ceph: use complete_all and wake_up_all
[linux-2.6.git] / fs / ceph / inode.c
1 #include "ceph_debug.h"
2
3 #include <linux/module.h>
4 #include <linux/fs.h>
5 #include <linux/smp_lock.h>
6 #include <linux/slab.h>
7 #include <linux/string.h>
8 #include <linux/uaccess.h>
9 #include <linux/kernel.h>
10 #include <linux/namei.h>
11 #include <linux/writeback.h>
12 #include <linux/vmalloc.h>
13 #include <linux/pagevec.h>
14
15 #include "super.h"
16 #include "decode.h"
17
18 /*
19  * Ceph inode operations
20  *
21  * Implement basic inode helpers (get, alloc) and inode ops (getattr,
22  * setattr, etc.), xattr helpers, and helpers for assimilating
23  * metadata returned by the MDS into our cache.
24  *
25  * Also define helpers for doing asynchronous writeback, invalidation,
26  * and truncation for the benefit of those who can't afford to block
27  * (typically because they are in the message handler path).
28  */
29
30 static const struct inode_operations ceph_symlink_iops;
31
32 static void ceph_invalidate_work(struct work_struct *work);
33 static void ceph_writeback_work(struct work_struct *work);
34 static void ceph_vmtruncate_work(struct work_struct *work);
35
36 /*
37  * find or create an inode, given the ceph ino number
38  */
39 struct inode *ceph_get_inode(struct super_block *sb, struct ceph_vino vino)
40 {
41         struct inode *inode;
42         ino_t t = ceph_vino_to_ino(vino);
43
44         inode = iget5_locked(sb, t, ceph_ino_compare, ceph_set_ino_cb, &vino);
45         if (inode == NULL)
46                 return ERR_PTR(-ENOMEM);
47         if (inode->i_state & I_NEW) {
48                 dout("get_inode created new inode %p %llx.%llx ino %llx\n",
49                      inode, ceph_vinop(inode), (u64)inode->i_ino);
50                 unlock_new_inode(inode);
51         }
52
53         dout("get_inode on %lu=%llx.%llx got %p\n", inode->i_ino, vino.ino,
54              vino.snap, inode);
55         return inode;
56 }
57
58 /*
59  * get/constuct snapdir inode for a given directory
60  */
61 struct inode *ceph_get_snapdir(struct inode *parent)
62 {
63         struct ceph_vino vino = {
64                 .ino = ceph_ino(parent),
65                 .snap = CEPH_SNAPDIR,
66         };
67         struct inode *inode = ceph_get_inode(parent->i_sb, vino);
68         struct ceph_inode_info *ci = ceph_inode(inode);
69
70         BUG_ON(!S_ISDIR(parent->i_mode));
71         if (IS_ERR(inode))
72                 return inode;
73         inode->i_mode = parent->i_mode;
74         inode->i_uid = parent->i_uid;
75         inode->i_gid = parent->i_gid;
76         inode->i_op = &ceph_dir_iops;
77         inode->i_fop = &ceph_dir_fops;
78         ci->i_snap_caps = CEPH_CAP_PIN; /* so we can open */
79         ci->i_rbytes = 0;
80         return inode;
81 }
82
83 const struct inode_operations ceph_file_iops = {
84         .permission = ceph_permission,
85         .setattr = ceph_setattr,
86         .getattr = ceph_getattr,
87         .setxattr = ceph_setxattr,
88         .getxattr = ceph_getxattr,
89         .listxattr = ceph_listxattr,
90         .removexattr = ceph_removexattr,
91 };
92
93
94 /*
95  * We use a 'frag tree' to keep track of the MDS's directory fragments
96  * for a given inode (usually there is just a single fragment).  We
97  * need to know when a child frag is delegated to a new MDS, or when
98  * it is flagged as replicated, so we can direct our requests
99  * accordingly.
100  */
101
102 /*
103  * find/create a frag in the tree
104  */
105 static struct ceph_inode_frag *__get_or_create_frag(struct ceph_inode_info *ci,
106                                                     u32 f)
107 {
108         struct rb_node **p;
109         struct rb_node *parent = NULL;
110         struct ceph_inode_frag *frag;
111         int c;
112
113         p = &ci->i_fragtree.rb_node;
114         while (*p) {
115                 parent = *p;
116                 frag = rb_entry(parent, struct ceph_inode_frag, node);
117                 c = ceph_frag_compare(f, frag->frag);
118                 if (c < 0)
119                         p = &(*p)->rb_left;
120                 else if (c > 0)
121                         p = &(*p)->rb_right;
122                 else
123                         return frag;
124         }
125
126         frag = kmalloc(sizeof(*frag), GFP_NOFS);
127         if (!frag) {
128                 pr_err("__get_or_create_frag ENOMEM on %p %llx.%llx "
129                        "frag %x\n", &ci->vfs_inode,
130                        ceph_vinop(&ci->vfs_inode), f);
131                 return ERR_PTR(-ENOMEM);
132         }
133         frag->frag = f;
134         frag->split_by = 0;
135         frag->mds = -1;
136         frag->ndist = 0;
137
138         rb_link_node(&frag->node, parent, p);
139         rb_insert_color(&frag->node, &ci->i_fragtree);
140
141         dout("get_or_create_frag added %llx.%llx frag %x\n",
142              ceph_vinop(&ci->vfs_inode), f);
143         return frag;
144 }
145
146 /*
147  * find a specific frag @f
148  */
149 struct ceph_inode_frag *__ceph_find_frag(struct ceph_inode_info *ci, u32 f)
150 {
151         struct rb_node *n = ci->i_fragtree.rb_node;
152
153         while (n) {
154                 struct ceph_inode_frag *frag =
155                         rb_entry(n, struct ceph_inode_frag, node);
156                 int c = ceph_frag_compare(f, frag->frag);
157                 if (c < 0)
158                         n = n->rb_left;
159                 else if (c > 0)
160                         n = n->rb_right;
161                 else
162                         return frag;
163         }
164         return NULL;
165 }
166
167 /*
168  * Choose frag containing the given value @v.  If @pfrag is
169  * specified, copy the frag delegation info to the caller if
170  * it is present.
171  */
172 u32 ceph_choose_frag(struct ceph_inode_info *ci, u32 v,
173                      struct ceph_inode_frag *pfrag,
174                      int *found)
175 {
176         u32 t = ceph_frag_make(0, 0);
177         struct ceph_inode_frag *frag;
178         unsigned nway, i;
179         u32 n;
180
181         if (found)
182                 *found = 0;
183
184         mutex_lock(&ci->i_fragtree_mutex);
185         while (1) {
186                 WARN_ON(!ceph_frag_contains_value(t, v));
187                 frag = __ceph_find_frag(ci, t);
188                 if (!frag)
189                         break; /* t is a leaf */
190                 if (frag->split_by == 0) {
191                         if (pfrag)
192                                 memcpy(pfrag, frag, sizeof(*pfrag));
193                         if (found)
194                                 *found = 1;
195                         break;
196                 }
197
198                 /* choose child */
199                 nway = 1 << frag->split_by;
200                 dout("choose_frag(%x) %x splits by %d (%d ways)\n", v, t,
201                      frag->split_by, nway);
202                 for (i = 0; i < nway; i++) {
203                         n = ceph_frag_make_child(t, frag->split_by, i);
204                         if (ceph_frag_contains_value(n, v)) {
205                                 t = n;
206                                 break;
207                         }
208                 }
209                 BUG_ON(i == nway);
210         }
211         dout("choose_frag(%x) = %x\n", v, t);
212
213         mutex_unlock(&ci->i_fragtree_mutex);
214         return t;
215 }
216
217 /*
218  * Process dirfrag (delegation) info from the mds.  Include leaf
219  * fragment in tree ONLY if ndist > 0.  Otherwise, only
220  * branches/splits are included in i_fragtree)
221  */
222 static int ceph_fill_dirfrag(struct inode *inode,
223                              struct ceph_mds_reply_dirfrag *dirinfo)
224 {
225         struct ceph_inode_info *ci = ceph_inode(inode);
226         struct ceph_inode_frag *frag;
227         u32 id = le32_to_cpu(dirinfo->frag);
228         int mds = le32_to_cpu(dirinfo->auth);
229         int ndist = le32_to_cpu(dirinfo->ndist);
230         int i;
231         int err = 0;
232
233         mutex_lock(&ci->i_fragtree_mutex);
234         if (ndist == 0) {
235                 /* no delegation info needed. */
236                 frag = __ceph_find_frag(ci, id);
237                 if (!frag)
238                         goto out;
239                 if (frag->split_by == 0) {
240                         /* tree leaf, remove */
241                         dout("fill_dirfrag removed %llx.%llx frag %x"
242                              " (no ref)\n", ceph_vinop(inode), id);
243                         rb_erase(&frag->node, &ci->i_fragtree);
244                         kfree(frag);
245                 } else {
246                         /* tree branch, keep and clear */
247                         dout("fill_dirfrag cleared %llx.%llx frag %x"
248                              " referral\n", ceph_vinop(inode), id);
249                         frag->mds = -1;
250                         frag->ndist = 0;
251                 }
252                 goto out;
253         }
254
255
256         /* find/add this frag to store mds delegation info */
257         frag = __get_or_create_frag(ci, id);
258         if (IS_ERR(frag)) {
259                 /* this is not the end of the world; we can continue
260                    with bad/inaccurate delegation info */
261                 pr_err("fill_dirfrag ENOMEM on mds ref %llx.%llx fg %x\n",
262                        ceph_vinop(inode), le32_to_cpu(dirinfo->frag));
263                 err = -ENOMEM;
264                 goto out;
265         }
266
267         frag->mds = mds;
268         frag->ndist = min_t(u32, ndist, CEPH_MAX_DIRFRAG_REP);
269         for (i = 0; i < frag->ndist; i++)
270                 frag->dist[i] = le32_to_cpu(dirinfo->dist[i]);
271         dout("fill_dirfrag %llx.%llx frag %x ndist=%d\n",
272              ceph_vinop(inode), frag->frag, frag->ndist);
273
274 out:
275         mutex_unlock(&ci->i_fragtree_mutex);
276         return err;
277 }
278
279
280 /*
281  * initialize a newly allocated inode.
282  */
283 struct inode *ceph_alloc_inode(struct super_block *sb)
284 {
285         struct ceph_inode_info *ci;
286         int i;
287
288         ci = kmem_cache_alloc(ceph_inode_cachep, GFP_NOFS);
289         if (!ci)
290                 return NULL;
291
292         dout("alloc_inode %p\n", &ci->vfs_inode);
293
294         ci->i_version = 0;
295         ci->i_time_warp_seq = 0;
296         ci->i_ceph_flags = 0;
297         ci->i_release_count = 0;
298         ci->i_symlink = NULL;
299
300         ci->i_fragtree = RB_ROOT;
301         mutex_init(&ci->i_fragtree_mutex);
302
303         ci->i_xattrs.blob = NULL;
304         ci->i_xattrs.prealloc_blob = NULL;
305         ci->i_xattrs.dirty = false;
306         ci->i_xattrs.index = RB_ROOT;
307         ci->i_xattrs.count = 0;
308         ci->i_xattrs.names_size = 0;
309         ci->i_xattrs.vals_size = 0;
310         ci->i_xattrs.version = 0;
311         ci->i_xattrs.index_version = 0;
312
313         ci->i_caps = RB_ROOT;
314         ci->i_auth_cap = NULL;
315         ci->i_dirty_caps = 0;
316         ci->i_flushing_caps = 0;
317         INIT_LIST_HEAD(&ci->i_dirty_item);
318         INIT_LIST_HEAD(&ci->i_flushing_item);
319         ci->i_cap_flush_seq = 0;
320         ci->i_cap_flush_last_tid = 0;
321         memset(&ci->i_cap_flush_tid, 0, sizeof(ci->i_cap_flush_tid));
322         init_waitqueue_head(&ci->i_cap_wq);
323         ci->i_hold_caps_min = 0;
324         ci->i_hold_caps_max = 0;
325         INIT_LIST_HEAD(&ci->i_cap_delay_list);
326         ci->i_cap_exporting_mds = 0;
327         ci->i_cap_exporting_mseq = 0;
328         ci->i_cap_exporting_issued = 0;
329         INIT_LIST_HEAD(&ci->i_cap_snaps);
330         ci->i_head_snapc = NULL;
331         ci->i_snap_caps = 0;
332
333         for (i = 0; i < CEPH_FILE_MODE_NUM; i++)
334                 ci->i_nr_by_mode[i] = 0;
335
336         ci->i_truncate_seq = 0;
337         ci->i_truncate_size = 0;
338         ci->i_truncate_pending = 0;
339
340         ci->i_max_size = 0;
341         ci->i_reported_size = 0;
342         ci->i_wanted_max_size = 0;
343         ci->i_requested_max_size = 0;
344
345         ci->i_pin_ref = 0;
346         ci->i_rd_ref = 0;
347         ci->i_rdcache_ref = 0;
348         ci->i_wr_ref = 0;
349         ci->i_wrbuffer_ref = 0;
350         ci->i_wrbuffer_ref_head = 0;
351         ci->i_shared_gen = 0;
352         ci->i_rdcache_gen = 0;
353         ci->i_rdcache_revoking = 0;
354
355         INIT_LIST_HEAD(&ci->i_unsafe_writes);
356         INIT_LIST_HEAD(&ci->i_unsafe_dirops);
357         spin_lock_init(&ci->i_unsafe_lock);
358
359         ci->i_snap_realm = NULL;
360         INIT_LIST_HEAD(&ci->i_snap_realm_item);
361         INIT_LIST_HEAD(&ci->i_snap_flush_item);
362
363         INIT_WORK(&ci->i_wb_work, ceph_writeback_work);
364         INIT_WORK(&ci->i_pg_inv_work, ceph_invalidate_work);
365
366         INIT_WORK(&ci->i_vmtruncate_work, ceph_vmtruncate_work);
367
368         return &ci->vfs_inode;
369 }
370
371 void ceph_destroy_inode(struct inode *inode)
372 {
373         struct ceph_inode_info *ci = ceph_inode(inode);
374         struct ceph_inode_frag *frag;
375         struct rb_node *n;
376
377         dout("destroy_inode %p ino %llx.%llx\n", inode, ceph_vinop(inode));
378
379         ceph_queue_caps_release(inode);
380
381         /*
382          * we may still have a snap_realm reference if there are stray
383          * caps in i_cap_exporting_issued or i_snap_caps.
384          */
385         if (ci->i_snap_realm) {
386                 struct ceph_mds_client *mdsc =
387                         &ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
388                 struct ceph_snap_realm *realm = ci->i_snap_realm;
389
390                 dout(" dropping residual ref to snap realm %p\n", realm);
391                 spin_lock(&realm->inodes_with_caps_lock);
392                 list_del_init(&ci->i_snap_realm_item);
393                 spin_unlock(&realm->inodes_with_caps_lock);
394                 ceph_put_snap_realm(mdsc, realm);
395         }
396
397         kfree(ci->i_symlink);
398         while ((n = rb_first(&ci->i_fragtree)) != NULL) {
399                 frag = rb_entry(n, struct ceph_inode_frag, node);
400                 rb_erase(n, &ci->i_fragtree);
401                 kfree(frag);
402         }
403
404         __ceph_destroy_xattrs(ci);
405         if (ci->i_xattrs.blob)
406                 ceph_buffer_put(ci->i_xattrs.blob);
407         if (ci->i_xattrs.prealloc_blob)
408                 ceph_buffer_put(ci->i_xattrs.prealloc_blob);
409
410         kmem_cache_free(ceph_inode_cachep, ci);
411 }
412
413
414 /*
415  * Helpers to fill in size, ctime, mtime, and atime.  We have to be
416  * careful because either the client or MDS may have more up to date
417  * info, depending on which capabilities are held, and whether
418  * time_warp_seq or truncate_seq have increased.  (Ordinarily, mtime
419  * and size are monotonically increasing, except when utimes() or
420  * truncate() increments the corresponding _seq values.)
421  */
422 int ceph_fill_file_size(struct inode *inode, int issued,
423                         u32 truncate_seq, u64 truncate_size, u64 size)
424 {
425         struct ceph_inode_info *ci = ceph_inode(inode);
426         int queue_trunc = 0;
427
428         if (ceph_seq_cmp(truncate_seq, ci->i_truncate_seq) > 0 ||
429             (truncate_seq == ci->i_truncate_seq && size > inode->i_size)) {
430                 dout("size %lld -> %llu\n", inode->i_size, size);
431                 inode->i_size = size;
432                 inode->i_blocks = (size + (1<<9) - 1) >> 9;
433                 ci->i_reported_size = size;
434                 if (truncate_seq != ci->i_truncate_seq) {
435                         dout("truncate_seq %u -> %u\n",
436                              ci->i_truncate_seq, truncate_seq);
437                         ci->i_truncate_seq = truncate_seq;
438                         /*
439                          * If we hold relevant caps, or in the case where we're
440                          * not the only client referencing this file and we
441                          * don't hold those caps, then we need to check whether
442                          * the file is either opened or mmaped
443                          */
444                         if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_RD|
445                                       CEPH_CAP_FILE_WR|CEPH_CAP_FILE_BUFFER|
446                                       CEPH_CAP_FILE_EXCL)) ||
447                             mapping_mapped(inode->i_mapping) ||
448                             __ceph_caps_file_wanted(ci)) {
449                                 ci->i_truncate_pending++;
450                                 queue_trunc = 1;
451                         }
452                 }
453         }
454         if (ceph_seq_cmp(truncate_seq, ci->i_truncate_seq) >= 0 &&
455             ci->i_truncate_size != truncate_size) {
456                 dout("truncate_size %lld -> %llu\n", ci->i_truncate_size,
457                      truncate_size);
458                 ci->i_truncate_size = truncate_size;
459         }
460         return queue_trunc;
461 }
462
463 void ceph_fill_file_time(struct inode *inode, int issued,
464                          u64 time_warp_seq, struct timespec *ctime,
465                          struct timespec *mtime, struct timespec *atime)
466 {
467         struct ceph_inode_info *ci = ceph_inode(inode);
468         int warn = 0;
469
470         if (issued & (CEPH_CAP_FILE_EXCL|
471                       CEPH_CAP_FILE_WR|
472                       CEPH_CAP_FILE_BUFFER)) {
473                 if (timespec_compare(ctime, &inode->i_ctime) > 0) {
474                         dout("ctime %ld.%09ld -> %ld.%09ld inc w/ cap\n",
475                              inode->i_ctime.tv_sec, inode->i_ctime.tv_nsec,
476                              ctime->tv_sec, ctime->tv_nsec);
477                         inode->i_ctime = *ctime;
478                 }
479                 if (ceph_seq_cmp(time_warp_seq, ci->i_time_warp_seq) > 0) {
480                         /* the MDS did a utimes() */
481                         dout("mtime %ld.%09ld -> %ld.%09ld "
482                              "tw %d -> %d\n",
483                              inode->i_mtime.tv_sec, inode->i_mtime.tv_nsec,
484                              mtime->tv_sec, mtime->tv_nsec,
485                              ci->i_time_warp_seq, (int)time_warp_seq);
486
487                         inode->i_mtime = *mtime;
488                         inode->i_atime = *atime;
489                         ci->i_time_warp_seq = time_warp_seq;
490                 } else if (time_warp_seq == ci->i_time_warp_seq) {
491                         /* nobody did utimes(); take the max */
492                         if (timespec_compare(mtime, &inode->i_mtime) > 0) {
493                                 dout("mtime %ld.%09ld -> %ld.%09ld inc\n",
494                                      inode->i_mtime.tv_sec,
495                                      inode->i_mtime.tv_nsec,
496                                      mtime->tv_sec, mtime->tv_nsec);
497                                 inode->i_mtime = *mtime;
498                         }
499                         if (timespec_compare(atime, &inode->i_atime) > 0) {
500                                 dout("atime %ld.%09ld -> %ld.%09ld inc\n",
501                                      inode->i_atime.tv_sec,
502                                      inode->i_atime.tv_nsec,
503                                      atime->tv_sec, atime->tv_nsec);
504                                 inode->i_atime = *atime;
505                         }
506                 } else if (issued & CEPH_CAP_FILE_EXCL) {
507                         /* we did a utimes(); ignore mds values */
508                 } else {
509                         warn = 1;
510                 }
511         } else {
512                 /* we have no write caps; whatever the MDS says is true */
513                 if (ceph_seq_cmp(time_warp_seq, ci->i_time_warp_seq) >= 0) {
514                         inode->i_ctime = *ctime;
515                         inode->i_mtime = *mtime;
516                         inode->i_atime = *atime;
517                         ci->i_time_warp_seq = time_warp_seq;
518                 } else {
519                         warn = 1;
520                 }
521         }
522         if (warn) /* time_warp_seq shouldn't go backwards */
523                 dout("%p mds time_warp_seq %llu < %u\n",
524                      inode, time_warp_seq, ci->i_time_warp_seq);
525 }
526
527 /*
528  * Populate an inode based on info from mds.  May be called on new or
529  * existing inodes.
530  */
531 static int fill_inode(struct inode *inode,
532                       struct ceph_mds_reply_info_in *iinfo,
533                       struct ceph_mds_reply_dirfrag *dirinfo,
534                       struct ceph_mds_session *session,
535                       unsigned long ttl_from, int cap_fmode,
536                       struct ceph_cap_reservation *caps_reservation)
537 {
538         struct ceph_mds_reply_inode *info = iinfo->in;
539         struct ceph_inode_info *ci = ceph_inode(inode);
540         int i;
541         int issued, implemented;
542         struct timespec mtime, atime, ctime;
543         u32 nsplits;
544         struct ceph_buffer *xattr_blob = NULL;
545         int err = 0;
546         int queue_trunc = 0;
547
548         dout("fill_inode %p ino %llx.%llx v %llu had %llu\n",
549              inode, ceph_vinop(inode), le64_to_cpu(info->version),
550              ci->i_version);
551
552         /*
553          * prealloc xattr data, if it looks like we'll need it.  only
554          * if len > 4 (meaning there are actually xattrs; the first 4
555          * bytes are the xattr count).
556          */
557         if (iinfo->xattr_len > 4) {
558                 xattr_blob = ceph_buffer_new(iinfo->xattr_len, GFP_NOFS);
559                 if (!xattr_blob)
560                         pr_err("fill_inode ENOMEM xattr blob %d bytes\n",
561                                iinfo->xattr_len);
562         }
563
564         spin_lock(&inode->i_lock);
565
566         /*
567          * provided version will be odd if inode value is projected,
568          * even if stable.  skip the update if we have a newer info
569          * (e.g., due to inode info racing form multiple MDSs), or if
570          * we are getting projected (unstable) inode info.
571          */
572         if (le64_to_cpu(info->version) > 0 &&
573             (ci->i_version & ~1) > le64_to_cpu(info->version))
574                 goto no_change;
575
576         issued = __ceph_caps_issued(ci, &implemented);
577         issued |= implemented | __ceph_caps_dirty(ci);
578
579         /* update inode */
580         ci->i_version = le64_to_cpu(info->version);
581         inode->i_version++;
582         inode->i_rdev = le32_to_cpu(info->rdev);
583
584         if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
585                 inode->i_mode = le32_to_cpu(info->mode);
586                 inode->i_uid = le32_to_cpu(info->uid);
587                 inode->i_gid = le32_to_cpu(info->gid);
588                 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
589                      inode->i_uid, inode->i_gid);
590         }
591
592         if ((issued & CEPH_CAP_LINK_EXCL) == 0)
593                 inode->i_nlink = le32_to_cpu(info->nlink);
594
595         /* be careful with mtime, atime, size */
596         ceph_decode_timespec(&atime, &info->atime);
597         ceph_decode_timespec(&mtime, &info->mtime);
598         ceph_decode_timespec(&ctime, &info->ctime);
599         queue_trunc = ceph_fill_file_size(inode, issued,
600                                           le32_to_cpu(info->truncate_seq),
601                                           le64_to_cpu(info->truncate_size),
602                                           le64_to_cpu(info->size));
603         ceph_fill_file_time(inode, issued,
604                             le32_to_cpu(info->time_warp_seq),
605                             &ctime, &mtime, &atime);
606
607         ci->i_max_size = le64_to_cpu(info->max_size);
608         ci->i_layout = info->layout;
609         inode->i_blkbits = fls(le32_to_cpu(info->layout.fl_stripe_unit)) - 1;
610
611         /* xattrs */
612         /* note that if i_xattrs.len <= 4, i_xattrs.data will still be NULL. */
613         if ((issued & CEPH_CAP_XATTR_EXCL) == 0 &&
614             le64_to_cpu(info->xattr_version) > ci->i_xattrs.version) {
615                 if (ci->i_xattrs.blob)
616                         ceph_buffer_put(ci->i_xattrs.blob);
617                 ci->i_xattrs.blob = xattr_blob;
618                 if (xattr_blob)
619                         memcpy(ci->i_xattrs.blob->vec.iov_base,
620                                iinfo->xattr_data, iinfo->xattr_len);
621                 ci->i_xattrs.version = le64_to_cpu(info->xattr_version);
622                 xattr_blob = NULL;
623         }
624
625         inode->i_mapping->a_ops = &ceph_aops;
626         inode->i_mapping->backing_dev_info =
627                 &ceph_sb_to_client(inode->i_sb)->backing_dev_info;
628
629         switch (inode->i_mode & S_IFMT) {
630         case S_IFIFO:
631         case S_IFBLK:
632         case S_IFCHR:
633         case S_IFSOCK:
634                 init_special_inode(inode, inode->i_mode, inode->i_rdev);
635                 inode->i_op = &ceph_file_iops;
636                 break;
637         case S_IFREG:
638                 inode->i_op = &ceph_file_iops;
639                 inode->i_fop = &ceph_file_fops;
640                 break;
641         case S_IFLNK:
642                 inode->i_op = &ceph_symlink_iops;
643                 if (!ci->i_symlink) {
644                         int symlen = iinfo->symlink_len;
645                         char *sym;
646
647                         BUG_ON(symlen != inode->i_size);
648                         spin_unlock(&inode->i_lock);
649
650                         err = -ENOMEM;
651                         sym = kmalloc(symlen+1, GFP_NOFS);
652                         if (!sym)
653                                 goto out;
654                         memcpy(sym, iinfo->symlink, symlen);
655                         sym[symlen] = 0;
656
657                         spin_lock(&inode->i_lock);
658                         if (!ci->i_symlink)
659                                 ci->i_symlink = sym;
660                         else
661                                 kfree(sym); /* lost a race */
662                 }
663                 break;
664         case S_IFDIR:
665                 inode->i_op = &ceph_dir_iops;
666                 inode->i_fop = &ceph_dir_fops;
667
668                 ci->i_files = le64_to_cpu(info->files);
669                 ci->i_subdirs = le64_to_cpu(info->subdirs);
670                 ci->i_rbytes = le64_to_cpu(info->rbytes);
671                 ci->i_rfiles = le64_to_cpu(info->rfiles);
672                 ci->i_rsubdirs = le64_to_cpu(info->rsubdirs);
673                 ceph_decode_timespec(&ci->i_rctime, &info->rctime);
674
675                 /* set dir completion flag? */
676                 if (ci->i_files == 0 && ci->i_subdirs == 0 &&
677                     ceph_snap(inode) == CEPH_NOSNAP &&
678                     (le32_to_cpu(info->cap.caps) & CEPH_CAP_FILE_SHARED) &&
679                     (ci->i_ceph_flags & CEPH_I_COMPLETE) == 0) {
680                         dout(" marking %p complete (empty)\n", inode);
681                         ci->i_ceph_flags |= CEPH_I_COMPLETE;
682                         ci->i_max_offset = 2;
683                 }
684
685                 /* it may be better to set st_size in getattr instead? */
686                 if (ceph_test_opt(ceph_sb_to_client(inode->i_sb), RBYTES))
687                         inode->i_size = ci->i_rbytes;
688                 break;
689         default:
690                 pr_err("fill_inode %llx.%llx BAD mode 0%o\n",
691                        ceph_vinop(inode), inode->i_mode);
692         }
693
694 no_change:
695         spin_unlock(&inode->i_lock);
696
697         /* queue truncate if we saw i_size decrease */
698         if (queue_trunc)
699                 ceph_queue_vmtruncate(inode);
700
701         /* populate frag tree */
702         /* FIXME: move me up, if/when version reflects fragtree changes */
703         nsplits = le32_to_cpu(info->fragtree.nsplits);
704         mutex_lock(&ci->i_fragtree_mutex);
705         for (i = 0; i < nsplits; i++) {
706                 u32 id = le32_to_cpu(info->fragtree.splits[i].frag);
707                 struct ceph_inode_frag *frag = __get_or_create_frag(ci, id);
708
709                 if (IS_ERR(frag))
710                         continue;
711                 frag->split_by = le32_to_cpu(info->fragtree.splits[i].by);
712                 dout(" frag %x split by %d\n", frag->frag, frag->split_by);
713         }
714         mutex_unlock(&ci->i_fragtree_mutex);
715
716         /* were we issued a capability? */
717         if (info->cap.caps) {
718                 if (ceph_snap(inode) == CEPH_NOSNAP) {
719                         ceph_add_cap(inode, session,
720                                      le64_to_cpu(info->cap.cap_id),
721                                      cap_fmode,
722                                      le32_to_cpu(info->cap.caps),
723                                      le32_to_cpu(info->cap.wanted),
724                                      le32_to_cpu(info->cap.seq),
725                                      le32_to_cpu(info->cap.mseq),
726                                      le64_to_cpu(info->cap.realm),
727                                      info->cap.flags,
728                                      caps_reservation);
729                 } else {
730                         spin_lock(&inode->i_lock);
731                         dout(" %p got snap_caps %s\n", inode,
732                              ceph_cap_string(le32_to_cpu(info->cap.caps)));
733                         ci->i_snap_caps |= le32_to_cpu(info->cap.caps);
734                         if (cap_fmode >= 0)
735                                 __ceph_get_fmode(ci, cap_fmode);
736                         spin_unlock(&inode->i_lock);
737                 }
738         } else if (cap_fmode >= 0) {
739                 pr_warning("mds issued no caps on %llx.%llx\n",
740                            ceph_vinop(inode));
741                 __ceph_get_fmode(ci, cap_fmode);
742         }
743
744         /* update delegation info? */
745         if (dirinfo)
746                 ceph_fill_dirfrag(inode, dirinfo);
747
748         err = 0;
749
750 out:
751         if (xattr_blob)
752                 ceph_buffer_put(xattr_blob);
753         return err;
754 }
755
756 /*
757  * caller should hold session s_mutex.
758  */
759 static void update_dentry_lease(struct dentry *dentry,
760                                 struct ceph_mds_reply_lease *lease,
761                                 struct ceph_mds_session *session,
762                                 unsigned long from_time)
763 {
764         struct ceph_dentry_info *di = ceph_dentry(dentry);
765         long unsigned duration = le32_to_cpu(lease->duration_ms);
766         long unsigned ttl = from_time + (duration * HZ) / 1000;
767         long unsigned half_ttl = from_time + (duration * HZ / 2) / 1000;
768         struct inode *dir;
769
770         /* only track leases on regular dentries */
771         if (dentry->d_op != &ceph_dentry_ops)
772                 return;
773
774         spin_lock(&dentry->d_lock);
775         dout("update_dentry_lease %p mask %d duration %lu ms ttl %lu\n",
776              dentry, le16_to_cpu(lease->mask), duration, ttl);
777
778         /* make lease_rdcache_gen match directory */
779         dir = dentry->d_parent->d_inode;
780         di->lease_shared_gen = ceph_inode(dir)->i_shared_gen;
781
782         if (lease->mask == 0)
783                 goto out_unlock;
784
785         if (di->lease_gen == session->s_cap_gen &&
786             time_before(ttl, dentry->d_time))
787                 goto out_unlock;  /* we already have a newer lease. */
788
789         if (di->lease_session && di->lease_session != session)
790                 goto out_unlock;
791
792         ceph_dentry_lru_touch(dentry);
793
794         if (!di->lease_session)
795                 di->lease_session = ceph_get_mds_session(session);
796         di->lease_gen = session->s_cap_gen;
797         di->lease_seq = le32_to_cpu(lease->seq);
798         di->lease_renew_after = half_ttl;
799         di->lease_renew_from = 0;
800         dentry->d_time = ttl;
801 out_unlock:
802         spin_unlock(&dentry->d_lock);
803         return;
804 }
805
806 /*
807  * Set dentry's directory position based on the current dir's max, and
808  * order it in d_subdirs, so that dcache_readdir behaves.
809  */
810 static void ceph_set_dentry_offset(struct dentry *dn)
811 {
812         struct dentry *dir = dn->d_parent;
813         struct inode *inode = dn->d_parent->d_inode;
814         struct ceph_dentry_info *di;
815
816         BUG_ON(!inode);
817
818         di = ceph_dentry(dn);
819
820         spin_lock(&inode->i_lock);
821         if ((ceph_inode(inode)->i_ceph_flags & CEPH_I_COMPLETE) == 0) {
822                 spin_unlock(&inode->i_lock);
823                 return;
824         }
825         di->offset = ceph_inode(inode)->i_max_offset++;
826         spin_unlock(&inode->i_lock);
827
828         spin_lock(&dcache_lock);
829         spin_lock(&dn->d_lock);
830         list_move(&dn->d_u.d_child, &dir->d_subdirs);
831         dout("set_dentry_offset %p %lld (%p %p)\n", dn, di->offset,
832              dn->d_u.d_child.prev, dn->d_u.d_child.next);
833         spin_unlock(&dn->d_lock);
834         spin_unlock(&dcache_lock);
835 }
836
837 /*
838  * splice a dentry to an inode.
839  * caller must hold directory i_mutex for this to be safe.
840  *
841  * we will only rehash the resulting dentry if @prehash is
842  * true; @prehash will be set to false (for the benefit of
843  * the caller) if we fail.
844  */
845 static struct dentry *splice_dentry(struct dentry *dn, struct inode *in,
846                                     bool *prehash)
847 {
848         struct dentry *realdn;
849
850         BUG_ON(dn->d_inode);
851
852         /* dn must be unhashed */
853         if (!d_unhashed(dn))
854                 d_drop(dn);
855         realdn = d_materialise_unique(dn, in);
856         if (IS_ERR(realdn)) {
857                 pr_err("splice_dentry error %ld %p inode %p ino %llx.%llx\n",
858                        PTR_ERR(realdn), dn, in, ceph_vinop(in));
859                 if (prehash)
860                         *prehash = false; /* don't rehash on error */
861                 dn = realdn; /* note realdn contains the error */
862                 goto out;
863         } else if (realdn) {
864                 dout("dn %p (%d) spliced with %p (%d) "
865                      "inode %p ino %llx.%llx\n",
866                      dn, atomic_read(&dn->d_count),
867                      realdn, atomic_read(&realdn->d_count),
868                      realdn->d_inode, ceph_vinop(realdn->d_inode));
869                 dput(dn);
870                 dn = realdn;
871         } else {
872                 BUG_ON(!ceph_dentry(dn));
873                 dout("dn %p attached to %p ino %llx.%llx\n",
874                      dn, dn->d_inode, ceph_vinop(dn->d_inode));
875         }
876         if ((!prehash || *prehash) && d_unhashed(dn))
877                 d_rehash(dn);
878         ceph_set_dentry_offset(dn);
879 out:
880         return dn;
881 }
882
883 /*
884  * Incorporate results into the local cache.  This is either just
885  * one inode, or a directory, dentry, and possibly linked-to inode (e.g.,
886  * after a lookup).
887  *
888  * A reply may contain
889  *         a directory inode along with a dentry.
890  *  and/or a target inode
891  *
892  * Called with snap_rwsem (read).
893  */
894 int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req,
895                     struct ceph_mds_session *session)
896 {
897         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
898         struct inode *in = NULL;
899         struct ceph_mds_reply_inode *ininfo;
900         struct ceph_vino vino;
901         struct ceph_client *client = ceph_sb_to_client(sb);
902         int i = 0;
903         int err = 0;
904
905         dout("fill_trace %p is_dentry %d is_target %d\n", req,
906              rinfo->head->is_dentry, rinfo->head->is_target);
907
908 #if 0
909         /*
910          * Debugging hook:
911          *
912          * If we resend completed ops to a recovering mds, we get no
913          * trace.  Since that is very rare, pretend this is the case
914          * to ensure the 'no trace' handlers in the callers behave.
915          *
916          * Fill in inodes unconditionally to avoid breaking cap
917          * invariants.
918          */
919         if (rinfo->head->op & CEPH_MDS_OP_WRITE) {
920                 pr_info("fill_trace faking empty trace on %lld %s\n",
921                         req->r_tid, ceph_mds_op_name(rinfo->head->op));
922                 if (rinfo->head->is_dentry) {
923                         rinfo->head->is_dentry = 0;
924                         err = fill_inode(req->r_locked_dir,
925                                          &rinfo->diri, rinfo->dirfrag,
926                                          session, req->r_request_started, -1);
927                 }
928                 if (rinfo->head->is_target) {
929                         rinfo->head->is_target = 0;
930                         ininfo = rinfo->targeti.in;
931                         vino.ino = le64_to_cpu(ininfo->ino);
932                         vino.snap = le64_to_cpu(ininfo->snapid);
933                         in = ceph_get_inode(sb, vino);
934                         err = fill_inode(in, &rinfo->targeti, NULL,
935                                          session, req->r_request_started,
936                                          req->r_fmode);
937                         iput(in);
938                 }
939         }
940 #endif
941
942         if (!rinfo->head->is_target && !rinfo->head->is_dentry) {
943                 dout("fill_trace reply is empty!\n");
944                 if (rinfo->head->result == 0 && req->r_locked_dir)
945                         ceph_invalidate_dir_request(req);
946                 return 0;
947         }
948
949         if (rinfo->head->is_dentry) {
950                 struct inode *dir = req->r_locked_dir;
951
952                 err = fill_inode(dir, &rinfo->diri, rinfo->dirfrag,
953                                  session, req->r_request_started, -1,
954                                  &req->r_caps_reservation);
955                 if (err < 0)
956                         return err;
957         }
958
959         /*
960          * ignore null lease/binding on snapdir ENOENT, or else we
961          * will have trouble splicing in the virtual snapdir later
962          */
963         if (rinfo->head->is_dentry && !req->r_aborted &&
964             (rinfo->head->is_target || strncmp(req->r_dentry->d_name.name,
965                                                client->mount_args->snapdir_name,
966                                                req->r_dentry->d_name.len))) {
967                 /*
968                  * lookup link rename   : null -> possibly existing inode
969                  * mknod symlink mkdir  : null -> new inode
970                  * unlink               : linked -> null
971                  */
972                 struct inode *dir = req->r_locked_dir;
973                 struct dentry *dn = req->r_dentry;
974                 bool have_dir_cap, have_lease;
975
976                 BUG_ON(!dn);
977                 BUG_ON(!dir);
978                 BUG_ON(dn->d_parent->d_inode != dir);
979                 BUG_ON(ceph_ino(dir) !=
980                        le64_to_cpu(rinfo->diri.in->ino));
981                 BUG_ON(ceph_snap(dir) !=
982                        le64_to_cpu(rinfo->diri.in->snapid));
983
984                 /* do we have a lease on the whole dir? */
985                 have_dir_cap =
986                         (le32_to_cpu(rinfo->diri.in->cap.caps) &
987                          CEPH_CAP_FILE_SHARED);
988
989                 /* do we have a dn lease? */
990                 have_lease = have_dir_cap ||
991                         (le16_to_cpu(rinfo->dlease->mask) &
992                          CEPH_LOCK_DN);
993
994                 if (!have_lease)
995                         dout("fill_trace  no dentry lease or dir cap\n");
996
997                 /* rename? */
998                 if (req->r_old_dentry && req->r_op == CEPH_MDS_OP_RENAME) {
999                         dout(" src %p '%.*s' dst %p '%.*s'\n",
1000                              req->r_old_dentry,
1001                              req->r_old_dentry->d_name.len,
1002                              req->r_old_dentry->d_name.name,
1003                              dn, dn->d_name.len, dn->d_name.name);
1004                         dout("fill_trace doing d_move %p -> %p\n",
1005                              req->r_old_dentry, dn);
1006
1007                         /* d_move screws up d_subdirs order */
1008                         ceph_i_clear(dir, CEPH_I_COMPLETE);
1009
1010                         d_move(req->r_old_dentry, dn);
1011                         dout(" src %p '%.*s' dst %p '%.*s'\n",
1012                              req->r_old_dentry,
1013                              req->r_old_dentry->d_name.len,
1014                              req->r_old_dentry->d_name.name,
1015                              dn, dn->d_name.len, dn->d_name.name);
1016
1017                         /* ensure target dentry is invalidated, despite
1018                            rehashing bug in vfs_rename_dir */
1019                         ceph_invalidate_dentry_lease(dn);
1020
1021                         /* take overwritten dentry's readdir offset */
1022                         dout("dn %p gets %p offset %lld (old offset %lld)\n",
1023                              req->r_old_dentry, dn, ceph_dentry(dn)->offset,
1024                              ceph_dentry(req->r_old_dentry)->offset);
1025                         ceph_dentry(req->r_old_dentry)->offset =
1026                                 ceph_dentry(dn)->offset;
1027
1028                         dn = req->r_old_dentry;  /* use old_dentry */
1029                         in = dn->d_inode;
1030                 }
1031
1032                 /* null dentry? */
1033                 if (!rinfo->head->is_target) {
1034                         dout("fill_trace null dentry\n");
1035                         if (dn->d_inode) {
1036                                 dout("d_delete %p\n", dn);
1037                                 d_delete(dn);
1038                         } else {
1039                                 dout("d_instantiate %p NULL\n", dn);
1040                                 d_instantiate(dn, NULL);
1041                                 if (have_lease && d_unhashed(dn))
1042                                         d_rehash(dn);
1043                                 update_dentry_lease(dn, rinfo->dlease,
1044                                                     session,
1045                                                     req->r_request_started);
1046                         }
1047                         goto done;
1048                 }
1049
1050                 /* attach proper inode */
1051                 ininfo = rinfo->targeti.in;
1052                 vino.ino = le64_to_cpu(ininfo->ino);
1053                 vino.snap = le64_to_cpu(ininfo->snapid);
1054                 if (!dn->d_inode) {
1055                         in = ceph_get_inode(sb, vino);
1056                         if (IS_ERR(in)) {
1057                                 pr_err("fill_trace bad get_inode "
1058                                        "%llx.%llx\n", vino.ino, vino.snap);
1059                                 err = PTR_ERR(in);
1060                                 d_delete(dn);
1061                                 goto done;
1062                         }
1063                         dn = splice_dentry(dn, in, &have_lease);
1064                         if (IS_ERR(dn)) {
1065                                 err = PTR_ERR(dn);
1066                                 goto done;
1067                         }
1068                         req->r_dentry = dn;  /* may have spliced */
1069                         igrab(in);
1070                 } else if (ceph_ino(in) == vino.ino &&
1071                            ceph_snap(in) == vino.snap) {
1072                         igrab(in);
1073                 } else {
1074                         dout(" %p links to %p %llx.%llx, not %llx.%llx\n",
1075                              dn, in, ceph_ino(in), ceph_snap(in),
1076                              vino.ino, vino.snap);
1077                         have_lease = false;
1078                         in = NULL;
1079                 }
1080
1081                 if (have_lease)
1082                         update_dentry_lease(dn, rinfo->dlease, session,
1083                                             req->r_request_started);
1084                 dout(" final dn %p\n", dn);
1085                 i++;
1086         } else if (req->r_op == CEPH_MDS_OP_LOOKUPSNAP ||
1087                    req->r_op == CEPH_MDS_OP_MKSNAP) {
1088                 struct dentry *dn = req->r_dentry;
1089
1090                 /* fill out a snapdir LOOKUPSNAP dentry */
1091                 BUG_ON(!dn);
1092                 BUG_ON(!req->r_locked_dir);
1093                 BUG_ON(ceph_snap(req->r_locked_dir) != CEPH_SNAPDIR);
1094                 ininfo = rinfo->targeti.in;
1095                 vino.ino = le64_to_cpu(ininfo->ino);
1096                 vino.snap = le64_to_cpu(ininfo->snapid);
1097                 in = ceph_get_inode(sb, vino);
1098                 if (IS_ERR(in)) {
1099                         pr_err("fill_inode get_inode badness %llx.%llx\n",
1100                                vino.ino, vino.snap);
1101                         err = PTR_ERR(in);
1102                         d_delete(dn);
1103                         goto done;
1104                 }
1105                 dout(" linking snapped dir %p to dn %p\n", in, dn);
1106                 dn = splice_dentry(dn, in, NULL);
1107                 if (IS_ERR(dn)) {
1108                         err = PTR_ERR(dn);
1109                         goto done;
1110                 }
1111                 req->r_dentry = dn;  /* may have spliced */
1112                 igrab(in);
1113                 rinfo->head->is_dentry = 1;  /* fool notrace handlers */
1114         }
1115
1116         if (rinfo->head->is_target) {
1117                 vino.ino = le64_to_cpu(rinfo->targeti.in->ino);
1118                 vino.snap = le64_to_cpu(rinfo->targeti.in->snapid);
1119
1120                 if (in == NULL || ceph_ino(in) != vino.ino ||
1121                     ceph_snap(in) != vino.snap) {
1122                         in = ceph_get_inode(sb, vino);
1123                         if (IS_ERR(in)) {
1124                                 err = PTR_ERR(in);
1125                                 goto done;
1126                         }
1127                 }
1128                 req->r_target_inode = in;
1129
1130                 err = fill_inode(in,
1131                                  &rinfo->targeti, NULL,
1132                                  session, req->r_request_started,
1133                                  (le32_to_cpu(rinfo->head->result) == 0) ?
1134                                  req->r_fmode : -1,
1135                                  &req->r_caps_reservation);
1136                 if (err < 0) {
1137                         pr_err("fill_inode badness %p %llx.%llx\n",
1138                                in, ceph_vinop(in));
1139                         goto done;
1140                 }
1141         }
1142
1143 done:
1144         dout("fill_trace done err=%d\n", err);
1145         return err;
1146 }
1147
1148 /*
1149  * Prepopulate our cache with readdir results, leases, etc.
1150  */
1151 int ceph_readdir_prepopulate(struct ceph_mds_request *req,
1152                              struct ceph_mds_session *session)
1153 {
1154         struct dentry *parent = req->r_dentry;
1155         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1156         struct qstr dname;
1157         struct dentry *dn;
1158         struct inode *in;
1159         int err = 0, i;
1160         struct inode *snapdir = NULL;
1161         struct ceph_mds_request_head *rhead = req->r_request->front.iov_base;
1162         u64 frag = le32_to_cpu(rhead->args.readdir.frag);
1163         struct ceph_dentry_info *di;
1164
1165         if (le32_to_cpu(rinfo->head->op) == CEPH_MDS_OP_LSSNAP) {
1166                 snapdir = ceph_get_snapdir(parent->d_inode);
1167                 parent = d_find_alias(snapdir);
1168                 dout("readdir_prepopulate %d items under SNAPDIR dn %p\n",
1169                      rinfo->dir_nr, parent);
1170         } else {
1171                 dout("readdir_prepopulate %d items under dn %p\n",
1172                      rinfo->dir_nr, parent);
1173                 if (rinfo->dir_dir)
1174                         ceph_fill_dirfrag(parent->d_inode, rinfo->dir_dir);
1175         }
1176
1177         for (i = 0; i < rinfo->dir_nr; i++) {
1178                 struct ceph_vino vino;
1179
1180                 dname.name = rinfo->dir_dname[i];
1181                 dname.len = rinfo->dir_dname_len[i];
1182                 dname.hash = full_name_hash(dname.name, dname.len);
1183
1184                 vino.ino = le64_to_cpu(rinfo->dir_in[i].in->ino);
1185                 vino.snap = le64_to_cpu(rinfo->dir_in[i].in->snapid);
1186
1187 retry_lookup:
1188                 dn = d_lookup(parent, &dname);
1189                 dout("d_lookup on parent=%p name=%.*s got %p\n",
1190                      parent, dname.len, dname.name, dn);
1191
1192                 if (!dn) {
1193                         dn = d_alloc(parent, &dname);
1194                         dout("d_alloc %p '%.*s' = %p\n", parent,
1195                              dname.len, dname.name, dn);
1196                         if (dn == NULL) {
1197                                 dout("d_alloc badness\n");
1198                                 err = -ENOMEM;
1199                                 goto out;
1200                         }
1201                         err = ceph_init_dentry(dn);
1202                         if (err < 0) {
1203                                 dput(dn);
1204                                 goto out;
1205                         }
1206                 } else if (dn->d_inode &&
1207                            (ceph_ino(dn->d_inode) != vino.ino ||
1208                             ceph_snap(dn->d_inode) != vino.snap)) {
1209                         dout(" dn %p points to wrong inode %p\n",
1210                              dn, dn->d_inode);
1211                         d_delete(dn);
1212                         dput(dn);
1213                         goto retry_lookup;
1214                 } else {
1215                         /* reorder parent's d_subdirs */
1216                         spin_lock(&dcache_lock);
1217                         spin_lock(&dn->d_lock);
1218                         list_move(&dn->d_u.d_child, &parent->d_subdirs);
1219                         spin_unlock(&dn->d_lock);
1220                         spin_unlock(&dcache_lock);
1221                 }
1222
1223                 di = dn->d_fsdata;
1224                 di->offset = ceph_make_fpos(frag, i + req->r_readdir_offset);
1225
1226                 /* inode */
1227                 if (dn->d_inode) {
1228                         in = dn->d_inode;
1229                 } else {
1230                         in = ceph_get_inode(parent->d_sb, vino);
1231                         if (in == NULL) {
1232                                 dout("new_inode badness\n");
1233                                 d_delete(dn);
1234                                 dput(dn);
1235                                 err = -ENOMEM;
1236                                 goto out;
1237                         }
1238                         dn = splice_dentry(dn, in, NULL);
1239                         if (IS_ERR(dn))
1240                                 dn = NULL;
1241                 }
1242
1243                 if (fill_inode(in, &rinfo->dir_in[i], NULL, session,
1244                                req->r_request_started, -1,
1245                                &req->r_caps_reservation) < 0) {
1246                         pr_err("fill_inode badness on %p\n", in);
1247                         goto next_item;
1248                 }
1249                 if (dn)
1250                         update_dentry_lease(dn, rinfo->dir_dlease[i],
1251                                             req->r_session,
1252                                             req->r_request_started);
1253 next_item:
1254                 if (dn)
1255                         dput(dn);
1256         }
1257         req->r_did_prepopulate = true;
1258
1259 out:
1260         if (snapdir) {
1261                 iput(snapdir);
1262                 dput(parent);
1263         }
1264         dout("readdir_prepopulate done\n");
1265         return err;
1266 }
1267
1268 int ceph_inode_set_size(struct inode *inode, loff_t size)
1269 {
1270         struct ceph_inode_info *ci = ceph_inode(inode);
1271         int ret = 0;
1272
1273         spin_lock(&inode->i_lock);
1274         dout("set_size %p %llu -> %llu\n", inode, inode->i_size, size);
1275         inode->i_size = size;
1276         inode->i_blocks = (size + (1 << 9) - 1) >> 9;
1277
1278         /* tell the MDS if we are approaching max_size */
1279         if ((size << 1) >= ci->i_max_size &&
1280             (ci->i_reported_size << 1) < ci->i_max_size)
1281                 ret = 1;
1282
1283         spin_unlock(&inode->i_lock);
1284         return ret;
1285 }
1286
1287 /*
1288  * Write back inode data in a worker thread.  (This can't be done
1289  * in the message handler context.)
1290  */
1291 void ceph_queue_writeback(struct inode *inode)
1292 {
1293         if (queue_work(ceph_inode_to_client(inode)->wb_wq,
1294                        &ceph_inode(inode)->i_wb_work)) {
1295                 dout("ceph_queue_writeback %p\n", inode);
1296                 igrab(inode);
1297         } else {
1298                 dout("ceph_queue_writeback %p failed\n", inode);
1299         }
1300 }
1301
1302 static void ceph_writeback_work(struct work_struct *work)
1303 {
1304         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1305                                                   i_wb_work);
1306         struct inode *inode = &ci->vfs_inode;
1307
1308         dout("writeback %p\n", inode);
1309         filemap_fdatawrite(&inode->i_data);
1310         iput(inode);
1311 }
1312
1313 /*
1314  * queue an async invalidation
1315  */
1316 void ceph_queue_invalidate(struct inode *inode)
1317 {
1318         if (queue_work(ceph_inode_to_client(inode)->pg_inv_wq,
1319                        &ceph_inode(inode)->i_pg_inv_work)) {
1320                 dout("ceph_queue_invalidate %p\n", inode);
1321                 igrab(inode);
1322         } else {
1323                 dout("ceph_queue_invalidate %p failed\n", inode);
1324         }
1325 }
1326
1327 /*
1328  * invalidate any pages that are not dirty or under writeback.  this
1329  * includes pages that are clean and mapped.
1330  */
1331 static void ceph_invalidate_nondirty_pages(struct address_space *mapping)
1332 {
1333         struct pagevec pvec;
1334         pgoff_t next = 0;
1335         int i;
1336
1337         pagevec_init(&pvec, 0);
1338         while (pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
1339                 for (i = 0; i < pagevec_count(&pvec); i++) {
1340                         struct page *page = pvec.pages[i];
1341                         pgoff_t index;
1342                         int skip_page =
1343                                 (PageDirty(page) || PageWriteback(page));
1344
1345                         if (!skip_page)
1346                                 skip_page = !trylock_page(page);
1347
1348                         /*
1349                          * We really shouldn't be looking at the ->index of an
1350                          * unlocked page.  But we're not allowed to lock these
1351                          * pages.  So we rely upon nobody altering the ->index
1352                          * of this (pinned-by-us) page.
1353                          */
1354                         index = page->index;
1355                         if (index > next)
1356                                 next = index;
1357                         next++;
1358
1359                         if (skip_page)
1360                                 continue;
1361
1362                         generic_error_remove_page(mapping, page);
1363                         unlock_page(page);
1364                 }
1365                 pagevec_release(&pvec);
1366                 cond_resched();
1367         }
1368 }
1369
1370 /*
1371  * Invalidate inode pages in a worker thread.  (This can't be done
1372  * in the message handler context.)
1373  */
1374 static void ceph_invalidate_work(struct work_struct *work)
1375 {
1376         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1377                                                   i_pg_inv_work);
1378         struct inode *inode = &ci->vfs_inode;
1379         u32 orig_gen;
1380         int check = 0;
1381
1382         spin_lock(&inode->i_lock);
1383         dout("invalidate_pages %p gen %d revoking %d\n", inode,
1384              ci->i_rdcache_gen, ci->i_rdcache_revoking);
1385         if (ci->i_rdcache_gen == 0 ||
1386             ci->i_rdcache_revoking != ci->i_rdcache_gen) {
1387                 BUG_ON(ci->i_rdcache_revoking > ci->i_rdcache_gen);
1388                 /* nevermind! */
1389                 ci->i_rdcache_revoking = 0;
1390                 spin_unlock(&inode->i_lock);
1391                 goto out;
1392         }
1393         orig_gen = ci->i_rdcache_gen;
1394         spin_unlock(&inode->i_lock);
1395
1396         ceph_invalidate_nondirty_pages(inode->i_mapping);
1397
1398         spin_lock(&inode->i_lock);
1399         if (orig_gen == ci->i_rdcache_gen) {
1400                 dout("invalidate_pages %p gen %d successful\n", inode,
1401                      ci->i_rdcache_gen);
1402                 ci->i_rdcache_gen = 0;
1403                 ci->i_rdcache_revoking = 0;
1404                 check = 1;
1405         } else {
1406                 dout("invalidate_pages %p gen %d raced, gen now %d\n",
1407                      inode, orig_gen, ci->i_rdcache_gen);
1408         }
1409         spin_unlock(&inode->i_lock);
1410
1411         if (check)
1412                 ceph_check_caps(ci, 0, NULL);
1413 out:
1414         iput(inode);
1415 }
1416
1417
1418 /*
1419  * called by trunc_wq; take i_mutex ourselves
1420  *
1421  * We also truncate in a separate thread as well.
1422  */
1423 static void ceph_vmtruncate_work(struct work_struct *work)
1424 {
1425         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1426                                                   i_vmtruncate_work);
1427         struct inode *inode = &ci->vfs_inode;
1428
1429         dout("vmtruncate_work %p\n", inode);
1430         mutex_lock(&inode->i_mutex);
1431         __ceph_do_pending_vmtruncate(inode);
1432         mutex_unlock(&inode->i_mutex);
1433         iput(inode);
1434 }
1435
1436 /*
1437  * Queue an async vmtruncate.  If we fail to queue work, we will handle
1438  * the truncation the next time we call __ceph_do_pending_vmtruncate.
1439  */
1440 void ceph_queue_vmtruncate(struct inode *inode)
1441 {
1442         struct ceph_inode_info *ci = ceph_inode(inode);
1443
1444         if (queue_work(ceph_sb_to_client(inode->i_sb)->trunc_wq,
1445                        &ci->i_vmtruncate_work)) {
1446                 dout("ceph_queue_vmtruncate %p\n", inode);
1447                 igrab(inode);
1448         } else {
1449                 dout("ceph_queue_vmtruncate %p failed, pending=%d\n",
1450                      inode, ci->i_truncate_pending);
1451         }
1452 }
1453
1454 /*
1455  * called with i_mutex held.
1456  *
1457  * Make sure any pending truncation is applied before doing anything
1458  * that may depend on it.
1459  */
1460 void __ceph_do_pending_vmtruncate(struct inode *inode)
1461 {
1462         struct ceph_inode_info *ci = ceph_inode(inode);
1463         u64 to;
1464         int wrbuffer_refs, wake = 0;
1465
1466 retry:
1467         spin_lock(&inode->i_lock);
1468         if (ci->i_truncate_pending == 0) {
1469                 dout("__do_pending_vmtruncate %p none pending\n", inode);
1470                 spin_unlock(&inode->i_lock);
1471                 return;
1472         }
1473
1474         /*
1475          * make sure any dirty snapped pages are flushed before we
1476          * possibly truncate them.. so write AND block!
1477          */
1478         if (ci->i_wrbuffer_ref_head < ci->i_wrbuffer_ref) {
1479                 dout("__do_pending_vmtruncate %p flushing snaps first\n",
1480                      inode);
1481                 spin_unlock(&inode->i_lock);
1482                 filemap_write_and_wait_range(&inode->i_data, 0,
1483                                              inode->i_sb->s_maxbytes);
1484                 goto retry;
1485         }
1486
1487         to = ci->i_truncate_size;
1488         wrbuffer_refs = ci->i_wrbuffer_ref;
1489         dout("__do_pending_vmtruncate %p (%d) to %lld\n", inode,
1490              ci->i_truncate_pending, to);
1491         spin_unlock(&inode->i_lock);
1492
1493         truncate_inode_pages(inode->i_mapping, to);
1494
1495         spin_lock(&inode->i_lock);
1496         ci->i_truncate_pending--;
1497         if (ci->i_truncate_pending == 0)
1498                 wake = 1;
1499         spin_unlock(&inode->i_lock);
1500
1501         if (wrbuffer_refs == 0)
1502                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1503         if (wake)
1504                 wake_up_all(&ci->i_cap_wq);
1505 }
1506
1507
1508 /*
1509  * symlinks
1510  */
1511 static void *ceph_sym_follow_link(struct dentry *dentry, struct nameidata *nd)
1512 {
1513         struct ceph_inode_info *ci = ceph_inode(dentry->d_inode);
1514         nd_set_link(nd, ci->i_symlink);
1515         return NULL;
1516 }
1517
1518 static const struct inode_operations ceph_symlink_iops = {
1519         .readlink = generic_readlink,
1520         .follow_link = ceph_sym_follow_link,
1521 };
1522
1523 /*
1524  * setattr
1525  */
1526 int ceph_setattr(struct dentry *dentry, struct iattr *attr)
1527 {
1528         struct inode *inode = dentry->d_inode;
1529         struct ceph_inode_info *ci = ceph_inode(inode);
1530         struct inode *parent_inode = dentry->d_parent->d_inode;
1531         const unsigned int ia_valid = attr->ia_valid;
1532         struct ceph_mds_request *req;
1533         struct ceph_mds_client *mdsc = &ceph_sb_to_client(dentry->d_sb)->mdsc;
1534         int issued;
1535         int release = 0, dirtied = 0;
1536         int mask = 0;
1537         int err = 0;
1538
1539         if (ceph_snap(inode) != CEPH_NOSNAP)
1540                 return -EROFS;
1541
1542         __ceph_do_pending_vmtruncate(inode);
1543
1544         err = inode_change_ok(inode, attr);
1545         if (err != 0)
1546                 return err;
1547
1548         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETATTR,
1549                                        USE_AUTH_MDS);
1550         if (IS_ERR(req))
1551                 return PTR_ERR(req);
1552
1553         spin_lock(&inode->i_lock);
1554         issued = __ceph_caps_issued(ci, NULL);
1555         dout("setattr %p issued %s\n", inode, ceph_cap_string(issued));
1556
1557         if (ia_valid & ATTR_UID) {
1558                 dout("setattr %p uid %d -> %d\n", inode,
1559                      inode->i_uid, attr->ia_uid);
1560                 if (issued & CEPH_CAP_AUTH_EXCL) {
1561                         inode->i_uid = attr->ia_uid;
1562                         dirtied |= CEPH_CAP_AUTH_EXCL;
1563                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
1564                            attr->ia_uid != inode->i_uid) {
1565                         req->r_args.setattr.uid = cpu_to_le32(attr->ia_uid);
1566                         mask |= CEPH_SETATTR_UID;
1567                         release |= CEPH_CAP_AUTH_SHARED;
1568                 }
1569         }
1570         if (ia_valid & ATTR_GID) {
1571                 dout("setattr %p gid %d -> %d\n", inode,
1572                      inode->i_gid, attr->ia_gid);
1573                 if (issued & CEPH_CAP_AUTH_EXCL) {
1574                         inode->i_gid = attr->ia_gid;
1575                         dirtied |= CEPH_CAP_AUTH_EXCL;
1576                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
1577                            attr->ia_gid != inode->i_gid) {
1578                         req->r_args.setattr.gid = cpu_to_le32(attr->ia_gid);
1579                         mask |= CEPH_SETATTR_GID;
1580                         release |= CEPH_CAP_AUTH_SHARED;
1581                 }
1582         }
1583         if (ia_valid & ATTR_MODE) {
1584                 dout("setattr %p mode 0%o -> 0%o\n", inode, inode->i_mode,
1585                      attr->ia_mode);
1586                 if (issued & CEPH_CAP_AUTH_EXCL) {
1587                         inode->i_mode = attr->ia_mode;
1588                         dirtied |= CEPH_CAP_AUTH_EXCL;
1589                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
1590                            attr->ia_mode != inode->i_mode) {
1591                         req->r_args.setattr.mode = cpu_to_le32(attr->ia_mode);
1592                         mask |= CEPH_SETATTR_MODE;
1593                         release |= CEPH_CAP_AUTH_SHARED;
1594                 }
1595         }
1596
1597         if (ia_valid & ATTR_ATIME) {
1598                 dout("setattr %p atime %ld.%ld -> %ld.%ld\n", inode,
1599                      inode->i_atime.tv_sec, inode->i_atime.tv_nsec,
1600                      attr->ia_atime.tv_sec, attr->ia_atime.tv_nsec);
1601                 if (issued & CEPH_CAP_FILE_EXCL) {
1602                         ci->i_time_warp_seq++;
1603                         inode->i_atime = attr->ia_atime;
1604                         dirtied |= CEPH_CAP_FILE_EXCL;
1605                 } else if ((issued & CEPH_CAP_FILE_WR) &&
1606                            timespec_compare(&inode->i_atime,
1607                                             &attr->ia_atime) < 0) {
1608                         inode->i_atime = attr->ia_atime;
1609                         dirtied |= CEPH_CAP_FILE_WR;
1610                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
1611                            !timespec_equal(&inode->i_atime, &attr->ia_atime)) {
1612                         ceph_encode_timespec(&req->r_args.setattr.atime,
1613                                              &attr->ia_atime);
1614                         mask |= CEPH_SETATTR_ATIME;
1615                         release |= CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_RD |
1616                                 CEPH_CAP_FILE_WR;
1617                 }
1618         }
1619         if (ia_valid & ATTR_MTIME) {
1620                 dout("setattr %p mtime %ld.%ld -> %ld.%ld\n", inode,
1621                      inode->i_mtime.tv_sec, inode->i_mtime.tv_nsec,
1622                      attr->ia_mtime.tv_sec, attr->ia_mtime.tv_nsec);
1623                 if (issued & CEPH_CAP_FILE_EXCL) {
1624                         ci->i_time_warp_seq++;
1625                         inode->i_mtime = attr->ia_mtime;
1626                         dirtied |= CEPH_CAP_FILE_EXCL;
1627                 } else if ((issued & CEPH_CAP_FILE_WR) &&
1628                            timespec_compare(&inode->i_mtime,
1629                                             &attr->ia_mtime) < 0) {
1630                         inode->i_mtime = attr->ia_mtime;
1631                         dirtied |= CEPH_CAP_FILE_WR;
1632                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
1633                            !timespec_equal(&inode->i_mtime, &attr->ia_mtime)) {
1634                         ceph_encode_timespec(&req->r_args.setattr.mtime,
1635                                              &attr->ia_mtime);
1636                         mask |= CEPH_SETATTR_MTIME;
1637                         release |= CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_RD |
1638                                 CEPH_CAP_FILE_WR;
1639                 }
1640         }
1641         if (ia_valid & ATTR_SIZE) {
1642                 dout("setattr %p size %lld -> %lld\n", inode,
1643                      inode->i_size, attr->ia_size);
1644                 if (attr->ia_size > inode->i_sb->s_maxbytes) {
1645                         err = -EINVAL;
1646                         goto out;
1647                 }
1648                 if ((issued & CEPH_CAP_FILE_EXCL) &&
1649                     attr->ia_size > inode->i_size) {
1650                         inode->i_size = attr->ia_size;
1651                         inode->i_blocks =
1652                                 (attr->ia_size + (1 << 9) - 1) >> 9;
1653                         inode->i_ctime = attr->ia_ctime;
1654                         ci->i_reported_size = attr->ia_size;
1655                         dirtied |= CEPH_CAP_FILE_EXCL;
1656                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
1657                            attr->ia_size != inode->i_size) {
1658                         req->r_args.setattr.size = cpu_to_le64(attr->ia_size);
1659                         req->r_args.setattr.old_size =
1660                                 cpu_to_le64(inode->i_size);
1661                         mask |= CEPH_SETATTR_SIZE;
1662                         release |= CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_RD |
1663                                 CEPH_CAP_FILE_WR;
1664                 }
1665         }
1666
1667         /* these do nothing */
1668         if (ia_valid & ATTR_CTIME) {
1669                 bool only = (ia_valid & (ATTR_SIZE|ATTR_MTIME|ATTR_ATIME|
1670                                          ATTR_MODE|ATTR_UID|ATTR_GID)) == 0;
1671                 dout("setattr %p ctime %ld.%ld -> %ld.%ld (%s)\n", inode,
1672                      inode->i_ctime.tv_sec, inode->i_ctime.tv_nsec,
1673                      attr->ia_ctime.tv_sec, attr->ia_ctime.tv_nsec,
1674                      only ? "ctime only" : "ignored");
1675                 inode->i_ctime = attr->ia_ctime;
1676                 if (only) {
1677                         /*
1678                          * if kernel wants to dirty ctime but nothing else,
1679                          * we need to choose a cap to dirty under, or do
1680                          * a almost-no-op setattr
1681                          */
1682                         if (issued & CEPH_CAP_AUTH_EXCL)
1683                                 dirtied |= CEPH_CAP_AUTH_EXCL;
1684                         else if (issued & CEPH_CAP_FILE_EXCL)
1685                                 dirtied |= CEPH_CAP_FILE_EXCL;
1686                         else if (issued & CEPH_CAP_XATTR_EXCL)
1687                                 dirtied |= CEPH_CAP_XATTR_EXCL;
1688                         else
1689                                 mask |= CEPH_SETATTR_CTIME;
1690                 }
1691         }
1692         if (ia_valid & ATTR_FILE)
1693                 dout("setattr %p ATTR_FILE ... hrm!\n", inode);
1694
1695         if (dirtied) {
1696                 __ceph_mark_dirty_caps(ci, dirtied);
1697                 inode->i_ctime = CURRENT_TIME;
1698         }
1699
1700         release &= issued;
1701         spin_unlock(&inode->i_lock);
1702
1703         if (mask) {
1704                 req->r_inode = igrab(inode);
1705                 req->r_inode_drop = release;
1706                 req->r_args.setattr.mask = cpu_to_le32(mask);
1707                 req->r_num_caps = 1;
1708                 err = ceph_mdsc_do_request(mdsc, parent_inode, req);
1709         }
1710         dout("setattr %p result=%d (%s locally, %d remote)\n", inode, err,
1711              ceph_cap_string(dirtied), mask);
1712
1713         ceph_mdsc_put_request(req);
1714         __ceph_do_pending_vmtruncate(inode);
1715         return err;
1716 out:
1717         spin_unlock(&inode->i_lock);
1718         ceph_mdsc_put_request(req);
1719         return err;
1720 }
1721
1722 /*
1723  * Verify that we have a lease on the given mask.  If not,
1724  * do a getattr against an mds.
1725  */
1726 int ceph_do_getattr(struct inode *inode, int mask)
1727 {
1728         struct ceph_client *client = ceph_sb_to_client(inode->i_sb);
1729         struct ceph_mds_client *mdsc = &client->mdsc;
1730         struct ceph_mds_request *req;
1731         int err;
1732
1733         if (ceph_snap(inode) == CEPH_SNAPDIR) {
1734                 dout("do_getattr inode %p SNAPDIR\n", inode);
1735                 return 0;
1736         }
1737
1738         dout("do_getattr inode %p mask %s\n", inode, ceph_cap_string(mask));
1739         if (ceph_caps_issued_mask(ceph_inode(inode), mask, 1))
1740                 return 0;
1741
1742         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, USE_ANY_MDS);
1743         if (IS_ERR(req))
1744                 return PTR_ERR(req);
1745         req->r_inode = igrab(inode);
1746         req->r_num_caps = 1;
1747         req->r_args.getattr.mask = cpu_to_le32(mask);
1748         err = ceph_mdsc_do_request(mdsc, NULL, req);
1749         ceph_mdsc_put_request(req);
1750         dout("do_getattr result=%d\n", err);
1751         return err;
1752 }
1753
1754
1755 /*
1756  * Check inode permissions.  We verify we have a valid value for
1757  * the AUTH cap, then call the generic handler.
1758  */
1759 int ceph_permission(struct inode *inode, int mask)
1760 {
1761         int err = ceph_do_getattr(inode, CEPH_CAP_AUTH_SHARED);
1762
1763         if (!err)
1764                 err = generic_permission(inode, mask, NULL);
1765         return err;
1766 }
1767
1768 /*
1769  * Get all attributes.  Hopefully somedata we'll have a statlite()
1770  * and can limit the fields we require to be accurate.
1771  */
1772 int ceph_getattr(struct vfsmount *mnt, struct dentry *dentry,
1773                  struct kstat *stat)
1774 {
1775         struct inode *inode = dentry->d_inode;
1776         struct ceph_inode_info *ci = ceph_inode(inode);
1777         int err;
1778
1779         err = ceph_do_getattr(inode, CEPH_STAT_CAP_INODE_ALL);
1780         if (!err) {
1781                 generic_fillattr(inode, stat);
1782                 stat->ino = inode->i_ino;
1783                 if (ceph_snap(inode) != CEPH_NOSNAP)
1784                         stat->dev = ceph_snap(inode);
1785                 else
1786                         stat->dev = 0;
1787                 if (S_ISDIR(inode->i_mode)) {
1788                         stat->size = ci->i_rbytes;
1789                         stat->blocks = 0;
1790                         stat->blksize = 65536;
1791                 }
1792         }
1793         return err;
1794 }