ABD optimized page allocation code

* Convert ABD to use the Linux Kernel scatterlist implementation
  instead of the hand rolled one from illumos.

* Scatter ABDs are preferentially populated with higher order
  compound pages from a single zone.  Allocation size is
  progressively decreased until it can be satisfied without
  performing reclaim or compaction.

* An alternate page allocator is provided for kernels older
  than 3.6 and for CONFIG_HIGHMEM systems.  This allocator
  is designed as a fallback for maximum compatibility.

* Extended abdstats to provide visibility in the the allocator.

* Add cached value for PAGESIZE in userspace.

Contributions-by:
Chunwei Chen <david.chen@osnexus.com>
Gvozden Neskovic <neskovic@gmail.com>
Jinshan Xiong <jinshan.xiong@intel.com>
Isaac Huang <he.huang@intel.com>
David Quigley <david.quigley@intel.com>
Brian Behlendorf <behlendorf1@llnl.gov>
This commit is contained in:
Chunwei Chen 2016-10-26 00:32:23 -04:00 committed by Brian Behlendorf
parent 4f60152910
commit 9829574834
5 changed files with 465 additions and 144 deletions

View File

@ -43,7 +43,9 @@ extern "C" {
typedef enum abd_flags { typedef enum abd_flags {
ABD_FLAG_LINEAR = 1 << 0, /* is buffer linear (or scattered)? */ ABD_FLAG_LINEAR = 1 << 0, /* is buffer linear (or scattered)? */
ABD_FLAG_OWNER = 1 << 1, /* does it own its data buffers? */ ABD_FLAG_OWNER = 1 << 1, /* does it own its data buffers? */
ABD_FLAG_META = 1 << 2 /* does this represent FS metadata? */ ABD_FLAG_META = 1 << 2, /* does this represent FS metadata? */
ABD_FLAG_MULTI_ZONE = 1 << 3, /* pages split over memory zones */
ABD_FLAG_MULTI_CHUNK = 1 << 4 /* pages split over multiple chunks */
} abd_flags_t; } abd_flags_t;
typedef struct abd { typedef struct abd {
@ -54,8 +56,8 @@ typedef struct abd {
union { union {
struct abd_scatter { struct abd_scatter {
uint_t abd_offset; uint_t abd_offset;
uint_t abd_chunk_size; uint_t abd_nents;
struct page *abd_chunks[]; struct scatterlist *abd_sgl;
} abd_scatter; } abd_scatter;
struct abd_linear { struct abd_linear {
void *abd_buf; void *abd_buf;

View File

@ -24,6 +24,7 @@ USER_C = \
getmntany.c \ getmntany.c \
list.c \ list.c \
mkdirp.c \ mkdirp.c \
page.c \
strlcat.c \ strlcat.c \
strlcpy.c \ strlcpy.c \
strnlen.c \ strnlen.c \

View File

@ -57,8 +57,11 @@
#define MAXUID UINT32_MAX /* max user id */ #define MAXUID UINT32_MAX /* max user id */
#define MAXPROJID MAXUID /* max project id */ #define MAXPROJID MAXUID /* max project id */
#ifndef PAGESIZE #ifdef PAGESIZE
#define PAGESIZE (sysconf(_SC_PAGESIZE)) #undef PAGESIZE
#endif /* PAGESIZE */ #endif /* PAGESIZE */
extern size_t spl_pagesize(void);
#define PAGESIZE (spl_pagesize())
#endif #endif

34
lib/libspl/page.c Normal file
View File

@ -0,0 +1,34 @@
/*
* CDDL HEADER START
*
* The contents of this file are subject to the terms of the
* Common Development and Distribution License, Version 1.0 only
* (the "License"). You may not use this file except in compliance
* with the License.
*
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
* or http://www.opensolaris.org/os/licensing.
* See the License for the specific language governing permissions
* and limitations under the License.
*
* When distributing Covered Code, include this CDDL HEADER in each
* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
* If applicable, add the following below this CDDL HEADER, with the
* fields enclosed by brackets "[]" replaced with your own identifying
* information: Portions Copyright [yyyy] [name of copyright owner]
*
* CDDL HEADER END
*/
#include <unistd.h>
size_t pagesize = 0;
size_t
spl_pagesize(void)
{
if (pagesize == 0)
pagesize = sysconf(_SC_PAGESIZE);
return (pagesize);
}

View File

@ -120,25 +120,38 @@
#include <sys/zfs_context.h> #include <sys/zfs_context.h>
#include <sys/zfs_znode.h> #include <sys/zfs_znode.h>
#ifdef _KERNEL #ifdef _KERNEL
#include <linux/scatterlist.h>
#include <linux/kmap_compat.h> #include <linux/kmap_compat.h>
#endif #else
#define MAX_ORDER 1
#ifndef KMC_NOTOUCH
#define KMC_NOTOUCH 0
#endif #endif
typedef struct abd_stats { typedef struct abd_stats {
kstat_named_t abdstat_struct_size; kstat_named_t abdstat_struct_size;
kstat_named_t abdstat_linear_cnt;
kstat_named_t abdstat_linear_data_size;
kstat_named_t abdstat_scatter_cnt; kstat_named_t abdstat_scatter_cnt;
kstat_named_t abdstat_scatter_data_size; kstat_named_t abdstat_scatter_data_size;
kstat_named_t abdstat_scatter_chunk_waste; kstat_named_t abdstat_scatter_chunk_waste;
kstat_named_t abdstat_linear_cnt; kstat_named_t abdstat_scatter_orders[MAX_ORDER];
kstat_named_t abdstat_linear_data_size; kstat_named_t abdstat_scatter_page_multi_chunk;
kstat_named_t abdstat_scatter_page_multi_zone;
kstat_named_t abdstat_scatter_page_alloc_retry;
kstat_named_t abdstat_scatter_sg_table_retry;
} abd_stats_t; } abd_stats_t;
static abd_stats_t abd_stats = { static abd_stats_t abd_stats = {
/* Amount of memory occupied by all of the abd_t struct allocations */ /* Amount of memory occupied by all of the abd_t struct allocations */
{ "struct_size", KSTAT_DATA_UINT64 }, { "struct_size", KSTAT_DATA_UINT64 },
/*
* The number of linear ABDs which are currently allocated, excluding
* ABDs which don't own their data (for instance the ones which were
* allocated through abd_get_offset() and abd_get_from_buf()). If an
* ABD takes ownership of its buf then it will become tracked.
*/
{ "linear_cnt", KSTAT_DATA_UINT64 },
/* Amount of data stored in all linear ABDs tracked by linear_cnt */
{ "linear_data_size", KSTAT_DATA_UINT64 },
/* /*
* The number of scatter ABDs which are currently allocated, excluding * The number of scatter ABDs which are currently allocated, excluding
* ABDs which don't own their data (for instance the ones which were * ABDs which don't own their data (for instance the ones which were
@ -153,14 +166,32 @@ static abd_stats_t abd_stats = {
*/ */
{ "scatter_chunk_waste", KSTAT_DATA_UINT64 }, { "scatter_chunk_waste", KSTAT_DATA_UINT64 },
/* /*
* The number of linear ABDs which are currently allocated, excluding * The number of compound allocations of a given order. These
* ABDs which don't own their data (for instance the ones which were * allocations are spread over all currently allocated ABDs, and
* allocated through abd_get_offset() and abd_get_from_buf()). If an * act as a measure of memory fragmentation.
* ABD takes ownership of its buf then it will become tracked.
*/ */
{ "linear_cnt", KSTAT_DATA_UINT64 }, { { "scatter_order_N", KSTAT_DATA_UINT64 } },
/* Amount of data stored in all linear ABDs tracked by linear_cnt */ /*
{ "linear_data_size", KSTAT_DATA_UINT64 }, * The number of scatter ABDs which contain multiple chunks.
* ABDs are preferentially allocated from the minimum number of
* contiguous multi-page chunks, a single chunk is optimal.
*/
{ "scatter_page_multi_chunk", KSTAT_DATA_UINT64 },
/*
* The number of scatter ABDs which are split across memory zones.
* ABDs are preferentially allocated using pages from a single zone.
*/
{ "scatter_page_multi_zone", KSTAT_DATA_UINT64 },
/*
* The total number of retries encountered when attempting to
* allocate the pages to populate the scatter ABD.
*/
{ "scatter_page_alloc_retry", KSTAT_DATA_UINT64 },
/*
* The total number of retries encountered when attempting to
* allocate the sg table for an ABD.
*/
{ "scatter_sg_table_retry", KSTAT_DATA_UINT64 },
}; };
#define ABDSTAT(stat) (abd_stats.stat.value.ui64) #define ABDSTAT(stat) (abd_stats.stat.value.ui64)
@ -169,35 +200,319 @@ static abd_stats_t abd_stats = {
#define ABDSTAT_BUMP(stat) ABDSTAT_INCR(stat, 1) #define ABDSTAT_BUMP(stat) ABDSTAT_INCR(stat, 1)
#define ABDSTAT_BUMPDOWN(stat) ABDSTAT_INCR(stat, -1) #define ABDSTAT_BUMPDOWN(stat) ABDSTAT_INCR(stat, -1)
#define ABD_SCATTER(abd) (abd->abd_u.abd_scatter)
#define ABD_BUF(abd) (abd->abd_u.abd_linear.abd_buf)
#define abd_for_each_sg(abd, sg, n, i) \
for_each_sg(ABD_SCATTER(abd).abd_sgl, sg, n, i)
/* see block comment above for description */ /* see block comment above for description */
int zfs_abd_scatter_enabled = B_TRUE; int zfs_abd_scatter_enabled = B_TRUE;
unsigned zfs_abd_scatter_max_order = MAX_ORDER - 1;
static kmem_cache_t *abd_cache = NULL;
#ifdef _KERNEL
static kstat_t *abd_ksp; static kstat_t *abd_ksp;
static struct page * static inline size_t
abd_alloc_chunk(void) abd_chunkcnt_for_bytes(size_t size)
{ {
struct page *c = alloc_page(kmem_flags_convert(KM_SLEEP)); return (P2ROUNDUP(size, PAGESIZE) / PAGESIZE);
ASSERT3P(c, !=, NULL); }
return (c);
#ifdef _KERNEL
#ifndef CONFIG_HIGHMEM
#ifndef __GFP_RECLAIM
#define __GFP_RECLAIM __GFP_WAIT
#endif
static unsigned long
abd_alloc_chunk(int nid, gfp_t gfp, unsigned int order)
{
struct page *page;
page = alloc_pages_node(nid, gfp, order);
if (!page)
return (0);
return ((unsigned long) page_address(page));
}
/*
* The goal is to minimize fragmentation by preferentially populating ABDs
* with higher order compound pages from a single zone. Allocation size is
* progressively decreased until it can be satisfied without performing
* reclaim or compaction. When necessary this function will degenerate to
* allocating individual pages and allowing reclaim to satisfy allocations.
*/
static void
abd_alloc_pages(abd_t *abd, size_t size)
{
struct list_head pages;
struct sg_table table;
struct scatterlist *sg;
struct page *page, *tmp_page;
gfp_t gfp = __GFP_NOWARN | GFP_NOIO;
gfp_t gfp_comp = (gfp | __GFP_NORETRY | __GFP_COMP) & ~__GFP_RECLAIM;
int max_order = MIN(zfs_abd_scatter_max_order, MAX_ORDER - 1);
int nr_pages = abd_chunkcnt_for_bytes(size);
int chunks = 0, zones = 0;
size_t remaining_size;
int nid = NUMA_NO_NODE;
int alloc_pages = 0;
int order;
INIT_LIST_HEAD(&pages);
while (alloc_pages < nr_pages) {
unsigned long paddr;
unsigned chunk_pages;
order = MIN(highbit64(nr_pages - alloc_pages) - 1, max_order);
chunk_pages = (1U << order);
paddr = abd_alloc_chunk(nid, order ? gfp_comp : gfp, order);
if (paddr == 0) {
if (order == 0) {
ABDSTAT_BUMP(abdstat_scatter_page_alloc_retry);
schedule_timeout_interruptible(1);
} else {
max_order = MAX(0, order - 1);
}
continue;
}
page = virt_to_page(paddr);
list_add_tail(&page->lru, &pages);
if ((nid != NUMA_NO_NODE) && (page_to_nid(page) != nid))
zones++;
nid = page_to_nid(page);
ABDSTAT_BUMP(abdstat_scatter_orders[order]);
chunks++;
alloc_pages += chunk_pages;
}
ASSERT3S(alloc_pages, ==, nr_pages);
while (sg_alloc_table(&table, chunks, gfp)) {
ABDSTAT_BUMP(abdstat_scatter_sg_table_retry);
schedule_timeout_interruptible(1);
}
sg = table.sgl;
remaining_size = size;
list_for_each_entry_safe(page, tmp_page, &pages, lru) {
size_t sg_size = MIN(PAGESIZE << compound_order(page),
remaining_size);
sg_set_page(sg, page, sg_size, 0);
remaining_size -= sg_size;
sg = sg_next(sg);
list_del(&page->lru);
}
if (chunks > 1) {
ABDSTAT_BUMP(abdstat_scatter_page_multi_chunk);
abd->abd_flags |= ABD_FLAG_MULTI_CHUNK;
if (zones) {
ABDSTAT_BUMP(abdstat_scatter_page_multi_zone);
abd->abd_flags |= ABD_FLAG_MULTI_ZONE;
}
}
ABD_SCATTER(abd).abd_sgl = table.sgl;
ABD_SCATTER(abd).abd_nents = table.nents;
}
#else
/*
* Allocate N individual pages to construct a scatter ABD. This function
* makes no attempt to request contiguous pages and requires the minimal
* number of kernel interfaces. It's designed for maximum compatibility.
*/
static void
abd_alloc_pages(abd_t *abd, size_t size)
{
struct scatterlist *sg;
struct sg_table table;
struct page *page;
gfp_t gfp = __GFP_NOWARN | GFP_NOIO;
int nr_pages = abd_chunkcnt_for_bytes(size);
int i;
while (sg_alloc_table(&table, nr_pages, gfp)) {
ABDSTAT_BUMP(abdstat_scatter_sg_table_retry);
schedule_timeout_interruptible(1);
}
ASSERT3U(table.nents, ==, nr_pages);
ABD_SCATTER(abd).abd_sgl = table.sgl;
ABD_SCATTER(abd).abd_nents = nr_pages;
abd_for_each_sg(abd, sg, nr_pages, i) {
while ((page = __page_cache_alloc(gfp)) == NULL) {
ABDSTAT_BUMP(abdstat_scatter_page_alloc_retry);
schedule_timeout_interruptible(1);
}
ABDSTAT_BUMP(abdstat_scatter_orders[0]);
sg_set_page(sg, page, PAGESIZE, 0);
}
if (nr_pages > 1) {
ABDSTAT_BUMP(abdstat_scatter_page_multi_chunk);
abd->abd_flags |= ABD_FLAG_MULTI_CHUNK;
}
}
#endif /* !CONFIG_HIGHMEM */
static void
abd_free_pages(abd_t *abd)
{
struct scatterlist *sg;
struct sg_table table;
struct page *page;
int nr_pages = ABD_SCATTER(abd).abd_nents;
int order, i, j;
if (abd->abd_flags & ABD_FLAG_MULTI_ZONE)
ABDSTAT_BUMPDOWN(abdstat_scatter_page_multi_zone);
if (abd->abd_flags & ABD_FLAG_MULTI_CHUNK)
ABDSTAT_BUMPDOWN(abdstat_scatter_page_multi_chunk);
abd_for_each_sg(abd, sg, nr_pages, i) {
for (j = 0; j < sg->length; ) {
page = nth_page(sg_page(sg), j >> PAGE_SHIFT);
order = compound_order(page);
__free_pages(page, order);
j += (PAGESIZE << order);
ABDSTAT_BUMPDOWN(abdstat_scatter_orders[order]);
}
}
table.sgl = ABD_SCATTER(abd).abd_sgl;
table.nents = table.orig_nents = nr_pages;
sg_free_table(&table);
}
#else /* _KERNEL */
#ifndef PAGE_SHIFT
#define PAGE_SHIFT (highbit64(PAGESIZE)-1)
#endif
struct page;
#define kpm_enable 1
#define abd_alloc_chunk(o) \
((struct page *) umem_alloc_aligned(PAGESIZE << (o), 64, KM_SLEEP))
#define abd_free_chunk(chunk, o) umem_free(chunk, PAGESIZE << (o))
#define zfs_kmap_atomic(chunk, km) ((void *)chunk)
#define zfs_kunmap_atomic(addr, km) do { (void)(addr); } while (0)
#define local_irq_save(flags) do { (void)(flags); } while (0)
#define local_irq_restore(flags) do { (void)(flags); } while (0)
#define nth_page(pg, i) \
((struct page *)((void *)(pg) + (i) * PAGESIZE))
struct scatterlist {
struct page *page;
int length;
int end;
};
static void
sg_init_table(struct scatterlist *sg, int nr) {
memset(sg, 0, nr * sizeof (struct scatterlist));
sg[nr - 1].end = 1;
}
#define for_each_sg(sgl, sg, nr, i) \
for ((i) = 0, (sg) = (sgl); (i) < (nr); (i)++, (sg) = sg_next(sg))
static inline void
sg_set_page(struct scatterlist *sg, struct page *page, unsigned int len,
unsigned int offset)
{
/* currently we don't use offset */
ASSERT(offset == 0);
sg->page = page;
sg->length = len;
}
static inline struct page *
sg_page(struct scatterlist *sg)
{
return (sg->page);
}
static inline struct scatterlist *
sg_next(struct scatterlist *sg)
{
if (sg->end)
return (NULL);
return (sg + 1);
} }
static void static void
abd_free_chunk(struct page *c) abd_alloc_pages(abd_t *abd, size_t size)
{ {
__free_pages(c, 0); unsigned nr_pages = abd_chunkcnt_for_bytes(size);
struct scatterlist *sg;
int i;
ABD_SCATTER(abd).abd_sgl = vmem_alloc(nr_pages *
sizeof (struct scatterlist), KM_SLEEP);
sg_init_table(ABD_SCATTER(abd).abd_sgl, nr_pages);
abd_for_each_sg(abd, sg, nr_pages, i) {
struct page *p = abd_alloc_chunk(0);
sg_set_page(sg, p, PAGESIZE, 0);
}
ABD_SCATTER(abd).abd_nents = nr_pages;
} }
static void
abd_free_pages(abd_t *abd)
{
int i, n = ABD_SCATTER(abd).abd_nents;
struct scatterlist *sg;
int j;
abd_for_each_sg(abd, sg, n, i) {
for (j = 0; j < sg->length; j += PAGESIZE) {
struct page *p = nth_page(sg_page(sg), j>>PAGE_SHIFT);
abd_free_chunk(p, 0);
}
}
vmem_free(ABD_SCATTER(abd).abd_sgl, n * sizeof (struct scatterlist));
}
#endif /* _KERNEL */
void void
abd_init(void) abd_init(void)
{ {
int i;
abd_cache = kmem_cache_create("abd_t", sizeof (abd_t),
0, NULL, NULL, NULL, NULL, NULL, 0);
abd_ksp = kstat_create("zfs", 0, "abdstats", "misc", KSTAT_TYPE_NAMED, abd_ksp = kstat_create("zfs", 0, "abdstats", "misc", KSTAT_TYPE_NAMED,
sizeof (abd_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); sizeof (abd_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
if (abd_ksp != NULL) { if (abd_ksp != NULL) {
abd_ksp->ks_data = &abd_stats; abd_ksp->ks_data = &abd_stats;
kstat_install(abd_ksp); kstat_install(abd_ksp);
for (i = 0; i < MAX_ORDER; i++) {
snprintf(abd_stats.abdstat_scatter_orders[i].name,
KSTAT_STRLEN, "scatter_order_%d", i);
abd_stats.abdstat_scatter_orders[i].data_type =
KSTAT_DATA_UINT64;
}
} }
} }
@ -208,44 +523,11 @@ abd_fini(void)
kstat_delete(abd_ksp); kstat_delete(abd_ksp);
abd_ksp = NULL; abd_ksp = NULL;
} }
}
#else if (abd_cache) {
kmem_cache_destroy(abd_cache);
struct page; abd_cache = NULL;
#define kpm_enable 1 }
#define abd_alloc_chunk() \
((struct page *) umem_alloc_aligned(PAGESIZE, 64, KM_SLEEP))
#define abd_free_chunk(chunk) umem_free(chunk, PAGESIZE)
#define zfs_kmap_atomic(chunk, km) ((void *)chunk)
#define zfs_kunmap_atomic(addr, km) do { (void)(addr); } while (0)
#define local_irq_save(flags) do { (void)(flags); } while (0)
#define local_irq_restore(flags) do { (void)(flags); } while (0)
void
abd_init(void)
{
}
void
abd_fini(void)
{
}
#endif /* _KERNEL */
static inline size_t
abd_chunkcnt_for_bytes(size_t size)
{
return (P2ROUNDUP(size, PAGESIZE) / PAGESIZE);
}
static inline size_t
abd_scatter_chunkcnt(abd_t *abd)
{
ASSERT(!abd_is_linear(abd));
return (abd_chunkcnt_for_bytes(
abd->abd_u.abd_scatter.abd_offset + abd->abd_size));
} }
static inline void static inline void
@ -254,7 +536,8 @@ abd_verify(abd_t *abd)
ASSERT3U(abd->abd_size, >, 0); ASSERT3U(abd->abd_size, >, 0);
ASSERT3U(abd->abd_size, <=, SPA_MAXBLOCKSIZE); ASSERT3U(abd->abd_size, <=, SPA_MAXBLOCKSIZE);
ASSERT3U(abd->abd_flags, ==, abd->abd_flags & (ABD_FLAG_LINEAR | ASSERT3U(abd->abd_flags, ==, abd->abd_flags & (ABD_FLAG_LINEAR |
ABD_FLAG_OWNER | ABD_FLAG_META)); ABD_FLAG_OWNER | ABD_FLAG_META | ABD_FLAG_MULTI_ZONE |
ABD_FLAG_MULTI_CHUNK));
IMPLY(abd->abd_parent != NULL, !(abd->abd_flags & ABD_FLAG_OWNER)); IMPLY(abd->abd_parent != NULL, !(abd->abd_flags & ABD_FLAG_OWNER));
IMPLY(abd->abd_flags & ABD_FLAG_META, abd->abd_flags & ABD_FLAG_OWNER); IMPLY(abd->abd_flags & ABD_FLAG_META, abd->abd_flags & ABD_FLAG_OWNER);
if (abd_is_linear(abd)) { if (abd_is_linear(abd)) {
@ -262,23 +545,25 @@ abd_verify(abd_t *abd)
} else { } else {
size_t n; size_t n;
int i; int i;
struct scatterlist *sg;
ASSERT3U(abd->abd_u.abd_scatter.abd_offset, <, PAGESIZE); ASSERT3U(ABD_SCATTER(abd).abd_nents, >, 0);
n = abd_scatter_chunkcnt(abd); ASSERT3U(ABD_SCATTER(abd).abd_offset, <,
for (i = 0; i < n; i++) { ABD_SCATTER(abd).abd_sgl->length);
ASSERT3P( n = ABD_SCATTER(abd).abd_nents;
abd->abd_u.abd_scatter.abd_chunks[i], !=, NULL); abd_for_each_sg(abd, sg, n, i) {
ASSERT3P(sg_page(sg), !=, NULL);
} }
} }
} }
static inline abd_t * static inline abd_t *
abd_alloc_struct(size_t chunkcnt) abd_alloc_struct(void)
{ {
size_t size = offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]); abd_t *abd = kmem_cache_alloc(abd_cache, KM_PUSHPAGE);
abd_t *abd = kmem_alloc(size, KM_PUSHPAGE);
ASSERT3P(abd, !=, NULL); ASSERT3P(abd, !=, NULL);
ABDSTAT_INCR(abdstat_struct_size, size); ABDSTAT_INCR(abdstat_struct_size, sizeof (abd_t));
return (abd); return (abd);
} }
@ -286,10 +571,8 @@ abd_alloc_struct(size_t chunkcnt)
static inline void static inline void
abd_free_struct(abd_t *abd) abd_free_struct(abd_t *abd)
{ {
size_t chunkcnt = abd_is_linear(abd) ? 0 : abd_scatter_chunkcnt(abd); kmem_cache_free(abd_cache, abd);
int size = offsetof(abd_t, abd_u.abd_scatter.abd_chunks[chunkcnt]); ABDSTAT_INCR(abdstat_struct_size, -sizeof (abd_t));
kmem_free(abd, size);
ABDSTAT_INCR(abdstat_struct_size, -size);
} }
/* /*
@ -299,19 +582,17 @@ abd_free_struct(abd_t *abd)
abd_t * abd_t *
abd_alloc(size_t size, boolean_t is_metadata) abd_alloc(size_t size, boolean_t is_metadata)
{ {
int i;
size_t n;
abd_t *abd; abd_t *abd;
if (!zfs_abd_scatter_enabled) if (!zfs_abd_scatter_enabled || size <= PAGESIZE)
return (abd_alloc_linear(size, is_metadata)); return (abd_alloc_linear(size, is_metadata));
VERIFY3U(size, <=, SPA_MAXBLOCKSIZE); VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
n = abd_chunkcnt_for_bytes(size); abd = abd_alloc_struct();
abd = abd_alloc_struct(n);
abd->abd_flags = ABD_FLAG_OWNER; abd->abd_flags = ABD_FLAG_OWNER;
abd_alloc_pages(abd, size);
if (is_metadata) { if (is_metadata) {
abd->abd_flags |= ABD_FLAG_META; abd->abd_flags |= ABD_FLAG_META;
} }
@ -320,18 +601,11 @@ abd_alloc(size_t size, boolean_t is_metadata)
refcount_create(&abd->abd_children); refcount_create(&abd->abd_children);
abd->abd_u.abd_scatter.abd_offset = 0; abd->abd_u.abd_scatter.abd_offset = 0;
abd->abd_u.abd_scatter.abd_chunk_size = PAGESIZE;
for (i = 0; i < n; i++) {
void *c = abd_alloc_chunk();
ASSERT3P(c, !=, NULL);
abd->abd_u.abd_scatter.abd_chunks[i] = c;
}
ABDSTAT_BUMP(abdstat_scatter_cnt); ABDSTAT_BUMP(abdstat_scatter_cnt);
ABDSTAT_INCR(abdstat_scatter_data_size, size); ABDSTAT_INCR(abdstat_scatter_data_size, size);
ABDSTAT_INCR(abdstat_scatter_chunk_waste, ABDSTAT_INCR(abdstat_scatter_chunk_waste,
n * PAGESIZE - size); P2ROUNDUP(size, PAGESIZE) - size);
return (abd); return (abd);
} }
@ -339,18 +613,13 @@ abd_alloc(size_t size, boolean_t is_metadata)
static void static void
abd_free_scatter(abd_t *abd) abd_free_scatter(abd_t *abd)
{ {
size_t n = abd_scatter_chunkcnt(abd); abd_free_pages(abd);
int i;
for (i = 0; i < n; i++) {
abd_free_chunk(abd->abd_u.abd_scatter.abd_chunks[i]);
}
refcount_destroy(&abd->abd_children); refcount_destroy(&abd->abd_children);
ABDSTAT_BUMPDOWN(abdstat_scatter_cnt); ABDSTAT_BUMPDOWN(abdstat_scatter_cnt);
ABDSTAT_INCR(abdstat_scatter_data_size, -(int)abd->abd_size); ABDSTAT_INCR(abdstat_scatter_data_size, -(int)abd->abd_size);
ABDSTAT_INCR(abdstat_scatter_chunk_waste, ABDSTAT_INCR(abdstat_scatter_chunk_waste,
abd->abd_size - n * PAGESIZE); abd->abd_size - P2ROUNDUP(abd->abd_size, PAGESIZE));
abd_free_struct(abd); abd_free_struct(abd);
} }
@ -363,7 +632,7 @@ abd_free_scatter(abd_t *abd)
abd_t * abd_t *
abd_alloc_linear(size_t size, boolean_t is_metadata) abd_alloc_linear(size_t size, boolean_t is_metadata)
{ {
abd_t *abd = abd_alloc_struct(0); abd_t *abd = abd_alloc_struct();
VERIFY3U(size, <=, SPA_MAXBLOCKSIZE); VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
@ -469,7 +738,7 @@ abd_get_offset_impl(abd_t *sabd, size_t off, size_t size)
ASSERT3U(off, <=, sabd->abd_size); ASSERT3U(off, <=, sabd->abd_size);
if (abd_is_linear(sabd)) { if (abd_is_linear(sabd)) {
abd = abd_alloc_struct(0); abd = abd_alloc_struct();
/* /*
* Even if this buf is filesystem metadata, we only track that * Even if this buf is filesystem metadata, we only track that
@ -481,11 +750,11 @@ abd_get_offset_impl(abd_t *sabd, size_t off, size_t size)
abd->abd_u.abd_linear.abd_buf = abd->abd_u.abd_linear.abd_buf =
(char *)sabd->abd_u.abd_linear.abd_buf + off; (char *)sabd->abd_u.abd_linear.abd_buf + off;
} else { } else {
int i;
struct scatterlist *sg;
size_t new_offset = sabd->abd_u.abd_scatter.abd_offset + off; size_t new_offset = sabd->abd_u.abd_scatter.abd_offset + off;
size_t chunkcnt = abd_chunkcnt_for_bytes(size +
new_offset % PAGESIZE);
abd = abd_alloc_struct(chunkcnt); abd = abd_alloc_struct();
/* /*
* Even if this buf is filesystem metadata, we only track that * Even if this buf is filesystem metadata, we only track that
@ -494,13 +763,15 @@ abd_get_offset_impl(abd_t *sabd, size_t off, size_t size)
*/ */
abd->abd_flags = 0; abd->abd_flags = 0;
abd->abd_u.abd_scatter.abd_offset = new_offset % PAGESIZE; abd_for_each_sg(sabd, sg, ABD_SCATTER(sabd).abd_nents, i) {
abd->abd_u.abd_scatter.abd_chunk_size = PAGESIZE; if (new_offset < sg->length)
break;
new_offset -= sg->length;
}
/* Copy the scatterlist starting at the correct offset */ ABD_SCATTER(abd).abd_sgl = sg;
(void) memcpy(&abd->abd_u.abd_scatter.abd_chunks, ABD_SCATTER(abd).abd_offset = new_offset;
&sabd->abd_u.abd_scatter.abd_chunks[new_offset / PAGESIZE], ABD_SCATTER(abd).abd_nents = ABD_SCATTER(sabd).abd_nents - i;
chunkcnt * sizeof (void *));
} }
abd->abd_size = size; abd->abd_size = size;
@ -536,7 +807,7 @@ abd_get_offset_size(abd_t *sabd, size_t off, size_t size)
abd_t * abd_t *
abd_get_from_buf(void *buf, size_t size) abd_get_from_buf(void *buf, size_t size)
{ {
abd_t *abd = abd_alloc_struct(0); abd_t *abd = abd_alloc_struct();
VERIFY3U(size, <=, SPA_MAXBLOCKSIZE); VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
@ -698,31 +969,21 @@ int km_table[NR_KM_TYPE] = {
#endif #endif
struct abd_iter { struct abd_iter {
abd_t *iter_abd; /* ABD being iterated through */ /* public interface */
size_t iter_pos; /* position (relative to abd_offset) */
void *iter_mapaddr; /* addr corresponding to iter_pos */ void *iter_mapaddr; /* addr corresponding to iter_pos */
size_t iter_mapsize; /* length of data valid at mapaddr */ size_t iter_mapsize; /* length of data valid at mapaddr */
/* private */
abd_t *iter_abd; /* ABD being iterated through */
size_t iter_pos;
size_t iter_offset; /* offset in current sg/abd_buf, */
/* abd_offset included */
struct scatterlist *iter_sg; /* current sg */
#ifndef HAVE_1ARG_KMAP_ATOMIC #ifndef HAVE_1ARG_KMAP_ATOMIC
int iter_km; /* KM_* for kmap_atomic */ int iter_km; /* KM_* for kmap_atomic */
#endif #endif
}; };
static inline size_t
abd_iter_scatter_chunk_offset(struct abd_iter *aiter)
{
ASSERT(!abd_is_linear(aiter->iter_abd));
return ((aiter->iter_abd->abd_u.abd_scatter.abd_offset +
aiter->iter_pos) % PAGESIZE);
}
static inline size_t
abd_iter_scatter_chunk_index(struct abd_iter *aiter)
{
ASSERT(!abd_is_linear(aiter->iter_abd));
return ((aiter->iter_abd->abd_u.abd_scatter.abd_offset +
aiter->iter_pos) / PAGESIZE);
}
/* /*
* Initialize the abd_iter. * Initialize the abd_iter.
*/ */
@ -731,9 +992,16 @@ abd_iter_init(struct abd_iter *aiter, abd_t *abd, int km_type)
{ {
abd_verify(abd); abd_verify(abd);
aiter->iter_abd = abd; aiter->iter_abd = abd;
aiter->iter_pos = 0;
aiter->iter_mapaddr = NULL; aiter->iter_mapaddr = NULL;
aiter->iter_mapsize = 0; aiter->iter_mapsize = 0;
aiter->iter_pos = 0;
if (abd_is_linear(abd)) {
aiter->iter_offset = 0;
aiter->iter_sg = NULL;
} else {
aiter->iter_offset = ABD_SCATTER(abd).abd_offset;
aiter->iter_sg = ABD_SCATTER(abd).abd_sgl;
}
#ifndef HAVE_1ARG_KMAP_ATOMIC #ifndef HAVE_1ARG_KMAP_ATOMIC
ASSERT3U(km_type, <, NR_KM_TYPE); ASSERT3U(km_type, <, NR_KM_TYPE);
aiter->iter_km = km_type; aiter->iter_km = km_type;
@ -756,6 +1024,17 @@ abd_iter_advance(struct abd_iter *aiter, size_t amount)
return; return;
aiter->iter_pos += amount; aiter->iter_pos += amount;
aiter->iter_offset += amount;
if (!abd_is_linear(aiter->iter_abd)) {
while (aiter->iter_offset >= aiter->iter_sg->length) {
aiter->iter_offset -= aiter->iter_sg->length;
aiter->iter_sg = sg_next(aiter->iter_sg);
if (aiter->iter_sg == NULL) {
ASSERT0(aiter->iter_offset);
break;
}
}
}
} }
/* /*
@ -776,18 +1055,16 @@ abd_iter_map(struct abd_iter *aiter)
return; return;
if (abd_is_linear(aiter->iter_abd)) { if (abd_is_linear(aiter->iter_abd)) {
offset = aiter->iter_pos; ASSERT3U(aiter->iter_pos, ==, aiter->iter_offset);
offset = aiter->iter_offset;
aiter->iter_mapsize = aiter->iter_abd->abd_size - offset; aiter->iter_mapsize = aiter->iter_abd->abd_size - offset;
paddr = aiter->iter_abd->abd_u.abd_linear.abd_buf; paddr = aiter->iter_abd->abd_u.abd_linear.abd_buf;
} else { } else {
size_t index = abd_iter_scatter_chunk_index(aiter); offset = aiter->iter_offset;
offset = abd_iter_scatter_chunk_offset(aiter); aiter->iter_mapsize = MIN(aiter->iter_sg->length - offset,
aiter->iter_mapsize = MIN(PAGESIZE - offset,
aiter->iter_abd->abd_size - aiter->iter_pos); aiter->iter_abd->abd_size - aiter->iter_pos);
paddr = zfs_kmap_atomic( paddr = zfs_kmap_atomic(sg_page(aiter->iter_sg),
aiter->iter_abd->abd_u.abd_scatter.abd_chunks[index],
km_table[aiter->iter_km]); km_table[aiter->iter_km]);
} }
@ -807,8 +1084,7 @@ abd_iter_unmap(struct abd_iter *aiter)
if (!abd_is_linear(aiter->iter_abd)) { if (!abd_is_linear(aiter->iter_abd)) {
/* LINTED E_FUNC_SET_NOT_USED */ /* LINTED E_FUNC_SET_NOT_USED */
zfs_kunmap_atomic(aiter->iter_mapaddr - zfs_kunmap_atomic(aiter->iter_mapaddr - aiter->iter_offset,
abd_iter_scatter_chunk_offset(aiter),
km_table[aiter->iter_km]); km_table[aiter->iter_km]);
} }
@ -1234,17 +1510,19 @@ abd_scatter_bio_map_off(struct bio *bio, abd_t *abd,
for (i = 0; i < bio->bi_max_vecs; i++) { for (i = 0; i < bio->bi_max_vecs; i++) {
struct page *pg; struct page *pg;
size_t len, pgoff, index; size_t len, sgoff, pgoff;
struct scatterlist *sg;
if (io_size <= 0) if (io_size <= 0)
break; break;
pgoff = abd_iter_scatter_chunk_offset(&aiter); sg = aiter.iter_sg;
sgoff = aiter.iter_offset;
pgoff = sgoff & (PAGESIZE - 1);
len = MIN(io_size, PAGESIZE - pgoff); len = MIN(io_size, PAGESIZE - pgoff);
ASSERT(len > 0); ASSERT(len > 0);
index = abd_iter_scatter_chunk_index(&aiter); pg = nth_page(sg_page(sg), sgoff >> PAGE_SHIFT);
pg = abd->abd_u.abd_scatter.abd_chunks[index];
if (bio_add_page(bio, pg, len, pgoff) != len) if (bio_add_page(bio, pg, len, pgoff) != len)
break; break;
@ -1259,4 +1537,7 @@ abd_scatter_bio_map_off(struct bio *bio, abd_t *abd,
module_param(zfs_abd_scatter_enabled, int, 0644); module_param(zfs_abd_scatter_enabled, int, 0644);
MODULE_PARM_DESC(zfs_abd_scatter_enabled, MODULE_PARM_DESC(zfs_abd_scatter_enabled,
"Toggle whether ABD allocations must be linear."); "Toggle whether ABD allocations must be linear.");
module_param(zfs_abd_scatter_max_order, uint, 0644);
MODULE_PARM_DESC(zfs_abd_scatter_max_order,
"Maximum order allocation used for a scatter ABD.");
#endif #endif