xref: /openbmc/linux/fs/ceph/dir.c (revision 6aa7de05)
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 		/* last_name no longer match cache index */
275 		if (fi->readdir_cache_idx >= 0) {
276 			fi->readdir_cache_idx = -1;
277 			fi->dir_release_count = 0;
278 		}
279 	}
280 	return err;
281 }
282 
283 static bool need_send_readdir(struct ceph_file_info *fi, loff_t pos)
284 {
285 	if (!fi->last_readdir)
286 		return true;
287 	if (is_hash_order(pos))
288 		return !ceph_frag_contains_value(fi->frag, fpos_hash(pos));
289 	else
290 		return fi->frag != fpos_frag(pos);
291 }
292 
293 static int ceph_readdir(struct file *file, struct dir_context *ctx)
294 {
295 	struct ceph_file_info *fi = file->private_data;
296 	struct inode *inode = file_inode(file);
297 	struct ceph_inode_info *ci = ceph_inode(inode);
298 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
299 	struct ceph_mds_client *mdsc = fsc->mdsc;
300 	int i;
301 	int err;
302 	unsigned frag = -1;
303 	struct ceph_mds_reply_info_parsed *rinfo;
304 
305 	dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
306 	if (fi->flags & CEPH_F_ATEND)
307 		return 0;
308 
309 	/* always start with . and .. */
310 	if (ctx->pos == 0) {
311 		dout("readdir off 0 -> '.'\n");
312 		if (!dir_emit(ctx, ".", 1,
313 			    ceph_translate_ino(inode->i_sb, inode->i_ino),
314 			    inode->i_mode >> 12))
315 			return 0;
316 		ctx->pos = 1;
317 	}
318 	if (ctx->pos == 1) {
319 		ino_t ino = parent_ino(file->f_path.dentry);
320 		dout("readdir off 1 -> '..'\n");
321 		if (!dir_emit(ctx, "..", 2,
322 			    ceph_translate_ino(inode->i_sb, ino),
323 			    inode->i_mode >> 12))
324 			return 0;
325 		ctx->pos = 2;
326 	}
327 
328 	/* can we use the dcache? */
329 	spin_lock(&ci->i_ceph_lock);
330 	if (ceph_test_mount_opt(fsc, DCACHE) &&
331 	    !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
332 	    ceph_snap(inode) != CEPH_SNAPDIR &&
333 	    __ceph_dir_is_complete_ordered(ci) &&
334 	    __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) {
335 		u32 shared_gen = ci->i_shared_gen;
336 		spin_unlock(&ci->i_ceph_lock);
337 		err = __dcache_readdir(file, ctx, shared_gen);
338 		if (err != -EAGAIN)
339 			return err;
340 	} else {
341 		spin_unlock(&ci->i_ceph_lock);
342 	}
343 
344 	/* proceed with a normal readdir */
345 more:
346 	/* do we have the correct frag content buffered? */
347 	if (need_send_readdir(fi, ctx->pos)) {
348 		struct ceph_mds_request *req;
349 		int op = ceph_snap(inode) == CEPH_SNAPDIR ?
350 			CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
351 
352 		/* discard old result, if any */
353 		if (fi->last_readdir) {
354 			ceph_mdsc_put_request(fi->last_readdir);
355 			fi->last_readdir = NULL;
356 		}
357 
358 		if (is_hash_order(ctx->pos)) {
359 			/* fragtree isn't always accurate. choose frag
360 			 * based on previous reply when possible. */
361 			if (frag == (unsigned)-1)
362 				frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
363 							NULL, NULL);
364 		} else {
365 			frag = fpos_frag(ctx->pos);
366 		}
367 
368 		dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
369 		     ceph_vinop(inode), frag, fi->last_name);
370 		req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
371 		if (IS_ERR(req))
372 			return PTR_ERR(req);
373 		err = ceph_alloc_readdir_reply_buffer(req, inode);
374 		if (err) {
375 			ceph_mdsc_put_request(req);
376 			return err;
377 		}
378 		/* hints to request -> mds selection code */
379 		req->r_direct_mode = USE_AUTH_MDS;
380 		if (op == CEPH_MDS_OP_READDIR) {
381 			req->r_direct_hash = ceph_frag_value(frag);
382 			__set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
383 		}
384 		if (fi->last_name) {
385 			req->r_path2 = kstrdup(fi->last_name, GFP_KERNEL);
386 			if (!req->r_path2) {
387 				ceph_mdsc_put_request(req);
388 				return -ENOMEM;
389 			}
390 		} else if (is_hash_order(ctx->pos)) {
391 			req->r_args.readdir.offset_hash =
392 				cpu_to_le32(fpos_hash(ctx->pos));
393 		}
394 
395 		req->r_dir_release_cnt = fi->dir_release_count;
396 		req->r_dir_ordered_cnt = fi->dir_ordered_count;
397 		req->r_readdir_cache_idx = fi->readdir_cache_idx;
398 		req->r_readdir_offset = fi->next_offset;
399 		req->r_args.readdir.frag = cpu_to_le32(frag);
400 		req->r_args.readdir.flags =
401 				cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
402 
403 		req->r_inode = inode;
404 		ihold(inode);
405 		req->r_dentry = dget(file->f_path.dentry);
406 		err = ceph_mdsc_do_request(mdsc, NULL, req);
407 		if (err < 0) {
408 			ceph_mdsc_put_request(req);
409 			return err;
410 		}
411 		dout("readdir got and parsed readdir result=%d on "
412 		     "frag %x, end=%d, complete=%d, hash_order=%d\n",
413 		     err, frag,
414 		     (int)req->r_reply_info.dir_end,
415 		     (int)req->r_reply_info.dir_complete,
416 		     (int)req->r_reply_info.hash_order);
417 
418 		rinfo = &req->r_reply_info;
419 		if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
420 			frag = le32_to_cpu(rinfo->dir_dir->frag);
421 			if (!rinfo->hash_order) {
422 				fi->next_offset = req->r_readdir_offset;
423 				/* adjust ctx->pos to beginning of frag */
424 				ctx->pos = ceph_make_fpos(frag,
425 							  fi->next_offset,
426 							  false);
427 			}
428 		}
429 
430 		fi->frag = frag;
431 		fi->last_readdir = req;
432 
433 		if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
434 			fi->readdir_cache_idx = req->r_readdir_cache_idx;
435 			if (fi->readdir_cache_idx < 0) {
436 				/* preclude from marking dir ordered */
437 				fi->dir_ordered_count = 0;
438 			} else if (ceph_frag_is_leftmost(frag) &&
439 				   fi->next_offset == 2) {
440 				/* note dir version at start of readdir so
441 				 * we can tell if any dentries get dropped */
442 				fi->dir_release_count = req->r_dir_release_cnt;
443 				fi->dir_ordered_count = req->r_dir_ordered_cnt;
444 			}
445 		} else {
446 			dout("readdir !did_prepopulate");
447 			/* disable readdir cache */
448 			fi->readdir_cache_idx = -1;
449 			/* preclude from marking dir complete */
450 			fi->dir_release_count = 0;
451 		}
452 
453 		/* note next offset and last dentry name */
454 		if (rinfo->dir_nr > 0) {
455 			struct ceph_mds_reply_dir_entry *rde =
456 					rinfo->dir_entries + (rinfo->dir_nr-1);
457 			unsigned next_offset = req->r_reply_info.dir_end ?
458 					2 : (fpos_off(rde->offset) + 1);
459 			err = note_last_dentry(fi, rde->name, rde->name_len,
460 					       next_offset);
461 			if (err)
462 				return err;
463 		} else if (req->r_reply_info.dir_end) {
464 			fi->next_offset = 2;
465 			/* keep last name */
466 		}
467 	}
468 
469 	rinfo = &fi->last_readdir->r_reply_info;
470 	dout("readdir frag %x num %d pos %llx chunk first %llx\n",
471 	     fi->frag, rinfo->dir_nr, ctx->pos,
472 	     rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
473 
474 	i = 0;
475 	/* search start position */
476 	if (rinfo->dir_nr > 0) {
477 		int step, nr = rinfo->dir_nr;
478 		while (nr > 0) {
479 			step = nr >> 1;
480 			if (rinfo->dir_entries[i + step].offset < ctx->pos) {
481 				i +=  step + 1;
482 				nr -= step + 1;
483 			} else {
484 				nr = step;
485 			}
486 		}
487 	}
488 	for (; i < rinfo->dir_nr; i++) {
489 		struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
490 		struct ceph_vino vino;
491 		ino_t ino;
492 		u32 ftype;
493 
494 		BUG_ON(rde->offset < ctx->pos);
495 
496 		ctx->pos = rde->offset;
497 		dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
498 		     i, rinfo->dir_nr, ctx->pos,
499 		     rde->name_len, rde->name, &rde->inode.in);
500 
501 		BUG_ON(!rde->inode.in);
502 		ftype = le32_to_cpu(rde->inode.in->mode) >> 12;
503 		vino.ino = le64_to_cpu(rde->inode.in->ino);
504 		vino.snap = le64_to_cpu(rde->inode.in->snapid);
505 		ino = ceph_vino_to_ino(vino);
506 
507 		if (!dir_emit(ctx, rde->name, rde->name_len,
508 			      ceph_translate_ino(inode->i_sb, ino), ftype)) {
509 			dout("filldir stopping us...\n");
510 			return 0;
511 		}
512 		ctx->pos++;
513 	}
514 
515 	ceph_mdsc_put_request(fi->last_readdir);
516 	fi->last_readdir = NULL;
517 
518 	if (fi->next_offset > 2) {
519 		frag = fi->frag;
520 		goto more;
521 	}
522 
523 	/* more frags? */
524 	if (!ceph_frag_is_rightmost(fi->frag)) {
525 		frag = ceph_frag_next(fi->frag);
526 		if (is_hash_order(ctx->pos)) {
527 			loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
528 							fi->next_offset, true);
529 			if (new_pos > ctx->pos)
530 				ctx->pos = new_pos;
531 			/* keep last_name */
532 		} else {
533 			ctx->pos = ceph_make_fpos(frag, fi->next_offset, false);
534 			kfree(fi->last_name);
535 			fi->last_name = NULL;
536 		}
537 		dout("readdir next frag is %x\n", frag);
538 		goto more;
539 	}
540 	fi->flags |= CEPH_F_ATEND;
541 
542 	/*
543 	 * if dir_release_count still matches the dir, no dentries
544 	 * were released during the whole readdir, and we should have
545 	 * the complete dir contents in our cache.
546 	 */
547 	if (atomic64_read(&ci->i_release_count) == fi->dir_release_count) {
548 		spin_lock(&ci->i_ceph_lock);
549 		if (fi->dir_ordered_count == atomic64_read(&ci->i_ordered_count)) {
550 			dout(" marking %p complete and ordered\n", inode);
551 			/* use i_size to track number of entries in
552 			 * readdir cache */
553 			BUG_ON(fi->readdir_cache_idx < 0);
554 			i_size_write(inode, fi->readdir_cache_idx *
555 				     sizeof(struct dentry*));
556 		} else {
557 			dout(" marking %p complete\n", inode);
558 		}
559 		__ceph_dir_set_complete(ci, fi->dir_release_count,
560 					fi->dir_ordered_count);
561 		spin_unlock(&ci->i_ceph_lock);
562 	}
563 
564 	dout("readdir %p file %p done.\n", inode, file);
565 	return 0;
566 }
567 
568 static void reset_readdir(struct ceph_file_info *fi)
569 {
570 	if (fi->last_readdir) {
571 		ceph_mdsc_put_request(fi->last_readdir);
572 		fi->last_readdir = NULL;
573 	}
574 	kfree(fi->last_name);
575 	fi->last_name = NULL;
576 	fi->dir_release_count = 0;
577 	fi->readdir_cache_idx = -1;
578 	fi->next_offset = 2;  /* compensate for . and .. */
579 	fi->flags &= ~CEPH_F_ATEND;
580 }
581 
582 /*
583  * discard buffered readdir content on seekdir(0), or seek to new frag,
584  * or seek prior to current chunk
585  */
586 static bool need_reset_readdir(struct ceph_file_info *fi, loff_t new_pos)
587 {
588 	struct ceph_mds_reply_info_parsed *rinfo;
589 	loff_t chunk_offset;
590 	if (new_pos == 0)
591 		return true;
592 	if (is_hash_order(new_pos)) {
593 		/* no need to reset last_name for a forward seek when
594 		 * dentries are sotred in hash order */
595 	} else if (fi->frag != fpos_frag(new_pos)) {
596 		return true;
597 	}
598 	rinfo = fi->last_readdir ? &fi->last_readdir->r_reply_info : NULL;
599 	if (!rinfo || !rinfo->dir_nr)
600 		return true;
601 	chunk_offset = rinfo->dir_entries[0].offset;
602 	return new_pos < chunk_offset ||
603 	       is_hash_order(new_pos) != is_hash_order(chunk_offset);
604 }
605 
606 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
607 {
608 	struct ceph_file_info *fi = file->private_data;
609 	struct inode *inode = file->f_mapping->host;
610 	loff_t retval;
611 
612 	inode_lock(inode);
613 	retval = -EINVAL;
614 	switch (whence) {
615 	case SEEK_CUR:
616 		offset += file->f_pos;
617 	case SEEK_SET:
618 		break;
619 	case SEEK_END:
620 		retval = -EOPNOTSUPP;
621 	default:
622 		goto out;
623 	}
624 
625 	if (offset >= 0) {
626 		if (need_reset_readdir(fi, offset)) {
627 			dout("dir_llseek dropping %p content\n", file);
628 			reset_readdir(fi);
629 		} else if (is_hash_order(offset) && offset > file->f_pos) {
630 			/* for hash offset, we don't know if a forward seek
631 			 * is within same frag */
632 			fi->dir_release_count = 0;
633 			fi->readdir_cache_idx = -1;
634 		}
635 
636 		if (offset != file->f_pos) {
637 			file->f_pos = offset;
638 			file->f_version = 0;
639 			fi->flags &= ~CEPH_F_ATEND;
640 		}
641 		retval = offset;
642 	}
643 out:
644 	inode_unlock(inode);
645 	return retval;
646 }
647 
648 /*
649  * Handle lookups for the hidden .snap directory.
650  */
651 int ceph_handle_snapdir(struct ceph_mds_request *req,
652 			struct dentry *dentry, int err)
653 {
654 	struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
655 	struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
656 
657 	/* .snap dir? */
658 	if (err == -ENOENT &&
659 	    ceph_snap(parent) == CEPH_NOSNAP &&
660 	    strcmp(dentry->d_name.name,
661 		   fsc->mount_options->snapdir_name) == 0) {
662 		struct inode *inode = ceph_get_snapdir(parent);
663 		dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
664 		     dentry, dentry, inode);
665 		BUG_ON(!d_unhashed(dentry));
666 		d_add(dentry, inode);
667 		err = 0;
668 	}
669 	return err;
670 }
671 
672 /*
673  * Figure out final result of a lookup/open request.
674  *
675  * Mainly, make sure we return the final req->r_dentry (if it already
676  * existed) in place of the original VFS-provided dentry when they
677  * differ.
678  *
679  * Gracefully handle the case where the MDS replies with -ENOENT and
680  * no trace (which it may do, at its discretion, e.g., if it doesn't
681  * care to issue a lease on the negative dentry).
682  */
683 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
684 				  struct dentry *dentry, int err)
685 {
686 	if (err == -ENOENT) {
687 		/* no trace? */
688 		err = 0;
689 		if (!req->r_reply_info.head->is_dentry) {
690 			dout("ENOENT and no trace, dentry %p inode %p\n",
691 			     dentry, d_inode(dentry));
692 			if (d_really_is_positive(dentry)) {
693 				d_drop(dentry);
694 				err = -ENOENT;
695 			} else {
696 				d_add(dentry, NULL);
697 			}
698 		}
699 	}
700 	if (err)
701 		dentry = ERR_PTR(err);
702 	else if (dentry != req->r_dentry)
703 		dentry = dget(req->r_dentry);   /* we got spliced */
704 	else
705 		dentry = NULL;
706 	return dentry;
707 }
708 
709 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
710 {
711 	return ceph_ino(inode) == CEPH_INO_ROOT &&
712 		strncmp(dentry->d_name.name, ".ceph", 5) == 0;
713 }
714 
715 /*
716  * Look up a single dir entry.  If there is a lookup intent, inform
717  * the MDS so that it gets our 'caps wanted' value in a single op.
718  */
719 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
720 				  unsigned int flags)
721 {
722 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
723 	struct ceph_mds_client *mdsc = fsc->mdsc;
724 	struct ceph_mds_request *req;
725 	int op;
726 	int mask;
727 	int err;
728 
729 	dout("lookup %p dentry %p '%pd'\n",
730 	     dir, dentry, dentry);
731 
732 	if (dentry->d_name.len > NAME_MAX)
733 		return ERR_PTR(-ENAMETOOLONG);
734 
735 	/* can we conclude ENOENT locally? */
736 	if (d_really_is_negative(dentry)) {
737 		struct ceph_inode_info *ci = ceph_inode(dir);
738 		struct ceph_dentry_info *di = ceph_dentry(dentry);
739 
740 		spin_lock(&ci->i_ceph_lock);
741 		dout(" dir %p flags are %d\n", dir, ci->i_ceph_flags);
742 		if (strncmp(dentry->d_name.name,
743 			    fsc->mount_options->snapdir_name,
744 			    dentry->d_name.len) &&
745 		    !is_root_ceph_dentry(dir, dentry) &&
746 		    ceph_test_mount_opt(fsc, DCACHE) &&
747 		    __ceph_dir_is_complete(ci) &&
748 		    (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) {
749 			spin_unlock(&ci->i_ceph_lock);
750 			dout(" dir %p complete, -ENOENT\n", dir);
751 			d_add(dentry, NULL);
752 			di->lease_shared_gen = ci->i_shared_gen;
753 			return NULL;
754 		}
755 		spin_unlock(&ci->i_ceph_lock);
756 	}
757 
758 	op = ceph_snap(dir) == CEPH_SNAPDIR ?
759 		CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
760 	req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
761 	if (IS_ERR(req))
762 		return ERR_CAST(req);
763 	req->r_dentry = dget(dentry);
764 	req->r_num_caps = 2;
765 
766 	mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
767 	if (ceph_security_xattr_wanted(dir))
768 		mask |= CEPH_CAP_XATTR_SHARED;
769 	req->r_args.getattr.mask = cpu_to_le32(mask);
770 
771 	req->r_parent = dir;
772 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
773 	err = ceph_mdsc_do_request(mdsc, NULL, req);
774 	err = ceph_handle_snapdir(req, dentry, err);
775 	dentry = ceph_finish_lookup(req, dentry, err);
776 	ceph_mdsc_put_request(req);  /* will dput(dentry) */
777 	dout("lookup result=%p\n", dentry);
778 	return dentry;
779 }
780 
781 /*
782  * If we do a create but get no trace back from the MDS, follow up with
783  * a lookup (the VFS expects us to link up the provided dentry).
784  */
785 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
786 {
787 	struct dentry *result = ceph_lookup(dir, dentry, 0);
788 
789 	if (result && !IS_ERR(result)) {
790 		/*
791 		 * We created the item, then did a lookup, and found
792 		 * it was already linked to another inode we already
793 		 * had in our cache (and thus got spliced). To not
794 		 * confuse VFS (especially when inode is a directory),
795 		 * we don't link our dentry to that inode, return an
796 		 * error instead.
797 		 *
798 		 * This event should be rare and it happens only when
799 		 * we talk to old MDS. Recent MDS does not send traceless
800 		 * reply for request that creates new inode.
801 		 */
802 		d_drop(result);
803 		return -ESTALE;
804 	}
805 	return PTR_ERR(result);
806 }
807 
808 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
809 		      umode_t mode, dev_t rdev)
810 {
811 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
812 	struct ceph_mds_client *mdsc = fsc->mdsc;
813 	struct ceph_mds_request *req;
814 	struct ceph_acls_info acls = {};
815 	int err;
816 
817 	if (ceph_snap(dir) != CEPH_NOSNAP)
818 		return -EROFS;
819 
820 	err = ceph_pre_init_acls(dir, &mode, &acls);
821 	if (err < 0)
822 		return err;
823 
824 	dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
825 	     dir, dentry, mode, rdev);
826 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
827 	if (IS_ERR(req)) {
828 		err = PTR_ERR(req);
829 		goto out;
830 	}
831 	req->r_dentry = dget(dentry);
832 	req->r_num_caps = 2;
833 	req->r_parent = dir;
834 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
835 	req->r_args.mknod.mode = cpu_to_le32(mode);
836 	req->r_args.mknod.rdev = cpu_to_le32(rdev);
837 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
838 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
839 	if (acls.pagelist) {
840 		req->r_pagelist = acls.pagelist;
841 		acls.pagelist = NULL;
842 	}
843 	err = ceph_mdsc_do_request(mdsc, dir, req);
844 	if (!err && !req->r_reply_info.head->is_dentry)
845 		err = ceph_handle_notrace_create(dir, dentry);
846 	ceph_mdsc_put_request(req);
847 out:
848 	if (!err)
849 		ceph_init_inode_acls(d_inode(dentry), &acls);
850 	else
851 		d_drop(dentry);
852 	ceph_release_acls_info(&acls);
853 	return err;
854 }
855 
856 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
857 		       bool excl)
858 {
859 	return ceph_mknod(dir, dentry, mode, 0);
860 }
861 
862 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
863 			    const char *dest)
864 {
865 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
866 	struct ceph_mds_client *mdsc = fsc->mdsc;
867 	struct ceph_mds_request *req;
868 	int err;
869 
870 	if (ceph_snap(dir) != CEPH_NOSNAP)
871 		return -EROFS;
872 
873 	dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
874 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
875 	if (IS_ERR(req)) {
876 		err = PTR_ERR(req);
877 		goto out;
878 	}
879 	req->r_path2 = kstrdup(dest, GFP_KERNEL);
880 	if (!req->r_path2) {
881 		err = -ENOMEM;
882 		ceph_mdsc_put_request(req);
883 		goto out;
884 	}
885 	req->r_parent = dir;
886 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
887 	req->r_dentry = dget(dentry);
888 	req->r_num_caps = 2;
889 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
890 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
891 	err = ceph_mdsc_do_request(mdsc, dir, req);
892 	if (!err && !req->r_reply_info.head->is_dentry)
893 		err = ceph_handle_notrace_create(dir, dentry);
894 	ceph_mdsc_put_request(req);
895 out:
896 	if (err)
897 		d_drop(dentry);
898 	return err;
899 }
900 
901 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
902 {
903 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
904 	struct ceph_mds_client *mdsc = fsc->mdsc;
905 	struct ceph_mds_request *req;
906 	struct ceph_acls_info acls = {};
907 	int err = -EROFS;
908 	int op;
909 
910 	if (ceph_snap(dir) == CEPH_SNAPDIR) {
911 		/* mkdir .snap/foo is a MKSNAP */
912 		op = CEPH_MDS_OP_MKSNAP;
913 		dout("mksnap dir %p snap '%pd' dn %p\n", dir,
914 		     dentry, dentry);
915 	} else if (ceph_snap(dir) == CEPH_NOSNAP) {
916 		dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
917 		op = CEPH_MDS_OP_MKDIR;
918 	} else {
919 		goto out;
920 	}
921 
922 	mode |= S_IFDIR;
923 	err = ceph_pre_init_acls(dir, &mode, &acls);
924 	if (err < 0)
925 		goto out;
926 
927 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
928 	if (IS_ERR(req)) {
929 		err = PTR_ERR(req);
930 		goto out;
931 	}
932 
933 	req->r_dentry = dget(dentry);
934 	req->r_num_caps = 2;
935 	req->r_parent = dir;
936 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
937 	req->r_args.mkdir.mode = cpu_to_le32(mode);
938 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
939 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
940 	if (acls.pagelist) {
941 		req->r_pagelist = acls.pagelist;
942 		acls.pagelist = NULL;
943 	}
944 	err = ceph_mdsc_do_request(mdsc, dir, req);
945 	if (!err &&
946 	    !req->r_reply_info.head->is_target &&
947 	    !req->r_reply_info.head->is_dentry)
948 		err = ceph_handle_notrace_create(dir, dentry);
949 	ceph_mdsc_put_request(req);
950 out:
951 	if (!err)
952 		ceph_init_inode_acls(d_inode(dentry), &acls);
953 	else
954 		d_drop(dentry);
955 	ceph_release_acls_info(&acls);
956 	return err;
957 }
958 
959 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
960 		     struct dentry *dentry)
961 {
962 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
963 	struct ceph_mds_client *mdsc = fsc->mdsc;
964 	struct ceph_mds_request *req;
965 	int err;
966 
967 	if (ceph_snap(dir) != CEPH_NOSNAP)
968 		return -EROFS;
969 
970 	dout("link in dir %p old_dentry %p dentry %p\n", dir,
971 	     old_dentry, dentry);
972 	req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
973 	if (IS_ERR(req)) {
974 		d_drop(dentry);
975 		return PTR_ERR(req);
976 	}
977 	req->r_dentry = dget(dentry);
978 	req->r_num_caps = 2;
979 	req->r_old_dentry = dget(old_dentry);
980 	req->r_parent = dir;
981 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
982 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
983 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
984 	/* release LINK_SHARED on source inode (mds will lock it) */
985 	req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
986 	err = ceph_mdsc_do_request(mdsc, dir, req);
987 	if (err) {
988 		d_drop(dentry);
989 	} else if (!req->r_reply_info.head->is_dentry) {
990 		ihold(d_inode(old_dentry));
991 		d_instantiate(dentry, d_inode(old_dentry));
992 	}
993 	ceph_mdsc_put_request(req);
994 	return err;
995 }
996 
997 /*
998  * For a soon-to-be unlinked file, drop the AUTH_RDCACHE caps.  If it
999  * looks like the link count will hit 0, drop any other caps (other
1000  * than PIN) we don't specifically want (due to the file still being
1001  * open).
1002  */
1003 static int drop_caps_for_unlink(struct inode *inode)
1004 {
1005 	struct ceph_inode_info *ci = ceph_inode(inode);
1006 	int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1007 
1008 	spin_lock(&ci->i_ceph_lock);
1009 	if (inode->i_nlink == 1) {
1010 		drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
1011 		ci->i_ceph_flags |= CEPH_I_NODELAY;
1012 	}
1013 	spin_unlock(&ci->i_ceph_lock);
1014 	return drop;
1015 }
1016 
1017 /*
1018  * rmdir and unlink are differ only by the metadata op code
1019  */
1020 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1021 {
1022 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1023 	struct ceph_mds_client *mdsc = fsc->mdsc;
1024 	struct inode *inode = d_inode(dentry);
1025 	struct ceph_mds_request *req;
1026 	int err = -EROFS;
1027 	int op;
1028 
1029 	if (ceph_snap(dir) == CEPH_SNAPDIR) {
1030 		/* rmdir .snap/foo is RMSNAP */
1031 		dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1032 		op = CEPH_MDS_OP_RMSNAP;
1033 	} else if (ceph_snap(dir) == CEPH_NOSNAP) {
1034 		dout("unlink/rmdir dir %p dn %p inode %p\n",
1035 		     dir, dentry, inode);
1036 		op = d_is_dir(dentry) ?
1037 			CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1038 	} else
1039 		goto out;
1040 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1041 	if (IS_ERR(req)) {
1042 		err = PTR_ERR(req);
1043 		goto out;
1044 	}
1045 	req->r_dentry = dget(dentry);
1046 	req->r_num_caps = 2;
1047 	req->r_parent = dir;
1048 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1049 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1050 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1051 	req->r_inode_drop = drop_caps_for_unlink(inode);
1052 	err = ceph_mdsc_do_request(mdsc, dir, req);
1053 	if (!err && !req->r_reply_info.head->is_dentry)
1054 		d_delete(dentry);
1055 	ceph_mdsc_put_request(req);
1056 out:
1057 	return err;
1058 }
1059 
1060 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
1061 		       struct inode *new_dir, struct dentry *new_dentry,
1062 		       unsigned int flags)
1063 {
1064 	struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
1065 	struct ceph_mds_client *mdsc = fsc->mdsc;
1066 	struct ceph_mds_request *req;
1067 	int op = CEPH_MDS_OP_RENAME;
1068 	int err;
1069 
1070 	if (flags)
1071 		return -EINVAL;
1072 
1073 	if (ceph_snap(old_dir) != ceph_snap(new_dir))
1074 		return -EXDEV;
1075 	if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1076 		if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1077 			op = CEPH_MDS_OP_RENAMESNAP;
1078 		else
1079 			return -EROFS;
1080 	}
1081 	dout("rename dir %p dentry %p to dir %p dentry %p\n",
1082 	     old_dir, old_dentry, new_dir, new_dentry);
1083 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1084 	if (IS_ERR(req))
1085 		return PTR_ERR(req);
1086 	ihold(old_dir);
1087 	req->r_dentry = dget(new_dentry);
1088 	req->r_num_caps = 2;
1089 	req->r_old_dentry = dget(old_dentry);
1090 	req->r_old_dentry_dir = old_dir;
1091 	req->r_parent = new_dir;
1092 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1093 	req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1094 	req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1095 	req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1096 	req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1097 	/* release LINK_RDCACHE on source inode (mds will lock it) */
1098 	req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
1099 	if (d_really_is_positive(new_dentry))
1100 		req->r_inode_drop = drop_caps_for_unlink(d_inode(new_dentry));
1101 	err = ceph_mdsc_do_request(mdsc, old_dir, req);
1102 	if (!err && !req->r_reply_info.head->is_dentry) {
1103 		/*
1104 		 * Normally d_move() is done by fill_trace (called by
1105 		 * do_request, above).  If there is no trace, we need
1106 		 * to do it here.
1107 		 */
1108 
1109 		/* d_move screws up sibling dentries' offsets */
1110 		ceph_dir_clear_complete(old_dir);
1111 		ceph_dir_clear_complete(new_dir);
1112 
1113 		d_move(old_dentry, new_dentry);
1114 
1115 		/* ensure target dentry is invalidated, despite
1116 		   rehashing bug in vfs_rename_dir */
1117 		ceph_invalidate_dentry_lease(new_dentry);
1118 	}
1119 	ceph_mdsc_put_request(req);
1120 	return err;
1121 }
1122 
1123 /*
1124  * Ensure a dentry lease will no longer revalidate.
1125  */
1126 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1127 {
1128 	spin_lock(&dentry->d_lock);
1129 	ceph_dentry(dentry)->time = jiffies;
1130 	ceph_dentry(dentry)->lease_shared_gen = 0;
1131 	spin_unlock(&dentry->d_lock);
1132 }
1133 
1134 /*
1135  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1136  * renew if the least is more than half up.
1137  */
1138 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags,
1139 				 struct inode *dir)
1140 {
1141 	struct ceph_dentry_info *di;
1142 	struct ceph_mds_session *s;
1143 	int valid = 0;
1144 	u32 gen;
1145 	unsigned long ttl;
1146 	struct ceph_mds_session *session = NULL;
1147 	u32 seq = 0;
1148 
1149 	spin_lock(&dentry->d_lock);
1150 	di = ceph_dentry(dentry);
1151 	if (di && di->lease_session) {
1152 		s = di->lease_session;
1153 		spin_lock(&s->s_gen_ttl_lock);
1154 		gen = s->s_cap_gen;
1155 		ttl = s->s_cap_ttl;
1156 		spin_unlock(&s->s_gen_ttl_lock);
1157 
1158 		if (di->lease_gen == gen &&
1159 		    time_before(jiffies, di->time) &&
1160 		    time_before(jiffies, ttl)) {
1161 			valid = 1;
1162 			if (di->lease_renew_after &&
1163 			    time_after(jiffies, di->lease_renew_after)) {
1164 				/*
1165 				 * We should renew. If we're in RCU walk mode
1166 				 * though, we can't do that so just return
1167 				 * -ECHILD.
1168 				 */
1169 				if (flags & LOOKUP_RCU) {
1170 					valid = -ECHILD;
1171 				} else {
1172 					session = ceph_get_mds_session(s);
1173 					seq = di->lease_seq;
1174 					di->lease_renew_after = 0;
1175 					di->lease_renew_from = jiffies;
1176 				}
1177 			}
1178 		}
1179 	}
1180 	spin_unlock(&dentry->d_lock);
1181 
1182 	if (session) {
1183 		ceph_mdsc_lease_send_msg(session, dir, dentry,
1184 					 CEPH_MDS_LEASE_RENEW, seq);
1185 		ceph_put_mds_session(session);
1186 	}
1187 	dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1188 	return valid;
1189 }
1190 
1191 /*
1192  * Check if directory-wide content lease/cap is valid.
1193  */
1194 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry)
1195 {
1196 	struct ceph_inode_info *ci = ceph_inode(dir);
1197 	struct ceph_dentry_info *di = ceph_dentry(dentry);
1198 	int valid = 0;
1199 
1200 	spin_lock(&ci->i_ceph_lock);
1201 	if (ci->i_shared_gen == di->lease_shared_gen)
1202 		valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1203 	spin_unlock(&ci->i_ceph_lock);
1204 	dout("dir_lease_is_valid dir %p v%u dentry %p v%u = %d\n",
1205 	     dir, (unsigned)ci->i_shared_gen, dentry,
1206 	     (unsigned)di->lease_shared_gen, valid);
1207 	return valid;
1208 }
1209 
1210 /*
1211  * Check if cached dentry can be trusted.
1212  */
1213 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1214 {
1215 	int valid = 0;
1216 	struct dentry *parent;
1217 	struct inode *dir;
1218 
1219 	if (flags & LOOKUP_RCU) {
1220 		parent = READ_ONCE(dentry->d_parent);
1221 		dir = d_inode_rcu(parent);
1222 		if (!dir)
1223 			return -ECHILD;
1224 	} else {
1225 		parent = dget_parent(dentry);
1226 		dir = d_inode(parent);
1227 	}
1228 
1229 	dout("d_revalidate %p '%pd' inode %p offset %lld\n", dentry,
1230 	     dentry, d_inode(dentry), ceph_dentry(dentry)->offset);
1231 
1232 	/* always trust cached snapped dentries, snapdir dentry */
1233 	if (ceph_snap(dir) != CEPH_NOSNAP) {
1234 		dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1235 		     dentry, d_inode(dentry));
1236 		valid = 1;
1237 	} else if (d_really_is_positive(dentry) &&
1238 		   ceph_snap(d_inode(dentry)) == CEPH_SNAPDIR) {
1239 		valid = 1;
1240 	} else {
1241 		valid = dentry_lease_is_valid(dentry, flags, dir);
1242 		if (valid == -ECHILD)
1243 			return valid;
1244 		if (valid || dir_lease_is_valid(dir, dentry)) {
1245 			if (d_really_is_positive(dentry))
1246 				valid = ceph_is_any_caps(d_inode(dentry));
1247 			else
1248 				valid = 1;
1249 		}
1250 	}
1251 
1252 	if (!valid) {
1253 		struct ceph_mds_client *mdsc =
1254 			ceph_sb_to_client(dir->i_sb)->mdsc;
1255 		struct ceph_mds_request *req;
1256 		int op, err;
1257 		u32 mask;
1258 
1259 		if (flags & LOOKUP_RCU)
1260 			return -ECHILD;
1261 
1262 		op = ceph_snap(dir) == CEPH_SNAPDIR ?
1263 			CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1264 		req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1265 		if (!IS_ERR(req)) {
1266 			req->r_dentry = dget(dentry);
1267 			req->r_num_caps = 2;
1268 			req->r_parent = dir;
1269 
1270 			mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1271 			if (ceph_security_xattr_wanted(dir))
1272 				mask |= CEPH_CAP_XATTR_SHARED;
1273 			req->r_args.getattr.mask = cpu_to_le32(mask);
1274 
1275 			err = ceph_mdsc_do_request(mdsc, NULL, req);
1276 			switch (err) {
1277 			case 0:
1278 				if (d_really_is_positive(dentry) &&
1279 				    d_inode(dentry) == req->r_target_inode)
1280 					valid = 1;
1281 				break;
1282 			case -ENOENT:
1283 				if (d_really_is_negative(dentry))
1284 					valid = 1;
1285 				/* Fallthrough */
1286 			default:
1287 				break;
1288 			}
1289 			ceph_mdsc_put_request(req);
1290 			dout("d_revalidate %p lookup result=%d\n",
1291 			     dentry, err);
1292 		}
1293 	}
1294 
1295 	dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1296 	if (valid) {
1297 		ceph_dentry_lru_touch(dentry);
1298 	} else {
1299 		ceph_dir_clear_complete(dir);
1300 	}
1301 
1302 	if (!(flags & LOOKUP_RCU))
1303 		dput(parent);
1304 	return valid;
1305 }
1306 
1307 /*
1308  * Release our ceph_dentry_info.
1309  */
1310 static void ceph_d_release(struct dentry *dentry)
1311 {
1312 	struct ceph_dentry_info *di = ceph_dentry(dentry);
1313 
1314 	dout("d_release %p\n", dentry);
1315 	ceph_dentry_lru_del(dentry);
1316 
1317 	spin_lock(&dentry->d_lock);
1318 	dentry->d_fsdata = NULL;
1319 	spin_unlock(&dentry->d_lock);
1320 
1321 	if (di->lease_session)
1322 		ceph_put_mds_session(di->lease_session);
1323 	kmem_cache_free(ceph_dentry_cachep, di);
1324 }
1325 
1326 /*
1327  * When the VFS prunes a dentry from the cache, we need to clear the
1328  * complete flag on the parent directory.
1329  *
1330  * Called under dentry->d_lock.
1331  */
1332 static void ceph_d_prune(struct dentry *dentry)
1333 {
1334 	dout("ceph_d_prune %p\n", dentry);
1335 
1336 	/* do we have a valid parent? */
1337 	if (IS_ROOT(dentry))
1338 		return;
1339 
1340 	/* if we are not hashed, we don't affect dir's completeness */
1341 	if (d_unhashed(dentry))
1342 		return;
1343 
1344 	if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_SNAPDIR)
1345 		return;
1346 
1347 	/*
1348 	 * we hold d_lock, so d_parent is stable, and d_fsdata is never
1349 	 * cleared until d_release
1350 	 */
1351 	ceph_dir_clear_complete(d_inode(dentry->d_parent));
1352 }
1353 
1354 /*
1355  * read() on a dir.  This weird interface hack only works if mounted
1356  * with '-o dirstat'.
1357  */
1358 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1359 			     loff_t *ppos)
1360 {
1361 	struct ceph_file_info *cf = file->private_data;
1362 	struct inode *inode = file_inode(file);
1363 	struct ceph_inode_info *ci = ceph_inode(inode);
1364 	int left;
1365 	const int bufsize = 1024;
1366 
1367 	if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1368 		return -EISDIR;
1369 
1370 	if (!cf->dir_info) {
1371 		cf->dir_info = kmalloc(bufsize, GFP_KERNEL);
1372 		if (!cf->dir_info)
1373 			return -ENOMEM;
1374 		cf->dir_info_len =
1375 			snprintf(cf->dir_info, bufsize,
1376 				"entries:   %20lld\n"
1377 				" files:    %20lld\n"
1378 				" subdirs:  %20lld\n"
1379 				"rentries:  %20lld\n"
1380 				" rfiles:   %20lld\n"
1381 				" rsubdirs: %20lld\n"
1382 				"rbytes:    %20lld\n"
1383 				"rctime:    %10ld.%09ld\n",
1384 				ci->i_files + ci->i_subdirs,
1385 				ci->i_files,
1386 				ci->i_subdirs,
1387 				ci->i_rfiles + ci->i_rsubdirs,
1388 				ci->i_rfiles,
1389 				ci->i_rsubdirs,
1390 				ci->i_rbytes,
1391 				(long)ci->i_rctime.tv_sec,
1392 				(long)ci->i_rctime.tv_nsec);
1393 	}
1394 
1395 	if (*ppos >= cf->dir_info_len)
1396 		return 0;
1397 	size = min_t(unsigned, size, cf->dir_info_len-*ppos);
1398 	left = copy_to_user(buf, cf->dir_info + *ppos, size);
1399 	if (left == size)
1400 		return -EFAULT;
1401 	*ppos += (size - left);
1402 	return size - left;
1403 }
1404 
1405 /*
1406  * We maintain a private dentry LRU.
1407  *
1408  * FIXME: this needs to be changed to a per-mds lru to be useful.
1409  */
1410 void ceph_dentry_lru_add(struct dentry *dn)
1411 {
1412 	struct ceph_dentry_info *di = ceph_dentry(dn);
1413 	struct ceph_mds_client *mdsc;
1414 
1415 	dout("dentry_lru_add %p %p '%pd'\n", di, dn, dn);
1416 	mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1417 	spin_lock(&mdsc->dentry_lru_lock);
1418 	list_add_tail(&di->lru, &mdsc->dentry_lru);
1419 	mdsc->num_dentry++;
1420 	spin_unlock(&mdsc->dentry_lru_lock);
1421 }
1422 
1423 void ceph_dentry_lru_touch(struct dentry *dn)
1424 {
1425 	struct ceph_dentry_info *di = ceph_dentry(dn);
1426 	struct ceph_mds_client *mdsc;
1427 
1428 	dout("dentry_lru_touch %p %p '%pd' (offset %lld)\n", di, dn, dn,
1429 	     di->offset);
1430 	mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1431 	spin_lock(&mdsc->dentry_lru_lock);
1432 	list_move_tail(&di->lru, &mdsc->dentry_lru);
1433 	spin_unlock(&mdsc->dentry_lru_lock);
1434 }
1435 
1436 void ceph_dentry_lru_del(struct dentry *dn)
1437 {
1438 	struct ceph_dentry_info *di = ceph_dentry(dn);
1439 	struct ceph_mds_client *mdsc;
1440 
1441 	dout("dentry_lru_del %p %p '%pd'\n", di, dn, dn);
1442 	mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1443 	spin_lock(&mdsc->dentry_lru_lock);
1444 	list_del_init(&di->lru);
1445 	mdsc->num_dentry--;
1446 	spin_unlock(&mdsc->dentry_lru_lock);
1447 }
1448 
1449 /*
1450  * Return name hash for a given dentry.  This is dependent on
1451  * the parent directory's hash function.
1452  */
1453 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1454 {
1455 	struct ceph_inode_info *dci = ceph_inode(dir);
1456 
1457 	switch (dci->i_dir_layout.dl_dir_hash) {
1458 	case 0:	/* for backward compat */
1459 	case CEPH_STR_HASH_LINUX:
1460 		return dn->d_name.hash;
1461 
1462 	default:
1463 		return ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1464 				     dn->d_name.name, dn->d_name.len);
1465 	}
1466 }
1467 
1468 const struct file_operations ceph_dir_fops = {
1469 	.read = ceph_read_dir,
1470 	.iterate = ceph_readdir,
1471 	.llseek = ceph_dir_llseek,
1472 	.open = ceph_open,
1473 	.release = ceph_release,
1474 	.unlocked_ioctl = ceph_ioctl,
1475 	.fsync = ceph_fsync,
1476 };
1477 
1478 const struct file_operations ceph_snapdir_fops = {
1479 	.iterate = ceph_readdir,
1480 	.llseek = ceph_dir_llseek,
1481 	.open = ceph_open,
1482 	.release = ceph_release,
1483 };
1484 
1485 const struct inode_operations ceph_dir_iops = {
1486 	.lookup = ceph_lookup,
1487 	.permission = ceph_permission,
1488 	.getattr = ceph_getattr,
1489 	.setattr = ceph_setattr,
1490 	.listxattr = ceph_listxattr,
1491 	.get_acl = ceph_get_acl,
1492 	.set_acl = ceph_set_acl,
1493 	.mknod = ceph_mknod,
1494 	.symlink = ceph_symlink,
1495 	.mkdir = ceph_mkdir,
1496 	.link = ceph_link,
1497 	.unlink = ceph_unlink,
1498 	.rmdir = ceph_unlink,
1499 	.rename = ceph_rename,
1500 	.create = ceph_create,
1501 	.atomic_open = ceph_atomic_open,
1502 };
1503 
1504 const struct inode_operations ceph_snapdir_iops = {
1505 	.lookup = ceph_lookup,
1506 	.permission = ceph_permission,
1507 	.getattr = ceph_getattr,
1508 	.mkdir = ceph_mkdir,
1509 	.rmdir = ceph_unlink,
1510 	.rename = ceph_rename,
1511 };
1512 
1513 const struct dentry_operations ceph_dentry_ops = {
1514 	.d_revalidate = ceph_d_revalidate,
1515 	.d_release = ceph_d_release,
1516 	.d_prune = ceph_d_prune,
1517 	.d_init = ceph_d_init,
1518 };
1519