xref: /illumos-gate/usr/src/uts/common/fs/zfs/zap.c (revision 4f364e7c)
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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2013 by Delphix. All rights reserved.
24  */
25 
26 /*
27  * This file contains the top half of the zfs directory structure
28  * implementation. The bottom half is in zap_leaf.c.
29  *
30  * The zdir is an extendable hash data structure. There is a table of
31  * pointers to buckets (zap_t->zd_data->zd_leafs). The buckets are
32  * each a constant size and hold a variable number of directory entries.
33  * The buckets (aka "leaf nodes") are implemented in zap_leaf.c.
34  *
35  * The pointer table holds a power of 2 number of pointers.
36  * (1<<zap_t->zd_data->zd_phys->zd_prefix_len).  The bucket pointed to
37  * by the pointer at index i in the table holds entries whose hash value
38  * has a zd_prefix_len - bit prefix
39  */
40 
41 #include <sys/spa.h>
42 #include <sys/dmu.h>
43 #include <sys/zfs_context.h>
44 #include <sys/zfs_znode.h>
45 #include <sys/fs/zfs.h>
46 #include <sys/zap.h>
47 #include <sys/refcount.h>
48 #include <sys/zap_impl.h>
49 #include <sys/zap_leaf.h>
50 
51 int fzap_default_block_shift = 14; /* 16k blocksize */
52 
53 static void zap_leaf_pageout(dmu_buf_t *db, void *vl);
54 static uint64_t zap_allocate_blocks(zap_t *zap, int nblocks);
55 
56 
57 void
58 fzap_byteswap(void *vbuf, size_t size)
59 {
60 	uint64_t block_type;
61 
62 	block_type = *(uint64_t *)vbuf;
63 
64 	if (block_type == ZBT_LEAF || block_type == BSWAP_64(ZBT_LEAF))
65 		zap_leaf_byteswap(vbuf, size);
66 	else {
67 		/* it's a ptrtbl block */
68 		byteswap_uint64_array(vbuf, size);
69 	}
70 }
71 
72 void
73 fzap_upgrade(zap_t *zap, dmu_tx_t *tx, zap_flags_t flags)
74 {
75 	dmu_buf_t *db;
76 	zap_leaf_t *l;
77 	int i;
78 	zap_phys_t *zp;
79 
80 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
81 	zap->zap_ismicro = FALSE;
82 
83 	(void) dmu_buf_update_user(zap->zap_dbuf, zap, zap,
84 	    &zap->zap_f.zap_phys, zap_evict);
85 
86 	mutex_init(&zap->zap_f.zap_num_entries_mtx, 0, 0, 0);
87 	zap->zap_f.zap_block_shift = highbit(zap->zap_dbuf->db_size) - 1;
88 
89 	zp = zap->zap_f.zap_phys;
90 	/*
91 	 * explicitly zero it since it might be coming from an
92 	 * initialized microzap
93 	 */
94 	bzero(zap->zap_dbuf->db_data, zap->zap_dbuf->db_size);
95 	zp->zap_block_type = ZBT_HEADER;
96 	zp->zap_magic = ZAP_MAGIC;
97 
98 	zp->zap_ptrtbl.zt_shift = ZAP_EMBEDDED_PTRTBL_SHIFT(zap);
99 
100 	zp->zap_freeblk = 2;		/* block 1 will be the first leaf */
101 	zp->zap_num_leafs = 1;
102 	zp->zap_num_entries = 0;
103 	zp->zap_salt = zap->zap_salt;
104 	zp->zap_normflags = zap->zap_normflags;
105 	zp->zap_flags = flags;
106 
107 	/* block 1 will be the first leaf */
108 	for (i = 0; i < (1<<zp->zap_ptrtbl.zt_shift); i++)
109 		ZAP_EMBEDDED_PTRTBL_ENT(zap, i) = 1;
110 
111 	/*
112 	 * set up block 1 - the first leaf
113 	 */
114 	VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object,
115 	    1<<FZAP_BLOCK_SHIFT(zap), FTAG, &db, DMU_READ_NO_PREFETCH));
116 	dmu_buf_will_dirty(db, tx);
117 
118 	l = kmem_zalloc(sizeof (zap_leaf_t), KM_SLEEP);
119 	l->l_dbuf = db;
120 	l->l_phys = db->db_data;
121 
122 	zap_leaf_init(l, zp->zap_normflags != 0);
123 
124 	kmem_free(l, sizeof (zap_leaf_t));
125 	dmu_buf_rele(db, FTAG);
126 }
127 
128 static int
129 zap_tryupgradedir(zap_t *zap, dmu_tx_t *tx)
130 {
131 	if (RW_WRITE_HELD(&zap->zap_rwlock))
132 		return (1);
133 	if (rw_tryupgrade(&zap->zap_rwlock)) {
134 		dmu_buf_will_dirty(zap->zap_dbuf, tx);
135 		return (1);
136 	}
137 	return (0);
138 }
139 
140 /*
141  * Generic routines for dealing with the pointer & cookie tables.
142  */
143 
144 static int
145 zap_table_grow(zap_t *zap, zap_table_phys_t *tbl,
146     void (*transfer_func)(const uint64_t *src, uint64_t *dst, int n),
147     dmu_tx_t *tx)
148 {
149 	uint64_t b, newblk;
150 	dmu_buf_t *db_old, *db_new;
151 	int err;
152 	int bs = FZAP_BLOCK_SHIFT(zap);
153 	int hepb = 1<<(bs-4);
154 	/* hepb = half the number of entries in a block */
155 
156 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
157 	ASSERT(tbl->zt_blk != 0);
158 	ASSERT(tbl->zt_numblks > 0);
159 
160 	if (tbl->zt_nextblk != 0) {
161 		newblk = tbl->zt_nextblk;
162 	} else {
163 		newblk = zap_allocate_blocks(zap, tbl->zt_numblks * 2);
164 		tbl->zt_nextblk = newblk;
165 		ASSERT0(tbl->zt_blks_copied);
166 		dmu_prefetch(zap->zap_objset, zap->zap_object,
167 		    tbl->zt_blk << bs, tbl->zt_numblks << bs);
168 	}
169 
170 	/*
171 	 * Copy the ptrtbl from the old to new location.
172 	 */
173 
174 	b = tbl->zt_blks_copied;
175 	err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
176 	    (tbl->zt_blk + b) << bs, FTAG, &db_old, DMU_READ_NO_PREFETCH);
177 	if (err)
178 		return (err);
179 
180 	/* first half of entries in old[b] go to new[2*b+0] */
181 	VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object,
182 	    (newblk + 2*b+0) << bs, FTAG, &db_new, DMU_READ_NO_PREFETCH));
183 	dmu_buf_will_dirty(db_new, tx);
184 	transfer_func(db_old->db_data, db_new->db_data, hepb);
185 	dmu_buf_rele(db_new, FTAG);
186 
187 	/* second half of entries in old[b] go to new[2*b+1] */
188 	VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object,
189 	    (newblk + 2*b+1) << bs, FTAG, &db_new, DMU_READ_NO_PREFETCH));
190 	dmu_buf_will_dirty(db_new, tx);
191 	transfer_func((uint64_t *)db_old->db_data + hepb,
192 	    db_new->db_data, hepb);
193 	dmu_buf_rele(db_new, FTAG);
194 
195 	dmu_buf_rele(db_old, FTAG);
196 
197 	tbl->zt_blks_copied++;
198 
199 	dprintf("copied block %llu of %llu\n",
200 	    tbl->zt_blks_copied, tbl->zt_numblks);
201 
202 	if (tbl->zt_blks_copied == tbl->zt_numblks) {
203 		(void) dmu_free_range(zap->zap_objset, zap->zap_object,
204 		    tbl->zt_blk << bs, tbl->zt_numblks << bs, tx);
205 
206 		tbl->zt_blk = newblk;
207 		tbl->zt_numblks *= 2;
208 		tbl->zt_shift++;
209 		tbl->zt_nextblk = 0;
210 		tbl->zt_blks_copied = 0;
211 
212 		dprintf("finished; numblocks now %llu (%lluk entries)\n",
213 		    tbl->zt_numblks, 1<<(tbl->zt_shift-10));
214 	}
215 
216 	return (0);
217 }
218 
219 static int
220 zap_table_store(zap_t *zap, zap_table_phys_t *tbl, uint64_t idx, uint64_t val,
221     dmu_tx_t *tx)
222 {
223 	int err;
224 	uint64_t blk, off;
225 	int bs = FZAP_BLOCK_SHIFT(zap);
226 	dmu_buf_t *db;
227 
228 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
229 	ASSERT(tbl->zt_blk != 0);
230 
231 	dprintf("storing %llx at index %llx\n", val, idx);
232 
233 	blk = idx >> (bs-3);
234 	off = idx & ((1<<(bs-3))-1);
235 
236 	err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
237 	    (tbl->zt_blk + blk) << bs, FTAG, &db, DMU_READ_NO_PREFETCH);
238 	if (err)
239 		return (err);
240 	dmu_buf_will_dirty(db, tx);
241 
242 	if (tbl->zt_nextblk != 0) {
243 		uint64_t idx2 = idx * 2;
244 		uint64_t blk2 = idx2 >> (bs-3);
245 		uint64_t off2 = idx2 & ((1<<(bs-3))-1);
246 		dmu_buf_t *db2;
247 
248 		err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
249 		    (tbl->zt_nextblk + blk2) << bs, FTAG, &db2,
250 		    DMU_READ_NO_PREFETCH);
251 		if (err) {
252 			dmu_buf_rele(db, FTAG);
253 			return (err);
254 		}
255 		dmu_buf_will_dirty(db2, tx);
256 		((uint64_t *)db2->db_data)[off2] = val;
257 		((uint64_t *)db2->db_data)[off2+1] = val;
258 		dmu_buf_rele(db2, FTAG);
259 	}
260 
261 	((uint64_t *)db->db_data)[off] = val;
262 	dmu_buf_rele(db, FTAG);
263 
264 	return (0);
265 }
266 
267 static int
268 zap_table_load(zap_t *zap, zap_table_phys_t *tbl, uint64_t idx, uint64_t *valp)
269 {
270 	uint64_t blk, off;
271 	int err;
272 	dmu_buf_t *db;
273 	int bs = FZAP_BLOCK_SHIFT(zap);
274 
275 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
276 
277 	blk = idx >> (bs-3);
278 	off = idx & ((1<<(bs-3))-1);
279 
280 	err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
281 	    (tbl->zt_blk + blk) << bs, FTAG, &db, DMU_READ_NO_PREFETCH);
282 	if (err)
283 		return (err);
284 	*valp = ((uint64_t *)db->db_data)[off];
285 	dmu_buf_rele(db, FTAG);
286 
287 	if (tbl->zt_nextblk != 0) {
288 		/*
289 		 * read the nextblk for the sake of i/o error checking,
290 		 * so that zap_table_load() will catch errors for
291 		 * zap_table_store.
292 		 */
293 		blk = (idx*2) >> (bs-3);
294 
295 		err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
296 		    (tbl->zt_nextblk + blk) << bs, FTAG, &db,
297 		    DMU_READ_NO_PREFETCH);
298 		if (err == 0)
299 			dmu_buf_rele(db, FTAG);
300 	}
301 	return (err);
302 }
303 
304 /*
305  * Routines for growing the ptrtbl.
306  */
307 
308 static void
309 zap_ptrtbl_transfer(const uint64_t *src, uint64_t *dst, int n)
310 {
311 	int i;
312 	for (i = 0; i < n; i++) {
313 		uint64_t lb = src[i];
314 		dst[2*i+0] = lb;
315 		dst[2*i+1] = lb;
316 	}
317 }
318 
319 static int
320 zap_grow_ptrtbl(zap_t *zap, dmu_tx_t *tx)
321 {
322 	/*
323 	 * The pointer table should never use more hash bits than we
324 	 * have (otherwise we'd be using useless zero bits to index it).
325 	 * If we are within 2 bits of running out, stop growing, since
326 	 * this is already an aberrant condition.
327 	 */
328 	if (zap->zap_f.zap_phys->zap_ptrtbl.zt_shift >= zap_hashbits(zap) - 2)
329 		return (SET_ERROR(ENOSPC));
330 
331 	if (zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks == 0) {
332 		/*
333 		 * We are outgrowing the "embedded" ptrtbl (the one
334 		 * stored in the header block).  Give it its own entire
335 		 * block, which will double the size of the ptrtbl.
336 		 */
337 		uint64_t newblk;
338 		dmu_buf_t *db_new;
339 		int err;
340 
341 		ASSERT3U(zap->zap_f.zap_phys->zap_ptrtbl.zt_shift, ==,
342 		    ZAP_EMBEDDED_PTRTBL_SHIFT(zap));
343 		ASSERT0(zap->zap_f.zap_phys->zap_ptrtbl.zt_blk);
344 
345 		newblk = zap_allocate_blocks(zap, 1);
346 		err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
347 		    newblk << FZAP_BLOCK_SHIFT(zap), FTAG, &db_new,
348 		    DMU_READ_NO_PREFETCH);
349 		if (err)
350 			return (err);
351 		dmu_buf_will_dirty(db_new, tx);
352 		zap_ptrtbl_transfer(&ZAP_EMBEDDED_PTRTBL_ENT(zap, 0),
353 		    db_new->db_data, 1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap));
354 		dmu_buf_rele(db_new, FTAG);
355 
356 		zap->zap_f.zap_phys->zap_ptrtbl.zt_blk = newblk;
357 		zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks = 1;
358 		zap->zap_f.zap_phys->zap_ptrtbl.zt_shift++;
359 
360 		ASSERT3U(1ULL << zap->zap_f.zap_phys->zap_ptrtbl.zt_shift, ==,
361 		    zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks <<
362 		    (FZAP_BLOCK_SHIFT(zap)-3));
363 
364 		return (0);
365 	} else {
366 		return (zap_table_grow(zap, &zap->zap_f.zap_phys->zap_ptrtbl,
367 		    zap_ptrtbl_transfer, tx));
368 	}
369 }
370 
371 static void
372 zap_increment_num_entries(zap_t *zap, int delta, dmu_tx_t *tx)
373 {
374 	dmu_buf_will_dirty(zap->zap_dbuf, tx);
375 	mutex_enter(&zap->zap_f.zap_num_entries_mtx);
376 	ASSERT(delta > 0 || zap->zap_f.zap_phys->zap_num_entries >= -delta);
377 	zap->zap_f.zap_phys->zap_num_entries += delta;
378 	mutex_exit(&zap->zap_f.zap_num_entries_mtx);
379 }
380 
381 static uint64_t
382 zap_allocate_blocks(zap_t *zap, int nblocks)
383 {
384 	uint64_t newblk;
385 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
386 	newblk = zap->zap_f.zap_phys->zap_freeblk;
387 	zap->zap_f.zap_phys->zap_freeblk += nblocks;
388 	return (newblk);
389 }
390 
391 static zap_leaf_t *
392 zap_create_leaf(zap_t *zap, dmu_tx_t *tx)
393 {
394 	void *winner;
395 	zap_leaf_t *l = kmem_alloc(sizeof (zap_leaf_t), KM_SLEEP);
396 
397 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
398 
399 	rw_init(&l->l_rwlock, 0, 0, 0);
400 	rw_enter(&l->l_rwlock, RW_WRITER);
401 	l->l_blkid = zap_allocate_blocks(zap, 1);
402 	l->l_dbuf = NULL;
403 	l->l_phys = NULL;
404 
405 	VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object,
406 	    l->l_blkid << FZAP_BLOCK_SHIFT(zap), NULL, &l->l_dbuf,
407 	    DMU_READ_NO_PREFETCH));
408 	winner = dmu_buf_set_user(l->l_dbuf, l, &l->l_phys, zap_leaf_pageout);
409 	ASSERT(winner == NULL);
410 	dmu_buf_will_dirty(l->l_dbuf, tx);
411 
412 	zap_leaf_init(l, zap->zap_normflags != 0);
413 
414 	zap->zap_f.zap_phys->zap_num_leafs++;
415 
416 	return (l);
417 }
418 
419 int
420 fzap_count(zap_t *zap, uint64_t *count)
421 {
422 	ASSERT(!zap->zap_ismicro);
423 	mutex_enter(&zap->zap_f.zap_num_entries_mtx); /* unnecessary */
424 	*count = zap->zap_f.zap_phys->zap_num_entries;
425 	mutex_exit(&zap->zap_f.zap_num_entries_mtx);
426 	return (0);
427 }
428 
429 /*
430  * Routines for obtaining zap_leaf_t's
431  */
432 
433 void
434 zap_put_leaf(zap_leaf_t *l)
435 {
436 	rw_exit(&l->l_rwlock);
437 	dmu_buf_rele(l->l_dbuf, NULL);
438 }
439 
440 _NOTE(ARGSUSED(0))
441 static void
442 zap_leaf_pageout(dmu_buf_t *db, void *vl)
443 {
444 	zap_leaf_t *l = vl;
445 
446 	rw_destroy(&l->l_rwlock);
447 	kmem_free(l, sizeof (zap_leaf_t));
448 }
449 
450 static zap_leaf_t *
451 zap_open_leaf(uint64_t blkid, dmu_buf_t *db)
452 {
453 	zap_leaf_t *l, *winner;
454 
455 	ASSERT(blkid != 0);
456 
457 	l = kmem_alloc(sizeof (zap_leaf_t), KM_SLEEP);
458 	rw_init(&l->l_rwlock, 0, 0, 0);
459 	rw_enter(&l->l_rwlock, RW_WRITER);
460 	l->l_blkid = blkid;
461 	l->l_bs = highbit(db->db_size)-1;
462 	l->l_dbuf = db;
463 	l->l_phys = NULL;
464 
465 	winner = dmu_buf_set_user(db, l, &l->l_phys, zap_leaf_pageout);
466 
467 	rw_exit(&l->l_rwlock);
468 	if (winner != NULL) {
469 		/* someone else set it first */
470 		zap_leaf_pageout(NULL, l);
471 		l = winner;
472 	}
473 
474 	/*
475 	 * lhr_pad was previously used for the next leaf in the leaf
476 	 * chain.  There should be no chained leafs (as we have removed
477 	 * support for them).
478 	 */
479 	ASSERT0(l->l_phys->l_hdr.lh_pad1);
480 
481 	/*
482 	 * There should be more hash entries than there can be
483 	 * chunks to put in the hash table
484 	 */
485 	ASSERT3U(ZAP_LEAF_HASH_NUMENTRIES(l), >, ZAP_LEAF_NUMCHUNKS(l) / 3);
486 
487 	/* The chunks should begin at the end of the hash table */
488 	ASSERT3P(&ZAP_LEAF_CHUNK(l, 0), ==,
489 	    &l->l_phys->l_hash[ZAP_LEAF_HASH_NUMENTRIES(l)]);
490 
491 	/* The chunks should end at the end of the block */
492 	ASSERT3U((uintptr_t)&ZAP_LEAF_CHUNK(l, ZAP_LEAF_NUMCHUNKS(l)) -
493 	    (uintptr_t)l->l_phys, ==, l->l_dbuf->db_size);
494 
495 	return (l);
496 }
497 
498 static int
499 zap_get_leaf_byblk(zap_t *zap, uint64_t blkid, dmu_tx_t *tx, krw_t lt,
500     zap_leaf_t **lp)
501 {
502 	dmu_buf_t *db;
503 	zap_leaf_t *l;
504 	int bs = FZAP_BLOCK_SHIFT(zap);
505 	int err;
506 
507 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
508 
509 	err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
510 	    blkid << bs, NULL, &db, DMU_READ_NO_PREFETCH);
511 	if (err)
512 		return (err);
513 
514 	ASSERT3U(db->db_object, ==, zap->zap_object);
515 	ASSERT3U(db->db_offset, ==, blkid << bs);
516 	ASSERT3U(db->db_size, ==, 1 << bs);
517 	ASSERT(blkid != 0);
518 
519 	l = dmu_buf_get_user(db);
520 
521 	if (l == NULL)
522 		l = zap_open_leaf(blkid, db);
523 
524 	rw_enter(&l->l_rwlock, lt);
525 	/*
526 	 * Must lock before dirtying, otherwise l->l_phys could change,
527 	 * causing ASSERT below to fail.
528 	 */
529 	if (lt == RW_WRITER)
530 		dmu_buf_will_dirty(db, tx);
531 	ASSERT3U(l->l_blkid, ==, blkid);
532 	ASSERT3P(l->l_dbuf, ==, db);
533 	ASSERT3P(l->l_phys, ==, l->l_dbuf->db_data);
534 	ASSERT3U(l->l_phys->l_hdr.lh_block_type, ==, ZBT_LEAF);
535 	ASSERT3U(l->l_phys->l_hdr.lh_magic, ==, ZAP_LEAF_MAGIC);
536 
537 	*lp = l;
538 	return (0);
539 }
540 
541 static int
542 zap_idx_to_blk(zap_t *zap, uint64_t idx, uint64_t *valp)
543 {
544 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
545 
546 	if (zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks == 0) {
547 		ASSERT3U(idx, <,
548 		    (1ULL << zap->zap_f.zap_phys->zap_ptrtbl.zt_shift));
549 		*valp = ZAP_EMBEDDED_PTRTBL_ENT(zap, idx);
550 		return (0);
551 	} else {
552 		return (zap_table_load(zap, &zap->zap_f.zap_phys->zap_ptrtbl,
553 		    idx, valp));
554 	}
555 }
556 
557 static int
558 zap_set_idx_to_blk(zap_t *zap, uint64_t idx, uint64_t blk, dmu_tx_t *tx)
559 {
560 	ASSERT(tx != NULL);
561 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
562 
563 	if (zap->zap_f.zap_phys->zap_ptrtbl.zt_blk == 0) {
564 		ZAP_EMBEDDED_PTRTBL_ENT(zap, idx) = blk;
565 		return (0);
566 	} else {
567 		return (zap_table_store(zap, &zap->zap_f.zap_phys->zap_ptrtbl,
568 		    idx, blk, tx));
569 	}
570 }
571 
572 static int
573 zap_deref_leaf(zap_t *zap, uint64_t h, dmu_tx_t *tx, krw_t lt, zap_leaf_t **lp)
574 {
575 	uint64_t idx, blk;
576 	int err;
577 
578 	ASSERT(zap->zap_dbuf == NULL ||
579 	    zap->zap_f.zap_phys == zap->zap_dbuf->db_data);
580 	ASSERT3U(zap->zap_f.zap_phys->zap_magic, ==, ZAP_MAGIC);
581 	idx = ZAP_HASH_IDX(h, zap->zap_f.zap_phys->zap_ptrtbl.zt_shift);
582 	err = zap_idx_to_blk(zap, idx, &blk);
583 	if (err != 0)
584 		return (err);
585 	err = zap_get_leaf_byblk(zap, blk, tx, lt, lp);
586 
587 	ASSERT(err || ZAP_HASH_IDX(h, (*lp)->l_phys->l_hdr.lh_prefix_len) ==
588 	    (*lp)->l_phys->l_hdr.lh_prefix);
589 	return (err);
590 }
591 
592 static int
593 zap_expand_leaf(zap_name_t *zn, zap_leaf_t *l, dmu_tx_t *tx, zap_leaf_t **lp)
594 {
595 	zap_t *zap = zn->zn_zap;
596 	uint64_t hash = zn->zn_hash;
597 	zap_leaf_t *nl;
598 	int prefix_diff, i, err;
599 	uint64_t sibling;
600 	int old_prefix_len = l->l_phys->l_hdr.lh_prefix_len;
601 
602 	ASSERT3U(old_prefix_len, <=, zap->zap_f.zap_phys->zap_ptrtbl.zt_shift);
603 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
604 
605 	ASSERT3U(ZAP_HASH_IDX(hash, old_prefix_len), ==,
606 	    l->l_phys->l_hdr.lh_prefix);
607 
608 	if (zap_tryupgradedir(zap, tx) == 0 ||
609 	    old_prefix_len == zap->zap_f.zap_phys->zap_ptrtbl.zt_shift) {
610 		/* We failed to upgrade, or need to grow the pointer table */
611 		objset_t *os = zap->zap_objset;
612 		uint64_t object = zap->zap_object;
613 
614 		zap_put_leaf(l);
615 		zap_unlockdir(zap);
616 		err = zap_lockdir(os, object, tx, RW_WRITER,
617 		    FALSE, FALSE, &zn->zn_zap);
618 		zap = zn->zn_zap;
619 		if (err)
620 			return (err);
621 		ASSERT(!zap->zap_ismicro);
622 
623 		while (old_prefix_len ==
624 		    zap->zap_f.zap_phys->zap_ptrtbl.zt_shift) {
625 			err = zap_grow_ptrtbl(zap, tx);
626 			if (err)
627 				return (err);
628 		}
629 
630 		err = zap_deref_leaf(zap, hash, tx, RW_WRITER, &l);
631 		if (err)
632 			return (err);
633 
634 		if (l->l_phys->l_hdr.lh_prefix_len != old_prefix_len) {
635 			/* it split while our locks were down */
636 			*lp = l;
637 			return (0);
638 		}
639 	}
640 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
641 	ASSERT3U(old_prefix_len, <, zap->zap_f.zap_phys->zap_ptrtbl.zt_shift);
642 	ASSERT3U(ZAP_HASH_IDX(hash, old_prefix_len), ==,
643 	    l->l_phys->l_hdr.lh_prefix);
644 
645 	prefix_diff = zap->zap_f.zap_phys->zap_ptrtbl.zt_shift -
646 	    (old_prefix_len + 1);
647 	sibling = (ZAP_HASH_IDX(hash, old_prefix_len + 1) | 1) << prefix_diff;
648 
649 	/* check for i/o errors before doing zap_leaf_split */
650 	for (i = 0; i < (1ULL<<prefix_diff); i++) {
651 		uint64_t blk;
652 		err = zap_idx_to_blk(zap, sibling+i, &blk);
653 		if (err)
654 			return (err);
655 		ASSERT3U(blk, ==, l->l_blkid);
656 	}
657 
658 	nl = zap_create_leaf(zap, tx);
659 	zap_leaf_split(l, nl, zap->zap_normflags != 0);
660 
661 	/* set sibling pointers */
662 	for (i = 0; i < (1ULL << prefix_diff); i++) {
663 		err = zap_set_idx_to_blk(zap, sibling+i, nl->l_blkid, tx);
664 		ASSERT0(err); /* we checked for i/o errors above */
665 	}
666 
667 	if (hash & (1ULL << (64 - l->l_phys->l_hdr.lh_prefix_len))) {
668 		/* we want the sibling */
669 		zap_put_leaf(l);
670 		*lp = nl;
671 	} else {
672 		zap_put_leaf(nl);
673 		*lp = l;
674 	}
675 
676 	return (0);
677 }
678 
679 static void
680 zap_put_leaf_maybe_grow_ptrtbl(zap_name_t *zn, zap_leaf_t *l, dmu_tx_t *tx)
681 {
682 	zap_t *zap = zn->zn_zap;
683 	int shift = zap->zap_f.zap_phys->zap_ptrtbl.zt_shift;
684 	int leaffull = (l->l_phys->l_hdr.lh_prefix_len == shift &&
685 	    l->l_phys->l_hdr.lh_nfree < ZAP_LEAF_LOW_WATER);
686 
687 	zap_put_leaf(l);
688 
689 	if (leaffull || zap->zap_f.zap_phys->zap_ptrtbl.zt_nextblk) {
690 		int err;
691 
692 		/*
693 		 * We are in the middle of growing the pointer table, or
694 		 * this leaf will soon make us grow it.
695 		 */
696 		if (zap_tryupgradedir(zap, tx) == 0) {
697 			objset_t *os = zap->zap_objset;
698 			uint64_t zapobj = zap->zap_object;
699 
700 			zap_unlockdir(zap);
701 			err = zap_lockdir(os, zapobj, tx,
702 			    RW_WRITER, FALSE, FALSE, &zn->zn_zap);
703 			zap = zn->zn_zap;
704 			if (err)
705 				return;
706 		}
707 
708 		/* could have finished growing while our locks were down */
709 		if (zap->zap_f.zap_phys->zap_ptrtbl.zt_shift == shift)
710 			(void) zap_grow_ptrtbl(zap, tx);
711 	}
712 }
713 
714 static int
715 fzap_checkname(zap_name_t *zn)
716 {
717 	if (zn->zn_key_orig_numints * zn->zn_key_intlen > ZAP_MAXNAMELEN)
718 		return (SET_ERROR(ENAMETOOLONG));
719 	return (0);
720 }
721 
722 static int
723 fzap_checksize(uint64_t integer_size, uint64_t num_integers)
724 {
725 	/* Only integer sizes supported by C */
726 	switch (integer_size) {
727 	case 1:
728 	case 2:
729 	case 4:
730 	case 8:
731 		break;
732 	default:
733 		return (SET_ERROR(EINVAL));
734 	}
735 
736 	if (integer_size * num_integers > ZAP_MAXVALUELEN)
737 		return (E2BIG);
738 
739 	return (0);
740 }
741 
742 static int
743 fzap_check(zap_name_t *zn, uint64_t integer_size, uint64_t num_integers)
744 {
745 	int err;
746 
747 	if ((err = fzap_checkname(zn)) != 0)
748 		return (err);
749 	return (fzap_checksize(integer_size, num_integers));
750 }
751 
752 /*
753  * Routines for manipulating attributes.
754  */
755 int
756 fzap_lookup(zap_name_t *zn,
757     uint64_t integer_size, uint64_t num_integers, void *buf,
758     char *realname, int rn_len, boolean_t *ncp)
759 {
760 	zap_leaf_t *l;
761 	int err;
762 	zap_entry_handle_t zeh;
763 
764 	if ((err = fzap_checkname(zn)) != 0)
765 		return (err);
766 
767 	err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, NULL, RW_READER, &l);
768 	if (err != 0)
769 		return (err);
770 	err = zap_leaf_lookup(l, zn, &zeh);
771 	if (err == 0) {
772 		if ((err = fzap_checksize(integer_size, num_integers)) != 0) {
773 			zap_put_leaf(l);
774 			return (err);
775 		}
776 
777 		err = zap_entry_read(&zeh, integer_size, num_integers, buf);
778 		(void) zap_entry_read_name(zn->zn_zap, &zeh, rn_len, realname);
779 		if (ncp) {
780 			*ncp = zap_entry_normalization_conflict(&zeh,
781 			    zn, NULL, zn->zn_zap);
782 		}
783 	}
784 
785 	zap_put_leaf(l);
786 	return (err);
787 }
788 
789 int
790 fzap_add_cd(zap_name_t *zn,
791     uint64_t integer_size, uint64_t num_integers,
792     const void *val, uint32_t cd, dmu_tx_t *tx)
793 {
794 	zap_leaf_t *l;
795 	int err;
796 	zap_entry_handle_t zeh;
797 	zap_t *zap = zn->zn_zap;
798 
799 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
800 	ASSERT(!zap->zap_ismicro);
801 	ASSERT(fzap_check(zn, integer_size, num_integers) == 0);
802 
803 	err = zap_deref_leaf(zap, zn->zn_hash, tx, RW_WRITER, &l);
804 	if (err != 0)
805 		return (err);
806 retry:
807 	err = zap_leaf_lookup(l, zn, &zeh);
808 	if (err == 0) {
809 		err = SET_ERROR(EEXIST);
810 		goto out;
811 	}
812 	if (err != ENOENT)
813 		goto out;
814 
815 	err = zap_entry_create(l, zn, cd,
816 	    integer_size, num_integers, val, &zeh);
817 
818 	if (err == 0) {
819 		zap_increment_num_entries(zap, 1, tx);
820 	} else if (err == EAGAIN) {
821 		err = zap_expand_leaf(zn, l, tx, &l);
822 		zap = zn->zn_zap;	/* zap_expand_leaf() may change zap */
823 		if (err == 0)
824 			goto retry;
825 	}
826 
827 out:
828 	if (zap != NULL)
829 		zap_put_leaf_maybe_grow_ptrtbl(zn, l, tx);
830 	return (err);
831 }
832 
833 int
834 fzap_add(zap_name_t *zn,
835     uint64_t integer_size, uint64_t num_integers,
836     const void *val, dmu_tx_t *tx)
837 {
838 	int err = fzap_check(zn, integer_size, num_integers);
839 	if (err != 0)
840 		return (err);
841 
842 	return (fzap_add_cd(zn, integer_size, num_integers,
843 	    val, ZAP_NEED_CD, tx));
844 }
845 
846 int
847 fzap_update(zap_name_t *zn,
848     int integer_size, uint64_t num_integers, const void *val, dmu_tx_t *tx)
849 {
850 	zap_leaf_t *l;
851 	int err, create;
852 	zap_entry_handle_t zeh;
853 	zap_t *zap = zn->zn_zap;
854 
855 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
856 	err = fzap_check(zn, integer_size, num_integers);
857 	if (err != 0)
858 		return (err);
859 
860 	err = zap_deref_leaf(zap, zn->zn_hash, tx, RW_WRITER, &l);
861 	if (err != 0)
862 		return (err);
863 retry:
864 	err = zap_leaf_lookup(l, zn, &zeh);
865 	create = (err == ENOENT);
866 	ASSERT(err == 0 || err == ENOENT);
867 
868 	if (create) {
869 		err = zap_entry_create(l, zn, ZAP_NEED_CD,
870 		    integer_size, num_integers, val, &zeh);
871 		if (err == 0)
872 			zap_increment_num_entries(zap, 1, tx);
873 	} else {
874 		err = zap_entry_update(&zeh, integer_size, num_integers, val);
875 	}
876 
877 	if (err == EAGAIN) {
878 		err = zap_expand_leaf(zn, l, tx, &l);
879 		zap = zn->zn_zap;	/* zap_expand_leaf() may change zap */
880 		if (err == 0)
881 			goto retry;
882 	}
883 
884 	if (zap != NULL)
885 		zap_put_leaf_maybe_grow_ptrtbl(zn, l, tx);
886 	return (err);
887 }
888 
889 int
890 fzap_length(zap_name_t *zn,
891     uint64_t *integer_size, uint64_t *num_integers)
892 {
893 	zap_leaf_t *l;
894 	int err;
895 	zap_entry_handle_t zeh;
896 
897 	err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, NULL, RW_READER, &l);
898 	if (err != 0)
899 		return (err);
900 	err = zap_leaf_lookup(l, zn, &zeh);
901 	if (err != 0)
902 		goto out;
903 
904 	if (integer_size)
905 		*integer_size = zeh.zeh_integer_size;
906 	if (num_integers)
907 		*num_integers = zeh.zeh_num_integers;
908 out:
909 	zap_put_leaf(l);
910 	return (err);
911 }
912 
913 int
914 fzap_remove(zap_name_t *zn, dmu_tx_t *tx)
915 {
916 	zap_leaf_t *l;
917 	int err;
918 	zap_entry_handle_t zeh;
919 
920 	err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, tx, RW_WRITER, &l);
921 	if (err != 0)
922 		return (err);
923 	err = zap_leaf_lookup(l, zn, &zeh);
924 	if (err == 0) {
925 		zap_entry_remove(&zeh);
926 		zap_increment_num_entries(zn->zn_zap, -1, tx);
927 	}
928 	zap_put_leaf(l);
929 	return (err);
930 }
931 
932 void
933 fzap_prefetch(zap_name_t *zn)
934 {
935 	uint64_t idx, blk;
936 	zap_t *zap = zn->zn_zap;
937 	int bs;
938 
939 	idx = ZAP_HASH_IDX(zn->zn_hash,
940 	    zap->zap_f.zap_phys->zap_ptrtbl.zt_shift);
941 	if (zap_idx_to_blk(zap, idx, &blk) != 0)
942 		return;
943 	bs = FZAP_BLOCK_SHIFT(zap);
944 	dmu_prefetch(zap->zap_objset, zap->zap_object, blk << bs, 1 << bs);
945 }
946 
947 /*
948  * Helper functions for consumers.
949  */
950 
951 uint64_t
952 zap_create_link(objset_t *os, dmu_object_type_t ot, uint64_t parent_obj,
953     const char *name, dmu_tx_t *tx)
954 {
955 	uint64_t new_obj;
956 
957 	VERIFY((new_obj = zap_create(os, ot, DMU_OT_NONE, 0, tx)) > 0);
958 	VERIFY(zap_add(os, parent_obj, name, sizeof (uint64_t), 1, &new_obj,
959 	    tx) == 0);
960 
961 	return (new_obj);
962 }
963 
964 int
965 zap_value_search(objset_t *os, uint64_t zapobj, uint64_t value, uint64_t mask,
966     char *name)
967 {
968 	zap_cursor_t zc;
969 	zap_attribute_t *za;
970 	int err;
971 
972 	if (mask == 0)
973 		mask = -1ULL;
974 
975 	za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
976 	for (zap_cursor_init(&zc, os, zapobj);
977 	    (err = zap_cursor_retrieve(&zc, za)) == 0;
978 	    zap_cursor_advance(&zc)) {
979 		if ((za->za_first_integer & mask) == (value & mask)) {
980 			(void) strcpy(name, za->za_name);
981 			break;
982 		}
983 	}
984 	zap_cursor_fini(&zc);
985 	kmem_free(za, sizeof (zap_attribute_t));
986 	return (err);
987 }
988 
989 int
990 zap_join(objset_t *os, uint64_t fromobj, uint64_t intoobj, dmu_tx_t *tx)
991 {
992 	zap_cursor_t zc;
993 	zap_attribute_t za;
994 	int err;
995 
996 	err = 0;
997 	for (zap_cursor_init(&zc, os, fromobj);
998 	    zap_cursor_retrieve(&zc, &za) == 0;
999 	    (void) zap_cursor_advance(&zc)) {
1000 		if (za.za_integer_length != 8 || za.za_num_integers != 1) {
1001 			err = SET_ERROR(EINVAL);
1002 			break;
1003 		}
1004 		err = zap_add(os, intoobj, za.za_name,
1005 		    8, 1, &za.za_first_integer, tx);
1006 		if (err)
1007 			break;
1008 	}
1009 	zap_cursor_fini(&zc);
1010 	return (err);
1011 }
1012 
1013 int
1014 zap_join_key(objset_t *os, uint64_t fromobj, uint64_t intoobj,
1015     uint64_t value, dmu_tx_t *tx)
1016 {
1017 	zap_cursor_t zc;
1018 	zap_attribute_t za;
1019 	int err;
1020 
1021 	err = 0;
1022 	for (zap_cursor_init(&zc, os, fromobj);
1023 	    zap_cursor_retrieve(&zc, &za) == 0;
1024 	    (void) zap_cursor_advance(&zc)) {
1025 		if (za.za_integer_length != 8 || za.za_num_integers != 1) {
1026 			err = SET_ERROR(EINVAL);
1027 			break;
1028 		}
1029 		err = zap_add(os, intoobj, za.za_name,
1030 		    8, 1, &value, tx);
1031 		if (err)
1032 			break;
1033 	}
1034 	zap_cursor_fini(&zc);
1035 	return (err);
1036 }
1037 
1038 int
1039 zap_join_increment(objset_t *os, uint64_t fromobj, uint64_t intoobj,
1040     dmu_tx_t *tx)
1041 {
1042 	zap_cursor_t zc;
1043 	zap_attribute_t za;
1044 	int err;
1045 
1046 	err = 0;
1047 	for (zap_cursor_init(&zc, os, fromobj);
1048 	    zap_cursor_retrieve(&zc, &za) == 0;
1049 	    (void) zap_cursor_advance(&zc)) {
1050 		uint64_t delta = 0;
1051 
1052 		if (za.za_integer_length != 8 || za.za_num_integers != 1) {
1053 			err = SET_ERROR(EINVAL);
1054 			break;
1055 		}
1056 
1057 		err = zap_lookup(os, intoobj, za.za_name, 8, 1, &delta);
1058 		if (err != 0 && err != ENOENT)
1059 			break;
1060 		delta += za.za_first_integer;
1061 		err = zap_update(os, intoobj, za.za_name, 8, 1, &delta, tx);
1062 		if (err)
1063 			break;
1064 	}
1065 	zap_cursor_fini(&zc);
1066 	return (err);
1067 }
1068 
1069 int
1070 zap_add_int(objset_t *os, uint64_t obj, uint64_t value, dmu_tx_t *tx)
1071 {
1072 	char name[20];
1073 
1074 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)value);
1075 	return (zap_add(os, obj, name, 8, 1, &value, tx));
1076 }
1077 
1078 int
1079 zap_remove_int(objset_t *os, uint64_t obj, uint64_t value, dmu_tx_t *tx)
1080 {
1081 	char name[20];
1082 
1083 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)value);
1084 	return (zap_remove(os, obj, name, tx));
1085 }
1086 
1087 int
1088 zap_lookup_int(objset_t *os, uint64_t obj, uint64_t value)
1089 {
1090 	char name[20];
1091 
1092 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)value);
1093 	return (zap_lookup(os, obj, name, 8, 1, &value));
1094 }
1095 
1096 int
1097 zap_add_int_key(objset_t *os, uint64_t obj,
1098     uint64_t key, uint64_t value, dmu_tx_t *tx)
1099 {
1100 	char name[20];
1101 
1102 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)key);
1103 	return (zap_add(os, obj, name, 8, 1, &value, tx));
1104 }
1105 
1106 int
1107 zap_update_int_key(objset_t *os, uint64_t obj,
1108     uint64_t key, uint64_t value, dmu_tx_t *tx)
1109 {
1110 	char name[20];
1111 
1112 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)key);
1113 	return (zap_update(os, obj, name, 8, 1, &value, tx));
1114 }
1115 
1116 int
1117 zap_lookup_int_key(objset_t *os, uint64_t obj, uint64_t key, uint64_t *valuep)
1118 {
1119 	char name[20];
1120 
1121 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)key);
1122 	return (zap_lookup(os, obj, name, 8, 1, valuep));
1123 }
1124 
1125 int
1126 zap_increment(objset_t *os, uint64_t obj, const char *name, int64_t delta,
1127     dmu_tx_t *tx)
1128 {
1129 	uint64_t value = 0;
1130 	int err;
1131 
1132 	if (delta == 0)
1133 		return (0);
1134 
1135 	err = zap_lookup(os, obj, name, 8, 1, &value);
1136 	if (err != 0 && err != ENOENT)
1137 		return (err);
1138 	value += delta;
1139 	if (value == 0)
1140 		err = zap_remove(os, obj, name, tx);
1141 	else
1142 		err = zap_update(os, obj, name, 8, 1, &value, tx);
1143 	return (err);
1144 }
1145 
1146 int
1147 zap_increment_int(objset_t *os, uint64_t obj, uint64_t key, int64_t delta,
1148     dmu_tx_t *tx)
1149 {
1150 	char name[20];
1151 
1152 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)key);
1153 	return (zap_increment(os, obj, name, delta, tx));
1154 }
1155 
1156 /*
1157  * Routines for iterating over the attributes.
1158  */
1159 
1160 int
1161 fzap_cursor_retrieve(zap_t *zap, zap_cursor_t *zc, zap_attribute_t *za)
1162 {
1163 	int err = ENOENT;
1164 	zap_entry_handle_t zeh;
1165 	zap_leaf_t *l;
1166 
1167 	/* retrieve the next entry at or after zc_hash/zc_cd */
1168 	/* if no entry, return ENOENT */
1169 
1170 	if (zc->zc_leaf &&
1171 	    (ZAP_HASH_IDX(zc->zc_hash,
1172 	    zc->zc_leaf->l_phys->l_hdr.lh_prefix_len) !=
1173 	    zc->zc_leaf->l_phys->l_hdr.lh_prefix)) {
1174 		rw_enter(&zc->zc_leaf->l_rwlock, RW_READER);
1175 		zap_put_leaf(zc->zc_leaf);
1176 		zc->zc_leaf = NULL;
1177 	}
1178 
1179 again:
1180 	if (zc->zc_leaf == NULL) {
1181 		err = zap_deref_leaf(zap, zc->zc_hash, NULL, RW_READER,
1182 		    &zc->zc_leaf);
1183 		if (err != 0)
1184 			return (err);
1185 	} else {
1186 		rw_enter(&zc->zc_leaf->l_rwlock, RW_READER);
1187 	}
1188 	l = zc->zc_leaf;
1189 
1190 	err = zap_leaf_lookup_closest(l, zc->zc_hash, zc->zc_cd, &zeh);
1191 
1192 	if (err == ENOENT) {
1193 		uint64_t nocare =
1194 		    (1ULL << (64 - l->l_phys->l_hdr.lh_prefix_len)) - 1;
1195 		zc->zc_hash = (zc->zc_hash & ~nocare) + nocare + 1;
1196 		zc->zc_cd = 0;
1197 		if (l->l_phys->l_hdr.lh_prefix_len == 0 || zc->zc_hash == 0) {
1198 			zc->zc_hash = -1ULL;
1199 		} else {
1200 			zap_put_leaf(zc->zc_leaf);
1201 			zc->zc_leaf = NULL;
1202 			goto again;
1203 		}
1204 	}
1205 
1206 	if (err == 0) {
1207 		zc->zc_hash = zeh.zeh_hash;
1208 		zc->zc_cd = zeh.zeh_cd;
1209 		za->za_integer_length = zeh.zeh_integer_size;
1210 		za->za_num_integers = zeh.zeh_num_integers;
1211 		if (zeh.zeh_num_integers == 0) {
1212 			za->za_first_integer = 0;
1213 		} else {
1214 			err = zap_entry_read(&zeh, 8, 1, &za->za_first_integer);
1215 			ASSERT(err == 0 || err == EOVERFLOW);
1216 		}
1217 		err = zap_entry_read_name(zap, &zeh,
1218 		    sizeof (za->za_name), za->za_name);
1219 		ASSERT(err == 0);
1220 
1221 		za->za_normalization_conflict =
1222 		    zap_entry_normalization_conflict(&zeh,
1223 		    NULL, za->za_name, zap);
1224 	}
1225 	rw_exit(&zc->zc_leaf->l_rwlock);
1226 	return (err);
1227 }
1228 
1229 static void
1230 zap_stats_ptrtbl(zap_t *zap, uint64_t *tbl, int len, zap_stats_t *zs)
1231 {
1232 	int i, err;
1233 	uint64_t lastblk = 0;
1234 
1235 	/*
1236 	 * NB: if a leaf has more pointers than an entire ptrtbl block
1237 	 * can hold, then it'll be accounted for more than once, since
1238 	 * we won't have lastblk.
1239 	 */
1240 	for (i = 0; i < len; i++) {
1241 		zap_leaf_t *l;
1242 
1243 		if (tbl[i] == lastblk)
1244 			continue;
1245 		lastblk = tbl[i];
1246 
1247 		err = zap_get_leaf_byblk(zap, tbl[i], NULL, RW_READER, &l);
1248 		if (err == 0) {
1249 			zap_leaf_stats(zap, l, zs);
1250 			zap_put_leaf(l);
1251 		}
1252 	}
1253 }
1254 
1255 void
1256 fzap_get_stats(zap_t *zap, zap_stats_t *zs)
1257 {
1258 	int bs = FZAP_BLOCK_SHIFT(zap);
1259 	zs->zs_blocksize = 1ULL << bs;
1260 
1261 	/*
1262 	 * Set zap_phys_t fields
1263 	 */
1264 	zs->zs_num_leafs = zap->zap_f.zap_phys->zap_num_leafs;
1265 	zs->zs_num_entries = zap->zap_f.zap_phys->zap_num_entries;
1266 	zs->zs_num_blocks = zap->zap_f.zap_phys->zap_freeblk;
1267 	zs->zs_block_type = zap->zap_f.zap_phys->zap_block_type;
1268 	zs->zs_magic = zap->zap_f.zap_phys->zap_magic;
1269 	zs->zs_salt = zap->zap_f.zap_phys->zap_salt;
1270 
1271 	/*
1272 	 * Set zap_ptrtbl fields
1273 	 */
1274 	zs->zs_ptrtbl_len = 1ULL << zap->zap_f.zap_phys->zap_ptrtbl.zt_shift;
1275 	zs->zs_ptrtbl_nextblk = zap->zap_f.zap_phys->zap_ptrtbl.zt_nextblk;
1276 	zs->zs_ptrtbl_blks_copied =
1277 	    zap->zap_f.zap_phys->zap_ptrtbl.zt_blks_copied;
1278 	zs->zs_ptrtbl_zt_blk = zap->zap_f.zap_phys->zap_ptrtbl.zt_blk;
1279 	zs->zs_ptrtbl_zt_numblks = zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks;
1280 	zs->zs_ptrtbl_zt_shift = zap->zap_f.zap_phys->zap_ptrtbl.zt_shift;
1281 
1282 	if (zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks == 0) {
1283 		/* the ptrtbl is entirely in the header block. */
1284 		zap_stats_ptrtbl(zap, &ZAP_EMBEDDED_PTRTBL_ENT(zap, 0),
1285 		    1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap), zs);
1286 	} else {
1287 		int b;
1288 
1289 		dmu_prefetch(zap->zap_objset, zap->zap_object,
1290 		    zap->zap_f.zap_phys->zap_ptrtbl.zt_blk << bs,
1291 		    zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks << bs);
1292 
1293 		for (b = 0; b < zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks;
1294 		    b++) {
1295 			dmu_buf_t *db;
1296 			int err;
1297 
1298 			err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
1299 			    (zap->zap_f.zap_phys->zap_ptrtbl.zt_blk + b) << bs,
1300 			    FTAG, &db, DMU_READ_NO_PREFETCH);
1301 			if (err == 0) {
1302 				zap_stats_ptrtbl(zap, db->db_data,
1303 				    1<<(bs-3), zs);
1304 				dmu_buf_rele(db, FTAG);
1305 			}
1306 		}
1307 	}
1308 }
1309 
1310 int
1311 fzap_count_write(zap_name_t *zn, int add, uint64_t *towrite,
1312     uint64_t *tooverwrite)
1313 {
1314 	zap_t *zap = zn->zn_zap;
1315 	zap_leaf_t *l;
1316 	int err;
1317 
1318 	/*
1319 	 * Account for the header block of the fatzap.
1320 	 */
1321 	if (!add && dmu_buf_freeable(zap->zap_dbuf)) {
1322 		*tooverwrite += zap->zap_dbuf->db_size;
1323 	} else {
1324 		*towrite += zap->zap_dbuf->db_size;
1325 	}
1326 
1327 	/*
1328 	 * Account for the pointer table blocks.
1329 	 * If we are adding we need to account for the following cases :
1330 	 * - If the pointer table is embedded, this operation could force an
1331 	 *   external pointer table.
1332 	 * - If this already has an external pointer table this operation
1333 	 *   could extend the table.
1334 	 */
1335 	if (add) {
1336 		if (zap->zap_f.zap_phys->zap_ptrtbl.zt_blk == 0)
1337 			*towrite += zap->zap_dbuf->db_size;
1338 		else
1339 			*towrite += (zap->zap_dbuf->db_size * 3);
1340 	}
1341 
1342 	/*
1343 	 * Now, check if the block containing leaf is freeable
1344 	 * and account accordingly.
1345 	 */
1346 	err = zap_deref_leaf(zap, zn->zn_hash, NULL, RW_READER, &l);
1347 	if (err != 0) {
1348 		return (err);
1349 	}
1350 
1351 	if (!add && dmu_buf_freeable(l->l_dbuf)) {
1352 		*tooverwrite += l->l_dbuf->db_size;
1353 	} else {
1354 		/*
1355 		 * If this an add operation, the leaf block could split.
1356 		 * Hence, we need to account for an additional leaf block.
1357 		 */
1358 		*towrite += (add ? 2 : 1) * l->l_dbuf->db_size;
1359 	}
1360 
1361 	zap_put_leaf(l);
1362 	return (0);
1363 }
1364