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);
899 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
901 struct ceph_snap_realm *realm = ci->i_snap_realm;
902 spin_lock(&realm->inodes_with_caps_lock);
903 list_del_init(&ci->i_snap_realm_item);
904 ci->i_snap_realm_counter++;
905 ci->i_snap_realm = NULL;
906 spin_unlock(&realm->inodes_with_caps_lock);
907 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
912 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
914 * caller should hold i_ceph_lock.
915 * caller will not hold session s_mutex if called from destroy_inode.
917 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
919 struct ceph_mds_session *session = cap->session;
920 struct ceph_inode_info *ci = cap->ci;
921 struct ceph_mds_client *mdsc =
922 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
925 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
927 /* remove from session list */
928 spin_lock(&session->s_cap_lock);
930 * s_cap_reconnect is protected by s_cap_lock. no one changes
931 * s_cap_gen while session is in the reconnect state.
934 (!session->s_cap_reconnect ||
935 cap->cap_gen == session->s_cap_gen))
936 __queue_cap_release(session, ci->i_vino.ino, cap->cap_id,
937 cap->mseq, cap->issue_seq);
939 if (session->s_cap_iterator == cap) {
940 /* not yet, we are iterating over this very cap */
941 dout("__ceph_remove_cap delaying %p removal from session %p\n",
944 list_del_init(&cap->session_caps);
945 session->s_nr_caps--;
949 /* protect backpointer with s_cap_lock: see iterate_session_caps */
951 spin_unlock(&session->s_cap_lock);
953 /* remove from inode list */
954 rb_erase(&cap->ci_node, &ci->i_caps);
955 if (ci->i_auth_cap == cap)
956 ci->i_auth_cap = NULL;
959 ceph_put_cap(mdsc, cap);
961 /* when reconnect denied, we remove session caps forcibly,
962 * i_wr_ref can be non-zero. If there are ongoing write,
965 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
966 drop_inode_snap_realm(ci);
968 if (!__ceph_is_any_real_caps(ci))
969 __cap_delay_cancel(mdsc, ci);
973 * Build and send a cap message to the given MDS.
975 * Caller should be holding s_mutex.
977 static int send_cap_msg(struct ceph_mds_session *session,
978 u64 ino, u64 cid, int op,
979 int caps, int wanted, int dirty,
980 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
981 u64 size, u64 max_size,
982 struct timespec *mtime, struct timespec *atime,
984 kuid_t uid, kgid_t gid, umode_t mode,
986 struct ceph_buffer *xattrs_buf,
987 u64 follows, bool inline_data)
989 struct ceph_mds_caps *fc;
990 struct ceph_msg *msg;
994 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
995 " seq %u/%u mseq %u follows %lld size %llu/%llu"
996 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
997 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
998 ceph_cap_string(dirty),
999 seq, issue_seq, mseq, follows, size, max_size,
1000 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
1002 /* flock buffer size + inline version + inline data size */
1003 extra_len = 4 + 8 + 4;
1004 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1009 msg->hdr.tid = cpu_to_le64(flush_tid);
1011 fc = msg->front.iov_base;
1012 memset(fc, 0, sizeof(*fc));
1014 fc->cap_id = cpu_to_le64(cid);
1015 fc->op = cpu_to_le32(op);
1016 fc->seq = cpu_to_le32(seq);
1017 fc->issue_seq = cpu_to_le32(issue_seq);
1018 fc->migrate_seq = cpu_to_le32(mseq);
1019 fc->caps = cpu_to_le32(caps);
1020 fc->wanted = cpu_to_le32(wanted);
1021 fc->dirty = cpu_to_le32(dirty);
1022 fc->ino = cpu_to_le64(ino);
1023 fc->snap_follows = cpu_to_le64(follows);
1025 fc->size = cpu_to_le64(size);
1026 fc->max_size = cpu_to_le64(max_size);
1028 ceph_encode_timespec(&fc->mtime, mtime);
1030 ceph_encode_timespec(&fc->atime, atime);
1031 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
1033 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid));
1034 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid));
1035 fc->mode = cpu_to_le32(mode);
1038 /* flock buffer size */
1039 ceph_encode_32(&p, 0);
1040 /* inline version */
1041 ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE);
1042 /* inline data size */
1043 ceph_encode_32(&p, 0);
1045 fc->xattr_version = cpu_to_le64(xattr_version);
1047 msg->middle = ceph_buffer_get(xattrs_buf);
1048 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1049 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1052 ceph_con_send(&session->s_con, msg);
1056 void __queue_cap_release(struct ceph_mds_session *session,
1057 u64 ino, u64 cap_id, u32 migrate_seq,
1060 struct ceph_msg *msg;
1061 struct ceph_mds_cap_release *head;
1062 struct ceph_mds_cap_item *item;
1064 BUG_ON(!session->s_num_cap_releases);
1065 msg = list_first_entry(&session->s_cap_releases,
1066 struct ceph_msg, list_head);
1068 dout(" adding %llx release to mds%d msg %p (%d left)\n",
1069 ino, session->s_mds, msg, session->s_num_cap_releases);
1071 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1072 head = msg->front.iov_base;
1073 le32_add_cpu(&head->num, 1);
1074 item = msg->front.iov_base + msg->front.iov_len;
1075 item->ino = cpu_to_le64(ino);
1076 item->cap_id = cpu_to_le64(cap_id);
1077 item->migrate_seq = cpu_to_le32(migrate_seq);
1078 item->seq = cpu_to_le32(issue_seq);
1080 session->s_num_cap_releases--;
1082 msg->front.iov_len += sizeof(*item);
1083 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1084 dout(" release msg %p full\n", msg);
1085 list_move_tail(&msg->list_head, &session->s_cap_releases_done);
1087 dout(" release msg %p at %d/%d (%d)\n", msg,
1088 (int)le32_to_cpu(head->num),
1089 (int)CEPH_CAPS_PER_RELEASE,
1090 (int)msg->front.iov_len);
1095 * Queue cap releases when an inode is dropped from our cache. Since
1096 * inode is about to be destroyed, there is no need for i_ceph_lock.
1098 void ceph_queue_caps_release(struct inode *inode)
1100 struct ceph_inode_info *ci = ceph_inode(inode);
1103 p = rb_first(&ci->i_caps);
1105 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1107 __ceph_remove_cap(cap, true);
1112 * Send a cap msg on the given inode. Update our caps state, then
1113 * drop i_ceph_lock and send the message.
1115 * Make note of max_size reported/requested from mds, revoked caps
1116 * that have now been implemented.
1118 * Make half-hearted attempt ot to invalidate page cache if we are
1119 * dropping RDCACHE. Note that this will leave behind locked pages
1120 * that we'll then need to deal with elsewhere.
1122 * Return non-zero if delayed release, or we experienced an error
1123 * such that the caller should requeue + retry later.
1125 * called with i_ceph_lock, then drops it.
1126 * caller should hold snap_rwsem (read), s_mutex.
1128 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1129 int op, int used, int want, int retain, int flushing,
1130 unsigned *pflush_tid)
1131 __releases(cap->ci->i_ceph_lock)
1133 struct ceph_inode_info *ci = cap->ci;
1134 struct inode *inode = &ci->vfs_inode;
1135 u64 cap_id = cap->cap_id;
1136 int held, revoking, dropping, keep;
1137 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1139 struct timespec mtime, atime;
1144 struct ceph_mds_session *session;
1145 u64 xattr_version = 0;
1146 struct ceph_buffer *xattr_blob = NULL;
1153 held = cap->issued | cap->implemented;
1154 revoking = cap->implemented & ~cap->issued;
1155 retain &= ~revoking;
1156 dropping = cap->issued & ~retain;
1158 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1159 inode, cap, cap->session,
1160 ceph_cap_string(held), ceph_cap_string(held & retain),
1161 ceph_cap_string(revoking));
1162 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1164 session = cap->session;
1166 /* don't release wanted unless we've waited a bit. */
1167 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1168 time_before(jiffies, ci->i_hold_caps_min)) {
1169 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1170 ceph_cap_string(cap->issued),
1171 ceph_cap_string(cap->issued & retain),
1172 ceph_cap_string(cap->mds_wanted),
1173 ceph_cap_string(want));
1174 want |= cap->mds_wanted;
1175 retain |= cap->issued;
1178 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1180 cap->issued &= retain; /* drop bits we don't want */
1181 if (cap->implemented & ~cap->issued) {
1183 * Wake up any waiters on wanted -> needed transition.
1184 * This is due to the weird transition from buffered
1185 * to sync IO... we need to flush dirty pages _before_
1186 * allowing sync writes to avoid reordering.
1190 cap->implemented &= cap->issued | used;
1191 cap->mds_wanted = want;
1195 * assign a tid for flush operations so we can avoid
1196 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1197 * clean type races. track latest tid for every bit
1198 * so we can handle flush AxFw, flush Fw, and have the
1199 * first ack clean Ax.
1201 flush_tid = ++ci->i_cap_flush_last_tid;
1203 *pflush_tid = flush_tid;
1204 dout(" cap_flush_tid %d\n", (int)flush_tid);
1205 for (i = 0; i < CEPH_CAP_BITS; i++)
1206 if (flushing & (1 << i))
1207 ci->i_cap_flush_tid[i] = flush_tid;
1209 follows = ci->i_head_snapc->seq;
1214 keep = cap->implemented;
1216 issue_seq = cap->issue_seq;
1218 size = inode->i_size;
1219 ci->i_reported_size = size;
1220 max_size = ci->i_wanted_max_size;
1221 ci->i_requested_max_size = max_size;
1222 mtime = inode->i_mtime;
1223 atime = inode->i_atime;
1224 time_warp_seq = ci->i_time_warp_seq;
1227 mode = inode->i_mode;
1229 if (flushing & CEPH_CAP_XATTR_EXCL) {
1230 __ceph_build_xattrs_blob(ci);
1231 xattr_blob = ci->i_xattrs.blob;
1232 xattr_version = ci->i_xattrs.version;
1235 inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1237 spin_unlock(&ci->i_ceph_lock);
1239 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1240 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1241 size, max_size, &mtime, &atime, time_warp_seq,
1242 uid, gid, mode, xattr_version, xattr_blob,
1243 follows, inline_data);
1245 dout("error sending cap msg, must requeue %p\n", inode);
1250 wake_up_all(&ci->i_cap_wq);
1256 * When a snapshot is taken, clients accumulate dirty metadata on
1257 * inodes with capabilities in ceph_cap_snaps to describe the file
1258 * state at the time the snapshot was taken. This must be flushed
1259 * asynchronously back to the MDS once sync writes complete and dirty
1260 * data is written out.
1262 * Unless @kick is true, skip cap_snaps that were already sent to
1263 * the MDS (i.e., during this session).
1265 * Called under i_ceph_lock. Takes s_mutex as needed.
1267 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1268 struct ceph_mds_session **psession,
1270 __releases(ci->i_ceph_lock)
1271 __acquires(ci->i_ceph_lock)
1273 struct inode *inode = &ci->vfs_inode;
1275 struct ceph_cap_snap *capsnap;
1277 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1278 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1280 u64 next_follows = 0; /* keep track of how far we've gotten through the
1281 i_cap_snaps list, and skip these entries next time
1282 around to avoid an infinite loop */
1285 session = *psession;
1287 dout("__flush_snaps %p\n", inode);
1289 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1290 /* avoid an infiniute loop after retry */
1291 if (capsnap->follows < next_follows)
1294 * we need to wait for sync writes to complete and for dirty
1295 * pages to be written out.
1297 if (capsnap->dirty_pages || capsnap->writing)
1300 /* should be removed by ceph_try_drop_cap_snap() */
1301 BUG_ON(!capsnap->need_flush);
1303 /* pick mds, take s_mutex */
1304 if (ci->i_auth_cap == NULL) {
1305 dout("no auth cap (migrating?), doing nothing\n");
1309 /* only flush each capsnap once */
1310 if (!kick && !list_empty(&capsnap->flushing_item)) {
1311 dout("already flushed %p, skipping\n", capsnap);
1315 mds = ci->i_auth_cap->session->s_mds;
1316 mseq = ci->i_auth_cap->mseq;
1318 if (session && session->s_mds != mds) {
1319 dout("oops, wrong session %p mutex\n", session);
1323 mutex_unlock(&session->s_mutex);
1324 ceph_put_mds_session(session);
1328 spin_unlock(&ci->i_ceph_lock);
1329 mutex_lock(&mdsc->mutex);
1330 session = __ceph_lookup_mds_session(mdsc, mds);
1331 mutex_unlock(&mdsc->mutex);
1333 dout("inverting session/ino locks on %p\n",
1335 mutex_lock(&session->s_mutex);
1338 * if session == NULL, we raced against a cap
1339 * deletion or migration. retry, and we'll
1340 * get a better @mds value next time.
1342 spin_lock(&ci->i_ceph_lock);
1346 capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1347 atomic_inc(&capsnap->nref);
1348 if (list_empty(&capsnap->flushing_item))
1349 list_add_tail(&capsnap->flushing_item,
1350 &session->s_cap_snaps_flushing);
1351 spin_unlock(&ci->i_ceph_lock);
1353 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1354 inode, capsnap, capsnap->follows, capsnap->flush_tid);
1355 send_cap_msg(session, ceph_vino(inode).ino, 0,
1356 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1357 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1359 &capsnap->mtime, &capsnap->atime,
1360 capsnap->time_warp_seq,
1361 capsnap->uid, capsnap->gid, capsnap->mode,
1362 capsnap->xattr_version, capsnap->xattr_blob,
1363 capsnap->follows, capsnap->inline_data);
1365 next_follows = capsnap->follows + 1;
1366 ceph_put_cap_snap(capsnap);
1368 spin_lock(&ci->i_ceph_lock);
1372 /* we flushed them all; remove this inode from the queue */
1373 spin_lock(&mdsc->snap_flush_lock);
1374 list_del_init(&ci->i_snap_flush_item);
1375 spin_unlock(&mdsc->snap_flush_lock);
1379 *psession = session;
1381 mutex_unlock(&session->s_mutex);
1382 ceph_put_mds_session(session);
1386 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1388 spin_lock(&ci->i_ceph_lock);
1389 __ceph_flush_snaps(ci, NULL, 0);
1390 spin_unlock(&ci->i_ceph_lock);
1394 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1395 * Caller is then responsible for calling __mark_inode_dirty with the
1396 * returned flags value.
1398 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1400 struct ceph_mds_client *mdsc =
1401 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1402 struct inode *inode = &ci->vfs_inode;
1403 int was = ci->i_dirty_caps;
1406 if (!ci->i_auth_cap) {
1407 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1408 "but no auth cap (session was closed?)\n",
1409 inode, ceph_ino(inode), ceph_cap_string(mask));
1413 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1414 ceph_cap_string(mask), ceph_cap_string(was),
1415 ceph_cap_string(was | mask));
1416 ci->i_dirty_caps |= mask;
1418 if (!ci->i_head_snapc) {
1419 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1420 ci->i_head_snapc = ceph_get_snap_context(
1421 ci->i_snap_realm->cached_context);
1423 dout(" inode %p now dirty snapc %p auth cap %p\n",
1424 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1425 BUG_ON(!list_empty(&ci->i_dirty_item));
1426 spin_lock(&mdsc->cap_dirty_lock);
1427 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1428 spin_unlock(&mdsc->cap_dirty_lock);
1429 if (ci->i_flushing_caps == 0) {
1431 dirty |= I_DIRTY_SYNC;
1434 BUG_ON(list_empty(&ci->i_dirty_item));
1435 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1436 (mask & CEPH_CAP_FILE_BUFFER))
1437 dirty |= I_DIRTY_DATASYNC;
1438 __cap_delay_requeue(mdsc, ci);
1443 * Add dirty inode to the flushing list. Assigned a seq number so we
1444 * can wait for caps to flush without starving.
1446 * Called under i_ceph_lock.
1448 static int __mark_caps_flushing(struct inode *inode,
1449 struct ceph_mds_session *session)
1451 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1452 struct ceph_inode_info *ci = ceph_inode(inode);
1455 BUG_ON(ci->i_dirty_caps == 0);
1456 BUG_ON(list_empty(&ci->i_dirty_item));
1458 flushing = ci->i_dirty_caps;
1459 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1460 ceph_cap_string(flushing),
1461 ceph_cap_string(ci->i_flushing_caps),
1462 ceph_cap_string(ci->i_flushing_caps | flushing));
1463 ci->i_flushing_caps |= flushing;
1464 ci->i_dirty_caps = 0;
1465 dout(" inode %p now !dirty\n", inode);
1467 spin_lock(&mdsc->cap_dirty_lock);
1468 list_del_init(&ci->i_dirty_item);
1470 if (list_empty(&ci->i_flushing_item)) {
1471 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1472 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1473 mdsc->num_cap_flushing++;
1474 dout(" inode %p now flushing seq %lld\n", inode,
1475 ci->i_cap_flush_seq);
1477 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1478 dout(" inode %p now flushing (more) seq %lld\n", inode,
1479 ci->i_cap_flush_seq);
1481 spin_unlock(&mdsc->cap_dirty_lock);
1487 * try to invalidate mapping pages without blocking.
1489 static int try_nonblocking_invalidate(struct inode *inode)
1491 struct ceph_inode_info *ci = ceph_inode(inode);
1492 u32 invalidating_gen = ci->i_rdcache_gen;
1494 spin_unlock(&ci->i_ceph_lock);
1495 invalidate_mapping_pages(&inode->i_data, 0, -1);
1496 spin_lock(&ci->i_ceph_lock);
1498 if (inode->i_data.nrpages == 0 &&
1499 invalidating_gen == ci->i_rdcache_gen) {
1501 dout("try_nonblocking_invalidate %p success\n", inode);
1502 /* save any racing async invalidate some trouble */
1503 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1506 dout("try_nonblocking_invalidate %p failed\n", inode);
1511 * Swiss army knife function to examine currently used and wanted
1512 * versus held caps. Release, flush, ack revoked caps to mds as
1515 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1516 * cap release further.
1517 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1518 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1521 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1522 struct ceph_mds_session *session)
1524 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1525 struct ceph_mds_client *mdsc = fsc->mdsc;
1526 struct inode *inode = &ci->vfs_inode;
1527 struct ceph_cap *cap;
1528 int file_wanted, used, cap_used;
1529 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1530 int issued, implemented, want, retain, revoking, flushing = 0;
1531 int mds = -1; /* keep track of how far we've gone through i_caps list
1532 to avoid an infinite loop on retry */
1534 int tried_invalidate = 0;
1535 int delayed = 0, sent = 0, force_requeue = 0, num;
1536 int queue_invalidate = 0;
1537 int is_delayed = flags & CHECK_CAPS_NODELAY;
1539 /* if we are unmounting, flush any unused caps immediately. */
1543 spin_lock(&ci->i_ceph_lock);
1545 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1546 flags |= CHECK_CAPS_FLUSH;
1548 /* flush snaps first time around only */
1549 if (!list_empty(&ci->i_cap_snaps))
1550 __ceph_flush_snaps(ci, &session, 0);
1553 spin_lock(&ci->i_ceph_lock);
1555 file_wanted = __ceph_caps_file_wanted(ci);
1556 used = __ceph_caps_used(ci);
1557 want = file_wanted | used;
1558 issued = __ceph_caps_issued(ci, &implemented);
1559 revoking = implemented & ~issued;
1561 retain = want | CEPH_CAP_PIN;
1562 if (!mdsc->stopping && inode->i_nlink > 0) {
1564 retain |= CEPH_CAP_ANY; /* be greedy */
1565 } else if (S_ISDIR(inode->i_mode) &&
1566 (issued & CEPH_CAP_FILE_SHARED) &&
1567 __ceph_dir_is_complete(ci)) {
1569 * If a directory is complete, we want to keep
1570 * the exclusive cap. So that MDS does not end up
1571 * revoking the shared cap on every create/unlink
1574 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1578 retain |= CEPH_CAP_ANY_SHARED;
1580 * keep RD only if we didn't have the file open RW,
1581 * because then the mds would revoke it anyway to
1582 * journal max_size=0.
1584 if (ci->i_max_size == 0)
1585 retain |= CEPH_CAP_ANY_RD;
1589 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1590 " issued %s revoking %s retain %s %s%s%s\n", inode,
1591 ceph_cap_string(file_wanted),
1592 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1593 ceph_cap_string(ci->i_flushing_caps),
1594 ceph_cap_string(issued), ceph_cap_string(revoking),
1595 ceph_cap_string(retain),
1596 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1597 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1598 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1601 * If we no longer need to hold onto old our caps, and we may
1602 * have cached pages, but don't want them, then try to invalidate.
1603 * If we fail, it's because pages are locked.... try again later.
1605 if ((!is_delayed || mdsc->stopping) &&
1606 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1607 inode->i_data.nrpages && /* have cached pages */
1608 (file_wanted == 0 || /* no open files */
1609 (revoking & (CEPH_CAP_FILE_CACHE|
1610 CEPH_CAP_FILE_LAZYIO))) && /* or revoking cache */
1611 !tried_invalidate) {
1612 dout("check_caps trying to invalidate on %p\n", inode);
1613 if (try_nonblocking_invalidate(inode) < 0) {
1614 if (revoking & (CEPH_CAP_FILE_CACHE|
1615 CEPH_CAP_FILE_LAZYIO)) {
1616 dout("check_caps queuing invalidate\n");
1617 queue_invalidate = 1;
1618 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1620 dout("check_caps failed to invalidate pages\n");
1621 /* we failed to invalidate pages. check these
1622 caps again later. */
1624 __cap_set_timeouts(mdsc, ci);
1627 tried_invalidate = 1;
1632 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1633 cap = rb_entry(p, struct ceph_cap, ci_node);
1636 /* avoid looping forever */
1637 if (mds >= cap->mds ||
1638 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1641 /* NOTE: no side-effects allowed, until we take s_mutex */
1644 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1645 cap_used &= ~ci->i_auth_cap->issued;
1647 revoking = cap->implemented & ~cap->issued;
1648 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1649 cap->mds, cap, ceph_cap_string(cap->issued),
1650 ceph_cap_string(cap_used),
1651 ceph_cap_string(cap->implemented),
1652 ceph_cap_string(revoking));
1654 if (cap == ci->i_auth_cap &&
1655 (cap->issued & CEPH_CAP_FILE_WR)) {
1656 /* request larger max_size from MDS? */
1657 if (ci->i_wanted_max_size > ci->i_max_size &&
1658 ci->i_wanted_max_size > ci->i_requested_max_size) {
1659 dout("requesting new max_size\n");
1663 /* approaching file_max? */
1664 if ((inode->i_size << 1) >= ci->i_max_size &&
1665 (ci->i_reported_size << 1) < ci->i_max_size) {
1666 dout("i_size approaching max_size\n");
1670 /* flush anything dirty? */
1671 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1673 dout("flushing dirty caps\n");
1677 /* completed revocation? going down and there are no caps? */
1678 if (revoking && (revoking & cap_used) == 0) {
1679 dout("completed revocation of %s\n",
1680 ceph_cap_string(cap->implemented & ~cap->issued));
1684 /* want more caps from mds? */
1685 if (want & ~(cap->mds_wanted | cap->issued))
1688 /* things we might delay */
1689 if ((cap->issued & ~retain) == 0 &&
1690 cap->mds_wanted == want)
1691 continue; /* nope, all good */
1697 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1698 time_before(jiffies, ci->i_hold_caps_max)) {
1699 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1700 ceph_cap_string(cap->issued),
1701 ceph_cap_string(cap->issued & retain),
1702 ceph_cap_string(cap->mds_wanted),
1703 ceph_cap_string(want));
1709 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1710 dout(" skipping %p I_NOFLUSH set\n", inode);
1714 if (session && session != cap->session) {
1715 dout("oops, wrong session %p mutex\n", session);
1716 mutex_unlock(&session->s_mutex);
1720 session = cap->session;
1721 if (mutex_trylock(&session->s_mutex) == 0) {
1722 dout("inverting session/ino locks on %p\n",
1724 spin_unlock(&ci->i_ceph_lock);
1725 if (took_snap_rwsem) {
1726 up_read(&mdsc->snap_rwsem);
1727 took_snap_rwsem = 0;
1729 mutex_lock(&session->s_mutex);
1733 /* take snap_rwsem after session mutex */
1734 if (!took_snap_rwsem) {
1735 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1736 dout("inverting snap/in locks on %p\n",
1738 spin_unlock(&ci->i_ceph_lock);
1739 down_read(&mdsc->snap_rwsem);
1740 took_snap_rwsem = 1;
1743 took_snap_rwsem = 1;
1746 if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1747 flushing = __mark_caps_flushing(inode, session);
1751 mds = cap->mds; /* remember mds, so we don't repeat */
1754 /* __send_cap drops i_ceph_lock */
1755 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1756 want, retain, flushing, NULL);
1757 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1761 * Reschedule delayed caps release if we delayed anything,
1764 if (delayed && is_delayed)
1765 force_requeue = 1; /* __send_cap delayed release; requeue */
1766 if (!delayed && !is_delayed)
1767 __cap_delay_cancel(mdsc, ci);
1768 else if (!is_delayed || force_requeue)
1769 __cap_delay_requeue(mdsc, ci);
1771 spin_unlock(&ci->i_ceph_lock);
1773 if (queue_invalidate)
1774 ceph_queue_invalidate(inode);
1777 mutex_unlock(&session->s_mutex);
1778 if (took_snap_rwsem)
1779 up_read(&mdsc->snap_rwsem);
1783 * Try to flush dirty caps back to the auth mds.
1785 static int try_flush_caps(struct inode *inode, unsigned *flush_tid)
1787 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1788 struct ceph_inode_info *ci = ceph_inode(inode);
1790 struct ceph_mds_session *session = NULL;
1793 spin_lock(&ci->i_ceph_lock);
1794 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1795 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1798 if (ci->i_dirty_caps && ci->i_auth_cap) {
1799 struct ceph_cap *cap = ci->i_auth_cap;
1800 int used = __ceph_caps_used(ci);
1801 int want = __ceph_caps_wanted(ci);
1804 if (!session || session != cap->session) {
1805 spin_unlock(&ci->i_ceph_lock);
1807 mutex_unlock(&session->s_mutex);
1808 session = cap->session;
1809 mutex_lock(&session->s_mutex);
1812 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1815 flushing = __mark_caps_flushing(inode, session);
1817 /* __send_cap drops i_ceph_lock */
1818 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1819 cap->issued | cap->implemented, flushing,
1824 spin_lock(&ci->i_ceph_lock);
1825 __cap_delay_requeue(mdsc, ci);
1828 spin_unlock(&ci->i_ceph_lock);
1831 mutex_unlock(&session->s_mutex);
1836 * Return true if we've flushed caps through the given flush_tid.
1838 static int caps_are_flushed(struct inode *inode, unsigned tid)
1840 struct ceph_inode_info *ci = ceph_inode(inode);
1843 spin_lock(&ci->i_ceph_lock);
1844 for (i = 0; i < CEPH_CAP_BITS; i++)
1845 if ((ci->i_flushing_caps & (1 << i)) &&
1846 ci->i_cap_flush_tid[i] <= tid) {
1847 /* still flushing this bit */
1851 spin_unlock(&ci->i_ceph_lock);
1856 * Wait on any unsafe replies for the given inode. First wait on the
1857 * newest request, and make that the upper bound. Then, if there are
1858 * more requests, keep waiting on the oldest as long as it is still older
1859 * than the original request.
1861 static void sync_write_wait(struct inode *inode)
1863 struct ceph_inode_info *ci = ceph_inode(inode);
1864 struct list_head *head = &ci->i_unsafe_writes;
1865 struct ceph_osd_request *req;
1868 spin_lock(&ci->i_unsafe_lock);
1869 if (list_empty(head))
1872 /* set upper bound as _last_ entry in chain */
1873 req = list_entry(head->prev, struct ceph_osd_request,
1875 last_tid = req->r_tid;
1878 ceph_osdc_get_request(req);
1879 spin_unlock(&ci->i_unsafe_lock);
1880 dout("sync_write_wait on tid %llu (until %llu)\n",
1881 req->r_tid, last_tid);
1882 wait_for_completion(&req->r_safe_completion);
1883 spin_lock(&ci->i_unsafe_lock);
1884 ceph_osdc_put_request(req);
1887 * from here on look at first entry in chain, since we
1888 * only want to wait for anything older than last_tid
1890 if (list_empty(head))
1892 req = list_entry(head->next, struct ceph_osd_request,
1894 } while (req->r_tid < last_tid);
1896 spin_unlock(&ci->i_unsafe_lock);
1899 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1901 struct inode *inode = file->f_mapping->host;
1902 struct ceph_inode_info *ci = ceph_inode(inode);
1907 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1908 sync_write_wait(inode);
1910 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1913 mutex_lock(&inode->i_mutex);
1915 dirty = try_flush_caps(inode, &flush_tid);
1916 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1919 * only wait on non-file metadata writeback (the mds
1920 * can recover size and mtime, so we don't need to
1923 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1924 dout("fsync waiting for flush_tid %u\n", flush_tid);
1925 ret = wait_event_interruptible(ci->i_cap_wq,
1926 caps_are_flushed(inode, flush_tid));
1929 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1930 mutex_unlock(&inode->i_mutex);
1935 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1936 * queue inode for flush but don't do so immediately, because we can
1937 * get by with fewer MDS messages if we wait for data writeback to
1940 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1942 struct ceph_inode_info *ci = ceph_inode(inode);
1946 int wait = wbc->sync_mode == WB_SYNC_ALL;
1948 dout("write_inode %p wait=%d\n", inode, wait);
1950 dirty = try_flush_caps(inode, &flush_tid);
1952 err = wait_event_interruptible(ci->i_cap_wq,
1953 caps_are_flushed(inode, flush_tid));
1955 struct ceph_mds_client *mdsc =
1956 ceph_sb_to_client(inode->i_sb)->mdsc;
1958 spin_lock(&ci->i_ceph_lock);
1959 if (__ceph_caps_dirty(ci))
1960 __cap_delay_requeue_front(mdsc, ci);
1961 spin_unlock(&ci->i_ceph_lock);
1967 * After a recovering MDS goes active, we need to resend any caps
1970 * Caller holds session->s_mutex.
1972 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1973 struct ceph_mds_session *session)
1975 struct ceph_cap_snap *capsnap;
1977 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1978 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1980 struct ceph_inode_info *ci = capsnap->ci;
1981 struct inode *inode = &ci->vfs_inode;
1982 struct ceph_cap *cap;
1984 spin_lock(&ci->i_ceph_lock);
1985 cap = ci->i_auth_cap;
1986 if (cap && cap->session == session) {
1987 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1989 __ceph_flush_snaps(ci, &session, 1);
1991 pr_err("%p auth cap %p not mds%d ???\n", inode,
1992 cap, session->s_mds);
1994 spin_unlock(&ci->i_ceph_lock);
1998 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1999 struct ceph_mds_session *session)
2001 struct ceph_inode_info *ci;
2003 kick_flushing_capsnaps(mdsc, session);
2005 dout("kick_flushing_caps mds%d\n", session->s_mds);
2006 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2007 struct inode *inode = &ci->vfs_inode;
2008 struct ceph_cap *cap;
2011 spin_lock(&ci->i_ceph_lock);
2012 cap = ci->i_auth_cap;
2013 if (cap && cap->session == session) {
2014 dout("kick_flushing_caps %p cap %p %s\n", inode,
2015 cap, ceph_cap_string(ci->i_flushing_caps));
2016 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2017 __ceph_caps_used(ci),
2018 __ceph_caps_wanted(ci),
2019 cap->issued | cap->implemented,
2020 ci->i_flushing_caps, NULL);
2022 spin_lock(&ci->i_ceph_lock);
2023 __cap_delay_requeue(mdsc, ci);
2024 spin_unlock(&ci->i_ceph_lock);
2027 pr_err("%p auth cap %p not mds%d ???\n", inode,
2028 cap, session->s_mds);
2029 spin_unlock(&ci->i_ceph_lock);
2034 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2035 struct ceph_mds_session *session,
2036 struct inode *inode)
2038 struct ceph_inode_info *ci = ceph_inode(inode);
2039 struct ceph_cap *cap;
2042 spin_lock(&ci->i_ceph_lock);
2043 cap = ci->i_auth_cap;
2044 dout("kick_flushing_inode_caps %p flushing %s flush_seq %lld\n", inode,
2045 ceph_cap_string(ci->i_flushing_caps), ci->i_cap_flush_seq);
2047 __ceph_flush_snaps(ci, &session, 1);
2049 if (ci->i_flushing_caps) {
2050 spin_lock(&mdsc->cap_dirty_lock);
2051 list_move_tail(&ci->i_flushing_item,
2052 &cap->session->s_cap_flushing);
2053 spin_unlock(&mdsc->cap_dirty_lock);
2055 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2056 __ceph_caps_used(ci),
2057 __ceph_caps_wanted(ci),
2058 cap->issued | cap->implemented,
2059 ci->i_flushing_caps, NULL);
2061 spin_lock(&ci->i_ceph_lock);
2062 __cap_delay_requeue(mdsc, ci);
2063 spin_unlock(&ci->i_ceph_lock);
2066 spin_unlock(&ci->i_ceph_lock);
2072 * Take references to capabilities we hold, so that we don't release
2073 * them to the MDS prematurely.
2075 * Protected by i_ceph_lock.
2077 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2078 bool snap_rwsem_locked)
2080 if (got & CEPH_CAP_PIN)
2082 if (got & CEPH_CAP_FILE_RD)
2084 if (got & CEPH_CAP_FILE_CACHE)
2085 ci->i_rdcache_ref++;
2086 if (got & CEPH_CAP_FILE_WR) {
2087 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2088 BUG_ON(!snap_rwsem_locked);
2089 ci->i_head_snapc = ceph_get_snap_context(
2090 ci->i_snap_realm->cached_context);
2094 if (got & CEPH_CAP_FILE_BUFFER) {
2095 if (ci->i_wb_ref == 0)
2096 ihold(&ci->vfs_inode);
2098 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2099 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2104 * Try to grab cap references. Specify those refs we @want, and the
2105 * minimal set we @need. Also include the larger offset we are writing
2106 * to (when applicable), and check against max_size here as well.
2107 * Note that caller is responsible for ensuring max_size increases are
2108 * requested from the MDS.
2110 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2111 loff_t endoff, bool nonblock, int *got, int *err)
2113 struct inode *inode = &ci->vfs_inode;
2114 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2116 int have, implemented;
2118 bool snap_rwsem_locked = false;
2120 dout("get_cap_refs %p need %s want %s\n", inode,
2121 ceph_cap_string(need), ceph_cap_string(want));
2124 spin_lock(&ci->i_ceph_lock);
2126 /* make sure file is actually open */
2127 file_wanted = __ceph_caps_file_wanted(ci);
2128 if ((file_wanted & need) == 0) {
2129 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2130 ceph_cap_string(need), ceph_cap_string(file_wanted));
2136 /* finish pending truncate */
2137 while (ci->i_truncate_pending) {
2138 spin_unlock(&ci->i_ceph_lock);
2139 if (snap_rwsem_locked) {
2140 up_read(&mdsc->snap_rwsem);
2141 snap_rwsem_locked = false;
2143 __ceph_do_pending_vmtruncate(inode);
2144 spin_lock(&ci->i_ceph_lock);
2147 have = __ceph_caps_issued(ci, &implemented);
2149 if (have & need & CEPH_CAP_FILE_WR) {
2150 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2151 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2152 inode, endoff, ci->i_max_size);
2153 if (endoff > ci->i_requested_max_size) {
2160 * If a sync write is in progress, we must wait, so that we
2161 * can get a final snapshot value for size+mtime.
2163 if (__ceph_have_pending_cap_snap(ci)) {
2164 dout("get_cap_refs %p cap_snap_pending\n", inode);
2169 if ((have & need) == need) {
2171 * Look at (implemented & ~have & not) so that we keep waiting
2172 * on transition from wanted -> needed caps. This is needed
2173 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2174 * going before a prior buffered writeback happens.
2176 int not = want & ~(have & need);
2177 int revoking = implemented & ~have;
2178 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2179 inode, ceph_cap_string(have), ceph_cap_string(not),
2180 ceph_cap_string(revoking));
2181 if ((revoking & not) == 0) {
2182 if (!snap_rwsem_locked &&
2183 !ci->i_head_snapc &&
2184 (need & CEPH_CAP_FILE_WR)) {
2185 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2187 * we can not call down_read() when
2188 * task isn't in TASK_RUNNING state
2196 spin_unlock(&ci->i_ceph_lock);
2197 down_read(&mdsc->snap_rwsem);
2198 snap_rwsem_locked = true;
2201 snap_rwsem_locked = true;
2203 *got = need | (have & want);
2204 __take_cap_refs(ci, *got, true);
2208 int session_readonly = false;
2209 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2210 struct ceph_mds_session *s = ci->i_auth_cap->session;
2211 spin_lock(&s->s_cap_lock);
2212 session_readonly = s->s_readonly;
2213 spin_unlock(&s->s_cap_lock);
2215 if (session_readonly) {
2216 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2217 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2223 dout("get_cap_refs %p have %s needed %s\n", inode,
2224 ceph_cap_string(have), ceph_cap_string(need));
2227 spin_unlock(&ci->i_ceph_lock);
2228 if (snap_rwsem_locked)
2229 up_read(&mdsc->snap_rwsem);
2231 dout("get_cap_refs %p ret %d got %s\n", inode,
2232 ret, ceph_cap_string(*got));
2237 * Check the offset we are writing up to against our current
2238 * max_size. If necessary, tell the MDS we want to write to
2241 static void check_max_size(struct inode *inode, loff_t endoff)
2243 struct ceph_inode_info *ci = ceph_inode(inode);
2246 /* do we need to explicitly request a larger max_size? */
2247 spin_lock(&ci->i_ceph_lock);
2248 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2249 dout("write %p at large endoff %llu, req max_size\n",
2251 ci->i_wanted_max_size = endoff;
2253 /* duplicate ceph_check_caps()'s logic */
2254 if (ci->i_auth_cap &&
2255 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2256 ci->i_wanted_max_size > ci->i_max_size &&
2257 ci->i_wanted_max_size > ci->i_requested_max_size)
2259 spin_unlock(&ci->i_ceph_lock);
2261 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2265 * Wait for caps, and take cap references. If we can't get a WR cap
2266 * due to a small max_size, make sure we check_max_size (and possibly
2267 * ask the mds) so we don't get hung up indefinitely.
2269 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2270 loff_t endoff, int *got, struct page **pinned_page)
2272 int _got, ret, err = 0;
2274 ret = ceph_pool_perm_check(ci, need);
2280 check_max_size(&ci->vfs_inode, endoff);
2284 ret = try_get_cap_refs(ci, need, want, endoff,
2285 false, &_got, &err);
2292 ret = wait_event_interruptible(ci->i_cap_wq,
2293 try_get_cap_refs(ci, need, want, endoff,
2294 true, &_got, &err));
2303 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2304 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2305 i_size_read(&ci->vfs_inode) > 0) {
2307 find_get_page(ci->vfs_inode.i_mapping, 0);
2309 if (PageUptodate(page)) {
2310 *pinned_page = page;
2313 page_cache_release(page);
2316 * drop cap refs first because getattr while
2317 * holding * caps refs can cause deadlock.
2319 ceph_put_cap_refs(ci, _got);
2323 * getattr request will bring inline data into
2326 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2327 CEPH_STAT_CAP_INLINE_DATA,
2341 * Take cap refs. Caller must already know we hold at least one ref
2342 * on the caps in question or we don't know this is safe.
2344 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2346 spin_lock(&ci->i_ceph_lock);
2347 __take_cap_refs(ci, caps, false);
2348 spin_unlock(&ci->i_ceph_lock);
2353 * drop cap_snap that is not associated with any snapshot.
2354 * we don't need to send FLUSHSNAP message for it.
2356 static int ceph_try_drop_cap_snap(struct ceph_cap_snap *capsnap)
2358 if (!capsnap->need_flush &&
2359 !capsnap->writing && !capsnap->dirty_pages) {
2361 dout("dropping cap_snap %p follows %llu\n",
2362 capsnap, capsnap->follows);
2363 ceph_put_snap_context(capsnap->context);
2364 list_del(&capsnap->ci_item);
2365 list_del(&capsnap->flushing_item);
2366 ceph_put_cap_snap(capsnap);
2375 * If we released the last ref on any given cap, call ceph_check_caps
2376 * to release (or schedule a release).
2378 * If we are releasing a WR cap (from a sync write), finalize any affected
2379 * cap_snap, and wake up any waiters.
2381 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2383 struct inode *inode = &ci->vfs_inode;
2384 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2386 spin_lock(&ci->i_ceph_lock);
2387 if (had & CEPH_CAP_PIN)
2389 if (had & CEPH_CAP_FILE_RD)
2390 if (--ci->i_rd_ref == 0)
2392 if (had & CEPH_CAP_FILE_CACHE)
2393 if (--ci->i_rdcache_ref == 0)
2395 if (had & CEPH_CAP_FILE_BUFFER) {
2396 if (--ci->i_wb_ref == 0) {
2400 dout("put_cap_refs %p wb %d -> %d (?)\n",
2401 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2403 if (had & CEPH_CAP_FILE_WR)
2404 if (--ci->i_wr_ref == 0) {
2406 if (__ceph_have_pending_cap_snap(ci)) {
2407 struct ceph_cap_snap *capsnap =
2408 list_last_entry(&ci->i_cap_snaps,
2409 struct ceph_cap_snap,
2411 capsnap->writing = 0;
2412 if (ceph_try_drop_cap_snap(capsnap))
2414 else if (__ceph_finish_cap_snap(ci, capsnap))
2418 if (ci->i_wrbuffer_ref_head == 0 &&
2419 ci->i_dirty_caps == 0 &&
2420 ci->i_flushing_caps == 0) {
2421 BUG_ON(!ci->i_head_snapc);
2422 ceph_put_snap_context(ci->i_head_snapc);
2423 ci->i_head_snapc = NULL;
2425 /* see comment in __ceph_remove_cap() */
2426 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2427 drop_inode_snap_realm(ci);
2429 spin_unlock(&ci->i_ceph_lock);
2431 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2432 last ? " last" : "", put ? " put" : "");
2434 if (last && !flushsnaps)
2435 ceph_check_caps(ci, 0, NULL);
2436 else if (flushsnaps)
2437 ceph_flush_snaps(ci);
2439 wake_up_all(&ci->i_cap_wq);
2445 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2446 * context. Adjust per-snap dirty page accounting as appropriate.
2447 * Once all dirty data for a cap_snap is flushed, flush snapped file
2448 * metadata back to the MDS. If we dropped the last ref, call
2451 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2452 struct ceph_snap_context *snapc)
2454 struct inode *inode = &ci->vfs_inode;
2456 int complete_capsnap = 0;
2457 int drop_capsnap = 0;
2459 struct ceph_cap_snap *capsnap = NULL;
2461 spin_lock(&ci->i_ceph_lock);
2462 ci->i_wrbuffer_ref -= nr;
2463 last = !ci->i_wrbuffer_ref;
2465 if (ci->i_head_snapc == snapc) {
2466 ci->i_wrbuffer_ref_head -= nr;
2467 if (ci->i_wrbuffer_ref_head == 0 &&
2468 ci->i_wr_ref == 0 &&
2469 ci->i_dirty_caps == 0 &&
2470 ci->i_flushing_caps == 0) {
2471 BUG_ON(!ci->i_head_snapc);
2472 ceph_put_snap_context(ci->i_head_snapc);
2473 ci->i_head_snapc = NULL;
2475 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2477 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2478 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2479 last ? " LAST" : "");
2481 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2482 if (capsnap->context == snapc) {
2488 capsnap->dirty_pages -= nr;
2489 if (capsnap->dirty_pages == 0) {
2490 complete_capsnap = 1;
2491 drop_capsnap = ceph_try_drop_cap_snap(capsnap);
2493 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2494 " snap %lld %d/%d -> %d/%d %s%s\n",
2495 inode, capsnap, capsnap->context->seq,
2496 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2497 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2498 last ? " (wrbuffer last)" : "",
2499 complete_capsnap ? " (complete capsnap)" : "");
2502 spin_unlock(&ci->i_ceph_lock);
2505 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2507 } else if (complete_capsnap) {
2508 ceph_flush_snaps(ci);
2509 wake_up_all(&ci->i_cap_wq);
2516 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2518 static void invalidate_aliases(struct inode *inode)
2520 struct dentry *dn, *prev = NULL;
2522 dout("invalidate_aliases inode %p\n", inode);
2523 d_prune_aliases(inode);
2525 * For non-directory inode, d_find_alias() only returns
2526 * hashed dentry. After calling d_invalidate(), the
2527 * dentry becomes unhashed.
2529 * For directory inode, d_find_alias() can return
2530 * unhashed dentry. But directory inode should have
2531 * one alias at most.
2533 while ((dn = d_find_alias(inode))) {
2548 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2549 * actually be a revocation if it specifies a smaller cap set.)
2551 * caller holds s_mutex and i_ceph_lock, we drop both.
2553 static void handle_cap_grant(struct ceph_mds_client *mdsc,
2554 struct inode *inode, struct ceph_mds_caps *grant,
2556 void *inline_data, int inline_len,
2557 struct ceph_buffer *xattr_buf,
2558 struct ceph_mds_session *session,
2559 struct ceph_cap *cap, int issued)
2560 __releases(ci->i_ceph_lock)
2561 __releases(mdsc->snap_rwsem)
2563 struct ceph_inode_info *ci = ceph_inode(inode);
2564 int mds = session->s_mds;
2565 int seq = le32_to_cpu(grant->seq);
2566 int newcaps = le32_to_cpu(grant->caps);
2567 int used, wanted, dirty;
2568 u64 size = le64_to_cpu(grant->size);
2569 u64 max_size = le64_to_cpu(grant->max_size);
2570 struct timespec mtime, atime, ctime;
2573 bool writeback = false;
2574 bool queue_trunc = false;
2575 bool queue_invalidate = false;
2576 bool queue_revalidate = false;
2577 bool deleted_inode = false;
2578 bool fill_inline = false;
2580 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2581 inode, cap, mds, seq, ceph_cap_string(newcaps));
2582 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2587 * auth mds of the inode changed. we received the cap export message,
2588 * but still haven't received the cap import message. handle_cap_export
2589 * updated the new auth MDS' cap.
2591 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2592 * that was sent before the cap import message. So don't remove caps.
2594 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2595 WARN_ON(cap != ci->i_auth_cap);
2596 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2598 newcaps |= cap->issued;
2602 * If CACHE is being revoked, and we have no dirty buffers,
2603 * try to invalidate (once). (If there are dirty buffers, we
2604 * will invalidate _after_ writeback.)
2606 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2607 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2608 !ci->i_wrbuffer_ref) {
2609 if (try_nonblocking_invalidate(inode)) {
2610 /* there were locked pages.. invalidate later
2611 in a separate thread. */
2612 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2613 queue_invalidate = true;
2614 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2618 ceph_fscache_invalidate(inode);
2621 /* side effects now are allowed */
2622 cap->cap_gen = session->s_cap_gen;
2625 __check_cap_issue(ci, cap, newcaps);
2627 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2628 (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2629 inode->i_mode = le32_to_cpu(grant->mode);
2630 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2631 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2632 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2633 from_kuid(&init_user_ns, inode->i_uid),
2634 from_kgid(&init_user_ns, inode->i_gid));
2637 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2638 (issued & CEPH_CAP_LINK_EXCL) == 0) {
2639 set_nlink(inode, le32_to_cpu(grant->nlink));
2640 if (inode->i_nlink == 0 &&
2641 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2642 deleted_inode = true;
2645 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2646 int len = le32_to_cpu(grant->xattr_len);
2647 u64 version = le64_to_cpu(grant->xattr_version);
2649 if (version > ci->i_xattrs.version) {
2650 dout(" got new xattrs v%llu on %p len %d\n",
2651 version, inode, len);
2652 if (ci->i_xattrs.blob)
2653 ceph_buffer_put(ci->i_xattrs.blob);
2654 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2655 ci->i_xattrs.version = version;
2656 ceph_forget_all_cached_acls(inode);
2660 /* Do we need to revalidate our fscache cookie. Don't bother on the
2661 * first cache cap as we already validate at cookie creation time. */
2662 if ((issued & CEPH_CAP_FILE_CACHE) && ci->i_rdcache_gen > 1)
2663 queue_revalidate = true;
2665 if (newcaps & CEPH_CAP_ANY_RD) {
2666 /* ctime/mtime/atime? */
2667 ceph_decode_timespec(&mtime, &grant->mtime);
2668 ceph_decode_timespec(&atime, &grant->atime);
2669 ceph_decode_timespec(&ctime, &grant->ctime);
2670 ceph_fill_file_time(inode, issued,
2671 le32_to_cpu(grant->time_warp_seq),
2672 &ctime, &mtime, &atime);
2675 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
2676 /* file layout may have changed */
2677 ci->i_layout = grant->layout;
2678 /* size/truncate_seq? */
2679 queue_trunc = ceph_fill_file_size(inode, issued,
2680 le32_to_cpu(grant->truncate_seq),
2681 le64_to_cpu(grant->truncate_size),
2683 /* max size increase? */
2684 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
2685 dout("max_size %lld -> %llu\n",
2686 ci->i_max_size, max_size);
2687 ci->i_max_size = max_size;
2688 if (max_size >= ci->i_wanted_max_size) {
2689 ci->i_wanted_max_size = 0; /* reset */
2690 ci->i_requested_max_size = 0;
2696 /* check cap bits */
2697 wanted = __ceph_caps_wanted(ci);
2698 used = __ceph_caps_used(ci);
2699 dirty = __ceph_caps_dirty(ci);
2700 dout(" my wanted = %s, used = %s, dirty %s\n",
2701 ceph_cap_string(wanted),
2702 ceph_cap_string(used),
2703 ceph_cap_string(dirty));
2704 if (wanted != le32_to_cpu(grant->wanted)) {
2705 dout("mds wanted %s -> %s\n",
2706 ceph_cap_string(le32_to_cpu(grant->wanted)),
2707 ceph_cap_string(wanted));
2708 /* imported cap may not have correct mds_wanted */
2709 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
2713 /* revocation, grant, or no-op? */
2714 if (cap->issued & ~newcaps) {
2715 int revoking = cap->issued & ~newcaps;
2717 dout("revocation: %s -> %s (revoking %s)\n",
2718 ceph_cap_string(cap->issued),
2719 ceph_cap_string(newcaps),
2720 ceph_cap_string(revoking));
2721 if (revoking & used & CEPH_CAP_FILE_BUFFER)
2722 writeback = true; /* initiate writeback; will delay ack */
2723 else if (revoking == CEPH_CAP_FILE_CACHE &&
2724 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2726 ; /* do nothing yet, invalidation will be queued */
2727 else if (cap == ci->i_auth_cap)
2728 check_caps = 1; /* check auth cap only */
2730 check_caps = 2; /* check all caps */
2731 cap->issued = newcaps;
2732 cap->implemented |= newcaps;
2733 } else if (cap->issued == newcaps) {
2734 dout("caps unchanged: %s -> %s\n",
2735 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2737 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2738 ceph_cap_string(newcaps));
2739 /* non-auth MDS is revoking the newly grant caps ? */
2740 if (cap == ci->i_auth_cap &&
2741 __ceph_caps_revoking_other(ci, cap, newcaps))
2744 cap->issued = newcaps;
2745 cap->implemented |= newcaps; /* add bits only, to
2746 * avoid stepping on a
2747 * pending revocation */
2750 BUG_ON(cap->issued & ~cap->implemented);
2752 if (inline_version > 0 && inline_version >= ci->i_inline_version) {
2753 ci->i_inline_version = inline_version;
2754 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2755 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
2759 spin_unlock(&ci->i_ceph_lock);
2761 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
2762 kick_flushing_inode_caps(mdsc, session, inode);
2763 up_read(&mdsc->snap_rwsem);
2764 if (newcaps & ~issued)
2769 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
2772 ceph_queue_vmtruncate(inode);
2773 ceph_queue_revalidate(inode);
2774 } else if (queue_revalidate)
2775 ceph_queue_revalidate(inode);
2779 * queue inode for writeback: we can't actually call
2780 * filemap_write_and_wait, etc. from message handler
2783 ceph_queue_writeback(inode);
2784 if (queue_invalidate)
2785 ceph_queue_invalidate(inode);
2787 invalidate_aliases(inode);
2789 wake_up_all(&ci->i_cap_wq);
2791 if (check_caps == 1)
2792 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2794 else if (check_caps == 2)
2795 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2797 mutex_unlock(&session->s_mutex);
2801 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2802 * MDS has been safely committed.
2804 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2805 struct ceph_mds_caps *m,
2806 struct ceph_mds_session *session,
2807 struct ceph_cap *cap)
2808 __releases(ci->i_ceph_lock)
2810 struct ceph_inode_info *ci = ceph_inode(inode);
2811 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2812 unsigned seq = le32_to_cpu(m->seq);
2813 int dirty = le32_to_cpu(m->dirty);
2818 for (i = 0; i < CEPH_CAP_BITS; i++)
2819 if ((dirty & (1 << i)) &&
2820 (u16)flush_tid == ci->i_cap_flush_tid[i])
2823 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2824 " flushing %s -> %s\n",
2825 inode, session->s_mds, seq, ceph_cap_string(dirty),
2826 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2827 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2829 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2832 ci->i_flushing_caps &= ~cleaned;
2834 spin_lock(&mdsc->cap_dirty_lock);
2835 if (ci->i_flushing_caps == 0) {
2836 list_del_init(&ci->i_flushing_item);
2837 if (!list_empty(&session->s_cap_flushing))
2838 dout(" mds%d still flushing cap on %p\n",
2840 &list_entry(session->s_cap_flushing.next,
2841 struct ceph_inode_info,
2842 i_flushing_item)->vfs_inode);
2843 mdsc->num_cap_flushing--;
2844 wake_up_all(&mdsc->cap_flushing_wq);
2845 dout(" inode %p now !flushing\n", inode);
2847 if (ci->i_dirty_caps == 0) {
2848 dout(" inode %p now clean\n", inode);
2849 BUG_ON(!list_empty(&ci->i_dirty_item));
2851 if (ci->i_wr_ref == 0 &&
2852 ci->i_wrbuffer_ref_head == 0) {
2853 BUG_ON(!ci->i_head_snapc);
2854 ceph_put_snap_context(ci->i_head_snapc);
2855 ci->i_head_snapc = NULL;
2858 BUG_ON(list_empty(&ci->i_dirty_item));
2861 spin_unlock(&mdsc->cap_dirty_lock);
2862 wake_up_all(&ci->i_cap_wq);
2865 spin_unlock(&ci->i_ceph_lock);
2871 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
2872 * throw away our cap_snap.
2874 * Caller hold s_mutex.
2876 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2877 struct ceph_mds_caps *m,
2878 struct ceph_mds_session *session)
2880 struct ceph_inode_info *ci = ceph_inode(inode);
2881 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2882 u64 follows = le64_to_cpu(m->snap_follows);
2883 struct ceph_cap_snap *capsnap;
2886 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2887 inode, ci, session->s_mds, follows);
2889 spin_lock(&ci->i_ceph_lock);
2890 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2891 if (capsnap->follows == follows) {
2892 if (capsnap->flush_tid != flush_tid) {
2893 dout(" cap_snap %p follows %lld tid %lld !="
2894 " %lld\n", capsnap, follows,
2895 flush_tid, capsnap->flush_tid);
2898 WARN_ON(capsnap->dirty_pages || capsnap->writing);
2899 dout(" removing %p cap_snap %p follows %lld\n",
2900 inode, capsnap, follows);
2901 ceph_put_snap_context(capsnap->context);
2902 list_del(&capsnap->ci_item);
2903 list_del(&capsnap->flushing_item);
2904 ceph_put_cap_snap(capsnap);
2905 wake_up_all(&mdsc->cap_flushing_wq);
2909 dout(" skipping cap_snap %p follows %lld\n",
2910 capsnap, capsnap->follows);
2913 spin_unlock(&ci->i_ceph_lock);
2919 * Handle TRUNC from MDS, indicating file truncation.
2921 * caller hold s_mutex.
2923 static void handle_cap_trunc(struct inode *inode,
2924 struct ceph_mds_caps *trunc,
2925 struct ceph_mds_session *session)
2926 __releases(ci->i_ceph_lock)
2928 struct ceph_inode_info *ci = ceph_inode(inode);
2929 int mds = session->s_mds;
2930 int seq = le32_to_cpu(trunc->seq);
2931 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2932 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2933 u64 size = le64_to_cpu(trunc->size);
2934 int implemented = 0;
2935 int dirty = __ceph_caps_dirty(ci);
2936 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2937 int queue_trunc = 0;
2939 issued |= implemented | dirty;
2941 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2942 inode, mds, seq, truncate_size, truncate_seq);
2943 queue_trunc = ceph_fill_file_size(inode, issued,
2944 truncate_seq, truncate_size, size);
2945 spin_unlock(&ci->i_ceph_lock);
2948 ceph_queue_vmtruncate(inode);
2949 ceph_fscache_invalidate(inode);
2954 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
2955 * different one. If we are the most recent migration we've seen (as
2956 * indicated by mseq), make note of the migrating cap bits for the
2957 * duration (until we see the corresponding IMPORT).
2959 * caller holds s_mutex
2961 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2962 struct ceph_mds_cap_peer *ph,
2963 struct ceph_mds_session *session)
2965 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2966 struct ceph_mds_session *tsession = NULL;
2967 struct ceph_cap *cap, *tcap, *new_cap = NULL;
2968 struct ceph_inode_info *ci = ceph_inode(inode);
2970 unsigned mseq = le32_to_cpu(ex->migrate_seq);
2971 unsigned t_seq, t_mseq;
2973 int mds = session->s_mds;
2976 t_cap_id = le64_to_cpu(ph->cap_id);
2977 t_seq = le32_to_cpu(ph->seq);
2978 t_mseq = le32_to_cpu(ph->mseq);
2979 target = le32_to_cpu(ph->mds);
2981 t_cap_id = t_seq = t_mseq = 0;
2985 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
2986 inode, ci, mds, mseq, target);
2988 spin_lock(&ci->i_ceph_lock);
2989 cap = __get_cap_for_mds(ci, mds);
2990 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
2994 __ceph_remove_cap(cap, false);
2999 * now we know we haven't received the cap import message yet
3000 * because the exported cap still exist.
3003 issued = cap->issued;
3004 WARN_ON(issued != cap->implemented);
3006 tcap = __get_cap_for_mds(ci, target);
3008 /* already have caps from the target */
3009 if (tcap->cap_id != t_cap_id ||
3010 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3011 dout(" updating import cap %p mds%d\n", tcap, target);
3012 tcap->cap_id = t_cap_id;
3013 tcap->seq = t_seq - 1;
3014 tcap->issue_seq = t_seq - 1;
3015 tcap->mseq = t_mseq;
3016 tcap->issued |= issued;
3017 tcap->implemented |= issued;
3018 if (cap == ci->i_auth_cap)
3019 ci->i_auth_cap = tcap;
3020 if (ci->i_flushing_caps && ci->i_auth_cap == tcap) {
3021 spin_lock(&mdsc->cap_dirty_lock);
3022 list_move_tail(&ci->i_flushing_item,
3023 &tcap->session->s_cap_flushing);
3024 spin_unlock(&mdsc->cap_dirty_lock);
3027 __ceph_remove_cap(cap, false);
3029 } else if (tsession) {
3030 /* add placeholder for the export tagert */
3031 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3032 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3033 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3035 __ceph_remove_cap(cap, false);
3039 spin_unlock(&ci->i_ceph_lock);
3040 mutex_unlock(&session->s_mutex);
3042 /* open target session */
3043 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3044 if (!IS_ERR(tsession)) {
3046 mutex_lock(&session->s_mutex);
3047 mutex_lock_nested(&tsession->s_mutex,
3048 SINGLE_DEPTH_NESTING);
3050 mutex_lock(&tsession->s_mutex);
3051 mutex_lock_nested(&session->s_mutex,
3052 SINGLE_DEPTH_NESTING);
3054 ceph_add_cap_releases(mdsc, tsession);
3055 new_cap = ceph_get_cap(mdsc, NULL);
3064 spin_unlock(&ci->i_ceph_lock);
3065 mutex_unlock(&session->s_mutex);
3067 mutex_unlock(&tsession->s_mutex);
3068 ceph_put_mds_session(tsession);
3071 ceph_put_cap(mdsc, new_cap);
3075 * Handle cap IMPORT.
3077 * caller holds s_mutex. acquires i_ceph_lock
3079 static void handle_cap_import(struct ceph_mds_client *mdsc,
3080 struct inode *inode, struct ceph_mds_caps *im,
3081 struct ceph_mds_cap_peer *ph,
3082 struct ceph_mds_session *session,
3083 struct ceph_cap **target_cap, int *old_issued)
3084 __acquires(ci->i_ceph_lock)
3086 struct ceph_inode_info *ci = ceph_inode(inode);
3087 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3088 int mds = session->s_mds;
3090 unsigned caps = le32_to_cpu(im->caps);
3091 unsigned wanted = le32_to_cpu(im->wanted);
3092 unsigned seq = le32_to_cpu(im->seq);
3093 unsigned mseq = le32_to_cpu(im->migrate_seq);
3094 u64 realmino = le64_to_cpu(im->realm);
3095 u64 cap_id = le64_to_cpu(im->cap_id);
3100 p_cap_id = le64_to_cpu(ph->cap_id);
3101 peer = le32_to_cpu(ph->mds);
3107 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3108 inode, ci, mds, mseq, peer);
3111 spin_lock(&ci->i_ceph_lock);
3112 cap = __get_cap_for_mds(ci, mds);
3115 spin_unlock(&ci->i_ceph_lock);
3116 new_cap = ceph_get_cap(mdsc, NULL);
3122 ceph_put_cap(mdsc, new_cap);
3127 __ceph_caps_issued(ci, &issued);
3128 issued |= __ceph_caps_dirty(ci);
3130 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3131 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3133 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3134 if (ocap && ocap->cap_id == p_cap_id) {
3135 dout(" remove export cap %p mds%d flags %d\n",
3136 ocap, peer, ph->flags);
3137 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3138 (ocap->seq != le32_to_cpu(ph->seq) ||
3139 ocap->mseq != le32_to_cpu(ph->mseq))) {
3140 pr_err("handle_cap_import: mismatched seq/mseq: "
3141 "ino (%llx.%llx) mds%d seq %d mseq %d "
3142 "importer mds%d has peer seq %d mseq %d\n",
3143 ceph_vinop(inode), peer, ocap->seq,
3144 ocap->mseq, mds, le32_to_cpu(ph->seq),
3145 le32_to_cpu(ph->mseq));
3147 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3150 /* make sure we re-request max_size, if necessary */
3151 ci->i_wanted_max_size = 0;
3152 ci->i_requested_max_size = 0;
3154 *old_issued = issued;
3159 * Handle a caps message from the MDS.
3161 * Identify the appropriate session, inode, and call the right handler
3162 * based on the cap op.
3164 void ceph_handle_caps(struct ceph_mds_session *session,
3165 struct ceph_msg *msg)
3167 struct ceph_mds_client *mdsc = session->s_mdsc;
3168 struct super_block *sb = mdsc->fsc->sb;
3169 struct inode *inode;
3170 struct ceph_inode_info *ci;
3171 struct ceph_cap *cap;
3172 struct ceph_mds_caps *h;
3173 struct ceph_mds_cap_peer *peer = NULL;
3174 struct ceph_snap_realm *realm;
3175 int mds = session->s_mds;
3178 struct ceph_vino vino;
3182 u64 inline_version = 0;
3183 void *inline_data = NULL;
3186 size_t snaptrace_len;
3189 dout("handle_caps from mds%d\n", mds);
3192 end = msg->front.iov_base + msg->front.iov_len;
3193 tid = le64_to_cpu(msg->hdr.tid);
3194 if (msg->front.iov_len < sizeof(*h))
3196 h = msg->front.iov_base;
3197 op = le32_to_cpu(h->op);
3198 vino.ino = le64_to_cpu(h->ino);
3199 vino.snap = CEPH_NOSNAP;
3200 cap_id = le64_to_cpu(h->cap_id);
3201 seq = le32_to_cpu(h->seq);
3202 mseq = le32_to_cpu(h->migrate_seq);
3203 size = le64_to_cpu(h->size);
3204 max_size = le64_to_cpu(h->max_size);
3207 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3208 p = snaptrace + snaptrace_len;
3210 if (le16_to_cpu(msg->hdr.version) >= 2) {
3212 ceph_decode_32_safe(&p, end, flock_len, bad);
3213 if (p + flock_len > end)
3218 if (le16_to_cpu(msg->hdr.version) >= 3) {
3219 if (op == CEPH_CAP_OP_IMPORT) {
3220 if (p + sizeof(*peer) > end)
3224 } else if (op == CEPH_CAP_OP_EXPORT) {
3225 /* recorded in unused fields */
3226 peer = (void *)&h->size;
3230 if (le16_to_cpu(msg->hdr.version) >= 4) {
3231 ceph_decode_64_safe(&p, end, inline_version, bad);
3232 ceph_decode_32_safe(&p, end, inline_len, bad);
3233 if (p + inline_len > end)
3240 inode = ceph_find_inode(sb, vino);
3241 ci = ceph_inode(inode);
3242 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3245 mutex_lock(&session->s_mutex);
3247 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3250 if (op == CEPH_CAP_OP_IMPORT)
3251 ceph_add_cap_releases(mdsc, session);
3254 dout(" i don't have ino %llx\n", vino.ino);
3256 if (op == CEPH_CAP_OP_IMPORT) {
3257 spin_lock(&session->s_cap_lock);
3258 __queue_cap_release(session, vino.ino, cap_id,
3260 spin_unlock(&session->s_cap_lock);
3262 goto flush_cap_releases;
3265 /* these will work even if we don't have a cap yet */
3267 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3268 handle_cap_flushsnap_ack(inode, tid, h, session);
3271 case CEPH_CAP_OP_EXPORT:
3272 handle_cap_export(inode, h, peer, session);
3275 case CEPH_CAP_OP_IMPORT:
3277 if (snaptrace_len) {
3278 down_write(&mdsc->snap_rwsem);
3279 ceph_update_snap_trace(mdsc, snaptrace,
3280 snaptrace + snaptrace_len,
3282 downgrade_write(&mdsc->snap_rwsem);
3284 down_read(&mdsc->snap_rwsem);
3286 handle_cap_import(mdsc, inode, h, peer, session,
3288 handle_cap_grant(mdsc, inode, h,
3289 inline_version, inline_data, inline_len,
3290 msg->middle, session, cap, issued);
3292 ceph_put_snap_realm(mdsc, realm);
3296 /* the rest require a cap */
3297 spin_lock(&ci->i_ceph_lock);
3298 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3300 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3301 inode, ceph_ino(inode), ceph_snap(inode), mds);
3302 spin_unlock(&ci->i_ceph_lock);
3303 goto flush_cap_releases;
3306 /* note that each of these drops i_ceph_lock for us */
3308 case CEPH_CAP_OP_REVOKE:
3309 case CEPH_CAP_OP_GRANT:
3310 __ceph_caps_issued(ci, &issued);
3311 issued |= __ceph_caps_dirty(ci);
3312 handle_cap_grant(mdsc, inode, h,
3313 inline_version, inline_data, inline_len,
3314 msg->middle, session, cap, issued);
3317 case CEPH_CAP_OP_FLUSH_ACK:
3318 handle_cap_flush_ack(inode, tid, h, session, cap);
3321 case CEPH_CAP_OP_TRUNC:
3322 handle_cap_trunc(inode, h, session);
3326 spin_unlock(&ci->i_ceph_lock);
3327 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3328 ceph_cap_op_name(op));
3335 * send any full release message to try to move things
3336 * along for the mds (who clearly thinks we still have this
3339 ceph_add_cap_releases(mdsc, session);
3340 ceph_send_cap_releases(mdsc, session);
3343 mutex_unlock(&session->s_mutex);
3349 pr_err("ceph_handle_caps: corrupt message\n");
3355 * Delayed work handler to process end of delayed cap release LRU list.
3357 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3359 struct ceph_inode_info *ci;
3360 int flags = CHECK_CAPS_NODELAY;
3362 dout("check_delayed_caps\n");
3364 spin_lock(&mdsc->cap_delay_lock);
3365 if (list_empty(&mdsc->cap_delay_list))
3367 ci = list_first_entry(&mdsc->cap_delay_list,
3368 struct ceph_inode_info,
3370 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3371 time_before(jiffies, ci->i_hold_caps_max))
3373 list_del_init(&ci->i_cap_delay_list);
3374 spin_unlock(&mdsc->cap_delay_lock);
3375 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3376 ceph_check_caps(ci, flags, NULL);
3378 spin_unlock(&mdsc->cap_delay_lock);
3382 * Flush all dirty caps to the mds
3384 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3386 struct ceph_inode_info *ci;
3387 struct inode *inode;
3389 dout("flush_dirty_caps\n");
3390 spin_lock(&mdsc->cap_dirty_lock);
3391 while (!list_empty(&mdsc->cap_dirty)) {
3392 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3394 inode = &ci->vfs_inode;
3396 dout("flush_dirty_caps %p\n", inode);
3397 spin_unlock(&mdsc->cap_dirty_lock);
3398 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3400 spin_lock(&mdsc->cap_dirty_lock);
3402 spin_unlock(&mdsc->cap_dirty_lock);
3403 dout("flush_dirty_caps done\n");
3407 * Drop open file reference. If we were the last open file,
3408 * we may need to release capabilities to the MDS (or schedule
3409 * their delayed release).
3411 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3413 struct inode *inode = &ci->vfs_inode;
3416 spin_lock(&ci->i_ceph_lock);
3417 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
3418 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
3419 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
3420 if (--ci->i_nr_by_mode[fmode] == 0)
3422 spin_unlock(&ci->i_ceph_lock);
3424 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3425 ceph_check_caps(ci, 0, NULL);
3429 * Helpers for embedding cap and dentry lease releases into mds
3432 * @force is used by dentry_release (below) to force inclusion of a
3433 * record for the directory inode, even when there aren't any caps to
3436 int ceph_encode_inode_release(void **p, struct inode *inode,
3437 int mds, int drop, int unless, int force)
3439 struct ceph_inode_info *ci = ceph_inode(inode);
3440 struct ceph_cap *cap;
3441 struct ceph_mds_request_release *rel = *p;
3445 spin_lock(&ci->i_ceph_lock);
3446 used = __ceph_caps_used(ci);
3447 dirty = __ceph_caps_dirty(ci);
3449 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3450 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3451 ceph_cap_string(unless));
3453 /* only drop unused, clean caps */
3454 drop &= ~(used | dirty);
3456 cap = __get_cap_for_mds(ci, mds);
3457 if (cap && __cap_is_valid(cap)) {
3459 ((cap->issued & drop) &&
3460 (cap->issued & unless) == 0)) {
3461 if ((cap->issued & drop) &&
3462 (cap->issued & unless) == 0) {
3463 int wanted = __ceph_caps_wanted(ci);
3464 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3465 wanted |= cap->mds_wanted;
3466 dout("encode_inode_release %p cap %p "
3467 "%s -> %s, wanted %s -> %s\n", inode, cap,
3468 ceph_cap_string(cap->issued),
3469 ceph_cap_string(cap->issued & ~drop),
3470 ceph_cap_string(cap->mds_wanted),
3471 ceph_cap_string(wanted));
3473 cap->issued &= ~drop;
3474 cap->implemented &= ~drop;
3475 cap->mds_wanted = wanted;
3477 dout("encode_inode_release %p cap %p %s"
3478 " (force)\n", inode, cap,
3479 ceph_cap_string(cap->issued));
3482 rel->ino = cpu_to_le64(ceph_ino(inode));
3483 rel->cap_id = cpu_to_le64(cap->cap_id);
3484 rel->seq = cpu_to_le32(cap->seq);
3485 rel->issue_seq = cpu_to_le32(cap->issue_seq);
3486 rel->mseq = cpu_to_le32(cap->mseq);
3487 rel->caps = cpu_to_le32(cap->implemented);
3488 rel->wanted = cpu_to_le32(cap->mds_wanted);
3494 dout("encode_inode_release %p cap %p %s\n",
3495 inode, cap, ceph_cap_string(cap->issued));
3498 spin_unlock(&ci->i_ceph_lock);
3502 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3503 int mds, int drop, int unless)
3505 struct inode *dir = d_inode(dentry->d_parent);
3506 struct ceph_mds_request_release *rel = *p;
3507 struct ceph_dentry_info *di = ceph_dentry(dentry);
3512 * force an record for the directory caps if we have a dentry lease.
3513 * this is racy (can't take i_ceph_lock and d_lock together), but it
3514 * doesn't have to be perfect; the mds will revoke anything we don't
3517 spin_lock(&dentry->d_lock);
3518 if (di->lease_session && di->lease_session->s_mds == mds)
3520 spin_unlock(&dentry->d_lock);
3522 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3524 spin_lock(&dentry->d_lock);
3525 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3526 dout("encode_dentry_release %p mds%d seq %d\n",
3527 dentry, mds, (int)di->lease_seq);
3528 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3529 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3530 *p += dentry->d_name.len;
3531 rel->dname_seq = cpu_to_le32(di->lease_seq);
3532 __ceph_mdsc_drop_dentry_lease(dentry);
3534 spin_unlock(&dentry->d_lock);