xref: /freebsd/sys/contrib/openzfs/module/os/linux/zfs/abd_os.c (revision f374ba41f55c1a127303d92d830dd58eef2f5243)
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 https://opensource.org/licenses/CDDL-1.0.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2014 by Chunwei Chen. All rights reserved.
23  * Copyright (c) 2019 by Delphix. All rights reserved.
24  */
25 
26 /*
27  * See abd.c for a general overview of the arc buffered data (ABD).
28  *
29  * Linear buffers act exactly like normal buffers and are always mapped into the
30  * kernel's virtual memory space, while scattered ABD data chunks are allocated
31  * as physical pages and then mapped in only while they are actually being
32  * accessed through one of the abd_* library functions. Using scattered ABDs
33  * provides several benefits:
34  *
35  *  (1) They avoid use of kmem_*, preventing performance problems where running
36  *      kmem_reap on very large memory systems never finishes and causes
37  *      constant TLB shootdowns.
38  *
39  *  (2) Fragmentation is less of an issue since when we are at the limit of
40  *      allocatable space, we won't have to search around for a long free
41  *      hole in the VA space for large ARC allocations. Each chunk is mapped in
42  *      individually, so even if we are using HIGHMEM (see next point) we
43  *      wouldn't need to worry about finding a contiguous address range.
44  *
45  *  (3) If we are not using HIGHMEM, then all physical memory is always
46  *      mapped into the kernel's address space, so we also avoid the map /
47  *      unmap costs on each ABD access.
48  *
49  * If we are not using HIGHMEM, scattered buffers which have only one chunk
50  * can be treated as linear buffers, because they are contiguous in the
51  * kernel's virtual address space.  See abd_alloc_chunks() for details.
52  */
53 
54 #include <sys/abd_impl.h>
55 #include <sys/param.h>
56 #include <sys/zio.h>
57 #include <sys/arc.h>
58 #include <sys/zfs_context.h>
59 #include <sys/zfs_znode.h>
60 #ifdef _KERNEL
61 #include <linux/kmap_compat.h>
62 #include <linux/scatterlist.h>
63 #else
64 #define	MAX_ORDER	1
65 #endif
66 
67 typedef struct abd_stats {
68 	kstat_named_t abdstat_struct_size;
69 	kstat_named_t abdstat_linear_cnt;
70 	kstat_named_t abdstat_linear_data_size;
71 	kstat_named_t abdstat_scatter_cnt;
72 	kstat_named_t abdstat_scatter_data_size;
73 	kstat_named_t abdstat_scatter_chunk_waste;
74 	kstat_named_t abdstat_scatter_orders[MAX_ORDER];
75 	kstat_named_t abdstat_scatter_page_multi_chunk;
76 	kstat_named_t abdstat_scatter_page_multi_zone;
77 	kstat_named_t abdstat_scatter_page_alloc_retry;
78 	kstat_named_t abdstat_scatter_sg_table_retry;
79 } abd_stats_t;
80 
81 static abd_stats_t abd_stats = {
82 	/* Amount of memory occupied by all of the abd_t struct allocations */
83 	{ "struct_size",			KSTAT_DATA_UINT64 },
84 	/*
85 	 * The number of linear ABDs which are currently allocated, excluding
86 	 * ABDs which don't own their data (for instance the ones which were
87 	 * allocated through abd_get_offset() and abd_get_from_buf()). If an
88 	 * ABD takes ownership of its buf then it will become tracked.
89 	 */
90 	{ "linear_cnt",				KSTAT_DATA_UINT64 },
91 	/* Amount of data stored in all linear ABDs tracked by linear_cnt */
92 	{ "linear_data_size",			KSTAT_DATA_UINT64 },
93 	/*
94 	 * The number of scatter ABDs which are currently allocated, excluding
95 	 * ABDs which don't own their data (for instance the ones which were
96 	 * allocated through abd_get_offset()).
97 	 */
98 	{ "scatter_cnt",			KSTAT_DATA_UINT64 },
99 	/* Amount of data stored in all scatter ABDs tracked by scatter_cnt */
100 	{ "scatter_data_size",			KSTAT_DATA_UINT64 },
101 	/*
102 	 * The amount of space wasted at the end of the last chunk across all
103 	 * scatter ABDs tracked by scatter_cnt.
104 	 */
105 	{ "scatter_chunk_waste",		KSTAT_DATA_UINT64 },
106 	/*
107 	 * The number of compound allocations of a given order.  These
108 	 * allocations are spread over all currently allocated ABDs, and
109 	 * act as a measure of memory fragmentation.
110 	 */
111 	{ { "scatter_order_N",			KSTAT_DATA_UINT64 } },
112 	/*
113 	 * The number of scatter ABDs which contain multiple chunks.
114 	 * ABDs are preferentially allocated from the minimum number of
115 	 * contiguous multi-page chunks, a single chunk is optimal.
116 	 */
117 	{ "scatter_page_multi_chunk",		KSTAT_DATA_UINT64 },
118 	/*
119 	 * The number of scatter ABDs which are split across memory zones.
120 	 * ABDs are preferentially allocated using pages from a single zone.
121 	 */
122 	{ "scatter_page_multi_zone",		KSTAT_DATA_UINT64 },
123 	/*
124 	 *  The total number of retries encountered when attempting to
125 	 *  allocate the pages to populate the scatter ABD.
126 	 */
127 	{ "scatter_page_alloc_retry",		KSTAT_DATA_UINT64 },
128 	/*
129 	 *  The total number of retries encountered when attempting to
130 	 *  allocate the sg table for an ABD.
131 	 */
132 	{ "scatter_sg_table_retry",		KSTAT_DATA_UINT64 },
133 };
134 
135 static struct {
136 	wmsum_t abdstat_struct_size;
137 	wmsum_t abdstat_linear_cnt;
138 	wmsum_t abdstat_linear_data_size;
139 	wmsum_t abdstat_scatter_cnt;
140 	wmsum_t abdstat_scatter_data_size;
141 	wmsum_t abdstat_scatter_chunk_waste;
142 	wmsum_t abdstat_scatter_orders[MAX_ORDER];
143 	wmsum_t abdstat_scatter_page_multi_chunk;
144 	wmsum_t abdstat_scatter_page_multi_zone;
145 	wmsum_t abdstat_scatter_page_alloc_retry;
146 	wmsum_t abdstat_scatter_sg_table_retry;
147 } abd_sums;
148 
149 #define	abd_for_each_sg(abd, sg, n, i)	\
150 	for_each_sg(ABD_SCATTER(abd).abd_sgl, sg, n, i)
151 
152 /*
153  * zfs_abd_scatter_min_size is the minimum allocation size to use scatter
154  * ABD's.  Smaller allocations will use linear ABD's which uses
155  * zio_[data_]buf_alloc().
156  *
157  * Scatter ABD's use at least one page each, so sub-page allocations waste
158  * some space when allocated as scatter (e.g. 2KB scatter allocation wastes
159  * half of each page).  Using linear ABD's for small allocations means that
160  * they will be put on slabs which contain many allocations.  This can
161  * improve memory efficiency, but it also makes it much harder for ARC
162  * evictions to actually free pages, because all the buffers on one slab need
163  * to be freed in order for the slab (and underlying pages) to be freed.
164  * Typically, 512B and 1KB kmem caches have 16 buffers per slab, so it's
165  * possible for them to actually waste more memory than scatter (one page per
166  * buf = wasting 3/4 or 7/8th; one buf per slab = wasting 15/16th).
167  *
168  * Spill blocks are typically 512B and are heavily used on systems running
169  * selinux with the default dnode size and the `xattr=sa` property set.
170  *
171  * By default we use linear allocations for 512B and 1KB, and scatter
172  * allocations for larger (1.5KB and up).
173  */
174 static int zfs_abd_scatter_min_size = 512 * 3;
175 
176 /*
177  * We use a scattered SPA_MAXBLOCKSIZE sized ABD whose pages are
178  * just a single zero'd page. This allows us to conserve memory by
179  * only using a single zero page for the scatterlist.
180  */
181 abd_t *abd_zero_scatter = NULL;
182 
183 struct page;
184 /*
185  * _KERNEL   - Will point to ZERO_PAGE if it is available or it will be
186  *             an allocated zero'd PAGESIZE buffer.
187  * Userspace - Will be an allocated zero'ed PAGESIZE buffer.
188  *
189  * abd_zero_page is assigned to each of the pages of abd_zero_scatter.
190  */
191 static struct page *abd_zero_page = NULL;
192 
193 static kmem_cache_t *abd_cache = NULL;
194 static kstat_t *abd_ksp;
195 
196 static uint_t
197 abd_chunkcnt_for_bytes(size_t size)
198 {
199 	return (P2ROUNDUP(size, PAGESIZE) / PAGESIZE);
200 }
201 
202 abd_t *
203 abd_alloc_struct_impl(size_t size)
204 {
205 	/*
206 	 * In Linux we do not use the size passed in during ABD
207 	 * allocation, so we just ignore it.
208 	 */
209 	(void) size;
210 	abd_t *abd = kmem_cache_alloc(abd_cache, KM_PUSHPAGE);
211 	ASSERT3P(abd, !=, NULL);
212 	ABDSTAT_INCR(abdstat_struct_size, sizeof (abd_t));
213 
214 	return (abd);
215 }
216 
217 void
218 abd_free_struct_impl(abd_t *abd)
219 {
220 	kmem_cache_free(abd_cache, abd);
221 	ABDSTAT_INCR(abdstat_struct_size, -(int)sizeof (abd_t));
222 }
223 
224 #ifdef _KERNEL
225 static unsigned zfs_abd_scatter_max_order = MAX_ORDER - 1;
226 
227 /*
228  * Mark zfs data pages so they can be excluded from kernel crash dumps
229  */
230 #ifdef _LP64
231 #define	ABD_FILE_CACHE_PAGE	0x2F5ABDF11ECAC4E
232 
233 static inline void
234 abd_mark_zfs_page(struct page *page)
235 {
236 	get_page(page);
237 	SetPagePrivate(page);
238 	set_page_private(page, ABD_FILE_CACHE_PAGE);
239 }
240 
241 static inline void
242 abd_unmark_zfs_page(struct page *page)
243 {
244 	set_page_private(page, 0UL);
245 	ClearPagePrivate(page);
246 	put_page(page);
247 }
248 #else
249 #define	abd_mark_zfs_page(page)
250 #define	abd_unmark_zfs_page(page)
251 #endif /* _LP64 */
252 
253 #ifndef CONFIG_HIGHMEM
254 
255 #ifndef __GFP_RECLAIM
256 #define	__GFP_RECLAIM		__GFP_WAIT
257 #endif
258 
259 /*
260  * The goal is to minimize fragmentation by preferentially populating ABDs
261  * with higher order compound pages from a single zone.  Allocation size is
262  * progressively decreased until it can be satisfied without performing
263  * reclaim or compaction.  When necessary this function will degenerate to
264  * allocating individual pages and allowing reclaim to satisfy allocations.
265  */
266 void
267 abd_alloc_chunks(abd_t *abd, size_t size)
268 {
269 	struct list_head pages;
270 	struct sg_table table;
271 	struct scatterlist *sg;
272 	struct page *page, *tmp_page = NULL;
273 	gfp_t gfp = __GFP_NOWARN | GFP_NOIO;
274 	gfp_t gfp_comp = (gfp | __GFP_NORETRY | __GFP_COMP) & ~__GFP_RECLAIM;
275 	int max_order = MIN(zfs_abd_scatter_max_order, MAX_ORDER - 1);
276 	int nr_pages = abd_chunkcnt_for_bytes(size);
277 	int chunks = 0, zones = 0;
278 	size_t remaining_size;
279 	int nid = NUMA_NO_NODE;
280 	int alloc_pages = 0;
281 
282 	INIT_LIST_HEAD(&pages);
283 
284 	while (alloc_pages < nr_pages) {
285 		unsigned chunk_pages;
286 		int order;
287 
288 		order = MIN(highbit64(nr_pages - alloc_pages) - 1, max_order);
289 		chunk_pages = (1U << order);
290 
291 		page = alloc_pages_node(nid, order ? gfp_comp : gfp, order);
292 		if (page == NULL) {
293 			if (order == 0) {
294 				ABDSTAT_BUMP(abdstat_scatter_page_alloc_retry);
295 				schedule_timeout_interruptible(1);
296 			} else {
297 				max_order = MAX(0, order - 1);
298 			}
299 			continue;
300 		}
301 
302 		list_add_tail(&page->lru, &pages);
303 
304 		if ((nid != NUMA_NO_NODE) && (page_to_nid(page) != nid))
305 			zones++;
306 
307 		nid = page_to_nid(page);
308 		ABDSTAT_BUMP(abdstat_scatter_orders[order]);
309 		chunks++;
310 		alloc_pages += chunk_pages;
311 	}
312 
313 	ASSERT3S(alloc_pages, ==, nr_pages);
314 
315 	while (sg_alloc_table(&table, chunks, gfp)) {
316 		ABDSTAT_BUMP(abdstat_scatter_sg_table_retry);
317 		schedule_timeout_interruptible(1);
318 	}
319 
320 	sg = table.sgl;
321 	remaining_size = size;
322 	list_for_each_entry_safe(page, tmp_page, &pages, lru) {
323 		size_t sg_size = MIN(PAGESIZE << compound_order(page),
324 		    remaining_size);
325 		sg_set_page(sg, page, sg_size, 0);
326 		abd_mark_zfs_page(page);
327 		remaining_size -= sg_size;
328 
329 		sg = sg_next(sg);
330 		list_del(&page->lru);
331 	}
332 
333 	/*
334 	 * These conditions ensure that a possible transformation to a linear
335 	 * ABD would be valid.
336 	 */
337 	ASSERT(!PageHighMem(sg_page(table.sgl)));
338 	ASSERT0(ABD_SCATTER(abd).abd_offset);
339 
340 	if (table.nents == 1) {
341 		/*
342 		 * Since there is only one entry, this ABD can be represented
343 		 * as a linear buffer.  All single-page (4K) ABD's can be
344 		 * represented this way.  Some multi-page ABD's can also be
345 		 * represented this way, if we were able to allocate a single
346 		 * "chunk" (higher-order "page" which represents a power-of-2
347 		 * series of physically-contiguous pages).  This is often the
348 		 * case for 2-page (8K) ABD's.
349 		 *
350 		 * Representing a single-entry scatter ABD as a linear ABD
351 		 * has the performance advantage of avoiding the copy (and
352 		 * allocation) in abd_borrow_buf_copy / abd_return_buf_copy.
353 		 * A performance increase of around 5% has been observed for
354 		 * ARC-cached reads (of small blocks which can take advantage
355 		 * of this).
356 		 *
357 		 * Note that this optimization is only possible because the
358 		 * pages are always mapped into the kernel's address space.
359 		 * This is not the case for highmem pages, so the
360 		 * optimization can not be made there.
361 		 */
362 		abd->abd_flags |= ABD_FLAG_LINEAR;
363 		abd->abd_flags |= ABD_FLAG_LINEAR_PAGE;
364 		abd->abd_u.abd_linear.abd_sgl = table.sgl;
365 		ABD_LINEAR_BUF(abd) = page_address(sg_page(table.sgl));
366 	} else if (table.nents > 1) {
367 		ABDSTAT_BUMP(abdstat_scatter_page_multi_chunk);
368 		abd->abd_flags |= ABD_FLAG_MULTI_CHUNK;
369 
370 		if (zones) {
371 			ABDSTAT_BUMP(abdstat_scatter_page_multi_zone);
372 			abd->abd_flags |= ABD_FLAG_MULTI_ZONE;
373 		}
374 
375 		ABD_SCATTER(abd).abd_sgl = table.sgl;
376 		ABD_SCATTER(abd).abd_nents = table.nents;
377 	}
378 }
379 #else
380 
381 /*
382  * Allocate N individual pages to construct a scatter ABD.  This function
383  * makes no attempt to request contiguous pages and requires the minimal
384  * number of kernel interfaces.  It's designed for maximum compatibility.
385  */
386 void
387 abd_alloc_chunks(abd_t *abd, size_t size)
388 {
389 	struct scatterlist *sg = NULL;
390 	struct sg_table table;
391 	struct page *page;
392 	gfp_t gfp = __GFP_NOWARN | GFP_NOIO;
393 	int nr_pages = abd_chunkcnt_for_bytes(size);
394 	int i = 0;
395 
396 	while (sg_alloc_table(&table, nr_pages, gfp)) {
397 		ABDSTAT_BUMP(abdstat_scatter_sg_table_retry);
398 		schedule_timeout_interruptible(1);
399 	}
400 
401 	ASSERT3U(table.nents, ==, nr_pages);
402 	ABD_SCATTER(abd).abd_sgl = table.sgl;
403 	ABD_SCATTER(abd).abd_nents = nr_pages;
404 
405 	abd_for_each_sg(abd, sg, nr_pages, i) {
406 		while ((page = __page_cache_alloc(gfp)) == NULL) {
407 			ABDSTAT_BUMP(abdstat_scatter_page_alloc_retry);
408 			schedule_timeout_interruptible(1);
409 		}
410 
411 		ABDSTAT_BUMP(abdstat_scatter_orders[0]);
412 		sg_set_page(sg, page, PAGESIZE, 0);
413 		abd_mark_zfs_page(page);
414 	}
415 
416 	if (nr_pages > 1) {
417 		ABDSTAT_BUMP(abdstat_scatter_page_multi_chunk);
418 		abd->abd_flags |= ABD_FLAG_MULTI_CHUNK;
419 	}
420 }
421 #endif /* !CONFIG_HIGHMEM */
422 
423 /*
424  * This must be called if any of the sg_table allocation functions
425  * are called.
426  */
427 static void
428 abd_free_sg_table(abd_t *abd)
429 {
430 	struct sg_table table;
431 
432 	table.sgl = ABD_SCATTER(abd).abd_sgl;
433 	table.nents = table.orig_nents = ABD_SCATTER(abd).abd_nents;
434 	sg_free_table(&table);
435 }
436 
437 void
438 abd_free_chunks(abd_t *abd)
439 {
440 	struct scatterlist *sg = NULL;
441 	struct page *page;
442 	int nr_pages = ABD_SCATTER(abd).abd_nents;
443 	int order, i = 0;
444 
445 	if (abd->abd_flags & ABD_FLAG_MULTI_ZONE)
446 		ABDSTAT_BUMPDOWN(abdstat_scatter_page_multi_zone);
447 
448 	if (abd->abd_flags & ABD_FLAG_MULTI_CHUNK)
449 		ABDSTAT_BUMPDOWN(abdstat_scatter_page_multi_chunk);
450 
451 	abd_for_each_sg(abd, sg, nr_pages, i) {
452 		page = sg_page(sg);
453 		abd_unmark_zfs_page(page);
454 		order = compound_order(page);
455 		__free_pages(page, order);
456 		ASSERT3U(sg->length, <=, PAGE_SIZE << order);
457 		ABDSTAT_BUMPDOWN(abdstat_scatter_orders[order]);
458 	}
459 	abd_free_sg_table(abd);
460 }
461 
462 /*
463  * Allocate scatter ABD of size SPA_MAXBLOCKSIZE, where each page in
464  * the scatterlist will be set to the zero'd out buffer abd_zero_page.
465  */
466 static void
467 abd_alloc_zero_scatter(void)
468 {
469 	struct scatterlist *sg = NULL;
470 	struct sg_table table;
471 	gfp_t gfp = __GFP_NOWARN | GFP_NOIO;
472 	int nr_pages = abd_chunkcnt_for_bytes(SPA_MAXBLOCKSIZE);
473 	int i = 0;
474 
475 #if defined(HAVE_ZERO_PAGE_GPL_ONLY)
476 	gfp_t gfp_zero_page = gfp | __GFP_ZERO;
477 	while ((abd_zero_page = __page_cache_alloc(gfp_zero_page)) == NULL) {
478 		ABDSTAT_BUMP(abdstat_scatter_page_alloc_retry);
479 		schedule_timeout_interruptible(1);
480 	}
481 	abd_mark_zfs_page(abd_zero_page);
482 #else
483 	abd_zero_page = ZERO_PAGE(0);
484 #endif /* HAVE_ZERO_PAGE_GPL_ONLY */
485 
486 	while (sg_alloc_table(&table, nr_pages, gfp)) {
487 		ABDSTAT_BUMP(abdstat_scatter_sg_table_retry);
488 		schedule_timeout_interruptible(1);
489 	}
490 	ASSERT3U(table.nents, ==, nr_pages);
491 
492 	abd_zero_scatter = abd_alloc_struct(SPA_MAXBLOCKSIZE);
493 	abd_zero_scatter->abd_flags |= ABD_FLAG_OWNER;
494 	ABD_SCATTER(abd_zero_scatter).abd_offset = 0;
495 	ABD_SCATTER(abd_zero_scatter).abd_sgl = table.sgl;
496 	ABD_SCATTER(abd_zero_scatter).abd_nents = nr_pages;
497 	abd_zero_scatter->abd_size = SPA_MAXBLOCKSIZE;
498 	abd_zero_scatter->abd_flags |= ABD_FLAG_MULTI_CHUNK | ABD_FLAG_ZEROS;
499 
500 	abd_for_each_sg(abd_zero_scatter, sg, nr_pages, i) {
501 		sg_set_page(sg, abd_zero_page, PAGESIZE, 0);
502 	}
503 
504 	ABDSTAT_BUMP(abdstat_scatter_cnt);
505 	ABDSTAT_INCR(abdstat_scatter_data_size, PAGESIZE);
506 	ABDSTAT_BUMP(abdstat_scatter_page_multi_chunk);
507 }
508 
509 #else /* _KERNEL */
510 
511 #ifndef PAGE_SHIFT
512 #define	PAGE_SHIFT (highbit64(PAGESIZE)-1)
513 #endif
514 
515 #define	zfs_kmap_atomic(chunk)		((void *)chunk)
516 #define	zfs_kunmap_atomic(addr)		do { (void)(addr); } while (0)
517 #define	local_irq_save(flags)		do { (void)(flags); } while (0)
518 #define	local_irq_restore(flags)	do { (void)(flags); } while (0)
519 #define	nth_page(pg, i) \
520 	((struct page *)((void *)(pg) + (i) * PAGESIZE))
521 
522 struct scatterlist {
523 	struct page *page;
524 	int length;
525 	int end;
526 };
527 
528 static void
529 sg_init_table(struct scatterlist *sg, int nr)
530 {
531 	memset(sg, 0, nr * sizeof (struct scatterlist));
532 	sg[nr - 1].end = 1;
533 }
534 
535 /*
536  * This must be called if any of the sg_table allocation functions
537  * are called.
538  */
539 static void
540 abd_free_sg_table(abd_t *abd)
541 {
542 	int nents = ABD_SCATTER(abd).abd_nents;
543 	vmem_free(ABD_SCATTER(abd).abd_sgl,
544 	    nents * sizeof (struct scatterlist));
545 }
546 
547 #define	for_each_sg(sgl, sg, nr, i)	\
548 	for ((i) = 0, (sg) = (sgl); (i) < (nr); (i)++, (sg) = sg_next(sg))
549 
550 static inline void
551 sg_set_page(struct scatterlist *sg, struct page *page, unsigned int len,
552     unsigned int offset)
553 {
554 	/* currently we don't use offset */
555 	ASSERT(offset == 0);
556 	sg->page = page;
557 	sg->length = len;
558 }
559 
560 static inline struct page *
561 sg_page(struct scatterlist *sg)
562 {
563 	return (sg->page);
564 }
565 
566 static inline struct scatterlist *
567 sg_next(struct scatterlist *sg)
568 {
569 	if (sg->end)
570 		return (NULL);
571 
572 	return (sg + 1);
573 }
574 
575 void
576 abd_alloc_chunks(abd_t *abd, size_t size)
577 {
578 	unsigned nr_pages = abd_chunkcnt_for_bytes(size);
579 	struct scatterlist *sg;
580 	int i;
581 
582 	ABD_SCATTER(abd).abd_sgl = vmem_alloc(nr_pages *
583 	    sizeof (struct scatterlist), KM_SLEEP);
584 	sg_init_table(ABD_SCATTER(abd).abd_sgl, nr_pages);
585 
586 	abd_for_each_sg(abd, sg, nr_pages, i) {
587 		struct page *p = umem_alloc_aligned(PAGESIZE, 64, KM_SLEEP);
588 		sg_set_page(sg, p, PAGESIZE, 0);
589 	}
590 	ABD_SCATTER(abd).abd_nents = nr_pages;
591 }
592 
593 void
594 abd_free_chunks(abd_t *abd)
595 {
596 	int i, n = ABD_SCATTER(abd).abd_nents;
597 	struct scatterlist *sg;
598 
599 	abd_for_each_sg(abd, sg, n, i) {
600 		struct page *p = nth_page(sg_page(sg), 0);
601 		umem_free_aligned(p, PAGESIZE);
602 	}
603 	abd_free_sg_table(abd);
604 }
605 
606 static void
607 abd_alloc_zero_scatter(void)
608 {
609 	unsigned nr_pages = abd_chunkcnt_for_bytes(SPA_MAXBLOCKSIZE);
610 	struct scatterlist *sg;
611 	int i;
612 
613 	abd_zero_page = umem_alloc_aligned(PAGESIZE, 64, KM_SLEEP);
614 	memset(abd_zero_page, 0, PAGESIZE);
615 	abd_zero_scatter = abd_alloc_struct(SPA_MAXBLOCKSIZE);
616 	abd_zero_scatter->abd_flags |= ABD_FLAG_OWNER;
617 	abd_zero_scatter->abd_flags |= ABD_FLAG_MULTI_CHUNK | ABD_FLAG_ZEROS;
618 	ABD_SCATTER(abd_zero_scatter).abd_offset = 0;
619 	ABD_SCATTER(abd_zero_scatter).abd_nents = nr_pages;
620 	abd_zero_scatter->abd_size = SPA_MAXBLOCKSIZE;
621 	ABD_SCATTER(abd_zero_scatter).abd_sgl = vmem_alloc(nr_pages *
622 	    sizeof (struct scatterlist), KM_SLEEP);
623 
624 	sg_init_table(ABD_SCATTER(abd_zero_scatter).abd_sgl, nr_pages);
625 
626 	abd_for_each_sg(abd_zero_scatter, sg, nr_pages, i) {
627 		sg_set_page(sg, abd_zero_page, PAGESIZE, 0);
628 	}
629 
630 	ABDSTAT_BUMP(abdstat_scatter_cnt);
631 	ABDSTAT_INCR(abdstat_scatter_data_size, PAGESIZE);
632 	ABDSTAT_BUMP(abdstat_scatter_page_multi_chunk);
633 }
634 
635 #endif /* _KERNEL */
636 
637 boolean_t
638 abd_size_alloc_linear(size_t size)
639 {
640 	return (!zfs_abd_scatter_enabled || size < zfs_abd_scatter_min_size);
641 }
642 
643 void
644 abd_update_scatter_stats(abd_t *abd, abd_stats_op_t op)
645 {
646 	ASSERT(op == ABDSTAT_INCR || op == ABDSTAT_DECR);
647 	int waste = P2ROUNDUP(abd->abd_size, PAGESIZE) - abd->abd_size;
648 	if (op == ABDSTAT_INCR) {
649 		ABDSTAT_BUMP(abdstat_scatter_cnt);
650 		ABDSTAT_INCR(abdstat_scatter_data_size, abd->abd_size);
651 		ABDSTAT_INCR(abdstat_scatter_chunk_waste, waste);
652 		arc_space_consume(waste, ARC_SPACE_ABD_CHUNK_WASTE);
653 	} else {
654 		ABDSTAT_BUMPDOWN(abdstat_scatter_cnt);
655 		ABDSTAT_INCR(abdstat_scatter_data_size, -(int)abd->abd_size);
656 		ABDSTAT_INCR(abdstat_scatter_chunk_waste, -waste);
657 		arc_space_return(waste, ARC_SPACE_ABD_CHUNK_WASTE);
658 	}
659 }
660 
661 void
662 abd_update_linear_stats(abd_t *abd, abd_stats_op_t op)
663 {
664 	ASSERT(op == ABDSTAT_INCR || op == ABDSTAT_DECR);
665 	if (op == ABDSTAT_INCR) {
666 		ABDSTAT_BUMP(abdstat_linear_cnt);
667 		ABDSTAT_INCR(abdstat_linear_data_size, abd->abd_size);
668 	} else {
669 		ABDSTAT_BUMPDOWN(abdstat_linear_cnt);
670 		ABDSTAT_INCR(abdstat_linear_data_size, -(int)abd->abd_size);
671 	}
672 }
673 
674 void
675 abd_verify_scatter(abd_t *abd)
676 {
677 	size_t n;
678 	int i = 0;
679 	struct scatterlist *sg = NULL;
680 
681 	ASSERT3U(ABD_SCATTER(abd).abd_nents, >, 0);
682 	ASSERT3U(ABD_SCATTER(abd).abd_offset, <,
683 	    ABD_SCATTER(abd).abd_sgl->length);
684 	n = ABD_SCATTER(abd).abd_nents;
685 	abd_for_each_sg(abd, sg, n, i) {
686 		ASSERT3P(sg_page(sg), !=, NULL);
687 	}
688 }
689 
690 static void
691 abd_free_zero_scatter(void)
692 {
693 	ABDSTAT_BUMPDOWN(abdstat_scatter_cnt);
694 	ABDSTAT_INCR(abdstat_scatter_data_size, -(int)PAGESIZE);
695 	ABDSTAT_BUMPDOWN(abdstat_scatter_page_multi_chunk);
696 
697 	abd_free_sg_table(abd_zero_scatter);
698 	abd_free_struct(abd_zero_scatter);
699 	abd_zero_scatter = NULL;
700 	ASSERT3P(abd_zero_page, !=, NULL);
701 #if defined(_KERNEL)
702 #if defined(HAVE_ZERO_PAGE_GPL_ONLY)
703 	abd_unmark_zfs_page(abd_zero_page);
704 	__free_page(abd_zero_page);
705 #endif /* HAVE_ZERO_PAGE_GPL_ONLY */
706 #else
707 	umem_free_aligned(abd_zero_page, PAGESIZE);
708 #endif /* _KERNEL */
709 }
710 
711 static int
712 abd_kstats_update(kstat_t *ksp, int rw)
713 {
714 	abd_stats_t *as = ksp->ks_data;
715 
716 	if (rw == KSTAT_WRITE)
717 		return (EACCES);
718 	as->abdstat_struct_size.value.ui64 =
719 	    wmsum_value(&abd_sums.abdstat_struct_size);
720 	as->abdstat_linear_cnt.value.ui64 =
721 	    wmsum_value(&abd_sums.abdstat_linear_cnt);
722 	as->abdstat_linear_data_size.value.ui64 =
723 	    wmsum_value(&abd_sums.abdstat_linear_data_size);
724 	as->abdstat_scatter_cnt.value.ui64 =
725 	    wmsum_value(&abd_sums.abdstat_scatter_cnt);
726 	as->abdstat_scatter_data_size.value.ui64 =
727 	    wmsum_value(&abd_sums.abdstat_scatter_data_size);
728 	as->abdstat_scatter_chunk_waste.value.ui64 =
729 	    wmsum_value(&abd_sums.abdstat_scatter_chunk_waste);
730 	for (int i = 0; i < MAX_ORDER; i++) {
731 		as->abdstat_scatter_orders[i].value.ui64 =
732 		    wmsum_value(&abd_sums.abdstat_scatter_orders[i]);
733 	}
734 	as->abdstat_scatter_page_multi_chunk.value.ui64 =
735 	    wmsum_value(&abd_sums.abdstat_scatter_page_multi_chunk);
736 	as->abdstat_scatter_page_multi_zone.value.ui64 =
737 	    wmsum_value(&abd_sums.abdstat_scatter_page_multi_zone);
738 	as->abdstat_scatter_page_alloc_retry.value.ui64 =
739 	    wmsum_value(&abd_sums.abdstat_scatter_page_alloc_retry);
740 	as->abdstat_scatter_sg_table_retry.value.ui64 =
741 	    wmsum_value(&abd_sums.abdstat_scatter_sg_table_retry);
742 	return (0);
743 }
744 
745 void
746 abd_init(void)
747 {
748 	int i;
749 
750 	abd_cache = kmem_cache_create("abd_t", sizeof (abd_t),
751 	    0, NULL, NULL, NULL, NULL, NULL, 0);
752 
753 	wmsum_init(&abd_sums.abdstat_struct_size, 0);
754 	wmsum_init(&abd_sums.abdstat_linear_cnt, 0);
755 	wmsum_init(&abd_sums.abdstat_linear_data_size, 0);
756 	wmsum_init(&abd_sums.abdstat_scatter_cnt, 0);
757 	wmsum_init(&abd_sums.abdstat_scatter_data_size, 0);
758 	wmsum_init(&abd_sums.abdstat_scatter_chunk_waste, 0);
759 	for (i = 0; i < MAX_ORDER; i++)
760 		wmsum_init(&abd_sums.abdstat_scatter_orders[i], 0);
761 	wmsum_init(&abd_sums.abdstat_scatter_page_multi_chunk, 0);
762 	wmsum_init(&abd_sums.abdstat_scatter_page_multi_zone, 0);
763 	wmsum_init(&abd_sums.abdstat_scatter_page_alloc_retry, 0);
764 	wmsum_init(&abd_sums.abdstat_scatter_sg_table_retry, 0);
765 
766 	abd_ksp = kstat_create("zfs", 0, "abdstats", "misc", KSTAT_TYPE_NAMED,
767 	    sizeof (abd_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
768 	if (abd_ksp != NULL) {
769 		for (i = 0; i < MAX_ORDER; i++) {
770 			snprintf(abd_stats.abdstat_scatter_orders[i].name,
771 			    KSTAT_STRLEN, "scatter_order_%d", i);
772 			abd_stats.abdstat_scatter_orders[i].data_type =
773 			    KSTAT_DATA_UINT64;
774 		}
775 		abd_ksp->ks_data = &abd_stats;
776 		abd_ksp->ks_update = abd_kstats_update;
777 		kstat_install(abd_ksp);
778 	}
779 
780 	abd_alloc_zero_scatter();
781 }
782 
783 void
784 abd_fini(void)
785 {
786 	abd_free_zero_scatter();
787 
788 	if (abd_ksp != NULL) {
789 		kstat_delete(abd_ksp);
790 		abd_ksp = NULL;
791 	}
792 
793 	wmsum_fini(&abd_sums.abdstat_struct_size);
794 	wmsum_fini(&abd_sums.abdstat_linear_cnt);
795 	wmsum_fini(&abd_sums.abdstat_linear_data_size);
796 	wmsum_fini(&abd_sums.abdstat_scatter_cnt);
797 	wmsum_fini(&abd_sums.abdstat_scatter_data_size);
798 	wmsum_fini(&abd_sums.abdstat_scatter_chunk_waste);
799 	for (int i = 0; i < MAX_ORDER; i++)
800 		wmsum_fini(&abd_sums.abdstat_scatter_orders[i]);
801 	wmsum_fini(&abd_sums.abdstat_scatter_page_multi_chunk);
802 	wmsum_fini(&abd_sums.abdstat_scatter_page_multi_zone);
803 	wmsum_fini(&abd_sums.abdstat_scatter_page_alloc_retry);
804 	wmsum_fini(&abd_sums.abdstat_scatter_sg_table_retry);
805 
806 	if (abd_cache) {
807 		kmem_cache_destroy(abd_cache);
808 		abd_cache = NULL;
809 	}
810 }
811 
812 void
813 abd_free_linear_page(abd_t *abd)
814 {
815 	/* Transform it back into a scatter ABD for freeing */
816 	struct scatterlist *sg = abd->abd_u.abd_linear.abd_sgl;
817 	abd->abd_flags &= ~ABD_FLAG_LINEAR;
818 	abd->abd_flags &= ~ABD_FLAG_LINEAR_PAGE;
819 	ABD_SCATTER(abd).abd_nents = 1;
820 	ABD_SCATTER(abd).abd_offset = 0;
821 	ABD_SCATTER(abd).abd_sgl = sg;
822 	abd_free_chunks(abd);
823 
824 	abd_update_scatter_stats(abd, ABDSTAT_DECR);
825 }
826 
827 /*
828  * If we're going to use this ABD for doing I/O using the block layer, the
829  * consumer of the ABD data doesn't care if it's scattered or not, and we don't
830  * plan to store this ABD in memory for a long period of time, we should
831  * allocate the ABD type that requires the least data copying to do the I/O.
832  *
833  * On Linux the optimal thing to do would be to use abd_get_offset() and
834  * construct a new ABD which shares the original pages thereby eliminating
835  * the copy.  But for the moment a new linear ABD is allocated until this
836  * performance optimization can be implemented.
837  */
838 abd_t *
839 abd_alloc_for_io(size_t size, boolean_t is_metadata)
840 {
841 	return (abd_alloc(size, is_metadata));
842 }
843 
844 abd_t *
845 abd_get_offset_scatter(abd_t *abd, abd_t *sabd, size_t off,
846     size_t size)
847 {
848 	(void) size;
849 	int i = 0;
850 	struct scatterlist *sg = NULL;
851 
852 	abd_verify(sabd);
853 	ASSERT3U(off, <=, sabd->abd_size);
854 
855 	size_t new_offset = ABD_SCATTER(sabd).abd_offset + off;
856 
857 	if (abd == NULL)
858 		abd = abd_alloc_struct(0);
859 
860 	/*
861 	 * Even if this buf is filesystem metadata, we only track that
862 	 * if we own the underlying data buffer, which is not true in
863 	 * this case. Therefore, we don't ever use ABD_FLAG_META here.
864 	 */
865 
866 	abd_for_each_sg(sabd, sg, ABD_SCATTER(sabd).abd_nents, i) {
867 		if (new_offset < sg->length)
868 			break;
869 		new_offset -= sg->length;
870 	}
871 
872 	ABD_SCATTER(abd).abd_sgl = sg;
873 	ABD_SCATTER(abd).abd_offset = new_offset;
874 	ABD_SCATTER(abd).abd_nents = ABD_SCATTER(sabd).abd_nents - i;
875 
876 	return (abd);
877 }
878 
879 /*
880  * Initialize the abd_iter.
881  */
882 void
883 abd_iter_init(struct abd_iter *aiter, abd_t *abd)
884 {
885 	ASSERT(!abd_is_gang(abd));
886 	abd_verify(abd);
887 	aiter->iter_abd = abd;
888 	aiter->iter_mapaddr = NULL;
889 	aiter->iter_mapsize = 0;
890 	aiter->iter_pos = 0;
891 	if (abd_is_linear(abd)) {
892 		aiter->iter_offset = 0;
893 		aiter->iter_sg = NULL;
894 	} else {
895 		aiter->iter_offset = ABD_SCATTER(abd).abd_offset;
896 		aiter->iter_sg = ABD_SCATTER(abd).abd_sgl;
897 	}
898 }
899 
900 /*
901  * This is just a helper function to see if we have exhausted the
902  * abd_iter and reached the end.
903  */
904 boolean_t
905 abd_iter_at_end(struct abd_iter *aiter)
906 {
907 	return (aiter->iter_pos == aiter->iter_abd->abd_size);
908 }
909 
910 /*
911  * Advance the iterator by a certain amount. Cannot be called when a chunk is
912  * in use. This can be safely called when the aiter has already exhausted, in
913  * which case this does nothing.
914  */
915 void
916 abd_iter_advance(struct abd_iter *aiter, size_t amount)
917 {
918 	ASSERT3P(aiter->iter_mapaddr, ==, NULL);
919 	ASSERT0(aiter->iter_mapsize);
920 
921 	/* There's nothing left to advance to, so do nothing */
922 	if (abd_iter_at_end(aiter))
923 		return;
924 
925 	aiter->iter_pos += amount;
926 	aiter->iter_offset += amount;
927 	if (!abd_is_linear(aiter->iter_abd)) {
928 		while (aiter->iter_offset >= aiter->iter_sg->length) {
929 			aiter->iter_offset -= aiter->iter_sg->length;
930 			aiter->iter_sg = sg_next(aiter->iter_sg);
931 			if (aiter->iter_sg == NULL) {
932 				ASSERT0(aiter->iter_offset);
933 				break;
934 			}
935 		}
936 	}
937 }
938 
939 /*
940  * Map the current chunk into aiter. This can be safely called when the aiter
941  * has already exhausted, in which case this does nothing.
942  */
943 void
944 abd_iter_map(struct abd_iter *aiter)
945 {
946 	void *paddr;
947 	size_t offset = 0;
948 
949 	ASSERT3P(aiter->iter_mapaddr, ==, NULL);
950 	ASSERT0(aiter->iter_mapsize);
951 
952 	/* There's nothing left to iterate over, so do nothing */
953 	if (abd_iter_at_end(aiter))
954 		return;
955 
956 	if (abd_is_linear(aiter->iter_abd)) {
957 		ASSERT3U(aiter->iter_pos, ==, aiter->iter_offset);
958 		offset = aiter->iter_offset;
959 		aiter->iter_mapsize = aiter->iter_abd->abd_size - offset;
960 		paddr = ABD_LINEAR_BUF(aiter->iter_abd);
961 	} else {
962 		offset = aiter->iter_offset;
963 		aiter->iter_mapsize = MIN(aiter->iter_sg->length - offset,
964 		    aiter->iter_abd->abd_size - aiter->iter_pos);
965 
966 		paddr = zfs_kmap_atomic(sg_page(aiter->iter_sg));
967 	}
968 
969 	aiter->iter_mapaddr = (char *)paddr + offset;
970 }
971 
972 /*
973  * Unmap the current chunk from aiter. This can be safely called when the aiter
974  * has already exhausted, in which case this does nothing.
975  */
976 void
977 abd_iter_unmap(struct abd_iter *aiter)
978 {
979 	/* There's nothing left to unmap, so do nothing */
980 	if (abd_iter_at_end(aiter))
981 		return;
982 
983 	if (!abd_is_linear(aiter->iter_abd)) {
984 		/* LINTED E_FUNC_SET_NOT_USED */
985 		zfs_kunmap_atomic(aiter->iter_mapaddr - aiter->iter_offset);
986 	}
987 
988 	ASSERT3P(aiter->iter_mapaddr, !=, NULL);
989 	ASSERT3U(aiter->iter_mapsize, >, 0);
990 
991 	aiter->iter_mapaddr = NULL;
992 	aiter->iter_mapsize = 0;
993 }
994 
995 void
996 abd_cache_reap_now(void)
997 {
998 }
999 
1000 #if defined(_KERNEL)
1001 /*
1002  * bio_nr_pages for ABD.
1003  * @off is the offset in @abd
1004  */
1005 unsigned long
1006 abd_nr_pages_off(abd_t *abd, unsigned int size, size_t off)
1007 {
1008 	unsigned long pos;
1009 
1010 	if (abd_is_gang(abd)) {
1011 		unsigned long count = 0;
1012 
1013 		for (abd_t *cabd = abd_gang_get_offset(abd, &off);
1014 		    cabd != NULL && size != 0;
1015 		    cabd = list_next(&ABD_GANG(abd).abd_gang_chain, cabd)) {
1016 			ASSERT3U(off, <, cabd->abd_size);
1017 			int mysize = MIN(size, cabd->abd_size - off);
1018 			count += abd_nr_pages_off(cabd, mysize, off);
1019 			size -= mysize;
1020 			off = 0;
1021 		}
1022 		return (count);
1023 	}
1024 
1025 	if (abd_is_linear(abd))
1026 		pos = (unsigned long)abd_to_buf(abd) + off;
1027 	else
1028 		pos = ABD_SCATTER(abd).abd_offset + off;
1029 
1030 	return (((pos + size + PAGESIZE - 1) >> PAGE_SHIFT) -
1031 	    (pos >> PAGE_SHIFT));
1032 }
1033 
1034 static unsigned int
1035 bio_map(struct bio *bio, void *buf_ptr, unsigned int bio_size)
1036 {
1037 	unsigned int offset, size, i;
1038 	struct page *page;
1039 
1040 	offset = offset_in_page(buf_ptr);
1041 	for (i = 0; i < bio->bi_max_vecs; i++) {
1042 		size = PAGE_SIZE - offset;
1043 
1044 		if (bio_size <= 0)
1045 			break;
1046 
1047 		if (size > bio_size)
1048 			size = bio_size;
1049 
1050 		if (is_vmalloc_addr(buf_ptr))
1051 			page = vmalloc_to_page(buf_ptr);
1052 		else
1053 			page = virt_to_page(buf_ptr);
1054 
1055 		/*
1056 		 * Some network related block device uses tcp_sendpage, which
1057 		 * doesn't behave well when using 0-count page, this is a
1058 		 * safety net to catch them.
1059 		 */
1060 		ASSERT3S(page_count(page), >, 0);
1061 
1062 		if (bio_add_page(bio, page, size, offset) != size)
1063 			break;
1064 
1065 		buf_ptr += size;
1066 		bio_size -= size;
1067 		offset = 0;
1068 	}
1069 
1070 	return (bio_size);
1071 }
1072 
1073 /*
1074  * bio_map for gang ABD.
1075  */
1076 static unsigned int
1077 abd_gang_bio_map_off(struct bio *bio, abd_t *abd,
1078     unsigned int io_size, size_t off)
1079 {
1080 	ASSERT(abd_is_gang(abd));
1081 
1082 	for (abd_t *cabd = abd_gang_get_offset(abd, &off);
1083 	    cabd != NULL;
1084 	    cabd = list_next(&ABD_GANG(abd).abd_gang_chain, cabd)) {
1085 		ASSERT3U(off, <, cabd->abd_size);
1086 		int size = MIN(io_size, cabd->abd_size - off);
1087 		int remainder = abd_bio_map_off(bio, cabd, size, off);
1088 		io_size -= (size - remainder);
1089 		if (io_size == 0 || remainder > 0)
1090 			return (io_size);
1091 		off = 0;
1092 	}
1093 	ASSERT0(io_size);
1094 	return (io_size);
1095 }
1096 
1097 /*
1098  * bio_map for ABD.
1099  * @off is the offset in @abd
1100  * Remaining IO size is returned
1101  */
1102 unsigned int
1103 abd_bio_map_off(struct bio *bio, abd_t *abd,
1104     unsigned int io_size, size_t off)
1105 {
1106 	struct abd_iter aiter;
1107 
1108 	ASSERT3U(io_size, <=, abd->abd_size - off);
1109 	if (abd_is_linear(abd))
1110 		return (bio_map(bio, ((char *)abd_to_buf(abd)) + off, io_size));
1111 
1112 	ASSERT(!abd_is_linear(abd));
1113 	if (abd_is_gang(abd))
1114 		return (abd_gang_bio_map_off(bio, abd, io_size, off));
1115 
1116 	abd_iter_init(&aiter, abd);
1117 	abd_iter_advance(&aiter, off);
1118 
1119 	for (int i = 0; i < bio->bi_max_vecs; i++) {
1120 		struct page *pg;
1121 		size_t len, sgoff, pgoff;
1122 		struct scatterlist *sg;
1123 
1124 		if (io_size <= 0)
1125 			break;
1126 
1127 		sg = aiter.iter_sg;
1128 		sgoff = aiter.iter_offset;
1129 		pgoff = sgoff & (PAGESIZE - 1);
1130 		len = MIN(io_size, PAGESIZE - pgoff);
1131 		ASSERT(len > 0);
1132 
1133 		pg = nth_page(sg_page(sg), sgoff >> PAGE_SHIFT);
1134 		if (bio_add_page(bio, pg, len, pgoff) != len)
1135 			break;
1136 
1137 		io_size -= len;
1138 		abd_iter_advance(&aiter, len);
1139 	}
1140 
1141 	return (io_size);
1142 }
1143 
1144 /* Tunable Parameters */
1145 module_param(zfs_abd_scatter_enabled, int, 0644);
1146 MODULE_PARM_DESC(zfs_abd_scatter_enabled,
1147 	"Toggle whether ABD allocations must be linear.");
1148 module_param(zfs_abd_scatter_min_size, int, 0644);
1149 MODULE_PARM_DESC(zfs_abd_scatter_min_size,
1150 	"Minimum size of scatter allocations.");
1151 /* CSTYLED */
1152 module_param(zfs_abd_scatter_max_order, uint, 0644);
1153 MODULE_PARM_DESC(zfs_abd_scatter_max_order,
1154 	"Maximum order allocation used for a scatter ABD.");
1155 #endif
1156