xref: /linux/fs/ceph/dir.c (revision 827634added7f38b7d724cab1dccdb2b004c13c3)
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 dentry_operations ceph_dentry_ops;
30 
31 /*
32  * Initialize ceph dentry state.
33  */
34 int ceph_init_dentry(struct dentry *dentry)
35 {
36 	struct ceph_dentry_info *di;
37 
38 	if (dentry->d_fsdata)
39 		return 0;
40 
41 	di = kmem_cache_alloc(ceph_dentry_cachep, GFP_NOFS | __GFP_ZERO);
42 	if (!di)
43 		return -ENOMEM;          /* oh well */
44 
45 	spin_lock(&dentry->d_lock);
46 	if (dentry->d_fsdata) {
47 		/* lost a race */
48 		kmem_cache_free(ceph_dentry_cachep, di);
49 		goto out_unlock;
50 	}
51 
52 	if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP)
53 		d_set_d_op(dentry, &ceph_dentry_ops);
54 	else if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_SNAPDIR)
55 		d_set_d_op(dentry, &ceph_snapdir_dentry_ops);
56 	else
57 		d_set_d_op(dentry, &ceph_snap_dentry_ops);
58 
59 	di->dentry = dentry;
60 	di->lease_session = NULL;
61 	dentry->d_time = jiffies;
62 	/* avoid reordering d_fsdata setup so that the check above is safe */
63 	smp_mb();
64 	dentry->d_fsdata = di;
65 	ceph_dentry_lru_add(dentry);
66 out_unlock:
67 	spin_unlock(&dentry->d_lock);
68 	return 0;
69 }
70 
71 struct inode *ceph_get_dentry_parent_inode(struct dentry *dentry)
72 {
73 	struct inode *inode = NULL;
74 
75 	if (!dentry)
76 		return NULL;
77 
78 	spin_lock(&dentry->d_lock);
79 	if (!IS_ROOT(dentry)) {
80 		inode = d_inode(dentry->d_parent);
81 		ihold(inode);
82 	}
83 	spin_unlock(&dentry->d_lock);
84 	return inode;
85 }
86 
87 
88 /*
89  * for readdir, we encode the directory frag and offset within that
90  * frag into f_pos.
91  */
92 static unsigned fpos_frag(loff_t p)
93 {
94 	return p >> 32;
95 }
96 static unsigned fpos_off(loff_t p)
97 {
98 	return p & 0xffffffff;
99 }
100 
101 static int fpos_cmp(loff_t l, loff_t r)
102 {
103 	int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
104 	if (v)
105 		return v;
106 	return (int)(fpos_off(l) - fpos_off(r));
107 }
108 
109 /*
110  * When possible, we try to satisfy a readdir by peeking at the
111  * dcache.  We make this work by carefully ordering dentries on
112  * d_child when we initially get results back from the MDS, and
113  * falling back to a "normal" sync readdir if any dentries in the dir
114  * are dropped.
115  *
116  * Complete dir indicates that we have all dentries in the dir.  It is
117  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
118  * the MDS if/when the directory is modified).
119  */
120 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
121 			    u32 shared_gen)
122 {
123 	struct ceph_file_info *fi = file->private_data;
124 	struct dentry *parent = file->f_path.dentry;
125 	struct inode *dir = d_inode(parent);
126 	struct list_head *p;
127 	struct dentry *dentry, *last;
128 	struct ceph_dentry_info *di;
129 	int err = 0;
130 
131 	/* claim ref on last dentry we returned */
132 	last = fi->dentry;
133 	fi->dentry = NULL;
134 
135 	dout("__dcache_readdir %p v%u at %llu (last %p)\n",
136 	     dir, shared_gen, ctx->pos, last);
137 
138 	spin_lock(&parent->d_lock);
139 
140 	/* start at beginning? */
141 	if (ctx->pos == 2 || last == NULL ||
142 	    fpos_cmp(ctx->pos, ceph_dentry(last)->offset) < 0) {
143 		if (list_empty(&parent->d_subdirs))
144 			goto out_unlock;
145 		p = parent->d_subdirs.prev;
146 		dout(" initial p %p/%p\n", p->prev, p->next);
147 	} else {
148 		p = last->d_child.prev;
149 	}
150 
151 more:
152 	dentry = list_entry(p, struct dentry, d_child);
153 	di = ceph_dentry(dentry);
154 	while (1) {
155 		dout(" p %p/%p %s d_subdirs %p/%p\n", p->prev, p->next,
156 		     d_unhashed(dentry) ? "!hashed" : "hashed",
157 		     parent->d_subdirs.prev, parent->d_subdirs.next);
158 		if (p == &parent->d_subdirs) {
159 			fi->flags |= CEPH_F_ATEND;
160 			goto out_unlock;
161 		}
162 		spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
163 		if (di->lease_shared_gen == shared_gen &&
164 		    !d_unhashed(dentry) && d_really_is_positive(dentry) &&
165 		    ceph_snap(d_inode(dentry)) != CEPH_SNAPDIR &&
166 		    ceph_ino(d_inode(dentry)) != CEPH_INO_CEPH &&
167 		    fpos_cmp(ctx->pos, di->offset) <= 0)
168 			break;
169 		dout(" skipping %p %pd at %llu (%llu)%s%s\n", dentry,
170 		     dentry, di->offset,
171 		     ctx->pos, d_unhashed(dentry) ? " unhashed" : "",
172 		     !d_inode(dentry) ? " null" : "");
173 		spin_unlock(&dentry->d_lock);
174 		p = p->prev;
175 		dentry = list_entry(p, struct dentry, d_child);
176 		di = ceph_dentry(dentry);
177 	}
178 
179 	dget_dlock(dentry);
180 	spin_unlock(&dentry->d_lock);
181 	spin_unlock(&parent->d_lock);
182 
183 	/* make sure a dentry wasn't dropped while we didn't have parent lock */
184 	if (!ceph_dir_is_complete_ordered(dir)) {
185 		dout(" lost dir complete on %p; falling back to mds\n", dir);
186 		dput(dentry);
187 		err = -EAGAIN;
188 		goto out;
189 	}
190 
191 	dout(" %llu (%llu) dentry %p %pd %p\n", di->offset, ctx->pos,
192 	     dentry, dentry, d_inode(dentry));
193 	if (!dir_emit(ctx, dentry->d_name.name,
194 		      dentry->d_name.len,
195 		      ceph_translate_ino(dentry->d_sb, d_inode(dentry)->i_ino),
196 		      d_inode(dentry)->i_mode >> 12)) {
197 		if (last) {
198 			/* remember our position */
199 			fi->dentry = last;
200 			fi->next_offset = fpos_off(di->offset);
201 		}
202 		dput(dentry);
203 		return 0;
204 	}
205 
206 	ctx->pos = di->offset + 1;
207 
208 	if (last)
209 		dput(last);
210 	last = dentry;
211 
212 	spin_lock(&parent->d_lock);
213 	p = p->prev;	/* advance to next dentry */
214 	goto more;
215 
216 out_unlock:
217 	spin_unlock(&parent->d_lock);
218 out:
219 	if (last)
220 		dput(last);
221 	return err;
222 }
223 
224 /*
225  * make note of the last dentry we read, so we can
226  * continue at the same lexicographical point,
227  * regardless of what dir changes take place on the
228  * server.
229  */
230 static int note_last_dentry(struct ceph_file_info *fi, const char *name,
231 			    int len)
232 {
233 	kfree(fi->last_name);
234 	fi->last_name = kmalloc(len+1, GFP_NOFS);
235 	if (!fi->last_name)
236 		return -ENOMEM;
237 	memcpy(fi->last_name, name, len);
238 	fi->last_name[len] = 0;
239 	dout("note_last_dentry '%s'\n", fi->last_name);
240 	return 0;
241 }
242 
243 static int ceph_readdir(struct file *file, struct dir_context *ctx)
244 {
245 	struct ceph_file_info *fi = file->private_data;
246 	struct inode *inode = file_inode(file);
247 	struct ceph_inode_info *ci = ceph_inode(inode);
248 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
249 	struct ceph_mds_client *mdsc = fsc->mdsc;
250 	unsigned frag = fpos_frag(ctx->pos);
251 	int off = fpos_off(ctx->pos);
252 	int err;
253 	u32 ftype;
254 	struct ceph_mds_reply_info_parsed *rinfo;
255 
256 	dout("readdir %p file %p frag %u off %u\n", inode, file, frag, off);
257 	if (fi->flags & CEPH_F_ATEND)
258 		return 0;
259 
260 	/* always start with . and .. */
261 	if (ctx->pos == 0) {
262 		dout("readdir off 0 -> '.'\n");
263 		if (!dir_emit(ctx, ".", 1,
264 			    ceph_translate_ino(inode->i_sb, inode->i_ino),
265 			    inode->i_mode >> 12))
266 			return 0;
267 		ctx->pos = 1;
268 		off = 1;
269 	}
270 	if (ctx->pos == 1) {
271 		ino_t ino = parent_ino(file->f_path.dentry);
272 		dout("readdir off 1 -> '..'\n");
273 		if (!dir_emit(ctx, "..", 2,
274 			    ceph_translate_ino(inode->i_sb, ino),
275 			    inode->i_mode >> 12))
276 			return 0;
277 		ctx->pos = 2;
278 		off = 2;
279 	}
280 
281 	/* can we use the dcache? */
282 	spin_lock(&ci->i_ceph_lock);
283 	if ((ctx->pos == 2 || fi->dentry) &&
284 	    ceph_test_mount_opt(fsc, DCACHE) &&
285 	    !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
286 	    ceph_snap(inode) != CEPH_SNAPDIR &&
287 	    __ceph_dir_is_complete_ordered(ci) &&
288 	    __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) {
289 		u32 shared_gen = ci->i_shared_gen;
290 		spin_unlock(&ci->i_ceph_lock);
291 		err = __dcache_readdir(file, ctx, shared_gen);
292 		if (err != -EAGAIN)
293 			return err;
294 		frag = fpos_frag(ctx->pos);
295 		off = fpos_off(ctx->pos);
296 	} else {
297 		spin_unlock(&ci->i_ceph_lock);
298 	}
299 	if (fi->dentry) {
300 		err = note_last_dentry(fi, fi->dentry->d_name.name,
301 				       fi->dentry->d_name.len);
302 		if (err)
303 			return err;
304 		dput(fi->dentry);
305 		fi->dentry = NULL;
306 	}
307 
308 	/* proceed with a normal readdir */
309 
310 	if (ctx->pos == 2) {
311 		/* note dir version at start of readdir so we can tell
312 		 * if any dentries get dropped */
313 		fi->dir_release_count = atomic_read(&ci->i_release_count);
314 		fi->dir_ordered_count = ci->i_ordered_count;
315 	}
316 
317 more:
318 	/* do we have the correct frag content buffered? */
319 	if (fi->frag != frag || fi->last_readdir == NULL) {
320 		struct ceph_mds_request *req;
321 		int op = ceph_snap(inode) == CEPH_SNAPDIR ?
322 			CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
323 
324 		/* discard old result, if any */
325 		if (fi->last_readdir) {
326 			ceph_mdsc_put_request(fi->last_readdir);
327 			fi->last_readdir = NULL;
328 		}
329 
330 		dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
331 		     ceph_vinop(inode), frag, fi->last_name);
332 		req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
333 		if (IS_ERR(req))
334 			return PTR_ERR(req);
335 		err = ceph_alloc_readdir_reply_buffer(req, inode);
336 		if (err) {
337 			ceph_mdsc_put_request(req);
338 			return err;
339 		}
340 		/* hints to request -> mds selection code */
341 		req->r_direct_mode = USE_AUTH_MDS;
342 		req->r_direct_hash = ceph_frag_value(frag);
343 		req->r_direct_is_hash = true;
344 		if (fi->last_name) {
345 			req->r_path2 = kstrdup(fi->last_name, GFP_NOFS);
346 			if (!req->r_path2) {
347 				ceph_mdsc_put_request(req);
348 				return -ENOMEM;
349 			}
350 		}
351 		req->r_readdir_offset = fi->next_offset;
352 		req->r_args.readdir.frag = cpu_to_le32(frag);
353 
354 		req->r_inode = inode;
355 		ihold(inode);
356 		req->r_dentry = dget(file->f_path.dentry);
357 		err = ceph_mdsc_do_request(mdsc, NULL, req);
358 		if (err < 0) {
359 			ceph_mdsc_put_request(req);
360 			return err;
361 		}
362 		dout("readdir got and parsed readdir result=%d"
363 		     " on frag %x, end=%d, complete=%d\n", err, frag,
364 		     (int)req->r_reply_info.dir_end,
365 		     (int)req->r_reply_info.dir_complete);
366 
367 		if (!req->r_did_prepopulate) {
368 			dout("readdir !did_prepopulate");
369 			/* preclude from marking dir complete */
370 			fi->dir_release_count--;
371 		}
372 
373 		/* note next offset and last dentry name */
374 		rinfo = &req->r_reply_info;
375 		if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
376 			frag = le32_to_cpu(rinfo->dir_dir->frag);
377 			if (ceph_frag_is_leftmost(frag))
378 				fi->next_offset = 2;
379 			else
380 				fi->next_offset = 0;
381 			off = fi->next_offset;
382 		}
383 		fi->frag = frag;
384 		fi->offset = fi->next_offset;
385 		fi->last_readdir = req;
386 
387 		if (req->r_reply_info.dir_end) {
388 			kfree(fi->last_name);
389 			fi->last_name = NULL;
390 			if (ceph_frag_is_rightmost(frag))
391 				fi->next_offset = 2;
392 			else
393 				fi->next_offset = 0;
394 		} else {
395 			err = note_last_dentry(fi,
396 				       rinfo->dir_dname[rinfo->dir_nr-1],
397 				       rinfo->dir_dname_len[rinfo->dir_nr-1]);
398 			if (err)
399 				return err;
400 			fi->next_offset += rinfo->dir_nr;
401 		}
402 	}
403 
404 	rinfo = &fi->last_readdir->r_reply_info;
405 	dout("readdir frag %x num %d off %d chunkoff %d\n", frag,
406 	     rinfo->dir_nr, off, fi->offset);
407 
408 	ctx->pos = ceph_make_fpos(frag, off);
409 	while (off >= fi->offset && off - fi->offset < rinfo->dir_nr) {
410 		struct ceph_mds_reply_inode *in =
411 			rinfo->dir_in[off - fi->offset].in;
412 		struct ceph_vino vino;
413 		ino_t ino;
414 
415 		dout("readdir off %d (%d/%d) -> %lld '%.*s' %p\n",
416 		     off, off - fi->offset, rinfo->dir_nr, ctx->pos,
417 		     rinfo->dir_dname_len[off - fi->offset],
418 		     rinfo->dir_dname[off - fi->offset], in);
419 		BUG_ON(!in);
420 		ftype = le32_to_cpu(in->mode) >> 12;
421 		vino.ino = le64_to_cpu(in->ino);
422 		vino.snap = le64_to_cpu(in->snapid);
423 		ino = ceph_vino_to_ino(vino);
424 		if (!dir_emit(ctx,
425 			    rinfo->dir_dname[off - fi->offset],
426 			    rinfo->dir_dname_len[off - fi->offset],
427 			    ceph_translate_ino(inode->i_sb, ino), ftype)) {
428 			dout("filldir stopping us...\n");
429 			return 0;
430 		}
431 		off++;
432 		ctx->pos++;
433 	}
434 
435 	if (fi->last_name) {
436 		ceph_mdsc_put_request(fi->last_readdir);
437 		fi->last_readdir = NULL;
438 		goto more;
439 	}
440 
441 	/* more frags? */
442 	if (!ceph_frag_is_rightmost(frag)) {
443 		frag = ceph_frag_next(frag);
444 		off = 0;
445 		ctx->pos = ceph_make_fpos(frag, off);
446 		dout("readdir next frag is %x\n", frag);
447 		goto more;
448 	}
449 	fi->flags |= CEPH_F_ATEND;
450 
451 	/*
452 	 * if dir_release_count still matches the dir, no dentries
453 	 * were released during the whole readdir, and we should have
454 	 * the complete dir contents in our cache.
455 	 */
456 	spin_lock(&ci->i_ceph_lock);
457 	if (atomic_read(&ci->i_release_count) == fi->dir_release_count) {
458 		if (ci->i_ordered_count == fi->dir_ordered_count)
459 			dout(" marking %p complete and ordered\n", inode);
460 		else
461 			dout(" marking %p complete\n", inode);
462 		__ceph_dir_set_complete(ci, fi->dir_release_count,
463 					fi->dir_ordered_count);
464 	}
465 	spin_unlock(&ci->i_ceph_lock);
466 
467 	dout("readdir %p file %p done.\n", inode, file);
468 	return 0;
469 }
470 
471 static void reset_readdir(struct ceph_file_info *fi, unsigned frag)
472 {
473 	if (fi->last_readdir) {
474 		ceph_mdsc_put_request(fi->last_readdir);
475 		fi->last_readdir = NULL;
476 	}
477 	kfree(fi->last_name);
478 	fi->last_name = NULL;
479 	if (ceph_frag_is_leftmost(frag))
480 		fi->next_offset = 2;  /* compensate for . and .. */
481 	else
482 		fi->next_offset = 0;
483 	if (fi->dentry) {
484 		dput(fi->dentry);
485 		fi->dentry = NULL;
486 	}
487 	fi->flags &= ~CEPH_F_ATEND;
488 }
489 
490 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
491 {
492 	struct ceph_file_info *fi = file->private_data;
493 	struct inode *inode = file->f_mapping->host;
494 	loff_t old_offset = ceph_make_fpos(fi->frag, fi->next_offset);
495 	loff_t retval;
496 
497 	mutex_lock(&inode->i_mutex);
498 	retval = -EINVAL;
499 	switch (whence) {
500 	case SEEK_END:
501 		offset += inode->i_size + 2;   /* FIXME */
502 		break;
503 	case SEEK_CUR:
504 		offset += file->f_pos;
505 	case SEEK_SET:
506 		break;
507 	default:
508 		goto out;
509 	}
510 
511 	if (offset >= 0) {
512 		if (offset != file->f_pos) {
513 			file->f_pos = offset;
514 			file->f_version = 0;
515 			fi->flags &= ~CEPH_F_ATEND;
516 		}
517 		retval = offset;
518 
519 		/*
520 		 * discard buffered readdir content on seekdir(0), or
521 		 * seek to new frag, or seek prior to current chunk.
522 		 */
523 		if (offset == 0 ||
524 		    fpos_frag(offset) != fi->frag ||
525 		    fpos_off(offset) < fi->offset) {
526 			dout("dir_llseek dropping %p content\n", file);
527 			reset_readdir(fi, fpos_frag(offset));
528 		}
529 
530 		/* bump dir_release_count if we did a forward seek */
531 		if (fpos_cmp(offset, old_offset) > 0)
532 			fi->dir_release_count--;
533 	}
534 out:
535 	mutex_unlock(&inode->i_mutex);
536 	return retval;
537 }
538 
539 /*
540  * Handle lookups for the hidden .snap directory.
541  */
542 int ceph_handle_snapdir(struct ceph_mds_request *req,
543 			struct dentry *dentry, int err)
544 {
545 	struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
546 	struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
547 
548 	/* .snap dir? */
549 	if (err == -ENOENT &&
550 	    ceph_snap(parent) == CEPH_NOSNAP &&
551 	    strcmp(dentry->d_name.name,
552 		   fsc->mount_options->snapdir_name) == 0) {
553 		struct inode *inode = ceph_get_snapdir(parent);
554 		dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
555 		     dentry, dentry, inode);
556 		BUG_ON(!d_unhashed(dentry));
557 		d_add(dentry, inode);
558 		err = 0;
559 	}
560 	return err;
561 }
562 
563 /*
564  * Figure out final result of a lookup/open request.
565  *
566  * Mainly, make sure we return the final req->r_dentry (if it already
567  * existed) in place of the original VFS-provided dentry when they
568  * differ.
569  *
570  * Gracefully handle the case where the MDS replies with -ENOENT and
571  * no trace (which it may do, at its discretion, e.g., if it doesn't
572  * care to issue a lease on the negative dentry).
573  */
574 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
575 				  struct dentry *dentry, int err)
576 {
577 	if (err == -ENOENT) {
578 		/* no trace? */
579 		err = 0;
580 		if (!req->r_reply_info.head->is_dentry) {
581 			dout("ENOENT and no trace, dentry %p inode %p\n",
582 			     dentry, d_inode(dentry));
583 			if (d_really_is_positive(dentry)) {
584 				d_drop(dentry);
585 				err = -ENOENT;
586 			} else {
587 				d_add(dentry, NULL);
588 			}
589 		}
590 	}
591 	if (err)
592 		dentry = ERR_PTR(err);
593 	else if (dentry != req->r_dentry)
594 		dentry = dget(req->r_dentry);   /* we got spliced */
595 	else
596 		dentry = NULL;
597 	return dentry;
598 }
599 
600 static int is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
601 {
602 	return ceph_ino(inode) == CEPH_INO_ROOT &&
603 		strncmp(dentry->d_name.name, ".ceph", 5) == 0;
604 }
605 
606 /*
607  * Look up a single dir entry.  If there is a lookup intent, inform
608  * the MDS so that it gets our 'caps wanted' value in a single op.
609  */
610 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
611 				  unsigned int flags)
612 {
613 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
614 	struct ceph_mds_client *mdsc = fsc->mdsc;
615 	struct ceph_mds_request *req;
616 	int op;
617 	int err;
618 
619 	dout("lookup %p dentry %p '%pd'\n",
620 	     dir, dentry, dentry);
621 
622 	if (dentry->d_name.len > NAME_MAX)
623 		return ERR_PTR(-ENAMETOOLONG);
624 
625 	err = ceph_init_dentry(dentry);
626 	if (err < 0)
627 		return ERR_PTR(err);
628 
629 	/* can we conclude ENOENT locally? */
630 	if (d_really_is_negative(dentry)) {
631 		struct ceph_inode_info *ci = ceph_inode(dir);
632 		struct ceph_dentry_info *di = ceph_dentry(dentry);
633 
634 		spin_lock(&ci->i_ceph_lock);
635 		dout(" dir %p flags are %d\n", dir, ci->i_ceph_flags);
636 		if (strncmp(dentry->d_name.name,
637 			    fsc->mount_options->snapdir_name,
638 			    dentry->d_name.len) &&
639 		    !is_root_ceph_dentry(dir, dentry) &&
640 		    ceph_test_mount_opt(fsc, DCACHE) &&
641 		    __ceph_dir_is_complete(ci) &&
642 		    (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) {
643 			spin_unlock(&ci->i_ceph_lock);
644 			dout(" dir %p complete, -ENOENT\n", dir);
645 			d_add(dentry, NULL);
646 			di->lease_shared_gen = ci->i_shared_gen;
647 			return NULL;
648 		}
649 		spin_unlock(&ci->i_ceph_lock);
650 	}
651 
652 	op = ceph_snap(dir) == CEPH_SNAPDIR ?
653 		CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
654 	req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
655 	if (IS_ERR(req))
656 		return ERR_CAST(req);
657 	req->r_dentry = dget(dentry);
658 	req->r_num_caps = 2;
659 	/* we only need inode linkage */
660 	req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INODE);
661 	req->r_locked_dir = dir;
662 	err = ceph_mdsc_do_request(mdsc, NULL, req);
663 	err = ceph_handle_snapdir(req, dentry, err);
664 	dentry = ceph_finish_lookup(req, dentry, err);
665 	ceph_mdsc_put_request(req);  /* will dput(dentry) */
666 	dout("lookup result=%p\n", dentry);
667 	return dentry;
668 }
669 
670 /*
671  * If we do a create but get no trace back from the MDS, follow up with
672  * a lookup (the VFS expects us to link up the provided dentry).
673  */
674 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
675 {
676 	struct dentry *result = ceph_lookup(dir, dentry, 0);
677 
678 	if (result && !IS_ERR(result)) {
679 		/*
680 		 * We created the item, then did a lookup, and found
681 		 * it was already linked to another inode we already
682 		 * had in our cache (and thus got spliced). To not
683 		 * confuse VFS (especially when inode is a directory),
684 		 * we don't link our dentry to that inode, return an
685 		 * error instead.
686 		 *
687 		 * This event should be rare and it happens only when
688 		 * we talk to old MDS. Recent MDS does not send traceless
689 		 * reply for request that creates new inode.
690 		 */
691 		d_drop(result);
692 		return -ESTALE;
693 	}
694 	return PTR_ERR(result);
695 }
696 
697 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
698 		      umode_t mode, dev_t rdev)
699 {
700 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
701 	struct ceph_mds_client *mdsc = fsc->mdsc;
702 	struct ceph_mds_request *req;
703 	struct ceph_acls_info acls = {};
704 	int err;
705 
706 	if (ceph_snap(dir) != CEPH_NOSNAP)
707 		return -EROFS;
708 
709 	err = ceph_pre_init_acls(dir, &mode, &acls);
710 	if (err < 0)
711 		return err;
712 
713 	dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
714 	     dir, dentry, mode, rdev);
715 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
716 	if (IS_ERR(req)) {
717 		err = PTR_ERR(req);
718 		goto out;
719 	}
720 	req->r_dentry = dget(dentry);
721 	req->r_num_caps = 2;
722 	req->r_locked_dir = dir;
723 	req->r_args.mknod.mode = cpu_to_le32(mode);
724 	req->r_args.mknod.rdev = cpu_to_le32(rdev);
725 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
726 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
727 	if (acls.pagelist) {
728 		req->r_pagelist = acls.pagelist;
729 		acls.pagelist = NULL;
730 	}
731 	err = ceph_mdsc_do_request(mdsc, dir, req);
732 	if (!err && !req->r_reply_info.head->is_dentry)
733 		err = ceph_handle_notrace_create(dir, dentry);
734 	ceph_mdsc_put_request(req);
735 out:
736 	if (!err)
737 		ceph_init_inode_acls(d_inode(dentry), &acls);
738 	else
739 		d_drop(dentry);
740 	ceph_release_acls_info(&acls);
741 	return err;
742 }
743 
744 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
745 		       bool excl)
746 {
747 	return ceph_mknod(dir, dentry, mode, 0);
748 }
749 
750 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
751 			    const char *dest)
752 {
753 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
754 	struct ceph_mds_client *mdsc = fsc->mdsc;
755 	struct ceph_mds_request *req;
756 	int err;
757 
758 	if (ceph_snap(dir) != CEPH_NOSNAP)
759 		return -EROFS;
760 
761 	dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
762 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
763 	if (IS_ERR(req)) {
764 		err = PTR_ERR(req);
765 		goto out;
766 	}
767 	req->r_path2 = kstrdup(dest, GFP_NOFS);
768 	if (!req->r_path2) {
769 		err = -ENOMEM;
770 		ceph_mdsc_put_request(req);
771 		goto out;
772 	}
773 	req->r_locked_dir = dir;
774 	req->r_dentry = dget(dentry);
775 	req->r_num_caps = 2;
776 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
777 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
778 	err = ceph_mdsc_do_request(mdsc, dir, req);
779 	if (!err && !req->r_reply_info.head->is_dentry)
780 		err = ceph_handle_notrace_create(dir, dentry);
781 	ceph_mdsc_put_request(req);
782 out:
783 	if (err)
784 		d_drop(dentry);
785 	return err;
786 }
787 
788 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
789 {
790 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
791 	struct ceph_mds_client *mdsc = fsc->mdsc;
792 	struct ceph_mds_request *req;
793 	struct ceph_acls_info acls = {};
794 	int err = -EROFS;
795 	int op;
796 
797 	if (ceph_snap(dir) == CEPH_SNAPDIR) {
798 		/* mkdir .snap/foo is a MKSNAP */
799 		op = CEPH_MDS_OP_MKSNAP;
800 		dout("mksnap dir %p snap '%pd' dn %p\n", dir,
801 		     dentry, dentry);
802 	} else if (ceph_snap(dir) == CEPH_NOSNAP) {
803 		dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
804 		op = CEPH_MDS_OP_MKDIR;
805 	} else {
806 		goto out;
807 	}
808 
809 	mode |= S_IFDIR;
810 	err = ceph_pre_init_acls(dir, &mode, &acls);
811 	if (err < 0)
812 		goto out;
813 
814 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
815 	if (IS_ERR(req)) {
816 		err = PTR_ERR(req);
817 		goto out;
818 	}
819 
820 	req->r_dentry = dget(dentry);
821 	req->r_num_caps = 2;
822 	req->r_locked_dir = dir;
823 	req->r_args.mkdir.mode = cpu_to_le32(mode);
824 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
825 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
826 	if (acls.pagelist) {
827 		req->r_pagelist = acls.pagelist;
828 		acls.pagelist = NULL;
829 	}
830 	err = ceph_mdsc_do_request(mdsc, dir, req);
831 	if (!err &&
832 	    !req->r_reply_info.head->is_target &&
833 	    !req->r_reply_info.head->is_dentry)
834 		err = ceph_handle_notrace_create(dir, dentry);
835 	ceph_mdsc_put_request(req);
836 out:
837 	if (!err)
838 		ceph_init_inode_acls(d_inode(dentry), &acls);
839 	else
840 		d_drop(dentry);
841 	ceph_release_acls_info(&acls);
842 	return err;
843 }
844 
845 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
846 		     struct dentry *dentry)
847 {
848 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
849 	struct ceph_mds_client *mdsc = fsc->mdsc;
850 	struct ceph_mds_request *req;
851 	int err;
852 
853 	if (ceph_snap(dir) != CEPH_NOSNAP)
854 		return -EROFS;
855 
856 	dout("link in dir %p old_dentry %p dentry %p\n", dir,
857 	     old_dentry, dentry);
858 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
859 	if (IS_ERR(req)) {
860 		d_drop(dentry);
861 		return PTR_ERR(req);
862 	}
863 	req->r_dentry = dget(dentry);
864 	req->r_num_caps = 2;
865 	req->r_old_dentry = dget(old_dentry);
866 	req->r_locked_dir = dir;
867 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
868 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
869 	/* release LINK_SHARED on source inode (mds will lock it) */
870 	req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
871 	err = ceph_mdsc_do_request(mdsc, dir, req);
872 	if (err) {
873 		d_drop(dentry);
874 	} else if (!req->r_reply_info.head->is_dentry) {
875 		ihold(d_inode(old_dentry));
876 		d_instantiate(dentry, d_inode(old_dentry));
877 	}
878 	ceph_mdsc_put_request(req);
879 	return err;
880 }
881 
882 /*
883  * For a soon-to-be unlinked file, drop the AUTH_RDCACHE caps.  If it
884  * looks like the link count will hit 0, drop any other caps (other
885  * than PIN) we don't specifically want (due to the file still being
886  * open).
887  */
888 static int drop_caps_for_unlink(struct inode *inode)
889 {
890 	struct ceph_inode_info *ci = ceph_inode(inode);
891 	int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
892 
893 	spin_lock(&ci->i_ceph_lock);
894 	if (inode->i_nlink == 1) {
895 		drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
896 		ci->i_ceph_flags |= CEPH_I_NODELAY;
897 	}
898 	spin_unlock(&ci->i_ceph_lock);
899 	return drop;
900 }
901 
902 /*
903  * rmdir and unlink are differ only by the metadata op code
904  */
905 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
906 {
907 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
908 	struct ceph_mds_client *mdsc = fsc->mdsc;
909 	struct inode *inode = d_inode(dentry);
910 	struct ceph_mds_request *req;
911 	int err = -EROFS;
912 	int op;
913 
914 	if (ceph_snap(dir) == CEPH_SNAPDIR) {
915 		/* rmdir .snap/foo is RMSNAP */
916 		dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
917 		op = CEPH_MDS_OP_RMSNAP;
918 	} else if (ceph_snap(dir) == CEPH_NOSNAP) {
919 		dout("unlink/rmdir dir %p dn %p inode %p\n",
920 		     dir, dentry, inode);
921 		op = d_is_dir(dentry) ?
922 			CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
923 	} else
924 		goto out;
925 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
926 	if (IS_ERR(req)) {
927 		err = PTR_ERR(req);
928 		goto out;
929 	}
930 	req->r_dentry = dget(dentry);
931 	req->r_num_caps = 2;
932 	req->r_locked_dir = dir;
933 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
934 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
935 	req->r_inode_drop = drop_caps_for_unlink(inode);
936 	err = ceph_mdsc_do_request(mdsc, dir, req);
937 	if (!err && !req->r_reply_info.head->is_dentry)
938 		d_delete(dentry);
939 	ceph_mdsc_put_request(req);
940 out:
941 	return err;
942 }
943 
944 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
945 		       struct inode *new_dir, struct dentry *new_dentry)
946 {
947 	struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
948 	struct ceph_mds_client *mdsc = fsc->mdsc;
949 	struct ceph_mds_request *req;
950 	int op = CEPH_MDS_OP_RENAME;
951 	int err;
952 
953 	if (ceph_snap(old_dir) != ceph_snap(new_dir))
954 		return -EXDEV;
955 	if (ceph_snap(old_dir) != CEPH_NOSNAP) {
956 		if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
957 			op = CEPH_MDS_OP_RENAMESNAP;
958 		else
959 			return -EROFS;
960 	}
961 	dout("rename dir %p dentry %p to dir %p dentry %p\n",
962 	     old_dir, old_dentry, new_dir, new_dentry);
963 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
964 	if (IS_ERR(req))
965 		return PTR_ERR(req);
966 	ihold(old_dir);
967 	req->r_dentry = dget(new_dentry);
968 	req->r_num_caps = 2;
969 	req->r_old_dentry = dget(old_dentry);
970 	req->r_old_dentry_dir = old_dir;
971 	req->r_locked_dir = new_dir;
972 	req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
973 	req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
974 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
975 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
976 	/* release LINK_RDCACHE on source inode (mds will lock it) */
977 	req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
978 	if (d_really_is_positive(new_dentry))
979 		req->r_inode_drop = drop_caps_for_unlink(d_inode(new_dentry));
980 	err = ceph_mdsc_do_request(mdsc, old_dir, req);
981 	if (!err && !req->r_reply_info.head->is_dentry) {
982 		/*
983 		 * Normally d_move() is done by fill_trace (called by
984 		 * do_request, above).  If there is no trace, we need
985 		 * to do it here.
986 		 */
987 
988 		d_move(old_dentry, new_dentry);
989 
990 		/* ensure target dentry is invalidated, despite
991 		   rehashing bug in vfs_rename_dir */
992 		ceph_invalidate_dentry_lease(new_dentry);
993 
994 		/* d_move screws up sibling dentries' offsets */
995 		ceph_dir_clear_complete(old_dir);
996 		ceph_dir_clear_complete(new_dir);
997 
998 	}
999 	ceph_mdsc_put_request(req);
1000 	return err;
1001 }
1002 
1003 /*
1004  * Ensure a dentry lease will no longer revalidate.
1005  */
1006 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1007 {
1008 	spin_lock(&dentry->d_lock);
1009 	dentry->d_time = jiffies;
1010 	ceph_dentry(dentry)->lease_shared_gen = 0;
1011 	spin_unlock(&dentry->d_lock);
1012 }
1013 
1014 /*
1015  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1016  * renew if the least is more than half up.
1017  */
1018 static int dentry_lease_is_valid(struct dentry *dentry)
1019 {
1020 	struct ceph_dentry_info *di;
1021 	struct ceph_mds_session *s;
1022 	int valid = 0;
1023 	u32 gen;
1024 	unsigned long ttl;
1025 	struct ceph_mds_session *session = NULL;
1026 	struct inode *dir = NULL;
1027 	u32 seq = 0;
1028 
1029 	spin_lock(&dentry->d_lock);
1030 	di = ceph_dentry(dentry);
1031 	if (di->lease_session) {
1032 		s = di->lease_session;
1033 		spin_lock(&s->s_gen_ttl_lock);
1034 		gen = s->s_cap_gen;
1035 		ttl = s->s_cap_ttl;
1036 		spin_unlock(&s->s_gen_ttl_lock);
1037 
1038 		if (di->lease_gen == gen &&
1039 		    time_before(jiffies, dentry->d_time) &&
1040 		    time_before(jiffies, ttl)) {
1041 			valid = 1;
1042 			if (di->lease_renew_after &&
1043 			    time_after(jiffies, di->lease_renew_after)) {
1044 				/* we should renew */
1045 				dir = d_inode(dentry->d_parent);
1046 				session = ceph_get_mds_session(s);
1047 				seq = di->lease_seq;
1048 				di->lease_renew_after = 0;
1049 				di->lease_renew_from = jiffies;
1050 			}
1051 		}
1052 	}
1053 	spin_unlock(&dentry->d_lock);
1054 
1055 	if (session) {
1056 		ceph_mdsc_lease_send_msg(session, dir, dentry,
1057 					 CEPH_MDS_LEASE_RENEW, seq);
1058 		ceph_put_mds_session(session);
1059 	}
1060 	dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1061 	return valid;
1062 }
1063 
1064 /*
1065  * Check if directory-wide content lease/cap is valid.
1066  */
1067 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry)
1068 {
1069 	struct ceph_inode_info *ci = ceph_inode(dir);
1070 	struct ceph_dentry_info *di = ceph_dentry(dentry);
1071 	int valid = 0;
1072 
1073 	spin_lock(&ci->i_ceph_lock);
1074 	if (ci->i_shared_gen == di->lease_shared_gen)
1075 		valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1076 	spin_unlock(&ci->i_ceph_lock);
1077 	dout("dir_lease_is_valid dir %p v%u dentry %p v%u = %d\n",
1078 	     dir, (unsigned)ci->i_shared_gen, dentry,
1079 	     (unsigned)di->lease_shared_gen, valid);
1080 	return valid;
1081 }
1082 
1083 /*
1084  * Check if cached dentry can be trusted.
1085  */
1086 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1087 {
1088 	int valid = 0;
1089 	struct inode *dir;
1090 
1091 	if (flags & LOOKUP_RCU)
1092 		return -ECHILD;
1093 
1094 	dout("d_revalidate %p '%pd' inode %p offset %lld\n", dentry,
1095 	     dentry, d_inode(dentry), ceph_dentry(dentry)->offset);
1096 
1097 	dir = ceph_get_dentry_parent_inode(dentry);
1098 
1099 	/* always trust cached snapped dentries, snapdir dentry */
1100 	if (ceph_snap(dir) != CEPH_NOSNAP) {
1101 		dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1102 		     dentry, d_inode(dentry));
1103 		valid = 1;
1104 	} else if (d_really_is_positive(dentry) &&
1105 		   ceph_snap(d_inode(dentry)) == CEPH_SNAPDIR) {
1106 		valid = 1;
1107 	} else if (dentry_lease_is_valid(dentry) ||
1108 		   dir_lease_is_valid(dir, dentry)) {
1109 		if (d_really_is_positive(dentry))
1110 			valid = ceph_is_any_caps(d_inode(dentry));
1111 		else
1112 			valid = 1;
1113 	}
1114 
1115 	dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1116 	if (valid) {
1117 		ceph_dentry_lru_touch(dentry);
1118 	} else {
1119 		ceph_dir_clear_complete(dir);
1120 	}
1121 	iput(dir);
1122 	return valid;
1123 }
1124 
1125 /*
1126  * Release our ceph_dentry_info.
1127  */
1128 static void ceph_d_release(struct dentry *dentry)
1129 {
1130 	struct ceph_dentry_info *di = ceph_dentry(dentry);
1131 
1132 	dout("d_release %p\n", dentry);
1133 	ceph_dentry_lru_del(dentry);
1134 	if (di->lease_session)
1135 		ceph_put_mds_session(di->lease_session);
1136 	kmem_cache_free(ceph_dentry_cachep, di);
1137 	dentry->d_fsdata = NULL;
1138 }
1139 
1140 static int ceph_snapdir_d_revalidate(struct dentry *dentry,
1141 					  unsigned int flags)
1142 {
1143 	/*
1144 	 * Eventually, we'll want to revalidate snapped metadata
1145 	 * too... probably...
1146 	 */
1147 	return 1;
1148 }
1149 
1150 /*
1151  * When the VFS prunes a dentry from the cache, we need to clear the
1152  * complete flag on the parent directory.
1153  *
1154  * Called under dentry->d_lock.
1155  */
1156 static void ceph_d_prune(struct dentry *dentry)
1157 {
1158 	dout("ceph_d_prune %p\n", dentry);
1159 
1160 	/* do we have a valid parent? */
1161 	if (IS_ROOT(dentry))
1162 		return;
1163 
1164 	/* if we are not hashed, we don't affect dir's completeness */
1165 	if (d_unhashed(dentry))
1166 		return;
1167 
1168 	/*
1169 	 * we hold d_lock, so d_parent is stable, and d_fsdata is never
1170 	 * cleared until d_release
1171 	 */
1172 	ceph_dir_clear_complete(d_inode(dentry->d_parent));
1173 }
1174 
1175 /*
1176  * read() on a dir.  This weird interface hack only works if mounted
1177  * with '-o dirstat'.
1178  */
1179 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1180 			     loff_t *ppos)
1181 {
1182 	struct ceph_file_info *cf = file->private_data;
1183 	struct inode *inode = file_inode(file);
1184 	struct ceph_inode_info *ci = ceph_inode(inode);
1185 	int left;
1186 	const int bufsize = 1024;
1187 
1188 	if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1189 		return -EISDIR;
1190 
1191 	if (!cf->dir_info) {
1192 		cf->dir_info = kmalloc(bufsize, GFP_NOFS);
1193 		if (!cf->dir_info)
1194 			return -ENOMEM;
1195 		cf->dir_info_len =
1196 			snprintf(cf->dir_info, bufsize,
1197 				"entries:   %20lld\n"
1198 				" files:    %20lld\n"
1199 				" subdirs:  %20lld\n"
1200 				"rentries:  %20lld\n"
1201 				" rfiles:   %20lld\n"
1202 				" rsubdirs: %20lld\n"
1203 				"rbytes:    %20lld\n"
1204 				"rctime:    %10ld.%09ld\n",
1205 				ci->i_files + ci->i_subdirs,
1206 				ci->i_files,
1207 				ci->i_subdirs,
1208 				ci->i_rfiles + ci->i_rsubdirs,
1209 				ci->i_rfiles,
1210 				ci->i_rsubdirs,
1211 				ci->i_rbytes,
1212 				(long)ci->i_rctime.tv_sec,
1213 				(long)ci->i_rctime.tv_nsec);
1214 	}
1215 
1216 	if (*ppos >= cf->dir_info_len)
1217 		return 0;
1218 	size = min_t(unsigned, size, cf->dir_info_len-*ppos);
1219 	left = copy_to_user(buf, cf->dir_info + *ppos, size);
1220 	if (left == size)
1221 		return -EFAULT;
1222 	*ppos += (size - left);
1223 	return size - left;
1224 }
1225 
1226 /*
1227  * an fsync() on a dir will wait for any uncommitted directory
1228  * operations to commit.
1229  */
1230 static int ceph_dir_fsync(struct file *file, loff_t start, loff_t end,
1231 			  int datasync)
1232 {
1233 	struct inode *inode = file_inode(file);
1234 	struct ceph_inode_info *ci = ceph_inode(inode);
1235 	struct list_head *head = &ci->i_unsafe_dirops;
1236 	struct ceph_mds_request *req;
1237 	u64 last_tid;
1238 	int ret = 0;
1239 
1240 	dout("dir_fsync %p\n", inode);
1241 	ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1242 	if (ret)
1243 		return ret;
1244 	mutex_lock(&inode->i_mutex);
1245 
1246 	spin_lock(&ci->i_unsafe_lock);
1247 	if (list_empty(head))
1248 		goto out;
1249 
1250 	req = list_entry(head->prev,
1251 			 struct ceph_mds_request, r_unsafe_dir_item);
1252 	last_tid = req->r_tid;
1253 
1254 	do {
1255 		ceph_mdsc_get_request(req);
1256 		spin_unlock(&ci->i_unsafe_lock);
1257 
1258 		dout("dir_fsync %p wait on tid %llu (until %llu)\n",
1259 		     inode, req->r_tid, last_tid);
1260 		if (req->r_timeout) {
1261 			unsigned long time_left = wait_for_completion_timeout(
1262 							&req->r_safe_completion,
1263 							req->r_timeout);
1264 			if (time_left > 0)
1265 				ret = 0;
1266 			else
1267 				ret = -EIO;  /* timed out */
1268 		} else {
1269 			wait_for_completion(&req->r_safe_completion);
1270 		}
1271 		ceph_mdsc_put_request(req);
1272 
1273 		spin_lock(&ci->i_unsafe_lock);
1274 		if (ret || list_empty(head))
1275 			break;
1276 		req = list_entry(head->next,
1277 				 struct ceph_mds_request, r_unsafe_dir_item);
1278 	} while (req->r_tid < last_tid);
1279 out:
1280 	spin_unlock(&ci->i_unsafe_lock);
1281 	mutex_unlock(&inode->i_mutex);
1282 
1283 	return ret;
1284 }
1285 
1286 /*
1287  * We maintain a private dentry LRU.
1288  *
1289  * FIXME: this needs to be changed to a per-mds lru to be useful.
1290  */
1291 void ceph_dentry_lru_add(struct dentry *dn)
1292 {
1293 	struct ceph_dentry_info *di = ceph_dentry(dn);
1294 	struct ceph_mds_client *mdsc;
1295 
1296 	dout("dentry_lru_add %p %p '%pd'\n", di, dn, dn);
1297 	mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1298 	spin_lock(&mdsc->dentry_lru_lock);
1299 	list_add_tail(&di->lru, &mdsc->dentry_lru);
1300 	mdsc->num_dentry++;
1301 	spin_unlock(&mdsc->dentry_lru_lock);
1302 }
1303 
1304 void ceph_dentry_lru_touch(struct dentry *dn)
1305 {
1306 	struct ceph_dentry_info *di = ceph_dentry(dn);
1307 	struct ceph_mds_client *mdsc;
1308 
1309 	dout("dentry_lru_touch %p %p '%pd' (offset %lld)\n", di, dn, dn,
1310 	     di->offset);
1311 	mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1312 	spin_lock(&mdsc->dentry_lru_lock);
1313 	list_move_tail(&di->lru, &mdsc->dentry_lru);
1314 	spin_unlock(&mdsc->dentry_lru_lock);
1315 }
1316 
1317 void ceph_dentry_lru_del(struct dentry *dn)
1318 {
1319 	struct ceph_dentry_info *di = ceph_dentry(dn);
1320 	struct ceph_mds_client *mdsc;
1321 
1322 	dout("dentry_lru_del %p %p '%pd'\n", di, dn, dn);
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 	.iterate = 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 file_operations ceph_snapdir_fops = {
1360 	.iterate = ceph_readdir,
1361 	.llseek = ceph_dir_llseek,
1362 	.open = ceph_open,
1363 	.release = ceph_release,
1364 };
1365 
1366 const struct inode_operations ceph_dir_iops = {
1367 	.lookup = ceph_lookup,
1368 	.permission = ceph_permission,
1369 	.getattr = ceph_getattr,
1370 	.setattr = ceph_setattr,
1371 	.setxattr = ceph_setxattr,
1372 	.getxattr = ceph_getxattr,
1373 	.listxattr = ceph_listxattr,
1374 	.removexattr = ceph_removexattr,
1375 	.get_acl = ceph_get_acl,
1376 	.set_acl = ceph_set_acl,
1377 	.mknod = ceph_mknod,
1378 	.symlink = ceph_symlink,
1379 	.mkdir = ceph_mkdir,
1380 	.link = ceph_link,
1381 	.unlink = ceph_unlink,
1382 	.rmdir = ceph_unlink,
1383 	.rename = ceph_rename,
1384 	.create = ceph_create,
1385 	.atomic_open = ceph_atomic_open,
1386 };
1387 
1388 const struct inode_operations ceph_snapdir_iops = {
1389 	.lookup = ceph_lookup,
1390 	.permission = ceph_permission,
1391 	.getattr = ceph_getattr,
1392 	.mkdir = ceph_mkdir,
1393 	.rmdir = ceph_unlink,
1394 	.rename = ceph_rename,
1395 };
1396 
1397 const struct dentry_operations ceph_dentry_ops = {
1398 	.d_revalidate = ceph_d_revalidate,
1399 	.d_release = ceph_d_release,
1400 	.d_prune = ceph_d_prune,
1401 };
1402 
1403 const struct dentry_operations ceph_snapdir_dentry_ops = {
1404 	.d_revalidate = ceph_snapdir_d_revalidate,
1405 	.d_release = ceph_d_release,
1406 };
1407 
1408 const struct dentry_operations ceph_snap_dentry_ops = {
1409 	.d_release = ceph_d_release,
1410 	.d_prune = ceph_d_prune,
1411 };
1412