xref: /illumos-gate/usr/src/uts/common/fs/zfs/abd.c (revision 0dbcca9b391b6a95b983f29699611dedbcb5d262)
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 static void *
161 abd_alloc_chunk()
162 {
163 	void *c = kmem_cache_alloc(abd_chunk_cache, KM_PUSHPAGE);
164 	ASSERT3P(c, !=, NULL);
165 	return (c);
166 }
167 
168 static void
169 abd_free_chunk(void *c)
170 {
171 	kmem_cache_free(abd_chunk_cache, c);
172 }
173 
174 void
175 abd_init(void)
176 {
177 	vmem_t *data_alloc_arena = NULL;
178 
179 #ifdef _KERNEL
180 	data_alloc_arena = zio_alloc_arena;
181 #endif
182 
183 	/*
184 	 * Since ABD chunks do not appear in crash dumps, we pass KMC_NOTOUCH
185 	 * so that no allocator metadata is stored with the buffers.
186 	 */
187 	abd_chunk_cache = kmem_cache_create("abd_chunk", zfs_abd_chunk_size, 0,
188 	    NULL, NULL, NULL, NULL, data_alloc_arena, KMC_NOTOUCH);
189 
190 	abd_ksp = kstat_create("zfs", 0, "abdstats", "misc", KSTAT_TYPE_NAMED,
191 	    sizeof (abd_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
192 	if (abd_ksp != NULL) {
193 		abd_ksp->ks_data = &abd_stats;
194 		kstat_install(abd_ksp);
195 	}
196 }
197 
198 void
199 abd_fini(void)
200 {
201 	if (abd_ksp != NULL) {
202 		kstat_delete(abd_ksp);
203 		abd_ksp = NULL;
204 	}
205 
206 	kmem_cache_destroy(abd_chunk_cache);
207 	abd_chunk_cache = NULL;
208 }
209 
210 static inline size_t
211 abd_chunkcnt_for_bytes(size_t size)
212 {
213 	return (P2ROUNDUP(size, zfs_abd_chunk_size) / zfs_abd_chunk_size);
214 }
215 
216 static inline size_t
217 abd_scatter_chunkcnt(abd_t *abd)
218 {
219 	ASSERT(!abd_is_linear(abd));
220 	return (abd_chunkcnt_for_bytes(
221 	    abd->abd_u.abd_scatter.abd_offset + abd->abd_size));
222 }
223 
224 static inline void
225 abd_verify(abd_t *abd)
226 {
227 	ASSERT3U(abd->abd_size, >, 0);
228 	ASSERT3U(abd->abd_size, <=, SPA_MAXBLOCKSIZE);
229 	ASSERT3U(abd->abd_flags, ==, abd->abd_flags & (ABD_FLAG_LINEAR |
230 	    ABD_FLAG_OWNER | ABD_FLAG_META));
231 	IMPLY(abd->abd_parent != NULL, !(abd->abd_flags & ABD_FLAG_OWNER));
232 	IMPLY(abd->abd_flags & ABD_FLAG_META, abd->abd_flags & ABD_FLAG_OWNER);
233 	if (abd_is_linear(abd)) {
234 		ASSERT3P(abd->abd_u.abd_linear.abd_buf, !=, NULL);
235 	} else {
236 		ASSERT3U(abd->abd_u.abd_scatter.abd_offset, <,
237 		    zfs_abd_chunk_size);
238 		size_t n = abd_scatter_chunkcnt(abd);
239 		for (int i = 0; i < n; i++) {
240 			ASSERT3P(
241 			    abd->abd_u.abd_scatter.abd_chunks[i], !=, NULL);
242 		}
243 	}
244 }
245 
246 static inline abd_t *
247 abd_alloc_struct(size_t chunkcnt)
248 {
249 	size_t size = offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]);
250 	abd_t *abd = kmem_alloc(size, KM_PUSHPAGE);
251 	ASSERT3P(abd, !=, NULL);
252 	ABDSTAT_INCR(abdstat_struct_size, size);
253 
254 	return (abd);
255 }
256 
257 static inline void
258 abd_free_struct(abd_t *abd)
259 {
260 	size_t chunkcnt = abd_is_linear(abd) ? 0 : abd_scatter_chunkcnt(abd);
261 	int size = offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]);
262 	kmem_free(abd, size);
263 	ABDSTAT_INCR(abdstat_struct_size, -size);
264 }
265 
266 /*
267  * Allocate an ABD, along with its own underlying data buffers. Use this if you
268  * don't care whether the ABD is linear or not.
269  */
270 abd_t *
271 abd_alloc(size_t size, boolean_t is_metadata)
272 {
273 	if (!zfs_abd_scatter_enabled)
274 		return (abd_alloc_linear(size, is_metadata));
275 
276 	VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
277 
278 	size_t n = abd_chunkcnt_for_bytes(size);
279 	abd_t *abd = abd_alloc_struct(n);
280 
281 	abd->abd_flags = ABD_FLAG_OWNER;
282 	if (is_metadata) {
283 		abd->abd_flags |= ABD_FLAG_META;
284 	}
285 	abd->abd_size = size;
286 	abd->abd_parent = NULL;
287 	refcount_create(&abd->abd_children);
288 
289 	abd->abd_u.abd_scatter.abd_offset = 0;
290 	abd->abd_u.abd_scatter.abd_chunk_size = zfs_abd_chunk_size;
291 
292 	for (int i = 0; i < n; i++) {
293 		void *c = abd_alloc_chunk();
294 		ASSERT3P(c, !=, NULL);
295 		abd->abd_u.abd_scatter.abd_chunks[i] = c;
296 	}
297 
298 	ABDSTAT_BUMP(abdstat_scatter_cnt);
299 	ABDSTAT_INCR(abdstat_scatter_data_size, size);
300 	ABDSTAT_INCR(abdstat_scatter_chunk_waste,
301 	    n * zfs_abd_chunk_size - size);
302 
303 	return (abd);
304 }
305 
306 static void
307 abd_free_scatter(abd_t *abd)
308 {
309 	size_t n = abd_scatter_chunkcnt(abd);
310 	for (int i = 0; i < n; i++) {
311 		abd_free_chunk(abd->abd_u.abd_scatter.abd_chunks[i]);
312 	}
313 
314 	refcount_destroy(&abd->abd_children);
315 	ABDSTAT_BUMPDOWN(abdstat_scatter_cnt);
316 	ABDSTAT_INCR(abdstat_scatter_data_size, -(int)abd->abd_size);
317 	ABDSTAT_INCR(abdstat_scatter_chunk_waste,
318 	    abd->abd_size - n * zfs_abd_chunk_size);
319 
320 	abd_free_struct(abd);
321 }
322 
323 /*
324  * Allocate an ABD that must be linear, along with its own underlying data
325  * buffer. Only use this when it would be very annoying to write your ABD
326  * consumer with a scattered ABD.
327  */
328 abd_t *
329 abd_alloc_linear(size_t size, boolean_t is_metadata)
330 {
331 	abd_t *abd = abd_alloc_struct(0);
332 
333 	VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
334 
335 	abd->abd_flags = ABD_FLAG_LINEAR | ABD_FLAG_OWNER;
336 	if (is_metadata) {
337 		abd->abd_flags |= ABD_FLAG_META;
338 	}
339 	abd->abd_size = size;
340 	abd->abd_parent = NULL;
341 	refcount_create(&abd->abd_children);
342 
343 	if (is_metadata) {
344 		abd->abd_u.abd_linear.abd_buf = zio_buf_alloc(size);
345 	} else {
346 		abd->abd_u.abd_linear.abd_buf = zio_data_buf_alloc(size);
347 	}
348 
349 	ABDSTAT_BUMP(abdstat_linear_cnt);
350 	ABDSTAT_INCR(abdstat_linear_data_size, size);
351 
352 	return (abd);
353 }
354 
355 static void
356 abd_free_linear(abd_t *abd)
357 {
358 	if (abd->abd_flags & ABD_FLAG_META) {
359 		zio_buf_free(abd->abd_u.abd_linear.abd_buf, abd->abd_size);
360 	} else {
361 		zio_data_buf_free(abd->abd_u.abd_linear.abd_buf, abd->abd_size);
362 	}
363 
364 	refcount_destroy(&abd->abd_children);
365 	ABDSTAT_BUMPDOWN(abdstat_linear_cnt);
366 	ABDSTAT_INCR(abdstat_linear_data_size, -(int)abd->abd_size);
367 
368 	abd_free_struct(abd);
369 }
370 
371 /*
372  * Free an ABD. Only use this on ABDs allocated with abd_alloc() or
373  * abd_alloc_linear().
374  */
375 void
376 abd_free(abd_t *abd)
377 {
378 	abd_verify(abd);
379 	ASSERT3P(abd->abd_parent, ==, NULL);
380 	ASSERT(abd->abd_flags & ABD_FLAG_OWNER);
381 	if (abd_is_linear(abd))
382 		abd_free_linear(abd);
383 	else
384 		abd_free_scatter(abd);
385 }
386 
387 /*
388  * Allocate an ABD of the same format (same metadata flag, same scatterize
389  * setting) as another ABD.
390  */
391 abd_t *
392 abd_alloc_sametype(abd_t *sabd, size_t size)
393 {
394 	boolean_t is_metadata = (sabd->abd_flags & ABD_FLAG_META) != 0;
395 	if (abd_is_linear(sabd)) {
396 		return (abd_alloc_linear(size, is_metadata));
397 	} else {
398 		return (abd_alloc(size, is_metadata));
399 	}
400 }
401 
402 /*
403  * If we're going to use this ABD for doing I/O using the block layer, the
404  * consumer of the ABD data doesn't care if it's scattered or not, and we don't
405  * plan to store this ABD in memory for a long period of time, we should
406  * allocate the ABD type that requires the least data copying to do the I/O.
407  *
408  * Currently this is linear ABDs, however if ldi_strategy() can ever issue I/Os
409  * using a scatter/gather list we should switch to that and replace this call
410  * with vanilla abd_alloc().
411  */
412 abd_t *
413 abd_alloc_for_io(size_t size, boolean_t is_metadata)
414 {
415 	return (abd_alloc_linear(size, is_metadata));
416 }
417 
418 /*
419  * Allocate a new ABD to point to offset off of sabd. It shares the underlying
420  * buffer data with sabd. Use abd_put() to free. sabd must not be freed while
421  * any derived ABDs exist.
422  */
423 abd_t *
424 abd_get_offset(abd_t *sabd, size_t off)
425 {
426 	abd_t *abd;
427 
428 	abd_verify(sabd);
429 	ASSERT3U(off, <=, sabd->abd_size);
430 
431 	if (abd_is_linear(sabd)) {
432 		abd = abd_alloc_struct(0);
433 
434 		/*
435 		 * Even if this buf is filesystem metadata, we only track that
436 		 * if we own the underlying data buffer, which is not true in
437 		 * this case. Therefore, we don't ever use ABD_FLAG_META here.
438 		 */
439 		abd->abd_flags = ABD_FLAG_LINEAR;
440 
441 		abd->abd_u.abd_linear.abd_buf =
442 		    (char *)sabd->abd_u.abd_linear.abd_buf + off;
443 	} else {
444 		size_t new_offset = sabd->abd_u.abd_scatter.abd_offset + off;
445 		size_t chunkcnt = abd_scatter_chunkcnt(sabd) -
446 		    (new_offset / zfs_abd_chunk_size);
447 
448 		abd = abd_alloc_struct(chunkcnt);
449 
450 		/*
451 		 * Even if this buf is filesystem metadata, we only track that
452 		 * if we own the underlying data buffer, which is not true in
453 		 * this case. Therefore, we don't ever use ABD_FLAG_META here.
454 		 */
455 		abd->abd_flags = 0;
456 
457 		abd->abd_u.abd_scatter.abd_offset =
458 		    new_offset % zfs_abd_chunk_size;
459 		abd->abd_u.abd_scatter.abd_chunk_size = zfs_abd_chunk_size;
460 
461 		/* Copy the scatterlist starting at the correct offset */
462 		(void) memcpy(&abd->abd_u.abd_scatter.abd_chunks,
463 		    &sabd->abd_u.abd_scatter.abd_chunks[new_offset /
464 		    zfs_abd_chunk_size],
465 		    chunkcnt * sizeof (void *));
466 	}
467 
468 	abd->abd_size = sabd->abd_size - off;
469 	abd->abd_parent = sabd;
470 	refcount_create(&abd->abd_children);
471 	(void) refcount_add_many(&sabd->abd_children, abd->abd_size, abd);
472 
473 	return (abd);
474 }
475 
476 /*
477  * Allocate a linear ABD structure for buf. You must free this with abd_put()
478  * since the resulting ABD doesn't own its own buffer.
479  */
480 abd_t *
481 abd_get_from_buf(void *buf, size_t size)
482 {
483 	abd_t *abd = abd_alloc_struct(0);
484 
485 	VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
486 
487 	/*
488 	 * Even if this buf is filesystem metadata, we only track that if we
489 	 * own the underlying data buffer, which is not true in this case.
490 	 * Therefore, we don't ever use ABD_FLAG_META here.
491 	 */
492 	abd->abd_flags = ABD_FLAG_LINEAR;
493 	abd->abd_size = size;
494 	abd->abd_parent = NULL;
495 	refcount_create(&abd->abd_children);
496 
497 	abd->abd_u.abd_linear.abd_buf = buf;
498 
499 	return (abd);
500 }
501 
502 /*
503  * Free an ABD allocated from abd_get_offset() or abd_get_from_buf(). Will not
504  * free the underlying scatterlist or buffer.
505  */
506 void
507 abd_put(abd_t *abd)
508 {
509 	abd_verify(abd);
510 	ASSERT(!(abd->abd_flags & ABD_FLAG_OWNER));
511 
512 	if (abd->abd_parent != NULL) {
513 		(void) refcount_remove_many(&abd->abd_parent->abd_children,
514 		    abd->abd_size, abd);
515 	}
516 
517 	refcount_destroy(&abd->abd_children);
518 	abd_free_struct(abd);
519 }
520 
521 /*
522  * Get the raw buffer associated with a linear ABD.
523  */
524 void *
525 abd_to_buf(abd_t *abd)
526 {
527 	ASSERT(abd_is_linear(abd));
528 	abd_verify(abd);
529 	return (abd->abd_u.abd_linear.abd_buf);
530 }
531 
532 /*
533  * Borrow a raw buffer from an ABD without copying the contents of the ABD
534  * into the buffer. If the ABD is scattered, this will allocate a raw buffer
535  * whose contents are undefined. To copy over the existing data in the ABD, use
536  * abd_borrow_buf_copy() instead.
537  */
538 void *
539 abd_borrow_buf(abd_t *abd, size_t n)
540 {
541 	void *buf;
542 	abd_verify(abd);
543 	ASSERT3U(abd->abd_size, >=, n);
544 	if (abd_is_linear(abd)) {
545 		buf = abd_to_buf(abd);
546 	} else {
547 		buf = zio_buf_alloc(n);
548 	}
549 	(void) refcount_add_many(&abd->abd_children, n, buf);
550 
551 	return (buf);
552 }
553 
554 void *
555 abd_borrow_buf_copy(abd_t *abd, size_t n)
556 {
557 	void *buf = abd_borrow_buf(abd, n);
558 	if (!abd_is_linear(abd)) {
559 		abd_copy_to_buf(buf, abd, n);
560 	}
561 	return (buf);
562 }
563 
564 /*
565  * Return a borrowed raw buffer to an ABD. If the ABD is scattered, this will
566  * not change the contents of the ABD and will ASSERT that you didn't modify
567  * the buffer since it was borrowed. If you want any changes you made to buf to
568  * be copied back to abd, use abd_return_buf_copy() instead.
569  */
570 void
571 abd_return_buf(abd_t *abd, void *buf, size_t n)
572 {
573 	abd_verify(abd);
574 	ASSERT3U(abd->abd_size, >=, n);
575 	if (abd_is_linear(abd)) {
576 		ASSERT3P(buf, ==, abd_to_buf(abd));
577 	} else {
578 		ASSERT0(abd_cmp_buf(abd, buf, n));
579 		zio_buf_free(buf, n);
580 	}
581 	(void) refcount_remove_many(&abd->abd_children, n, buf);
582 }
583 
584 void
585 abd_return_buf_copy(abd_t *abd, void *buf, size_t n)
586 {
587 	if (!abd_is_linear(abd)) {
588 		abd_copy_from_buf(abd, buf, n);
589 	}
590 	abd_return_buf(abd, buf, n);
591 }
592 
593 /*
594  * Give this ABD ownership of the buffer that it's storing. Can only be used on
595  * linear ABDs which were allocated via abd_get_from_buf(), or ones allocated
596  * with abd_alloc_linear() which subsequently released ownership of their buf
597  * with abd_release_ownership_of_buf().
598  */
599 void
600 abd_take_ownership_of_buf(abd_t *abd, boolean_t is_metadata)
601 {
602 	ASSERT(abd_is_linear(abd));
603 	ASSERT(!(abd->abd_flags & ABD_FLAG_OWNER));
604 	abd_verify(abd);
605 
606 	abd->abd_flags |= ABD_FLAG_OWNER;
607 	if (is_metadata) {
608 		abd->abd_flags |= ABD_FLAG_META;
609 	}
610 
611 	ABDSTAT_BUMP(abdstat_linear_cnt);
612 	ABDSTAT_INCR(abdstat_linear_data_size, abd->abd_size);
613 }
614 
615 void
616 abd_release_ownership_of_buf(abd_t *abd)
617 {
618 	ASSERT(abd_is_linear(abd));
619 	ASSERT(abd->abd_flags & ABD_FLAG_OWNER);
620 	abd_verify(abd);
621 
622 	abd->abd_flags &= ~ABD_FLAG_OWNER;
623 	/* Disable this flag since we no longer own the data buffer */
624 	abd->abd_flags &= ~ABD_FLAG_META;
625 
626 	ABDSTAT_BUMPDOWN(abdstat_linear_cnt);
627 	ABDSTAT_INCR(abdstat_linear_data_size, -(int)abd->abd_size);
628 }
629 
630 struct abd_iter {
631 	abd_t		*iter_abd;	/* ABD being iterated through */
632 	size_t		iter_pos;	/* position (relative to abd_offset) */
633 	void		*iter_mapaddr;	/* addr corresponding to iter_pos */
634 	size_t		iter_mapsize;	/* length of data valid at mapaddr */
635 };
636 
637 static inline size_t
638 abd_iter_scatter_chunk_offset(struct abd_iter *aiter)
639 {
640 	ASSERT(!abd_is_linear(aiter->iter_abd));
641 	return ((aiter->iter_abd->abd_u.abd_scatter.abd_offset +
642 	    aiter->iter_pos) % zfs_abd_chunk_size);
643 }
644 
645 static inline size_t
646 abd_iter_scatter_chunk_index(struct abd_iter *aiter)
647 {
648 	ASSERT(!abd_is_linear(aiter->iter_abd));
649 	return ((aiter->iter_abd->abd_u.abd_scatter.abd_offset +
650 	    aiter->iter_pos) / zfs_abd_chunk_size);
651 }
652 
653 /*
654  * Initialize the abd_iter.
655  */
656 static void
657 abd_iter_init(struct abd_iter *aiter, abd_t *abd)
658 {
659 	abd_verify(abd);
660 	aiter->iter_abd = abd;
661 	aiter->iter_pos = 0;
662 	aiter->iter_mapaddr = NULL;
663 	aiter->iter_mapsize = 0;
664 }
665 
666 /*
667  * Advance the iterator by a certain amount. Cannot be called when a chunk is
668  * in use. This can be safely called when the aiter has already exhausted, in
669  * which case this does nothing.
670  */
671 static void
672 abd_iter_advance(struct abd_iter *aiter, size_t amount)
673 {
674 	ASSERT3P(aiter->iter_mapaddr, ==, NULL);
675 	ASSERT0(aiter->iter_mapsize);
676 
677 	/* There's nothing left to advance to, so do nothing */
678 	if (aiter->iter_pos == aiter->iter_abd->abd_size)
679 		return;
680 
681 	aiter->iter_pos += amount;
682 }
683 
684 /*
685  * Map the current chunk into aiter. This can be safely called when the aiter
686  * has already exhausted, in which case this does nothing.
687  */
688 static void
689 abd_iter_map(struct abd_iter *aiter)
690 {
691 	void *paddr;
692 	size_t offset = 0;
693 
694 	ASSERT3P(aiter->iter_mapaddr, ==, NULL);
695 	ASSERT0(aiter->iter_mapsize);
696 
697 	/* Panic if someone has changed zfs_abd_chunk_size */
698 	IMPLY(!abd_is_linear(aiter->iter_abd), zfs_abd_chunk_size ==
699 	    aiter->iter_abd->abd_u.abd_scatter.abd_chunk_size);
700 
701 	/* There's nothing left to iterate over, so do nothing */
702 	if (aiter->iter_pos == aiter->iter_abd->abd_size)
703 		return;
704 
705 	if (abd_is_linear(aiter->iter_abd)) {
706 		offset = aiter->iter_pos;
707 		aiter->iter_mapsize = aiter->iter_abd->abd_size - offset;
708 		paddr = aiter->iter_abd->abd_u.abd_linear.abd_buf;
709 	} else {
710 		size_t index = abd_iter_scatter_chunk_index(aiter);
711 		offset = abd_iter_scatter_chunk_offset(aiter);
712 		aiter->iter_mapsize = zfs_abd_chunk_size - offset;
713 		paddr = aiter->iter_abd->abd_u.abd_scatter.abd_chunks[index];
714 	}
715 	aiter->iter_mapaddr = (char *)paddr + offset;
716 }
717 
718 /*
719  * Unmap the current chunk from aiter. This can be safely called when the aiter
720  * has already exhausted, in which case this does nothing.
721  */
722 static void
723 abd_iter_unmap(struct abd_iter *aiter)
724 {
725 	/* There's nothing left to unmap, so do nothing */
726 	if (aiter->iter_pos == aiter->iter_abd->abd_size)
727 		return;
728 
729 	ASSERT3P(aiter->iter_mapaddr, !=, NULL);
730 	ASSERT3U(aiter->iter_mapsize, >, 0);
731 
732 	aiter->iter_mapaddr = NULL;
733 	aiter->iter_mapsize = 0;
734 }
735 
736 int
737 abd_iterate_func(abd_t *abd, size_t off, size_t size,
738     abd_iter_func_t *func, void *private)
739 {
740 	int ret = 0;
741 	struct abd_iter aiter;
742 
743 	abd_verify(abd);
744 	ASSERT3U(off + size, <=, abd->abd_size);
745 
746 	abd_iter_init(&aiter, abd);
747 	abd_iter_advance(&aiter, off);
748 
749 	while (size > 0) {
750 		abd_iter_map(&aiter);
751 
752 		size_t len = MIN(aiter.iter_mapsize, size);
753 		ASSERT3U(len, >, 0);
754 
755 		ret = func(aiter.iter_mapaddr, len, private);
756 
757 		abd_iter_unmap(&aiter);
758 
759 		if (ret != 0)
760 			break;
761 
762 		size -= len;
763 		abd_iter_advance(&aiter, len);
764 	}
765 
766 	return (ret);
767 }
768 
769 struct buf_arg {
770 	void *arg_buf;
771 };
772 
773 static int
774 abd_copy_to_buf_off_cb(void *buf, size_t size, void *private)
775 {
776 	struct buf_arg *ba_ptr = private;
777 
778 	(void) memcpy(ba_ptr->arg_buf, buf, size);
779 	ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size;
780 
781 	return (0);
782 }
783 
784 /*
785  * Copy abd to buf. (off is the offset in abd.)
786  */
787 void
788 abd_copy_to_buf_off(void *buf, abd_t *abd, size_t off, size_t size)
789 {
790 	struct buf_arg ba_ptr = { buf };
791 
792 	(void) abd_iterate_func(abd, off, size, abd_copy_to_buf_off_cb,
793 	    &ba_ptr);
794 }
795 
796 static int
797 abd_cmp_buf_off_cb(void *buf, size_t size, void *private)
798 {
799 	int ret;
800 	struct buf_arg *ba_ptr = private;
801 
802 	ret = memcmp(buf, ba_ptr->arg_buf, size);
803 	ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size;
804 
805 	return (ret);
806 }
807 
808 /*
809  * Compare the contents of abd to buf. (off is the offset in abd.)
810  */
811 int
812 abd_cmp_buf_off(abd_t *abd, const void *buf, size_t off, size_t size)
813 {
814 	struct buf_arg ba_ptr = { (void *) buf };
815 
816 	return (abd_iterate_func(abd, off, size, abd_cmp_buf_off_cb, &ba_ptr));
817 }
818 
819 static int
820 abd_copy_from_buf_off_cb(void *buf, size_t size, void *private)
821 {
822 	struct buf_arg *ba_ptr = private;
823 
824 	(void) memcpy(buf, ba_ptr->arg_buf, size);
825 	ba_ptr->arg_buf = (char *)ba_ptr->arg_buf + size;
826 
827 	return (0);
828 }
829 
830 /*
831  * Copy from buf to abd. (off is the offset in abd.)
832  */
833 void
834 abd_copy_from_buf_off(abd_t *abd, const void *buf, size_t off, size_t size)
835 {
836 	struct buf_arg ba_ptr = { (void *) buf };
837 
838 	(void) abd_iterate_func(abd, off, size, abd_copy_from_buf_off_cb,
839 	    &ba_ptr);
840 }
841 
842 /*ARGSUSED*/
843 static int
844 abd_zero_off_cb(void *buf, size_t size, void *private)
845 {
846 	(void) memset(buf, 0, size);
847 	return (0);
848 }
849 
850 /*
851  * Zero out the abd from a particular offset to the end.
852  */
853 void
854 abd_zero_off(abd_t *abd, size_t off, size_t size)
855 {
856 	(void) abd_iterate_func(abd, off, size, abd_zero_off_cb, NULL);
857 }
858 
859 /*
860  * Iterate over two ABDs and call func incrementally on the two ABDs' data in
861  * equal-sized chunks (passed to func as raw buffers). func could be called many
862  * times during this iteration.
863  */
864 int
865 abd_iterate_func2(abd_t *dabd, abd_t *sabd, size_t doff, size_t soff,
866     size_t size, abd_iter_func2_t *func, void *private)
867 {
868 	int ret = 0;
869 	struct abd_iter daiter, saiter;
870 
871 	abd_verify(dabd);
872 	abd_verify(sabd);
873 
874 	ASSERT3U(doff + size, <=, dabd->abd_size);
875 	ASSERT3U(soff + size, <=, sabd->abd_size);
876 
877 	abd_iter_init(&daiter, dabd);
878 	abd_iter_init(&saiter, sabd);
879 	abd_iter_advance(&daiter, doff);
880 	abd_iter_advance(&saiter, soff);
881 
882 	while (size > 0) {
883 		abd_iter_map(&daiter);
884 		abd_iter_map(&saiter);
885 
886 		size_t dlen = MIN(daiter.iter_mapsize, size);
887 		size_t slen = MIN(saiter.iter_mapsize, size);
888 		size_t len = MIN(dlen, slen);
889 		ASSERT(dlen > 0 || slen > 0);
890 
891 		ret = func(daiter.iter_mapaddr, saiter.iter_mapaddr, len,
892 		    private);
893 
894 		abd_iter_unmap(&saiter);
895 		abd_iter_unmap(&daiter);
896 
897 		if (ret != 0)
898 			break;
899 
900 		size -= len;
901 		abd_iter_advance(&daiter, len);
902 		abd_iter_advance(&saiter, len);
903 	}
904 
905 	return (ret);
906 }
907 
908 /*ARGSUSED*/
909 static int
910 abd_copy_off_cb(void *dbuf, void *sbuf, size_t size, void *private)
911 {
912 	(void) memcpy(dbuf, sbuf, size);
913 	return (0);
914 }
915 
916 /*
917  * Copy from sabd to dabd starting from soff and doff.
918  */
919 void
920 abd_copy_off(abd_t *dabd, abd_t *sabd, size_t doff, size_t soff, size_t size)
921 {
922 	(void) abd_iterate_func2(dabd, sabd, doff, soff, size,
923 	    abd_copy_off_cb, NULL);
924 }
925 
926 /*ARGSUSED*/
927 static int
928 abd_cmp_cb(void *bufa, void *bufb, size_t size, void *private)
929 {
930 	return (memcmp(bufa, bufb, size));
931 }
932 
933 /*
934  * Compares the first size bytes of two ABDs.
935  */
936 int
937 abd_cmp(abd_t *dabd, abd_t *sabd, size_t size)
938 {
939 	return (abd_iterate_func2(dabd, sabd, 0, 0, size, abd_cmp_cb, NULL));
940 }
941