			list_add(&cap->caps_item, &newcaps);
232 			alloc++;
233 			i++;
234 			continue;
235 		}
236 
237 		if (!trimmed) {
238 			for (j = 0; j < mdsc->max_sessions; j++) {
239 				s = __ceph_lookup_mds_session(mdsc, j);
240 				if (!s)
241 					continue;
242 				mutex_unlock(&mdsc->mutex);
243 
244 				mutex_lock(&s->s_mutex);
245 				max_caps = s->s_nr_caps - (need - i);
246 				ceph_trim_caps(mdsc, s, max_caps);
247 				mutex_unlock(&s->s_mutex);
248 
249 				ceph_put_mds_session(s);
250 				mutex_lock(&mdsc->mutex);
251 			}
252 			trimmed = true;
253 
254 			spin_lock(&mdsc->caps_list_lock);
255 			if (mdsc->caps_avail_count) {
256 				int more_have;
257 				if (mdsc->caps_avail_count >= need - i)
258 					more_have = need - i;
259 				else
260 					more_have = mdsc->caps_avail_count;
261 
262 				i += more_have;
263 				have += more_have;
264 				mdsc->caps_avail_count -= more_have;
265 				mdsc->caps_reserve_count += more_have;
266 
267 			}
268 			spin_unlock(&mdsc->caps_list_lock);
269 
270 			continue;
271 		}
272 
273 		pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
274 			ctx, need, have + alloc);
275 		err = -ENOMEM;
276 		break;
277 	}
278 
279 	if (!err) {
280 		BUG_ON(have + alloc != need);
281 		ctx->count = need;
282 		ctx->used = 0;
283 	}
284 
285 	spin_lock(&mdsc->caps_list_lock);
286 	mdsc->caps_total_count += alloc;
287 	mdsc->caps_reserve_count += alloc;
288 	list_splice(&newcaps, &mdsc->caps_list);
289 
290 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
291 					 mdsc->caps_reserve_count +
292 					 mdsc->caps_avail_count);
293 
294 	if (err)
295 		__ceph_unreserve_caps(mdsc, have + alloc);
296 
297 	spin_unlock(&mdsc->caps_list_lock);
298 
299 	dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
300 	     ctx, mdsc->caps_total_count, mdsc->caps_use_count,
301 	     mdsc->caps_reserve_count, mdsc->caps_avail_count);
302 	return err;
303 }
304 
305 void ceph_unreserve_caps(struct ceph_mds_client *mdsc,
306 			 struct ceph_cap_reservation *ctx)
307 {
308 	bool reclaim = false;
309 	if (!ctx->count)
310 		return;
311 
312 	dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
313 	spin_lock(&mdsc->caps_list_lock);
314 	__ceph_unreserve_caps(mdsc, ctx->count);
315 	ctx->count = 0;
316 
317 	if (mdsc->caps_use_max > 0 &&
318 	    mdsc->caps_use_count > mdsc->caps_use_max)
319 		reclaim = true;
320 	spin_unlock(&mdsc->caps_list_lock);
321 
322 	if (reclaim)
323 		ceph_reclaim_caps_nr(mdsc, ctx->used);
324 }
325 
326 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
327 			      struct ceph_cap_reservation *ctx)
328 {
329 	struct ceph_cap *cap = NULL;
330 
331 	/* temporary, until we do something about cap import/export */
332 	if (!ctx) {
333 		cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
334 		if (cap) {
335 			spin_lock(&mdsc->caps_list_lock);
336 			mdsc->caps_use_count++;
337 			mdsc->caps_total_count++;
338 			spin_unlock(&mdsc->caps_list_lock);
339 		} else {
340 			spin_lock(&mdsc->caps_list_lock);
341 			if (mdsc->caps_avail_count) {
342 				BUG_ON(list_empty(&mdsc->caps_list));
343 
344 				mdsc->caps_avail_count--;
345 				mdsc->caps_use_count++;
346 				cap = list_first_entry(&mdsc->caps_list,
347 						struct ceph_cap, caps_item);
348 				list_del(&cap->caps_item);
349 
350 				BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
351 				       mdsc->caps_reserve_count + mdsc->caps_avail_count);
352 			}
353 			spin_unlock(&mdsc->caps_list_lock);
354 		}
355 
356 		return cap;
357 	}
358 
359 	spin_lock(&mdsc->caps_list_lock);
360 	dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
361 	     ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
362 	     mdsc->caps_reserve_count, mdsc->caps_avail_count);
363 	BUG_ON(!ctx->count);
364 	BUG_ON(ctx->count > mdsc->caps_reserve_count);
365 	BUG_ON(list_empty(&mdsc->caps_list));
366 
367 	ctx->count--;
368 	ctx->used++;
369 	mdsc->caps_reserve_count--;
370 	mdsc->caps_use_count++;
371 
372 	cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
373 	list_del(&cap->caps_item);
374 
375 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
376 	       mdsc->caps_reserve_count + mdsc->caps_avail_count);
377 	spin_unlock(&mdsc->caps_list_lock);
378 	return cap;
379 }
380 
381 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
382 {
383 	spin_lock(&mdsc->caps_list_lock);
384 	dout("put_cap %p %d = %d used + %d resv + %d avail\n",
385 	     cap, mdsc->caps_total_count, mdsc->caps_use_count,
386 	     mdsc->caps_reserve_count, mdsc->caps_avail_count);
387 	mdsc->caps_use_count--;
388 	/*
389 	 * Keep some preallocated caps around (ceph_min_count), to
390 	 * avoid lots of free/alloc churn.
391 	 */
392 	if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
393 				      mdsc->caps_min_count) {
394 		mdsc->caps_total_count--;
395 		kmem_cache_free(ceph_cap_cachep, cap);
396 	} else {
397 		mdsc->caps_avail_count++;
398 		list_add(&cap->caps_item, &mdsc->caps_list);
399 	}
400 
401 	BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
402 	       mdsc->caps_reserve_count + mdsc->caps_avail_count);
403 	spin_unlock(&mdsc->caps_list_lock);
404 }
405 
406 void ceph_reservation_status(struct ceph_fs_client *fsc,
407 			     int *total, int *avail, int *used, int *reserved,
408 			     int *min)
409 {
410 	struct ceph_mds_client *mdsc = fsc->mdsc;
411 
412 	spin_lock(&mdsc->caps_list_lock);
413 
414 	if (total)
415 		*total = mdsc->caps_total_count;
416 	if (avail)
417 		*avail = mdsc->caps_avail_count;
418 	if (used)
419 		*used = mdsc->caps_use_count;
420 	if (reserved)
421 		*reserved = mdsc->caps_reserve_count;
422 	if (min)
423 		*min = mdsc->caps_min_count;
424 
425 	spin_unlock(&mdsc->caps_list_lock);
426 }
427 
428 /*
429  * Find ceph_cap for given mds, if any.
430  *
431  * Called with i_ceph_lock held.
432  */
433 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
434 {
435 	struct ceph_cap *cap;
436 	struct rb_node *n = ci->i_caps.rb_node;
437 
438 	while (n) {
439 		cap = rb_entry(n, struct ceph_cap, ci_node);
440 		if (mds < cap->mds)
441 			n = n->rb_left;
442 		else if (mds > cap->mds)
443 			n = n->rb_right;
444 		else
445 			return cap;
446 	}
447 	return NULL;
448 }
449 
450 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
451 {
452 	struct ceph_cap *cap;
453 
454 	spin_lock(&ci->i_ceph_lock);
455 	cap = __get_cap_for_mds(ci, mds);
456 	spin_unlock(&ci->i_ceph_lock);
457 	return cap;
458 }
459 
460 /*
461  * Called under i_ceph_lock.
462  */
463 static void __insert_cap_node(struct ceph_inode_info *ci,
464 			      struct ceph_cap *new)
465 {
466 	struct rb_node **p = &ci->i_caps.rb_node;
467 	struct rb_node *parent = NULL;
468 	struct ceph_cap *cap = NULL;
469 
470 	while (*p) {
471 		parent = *p;
472 		cap = rb_entry(parent, struct ceph_cap, ci_node);
473 		if (new->mds < cap->mds)
474 			p = &(*p)->rb_left;
475 		else if (new->mds > cap->mds)
476 			p = &(*p)->rb_right;
477 		else
478 			BUG();
479 	}
480 
481 	rb_link_node(&new->ci_node, parent, p);
482 	rb_insert_color(&new->ci_node, &ci->i_caps);
483 }
484 
485 /*
486  * (re)set cap hold timeouts, which control the delayed release
487  * of unused caps back to the MDS.  Should be called on cap use.
488  */
489 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
490 			       struct ceph_inode_info *ci)
491 {
492 	struct ceph_mount_options *opt = mdsc->fsc->mount_options;
493 	ci->i_hold_caps_max = round_jiffies(jiffies +
494 					    opt->caps_wanted_delay_max * HZ);
495 	dout("__cap_set_timeouts %p %lu\n", &ci->vfs_inode,
496 	     ci->i_hold_caps_max - jiffies);
497 }
498 
499 /*
500  * (Re)queue cap at the end of the delayed cap release list.
501  *
502  * If I_FLUSH is set, leave the inode at the front of the list.
503  *
504  * Caller holds i_ceph_lock
505  *    -> we take mdsc->cap_delay_lock
506  */
507 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
508 				struct ceph_inode_info *ci)
509 {
510 	dout("__cap_delay_requeue %p flags 0x%lx at %lu\n", &ci->vfs_inode,
511 	     ci->i_ceph_flags, ci->i_hold_caps_max);
512 	if (!mdsc->stopping) {
513 		spin_lock(&mdsc->cap_delay_lock);
514 		if (!list_empty(&ci->i_cap_delay_list)) {
515 			if (ci->i_ceph_flags & CEPH_I_FLUSH)
516 				goto no_change;
517 			list_del_init(&ci->i_cap_delay_list);
518 		}
519 		__cap_set_timeouts(mdsc, ci);
520 		list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
521 no_change:
522 		spin_unlock(&mdsc->cap_delay_lock);
523 	}
524 }
525 
526 /*
527  * Queue an inode for immediate writeback.  Mark inode with I_FLUSH,
528  * indicating we should send a cap message to flush dirty metadata
529  * asap, and move to the front of the delayed cap list.
530  */
531 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
532 				      struct ceph_inode_info *ci)
533 {
534 	dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
535 	spin_lock(&mdsc->cap_delay_lock);
536 	ci->i_ceph_flags |= CEPH_I_FLUSH;
537 	if (!list_empty(&ci->i_cap_delay_list))
538 		list_del_init(&ci->i_cap_delay_list);
539 	list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
540 	spin_unlock(&mdsc->cap_delay_lock);
541 }
542 
543 /*
544  * Cancel delayed work on cap.
545  *
546  * Caller must hold i_ceph_lock.
547  */
548 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
549 			       struct ceph_inode_info *ci)
550 {
551 	dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
552 	if (list_empty(&ci->i_cap_delay_list))
553 		return;
554 	spin_lock(&mdsc->cap_delay_lock);
555 	list_del_init(&ci->i_cap_delay_list);
556 	spin_unlock(&mdsc->cap_delay_lock);
557 }
558 
559 /* Common issue checks for add_cap, handle_cap_grant. */
560 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
561 			      unsigned issued)
562 {
563 	unsigned had = __ceph_caps_issued(ci, NULL);
564 
565 	lockdep_assert_held(&ci->i_ceph_lock);
566 
567 	/*
568 	 * Each time we receive FILE_CACHE anew, we increment
569 	 * i_rdcache_gen.
570 	 */
571 	if (S_ISREG(ci->vfs_inode.i_mode) &&
572 	    (issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
573 	    (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
574 		ci->i_rdcache_gen++;
575 	}
576 
577 	/*
578 	 * If FILE_SHARED is newly issued, mark dir not complete. We don't
579 	 * know what happened to this directory while we didn't have the cap.
580 	 * If FILE_SHARED is being revoked, also mark dir not complete. It
581 	 * stops on-going cached readdir.
582 	 */
583 	if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
584 		if (issued & CEPH_CAP_FILE_SHARED)
585 			atomic_inc(&ci->i_shared_gen);
586 		if (S_ISDIR(ci->vfs_inode.i_mode)) {
587 			dout(" marking %p NOT complete\n", &ci->vfs_inode);
588 			__ceph_dir_clear_complete(ci);
589 		}
590 	}
591 
592 	/* Wipe saved layout if we're losing DIR_CREATE caps */
593 	if (S_ISDIR(ci->vfs_inode.i_mode) && (had & CEPH_CAP_DIR_CREATE) &&
594 		!(issued & CEPH_CAP_DIR_CREATE)) {
595 	     ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
596 	     memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
597 	}
598 }
599 
600 /**
601  * change_auth_cap_ses - move inode to appropriate lists when auth caps change
602  * @ci: inode to be moved
603  * @session: new auth caps session
604  */
605 static void change_auth_cap_ses(struct ceph_inode_info *ci,
606 				struct ceph_mds_session *session)
607 {
608 	lockdep_assert_held(&ci->i_ceph_lock);
609 
610 	if (list_empty(&ci->i_dirty_item) && list_empty(&ci->i_flushing_item))
611 		return;
612 
613 	spin_lock(&session->s_mdsc->cap_dirty_lock);
614 	if (!list_empty(&ci->i_dirty_item))
615 		list_move(&ci->i_dirty_item, &session->s_cap_dirty);
616 	if (!list_empty(&ci->i_flushing_item))
617 		list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
618 	spin_unlock(&session->s_mdsc->cap_dirty_lock);
619 }
620 
621 /*
622  * Add a capability under the given MDS session.
623  *
624  * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock
625  *
626  * @fmode is the open file mode, if we are opening a file, otherwise
627  * it is < 0.  (This is so we can atomically add the cap and add an
628  * open file reference to it.)
629  */
630 void ceph_add_cap(struct inode *inode,
631 		  struct ceph_mds_session *session, u64 cap_id,
632 		  unsigned issued, unsigned wanted,
633 		  unsigned seq, unsigned mseq, u64 realmino, int flags,
634 		  struct ceph_cap **new_cap)
635 {
636 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
637 	struct ceph_inode_info *ci = ceph_inode(inode);
638 	struct ceph_cap *cap;
639 	int mds = session->s_mds;
640 	int actual_wanted;
641 	u32 gen;
642 
643 	lockdep_assert_held(&ci->i_ceph_lock);
644 
645 	dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
646 	     session->s_mds, cap_id, ceph_cap_string(issued), seq);
647 
648 	gen = atomic_read(&session->s_cap_gen);
649 
650 	cap = __get_cap_for_mds(ci, mds);
651 	if (!cap) {
652 		cap = *new_cap;
653 		*new_cap = NULL;
654 
655 		cap->issued = 0;
656 		cap->implemented = 0;
657 		cap->mds = mds;
658 		cap->mds_wanted = 0;
659 		cap->mseq = 0;
660 
661 		cap->ci = ci;
662 		__insert_cap_node(ci, cap);
663 
664 		/* add to session cap list */
665 		cap->session = session;
666 		spin_lock(&session->s_cap_lock);
667 		list_add_tail(&cap->session_caps, &session->s_caps);
668 		session->s_nr_caps++;
669 		atomic64_inc(&mdsc->metric.total_caps);
670 		spin_unlock(&session->s_cap_lock);
671 	} else {
672 		spin_lock(&session->s_cap_lock);
673 		list_move_tail(&cap->session_caps, &session->s_caps);
674 		spin_unlock(&session->s_cap_lock);
675 
676 		if (cap->cap_gen < gen)
677 			cap->issued = cap->implemented = CEPH_CAP_PIN;
678 
679 		/*
680 		 * auth mds of the inode changed. we received the cap export
681 		 * message, but still haven't received the cap import message.
682 		 * handle_cap_export() updated the new auth MDS' cap.
683 		 *
684 		 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
685 		 * a message that was send before the cap import message. So
686 		 * don't remove caps.
687 		 */
688 		if (ceph_seq_cmp(seq, cap->seq) <= 0) {
689 			WARN_ON(cap != ci->i_auth_cap);
690 			WARN_ON(cap->cap_id != cap_id);
691 			seq = cap->seq;
692 			mseq = cap->mseq;
693 			issued |= cap->issued;
694 			flags |= CEPH_CAP_FLAG_AUTH;
695 		}
696 	}
697 
698 	if (!ci->i_snap_realm ||
699 	    ((flags & CEPH_CAP_FLAG_AUTH) &&
700 	     realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
701 		/*
702 		 * add this inode to the appropriate snap realm
703 		 */
704 		struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
705 							       realmino);
706 		if (realm) {
707 			struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
708 			if (oldrealm) {
709 				spin_lock(&oldrealm->inodes_with_caps_lock);
710 				list_del_init(&ci->i_snap_realm_item);
711 				spin_unlock(&oldrealm->inodes_with_caps_lock);
712 			}
713 
714 			spin_lock(&realm->inodes_with_caps_lock);
715 			list_add(&ci->i_snap_realm_item,
716 				 &realm->inodes_with_caps);
717 			ci->i_snap_realm = realm;
718 			if (realm->ino == ci->i_vino.ino)
719 				realm->inode = inode;
720 			spin_unlock(&realm->inodes_with_caps_lock);
721 
722 			if (oldrealm)
723 				ceph_put_snap_realm(mdsc, oldrealm);
724 		} else {
725 			pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
726 			       realmino);
727 			WARN_ON(!realm);
728 		}
729 	}
730 
731 	__check_cap_issue(ci, cap, issued);
732 
733 	/*
734 	 * If we are issued caps we don't want, or the mds' wanted
735 	 * value appears to be off, queue a check so we'll release
736 	 * later and/or update the mds wanted value.
737 	 */
738 	actual_wanted = __ceph_caps_wanted(ci);
739 	if ((wanted & ~actual_wanted) ||
740 	    (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
741 		dout(" issued %s, mds wanted %s, actual %s, queueing\n",
742 		     ceph_cap_string(issued), ceph_cap_string(wanted),
743 		     ceph_cap_string(actual_wanted));
744 		__cap_delay_requeue(mdsc, ci);
745 	}
746 
747 	if (flags & CEPH_CAP_FLAG_AUTH) {
748 		if (!ci->i_auth_cap ||
749 		    ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
750 			if (ci->i_auth_cap &&
751 			    ci->i_auth_cap->session != cap->session)
752 				change_auth_cap_ses(ci, cap->session);
753 			ci->i_auth_cap = cap;
754 			cap->mds_wanted = wanted;
755 		}
756 	} else {
757 		WARN_ON(ci->i_auth_cap == cap);
758 	}
759 
760 	dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
761 	     inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
762 	     ceph_cap_string(issued|cap->issued), seq, mds);
763 	cap->cap_id = cap_id;
764 	cap->issued = issued;
765 	cap->implemented |= issued;
766 	if (ceph_seq_cmp(mseq, cap->mseq) > 0)
767 		cap->mds_wanted = wanted;
768 	else
769 		cap->mds_wanted |= wanted;
770 	cap->seq = seq;
771 	cap->issue_seq = seq;
772 	cap->mseq = mseq;
773 	cap->cap_gen = gen;
774 }
775 
776 /*
777  * Return true if cap has not timed out and belongs to the current
778  * generation of the MDS session (i.e. has not gone 'stale' due to
779  * us losing touch with the mds).
780  */
781 static int __cap_is_valid(struct ceph_cap *cap)
782 {
783 	unsigned long ttl;
784 	u32 gen;
785 
786 	gen = atomic_read(&cap->session->s_cap_gen);
787 	ttl = cap->session->s_cap_ttl;
788 
789 	if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
790 		dout("__cap_is_valid %p cap %p issued %s "
791 		     "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
792 		     cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
793 		return 0;
794 	}
795 
796 	return 1;
797 }
798 
799 /*
800  * Return set of valid cap bits issued to us.  Note that caps time
801  * out, and may be invalidated in bulk if the client session times out
802  * and session->s_cap_gen is bumped.
803  */
804 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
805 {
806 	int have = ci->i_snap_caps;
807 	struct ceph_cap *cap;
808 	struct rb_node *p;
809 
810 	if (implemented)
811 		*implemented = 0;
812 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
813 		cap = rb_entry(p, struct ceph_cap, ci_node);
814 		if (!__cap_is_valid(cap))
815 			continue;
816 		dout("__ceph_caps_issued %p cap %p issued %s\n",
817 		     &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
818 		have |= cap->issued;
819 		if (implemented)
820 			*implemented |= cap->implemented;
821 	}
822 	/*
823 	 * exclude caps issued by non-auth MDS, but are been revoking
824 	 * by the auth MDS. The non-auth MDS should be revoking/exporting
825 	 * these caps, but the message is delayed.
826 	 */
827 	if (ci->i_auth_cap) {
828 		cap = ci->i_auth_cap;
829 		have &= ~cap->implemented | cap->issued;
830 	}
831 	return have;
832 }
833 
834 /*
835  * Get cap bits issued by caps other than @ocap
836  */
837 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
838 {
839 	int have = ci->i_snap_caps;
840 	struct ceph_cap *cap;
841 	struct rb_node *p;
842 
843 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
844 		cap = rb_entry(p, struct ceph_cap, ci_node);
845 		if (cap == ocap)
846 			continue;
847 		if (!__cap_is_valid(cap))
848 			continue;
849 		have |= cap->issued;
850 	}
851 	return have;
852 }
853 
854 /*
855  * Move a cap to the end of the LRU (oldest caps at list head, newest
856  * at list tail).
857  */
858 static void __touch_cap(struct ceph_cap *cap)
859 {
860 	struct ceph_mds_session *s = cap->session;
861 
862 	spin_lock(&s->s_cap_lock);
863 	if (!s->s_cap_iterator) {
864 		dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
865 		     s->s_mds);
866 		list_move_tail(&cap->session_caps, &s->s_caps);
867 	} else {
868 		dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
869 		     &cap->ci->vfs_inode, cap, s->s_mds);
870 	}
871 	spin_unlock(&s->s_cap_lock);
872 }
873 
874 /*
875  * Check if we hold the given mask.  If so, move the cap(s) to the
876  * front of their respective LRUs.  (This is the preferred way for
877  * callers to check for caps they want.)
878  */
879 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
880 {
881 	struct ceph_cap *cap;
882 	struct rb_node *p;
883 	int have = ci->i_snap_caps;
884 
885 	if ((have & mask) == mask) {
886 		dout("__ceph_caps_issued_mask ino 0x%llx snap issued %s"
887 		     " (mask %s)\n", ceph_ino(&ci->vfs_inode),
888 		     ceph_cap_string(have),
889 		     ceph_cap_string(mask));
890 		return 1;
891 	}
892 
893 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
894 		cap = rb_entry(p, struct ceph_cap, ci_node);
895 		if (!__cap_is_valid(cap))
896 			continue;
897 		if ((cap->issued & mask) == mask) {
898 			dout("__ceph_caps_issued_mask ino 0x%llx cap %p issued %s"
899 			     " (mask %s)\n", ceph_ino(&ci->vfs_inode), cap,
900 			     ceph_cap_string(cap->issued),
901 			     ceph_cap_string(mask));
902 			if (touch)
903 				__touch_cap(cap);
904 			return 1;
905 		}
906 
907 		/* does a combination of caps satisfy mask? */
908 		have |= cap->issued;
909 		if ((have & mask) == mask) {
910 			dout("__ceph_caps_issued_mask ino 0x%llx combo issued %s"
911 			     " (mask %s)\n", ceph_ino(&ci->vfs_inode),
912 			     ceph_cap_string(cap->issued),
913 			     ceph_cap_string(mask));
914 			if (touch) {
915 				struct rb_node *q;
916 
917 				/* touch this + preceding caps */
918 				__touch_cap(cap);
919 				for (q = rb_first(&ci->i_caps); q != p;
920 				     q = rb_next(q)) {
921 					cap = rb_entry(q, struct ceph_cap,
922 						       ci_node);
923 					if (!__cap_is_valid(cap))
924 						continue;
925 					if (cap->issued & mask)
926 						__touch_cap(cap);
927 				}
928 			}
929 			return 1;
930 		}
931 	}
932 
933 	return 0;
934 }
935 
936 int __ceph_caps_issued_mask_metric(struct ceph_inode_info *ci, int mask,
937 				   int touch)
938 {
939 	struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb);
940 	int r;
941 
942 	r = __ceph_caps_issued_mask(ci, mask, touch);
943 	if (r)
944 		ceph_update_cap_hit(&fsc->mdsc->metric);
945 	else
946 		ceph_update_cap_mis(&fsc->mdsc->metric);
947 	return r;
948 }
949 
950 /*
951  * Return true if mask caps are currently being revoked by an MDS.
952  */
953 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
954 			       struct ceph_cap *ocap, int mask)
955 {
956 	struct ceph_cap *cap;
957 	struct rb_node *p;
958 
959 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
960 		cap = rb_entry(p, struct ceph_cap, ci_node);
961 		if (cap != ocap &&
962 		    (cap->implemented & ~cap->issued & mask))
963 			return 1;
964 	}
965 	return 0;
966 }
967 
968 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
969 {
970 	struct inode *inode = &ci->vfs_inode;
971 	int ret;
972 
973 	spin_lock(&ci->i_ceph_lock);
974 	ret = __ceph_caps_revoking_other(ci, NULL, mask);
975 	spin_unlock(&ci->i_ceph_lock);
976 	dout("ceph_caps_revoking %p %s = %d\n", inode,
977 	     ceph_cap_string(mask), ret);
978 	return ret;
979 }
980 
981 int __ceph_caps_used(struct ceph_inode_info *ci)
982 {
983 	int used = 0;
984 	if (ci->i_pin_ref)
985 		used |= CEPH_CAP_PIN;
986 	if (ci->i_rd_ref)
987 		used |= CEPH_CAP_FILE_RD;
988 	if (ci->i_rdcache_ref ||
989 	    (S_ISREG(ci->vfs_inode.i_mode) &&
990 	     ci->vfs_inode.i_data.nrpages))
991 		used |= CEPH_CAP_FILE_CACHE;
992 	if (ci->i_wr_ref)
993 		used |= CEPH_CAP_FILE_WR;
994 	if (ci->i_wb_ref || ci->i_wrbuffer_ref)
995 		used |= CEPH_CAP_FILE_BUFFER;
996 	if (ci->i_fx_ref)
997 		used |= CEPH_CAP_FILE_EXCL;
998 	return used;
999 }
1000 
1001 #define FMODE_WAIT_BIAS 1000
1002 
1003 /*
1004  * wanted, by virtue of open file modes
1005  */
1006 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
1007 {
1008 	const int PIN_SHIFT = ffs(CEPH_FILE_MODE_PIN);
1009 	const int RD_SHIFT = ffs(CEPH_FILE_MODE_RD);
1010 	const int WR_SHIFT = ffs(CEPH_FILE_MODE_WR);
1011 	const int LAZY_SHIFT = ffs(CEPH_FILE_MODE_LAZY);
1012 	struct ceph_mount_options *opt =
1013 		ceph_inode_to_client(&ci->vfs_inode)->mount_options;
1014 	unsigned long used_cutoff = jiffies - opt->caps_wanted_delay_max * HZ;
1015 	unsigned long idle_cutoff = jiffies - opt->caps_wanted_delay_min * HZ;
1016 
1017 	if (S_ISDIR(ci->vfs_inode.i_mode)) {
1018 		int want = 0;
1019 
1020 		/* use used_cutoff here, to keep dir's wanted caps longer */
1021 		if (ci->i_nr_by_mode[RD_SHIFT] > 0 ||
1022 		    time_after(ci->i_last_rd, used_cutoff))
1023 			want |= CEPH_CAP_ANY_SHARED;
1024 
1025 		if (ci->i_nr_by_mode[WR_SHIFT] > 0 ||
1026 		    time_after(ci->i_last_wr, used_cutoff)) {
1027 			want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1028 			if (opt->flags & CEPH_MOUNT_OPT_ASYNC_DIROPS)
1029 				want |= CEPH_CAP_ANY_DIR_OPS;
1030 		}
1031 
1032 		if (want || ci->i_nr_by_mode[PIN_SHIFT] > 0)
1033 			want |= CEPH_CAP_PIN;
1034 
1035 		return want;
1036 	} else {
1037 		int bits = 0;
1038 
1039 		if (ci->i_nr_by_mode[RD_SHIFT] > 0) {
1040 			if (ci->i_nr_by_mode[RD_SHIFT] >= FMODE_WAIT_BIAS ||
1041 			    time_after(ci->i_last_rd, used_cutoff))
1042 				bits |= 1 << RD_SHIFT;
1043 		} else if (time_after(ci->i_last_rd, idle_cutoff)) {
1044 			bits |= 1 << RD_SHIFT;
1045 		}
1046 
1047 		if (ci->i_nr_by_mode[WR_SHIFT] > 0) {
1048 			if (ci->i_nr_by_mode[WR_SHIFT] >= FMODE_WAIT_BIAS ||
1049 			    time_after(ci->i_last_wr, used_cutoff))
1050 				bits |= 1 << WR_SHIFT;
1051 		} else if (time_after(ci->i_last_wr, idle_cutoff)) {
1052 			bits |= 1 << WR_SHIFT;
1053 		}
1054 
1055 		/* check lazyio only when read/write is wanted */
1056 		if ((bits & (CEPH_FILE_MODE_RDWR << 1)) &&
1057 		    ci->i_nr_by_mode[LAZY_SHIFT] > 0)
1058 			bits |= 1 << LAZY_SHIFT;
1059 
1060 		return bits ? ceph_caps_for_mode(bits >> 1) : 0;
1061 	}
1062 }
1063 
1064 /*
1065  * wanted, by virtue of open file modes AND cap refs (buffered/cached data)
1066  */
1067 int __ceph_caps_wanted(struct ceph_inode_info *ci)
1068 {
1069 	int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci);
1070 	if (S_ISDIR(ci->vfs_inode.i_mode)) {
1071 		/* we want EXCL if holding caps of dir ops */
1072 		if (w & CEPH_CAP_ANY_DIR_OPS)
1073 			w |= CEPH_CAP_FILE_EXCL;
1074 	} else {
1075 		/* we want EXCL if dirty data */
1076 		if (w & CEPH_CAP_FILE_BUFFER)
1077 			w |= CEPH_CAP_FILE_EXCL;
1078 	}
1079 	return w;
1080 }
1081 
1082 /*
1083  * Return caps we have registered with the MDS(s) as 'wanted'.
1084  */
1085 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
1086 {
1087 	struct ceph_cap *cap;
1088 	struct rb_node *p;
1089 	int mds_wanted = 0;
1090 
1091 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1092 		cap = rb_entry(p, struct ceph_cap, ci_node);
1093 		if (check && !__cap_is_valid(cap))
1094 			continue;
1095 		if (cap == ci->i_auth_cap)
1096 			mds_wanted |= cap->mds_wanted;
1097 		else
1098 			mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1099 	}
1100 	return mds_wanted;
1101 }
1102 
1103 int ceph_is_any_caps(struct inode *inode)
1104 {
1105 	struct ceph_inode_info *ci = ceph_inode(inode);
1106 	int ret;
1107 
1108 	spin_lock(&ci->i_ceph_lock);
1109 	ret = __ceph_is_any_real_caps(ci);
1110 	spin_unlock(&ci->i_ceph_lock);
1111 
1112 	return ret;
1113 }
1114 
1115 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
1116 {
1117 	struct ceph_snap_realm *realm = ci->i_snap_realm;
1118 	spin_lock(&realm->inodes_with_caps_lock);
1119 	list_del_init(&ci->i_snap_realm_item);
1120 	ci->i_snap_realm_counter++;
1121 	ci->i_snap_realm = NULL;
1122 	if (realm->ino == ci->i_vino.ino)
1123 		realm->inode = NULL;
1124 	spin_unlock(&realm->inodes_with_caps_lock);
1125 	ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
1126 			    realm);
1127 }
1128 
1129 /*
1130  * Remove a cap.  Take steps to deal with a racing iterate_session_caps.
1131  *
1132  * caller should hold i_ceph_lock.
1133  * caller will not hold session s_mutex if called from destroy_inode.
1134  */
1135 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1136 {
1137 	struct ceph_mds_session *session = cap->session;
1138 	struct ceph_inode_info *ci = cap->ci;
1139 	struct ceph_mds_client *mdsc;
1140 	int removed = 0;
1141 
1142 	/* 'ci' being NULL means the remove have already occurred */
1143 	if (!ci) {
1144 		dout("%s: cap inode is NULL\n", __func__);
1145 		return;
1146 	}
1147 
1148 	dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1149 
1150 	mdsc = ceph_inode_to_client(&ci->vfs_inode)->mdsc;
1151 
1152 	/* remove from inode's cap rbtree, and clear auth cap */
1153 	rb_erase(&cap->ci_node, &ci->i_caps);
1154 	if (ci->i_auth_cap == cap) {
1155 		WARN_ON_ONCE(!list_empty(&ci->i_dirty_item) &&
1156 			     !mdsc->fsc->blocklisted);
1157 		ci->i_auth_cap = NULL;
1158 	}
1159 
1160 	/* remove from session list */
1161 	spin_lock(&session->s_cap_lock);
1162 	if (session->s_cap_iterator == cap) {
1163 		/* not yet, we are iterating over this very cap */
1164 		dout("__ceph_remove_cap  delaying %p removal from session %p\n",
1165 		     cap, cap->session);
1166 	} else {
1167 		list_del_init(&cap->session_caps);
1168 		session->s_nr_caps--;
1169 		atomic64_dec(&mdsc->metric.total_caps);
1170 		cap->session = NULL;
1171 		removed = 1;
1172 	}
1173 	/* protect backpointer with s_cap_lock: see iterate_session_caps */
1174 	cap->ci = NULL;
1175 
1176 	/*
1177 	 * s_cap_reconnect is protected by s_cap_lock. no one changes
1178 	 * s_cap_gen while session is in the reconnect state.
1179 	 */
1180 	if (queue_release &&
1181 	    (!session->s_cap_reconnect ||
1182 	     cap->cap_gen == atomic_read(&session->s_cap_gen))) {
1183 		cap->queue_release = 1;
1184 		if (removed) {
1185 			__ceph_queue_cap_release(session, cap);
1186 			removed = 0;
1187 		}
1188 	} else {
1189 		cap->queue_release = 0;
1190 	}
1191 	cap->cap_ino = ci->i_vino.ino;
1192 
1193 	spin_unlock(&session->s_cap_lock);
1194 
1195 	if (removed)
1196 		ceph_put_cap(mdsc, cap);
1197 
1198 	if (!__ceph_is_any_real_caps(ci)) {
1199 		/* when reconnect denied, we remove session caps forcibly,
1200 		 * i_wr_ref can be non-zero. If there are ongoing write,
1201 		 * keep i_snap_realm.
1202 		 */
1203 		if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1204 			drop_inode_snap_realm(ci);
1205 
1206 		__cap_delay_cancel(mdsc, ci);
1207 	}
1208 }
1209 
1210 struct cap_msg_args {
1211 	struct ceph_mds_session	*session;
1212 	u64			ino, cid, follows;
1213 	u64			flush_tid, oldest_flush_tid, size, max_size;
1214 	u64			xattr_version;
1215 	u64			change_attr;
1216 	struct ceph_buffer	*xattr_buf;
1217 	struct ceph_buffer	*old_xattr_buf;
1218 	struct timespec64	atime, mtime, ctime, btime;
1219 	int			op, caps, wanted, dirty;
1220 	u32			seq, issue_seq, mseq, time_warp_seq;
1221 	u32			flags;
1222 	kuid_t			uid;
1223 	kgid_t			gid;
1224 	umode_t			mode;
1225 	bool			inline_data;
1226 	bool			wake;
1227 };
1228 
1229 /*
1230  * cap struct size + flock buffer size + inline version + inline data size +
1231  * osd_epoch_barrier + oldest_flush_tid
1232  */
1233 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1234 		      4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1235 
1236 /* Marshal up the cap msg to the MDS */
1237 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1238 {
1239 	struct ceph_mds_caps *fc;
1240 	void *p;
1241 	struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1242 
1243 	dout("%s %s %llx %llx caps %s wanted %s dirty %s seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu xattr_ver %llu xattr_len %d\n",
1244 	     __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1245 	     ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1246 	     ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1247 	     arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1248 	     arg->size, arg->max_size, arg->xattr_version,
1249 	     arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1250 
1251 	msg->hdr.version = cpu_to_le16(10);
1252 	msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1253 
1254 	fc = msg->front.iov_base;
1255 	memset(fc, 0, sizeof(*fc));
1256 
1257 	fc->cap_id = cpu_to_le64(arg->cid);
1258 	fc->op = cpu_to_le32(arg->op);
1259 	fc->seq = cpu_to_le32(arg->seq);
1260 	fc->issue_seq = cpu_to_le32(arg->issue_seq);
1261 	fc->migrate_seq = cpu_to_le32(arg->mseq);
1262 	fc->caps = cpu_to_le32(arg->caps);
1263 	fc->wanted = cpu_to_le32(arg->wanted);
1264 	fc->dirty = cpu_to_le32(arg->dirty);
1265 	fc->ino = cpu_to_le64(arg->ino);
1266 	fc->snap_follows = cpu_to_le64(arg->follows);
1267 
1268 	fc->size = cpu_to_le64(arg->size);
1269 	fc->max_size = cpu_to_le64(arg->max_size);
1270 	ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1271 	ceph_encode_timespec64(&fc->atime, &arg->atime);
1272 	ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1273 	fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1274 
1275 	fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1276 	fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1277 	fc->mode = cpu_to_le32(arg->mode);
1278 
1279 	fc->xattr_version = cpu_to_le64(arg->xattr_version);
1280 	if (arg->xattr_buf) {
1281 		msg->middle = ceph_buffer_get(arg->xattr_buf);
1282 		fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1283 		msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1284 	}
1285 
1286 	p = fc + 1;
1287 	/* flock buffer size (version 2) */
1288 	ceph_encode_32(&p, 0);
1289 	/* inline version (version 4) */
1290 	ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1291 	/* inline data size */
1292 	ceph_encode_32(&p, 0);
1293 	/*
1294 	 * osd_epoch_barrier (version 5)
1295 	 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1296 	 * case it was recently changed
1297 	 */
1298 	ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1299 	/* oldest_flush_tid (version 6) */
1300 	ceph_encode_64(&p, arg->oldest_flush_tid);
1301 
1302 	/*
1303 	 * caller_uid/caller_gid (version 7)
1304 	 *
1305 	 * Currently, we don't properly track which caller dirtied the caps
1306 	 * last, and force a flush of them when there is a conflict. For now,
1307 	 * just set this to 0:0, to emulate how the MDS has worked up to now.
1308 	 */
1309 	ceph_encode_32(&p, 0);
1310 	ceph_encode_32(&p, 0);
1311 
1312 	/* pool namespace (version 8) (mds always ignores this) */
1313 	ceph_encode_32(&p, 0);
1314 
1315 	/* btime and change_attr (version 9) */
1316 	ceph_encode_timespec64(p, &arg->btime);
1317 	p += sizeof(struct ceph_timespec);
1318 	ceph_encode_64(&p, arg->change_attr);
1319 
1320 	/* Advisory flags (version 10) */
1321 	ceph_encode_32(&p, arg->flags);
1322 }
1323 
1324 /*
1325  * Queue cap releases when an inode is dropped from our cache.
1326  */
1327 void __ceph_remove_caps(struct ceph_inode_info *ci)
1328 {
1329 	struct rb_node *p;
1330 
1331 	/* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1332 	 * may call __ceph_caps_issued_mask() on a freeing inode. */
1333 	spin_lock(&ci->i_ceph_lock);
1334 	p = rb_first(&ci->i_caps);
1335 	while (p) {
1336 		struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1337 		p = rb_next(p);
1338 		__ceph_remove_cap(cap, true);
1339 	}
1340 	spin_unlock(&ci->i_ceph_lock);
1341 }
1342 
1343 /*
1344  * Prepare to send a cap message to an MDS. Update the cap state, and populate
1345  * the arg struct with the parameters that will need to be sent. This should
1346  * be done under the i_ceph_lock to guard against changes to cap state.
1347  *
1348  * Make note of max_size reported/requested from mds, revoked caps
1349  * that have now been implemented.
1350  */
1351 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1352 		       int op, int flags, int used, int want, int retain,
1353 		       int flushing, u64 flush_tid, u64 oldest_flush_tid)
1354 {
1355 	struct ceph_inode_info *ci = cap->ci;
1356 	struct inode *inode = &ci->vfs_inode;
1357 	int held, revoking;
1358 
1359 	lockdep_assert_held(&ci->i_ceph_lock);
1360 
1361 	held = cap->issued | cap->implemented;
1362 	revoking = cap->implemented & ~cap->issued;
1363 	retain &= ~revoking;
1364 
1365 	dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1366 	     __func__, inode, cap, cap->session,
1367 	     ceph_cap_string(held), ceph_cap_string(held & retain),
1368 	     ceph_cap_string(revoking));
1369 	BUG_ON((retain & CEPH_CAP_PIN) == 0);
1370 
1371 	ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1372 
1373 	cap->issued &= retain;  /* drop bits we don't want */
1374 	/*
1375 	 * Wake up any waiters on wanted -> needed transition. This is due to
1376 	 * the weird transition from buffered to sync IO... we need to flush
1377 	 * dirty pages _before_ allowing sync writes to avoid reordering.
1378 	 */
1379 	arg->wake = cap->implemented & ~cap->issued;
1380 	cap->implemented &= cap->issued | used;
1381 	cap->mds_wanted = want;
1382 
1383 	arg->session = cap->session;
1384 	arg->ino = ceph_vino(inode).ino;
1385 	arg->cid = cap->cap_id;
1386 	arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1387 	arg->flush_tid = flush_tid;
1388 	arg->oldest_flush_tid = oldest_flush_tid;
1389 
1390 	arg->size = i_size_read(inode);
1391 	ci->i_reported_size = arg->size;
1392 	arg->max_size = ci->i_wanted_max_size;
1393 	if (cap == ci->i_auth_cap) {
1394 		if (want & CEPH_CAP_ANY_FILE_WR)
1395 			ci->i_requested_max_size = arg->max_size;
1396 		else
1397 			ci->i_requested_max_size = 0;
1398 	}
1399 
1400 	if (flushing & CEPH_CAP_XATTR_EXCL) {
1401 		arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1402 		arg->xattr_version = ci->i_xattrs.version;
1403 		arg->xattr_buf = ci->i_xattrs.blob;
1404 	} else {
1405 		arg->xattr_buf = NULL;
1406 		arg->old_xattr_buf = NULL;
1407 	}
1408 
1409 	arg->mtime = inode->i_mtime;
1410 	arg->atime = inode->i_atime;
1411 	arg->ctime = inode->i_ctime;
1412 	arg->btime = ci->i_btime;
1413 	arg->change_attr = inode_peek_iversion_raw(inode);
1414 
1415 	arg->op = op;
1416 	arg->caps = cap->implemented;
1417 	arg->wanted = want;
1418 	arg->dirty = flushing;
1419 
1420 	arg->seq = cap->seq;
1421 	arg->issue_seq = cap->issue_seq;
1422 	arg->mseq = cap->mseq;
1423 	arg->time_warp_seq = ci->i_time_warp_seq;
1424 
1425 	arg->uid = inode->i_uid;
1426 	arg->gid = inode->i_gid;
1427 	arg->mode = inode->i_mode;
1428 
1429 	arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1430 	if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1431 	    !list_empty(&ci->i_cap_snaps)) {
1432 		struct ceph_cap_snap *capsnap;
1433 		list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1434 			if (capsnap->cap_flush.tid)
1435 				break;
1436 			if (capsnap->need_flush) {
1437 				flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1438 				break;
1439 			}
1440 		}
1441 	}
1442 	arg->flags = flags;
1443 }
1444 
1445 /*
1446  * Send a cap msg on the given inode.
1447  *
1448  * Caller should hold snap_rwsem (read), s_mutex.
1449  */
1450 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1451 {
1452 	struct ceph_msg *msg;
1453 	struct inode *inode = &ci->vfs_inode;
1454 
1455 	msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1456 	if (!msg) {
1457 		pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1458 		       ceph_vinop(inode), ceph_cap_string(arg->dirty),
1459 		       arg->flush_tid);
1460 		spin_lock(&ci->i_ceph_lock);
1461 		__cap_delay_requeue(arg->session->s_mdsc, ci);
1462 		spin_unlock(&ci->i_ceph_lock);
1463 		return;
1464 	}
1465 
1466 	encode_cap_msg(msg, arg);
1467 	ceph_con_send(&arg->session->s_con, msg);
1468 	ceph_buffer_put(arg->old_xattr_buf);
1469 	if (arg->wake)
1470 		wake_up_all(&ci->i_cap_wq);
1471 }
1472 
1473 static inline int __send_flush_snap(struct inode *inode,
1474 				    struct ceph_mds_session *session,
1475 				    struct ceph_cap_snap *capsnap,
1476 				    u32 mseq, u64 oldest_flush_tid)
1477 {
1478 	struct cap_msg_args	arg;
1479 	struct ceph_msg		*msg;
1480 
1481 	msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1482 	if (!msg)
1483 		return -ENOMEM;
1484 
1485 	arg.session = session;
1486 	arg.ino = ceph_vino(inode).ino;
1487 	arg.cid = 0;
1488 	arg.follows = capsnap->follows;
1489 	arg.flush_tid = capsnap->cap_flush.tid;
1490 	arg.oldest_flush_tid = oldest_flush_tid;
1491 
1492 	arg.size = capsnap->size;
1493 	arg.max_size = 0;
1494 	arg.xattr_version = capsnap->xattr_version;
1495 	arg.xattr_buf = capsnap->xattr_blob;
1496 	arg.old_xattr_buf = NULL;
1497 
1498 	arg.atime = capsnap->atime;
1499 	arg.mtime = capsnap->mtime;
1500 	arg.ctime = capsnap->ctime;
1501 	arg.btime = capsnap->btime;
1502 	arg.change_attr = capsnap->change_attr;
1503 
1504 	arg.op = CEPH_CAP_OP_FLUSHSNAP;
1505 	arg.caps = capsnap->issued;
1506 	arg.wanted = 0;
1507 	arg.dirty = capsnap->dirty;
1508 
1509 	arg.seq = 0;
1510 	arg.issue_seq = 0;
1511 	arg.mseq = mseq;
1512 	arg.time_warp_seq = capsnap->time_warp_seq;
1513 
1514 	arg.uid = capsnap->uid;
1515 	arg.gid = capsnap->gid;
1516 	arg.mode = capsnap->mode;
1517 
1518 	arg.inline_data = capsnap->inline_data;
1519 	arg.flags = 0;
1520 	arg.wake = false;
1521 
1522 	encode_cap_msg(msg, &arg);
1523 	ceph_con_send(&arg.session->s_con, msg);
1524 	return 0;
1525 }
1526 
1527 /*
1528  * When a snapshot is taken, clients accumulate dirty metadata on
1529  * inodes with capabilities in ceph_cap_snaps to describe the file
1530  * state at the time the snapshot was taken.  This must be flushed
1531  * asynchronously back to the MDS once sync writes complete and dirty
1532  * data is written out.
1533  *
1534  * Called under i_ceph_lock.
1535  */
1536 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1537 			       struct ceph_mds_session *session)
1538 		__releases(ci->i_ceph_lock)
1539 		__acquires(ci->i_ceph_lock)
1540 {
1541 	struct inode *inode = &ci->vfs_inode;
1542 	struct ceph_mds_client *mdsc = session->s_mdsc;
1543 	struct ceph_cap_snap *capsnap;
1544 	u64 oldest_flush_tid = 0;
1545 	u64 first_tid = 1, last_tid = 0;
1546 
1547 	dout("__flush_snaps %p session %p\n", inode, session);
1548 
1549 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1550 		/*
1551 		 * we need to wait for sync writes to complete and for dirty
1552 		 * pages to be written out.
1553 		 */
1554 		if (capsnap->dirty_pages || capsnap->writing)
1555 			break;
1556 
1557 		/* should be removed by ceph_try_drop_cap_snap() */
1558 		BUG_ON(!capsnap->need_flush);
1559 
1560 		/* only flush each capsnap once */
1561 		if (capsnap->cap_flush.tid > 0) {
1562 			dout(" already flushed %p, skipping\n", capsnap);
1563 			continue;
1564 		}
1565 
1566 		spin_lock(&mdsc->cap_dirty_lock);
1567 		capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1568 		list_add_tail(&capsnap->cap_flush.g_list,
1569 			      &mdsc->cap_flush_list);
1570 		if (oldest_flush_tid == 0)
1571 			oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1572 		if (list_empty(&ci->i_flushing_item)) {
1573 			list_add_tail(&ci->i_flushing_item,
1574 				      &session->s_cap_flushing);
1575 		}
1576 		spin_unlock(&mdsc->cap_dirty_lock);
1577 
1578 		list_add_tail(&capsnap->cap_flush.i_list,
1579 			      &ci->i_cap_flush_list);
1580 
1581 		if (first_tid == 1)
1582 			first_tid = capsnap->cap_flush.tid;
1583 		last_tid = capsnap->cap_flush.tid;
1584 	}
1585 
1586 	ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1587 
1588 	while (first_tid <= last_tid) {
1589 		struct ceph_cap *cap = ci->i_auth_cap;
1590 		struct ceph_cap_flush *cf;
1591 		int ret;
1592 
1593 		if (!(cap && cap->session == session)) {
1594 			dout("__flush_snaps %p auth cap %p not mds%d, "
1595 			     "stop\n", inode, cap, session->s_mds);
1596 			break;
1597 		}
1598 
1599 		ret = -ENOENT;
1600 		list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1601 			if (cf->tid >= first_tid) {
1602 				ret = 0;
1603 				break;
1604 			}
1605 		}
1606 		if (ret < 0)
1607 			break;
1608 
1609 		first_tid = cf->tid + 1;
1610 
1611 		capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1612 		refcount_inc(&capsnap->nref);
1613 		spin_unlock(&ci->i_ceph_lock);
1614 
1615 		dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1616 		     inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1617 
1618 		ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1619 					oldest_flush_tid);
1620 		if (ret < 0) {
1621 			pr_err("__flush_snaps: error sending cap flushsnap, "
1622 			       "ino (%llx.%llx) tid %llu follows %llu\n",
1623 				ceph_vinop(inode), cf->tid, capsnap->follows);
1624 		}
1625 
1626 		ceph_put_cap_snap(capsnap);
1627 		spin_lock(&ci->i_ceph_lock);
1628 	}
1629 }
1630 
1631 void ceph_flush_snaps(struct ceph_inode_info *ci,
1632 		      struct ceph_mds_session **psession)
1633 {
1634 	struct inode *inode = &ci->vfs_inode;
1635 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1636 	struct ceph_mds_session *session = NULL;
1637 	int mds;
1638 
1639 	dout("ceph_flush_snaps %p\n", inode);
1640 	if (psession)
1641 		session = *psession;
1642 retry:
1643 	spin_lock(&ci->i_ceph_lock);
1644 	if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1645 		dout(" no capsnap needs flush, doing nothing\n");
1646 		goto out;
1647 	}
1648 	if (!ci->i_auth_cap) {
1649 		dout(" no auth cap (migrating?), doing nothing\n");
1650 		goto out;
1651 	}
1652 
1653 	mds = ci->i_auth_cap->session->s_mds;
1654 	if (session && session->s_mds != mds) {
1655 		dout(" oops, wrong session %p mutex\n", session);
1656 		ceph_put_mds_session(session);
1657 		session = NULL;
1658 	}
1659 	if (!session) {
1660 		spin_unlock(&ci->i_ceph_lock);
1661 		mutex_lock(&mdsc->mutex);
1662 		session = __ceph_lookup_mds_session(mdsc, mds);
1663 		mutex_unlock(&mdsc->mutex);
1664 		goto retry;
1665 	}
1666 
1667 	// make sure flushsnap messages are sent in proper order.
1668 	if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1669 		__kick_flushing_caps(mdsc, session, ci, 0);
1670 
1671 	__ceph_flush_snaps(ci, session);
1672 out:
1673 	spin_unlock(&ci->i_ceph_lock);
1674 
1675 	if (psession)
1676 		*psession = session;
1677 	else
1678 		ceph_put_mds_session(session);
1679 	/* we flushed them all; remove this inode from the queue */
1680 	spin_lock(&mdsc->snap_flush_lock);
1681 	list_del_init(&ci->i_snap_flush_item);
1682 	spin_unlock(&mdsc->snap_flush_lock);
1683 }
1684 
1685 /*
1686  * Mark caps dirty.  If inode is newly dirty, return the dirty flags.
1687  * Caller is then responsible for calling __mark_inode_dirty with the
1688  * returned flags value.
1689  */
1690 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1691 			   struct ceph_cap_flush **pcf)
1692 {
1693 	struct ceph_mds_client *mdsc =
1694 		ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1695 	struct inode *inode = &ci->vfs_inode;
1696 	int was = ci->i_dirty_caps;
1697 	int dirty = 0;
1698 
1699 	lockdep_assert_held(&ci->i_ceph_lock);
1700 
1701 	if (!ci->i_auth_cap) {
1702 		pr_warn("__mark_dirty_caps %p %llx mask %s, "
1703 			"but no auth cap (session was closed?)\n",
1704 			inode, ceph_ino(inode), ceph_cap_string(mask));
1705 		return 0;
1706 	}
1707 
1708 	dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1709 	     ceph_cap_string(mask), ceph_cap_string(was),
1710 	     ceph_cap_string(was | mask));
1711 	ci->i_dirty_caps |= mask;
1712 	if (was == 0) {
1713 		struct ceph_mds_session *session = ci->i_auth_cap->session;
1714 
1715 		WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1716 		swap(ci->i_prealloc_cap_flush, *pcf);
1717 
1718 		if (!ci->i_head_snapc) {
1719 			WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1720 			ci->i_head_snapc = ceph_get_snap_context(
1721 				ci->i_snap_realm->cached_context);
1722 		}
1723 		dout(" inode %p now dirty snapc %p auth cap %p\n",
1724 		     &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1725 		BUG_ON(!list_empty(&ci->i_dirty_item));
1726 		spin_lock(&mdsc->cap_dirty_lock);
1727 		list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1728 		spin_unlock(&mdsc->cap_dirty_lock);
1729 		if (ci->i_flushing_caps == 0) {
1730 			ihold(inode);
1731 			dirty |= I_DIRTY_SYNC;
1732 		}
1733 	} else {
1734 		WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1735 	}
1736 	BUG_ON(list_empty(&ci->i_dirty_item));
1737 	if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1738 	    (mask & CEPH_CAP_FILE_BUFFER))
1739 		dirty |= I_DIRTY_DATASYNC;
1740 	__cap_delay_requeue(mdsc, ci);
1741 	return dirty;
1742 }
1743 
1744 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1745 {
1746 	struct ceph_cap_flush *cf;
1747 
1748 	cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1749 	cf->is_capsnap = false;
1750 	return cf;
1751 }
1752 
1753 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1754 {
1755 	if (cf)
1756 		kmem_cache_free(ceph_cap_flush_cachep, cf);
1757 }
1758 
1759 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1760 {
1761 	if (!list_empty(&mdsc->cap_flush_list)) {
1762 		struct ceph_cap_flush *cf =
1763 			list_first_entry(&mdsc->cap_flush_list,
1764 					 struct ceph_cap_flush, g_list);
1765 		return cf->tid;
1766 	}
1767 	return 0;
1768 }
1769 
1770 /*
1771  * Remove cap_flush from the mdsc's or inode's flushing cap list.
1772  * Return true if caller needs to wake up flush waiters.
1773  */
1774 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1775 					 struct ceph_cap_flush *cf)
1776 {
1777 	struct ceph_cap_flush *prev;
1778 	bool wake = cf->wake;
1779 
1780 	if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1781 		prev = list_prev_entry(cf, g_list);
1782 		prev->wake = true;
1783 		wake = false;
1784 	}
1785 	list_del_init(&cf->g_list);
1786 	return wake;
1787 }
1788 
1789 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1790 				       struct ceph_cap_flush *cf)
1791 {
1792 	struct ceph_cap_flush *prev;
1793 	bool wake = cf->wake;
1794 
1795 	if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1796 		prev = list_prev_entry(cf, i_list);
1797 		prev->wake = true;
1798 		wake = false;
1799 	}
1800 	list_del_init(&cf->i_list);
1801 	return wake;
1802 }
1803 
1804 /*
1805  * Add dirty inode to the flushing list.  Assigned a seq number so we
1806  * can wait for caps to flush without starving.
1807  *
1808  * Called under i_ceph_lock. Returns the flush tid.
1809  */
1810 static u64 __mark_caps_flushing(struct inode *inode,
1811 				struct ceph_mds_session *session, bool wake,
1812 				u64 *oldest_flush_tid)
1813 {
1814 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1815 	struct ceph_inode_info *ci = ceph_inode(inode);
1816 	struct ceph_cap_flush *cf = NULL;
1817 	int flushing;
1818 
1819 	lockdep_assert_held(&ci->i_ceph_lock);
1820 	BUG_ON(ci->i_dirty_caps == 0);
1821 	BUG_ON(list_empty(&ci->i_dirty_item));
1822 	BUG_ON(!ci->i_prealloc_cap_flush);
1823 
1824 	flushing = ci->i_dirty_caps;
1825 	dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1826 	     ceph_cap_string(flushing),
1827 	     ceph_cap_string(ci->i_flushing_caps),
1828 	     ceph_cap_string(ci->i_flushing_caps | flushing));
1829 	ci->i_flushing_caps |= flushing;
1830 	ci->i_dirty_caps = 0;
1831 	dout(" inode %p now !dirty\n", inode);
1832 
1833 	swap(cf, ci->i_prealloc_cap_flush);
1834 	cf->caps = flushing;
1835 	cf->wake = wake;
1836 
1837 	spin_lock(&mdsc->cap_dirty_lock);
1838 	list_del_init(&ci->i_dirty_item);
1839 
1840 	cf->tid = ++mdsc->last_cap_flush_tid;
1841 	list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1842 	*oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1843 
1844 	if (list_empty(&ci->i_flushing_item)) {
1845 		list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1846 		mdsc->num_cap_flushing++;
1847 	}
1848 	spin_unlock(&mdsc->cap_dirty_lock);
1849 
1850 	list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1851 
1852 	return cf->tid;
1853 }
1854 
1855 /*
1856  * try to invalidate mapping pages without blocking.
1857  */
1858 static int try_nonblocking_invalidate(struct inode *inode)
1859 {
1860 	struct ceph_inode_info *ci = ceph_inode(inode);
1861 	u32 invalidating_gen = ci->i_rdcache_gen;
1862 
1863 	spin_unlock(&ci->i_ceph_lock);
1864 	ceph_fscache_invalidate(inode);
1865 	invalidate_mapping_pages(&inode->i_data, 0, -1);
1866 	spin_lock(&ci->i_ceph_lock);
1867 
1868 	if (inode->i_data.nrpages == 0 &&
1869 	    invalidating_gen == ci->i_rdcache_gen) {
1870 		/* success. */
1871 		dout("try_nonblocking_invalidate %p success\n", inode);
1872 		/* save any racing async invalidate some trouble */
1873 		ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1874 		return 0;
1875 	}
1876 	dout("try_nonblocking_invalidate %p failed\n", inode);
1877 	return -1;
1878 }
1879 
1880 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1881 {
1882 	loff_t size = i_size_read(&ci->vfs_inode);
1883 	/* mds will adjust max size according to the reported size */
1884 	if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1885 		return false;
1886 	if (size >= ci->i_max_size)
1887 		return true;
1888 	/* half of previous max_size increment has been used */
1889 	if (ci->i_max_size > ci->i_reported_size &&
1890 	    (size << 1) >= ci->i_max_size + ci->i_reported_size)
1891 		return true;
1892 	return false;
1893 }
1894 
1895 /*
1896  * Swiss army knife function to examine currently used and wanted
1897  * versus held caps.  Release, flush, ack revoked caps to mds as
1898  * appropriate.
1899  *
1900  *  CHECK_CAPS_AUTHONLY - we should only check the auth cap
1901  *  CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1902  *    further delay.
1903  */
1904 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1905 		     struct ceph_mds_session *session)
1906 {
1907 	struct inode *inode = &ci->vfs_inode;
1908 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1909 	struct ceph_cap *cap;
1910 	u64 flush_tid, oldest_flush_tid;
1911 	int file_wanted, used, cap_used;
1912 	int issued, implemented, want, retain, revoking, flushing = 0;
1913 	int mds = -1;   /* keep track of how far we've gone through i_caps list
1914 			   to avoid an infinite loop on retry */
1915 	struct rb_node *p;
1916 	bool queue_invalidate = false;
1917 	bool tried_invalidate = false;
1918 
1919 	if (session)
1920 		ceph_get_mds_session(session);
1921 
1922 	spin_lock(&ci->i_ceph_lock);
1923 	if (ci->i_ceph_flags & CEPH_I_FLUSH)
1924 		flags |= CHECK_CAPS_FLUSH;
1925 retry:
1926 	/* Caps wanted by virtue of active open files. */
1927 	file_wanted = __ceph_caps_file_wanted(ci);
1928 
1929 	/* Caps which have active references against them */
1930 	used = __ceph_caps_used(ci);
1931 
1932 	/*
1933 	 * "issued" represents the current caps that the MDS wants us to have.
1934 	 * "implemented" is the set that we have been granted, and includes the
1935 	 * ones that have not yet been returned to the MDS (the "revoking" set,
1936 	 * usually because they have outstanding references).
1937 	 */
1938 	issued = __ceph_caps_issued(ci, &implemented);
1939 	revoking = implemented & ~issued;
1940 
1941 	want = file_wanted;
1942 
1943 	/* The ones we currently want to retain (may be adjusted below) */
1944 	retain = file_wanted | used | CEPH_CAP_PIN;
1945 	if (!mdsc->stopping && inode->i_nlink > 0) {
1946 		if (file_wanted) {
1947 			retain |= CEPH_CAP_ANY;       /* be greedy */
1948 		} else if (S_ISDIR(inode->i_mode) &&
1949 			   (issued & CEPH_CAP_FILE_SHARED) &&
1950 			   __ceph_dir_is_complete(ci)) {
1951 			/*
1952 			 * If a directory is complete, we want to keep
1953 			 * the exclusive cap. So that MDS does not end up
1954 			 * revoking the shared cap on every create/unlink
1955 			 * operation.
1956 			 */
1957 			if (IS_RDONLY(inode)) {
1958 				want = CEPH_CAP_ANY_SHARED;
1959 			} else {
1960 				want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1961 			}
1962 			retain |= want;
1963 		} else {
1964 
1965 			retain |= CEPH_CAP_ANY_SHARED;
1966 			/*
1967 			 * keep RD only if we didn't have the file open RW,
1968 			 * because then the mds would revoke it anyway to
1969 			 * journal max_size=0.
1970 			 */
1971 			if (ci->i_max_size == 0)
1972 				retain |= CEPH_CAP_ANY_RD;
1973 		}
1974 	}
1975 
1976 	dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1977 	     " issued %s revoking %s retain %s %s%s\n", inode,
1978 	     ceph_cap_string(file_wanted),
1979 	     ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1980 	     ceph_cap_string(ci->i_flushing_caps),
1981 	     ceph_cap_string(issued), ceph_cap_string(revoking),
1982 	     ceph_cap_string(retain),
1983 	     (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1984 	     (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1985 
1986 	/*
1987 	 * If we no longer need to hold onto old our caps, and we may
1988 	 * have cached pages, but don't want them, then try to invalidate.
1989 	 * If we fail, it's because pages are locked.... try again later.
1990 	 */
1991 	if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
1992 	    S_ISREG(inode->i_mode) &&
1993 	    !(ci->i_wb_ref || ci->i_wrbuffer_ref) &&   /* no dirty pages... */
1994 	    inode->i_data.nrpages &&		/* have cached pages */
1995 	    (revoking & (CEPH_CAP_FILE_CACHE|
1996 			 CEPH_CAP_FILE_LAZYIO)) && /*  or revoking cache */
1997 	    !tried_invalidate) {
1998 		dout("check_caps trying to invalidate on %p\n", inode);
1999 		if (try_nonblocking_invalidate(inode) < 0) {
2000 			dout("check_caps queuing invalidate\n");
2001 			queue_invalidate = true;
2002 			ci->i_rdcache_revoking = ci->i_rdcache_gen;
2003 		}
2004 		tried_invalidate = true;
2005 		goto retry;
2006 	}
2007 
2008 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2009 		int mflags = 0;
2010 		struct cap_msg_args arg;
2011 
2012 		cap = rb_entry(p, struct ceph_cap, ci_node);
2013 
2014 		/* avoid looping forever */
2015 		if (mds >= cap->mds ||
2016 		    ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2017 			continue;
2018 
2019 		/*
2020 		 * If we have an auth cap, we don't need to consider any
2021 		 * overlapping caps as used.
2022 		 */
2023 		cap_used = used;
2024 		if (ci->i_auth_cap && cap != ci->i_auth_cap)
2025 			cap_used &= ~ci->i_auth_cap->issued;
2026 
2027 		revoking = cap->implemented & ~cap->issued;
2028 		dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2029 		     cap->mds, cap, ceph_cap_string(cap_used),
2030 		     ceph_cap_string(cap->issued),
2031 		     ceph_cap_string(cap->implemented),
2032 		     ceph_cap_string(revoking));
2033 
2034 		if (cap == ci->i_auth_cap &&
2035 		    (cap->issued & CEPH_CAP_FILE_WR)) {
2036 			/* request larger max_size from MDS? */
2037 			if (ci->i_wanted_max_size > ci->i_max_size &&
2038 			    ci->i_wanted_max_size > ci->i_requested_max_size) {
2039 				dout("requesting new max_size\n");
2040 				goto ack;
2041 			}
2042 
2043 			/* approaching file_max? */
2044 			if (__ceph_should_report_size(ci)) {
2045 				dout("i_size approaching max_size\n");
2046 				goto ack;
2047 			}
2048 		}
2049 		/* flush anything dirty? */
2050 		if (cap == ci->i_auth_cap) {
2051 			if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2052 				dout("flushing dirty caps\n");
2053 				goto ack;
2054 			}
2055 			if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2056 				dout("flushing snap caps\n");
2057 				goto ack;
2058 			}
2059 		}
2060 
2061 		/* completed revocation? going down and there are no caps? */
2062 		if (revoking && (revoking & cap_used) == 0) {
2063 			dout("completed revocation of %s\n",
2064 			     ceph_cap_string(cap->implemented & ~cap->issued));
2065 			goto ack;
2066 		}
2067 
2068 		/* want more caps from mds? */
2069 		if (want & ~cap->mds_wanted) {
2070 			if (want & ~(cap->mds_wanted | cap->issued))
2071 				goto ack;
2072 			if (!__cap_is_valid(cap))
2073 				goto ack;
2074 		}
2075 
2076 		/* things we might delay */
2077 		if ((cap->issued & ~retain) == 0)
2078 			continue;     /* nope, all good */
2079 
2080 ack:
2081 		ceph_put_mds_session(session);
2082 		session = ceph_get_mds_session(cap->session);
2083 
2084 		/* kick flushing and flush snaps before sending normal
2085 		 * cap message */
2086 		if (cap == ci->i_auth_cap &&
2087 		    (ci->i_ceph_flags &
2088 		     (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2089 			if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2090 				__kick_flushing_caps(mdsc, session, ci, 0);
2091 			if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2092 				__ceph_flush_snaps(ci, session);
2093 
2094 			goto retry;
2095 		}
2096 
2097 		if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2098 			flushing = ci->i_dirty_caps;
2099 			flush_tid = __mark_caps_flushing(inode, session, false,
2100 							 &oldest_flush_tid);
2101 			if (flags & CHECK_CAPS_FLUSH &&
2102 			    list_empty(&session->s_cap_dirty))
2103 				mflags |= CEPH_CLIENT_CAPS_SYNC;
2104 		} else {
2105 			flushing = 0;
2106 			flush_tid = 0;
2107 			spin_lock(&mdsc->cap_dirty_lock);
2108 			oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2109 			spin_unlock(&mdsc->cap_dirty_lock);
2110 		}
2111 
2112 		mds = cap->mds;  /* remember mds, so we don't repeat */
2113 
2114 		__prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2115 			   want, retain, flushing, flush_tid, oldest_flush_tid);
2116 
2117 		spin_unlock(&ci->i_ceph_lock);
2118 		__send_cap(&arg, ci);
2119 		spin_lock(&ci->i_ceph_lock);
2120 
2121 		goto retry; /* retake i_ceph_lock and restart our cap scan. */
2122 	}
2123 
2124 	/* periodically re-calculate caps wanted by open files */
2125 	if (__ceph_is_any_real_caps(ci) &&
2126 	    list_empty(&ci->i_cap_delay_list) &&
2127 	    (file_wanted & ~CEPH_CAP_PIN) &&
2128 	    !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2129 		__cap_delay_requeue(mdsc, ci);
2130 	}
2131 
2132 	spin_unlock(&ci->i_ceph_lock);
2133 
2134 	ceph_put_mds_session(session);
2135 	if (queue_invalidate)
2136 		ceph_queue_invalidate(inode);
2137 }
2138 
2139 /*
2140  * Try to flush dirty caps back to the auth mds.
2141  */
2142 static int try_flush_caps(struct inode *inode, u64 *ptid)
2143 {
2144 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2145 	struct ceph_inode_info *ci = ceph_inode(inode);
2146 	int flushing = 0;
2147 	u64 flush_tid = 0, oldest_flush_tid = 0;
2148 
2149 	spin_lock(&ci->i_ceph_lock);
2150 retry_locked:
2151 	if (ci->i_dirty_caps && ci->i_auth_cap) {
2152 		struct ceph_cap *cap = ci->i_auth_cap;
2153 		struct cap_msg_args arg;
2154 		struct ceph_mds_session *session = cap->session;
2155 
2156 		if (session->s_state < CEPH_MDS_SESSION_OPEN) {
2157 			spin_unlock(&ci->i_ceph_lock);
2158 			goto out;
2159 		}
2160 
2161 		if (ci->i_ceph_flags &
2162 		    (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2163 			if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2164 				__kick_flushing_caps(mdsc, session, ci, 0);
2165 			if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2166 				__ceph_flush_snaps(ci, session);
2167 			goto retry_locked;
2168 		}
2169 
2170 		flushing = ci->i_dirty_caps;
2171 		flush_tid = __mark_caps_flushing(inode, session, true,
2172 						 &oldest_flush_tid);
2173 
2174 		__prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2175 			   __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2176 			   (cap->issued | cap->implemented),
2177 			   flushing, flush_tid, oldest_flush_tid);
2178 		spin_unlock(&ci->i_ceph_lock);
2179 
2180 		__send_cap(&arg, ci);
2181 	} else {
2182 		if (!list_empty(&ci->i_cap_flush_list)) {
2183 			struct ceph_cap_flush *cf =
2184 				list_last_entry(&ci->i_cap_flush_list,
2185 						struct ceph_cap_flush, i_list);
2186 			cf->wake = true;
2187 			flush_tid = cf->tid;
2188 		}
2189 		flushing = ci->i_flushing_caps;
2190 		spin_unlock(&ci->i_ceph_lock);
2191 	}
2192 out:
2193 	*ptid = flush_tid;
2194 	return flushing;
2195 }
2196 
2197 /*
2198  * Return true if we've flushed caps through the given flush_tid.
2199  */
2200 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2201 {
2202 	struct ceph_inode_info *ci = ceph_inode(inode);
2203 	int ret = 1;
2204 
2205 	spin_lock(&ci->i_ceph_lock);
2206 	if (!list_empty(&ci->i_cap_flush_list)) {
2207 		struct ceph_cap_flush * cf =
2208 			list_first_entry(&ci->i_cap_flush_list,
2209 					 struct ceph_cap_flush, i_list);
2210 		if (cf->tid <= flush_tid)
2211 			ret = 0;
2212 	}
2213 	spin_unlock(&ci->i_ceph_lock);
2214 	return ret;
2215 }
2216 
2217 /*
2218  * wait for any unsafe requests to complete.
2219  */
2220 static int unsafe_request_wait(struct inode *inode)
2221 {
2222 	struct ceph_inode_info *ci = ceph_inode(inode);
2223 	struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2224 	int ret, err = 0;
2225 
2226 	spin_lock(&ci->i_unsafe_lock);
2227 	if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2228 		req1 = list_last_entry(&ci->i_unsafe_dirops,
2229 					struct ceph_mds_request,
2230 					r_unsafe_dir_item);
2231 		ceph_mdsc_get_request(req1);
2232 	}
2233 	if (!list_empty(&ci->i_unsafe_iops)) {
2234 		req2 = list_last_entry(&ci->i_unsafe_iops,
2235 					struct ceph_mds_request,
2236 					r_unsafe_target_item);
2237 		ceph_mdsc_get_request(req2);
2238 	}
2239 	spin_unlock(&ci->i_unsafe_lock);
2240 
2241 	dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2242 	     inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2243 	if (req1) {
2244 		ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2245 					ceph_timeout_jiffies(req1->r_timeout));
2246 		if (ret)
2247 			err = -EIO;
2248 		ceph_mdsc_put_request(req1);
2249 	}
2250 	if (req2) {
2251 		ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2252 					ceph_timeout_jiffies(req2->r_timeout));
2253 		if (ret)
2254 			err = -EIO;
2255 		ceph_mdsc_put_request(req2);
2256 	}
2257 	return err;
2258 }
2259 
2260 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2261 {
2262 	struct ceph_file_info *fi = file->private_data;
2263 	struct inode *inode = file->f_mapping->host;
2264 	struct ceph_inode_info *ci = ceph_inode(inode);
2265 	u64 flush_tid;
2266 	int ret, err;
2267 	int dirty;
2268 
2269 	dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2270 
2271 	ret = file_write_and_wait_range(file, start, end);
2272 	if (datasync)
2273 		goto out;
2274 
2275 	ret = ceph_wait_on_async_create(inode);
2276 	if (ret)
2277 		goto out;
2278 
2279 	dirty = try_flush_caps(inode, &flush_tid);
2280 	dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2281 
2282 	err = unsafe_request_wait(inode);
2283 
2284 	/*
2285 	 * only wait on non-file metadata writeback (the mds
2286 	 * can recover size and mtime, so we don't need to
2287 	 * wait for that)
2288 	 */
2289 	if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2290 		err = wait_event_interruptible(ci->i_cap_wq,
2291 					caps_are_flushed(inode, flush_tid));
2292 	}
2293 
2294 	if (err < 0)
2295 		ret = err;
2296 
2297 	if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) {
2298 		spin_lock(&file->f_lock);
2299 		err = errseq_check_and_advance(&ci->i_meta_err,
2300 					       &fi->meta_err);
2301 		spin_unlock(&file->f_lock);
2302 		if (err < 0)
2303 			ret = err;
2304 	}
2305 out:
2306 	dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2307 	return ret;
2308 }
2309 
2310 /*
2311  * Flush any dirty caps back to the mds.  If we aren't asked to wait,
2312  * queue inode for flush but don't do so immediately, because we can
2313  * get by with fewer MDS messages if we wait for data writeback to
2314  * complete first.
2315  */
2316 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2317 {
2318 	struct ceph_inode_info *ci = ceph_inode(inode);
2319 	u64 flush_tid;
2320 	int err = 0;
2321 	int dirty;
2322 	int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2323 
2324 	dout("write_inode %p wait=%d\n", inode, wait);
2325 	if (wait) {
2326 		dirty = try_flush_caps(inode, &flush_tid);
2327 		if (dirty)
2328 			err = wait_event_interruptible(ci->i_cap_wq,
2329 				       caps_are_flushed(inode, flush_tid));
2330 	} else {
2331 		struct ceph_mds_client *mdsc =
2332 			ceph_sb_to_client(inode->i_sb)->mdsc;
2333 
2334 		spin_lock(&ci->i_ceph_lock);
2335 		if (__ceph_caps_dirty(ci))
2336 			__cap_delay_requeue_front(mdsc, ci);
2337 		spin_unlock(&ci->i_ceph_lock);
2338 	}
2339 	return err;
2340 }
2341 
2342 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2343 				 struct ceph_mds_session *session,
2344 				 struct ceph_inode_info *ci,
2345 				 u64 oldest_flush_tid)
2346 	__releases(ci->i_ceph_lock)
2347 	__acquires(ci->i_ceph_lock)
2348 {
2349 	struct inode *inode = &ci->vfs_inode;
2350 	struct ceph_cap *cap;
2351 	struct ceph_cap_flush *cf;
2352 	int ret;
2353 	u64 first_tid = 0;
2354 	u64 last_snap_flush = 0;
2355 
2356 	ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2357 
2358 	list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2359 		if (cf->is_capsnap) {
2360 			last_snap_flush = cf->tid;
2361 			break;
2362 		}
2363 	}
2364 
2365 	list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2366 		if (cf->tid < first_tid)
2367 			continue;
2368 
2369 		cap = ci->i_auth_cap;
2370 		if (!(cap && cap->session == session)) {
2371 			pr_err("%p auth cap %p not mds%d ???\n",
2372 			       inode, cap, session->s_mds);
2373 			break;
2374 		}
2375 
2376 		first_tid = cf->tid + 1;
2377 
2378 		if (!cf->is_capsnap) {
2379 			struct cap_msg_args arg;
2380 
2381 			dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2382 			     inode, cap, cf->tid, ceph_cap_string(cf->caps));
2383 			__prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2384 					 (cf->tid < last_snap_flush ?
2385 					  CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2386 					  __ceph_caps_used(ci),
2387 					  __ceph_caps_wanted(ci),
2388 					  (cap->issued | cap->implemented),
2389 					  cf->caps, cf->tid, oldest_flush_tid);
2390 			spin_unlock(&ci->i_ceph_lock);
2391 			__send_cap(&arg, ci);
2392 		} else {
2393 			struct ceph_cap_snap *capsnap =
2394 					container_of(cf, struct ceph_cap_snap,
2395 						    cap_flush);
2396 			dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2397 			     inode, capsnap, cf->tid,
2398 			     ceph_cap_string(capsnap->dirty));
2399 
2400 			refcount_inc(&capsnap->nref);
2401 			spin_unlock(&ci->i_ceph_lock);
2402 
2403 			ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2404 						oldest_flush_tid);
2405 			if (ret < 0) {
2406 				pr_err("kick_flushing_caps: error sending "
2407 					"cap flushsnap, ino (%llx.%llx) "
2408 					"tid %llu follows %llu\n",
2409 					ceph_vinop(inode), cf->tid,
2410 					capsnap->follows);
2411 			}
2412 
2413 			ceph_put_cap_snap(capsnap);
2414 		}
2415 
2416 		spin_lock(&ci->i_ceph_lock);
2417 	}
2418 }
2419 
2420 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2421 				   struct ceph_mds_session *session)
2422 {
2423 	struct ceph_inode_info *ci;
2424 	struct ceph_cap *cap;
2425 	u64 oldest_flush_tid;
2426 
2427 	dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2428 
2429 	spin_lock(&mdsc->cap_dirty_lock);
2430 	oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2431 	spin_unlock(&mdsc->cap_dirty_lock);
2432 
2433 	list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2434 		spin_lock(&ci->i_ceph_lock);
2435 		cap = ci->i_auth_cap;
2436 		if (!(cap && cap->session == session)) {
2437 			pr_err("%p auth cap %p not mds%d ???\n",
2438 				&ci->vfs_inode, cap, session->s_mds);
2439 			spin_unlock(&ci->i_ceph_lock);
2440 			continue;
2441 		}
2442 
2443 
2444 		/*
2445 		 * if flushing caps were revoked, we re-send the cap flush
2446 		 * in client reconnect stage. This guarantees MDS * processes
2447 		 * the cap flush message before issuing the flushing caps to
2448 		 * other client.
2449 		 */
2450 		if ((cap->issued & ci->i_flushing_caps) !=
2451 		    ci->i_flushing_caps) {
2452 			/* encode_caps_cb() also will reset these sequence
2453 			 * numbers. make sure sequence numbers in cap flush
2454 			 * message match later reconnect message */
2455 			cap->seq = 0;
2456 			cap->issue_seq = 0;
2457 			cap->mseq = 0;
2458 			__kick_flushing_caps(mdsc, session, ci,
2459 					     oldest_flush_tid);
2460 		} else {
2461 			ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2462 		}
2463 
2464 		spin_unlock(&ci->i_ceph_lock);
2465 	}
2466 }
2467 
2468 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2469 			     struct ceph_mds_session *session)
2470 {
2471 	struct ceph_inode_info *ci;
2472 	struct ceph_cap *cap;
2473 	u64 oldest_flush_tid;
2474 
2475 	lockdep_assert_held(&session->s_mutex);
2476 
2477 	dout("kick_flushing_caps mds%d\n", session->s_mds);
2478 
2479 	spin_lock(&mdsc->cap_dirty_lock);
2480 	oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2481 	spin_unlock(&mdsc->cap_dirty_lock);
2482 
2483 	list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2484 		spin_lock(&ci->i_ceph_lock);
2485 		cap = ci->i_auth_cap;
2486 		if (!(cap && cap->session == session)) {
2487 			pr_err("%p auth cap %p not mds%d ???\n",
2488 				&ci->vfs_inode, cap, session->s_mds);
2489 			spin_unlock(&ci->i_ceph_lock);
2490 			continue;
2491 		}
2492 		if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2493 			__kick_flushing_caps(mdsc, session, ci,
2494 					     oldest_flush_tid);
2495 		}
2496 		spin_unlock(&ci->i_ceph_lock);
2497 	}
2498 }
2499 
2500 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2501 				   struct ceph_inode_info *ci)
2502 {
2503 	struct ceph_mds_client *mdsc = session->s_mdsc;
2504 	struct ceph_cap *cap = ci->i_auth_cap;
2505 
2506 	lockdep_assert_held(&ci->i_ceph_lock);
2507 
2508 	dout("%s %p flushing %s\n", __func__, &ci->vfs_inode,
2509 	     ceph_cap_string(ci->i_flushing_caps));
2510 
2511 	if (!list_empty(&ci->i_cap_flush_list)) {
2512 		u64 oldest_flush_tid;
2513 		spin_lock(&mdsc->cap_dirty_lock);
2514 		list_move_tail(&ci->i_flushing_item,
2515 			       &cap->session->s_cap_flushing);
2516 		oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2517 		spin_unlock(&mdsc->cap_dirty_lock);
2518 
2519 		__kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2520 	}
2521 }
2522 
2523 
2524 /*
2525  * Take references to capabilities we hold, so that we don't release
2526  * them to the MDS prematurely.
2527  */
2528 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2529 			    bool snap_rwsem_locked)
2530 {
2531 	lockdep_assert_held(&ci->i_ceph_lock);
2532 
2533 	if (got & CEPH_CAP_PIN)
2534 		ci->i_pin_ref++;
2535 	if (got & CEPH_CAP_FILE_RD)
2536 		ci->i_rd_ref++;
2537 	if (got & CEPH_CAP_FILE_CACHE)
2538 		ci->i_rdcache_ref++;
2539 	if (got & CEPH_CAP_FILE_EXCL)
2540 		ci->i_fx_ref++;
2541 	if (got & CEPH_CAP_FILE_WR) {
2542 		if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2543 			BUG_ON(!snap_rwsem_locked);
2544 			ci->i_head_snapc = ceph_get_snap_context(
2545 					ci->i_snap_realm->cached_context);
2546 		}
2547 		ci->i_wr_ref++;
2548 	}
2549 	if (got & CEPH_CAP_FILE_BUFFER) {
2550 		if (ci->i_wb_ref == 0)
2551 			ihold(&ci->vfs_inode);
2552 		ci->i_wb_ref++;
2553 		dout("%s %p wb %d -> %d (?)\n", __func__,
2554 		     &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2555 	}
2556 }
2557 
2558 /*
2559  * Try to grab cap references.  Specify those refs we @want, and the
2560  * minimal set we @need.  Also include the larger offset we are writing
2561  * to (when applicable), and check against max_size here as well.
2562  * Note that caller is responsible for ensuring max_size increases are
2563  * requested from the MDS.
2564  *
2565  * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2566  * or a negative error code. There are 3 speical error codes:
2567  *  -EAGAIN: need to sleep but non-blocking is specified
2568  *  -EFBIG:  ask caller to call check_max_size() and try again.
2569  *  -ESTALE: ask caller to call ceph_renew_caps() and try again.
2570  */
2571 enum {
2572 	/* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2573 	NON_BLOCKING	= (1 << 8),
2574 	CHECK_FILELOCK	= (1 << 9),
2575 };
2576 
2577 static int try_get_cap_refs(struct inode *inode, int need, int want,
2578 			    loff_t endoff, int flags, int *got)
2579 {
2580 	struct ceph_inode_info *ci = ceph_inode(inode);
2581 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2582 	int ret = 0;
2583 	int have, implemented;
2584 	bool snap_rwsem_locked = false;
2585 
2586 	dout("get_cap_refs %p need %s want %s\n", inode,
2587 	     ceph_cap_string(need), ceph_cap_string(want));
2588 
2589 again:
2590 	spin_lock(&ci->i_ceph_lock);
2591 
2592 	if ((flags & CHECK_FILELOCK) &&
2593 	    (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2594 		dout("try_get_cap_refs %p error filelock\n", inode);
2595 		ret = -EIO;
2596 		goto out_unlock;
2597 	}
2598 
2599 	/* finish pending truncate */
2600 	while (ci->i_truncate_pending) {
2601 		spin_unlock(&ci->i_ceph_lock);
2602 		if (snap_rwsem_locked) {
2603 			up_read(&mdsc->snap_rwsem);
2604 			snap_rwsem_locked = false;
2605 		}
2606 		__ceph_do_pending_vmtruncate(inode);
2607 		spin_lock(&ci->i_ceph_lock);
2608 	}
2609 
2610 	have = __ceph_caps_issued(ci, &implemented);
2611 
2612 	if (have & need & CEPH_CAP_FILE_WR) {
2613 		if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2614 			dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2615 			     inode, endoff, ci->i_max_size);
2616 			if (endoff > ci->i_requested_max_size)
2617 				ret = ci->i_auth_cap ? -EFBIG : -ESTALE;
2618 			goto out_unlock;
2619 		}
2620 		/*
2621 		 * If a sync write is in progress, we must wait, so that we
2622 		 * can get a final snapshot value for size+mtime.
2623 		 */
2624 		if (__ceph_have_pending_cap_snap(ci)) {
2625 			dout("get_cap_refs %p cap_snap_pending\n", inode);
2626 			goto out_unlock;
2627 		}
2628 	}
2629 
2630 	if ((have & need) == need) {
2631 		/*
2632 		 * Look at (implemented & ~have & not) so that we keep waiting
2633 		 * on transition from wanted -> needed caps.  This is needed
2634 		 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2635 		 * going before a prior buffered writeback happens.
2636 		 */
2637 		int not = want & ~(have & need);
2638 		int revoking = implemented & ~have;
2639 		dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2640 		     inode, ceph_cap_string(have), ceph_cap_string(not),
2641 		     ceph_cap_string(revoking));
2642 		if ((revoking & not) == 0) {
2643 			if (!snap_rwsem_locked &&
2644 			    !ci->i_head_snapc &&
2645 			    (need & CEPH_CAP_FILE_WR)) {
2646 				if (!down_read_trylock(&mdsc->snap_rwsem)) {
2647 					/*
2648 					 * we can not call down_read() when
2649 					 * task isn't in TASK_RUNNING state
2650 					 */
2651 					if (flags & NON_BLOCKING) {
2652 						ret = -EAGAIN;
2653 						goto out_unlock;
2654 					}
2655 
2656 					spin_unlock(&ci->i_ceph_lock);
2657 					down_read(&mdsc->snap_rwsem);
2658 					snap_rwsem_locked = true;
2659 					goto again;
2660 				}
2661 				snap_rwsem_locked = true;
2662 			}
2663 			if ((have & want) == want)
2664 				*got = need | want;
2665 			else
2666 				*got = need;
2667 			ceph_take_cap_refs(ci, *got, true);
2668 			ret = 1;
2669 		}
2670 	} else {
2671 		int session_readonly = false;
2672 		int mds_wanted;
2673 		if (ci->i_auth_cap &&
2674 		    (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2675 			struct ceph_mds_session *s = ci->i_auth_cap->session;
2676 			spin_lock(&s->s_cap_lock);
2677 			session_readonly = s->s_readonly;
2678 			spin_unlock(&s->s_cap_lock);
2679 		}
2680 		if (session_readonly) {
2681 			dout("get_cap_refs %p need %s but mds%d readonly\n",
2682 			     inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2683 			ret = -EROFS;
2684 			goto out_unlock;
2685 		}
2686 
2687 		if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
2688 			dout("get_cap_refs %p forced umount\n", inode);
2689 			ret = -EIO;
2690 			goto out_unlock;
2691 		}
2692 		mds_wanted = __ceph_caps_mds_wanted(ci, false);
2693 		if (need & ~mds_wanted) {
2694 			dout("get_cap_refs %p need %s > mds_wanted %s\n",
2695 			     inode, ceph_cap_string(need),
2696 			     ceph_cap_string(mds_wanted));
2697 			ret = -ESTALE;
2698 			goto out_unlock;
2699 		}
2700 
2701 		dout("get_cap_refs %p have %s need %s\n", inode,
2702 		     ceph_cap_string(have), ceph_cap_string(need));
2703 	}
2704 out_unlock:
2705 
2706 	__ceph_touch_fmode(ci, mdsc, flags);
2707 
2708 	spin_unlock(&ci->i_ceph_lock);
2709 	if (snap_rwsem_locked)
2710 		up_read(&mdsc->snap_rwsem);
2711 
2712 	if (!ret)
2713 		ceph_update_cap_mis(&mdsc->metric);
2714 	else if (ret == 1)
2715 		ceph_update_cap_hit(&mdsc->metric);
2716 
2717 	dout("get_cap_refs %p ret %d got %s\n", inode,
2718 	     ret, ceph_cap_string(*got));
2719 	return ret;
2720 }
2721 
2722 /*
2723  * Check the offset we are writing up to against our current
2724  * max_size.  If necessary, tell the MDS we want to write to
2725  * a larger offset.
2726  */
2727 static void check_max_size(struct inode *inode, loff_t endoff)
2728 {
2729 	struct ceph_inode_info *ci = ceph_inode(inode);
2730 	int check = 0;
2731 
2732 	/* do we need to explicitly request a larger max_size? */
2733 	spin_lock(&ci->i_ceph_lock);
2734 	if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2735 		dout("write %p at large endoff %llu, req max_size\n",
2736 		     inode, endoff);
2737 		ci->i_wanted_max_size = endoff;
2738 	}
2739 	/* duplicate ceph_check_caps()'s logic */
2740 	if (ci->i_auth_cap &&
2741 	    (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2742 	    ci->i_wanted_max_size > ci->i_max_size &&
2743 	    ci->i_wanted_max_size > ci->i_requested_max_size)
2744 		check = 1;
2745 	spin_unlock(&ci->i_ceph_lock);
2746 	if (check)
2747 		ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2748 }
2749 
2750 static inline int get_used_fmode(int caps)
2751 {
2752 	int fmode = 0;
2753 	if (caps & CEPH_CAP_FILE_RD)
2754 		fmode |= CEPH_FILE_MODE_RD;
2755 	if (caps & CEPH_CAP_FILE_WR)
2756 		fmode |= CEPH_FILE_MODE_WR;
2757 	return fmode;
2758 }
2759 
2760 int ceph_try_get_caps(struct inode *inode, int need, int want,
2761 		      bool nonblock, int *got)
2762 {
2763 	int ret, flags;
2764 
2765 	BUG_ON(need & ~CEPH_CAP_FILE_RD);
2766 	BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2767 			CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2768 			CEPH_CAP_ANY_DIR_OPS));
2769 	if (need) {
2770 		ret = ceph_pool_perm_check(inode, need);
2771 		if (ret < 0)
2772 			return ret;
2773 	}
2774 
2775 	flags = get_used_fmode(need | want);
2776 	if (nonblock)
2777 		flags |= NON_BLOCKING;
2778 
2779 	ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2780 	/* three special error codes */
2781 	if (ret == -EAGAIN || ret == -EFBIG || ret == -ESTALE)
2782 		ret = 0;
2783 	return ret;
2784 }
2785 
2786 /*
2787  * Wait for caps, and take cap references.  If we can't get a WR cap
2788  * due to a small max_size, make sure we check_max_size (and possibly
2789  * ask the mds) so we don't get hung up indefinitely.
2790  */
2791 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2792 {
2793 	struct ceph_file_info *fi = filp->private_data;
2794 	struct inode *inode = file_inode(filp);
2795 	struct ceph_inode_info *ci = ceph_inode(inode);
2796 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2797 	int ret, _got, flags;
2798 
2799 	ret = ceph_pool_perm_check(inode, need);
2800 	if (ret < 0)
2801 		return ret;
2802 
2803 	if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2804 	    fi->filp_gen != READ_ONCE(fsc->filp_gen))
2805 		return -EBADF;
2806 
2807 	flags = get_used_fmode(need | want);
2808 
2809 	while (true) {
2810 		flags &= CEPH_FILE_MODE_MASK;
2811 		if (atomic_read(&fi->num_locks))
2812 			flags |= CHECK_FILELOCK;
2813 		_got = 0;
2814 		ret = try_get_cap_refs(inode, need, want, endoff,
2815 				       flags, &_got);
2816 		WARN_ON_ONCE(ret == -EAGAIN);
2817 		if (!ret) {
2818 			struct ceph_mds_client *mdsc = fsc->mdsc;
2819 			struct cap_wait cw;
2820 			DEFINE_WAIT_FUNC(wait, woken_wake_function);
2821 
2822 			cw.ino = ceph_ino(inode);
2823 			cw.tgid = current->tgid;
2824 			cw.need = need;
2825 			cw.want = want;
2826 
2827 			spin_lock(&mdsc->caps_list_lock);
2828 			list_add(&cw.list, &mdsc->cap_wait_list);
2829 			spin_unlock(&mdsc->caps_list_lock);
2830 
2831 			/* make sure used fmode not timeout */
2832 			ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2833 			add_wait_queue(&ci->i_cap_wq, &wait);
2834 
2835 			flags |= NON_BLOCKING;
2836 			while (!(ret = try_get_cap_refs(inode, need, want,
2837 							endoff, flags, &_got))) {
2838 				if (signal_pending(current)) {
2839 					ret = -ERESTARTSYS;
2840 					break;
2841 				}
2842 				wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2843 			}
2844 
2845 			remove_wait_queue(&ci->i_cap_wq, &wait);
2846 			ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2847 
2848 			spin_lock(&mdsc->caps_list_lock);
2849 			list_del(&cw.list);
2850 			spin_unlock(&mdsc->caps_list_lock);
2851 
2852 			if (ret == -EAGAIN)
2853 				continue;
2854 		}
2855 
2856 		if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2857 		    fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2858 			if (ret >= 0 && _got)
2859 				ceph_put_cap_refs(ci, _got);
2860 			return -EBADF;
2861 		}
2862 
2863 		if (ret < 0) {
2864 			if (ret == -EFBIG || ret == -ESTALE) {
2865 				int ret2 = ceph_wait_on_async_create(inode);
2866 				if (ret2 < 0)
2867 					return ret2;
2868 			}
2869 			if (ret == -EFBIG) {
2870 				check_max_size(inode, endoff);
2871 				continue;
2872 			}
2873 			if (ret == -ESTALE) {
2874 				/* session was killed, try renew caps */
2875 				ret = ceph_renew_caps(inode, flags);
2876 				if (ret == 0)
2877 					continue;
2878 			}
2879 			return ret;
2880 		}
2881 
2882 		if (S_ISREG(ci->vfs_inode.i_mode) &&
2883 		    ci->i_inline_version != CEPH_INLINE_NONE &&
2884 		    (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2885 		    i_size_read(inode) > 0) {
2886 			struct page *page =
2887 				find_get_page(inode->i_mapping, 0);
2888 			if (page) {
2889 				bool uptodate = PageUptodate(page);
2890 
2891 				put_page(page);
2892 				if (uptodate)
2893 					break;
2894 			}
2895 			/*
2896 			 * drop cap refs first because getattr while
2897 			 * holding * caps refs can cause deadlock.
2898 			 */
2899 			ceph_put_cap_refs(ci, _got);
2900 			_got = 0;
2901 
2902 			/*
2903 			 * getattr request will bring inline data into
2904 			 * page cache
2905 			 */
2906 			ret = __ceph_do_getattr(inode, NULL,
2907 						CEPH_STAT_CAP_INLINE_DATA,
2908 						true);
2909 			if (ret < 0)
2910 				return ret;
2911 			continue;
2912 		}
2913 		break;
2914 	}
2915 	*got = _got;
2916 	return 0;
2917 }
2918 
2919 /*
2920  * Take cap refs.  Caller must already know we hold at least one ref
2921  * on the caps in question or we don't know this is safe.
2922  */
2923 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2924 {
2925 	spin_lock(&ci->i_ceph_lock);
2926 	ceph_take_cap_refs(ci, caps, false);
2927 	spin_unlock(&ci->i_ceph_lock);
2928 }
2929 
2930 
2931 /*
2932  * drop cap_snap that is not associated with any snapshot.
2933  * we don't need to send FLUSHSNAP message for it.
2934  */
2935 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2936 				  struct ceph_cap_snap *capsnap)
2937 {
2938 	if (!capsnap->need_flush &&
2939 	    !capsnap->writing && !capsnap->dirty_pages) {
2940 		dout("dropping cap_snap %p follows %llu\n",
2941 		     capsnap, capsnap->follows);
2942 		BUG_ON(capsnap->cap_flush.tid > 0);
2943 		ceph_put_snap_context(capsnap->context);
2944 		if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2945 			ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2946 
2947 		list_del(&capsnap->ci_item);
2948 		ceph_put_cap_snap(capsnap);
2949 		return 1;
2950 	}
2951 	return 0;
2952 }
2953 
2954 enum put_cap_refs_mode {
2955 	PUT_CAP_REFS_SYNC = 0,
2956 	PUT_CAP_REFS_NO_CHECK,
2957 	PUT_CAP_REFS_ASYNC,
2958 };
2959 
2960 /*
2961  * Release cap refs.
2962  *
2963  * If we released the last ref on any given cap, call ceph_check_caps
2964  * to release (or schedule a release).
2965  *
2966  * If we are releasing a WR cap (from a sync write), finalize any affected
2967  * cap_snap, and wake up any waiters.
2968  */
2969 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
2970 				enum put_cap_refs_mode mode)
2971 {
2972 	struct inode *inode = &ci->vfs_inode;
2973 	int last = 0, put = 0, flushsnaps = 0, wake = 0;
2974 	bool check_flushsnaps = false;
2975 
2976 	spin_lock(&ci->i_ceph_lock);
2977 	if (had & CEPH_CAP_PIN)
2978 		--ci->i_pin_ref;
2979 	if (had & CEPH_CAP_FILE_RD)
2980 		if (--ci->i_rd_ref == 0)
2981 			last++;
2982 	if (had & CEPH_CAP_FILE_CACHE)
2983 		if (--ci->i_rdcache_ref == 0)
2984 			last++;
2985 	if (had & CEPH_CAP_FILE_EXCL)
2986 		if (--ci->i_fx_ref == 0)
2987 			last++;
2988 	if (had & CEPH_CAP_FILE_BUFFER) {
2989 		if (--ci->i_wb_ref == 0) {
2990 			last++;
2991 			/* put the ref held by ceph_take_cap_refs() */
2992 			put++;
2993 			check_flushsnaps = true;
2994 		}
2995 		dout("put_cap_refs %p wb %d -> %d (?)\n",
2996 		     inode, ci->i_wb_ref+1, ci->i_wb_ref);
2997 	}
2998 	if (had & CEPH_CAP_FILE_WR) {
2999 		if (--ci->i_wr_ref == 0) {
3000 			last++;
3001 			check_flushsnaps = true;
3002 			if (ci->i_wrbuffer_ref_head == 0 &&
3003 			    ci->i_dirty_caps == 0 &&
3004 			    ci->i_flushing_caps == 0) {
3005 				BUG_ON(!ci->i_head_snapc);
3006 				ceph_put_snap_context(ci->i_head_snapc);
3007 				ci->i_head_snapc = NULL;
3008 			}
3009 			/* see comment in __ceph_remove_cap() */
3010 			if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3011 				drop_inode_snap_realm(ci);
3012 		}
3013 	}
3014 	if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3015 		struct ceph_cap_snap *capsnap =
3016 			list_last_entry(&ci->i_cap_snaps,
3017 					struct ceph_cap_snap,
3018 					ci_item);
3019 
3020 		capsnap->writing = 0;
3021 		if (ceph_try_drop_cap_snap(ci, capsnap))
3022 			/* put the ref held by ceph_queue_cap_snap() */
3023 			put++;
3024 		else if (__ceph_finish_cap_snap(ci, capsnap))
3025 			flushsnaps = 1;
3026 		wake = 1;
3027 	}
3028 	spin_unlock(&ci->i_ceph_lock);
3029 
3030 	dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3031 	     last ? " last" : "", put ? " put" : "");
3032 
3033 	switch (mode) {
3034 	case PUT_CAP_REFS_SYNC:
3035 		if (last)
3036 			ceph_check_caps(ci, 0, NULL);
3037 		else if (flushsnaps)
3038 			ceph_flush_snaps(ci, NULL);
3039 		break;
3040 	case PUT_CAP_REFS_ASYNC:
3041 		if (last)
3042 			ceph_queue_check_caps(inode);
3043 		else if (flushsnaps)
3044 			ceph_queue_flush_snaps(inode);
3045 		break;
3046 	default:
3047 		break;
3048 	}
3049 	if (wake)
3050 		wake_up_all(&ci->i_cap_wq);
3051 	while (put-- > 0)
3052 		iput(inode);
3053 }
3054 
3055 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3056 {
3057 	__ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3058 }
3059 
3060 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3061 {
3062 	__ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3063 }
3064 
3065 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3066 {
3067 	__ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3068 }
3069 
3070 /*
3071  * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3072  * context.  Adjust per-snap dirty page accounting as appropriate.
3073  * Once all dirty data for a cap_snap is flushed, flush snapped file
3074  * metadata back to the MDS.  If we dropped the last ref, call
3075  * ceph_check_caps.
3076  */
3077 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3078 				struct ceph_snap_context *snapc)
3079 {
3080 	struct inode *inode = &ci->vfs_inode;
3081 	struct ceph_cap_snap *capsnap = NULL;
3082 	int put = 0;
3083 	bool last = false;
3084 	bool found = false;
3085 	bool flush_snaps = false;
3086 	bool complete_capsnap = false;
3087 
3088 	spin_lock(&ci->i_ceph_lock);
3089 	ci->i_wrbuffer_ref -= nr;
3090 	if (ci->i_wrbuffer_ref == 0) {
3091 		last = true;
3092 		put++;
3093 	}
3094 
3095 	if (ci->i_head_snapc == snapc) {
3096 		ci->i_wrbuffer_ref_head -= nr;
3097 		if (ci->i_wrbuffer_ref_head == 0 &&
3098 		    ci->i_wr_ref == 0 &&
3099 		    ci->i_dirty_caps == 0 &&
3100 		    ci->i_flushing_caps == 0) {
3101 			BUG_ON(!ci->i_head_snapc);
3102 			ceph_put_snap_context(ci->i_head_snapc);
3103 			ci->i_head_snapc = NULL;
3104 		}
3105 		dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3106 		     inode,
3107 		     ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3108 		     ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3109 		     last ? " LAST" : "");
3110 	} else {
3111 		list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3112 			if (capsnap->context == snapc) {
3113 				found = true;
3114 				break;
3115 			}
3116 		}
3117 		BUG_ON(!found);
3118 		capsnap->dirty_pages -= nr;
3119 		if (capsnap->dirty_pages == 0) {
3120 			complete_capsnap = true;
3121 			if (!capsnap->writing) {
3122 				if (ceph_try_drop_cap_snap(ci, capsnap)) {
3123 					put++;
3124 				} else {
3125 					ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3126 					flush_snaps = true;
3127 				}
3128 			}
3129 		}
3130 		dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3131 		     " snap %lld %d/%d -> %d/%d %s%s\n",
3132 		     inode, capsnap, capsnap->context->seq,
3133 		     ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3134 		     ci->i_wrbuffer_ref, capsnap->dirty_pages,
3135 		     last ? " (wrbuffer last)" : "",
3136 		     complete_capsnap ? " (complete capsnap)" : "");
3137 	}
3138 
3139 	spin_unlock(&ci->i_ceph_lock);
3140 
3141 	if (last) {
3142 		ceph_check_caps(ci, 0, NULL);
3143 	} else if (flush_snaps) {
3144 		ceph_flush_snaps(ci, NULL);
3145 	}
3146 	if (complete_capsnap)
3147 		wake_up_all(&ci->i_cap_wq);
3148 	while (put-- > 0) {
3149 		iput(inode);
3150 	}
3151 }
3152 
3153 /*
3154  * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3155  */
3156 static void invalidate_aliases(struct inode *inode)
3157 {
3158 	struct dentry *dn, *prev = NULL;
3159 
3160 	dout("invalidate_aliases inode %p\n", inode);
3161 	d_prune_aliases(inode);
3162 	/*
3163 	 * For non-directory inode, d_find_alias() only returns
3164 	 * hashed dentry. After calling d_invalidate(), the
3165 	 * dentry becomes unhashed.
3166 	 *
3167 	 * For directory inode, d_find_alias() can return
3168 	 * unhashed dentry. But directory inode should have
3169 	 * one alias at most.
3170 	 */
3171 	while ((dn = d_find_alias(inode))) {
3172 		if (dn == prev) {
3173 			dput(dn);
3174 			break;
3175 		}
3176 		d_invalidate(dn);
3177 		if (prev)
3178 			dput(prev);
3179 		prev = dn;
3180 	}
3181 	if (prev)
3182 		dput(prev);
3183 }
3184 
3185 struct cap_extra_info {
3186 	struct ceph_string *pool_ns;
3187 	/* inline data */
3188 	u64 inline_version;
3189 	void *inline_data;
3190 	u32 inline_len;
3191 	/* dirstat */
3192 	bool dirstat_valid;
3193 	u64 nfiles;
3194 	u64 nsubdirs;
3195 	u64 change_attr;
3196 	/* currently issued */
3197 	int issued;
3198 	struct timespec64 btime;
3199 };
3200 
3201 /*
3202  * Handle a cap GRANT message from the MDS.  (Note that a GRANT may
3203  * actually be a revocation if it specifies a smaller cap set.)
3204  *
3205  * caller holds s_mutex and i_ceph_lock, we drop both.
3206  */
3207 static void handle_cap_grant(struct inode *inode,
3208 			     struct ceph_mds_session *session,
3209 			     struct ceph_cap *cap,
3210 			     struct ceph_mds_caps *grant,
3211 			     struct ceph_buffer *xattr_buf,
3212 			     struct cap_extra_info *extra_info)
3213 	__releases(ci->i_ceph_lock)
3214 	__releases(session->s_mdsc->snap_rwsem)
3215 {
3216 	struct ceph_inode_info *ci = ceph_inode(inode);
3217 	int seq = le32_to_cpu(grant->seq);
3218 	int newcaps = le32_to_cpu(grant->caps);
3219 	int used, wanted, dirty;
3220 	u64 size = le64_to_cpu(grant->size);
3221 	u64 max_size = le64_to_cpu(grant->max_size);
3222 	unsigned char check_caps = 0;
3223 	bool was_stale = cap->cap_gen < atomic_read(&session->s_cap_gen);
3224 	bool wake = false;
3225 	bool writeback = false;
3226 	bool queue_trunc = false;
3227 	bool queue_invalidate = false;
3228 	bool deleted_inode = false;
3229 	bool fill_inline = false;
3230 
3231 	dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3232 	     inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3233 	dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3234 		i_size_read(inode));
3235 
3236 
3237 	/*
3238 	 * If CACHE is being revoked, and we have no dirty buffers,
3239 	 * try to invalidate (once).  (If there are dirty buffers, we
3240 	 * will invalidate _after_ writeback.)
3241 	 */
3242 	if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3243 	    ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3244 	    (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3245 	    !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3246 		if (try_nonblocking_invalidate(inode)) {
3247 			/* there were locked pages.. invalidate later
3248 			   in a separate thread. */
3249 			if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3250 				queue_invalidate = true;
3251 				ci->i_rdcache_revoking = ci->i_rdcache_gen;
3252 			}
3253 		}
3254 	}
3255 
3256 	if (was_stale)
3257 		cap->issued = cap->implemented = CEPH_CAP_PIN;
3258 
3259 	/*
3260 	 * auth mds of the inode changed. we received the cap export message,
3261 	 * but still haven't received the cap import message. handle_cap_export
3262 	 * updated the new auth MDS' cap.
3263 	 *
3264 	 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3265 	 * that was sent before the cap import message. So don't remove caps.
3266 	 */
3267 	if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3268 		WARN_ON(cap != ci->i_auth_cap);
3269 		WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3270 		seq = cap->seq;
3271 		newcaps |= cap->issued;
3272 	}
3273 
3274 	/* side effects now are allowed */
3275 	cap->cap_gen = atomic_read(&session->s_cap_gen);
3276 	cap->seq = seq;
3277 
3278 	__check_cap_issue(ci, cap, newcaps);
3279 
3280 	inode_set_max_iversion_raw(inode, extra_info->change_attr);
3281 
3282 	if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3283 	    (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3284 		umode_t mode = le32_to_cpu(grant->mode);
3285 
3286 		if (inode_wrong_type(inode, mode))
3287 			pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3288 				     ceph_vinop(inode), inode->i_mode, mode);
3289 		else
3290 			inode->i_mode = mode;
3291 		inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3292 		inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3293 		ci->i_btime = extra_info->btime;
3294 		dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3295 		     from_kuid(&init_user_ns, inode->i_uid),
3296 		     from_kgid(&init_user_ns, inode->i_gid));
3297 	}
3298 
3299 	if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3300 	    (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3301 		set_nlink(inode, le32_to_cpu(grant->nlink));
3302 		if (inode->i_nlink == 0 &&
3303 		    (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3304 			deleted_inode = true;
3305 	}
3306 
3307 	if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3308 	    grant->xattr_len) {
3309 		int len = le32_to_cpu(grant->xattr_len);
3310 		u64 version = le64_to_cpu(grant->xattr_version);
3311 
3312 		if (version > ci->i_xattrs.version) {
3313 			dout(" got new xattrs v%llu on %p len %d\n",
3314 			     version, inode, len);
3315 			if (ci->i_xattrs.blob)
3316 				ceph_buffer_put(ci->i_xattrs.blob);
3317 			ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3318 			ci->i_xattrs.version = version;
3319 			ceph_forget_all_cached_acls(inode);
3320 			ceph_security_invalidate_secctx(inode);
3321 		}
3322 	}
3323 
3324 	if (newcaps & CEPH_CAP_ANY_RD) {
3325 		struct timespec64 mtime, atime, ctime;
3326 		/* ctime/mtime/atime? */
3327 		ceph_decode_timespec64(&mtime, &grant->mtime);
3328 		ceph_decode_timespec64(&atime, &grant->atime);
3329 		ceph_decode_timespec64(&ctime, &grant->ctime);
3330 		ceph_fill_file_time(inode, extra_info->issued,
3331 				    le32_to_cpu(grant->time_warp_seq),
3332 				    &ctime, &mtime, &atime);
3333 	}
3334 
3335 	if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3336 		ci->i_files = extra_info->nfiles;
3337 		ci->i_subdirs = extra_info->nsubdirs;
3338 	}
3339 
3340 	if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3341 		/* file layout may have changed */
3342 		s64 old_pool = ci->i_layout.pool_id;
3343 		struct ceph_string *old_ns;
3344 
3345 		ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3346 		old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3347 					lockdep_is_held(&ci->i_ceph_lock));
3348 		rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3349 
3350 		if (ci->i_layout.pool_id != old_pool ||
3351 		    extra_info->pool_ns != old_ns)
3352 			ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3353 
3354 		extra_info->pool_ns = old_ns;
3355 
3356 		/* size/truncate_seq? */
3357 		queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3358 					le32_to_cpu(grant->truncate_seq),
3359 					le64_to_cpu(grant->truncate_size),
3360 					size);
3361 	}
3362 
3363 	if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3364 		if (max_size != ci->i_max_size) {
3365 			dout("max_size %lld -> %llu\n",
3366 			     ci->i_max_size, max_size);
3367 			ci->i_max_size = max_size;
3368 			if (max_size >= ci->i_wanted_max_size) {
3369 				ci->i_wanted_max_size = 0;  /* reset */
3370 				ci->i_requested_max_size = 0;
3371 			}
3372 			wake = true;
3373 		}
3374 	}
3375 
3376 	/* check cap bits */
3377 	wanted = __ceph_caps_wanted(ci);
3378 	used = __ceph_caps_used(ci);
3379 	dirty = __ceph_caps_dirty(ci);
3380 	dout(" my wanted = %s, used = %s, dirty %s\n",
3381 	     ceph_cap_string(wanted),
3382 	     ceph_cap_string(used),
3383 	     ceph_cap_string(dirty));
3384 
3385 	if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3386 	    (wanted & ~(cap->mds_wanted | newcaps))) {
3387 		/*
3388 		 * If mds is importing cap, prior cap messages that update
3389 		 * 'wanted' may get dropped by mds (migrate seq mismatch).
3390 		 *
3391 		 * We don't send cap message to update 'wanted' if what we
3392 		 * want are already issued. If mds revokes caps, cap message
3393 		 * that releases caps also tells mds what we want. But if
3394 		 * caps got revoked by mds forcedly (session stale). We may
3395 		 * haven't told mds what we want.
3396 		 */
3397 		check_caps = 1;
3398 	}
3399 
3400 	/* revocation, grant, or no-op? */
3401 	if (cap->issued & ~newcaps) {
3402 		int revoking = cap->issued & ~newcaps;
3403 
3404 		dout("revocation: %s -> %s (revoking %s)\n",
3405 		     ceph_cap_string(cap->issued),
3406 		     ceph_cap_string(newcaps),
3407 		     ceph_cap_string(revoking));
3408 		if (S_ISREG(inode->i_mode) &&
3409 		    (revoking & used & CEPH_CAP_FILE_BUFFER))
3410 			writeback = true;  /* initiate writeback; will delay ack */
3411 		else if (queue_invalidate &&
3412 			 revoking == CEPH_CAP_FILE_CACHE &&
3413 			 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3414 			; /* do nothing yet, invalidation will be queued */
3415 		else if (cap == ci->i_auth_cap)
3416 			check_caps = 1; /* check auth cap only */
3417 		else
3418 			check_caps = 2; /* check all caps */
3419 		cap->issued = newcaps;
3420 		cap->implemented |= newcaps;
3421 	} else if (cap->issued == newcaps) {
3422 		dout("caps unchanged: %s -> %s\n",
3423 		     ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3424 	} else {
3425 		dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3426 		     ceph_cap_string(newcaps));
3427 		/* non-auth MDS is revoking the newly grant caps ? */
3428 		if (cap == ci->i_auth_cap &&
3429 		    __ceph_caps_revoking_other(ci, cap, newcaps))
3430 		    check_caps = 2;
3431 
3432 		cap->issued = newcaps;
3433 		cap->implemented |= newcaps; /* add bits only, to
3434 					      * avoid stepping on a
3435 					      * pending revocation */
3436 		wake = true;
3437 	}
3438 	BUG_ON(cap->issued & ~cap->implemented);
3439 
3440 	if (extra_info->inline_version > 0 &&
3441 	    extra_info->inline_version >= ci->i_inline_version) {
3442 		ci->i_inline_version = extra_info->inline_version;
3443 		if (ci->i_inline_version != CEPH_INLINE_NONE &&
3444 		    (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3445 			fill_inline = true;
3446 	}
3447 
3448 	if (ci->i_auth_cap == cap &&
3449 	    le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3450 		if (newcaps & ~extra_info->issued)
3451 			wake = true;
3452 
3453 		if (ci->i_requested_max_size > max_size ||
3454 		    !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3455 			/* re-request max_size if necessary */
3456 			ci->i_requested_max_size = 0;
3457 			wake = true;
3458 		}
3459 
3460 		ceph_kick_flushing_inode_caps(session, ci);
3461 		spin_unlock(&ci->i_ceph_lock);
3462 		up_read(&session->s_mdsc->snap_rwsem);
3463 	} else {
3464 		spin_unlock(&ci->i_ceph_lock);
3465 	}
3466 
3467 	if (fill_inline)
3468 		ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3469 				      extra_info->inline_len);
3470 
3471 	if (queue_trunc)
3472 		ceph_queue_vmtruncate(inode);
3473 
3474 	if (writeback)
3475 		/*
3476 		 * queue inode for writeback: we can't actually call
3477 		 * filemap_write_and_wait, etc. from message handler
3478 		 * context.
3479 		 */
3480 		ceph_queue_writeback(inode);
3481 	if (queue_invalidate)
3482 		ceph_queue_invalidate(inode);
3483 	if (deleted_inode)
3484 		invalidate_aliases(inode);
3485 	if (wake)
3486 		wake_up_all(&ci->i_cap_wq);
3487 
3488 	mutex_unlock(&session->s_mutex);
3489 	if (check_caps == 1)
3490 		ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3491 				session);
3492 	else if (check_caps == 2)
3493 		ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3494 }
3495 
3496 /*
3497  * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3498  * MDS has been safely committed.
3499  */
3500 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3501 				 struct ceph_mds_caps *m,
3502 				 struct ceph_mds_session *session,
3503 				 struct ceph_cap *cap)
3504 	__releases(ci->i_ceph_lock)
3505 {
3506 	struct ceph_inode_info *ci = ceph_inode(inode);
3507 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3508 	struct ceph_cap_flush *cf, *tmp_cf;
3509 	LIST_HEAD(to_remove);
3510 	unsigned seq = le32_to_cpu(m->seq);
3511 	int dirty = le32_to_cpu(m->dirty);
3512 	int cleaned = 0;
3513 	bool drop = false;
3514 	bool wake_ci = false;
3515 	bool wake_mdsc = false;
3516 
3517 	list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3518 		/* Is this the one that was flushed? */
3519 		if (cf->tid == flush_tid)
3520 			cleaned = cf->caps;
3521 
3522 		/* Is this a capsnap? */
3523 		if (cf->is_capsnap)
3524 			continue;
3525 
3526 		if (cf->tid <= flush_tid) {
3527 			/*
3528 			 * An earlier or current tid. The FLUSH_ACK should
3529 			 * represent a superset of this flush's caps.
3530 			 */
3531 			wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3532 			list_add_tail(&cf->i_list, &to_remove);
3533 		} else {
3534 			/*
3535 			 * This is a later one. Any caps in it are still dirty
3536 			 * so don't count them as cleaned.
3537 			 */
3538 			cleaned &= ~cf->caps;
3539 			if (!cleaned)
3540 				break;
3541 		}
3542 	}
3543 
3544 	dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3545 	     " flushing %s -> %s\n",
3546 	     inode, session->s_mds, seq, ceph_cap_string(dirty),
3547 	     ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3548 	     ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3549 
3550 	if (list_empty(&to_remove) && !cleaned)
3551 		goto out;
3552 
3553 	ci->i_flushing_caps &= ~cleaned;
3554 
3555 	spin_lock(&mdsc->cap_dirty_lock);
3556 
3557 	list_for_each_entry(cf, &to_remove, i_list)
3558 		wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3559 
3560 	if (ci->i_flushing_caps == 0) {
3561 		if (list_empty(&ci->i_cap_flush_list)) {
3562 			list_del_init(&ci->i_flushing_item);
3563 			if (!list_empty(&session->s_cap_flushing)) {
3564 				dout(" mds%d still flushing cap on %p\n",
3565 				     session->s_mds,
3566 				     &list_first_entry(&session->s_cap_flushing,
3567 						struct ceph_inode_info,
3568 						i_flushing_item)->vfs_inode);
3569 			}
3570 		}
3571 		mdsc->num_cap_flushing--;
3572 		dout(" inode %p now !flushing\n", inode);
3573 
3574 		if (ci->i_dirty_caps == 0) {
3575 			dout(" inode %p now clean\n", inode);
3576 			BUG_ON(!list_empty(&ci->i_dirty_item));
3577 			drop = true;
3578 			if (ci->i_wr_ref == 0 &&
3579 			    ci->i_wrbuffer_ref_head == 0) {
3580 				BUG_ON(!ci->i_head_snapc);
3581 				ceph_put_snap_context(ci->i_head_snapc);
3582 				ci->i_head_snapc = NULL;
3583 			}
3584 		} else {
3585 			BUG_ON(list_empty(&ci->i_dirty_item));
3586 		}
3587 	}
3588 	spin_unlock(&mdsc->cap_dirty_lock);
3589 
3590 out:
3591 	spin_unlock(&ci->i_ceph_lock);
3592 
3593 	while (!list_empty(&to_remove)) {
3594 		cf = list_first_entry(&to_remove,
3595 				      struct ceph_cap_flush, i_list);
3596 		list_del_init(&cf->i_list);
3597 		if (!cf->is_capsnap)
3598 			ceph_free_cap_flush(cf);
3599 	}
3600 
3601 	if (wake_ci)
3602 		wake_up_all(&ci->i_cap_wq);
3603 	if (wake_mdsc)
3604 		wake_up_all(&mdsc->cap_flushing_wq);
3605 	if (drop)
3606 		iput(inode);
3607 }
3608 
3609 /*
3610  * Handle FLUSHSNAP_ACK.  MDS has flushed snap data to disk and we can
3611  * throw away our cap_snap.
3612  *
3613  * Caller hold s_mutex.
3614  */
3615 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3616 				     struct ceph_mds_caps *m,
3617 				     struct ceph_mds_session *session)
3618 {
3619 	struct ceph_inode_info *ci = ceph_inode(inode);
3620 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3621 	u64 follows = le64_to_cpu(m->snap_follows);
3622 	struct ceph_cap_snap *capsnap;
3623 	bool flushed = false;
3624 	bool wake_ci = false;
3625 	bool wake_mdsc = false;
3626 
3627 	dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3628 	     inode, ci, session->s_mds, follows);
3629 
3630 	spin_lock(&ci->i_ceph_lock);
3631 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3632 		if (capsnap->follows == follows) {
3633 			if (capsnap->cap_flush.tid != flush_tid) {
3634 				dout(" cap_snap %p follows %lld tid %lld !="
3635 				     " %lld\n", capsnap, follows,
3636 				     flush_tid, capsnap->cap_flush.tid);
3637 				break;
3638 			}
3639 			flushed = true;
3640 			break;
3641 		} else {
3642 			dout(" skipping cap_snap %p follows %lld\n",
3643 			     capsnap, capsnap->follows);
3644 		}
3645 	}
3646 	if (flushed) {
3647 		WARN_ON(capsnap->dirty_pages || capsnap->writing);
3648 		dout(" removing %p cap_snap %p follows %lld\n",
3649 		     inode, capsnap, follows);
3650 		list_del(&capsnap->ci_item);
3651 		wake_ci |= __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3652 
3653 		spin_lock(&mdsc->cap_dirty_lock);
3654 
3655 		if (list_empty(&ci->i_cap_flush_list))
3656 			list_del_init(&ci->i_flushing_item);
3657 
3658 		wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc,
3659 							  &capsnap->cap_flush);
3660 		spin_unlock(&mdsc->cap_dirty_lock);
3661 	}
3662 	spin_unlock(&ci->i_ceph_lock);
3663 	if (flushed) {
3664 		ceph_put_snap_context(capsnap->context);
3665 		ceph_put_cap_snap(capsnap);
3666 		if (wake_ci)
3667 			wake_up_all(&ci->i_cap_wq);
3668 		if (wake_mdsc)
3669 			wake_up_all(&mdsc->cap_flushing_wq);
3670 		iput(inode);
3671 	}
3672 }
3673 
3674 /*
3675  * Handle TRUNC from MDS, indicating file truncation.
3676  *
3677  * caller hold s_mutex.
3678  */
3679 static bool handle_cap_trunc(struct inode *inode,
3680 			     struct ceph_mds_caps *trunc,
3681 			     struct ceph_mds_session *session)
3682 {
3683 	struct ceph_inode_info *ci = ceph_inode(inode);
3684 	int mds = session->s_mds;
3685 	int seq = le32_to_cpu(trunc->seq);
3686 	u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3687 	u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3688 	u64 size = le64_to_cpu(trunc->size);
3689 	int implemented = 0;
3690 	int dirty = __ceph_caps_dirty(ci);
3691 	int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3692 	bool queue_trunc = false;
3693 
3694 	lockdep_assert_held(&ci->i_ceph_lock);
3695 
3696 	issued |= implemented | dirty;
3697 
3698 	dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3699 	     inode, mds, seq, truncate_size, truncate_seq);
3700 	queue_trunc = ceph_fill_file_size(inode, issued,
3701 					  truncate_seq, truncate_size, size);
3702 	return queue_trunc;
3703 }
3704 
3705 /*
3706  * Handle EXPORT from MDS.  Cap is being migrated _from_ this mds to a
3707  * different one.  If we are the most recent migration we've seen (as
3708  * indicated by mseq), make note of the migrating cap bits for the
3709  * duration (until we see the corresponding IMPORT).
3710  *
3711  * caller holds s_mutex
3712  */
3713 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3714 			      struct ceph_mds_cap_peer *ph,
3715 			      struct ceph_mds_session *session)
3716 {
3717 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3718 	struct ceph_mds_session *tsession = NULL;
3719 	struct ceph_cap *cap, *tcap, *new_cap = NULL;
3720 	struct ceph_inode_info *ci = ceph_inode(inode);
3721 	u64 t_cap_id;
3722 	unsigned mseq = le32_to_cpu(ex->migrate_seq);
3723 	unsigned t_seq, t_mseq;
3724 	int target, issued;
3725 	int mds = session->s_mds;
3726 
3727 	if (ph) {
3728 		t_cap_id = le64_to_cpu(ph->cap_id);
3729 		t_seq = le32_to_cpu(ph->seq);
3730 		t_mseq = le32_to_cpu(ph->mseq);
3731 		target = le32_to_cpu(ph->mds);
3732 	} else {
3733 		t_cap_id = t_seq = t_mseq = 0;
3734 		target = -1;
3735 	}
3736 
3737 	dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3738 	     inode, ci, mds, mseq, target);
3739 retry:
3740 	spin_lock(&ci->i_ceph_lock);
3741 	cap = __get_cap_for_mds(ci, mds);
3742 	if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3743 		goto out_unlock;
3744 
3745 	if (target < 0) {
3746 		__ceph_remove_cap(cap, false);
3747 		goto out_unlock;
3748 	}
3749 
3750 	/*
3751 	 * now we know we haven't received the cap import message yet
3752 	 * because the exported cap still exist.
3753 	 */
3754 
3755 	issued = cap->issued;
3756 	if (issued != cap->implemented)
3757 		pr_err_ratelimited("handle_cap_export: issued != implemented: "
3758 				"ino (%llx.%llx) mds%d seq %d mseq %d "
3759 				"issued %s implemented %s\n",
3760 				ceph_vinop(inode), mds, cap->seq, cap->mseq,
3761 				ceph_cap_string(issued),
3762 				ceph_cap_string(cap->implemented));
3763 
3764 
3765 	tcap = __get_cap_for_mds(ci, target);
3766 	if (tcap) {
3767 		/* already have caps from the target */
3768 		if (tcap->cap_id == t_cap_id &&
3769 		    ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3770 			dout(" updating import cap %p mds%d\n", tcap, target);
3771 			tcap->cap_id = t_cap_id;
3772 			tcap->seq = t_seq - 1;
3773 			tcap->issue_seq = t_seq - 1;
3774 			tcap->issued |= issued;
3775 			tcap->implemented |= issued;
3776 			if (cap == ci->i_auth_cap) {
3777 				ci->i_auth_cap = tcap;
3778 				change_auth_cap_ses(ci, tcap->session);
3779 			}
3780 		}
3781 		__ceph_remove_cap(cap, false);
3782 		goto out_unlock;
3783 	} else if (tsession) {
3784 		/* add placeholder for the export tagert */
3785 		int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3786 		tcap = new_cap;
3787 		ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3788 			     t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3789 
3790 		if (!list_empty(&ci->i_cap_flush_list) &&
3791 		    ci->i_auth_cap == tcap) {
3792 			spin_lock(&mdsc->cap_dirty_lock);
3793 			list_move_tail(&ci->i_flushing_item,
3794 				       &tcap->session->s_cap_flushing);
3795 			spin_unlock(&mdsc->cap_dirty_lock);
3796 		}
3797 
3798 		__ceph_remove_cap(cap, false);
3799 		goto out_unlock;
3800 	}
3801 
3802 	spin_unlock(&ci->i_ceph_lock);
3803 	mutex_unlock(&session->s_mutex);
3804 
3805 	/* open target session */
3806 	tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3807 	if (!IS_ERR(tsession)) {
3808 		if (mds > target) {
3809 			mutex_lock(&session->s_mutex);
3810 			mutex_lock_nested(&tsession->s_mutex,
3811 					  SINGLE_DEPTH_NESTING);
3812 		} else {
3813 			mutex_lock(&tsession->s_mutex);
3814 			mutex_lock_nested(&session->s_mutex,
3815 					  SINGLE_DEPTH_NESTING);
3816 		}
3817 		new_cap = ceph_get_cap(mdsc, NULL);
3818 	} else {
3819 		WARN_ON(1);
3820 		tsession = NULL;
3821 		target = -1;
3822 		mutex_lock(&session->s_mutex);
3823 	}
3824 	goto retry;
3825 
3826 out_unlock:
3827 	spin_unlock(&ci->i_ceph_lock);
3828 	mutex_unlock(&session->s_mutex);
3829 	if (tsession) {
3830 		mutex_unlock(&tsession->s_mutex);
3831 		ceph_put_mds_session(tsession);
3832 	}
3833 	if (new_cap)
3834 		ceph_put_cap(mdsc, new_cap);
3835 }
3836 
3837 /*
3838  * Handle cap IMPORT.
3839  *
3840  * caller holds s_mutex. acquires i_ceph_lock
3841  */
3842 static void handle_cap_import(struct ceph_mds_client *mdsc,
3843 			      struct inode *inode, struct ceph_mds_caps *im,
3844 			      struct ceph_mds_cap_peer *ph,
3845 			      struct ceph_mds_session *session,
3846 			      struct ceph_cap **target_cap, int *old_issued)
3847 {
3848 	struct ceph_inode_info *ci = ceph_inode(inode);
3849 	struct ceph_cap *cap, *ocap, *new_cap = NULL;
3850 	int mds = session->s_mds;
3851 	int issued;
3852 	unsigned caps = le32_to_cpu(im->caps);
3853 	unsigned wanted = le32_to_cpu(im->wanted);
3854 	unsigned seq = le32_to_cpu(im->seq);
3855 	unsigned mseq = le32_to_cpu(im->migrate_seq);
3856 	u64 realmino = le64_to_cpu(im->realm);
3857 	u64 cap_id = le64_to_cpu(im->cap_id);
3858 	u64 p_cap_id;
3859 	int peer;
3860 
3861 	if (ph) {
3862 		p_cap_id = le64_to_cpu(ph->cap_id);
3863 		peer = le32_to_cpu(ph->mds);
3864 	} else {
3865 		p_cap_id = 0;
3866 		peer = -1;
3867 	}
3868 
3869 	dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3870 	     inode, ci, mds, mseq, peer);
3871 retry:
3872 	cap = __get_cap_for_mds(ci, mds);
3873 	if (!cap) {
3874 		if (!new_cap) {
3875 			spin_unlock(&ci->i_ceph_lock);
3876 			new_cap = ceph_get_cap(mdsc, NULL);
3877 			spin_lock(&ci->i_ceph_lock);
3878 			goto retry;
3879 		}
3880 		cap = new_cap;
3881 	} else {
3882 		if (new_cap) {
3883 			ceph_put_cap(mdsc, new_cap);
3884 			new_cap = NULL;
3885 		}
3886 	}
3887 
3888 	__ceph_caps_issued(ci, &issued);
3889 	issued |= __ceph_caps_dirty(ci);
3890 
3891 	ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
3892 		     realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3893 
3894 	ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3895 	if (ocap && ocap->cap_id == p_cap_id) {
3896 		dout(" remove export cap %p mds%d flags %d\n",
3897 		     ocap, peer, ph->flags);
3898 		if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3899 		    (ocap->seq != le32_to_cpu(ph->seq) ||
3900 		     ocap->mseq != le32_to_cpu(ph->mseq))) {
3901 			pr_err_ratelimited("handle_cap_import: "
3902 					"mismatched seq/mseq: ino (%llx.%llx) "
3903 					"mds%d seq %d mseq %d importer mds%d "
3904 					"has peer seq %d mseq %d\n",
3905 					ceph_vinop(inode), peer, ocap->seq,
3906 					ocap->mseq, mds, le32_to_cpu(ph->seq),
3907 					le32_to_cpu(ph->mseq));
3908 		}
3909 		__ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3910 	}
3911 
3912 	*old_issued = issued;
3913 	*target_cap = cap;
3914 }
3915 
3916 /*
3917  * Handle a caps message from the MDS.
3918  *
3919  * Identify the appropriate session, inode, and call the right handler
3920  * based on the cap op.
3921  */
3922 void ceph_handle_caps(struct ceph_mds_session *session,
3923 		      struct ceph_msg *msg)
3924 {
3925 	struct ceph_mds_client *mdsc = session->s_mdsc;
3926 	struct inode *inode;
3927 	struct ceph_inode_info *ci;
3928 	struct ceph_cap *cap;
3929 	struct ceph_mds_caps *h;
3930 	struct ceph_mds_cap_peer *peer = NULL;
3931 	struct ceph_snap_realm *realm = NULL;
3932 	int op;
3933 	int msg_version = le16_to_cpu(msg->hdr.version);
3934 	u32 seq, mseq;
3935 	struct ceph_vino vino;
3936 	void *snaptrace;
3937 	size_t snaptrace_len;
3938 	void *p, *end;
3939 	struct cap_extra_info extra_info = {};
3940 	bool queue_trunc;
3941 
3942 	dout("handle_caps from mds%d\n", session->s_mds);
3943 
3944 	/* decode */
3945 	end = msg->front.iov_base + msg->front.iov_len;
3946 	if (msg->front.iov_len < sizeof(*h))
3947 		goto bad;
3948 	h = msg->front.iov_base;
3949 	op = le32_to_cpu(h->op);
3950 	vino.ino = le64_to_cpu(h->ino);
3951 	vino.snap = CEPH_NOSNAP;
3952 	seq = le32_to_cpu(h->seq);
3953 	mseq = le32_to_cpu(h->migrate_seq);
3954 
3955 	snaptrace = h + 1;
3956 	snaptrace_len = le32_to_cpu(h->snap_trace_len);
3957 	p = snaptrace + snaptrace_len;
3958 
3959 	if (msg_version >= 2) {
3960 		u32 flock_len;
3961 		ceph_decode_32_safe(&p, end, flock_len, bad);
3962 		if (p + flock_len > end)
3963 			goto bad;
3964 		p += flock_len;
3965 	}
3966 
3967 	if (msg_version >= 3) {
3968 		if (op == CEPH_CAP_OP_IMPORT) {
3969 			if (p + sizeof(*peer) > end)
3970 				goto bad;
3971 			peer = p;
3972 			p += sizeof(*peer);
3973 		} else if (op == CEPH_CAP_OP_EXPORT) {
3974 			/* recorded in unused fields */
3975 			peer = (void *)&h->size;
3976 		}
3977 	}
3978 
3979 	if (msg_version >= 4) {
3980 		ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
3981 		ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
3982 		if (p + extra_info.inline_len > end)
3983 			goto bad;
3984 		extra_info.inline_data = p;
3985 		p += extra_info.inline_len;
3986 	}
3987 
3988 	if (msg_version >= 5) {
3989 		struct ceph_osd_client	*osdc = &mdsc->fsc->client->osdc;
3990 		u32			epoch_barrier;
3991 
3992 		ceph_decode_32_safe(&p, end, epoch_barrier, bad);
3993 		ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
3994 	}
3995 
3996 	if (msg_version >= 8) {
3997 		u32 pool_ns_len;
3998 
3999 		/* version >= 6 */
4000 		ceph_decode_skip_64(&p, end, bad);	// flush_tid
4001 		/* version >= 7 */
4002 		ceph_decode_skip_32(&p, end, bad);	// caller_uid
4003 		ceph_decode_skip_32(&p, end, bad);	// caller_gid
4004 		/* version >= 8 */
4005 		ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4006 		if (pool_ns_len > 0) {
4007 			ceph_decode_need(&p, end, pool_ns_len, bad);
4008 			extra_info.pool_ns =
4009 				ceph_find_or_create_string(p, pool_ns_len);
4010 			p += pool_ns_len;
4011 		}
4012 	}
4013 
4014 	if (msg_version >= 9) {
4015 		struct ceph_timespec *btime;
4016 
4017 		if (p + sizeof(*btime) > end)
4018 			goto bad;
4019 		btime = p;
4020 		ceph_decode_timespec64(&extra_info.btime, btime);
4021 		p += sizeof(*btime);
4022 		ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4023 	}
4024 
4025 	if (msg_version >= 11) {
4026 		/* version >= 10 */
4027 		ceph_decode_skip_32(&p, end, bad); // flags
4028 		/* version >= 11 */
4029 		extra_info.dirstat_valid = true;
4030 		ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4031 		ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4032 	}
4033 
4034 	/* lookup ino */
4035 	inode = ceph_find_inode(mdsc->fsc->sb, vino);
4036 	ci = ceph_inode(inode);
4037 	dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4038 	     vino.snap, inode);
4039 
4040 	mutex_lock(&session->s_mutex);
4041 	inc_session_sequence(session);
4042 	dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4043 	     (unsigned)seq);
4044 
4045 	if (!inode) {
4046 		dout(" i don't have ino %llx\n", vino.ino);
4047 
4048 		if (op == CEPH_CAP_OP_IMPORT) {
4049 			cap = ceph_get_cap(mdsc, NULL);
4050 			cap->cap_ino = vino.ino;
4051 			cap->queue_release = 1;
4052 			cap->cap_id = le64_to_cpu(h->cap_id);
4053 			cap->mseq = mseq;
4054 			cap->seq = seq;
4055 			cap->issue_seq = seq;
4056 			spin_lock(&session->s_cap_lock);
4057 			__ceph_queue_cap_release(session, cap);
4058 			spin_unlock(&session->s_cap_lock);
4059 		}
4060 		goto flush_cap_releases;
4061 	}
4062 
4063 	/* these will work even if we don't have a cap yet */
4064 	switch (op) {
4065 	case CEPH_CAP_OP_FLUSHSNAP_ACK:
4066 		handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4067 					 h, session);
4068 		goto done;
4069 
4070 	case CEPH_CAP_OP_EXPORT:
4071 		handle_cap_export(inode, h, peer, session);
4072 		goto done_unlocked;
4073 
4074 	case CEPH_CAP_OP_IMPORT:
4075 		realm = NULL;
4076 		if (snaptrace_len) {
4077 			down_write(&mdsc->snap_rwsem);
4078 			ceph_update_snap_trace(mdsc, snaptrace,
4079 					       snaptrace + snaptrace_len,
4080 					       false, &realm);
4081 			downgrade_write(&mdsc->snap_rwsem);
4082 		} else {
4083 			down_read(&mdsc->snap_rwsem);
4084 		}
4085 		spin_lock(&ci->i_ceph_lock);
4086 		handle_cap_import(mdsc, inode, h, peer, session,
4087 				  &cap, &extra_info.issued);
4088 		handle_cap_grant(inode, session, cap,
4089 				 h, msg->middle, &extra_info);
4090 		if (realm)
4091 			ceph_put_snap_realm(mdsc, realm);
4092 		goto done_unlocked;
4093 	}
4094 
4095 	/* the rest require a cap */
4096 	spin_lock(&ci->i_ceph_lock);
4097 	cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4098 	if (!cap) {
4099 		dout(" no cap on %p ino %llx.%llx from mds%d\n",
4100 		     inode, ceph_ino(inode), ceph_snap(inode),
4101 		     session->s_mds);
4102 		spin_unlock(&ci->i_ceph_lock);
4103 		goto flush_cap_releases;
4104 	}
4105 
4106 	/* note that each of these drops i_ceph_lock for us */
4107 	switch (op) {
4108 	case CEPH_CAP_OP_REVOKE:
4109 	case CEPH_CAP_OP_GRANT:
4110 		__ceph_caps_issued(ci, &extra_info.issued);
4111 		extra_info.issued |= __ceph_caps_dirty(ci);
4112 		handle_cap_grant(inode, session, cap,
4113 				 h, msg->middle, &extra_info);
4114 		goto done_unlocked;
4115 
4116 	case CEPH_CAP_OP_FLUSH_ACK:
4117 		handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4118 				     h, session, cap);
4119 		break;
4120 
4121 	case CEPH_CAP_OP_TRUNC:
4122 		queue_trunc = handle_cap_trunc(inode, h, session);
4123 		spin_unlock(&ci->i_ceph_lock);
4124 		if (queue_trunc)
4125 			ceph_queue_vmtruncate(inode);
4126 		break;
4127 
4128 	default:
4129 		spin_unlock(&ci->i_ceph_lock);
4130 		pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4131 		       ceph_cap_op_name(op));
4132 	}
4133 
4134 done:
4135 	mutex_unlock(&session->s_mutex);
4136 done_unlocked:
4137 	ceph_put_string(extra_info.pool_ns);
4138 	iput(inode);
4139 	return;
4140 
4141 flush_cap_releases:
4142 	/*
4143 	 * send any cap release message to try to move things
4144 	 * along for the mds (who clearly thinks we still have this
4145 	 * cap).
4146 	 */
4147 	ceph_flush_cap_releases(mdsc, session);
4148 	goto done;
4149 
4150 bad:
4151 	pr_err("ceph_handle_caps: corrupt message\n");
4152 	ceph_msg_dump(msg);
4153 	return;
4154 }
4155 
4156 /*
4157  * Delayed work handler to process end of delayed cap release LRU list.
4158  *
4159  * If new caps are added to the list while processing it, these won't get
4160  * processed in this run.  In this case, the ci->i_hold_caps_max will be
4161  * returned so that the work can be scheduled accordingly.
4162  */
4163 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4164 {
4165 	struct inode *inode;
4166 	struct ceph_inode_info *ci;
4167 	struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4168 	unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4169 	unsigned long loop_start = jiffies;
4170 	unsigned long delay = 0;
4171 
4172 	dout("check_delayed_caps\n");
4173 	spin_lock(&mdsc->cap_delay_lock);
4174 	while (!list_empty(&mdsc->cap_delay_list)) {
4175 		ci = list_first_entry(&mdsc->cap_delay_list,
4176 				      struct ceph_inode_info,
4177 				      i_cap_delay_list);
4178 		if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4179 			dout("%s caps added recently.  Exiting loop", __func__);
4180 			delay = ci->i_hold_caps_max;
4181 			break;
4182 		}
4183 		if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4184 		    time_before(jiffies, ci->i_hold_caps_max))
4185 			break;
4186 		list_del_init(&ci->i_cap_delay_list);
4187 
4188 		inode = igrab(&ci->vfs_inode);
4189 		if (inode) {
4190 			spin_unlock(&mdsc->cap_delay_lock);
4191 			dout("check_delayed_caps on %p\n", inode);
4192 			ceph_check_caps(ci, 0, NULL);
4193 			iput(inode);
4194 			spin_lock(&mdsc->cap_delay_lock);
4195 		}
4196 	}
4197 	spin_unlock(&mdsc->cap_delay_lock);
4198 
4199 	return delay;
4200 }
4201 
4202 /*
4203  * Flush all dirty caps to the mds
4204  */
4205 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4206 {
4207 	struct ceph_mds_client *mdsc = s->s_mdsc;
4208 	struct ceph_inode_info *ci;
4209 	struct inode *inode;
4210 
4211 	dout("flush_dirty_caps\n");
4212 	spin_lock(&mdsc->cap_dirty_lock);
4213 	while (!list_empty(&s->s_cap_dirty)) {
4214 		ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4215 				      i_dirty_item);
4216 		inode = &ci->vfs_inode;
4217 		ihold(inode);
4218 		dout("flush_dirty_caps %p\n", inode);
4219 		spin_unlock(&mdsc->cap_dirty_lock);
4220 		ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4221 		iput(inode);
4222 		spin_lock(&mdsc->cap_dirty_lock);
4223 	}
4224 	spin_unlock(&mdsc->cap_dirty_lock);
4225 	dout("flush_dirty_caps done\n");
4226 }
4227 
4228 static void iterate_sessions(struct ceph_mds_client *mdsc,
4229 			     void (*cb)(struct ceph_mds_session *))
4230 {
4231 	int mds;
4232 
4233 	mutex_lock(&mdsc->mutex);
4234 	for (mds = 0; mds < mdsc->max_sessions; ++mds) {
4235 		struct ceph_mds_session *s;
4236 
4237 		if (!mdsc->sessions[mds])
4238 			continue;
4239 
4240 		s = ceph_get_mds_session(mdsc->sessions[mds]);
4241 		if (!s)
4242 			continue;
4243 
4244 		mutex_unlock(&mdsc->mutex);
4245 		cb(s);
4246 		ceph_put_mds_session(s);
4247 		mutex_lock(&mdsc->mutex);
4248 	}
4249 	mutex_unlock(&mdsc->mutex);
4250 }
4251 
4252 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4253 {
4254 	iterate_sessions(mdsc, flush_dirty_session_caps);
4255 }
4256 
4257 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4258 			struct ceph_mds_client *mdsc, int fmode)
4259 {
4260 	unsigned long now = jiffies;
4261 	if (fmode & CEPH_FILE_MODE_RD)
4262 		ci->i_last_rd = now;
4263 	if (fmode & CEPH_FILE_MODE_WR)
4264 		ci->i_last_wr = now;
4265 	/* queue periodic check */
4266 	if (fmode &&
4267 	    __ceph_is_any_real_caps(ci) &&
4268 	    list_empty(&ci->i_cap_delay_list))
4269 		__cap_delay_requeue(mdsc, ci);
4270 }
4271 
4272 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4273 {
4274 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4275 	int bits = (fmode << 1) | 1;
4276 	bool is_opened = false;
4277 	int i;
4278 
4279 	if (count == 1)
4280 		atomic64_inc(&mdsc->metric.opened_files);
4281 
4282 	spin_lock(&ci->i_ceph_lock);
4283 	for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4284 		if (bits & (1 << i))
4285 			ci->i_nr_by_mode[i] += count;
4286 
4287 		/*
4288 		 * If any of the mode ref is larger than 1,
4289 		 * that means it has been already opened by
4290 		 * others. Just skip checking the PIN ref.
4291 		 */
4292 		if (i && ci->i_nr_by_mode[i] > 1)
4293 			is_opened = true;
4294 	}
4295 
4296 	if (!is_opened)
4297 		percpu_counter_inc(&mdsc->metric.opened_inodes);
4298 	spin_unlock(&ci->i_ceph_lock);
4299 }
4300 
4301 /*
4302  * Drop open file reference.  If we were the last open file,
4303  * we may need to release capabilities to the MDS (or schedule
4304  * their delayed release).
4305  */
4306 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4307 {
4308 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4309 	int bits = (fmode << 1) | 1;
4310 	bool is_closed = true;
4311 	int i;
4312 
4313 	if (count == 1)
4314 		atomic64_dec(&mdsc->metric.opened_files);
4315 
4316 	spin_lock(&ci->i_ceph_lock);
4317 	for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4318 		if (bits & (1 << i)) {
4319 			BUG_ON(ci->i_nr_by_mode[i] < count);
4320 			ci->i_nr_by_mode[i] -= count;
4321 		}
4322 
4323 		/*
4324 		 * If any of the mode ref is not 0 after
4325 		 * decreased, that means it is still opened
4326 		 * by others. Just skip checking the PIN ref.
4327 		 */
4328 		if (i && ci->i_nr_by_mode[i])
4329 			is_closed = false;
4330 	}
4331 
4332 	if (is_closed)
4333 		percpu_counter_dec(&mdsc->metric.opened_inodes);
4334 	spin_unlock(&ci->i_ceph_lock);
4335 }
4336 
4337 /*
4338  * For a soon-to-be unlinked file, drop the LINK caps. If it
4339  * looks like the link count will hit 0, drop any other caps (other
4340  * than PIN) we don't specifically want (due to the file still being
4341  * open).
4342  */
4343 int ceph_drop_caps_for_unlink(struct inode *inode)
4344 {
4345 	struct ceph_inode_info *ci = ceph_inode(inode);
4346 	int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4347 
4348 	spin_lock(&ci->i_ceph_lock);
4349 	if (inode->i_nlink == 1) {
4350 		drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4351 
4352 		if (__ceph_caps_dirty(ci)) {
4353 			struct ceph_mds_client *mdsc =
4354 				ceph_inode_to_client(inode)->mdsc;
4355 			__cap_delay_requeue_front(mdsc, ci);
4356 		}
4357 	}
4358 	spin_unlock(&ci->i_ceph_lock);
4359 	return drop;
4360 }
4361 
4362 /*
4363  * Helpers for embedding cap and dentry lease releases into mds
4364  * requests.
4365  *
4366  * @force is used by dentry_release (below) to force inclusion of a
4367  * record for the directory inode, even when there aren't any caps to
4368  * drop.
4369  */
4370 int ceph_encode_inode_release(void **p, struct inode *inode,
4371 			      int mds, int drop, int unless, int force)
4372 {
4373 	struct ceph_inode_info *ci = ceph_inode(inode);
4374 	struct ceph_cap *cap;
4375 	struct ceph_mds_request_release *rel = *p;
4376 	int used, dirty;
4377 	int ret = 0;
4378 
4379 	spin_lock(&ci->i_ceph_lock);
4380 	used = __ceph_caps_used(ci);
4381 	dirty = __ceph_caps_dirty(ci);
4382 
4383 	dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4384 	     inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4385 	     ceph_cap_string(unless));
4386 
4387 	/* only drop unused, clean caps */
4388 	drop &= ~(used | dirty);
4389 
4390 	cap = __get_cap_for_mds(ci, mds);
4391 	if (cap && __cap_is_valid(cap)) {
4392 		unless &= cap->issued;
4393 		if (unless) {
4394 			if (unless & CEPH_CAP_AUTH_EXCL)
4395 				drop &= ~CEPH_CAP_AUTH_SHARED;
4396 			if (unless & CEPH_CAP_LINK_EXCL)
4397 				drop &= ~CEPH_CAP_LINK_SHARED;
4398 			if (unless & CEPH_CAP_XATTR_EXCL)
4399 				drop &= ~CEPH_CAP_XATTR_SHARED;
4400 			if (unless & CEPH_CAP_FILE_EXCL)
4401 				drop &= ~CEPH_CAP_FILE_SHARED;
4402 		}
4403 
4404 		if (force || (cap->issued & drop)) {
4405 			if (cap->issued & drop) {
4406 				int wanted = __ceph_caps_wanted(ci);
4407 				dout("encode_inode_release %p cap %p "
4408 				     "%s -> %s, wanted %s -> %s\n", inode, cap,
4409 				     ceph_cap_string(cap->issued),
4410 				     ceph_cap_string(cap->issued & ~drop),
4411 				     ceph_cap_string(cap->mds_wanted),
4412 				     ceph_cap_string(wanted));
4413 
4414 				cap->issued &= ~drop;
4415 				cap->implemented &= ~drop;
4416 				cap->mds_wanted = wanted;
4417 				if (cap == ci->i_auth_cap &&
4418 				    !(wanted & CEPH_CAP_ANY_FILE_WR))
4419 					ci->i_requested_max_size = 0;
4420 			} else {
4421 				dout("encode_inode_release %p cap %p %s"
4422 				     " (force)\n", inode, cap,
4423 				     ceph_cap_string(cap->issued));
4424 			}
4425 
4426 			rel->ino = cpu_to_le64(ceph_ino(inode));
4427 			rel->cap_id = cpu_to_le64(cap->cap_id);
4428 			rel->seq = cpu_to_le32(cap->seq);
4429 			rel->issue_seq = cpu_to_le32(cap->issue_seq);
4430 			rel->mseq = cpu_to_le32(cap->mseq);
4431 			rel->caps = cpu_to_le32(cap->implemented);
4432 			rel->wanted = cpu_to_le32(cap->mds_wanted);
4433 			rel->dname_len = 0;
4434 			rel->dname_seq = 0;
4435 			*p += sizeof(*rel);
4436 			ret = 1;
4437 		} else {
4438 			dout("encode_inode_release %p cap %p %s (noop)\n",
4439 			     inode, cap, ceph_cap_string(cap->issued));
4440 		}
4441 	}
4442 	spin_unlock(&ci->i_ceph_lock);
4443 	return ret;
4444 }
4445 
4446 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4447 			       struct inode *dir,
4448 			       int mds, int drop, int unless)
4449 {
4450 	struct dentry *parent = NULL;
4451 	struct ceph_mds_request_release *rel = *p;
4452 	struct ceph_dentry_info *di = ceph_dentry(dentry);
4453 	int force = 0;
4454 	int ret;
4455 
4456 	/*
4457 	 * force an record for the directory caps if we have a dentry lease.
4458 	 * this is racy (can't take i_ceph_lock and d_lock together), but it
4459 	 * doesn't have to be perfect; the mds will revoke anything we don't
4460 	 * release.
4461 	 */
4462 	spin_lock(&dentry->d_lock);
4463 	if (di->lease_session && di->lease_session->s_mds == mds)
4464 		force = 1;
4465 	if (!dir) {
4466 		parent = dget(dentry->d_parent);
4467 		dir = d_inode(parent);
4468 	}
4469 	spin_unlock(&dentry->d_lock);
4470 
4471 	ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4472 	dput(parent);
4473 
4474 	spin_lock(&dentry->d_lock);
4475 	if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4476 		dout("encode_dentry_release %p mds%d seq %d\n",
4477 		     dentry, mds, (int)di->lease_seq);
4478 		rel->dname_len = cpu_to_le32(dentry->d_name.len);
4479 		memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4480 		*p += dentry->d_name.len;
4481 		rel->dname_seq = cpu_to_le32(di->lease_seq);
4482 		__ceph_mdsc_drop_dentry_lease(dentry);
4483 	}
4484 	spin_unlock(&dentry->d_lock);
4485 	return ret;
4486 }
4487