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 
gcap_string(char * s,int c)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 
ceph_cap_string(int caps)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 
ceph_caps_init(struct ceph_mds_client * mdsc)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 
ceph_caps_finalize(struct ceph_mds_client * mdsc)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 
ceph_adjust_caps_max_min(struct ceph_mds_client * mdsc,struct ceph_mount_options * fsopt)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 
__ceph_unreserve_caps(struct ceph_mds_client * mdsc,int nr_caps)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  */
ceph_reserve_caps(struct ceph_mds_client * mdsc,struct ceph_cap_reservation * ctx,int need)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 
ceph_unreserve_caps(struct ceph_mds_client * mdsc,struct ceph_cap_reservation * ctx)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 
ceph_get_cap(struct ceph_mds_client * mdsc,struct ceph_cap_reservation * ctx)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 
ceph_put_cap(struct ceph_mds_client * mdsc,struct ceph_cap * cap)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 
ceph_reservation_status(struct ceph_fs_client * fsc,int * total,int * avail,int * used,int * reserved,int * min)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  */
__get_cap_for_mds(struct ceph_inode_info * ci,int mds)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 
ceph_get_cap_for_mds(struct ceph_inode_info * ci,int mds)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  */
__insert_cap_node(struct ceph_inode_info * ci,struct ceph_cap * new)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  */
__cap_set_timeouts(struct ceph_mds_client * mdsc,struct ceph_inode_info * ci)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->netfs.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  */
__cap_delay_requeue(struct ceph_mds_client * mdsc,struct ceph_inode_info * ci)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->netfs.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  */
__cap_delay_requeue_front(struct ceph_mds_client * mdsc,struct ceph_inode_info * ci)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->netfs.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  */
__cap_delay_cancel(struct ceph_mds_client * mdsc,struct ceph_inode_info * ci)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->netfs.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. */
__check_cap_issue(struct ceph_inode_info * ci,struct ceph_cap * cap,unsigned issued)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->netfs.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->netfs.inode.i_mode)) {
587 			dout(" marking %p NOT complete\n", &ci->netfs.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->netfs.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  */
change_auth_cap_ses(struct ceph_inode_info * ci,struct ceph_mds_session * session)605 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  */
ceph_add_cap(struct inode * inode,struct ceph_mds_session * session,u64 cap_id,unsigned issued,unsigned wanted,unsigned seq,unsigned mseq,u64 realmino,int flags,struct ceph_cap ** new_cap)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 			ceph_change_snap_realm(inode, realm);
708 		else
709 			WARN(1, "%s: couldn't find snap realm 0x%llx (ino 0x%llx oldrealm 0x%llx)\n",
710 			     __func__, realmino, ci->i_vino.ino,
711 			     ci->i_snap_realm ? ci->i_snap_realm->ino : 0);
712 	}
713 
714 	__check_cap_issue(ci, cap, issued);
715 
716 	/*
717 	 * If we are issued caps we don't want, or the mds' wanted
718 	 * value appears to be off, queue a check so we'll release
719 	 * later and/or update the mds wanted value.
720 	 */
721 	actual_wanted = __ceph_caps_wanted(ci);
722 	if ((wanted & ~actual_wanted) ||
723 	    (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
724 		dout(" issued %s, mds wanted %s, actual %s, queueing\n",
725 		     ceph_cap_string(issued), ceph_cap_string(wanted),
726 		     ceph_cap_string(actual_wanted));
727 		__cap_delay_requeue(mdsc, ci);
728 	}
729 
730 	if (flags & CEPH_CAP_FLAG_AUTH) {
731 		if (!ci->i_auth_cap ||
732 		    ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
733 			if (ci->i_auth_cap &&
734 			    ci->i_auth_cap->session != cap->session)
735 				change_auth_cap_ses(ci, cap->session);
736 			ci->i_auth_cap = cap;
737 			cap->mds_wanted = wanted;
738 		}
739 	} else {
740 		WARN_ON(ci->i_auth_cap == cap);
741 	}
742 
743 	dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
744 	     inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
745 	     ceph_cap_string(issued|cap->issued), seq, mds);
746 	cap->cap_id = cap_id;
747 	cap->issued = issued;
748 	cap->implemented |= issued;
749 	if (ceph_seq_cmp(mseq, cap->mseq) > 0)
750 		cap->mds_wanted = wanted;
751 	else
752 		cap->mds_wanted |= wanted;
753 	cap->seq = seq;
754 	cap->issue_seq = seq;
755 	cap->mseq = mseq;
756 	cap->cap_gen = gen;
757 	wake_up_all(&ci->i_cap_wq);
758 }
759 
760 /*
761  * Return true if cap has not timed out and belongs to the current
762  * generation of the MDS session (i.e. has not gone 'stale' due to
763  * us losing touch with the mds).
764  */
__cap_is_valid(struct ceph_cap * cap)765 static int __cap_is_valid(struct ceph_cap *cap)
766 {
767 	unsigned long ttl;
768 	u32 gen;
769 
770 	gen = atomic_read(&cap->session->s_cap_gen);
771 	ttl = cap->session->s_cap_ttl;
772 
773 	if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
774 		dout("__cap_is_valid %p cap %p issued %s "
775 		     "but STALE (gen %u vs %u)\n", &cap->ci->netfs.inode,
776 		     cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
777 		return 0;
778 	}
779 
780 	return 1;
781 }
782 
783 /*
784  * Return set of valid cap bits issued to us.  Note that caps time
785  * out, and may be invalidated in bulk if the client session times out
786  * and session->s_cap_gen is bumped.
787  */
__ceph_caps_issued(struct ceph_inode_info * ci,int * implemented)788 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
789 {
790 	int have = ci->i_snap_caps;
791 	struct ceph_cap *cap;
792 	struct rb_node *p;
793 
794 	if (implemented)
795 		*implemented = 0;
796 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
797 		cap = rb_entry(p, struct ceph_cap, ci_node);
798 		if (!__cap_is_valid(cap))
799 			continue;
800 		dout("__ceph_caps_issued %p cap %p issued %s\n",
801 		     &ci->netfs.inode, cap, ceph_cap_string(cap->issued));
802 		have |= cap->issued;
803 		if (implemented)
804 			*implemented |= cap->implemented;
805 	}
806 	/*
807 	 * exclude caps issued by non-auth MDS, but are been revoking
808 	 * by the auth MDS. The non-auth MDS should be revoking/exporting
809 	 * these caps, but the message is delayed.
810 	 */
811 	if (ci->i_auth_cap) {
812 		cap = ci->i_auth_cap;
813 		have &= ~cap->implemented | cap->issued;
814 	}
815 	return have;
816 }
817 
818 /*
819  * Get cap bits issued by caps other than @ocap
820  */
__ceph_caps_issued_other(struct ceph_inode_info * ci,struct ceph_cap * ocap)821 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
822 {
823 	int have = ci->i_snap_caps;
824 	struct ceph_cap *cap;
825 	struct rb_node *p;
826 
827 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
828 		cap = rb_entry(p, struct ceph_cap, ci_node);
829 		if (cap == ocap)
830 			continue;
831 		if (!__cap_is_valid(cap))
832 			continue;
833 		have |= cap->issued;
834 	}
835 	return have;
836 }
837 
838 /*
839  * Move a cap to the end of the LRU (oldest caps at list head, newest
840  * at list tail).
841  */
__touch_cap(struct ceph_cap * cap)842 static void __touch_cap(struct ceph_cap *cap)
843 {
844 	struct ceph_mds_session *s = cap->session;
845 
846 	spin_lock(&s->s_cap_lock);
847 	if (!s->s_cap_iterator) {
848 		dout("__touch_cap %p cap %p mds%d\n", &cap->ci->netfs.inode, cap,
849 		     s->s_mds);
850 		list_move_tail(&cap->session_caps, &s->s_caps);
851 	} else {
852 		dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
853 		     &cap->ci->netfs.inode, cap, s->s_mds);
854 	}
855 	spin_unlock(&s->s_cap_lock);
856 }
857 
858 /*
859  * Check if we hold the given mask.  If so, move the cap(s) to the
860  * front of their respective LRUs.  (This is the preferred way for
861  * callers to check for caps they want.)
862  */
__ceph_caps_issued_mask(struct ceph_inode_info * ci,int mask,int touch)863 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
864 {
865 	struct ceph_cap *cap;
866 	struct rb_node *p;
867 	int have = ci->i_snap_caps;
868 
869 	if ((have & mask) == mask) {
870 		dout("__ceph_caps_issued_mask ino 0x%llx snap issued %s"
871 		     " (mask %s)\n", ceph_ino(&ci->netfs.inode),
872 		     ceph_cap_string(have),
873 		     ceph_cap_string(mask));
874 		return 1;
875 	}
876 
877 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
878 		cap = rb_entry(p, struct ceph_cap, ci_node);
879 		if (!__cap_is_valid(cap))
880 			continue;
881 		if ((cap->issued & mask) == mask) {
882 			dout("__ceph_caps_issued_mask ino 0x%llx cap %p issued %s"
883 			     " (mask %s)\n", ceph_ino(&ci->netfs.inode), cap,
884 			     ceph_cap_string(cap->issued),
885 			     ceph_cap_string(mask));
886 			if (touch)
887 				__touch_cap(cap);
888 			return 1;
889 		}
890 
891 		/* does a combination of caps satisfy mask? */
892 		have |= cap->issued;
893 		if ((have & mask) == mask) {
894 			dout("__ceph_caps_issued_mask ino 0x%llx combo issued %s"
895 			     " (mask %s)\n", ceph_ino(&ci->netfs.inode),
896 			     ceph_cap_string(cap->issued),
897 			     ceph_cap_string(mask));
898 			if (touch) {
899 				struct rb_node *q;
900 
901 				/* touch this + preceding caps */
902 				__touch_cap(cap);
903 				for (q = rb_first(&ci->i_caps); q != p;
904 				     q = rb_next(q)) {
905 					cap = rb_entry(q, struct ceph_cap,
906 						       ci_node);
907 					if (!__cap_is_valid(cap))
908 						continue;
909 					if (cap->issued & mask)
910 						__touch_cap(cap);
911 				}
912 			}
913 			return 1;
914 		}
915 	}
916 
917 	return 0;
918 }
919 
__ceph_caps_issued_mask_metric(struct ceph_inode_info * ci,int mask,int touch)920 int __ceph_caps_issued_mask_metric(struct ceph_inode_info *ci, int mask,
921 				   int touch)
922 {
923 	struct ceph_fs_client *fsc = ceph_sb_to_client(ci->netfs.inode.i_sb);
924 	int r;
925 
926 	r = __ceph_caps_issued_mask(ci, mask, touch);
927 	if (r)
928 		ceph_update_cap_hit(&fsc->mdsc->metric);
929 	else
930 		ceph_update_cap_mis(&fsc->mdsc->metric);
931 	return r;
932 }
933 
934 /*
935  * Return true if mask caps are currently being revoked by an MDS.
936  */
__ceph_caps_revoking_other(struct ceph_inode_info * ci,struct ceph_cap * ocap,int mask)937 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
938 			       struct ceph_cap *ocap, int mask)
939 {
940 	struct ceph_cap *cap;
941 	struct rb_node *p;
942 
943 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
944 		cap = rb_entry(p, struct ceph_cap, ci_node);
945 		if (cap != ocap &&
946 		    (cap->implemented & ~cap->issued & mask))
947 			return 1;
948 	}
949 	return 0;
950 }
951 
ceph_caps_revoking(struct ceph_inode_info * ci,int mask)952 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
953 {
954 	struct inode *inode = &ci->netfs.inode;
955 	int ret;
956 
957 	spin_lock(&ci->i_ceph_lock);
958 	ret = __ceph_caps_revoking_other(ci, NULL, mask);
959 	spin_unlock(&ci->i_ceph_lock);
960 	dout("ceph_caps_revoking %p %s = %d\n", inode,
961 	     ceph_cap_string(mask), ret);
962 	return ret;
963 }
964 
__ceph_caps_used(struct ceph_inode_info * ci)965 int __ceph_caps_used(struct ceph_inode_info *ci)
966 {
967 	int used = 0;
968 	if (ci->i_pin_ref)
969 		used |= CEPH_CAP_PIN;
970 	if (ci->i_rd_ref)
971 		used |= CEPH_CAP_FILE_RD;
972 	if (ci->i_rdcache_ref ||
973 	    (S_ISREG(ci->netfs.inode.i_mode) &&
974 	     ci->netfs.inode.i_data.nrpages))
975 		used |= CEPH_CAP_FILE_CACHE;
976 	if (ci->i_wr_ref)
977 		used |= CEPH_CAP_FILE_WR;
978 	if (ci->i_wb_ref || ci->i_wrbuffer_ref)
979 		used |= CEPH_CAP_FILE_BUFFER;
980 	if (ci->i_fx_ref)
981 		used |= CEPH_CAP_FILE_EXCL;
982 	return used;
983 }
984 
985 #define FMODE_WAIT_BIAS 1000
986 
987 /*
988  * wanted, by virtue of open file modes
989  */
__ceph_caps_file_wanted(struct ceph_inode_info * ci)990 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
991 {
992 	const int PIN_SHIFT = ffs(CEPH_FILE_MODE_PIN);
993 	const int RD_SHIFT = ffs(CEPH_FILE_MODE_RD);
994 	const int WR_SHIFT = ffs(CEPH_FILE_MODE_WR);
995 	const int LAZY_SHIFT = ffs(CEPH_FILE_MODE_LAZY);
996 	struct ceph_mount_options *opt =
997 		ceph_inode_to_client(&ci->netfs.inode)->mount_options;
998 	unsigned long used_cutoff = jiffies - opt->caps_wanted_delay_max * HZ;
999 	unsigned long idle_cutoff = jiffies - opt->caps_wanted_delay_min * HZ;
1000 
1001 	if (S_ISDIR(ci->netfs.inode.i_mode)) {
1002 		int want = 0;
1003 
1004 		/* use used_cutoff here, to keep dir's wanted caps longer */
1005 		if (ci->i_nr_by_mode[RD_SHIFT] > 0 ||
1006 		    time_after(ci->i_last_rd, used_cutoff))
1007 			want |= CEPH_CAP_ANY_SHARED;
1008 
1009 		if (ci->i_nr_by_mode[WR_SHIFT] > 0 ||
1010 		    time_after(ci->i_last_wr, used_cutoff)) {
1011 			want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1012 			if (opt->flags & CEPH_MOUNT_OPT_ASYNC_DIROPS)
1013 				want |= CEPH_CAP_ANY_DIR_OPS;
1014 		}
1015 
1016 		if (want || ci->i_nr_by_mode[PIN_SHIFT] > 0)
1017 			want |= CEPH_CAP_PIN;
1018 
1019 		return want;
1020 	} else {
1021 		int bits = 0;
1022 
1023 		if (ci->i_nr_by_mode[RD_SHIFT] > 0) {
1024 			if (ci->i_nr_by_mode[RD_SHIFT] >= FMODE_WAIT_BIAS ||
1025 			    time_after(ci->i_last_rd, used_cutoff))
1026 				bits |= 1 << RD_SHIFT;
1027 		} else if (time_after(ci->i_last_rd, idle_cutoff)) {
1028 			bits |= 1 << RD_SHIFT;
1029 		}
1030 
1031 		if (ci->i_nr_by_mode[WR_SHIFT] > 0) {
1032 			if (ci->i_nr_by_mode[WR_SHIFT] >= FMODE_WAIT_BIAS ||
1033 			    time_after(ci->i_last_wr, used_cutoff))
1034 				bits |= 1 << WR_SHIFT;
1035 		} else if (time_after(ci->i_last_wr, idle_cutoff)) {
1036 			bits |= 1 << WR_SHIFT;
1037 		}
1038 
1039 		/* check lazyio only when read/write is wanted */
1040 		if ((bits & (CEPH_FILE_MODE_RDWR << 1)) &&
1041 		    ci->i_nr_by_mode[LAZY_SHIFT] > 0)
1042 			bits |= 1 << LAZY_SHIFT;
1043 
1044 		return bits ? ceph_caps_for_mode(bits >> 1) : 0;
1045 	}
1046 }
1047 
1048 /*
1049  * wanted, by virtue of open file modes AND cap refs (buffered/cached data)
1050  */
__ceph_caps_wanted(struct ceph_inode_info * ci)1051 int __ceph_caps_wanted(struct ceph_inode_info *ci)
1052 {
1053 	int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci);
1054 	if (S_ISDIR(ci->netfs.inode.i_mode)) {
1055 		/* we want EXCL if holding caps of dir ops */
1056 		if (w & CEPH_CAP_ANY_DIR_OPS)
1057 			w |= CEPH_CAP_FILE_EXCL;
1058 	} else {
1059 		/* we want EXCL if dirty data */
1060 		if (w & CEPH_CAP_FILE_BUFFER)
1061 			w |= CEPH_CAP_FILE_EXCL;
1062 	}
1063 	return w;
1064 }
1065 
1066 /*
1067  * Return caps we have registered with the MDS(s) as 'wanted'.
1068  */
__ceph_caps_mds_wanted(struct ceph_inode_info * ci,bool check)1069 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
1070 {
1071 	struct ceph_cap *cap;
1072 	struct rb_node *p;
1073 	int mds_wanted = 0;
1074 
1075 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1076 		cap = rb_entry(p, struct ceph_cap, ci_node);
1077 		if (check && !__cap_is_valid(cap))
1078 			continue;
1079 		if (cap == ci->i_auth_cap)
1080 			mds_wanted |= cap->mds_wanted;
1081 		else
1082 			mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1083 	}
1084 	return mds_wanted;
1085 }
1086 
ceph_is_any_caps(struct inode * inode)1087 int ceph_is_any_caps(struct inode *inode)
1088 {
1089 	struct ceph_inode_info *ci = ceph_inode(inode);
1090 	int ret;
1091 
1092 	spin_lock(&ci->i_ceph_lock);
1093 	ret = __ceph_is_any_real_caps(ci);
1094 	spin_unlock(&ci->i_ceph_lock);
1095 
1096 	return ret;
1097 }
1098 
1099 /*
1100  * Remove a cap.  Take steps to deal with a racing iterate_session_caps.
1101  *
1102  * caller should hold i_ceph_lock.
1103  * caller will not hold session s_mutex if called from destroy_inode.
1104  */
__ceph_remove_cap(struct ceph_cap * cap,bool queue_release)1105 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1106 {
1107 	struct ceph_mds_session *session = cap->session;
1108 	struct ceph_inode_info *ci = cap->ci;
1109 	struct ceph_mds_client *mdsc;
1110 	int removed = 0;
1111 
1112 	/* 'ci' being NULL means the remove have already occurred */
1113 	if (!ci) {
1114 		dout("%s: cap inode is NULL\n", __func__);
1115 		return;
1116 	}
1117 
1118 	lockdep_assert_held(&ci->i_ceph_lock);
1119 
1120 	dout("__ceph_remove_cap %p from %p\n", cap, &ci->netfs.inode);
1121 
1122 	mdsc = ceph_inode_to_client(&ci->netfs.inode)->mdsc;
1123 
1124 	/* remove from inode's cap rbtree, and clear auth cap */
1125 	rb_erase(&cap->ci_node, &ci->i_caps);
1126 	if (ci->i_auth_cap == cap)
1127 		ci->i_auth_cap = NULL;
1128 
1129 	/* remove from session list */
1130 	spin_lock(&session->s_cap_lock);
1131 	if (session->s_cap_iterator == cap) {
1132 		/* not yet, we are iterating over this very cap */
1133 		dout("__ceph_remove_cap  delaying %p removal from session %p\n",
1134 		     cap, cap->session);
1135 	} else {
1136 		list_del_init(&cap->session_caps);
1137 		session->s_nr_caps--;
1138 		atomic64_dec(&mdsc->metric.total_caps);
1139 		cap->session = NULL;
1140 		removed = 1;
1141 	}
1142 	/* protect backpointer with s_cap_lock: see iterate_session_caps */
1143 	cap->ci = NULL;
1144 
1145 	/*
1146 	 * s_cap_reconnect is protected by s_cap_lock. no one changes
1147 	 * s_cap_gen while session is in the reconnect state.
1148 	 */
1149 	if (queue_release &&
1150 	    (!session->s_cap_reconnect ||
1151 	     cap->cap_gen == atomic_read(&session->s_cap_gen))) {
1152 		cap->queue_release = 1;
1153 		if (removed) {
1154 			__ceph_queue_cap_release(session, cap);
1155 			removed = 0;
1156 		}
1157 	} else {
1158 		cap->queue_release = 0;
1159 	}
1160 	cap->cap_ino = ci->i_vino.ino;
1161 
1162 	spin_unlock(&session->s_cap_lock);
1163 
1164 	if (removed)
1165 		ceph_put_cap(mdsc, cap);
1166 
1167 	if (!__ceph_is_any_real_caps(ci)) {
1168 		/* when reconnect denied, we remove session caps forcibly,
1169 		 * i_wr_ref can be non-zero. If there are ongoing write,
1170 		 * keep i_snap_realm.
1171 		 */
1172 		if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1173 			ceph_change_snap_realm(&ci->netfs.inode, NULL);
1174 
1175 		__cap_delay_cancel(mdsc, ci);
1176 	}
1177 }
1178 
ceph_remove_cap(struct ceph_cap * cap,bool queue_release)1179 void ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1180 {
1181 	struct ceph_inode_info *ci = cap->ci;
1182 	struct ceph_fs_client *fsc;
1183 
1184 	/* 'ci' being NULL means the remove have already occurred */
1185 	if (!ci) {
1186 		dout("%s: cap inode is NULL\n", __func__);
1187 		return;
1188 	}
1189 
1190 	lockdep_assert_held(&ci->i_ceph_lock);
1191 
1192 	fsc = ceph_inode_to_client(&ci->netfs.inode);
1193 	WARN_ON_ONCE(ci->i_auth_cap == cap &&
1194 		     !list_empty(&ci->i_dirty_item) &&
1195 		     !fsc->blocklisted &&
1196 		     !ceph_inode_is_shutdown(&ci->netfs.inode));
1197 
1198 	__ceph_remove_cap(cap, queue_release);
1199 }
1200 
1201 struct cap_msg_args {
1202 	struct ceph_mds_session	*session;
1203 	u64			ino, cid, follows;
1204 	u64			flush_tid, oldest_flush_tid, size, max_size;
1205 	u64			xattr_version;
1206 	u64			change_attr;
1207 	struct ceph_buffer	*xattr_buf;
1208 	struct ceph_buffer	*old_xattr_buf;
1209 	struct timespec64	atime, mtime, ctime, btime;
1210 	int			op, caps, wanted, dirty;
1211 	u32			seq, issue_seq, mseq, time_warp_seq;
1212 	u32			flags;
1213 	kuid_t			uid;
1214 	kgid_t			gid;
1215 	umode_t			mode;
1216 	bool			inline_data;
1217 	bool			wake;
1218 };
1219 
1220 /*
1221  * cap struct size + flock buffer size + inline version + inline data size +
1222  * osd_epoch_barrier + oldest_flush_tid
1223  */
1224 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1225 		      4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1226 
1227 /* Marshal up the cap msg to the MDS */
encode_cap_msg(struct ceph_msg * msg,struct cap_msg_args * arg)1228 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1229 {
1230 	struct ceph_mds_caps *fc;
1231 	void *p;
1232 	struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1233 
1234 	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",
1235 	     __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1236 	     ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1237 	     ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1238 	     arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1239 	     arg->size, arg->max_size, arg->xattr_version,
1240 	     arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1241 
1242 	msg->hdr.version = cpu_to_le16(10);
1243 	msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1244 
1245 	fc = msg->front.iov_base;
1246 	memset(fc, 0, sizeof(*fc));
1247 
1248 	fc->cap_id = cpu_to_le64(arg->cid);
1249 	fc->op = cpu_to_le32(arg->op);
1250 	fc->seq = cpu_to_le32(arg->seq);
1251 	fc->issue_seq = cpu_to_le32(arg->issue_seq);
1252 	fc->migrate_seq = cpu_to_le32(arg->mseq);
1253 	fc->caps = cpu_to_le32(arg->caps);
1254 	fc->wanted = cpu_to_le32(arg->wanted);
1255 	fc->dirty = cpu_to_le32(arg->dirty);
1256 	fc->ino = cpu_to_le64(arg->ino);
1257 	fc->snap_follows = cpu_to_le64(arg->follows);
1258 
1259 	fc->size = cpu_to_le64(arg->size);
1260 	fc->max_size = cpu_to_le64(arg->max_size);
1261 	ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1262 	ceph_encode_timespec64(&fc->atime, &arg->atime);
1263 	ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1264 	fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1265 
1266 	fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1267 	fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1268 	fc->mode = cpu_to_le32(arg->mode);
1269 
1270 	fc->xattr_version = cpu_to_le64(arg->xattr_version);
1271 	if (arg->xattr_buf) {
1272 		msg->middle = ceph_buffer_get(arg->xattr_buf);
1273 		fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1274 		msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1275 	}
1276 
1277 	p = fc + 1;
1278 	/* flock buffer size (version 2) */
1279 	ceph_encode_32(&p, 0);
1280 	/* inline version (version 4) */
1281 	ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1282 	/* inline data size */
1283 	ceph_encode_32(&p, 0);
1284 	/*
1285 	 * osd_epoch_barrier (version 5)
1286 	 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1287 	 * case it was recently changed
1288 	 */
1289 	ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1290 	/* oldest_flush_tid (version 6) */
1291 	ceph_encode_64(&p, arg->oldest_flush_tid);
1292 
1293 	/*
1294 	 * caller_uid/caller_gid (version 7)
1295 	 *
1296 	 * Currently, we don't properly track which caller dirtied the caps
1297 	 * last, and force a flush of them when there is a conflict. For now,
1298 	 * just set this to 0:0, to emulate how the MDS has worked up to now.
1299 	 */
1300 	ceph_encode_32(&p, 0);
1301 	ceph_encode_32(&p, 0);
1302 
1303 	/* pool namespace (version 8) (mds always ignores this) */
1304 	ceph_encode_32(&p, 0);
1305 
1306 	/* btime and change_attr (version 9) */
1307 	ceph_encode_timespec64(p, &arg->btime);
1308 	p += sizeof(struct ceph_timespec);
1309 	ceph_encode_64(&p, arg->change_attr);
1310 
1311 	/* Advisory flags (version 10) */
1312 	ceph_encode_32(&p, arg->flags);
1313 }
1314 
1315 /*
1316  * Queue cap releases when an inode is dropped from our cache.
1317  */
__ceph_remove_caps(struct ceph_inode_info * ci)1318 void __ceph_remove_caps(struct ceph_inode_info *ci)
1319 {
1320 	struct rb_node *p;
1321 
1322 	/* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1323 	 * may call __ceph_caps_issued_mask() on a freeing inode. */
1324 	spin_lock(&ci->i_ceph_lock);
1325 	p = rb_first(&ci->i_caps);
1326 	while (p) {
1327 		struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1328 		p = rb_next(p);
1329 		ceph_remove_cap(cap, true);
1330 	}
1331 	spin_unlock(&ci->i_ceph_lock);
1332 }
1333 
1334 /*
1335  * Prepare to send a cap message to an MDS. Update the cap state, and populate
1336  * the arg struct with the parameters that will need to be sent. This should
1337  * be done under the i_ceph_lock to guard against changes to cap state.
1338  *
1339  * Make note of max_size reported/requested from mds, revoked caps
1340  * that have now been implemented.
1341  */
__prep_cap(struct cap_msg_args * arg,struct ceph_cap * cap,int op,int flags,int used,int want,int retain,int flushing,u64 flush_tid,u64 oldest_flush_tid)1342 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1343 		       int op, int flags, int used, int want, int retain,
1344 		       int flushing, u64 flush_tid, u64 oldest_flush_tid)
1345 {
1346 	struct ceph_inode_info *ci = cap->ci;
1347 	struct inode *inode = &ci->netfs.inode;
1348 	int held, revoking;
1349 
1350 	lockdep_assert_held(&ci->i_ceph_lock);
1351 
1352 	held = cap->issued | cap->implemented;
1353 	revoking = cap->implemented & ~cap->issued;
1354 	retain &= ~revoking;
1355 
1356 	dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1357 	     __func__, inode, cap, cap->session,
1358 	     ceph_cap_string(held), ceph_cap_string(held & retain),
1359 	     ceph_cap_string(revoking));
1360 	BUG_ON((retain & CEPH_CAP_PIN) == 0);
1361 
1362 	ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1363 
1364 	cap->issued &= retain;  /* drop bits we don't want */
1365 	/*
1366 	 * Wake up any waiters on wanted -> needed transition. This is due to
1367 	 * the weird transition from buffered to sync IO... we need to flush
1368 	 * dirty pages _before_ allowing sync writes to avoid reordering.
1369 	 */
1370 	arg->wake = cap->implemented & ~cap->issued;
1371 	cap->implemented &= cap->issued | used;
1372 	cap->mds_wanted = want;
1373 
1374 	arg->session = cap->session;
1375 	arg->ino = ceph_vino(inode).ino;
1376 	arg->cid = cap->cap_id;
1377 	arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1378 	arg->flush_tid = flush_tid;
1379 	arg->oldest_flush_tid = oldest_flush_tid;
1380 
1381 	arg->size = i_size_read(inode);
1382 	ci->i_reported_size = arg->size;
1383 	arg->max_size = ci->i_wanted_max_size;
1384 	if (cap == ci->i_auth_cap) {
1385 		if (want & CEPH_CAP_ANY_FILE_WR)
1386 			ci->i_requested_max_size = arg->max_size;
1387 		else
1388 			ci->i_requested_max_size = 0;
1389 	}
1390 
1391 	if (flushing & CEPH_CAP_XATTR_EXCL) {
1392 		arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1393 		arg->xattr_version = ci->i_xattrs.version;
1394 		arg->xattr_buf = ci->i_xattrs.blob;
1395 	} else {
1396 		arg->xattr_buf = NULL;
1397 		arg->old_xattr_buf = NULL;
1398 	}
1399 
1400 	arg->mtime = inode->i_mtime;
1401 	arg->atime = inode->i_atime;
1402 	arg->ctime = inode->i_ctime;
1403 	arg->btime = ci->i_btime;
1404 	arg->change_attr = inode_peek_iversion_raw(inode);
1405 
1406 	arg->op = op;
1407 	arg->caps = cap->implemented;
1408 	arg->wanted = want;
1409 	arg->dirty = flushing;
1410 
1411 	arg->seq = cap->seq;
1412 	arg->issue_seq = cap->issue_seq;
1413 	arg->mseq = cap->mseq;
1414 	arg->time_warp_seq = ci->i_time_warp_seq;
1415 
1416 	arg->uid = inode->i_uid;
1417 	arg->gid = inode->i_gid;
1418 	arg->mode = inode->i_mode;
1419 
1420 	arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1421 	if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1422 	    !list_empty(&ci->i_cap_snaps)) {
1423 		struct ceph_cap_snap *capsnap;
1424 		list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1425 			if (capsnap->cap_flush.tid)
1426 				break;
1427 			if (capsnap->need_flush) {
1428 				flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1429 				break;
1430 			}
1431 		}
1432 	}
1433 	arg->flags = flags;
1434 }
1435 
1436 /*
1437  * Send a cap msg on the given inode.
1438  *
1439  * Caller should hold snap_rwsem (read), s_mutex.
1440  */
__send_cap(struct cap_msg_args * arg,struct ceph_inode_info * ci)1441 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1442 {
1443 	struct ceph_msg *msg;
1444 	struct inode *inode = &ci->netfs.inode;
1445 
1446 	msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1447 	if (!msg) {
1448 		pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1449 		       ceph_vinop(inode), ceph_cap_string(arg->dirty),
1450 		       arg->flush_tid);
1451 		spin_lock(&ci->i_ceph_lock);
1452 		__cap_delay_requeue(arg->session->s_mdsc, ci);
1453 		spin_unlock(&ci->i_ceph_lock);
1454 		return;
1455 	}
1456 
1457 	encode_cap_msg(msg, arg);
1458 	ceph_con_send(&arg->session->s_con, msg);
1459 	ceph_buffer_put(arg->old_xattr_buf);
1460 	if (arg->wake)
1461 		wake_up_all(&ci->i_cap_wq);
1462 }
1463 
__send_flush_snap(struct inode * inode,struct ceph_mds_session * session,struct ceph_cap_snap * capsnap,u32 mseq,u64 oldest_flush_tid)1464 static inline int __send_flush_snap(struct inode *inode,
1465 				    struct ceph_mds_session *session,
1466 				    struct ceph_cap_snap *capsnap,
1467 				    u32 mseq, u64 oldest_flush_tid)
1468 {
1469 	struct cap_msg_args	arg;
1470 	struct ceph_msg		*msg;
1471 
1472 	msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1473 	if (!msg)
1474 		return -ENOMEM;
1475 
1476 	arg.session = session;
1477 	arg.ino = ceph_vino(inode).ino;
1478 	arg.cid = 0;
1479 	arg.follows = capsnap->follows;
1480 	arg.flush_tid = capsnap->cap_flush.tid;
1481 	arg.oldest_flush_tid = oldest_flush_tid;
1482 
1483 	arg.size = capsnap->size;
1484 	arg.max_size = 0;
1485 	arg.xattr_version = capsnap->xattr_version;
1486 	arg.xattr_buf = capsnap->xattr_blob;
1487 	arg.old_xattr_buf = NULL;
1488 
1489 	arg.atime = capsnap->atime;
1490 	arg.mtime = capsnap->mtime;
1491 	arg.ctime = capsnap->ctime;
1492 	arg.btime = capsnap->btime;
1493 	arg.change_attr = capsnap->change_attr;
1494 
1495 	arg.op = CEPH_CAP_OP_FLUSHSNAP;
1496 	arg.caps = capsnap->issued;
1497 	arg.wanted = 0;
1498 	arg.dirty = capsnap->dirty;
1499 
1500 	arg.seq = 0;
1501 	arg.issue_seq = 0;
1502 	arg.mseq = mseq;
1503 	arg.time_warp_seq = capsnap->time_warp_seq;
1504 
1505 	arg.uid = capsnap->uid;
1506 	arg.gid = capsnap->gid;
1507 	arg.mode = capsnap->mode;
1508 
1509 	arg.inline_data = capsnap->inline_data;
1510 	arg.flags = 0;
1511 	arg.wake = false;
1512 
1513 	encode_cap_msg(msg, &arg);
1514 	ceph_con_send(&arg.session->s_con, msg);
1515 	return 0;
1516 }
1517 
1518 /*
1519  * When a snapshot is taken, clients accumulate dirty metadata on
1520  * inodes with capabilities in ceph_cap_snaps to describe the file
1521  * state at the time the snapshot was taken.  This must be flushed
1522  * asynchronously back to the MDS once sync writes complete and dirty
1523  * data is written out.
1524  *
1525  * Called under i_ceph_lock.
1526  */
__ceph_flush_snaps(struct ceph_inode_info * ci,struct ceph_mds_session * session)1527 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1528 			       struct ceph_mds_session *session)
1529 		__releases(ci->i_ceph_lock)
1530 		__acquires(ci->i_ceph_lock)
1531 {
1532 	struct inode *inode = &ci->netfs.inode;
1533 	struct ceph_mds_client *mdsc = session->s_mdsc;
1534 	struct ceph_cap_snap *capsnap;
1535 	u64 oldest_flush_tid = 0;
1536 	u64 first_tid = 1, last_tid = 0;
1537 
1538 	dout("__flush_snaps %p session %p\n", inode, session);
1539 
1540 	list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1541 		/*
1542 		 * we need to wait for sync writes to complete and for dirty
1543 		 * pages to be written out.
1544 		 */
1545 		if (capsnap->dirty_pages || capsnap->writing)
1546 			break;
1547 
1548 		/* should be removed by ceph_try_drop_cap_snap() */
1549 		BUG_ON(!capsnap->need_flush);
1550 
1551 		/* only flush each capsnap once */
1552 		if (capsnap->cap_flush.tid > 0) {
1553 			dout(" already flushed %p, skipping\n", capsnap);
1554 			continue;
1555 		}
1556 
1557 		spin_lock(&mdsc->cap_dirty_lock);
1558 		capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1559 		list_add_tail(&capsnap->cap_flush.g_list,
1560 			      &mdsc->cap_flush_list);
1561 		if (oldest_flush_tid == 0)
1562 			oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1563 		if (list_empty(&ci->i_flushing_item)) {
1564 			list_add_tail(&ci->i_flushing_item,
1565 				      &session->s_cap_flushing);
1566 		}
1567 		spin_unlock(&mdsc->cap_dirty_lock);
1568 
1569 		list_add_tail(&capsnap->cap_flush.i_list,
1570 			      &ci->i_cap_flush_list);
1571 
1572 		if (first_tid == 1)
1573 			first_tid = capsnap->cap_flush.tid;
1574 		last_tid = capsnap->cap_flush.tid;
1575 	}
1576 
1577 	ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1578 
1579 	while (first_tid <= last_tid) {
1580 		struct ceph_cap *cap = ci->i_auth_cap;
1581 		struct ceph_cap_flush *cf = NULL, *iter;
1582 		int ret;
1583 
1584 		if (!(cap && cap->session == session)) {
1585 			dout("__flush_snaps %p auth cap %p not mds%d, "
1586 			     "stop\n", inode, cap, session->s_mds);
1587 			break;
1588 		}
1589 
1590 		ret = -ENOENT;
1591 		list_for_each_entry(iter, &ci->i_cap_flush_list, i_list) {
1592 			if (iter->tid >= first_tid) {
1593 				cf = iter;
1594 				ret = 0;
1595 				break;
1596 			}
1597 		}
1598 		if (ret < 0)
1599 			break;
1600 
1601 		first_tid = cf->tid + 1;
1602 
1603 		capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1604 		refcount_inc(&capsnap->nref);
1605 		spin_unlock(&ci->i_ceph_lock);
1606 
1607 		dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1608 		     inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1609 
1610 		ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1611 					oldest_flush_tid);
1612 		if (ret < 0) {
1613 			pr_err("__flush_snaps: error sending cap flushsnap, "
1614 			       "ino (%llx.%llx) tid %llu follows %llu\n",
1615 				ceph_vinop(inode), cf->tid, capsnap->follows);
1616 		}
1617 
1618 		ceph_put_cap_snap(capsnap);
1619 		spin_lock(&ci->i_ceph_lock);
1620 	}
1621 }
1622 
ceph_flush_snaps(struct ceph_inode_info * ci,struct ceph_mds_session ** psession)1623 void ceph_flush_snaps(struct ceph_inode_info *ci,
1624 		      struct ceph_mds_session **psession)
1625 {
1626 	struct inode *inode = &ci->netfs.inode;
1627 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1628 	struct ceph_mds_session *session = NULL;
1629 	int mds;
1630 
1631 	dout("ceph_flush_snaps %p\n", inode);
1632 	if (psession)
1633 		session = *psession;
1634 retry:
1635 	spin_lock(&ci->i_ceph_lock);
1636 	if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1637 		dout(" no capsnap needs flush, doing nothing\n");
1638 		goto out;
1639 	}
1640 	if (!ci->i_auth_cap) {
1641 		dout(" no auth cap (migrating?), doing nothing\n");
1642 		goto out;
1643 	}
1644 
1645 	mds = ci->i_auth_cap->session->s_mds;
1646 	if (session && session->s_mds != mds) {
1647 		dout(" oops, wrong session %p mutex\n", session);
1648 		ceph_put_mds_session(session);
1649 		session = NULL;
1650 	}
1651 	if (!session) {
1652 		spin_unlock(&ci->i_ceph_lock);
1653 		mutex_lock(&mdsc->mutex);
1654 		session = __ceph_lookup_mds_session(mdsc, mds);
1655 		mutex_unlock(&mdsc->mutex);
1656 		goto retry;
1657 	}
1658 
1659 	// make sure flushsnap messages are sent in proper order.
1660 	if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1661 		__kick_flushing_caps(mdsc, session, ci, 0);
1662 
1663 	__ceph_flush_snaps(ci, session);
1664 out:
1665 	spin_unlock(&ci->i_ceph_lock);
1666 
1667 	if (psession)
1668 		*psession = session;
1669 	else
1670 		ceph_put_mds_session(session);
1671 	/* we flushed them all; remove this inode from the queue */
1672 	spin_lock(&mdsc->snap_flush_lock);
1673 	list_del_init(&ci->i_snap_flush_item);
1674 	spin_unlock(&mdsc->snap_flush_lock);
1675 }
1676 
1677 /*
1678  * Mark caps dirty.  If inode is newly dirty, return the dirty flags.
1679  * Caller is then responsible for calling __mark_inode_dirty with the
1680  * returned flags value.
1681  */
__ceph_mark_dirty_caps(struct ceph_inode_info * ci,int mask,struct ceph_cap_flush ** pcf)1682 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1683 			   struct ceph_cap_flush **pcf)
1684 {
1685 	struct ceph_mds_client *mdsc =
1686 		ceph_sb_to_client(ci->netfs.inode.i_sb)->mdsc;
1687 	struct inode *inode = &ci->netfs.inode;
1688 	int was = ci->i_dirty_caps;
1689 	int dirty = 0;
1690 
1691 	lockdep_assert_held(&ci->i_ceph_lock);
1692 
1693 	if (!ci->i_auth_cap) {
1694 		pr_warn("__mark_dirty_caps %p %llx mask %s, "
1695 			"but no auth cap (session was closed?)\n",
1696 			inode, ceph_ino(inode), ceph_cap_string(mask));
1697 		return 0;
1698 	}
1699 
1700 	dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->netfs.inode,
1701 	     ceph_cap_string(mask), ceph_cap_string(was),
1702 	     ceph_cap_string(was | mask));
1703 	ci->i_dirty_caps |= mask;
1704 	if (was == 0) {
1705 		struct ceph_mds_session *session = ci->i_auth_cap->session;
1706 
1707 		WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1708 		swap(ci->i_prealloc_cap_flush, *pcf);
1709 
1710 		if (!ci->i_head_snapc) {
1711 			WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1712 			ci->i_head_snapc = ceph_get_snap_context(
1713 				ci->i_snap_realm->cached_context);
1714 		}
1715 		dout(" inode %p now dirty snapc %p auth cap %p\n",
1716 		     &ci->netfs.inode, ci->i_head_snapc, ci->i_auth_cap);
1717 		BUG_ON(!list_empty(&ci->i_dirty_item));
1718 		spin_lock(&mdsc->cap_dirty_lock);
1719 		list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1720 		spin_unlock(&mdsc->cap_dirty_lock);
1721 		if (ci->i_flushing_caps == 0) {
1722 			ihold(inode);
1723 			dirty |= I_DIRTY_SYNC;
1724 		}
1725 	} else {
1726 		WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1727 	}
1728 	BUG_ON(list_empty(&ci->i_dirty_item));
1729 	if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1730 	    (mask & CEPH_CAP_FILE_BUFFER))
1731 		dirty |= I_DIRTY_DATASYNC;
1732 	__cap_delay_requeue(mdsc, ci);
1733 	return dirty;
1734 }
1735 
ceph_alloc_cap_flush(void)1736 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1737 {
1738 	struct ceph_cap_flush *cf;
1739 
1740 	cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1741 	if (!cf)
1742 		return NULL;
1743 
1744 	cf->is_capsnap = false;
1745 	return cf;
1746 }
1747 
ceph_free_cap_flush(struct ceph_cap_flush * cf)1748 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1749 {
1750 	if (cf)
1751 		kmem_cache_free(ceph_cap_flush_cachep, cf);
1752 }
1753 
__get_oldest_flush_tid(struct ceph_mds_client * mdsc)1754 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1755 {
1756 	if (!list_empty(&mdsc->cap_flush_list)) {
1757 		struct ceph_cap_flush *cf =
1758 			list_first_entry(&mdsc->cap_flush_list,
1759 					 struct ceph_cap_flush, g_list);
1760 		return cf->tid;
1761 	}
1762 	return 0;
1763 }
1764 
1765 /*
1766  * Remove cap_flush from the mdsc's or inode's flushing cap list.
1767  * Return true if caller needs to wake up flush waiters.
1768  */
__detach_cap_flush_from_mdsc(struct ceph_mds_client * mdsc,struct ceph_cap_flush * cf)1769 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1770 					 struct ceph_cap_flush *cf)
1771 {
1772 	struct ceph_cap_flush *prev;
1773 	bool wake = cf->wake;
1774 
1775 	if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1776 		prev = list_prev_entry(cf, g_list);
1777 		prev->wake = true;
1778 		wake = false;
1779 	}
1780 	list_del_init(&cf->g_list);
1781 	return wake;
1782 }
1783 
__detach_cap_flush_from_ci(struct ceph_inode_info * ci,struct ceph_cap_flush * cf)1784 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1785 				       struct ceph_cap_flush *cf)
1786 {
1787 	struct ceph_cap_flush *prev;
1788 	bool wake = cf->wake;
1789 
1790 	if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1791 		prev = list_prev_entry(cf, i_list);
1792 		prev->wake = true;
1793 		wake = false;
1794 	}
1795 	list_del_init(&cf->i_list);
1796 	return wake;
1797 }
1798 
1799 /*
1800  * Add dirty inode to the flushing list.  Assigned a seq number so we
1801  * can wait for caps to flush without starving.
1802  *
1803  * Called under i_ceph_lock. Returns the flush tid.
1804  */
__mark_caps_flushing(struct inode * inode,struct ceph_mds_session * session,bool wake,u64 * oldest_flush_tid)1805 static u64 __mark_caps_flushing(struct inode *inode,
1806 				struct ceph_mds_session *session, bool wake,
1807 				u64 *oldest_flush_tid)
1808 {
1809 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1810 	struct ceph_inode_info *ci = ceph_inode(inode);
1811 	struct ceph_cap_flush *cf = NULL;
1812 	int flushing;
1813 
1814 	lockdep_assert_held(&ci->i_ceph_lock);
1815 	BUG_ON(ci->i_dirty_caps == 0);
1816 	BUG_ON(list_empty(&ci->i_dirty_item));
1817 	BUG_ON(!ci->i_prealloc_cap_flush);
1818 
1819 	flushing = ci->i_dirty_caps;
1820 	dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1821 	     ceph_cap_string(flushing),
1822 	     ceph_cap_string(ci->i_flushing_caps),
1823 	     ceph_cap_string(ci->i_flushing_caps | flushing));
1824 	ci->i_flushing_caps |= flushing;
1825 	ci->i_dirty_caps = 0;
1826 	dout(" inode %p now !dirty\n", inode);
1827 
1828 	swap(cf, ci->i_prealloc_cap_flush);
1829 	cf->caps = flushing;
1830 	cf->wake = wake;
1831 
1832 	spin_lock(&mdsc->cap_dirty_lock);
1833 	list_del_init(&ci->i_dirty_item);
1834 
1835 	cf->tid = ++mdsc->last_cap_flush_tid;
1836 	list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1837 	*oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1838 
1839 	if (list_empty(&ci->i_flushing_item)) {
1840 		list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1841 		mdsc->num_cap_flushing++;
1842 	}
1843 	spin_unlock(&mdsc->cap_dirty_lock);
1844 
1845 	list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1846 
1847 	return cf->tid;
1848 }
1849 
1850 /*
1851  * try to invalidate mapping pages without blocking.
1852  */
try_nonblocking_invalidate(struct inode * inode)1853 static int try_nonblocking_invalidate(struct inode *inode)
1854 	__releases(ci->i_ceph_lock)
1855 	__acquires(ci->i_ceph_lock)
1856 {
1857 	struct ceph_inode_info *ci = ceph_inode(inode);
1858 	u32 invalidating_gen = ci->i_rdcache_gen;
1859 
1860 	spin_unlock(&ci->i_ceph_lock);
1861 	ceph_fscache_invalidate(inode, false);
1862 	invalidate_mapping_pages(&inode->i_data, 0, -1);
1863 	spin_lock(&ci->i_ceph_lock);
1864 
1865 	if (inode->i_data.nrpages == 0 &&
1866 	    invalidating_gen == ci->i_rdcache_gen) {
1867 		/* success. */
1868 		dout("try_nonblocking_invalidate %p success\n", inode);
1869 		/* save any racing async invalidate some trouble */
1870 		ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1871 		return 0;
1872 	}
1873 	dout("try_nonblocking_invalidate %p failed\n", inode);
1874 	return -1;
1875 }
1876 
__ceph_should_report_size(struct ceph_inode_info * ci)1877 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1878 {
1879 	loff_t size = i_size_read(&ci->netfs.inode);
1880 	/* mds will adjust max size according to the reported size */
1881 	if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1882 		return false;
1883 	if (size >= ci->i_max_size)
1884 		return true;
1885 	/* half of previous max_size increment has been used */
1886 	if (ci->i_max_size > ci->i_reported_size &&
1887 	    (size << 1) >= ci->i_max_size + ci->i_reported_size)
1888 		return true;
1889 	return false;
1890 }
1891 
1892 /*
1893  * Swiss army knife function to examine currently used and wanted
1894  * versus held caps.  Release, flush, ack revoked caps to mds as
1895  * appropriate.
1896  *
1897  *  CHECK_CAPS_AUTHONLY - we should only check the auth cap
1898  *  CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1899  *    further delay.
1900  */
ceph_check_caps(struct ceph_inode_info * ci,int flags,struct ceph_mds_session * session)1901 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1902 		     struct ceph_mds_session *session)
1903 {
1904 	struct inode *inode = &ci->netfs.inode;
1905 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1906 	struct ceph_cap *cap;
1907 	u64 flush_tid, oldest_flush_tid;
1908 	int file_wanted, used, cap_used;
1909 	int issued, implemented, want, retain, revoking, flushing = 0;
1910 	int mds = -1;   /* keep track of how far we've gone through i_caps list
1911 			   to avoid an infinite loop on retry */
1912 	struct rb_node *p;
1913 	bool queue_invalidate = false;
1914 	bool tried_invalidate = false;
1915 	bool queue_writeback = false;
1916 
1917 	if (session)
1918 		ceph_get_mds_session(session);
1919 
1920 	spin_lock(&ci->i_ceph_lock);
1921 	if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
1922 		/* Don't send messages until we get async create reply */
1923 		spin_unlock(&ci->i_ceph_lock);
1924 		ceph_put_mds_session(session);
1925 		return;
1926 	}
1927 
1928 	if (ci->i_ceph_flags & CEPH_I_FLUSH)
1929 		flags |= CHECK_CAPS_FLUSH;
1930 retry:
1931 	/* Caps wanted by virtue of active open files. */
1932 	file_wanted = __ceph_caps_file_wanted(ci);
1933 
1934 	/* Caps which have active references against them */
1935 	used = __ceph_caps_used(ci);
1936 
1937 	/*
1938 	 * "issued" represents the current caps that the MDS wants us to have.
1939 	 * "implemented" is the set that we have been granted, and includes the
1940 	 * ones that have not yet been returned to the MDS (the "revoking" set,
1941 	 * usually because they have outstanding references).
1942 	 */
1943 	issued = __ceph_caps_issued(ci, &implemented);
1944 	revoking = implemented & ~issued;
1945 
1946 	want = file_wanted;
1947 
1948 	/* The ones we currently want to retain (may be adjusted below) */
1949 	retain = file_wanted | used | CEPH_CAP_PIN;
1950 	if (!mdsc->stopping && inode->i_nlink > 0) {
1951 		if (file_wanted) {
1952 			retain |= CEPH_CAP_ANY;       /* be greedy */
1953 		} else if (S_ISDIR(inode->i_mode) &&
1954 			   (issued & CEPH_CAP_FILE_SHARED) &&
1955 			   __ceph_dir_is_complete(ci)) {
1956 			/*
1957 			 * If a directory is complete, we want to keep
1958 			 * the exclusive cap. So that MDS does not end up
1959 			 * revoking the shared cap on every create/unlink
1960 			 * operation.
1961 			 */
1962 			if (IS_RDONLY(inode)) {
1963 				want = CEPH_CAP_ANY_SHARED;
1964 			} else {
1965 				want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1966 			}
1967 			retain |= want;
1968 		} else {
1969 
1970 			retain |= CEPH_CAP_ANY_SHARED;
1971 			/*
1972 			 * keep RD only if we didn't have the file open RW,
1973 			 * because then the mds would revoke it anyway to
1974 			 * journal max_size=0.
1975 			 */
1976 			if (ci->i_max_size == 0)
1977 				retain |= CEPH_CAP_ANY_RD;
1978 		}
1979 	}
1980 
1981 	dout("check_caps %llx.%llx file_want %s used %s dirty %s flushing %s"
1982 	     " issued %s revoking %s retain %s %s%s%s\n", ceph_vinop(inode),
1983 	     ceph_cap_string(file_wanted),
1984 	     ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1985 	     ceph_cap_string(ci->i_flushing_caps),
1986 	     ceph_cap_string(issued), ceph_cap_string(revoking),
1987 	     ceph_cap_string(retain),
1988 	     (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1989 	     (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "",
1990 	     (flags & CHECK_CAPS_NOINVAL) ? " NOINVAL" : "");
1991 
1992 	/*
1993 	 * If we no longer need to hold onto old our caps, and we may
1994 	 * have cached pages, but don't want them, then try to invalidate.
1995 	 * If we fail, it's because pages are locked.... try again later.
1996 	 */
1997 	if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
1998 	    S_ISREG(inode->i_mode) &&
1999 	    !(ci->i_wb_ref || ci->i_wrbuffer_ref) &&   /* no dirty pages... */
2000 	    inode->i_data.nrpages &&		/* have cached pages */
2001 	    (revoking & (CEPH_CAP_FILE_CACHE|
2002 			 CEPH_CAP_FILE_LAZYIO)) && /*  or revoking cache */
2003 	    !tried_invalidate) {
2004 		dout("check_caps trying to invalidate on %llx.%llx\n",
2005 		     ceph_vinop(inode));
2006 		if (try_nonblocking_invalidate(inode) < 0) {
2007 			dout("check_caps queuing invalidate\n");
2008 			queue_invalidate = true;
2009 			ci->i_rdcache_revoking = ci->i_rdcache_gen;
2010 		}
2011 		tried_invalidate = true;
2012 		goto retry;
2013 	}
2014 
2015 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2016 		int mflags = 0;
2017 		struct cap_msg_args arg;
2018 
2019 		cap = rb_entry(p, struct ceph_cap, ci_node);
2020 
2021 		/* avoid looping forever */
2022 		if (mds >= cap->mds ||
2023 		    ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2024 			continue;
2025 
2026 		/*
2027 		 * If we have an auth cap, we don't need to consider any
2028 		 * overlapping caps as used.
2029 		 */
2030 		cap_used = used;
2031 		if (ci->i_auth_cap && cap != ci->i_auth_cap)
2032 			cap_used &= ~ci->i_auth_cap->issued;
2033 
2034 		revoking = cap->implemented & ~cap->issued;
2035 		dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2036 		     cap->mds, cap, ceph_cap_string(cap_used),
2037 		     ceph_cap_string(cap->issued),
2038 		     ceph_cap_string(cap->implemented),
2039 		     ceph_cap_string(revoking));
2040 
2041 		if (cap == ci->i_auth_cap &&
2042 		    (cap->issued & CEPH_CAP_FILE_WR)) {
2043 			/* request larger max_size from MDS? */
2044 			if (ci->i_wanted_max_size > ci->i_max_size &&
2045 			    ci->i_wanted_max_size > ci->i_requested_max_size) {
2046 				dout("requesting new max_size\n");
2047 				goto ack;
2048 			}
2049 
2050 			/* approaching file_max? */
2051 			if (__ceph_should_report_size(ci)) {
2052 				dout("i_size approaching max_size\n");
2053 				goto ack;
2054 			}
2055 		}
2056 		/* flush anything dirty? */
2057 		if (cap == ci->i_auth_cap) {
2058 			if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2059 				dout("flushing dirty caps\n");
2060 				goto ack;
2061 			}
2062 			if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2063 				dout("flushing snap caps\n");
2064 				goto ack;
2065 			}
2066 		}
2067 
2068 		/* completed revocation? going down and there are no caps? */
2069 		if (revoking) {
2070 			if ((revoking & cap_used) == 0) {
2071 				dout("completed revocation of %s\n",
2072 				      ceph_cap_string(cap->implemented & ~cap->issued));
2073 				goto ack;
2074 			}
2075 
2076 			/*
2077 			 * If the "i_wrbuffer_ref" was increased by mmap or generic
2078 			 * cache write just before the ceph_check_caps() is called,
2079 			 * the Fb capability revoking will fail this time. Then we
2080 			 * must wait for the BDI's delayed work to flush the dirty
2081 			 * pages and to release the "i_wrbuffer_ref", which will cost
2082 			 * at most 5 seconds. That means the MDS needs to wait at
2083 			 * most 5 seconds to finished the Fb capability's revocation.
2084 			 *
2085 			 * Let's queue a writeback for it.
2086 			 */
2087 			if (S_ISREG(inode->i_mode) && ci->i_wrbuffer_ref &&
2088 			    (revoking & CEPH_CAP_FILE_BUFFER))
2089 				queue_writeback = true;
2090 		}
2091 
2092 		/* want more caps from mds? */
2093 		if (want & ~cap->mds_wanted) {
2094 			if (want & ~(cap->mds_wanted | cap->issued))
2095 				goto ack;
2096 			if (!__cap_is_valid(cap))
2097 				goto ack;
2098 		}
2099 
2100 		/* things we might delay */
2101 		if ((cap->issued & ~retain) == 0)
2102 			continue;     /* nope, all good */
2103 
2104 ack:
2105 		ceph_put_mds_session(session);
2106 		session = ceph_get_mds_session(cap->session);
2107 
2108 		/* kick flushing and flush snaps before sending normal
2109 		 * cap message */
2110 		if (cap == ci->i_auth_cap &&
2111 		    (ci->i_ceph_flags &
2112 		     (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2113 			if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2114 				__kick_flushing_caps(mdsc, session, ci, 0);
2115 			if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2116 				__ceph_flush_snaps(ci, session);
2117 
2118 			goto retry;
2119 		}
2120 
2121 		if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2122 			flushing = ci->i_dirty_caps;
2123 			flush_tid = __mark_caps_flushing(inode, session, false,
2124 							 &oldest_flush_tid);
2125 			if (flags & CHECK_CAPS_FLUSH &&
2126 			    list_empty(&session->s_cap_dirty))
2127 				mflags |= CEPH_CLIENT_CAPS_SYNC;
2128 		} else {
2129 			flushing = 0;
2130 			flush_tid = 0;
2131 			spin_lock(&mdsc->cap_dirty_lock);
2132 			oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2133 			spin_unlock(&mdsc->cap_dirty_lock);
2134 		}
2135 
2136 		mds = cap->mds;  /* remember mds, so we don't repeat */
2137 
2138 		__prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2139 			   want, retain, flushing, flush_tid, oldest_flush_tid);
2140 
2141 		spin_unlock(&ci->i_ceph_lock);
2142 		__send_cap(&arg, ci);
2143 		spin_lock(&ci->i_ceph_lock);
2144 
2145 		goto retry; /* retake i_ceph_lock and restart our cap scan. */
2146 	}
2147 
2148 	/* periodically re-calculate caps wanted by open files */
2149 	if (__ceph_is_any_real_caps(ci) &&
2150 	    list_empty(&ci->i_cap_delay_list) &&
2151 	    (file_wanted & ~CEPH_CAP_PIN) &&
2152 	    !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2153 		__cap_delay_requeue(mdsc, ci);
2154 	}
2155 
2156 	spin_unlock(&ci->i_ceph_lock);
2157 
2158 	ceph_put_mds_session(session);
2159 	if (queue_writeback)
2160 		ceph_queue_writeback(inode);
2161 	if (queue_invalidate)
2162 		ceph_queue_invalidate(inode);
2163 }
2164 
2165 /*
2166  * Try to flush dirty caps back to the auth mds.
2167  */
try_flush_caps(struct inode * inode,u64 * ptid)2168 static int try_flush_caps(struct inode *inode, u64 *ptid)
2169 {
2170 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2171 	struct ceph_inode_info *ci = ceph_inode(inode);
2172 	int flushing = 0;
2173 	u64 flush_tid = 0, oldest_flush_tid = 0;
2174 
2175 	spin_lock(&ci->i_ceph_lock);
2176 retry_locked:
2177 	if (ci->i_dirty_caps && ci->i_auth_cap) {
2178 		struct ceph_cap *cap = ci->i_auth_cap;
2179 		struct cap_msg_args arg;
2180 		struct ceph_mds_session *session = cap->session;
2181 
2182 		if (session->s_state < CEPH_MDS_SESSION_OPEN) {
2183 			spin_unlock(&ci->i_ceph_lock);
2184 			goto out;
2185 		}
2186 
2187 		if (ci->i_ceph_flags &
2188 		    (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2189 			if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2190 				__kick_flushing_caps(mdsc, session, ci, 0);
2191 			if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2192 				__ceph_flush_snaps(ci, session);
2193 			goto retry_locked;
2194 		}
2195 
2196 		flushing = ci->i_dirty_caps;
2197 		flush_tid = __mark_caps_flushing(inode, session, true,
2198 						 &oldest_flush_tid);
2199 
2200 		__prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2201 			   __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2202 			   (cap->issued | cap->implemented),
2203 			   flushing, flush_tid, oldest_flush_tid);
2204 		spin_unlock(&ci->i_ceph_lock);
2205 
2206 		__send_cap(&arg, ci);
2207 	} else {
2208 		if (!list_empty(&ci->i_cap_flush_list)) {
2209 			struct ceph_cap_flush *cf =
2210 				list_last_entry(&ci->i_cap_flush_list,
2211 						struct ceph_cap_flush, i_list);
2212 			cf->wake = true;
2213 			flush_tid = cf->tid;
2214 		}
2215 		flushing = ci->i_flushing_caps;
2216 		spin_unlock(&ci->i_ceph_lock);
2217 	}
2218 out:
2219 	*ptid = flush_tid;
2220 	return flushing;
2221 }
2222 
2223 /*
2224  * Return true if we've flushed caps through the given flush_tid.
2225  */
caps_are_flushed(struct inode * inode,u64 flush_tid)2226 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2227 {
2228 	struct ceph_inode_info *ci = ceph_inode(inode);
2229 	int ret = 1;
2230 
2231 	spin_lock(&ci->i_ceph_lock);
2232 	if (!list_empty(&ci->i_cap_flush_list)) {
2233 		struct ceph_cap_flush * cf =
2234 			list_first_entry(&ci->i_cap_flush_list,
2235 					 struct ceph_cap_flush, i_list);
2236 		if (cf->tid <= flush_tid)
2237 			ret = 0;
2238 	}
2239 	spin_unlock(&ci->i_ceph_lock);
2240 	return ret;
2241 }
2242 
2243 /*
2244  * flush the mdlog and wait for any unsafe requests to complete.
2245  */
flush_mdlog_and_wait_inode_unsafe_requests(struct inode * inode)2246 static int flush_mdlog_and_wait_inode_unsafe_requests(struct inode *inode)
2247 {
2248 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2249 	struct ceph_inode_info *ci = ceph_inode(inode);
2250 	struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2251 	int ret, err = 0;
2252 
2253 	spin_lock(&ci->i_unsafe_lock);
2254 	if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2255 		req1 = list_last_entry(&ci->i_unsafe_dirops,
2256 					struct ceph_mds_request,
2257 					r_unsafe_dir_item);
2258 		ceph_mdsc_get_request(req1);
2259 	}
2260 	if (!list_empty(&ci->i_unsafe_iops)) {
2261 		req2 = list_last_entry(&ci->i_unsafe_iops,
2262 					struct ceph_mds_request,
2263 					r_unsafe_target_item);
2264 		ceph_mdsc_get_request(req2);
2265 	}
2266 	spin_unlock(&ci->i_unsafe_lock);
2267 
2268 	/*
2269 	 * Trigger to flush the journal logs in all the relevant MDSes
2270 	 * manually, or in the worst case we must wait at most 5 seconds
2271 	 * to wait the journal logs to be flushed by the MDSes periodically.
2272 	 */
2273 	if (req1 || req2) {
2274 		struct ceph_mds_request *req;
2275 		struct ceph_mds_session **sessions;
2276 		struct ceph_mds_session *s;
2277 		unsigned int max_sessions;
2278 		int i;
2279 
2280 		mutex_lock(&mdsc->mutex);
2281 		max_sessions = mdsc->max_sessions;
2282 
2283 		sessions = kcalloc(max_sessions, sizeof(s), GFP_KERNEL);
2284 		if (!sessions) {
2285 			mutex_unlock(&mdsc->mutex);
2286 			err = -ENOMEM;
2287 			goto out;
2288 		}
2289 
2290 		spin_lock(&ci->i_unsafe_lock);
2291 		if (req1) {
2292 			list_for_each_entry(req, &ci->i_unsafe_dirops,
2293 					    r_unsafe_dir_item) {
2294 				s = req->r_session;
2295 				if (!s)
2296 					continue;
2297 				if (!sessions[s->s_mds]) {
2298 					s = ceph_get_mds_session(s);
2299 					sessions[s->s_mds] = s;
2300 				}
2301 			}
2302 		}
2303 		if (req2) {
2304 			list_for_each_entry(req, &ci->i_unsafe_iops,
2305 					    r_unsafe_target_item) {
2306 				s = req->r_session;
2307 				if (!s)
2308 					continue;
2309 				if (!sessions[s->s_mds]) {
2310 					s = ceph_get_mds_session(s);
2311 					sessions[s->s_mds] = s;
2312 				}
2313 			}
2314 		}
2315 		spin_unlock(&ci->i_unsafe_lock);
2316 
2317 		/* the auth MDS */
2318 		spin_lock(&ci->i_ceph_lock);
2319 		if (ci->i_auth_cap) {
2320 			s = ci->i_auth_cap->session;
2321 			if (!sessions[s->s_mds])
2322 				sessions[s->s_mds] = ceph_get_mds_session(s);
2323 		}
2324 		spin_unlock(&ci->i_ceph_lock);
2325 		mutex_unlock(&mdsc->mutex);
2326 
2327 		/* send flush mdlog request to MDSes */
2328 		for (i = 0; i < max_sessions; i++) {
2329 			s = sessions[i];
2330 			if (s) {
2331 				send_flush_mdlog(s);
2332 				ceph_put_mds_session(s);
2333 			}
2334 		}
2335 		kfree(sessions);
2336 	}
2337 
2338 	dout("%s %p wait on tid %llu %llu\n", __func__,
2339 	     inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2340 	if (req1) {
2341 		ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2342 					ceph_timeout_jiffies(req1->r_timeout));
2343 		if (ret)
2344 			err = -EIO;
2345 	}
2346 	if (req2) {
2347 		ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2348 					ceph_timeout_jiffies(req2->r_timeout));
2349 		if (ret)
2350 			err = -EIO;
2351 	}
2352 
2353 out:
2354 	if (req1)
2355 		ceph_mdsc_put_request(req1);
2356 	if (req2)
2357 		ceph_mdsc_put_request(req2);
2358 	return err;
2359 }
2360 
ceph_fsync(struct file * file,loff_t start,loff_t end,int datasync)2361 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2362 {
2363 	struct inode *inode = file->f_mapping->host;
2364 	struct ceph_inode_info *ci = ceph_inode(inode);
2365 	u64 flush_tid;
2366 	int ret, err;
2367 	int dirty;
2368 
2369 	dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2370 
2371 	ret = file_write_and_wait_range(file, start, end);
2372 	if (datasync)
2373 		goto out;
2374 
2375 	ret = ceph_wait_on_async_create(inode);
2376 	if (ret)
2377 		goto out;
2378 
2379 	dirty = try_flush_caps(inode, &flush_tid);
2380 	dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2381 
2382 	err = flush_mdlog_and_wait_inode_unsafe_requests(inode);
2383 
2384 	/*
2385 	 * only wait on non-file metadata writeback (the mds
2386 	 * can recover size and mtime, so we don't need to
2387 	 * wait for that)
2388 	 */
2389 	if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2390 		err = wait_event_interruptible(ci->i_cap_wq,
2391 					caps_are_flushed(inode, flush_tid));
2392 	}
2393 
2394 	if (err < 0)
2395 		ret = err;
2396 
2397 	err = file_check_and_advance_wb_err(file);
2398 	if (err < 0)
2399 		ret = err;
2400 out:
2401 	dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2402 	return ret;
2403 }
2404 
2405 /*
2406  * Flush any dirty caps back to the mds.  If we aren't asked to wait,
2407  * queue inode for flush but don't do so immediately, because we can
2408  * get by with fewer MDS messages if we wait for data writeback to
2409  * complete first.
2410  */
ceph_write_inode(struct inode * inode,struct writeback_control * wbc)2411 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2412 {
2413 	struct ceph_inode_info *ci = ceph_inode(inode);
2414 	u64 flush_tid;
2415 	int err = 0;
2416 	int dirty;
2417 	int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2418 
2419 	dout("write_inode %p wait=%d\n", inode, wait);
2420 	ceph_fscache_unpin_writeback(inode, wbc);
2421 	if (wait) {
2422 		err = ceph_wait_on_async_create(inode);
2423 		if (err)
2424 			return err;
2425 		dirty = try_flush_caps(inode, &flush_tid);
2426 		if (dirty)
2427 			err = wait_event_interruptible(ci->i_cap_wq,
2428 				       caps_are_flushed(inode, flush_tid));
2429 	} else {
2430 		struct ceph_mds_client *mdsc =
2431 			ceph_sb_to_client(inode->i_sb)->mdsc;
2432 
2433 		spin_lock(&ci->i_ceph_lock);
2434 		if (__ceph_caps_dirty(ci))
2435 			__cap_delay_requeue_front(mdsc, ci);
2436 		spin_unlock(&ci->i_ceph_lock);
2437 	}
2438 	return err;
2439 }
2440 
__kick_flushing_caps(struct ceph_mds_client * mdsc,struct ceph_mds_session * session,struct ceph_inode_info * ci,u64 oldest_flush_tid)2441 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2442 				 struct ceph_mds_session *session,
2443 				 struct ceph_inode_info *ci,
2444 				 u64 oldest_flush_tid)
2445 	__releases(ci->i_ceph_lock)
2446 	__acquires(ci->i_ceph_lock)
2447 {
2448 	struct inode *inode = &ci->netfs.inode;
2449 	struct ceph_cap *cap;
2450 	struct ceph_cap_flush *cf;
2451 	int ret;
2452 	u64 first_tid = 0;
2453 	u64 last_snap_flush = 0;
2454 
2455 	/* Don't do anything until create reply comes in */
2456 	if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE)
2457 		return;
2458 
2459 	ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2460 
2461 	list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2462 		if (cf->is_capsnap) {
2463 			last_snap_flush = cf->tid;
2464 			break;
2465 		}
2466 	}
2467 
2468 	list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2469 		if (cf->tid < first_tid)
2470 			continue;
2471 
2472 		cap = ci->i_auth_cap;
2473 		if (!(cap && cap->session == session)) {
2474 			pr_err("%p auth cap %p not mds%d ???\n",
2475 			       inode, cap, session->s_mds);
2476 			break;
2477 		}
2478 
2479 		first_tid = cf->tid + 1;
2480 
2481 		if (!cf->is_capsnap) {
2482 			struct cap_msg_args arg;
2483 
2484 			dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2485 			     inode, cap, cf->tid, ceph_cap_string(cf->caps));
2486 			__prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2487 					 (cf->tid < last_snap_flush ?
2488 					  CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2489 					  __ceph_caps_used(ci),
2490 					  __ceph_caps_wanted(ci),
2491 					  (cap->issued | cap->implemented),
2492 					  cf->caps, cf->tid, oldest_flush_tid);
2493 			spin_unlock(&ci->i_ceph_lock);
2494 			__send_cap(&arg, ci);
2495 		} else {
2496 			struct ceph_cap_snap *capsnap =
2497 					container_of(cf, struct ceph_cap_snap,
2498 						    cap_flush);
2499 			dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2500 			     inode, capsnap, cf->tid,
2501 			     ceph_cap_string(capsnap->dirty));
2502 
2503 			refcount_inc(&capsnap->nref);
2504 			spin_unlock(&ci->i_ceph_lock);
2505 
2506 			ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2507 						oldest_flush_tid);
2508 			if (ret < 0) {
2509 				pr_err("kick_flushing_caps: error sending "
2510 					"cap flushsnap, ino (%llx.%llx) "
2511 					"tid %llu follows %llu\n",
2512 					ceph_vinop(inode), cf->tid,
2513 					capsnap->follows);
2514 			}
2515 
2516 			ceph_put_cap_snap(capsnap);
2517 		}
2518 
2519 		spin_lock(&ci->i_ceph_lock);
2520 	}
2521 }
2522 
ceph_early_kick_flushing_caps(struct ceph_mds_client * mdsc,struct ceph_mds_session * session)2523 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2524 				   struct ceph_mds_session *session)
2525 {
2526 	struct ceph_inode_info *ci;
2527 	struct ceph_cap *cap;
2528 	u64 oldest_flush_tid;
2529 
2530 	dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2531 
2532 	spin_lock(&mdsc->cap_dirty_lock);
2533 	oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2534 	spin_unlock(&mdsc->cap_dirty_lock);
2535 
2536 	list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2537 		spin_lock(&ci->i_ceph_lock);
2538 		cap = ci->i_auth_cap;
2539 		if (!(cap && cap->session == session)) {
2540 			pr_err("%p auth cap %p not mds%d ???\n",
2541 				&ci->netfs.inode, cap, session->s_mds);
2542 			spin_unlock(&ci->i_ceph_lock);
2543 			continue;
2544 		}
2545 
2546 
2547 		/*
2548 		 * if flushing caps were revoked, we re-send the cap flush
2549 		 * in client reconnect stage. This guarantees MDS * processes
2550 		 * the cap flush message before issuing the flushing caps to
2551 		 * other client.
2552 		 */
2553 		if ((cap->issued & ci->i_flushing_caps) !=
2554 		    ci->i_flushing_caps) {
2555 			/* encode_caps_cb() also will reset these sequence
2556 			 * numbers. make sure sequence numbers in cap flush
2557 			 * message match later reconnect message */
2558 			cap->seq = 0;
2559 			cap->issue_seq = 0;
2560 			cap->mseq = 0;
2561 			__kick_flushing_caps(mdsc, session, ci,
2562 					     oldest_flush_tid);
2563 		} else {
2564 			ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2565 		}
2566 
2567 		spin_unlock(&ci->i_ceph_lock);
2568 	}
2569 }
2570 
ceph_kick_flushing_caps(struct ceph_mds_client * mdsc,struct ceph_mds_session * session)2571 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2572 			     struct ceph_mds_session *session)
2573 {
2574 	struct ceph_inode_info *ci;
2575 	struct ceph_cap *cap;
2576 	u64 oldest_flush_tid;
2577 
2578 	lockdep_assert_held(&session->s_mutex);
2579 
2580 	dout("kick_flushing_caps mds%d\n", session->s_mds);
2581 
2582 	spin_lock(&mdsc->cap_dirty_lock);
2583 	oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2584 	spin_unlock(&mdsc->cap_dirty_lock);
2585 
2586 	list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2587 		spin_lock(&ci->i_ceph_lock);
2588 		cap = ci->i_auth_cap;
2589 		if (!(cap && cap->session == session)) {
2590 			pr_err("%p auth cap %p not mds%d ???\n",
2591 				&ci->netfs.inode, cap, session->s_mds);
2592 			spin_unlock(&ci->i_ceph_lock);
2593 			continue;
2594 		}
2595 		if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2596 			__kick_flushing_caps(mdsc, session, ci,
2597 					     oldest_flush_tid);
2598 		}
2599 		spin_unlock(&ci->i_ceph_lock);
2600 	}
2601 }
2602 
ceph_kick_flushing_inode_caps(struct ceph_mds_session * session,struct ceph_inode_info * ci)2603 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2604 				   struct ceph_inode_info *ci)
2605 {
2606 	struct ceph_mds_client *mdsc = session->s_mdsc;
2607 	struct ceph_cap *cap = ci->i_auth_cap;
2608 
2609 	lockdep_assert_held(&ci->i_ceph_lock);
2610 
2611 	dout("%s %p flushing %s\n", __func__, &ci->netfs.inode,
2612 	     ceph_cap_string(ci->i_flushing_caps));
2613 
2614 	if (!list_empty(&ci->i_cap_flush_list)) {
2615 		u64 oldest_flush_tid;
2616 		spin_lock(&mdsc->cap_dirty_lock);
2617 		list_move_tail(&ci->i_flushing_item,
2618 			       &cap->session->s_cap_flushing);
2619 		oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2620 		spin_unlock(&mdsc->cap_dirty_lock);
2621 
2622 		__kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2623 	}
2624 }
2625 
2626 
2627 /*
2628  * Take references to capabilities we hold, so that we don't release
2629  * them to the MDS prematurely.
2630  */
ceph_take_cap_refs(struct ceph_inode_info * ci,int got,bool snap_rwsem_locked)2631 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2632 			    bool snap_rwsem_locked)
2633 {
2634 	lockdep_assert_held(&ci->i_ceph_lock);
2635 
2636 	if (got & CEPH_CAP_PIN)
2637 		ci->i_pin_ref++;
2638 	if (got & CEPH_CAP_FILE_RD)
2639 		ci->i_rd_ref++;
2640 	if (got & CEPH_CAP_FILE_CACHE)
2641 		ci->i_rdcache_ref++;
2642 	if (got & CEPH_CAP_FILE_EXCL)
2643 		ci->i_fx_ref++;
2644 	if (got & CEPH_CAP_FILE_WR) {
2645 		if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2646 			BUG_ON(!snap_rwsem_locked);
2647 			ci->i_head_snapc = ceph_get_snap_context(
2648 					ci->i_snap_realm->cached_context);
2649 		}
2650 		ci->i_wr_ref++;
2651 	}
2652 	if (got & CEPH_CAP_FILE_BUFFER) {
2653 		if (ci->i_wb_ref == 0)
2654 			ihold(&ci->netfs.inode);
2655 		ci->i_wb_ref++;
2656 		dout("%s %p wb %d -> %d (?)\n", __func__,
2657 		     &ci->netfs.inode, ci->i_wb_ref-1, ci->i_wb_ref);
2658 	}
2659 }
2660 
2661 /*
2662  * Try to grab cap references.  Specify those refs we @want, and the
2663  * minimal set we @need.  Also include the larger offset we are writing
2664  * to (when applicable), and check against max_size here as well.
2665  * Note that caller is responsible for ensuring max_size increases are
2666  * requested from the MDS.
2667  *
2668  * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2669  * or a negative error code. There are 3 speical error codes:
2670  *  -EAGAIN:  need to sleep but non-blocking is specified
2671  *  -EFBIG:   ask caller to call check_max_size() and try again.
2672  *  -EUCLEAN: ask caller to call ceph_renew_caps() and try again.
2673  */
2674 enum {
2675 	/* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2676 	NON_BLOCKING	= (1 << 8),
2677 	CHECK_FILELOCK	= (1 << 9),
2678 };
2679 
try_get_cap_refs(struct inode * inode,int need,int want,loff_t endoff,int flags,int * got)2680 static int try_get_cap_refs(struct inode *inode, int need, int want,
2681 			    loff_t endoff, int flags, int *got)
2682 {
2683 	struct ceph_inode_info *ci = ceph_inode(inode);
2684 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2685 	int ret = 0;
2686 	int have, implemented;
2687 	bool snap_rwsem_locked = false;
2688 
2689 	dout("get_cap_refs %p need %s want %s\n", inode,
2690 	     ceph_cap_string(need), ceph_cap_string(want));
2691 
2692 again:
2693 	spin_lock(&ci->i_ceph_lock);
2694 
2695 	if ((flags & CHECK_FILELOCK) &&
2696 	    (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2697 		dout("try_get_cap_refs %p error filelock\n", inode);
2698 		ret = -EIO;
2699 		goto out_unlock;
2700 	}
2701 
2702 	/* finish pending truncate */
2703 	while (ci->i_truncate_pending) {
2704 		spin_unlock(&ci->i_ceph_lock);
2705 		if (snap_rwsem_locked) {
2706 			up_read(&mdsc->snap_rwsem);
2707 			snap_rwsem_locked = false;
2708 		}
2709 		__ceph_do_pending_vmtruncate(inode);
2710 		spin_lock(&ci->i_ceph_lock);
2711 	}
2712 
2713 	have = __ceph_caps_issued(ci, &implemented);
2714 
2715 	if (have & need & CEPH_CAP_FILE_WR) {
2716 		if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2717 			dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2718 			     inode, endoff, ci->i_max_size);
2719 			if (endoff > ci->i_requested_max_size)
2720 				ret = ci->i_auth_cap ? -EFBIG : -EUCLEAN;
2721 			goto out_unlock;
2722 		}
2723 		/*
2724 		 * If a sync write is in progress, we must wait, so that we
2725 		 * can get a final snapshot value for size+mtime.
2726 		 */
2727 		if (__ceph_have_pending_cap_snap(ci)) {
2728 			dout("get_cap_refs %p cap_snap_pending\n", inode);
2729 			goto out_unlock;
2730 		}
2731 	}
2732 
2733 	if ((have & need) == need) {
2734 		/*
2735 		 * Look at (implemented & ~have & not) so that we keep waiting
2736 		 * on transition from wanted -> needed caps.  This is needed
2737 		 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2738 		 * going before a prior buffered writeback happens.
2739 		 *
2740 		 * For RDCACHE|RD -> RD, there is not need to wait and we can
2741 		 * just exclude the revoking caps and force to sync read.
2742 		 */
2743 		int not = want & ~(have & need);
2744 		int revoking = implemented & ~have;
2745 		int exclude = revoking & not;
2746 		dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2747 		     inode, ceph_cap_string(have), ceph_cap_string(not),
2748 		     ceph_cap_string(revoking));
2749 		if (!exclude || !(exclude & CEPH_CAP_FILE_BUFFER)) {
2750 			if (!snap_rwsem_locked &&
2751 			    !ci->i_head_snapc &&
2752 			    (need & CEPH_CAP_FILE_WR)) {
2753 				if (!down_read_trylock(&mdsc->snap_rwsem)) {
2754 					/*
2755 					 * we can not call down_read() when
2756 					 * task isn't in TASK_RUNNING state
2757 					 */
2758 					if (flags & NON_BLOCKING) {
2759 						ret = -EAGAIN;
2760 						goto out_unlock;
2761 					}
2762 
2763 					spin_unlock(&ci->i_ceph_lock);
2764 					down_read(&mdsc->snap_rwsem);
2765 					snap_rwsem_locked = true;
2766 					goto again;
2767 				}
2768 				snap_rwsem_locked = true;
2769 			}
2770 			if ((have & want) == want)
2771 				*got = need | (want & ~exclude);
2772 			else
2773 				*got = need;
2774 			ceph_take_cap_refs(ci, *got, true);
2775 			ret = 1;
2776 		}
2777 	} else {
2778 		int session_readonly = false;
2779 		int mds_wanted;
2780 		if (ci->i_auth_cap &&
2781 		    (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2782 			struct ceph_mds_session *s = ci->i_auth_cap->session;
2783 			spin_lock(&s->s_cap_lock);
2784 			session_readonly = s->s_readonly;
2785 			spin_unlock(&s->s_cap_lock);
2786 		}
2787 		if (session_readonly) {
2788 			dout("get_cap_refs %p need %s but mds%d readonly\n",
2789 			     inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2790 			ret = -EROFS;
2791 			goto out_unlock;
2792 		}
2793 
2794 		if (ceph_inode_is_shutdown(inode)) {
2795 			dout("get_cap_refs %p inode is shutdown\n", inode);
2796 			ret = -ESTALE;
2797 			goto out_unlock;
2798 		}
2799 		mds_wanted = __ceph_caps_mds_wanted(ci, false);
2800 		if (need & ~mds_wanted) {
2801 			dout("get_cap_refs %p need %s > mds_wanted %s\n",
2802 			     inode, ceph_cap_string(need),
2803 			     ceph_cap_string(mds_wanted));
2804 			ret = -EUCLEAN;
2805 			goto out_unlock;
2806 		}
2807 
2808 		dout("get_cap_refs %p have %s need %s\n", inode,
2809 		     ceph_cap_string(have), ceph_cap_string(need));
2810 	}
2811 out_unlock:
2812 
2813 	__ceph_touch_fmode(ci, mdsc, flags);
2814 
2815 	spin_unlock(&ci->i_ceph_lock);
2816 	if (snap_rwsem_locked)
2817 		up_read(&mdsc->snap_rwsem);
2818 
2819 	if (!ret)
2820 		ceph_update_cap_mis(&mdsc->metric);
2821 	else if (ret == 1)
2822 		ceph_update_cap_hit(&mdsc->metric);
2823 
2824 	dout("get_cap_refs %p ret %d got %s\n", inode,
2825 	     ret, ceph_cap_string(*got));
2826 	return ret;
2827 }
2828 
2829 /*
2830  * Check the offset we are writing up to against our current
2831  * max_size.  If necessary, tell the MDS we want to write to
2832  * a larger offset.
2833  */
check_max_size(struct inode * inode,loff_t endoff)2834 static void check_max_size(struct inode *inode, loff_t endoff)
2835 {
2836 	struct ceph_inode_info *ci = ceph_inode(inode);
2837 	int check = 0;
2838 
2839 	/* do we need to explicitly request a larger max_size? */
2840 	spin_lock(&ci->i_ceph_lock);
2841 	if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2842 		dout("write %p at large endoff %llu, req max_size\n",
2843 		     inode, endoff);
2844 		ci->i_wanted_max_size = endoff;
2845 	}
2846 	/* duplicate ceph_check_caps()'s logic */
2847 	if (ci->i_auth_cap &&
2848 	    (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2849 	    ci->i_wanted_max_size > ci->i_max_size &&
2850 	    ci->i_wanted_max_size > ci->i_requested_max_size)
2851 		check = 1;
2852 	spin_unlock(&ci->i_ceph_lock);
2853 	if (check)
2854 		ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2855 }
2856 
get_used_fmode(int caps)2857 static inline int get_used_fmode(int caps)
2858 {
2859 	int fmode = 0;
2860 	if (caps & CEPH_CAP_FILE_RD)
2861 		fmode |= CEPH_FILE_MODE_RD;
2862 	if (caps & CEPH_CAP_FILE_WR)
2863 		fmode |= CEPH_FILE_MODE_WR;
2864 	return fmode;
2865 }
2866 
ceph_try_get_caps(struct inode * inode,int need,int want,bool nonblock,int * got)2867 int ceph_try_get_caps(struct inode *inode, int need, int want,
2868 		      bool nonblock, int *got)
2869 {
2870 	int ret, flags;
2871 
2872 	BUG_ON(need & ~CEPH_CAP_FILE_RD);
2873 	BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2874 			CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2875 			CEPH_CAP_ANY_DIR_OPS));
2876 	if (need) {
2877 		ret = ceph_pool_perm_check(inode, need);
2878 		if (ret < 0)
2879 			return ret;
2880 	}
2881 
2882 	flags = get_used_fmode(need | want);
2883 	if (nonblock)
2884 		flags |= NON_BLOCKING;
2885 
2886 	ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2887 	/* three special error codes */
2888 	if (ret == -EAGAIN || ret == -EFBIG || ret == -EUCLEAN)
2889 		ret = 0;
2890 	return ret;
2891 }
2892 
2893 /*
2894  * Wait for caps, and take cap references.  If we can't get a WR cap
2895  * due to a small max_size, make sure we check_max_size (and possibly
2896  * ask the mds) so we don't get hung up indefinitely.
2897  */
ceph_get_caps(struct file * filp,int need,int want,loff_t endoff,int * got)2898 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2899 {
2900 	struct ceph_file_info *fi = filp->private_data;
2901 	struct inode *inode = file_inode(filp);
2902 	struct ceph_inode_info *ci = ceph_inode(inode);
2903 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2904 	int ret, _got, flags;
2905 
2906 	ret = ceph_pool_perm_check(inode, need);
2907 	if (ret < 0)
2908 		return ret;
2909 
2910 	if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2911 	    fi->filp_gen != READ_ONCE(fsc->filp_gen))
2912 		return -EBADF;
2913 
2914 	flags = get_used_fmode(need | want);
2915 
2916 	while (true) {
2917 		flags &= CEPH_FILE_MODE_MASK;
2918 		if (vfs_inode_has_locks(inode))
2919 			flags |= CHECK_FILELOCK;
2920 		_got = 0;
2921 		ret = try_get_cap_refs(inode, need, want, endoff,
2922 				       flags, &_got);
2923 		WARN_ON_ONCE(ret == -EAGAIN);
2924 		if (!ret) {
2925 			struct ceph_mds_client *mdsc = fsc->mdsc;
2926 			struct cap_wait cw;
2927 			DEFINE_WAIT_FUNC(wait, woken_wake_function);
2928 
2929 			cw.ino = ceph_ino(inode);
2930 			cw.tgid = current->tgid;
2931 			cw.need = need;
2932 			cw.want = want;
2933 
2934 			spin_lock(&mdsc->caps_list_lock);
2935 			list_add(&cw.list, &mdsc->cap_wait_list);
2936 			spin_unlock(&mdsc->caps_list_lock);
2937 
2938 			/* make sure used fmode not timeout */
2939 			ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2940 			add_wait_queue(&ci->i_cap_wq, &wait);
2941 
2942 			flags |= NON_BLOCKING;
2943 			while (!(ret = try_get_cap_refs(inode, need, want,
2944 							endoff, flags, &_got))) {
2945 				if (signal_pending(current)) {
2946 					ret = -ERESTARTSYS;
2947 					break;
2948 				}
2949 				wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2950 			}
2951 
2952 			remove_wait_queue(&ci->i_cap_wq, &wait);
2953 			ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2954 
2955 			spin_lock(&mdsc->caps_list_lock);
2956 			list_del(&cw.list);
2957 			spin_unlock(&mdsc->caps_list_lock);
2958 
2959 			if (ret == -EAGAIN)
2960 				continue;
2961 		}
2962 
2963 		if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2964 		    fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2965 			if (ret >= 0 && _got)
2966 				ceph_put_cap_refs(ci, _got);
2967 			return -EBADF;
2968 		}
2969 
2970 		if (ret < 0) {
2971 			if (ret == -EFBIG || ret == -EUCLEAN) {
2972 				int ret2 = ceph_wait_on_async_create(inode);
2973 				if (ret2 < 0)
2974 					return ret2;
2975 			}
2976 			if (ret == -EFBIG) {
2977 				check_max_size(inode, endoff);
2978 				continue;
2979 			}
2980 			if (ret == -EUCLEAN) {
2981 				/* session was killed, try renew caps */
2982 				ret = ceph_renew_caps(inode, flags);
2983 				if (ret == 0)
2984 					continue;
2985 			}
2986 			return ret;
2987 		}
2988 
2989 		if (S_ISREG(ci->netfs.inode.i_mode) &&
2990 		    ceph_has_inline_data(ci) &&
2991 		    (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2992 		    i_size_read(inode) > 0) {
2993 			struct page *page =
2994 				find_get_page(inode->i_mapping, 0);
2995 			if (page) {
2996 				bool uptodate = PageUptodate(page);
2997 
2998 				put_page(page);
2999 				if (uptodate)
3000 					break;
3001 			}
3002 			/*
3003 			 * drop cap refs first because getattr while
3004 			 * holding * caps refs can cause deadlock.
3005 			 */
3006 			ceph_put_cap_refs(ci, _got);
3007 			_got = 0;
3008 
3009 			/*
3010 			 * getattr request will bring inline data into
3011 			 * page cache
3012 			 */
3013 			ret = __ceph_do_getattr(inode, NULL,
3014 						CEPH_STAT_CAP_INLINE_DATA,
3015 						true);
3016 			if (ret < 0)
3017 				return ret;
3018 			continue;
3019 		}
3020 		break;
3021 	}
3022 	*got = _got;
3023 	return 0;
3024 }
3025 
3026 /*
3027  * Take cap refs.  Caller must already know we hold at least one ref
3028  * on the caps in question or we don't know this is safe.
3029  */
ceph_get_cap_refs(struct ceph_inode_info * ci,int caps)3030 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
3031 {
3032 	spin_lock(&ci->i_ceph_lock);
3033 	ceph_take_cap_refs(ci, caps, false);
3034 	spin_unlock(&ci->i_ceph_lock);
3035 }
3036 
3037 
3038 /*
3039  * drop cap_snap that is not associated with any snapshot.
3040  * we don't need to send FLUSHSNAP message for it.
3041  */
ceph_try_drop_cap_snap(struct ceph_inode_info * ci,struct ceph_cap_snap * capsnap)3042 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
3043 				  struct ceph_cap_snap *capsnap)
3044 {
3045 	if (!capsnap->need_flush &&
3046 	    !capsnap->writing && !capsnap->dirty_pages) {
3047 		dout("dropping cap_snap %p follows %llu\n",
3048 		     capsnap, capsnap->follows);
3049 		BUG_ON(capsnap->cap_flush.tid > 0);
3050 		ceph_put_snap_context(capsnap->context);
3051 		if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
3052 			ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3053 
3054 		list_del(&capsnap->ci_item);
3055 		ceph_put_cap_snap(capsnap);
3056 		return 1;
3057 	}
3058 	return 0;
3059 }
3060 
3061 enum put_cap_refs_mode {
3062 	PUT_CAP_REFS_SYNC = 0,
3063 	PUT_CAP_REFS_NO_CHECK,
3064 	PUT_CAP_REFS_ASYNC,
3065 };
3066 
3067 /*
3068  * Release cap refs.
3069  *
3070  * If we released the last ref on any given cap, call ceph_check_caps
3071  * to release (or schedule a release).
3072  *
3073  * If we are releasing a WR cap (from a sync write), finalize any affected
3074  * cap_snap, and wake up any waiters.
3075  */
__ceph_put_cap_refs(struct ceph_inode_info * ci,int had,enum put_cap_refs_mode mode)3076 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3077 				enum put_cap_refs_mode mode)
3078 {
3079 	struct inode *inode = &ci->netfs.inode;
3080 	int last = 0, put = 0, flushsnaps = 0, wake = 0;
3081 	bool check_flushsnaps = false;
3082 
3083 	spin_lock(&ci->i_ceph_lock);
3084 	if (had & CEPH_CAP_PIN)
3085 		--ci->i_pin_ref;
3086 	if (had & CEPH_CAP_FILE_RD)
3087 		if (--ci->i_rd_ref == 0)
3088 			last++;
3089 	if (had & CEPH_CAP_FILE_CACHE)
3090 		if (--ci->i_rdcache_ref == 0)
3091 			last++;
3092 	if (had & CEPH_CAP_FILE_EXCL)
3093 		if (--ci->i_fx_ref == 0)
3094 			last++;
3095 	if (had & CEPH_CAP_FILE_BUFFER) {
3096 		if (--ci->i_wb_ref == 0) {
3097 			last++;
3098 			/* put the ref held by ceph_take_cap_refs() */
3099 			put++;
3100 			check_flushsnaps = true;
3101 		}
3102 		dout("put_cap_refs %p wb %d -> %d (?)\n",
3103 		     inode, ci->i_wb_ref+1, ci->i_wb_ref);
3104 	}
3105 	if (had & CEPH_CAP_FILE_WR) {
3106 		if (--ci->i_wr_ref == 0) {
3107 			last++;
3108 			check_flushsnaps = true;
3109 			if (ci->i_wrbuffer_ref_head == 0 &&
3110 			    ci->i_dirty_caps == 0 &&
3111 			    ci->i_flushing_caps == 0) {
3112 				BUG_ON(!ci->i_head_snapc);
3113 				ceph_put_snap_context(ci->i_head_snapc);
3114 				ci->i_head_snapc = NULL;
3115 			}
3116 			/* see comment in __ceph_remove_cap() */
3117 			if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3118 				ceph_change_snap_realm(inode, NULL);
3119 		}
3120 	}
3121 	if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3122 		struct ceph_cap_snap *capsnap =
3123 			list_last_entry(&ci->i_cap_snaps,
3124 					struct ceph_cap_snap,
3125 					ci_item);
3126 
3127 		capsnap->writing = 0;
3128 		if (ceph_try_drop_cap_snap(ci, capsnap))
3129 			/* put the ref held by ceph_queue_cap_snap() */
3130 			put++;
3131 		else if (__ceph_finish_cap_snap(ci, capsnap))
3132 			flushsnaps = 1;
3133 		wake = 1;
3134 	}
3135 	spin_unlock(&ci->i_ceph_lock);
3136 
3137 	dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3138 	     last ? " last" : "", put ? " put" : "");
3139 
3140 	switch (mode) {
3141 	case PUT_CAP_REFS_SYNC:
3142 		if (last)
3143 			ceph_check_caps(ci, 0, NULL);
3144 		else if (flushsnaps)
3145 			ceph_flush_snaps(ci, NULL);
3146 		break;
3147 	case PUT_CAP_REFS_ASYNC:
3148 		if (last)
3149 			ceph_queue_check_caps(inode);
3150 		else if (flushsnaps)
3151 			ceph_queue_flush_snaps(inode);
3152 		break;
3153 	default:
3154 		break;
3155 	}
3156 	if (wake)
3157 		wake_up_all(&ci->i_cap_wq);
3158 	while (put-- > 0)
3159 		iput(inode);
3160 }
3161 
ceph_put_cap_refs(struct ceph_inode_info * ci,int had)3162 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3163 {
3164 	__ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3165 }
3166 
ceph_put_cap_refs_async(struct ceph_inode_info * ci,int had)3167 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3168 {
3169 	__ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3170 }
3171 
ceph_put_cap_refs_no_check_caps(struct ceph_inode_info * ci,int had)3172 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3173 {
3174 	__ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3175 }
3176 
3177 /*
3178  * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3179  * context.  Adjust per-snap dirty page accounting as appropriate.
3180  * Once all dirty data for a cap_snap is flushed, flush snapped file
3181  * metadata back to the MDS.  If we dropped the last ref, call
3182  * ceph_check_caps.
3183  */
ceph_put_wrbuffer_cap_refs(struct ceph_inode_info * ci,int nr,struct ceph_snap_context * snapc)3184 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3185 				struct ceph_snap_context *snapc)
3186 {
3187 	struct inode *inode = &ci->netfs.inode;
3188 	struct ceph_cap_snap *capsnap = NULL, *iter;
3189 	int put = 0;
3190 	bool last = false;
3191 	bool flush_snaps = false;
3192 	bool complete_capsnap = false;
3193 
3194 	spin_lock(&ci->i_ceph_lock);
3195 	ci->i_wrbuffer_ref -= nr;
3196 	if (ci->i_wrbuffer_ref == 0) {
3197 		last = true;
3198 		put++;
3199 	}
3200 
3201 	if (ci->i_head_snapc == snapc) {
3202 		ci->i_wrbuffer_ref_head -= nr;
3203 		if (ci->i_wrbuffer_ref_head == 0 &&
3204 		    ci->i_wr_ref == 0 &&
3205 		    ci->i_dirty_caps == 0 &&
3206 		    ci->i_flushing_caps == 0) {
3207 			BUG_ON(!ci->i_head_snapc);
3208 			ceph_put_snap_context(ci->i_head_snapc);
3209 			ci->i_head_snapc = NULL;
3210 		}
3211 		dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3212 		     inode,
3213 		     ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3214 		     ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3215 		     last ? " LAST" : "");
3216 	} else {
3217 		list_for_each_entry(iter, &ci->i_cap_snaps, ci_item) {
3218 			if (iter->context == snapc) {
3219 				capsnap = iter;
3220 				break;
3221 			}
3222 		}
3223 
3224 		if (!capsnap) {
3225 			/*
3226 			 * The capsnap should already be removed when removing
3227 			 * auth cap in the case of a forced unmount.
3228 			 */
3229 			WARN_ON_ONCE(ci->i_auth_cap);
3230 			goto unlock;
3231 		}
3232 
3233 		capsnap->dirty_pages -= nr;
3234 		if (capsnap->dirty_pages == 0) {
3235 			complete_capsnap = true;
3236 			if (!capsnap->writing) {
3237 				if (ceph_try_drop_cap_snap(ci, capsnap)) {
3238 					put++;
3239 				} else {
3240 					ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3241 					flush_snaps = true;
3242 				}
3243 			}
3244 		}
3245 		dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3246 		     " snap %lld %d/%d -> %d/%d %s%s\n",
3247 		     inode, capsnap, capsnap->context->seq,
3248 		     ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3249 		     ci->i_wrbuffer_ref, capsnap->dirty_pages,
3250 		     last ? " (wrbuffer last)" : "",
3251 		     complete_capsnap ? " (complete capsnap)" : "");
3252 	}
3253 
3254 unlock:
3255 	spin_unlock(&ci->i_ceph_lock);
3256 
3257 	if (last) {
3258 		ceph_check_caps(ci, 0, NULL);
3259 	} else if (flush_snaps) {
3260 		ceph_flush_snaps(ci, NULL);
3261 	}
3262 	if (complete_capsnap)
3263 		wake_up_all(&ci->i_cap_wq);
3264 	while (put-- > 0) {
3265 		iput(inode);
3266 	}
3267 }
3268 
3269 /*
3270  * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3271  */
invalidate_aliases(struct inode * inode)3272 static void invalidate_aliases(struct inode *inode)
3273 {
3274 	struct dentry *dn, *prev = NULL;
3275 
3276 	dout("invalidate_aliases inode %p\n", inode);
3277 	d_prune_aliases(inode);
3278 	/*
3279 	 * For non-directory inode, d_find_alias() only returns
3280 	 * hashed dentry. After calling d_invalidate(), the
3281 	 * dentry becomes unhashed.
3282 	 *
3283 	 * For directory inode, d_find_alias() can return
3284 	 * unhashed dentry. But directory inode should have
3285 	 * one alias at most.
3286 	 */
3287 	while ((dn = d_find_alias(inode))) {
3288 		if (dn == prev) {
3289 			dput(dn);
3290 			break;
3291 		}
3292 		d_invalidate(dn);
3293 		if (prev)
3294 			dput(prev);
3295 		prev = dn;
3296 	}
3297 	if (prev)
3298 		dput(prev);
3299 }
3300 
3301 struct cap_extra_info {
3302 	struct ceph_string *pool_ns;
3303 	/* inline data */
3304 	u64 inline_version;
3305 	void *inline_data;
3306 	u32 inline_len;
3307 	/* dirstat */
3308 	bool dirstat_valid;
3309 	u64 nfiles;
3310 	u64 nsubdirs;
3311 	u64 change_attr;
3312 	/* currently issued */
3313 	int issued;
3314 	struct timespec64 btime;
3315 };
3316 
3317 /*
3318  * Handle a cap GRANT message from the MDS.  (Note that a GRANT may
3319  * actually be a revocation if it specifies a smaller cap set.)
3320  *
3321  * caller holds s_mutex and i_ceph_lock, we drop both.
3322  */
handle_cap_grant(struct inode * inode,struct ceph_mds_session * session,struct ceph_cap * cap,struct ceph_mds_caps * grant,struct ceph_buffer * xattr_buf,struct cap_extra_info * extra_info)3323 static void handle_cap_grant(struct inode *inode,
3324 			     struct ceph_mds_session *session,
3325 			     struct ceph_cap *cap,
3326 			     struct ceph_mds_caps *grant,
3327 			     struct ceph_buffer *xattr_buf,
3328 			     struct cap_extra_info *extra_info)
3329 	__releases(ci->i_ceph_lock)
3330 	__releases(session->s_mdsc->snap_rwsem)
3331 {
3332 	struct ceph_inode_info *ci = ceph_inode(inode);
3333 	int seq = le32_to_cpu(grant->seq);
3334 	int newcaps = le32_to_cpu(grant->caps);
3335 	int used, wanted, dirty;
3336 	u64 size = le64_to_cpu(grant->size);
3337 	u64 max_size = le64_to_cpu(grant->max_size);
3338 	unsigned char check_caps = 0;
3339 	bool was_stale = cap->cap_gen < atomic_read(&session->s_cap_gen);
3340 	bool wake = false;
3341 	bool writeback = false;
3342 	bool queue_trunc = false;
3343 	bool queue_invalidate = false;
3344 	bool deleted_inode = false;
3345 	bool fill_inline = false;
3346 
3347 	dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3348 	     inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3349 	dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3350 		i_size_read(inode));
3351 
3352 
3353 	/*
3354 	 * If CACHE is being revoked, and we have no dirty buffers,
3355 	 * try to invalidate (once).  (If there are dirty buffers, we
3356 	 * will invalidate _after_ writeback.)
3357 	 */
3358 	if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3359 	    ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3360 	    (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3361 	    !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3362 		if (try_nonblocking_invalidate(inode)) {
3363 			/* there were locked pages.. invalidate later
3364 			   in a separate thread. */
3365 			if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3366 				queue_invalidate = true;
3367 				ci->i_rdcache_revoking = ci->i_rdcache_gen;
3368 			}
3369 		}
3370 	}
3371 
3372 	if (was_stale)
3373 		cap->issued = cap->implemented = CEPH_CAP_PIN;
3374 
3375 	/*
3376 	 * auth mds of the inode changed. we received the cap export message,
3377 	 * but still haven't received the cap import message. handle_cap_export
3378 	 * updated the new auth MDS' cap.
3379 	 *
3380 	 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3381 	 * that was sent before the cap import message. So don't remove caps.
3382 	 */
3383 	if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3384 		WARN_ON(cap != ci->i_auth_cap);
3385 		WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3386 		seq = cap->seq;
3387 		newcaps |= cap->issued;
3388 	}
3389 
3390 	/* side effects now are allowed */
3391 	cap->cap_gen = atomic_read(&session->s_cap_gen);
3392 	cap->seq = seq;
3393 
3394 	__check_cap_issue(ci, cap, newcaps);
3395 
3396 	inode_set_max_iversion_raw(inode, extra_info->change_attr);
3397 
3398 	if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3399 	    (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3400 		umode_t mode = le32_to_cpu(grant->mode);
3401 
3402 		if (inode_wrong_type(inode, mode))
3403 			pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3404 				     ceph_vinop(inode), inode->i_mode, mode);
3405 		else
3406 			inode->i_mode = mode;
3407 		inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3408 		inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3409 		ci->i_btime = extra_info->btime;
3410 		dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3411 		     from_kuid(&init_user_ns, inode->i_uid),
3412 		     from_kgid(&init_user_ns, inode->i_gid));
3413 	}
3414 
3415 	if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3416 	    (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3417 		set_nlink(inode, le32_to_cpu(grant->nlink));
3418 		if (inode->i_nlink == 0)
3419 			deleted_inode = true;
3420 	}
3421 
3422 	if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3423 	    grant->xattr_len) {
3424 		int len = le32_to_cpu(grant->xattr_len);
3425 		u64 version = le64_to_cpu(grant->xattr_version);
3426 
3427 		if (version > ci->i_xattrs.version) {
3428 			dout(" got new xattrs v%llu on %p len %d\n",
3429 			     version, inode, len);
3430 			if (ci->i_xattrs.blob)
3431 				ceph_buffer_put(ci->i_xattrs.blob);
3432 			ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3433 			ci->i_xattrs.version = version;
3434 			ceph_forget_all_cached_acls(inode);
3435 			ceph_security_invalidate_secctx(inode);
3436 		}
3437 	}
3438 
3439 	if (newcaps & CEPH_CAP_ANY_RD) {
3440 		struct timespec64 mtime, atime, ctime;
3441 		/* ctime/mtime/atime? */
3442 		ceph_decode_timespec64(&mtime, &grant->mtime);
3443 		ceph_decode_timespec64(&atime, &grant->atime);
3444 		ceph_decode_timespec64(&ctime, &grant->ctime);
3445 		ceph_fill_file_time(inode, extra_info->issued,
3446 				    le32_to_cpu(grant->time_warp_seq),
3447 				    &ctime, &mtime, &atime);
3448 	}
3449 
3450 	if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3451 		ci->i_files = extra_info->nfiles;
3452 		ci->i_subdirs = extra_info->nsubdirs;
3453 	}
3454 
3455 	if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3456 		/* file layout may have changed */
3457 		s64 old_pool = ci->i_layout.pool_id;
3458 		struct ceph_string *old_ns;
3459 
3460 		ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3461 		old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3462 					lockdep_is_held(&ci->i_ceph_lock));
3463 		rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3464 
3465 		if (ci->i_layout.pool_id != old_pool ||
3466 		    extra_info->pool_ns != old_ns)
3467 			ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3468 
3469 		extra_info->pool_ns = old_ns;
3470 
3471 		/* size/truncate_seq? */
3472 		queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3473 					le32_to_cpu(grant->truncate_seq),
3474 					le64_to_cpu(grant->truncate_size),
3475 					size);
3476 	}
3477 
3478 	if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3479 		if (max_size != ci->i_max_size) {
3480 			dout("max_size %lld -> %llu\n",
3481 			     ci->i_max_size, max_size);
3482 			ci->i_max_size = max_size;
3483 			if (max_size >= ci->i_wanted_max_size) {
3484 				ci->i_wanted_max_size = 0;  /* reset */
3485 				ci->i_requested_max_size = 0;
3486 			}
3487 			wake = true;
3488 		}
3489 	}
3490 
3491 	/* check cap bits */
3492 	wanted = __ceph_caps_wanted(ci);
3493 	used = __ceph_caps_used(ci);
3494 	dirty = __ceph_caps_dirty(ci);
3495 	dout(" my wanted = %s, used = %s, dirty %s\n",
3496 	     ceph_cap_string(wanted),
3497 	     ceph_cap_string(used),
3498 	     ceph_cap_string(dirty));
3499 
3500 	if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3501 	    (wanted & ~(cap->mds_wanted | newcaps))) {
3502 		/*
3503 		 * If mds is importing cap, prior cap messages that update
3504 		 * 'wanted' may get dropped by mds (migrate seq mismatch).
3505 		 *
3506 		 * We don't send cap message to update 'wanted' if what we
3507 		 * want are already issued. If mds revokes caps, cap message
3508 		 * that releases caps also tells mds what we want. But if
3509 		 * caps got revoked by mds forcedly (session stale). We may
3510 		 * haven't told mds what we want.
3511 		 */
3512 		check_caps = 1;
3513 	}
3514 
3515 	/* revocation, grant, or no-op? */
3516 	if (cap->issued & ~newcaps) {
3517 		int revoking = cap->issued & ~newcaps;
3518 
3519 		dout("revocation: %s -> %s (revoking %s)\n",
3520 		     ceph_cap_string(cap->issued),
3521 		     ceph_cap_string(newcaps),
3522 		     ceph_cap_string(revoking));
3523 		if (S_ISREG(inode->i_mode) &&
3524 		    (revoking & used & CEPH_CAP_FILE_BUFFER))
3525 			writeback = true;  /* initiate writeback; will delay ack */
3526 		else if (queue_invalidate &&
3527 			 revoking == CEPH_CAP_FILE_CACHE &&
3528 			 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3529 			; /* do nothing yet, invalidation will be queued */
3530 		else if (cap == ci->i_auth_cap)
3531 			check_caps = 1; /* check auth cap only */
3532 		else
3533 			check_caps = 2; /* check all caps */
3534 		/* If there is new caps, try to wake up the waiters */
3535 		if (~cap->issued & newcaps)
3536 			wake = true;
3537 		cap->issued = newcaps;
3538 		cap->implemented |= newcaps;
3539 	} else if (cap->issued == newcaps) {
3540 		dout("caps unchanged: %s -> %s\n",
3541 		     ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3542 	} else {
3543 		dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3544 		     ceph_cap_string(newcaps));
3545 		/* non-auth MDS is revoking the newly grant caps ? */
3546 		if (cap == ci->i_auth_cap &&
3547 		    __ceph_caps_revoking_other(ci, cap, newcaps))
3548 		    check_caps = 2;
3549 
3550 		cap->issued = newcaps;
3551 		cap->implemented |= newcaps; /* add bits only, to
3552 					      * avoid stepping on a
3553 					      * pending revocation */
3554 		wake = true;
3555 	}
3556 	BUG_ON(cap->issued & ~cap->implemented);
3557 
3558 	if (extra_info->inline_version > 0 &&
3559 	    extra_info->inline_version >= ci->i_inline_version) {
3560 		ci->i_inline_version = extra_info->inline_version;
3561 		if (ci->i_inline_version != CEPH_INLINE_NONE &&
3562 		    (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3563 			fill_inline = true;
3564 	}
3565 
3566 	if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3567 		if (ci->i_auth_cap == cap) {
3568 			if (newcaps & ~extra_info->issued)
3569 				wake = true;
3570 
3571 			if (ci->i_requested_max_size > max_size ||
3572 			    !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3573 				/* re-request max_size if necessary */
3574 				ci->i_requested_max_size = 0;
3575 				wake = true;
3576 			}
3577 
3578 			ceph_kick_flushing_inode_caps(session, ci);
3579 		}
3580 		up_read(&session->s_mdsc->snap_rwsem);
3581 	}
3582 	spin_unlock(&ci->i_ceph_lock);
3583 
3584 	if (fill_inline)
3585 		ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3586 				      extra_info->inline_len);
3587 
3588 	if (queue_trunc)
3589 		ceph_queue_vmtruncate(inode);
3590 
3591 	if (writeback)
3592 		/*
3593 		 * queue inode for writeback: we can't actually call
3594 		 * filemap_write_and_wait, etc. from message handler
3595 		 * context.
3596 		 */
3597 		ceph_queue_writeback(inode);
3598 	if (queue_invalidate)
3599 		ceph_queue_invalidate(inode);
3600 	if (deleted_inode)
3601 		invalidate_aliases(inode);
3602 	if (wake)
3603 		wake_up_all(&ci->i_cap_wq);
3604 
3605 	mutex_unlock(&session->s_mutex);
3606 	if (check_caps == 1)
3607 		ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3608 				session);
3609 	else if (check_caps == 2)
3610 		ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3611 }
3612 
3613 /*
3614  * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3615  * MDS has been safely committed.
3616  */
handle_cap_flush_ack(struct inode * inode,u64 flush_tid,struct ceph_mds_caps * m,struct ceph_mds_session * session,struct ceph_cap * cap)3617 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3618 				 struct ceph_mds_caps *m,
3619 				 struct ceph_mds_session *session,
3620 				 struct ceph_cap *cap)
3621 	__releases(ci->i_ceph_lock)
3622 {
3623 	struct ceph_inode_info *ci = ceph_inode(inode);
3624 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3625 	struct ceph_cap_flush *cf, *tmp_cf;
3626 	LIST_HEAD(to_remove);
3627 	unsigned seq = le32_to_cpu(m->seq);
3628 	int dirty = le32_to_cpu(m->dirty);
3629 	int cleaned = 0;
3630 	bool drop = false;
3631 	bool wake_ci = false;
3632 	bool wake_mdsc = false;
3633 
3634 	list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3635 		/* Is this the one that was flushed? */
3636 		if (cf->tid == flush_tid)
3637 			cleaned = cf->caps;
3638 
3639 		/* Is this a capsnap? */
3640 		if (cf->is_capsnap)
3641 			continue;
3642 
3643 		if (cf->tid <= flush_tid) {
3644 			/*
3645 			 * An earlier or current tid. The FLUSH_ACK should
3646 			 * represent a superset of this flush's caps.
3647 			 */
3648 			wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3649 			list_add_tail(&cf->i_list, &to_remove);
3650 		} else {
3651 			/*
3652 			 * This is a later one. Any caps in it are still dirty
3653 			 * so don't count them as cleaned.
3654 			 */
3655 			cleaned &= ~cf->caps;
3656 			if (!cleaned)
3657 				break;
3658 		}
3659 	}
3660 
3661 	dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3662 	     " flushing %s -> %s\n",
3663 	     inode, session->s_mds, seq, ceph_cap_string(dirty),
3664 	     ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3665 	     ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3666 
3667 	if (list_empty(&to_remove) && !cleaned)
3668 		goto out;
3669 
3670 	ci->i_flushing_caps &= ~cleaned;
3671 
3672 	spin_lock(&mdsc->cap_dirty_lock);
3673 
3674 	list_for_each_entry(cf, &to_remove, i_list)
3675 		wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3676 
3677 	if (ci->i_flushing_caps == 0) {
3678 		if (list_empty(&ci->i_cap_flush_list)) {
3679 			list_del_init(&ci->i_flushing_item);
3680 			if (!list_empty(&session->s_cap_flushing)) {
3681 				dout(" mds%d still flushing cap on %p\n",
3682 				     session->s_mds,
3683 				     &list_first_entry(&session->s_cap_flushing,
3684 						struct ceph_inode_info,
3685 						i_flushing_item)->netfs.inode);
3686 			}
3687 		}
3688 		mdsc->num_cap_flushing--;
3689 		dout(" inode %p now !flushing\n", inode);
3690 
3691 		if (ci->i_dirty_caps == 0) {
3692 			dout(" inode %p now clean\n", inode);
3693 			BUG_ON(!list_empty(&ci->i_dirty_item));
3694 			drop = true;
3695 			if (ci->i_wr_ref == 0 &&
3696 			    ci->i_wrbuffer_ref_head == 0) {
3697 				BUG_ON(!ci->i_head_snapc);
3698 				ceph_put_snap_context(ci->i_head_snapc);
3699 				ci->i_head_snapc = NULL;
3700 			}
3701 		} else {
3702 			BUG_ON(list_empty(&ci->i_dirty_item));
3703 		}
3704 	}
3705 	spin_unlock(&mdsc->cap_dirty_lock);
3706 
3707 out:
3708 	spin_unlock(&ci->i_ceph_lock);
3709 
3710 	while (!list_empty(&to_remove)) {
3711 		cf = list_first_entry(&to_remove,
3712 				      struct ceph_cap_flush, i_list);
3713 		list_del_init(&cf->i_list);
3714 		if (!cf->is_capsnap)
3715 			ceph_free_cap_flush(cf);
3716 	}
3717 
3718 	if (wake_ci)
3719 		wake_up_all(&ci->i_cap_wq);
3720 	if (wake_mdsc)
3721 		wake_up_all(&mdsc->cap_flushing_wq);
3722 	if (drop)
3723 		iput(inode);
3724 }
3725 
__ceph_remove_capsnap(struct inode * inode,struct ceph_cap_snap * capsnap,bool * wake_ci,bool * wake_mdsc)3726 void __ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3727 			   bool *wake_ci, bool *wake_mdsc)
3728 {
3729 	struct ceph_inode_info *ci = ceph_inode(inode);
3730 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3731 	bool ret;
3732 
3733 	lockdep_assert_held(&ci->i_ceph_lock);
3734 
3735 	dout("removing capsnap %p, inode %p ci %p\n", capsnap, inode, ci);
3736 
3737 	list_del_init(&capsnap->ci_item);
3738 	ret = __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3739 	if (wake_ci)
3740 		*wake_ci = ret;
3741 
3742 	spin_lock(&mdsc->cap_dirty_lock);
3743 	if (list_empty(&ci->i_cap_flush_list))
3744 		list_del_init(&ci->i_flushing_item);
3745 
3746 	ret = __detach_cap_flush_from_mdsc(mdsc, &capsnap->cap_flush);
3747 	if (wake_mdsc)
3748 		*wake_mdsc = ret;
3749 	spin_unlock(&mdsc->cap_dirty_lock);
3750 }
3751 
ceph_remove_capsnap(struct inode * inode,struct ceph_cap_snap * capsnap,bool * wake_ci,bool * wake_mdsc)3752 void ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3753 			 bool *wake_ci, bool *wake_mdsc)
3754 {
3755 	struct ceph_inode_info *ci = ceph_inode(inode);
3756 
3757 	lockdep_assert_held(&ci->i_ceph_lock);
3758 
3759 	WARN_ON_ONCE(capsnap->dirty_pages || capsnap->writing);
3760 	__ceph_remove_capsnap(inode, capsnap, wake_ci, wake_mdsc);
3761 }
3762 
3763 /*
3764  * Handle FLUSHSNAP_ACK.  MDS has flushed snap data to disk and we can
3765  * throw away our cap_snap.
3766  *
3767  * Caller hold s_mutex.
3768  */
handle_cap_flushsnap_ack(struct inode * inode,u64 flush_tid,struct ceph_mds_caps * m,struct ceph_mds_session * session)3769 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3770 				     struct ceph_mds_caps *m,
3771 				     struct ceph_mds_session *session)
3772 {
3773 	struct ceph_inode_info *ci = ceph_inode(inode);
3774 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3775 	u64 follows = le64_to_cpu(m->snap_follows);
3776 	struct ceph_cap_snap *capsnap = NULL, *iter;
3777 	bool wake_ci = false;
3778 	bool wake_mdsc = false;
3779 
3780 	dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3781 	     inode, ci, session->s_mds, follows);
3782 
3783 	spin_lock(&ci->i_ceph_lock);
3784 	list_for_each_entry(iter, &ci->i_cap_snaps, ci_item) {
3785 		if (iter->follows == follows) {
3786 			if (iter->cap_flush.tid != flush_tid) {
3787 				dout(" cap_snap %p follows %lld tid %lld !="
3788 				     " %lld\n", iter, follows,
3789 				     flush_tid, iter->cap_flush.tid);
3790 				break;
3791 			}
3792 			capsnap = iter;
3793 			break;
3794 		} else {
3795 			dout(" skipping cap_snap %p follows %lld\n",
3796 			     iter, iter->follows);
3797 		}
3798 	}
3799 	if (capsnap)
3800 		ceph_remove_capsnap(inode, capsnap, &wake_ci, &wake_mdsc);
3801 	spin_unlock(&ci->i_ceph_lock);
3802 
3803 	if (capsnap) {
3804 		ceph_put_snap_context(capsnap->context);
3805 		ceph_put_cap_snap(capsnap);
3806 		if (wake_ci)
3807 			wake_up_all(&ci->i_cap_wq);
3808 		if (wake_mdsc)
3809 			wake_up_all(&mdsc->cap_flushing_wq);
3810 		iput(inode);
3811 	}
3812 }
3813 
3814 /*
3815  * Handle TRUNC from MDS, indicating file truncation.
3816  *
3817  * caller hold s_mutex.
3818  */
handle_cap_trunc(struct inode * inode,struct ceph_mds_caps * trunc,struct ceph_mds_session * session)3819 static bool handle_cap_trunc(struct inode *inode,
3820 			     struct ceph_mds_caps *trunc,
3821 			     struct ceph_mds_session *session)
3822 {
3823 	struct ceph_inode_info *ci = ceph_inode(inode);
3824 	int mds = session->s_mds;
3825 	int seq = le32_to_cpu(trunc->seq);
3826 	u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3827 	u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3828 	u64 size = le64_to_cpu(trunc->size);
3829 	int implemented = 0;
3830 	int dirty = __ceph_caps_dirty(ci);
3831 	int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3832 	bool queue_trunc = false;
3833 
3834 	lockdep_assert_held(&ci->i_ceph_lock);
3835 
3836 	issued |= implemented | dirty;
3837 
3838 	dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3839 	     inode, mds, seq, truncate_size, truncate_seq);
3840 	queue_trunc = ceph_fill_file_size(inode, issued,
3841 					  truncate_seq, truncate_size, size);
3842 	return queue_trunc;
3843 }
3844 
3845 /*
3846  * Handle EXPORT from MDS.  Cap is being migrated _from_ this mds to a
3847  * different one.  If we are the most recent migration we've seen (as
3848  * indicated by mseq), make note of the migrating cap bits for the
3849  * duration (until we see the corresponding IMPORT).
3850  *
3851  * caller holds s_mutex
3852  */
handle_cap_export(struct inode * inode,struct ceph_mds_caps * ex,struct ceph_mds_cap_peer * ph,struct ceph_mds_session * session)3853 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3854 			      struct ceph_mds_cap_peer *ph,
3855 			      struct ceph_mds_session *session)
3856 {
3857 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3858 	struct ceph_mds_session *tsession = NULL;
3859 	struct ceph_cap *cap, *tcap, *new_cap = NULL;
3860 	struct ceph_inode_info *ci = ceph_inode(inode);
3861 	u64 t_cap_id;
3862 	unsigned mseq = le32_to_cpu(ex->migrate_seq);
3863 	unsigned t_seq, t_mseq;
3864 	int target, issued;
3865 	int mds = session->s_mds;
3866 
3867 	if (ph) {
3868 		t_cap_id = le64_to_cpu(ph->cap_id);
3869 		t_seq = le32_to_cpu(ph->seq);
3870 		t_mseq = le32_to_cpu(ph->mseq);
3871 		target = le32_to_cpu(ph->mds);
3872 	} else {
3873 		t_cap_id = t_seq = t_mseq = 0;
3874 		target = -1;
3875 	}
3876 
3877 	dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3878 	     inode, ci, mds, mseq, target);
3879 retry:
3880 	down_read(&mdsc->snap_rwsem);
3881 	spin_lock(&ci->i_ceph_lock);
3882 	cap = __get_cap_for_mds(ci, mds);
3883 	if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3884 		goto out_unlock;
3885 
3886 	if (target < 0) {
3887 		ceph_remove_cap(cap, false);
3888 		goto out_unlock;
3889 	}
3890 
3891 	/*
3892 	 * now we know we haven't received the cap import message yet
3893 	 * because the exported cap still exist.
3894 	 */
3895 
3896 	issued = cap->issued;
3897 	if (issued != cap->implemented)
3898 		pr_err_ratelimited("handle_cap_export: issued != implemented: "
3899 				"ino (%llx.%llx) mds%d seq %d mseq %d "
3900 				"issued %s implemented %s\n",
3901 				ceph_vinop(inode), mds, cap->seq, cap->mseq,
3902 				ceph_cap_string(issued),
3903 				ceph_cap_string(cap->implemented));
3904 
3905 
3906 	tcap = __get_cap_for_mds(ci, target);
3907 	if (tcap) {
3908 		/* already have caps from the target */
3909 		if (tcap->cap_id == t_cap_id &&
3910 		    ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3911 			dout(" updating import cap %p mds%d\n", tcap, target);
3912 			tcap->cap_id = t_cap_id;
3913 			tcap->seq = t_seq - 1;
3914 			tcap->issue_seq = t_seq - 1;
3915 			tcap->issued |= issued;
3916 			tcap->implemented |= issued;
3917 			if (cap == ci->i_auth_cap) {
3918 				ci->i_auth_cap = tcap;
3919 				change_auth_cap_ses(ci, tcap->session);
3920 			}
3921 		}
3922 		ceph_remove_cap(cap, false);
3923 		goto out_unlock;
3924 	} else if (tsession) {
3925 		/* add placeholder for the export tagert */
3926 		int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3927 		tcap = new_cap;
3928 		ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3929 			     t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3930 
3931 		if (!list_empty(&ci->i_cap_flush_list) &&
3932 		    ci->i_auth_cap == tcap) {
3933 			spin_lock(&mdsc->cap_dirty_lock);
3934 			list_move_tail(&ci->i_flushing_item,
3935 				       &tcap->session->s_cap_flushing);
3936 			spin_unlock(&mdsc->cap_dirty_lock);
3937 		}
3938 
3939 		ceph_remove_cap(cap, false);
3940 		goto out_unlock;
3941 	}
3942 
3943 	spin_unlock(&ci->i_ceph_lock);
3944 	up_read(&mdsc->snap_rwsem);
3945 	mutex_unlock(&session->s_mutex);
3946 
3947 	/* open target session */
3948 	tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3949 	if (!IS_ERR(tsession)) {
3950 		if (mds > target) {
3951 			mutex_lock(&session->s_mutex);
3952 			mutex_lock_nested(&tsession->s_mutex,
3953 					  SINGLE_DEPTH_NESTING);
3954 		} else {
3955 			mutex_lock(&tsession->s_mutex);
3956 			mutex_lock_nested(&session->s_mutex,
3957 					  SINGLE_DEPTH_NESTING);
3958 		}
3959 		new_cap = ceph_get_cap(mdsc, NULL);
3960 	} else {
3961 		WARN_ON(1);
3962 		tsession = NULL;
3963 		target = -1;
3964 		mutex_lock(&session->s_mutex);
3965 	}
3966 	goto retry;
3967 
3968 out_unlock:
3969 	spin_unlock(&ci->i_ceph_lock);
3970 	up_read(&mdsc->snap_rwsem);
3971 	mutex_unlock(&session->s_mutex);
3972 	if (tsession) {
3973 		mutex_unlock(&tsession->s_mutex);
3974 		ceph_put_mds_session(tsession);
3975 	}
3976 	if (new_cap)
3977 		ceph_put_cap(mdsc, new_cap);
3978 }
3979 
3980 /*
3981  * Handle cap IMPORT.
3982  *
3983  * caller holds s_mutex. acquires i_ceph_lock
3984  */
handle_cap_import(struct ceph_mds_client * mdsc,struct inode * inode,struct ceph_mds_caps * im,struct ceph_mds_cap_peer * ph,struct ceph_mds_session * session,struct ceph_cap ** target_cap,int * old_issued)3985 static void handle_cap_import(struct ceph_mds_client *mdsc,
3986 			      struct inode *inode, struct ceph_mds_caps *im,
3987 			      struct ceph_mds_cap_peer *ph,
3988 			      struct ceph_mds_session *session,
3989 			      struct ceph_cap **target_cap, int *old_issued)
3990 {
3991 	struct ceph_inode_info *ci = ceph_inode(inode);
3992 	struct ceph_cap *cap, *ocap, *new_cap = NULL;
3993 	int mds = session->s_mds;
3994 	int issued;
3995 	unsigned caps = le32_to_cpu(im->caps);
3996 	unsigned wanted = le32_to_cpu(im->wanted);
3997 	unsigned seq = le32_to_cpu(im->seq);
3998 	unsigned mseq = le32_to_cpu(im->migrate_seq);
3999 	u64 realmino = le64_to_cpu(im->realm);
4000 	u64 cap_id = le64_to_cpu(im->cap_id);
4001 	u64 p_cap_id;
4002 	int peer;
4003 
4004 	if (ph) {
4005 		p_cap_id = le64_to_cpu(ph->cap_id);
4006 		peer = le32_to_cpu(ph->mds);
4007 	} else {
4008 		p_cap_id = 0;
4009 		peer = -1;
4010 	}
4011 
4012 	dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
4013 	     inode, ci, mds, mseq, peer);
4014 retry:
4015 	cap = __get_cap_for_mds(ci, mds);
4016 	if (!cap) {
4017 		if (!new_cap) {
4018 			spin_unlock(&ci->i_ceph_lock);
4019 			new_cap = ceph_get_cap(mdsc, NULL);
4020 			spin_lock(&ci->i_ceph_lock);
4021 			goto retry;
4022 		}
4023 		cap = new_cap;
4024 	} else {
4025 		if (new_cap) {
4026 			ceph_put_cap(mdsc, new_cap);
4027 			new_cap = NULL;
4028 		}
4029 	}
4030 
4031 	__ceph_caps_issued(ci, &issued);
4032 	issued |= __ceph_caps_dirty(ci);
4033 
4034 	ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
4035 		     realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
4036 
4037 	ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
4038 	if (ocap && ocap->cap_id == p_cap_id) {
4039 		dout(" remove export cap %p mds%d flags %d\n",
4040 		     ocap, peer, ph->flags);
4041 		if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
4042 		    (ocap->seq != le32_to_cpu(ph->seq) ||
4043 		     ocap->mseq != le32_to_cpu(ph->mseq))) {
4044 			pr_err_ratelimited("handle_cap_import: "
4045 					"mismatched seq/mseq: ino (%llx.%llx) "
4046 					"mds%d seq %d mseq %d importer mds%d "
4047 					"has peer seq %d mseq %d\n",
4048 					ceph_vinop(inode), peer, ocap->seq,
4049 					ocap->mseq, mds, le32_to_cpu(ph->seq),
4050 					le32_to_cpu(ph->mseq));
4051 		}
4052 		ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
4053 	}
4054 
4055 	*old_issued = issued;
4056 	*target_cap = cap;
4057 }
4058 
4059 /*
4060  * Handle a caps message from the MDS.
4061  *
4062  * Identify the appropriate session, inode, and call the right handler
4063  * based on the cap op.
4064  */
ceph_handle_caps(struct ceph_mds_session * session,struct ceph_msg * msg)4065 void ceph_handle_caps(struct ceph_mds_session *session,
4066 		      struct ceph_msg *msg)
4067 {
4068 	struct ceph_mds_client *mdsc = session->s_mdsc;
4069 	struct inode *inode;
4070 	struct ceph_inode_info *ci;
4071 	struct ceph_cap *cap;
4072 	struct ceph_mds_caps *h;
4073 	struct ceph_mds_cap_peer *peer = NULL;
4074 	struct ceph_snap_realm *realm = NULL;
4075 	int op;
4076 	int msg_version = le16_to_cpu(msg->hdr.version);
4077 	u32 seq, mseq;
4078 	struct ceph_vino vino;
4079 	void *snaptrace;
4080 	size_t snaptrace_len;
4081 	void *p, *end;
4082 	struct cap_extra_info extra_info = {};
4083 	bool queue_trunc;
4084 
4085 	dout("handle_caps from mds%d\n", session->s_mds);
4086 
4087 	/* decode */
4088 	end = msg->front.iov_base + msg->front.iov_len;
4089 	if (msg->front.iov_len < sizeof(*h))
4090 		goto bad;
4091 	h = msg->front.iov_base;
4092 	op = le32_to_cpu(h->op);
4093 	vino.ino = le64_to_cpu(h->ino);
4094 	vino.snap = CEPH_NOSNAP;
4095 	seq = le32_to_cpu(h->seq);
4096 	mseq = le32_to_cpu(h->migrate_seq);
4097 
4098 	snaptrace = h + 1;
4099 	snaptrace_len = le32_to_cpu(h->snap_trace_len);
4100 	p = snaptrace + snaptrace_len;
4101 
4102 	if (msg_version >= 2) {
4103 		u32 flock_len;
4104 		ceph_decode_32_safe(&p, end, flock_len, bad);
4105 		if (p + flock_len > end)
4106 			goto bad;
4107 		p += flock_len;
4108 	}
4109 
4110 	if (msg_version >= 3) {
4111 		if (op == CEPH_CAP_OP_IMPORT) {
4112 			if (p + sizeof(*peer) > end)
4113 				goto bad;
4114 			peer = p;
4115 			p += sizeof(*peer);
4116 		} else if (op == CEPH_CAP_OP_EXPORT) {
4117 			/* recorded in unused fields */
4118 			peer = (void *)&h->size;
4119 		}
4120 	}
4121 
4122 	if (msg_version >= 4) {
4123 		ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4124 		ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4125 		if (p + extra_info.inline_len > end)
4126 			goto bad;
4127 		extra_info.inline_data = p;
4128 		p += extra_info.inline_len;
4129 	}
4130 
4131 	if (msg_version >= 5) {
4132 		struct ceph_osd_client	*osdc = &mdsc->fsc->client->osdc;
4133 		u32			epoch_barrier;
4134 
4135 		ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4136 		ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4137 	}
4138 
4139 	if (msg_version >= 8) {
4140 		u32 pool_ns_len;
4141 
4142 		/* version >= 6 */
4143 		ceph_decode_skip_64(&p, end, bad);	// flush_tid
4144 		/* version >= 7 */
4145 		ceph_decode_skip_32(&p, end, bad);	// caller_uid
4146 		ceph_decode_skip_32(&p, end, bad);	// caller_gid
4147 		/* version >= 8 */
4148 		ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4149 		if (pool_ns_len > 0) {
4150 			ceph_decode_need(&p, end, pool_ns_len, bad);
4151 			extra_info.pool_ns =
4152 				ceph_find_or_create_string(p, pool_ns_len);
4153 			p += pool_ns_len;
4154 		}
4155 	}
4156 
4157 	if (msg_version >= 9) {
4158 		struct ceph_timespec *btime;
4159 
4160 		if (p + sizeof(*btime) > end)
4161 			goto bad;
4162 		btime = p;
4163 		ceph_decode_timespec64(&extra_info.btime, btime);
4164 		p += sizeof(*btime);
4165 		ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4166 	}
4167 
4168 	if (msg_version >= 11) {
4169 		/* version >= 10 */
4170 		ceph_decode_skip_32(&p, end, bad); // flags
4171 		/* version >= 11 */
4172 		extra_info.dirstat_valid = true;
4173 		ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4174 		ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4175 	}
4176 
4177 	/* lookup ino */
4178 	inode = ceph_find_inode(mdsc->fsc->sb, vino);
4179 	dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4180 	     vino.snap, inode);
4181 
4182 	mutex_lock(&session->s_mutex);
4183 	inc_session_sequence(session);
4184 	dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4185 	     (unsigned)seq);
4186 
4187 	if (!inode) {
4188 		dout(" i don't have ino %llx\n", vino.ino);
4189 
4190 		if (op == CEPH_CAP_OP_IMPORT) {
4191 			cap = ceph_get_cap(mdsc, NULL);
4192 			cap->cap_ino = vino.ino;
4193 			cap->queue_release = 1;
4194 			cap->cap_id = le64_to_cpu(h->cap_id);
4195 			cap->mseq = mseq;
4196 			cap->seq = seq;
4197 			cap->issue_seq = seq;
4198 			spin_lock(&session->s_cap_lock);
4199 			__ceph_queue_cap_release(session, cap);
4200 			spin_unlock(&session->s_cap_lock);
4201 		}
4202 		goto flush_cap_releases;
4203 	}
4204 	ci = ceph_inode(inode);
4205 
4206 	/* these will work even if we don't have a cap yet */
4207 	switch (op) {
4208 	case CEPH_CAP_OP_FLUSHSNAP_ACK:
4209 		handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4210 					 h, session);
4211 		goto done;
4212 
4213 	case CEPH_CAP_OP_EXPORT:
4214 		handle_cap_export(inode, h, peer, session);
4215 		goto done_unlocked;
4216 
4217 	case CEPH_CAP_OP_IMPORT:
4218 		realm = NULL;
4219 		if (snaptrace_len) {
4220 			down_write(&mdsc->snap_rwsem);
4221 			ceph_update_snap_trace(mdsc, snaptrace,
4222 					       snaptrace + snaptrace_len,
4223 					       false, &realm);
4224 			downgrade_write(&mdsc->snap_rwsem);
4225 		} else {
4226 			down_read(&mdsc->snap_rwsem);
4227 		}
4228 		spin_lock(&ci->i_ceph_lock);
4229 		handle_cap_import(mdsc, inode, h, peer, session,
4230 				  &cap, &extra_info.issued);
4231 		handle_cap_grant(inode, session, cap,
4232 				 h, msg->middle, &extra_info);
4233 		if (realm)
4234 			ceph_put_snap_realm(mdsc, realm);
4235 		goto done_unlocked;
4236 	}
4237 
4238 	/* the rest require a cap */
4239 	spin_lock(&ci->i_ceph_lock);
4240 	cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4241 	if (!cap) {
4242 		dout(" no cap on %p ino %llx.%llx from mds%d\n",
4243 		     inode, ceph_ino(inode), ceph_snap(inode),
4244 		     session->s_mds);
4245 		spin_unlock(&ci->i_ceph_lock);
4246 		goto flush_cap_releases;
4247 	}
4248 
4249 	/* note that each of these drops i_ceph_lock for us */
4250 	switch (op) {
4251 	case CEPH_CAP_OP_REVOKE:
4252 	case CEPH_CAP_OP_GRANT:
4253 		__ceph_caps_issued(ci, &extra_info.issued);
4254 		extra_info.issued |= __ceph_caps_dirty(ci);
4255 		handle_cap_grant(inode, session, cap,
4256 				 h, msg->middle, &extra_info);
4257 		goto done_unlocked;
4258 
4259 	case CEPH_CAP_OP_FLUSH_ACK:
4260 		handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4261 				     h, session, cap);
4262 		break;
4263 
4264 	case CEPH_CAP_OP_TRUNC:
4265 		queue_trunc = handle_cap_trunc(inode, h, session);
4266 		spin_unlock(&ci->i_ceph_lock);
4267 		if (queue_trunc)
4268 			ceph_queue_vmtruncate(inode);
4269 		break;
4270 
4271 	default:
4272 		spin_unlock(&ci->i_ceph_lock);
4273 		pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4274 		       ceph_cap_op_name(op));
4275 	}
4276 
4277 done:
4278 	mutex_unlock(&session->s_mutex);
4279 done_unlocked:
4280 	iput(inode);
4281 out:
4282 	ceph_put_string(extra_info.pool_ns);
4283 	return;
4284 
4285 flush_cap_releases:
4286 	/*
4287 	 * send any cap release message to try to move things
4288 	 * along for the mds (who clearly thinks we still have this
4289 	 * cap).
4290 	 */
4291 	ceph_flush_cap_releases(mdsc, session);
4292 	goto done;
4293 
4294 bad:
4295 	pr_err("ceph_handle_caps: corrupt message\n");
4296 	ceph_msg_dump(msg);
4297 	goto out;
4298 }
4299 
4300 /*
4301  * Delayed work handler to process end of delayed cap release LRU list.
4302  *
4303  * If new caps are added to the list while processing it, these won't get
4304  * processed in this run.  In this case, the ci->i_hold_caps_max will be
4305  * returned so that the work can be scheduled accordingly.
4306  */
ceph_check_delayed_caps(struct ceph_mds_client * mdsc)4307 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4308 {
4309 	struct inode *inode;
4310 	struct ceph_inode_info *ci;
4311 	struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4312 	unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4313 	unsigned long loop_start = jiffies;
4314 	unsigned long delay = 0;
4315 
4316 	dout("check_delayed_caps\n");
4317 	spin_lock(&mdsc->cap_delay_lock);
4318 	while (!list_empty(&mdsc->cap_delay_list)) {
4319 		ci = list_first_entry(&mdsc->cap_delay_list,
4320 				      struct ceph_inode_info,
4321 				      i_cap_delay_list);
4322 		if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4323 			dout("%s caps added recently.  Exiting loop", __func__);
4324 			delay = ci->i_hold_caps_max;
4325 			break;
4326 		}
4327 		if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4328 		    time_before(jiffies, ci->i_hold_caps_max))
4329 			break;
4330 		list_del_init(&ci->i_cap_delay_list);
4331 
4332 		inode = igrab(&ci->netfs.inode);
4333 		if (inode) {
4334 			spin_unlock(&mdsc->cap_delay_lock);
4335 			dout("check_delayed_caps on %p\n", inode);
4336 			ceph_check_caps(ci, 0, NULL);
4337 			iput(inode);
4338 			spin_lock(&mdsc->cap_delay_lock);
4339 		}
4340 	}
4341 	spin_unlock(&mdsc->cap_delay_lock);
4342 
4343 	return delay;
4344 }
4345 
4346 /*
4347  * Flush all dirty caps to the mds
4348  */
flush_dirty_session_caps(struct ceph_mds_session * s)4349 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4350 {
4351 	struct ceph_mds_client *mdsc = s->s_mdsc;
4352 	struct ceph_inode_info *ci;
4353 	struct inode *inode;
4354 
4355 	dout("flush_dirty_caps\n");
4356 	spin_lock(&mdsc->cap_dirty_lock);
4357 	while (!list_empty(&s->s_cap_dirty)) {
4358 		ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4359 				      i_dirty_item);
4360 		inode = &ci->netfs.inode;
4361 		ihold(inode);
4362 		dout("flush_dirty_caps %llx.%llx\n", ceph_vinop(inode));
4363 		spin_unlock(&mdsc->cap_dirty_lock);
4364 		ceph_wait_on_async_create(inode);
4365 		ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4366 		iput(inode);
4367 		spin_lock(&mdsc->cap_dirty_lock);
4368 	}
4369 	spin_unlock(&mdsc->cap_dirty_lock);
4370 	dout("flush_dirty_caps done\n");
4371 }
4372 
ceph_flush_dirty_caps(struct ceph_mds_client * mdsc)4373 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4374 {
4375 	ceph_mdsc_iterate_sessions(mdsc, flush_dirty_session_caps, true);
4376 }
4377 
__ceph_touch_fmode(struct ceph_inode_info * ci,struct ceph_mds_client * mdsc,int fmode)4378 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4379 			struct ceph_mds_client *mdsc, int fmode)
4380 {
4381 	unsigned long now = jiffies;
4382 	if (fmode & CEPH_FILE_MODE_RD)
4383 		ci->i_last_rd = now;
4384 	if (fmode & CEPH_FILE_MODE_WR)
4385 		ci->i_last_wr = now;
4386 	/* queue periodic check */
4387 	if (fmode &&
4388 	    __ceph_is_any_real_caps(ci) &&
4389 	    list_empty(&ci->i_cap_delay_list))
4390 		__cap_delay_requeue(mdsc, ci);
4391 }
4392 
ceph_get_fmode(struct ceph_inode_info * ci,int fmode,int count)4393 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4394 {
4395 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->netfs.inode.i_sb);
4396 	int bits = (fmode << 1) | 1;
4397 	bool already_opened = false;
4398 	int i;
4399 
4400 	if (count == 1)
4401 		atomic64_inc(&mdsc->metric.opened_files);
4402 
4403 	spin_lock(&ci->i_ceph_lock);
4404 	for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4405 		/*
4406 		 * If any of the mode ref is larger than 0,
4407 		 * that means it has been already opened by
4408 		 * others. Just skip checking the PIN ref.
4409 		 */
4410 		if (i && ci->i_nr_by_mode[i])
4411 			already_opened = true;
4412 
4413 		if (bits & (1 << i))
4414 			ci->i_nr_by_mode[i] += count;
4415 	}
4416 
4417 	if (!already_opened)
4418 		percpu_counter_inc(&mdsc->metric.opened_inodes);
4419 	spin_unlock(&ci->i_ceph_lock);
4420 }
4421 
4422 /*
4423  * Drop open file reference.  If we were the last open file,
4424  * we may need to release capabilities to the MDS (or schedule
4425  * their delayed release).
4426  */
ceph_put_fmode(struct ceph_inode_info * ci,int fmode,int count)4427 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4428 {
4429 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->netfs.inode.i_sb);
4430 	int bits = (fmode << 1) | 1;
4431 	bool is_closed = true;
4432 	int i;
4433 
4434 	if (count == 1)
4435 		atomic64_dec(&mdsc->metric.opened_files);
4436 
4437 	spin_lock(&ci->i_ceph_lock);
4438 	for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4439 		if (bits & (1 << i)) {
4440 			BUG_ON(ci->i_nr_by_mode[i] < count);
4441 			ci->i_nr_by_mode[i] -= count;
4442 		}
4443 
4444 		/*
4445 		 * If any of the mode ref is not 0 after
4446 		 * decreased, that means it is still opened
4447 		 * by others. Just skip checking the PIN ref.
4448 		 */
4449 		if (i && ci->i_nr_by_mode[i])
4450 			is_closed = false;
4451 	}
4452 
4453 	if (is_closed)
4454 		percpu_counter_dec(&mdsc->metric.opened_inodes);
4455 	spin_unlock(&ci->i_ceph_lock);
4456 }
4457 
4458 /*
4459  * For a soon-to-be unlinked file, drop the LINK caps. If it
4460  * looks like the link count will hit 0, drop any other caps (other
4461  * than PIN) we don't specifically want (due to the file still being
4462  * open).
4463  */
ceph_drop_caps_for_unlink(struct inode * inode)4464 int ceph_drop_caps_for_unlink(struct inode *inode)
4465 {
4466 	struct ceph_inode_info *ci = ceph_inode(inode);
4467 	int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4468 
4469 	spin_lock(&ci->i_ceph_lock);
4470 	if (inode->i_nlink == 1) {
4471 		drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4472 
4473 		if (__ceph_caps_dirty(ci)) {
4474 			struct ceph_mds_client *mdsc =
4475 				ceph_inode_to_client(inode)->mdsc;
4476 			__cap_delay_requeue_front(mdsc, ci);
4477 		}
4478 	}
4479 	spin_unlock(&ci->i_ceph_lock);
4480 	return drop;
4481 }
4482 
4483 /*
4484  * Helpers for embedding cap and dentry lease releases into mds
4485  * requests.
4486  *
4487  * @force is used by dentry_release (below) to force inclusion of a
4488  * record for the directory inode, even when there aren't any caps to
4489  * drop.
4490  */
ceph_encode_inode_release(void ** p,struct inode * inode,int mds,int drop,int unless,int force)4491 int ceph_encode_inode_release(void **p, struct inode *inode,
4492 			      int mds, int drop, int unless, int force)
4493 {
4494 	struct ceph_inode_info *ci = ceph_inode(inode);
4495 	struct ceph_cap *cap;
4496 	struct ceph_mds_request_release *rel = *p;
4497 	int used, dirty;
4498 	int ret = 0;
4499 
4500 	spin_lock(&ci->i_ceph_lock);
4501 	used = __ceph_caps_used(ci);
4502 	dirty = __ceph_caps_dirty(ci);
4503 
4504 	dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4505 	     inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4506 	     ceph_cap_string(unless));
4507 
4508 	/* only drop unused, clean caps */
4509 	drop &= ~(used | dirty);
4510 
4511 	cap = __get_cap_for_mds(ci, mds);
4512 	if (cap && __cap_is_valid(cap)) {
4513 		unless &= cap->issued;
4514 		if (unless) {
4515 			if (unless & CEPH_CAP_AUTH_EXCL)
4516 				drop &= ~CEPH_CAP_AUTH_SHARED;
4517 			if (unless & CEPH_CAP_LINK_EXCL)
4518 				drop &= ~CEPH_CAP_LINK_SHARED;
4519 			if (unless & CEPH_CAP_XATTR_EXCL)
4520 				drop &= ~CEPH_CAP_XATTR_SHARED;
4521 			if (unless & CEPH_CAP_FILE_EXCL)
4522 				drop &= ~CEPH_CAP_FILE_SHARED;
4523 		}
4524 
4525 		if (force || (cap->issued & drop)) {
4526 			if (cap->issued & drop) {
4527 				int wanted = __ceph_caps_wanted(ci);
4528 				dout("encode_inode_release %p cap %p "
4529 				     "%s -> %s, wanted %s -> %s\n", inode, cap,
4530 				     ceph_cap_string(cap->issued),
4531 				     ceph_cap_string(cap->issued & ~drop),
4532 				     ceph_cap_string(cap->mds_wanted),
4533 				     ceph_cap_string(wanted));
4534 
4535 				cap->issued &= ~drop;
4536 				cap->implemented &= ~drop;
4537 				cap->mds_wanted = wanted;
4538 				if (cap == ci->i_auth_cap &&
4539 				    !(wanted & CEPH_CAP_ANY_FILE_WR))
4540 					ci->i_requested_max_size = 0;
4541 			} else {
4542 				dout("encode_inode_release %p cap %p %s"
4543 				     " (force)\n", inode, cap,
4544 				     ceph_cap_string(cap->issued));
4545 			}
4546 
4547 			rel->ino = cpu_to_le64(ceph_ino(inode));
4548 			rel->cap_id = cpu_to_le64(cap->cap_id);
4549 			rel->seq = cpu_to_le32(cap->seq);
4550 			rel->issue_seq = cpu_to_le32(cap->issue_seq);
4551 			rel->mseq = cpu_to_le32(cap->mseq);
4552 			rel->caps = cpu_to_le32(cap->implemented);
4553 			rel->wanted = cpu_to_le32(cap->mds_wanted);
4554 			rel->dname_len = 0;
4555 			rel->dname_seq = 0;
4556 			*p += sizeof(*rel);
4557 			ret = 1;
4558 		} else {
4559 			dout("encode_inode_release %p cap %p %s (noop)\n",
4560 			     inode, cap, ceph_cap_string(cap->issued));
4561 		}
4562 	}
4563 	spin_unlock(&ci->i_ceph_lock);
4564 	return ret;
4565 }
4566 
ceph_encode_dentry_release(void ** p,struct dentry * dentry,struct inode * dir,int mds,int drop,int unless)4567 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4568 			       struct inode *dir,
4569 			       int mds, int drop, int unless)
4570 {
4571 	struct dentry *parent = NULL;
4572 	struct ceph_mds_request_release *rel = *p;
4573 	struct ceph_dentry_info *di = ceph_dentry(dentry);
4574 	int force = 0;
4575 	int ret;
4576 
4577 	/*
4578 	 * force an record for the directory caps if we have a dentry lease.
4579 	 * this is racy (can't take i_ceph_lock and d_lock together), but it
4580 	 * doesn't have to be perfect; the mds will revoke anything we don't
4581 	 * release.
4582 	 */
4583 	spin_lock(&dentry->d_lock);
4584 	if (di->lease_session && di->lease_session->s_mds == mds)
4585 		force = 1;
4586 	if (!dir) {
4587 		parent = dget(dentry->d_parent);
4588 		dir = d_inode(parent);
4589 	}
4590 	spin_unlock(&dentry->d_lock);
4591 
4592 	ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4593 	dput(parent);
4594 
4595 	spin_lock(&dentry->d_lock);
4596 	if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4597 		dout("encode_dentry_release %p mds%d seq %d\n",
4598 		     dentry, mds, (int)di->lease_seq);
4599 		rel->dname_len = cpu_to_le32(dentry->d_name.len);
4600 		memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4601 		*p += dentry->d_name.len;
4602 		rel->dname_seq = cpu_to_le32(di->lease_seq);
4603 		__ceph_mdsc_drop_dentry_lease(dentry);
4604 	}
4605 	spin_unlock(&dentry->d_lock);
4606 	return ret;
4607 }
4608 
remove_capsnaps(struct ceph_mds_client * mdsc,struct inode * inode)4609 static int remove_capsnaps(struct ceph_mds_client *mdsc, struct inode *inode)
4610 {
4611 	struct ceph_inode_info *ci = ceph_inode(inode);
4612 	struct ceph_cap_snap *capsnap;
4613 	int capsnap_release = 0;
4614 
4615 	lockdep_assert_held(&ci->i_ceph_lock);
4616 
4617 	dout("removing capsnaps, ci is %p, inode is %p\n", ci, inode);
4618 
4619 	while (!list_empty(&ci->i_cap_snaps)) {
4620 		capsnap = list_first_entry(&ci->i_cap_snaps,
4621 					   struct ceph_cap_snap, ci_item);
4622 		__ceph_remove_capsnap(inode, capsnap, NULL, NULL);
4623 		ceph_put_snap_context(capsnap->context);
4624 		ceph_put_cap_snap(capsnap);
4625 		capsnap_release++;
4626 	}
4627 	wake_up_all(&ci->i_cap_wq);
4628 	wake_up_all(&mdsc->cap_flushing_wq);
4629 	return capsnap_release;
4630 }
4631 
ceph_purge_inode_cap(struct inode * inode,struct ceph_cap * cap,bool * invalidate)4632 int ceph_purge_inode_cap(struct inode *inode, struct ceph_cap *cap, bool *invalidate)
4633 {
4634 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
4635 	struct ceph_mds_client *mdsc = fsc->mdsc;
4636 	struct ceph_inode_info *ci = ceph_inode(inode);
4637 	bool is_auth;
4638 	bool dirty_dropped = false;
4639 	int iputs = 0;
4640 
4641 	lockdep_assert_held(&ci->i_ceph_lock);
4642 
4643 	dout("removing cap %p, ci is %p, inode is %p\n",
4644 	     cap, ci, &ci->netfs.inode);
4645 
4646 	is_auth = (cap == ci->i_auth_cap);
4647 	__ceph_remove_cap(cap, false);
4648 	if (is_auth) {
4649 		struct ceph_cap_flush *cf;
4650 
4651 		if (ceph_inode_is_shutdown(inode)) {
4652 			if (inode->i_data.nrpages > 0)
4653 				*invalidate = true;
4654 			if (ci->i_wrbuffer_ref > 0)
4655 				mapping_set_error(&inode->i_data, -EIO);
4656 		}
4657 
4658 		spin_lock(&mdsc->cap_dirty_lock);
4659 
4660 		/* trash all of the cap flushes for this inode */
4661 		while (!list_empty(&ci->i_cap_flush_list)) {
4662 			cf = list_first_entry(&ci->i_cap_flush_list,
4663 					      struct ceph_cap_flush, i_list);
4664 			list_del_init(&cf->g_list);
4665 			list_del_init(&cf->i_list);
4666 			if (!cf->is_capsnap)
4667 				ceph_free_cap_flush(cf);
4668 		}
4669 
4670 		if (!list_empty(&ci->i_dirty_item)) {
4671 			pr_warn_ratelimited(
4672 				" dropping dirty %s state for %p %lld\n",
4673 				ceph_cap_string(ci->i_dirty_caps),
4674 				inode, ceph_ino(inode));
4675 			ci->i_dirty_caps = 0;
4676 			list_del_init(&ci->i_dirty_item);
4677 			dirty_dropped = true;
4678 		}
4679 		if (!list_empty(&ci->i_flushing_item)) {
4680 			pr_warn_ratelimited(
4681 				" dropping dirty+flushing %s state for %p %lld\n",
4682 				ceph_cap_string(ci->i_flushing_caps),
4683 				inode, ceph_ino(inode));
4684 			ci->i_flushing_caps = 0;
4685 			list_del_init(&ci->i_flushing_item);
4686 			mdsc->num_cap_flushing--;
4687 			dirty_dropped = true;
4688 		}
4689 		spin_unlock(&mdsc->cap_dirty_lock);
4690 
4691 		if (dirty_dropped) {
4692 			mapping_set_error(inode->i_mapping, -EIO);
4693 
4694 			if (ci->i_wrbuffer_ref_head == 0 &&
4695 			    ci->i_wr_ref == 0 &&
4696 			    ci->i_dirty_caps == 0 &&
4697 			    ci->i_flushing_caps == 0) {
4698 				ceph_put_snap_context(ci->i_head_snapc);
4699 				ci->i_head_snapc = NULL;
4700 			}
4701 		}
4702 
4703 		if (atomic_read(&ci->i_filelock_ref) > 0) {
4704 			/* make further file lock syscall return -EIO */
4705 			ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
4706 			pr_warn_ratelimited(" dropping file locks for %p %lld\n",
4707 					    inode, ceph_ino(inode));
4708 		}
4709 
4710 		if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
4711 			cf = ci->i_prealloc_cap_flush;
4712 			ci->i_prealloc_cap_flush = NULL;
4713 			if (!cf->is_capsnap)
4714 				ceph_free_cap_flush(cf);
4715 		}
4716 
4717 		if (!list_empty(&ci->i_cap_snaps))
4718 			iputs = remove_capsnaps(mdsc, inode);
4719 	}
4720 	if (dirty_dropped)
4721 		++iputs;
4722 	return iputs;
4723 }
4724