BKL: remove extraneous #include <smp_lock.h>
[linux-2.6.git] / fs / ceph / inode.c
1 #include <linux/ceph/ceph_debug.h>
2
3 #include <linux/module.h>
4 #include <linux/fs.h>
5 #include <linux/slab.h>
6 #include <linux/string.h>
7 #include <linux/uaccess.h>
8 #include <linux/kernel.h>
9 #include <linux/namei.h>
10 #include <linux/writeback.h>
11 #include <linux/vmalloc.h>
12 #include <linux/pagevec.h>
13
14 #include "super.h"
15 #include "mds_client.h"
16 #include <linux/ceph/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                                        CEPH_CAP_FILE_LAZYIO)) ||
448                             mapping_mapped(inode->i_mapping) ||
449                             __ceph_caps_file_wanted(ci)) {
450                                 ci->i_truncate_pending++;
451                                 queue_trunc = 1;
452                         }
453                 }
454         }
455         if (ceph_seq_cmp(truncate_seq, ci->i_truncate_seq) >= 0 &&
456             ci->i_truncate_size != truncate_size) {
457                 dout("truncate_size %lld -> %llu\n", ci->i_truncate_size,
458                      truncate_size);
459                 ci->i_truncate_size = truncate_size;
460         }
461         return queue_trunc;
462 }
463
464 void ceph_fill_file_time(struct inode *inode, int issued,
465                          u64 time_warp_seq, struct timespec *ctime,
466                          struct timespec *mtime, struct timespec *atime)
467 {
468         struct ceph_inode_info *ci = ceph_inode(inode);
469         int warn = 0;
470
471         if (issued & (CEPH_CAP_FILE_EXCL|
472                       CEPH_CAP_FILE_WR|
473                       CEPH_CAP_FILE_BUFFER)) {
474                 if (timespec_compare(ctime, &inode->i_ctime) > 0) {
475                         dout("ctime %ld.%09ld -> %ld.%09ld inc w/ cap\n",
476                              inode->i_ctime.tv_sec, inode->i_ctime.tv_nsec,
477                              ctime->tv_sec, ctime->tv_nsec);
478                         inode->i_ctime = *ctime;
479                 }
480                 if (ceph_seq_cmp(time_warp_seq, ci->i_time_warp_seq) > 0) {
481                         /* the MDS did a utimes() */
482                         dout("mtime %ld.%09ld -> %ld.%09ld "
483                              "tw %d -> %d\n",
484                              inode->i_mtime.tv_sec, inode->i_mtime.tv_nsec,
485                              mtime->tv_sec, mtime->tv_nsec,
486                              ci->i_time_warp_seq, (int)time_warp_seq);
487
488                         inode->i_mtime = *mtime;
489                         inode->i_atime = *atime;
490                         ci->i_time_warp_seq = time_warp_seq;
491                 } else if (time_warp_seq == ci->i_time_warp_seq) {
492                         /* nobody did utimes(); take the max */
493                         if (timespec_compare(mtime, &inode->i_mtime) > 0) {
494                                 dout("mtime %ld.%09ld -> %ld.%09ld inc\n",
495                                      inode->i_mtime.tv_sec,
496                                      inode->i_mtime.tv_nsec,
497                                      mtime->tv_sec, mtime->tv_nsec);
498                                 inode->i_mtime = *mtime;
499                         }
500                         if (timespec_compare(atime, &inode->i_atime) > 0) {
501                                 dout("atime %ld.%09ld -> %ld.%09ld inc\n",
502                                      inode->i_atime.tv_sec,
503                                      inode->i_atime.tv_nsec,
504                                      atime->tv_sec, atime->tv_nsec);
505                                 inode->i_atime = *atime;
506                         }
507                 } else if (issued & CEPH_CAP_FILE_EXCL) {
508                         /* we did a utimes(); ignore mds values */
509                 } else {
510                         warn = 1;
511                 }
512         } else {
513                 /* we have no write caps; whatever the MDS says is true */
514                 if (ceph_seq_cmp(time_warp_seq, ci->i_time_warp_seq) >= 0) {
515                         inode->i_ctime = *ctime;
516                         inode->i_mtime = *mtime;
517                         inode->i_atime = *atime;
518                         ci->i_time_warp_seq = time_warp_seq;
519                 } else {
520                         warn = 1;
521                 }
522         }
523         if (warn) /* time_warp_seq shouldn't go backwards */
524                 dout("%p mds time_warp_seq %llu < %u\n",
525                      inode, time_warp_seq, ci->i_time_warp_seq);
526 }
527
528 /*
529  * Populate an inode based on info from mds.  May be called on new or
530  * existing inodes.
531  */
532 static int fill_inode(struct inode *inode,
533                       struct ceph_mds_reply_info_in *iinfo,
534                       struct ceph_mds_reply_dirfrag *dirinfo,
535                       struct ceph_mds_session *session,
536                       unsigned long ttl_from, int cap_fmode,
537                       struct ceph_cap_reservation *caps_reservation)
538 {
539         struct ceph_mds_reply_inode *info = iinfo->in;
540         struct ceph_inode_info *ci = ceph_inode(inode);
541         int i;
542         int issued, implemented;
543         struct timespec mtime, atime, ctime;
544         u32 nsplits;
545         struct ceph_buffer *xattr_blob = NULL;
546         int err = 0;
547         int queue_trunc = 0;
548
549         dout("fill_inode %p ino %llx.%llx v %llu had %llu\n",
550              inode, ceph_vinop(inode), le64_to_cpu(info->version),
551              ci->i_version);
552
553         /*
554          * prealloc xattr data, if it looks like we'll need it.  only
555          * if len > 4 (meaning there are actually xattrs; the first 4
556          * bytes are the xattr count).
557          */
558         if (iinfo->xattr_len > 4) {
559                 xattr_blob = ceph_buffer_new(iinfo->xattr_len, GFP_NOFS);
560                 if (!xattr_blob)
561                         pr_err("fill_inode ENOMEM xattr blob %d bytes\n",
562                                iinfo->xattr_len);
563         }
564
565         spin_lock(&inode->i_lock);
566
567         /*
568          * provided version will be odd if inode value is projected,
569          * even if stable.  skip the update if we have a newer info
570          * (e.g., due to inode info racing form multiple MDSs), or if
571          * we are getting projected (unstable) inode info.
572          */
573         if (le64_to_cpu(info->version) > 0 &&
574             (ci->i_version & ~1) > le64_to_cpu(info->version))
575                 goto no_change;
576
577         issued = __ceph_caps_issued(ci, &implemented);
578         issued |= implemented | __ceph_caps_dirty(ci);
579
580         /* update inode */
581         ci->i_version = le64_to_cpu(info->version);
582         inode->i_version++;
583         inode->i_rdev = le32_to_cpu(info->rdev);
584
585         if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
586                 inode->i_mode = le32_to_cpu(info->mode);
587                 inode->i_uid = le32_to_cpu(info->uid);
588                 inode->i_gid = le32_to_cpu(info->gid);
589                 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
590                      inode->i_uid, inode->i_gid);
591         }
592
593         if ((issued & CEPH_CAP_LINK_EXCL) == 0)
594                 inode->i_nlink = le32_to_cpu(info->nlink);
595
596         /* be careful with mtime, atime, size */
597         ceph_decode_timespec(&atime, &info->atime);
598         ceph_decode_timespec(&mtime, &info->mtime);
599         ceph_decode_timespec(&ctime, &info->ctime);
600         queue_trunc = ceph_fill_file_size(inode, issued,
601                                           le32_to_cpu(info->truncate_seq),
602                                           le64_to_cpu(info->truncate_size),
603                                           le64_to_cpu(info->size));
604         ceph_fill_file_time(inode, issued,
605                             le32_to_cpu(info->time_warp_seq),
606                             &ctime, &mtime, &atime);
607
608         ci->i_max_size = le64_to_cpu(info->max_size);
609         ci->i_layout = info->layout;
610         inode->i_blkbits = fls(le32_to_cpu(info->layout.fl_stripe_unit)) - 1;
611
612         /* xattrs */
613         /* note that if i_xattrs.len <= 4, i_xattrs.data will still be NULL. */
614         if ((issued & CEPH_CAP_XATTR_EXCL) == 0 &&
615             le64_to_cpu(info->xattr_version) > ci->i_xattrs.version) {
616                 if (ci->i_xattrs.blob)
617                         ceph_buffer_put(ci->i_xattrs.blob);
618                 ci->i_xattrs.blob = xattr_blob;
619                 if (xattr_blob)
620                         memcpy(ci->i_xattrs.blob->vec.iov_base,
621                                iinfo->xattr_data, iinfo->xattr_len);
622                 ci->i_xattrs.version = le64_to_cpu(info->xattr_version);
623                 xattr_blob = NULL;
624         }
625
626         inode->i_mapping->a_ops = &ceph_aops;
627         inode->i_mapping->backing_dev_info =
628                 &ceph_sb_to_client(inode->i_sb)->backing_dev_info;
629
630         switch (inode->i_mode & S_IFMT) {
631         case S_IFIFO:
632         case S_IFBLK:
633         case S_IFCHR:
634         case S_IFSOCK:
635                 init_special_inode(inode, inode->i_mode, inode->i_rdev);
636                 inode->i_op = &ceph_file_iops;
637                 break;
638         case S_IFREG:
639                 inode->i_op = &ceph_file_iops;
640                 inode->i_fop = &ceph_file_fops;
641                 break;
642         case S_IFLNK:
643                 inode->i_op = &ceph_symlink_iops;
644                 if (!ci->i_symlink) {
645                         int symlen = iinfo->symlink_len;
646                         char *sym;
647
648                         BUG_ON(symlen != inode->i_size);
649                         spin_unlock(&inode->i_lock);
650
651                         err = -ENOMEM;
652                         sym = kmalloc(symlen+1, GFP_NOFS);
653                         if (!sym)
654                                 goto out;
655                         memcpy(sym, iinfo->symlink, symlen);
656                         sym[symlen] = 0;
657
658                         spin_lock(&inode->i_lock);
659                         if (!ci->i_symlink)
660                                 ci->i_symlink = sym;
661                         else
662                                 kfree(sym); /* lost a race */
663                 }
664                 break;
665         case S_IFDIR:
666                 inode->i_op = &ceph_dir_iops;
667                 inode->i_fop = &ceph_dir_fops;
668
669                 ci->i_files = le64_to_cpu(info->files);
670                 ci->i_subdirs = le64_to_cpu(info->subdirs);
671                 ci->i_rbytes = le64_to_cpu(info->rbytes);
672                 ci->i_rfiles = le64_to_cpu(info->rfiles);
673                 ci->i_rsubdirs = le64_to_cpu(info->rsubdirs);
674                 ceph_decode_timespec(&ci->i_rctime, &info->rctime);
675
676                 /* set dir completion flag? */
677                 if (ci->i_files == 0 && ci->i_subdirs == 0 &&
678                     ceph_snap(inode) == CEPH_NOSNAP &&
679                     (le32_to_cpu(info->cap.caps) & CEPH_CAP_FILE_SHARED) &&
680                     (issued & CEPH_CAP_FILE_EXCL) == 0 &&
681                     (ci->i_ceph_flags & CEPH_I_COMPLETE) == 0) {
682                         dout(" marking %p complete (empty)\n", inode);
683                         ci->i_ceph_flags |= CEPH_I_COMPLETE;
684                         ci->i_max_offset = 2;
685                 }
686
687                 /* it may be better to set st_size in getattr instead? */
688                 if (ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), RBYTES))
689                         inode->i_size = ci->i_rbytes;
690                 break;
691         default:
692                 pr_err("fill_inode %llx.%llx BAD mode 0%o\n",
693                        ceph_vinop(inode), inode->i_mode);
694         }
695
696 no_change:
697         spin_unlock(&inode->i_lock);
698
699         /* queue truncate if we saw i_size decrease */
700         if (queue_trunc)
701                 ceph_queue_vmtruncate(inode);
702
703         /* populate frag tree */
704         /* FIXME: move me up, if/when version reflects fragtree changes */
705         nsplits = le32_to_cpu(info->fragtree.nsplits);
706         mutex_lock(&ci->i_fragtree_mutex);
707         for (i = 0; i < nsplits; i++) {
708                 u32 id = le32_to_cpu(info->fragtree.splits[i].frag);
709                 struct ceph_inode_frag *frag = __get_or_create_frag(ci, id);
710
711                 if (IS_ERR(frag))
712                         continue;
713                 frag->split_by = le32_to_cpu(info->fragtree.splits[i].by);
714                 dout(" frag %x split by %d\n", frag->frag, frag->split_by);
715         }
716         mutex_unlock(&ci->i_fragtree_mutex);
717
718         /* were we issued a capability? */
719         if (info->cap.caps) {
720                 if (ceph_snap(inode) == CEPH_NOSNAP) {
721                         ceph_add_cap(inode, session,
722                                      le64_to_cpu(info->cap.cap_id),
723                                      cap_fmode,
724                                      le32_to_cpu(info->cap.caps),
725                                      le32_to_cpu(info->cap.wanted),
726                                      le32_to_cpu(info->cap.seq),
727                                      le32_to_cpu(info->cap.mseq),
728                                      le64_to_cpu(info->cap.realm),
729                                      info->cap.flags,
730                                      caps_reservation);
731                 } else {
732                         spin_lock(&inode->i_lock);
733                         dout(" %p got snap_caps %s\n", inode,
734                              ceph_cap_string(le32_to_cpu(info->cap.caps)));
735                         ci->i_snap_caps |= le32_to_cpu(info->cap.caps);
736                         if (cap_fmode >= 0)
737                                 __ceph_get_fmode(ci, cap_fmode);
738                         spin_unlock(&inode->i_lock);
739                 }
740         } else if (cap_fmode >= 0) {
741                 pr_warning("mds issued no caps on %llx.%llx\n",
742                            ceph_vinop(inode));
743                 __ceph_get_fmode(ci, cap_fmode);
744         }
745
746         /* update delegation info? */
747         if (dirinfo)
748                 ceph_fill_dirfrag(inode, dirinfo);
749
750         err = 0;
751
752 out:
753         if (xattr_blob)
754                 ceph_buffer_put(xattr_blob);
755         return err;
756 }
757
758 /*
759  * caller should hold session s_mutex.
760  */
761 static void update_dentry_lease(struct dentry *dentry,
762                                 struct ceph_mds_reply_lease *lease,
763                                 struct ceph_mds_session *session,
764                                 unsigned long from_time)
765 {
766         struct ceph_dentry_info *di = ceph_dentry(dentry);
767         long unsigned duration = le32_to_cpu(lease->duration_ms);
768         long unsigned ttl = from_time + (duration * HZ) / 1000;
769         long unsigned half_ttl = from_time + (duration * HZ / 2) / 1000;
770         struct inode *dir;
771
772         /* only track leases on regular dentries */
773         if (dentry->d_op != &ceph_dentry_ops)
774                 return;
775
776         spin_lock(&dentry->d_lock);
777         dout("update_dentry_lease %p mask %d duration %lu ms ttl %lu\n",
778              dentry, le16_to_cpu(lease->mask), duration, ttl);
779
780         /* make lease_rdcache_gen match directory */
781         dir = dentry->d_parent->d_inode;
782         di->lease_shared_gen = ceph_inode(dir)->i_shared_gen;
783
784         if (lease->mask == 0)
785                 goto out_unlock;
786
787         if (di->lease_gen == session->s_cap_gen &&
788             time_before(ttl, dentry->d_time))
789                 goto out_unlock;  /* we already have a newer lease. */
790
791         if (di->lease_session && di->lease_session != session)
792                 goto out_unlock;
793
794         ceph_dentry_lru_touch(dentry);
795
796         if (!di->lease_session)
797                 di->lease_session = ceph_get_mds_session(session);
798         di->lease_gen = session->s_cap_gen;
799         di->lease_seq = le32_to_cpu(lease->seq);
800         di->lease_renew_after = half_ttl;
801         di->lease_renew_from = 0;
802         dentry->d_time = ttl;
803 out_unlock:
804         spin_unlock(&dentry->d_lock);
805         return;
806 }
807
808 /*
809  * Set dentry's directory position based on the current dir's max, and
810  * order it in d_subdirs, so that dcache_readdir behaves.
811  */
812 static void ceph_set_dentry_offset(struct dentry *dn)
813 {
814         struct dentry *dir = dn->d_parent;
815         struct inode *inode = dn->d_parent->d_inode;
816         struct ceph_dentry_info *di;
817
818         BUG_ON(!inode);
819
820         di = ceph_dentry(dn);
821
822         spin_lock(&inode->i_lock);
823         if ((ceph_inode(inode)->i_ceph_flags & CEPH_I_COMPLETE) == 0) {
824                 spin_unlock(&inode->i_lock);
825                 return;
826         }
827         di->offset = ceph_inode(inode)->i_max_offset++;
828         spin_unlock(&inode->i_lock);
829
830         spin_lock(&dcache_lock);
831         spin_lock(&dn->d_lock);
832         list_move(&dn->d_u.d_child, &dir->d_subdirs);
833         dout("set_dentry_offset %p %lld (%p %p)\n", dn, di->offset,
834              dn->d_u.d_child.prev, dn->d_u.d_child.next);
835         spin_unlock(&dn->d_lock);
836         spin_unlock(&dcache_lock);
837 }
838
839 /*
840  * splice a dentry to an inode.
841  * caller must hold directory i_mutex for this to be safe.
842  *
843  * we will only rehash the resulting dentry if @prehash is
844  * true; @prehash will be set to false (for the benefit of
845  * the caller) if we fail.
846  */
847 static struct dentry *splice_dentry(struct dentry *dn, struct inode *in,
848                                     bool *prehash, bool set_offset)
849 {
850         struct dentry *realdn;
851
852         BUG_ON(dn->d_inode);
853
854         /* dn must be unhashed */
855         if (!d_unhashed(dn))
856                 d_drop(dn);
857         realdn = d_materialise_unique(dn, in);
858         if (IS_ERR(realdn)) {
859                 pr_err("splice_dentry error %ld %p inode %p ino %llx.%llx\n",
860                        PTR_ERR(realdn), dn, in, ceph_vinop(in));
861                 if (prehash)
862                         *prehash = false; /* don't rehash on error */
863                 dn = realdn; /* note realdn contains the error */
864                 goto out;
865         } else if (realdn) {
866                 dout("dn %p (%d) spliced with %p (%d) "
867                      "inode %p ino %llx.%llx\n",
868                      dn, atomic_read(&dn->d_count),
869                      realdn, atomic_read(&realdn->d_count),
870                      realdn->d_inode, ceph_vinop(realdn->d_inode));
871                 dput(dn);
872                 dn = realdn;
873         } else {
874                 BUG_ON(!ceph_dentry(dn));
875                 dout("dn %p attached to %p ino %llx.%llx\n",
876                      dn, dn->d_inode, ceph_vinop(dn->d_inode));
877         }
878         if ((!prehash || *prehash) && d_unhashed(dn))
879                 d_rehash(dn);
880         if (set_offset)
881                 ceph_set_dentry_offset(dn);
882 out:
883         return dn;
884 }
885
886 /*
887  * Incorporate results into the local cache.  This is either just
888  * one inode, or a directory, dentry, and possibly linked-to inode (e.g.,
889  * after a lookup).
890  *
891  * A reply may contain
892  *         a directory inode along with a dentry.
893  *  and/or a target inode
894  *
895  * Called with snap_rwsem (read).
896  */
897 int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req,
898                     struct ceph_mds_session *session)
899 {
900         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
901         struct inode *in = NULL;
902         struct ceph_mds_reply_inode *ininfo;
903         struct ceph_vino vino;
904         struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
905         int i = 0;
906         int err = 0;
907
908         dout("fill_trace %p is_dentry %d is_target %d\n", req,
909              rinfo->head->is_dentry, rinfo->head->is_target);
910
911 #if 0
912         /*
913          * Debugging hook:
914          *
915          * If we resend completed ops to a recovering mds, we get no
916          * trace.  Since that is very rare, pretend this is the case
917          * to ensure the 'no trace' handlers in the callers behave.
918          *
919          * Fill in inodes unconditionally to avoid breaking cap
920          * invariants.
921          */
922         if (rinfo->head->op & CEPH_MDS_OP_WRITE) {
923                 pr_info("fill_trace faking empty trace on %lld %s\n",
924                         req->r_tid, ceph_mds_op_name(rinfo->head->op));
925                 if (rinfo->head->is_dentry) {
926                         rinfo->head->is_dentry = 0;
927                         err = fill_inode(req->r_locked_dir,
928                                          &rinfo->diri, rinfo->dirfrag,
929                                          session, req->r_request_started, -1);
930                 }
931                 if (rinfo->head->is_target) {
932                         rinfo->head->is_target = 0;
933                         ininfo = rinfo->targeti.in;
934                         vino.ino = le64_to_cpu(ininfo->ino);
935                         vino.snap = le64_to_cpu(ininfo->snapid);
936                         in = ceph_get_inode(sb, vino);
937                         err = fill_inode(in, &rinfo->targeti, NULL,
938                                          session, req->r_request_started,
939                                          req->r_fmode);
940                         iput(in);
941                 }
942         }
943 #endif
944
945         if (!rinfo->head->is_target && !rinfo->head->is_dentry) {
946                 dout("fill_trace reply is empty!\n");
947                 if (rinfo->head->result == 0 && req->r_locked_dir)
948                         ceph_invalidate_dir_request(req);
949                 return 0;
950         }
951
952         if (rinfo->head->is_dentry) {
953                 struct inode *dir = req->r_locked_dir;
954
955                 err = fill_inode(dir, &rinfo->diri, rinfo->dirfrag,
956                                  session, req->r_request_started, -1,
957                                  &req->r_caps_reservation);
958                 if (err < 0)
959                         return err;
960         }
961
962         /*
963          * ignore null lease/binding on snapdir ENOENT, or else we
964          * will have trouble splicing in the virtual snapdir later
965          */
966         if (rinfo->head->is_dentry && !req->r_aborted &&
967             (rinfo->head->is_target || strncmp(req->r_dentry->d_name.name,
968                                                fsc->mount_options->snapdir_name,
969                                                req->r_dentry->d_name.len))) {
970                 /*
971                  * lookup link rename   : null -> possibly existing inode
972                  * mknod symlink mkdir  : null -> new inode
973                  * unlink               : linked -> null
974                  */
975                 struct inode *dir = req->r_locked_dir;
976                 struct dentry *dn = req->r_dentry;
977                 bool have_dir_cap, have_lease;
978
979                 BUG_ON(!dn);
980                 BUG_ON(!dir);
981                 BUG_ON(dn->d_parent->d_inode != dir);
982                 BUG_ON(ceph_ino(dir) !=
983                        le64_to_cpu(rinfo->diri.in->ino));
984                 BUG_ON(ceph_snap(dir) !=
985                        le64_to_cpu(rinfo->diri.in->snapid));
986
987                 /* do we have a lease on the whole dir? */
988                 have_dir_cap =
989                         (le32_to_cpu(rinfo->diri.in->cap.caps) &
990                          CEPH_CAP_FILE_SHARED);
991
992                 /* do we have a dn lease? */
993                 have_lease = have_dir_cap ||
994                         (le16_to_cpu(rinfo->dlease->mask) &
995                          CEPH_LOCK_DN);
996
997                 if (!have_lease)
998                         dout("fill_trace  no dentry lease or dir cap\n");
999
1000                 /* rename? */
1001                 if (req->r_old_dentry && req->r_op == CEPH_MDS_OP_RENAME) {
1002                         dout(" src %p '%.*s' dst %p '%.*s'\n",
1003                              req->r_old_dentry,
1004                              req->r_old_dentry->d_name.len,
1005                              req->r_old_dentry->d_name.name,
1006                              dn, dn->d_name.len, dn->d_name.name);
1007                         dout("fill_trace doing d_move %p -> %p\n",
1008                              req->r_old_dentry, dn);
1009
1010                         /* d_move screws up d_subdirs order */
1011                         ceph_i_clear(dir, CEPH_I_COMPLETE);
1012
1013                         d_move(req->r_old_dentry, dn);
1014                         dout(" src %p '%.*s' dst %p '%.*s'\n",
1015                              req->r_old_dentry,
1016                              req->r_old_dentry->d_name.len,
1017                              req->r_old_dentry->d_name.name,
1018                              dn, dn->d_name.len, dn->d_name.name);
1019
1020                         /* ensure target dentry is invalidated, despite
1021                            rehashing bug in vfs_rename_dir */
1022                         ceph_invalidate_dentry_lease(dn);
1023
1024                         /* take overwritten dentry's readdir offset */
1025                         dout("dn %p gets %p offset %lld (old offset %lld)\n",
1026                              req->r_old_dentry, dn, ceph_dentry(dn)->offset,
1027                              ceph_dentry(req->r_old_dentry)->offset);
1028                         ceph_dentry(req->r_old_dentry)->offset =
1029                                 ceph_dentry(dn)->offset;
1030
1031                         dn = req->r_old_dentry;  /* use old_dentry */
1032                         in = dn->d_inode;
1033                 }
1034
1035                 /* null dentry? */
1036                 if (!rinfo->head->is_target) {
1037                         dout("fill_trace null dentry\n");
1038                         if (dn->d_inode) {
1039                                 dout("d_delete %p\n", dn);
1040                                 d_delete(dn);
1041                         } else {
1042                                 dout("d_instantiate %p NULL\n", dn);
1043                                 d_instantiate(dn, NULL);
1044                                 if (have_lease && d_unhashed(dn))
1045                                         d_rehash(dn);
1046                                 update_dentry_lease(dn, rinfo->dlease,
1047                                                     session,
1048                                                     req->r_request_started);
1049                         }
1050                         goto done;
1051                 }
1052
1053                 /* attach proper inode */
1054                 ininfo = rinfo->targeti.in;
1055                 vino.ino = le64_to_cpu(ininfo->ino);
1056                 vino.snap = le64_to_cpu(ininfo->snapid);
1057                 if (!dn->d_inode) {
1058                         in = ceph_get_inode(sb, vino);
1059                         if (IS_ERR(in)) {
1060                                 pr_err("fill_trace bad get_inode "
1061                                        "%llx.%llx\n", vino.ino, vino.snap);
1062                                 err = PTR_ERR(in);
1063                                 d_delete(dn);
1064                                 goto done;
1065                         }
1066                         dn = splice_dentry(dn, in, &have_lease, true);
1067                         if (IS_ERR(dn)) {
1068                                 err = PTR_ERR(dn);
1069                                 goto done;
1070                         }
1071                         req->r_dentry = dn;  /* may have spliced */
1072                         igrab(in);
1073                 } else if (ceph_ino(in) == vino.ino &&
1074                            ceph_snap(in) == vino.snap) {
1075                         igrab(in);
1076                 } else {
1077                         dout(" %p links to %p %llx.%llx, not %llx.%llx\n",
1078                              dn, in, ceph_ino(in), ceph_snap(in),
1079                              vino.ino, vino.snap);
1080                         have_lease = false;
1081                         in = NULL;
1082                 }
1083
1084                 if (have_lease)
1085                         update_dentry_lease(dn, rinfo->dlease, session,
1086                                             req->r_request_started);
1087                 dout(" final dn %p\n", dn);
1088                 i++;
1089         } else if (req->r_op == CEPH_MDS_OP_LOOKUPSNAP ||
1090                    req->r_op == CEPH_MDS_OP_MKSNAP) {
1091                 struct dentry *dn = req->r_dentry;
1092
1093                 /* fill out a snapdir LOOKUPSNAP dentry */
1094                 BUG_ON(!dn);
1095                 BUG_ON(!req->r_locked_dir);
1096                 BUG_ON(ceph_snap(req->r_locked_dir) != CEPH_SNAPDIR);
1097                 ininfo = rinfo->targeti.in;
1098                 vino.ino = le64_to_cpu(ininfo->ino);
1099                 vino.snap = le64_to_cpu(ininfo->snapid);
1100                 in = ceph_get_inode(sb, vino);
1101                 if (IS_ERR(in)) {
1102                         pr_err("fill_inode get_inode badness %llx.%llx\n",
1103                                vino.ino, vino.snap);
1104                         err = PTR_ERR(in);
1105                         d_delete(dn);
1106                         goto done;
1107                 }
1108                 dout(" linking snapped dir %p to dn %p\n", in, dn);
1109                 dn = splice_dentry(dn, in, NULL, true);
1110                 if (IS_ERR(dn)) {
1111                         err = PTR_ERR(dn);
1112                         goto done;
1113                 }
1114                 req->r_dentry = dn;  /* may have spliced */
1115                 igrab(in);
1116                 rinfo->head->is_dentry = 1;  /* fool notrace handlers */
1117         }
1118
1119         if (rinfo->head->is_target) {
1120                 vino.ino = le64_to_cpu(rinfo->targeti.in->ino);
1121                 vino.snap = le64_to_cpu(rinfo->targeti.in->snapid);
1122
1123                 if (in == NULL || ceph_ino(in) != vino.ino ||
1124                     ceph_snap(in) != vino.snap) {
1125                         in = ceph_get_inode(sb, vino);
1126                         if (IS_ERR(in)) {
1127                                 err = PTR_ERR(in);
1128                                 goto done;
1129                         }
1130                 }
1131                 req->r_target_inode = in;
1132
1133                 err = fill_inode(in,
1134                                  &rinfo->targeti, NULL,
1135                                  session, req->r_request_started,
1136                                  (le32_to_cpu(rinfo->head->result) == 0) ?
1137                                  req->r_fmode : -1,
1138                                  &req->r_caps_reservation);
1139                 if (err < 0) {
1140                         pr_err("fill_inode badness %p %llx.%llx\n",
1141                                in, ceph_vinop(in));
1142                         goto done;
1143                 }
1144         }
1145
1146 done:
1147         dout("fill_trace done err=%d\n", err);
1148         return err;
1149 }
1150
1151 /*
1152  * Prepopulate our cache with readdir results, leases, etc.
1153  */
1154 int ceph_readdir_prepopulate(struct ceph_mds_request *req,
1155                              struct ceph_mds_session *session)
1156 {
1157         struct dentry *parent = req->r_dentry;
1158         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1159         struct qstr dname;
1160         struct dentry *dn;
1161         struct inode *in;
1162         int err = 0, i;
1163         struct inode *snapdir = NULL;
1164         struct ceph_mds_request_head *rhead = req->r_request->front.iov_base;
1165         u64 frag = le32_to_cpu(rhead->args.readdir.frag);
1166         struct ceph_dentry_info *di;
1167
1168         if (le32_to_cpu(rinfo->head->op) == CEPH_MDS_OP_LSSNAP) {
1169                 snapdir = ceph_get_snapdir(parent->d_inode);
1170                 parent = d_find_alias(snapdir);
1171                 dout("readdir_prepopulate %d items under SNAPDIR dn %p\n",
1172                      rinfo->dir_nr, parent);
1173         } else {
1174                 dout("readdir_prepopulate %d items under dn %p\n",
1175                      rinfo->dir_nr, parent);
1176                 if (rinfo->dir_dir)
1177                         ceph_fill_dirfrag(parent->d_inode, rinfo->dir_dir);
1178         }
1179
1180         for (i = 0; i < rinfo->dir_nr; i++) {
1181                 struct ceph_vino vino;
1182
1183                 dname.name = rinfo->dir_dname[i];
1184                 dname.len = rinfo->dir_dname_len[i];
1185                 dname.hash = full_name_hash(dname.name, dname.len);
1186
1187                 vino.ino = le64_to_cpu(rinfo->dir_in[i].in->ino);
1188                 vino.snap = le64_to_cpu(rinfo->dir_in[i].in->snapid);
1189
1190 retry_lookup:
1191                 dn = d_lookup(parent, &dname);
1192                 dout("d_lookup on parent=%p name=%.*s got %p\n",
1193                      parent, dname.len, dname.name, dn);
1194
1195                 if (!dn) {
1196                         dn = d_alloc(parent, &dname);
1197                         dout("d_alloc %p '%.*s' = %p\n", parent,
1198                              dname.len, dname.name, dn);
1199                         if (dn == NULL) {
1200                                 dout("d_alloc badness\n");
1201                                 err = -ENOMEM;
1202                                 goto out;
1203                         }
1204                         err = ceph_init_dentry(dn);
1205                         if (err < 0) {
1206                                 dput(dn);
1207                                 goto out;
1208                         }
1209                 } else if (dn->d_inode &&
1210                            (ceph_ino(dn->d_inode) != vino.ino ||
1211                             ceph_snap(dn->d_inode) != vino.snap)) {
1212                         dout(" dn %p points to wrong inode %p\n",
1213                              dn, dn->d_inode);
1214                         d_delete(dn);
1215                         dput(dn);
1216                         goto retry_lookup;
1217                 } else {
1218                         /* reorder parent's d_subdirs */
1219                         spin_lock(&dcache_lock);
1220                         spin_lock(&dn->d_lock);
1221                         list_move(&dn->d_u.d_child, &parent->d_subdirs);
1222                         spin_unlock(&dn->d_lock);
1223                         spin_unlock(&dcache_lock);
1224                 }
1225
1226                 di = dn->d_fsdata;
1227                 di->offset = ceph_make_fpos(frag, i + req->r_readdir_offset);
1228
1229                 /* inode */
1230                 if (dn->d_inode) {
1231                         in = dn->d_inode;
1232                 } else {
1233                         in = ceph_get_inode(parent->d_sb, vino);
1234                         if (IS_ERR(in)) {
1235                                 dout("new_inode badness\n");
1236                                 d_delete(dn);
1237                                 dput(dn);
1238                                 err = PTR_ERR(in);
1239                                 goto out;
1240                         }
1241                         dn = splice_dentry(dn, in, NULL, false);
1242                         if (IS_ERR(dn))
1243                                 dn = NULL;
1244                 }
1245
1246                 if (fill_inode(in, &rinfo->dir_in[i], NULL, session,
1247                                req->r_request_started, -1,
1248                                &req->r_caps_reservation) < 0) {
1249                         pr_err("fill_inode badness on %p\n", in);
1250                         goto next_item;
1251                 }
1252                 if (dn)
1253                         update_dentry_lease(dn, rinfo->dir_dlease[i],
1254                                             req->r_session,
1255                                             req->r_request_started);
1256 next_item:
1257                 if (dn)
1258                         dput(dn);
1259         }
1260         req->r_did_prepopulate = true;
1261
1262 out:
1263         if (snapdir) {
1264                 iput(snapdir);
1265                 dput(parent);
1266         }
1267         dout("readdir_prepopulate done\n");
1268         return err;
1269 }
1270
1271 int ceph_inode_set_size(struct inode *inode, loff_t size)
1272 {
1273         struct ceph_inode_info *ci = ceph_inode(inode);
1274         int ret = 0;
1275
1276         spin_lock(&inode->i_lock);
1277         dout("set_size %p %llu -> %llu\n", inode, inode->i_size, size);
1278         inode->i_size = size;
1279         inode->i_blocks = (size + (1 << 9) - 1) >> 9;
1280
1281         /* tell the MDS if we are approaching max_size */
1282         if ((size << 1) >= ci->i_max_size &&
1283             (ci->i_reported_size << 1) < ci->i_max_size)
1284                 ret = 1;
1285
1286         spin_unlock(&inode->i_lock);
1287         return ret;
1288 }
1289
1290 /*
1291  * Write back inode data in a worker thread.  (This can't be done
1292  * in the message handler context.)
1293  */
1294 void ceph_queue_writeback(struct inode *inode)
1295 {
1296         if (queue_work(ceph_inode_to_client(inode)->wb_wq,
1297                        &ceph_inode(inode)->i_wb_work)) {
1298                 dout("ceph_queue_writeback %p\n", inode);
1299                 igrab(inode);
1300         } else {
1301                 dout("ceph_queue_writeback %p failed\n", inode);
1302         }
1303 }
1304
1305 static void ceph_writeback_work(struct work_struct *work)
1306 {
1307         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1308                                                   i_wb_work);
1309         struct inode *inode = &ci->vfs_inode;
1310
1311         dout("writeback %p\n", inode);
1312         filemap_fdatawrite(&inode->i_data);
1313         iput(inode);
1314 }
1315
1316 /*
1317  * queue an async invalidation
1318  */
1319 void ceph_queue_invalidate(struct inode *inode)
1320 {
1321         if (queue_work(ceph_inode_to_client(inode)->pg_inv_wq,
1322                        &ceph_inode(inode)->i_pg_inv_work)) {
1323                 dout("ceph_queue_invalidate %p\n", inode);
1324                 igrab(inode);
1325         } else {
1326                 dout("ceph_queue_invalidate %p failed\n", inode);
1327         }
1328 }
1329
1330 /*
1331  * invalidate any pages that are not dirty or under writeback.  this
1332  * includes pages that are clean and mapped.
1333  */
1334 static void ceph_invalidate_nondirty_pages(struct address_space *mapping)
1335 {
1336         struct pagevec pvec;
1337         pgoff_t next = 0;
1338         int i;
1339
1340         pagevec_init(&pvec, 0);
1341         while (pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
1342                 for (i = 0; i < pagevec_count(&pvec); i++) {
1343                         struct page *page = pvec.pages[i];
1344                         pgoff_t index;
1345                         int skip_page =
1346                                 (PageDirty(page) || PageWriteback(page));
1347
1348                         if (!skip_page)
1349                                 skip_page = !trylock_page(page);
1350
1351                         /*
1352                          * We really shouldn't be looking at the ->index of an
1353                          * unlocked page.  But we're not allowed to lock these
1354                          * pages.  So we rely upon nobody altering the ->index
1355                          * of this (pinned-by-us) page.
1356                          */
1357                         index = page->index;
1358                         if (index > next)
1359                                 next = index;
1360                         next++;
1361
1362                         if (skip_page)
1363                                 continue;
1364
1365                         generic_error_remove_page(mapping, page);
1366                         unlock_page(page);
1367                 }
1368                 pagevec_release(&pvec);
1369                 cond_resched();
1370         }
1371 }
1372
1373 /*
1374  * Invalidate inode pages in a worker thread.  (This can't be done
1375  * in the message handler context.)
1376  */
1377 static void ceph_invalidate_work(struct work_struct *work)
1378 {
1379         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1380                                                   i_pg_inv_work);
1381         struct inode *inode = &ci->vfs_inode;
1382         u32 orig_gen;
1383         int check = 0;
1384
1385         spin_lock(&inode->i_lock);
1386         dout("invalidate_pages %p gen %d revoking %d\n", inode,
1387              ci->i_rdcache_gen, ci->i_rdcache_revoking);
1388         if (ci->i_rdcache_gen == 0 ||
1389             ci->i_rdcache_revoking != ci->i_rdcache_gen) {
1390                 BUG_ON(ci->i_rdcache_revoking > ci->i_rdcache_gen);
1391                 /* nevermind! */
1392                 ci->i_rdcache_revoking = 0;
1393                 spin_unlock(&inode->i_lock);
1394                 goto out;
1395         }
1396         orig_gen = ci->i_rdcache_gen;
1397         spin_unlock(&inode->i_lock);
1398
1399         ceph_invalidate_nondirty_pages(inode->i_mapping);
1400
1401         spin_lock(&inode->i_lock);
1402         if (orig_gen == ci->i_rdcache_gen) {
1403                 dout("invalidate_pages %p gen %d successful\n", inode,
1404                      ci->i_rdcache_gen);
1405                 ci->i_rdcache_gen = 0;
1406                 ci->i_rdcache_revoking = 0;
1407                 check = 1;
1408         } else {
1409                 dout("invalidate_pages %p gen %d raced, gen now %d\n",
1410                      inode, orig_gen, ci->i_rdcache_gen);
1411         }
1412         spin_unlock(&inode->i_lock);
1413
1414         if (check)
1415                 ceph_check_caps(ci, 0, NULL);
1416 out:
1417         iput(inode);
1418 }
1419
1420
1421 /*
1422  * called by trunc_wq; take i_mutex ourselves
1423  *
1424  * We also truncate in a separate thread as well.
1425  */
1426 static void ceph_vmtruncate_work(struct work_struct *work)
1427 {
1428         struct ceph_inode_info *ci = container_of(work, struct ceph_inode_info,
1429                                                   i_vmtruncate_work);
1430         struct inode *inode = &ci->vfs_inode;
1431
1432         dout("vmtruncate_work %p\n", inode);
1433         mutex_lock(&inode->i_mutex);
1434         __ceph_do_pending_vmtruncate(inode);
1435         mutex_unlock(&inode->i_mutex);
1436         iput(inode);
1437 }
1438
1439 /*
1440  * Queue an async vmtruncate.  If we fail to queue work, we will handle
1441  * the truncation the next time we call __ceph_do_pending_vmtruncate.
1442  */
1443 void ceph_queue_vmtruncate(struct inode *inode)
1444 {
1445         struct ceph_inode_info *ci = ceph_inode(inode);
1446
1447         if (queue_work(ceph_sb_to_client(inode->i_sb)->trunc_wq,
1448                        &ci->i_vmtruncate_work)) {
1449                 dout("ceph_queue_vmtruncate %p\n", inode);
1450                 igrab(inode);
1451         } else {
1452                 dout("ceph_queue_vmtruncate %p failed, pending=%d\n",
1453                      inode, ci->i_truncate_pending);
1454         }
1455 }
1456
1457 /*
1458  * called with i_mutex held.
1459  *
1460  * Make sure any pending truncation is applied before doing anything
1461  * that may depend on it.
1462  */
1463 void __ceph_do_pending_vmtruncate(struct inode *inode)
1464 {
1465         struct ceph_inode_info *ci = ceph_inode(inode);
1466         u64 to;
1467         int wrbuffer_refs, wake = 0;
1468
1469 retry:
1470         spin_lock(&inode->i_lock);
1471         if (ci->i_truncate_pending == 0) {
1472                 dout("__do_pending_vmtruncate %p none pending\n", inode);
1473                 spin_unlock(&inode->i_lock);
1474                 return;
1475         }
1476
1477         /*
1478          * make sure any dirty snapped pages are flushed before we
1479          * possibly truncate them.. so write AND block!
1480          */
1481         if (ci->i_wrbuffer_ref_head < ci->i_wrbuffer_ref) {
1482                 dout("__do_pending_vmtruncate %p flushing snaps first\n",
1483                      inode);
1484                 spin_unlock(&inode->i_lock);
1485                 filemap_write_and_wait_range(&inode->i_data, 0,
1486                                              inode->i_sb->s_maxbytes);
1487                 goto retry;
1488         }
1489
1490         to = ci->i_truncate_size;
1491         wrbuffer_refs = ci->i_wrbuffer_ref;
1492         dout("__do_pending_vmtruncate %p (%d) to %lld\n", inode,
1493              ci->i_truncate_pending, to);
1494         spin_unlock(&inode->i_lock);
1495
1496         truncate_inode_pages(inode->i_mapping, to);
1497
1498         spin_lock(&inode->i_lock);
1499         ci->i_truncate_pending--;
1500         if (ci->i_truncate_pending == 0)
1501                 wake = 1;
1502         spin_unlock(&inode->i_lock);
1503
1504         if (wrbuffer_refs == 0)
1505                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1506         if (wake)
1507                 wake_up_all(&ci->i_cap_wq);
1508 }
1509
1510
1511 /*
1512  * symlinks
1513  */
1514 static void *ceph_sym_follow_link(struct dentry *dentry, struct nameidata *nd)
1515 {
1516         struct ceph_inode_info *ci = ceph_inode(dentry->d_inode);
1517         nd_set_link(nd, ci->i_symlink);
1518         return NULL;
1519 }
1520
1521 static const struct inode_operations ceph_symlink_iops = {
1522         .readlink = generic_readlink,
1523         .follow_link = ceph_sym_follow_link,
1524 };
1525
1526 /*
1527  * setattr
1528  */
1529 int ceph_setattr(struct dentry *dentry, struct iattr *attr)
1530 {
1531         struct inode *inode = dentry->d_inode;
1532         struct ceph_inode_info *ci = ceph_inode(inode);
1533         struct inode *parent_inode = dentry->d_parent->d_inode;
1534         const unsigned int ia_valid = attr->ia_valid;
1535         struct ceph_mds_request *req;
1536         struct ceph_mds_client *mdsc = ceph_sb_to_client(dentry->d_sb)->mdsc;
1537         int issued;
1538         int release = 0, dirtied = 0;
1539         int mask = 0;
1540         int err = 0;
1541
1542         if (ceph_snap(inode) != CEPH_NOSNAP)
1543                 return -EROFS;
1544
1545         __ceph_do_pending_vmtruncate(inode);
1546
1547         err = inode_change_ok(inode, attr);
1548         if (err != 0)
1549                 return err;
1550
1551         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETATTR,
1552                                        USE_AUTH_MDS);
1553         if (IS_ERR(req))
1554                 return PTR_ERR(req);
1555
1556         spin_lock(&inode->i_lock);
1557         issued = __ceph_caps_issued(ci, NULL);
1558         dout("setattr %p issued %s\n", inode, ceph_cap_string(issued));
1559
1560         if (ia_valid & ATTR_UID) {
1561                 dout("setattr %p uid %d -> %d\n", inode,
1562                      inode->i_uid, attr->ia_uid);
1563                 if (issued & CEPH_CAP_AUTH_EXCL) {
1564                         inode->i_uid = attr->ia_uid;
1565                         dirtied |= CEPH_CAP_AUTH_EXCL;
1566                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
1567                            attr->ia_uid != inode->i_uid) {
1568                         req->r_args.setattr.uid = cpu_to_le32(attr->ia_uid);
1569                         mask |= CEPH_SETATTR_UID;
1570                         release |= CEPH_CAP_AUTH_SHARED;
1571                 }
1572         }
1573         if (ia_valid & ATTR_GID) {
1574                 dout("setattr %p gid %d -> %d\n", inode,
1575                      inode->i_gid, attr->ia_gid);
1576                 if (issued & CEPH_CAP_AUTH_EXCL) {
1577                         inode->i_gid = attr->ia_gid;
1578                         dirtied |= CEPH_CAP_AUTH_EXCL;
1579                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
1580                            attr->ia_gid != inode->i_gid) {
1581                         req->r_args.setattr.gid = cpu_to_le32(attr->ia_gid);
1582                         mask |= CEPH_SETATTR_GID;
1583                         release |= CEPH_CAP_AUTH_SHARED;
1584                 }
1585         }
1586         if (ia_valid & ATTR_MODE) {
1587                 dout("setattr %p mode 0%o -> 0%o\n", inode, inode->i_mode,
1588                      attr->ia_mode);
1589                 if (issued & CEPH_CAP_AUTH_EXCL) {
1590                         inode->i_mode = attr->ia_mode;
1591                         dirtied |= CEPH_CAP_AUTH_EXCL;
1592                 } else if ((issued & CEPH_CAP_AUTH_SHARED) == 0 ||
1593                            attr->ia_mode != inode->i_mode) {
1594                         req->r_args.setattr.mode = cpu_to_le32(attr->ia_mode);
1595                         mask |= CEPH_SETATTR_MODE;
1596                         release |= CEPH_CAP_AUTH_SHARED;
1597                 }
1598         }
1599
1600         if (ia_valid & ATTR_ATIME) {
1601                 dout("setattr %p atime %ld.%ld -> %ld.%ld\n", inode,
1602                      inode->i_atime.tv_sec, inode->i_atime.tv_nsec,
1603                      attr->ia_atime.tv_sec, attr->ia_atime.tv_nsec);
1604                 if (issued & CEPH_CAP_FILE_EXCL) {
1605                         ci->i_time_warp_seq++;
1606                         inode->i_atime = attr->ia_atime;
1607                         dirtied |= CEPH_CAP_FILE_EXCL;
1608                 } else if ((issued & CEPH_CAP_FILE_WR) &&
1609                            timespec_compare(&inode->i_atime,
1610                                             &attr->ia_atime) < 0) {
1611                         inode->i_atime = attr->ia_atime;
1612                         dirtied |= CEPH_CAP_FILE_WR;
1613                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
1614                            !timespec_equal(&inode->i_atime, &attr->ia_atime)) {
1615                         ceph_encode_timespec(&req->r_args.setattr.atime,
1616                                              &attr->ia_atime);
1617                         mask |= CEPH_SETATTR_ATIME;
1618                         release |= CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_RD |
1619                                 CEPH_CAP_FILE_WR;
1620                 }
1621         }
1622         if (ia_valid & ATTR_MTIME) {
1623                 dout("setattr %p mtime %ld.%ld -> %ld.%ld\n", inode,
1624                      inode->i_mtime.tv_sec, inode->i_mtime.tv_nsec,
1625                      attr->ia_mtime.tv_sec, attr->ia_mtime.tv_nsec);
1626                 if (issued & CEPH_CAP_FILE_EXCL) {
1627                         ci->i_time_warp_seq++;
1628                         inode->i_mtime = attr->ia_mtime;
1629                         dirtied |= CEPH_CAP_FILE_EXCL;
1630                 } else if ((issued & CEPH_CAP_FILE_WR) &&
1631                            timespec_compare(&inode->i_mtime,
1632                                             &attr->ia_mtime) < 0) {
1633                         inode->i_mtime = attr->ia_mtime;
1634                         dirtied |= CEPH_CAP_FILE_WR;
1635                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
1636                            !timespec_equal(&inode->i_mtime, &attr->ia_mtime)) {
1637                         ceph_encode_timespec(&req->r_args.setattr.mtime,
1638                                              &attr->ia_mtime);
1639                         mask |= CEPH_SETATTR_MTIME;
1640                         release |= CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_RD |
1641                                 CEPH_CAP_FILE_WR;
1642                 }
1643         }
1644         if (ia_valid & ATTR_SIZE) {
1645                 dout("setattr %p size %lld -> %lld\n", inode,
1646                      inode->i_size, attr->ia_size);
1647                 if (attr->ia_size > inode->i_sb->s_maxbytes) {
1648                         err = -EINVAL;
1649                         goto out;
1650                 }
1651                 if ((issued & CEPH_CAP_FILE_EXCL) &&
1652                     attr->ia_size > inode->i_size) {
1653                         inode->i_size = attr->ia_size;
1654                         inode->i_blocks =
1655                                 (attr->ia_size + (1 << 9) - 1) >> 9;
1656                         inode->i_ctime = attr->ia_ctime;
1657                         ci->i_reported_size = attr->ia_size;
1658                         dirtied |= CEPH_CAP_FILE_EXCL;
1659                 } else if ((issued & CEPH_CAP_FILE_SHARED) == 0 ||
1660                            attr->ia_size != inode->i_size) {
1661                         req->r_args.setattr.size = cpu_to_le64(attr->ia_size);
1662                         req->r_args.setattr.old_size =
1663                                 cpu_to_le64(inode->i_size);
1664                         mask |= CEPH_SETATTR_SIZE;
1665                         release |= CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_RD |
1666                                 CEPH_CAP_FILE_WR;
1667                 }
1668         }
1669
1670         /* these do nothing */
1671         if (ia_valid & ATTR_CTIME) {
1672                 bool only = (ia_valid & (ATTR_SIZE|ATTR_MTIME|ATTR_ATIME|
1673                                          ATTR_MODE|ATTR_UID|ATTR_GID)) == 0;
1674                 dout("setattr %p ctime %ld.%ld -> %ld.%ld (%s)\n", inode,
1675                      inode->i_ctime.tv_sec, inode->i_ctime.tv_nsec,
1676                      attr->ia_ctime.tv_sec, attr->ia_ctime.tv_nsec,
1677                      only ? "ctime only" : "ignored");
1678                 inode->i_ctime = attr->ia_ctime;
1679                 if (only) {
1680                         /*
1681                          * if kernel wants to dirty ctime but nothing else,
1682                          * we need to choose a cap to dirty under, or do
1683                          * a almost-no-op setattr
1684                          */
1685                         if (issued & CEPH_CAP_AUTH_EXCL)
1686                                 dirtied |= CEPH_CAP_AUTH_EXCL;
1687                         else if (issued & CEPH_CAP_FILE_EXCL)
1688                                 dirtied |= CEPH_CAP_FILE_EXCL;
1689                         else if (issued & CEPH_CAP_XATTR_EXCL)
1690                                 dirtied |= CEPH_CAP_XATTR_EXCL;
1691                         else
1692                                 mask |= CEPH_SETATTR_CTIME;
1693                 }
1694         }
1695         if (ia_valid & ATTR_FILE)
1696                 dout("setattr %p ATTR_FILE ... hrm!\n", inode);
1697
1698         if (dirtied) {
1699                 __ceph_mark_dirty_caps(ci, dirtied);
1700                 inode->i_ctime = CURRENT_TIME;
1701         }
1702
1703         release &= issued;
1704         spin_unlock(&inode->i_lock);
1705
1706         if (mask) {
1707                 req->r_inode = igrab(inode);
1708                 req->r_inode_drop = release;
1709                 req->r_args.setattr.mask = cpu_to_le32(mask);
1710                 req->r_num_caps = 1;
1711                 err = ceph_mdsc_do_request(mdsc, parent_inode, req);
1712         }
1713         dout("setattr %p result=%d (%s locally, %d remote)\n", inode, err,
1714              ceph_cap_string(dirtied), mask);
1715
1716         ceph_mdsc_put_request(req);
1717         __ceph_do_pending_vmtruncate(inode);
1718         return err;
1719 out:
1720         spin_unlock(&inode->i_lock);
1721         ceph_mdsc_put_request(req);
1722         return err;
1723 }
1724
1725 /*
1726  * Verify that we have a lease on the given mask.  If not,
1727  * do a getattr against an mds.
1728  */
1729 int ceph_do_getattr(struct inode *inode, int mask)
1730 {
1731         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
1732         struct ceph_mds_client *mdsc = fsc->mdsc;
1733         struct ceph_mds_request *req;
1734         int err;
1735
1736         if (ceph_snap(inode) == CEPH_SNAPDIR) {
1737                 dout("do_getattr inode %p SNAPDIR\n", inode);
1738                 return 0;
1739         }
1740
1741         dout("do_getattr inode %p mask %s\n", inode, ceph_cap_string(mask));
1742         if (ceph_caps_issued_mask(ceph_inode(inode), mask, 1))
1743                 return 0;
1744
1745         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, USE_ANY_MDS);
1746         if (IS_ERR(req))
1747                 return PTR_ERR(req);
1748         req->r_inode = igrab(inode);
1749         req->r_num_caps = 1;
1750         req->r_args.getattr.mask = cpu_to_le32(mask);
1751         err = ceph_mdsc_do_request(mdsc, NULL, req);
1752         ceph_mdsc_put_request(req);
1753         dout("do_getattr result=%d\n", err);
1754         return err;
1755 }
1756
1757
1758 /*
1759  * Check inode permissions.  We verify we have a valid value for
1760  * the AUTH cap, then call the generic handler.
1761  */
1762 int ceph_permission(struct inode *inode, int mask)
1763 {
1764         int err = ceph_do_getattr(inode, CEPH_CAP_AUTH_SHARED);
1765
1766         if (!err)
1767                 err = generic_permission(inode, mask, NULL);
1768         return err;
1769 }
1770
1771 /*
1772  * Get all attributes.  Hopefully somedata we'll have a statlite()
1773  * and can limit the fields we require to be accurate.
1774  */
1775 int ceph_getattr(struct vfsmount *mnt, struct dentry *dentry,
1776                  struct kstat *stat)
1777 {
1778         struct inode *inode = dentry->d_inode;
1779         struct ceph_inode_info *ci = ceph_inode(inode);
1780         int err;
1781
1782         err = ceph_do_getattr(inode, CEPH_STAT_CAP_INODE_ALL);
1783         if (!err) {
1784                 generic_fillattr(inode, stat);
1785                 stat->ino = inode->i_ino;
1786                 if (ceph_snap(inode) != CEPH_NOSNAP)
1787                         stat->dev = ceph_snap(inode);
1788                 else
1789                         stat->dev = 0;
1790                 if (S_ISDIR(inode->i_mode)) {
1791                         stat->size = ci->i_rbytes;
1792                         stat->blocks = 0;
1793                         stat->blksize = 65536;
1794                 }
1795         }
1796         return err;
1797 }