server-skynet-source-3rd-je.../src/huge.c

420 lines
9.2 KiB
C
Raw Normal View History

#define JEMALLOC_HUGE_C_
2010-02-12 06:45:59 +08:00
#include "jemalloc/internal/jemalloc_internal.h"
/******************************************************************************/
/* Data. */
/* Protects chunk-related data structures. */
static malloc_mutex_t huge_mtx;
/******************************************************************************/
/* Tree of chunks that are stand-alone huge allocations. */
static extent_tree_t huge;
void *
huge_malloc(tsd_t *tsd, arena_t *arena, size_t size, bool zero)
{
size_t usize;
usize = s2u(size);
if (usize == 0) {
/* size_t overflow. */
return (NULL);
}
return (huge_palloc(tsd, arena, usize, chunksize, zero));
}
void *
huge_palloc(tsd_t *tsd, arena_t *arena, size_t usize, size_t alignment,
bool zero)
{
void *ret;
size_t csize;
extent_node_t *node;
bool is_zeroed;
/* Allocate one or more contiguous chunks for this request. */
csize = CHUNK_CEILING(usize);
assert(csize >= usize);
/* Allocate an extent node with which to track the chunk. */
node = base_node_alloc();
if (node == NULL)
return (NULL);
/*
* Copy zero into is_zeroed and pass the copy to chunk_alloc(), so that
* it is possible to make correct junk/zero fill decisions below.
*/
is_zeroed = zero;
arena = choose_arena(tsd, arena);
Attempt to expand huge allocations in-place. This adds support for expanding huge allocations in-place by requesting memory at a specific address from the chunk allocator. It's currently only implemented for the chunk recycling path, although in theory it could also be done by optimistically allocating new chunks. On Linux, it could attempt an in-place mremap. However, that won't work in practice since the heap is grown downwards and memory is not unmapped (in a normal build, at least). Repeated vector reallocation micro-benchmark: #include <string.h> #include <stdlib.h> int main(void) { for (size_t i = 0; i < 100; i++) { void *ptr = NULL; size_t old_size = 0; for (size_t size = 4; size < (1 << 30); size *= 2) { ptr = realloc(ptr, size); if (!ptr) return 1; memset(ptr + old_size, 0xff, size - old_size); old_size = size; } free(ptr); } } The glibc allocator fails to do any in-place reallocations on this benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it elides the cost of copies via mremap, which is currently not something that jemalloc can use. With this improvement, jemalloc still fails to do any in-place huge reallocations for the first outer loop, but then succeeds 100% of the time for the remaining 99 iterations. The time spent doing allocations and copies drops down to under 5%, with nearly all of it spent doing purging + faulting (when huge pages are disabled) and the array memset. An improved mremap API (MREMAP_RETAIN - #138) would be far more general but this is a portable optimization and would still be useful on Linux for xallocx. Numbers with transparent huge pages enabled: glibc (copies elided via MREMAP_MAYMOVE): 8.471s jemalloc: 17.816s jemalloc + no-op madvise: 13.236s jemalloc + this commit: 6.787s jemalloc + this commit + no-op madvise: 6.144s Numbers with transparent huge pages disabled: glibc (copies elided via MREMAP_MAYMOVE): 15.403s jemalloc: 39.456s jemalloc + no-op madvise: 12.768s jemalloc + this commit: 15.534s jemalloc + this commit + no-op madvise: 6.354s Closes #137
2014-10-04 13:39:32 +08:00
ret = arena_chunk_alloc_huge(arena, NULL, csize, alignment, &is_zeroed);
if (ret == NULL) {
base_node_dalloc(node);
return (NULL);
}
/* Insert node into huge. */
node->addr = ret;
node->size = usize;
node->arena = arena;
malloc_mutex_lock(&huge_mtx);
extent_tree_ad_insert(&huge, node);
malloc_mutex_unlock(&huge_mtx);
if (config_fill && !zero) {
if (unlikely(opt_junk))
memset(ret, 0xa5, usize);
else if (unlikely(opt_zero) && !is_zeroed)
memset(ret, 0, usize);
}
return (ret);
}
#ifdef JEMALLOC_JET
#undef huge_dalloc_junk
#define huge_dalloc_junk JEMALLOC_N(huge_dalloc_junk_impl)
#endif
static void
huge_dalloc_junk(void *ptr, size_t usize)
{
if (config_fill && have_dss && unlikely(opt_junk)) {
/*
* Only bother junk filling if the chunk isn't about to be
* unmapped.
*/
if (!config_munmap || (have_dss && chunk_in_dss(ptr)))
memset(ptr, 0x5a, usize);
}
}
#ifdef JEMALLOC_JET
#undef huge_dalloc_junk
#define huge_dalloc_junk JEMALLOC_N(huge_dalloc_junk)
huge_dalloc_junk_t *huge_dalloc_junk = JEMALLOC_N(huge_dalloc_junk_impl);
#endif
Attempt to expand huge allocations in-place. This adds support for expanding huge allocations in-place by requesting memory at a specific address from the chunk allocator. It's currently only implemented for the chunk recycling path, although in theory it could also be done by optimistically allocating new chunks. On Linux, it could attempt an in-place mremap. However, that won't work in practice since the heap is grown downwards and memory is not unmapped (in a normal build, at least). Repeated vector reallocation micro-benchmark: #include <string.h> #include <stdlib.h> int main(void) { for (size_t i = 0; i < 100; i++) { void *ptr = NULL; size_t old_size = 0; for (size_t size = 4; size < (1 << 30); size *= 2) { ptr = realloc(ptr, size); if (!ptr) return 1; memset(ptr + old_size, 0xff, size - old_size); old_size = size; } free(ptr); } } The glibc allocator fails to do any in-place reallocations on this benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it elides the cost of copies via mremap, which is currently not something that jemalloc can use. With this improvement, jemalloc still fails to do any in-place huge reallocations for the first outer loop, but then succeeds 100% of the time for the remaining 99 iterations. The time spent doing allocations and copies drops down to under 5%, with nearly all of it spent doing purging + faulting (when huge pages are disabled) and the array memset. An improved mremap API (MREMAP_RETAIN - #138) would be far more general but this is a portable optimization and would still be useful on Linux for xallocx. Numbers with transparent huge pages enabled: glibc (copies elided via MREMAP_MAYMOVE): 8.471s jemalloc: 17.816s jemalloc + no-op madvise: 13.236s jemalloc + this commit: 6.787s jemalloc + this commit + no-op madvise: 6.144s Numbers with transparent huge pages disabled: glibc (copies elided via MREMAP_MAYMOVE): 15.403s jemalloc: 39.456s jemalloc + no-op madvise: 12.768s jemalloc + this commit: 15.534s jemalloc + this commit + no-op madvise: 6.354s Closes #137
2014-10-04 13:39:32 +08:00
static bool
huge_ralloc_no_move_expand(void *ptr, size_t oldsize, size_t size, bool zero) {
size_t usize;
Attempt to expand huge allocations in-place. This adds support for expanding huge allocations in-place by requesting memory at a specific address from the chunk allocator. It's currently only implemented for the chunk recycling path, although in theory it could also be done by optimistically allocating new chunks. On Linux, it could attempt an in-place mremap. However, that won't work in practice since the heap is grown downwards and memory is not unmapped (in a normal build, at least). Repeated vector reallocation micro-benchmark: #include <string.h> #include <stdlib.h> int main(void) { for (size_t i = 0; i < 100; i++) { void *ptr = NULL; size_t old_size = 0; for (size_t size = 4; size < (1 << 30); size *= 2) { ptr = realloc(ptr, size); if (!ptr) return 1; memset(ptr + old_size, 0xff, size - old_size); old_size = size; } free(ptr); } } The glibc allocator fails to do any in-place reallocations on this benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it elides the cost of copies via mremap, which is currently not something that jemalloc can use. With this improvement, jemalloc still fails to do any in-place huge reallocations for the first outer loop, but then succeeds 100% of the time for the remaining 99 iterations. The time spent doing allocations and copies drops down to under 5%, with nearly all of it spent doing purging + faulting (when huge pages are disabled) and the array memset. An improved mremap API (MREMAP_RETAIN - #138) would be far more general but this is a portable optimization and would still be useful on Linux for xallocx. Numbers with transparent huge pages enabled: glibc (copies elided via MREMAP_MAYMOVE): 8.471s jemalloc: 17.816s jemalloc + no-op madvise: 13.236s jemalloc + this commit: 6.787s jemalloc + this commit + no-op madvise: 6.144s Numbers with transparent huge pages disabled: glibc (copies elided via MREMAP_MAYMOVE): 15.403s jemalloc: 39.456s jemalloc + no-op madvise: 12.768s jemalloc + this commit: 15.534s jemalloc + this commit + no-op madvise: 6.354s Closes #137
2014-10-04 13:39:32 +08:00
void *expand_addr;
size_t expand_size;
extent_node_t *node, key;
arena_t *arena;
bool is_zeroed;
void *ret;
usize = s2u(size);
if (usize == 0) {
/* size_t overflow. */
Attempt to expand huge allocations in-place. This adds support for expanding huge allocations in-place by requesting memory at a specific address from the chunk allocator. It's currently only implemented for the chunk recycling path, although in theory it could also be done by optimistically allocating new chunks. On Linux, it could attempt an in-place mremap. However, that won't work in practice since the heap is grown downwards and memory is not unmapped (in a normal build, at least). Repeated vector reallocation micro-benchmark: #include <string.h> #include <stdlib.h> int main(void) { for (size_t i = 0; i < 100; i++) { void *ptr = NULL; size_t old_size = 0; for (size_t size = 4; size < (1 << 30); size *= 2) { ptr = realloc(ptr, size); if (!ptr) return 1; memset(ptr + old_size, 0xff, size - old_size); old_size = size; } free(ptr); } } The glibc allocator fails to do any in-place reallocations on this benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it elides the cost of copies via mremap, which is currently not something that jemalloc can use. With this improvement, jemalloc still fails to do any in-place huge reallocations for the first outer loop, but then succeeds 100% of the time for the remaining 99 iterations. The time spent doing allocations and copies drops down to under 5%, with nearly all of it spent doing purging + faulting (when huge pages are disabled) and the array memset. An improved mremap API (MREMAP_RETAIN - #138) would be far more general but this is a portable optimization and would still be useful on Linux for xallocx. Numbers with transparent huge pages enabled: glibc (copies elided via MREMAP_MAYMOVE): 8.471s jemalloc: 17.816s jemalloc + no-op madvise: 13.236s jemalloc + this commit: 6.787s jemalloc + this commit + no-op madvise: 6.144s Numbers with transparent huge pages disabled: glibc (copies elided via MREMAP_MAYMOVE): 15.403s jemalloc: 39.456s jemalloc + no-op madvise: 12.768s jemalloc + this commit: 15.534s jemalloc + this commit + no-op madvise: 6.354s Closes #137
2014-10-04 13:39:32 +08:00
return (true);
}
expand_addr = ptr + CHUNK_CEILING(oldsize);
expand_size = CHUNK_CEILING(usize) - CHUNK_CEILING(oldsize);
Attempt to expand huge allocations in-place. This adds support for expanding huge allocations in-place by requesting memory at a specific address from the chunk allocator. It's currently only implemented for the chunk recycling path, although in theory it could also be done by optimistically allocating new chunks. On Linux, it could attempt an in-place mremap. However, that won't work in practice since the heap is grown downwards and memory is not unmapped (in a normal build, at least). Repeated vector reallocation micro-benchmark: #include <string.h> #include <stdlib.h> int main(void) { for (size_t i = 0; i < 100; i++) { void *ptr = NULL; size_t old_size = 0; for (size_t size = 4; size < (1 << 30); size *= 2) { ptr = realloc(ptr, size); if (!ptr) return 1; memset(ptr + old_size, 0xff, size - old_size); old_size = size; } free(ptr); } } The glibc allocator fails to do any in-place reallocations on this benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it elides the cost of copies via mremap, which is currently not something that jemalloc can use. With this improvement, jemalloc still fails to do any in-place huge reallocations for the first outer loop, but then succeeds 100% of the time for the remaining 99 iterations. The time spent doing allocations and copies drops down to under 5%, with nearly all of it spent doing purging + faulting (when huge pages are disabled) and the array memset. An improved mremap API (MREMAP_RETAIN - #138) would be far more general but this is a portable optimization and would still be useful on Linux for xallocx. Numbers with transparent huge pages enabled: glibc (copies elided via MREMAP_MAYMOVE): 8.471s jemalloc: 17.816s jemalloc + no-op madvise: 13.236s jemalloc + this commit: 6.787s jemalloc + this commit + no-op madvise: 6.144s Numbers with transparent huge pages disabled: glibc (copies elided via MREMAP_MAYMOVE): 15.403s jemalloc: 39.456s jemalloc + no-op madvise: 12.768s jemalloc + this commit: 15.534s jemalloc + this commit + no-op madvise: 6.354s Closes #137
2014-10-04 13:39:32 +08:00
malloc_mutex_lock(&huge_mtx);
key.addr = ptr;
node = extent_tree_ad_search(&huge, &key);
assert(node != NULL);
assert(node->addr == ptr);
/* Find the current arena. */
arena = node->arena;
malloc_mutex_unlock(&huge_mtx);
/*
* Copy zero into is_zeroed and pass the copy to chunk_alloc(), so that
* it is possible to make correct junk/zero fill decisions below.
*/
is_zeroed = zero;
ret = arena_chunk_alloc_huge(arena, expand_addr, expand_size, chunksize,
&is_zeroed);
if (ret == NULL)
return (true);
assert(ret == expand_addr);
malloc_mutex_lock(&huge_mtx);
/* Update the size of the huge allocation. */
node->size = usize;
Attempt to expand huge allocations in-place. This adds support for expanding huge allocations in-place by requesting memory at a specific address from the chunk allocator. It's currently only implemented for the chunk recycling path, although in theory it could also be done by optimistically allocating new chunks. On Linux, it could attempt an in-place mremap. However, that won't work in practice since the heap is grown downwards and memory is not unmapped (in a normal build, at least). Repeated vector reallocation micro-benchmark: #include <string.h> #include <stdlib.h> int main(void) { for (size_t i = 0; i < 100; i++) { void *ptr = NULL; size_t old_size = 0; for (size_t size = 4; size < (1 << 30); size *= 2) { ptr = realloc(ptr, size); if (!ptr) return 1; memset(ptr + old_size, 0xff, size - old_size); old_size = size; } free(ptr); } } The glibc allocator fails to do any in-place reallocations on this benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it elides the cost of copies via mremap, which is currently not something that jemalloc can use. With this improvement, jemalloc still fails to do any in-place huge reallocations for the first outer loop, but then succeeds 100% of the time for the remaining 99 iterations. The time spent doing allocations and copies drops down to under 5%, with nearly all of it spent doing purging + faulting (when huge pages are disabled) and the array memset. An improved mremap API (MREMAP_RETAIN - #138) would be far more general but this is a portable optimization and would still be useful on Linux for xallocx. Numbers with transparent huge pages enabled: glibc (copies elided via MREMAP_MAYMOVE): 8.471s jemalloc: 17.816s jemalloc + no-op madvise: 13.236s jemalloc + this commit: 6.787s jemalloc + this commit + no-op madvise: 6.144s Numbers with transparent huge pages disabled: glibc (copies elided via MREMAP_MAYMOVE): 15.403s jemalloc: 39.456s jemalloc + no-op madvise: 12.768s jemalloc + this commit: 15.534s jemalloc + this commit + no-op madvise: 6.354s Closes #137
2014-10-04 13:39:32 +08:00
malloc_mutex_unlock(&huge_mtx);
if (config_fill && !zero) {
if (unlikely(opt_junk))
memset(ptr + oldsize, 0xa5, usize - oldsize);
Attempt to expand huge allocations in-place. This adds support for expanding huge allocations in-place by requesting memory at a specific address from the chunk allocator. It's currently only implemented for the chunk recycling path, although in theory it could also be done by optimistically allocating new chunks. On Linux, it could attempt an in-place mremap. However, that won't work in practice since the heap is grown downwards and memory is not unmapped (in a normal build, at least). Repeated vector reallocation micro-benchmark: #include <string.h> #include <stdlib.h> int main(void) { for (size_t i = 0; i < 100; i++) { void *ptr = NULL; size_t old_size = 0; for (size_t size = 4; size < (1 << 30); size *= 2) { ptr = realloc(ptr, size); if (!ptr) return 1; memset(ptr + old_size, 0xff, size - old_size); old_size = size; } free(ptr); } } The glibc allocator fails to do any in-place reallocations on this benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it elides the cost of copies via mremap, which is currently not something that jemalloc can use. With this improvement, jemalloc still fails to do any in-place huge reallocations for the first outer loop, but then succeeds 100% of the time for the remaining 99 iterations. The time spent doing allocations and copies drops down to under 5%, with nearly all of it spent doing purging + faulting (when huge pages are disabled) and the array memset. An improved mremap API (MREMAP_RETAIN - #138) would be far more general but this is a portable optimization and would still be useful on Linux for xallocx. Numbers with transparent huge pages enabled: glibc (copies elided via MREMAP_MAYMOVE): 8.471s jemalloc: 17.816s jemalloc + no-op madvise: 13.236s jemalloc + this commit: 6.787s jemalloc + this commit + no-op madvise: 6.144s Numbers with transparent huge pages disabled: glibc (copies elided via MREMAP_MAYMOVE): 15.403s jemalloc: 39.456s jemalloc + no-op madvise: 12.768s jemalloc + this commit: 15.534s jemalloc + this commit + no-op madvise: 6.354s Closes #137
2014-10-04 13:39:32 +08:00
else if (unlikely(opt_zero) && !is_zeroed)
memset(ptr + oldsize, 0, usize - oldsize);
Attempt to expand huge allocations in-place. This adds support for expanding huge allocations in-place by requesting memory at a specific address from the chunk allocator. It's currently only implemented for the chunk recycling path, although in theory it could also be done by optimistically allocating new chunks. On Linux, it could attempt an in-place mremap. However, that won't work in practice since the heap is grown downwards and memory is not unmapped (in a normal build, at least). Repeated vector reallocation micro-benchmark: #include <string.h> #include <stdlib.h> int main(void) { for (size_t i = 0; i < 100; i++) { void *ptr = NULL; size_t old_size = 0; for (size_t size = 4; size < (1 << 30); size *= 2) { ptr = realloc(ptr, size); if (!ptr) return 1; memset(ptr + old_size, 0xff, size - old_size); old_size = size; } free(ptr); } } The glibc allocator fails to do any in-place reallocations on this benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it elides the cost of copies via mremap, which is currently not something that jemalloc can use. With this improvement, jemalloc still fails to do any in-place huge reallocations for the first outer loop, but then succeeds 100% of the time for the remaining 99 iterations. The time spent doing allocations and copies drops down to under 5%, with nearly all of it spent doing purging + faulting (when huge pages are disabled) and the array memset. An improved mremap API (MREMAP_RETAIN - #138) would be far more general but this is a portable optimization and would still be useful on Linux for xallocx. Numbers with transparent huge pages enabled: glibc (copies elided via MREMAP_MAYMOVE): 8.471s jemalloc: 17.816s jemalloc + no-op madvise: 13.236s jemalloc + this commit: 6.787s jemalloc + this commit + no-op madvise: 6.144s Numbers with transparent huge pages disabled: glibc (copies elided via MREMAP_MAYMOVE): 15.403s jemalloc: 39.456s jemalloc + no-op madvise: 12.768s jemalloc + this commit: 15.534s jemalloc + this commit + no-op madvise: 6.354s Closes #137
2014-10-04 13:39:32 +08:00
}
return (false);
}
bool
Attempt to expand huge allocations in-place. This adds support for expanding huge allocations in-place by requesting memory at a specific address from the chunk allocator. It's currently only implemented for the chunk recycling path, although in theory it could also be done by optimistically allocating new chunks. On Linux, it could attempt an in-place mremap. However, that won't work in practice since the heap is grown downwards and memory is not unmapped (in a normal build, at least). Repeated vector reallocation micro-benchmark: #include <string.h> #include <stdlib.h> int main(void) { for (size_t i = 0; i < 100; i++) { void *ptr = NULL; size_t old_size = 0; for (size_t size = 4; size < (1 << 30); size *= 2) { ptr = realloc(ptr, size); if (!ptr) return 1; memset(ptr + old_size, 0xff, size - old_size); old_size = size; } free(ptr); } } The glibc allocator fails to do any in-place reallocations on this benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it elides the cost of copies via mremap, which is currently not something that jemalloc can use. With this improvement, jemalloc still fails to do any in-place huge reallocations for the first outer loop, but then succeeds 100% of the time for the remaining 99 iterations. The time spent doing allocations and copies drops down to under 5%, with nearly all of it spent doing purging + faulting (when huge pages are disabled) and the array memset. An improved mremap API (MREMAP_RETAIN - #138) would be far more general but this is a portable optimization and would still be useful on Linux for xallocx. Numbers with transparent huge pages enabled: glibc (copies elided via MREMAP_MAYMOVE): 8.471s jemalloc: 17.816s jemalloc + no-op madvise: 13.236s jemalloc + this commit: 6.787s jemalloc + this commit + no-op madvise: 6.144s Numbers with transparent huge pages disabled: glibc (copies elided via MREMAP_MAYMOVE): 15.403s jemalloc: 39.456s jemalloc + no-op madvise: 12.768s jemalloc + this commit: 15.534s jemalloc + this commit + no-op madvise: 6.354s Closes #137
2014-10-04 13:39:32 +08:00
huge_ralloc_no_move(void *ptr, size_t oldsize, size_t size, size_t extra,
bool zero)
{
size_t usize;
/* Both allocations must be huge to avoid a move. */
if (oldsize < chunksize)
return (true);
assert(s2u(oldsize) == oldsize);
usize = s2u(size);
if (usize == 0) {
/* size_t overflow. */
return (true);
}
/*
* Avoid moving the allocation if the existing chunk size accommodates
* the new size.
*/
if (CHUNK_CEILING(oldsize) >= CHUNK_CEILING(usize)
&& CHUNK_CEILING(oldsize) <= CHUNK_CEILING(size+extra)) {
size_t usize_next;
/* Increase usize to incorporate extra. */
while (usize < s2u(size+extra) && (usize_next = s2u(usize+1)) <
oldsize)
usize = usize_next;
/* Update the size of the huge allocation if it changed. */
if (oldsize != usize) {
extent_node_t *node, key;
malloc_mutex_lock(&huge_mtx);
key.addr = ptr;
node = extent_tree_ad_search(&huge, &key);
assert(node != NULL);
assert(node->addr == ptr);
assert(node->size != usize);
node->size = usize;
malloc_mutex_unlock(&huge_mtx);
if (oldsize < usize) {
if (zero || (config_fill &&
unlikely(opt_zero))) {
memset(ptr + oldsize, 0, usize -
oldsize);
} else if (config_fill && unlikely(opt_junk)) {
memset(ptr + oldsize, 0xa5, usize -
oldsize);
}
} else if (config_fill && unlikely(opt_junk) && oldsize
> usize)
memset(ptr + usize, 0x5a, oldsize - usize);
}
return (false);
}
if (CHUNK_CEILING(oldsize) >= CHUNK_CEILING(size)
&& CHUNK_CEILING(oldsize) <= CHUNK_CEILING(size+extra)) {
return (false);
}
/* Shrink the allocation in-place. */
if (CHUNK_CEILING(oldsize) > CHUNK_CEILING(usize)) {
extent_node_t *node, key;
void *excess_addr;
size_t excess_size;
malloc_mutex_lock(&huge_mtx);
key.addr = ptr;
node = extent_tree_ad_search(&huge, &key);
assert(node != NULL);
assert(node->addr == ptr);
/* Update the size of the huge allocation. */
node->size = usize;
malloc_mutex_unlock(&huge_mtx);
excess_addr = node->addr + CHUNK_CEILING(usize);
excess_size = CHUNK_CEILING(oldsize) - CHUNK_CEILING(usize);
/* Zap the excess chunks. */
huge_dalloc_junk(ptr + usize, oldsize - usize);
arena_chunk_dalloc_huge(node->arena, excess_addr, excess_size);
return (false);
}
Attempt to expand huge allocations in-place. This adds support for expanding huge allocations in-place by requesting memory at a specific address from the chunk allocator. It's currently only implemented for the chunk recycling path, although in theory it could also be done by optimistically allocating new chunks. On Linux, it could attempt an in-place mremap. However, that won't work in practice since the heap is grown downwards and memory is not unmapped (in a normal build, at least). Repeated vector reallocation micro-benchmark: #include <string.h> #include <stdlib.h> int main(void) { for (size_t i = 0; i < 100; i++) { void *ptr = NULL; size_t old_size = 0; for (size_t size = 4; size < (1 << 30); size *= 2) { ptr = realloc(ptr, size); if (!ptr) return 1; memset(ptr + old_size, 0xff, size - old_size); old_size = size; } free(ptr); } } The glibc allocator fails to do any in-place reallocations on this benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it elides the cost of copies via mremap, which is currently not something that jemalloc can use. With this improvement, jemalloc still fails to do any in-place huge reallocations for the first outer loop, but then succeeds 100% of the time for the remaining 99 iterations. The time spent doing allocations and copies drops down to under 5%, with nearly all of it spent doing purging + faulting (when huge pages are disabled) and the array memset. An improved mremap API (MREMAP_RETAIN - #138) would be far more general but this is a portable optimization and would still be useful on Linux for xallocx. Numbers with transparent huge pages enabled: glibc (copies elided via MREMAP_MAYMOVE): 8.471s jemalloc: 17.816s jemalloc + no-op madvise: 13.236s jemalloc + this commit: 6.787s jemalloc + this commit + no-op madvise: 6.144s Numbers with transparent huge pages disabled: glibc (copies elided via MREMAP_MAYMOVE): 15.403s jemalloc: 39.456s jemalloc + no-op madvise: 12.768s jemalloc + this commit: 15.534s jemalloc + this commit + no-op madvise: 6.354s Closes #137
2014-10-04 13:39:32 +08:00
/* Attempt to expand the allocation in-place. */
if (huge_ralloc_no_move_expand(ptr, oldsize, size + extra, zero)) {
if (extra == 0)
return (true);
/* Try again, this time without extra. */
return (huge_ralloc_no_move_expand(ptr, oldsize, size, zero));
}
return (false);
}
void *
huge_ralloc(tsd_t *tsd, arena_t *arena, void *ptr, size_t oldsize, size_t size,
size_t extra, size_t alignment, bool zero, bool try_tcache_dalloc)
{
void *ret;
size_t copysize;
/* Try to avoid moving the allocation. */
Attempt to expand huge allocations in-place. This adds support for expanding huge allocations in-place by requesting memory at a specific address from the chunk allocator. It's currently only implemented for the chunk recycling path, although in theory it could also be done by optimistically allocating new chunks. On Linux, it could attempt an in-place mremap. However, that won't work in practice since the heap is grown downwards and memory is not unmapped (in a normal build, at least). Repeated vector reallocation micro-benchmark: #include <string.h> #include <stdlib.h> int main(void) { for (size_t i = 0; i < 100; i++) { void *ptr = NULL; size_t old_size = 0; for (size_t size = 4; size < (1 << 30); size *= 2) { ptr = realloc(ptr, size); if (!ptr) return 1; memset(ptr + old_size, 0xff, size - old_size); old_size = size; } free(ptr); } } The glibc allocator fails to do any in-place reallocations on this benchmark once it passes the M_MMAP_THRESHOLD (default 128k) but it elides the cost of copies via mremap, which is currently not something that jemalloc can use. With this improvement, jemalloc still fails to do any in-place huge reallocations for the first outer loop, but then succeeds 100% of the time for the remaining 99 iterations. The time spent doing allocations and copies drops down to under 5%, with nearly all of it spent doing purging + faulting (when huge pages are disabled) and the array memset. An improved mremap API (MREMAP_RETAIN - #138) would be far more general but this is a portable optimization and would still be useful on Linux for xallocx. Numbers with transparent huge pages enabled: glibc (copies elided via MREMAP_MAYMOVE): 8.471s jemalloc: 17.816s jemalloc + no-op madvise: 13.236s jemalloc + this commit: 6.787s jemalloc + this commit + no-op madvise: 6.144s Numbers with transparent huge pages disabled: glibc (copies elided via MREMAP_MAYMOVE): 15.403s jemalloc: 39.456s jemalloc + no-op madvise: 12.768s jemalloc + this commit: 15.534s jemalloc + this commit + no-op madvise: 6.354s Closes #137
2014-10-04 13:39:32 +08:00
if (!huge_ralloc_no_move(ptr, oldsize, size, extra, zero))
return (ptr);
/*
* size and oldsize are different enough that we need to use a
* different size class. In that case, fall back to allocating new
* space and copying.
*/
if (alignment > chunksize)
ret = huge_palloc(tsd, arena, size + extra, alignment, zero);
else
ret = huge_malloc(tsd, arena, size + extra, zero);
if (ret == NULL) {
if (extra == 0)
return (NULL);
/* Try again, this time without extra. */
if (alignment > chunksize)
ret = huge_palloc(tsd, arena, size, alignment, zero);
else
ret = huge_malloc(tsd, arena, size, zero);
if (ret == NULL)
return (NULL);
}
/*
* Copy at most size bytes (not size+extra), since the caller has no
* expectation that the extra bytes will be reliably preserved.
*/
copysize = (size < oldsize) ? size : oldsize;
memcpy(ret, ptr, copysize);
iqalloc(tsd, ptr, try_tcache_dalloc);
return (ret);
}
void
huge_dalloc(void *ptr)
{
extent_node_t *node, key;
malloc_mutex_lock(&huge_mtx);
/* Extract from tree of huge allocations. */
key.addr = ptr;
node = extent_tree_ad_search(&huge, &key);
assert(node != NULL);
assert(node->addr == ptr);
extent_tree_ad_remove(&huge, node);
malloc_mutex_unlock(&huge_mtx);
huge_dalloc_junk(node->addr, node->size);
arena_chunk_dalloc_huge(node->arena, node->addr,
CHUNK_CEILING(node->size));
base_node_dalloc(node);
}
size_t
huge_salloc(const void *ptr)
{
size_t ret;
extent_node_t *node, key;
malloc_mutex_lock(&huge_mtx);
/* Extract from tree of huge allocations. */
key.addr = __DECONST(void *, ptr);
node = extent_tree_ad_search(&huge, &key);
assert(node != NULL);
ret = node->size;
malloc_mutex_unlock(&huge_mtx);
return (ret);
}
prof_tctx_t *
huge_prof_tctx_get(const void *ptr)
{
prof_tctx_t *ret;
extent_node_t *node, key;
malloc_mutex_lock(&huge_mtx);
/* Extract from tree of huge allocations. */
key.addr = __DECONST(void *, ptr);
node = extent_tree_ad_search(&huge, &key);
assert(node != NULL);
ret = node->prof_tctx;
malloc_mutex_unlock(&huge_mtx);
return (ret);
}
void
huge_prof_tctx_set(const void *ptr, prof_tctx_t *tctx)
{
extent_node_t *node, key;
malloc_mutex_lock(&huge_mtx);
/* Extract from tree of huge allocations. */
key.addr = __DECONST(void *, ptr);
node = extent_tree_ad_search(&huge, &key);
assert(node != NULL);
node->prof_tctx = tctx;
malloc_mutex_unlock(&huge_mtx);
}
bool
huge_boot(void)
{
/* Initialize chunks data. */
if (malloc_mutex_init(&huge_mtx))
return (true);
extent_tree_ad_new(&huge);
return (false);
}
void
huge_prefork(void)
{
malloc_mutex_prefork(&huge_mtx);
}
void
huge_postfork_parent(void)
{
malloc_mutex_postfork_parent(&huge_mtx);
}
void
huge_postfork_child(void)
{
malloc_mutex_postfork_child(&huge_mtx);
}