1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/kernel.h>
5 #include <linux/sched.h>
6 #include <linux/slab.h>
7 #include <linux/vmalloc.h>
8 #include <linux/wait.h>
9 #include <linux/writeback.h>
12 #include "mds_client.h"
14 #include <linux/ceph/decode.h>
15 #include <linux/ceph/messenger.h>
18 * Capability management
20 * The Ceph metadata servers control client access to inode metadata
21 * and file data by issuing capabilities, granting clients permission
22 * to read and/or write both inode field and file data to OSDs
23 * (storage nodes). Each capability consists of a set of bits
24 * indicating which operations are allowed.
26 * If the client holds a *_SHARED cap, the client has a coherent value
27 * that can be safely read from the cached inode.
29 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
30 * client is allowed to change inode attributes (e.g., file size,
31 * mtime), note its dirty state in the ceph_cap, and asynchronously
32 * flush that metadata change to the MDS.
34 * In the event of a conflicting operation (perhaps by another
35 * client), the MDS will revoke the conflicting client capabilities.
37 * In order for a client to cache an inode, it must hold a capability
38 * with at least one MDS server. When inodes are released, release
39 * notifications are batched and periodically sent en masse to the MDS
40 * cluster to release server state.
45 * Generate readable cap strings for debugging output.
47 #define MAX_CAP_STR 20
48 static char cap_str[MAX_CAP_STR][40];
49 static DEFINE_SPINLOCK(cap_str_lock);
50 static int last_cap_str;
52 static char *gcap_string(char *s, int c)
54 if (c & CEPH_CAP_GSHARED)
56 if (c & CEPH_CAP_GEXCL)
58 if (c & CEPH_CAP_GCACHE)
64 if (c & CEPH_CAP_GBUFFER)
66 if (c & CEPH_CAP_GLAZYIO)
71 const char *ceph_cap_string(int caps)
77 spin_lock(&cap_str_lock);
79 if (last_cap_str == MAX_CAP_STR)
81 spin_unlock(&cap_str_lock);
85 if (caps & CEPH_CAP_PIN)
88 c = (caps >> CEPH_CAP_SAUTH) & 3;
91 s = gcap_string(s, c);
94 c = (caps >> CEPH_CAP_SLINK) & 3;
97 s = gcap_string(s, c);
100 c = (caps >> CEPH_CAP_SXATTR) & 3;
103 s = gcap_string(s, c);
106 c = caps >> CEPH_CAP_SFILE;
109 s = gcap_string(s, c);
118 void ceph_caps_init(struct ceph_mds_client *mdsc)
120 INIT_LIST_HEAD(&mdsc->caps_list);
121 spin_lock_init(&mdsc->caps_list_lock);
124 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
126 struct ceph_cap *cap;
128 spin_lock(&mdsc->caps_list_lock);
129 while (!list_empty(&mdsc->caps_list)) {
130 cap = list_first_entry(&mdsc->caps_list,
131 struct ceph_cap, caps_item);
132 list_del(&cap->caps_item);
133 kmem_cache_free(ceph_cap_cachep, cap);
135 mdsc->caps_total_count = 0;
136 mdsc->caps_avail_count = 0;
137 mdsc->caps_use_count = 0;
138 mdsc->caps_reserve_count = 0;
139 mdsc->caps_min_count = 0;
140 spin_unlock(&mdsc->caps_list_lock);
143 void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
145 spin_lock(&mdsc->caps_list_lock);
146 mdsc->caps_min_count += delta;
147 BUG_ON(mdsc->caps_min_count < 0);
148 spin_unlock(&mdsc->caps_list_lock);
151 void ceph_reserve_caps(struct ceph_mds_client *mdsc,
152 struct ceph_cap_reservation *ctx, int need)
155 struct ceph_cap *cap;
160 dout("reserve caps ctx=%p need=%d\n", ctx, need);
162 /* first reserve any caps that are already allocated */
163 spin_lock(&mdsc->caps_list_lock);
164 if (mdsc->caps_avail_count >= need)
167 have = mdsc->caps_avail_count;
168 mdsc->caps_avail_count -= have;
169 mdsc->caps_reserve_count += have;
170 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
171 mdsc->caps_reserve_count +
172 mdsc->caps_avail_count);
173 spin_unlock(&mdsc->caps_list_lock);
175 for (i = have; i < need; i++) {
176 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
179 list_add(&cap->caps_item, &newcaps);
182 /* we didn't manage to reserve as much as we needed */
183 if (have + alloc != need)
184 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
185 ctx, need, have + alloc);
187 spin_lock(&mdsc->caps_list_lock);
188 mdsc->caps_total_count += alloc;
189 mdsc->caps_reserve_count += alloc;
190 list_splice(&newcaps, &mdsc->caps_list);
192 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
193 mdsc->caps_reserve_count +
194 mdsc->caps_avail_count);
195 spin_unlock(&mdsc->caps_list_lock);
198 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
199 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
200 mdsc->caps_reserve_count, mdsc->caps_avail_count);
203 int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
204 struct ceph_cap_reservation *ctx)
206 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
208 spin_lock(&mdsc->caps_list_lock);
209 BUG_ON(mdsc->caps_reserve_count < ctx->count);
210 mdsc->caps_reserve_count -= ctx->count;
211 mdsc->caps_avail_count += ctx->count;
213 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
214 mdsc->caps_total_count, mdsc->caps_use_count,
215 mdsc->caps_reserve_count, mdsc->caps_avail_count);
216 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
217 mdsc->caps_reserve_count +
218 mdsc->caps_avail_count);
219 spin_unlock(&mdsc->caps_list_lock);
224 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
225 struct ceph_cap_reservation *ctx)
227 struct ceph_cap *cap = NULL;
229 /* temporary, until we do something about cap import/export */
231 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
233 spin_lock(&mdsc->caps_list_lock);
234 mdsc->caps_use_count++;
235 mdsc->caps_total_count++;
236 spin_unlock(&mdsc->caps_list_lock);
241 spin_lock(&mdsc->caps_list_lock);
242 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
243 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
244 mdsc->caps_reserve_count, mdsc->caps_avail_count);
246 BUG_ON(ctx->count > mdsc->caps_reserve_count);
247 BUG_ON(list_empty(&mdsc->caps_list));
250 mdsc->caps_reserve_count--;
251 mdsc->caps_use_count++;
253 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
254 list_del(&cap->caps_item);
256 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
257 mdsc->caps_reserve_count + mdsc->caps_avail_count);
258 spin_unlock(&mdsc->caps_list_lock);
262 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
264 spin_lock(&mdsc->caps_list_lock);
265 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
266 cap, mdsc->caps_total_count, mdsc->caps_use_count,
267 mdsc->caps_reserve_count, mdsc->caps_avail_count);
268 mdsc->caps_use_count--;
270 * Keep some preallocated caps around (ceph_min_count), to
271 * avoid lots of free/alloc churn.
273 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
274 mdsc->caps_min_count) {
275 mdsc->caps_total_count--;
276 kmem_cache_free(ceph_cap_cachep, cap);
278 mdsc->caps_avail_count++;
279 list_add(&cap->caps_item, &mdsc->caps_list);
282 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
283 mdsc->caps_reserve_count + mdsc->caps_avail_count);
284 spin_unlock(&mdsc->caps_list_lock);
287 void ceph_reservation_status(struct ceph_fs_client *fsc,
288 int *total, int *avail, int *used, int *reserved,
291 struct ceph_mds_client *mdsc = fsc->mdsc;
294 *total = mdsc->caps_total_count;
296 *avail = mdsc->caps_avail_count;
298 *used = mdsc->caps_use_count;
300 *reserved = mdsc->caps_reserve_count;
302 *min = mdsc->caps_min_count;
306 * Find ceph_cap for given mds, if any.
308 * Called with i_ceph_lock held.
310 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
312 struct ceph_cap *cap;
313 struct rb_node *n = ci->i_caps.rb_node;
316 cap = rb_entry(n, struct ceph_cap, ci_node);
319 else if (mds > cap->mds)
327 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
329 struct ceph_cap *cap;
331 spin_lock(&ci->i_ceph_lock);
332 cap = __get_cap_for_mds(ci, mds);
333 spin_unlock(&ci->i_ceph_lock);
338 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
340 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
342 struct ceph_cap *cap;
346 /* prefer mds with WR|BUFFER|EXCL caps */
347 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
348 cap = rb_entry(p, struct ceph_cap, ci_node);
350 if (cap->issued & (CEPH_CAP_FILE_WR |
351 CEPH_CAP_FILE_BUFFER |
358 int ceph_get_cap_mds(struct inode *inode)
360 struct ceph_inode_info *ci = ceph_inode(inode);
362 spin_lock(&ci->i_ceph_lock);
363 mds = __ceph_get_cap_mds(ceph_inode(inode));
364 spin_unlock(&ci->i_ceph_lock);
369 * Called under i_ceph_lock.
371 static void __insert_cap_node(struct ceph_inode_info *ci,
372 struct ceph_cap *new)
374 struct rb_node **p = &ci->i_caps.rb_node;
375 struct rb_node *parent = NULL;
376 struct ceph_cap *cap = NULL;
380 cap = rb_entry(parent, struct ceph_cap, ci_node);
381 if (new->mds < cap->mds)
383 else if (new->mds > cap->mds)
389 rb_link_node(&new->ci_node, parent, p);
390 rb_insert_color(&new->ci_node, &ci->i_caps);
394 * (re)set cap hold timeouts, which control the delayed release
395 * of unused caps back to the MDS. Should be called on cap use.
397 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
398 struct ceph_inode_info *ci)
400 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
402 ci->i_hold_caps_min = round_jiffies(jiffies +
403 ma->caps_wanted_delay_min * HZ);
404 ci->i_hold_caps_max = round_jiffies(jiffies +
405 ma->caps_wanted_delay_max * HZ);
406 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
407 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
411 * (Re)queue cap at the end of the delayed cap release list.
413 * If I_FLUSH is set, leave the inode at the front of the list.
415 * Caller holds i_ceph_lock
416 * -> we take mdsc->cap_delay_lock
418 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
419 struct ceph_inode_info *ci)
421 __cap_set_timeouts(mdsc, ci);
422 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
423 ci->i_ceph_flags, ci->i_hold_caps_max);
424 if (!mdsc->stopping) {
425 spin_lock(&mdsc->cap_delay_lock);
426 if (!list_empty(&ci->i_cap_delay_list)) {
427 if (ci->i_ceph_flags & CEPH_I_FLUSH)
429 list_del_init(&ci->i_cap_delay_list);
431 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
433 spin_unlock(&mdsc->cap_delay_lock);
438 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
439 * indicating we should send a cap message to flush dirty metadata
440 * asap, and move to the front of the delayed cap list.
442 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
443 struct ceph_inode_info *ci)
445 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
446 spin_lock(&mdsc->cap_delay_lock);
447 ci->i_ceph_flags |= CEPH_I_FLUSH;
448 if (!list_empty(&ci->i_cap_delay_list))
449 list_del_init(&ci->i_cap_delay_list);
450 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
451 spin_unlock(&mdsc->cap_delay_lock);
455 * Cancel delayed work on cap.
457 * Caller must hold i_ceph_lock.
459 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
460 struct ceph_inode_info *ci)
462 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
463 if (list_empty(&ci->i_cap_delay_list))
465 spin_lock(&mdsc->cap_delay_lock);
466 list_del_init(&ci->i_cap_delay_list);
467 spin_unlock(&mdsc->cap_delay_lock);
471 * Common issue checks for add_cap, handle_cap_grant.
473 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
476 unsigned had = __ceph_caps_issued(ci, NULL);
479 * Each time we receive FILE_CACHE anew, we increment
482 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
483 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
488 * if we are newly issued FILE_SHARED, mark dir not complete; we
489 * don't know what happened to this directory while we didn't
492 if ((issued & CEPH_CAP_FILE_SHARED) &&
493 (had & CEPH_CAP_FILE_SHARED) == 0) {
495 if (S_ISDIR(ci->vfs_inode.i_mode)) {
496 dout(" marking %p NOT complete\n", &ci->vfs_inode);
497 __ceph_dir_clear_complete(ci);
503 * Add a capability under the given MDS session.
505 * Caller should hold session snap_rwsem (read) and s_mutex.
507 * @fmode is the open file mode, if we are opening a file, otherwise
508 * it is < 0. (This is so we can atomically add the cap and add an
509 * open file reference to it.)
511 void ceph_add_cap(struct inode *inode,
512 struct ceph_mds_session *session, u64 cap_id,
513 int fmode, unsigned issued, unsigned wanted,
514 unsigned seq, unsigned mseq, u64 realmino, int flags,
515 struct ceph_cap **new_cap)
517 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
518 struct ceph_inode_info *ci = ceph_inode(inode);
519 struct ceph_cap *cap;
520 int mds = session->s_mds;
523 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
524 session->s_mds, cap_id, ceph_cap_string(issued), seq);
527 * If we are opening the file, include file mode wanted bits
531 wanted |= ceph_caps_for_mode(fmode);
533 cap = __get_cap_for_mds(ci, mds);
539 cap->implemented = 0;
545 __insert_cap_node(ci, cap);
547 /* add to session cap list */
548 cap->session = session;
549 spin_lock(&session->s_cap_lock);
550 list_add_tail(&cap->session_caps, &session->s_caps);
551 session->s_nr_caps++;
552 spin_unlock(&session->s_cap_lock);
555 * auth mds of the inode changed. we received the cap export
556 * message, but still haven't received the cap import message.
557 * handle_cap_export() updated the new auth MDS' cap.
559 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
560 * a message that was send before the cap import message. So
563 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
564 WARN_ON(cap != ci->i_auth_cap);
565 WARN_ON(cap->cap_id != cap_id);
568 issued |= cap->issued;
569 flags |= CEPH_CAP_FLAG_AUTH;
573 if (!ci->i_snap_realm) {
575 * add this inode to the appropriate snap realm
577 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
580 ceph_get_snap_realm(mdsc, realm);
581 spin_lock(&realm->inodes_with_caps_lock);
582 ci->i_snap_realm = realm;
583 list_add(&ci->i_snap_realm_item,
584 &realm->inodes_with_caps);
585 spin_unlock(&realm->inodes_with_caps_lock);
587 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
593 __check_cap_issue(ci, cap, issued);
596 * If we are issued caps we don't want, or the mds' wanted
597 * value appears to be off, queue a check so we'll release
598 * later and/or update the mds wanted value.
600 actual_wanted = __ceph_caps_wanted(ci);
601 if ((wanted & ~actual_wanted) ||
602 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
603 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
604 ceph_cap_string(issued), ceph_cap_string(wanted),
605 ceph_cap_string(actual_wanted));
606 __cap_delay_requeue(mdsc, ci);
609 if (flags & CEPH_CAP_FLAG_AUTH) {
610 if (ci->i_auth_cap == NULL ||
611 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
612 ci->i_auth_cap = cap;
613 cap->mds_wanted = wanted;
616 WARN_ON(ci->i_auth_cap == cap);
619 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
620 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
621 ceph_cap_string(issued|cap->issued), seq, mds);
622 cap->cap_id = cap_id;
623 cap->issued = issued;
624 cap->implemented |= issued;
625 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
626 cap->mds_wanted = wanted;
628 cap->mds_wanted |= wanted;
630 cap->issue_seq = seq;
632 cap->cap_gen = session->s_cap_gen;
635 __ceph_get_fmode(ci, fmode);
639 * Return true if cap has not timed out and belongs to the current
640 * generation of the MDS session (i.e. has not gone 'stale' due to
641 * us losing touch with the mds).
643 static int __cap_is_valid(struct ceph_cap *cap)
648 spin_lock(&cap->session->s_gen_ttl_lock);
649 gen = cap->session->s_cap_gen;
650 ttl = cap->session->s_cap_ttl;
651 spin_unlock(&cap->session->s_gen_ttl_lock);
653 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
654 dout("__cap_is_valid %p cap %p issued %s "
655 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
656 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
664 * Return set of valid cap bits issued to us. Note that caps time
665 * out, and may be invalidated in bulk if the client session times out
666 * and session->s_cap_gen is bumped.
668 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
670 int have = ci->i_snap_caps;
671 struct ceph_cap *cap;
676 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
677 cap = rb_entry(p, struct ceph_cap, ci_node);
678 if (!__cap_is_valid(cap))
680 dout("__ceph_caps_issued %p cap %p issued %s\n",
681 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
684 *implemented |= cap->implemented;
687 * exclude caps issued by non-auth MDS, but are been revoking
688 * by the auth MDS. The non-auth MDS should be revoking/exporting
689 * these caps, but the message is delayed.
691 if (ci->i_auth_cap) {
692 cap = ci->i_auth_cap;
693 have &= ~cap->implemented | cap->issued;
699 * Get cap bits issued by caps other than @ocap
701 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
703 int have = ci->i_snap_caps;
704 struct ceph_cap *cap;
707 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
708 cap = rb_entry(p, struct ceph_cap, ci_node);
711 if (!__cap_is_valid(cap))
719 * Move a cap to the end of the LRU (oldest caps at list head, newest
722 static void __touch_cap(struct ceph_cap *cap)
724 struct ceph_mds_session *s = cap->session;
726 spin_lock(&s->s_cap_lock);
727 if (s->s_cap_iterator == NULL) {
728 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
730 list_move_tail(&cap->session_caps, &s->s_caps);
732 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
733 &cap->ci->vfs_inode, cap, s->s_mds);
735 spin_unlock(&s->s_cap_lock);
739 * Check if we hold the given mask. If so, move the cap(s) to the
740 * front of their respective LRUs. (This is the preferred way for
741 * callers to check for caps they want.)
743 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
745 struct ceph_cap *cap;
747 int have = ci->i_snap_caps;
749 if ((have & mask) == mask) {
750 dout("__ceph_caps_issued_mask %p snap issued %s"
751 " (mask %s)\n", &ci->vfs_inode,
752 ceph_cap_string(have),
753 ceph_cap_string(mask));
757 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
758 cap = rb_entry(p, struct ceph_cap, ci_node);
759 if (!__cap_is_valid(cap))
761 if ((cap->issued & mask) == mask) {
762 dout("__ceph_caps_issued_mask %p cap %p issued %s"
763 " (mask %s)\n", &ci->vfs_inode, cap,
764 ceph_cap_string(cap->issued),
765 ceph_cap_string(mask));
771 /* does a combination of caps satisfy mask? */
773 if ((have & mask) == mask) {
774 dout("__ceph_caps_issued_mask %p combo issued %s"
775 " (mask %s)\n", &ci->vfs_inode,
776 ceph_cap_string(cap->issued),
777 ceph_cap_string(mask));
781 /* touch this + preceding caps */
783 for (q = rb_first(&ci->i_caps); q != p;
785 cap = rb_entry(q, struct ceph_cap,
787 if (!__cap_is_valid(cap))
800 * Return true if mask caps are currently being revoked by an MDS.
802 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
803 struct ceph_cap *ocap, int mask)
805 struct ceph_cap *cap;
808 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
809 cap = rb_entry(p, struct ceph_cap, ci_node);
811 (cap->implemented & ~cap->issued & mask))
817 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
819 struct inode *inode = &ci->vfs_inode;
822 spin_lock(&ci->i_ceph_lock);
823 ret = __ceph_caps_revoking_other(ci, NULL, mask);
824 spin_unlock(&ci->i_ceph_lock);
825 dout("ceph_caps_revoking %p %s = %d\n", inode,
826 ceph_cap_string(mask), ret);
830 int __ceph_caps_used(struct ceph_inode_info *ci)
834 used |= CEPH_CAP_PIN;
836 used |= CEPH_CAP_FILE_RD;
837 if (ci->i_rdcache_ref || ci->vfs_inode.i_data.nrpages)
838 used |= CEPH_CAP_FILE_CACHE;
840 used |= CEPH_CAP_FILE_WR;
841 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
842 used |= CEPH_CAP_FILE_BUFFER;
847 * wanted, by virtue of open file modes
849 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
853 for (mode = 0; mode < CEPH_FILE_MODE_NUM; mode++)
854 if (ci->i_nr_by_mode[mode])
855 want |= ceph_caps_for_mode(mode);
860 * Return caps we have registered with the MDS(s) as 'wanted'.
862 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
864 struct ceph_cap *cap;
868 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
869 cap = rb_entry(p, struct ceph_cap, ci_node);
870 if (!__cap_is_valid(cap))
872 if (cap == ci->i_auth_cap)
873 mds_wanted |= cap->mds_wanted;
875 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
881 * called under i_ceph_lock
883 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
885 return !RB_EMPTY_ROOT(&ci->i_caps);
888 int ceph_is_any_caps(struct inode *inode)
890 struct ceph_inode_info *ci = ceph_inode(inode);
893 spin_lock(&ci->i_ceph_lock);
894 ret = __ceph_is_any_caps(ci);
895 spin_unlock(&ci->i_ceph_lock);
901 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
903 * caller should hold i_ceph_lock.
904 * caller will not hold session s_mutex if called from destroy_inode.
906 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
908 struct ceph_mds_session *session = cap->session;
909 struct ceph_inode_info *ci = cap->ci;
910 struct ceph_mds_client *mdsc =
911 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
914 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
916 /* remove from session list */
917 spin_lock(&session->s_cap_lock);
919 * s_cap_reconnect is protected by s_cap_lock. no one changes
920 * s_cap_gen while session is in the reconnect state.
923 (!session->s_cap_reconnect ||
924 cap->cap_gen == session->s_cap_gen))
925 __queue_cap_release(session, ci->i_vino.ino, cap->cap_id,
926 cap->mseq, cap->issue_seq);
928 if (session->s_cap_iterator == cap) {
929 /* not yet, we are iterating over this very cap */
930 dout("__ceph_remove_cap delaying %p removal from session %p\n",
933 list_del_init(&cap->session_caps);
934 session->s_nr_caps--;
938 /* protect backpointer with s_cap_lock: see iterate_session_caps */
940 spin_unlock(&session->s_cap_lock);
942 /* remove from inode list */
943 rb_erase(&cap->ci_node, &ci->i_caps);
944 if (ci->i_auth_cap == cap)
945 ci->i_auth_cap = NULL;
948 ceph_put_cap(mdsc, cap);
950 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
951 struct ceph_snap_realm *realm = ci->i_snap_realm;
952 spin_lock(&realm->inodes_with_caps_lock);
953 list_del_init(&ci->i_snap_realm_item);
954 ci->i_snap_realm_counter++;
955 ci->i_snap_realm = NULL;
956 spin_unlock(&realm->inodes_with_caps_lock);
957 ceph_put_snap_realm(mdsc, realm);
959 if (!__ceph_is_any_real_caps(ci))
960 __cap_delay_cancel(mdsc, ci);
964 * Build and send a cap message to the given MDS.
966 * Caller should be holding s_mutex.
968 static int send_cap_msg(struct ceph_mds_session *session,
969 u64 ino, u64 cid, int op,
970 int caps, int wanted, int dirty,
971 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
972 u64 size, u64 max_size,
973 struct timespec *mtime, struct timespec *atime,
975 kuid_t uid, kgid_t gid, umode_t mode,
977 struct ceph_buffer *xattrs_buf,
978 u64 follows, bool inline_data)
980 struct ceph_mds_caps *fc;
981 struct ceph_msg *msg;
985 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
986 " seq %u/%u mseq %u follows %lld size %llu/%llu"
987 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
988 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
989 ceph_cap_string(dirty),
990 seq, issue_seq, mseq, follows, size, max_size,
991 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
993 /* flock buffer size + inline version + inline data size */
994 extra_len = 4 + 8 + 4;
995 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1000 msg->hdr.tid = cpu_to_le64(flush_tid);
1002 fc = msg->front.iov_base;
1003 memset(fc, 0, sizeof(*fc));
1005 fc->cap_id = cpu_to_le64(cid);
1006 fc->op = cpu_to_le32(op);
1007 fc->seq = cpu_to_le32(seq);
1008 fc->issue_seq = cpu_to_le32(issue_seq);
1009 fc->migrate_seq = cpu_to_le32(mseq);
1010 fc->caps = cpu_to_le32(caps);
1011 fc->wanted = cpu_to_le32(wanted);
1012 fc->dirty = cpu_to_le32(dirty);
1013 fc->ino = cpu_to_le64(ino);
1014 fc->snap_follows = cpu_to_le64(follows);
1016 fc->size = cpu_to_le64(size);
1017 fc->max_size = cpu_to_le64(max_size);
1019 ceph_encode_timespec(&fc->mtime, mtime);
1021 ceph_encode_timespec(&fc->atime, atime);
1022 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
1024 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid));
1025 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid));
1026 fc->mode = cpu_to_le32(mode);
1029 /* flock buffer size */
1030 ceph_encode_32(&p, 0);
1031 /* inline version */
1032 ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE);
1033 /* inline data size */
1034 ceph_encode_32(&p, 0);
1036 fc->xattr_version = cpu_to_le64(xattr_version);
1038 msg->middle = ceph_buffer_get(xattrs_buf);
1039 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1040 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1043 ceph_con_send(&session->s_con, msg);
1047 void __queue_cap_release(struct ceph_mds_session *session,
1048 u64 ino, u64 cap_id, u32 migrate_seq,
1051 struct ceph_msg *msg;
1052 struct ceph_mds_cap_release *head;
1053 struct ceph_mds_cap_item *item;
1055 BUG_ON(!session->s_num_cap_releases);
1056 msg = list_first_entry(&session->s_cap_releases,
1057 struct ceph_msg, list_head);
1059 dout(" adding %llx release to mds%d msg %p (%d left)\n",
1060 ino, session->s_mds, msg, session->s_num_cap_releases);
1062 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1063 head = msg->front.iov_base;
1064 le32_add_cpu(&head->num, 1);
1065 item = msg->front.iov_base + msg->front.iov_len;
1066 item->ino = cpu_to_le64(ino);
1067 item->cap_id = cpu_to_le64(cap_id);
1068 item->migrate_seq = cpu_to_le32(migrate_seq);
1069 item->seq = cpu_to_le32(issue_seq);
1071 session->s_num_cap_releases--;
1073 msg->front.iov_len += sizeof(*item);
1074 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1075 dout(" release msg %p full\n", msg);
1076 list_move_tail(&msg->list_head, &session->s_cap_releases_done);
1078 dout(" release msg %p at %d/%d (%d)\n", msg,
1079 (int)le32_to_cpu(head->num),
1080 (int)CEPH_CAPS_PER_RELEASE,
1081 (int)msg->front.iov_len);
1086 * Queue cap releases when an inode is dropped from our cache. Since
1087 * inode is about to be destroyed, there is no need for i_ceph_lock.
1089 void ceph_queue_caps_release(struct inode *inode)
1091 struct ceph_inode_info *ci = ceph_inode(inode);
1094 p = rb_first(&ci->i_caps);
1096 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1098 __ceph_remove_cap(cap, true);
1103 * Send a cap msg on the given inode. Update our caps state, then
1104 * drop i_ceph_lock and send the message.
1106 * Make note of max_size reported/requested from mds, revoked caps
1107 * that have now been implemented.
1109 * Make half-hearted attempt ot to invalidate page cache if we are
1110 * dropping RDCACHE. Note that this will leave behind locked pages
1111 * that we'll then need to deal with elsewhere.
1113 * Return non-zero if delayed release, or we experienced an error
1114 * such that the caller should requeue + retry later.
1116 * called with i_ceph_lock, then drops it.
1117 * caller should hold snap_rwsem (read), s_mutex.
1119 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1120 int op, int used, int want, int retain, int flushing,
1121 unsigned *pflush_tid)
1122 __releases(cap->ci->i_ceph_lock)
1124 struct ceph_inode_info *ci = cap->ci;
1125 struct inode *inode = &ci->vfs_inode;
1126 u64 cap_id = cap->cap_id;
1127 int held, revoking, dropping, keep;
1128 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1130 struct timespec mtime, atime;
1135 struct ceph_mds_session *session;
1136 u64 xattr_version = 0;
1137 struct ceph_buffer *xattr_blob = NULL;
1144 held = cap->issued | cap->implemented;
1145 revoking = cap->implemented & ~cap->issued;
1146 retain &= ~revoking;
1147 dropping = cap->issued & ~retain;
1149 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1150 inode, cap, cap->session,
1151 ceph_cap_string(held), ceph_cap_string(held & retain),
1152 ceph_cap_string(revoking));
1153 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1155 session = cap->session;
1157 /* don't release wanted unless we've waited a bit. */
1158 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1159 time_before(jiffies, ci->i_hold_caps_min)) {
1160 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1161 ceph_cap_string(cap->issued),
1162 ceph_cap_string(cap->issued & retain),
1163 ceph_cap_string(cap->mds_wanted),
1164 ceph_cap_string(want));
1165 want |= cap->mds_wanted;
1166 retain |= cap->issued;
1169 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1171 cap->issued &= retain; /* drop bits we don't want */
1172 if (cap->implemented & ~cap->issued) {
1174 * Wake up any waiters on wanted -> needed transition.
1175 * This is due to the weird transition from buffered
1176 * to sync IO... we need to flush dirty pages _before_
1177 * allowing sync writes to avoid reordering.
1181 cap->implemented &= cap->issued | used;
1182 cap->mds_wanted = want;
1186 * assign a tid for flush operations so we can avoid
1187 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1188 * clean type races. track latest tid for every bit
1189 * so we can handle flush AxFw, flush Fw, and have the
1190 * first ack clean Ax.
1192 flush_tid = ++ci->i_cap_flush_last_tid;
1194 *pflush_tid = flush_tid;
1195 dout(" cap_flush_tid %d\n", (int)flush_tid);
1196 for (i = 0; i < CEPH_CAP_BITS; i++)
1197 if (flushing & (1 << i))
1198 ci->i_cap_flush_tid[i] = flush_tid;
1200 follows = ci->i_head_snapc->seq;
1205 keep = cap->implemented;
1207 issue_seq = cap->issue_seq;
1209 size = inode->i_size;
1210 ci->i_reported_size = size;
1211 max_size = ci->i_wanted_max_size;
1212 ci->i_requested_max_size = max_size;
1213 mtime = inode->i_mtime;
1214 atime = inode->i_atime;
1215 time_warp_seq = ci->i_time_warp_seq;
1218 mode = inode->i_mode;
1220 if (flushing & CEPH_CAP_XATTR_EXCL) {
1221 __ceph_build_xattrs_blob(ci);
1222 xattr_blob = ci->i_xattrs.blob;
1223 xattr_version = ci->i_xattrs.version;
1226 inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1228 spin_unlock(&ci->i_ceph_lock);
1230 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1231 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1232 size, max_size, &mtime, &atime, time_warp_seq,
1233 uid, gid, mode, xattr_version, xattr_blob,
1234 follows, inline_data);
1236 dout("error sending cap msg, must requeue %p\n", inode);
1241 wake_up_all(&ci->i_cap_wq);
1247 * When a snapshot is taken, clients accumulate dirty metadata on
1248 * inodes with capabilities in ceph_cap_snaps to describe the file
1249 * state at the time the snapshot was taken. This must be flushed
1250 * asynchronously back to the MDS once sync writes complete and dirty
1251 * data is written out.
1253 * Unless @again is true, skip cap_snaps that were already sent to
1254 * the MDS (i.e., during this session).
1256 * Called under i_ceph_lock. Takes s_mutex as needed.
1258 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1259 struct ceph_mds_session **psession,
1261 __releases(ci->i_ceph_lock)
1262 __acquires(ci->i_ceph_lock)
1264 struct inode *inode = &ci->vfs_inode;
1266 struct ceph_cap_snap *capsnap;
1268 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1269 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1271 u64 next_follows = 0; /* keep track of how far we've gotten through the
1272 i_cap_snaps list, and skip these entries next time
1273 around to avoid an infinite loop */
1276 session = *psession;
1278 dout("__flush_snaps %p\n", inode);
1280 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1281 /* avoid an infiniute loop after retry */
1282 if (capsnap->follows < next_follows)
1285 * we need to wait for sync writes to complete and for dirty
1286 * pages to be written out.
1288 if (capsnap->dirty_pages || capsnap->writing)
1292 * if cap writeback already occurred, we should have dropped
1293 * the capsnap in ceph_put_wrbuffer_cap_refs.
1295 BUG_ON(capsnap->dirty == 0);
1297 /* pick mds, take s_mutex */
1298 if (ci->i_auth_cap == NULL) {
1299 dout("no auth cap (migrating?), doing nothing\n");
1303 /* only flush each capsnap once */
1304 if (!again && !list_empty(&capsnap->flushing_item)) {
1305 dout("already flushed %p, skipping\n", capsnap);
1309 mds = ci->i_auth_cap->session->s_mds;
1310 mseq = ci->i_auth_cap->mseq;
1312 if (session && session->s_mds != mds) {
1313 dout("oops, wrong session %p mutex\n", session);
1314 mutex_unlock(&session->s_mutex);
1315 ceph_put_mds_session(session);
1319 spin_unlock(&ci->i_ceph_lock);
1320 mutex_lock(&mdsc->mutex);
1321 session = __ceph_lookup_mds_session(mdsc, mds);
1322 mutex_unlock(&mdsc->mutex);
1324 dout("inverting session/ino locks on %p\n",
1326 mutex_lock(&session->s_mutex);
1329 * if session == NULL, we raced against a cap
1330 * deletion or migration. retry, and we'll
1331 * get a better @mds value next time.
1333 spin_lock(&ci->i_ceph_lock);
1337 capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1338 atomic_inc(&capsnap->nref);
1339 if (!list_empty(&capsnap->flushing_item))
1340 list_del_init(&capsnap->flushing_item);
1341 list_add_tail(&capsnap->flushing_item,
1342 &session->s_cap_snaps_flushing);
1343 spin_unlock(&ci->i_ceph_lock);
1345 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1346 inode, capsnap, capsnap->follows, capsnap->flush_tid);
1347 send_cap_msg(session, ceph_vino(inode).ino, 0,
1348 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1349 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1351 &capsnap->mtime, &capsnap->atime,
1352 capsnap->time_warp_seq,
1353 capsnap->uid, capsnap->gid, capsnap->mode,
1354 capsnap->xattr_version, capsnap->xattr_blob,
1355 capsnap->follows, capsnap->inline_data);
1357 next_follows = capsnap->follows + 1;
1358 ceph_put_cap_snap(capsnap);
1360 spin_lock(&ci->i_ceph_lock);
1364 /* we flushed them all; remove this inode from the queue */
1365 spin_lock(&mdsc->snap_flush_lock);
1366 list_del_init(&ci->i_snap_flush_item);
1367 spin_unlock(&mdsc->snap_flush_lock);
1371 *psession = session;
1373 mutex_unlock(&session->s_mutex);
1374 ceph_put_mds_session(session);
1378 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1380 spin_lock(&ci->i_ceph_lock);
1381 __ceph_flush_snaps(ci, NULL, 0);
1382 spin_unlock(&ci->i_ceph_lock);
1386 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1387 * Caller is then responsible for calling __mark_inode_dirty with the
1388 * returned flags value.
1390 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1392 struct ceph_mds_client *mdsc =
1393 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1394 struct inode *inode = &ci->vfs_inode;
1395 int was = ci->i_dirty_caps;
1398 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1399 ceph_cap_string(mask), ceph_cap_string(was),
1400 ceph_cap_string(was | mask));
1401 ci->i_dirty_caps |= mask;
1403 if (!ci->i_head_snapc)
1404 ci->i_head_snapc = ceph_get_snap_context(
1405 ci->i_snap_realm->cached_context);
1406 dout(" inode %p now dirty snapc %p auth cap %p\n",
1407 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1408 WARN_ON(!ci->i_auth_cap);
1409 BUG_ON(!list_empty(&ci->i_dirty_item));
1410 spin_lock(&mdsc->cap_dirty_lock);
1411 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1412 spin_unlock(&mdsc->cap_dirty_lock);
1413 if (ci->i_flushing_caps == 0) {
1415 dirty |= I_DIRTY_SYNC;
1418 BUG_ON(list_empty(&ci->i_dirty_item));
1419 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1420 (mask & CEPH_CAP_FILE_BUFFER))
1421 dirty |= I_DIRTY_DATASYNC;
1422 __cap_delay_requeue(mdsc, ci);
1427 * Add dirty inode to the flushing list. Assigned a seq number so we
1428 * can wait for caps to flush without starving.
1430 * Called under i_ceph_lock.
1432 static int __mark_caps_flushing(struct inode *inode,
1433 struct ceph_mds_session *session)
1435 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1436 struct ceph_inode_info *ci = ceph_inode(inode);
1439 BUG_ON(ci->i_dirty_caps == 0);
1440 BUG_ON(list_empty(&ci->i_dirty_item));
1442 flushing = ci->i_dirty_caps;
1443 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1444 ceph_cap_string(flushing),
1445 ceph_cap_string(ci->i_flushing_caps),
1446 ceph_cap_string(ci->i_flushing_caps | flushing));
1447 ci->i_flushing_caps |= flushing;
1448 ci->i_dirty_caps = 0;
1449 dout(" inode %p now !dirty\n", inode);
1451 spin_lock(&mdsc->cap_dirty_lock);
1452 list_del_init(&ci->i_dirty_item);
1454 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1455 if (list_empty(&ci->i_flushing_item)) {
1456 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1457 mdsc->num_cap_flushing++;
1458 dout(" inode %p now flushing seq %lld\n", inode,
1459 ci->i_cap_flush_seq);
1461 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1462 dout(" inode %p now flushing (more) seq %lld\n", inode,
1463 ci->i_cap_flush_seq);
1465 spin_unlock(&mdsc->cap_dirty_lock);
1471 * try to invalidate mapping pages without blocking.
1473 static int try_nonblocking_invalidate(struct inode *inode)
1475 struct ceph_inode_info *ci = ceph_inode(inode);
1476 u32 invalidating_gen = ci->i_rdcache_gen;
1478 spin_unlock(&ci->i_ceph_lock);
1479 invalidate_mapping_pages(&inode->i_data, 0, -1);
1480 spin_lock(&ci->i_ceph_lock);
1482 if (inode->i_data.nrpages == 0 &&
1483 invalidating_gen == ci->i_rdcache_gen) {
1485 dout("try_nonblocking_invalidate %p success\n", inode);
1486 /* save any racing async invalidate some trouble */
1487 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1490 dout("try_nonblocking_invalidate %p failed\n", inode);
1495 * Swiss army knife function to examine currently used and wanted
1496 * versus held caps. Release, flush, ack revoked caps to mds as
1499 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1500 * cap release further.
1501 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1502 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1505 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1506 struct ceph_mds_session *session)
1508 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1509 struct ceph_mds_client *mdsc = fsc->mdsc;
1510 struct inode *inode = &ci->vfs_inode;
1511 struct ceph_cap *cap;
1512 int file_wanted, used, cap_used;
1513 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1514 int issued, implemented, want, retain, revoking, flushing = 0;
1515 int mds = -1; /* keep track of how far we've gone through i_caps list
1516 to avoid an infinite loop on retry */
1518 int tried_invalidate = 0;
1519 int delayed = 0, sent = 0, force_requeue = 0, num;
1520 int queue_invalidate = 0;
1521 int is_delayed = flags & CHECK_CAPS_NODELAY;
1523 /* if we are unmounting, flush any unused caps immediately. */
1527 spin_lock(&ci->i_ceph_lock);
1529 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1530 flags |= CHECK_CAPS_FLUSH;
1532 /* flush snaps first time around only */
1533 if (!list_empty(&ci->i_cap_snaps))
1534 __ceph_flush_snaps(ci, &session, 0);
1537 spin_lock(&ci->i_ceph_lock);
1539 file_wanted = __ceph_caps_file_wanted(ci);
1540 used = __ceph_caps_used(ci);
1541 want = file_wanted | used;
1542 issued = __ceph_caps_issued(ci, &implemented);
1543 revoking = implemented & ~issued;
1545 retain = want | CEPH_CAP_PIN;
1546 if (!mdsc->stopping && inode->i_nlink > 0) {
1548 retain |= CEPH_CAP_ANY; /* be greedy */
1550 retain |= CEPH_CAP_ANY_SHARED;
1552 * keep RD only if we didn't have the file open RW,
1553 * because then the mds would revoke it anyway to
1554 * journal max_size=0.
1556 if (ci->i_max_size == 0)
1557 retain |= CEPH_CAP_ANY_RD;
1561 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1562 " issued %s revoking %s retain %s %s%s%s\n", inode,
1563 ceph_cap_string(file_wanted),
1564 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1565 ceph_cap_string(ci->i_flushing_caps),
1566 ceph_cap_string(issued), ceph_cap_string(revoking),
1567 ceph_cap_string(retain),
1568 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1569 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1570 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1573 * If we no longer need to hold onto old our caps, and we may
1574 * have cached pages, but don't want them, then try to invalidate.
1575 * If we fail, it's because pages are locked.... try again later.
1577 if ((!is_delayed || mdsc->stopping) &&
1578 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1579 inode->i_data.nrpages && /* have cached pages */
1580 (file_wanted == 0 || /* no open files */
1581 (revoking & (CEPH_CAP_FILE_CACHE|
1582 CEPH_CAP_FILE_LAZYIO))) && /* or revoking cache */
1583 !tried_invalidate) {
1584 dout("check_caps trying to invalidate on %p\n", inode);
1585 if (try_nonblocking_invalidate(inode) < 0) {
1586 if (revoking & (CEPH_CAP_FILE_CACHE|
1587 CEPH_CAP_FILE_LAZYIO)) {
1588 dout("check_caps queuing invalidate\n");
1589 queue_invalidate = 1;
1590 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1592 dout("check_caps failed to invalidate pages\n");
1593 /* we failed to invalidate pages. check these
1594 caps again later. */
1596 __cap_set_timeouts(mdsc, ci);
1599 tried_invalidate = 1;
1604 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1605 cap = rb_entry(p, struct ceph_cap, ci_node);
1608 /* avoid looping forever */
1609 if (mds >= cap->mds ||
1610 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1613 /* NOTE: no side-effects allowed, until we take s_mutex */
1616 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1617 cap_used &= ~ci->i_auth_cap->issued;
1619 revoking = cap->implemented & ~cap->issued;
1620 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1621 cap->mds, cap, ceph_cap_string(cap->issued),
1622 ceph_cap_string(cap_used),
1623 ceph_cap_string(cap->implemented),
1624 ceph_cap_string(revoking));
1626 if (cap == ci->i_auth_cap &&
1627 (cap->issued & CEPH_CAP_FILE_WR)) {
1628 /* request larger max_size from MDS? */
1629 if (ci->i_wanted_max_size > ci->i_max_size &&
1630 ci->i_wanted_max_size > ci->i_requested_max_size) {
1631 dout("requesting new max_size\n");
1635 /* approaching file_max? */
1636 if ((inode->i_size << 1) >= ci->i_max_size &&
1637 (ci->i_reported_size << 1) < ci->i_max_size) {
1638 dout("i_size approaching max_size\n");
1642 /* flush anything dirty? */
1643 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1645 dout("flushing dirty caps\n");
1649 /* completed revocation? going down and there are no caps? */
1650 if (revoking && (revoking & cap_used) == 0) {
1651 dout("completed revocation of %s\n",
1652 ceph_cap_string(cap->implemented & ~cap->issued));
1656 /* want more caps from mds? */
1657 if (want & ~(cap->mds_wanted | cap->issued))
1660 /* things we might delay */
1661 if ((cap->issued & ~retain) == 0 &&
1662 cap->mds_wanted == want)
1663 continue; /* nope, all good */
1669 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1670 time_before(jiffies, ci->i_hold_caps_max)) {
1671 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1672 ceph_cap_string(cap->issued),
1673 ceph_cap_string(cap->issued & retain),
1674 ceph_cap_string(cap->mds_wanted),
1675 ceph_cap_string(want));
1681 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1682 dout(" skipping %p I_NOFLUSH set\n", inode);
1686 if (session && session != cap->session) {
1687 dout("oops, wrong session %p mutex\n", session);
1688 mutex_unlock(&session->s_mutex);
1692 session = cap->session;
1693 if (mutex_trylock(&session->s_mutex) == 0) {
1694 dout("inverting session/ino locks on %p\n",
1696 spin_unlock(&ci->i_ceph_lock);
1697 if (took_snap_rwsem) {
1698 up_read(&mdsc->snap_rwsem);
1699 took_snap_rwsem = 0;
1701 mutex_lock(&session->s_mutex);
1705 /* take snap_rwsem after session mutex */
1706 if (!took_snap_rwsem) {
1707 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1708 dout("inverting snap/in locks on %p\n",
1710 spin_unlock(&ci->i_ceph_lock);
1711 down_read(&mdsc->snap_rwsem);
1712 took_snap_rwsem = 1;
1715 took_snap_rwsem = 1;
1718 if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1719 flushing = __mark_caps_flushing(inode, session);
1723 mds = cap->mds; /* remember mds, so we don't repeat */
1726 /* __send_cap drops i_ceph_lock */
1727 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1728 want, retain, flushing, NULL);
1729 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1733 * Reschedule delayed caps release if we delayed anything,
1736 if (delayed && is_delayed)
1737 force_requeue = 1; /* __send_cap delayed release; requeue */
1738 if (!delayed && !is_delayed)
1739 __cap_delay_cancel(mdsc, ci);
1740 else if (!is_delayed || force_requeue)
1741 __cap_delay_requeue(mdsc, ci);
1743 spin_unlock(&ci->i_ceph_lock);
1745 if (queue_invalidate)
1746 ceph_queue_invalidate(inode);
1749 mutex_unlock(&session->s_mutex);
1750 if (took_snap_rwsem)
1751 up_read(&mdsc->snap_rwsem);
1755 * Try to flush dirty caps back to the auth mds.
1757 static int try_flush_caps(struct inode *inode, unsigned *flush_tid)
1759 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1760 struct ceph_inode_info *ci = ceph_inode(inode);
1762 struct ceph_mds_session *session = NULL;
1765 spin_lock(&ci->i_ceph_lock);
1766 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1767 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1770 if (ci->i_dirty_caps && ci->i_auth_cap) {
1771 struct ceph_cap *cap = ci->i_auth_cap;
1772 int used = __ceph_caps_used(ci);
1773 int want = __ceph_caps_wanted(ci);
1776 if (!session || session != cap->session) {
1777 spin_unlock(&ci->i_ceph_lock);
1779 mutex_unlock(&session->s_mutex);
1780 session = cap->session;
1781 mutex_lock(&session->s_mutex);
1784 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1787 flushing = __mark_caps_flushing(inode, session);
1789 /* __send_cap drops i_ceph_lock */
1790 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1791 cap->issued | cap->implemented, flushing,
1796 spin_lock(&ci->i_ceph_lock);
1797 __cap_delay_requeue(mdsc, ci);
1800 spin_unlock(&ci->i_ceph_lock);
1803 mutex_unlock(&session->s_mutex);
1808 * Return true if we've flushed caps through the given flush_tid.
1810 static int caps_are_flushed(struct inode *inode, unsigned tid)
1812 struct ceph_inode_info *ci = ceph_inode(inode);
1815 spin_lock(&ci->i_ceph_lock);
1816 for (i = 0; i < CEPH_CAP_BITS; i++)
1817 if ((ci->i_flushing_caps & (1 << i)) &&
1818 ci->i_cap_flush_tid[i] <= tid) {
1819 /* still flushing this bit */
1823 spin_unlock(&ci->i_ceph_lock);
1828 * Wait on any unsafe replies for the given inode. First wait on the
1829 * newest request, and make that the upper bound. Then, if there are
1830 * more requests, keep waiting on the oldest as long as it is still older
1831 * than the original request.
1833 static void sync_write_wait(struct inode *inode)
1835 struct ceph_inode_info *ci = ceph_inode(inode);
1836 struct list_head *head = &ci->i_unsafe_writes;
1837 struct ceph_osd_request *req;
1840 spin_lock(&ci->i_unsafe_lock);
1841 if (list_empty(head))
1844 /* set upper bound as _last_ entry in chain */
1845 req = list_entry(head->prev, struct ceph_osd_request,
1847 last_tid = req->r_tid;
1850 ceph_osdc_get_request(req);
1851 spin_unlock(&ci->i_unsafe_lock);
1852 dout("sync_write_wait on tid %llu (until %llu)\n",
1853 req->r_tid, last_tid);
1854 wait_for_completion(&req->r_safe_completion);
1855 spin_lock(&ci->i_unsafe_lock);
1856 ceph_osdc_put_request(req);
1859 * from here on look at first entry in chain, since we
1860 * only want to wait for anything older than last_tid
1862 if (list_empty(head))
1864 req = list_entry(head->next, struct ceph_osd_request,
1866 } while (req->r_tid < last_tid);
1868 spin_unlock(&ci->i_unsafe_lock);
1871 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1873 struct inode *inode = file->f_mapping->host;
1874 struct ceph_inode_info *ci = ceph_inode(inode);
1879 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1880 sync_write_wait(inode);
1882 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1885 mutex_lock(&inode->i_mutex);
1887 dirty = try_flush_caps(inode, &flush_tid);
1888 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1891 * only wait on non-file metadata writeback (the mds
1892 * can recover size and mtime, so we don't need to
1895 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1896 dout("fsync waiting for flush_tid %u\n", flush_tid);
1897 ret = wait_event_interruptible(ci->i_cap_wq,
1898 caps_are_flushed(inode, flush_tid));
1901 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1902 mutex_unlock(&inode->i_mutex);
1907 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1908 * queue inode for flush but don't do so immediately, because we can
1909 * get by with fewer MDS messages if we wait for data writeback to
1912 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1914 struct ceph_inode_info *ci = ceph_inode(inode);
1918 int wait = wbc->sync_mode == WB_SYNC_ALL;
1920 dout("write_inode %p wait=%d\n", inode, wait);
1922 dirty = try_flush_caps(inode, &flush_tid);
1924 err = wait_event_interruptible(ci->i_cap_wq,
1925 caps_are_flushed(inode, flush_tid));
1927 struct ceph_mds_client *mdsc =
1928 ceph_sb_to_client(inode->i_sb)->mdsc;
1930 spin_lock(&ci->i_ceph_lock);
1931 if (__ceph_caps_dirty(ci))
1932 __cap_delay_requeue_front(mdsc, ci);
1933 spin_unlock(&ci->i_ceph_lock);
1939 * After a recovering MDS goes active, we need to resend any caps
1942 * Caller holds session->s_mutex.
1944 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1945 struct ceph_mds_session *session)
1947 struct ceph_cap_snap *capsnap;
1949 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1950 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1952 struct ceph_inode_info *ci = capsnap->ci;
1953 struct inode *inode = &ci->vfs_inode;
1954 struct ceph_cap *cap;
1956 spin_lock(&ci->i_ceph_lock);
1957 cap = ci->i_auth_cap;
1958 if (cap && cap->session == session) {
1959 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1961 __ceph_flush_snaps(ci, &session, 1);
1963 pr_err("%p auth cap %p not mds%d ???\n", inode,
1964 cap, session->s_mds);
1966 spin_unlock(&ci->i_ceph_lock);
1970 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1971 struct ceph_mds_session *session)
1973 struct ceph_inode_info *ci;
1975 kick_flushing_capsnaps(mdsc, session);
1977 dout("kick_flushing_caps mds%d\n", session->s_mds);
1978 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1979 struct inode *inode = &ci->vfs_inode;
1980 struct ceph_cap *cap;
1983 spin_lock(&ci->i_ceph_lock);
1984 cap = ci->i_auth_cap;
1985 if (cap && cap->session == session) {
1986 dout("kick_flushing_caps %p cap %p %s\n", inode,
1987 cap, ceph_cap_string(ci->i_flushing_caps));
1988 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1989 __ceph_caps_used(ci),
1990 __ceph_caps_wanted(ci),
1991 cap->issued | cap->implemented,
1992 ci->i_flushing_caps, NULL);
1994 spin_lock(&ci->i_ceph_lock);
1995 __cap_delay_requeue(mdsc, ci);
1996 spin_unlock(&ci->i_ceph_lock);
1999 pr_err("%p auth cap %p not mds%d ???\n", inode,
2000 cap, session->s_mds);
2001 spin_unlock(&ci->i_ceph_lock);
2006 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2007 struct ceph_mds_session *session,
2008 struct inode *inode)
2010 struct ceph_inode_info *ci = ceph_inode(inode);
2011 struct ceph_cap *cap;
2014 spin_lock(&ci->i_ceph_lock);
2015 cap = ci->i_auth_cap;
2016 dout("kick_flushing_inode_caps %p flushing %s flush_seq %lld\n", inode,
2017 ceph_cap_string(ci->i_flushing_caps), ci->i_cap_flush_seq);
2019 __ceph_flush_snaps(ci, &session, 1);
2021 if (ci->i_flushing_caps) {
2022 spin_lock(&mdsc->cap_dirty_lock);
2023 list_move_tail(&ci->i_flushing_item,
2024 &cap->session->s_cap_flushing);
2025 spin_unlock(&mdsc->cap_dirty_lock);
2027 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2028 __ceph_caps_used(ci),
2029 __ceph_caps_wanted(ci),
2030 cap->issued | cap->implemented,
2031 ci->i_flushing_caps, NULL);
2033 spin_lock(&ci->i_ceph_lock);
2034 __cap_delay_requeue(mdsc, ci);
2035 spin_unlock(&ci->i_ceph_lock);
2038 spin_unlock(&ci->i_ceph_lock);
2044 * Take references to capabilities we hold, so that we don't release
2045 * them to the MDS prematurely.
2047 * Protected by i_ceph_lock.
2049 static void __take_cap_refs(struct ceph_inode_info *ci, int got)
2051 if (got & CEPH_CAP_PIN)
2053 if (got & CEPH_CAP_FILE_RD)
2055 if (got & CEPH_CAP_FILE_CACHE)
2056 ci->i_rdcache_ref++;
2057 if (got & CEPH_CAP_FILE_WR)
2059 if (got & CEPH_CAP_FILE_BUFFER) {
2060 if (ci->i_wb_ref == 0)
2061 ihold(&ci->vfs_inode);
2063 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2064 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2069 * Try to grab cap references. Specify those refs we @want, and the
2070 * minimal set we @need. Also include the larger offset we are writing
2071 * to (when applicable), and check against max_size here as well.
2072 * Note that caller is responsible for ensuring max_size increases are
2073 * requested from the MDS.
2075 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2076 loff_t endoff, int *got, struct page **pinned_page,
2077 int *check_max, int *err)
2079 struct inode *inode = &ci->vfs_inode;
2081 int have, implemented, _got = 0;
2084 dout("get_cap_refs %p need %s want %s\n", inode,
2085 ceph_cap_string(need), ceph_cap_string(want));
2087 spin_lock(&ci->i_ceph_lock);
2089 /* make sure file is actually open */
2090 file_wanted = __ceph_caps_file_wanted(ci);
2091 if ((file_wanted & need) == 0) {
2092 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2093 ceph_cap_string(need), ceph_cap_string(file_wanted));
2099 /* finish pending truncate */
2100 while (ci->i_truncate_pending) {
2101 spin_unlock(&ci->i_ceph_lock);
2102 __ceph_do_pending_vmtruncate(inode);
2103 spin_lock(&ci->i_ceph_lock);
2106 have = __ceph_caps_issued(ci, &implemented);
2108 if (have & need & CEPH_CAP_FILE_WR) {
2109 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2110 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2111 inode, endoff, ci->i_max_size);
2112 if (endoff > ci->i_requested_max_size) {
2119 * If a sync write is in progress, we must wait, so that we
2120 * can get a final snapshot value for size+mtime.
2122 if (__ceph_have_pending_cap_snap(ci)) {
2123 dout("get_cap_refs %p cap_snap_pending\n", inode);
2128 if ((have & need) == need) {
2130 * Look at (implemented & ~have & not) so that we keep waiting
2131 * on transition from wanted -> needed caps. This is needed
2132 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2133 * going before a prior buffered writeback happens.
2135 int not = want & ~(have & need);
2136 int revoking = implemented & ~have;
2137 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2138 inode, ceph_cap_string(have), ceph_cap_string(not),
2139 ceph_cap_string(revoking));
2140 if ((revoking & not) == 0) {
2141 _got = need | (have & want);
2142 __take_cap_refs(ci, _got);
2146 dout("get_cap_refs %p have %s needed %s\n", inode,
2147 ceph_cap_string(have), ceph_cap_string(need));
2150 spin_unlock(&ci->i_ceph_lock);
2152 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2153 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2154 i_size_read(inode) > 0) {
2156 struct page *page = find_get_page(inode->i_mapping, 0);
2158 if (PageUptodate(page)) {
2159 *pinned_page = page;
2162 page_cache_release(page);
2165 * drop cap refs first because getattr while holding
2166 * caps refs can cause deadlock.
2168 ceph_put_cap_refs(ci, _got);
2171 /* getattr request will bring inline data into page cache */
2172 ret1 = __ceph_do_getattr(inode, NULL,
2173 CEPH_STAT_CAP_INLINE_DATA, true);
2182 dout("get_cap_refs %p ret %d got %s\n", inode,
2183 ret, ceph_cap_string(_got));
2189 * Check the offset we are writing up to against our current
2190 * max_size. If necessary, tell the MDS we want to write to
2193 static void check_max_size(struct inode *inode, loff_t endoff)
2195 struct ceph_inode_info *ci = ceph_inode(inode);
2198 /* do we need to explicitly request a larger max_size? */
2199 spin_lock(&ci->i_ceph_lock);
2200 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2201 dout("write %p at large endoff %llu, req max_size\n",
2203 ci->i_wanted_max_size = endoff;
2205 /* duplicate ceph_check_caps()'s logic */
2206 if (ci->i_auth_cap &&
2207 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2208 ci->i_wanted_max_size > ci->i_max_size &&
2209 ci->i_wanted_max_size > ci->i_requested_max_size)
2211 spin_unlock(&ci->i_ceph_lock);
2213 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2217 * Wait for caps, and take cap references. If we can't get a WR cap
2218 * due to a small max_size, make sure we check_max_size (and possibly
2219 * ask the mds) so we don't get hung up indefinitely.
2221 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2222 loff_t endoff, int *got, struct page **pinned_page)
2224 int check_max, ret, err;
2228 check_max_size(&ci->vfs_inode, endoff);
2231 ret = wait_event_interruptible(ci->i_cap_wq,
2232 try_get_cap_refs(ci, need, want, endoff,
2243 * Take cap refs. Caller must already know we hold at least one ref
2244 * on the caps in question or we don't know this is safe.
2246 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2248 spin_lock(&ci->i_ceph_lock);
2249 __take_cap_refs(ci, caps);
2250 spin_unlock(&ci->i_ceph_lock);
2256 * If we released the last ref on any given cap, call ceph_check_caps
2257 * to release (or schedule a release).
2259 * If we are releasing a WR cap (from a sync write), finalize any affected
2260 * cap_snap, and wake up any waiters.
2262 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2264 struct inode *inode = &ci->vfs_inode;
2265 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2266 struct ceph_cap_snap *capsnap;
2268 spin_lock(&ci->i_ceph_lock);
2269 if (had & CEPH_CAP_PIN)
2271 if (had & CEPH_CAP_FILE_RD)
2272 if (--ci->i_rd_ref == 0)
2274 if (had & CEPH_CAP_FILE_CACHE)
2275 if (--ci->i_rdcache_ref == 0)
2277 if (had & CEPH_CAP_FILE_BUFFER) {
2278 if (--ci->i_wb_ref == 0) {
2282 dout("put_cap_refs %p wb %d -> %d (?)\n",
2283 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2285 if (had & CEPH_CAP_FILE_WR)
2286 if (--ci->i_wr_ref == 0) {
2288 if (!list_empty(&ci->i_cap_snaps)) {
2289 capsnap = list_first_entry(&ci->i_cap_snaps,
2290 struct ceph_cap_snap,
2292 if (capsnap->writing) {
2293 capsnap->writing = 0;
2295 __ceph_finish_cap_snap(ci,
2301 spin_unlock(&ci->i_ceph_lock);
2303 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2304 last ? " last" : "", put ? " put" : "");
2306 if (last && !flushsnaps)
2307 ceph_check_caps(ci, 0, NULL);
2308 else if (flushsnaps)
2309 ceph_flush_snaps(ci);
2311 wake_up_all(&ci->i_cap_wq);
2317 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2318 * context. Adjust per-snap dirty page accounting as appropriate.
2319 * Once all dirty data for a cap_snap is flushed, flush snapped file
2320 * metadata back to the MDS. If we dropped the last ref, call
2323 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2324 struct ceph_snap_context *snapc)
2326 struct inode *inode = &ci->vfs_inode;
2328 int complete_capsnap = 0;
2329 int drop_capsnap = 0;
2331 struct ceph_cap_snap *capsnap = NULL;
2333 spin_lock(&ci->i_ceph_lock);
2334 ci->i_wrbuffer_ref -= nr;
2335 last = !ci->i_wrbuffer_ref;
2337 if (ci->i_head_snapc == snapc) {
2338 ci->i_wrbuffer_ref_head -= nr;
2339 if (ci->i_wrbuffer_ref_head == 0 &&
2340 ci->i_dirty_caps == 0 && ci->i_flushing_caps == 0) {
2341 BUG_ON(!ci->i_head_snapc);
2342 ceph_put_snap_context(ci->i_head_snapc);
2343 ci->i_head_snapc = NULL;
2345 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2347 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2348 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2349 last ? " LAST" : "");
2351 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2352 if (capsnap->context == snapc) {
2358 capsnap->dirty_pages -= nr;
2359 if (capsnap->dirty_pages == 0) {
2360 complete_capsnap = 1;
2361 if (capsnap->dirty == 0)
2362 /* cap writeback completed before we created
2363 * the cap_snap; no FLUSHSNAP is needed */
2366 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2367 " snap %lld %d/%d -> %d/%d %s%s%s\n",
2368 inode, capsnap, capsnap->context->seq,
2369 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2370 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2371 last ? " (wrbuffer last)" : "",
2372 complete_capsnap ? " (complete capsnap)" : "",
2373 drop_capsnap ? " (drop capsnap)" : "");
2375 ceph_put_snap_context(capsnap->context);
2376 list_del(&capsnap->ci_item);
2377 list_del(&capsnap->flushing_item);
2378 ceph_put_cap_snap(capsnap);
2382 spin_unlock(&ci->i_ceph_lock);
2385 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2387 } else if (complete_capsnap) {
2388 ceph_flush_snaps(ci);
2389 wake_up_all(&ci->i_cap_wq);
2396 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2398 static void invalidate_aliases(struct inode *inode)
2400 struct dentry *dn, *prev = NULL;
2402 dout("invalidate_aliases inode %p\n", inode);
2403 d_prune_aliases(inode);
2405 * For non-directory inode, d_find_alias() only returns
2406 * hashed dentry. After calling d_invalidate(), the
2407 * dentry becomes unhashed.
2409 * For directory inode, d_find_alias() can return
2410 * unhashed dentry. But directory inode should have
2411 * one alias at most.
2413 while ((dn = d_find_alias(inode))) {
2428 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2429 * actually be a revocation if it specifies a smaller cap set.)
2431 * caller holds s_mutex and i_ceph_lock, we drop both.
2433 static void handle_cap_grant(struct ceph_mds_client *mdsc,
2434 struct inode *inode, struct ceph_mds_caps *grant,
2435 void *snaptrace, int snaptrace_len,
2437 void *inline_data, int inline_len,
2438 struct ceph_buffer *xattr_buf,
2439 struct ceph_mds_session *session,
2440 struct ceph_cap *cap, int issued)
2441 __releases(ci->i_ceph_lock)
2443 struct ceph_inode_info *ci = ceph_inode(inode);
2444 int mds = session->s_mds;
2445 int seq = le32_to_cpu(grant->seq);
2446 int newcaps = le32_to_cpu(grant->caps);
2447 int used, wanted, dirty;
2448 u64 size = le64_to_cpu(grant->size);
2449 u64 max_size = le64_to_cpu(grant->max_size);
2450 struct timespec mtime, atime, ctime;
2453 bool writeback = false;
2454 bool queue_trunc = false;
2455 bool queue_invalidate = false;
2456 bool queue_revalidate = false;
2457 bool deleted_inode = false;
2458 bool fill_inline = false;
2460 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2461 inode, cap, mds, seq, ceph_cap_string(newcaps));
2462 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2467 * auth mds of the inode changed. we received the cap export message,
2468 * but still haven't received the cap import message. handle_cap_export
2469 * updated the new auth MDS' cap.
2471 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2472 * that was sent before the cap import message. So don't remove caps.
2474 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2475 WARN_ON(cap != ci->i_auth_cap);
2476 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2478 newcaps |= cap->issued;
2482 * If CACHE is being revoked, and we have no dirty buffers,
2483 * try to invalidate (once). (If there are dirty buffers, we
2484 * will invalidate _after_ writeback.)
2486 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2487 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2488 !ci->i_wrbuffer_ref) {
2489 if (try_nonblocking_invalidate(inode)) {
2490 /* there were locked pages.. invalidate later
2491 in a separate thread. */
2492 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2493 queue_invalidate = true;
2494 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2498 ceph_fscache_invalidate(inode);
2501 /* side effects now are allowed */
2502 cap->cap_gen = session->s_cap_gen;
2505 __check_cap_issue(ci, cap, newcaps);
2507 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2508 (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2509 inode->i_mode = le32_to_cpu(grant->mode);
2510 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2511 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2512 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2513 from_kuid(&init_user_ns, inode->i_uid),
2514 from_kgid(&init_user_ns, inode->i_gid));
2517 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2518 (issued & CEPH_CAP_LINK_EXCL) == 0) {
2519 set_nlink(inode, le32_to_cpu(grant->nlink));
2520 if (inode->i_nlink == 0 &&
2521 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2522 deleted_inode = true;
2525 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2526 int len = le32_to_cpu(grant->xattr_len);
2527 u64 version = le64_to_cpu(grant->xattr_version);
2529 if (version > ci->i_xattrs.version) {
2530 dout(" got new xattrs v%llu on %p len %d\n",
2531 version, inode, len);
2532 if (ci->i_xattrs.blob)
2533 ceph_buffer_put(ci->i_xattrs.blob);
2534 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2535 ci->i_xattrs.version = version;
2536 ceph_forget_all_cached_acls(inode);
2540 /* Do we need to revalidate our fscache cookie. Don't bother on the
2541 * first cache cap as we already validate at cookie creation time. */
2542 if ((issued & CEPH_CAP_FILE_CACHE) && ci->i_rdcache_gen > 1)
2543 queue_revalidate = true;
2545 if (newcaps & CEPH_CAP_ANY_RD) {
2546 /* ctime/mtime/atime? */
2547 ceph_decode_timespec(&mtime, &grant->mtime);
2548 ceph_decode_timespec(&atime, &grant->atime);
2549 ceph_decode_timespec(&ctime, &grant->ctime);
2550 ceph_fill_file_time(inode, issued,
2551 le32_to_cpu(grant->time_warp_seq),
2552 &ctime, &mtime, &atime);
2555 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
2556 /* file layout may have changed */
2557 ci->i_layout = grant->layout;
2558 /* size/truncate_seq? */
2559 queue_trunc = ceph_fill_file_size(inode, issued,
2560 le32_to_cpu(grant->truncate_seq),
2561 le64_to_cpu(grant->truncate_size),
2563 /* max size increase? */
2564 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
2565 dout("max_size %lld -> %llu\n",
2566 ci->i_max_size, max_size);
2567 ci->i_max_size = max_size;
2568 if (max_size >= ci->i_wanted_max_size) {
2569 ci->i_wanted_max_size = 0; /* reset */
2570 ci->i_requested_max_size = 0;
2576 /* check cap bits */
2577 wanted = __ceph_caps_wanted(ci);
2578 used = __ceph_caps_used(ci);
2579 dirty = __ceph_caps_dirty(ci);
2580 dout(" my wanted = %s, used = %s, dirty %s\n",
2581 ceph_cap_string(wanted),
2582 ceph_cap_string(used),
2583 ceph_cap_string(dirty));
2584 if (wanted != le32_to_cpu(grant->wanted)) {
2585 dout("mds wanted %s -> %s\n",
2586 ceph_cap_string(le32_to_cpu(grant->wanted)),
2587 ceph_cap_string(wanted));
2588 /* imported cap may not have correct mds_wanted */
2589 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
2593 /* revocation, grant, or no-op? */
2594 if (cap->issued & ~newcaps) {
2595 int revoking = cap->issued & ~newcaps;
2597 dout("revocation: %s -> %s (revoking %s)\n",
2598 ceph_cap_string(cap->issued),
2599 ceph_cap_string(newcaps),
2600 ceph_cap_string(revoking));
2601 if (revoking & used & CEPH_CAP_FILE_BUFFER)
2602 writeback = true; /* initiate writeback; will delay ack */
2603 else if (revoking == CEPH_CAP_FILE_CACHE &&
2604 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2606 ; /* do nothing yet, invalidation will be queued */
2607 else if (cap == ci->i_auth_cap)
2608 check_caps = 1; /* check auth cap only */
2610 check_caps = 2; /* check all caps */
2611 cap->issued = newcaps;
2612 cap->implemented |= newcaps;
2613 } else if (cap->issued == newcaps) {
2614 dout("caps unchanged: %s -> %s\n",
2615 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2617 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2618 ceph_cap_string(newcaps));
2619 /* non-auth MDS is revoking the newly grant caps ? */
2620 if (cap == ci->i_auth_cap &&
2621 __ceph_caps_revoking_other(ci, cap, newcaps))
2624 cap->issued = newcaps;
2625 cap->implemented |= newcaps; /* add bits only, to
2626 * avoid stepping on a
2627 * pending revocation */
2630 BUG_ON(cap->issued & ~cap->implemented);
2632 if (inline_version > 0 && inline_version >= ci->i_inline_version) {
2633 ci->i_inline_version = inline_version;
2634 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2635 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
2639 spin_unlock(&ci->i_ceph_lock);
2641 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
2642 down_write(&mdsc->snap_rwsem);
2643 ceph_update_snap_trace(mdsc, snaptrace,
2644 snaptrace + snaptrace_len, false);
2645 downgrade_write(&mdsc->snap_rwsem);
2646 kick_flushing_inode_caps(mdsc, session, inode);
2647 up_read(&mdsc->snap_rwsem);
2648 if (newcaps & ~issued)
2653 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
2656 ceph_queue_vmtruncate(inode);
2657 ceph_queue_revalidate(inode);
2658 } else if (queue_revalidate)
2659 ceph_queue_revalidate(inode);
2663 * queue inode for writeback: we can't actually call
2664 * filemap_write_and_wait, etc. from message handler
2667 ceph_queue_writeback(inode);
2668 if (queue_invalidate)
2669 ceph_queue_invalidate(inode);
2671 invalidate_aliases(inode);
2673 wake_up_all(&ci->i_cap_wq);
2675 if (check_caps == 1)
2676 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2678 else if (check_caps == 2)
2679 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2681 mutex_unlock(&session->s_mutex);
2685 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2686 * MDS has been safely committed.
2688 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2689 struct ceph_mds_caps *m,
2690 struct ceph_mds_session *session,
2691 struct ceph_cap *cap)
2692 __releases(ci->i_ceph_lock)
2694 struct ceph_inode_info *ci = ceph_inode(inode);
2695 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2696 unsigned seq = le32_to_cpu(m->seq);
2697 int dirty = le32_to_cpu(m->dirty);
2702 for (i = 0; i < CEPH_CAP_BITS; i++)
2703 if ((dirty & (1 << i)) &&
2704 (u16)flush_tid == ci->i_cap_flush_tid[i])
2707 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2708 " flushing %s -> %s\n",
2709 inode, session->s_mds, seq, ceph_cap_string(dirty),
2710 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2711 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2713 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2716 ci->i_flushing_caps &= ~cleaned;
2718 spin_lock(&mdsc->cap_dirty_lock);
2719 if (ci->i_flushing_caps == 0) {
2720 list_del_init(&ci->i_flushing_item);
2721 if (!list_empty(&session->s_cap_flushing))
2722 dout(" mds%d still flushing cap on %p\n",
2724 &list_entry(session->s_cap_flushing.next,
2725 struct ceph_inode_info,
2726 i_flushing_item)->vfs_inode);
2727 mdsc->num_cap_flushing--;
2728 wake_up_all(&mdsc->cap_flushing_wq);
2729 dout(" inode %p now !flushing\n", inode);
2731 if (ci->i_dirty_caps == 0) {
2732 dout(" inode %p now clean\n", inode);
2733 BUG_ON(!list_empty(&ci->i_dirty_item));
2735 if (ci->i_wrbuffer_ref_head == 0) {
2736 BUG_ON(!ci->i_head_snapc);
2737 ceph_put_snap_context(ci->i_head_snapc);
2738 ci->i_head_snapc = NULL;
2741 BUG_ON(list_empty(&ci->i_dirty_item));
2744 spin_unlock(&mdsc->cap_dirty_lock);
2745 wake_up_all(&ci->i_cap_wq);
2748 spin_unlock(&ci->i_ceph_lock);
2754 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
2755 * throw away our cap_snap.
2757 * Caller hold s_mutex.
2759 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2760 struct ceph_mds_caps *m,
2761 struct ceph_mds_session *session)
2763 struct ceph_inode_info *ci = ceph_inode(inode);
2764 u64 follows = le64_to_cpu(m->snap_follows);
2765 struct ceph_cap_snap *capsnap;
2768 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2769 inode, ci, session->s_mds, follows);
2771 spin_lock(&ci->i_ceph_lock);
2772 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2773 if (capsnap->follows == follows) {
2774 if (capsnap->flush_tid != flush_tid) {
2775 dout(" cap_snap %p follows %lld tid %lld !="
2776 " %lld\n", capsnap, follows,
2777 flush_tid, capsnap->flush_tid);
2780 WARN_ON(capsnap->dirty_pages || capsnap->writing);
2781 dout(" removing %p cap_snap %p follows %lld\n",
2782 inode, capsnap, follows);
2783 ceph_put_snap_context(capsnap->context);
2784 list_del(&capsnap->ci_item);
2785 list_del(&capsnap->flushing_item);
2786 ceph_put_cap_snap(capsnap);
2790 dout(" skipping cap_snap %p follows %lld\n",
2791 capsnap, capsnap->follows);
2794 spin_unlock(&ci->i_ceph_lock);
2800 * Handle TRUNC from MDS, indicating file truncation.
2802 * caller hold s_mutex.
2804 static void handle_cap_trunc(struct inode *inode,
2805 struct ceph_mds_caps *trunc,
2806 struct ceph_mds_session *session)
2807 __releases(ci->i_ceph_lock)
2809 struct ceph_inode_info *ci = ceph_inode(inode);
2810 int mds = session->s_mds;
2811 int seq = le32_to_cpu(trunc->seq);
2812 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2813 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2814 u64 size = le64_to_cpu(trunc->size);
2815 int implemented = 0;
2816 int dirty = __ceph_caps_dirty(ci);
2817 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2818 int queue_trunc = 0;
2820 issued |= implemented | dirty;
2822 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2823 inode, mds, seq, truncate_size, truncate_seq);
2824 queue_trunc = ceph_fill_file_size(inode, issued,
2825 truncate_seq, truncate_size, size);
2826 spin_unlock(&ci->i_ceph_lock);
2829 ceph_queue_vmtruncate(inode);
2830 ceph_fscache_invalidate(inode);
2835 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
2836 * different one. If we are the most recent migration we've seen (as
2837 * indicated by mseq), make note of the migrating cap bits for the
2838 * duration (until we see the corresponding IMPORT).
2840 * caller holds s_mutex
2842 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2843 struct ceph_mds_cap_peer *ph,
2844 struct ceph_mds_session *session)
2846 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2847 struct ceph_mds_session *tsession = NULL;
2848 struct ceph_cap *cap, *tcap, *new_cap = NULL;
2849 struct ceph_inode_info *ci = ceph_inode(inode);
2851 unsigned mseq = le32_to_cpu(ex->migrate_seq);
2852 unsigned t_seq, t_mseq;
2854 int mds = session->s_mds;
2857 t_cap_id = le64_to_cpu(ph->cap_id);
2858 t_seq = le32_to_cpu(ph->seq);
2859 t_mseq = le32_to_cpu(ph->mseq);
2860 target = le32_to_cpu(ph->mds);
2862 t_cap_id = t_seq = t_mseq = 0;
2866 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
2867 inode, ci, mds, mseq, target);
2869 spin_lock(&ci->i_ceph_lock);
2870 cap = __get_cap_for_mds(ci, mds);
2871 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
2875 __ceph_remove_cap(cap, false);
2880 * now we know we haven't received the cap import message yet
2881 * because the exported cap still exist.
2884 issued = cap->issued;
2885 WARN_ON(issued != cap->implemented);
2887 tcap = __get_cap_for_mds(ci, target);
2889 /* already have caps from the target */
2890 if (tcap->cap_id != t_cap_id ||
2891 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
2892 dout(" updating import cap %p mds%d\n", tcap, target);
2893 tcap->cap_id = t_cap_id;
2894 tcap->seq = t_seq - 1;
2895 tcap->issue_seq = t_seq - 1;
2896 tcap->mseq = t_mseq;
2897 tcap->issued |= issued;
2898 tcap->implemented |= issued;
2899 if (cap == ci->i_auth_cap)
2900 ci->i_auth_cap = tcap;
2901 if (ci->i_flushing_caps && ci->i_auth_cap == tcap) {
2902 spin_lock(&mdsc->cap_dirty_lock);
2903 list_move_tail(&ci->i_flushing_item,
2904 &tcap->session->s_cap_flushing);
2905 spin_unlock(&mdsc->cap_dirty_lock);
2908 __ceph_remove_cap(cap, false);
2910 } else if (tsession) {
2911 /* add placeholder for the export tagert */
2912 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
2913 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
2914 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
2916 __ceph_remove_cap(cap, false);
2920 spin_unlock(&ci->i_ceph_lock);
2921 mutex_unlock(&session->s_mutex);
2923 /* open target session */
2924 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
2925 if (!IS_ERR(tsession)) {
2927 mutex_lock(&session->s_mutex);
2928 mutex_lock_nested(&tsession->s_mutex,
2929 SINGLE_DEPTH_NESTING);
2931 mutex_lock(&tsession->s_mutex);
2932 mutex_lock_nested(&session->s_mutex,
2933 SINGLE_DEPTH_NESTING);
2935 ceph_add_cap_releases(mdsc, tsession);
2936 new_cap = ceph_get_cap(mdsc, NULL);
2945 spin_unlock(&ci->i_ceph_lock);
2946 mutex_unlock(&session->s_mutex);
2948 mutex_unlock(&tsession->s_mutex);
2949 ceph_put_mds_session(tsession);
2952 ceph_put_cap(mdsc, new_cap);
2956 * Handle cap IMPORT.
2958 * caller holds s_mutex. acquires i_ceph_lock
2960 static void handle_cap_import(struct ceph_mds_client *mdsc,
2961 struct inode *inode, struct ceph_mds_caps *im,
2962 struct ceph_mds_cap_peer *ph,
2963 struct ceph_mds_session *session,
2964 struct ceph_cap **target_cap, int *old_issued)
2965 __acquires(ci->i_ceph_lock)
2967 struct ceph_inode_info *ci = ceph_inode(inode);
2968 struct ceph_cap *cap, *ocap, *new_cap = NULL;
2969 int mds = session->s_mds;
2971 unsigned caps = le32_to_cpu(im->caps);
2972 unsigned wanted = le32_to_cpu(im->wanted);
2973 unsigned seq = le32_to_cpu(im->seq);
2974 unsigned mseq = le32_to_cpu(im->migrate_seq);
2975 u64 realmino = le64_to_cpu(im->realm);
2976 u64 cap_id = le64_to_cpu(im->cap_id);
2981 p_cap_id = le64_to_cpu(ph->cap_id);
2982 peer = le32_to_cpu(ph->mds);
2988 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
2989 inode, ci, mds, mseq, peer);
2992 spin_lock(&ci->i_ceph_lock);
2993 cap = __get_cap_for_mds(ci, mds);
2996 spin_unlock(&ci->i_ceph_lock);
2997 new_cap = ceph_get_cap(mdsc, NULL);
3003 ceph_put_cap(mdsc, new_cap);
3008 __ceph_caps_issued(ci, &issued);
3009 issued |= __ceph_caps_dirty(ci);
3011 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3012 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3014 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3015 if (ocap && ocap->cap_id == p_cap_id) {
3016 dout(" remove export cap %p mds%d flags %d\n",
3017 ocap, peer, ph->flags);
3018 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3019 (ocap->seq != le32_to_cpu(ph->seq) ||
3020 ocap->mseq != le32_to_cpu(ph->mseq))) {
3021 pr_err("handle_cap_import: mismatched seq/mseq: "
3022 "ino (%llx.%llx) mds%d seq %d mseq %d "
3023 "importer mds%d has peer seq %d mseq %d\n",
3024 ceph_vinop(inode), peer, ocap->seq,
3025 ocap->mseq, mds, le32_to_cpu(ph->seq),
3026 le32_to_cpu(ph->mseq));
3028 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3031 /* make sure we re-request max_size, if necessary */
3032 ci->i_wanted_max_size = 0;
3033 ci->i_requested_max_size = 0;
3035 *old_issued = issued;
3040 * Handle a caps message from the MDS.
3042 * Identify the appropriate session, inode, and call the right handler
3043 * based on the cap op.
3045 void ceph_handle_caps(struct ceph_mds_session *session,
3046 struct ceph_msg *msg)
3048 struct ceph_mds_client *mdsc = session->s_mdsc;
3049 struct super_block *sb = mdsc->fsc->sb;
3050 struct inode *inode;
3051 struct ceph_inode_info *ci;
3052 struct ceph_cap *cap;
3053 struct ceph_mds_caps *h;
3054 struct ceph_mds_cap_peer *peer = NULL;
3055 int mds = session->s_mds;
3058 struct ceph_vino vino;
3062 u64 inline_version = 0;
3063 void *inline_data = NULL;
3066 size_t snaptrace_len;
3069 dout("handle_caps from mds%d\n", mds);
3072 end = msg->front.iov_base + msg->front.iov_len;
3073 tid = le64_to_cpu(msg->hdr.tid);
3074 if (msg->front.iov_len < sizeof(*h))
3076 h = msg->front.iov_base;
3077 op = le32_to_cpu(h->op);
3078 vino.ino = le64_to_cpu(h->ino);
3079 vino.snap = CEPH_NOSNAP;
3080 cap_id = le64_to_cpu(h->cap_id);
3081 seq = le32_to_cpu(h->seq);
3082 mseq = le32_to_cpu(h->migrate_seq);
3083 size = le64_to_cpu(h->size);
3084 max_size = le64_to_cpu(h->max_size);
3087 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3088 p = snaptrace + snaptrace_len;
3090 if (le16_to_cpu(msg->hdr.version) >= 2) {
3092 ceph_decode_32_safe(&p, end, flock_len, bad);
3093 if (p + flock_len > end)
3098 if (le16_to_cpu(msg->hdr.version) >= 3) {
3099 if (op == CEPH_CAP_OP_IMPORT) {
3100 if (p + sizeof(*peer) > end)
3104 } else if (op == CEPH_CAP_OP_EXPORT) {
3105 /* recorded in unused fields */
3106 peer = (void *)&h->size;
3110 if (le16_to_cpu(msg->hdr.version) >= 4) {
3111 ceph_decode_64_safe(&p, end, inline_version, bad);
3112 ceph_decode_32_safe(&p, end, inline_len, bad);
3113 if (p + inline_len > end)
3120 inode = ceph_find_inode(sb, vino);
3121 ci = ceph_inode(inode);
3122 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3125 mutex_lock(&session->s_mutex);
3127 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3130 if (op == CEPH_CAP_OP_IMPORT)
3131 ceph_add_cap_releases(mdsc, session);
3134 dout(" i don't have ino %llx\n", vino.ino);
3136 if (op == CEPH_CAP_OP_IMPORT) {
3137 spin_lock(&session->s_cap_lock);
3138 __queue_cap_release(session, vino.ino, cap_id,
3140 spin_unlock(&session->s_cap_lock);
3142 goto flush_cap_releases;
3145 /* these will work even if we don't have a cap yet */
3147 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3148 handle_cap_flushsnap_ack(inode, tid, h, session);
3151 case CEPH_CAP_OP_EXPORT:
3152 handle_cap_export(inode, h, peer, session);
3155 case CEPH_CAP_OP_IMPORT:
3156 handle_cap_import(mdsc, inode, h, peer, session,
3158 handle_cap_grant(mdsc, inode, h, snaptrace, snaptrace_len,
3159 inline_version, inline_data, inline_len,
3160 msg->middle, session, cap, issued);
3164 /* the rest require a cap */
3165 spin_lock(&ci->i_ceph_lock);
3166 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3168 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3169 inode, ceph_ino(inode), ceph_snap(inode), mds);
3170 spin_unlock(&ci->i_ceph_lock);
3171 goto flush_cap_releases;
3174 /* note that each of these drops i_ceph_lock for us */
3176 case CEPH_CAP_OP_REVOKE:
3177 case CEPH_CAP_OP_GRANT:
3178 __ceph_caps_issued(ci, &issued);
3179 issued |= __ceph_caps_dirty(ci);
3180 handle_cap_grant(mdsc, inode, h, NULL, 0,
3181 inline_version, inline_data, inline_len,
3182 msg->middle, session, cap, issued);
3185 case CEPH_CAP_OP_FLUSH_ACK:
3186 handle_cap_flush_ack(inode, tid, h, session, cap);
3189 case CEPH_CAP_OP_TRUNC:
3190 handle_cap_trunc(inode, h, session);
3194 spin_unlock(&ci->i_ceph_lock);
3195 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3196 ceph_cap_op_name(op));
3203 * send any full release message to try to move things
3204 * along for the mds (who clearly thinks we still have this
3207 ceph_add_cap_releases(mdsc, session);
3208 ceph_send_cap_releases(mdsc, session);
3211 mutex_unlock(&session->s_mutex);
3217 pr_err("ceph_handle_caps: corrupt message\n");
3223 * Delayed work handler to process end of delayed cap release LRU list.
3225 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3227 struct ceph_inode_info *ci;
3228 int flags = CHECK_CAPS_NODELAY;
3230 dout("check_delayed_caps\n");
3232 spin_lock(&mdsc->cap_delay_lock);
3233 if (list_empty(&mdsc->cap_delay_list))
3235 ci = list_first_entry(&mdsc->cap_delay_list,
3236 struct ceph_inode_info,
3238 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3239 time_before(jiffies, ci->i_hold_caps_max))
3241 list_del_init(&ci->i_cap_delay_list);
3242 spin_unlock(&mdsc->cap_delay_lock);
3243 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3244 ceph_check_caps(ci, flags, NULL);
3246 spin_unlock(&mdsc->cap_delay_lock);
3250 * Flush all dirty caps to the mds
3252 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3254 struct ceph_inode_info *ci;
3255 struct inode *inode;
3257 dout("flush_dirty_caps\n");
3258 spin_lock(&mdsc->cap_dirty_lock);
3259 while (!list_empty(&mdsc->cap_dirty)) {
3260 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3262 inode = &ci->vfs_inode;
3264 dout("flush_dirty_caps %p\n", inode);
3265 spin_unlock(&mdsc->cap_dirty_lock);
3266 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3268 spin_lock(&mdsc->cap_dirty_lock);
3270 spin_unlock(&mdsc->cap_dirty_lock);
3271 dout("flush_dirty_caps done\n");
3275 * Drop open file reference. If we were the last open file,
3276 * we may need to release capabilities to the MDS (or schedule
3277 * their delayed release).
3279 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3281 struct inode *inode = &ci->vfs_inode;
3284 spin_lock(&ci->i_ceph_lock);
3285 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
3286 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
3287 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
3288 if (--ci->i_nr_by_mode[fmode] == 0)
3290 spin_unlock(&ci->i_ceph_lock);
3292 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3293 ceph_check_caps(ci, 0, NULL);
3297 * Helpers for embedding cap and dentry lease releases into mds
3300 * @force is used by dentry_release (below) to force inclusion of a
3301 * record for the directory inode, even when there aren't any caps to
3304 int ceph_encode_inode_release(void **p, struct inode *inode,
3305 int mds, int drop, int unless, int force)
3307 struct ceph_inode_info *ci = ceph_inode(inode);
3308 struct ceph_cap *cap;
3309 struct ceph_mds_request_release *rel = *p;
3313 spin_lock(&ci->i_ceph_lock);
3314 used = __ceph_caps_used(ci);
3315 dirty = __ceph_caps_dirty(ci);
3317 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3318 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3319 ceph_cap_string(unless));
3321 /* only drop unused, clean caps */
3322 drop &= ~(used | dirty);
3324 cap = __get_cap_for_mds(ci, mds);
3325 if (cap && __cap_is_valid(cap)) {
3327 ((cap->issued & drop) &&
3328 (cap->issued & unless) == 0)) {
3329 if ((cap->issued & drop) &&
3330 (cap->issued & unless) == 0) {
3331 int wanted = __ceph_caps_wanted(ci);
3332 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3333 wanted |= cap->mds_wanted;
3334 dout("encode_inode_release %p cap %p "
3335 "%s -> %s, wanted %s -> %s\n", inode, cap,
3336 ceph_cap_string(cap->issued),
3337 ceph_cap_string(cap->issued & ~drop),
3338 ceph_cap_string(cap->mds_wanted),
3339 ceph_cap_string(wanted));
3341 cap->issued &= ~drop;
3342 cap->implemented &= ~drop;
3343 cap->mds_wanted = wanted;
3345 dout("encode_inode_release %p cap %p %s"
3346 " (force)\n", inode, cap,
3347 ceph_cap_string(cap->issued));
3350 rel->ino = cpu_to_le64(ceph_ino(inode));
3351 rel->cap_id = cpu_to_le64(cap->cap_id);
3352 rel->seq = cpu_to_le32(cap->seq);
3353 rel->issue_seq = cpu_to_le32(cap->issue_seq);
3354 rel->mseq = cpu_to_le32(cap->mseq);
3355 rel->caps = cpu_to_le32(cap->implemented);
3356 rel->wanted = cpu_to_le32(cap->mds_wanted);
3362 dout("encode_inode_release %p cap %p %s\n",
3363 inode, cap, ceph_cap_string(cap->issued));
3366 spin_unlock(&ci->i_ceph_lock);
3370 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3371 int mds, int drop, int unless)
3373 struct inode *dir = dentry->d_parent->d_inode;
3374 struct ceph_mds_request_release *rel = *p;
3375 struct ceph_dentry_info *di = ceph_dentry(dentry);
3380 * force an record for the directory caps if we have a dentry lease.
3381 * this is racy (can't take i_ceph_lock and d_lock together), but it
3382 * doesn't have to be perfect; the mds will revoke anything we don't
3385 spin_lock(&dentry->d_lock);
3386 if (di->lease_session && di->lease_session->s_mds == mds)
3388 spin_unlock(&dentry->d_lock);
3390 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3392 spin_lock(&dentry->d_lock);
3393 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3394 dout("encode_dentry_release %p mds%d seq %d\n",
3395 dentry, mds, (int)di->lease_seq);
3396 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3397 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3398 *p += dentry->d_name.len;
3399 rel->dname_seq = cpu_to_le32(di->lease_seq);
3400 __ceph_mdsc_drop_dentry_lease(dentry);
3402 spin_unlock(&dentry->d_lock);