~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/ceph/caps.c

Version: ~ [ linux-5.15-rc7 ] ~ [ linux-5.14.14 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.75 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.155 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.213 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.252 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.287 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.289 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 #include <linux/ceph/ceph_debug.h>
  2 
  3 #include <linux/fs.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>
 10 
 11 #include "super.h"
 12 #include "mds_client.h"
 13 #include "cache.h"
 14 #include <linux/ceph/decode.h>
 15 #include <linux/ceph/messenger.h>
 16 
 17 /*
 18  * Capability management
 19  *
 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.
 25  *
 26  * If the client holds a *_SHARED cap, the client has a coherent value
 27  * that can be safely read from the cached inode.
 28  *
 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.
 33  *
 34  * In the event of a conflicting operation (perhaps by another
 35  * client), the MDS will revoke the conflicting client capabilities.
 36  *
 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.
 41  */
 42 
 43 
 44 /*
 45  * Generate readable cap strings for debugging output.
 46  */
 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;
 51 
 52 static char *gcap_string(char *s, int c)
 53 {
 54         if (c & CEPH_CAP_GSHARED)
 55                 *s++ = 's';
 56         if (c & CEPH_CAP_GEXCL)
 57                 *s++ = 'x';
 58         if (c & CEPH_CAP_GCACHE)
 59                 *s++ = 'c';
 60         if (c & CEPH_CAP_GRD)
 61                 *s++ = 'r';
 62         if (c & CEPH_CAP_GWR)
 63                 *s++ = 'w';
 64         if (c & CEPH_CAP_GBUFFER)
 65                 *s++ = 'b';
 66         if (c & CEPH_CAP_GLAZYIO)
 67                 *s++ = 'l';
 68         return s;
 69 }
 70 
 71 const char *ceph_cap_string(int caps)
 72 {
 73         int i;
 74         char *s;
 75         int c;
 76 
 77         spin_lock(&cap_str_lock);
 78         i = last_cap_str++;
 79         if (last_cap_str == MAX_CAP_STR)
 80                 last_cap_str = 0;
 81         spin_unlock(&cap_str_lock);
 82 
 83         s = cap_str[i];
 84 
 85         if (caps & CEPH_CAP_PIN)
 86                 *s++ = 'p';
 87 
 88         c = (caps >> CEPH_CAP_SAUTH) & 3;
 89         if (c) {
 90                 *s++ = 'A';
 91                 s = gcap_string(s, c);
 92         }
 93 
 94         c = (caps >> CEPH_CAP_SLINK) & 3;
 95         if (c) {
 96                 *s++ = 'L';
 97                 s = gcap_string(s, c);
 98         }
 99 
100         c = (caps >> CEPH_CAP_SXATTR) & 3;
101         if (c) {
102                 *s++ = 'X';
103                 s = gcap_string(s, c);
104         }
105 
106         c = caps >> CEPH_CAP_SFILE;
107         if (c) {
108                 *s++ = 'F';
109                 s = gcap_string(s, c);
110         }
111 
112         if (s == cap_str[i])
113                 *s++ = '-';
114         *s = 0;
115         return cap_str[i];
116 }
117 
118 void ceph_caps_init(struct ceph_mds_client *mdsc)
119 {
120         INIT_LIST_HEAD(&mdsc->caps_list);
121         spin_lock_init(&mdsc->caps_list_lock);
122 }
123 
124 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
125 {
126         struct ceph_cap *cap;
127 
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);
134         }
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);
141 }
142 
143 void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
144 {
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);
149 }
150 
151 void ceph_reserve_caps(struct ceph_mds_client *mdsc,
152                       struct ceph_cap_reservation *ctx, int need)
153 {
154         int i;
155         struct ceph_cap *cap;
156         int have;
157         int alloc = 0;
158         LIST_HEAD(newcaps);
159 
160         dout("reserve caps ctx=%p need=%d\n", ctx, need);
161 
162         /* first reserve any caps that are already allocated */
163         spin_lock(&mdsc->caps_list_lock);
164         if (mdsc->caps_avail_count >= need)
165                 have = need;
166         else
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);
174 
175         for (i = have; i < need; i++) {
176                 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
177                 if (!cap)
178                         break;
179                 list_add(&cap->caps_item, &newcaps);
180                 alloc++;
181         }
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);
186 
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);
191 
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);
196 
197         ctx->count = need;
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);
201 }
202 
203 int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
204                         struct ceph_cap_reservation *ctx)
205 {
206         dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
207         if (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;
212                 ctx->count = 0;
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);
220         }
221         return 0;
222 }
223 
224 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
225                               struct ceph_cap_reservation *ctx)
226 {
227         struct ceph_cap *cap = NULL;
228 
229         /* temporary, until we do something about cap import/export */
230         if (!ctx) {
231                 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
232                 if (cap) {
233                         spin_lock(&mdsc->caps_list_lock);
234                         mdsc->caps_use_count++;
235                         mdsc->caps_total_count++;
236                         spin_unlock(&mdsc->caps_list_lock);
237                 }
238                 return cap;
239         }
240 
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);
245         BUG_ON(!ctx->count);
246         BUG_ON(ctx->count > mdsc->caps_reserve_count);
247         BUG_ON(list_empty(&mdsc->caps_list));
248 
249         ctx->count--;
250         mdsc->caps_reserve_count--;
251         mdsc->caps_use_count++;
252 
253         cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
254         list_del(&cap->caps_item);
255 
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);
259         return cap;
260 }
261 
262 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
263 {
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--;
269         /*
270          * Keep some preallocated caps around (ceph_min_count), to
271          * avoid lots of free/alloc churn.
272          */
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);
277         } else {
278                 mdsc->caps_avail_count++;
279                 list_add(&cap->caps_item, &mdsc->caps_list);
280         }
281 
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);
285 }
286 
287 void ceph_reservation_status(struct ceph_fs_client *fsc,
288                              int *total, int *avail, int *used, int *reserved,
289                              int *min)
290 {
291         struct ceph_mds_client *mdsc = fsc->mdsc;
292 
293         if (total)
294                 *total = mdsc->caps_total_count;
295         if (avail)
296                 *avail = mdsc->caps_avail_count;
297         if (used)
298                 *used = mdsc->caps_use_count;
299         if (reserved)
300                 *reserved = mdsc->caps_reserve_count;
301         if (min)
302                 *min = mdsc->caps_min_count;
303 }
304 
305 /*
306  * Find ceph_cap for given mds, if any.
307  *
308  * Called with i_ceph_lock held.
309  */
310 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
311 {
312         struct ceph_cap *cap;
313         struct rb_node *n = ci->i_caps.rb_node;
314 
315         while (n) {
316                 cap = rb_entry(n, struct ceph_cap, ci_node);
317                 if (mds < cap->mds)
318                         n = n->rb_left;
319                 else if (mds > cap->mds)
320                         n = n->rb_right;
321                 else
322                         return cap;
323         }
324         return NULL;
325 }
326 
327 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
328 {
329         struct ceph_cap *cap;
330 
331         spin_lock(&ci->i_ceph_lock);
332         cap = __get_cap_for_mds(ci, mds);
333         spin_unlock(&ci->i_ceph_lock);
334         return cap;
335 }
336 
337 /*
338  * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
339  */
340 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
341 {
342         struct ceph_cap *cap;
343         int mds = -1;
344         struct rb_node *p;
345 
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);
349                 mds = cap->mds;
350                 if (cap->issued & (CEPH_CAP_FILE_WR |
351                                    CEPH_CAP_FILE_BUFFER |
352                                    CEPH_CAP_FILE_EXCL))
353                         break;
354         }
355         return mds;
356 }
357 
358 int ceph_get_cap_mds(struct inode *inode)
359 {
360         struct ceph_inode_info *ci = ceph_inode(inode);
361         int mds;
362         spin_lock(&ci->i_ceph_lock);
363         mds = __ceph_get_cap_mds(ceph_inode(inode));
364         spin_unlock(&ci->i_ceph_lock);
365         return mds;
366 }
367 
368 /*
369  * Called under i_ceph_lock.
370  */
371 static void __insert_cap_node(struct ceph_inode_info *ci,
372                               struct ceph_cap *new)
373 {
374         struct rb_node **p = &ci->i_caps.rb_node;
375         struct rb_node *parent = NULL;
376         struct ceph_cap *cap = NULL;
377 
378         while (*p) {
379                 parent = *p;
380                 cap = rb_entry(parent, struct ceph_cap, ci_node);
381                 if (new->mds < cap->mds)
382                         p = &(*p)->rb_left;
383                 else if (new->mds > cap->mds)
384                         p = &(*p)->rb_right;
385                 else
386                         BUG();
387         }
388 
389         rb_link_node(&new->ci_node, parent, p);
390         rb_insert_color(&new->ci_node, &ci->i_caps);
391 }
392 
393 /*
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.
396  */
397 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
398                                struct ceph_inode_info *ci)
399 {
400         struct ceph_mount_options *ma = mdsc->fsc->mount_options;
401 
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);
408 }
409 
410 /*
411  * (Re)queue cap at the end of the delayed cap release list.
412  *
413  * If I_FLUSH is set, leave the inode at the front of the list.
414  *
415  * Caller holds i_ceph_lock
416  *    -> we take mdsc->cap_delay_lock
417  */
418 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
419                                 struct ceph_inode_info *ci)
420 {
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)
428                                 goto no_change;
429                         list_del_init(&ci->i_cap_delay_list);
430                 }
431                 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
432 no_change:
433                 spin_unlock(&mdsc->cap_delay_lock);
434         }
435 }
436 
437 /*
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.
441  */
442 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
443                                       struct ceph_inode_info *ci)
444 {
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);
452 }
453 
454 /*
455  * Cancel delayed work on cap.
456  *
457  * Caller must hold i_ceph_lock.
458  */
459 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
460                                struct ceph_inode_info *ci)
461 {
462         dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
463         if (list_empty(&ci->i_cap_delay_list))
464                 return;
465         spin_lock(&mdsc->cap_delay_lock);
466         list_del_init(&ci->i_cap_delay_list);
467         spin_unlock(&mdsc->cap_delay_lock);
468 }
469 
470 /*
471  * Common issue checks for add_cap, handle_cap_grant.
472  */
473 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
474                               unsigned issued)
475 {
476         unsigned had = __ceph_caps_issued(ci, NULL);
477 
478         /*
479          * Each time we receive FILE_CACHE anew, we increment
480          * i_rdcache_gen.
481          */
482         if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
483             (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
484                 ci->i_rdcache_gen++;
485         }
486 
487         /*
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
490          * have the cap.
491          */
492         if ((issued & CEPH_CAP_FILE_SHARED) &&
493             (had & CEPH_CAP_FILE_SHARED) == 0) {
494                 ci->i_shared_gen++;
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);
498                 }
499         }
500 }
501 
502 /*
503  * Add a capability under the given MDS session.
504  *
505  * Caller should hold session snap_rwsem (read) and s_mutex.
506  *
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.)
510  */
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)
516 {
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;
521         int actual_wanted;
522 
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);
525 
526         /*
527          * If we are opening the file, include file mode wanted bits
528          * in wanted.
529          */
530         if (fmode >= 0)
531                 wanted |= ceph_caps_for_mode(fmode);
532 
533         cap = __get_cap_for_mds(ci, mds);
534         if (!cap) {
535                 cap = *new_cap;
536                 *new_cap = NULL;
537 
538                 cap->issued = 0;
539                 cap->implemented = 0;
540                 cap->mds = mds;
541                 cap->mds_wanted = 0;
542                 cap->mseq = 0;
543 
544                 cap->ci = ci;
545                 __insert_cap_node(ci, cap);
546 
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);
553         } else {
554                 /*
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.
558                  *
559                  * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
560                  * a message that was send before the cap import message. So
561                  * don't remove caps.
562                  */
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);
566                         seq = cap->seq;
567                         mseq = cap->mseq;
568                         issued |= cap->issued;
569                         flags |= CEPH_CAP_FLAG_AUTH;
570                 }
571         }
572 
573         if (!ci->i_snap_realm) {
574                 /*
575                  * add this inode to the appropriate snap realm
576                  */
577                 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
578                                                                realmino);
579                 if (realm) {
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);
585                 } else {
586                         pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
587                                realmino);
588                         WARN_ON(!realm);
589                 }
590         }
591 
592         __check_cap_issue(ci, cap, issued);
593 
594         /*
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.
598          */
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);
606         }
607 
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;
613                 }
614         } else {
615                 WARN_ON(ci->i_auth_cap == cap);
616         }
617 
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;
626         else
627                 cap->mds_wanted |= wanted;
628         cap->seq = seq;
629         cap->issue_seq = seq;
630         cap->mseq = mseq;
631         cap->cap_gen = session->s_cap_gen;
632 
633         if (fmode >= 0)
634                 __ceph_get_fmode(ci, fmode);
635 }
636 
637 /*
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).
641  */
642 static int __cap_is_valid(struct ceph_cap *cap)
643 {
644         unsigned long ttl;
645         u32 gen;
646 
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);
651 
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);
656                 return 0;
657         }
658 
659         return 1;
660 }
661 
662 /*
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.
666  */
667 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
668 {
669         int have = ci->i_snap_caps;
670         struct ceph_cap *cap;
671         struct rb_node *p;
672 
673         if (implemented)
674                 *implemented = 0;
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))
678                         continue;
679                 dout("__ceph_caps_issued %p cap %p issued %s\n",
680                      &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
681                 have |= cap->issued;
682                 if (implemented)
683                         *implemented |= cap->implemented;
684         }
685         /*
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.
689          */
690         if (ci->i_auth_cap) {
691                 cap = ci->i_auth_cap;
692                 have &= ~cap->implemented | cap->issued;
693         }
694         return have;
695 }
696 
697 /*
698  * Get cap bits issued by caps other than @ocap
699  */
700 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
701 {
702         int have = ci->i_snap_caps;
703         struct ceph_cap *cap;
704         struct rb_node *p;
705 
706         for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
707                 cap = rb_entry(p, struct ceph_cap, ci_node);
708                 if (cap == ocap)
709                         continue;
710                 if (!__cap_is_valid(cap))
711                         continue;
712                 have |= cap->issued;
713         }
714         return have;
715 }
716 
717 /*
718  * Move a cap to the end of the LRU (oldest caps at list head, newest
719  * at list tail).
720  */
721 static void __touch_cap(struct ceph_cap *cap)
722 {
723         struct ceph_mds_session *s = cap->session;
724 
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,
728                      s->s_mds);
729                 list_move_tail(&cap->session_caps, &s->s_caps);
730         } else {
731                 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
732                      &cap->ci->vfs_inode, cap, s->s_mds);
733         }
734         spin_unlock(&s->s_cap_lock);
735 }
736 
737 /*
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.)
741  */
742 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
743 {
744         struct ceph_cap *cap;
745         struct rb_node *p;
746         int have = ci->i_snap_caps;
747 
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));
753                 return 1;
754         }
755 
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))
759                         continue;
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));
765                         if (touch)
766                                 __touch_cap(cap);
767                         return 1;
768                 }
769 
770                 /* does a combination of caps satisfy mask? */
771                 have |= cap->issued;
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));
777                         if (touch) {
778                                 struct rb_node *q;
779 
780                                 /* touch this + preceding caps */
781                                 __touch_cap(cap);
782                                 for (q = rb_first(&ci->i_caps); q != p;
783                                      q = rb_next(q)) {
784                                         cap = rb_entry(q, struct ceph_cap,
785                                                        ci_node);
786                                         if (!__cap_is_valid(cap))
787                                                 continue;
788                                         __touch_cap(cap);
789                                 }
790                         }
791                         return 1;
792                 }
793         }
794 
795         return 0;
796 }
797 
798 /*
799  * Return true if mask caps are currently being revoked by an MDS.
800  */
801 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
802                                struct ceph_cap *ocap, int mask)
803 {
804         struct ceph_cap *cap;
805         struct rb_node *p;
806 
807         for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
808                 cap = rb_entry(p, struct ceph_cap, ci_node);
809                 if (cap != ocap &&
810                     (cap->implemented & ~cap->issued & mask))
811                         return 1;
812         }
813         return 0;
814 }
815 
816 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
817 {
818         struct inode *inode = &ci->vfs_inode;
819         int ret;
820 
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);
826         return ret;
827 }
828 
829 int __ceph_caps_used(struct ceph_inode_info *ci)
830 {
831         int used = 0;
832         if (ci->i_pin_ref)
833                 used |= CEPH_CAP_PIN;
834         if (ci->i_rd_ref)
835                 used |= CEPH_CAP_FILE_RD;
836         if (ci->i_rdcache_ref ||
837             (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
838              ci->vfs_inode.i_data.nrpages))
839                 used |= CEPH_CAP_FILE_CACHE;
840         if (ci->i_wr_ref)
841                 used |= CEPH_CAP_FILE_WR;
842         if (ci->i_wb_ref || ci->i_wrbuffer_ref)
843                 used |= CEPH_CAP_FILE_BUFFER;
844         return used;
845 }
846 
847 /*
848  * wanted, by virtue of open file modes
849  */
850 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
851 {
852         int want = 0;
853         int mode;
854         for (mode = 0; mode < CEPH_FILE_MODE_NUM; mode++)
855                 if (ci->i_nr_by_mode[mode])
856                         want |= ceph_caps_for_mode(mode);
857         return want;
858 }
859 
860 /*
861  * Return caps we have registered with the MDS(s) as 'wanted'.
862  */
863 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
864 {
865         struct ceph_cap *cap;
866         struct rb_node *p;
867         int mds_wanted = 0;
868 
869         for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
870                 cap = rb_entry(p, struct ceph_cap, ci_node);
871                 if (!__cap_is_valid(cap))
872                         continue;
873                 if (cap == ci->i_auth_cap)
874                         mds_wanted |= cap->mds_wanted;
875                 else
876                         mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
877         }
878         return mds_wanted;
879 }
880 
881 /*
882  * called under i_ceph_lock
883  */
884 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
885 {
886         return !RB_EMPTY_ROOT(&ci->i_caps);
887 }
888 
889 int ceph_is_any_caps(struct inode *inode)
890 {
891         struct ceph_inode_info *ci = ceph_inode(inode);
892         int ret;
893 
894         spin_lock(&ci->i_ceph_lock);
895         ret = __ceph_is_any_caps(ci);
896         spin_unlock(&ci->i_ceph_lock);
897 
898         return ret;
899 }
900 
901 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
902 {
903         struct ceph_snap_realm *realm = ci->i_snap_realm;
904         spin_lock(&realm->inodes_with_caps_lock);
905         list_del_init(&ci->i_snap_realm_item);
906         ci->i_snap_realm_counter++;
907         ci->i_snap_realm = NULL;
908         spin_unlock(&realm->inodes_with_caps_lock);
909         ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
910                             realm);
911 }
912 
913 /*
914  * Remove a cap.  Take steps to deal with a racing iterate_session_caps.
915  *
916  * caller should hold i_ceph_lock.
917  * caller will not hold session s_mutex if called from destroy_inode.
918  */
919 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
920 {
921         struct ceph_mds_session *session = cap->session;
922         struct ceph_inode_info *ci = cap->ci;
923         struct ceph_mds_client *mdsc =
924                 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
925         int removed = 0;
926 
927         dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
928 
929         /* remove from session list */
930         spin_lock(&session->s_cap_lock);
931         if (session->s_cap_iterator == cap) {
932                 /* not yet, we are iterating over this very cap */
933                 dout("__ceph_remove_cap  delaying %p removal from session %p\n",
934                      cap, cap->session);
935         } else {
936                 list_del_init(&cap->session_caps);
937                 session->s_nr_caps--;
938                 cap->session = NULL;
939                 removed = 1;
940         }
941         /* protect backpointer with s_cap_lock: see iterate_session_caps */
942         cap->ci = NULL;
943 
944         /*
945          * s_cap_reconnect is protected by s_cap_lock. no one changes
946          * s_cap_gen while session is in the reconnect state.
947          */
948         if (queue_release &&
949             (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
950                 cap->queue_release = 1;
951                 if (removed) {
952                         list_add_tail(&cap->session_caps,
953                                       &session->s_cap_releases);
954                         session->s_num_cap_releases++;
955                         removed = 0;
956                 }
957         } else {
958                 cap->queue_release = 0;
959         }
960         cap->cap_ino = ci->i_vino.ino;
961 
962         spin_unlock(&session->s_cap_lock);
963 
964         /* remove from inode list */
965         rb_erase(&cap->ci_node, &ci->i_caps);
966         if (ci->i_auth_cap == cap)
967                 ci->i_auth_cap = NULL;
968 
969         if (removed)
970                 ceph_put_cap(mdsc, cap);
971 
972         /* when reconnect denied, we remove session caps forcibly,
973          * i_wr_ref can be non-zero. If there are ongoing write,
974          * keep i_snap_realm.
975          */
976         if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
977                 drop_inode_snap_realm(ci);
978 
979         if (!__ceph_is_any_real_caps(ci))
980                 __cap_delay_cancel(mdsc, ci);
981 }
982 
983 /*
984  * Build and send a cap message to the given MDS.
985  *
986  * Caller should be holding s_mutex.
987  */
988 static int send_cap_msg(struct ceph_mds_session *session,
989                         u64 ino, u64 cid, int op,
990                         int caps, int wanted, int dirty,
991                         u32 seq, u64 flush_tid, u64 oldest_flush_tid,
992                         u32 issue_seq, u32 mseq, u64 size, u64 max_size,
993                         struct timespec *mtime, struct timespec *atime,
994                         u64 time_warp_seq,
995                         kuid_t uid, kgid_t gid, umode_t mode,
996                         u64 xattr_version,
997                         struct ceph_buffer *xattrs_buf,
998                         u64 follows, bool inline_data)
999 {
1000         struct ceph_mds_caps *fc;
1001         struct ceph_msg *msg;
1002         void *p;
1003         size_t extra_len;
1004 
1005         dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1006              " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1007              " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
1008              cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
1009              ceph_cap_string(dirty),
1010              seq, issue_seq, flush_tid, oldest_flush_tid,
1011              mseq, follows, size, max_size,
1012              xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
1013 
1014         /* flock buffer size + inline version + inline data size +
1015          * osd_epoch_barrier + oldest_flush_tid */
1016         extra_len = 4 + 8 + 4 + 4 + 8;
1017         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1018                            GFP_NOFS, false);
1019         if (!msg)
1020                 return -ENOMEM;
1021 
1022         msg->hdr.version = cpu_to_le16(6);
1023         msg->hdr.tid = cpu_to_le64(flush_tid);
1024 
1025         fc = msg->front.iov_base;
1026         memset(fc, 0, sizeof(*fc));
1027 
1028         fc->cap_id = cpu_to_le64(cid);
1029         fc->op = cpu_to_le32(op);
1030         fc->seq = cpu_to_le32(seq);
1031         fc->issue_seq = cpu_to_le32(issue_seq);
1032         fc->migrate_seq = cpu_to_le32(mseq);
1033         fc->caps = cpu_to_le32(caps);
1034         fc->wanted = cpu_to_le32(wanted);
1035         fc->dirty = cpu_to_le32(dirty);
1036         fc->ino = cpu_to_le64(ino);
1037         fc->snap_follows = cpu_to_le64(follows);
1038 
1039         fc->size = cpu_to_le64(size);
1040         fc->max_size = cpu_to_le64(max_size);
1041         if (mtime)
1042                 ceph_encode_timespec(&fc->mtime, mtime);
1043         if (atime)
1044                 ceph_encode_timespec(&fc->atime, atime);
1045         fc->time_warp_seq = cpu_to_le32(time_warp_seq);
1046 
1047         fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid));
1048         fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid));
1049         fc->mode = cpu_to_le32(mode);
1050 
1051         p = fc + 1;
1052         /* flock buffer size */
1053         ceph_encode_32(&p, 0);
1054         /* inline version */
1055         ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE);
1056         /* inline data size */
1057         ceph_encode_32(&p, 0);
1058         /* osd_epoch_barrier */
1059         ceph_encode_32(&p, 0);
1060         /* oldest_flush_tid */
1061         ceph_encode_64(&p, oldest_flush_tid);
1062 
1063         fc->xattr_version = cpu_to_le64(xattr_version);
1064         if (xattrs_buf) {
1065                 msg->middle = ceph_buffer_get(xattrs_buf);
1066                 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1067                 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1068         }
1069 
1070         ceph_con_send(&session->s_con, msg);
1071         return 0;
1072 }
1073 
1074 /*
1075  * Queue cap releases when an inode is dropped from our cache.  Since
1076  * inode is about to be destroyed, there is no need for i_ceph_lock.
1077  */
1078 void ceph_queue_caps_release(struct inode *inode)
1079 {
1080         struct ceph_inode_info *ci = ceph_inode(inode);
1081         struct rb_node *p;
1082 
1083         p = rb_first(&ci->i_caps);
1084         while (p) {
1085                 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1086                 p = rb_next(p);
1087                 __ceph_remove_cap(cap, true);
1088         }
1089 }
1090 
1091 /*
1092  * Send a cap msg on the given inode.  Update our caps state, then
1093  * drop i_ceph_lock and send the message.
1094  *
1095  * Make note of max_size reported/requested from mds, revoked caps
1096  * that have now been implemented.
1097  *
1098  * Make half-hearted attempt ot to invalidate page cache if we are
1099  * dropping RDCACHE.  Note that this will leave behind locked pages
1100  * that we'll then need to deal with elsewhere.
1101  *
1102  * Return non-zero if delayed release, or we experienced an error
1103  * such that the caller should requeue + retry later.
1104  *
1105  * called with i_ceph_lock, then drops it.
1106  * caller should hold snap_rwsem (read), s_mutex.
1107  */
1108 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1109                       int op, int used, int want, int retain, int flushing,
1110                       u64 flush_tid, u64 oldest_flush_tid)
1111         __releases(cap->ci->i_ceph_lock)
1112 {
1113         struct ceph_inode_info *ci = cap->ci;
1114         struct inode *inode = &ci->vfs_inode;
1115         u64 cap_id = cap->cap_id;
1116         int held, revoking, dropping, keep;
1117         u64 seq, issue_seq, mseq, time_warp_seq, follows;
1118         u64 size, max_size;
1119         struct timespec mtime, atime;
1120         int wake = 0;
1121         umode_t mode;
1122         kuid_t uid;
1123         kgid_t gid;
1124         struct ceph_mds_session *session;
1125         u64 xattr_version = 0;
1126         struct ceph_buffer *xattr_blob = NULL;
1127         int delayed = 0;
1128         int ret;
1129         bool inline_data;
1130 
1131         held = cap->issued | cap->implemented;
1132         revoking = cap->implemented & ~cap->issued;
1133         retain &= ~revoking;
1134         dropping = cap->issued & ~retain;
1135 
1136         dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1137              inode, cap, cap->session,
1138              ceph_cap_string(held), ceph_cap_string(held & retain),
1139              ceph_cap_string(revoking));
1140         BUG_ON((retain & CEPH_CAP_PIN) == 0);
1141 
1142         session = cap->session;
1143 
1144         /* don't release wanted unless we've waited a bit. */
1145         if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1146             time_before(jiffies, ci->i_hold_caps_min)) {
1147                 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1148                      ceph_cap_string(cap->issued),
1149                      ceph_cap_string(cap->issued & retain),
1150                      ceph_cap_string(cap->mds_wanted),
1151                      ceph_cap_string(want));
1152                 want |= cap->mds_wanted;
1153                 retain |= cap->issued;
1154                 delayed = 1;
1155         }
1156         ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1157 
1158         cap->issued &= retain;  /* drop bits we don't want */
1159         if (cap->implemented & ~cap->issued) {
1160                 /*
1161                  * Wake up any waiters on wanted -> needed transition.
1162                  * This is due to the weird transition from buffered
1163                  * to sync IO... we need to flush dirty pages _before_
1164                  * allowing sync writes to avoid reordering.
1165                  */
1166                 wake = 1;
1167         }
1168         cap->implemented &= cap->issued | used;
1169         cap->mds_wanted = want;
1170 
1171         follows = flushing ? ci->i_head_snapc->seq : 0;
1172 
1173         keep = cap->implemented;
1174         seq = cap->seq;
1175         issue_seq = cap->issue_seq;
1176         mseq = cap->mseq;
1177         size = inode->i_size;
1178         ci->i_reported_size = size;
1179         max_size = ci->i_wanted_max_size;
1180         ci->i_requested_max_size = max_size;
1181         mtime = inode->i_mtime;
1182         atime = inode->i_atime;
1183         time_warp_seq = ci->i_time_warp_seq;
1184         uid = inode->i_uid;
1185         gid = inode->i_gid;
1186         mode = inode->i_mode;
1187 
1188         if (flushing & CEPH_CAP_XATTR_EXCL) {
1189                 __ceph_build_xattrs_blob(ci);
1190                 xattr_blob = ci->i_xattrs.blob;
1191                 xattr_version = ci->i_xattrs.version;
1192         }
1193 
1194         inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1195 
1196         spin_unlock(&ci->i_ceph_lock);
1197 
1198         ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1199                 op, keep, want, flushing, seq,
1200                 flush_tid, oldest_flush_tid, issue_seq, mseq,
1201                 size, max_size, &mtime, &atime, time_warp_seq,
1202                 uid, gid, mode, xattr_version, xattr_blob,
1203                 follows, inline_data);
1204         if (ret < 0) {
1205                 dout("error sending cap msg, must requeue %p\n", inode);
1206                 delayed = 1;
1207         }
1208 
1209         if (wake)
1210                 wake_up_all(&ci->i_cap_wq);
1211 
1212         return delayed;
1213 }
1214 
1215 /*
1216  * When a snapshot is taken, clients accumulate dirty metadata on
1217  * inodes with capabilities in ceph_cap_snaps to describe the file
1218  * state at the time the snapshot was taken.  This must be flushed
1219  * asynchronously back to the MDS once sync writes complete and dirty
1220  * data is written out.
1221  *
1222  * Unless @kick is true, skip cap_snaps that were already sent to
1223  * the MDS (i.e., during this session).
1224  *
1225  * Called under i_ceph_lock.  Takes s_mutex as needed.
1226  */
1227 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1228                         struct ceph_mds_session **psession,
1229                         int kick)
1230                 __releases(ci->i_ceph_lock)
1231                 __acquires(ci->i_ceph_lock)
1232 {
1233         struct inode *inode = &ci->vfs_inode;
1234         int mds;
1235         struct ceph_cap_snap *capsnap;
1236         u32 mseq;
1237         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1238         struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1239                                                     session->s_mutex */
1240         u64 next_follows = 0;  /* keep track of how far we've gotten through the
1241                              i_cap_snaps list, and skip these entries next time
1242                              around to avoid an infinite loop */
1243 
1244         if (psession)
1245                 session = *psession;
1246 
1247         dout("__flush_snaps %p\n", inode);
1248 retry:
1249         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1250                 /* avoid an infiniute loop after retry */
1251                 if (capsnap->follows < next_follows)
1252                         continue;
1253                 /*
1254                  * we need to wait for sync writes to complete and for dirty
1255                  * pages to be written out.
1256                  */
1257                 if (capsnap->dirty_pages || capsnap->writing)
1258                         break;
1259 
1260                 /* should be removed by ceph_try_drop_cap_snap() */
1261                 BUG_ON(!capsnap->need_flush);
1262 
1263                 /* pick mds, take s_mutex */
1264                 if (ci->i_auth_cap == NULL) {
1265                         dout("no auth cap (migrating?), doing nothing\n");
1266                         goto out;
1267                 }
1268 
1269                 /* only flush each capsnap once */
1270                 if (!kick && !list_empty(&capsnap->flushing_item)) {
1271                         dout("already flushed %p, skipping\n", capsnap);
1272                         continue;
1273                 }
1274 
1275                 mds = ci->i_auth_cap->session->s_mds;
1276                 mseq = ci->i_auth_cap->mseq;
1277 
1278                 if (session && session->s_mds != mds) {
1279                         dout("oops, wrong session %p mutex\n", session);
1280                         if (kick)
1281                                 goto out;
1282 
1283                         mutex_unlock(&session->s_mutex);
1284                         ceph_put_mds_session(session);
1285                         session = NULL;
1286                 }
1287                 if (!session) {
1288                         spin_unlock(&ci->i_ceph_lock);
1289                         mutex_lock(&mdsc->mutex);
1290                         session = __ceph_lookup_mds_session(mdsc, mds);
1291                         mutex_unlock(&mdsc->mutex);
1292                         if (session) {
1293                                 dout("inverting session/ino locks on %p\n",
1294                                      session);
1295                                 mutex_lock(&session->s_mutex);
1296                         }
1297                         /*
1298                          * if session == NULL, we raced against a cap
1299                          * deletion or migration.  retry, and we'll
1300                          * get a better @mds value next time.
1301                          */
1302                         spin_lock(&ci->i_ceph_lock);
1303                         goto retry;
1304                 }
1305 
1306                 spin_lock(&mdsc->cap_dirty_lock);
1307                 capsnap->flush_tid = ++mdsc->last_cap_flush_tid;
1308                 spin_unlock(&mdsc->cap_dirty_lock);
1309 
1310                 atomic_inc(&capsnap->nref);
1311                 if (list_empty(&capsnap->flushing_item))
1312                         list_add_tail(&capsnap->flushing_item,
1313                                       &session->s_cap_snaps_flushing);
1314                 spin_unlock(&ci->i_ceph_lock);
1315 
1316                 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1317                      inode, capsnap, capsnap->follows, capsnap->flush_tid);
1318                 send_cap_msg(session, ceph_vino(inode).ino, 0,
1319                              CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1320                              capsnap->dirty, 0, capsnap->flush_tid, 0,
1321                              0, mseq, capsnap->size, 0,
1322                              &capsnap->mtime, &capsnap->atime,
1323                              capsnap->time_warp_seq,
1324                              capsnap->uid, capsnap->gid, capsnap->mode,
1325                              capsnap->xattr_version, capsnap->xattr_blob,
1326                              capsnap->follows, capsnap->inline_data);
1327 
1328                 next_follows = capsnap->follows + 1;
1329                 ceph_put_cap_snap(capsnap);
1330 
1331                 spin_lock(&ci->i_ceph_lock);
1332                 goto retry;
1333         }
1334 
1335         /* we flushed them all; remove this inode from the queue */
1336         spin_lock(&mdsc->snap_flush_lock);
1337         list_del_init(&ci->i_snap_flush_item);
1338         spin_unlock(&mdsc->snap_flush_lock);
1339 
1340 out:
1341         if (psession)
1342                 *psession = session;
1343         else if (session) {
1344                 mutex_unlock(&session->s_mutex);
1345                 ceph_put_mds_session(session);
1346         }
1347 }
1348 
1349 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1350 {
1351         spin_lock(&ci->i_ceph_lock);
1352         __ceph_flush_snaps(ci, NULL, 0);
1353         spin_unlock(&ci->i_ceph_lock);
1354 }
1355 
1356 /*
1357  * Mark caps dirty.  If inode is newly dirty, return the dirty flags.
1358  * Caller is then responsible for calling __mark_inode_dirty with the
1359  * returned flags value.
1360  */
1361 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1362                            struct ceph_cap_flush **pcf)
1363 {
1364         struct ceph_mds_client *mdsc =
1365                 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1366         struct inode *inode = &ci->vfs_inode;
1367         int was = ci->i_dirty_caps;
1368         int dirty = 0;
1369 
1370         if (!ci->i_auth_cap) {
1371                 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1372                         "but no auth cap (session was closed?)\n",
1373                         inode, ceph_ino(inode), ceph_cap_string(mask));
1374                 return 0;
1375         }
1376 
1377         dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1378              ceph_cap_string(mask), ceph_cap_string(was),
1379              ceph_cap_string(was | mask));
1380         ci->i_dirty_caps |= mask;
1381         if (was == 0) {
1382                 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1383                 swap(ci->i_prealloc_cap_flush, *pcf);
1384 
1385                 if (!ci->i_head_snapc) {
1386                         WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1387                         ci->i_head_snapc = ceph_get_snap_context(
1388                                 ci->i_snap_realm->cached_context);
1389                 }
1390                 dout(" inode %p now dirty snapc %p auth cap %p\n",
1391                      &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1392                 BUG_ON(!list_empty(&ci->i_dirty_item));
1393                 spin_lock(&mdsc->cap_dirty_lock);
1394                 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1395                 spin_unlock(&mdsc->cap_dirty_lock);
1396                 if (ci->i_flushing_caps == 0) {
1397                         ihold(inode);
1398                         dirty |= I_DIRTY_SYNC;
1399                 }
1400         } else {
1401                 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1402         }
1403         BUG_ON(list_empty(&ci->i_dirty_item));
1404         if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1405             (mask & CEPH_CAP_FILE_BUFFER))
1406                 dirty |= I_DIRTY_DATASYNC;
1407         __cap_delay_requeue(mdsc, ci);
1408         return dirty;
1409 }
1410 
1411 static void __add_cap_flushing_to_inode(struct ceph_inode_info *ci,
1412                                         struct ceph_cap_flush *cf)
1413 {
1414         struct rb_node **p = &ci->i_cap_flush_tree.rb_node;
1415         struct rb_node *parent = NULL;
1416         struct ceph_cap_flush *other = NULL;
1417 
1418         while (*p) {
1419                 parent = *p;
1420                 other = rb_entry(parent, struct ceph_cap_flush, i_node);
1421 
1422                 if (cf->tid < other->tid)
1423                         p = &(*p)->rb_left;
1424                 else if (cf->tid > other->tid)
1425                         p = &(*p)->rb_right;
1426                 else
1427                         BUG();
1428         }
1429 
1430         rb_link_node(&cf->i_node, parent, p);
1431         rb_insert_color(&cf->i_node, &ci->i_cap_flush_tree);
1432 }
1433 
1434 static void __add_cap_flushing_to_mdsc(struct ceph_mds_client *mdsc,
1435                                        struct ceph_cap_flush *cf)
1436 {
1437         struct rb_node **p = &mdsc->cap_flush_tree.rb_node;
1438         struct rb_node *parent = NULL;
1439         struct ceph_cap_flush *other = NULL;
1440 
1441         while (*p) {
1442                 parent = *p;
1443                 other = rb_entry(parent, struct ceph_cap_flush, g_node);
1444 
1445                 if (cf->tid < other->tid)
1446                         p = &(*p)->rb_left;
1447                 else if (cf->tid > other->tid)
1448                         p = &(*p)->rb_right;
1449                 else
1450                         BUG();
1451         }
1452 
1453         rb_link_node(&cf->g_node, parent, p);
1454         rb_insert_color(&cf->g_node, &mdsc->cap_flush_tree);
1455 }
1456 
1457 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1458 {
1459         return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1460 }
1461 
1462 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1463 {
1464         if (cf)
1465                 kmem_cache_free(ceph_cap_flush_cachep, cf);
1466 }
1467 
1468 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1469 {
1470         struct rb_node *n = rb_first(&mdsc->cap_flush_tree);
1471         if (n) {
1472                 struct ceph_cap_flush *cf =
1473                         rb_entry(n, struct ceph_cap_flush, g_node);
1474                 return cf->tid;
1475         }
1476         return 0;
1477 }
1478 
1479 /*
1480  * Add dirty inode to the flushing list.  Assigned a seq number so we
1481  * can wait for caps to flush without starving.
1482  *
1483  * Called under i_ceph_lock.
1484  */
1485 static int __mark_caps_flushing(struct inode *inode,
1486                                 struct ceph_mds_session *session,
1487                                 u64 *flush_tid, u64 *oldest_flush_tid)
1488 {
1489         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1490         struct ceph_inode_info *ci = ceph_inode(inode);
1491         struct ceph_cap_flush *cf = NULL;
1492         int flushing;
1493 
1494         BUG_ON(ci->i_dirty_caps == 0);
1495         BUG_ON(list_empty(&ci->i_dirty_item));
1496         BUG_ON(!ci->i_prealloc_cap_flush);
1497 
1498         flushing = ci->i_dirty_caps;
1499         dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1500              ceph_cap_string(flushing),
1501              ceph_cap_string(ci->i_flushing_caps),
1502              ceph_cap_string(ci->i_flushing_caps | flushing));
1503         ci->i_flushing_caps |= flushing;
1504         ci->i_dirty_caps = 0;
1505         dout(" inode %p now !dirty\n", inode);
1506 
1507         swap(cf, ci->i_prealloc_cap_flush);
1508         cf->caps = flushing;
1509 
1510         spin_lock(&mdsc->cap_dirty_lock);
1511         list_del_init(&ci->i_dirty_item);
1512 
1513         cf->tid = ++mdsc->last_cap_flush_tid;
1514         __add_cap_flushing_to_mdsc(mdsc, cf);
1515         *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1516 
1517         if (list_empty(&ci->i_flushing_item)) {
1518                 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1519                 mdsc->num_cap_flushing++;
1520                 dout(" inode %p now flushing tid %llu\n", inode, cf->tid);
1521         } else {
1522                 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1523                 dout(" inode %p now flushing (more) tid %llu\n",
1524                      inode, cf->tid);
1525         }
1526         spin_unlock(&mdsc->cap_dirty_lock);
1527 
1528         __add_cap_flushing_to_inode(ci, cf);
1529 
1530         *flush_tid = cf->tid;
1531         return flushing;
1532 }
1533 
1534 /*
1535  * try to invalidate mapping pages without blocking.
1536  */
1537 static int try_nonblocking_invalidate(struct inode *inode)
1538 {
1539         struct ceph_inode_info *ci = ceph_inode(inode);
1540         u32 invalidating_gen = ci->i_rdcache_gen;
1541 
1542         spin_unlock(&ci->i_ceph_lock);
1543         invalidate_mapping_pages(&inode->i_data, 0, -1);
1544         spin_lock(&ci->i_ceph_lock);
1545 
1546         if (inode->i_data.nrpages == 0 &&
1547             invalidating_gen == ci->i_rdcache_gen) {
1548                 /* success. */
1549                 dout("try_nonblocking_invalidate %p success\n", inode);
1550                 /* save any racing async invalidate some trouble */
1551                 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1552                 return 0;
1553         }
1554         dout("try_nonblocking_invalidate %p failed\n", inode);
1555         return -1;
1556 }
1557 
1558 /*
1559  * Swiss army knife function to examine currently used and wanted
1560  * versus held caps.  Release, flush, ack revoked caps to mds as
1561  * appropriate.
1562  *
1563  *  CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1564  *    cap release further.
1565  *  CHECK_CAPS_AUTHONLY - we should only check the auth cap
1566  *  CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1567  *    further delay.
1568  */
1569 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1570                      struct ceph_mds_session *session)
1571 {
1572         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1573         struct ceph_mds_client *mdsc = fsc->mdsc;
1574         struct inode *inode = &ci->vfs_inode;
1575         struct ceph_cap *cap;
1576         u64 flush_tid, oldest_flush_tid;
1577         int file_wanted, used, cap_used;
1578         int took_snap_rwsem = 0;             /* true if mdsc->snap_rwsem held */
1579         int issued, implemented, want, retain, revoking, flushing = 0;
1580         int mds = -1;   /* keep track of how far we've gone through i_caps list
1581                            to avoid an infinite loop on retry */
1582         struct rb_node *p;
1583         int tried_invalidate = 0;
1584         int delayed = 0, sent = 0, force_requeue = 0, num;
1585         int queue_invalidate = 0;
1586         int is_delayed = flags & CHECK_CAPS_NODELAY;
1587 
1588         /* if we are unmounting, flush any unused caps immediately. */
1589         if (mdsc->stopping)
1590                 is_delayed = 1;
1591 
1592         spin_lock(&ci->i_ceph_lock);
1593 
1594         if (ci->i_ceph_flags & CEPH_I_FLUSH)
1595                 flags |= CHECK_CAPS_FLUSH;
1596 
1597         /* flush snaps first time around only */
1598         if (!list_empty(&ci->i_cap_snaps))
1599                 __ceph_flush_snaps(ci, &session, 0);
1600         goto retry_locked;
1601 retry:
1602         spin_lock(&ci->i_ceph_lock);
1603 retry_locked:
1604         file_wanted = __ceph_caps_file_wanted(ci);
1605         used = __ceph_caps_used(ci);
1606         issued = __ceph_caps_issued(ci, &implemented);
1607         revoking = implemented & ~issued;
1608 
1609         want = file_wanted;
1610         retain = file_wanted | used | CEPH_CAP_PIN;
1611         if (!mdsc->stopping && inode->i_nlink > 0) {
1612                 if (file_wanted) {
1613                         retain |= CEPH_CAP_ANY;       /* be greedy */
1614                 } else if (S_ISDIR(inode->i_mode) &&
1615                            (issued & CEPH_CAP_FILE_SHARED) &&
1616                             __ceph_dir_is_complete(ci)) {
1617                         /*
1618                          * If a directory is complete, we want to keep
1619                          * the exclusive cap. So that MDS does not end up
1620                          * revoking the shared cap on every create/unlink
1621                          * operation.
1622                          */
1623                         want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1624                         retain |= want;
1625                 } else {
1626 
1627                         retain |= CEPH_CAP_ANY_SHARED;
1628                         /*
1629                          * keep RD only if we didn't have the file open RW,
1630                          * because then the mds would revoke it anyway to
1631                          * journal max_size=0.
1632                          */
1633                         if (ci->i_max_size == 0)
1634                                 retain |= CEPH_CAP_ANY_RD;
1635                 }
1636         }
1637 
1638         dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1639              " issued %s revoking %s retain %s %s%s%s\n", inode,
1640              ceph_cap_string(file_wanted),
1641              ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1642              ceph_cap_string(ci->i_flushing_caps),
1643              ceph_cap_string(issued), ceph_cap_string(revoking),
1644              ceph_cap_string(retain),
1645              (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1646              (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1647              (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1648 
1649         /*
1650          * If we no longer need to hold onto old our caps, and we may
1651          * have cached pages, but don't want them, then try to invalidate.
1652          * If we fail, it's because pages are locked.... try again later.
1653          */
1654         if ((!is_delayed || mdsc->stopping) &&
1655             !S_ISDIR(inode->i_mode) &&          /* ignore readdir cache */
1656             ci->i_wrbuffer_ref == 0 &&          /* no dirty pages... */
1657             inode->i_data.nrpages &&            /* have cached pages */
1658             (file_wanted == 0 ||                /* no open files */
1659              (revoking & (CEPH_CAP_FILE_CACHE|
1660                           CEPH_CAP_FILE_LAZYIO))) && /*  or revoking cache */
1661             !tried_invalidate) {
1662                 dout("check_caps trying to invalidate on %p\n", inode);
1663                 if (try_nonblocking_invalidate(inode) < 0) {
1664                         if (revoking & (CEPH_CAP_FILE_CACHE|
1665                                         CEPH_CAP_FILE_LAZYIO)) {
1666                                 dout("check_caps queuing invalidate\n");
1667                                 queue_invalidate = 1;
1668                                 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1669                         } else {
1670                                 dout("check_caps failed to invalidate pages\n");
1671                                 /* we failed to invalidate pages.  check these
1672                                    caps again later. */
1673                                 force_requeue = 1;
1674                                 __cap_set_timeouts(mdsc, ci);
1675                         }
1676                 }
1677                 tried_invalidate = 1;
1678                 goto retry_locked;
1679         }
1680 
1681         num = 0;
1682         for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1683                 cap = rb_entry(p, struct ceph_cap, ci_node);
1684                 num++;
1685 
1686                 /* avoid looping forever */
1687                 if (mds >= cap->mds ||
1688                     ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1689                         continue;
1690 
1691                 /* NOTE: no side-effects allowed, until we take s_mutex */
1692 
1693                 cap_used = used;
1694                 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1695                         cap_used &= ~ci->i_auth_cap->issued;
1696 
1697                 revoking = cap->implemented & ~cap->issued;
1698                 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1699                      cap->mds, cap, ceph_cap_string(cap->issued),
1700                      ceph_cap_string(cap_used),
1701                      ceph_cap_string(cap->implemented),
1702                      ceph_cap_string(revoking));
1703 
1704                 if (cap == ci->i_auth_cap &&
1705                     (cap->issued & CEPH_CAP_FILE_WR)) {
1706                         /* request larger max_size from MDS? */
1707                         if (ci->i_wanted_max_size > ci->i_max_size &&
1708                             ci->i_wanted_max_size > ci->i_requested_max_size) {
1709                                 dout("requesting new max_size\n");
1710                                 goto ack;
1711                         }
1712 
1713                         /* approaching file_max? */
1714                         if ((inode->i_size << 1) >= ci->i_max_size &&
1715                             (ci->i_reported_size << 1) < ci->i_max_size) {
1716                                 dout("i_size approaching max_size\n");
1717                                 goto ack;
1718                         }
1719                 }
1720                 /* flush anything dirty? */
1721                 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1722                     ci->i_dirty_caps) {
1723                         dout("flushing dirty caps\n");
1724                         goto ack;
1725                 }
1726 
1727                 /* completed revocation? going down and there are no caps? */
1728                 if (revoking && (revoking & cap_used) == 0) {
1729                         dout("completed revocation of %s\n",
1730                              ceph_cap_string(cap->implemented & ~cap->issued));
1731                         goto ack;
1732                 }
1733 
1734                 /* want more caps from mds? */
1735                 if (want & ~(cap->mds_wanted | cap->issued))
1736                         goto ack;
1737 
1738                 /* things we might delay */
1739                 if ((cap->issued & ~retain) == 0 &&
1740                     cap->mds_wanted == want)
1741                         continue;     /* nope, all good */
1742 
1743                 if (is_delayed)
1744                         goto ack;
1745 
1746                 /* delay? */
1747                 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1748                     time_before(jiffies, ci->i_hold_caps_max)) {
1749                         dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1750                              ceph_cap_string(cap->issued),
1751                              ceph_cap_string(cap->issued & retain),
1752                              ceph_cap_string(cap->mds_wanted),
1753                              ceph_cap_string(want));
1754                         delayed++;
1755                         continue;
1756                 }
1757 
1758 ack:
1759                 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1760                         dout(" skipping %p I_NOFLUSH set\n", inode);
1761                         continue;
1762                 }
1763 
1764                 if (session && session != cap->session) {
1765                         dout("oops, wrong session %p mutex\n", session);
1766                         mutex_unlock(&session->s_mutex);
1767                         session = NULL;
1768                 }
1769                 if (!session) {
1770                         session = cap->session;
1771                         if (mutex_trylock(&session->s_mutex) == 0) {
1772                                 dout("inverting session/ino locks on %p\n",
1773                                      session);
1774                                 spin_unlock(&ci->i_ceph_lock);
1775                                 if (took_snap_rwsem) {
1776                                         up_read(&mdsc->snap_rwsem);
1777                                         took_snap_rwsem = 0;
1778                                 }
1779                                 mutex_lock(&session->s_mutex);
1780                                 goto retry;
1781                         }
1782                 }
1783                 /* take snap_rwsem after session mutex */
1784                 if (!took_snap_rwsem) {
1785                         if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1786                                 dout("inverting snap/in locks on %p\n",
1787                                      inode);
1788                                 spin_unlock(&ci->i_ceph_lock);
1789                                 down_read(&mdsc->snap_rwsem);
1790                                 took_snap_rwsem = 1;
1791                                 goto retry;
1792                         }
1793                         took_snap_rwsem = 1;
1794                 }
1795 
1796                 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
1797                         flushing = __mark_caps_flushing(inode, session,
1798                                                         &flush_tid,
1799                                                         &oldest_flush_tid);
1800                 } else {
1801                         flushing = 0;
1802                         flush_tid = 0;
1803                         spin_lock(&mdsc->cap_dirty_lock);
1804                         oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1805                         spin_unlock(&mdsc->cap_dirty_lock);
1806                 }
1807 
1808                 mds = cap->mds;  /* remember mds, so we don't repeat */
1809                 sent++;
1810 
1811                 /* __send_cap drops i_ceph_lock */
1812                 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1813                                       want, retain, flushing,
1814                                       flush_tid, oldest_flush_tid);
1815                 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1816         }
1817 
1818         /*
1819          * Reschedule delayed caps release if we delayed anything,
1820          * otherwise cancel.
1821          */
1822         if (delayed && is_delayed)
1823                 force_requeue = 1;   /* __send_cap delayed release; requeue */
1824         if (!delayed && !is_delayed)
1825                 __cap_delay_cancel(mdsc, ci);
1826         else if (!is_delayed || force_requeue)
1827                 __cap_delay_requeue(mdsc, ci);
1828 
1829         spin_unlock(&ci->i_ceph_lock);
1830 
1831         if (queue_invalidate)
1832                 ceph_queue_invalidate(inode);
1833 
1834         if (session)
1835                 mutex_unlock(&session->s_mutex);
1836         if (took_snap_rwsem)
1837                 up_read(&mdsc->snap_rwsem);
1838 }
1839 
1840 /*
1841  * Try to flush dirty caps back to the auth mds.
1842  */
1843 static int try_flush_caps(struct inode *inode, u64 *ptid)
1844 {
1845         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1846         struct ceph_inode_info *ci = ceph_inode(inode);
1847         struct ceph_mds_session *session = NULL;
1848         int flushing = 0;
1849         u64 flush_tid = 0, oldest_flush_tid = 0;
1850 
1851 retry:
1852         spin_lock(&ci->i_ceph_lock);
1853         if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1854                 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1855                 goto out;
1856         }
1857         if (ci->i_dirty_caps && ci->i_auth_cap) {
1858                 struct ceph_cap *cap = ci->i_auth_cap;
1859                 int used = __ceph_caps_used(ci);
1860                 int want = __ceph_caps_wanted(ci);
1861                 int delayed;
1862 
1863                 if (!session || session != cap->session) {
1864                         spin_unlock(&ci->i_ceph_lock);
1865                         if (session)
1866                                 mutex_unlock(&session->s_mutex);
1867                         session = cap->session;
1868                         mutex_lock(&session->s_mutex);
1869                         goto retry;
1870                 }
1871                 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1872                         goto out;
1873 
1874                 flushing = __mark_caps_flushing(inode, session, &flush_tid,
1875                                                 &oldest_flush_tid);
1876 
1877                 /* __send_cap drops i_ceph_lock */
1878                 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1879                                      (cap->issued | cap->implemented),
1880                                      flushing, flush_tid, oldest_flush_tid);
1881 
1882                 if (delayed) {
1883                         spin_lock(&ci->i_ceph_lock);
1884                         __cap_delay_requeue(mdsc, ci);
1885                         spin_unlock(&ci->i_ceph_lock);
1886                 }
1887         } else {
1888                 struct rb_node *n = rb_last(&ci->i_cap_flush_tree);
1889                 if (n) {
1890                         struct ceph_cap_flush *cf =
1891                                 rb_entry(n, struct ceph_cap_flush, i_node);
1892                         flush_tid = cf->tid;
1893                 }
1894                 flushing = ci->i_flushing_caps;
1895                 spin_unlock(&ci->i_ceph_lock);
1896         }
1897 out:
1898         if (session)
1899                 mutex_unlock(&session->s_mutex);
1900 
1901         *ptid = flush_tid;
1902         return flushing;
1903 }
1904 
1905 /*
1906  * Return true if we've flushed caps through the given flush_tid.
1907  */
1908 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
1909 {
1910         struct ceph_inode_info *ci = ceph_inode(inode);
1911         struct ceph_cap_flush *cf;
1912         struct rb_node *n;
1913         int ret = 1;
1914 
1915         spin_lock(&ci->i_ceph_lock);
1916         n = rb_first(&ci->i_cap_flush_tree);
1917         if (n) {
1918                 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1919                 if (cf->tid <= flush_tid)
1920                         ret = 0;
1921         }
1922         spin_unlock(&ci->i_ceph_lock);
1923         return ret;
1924 }
1925 
1926 /*
1927  * Wait on any unsafe replies for the given inode.  First wait on the
1928  * newest request, and make that the upper bound.  Then, if there are
1929  * more requests, keep waiting on the oldest as long as it is still older
1930  * than the original request.
1931  */
1932 static void sync_write_wait(struct inode *inode)
1933 {
1934         struct ceph_inode_info *ci = ceph_inode(inode);
1935         struct list_head *head = &ci->i_unsafe_writes;
1936         struct ceph_osd_request *req;
1937         u64 last_tid;
1938 
1939         if (!S_ISREG(inode->i_mode))
1940                 return;
1941 
1942         spin_lock(&ci->i_unsafe_lock);
1943         if (list_empty(head))
1944                 goto out;
1945 
1946         /* set upper bound as _last_ entry in chain */
1947         req = list_last_entry(head, struct ceph_osd_request,
1948                               r_unsafe_item);
1949         last_tid = req->r_tid;
1950 
1951         do {
1952                 ceph_osdc_get_request(req);
1953                 spin_unlock(&ci->i_unsafe_lock);
1954                 dout("sync_write_wait on tid %llu (until %llu)\n",
1955                      req->r_tid, last_tid);
1956                 wait_for_completion(&req->r_safe_completion);
1957                 spin_lock(&ci->i_unsafe_lock);
1958                 ceph_osdc_put_request(req);
1959 
1960                 /*
1961                  * from here on look at first entry in chain, since we
1962                  * only want to wait for anything older than last_tid
1963                  */
1964                 if (list_empty(head))
1965                         break;
1966                 req = list_first_entry(head, struct ceph_osd_request,
1967                                        r_unsafe_item);
1968         } while (req->r_tid < last_tid);
1969 out:
1970         spin_unlock(&ci->i_unsafe_lock);
1971 }
1972 
1973 /*
1974  * wait for any uncommitted directory operations to commit.
1975  */
1976 static int unsafe_dirop_wait(struct inode *inode)
1977 {
1978         struct ceph_inode_info *ci = ceph_inode(inode);
1979         struct list_head *head = &ci->i_unsafe_dirops;
1980         struct ceph_mds_request *req;
1981         u64 last_tid;
1982         int ret = 0;
1983 
1984         if (!S_ISDIR(inode->i_mode))
1985                 return 0;
1986 
1987         spin_lock(&ci->i_unsafe_lock);
1988         if (list_empty(head))
1989                 goto out;
1990 
1991         req = list_last_entry(head, struct ceph_mds_request,
1992                               r_unsafe_dir_item);
1993         last_tid = req->r_tid;
1994 
1995         do {
1996                 ceph_mdsc_get_request(req);
1997                 spin_unlock(&ci->i_unsafe_lock);
1998 
1999                 dout("unsafe_dirop_wait %p wait on tid %llu (until %llu)\n",
2000                      inode, req->r_tid, last_tid);
2001                 ret = !wait_for_completion_timeout(&req->r_safe_completion,
2002                                         ceph_timeout_jiffies(req->r_timeout));
2003                 if (ret)
2004                         ret = -EIO;  /* timed out */
2005 
2006                 ceph_mdsc_put_request(req);
2007 
2008                 spin_lock(&ci->i_unsafe_lock);
2009                 if (ret || list_empty(head))
2010                         break;
2011                 req = list_first_entry(head, struct ceph_mds_request,
2012                                        r_unsafe_dir_item);
2013         } while (req->r_tid < last_tid);
2014 out:
2015         spin_unlock(&ci->i_unsafe_lock);
2016         return ret;
2017 }
2018 
2019 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2020 {
2021         struct inode *inode = file->f_mapping->host;
2022         struct ceph_inode_info *ci = ceph_inode(inode);
2023         u64 flush_tid;
2024         int ret;
2025         int dirty;
2026 
2027         dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2028         sync_write_wait(inode);
2029 
2030         ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
2031         if (ret < 0)
2032                 goto out;
2033 
2034         if (datasync)
2035                 goto out;
2036 
2037         mutex_lock(&inode->i_mutex);
2038 
2039         dirty = try_flush_caps(inode, &flush_tid);
2040         dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2041 
2042         ret = unsafe_dirop_wait(inode);
2043 
2044         /*
2045          * only wait on non-file metadata writeback (the mds
2046          * can recover size and mtime, so we don't need to
2047          * wait for that)
2048          */
2049         if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2050                 ret = wait_event_interruptible(ci->i_cap_wq,
2051                                         caps_are_flushed(inode, flush_tid));
2052         }
2053         mutex_unlock(&inode->i_mutex);
2054 out:
2055         dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2056         return ret;
2057 }
2058 
2059 /*
2060  * Flush any dirty caps back to the mds.  If we aren't asked to wait,
2061  * queue inode for flush but don't do so immediately, because we can
2062  * get by with fewer MDS messages if we wait for data writeback to
2063  * complete first.
2064  */
2065 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2066 {
2067         struct ceph_inode_info *ci = ceph_inode(inode);
2068         u64 flush_tid;
2069         int err = 0;
2070         int dirty;
2071         int wait = wbc->sync_mode == WB_SYNC_ALL;
2072 
2073         dout("write_inode %p wait=%d\n", inode, wait);
2074         if (wait) {
2075                 dirty = try_flush_caps(inode, &flush_tid);
2076                 if (dirty)
2077                         err = wait_event_interruptible(ci->i_cap_wq,
2078                                        caps_are_flushed(inode, flush_tid));
2079         } else {
2080                 struct ceph_mds_client *mdsc =
2081                         ceph_sb_to_client(inode->i_sb)->mdsc;
2082 
2083                 spin_lock(&ci->i_ceph_lock);
2084                 if (__ceph_caps_dirty(ci))
2085                         __cap_delay_requeue_front(mdsc, ci);
2086                 spin_unlock(&ci->i_ceph_lock);
2087         }
2088         return err;
2089 }
2090 
2091 /*
2092  * After a recovering MDS goes active, we need to resend any caps
2093  * we were flushing.
2094  *
2095  * Caller holds session->s_mutex.
2096  */
2097 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
2098                                    struct ceph_mds_session *session)
2099 {
2100         struct ceph_cap_snap *capsnap;
2101 
2102         dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
2103         list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
2104                             flushing_item) {
2105                 struct ceph_inode_info *ci = capsnap->ci;
2106                 struct inode *inode = &ci->vfs_inode;
2107                 struct ceph_cap *cap;
2108 
2109                 spin_lock(&ci->i_ceph_lock);
2110                 cap = ci->i_auth_cap;
2111                 if (cap && cap->session == session) {
2112                         dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
2113                              cap, capsnap);
2114                         __ceph_flush_snaps(ci, &session, 1);
2115                 } else {
2116                         pr_err("%p auth cap %p not mds%d ???\n", inode,
2117                                cap, session->s_mds);
2118                 }
2119                 spin_unlock(&ci->i_ceph_lock);
2120         }
2121 }
2122 
2123 static int __kick_flushing_caps(struct ceph_mds_client *mdsc,
2124                                 struct ceph_mds_session *session,
2125                                 struct ceph_inode_info *ci)
2126 {
2127         struct inode *inode = &ci->vfs_inode;
2128         struct ceph_cap *cap;
2129         struct ceph_cap_flush *cf;
2130         struct rb_node *n;
2131         int delayed = 0;
2132         u64 first_tid = 0;
2133         u64 oldest_flush_tid;
2134 
2135         spin_lock(&mdsc->cap_dirty_lock);
2136         oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2137         spin_unlock(&mdsc->cap_dirty_lock);
2138 
2139         while (true) {
2140                 spin_lock(&ci->i_ceph_lock);
2141                 cap = ci->i_auth_cap;
2142                 if (!(cap && cap->session == session)) {
2143                         pr_err("%p auth cap %p not mds%d ???\n", inode,
2144                                         cap, session->s_mds);
2145                         spin_unlock(&ci->i_ceph_lock);
2146                         break;
2147                 }
2148 
2149                 for (n = rb_first(&ci->i_cap_flush_tree); n; n = rb_next(n)) {
2150                         cf = rb_entry(n, struct ceph_cap_flush, i_node);
2151                         if (cf->tid >= first_tid)
2152                                 break;
2153                 }
2154                 if (!n) {
2155                         spin_unlock(&ci->i_ceph_lock);
2156                         break;
2157                 }
2158 
2159                 cf = rb_entry(n, struct ceph_cap_flush, i_node);
2160 
2161                 first_tid = cf->tid + 1;
2162 
2163                 dout("kick_flushing_caps %p cap %p tid %llu %s\n", inode,
2164                      cap, cf->tid, ceph_cap_string(cf->caps));
2165                 delayed |= __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2166                                       __ceph_caps_used(ci),
2167                                       __ceph_caps_wanted(ci),
2168                                       cap->issued | cap->implemented,
2169                                       cf->caps, cf->tid, oldest_flush_tid);
2170         }
2171         return delayed;
2172 }
2173 
2174 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2175                                    struct ceph_mds_session *session)
2176 {
2177         struct ceph_inode_info *ci;
2178         struct ceph_cap *cap;
2179 
2180         dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2181         list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2182                 spin_lock(&ci->i_ceph_lock);
2183                 cap = ci->i_auth_cap;
2184                 if (!(cap && cap->session == session)) {
2185                         pr_err("%p auth cap %p not mds%d ???\n",
2186                                 &ci->vfs_inode, cap, session->s_mds);
2187                         spin_unlock(&ci->i_ceph_lock);
2188                         continue;
2189                 }
2190 
2191 
2192                 /*
2193                  * if flushing caps were revoked, we re-send the cap flush
2194                  * in client reconnect stage. This guarantees MDS * processes
2195                  * the cap flush message before issuing the flushing caps to
2196                  * other client.
2197                  */
2198                 if ((cap->issued & ci->i_flushing_caps) !=
2199                     ci->i_flushing_caps) {
2200                         spin_unlock(&ci->i_ceph_lock);
2201                         if (!__kick_flushing_caps(mdsc, session, ci))
2202                                 continue;
2203                         spin_lock(&ci->i_ceph_lock);
2204                 }
2205 
2206                 spin_unlock(&ci->i_ceph_lock);
2207         }
2208 }
2209 
2210 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2211                              struct ceph_mds_session *session)
2212 {
2213         struct ceph_inode_info *ci;
2214 
2215         kick_flushing_capsnaps(mdsc, session);
2216 
2217         dout("kick_flushing_caps mds%d\n", session->s_mds);
2218         list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2219                 int delayed = __kick_flushing_caps(mdsc, session, ci);
2220                 if (delayed) {
2221                         spin_lock(&ci->i_ceph_lock);
2222                         __cap_delay_requeue(mdsc, ci);
2223                         spin_unlock(&ci->i_ceph_lock);
2224                 }
2225         }
2226 }
2227 
2228 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2229                                      struct ceph_mds_session *session,
2230                                      struct inode *inode)
2231 {
2232         struct ceph_inode_info *ci = ceph_inode(inode);
2233         struct ceph_cap *cap;
2234 
2235         spin_lock(&ci->i_ceph_lock);
2236         cap = ci->i_auth_cap;
2237         dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2238              ceph_cap_string(ci->i_flushing_caps));
2239 
2240         __ceph_flush_snaps(ci, &session, 1);
2241 
2242         if (ci->i_flushing_caps) {
2243                 int delayed;
2244 
2245                 spin_lock(&mdsc->cap_dirty_lock);
2246                 list_move_tail(&ci->i_flushing_item,
2247                                &cap->session->s_cap_flushing);
2248                 spin_unlock(&mdsc->cap_dirty_lock);
2249 
2250                 spin_unlock(&ci->i_ceph_lock);
2251 
2252                 delayed = __kick_flushing_caps(mdsc, session, ci);
2253                 if (delayed) {
2254                         spin_lock(&ci->i_ceph_lock);
2255                         __cap_delay_requeue(mdsc, ci);
2256                         spin_unlock(&ci->i_ceph_lock);
2257                 }
2258         } else {
2259                 spin_unlock(&ci->i_ceph_lock);
2260         }
2261 }
2262 
2263 
2264 /*
2265  * Take references to capabilities we hold, so that we don't release
2266  * them to the MDS prematurely.
2267  *
2268  * Protected by i_ceph_lock.
2269  */
2270 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2271                             bool snap_rwsem_locked)
2272 {
2273         if (got & CEPH_CAP_PIN)
2274                 ci->i_pin_ref++;
2275         if (got & CEPH_CAP_FILE_RD)
2276                 ci->i_rd_ref++;
2277         if (got & CEPH_CAP_FILE_CACHE)
2278                 ci->i_rdcache_ref++;
2279         if (got & CEPH_CAP_FILE_WR) {
2280                 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2281                         BUG_ON(!snap_rwsem_locked);
2282                         ci->i_head_snapc = ceph_get_snap_context(
2283                                         ci->i_snap_realm->cached_context);
2284                 }
2285                 ci->i_wr_ref++;
2286         }
2287         if (got & CEPH_CAP_FILE_BUFFER) {
2288                 if (ci->i_wb_ref == 0)
2289                         ihold(&ci->vfs_inode);
2290                 ci->i_wb_ref++;
2291                 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2292                      &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2293         }
2294 }
2295 
2296 /*
2297  * Try to grab cap references.  Specify those refs we @want, and the
2298  * minimal set we @need.  Also include the larger offset we are writing
2299  * to (when applicable), and check against max_size here as well.
2300  * Note that caller is responsible for ensuring max_size increases are
2301  * requested from the MDS.
2302  */
2303 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2304                             loff_t endoff, bool nonblock, int *got, int *err)
2305 {
2306         struct inode *inode = &ci->vfs_inode;
2307         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2308         int ret = 0;
2309         int have, implemented;
2310         int file_wanted;
2311         bool snap_rwsem_locked = false;
2312 
2313         dout("get_cap_refs %p need %s want %s\n", inode,
2314              ceph_cap_string(need), ceph_cap_string(want));
2315 
2316 again:
2317         spin_lock(&ci->i_ceph_lock);
2318 
2319         /* make sure file is actually open */
2320         file_wanted = __ceph_caps_file_wanted(ci);
2321         if ((file_wanted & need) == 0) {
2322                 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2323                      ceph_cap_string(need), ceph_cap_string(file_wanted));
2324                 *err = -EBADF;
2325                 ret = 1;
2326                 goto out_unlock;
2327         }
2328 
2329         /* finish pending truncate */
2330         while (ci->i_truncate_pending) {
2331                 spin_unlock(&ci->i_ceph_lock);
2332                 if (snap_rwsem_locked) {
2333                         up_read(&mdsc->snap_rwsem);
2334                         snap_rwsem_locked = false;
2335                 }
2336                 __ceph_do_pending_vmtruncate(inode);
2337                 spin_lock(&ci->i_ceph_lock);
2338         }
2339 
2340         have = __ceph_caps_issued(ci, &implemented);
2341 
2342         if (have & need & CEPH_CAP_FILE_WR) {
2343                 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2344                         dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2345                              inode, endoff, ci->i_max_size);
2346                         if (endoff > ci->i_requested_max_size) {
2347                                 *err = -EAGAIN;
2348                                 ret = 1;
2349                         }
2350                         goto out_unlock;
2351                 }
2352                 /*
2353                  * If a sync write is in progress, we must wait, so that we
2354                  * can get a final snapshot value for size+mtime.
2355                  */
2356                 if (__ceph_have_pending_cap_snap(ci)) {
2357                         dout("get_cap_refs %p cap_snap_pending\n", inode);
2358                         goto out_unlock;
2359                 }
2360         }
2361 
2362         if ((have & need) == need) {
2363                 /*
2364                  * Look at (implemented & ~have & not) so that we keep waiting
2365                  * on transition from wanted -> needed caps.  This is needed
2366                  * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2367                  * going before a prior buffered writeback happens.
2368                  */
2369                 int not = want & ~(have & need);
2370                 int revoking = implemented & ~have;
2371                 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2372                      inode, ceph_cap_string(have), ceph_cap_string(not),
2373                      ceph_cap_string(revoking));
2374                 if ((revoking & not) == 0) {
2375                         if (!snap_rwsem_locked &&
2376                             !ci->i_head_snapc &&
2377                             (need & CEPH_CAP_FILE_WR)) {
2378                                 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2379                                         /*
2380                                          * we can not call down_read() when
2381                                          * task isn't in TASK_RUNNING state
2382                                          */
2383                                         if (nonblock) {
2384                                                 *err = -EAGAIN;
2385                                                 ret = 1;
2386                                                 goto out_unlock;
2387                                         }
2388 
2389                                         spin_unlock(&ci->i_ceph_lock);
2390                                         down_read(&mdsc->snap_rwsem);
2391                                         snap_rwsem_locked = true;
2392                                         goto again;
2393                                 }
2394                                 snap_rwsem_locked = true;
2395                         }
2396                         *got = need | (have & want);
2397                         __take_cap_refs(ci, *got, true);
2398                         ret = 1;
2399                 }
2400         } else {
2401                 int session_readonly = false;
2402                 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2403                         struct ceph_mds_session *s = ci->i_auth_cap->session;
2404                         spin_lock(&s->s_cap_lock);
2405                         session_readonly = s->s_readonly;
2406                         spin_unlock(&s->s_cap_lock);
2407                 }
2408                 if (session_readonly) {
2409                         dout("get_cap_refs %p needed %s but mds%d readonly\n",
2410                              inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2411                         *err = -EROFS;
2412                         ret = 1;
2413                         goto out_unlock;
2414                 }
2415 
2416                 dout("get_cap_refs %p have %s needed %s\n", inode,
2417                      ceph_cap_string(have), ceph_cap_string(need));
2418         }
2419 out_unlock:
2420         spin_unlock(&ci->i_ceph_lock);
2421         if (snap_rwsem_locked)
2422                 up_read(&mdsc->snap_rwsem);
2423 
2424         dout("get_cap_refs %p ret %d got %s\n", inode,
2425              ret, ceph_cap_string(*got));
2426         return ret;
2427 }
2428 
2429 /*
2430  * Check the offset we are writing up to against our current
2431  * max_size.  If necessary, tell the MDS we want to write to
2432  * a larger offset.
2433  */
2434 static void check_max_size(struct inode *inode, loff_t endoff)
2435 {
2436         struct ceph_inode_info *ci = ceph_inode(inode);
2437         int check = 0;
2438 
2439         /* do we need to explicitly request a larger max_size? */
2440         spin_lock(&ci->i_ceph_lock);
2441         if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2442                 dout("write %p at large endoff %llu, req max_size\n",
2443                      inode, endoff);
2444                 ci->i_wanted_max_size = endoff;
2445         }
2446         /* duplicate ceph_check_caps()'s logic */
2447         if (ci->i_auth_cap &&
2448             (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2449             ci->i_wanted_max_size > ci->i_max_size &&
2450             ci->i_wanted_max_size > ci->i_requested_max_size)
2451                 check = 1;
2452         spin_unlock(&ci->i_ceph_lock);
2453         if (check)
2454                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2455 }
2456 
2457 /*
2458  * Wait for caps, and take cap references.  If we can't get a WR cap
2459  * due to a small max_size, make sure we check_max_size (and possibly
2460  * ask the mds) so we don't get hung up indefinitely.
2461  */
2462 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2463                   loff_t endoff, int *got, struct page **pinned_page)
2464 {
2465         int _got, ret, err = 0;
2466 
2467         ret = ceph_pool_perm_check(ci, need);
2468         if (ret < 0)
2469                 return ret;
2470 
2471         while (true) {
2472                 if (endoff > 0)
2473                         check_max_size(&ci->vfs_inode, endoff);
2474 
2475                 err = 0;
2476                 _got = 0;
2477                 ret = try_get_cap_refs(ci, need, want, endoff,
2478                                        false, &_got, &err);
2479                 if (ret) {
2480                         if (err == -EAGAIN)
2481                                 continue;
2482                         if (err < 0)
2483                                 return err;
2484                 } else {
2485                         ret = wait_event_interruptible(ci->i_cap_wq,
2486                                         try_get_cap_refs(ci, need, want, endoff,
2487                                                          true, &_got, &err));
2488                         if (err == -EAGAIN)
2489                                 continue;
2490                         if (err < 0)
2491                                 ret = err;
2492                         if (ret < 0)
2493                                 return ret;
2494                 }
2495 
2496                 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2497                     (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2498                     i_size_read(&ci->vfs_inode) > 0) {
2499                         struct page *page =
2500                                 find_get_page(ci->vfs_inode.i_mapping, 0);
2501                         if (page) {
2502                                 if (PageUptodate(page)) {
2503                                         *pinned_page = page;
2504                                         break;
2505                                 }
2506                                 page_cache_release(page);
2507                         }
2508                         /*
2509                          * drop cap refs first because getattr while
2510                          * holding * caps refs can cause deadlock.
2511                          */
2512                         ceph_put_cap_refs(ci, _got);
2513                         _got = 0;
2514 
2515                         /*
2516                          * getattr request will bring inline data into
2517                          * page cache
2518                          */
2519                         ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2520                                                 CEPH_STAT_CAP_INLINE_DATA,
2521                                                 true);
2522                         if (ret < 0)
2523                                 return ret;
2524                         continue;
2525                 }
2526                 break;
2527         }
2528 
2529         *got = _got;
2530         return 0;
2531 }
2532 
2533 /*
2534  * Take cap refs.  Caller must already know we hold at least one ref
2535  * on the caps in question or we don't know this is safe.
2536  */
2537 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2538 {
2539         spin_lock(&ci->i_ceph_lock);
2540         __take_cap_refs(ci, caps, false);
2541         spin_unlock(&ci->i_ceph_lock);
2542 }
2543 
2544 
2545 /*
2546  * drop cap_snap that is not associated with any snapshot.
2547  * we don't need to send FLUSHSNAP message for it.
2548  */
2549 static int ceph_try_drop_cap_snap(struct ceph_cap_snap *capsnap)
2550 {
2551         if (!capsnap->need_flush &&
2552             !capsnap->writing && !capsnap->dirty_pages) {
2553 
2554                 dout("dropping cap_snap %p follows %llu\n",
2555                      capsnap, capsnap->follows);
2556                 ceph_put_snap_context(capsnap->context);
2557                 list_del(&capsnap->ci_item);
2558                 list_del(&capsnap->flushing_item);
2559                 ceph_put_cap_snap(capsnap);
2560                 return 1;
2561         }
2562         return 0;
2563 }
2564 
2565 /*
2566  * Release cap refs.
2567  *
2568  * If we released the last ref on any given cap, call ceph_check_caps
2569  * to release (or schedule a release).
2570  *
2571  * If we are releasing a WR cap (from a sync write), finalize any affected
2572  * cap_snap, and wake up any waiters.
2573  */
2574 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2575 {
2576         struct inode *inode = &ci->vfs_inode;
2577         int last = 0, put = 0, flushsnaps = 0, wake = 0;
2578 
2579         spin_lock(&ci->i_ceph_lock);
2580         if (had & CEPH_CAP_PIN)
2581                 --ci->i_pin_ref;
2582         if (had & CEPH_CAP_FILE_RD)
2583                 if (--ci->i_rd_ref == 0)
2584                         last++;
2585         if (had & CEPH_CAP_FILE_CACHE)
2586                 if (--ci->i_rdcache_ref == 0)
2587                         last++;
2588         if (had & CEPH_CAP_FILE_BUFFER) {
2589                 if (--ci->i_wb_ref == 0) {
2590                         last++;
2591                         put++;
2592                 }
2593                 dout("put_cap_refs %p wb %d -> %d (?)\n",
2594                      inode, ci->i_wb_ref+1, ci->i_wb_ref);
2595         }
2596         if (had & CEPH_CAP_FILE_WR)
2597                 if (--ci->i_wr_ref == 0) {
2598                         last++;
2599                         if (__ceph_have_pending_cap_snap(ci)) {
2600                                 struct ceph_cap_snap *capsnap =
2601                                         list_last_entry(&ci->i_cap_snaps,
2602                                                         struct ceph_cap_snap,
2603                                                         ci_item);
2604                                 capsnap->writing = 0;
2605                                 if (ceph_try_drop_cap_snap(capsnap))
2606                                         put++;
2607                                 else if (__ceph_finish_cap_snap(ci, capsnap))
2608                                         flushsnaps = 1;
2609                                 wake = 1;
2610                         }
2611                         if (ci->i_wrbuffer_ref_head == 0 &&
2612                             ci->i_dirty_caps == 0 &&
2613                             ci->i_flushing_caps == 0) {
2614                                 BUG_ON(!ci->i_head_snapc);
2615                                 ceph_put_snap_context(ci->i_head_snapc);
2616                                 ci->i_head_snapc = NULL;
2617                         }
2618                         /* see comment in __ceph_remove_cap() */
2619                         if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2620                                 drop_inode_snap_realm(ci);
2621                 }
2622         spin_unlock(&ci->i_ceph_lock);
2623 
2624         dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2625              last ? " last" : "", put ? " put" : "");
2626 
2627         if (last && !flushsnaps)
2628                 ceph_check_caps(ci, 0, NULL);
2629         else if (flushsnaps)
2630                 ceph_flush_snaps(ci);
2631         if (wake)
2632                 wake_up_all(&ci->i_cap_wq);
2633         while (put-- > 0)
2634                 iput(inode);
2635 }
2636 
2637 /*
2638  * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2639  * context.  Adjust per-snap dirty page accounting as appropriate.
2640  * Once all dirty data for a cap_snap is flushed, flush snapped file
2641  * metadata back to the MDS.  If we dropped the last ref, call
2642  * ceph_check_caps.
2643  */
2644 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2645                                 struct ceph_snap_context *snapc)
2646 {
2647         struct inode *inode = &ci->vfs_inode;
2648         int last = 0;
2649         int complete_capsnap = 0;
2650         int drop_capsnap = 0;
2651         int found = 0;
2652         struct ceph_cap_snap *capsnap = NULL;
2653 
2654         spin_lock(&ci->i_ceph_lock);
2655         ci->i_wrbuffer_ref -= nr;
2656         last = !ci->i_wrbuffer_ref;
2657 
2658         if (ci->i_head_snapc == snapc) {
2659                 ci->i_wrbuffer_ref_head -= nr;
2660                 if (ci->i_wrbuffer_ref_head == 0 &&
2661                     ci->i_wr_ref == 0 &&
2662                     ci->i_dirty_caps == 0 &&
2663                     ci->i_flushing_caps == 0) {
2664                         BUG_ON(!ci->i_head_snapc);
2665                         ceph_put_snap_context(ci->i_head_snapc);
2666                         ci->i_head_snapc = NULL;
2667                 }
2668                 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2669                      inode,
2670                      ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2671                      ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2672                      last ? " LAST" : "");
2673         } else {
2674                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2675                         if (capsnap->context == snapc) {
2676                                 found = 1;
2677                                 break;
2678                         }
2679                 }
2680                 BUG_ON(!found);
2681                 capsnap->dirty_pages -= nr;
2682                 if (capsnap->dirty_pages == 0) {
2683                         complete_capsnap = 1;
2684                         drop_capsnap = ceph_try_drop_cap_snap(capsnap);
2685                 }
2686                 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2687                      " snap %lld %d/%d -> %d/%d %s%s\n",
2688                      inode, capsnap, capsnap->context->seq,
2689                      ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2690                      ci->i_wrbuffer_ref, capsnap->dirty_pages,
2691                      last ? " (wrbuffer last)" : "",
2692                      complete_capsnap ? " (complete capsnap)" : "");
2693         }
2694 
2695         spin_unlock(&ci->i_ceph_lock);
2696 
2697         if (last) {
2698                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2699                 iput(inode);
2700         } else if (complete_capsnap) {
2701                 ceph_flush_snaps(ci);
2702                 wake_up_all(&ci->i_cap_wq);
2703         }
2704         if (drop_capsnap)
2705                 iput(inode);
2706 }
2707 
2708 /*
2709  * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2710  */
2711 static void invalidate_aliases(struct inode *inode)
2712 {
2713         struct dentry *dn, *prev = NULL;
2714 
2715         dout("invalidate_aliases inode %p\n", inode);
2716         d_prune_aliases(inode);
2717         /*
2718          * For non-directory inode, d_find_alias() only returns
2719          * hashed dentry. After calling d_invalidate(), the
2720          * dentry becomes unhashed.
2721          *
2722          * For directory inode, d_find_alias() can return
2723          * unhashed dentry. But directory inode should have
2724          * one alias at most.
2725          */
2726         while ((dn = d_find_alias(inode))) {
2727                 if (dn == prev) {
2728                         dput(dn);
2729                         break;
2730                 }
2731                 d_invalidate(dn);
2732                 if (prev)
2733                         dput(prev);
2734                 prev = dn;
2735         }
2736         if (prev)
2737                 dput(prev);
2738 }
2739 
2740 /*
2741  * Handle a cap GRANT message from the MDS.  (Note that a GRANT may
2742  * actually be a revocation if it specifies a smaller cap set.)
2743  *
2744  * caller holds s_mutex and i_ceph_lock, we drop both.
2745  */
2746 static void handle_cap_grant(struct ceph_mds_client *mdsc,
2747                              struct inode *inode, struct ceph_mds_caps *grant,
2748                              u64 inline_version,
2749                              void *inline_data, int inline_len,
2750                              struct ceph_buffer *xattr_buf,
2751                              struct ceph_mds_session *session,
2752                              struct ceph_cap *cap, int issued)
2753         __releases(ci->i_ceph_lock)
2754         __releases(mdsc->snap_rwsem)
2755 {
2756         struct ceph_inode_info *ci = ceph_inode(inode);
2757         int mds = session->s_mds;
2758         int seq = le32_to_cpu(grant->seq);
2759         int newcaps = le32_to_cpu(grant->caps);
2760         int used, wanted, dirty;
2761         u64 size = le64_to_cpu(grant->size);
2762         u64 max_size = le64_to_cpu(grant->max_size);
2763         struct timespec mtime, atime, ctime;
2764         int check_caps = 0;
2765         bool wake = false;
2766         bool writeback = false;
2767         bool queue_trunc = false;
2768         bool queue_invalidate = false;
2769         bool queue_revalidate = false;
2770         bool deleted_inode = false;
2771         bool fill_inline = false;
2772 
2773         dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2774              inode, cap, mds, seq, ceph_cap_string(newcaps));
2775         dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2776                 inode->i_size);
2777 
2778 
2779         /*
2780          * auth mds of the inode changed. we received the cap export message,
2781          * but still haven't received the cap import message. handle_cap_export
2782          * updated the new auth MDS' cap.
2783          *
2784          * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2785          * that was sent before the cap import message. So don't remove caps.
2786          */
2787         if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2788                 WARN_ON(cap != ci->i_auth_cap);
2789                 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2790                 seq = cap->seq;
2791                 newcaps |= cap->issued;
2792         }
2793 
2794         /*
2795          * If CACHE is being revoked, and we have no dirty buffers,
2796          * try to invalidate (once).  (If there are dirty buffers, we
2797          * will invalidate _after_ writeback.)
2798          */
2799         if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
2800             ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2801             (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2802             !ci->i_wrbuffer_ref) {
2803                 if (try_nonblocking_invalidate(inode)) {
2804                         /* there were locked pages.. invalidate later
2805                            in a separate thread. */
2806                         if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2807                                 queue_invalidate = true;
2808                                 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2809                         }
2810                 }
2811 
2812                 ceph_fscache_invalidate(inode);
2813         }
2814 
2815         /* side effects now are allowed */
2816         cap->cap_gen = session->s_cap_gen;
2817         cap->seq = seq;
2818 
2819         __check_cap_issue(ci, cap, newcaps);
2820 
2821         if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2822             (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2823                 inode->i_mode = le32_to_cpu(grant->mode);
2824                 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2825                 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2826                 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2827                      from_kuid(&init_user_ns, inode->i_uid),
2828                      from_kgid(&init_user_ns, inode->i_gid));
2829         }
2830 
2831         if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2832             (issued & CEPH_CAP_LINK_EXCL) == 0) {
2833                 set_nlink(inode, le32_to_cpu(grant->nlink));
2834                 if (inode->i_nlink == 0 &&
2835                     (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2836                         deleted_inode = true;
2837         }
2838 
2839         if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2840                 int len = le32_to_cpu(grant->xattr_len);
2841                 u64 version = le64_to_cpu(grant->xattr_version);
2842 
2843                 if (version > ci->i_xattrs.version) {
2844                         dout(" got new xattrs v%llu on %p len %d\n",
2845                              version, inode, len);
2846                         if (ci->i_xattrs.blob)
2847                                 ceph_buffer_put(ci->i_xattrs.blob);
2848                         ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2849                         ci->i_xattrs.version = version;
2850                         ceph_forget_all_cached_acls(inode);
2851                 }
2852         }
2853 
2854         /* Do we need to revalidate our fscache cookie. Don't bother on the
2855          * first cache cap as we already validate at cookie creation time. */
2856         if ((issued & CEPH_CAP_FILE_CACHE) && ci->i_rdcache_gen > 1)
2857                 queue_revalidate = true;
2858 
2859         if (newcaps & CEPH_CAP_ANY_RD) {
2860                 /* ctime/mtime/atime? */
2861                 ceph_decode_timespec(&mtime, &grant->mtime);
2862                 ceph_decode_timespec(&atime, &grant->atime);
2863                 ceph_decode_timespec(&ctime, &grant->ctime);
2864                 ceph_fill_file_time(inode, issued,
2865                                     le32_to_cpu(grant->time_warp_seq),
2866                                     &ctime, &mtime, &atime);
2867         }
2868 
2869         if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
2870                 /* file layout may have changed */
2871                 ci->i_layout = grant->layout;
2872                 /* size/truncate_seq? */
2873                 queue_trunc = ceph_fill_file_size(inode, issued,
2874                                         le32_to_cpu(grant->truncate_seq),
2875                                         le64_to_cpu(grant->truncate_size),
2876                                         size);
2877                 /* max size increase? */
2878                 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
2879                         dout("max_size %lld -> %llu\n",
2880                              ci->i_max_size, max_size);
2881                         ci->i_max_size = max_size;
2882                         if (max_size >= ci->i_wanted_max_size) {
2883                                 ci->i_wanted_max_size = 0;  /* reset */
2884                                 ci->i_requested_max_size = 0;
2885                         }
2886                         wake = true;
2887                 }
2888         }
2889 
2890         /* check cap bits */
2891         wanted = __ceph_caps_wanted(ci);
2892         used = __ceph_caps_used(ci);
2893         dirty = __ceph_caps_dirty(ci);
2894         dout(" my wanted = %s, used = %s, dirty %s\n",
2895              ceph_cap_string(wanted),
2896              ceph_cap_string(used),
2897              ceph_cap_string(dirty));
2898         if (wanted != le32_to_cpu(grant->wanted)) {
2899                 dout("mds wanted %s -> %s\n",
2900                      ceph_cap_string(le32_to_cpu(grant->wanted)),
2901                      ceph_cap_string(wanted));
2902                 /* imported cap may not have correct mds_wanted */
2903                 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
2904                         check_caps = 1;
2905         }
2906 
2907         /* revocation, grant, or no-op? */
2908         if (cap->issued & ~newcaps) {
2909                 int revoking = cap->issued & ~newcaps;
2910 
2911                 dout("revocation: %s -> %s (revoking %s)\n",
2912                      ceph_cap_string(cap->issued),
2913                      ceph_cap_string(newcaps),
2914                      ceph_cap_string(revoking));
2915                 if (revoking & used & CEPH_CAP_FILE_BUFFER)
2916                         writeback = true;  /* initiate writeback; will delay ack */
2917                 else if (revoking == CEPH_CAP_FILE_CACHE &&
2918                          (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2919                          queue_invalidate)
2920                         ; /* do nothing yet, invalidation will be queued */
2921                 else if (cap == ci->i_auth_cap)
2922                         check_caps = 1; /* check auth cap only */
2923                 else
2924                         check_caps = 2; /* check all caps */
2925                 cap->issued = newcaps;
2926                 cap->implemented |= newcaps;
2927         } else if (cap->issued == newcaps) {
2928                 dout("caps unchanged: %s -> %s\n",
2929                      ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2930         } else {
2931                 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2932                      ceph_cap_string(newcaps));
2933                 /* non-auth MDS is revoking the newly grant caps ? */
2934                 if (cap == ci->i_auth_cap &&
2935                     __ceph_caps_revoking_other(ci, cap, newcaps))
2936                     check_caps = 2;
2937 
2938                 cap->issued = newcaps;
2939                 cap->implemented |= newcaps; /* add bits only, to
2940                                               * avoid stepping on a
2941                                               * pending revocation */
2942                 wake = true;
2943         }
2944         BUG_ON(cap->issued & ~cap->implemented);
2945 
2946         if (inline_version > 0 && inline_version >= ci->i_inline_version) {
2947                 ci->i_inline_version = inline_version;
2948                 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2949                     (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
2950                         fill_inline = true;
2951         }
2952 
2953         spin_unlock(&ci->i_ceph_lock);
2954 
2955         if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
2956                 kick_flushing_inode_caps(mdsc, session, inode);
2957                 up_read(&mdsc->snap_rwsem);
2958                 if (newcaps & ~issued)
2959                         wake = true;
2960         }
2961 
2962         if (fill_inline)
2963                 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
2964 
2965         if (queue_trunc) {
2966                 ceph_queue_vmtruncate(inode);
2967                 ceph_queue_revalidate(inode);
2968         } else if (queue_revalidate)
2969                 ceph_queue_revalidate(inode);
2970 
2971         if (writeback)
2972                 /*
2973                  * queue inode for writeback: we can't actually call
2974                  * filemap_write_and_wait, etc. from message handler
2975                  * context.
2976                  */
2977                 ceph_queue_writeback(inode);
2978         if (queue_invalidate)
2979                 ceph_queue_invalidate(inode);
2980         if (deleted_inode)
2981                 invalidate_aliases(inode);
2982         if (wake)
2983                 wake_up_all(&ci->i_cap_wq);
2984 
2985         if (check_caps == 1)
2986                 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2987                                 session);
2988         else if (check_caps == 2)
2989                 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2990         else
2991                 mutex_unlock(&session->s_mutex);
2992 }
2993 
2994 /*
2995  * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2996  * MDS has been safely committed.
2997  */
2998 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2999                                  struct ceph_mds_caps *m,
3000                                  struct ceph_mds_session *session,
3001                                  struct ceph_cap *cap)
3002         __releases(ci->i_ceph_lock)
3003 {
3004         struct ceph_inode_info *ci = ceph_inode(inode);
3005         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3006         struct ceph_cap_flush *cf;
3007         struct rb_node *n;
3008         LIST_HEAD(to_remove);
3009         unsigned seq = le32_to_cpu(m->seq);
3010         int dirty = le32_to_cpu(m->dirty);
3011         int cleaned = 0;
3012         int drop = 0;
3013 
3014         n = rb_first(&ci->i_cap_flush_tree);
3015         while (n) {
3016                 cf = rb_entry(n, struct ceph_cap_flush, i_node);
3017                 n = rb_next(&cf->i_node);
3018                 if (cf->tid == flush_tid)
3019                         cleaned = cf->caps;
3020                 if (cf->tid <= flush_tid) {
3021                         rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
3022                         list_add_tail(&cf->list, &to_remove);
3023                 } else {
3024                         cleaned &= ~cf->caps;
3025                         if (!cleaned)
3026                                 break;
3027                 }
3028         }
3029 
3030         dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3031              " flushing %s -> %s\n",
3032              inode, session->s_mds, seq, ceph_cap_string(dirty),
3033              ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3034              ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3035 
3036         if (list_empty(&to_remove) && !cleaned)
3037                 goto out;
3038 
3039         ci->i_flushing_caps &= ~cleaned;
3040 
3041         spin_lock(&mdsc->cap_dirty_lock);
3042 
3043         if (!list_empty(&to_remove)) {
3044                 list_for_each_entry(cf, &to_remove, list)
3045                         rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
3046 
3047                 n = rb_first(&mdsc->cap_flush_tree);
3048                 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
3049                 if (!cf || cf->tid > flush_tid)
3050                         wake_up_all(&mdsc->cap_flushing_wq);
3051         }
3052 
3053         if (ci->i_flushing_caps == 0) {
3054                 list_del_init(&ci->i_flushing_item);
3055                 if (!list_empty(&session->s_cap_flushing))
3056                         dout(" mds%d still flushing cap on %p\n",
3057                              session->s_mds,
3058                              &list_entry(session->s_cap_flushing.next,
3059                                          struct ceph_inode_info,
3060                                          i_flushing_item)->vfs_inode);
3061                 mdsc->num_cap_flushing--;
3062                 dout(" inode %p now !flushing\n", inode);
3063 
3064                 if (ci->i_dirty_caps == 0) {
3065                         dout(" inode %p now clean\n", inode);
3066                         BUG_ON(!list_empty(&ci->i_dirty_item));
3067                         drop = 1;
3068                         if (ci->i_wr_ref == 0 &&
3069                             ci->i_wrbuffer_ref_head == 0) {
3070                                 BUG_ON(!ci->i_head_snapc);
3071                                 ceph_put_snap_context(ci->i_head_snapc);
3072                                 ci->i_head_snapc = NULL;
3073                         }
3074                 } else {
3075                         BUG_ON(list_empty(&ci->i_dirty_item));
3076                 }
3077         }
3078         spin_unlock(&mdsc->cap_dirty_lock);
3079         wake_up_all(&ci->i_cap_wq);
3080 
3081 out:
3082         spin_unlock(&ci->i_ceph_lock);
3083 
3084         while (!list_empty(&to_remove)) {
3085                 cf = list_first_entry(&to_remove,
3086                                       struct ceph_cap_flush, list);
3087                 list_del(&cf->list);
3088                 ceph_free_cap_flush(cf);
3089         }
3090         if (drop)
3091                 iput(inode);
3092 }
3093 
3094 /*
3095  * Handle FLUSHSNAP_ACK.  MDS has flushed snap data to disk and we can
3096  * throw away our cap_snap.
3097  *
3098  * Caller hold s_mutex.
3099  */
3100 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3101                                      struct ceph_mds_caps *m,
3102                                      struct ceph_mds_session *session)
3103 {
3104         struct ceph_inode_info *ci = ceph_inode(inode);
3105         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3106         u64 follows = le64_to_cpu(m->snap_follows);
3107         struct ceph_cap_snap *capsnap;
3108         int drop = 0;
3109 
3110         dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3111              inode, ci, session->s_mds, follows);
3112 
3113         spin_lock(&ci->i_ceph_lock);
3114         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3115                 if (capsnap->follows == follows) {
3116                         if (capsnap->flush_tid != flush_tid) {
3117                                 dout(" cap_snap %p follows %lld tid %lld !="
3118                                      " %lld\n", capsnap, follows,
3119                                      flush_tid, capsnap->flush_tid);
3120                                 break;
3121                         }
3122                         WARN_ON(capsnap->dirty_pages || capsnap->writing);
3123                         dout(" removing %p cap_snap %p follows %lld\n",
3124                              inode, capsnap, follows);
3125                         ceph_put_snap_context(capsnap->context);
3126                         list_del(&capsnap->ci_item);
3127                         list_del(&capsnap->flushing_item);
3128                         ceph_put_cap_snap(capsnap);
3129                         wake_up_all(&mdsc->cap_flushing_wq);
3130                         drop = 1;
3131                         break;
3132                 } else {
3133                         dout(" skipping cap_snap %p follows %lld\n",
3134                              capsnap, capsnap->follows);
3135                 }
3136         }
3137         spin_unlock(&ci->i_ceph_lock);
3138         if (drop)
3139                 iput(inode);
3140 }
3141 
3142 /*
3143  * Handle TRUNC from MDS, indicating file truncation.
3144  *
3145  * caller hold s_mutex.
3146  */
3147 static void handle_cap_trunc(struct inode *inode,
3148                              struct ceph_mds_caps *trunc,
3149                              struct ceph_mds_session *session)
3150         __releases(ci->i_ceph_lock)
3151 {
3152         struct ceph_inode_info *ci = ceph_inode(inode);
3153         int mds = session->s_mds;
3154         int seq = le32_to_cpu(trunc->seq);
3155         u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3156         u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3157         u64 size = le64_to_cpu(trunc->size);
3158         int implemented = 0;
3159         int dirty = __ceph_caps_dirty(ci);
3160         int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3161         int queue_trunc = 0;
3162 
3163         issued |= implemented | dirty;
3164 
3165         dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3166              inode, mds, seq, truncate_size, truncate_seq);
3167         queue_trunc = ceph_fill_file_size(inode, issued,
3168                                           truncate_seq, truncate_size, size);
3169         spin_unlock(&ci->i_ceph_lock);
3170 
3171         if (queue_trunc) {
3172                 ceph_queue_vmtruncate(inode);
3173                 ceph_fscache_invalidate(inode);
3174         }
3175 }
3176 
3177 /*
3178  * Handle EXPORT from MDS.  Cap is being migrated _from_ this mds to a
3179  * different one.  If we are the most recent migration we've seen (as
3180  * indicated by mseq), make note of the migrating cap bits for the
3181  * duration (until we see the corresponding IMPORT).
3182  *
3183  * caller holds s_mutex
3184  */
3185 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3186                               struct ceph_mds_cap_peer *ph,
3187                               struct ceph_mds_session *session)
3188 {
3189         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3190         struct ceph_mds_session *tsession = NULL;
3191         struct ceph_cap *cap, *tcap, *new_cap = NULL;
3192         struct ceph_inode_info *ci = ceph_inode(inode);
3193         u64 t_cap_id;
3194         unsigned mseq = le32_to_cpu(ex->migrate_seq);
3195         unsigned t_seq, t_mseq;
3196         int target, issued;
3197         int mds = session->s_mds;
3198 
3199         if (ph) {
3200                 t_cap_id = le64_to_cpu(ph->cap_id);
3201                 t_seq = le32_to_cpu(ph->seq);
3202                 t_mseq = le32_to_cpu(ph->mseq);
3203                 target = le32_to_cpu(ph->mds);
3204         } else {
3205                 t_cap_id = t_seq = t_mseq = 0;
3206                 target = -1;
3207         }
3208 
3209         dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3210              inode, ci, mds, mseq, target);
3211 retry:
3212         spin_lock(&ci->i_ceph_lock);
3213         cap = __get_cap_for_mds(ci, mds);
3214         if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3215                 goto out_unlock;
3216 
3217         if (target < 0) {
3218                 __ceph_remove_cap(cap, false);
3219                 goto out_unlock;
3220         }
3221 
3222         /*
3223          * now we know we haven't received the cap import message yet
3224          * because the exported cap still exist.
3225          */
3226 
3227         issued = cap->issued;
3228         WARN_ON(issued != cap->implemented);
3229 
3230         tcap = __get_cap_for_mds(ci, target);
3231         if (tcap) {
3232                 /* already have caps from the target */
3233                 if (tcap->cap_id != t_cap_id ||
3234                     ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3235                         dout(" updating import cap %p mds%d\n", tcap, target);
3236                         tcap->cap_id = t_cap_id;
3237                         tcap->seq = t_seq - 1;
3238                         tcap->issue_seq = t_seq - 1;
3239                         tcap->mseq = t_mseq;
3240                         tcap->issued |= issued;
3241                         tcap->implemented |= issued;
3242                         if (cap == ci->i_auth_cap)
3243                                 ci->i_auth_cap = tcap;
3244                         if (ci->i_flushing_caps && ci->i_auth_cap == tcap) {
3245                                 spin_lock(&mdsc->cap_dirty_lock);
3246                                 list_move_tail(&ci->i_flushing_item,
3247                                                &tcap->session->s_cap_flushing);
3248                                 spin_unlock(&mdsc->cap_dirty_lock);
3249                         }
3250                 }
3251                 __ceph_remove_cap(cap, false);
3252                 goto out_unlock;
3253         } else if (tsession) {
3254                 /* add placeholder for the export tagert */
3255                 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3256                 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3257                              t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3258 
3259                 __ceph_remove_cap(cap, false);
3260                 goto out_unlock;
3261         }
3262 
3263         spin_unlock(&ci->i_ceph_lock);
3264         mutex_unlock(&session->s_mutex);
3265 
3266         /* open target session */
3267         tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3268         if (!IS_ERR(tsession)) {
3269                 if (mds > target) {
3270                         mutex_lock(&session->s_mutex);
3271                         mutex_lock_nested(&tsession->s_mutex,
3272                                           SINGLE_DEPTH_NESTING);
3273                 } else {
3274                         mutex_lock(&tsession->s_mutex);
3275                         mutex_lock_nested(&session->s_mutex,
3276                                           SINGLE_DEPTH_NESTING);
3277                 }
3278                 new_cap = ceph_get_cap(mdsc, NULL);
3279         } else {
3280                 WARN_ON(1);
3281                 tsession = NULL;
3282                 target = -1;
3283         }
3284         goto retry;
3285 
3286 out_unlock:
3287         spin_unlock(&ci->i_ceph_lock);
3288         mutex_unlock(&session->s_mutex);
3289         if (tsession) {
3290                 mutex_unlock(&tsession->s_mutex);
3291                 ceph_put_mds_session(tsession);
3292         }
3293         if (new_cap)
3294                 ceph_put_cap(mdsc, new_cap);
3295 }
3296 
3297 /*
3298  * Handle cap IMPORT.
3299  *
3300  * caller holds s_mutex. acquires i_ceph_lock
3301  */
3302 static void handle_cap_import(struct ceph_mds_client *mdsc,
3303                               struct inode *inode, struct ceph_mds_caps *im,
3304                               struct ceph_mds_cap_peer *ph,
3305                               struct ceph_mds_session *session,
3306                               struct ceph_cap **target_cap, int *old_issued)
3307         __acquires(ci->i_ceph_lock)
3308 {
3309         struct ceph_inode_info *ci = ceph_inode(inode);
3310         struct ceph_cap *cap, *ocap, *new_cap = NULL;
3311         int mds = session->s_mds;
3312         int issued;
3313         unsigned caps = le32_to_cpu(im->caps);
3314         unsigned wanted = le32_to_cpu(im->wanted);
3315         unsigned seq = le32_to_cpu(im->seq);
3316         unsigned mseq = le32_to_cpu(im->migrate_seq);
3317         u64 realmino = le64_to_cpu(im->realm);
3318         u64 cap_id = le64_to_cpu(im->cap_id);
3319         u64 p_cap_id;
3320         int peer;
3321 
3322         if (ph) {
3323                 p_cap_id = le64_to_cpu(ph->cap_id);
3324                 peer = le32_to_cpu(ph->mds);
3325         } else {
3326                 p_cap_id = 0;
3327                 peer = -1;
3328         }
3329 
3330         dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3331              inode, ci, mds, mseq, peer);
3332 
3333 retry:
3334         spin_lock(&ci->i_ceph_lock);
3335         cap = __get_cap_for_mds(ci, mds);
3336         if (!cap) {
3337                 if (!new_cap) {
3338                         spin_unlock(&ci->i_ceph_lock);
3339                         new_cap = ceph_get_cap(mdsc, NULL);
3340                         goto retry;
3341                 }
3342                 cap = new_cap;
3343         } else {
3344                 if (new_cap) {
3345                         ceph_put_cap(mdsc, new_cap);
3346                         new_cap = NULL;
3347                 }
3348         }
3349 
3350         __ceph_caps_issued(ci, &issued);
3351         issued |= __ceph_caps_dirty(ci);
3352 
3353         ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3354                      realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3355 
3356         ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3357         if (ocap && ocap->cap_id == p_cap_id) {
3358                 dout(" remove export cap %p mds%d flags %d\n",
3359                      ocap, peer, ph->flags);
3360                 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3361                     (ocap->seq != le32_to_cpu(ph->seq) ||
3362                      ocap->mseq != le32_to_cpu(ph->mseq))) {
3363                         pr_err("handle_cap_import: mismatched seq/mseq: "
3364                                "ino (%llx.%llx) mds%d seq %d mseq %d "
3365                                "importer mds%d has peer seq %d mseq %d\n",
3366                                ceph_vinop(inode), peer, ocap->seq,
3367                                ocap->mseq, mds, le32_to_cpu(ph->seq),
3368                                le32_to_cpu(ph->mseq));
3369                 }
3370                 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3371         }
3372 
3373         /* make sure we re-request max_size, if necessary */
3374         ci->i_wanted_max_size = 0;
3375         ci->i_requested_max_size = 0;
3376 
3377         *old_issued = issued;
3378         *target_cap = cap;
3379 }
3380 
3381 /*
3382  * Handle a caps message from the MDS.
3383  *
3384  * Identify the appropriate session, inode, and call the right handler
3385  * based on the cap op.
3386  */
3387 void ceph_handle_caps(struct ceph_mds_session *session,
3388                       struct ceph_msg *msg)
3389 {
3390         struct ceph_mds_client *mdsc = session->s_mdsc;
3391         struct super_block *sb = mdsc->fsc->sb;
3392         struct inode *inode;
3393         struct ceph_inode_info *ci;
3394         struct ceph_cap *cap;
3395         struct ceph_mds_caps *h;
3396         struct ceph_mds_cap_peer *peer = NULL;
3397         struct ceph_snap_realm *realm;
3398         int mds = session->s_mds;
3399         int op, issued;
3400         u32 seq, mseq;
3401         struct ceph_vino vino;
3402         u64 cap_id;
3403         u64 size, max_size;
3404         u64 tid;
3405         u64 inline_version = 0;
3406         void *inline_data = NULL;
3407         u32  inline_len = 0;
3408         void *snaptrace;
3409         size_t snaptrace_len;
3410         void *p, *end;
3411 
3412         dout("handle_caps from mds%d\n", mds);
3413 
3414         /* decode */
3415         end = msg->front.iov_base + msg->front.iov_len;
3416         tid = le64_to_cpu(msg->hdr.tid);
3417         if (msg->front.iov_len < sizeof(*h))
3418                 goto bad;
3419         h = msg->front.iov_base;
3420         op = le32_to_cpu(h->op);
3421         vino.ino = le64_to_cpu(h->ino);
3422         vino.snap = CEPH_NOSNAP;
3423         cap_id = le64_to_cpu(h->cap_id);
3424         seq = le32_to_cpu(h->seq);
3425         mseq = le32_to_cpu(h->migrate_seq);
3426         size = le64_to_cpu(h->size);
3427         max_size = le64_to_cpu(h->max_size);
3428 
3429         snaptrace = h + 1;
3430         snaptrace_len = le32_to_cpu(h->snap_trace_len);
3431         p = snaptrace + snaptrace_len;
3432 
3433         if (le16_to_cpu(msg->hdr.version) >= 2) {
3434                 u32 flock_len;
3435                 ceph_decode_32_safe(&p, end, flock_len, bad);
3436                 if (p + flock_len > end)
3437                         goto bad;
3438                 p += flock_len;
3439         }
3440 
3441         if (le16_to_cpu(msg->hdr.version) >= 3) {
3442                 if (op == CEPH_CAP_OP_IMPORT) {
3443                         if (p + sizeof(*peer) > end)
3444                                 goto bad;
3445                         peer = p;
3446                         p += sizeof(*peer);
3447                 } else if (op == CEPH_CAP_OP_EXPORT) {
3448                         /* recorded in unused fields */
3449                         peer = (void *)&h->size;
3450                 }
3451         }
3452 
3453         if (le16_to_cpu(msg->hdr.version) >= 4) {
3454                 ceph_decode_64_safe(&p, end, inline_version, bad);
3455                 ceph_decode_32_safe(&p, end, inline_len, bad);
3456                 if (p + inline_len > end)
3457                         goto bad;
3458                 inline_data = p;
3459                 p += inline_len;
3460         }
3461 
3462         /* lookup ino */
3463         inode = ceph_find_inode(sb, vino);
3464         ci = ceph_inode(inode);
3465         dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3466              vino.snap, inode);
3467 
3468         mutex_lock(&session->s_mutex);
3469         session->s_seq++;
3470         dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3471              (unsigned)seq);
3472 
3473         if (!inode) {
3474                 dout(" i don't have ino %llx\n", vino.ino);
3475 
3476                 if (op == CEPH_CAP_OP_IMPORT) {
3477                         cap = ceph_get_cap(mdsc, NULL);
3478                         cap->cap_ino = vino.ino;
3479                         cap->queue_release = 1;
3480                         cap->cap_id = cap_id;
3481                         cap->mseq = mseq;
3482                         cap->seq = seq;
3483                         spin_lock(&session->s_cap_lock);
3484                         list_add_tail(&cap->session_caps,
3485                                         &session->s_cap_releases);
3486                         session->s_num_cap_releases++;
3487                         spin_unlock(&session->s_cap_lock);
3488                 }
3489                 goto flush_cap_releases;
3490         }
3491 
3492         /* these will work even if we don't have a cap yet */
3493         switch (op) {
3494         case CEPH_CAP_OP_FLUSHSNAP_ACK:
3495                 handle_cap_flushsnap_ack(inode, tid, h, session);
3496                 goto done;
3497 
3498         case CEPH_CAP_OP_EXPORT:
3499                 handle_cap_export(inode, h, peer, session);
3500                 goto done_unlocked;
3501 
3502         case CEPH_CAP_OP_IMPORT:
3503                 realm = NULL;
3504                 if (snaptrace_len) {
3505                         down_write(&mdsc->snap_rwsem);
3506                         ceph_update_snap_trace(mdsc, snaptrace,
3507                                                snaptrace + snaptrace_len,
3508                                                false, &realm);
3509                         downgrade_write(&mdsc->snap_rwsem);
3510                 } else {
3511                         down_read(&mdsc->snap_rwsem);
3512                 }
3513                 handle_cap_import(mdsc, inode, h, peer, session,
3514                                   &cap, &issued);
3515                 handle_cap_grant(mdsc, inode, h,
3516                                  inline_version, inline_data, inline_len,
3517                                  msg->middle, session, cap, issued);
3518                 if (realm)
3519                         ceph_put_snap_realm(mdsc, realm);
3520                 goto done_unlocked;
3521         }
3522 
3523         /* the rest require a cap */
3524         spin_lock(&ci->i_ceph_lock);
3525         cap = __get_cap_for_mds(ceph_inode(inode), mds);
3526         if (!cap) {
3527                 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3528                      inode, ceph_ino(inode), ceph_snap(inode), mds);
3529                 spin_unlock(&ci->i_ceph_lock);
3530                 goto flush_cap_releases;
3531         }
3532 
3533         /* note that each of these drops i_ceph_lock for us */
3534         switch (op) {
3535         case CEPH_CAP_OP_REVOKE:
3536         case CEPH_CAP_OP_GRANT:
3537                 __ceph_caps_issued(ci, &issued);
3538                 issued |= __ceph_caps_dirty(ci);
3539                 handle_cap_grant(mdsc, inode, h,
3540                                  inline_version, inline_data, inline_len,
3541                                  msg->middle, session, cap, issued);
3542                 goto done_unlocked;
3543 
3544         case CEPH_CAP_OP_FLUSH_ACK:
3545                 handle_cap_flush_ack(inode, tid, h, session, cap);
3546                 break;
3547 
3548         case CEPH_CAP_OP_TRUNC:
3549                 handle_cap_trunc(inode, h, session);
3550                 break;
3551 
3552         default:
3553                 spin_unlock(&ci->i_ceph_lock);
3554                 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3555                        ceph_cap_op_name(op));
3556         }
3557 
3558         goto done;
3559 
3560 flush_cap_releases:
3561         /*
3562          * send any cap release message to try to move things
3563          * along for the mds (who clearly thinks we still have this
3564          * cap).
3565          */
3566         ceph_send_cap_releases(mdsc, session);
3567 
3568 done:
3569         mutex_unlock(&session->s_mutex);
3570 done_unlocked:
3571         iput(inode);
3572         return;
3573 
3574 bad:
3575         pr_err("ceph_handle_caps: corrupt message\n");
3576         ceph_msg_dump(msg);
3577         return;
3578 }
3579 
3580 /*
3581  * Delayed work handler to process end of delayed cap release LRU list.
3582  */
3583 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3584 {
3585         struct ceph_inode_info *ci;
3586         int flags = CHECK_CAPS_NODELAY;
3587 
3588         dout("check_delayed_caps\n");
3589         while (1) {
3590                 spin_lock(&mdsc->cap_delay_lock);
3591                 if (list_empty(&mdsc->cap_delay_list))
3592                         break;
3593                 ci = list_first_entry(&mdsc->cap_delay_list,
3594                                       struct ceph_inode_info,
3595                                       i_cap_delay_list);
3596                 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3597                     time_before(jiffies, ci->i_hold_caps_max))
3598                         break;
3599                 list_del_init(&ci->i_cap_delay_list);
3600                 spin_unlock(&mdsc->cap_delay_lock);
3601                 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3602                 ceph_check_caps(ci, flags, NULL);
3603         }
3604         spin_unlock(&mdsc->cap_delay_lock);
3605 }
3606 
3607 /*
3608  * Flush all dirty caps to the mds
3609  */
3610 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3611 {
3612         struct ceph_inode_info *ci;
3613         struct inode *inode;
3614 
3615         dout("flush_dirty_caps\n");
3616         spin_lock(&mdsc->cap_dirty_lock);
3617         while (!list_empty(&mdsc->cap_dirty)) {
3618                 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3619                                       i_dirty_item);
3620                 inode = &ci->vfs_inode;
3621                 ihold(inode);
3622                 dout("flush_dirty_caps %p\n", inode);
3623                 spin_unlock(&mdsc->cap_dirty_lock);
3624                 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3625                 iput(inode);
3626                 spin_lock(&mdsc->cap_dirty_lock);
3627         }
3628         spin_unlock(&mdsc->cap_dirty_lock);
3629         dout("flush_dirty_caps done\n");
3630 }
3631 
3632 /*
3633  * Drop open file reference.  If we were the last open file,
3634  * we may need to release capabilities to the MDS (or schedule
3635  * their delayed release).
3636  */
3637 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3638 {
3639         struct inode *inode = &ci->vfs_inode;
3640         int last = 0;
3641 
3642         spin_lock(&ci->i_ceph_lock);
3643         dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
3644              ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
3645         BUG_ON(ci->i_nr_by_mode[fmode] == 0);
3646         if (--ci->i_nr_by_mode[fmode] == 0)
3647                 last++;
3648         spin_unlock(&ci->i_ceph_lock);
3649 
3650         if (last && ci->i_vino.snap == CEPH_NOSNAP)
3651                 ceph_check_caps(ci, 0, NULL);
3652 }
3653 
3654 /*
3655  * Helpers for embedding cap and dentry lease releases into mds
3656  * requests.
3657  *
3658  * @force is used by dentry_release (below) to force inclusion of a
3659  * record for the directory inode, even when there aren't any caps to
3660  * drop.
3661  */
3662 int ceph_encode_inode_release(void **p, struct inode *inode,
3663                               int mds, int drop, int unless, int force)
3664 {
3665         struct ceph_inode_info *ci = ceph_inode(inode);
3666         struct ceph_cap *cap;
3667         struct ceph_mds_request_release *rel = *p;
3668         int used, dirty;
3669         int ret = 0;
3670 
3671         spin_lock(&ci->i_ceph_lock);
3672         used = __ceph_caps_used(ci);
3673         dirty = __ceph_caps_dirty(ci);
3674 
3675         dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3676              inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3677              ceph_cap_string(unless));
3678 
3679         /* only drop unused, clean caps */
3680         drop &= ~(used | dirty);
3681 
3682         cap = __get_cap_for_mds(ci, mds);
3683         if (cap && __cap_is_valid(cap)) {
3684                 if (force ||
3685                     ((cap->issued & drop) &&
3686                      (cap->issued & unless) == 0)) {
3687                         if ((cap->issued & drop) &&
3688                             (cap->issued & unless) == 0) {
3689                                 int wanted = __ceph_caps_wanted(ci);
3690                                 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3691                                         wanted |= cap->mds_wanted;
3692                                 dout("encode_inode_release %p cap %p "
3693                                      "%s -> %s, wanted %s -> %s\n", inode, cap,
3694                                      ceph_cap_string(cap->issued),
3695                                      ceph_cap_string(cap->issued & ~drop),
3696                                      ceph_cap_string(cap->mds_wanted),
3697                                      ceph_cap_string(wanted));
3698 
3699                                 cap->issued &= ~drop;
3700                                 cap->implemented &= ~drop;
3701                                 cap->mds_wanted = wanted;
3702                         } else {
3703                                 dout("encode_inode_release %p cap %p %s"
3704                                      " (force)\n", inode, cap,
3705                                      ceph_cap_string(cap->issued));
3706                         }
3707 
3708                         rel->ino = cpu_to_le64(ceph_ino(inode));
3709                         rel->cap_id = cpu_to_le64(cap->cap_id);
3710                         rel->seq = cpu_to_le32(cap->seq);
3711                         rel->issue_seq = cpu_to_le32(cap->issue_seq);
3712                         rel->mseq = cpu_to_le32(cap->mseq);
3713                         rel->caps = cpu_to_le32(cap->implemented);
3714                         rel->wanted = cpu_to_le32(cap->mds_wanted);
3715                         rel->dname_len = 0;
3716                         rel->dname_seq = 0;
3717                         *p += sizeof(*rel);
3718                         ret = 1;
3719                 } else {
3720                         dout("encode_inode_release %p cap %p %s\n",
3721                              inode, cap, ceph_cap_string(cap->issued));
3722                 }
3723         }
3724         spin_unlock(&ci->i_ceph_lock);
3725         return ret;
3726 }
3727 
3728 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3729                                int mds, int drop, int unless)
3730 {
3731         struct inode *dir = d_inode(dentry->d_parent);
3732         struct ceph_mds_request_release *rel = *p;
3733         struct ceph_dentry_info *di = ceph_dentry(dentry);
3734         int force = 0;
3735         int ret;
3736 
3737         /*
3738          * force an record for the directory caps if we have a dentry lease.
3739          * this is racy (can't take i_ceph_lock and d_lock together), but it
3740          * doesn't have to be perfect; the mds will revoke anything we don't
3741          * release.
3742          */
3743         spin_lock(&dentry->d_lock);
3744         if (di->lease_session && di->lease_session->s_mds == mds)
3745                 force = 1;
3746         spin_unlock(&dentry->d_lock);
3747 
3748         ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3749 
3750         spin_lock(&dentry->d_lock);
3751         if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3752                 dout("encode_dentry_release %p mds%d seq %d\n",
3753                      dentry, mds, (int)di->lease_seq);
3754                 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3755                 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3756                 *p += dentry->d_name.len;
3757                 rel->dname_seq = cpu_to_le32(di->lease_seq);
3758                 __ceph_mdsc_drop_dentry_lease(dentry);
3759         }
3760         spin_unlock(&dentry->d_lock);
3761         return ret;
3762 }
3763 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp