xref: /freebsd/sys/contrib/openzfs/module/zfs/zap.c (revision e92ffd9b626833ebdbf2742c8ffddc6cd94b963e)
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) 2012, 2018 by Delphix. All rights reserved.
24  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25  */
26 
27 /*
28  * This file contains the top half of the zfs directory structure
29  * implementation. The bottom half is in zap_leaf.c.
30  *
31  * The zdir is an extendable hash data structure. There is a table of
32  * pointers to buckets (zap_t->zd_data->zd_leafs). The buckets are
33  * each a constant size and hold a variable number of directory entries.
34  * The buckets (aka "leaf nodes") are implemented in zap_leaf.c.
35  *
36  * The pointer table holds a power of 2 number of pointers.
37  * (1<<zap_t->zd_data->zd_phys->zd_prefix_len).  The bucket pointed to
38  * by the pointer at index i in the table holds entries whose hash value
39  * has a zd_prefix_len - bit prefix
40  */
41 
42 #include <sys/spa.h>
43 #include <sys/dmu.h>
44 #include <sys/zfs_context.h>
45 #include <sys/zfs_znode.h>
46 #include <sys/fs/zfs.h>
47 #include <sys/zap.h>
48 #include <sys/zap_impl.h>
49 #include <sys/zap_leaf.h>
50 
51 /*
52  * If zap_iterate_prefetch is set, we will prefetch the entire ZAP object
53  * (all leaf blocks) when we start iterating over it.
54  *
55  * For zap_cursor_init(), the callers all intend to iterate through all the
56  * entries.  There are a few cases where an error (typically i/o error) could
57  * cause it to bail out early.
58  *
59  * For zap_cursor_init_serialized(), there are callers that do the iteration
60  * outside of ZFS.  Typically they would iterate over everything, but we
61  * don't have control of that.  E.g. zfs_ioc_snapshot_list_next(),
62  * zcp_snapshots_iter(), and other iterators over things in the MOS - these
63  * are called by /sbin/zfs and channel programs.  The other example is
64  * zfs_readdir() which iterates over directory entries for the getdents()
65  * syscall.  /sbin/ls iterates to the end (unless it receives a signal), but
66  * userland doesn't have to.
67  *
68  * Given that the ZAP entries aren't returned in a specific order, the only
69  * legitimate use cases for partial iteration would be:
70  *
71  * 1. Pagination: e.g. you only want to display 100 entries at a time, so you
72  *    get the first 100 and then wait for the user to hit "next page", which
73  *    they may never do).
74  *
75  * 2. You want to know if there are more than X entries, without relying on
76  *    the zfs-specific implementation of the directory's st_size (which is
77  *    the number of entries).
78  */
79 static int zap_iterate_prefetch = B_TRUE;
80 
81 int fzap_default_block_shift = 14; /* 16k blocksize */
82 
83 extern inline zap_phys_t *zap_f_phys(zap_t *zap);
84 
85 static uint64_t zap_allocate_blocks(zap_t *zap, int nblocks);
86 
87 void
88 fzap_byteswap(void *vbuf, size_t size)
89 {
90 	uint64_t block_type = *(uint64_t *)vbuf;
91 
92 	if (block_type == ZBT_LEAF || block_type == BSWAP_64(ZBT_LEAF))
93 		zap_leaf_byteswap(vbuf, size);
94 	else {
95 		/* it's a ptrtbl block */
96 		byteswap_uint64_array(vbuf, size);
97 	}
98 }
99 
100 void
101 fzap_upgrade(zap_t *zap, dmu_tx_t *tx, zap_flags_t flags)
102 {
103 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
104 	zap->zap_ismicro = FALSE;
105 
106 	zap->zap_dbu.dbu_evict_func_sync = zap_evict_sync;
107 	zap->zap_dbu.dbu_evict_func_async = NULL;
108 
109 	mutex_init(&zap->zap_f.zap_num_entries_mtx, 0, MUTEX_DEFAULT, 0);
110 	zap->zap_f.zap_block_shift = highbit64(zap->zap_dbuf->db_size) - 1;
111 
112 	zap_phys_t *zp = zap_f_phys(zap);
113 	/*
114 	 * explicitly zero it since it might be coming from an
115 	 * initialized microzap
116 	 */
117 	bzero(zap->zap_dbuf->db_data, zap->zap_dbuf->db_size);
118 	zp->zap_block_type = ZBT_HEADER;
119 	zp->zap_magic = ZAP_MAGIC;
120 
121 	zp->zap_ptrtbl.zt_shift = ZAP_EMBEDDED_PTRTBL_SHIFT(zap);
122 
123 	zp->zap_freeblk = 2;		/* block 1 will be the first leaf */
124 	zp->zap_num_leafs = 1;
125 	zp->zap_num_entries = 0;
126 	zp->zap_salt = zap->zap_salt;
127 	zp->zap_normflags = zap->zap_normflags;
128 	zp->zap_flags = flags;
129 
130 	/* block 1 will be the first leaf */
131 	for (int i = 0; i < (1<<zp->zap_ptrtbl.zt_shift); i++)
132 		ZAP_EMBEDDED_PTRTBL_ENT(zap, i) = 1;
133 
134 	/*
135 	 * set up block 1 - the first leaf
136 	 */
137 	dmu_buf_t *db;
138 	VERIFY0(dmu_buf_hold(zap->zap_objset, zap->zap_object,
139 	    1<<FZAP_BLOCK_SHIFT(zap), FTAG, &db, DMU_READ_NO_PREFETCH));
140 	dmu_buf_will_dirty(db, tx);
141 
142 	zap_leaf_t *l = kmem_zalloc(sizeof (zap_leaf_t), KM_SLEEP);
143 	l->l_dbuf = db;
144 
145 	zap_leaf_init(l, zp->zap_normflags != 0);
146 
147 	kmem_free(l, sizeof (zap_leaf_t));
148 	dmu_buf_rele(db, FTAG);
149 }
150 
151 static int
152 zap_tryupgradedir(zap_t *zap, dmu_tx_t *tx)
153 {
154 	if (RW_WRITE_HELD(&zap->zap_rwlock))
155 		return (1);
156 	if (rw_tryupgrade(&zap->zap_rwlock)) {
157 		dmu_buf_will_dirty(zap->zap_dbuf, tx);
158 		return (1);
159 	}
160 	return (0);
161 }
162 
163 /*
164  * Generic routines for dealing with the pointer & cookie tables.
165  */
166 
167 static int
168 zap_table_grow(zap_t *zap, zap_table_phys_t *tbl,
169     void (*transfer_func)(const uint64_t *src, uint64_t *dst, int n),
170     dmu_tx_t *tx)
171 {
172 	uint64_t newblk;
173 	int bs = FZAP_BLOCK_SHIFT(zap);
174 	int hepb = 1<<(bs-4);
175 	/* hepb = half the number of entries in a block */
176 
177 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
178 	ASSERT(tbl->zt_blk != 0);
179 	ASSERT(tbl->zt_numblks > 0);
180 
181 	if (tbl->zt_nextblk != 0) {
182 		newblk = tbl->zt_nextblk;
183 	} else {
184 		newblk = zap_allocate_blocks(zap, tbl->zt_numblks * 2);
185 		tbl->zt_nextblk = newblk;
186 		ASSERT0(tbl->zt_blks_copied);
187 		dmu_prefetch(zap->zap_objset, zap->zap_object, 0,
188 		    tbl->zt_blk << bs, tbl->zt_numblks << bs,
189 		    ZIO_PRIORITY_SYNC_READ);
190 	}
191 
192 	/*
193 	 * Copy the ptrtbl from the old to new location.
194 	 */
195 
196 	uint64_t b = tbl->zt_blks_copied;
197 	dmu_buf_t *db_old;
198 	int err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
199 	    (tbl->zt_blk + b) << bs, FTAG, &db_old, DMU_READ_NO_PREFETCH);
200 	if (err != 0)
201 		return (err);
202 
203 	/* first half of entries in old[b] go to new[2*b+0] */
204 	dmu_buf_t *db_new;
205 	VERIFY0(dmu_buf_hold(zap->zap_objset, zap->zap_object,
206 	    (newblk + 2*b+0) << bs, FTAG, &db_new, DMU_READ_NO_PREFETCH));
207 	dmu_buf_will_dirty(db_new, tx);
208 	transfer_func(db_old->db_data, db_new->db_data, hepb);
209 	dmu_buf_rele(db_new, FTAG);
210 
211 	/* second half of entries in old[b] go to new[2*b+1] */
212 	VERIFY0(dmu_buf_hold(zap->zap_objset, zap->zap_object,
213 	    (newblk + 2*b+1) << bs, FTAG, &db_new, DMU_READ_NO_PREFETCH));
214 	dmu_buf_will_dirty(db_new, tx);
215 	transfer_func((uint64_t *)db_old->db_data + hepb,
216 	    db_new->db_data, hepb);
217 	dmu_buf_rele(db_new, FTAG);
218 
219 	dmu_buf_rele(db_old, FTAG);
220 
221 	tbl->zt_blks_copied++;
222 
223 	dprintf("copied block %llu of %llu\n",
224 	    (u_longlong_t)tbl->zt_blks_copied,
225 	    (u_longlong_t)tbl->zt_numblks);
226 
227 	if (tbl->zt_blks_copied == tbl->zt_numblks) {
228 		(void) dmu_free_range(zap->zap_objset, zap->zap_object,
229 		    tbl->zt_blk << bs, tbl->zt_numblks << bs, tx);
230 
231 		tbl->zt_blk = newblk;
232 		tbl->zt_numblks *= 2;
233 		tbl->zt_shift++;
234 		tbl->zt_nextblk = 0;
235 		tbl->zt_blks_copied = 0;
236 
237 		dprintf("finished; numblocks now %llu (%uk entries)\n",
238 		    (u_longlong_t)tbl->zt_numblks, 1<<(tbl->zt_shift-10));
239 	}
240 
241 	return (0);
242 }
243 
244 static int
245 zap_table_store(zap_t *zap, zap_table_phys_t *tbl, uint64_t idx, uint64_t val,
246     dmu_tx_t *tx)
247 {
248 	int bs = FZAP_BLOCK_SHIFT(zap);
249 
250 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
251 	ASSERT(tbl->zt_blk != 0);
252 
253 	dprintf("storing %llx at index %llx\n", (u_longlong_t)val,
254 	    (u_longlong_t)idx);
255 
256 	uint64_t blk = idx >> (bs-3);
257 	uint64_t off = idx & ((1<<(bs-3))-1);
258 
259 	dmu_buf_t *db;
260 	int err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
261 	    (tbl->zt_blk + blk) << bs, FTAG, &db, DMU_READ_NO_PREFETCH);
262 	if (err != 0)
263 		return (err);
264 	dmu_buf_will_dirty(db, tx);
265 
266 	if (tbl->zt_nextblk != 0) {
267 		uint64_t idx2 = idx * 2;
268 		uint64_t blk2 = idx2 >> (bs-3);
269 		uint64_t off2 = idx2 & ((1<<(bs-3))-1);
270 		dmu_buf_t *db2;
271 
272 		err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
273 		    (tbl->zt_nextblk + blk2) << bs, FTAG, &db2,
274 		    DMU_READ_NO_PREFETCH);
275 		if (err != 0) {
276 			dmu_buf_rele(db, FTAG);
277 			return (err);
278 		}
279 		dmu_buf_will_dirty(db2, tx);
280 		((uint64_t *)db2->db_data)[off2] = val;
281 		((uint64_t *)db2->db_data)[off2+1] = val;
282 		dmu_buf_rele(db2, FTAG);
283 	}
284 
285 	((uint64_t *)db->db_data)[off] = val;
286 	dmu_buf_rele(db, FTAG);
287 
288 	return (0);
289 }
290 
291 static int
292 zap_table_load(zap_t *zap, zap_table_phys_t *tbl, uint64_t idx, uint64_t *valp)
293 {
294 	int bs = FZAP_BLOCK_SHIFT(zap);
295 
296 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
297 
298 	uint64_t blk = idx >> (bs-3);
299 	uint64_t off = idx & ((1<<(bs-3))-1);
300 
301 	/*
302 	 * Note: this is equivalent to dmu_buf_hold(), but we use
303 	 * _dnode_enter / _by_dnode because it's faster because we don't
304 	 * have to hold the dnode.
305 	 */
306 	dnode_t *dn = dmu_buf_dnode_enter(zap->zap_dbuf);
307 	dmu_buf_t *db;
308 	int err = dmu_buf_hold_by_dnode(dn,
309 	    (tbl->zt_blk + blk) << bs, FTAG, &db, DMU_READ_NO_PREFETCH);
310 	dmu_buf_dnode_exit(zap->zap_dbuf);
311 	if (err != 0)
312 		return (err);
313 	*valp = ((uint64_t *)db->db_data)[off];
314 	dmu_buf_rele(db, FTAG);
315 
316 	if (tbl->zt_nextblk != 0) {
317 		/*
318 		 * read the nextblk for the sake of i/o error checking,
319 		 * so that zap_table_load() will catch errors for
320 		 * zap_table_store.
321 		 */
322 		blk = (idx*2) >> (bs-3);
323 
324 		dn = dmu_buf_dnode_enter(zap->zap_dbuf);
325 		err = dmu_buf_hold_by_dnode(dn,
326 		    (tbl->zt_nextblk + blk) << bs, FTAG, &db,
327 		    DMU_READ_NO_PREFETCH);
328 		dmu_buf_dnode_exit(zap->zap_dbuf);
329 		if (err == 0)
330 			dmu_buf_rele(db, FTAG);
331 	}
332 	return (err);
333 }
334 
335 /*
336  * Routines for growing the ptrtbl.
337  */
338 
339 static void
340 zap_ptrtbl_transfer(const uint64_t *src, uint64_t *dst, int n)
341 {
342 	for (int i = 0; i < n; i++) {
343 		uint64_t lb = src[i];
344 		dst[2 * i + 0] = lb;
345 		dst[2 * i + 1] = lb;
346 	}
347 }
348 
349 static int
350 zap_grow_ptrtbl(zap_t *zap, dmu_tx_t *tx)
351 {
352 	/*
353 	 * The pointer table should never use more hash bits than we
354 	 * have (otherwise we'd be using useless zero bits to index it).
355 	 * If we are within 2 bits of running out, stop growing, since
356 	 * this is already an aberrant condition.
357 	 */
358 	if (zap_f_phys(zap)->zap_ptrtbl.zt_shift >= zap_hashbits(zap) - 2)
359 		return (SET_ERROR(ENOSPC));
360 
361 	if (zap_f_phys(zap)->zap_ptrtbl.zt_numblks == 0) {
362 		/*
363 		 * We are outgrowing the "embedded" ptrtbl (the one
364 		 * stored in the header block).  Give it its own entire
365 		 * block, which will double the size of the ptrtbl.
366 		 */
367 		ASSERT3U(zap_f_phys(zap)->zap_ptrtbl.zt_shift, ==,
368 		    ZAP_EMBEDDED_PTRTBL_SHIFT(zap));
369 		ASSERT0(zap_f_phys(zap)->zap_ptrtbl.zt_blk);
370 
371 		uint64_t newblk = zap_allocate_blocks(zap, 1);
372 		dmu_buf_t *db_new;
373 		int err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
374 		    newblk << FZAP_BLOCK_SHIFT(zap), FTAG, &db_new,
375 		    DMU_READ_NO_PREFETCH);
376 		if (err != 0)
377 			return (err);
378 		dmu_buf_will_dirty(db_new, tx);
379 		zap_ptrtbl_transfer(&ZAP_EMBEDDED_PTRTBL_ENT(zap, 0),
380 		    db_new->db_data, 1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap));
381 		dmu_buf_rele(db_new, FTAG);
382 
383 		zap_f_phys(zap)->zap_ptrtbl.zt_blk = newblk;
384 		zap_f_phys(zap)->zap_ptrtbl.zt_numblks = 1;
385 		zap_f_phys(zap)->zap_ptrtbl.zt_shift++;
386 
387 		ASSERT3U(1ULL << zap_f_phys(zap)->zap_ptrtbl.zt_shift, ==,
388 		    zap_f_phys(zap)->zap_ptrtbl.zt_numblks <<
389 		    (FZAP_BLOCK_SHIFT(zap)-3));
390 
391 		return (0);
392 	} else {
393 		return (zap_table_grow(zap, &zap_f_phys(zap)->zap_ptrtbl,
394 		    zap_ptrtbl_transfer, tx));
395 	}
396 }
397 
398 static void
399 zap_increment_num_entries(zap_t *zap, int delta, dmu_tx_t *tx)
400 {
401 	dmu_buf_will_dirty(zap->zap_dbuf, tx);
402 	mutex_enter(&zap->zap_f.zap_num_entries_mtx);
403 	ASSERT(delta > 0 || zap_f_phys(zap)->zap_num_entries >= -delta);
404 	zap_f_phys(zap)->zap_num_entries += delta;
405 	mutex_exit(&zap->zap_f.zap_num_entries_mtx);
406 }
407 
408 static uint64_t
409 zap_allocate_blocks(zap_t *zap, int nblocks)
410 {
411 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
412 	uint64_t newblk = zap_f_phys(zap)->zap_freeblk;
413 	zap_f_phys(zap)->zap_freeblk += nblocks;
414 	return (newblk);
415 }
416 
417 static void
418 zap_leaf_evict_sync(void *dbu)
419 {
420 	zap_leaf_t *l = dbu;
421 
422 	rw_destroy(&l->l_rwlock);
423 	kmem_free(l, sizeof (zap_leaf_t));
424 }
425 
426 static zap_leaf_t *
427 zap_create_leaf(zap_t *zap, dmu_tx_t *tx)
428 {
429 	zap_leaf_t *l = kmem_zalloc(sizeof (zap_leaf_t), KM_SLEEP);
430 
431 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
432 
433 	rw_init(&l->l_rwlock, NULL, RW_NOLOCKDEP, NULL);
434 	rw_enter(&l->l_rwlock, RW_WRITER);
435 	l->l_blkid = zap_allocate_blocks(zap, 1);
436 	l->l_dbuf = NULL;
437 
438 	VERIFY0(dmu_buf_hold(zap->zap_objset, zap->zap_object,
439 	    l->l_blkid << FZAP_BLOCK_SHIFT(zap), NULL, &l->l_dbuf,
440 	    DMU_READ_NO_PREFETCH));
441 	dmu_buf_init_user(&l->l_dbu, zap_leaf_evict_sync, NULL, &l->l_dbuf);
442 	VERIFY3P(NULL, ==, dmu_buf_set_user(l->l_dbuf, &l->l_dbu));
443 	dmu_buf_will_dirty(l->l_dbuf, tx);
444 
445 	zap_leaf_init(l, zap->zap_normflags != 0);
446 
447 	zap_f_phys(zap)->zap_num_leafs++;
448 
449 	return (l);
450 }
451 
452 int
453 fzap_count(zap_t *zap, uint64_t *count)
454 {
455 	ASSERT(!zap->zap_ismicro);
456 	mutex_enter(&zap->zap_f.zap_num_entries_mtx); /* unnecessary */
457 	*count = zap_f_phys(zap)->zap_num_entries;
458 	mutex_exit(&zap->zap_f.zap_num_entries_mtx);
459 	return (0);
460 }
461 
462 /*
463  * Routines for obtaining zap_leaf_t's
464  */
465 
466 void
467 zap_put_leaf(zap_leaf_t *l)
468 {
469 	rw_exit(&l->l_rwlock);
470 	dmu_buf_rele(l->l_dbuf, NULL);
471 }
472 
473 static zap_leaf_t *
474 zap_open_leaf(uint64_t blkid, dmu_buf_t *db)
475 {
476 	ASSERT(blkid != 0);
477 
478 	zap_leaf_t *l = kmem_zalloc(sizeof (zap_leaf_t), KM_SLEEP);
479 	rw_init(&l->l_rwlock, NULL, RW_DEFAULT, NULL);
480 	rw_enter(&l->l_rwlock, RW_WRITER);
481 	l->l_blkid = blkid;
482 	l->l_bs = highbit64(db->db_size) - 1;
483 	l->l_dbuf = db;
484 
485 	dmu_buf_init_user(&l->l_dbu, zap_leaf_evict_sync, NULL, &l->l_dbuf);
486 	zap_leaf_t *winner = dmu_buf_set_user(db, &l->l_dbu);
487 
488 	rw_exit(&l->l_rwlock);
489 	if (winner != NULL) {
490 		/* someone else set it first */
491 		zap_leaf_evict_sync(&l->l_dbu);
492 		l = winner;
493 	}
494 
495 	/*
496 	 * lhr_pad was previously used for the next leaf in the leaf
497 	 * chain.  There should be no chained leafs (as we have removed
498 	 * support for them).
499 	 */
500 	ASSERT0(zap_leaf_phys(l)->l_hdr.lh_pad1);
501 
502 	/*
503 	 * There should be more hash entries than there can be
504 	 * chunks to put in the hash table
505 	 */
506 	ASSERT3U(ZAP_LEAF_HASH_NUMENTRIES(l), >, ZAP_LEAF_NUMCHUNKS(l) / 3);
507 
508 	/* The chunks should begin at the end of the hash table */
509 	ASSERT3P(&ZAP_LEAF_CHUNK(l, 0), ==, (zap_leaf_chunk_t *)
510 	    &zap_leaf_phys(l)->l_hash[ZAP_LEAF_HASH_NUMENTRIES(l)]);
511 
512 	/* The chunks should end at the end of the block */
513 	ASSERT3U((uintptr_t)&ZAP_LEAF_CHUNK(l, ZAP_LEAF_NUMCHUNKS(l)) -
514 	    (uintptr_t)zap_leaf_phys(l), ==, l->l_dbuf->db_size);
515 
516 	return (l);
517 }
518 
519 static int
520 zap_get_leaf_byblk(zap_t *zap, uint64_t blkid, dmu_tx_t *tx, krw_t lt,
521     zap_leaf_t **lp)
522 {
523 	dmu_buf_t *db;
524 
525 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
526 
527 	/*
528 	 * If system crashed just after dmu_free_long_range in zfs_rmnode, we
529 	 * would be left with an empty xattr dir in delete queue. blkid=0
530 	 * would be passed in when doing zfs_purgedir. If that's the case we
531 	 * should just return immediately. The underlying objects should
532 	 * already be freed, so this should be perfectly fine.
533 	 */
534 	if (blkid == 0)
535 		return (SET_ERROR(ENOENT));
536 
537 	int bs = FZAP_BLOCK_SHIFT(zap);
538 	dnode_t *dn = dmu_buf_dnode_enter(zap->zap_dbuf);
539 	int err = dmu_buf_hold_by_dnode(dn,
540 	    blkid << bs, NULL, &db, DMU_READ_NO_PREFETCH);
541 	dmu_buf_dnode_exit(zap->zap_dbuf);
542 	if (err != 0)
543 		return (err);
544 
545 	ASSERT3U(db->db_object, ==, zap->zap_object);
546 	ASSERT3U(db->db_offset, ==, blkid << bs);
547 	ASSERT3U(db->db_size, ==, 1 << bs);
548 	ASSERT(blkid != 0);
549 
550 	zap_leaf_t *l = dmu_buf_get_user(db);
551 
552 	if (l == NULL)
553 		l = zap_open_leaf(blkid, db);
554 
555 	rw_enter(&l->l_rwlock, lt);
556 	/*
557 	 * Must lock before dirtying, otherwise zap_leaf_phys(l) could change,
558 	 * causing ASSERT below to fail.
559 	 */
560 	if (lt == RW_WRITER)
561 		dmu_buf_will_dirty(db, tx);
562 	ASSERT3U(l->l_blkid, ==, blkid);
563 	ASSERT3P(l->l_dbuf, ==, db);
564 	ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_block_type, ==, ZBT_LEAF);
565 	ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_magic, ==, ZAP_LEAF_MAGIC);
566 
567 	*lp = l;
568 	return (0);
569 }
570 
571 static int
572 zap_idx_to_blk(zap_t *zap, uint64_t idx, uint64_t *valp)
573 {
574 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
575 
576 	if (zap_f_phys(zap)->zap_ptrtbl.zt_numblks == 0) {
577 		ASSERT3U(idx, <,
578 		    (1ULL << zap_f_phys(zap)->zap_ptrtbl.zt_shift));
579 		*valp = ZAP_EMBEDDED_PTRTBL_ENT(zap, idx);
580 		return (0);
581 	} else {
582 		return (zap_table_load(zap, &zap_f_phys(zap)->zap_ptrtbl,
583 		    idx, valp));
584 	}
585 }
586 
587 static int
588 zap_set_idx_to_blk(zap_t *zap, uint64_t idx, uint64_t blk, dmu_tx_t *tx)
589 {
590 	ASSERT(tx != NULL);
591 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
592 
593 	if (zap_f_phys(zap)->zap_ptrtbl.zt_blk == 0) {
594 		ZAP_EMBEDDED_PTRTBL_ENT(zap, idx) = blk;
595 		return (0);
596 	} else {
597 		return (zap_table_store(zap, &zap_f_phys(zap)->zap_ptrtbl,
598 		    idx, blk, tx));
599 	}
600 }
601 
602 static int
603 zap_deref_leaf(zap_t *zap, uint64_t h, dmu_tx_t *tx, krw_t lt, zap_leaf_t **lp)
604 {
605 	uint64_t blk;
606 
607 	ASSERT(zap->zap_dbuf == NULL ||
608 	    zap_f_phys(zap) == zap->zap_dbuf->db_data);
609 
610 	/* Reality check for corrupt zap objects (leaf or header). */
611 	if ((zap_f_phys(zap)->zap_block_type != ZBT_LEAF &&
612 	    zap_f_phys(zap)->zap_block_type != ZBT_HEADER) ||
613 	    zap_f_phys(zap)->zap_magic != ZAP_MAGIC) {
614 		return (SET_ERROR(EIO));
615 	}
616 
617 	uint64_t idx = ZAP_HASH_IDX(h, zap_f_phys(zap)->zap_ptrtbl.zt_shift);
618 	int err = zap_idx_to_blk(zap, idx, &blk);
619 	if (err != 0)
620 		return (err);
621 	err = zap_get_leaf_byblk(zap, blk, tx, lt, lp);
622 
623 	ASSERT(err ||
624 	    ZAP_HASH_IDX(h, zap_leaf_phys(*lp)->l_hdr.lh_prefix_len) ==
625 	    zap_leaf_phys(*lp)->l_hdr.lh_prefix);
626 	return (err);
627 }
628 
629 static int
630 zap_expand_leaf(zap_name_t *zn, zap_leaf_t *l,
631     void *tag, dmu_tx_t *tx, zap_leaf_t **lp)
632 {
633 	zap_t *zap = zn->zn_zap;
634 	uint64_t hash = zn->zn_hash;
635 	int err;
636 	int old_prefix_len = zap_leaf_phys(l)->l_hdr.lh_prefix_len;
637 
638 	ASSERT3U(old_prefix_len, <=, zap_f_phys(zap)->zap_ptrtbl.zt_shift);
639 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
640 
641 	ASSERT3U(ZAP_HASH_IDX(hash, old_prefix_len), ==,
642 	    zap_leaf_phys(l)->l_hdr.lh_prefix);
643 
644 	if (zap_tryupgradedir(zap, tx) == 0 ||
645 	    old_prefix_len == zap_f_phys(zap)->zap_ptrtbl.zt_shift) {
646 		/* We failed to upgrade, or need to grow the pointer table */
647 		objset_t *os = zap->zap_objset;
648 		uint64_t object = zap->zap_object;
649 
650 		zap_put_leaf(l);
651 		zap_unlockdir(zap, tag);
652 		err = zap_lockdir(os, object, tx, RW_WRITER,
653 		    FALSE, FALSE, tag, &zn->zn_zap);
654 		zap = zn->zn_zap;
655 		if (err != 0)
656 			return (err);
657 		ASSERT(!zap->zap_ismicro);
658 
659 		while (old_prefix_len ==
660 		    zap_f_phys(zap)->zap_ptrtbl.zt_shift) {
661 			err = zap_grow_ptrtbl(zap, tx);
662 			if (err != 0)
663 				return (err);
664 		}
665 
666 		err = zap_deref_leaf(zap, hash, tx, RW_WRITER, &l);
667 		if (err != 0)
668 			return (err);
669 
670 		if (zap_leaf_phys(l)->l_hdr.lh_prefix_len != old_prefix_len) {
671 			/* it split while our locks were down */
672 			*lp = l;
673 			return (0);
674 		}
675 	}
676 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
677 	ASSERT3U(old_prefix_len, <, zap_f_phys(zap)->zap_ptrtbl.zt_shift);
678 	ASSERT3U(ZAP_HASH_IDX(hash, old_prefix_len), ==,
679 	    zap_leaf_phys(l)->l_hdr.lh_prefix);
680 
681 	int prefix_diff = zap_f_phys(zap)->zap_ptrtbl.zt_shift -
682 	    (old_prefix_len + 1);
683 	uint64_t sibling =
684 	    (ZAP_HASH_IDX(hash, old_prefix_len + 1) | 1) << prefix_diff;
685 
686 	/* check for i/o errors before doing zap_leaf_split */
687 	for (int i = 0; i < (1ULL << prefix_diff); i++) {
688 		uint64_t blk;
689 		err = zap_idx_to_blk(zap, sibling + i, &blk);
690 		if (err != 0)
691 			return (err);
692 		ASSERT3U(blk, ==, l->l_blkid);
693 	}
694 
695 	zap_leaf_t *nl = zap_create_leaf(zap, tx);
696 	zap_leaf_split(l, nl, zap->zap_normflags != 0);
697 
698 	/* set sibling pointers */
699 	for (int i = 0; i < (1ULL << prefix_diff); i++) {
700 		err = zap_set_idx_to_blk(zap, sibling + i, nl->l_blkid, tx);
701 		ASSERT0(err); /* we checked for i/o errors above */
702 	}
703 
704 	ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_prefix_len, >, 0);
705 
706 	if (hash & (1ULL << (64 - zap_leaf_phys(l)->l_hdr.lh_prefix_len))) {
707 		/* we want the sibling */
708 		zap_put_leaf(l);
709 		*lp = nl;
710 	} else {
711 		zap_put_leaf(nl);
712 		*lp = l;
713 	}
714 
715 	return (0);
716 }
717 
718 static void
719 zap_put_leaf_maybe_grow_ptrtbl(zap_name_t *zn, zap_leaf_t *l,
720     void *tag, dmu_tx_t *tx)
721 {
722 	zap_t *zap = zn->zn_zap;
723 	int shift = zap_f_phys(zap)->zap_ptrtbl.zt_shift;
724 	int leaffull = (zap_leaf_phys(l)->l_hdr.lh_prefix_len == shift &&
725 	    zap_leaf_phys(l)->l_hdr.lh_nfree < ZAP_LEAF_LOW_WATER);
726 
727 	zap_put_leaf(l);
728 
729 	if (leaffull || zap_f_phys(zap)->zap_ptrtbl.zt_nextblk) {
730 		/*
731 		 * We are in the middle of growing the pointer table, or
732 		 * this leaf will soon make us grow it.
733 		 */
734 		if (zap_tryupgradedir(zap, tx) == 0) {
735 			objset_t *os = zap->zap_objset;
736 			uint64_t zapobj = zap->zap_object;
737 
738 			zap_unlockdir(zap, tag);
739 			int err = zap_lockdir(os, zapobj, tx,
740 			    RW_WRITER, FALSE, FALSE, tag, &zn->zn_zap);
741 			zap = zn->zn_zap;
742 			if (err != 0)
743 				return;
744 		}
745 
746 		/* could have finished growing while our locks were down */
747 		if (zap_f_phys(zap)->zap_ptrtbl.zt_shift == shift)
748 			(void) zap_grow_ptrtbl(zap, tx);
749 	}
750 }
751 
752 static int
753 fzap_checkname(zap_name_t *zn)
754 {
755 	if (zn->zn_key_orig_numints * zn->zn_key_intlen > ZAP_MAXNAMELEN)
756 		return (SET_ERROR(ENAMETOOLONG));
757 	return (0);
758 }
759 
760 static int
761 fzap_checksize(uint64_t integer_size, uint64_t num_integers)
762 {
763 	/* Only integer sizes supported by C */
764 	switch (integer_size) {
765 	case 1:
766 	case 2:
767 	case 4:
768 	case 8:
769 		break;
770 	default:
771 		return (SET_ERROR(EINVAL));
772 	}
773 
774 	if (integer_size * num_integers > ZAP_MAXVALUELEN)
775 		return (SET_ERROR(E2BIG));
776 
777 	return (0);
778 }
779 
780 static int
781 fzap_check(zap_name_t *zn, uint64_t integer_size, uint64_t num_integers)
782 {
783 	int err = fzap_checkname(zn);
784 	if (err != 0)
785 		return (err);
786 	return (fzap_checksize(integer_size, num_integers));
787 }
788 
789 /*
790  * Routines for manipulating attributes.
791  */
792 int
793 fzap_lookup(zap_name_t *zn,
794     uint64_t integer_size, uint64_t num_integers, void *buf,
795     char *realname, int rn_len, boolean_t *ncp)
796 {
797 	zap_leaf_t *l;
798 	zap_entry_handle_t zeh;
799 
800 	int err = fzap_checkname(zn);
801 	if (err != 0)
802 		return (err);
803 
804 	err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, NULL, RW_READER, &l);
805 	if (err != 0)
806 		return (err);
807 	err = zap_leaf_lookup(l, zn, &zeh);
808 	if (err == 0) {
809 		if ((err = fzap_checksize(integer_size, num_integers)) != 0) {
810 			zap_put_leaf(l);
811 			return (err);
812 		}
813 
814 		err = zap_entry_read(&zeh, integer_size, num_integers, buf);
815 		(void) zap_entry_read_name(zn->zn_zap, &zeh, rn_len, realname);
816 		if (ncp) {
817 			*ncp = zap_entry_normalization_conflict(&zeh,
818 			    zn, NULL, zn->zn_zap);
819 		}
820 	}
821 
822 	zap_put_leaf(l);
823 	return (err);
824 }
825 
826 int
827 fzap_add_cd(zap_name_t *zn,
828     uint64_t integer_size, uint64_t num_integers,
829     const void *val, uint32_t cd, void *tag, dmu_tx_t *tx)
830 {
831 	zap_leaf_t *l;
832 	int err;
833 	zap_entry_handle_t zeh;
834 	zap_t *zap = zn->zn_zap;
835 
836 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
837 	ASSERT(!zap->zap_ismicro);
838 	ASSERT(fzap_check(zn, integer_size, num_integers) == 0);
839 
840 	err = zap_deref_leaf(zap, zn->zn_hash, tx, RW_WRITER, &l);
841 	if (err != 0)
842 		return (err);
843 retry:
844 	err = zap_leaf_lookup(l, zn, &zeh);
845 	if (err == 0) {
846 		err = SET_ERROR(EEXIST);
847 		goto out;
848 	}
849 	if (err != ENOENT)
850 		goto out;
851 
852 	err = zap_entry_create(l, zn, cd,
853 	    integer_size, num_integers, val, &zeh);
854 
855 	if (err == 0) {
856 		zap_increment_num_entries(zap, 1, tx);
857 	} else if (err == EAGAIN) {
858 		err = zap_expand_leaf(zn, l, tag, tx, &l);
859 		zap = zn->zn_zap;	/* zap_expand_leaf() may change zap */
860 		if (err == 0) {
861 			goto retry;
862 		} else if (err == ENOSPC) {
863 			/*
864 			 * If we failed to expand the leaf, then bailout
865 			 * as there is no point trying
866 			 * zap_put_leaf_maybe_grow_ptrtbl().
867 			 */
868 			return (err);
869 		}
870 	}
871 
872 out:
873 	if (zap != NULL)
874 		zap_put_leaf_maybe_grow_ptrtbl(zn, l, tag, tx);
875 	return (err);
876 }
877 
878 int
879 fzap_add(zap_name_t *zn,
880     uint64_t integer_size, uint64_t num_integers,
881     const void *val, void *tag, dmu_tx_t *tx)
882 {
883 	int err = fzap_check(zn, integer_size, num_integers);
884 	if (err != 0)
885 		return (err);
886 
887 	return (fzap_add_cd(zn, integer_size, num_integers,
888 	    val, ZAP_NEED_CD, tag, tx));
889 }
890 
891 int
892 fzap_update(zap_name_t *zn,
893     int integer_size, uint64_t num_integers, const void *val,
894     void *tag, dmu_tx_t *tx)
895 {
896 	zap_leaf_t *l;
897 	int err;
898 	boolean_t create;
899 	zap_entry_handle_t zeh;
900 	zap_t *zap = zn->zn_zap;
901 
902 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
903 	err = fzap_check(zn, integer_size, num_integers);
904 	if (err != 0)
905 		return (err);
906 
907 	err = zap_deref_leaf(zap, zn->zn_hash, tx, RW_WRITER, &l);
908 	if (err != 0)
909 		return (err);
910 retry:
911 	err = zap_leaf_lookup(l, zn, &zeh);
912 	create = (err == ENOENT);
913 	ASSERT(err == 0 || err == ENOENT);
914 
915 	if (create) {
916 		err = zap_entry_create(l, zn, ZAP_NEED_CD,
917 		    integer_size, num_integers, val, &zeh);
918 		if (err == 0)
919 			zap_increment_num_entries(zap, 1, tx);
920 	} else {
921 		err = zap_entry_update(&zeh, integer_size, num_integers, val);
922 	}
923 
924 	if (err == EAGAIN) {
925 		err = zap_expand_leaf(zn, l, tag, tx, &l);
926 		zap = zn->zn_zap;	/* zap_expand_leaf() may change zap */
927 		if (err == 0)
928 			goto retry;
929 	}
930 
931 	if (zap != NULL)
932 		zap_put_leaf_maybe_grow_ptrtbl(zn, l, tag, tx);
933 	return (err);
934 }
935 
936 int
937 fzap_length(zap_name_t *zn,
938     uint64_t *integer_size, uint64_t *num_integers)
939 {
940 	zap_leaf_t *l;
941 	int err;
942 	zap_entry_handle_t zeh;
943 
944 	err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, NULL, RW_READER, &l);
945 	if (err != 0)
946 		return (err);
947 	err = zap_leaf_lookup(l, zn, &zeh);
948 	if (err != 0)
949 		goto out;
950 
951 	if (integer_size != 0)
952 		*integer_size = zeh.zeh_integer_size;
953 	if (num_integers != 0)
954 		*num_integers = zeh.zeh_num_integers;
955 out:
956 	zap_put_leaf(l);
957 	return (err);
958 }
959 
960 int
961 fzap_remove(zap_name_t *zn, dmu_tx_t *tx)
962 {
963 	zap_leaf_t *l;
964 	int err;
965 	zap_entry_handle_t zeh;
966 
967 	err = zap_deref_leaf(zn->zn_zap, zn->zn_hash, tx, RW_WRITER, &l);
968 	if (err != 0)
969 		return (err);
970 	err = zap_leaf_lookup(l, zn, &zeh);
971 	if (err == 0) {
972 		zap_entry_remove(&zeh);
973 		zap_increment_num_entries(zn->zn_zap, -1, tx);
974 	}
975 	zap_put_leaf(l);
976 	return (err);
977 }
978 
979 void
980 fzap_prefetch(zap_name_t *zn)
981 {
982 	uint64_t blk;
983 	zap_t *zap = zn->zn_zap;
984 
985 	uint64_t idx = ZAP_HASH_IDX(zn->zn_hash,
986 	    zap_f_phys(zap)->zap_ptrtbl.zt_shift);
987 	if (zap_idx_to_blk(zap, idx, &blk) != 0)
988 		return;
989 	int bs = FZAP_BLOCK_SHIFT(zap);
990 	dmu_prefetch(zap->zap_objset, zap->zap_object, 0, blk << bs, 1 << bs,
991 	    ZIO_PRIORITY_SYNC_READ);
992 }
993 
994 /*
995  * Helper functions for consumers.
996  */
997 
998 uint64_t
999 zap_create_link(objset_t *os, dmu_object_type_t ot, uint64_t parent_obj,
1000     const char *name, dmu_tx_t *tx)
1001 {
1002 	return (zap_create_link_dnsize(os, ot, parent_obj, name, 0, tx));
1003 }
1004 
1005 uint64_t
1006 zap_create_link_dnsize(objset_t *os, dmu_object_type_t ot, uint64_t parent_obj,
1007     const char *name, int dnodesize, dmu_tx_t *tx)
1008 {
1009 	uint64_t new_obj;
1010 
1011 	new_obj = zap_create_dnsize(os, ot, DMU_OT_NONE, 0, dnodesize, tx);
1012 	VERIFY(new_obj != 0);
1013 	VERIFY0(zap_add(os, parent_obj, name, sizeof (uint64_t), 1, &new_obj,
1014 	    tx));
1015 
1016 	return (new_obj);
1017 }
1018 
1019 int
1020 zap_value_search(objset_t *os, uint64_t zapobj, uint64_t value, uint64_t mask,
1021     char *name)
1022 {
1023 	zap_cursor_t zc;
1024 	int err;
1025 
1026 	if (mask == 0)
1027 		mask = -1ULL;
1028 
1029 	zap_attribute_t *za = kmem_alloc(sizeof (*za), KM_SLEEP);
1030 	for (zap_cursor_init(&zc, os, zapobj);
1031 	    (err = zap_cursor_retrieve(&zc, za)) == 0;
1032 	    zap_cursor_advance(&zc)) {
1033 		if ((za->za_first_integer & mask) == (value & mask)) {
1034 			(void) strlcpy(name, za->za_name, MAXNAMELEN);
1035 			break;
1036 		}
1037 	}
1038 	zap_cursor_fini(&zc);
1039 	kmem_free(za, sizeof (*za));
1040 	return (err);
1041 }
1042 
1043 int
1044 zap_join(objset_t *os, uint64_t fromobj, uint64_t intoobj, dmu_tx_t *tx)
1045 {
1046 	zap_cursor_t zc;
1047 	int err = 0;
1048 
1049 	zap_attribute_t *za = kmem_alloc(sizeof (*za), KM_SLEEP);
1050 	for (zap_cursor_init(&zc, os, fromobj);
1051 	    zap_cursor_retrieve(&zc, za) == 0;
1052 	    (void) zap_cursor_advance(&zc)) {
1053 		if (za->za_integer_length != 8 || za->za_num_integers != 1) {
1054 			err = SET_ERROR(EINVAL);
1055 			break;
1056 		}
1057 		err = zap_add(os, intoobj, za->za_name,
1058 		    8, 1, &za->za_first_integer, tx);
1059 		if (err != 0)
1060 			break;
1061 	}
1062 	zap_cursor_fini(&zc);
1063 	kmem_free(za, sizeof (*za));
1064 	return (err);
1065 }
1066 
1067 int
1068 zap_join_key(objset_t *os, uint64_t fromobj, uint64_t intoobj,
1069     uint64_t value, dmu_tx_t *tx)
1070 {
1071 	zap_cursor_t zc;
1072 	int err = 0;
1073 
1074 	zap_attribute_t *za = kmem_alloc(sizeof (*za), KM_SLEEP);
1075 	for (zap_cursor_init(&zc, os, fromobj);
1076 	    zap_cursor_retrieve(&zc, za) == 0;
1077 	    (void) zap_cursor_advance(&zc)) {
1078 		if (za->za_integer_length != 8 || za->za_num_integers != 1) {
1079 			err = SET_ERROR(EINVAL);
1080 			break;
1081 		}
1082 		err = zap_add(os, intoobj, za->za_name,
1083 		    8, 1, &value, tx);
1084 		if (err != 0)
1085 			break;
1086 	}
1087 	zap_cursor_fini(&zc);
1088 	kmem_free(za, sizeof (*za));
1089 	return (err);
1090 }
1091 
1092 int
1093 zap_join_increment(objset_t *os, uint64_t fromobj, uint64_t intoobj,
1094     dmu_tx_t *tx)
1095 {
1096 	zap_cursor_t zc;
1097 	int err = 0;
1098 
1099 	zap_attribute_t *za = kmem_alloc(sizeof (*za), KM_SLEEP);
1100 	for (zap_cursor_init(&zc, os, fromobj);
1101 	    zap_cursor_retrieve(&zc, za) == 0;
1102 	    (void) zap_cursor_advance(&zc)) {
1103 		uint64_t delta = 0;
1104 
1105 		if (za->za_integer_length != 8 || za->za_num_integers != 1) {
1106 			err = SET_ERROR(EINVAL);
1107 			break;
1108 		}
1109 
1110 		err = zap_lookup(os, intoobj, za->za_name, 8, 1, &delta);
1111 		if (err != 0 && err != ENOENT)
1112 			break;
1113 		delta += za->za_first_integer;
1114 		err = zap_update(os, intoobj, za->za_name, 8, 1, &delta, tx);
1115 		if (err != 0)
1116 			break;
1117 	}
1118 	zap_cursor_fini(&zc);
1119 	kmem_free(za, sizeof (*za));
1120 	return (err);
1121 }
1122 
1123 int
1124 zap_add_int(objset_t *os, uint64_t obj, uint64_t value, dmu_tx_t *tx)
1125 {
1126 	char name[20];
1127 
1128 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)value);
1129 	return (zap_add(os, obj, name, 8, 1, &value, tx));
1130 }
1131 
1132 int
1133 zap_remove_int(objset_t *os, uint64_t obj, uint64_t value, dmu_tx_t *tx)
1134 {
1135 	char name[20];
1136 
1137 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)value);
1138 	return (zap_remove(os, obj, name, tx));
1139 }
1140 
1141 int
1142 zap_lookup_int(objset_t *os, uint64_t obj, uint64_t value)
1143 {
1144 	char name[20];
1145 
1146 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)value);
1147 	return (zap_lookup(os, obj, name, 8, 1, &value));
1148 }
1149 
1150 int
1151 zap_add_int_key(objset_t *os, uint64_t obj,
1152     uint64_t key, uint64_t value, dmu_tx_t *tx)
1153 {
1154 	char name[20];
1155 
1156 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)key);
1157 	return (zap_add(os, obj, name, 8, 1, &value, tx));
1158 }
1159 
1160 int
1161 zap_update_int_key(objset_t *os, uint64_t obj,
1162     uint64_t key, uint64_t value, dmu_tx_t *tx)
1163 {
1164 	char name[20];
1165 
1166 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)key);
1167 	return (zap_update(os, obj, name, 8, 1, &value, tx));
1168 }
1169 
1170 int
1171 zap_lookup_int_key(objset_t *os, uint64_t obj, uint64_t key, uint64_t *valuep)
1172 {
1173 	char name[20];
1174 
1175 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)key);
1176 	return (zap_lookup(os, obj, name, 8, 1, valuep));
1177 }
1178 
1179 int
1180 zap_increment(objset_t *os, uint64_t obj, const char *name, int64_t delta,
1181     dmu_tx_t *tx)
1182 {
1183 	uint64_t value = 0;
1184 
1185 	if (delta == 0)
1186 		return (0);
1187 
1188 	int err = zap_lookup(os, obj, name, 8, 1, &value);
1189 	if (err != 0 && err != ENOENT)
1190 		return (err);
1191 	value += delta;
1192 	if (value == 0)
1193 		err = zap_remove(os, obj, name, tx);
1194 	else
1195 		err = zap_update(os, obj, name, 8, 1, &value, tx);
1196 	return (err);
1197 }
1198 
1199 int
1200 zap_increment_int(objset_t *os, uint64_t obj, uint64_t key, int64_t delta,
1201     dmu_tx_t *tx)
1202 {
1203 	char name[20];
1204 
1205 	(void) snprintf(name, sizeof (name), "%llx", (longlong_t)key);
1206 	return (zap_increment(os, obj, name, delta, tx));
1207 }
1208 
1209 /*
1210  * Routines for iterating over the attributes.
1211  */
1212 
1213 int
1214 fzap_cursor_retrieve(zap_t *zap, zap_cursor_t *zc, zap_attribute_t *za)
1215 {
1216 	int err = ENOENT;
1217 	zap_entry_handle_t zeh;
1218 	zap_leaf_t *l;
1219 
1220 	/* retrieve the next entry at or after zc_hash/zc_cd */
1221 	/* if no entry, return ENOENT */
1222 
1223 	/*
1224 	 * If we are reading from the beginning, we're almost certain to
1225 	 * iterate over the entire ZAP object.  If there are multiple leaf
1226 	 * blocks (freeblk > 2), prefetch the whole object (up to
1227 	 * dmu_prefetch_max bytes), so that we read the leaf blocks
1228 	 * concurrently. (Unless noprefetch was requested via
1229 	 * zap_cursor_init_noprefetch()).
1230 	 */
1231 	if (zc->zc_hash == 0 && zap_iterate_prefetch &&
1232 	    zc->zc_prefetch && zap_f_phys(zap)->zap_freeblk > 2) {
1233 		dmu_prefetch(zc->zc_objset, zc->zc_zapobj, 0, 0,
1234 		    zap_f_phys(zap)->zap_freeblk << FZAP_BLOCK_SHIFT(zap),
1235 		    ZIO_PRIORITY_ASYNC_READ);
1236 	}
1237 
1238 	if (zc->zc_leaf &&
1239 	    (ZAP_HASH_IDX(zc->zc_hash,
1240 	    zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_prefix_len) !=
1241 	    zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_prefix)) {
1242 		rw_enter(&zc->zc_leaf->l_rwlock, RW_READER);
1243 		zap_put_leaf(zc->zc_leaf);
1244 		zc->zc_leaf = NULL;
1245 	}
1246 
1247 again:
1248 	if (zc->zc_leaf == NULL) {
1249 		err = zap_deref_leaf(zap, zc->zc_hash, NULL, RW_READER,
1250 		    &zc->zc_leaf);
1251 		if (err != 0)
1252 			return (err);
1253 	} else {
1254 		rw_enter(&zc->zc_leaf->l_rwlock, RW_READER);
1255 	}
1256 	l = zc->zc_leaf;
1257 
1258 	err = zap_leaf_lookup_closest(l, zc->zc_hash, zc->zc_cd, &zeh);
1259 
1260 	if (err == ENOENT) {
1261 		if (zap_leaf_phys(l)->l_hdr.lh_prefix_len == 0) {
1262 			zc->zc_hash = -1ULL;
1263 			zc->zc_cd = 0;
1264 		} else {
1265 			uint64_t nocare = (1ULL <<
1266 			    (64 - zap_leaf_phys(l)->l_hdr.lh_prefix_len)) - 1;
1267 
1268 			zc->zc_hash = (zc->zc_hash & ~nocare) + nocare + 1;
1269 			zc->zc_cd = 0;
1270 
1271 			if (zc->zc_hash == 0) {
1272 				zc->zc_hash = -1ULL;
1273 			} else {
1274 				zap_put_leaf(zc->zc_leaf);
1275 				zc->zc_leaf = NULL;
1276 				goto again;
1277 			}
1278 		}
1279 	}
1280 
1281 	if (err == 0) {
1282 		zc->zc_hash = zeh.zeh_hash;
1283 		zc->zc_cd = zeh.zeh_cd;
1284 		za->za_integer_length = zeh.zeh_integer_size;
1285 		za->za_num_integers = zeh.zeh_num_integers;
1286 		if (zeh.zeh_num_integers == 0) {
1287 			za->za_first_integer = 0;
1288 		} else {
1289 			err = zap_entry_read(&zeh, 8, 1, &za->za_first_integer);
1290 			ASSERT(err == 0 || err == EOVERFLOW);
1291 		}
1292 		err = zap_entry_read_name(zap, &zeh,
1293 		    sizeof (za->za_name), za->za_name);
1294 		ASSERT(err == 0);
1295 
1296 		za->za_normalization_conflict =
1297 		    zap_entry_normalization_conflict(&zeh,
1298 		    NULL, za->za_name, zap);
1299 	}
1300 	rw_exit(&zc->zc_leaf->l_rwlock);
1301 	return (err);
1302 }
1303 
1304 static void
1305 zap_stats_ptrtbl(zap_t *zap, uint64_t *tbl, int len, zap_stats_t *zs)
1306 {
1307 	uint64_t lastblk = 0;
1308 
1309 	/*
1310 	 * NB: if a leaf has more pointers than an entire ptrtbl block
1311 	 * can hold, then it'll be accounted for more than once, since
1312 	 * we won't have lastblk.
1313 	 */
1314 	for (int i = 0; i < len; i++) {
1315 		zap_leaf_t *l;
1316 
1317 		if (tbl[i] == lastblk)
1318 			continue;
1319 		lastblk = tbl[i];
1320 
1321 		int err = zap_get_leaf_byblk(zap, tbl[i], NULL, RW_READER, &l);
1322 		if (err == 0) {
1323 			zap_leaf_stats(zap, l, zs);
1324 			zap_put_leaf(l);
1325 		}
1326 	}
1327 }
1328 
1329 void
1330 fzap_get_stats(zap_t *zap, zap_stats_t *zs)
1331 {
1332 	int bs = FZAP_BLOCK_SHIFT(zap);
1333 	zs->zs_blocksize = 1ULL << bs;
1334 
1335 	/*
1336 	 * Set zap_phys_t fields
1337 	 */
1338 	zs->zs_num_leafs = zap_f_phys(zap)->zap_num_leafs;
1339 	zs->zs_num_entries = zap_f_phys(zap)->zap_num_entries;
1340 	zs->zs_num_blocks = zap_f_phys(zap)->zap_freeblk;
1341 	zs->zs_block_type = zap_f_phys(zap)->zap_block_type;
1342 	zs->zs_magic = zap_f_phys(zap)->zap_magic;
1343 	zs->zs_salt = zap_f_phys(zap)->zap_salt;
1344 
1345 	/*
1346 	 * Set zap_ptrtbl fields
1347 	 */
1348 	zs->zs_ptrtbl_len = 1ULL << zap_f_phys(zap)->zap_ptrtbl.zt_shift;
1349 	zs->zs_ptrtbl_nextblk = zap_f_phys(zap)->zap_ptrtbl.zt_nextblk;
1350 	zs->zs_ptrtbl_blks_copied =
1351 	    zap_f_phys(zap)->zap_ptrtbl.zt_blks_copied;
1352 	zs->zs_ptrtbl_zt_blk = zap_f_phys(zap)->zap_ptrtbl.zt_blk;
1353 	zs->zs_ptrtbl_zt_numblks = zap_f_phys(zap)->zap_ptrtbl.zt_numblks;
1354 	zs->zs_ptrtbl_zt_shift = zap_f_phys(zap)->zap_ptrtbl.zt_shift;
1355 
1356 	if (zap_f_phys(zap)->zap_ptrtbl.zt_numblks == 0) {
1357 		/* the ptrtbl is entirely in the header block. */
1358 		zap_stats_ptrtbl(zap, &ZAP_EMBEDDED_PTRTBL_ENT(zap, 0),
1359 		    1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap), zs);
1360 	} else {
1361 		dmu_prefetch(zap->zap_objset, zap->zap_object, 0,
1362 		    zap_f_phys(zap)->zap_ptrtbl.zt_blk << bs,
1363 		    zap_f_phys(zap)->zap_ptrtbl.zt_numblks << bs,
1364 		    ZIO_PRIORITY_SYNC_READ);
1365 
1366 		for (int b = 0; b < zap_f_phys(zap)->zap_ptrtbl.zt_numblks;
1367 		    b++) {
1368 			dmu_buf_t *db;
1369 			int err;
1370 
1371 			err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
1372 			    (zap_f_phys(zap)->zap_ptrtbl.zt_blk + b) << bs,
1373 			    FTAG, &db, DMU_READ_NO_PREFETCH);
1374 			if (err == 0) {
1375 				zap_stats_ptrtbl(zap, db->db_data,
1376 				    1<<(bs-3), zs);
1377 				dmu_buf_rele(db, FTAG);
1378 			}
1379 		}
1380 	}
1381 }
1382 
1383 /* BEGIN CSTYLED */
1384 ZFS_MODULE_PARAM(zfs, , zap_iterate_prefetch, INT, ZMOD_RW,
1385 	"When iterating ZAP object, prefetch it");
1386 /* END CSTYLED */
1387