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 https://opensource.org/licenses/CDDL-1.0.
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 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
26  * Copyright 2017 Nexenta Systems, Inc.
27  */
28 
29 #include <sys/zio.h>
30 #include <sys/spa.h>
31 #include <sys/dmu.h>
32 #include <sys/zfs_context.h>
33 #include <sys/zap.h>
34 #include <sys/zap_impl.h>
35 #include <sys/zap_leaf.h>
36 #include <sys/avl.h>
37 #include <sys/arc.h>
38 #include <sys/dmu_objset.h>
39 
40 #ifdef _KERNEL
41 #include <sys/sunddi.h>
42 #endif
43 
44 static int mzap_upgrade(zap_t **zapp,
45     const void *tag, dmu_tx_t *tx, zap_flags_t flags);
46 
47 uint64_t
48 zap_getflags(zap_t *zap)
49 {
50 	if (zap->zap_ismicro)
51 		return (0);
52 	return (zap_f_phys(zap)->zap_flags);
53 }
54 
55 int
56 zap_hashbits(zap_t *zap)
57 {
58 	if (zap_getflags(zap) & ZAP_FLAG_HASH64)
59 		return (48);
60 	else
61 		return (28);
62 }
63 
64 uint32_t
65 zap_maxcd(zap_t *zap)
66 {
67 	if (zap_getflags(zap) & ZAP_FLAG_HASH64)
68 		return ((1<<16)-1);
69 	else
70 		return (-1U);
71 }
72 
73 static uint64_t
74 zap_hash(zap_name_t *zn)
75 {
76 	zap_t *zap = zn->zn_zap;
77 	uint64_t h = 0;
78 
79 	if (zap_getflags(zap) & ZAP_FLAG_PRE_HASHED_KEY) {
80 		ASSERT(zap_getflags(zap) & ZAP_FLAG_UINT64_KEY);
81 		h = *(uint64_t *)zn->zn_key_orig;
82 	} else {
83 		h = zap->zap_salt;
84 		ASSERT(h != 0);
85 		ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
86 
87 		if (zap_getflags(zap) & ZAP_FLAG_UINT64_KEY) {
88 			const uint64_t *wp = zn->zn_key_norm;
89 
90 			ASSERT(zn->zn_key_intlen == 8);
91 			for (int i = 0; i < zn->zn_key_norm_numints;
92 			    wp++, i++) {
93 				uint64_t word = *wp;
94 
95 				for (int j = 0; j < zn->zn_key_intlen; j++) {
96 					h = (h >> 8) ^
97 					    zfs_crc64_table[(h ^ word) & 0xFF];
98 					word >>= NBBY;
99 				}
100 			}
101 		} else {
102 			const uint8_t *cp = zn->zn_key_norm;
103 
104 			/*
105 			 * We previously stored the terminating null on
106 			 * disk, but didn't hash it, so we need to
107 			 * continue to not hash it.  (The
108 			 * zn_key_*_numints includes the terminating
109 			 * null for non-binary keys.)
110 			 */
111 			int len = zn->zn_key_norm_numints - 1;
112 
113 			ASSERT(zn->zn_key_intlen == 1);
114 			for (int i = 0; i < len; cp++, i++) {
115 				h = (h >> 8) ^
116 				    zfs_crc64_table[(h ^ *cp) & 0xFF];
117 			}
118 		}
119 	}
120 	/*
121 	 * Don't use all 64 bits, since we need some in the cookie for
122 	 * the collision differentiator.  We MUST use the high bits,
123 	 * since those are the ones that we first pay attention to when
124 	 * choosing the bucket.
125 	 */
126 	h &= ~((1ULL << (64 - zap_hashbits(zap))) - 1);
127 
128 	return (h);
129 }
130 
131 static int
132 zap_normalize(zap_t *zap, const char *name, char *namenorm, int normflags)
133 {
134 	ASSERT(!(zap_getflags(zap) & ZAP_FLAG_UINT64_KEY));
135 
136 	size_t inlen = strlen(name) + 1;
137 	size_t outlen = ZAP_MAXNAMELEN;
138 
139 	int err = 0;
140 	(void) u8_textprep_str((char *)name, &inlen, namenorm, &outlen,
141 	    normflags | U8_TEXTPREP_IGNORE_NULL | U8_TEXTPREP_IGNORE_INVALID,
142 	    U8_UNICODE_LATEST, &err);
143 
144 	return (err);
145 }
146 
147 boolean_t
148 zap_match(zap_name_t *zn, const char *matchname)
149 {
150 	ASSERT(!(zap_getflags(zn->zn_zap) & ZAP_FLAG_UINT64_KEY));
151 
152 	if (zn->zn_matchtype & MT_NORMALIZE) {
153 		char norm[ZAP_MAXNAMELEN];
154 
155 		if (zap_normalize(zn->zn_zap, matchname, norm,
156 		    zn->zn_normflags) != 0)
157 			return (B_FALSE);
158 
159 		return (strcmp(zn->zn_key_norm, norm) == 0);
160 	} else {
161 		return (strcmp(zn->zn_key_orig, matchname) == 0);
162 	}
163 }
164 
165 void
166 zap_name_free(zap_name_t *zn)
167 {
168 	kmem_free(zn, sizeof (zap_name_t));
169 }
170 
171 zap_name_t *
172 zap_name_alloc(zap_t *zap, const char *key, matchtype_t mt)
173 {
174 	zap_name_t *zn = kmem_alloc(sizeof (zap_name_t), KM_SLEEP);
175 
176 	zn->zn_zap = zap;
177 	zn->zn_key_intlen = sizeof (*key);
178 	zn->zn_key_orig = key;
179 	zn->zn_key_orig_numints = strlen(zn->zn_key_orig) + 1;
180 	zn->zn_matchtype = mt;
181 	zn->zn_normflags = zap->zap_normflags;
182 
183 	/*
184 	 * If we're dealing with a case sensitive lookup on a mixed or
185 	 * insensitive fs, remove U8_TEXTPREP_TOUPPER or the lookup
186 	 * will fold case to all caps overriding the lookup request.
187 	 */
188 	if (mt & MT_MATCH_CASE)
189 		zn->zn_normflags &= ~U8_TEXTPREP_TOUPPER;
190 
191 	if (zap->zap_normflags) {
192 		/*
193 		 * We *must* use zap_normflags because this normalization is
194 		 * what the hash is computed from.
195 		 */
196 		if (zap_normalize(zap, key, zn->zn_normbuf,
197 		    zap->zap_normflags) != 0) {
198 			zap_name_free(zn);
199 			return (NULL);
200 		}
201 		zn->zn_key_norm = zn->zn_normbuf;
202 		zn->zn_key_norm_numints = strlen(zn->zn_key_norm) + 1;
203 	} else {
204 		if (mt != 0) {
205 			zap_name_free(zn);
206 			return (NULL);
207 		}
208 		zn->zn_key_norm = zn->zn_key_orig;
209 		zn->zn_key_norm_numints = zn->zn_key_orig_numints;
210 	}
211 
212 	zn->zn_hash = zap_hash(zn);
213 
214 	if (zap->zap_normflags != zn->zn_normflags) {
215 		/*
216 		 * We *must* use zn_normflags because this normalization is
217 		 * what the matching is based on.  (Not the hash!)
218 		 */
219 		if (zap_normalize(zap, key, zn->zn_normbuf,
220 		    zn->zn_normflags) != 0) {
221 			zap_name_free(zn);
222 			return (NULL);
223 		}
224 		zn->zn_key_norm_numints = strlen(zn->zn_key_norm) + 1;
225 	}
226 
227 	return (zn);
228 }
229 
230 static zap_name_t *
231 zap_name_alloc_uint64(zap_t *zap, const uint64_t *key, int numints)
232 {
233 	zap_name_t *zn = kmem_alloc(sizeof (zap_name_t), KM_SLEEP);
234 
235 	ASSERT(zap->zap_normflags == 0);
236 	zn->zn_zap = zap;
237 	zn->zn_key_intlen = sizeof (*key);
238 	zn->zn_key_orig = zn->zn_key_norm = key;
239 	zn->zn_key_orig_numints = zn->zn_key_norm_numints = numints;
240 	zn->zn_matchtype = 0;
241 
242 	zn->zn_hash = zap_hash(zn);
243 	return (zn);
244 }
245 
246 static void
247 mzap_byteswap(mzap_phys_t *buf, size_t size)
248 {
249 	buf->mz_block_type = BSWAP_64(buf->mz_block_type);
250 	buf->mz_salt = BSWAP_64(buf->mz_salt);
251 	buf->mz_normflags = BSWAP_64(buf->mz_normflags);
252 	int max = (size / MZAP_ENT_LEN) - 1;
253 	for (int i = 0; i < max; i++) {
254 		buf->mz_chunk[i].mze_value =
255 		    BSWAP_64(buf->mz_chunk[i].mze_value);
256 		buf->mz_chunk[i].mze_cd =
257 		    BSWAP_32(buf->mz_chunk[i].mze_cd);
258 	}
259 }
260 
261 void
262 zap_byteswap(void *buf, size_t size)
263 {
264 	uint64_t block_type = *(uint64_t *)buf;
265 
266 	if (block_type == ZBT_MICRO || block_type == BSWAP_64(ZBT_MICRO)) {
267 		/* ASSERT(magic == ZAP_LEAF_MAGIC); */
268 		mzap_byteswap(buf, size);
269 	} else {
270 		fzap_byteswap(buf, size);
271 	}
272 }
273 
274 static int
275 mze_compare(const void *arg1, const void *arg2)
276 {
277 	const mzap_ent_t *mze1 = arg1;
278 	const mzap_ent_t *mze2 = arg2;
279 
280 	int cmp = TREE_CMP(mze1->mze_hash, mze2->mze_hash);
281 	if (likely(cmp))
282 		return (cmp);
283 
284 	return (TREE_CMP(mze1->mze_cd, mze2->mze_cd));
285 }
286 
287 static void
288 mze_insert(zap_t *zap, int chunkid, uint64_t hash)
289 {
290 	ASSERT(zap->zap_ismicro);
291 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
292 
293 	mzap_ent_t *mze = kmem_alloc(sizeof (mzap_ent_t), KM_SLEEP);
294 	mze->mze_chunkid = chunkid;
295 	mze->mze_hash = hash;
296 	mze->mze_cd = MZE_PHYS(zap, mze)->mze_cd;
297 	ASSERT(MZE_PHYS(zap, mze)->mze_name[0] != 0);
298 	avl_add(&zap->zap_m.zap_avl, mze);
299 }
300 
301 static mzap_ent_t *
302 mze_find(zap_name_t *zn)
303 {
304 	mzap_ent_t mze_tofind;
305 	mzap_ent_t *mze;
306 	avl_index_t idx;
307 	avl_tree_t *avl = &zn->zn_zap->zap_m.zap_avl;
308 
309 	ASSERT(zn->zn_zap->zap_ismicro);
310 	ASSERT(RW_LOCK_HELD(&zn->zn_zap->zap_rwlock));
311 
312 	mze_tofind.mze_hash = zn->zn_hash;
313 	mze_tofind.mze_cd = 0;
314 
315 	mze = avl_find(avl, &mze_tofind, &idx);
316 	if (mze == NULL)
317 		mze = avl_nearest(avl, idx, AVL_AFTER);
318 	for (; mze && mze->mze_hash == zn->zn_hash; mze = AVL_NEXT(avl, mze)) {
319 		ASSERT3U(mze->mze_cd, ==, MZE_PHYS(zn->zn_zap, mze)->mze_cd);
320 		if (zap_match(zn, MZE_PHYS(zn->zn_zap, mze)->mze_name))
321 			return (mze);
322 	}
323 
324 	return (NULL);
325 }
326 
327 static uint32_t
328 mze_find_unused_cd(zap_t *zap, uint64_t hash)
329 {
330 	mzap_ent_t mze_tofind;
331 	avl_index_t idx;
332 	avl_tree_t *avl = &zap->zap_m.zap_avl;
333 
334 	ASSERT(zap->zap_ismicro);
335 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
336 
337 	mze_tofind.mze_hash = hash;
338 	mze_tofind.mze_cd = 0;
339 
340 	uint32_t cd = 0;
341 	for (mzap_ent_t *mze = avl_find(avl, &mze_tofind, &idx);
342 	    mze && mze->mze_hash == hash; mze = AVL_NEXT(avl, mze)) {
343 		if (mze->mze_cd != cd)
344 			break;
345 		cd++;
346 	}
347 
348 	return (cd);
349 }
350 
351 /*
352  * Each mzap entry requires at max : 4 chunks
353  * 3 chunks for names + 1 chunk for value.
354  */
355 #define	MZAP_ENT_CHUNKS	(1 + ZAP_LEAF_ARRAY_NCHUNKS(MZAP_NAME_LEN) + \
356 	ZAP_LEAF_ARRAY_NCHUNKS(sizeof (uint64_t)))
357 
358 /*
359  * Check if the current entry keeps the colliding entries under the fatzap leaf
360  * size.
361  */
362 static boolean_t
363 mze_canfit_fzap_leaf(zap_name_t *zn, uint64_t hash)
364 {
365 	zap_t *zap = zn->zn_zap;
366 	mzap_ent_t mze_tofind;
367 	mzap_ent_t *mze;
368 	avl_index_t idx;
369 	avl_tree_t *avl = &zap->zap_m.zap_avl;
370 	uint32_t mzap_ents = 0;
371 
372 	mze_tofind.mze_hash = hash;
373 	mze_tofind.mze_cd = 0;
374 
375 	for (mze = avl_find(avl, &mze_tofind, &idx);
376 	    mze && mze->mze_hash == hash; mze = AVL_NEXT(avl, mze)) {
377 		mzap_ents++;
378 	}
379 
380 	/* Include the new entry being added */
381 	mzap_ents++;
382 
383 	return (ZAP_LEAF_NUMCHUNKS_DEF > (mzap_ents * MZAP_ENT_CHUNKS));
384 }
385 
386 static void
387 mze_remove(zap_t *zap, mzap_ent_t *mze)
388 {
389 	ASSERT(zap->zap_ismicro);
390 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
391 
392 	avl_remove(&zap->zap_m.zap_avl, mze);
393 	kmem_free(mze, sizeof (mzap_ent_t));
394 }
395 
396 static void
397 mze_destroy(zap_t *zap)
398 {
399 	mzap_ent_t *mze;
400 	void *avlcookie = NULL;
401 
402 	while ((mze = avl_destroy_nodes(&zap->zap_m.zap_avl, &avlcookie)))
403 		kmem_free(mze, sizeof (mzap_ent_t));
404 	avl_destroy(&zap->zap_m.zap_avl);
405 }
406 
407 static zap_t *
408 mzap_open(objset_t *os, uint64_t obj, dmu_buf_t *db)
409 {
410 	zap_t *winner;
411 	uint64_t *zap_hdr = (uint64_t *)db->db_data;
412 	uint64_t zap_block_type = zap_hdr[0];
413 	uint64_t zap_magic = zap_hdr[1];
414 
415 	ASSERT3U(MZAP_ENT_LEN, ==, sizeof (mzap_ent_phys_t));
416 
417 	zap_t *zap = kmem_zalloc(sizeof (zap_t), KM_SLEEP);
418 	rw_init(&zap->zap_rwlock, NULL, RW_DEFAULT, NULL);
419 	rw_enter(&zap->zap_rwlock, RW_WRITER);
420 	zap->zap_objset = os;
421 	zap->zap_object = obj;
422 	zap->zap_dbuf = db;
423 
424 	if (zap_block_type != ZBT_MICRO) {
425 		mutex_init(&zap->zap_f.zap_num_entries_mtx, 0, MUTEX_DEFAULT,
426 		    0);
427 		zap->zap_f.zap_block_shift = highbit64(db->db_size) - 1;
428 		if (zap_block_type != ZBT_HEADER || zap_magic != ZAP_MAGIC) {
429 			winner = NULL;	/* No actual winner here... */
430 			goto handle_winner;
431 		}
432 	} else {
433 		zap->zap_ismicro = TRUE;
434 	}
435 
436 	/*
437 	 * Make sure that zap_ismicro is set before we let others see
438 	 * it, because zap_lockdir() checks zap_ismicro without the lock
439 	 * held.
440 	 */
441 	dmu_buf_init_user(&zap->zap_dbu, zap_evict_sync, NULL, &zap->zap_dbuf);
442 	winner = dmu_buf_set_user(db, &zap->zap_dbu);
443 
444 	if (winner != NULL)
445 		goto handle_winner;
446 
447 	if (zap->zap_ismicro) {
448 		zap->zap_salt = zap_m_phys(zap)->mz_salt;
449 		zap->zap_normflags = zap_m_phys(zap)->mz_normflags;
450 		zap->zap_m.zap_num_chunks = db->db_size / MZAP_ENT_LEN - 1;
451 		avl_create(&zap->zap_m.zap_avl, mze_compare,
452 		    sizeof (mzap_ent_t), offsetof(mzap_ent_t, mze_node));
453 
454 		for (int i = 0; i < zap->zap_m.zap_num_chunks; i++) {
455 			mzap_ent_phys_t *mze =
456 			    &zap_m_phys(zap)->mz_chunk[i];
457 			if (mze->mze_name[0]) {
458 				zap_name_t *zn;
459 
460 				zap->zap_m.zap_num_entries++;
461 				zn = zap_name_alloc(zap, mze->mze_name, 0);
462 				mze_insert(zap, i, zn->zn_hash);
463 				zap_name_free(zn);
464 			}
465 		}
466 	} else {
467 		zap->zap_salt = zap_f_phys(zap)->zap_salt;
468 		zap->zap_normflags = zap_f_phys(zap)->zap_normflags;
469 
470 		ASSERT3U(sizeof (struct zap_leaf_header), ==,
471 		    2*ZAP_LEAF_CHUNKSIZE);
472 
473 		/*
474 		 * The embedded pointer table should not overlap the
475 		 * other members.
476 		 */
477 		ASSERT3P(&ZAP_EMBEDDED_PTRTBL_ENT(zap, 0), >,
478 		    &zap_f_phys(zap)->zap_salt);
479 
480 		/*
481 		 * The embedded pointer table should end at the end of
482 		 * the block
483 		 */
484 		ASSERT3U((uintptr_t)&ZAP_EMBEDDED_PTRTBL_ENT(zap,
485 		    1<<ZAP_EMBEDDED_PTRTBL_SHIFT(zap)) -
486 		    (uintptr_t)zap_f_phys(zap), ==,
487 		    zap->zap_dbuf->db_size);
488 	}
489 	rw_exit(&zap->zap_rwlock);
490 	return (zap);
491 
492 handle_winner:
493 	rw_exit(&zap->zap_rwlock);
494 	rw_destroy(&zap->zap_rwlock);
495 	if (!zap->zap_ismicro)
496 		mutex_destroy(&zap->zap_f.zap_num_entries_mtx);
497 	kmem_free(zap, sizeof (zap_t));
498 	return (winner);
499 }
500 
501 /*
502  * This routine "consumes" the caller's hold on the dbuf, which must
503  * have the specified tag.
504  */
505 static int
506 zap_lockdir_impl(dmu_buf_t *db, const void *tag, dmu_tx_t *tx,
507     krw_t lti, boolean_t fatreader, boolean_t adding, zap_t **zapp)
508 {
509 	ASSERT0(db->db_offset);
510 	objset_t *os = dmu_buf_get_objset(db);
511 	uint64_t obj = db->db_object;
512 	dmu_object_info_t doi;
513 
514 	*zapp = NULL;
515 
516 	dmu_object_info_from_db(db, &doi);
517 	if (DMU_OT_BYTESWAP(doi.doi_type) != DMU_BSWAP_ZAP)
518 		return (SET_ERROR(EINVAL));
519 
520 	zap_t *zap = dmu_buf_get_user(db);
521 	if (zap == NULL) {
522 		zap = mzap_open(os, obj, db);
523 		if (zap == NULL) {
524 			/*
525 			 * mzap_open() didn't like what it saw on-disk.
526 			 * Check for corruption!
527 			 */
528 			return (SET_ERROR(EIO));
529 		}
530 	}
531 
532 	/*
533 	 * We're checking zap_ismicro without the lock held, in order to
534 	 * tell what type of lock we want.  Once we have some sort of
535 	 * lock, see if it really is the right type.  In practice this
536 	 * can only be different if it was upgraded from micro to fat,
537 	 * and micro wanted WRITER but fat only needs READER.
538 	 */
539 	krw_t lt = (!zap->zap_ismicro && fatreader) ? RW_READER : lti;
540 	rw_enter(&zap->zap_rwlock, lt);
541 	if (lt != ((!zap->zap_ismicro && fatreader) ? RW_READER : lti)) {
542 		/* it was upgraded, now we only need reader */
543 		ASSERT(lt == RW_WRITER);
544 		ASSERT(RW_READER ==
545 		    ((!zap->zap_ismicro && fatreader) ? RW_READER : lti));
546 		rw_downgrade(&zap->zap_rwlock);
547 		lt = RW_READER;
548 	}
549 
550 	zap->zap_objset = os;
551 
552 	if (lt == RW_WRITER)
553 		dmu_buf_will_dirty(db, tx);
554 
555 	ASSERT3P(zap->zap_dbuf, ==, db);
556 
557 	ASSERT(!zap->zap_ismicro ||
558 	    zap->zap_m.zap_num_entries <= zap->zap_m.zap_num_chunks);
559 	if (zap->zap_ismicro && tx && adding &&
560 	    zap->zap_m.zap_num_entries == zap->zap_m.zap_num_chunks) {
561 		uint64_t newsz = db->db_size + SPA_MINBLOCKSIZE;
562 		if (newsz > MZAP_MAX_BLKSZ) {
563 			dprintf("upgrading obj %llu: num_entries=%u\n",
564 			    (u_longlong_t)obj, zap->zap_m.zap_num_entries);
565 			*zapp = zap;
566 			int err = mzap_upgrade(zapp, tag, tx, 0);
567 			if (err != 0)
568 				rw_exit(&zap->zap_rwlock);
569 			return (err);
570 		}
571 		VERIFY0(dmu_object_set_blocksize(os, obj, newsz, 0, tx));
572 		zap->zap_m.zap_num_chunks =
573 		    db->db_size / MZAP_ENT_LEN - 1;
574 	}
575 
576 	*zapp = zap;
577 	return (0);
578 }
579 
580 static int
581 zap_lockdir_by_dnode(dnode_t *dn, dmu_tx_t *tx,
582     krw_t lti, boolean_t fatreader, boolean_t adding, const void *tag,
583     zap_t **zapp)
584 {
585 	dmu_buf_t *db;
586 
587 	int err = dmu_buf_hold_by_dnode(dn, 0, tag, &db, DMU_READ_NO_PREFETCH);
588 	if (err != 0) {
589 		return (err);
590 	}
591 #ifdef ZFS_DEBUG
592 	{
593 		dmu_object_info_t doi;
594 		dmu_object_info_from_db(db, &doi);
595 		ASSERT3U(DMU_OT_BYTESWAP(doi.doi_type), ==, DMU_BSWAP_ZAP);
596 	}
597 #endif
598 
599 	err = zap_lockdir_impl(db, tag, tx, lti, fatreader, adding, zapp);
600 	if (err != 0) {
601 		dmu_buf_rele(db, tag);
602 	}
603 	return (err);
604 }
605 
606 int
607 zap_lockdir(objset_t *os, uint64_t obj, dmu_tx_t *tx,
608     krw_t lti, boolean_t fatreader, boolean_t adding, const void *tag,
609     zap_t **zapp)
610 {
611 	dmu_buf_t *db;
612 
613 	int err = dmu_buf_hold(os, obj, 0, tag, &db, DMU_READ_NO_PREFETCH);
614 	if (err != 0)
615 		return (err);
616 #ifdef ZFS_DEBUG
617 	{
618 		dmu_object_info_t doi;
619 		dmu_object_info_from_db(db, &doi);
620 		ASSERT3U(DMU_OT_BYTESWAP(doi.doi_type), ==, DMU_BSWAP_ZAP);
621 	}
622 #endif
623 	err = zap_lockdir_impl(db, tag, tx, lti, fatreader, adding, zapp);
624 	if (err != 0)
625 		dmu_buf_rele(db, tag);
626 	return (err);
627 }
628 
629 void
630 zap_unlockdir(zap_t *zap, const void *tag)
631 {
632 	rw_exit(&zap->zap_rwlock);
633 	dmu_buf_rele(zap->zap_dbuf, tag);
634 }
635 
636 static int
637 mzap_upgrade(zap_t **zapp, const void *tag, dmu_tx_t *tx, zap_flags_t flags)
638 {
639 	int err = 0;
640 	zap_t *zap = *zapp;
641 
642 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
643 
644 	int sz = zap->zap_dbuf->db_size;
645 	mzap_phys_t *mzp = vmem_alloc(sz, KM_SLEEP);
646 	memcpy(mzp, zap->zap_dbuf->db_data, sz);
647 	int nchunks = zap->zap_m.zap_num_chunks;
648 
649 	if (!flags) {
650 		err = dmu_object_set_blocksize(zap->zap_objset, zap->zap_object,
651 		    1ULL << fzap_default_block_shift, 0, tx);
652 		if (err != 0) {
653 			vmem_free(mzp, sz);
654 			return (err);
655 		}
656 	}
657 
658 	dprintf("upgrading obj=%llu with %u chunks\n",
659 	    (u_longlong_t)zap->zap_object, nchunks);
660 	/* XXX destroy the avl later, so we can use the stored hash value */
661 	mze_destroy(zap);
662 
663 	fzap_upgrade(zap, tx, flags);
664 
665 	for (int i = 0; i < nchunks; i++) {
666 		mzap_ent_phys_t *mze = &mzp->mz_chunk[i];
667 		if (mze->mze_name[0] == 0)
668 			continue;
669 		dprintf("adding %s=%llu\n",
670 		    mze->mze_name, (u_longlong_t)mze->mze_value);
671 		zap_name_t *zn = zap_name_alloc(zap, mze->mze_name, 0);
672 		/* If we fail here, we would end up losing entries */
673 		VERIFY0(fzap_add_cd(zn, 8, 1, &mze->mze_value, mze->mze_cd,
674 		    tag, tx));
675 		zap = zn->zn_zap;	/* fzap_add_cd() may change zap */
676 		zap_name_free(zn);
677 	}
678 	vmem_free(mzp, sz);
679 	*zapp = zap;
680 	return (0);
681 }
682 
683 /*
684  * The "normflags" determine the behavior of the matchtype_t which is
685  * passed to zap_lookup_norm().  Names which have the same normalized
686  * version will be stored with the same hash value, and therefore we can
687  * perform normalization-insensitive lookups.  We can be Unicode form-
688  * insensitive and/or case-insensitive.  The following flags are valid for
689  * "normflags":
690  *
691  * U8_TEXTPREP_NFC
692  * U8_TEXTPREP_NFD
693  * U8_TEXTPREP_NFKC
694  * U8_TEXTPREP_NFKD
695  * U8_TEXTPREP_TOUPPER
696  *
697  * The *_NF* (Normalization Form) flags are mutually exclusive; at most one
698  * of them may be supplied.
699  */
700 void
701 mzap_create_impl(dnode_t *dn, int normflags, zap_flags_t flags, dmu_tx_t *tx)
702 {
703 	dmu_buf_t *db;
704 
705 	VERIFY0(dmu_buf_hold_by_dnode(dn, 0, FTAG, &db, DMU_READ_NO_PREFETCH));
706 
707 	dmu_buf_will_dirty(db, tx);
708 	mzap_phys_t *zp = db->db_data;
709 	zp->mz_block_type = ZBT_MICRO;
710 	zp->mz_salt =
711 	    ((uintptr_t)db ^ (uintptr_t)tx ^ (dn->dn_object << 1)) | 1ULL;
712 	zp->mz_normflags = normflags;
713 
714 	if (flags != 0) {
715 		zap_t *zap;
716 		/* Only fat zap supports flags; upgrade immediately. */
717 		VERIFY0(zap_lockdir_impl(db, FTAG, tx, RW_WRITER,
718 		    B_FALSE, B_FALSE, &zap));
719 		VERIFY0(mzap_upgrade(&zap, FTAG, tx, flags));
720 		zap_unlockdir(zap, FTAG);
721 	} else {
722 		dmu_buf_rele(db, FTAG);
723 	}
724 }
725 
726 static uint64_t
727 zap_create_impl(objset_t *os, int normflags, zap_flags_t flags,
728     dmu_object_type_t ot, int leaf_blockshift, int indirect_blockshift,
729     dmu_object_type_t bonustype, int bonuslen, int dnodesize,
730     dnode_t **allocated_dnode, const void *tag, dmu_tx_t *tx)
731 {
732 	uint64_t obj;
733 
734 	ASSERT3U(DMU_OT_BYTESWAP(ot), ==, DMU_BSWAP_ZAP);
735 
736 	if (allocated_dnode == NULL) {
737 		dnode_t *dn;
738 		obj = dmu_object_alloc_hold(os, ot, 1ULL << leaf_blockshift,
739 		    indirect_blockshift, bonustype, bonuslen, dnodesize,
740 		    &dn, FTAG, tx);
741 		mzap_create_impl(dn, normflags, flags, tx);
742 		dnode_rele(dn, FTAG);
743 	} else {
744 		obj = dmu_object_alloc_hold(os, ot, 1ULL << leaf_blockshift,
745 		    indirect_blockshift, bonustype, bonuslen, dnodesize,
746 		    allocated_dnode, tag, tx);
747 		mzap_create_impl(*allocated_dnode, normflags, flags, tx);
748 	}
749 
750 	return (obj);
751 }
752 
753 int
754 zap_create_claim(objset_t *os, uint64_t obj, dmu_object_type_t ot,
755     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
756 {
757 	return (zap_create_claim_dnsize(os, obj, ot, bonustype, bonuslen,
758 	    0, tx));
759 }
760 
761 int
762 zap_create_claim_dnsize(objset_t *os, uint64_t obj, dmu_object_type_t ot,
763     dmu_object_type_t bonustype, int bonuslen, int dnodesize, dmu_tx_t *tx)
764 {
765 	return (zap_create_claim_norm_dnsize(os, obj,
766 	    0, ot, bonustype, bonuslen, dnodesize, tx));
767 }
768 
769 int
770 zap_create_claim_norm(objset_t *os, uint64_t obj, int normflags,
771     dmu_object_type_t ot,
772     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
773 {
774 	return (zap_create_claim_norm_dnsize(os, obj, normflags, ot, bonustype,
775 	    bonuslen, 0, tx));
776 }
777 
778 int
779 zap_create_claim_norm_dnsize(objset_t *os, uint64_t obj, int normflags,
780     dmu_object_type_t ot, dmu_object_type_t bonustype, int bonuslen,
781     int dnodesize, dmu_tx_t *tx)
782 {
783 	dnode_t *dn;
784 	int error;
785 
786 	ASSERT3U(DMU_OT_BYTESWAP(ot), ==, DMU_BSWAP_ZAP);
787 	error = dmu_object_claim_dnsize(os, obj, ot, 0, bonustype, bonuslen,
788 	    dnodesize, tx);
789 	if (error != 0)
790 		return (error);
791 
792 	error = dnode_hold(os, obj, FTAG, &dn);
793 	if (error != 0)
794 		return (error);
795 
796 	mzap_create_impl(dn, normflags, 0, tx);
797 
798 	dnode_rele(dn, FTAG);
799 
800 	return (0);
801 }
802 
803 uint64_t
804 zap_create(objset_t *os, dmu_object_type_t ot,
805     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
806 {
807 	return (zap_create_norm(os, 0, ot, bonustype, bonuslen, tx));
808 }
809 
810 uint64_t
811 zap_create_dnsize(objset_t *os, dmu_object_type_t ot,
812     dmu_object_type_t bonustype, int bonuslen, int dnodesize, dmu_tx_t *tx)
813 {
814 	return (zap_create_norm_dnsize(os, 0, ot, bonustype, bonuslen,
815 	    dnodesize, tx));
816 }
817 
818 uint64_t
819 zap_create_norm(objset_t *os, int normflags, dmu_object_type_t ot,
820     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
821 {
822 	return (zap_create_norm_dnsize(os, normflags, ot, bonustype, bonuslen,
823 	    0, tx));
824 }
825 
826 uint64_t
827 zap_create_norm_dnsize(objset_t *os, int normflags, dmu_object_type_t ot,
828     dmu_object_type_t bonustype, int bonuslen, int dnodesize, dmu_tx_t *tx)
829 {
830 	return (zap_create_impl(os, normflags, 0, ot, 0, 0,
831 	    bonustype, bonuslen, dnodesize, NULL, NULL, tx));
832 }
833 
834 uint64_t
835 zap_create_flags(objset_t *os, int normflags, zap_flags_t flags,
836     dmu_object_type_t ot, int leaf_blockshift, int indirect_blockshift,
837     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
838 {
839 	return (zap_create_flags_dnsize(os, normflags, flags, ot,
840 	    leaf_blockshift, indirect_blockshift, bonustype, bonuslen, 0, tx));
841 }
842 
843 uint64_t
844 zap_create_flags_dnsize(objset_t *os, int normflags, zap_flags_t flags,
845     dmu_object_type_t ot, int leaf_blockshift, int indirect_blockshift,
846     dmu_object_type_t bonustype, int bonuslen, int dnodesize, dmu_tx_t *tx)
847 {
848 	return (zap_create_impl(os, normflags, flags, ot, leaf_blockshift,
849 	    indirect_blockshift, bonustype, bonuslen, dnodesize, NULL, NULL,
850 	    tx));
851 }
852 
853 /*
854  * Create a zap object and return a pointer to the newly allocated dnode via
855  * the allocated_dnode argument.  The returned dnode will be held and the
856  * caller is responsible for releasing the hold by calling dnode_rele().
857  */
858 uint64_t
859 zap_create_hold(objset_t *os, int normflags, zap_flags_t flags,
860     dmu_object_type_t ot, int leaf_blockshift, int indirect_blockshift,
861     dmu_object_type_t bonustype, int bonuslen, int dnodesize,
862     dnode_t **allocated_dnode, const void *tag, dmu_tx_t *tx)
863 {
864 	return (zap_create_impl(os, normflags, flags, ot, leaf_blockshift,
865 	    indirect_blockshift, bonustype, bonuslen, dnodesize,
866 	    allocated_dnode, tag, tx));
867 }
868 
869 int
870 zap_destroy(objset_t *os, uint64_t zapobj, dmu_tx_t *tx)
871 {
872 	/*
873 	 * dmu_object_free will free the object number and free the
874 	 * data.  Freeing the data will cause our pageout function to be
875 	 * called, which will destroy our data (zap_leaf_t's and zap_t).
876 	 */
877 
878 	return (dmu_object_free(os, zapobj, tx));
879 }
880 
881 void
882 zap_evict_sync(void *dbu)
883 {
884 	zap_t *zap = dbu;
885 
886 	rw_destroy(&zap->zap_rwlock);
887 
888 	if (zap->zap_ismicro)
889 		mze_destroy(zap);
890 	else
891 		mutex_destroy(&zap->zap_f.zap_num_entries_mtx);
892 
893 	kmem_free(zap, sizeof (zap_t));
894 }
895 
896 int
897 zap_count(objset_t *os, uint64_t zapobj, uint64_t *count)
898 {
899 	zap_t *zap;
900 
901 	int err =
902 	    zap_lockdir(os, zapobj, NULL, RW_READER, TRUE, FALSE, FTAG, &zap);
903 	if (err != 0)
904 		return (err);
905 	if (!zap->zap_ismicro) {
906 		err = fzap_count(zap, count);
907 	} else {
908 		*count = zap->zap_m.zap_num_entries;
909 	}
910 	zap_unlockdir(zap, FTAG);
911 	return (err);
912 }
913 
914 /*
915  * zn may be NULL; if not specified, it will be computed if needed.
916  * See also the comment above zap_entry_normalization_conflict().
917  */
918 static boolean_t
919 mzap_normalization_conflict(zap_t *zap, zap_name_t *zn, mzap_ent_t *mze)
920 {
921 	int direction = AVL_BEFORE;
922 	boolean_t allocdzn = B_FALSE;
923 
924 	if (zap->zap_normflags == 0)
925 		return (B_FALSE);
926 
927 again:
928 	for (mzap_ent_t *other = avl_walk(&zap->zap_m.zap_avl, mze, direction);
929 	    other && other->mze_hash == mze->mze_hash;
930 	    other = avl_walk(&zap->zap_m.zap_avl, other, direction)) {
931 
932 		if (zn == NULL) {
933 			zn = zap_name_alloc(zap, MZE_PHYS(zap, mze)->mze_name,
934 			    MT_NORMALIZE);
935 			allocdzn = B_TRUE;
936 		}
937 		if (zap_match(zn, MZE_PHYS(zap, other)->mze_name)) {
938 			if (allocdzn)
939 				zap_name_free(zn);
940 			return (B_TRUE);
941 		}
942 	}
943 
944 	if (direction == AVL_BEFORE) {
945 		direction = AVL_AFTER;
946 		goto again;
947 	}
948 
949 	if (allocdzn)
950 		zap_name_free(zn);
951 	return (B_FALSE);
952 }
953 
954 /*
955  * Routines for manipulating attributes.
956  */
957 
958 int
959 zap_lookup(objset_t *os, uint64_t zapobj, const char *name,
960     uint64_t integer_size, uint64_t num_integers, void *buf)
961 {
962 	return (zap_lookup_norm(os, zapobj, name, integer_size,
963 	    num_integers, buf, 0, NULL, 0, NULL));
964 }
965 
966 static int
967 zap_lookup_impl(zap_t *zap, const char *name,
968     uint64_t integer_size, uint64_t num_integers, void *buf,
969     matchtype_t mt, char *realname, int rn_len,
970     boolean_t *ncp)
971 {
972 	int err = 0;
973 
974 	zap_name_t *zn = zap_name_alloc(zap, name, mt);
975 	if (zn == NULL)
976 		return (SET_ERROR(ENOTSUP));
977 
978 	if (!zap->zap_ismicro) {
979 		err = fzap_lookup(zn, integer_size, num_integers, buf,
980 		    realname, rn_len, ncp);
981 	} else {
982 		mzap_ent_t *mze = mze_find(zn);
983 		if (mze == NULL) {
984 			err = SET_ERROR(ENOENT);
985 		} else {
986 			if (num_integers < 1) {
987 				err = SET_ERROR(EOVERFLOW);
988 			} else if (integer_size != 8) {
989 				err = SET_ERROR(EINVAL);
990 			} else {
991 				*(uint64_t *)buf =
992 				    MZE_PHYS(zap, mze)->mze_value;
993 				(void) strlcpy(realname,
994 				    MZE_PHYS(zap, mze)->mze_name, rn_len);
995 				if (ncp) {
996 					*ncp = mzap_normalization_conflict(zap,
997 					    zn, mze);
998 				}
999 			}
1000 		}
1001 	}
1002 	zap_name_free(zn);
1003 	return (err);
1004 }
1005 
1006 int
1007 zap_lookup_norm(objset_t *os, uint64_t zapobj, const char *name,
1008     uint64_t integer_size, uint64_t num_integers, void *buf,
1009     matchtype_t mt, char *realname, int rn_len,
1010     boolean_t *ncp)
1011 {
1012 	zap_t *zap;
1013 
1014 	int err =
1015 	    zap_lockdir(os, zapobj, NULL, RW_READER, TRUE, FALSE, FTAG, &zap);
1016 	if (err != 0)
1017 		return (err);
1018 	err = zap_lookup_impl(zap, name, integer_size,
1019 	    num_integers, buf, mt, realname, rn_len, ncp);
1020 	zap_unlockdir(zap, FTAG);
1021 	return (err);
1022 }
1023 
1024 int
1025 zap_prefetch(objset_t *os, uint64_t zapobj, const char *name)
1026 {
1027 	zap_t *zap;
1028 	int err;
1029 	zap_name_t *zn;
1030 
1031 	err = zap_lockdir(os, zapobj, NULL, RW_READER, TRUE, FALSE, FTAG, &zap);
1032 	if (err)
1033 		return (err);
1034 	zn = zap_name_alloc(zap, name, 0);
1035 	if (zn == NULL) {
1036 		zap_unlockdir(zap, FTAG);
1037 		return (SET_ERROR(ENOTSUP));
1038 	}
1039 
1040 	fzap_prefetch(zn);
1041 	zap_name_free(zn);
1042 	zap_unlockdir(zap, FTAG);
1043 	return (err);
1044 }
1045 
1046 int
1047 zap_lookup_by_dnode(dnode_t *dn, const char *name,
1048     uint64_t integer_size, uint64_t num_integers, void *buf)
1049 {
1050 	return (zap_lookup_norm_by_dnode(dn, name, integer_size,
1051 	    num_integers, buf, 0, NULL, 0, NULL));
1052 }
1053 
1054 int
1055 zap_lookup_norm_by_dnode(dnode_t *dn, const char *name,
1056     uint64_t integer_size, uint64_t num_integers, void *buf,
1057     matchtype_t mt, char *realname, int rn_len,
1058     boolean_t *ncp)
1059 {
1060 	zap_t *zap;
1061 
1062 	int err = zap_lockdir_by_dnode(dn, NULL, RW_READER, TRUE, FALSE,
1063 	    FTAG, &zap);
1064 	if (err != 0)
1065 		return (err);
1066 	err = zap_lookup_impl(zap, name, integer_size,
1067 	    num_integers, buf, mt, realname, rn_len, ncp);
1068 	zap_unlockdir(zap, FTAG);
1069 	return (err);
1070 }
1071 
1072 int
1073 zap_prefetch_uint64(objset_t *os, uint64_t zapobj, const uint64_t *key,
1074     int key_numints)
1075 {
1076 	zap_t *zap;
1077 
1078 	int err =
1079 	    zap_lockdir(os, zapobj, NULL, RW_READER, TRUE, FALSE, FTAG, &zap);
1080 	if (err != 0)
1081 		return (err);
1082 	zap_name_t *zn = zap_name_alloc_uint64(zap, key, key_numints);
1083 	if (zn == NULL) {
1084 		zap_unlockdir(zap, FTAG);
1085 		return (SET_ERROR(ENOTSUP));
1086 	}
1087 
1088 	fzap_prefetch(zn);
1089 	zap_name_free(zn);
1090 	zap_unlockdir(zap, FTAG);
1091 	return (err);
1092 }
1093 
1094 int
1095 zap_lookup_uint64(objset_t *os, uint64_t zapobj, const uint64_t *key,
1096     int key_numints, uint64_t integer_size, uint64_t num_integers, void *buf)
1097 {
1098 	zap_t *zap;
1099 
1100 	int err =
1101 	    zap_lockdir(os, zapobj, NULL, RW_READER, TRUE, FALSE, FTAG, &zap);
1102 	if (err != 0)
1103 		return (err);
1104 	zap_name_t *zn = zap_name_alloc_uint64(zap, key, key_numints);
1105 	if (zn == NULL) {
1106 		zap_unlockdir(zap, FTAG);
1107 		return (SET_ERROR(ENOTSUP));
1108 	}
1109 
1110 	err = fzap_lookup(zn, integer_size, num_integers, buf,
1111 	    NULL, 0, NULL);
1112 	zap_name_free(zn);
1113 	zap_unlockdir(zap, FTAG);
1114 	return (err);
1115 }
1116 
1117 int
1118 zap_contains(objset_t *os, uint64_t zapobj, const char *name)
1119 {
1120 	int err = zap_lookup_norm(os, zapobj, name, 0,
1121 	    0, NULL, 0, NULL, 0, NULL);
1122 	if (err == EOVERFLOW || err == EINVAL)
1123 		err = 0; /* found, but skipped reading the value */
1124 	return (err);
1125 }
1126 
1127 int
1128 zap_length(objset_t *os, uint64_t zapobj, const char *name,
1129     uint64_t *integer_size, uint64_t *num_integers)
1130 {
1131 	zap_t *zap;
1132 
1133 	int err =
1134 	    zap_lockdir(os, zapobj, NULL, RW_READER, TRUE, FALSE, FTAG, &zap);
1135 	if (err != 0)
1136 		return (err);
1137 	zap_name_t *zn = zap_name_alloc(zap, name, 0);
1138 	if (zn == NULL) {
1139 		zap_unlockdir(zap, FTAG);
1140 		return (SET_ERROR(ENOTSUP));
1141 	}
1142 	if (!zap->zap_ismicro) {
1143 		err = fzap_length(zn, integer_size, num_integers);
1144 	} else {
1145 		mzap_ent_t *mze = mze_find(zn);
1146 		if (mze == NULL) {
1147 			err = SET_ERROR(ENOENT);
1148 		} else {
1149 			if (integer_size)
1150 				*integer_size = 8;
1151 			if (num_integers)
1152 				*num_integers = 1;
1153 		}
1154 	}
1155 	zap_name_free(zn);
1156 	zap_unlockdir(zap, FTAG);
1157 	return (err);
1158 }
1159 
1160 int
1161 zap_length_uint64(objset_t *os, uint64_t zapobj, const uint64_t *key,
1162     int key_numints, uint64_t *integer_size, uint64_t *num_integers)
1163 {
1164 	zap_t *zap;
1165 
1166 	int err =
1167 	    zap_lockdir(os, zapobj, NULL, RW_READER, TRUE, FALSE, FTAG, &zap);
1168 	if (err != 0)
1169 		return (err);
1170 	zap_name_t *zn = zap_name_alloc_uint64(zap, key, key_numints);
1171 	if (zn == NULL) {
1172 		zap_unlockdir(zap, FTAG);
1173 		return (SET_ERROR(ENOTSUP));
1174 	}
1175 	err = fzap_length(zn, integer_size, num_integers);
1176 	zap_name_free(zn);
1177 	zap_unlockdir(zap, FTAG);
1178 	return (err);
1179 }
1180 
1181 static void
1182 mzap_addent(zap_name_t *zn, uint64_t value)
1183 {
1184 	zap_t *zap = zn->zn_zap;
1185 	int start = zap->zap_m.zap_alloc_next;
1186 
1187 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
1188 
1189 #ifdef ZFS_DEBUG
1190 	for (int i = 0; i < zap->zap_m.zap_num_chunks; i++) {
1191 		mzap_ent_phys_t *mze = &zap_m_phys(zap)->mz_chunk[i];
1192 		ASSERT(strcmp(zn->zn_key_orig, mze->mze_name) != 0);
1193 	}
1194 #endif
1195 
1196 	uint32_t cd = mze_find_unused_cd(zap, zn->zn_hash);
1197 	/* given the limited size of the microzap, this can't happen */
1198 	ASSERT(cd < zap_maxcd(zap));
1199 
1200 again:
1201 	for (int i = start; i < zap->zap_m.zap_num_chunks; i++) {
1202 		mzap_ent_phys_t *mze = &zap_m_phys(zap)->mz_chunk[i];
1203 		if (mze->mze_name[0] == 0) {
1204 			mze->mze_value = value;
1205 			mze->mze_cd = cd;
1206 			(void) strlcpy(mze->mze_name, zn->zn_key_orig,
1207 			    sizeof (mze->mze_name));
1208 			zap->zap_m.zap_num_entries++;
1209 			zap->zap_m.zap_alloc_next = i+1;
1210 			if (zap->zap_m.zap_alloc_next ==
1211 			    zap->zap_m.zap_num_chunks)
1212 				zap->zap_m.zap_alloc_next = 0;
1213 			mze_insert(zap, i, zn->zn_hash);
1214 			return;
1215 		}
1216 	}
1217 	if (start != 0) {
1218 		start = 0;
1219 		goto again;
1220 	}
1221 	cmn_err(CE_PANIC, "out of entries!");
1222 }
1223 
1224 static int
1225 zap_add_impl(zap_t *zap, const char *key,
1226     int integer_size, uint64_t num_integers,
1227     const void *val, dmu_tx_t *tx, const void *tag)
1228 {
1229 	const uint64_t *intval = val;
1230 	int err = 0;
1231 
1232 	zap_name_t *zn = zap_name_alloc(zap, key, 0);
1233 	if (zn == NULL) {
1234 		zap_unlockdir(zap, tag);
1235 		return (SET_ERROR(ENOTSUP));
1236 	}
1237 	if (!zap->zap_ismicro) {
1238 		err = fzap_add(zn, integer_size, num_integers, val, tag, tx);
1239 		zap = zn->zn_zap;	/* fzap_add() may change zap */
1240 	} else if (integer_size != 8 || num_integers != 1 ||
1241 	    strlen(key) >= MZAP_NAME_LEN ||
1242 	    !mze_canfit_fzap_leaf(zn, zn->zn_hash)) {
1243 		err = mzap_upgrade(&zn->zn_zap, tag, tx, 0);
1244 		if (err == 0) {
1245 			err = fzap_add(zn, integer_size, num_integers, val,
1246 			    tag, tx);
1247 		}
1248 		zap = zn->zn_zap;	/* fzap_add() may change zap */
1249 	} else {
1250 		if (mze_find(zn) != NULL) {
1251 			err = SET_ERROR(EEXIST);
1252 		} else {
1253 			mzap_addent(zn, *intval);
1254 		}
1255 	}
1256 	ASSERT(zap == zn->zn_zap);
1257 	zap_name_free(zn);
1258 	if (zap != NULL)	/* may be NULL if fzap_add() failed */
1259 		zap_unlockdir(zap, tag);
1260 	return (err);
1261 }
1262 
1263 int
1264 zap_add(objset_t *os, uint64_t zapobj, const char *key,
1265     int integer_size, uint64_t num_integers,
1266     const void *val, dmu_tx_t *tx)
1267 {
1268 	zap_t *zap;
1269 	int err;
1270 
1271 	err = zap_lockdir(os, zapobj, tx, RW_WRITER, TRUE, TRUE, FTAG, &zap);
1272 	if (err != 0)
1273 		return (err);
1274 	err = zap_add_impl(zap, key, integer_size, num_integers, val, tx, FTAG);
1275 	/* zap_add_impl() calls zap_unlockdir() */
1276 	return (err);
1277 }
1278 
1279 int
1280 zap_add_by_dnode(dnode_t *dn, const char *key,
1281     int integer_size, uint64_t num_integers,
1282     const void *val, dmu_tx_t *tx)
1283 {
1284 	zap_t *zap;
1285 	int err;
1286 
1287 	err = zap_lockdir_by_dnode(dn, tx, RW_WRITER, TRUE, TRUE, FTAG, &zap);
1288 	if (err != 0)
1289 		return (err);
1290 	err = zap_add_impl(zap, key, integer_size, num_integers, val, tx, FTAG);
1291 	/* zap_add_impl() calls zap_unlockdir() */
1292 	return (err);
1293 }
1294 
1295 int
1296 zap_add_uint64(objset_t *os, uint64_t zapobj, const uint64_t *key,
1297     int key_numints, int integer_size, uint64_t num_integers,
1298     const void *val, dmu_tx_t *tx)
1299 {
1300 	zap_t *zap;
1301 
1302 	int err =
1303 	    zap_lockdir(os, zapobj, tx, RW_WRITER, TRUE, TRUE, FTAG, &zap);
1304 	if (err != 0)
1305 		return (err);
1306 	zap_name_t *zn = zap_name_alloc_uint64(zap, key, key_numints);
1307 	if (zn == NULL) {
1308 		zap_unlockdir(zap, FTAG);
1309 		return (SET_ERROR(ENOTSUP));
1310 	}
1311 	err = fzap_add(zn, integer_size, num_integers, val, FTAG, tx);
1312 	zap = zn->zn_zap;	/* fzap_add() may change zap */
1313 	zap_name_free(zn);
1314 	if (zap != NULL)	/* may be NULL if fzap_add() failed */
1315 		zap_unlockdir(zap, FTAG);
1316 	return (err);
1317 }
1318 
1319 int
1320 zap_update(objset_t *os, uint64_t zapobj, const char *name,
1321     int integer_size, uint64_t num_integers, const void *val, dmu_tx_t *tx)
1322 {
1323 	zap_t *zap;
1324 	const uint64_t *intval = val;
1325 
1326 	int err =
1327 	    zap_lockdir(os, zapobj, tx, RW_WRITER, TRUE, TRUE, FTAG, &zap);
1328 	if (err != 0)
1329 		return (err);
1330 	zap_name_t *zn = zap_name_alloc(zap, name, 0);
1331 	if (zn == NULL) {
1332 		zap_unlockdir(zap, FTAG);
1333 		return (SET_ERROR(ENOTSUP));
1334 	}
1335 	if (!zap->zap_ismicro) {
1336 		err = fzap_update(zn, integer_size, num_integers, val,
1337 		    FTAG, tx);
1338 		zap = zn->zn_zap;	/* fzap_update() may change zap */
1339 	} else if (integer_size != 8 || num_integers != 1 ||
1340 	    strlen(name) >= MZAP_NAME_LEN) {
1341 		dprintf("upgrading obj %llu: intsz=%u numint=%llu name=%s\n",
1342 		    (u_longlong_t)zapobj, integer_size,
1343 		    (u_longlong_t)num_integers, name);
1344 		err = mzap_upgrade(&zn->zn_zap, FTAG, tx, 0);
1345 		if (err == 0) {
1346 			err = fzap_update(zn, integer_size, num_integers,
1347 			    val, FTAG, tx);
1348 		}
1349 		zap = zn->zn_zap;	/* fzap_update() may change zap */
1350 	} else {
1351 		mzap_ent_t *mze = mze_find(zn);
1352 		if (mze != NULL) {
1353 			MZE_PHYS(zap, mze)->mze_value = *intval;
1354 		} else {
1355 			mzap_addent(zn, *intval);
1356 		}
1357 	}
1358 	ASSERT(zap == zn->zn_zap);
1359 	zap_name_free(zn);
1360 	if (zap != NULL)	/* may be NULL if fzap_upgrade() failed */
1361 		zap_unlockdir(zap, FTAG);
1362 	return (err);
1363 }
1364 
1365 int
1366 zap_update_uint64(objset_t *os, uint64_t zapobj, const uint64_t *key,
1367     int key_numints,
1368     int integer_size, uint64_t num_integers, const void *val, dmu_tx_t *tx)
1369 {
1370 	zap_t *zap;
1371 
1372 	int err =
1373 	    zap_lockdir(os, zapobj, tx, RW_WRITER, TRUE, TRUE, FTAG, &zap);
1374 	if (err != 0)
1375 		return (err);
1376 	zap_name_t *zn = zap_name_alloc_uint64(zap, key, key_numints);
1377 	if (zn == NULL) {
1378 		zap_unlockdir(zap, FTAG);
1379 		return (SET_ERROR(ENOTSUP));
1380 	}
1381 	err = fzap_update(zn, integer_size, num_integers, val, FTAG, tx);
1382 	zap = zn->zn_zap;	/* fzap_update() may change zap */
1383 	zap_name_free(zn);
1384 	if (zap != NULL)	/* may be NULL if fzap_upgrade() failed */
1385 		zap_unlockdir(zap, FTAG);
1386 	return (err);
1387 }
1388 
1389 int
1390 zap_remove(objset_t *os, uint64_t zapobj, const char *name, dmu_tx_t *tx)
1391 {
1392 	return (zap_remove_norm(os, zapobj, name, 0, tx));
1393 }
1394 
1395 static int
1396 zap_remove_impl(zap_t *zap, const char *name,
1397     matchtype_t mt, dmu_tx_t *tx)
1398 {
1399 	int err = 0;
1400 
1401 	zap_name_t *zn = zap_name_alloc(zap, name, mt);
1402 	if (zn == NULL)
1403 		return (SET_ERROR(ENOTSUP));
1404 	if (!zap->zap_ismicro) {
1405 		err = fzap_remove(zn, tx);
1406 	} else {
1407 		mzap_ent_t *mze = mze_find(zn);
1408 		if (mze == NULL) {
1409 			err = SET_ERROR(ENOENT);
1410 		} else {
1411 			zap->zap_m.zap_num_entries--;
1412 			memset(&zap_m_phys(zap)->mz_chunk[mze->mze_chunkid], 0,
1413 			    sizeof (mzap_ent_phys_t));
1414 			mze_remove(zap, mze);
1415 		}
1416 	}
1417 	zap_name_free(zn);
1418 	return (err);
1419 }
1420 
1421 int
1422 zap_remove_norm(objset_t *os, uint64_t zapobj, const char *name,
1423     matchtype_t mt, dmu_tx_t *tx)
1424 {
1425 	zap_t *zap;
1426 	int err;
1427 
1428 	err = zap_lockdir(os, zapobj, tx, RW_WRITER, TRUE, FALSE, FTAG, &zap);
1429 	if (err)
1430 		return (err);
1431 	err = zap_remove_impl(zap, name, mt, tx);
1432 	zap_unlockdir(zap, FTAG);
1433 	return (err);
1434 }
1435 
1436 int
1437 zap_remove_by_dnode(dnode_t *dn, const char *name, dmu_tx_t *tx)
1438 {
1439 	zap_t *zap;
1440 	int err;
1441 
1442 	err = zap_lockdir_by_dnode(dn, tx, RW_WRITER, TRUE, FALSE, FTAG, &zap);
1443 	if (err)
1444 		return (err);
1445 	err = zap_remove_impl(zap, name, 0, tx);
1446 	zap_unlockdir(zap, FTAG);
1447 	return (err);
1448 }
1449 
1450 int
1451 zap_remove_uint64(objset_t *os, uint64_t zapobj, const uint64_t *key,
1452     int key_numints, dmu_tx_t *tx)
1453 {
1454 	zap_t *zap;
1455 
1456 	int err =
1457 	    zap_lockdir(os, zapobj, tx, RW_WRITER, TRUE, FALSE, FTAG, &zap);
1458 	if (err != 0)
1459 		return (err);
1460 	zap_name_t *zn = zap_name_alloc_uint64(zap, key, key_numints);
1461 	if (zn == NULL) {
1462 		zap_unlockdir(zap, FTAG);
1463 		return (SET_ERROR(ENOTSUP));
1464 	}
1465 	err = fzap_remove(zn, tx);
1466 	zap_name_free(zn);
1467 	zap_unlockdir(zap, FTAG);
1468 	return (err);
1469 }
1470 
1471 /*
1472  * Routines for iterating over the attributes.
1473  */
1474 
1475 static void
1476 zap_cursor_init_impl(zap_cursor_t *zc, objset_t *os, uint64_t zapobj,
1477     uint64_t serialized, boolean_t prefetch)
1478 {
1479 	zc->zc_objset = os;
1480 	zc->zc_zap = NULL;
1481 	zc->zc_leaf = NULL;
1482 	zc->zc_zapobj = zapobj;
1483 	zc->zc_serialized = serialized;
1484 	zc->zc_hash = 0;
1485 	zc->zc_cd = 0;
1486 	zc->zc_prefetch = prefetch;
1487 }
1488 void
1489 zap_cursor_init_serialized(zap_cursor_t *zc, objset_t *os, uint64_t zapobj,
1490     uint64_t serialized)
1491 {
1492 	zap_cursor_init_impl(zc, os, zapobj, serialized, B_TRUE);
1493 }
1494 
1495 /*
1496  * Initialize a cursor at the beginning of the ZAP object.  The entire
1497  * ZAP object will be prefetched.
1498  */
1499 void
1500 zap_cursor_init(zap_cursor_t *zc, objset_t *os, uint64_t zapobj)
1501 {
1502 	zap_cursor_init_impl(zc, os, zapobj, 0, B_TRUE);
1503 }
1504 
1505 /*
1506  * Initialize a cursor at the beginning, but request that we not prefetch
1507  * the entire ZAP object.
1508  */
1509 void
1510 zap_cursor_init_noprefetch(zap_cursor_t *zc, objset_t *os, uint64_t zapobj)
1511 {
1512 	zap_cursor_init_impl(zc, os, zapobj, 0, B_FALSE);
1513 }
1514 
1515 void
1516 zap_cursor_fini(zap_cursor_t *zc)
1517 {
1518 	if (zc->zc_zap) {
1519 		rw_enter(&zc->zc_zap->zap_rwlock, RW_READER);
1520 		zap_unlockdir(zc->zc_zap, NULL);
1521 		zc->zc_zap = NULL;
1522 	}
1523 	if (zc->zc_leaf) {
1524 		rw_enter(&zc->zc_leaf->l_rwlock, RW_READER);
1525 		zap_put_leaf(zc->zc_leaf);
1526 		zc->zc_leaf = NULL;
1527 	}
1528 	zc->zc_objset = NULL;
1529 }
1530 
1531 uint64_t
1532 zap_cursor_serialize(zap_cursor_t *zc)
1533 {
1534 	if (zc->zc_hash == -1ULL)
1535 		return (-1ULL);
1536 	if (zc->zc_zap == NULL)
1537 		return (zc->zc_serialized);
1538 	ASSERT((zc->zc_hash & zap_maxcd(zc->zc_zap)) == 0);
1539 	ASSERT(zc->zc_cd < zap_maxcd(zc->zc_zap));
1540 
1541 	/*
1542 	 * We want to keep the high 32 bits of the cursor zero if we can, so
1543 	 * that 32-bit programs can access this.  So usually use a small
1544 	 * (28-bit) hash value so we can fit 4 bits of cd into the low 32-bits
1545 	 * of the cursor.
1546 	 *
1547 	 * [ collision differentiator | zap_hashbits()-bit hash value ]
1548 	 */
1549 	return ((zc->zc_hash >> (64 - zap_hashbits(zc->zc_zap))) |
1550 	    ((uint64_t)zc->zc_cd << zap_hashbits(zc->zc_zap)));
1551 }
1552 
1553 int
1554 zap_cursor_retrieve(zap_cursor_t *zc, zap_attribute_t *za)
1555 {
1556 	int err;
1557 
1558 	if (zc->zc_hash == -1ULL)
1559 		return (SET_ERROR(ENOENT));
1560 
1561 	if (zc->zc_zap == NULL) {
1562 		int hb;
1563 		err = zap_lockdir(zc->zc_objset, zc->zc_zapobj, NULL,
1564 		    RW_READER, TRUE, FALSE, NULL, &zc->zc_zap);
1565 		if (err != 0)
1566 			return (err);
1567 
1568 		/*
1569 		 * To support zap_cursor_init_serialized, advance, retrieve,
1570 		 * we must add to the existing zc_cd, which may already
1571 		 * be 1 due to the zap_cursor_advance.
1572 		 */
1573 		ASSERT(zc->zc_hash == 0);
1574 		hb = zap_hashbits(zc->zc_zap);
1575 		zc->zc_hash = zc->zc_serialized << (64 - hb);
1576 		zc->zc_cd += zc->zc_serialized >> hb;
1577 		if (zc->zc_cd >= zap_maxcd(zc->zc_zap)) /* corrupt serialized */
1578 			zc->zc_cd = 0;
1579 	} else {
1580 		rw_enter(&zc->zc_zap->zap_rwlock, RW_READER);
1581 	}
1582 	if (!zc->zc_zap->zap_ismicro) {
1583 		err = fzap_cursor_retrieve(zc->zc_zap, zc, za);
1584 	} else {
1585 		avl_index_t idx;
1586 		mzap_ent_t mze_tofind;
1587 
1588 		mze_tofind.mze_hash = zc->zc_hash;
1589 		mze_tofind.mze_cd = zc->zc_cd;
1590 
1591 		mzap_ent_t *mze =
1592 		    avl_find(&zc->zc_zap->zap_m.zap_avl, &mze_tofind, &idx);
1593 		if (mze == NULL) {
1594 			mze = avl_nearest(&zc->zc_zap->zap_m.zap_avl,
1595 			    idx, AVL_AFTER);
1596 		}
1597 		if (mze) {
1598 			mzap_ent_phys_t *mzep = MZE_PHYS(zc->zc_zap, mze);
1599 			ASSERT3U(mze->mze_cd, ==, mzep->mze_cd);
1600 			za->za_normalization_conflict =
1601 			    mzap_normalization_conflict(zc->zc_zap, NULL, mze);
1602 			za->za_integer_length = 8;
1603 			za->za_num_integers = 1;
1604 			za->za_first_integer = mzep->mze_value;
1605 			(void) strlcpy(za->za_name, mzep->mze_name,
1606 			    sizeof (za->za_name));
1607 			zc->zc_hash = mze->mze_hash;
1608 			zc->zc_cd = mze->mze_cd;
1609 			err = 0;
1610 		} else {
1611 			zc->zc_hash = -1ULL;
1612 			err = SET_ERROR(ENOENT);
1613 		}
1614 	}
1615 	rw_exit(&zc->zc_zap->zap_rwlock);
1616 	return (err);
1617 }
1618 
1619 void
1620 zap_cursor_advance(zap_cursor_t *zc)
1621 {
1622 	if (zc->zc_hash == -1ULL)
1623 		return;
1624 	zc->zc_cd++;
1625 }
1626 
1627 int
1628 zap_get_stats(objset_t *os, uint64_t zapobj, zap_stats_t *zs)
1629 {
1630 	zap_t *zap;
1631 
1632 	int err =
1633 	    zap_lockdir(os, zapobj, NULL, RW_READER, TRUE, FALSE, FTAG, &zap);
1634 	if (err != 0)
1635 		return (err);
1636 
1637 	memset(zs, 0, sizeof (zap_stats_t));
1638 
1639 	if (zap->zap_ismicro) {
1640 		zs->zs_blocksize = zap->zap_dbuf->db_size;
1641 		zs->zs_num_entries = zap->zap_m.zap_num_entries;
1642 		zs->zs_num_blocks = 1;
1643 	} else {
1644 		fzap_get_stats(zap, zs);
1645 	}
1646 	zap_unlockdir(zap, FTAG);
1647 	return (0);
1648 }
1649 
1650 #if defined(_KERNEL)
1651 EXPORT_SYMBOL(zap_create);
1652 EXPORT_SYMBOL(zap_create_dnsize);
1653 EXPORT_SYMBOL(zap_create_norm);
1654 EXPORT_SYMBOL(zap_create_norm_dnsize);
1655 EXPORT_SYMBOL(zap_create_flags);
1656 EXPORT_SYMBOL(zap_create_flags_dnsize);
1657 EXPORT_SYMBOL(zap_create_claim);
1658 EXPORT_SYMBOL(zap_create_claim_norm);
1659 EXPORT_SYMBOL(zap_create_claim_norm_dnsize);
1660 EXPORT_SYMBOL(zap_create_hold);
1661 EXPORT_SYMBOL(zap_destroy);
1662 EXPORT_SYMBOL(zap_lookup);
1663 EXPORT_SYMBOL(zap_lookup_by_dnode);
1664 EXPORT_SYMBOL(zap_lookup_norm);
1665 EXPORT_SYMBOL(zap_lookup_uint64);
1666 EXPORT_SYMBOL(zap_contains);
1667 EXPORT_SYMBOL(zap_prefetch);
1668 EXPORT_SYMBOL(zap_prefetch_uint64);
1669 EXPORT_SYMBOL(zap_add);
1670 EXPORT_SYMBOL(zap_add_by_dnode);
1671 EXPORT_SYMBOL(zap_add_uint64);
1672 EXPORT_SYMBOL(zap_update);
1673 EXPORT_SYMBOL(zap_update_uint64);
1674 EXPORT_SYMBOL(zap_length);
1675 EXPORT_SYMBOL(zap_length_uint64);
1676 EXPORT_SYMBOL(zap_remove);
1677 EXPORT_SYMBOL(zap_remove_by_dnode);
1678 EXPORT_SYMBOL(zap_remove_norm);
1679 EXPORT_SYMBOL(zap_remove_uint64);
1680 EXPORT_SYMBOL(zap_count);
1681 EXPORT_SYMBOL(zap_value_search);
1682 EXPORT_SYMBOL(zap_join);
1683 EXPORT_SYMBOL(zap_join_increment);
1684 EXPORT_SYMBOL(zap_add_int);
1685 EXPORT_SYMBOL(zap_remove_int);
1686 EXPORT_SYMBOL(zap_lookup_int);
1687 EXPORT_SYMBOL(zap_increment_int);
1688 EXPORT_SYMBOL(zap_add_int_key);
1689 EXPORT_SYMBOL(zap_lookup_int_key);
1690 EXPORT_SYMBOL(zap_increment);
1691 EXPORT_SYMBOL(zap_cursor_init);
1692 EXPORT_SYMBOL(zap_cursor_fini);
1693 EXPORT_SYMBOL(zap_cursor_retrieve);
1694 EXPORT_SYMBOL(zap_cursor_advance);
1695 EXPORT_SYMBOL(zap_cursor_serialize);
1696 EXPORT_SYMBOL(zap_cursor_init_serialized);
1697 EXPORT_SYMBOL(zap_get_stats);
1698 #endif
1699