xref: /illumos-gate/usr/src/uts/common/fs/zfs/abd.c (revision eb633035c80613ec93d62f90482837adaaf21a0a)
1 /*
2  * This file and its contents are supplied under the terms of the
3  * Common Development and Distribution License ("CDDL"), version 1.0.
4  * You may only use this file in accordance with the terms of version
5  * 1.0 of the CDDL.
6  *
7  * A full copy of the text of the CDDL should have accompanied this
8  * source.  A copy of the CDDL is also available via the Internet at
9  * http://www.illumos.org/license/CDDL.
10  */
11 
12 /*
13  * Copyright (c) 2014 by Chunwei Chen. All rights reserved.
14  * Copyright (c) 2016 by Delphix. All rights reserved.
15  */
16 
17 /*
18  * ARC buffer data (ABD).
19  *
20  * ABDs are an abstract data structure for the ARC which can use two
21  * different ways of storing the underlying data:
22  *
23  * (a) Linear buffer. In this case, all the data in the ABD is stored in one
24  *     contiguous buffer in memory (from a zio_[data_]buf_* kmem cache).
25  *
26  *         +-------------------+
27  *         | ABD (linear)      |
28  *         |   abd_flags = ... |
29  *         |   abd_size = ...  |     +--------------------------------+
30  *         |   abd_buf ------------->| raw buffer of size abd_size    |
31  *         +-------------------+     +--------------------------------+
32  *              no abd_chunks
33  *
34  * (b) Scattered buffer. In this case, the data in the ABD is split into
35  *     equal-sized chunks (from the abd_chunk_cache kmem_cache), with pointers
36  *     to the chunks recorded in an array at the end of the ABD structure.
37  *
38  *         +-------------------+
39  *         | ABD (scattered)   |
40  *         |   abd_flags = ... |
41  *         |   abd_size = ...  |
42  *         |   abd_offset = 0  |                           +-----------+
43  *         |   abd_chunks[0] ----------------------------->| chunk 0   |
44  *         |   abd_chunks[1] ---------------------+        +-----------+
45  *         |   ...             |                  |        +-----------+
46  *         |   abd_chunks[N-1] ---------+         +------->| chunk 1   |
47  *         +-------------------+        |                  +-----------+
48  *                                      |                      ...
49  *                                      |                  +-----------+
50  *                                      +----------------->| chunk N-1 |
51  *                                                         +-----------+
52  *
53  * Using a large proportion of scattered ABDs decreases ARC fragmentation since
54  * when we are at the limit of allocatable space, using equal-size chunks will
55  * allow us to quickly reclaim enough space for a new large allocation (assuming
56  * it is also scattered).
57  *
58  * In addition to directly allocating a linear or scattered ABD, it is also
59  * possible to create an ABD by requesting the "sub-ABD" starting at an offset
60  * within an existing ABD. In linear buffers this is simple (set abd_buf of
61  * the new ABD to the starting point within the original raw buffer), but
62  * scattered ABDs are a little more complex. The new ABD makes a copy of the
63  * relevant abd_chunks pointers (but not the underlying data). However, to
64  * provide arbitrary rather than only chunk-aligned starting offsets, it also
65  * tracks an abd_offset field which represents the starting point of the data
66  * within the first chunk in abd_chunks. For both linear and scattered ABDs,
67  * creating an offset ABD marks the original ABD as the offset's parent, and the
68  * original ABD's abd_children refcount is incremented. This data allows us to
69  * ensure the root ABD isn't deleted before its children.
70  *
71  * Most consumers should never need to know what type of ABD they're using --
72  * the ABD public API ensures that it's possible to transparently switch from
73  * using a linear ABD to a scattered one when doing so would be beneficial.
74  *
75  * If you need to use the data within an ABD directly, if you know it's linear
76  * (because you allocated it) you can use abd_to_buf() to access the underlying
77  * raw buffer. Otherwise, you should use one of the abd_borrow_buf* functions
78  * which will allocate a raw buffer if necessary. Use the abd_return_buf*
79  * functions to return any raw buffers that are no longer necessary when you're
80  * done using them.
81  *
82  * There are a variety of ABD APIs that implement basic buffer operations:
83  * compare, copy, read, write, and fill with zeroes. If you need a custom
84  * function which progressively accesses the whole ABD, use the abd_iterate_*
85  * functions.
86  */
87 
88 #include <sys/abd.h>
89 #include <sys/param.h>
90 #include <sys/zio.h>
91 #include <sys/zfs_context.h>
92 #include <sys/zfs_znode.h>
93 
94 typedef struct abd_stats {
95 	kstat_named_t abdstat_struct_size;
96 	kstat_named_t abdstat_scatter_cnt;
97 	kstat_named_t abdstat_scatter_data_size;
98 	kstat_named_t abdstat_scatter_chunk_waste;
99 	kstat_named_t abdstat_linear_cnt;
100 	kstat_named_t abdstat_linear_data_size;
101 } abd_stats_t;
102 
103 static abd_stats_t abd_stats = {
104 	/* Amount of memory occupied by all of the abd_t struct allocations */
105 	{ "struct_size",			KSTAT_DATA_UINT64 },
106 	/*
107 	 * The number of scatter ABDs which are currently allocated, excluding
108 	 * ABDs which don't own their data (for instance the ones which were
109 	 * allocated through abd_get_offset()).
110 	 */
111 	{ "scatter_cnt",			KSTAT_DATA_UINT64 },
112 	/* Amount of data stored in all scatter ABDs tracked by scatter_cnt */
113 	{ "scatter_data_size",			KSTAT_DATA_UINT64 },
114 	/*
115 	 * The amount of space wasted at the end of the last chunk across all
116 	 * scatter ABDs tracked by scatter_cnt.
117 	 */
118 	{ "scatter_chunk_waste",		KSTAT_DATA_UINT64 },
119 	/*
120 	 * The number of linear ABDs which are currently allocated, excluding
121 	 * ABDs which don't own their data (for instance the ones which were
122 	 * allocated through abd_get_offset() and abd_get_from_buf()). If an
123 	 * ABD takes ownership of its buf then it will become tracked.
124 	 */
125 	{ "linear_cnt",				KSTAT_DATA_UINT64 },
126 	/* Amount of data stored in all linear ABDs tracked by linear_cnt */
127 	{ "linear_data_size",			KSTAT_DATA_UINT64 },
128 };
129 
130 #define	ABDSTAT(stat)		(abd_stats.stat.value.ui64)
131 #define	ABDSTAT_INCR(stat, val) \
132 	atomic_add_64(&abd_stats.stat.value.ui64, (val))
133 #define	ABDSTAT_BUMP(stat)	ABDSTAT_INCR(stat, 1)
134 #define	ABDSTAT_BUMPDOWN(stat)	ABDSTAT_INCR(stat, -1)
135 
136 /*
137  * It is possible to make all future ABDs be linear by setting this to B_FALSE.
138  * Otherwise, ABDs are allocated scattered by default unless the caller uses
139  * abd_alloc_linear().
140  */
141 boolean_t zfs_abd_scatter_enabled = B_TRUE;
142 
143 /*
144  * The size of the chunks ABD allocates. Because the sizes allocated from the
145  * kmem_cache can't change, this tunable can only be modified at boot. Changing
146  * it at runtime would cause ABD iteration to work incorrectly for ABDs which
147  * were allocated with the old size, so a safeguard has been put in place which
148  * will cause the machine to panic if you change it and try to access the data
149  * within a scattered ABD.
150  */
151 size_t zfs_abd_chunk_size = 4096;
152 
153 #ifdef _KERNEL
154 extern vmem_t *zio_alloc_arena;
155 #endif
156 
157 kmem_cache_t *abd_chunk_cache;
158 static kstat_t *abd_ksp;
159 
160 extern inline boolean_t abd_is_linear(abd_t *abd);
161 extern inline void abd_copy(abd_t *dabd, abd_t *sabd, size_t size);
162 extern inline void abd_copy_from_buf(abd_t *abd, const void *buf, size_t size);
163 extern inline void abd_copy_to_buf(void* buf, abd_t *abd, size_t size);
164 extern inline int abd_cmp_buf(abd_t *abd, const void *buf, size_t size);
165 extern inline void abd_zero(abd_t *abd, size_t size);
166 
167 static void *
168 abd_alloc_chunk()
169 {
170 	void *c = kmem_cache_alloc(abd_chunk_cache, KM_PUSHPAGE);
171 	ASSERT3P(c, !=, NULL);
172 	return (c);
173 }
174 
175 static void
176 abd_free_chunk(void *c)
177 {
178 	kmem_cache_free(abd_chunk_cache, c);
179 }
180 
181 void
182 abd_init(void)
183 {
184 	vmem_t *data_alloc_arena = NULL;
185 
186 #ifdef _KERNEL
187 	data_alloc_arena = zio_alloc_arena;
188 #endif
189 
190 	/*
191 	 * Since ABD chunks do not appear in crash dumps, we pass KMC_NOTOUCH
192 	 * so that no allocator metadata is stored with the buffers.
193 	 */
194 	abd_chunk_cache = kmem_cache_create("abd_chunk", zfs_abd_chunk_size, 0,
195 	    NULL, NULL, NULL, NULL, data_alloc_arena, KMC_NOTOUCH);
196 
197 	abd_ksp = kstat_create("zfs", 0, "abdstats", "misc", KSTAT_TYPE_NAMED,
198 	    sizeof (abd_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
199 	if (abd_ksp != NULL) {
200 		abd_ksp->ks_data = &abd_stats;
201 		kstat_install(abd_ksp);
202 	}
203 }
204 
205 void
206 abd_fini(void)
207 {
208 	if (abd_ksp != NULL) {
209 		kstat_delete(abd_ksp);
210 		abd_ksp = NULL;
211 	}
212 
213 	kmem_cache_destroy(abd_chunk_cache);
214 	abd_chunk_cache = NULL;
215 }
216 
217 static inline size_t
218 abd_chunkcnt_for_bytes(size_t size)
219 {
220 	return (P2ROUNDUP(size, zfs_abd_chunk_size) / zfs_abd_chunk_size);
221 }
222 
223 static inline size_t
224 abd_scatter_chunkcnt(abd_t *abd)
225 {
226 	ASSERT(!abd_is_linear(abd));
227 	return (abd_chunkcnt_for_bytes(
228 	    abd->abd_u.abd_scatter.abd_offset + abd->abd_size));
229 }
230 
231 static inline void
232 abd_verify(abd_t *abd)
233 {
234 	ASSERT3U(abd->abd_size, >, 0);
235 	ASSERT3U(abd->abd_size, <=, SPA_MAXBLOCKSIZE);
236 	ASSERT3U(abd->abd_flags, ==, abd->abd_flags & (ABD_FLAG_LINEAR |
237 	    ABD_FLAG_OWNER | ABD_FLAG_META));
238 	IMPLY(abd->abd_parent != NULL, !(abd->abd_flags & ABD_FLAG_OWNER));
239 	IMPLY(abd->abd_flags & ABD_FLAG_META, abd->abd_flags & ABD_FLAG_OWNER);
240 	if (abd_is_linear(abd)) {
241 		ASSERT3P(abd->abd_u.abd_linear.abd_buf, !=, NULL);
242 	} else {
243 		ASSERT3U(abd->abd_u.abd_scatter.abd_offset, <,
244 		    zfs_abd_chunk_size);
245 		size_t n = abd_scatter_chunkcnt(abd);
246 		for (int i = 0; i < n; i++) {
247 			ASSERT3P(
248 			    abd->abd_u.abd_scatter.abd_chunks[i], !=, NULL);
249 		}
250 	}
251 }
252 
253 static inline abd_t *
254 abd_alloc_struct(size_t chunkcnt)
255 {
256 	size_t size = offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]);
257 	abd_t *abd = kmem_alloc(size, KM_PUSHPAGE);
258 	ASSERT3P(abd, !=, NULL);
259 	ABDSTAT_INCR(abdstat_struct_size, size);
260 
261 	return (abd);
262 }
263 
264 static inline void
265 abd_free_struct(abd_t *abd)
266 {
267 	size_t chunkcnt = abd_is_linear(abd) ? 0 : abd_scatter_chunkcnt(abd);
268 	int size = offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]);
269 	kmem_free(abd, size);
270 	ABDSTAT_INCR(abdstat_struct_size, -size);
271 }
272 
273 /*
274  * Allocate an ABD, along with its own underlying data buffers. Use this if you
275  * don't care whether the ABD is linear or not.
276  */
277 abd_t *
278 abd_alloc(size_t size, boolean_t is_metadata)
279 {
280 	if (!zfs_abd_scatter_enabled)
281 		return (abd_alloc_linear(size, is_metadata));
282 
283 	VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
284 
285 	size_t n = abd_chunkcnt_for_bytes(size);
286 	abd_t *abd = abd_alloc_struct(n);
287 
288 	abd->abd_flags = ABD_FLAG_OWNER;
289 	if (is_metadata) {
290 		abd->abd_flags |= ABD_FLAG_META;
291 	}
292 	abd->abd_size = size;
293 	abd->abd_parent = NULL;
294 	zfs_refcount_create(&abd->abd_children);
295 
296 	abd->abd_u.abd_scatter.abd_offset = 0;
297 	abd->abd_u.abd_scatter.abd_chunk_size = zfs_abd_chunk_size;
298 
299 	for (int i = 0; i < n; i++) {
300 		void *c = abd_alloc_chunk();
301 		ASSERT3P(c, !=, NULL);
302 		abd->abd_u.abd_scatter.abd_chunks[i] = c;
303 	}
304 
305 	ABDSTAT_BUMP(abdstat_scatter_cnt);
306 	ABDSTAT_INCR(abdstat_scatter_data_size, size);
307 	ABDSTAT_INCR(abdstat_scatter_chunk_waste,
308 	    n * zfs_abd_chunk_size - size);
309 
310 	return (abd);
311 }
312 
313 static void
314 abd_free_scatter(abd_t *abd)
315 {
316 	size_t n = abd_scatter_chunkcnt(abd);
317 	for (int i = 0; i < n; i++) {
318 		abd_free_chunk(abd->abd_u.abd_scatter.abd_chunks[i]);
319 	}
320 
321 	zfs_refcount_destroy(&abd->abd_children);
322 	ABDSTAT_BUMPDOWN(abdstat_scatter_cnt);
323 	ABDSTAT_INCR(abdstat_scatter_data_size, -(int)abd->abd_size);
324 	ABDSTAT_INCR(abdstat_scatter_chunk_waste,
325 	    abd->abd_size - n * zfs_abd_chunk_size);
326 
327 	abd_free_struct(abd);
328 }
329 
330 /*
331  * Allocate an ABD that must be linear, along with its own underlying data
332  * buffer. Only use this when it would be very annoying to write your ABD
333  * consumer with a scattered ABD.
334  */
335 abd_t *
336 abd_alloc_linear(size_t size, boolean_t is_metadata)
337 {
338 	abd_t *abd = abd_alloc_struct(0);
339 
340 	VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
341 
342 	abd->abd_flags = ABD_FLAG_LINEAR | ABD_FLAG_OWNER;
343 	if (is_metadata) {
344 		abd->abd_flags |= ABD_FLAG_META;
345 	}
346 	abd->abd_size = size;
347 	abd->abd_parent = NULL;
348 	zfs_refcount_create(&abd->abd_children);
349 
350 	if (is_metadata) {
351 		abd->abd_u.abd_linear.abd_buf = zio_buf_alloc(size);
352 	} else {
353 		abd->abd_u.abd_linear.abd_buf = zio_data_buf_alloc(size);
354 	}
355 
356 	ABDSTAT_BUMP(abdstat_linear_cnt);
357 	ABDSTAT_INCR(abdstat_linear_data_size, size);
358 
359 	return (abd);
360 }
361 
362 static void
363 abd_free_linear(abd_t *abd)
364 {
365 	if (abd->abd_flags & ABD_FLAG_META) {
366 		zio_buf_free(abd->abd_u.abd_linear.abd_buf, abd->abd_size);
367 	} else {
368 		zio_data_buf_free(abd->abd_u.abd_linear.abd_buf, abd->abd_size);
369 	}
370 
371 	zfs_refcount_destroy(&abd->abd_children);
372 	ABDSTAT_BUMPDOWN(abdstat_linear_cnt);
373 	ABDSTAT_INCR(abdstat_linear_data_size, -(int)abd->abd_size);
374 
375 	abd_free_struct(abd);
376 }
377 
378 /*
379  * Free an ABD. Only use this on ABDs allocated with abd_alloc() or
380  * abd_alloc_linear().
381  */
382 void
383 abd_free(abd_t *abd)
384 {
385 	abd_verify(abd);
386 	ASSERT3P(abd->abd_parent, ==, NULL);
387 	ASSERT(abd->abd_flags & ABD_FLAG_OWNER);
388 	if (abd_is_linear(abd))
389 		abd_free_linear(abd);
390 	else
391 		abd_free_scatter(abd);
392 }
393 
394 /*
395  * Allocate an ABD of the same format (same metadata flag, same scatterize
396  * setting) as another ABD.
397  */
398 abd_t *
399 abd_alloc_sametype(abd_t *sabd, size_t size)
400 {
401 	boolean_t is_metadata = (sabd->abd_flags & ABD_FLAG_META) != 0;
402 	if (abd_is_linear(sabd)) {
403 		return (abd_alloc_linear(size, is_metadata));
404 	} else {
405 		return (abd_alloc(size, is_metadata));
406 	}
407 }
408 
409 /*
410  * If we're going to use this ABD for doing I/O using the block layer, the
411  * consumer of the ABD data doesn't care if it's scattered or not, and we don't
412  * plan to store this ABD in memory for a long period of time, we should
413  * allocate the ABD type that requires the least data copying to do the I/O.
414  *
415  * Currently this is linear ABDs, however if ldi_strategy() can ever issue I/Os
416  * using a scatter/gather list we should switch to that and replace this call
417  * with vanilla abd_alloc().
418  */
419 abd_t *
420 abd_alloc_for_io(size_t size, boolean_t is_metadata)
421 {
422 	return (abd_alloc_linear(size, is_metadata));
423 }
424 
425 /*
426  * Allocate a new ABD to point to offset off of sabd. It shares the underlying
427  * buffer data with sabd. Use abd_put() to free. sabd must not be freed while
428  * any derived ABDs exist.
429  */
430 /* ARGSUSED */
431 static inline abd_t *
432 abd_get_offset_impl(abd_t *sabd, size_t off, size_t size)
433 {
434 	abd_t *abd;
435 
436 	abd_verify(sabd);
437 	ASSERT3U(off, <=, sabd->abd_size);
438 
439 	if (abd_is_linear(sabd)) {
440 		abd = abd_alloc_struct(0);
441 
442 		/*
443 		 * Even if this buf is filesystem metadata, we only track that
444 		 * if we own the underlying data buffer, which is not true in
445 		 * this case. Therefore, we don't ever use ABD_FLAG_META here.
446 		 */
447 		abd->abd_flags = ABD_FLAG_LINEAR;
448 
449 		abd->abd_u.abd_linear.abd_buf =
450 		    (char *)sabd->abd_u.abd_linear.abd_buf + off;
451 	} else {
452 		size_t new_offset = sabd->abd_u.abd_scatter.abd_offset + off;
453 		size_t chunkcnt = abd_scatter_chunkcnt(sabd) -
454 		    (new_offset / zfs_abd_chunk_size);
455 
456 		abd = abd_alloc_struct(chunkcnt);
457 
458 		/*
459 		 * Even if this buf is filesystem metadata, we only track that
460 		 * if we own the underlying data buffer, which is not true in
461 		 * this case. Therefore, we don't ever use ABD_FLAG_META here.
462 		 */
463 		abd->abd_flags = 0;
464 
465 		abd->abd_u.abd_scatter.abd_offset =
466 		    new_offset % zfs_abd_chunk_size;
467 		abd->abd_u.abd_scatter.abd_chunk_size = zfs_abd_chunk_size;
468 
469 		/* Copy the scatterlist starting at the correct offset */
470 		(void) memcpy(&abd->abd_u.abd_scatter.abd_chunks,
471 		    &sabd->abd_u.abd_scatter.abd_chunks[new_offset /
472 		    zfs_abd_chunk_size],
473 		    chunkcnt * sizeof (void *));
474 	}
475 
476 	abd->abd_size = sabd->abd_size - off;
477 	abd->abd_parent = sabd;
478 	zfs_refcount_create(&abd->abd_children);
479 	(void) zfs_refcount_add_many(&sabd->abd_children, abd->abd_size, abd);
480 
481 	return (abd);
482 }
483 
484 abd_t *
485 abd_get_offset(abd_t *sabd, size_t off)
486 {
487 	size_t size = sabd->abd_size > off ? sabd->abd_size - off : 0;
488 
489 	VERIFY3U(size, >, 0);
490 
491 	return (abd_get_offset_impl(sabd, off, size));
492 }
493 
494 abd_t *
495 abd_get_offset_size(abd_t *sabd, size_t off, size_t size)
496 {
497 	ASSERT3U(off + size, <=, sabd->abd_size);
498 
499 	return (abd_get_offset_impl(sabd, off, size));
500 }
501 
502 
503 /*
504  * Allocate a linear ABD structure for buf. You must free this with abd_put()
505  * since the resulting ABD doesn't own its own buffer.
506  */
507 abd_t *
508 abd_get_from_buf(void *buf, size_t size)
509 {
510 	abd_t *abd = abd_alloc_struct(0);
511 
512 	VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
513 
514 	/*
515 	 * Even if this buf is filesystem metadata, we only track that if we
516 	 * own the underlying data buffer, which is not true in this case.
517 	 * Therefore, we don't ever use ABD_FLAG_META here.
518 	 */
519 	abd->abd_flags = ABD_FLAG_LINEAR;
520 	abd->abd_size = size;
521 	abd->abd_parent = NULL;
522 	zfs_refcount_create(&abd->abd_children);
523 
524 	abd->abd_u.abd_linear.abd_buf = buf;
525 
526 	return (abd);
527 }
528 
529 /*
530  * Free an ABD allocated from abd_get_offset() or abd_get_from_buf(). Will not
531  * free the underlying scatterlist or buffer.
532  */
533 void
534 abd_put(abd_t *abd)
535 {
536 	abd_verify(abd);
537 	ASSERT(!(abd->abd_flags & ABD_FLAG_OWNER));
538 
539 	if (abd->abd_parent != NULL) {
540 		(void) zfs_refcount_remove_many(&abd->abd_parent->abd_children,
541 		    abd->abd_size, abd);
542 	}
543 
544 	zfs_refcount_destroy(&abd->abd_children);
545 	abd_free_struct(abd);
546 }
547 
548 /*
549  * Get the raw buffer associated with a linear ABD.
550  */
551 void *
552 abd_to_buf(abd_t *abd)
553 {
554 	ASSERT(abd_is_linear(abd));
555 	abd_verify(abd);
556 	return (abd->abd_u.abd_linear.abd_buf);
557 }
558 
559 /*
560  * Borrow a raw buffer from an ABD without copying the contents of the ABD
561  * into the buffer. If the ABD is scattered, this will allocate a raw buffer
562  * whose contents are undefined. To copy over the existing data in the ABD, use
563  * abd_borrow_buf_copy() instead.
564  */
565 void *
566 abd_borrow_buf(abd_t *abd, size_t n)
567 {
568 	void *buf;
569 	abd_verify(abd);
570 	ASSERT3U(abd->abd_size, >=, n);
571 	if (abd_is_linear(abd)) {
572 		buf = abd_to_buf(abd);
573 	} else {
574 		buf = zio_buf_alloc(n);
575 	}
576 	(void) zfs_refcount_add_many(&abd->abd_children, n, buf);
577 
578 	return (buf);
579 }
580 
581 void *
582 abd_borrow_buf_copy(abd_t *abd, size_t n)
583 {
584 	void *buf = abd_borrow_buf(abd, n);
585 	if (!abd_is_linear(abd)) {
586 		abd_copy_to_buf(buf, abd, n);
587 	}
588 	return (buf);
589 }
590 
591 /*
592  * Return a borrowed raw buffer to an ABD. If the ABD is scattered, this will
593  * not change the contents of the ABD and will ASSERT that you didn't modify
594  * the buffer since it was borrowed. If you want any changes you made to buf to
595  * be copied back to abd, use abd_return_buf_copy() instead.
596  */
597 void
598 abd_return_buf(abd_t *abd, void *buf, size_t n)
599 {
600 	abd_verify(abd);
601 	ASSERT3U(abd->abd_size, >=, n);
602 	if (abd_is_linear(abd)) {
603 		ASSERT3P(buf, ==, abd_to_buf(abd));
604 	} else {
605 		ASSERT0(abd_cmp_buf(abd, buf, n));
606 		zio_buf_free(buf, n);
607 	}
608 	(void) zfs_refcount_remove_many(&abd->abd_children, n, buf);
609 }
610 
611 void
612 abd_return_buf_copy(abd_t *abd, void *buf, size_t n)
613 {
614 	if (!abd_is_linear(abd)) {
615 		abd_copy_from_buf(abd, buf, n);
616 	}
617 	abd_return_buf(abd, buf, n);
618 }
619 
620 /*
621  * Give this ABD ownership of the buffer that it's storing. Can only be used on
622  * linear ABDs which were allocated via abd_get_from_buf(), or ones allocated
623  * with abd_alloc_linear() which subsequently released ownership of their buf
624  * with abd_release_ownership_of_buf().
625  */
626 void
627 abd_take_ownership_of_buf(abd_t *abd, boolean_t is_metadata)
628 {
629 	ASSERT(abd_is_linear(abd));
630 	ASSERT(!(abd->abd_flags & ABD_FLAG_OWNER));
631 	abd_verify(abd);
632 
633 	abd->abd_flags |= ABD_FLAG_OWNER;
634 	if (is_metadata) {
635 		abd->abd_flags |= ABD_FLAG_META;
636 	}
637 
638 	ABDSTAT_BUMP(abdstat_linear_cnt);
639 	ABDSTAT_INCR(abdstat_linear_data_size, abd->abd_size);
640 }
641 
642 void
643 abd_release_ownership_of_buf(abd_t *abd)
644 {
645 	ASSERT(abd_is_linear(abd));
646 	ASSERT(abd->abd_flags & ABD_FLAG_OWNER);
647 	abd_verify(abd);
648 
649 	abd->abd_flags &= ~ABD_FLAG_OWNER;
650 	/* Disable this flag since we no longer own the data buffer */
651 	abd->abd_flags &= ~ABD_FLAG_META;
652 
653 	ABDSTAT_BUMPDOWN(abdstat_linear_cnt);
654 	ABDSTAT_INCR(abdstat_linear_data_size, -(int)abd->abd_size);
655 }
656 
657 struct abd_iter {
658 	abd_t		*iter_abd;	/* ABD being iterated through */
659 	size_t		iter_pos;	/* position (relative to abd_offset) */
660 	void		*iter_mapaddr;	/* addr corresponding to iter_pos */
661 	size_t		iter_mapsize;	/* length of data valid at mapaddr */
662 };
663 
664 static inline size_t
665 abd_iter_scatter_chunk_offset(struct abd_iter *aiter)
666 {
667 	ASSERT(!abd_is_linear(aiter->iter_abd));
668 	return ((aiter->iter_abd->abd_u.abd_scatter.abd_offset +
669 	    aiter->iter_pos) % zfs_abd_chunk_size);
670 }
671 
672 static inline size_t
673 abd_iter_scatter_chunk_index(struct abd_iter *aiter)
674 {
675 	ASSERT(!abd_is_linear(aiter->iter_abd));
676 	return ((aiter->iter_abd->abd_u.abd_scatter.abd_offset +
677 	    aiter->iter_pos) / zfs_abd_chunk_size);
678 }
679 
680 /*
681  * Initialize the abd_iter.
682  */
683 static void
684 abd_iter_init(struct abd_iter *aiter, abd_t *abd)
685 {
686 	abd_verify(abd);
687 	aiter->iter_abd = abd;
688 	aiter->iter_pos = 0;
689 	aiter->iter_mapaddr = NULL;
690 	aiter->iter_mapsize = 0;
691 }
692 
693 /*
694  * Advance the iterator by a certain amount. Cannot be called when a chunk is
695  * in use. This can be safely called when the aiter has already exhausted, in
696  * which case this does nothing.
697  */
698 static void
699 abd_iter_advance(struct abd_iter *aiter, size_t amount)
700 {
701 	ASSERT3P(aiter->iter_mapaddr, ==, NULL);
702 	ASSERT0(aiter->iter_mapsize);
703 
704 	/* There's nothing left to advance to, so do nothing */
705 	if (aiter->iter_pos == aiter->iter_abd->abd_size)
706 		return;
707 
708 	aiter->iter_pos += amount;
709 }
710 
711 /*
712  * Map the current chunk into aiter. This can be safely called when the aiter
713  * has already exhausted, in which case this does nothing.
714  */
715 static void
716 abd_iter_map(struct abd_iter *aiter)
717 {
718 	void *paddr;
719 	size_t offset = 0;
720 
721 	ASSERT3P(aiter->iter_mapaddr, ==, NULL);
722 	ASSERT0(aiter->iter_mapsize);
723 
724 	/* Panic if someone has changed zfs_abd_chunk_size */
725 	IMPLY(!abd_is_linear(aiter->iter_abd), zfs_abd_chunk_size ==
726 	    aiter->iter_abd->abd_u.abd_scatter.abd_chunk_size);
727 
728 	/* There's nothing left to iterate over, so do nothing */
729 	if (aiter->iter_pos == aiter->iter_abd->abd_size)
730 		return;
731 
732 	if (abd_is_linear(aiter->iter_abd)) {
733 		offset = aiter->iter_pos;
734 		aiter->iter_mapsize = aiter->iter_abd->abd_size - offset;
735 		paddr = aiter->iter_abd->abd_u.abd_linear.abd_buf;
736 	} else {
737 		size_t index = abd_iter_scatter_chunk_index(aiter);
738 		offset = abd_iter_scatter_chunk_offset(aiter);
739 		aiter->iter_mapsize = zfs_abd_chunk_size - offset;
740 		paddr = aiter->iter_abd->abd_u.abd_scatter.abd_chunks[index];
741 	}
742 	aiter->iter_mapaddr = (char *)paddr + offset;
743 }
744 
745 /*
746  * Unmap the current chunk from aiter. This can be safely called when the aiter
747  * has already exhausted, in which case this does nothing.
748  */
749 static void
750 abd_iter_unmap(struct abd_iter *aiter)
751 {
752 	/* There's nothing left to unmap, so do nothing */
753 	if (aiter->iter_pos == aiter->iter_abd->abd_size)
754 		return;
755 
756 	ASSERT3P(aiter->iter_mapaddr, !=, NULL);
757 	ASSERT3U(aiter->iter_mapsize, >, 0);
758 
759 	aiter->iter_mapaddr = NULL;
760 	aiter->iter_mapsize = 0;
761 }
762 
763 int
764 abd_iterate_func(abd_t *abd, size_t off, size_t size,
765     abd_iter_func_t *func, void *private)
766 {
767 	int ret = 0;
768 	struct abd_iter aiter;
769 
770 	abd_verify(abd);
771 	ASSERT3U(off + size, <=, abd->abd_size);
772 
773 	abd_iter_init(&aiter, abd);
774 	abd_iter_advance(&aiter, off);
775 
776 	while (size > 0) {
777 		abd_iter_map(&aiter);
778 
779 		size_t len = MIN(aiter.iter_mapsize, size);
780 		ASSERT3U(len, >, 0);
781 
782 		ret = func(aiter.iter_mapaddr, len, private);
783 
784 		abd_iter_unmap(&aiter);
785 
786 		if (ret != 0)
787 			break;
788 
789 		size -= len;
790 		abd_iter_advance(&aiter, len);
791 	}
792 
793 	return (ret);
794 }
795 
796 struct buf_arg {
797 	void *arg_buf;
798 };
799 
800 static int
801 abd_copy_to_buf_off_cb(void *buf, size_t size, void *private)
802 {
803 	struct buf_arg *ba_ptr = private;
804 
805 	(void) memcpy(ba_ptr->arg_buf, buf, size);
806 	ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size;
807 
808 	return (0);
809 }
810 
811 /*
812  * Copy abd to buf. (off is the offset in abd.)
813  */
814 void
815 abd_copy_to_buf_off(void *buf, abd_t *abd, size_t off, size_t size)
816 {
817 	struct buf_arg ba_ptr = { buf };
818 
819 	(void) abd_iterate_func(abd, off, size, abd_copy_to_buf_off_cb,
820 	    &ba_ptr);
821 }
822 
823 static int
824 abd_cmp_buf_off_cb(void *buf, size_t size, void *private)
825 {
826 	int ret;
827 	struct buf_arg *ba_ptr = private;
828 
829 	ret = memcmp(buf, ba_ptr->arg_buf, size);
830 	ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size;
831 
832 	return (ret);
833 }
834 
835 /*
836  * Compare the contents of abd to buf. (off is the offset in abd.)
837  */
838 int
839 abd_cmp_buf_off(abd_t *abd, const void *buf, size_t off, size_t size)
840 {
841 	struct buf_arg ba_ptr = { (void *) buf };
842 
843 	return (abd_iterate_func(abd, off, size, abd_cmp_buf_off_cb, &ba_ptr));
844 }
845 
846 static int
847 abd_copy_from_buf_off_cb(void *buf, size_t size, void *private)
848 {
849 	struct buf_arg *ba_ptr = private;
850 
851 	(void) memcpy(buf, ba_ptr->arg_buf, size);
852 	ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size;
853 
854 	return (0);
855 }
856 
857 /*
858  * Copy from buf to abd. (off is the offset in abd.)
859  */
860 void
861 abd_copy_from_buf_off(abd_t *abd, const void *buf, size_t off, size_t size)
862 {
863 	struct buf_arg ba_ptr = { (void *) buf };
864 
865 	(void) abd_iterate_func(abd, off, size, abd_copy_from_buf_off_cb,
866 	    &ba_ptr);
867 }
868 
869 /*ARGSUSED*/
870 static int
871 abd_zero_off_cb(void *buf, size_t size, void *private)
872 {
873 	(void) memset(buf, 0, size);
874 	return (0);
875 }
876 
877 /*
878  * Zero out the abd from a particular offset to the end.
879  */
880 void
881 abd_zero_off(abd_t *abd, size_t off, size_t size)
882 {
883 	(void) abd_iterate_func(abd, off, size, abd_zero_off_cb, NULL);
884 }
885 
886 /*
887  * Iterate over two ABDs and call func incrementally on the two ABDs' data in
888  * equal-sized chunks (passed to func as raw buffers). func could be called many
889  * times during this iteration.
890  */
891 int
892 abd_iterate_func2(abd_t *dabd, abd_t *sabd, size_t doff, size_t soff,
893     size_t size, abd_iter_func2_t *func, void *private)
894 {
895 	int ret = 0;
896 	struct abd_iter daiter, saiter;
897 
898 	abd_verify(dabd);
899 	abd_verify(sabd);
900 
901 	ASSERT3U(doff + size, <=, dabd->abd_size);
902 	ASSERT3U(soff + size, <=, sabd->abd_size);
903 
904 	abd_iter_init(&daiter, dabd);
905 	abd_iter_init(&saiter, sabd);
906 	abd_iter_advance(&daiter, doff);
907 	abd_iter_advance(&saiter, soff);
908 
909 	while (size > 0) {
910 		abd_iter_map(&daiter);
911 		abd_iter_map(&saiter);
912 
913 		size_t dlen = MIN(daiter.iter_mapsize, size);
914 		size_t slen = MIN(saiter.iter_mapsize, size);
915 		size_t len = MIN(dlen, slen);
916 		ASSERT(dlen > 0 || slen > 0);
917 
918 		ret = func(daiter.iter_mapaddr, saiter.iter_mapaddr, len,
919 		    private);
920 
921 		abd_iter_unmap(&saiter);
922 		abd_iter_unmap(&daiter);
923 
924 		if (ret != 0)
925 			break;
926 
927 		size -= len;
928 		abd_iter_advance(&daiter, len);
929 		abd_iter_advance(&saiter, len);
930 	}
931 
932 	return (ret);
933 }
934 
935 /*ARGSUSED*/
936 static int
937 abd_copy_off_cb(void *dbuf, void *sbuf, size_t size, void *private)
938 {
939 	(void) memcpy(dbuf, sbuf, size);
940 	return (0);
941 }
942 
943 /*
944  * Copy from sabd to dabd starting from soff and doff.
945  */
946 void
947 abd_copy_off(abd_t *dabd, abd_t *sabd, size_t doff, size_t soff, size_t size)
948 {
949 	(void) abd_iterate_func2(dabd, sabd, doff, soff, size,
950 	    abd_copy_off_cb, NULL);
951 }
952 
953 /*ARGSUSED*/
954 static int
955 abd_cmp_cb(void *bufa, void *bufb, size_t size, void *private)
956 {
957 	return (memcmp(bufa, bufb, size));
958 }
959 
960 /*
961  * Compares the first size bytes of two ABDs.
962  */
963 int
964 abd_cmp(abd_t *dabd, abd_t *sabd, size_t size)
965 {
966 	return (abd_iterate_func2(dabd, sabd, 0, 0, size, abd_cmp_cb, NULL));
967 }
968