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 spin_lock(&realm->inodes_with_caps_lock);
581 ci->i_snap_realm = realm;
582 list_add(&ci->i_snap_realm_item,
583 &realm->inodes_with_caps);
584 spin_unlock(&realm->inodes_with_caps_lock);
586 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
592 __check_cap_issue(ci, cap, issued);
595 * If we are issued caps we don't want, or the mds' wanted
596 * value appears to be off, queue a check so we'll release
597 * later and/or update the mds wanted value.
599 actual_wanted = __ceph_caps_wanted(ci);
600 if ((wanted & ~actual_wanted) ||
601 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
602 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
603 ceph_cap_string(issued), ceph_cap_string(wanted),
604 ceph_cap_string(actual_wanted));
605 __cap_delay_requeue(mdsc, ci);
608 if (flags & CEPH_CAP_FLAG_AUTH) {
609 if (ci->i_auth_cap == NULL ||
610 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
611 ci->i_auth_cap = cap;
612 cap->mds_wanted = wanted;
615 WARN_ON(ci->i_auth_cap == cap);
618 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
619 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
620 ceph_cap_string(issued|cap->issued), seq, mds);
621 cap->cap_id = cap_id;
622 cap->issued = issued;
623 cap->implemented |= issued;
624 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
625 cap->mds_wanted = wanted;
627 cap->mds_wanted |= wanted;
629 cap->issue_seq = seq;
631 cap->cap_gen = session->s_cap_gen;
634 __ceph_get_fmode(ci, fmode);
638 * Return true if cap has not timed out and belongs to the current
639 * generation of the MDS session (i.e. has not gone 'stale' due to
640 * us losing touch with the mds).
642 static int __cap_is_valid(struct ceph_cap *cap)
647 spin_lock(&cap->session->s_gen_ttl_lock);
648 gen = cap->session->s_cap_gen;
649 ttl = cap->session->s_cap_ttl;
650 spin_unlock(&cap->session->s_gen_ttl_lock);
652 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
653 dout("__cap_is_valid %p cap %p issued %s "
654 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
655 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
663 * Return set of valid cap bits issued to us. Note that caps time
664 * out, and may be invalidated in bulk if the client session times out
665 * and session->s_cap_gen is bumped.
667 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
669 int have = ci->i_snap_caps;
670 struct ceph_cap *cap;
675 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
676 cap = rb_entry(p, struct ceph_cap, ci_node);
677 if (!__cap_is_valid(cap))
679 dout("__ceph_caps_issued %p cap %p issued %s\n",
680 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
683 *implemented |= cap->implemented;
686 * exclude caps issued by non-auth MDS, but are been revoking
687 * by the auth MDS. The non-auth MDS should be revoking/exporting
688 * these caps, but the message is delayed.
690 if (ci->i_auth_cap) {
691 cap = ci->i_auth_cap;
692 have &= ~cap->implemented | cap->issued;
698 * Get cap bits issued by caps other than @ocap
700 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
702 int have = ci->i_snap_caps;
703 struct ceph_cap *cap;
706 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
707 cap = rb_entry(p, struct ceph_cap, ci_node);
710 if (!__cap_is_valid(cap))
718 * Move a cap to the end of the LRU (oldest caps at list head, newest
721 static void __touch_cap(struct ceph_cap *cap)
723 struct ceph_mds_session *s = cap->session;
725 spin_lock(&s->s_cap_lock);
726 if (s->s_cap_iterator == NULL) {
727 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
729 list_move_tail(&cap->session_caps, &s->s_caps);
731 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
732 &cap->ci->vfs_inode, cap, s->s_mds);
734 spin_unlock(&s->s_cap_lock);
738 * Check if we hold the given mask. If so, move the cap(s) to the
739 * front of their respective LRUs. (This is the preferred way for
740 * callers to check for caps they want.)
742 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
744 struct ceph_cap *cap;
746 int have = ci->i_snap_caps;
748 if ((have & mask) == mask) {
749 dout("__ceph_caps_issued_mask %p snap issued %s"
750 " (mask %s)\n", &ci->vfs_inode,
751 ceph_cap_string(have),
752 ceph_cap_string(mask));
756 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
757 cap = rb_entry(p, struct ceph_cap, ci_node);
758 if (!__cap_is_valid(cap))
760 if ((cap->issued & mask) == mask) {
761 dout("__ceph_caps_issued_mask %p cap %p issued %s"
762 " (mask %s)\n", &ci->vfs_inode, cap,
763 ceph_cap_string(cap->issued),
764 ceph_cap_string(mask));
770 /* does a combination of caps satisfy mask? */
772 if ((have & mask) == mask) {
773 dout("__ceph_caps_issued_mask %p combo issued %s"
774 " (mask %s)\n", &ci->vfs_inode,
775 ceph_cap_string(cap->issued),
776 ceph_cap_string(mask));
780 /* touch this + preceding caps */
782 for (q = rb_first(&ci->i_caps); q != p;
784 cap = rb_entry(q, struct ceph_cap,
786 if (!__cap_is_valid(cap))
799 * Return true if mask caps are currently being revoked by an MDS.
801 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
802 struct ceph_cap *ocap, int mask)
804 struct ceph_cap *cap;
807 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
808 cap = rb_entry(p, struct ceph_cap, ci_node);
810 (cap->implemented & ~cap->issued & mask))
816 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
818 struct inode *inode = &ci->vfs_inode;
821 spin_lock(&ci->i_ceph_lock);
822 ret = __ceph_caps_revoking_other(ci, NULL, mask);
823 spin_unlock(&ci->i_ceph_lock);
824 dout("ceph_caps_revoking %p %s = %d\n", inode,
825 ceph_cap_string(mask), ret);
829 int __ceph_caps_used(struct ceph_inode_info *ci)
833 used |= CEPH_CAP_PIN;
835 used |= CEPH_CAP_FILE_RD;
836 if (ci->i_rdcache_ref || ci->vfs_inode.i_data.nrpages)
837 used |= CEPH_CAP_FILE_CACHE;
839 used |= CEPH_CAP_FILE_WR;
840 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
841 used |= CEPH_CAP_FILE_BUFFER;
846 * wanted, by virtue of open file modes
848 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
852 for (mode = 0; mode < CEPH_FILE_MODE_NUM; mode++)
853 if (ci->i_nr_by_mode[mode])
854 want |= ceph_caps_for_mode(mode);
859 * Return caps we have registered with the MDS(s) as 'wanted'.
861 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
863 struct ceph_cap *cap;
867 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
868 cap = rb_entry(p, struct ceph_cap, ci_node);
869 if (!__cap_is_valid(cap))
871 if (cap == ci->i_auth_cap)
872 mds_wanted |= cap->mds_wanted;
874 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
880 * called under i_ceph_lock
882 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
884 return !RB_EMPTY_ROOT(&ci->i_caps);
887 int ceph_is_any_caps(struct inode *inode)
889 struct ceph_inode_info *ci = ceph_inode(inode);
892 spin_lock(&ci->i_ceph_lock);
893 ret = __ceph_is_any_caps(ci);
894 spin_unlock(&ci->i_ceph_lock);
900 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
902 * caller should hold i_ceph_lock.
903 * caller will not hold session s_mutex if called from destroy_inode.
905 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
907 struct ceph_mds_session *session = cap->session;
908 struct ceph_inode_info *ci = cap->ci;
909 struct ceph_mds_client *mdsc =
910 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
913 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
915 /* remove from session list */
916 spin_lock(&session->s_cap_lock);
918 * s_cap_reconnect is protected by s_cap_lock. no one changes
919 * s_cap_gen while session is in the reconnect state.
922 (!session->s_cap_reconnect ||
923 cap->cap_gen == session->s_cap_gen))
924 __queue_cap_release(session, ci->i_vino.ino, cap->cap_id,
925 cap->mseq, cap->issue_seq);
927 if (session->s_cap_iterator == cap) {
928 /* not yet, we are iterating over this very cap */
929 dout("__ceph_remove_cap delaying %p removal from session %p\n",
932 list_del_init(&cap->session_caps);
933 session->s_nr_caps--;
937 /* protect backpointer with s_cap_lock: see iterate_session_caps */
939 spin_unlock(&session->s_cap_lock);
941 /* remove from inode list */
942 rb_erase(&cap->ci_node, &ci->i_caps);
943 if (ci->i_auth_cap == cap)
944 ci->i_auth_cap = NULL;
947 ceph_put_cap(mdsc, cap);
949 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
950 struct ceph_snap_realm *realm = ci->i_snap_realm;
951 spin_lock(&realm->inodes_with_caps_lock);
952 list_del_init(&ci->i_snap_realm_item);
953 ci->i_snap_realm_counter++;
954 ci->i_snap_realm = NULL;
955 spin_unlock(&realm->inodes_with_caps_lock);
956 ceph_put_snap_realm(mdsc, realm);
958 if (!__ceph_is_any_real_caps(ci))
959 __cap_delay_cancel(mdsc, ci);
963 * Build and send a cap message to the given MDS.
965 * Caller should be holding s_mutex.
967 static int send_cap_msg(struct ceph_mds_session *session,
968 u64 ino, u64 cid, int op,
969 int caps, int wanted, int dirty,
970 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
971 u64 size, u64 max_size,
972 struct timespec *mtime, struct timespec *atime,
974 kuid_t uid, kgid_t gid, umode_t mode,
976 struct ceph_buffer *xattrs_buf,
977 u64 follows, bool inline_data)
979 struct ceph_mds_caps *fc;
980 struct ceph_msg *msg;
984 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
985 " seq %u/%u mseq %u follows %lld size %llu/%llu"
986 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
987 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
988 ceph_cap_string(dirty),
989 seq, issue_seq, mseq, follows, size, max_size,
990 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
992 /* flock buffer size + inline version + inline data size */
993 extra_len = 4 + 8 + 4;
994 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
999 msg->hdr.tid = cpu_to_le64(flush_tid);
1001 fc = msg->front.iov_base;
1002 memset(fc, 0, sizeof(*fc));
1004 fc->cap_id = cpu_to_le64(cid);
1005 fc->op = cpu_to_le32(op);
1006 fc->seq = cpu_to_le32(seq);
1007 fc->issue_seq = cpu_to_le32(issue_seq);
1008 fc->migrate_seq = cpu_to_le32(mseq);
1009 fc->caps = cpu_to_le32(caps);
1010 fc->wanted = cpu_to_le32(wanted);
1011 fc->dirty = cpu_to_le32(dirty);
1012 fc->ino = cpu_to_le64(ino);
1013 fc->snap_follows = cpu_to_le64(follows);
1015 fc->size = cpu_to_le64(size);
1016 fc->max_size = cpu_to_le64(max_size);
1018 ceph_encode_timespec(&fc->mtime, mtime);
1020 ceph_encode_timespec(&fc->atime, atime);
1021 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
1023 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid));
1024 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid));
1025 fc->mode = cpu_to_le32(mode);
1028 /* flock buffer size */
1029 ceph_encode_32(&p, 0);
1030 /* inline version */
1031 ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE);
1032 /* inline data size */
1033 ceph_encode_32(&p, 0);
1035 fc->xattr_version = cpu_to_le64(xattr_version);
1037 msg->middle = ceph_buffer_get(xattrs_buf);
1038 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1039 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1042 ceph_con_send(&session->s_con, msg);
1046 void __queue_cap_release(struct ceph_mds_session *session,
1047 u64 ino, u64 cap_id, u32 migrate_seq,
1050 struct ceph_msg *msg;
1051 struct ceph_mds_cap_release *head;
1052 struct ceph_mds_cap_item *item;
1054 BUG_ON(!session->s_num_cap_releases);
1055 msg = list_first_entry(&session->s_cap_releases,
1056 struct ceph_msg, list_head);
1058 dout(" adding %llx release to mds%d msg %p (%d left)\n",
1059 ino, session->s_mds, msg, session->s_num_cap_releases);
1061 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1062 head = msg->front.iov_base;
1063 le32_add_cpu(&head->num, 1);
1064 item = msg->front.iov_base + msg->front.iov_len;
1065 item->ino = cpu_to_le64(ino);
1066 item->cap_id = cpu_to_le64(cap_id);
1067 item->migrate_seq = cpu_to_le32(migrate_seq);
1068 item->seq = cpu_to_le32(issue_seq);
1070 session->s_num_cap_releases--;
1072 msg->front.iov_len += sizeof(*item);
1073 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1074 dout(" release msg %p full\n", msg);
1075 list_move_tail(&msg->list_head, &session->s_cap_releases_done);
1077 dout(" release msg %p at %d/%d (%d)\n", msg,
1078 (int)le32_to_cpu(head->num),
1079 (int)CEPH_CAPS_PER_RELEASE,
1080 (int)msg->front.iov_len);
1085 * Queue cap releases when an inode is dropped from our cache. Since
1086 * inode is about to be destroyed, there is no need for i_ceph_lock.
1088 void ceph_queue_caps_release(struct inode *inode)
1090 struct ceph_inode_info *ci = ceph_inode(inode);
1093 p = rb_first(&ci->i_caps);
1095 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1097 __ceph_remove_cap(cap, true);
1102 * Send a cap msg on the given inode. Update our caps state, then
1103 * drop i_ceph_lock and send the message.
1105 * Make note of max_size reported/requested from mds, revoked caps
1106 * that have now been implemented.
1108 * Make half-hearted attempt ot to invalidate page cache if we are
1109 * dropping RDCACHE. Note that this will leave behind locked pages
1110 * that we'll then need to deal with elsewhere.
1112 * Return non-zero if delayed release, or we experienced an error
1113 * such that the caller should requeue + retry later.
1115 * called with i_ceph_lock, then drops it.
1116 * caller should hold snap_rwsem (read), s_mutex.
1118 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1119 int op, int used, int want, int retain, int flushing,
1120 unsigned *pflush_tid)
1121 __releases(cap->ci->i_ceph_lock)
1123 struct ceph_inode_info *ci = cap->ci;
1124 struct inode *inode = &ci->vfs_inode;
1125 u64 cap_id = cap->cap_id;
1126 int held, revoking, dropping, keep;
1127 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1129 struct timespec mtime, atime;
1134 struct ceph_mds_session *session;
1135 u64 xattr_version = 0;
1136 struct ceph_buffer *xattr_blob = NULL;
1143 held = cap->issued | cap->implemented;
1144 revoking = cap->implemented & ~cap->issued;
1145 retain &= ~revoking;
1146 dropping = cap->issued & ~retain;
1148 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1149 inode, cap, cap->session,
1150 ceph_cap_string(held), ceph_cap_string(held & retain),
1151 ceph_cap_string(revoking));
1152 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1154 session = cap->session;
1156 /* don't release wanted unless we've waited a bit. */
1157 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1158 time_before(jiffies, ci->i_hold_caps_min)) {
1159 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1160 ceph_cap_string(cap->issued),
1161 ceph_cap_string(cap->issued & retain),
1162 ceph_cap_string(cap->mds_wanted),
1163 ceph_cap_string(want));
1164 want |= cap->mds_wanted;
1165 retain |= cap->issued;
1168 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1170 cap->issued &= retain; /* drop bits we don't want */
1171 if (cap->implemented & ~cap->issued) {
1173 * Wake up any waiters on wanted -> needed transition.
1174 * This is due to the weird transition from buffered
1175 * to sync IO... we need to flush dirty pages _before_
1176 * allowing sync writes to avoid reordering.
1180 cap->implemented &= cap->issued | used;
1181 cap->mds_wanted = want;
1185 * assign a tid for flush operations so we can avoid
1186 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1187 * clean type races. track latest tid for every bit
1188 * so we can handle flush AxFw, flush Fw, and have the
1189 * first ack clean Ax.
1191 flush_tid = ++ci->i_cap_flush_last_tid;
1193 *pflush_tid = flush_tid;
1194 dout(" cap_flush_tid %d\n", (int)flush_tid);
1195 for (i = 0; i < CEPH_CAP_BITS; i++)
1196 if (flushing & (1 << i))
1197 ci->i_cap_flush_tid[i] = flush_tid;
1199 follows = ci->i_head_snapc->seq;
1204 keep = cap->implemented;
1206 issue_seq = cap->issue_seq;
1208 size = inode->i_size;
1209 ci->i_reported_size = size;
1210 max_size = ci->i_wanted_max_size;
1211 ci->i_requested_max_size = max_size;
1212 mtime = inode->i_mtime;
1213 atime = inode->i_atime;
1214 time_warp_seq = ci->i_time_warp_seq;
1217 mode = inode->i_mode;
1219 if (flushing & CEPH_CAP_XATTR_EXCL) {
1220 __ceph_build_xattrs_blob(ci);
1221 xattr_blob = ci->i_xattrs.blob;
1222 xattr_version = ci->i_xattrs.version;
1225 inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1227 spin_unlock(&ci->i_ceph_lock);
1229 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1230 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1231 size, max_size, &mtime, &atime, time_warp_seq,
1232 uid, gid, mode, xattr_version, xattr_blob,
1233 follows, inline_data);
1235 dout("error sending cap msg, must requeue %p\n", inode);
1240 wake_up_all(&ci->i_cap_wq);
1246 * When a snapshot is taken, clients accumulate dirty metadata on
1247 * inodes with capabilities in ceph_cap_snaps to describe the file
1248 * state at the time the snapshot was taken. This must be flushed
1249 * asynchronously back to the MDS once sync writes complete and dirty
1250 * data is written out.
1252 * Unless @again is true, skip cap_snaps that were already sent to
1253 * the MDS (i.e., during this session).
1255 * Called under i_ceph_lock. Takes s_mutex as needed.
1257 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1258 struct ceph_mds_session **psession,
1260 __releases(ci->i_ceph_lock)
1261 __acquires(ci->i_ceph_lock)
1263 struct inode *inode = &ci->vfs_inode;
1265 struct ceph_cap_snap *capsnap;
1267 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1268 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1270 u64 next_follows = 0; /* keep track of how far we've gotten through the
1271 i_cap_snaps list, and skip these entries next time
1272 around to avoid an infinite loop */
1275 session = *psession;
1277 dout("__flush_snaps %p\n", inode);
1279 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1280 /* avoid an infiniute loop after retry */
1281 if (capsnap->follows < next_follows)
1284 * we need to wait for sync writes to complete and for dirty
1285 * pages to be written out.
1287 if (capsnap->dirty_pages || capsnap->writing)
1291 * if cap writeback already occurred, we should have dropped
1292 * the capsnap in ceph_put_wrbuffer_cap_refs.
1294 BUG_ON(capsnap->dirty == 0);
1296 /* pick mds, take s_mutex */
1297 if (ci->i_auth_cap == NULL) {
1298 dout("no auth cap (migrating?), doing nothing\n");
1302 /* only flush each capsnap once */
1303 if (!again && !list_empty(&capsnap->flushing_item)) {
1304 dout("already flushed %p, skipping\n", capsnap);
1308 mds = ci->i_auth_cap->session->s_mds;
1309 mseq = ci->i_auth_cap->mseq;
1311 if (session && session->s_mds != mds) {
1312 dout("oops, wrong session %p mutex\n", session);
1313 mutex_unlock(&session->s_mutex);
1314 ceph_put_mds_session(session);
1318 spin_unlock(&ci->i_ceph_lock);
1319 mutex_lock(&mdsc->mutex);
1320 session = __ceph_lookup_mds_session(mdsc, mds);
1321 mutex_unlock(&mdsc->mutex);
1323 dout("inverting session/ino locks on %p\n",
1325 mutex_lock(&session->s_mutex);
1328 * if session == NULL, we raced against a cap
1329 * deletion or migration. retry, and we'll
1330 * get a better @mds value next time.
1332 spin_lock(&ci->i_ceph_lock);
1336 capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1337 atomic_inc(&capsnap->nref);
1338 if (!list_empty(&capsnap->flushing_item))
1339 list_del_init(&capsnap->flushing_item);
1340 list_add_tail(&capsnap->flushing_item,
1341 &session->s_cap_snaps_flushing);
1342 spin_unlock(&ci->i_ceph_lock);
1344 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1345 inode, capsnap, capsnap->follows, capsnap->flush_tid);
1346 send_cap_msg(session, ceph_vino(inode).ino, 0,
1347 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1348 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1350 &capsnap->mtime, &capsnap->atime,
1351 capsnap->time_warp_seq,
1352 capsnap->uid, capsnap->gid, capsnap->mode,
1353 capsnap->xattr_version, capsnap->xattr_blob,
1354 capsnap->follows, capsnap->inline_data);
1356 next_follows = capsnap->follows + 1;
1357 ceph_put_cap_snap(capsnap);
1359 spin_lock(&ci->i_ceph_lock);
1363 /* we flushed them all; remove this inode from the queue */
1364 spin_lock(&mdsc->snap_flush_lock);
1365 list_del_init(&ci->i_snap_flush_item);
1366 spin_unlock(&mdsc->snap_flush_lock);
1370 *psession = session;
1372 mutex_unlock(&session->s_mutex);
1373 ceph_put_mds_session(session);
1377 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1379 spin_lock(&ci->i_ceph_lock);
1380 __ceph_flush_snaps(ci, NULL, 0);
1381 spin_unlock(&ci->i_ceph_lock);
1385 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1386 * Caller is then responsible for calling __mark_inode_dirty with the
1387 * returned flags value.
1389 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1391 struct ceph_mds_client *mdsc =
1392 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1393 struct inode *inode = &ci->vfs_inode;
1394 int was = ci->i_dirty_caps;
1397 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1398 ceph_cap_string(mask), ceph_cap_string(was),
1399 ceph_cap_string(was | mask));
1400 ci->i_dirty_caps |= mask;
1402 if (!ci->i_head_snapc)
1403 ci->i_head_snapc = ceph_get_snap_context(
1404 ci->i_snap_realm->cached_context);
1405 dout(" inode %p now dirty snapc %p auth cap %p\n",
1406 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1407 WARN_ON(!ci->i_auth_cap);
1408 BUG_ON(!list_empty(&ci->i_dirty_item));
1409 spin_lock(&mdsc->cap_dirty_lock);
1410 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1411 spin_unlock(&mdsc->cap_dirty_lock);
1412 if (ci->i_flushing_caps == 0) {
1414 dirty |= I_DIRTY_SYNC;
1417 BUG_ON(list_empty(&ci->i_dirty_item));
1418 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1419 (mask & CEPH_CAP_FILE_BUFFER))
1420 dirty |= I_DIRTY_DATASYNC;
1421 __cap_delay_requeue(mdsc, ci);
1426 * Add dirty inode to the flushing list. Assigned a seq number so we
1427 * can wait for caps to flush without starving.
1429 * Called under i_ceph_lock.
1431 static int __mark_caps_flushing(struct inode *inode,
1432 struct ceph_mds_session *session)
1434 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1435 struct ceph_inode_info *ci = ceph_inode(inode);
1438 BUG_ON(ci->i_dirty_caps == 0);
1439 BUG_ON(list_empty(&ci->i_dirty_item));
1441 flushing = ci->i_dirty_caps;
1442 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1443 ceph_cap_string(flushing),
1444 ceph_cap_string(ci->i_flushing_caps),
1445 ceph_cap_string(ci->i_flushing_caps | flushing));
1446 ci->i_flushing_caps |= flushing;
1447 ci->i_dirty_caps = 0;
1448 dout(" inode %p now !dirty\n", inode);
1450 spin_lock(&mdsc->cap_dirty_lock);
1451 list_del_init(&ci->i_dirty_item);
1453 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1454 if (list_empty(&ci->i_flushing_item)) {
1455 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1456 mdsc->num_cap_flushing++;
1457 dout(" inode %p now flushing seq %lld\n", inode,
1458 ci->i_cap_flush_seq);
1460 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1461 dout(" inode %p now flushing (more) seq %lld\n", inode,
1462 ci->i_cap_flush_seq);
1464 spin_unlock(&mdsc->cap_dirty_lock);
1470 * try to invalidate mapping pages without blocking.
1472 static int try_nonblocking_invalidate(struct inode *inode)
1474 struct ceph_inode_info *ci = ceph_inode(inode);
1475 u32 invalidating_gen = ci->i_rdcache_gen;
1477 spin_unlock(&ci->i_ceph_lock);
1478 invalidate_mapping_pages(&inode->i_data, 0, -1);
1479 spin_lock(&ci->i_ceph_lock);
1481 if (inode->i_data.nrpages == 0 &&
1482 invalidating_gen == ci->i_rdcache_gen) {
1484 dout("try_nonblocking_invalidate %p success\n", inode);
1485 /* save any racing async invalidate some trouble */
1486 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1489 dout("try_nonblocking_invalidate %p failed\n", inode);
1494 * Swiss army knife function to examine currently used and wanted
1495 * versus held caps. Release, flush, ack revoked caps to mds as
1498 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1499 * cap release further.
1500 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1501 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1504 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1505 struct ceph_mds_session *session)
1507 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1508 struct ceph_mds_client *mdsc = fsc->mdsc;
1509 struct inode *inode = &ci->vfs_inode;
1510 struct ceph_cap *cap;
1511 int file_wanted, used, cap_used;
1512 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1513 int issued, implemented, want, retain, revoking, flushing = 0;
1514 int mds = -1; /* keep track of how far we've gone through i_caps list
1515 to avoid an infinite loop on retry */
1517 int tried_invalidate = 0;
1518 int delayed = 0, sent = 0, force_requeue = 0, num;
1519 int queue_invalidate = 0;
1520 int is_delayed = flags & CHECK_CAPS_NODELAY;
1522 /* if we are unmounting, flush any unused caps immediately. */
1526 spin_lock(&ci->i_ceph_lock);
1528 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1529 flags |= CHECK_CAPS_FLUSH;
1531 /* flush snaps first time around only */
1532 if (!list_empty(&ci->i_cap_snaps))
1533 __ceph_flush_snaps(ci, &session, 0);
1536 spin_lock(&ci->i_ceph_lock);
1538 file_wanted = __ceph_caps_file_wanted(ci);
1539 used = __ceph_caps_used(ci);
1540 want = file_wanted | used;
1541 issued = __ceph_caps_issued(ci, &implemented);
1542 revoking = implemented & ~issued;
1544 retain = want | CEPH_CAP_PIN;
1545 if (!mdsc->stopping && inode->i_nlink > 0) {
1547 retain |= CEPH_CAP_ANY; /* be greedy */
1549 retain |= CEPH_CAP_ANY_SHARED;
1551 * keep RD only if we didn't have the file open RW,
1552 * because then the mds would revoke it anyway to
1553 * journal max_size=0.
1555 if (ci->i_max_size == 0)
1556 retain |= CEPH_CAP_ANY_RD;
1560 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1561 " issued %s revoking %s retain %s %s%s%s\n", inode,
1562 ceph_cap_string(file_wanted),
1563 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1564 ceph_cap_string(ci->i_flushing_caps),
1565 ceph_cap_string(issued), ceph_cap_string(revoking),
1566 ceph_cap_string(retain),
1567 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1568 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1569 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1572 * If we no longer need to hold onto old our caps, and we may
1573 * have cached pages, but don't want them, then try to invalidate.
1574 * If we fail, it's because pages are locked.... try again later.
1576 if ((!is_delayed || mdsc->stopping) &&
1577 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1578 inode->i_data.nrpages && /* have cached pages */
1579 (file_wanted == 0 || /* no open files */
1580 (revoking & (CEPH_CAP_FILE_CACHE|
1581 CEPH_CAP_FILE_LAZYIO))) && /* or revoking cache */
1582 !tried_invalidate) {
1583 dout("check_caps trying to invalidate on %p\n", inode);
1584 if (try_nonblocking_invalidate(inode) < 0) {
1585 if (revoking & (CEPH_CAP_FILE_CACHE|
1586 CEPH_CAP_FILE_LAZYIO)) {
1587 dout("check_caps queuing invalidate\n");
1588 queue_invalidate = 1;
1589 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1591 dout("check_caps failed to invalidate pages\n");
1592 /* we failed to invalidate pages. check these
1593 caps again later. */
1595 __cap_set_timeouts(mdsc, ci);
1598 tried_invalidate = 1;
1603 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1604 cap = rb_entry(p, struct ceph_cap, ci_node);
1607 /* avoid looping forever */
1608 if (mds >= cap->mds ||
1609 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1612 /* NOTE: no side-effects allowed, until we take s_mutex */
1615 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1616 cap_used &= ~ci->i_auth_cap->issued;
1618 revoking = cap->implemented & ~cap->issued;
1619 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1620 cap->mds, cap, ceph_cap_string(cap->issued),
1621 ceph_cap_string(cap_used),
1622 ceph_cap_string(cap->implemented),
1623 ceph_cap_string(revoking));
1625 if (cap == ci->i_auth_cap &&
1626 (cap->issued & CEPH_CAP_FILE_WR)) {
1627 /* request larger max_size from MDS? */
1628 if (ci->i_wanted_max_size > ci->i_max_size &&
1629 ci->i_wanted_max_size > ci->i_requested_max_size) {
1630 dout("requesting new max_size\n");
1634 /* approaching file_max? */
1635 if ((inode->i_size << 1) >= ci->i_max_size &&
1636 (ci->i_reported_size << 1) < ci->i_max_size) {
1637 dout("i_size approaching max_size\n");
1641 /* flush anything dirty? */
1642 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1644 dout("flushing dirty caps\n");
1648 /* completed revocation? going down and there are no caps? */
1649 if (revoking && (revoking & cap_used) == 0) {
1650 dout("completed revocation of %s\n",
1651 ceph_cap_string(cap->implemented & ~cap->issued));
1655 /* want more caps from mds? */
1656 if (want & ~(cap->mds_wanted | cap->issued))
1659 /* things we might delay */
1660 if ((cap->issued & ~retain) == 0 &&
1661 cap->mds_wanted == want)
1662 continue; /* nope, all good */
1668 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1669 time_before(jiffies, ci->i_hold_caps_max)) {
1670 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1671 ceph_cap_string(cap->issued),
1672 ceph_cap_string(cap->issued & retain),
1673 ceph_cap_string(cap->mds_wanted),
1674 ceph_cap_string(want));
1680 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1681 dout(" skipping %p I_NOFLUSH set\n", inode);
1685 if (session && session != cap->session) {
1686 dout("oops, wrong session %p mutex\n", session);
1687 mutex_unlock(&session->s_mutex);
1691 session = cap->session;
1692 if (mutex_trylock(&session->s_mutex) == 0) {
1693 dout("inverting session/ino locks on %p\n",
1695 spin_unlock(&ci->i_ceph_lock);
1696 if (took_snap_rwsem) {
1697 up_read(&mdsc->snap_rwsem);
1698 took_snap_rwsem = 0;
1700 mutex_lock(&session->s_mutex);
1704 /* take snap_rwsem after session mutex */
1705 if (!took_snap_rwsem) {
1706 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1707 dout("inverting snap/in locks on %p\n",
1709 spin_unlock(&ci->i_ceph_lock);
1710 down_read(&mdsc->snap_rwsem);
1711 took_snap_rwsem = 1;
1714 took_snap_rwsem = 1;
1717 if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1718 flushing = __mark_caps_flushing(inode, session);
1722 mds = cap->mds; /* remember mds, so we don't repeat */
1725 /* __send_cap drops i_ceph_lock */
1726 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1727 want, retain, flushing, NULL);
1728 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1732 * Reschedule delayed caps release if we delayed anything,
1735 if (delayed && is_delayed)
1736 force_requeue = 1; /* __send_cap delayed release; requeue */
1737 if (!delayed && !is_delayed)
1738 __cap_delay_cancel(mdsc, ci);
1739 else if (!is_delayed || force_requeue)
1740 __cap_delay_requeue(mdsc, ci);
1742 spin_unlock(&ci->i_ceph_lock);
1744 if (queue_invalidate)
1745 ceph_queue_invalidate(inode);
1748 mutex_unlock(&session->s_mutex);
1749 if (took_snap_rwsem)
1750 up_read(&mdsc->snap_rwsem);
1754 * Try to flush dirty caps back to the auth mds.
1756 static int try_flush_caps(struct inode *inode, unsigned *flush_tid)
1758 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1759 struct ceph_inode_info *ci = ceph_inode(inode);
1761 struct ceph_mds_session *session = NULL;
1764 spin_lock(&ci->i_ceph_lock);
1765 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1766 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1769 if (ci->i_dirty_caps && ci->i_auth_cap) {
1770 struct ceph_cap *cap = ci->i_auth_cap;
1771 int used = __ceph_caps_used(ci);
1772 int want = __ceph_caps_wanted(ci);
1775 if (!session || session != cap->session) {
1776 spin_unlock(&ci->i_ceph_lock);
1778 mutex_unlock(&session->s_mutex);
1779 session = cap->session;
1780 mutex_lock(&session->s_mutex);
1783 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1786 flushing = __mark_caps_flushing(inode, session);
1788 /* __send_cap drops i_ceph_lock */
1789 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1790 cap->issued | cap->implemented, flushing,
1795 spin_lock(&ci->i_ceph_lock);
1796 __cap_delay_requeue(mdsc, ci);
1799 spin_unlock(&ci->i_ceph_lock);
1802 mutex_unlock(&session->s_mutex);
1807 * Return true if we've flushed caps through the given flush_tid.
1809 static int caps_are_flushed(struct inode *inode, unsigned tid)
1811 struct ceph_inode_info *ci = ceph_inode(inode);
1814 spin_lock(&ci->i_ceph_lock);
1815 for (i = 0; i < CEPH_CAP_BITS; i++)
1816 if ((ci->i_flushing_caps & (1 << i)) &&
1817 ci->i_cap_flush_tid[i] <= tid) {
1818 /* still flushing this bit */
1822 spin_unlock(&ci->i_ceph_lock);
1827 * Wait on any unsafe replies for the given inode. First wait on the
1828 * newest request, and make that the upper bound. Then, if there are
1829 * more requests, keep waiting on the oldest as long as it is still older
1830 * than the original request.
1832 static void sync_write_wait(struct inode *inode)
1834 struct ceph_inode_info *ci = ceph_inode(inode);
1835 struct list_head *head = &ci->i_unsafe_writes;
1836 struct ceph_osd_request *req;
1839 spin_lock(&ci->i_unsafe_lock);
1840 if (list_empty(head))
1843 /* set upper bound as _last_ entry in chain */
1844 req = list_entry(head->prev, struct ceph_osd_request,
1846 last_tid = req->r_tid;
1849 ceph_osdc_get_request(req);
1850 spin_unlock(&ci->i_unsafe_lock);
1851 dout("sync_write_wait on tid %llu (until %llu)\n",
1852 req->r_tid, last_tid);
1853 wait_for_completion(&req->r_safe_completion);
1854 spin_lock(&ci->i_unsafe_lock);
1855 ceph_osdc_put_request(req);
1858 * from here on look at first entry in chain, since we
1859 * only want to wait for anything older than last_tid
1861 if (list_empty(head))
1863 req = list_entry(head->next, struct ceph_osd_request,
1865 } while (req->r_tid < last_tid);
1867 spin_unlock(&ci->i_unsafe_lock);
1870 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1872 struct inode *inode = file->f_mapping->host;
1873 struct ceph_inode_info *ci = ceph_inode(inode);
1878 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1879 sync_write_wait(inode);
1881 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1884 mutex_lock(&inode->i_mutex);
1886 dirty = try_flush_caps(inode, &flush_tid);
1887 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1890 * only wait on non-file metadata writeback (the mds
1891 * can recover size and mtime, so we don't need to
1894 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1895 dout("fsync waiting for flush_tid %u\n", flush_tid);
1896 ret = wait_event_interruptible(ci->i_cap_wq,
1897 caps_are_flushed(inode, flush_tid));
1900 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1901 mutex_unlock(&inode->i_mutex);
1906 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1907 * queue inode for flush but don't do so immediately, because we can
1908 * get by with fewer MDS messages if we wait for data writeback to
1911 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1913 struct ceph_inode_info *ci = ceph_inode(inode);
1917 int wait = wbc->sync_mode == WB_SYNC_ALL;
1919 dout("write_inode %p wait=%d\n", inode, wait);
1921 dirty = try_flush_caps(inode, &flush_tid);
1923 err = wait_event_interruptible(ci->i_cap_wq,
1924 caps_are_flushed(inode, flush_tid));
1926 struct ceph_mds_client *mdsc =
1927 ceph_sb_to_client(inode->i_sb)->mdsc;
1929 spin_lock(&ci->i_ceph_lock);
1930 if (__ceph_caps_dirty(ci))
1931 __cap_delay_requeue_front(mdsc, ci);
1932 spin_unlock(&ci->i_ceph_lock);
1938 * After a recovering MDS goes active, we need to resend any caps
1941 * Caller holds session->s_mutex.
1943 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1944 struct ceph_mds_session *session)
1946 struct ceph_cap_snap *capsnap;
1948 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1949 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1951 struct ceph_inode_info *ci = capsnap->ci;
1952 struct inode *inode = &ci->vfs_inode;
1953 struct ceph_cap *cap;
1955 spin_lock(&ci->i_ceph_lock);
1956 cap = ci->i_auth_cap;
1957 if (cap && cap->session == session) {
1958 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1960 __ceph_flush_snaps(ci, &session, 1);
1962 pr_err("%p auth cap %p not mds%d ???\n", inode,
1963 cap, session->s_mds);
1965 spin_unlock(&ci->i_ceph_lock);
1969 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1970 struct ceph_mds_session *session)
1972 struct ceph_inode_info *ci;
1974 kick_flushing_capsnaps(mdsc, session);
1976 dout("kick_flushing_caps mds%d\n", session->s_mds);
1977 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1978 struct inode *inode = &ci->vfs_inode;
1979 struct ceph_cap *cap;
1982 spin_lock(&ci->i_ceph_lock);
1983 cap = ci->i_auth_cap;
1984 if (cap && cap->session == session) {
1985 dout("kick_flushing_caps %p cap %p %s\n", inode,
1986 cap, ceph_cap_string(ci->i_flushing_caps));
1987 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1988 __ceph_caps_used(ci),
1989 __ceph_caps_wanted(ci),
1990 cap->issued | cap->implemented,
1991 ci->i_flushing_caps, NULL);
1993 spin_lock(&ci->i_ceph_lock);
1994 __cap_delay_requeue(mdsc, ci);
1995 spin_unlock(&ci->i_ceph_lock);
1998 pr_err("%p auth cap %p not mds%d ???\n", inode,
1999 cap, session->s_mds);
2000 spin_unlock(&ci->i_ceph_lock);
2005 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2006 struct ceph_mds_session *session,
2007 struct inode *inode)
2009 struct ceph_inode_info *ci = ceph_inode(inode);
2010 struct ceph_cap *cap;
2013 spin_lock(&ci->i_ceph_lock);
2014 cap = ci->i_auth_cap;
2015 dout("kick_flushing_inode_caps %p flushing %s flush_seq %lld\n", inode,
2016 ceph_cap_string(ci->i_flushing_caps), ci->i_cap_flush_seq);
2018 __ceph_flush_snaps(ci, &session, 1);
2020 if (ci->i_flushing_caps) {
2021 spin_lock(&mdsc->cap_dirty_lock);
2022 list_move_tail(&ci->i_flushing_item,
2023 &cap->session->s_cap_flushing);
2024 spin_unlock(&mdsc->cap_dirty_lock);
2026 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2027 __ceph_caps_used(ci),
2028 __ceph_caps_wanted(ci),
2029 cap->issued | cap->implemented,
2030 ci->i_flushing_caps, NULL);
2032 spin_lock(&ci->i_ceph_lock);
2033 __cap_delay_requeue(mdsc, ci);
2034 spin_unlock(&ci->i_ceph_lock);
2037 spin_unlock(&ci->i_ceph_lock);
2043 * Take references to capabilities we hold, so that we don't release
2044 * them to the MDS prematurely.
2046 * Protected by i_ceph_lock.
2048 static void __take_cap_refs(struct ceph_inode_info *ci, int got)
2050 if (got & CEPH_CAP_PIN)
2052 if (got & CEPH_CAP_FILE_RD)
2054 if (got & CEPH_CAP_FILE_CACHE)
2055 ci->i_rdcache_ref++;
2056 if (got & CEPH_CAP_FILE_WR)
2058 if (got & CEPH_CAP_FILE_BUFFER) {
2059 if (ci->i_wb_ref == 0)
2060 ihold(&ci->vfs_inode);
2062 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2063 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2068 * Try to grab cap references. Specify those refs we @want, and the
2069 * minimal set we @need. Also include the larger offset we are writing
2070 * to (when applicable), and check against max_size here as well.
2071 * Note that caller is responsible for ensuring max_size increases are
2072 * requested from the MDS.
2074 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2075 loff_t endoff, int *got, struct page **pinned_page,
2076 int *check_max, int *err)
2078 struct inode *inode = &ci->vfs_inode;
2080 int have, implemented, _got = 0;
2083 dout("get_cap_refs %p need %s want %s\n", inode,
2084 ceph_cap_string(need), ceph_cap_string(want));
2086 spin_lock(&ci->i_ceph_lock);
2088 /* make sure file is actually open */
2089 file_wanted = __ceph_caps_file_wanted(ci);
2090 if ((file_wanted & need) == 0) {
2091 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2092 ceph_cap_string(need), ceph_cap_string(file_wanted));
2098 /* finish pending truncate */
2099 while (ci->i_truncate_pending) {
2100 spin_unlock(&ci->i_ceph_lock);
2101 __ceph_do_pending_vmtruncate(inode);
2102 spin_lock(&ci->i_ceph_lock);
2105 have = __ceph_caps_issued(ci, &implemented);
2107 if (have & need & CEPH_CAP_FILE_WR) {
2108 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2109 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2110 inode, endoff, ci->i_max_size);
2111 if (endoff > ci->i_requested_max_size) {
2118 * If a sync write is in progress, we must wait, so that we
2119 * can get a final snapshot value for size+mtime.
2121 if (__ceph_have_pending_cap_snap(ci)) {
2122 dout("get_cap_refs %p cap_snap_pending\n", inode);
2127 if ((have & need) == need) {
2129 * Look at (implemented & ~have & not) so that we keep waiting
2130 * on transition from wanted -> needed caps. This is needed
2131 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2132 * going before a prior buffered writeback happens.
2134 int not = want & ~(have & need);
2135 int revoking = implemented & ~have;
2136 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2137 inode, ceph_cap_string(have), ceph_cap_string(not),
2138 ceph_cap_string(revoking));
2139 if ((revoking & not) == 0) {
2140 _got = need | (have & want);
2141 __take_cap_refs(ci, _got);
2145 int session_readonly = false;
2146 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2147 struct ceph_mds_session *s = ci->i_auth_cap->session;
2148 spin_lock(&s->s_cap_lock);
2149 session_readonly = s->s_readonly;
2150 spin_unlock(&s->s_cap_lock);
2152 if (session_readonly) {
2153 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2154 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2160 dout("get_cap_refs %p have %s needed %s\n", inode,
2161 ceph_cap_string(have), ceph_cap_string(need));
2164 spin_unlock(&ci->i_ceph_lock);
2166 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2167 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2168 i_size_read(inode) > 0) {
2170 struct page *page = find_get_page(inode->i_mapping, 0);
2172 if (PageUptodate(page)) {
2173 *pinned_page = page;
2176 page_cache_release(page);
2179 * drop cap refs first because getattr while holding
2180 * caps refs can cause deadlock.
2182 ceph_put_cap_refs(ci, _got);
2185 /* getattr request will bring inline data into page cache */
2186 ret1 = __ceph_do_getattr(inode, NULL,
2187 CEPH_STAT_CAP_INLINE_DATA, true);
2196 dout("get_cap_refs %p ret %d got %s\n", inode,
2197 ret, ceph_cap_string(_got));
2203 * Check the offset we are writing up to against our current
2204 * max_size. If necessary, tell the MDS we want to write to
2207 static void check_max_size(struct inode *inode, loff_t endoff)
2209 struct ceph_inode_info *ci = ceph_inode(inode);
2212 /* do we need to explicitly request a larger max_size? */
2213 spin_lock(&ci->i_ceph_lock);
2214 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2215 dout("write %p at large endoff %llu, req max_size\n",
2217 ci->i_wanted_max_size = endoff;
2219 /* duplicate ceph_check_caps()'s logic */
2220 if (ci->i_auth_cap &&
2221 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2222 ci->i_wanted_max_size > ci->i_max_size &&
2223 ci->i_wanted_max_size > ci->i_requested_max_size)
2225 spin_unlock(&ci->i_ceph_lock);
2227 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2231 * Wait for caps, and take cap references. If we can't get a WR cap
2232 * due to a small max_size, make sure we check_max_size (and possibly
2233 * ask the mds) so we don't get hung up indefinitely.
2235 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2236 loff_t endoff, int *got, struct page **pinned_page)
2238 int check_max, ret, err;
2242 check_max_size(&ci->vfs_inode, endoff);
2245 ret = wait_event_interruptible(ci->i_cap_wq,
2246 try_get_cap_refs(ci, need, want, endoff,
2257 * Take cap refs. Caller must already know we hold at least one ref
2258 * on the caps in question or we don't know this is safe.
2260 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2262 spin_lock(&ci->i_ceph_lock);
2263 __take_cap_refs(ci, caps);
2264 spin_unlock(&ci->i_ceph_lock);
2270 * If we released the last ref on any given cap, call ceph_check_caps
2271 * to release (or schedule a release).
2273 * If we are releasing a WR cap (from a sync write), finalize any affected
2274 * cap_snap, and wake up any waiters.
2276 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2278 struct inode *inode = &ci->vfs_inode;
2279 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2280 struct ceph_cap_snap *capsnap;
2282 spin_lock(&ci->i_ceph_lock);
2283 if (had & CEPH_CAP_PIN)
2285 if (had & CEPH_CAP_FILE_RD)
2286 if (--ci->i_rd_ref == 0)
2288 if (had & CEPH_CAP_FILE_CACHE)
2289 if (--ci->i_rdcache_ref == 0)
2291 if (had & CEPH_CAP_FILE_BUFFER) {
2292 if (--ci->i_wb_ref == 0) {
2296 dout("put_cap_refs %p wb %d -> %d (?)\n",
2297 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2299 if (had & CEPH_CAP_FILE_WR)
2300 if (--ci->i_wr_ref == 0) {
2302 if (!list_empty(&ci->i_cap_snaps)) {
2303 capsnap = list_first_entry(&ci->i_cap_snaps,
2304 struct ceph_cap_snap,
2306 if (capsnap->writing) {
2307 capsnap->writing = 0;
2309 __ceph_finish_cap_snap(ci,
2315 spin_unlock(&ci->i_ceph_lock);
2317 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2318 last ? " last" : "", put ? " put" : "");
2320 if (last && !flushsnaps)
2321 ceph_check_caps(ci, 0, NULL);
2322 else if (flushsnaps)
2323 ceph_flush_snaps(ci);
2325 wake_up_all(&ci->i_cap_wq);
2331 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2332 * context. Adjust per-snap dirty page accounting as appropriate.
2333 * Once all dirty data for a cap_snap is flushed, flush snapped file
2334 * metadata back to the MDS. If we dropped the last ref, call
2337 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2338 struct ceph_snap_context *snapc)
2340 struct inode *inode = &ci->vfs_inode;
2342 int complete_capsnap = 0;
2343 int drop_capsnap = 0;
2345 struct ceph_cap_snap *capsnap = NULL;
2347 spin_lock(&ci->i_ceph_lock);
2348 ci->i_wrbuffer_ref -= nr;
2349 last = !ci->i_wrbuffer_ref;
2351 if (ci->i_head_snapc == snapc) {
2352 ci->i_wrbuffer_ref_head -= nr;
2353 if (ci->i_wrbuffer_ref_head == 0 &&
2354 ci->i_dirty_caps == 0 && ci->i_flushing_caps == 0) {
2355 BUG_ON(!ci->i_head_snapc);
2356 ceph_put_snap_context(ci->i_head_snapc);
2357 ci->i_head_snapc = NULL;
2359 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2361 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2362 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2363 last ? " LAST" : "");
2365 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2366 if (capsnap->context == snapc) {
2372 capsnap->dirty_pages -= nr;
2373 if (capsnap->dirty_pages == 0) {
2374 complete_capsnap = 1;
2375 if (capsnap->dirty == 0)
2376 /* cap writeback completed before we created
2377 * the cap_snap; no FLUSHSNAP is needed */
2380 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2381 " snap %lld %d/%d -> %d/%d %s%s%s\n",
2382 inode, capsnap, capsnap->context->seq,
2383 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2384 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2385 last ? " (wrbuffer last)" : "",
2386 complete_capsnap ? " (complete capsnap)" : "",
2387 drop_capsnap ? " (drop capsnap)" : "");
2389 ceph_put_snap_context(capsnap->context);
2390 list_del(&capsnap->ci_item);
2391 list_del(&capsnap->flushing_item);
2392 ceph_put_cap_snap(capsnap);
2396 spin_unlock(&ci->i_ceph_lock);
2399 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2401 } else if (complete_capsnap) {
2402 ceph_flush_snaps(ci);
2403 wake_up_all(&ci->i_cap_wq);
2410 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2412 static void invalidate_aliases(struct inode *inode)
2414 struct dentry *dn, *prev = NULL;
2416 dout("invalidate_aliases inode %p\n", inode);
2417 d_prune_aliases(inode);
2419 * For non-directory inode, d_find_alias() only returns
2420 * hashed dentry. After calling d_invalidate(), the
2421 * dentry becomes unhashed.
2423 * For directory inode, d_find_alias() can return
2424 * unhashed dentry. But directory inode should have
2425 * one alias at most.
2427 while ((dn = d_find_alias(inode))) {
2442 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2443 * actually be a revocation if it specifies a smaller cap set.)
2445 * caller holds s_mutex and i_ceph_lock, we drop both.
2447 static void handle_cap_grant(struct ceph_mds_client *mdsc,
2448 struct inode *inode, struct ceph_mds_caps *grant,
2450 void *inline_data, int inline_len,
2451 struct ceph_buffer *xattr_buf,
2452 struct ceph_mds_session *session,
2453 struct ceph_cap *cap, int issued)
2454 __releases(ci->i_ceph_lock)
2455 __releases(mdsc->snap_rwsem)
2457 struct ceph_inode_info *ci = ceph_inode(inode);
2458 int mds = session->s_mds;
2459 int seq = le32_to_cpu(grant->seq);
2460 int newcaps = le32_to_cpu(grant->caps);
2461 int used, wanted, dirty;
2462 u64 size = le64_to_cpu(grant->size);
2463 u64 max_size = le64_to_cpu(grant->max_size);
2464 struct timespec mtime, atime, ctime;
2467 bool writeback = false;
2468 bool queue_trunc = false;
2469 bool queue_invalidate = false;
2470 bool queue_revalidate = false;
2471 bool deleted_inode = false;
2472 bool fill_inline = false;
2474 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2475 inode, cap, mds, seq, ceph_cap_string(newcaps));
2476 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2481 * auth mds of the inode changed. we received the cap export message,
2482 * but still haven't received the cap import message. handle_cap_export
2483 * updated the new auth MDS' cap.
2485 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2486 * that was sent before the cap import message. So don't remove caps.
2488 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2489 WARN_ON(cap != ci->i_auth_cap);
2490 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2492 newcaps |= cap->issued;
2496 * If CACHE is being revoked, and we have no dirty buffers,
2497 * try to invalidate (once). (If there are dirty buffers, we
2498 * will invalidate _after_ writeback.)
2500 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2501 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2502 !ci->i_wrbuffer_ref) {
2503 if (try_nonblocking_invalidate(inode)) {
2504 /* there were locked pages.. invalidate later
2505 in a separate thread. */
2506 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2507 queue_invalidate = true;
2508 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2512 ceph_fscache_invalidate(inode);
2515 /* side effects now are allowed */
2516 cap->cap_gen = session->s_cap_gen;
2519 __check_cap_issue(ci, cap, newcaps);
2521 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2522 (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2523 inode->i_mode = le32_to_cpu(grant->mode);
2524 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2525 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2526 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2527 from_kuid(&init_user_ns, inode->i_uid),
2528 from_kgid(&init_user_ns, inode->i_gid));
2531 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2532 (issued & CEPH_CAP_LINK_EXCL) == 0) {
2533 set_nlink(inode, le32_to_cpu(grant->nlink));
2534 if (inode->i_nlink == 0 &&
2535 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2536 deleted_inode = true;
2539 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2540 int len = le32_to_cpu(grant->xattr_len);
2541 u64 version = le64_to_cpu(grant->xattr_version);
2543 if (version > ci->i_xattrs.version) {
2544 dout(" got new xattrs v%llu on %p len %d\n",
2545 version, inode, len);
2546 if (ci->i_xattrs.blob)
2547 ceph_buffer_put(ci->i_xattrs.blob);
2548 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2549 ci->i_xattrs.version = version;
2550 ceph_forget_all_cached_acls(inode);
2554 /* Do we need to revalidate our fscache cookie. Don't bother on the
2555 * first cache cap as we already validate at cookie creation time. */
2556 if ((issued & CEPH_CAP_FILE_CACHE) && ci->i_rdcache_gen > 1)
2557 queue_revalidate = true;
2559 if (newcaps & CEPH_CAP_ANY_RD) {
2560 /* ctime/mtime/atime? */
2561 ceph_decode_timespec(&mtime, &grant->mtime);
2562 ceph_decode_timespec(&atime, &grant->atime);
2563 ceph_decode_timespec(&ctime, &grant->ctime);
2564 ceph_fill_file_time(inode, issued,
2565 le32_to_cpu(grant->time_warp_seq),
2566 &ctime, &mtime, &atime);
2569 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
2570 /* file layout may have changed */
2571 ci->i_layout = grant->layout;
2572 /* size/truncate_seq? */
2573 queue_trunc = ceph_fill_file_size(inode, issued,
2574 le32_to_cpu(grant->truncate_seq),
2575 le64_to_cpu(grant->truncate_size),
2577 /* max size increase? */
2578 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
2579 dout("max_size %lld -> %llu\n",
2580 ci->i_max_size, max_size);
2581 ci->i_max_size = max_size;
2582 if (max_size >= ci->i_wanted_max_size) {
2583 ci->i_wanted_max_size = 0; /* reset */
2584 ci->i_requested_max_size = 0;
2590 /* check cap bits */
2591 wanted = __ceph_caps_wanted(ci);
2592 used = __ceph_caps_used(ci);
2593 dirty = __ceph_caps_dirty(ci);
2594 dout(" my wanted = %s, used = %s, dirty %s\n",
2595 ceph_cap_string(wanted),
2596 ceph_cap_string(used),
2597 ceph_cap_string(dirty));
2598 if (wanted != le32_to_cpu(grant->wanted)) {
2599 dout("mds wanted %s -> %s\n",
2600 ceph_cap_string(le32_to_cpu(grant->wanted)),
2601 ceph_cap_string(wanted));
2602 /* imported cap may not have correct mds_wanted */
2603 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
2607 /* revocation, grant, or no-op? */
2608 if (cap->issued & ~newcaps) {
2609 int revoking = cap->issued & ~newcaps;
2611 dout("revocation: %s -> %s (revoking %s)\n",
2612 ceph_cap_string(cap->issued),
2613 ceph_cap_string(newcaps),
2614 ceph_cap_string(revoking));
2615 if (revoking & used & CEPH_CAP_FILE_BUFFER)
2616 writeback = true; /* initiate writeback; will delay ack */
2617 else if (revoking == CEPH_CAP_FILE_CACHE &&
2618 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2620 ; /* do nothing yet, invalidation will be queued */
2621 else if (cap == ci->i_auth_cap)
2622 check_caps = 1; /* check auth cap only */
2624 check_caps = 2; /* check all caps */
2625 cap->issued = newcaps;
2626 cap->implemented |= newcaps;
2627 } else if (cap->issued == newcaps) {
2628 dout("caps unchanged: %s -> %s\n",
2629 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2631 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2632 ceph_cap_string(newcaps));
2633 /* non-auth MDS is revoking the newly grant caps ? */
2634 if (cap == ci->i_auth_cap &&
2635 __ceph_caps_revoking_other(ci, cap, newcaps))
2638 cap->issued = newcaps;
2639 cap->implemented |= newcaps; /* add bits only, to
2640 * avoid stepping on a
2641 * pending revocation */
2644 BUG_ON(cap->issued & ~cap->implemented);
2646 if (inline_version > 0 && inline_version >= ci->i_inline_version) {
2647 ci->i_inline_version = inline_version;
2648 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2649 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
2653 spin_unlock(&ci->i_ceph_lock);
2655 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
2656 kick_flushing_inode_caps(mdsc, session, inode);
2657 up_read(&mdsc->snap_rwsem);
2658 if (newcaps & ~issued)
2663 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
2666 ceph_queue_vmtruncate(inode);
2667 ceph_queue_revalidate(inode);
2668 } else if (queue_revalidate)
2669 ceph_queue_revalidate(inode);
2673 * queue inode for writeback: we can't actually call
2674 * filemap_write_and_wait, etc. from message handler
2677 ceph_queue_writeback(inode);
2678 if (queue_invalidate)
2679 ceph_queue_invalidate(inode);
2681 invalidate_aliases(inode);
2683 wake_up_all(&ci->i_cap_wq);
2685 if (check_caps == 1)
2686 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2688 else if (check_caps == 2)
2689 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2691 mutex_unlock(&session->s_mutex);
2695 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2696 * MDS has been safely committed.
2698 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2699 struct ceph_mds_caps *m,
2700 struct ceph_mds_session *session,
2701 struct ceph_cap *cap)
2702 __releases(ci->i_ceph_lock)
2704 struct ceph_inode_info *ci = ceph_inode(inode);
2705 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2706 unsigned seq = le32_to_cpu(m->seq);
2707 int dirty = le32_to_cpu(m->dirty);
2712 for (i = 0; i < CEPH_CAP_BITS; i++)
2713 if ((dirty & (1 << i)) &&
2714 (u16)flush_tid == ci->i_cap_flush_tid[i])
2717 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2718 " flushing %s -> %s\n",
2719 inode, session->s_mds, seq, ceph_cap_string(dirty),
2720 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2721 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2723 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2726 ci->i_flushing_caps &= ~cleaned;
2728 spin_lock(&mdsc->cap_dirty_lock);
2729 if (ci->i_flushing_caps == 0) {
2730 list_del_init(&ci->i_flushing_item);
2731 if (!list_empty(&session->s_cap_flushing))
2732 dout(" mds%d still flushing cap on %p\n",
2734 &list_entry(session->s_cap_flushing.next,
2735 struct ceph_inode_info,
2736 i_flushing_item)->vfs_inode);
2737 mdsc->num_cap_flushing--;
2738 wake_up_all(&mdsc->cap_flushing_wq);
2739 dout(" inode %p now !flushing\n", inode);
2741 if (ci->i_dirty_caps == 0) {
2742 dout(" inode %p now clean\n", inode);
2743 BUG_ON(!list_empty(&ci->i_dirty_item));
2745 if (ci->i_wrbuffer_ref_head == 0) {
2746 BUG_ON(!ci->i_head_snapc);
2747 ceph_put_snap_context(ci->i_head_snapc);
2748 ci->i_head_snapc = NULL;
2751 BUG_ON(list_empty(&ci->i_dirty_item));
2754 spin_unlock(&mdsc->cap_dirty_lock);
2755 wake_up_all(&ci->i_cap_wq);
2758 spin_unlock(&ci->i_ceph_lock);
2764 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
2765 * throw away our cap_snap.
2767 * Caller hold s_mutex.
2769 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2770 struct ceph_mds_caps *m,
2771 struct ceph_mds_session *session)
2773 struct ceph_inode_info *ci = ceph_inode(inode);
2774 u64 follows = le64_to_cpu(m->snap_follows);
2775 struct ceph_cap_snap *capsnap;
2778 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2779 inode, ci, session->s_mds, follows);
2781 spin_lock(&ci->i_ceph_lock);
2782 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2783 if (capsnap->follows == follows) {
2784 if (capsnap->flush_tid != flush_tid) {
2785 dout(" cap_snap %p follows %lld tid %lld !="
2786 " %lld\n", capsnap, follows,
2787 flush_tid, capsnap->flush_tid);
2790 WARN_ON(capsnap->dirty_pages || capsnap->writing);
2791 dout(" removing %p cap_snap %p follows %lld\n",
2792 inode, capsnap, follows);
2793 ceph_put_snap_context(capsnap->context);
2794 list_del(&capsnap->ci_item);
2795 list_del(&capsnap->flushing_item);
2796 ceph_put_cap_snap(capsnap);
2800 dout(" skipping cap_snap %p follows %lld\n",
2801 capsnap, capsnap->follows);
2804 spin_unlock(&ci->i_ceph_lock);
2810 * Handle TRUNC from MDS, indicating file truncation.
2812 * caller hold s_mutex.
2814 static void handle_cap_trunc(struct inode *inode,
2815 struct ceph_mds_caps *trunc,
2816 struct ceph_mds_session *session)
2817 __releases(ci->i_ceph_lock)
2819 struct ceph_inode_info *ci = ceph_inode(inode);
2820 int mds = session->s_mds;
2821 int seq = le32_to_cpu(trunc->seq);
2822 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2823 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2824 u64 size = le64_to_cpu(trunc->size);
2825 int implemented = 0;
2826 int dirty = __ceph_caps_dirty(ci);
2827 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2828 int queue_trunc = 0;
2830 issued |= implemented | dirty;
2832 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2833 inode, mds, seq, truncate_size, truncate_seq);
2834 queue_trunc = ceph_fill_file_size(inode, issued,
2835 truncate_seq, truncate_size, size);
2836 spin_unlock(&ci->i_ceph_lock);
2839 ceph_queue_vmtruncate(inode);
2840 ceph_fscache_invalidate(inode);
2845 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
2846 * different one. If we are the most recent migration we've seen (as
2847 * indicated by mseq), make note of the migrating cap bits for the
2848 * duration (until we see the corresponding IMPORT).
2850 * caller holds s_mutex
2852 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2853 struct ceph_mds_cap_peer *ph,
2854 struct ceph_mds_session *session)
2856 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2857 struct ceph_mds_session *tsession = NULL;
2858 struct ceph_cap *cap, *tcap, *new_cap = NULL;
2859 struct ceph_inode_info *ci = ceph_inode(inode);
2861 unsigned mseq = le32_to_cpu(ex->migrate_seq);
2862 unsigned t_seq, t_mseq;
2864 int mds = session->s_mds;
2867 t_cap_id = le64_to_cpu(ph->cap_id);
2868 t_seq = le32_to_cpu(ph->seq);
2869 t_mseq = le32_to_cpu(ph->mseq);
2870 target = le32_to_cpu(ph->mds);
2872 t_cap_id = t_seq = t_mseq = 0;
2876 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
2877 inode, ci, mds, mseq, target);
2879 spin_lock(&ci->i_ceph_lock);
2880 cap = __get_cap_for_mds(ci, mds);
2881 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
2885 __ceph_remove_cap(cap, false);
2890 * now we know we haven't received the cap import message yet
2891 * because the exported cap still exist.
2894 issued = cap->issued;
2895 WARN_ON(issued != cap->implemented);
2897 tcap = __get_cap_for_mds(ci, target);
2899 /* already have caps from the target */
2900 if (tcap->cap_id != t_cap_id ||
2901 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
2902 dout(" updating import cap %p mds%d\n", tcap, target);
2903 tcap->cap_id = t_cap_id;
2904 tcap->seq = t_seq - 1;
2905 tcap->issue_seq = t_seq - 1;
2906 tcap->mseq = t_mseq;
2907 tcap->issued |= issued;
2908 tcap->implemented |= issued;
2909 if (cap == ci->i_auth_cap)
2910 ci->i_auth_cap = tcap;
2911 if (ci->i_flushing_caps && ci->i_auth_cap == tcap) {
2912 spin_lock(&mdsc->cap_dirty_lock);
2913 list_move_tail(&ci->i_flushing_item,
2914 &tcap->session->s_cap_flushing);
2915 spin_unlock(&mdsc->cap_dirty_lock);
2918 __ceph_remove_cap(cap, false);
2920 } else if (tsession) {
2921 /* add placeholder for the export tagert */
2922 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
2923 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
2924 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
2926 __ceph_remove_cap(cap, false);
2930 spin_unlock(&ci->i_ceph_lock);
2931 mutex_unlock(&session->s_mutex);
2933 /* open target session */
2934 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
2935 if (!IS_ERR(tsession)) {
2937 mutex_lock(&session->s_mutex);
2938 mutex_lock_nested(&tsession->s_mutex,
2939 SINGLE_DEPTH_NESTING);
2941 mutex_lock(&tsession->s_mutex);
2942 mutex_lock_nested(&session->s_mutex,
2943 SINGLE_DEPTH_NESTING);
2945 ceph_add_cap_releases(mdsc, tsession);
2946 new_cap = ceph_get_cap(mdsc, NULL);
2955 spin_unlock(&ci->i_ceph_lock);
2956 mutex_unlock(&session->s_mutex);
2958 mutex_unlock(&tsession->s_mutex);
2959 ceph_put_mds_session(tsession);
2962 ceph_put_cap(mdsc, new_cap);
2966 * Handle cap IMPORT.
2968 * caller holds s_mutex. acquires i_ceph_lock
2970 static void handle_cap_import(struct ceph_mds_client *mdsc,
2971 struct inode *inode, struct ceph_mds_caps *im,
2972 struct ceph_mds_cap_peer *ph,
2973 struct ceph_mds_session *session,
2974 struct ceph_cap **target_cap, int *old_issued)
2975 __acquires(ci->i_ceph_lock)
2977 struct ceph_inode_info *ci = ceph_inode(inode);
2978 struct ceph_cap *cap, *ocap, *new_cap = NULL;
2979 int mds = session->s_mds;
2981 unsigned caps = le32_to_cpu(im->caps);
2982 unsigned wanted = le32_to_cpu(im->wanted);
2983 unsigned seq = le32_to_cpu(im->seq);
2984 unsigned mseq = le32_to_cpu(im->migrate_seq);
2985 u64 realmino = le64_to_cpu(im->realm);
2986 u64 cap_id = le64_to_cpu(im->cap_id);
2991 p_cap_id = le64_to_cpu(ph->cap_id);
2992 peer = le32_to_cpu(ph->mds);
2998 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
2999 inode, ci, mds, mseq, peer);
3002 spin_lock(&ci->i_ceph_lock);
3003 cap = __get_cap_for_mds(ci, mds);
3006 spin_unlock(&ci->i_ceph_lock);
3007 new_cap = ceph_get_cap(mdsc, NULL);
3013 ceph_put_cap(mdsc, new_cap);
3018 __ceph_caps_issued(ci, &issued);
3019 issued |= __ceph_caps_dirty(ci);
3021 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3022 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3024 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3025 if (ocap && ocap->cap_id == p_cap_id) {
3026 dout(" remove export cap %p mds%d flags %d\n",
3027 ocap, peer, ph->flags);
3028 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3029 (ocap->seq != le32_to_cpu(ph->seq) ||
3030 ocap->mseq != le32_to_cpu(ph->mseq))) {
3031 pr_err("handle_cap_import: mismatched seq/mseq: "
3032 "ino (%llx.%llx) mds%d seq %d mseq %d "
3033 "importer mds%d has peer seq %d mseq %d\n",
3034 ceph_vinop(inode), peer, ocap->seq,
3035 ocap->mseq, mds, le32_to_cpu(ph->seq),
3036 le32_to_cpu(ph->mseq));
3038 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3041 /* make sure we re-request max_size, if necessary */
3042 ci->i_wanted_max_size = 0;
3043 ci->i_requested_max_size = 0;
3045 *old_issued = issued;
3050 * Handle a caps message from the MDS.
3052 * Identify the appropriate session, inode, and call the right handler
3053 * based on the cap op.
3055 void ceph_handle_caps(struct ceph_mds_session *session,
3056 struct ceph_msg *msg)
3058 struct ceph_mds_client *mdsc = session->s_mdsc;
3059 struct super_block *sb = mdsc->fsc->sb;
3060 struct inode *inode;
3061 struct ceph_inode_info *ci;
3062 struct ceph_cap *cap;
3063 struct ceph_mds_caps *h;
3064 struct ceph_mds_cap_peer *peer = NULL;
3065 struct ceph_snap_realm *realm;
3066 int mds = session->s_mds;
3069 struct ceph_vino vino;
3073 u64 inline_version = 0;
3074 void *inline_data = NULL;
3077 size_t snaptrace_len;
3080 dout("handle_caps from mds%d\n", mds);
3083 end = msg->front.iov_base + msg->front.iov_len;
3084 tid = le64_to_cpu(msg->hdr.tid);
3085 if (msg->front.iov_len < sizeof(*h))
3087 h = msg->front.iov_base;
3088 op = le32_to_cpu(h->op);
3089 vino.ino = le64_to_cpu(h->ino);
3090 vino.snap = CEPH_NOSNAP;
3091 cap_id = le64_to_cpu(h->cap_id);
3092 seq = le32_to_cpu(h->seq);
3093 mseq = le32_to_cpu(h->migrate_seq);
3094 size = le64_to_cpu(h->size);
3095 max_size = le64_to_cpu(h->max_size);
3098 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3099 p = snaptrace + snaptrace_len;
3101 if (le16_to_cpu(msg->hdr.version) >= 2) {
3103 ceph_decode_32_safe(&p, end, flock_len, bad);
3104 if (p + flock_len > end)
3109 if (le16_to_cpu(msg->hdr.version) >= 3) {
3110 if (op == CEPH_CAP_OP_IMPORT) {
3111 if (p + sizeof(*peer) > end)
3115 } else if (op == CEPH_CAP_OP_EXPORT) {
3116 /* recorded in unused fields */
3117 peer = (void *)&h->size;
3121 if (le16_to_cpu(msg->hdr.version) >= 4) {
3122 ceph_decode_64_safe(&p, end, inline_version, bad);
3123 ceph_decode_32_safe(&p, end, inline_len, bad);
3124 if (p + inline_len > end)
3131 inode = ceph_find_inode(sb, vino);
3132 ci = ceph_inode(inode);
3133 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3136 mutex_lock(&session->s_mutex);
3138 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3141 if (op == CEPH_CAP_OP_IMPORT)
3142 ceph_add_cap_releases(mdsc, session);
3145 dout(" i don't have ino %llx\n", vino.ino);
3147 if (op == CEPH_CAP_OP_IMPORT) {
3148 spin_lock(&session->s_cap_lock);
3149 __queue_cap_release(session, vino.ino, cap_id,
3151 spin_unlock(&session->s_cap_lock);
3153 goto flush_cap_releases;
3156 /* these will work even if we don't have a cap yet */
3158 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3159 handle_cap_flushsnap_ack(inode, tid, h, session);
3162 case CEPH_CAP_OP_EXPORT:
3163 handle_cap_export(inode, h, peer, session);
3166 case CEPH_CAP_OP_IMPORT:
3168 if (snaptrace_len) {
3169 down_write(&mdsc->snap_rwsem);
3170 ceph_update_snap_trace(mdsc, snaptrace,
3171 snaptrace + snaptrace_len,
3173 downgrade_write(&mdsc->snap_rwsem);
3175 down_read(&mdsc->snap_rwsem);
3177 handle_cap_import(mdsc, inode, h, peer, session,
3179 handle_cap_grant(mdsc, inode, h,
3180 inline_version, inline_data, inline_len,
3181 msg->middle, session, cap, issued);
3183 ceph_put_snap_realm(mdsc, realm);
3187 /* the rest require a cap */
3188 spin_lock(&ci->i_ceph_lock);
3189 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3191 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3192 inode, ceph_ino(inode), ceph_snap(inode), mds);
3193 spin_unlock(&ci->i_ceph_lock);
3194 goto flush_cap_releases;
3197 /* note that each of these drops i_ceph_lock for us */
3199 case CEPH_CAP_OP_REVOKE:
3200 case CEPH_CAP_OP_GRANT:
3201 __ceph_caps_issued(ci, &issued);
3202 issued |= __ceph_caps_dirty(ci);
3203 handle_cap_grant(mdsc, inode, h,
3204 inline_version, inline_data, inline_len,
3205 msg->middle, session, cap, issued);
3208 case CEPH_CAP_OP_FLUSH_ACK:
3209 handle_cap_flush_ack(inode, tid, h, session, cap);
3212 case CEPH_CAP_OP_TRUNC:
3213 handle_cap_trunc(inode, h, session);
3217 spin_unlock(&ci->i_ceph_lock);
3218 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3219 ceph_cap_op_name(op));
3226 * send any full release message to try to move things
3227 * along for the mds (who clearly thinks we still have this
3230 ceph_add_cap_releases(mdsc, session);
3231 ceph_send_cap_releases(mdsc, session);
3234 mutex_unlock(&session->s_mutex);
3240 pr_err("ceph_handle_caps: corrupt message\n");
3246 * Delayed work handler to process end of delayed cap release LRU list.
3248 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3250 struct ceph_inode_info *ci;
3251 int flags = CHECK_CAPS_NODELAY;
3253 dout("check_delayed_caps\n");
3255 spin_lock(&mdsc->cap_delay_lock);
3256 if (list_empty(&mdsc->cap_delay_list))
3258 ci = list_first_entry(&mdsc->cap_delay_list,
3259 struct ceph_inode_info,
3261 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3262 time_before(jiffies, ci->i_hold_caps_max))
3264 list_del_init(&ci->i_cap_delay_list);
3265 spin_unlock(&mdsc->cap_delay_lock);
3266 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3267 ceph_check_caps(ci, flags, NULL);
3269 spin_unlock(&mdsc->cap_delay_lock);
3273 * Flush all dirty caps to the mds
3275 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3277 struct ceph_inode_info *ci;
3278 struct inode *inode;
3280 dout("flush_dirty_caps\n");
3281 spin_lock(&mdsc->cap_dirty_lock);
3282 while (!list_empty(&mdsc->cap_dirty)) {
3283 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3285 inode = &ci->vfs_inode;
3287 dout("flush_dirty_caps %p\n", inode);
3288 spin_unlock(&mdsc->cap_dirty_lock);
3289 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3291 spin_lock(&mdsc->cap_dirty_lock);
3293 spin_unlock(&mdsc->cap_dirty_lock);
3294 dout("flush_dirty_caps done\n");
3298 * Drop open file reference. If we were the last open file,
3299 * we may need to release capabilities to the MDS (or schedule
3300 * their delayed release).
3302 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3304 struct inode *inode = &ci->vfs_inode;
3307 spin_lock(&ci->i_ceph_lock);
3308 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
3309 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
3310 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
3311 if (--ci->i_nr_by_mode[fmode] == 0)
3313 spin_unlock(&ci->i_ceph_lock);
3315 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3316 ceph_check_caps(ci, 0, NULL);
3320 * Helpers for embedding cap and dentry lease releases into mds
3323 * @force is used by dentry_release (below) to force inclusion of a
3324 * record for the directory inode, even when there aren't any caps to
3327 int ceph_encode_inode_release(void **p, struct inode *inode,
3328 int mds, int drop, int unless, int force)
3330 struct ceph_inode_info *ci = ceph_inode(inode);
3331 struct ceph_cap *cap;
3332 struct ceph_mds_request_release *rel = *p;
3336 spin_lock(&ci->i_ceph_lock);
3337 used = __ceph_caps_used(ci);
3338 dirty = __ceph_caps_dirty(ci);
3340 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3341 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3342 ceph_cap_string(unless));
3344 /* only drop unused, clean caps */
3345 drop &= ~(used | dirty);
3347 cap = __get_cap_for_mds(ci, mds);
3348 if (cap && __cap_is_valid(cap)) {
3350 ((cap->issued & drop) &&
3351 (cap->issued & unless) == 0)) {
3352 if ((cap->issued & drop) &&
3353 (cap->issued & unless) == 0) {
3354 int wanted = __ceph_caps_wanted(ci);
3355 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3356 wanted |= cap->mds_wanted;
3357 dout("encode_inode_release %p cap %p "
3358 "%s -> %s, wanted %s -> %s\n", inode, cap,
3359 ceph_cap_string(cap->issued),
3360 ceph_cap_string(cap->issued & ~drop),
3361 ceph_cap_string(cap->mds_wanted),
3362 ceph_cap_string(wanted));
3364 cap->issued &= ~drop;
3365 cap->implemented &= ~drop;
3366 cap->mds_wanted = wanted;
3368 dout("encode_inode_release %p cap %p %s"
3369 " (force)\n", inode, cap,
3370 ceph_cap_string(cap->issued));
3373 rel->ino = cpu_to_le64(ceph_ino(inode));
3374 rel->cap_id = cpu_to_le64(cap->cap_id);
3375 rel->seq = cpu_to_le32(cap->seq);
3376 rel->issue_seq = cpu_to_le32(cap->issue_seq);
3377 rel->mseq = cpu_to_le32(cap->mseq);
3378 rel->caps = cpu_to_le32(cap->implemented);
3379 rel->wanted = cpu_to_le32(cap->mds_wanted);
3385 dout("encode_inode_release %p cap %p %s\n",
3386 inode, cap, ceph_cap_string(cap->issued));
3389 spin_unlock(&ci->i_ceph_lock);
3393 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3394 int mds, int drop, int unless)
3396 struct inode *dir = dentry->d_parent->d_inode;
3397 struct ceph_mds_request_release *rel = *p;
3398 struct ceph_dentry_info *di = ceph_dentry(dentry);
3403 * force an record for the directory caps if we have a dentry lease.
3404 * this is racy (can't take i_ceph_lock and d_lock together), but it
3405 * doesn't have to be perfect; the mds will revoke anything we don't
3408 spin_lock(&dentry->d_lock);
3409 if (di->lease_session && di->lease_session->s_mds == mds)
3411 spin_unlock(&dentry->d_lock);
3413 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3415 spin_lock(&dentry->d_lock);
3416 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3417 dout("encode_dentry_release %p mds%d seq %d\n",
3418 dentry, mds, (int)di->lease_seq);
3419 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3420 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3421 *p += dentry->d_name.len;
3422 rel->dname_seq = cpu_to_le32(di->lease_seq);
3423 __ceph_mdsc_drop_dentry_lease(dentry);
3425 spin_unlock(&dentry->d_lock);