xref: /freebsd/sys/contrib/openzfs/module/zfs/zap_leaf.c (revision 22649d4dba730d46244fd2dff4fd174903c8379f)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * This file and its contents are supplied under the terms of the
4  * Common Development and Distribution License ("CDDL"), version 1.0.
5  * You may only use this file in accordance with the terms of version
6  * 1.0 of the CDDL.
7  *
8  * A full copy of the text of the CDDL should have accompanied this
9  * source.  A copy of the CDDL is also available via the Internet at
10  * https://opensource.org/license/CDDL-1.0.
11  */
12 
13 /*
14  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
15  * Copyright (c) 2013, 2016 by Delphix. All rights reserved.
16  * Copyright 2017 Nexenta Systems, Inc.
17  */
18 
19 /*
20  * The 512-byte leaf is broken into 32 16-byte chunks.
21  * chunk number n means l_chunk[n], even though the header precedes it.
22  * the names are stored null-terminated.
23  */
24 
25 #include <sys/zio.h>
26 #include <sys/spa.h>
27 #include <sys/dmu.h>
28 #include <sys/zfs_context.h>
29 #include <sys/fs/zfs.h>
30 #include <sys/zap.h>
31 #include <sys/zap_impl.h>
32 #include <sys/zap_leaf.h>
33 #include <sys/arc.h>
34 
35 static uint16_t *zap_leaf_rehash_entry(zap_leaf_t *l, struct zap_leaf_entry *le,
36     uint16_t entry);
37 
38 #define	CHAIN_END 0xffff /* end of the chunk chain */
39 
40 #define	LEAF_HASH(l, h) \
41 	((ZAP_LEAF_HASH_NUMENTRIES(l)-1) & \
42 	((h) >> \
43 	(64 - ZAP_LEAF_HASH_SHIFT(l) - zap_leaf_phys(l)->l_hdr.lh_prefix_len)))
44 
45 #define	LEAF_HASH_ENTPTR(l, h)	(&zap_leaf_phys(l)->l_hash[LEAF_HASH(l, h)])
46 
47 static void
stv(int len,void * addr,uint64_t value)48 stv(int len, void *addr, uint64_t value)
49 {
50 	switch (len) {
51 	case 1:
52 		*(uint8_t *)addr = value;
53 		return;
54 	case 2:
55 		*(uint16_t *)addr = value;
56 		return;
57 	case 4:
58 		*(uint32_t *)addr = value;
59 		return;
60 	case 8:
61 		*(uint64_t *)addr = value;
62 		return;
63 	default:
64 		PANIC("bad int len %d", len);
65 	}
66 }
67 
68 static uint64_t
ldv(int len,const void * addr)69 ldv(int len, const void *addr)
70 {
71 	switch (len) {
72 	case 1:
73 		return (*(uint8_t *)addr);
74 	case 2:
75 		return (*(uint16_t *)addr);
76 	case 4:
77 		return (*(uint32_t *)addr);
78 	case 8:
79 		return (*(uint64_t *)addr);
80 	default:
81 		PANIC("bad int len %d", len);
82 	}
83 	return (0xFEEDFACEDEADBEEFULL);
84 }
85 
86 void
zap_leaf_byteswap(zap_leaf_phys_t * buf,size_t size)87 zap_leaf_byteswap(zap_leaf_phys_t *buf, size_t size)
88 {
89 	zap_leaf_t l;
90 	dmu_buf_t l_dbuf;
91 
92 	l_dbuf.db_data = buf;
93 	l.l_bs = highbit64(size) - 1;
94 	l.l_dbuf = &l_dbuf;
95 
96 	buf->l_hdr.lh_block_type =	BSWAP_64(buf->l_hdr.lh_block_type);
97 	buf->l_hdr.lh_prefix =		BSWAP_64(buf->l_hdr.lh_prefix);
98 	buf->l_hdr.lh_magic =		BSWAP_32(buf->l_hdr.lh_magic);
99 	buf->l_hdr.lh_nfree =		BSWAP_16(buf->l_hdr.lh_nfree);
100 	buf->l_hdr.lh_nentries =	BSWAP_16(buf->l_hdr.lh_nentries);
101 	buf->l_hdr.lh_prefix_len =	BSWAP_16(buf->l_hdr.lh_prefix_len);
102 	buf->l_hdr.lh_freelist =	BSWAP_16(buf->l_hdr.lh_freelist);
103 
104 	for (uint_t i = 0; i < ZAP_LEAF_HASH_NUMENTRIES(&l); i++)
105 		buf->l_hash[i] = BSWAP_16(buf->l_hash[i]);
106 
107 	for (uint_t i = 0; i < ZAP_LEAF_NUMCHUNKS(&l); i++) {
108 		zap_leaf_chunk_t *lc = &ZAP_LEAF_CHUNK(&l, i);
109 		struct zap_leaf_entry *le;
110 
111 		switch (lc->l_free.lf_type) {
112 		case ZAP_CHUNK_ENTRY:
113 			le = &lc->l_entry;
114 
115 			le->le_type =		BSWAP_8(le->le_type);
116 			le->le_value_intlen =	BSWAP_8(le->le_value_intlen);
117 			le->le_next =		BSWAP_16(le->le_next);
118 			le->le_name_chunk =	BSWAP_16(le->le_name_chunk);
119 			le->le_name_numints =	BSWAP_16(le->le_name_numints);
120 			le->le_value_chunk =	BSWAP_16(le->le_value_chunk);
121 			le->le_value_numints =	BSWAP_16(le->le_value_numints);
122 			le->le_cd =		BSWAP_32(le->le_cd);
123 			le->le_hash =		BSWAP_64(le->le_hash);
124 			break;
125 		case ZAP_CHUNK_FREE:
126 			lc->l_free.lf_type =	BSWAP_8(lc->l_free.lf_type);
127 			lc->l_free.lf_next =	BSWAP_16(lc->l_free.lf_next);
128 			break;
129 		case ZAP_CHUNK_ARRAY:
130 			lc->l_array.la_type =	BSWAP_8(lc->l_array.la_type);
131 			lc->l_array.la_next =	BSWAP_16(lc->l_array.la_next);
132 			/* la_array doesn't need swapping */
133 			break;
134 		default:
135 			cmn_err(CE_PANIC, "bad leaf type %d",
136 			    lc->l_free.lf_type);
137 		}
138 	}
139 }
140 
141 void
zap_leaf_init(zap_leaf_t * l,boolean_t sort)142 zap_leaf_init(zap_leaf_t *l, boolean_t sort)
143 {
144 	l->l_bs = highbit64(l->l_dbuf->db_size) - 1;
145 	memset(&zap_leaf_phys(l)->l_hdr, 0,
146 	    sizeof (struct zap_leaf_header));
147 	memset(zap_leaf_phys(l)->l_hash, CHAIN_END,
148 	    2*ZAP_LEAF_HASH_NUMENTRIES(l));
149 	for (uint_t i = 0; i < ZAP_LEAF_NUMCHUNKS(l); i++) {
150 		ZAP_LEAF_CHUNK(l, i).l_free.lf_type = ZAP_CHUNK_FREE;
151 		ZAP_LEAF_CHUNK(l, i).l_free.lf_next = i+1;
152 	}
153 	ZAP_LEAF_CHUNK(l, ZAP_LEAF_NUMCHUNKS(l)-1).l_free.lf_next = CHAIN_END;
154 	zap_leaf_phys(l)->l_hdr.lh_block_type = ZBT_LEAF;
155 	zap_leaf_phys(l)->l_hdr.lh_magic = ZAP_LEAF_MAGIC;
156 	zap_leaf_phys(l)->l_hdr.lh_nfree = ZAP_LEAF_NUMCHUNKS(l);
157 	if (sort)
158 		zap_leaf_phys(l)->l_hdr.lh_flags |= ZLF_ENTRIES_CDSORTED;
159 }
160 
161 /*
162  * Routines which manipulate leaf chunks (l_chunk[]).
163  */
164 
165 static uint16_t
zap_leaf_chunk_alloc(zap_leaf_t * l)166 zap_leaf_chunk_alloc(zap_leaf_t *l)
167 {
168 	ASSERT(zap_leaf_phys(l)->l_hdr.lh_nfree > 0);
169 
170 	uint_t chunk = zap_leaf_phys(l)->l_hdr.lh_freelist;
171 	ASSERT3U(chunk, <, ZAP_LEAF_NUMCHUNKS(l));
172 	ASSERT3U(ZAP_LEAF_CHUNK(l, chunk).l_free.lf_type, ==, ZAP_CHUNK_FREE);
173 
174 	zap_leaf_phys(l)->l_hdr.lh_freelist =
175 	    ZAP_LEAF_CHUNK(l, chunk).l_free.lf_next;
176 
177 	zap_leaf_phys(l)->l_hdr.lh_nfree--;
178 
179 	return (chunk);
180 }
181 
182 static void
zap_leaf_chunk_free(zap_leaf_t * l,uint16_t chunk)183 zap_leaf_chunk_free(zap_leaf_t *l, uint16_t chunk)
184 {
185 	struct zap_leaf_free *zlf = &ZAP_LEAF_CHUNK(l, chunk).l_free;
186 	ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_nfree, <, ZAP_LEAF_NUMCHUNKS(l));
187 	ASSERT3U(chunk, <, ZAP_LEAF_NUMCHUNKS(l));
188 	ASSERT(zlf->lf_type != ZAP_CHUNK_FREE);
189 
190 	zlf->lf_type = ZAP_CHUNK_FREE;
191 	zlf->lf_next = zap_leaf_phys(l)->l_hdr.lh_freelist;
192 	memset(zlf->lf_pad, 0, sizeof (zlf->lf_pad)); /* help it to compress */
193 	zap_leaf_phys(l)->l_hdr.lh_freelist = chunk;
194 
195 	zap_leaf_phys(l)->l_hdr.lh_nfree++;
196 }
197 
198 /*
199  * Routines which manipulate leaf arrays (zap_leaf_array type chunks).
200  */
201 
202 static uint16_t
zap_leaf_array_create(zap_leaf_t * l,const char * buf,int integer_size,int num_integers)203 zap_leaf_array_create(zap_leaf_t *l, const char *buf,
204     int integer_size, int num_integers)
205 {
206 	uint16_t chunk_head;
207 	uint16_t *chunkp = &chunk_head;
208 	int byten = integer_size;
209 	uint64_t value = 0;
210 	int shift = (integer_size - 1) * 8;
211 	int len = num_integers;
212 
213 	ASSERT3U(num_integers * integer_size, <=, ZAP_MAXVALUELEN);
214 
215 	if (len > 0)
216 		value = ldv(integer_size, buf);
217 	while (len > 0) {
218 		uint16_t chunk = zap_leaf_chunk_alloc(l);
219 		struct zap_leaf_array *la = &ZAP_LEAF_CHUNK(l, chunk).l_array;
220 
221 		la->la_type = ZAP_CHUNK_ARRAY;
222 		for (int i = 0; i < ZAP_LEAF_ARRAY_BYTES; i++) {
223 			la->la_array[i] = value >> shift;
224 			value <<= 8;
225 			if (--byten == 0) {
226 				if (--len == 0)
227 					break;
228 				byten = integer_size;
229 				buf += integer_size;
230 				value = ldv(integer_size, buf);
231 			}
232 		}
233 
234 		*chunkp = chunk;
235 		chunkp = &la->la_next;
236 	}
237 	*chunkp = CHAIN_END;
238 
239 	return (chunk_head);
240 }
241 
242 /*
243  * Non-destructively copy array between leaves.
244  */
245 static uint16_t
zap_leaf_array_copy(zap_leaf_t * l,uint16_t chunk,zap_leaf_t * nl)246 zap_leaf_array_copy(zap_leaf_t *l, uint16_t chunk, zap_leaf_t *nl)
247 {
248 	uint16_t new_chunk;
249 	uint16_t *nchunkp = &new_chunk;
250 
251 	while (chunk != CHAIN_END) {
252 		ASSERT3U(chunk, <, ZAP_LEAF_NUMCHUNKS(l));
253 		uint16_t nchunk = zap_leaf_chunk_alloc(nl);
254 
255 		struct zap_leaf_array *la =
256 		    &ZAP_LEAF_CHUNK(l, chunk).l_array;
257 		struct zap_leaf_array *nla =
258 		    &ZAP_LEAF_CHUNK(nl, nchunk).l_array;
259 		ASSERT3U(la->la_type, ==, ZAP_CHUNK_ARRAY);
260 
261 		*nla = *la; /* structure assignment */
262 
263 		chunk = la->la_next;
264 		*nchunkp = nchunk;
265 		nchunkp = &nla->la_next;
266 	}
267 	*nchunkp = CHAIN_END;
268 	return (new_chunk);
269 }
270 
271 /*
272  * Free array.  Unlike trivial loop of zap_leaf_chunk_free() this does
273  * not reverse order of chunks in the free list, reducing fragmentation.
274  */
275 static void
zap_leaf_array_free(zap_leaf_t * l,uint16_t chunk)276 zap_leaf_array_free(zap_leaf_t *l, uint16_t chunk)
277 {
278 	struct zap_leaf_header *hdr = &zap_leaf_phys(l)->l_hdr;
279 	uint16_t *tailp = &hdr->lh_freelist;
280 	uint16_t oldfree = *tailp;
281 
282 	while (chunk != CHAIN_END) {
283 		ASSERT3U(chunk, <, ZAP_LEAF_NUMCHUNKS(l));
284 		zap_leaf_chunk_t *c = &ZAP_LEAF_CHUNK(l, chunk);
285 		ASSERT3U(c->l_array.la_type, ==, ZAP_CHUNK_ARRAY);
286 
287 		*tailp = chunk;
288 		chunk = c->l_array.la_next;
289 
290 		c->l_free.lf_type = ZAP_CHUNK_FREE;
291 		memset(c->l_free.lf_pad, 0, sizeof (c->l_free.lf_pad));
292 		tailp = &c->l_free.lf_next;
293 
294 		ASSERT3U(hdr->lh_nfree, <, ZAP_LEAF_NUMCHUNKS(l));
295 		hdr->lh_nfree++;
296 	}
297 
298 	*tailp = oldfree;
299 }
300 
301 /* array_len and buf_len are in integers, not bytes */
302 static void
zap_leaf_array_read(zap_leaf_t * l,uint16_t chunk,int array_int_len,int array_len,int buf_int_len,uint64_t buf_len,void * buf)303 zap_leaf_array_read(zap_leaf_t *l, uint16_t chunk,
304     int array_int_len, int array_len, int buf_int_len, uint64_t buf_len,
305     void *buf)
306 {
307 	int len = MIN(array_len, buf_len);
308 	int byten = 0;
309 	uint64_t value = 0;
310 	char *p = buf;
311 
312 	ASSERT3U(array_int_len, <=, buf_int_len);
313 
314 	/* Fast path for one 8-byte integer */
315 	if (array_int_len == 8 && buf_int_len == 8 && len == 1) {
316 		struct zap_leaf_array *la = &ZAP_LEAF_CHUNK(l, chunk).l_array;
317 		uint8_t *ip = la->la_array;
318 		uint64_t *buf64 = buf;
319 
320 		*buf64 = (uint64_t)ip[0] << 56 | (uint64_t)ip[1] << 48 |
321 		    (uint64_t)ip[2] << 40 | (uint64_t)ip[3] << 32 |
322 		    (uint64_t)ip[4] << 24 | (uint64_t)ip[5] << 16 |
323 		    (uint64_t)ip[6] << 8 | (uint64_t)ip[7];
324 		return;
325 	}
326 
327 	/* Fast path for an array of 1-byte integers (eg. the entry name) */
328 	if (array_int_len == 1 && buf_int_len == 1 &&
329 	    buf_len > array_len + ZAP_LEAF_ARRAY_BYTES) {
330 		while (chunk != CHAIN_END) {
331 			struct zap_leaf_array *la =
332 			    &ZAP_LEAF_CHUNK(l, chunk).l_array;
333 			memcpy(p, la->la_array, ZAP_LEAF_ARRAY_BYTES);
334 			p += ZAP_LEAF_ARRAY_BYTES;
335 			chunk = la->la_next;
336 		}
337 		return;
338 	}
339 
340 	while (len > 0) {
341 		struct zap_leaf_array *la = &ZAP_LEAF_CHUNK(l, chunk).l_array;
342 
343 		ASSERT3U(chunk, <, ZAP_LEAF_NUMCHUNKS(l));
344 		for (int i = 0; i < ZAP_LEAF_ARRAY_BYTES; i++) {
345 			value = (value << 8) | la->la_array[i];
346 			byten++;
347 			if (byten == array_int_len) {
348 				stv(buf_int_len, p, value);
349 				byten = 0;
350 				len--;
351 				if (len == 0)
352 					return;
353 				p += buf_int_len;
354 			}
355 		}
356 		chunk = la->la_next;
357 	}
358 }
359 
360 static boolean_t
zap_leaf_array_match(zap_leaf_t * l,zap_name_t * zn,uint_t chunk,int array_numints)361 zap_leaf_array_match(zap_leaf_t *l, zap_name_t *zn,
362     uint_t chunk, int array_numints)
363 {
364 	int bseen = 0;
365 
366 	if (zap_getflags(zn->zn_zap) & ZAP_FLAG_UINT64_KEY) {
367 		uint64_t *thiskey =
368 		    kmem_alloc(array_numints * sizeof (*thiskey), KM_SLEEP);
369 		ASSERT(zn->zn_key_intlen == sizeof (*thiskey));
370 
371 		zap_leaf_array_read(l, chunk, sizeof (*thiskey), array_numints,
372 		    sizeof (*thiskey), array_numints, thiskey);
373 		boolean_t match = memcmp(thiskey, zn->zn_key_orig,
374 		    array_numints * sizeof (*thiskey)) == 0;
375 		kmem_free(thiskey, array_numints * sizeof (*thiskey));
376 		return (match);
377 	}
378 
379 	ASSERT(zn->zn_key_intlen == 1);
380 	if (zn->zn_matchtype & MT_NORMALIZE) {
381 		char *thisname = kmem_alloc(array_numints, KM_SLEEP);
382 
383 		zap_leaf_array_read(l, chunk, sizeof (char), array_numints,
384 		    sizeof (char), array_numints, thisname);
385 		boolean_t match = zap_match(zn, thisname);
386 		kmem_free(thisname, array_numints);
387 		return (match);
388 	}
389 
390 	/*
391 	 * Fast path for exact matching.
392 	 * First check that the lengths match, so that we don't read
393 	 * past the end of the zn_key_orig array.
394 	 */
395 	if (array_numints != zn->zn_key_orig_numints)
396 		return (B_FALSE);
397 	while (bseen < array_numints) {
398 		struct zap_leaf_array *la = &ZAP_LEAF_CHUNK(l, chunk).l_array;
399 		int toread = MIN(array_numints - bseen, ZAP_LEAF_ARRAY_BYTES);
400 		ASSERT3U(chunk, <, ZAP_LEAF_NUMCHUNKS(l));
401 		if (memcmp(la->la_array, (char *)zn->zn_key_orig + bseen,
402 		    toread))
403 			break;
404 		chunk = la->la_next;
405 		bseen += toread;
406 	}
407 	return (bseen == array_numints);
408 }
409 
410 /*
411  * Routines which manipulate leaf entries.
412  */
413 
414 int
zap_leaf_lookup(zap_leaf_t * l,zap_name_t * zn,zap_entry_handle_t * zeh)415 zap_leaf_lookup(zap_leaf_t *l, zap_name_t *zn, zap_entry_handle_t *zeh)
416 {
417 	struct zap_leaf_entry *le;
418 
419 	ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_magic, ==, ZAP_LEAF_MAGIC);
420 
421 	for (uint16_t *chunkp = LEAF_HASH_ENTPTR(l, zn->zn_hash);
422 	    *chunkp != CHAIN_END; chunkp = &le->le_next) {
423 		uint16_t chunk = *chunkp;
424 		le = ZAP_LEAF_ENTRY(l, chunk);
425 
426 		ASSERT3U(chunk, <, ZAP_LEAF_NUMCHUNKS(l));
427 		ASSERT3U(le->le_type, ==, ZAP_CHUNK_ENTRY);
428 
429 		if (le->le_hash != zn->zn_hash)
430 			continue;
431 
432 		/*
433 		 * NB: the entry chain is always sorted by cd on
434 		 * normalized zap objects, so this will find the
435 		 * lowest-cd match for MT_NORMALIZE.
436 		 */
437 		ASSERT((zn->zn_matchtype == 0) ||
438 		    (zap_leaf_phys(l)->l_hdr.lh_flags & ZLF_ENTRIES_CDSORTED));
439 		if (zap_leaf_array_match(l, zn, le->le_name_chunk,
440 		    le->le_name_numints)) {
441 			zeh->zeh_num_integers = le->le_value_numints;
442 			zeh->zeh_integer_size = le->le_value_intlen;
443 			zeh->zeh_cd = le->le_cd;
444 			zeh->zeh_hash = le->le_hash;
445 			zeh->zeh_chunkp = chunkp;
446 			zeh->zeh_leaf = l;
447 			return (0);
448 		}
449 	}
450 
451 	return (SET_ERROR(ENOENT));
452 }
453 
454 /* Return (h1,cd1 >= h2,cd2) */
455 #define	HCD_GTEQ(h1, cd1, h2, cd2) \
456 	((h1 > h2) ? TRUE : ((h1 == h2 && cd1 >= cd2) ? TRUE : FALSE))
457 
458 int
zap_leaf_lookup_closest(zap_leaf_t * l,uint64_t h,uint32_t cd,zap_entry_handle_t * zeh)459 zap_leaf_lookup_closest(zap_leaf_t *l,
460     uint64_t h, uint32_t cd, zap_entry_handle_t *zeh)
461 {
462 	uint64_t besth = -1ULL;
463 	uint32_t bestcd = -1U;
464 	uint16_t bestlh = ZAP_LEAF_HASH_NUMENTRIES(l)-1;
465 	struct zap_leaf_entry *le;
466 
467 	ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_magic, ==, ZAP_LEAF_MAGIC);
468 
469 	for (uint16_t lh = LEAF_HASH(l, h); lh <= bestlh; lh++) {
470 		for (uint16_t chunk = zap_leaf_phys(l)->l_hash[lh];
471 		    chunk != CHAIN_END; chunk = le->le_next) {
472 			le = ZAP_LEAF_ENTRY(l, chunk);
473 
474 			ASSERT3U(chunk, <, ZAP_LEAF_NUMCHUNKS(l));
475 			ASSERT3U(le->le_type, ==, ZAP_CHUNK_ENTRY);
476 
477 			if (HCD_GTEQ(le->le_hash, le->le_cd, h, cd) &&
478 			    HCD_GTEQ(besth, bestcd, le->le_hash, le->le_cd)) {
479 				ASSERT3U(bestlh, >=, lh);
480 				bestlh = lh;
481 				besth = le->le_hash;
482 				bestcd = le->le_cd;
483 
484 				zeh->zeh_num_integers = le->le_value_numints;
485 				zeh->zeh_integer_size = le->le_value_intlen;
486 				zeh->zeh_cd = le->le_cd;
487 				zeh->zeh_hash = le->le_hash;
488 				zeh->zeh_fakechunk = chunk;
489 				zeh->zeh_chunkp = &zeh->zeh_fakechunk;
490 				zeh->zeh_leaf = l;
491 			}
492 		}
493 	}
494 
495 	return (bestcd == -1U ? SET_ERROR(ENOENT) : 0);
496 }
497 
498 int
zap_entry_read(const zap_entry_handle_t * zeh,uint8_t integer_size,uint64_t num_integers,void * buf)499 zap_entry_read(const zap_entry_handle_t *zeh,
500     uint8_t integer_size, uint64_t num_integers, void *buf)
501 {
502 	struct zap_leaf_entry *le =
503 	    ZAP_LEAF_ENTRY(zeh->zeh_leaf, *zeh->zeh_chunkp);
504 	ASSERT3U(le->le_type, ==, ZAP_CHUNK_ENTRY);
505 
506 	if (le->le_value_intlen > integer_size)
507 		return (SET_ERROR(EINVAL));
508 
509 	zap_leaf_array_read(zeh->zeh_leaf, le->le_value_chunk,
510 	    le->le_value_intlen, le->le_value_numints,
511 	    integer_size, num_integers, buf);
512 
513 	if (zeh->zeh_num_integers > num_integers)
514 		return (SET_ERROR(EOVERFLOW));
515 	return (0);
516 
517 }
518 
519 int
zap_entry_read_name(zap_t * zap,const zap_entry_handle_t * zeh,uint16_t buflen,char * buf)520 zap_entry_read_name(zap_t *zap, const zap_entry_handle_t *zeh, uint16_t buflen,
521     char *buf)
522 {
523 	struct zap_leaf_entry *le =
524 	    ZAP_LEAF_ENTRY(zeh->zeh_leaf, *zeh->zeh_chunkp);
525 	ASSERT3U(le->le_type, ==, ZAP_CHUNK_ENTRY);
526 
527 	if (zap_getflags(zap) & ZAP_FLAG_UINT64_KEY) {
528 		zap_leaf_array_read(zeh->zeh_leaf, le->le_name_chunk, 8,
529 		    le->le_name_numints, 8, buflen / 8, buf);
530 	} else {
531 		zap_leaf_array_read(zeh->zeh_leaf, le->le_name_chunk, 1,
532 		    le->le_name_numints, 1, buflen, buf);
533 	}
534 	if (le->le_name_numints > buflen)
535 		return (SET_ERROR(EOVERFLOW));
536 	return (0);
537 }
538 
539 int
zap_entry_update(zap_entry_handle_t * zeh,uint8_t integer_size,uint64_t num_integers,const void * buf)540 zap_entry_update(zap_entry_handle_t *zeh,
541     uint8_t integer_size, uint64_t num_integers, const void *buf)
542 {
543 	zap_leaf_t *l = zeh->zeh_leaf;
544 	struct zap_leaf_entry *le = ZAP_LEAF_ENTRY(l, *zeh->zeh_chunkp);
545 
546 	int delta_chunks = ZAP_LEAF_ARRAY_NCHUNKS(num_integers * integer_size) -
547 	    ZAP_LEAF_ARRAY_NCHUNKS(le->le_value_numints * le->le_value_intlen);
548 
549 	if ((int)zap_leaf_phys(l)->l_hdr.lh_nfree < delta_chunks)
550 		return (SET_ERROR(EAGAIN));
551 
552 	zap_leaf_array_free(l, le->le_value_chunk);
553 	le->le_value_chunk =
554 	    zap_leaf_array_create(l, buf, integer_size, num_integers);
555 	le->le_value_numints = num_integers;
556 	le->le_value_intlen = integer_size;
557 	return (0);
558 }
559 
560 void
zap_entry_remove(zap_entry_handle_t * zeh)561 zap_entry_remove(zap_entry_handle_t *zeh)
562 {
563 	zap_leaf_t *l = zeh->zeh_leaf;
564 
565 	ASSERT3P(zeh->zeh_chunkp, !=, &zeh->zeh_fakechunk);
566 
567 	uint16_t entry_chunk = *zeh->zeh_chunkp;
568 	struct zap_leaf_entry *le = ZAP_LEAF_ENTRY(l, entry_chunk);
569 	ASSERT3U(le->le_type, ==, ZAP_CHUNK_ENTRY);
570 
571 	*zeh->zeh_chunkp = le->le_next;
572 
573 	/* Free in opposite order to reduce fragmentation. */
574 	zap_leaf_array_free(l, le->le_value_chunk);
575 	zap_leaf_array_free(l, le->le_name_chunk);
576 	zap_leaf_chunk_free(l, entry_chunk);
577 
578 	zap_leaf_phys(l)->l_hdr.lh_nentries--;
579 }
580 
581 int
zap_entry_create(zap_leaf_t * l,zap_name_t * zn,uint32_t cd,uint8_t integer_size,uint64_t num_integers,const void * buf,zap_entry_handle_t * zeh)582 zap_entry_create(zap_leaf_t *l, zap_name_t *zn, uint32_t cd,
583     uint8_t integer_size, uint64_t num_integers, const void *buf,
584     zap_entry_handle_t *zeh)
585 {
586 	uint16_t chunk;
587 	struct zap_leaf_entry *le;
588 	uint64_t h = zn->zn_hash;
589 
590 	uint64_t valuelen = integer_size * num_integers;
591 
592 	uint_t numchunks = 1 + ZAP_LEAF_ARRAY_NCHUNKS(zn->zn_key_orig_numints *
593 	    zn->zn_key_intlen) + ZAP_LEAF_ARRAY_NCHUNKS(valuelen);
594 	if (numchunks > ZAP_LEAF_NUMCHUNKS(l))
595 		return (SET_ERROR(E2BIG));
596 
597 	if (cd == ZAP_NEED_CD) {
598 		/* find the lowest unused cd */
599 		if (zap_leaf_phys(l)->l_hdr.lh_flags & ZLF_ENTRIES_CDSORTED) {
600 			cd = 0;
601 
602 			for (chunk = *LEAF_HASH_ENTPTR(l, h);
603 			    chunk != CHAIN_END; chunk = le->le_next) {
604 				le = ZAP_LEAF_ENTRY(l, chunk);
605 				if (le->le_cd > cd)
606 					break;
607 				if (le->le_hash == h) {
608 					ASSERT3U(cd, ==, le->le_cd);
609 					cd++;
610 				}
611 			}
612 		} else {
613 			/* old unsorted format; do it the O(n^2) way */
614 			for (cd = 0; ; cd++) {
615 				for (chunk = *LEAF_HASH_ENTPTR(l, h);
616 				    chunk != CHAIN_END; chunk = le->le_next) {
617 					le = ZAP_LEAF_ENTRY(l, chunk);
618 					if (le->le_hash == h &&
619 					    le->le_cd == cd) {
620 						break;
621 					}
622 				}
623 				/* If this cd is not in use, we are good. */
624 				if (chunk == CHAIN_END)
625 					break;
626 			}
627 		}
628 		/*
629 		 * We would run out of space in a block before we could
630 		 * store enough entries to run out of CD values.
631 		 */
632 		ASSERT3U(cd, <, zap_maxcd(zn->zn_zap));
633 	}
634 
635 	if (zap_leaf_phys(l)->l_hdr.lh_nfree < numchunks)
636 		return (SET_ERROR(EAGAIN));
637 
638 	/* make the entry */
639 	chunk = zap_leaf_chunk_alloc(l);
640 	le = ZAP_LEAF_ENTRY(l, chunk);
641 	le->le_type = ZAP_CHUNK_ENTRY;
642 	le->le_name_chunk = zap_leaf_array_create(l, zn->zn_key_orig,
643 	    zn->zn_key_intlen, zn->zn_key_orig_numints);
644 	le->le_name_numints = zn->zn_key_orig_numints;
645 	le->le_value_chunk =
646 	    zap_leaf_array_create(l, buf, integer_size, num_integers);
647 	le->le_value_numints = num_integers;
648 	le->le_value_intlen = integer_size;
649 	le->le_hash = h;
650 	le->le_cd = cd;
651 
652 	/* link it into the hash chain */
653 	/* XXX if we did the search above, we could just use that */
654 	uint16_t *chunkp = zap_leaf_rehash_entry(l, le, chunk);
655 
656 	zap_leaf_phys(l)->l_hdr.lh_nentries++;
657 
658 	zeh->zeh_leaf = l;
659 	zeh->zeh_num_integers = num_integers;
660 	zeh->zeh_integer_size = le->le_value_intlen;
661 	zeh->zeh_cd = le->le_cd;
662 	zeh->zeh_hash = le->le_hash;
663 	zeh->zeh_chunkp = chunkp;
664 
665 	return (0);
666 }
667 
668 /*
669  * Determine if there is another entry with the same normalized form.
670  * For performance purposes, either zn or name must be provided (the
671  * other can be NULL).  Note, there usually won't be any hash
672  * conflicts, in which case we don't need the concatenated/normalized
673  * form of the name.  But all callers have one of these on hand anyway,
674  * so might as well take advantage.  A cleaner but slower interface
675  * would accept neither argument, and compute the normalized name as
676  * needed (using zap_name_alloc_str(zap_entry_read_name(zeh))).
677  */
678 boolean_t
zap_entry_normalization_conflict(zap_entry_handle_t * zeh,zap_name_t * zn,const char * name,zap_t * zap)679 zap_entry_normalization_conflict(zap_entry_handle_t *zeh, zap_name_t *zn,
680     const char *name, zap_t *zap)
681 {
682 	struct zap_leaf_entry *le;
683 	boolean_t allocdzn = B_FALSE;
684 
685 	if (zap->zap_normflags == 0)
686 		return (B_FALSE);
687 
688 	for (uint16_t chunk = *LEAF_HASH_ENTPTR(zeh->zeh_leaf, zeh->zeh_hash);
689 	    chunk != CHAIN_END; chunk = le->le_next) {
690 		le = ZAP_LEAF_ENTRY(zeh->zeh_leaf, chunk);
691 		if (le->le_hash != zeh->zeh_hash)
692 			continue;
693 		if (le->le_cd == zeh->zeh_cd)
694 			continue;
695 
696 		if (zn == NULL) {
697 			zn = zap_name_alloc_str(zap, name, MT_NORMALIZE);
698 			allocdzn = B_TRUE;
699 		}
700 		if (zap_leaf_array_match(zeh->zeh_leaf, zn,
701 		    le->le_name_chunk, le->le_name_numints)) {
702 			if (allocdzn)
703 				zap_name_free(zn);
704 			return (B_TRUE);
705 		}
706 	}
707 	if (allocdzn)
708 		zap_name_free(zn);
709 	return (B_FALSE);
710 }
711 
712 /*
713  * Routines for transferring entries between leafs.
714  */
715 
716 static uint16_t *
zap_leaf_rehash_entry(zap_leaf_t * l,struct zap_leaf_entry * le,uint16_t entry)717 zap_leaf_rehash_entry(zap_leaf_t *l, struct zap_leaf_entry *le, uint16_t entry)
718 {
719 	struct zap_leaf_entry *le2;
720 	uint16_t *chunkp;
721 
722 	/*
723 	 * keep the entry chain sorted by cd
724 	 * NB: this will not cause problems for unsorted leafs, though
725 	 * it is unnecessary there.
726 	 */
727 	for (chunkp = LEAF_HASH_ENTPTR(l, le->le_hash);
728 	    *chunkp != CHAIN_END; chunkp = &le2->le_next) {
729 		le2 = ZAP_LEAF_ENTRY(l, *chunkp);
730 		if (le2->le_cd > le->le_cd)
731 			break;
732 	}
733 
734 	le->le_next = *chunkp;
735 	*chunkp = entry;
736 	return (chunkp);
737 }
738 
739 static void
zap_leaf_transfer_entry(zap_leaf_t * l,uint_t entry,zap_leaf_t * nl)740 zap_leaf_transfer_entry(zap_leaf_t *l, uint_t entry, zap_leaf_t *nl)
741 {
742 	struct zap_leaf_entry *le = ZAP_LEAF_ENTRY(l, entry);
743 	ASSERT3U(le->le_type, ==, ZAP_CHUNK_ENTRY);
744 
745 	uint16_t chunk = zap_leaf_chunk_alloc(nl);
746 	struct zap_leaf_entry *nle = ZAP_LEAF_ENTRY(nl, chunk);
747 	*nle = *le; /* structure assignment */
748 
749 	(void) zap_leaf_rehash_entry(nl, nle, chunk);
750 
751 	nle->le_name_chunk = zap_leaf_array_copy(l, le->le_name_chunk, nl);
752 	nle->le_value_chunk = zap_leaf_array_copy(l, le->le_value_chunk, nl);
753 
754 	/* Free in opposite order to reduce fragmentation. */
755 	zap_leaf_array_free(l, le->le_value_chunk);
756 	zap_leaf_array_free(l, le->le_name_chunk);
757 	zap_leaf_chunk_free(l, entry);
758 
759 	zap_leaf_phys(l)->l_hdr.lh_nentries--;
760 	zap_leaf_phys(nl)->l_hdr.lh_nentries++;
761 }
762 
763 /*
764  * Transfer the entries whose hash prefix ends in 1 to the new leaf.
765  */
766 void
zap_leaf_split(zap_leaf_t * l,zap_leaf_t * nl,boolean_t sort)767 zap_leaf_split(zap_leaf_t *l, zap_leaf_t *nl, boolean_t sort)
768 {
769 	uint_t bit = 64 - 1 - zap_leaf_phys(l)->l_hdr.lh_prefix_len;
770 
771 	/* set new prefix and prefix_len */
772 	zap_leaf_phys(l)->l_hdr.lh_prefix <<= 1;
773 	zap_leaf_phys(l)->l_hdr.lh_prefix_len++;
774 	zap_leaf_phys(nl)->l_hdr.lh_prefix =
775 	    zap_leaf_phys(l)->l_hdr.lh_prefix | 1;
776 	zap_leaf_phys(nl)->l_hdr.lh_prefix_len =
777 	    zap_leaf_phys(l)->l_hdr.lh_prefix_len;
778 
779 	/* break existing hash chains */
780 	memset(zap_leaf_phys(l)->l_hash, CHAIN_END,
781 	    2*ZAP_LEAF_HASH_NUMENTRIES(l));
782 
783 	if (sort)
784 		zap_leaf_phys(l)->l_hdr.lh_flags |= ZLF_ENTRIES_CDSORTED;
785 
786 	/*
787 	 * Transfer entries whose hash bit 'bit' is set to nl; rehash
788 	 * the remaining entries
789 	 *
790 	 * NB: We could find entries via the hashtable instead. That
791 	 * would be O(hashents+numents) rather than O(numblks+numents),
792 	 * but this accesses memory more sequentially, and when we're
793 	 * called, the block is usually pretty full.
794 	 */
795 	for (uint_t i = 0; i < ZAP_LEAF_NUMCHUNKS(l); i++) {
796 		struct zap_leaf_entry *le = ZAP_LEAF_ENTRY(l, i);
797 		if (le->le_type != ZAP_CHUNK_ENTRY)
798 			continue;
799 
800 		if (le->le_hash & (1ULL << bit))
801 			zap_leaf_transfer_entry(l, i, nl);
802 		else
803 			(void) zap_leaf_rehash_entry(l, le, i);
804 	}
805 }
806 
807 void
zap_leaf_stats(zap_t * zap,zap_leaf_t * l,zap_stats_t * zs)808 zap_leaf_stats(zap_t *zap, zap_leaf_t *l, zap_stats_t *zs)
809 {
810 	uint_t n = zap_f_phys(zap)->zap_ptrtbl.zt_shift -
811 	    zap_leaf_phys(l)->l_hdr.lh_prefix_len;
812 	n = MIN(n, ZAP_HISTOGRAM_SIZE-1);
813 	zs->zs_leafs_with_2n_pointers[n]++;
814 
815 
816 	n = zap_leaf_phys(l)->l_hdr.lh_nentries/5;
817 	n = MIN(n, ZAP_HISTOGRAM_SIZE-1);
818 	zs->zs_blocks_with_n5_entries[n]++;
819 
820 	n = ((1<<FZAP_BLOCK_SHIFT(zap)) -
821 	    zap_leaf_phys(l)->l_hdr.lh_nfree * (ZAP_LEAF_ARRAY_BYTES+1))*10 /
822 	    (1<<FZAP_BLOCK_SHIFT(zap));
823 	n = MIN(n, ZAP_HISTOGRAM_SIZE-1);
824 	zs->zs_blocks_n_tenths_full[n]++;
825 
826 	for (uint_t i = 0; i < ZAP_LEAF_HASH_NUMENTRIES(l); i++) {
827 		uint_t nentries = 0;
828 		uint_t chunk = zap_leaf_phys(l)->l_hash[i];
829 
830 		while (chunk != CHAIN_END) {
831 			struct zap_leaf_entry *le =
832 			    ZAP_LEAF_ENTRY(l, chunk);
833 
834 			n = 1 + ZAP_LEAF_ARRAY_NCHUNKS(le->le_name_numints) +
835 			    ZAP_LEAF_ARRAY_NCHUNKS(le->le_value_numints *
836 			    le->le_value_intlen);
837 			n = MIN(n, ZAP_HISTOGRAM_SIZE-1);
838 			zs->zs_entries_using_n_chunks[n]++;
839 
840 			chunk = le->le_next;
841 			nentries++;
842 		}
843 
844 		n = nentries;
845 		n = MIN(n, ZAP_HISTOGRAM_SIZE-1);
846 		zs->zs_buckets_with_n_entries[n]++;
847 	}
848 }
849