xref: /titanic_52/usr/src/uts/common/fs/zfs/zfs_dir.c (revision 74e20cfe817b82802b16fac8690dadcda76f54f5)
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 2006 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/dnlc.h>
56 
57 /*
58  * Lock a directory entry.  A dirlock on <dzp, name> protects that name
59  * in dzp's directory zap object.  As long as you hold a dirlock, you can
60  * assume two things: (1) dzp cannot be reaped, and (2) no other thread
61  * can change the zap entry for (i.e. link or unlink) this name.
62  *
63  * Input arguments:
64  *	dzp	- znode for directory
65  *	name	- name of entry to lock
66  *	flag	- ZNEW: if the entry already exists, fail with EEXIST.
67  *		  ZEXISTS: if the entry does not exist, fail with ENOENT.
68  *		  ZSHARED: allow concurrent access with other ZSHARED callers.
69  *		  ZXATTR: we want dzp's xattr directory
70  *
71  * Output arguments:
72  *	zpp	- pointer to the znode for the entry (NULL if there isn't one)
73  *	dlpp	- pointer to the dirlock for this entry (NULL on error)
74  *
75  * Return value: 0 on success or errno on failure.
76  *
77  * NOTE: Always checks for, and rejects, '.' and '..'.
78  */
79 int
80 zfs_dirent_lock(zfs_dirlock_t **dlpp, znode_t *dzp, char *name, znode_t **zpp,
81 	int flag)
82 {
83 	zfsvfs_t	*zfsvfs = dzp->z_zfsvfs;
84 	zfs_dirlock_t	*dl;
85 	uint64_t	zoid;
86 	int		error;
87 	vnode_t		*vp;
88 
89 	*zpp = NULL;
90 	*dlpp = NULL;
91 
92 	/*
93 	 * Verify that we are not trying to lock '.', '..', or '.zfs'
94 	 */
95 	if (name[0] == '.' &&
96 	    (name[1] == '\0' || (name[1] == '.' && name[2] == '\0')) ||
97 	    zfs_has_ctldir(dzp) && strcmp(name, ZFS_CTLDIR_NAME) == 0)
98 		return (EEXIST);
99 
100 	/*
101 	 * Wait until there are no locks on this name.
102 	 */
103 	mutex_enter(&dzp->z_lock);
104 	for (;;) {
105 		if (dzp->z_reap) {
106 			mutex_exit(&dzp->z_lock);
107 			return (ENOENT);
108 		}
109 		for (dl = dzp->z_dirlocks; dl != NULL; dl = dl->dl_next)
110 			if (strcmp(name, dl->dl_name) == 0)
111 				break;
112 		if (dl == NULL)	{
113 			/*
114 			 * Allocate a new dirlock and add it to the list.
115 			 */
116 			dl = kmem_alloc(sizeof (zfs_dirlock_t), KM_SLEEP);
117 			cv_init(&dl->dl_cv, NULL, CV_DEFAULT, NULL);
118 			dl->dl_name = name;
119 			dl->dl_sharecnt = 0;
120 			dl->dl_namesize = 0;
121 			dl->dl_dzp = dzp;
122 			dl->dl_next = dzp->z_dirlocks;
123 			dzp->z_dirlocks = dl;
124 			break;
125 		}
126 		if ((flag & ZSHARED) && dl->dl_sharecnt != 0)
127 			break;
128 		cv_wait(&dl->dl_cv, &dzp->z_lock);
129 	}
130 
131 	if ((flag & ZSHARED) && ++dl->dl_sharecnt > 1 && dl->dl_namesize == 0) {
132 		/*
133 		 * We're the second shared reference to dl.  Make a copy of
134 		 * dl_name in case the first thread goes away before we do.
135 		 * Note that we initialize the new name before storing its
136 		 * pointer into dl_name, because the first thread may load
137 		 * dl->dl_name at any time.  He'll either see the old value,
138 		 * which is his, or the new shared copy; either is OK.
139 		 */
140 		dl->dl_namesize = strlen(dl->dl_name) + 1;
141 		name = kmem_alloc(dl->dl_namesize, KM_SLEEP);
142 		bcopy(dl->dl_name, name, dl->dl_namesize);
143 		dl->dl_name = name;
144 	}
145 
146 	mutex_exit(&dzp->z_lock);
147 
148 	/*
149 	 * We have a dirlock on the name.  (Note that it is the dirlock,
150 	 * not the dzp's z_lock, that protects the name in the zap object.)
151 	 * See if there's an object by this name; if so, put a hold on it.
152 	 */
153 	if (flag & ZXATTR) {
154 		zoid = dzp->z_phys->zp_xattr;
155 		error = (zoid == 0 ? ENOENT : 0);
156 	} else {
157 		vp = dnlc_lookup(ZTOV(dzp), name);
158 		if (vp == DNLC_NO_VNODE) {
159 			VN_RELE(vp);
160 			error = ENOENT;
161 		} else if (vp) {
162 			if (flag & ZNEW) {
163 				zfs_dirent_unlock(dl);
164 				VN_RELE(vp);
165 				return (EEXIST);
166 			}
167 			*dlpp = dl;
168 			*zpp = VTOZ(vp);
169 			return (0);
170 		} else {
171 			error = zap_lookup(zfsvfs->z_os, dzp->z_id, name,
172 			    8, 1, &zoid);
173 			if (error == ENOENT)
174 				dnlc_update(ZTOV(dzp), name, DNLC_NO_VNODE);
175 		}
176 	}
177 	if (error) {
178 		if (error != ENOENT || (flag & ZEXISTS)) {
179 			zfs_dirent_unlock(dl);
180 			return (error);
181 		}
182 	} else {
183 		if (flag & ZNEW) {
184 			zfs_dirent_unlock(dl);
185 			return (EEXIST);
186 		}
187 		error = zfs_zget(zfsvfs, zoid, zpp);
188 		if (error) {
189 			zfs_dirent_unlock(dl);
190 			return (error);
191 		}
192 		if (!(flag & ZXATTR))
193 			dnlc_update(ZTOV(dzp), name, ZTOV(*zpp));
194 	}
195 
196 	*dlpp = dl;
197 
198 	return (0);
199 }
200 
201 /*
202  * Unlock this directory entry and wake anyone who was waiting for it.
203  */
204 void
205 zfs_dirent_unlock(zfs_dirlock_t *dl)
206 {
207 	znode_t *dzp = dl->dl_dzp;
208 	zfs_dirlock_t **prev_dl, *cur_dl;
209 
210 	mutex_enter(&dzp->z_lock);
211 	if (dl->dl_sharecnt > 1) {
212 		dl->dl_sharecnt--;
213 		mutex_exit(&dzp->z_lock);
214 		return;
215 	}
216 	prev_dl = &dzp->z_dirlocks;
217 	while ((cur_dl = *prev_dl) != dl)
218 		prev_dl = &cur_dl->dl_next;
219 	*prev_dl = dl->dl_next;
220 	cv_broadcast(&dl->dl_cv);
221 	mutex_exit(&dzp->z_lock);
222 
223 	if (dl->dl_namesize != 0)
224 		kmem_free(dl->dl_name, dl->dl_namesize);
225 	cv_destroy(&dl->dl_cv);
226 	kmem_free(dl, sizeof (*dl));
227 }
228 
229 /*
230  * Look up an entry in a directory.
231  *
232  * NOTE: '.' and '..' are handled as special cases because
233  *	no directory entries are actually stored for them.  If this is
234  *	the root of a filesystem, then '.zfs' is also treated as a
235  *	special pseudo-directory.
236  */
237 int
238 zfs_dirlook(znode_t *dzp, char *name, vnode_t **vpp)
239 {
240 	zfs_dirlock_t *dl;
241 	znode_t *zp;
242 	int error = 0;
243 
244 	if (name[0] == 0 || (name[0] == '.' && name[1] == 0)) {
245 		*vpp = ZTOV(dzp);
246 		VN_HOLD(*vpp);
247 	} else if (name[0] == '.' && name[1] == '.' && name[2] == 0) {
248 		zfsvfs_t *zfsvfs = dzp->z_zfsvfs;
249 		/*
250 		 * If we are a snapshot mounted under .zfs, return
251 		 * the vp for the snapshot directory.
252 		 */
253 		if (dzp->z_phys->zp_parent == dzp->z_id &&
254 		    zfsvfs->z_parent != zfsvfs) {
255 			error = zfsctl_root_lookup(zfsvfs->z_parent->z_ctldir,
256 			    "snapshot", vpp, NULL, 0, NULL, kcred);
257 			return (error);
258 		}
259 		rw_enter(&dzp->z_parent_lock, RW_READER);
260 		error = zfs_zget(zfsvfs, dzp->z_phys->zp_parent, &zp);
261 		if (error == 0)
262 			*vpp = ZTOV(zp);
263 		rw_exit(&dzp->z_parent_lock);
264 	} else if (zfs_has_ctldir(dzp) && strcmp(name, ZFS_CTLDIR_NAME) == 0) {
265 		*vpp = zfsctl_root(dzp);
266 	} else {
267 		error = zfs_dirent_lock(&dl, dzp, name, &zp, ZEXISTS | ZSHARED);
268 		if (error == 0) {
269 			*vpp = ZTOV(zp);
270 			zfs_dirent_unlock(dl);
271 			dzp->z_zn_prefetch = B_TRUE; /* enable prefetching */
272 		}
273 	}
274 
275 	return (error);
276 }
277 
278 static char *
279 zfs_dq_hexname(char namebuf[17], uint64_t x)
280 {
281 	char *name = &namebuf[16];
282 	const char digits[16] = "0123456789abcdef";
283 
284 	*name = '\0';
285 	do {
286 		*--name = digits[x & 0xf];
287 		x >>= 4;
288 	} while (x != 0);
289 
290 	return (name);
291 }
292 
293 /*
294  * Delete Queue Error Handling
295  *
296  * When dealing with the delete queue, we dmu_tx_hold_zap(), but we
297  * don't specify the name of the entry that we will be manipulating.  We
298  * also fib and say that we won't be adding any new entries to the
299  * delete queue, even though we might (this is to lower the minimum file
300  * size that can be deleted in a full filesystem).  So on the small
301  * chance that the delete queue is using a fat zap (ie. has more than
302  * 2000 entries), we *may* not pre-read a block that's needed.
303  * Therefore it is remotely possible for some of the assertions
304  * regarding the delete queue below to fail due to i/o error.  On a
305  * nondebug system, this will result in the space being leaked.
306  */
307 
308 void
309 zfs_dq_add(znode_t *zp, dmu_tx_t *tx)
310 {
311 	zfsvfs_t *zfsvfs = zp->z_zfsvfs;
312 	char obj_name[17];
313 	int error;
314 
315 	ASSERT(zp->z_reap);
316 	ASSERT3U(zp->z_phys->zp_links, ==, 0);
317 
318 	error = zap_add(zfsvfs->z_os, zfsvfs->z_dqueue,
319 	    zfs_dq_hexname(obj_name, zp->z_id), 8, 1, &zp->z_id, tx);
320 	ASSERT3U(error, ==, 0);
321 }
322 
323 /*
324  * Delete the entire contents of a directory.  Return a count
325  * of the number of entries that could not be deleted.
326  *
327  * NOTE: this function assumes that the directory is inactive,
328  *	so there is no need to lock its entries before deletion.
329  *	Also, it assumes the directory contents is *only* regular
330  *	files.
331  */
332 static int
333 zfs_purgedir(znode_t *dzp)
334 {
335 	zap_cursor_t	zc;
336 	zap_attribute_t	zap;
337 	znode_t		*xzp;
338 	dmu_tx_t	*tx;
339 	zfsvfs_t	*zfsvfs = dzp->z_zfsvfs;
340 	zfs_dirlock_t	dl;
341 	int skipped = 0;
342 	int error;
343 
344 
345 	for (zap_cursor_init(&zc, zfsvfs->z_os, dzp->z_id);
346 	    (error = zap_cursor_retrieve(&zc, &zap)) == 0;
347 	    zap_cursor_advance(&zc)) {
348 		error = zfs_zget(zfsvfs, zap.za_first_integer, &xzp);
349 		ASSERT3U(error, ==, 0);
350 
351 		ASSERT((ZTOV(xzp)->v_type == VREG) ||
352 		    (ZTOV(xzp)->v_type == VLNK));
353 
354 		tx = dmu_tx_create(zfsvfs->z_os);
355 		dmu_tx_hold_bonus(tx, dzp->z_id);
356 		dmu_tx_hold_zap(tx, dzp->z_id, FALSE, zap.za_name);
357 		dmu_tx_hold_bonus(tx, xzp->z_id);
358 		dmu_tx_hold_zap(tx, zfsvfs->z_dqueue, FALSE, NULL);
359 		error = dmu_tx_assign(tx, TXG_WAIT);
360 		if (error) {
361 			dmu_tx_abort(tx);
362 			VN_RELE(ZTOV(xzp));
363 			skipped += 1;
364 			continue;
365 		}
366 		bzero(&dl, sizeof (dl));
367 		dl.dl_dzp = dzp;
368 		dl.dl_name = zap.za_name;
369 
370 		error = zfs_link_destroy(&dl, xzp, tx, 0, NULL);
371 		ASSERT3U(error, ==, 0);
372 		dmu_tx_commit(tx);
373 
374 		VN_RELE(ZTOV(xzp));
375 	}
376 	zap_cursor_fini(&zc);
377 	ASSERT(error == ENOENT);
378 	return (skipped);
379 }
380 
381 /*
382  * Special function to requeue the znodes for deletion that were
383  * in progress when we either crashed or umounted the file system.
384  *
385  * returns 1 if queue was drained.
386  */
387 static int
388 zfs_drain_dq(zfsvfs_t *zfsvfs)
389 {
390 	zap_cursor_t	zc;
391 	zap_attribute_t zap;
392 	dmu_object_info_t doi;
393 	znode_t		*zp;
394 	int		error;
395 
396 	/*
397 	 * Interate over the contents of the delete queue.
398 	 */
399 	for (zap_cursor_init(&zc, zfsvfs->z_os, zfsvfs->z_dqueue);
400 	    zap_cursor_retrieve(&zc, &zap) == 0;
401 	    zap_cursor_advance(&zc)) {
402 
403 		/*
404 		 * Create more threads if necessary to balance the load.
405 		 * quit if the delete threads have been shut down.
406 		 */
407 		if (zfs_delete_thread_target(zfsvfs, -1) != 0)
408 			return (0);
409 
410 		/*
411 		 * See what kind of object we have in queue
412 		 */
413 
414 		error = dmu_object_info(zfsvfs->z_os,
415 		    zap.za_first_integer, &doi);
416 		if (error != 0)
417 			continue;
418 
419 		ASSERT((doi.doi_type == DMU_OT_PLAIN_FILE_CONTENTS) ||
420 		    (doi.doi_type == DMU_OT_DIRECTORY_CONTENTS));
421 		/*
422 		 * We need to re-mark these queue entries for reaping,
423 		 * so we pull them back into core and set zp->z_reap.
424 		 */
425 		error = zfs_zget(zfsvfs, zap.za_first_integer, &zp);
426 
427 		/*
428 		 * We may pick up znodes that are already marked for reaping.
429 		 * This could happen during the purge of an extended attribute
430 		 * directory.  All we need to do is skip over them, since they
431 		 * are already in the system to be processed by the delete
432 		 * thread(s).
433 		 */
434 		if (error != 0) {
435 			continue;
436 		}
437 
438 		zp->z_reap = 1;
439 		VN_RELE(ZTOV(zp));
440 	}
441 	zap_cursor_fini(&zc);
442 	return (1);
443 }
444 
445 void
446 zfs_delete_thread(void *arg)
447 {
448 	zfsvfs_t	*zfsvfs = arg;
449 	zfs_delete_t 	*zd = &zfsvfs->z_delete_head;
450 	znode_t		*zp;
451 	callb_cpr_t	cprinfo;
452 	int		drained;
453 
454 	CALLB_CPR_INIT(&cprinfo, &zd->z_mutex, callb_generic_cpr, "zfs_delete");
455 
456 	mutex_enter(&zd->z_mutex);
457 
458 	if (!zd->z_drained && !zd->z_draining) {
459 		zd->z_draining = B_TRUE;
460 		mutex_exit(&zd->z_mutex);
461 		drained = zfs_drain_dq(zfsvfs);
462 		mutex_enter(&zd->z_mutex);
463 		zd->z_draining = B_FALSE;
464 		zd->z_drained = drained;
465 		cv_broadcast(&zd->z_quiesce_cv);
466 	}
467 
468 	while (zd->z_thread_count <= zd->z_thread_target) {
469 		zp = list_head(&zd->z_znodes);
470 		if (zp == NULL) {
471 			ASSERT(zd->z_znode_count == 0);
472 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
473 			cv_wait(&zd->z_cv, &zd->z_mutex);
474 			CALLB_CPR_SAFE_END(&cprinfo, &zd->z_mutex);
475 			continue;
476 		}
477 		ASSERT(zd->z_znode_count != 0);
478 		list_remove(&zd->z_znodes, zp);
479 		if (--zd->z_znode_count == 0)
480 			cv_broadcast(&zd->z_quiesce_cv);
481 		mutex_exit(&zd->z_mutex);
482 		zfs_rmnode(zp);
483 		(void) zfs_delete_thread_target(zfsvfs, -1);
484 		mutex_enter(&zd->z_mutex);
485 	}
486 
487 	ASSERT(zd->z_thread_count != 0);
488 	if (--zd->z_thread_count == 0)
489 		cv_broadcast(&zd->z_cv);
490 
491 	CALLB_CPR_EXIT(&cprinfo);	/* NB: drops z_mutex */
492 	thread_exit();
493 }
494 
495 static int zfs_work_per_thread_shift = 11;	/* 2048 (2^11) per thread */
496 
497 /*
498  * Set the target number of delete threads to 'nthreads'.
499  * If nthreads == -1, choose a number based on current workload.
500  * If nthreads == 0, don't return until the threads have exited.
501  */
502 int
503 zfs_delete_thread_target(zfsvfs_t *zfsvfs, int nthreads)
504 {
505 	zfs_delete_t *zd = &zfsvfs->z_delete_head;
506 
507 	mutex_enter(&zd->z_mutex);
508 
509 	if (nthreads == -1) {
510 		if (zd->z_thread_target == 0) {
511 			mutex_exit(&zd->z_mutex);
512 			return (EBUSY);
513 		}
514 		nthreads = zd->z_znode_count >> zfs_work_per_thread_shift;
515 		nthreads = MIN(nthreads, ncpus << 1);
516 		nthreads = MAX(nthreads, 1);
517 		nthreads += !!zd->z_draining;
518 	}
519 
520 	zd->z_thread_target = nthreads;
521 
522 	while (zd->z_thread_count < zd->z_thread_target) {
523 		(void) thread_create(NULL, 0, zfs_delete_thread, zfsvfs,
524 		    0, &p0, TS_RUN, minclsyspri);
525 		zd->z_thread_count++;
526 	}
527 
528 	while (zd->z_thread_count > zd->z_thread_target && nthreads == 0) {
529 		cv_broadcast(&zd->z_cv);
530 		cv_wait(&zd->z_cv, &zd->z_mutex);
531 	}
532 
533 	mutex_exit(&zd->z_mutex);
534 
535 	return (0);
536 }
537 
538 /*
539  * Wait until everything that's been queued has been deleted.
540  */
541 void
542 zfs_delete_wait_empty(zfsvfs_t *zfsvfs)
543 {
544 	zfs_delete_t *zd = &zfsvfs->z_delete_head;
545 
546 	mutex_enter(&zd->z_mutex);
547 	ASSERT(zd->z_thread_target != 0);
548 	while (!zd->z_drained || zd->z_znode_count != 0) {
549 		ASSERT(zd->z_thread_target != 0);
550 		cv_wait(&zd->z_quiesce_cv, &zd->z_mutex);
551 	}
552 	mutex_exit(&zd->z_mutex);
553 }
554 
555 void
556 zfs_rmnode(znode_t *zp)
557 {
558 	zfsvfs_t	*zfsvfs = zp->z_zfsvfs;
559 	objset_t	*os = zfsvfs->z_os;
560 	znode_t		*xzp = NULL;
561 	char		obj_name[17];
562 	dmu_tx_t	*tx;
563 	uint64_t	acl_obj;
564 	int		error;
565 
566 	ASSERT(ZTOV(zp)->v_count == 0);
567 	ASSERT(zp->z_phys->zp_links == 0);
568 
569 	/*
570 	 * If this is an attribute directory, purge its contents.
571 	 */
572 	if (ZTOV(zp)->v_type == VDIR && (zp->z_phys->zp_flags & ZFS_XATTR))
573 		if (zfs_purgedir(zp) != 0) {
574 			zfs_delete_t *delq = &zfsvfs->z_delete_head;
575 			/*
576 			 * Add this back to the delete list to be retried later.
577 			 *
578 			 * XXX - this could just busy loop on us...
579 			 */
580 			mutex_enter(&delq->z_mutex);
581 			list_insert_tail(&delq->z_znodes, zp);
582 			delq->z_znode_count++;
583 			mutex_exit(&delq->z_mutex);
584 			return;
585 		}
586 
587 	/*
588 	 * If the file has extended attributes, unlink the xattr dir.
589 	 */
590 	if (zp->z_phys->zp_xattr) {
591 		error = zfs_zget(zfsvfs, zp->z_phys->zp_xattr, &xzp);
592 		ASSERT(error == 0);
593 	}
594 
595 	acl_obj = zp->z_phys->zp_acl.z_acl_extern_obj;
596 
597 	/*
598 	 * Set up the transaction.
599 	 */
600 	tx = dmu_tx_create(os);
601 	dmu_tx_hold_free(tx, zp->z_id, 0, DMU_OBJECT_END);
602 	dmu_tx_hold_zap(tx, zfsvfs->z_dqueue, FALSE, NULL);
603 	if (xzp) {
604 		dmu_tx_hold_bonus(tx, xzp->z_id);
605 		dmu_tx_hold_zap(tx, zfsvfs->z_dqueue, TRUE, NULL);
606 	}
607 	if (acl_obj)
608 		dmu_tx_hold_free(tx, acl_obj, 0, DMU_OBJECT_END);
609 	error = dmu_tx_assign(tx, TXG_WAIT);
610 	if (error) {
611 		zfs_delete_t *delq = &zfsvfs->z_delete_head;
612 
613 		dmu_tx_abort(tx);
614 		/*
615 		 * Add this back to the delete list to be retried later.
616 		 *
617 		 * XXX - this could just busy loop on us...
618 		 */
619 		mutex_enter(&delq->z_mutex);
620 		list_insert_tail(&delq->z_znodes, zp);
621 		delq->z_znode_count++;
622 		mutex_exit(&delq->z_mutex);
623 		return;
624 	}
625 
626 	if (xzp) {
627 		dmu_buf_will_dirty(xzp->z_dbuf, tx);
628 		mutex_enter(&xzp->z_lock);
629 		xzp->z_reap = 1;		/* mark xzp for deletion */
630 		xzp->z_phys->zp_links = 0;	/* no more links to it */
631 		mutex_exit(&xzp->z_lock);
632 		zfs_dq_add(xzp, tx);		/* add xzp to delete queue */
633 	}
634 
635 	/*
636 	 * Remove this znode from delete queue
637 	 */
638 	error = zap_remove(os, zfsvfs->z_dqueue,
639 	    zfs_dq_hexname(obj_name, zp->z_id), tx);
640 	ASSERT3U(error, ==, 0);
641 
642 	zfs_znode_delete(zp, tx);
643 
644 	dmu_tx_commit(tx);
645 
646 	if (xzp)
647 		VN_RELE(ZTOV(xzp));
648 }
649 
650 /*
651  * Link zp into dl.  Can only fail if zp has been reaped.
652  */
653 int
654 zfs_link_create(zfs_dirlock_t *dl, znode_t *zp, dmu_tx_t *tx, int flag)
655 {
656 	znode_t *dzp = dl->dl_dzp;
657 	vnode_t *vp = ZTOV(zp);
658 	int zp_is_dir = (vp->v_type == VDIR);
659 	int error;
660 
661 	dmu_buf_will_dirty(zp->z_dbuf, tx);
662 	mutex_enter(&zp->z_lock);
663 
664 	if (!(flag & ZRENAMING)) {
665 		if (zp->z_reap) {	/* no new links to reaped zp */
666 			ASSERT(!(flag & (ZNEW | ZEXISTS)));
667 			mutex_exit(&zp->z_lock);
668 			return (ENOENT);
669 		}
670 		zp->z_phys->zp_links++;
671 	}
672 	zp->z_phys->zp_parent = dzp->z_id;	/* dzp is now zp's parent */
673 
674 	if (!(flag & ZNEW))
675 		zfs_time_stamper_locked(zp, STATE_CHANGED, tx);
676 	mutex_exit(&zp->z_lock);
677 
678 	dmu_buf_will_dirty(dzp->z_dbuf, tx);
679 	mutex_enter(&dzp->z_lock);
680 	dzp->z_phys->zp_size++;			/* one dirent added */
681 	dzp->z_phys->zp_links += zp_is_dir;	/* ".." link from zp */
682 	zfs_time_stamper_locked(dzp, CONTENT_MODIFIED, tx);
683 	mutex_exit(&dzp->z_lock);
684 
685 	error = zap_add(zp->z_zfsvfs->z_os, dzp->z_id, dl->dl_name,
686 	    8, 1, &zp->z_id, tx);
687 	ASSERT(error == 0);
688 
689 	dnlc_update(ZTOV(dzp), dl->dl_name, vp);
690 
691 	return (0);
692 }
693 
694 /*
695  * Unlink zp from dl, and mark zp for reaping if this was the last link.
696  * Can fail if zp is a mount point (EBUSY) or a non-empty directory (EEXIST).
697  * If 'reaped_ptr' is NULL, we put reaped znodes on the delete queue.
698  * If it's non-NULL, we use it to indicate whether the znode needs reaping,
699  * and it's the caller's job to do it.
700  */
701 int
702 zfs_link_destroy(zfs_dirlock_t *dl, znode_t *zp, dmu_tx_t *tx, int flag,
703 	int *reaped_ptr)
704 {
705 	znode_t *dzp = dl->dl_dzp;
706 	vnode_t *vp = ZTOV(zp);
707 	int zp_is_dir = (vp->v_type == VDIR);
708 	int reaped = 0;
709 	int error;
710 
711 	dnlc_remove(ZTOV(dzp), dl->dl_name);
712 
713 	if (!(flag & ZRENAMING)) {
714 		dmu_buf_will_dirty(zp->z_dbuf, tx);
715 
716 		if (vn_vfswlock(vp))		/* prevent new mounts on zp */
717 			return (EBUSY);
718 
719 		if (vn_ismntpt(vp)) {		/* don't remove mount point */
720 			vn_vfsunlock(vp);
721 			return (EBUSY);
722 		}
723 
724 		mutex_enter(&zp->z_lock);
725 		if (zp_is_dir && !zfs_dirempty(zp)) {	/* dir not empty */
726 			mutex_exit(&zp->z_lock);
727 			vn_vfsunlock(vp);
728 			return (EEXIST);
729 		}
730 		ASSERT(zp->z_phys->zp_links > zp_is_dir);
731 		if (--zp->z_phys->zp_links == zp_is_dir) {
732 			zp->z_reap = 1;
733 			zp->z_phys->zp_links = 0;
734 			reaped = 1;
735 		} else {
736 			zfs_time_stamper_locked(zp, STATE_CHANGED, tx);
737 		}
738 		mutex_exit(&zp->z_lock);
739 		vn_vfsunlock(vp);
740 	}
741 
742 	dmu_buf_will_dirty(dzp->z_dbuf, tx);
743 	mutex_enter(&dzp->z_lock);
744 	dzp->z_phys->zp_size--;			/* one dirent removed */
745 	dzp->z_phys->zp_links -= zp_is_dir;	/* ".." link from zp */
746 	zfs_time_stamper_locked(dzp, CONTENT_MODIFIED, tx);
747 	mutex_exit(&dzp->z_lock);
748 
749 	error = zap_remove(zp->z_zfsvfs->z_os, dzp->z_id, dl->dl_name, tx);
750 	ASSERT(error == 0);
751 
752 	if (reaped_ptr != NULL)
753 		*reaped_ptr = reaped;
754 	else if (reaped)
755 		zfs_dq_add(zp, tx);
756 
757 	return (0);
758 }
759 
760 /*
761  * Indicate whether the directory is empty.  Works with or without z_lock
762  * held, but can only be consider a hint in the latter case.  Returns true
763  * if only "." and ".." remain and there's no work in progress.
764  */
765 boolean_t
766 zfs_dirempty(znode_t *dzp)
767 {
768 	return (dzp->z_phys->zp_size == 2 && dzp->z_dirlocks == 0);
769 }
770 
771 int
772 zfs_make_xattrdir(znode_t *zp, vattr_t *vap, vnode_t **xvpp, cred_t *cr)
773 {
774 	zfsvfs_t *zfsvfs = zp->z_zfsvfs;
775 	znode_t *xzp;
776 	dmu_tx_t *tx;
777 	uint64_t xoid;
778 	int error;
779 
780 	*xvpp = NULL;
781 
782 	if (error = zfs_zaccess(zp, ACE_WRITE_NAMED_ATTRS, cr))
783 		return (error);
784 
785 	tx = dmu_tx_create(zfsvfs->z_os);
786 	dmu_tx_hold_bonus(tx, zp->z_id);
787 	dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, FALSE, NULL);
788 	error = dmu_tx_assign(tx, zfsvfs->z_assign);
789 	if (error) {
790 		if (error == ERESTART && zfsvfs->z_assign == TXG_NOWAIT)
791 			dmu_tx_wait(tx);
792 		dmu_tx_abort(tx);
793 		return (error);
794 	}
795 	zfs_mknode(zp, vap, &xoid, tx, cr, IS_XATTR, &xzp, 0);
796 	ASSERT(xzp->z_id == xoid);
797 	ASSERT(xzp->z_phys->zp_parent == zp->z_id);
798 	dmu_buf_will_dirty(zp->z_dbuf, tx);
799 	zp->z_phys->zp_xattr = xoid;
800 
801 	(void) zfs_log_create(zfsvfs->z_log, tx, TX_MKXATTR, zp, xzp, "");
802 	dmu_tx_commit(tx);
803 
804 	*xvpp = ZTOV(xzp);
805 
806 	return (0);
807 }
808 
809 /*
810  * Return a znode for the extended attribute directory for zp.
811  * ** If the directory does not already exist, it is created **
812  *
813  *	IN:	zp	- znode to obtain attribute directory from
814  *		cr	- credentials of caller
815  *		flags	- flags from the VOP_LOOKUP call
816  *
817  *	OUT:	xzpp	- pointer to extended attribute znode
818  *
819  *	RETURN:	0 on success
820  *		error number on failure
821  */
822 int
823 zfs_get_xattrdir(znode_t *zp, vnode_t **xvpp, cred_t *cr, int flags)
824 {
825 	zfsvfs_t	*zfsvfs = zp->z_zfsvfs;
826 	znode_t		*xzp;
827 	zfs_dirlock_t	*dl;
828 	vattr_t		va;
829 	int		error;
830 top:
831 	error = zfs_dirent_lock(&dl, zp, "", &xzp, ZXATTR);
832 	if (error)
833 		return (error);
834 
835 	if (xzp != NULL) {
836 		*xvpp = ZTOV(xzp);
837 		zfs_dirent_unlock(dl);
838 		return (0);
839 	}
840 
841 	ASSERT(zp->z_phys->zp_xattr == 0);
842 
843 	if (!(flags & CREATE_XATTR_DIR)) {
844 		zfs_dirent_unlock(dl);
845 		return (ENOENT);
846 	}
847 
848 	if (zfsvfs->z_vfs->vfs_flag & VFS_RDONLY) {
849 		zfs_dirent_unlock(dl);
850 		return (EROFS);
851 	}
852 
853 	/*
854 	 * The ability to 'create' files in an attribute
855 	 * directory comes from the write_xattr permission on the base file.
856 	 *
857 	 * The ability to 'search' an attribute directory requires
858 	 * read_xattr permission on the base file.
859 	 *
860 	 * Once in a directory the ability to read/write attributes
861 	 * is controlled by the permissions on the attribute file.
862 	 */
863 	va.va_mask = AT_TYPE | AT_MODE | AT_UID | AT_GID;
864 	va.va_type = VDIR;
865 	va.va_mode = S_IFDIR | S_ISVTX | 0777;
866 	va.va_uid = (uid_t)zp->z_phys->zp_uid;
867 	va.va_gid = (gid_t)zp->z_phys->zp_gid;
868 
869 	error = zfs_make_xattrdir(zp, &va, xvpp, cr);
870 	zfs_dirent_unlock(dl);
871 
872 	if (error == ERESTART && zfsvfs->z_assign == TXG_NOWAIT) {
873 		/* NB: we already did dmu_tx_wait() if necessary */
874 		goto top;
875 	}
876 
877 	return (error);
878 }
879 
880 /*
881  * Decide whether it is okay to remove within a sticky directory.
882  *
883  * In sticky directories, write access is not sufficient;
884  * you can remove entries from a directory only if:
885  *
886  *	you own the directory,
887  *	you own the entry,
888  *	the entry is a plain file and you have write access,
889  *	or you are privileged (checked in secpolicy...).
890  *
891  * The function returns 0 if remove access is granted.
892  */
893 int
894 zfs_sticky_remove_access(znode_t *zdp, znode_t *zp, cred_t *cr)
895 {
896 	uid_t  		uid;
897 
898 	if (zdp->z_zfsvfs->z_assign >= TXG_INITIAL)	/* ZIL replay */
899 		return (0);
900 
901 	if ((zdp->z_phys->zp_mode & S_ISVTX) == 0 ||
902 	    (uid = crgetuid(cr)) == zdp->z_phys->zp_uid ||
903 	    uid == zp->z_phys->zp_uid ||
904 	    (ZTOV(zp)->v_type == VREG &&
905 	    zfs_zaccess(zp, ACE_WRITE_DATA, cr) == 0))
906 		return (0);
907 	else
908 		return (secpolicy_vnode_remove(cr));
909 }
910