xref: /titanic_51/usr/src/uts/common/fs/zfs/zap.c (revision 5aefb6555731130ca4fd295960123d71f2d21fe8)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 
29 /*
30  * This file contains the top half of the zfs directory structure
31  * implementation. The bottom half is in zap_leaf.c.
32  *
33  * The zdir is an extendable hash data structure. There is a table of
34  * pointers to buckets (zap_t->zd_data->zd_leafs). The buckets are
35  * each a constant size and hold a variable number of directory entries.
36  * The buckets (aka "leaf nodes") are implemented in zap_leaf.c.
37  *
38  * The pointer table holds a power of 2 number of pointers.
39  * (1<<zap_t->zd_data->zd_phys->zd_prefix_len).  The bucket pointed to
40  * by the pointer at index i in the table holds entries whose hash value
41  * has a zd_prefix_len - bit prefix
42  */
43 
44 #include <sys/spa.h>
45 #include <sys/dmu.h>
46 #include <sys/zfs_context.h>
47 #include <sys/zap.h>
48 #include <sys/refcount.h>
49 #include <sys/zap_impl.h>
50 #include <sys/zap_leaf.h>
51 
52 int fzap_default_block_shift = 14; /* 16k blocksize */
53 
54 static void zap_leaf_pageout(dmu_buf_t *db, void *vl);
55 static uint64_t zap_allocate_blocks(zap_t *zap, int nblocks);
56 
57 
58 void
59 fzap_byteswap(void *vbuf, size_t size)
60 {
61 	uint64_t block_type;
62 
63 	block_type = *(uint64_t *)vbuf;
64 
65 	switch (block_type) {
66 	case ZBT_LEAF:
67 	case BSWAP_64(ZBT_LEAF):
68 		zap_leaf_byteswap(vbuf, size);
69 		return;
70 	case ZBT_HEADER:
71 	case BSWAP_64(ZBT_HEADER):
72 	default:
73 		/* it's a ptrtbl block */
74 		byteswap_uint64_array(vbuf, size);
75 		return;
76 	}
77 }
78 
79 void
80 fzap_upgrade(zap_t *zap, dmu_tx_t *tx)
81 {
82 	dmu_buf_t *db;
83 	zap_leaf_t *l;
84 	int i;
85 	zap_phys_t *zp;
86 
87 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
88 	zap->zap_ismicro = FALSE;
89 
90 	(void) dmu_buf_update_user(zap->zap_dbuf, zap, zap,
91 	    &zap->zap_f.zap_phys, zap_pageout);
92 
93 	mutex_init(&zap->zap_f.zap_num_entries_mtx, 0, 0, 0);
94 	zap->zap_f.zap_block_shift = highbit(zap->zap_dbuf->db_size) - 1;
95 
96 	zp = zap->zap_f.zap_phys;
97 	/*
98 	 * explicitly zero it since it might be coming from an
99 	 * initialized microzap
100 	 */
101 	bzero(zap->zap_dbuf->db_data, zap->zap_dbuf->db_size);
102 	zp->zap_block_type = ZBT_HEADER;
103 	zp->zap_magic = ZAP_MAGIC;
104 
105 	zp->zap_ptrtbl.zt_shift = ZAP_EMBEDDED_PTRTBL_SHIFT(zap);
106 
107 	zp->zap_freeblk = 2;		/* block 1 will be the first leaf */
108 	zp->zap_num_leafs = 1;
109 	zp->zap_num_entries = 0;
110 	zp->zap_salt = zap->zap_salt;
111 
112 	/* block 1 will be the first leaf */
113 	for (i = 0; i < (1<<zp->zap_ptrtbl.zt_shift); i++)
114 		ZAP_EMBEDDED_PTRTBL_ENT(zap, i) = 1;
115 
116 	/*
117 	 * set up block 1 - the first leaf
118 	 */
119 	VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object,
120 	    1<<FZAP_BLOCK_SHIFT(zap), FTAG, &db));
121 	dmu_buf_will_dirty(db, tx);
122 
123 	l = kmem_zalloc(sizeof (zap_leaf_t), KM_SLEEP);
124 	l->l_dbuf = db;
125 	l->l_phys = db->db_data;
126 
127 	zap_leaf_init(l);
128 
129 	kmem_free(l, sizeof (zap_leaf_t));
130 	dmu_buf_rele(db, FTAG);
131 }
132 
133 static int
134 zap_tryupgradedir(zap_t *zap, dmu_tx_t *tx)
135 {
136 	if (RW_WRITE_HELD(&zap->zap_rwlock))
137 		return (1);
138 	if (rw_tryupgrade(&zap->zap_rwlock)) {
139 		dmu_buf_will_dirty(zap->zap_dbuf, tx);
140 		return (1);
141 	}
142 	return (0);
143 }
144 
145 /*
146  * Generic routines for dealing with the pointer & cookie tables.
147  */
148 
149 static int
150 zap_table_grow(zap_t *zap, zap_table_phys_t *tbl,
151     void (*transfer_func)(const uint64_t *src, uint64_t *dst, int n),
152     dmu_tx_t *tx)
153 {
154 	uint64_t b, newblk;
155 	dmu_buf_t *db_old, *db_new;
156 	int err;
157 	int bs = FZAP_BLOCK_SHIFT(zap);
158 	int hepb = 1<<(bs-4);
159 	/* hepb = half the number of entries in a block */
160 
161 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
162 	ASSERT(tbl->zt_blk != 0);
163 	ASSERT(tbl->zt_numblks > 0);
164 
165 	if (tbl->zt_nextblk != 0) {
166 		newblk = tbl->zt_nextblk;
167 	} else {
168 		newblk = zap_allocate_blocks(zap, tbl->zt_numblks * 2);
169 		tbl->zt_nextblk = newblk;
170 		ASSERT3U(tbl->zt_blks_copied, ==, 0);
171 		dmu_prefetch(zap->zap_objset, zap->zap_object,
172 		    tbl->zt_blk << bs, tbl->zt_numblks << bs);
173 	}
174 
175 	/*
176 	 * Copy the ptrtbl from the old to new location.
177 	 */
178 
179 	b = tbl->zt_blks_copied;
180 	err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
181 	    (tbl->zt_blk + b) << bs, FTAG, &db_old);
182 	if (err)
183 		return (err);
184 
185 	/* first half of entries in old[b] go to new[2*b+0] */
186 	VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object,
187 	    (newblk + 2*b+0) << bs, FTAG, &db_new));
188 	dmu_buf_will_dirty(db_new, tx);
189 	transfer_func(db_old->db_data, db_new->db_data, hepb);
190 	dmu_buf_rele(db_new, FTAG);
191 
192 	/* second half of entries in old[b] go to new[2*b+1] */
193 	VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object,
194 	    (newblk + 2*b+1) << bs, FTAG, &db_new));
195 	dmu_buf_will_dirty(db_new, tx);
196 	transfer_func((uint64_t *)db_old->db_data + hepb,
197 	    db_new->db_data, hepb);
198 	dmu_buf_rele(db_new, FTAG);
199 
200 	dmu_buf_rele(db_old, FTAG);
201 
202 	tbl->zt_blks_copied++;
203 
204 	dprintf("copied block %llu of %llu\n",
205 	    tbl->zt_blks_copied, tbl->zt_numblks);
206 
207 	if (tbl->zt_blks_copied == tbl->zt_numblks) {
208 		(void) dmu_free_range(zap->zap_objset, zap->zap_object,
209 		    tbl->zt_blk << bs, tbl->zt_numblks << bs, tx);
210 
211 		tbl->zt_blk = newblk;
212 		tbl->zt_numblks *= 2;
213 		tbl->zt_shift++;
214 		tbl->zt_nextblk = 0;
215 		tbl->zt_blks_copied = 0;
216 
217 		dprintf("finished; numblocks now %llu (%lluk entries)\n",
218 		    tbl->zt_numblks, 1<<(tbl->zt_shift-10));
219 	}
220 
221 	return (0);
222 }
223 
224 static int
225 zap_table_store(zap_t *zap, zap_table_phys_t *tbl, uint64_t idx, uint64_t val,
226     dmu_tx_t *tx)
227 {
228 	int err;
229 	uint64_t blk, off;
230 	int bs = FZAP_BLOCK_SHIFT(zap);
231 	dmu_buf_t *db;
232 
233 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
234 	ASSERT(tbl->zt_blk != 0);
235 
236 	dprintf("storing %llx at index %llx\n", val, idx);
237 
238 	blk = idx >> (bs-3);
239 	off = idx & ((1<<(bs-3))-1);
240 
241 	err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
242 	    (tbl->zt_blk + blk) << bs, FTAG, &db);
243 	if (err)
244 		return (err);
245 	dmu_buf_will_dirty(db, tx);
246 
247 	if (tbl->zt_nextblk != 0) {
248 		uint64_t idx2 = idx * 2;
249 		uint64_t blk2 = idx2 >> (bs-3);
250 		uint64_t off2 = idx2 & ((1<<(bs-3))-1);
251 		dmu_buf_t *db2;
252 
253 		err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
254 		    (tbl->zt_nextblk + blk2) << bs, FTAG, &db2);
255 		if (err) {
256 			dmu_buf_rele(db, FTAG);
257 			return (err);
258 		}
259 		dmu_buf_will_dirty(db2, tx);
260 		((uint64_t *)db2->db_data)[off2] = val;
261 		((uint64_t *)db2->db_data)[off2+1] = val;
262 		dmu_buf_rele(db2, FTAG);
263 	}
264 
265 	((uint64_t *)db->db_data)[off] = val;
266 	dmu_buf_rele(db, FTAG);
267 
268 	return (0);
269 }
270 
271 static int
272 zap_table_load(zap_t *zap, zap_table_phys_t *tbl, uint64_t idx, uint64_t *valp)
273 {
274 	uint64_t blk, off;
275 	int err;
276 	dmu_buf_t *db;
277 	int bs = FZAP_BLOCK_SHIFT(zap);
278 
279 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
280 
281 	blk = idx >> (bs-3);
282 	off = idx & ((1<<(bs-3))-1);
283 
284 	err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
285 	    (tbl->zt_blk + blk) << bs, FTAG, &db);
286 	if (err)
287 		return (err);
288 	*valp = ((uint64_t *)db->db_data)[off];
289 	dmu_buf_rele(db, FTAG);
290 
291 	if (tbl->zt_nextblk != 0) {
292 		/*
293 		 * read the nextblk for the sake of i/o error checking,
294 		 * so that zap_table_load() will catch errors for
295 		 * zap_table_store.
296 		 */
297 		blk = (idx*2) >> (bs-3);
298 
299 		err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
300 		    (tbl->zt_nextblk + blk) << bs, FTAG, &db);
301 		dmu_buf_rele(db, FTAG);
302 	}
303 	return (err);
304 }
305 
306 /*
307  * Routines for growing the ptrtbl.
308  */
309 
310 static void
311 zap_ptrtbl_transfer(const uint64_t *src, uint64_t *dst, int n)
312 {
313 	int i;
314 	for (i = 0; i < n; i++) {
315 		uint64_t lb = src[i];
316 		dst[2*i+0] = lb;
317 		dst[2*i+1] = lb;
318 	}
319 }
320 
321 static int
322 zap_grow_ptrtbl(zap_t *zap, dmu_tx_t *tx)
323 {
324 	/* In case things go horribly wrong. */
325 	if (zap->zap_f.zap_phys->zap_ptrtbl.zt_shift >= ZAP_HASHBITS-2)
326 		return (ENOSPC);
327 
328 	if (zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks == 0) {
329 		/*
330 		 * We are outgrowing the "embedded" ptrtbl (the one
331 		 * stored in the header block).  Give it its own entire
332 		 * block, which will double the size of the ptrtbl.
333 		 */
334 		uint64_t newblk;
335 		dmu_buf_t *db_new;
336 		int err;
337 
338 		ASSERT3U(zap->zap_f.zap_phys->zap_ptrtbl.zt_shift, ==,
339 		    ZAP_EMBEDDED_PTRTBL_SHIFT(zap));
340 		ASSERT3U(zap->zap_f.zap_phys->zap_ptrtbl.zt_blk, ==, 0);
341 
342 		newblk = zap_allocate_blocks(zap, 1);
343 		err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
344 		    newblk << FZAP_BLOCK_SHIFT(zap), FTAG, &db_new);
345 		if (err)
346 			return (err);
347 		dmu_buf_will_dirty(db_new, tx);
348 		zap_ptrtbl_transfer(&ZAP_EMBEDDED_PTRTBL_ENT(zap, 0),
349 		    db_new->db_data, 1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap));
350 		dmu_buf_rele(db_new, FTAG);
351 
352 		zap->zap_f.zap_phys->zap_ptrtbl.zt_blk = newblk;
353 		zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks = 1;
354 		zap->zap_f.zap_phys->zap_ptrtbl.zt_shift++;
355 
356 		ASSERT3U(1ULL << zap->zap_f.zap_phys->zap_ptrtbl.zt_shift, ==,
357 		    zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks <<
358 		    (FZAP_BLOCK_SHIFT(zap)-3));
359 
360 		return (0);
361 	} else {
362 		return (zap_table_grow(zap, &zap->zap_f.zap_phys->zap_ptrtbl,
363 		    zap_ptrtbl_transfer, tx));
364 	}
365 }
366 
367 static void
368 zap_increment_num_entries(zap_t *zap, int delta, dmu_tx_t *tx)
369 {
370 	dmu_buf_will_dirty(zap->zap_dbuf, tx);
371 	mutex_enter(&zap->zap_f.zap_num_entries_mtx);
372 	ASSERT(delta > 0 || zap->zap_f.zap_phys->zap_num_entries >= -delta);
373 	zap->zap_f.zap_phys->zap_num_entries += delta;
374 	mutex_exit(&zap->zap_f.zap_num_entries_mtx);
375 }
376 
377 static uint64_t
378 zap_allocate_blocks(zap_t *zap, int nblocks)
379 {
380 	uint64_t newblk;
381 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
382 	newblk = zap->zap_f.zap_phys->zap_freeblk;
383 	zap->zap_f.zap_phys->zap_freeblk += nblocks;
384 	return (newblk);
385 }
386 
387 static zap_leaf_t *
388 zap_create_leaf(zap_t *zap, dmu_tx_t *tx)
389 {
390 	void *winner;
391 	zap_leaf_t *l = kmem_alloc(sizeof (zap_leaf_t), KM_SLEEP);
392 
393 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
394 
395 	rw_init(&l->l_rwlock, 0, 0, 0);
396 	rw_enter(&l->l_rwlock, RW_WRITER);
397 	l->l_blkid = zap_allocate_blocks(zap, 1);
398 	l->l_dbuf = NULL;
399 	l->l_phys = NULL;
400 
401 	VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object,
402 	    l->l_blkid << FZAP_BLOCK_SHIFT(zap), NULL, &l->l_dbuf));
403 	winner = dmu_buf_set_user(l->l_dbuf, l, &l->l_phys, zap_leaf_pageout);
404 	ASSERT(winner == NULL);
405 	dmu_buf_will_dirty(l->l_dbuf, tx);
406 
407 	zap_leaf_init(l);
408 
409 	zap->zap_f.zap_phys->zap_num_leafs++;
410 
411 	return (l);
412 }
413 
414 int
415 fzap_count(zap_t *zap, uint64_t *count)
416 {
417 	ASSERT(!zap->zap_ismicro);
418 	mutex_enter(&zap->zap_f.zap_num_entries_mtx); /* unnecessary */
419 	*count = zap->zap_f.zap_phys->zap_num_entries;
420 	mutex_exit(&zap->zap_f.zap_num_entries_mtx);
421 	return (0);
422 }
423 
424 /*
425  * Routines for obtaining zap_leaf_t's
426  */
427 
428 void
429 zap_put_leaf(zap_leaf_t *l)
430 {
431 	rw_exit(&l->l_rwlock);
432 	dmu_buf_rele(l->l_dbuf, NULL);
433 }
434 
435 _NOTE(ARGSUSED(0))
436 static void
437 zap_leaf_pageout(dmu_buf_t *db, void *vl)
438 {
439 	zap_leaf_t *l = vl;
440 
441 	rw_destroy(&l->l_rwlock);
442 	kmem_free(l, sizeof (zap_leaf_t));
443 }
444 
445 static zap_leaf_t *
446 zap_open_leaf(uint64_t blkid, dmu_buf_t *db)
447 {
448 	zap_leaf_t *l, *winner;
449 
450 	ASSERT(blkid != 0);
451 
452 	l = kmem_alloc(sizeof (zap_leaf_t), KM_SLEEP);
453 	rw_init(&l->l_rwlock, 0, 0, 0);
454 	rw_enter(&l->l_rwlock, RW_WRITER);
455 	l->l_blkid = blkid;
456 	l->l_bs = highbit(db->db_size)-1;
457 	l->l_dbuf = db;
458 	l->l_phys = NULL;
459 
460 	winner = dmu_buf_set_user(db, l, &l->l_phys, zap_leaf_pageout);
461 
462 	rw_exit(&l->l_rwlock);
463 	if (winner != NULL) {
464 		/* someone else set it first */
465 		zap_leaf_pageout(NULL, l);
466 		l = winner;
467 	}
468 
469 	/*
470 	 * lhr_pad was previously used for the next leaf in the leaf
471 	 * chain.  There should be no chained leafs (as we have removed
472 	 * support for them).
473 	 */
474 	ASSERT3U(l->l_phys->l_hdr.lh_pad1, ==, 0);
475 
476 	/*
477 	 * There should be more hash entries than there can be
478 	 * chunks to put in the hash table
479 	 */
480 	ASSERT3U(ZAP_LEAF_HASH_NUMENTRIES(l), >, ZAP_LEAF_NUMCHUNKS(l) / 3);
481 
482 	/* The chunks should begin at the end of the hash table */
483 	ASSERT3P(&ZAP_LEAF_CHUNK(l, 0), ==,
484 	    &l->l_phys->l_hash[ZAP_LEAF_HASH_NUMENTRIES(l)]);
485 
486 	/* The chunks should end at the end of the block */
487 	ASSERT3U((uintptr_t)&ZAP_LEAF_CHUNK(l, ZAP_LEAF_NUMCHUNKS(l)) -
488 	    (uintptr_t)l->l_phys, ==, l->l_dbuf->db_size);
489 
490 	return (l);
491 }
492 
493 static int
494 zap_get_leaf_byblk(zap_t *zap, uint64_t blkid, dmu_tx_t *tx, krw_t lt,
495     zap_leaf_t **lp)
496 {
497 	dmu_buf_t *db;
498 	zap_leaf_t *l;
499 	int bs = FZAP_BLOCK_SHIFT(zap);
500 	int err;
501 
502 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
503 
504 	err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
505 	    blkid << bs, NULL, &db);
506 	if (err)
507 		return (err);
508 
509 	ASSERT3U(db->db_object, ==, zap->zap_object);
510 	ASSERT3U(db->db_offset, ==, blkid << bs);
511 	ASSERT3U(db->db_size, ==, 1 << bs);
512 	ASSERT(blkid != 0);
513 
514 	l = dmu_buf_get_user(db);
515 
516 	if (l == NULL)
517 		l = zap_open_leaf(blkid, db);
518 
519 	rw_enter(&l->l_rwlock, lt);
520 	/*
521 	 * Must lock before dirtying, otherwise l->l_phys could change,
522 	 * causing ASSERT below to fail.
523 	 */
524 	if (lt == RW_WRITER)
525 		dmu_buf_will_dirty(db, tx);
526 	ASSERT3U(l->l_blkid, ==, blkid);
527 	ASSERT3P(l->l_dbuf, ==, db);
528 	ASSERT3P(l->l_phys, ==, l->l_dbuf->db_data);
529 	ASSERT3U(l->l_phys->l_hdr.lh_block_type, ==, ZBT_LEAF);
530 	ASSERT3U(l->l_phys->l_hdr.lh_magic, ==, ZAP_LEAF_MAGIC);
531 
532 	*lp = l;
533 	return (0);
534 }
535 
536 static int
537 zap_idx_to_blk(zap_t *zap, uint64_t idx, uint64_t *valp)
538 {
539 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
540 
541 	if (zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks == 0) {
542 		ASSERT3U(idx, <,
543 		    (1ULL << zap->zap_f.zap_phys->zap_ptrtbl.zt_shift));
544 		*valp = ZAP_EMBEDDED_PTRTBL_ENT(zap, idx);
545 		return (0);
546 	} else {
547 		return (zap_table_load(zap, &zap->zap_f.zap_phys->zap_ptrtbl,
548 		    idx, valp));
549 	}
550 }
551 
552 static int
553 zap_set_idx_to_blk(zap_t *zap, uint64_t idx, uint64_t blk, dmu_tx_t *tx)
554 {
555 	ASSERT(tx != NULL);
556 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
557 
558 	if (zap->zap_f.zap_phys->zap_ptrtbl.zt_blk == 0) {
559 		ZAP_EMBEDDED_PTRTBL_ENT(zap, idx) = blk;
560 		return (0);
561 	} else {
562 		return (zap_table_store(zap, &zap->zap_f.zap_phys->zap_ptrtbl,
563 		    idx, blk, tx));
564 	}
565 }
566 
567 static int
568 zap_deref_leaf(zap_t *zap, uint64_t h, dmu_tx_t *tx, krw_t lt, zap_leaf_t **lp)
569 {
570 	uint64_t idx, blk;
571 	int err;
572 
573 	ASSERT(zap->zap_dbuf == NULL ||
574 	    zap->zap_f.zap_phys == zap->zap_dbuf->db_data);
575 	ASSERT3U(zap->zap_f.zap_phys->zap_magic, ==, ZAP_MAGIC);
576 	idx = ZAP_HASH_IDX(h, zap->zap_f.zap_phys->zap_ptrtbl.zt_shift);
577 	err = zap_idx_to_blk(zap, idx, &blk);
578 	if (err != 0)
579 		return (err);
580 	err = zap_get_leaf_byblk(zap, blk, tx, lt, lp);
581 
582 	ASSERT(err || ZAP_HASH_IDX(h, (*lp)->l_phys->l_hdr.lh_prefix_len) ==
583 	    (*lp)->l_phys->l_hdr.lh_prefix);
584 	return (err);
585 }
586 
587 static int
588 zap_expand_leaf(zap_t *zap, zap_leaf_t *l, uint64_t hash, dmu_tx_t *tx,
589     zap_leaf_t **lp)
590 {
591 	zap_leaf_t *nl;
592 	int prefix_diff, i, err;
593 	uint64_t sibling;
594 	int old_prefix_len = l->l_phys->l_hdr.lh_prefix_len;
595 
596 	ASSERT3U(old_prefix_len, <=, zap->zap_f.zap_phys->zap_ptrtbl.zt_shift);
597 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
598 
599 	ASSERT3U(ZAP_HASH_IDX(hash, old_prefix_len), ==,
600 	    l->l_phys->l_hdr.lh_prefix);
601 
602 	if (zap_tryupgradedir(zap, tx) == 0 ||
603 	    old_prefix_len == zap->zap_f.zap_phys->zap_ptrtbl.zt_shift) {
604 		/* We failed to upgrade, or need to grow the pointer table */
605 		objset_t *os = zap->zap_objset;
606 		uint64_t object = zap->zap_object;
607 
608 		zap_put_leaf(l);
609 		zap_unlockdir(zap);
610 		err = zap_lockdir(os, object, tx, RW_WRITER, FALSE, &zap);
611 		if (err)
612 			return (err);
613 		ASSERT(!zap->zap_ismicro);
614 
615 		while (old_prefix_len ==
616 		    zap->zap_f.zap_phys->zap_ptrtbl.zt_shift) {
617 			err = zap_grow_ptrtbl(zap, tx);
618 			if (err)
619 				return (err);
620 		}
621 
622 		err = zap_deref_leaf(zap, hash, tx, RW_WRITER, &l);
623 		if (err)
624 			return (err);
625 
626 		if (l->l_phys->l_hdr.lh_prefix_len != old_prefix_len) {
627 			/* it split while our locks were down */
628 			*lp = l;
629 			return (0);
630 		}
631 	}
632 	ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
633 	ASSERT3U(old_prefix_len, <, zap->zap_f.zap_phys->zap_ptrtbl.zt_shift);
634 	ASSERT3U(ZAP_HASH_IDX(hash, old_prefix_len), ==,
635 	    l->l_phys->l_hdr.lh_prefix);
636 
637 	prefix_diff = zap->zap_f.zap_phys->zap_ptrtbl.zt_shift -
638 	    (old_prefix_len + 1);
639 	sibling = (ZAP_HASH_IDX(hash, old_prefix_len + 1) | 1) << prefix_diff;
640 
641 	/* check for i/o errors before doing zap_leaf_split */
642 	for (i = 0; i < (1ULL<<prefix_diff); i++) {
643 		uint64_t blk;
644 		err = zap_idx_to_blk(zap, sibling+i, &blk);
645 		if (err)
646 			return (err);
647 		ASSERT3U(blk, ==, l->l_blkid);
648 	}
649 
650 	nl = zap_create_leaf(zap, tx);
651 	zap_leaf_split(l, nl);
652 
653 	/* set sibling pointers */
654 	for (i = 0; i < (1ULL<<prefix_diff); i++) {
655 		err = zap_set_idx_to_blk(zap, sibling+i, nl->l_blkid, tx);
656 		ASSERT3U(err, ==, 0); /* we checked for i/o errors above */
657 	}
658 
659 	if (hash & (1ULL << (64 - l->l_phys->l_hdr.lh_prefix_len))) {
660 		/* we want the sibling */
661 		zap_put_leaf(l);
662 		*lp = nl;
663 	} else {
664 		zap_put_leaf(nl);
665 		*lp = l;
666 	}
667 
668 	return (0);
669 }
670 
671 static void
672 zap_put_leaf_maybe_grow_ptrtbl(zap_t *zap, zap_leaf_t *l, dmu_tx_t *tx)
673 {
674 	int shift = zap->zap_f.zap_phys->zap_ptrtbl.zt_shift;
675 	int leaffull = (l->l_phys->l_hdr.lh_prefix_len == shift &&
676 	    l->l_phys->l_hdr.lh_nfree < ZAP_LEAF_LOW_WATER);
677 
678 	zap_put_leaf(l);
679 
680 	if (leaffull || zap->zap_f.zap_phys->zap_ptrtbl.zt_nextblk) {
681 		int err;
682 
683 		/*
684 		 * We are in the middle of growing the pointer table, or
685 		 * this leaf will soon make us grow it.
686 		 */
687 		if (zap_tryupgradedir(zap, tx) == 0) {
688 			objset_t *os = zap->zap_objset;
689 			uint64_t zapobj = zap->zap_object;
690 
691 			zap_unlockdir(zap);
692 			err = zap_lockdir(os, zapobj, tx,
693 			    RW_WRITER, FALSE, &zap);
694 			if (err)
695 				return;
696 		}
697 
698 		/* could have finished growing while our locks were down */
699 		if (zap->zap_f.zap_phys->zap_ptrtbl.zt_shift == shift)
700 			(void) zap_grow_ptrtbl(zap, tx);
701 	}
702 }
703 
704 
705 static int
706 fzap_checksize(const char *name, uint64_t integer_size, uint64_t num_integers)
707 {
708 	if (name && strlen(name) > ZAP_MAXNAMELEN)
709 		return (E2BIG);
710 
711 	/* Only integer sizes supported by C */
712 	switch (integer_size) {
713 	case 1:
714 	case 2:
715 	case 4:
716 	case 8:
717 		break;
718 	default:
719 		return (EINVAL);
720 	}
721 
722 	if (integer_size * num_integers > ZAP_MAXVALUELEN)
723 		return (E2BIG);
724 
725 	return (0);
726 }
727 
728 /*
729  * Routines for maniplulating attributes.
730  */
731 int
732 fzap_lookup(zap_t *zap, const char *name,
733     uint64_t integer_size, uint64_t num_integers, void *buf)
734 {
735 	zap_leaf_t *l;
736 	int err;
737 	uint64_t hash;
738 	zap_entry_handle_t zeh;
739 
740 	err = fzap_checksize(name, integer_size, num_integers);
741 	if (err != 0)
742 		return (err);
743 
744 	hash = zap_hash(zap, name);
745 	err = zap_deref_leaf(zap, hash, NULL, RW_READER, &l);
746 	if (err != 0)
747 		return (err);
748 	err = zap_leaf_lookup(l, name, hash, &zeh);
749 	if (err == 0)
750 		err = zap_entry_read(&zeh, integer_size, num_integers, buf);
751 
752 	zap_put_leaf(l);
753 	return (err);
754 }
755 
756 int
757 fzap_add_cd(zap_t *zap, const char *name,
758     uint64_t integer_size, uint64_t num_integers,
759     const void *val, uint32_t cd, dmu_tx_t *tx)
760 {
761 	zap_leaf_t *l;
762 	uint64_t hash;
763 	int err;
764 	zap_entry_handle_t zeh;
765 
766 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
767 	ASSERT(!zap->zap_ismicro);
768 	ASSERT(fzap_checksize(name, integer_size, num_integers) == 0);
769 
770 	hash = zap_hash(zap, name);
771 	err = zap_deref_leaf(zap, hash, tx, RW_WRITER, &l);
772 	if (err != 0)
773 		return (err);
774 retry:
775 	err = zap_leaf_lookup(l, name, hash, &zeh);
776 	if (err == 0) {
777 		err = EEXIST;
778 		goto out;
779 	}
780 	if (err != ENOENT)
781 		goto out;
782 
783 	err = zap_entry_create(l, name, hash, cd,
784 	    integer_size, num_integers, val, &zeh);
785 
786 	if (err == 0) {
787 		zap_increment_num_entries(zap, 1, tx);
788 	} else if (err == EAGAIN) {
789 		err = zap_expand_leaf(zap, l, hash, tx, &l);
790 		if (err == 0)
791 			goto retry;
792 	}
793 
794 out:
795 	zap_put_leaf_maybe_grow_ptrtbl(zap, l, tx);
796 	return (err);
797 }
798 
799 int
800 fzap_add(zap_t *zap, const char *name,
801     uint64_t integer_size, uint64_t num_integers,
802     const void *val, dmu_tx_t *tx)
803 {
804 	int err = fzap_checksize(name, integer_size, num_integers);
805 	if (err != 0)
806 		return (err);
807 
808 	return (fzap_add_cd(zap, name, integer_size, num_integers,
809 	    val, ZAP_MAXCD, tx));
810 }
811 
812 int
813 fzap_update(zap_t *zap, const char *name,
814     int integer_size, uint64_t num_integers, const void *val, dmu_tx_t *tx)
815 {
816 	zap_leaf_t *l;
817 	uint64_t hash;
818 	int err, create;
819 	zap_entry_handle_t zeh;
820 
821 	ASSERT(RW_LOCK_HELD(&zap->zap_rwlock));
822 	err = fzap_checksize(name, integer_size, num_integers);
823 	if (err != 0)
824 		return (err);
825 
826 	hash = zap_hash(zap, name);
827 	err = zap_deref_leaf(zap, hash, tx, RW_WRITER, &l);
828 	if (err != 0)
829 		return (err);
830 retry:
831 	err = zap_leaf_lookup(l, name, hash, &zeh);
832 	create = (err == ENOENT);
833 	ASSERT(err == 0 || err == ENOENT);
834 
835 	/* XXX If this leaf is chained, split it if we can. */
836 
837 	if (create) {
838 		err = zap_entry_create(l, name, hash, ZAP_MAXCD,
839 		    integer_size, num_integers, val, &zeh);
840 		if (err == 0)
841 			zap_increment_num_entries(zap, 1, tx);
842 	} else {
843 		err = zap_entry_update(&zeh, integer_size, num_integers, val);
844 	}
845 
846 	if (err == EAGAIN) {
847 		err = zap_expand_leaf(zap, l, hash, tx, &l);
848 		if (err == 0)
849 			goto retry;
850 	}
851 
852 out:
853 	zap_put_leaf_maybe_grow_ptrtbl(zap, l, tx);
854 	return (err);
855 }
856 
857 int
858 fzap_length(zap_t *zap, const char *name,
859     uint64_t *integer_size, uint64_t *num_integers)
860 {
861 	zap_leaf_t *l;
862 	int err;
863 	uint64_t hash;
864 	zap_entry_handle_t zeh;
865 
866 	hash = zap_hash(zap, name);
867 	err = zap_deref_leaf(zap, hash, NULL, RW_READER, &l);
868 	if (err != 0)
869 		return (err);
870 	err = zap_leaf_lookup(l, name, hash, &zeh);
871 	if (err != 0)
872 		goto out;
873 
874 	if (integer_size)
875 		*integer_size = zeh.zeh_integer_size;
876 	if (num_integers)
877 		*num_integers = zeh.zeh_num_integers;
878 out:
879 	zap_put_leaf(l);
880 	return (err);
881 }
882 
883 int
884 fzap_remove(zap_t *zap, const char *name, dmu_tx_t *tx)
885 {
886 	zap_leaf_t *l;
887 	uint64_t hash;
888 	int err;
889 	zap_entry_handle_t zeh;
890 
891 	hash = zap_hash(zap, name);
892 	err = zap_deref_leaf(zap, hash, tx, RW_WRITER, &l);
893 	if (err != 0)
894 		return (err);
895 	err = zap_leaf_lookup(l, name, hash, &zeh);
896 	if (err == 0) {
897 		zap_entry_remove(&zeh);
898 		zap_increment_num_entries(zap, -1, tx);
899 	}
900 	zap_put_leaf(l);
901 	dprintf("fzap_remove: ds=%p obj=%llu name=%s err=%d\n",
902 	    zap->zap_objset, zap->zap_object, name, err);
903 	return (err);
904 }
905 
906 int
907 zap_value_search(objset_t *os, uint64_t zapobj, uint64_t value, char *name)
908 {
909 	zap_cursor_t zc;
910 	zap_attribute_t *za;
911 	int err;
912 
913 	za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
914 	for (zap_cursor_init(&zc, os, zapobj);
915 	    (err = zap_cursor_retrieve(&zc, za)) == 0;
916 	    zap_cursor_advance(&zc)) {
917 		if (za->za_first_integer == value) {
918 			(void) strcpy(name, za->za_name);
919 			break;
920 		}
921 	}
922 	zap_cursor_fini(&zc);
923 	kmem_free(za, sizeof (zap_attribute_t));
924 	return (err);
925 }
926 
927 
928 /*
929  * Routines for iterating over the attributes.
930  */
931 
932 int
933 fzap_cursor_retrieve(zap_t *zap, zap_cursor_t *zc, zap_attribute_t *za)
934 {
935 	int err = ENOENT;
936 	zap_entry_handle_t zeh;
937 	zap_leaf_t *l;
938 
939 	/* retrieve the next entry at or after zc_hash/zc_cd */
940 	/* if no entry, return ENOENT */
941 
942 	if (zc->zc_leaf &&
943 	    (ZAP_HASH_IDX(zc->zc_hash,
944 	    zc->zc_leaf->l_phys->l_hdr.lh_prefix_len) !=
945 	    zc->zc_leaf->l_phys->l_hdr.lh_prefix)) {
946 		rw_enter(&zc->zc_leaf->l_rwlock, RW_READER);
947 		zap_put_leaf(zc->zc_leaf);
948 		zc->zc_leaf = NULL;
949 	}
950 
951 again:
952 	if (zc->zc_leaf == NULL) {
953 		err = zap_deref_leaf(zap, zc->zc_hash, NULL, RW_READER,
954 		    &zc->zc_leaf);
955 		if (err != 0)
956 			return (err);
957 	} else {
958 		rw_enter(&zc->zc_leaf->l_rwlock, RW_READER);
959 	}
960 	l = zc->zc_leaf;
961 
962 	err = zap_leaf_lookup_closest(l, zc->zc_hash, zc->zc_cd, &zeh);
963 
964 	if (err == ENOENT) {
965 		uint64_t nocare =
966 		    (1ULL << (64 - l->l_phys->l_hdr.lh_prefix_len)) - 1;
967 		zc->zc_hash = (zc->zc_hash & ~nocare) + nocare + 1;
968 		zc->zc_cd = 0;
969 		if (l->l_phys->l_hdr.lh_prefix_len == 0 || zc->zc_hash == 0) {
970 			zc->zc_hash = -1ULL;
971 		} else {
972 			zap_put_leaf(zc->zc_leaf);
973 			zc->zc_leaf = NULL;
974 			goto again;
975 		}
976 	}
977 
978 	if (err == 0) {
979 		zc->zc_hash = zeh.zeh_hash;
980 		zc->zc_cd = zeh.zeh_cd;
981 		za->za_integer_length = zeh.zeh_integer_size;
982 		za->za_num_integers = zeh.zeh_num_integers;
983 		if (zeh.zeh_num_integers == 0) {
984 			za->za_first_integer = 0;
985 		} else {
986 			err = zap_entry_read(&zeh, 8, 1, &za->za_first_integer);
987 			ASSERT(err == 0 || err == EOVERFLOW);
988 		}
989 		err = zap_entry_read_name(&zeh,
990 		    sizeof (za->za_name), za->za_name);
991 		ASSERT(err == 0);
992 	}
993 	rw_exit(&zc->zc_leaf->l_rwlock);
994 	return (err);
995 }
996 
997 
998 static void
999 zap_stats_ptrtbl(zap_t *zap, uint64_t *tbl, int len, zap_stats_t *zs)
1000 {
1001 	int i, err;
1002 	uint64_t lastblk = 0;
1003 
1004 	/*
1005 	 * NB: if a leaf has more pointers than an entire ptrtbl block
1006 	 * can hold, then it'll be accounted for more than once, since
1007 	 * we won't have lastblk.
1008 	 */
1009 	for (i = 0; i < len; i++) {
1010 		zap_leaf_t *l;
1011 
1012 		if (tbl[i] == lastblk)
1013 			continue;
1014 		lastblk = tbl[i];
1015 
1016 		err = zap_get_leaf_byblk(zap, tbl[i], NULL, RW_READER, &l);
1017 		if (err == 0) {
1018 			zap_leaf_stats(zap, l, zs);
1019 			zap_put_leaf(l);
1020 		}
1021 	}
1022 }
1023 
1024 void
1025 fzap_get_stats(zap_t *zap, zap_stats_t *zs)
1026 {
1027 	int bs = FZAP_BLOCK_SHIFT(zap);
1028 	zs->zs_ptrtbl_len = 1ULL << zap->zap_f.zap_phys->zap_ptrtbl.zt_shift;
1029 	zs->zs_blocksize = 1ULL << bs;
1030 	zs->zs_num_leafs = zap->zap_f.zap_phys->zap_num_leafs;
1031 	zs->zs_num_entries = zap->zap_f.zap_phys->zap_num_entries;
1032 	zs->zs_num_blocks = zap->zap_f.zap_phys->zap_freeblk;
1033 
1034 	if (zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks == 0) {
1035 		/* the ptrtbl is entirely in the header block. */
1036 		zap_stats_ptrtbl(zap, &ZAP_EMBEDDED_PTRTBL_ENT(zap, 0),
1037 		    1 << ZAP_EMBEDDED_PTRTBL_SHIFT(zap), zs);
1038 	} else {
1039 		int b;
1040 
1041 		dmu_prefetch(zap->zap_objset, zap->zap_object,
1042 		    zap->zap_f.zap_phys->zap_ptrtbl.zt_blk << bs,
1043 		    zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks << bs);
1044 
1045 		for (b = 0; b < zap->zap_f.zap_phys->zap_ptrtbl.zt_numblks;
1046 		    b++) {
1047 			dmu_buf_t *db;
1048 			int err;
1049 
1050 			err = dmu_buf_hold(zap->zap_objset, zap->zap_object,
1051 			    (zap->zap_f.zap_phys->zap_ptrtbl.zt_blk + b) << bs,
1052 			    FTAG, &db);
1053 			if (err == 0) {
1054 				zap_stats_ptrtbl(zap, db->db_data,
1055 				    1<<(bs-3), zs);
1056 				dmu_buf_rele(db, FTAG);
1057 			}
1058 		}
1059 	}
1060 }
1061