xref: /titanic_50/usr/src/uts/common/fs/zfs/zfs_dir.c (revision 380789fc80376bd1573770361cb177a08c7e3524)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/types.h>
29 #include <sys/param.h>
30 #include <sys/time.h>
31 #include <sys/systm.h>
32 #include <sys/sysmacros.h>
33 #include <sys/resource.h>
34 #include <sys/vfs.h>
35 #include <sys/vnode.h>
36 #include <sys/file.h>
37 #include <sys/mode.h>
38 #include <sys/kmem.h>
39 #include <sys/uio.h>
40 #include <sys/pathname.h>
41 #include <sys/cmn_err.h>
42 #include <sys/errno.h>
43 #include <sys/stat.h>
44 #include <sys/unistd.h>
45 #include <sys/random.h>
46 #include <sys/policy.h>
47 #include <sys/zfs_dir.h>
48 #include <sys/zfs_acl.h>
49 #include <sys/fs/zfs.h>
50 #include "fs/fs_subr.h"
51 #include <sys/zap.h>
52 #include <sys/dmu.h>
53 #include <sys/atomic.h>
54 #include <sys/zfs_ctldir.h>
55 #include <sys/zfs_fuid.h>
56 #include <sys/dnlc.h>
57 #include <sys/extdirent.h>
58 #include <sys/zfs_i18n.h>
59 
60 /*
61  * zfs_match_find() is used by zfs_dirent_lock() to peform zap lookups
62  * of names after deciding which is the appropriate lookup interface.
63  */
64 static int
65 zfs_match_find(zfsvfs_t *zfsvfs, znode_t *dzp, char *name, boolean_t exact,
66     boolean_t update, int *deflags, pathname_t *rpnp, uint64_t *zoid)
67 {
68 	int error;
69 
70 	if (zfsvfs->z_norm) {
71 		matchtype_t mt = MT_FIRST;
72 		boolean_t conflict = B_FALSE;
73 		size_t bufsz = 0;
74 		char *buf = NULL;
75 
76 		if (rpnp) {
77 			buf = rpnp->pn_path;
78 			bufsz = rpnp->pn_bufsize;
79 		}
80 		if (exact)
81 			mt = MT_EXACT;
82 		/*
83 		 * In the non-mixed case we only expect there would ever
84 		 * be one match, but we need to use the normalizing lookup.
85 		 */
86 		error = zap_lookup_norm(zfsvfs->z_os, dzp->z_id, name, 8, 1,
87 		    zoid, mt, buf, bufsz, &conflict);
88 		if (deflags)
89 			*deflags = conflict ? ED_CASE_CONFLICT : 0;
90 	} else {
91 		error = zap_lookup(zfsvfs->z_os, dzp->z_id, name, 8, 1, zoid);
92 	}
93 	*zoid = ZFS_DIRENT_OBJ(*zoid);
94 
95 	if (error == ENOENT && update)
96 		dnlc_update(ZTOV(dzp), name, DNLC_NO_VNODE);
97 
98 	return (error);
99 }
100 
101 /*
102  * Lock a directory entry.  A dirlock on <dzp, name> protects that name
103  * in dzp's directory zap object.  As long as you hold a dirlock, you can
104  * assume two things: (1) dzp cannot be reaped, and (2) no other thread
105  * can change the zap entry for (i.e. link or unlink) this name.
106  *
107  * Input arguments:
108  *	dzp	- znode for directory
109  *	name	- name of entry to lock
110  *	flag	- ZNEW: if the entry already exists, fail with EEXIST.
111  *		  ZEXISTS: if the entry does not exist, fail with ENOENT.
112  *		  ZSHARED: allow concurrent access with other ZSHARED callers.
113  *		  ZXATTR: we want dzp's xattr directory
114  *		  ZCILOOK: On a mixed sensitivity file system,
115  *			   this lookup should be case-insensitive.
116  *		  ZCIEXACT: On a purely case-insensitive file system,
117  *			    this lookup should be case-sensitive.
118  *		  ZRENAMING: we are locking for renaming, force narrow locks
119  *
120  * Output arguments:
121  *	zpp	- pointer to the znode for the entry (NULL if there isn't one)
122  *	dlpp	- pointer to the dirlock for this entry (NULL on error)
123  *      direntflags - (case-insensitive lookup only)
124  *		flags if multiple case-sensitive matches exist in directory
125  *      realpnp     - (case-insensitive lookup only)
126  *		actual name matched within the directory
127  *
128  * Return value: 0 on success or errno on failure.
129  *
130  * NOTE: Always checks for, and rejects, '.' and '..'.
131  * NOTE: For case-insensitive file systems we take wide locks (see below),
132  *	 but return znode pointers to a single match.
133  */
134 int
135 zfs_dirent_lock(zfs_dirlock_t **dlpp, znode_t *dzp, char *name, znode_t **zpp,
136     int flag, int *direntflags, pathname_t *realpnp)
137 {
138 	zfsvfs_t	*zfsvfs = dzp->z_zfsvfs;
139 	zfs_dirlock_t	*dl;
140 	boolean_t	update;
141 	boolean_t	exact;
142 	uint64_t	zoid;
143 	vnode_t		*vp = NULL;
144 	int		error = 0;
145 	int		cmpflags;
146 
147 	*zpp = NULL;
148 	*dlpp = NULL;
149 
150 	/*
151 	 * Verify that we are not trying to lock '.', '..', or '.zfs'
152 	 */
153 	if (name[0] == '.' &&
154 	    (name[1] == '\0' || (name[1] == '.' && name[2] == '\0')) ||
155 	    zfs_has_ctldir(dzp) && strcmp(name, ZFS_CTLDIR_NAME) == 0)
156 		return (EEXIST);
157 
158 	/*
159 	 * Case sensitivity and normalization preferences are set when
160 	 * the file system is created.  These are stored in the
161 	 * zfsvfs->z_case and zfsvfs->z_norm fields.  These choices
162 	 * affect what vnodes can be cached in the DNLC, how we
163 	 * perform zap lookups, and the "width" of our dirlocks.
164 	 *
165 	 * A normal dirlock locks a single name.  Note that with
166 	 * normalization a name can be composed multiple ways, but
167 	 * when normalized, these names all compare equal.  A wide
168 	 * dirlock locks multiple names.  We need these when the file
169 	 * system is supporting mixed-mode access.  It is sometimes
170 	 * necessary to lock all case permutations of file name at
171 	 * once so that simultaneous case-insensitive/case-sensitive
172 	 * behaves as rationally as possible.
173 	 */
174 
175 	/*
176 	 * Decide if exact matches should be requested when performing
177 	 * a zap lookup on file systems supporting case-insensitive
178 	 * access.
179 	 */
180 	exact = ((zfsvfs->z_case & ZFS_CI_ONLY) && (flag & ZCIEXACT)) ||
181 	    ((zfsvfs->z_case & ZFS_CI_MIXD) && !(flag & ZCILOOK));
182 
183 	/*
184 	 * Only look in or update the DNLC if we are looking for the
185 	 * name on a file system that does not require normalization
186 	 * or case folding.  We can also look there if we happen to be
187 	 * on a non-normalizing, mixed sensitivity file system IF we
188 	 * are looking for the exact name.
189 	 *
190 	 * Maybe can add TO-UPPERed version of name to dnlc in ci-only
191 	 * case for performance improvement?
192 	 */
193 	update = !zfsvfs->z_norm ||
194 	    ((zfsvfs->z_case & ZFS_CI_MIXD) &&
195 	    !(zfsvfs->z_norm & ~U8_TEXTPREP_TOUPPER) && !(flag & ZCILOOK));
196 
197 	/*
198 	 * ZRENAMING indicates we are in a situation where we should
199 	 * take narrow locks regardless of the file system's
200 	 * preferences for normalizing and case folding.  This will
201 	 * prevent us deadlocking trying to grab the same wide lock
202 	 * twice if the two names happen to be case-insensitive
203 	 * matches.
204 	 */
205 	if (flag & ZRENAMING)
206 		cmpflags = 0;
207 	else
208 		cmpflags = zfsvfs->z_norm;
209 
210 	/*
211 	 * Wait until there are no locks on this name.
212 	 */
213 	rw_enter(&dzp->z_name_lock, RW_READER);
214 	mutex_enter(&dzp->z_lock);
215 	for (;;) {
216 		if (dzp->z_unlinked) {
217 			mutex_exit(&dzp->z_lock);
218 			rw_exit(&dzp->z_name_lock);
219 			return (ENOENT);
220 		}
221 		for (dl = dzp->z_dirlocks; dl != NULL; dl = dl->dl_next) {
222 			if ((u8_strcmp(name, dl->dl_name, 0, cmpflags,
223 			    U8_UNICODE_LATEST, &error) == 0) || error != 0)
224 				break;
225 		}
226 		if (error != 0) {
227 			mutex_exit(&dzp->z_lock);
228 			rw_exit(&dzp->z_name_lock);
229 			return (ENOENT);
230 		}
231 		if (dl == NULL)	{
232 			/*
233 			 * Allocate a new dirlock and add it to the list.
234 			 */
235 			dl = kmem_alloc(sizeof (zfs_dirlock_t), KM_SLEEP);
236 			cv_init(&dl->dl_cv, NULL, CV_DEFAULT, NULL);
237 			dl->dl_name = name;
238 			dl->dl_sharecnt = 0;
239 			dl->dl_namesize = 0;
240 			dl->dl_dzp = dzp;
241 			dl->dl_next = dzp->z_dirlocks;
242 			dzp->z_dirlocks = dl;
243 			break;
244 		}
245 		if ((flag & ZSHARED) && dl->dl_sharecnt != 0)
246 			break;
247 		cv_wait(&dl->dl_cv, &dzp->z_lock);
248 	}
249 
250 	if ((flag & ZSHARED) && ++dl->dl_sharecnt > 1 && dl->dl_namesize == 0) {
251 		/*
252 		 * We're the second shared reference to dl.  Make a copy of
253 		 * dl_name in case the first thread goes away before we do.
254 		 * Note that we initialize the new name before storing its
255 		 * pointer into dl_name, because the first thread may load
256 		 * dl->dl_name at any time.  He'll either see the old value,
257 		 * which is his, or the new shared copy; either is OK.
258 		 */
259 		dl->dl_namesize = strlen(dl->dl_name) + 1;
260 		name = kmem_alloc(dl->dl_namesize, KM_SLEEP);
261 		bcopy(dl->dl_name, name, dl->dl_namesize);
262 		dl->dl_name = name;
263 	}
264 
265 	mutex_exit(&dzp->z_lock);
266 
267 	/*
268 	 * We have a dirlock on the name.  (Note that it is the dirlock,
269 	 * not the dzp's z_lock, that protects the name in the zap object.)
270 	 * See if there's an object by this name; if so, put a hold on it.
271 	 */
272 	if (flag & ZXATTR) {
273 		zoid = dzp->z_phys->zp_xattr;
274 		error = (zoid == 0 ? ENOENT : 0);
275 	} else {
276 		if (update)
277 			vp = dnlc_lookup(ZTOV(dzp), name);
278 		if (vp == DNLC_NO_VNODE) {
279 			VN_RELE(vp);
280 			error = ENOENT;
281 		} else if (vp) {
282 			if (flag & ZNEW) {
283 				zfs_dirent_unlock(dl);
284 				VN_RELE(vp);
285 				return (EEXIST);
286 			}
287 			*dlpp = dl;
288 			*zpp = VTOZ(vp);
289 			return (0);
290 		} else {
291 			error = zfs_match_find(zfsvfs, dzp, name, exact,
292 			    update, direntflags, realpnp, &zoid);
293 		}
294 	}
295 	if (error) {
296 		if (error != ENOENT || (flag & ZEXISTS)) {
297 			zfs_dirent_unlock(dl);
298 			return (error);
299 		}
300 	} else {
301 		if (flag & ZNEW) {
302 			zfs_dirent_unlock(dl);
303 			return (EEXIST);
304 		}
305 		error = zfs_zget(zfsvfs, zoid, zpp);
306 		if (error) {
307 			zfs_dirent_unlock(dl);
308 			return (error);
309 		}
310 		if (!(flag & ZXATTR) && update)
311 			dnlc_update(ZTOV(dzp), name, ZTOV(*zpp));
312 	}
313 
314 	*dlpp = dl;
315 
316 	return (0);
317 }
318 
319 /*
320  * Unlock this directory entry and wake anyone who was waiting for it.
321  */
322 void
323 zfs_dirent_unlock(zfs_dirlock_t *dl)
324 {
325 	znode_t *dzp = dl->dl_dzp;
326 	zfs_dirlock_t **prev_dl, *cur_dl;
327 
328 	mutex_enter(&dzp->z_lock);
329 	rw_exit(&dzp->z_name_lock);
330 	if (dl->dl_sharecnt > 1) {
331 		dl->dl_sharecnt--;
332 		mutex_exit(&dzp->z_lock);
333 		return;
334 	}
335 	prev_dl = &dzp->z_dirlocks;
336 	while ((cur_dl = *prev_dl) != dl)
337 		prev_dl = &cur_dl->dl_next;
338 	*prev_dl = dl->dl_next;
339 	cv_broadcast(&dl->dl_cv);
340 	mutex_exit(&dzp->z_lock);
341 
342 	if (dl->dl_namesize != 0)
343 		kmem_free(dl->dl_name, dl->dl_namesize);
344 	cv_destroy(&dl->dl_cv);
345 	kmem_free(dl, sizeof (*dl));
346 }
347 
348 /*
349  * Look up an entry in a directory.
350  *
351  * NOTE: '.' and '..' are handled as special cases because
352  *	no directory entries are actually stored for them.  If this is
353  *	the root of a filesystem, then '.zfs' is also treated as a
354  *	special pseudo-directory.
355  */
356 int
357 zfs_dirlook(znode_t *dzp, char *name, vnode_t **vpp, int flags,
358     int *deflg, pathname_t *rpnp)
359 {
360 	zfs_dirlock_t *dl;
361 	znode_t *zp;
362 	int error = 0;
363 
364 	if (name[0] == 0 || (name[0] == '.' && name[1] == 0)) {
365 		*vpp = ZTOV(dzp);
366 		VN_HOLD(*vpp);
367 	} else if (name[0] == '.' && name[1] == '.' && name[2] == 0) {
368 		zfsvfs_t *zfsvfs = dzp->z_zfsvfs;
369 		/*
370 		 * If we are a snapshot mounted under .zfs, return
371 		 * the vp for the snapshot directory.
372 		 */
373 		if (dzp->z_phys->zp_parent == dzp->z_id &&
374 		    zfsvfs->z_parent != zfsvfs) {
375 			error = zfsctl_root_lookup(zfsvfs->z_parent->z_ctldir,
376 			    "snapshot", vpp, NULL, 0, NULL, kcred,
377 			    NULL, NULL, NULL);
378 			return (error);
379 		}
380 		rw_enter(&dzp->z_parent_lock, RW_READER);
381 		error = zfs_zget(zfsvfs, dzp->z_phys->zp_parent, &zp);
382 		if (error == 0)
383 			*vpp = ZTOV(zp);
384 		rw_exit(&dzp->z_parent_lock);
385 	} else if (zfs_has_ctldir(dzp) && strcmp(name, ZFS_CTLDIR_NAME) == 0) {
386 		*vpp = zfsctl_root(dzp);
387 	} else {
388 		int zf;
389 
390 		zf = ZEXISTS | ZSHARED;
391 		if (flags & FIGNORECASE)
392 			zf |= ZCILOOK;
393 
394 		error = zfs_dirent_lock(&dl, dzp, name, &zp, zf, deflg, rpnp);
395 		if (error == 0) {
396 			*vpp = ZTOV(zp);
397 			zfs_dirent_unlock(dl);
398 			dzp->z_zn_prefetch = B_TRUE; /* enable prefetching */
399 		}
400 		rpnp = NULL;
401 	}
402 
403 	if ((flags & FIGNORECASE) && rpnp)
404 		(void) strlcpy(rpnp->pn_path, name, rpnp->pn_bufsize);
405 
406 	return (error);
407 }
408 
409 static char *
410 zfs_unlinked_hexname(char namebuf[17], uint64_t x)
411 {
412 	char *name = &namebuf[16];
413 	const char digits[16] = "0123456789abcdef";
414 
415 	*name = '\0';
416 	do {
417 		*--name = digits[x & 0xf];
418 		x >>= 4;
419 	} while (x != 0);
420 
421 	return (name);
422 }
423 
424 /*
425  * unlinked Set (formerly known as the "delete queue") Error Handling
426  *
427  * When dealing with the unlinked set, we dmu_tx_hold_zap(), but we
428  * don't specify the name of the entry that we will be manipulating.  We
429  * also fib and say that we won't be adding any new entries to the
430  * unlinked set, even though we might (this is to lower the minimum file
431  * size that can be deleted in a full filesystem).  So on the small
432  * chance that the nlink list is using a fat zap (ie. has more than
433  * 2000 entries), we *may* not pre-read a block that's needed.
434  * Therefore it is remotely possible for some of the assertions
435  * regarding the unlinked set below to fail due to i/o error.  On a
436  * nondebug system, this will result in the space being leaked.
437  */
438 void
439 zfs_unlinked_add(znode_t *zp, dmu_tx_t *tx)
440 {
441 	zfsvfs_t *zfsvfs = zp->z_zfsvfs;
442 	char obj_name[17];
443 	int error;
444 
445 	ASSERT(zp->z_unlinked);
446 	ASSERT3U(zp->z_phys->zp_links, ==, 0);
447 
448 	error = zap_add(zfsvfs->z_os, zfsvfs->z_unlinkedobj,
449 	    zfs_unlinked_hexname(obj_name, zp->z_id), 8, 1, &zp->z_id, tx);
450 	ASSERT3U(error, ==, 0);
451 }
452 
453 /*
454  * Clean up any znodes that had no links when we either crashed or
455  * (force) umounted the file system.
456  */
457 void
458 zfs_unlinked_drain(zfsvfs_t *zfsvfs)
459 {
460 	zap_cursor_t	zc;
461 	zap_attribute_t zap;
462 	dmu_object_info_t doi;
463 	znode_t		*zp;
464 	int		error;
465 
466 	/*
467 	 * Interate over the contents of the unlinked set.
468 	 */
469 	for (zap_cursor_init(&zc, zfsvfs->z_os, zfsvfs->z_unlinkedobj);
470 	    zap_cursor_retrieve(&zc, &zap) == 0;
471 	    zap_cursor_advance(&zc)) {
472 
473 		/*
474 		 * See what kind of object we have in list
475 		 */
476 
477 		error = dmu_object_info(zfsvfs->z_os,
478 		    zap.za_first_integer, &doi);
479 		if (error != 0)
480 			continue;
481 
482 		ASSERT((doi.doi_type == DMU_OT_PLAIN_FILE_CONTENTS) ||
483 		    (doi.doi_type == DMU_OT_DIRECTORY_CONTENTS));
484 		/*
485 		 * We need to re-mark these list entries for deletion,
486 		 * so we pull them back into core and set zp->z_unlinked.
487 		 */
488 		error = zfs_zget(zfsvfs, zap.za_first_integer, &zp);
489 
490 		/*
491 		 * We may pick up znodes that are already marked for deletion.
492 		 * This could happen during the purge of an extended attribute
493 		 * directory.  All we need to do is skip over them, since they
494 		 * are already in the system marked z_unlinked.
495 		 */
496 		if (error != 0)
497 			continue;
498 
499 		zp->z_unlinked = B_TRUE;
500 		VN_RELE(ZTOV(zp));
501 	}
502 	zap_cursor_fini(&zc);
503 }
504 
505 /*
506  * Delete the entire contents of a directory.  Return a count
507  * of the number of entries that could not be deleted.
508  *
509  * NOTE: this function assumes that the directory is inactive,
510  *	so there is no need to lock its entries before deletion.
511  *	Also, it assumes the directory contents is *only* regular
512  *	files.
513  */
514 static int
515 zfs_purgedir(znode_t *dzp)
516 {
517 	zap_cursor_t	zc;
518 	zap_attribute_t	zap;
519 	znode_t		*xzp;
520 	dmu_tx_t	*tx;
521 	zfsvfs_t	*zfsvfs = dzp->z_zfsvfs;
522 	zfs_dirlock_t	dl;
523 	int skipped = 0;
524 	int error;
525 
526 	for (zap_cursor_init(&zc, zfsvfs->z_os, dzp->z_id);
527 	    (error = zap_cursor_retrieve(&zc, &zap)) == 0;
528 	    zap_cursor_advance(&zc)) {
529 		error = zfs_zget(zfsvfs,
530 		    ZFS_DIRENT_OBJ(zap.za_first_integer), &xzp);
531 		ASSERT3U(error, ==, 0);
532 
533 		ASSERT((ZTOV(xzp)->v_type == VREG) ||
534 		    (ZTOV(xzp)->v_type == VLNK));
535 
536 		tx = dmu_tx_create(zfsvfs->z_os);
537 		dmu_tx_hold_bonus(tx, dzp->z_id);
538 		dmu_tx_hold_zap(tx, dzp->z_id, FALSE, zap.za_name);
539 		dmu_tx_hold_bonus(tx, xzp->z_id);
540 		dmu_tx_hold_zap(tx, zfsvfs->z_unlinkedobj, FALSE, NULL);
541 		error = dmu_tx_assign(tx, TXG_WAIT);
542 		if (error) {
543 			dmu_tx_abort(tx);
544 			VN_RELE(ZTOV(xzp));
545 			skipped += 1;
546 			continue;
547 		}
548 		bzero(&dl, sizeof (dl));
549 		dl.dl_dzp = dzp;
550 		dl.dl_name = zap.za_name;
551 
552 		error = zfs_link_destroy(&dl, xzp, tx, 0, NULL);
553 		ASSERT3U(error, ==, 0);
554 		dmu_tx_commit(tx);
555 
556 		VN_RELE(ZTOV(xzp));
557 	}
558 	zap_cursor_fini(&zc);
559 	ASSERT(error == ENOENT);
560 	return (skipped);
561 }
562 
563 void
564 zfs_rmnode(znode_t *zp)
565 {
566 	zfsvfs_t	*zfsvfs = zp->z_zfsvfs;
567 	objset_t	*os = zfsvfs->z_os;
568 	znode_t		*xzp = NULL;
569 	char		obj_name[17];
570 	dmu_tx_t	*tx;
571 	uint64_t	acl_obj;
572 	int		error;
573 
574 	ASSERT(ZTOV(zp)->v_count == 0);
575 	ASSERT(zp->z_phys->zp_links == 0);
576 
577 	/*
578 	 * If this is an attribute directory, purge its contents.
579 	 */
580 	if (ZTOV(zp)->v_type == VDIR && (zp->z_phys->zp_flags & ZFS_XATTR)) {
581 		if (zfs_purgedir(zp) != 0) {
582 			/*
583 			 * Not enough space to delete some xattrs.
584 			 * Leave it on the unlinked set.
585 			 */
586 			return;
587 		}
588 	}
589 
590 	/*
591 	 * If the file has extended attributes, we're going to unlink
592 	 * the xattr dir.
593 	 */
594 	if (zp->z_phys->zp_xattr) {
595 		error = zfs_zget(zfsvfs, zp->z_phys->zp_xattr, &xzp);
596 		ASSERT(error == 0);
597 	}
598 
599 	acl_obj = zp->z_phys->zp_acl.z_acl_extern_obj;
600 
601 	/*
602 	 * Set up the transaction.
603 	 */
604 	tx = dmu_tx_create(os);
605 	dmu_tx_hold_free(tx, zp->z_id, 0, DMU_OBJECT_END);
606 	dmu_tx_hold_zap(tx, zfsvfs->z_unlinkedobj, FALSE, NULL);
607 	if (xzp) {
608 		dmu_tx_hold_bonus(tx, xzp->z_id);
609 		dmu_tx_hold_zap(tx, zfsvfs->z_unlinkedobj, TRUE, NULL);
610 	}
611 	if (acl_obj)
612 		dmu_tx_hold_free(tx, acl_obj, 0, DMU_OBJECT_END);
613 	error = dmu_tx_assign(tx, TXG_WAIT);
614 	if (error) {
615 		/*
616 		 * Not enough space to delete the file.  Leave it in the
617 		 * unlinked set, leaking it until the fs is remounted (at
618 		 * which point we'll call zfs_unlinked_drain() to process it).
619 		 */
620 		dmu_tx_abort(tx);
621 		return;
622 	}
623 
624 	if (xzp) {
625 		dmu_buf_will_dirty(xzp->z_dbuf, tx);
626 		mutex_enter(&xzp->z_lock);
627 		xzp->z_unlinked = B_TRUE;	/* mark xzp for deletion */
628 		xzp->z_phys->zp_links = 0;	/* no more links to it */
629 		mutex_exit(&xzp->z_lock);
630 		zfs_unlinked_add(xzp, tx);
631 	}
632 
633 	/* Remove this znode from the unlinked set */
634 	error = zap_remove(os, zfsvfs->z_unlinkedobj,
635 	    zfs_unlinked_hexname(obj_name, zp->z_id), tx);
636 	ASSERT3U(error, ==, 0);
637 
638 	zfs_znode_delete(zp, tx);
639 
640 	dmu_tx_commit(tx);
641 
642 	if (xzp)
643 		VN_RELE(ZTOV(xzp));
644 }
645 
646 static uint64_t
647 zfs_dirent(znode_t *zp)
648 {
649 	uint64_t de = zp->z_id;
650 	if (zp->z_zfsvfs->z_version >= ZPL_VERSION_DIRENT_TYPE)
651 		de |= IFTODT((zp)->z_phys->zp_mode) << 60;
652 	return (de);
653 }
654 
655 /*
656  * Link zp into dl.  Can only fail if zp has been unlinked.
657  */
658 int
659 zfs_link_create(zfs_dirlock_t *dl, znode_t *zp, dmu_tx_t *tx, int flag)
660 {
661 	znode_t *dzp = dl->dl_dzp;
662 	vnode_t *vp = ZTOV(zp);
663 	uint64_t value;
664 	int zp_is_dir = (vp->v_type == VDIR);
665 	int error;
666 
667 	dmu_buf_will_dirty(zp->z_dbuf, tx);
668 	mutex_enter(&zp->z_lock);
669 
670 	if (!(flag & ZRENAMING)) {
671 		if (zp->z_unlinked) {	/* no new links to unlinked zp */
672 			ASSERT(!(flag & (ZNEW | ZEXISTS)));
673 			mutex_exit(&zp->z_lock);
674 			return (ENOENT);
675 		}
676 		zp->z_phys->zp_links++;
677 	}
678 	zp->z_phys->zp_parent = dzp->z_id;	/* dzp is now zp's parent */
679 
680 	if (!(flag & ZNEW))
681 		zfs_time_stamper_locked(zp, STATE_CHANGED, tx);
682 	mutex_exit(&zp->z_lock);
683 
684 	dmu_buf_will_dirty(dzp->z_dbuf, tx);
685 	mutex_enter(&dzp->z_lock);
686 	dzp->z_phys->zp_size++;			/* one dirent added */
687 	dzp->z_phys->zp_links += zp_is_dir;	/* ".." link from zp */
688 	zfs_time_stamper_locked(dzp, CONTENT_MODIFIED, tx);
689 	mutex_exit(&dzp->z_lock);
690 
691 	value = zfs_dirent(zp);
692 	error = zap_add(zp->z_zfsvfs->z_os, dzp->z_id, dl->dl_name,
693 	    8, 1, &value, tx);
694 	ASSERT(error == 0);
695 
696 	dnlc_update(ZTOV(dzp), dl->dl_name, vp);
697 
698 	return (0);
699 }
700 
701 /*
702  * Unlink zp from dl, and mark zp for deletion if this was the last link.
703  * Can fail if zp is a mount point (EBUSY) or a non-empty directory (EEXIST).
704  * If 'unlinkedp' is NULL, we put unlinked znodes on the unlinked list.
705  * If it's non-NULL, we use it to indicate whether the znode needs deletion,
706  * and it's the caller's job to do it.
707  */
708 int
709 zfs_link_destroy(zfs_dirlock_t *dl, znode_t *zp, dmu_tx_t *tx, int flag,
710 	boolean_t *unlinkedp)
711 {
712 	znode_t *dzp = dl->dl_dzp;
713 	vnode_t *vp = ZTOV(zp);
714 	int zp_is_dir = (vp->v_type == VDIR);
715 	boolean_t unlinked = B_FALSE;
716 	int error;
717 
718 	dnlc_remove(ZTOV(dzp), dl->dl_name);
719 
720 	if (!(flag & ZRENAMING)) {
721 		dmu_buf_will_dirty(zp->z_dbuf, tx);
722 
723 		if (vn_vfswlock(vp))		/* prevent new mounts on zp */
724 			return (EBUSY);
725 
726 		if (vn_ismntpt(vp)) {		/* don't remove mount point */
727 			vn_vfsunlock(vp);
728 			return (EBUSY);
729 		}
730 
731 		mutex_enter(&zp->z_lock);
732 		if (zp_is_dir && !zfs_dirempty(zp)) {	/* dir not empty */
733 			mutex_exit(&zp->z_lock);
734 			vn_vfsunlock(vp);
735 			return (EEXIST);
736 		}
737 		if (zp->z_phys->zp_links <= zp_is_dir) {
738 			zfs_panic_recover("zfs: link count on %s is %u, "
739 			    "should be at least %u",
740 			    zp->z_vnode->v_path ? zp->z_vnode->v_path :
741 			    "<unknown>", (int)zp->z_phys->zp_links,
742 			    zp_is_dir + 1);
743 			zp->z_phys->zp_links = zp_is_dir + 1;
744 		}
745 		if (--zp->z_phys->zp_links == zp_is_dir) {
746 			zp->z_unlinked = B_TRUE;
747 			zp->z_phys->zp_links = 0;
748 			unlinked = B_TRUE;
749 		} else {
750 			zfs_time_stamper_locked(zp, STATE_CHANGED, tx);
751 		}
752 		mutex_exit(&zp->z_lock);
753 		vn_vfsunlock(vp);
754 	}
755 
756 	dmu_buf_will_dirty(dzp->z_dbuf, tx);
757 	mutex_enter(&dzp->z_lock);
758 	dzp->z_phys->zp_size--;			/* one dirent removed */
759 	dzp->z_phys->zp_links -= zp_is_dir;	/* ".." link from zp */
760 	zfs_time_stamper_locked(dzp, CONTENT_MODIFIED, tx);
761 	mutex_exit(&dzp->z_lock);
762 
763 	if (zp->z_zfsvfs->z_norm) {
764 		if (((zp->z_zfsvfs->z_case & ZFS_CI_ONLY) &&
765 		    (flag & ZCIEXACT)) ||
766 		    ((zp->z_zfsvfs->z_case & ZFS_CI_MIXD) &&
767 		    !(flag & ZCILOOK)))
768 			error = zap_remove_norm(zp->z_zfsvfs->z_os,
769 			    dzp->z_id, dl->dl_name, MT_EXACT, tx);
770 		else
771 			error = zap_remove_norm(zp->z_zfsvfs->z_os,
772 			    dzp->z_id, dl->dl_name, MT_FIRST, tx);
773 	} else {
774 		error = zap_remove(zp->z_zfsvfs->z_os,
775 		    dzp->z_id, dl->dl_name, tx);
776 	}
777 	ASSERT(error == 0);
778 
779 	if (unlinkedp != NULL)
780 		*unlinkedp = unlinked;
781 	else if (unlinked)
782 		zfs_unlinked_add(zp, tx);
783 
784 	return (0);
785 }
786 
787 /*
788  * Indicate whether the directory is empty.  Works with or without z_lock
789  * held, but can only be consider a hint in the latter case.  Returns true
790  * if only "." and ".." remain and there's no work in progress.
791  */
792 boolean_t
793 zfs_dirempty(znode_t *dzp)
794 {
795 	return (dzp->z_phys->zp_size == 2 && dzp->z_dirlocks == 0);
796 }
797 
798 int
799 zfs_make_xattrdir(znode_t *zp, vattr_t *vap, vnode_t **xvpp, cred_t *cr)
800 {
801 	zfsvfs_t *zfsvfs = zp->z_zfsvfs;
802 	znode_t *xzp;
803 	dmu_tx_t *tx;
804 	int error;
805 	zfs_fuid_info_t *fuidp = NULL;
806 
807 	*xvpp = NULL;
808 
809 	if (error = zfs_zaccess(zp, ACE_WRITE_NAMED_ATTRS, 0, B_FALSE, cr))
810 		return (error);
811 
812 	tx = dmu_tx_create(zfsvfs->z_os);
813 	dmu_tx_hold_bonus(tx, zp->z_id);
814 	dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, FALSE, NULL);
815 	if (zfsvfs->z_fuid_obj == 0) {
816 		dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
817 		dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, SPA_MAXBLOCKSIZE);
818 		dmu_tx_hold_zap(tx, MASTER_NODE_OBJ, FALSE, NULL);
819 	} else {
820 		dmu_tx_hold_bonus(tx, zfsvfs->z_fuid_obj);
821 		dmu_tx_hold_write(tx, zfsvfs->z_fuid_obj, 0, SPA_MAXBLOCKSIZE);
822 	}
823 	error = dmu_tx_assign(tx, zfsvfs->z_assign);
824 	if (error) {
825 		if (error == ERESTART && zfsvfs->z_assign == TXG_NOWAIT)
826 			dmu_tx_wait(tx);
827 		dmu_tx_abort(tx);
828 		return (error);
829 	}
830 	zfs_mknode(zp, vap, tx, cr, IS_XATTR, &xzp, 0, NULL, &fuidp);
831 	ASSERT(xzp->z_phys->zp_parent == zp->z_id);
832 	dmu_buf_will_dirty(zp->z_dbuf, tx);
833 	zp->z_phys->zp_xattr = xzp->z_id;
834 
835 	(void) zfs_log_create(zfsvfs->z_log, tx, TX_MKXATTR, zp,
836 	    xzp, "", NULL, fuidp, vap);
837 	if (fuidp)
838 		zfs_fuid_info_free(fuidp);
839 	dmu_tx_commit(tx);
840 
841 	*xvpp = ZTOV(xzp);
842 
843 	return (0);
844 }
845 
846 /*
847  * Return a znode for the extended attribute directory for zp.
848  * ** If the directory does not already exist, it is created **
849  *
850  *	IN:	zp	- znode to obtain attribute directory from
851  *		cr	- credentials of caller
852  *		flags	- flags from the VOP_LOOKUP call
853  *
854  *	OUT:	xzpp	- pointer to extended attribute znode
855  *
856  *	RETURN:	0 on success
857  *		error number on failure
858  */
859 int
860 zfs_get_xattrdir(znode_t *zp, vnode_t **xvpp, cred_t *cr, int flags)
861 {
862 	zfsvfs_t	*zfsvfs = zp->z_zfsvfs;
863 	znode_t		*xzp;
864 	zfs_dirlock_t	*dl;
865 	vattr_t		va;
866 	int		error;
867 top:
868 	error = zfs_dirent_lock(&dl, zp, "", &xzp, ZXATTR, NULL, NULL);
869 	if (error)
870 		return (error);
871 
872 	if (xzp != NULL) {
873 		*xvpp = ZTOV(xzp);
874 		zfs_dirent_unlock(dl);
875 		return (0);
876 	}
877 
878 	ASSERT(zp->z_phys->zp_xattr == 0);
879 
880 	if (!(flags & CREATE_XATTR_DIR)) {
881 		zfs_dirent_unlock(dl);
882 		return (ENOENT);
883 	}
884 
885 	if (zfsvfs->z_vfs->vfs_flag & VFS_RDONLY) {
886 		zfs_dirent_unlock(dl);
887 		return (EROFS);
888 	}
889 
890 	/*
891 	 * The ability to 'create' files in an attribute
892 	 * directory comes from the write_xattr permission on the base file.
893 	 *
894 	 * The ability to 'search' an attribute directory requires
895 	 * read_xattr permission on the base file.
896 	 *
897 	 * Once in a directory the ability to read/write attributes
898 	 * is controlled by the permissions on the attribute file.
899 	 */
900 	va.va_mask = AT_TYPE | AT_MODE | AT_UID | AT_GID;
901 	va.va_type = VDIR;
902 	va.va_mode = S_IFDIR | S_ISVTX | 0777;
903 	zfs_fuid_map_ids(zp, &va.va_uid, &va.va_gid);
904 
905 	error = zfs_make_xattrdir(zp, &va, xvpp, cr);
906 	zfs_dirent_unlock(dl);
907 
908 	if (error == ERESTART && zfsvfs->z_assign == TXG_NOWAIT) {
909 		/* NB: we already did dmu_tx_wait() if necessary */
910 		goto top;
911 	}
912 
913 	return (error);
914 }
915 
916 /*
917  * Decide whether it is okay to remove within a sticky directory.
918  *
919  * In sticky directories, write access is not sufficient;
920  * you can remove entries from a directory only if:
921  *
922  *	you own the directory,
923  *	you own the entry,
924  *	the entry is a plain file and you have write access,
925  *	or you are privileged (checked in secpolicy...).
926  *
927  * The function returns 0 if remove access is granted.
928  */
929 int
930 zfs_sticky_remove_access(znode_t *zdp, znode_t *zp, cred_t *cr)
931 {
932 	uid_t  		uid;
933 	uid_t		downer;
934 	uid_t		fowner;
935 	zfsvfs_t	*zfsvfs = zdp->z_zfsvfs;
936 
937 	if (zdp->z_zfsvfs->z_assign >= TXG_INITIAL)	/* ZIL replay */
938 		return (0);
939 
940 	if ((zdp->z_phys->zp_mode & S_ISVTX) == 0)
941 		return (0);
942 
943 	zfs_fuid_map_id(zfsvfs, zdp->z_phys->zp_uid, ZFS_OWNER, &downer);
944 	zfs_fuid_map_id(zfsvfs, zp->z_phys->zp_uid, ZFS_OWNER, &fowner);
945 
946 	if ((uid = crgetuid(cr)) == downer || uid == fowner ||
947 	    (ZTOV(zp)->v_type == VREG &&
948 	    zfs_zaccess(zp, ACE_WRITE_DATA, 0, B_FALSE, cr) == 0))
949 		return (0);
950 	else
951 		return (secpolicy_vnode_remove(cr));
952 }
953