xref: /openbmc/linux/fs/ceph/dir.c (revision d95852777bc8ba6b3ad3397d495c5f9dd8ca8383)
1 #include <linux/ceph/ceph_debug.h>
2 
3 #include <linux/spinlock.h>
4 #include <linux/fs_struct.h>
5 #include <linux/namei.h>
6 #include <linux/slab.h>
7 #include <linux/sched.h>
8 
9 #include "super.h"
10 #include "mds_client.h"
11 
12 /*
13  * Directory operations: readdir, lookup, create, link, unlink,
14  * rename, etc.
15  */
16 
17 /*
18  * Ceph MDS operations are specified in terms of a base ino and
19  * relative path.  Thus, the client can specify an operation on a
20  * specific inode (e.g., a getattr due to fstat(2)), or as a path
21  * relative to, say, the root directory.
22  *
23  * Normally, we limit ourselves to strict inode ops (no path component)
24  * or dentry operations (a single path component relative to an ino).  The
25  * exception to this is open_root_dentry(), which will open the mount
26  * point by name.
27  */
28 
29 const struct inode_operations ceph_dir_iops;
30 const struct file_operations ceph_dir_fops;
31 const struct dentry_operations ceph_dentry_ops;
32 
33 /*
34  * Initialize ceph dentry state.
35  */
36 int ceph_init_dentry(struct dentry *dentry)
37 {
38 	struct ceph_dentry_info *di;
39 
40 	if (dentry->d_fsdata)
41 		return 0;
42 
43 	di = kmem_cache_alloc(ceph_dentry_cachep, GFP_NOFS | __GFP_ZERO);
44 	if (!di)
45 		return -ENOMEM;          /* oh well */
46 
47 	spin_lock(&dentry->d_lock);
48 	if (dentry->d_fsdata) {
49 		/* lost a race */
50 		kmem_cache_free(ceph_dentry_cachep, di);
51 		goto out_unlock;
52 	}
53 
54 	if (dentry->d_parent == NULL ||   /* nfs fh_to_dentry */
55 	    ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP)
56 		d_set_d_op(dentry, &ceph_dentry_ops);
57 	else if (ceph_snap(dentry->d_parent->d_inode) == CEPH_SNAPDIR)
58 		d_set_d_op(dentry, &ceph_snapdir_dentry_ops);
59 	else
60 		d_set_d_op(dentry, &ceph_snap_dentry_ops);
61 
62 	di->dentry = dentry;
63 	di->lease_session = NULL;
64 	dentry->d_time = jiffies;
65 	/* avoid reordering d_fsdata setup so that the check above is safe */
66 	smp_mb();
67 	dentry->d_fsdata = di;
68 	ceph_dentry_lru_add(dentry);
69 out_unlock:
70 	spin_unlock(&dentry->d_lock);
71 	return 0;
72 }
73 
74 struct inode *ceph_get_dentry_parent_inode(struct dentry *dentry)
75 {
76 	struct inode *inode = NULL;
77 
78 	if (!dentry)
79 		return NULL;
80 
81 	spin_lock(&dentry->d_lock);
82 	if (dentry->d_parent) {
83 		inode = dentry->d_parent->d_inode;
84 		ihold(inode);
85 	}
86 	spin_unlock(&dentry->d_lock);
87 	return inode;
88 }
89 
90 
91 /*
92  * for readdir, we encode the directory frag and offset within that
93  * frag into f_pos.
94  */
95 static unsigned fpos_frag(loff_t p)
96 {
97 	return p >> 32;
98 }
99 static unsigned fpos_off(loff_t p)
100 {
101 	return p & 0xffffffff;
102 }
103 
104 /*
105  * When possible, we try to satisfy a readdir by peeking at the
106  * dcache.  We make this work by carefully ordering dentries on
107  * d_u.d_child when we initially get results back from the MDS, and
108  * falling back to a "normal" sync readdir if any dentries in the dir
109  * are dropped.
110  *
111  * D_COMPLETE tells indicates we have all dentries in the dir.  It is
112  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
113  * the MDS if/when the directory is modified).
114  */
115 static int __dcache_readdir(struct file *filp,
116 			    void *dirent, filldir_t filldir)
117 {
118 	struct ceph_file_info *fi = filp->private_data;
119 	struct dentry *parent = filp->f_dentry;
120 	struct inode *dir = parent->d_inode;
121 	struct list_head *p;
122 	struct dentry *dentry, *last;
123 	struct ceph_dentry_info *di;
124 	int err = 0;
125 
126 	/* claim ref on last dentry we returned */
127 	last = fi->dentry;
128 	fi->dentry = NULL;
129 
130 	dout("__dcache_readdir %p at %llu (last %p)\n", dir, filp->f_pos,
131 	     last);
132 
133 	spin_lock(&parent->d_lock);
134 
135 	/* start at beginning? */
136 	if (filp->f_pos == 2 || last == NULL ||
137 	    filp->f_pos < ceph_dentry(last)->offset) {
138 		if (list_empty(&parent->d_subdirs))
139 			goto out_unlock;
140 		p = parent->d_subdirs.prev;
141 		dout(" initial p %p/%p\n", p->prev, p->next);
142 	} else {
143 		p = last->d_u.d_child.prev;
144 	}
145 
146 more:
147 	dentry = list_entry(p, struct dentry, d_u.d_child);
148 	di = ceph_dentry(dentry);
149 	while (1) {
150 		dout(" p %p/%p %s d_subdirs %p/%p\n", p->prev, p->next,
151 		     d_unhashed(dentry) ? "!hashed" : "hashed",
152 		     parent->d_subdirs.prev, parent->d_subdirs.next);
153 		if (p == &parent->d_subdirs) {
154 			fi->flags |= CEPH_F_ATEND;
155 			goto out_unlock;
156 		}
157 		spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
158 		if (!d_unhashed(dentry) && dentry->d_inode &&
159 		    ceph_snap(dentry->d_inode) != CEPH_SNAPDIR &&
160 		    ceph_ino(dentry->d_inode) != CEPH_INO_CEPH &&
161 		    filp->f_pos <= di->offset)
162 			break;
163 		dout(" skipping %p %.*s at %llu (%llu)%s%s\n", dentry,
164 		     dentry->d_name.len, dentry->d_name.name, di->offset,
165 		     filp->f_pos, d_unhashed(dentry) ? " unhashed" : "",
166 		     !dentry->d_inode ? " null" : "");
167 		spin_unlock(&dentry->d_lock);
168 		p = p->prev;
169 		dentry = list_entry(p, struct dentry, d_u.d_child);
170 		di = ceph_dentry(dentry);
171 	}
172 
173 	dget_dlock(dentry);
174 	spin_unlock(&dentry->d_lock);
175 	spin_unlock(&parent->d_lock);
176 
177 	dout(" %llu (%llu) dentry %p %.*s %p\n", di->offset, filp->f_pos,
178 	     dentry, dentry->d_name.len, dentry->d_name.name, dentry->d_inode);
179 	filp->f_pos = di->offset;
180 	err = filldir(dirent, dentry->d_name.name,
181 		      dentry->d_name.len, di->offset,
182 		      ceph_translate_ino(dentry->d_sb, dentry->d_inode->i_ino),
183 		      dentry->d_inode->i_mode >> 12);
184 
185 	if (last) {
186 		if (err < 0) {
187 			/* remember our position */
188 			fi->dentry = last;
189 			fi->next_offset = di->offset;
190 		} else {
191 			dput(last);
192 		}
193 	}
194 	last = dentry;
195 
196 	if (err < 0)
197 		goto out;
198 
199 	filp->f_pos++;
200 
201 	/* make sure a dentry wasn't dropped while we didn't have parent lock */
202 	if (!ceph_dir_test_complete(dir)) {
203 		dout(" lost D_COMPLETE on %p; falling back to mds\n", dir);
204 		err = -EAGAIN;
205 		goto out;
206 	}
207 
208 	spin_lock(&parent->d_lock);
209 	p = p->prev;	/* advance to next dentry */
210 	goto more;
211 
212 out_unlock:
213 	spin_unlock(&parent->d_lock);
214 out:
215 	if (last)
216 		dput(last);
217 	return err;
218 }
219 
220 /*
221  * make note of the last dentry we read, so we can
222  * continue at the same lexicographical point,
223  * regardless of what dir changes take place on the
224  * server.
225  */
226 static int note_last_dentry(struct ceph_file_info *fi, const char *name,
227 			    int len)
228 {
229 	kfree(fi->last_name);
230 	fi->last_name = kmalloc(len+1, GFP_NOFS);
231 	if (!fi->last_name)
232 		return -ENOMEM;
233 	memcpy(fi->last_name, name, len);
234 	fi->last_name[len] = 0;
235 	dout("note_last_dentry '%s'\n", fi->last_name);
236 	return 0;
237 }
238 
239 static int ceph_readdir(struct file *filp, void *dirent, filldir_t filldir)
240 {
241 	struct ceph_file_info *fi = filp->private_data;
242 	struct inode *inode = filp->f_dentry->d_inode;
243 	struct ceph_inode_info *ci = ceph_inode(inode);
244 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
245 	struct ceph_mds_client *mdsc = fsc->mdsc;
246 	unsigned frag = fpos_frag(filp->f_pos);
247 	int off = fpos_off(filp->f_pos);
248 	int err;
249 	u32 ftype;
250 	struct ceph_mds_reply_info_parsed *rinfo;
251 	const int max_entries = fsc->mount_options->max_readdir;
252 	const int max_bytes = fsc->mount_options->max_readdir_bytes;
253 
254 	dout("readdir %p filp %p frag %u off %u\n", inode, filp, frag, off);
255 	if (fi->flags & CEPH_F_ATEND)
256 		return 0;
257 
258 	/* always start with . and .. */
259 	if (filp->f_pos == 0) {
260 		/* note dir version at start of readdir so we can tell
261 		 * if any dentries get dropped */
262 		fi->dir_release_count = ci->i_release_count;
263 
264 		dout("readdir off 0 -> '.'\n");
265 		if (filldir(dirent, ".", 1, ceph_make_fpos(0, 0),
266 			    ceph_translate_ino(inode->i_sb, inode->i_ino),
267 			    inode->i_mode >> 12) < 0)
268 			return 0;
269 		filp->f_pos = 1;
270 		off = 1;
271 	}
272 	if (filp->f_pos == 1) {
273 		ino_t ino = parent_ino(filp->f_dentry);
274 		dout("readdir off 1 -> '..'\n");
275 		if (filldir(dirent, "..", 2, ceph_make_fpos(0, 1),
276 			    ceph_translate_ino(inode->i_sb, ino),
277 			    inode->i_mode >> 12) < 0)
278 			return 0;
279 		filp->f_pos = 2;
280 		off = 2;
281 	}
282 
283 	/* can we use the dcache? */
284 	spin_lock(&ci->i_ceph_lock);
285 	if ((filp->f_pos == 2 || fi->dentry) &&
286 	    !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
287 	    ceph_snap(inode) != CEPH_SNAPDIR &&
288 	    ceph_dir_test_complete(inode) &&
289 	    __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) {
290 		spin_unlock(&ci->i_ceph_lock);
291 		err = __dcache_readdir(filp, dirent, filldir);
292 		if (err != -EAGAIN)
293 			return err;
294 	} else {
295 		spin_unlock(&ci->i_ceph_lock);
296 	}
297 	if (fi->dentry) {
298 		err = note_last_dentry(fi, fi->dentry->d_name.name,
299 				       fi->dentry->d_name.len);
300 		if (err)
301 			return err;
302 		dput(fi->dentry);
303 		fi->dentry = NULL;
304 	}
305 
306 	/* proceed with a normal readdir */
307 
308 more:
309 	/* do we have the correct frag content buffered? */
310 	if (fi->frag != frag || fi->last_readdir == NULL) {
311 		struct ceph_mds_request *req;
312 		int op = ceph_snap(inode) == CEPH_SNAPDIR ?
313 			CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
314 
315 		/* discard old result, if any */
316 		if (fi->last_readdir) {
317 			ceph_mdsc_put_request(fi->last_readdir);
318 			fi->last_readdir = NULL;
319 		}
320 
321 		/* requery frag tree, as the frag topology may have changed */
322 		frag = ceph_choose_frag(ceph_inode(inode), frag, NULL, NULL);
323 
324 		dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
325 		     ceph_vinop(inode), frag, fi->last_name);
326 		req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
327 		if (IS_ERR(req))
328 			return PTR_ERR(req);
329 		req->r_inode = inode;
330 		ihold(inode);
331 		req->r_dentry = dget(filp->f_dentry);
332 		/* hints to request -> mds selection code */
333 		req->r_direct_mode = USE_AUTH_MDS;
334 		req->r_direct_hash = ceph_frag_value(frag);
335 		req->r_direct_is_hash = true;
336 		req->r_path2 = kstrdup(fi->last_name, GFP_NOFS);
337 		req->r_readdir_offset = fi->next_offset;
338 		req->r_args.readdir.frag = cpu_to_le32(frag);
339 		req->r_args.readdir.max_entries = cpu_to_le32(max_entries);
340 		req->r_args.readdir.max_bytes = cpu_to_le32(max_bytes);
341 		req->r_num_caps = max_entries + 1;
342 		err = ceph_mdsc_do_request(mdsc, NULL, req);
343 		if (err < 0) {
344 			ceph_mdsc_put_request(req);
345 			return err;
346 		}
347 		dout("readdir got and parsed readdir result=%d"
348 		     " on frag %x, end=%d, complete=%d\n", err, frag,
349 		     (int)req->r_reply_info.dir_end,
350 		     (int)req->r_reply_info.dir_complete);
351 
352 		if (!req->r_did_prepopulate) {
353 			dout("readdir !did_prepopulate");
354 			fi->dir_release_count--;    /* preclude D_COMPLETE */
355 		}
356 
357 		/* note next offset and last dentry name */
358 		fi->offset = fi->next_offset;
359 		fi->last_readdir = req;
360 
361 		if (req->r_reply_info.dir_end) {
362 			kfree(fi->last_name);
363 			fi->last_name = NULL;
364 			if (ceph_frag_is_rightmost(frag))
365 				fi->next_offset = 2;
366 			else
367 				fi->next_offset = 0;
368 		} else {
369 			rinfo = &req->r_reply_info;
370 			err = note_last_dentry(fi,
371 				       rinfo->dir_dname[rinfo->dir_nr-1],
372 				       rinfo->dir_dname_len[rinfo->dir_nr-1]);
373 			if (err)
374 				return err;
375 			fi->next_offset += rinfo->dir_nr;
376 		}
377 	}
378 
379 	rinfo = &fi->last_readdir->r_reply_info;
380 	dout("readdir frag %x num %d off %d chunkoff %d\n", frag,
381 	     rinfo->dir_nr, off, fi->offset);
382 	while (off >= fi->offset && off - fi->offset < rinfo->dir_nr) {
383 		u64 pos = ceph_make_fpos(frag, off);
384 		struct ceph_mds_reply_inode *in =
385 			rinfo->dir_in[off - fi->offset].in;
386 		struct ceph_vino vino;
387 		ino_t ino;
388 
389 		dout("readdir off %d (%d/%d) -> %lld '%.*s' %p\n",
390 		     off, off - fi->offset, rinfo->dir_nr, pos,
391 		     rinfo->dir_dname_len[off - fi->offset],
392 		     rinfo->dir_dname[off - fi->offset], in);
393 		BUG_ON(!in);
394 		ftype = le32_to_cpu(in->mode) >> 12;
395 		vino.ino = le64_to_cpu(in->ino);
396 		vino.snap = le64_to_cpu(in->snapid);
397 		ino = ceph_vino_to_ino(vino);
398 		if (filldir(dirent,
399 			    rinfo->dir_dname[off - fi->offset],
400 			    rinfo->dir_dname_len[off - fi->offset],
401 			    pos,
402 			    ceph_translate_ino(inode->i_sb, ino), ftype) < 0) {
403 			dout("filldir stopping us...\n");
404 			return 0;
405 		}
406 		off++;
407 		filp->f_pos = pos + 1;
408 	}
409 
410 	if (fi->last_name) {
411 		ceph_mdsc_put_request(fi->last_readdir);
412 		fi->last_readdir = NULL;
413 		goto more;
414 	}
415 
416 	/* more frags? */
417 	if (!ceph_frag_is_rightmost(frag)) {
418 		frag = ceph_frag_next(frag);
419 		off = 0;
420 		filp->f_pos = ceph_make_fpos(frag, off);
421 		dout("readdir next frag is %x\n", frag);
422 		goto more;
423 	}
424 	fi->flags |= CEPH_F_ATEND;
425 
426 	/*
427 	 * if dir_release_count still matches the dir, no dentries
428 	 * were released during the whole readdir, and we should have
429 	 * the complete dir contents in our cache.
430 	 */
431 	spin_lock(&ci->i_ceph_lock);
432 	if (ci->i_release_count == fi->dir_release_count) {
433 		ceph_dir_set_complete(inode);
434 		ci->i_max_offset = filp->f_pos;
435 	}
436 	spin_unlock(&ci->i_ceph_lock);
437 
438 	dout("readdir %p filp %p done.\n", inode, filp);
439 	return 0;
440 }
441 
442 static void reset_readdir(struct ceph_file_info *fi)
443 {
444 	if (fi->last_readdir) {
445 		ceph_mdsc_put_request(fi->last_readdir);
446 		fi->last_readdir = NULL;
447 	}
448 	kfree(fi->last_name);
449 	fi->last_name = NULL;
450 	fi->next_offset = 2;  /* compensate for . and .. */
451 	if (fi->dentry) {
452 		dput(fi->dentry);
453 		fi->dentry = NULL;
454 	}
455 	fi->flags &= ~CEPH_F_ATEND;
456 }
457 
458 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int origin)
459 {
460 	struct ceph_file_info *fi = file->private_data;
461 	struct inode *inode = file->f_mapping->host;
462 	loff_t old_offset = offset;
463 	loff_t retval;
464 
465 	mutex_lock(&inode->i_mutex);
466 	retval = -EINVAL;
467 	switch (origin) {
468 	case SEEK_END:
469 		offset += inode->i_size + 2;   /* FIXME */
470 		break;
471 	case SEEK_CUR:
472 		offset += file->f_pos;
473 	case SEEK_SET:
474 		break;
475 	default:
476 		goto out;
477 	}
478 
479 	if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
480 		if (offset != file->f_pos) {
481 			file->f_pos = offset;
482 			file->f_version = 0;
483 			fi->flags &= ~CEPH_F_ATEND;
484 		}
485 		retval = offset;
486 
487 		/*
488 		 * discard buffered readdir content on seekdir(0), or
489 		 * seek to new frag, or seek prior to current chunk.
490 		 */
491 		if (offset == 0 ||
492 		    fpos_frag(offset) != fpos_frag(old_offset) ||
493 		    fpos_off(offset) < fi->offset) {
494 			dout("dir_llseek dropping %p content\n", file);
495 			reset_readdir(fi);
496 		}
497 
498 		/* bump dir_release_count if we did a forward seek */
499 		if (offset > old_offset)
500 			fi->dir_release_count--;
501 	}
502 out:
503 	mutex_unlock(&inode->i_mutex);
504 	return retval;
505 }
506 
507 /*
508  * Handle lookups for the hidden .snap directory.
509  */
510 int ceph_handle_snapdir(struct ceph_mds_request *req,
511 			struct dentry *dentry, int err)
512 {
513 	struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
514 	struct inode *parent = dentry->d_parent->d_inode; /* we hold i_mutex */
515 
516 	/* .snap dir? */
517 	if (err == -ENOENT &&
518 	    ceph_snap(parent) == CEPH_NOSNAP &&
519 	    strcmp(dentry->d_name.name,
520 		   fsc->mount_options->snapdir_name) == 0) {
521 		struct inode *inode = ceph_get_snapdir(parent);
522 		dout("ENOENT on snapdir %p '%.*s', linking to snapdir %p\n",
523 		     dentry, dentry->d_name.len, dentry->d_name.name, inode);
524 		BUG_ON(!d_unhashed(dentry));
525 		d_add(dentry, inode);
526 		err = 0;
527 	}
528 	return err;
529 }
530 
531 /*
532  * Figure out final result of a lookup/open request.
533  *
534  * Mainly, make sure we return the final req->r_dentry (if it already
535  * existed) in place of the original VFS-provided dentry when they
536  * differ.
537  *
538  * Gracefully handle the case where the MDS replies with -ENOENT and
539  * no trace (which it may do, at its discretion, e.g., if it doesn't
540  * care to issue a lease on the negative dentry).
541  */
542 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
543 				  struct dentry *dentry, int err)
544 {
545 	if (err == -ENOENT) {
546 		/* no trace? */
547 		err = 0;
548 		if (!req->r_reply_info.head->is_dentry) {
549 			dout("ENOENT and no trace, dentry %p inode %p\n",
550 			     dentry, dentry->d_inode);
551 			if (dentry->d_inode) {
552 				d_drop(dentry);
553 				err = -ENOENT;
554 			} else {
555 				d_add(dentry, NULL);
556 			}
557 		}
558 	}
559 	if (err)
560 		dentry = ERR_PTR(err);
561 	else if (dentry != req->r_dentry)
562 		dentry = dget(req->r_dentry);   /* we got spliced */
563 	else
564 		dentry = NULL;
565 	return dentry;
566 }
567 
568 static int is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
569 {
570 	return ceph_ino(inode) == CEPH_INO_ROOT &&
571 		strncmp(dentry->d_name.name, ".ceph", 5) == 0;
572 }
573 
574 /*
575  * Look up a single dir entry.  If there is a lookup intent, inform
576  * the MDS so that it gets our 'caps wanted' value in a single op.
577  */
578 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
579 				  struct nameidata *nd)
580 {
581 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
582 	struct ceph_mds_client *mdsc = fsc->mdsc;
583 	struct ceph_mds_request *req;
584 	int op;
585 	int err;
586 
587 	dout("lookup %p dentry %p '%.*s'\n",
588 	     dir, dentry, dentry->d_name.len, dentry->d_name.name);
589 
590 	if (dentry->d_name.len > NAME_MAX)
591 		return ERR_PTR(-ENAMETOOLONG);
592 
593 	err = ceph_init_dentry(dentry);
594 	if (err < 0)
595 		return ERR_PTR(err);
596 
597 	/* can we conclude ENOENT locally? */
598 	if (dentry->d_inode == NULL) {
599 		struct ceph_inode_info *ci = ceph_inode(dir);
600 		struct ceph_dentry_info *di = ceph_dentry(dentry);
601 
602 		spin_lock(&ci->i_ceph_lock);
603 		dout(" dir %p flags are %d\n", dir, ci->i_ceph_flags);
604 		if (strncmp(dentry->d_name.name,
605 			    fsc->mount_options->snapdir_name,
606 			    dentry->d_name.len) &&
607 		    !is_root_ceph_dentry(dir, dentry) &&
608 		    ceph_dir_test_complete(dir) &&
609 		    (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) {
610 			spin_unlock(&ci->i_ceph_lock);
611 			dout(" dir %p complete, -ENOENT\n", dir);
612 			d_add(dentry, NULL);
613 			di->lease_shared_gen = ci->i_shared_gen;
614 			return NULL;
615 		}
616 		spin_unlock(&ci->i_ceph_lock);
617 	}
618 
619 	op = ceph_snap(dir) == CEPH_SNAPDIR ?
620 		CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
621 	req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
622 	if (IS_ERR(req))
623 		return ERR_CAST(req);
624 	req->r_dentry = dget(dentry);
625 	req->r_num_caps = 2;
626 	/* we only need inode linkage */
627 	req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INODE);
628 	req->r_locked_dir = dir;
629 	err = ceph_mdsc_do_request(mdsc, NULL, req);
630 	err = ceph_handle_snapdir(req, dentry, err);
631 	dentry = ceph_finish_lookup(req, dentry, err);
632 	ceph_mdsc_put_request(req);  /* will dput(dentry) */
633 	dout("lookup result=%p\n", dentry);
634 	return dentry;
635 }
636 
637 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
638 		     struct opendata *od, unsigned flags, umode_t mode,
639 		     int *opened)
640 {
641 	int err;
642 	struct dentry *res = NULL;
643 
644 	if (!(flags & O_CREAT)) {
645 		if (dentry->d_name.len > NAME_MAX)
646 			return -ENAMETOOLONG;
647 
648 		err = ceph_init_dentry(dentry);
649 		if (err < 0)
650 			return err;
651 
652 		return ceph_lookup_open(dir, dentry, od, flags, mode, opened);
653 	}
654 
655 	if (d_unhashed(dentry)) {
656 		res = ceph_lookup(dir, dentry, NULL);
657 		if (IS_ERR(res))
658 			return PTR_ERR(res);
659 
660 		if (res)
661 			dentry = res;
662 	}
663 
664 	/* We don't deal with positive dentries here */
665 	if (dentry->d_inode) {
666 		finish_no_open(od, res);
667 		return 1;
668 	}
669 
670 	*opened |= FILE_CREATED;
671 	err = ceph_lookup_open(dir, dentry, od, flags, mode, opened);
672 	dput(res);
673 
674 	return err;
675 }
676 
677 /*
678  * If we do a create but get no trace back from the MDS, follow up with
679  * a lookup (the VFS expects us to link up the provided dentry).
680  */
681 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
682 {
683 	struct dentry *result = ceph_lookup(dir, dentry, NULL);
684 
685 	if (result && !IS_ERR(result)) {
686 		/*
687 		 * We created the item, then did a lookup, and found
688 		 * it was already linked to another inode we already
689 		 * had in our cache (and thus got spliced).  Link our
690 		 * dentry to that inode, but don't hash it, just in
691 		 * case the VFS wants to dereference it.
692 		 */
693 		BUG_ON(!result->d_inode);
694 		d_instantiate(dentry, result->d_inode);
695 		return 0;
696 	}
697 	return PTR_ERR(result);
698 }
699 
700 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
701 		      umode_t mode, dev_t rdev)
702 {
703 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
704 	struct ceph_mds_client *mdsc = fsc->mdsc;
705 	struct ceph_mds_request *req;
706 	int err;
707 
708 	if (ceph_snap(dir) != CEPH_NOSNAP)
709 		return -EROFS;
710 
711 	dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
712 	     dir, dentry, mode, rdev);
713 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
714 	if (IS_ERR(req)) {
715 		d_drop(dentry);
716 		return PTR_ERR(req);
717 	}
718 	req->r_dentry = dget(dentry);
719 	req->r_num_caps = 2;
720 	req->r_locked_dir = dir;
721 	req->r_args.mknod.mode = cpu_to_le32(mode);
722 	req->r_args.mknod.rdev = cpu_to_le32(rdev);
723 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
724 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
725 	err = ceph_mdsc_do_request(mdsc, dir, req);
726 	if (!err && !req->r_reply_info.head->is_dentry)
727 		err = ceph_handle_notrace_create(dir, dentry);
728 	ceph_mdsc_put_request(req);
729 	if (err)
730 		d_drop(dentry);
731 	return err;
732 }
733 
734 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
735 		       struct nameidata *nd)
736 {
737 	return ceph_mknod(dir, dentry, mode, 0);
738 }
739 
740 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
741 			    const char *dest)
742 {
743 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
744 	struct ceph_mds_client *mdsc = fsc->mdsc;
745 	struct ceph_mds_request *req;
746 	int err;
747 
748 	if (ceph_snap(dir) != CEPH_NOSNAP)
749 		return -EROFS;
750 
751 	dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
752 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
753 	if (IS_ERR(req)) {
754 		d_drop(dentry);
755 		return PTR_ERR(req);
756 	}
757 	req->r_dentry = dget(dentry);
758 	req->r_num_caps = 2;
759 	req->r_path2 = kstrdup(dest, GFP_NOFS);
760 	req->r_locked_dir = dir;
761 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
762 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
763 	err = ceph_mdsc_do_request(mdsc, dir, req);
764 	if (!err && !req->r_reply_info.head->is_dentry)
765 		err = ceph_handle_notrace_create(dir, dentry);
766 	ceph_mdsc_put_request(req);
767 	if (err)
768 		d_drop(dentry);
769 	return err;
770 }
771 
772 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
773 {
774 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
775 	struct ceph_mds_client *mdsc = fsc->mdsc;
776 	struct ceph_mds_request *req;
777 	int err = -EROFS;
778 	int op;
779 
780 	if (ceph_snap(dir) == CEPH_SNAPDIR) {
781 		/* mkdir .snap/foo is a MKSNAP */
782 		op = CEPH_MDS_OP_MKSNAP;
783 		dout("mksnap dir %p snap '%.*s' dn %p\n", dir,
784 		     dentry->d_name.len, dentry->d_name.name, dentry);
785 	} else if (ceph_snap(dir) == CEPH_NOSNAP) {
786 		dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
787 		op = CEPH_MDS_OP_MKDIR;
788 	} else {
789 		goto out;
790 	}
791 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
792 	if (IS_ERR(req)) {
793 		err = PTR_ERR(req);
794 		goto out;
795 	}
796 
797 	req->r_dentry = dget(dentry);
798 	req->r_num_caps = 2;
799 	req->r_locked_dir = dir;
800 	req->r_args.mkdir.mode = cpu_to_le32(mode);
801 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
802 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
803 	err = ceph_mdsc_do_request(mdsc, dir, req);
804 	if (!err && !req->r_reply_info.head->is_dentry)
805 		err = ceph_handle_notrace_create(dir, dentry);
806 	ceph_mdsc_put_request(req);
807 out:
808 	if (err < 0)
809 		d_drop(dentry);
810 	return err;
811 }
812 
813 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
814 		     struct dentry *dentry)
815 {
816 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
817 	struct ceph_mds_client *mdsc = fsc->mdsc;
818 	struct ceph_mds_request *req;
819 	int err;
820 
821 	if (ceph_snap(dir) != CEPH_NOSNAP)
822 		return -EROFS;
823 
824 	dout("link in dir %p old_dentry %p dentry %p\n", dir,
825 	     old_dentry, dentry);
826 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
827 	if (IS_ERR(req)) {
828 		d_drop(dentry);
829 		return PTR_ERR(req);
830 	}
831 	req->r_dentry = dget(dentry);
832 	req->r_num_caps = 2;
833 	req->r_old_dentry = dget(old_dentry); /* or inode? hrm. */
834 	req->r_old_dentry_dir = ceph_get_dentry_parent_inode(old_dentry);
835 	req->r_locked_dir = dir;
836 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
837 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
838 	err = ceph_mdsc_do_request(mdsc, dir, req);
839 	if (err) {
840 		d_drop(dentry);
841 	} else if (!req->r_reply_info.head->is_dentry) {
842 		ihold(old_dentry->d_inode);
843 		d_instantiate(dentry, old_dentry->d_inode);
844 	}
845 	ceph_mdsc_put_request(req);
846 	return err;
847 }
848 
849 /*
850  * For a soon-to-be unlinked file, drop the AUTH_RDCACHE caps.  If it
851  * looks like the link count will hit 0, drop any other caps (other
852  * than PIN) we don't specifically want (due to the file still being
853  * open).
854  */
855 static int drop_caps_for_unlink(struct inode *inode)
856 {
857 	struct ceph_inode_info *ci = ceph_inode(inode);
858 	int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
859 
860 	spin_lock(&ci->i_ceph_lock);
861 	if (inode->i_nlink == 1) {
862 		drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
863 		ci->i_ceph_flags |= CEPH_I_NODELAY;
864 	}
865 	spin_unlock(&ci->i_ceph_lock);
866 	return drop;
867 }
868 
869 /*
870  * rmdir and unlink are differ only by the metadata op code
871  */
872 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
873 {
874 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
875 	struct ceph_mds_client *mdsc = fsc->mdsc;
876 	struct inode *inode = dentry->d_inode;
877 	struct ceph_mds_request *req;
878 	int err = -EROFS;
879 	int op;
880 
881 	if (ceph_snap(dir) == CEPH_SNAPDIR) {
882 		/* rmdir .snap/foo is RMSNAP */
883 		dout("rmsnap dir %p '%.*s' dn %p\n", dir, dentry->d_name.len,
884 		     dentry->d_name.name, dentry);
885 		op = CEPH_MDS_OP_RMSNAP;
886 	} else if (ceph_snap(dir) == CEPH_NOSNAP) {
887 		dout("unlink/rmdir dir %p dn %p inode %p\n",
888 		     dir, dentry, inode);
889 		op = S_ISDIR(dentry->d_inode->i_mode) ?
890 			CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
891 	} else
892 		goto out;
893 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
894 	if (IS_ERR(req)) {
895 		err = PTR_ERR(req);
896 		goto out;
897 	}
898 	req->r_dentry = dget(dentry);
899 	req->r_num_caps = 2;
900 	req->r_locked_dir = dir;
901 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
902 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
903 	req->r_inode_drop = drop_caps_for_unlink(inode);
904 	err = ceph_mdsc_do_request(mdsc, dir, req);
905 	if (!err && !req->r_reply_info.head->is_dentry)
906 		d_delete(dentry);
907 	ceph_mdsc_put_request(req);
908 out:
909 	return err;
910 }
911 
912 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
913 		       struct inode *new_dir, struct dentry *new_dentry)
914 {
915 	struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
916 	struct ceph_mds_client *mdsc = fsc->mdsc;
917 	struct ceph_mds_request *req;
918 	int err;
919 
920 	if (ceph_snap(old_dir) != ceph_snap(new_dir))
921 		return -EXDEV;
922 	if (ceph_snap(old_dir) != CEPH_NOSNAP ||
923 	    ceph_snap(new_dir) != CEPH_NOSNAP)
924 		return -EROFS;
925 	dout("rename dir %p dentry %p to dir %p dentry %p\n",
926 	     old_dir, old_dentry, new_dir, new_dentry);
927 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_RENAME, USE_AUTH_MDS);
928 	if (IS_ERR(req))
929 		return PTR_ERR(req);
930 	req->r_dentry = dget(new_dentry);
931 	req->r_num_caps = 2;
932 	req->r_old_dentry = dget(old_dentry);
933 	req->r_old_dentry_dir = ceph_get_dentry_parent_inode(old_dentry);
934 	req->r_locked_dir = new_dir;
935 	req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
936 	req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
937 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
938 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
939 	/* release LINK_RDCACHE on source inode (mds will lock it) */
940 	req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
941 	if (new_dentry->d_inode)
942 		req->r_inode_drop = drop_caps_for_unlink(new_dentry->d_inode);
943 	err = ceph_mdsc_do_request(mdsc, old_dir, req);
944 	if (!err && !req->r_reply_info.head->is_dentry) {
945 		/*
946 		 * Normally d_move() is done by fill_trace (called by
947 		 * do_request, above).  If there is no trace, we need
948 		 * to do it here.
949 		 */
950 
951 		/* d_move screws up d_subdirs order */
952 		ceph_dir_clear_complete(new_dir);
953 
954 		d_move(old_dentry, new_dentry);
955 
956 		/* ensure target dentry is invalidated, despite
957 		   rehashing bug in vfs_rename_dir */
958 		ceph_invalidate_dentry_lease(new_dentry);
959 	}
960 	ceph_mdsc_put_request(req);
961 	return err;
962 }
963 
964 /*
965  * Ensure a dentry lease will no longer revalidate.
966  */
967 void ceph_invalidate_dentry_lease(struct dentry *dentry)
968 {
969 	spin_lock(&dentry->d_lock);
970 	dentry->d_time = jiffies;
971 	ceph_dentry(dentry)->lease_shared_gen = 0;
972 	spin_unlock(&dentry->d_lock);
973 }
974 
975 /*
976  * Check if dentry lease is valid.  If not, delete the lease.  Try to
977  * renew if the least is more than half up.
978  */
979 static int dentry_lease_is_valid(struct dentry *dentry)
980 {
981 	struct ceph_dentry_info *di;
982 	struct ceph_mds_session *s;
983 	int valid = 0;
984 	u32 gen;
985 	unsigned long ttl;
986 	struct ceph_mds_session *session = NULL;
987 	struct inode *dir = NULL;
988 	u32 seq = 0;
989 
990 	spin_lock(&dentry->d_lock);
991 	di = ceph_dentry(dentry);
992 	if (di->lease_session) {
993 		s = di->lease_session;
994 		spin_lock(&s->s_gen_ttl_lock);
995 		gen = s->s_cap_gen;
996 		ttl = s->s_cap_ttl;
997 		spin_unlock(&s->s_gen_ttl_lock);
998 
999 		if (di->lease_gen == gen &&
1000 		    time_before(jiffies, dentry->d_time) &&
1001 		    time_before(jiffies, ttl)) {
1002 			valid = 1;
1003 			if (di->lease_renew_after &&
1004 			    time_after(jiffies, di->lease_renew_after)) {
1005 				/* we should renew */
1006 				dir = dentry->d_parent->d_inode;
1007 				session = ceph_get_mds_session(s);
1008 				seq = di->lease_seq;
1009 				di->lease_renew_after = 0;
1010 				di->lease_renew_from = jiffies;
1011 			}
1012 		}
1013 	}
1014 	spin_unlock(&dentry->d_lock);
1015 
1016 	if (session) {
1017 		ceph_mdsc_lease_send_msg(session, dir, dentry,
1018 					 CEPH_MDS_LEASE_RENEW, seq);
1019 		ceph_put_mds_session(session);
1020 	}
1021 	dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1022 	return valid;
1023 }
1024 
1025 /*
1026  * Check if directory-wide content lease/cap is valid.
1027  */
1028 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry)
1029 {
1030 	struct ceph_inode_info *ci = ceph_inode(dir);
1031 	struct ceph_dentry_info *di = ceph_dentry(dentry);
1032 	int valid = 0;
1033 
1034 	spin_lock(&ci->i_ceph_lock);
1035 	if (ci->i_shared_gen == di->lease_shared_gen)
1036 		valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1037 	spin_unlock(&ci->i_ceph_lock);
1038 	dout("dir_lease_is_valid dir %p v%u dentry %p v%u = %d\n",
1039 	     dir, (unsigned)ci->i_shared_gen, dentry,
1040 	     (unsigned)di->lease_shared_gen, valid);
1041 	return valid;
1042 }
1043 
1044 /*
1045  * Check if cached dentry can be trusted.
1046  */
1047 static int ceph_d_revalidate(struct dentry *dentry, struct nameidata *nd)
1048 {
1049 	int valid = 0;
1050 	struct inode *dir;
1051 
1052 	if (nd && nd->flags & LOOKUP_RCU)
1053 		return -ECHILD;
1054 
1055 	dout("d_revalidate %p '%.*s' inode %p offset %lld\n", dentry,
1056 	     dentry->d_name.len, dentry->d_name.name, dentry->d_inode,
1057 	     ceph_dentry(dentry)->offset);
1058 
1059 	dir = ceph_get_dentry_parent_inode(dentry);
1060 
1061 	/* always trust cached snapped dentries, snapdir dentry */
1062 	if (ceph_snap(dir) != CEPH_NOSNAP) {
1063 		dout("d_revalidate %p '%.*s' inode %p is SNAPPED\n", dentry,
1064 		     dentry->d_name.len, dentry->d_name.name, dentry->d_inode);
1065 		valid = 1;
1066 	} else if (dentry->d_inode &&
1067 		   ceph_snap(dentry->d_inode) == CEPH_SNAPDIR) {
1068 		valid = 1;
1069 	} else if (dentry_lease_is_valid(dentry) ||
1070 		   dir_lease_is_valid(dir, dentry)) {
1071 		valid = 1;
1072 	}
1073 
1074 	dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1075 	if (valid)
1076 		ceph_dentry_lru_touch(dentry);
1077 	else
1078 		d_drop(dentry);
1079 	iput(dir);
1080 	return valid;
1081 }
1082 
1083 /*
1084  * Release our ceph_dentry_info.
1085  */
1086 static void ceph_d_release(struct dentry *dentry)
1087 {
1088 	struct ceph_dentry_info *di = ceph_dentry(dentry);
1089 
1090 	dout("d_release %p\n", dentry);
1091 	ceph_dentry_lru_del(dentry);
1092 	if (di->lease_session)
1093 		ceph_put_mds_session(di->lease_session);
1094 	kmem_cache_free(ceph_dentry_cachep, di);
1095 	dentry->d_fsdata = NULL;
1096 }
1097 
1098 static int ceph_snapdir_d_revalidate(struct dentry *dentry,
1099 					  struct nameidata *nd)
1100 {
1101 	/*
1102 	 * Eventually, we'll want to revalidate snapped metadata
1103 	 * too... probably...
1104 	 */
1105 	return 1;
1106 }
1107 
1108 /*
1109  * Set/clear/test dir complete flag on the dir's dentry.
1110  */
1111 void ceph_dir_set_complete(struct inode *inode)
1112 {
1113 	struct dentry *dentry = d_find_any_alias(inode);
1114 
1115 	if (dentry && ceph_dentry(dentry) &&
1116 	    ceph_test_mount_opt(ceph_sb_to_client(dentry->d_sb), DCACHE)) {
1117 		dout(" marking %p (%p) complete\n", inode, dentry);
1118 		set_bit(CEPH_D_COMPLETE, &ceph_dentry(dentry)->flags);
1119 	}
1120 	dput(dentry);
1121 }
1122 
1123 void ceph_dir_clear_complete(struct inode *inode)
1124 {
1125 	struct dentry *dentry = d_find_any_alias(inode);
1126 
1127 	if (dentry && ceph_dentry(dentry)) {
1128 		dout(" marking %p (%p) complete\n", inode, dentry);
1129 		set_bit(CEPH_D_COMPLETE, &ceph_dentry(dentry)->flags);
1130 	}
1131 	dput(dentry);
1132 }
1133 
1134 bool ceph_dir_test_complete(struct inode *inode)
1135 {
1136 	struct dentry *dentry = d_find_any_alias(inode);
1137 
1138 	if (dentry && ceph_dentry(dentry)) {
1139 		dout(" marking %p (%p) NOT complete\n", inode, dentry);
1140 		clear_bit(CEPH_D_COMPLETE, &ceph_dentry(dentry)->flags);
1141 	}
1142 	dput(dentry);
1143 	return false;
1144 }
1145 
1146 /*
1147  * When the VFS prunes a dentry from the cache, we need to clear the
1148  * complete flag on the parent directory.
1149  *
1150  * Called under dentry->d_lock.
1151  */
1152 static void ceph_d_prune(struct dentry *dentry)
1153 {
1154 	struct ceph_dentry_info *di;
1155 
1156 	dout("ceph_d_prune %p\n", dentry);
1157 
1158 	/* do we have a valid parent? */
1159 	if (!dentry->d_parent || IS_ROOT(dentry))
1160 		return;
1161 
1162 	/* if we are not hashed, we don't affect D_COMPLETE */
1163 	if (d_unhashed(dentry))
1164 		return;
1165 
1166 	/*
1167 	 * we hold d_lock, so d_parent is stable, and d_fsdata is never
1168 	 * cleared until d_release
1169 	 */
1170 	di = ceph_dentry(dentry->d_parent);
1171 	clear_bit(CEPH_D_COMPLETE, &di->flags);
1172 }
1173 
1174 /*
1175  * read() on a dir.  This weird interface hack only works if mounted
1176  * with '-o dirstat'.
1177  */
1178 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1179 			     loff_t *ppos)
1180 {
1181 	struct ceph_file_info *cf = file->private_data;
1182 	struct inode *inode = file->f_dentry->d_inode;
1183 	struct ceph_inode_info *ci = ceph_inode(inode);
1184 	int left;
1185 	const int bufsize = 1024;
1186 
1187 	if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1188 		return -EISDIR;
1189 
1190 	if (!cf->dir_info) {
1191 		cf->dir_info = kmalloc(bufsize, GFP_NOFS);
1192 		if (!cf->dir_info)
1193 			return -ENOMEM;
1194 		cf->dir_info_len =
1195 			snprintf(cf->dir_info, bufsize,
1196 				"entries:   %20lld\n"
1197 				" files:    %20lld\n"
1198 				" subdirs:  %20lld\n"
1199 				"rentries:  %20lld\n"
1200 				" rfiles:   %20lld\n"
1201 				" rsubdirs: %20lld\n"
1202 				"rbytes:    %20lld\n"
1203 				"rctime:    %10ld.%09ld\n",
1204 				ci->i_files + ci->i_subdirs,
1205 				ci->i_files,
1206 				ci->i_subdirs,
1207 				ci->i_rfiles + ci->i_rsubdirs,
1208 				ci->i_rfiles,
1209 				ci->i_rsubdirs,
1210 				ci->i_rbytes,
1211 				(long)ci->i_rctime.tv_sec,
1212 				(long)ci->i_rctime.tv_nsec);
1213 	}
1214 
1215 	if (*ppos >= cf->dir_info_len)
1216 		return 0;
1217 	size = min_t(unsigned, size, cf->dir_info_len-*ppos);
1218 	left = copy_to_user(buf, cf->dir_info + *ppos, size);
1219 	if (left == size)
1220 		return -EFAULT;
1221 	*ppos += (size - left);
1222 	return size - left;
1223 }
1224 
1225 /*
1226  * an fsync() on a dir will wait for any uncommitted directory
1227  * operations to commit.
1228  */
1229 static int ceph_dir_fsync(struct file *file, loff_t start, loff_t end,
1230 			  int datasync)
1231 {
1232 	struct inode *inode = file->f_path.dentry->d_inode;
1233 	struct ceph_inode_info *ci = ceph_inode(inode);
1234 	struct list_head *head = &ci->i_unsafe_dirops;
1235 	struct ceph_mds_request *req;
1236 	u64 last_tid;
1237 	int ret = 0;
1238 
1239 	dout("dir_fsync %p\n", inode);
1240 	ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1241 	if (ret)
1242 		return ret;
1243 	mutex_lock(&inode->i_mutex);
1244 
1245 	spin_lock(&ci->i_unsafe_lock);
1246 	if (list_empty(head))
1247 		goto out;
1248 
1249 	req = list_entry(head->prev,
1250 			 struct ceph_mds_request, r_unsafe_dir_item);
1251 	last_tid = req->r_tid;
1252 
1253 	do {
1254 		ceph_mdsc_get_request(req);
1255 		spin_unlock(&ci->i_unsafe_lock);
1256 
1257 		dout("dir_fsync %p wait on tid %llu (until %llu)\n",
1258 		     inode, req->r_tid, last_tid);
1259 		if (req->r_timeout) {
1260 			ret = wait_for_completion_timeout(
1261 				&req->r_safe_completion, req->r_timeout);
1262 			if (ret > 0)
1263 				ret = 0;
1264 			else if (ret == 0)
1265 				ret = -EIO;  /* timed out */
1266 		} else {
1267 			wait_for_completion(&req->r_safe_completion);
1268 		}
1269 		ceph_mdsc_put_request(req);
1270 
1271 		spin_lock(&ci->i_unsafe_lock);
1272 		if (ret || list_empty(head))
1273 			break;
1274 		req = list_entry(head->next,
1275 				 struct ceph_mds_request, r_unsafe_dir_item);
1276 	} while (req->r_tid < last_tid);
1277 out:
1278 	spin_unlock(&ci->i_unsafe_lock);
1279 	mutex_unlock(&inode->i_mutex);
1280 
1281 	return ret;
1282 }
1283 
1284 /*
1285  * We maintain a private dentry LRU.
1286  *
1287  * FIXME: this needs to be changed to a per-mds lru to be useful.
1288  */
1289 void ceph_dentry_lru_add(struct dentry *dn)
1290 {
1291 	struct ceph_dentry_info *di = ceph_dentry(dn);
1292 	struct ceph_mds_client *mdsc;
1293 
1294 	dout("dentry_lru_add %p %p '%.*s'\n", di, dn,
1295 	     dn->d_name.len, dn->d_name.name);
1296 	mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1297 	spin_lock(&mdsc->dentry_lru_lock);
1298 	list_add_tail(&di->lru, &mdsc->dentry_lru);
1299 	mdsc->num_dentry++;
1300 	spin_unlock(&mdsc->dentry_lru_lock);
1301 }
1302 
1303 void ceph_dentry_lru_touch(struct dentry *dn)
1304 {
1305 	struct ceph_dentry_info *di = ceph_dentry(dn);
1306 	struct ceph_mds_client *mdsc;
1307 
1308 	dout("dentry_lru_touch %p %p '%.*s' (offset %lld)\n", di, dn,
1309 	     dn->d_name.len, dn->d_name.name, di->offset);
1310 	mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1311 	spin_lock(&mdsc->dentry_lru_lock);
1312 	list_move_tail(&di->lru, &mdsc->dentry_lru);
1313 	spin_unlock(&mdsc->dentry_lru_lock);
1314 }
1315 
1316 void ceph_dentry_lru_del(struct dentry *dn)
1317 {
1318 	struct ceph_dentry_info *di = ceph_dentry(dn);
1319 	struct ceph_mds_client *mdsc;
1320 
1321 	dout("dentry_lru_del %p %p '%.*s'\n", di, dn,
1322 	     dn->d_name.len, dn->d_name.name);
1323 	mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1324 	spin_lock(&mdsc->dentry_lru_lock);
1325 	list_del_init(&di->lru);
1326 	mdsc->num_dentry--;
1327 	spin_unlock(&mdsc->dentry_lru_lock);
1328 }
1329 
1330 /*
1331  * Return name hash for a given dentry.  This is dependent on
1332  * the parent directory's hash function.
1333  */
1334 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1335 {
1336 	struct ceph_inode_info *dci = ceph_inode(dir);
1337 
1338 	switch (dci->i_dir_layout.dl_dir_hash) {
1339 	case 0:	/* for backward compat */
1340 	case CEPH_STR_HASH_LINUX:
1341 		return dn->d_name.hash;
1342 
1343 	default:
1344 		return ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1345 				     dn->d_name.name, dn->d_name.len);
1346 	}
1347 }
1348 
1349 const struct file_operations ceph_dir_fops = {
1350 	.read = ceph_read_dir,
1351 	.readdir = ceph_readdir,
1352 	.llseek = ceph_dir_llseek,
1353 	.open = ceph_open,
1354 	.release = ceph_release,
1355 	.unlocked_ioctl = ceph_ioctl,
1356 	.fsync = ceph_dir_fsync,
1357 };
1358 
1359 const struct inode_operations ceph_dir_iops = {
1360 	.lookup = ceph_lookup,
1361 	.permission = ceph_permission,
1362 	.getattr = ceph_getattr,
1363 	.setattr = ceph_setattr,
1364 	.setxattr = ceph_setxattr,
1365 	.getxattr = ceph_getxattr,
1366 	.listxattr = ceph_listxattr,
1367 	.removexattr = ceph_removexattr,
1368 	.mknod = ceph_mknod,
1369 	.symlink = ceph_symlink,
1370 	.mkdir = ceph_mkdir,
1371 	.link = ceph_link,
1372 	.unlink = ceph_unlink,
1373 	.rmdir = ceph_unlink,
1374 	.rename = ceph_rename,
1375 	.create = ceph_create,
1376 	.atomic_open = ceph_atomic_open,
1377 };
1378 
1379 const struct dentry_operations ceph_dentry_ops = {
1380 	.d_revalidate = ceph_d_revalidate,
1381 	.d_release = ceph_d_release,
1382 	.d_prune = ceph_d_prune,
1383 };
1384 
1385 const struct dentry_operations ceph_snapdir_dentry_ops = {
1386 	.d_revalidate = ceph_snapdir_d_revalidate,
1387 	.d_release = ceph_d_release,
1388 };
1389 
1390 const struct dentry_operations ceph_snap_dentry_ops = {
1391 	.d_release = ceph_d_release,
1392 	.d_prune = ceph_d_prune,
1393 };
1394