xref: /linux/fs/ceph/dir.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
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 #include <linux/xattr.h>
9 
10 #include "super.h"
11 #include "mds_client.h"
12 
13 /*
14  * Directory operations: readdir, lookup, create, link, unlink,
15  * rename, etc.
16  */
17 
18 /*
19  * Ceph MDS operations are specified in terms of a base ino and
20  * relative path.  Thus, the client can specify an operation on a
21  * specific inode (e.g., a getattr due to fstat(2)), or as a path
22  * relative to, say, the root directory.
23  *
24  * Normally, we limit ourselves to strict inode ops (no path component)
25  * or dentry operations (a single path component relative to an ino).  The
26  * exception to this is open_root_dentry(), which will open the mount
27  * point by name.
28  */
29 
30 const struct dentry_operations ceph_dentry_ops;
31 
32 /*
33  * Initialize ceph dentry state.
34  */
35 static int ceph_d_init(struct dentry *dentry)
36 {
37 	struct ceph_dentry_info *di;
38 
39 	di = kmem_cache_zalloc(ceph_dentry_cachep, GFP_KERNEL);
40 	if (!di)
41 		return -ENOMEM;          /* oh well */
42 
43 	di->dentry = dentry;
44 	di->lease_session = NULL;
45 	di->time = jiffies;
46 	dentry->d_fsdata = di;
47 	ceph_dentry_lru_add(dentry);
48 	return 0;
49 }
50 
51 /*
52  * for f_pos for readdir:
53  * - hash order:
54  *	(0xff << 52) | ((24 bits hash) << 28) |
55  *	(the nth entry has hash collision);
56  * - frag+name order;
57  *	((frag value) << 28) | (the nth entry in frag);
58  */
59 #define OFFSET_BITS	28
60 #define OFFSET_MASK	((1 << OFFSET_BITS) - 1)
61 #define HASH_ORDER	(0xffull << (OFFSET_BITS + 24))
62 loff_t ceph_make_fpos(unsigned high, unsigned off, bool hash_order)
63 {
64 	loff_t fpos = ((loff_t)high << 28) | (loff_t)off;
65 	if (hash_order)
66 		fpos |= HASH_ORDER;
67 	return fpos;
68 }
69 
70 static bool is_hash_order(loff_t p)
71 {
72 	return (p & HASH_ORDER) == HASH_ORDER;
73 }
74 
75 static unsigned fpos_frag(loff_t p)
76 {
77 	return p >> OFFSET_BITS;
78 }
79 
80 static unsigned fpos_hash(loff_t p)
81 {
82 	return ceph_frag_value(fpos_frag(p));
83 }
84 
85 static unsigned fpos_off(loff_t p)
86 {
87 	return p & OFFSET_MASK;
88 }
89 
90 static int fpos_cmp(loff_t l, loff_t r)
91 {
92 	int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
93 	if (v)
94 		return v;
95 	return (int)(fpos_off(l) - fpos_off(r));
96 }
97 
98 /*
99  * make note of the last dentry we read, so we can
100  * continue at the same lexicographical point,
101  * regardless of what dir changes take place on the
102  * server.
103  */
104 static int note_last_dentry(struct ceph_file_info *fi, const char *name,
105 		            int len, unsigned next_offset)
106 {
107 	char *buf = kmalloc(len+1, GFP_KERNEL);
108 	if (!buf)
109 		return -ENOMEM;
110 	kfree(fi->last_name);
111 	fi->last_name = buf;
112 	memcpy(fi->last_name, name, len);
113 	fi->last_name[len] = 0;
114 	fi->next_offset = next_offset;
115 	dout("note_last_dentry '%s'\n", fi->last_name);
116 	return 0;
117 }
118 
119 
120 static struct dentry *
121 __dcache_find_get_entry(struct dentry *parent, u64 idx,
122 			struct ceph_readdir_cache_control *cache_ctl)
123 {
124 	struct inode *dir = d_inode(parent);
125 	struct dentry *dentry;
126 	unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1;
127 	loff_t ptr_pos = idx * sizeof(struct dentry *);
128 	pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT;
129 
130 	if (ptr_pos >= i_size_read(dir))
131 		return NULL;
132 
133 	if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) {
134 		ceph_readdir_cache_release(cache_ctl);
135 		cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff);
136 		if (!cache_ctl->page) {
137 			dout(" page %lu not found\n", ptr_pgoff);
138 			return ERR_PTR(-EAGAIN);
139 		}
140 		/* reading/filling the cache are serialized by
141 		   i_mutex, no need to use page lock */
142 		unlock_page(cache_ctl->page);
143 		cache_ctl->dentries = kmap(cache_ctl->page);
144 	}
145 
146 	cache_ctl->index = idx & idx_mask;
147 
148 	rcu_read_lock();
149 	spin_lock(&parent->d_lock);
150 	/* check i_size again here, because empty directory can be
151 	 * marked as complete while not holding the i_mutex. */
152 	if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
153 		dentry = cache_ctl->dentries[cache_ctl->index];
154 	else
155 		dentry = NULL;
156 	spin_unlock(&parent->d_lock);
157 	if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
158 		dentry = NULL;
159 	rcu_read_unlock();
160 	return dentry ? : ERR_PTR(-EAGAIN);
161 }
162 
163 /*
164  * When possible, we try to satisfy a readdir by peeking at the
165  * dcache.  We make this work by carefully ordering dentries on
166  * d_child when we initially get results back from the MDS, and
167  * falling back to a "normal" sync readdir if any dentries in the dir
168  * are dropped.
169  *
170  * Complete dir indicates that we have all dentries in the dir.  It is
171  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
172  * the MDS if/when the directory is modified).
173  */
174 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
175 			    u32 shared_gen)
176 {
177 	struct ceph_file_info *fi = file->private_data;
178 	struct dentry *parent = file->f_path.dentry;
179 	struct inode *dir = d_inode(parent);
180 	struct dentry *dentry, *last = NULL;
181 	struct ceph_dentry_info *di;
182 	struct ceph_readdir_cache_control cache_ctl = {};
183 	u64 idx = 0;
184 	int err = 0;
185 
186 	dout("__dcache_readdir %p v%u at %llx\n", dir, shared_gen, ctx->pos);
187 
188 	/* search start position */
189 	if (ctx->pos > 2) {
190 		u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
191 		while (count > 0) {
192 			u64 step = count >> 1;
193 			dentry = __dcache_find_get_entry(parent, idx + step,
194 							 &cache_ctl);
195 			if (!dentry) {
196 				/* use linar search */
197 				idx = 0;
198 				break;
199 			}
200 			if (IS_ERR(dentry)) {
201 				err = PTR_ERR(dentry);
202 				goto out;
203 			}
204 			di = ceph_dentry(dentry);
205 			spin_lock(&dentry->d_lock);
206 			if (fpos_cmp(di->offset, ctx->pos) < 0) {
207 				idx += step + 1;
208 				count -= step + 1;
209 			} else {
210 				count = step;
211 			}
212 			spin_unlock(&dentry->d_lock);
213 			dput(dentry);
214 		}
215 
216 		dout("__dcache_readdir %p cache idx %llu\n", dir, idx);
217 	}
218 
219 
220 	for (;;) {
221 		bool emit_dentry = false;
222 		dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
223 		if (!dentry) {
224 			fi->flags |= CEPH_F_ATEND;
225 			err = 0;
226 			break;
227 		}
228 		if (IS_ERR(dentry)) {
229 			err = PTR_ERR(dentry);
230 			goto out;
231 		}
232 
233 		di = ceph_dentry(dentry);
234 		spin_lock(&dentry->d_lock);
235 		if (di->lease_shared_gen == shared_gen &&
236 		    d_really_is_positive(dentry) &&
237 		    fpos_cmp(ctx->pos, di->offset) <= 0) {
238 			emit_dentry = true;
239 		}
240 		spin_unlock(&dentry->d_lock);
241 
242 		if (emit_dentry) {
243 			dout(" %llx dentry %p %pd %p\n", di->offset,
244 			     dentry, dentry, d_inode(dentry));
245 			ctx->pos = di->offset;
246 			if (!dir_emit(ctx, dentry->d_name.name,
247 				      dentry->d_name.len,
248 				      ceph_translate_ino(dentry->d_sb,
249 							 d_inode(dentry)->i_ino),
250 				      d_inode(dentry)->i_mode >> 12)) {
251 				dput(dentry);
252 				err = 0;
253 				break;
254 			}
255 			ctx->pos++;
256 
257 			if (last)
258 				dput(last);
259 			last = dentry;
260 		} else {
261 			dput(dentry);
262 		}
263 	}
264 out:
265 	ceph_readdir_cache_release(&cache_ctl);
266 	if (last) {
267 		int ret;
268 		di = ceph_dentry(last);
269 		ret = note_last_dentry(fi, last->d_name.name, last->d_name.len,
270 				       fpos_off(di->offset) + 1);
271 		if (ret < 0)
272 			err = ret;
273 		dput(last);
274 	}
275 	return err;
276 }
277 
278 static bool need_send_readdir(struct ceph_file_info *fi, loff_t pos)
279 {
280 	if (!fi->last_readdir)
281 		return true;
282 	if (is_hash_order(pos))
283 		return !ceph_frag_contains_value(fi->frag, fpos_hash(pos));
284 	else
285 		return fi->frag != fpos_frag(pos);
286 }
287 
288 static int ceph_readdir(struct file *file, struct dir_context *ctx)
289 {
290 	struct ceph_file_info *fi = file->private_data;
291 	struct inode *inode = file_inode(file);
292 	struct ceph_inode_info *ci = ceph_inode(inode);
293 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
294 	struct ceph_mds_client *mdsc = fsc->mdsc;
295 	int i;
296 	int err;
297 	u32 ftype;
298 	struct ceph_mds_reply_info_parsed *rinfo;
299 
300 	dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
301 	if (fi->flags & CEPH_F_ATEND)
302 		return 0;
303 
304 	/* always start with . and .. */
305 	if (ctx->pos == 0) {
306 		dout("readdir off 0 -> '.'\n");
307 		if (!dir_emit(ctx, ".", 1,
308 			    ceph_translate_ino(inode->i_sb, inode->i_ino),
309 			    inode->i_mode >> 12))
310 			return 0;
311 		ctx->pos = 1;
312 	}
313 	if (ctx->pos == 1) {
314 		ino_t ino = parent_ino(file->f_path.dentry);
315 		dout("readdir off 1 -> '..'\n");
316 		if (!dir_emit(ctx, "..", 2,
317 			    ceph_translate_ino(inode->i_sb, ino),
318 			    inode->i_mode >> 12))
319 			return 0;
320 		ctx->pos = 2;
321 	}
322 
323 	/* can we use the dcache? */
324 	spin_lock(&ci->i_ceph_lock);
325 	if (ceph_test_mount_opt(fsc, DCACHE) &&
326 	    !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
327 	    ceph_snap(inode) != CEPH_SNAPDIR &&
328 	    __ceph_dir_is_complete_ordered(ci) &&
329 	    __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) {
330 		u32 shared_gen = ci->i_shared_gen;
331 		spin_unlock(&ci->i_ceph_lock);
332 		err = __dcache_readdir(file, ctx, shared_gen);
333 		if (err != -EAGAIN)
334 			return err;
335 	} else {
336 		spin_unlock(&ci->i_ceph_lock);
337 	}
338 
339 	/* proceed with a normal readdir */
340 more:
341 	/* do we have the correct frag content buffered? */
342 	if (need_send_readdir(fi, ctx->pos)) {
343 		struct ceph_mds_request *req;
344 		unsigned frag;
345 		int op = ceph_snap(inode) == CEPH_SNAPDIR ?
346 			CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
347 
348 		/* discard old result, if any */
349 		if (fi->last_readdir) {
350 			ceph_mdsc_put_request(fi->last_readdir);
351 			fi->last_readdir = NULL;
352 		}
353 
354 		if (is_hash_order(ctx->pos)) {
355 			frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
356 						NULL, NULL);
357 		} else {
358 			frag = fpos_frag(ctx->pos);
359 		}
360 
361 		dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
362 		     ceph_vinop(inode), frag, fi->last_name);
363 		req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
364 		if (IS_ERR(req))
365 			return PTR_ERR(req);
366 		err = ceph_alloc_readdir_reply_buffer(req, inode);
367 		if (err) {
368 			ceph_mdsc_put_request(req);
369 			return err;
370 		}
371 		/* hints to request -> mds selection code */
372 		req->r_direct_mode = USE_AUTH_MDS;
373 		req->r_direct_hash = ceph_frag_value(frag);
374 		__set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
375 		if (fi->last_name) {
376 			req->r_path2 = kstrdup(fi->last_name, GFP_KERNEL);
377 			if (!req->r_path2) {
378 				ceph_mdsc_put_request(req);
379 				return -ENOMEM;
380 			}
381 		}
382 		req->r_dir_release_cnt = fi->dir_release_count;
383 		req->r_dir_ordered_cnt = fi->dir_ordered_count;
384 		req->r_readdir_cache_idx = fi->readdir_cache_idx;
385 		req->r_readdir_offset = fi->next_offset;
386 		req->r_args.readdir.frag = cpu_to_le32(frag);
387 		req->r_args.readdir.flags =
388 				cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
389 
390 		req->r_inode = inode;
391 		ihold(inode);
392 		req->r_dentry = dget(file->f_path.dentry);
393 		err = ceph_mdsc_do_request(mdsc, NULL, req);
394 		if (err < 0) {
395 			ceph_mdsc_put_request(req);
396 			return err;
397 		}
398 		dout("readdir got and parsed readdir result=%d on "
399 		     "frag %x, end=%d, complete=%d, hash_order=%d\n",
400 		     err, frag,
401 		     (int)req->r_reply_info.dir_end,
402 		     (int)req->r_reply_info.dir_complete,
403 		     (int)req->r_reply_info.hash_order);
404 
405 		rinfo = &req->r_reply_info;
406 		if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
407 			frag = le32_to_cpu(rinfo->dir_dir->frag);
408 			if (!rinfo->hash_order) {
409 				fi->next_offset = req->r_readdir_offset;
410 				/* adjust ctx->pos to beginning of frag */
411 				ctx->pos = ceph_make_fpos(frag,
412 							  fi->next_offset,
413 							  false);
414 			}
415 		}
416 
417 		fi->frag = frag;
418 		fi->last_readdir = req;
419 
420 		if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
421 			fi->readdir_cache_idx = req->r_readdir_cache_idx;
422 			if (fi->readdir_cache_idx < 0) {
423 				/* preclude from marking dir ordered */
424 				fi->dir_ordered_count = 0;
425 			} else if (ceph_frag_is_leftmost(frag) &&
426 				   fi->next_offset == 2) {
427 				/* note dir version at start of readdir so
428 				 * we can tell if any dentries get dropped */
429 				fi->dir_release_count = req->r_dir_release_cnt;
430 				fi->dir_ordered_count = req->r_dir_ordered_cnt;
431 			}
432 		} else {
433 			dout("readdir !did_prepopulate");
434 			/* disable readdir cache */
435 			fi->readdir_cache_idx = -1;
436 			/* preclude from marking dir complete */
437 			fi->dir_release_count = 0;
438 		}
439 
440 		/* note next offset and last dentry name */
441 		if (rinfo->dir_nr > 0) {
442 			struct ceph_mds_reply_dir_entry *rde =
443 					rinfo->dir_entries + (rinfo->dir_nr-1);
444 			unsigned next_offset = req->r_reply_info.dir_end ?
445 					2 : (fpos_off(rde->offset) + 1);
446 			err = note_last_dentry(fi, rde->name, rde->name_len,
447 					       next_offset);
448 			if (err)
449 				return err;
450 		} else if (req->r_reply_info.dir_end) {
451 			fi->next_offset = 2;
452 			/* keep last name */
453 		}
454 	}
455 
456 	rinfo = &fi->last_readdir->r_reply_info;
457 	dout("readdir frag %x num %d pos %llx chunk first %llx\n",
458 	     fi->frag, rinfo->dir_nr, ctx->pos,
459 	     rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
460 
461 	i = 0;
462 	/* search start position */
463 	if (rinfo->dir_nr > 0) {
464 		int step, nr = rinfo->dir_nr;
465 		while (nr > 0) {
466 			step = nr >> 1;
467 			if (rinfo->dir_entries[i + step].offset < ctx->pos) {
468 				i +=  step + 1;
469 				nr -= step + 1;
470 			} else {
471 				nr = step;
472 			}
473 		}
474 	}
475 	for (; i < rinfo->dir_nr; i++) {
476 		struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
477 		struct ceph_vino vino;
478 		ino_t ino;
479 
480 		BUG_ON(rde->offset < ctx->pos);
481 
482 		ctx->pos = rde->offset;
483 		dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
484 		     i, rinfo->dir_nr, ctx->pos,
485 		     rde->name_len, rde->name, &rde->inode.in);
486 
487 		BUG_ON(!rde->inode.in);
488 		ftype = le32_to_cpu(rde->inode.in->mode) >> 12;
489 		vino.ino = le64_to_cpu(rde->inode.in->ino);
490 		vino.snap = le64_to_cpu(rde->inode.in->snapid);
491 		ino = ceph_vino_to_ino(vino);
492 
493 		if (!dir_emit(ctx, rde->name, rde->name_len,
494 			      ceph_translate_ino(inode->i_sb, ino), ftype)) {
495 			dout("filldir stopping us...\n");
496 			return 0;
497 		}
498 		ctx->pos++;
499 	}
500 
501 	if (fi->next_offset > 2) {
502 		ceph_mdsc_put_request(fi->last_readdir);
503 		fi->last_readdir = NULL;
504 		goto more;
505 	}
506 
507 	/* more frags? */
508 	if (!ceph_frag_is_rightmost(fi->frag)) {
509 		unsigned frag = ceph_frag_next(fi->frag);
510 		if (is_hash_order(ctx->pos)) {
511 			loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
512 							fi->next_offset, true);
513 			if (new_pos > ctx->pos)
514 				ctx->pos = new_pos;
515 			/* keep last_name */
516 		} else {
517 			ctx->pos = ceph_make_fpos(frag, fi->next_offset, false);
518 			kfree(fi->last_name);
519 			fi->last_name = NULL;
520 		}
521 		dout("readdir next frag is %x\n", frag);
522 		goto more;
523 	}
524 	fi->flags |= CEPH_F_ATEND;
525 
526 	/*
527 	 * if dir_release_count still matches the dir, no dentries
528 	 * were released during the whole readdir, and we should have
529 	 * the complete dir contents in our cache.
530 	 */
531 	if (atomic64_read(&ci->i_release_count) == fi->dir_release_count) {
532 		spin_lock(&ci->i_ceph_lock);
533 		if (fi->dir_ordered_count == atomic64_read(&ci->i_ordered_count)) {
534 			dout(" marking %p complete and ordered\n", inode);
535 			/* use i_size to track number of entries in
536 			 * readdir cache */
537 			BUG_ON(fi->readdir_cache_idx < 0);
538 			i_size_write(inode, fi->readdir_cache_idx *
539 				     sizeof(struct dentry*));
540 		} else {
541 			dout(" marking %p complete\n", inode);
542 		}
543 		__ceph_dir_set_complete(ci, fi->dir_release_count,
544 					fi->dir_ordered_count);
545 		spin_unlock(&ci->i_ceph_lock);
546 	}
547 
548 	dout("readdir %p file %p done.\n", inode, file);
549 	return 0;
550 }
551 
552 static void reset_readdir(struct ceph_file_info *fi)
553 {
554 	if (fi->last_readdir) {
555 		ceph_mdsc_put_request(fi->last_readdir);
556 		fi->last_readdir = NULL;
557 	}
558 	kfree(fi->last_name);
559 	fi->last_name = NULL;
560 	fi->dir_release_count = 0;
561 	fi->readdir_cache_idx = -1;
562 	fi->next_offset = 2;  /* compensate for . and .. */
563 	fi->flags &= ~CEPH_F_ATEND;
564 }
565 
566 /*
567  * discard buffered readdir content on seekdir(0), or seek to new frag,
568  * or seek prior to current chunk
569  */
570 static bool need_reset_readdir(struct ceph_file_info *fi, loff_t new_pos)
571 {
572 	struct ceph_mds_reply_info_parsed *rinfo;
573 	loff_t chunk_offset;
574 	if (new_pos == 0)
575 		return true;
576 	if (is_hash_order(new_pos)) {
577 		/* no need to reset last_name for a forward seek when
578 		 * dentries are sotred in hash order */
579 	} else if (fi->frag != fpos_frag(new_pos)) {
580 		return true;
581 	}
582 	rinfo = fi->last_readdir ? &fi->last_readdir->r_reply_info : NULL;
583 	if (!rinfo || !rinfo->dir_nr)
584 		return true;
585 	chunk_offset = rinfo->dir_entries[0].offset;
586 	return new_pos < chunk_offset ||
587 	       is_hash_order(new_pos) != is_hash_order(chunk_offset);
588 }
589 
590 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
591 {
592 	struct ceph_file_info *fi = file->private_data;
593 	struct inode *inode = file->f_mapping->host;
594 	loff_t retval;
595 
596 	inode_lock(inode);
597 	retval = -EINVAL;
598 	switch (whence) {
599 	case SEEK_CUR:
600 		offset += file->f_pos;
601 	case SEEK_SET:
602 		break;
603 	case SEEK_END:
604 		retval = -EOPNOTSUPP;
605 	default:
606 		goto out;
607 	}
608 
609 	if (offset >= 0) {
610 		if (need_reset_readdir(fi, offset)) {
611 			dout("dir_llseek dropping %p content\n", file);
612 			reset_readdir(fi);
613 		} else if (is_hash_order(offset) && offset > file->f_pos) {
614 			/* for hash offset, we don't know if a forward seek
615 			 * is within same frag */
616 			fi->dir_release_count = 0;
617 			fi->readdir_cache_idx = -1;
618 		}
619 
620 		if (offset != file->f_pos) {
621 			file->f_pos = offset;
622 			file->f_version = 0;
623 			fi->flags &= ~CEPH_F_ATEND;
624 		}
625 		retval = offset;
626 	}
627 out:
628 	inode_unlock(inode);
629 	return retval;
630 }
631 
632 /*
633  * Handle lookups for the hidden .snap directory.
634  */
635 int ceph_handle_snapdir(struct ceph_mds_request *req,
636 			struct dentry *dentry, int err)
637 {
638 	struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
639 	struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
640 
641 	/* .snap dir? */
642 	if (err == -ENOENT &&
643 	    ceph_snap(parent) == CEPH_NOSNAP &&
644 	    strcmp(dentry->d_name.name,
645 		   fsc->mount_options->snapdir_name) == 0) {
646 		struct inode *inode = ceph_get_snapdir(parent);
647 		dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
648 		     dentry, dentry, inode);
649 		BUG_ON(!d_unhashed(dentry));
650 		d_add(dentry, inode);
651 		err = 0;
652 	}
653 	return err;
654 }
655 
656 /*
657  * Figure out final result of a lookup/open request.
658  *
659  * Mainly, make sure we return the final req->r_dentry (if it already
660  * existed) in place of the original VFS-provided dentry when they
661  * differ.
662  *
663  * Gracefully handle the case where the MDS replies with -ENOENT and
664  * no trace (which it may do, at its discretion, e.g., if it doesn't
665  * care to issue a lease on the negative dentry).
666  */
667 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
668 				  struct dentry *dentry, int err)
669 {
670 	if (err == -ENOENT) {
671 		/* no trace? */
672 		err = 0;
673 		if (!req->r_reply_info.head->is_dentry) {
674 			dout("ENOENT and no trace, dentry %p inode %p\n",
675 			     dentry, d_inode(dentry));
676 			if (d_really_is_positive(dentry)) {
677 				d_drop(dentry);
678 				err = -ENOENT;
679 			} else {
680 				d_add(dentry, NULL);
681 			}
682 		}
683 	}
684 	if (err)
685 		dentry = ERR_PTR(err);
686 	else if (dentry != req->r_dentry)
687 		dentry = dget(req->r_dentry);   /* we got spliced */
688 	else
689 		dentry = NULL;
690 	return dentry;
691 }
692 
693 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
694 {
695 	return ceph_ino(inode) == CEPH_INO_ROOT &&
696 		strncmp(dentry->d_name.name, ".ceph", 5) == 0;
697 }
698 
699 /*
700  * Look up a single dir entry.  If there is a lookup intent, inform
701  * the MDS so that it gets our 'caps wanted' value in a single op.
702  */
703 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
704 				  unsigned int flags)
705 {
706 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
707 	struct ceph_mds_client *mdsc = fsc->mdsc;
708 	struct ceph_mds_request *req;
709 	int op;
710 	int mask;
711 	int err;
712 
713 	dout("lookup %p dentry %p '%pd'\n",
714 	     dir, dentry, dentry);
715 
716 	if (dentry->d_name.len > NAME_MAX)
717 		return ERR_PTR(-ENAMETOOLONG);
718 
719 	/* can we conclude ENOENT locally? */
720 	if (d_really_is_negative(dentry)) {
721 		struct ceph_inode_info *ci = ceph_inode(dir);
722 		struct ceph_dentry_info *di = ceph_dentry(dentry);
723 
724 		spin_lock(&ci->i_ceph_lock);
725 		dout(" dir %p flags are %d\n", dir, ci->i_ceph_flags);
726 		if (strncmp(dentry->d_name.name,
727 			    fsc->mount_options->snapdir_name,
728 			    dentry->d_name.len) &&
729 		    !is_root_ceph_dentry(dir, dentry) &&
730 		    ceph_test_mount_opt(fsc, DCACHE) &&
731 		    __ceph_dir_is_complete(ci) &&
732 		    (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) {
733 			spin_unlock(&ci->i_ceph_lock);
734 			dout(" dir %p complete, -ENOENT\n", dir);
735 			d_add(dentry, NULL);
736 			di->lease_shared_gen = ci->i_shared_gen;
737 			return NULL;
738 		}
739 		spin_unlock(&ci->i_ceph_lock);
740 	}
741 
742 	op = ceph_snap(dir) == CEPH_SNAPDIR ?
743 		CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
744 	req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
745 	if (IS_ERR(req))
746 		return ERR_CAST(req);
747 	req->r_dentry = dget(dentry);
748 	req->r_num_caps = 2;
749 
750 	mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
751 	if (ceph_security_xattr_wanted(dir))
752 		mask |= CEPH_CAP_XATTR_SHARED;
753 	req->r_args.getattr.mask = cpu_to_le32(mask);
754 
755 	req->r_parent = dir;
756 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
757 	err = ceph_mdsc_do_request(mdsc, NULL, req);
758 	err = ceph_handle_snapdir(req, dentry, err);
759 	dentry = ceph_finish_lookup(req, dentry, err);
760 	ceph_mdsc_put_request(req);  /* will dput(dentry) */
761 	dout("lookup result=%p\n", dentry);
762 	return dentry;
763 }
764 
765 /*
766  * If we do a create but get no trace back from the MDS, follow up with
767  * a lookup (the VFS expects us to link up the provided dentry).
768  */
769 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
770 {
771 	struct dentry *result = ceph_lookup(dir, dentry, 0);
772 
773 	if (result && !IS_ERR(result)) {
774 		/*
775 		 * We created the item, then did a lookup, and found
776 		 * it was already linked to another inode we already
777 		 * had in our cache (and thus got spliced). To not
778 		 * confuse VFS (especially when inode is a directory),
779 		 * we don't link our dentry to that inode, return an
780 		 * error instead.
781 		 *
782 		 * This event should be rare and it happens only when
783 		 * we talk to old MDS. Recent MDS does not send traceless
784 		 * reply for request that creates new inode.
785 		 */
786 		d_drop(result);
787 		return -ESTALE;
788 	}
789 	return PTR_ERR(result);
790 }
791 
792 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
793 		      umode_t mode, dev_t rdev)
794 {
795 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
796 	struct ceph_mds_client *mdsc = fsc->mdsc;
797 	struct ceph_mds_request *req;
798 	struct ceph_acls_info acls = {};
799 	int err;
800 
801 	if (ceph_snap(dir) != CEPH_NOSNAP)
802 		return -EROFS;
803 
804 	err = ceph_pre_init_acls(dir, &mode, &acls);
805 	if (err < 0)
806 		return err;
807 
808 	dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
809 	     dir, dentry, mode, rdev);
810 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
811 	if (IS_ERR(req)) {
812 		err = PTR_ERR(req);
813 		goto out;
814 	}
815 	req->r_dentry = dget(dentry);
816 	req->r_num_caps = 2;
817 	req->r_parent = dir;
818 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
819 	req->r_args.mknod.mode = cpu_to_le32(mode);
820 	req->r_args.mknod.rdev = cpu_to_le32(rdev);
821 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
822 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
823 	if (acls.pagelist) {
824 		req->r_pagelist = acls.pagelist;
825 		acls.pagelist = NULL;
826 	}
827 	err = ceph_mdsc_do_request(mdsc, dir, req);
828 	if (!err && !req->r_reply_info.head->is_dentry)
829 		err = ceph_handle_notrace_create(dir, dentry);
830 	ceph_mdsc_put_request(req);
831 out:
832 	if (!err)
833 		ceph_init_inode_acls(d_inode(dentry), &acls);
834 	else
835 		d_drop(dentry);
836 	ceph_release_acls_info(&acls);
837 	return err;
838 }
839 
840 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
841 		       bool excl)
842 {
843 	return ceph_mknod(dir, dentry, mode, 0);
844 }
845 
846 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
847 			    const char *dest)
848 {
849 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
850 	struct ceph_mds_client *mdsc = fsc->mdsc;
851 	struct ceph_mds_request *req;
852 	int err;
853 
854 	if (ceph_snap(dir) != CEPH_NOSNAP)
855 		return -EROFS;
856 
857 	dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
858 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
859 	if (IS_ERR(req)) {
860 		err = PTR_ERR(req);
861 		goto out;
862 	}
863 	req->r_path2 = kstrdup(dest, GFP_KERNEL);
864 	if (!req->r_path2) {
865 		err = -ENOMEM;
866 		ceph_mdsc_put_request(req);
867 		goto out;
868 	}
869 	req->r_parent = dir;
870 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
871 	req->r_dentry = dget(dentry);
872 	req->r_num_caps = 2;
873 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
874 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
875 	err = ceph_mdsc_do_request(mdsc, dir, req);
876 	if (!err && !req->r_reply_info.head->is_dentry)
877 		err = ceph_handle_notrace_create(dir, dentry);
878 	ceph_mdsc_put_request(req);
879 out:
880 	if (err)
881 		d_drop(dentry);
882 	return err;
883 }
884 
885 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
886 {
887 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
888 	struct ceph_mds_client *mdsc = fsc->mdsc;
889 	struct ceph_mds_request *req;
890 	struct ceph_acls_info acls = {};
891 	int err = -EROFS;
892 	int op;
893 
894 	if (ceph_snap(dir) == CEPH_SNAPDIR) {
895 		/* mkdir .snap/foo is a MKSNAP */
896 		op = CEPH_MDS_OP_MKSNAP;
897 		dout("mksnap dir %p snap '%pd' dn %p\n", dir,
898 		     dentry, dentry);
899 	} else if (ceph_snap(dir) == CEPH_NOSNAP) {
900 		dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
901 		op = CEPH_MDS_OP_MKDIR;
902 	} else {
903 		goto out;
904 	}
905 
906 	mode |= S_IFDIR;
907 	err = ceph_pre_init_acls(dir, &mode, &acls);
908 	if (err < 0)
909 		goto out;
910 
911 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
912 	if (IS_ERR(req)) {
913 		err = PTR_ERR(req);
914 		goto out;
915 	}
916 
917 	req->r_dentry = dget(dentry);
918 	req->r_num_caps = 2;
919 	req->r_parent = dir;
920 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
921 	req->r_args.mkdir.mode = cpu_to_le32(mode);
922 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
923 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
924 	if (acls.pagelist) {
925 		req->r_pagelist = acls.pagelist;
926 		acls.pagelist = NULL;
927 	}
928 	err = ceph_mdsc_do_request(mdsc, dir, req);
929 	if (!err &&
930 	    !req->r_reply_info.head->is_target &&
931 	    !req->r_reply_info.head->is_dentry)
932 		err = ceph_handle_notrace_create(dir, dentry);
933 	ceph_mdsc_put_request(req);
934 out:
935 	if (!err)
936 		ceph_init_inode_acls(d_inode(dentry), &acls);
937 	else
938 		d_drop(dentry);
939 	ceph_release_acls_info(&acls);
940 	return err;
941 }
942 
943 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
944 		     struct dentry *dentry)
945 {
946 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
947 	struct ceph_mds_client *mdsc = fsc->mdsc;
948 	struct ceph_mds_request *req;
949 	int err;
950 
951 	if (ceph_snap(dir) != CEPH_NOSNAP)
952 		return -EROFS;
953 
954 	dout("link in dir %p old_dentry %p dentry %p\n", dir,
955 	     old_dentry, dentry);
956 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
957 	if (IS_ERR(req)) {
958 		d_drop(dentry);
959 		return PTR_ERR(req);
960 	}
961 	req->r_dentry = dget(dentry);
962 	req->r_num_caps = 2;
963 	req->r_old_dentry = dget(old_dentry);
964 	req->r_parent = dir;
965 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
966 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
967 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
968 	/* release LINK_SHARED on source inode (mds will lock it) */
969 	req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
970 	err = ceph_mdsc_do_request(mdsc, dir, req);
971 	if (err) {
972 		d_drop(dentry);
973 	} else if (!req->r_reply_info.head->is_dentry) {
974 		ihold(d_inode(old_dentry));
975 		d_instantiate(dentry, d_inode(old_dentry));
976 	}
977 	ceph_mdsc_put_request(req);
978 	return err;
979 }
980 
981 /*
982  * For a soon-to-be unlinked file, drop the AUTH_RDCACHE caps.  If it
983  * looks like the link count will hit 0, drop any other caps (other
984  * than PIN) we don't specifically want (due to the file still being
985  * open).
986  */
987 static int drop_caps_for_unlink(struct inode *inode)
988 {
989 	struct ceph_inode_info *ci = ceph_inode(inode);
990 	int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
991 
992 	spin_lock(&ci->i_ceph_lock);
993 	if (inode->i_nlink == 1) {
994 		drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
995 		ci->i_ceph_flags |= CEPH_I_NODELAY;
996 	}
997 	spin_unlock(&ci->i_ceph_lock);
998 	return drop;
999 }
1000 
1001 /*
1002  * rmdir and unlink are differ only by the metadata op code
1003  */
1004 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1005 {
1006 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1007 	struct ceph_mds_client *mdsc = fsc->mdsc;
1008 	struct inode *inode = d_inode(dentry);
1009 	struct ceph_mds_request *req;
1010 	int err = -EROFS;
1011 	int op;
1012 
1013 	if (ceph_snap(dir) == CEPH_SNAPDIR) {
1014 		/* rmdir .snap/foo is RMSNAP */
1015 		dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1016 		op = CEPH_MDS_OP_RMSNAP;
1017 	} else if (ceph_snap(dir) == CEPH_NOSNAP) {
1018 		dout("unlink/rmdir dir %p dn %p inode %p\n",
1019 		     dir, dentry, inode);
1020 		op = d_is_dir(dentry) ?
1021 			CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1022 	} else
1023 		goto out;
1024 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1025 	if (IS_ERR(req)) {
1026 		err = PTR_ERR(req);
1027 		goto out;
1028 	}
1029 	req->r_dentry = dget(dentry);
1030 	req->r_num_caps = 2;
1031 	req->r_parent = dir;
1032 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1033 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1034 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1035 	req->r_inode_drop = drop_caps_for_unlink(inode);
1036 	err = ceph_mdsc_do_request(mdsc, dir, req);
1037 	if (!err && !req->r_reply_info.head->is_dentry)
1038 		d_delete(dentry);
1039 	ceph_mdsc_put_request(req);
1040 out:
1041 	return err;
1042 }
1043 
1044 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
1045 		       struct inode *new_dir, struct dentry *new_dentry,
1046 		       unsigned int flags)
1047 {
1048 	struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
1049 	struct ceph_mds_client *mdsc = fsc->mdsc;
1050 	struct ceph_mds_request *req;
1051 	int op = CEPH_MDS_OP_RENAME;
1052 	int err;
1053 
1054 	if (flags)
1055 		return -EINVAL;
1056 
1057 	if (ceph_snap(old_dir) != ceph_snap(new_dir))
1058 		return -EXDEV;
1059 	if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1060 		if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1061 			op = CEPH_MDS_OP_RENAMESNAP;
1062 		else
1063 			return -EROFS;
1064 	}
1065 	dout("rename dir %p dentry %p to dir %p dentry %p\n",
1066 	     old_dir, old_dentry, new_dir, new_dentry);
1067 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1068 	if (IS_ERR(req))
1069 		return PTR_ERR(req);
1070 	ihold(old_dir);
1071 	req->r_dentry = dget(new_dentry);
1072 	req->r_num_caps = 2;
1073 	req->r_old_dentry = dget(old_dentry);
1074 	req->r_old_dentry_dir = old_dir;
1075 	req->r_parent = new_dir;
1076 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1077 	req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1078 	req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1079 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1080 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1081 	/* release LINK_RDCACHE on source inode (mds will lock it) */
1082 	req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
1083 	if (d_really_is_positive(new_dentry))
1084 		req->r_inode_drop = drop_caps_for_unlink(d_inode(new_dentry));
1085 	err = ceph_mdsc_do_request(mdsc, old_dir, req);
1086 	if (!err && !req->r_reply_info.head->is_dentry) {
1087 		/*
1088 		 * Normally d_move() is done by fill_trace (called by
1089 		 * do_request, above).  If there is no trace, we need
1090 		 * to do it here.
1091 		 */
1092 
1093 		/* d_move screws up sibling dentries' offsets */
1094 		ceph_dir_clear_complete(old_dir);
1095 		ceph_dir_clear_complete(new_dir);
1096 
1097 		d_move(old_dentry, new_dentry);
1098 
1099 		/* ensure target dentry is invalidated, despite
1100 		   rehashing bug in vfs_rename_dir */
1101 		ceph_invalidate_dentry_lease(new_dentry);
1102 	}
1103 	ceph_mdsc_put_request(req);
1104 	return err;
1105 }
1106 
1107 /*
1108  * Ensure a dentry lease will no longer revalidate.
1109  */
1110 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1111 {
1112 	spin_lock(&dentry->d_lock);
1113 	ceph_dentry(dentry)->time = jiffies;
1114 	ceph_dentry(dentry)->lease_shared_gen = 0;
1115 	spin_unlock(&dentry->d_lock);
1116 }
1117 
1118 /*
1119  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1120  * renew if the least is more than half up.
1121  */
1122 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags,
1123 				 struct inode *dir)
1124 {
1125 	struct ceph_dentry_info *di;
1126 	struct ceph_mds_session *s;
1127 	int valid = 0;
1128 	u32 gen;
1129 	unsigned long ttl;
1130 	struct ceph_mds_session *session = NULL;
1131 	u32 seq = 0;
1132 
1133 	spin_lock(&dentry->d_lock);
1134 	di = ceph_dentry(dentry);
1135 	if (di && di->lease_session) {
1136 		s = di->lease_session;
1137 		spin_lock(&s->s_gen_ttl_lock);
1138 		gen = s->s_cap_gen;
1139 		ttl = s->s_cap_ttl;
1140 		spin_unlock(&s->s_gen_ttl_lock);
1141 
1142 		if (di->lease_gen == gen &&
1143 		    time_before(jiffies, di->time) &&
1144 		    time_before(jiffies, ttl)) {
1145 			valid = 1;
1146 			if (di->lease_renew_after &&
1147 			    time_after(jiffies, di->lease_renew_after)) {
1148 				/*
1149 				 * We should renew. If we're in RCU walk mode
1150 				 * though, we can't do that so just return
1151 				 * -ECHILD.
1152 				 */
1153 				if (flags & LOOKUP_RCU) {
1154 					valid = -ECHILD;
1155 				} else {
1156 					session = ceph_get_mds_session(s);
1157 					seq = di->lease_seq;
1158 					di->lease_renew_after = 0;
1159 					di->lease_renew_from = jiffies;
1160 				}
1161 			}
1162 		}
1163 	}
1164 	spin_unlock(&dentry->d_lock);
1165 
1166 	if (session) {
1167 		ceph_mdsc_lease_send_msg(session, dir, dentry,
1168 					 CEPH_MDS_LEASE_RENEW, seq);
1169 		ceph_put_mds_session(session);
1170 	}
1171 	dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1172 	return valid;
1173 }
1174 
1175 /*
1176  * Check if directory-wide content lease/cap is valid.
1177  */
1178 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry)
1179 {
1180 	struct ceph_inode_info *ci = ceph_inode(dir);
1181 	struct ceph_dentry_info *di = ceph_dentry(dentry);
1182 	int valid = 0;
1183 
1184 	spin_lock(&ci->i_ceph_lock);
1185 	if (ci->i_shared_gen == di->lease_shared_gen)
1186 		valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1187 	spin_unlock(&ci->i_ceph_lock);
1188 	dout("dir_lease_is_valid dir %p v%u dentry %p v%u = %d\n",
1189 	     dir, (unsigned)ci->i_shared_gen, dentry,
1190 	     (unsigned)di->lease_shared_gen, valid);
1191 	return valid;
1192 }
1193 
1194 /*
1195  * Check if cached dentry can be trusted.
1196  */
1197 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1198 {
1199 	int valid = 0;
1200 	struct dentry *parent;
1201 	struct inode *dir;
1202 
1203 	if (flags & LOOKUP_RCU) {
1204 		parent = READ_ONCE(dentry->d_parent);
1205 		dir = d_inode_rcu(parent);
1206 		if (!dir)
1207 			return -ECHILD;
1208 	} else {
1209 		parent = dget_parent(dentry);
1210 		dir = d_inode(parent);
1211 	}
1212 
1213 	dout("d_revalidate %p '%pd' inode %p offset %lld\n", dentry,
1214 	     dentry, d_inode(dentry), ceph_dentry(dentry)->offset);
1215 
1216 	/* always trust cached snapped dentries, snapdir dentry */
1217 	if (ceph_snap(dir) != CEPH_NOSNAP) {
1218 		dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1219 		     dentry, d_inode(dentry));
1220 		valid = 1;
1221 	} else if (d_really_is_positive(dentry) &&
1222 		   ceph_snap(d_inode(dentry)) == CEPH_SNAPDIR) {
1223 		valid = 1;
1224 	} else {
1225 		valid = dentry_lease_is_valid(dentry, flags, dir);
1226 		if (valid == -ECHILD)
1227 			return valid;
1228 		if (valid || dir_lease_is_valid(dir, dentry)) {
1229 			if (d_really_is_positive(dentry))
1230 				valid = ceph_is_any_caps(d_inode(dentry));
1231 			else
1232 				valid = 1;
1233 		}
1234 	}
1235 
1236 	if (!valid) {
1237 		struct ceph_mds_client *mdsc =
1238 			ceph_sb_to_client(dir->i_sb)->mdsc;
1239 		struct ceph_mds_request *req;
1240 		int op, err;
1241 		u32 mask;
1242 
1243 		if (flags & LOOKUP_RCU)
1244 			return -ECHILD;
1245 
1246 		op = ceph_snap(dir) == CEPH_SNAPDIR ?
1247 			CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1248 		req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1249 		if (!IS_ERR(req)) {
1250 			req->r_dentry = dget(dentry);
1251 			req->r_num_caps = 2;
1252 			req->r_parent = dir;
1253 
1254 			mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1255 			if (ceph_security_xattr_wanted(dir))
1256 				mask |= CEPH_CAP_XATTR_SHARED;
1257 			req->r_args.getattr.mask = cpu_to_le32(mask);
1258 
1259 			err = ceph_mdsc_do_request(mdsc, NULL, req);
1260 			switch (err) {
1261 			case 0:
1262 				if (d_really_is_positive(dentry) &&
1263 				    d_inode(dentry) == req->r_target_inode)
1264 					valid = 1;
1265 				break;
1266 			case -ENOENT:
1267 				if (d_really_is_negative(dentry))
1268 					valid = 1;
1269 				/* Fallthrough */
1270 			default:
1271 				break;
1272 			}
1273 			ceph_mdsc_put_request(req);
1274 			dout("d_revalidate %p lookup result=%d\n",
1275 			     dentry, err);
1276 		}
1277 	}
1278 
1279 	dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1280 	if (valid) {
1281 		ceph_dentry_lru_touch(dentry);
1282 	} else {
1283 		ceph_dir_clear_complete(dir);
1284 	}
1285 
1286 	if (!(flags & LOOKUP_RCU))
1287 		dput(parent);
1288 	return valid;
1289 }
1290 
1291 /*
1292  * Release our ceph_dentry_info.
1293  */
1294 static void ceph_d_release(struct dentry *dentry)
1295 {
1296 	struct ceph_dentry_info *di = ceph_dentry(dentry);
1297 
1298 	dout("d_release %p\n", dentry);
1299 	ceph_dentry_lru_del(dentry);
1300 
1301 	spin_lock(&dentry->d_lock);
1302 	dentry->d_fsdata = NULL;
1303 	spin_unlock(&dentry->d_lock);
1304 
1305 	if (di->lease_session)
1306 		ceph_put_mds_session(di->lease_session);
1307 	kmem_cache_free(ceph_dentry_cachep, di);
1308 }
1309 
1310 /*
1311  * When the VFS prunes a dentry from the cache, we need to clear the
1312  * complete flag on the parent directory.
1313  *
1314  * Called under dentry->d_lock.
1315  */
1316 static void ceph_d_prune(struct dentry *dentry)
1317 {
1318 	dout("ceph_d_prune %p\n", dentry);
1319 
1320 	/* do we have a valid parent? */
1321 	if (IS_ROOT(dentry))
1322 		return;
1323 
1324 	/* if we are not hashed, we don't affect dir's completeness */
1325 	if (d_unhashed(dentry))
1326 		return;
1327 
1328 	if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_SNAPDIR)
1329 		return;
1330 
1331 	/*
1332 	 * we hold d_lock, so d_parent is stable, and d_fsdata is never
1333 	 * cleared until d_release
1334 	 */
1335 	ceph_dir_clear_complete(d_inode(dentry->d_parent));
1336 }
1337 
1338 /*
1339  * read() on a dir.  This weird interface hack only works if mounted
1340  * with '-o dirstat'.
1341  */
1342 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1343 			     loff_t *ppos)
1344 {
1345 	struct ceph_file_info *cf = file->private_data;
1346 	struct inode *inode = file_inode(file);
1347 	struct ceph_inode_info *ci = ceph_inode(inode);
1348 	int left;
1349 	const int bufsize = 1024;
1350 
1351 	if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1352 		return -EISDIR;
1353 
1354 	if (!cf->dir_info) {
1355 		cf->dir_info = kmalloc(bufsize, GFP_KERNEL);
1356 		if (!cf->dir_info)
1357 			return -ENOMEM;
1358 		cf->dir_info_len =
1359 			snprintf(cf->dir_info, bufsize,
1360 				"entries:   %20lld\n"
1361 				" files:    %20lld\n"
1362 				" subdirs:  %20lld\n"
1363 				"rentries:  %20lld\n"
1364 				" rfiles:   %20lld\n"
1365 				" rsubdirs: %20lld\n"
1366 				"rbytes:    %20lld\n"
1367 				"rctime:    %10ld.%09ld\n",
1368 				ci->i_files + ci->i_subdirs,
1369 				ci->i_files,
1370 				ci->i_subdirs,
1371 				ci->i_rfiles + ci->i_rsubdirs,
1372 				ci->i_rfiles,
1373 				ci->i_rsubdirs,
1374 				ci->i_rbytes,
1375 				(long)ci->i_rctime.tv_sec,
1376 				(long)ci->i_rctime.tv_nsec);
1377 	}
1378 
1379 	if (*ppos >= cf->dir_info_len)
1380 		return 0;
1381 	size = min_t(unsigned, size, cf->dir_info_len-*ppos);
1382 	left = copy_to_user(buf, cf->dir_info + *ppos, size);
1383 	if (left == size)
1384 		return -EFAULT;
1385 	*ppos += (size - left);
1386 	return size - left;
1387 }
1388 
1389 /*
1390  * We maintain a private dentry LRU.
1391  *
1392  * FIXME: this needs to be changed to a per-mds lru to be useful.
1393  */
1394 void ceph_dentry_lru_add(struct dentry *dn)
1395 {
1396 	struct ceph_dentry_info *di = ceph_dentry(dn);
1397 	struct ceph_mds_client *mdsc;
1398 
1399 	dout("dentry_lru_add %p %p '%pd'\n", di, dn, dn);
1400 	mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1401 	spin_lock(&mdsc->dentry_lru_lock);
1402 	list_add_tail(&di->lru, &mdsc->dentry_lru);
1403 	mdsc->num_dentry++;
1404 	spin_unlock(&mdsc->dentry_lru_lock);
1405 }
1406 
1407 void ceph_dentry_lru_touch(struct dentry *dn)
1408 {
1409 	struct ceph_dentry_info *di = ceph_dentry(dn);
1410 	struct ceph_mds_client *mdsc;
1411 
1412 	dout("dentry_lru_touch %p %p '%pd' (offset %lld)\n", di, dn, dn,
1413 	     di->offset);
1414 	mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1415 	spin_lock(&mdsc->dentry_lru_lock);
1416 	list_move_tail(&di->lru, &mdsc->dentry_lru);
1417 	spin_unlock(&mdsc->dentry_lru_lock);
1418 }
1419 
1420 void ceph_dentry_lru_del(struct dentry *dn)
1421 {
1422 	struct ceph_dentry_info *di = ceph_dentry(dn);
1423 	struct ceph_mds_client *mdsc;
1424 
1425 	dout("dentry_lru_del %p %p '%pd'\n", di, dn, dn);
1426 	mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1427 	spin_lock(&mdsc->dentry_lru_lock);
1428 	list_del_init(&di->lru);
1429 	mdsc->num_dentry--;
1430 	spin_unlock(&mdsc->dentry_lru_lock);
1431 }
1432 
1433 /*
1434  * Return name hash for a given dentry.  This is dependent on
1435  * the parent directory's hash function.
1436  */
1437 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1438 {
1439 	struct ceph_inode_info *dci = ceph_inode(dir);
1440 
1441 	switch (dci->i_dir_layout.dl_dir_hash) {
1442 	case 0:	/* for backward compat */
1443 	case CEPH_STR_HASH_LINUX:
1444 		return dn->d_name.hash;
1445 
1446 	default:
1447 		return ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1448 				     dn->d_name.name, dn->d_name.len);
1449 	}
1450 }
1451 
1452 const struct file_operations ceph_dir_fops = {
1453 	.read = ceph_read_dir,
1454 	.iterate = ceph_readdir,
1455 	.llseek = ceph_dir_llseek,
1456 	.open = ceph_open,
1457 	.release = ceph_release,
1458 	.unlocked_ioctl = ceph_ioctl,
1459 	.fsync = ceph_fsync,
1460 };
1461 
1462 const struct file_operations ceph_snapdir_fops = {
1463 	.iterate = ceph_readdir,
1464 	.llseek = ceph_dir_llseek,
1465 	.open = ceph_open,
1466 	.release = ceph_release,
1467 };
1468 
1469 const struct inode_operations ceph_dir_iops = {
1470 	.lookup = ceph_lookup,
1471 	.permission = ceph_permission,
1472 	.getattr = ceph_getattr,
1473 	.setattr = ceph_setattr,
1474 	.listxattr = ceph_listxattr,
1475 	.get_acl = ceph_get_acl,
1476 	.set_acl = ceph_set_acl,
1477 	.mknod = ceph_mknod,
1478 	.symlink = ceph_symlink,
1479 	.mkdir = ceph_mkdir,
1480 	.link = ceph_link,
1481 	.unlink = ceph_unlink,
1482 	.rmdir = ceph_unlink,
1483 	.rename = ceph_rename,
1484 	.create = ceph_create,
1485 	.atomic_open = ceph_atomic_open,
1486 };
1487 
1488 const struct inode_operations ceph_snapdir_iops = {
1489 	.lookup = ceph_lookup,
1490 	.permission = ceph_permission,
1491 	.getattr = ceph_getattr,
1492 	.mkdir = ceph_mkdir,
1493 	.rmdir = ceph_unlink,
1494 	.rename = ceph_rename,
1495 };
1496 
1497 const struct dentry_operations ceph_dentry_ops = {
1498 	.d_revalidate = ceph_d_revalidate,
1499 	.d_release = ceph_d_release,
1500 	.d_prune = ceph_d_prune,
1501 	.d_init = ceph_d_init,
1502 };
1503