mirror of
https://github.com/microsoft/mimalloc.git
synced 2024-12-28 22:05:40 +08:00
initial numa support for arenas
This commit is contained in:
daan
parent
aaf01620f4
commit
a6499be074
@ -56,6 +56,7 @@ void _mi_os_init(void); // called fro
|
||||
void* _mi_os_alloc(size_t size, mi_stats_t* stats); // to allocate thread local data
|
||||
void _mi_os_free(void* p, size_t size, mi_stats_t* stats); // to free thread local data
|
||||
size_t _mi_os_good_alloc_size(size_t size);
|
||||
int _mi_os_numa_node(void);
|
||||
|
||||
|
||||
// memory.c
|
||||
|
||||
@ -228,9 +228,14 @@ mi_decl_export bool mi_heap_visit_blocks(const mi_heap_t* heap, bool visit_all_b
|
||||
|
||||
// Experimental
|
||||
mi_decl_export bool mi_is_in_heap_region(const void* p) mi_attr_noexcept;
|
||||
mi_decl_export int mi_reserve_huge_os_pages(size_t pages, double max_secs, size_t* pages_reserved) mi_attr_noexcept;
|
||||
mi_decl_export bool mi_is_redirected() mi_attr_noexcept;
|
||||
|
||||
mi_decl_export int mi_reserve_huge_os_pages_interleave(size_t pages) mi_attr_noexcept;
|
||||
mi_decl_export int mi_reserve_huge_os_pages_at(size_t pages, int numa_node) mi_attr_noexcept;
|
||||
|
||||
// deprecated
|
||||
mi_decl_export int mi_reserve_huge_os_pages(size_t pages, double max_secs, size_t* pages_reserved) mi_attr_noexcept;
|
||||
|
||||
// ------------------------------------------------------
|
||||
// Convenience
|
||||
// ------------------------------------------------------
|
||||
@ -271,6 +276,7 @@ typedef enum mi_option_e {
|
||||
mi_option_eager_commit_delay,
|
||||
mi_option_segment_reset,
|
||||
mi_option_os_tag,
|
||||
mi_option_max_numa_node,
|
||||
_mi_option_last
|
||||
} mi_option_t;
|
||||
|
||||
|
||||
122
src/arena.c
122
src/arena.c
@ -25,8 +25,10 @@ with on-demand coalescing.
|
||||
|
||||
// os.c
|
||||
void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool* large, mi_os_tld_t* tld);
|
||||
int _mi_os_alloc_huge_os_pages(size_t pages, double max_secs, void** pstart, size_t* pages_reserved, size_t* psize) mi_attr_noexcept;
|
||||
//int _mi_os_alloc_huge_os_pages(size_t pages, double max_secs, void** pstart, size_t* pages_reserved, size_t* psize) mi_attr_noexcept;
|
||||
void _mi_os_free(void* p, size_t size, mi_stats_t* stats);
|
||||
void* _mi_os_alloc_huge_os_pages(size_t pages, int numa_node, size_t* psize);
|
||||
int _mi_os_numa_node_count(void);
|
||||
|
||||
/* -----------------------------------------------------------
|
||||
Arena allocation
|
||||
@ -44,6 +46,7 @@ typedef uintptr_t mi_block_info_t;
|
||||
typedef struct mi_arena_s {
|
||||
uint8_t* start; // the start of the memory area
|
||||
size_t block_count; // size of the area in arena blocks (of `MI_ARENA_BLOCK_SIZE`)
|
||||
int numa_node; // associated NUMA node
|
||||
bool is_zero_init; // is the arena zero initialized?
|
||||
bool is_large; // large OS page allocated
|
||||
_Atomic(uintptr_t) block_bottom; // optimization to start the search for free blocks
|
||||
@ -223,7 +226,31 @@ static void* mi_arena_alloc(mi_arena_t* arena, size_t needed_bcount, bool* is_ze
|
||||
Arena Allocation
|
||||
----------------------------------------------------------- */
|
||||
|
||||
void* _mi_arena_alloc_aligned(size_t size, size_t alignment, bool* commit, bool* large, bool* is_zero, size_t* memid, mi_os_tld_t* tld) {
|
||||
static void* mi_arena_alloc_from(mi_arena_t* arena, size_t arena_index, size_t needed_bcount,
|
||||
bool* commit, bool* large, bool* is_zero,
|
||||
size_t* memid)
|
||||
{
|
||||
size_t block_index = SIZE_MAX;
|
||||
void* p = mi_arena_alloc(arena, needed_bcount, is_zero, &block_index);
|
||||
if (p != NULL) {
|
||||
mi_assert_internal(block_index != SIZE_MAX);
|
||||
#if MI_DEBUG>=1
|
||||
_Atomic(mi_block_info_t)* block = &arena->blocks[block_index];
|
||||
mi_block_info_t binfo = mi_atomic_read(block);
|
||||
mi_assert_internal(mi_block_is_in_use(binfo));
|
||||
mi_assert_internal(mi_block_count(binfo) >= needed_bcount);
|
||||
#endif
|
||||
*memid = mi_memid_create(arena_index, block_index);
|
||||
*commit = true; // TODO: support commit on demand?
|
||||
*large = arena->is_large;
|
||||
}
|
||||
return p;
|
||||
}
|
||||
|
||||
void* _mi_arena_alloc_aligned(size_t size, size_t alignment,
|
||||
bool* commit, bool* large, bool* is_zero,
|
||||
size_t* memid, mi_os_tld_t* tld)
|
||||
{
|
||||
mi_assert_internal(memid != NULL && tld != NULL);
|
||||
mi_assert_internal(size > 0);
|
||||
*memid = MI_MEMID_OS;
|
||||
@ -240,33 +267,36 @@ void* _mi_arena_alloc_aligned(size_t size, size_t alignment, bool* commit, bool*
|
||||
{
|
||||
size_t asize = _mi_align_up(size, MI_ARENA_BLOCK_SIZE);
|
||||
size_t bcount = asize / MI_ARENA_BLOCK_SIZE;
|
||||
int numa_node = _mi_os_numa_node(); // current numa node
|
||||
|
||||
mi_assert_internal(size <= bcount*MI_ARENA_BLOCK_SIZE);
|
||||
// try numa affine allocation
|
||||
for (size_t i = 0; i < MI_MAX_ARENAS; i++) {
|
||||
mi_arena_t* arena = (mi_arena_t*)mi_atomic_read_ptr_relaxed(mi_atomic_cast(void*, &mi_arenas[i]));
|
||||
if (arena==NULL) break;
|
||||
if (*large || !arena->is_large) { // large OS pages allowed, or arena is not large OS pages
|
||||
size_t block_index = SIZE_MAX;
|
||||
void* p = mi_arena_alloc(arena, bcount, is_zero, &block_index);
|
||||
if (p != NULL) {
|
||||
mi_assert_internal(block_index != SIZE_MAX);
|
||||
#if MI_DEBUG>=1
|
||||
_Atomic(mi_block_info_t)* block = &arena->blocks[block_index];
|
||||
mi_block_info_t binfo = mi_atomic_read(block);
|
||||
mi_assert_internal(mi_block_is_in_use(binfo));
|
||||
mi_assert_internal(mi_block_count(binfo)*MI_ARENA_BLOCK_SIZE >= size);
|
||||
#endif
|
||||
*memid = mi_memid_create(i, block_index);
|
||||
*commit = true; // TODO: support commit on demand?
|
||||
*large = arena->is_large;
|
||||
if (arena==NULL) break; // end reached
|
||||
if ((arena->numa_node<0 || arena->numa_node==numa_node) && // numa local?
|
||||
(*large || !arena->is_large)) // large OS pages allowed, or arena is not large OS pages
|
||||
{
|
||||
void* p = mi_arena_alloc_from(arena, i, bcount, commit, large, is_zero, memid);
|
||||
mi_assert_internal((uintptr_t)p % alignment == 0);
|
||||
return p;
|
||||
if (p != NULL) return p;
|
||||
}
|
||||
}
|
||||
// try from another numa node instead..
|
||||
for (size_t i = 0; i < MI_MAX_ARENAS; i++) {
|
||||
mi_arena_t* arena = (mi_arena_t*)mi_atomic_read_ptr_relaxed(mi_atomic_cast(void*, &mi_arenas[i]));
|
||||
if (arena==NULL) break; // end reached
|
||||
if ((arena->numa_node>=0 && arena->numa_node!=numa_node) && // not numa local!
|
||||
(*large || !arena->is_large)) // large OS pages allowed, or arena is not large OS pages
|
||||
{
|
||||
void* p = mi_arena_alloc_from(arena, i, bcount, commit, large, is_zero, memid);
|
||||
mi_assert_internal((uintptr_t)p % alignment == 0);
|
||||
if (p != NULL) return p;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// fall back to the OS
|
||||
// finally, fall back to the OS
|
||||
*is_zero = true;
|
||||
*memid = MI_MEMID_OS;
|
||||
return _mi_os_alloc_aligned(size, alignment, *commit, large, tld);
|
||||
@ -350,31 +380,61 @@ static bool mi_arena_add(mi_arena_t* arena) {
|
||||
----------------------------------------------------------- */
|
||||
#include <errno.h> // ENOMEM
|
||||
|
||||
int mi_reserve_huge_os_pages(size_t pages, double max_secs, size_t* pages_reserved) mi_attr_noexcept {
|
||||
size_t pages_reserved_default = 0;
|
||||
if (pages_reserved==NULL) pages_reserved = &pages_reserved_default;
|
||||
// reserve at a specific numa node
|
||||
static int mi_reserve_huge_os_pages_at(size_t pages, int numa_node) mi_attr_noexcept {
|
||||
size_t hsize = 0;
|
||||
void* p = NULL;
|
||||
int err = _mi_os_alloc_huge_os_pages(pages, max_secs, &p, pages_reserved, &hsize);
|
||||
_mi_verbose_message("reserved %zu huge pages\n", *pages_reserved);
|
||||
if (p==NULL) return err;
|
||||
// err might be != 0 but that is fine, we just got less pages.
|
||||
mi_assert_internal(*pages_reserved > 0 && hsize > 0 && *pages_reserved <= pages);
|
||||
void* p = _mi_os_alloc_huge_os_pages(pages, numa_node, &hsize);
|
||||
if (p==NULL) return ENOMEM;
|
||||
_mi_verbose_message("reserved %zu huge (1GiB) pages\n", pages);
|
||||
|
||||
size_t bcount = hsize / MI_ARENA_BLOCK_SIZE;
|
||||
size_t asize = sizeof(mi_arena_t) + (bcount*sizeof(mi_block_info_t)); // one too much
|
||||
mi_arena_t* arena = (mi_arena_t*)_mi_os_alloc(asize, &_mi_stats_main);
|
||||
mi_arena_t* arena = (mi_arena_t*)_mi_os_alloc(asize, &_mi_stats_main); // TODO: can we avoid allocating from the OS?
|
||||
if (arena == NULL) {
|
||||
*pages_reserved = 0;
|
||||
_mi_os_free(p, hsize, &_mi_stats_main);
|
||||
return ENOMEM;
|
||||
}
|
||||
arena->block_count = bcount;
|
||||
arena->start = (uint8_t*)p;
|
||||
arena->block_bottom = 0;
|
||||
arena->numa_node = numa_node; // TODO: or get the current numa node if -1? (now it allows anyone to allocate on -1)
|
||||
arena->is_large = true;
|
||||
arena->is_zero_init = true;
|
||||
memset(arena->blocks, 0, bcount * sizeof(mi_block_info_t));
|
||||
//mi_atomic_write(&arena->blocks[0], mi_block_info_create(bcount, false));
|
||||
mi_arena_add(arena);
|
||||
return 0;
|
||||
}
|
||||
|
||||
|
||||
// reserve huge pages evenly among all numa nodes.
|
||||
int mi_reserve_huge_os_pages_interleave(size_t pages) mi_attr_noexcept {
|
||||
if (pages == 0) return 0;
|
||||
|
||||
// pages per numa node
|
||||
int numa_count = _mi_os_numa_node_count();
|
||||
if (numa_count <= 0) numa_count = 1;
|
||||
size_t pages_per = pages / numa_count;
|
||||
if (pages_per == 0) pages_per = 1;
|
||||
|
||||
// reserve evenly among numa nodes
|
||||
for (int numa_node = 0; numa_node < numa_count && pages > 0; numa_node++) {
|
||||
int err = mi_reserve_huge_os_pages_at((pages_per > pages ? pages : pages_per), numa_node);
|
||||
if (err) return err;
|
||||
if (pages < pages_per) {
|
||||
pages = 0;
|
||||
}
|
||||
else {
|
||||
pages -= pages_per;
|
||||
}
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
int mi_reserve_huge_os_pages(size_t pages, double max_secs, size_t* pages_reserved) mi_attr_noexcept {
|
||||
_mi_verbose_message("mi_reserve_huge_os_pages is deprecated: use mi_reserve_huge_os_pages_interleave/at instead\n");
|
||||
if (pages_reserved != NULL) *pages_reserved = 0;
|
||||
int err = mi_reserve_huge_os_pages_interleave(pages);
|
||||
if (err==0 && pages_reserved!=NULL) *pages_reserved = pages;
|
||||
return err;
|
||||
}
|
||||
|
||||
@ -435,7 +435,7 @@ static void mi_process_load(void) {
|
||||
if (mi_option_is_enabled(mi_option_reserve_huge_os_pages)) {
|
||||
size_t pages = mi_option_get(mi_option_reserve_huge_os_pages);
|
||||
double max_secs = (double)pages / 2.0; // 0.5s per page (1GiB)
|
||||
mi_reserve_huge_os_pages(pages, max_secs, NULL);
|
||||
mi_reserve_huge_os_pages_interleave(pages);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@ -66,7 +66,8 @@ static mi_option_desc_t options[_mi_option_last] =
|
||||
{ 0, UNINIT, MI_OPTION(reset_decommits) }, // note: cannot enable this if secure is on
|
||||
{ 0, UNINIT, MI_OPTION(eager_commit_delay) }, // the first N segments per thread are not eagerly committed
|
||||
{ 0, UNINIT, MI_OPTION(segment_reset) }, // reset segment memory on free (needs eager commit)
|
||||
{ 100, UNINIT, MI_OPTION(os_tag) } // only apple specific for now but might serve more or less related purpose
|
||||
{ 100, UNINIT, MI_OPTION(os_tag) }, // only apple specific for now but might serve more or less related purpose
|
||||
{ 256, UNINIT, MI_OPTION(max_numa_node) } // maximum allowed numa node
|
||||
};
|
||||
|
||||
static void mi_option_init(mi_option_desc_t* desc);
|
||||
|
||||
236
src/os.c
236
src/os.c
@ -170,7 +170,7 @@ void _mi_os_init() {
|
||||
os_alloc_granularity = os_page_size;
|
||||
}
|
||||
if (mi_option_is_enabled(mi_option_large_os_pages)) {
|
||||
large_os_page_size = (1UL << 21); // 2MiB
|
||||
large_os_page_size = 2*MiB;
|
||||
}
|
||||
}
|
||||
#endif
|
||||
@ -207,31 +207,6 @@ static void* mi_os_get_aligned_hint(size_t try_alignment, size_t size);
|
||||
|
||||
#ifdef _WIN32
|
||||
static void* mi_win_virtual_allocx(void* addr, size_t size, size_t try_alignment, DWORD flags) {
|
||||
#if defined(MEM_EXTENDED_PARAMETER_TYPE_BITS)
|
||||
// on modern Windows try use NtAllocateVirtualMemoryEx for 1GiB huge pages
|
||||
if ((size % ((uintptr_t)1 << 30)) == 0 /* 1GiB multiple */
|
||||
&& (flags & MEM_LARGE_PAGES) != 0 && (flags & MEM_COMMIT) != 0 && (flags & MEM_RESERVE) != 0
|
||||
&& (addr != NULL || try_alignment == 0 || try_alignment % _mi_os_page_size() == 0)
|
||||
&& pNtAllocateVirtualMemoryEx != NULL)
|
||||
{
|
||||
#ifndef MEM_EXTENDED_PARAMETER_NONPAGED_HUGE
|
||||
#define MEM_EXTENDED_PARAMETER_NONPAGED_HUGE (0x10)
|
||||
#endif
|
||||
MEM_EXTENDED_PARAMETER param = { 0, 0 };
|
||||
param.Type = 5; // == MemExtendedParameterAttributeFlags;
|
||||
param.ULong64 = MEM_EXTENDED_PARAMETER_NONPAGED_HUGE;
|
||||
SIZE_T psize = size;
|
||||
void* base = addr;
|
||||
NTSTATUS err = (*pNtAllocateVirtualMemoryEx)(GetCurrentProcess(), &base, &psize, flags, PAGE_READWRITE, ¶m, 1);
|
||||
if (err == 0) {
|
||||
return base;
|
||||
}
|
||||
else {
|
||||
// else fall back to regular large OS pages
|
||||
_mi_warning_message("unable to allocate huge (1GiB) page, trying large (2MiB) pages instead (error 0x%lx)\n", err);
|
||||
}
|
||||
}
|
||||
#endif
|
||||
#if (MI_INTPTR_SIZE >= 8)
|
||||
// on 64-bit systems, try to use the virtual address area after 4TiB for 4MiB aligned allocations
|
||||
void* hint;
|
||||
@ -364,7 +339,7 @@ static void* mi_unix_mmap(void* addr, size_t size, size_t try_alignment, int pro
|
||||
lflags |= MAP_HUGETLB;
|
||||
#endif
|
||||
#ifdef MAP_HUGE_1GB
|
||||
if ((size % ((uintptr_t)1 << 30)) == 0) {
|
||||
if ((size % GiB) == 0) {
|
||||
lflags |= MAP_HUGE_1GB;
|
||||
}
|
||||
else
|
||||
@ -400,10 +375,10 @@ static void* mi_unix_mmap(void* addr, size_t size, size_t try_alignment, int pro
|
||||
p = mi_unix_mmapx(addr, size, try_alignment, protect_flags, flags, fd);
|
||||
#if defined(MADV_HUGEPAGE)
|
||||
// Many Linux systems don't allow MAP_HUGETLB but they support instead
|
||||
// transparent huge pages (TPH). It is not required to call `madvise` with MADV_HUGE
|
||||
// transparent huge pages (THP). It is not required to call `madvise` with MADV_HUGE
|
||||
// though since properly aligned allocations will already use large pages if available
|
||||
// in that case -- in particular for our large regions (in `memory.c`).
|
||||
// However, some systems only allow TPH if called with explicit `madvise`, so
|
||||
// However, some systems only allow THP if called with explicit `madvise`, so
|
||||
// when large OS pages are enabled for mimalloc, we call `madvice` anyways.
|
||||
if (allow_large && use_large_os_page(size, try_alignment)) {
|
||||
if (madvise(p, size, MADV_HUGEPAGE) == 0) {
|
||||
@ -810,101 +785,146 @@ bool _mi_os_shrink(void* p, size_t oldsize, size_t newsize, mi_stats_t* stats) {
|
||||
|
||||
|
||||
/* ----------------------------------------------------------------------------
|
||||
Support for huge OS pages (1Gib) that are reserved up-front and never
|
||||
released. Only regions are allocated in here (see `memory.c`) so the memory
|
||||
will be reused.
|
||||
Support for allocating huge OS pages (1Gib) that are reserved up-front
|
||||
and possibly associated with a specific NUMA node. (use `numa_node>=0`)
|
||||
-----------------------------------------------------------------------------*/
|
||||
#define MI_HUGE_OS_PAGE_SIZE ((size_t)1 << 30) // 1GiB
|
||||
#define MI_HUGE_OS_PAGE_SIZE (GiB)
|
||||
|
||||
|
||||
#if !(MI_INTPTR_SIZE >= 8 && (defined(_WIN32) || defined(MI_OS_USE_MMAP)))
|
||||
int _mi_os_alloc_huge_os_pages(size_t pages, double max_secs, void** start, size_t* pages_reserved, size_t* size) mi_attr_noexcept {
|
||||
UNUSED(pages); UNUSED(max_secs);
|
||||
if (start != NULL) *start = NULL;
|
||||
if (pages_reserved != NULL) *pages_reserved = 0;
|
||||
if (size != NULL) *size = 0;
|
||||
return ENOMEM;
|
||||
}
|
||||
#else
|
||||
static _Atomic(uintptr_t) huge_top; // = 0
|
||||
|
||||
int _mi_os_alloc_huge_os_pages(size_t pages, double max_secs, void** pstart, size_t* pages_reserved, size_t* psize) mi_attr_noexcept
|
||||
#if defined(WIN32) && (MI_INTPTR_SIZE >= 8)
|
||||
static void* mi_os_alloc_huge_os_pagesx(size_t size, int numa_node)
|
||||
{
|
||||
*pstart = NULL;
|
||||
*pages_reserved = 0;
|
||||
*psize = 0;
|
||||
if (max_secs==0) return ETIMEDOUT; // timeout
|
||||
if (pages==0) return 0; // ok
|
||||
mi_assert_internal(size%GiB == 0);
|
||||
|
||||
// Atomically claim a huge address range
|
||||
size_t size = pages * MI_HUGE_OS_PAGE_SIZE;
|
||||
uint8_t* start;
|
||||
do {
|
||||
start = (uint8_t*)mi_atomic_addu(&huge_top, size);
|
||||
if (start == NULL) {
|
||||
uintptr_t top = ((uintptr_t)32 << 40); // 32TiB virtual start address
|
||||
#if (MI_SECURE>0 || MI_DEBUG==0) // security: randomize start of huge pages unless in debug mode
|
||||
uintptr_t r = _mi_random_init((uintptr_t)&_mi_os_alloc_huge_os_pages);
|
||||
top += ((uintptr_t)MI_HUGE_OS_PAGE_SIZE * ((r>>17) & 0x3FF)); // (randomly 0-1024)*1GiB == 0 to 1TiB
|
||||
#if defined(MEM_EXTENDED_PARAMETER_TYPE_BITS)
|
||||
DWORD flags = MEM_LARGE_PAGES | MEM_COMMIT | MEM_RESERVE;
|
||||
MEM_EXTENDED_PARAMETER params[4] = { {0,0},{0,0},{0,0},{0,0} };
|
||||
MEM_ADDRESS_REQUIREMENTS reqs = {0,0,0};
|
||||
reqs.HighestEndingAddress = NULL;
|
||||
reqs.LowestStartingAddress = NULL;
|
||||
reqs.Alignment = MI_SEGMENT_SIZE;
|
||||
|
||||
// on modern Windows try use NtAllocateVirtualMemoryEx for 1GiB huge pages
|
||||
if (pNtAllocateVirtualMemoryEx != NULL) {
|
||||
#ifndef MEM_EXTENDED_PARAMETER_NONPAGED_HUGE
|
||||
#define MEM_EXTENDED_PARAMETER_NONPAGED_HUGE (0x10)
|
||||
#endif
|
||||
mi_atomic_cas_strong(&huge_top, top, 0);
|
||||
params[0].Type = MemExtendedParameterAddressRequirements;
|
||||
params[0].Pointer = &reqs;
|
||||
params[1].Type = 5; // == MemExtendedParameterAttributeFlags;
|
||||
params[1].ULong64 = MEM_EXTENDED_PARAMETER_NONPAGED_HUGE;
|
||||
size_t param_count = 2;
|
||||
if (numa_node >= 0) {
|
||||
param_count++;
|
||||
params[2].Type = MemExtendedParameterNumaNode;
|
||||
params[2].ULong = (unsigned)numa_node;
|
||||
}
|
||||
} while (start == NULL);
|
||||
|
||||
|
||||
// Allocate one page at the time but try to place them contiguously
|
||||
// We allocate one page at the time to be able to abort if it takes too long
|
||||
double start_t = _mi_clock_start();
|
||||
uint8_t* addr = start; // current top of the allocations
|
||||
for (size_t page = 0; page < pages; page++, addr += MI_HUGE_OS_PAGE_SIZE ) {
|
||||
// allocate a page
|
||||
void* p = NULL;
|
||||
bool is_large = true;
|
||||
#ifdef _WIN32
|
||||
if (page==0) { mi_win_enable_large_os_pages(); }
|
||||
p = mi_win_virtual_alloc(addr, MI_HUGE_OS_PAGE_SIZE, 0, MEM_LARGE_PAGES | MEM_COMMIT | MEM_RESERVE, true, true, &is_large);
|
||||
#elif defined(MI_OS_USE_MMAP)
|
||||
p = mi_unix_mmap(addr, MI_HUGE_OS_PAGE_SIZE, 0, PROT_READ | PROT_WRITE, true, true, &is_large);
|
||||
#else
|
||||
// always fail
|
||||
#endif
|
||||
|
||||
// Did we succeed at a contiguous address?
|
||||
if (p != addr) {
|
||||
// no success, issue a warning and return with an error
|
||||
if (p != NULL) {
|
||||
_mi_warning_message("could not allocate contiguous huge page %zu at 0x%p\n", page, addr);
|
||||
_mi_os_free(p, MI_HUGE_OS_PAGE_SIZE, &_mi_stats_main );
|
||||
SIZE_T psize = size;
|
||||
void* base = NULL;
|
||||
NTSTATUS err = (*pNtAllocateVirtualMemoryEx)(GetCurrentProcess(), &base, &psize, flags, PAGE_READWRITE, params, param_count);
|
||||
if (err == 0) {
|
||||
return base;
|
||||
}
|
||||
else {
|
||||
#ifdef _WIN32
|
||||
int err = GetLastError();
|
||||
#else
|
||||
int err = errno;
|
||||
// fall back to regular huge pages
|
||||
_mi_warning_message("unable to allocate using huge (1GiB) pages, trying large (2MiB) pages instead (error 0x%lx)\n", err);
|
||||
}
|
||||
}
|
||||
// on modern Windows try use VirtualAlloc2 for aligned large OS page allocation
|
||||
if (pVirtualAlloc2 != NULL) {
|
||||
params[0].Type = MemExtendedParameterAddressRequirements;
|
||||
params[0].Pointer = &reqs;
|
||||
size_t param_count = 1;
|
||||
if (numa_node >= 0) {
|
||||
param_count++;
|
||||
params[1].Type = MemExtendedParameterNumaNode;
|
||||
params[1].ULong = (unsigned)numa_node;
|
||||
}
|
||||
return (*pVirtualAlloc2)(GetCurrentProcess(), NULL, size, flags, PAGE_READWRITE, params, param_count);
|
||||
}
|
||||
#endif
|
||||
_mi_warning_message("could not allocate huge page %zu at 0x%p, error: %i\n", page, addr, err);
|
||||
return NULL; // give up on older Windows..
|
||||
}
|
||||
return ENOMEM;
|
||||
#elif defined(MI_OS_USE_MMAP) && (MI_INTPTR_SIZE >= 8)
|
||||
#ifdef MI_HAS_NUMA
|
||||
#include <numaif.h> // mbind, and use -lnuma
|
||||
#endif
|
||||
static void* mi_os_alloc_huge_os_pagesx(size_t size, int numa_node) {
|
||||
mi_assert_internal(size%GiB == 0);
|
||||
bool is_large = true;
|
||||
void* p = mi_unix_mmap(NULL, MI_HUGE_OS_PAGE_SIZE, MI_SEGMENT_SIZE, PROT_READ | PROT_WRITE, true, true, &is_large);
|
||||
if (p == NULL) return NULL;
|
||||
#ifdef MI_HAS_NUMA
|
||||
if (numa_node >= 0 && numa_node < 8*MI_INTPTR_SIZE) {
|
||||
uintptr_t numa_mask = (1UL << numa_node);
|
||||
long err = mbind(p, size, MPOL_PREFERRED, &numa_mask, 8*MI_INTPTR_SIZE, 0);
|
||||
if (err != 0) {
|
||||
_mi_warning_message("failed to bind huge (1GiB) pages to NUMA node %d: %s\n", numa_node, strerror(errno));
|
||||
}
|
||||
// success, record it
|
||||
if (page==0) {
|
||||
*pstart = addr;
|
||||
}
|
||||
*psize += MI_HUGE_OS_PAGE_SIZE;
|
||||
*pages_reserved += 1;
|
||||
_mi_stat_increase(&_mi_stats_main.committed, MI_HUGE_OS_PAGE_SIZE);
|
||||
_mi_stat_increase(&_mi_stats_main.reserved, MI_HUGE_OS_PAGE_SIZE);
|
||||
#endif
|
||||
return p;
|
||||
}
|
||||
#else
|
||||
static void* mi_os_alloc_huge_os_pagesx(size_t size, int numa_node) {
|
||||
return NULL;
|
||||
}
|
||||
#endif
|
||||
|
||||
// check for timeout
|
||||
double elapsed = _mi_clock_end(start_t);
|
||||
if (elapsed > max_secs) return ETIMEDOUT;
|
||||
if (page >= 1) {
|
||||
double estimate = ((elapsed / (double)(page+1)) * (double)pages);
|
||||
if (estimate > 1.5*max_secs) return ETIMEDOUT; // seems like we are going to timeout
|
||||
void* _mi_os_alloc_huge_os_pages(size_t pages, int numa_node, size_t* psize) {
|
||||
if (psize != NULL) *psize = 0;
|
||||
size_t size = pages * MI_HUGE_OS_PAGE_SIZE;
|
||||
void* p = mi_os_alloc_huge_os_pagesx(size, numa_node);
|
||||
if (p==NULL) return NULL;
|
||||
if (psize != NULL) *psize = size;
|
||||
_mi_stat_increase(&_mi_stats_main.committed, size);
|
||||
_mi_stat_increase(&_mi_stats_main.reserved, size);
|
||||
return p;
|
||||
}
|
||||
|
||||
#ifdef WIN32
|
||||
static int mi_os_numa_nodex(void) {
|
||||
PROCESSOR_NUMBER pnum;
|
||||
USHORT numa_node = 0;
|
||||
GetCurrentProcessorNumberEx(&pnum);
|
||||
GetNumaProcessorNodeEx(&pnum,&numa_node);
|
||||
return (int)numa_node;
|
||||
}
|
||||
mi_assert_internal(*psize == size);
|
||||
|
||||
static int mi_os_numa_node_countx(void) {
|
||||
ULONG numa_max = 0;
|
||||
GetNumaHighestNodeNumber(&numa_max);
|
||||
return (int)(numa_max + 1);
|
||||
}
|
||||
#elif MI_HAS_NUMA
|
||||
#include <numa.h>
|
||||
static int mi_os_numa_nodex(void) {
|
||||
return numa_preferred();
|
||||
}
|
||||
static int mi_os_numa_node_countx(void) {
|
||||
return (numa_max_node() + 1);
|
||||
}
|
||||
#else
|
||||
static int mi_os_numa_nodex(void) {
|
||||
return 0;
|
||||
}
|
||||
static int mi_os_numa_node_countx(void) {
|
||||
return 1;
|
||||
}
|
||||
#endif
|
||||
|
||||
int _mi_os_numa_node_count(void) {
|
||||
long ncount = mi_os_numa_node_countx();
|
||||
// never more than max numa node and at least 1
|
||||
long nmax = 1 + mi_option_get(mi_option_max_numa_node);
|
||||
if (ncount > nmax) ncount = nmax;
|
||||
if (ncount <= 0) ncount = 1;
|
||||
return ncount;
|
||||
}
|
||||
|
||||
int _mi_os_numa_node(void) {
|
||||
int nnode = mi_os_numa_nodex();
|
||||
// never more than the node count
|
||||
int ncount = _mi_os_numa_node_count();
|
||||
if (nnode >= ncount) { nnode = nnode % ncount; }
|
||||
return nnode;
|
||||
}
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user