2019-06-20 02:26:12 +03:00
|
|
|
/* ----------------------------------------------------------------------------
|
|
|
|
Copyright (c) 2018, Microsoft Research, Daan Leijen
|
|
|
|
This is free software; you can redistribute it and/or modify it under the
|
|
|
|
terms of the MIT license. A copy of the license can be found in the file
|
2019-06-23 14:53:34 +03:00
|
|
|
"LICENSE" at the root of this distribution.
|
2019-06-20 02:26:12 +03:00
|
|
|
-----------------------------------------------------------------------------*/
|
|
|
|
#ifndef _DEFAULT_SOURCE
|
|
|
|
#define _DEFAULT_SOURCE // ensure mmap flags are defined
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#include "mimalloc.h"
|
|
|
|
#include "mimalloc-internal.h"
|
2019-07-22 03:08:09 +03:00
|
|
|
#include "mimalloc-atomic.h"
|
2019-06-20 02:26:12 +03:00
|
|
|
|
2019-07-21 18:21:14 +03:00
|
|
|
#include <string.h> // strerror
|
2020-01-18 06:59:55 +03:00
|
|
|
|
2019-06-20 02:26:12 +03:00
|
|
|
|
2019-07-04 00:52:32 +03:00
|
|
|
#if defined(_WIN32)
|
|
|
|
#include <windows.h>
|
2019-07-10 21:23:20 +03:00
|
|
|
#elif defined(__wasi__)
|
|
|
|
// stdlib.h is all we need, and has already been included in mimalloc.h
|
2019-07-04 00:52:32 +03:00
|
|
|
#else
|
|
|
|
#include <sys/mman.h> // mmap
|
|
|
|
#include <unistd.h> // sysconf
|
2019-08-14 17:07:13 +03:00
|
|
|
#if defined(__linux__)
|
|
|
|
#include <linux/mman.h> // linux mmap flags
|
|
|
|
#endif
|
2019-07-05 22:05:04 +03:00
|
|
|
#if defined(__APPLE__)
|
|
|
|
#include <mach/vm_statistics.h>
|
|
|
|
#endif
|
2019-07-04 00:52:32 +03:00
|
|
|
#endif
|
|
|
|
|
2019-06-20 02:26:12 +03:00
|
|
|
/* -----------------------------------------------------------
|
2019-06-29 08:35:57 +03:00
|
|
|
Initialization.
|
2019-07-03 08:49:12 +03:00
|
|
|
On windows initializes support for aligned allocation and
|
2019-06-29 08:35:57 +03:00
|
|
|
large OS pages (if MIMALLOC_LARGE_OS_PAGES is true).
|
2019-06-20 02:26:12 +03:00
|
|
|
----------------------------------------------------------- */
|
2019-08-28 21:58:30 +03:00
|
|
|
bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* stats);
|
2019-08-19 21:10:06 +03:00
|
|
|
|
2019-07-04 00:52:32 +03:00
|
|
|
static void* mi_align_up_ptr(void* p, size_t alignment) {
|
|
|
|
return (void*)_mi_align_up((uintptr_t)p, alignment);
|
|
|
|
}
|
|
|
|
|
|
|
|
static uintptr_t _mi_align_down(uintptr_t sz, size_t alignment) {
|
|
|
|
return (sz / alignment) * alignment;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void* mi_align_down_ptr(void* p, size_t alignment) {
|
|
|
|
return (void*)_mi_align_down((uintptr_t)p, alignment);
|
|
|
|
}
|
2019-06-20 02:26:12 +03:00
|
|
|
|
2019-07-02 17:23:24 +03:00
|
|
|
// page size (initialized properly in `os_init`)
|
|
|
|
static size_t os_page_size = 4096;
|
|
|
|
|
|
|
|
// minimal allocation granularity
|
|
|
|
static size_t os_alloc_granularity = 4096;
|
|
|
|
|
2019-06-29 08:35:57 +03:00
|
|
|
// if non-zero, use large page allocation
|
|
|
|
static size_t large_os_page_size = 0;
|
|
|
|
|
2019-07-02 17:23:24 +03:00
|
|
|
// OS (small) page size
|
|
|
|
size_t _mi_os_page_size() {
|
|
|
|
return os_page_size;
|
|
|
|
}
|
|
|
|
|
|
|
|
// if large OS pages are supported (2 or 4MiB), then return the size, otherwise return the small page size (4KiB)
|
|
|
|
size_t _mi_os_large_page_size() {
|
|
|
|
return (large_os_page_size != 0 ? large_os_page_size : _mi_os_page_size());
|
|
|
|
}
|
|
|
|
|
2019-06-29 08:35:57 +03:00
|
|
|
static bool use_large_os_page(size_t size, size_t alignment) {
|
|
|
|
// if we have access, check the size and alignment requirements
|
2019-08-19 21:17:00 +03:00
|
|
|
if (large_os_page_size == 0 || !mi_option_is_enabled(mi_option_large_os_pages)) return false;
|
2019-06-29 08:35:57 +03:00
|
|
|
return ((size % large_os_page_size) == 0 && (alignment % large_os_page_size) == 0);
|
|
|
|
}
|
|
|
|
|
2019-09-10 23:26:51 +03:00
|
|
|
// round to a good OS allocation size (bounded by max 12.5% waste)
|
|
|
|
size_t _mi_os_good_alloc_size(size_t size) {
|
|
|
|
size_t align_size;
|
2019-11-21 01:55:12 +03:00
|
|
|
if (size < 512*KiB) align_size = _mi_os_page_size();
|
|
|
|
else if (size < 2*MiB) align_size = 64*KiB;
|
|
|
|
else if (size < 8*MiB) align_size = 256*KiB;
|
|
|
|
else if (size < 32*MiB) align_size = 1*MiB;
|
|
|
|
else align_size = 4*MiB;
|
2019-09-10 23:26:51 +03:00
|
|
|
if (size >= (SIZE_MAX - align_size)) return size; // possible overflow?
|
|
|
|
return _mi_align_up(size, align_size);
|
2019-07-02 17:23:24 +03:00
|
|
|
}
|
2019-06-29 08:35:57 +03:00
|
|
|
|
|
|
|
#if defined(_WIN32)
|
|
|
|
// We use VirtualAlloc2 for aligned allocation, but it is only supported on Windows 10 and Windows Server 2016.
|
2019-07-04 00:52:32 +03:00
|
|
|
// So, we need to look it up dynamically to run on older systems. (use __stdcall for 32-bit compatibility)
|
2019-08-20 04:16:12 +03:00
|
|
|
// NtAllocateVirtualAllocEx is used for huge OS page allocation (1GiB)
|
|
|
|
// We hide MEM_EXTENDED_PARAMETER to compile with older SDK's.
|
|
|
|
#include <winternl.h>
|
2019-11-21 01:55:12 +03:00
|
|
|
typedef PVOID (__stdcall *PVirtualAlloc2)(HANDLE, PVOID, SIZE_T, ULONG, ULONG, /* MEM_EXTENDED_PARAMETER* */ void*, ULONG);
|
|
|
|
typedef NTSTATUS (__stdcall *PNtAllocateVirtualMemoryEx)(HANDLE, PVOID*, SIZE_T*, ULONG, ULONG, /* MEM_EXTENDED_PARAMETER* */ PVOID, ULONG);
|
2019-07-13 18:12:16 +03:00
|
|
|
static PVirtualAlloc2 pVirtualAlloc2 = NULL;
|
2019-08-20 04:16:12 +03:00
|
|
|
static PNtAllocateVirtualMemoryEx pNtAllocateVirtualMemoryEx = NULL;
|
2019-06-29 08:35:57 +03:00
|
|
|
|
2019-11-02 08:01:52 +03:00
|
|
|
static bool mi_win_enable_large_os_pages()
|
2019-09-15 01:23:28 +03:00
|
|
|
{
|
|
|
|
if (large_os_page_size > 0) return true;
|
|
|
|
|
|
|
|
// Try to see if large OS pages are supported
|
|
|
|
// To use large pages on Windows, we first need access permission
|
|
|
|
// Set "Lock pages in memory" permission in the group policy editor
|
|
|
|
// <https://devblogs.microsoft.com/oldnewthing/20110128-00/?p=11643>
|
|
|
|
unsigned long err = 0;
|
|
|
|
HANDLE token = NULL;
|
|
|
|
BOOL ok = OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &token);
|
|
|
|
if (ok) {
|
|
|
|
TOKEN_PRIVILEGES tp;
|
|
|
|
ok = LookupPrivilegeValue(NULL, TEXT("SeLockMemoryPrivilege"), &tp.Privileges[0].Luid);
|
|
|
|
if (ok) {
|
|
|
|
tp.PrivilegeCount = 1;
|
|
|
|
tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
|
|
|
|
ok = AdjustTokenPrivileges(token, FALSE, &tp, 0, (PTOKEN_PRIVILEGES)NULL, 0);
|
|
|
|
if (ok) {
|
|
|
|
err = GetLastError();
|
|
|
|
ok = (err == ERROR_SUCCESS);
|
|
|
|
if (ok) {
|
|
|
|
large_os_page_size = GetLargePageMinimum();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
CloseHandle(token);
|
|
|
|
}
|
|
|
|
if (!ok) {
|
|
|
|
if (err == 0) err = GetLastError();
|
|
|
|
_mi_warning_message("cannot enable large OS page support, error %lu\n", err);
|
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
return (ok!=0);
|
2019-09-15 01:23:28 +03:00
|
|
|
}
|
|
|
|
|
2019-06-29 08:35:57 +03:00
|
|
|
void _mi_os_init(void) {
|
2019-07-02 17:23:24 +03:00
|
|
|
// get the page size
|
|
|
|
SYSTEM_INFO si;
|
|
|
|
GetSystemInfo(&si);
|
|
|
|
if (si.dwPageSize > 0) os_page_size = si.dwPageSize;
|
|
|
|
if (si.dwAllocationGranularity > 0) os_alloc_granularity = si.dwAllocationGranularity;
|
|
|
|
// get the VirtualAlloc2 function
|
2019-06-29 08:35:57 +03:00
|
|
|
HINSTANCE hDll;
|
2019-07-10 09:49:12 +03:00
|
|
|
hDll = LoadLibrary(TEXT("kernelbase.dll"));
|
2019-07-04 00:52:32 +03:00
|
|
|
if (hDll != NULL) {
|
2019-07-13 18:12:16 +03:00
|
|
|
// use VirtualAlloc2FromApp if possible as it is available to Windows store apps
|
2019-10-16 23:43:57 +03:00
|
|
|
pVirtualAlloc2 = (PVirtualAlloc2)(void (*)(void))GetProcAddress(hDll, "VirtualAlloc2FromApp");
|
2019-11-21 01:55:12 +03:00
|
|
|
if (pVirtualAlloc2==NULL) pVirtualAlloc2 = (PVirtualAlloc2)(void (*)(void))GetProcAddress(hDll, "VirtualAlloc2");
|
2019-08-20 04:16:12 +03:00
|
|
|
FreeLibrary(hDll);
|
|
|
|
}
|
|
|
|
hDll = LoadLibrary(TEXT("ntdll.dll"));
|
2019-11-02 08:01:52 +03:00
|
|
|
if (hDll != NULL) {
|
2019-10-16 23:43:57 +03:00
|
|
|
pNtAllocateVirtualMemoryEx = (PNtAllocateVirtualMemoryEx)(void (*)(void))GetProcAddress(hDll, "NtAllocateVirtualMemoryEx");
|
2019-06-29 08:35:57 +03:00
|
|
|
FreeLibrary(hDll);
|
2019-11-02 08:01:52 +03:00
|
|
|
}
|
2019-09-15 01:23:28 +03:00
|
|
|
if (mi_option_is_enabled(mi_option_large_os_pages) || mi_option_is_enabled(mi_option_reserve_huge_os_pages)) {
|
|
|
|
mi_win_enable_large_os_pages();
|
2019-07-03 08:49:12 +03:00
|
|
|
}
|
2019-06-29 08:35:57 +03:00
|
|
|
}
|
2019-07-10 21:23:20 +03:00
|
|
|
#elif defined(__wasi__)
|
|
|
|
void _mi_os_init() {
|
|
|
|
os_page_size = 0x10000; // WebAssembly has a fixed page size: 64KB
|
|
|
|
os_alloc_granularity = 16;
|
|
|
|
}
|
2019-06-29 08:35:57 +03:00
|
|
|
#else
|
|
|
|
void _mi_os_init() {
|
2019-07-02 17:23:24 +03:00
|
|
|
// get the page size
|
|
|
|
long result = sysconf(_SC_PAGESIZE);
|
|
|
|
if (result > 0) {
|
|
|
|
os_page_size = (size_t)result;
|
|
|
|
os_alloc_granularity = os_page_size;
|
|
|
|
}
|
2019-12-24 21:32:44 +03:00
|
|
|
large_os_page_size = 2*MiB; // TODO: can we query the OS for this?
|
2019-06-29 08:35:57 +03:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
|
|
/* -----------------------------------------------------------
|
|
|
|
Raw allocation on Windows (VirtualAlloc) and Unix's (mmap).
|
|
|
|
----------------------------------------------------------- */
|
2019-06-20 02:26:12 +03:00
|
|
|
|
2019-08-28 04:08:03 +03:00
|
|
|
static bool mi_os_mem_free(void* addr, size_t size, bool was_committed, mi_stats_t* stats)
|
2019-06-20 02:26:12 +03:00
|
|
|
{
|
2019-11-01 05:39:49 +03:00
|
|
|
if (addr == NULL || size == 0) return true; // || _mi_os_is_huge_reserved(addr)
|
2019-06-20 02:26:12 +03:00
|
|
|
bool err = false;
|
|
|
|
#if defined(_WIN32)
|
|
|
|
err = (VirtualFree(addr, 0, MEM_RELEASE) == 0);
|
2019-07-10 21:23:20 +03:00
|
|
|
#elif defined(__wasi__)
|
|
|
|
err = 0; // WebAssembly's heap cannot be shrunk
|
2019-06-20 02:26:12 +03:00
|
|
|
#else
|
|
|
|
err = (munmap(addr, size) == -1);
|
|
|
|
#endif
|
2019-11-02 08:01:52 +03:00
|
|
|
if (was_committed) _mi_stat_decrease(&stats->committed, size);
|
2019-07-02 17:23:24 +03:00
|
|
|
_mi_stat_decrease(&stats->reserved, size);
|
2019-06-20 02:26:12 +03:00
|
|
|
if (err) {
|
2020-02-03 02:43:13 +03:00
|
|
|
#pragma warning(suppress:4996)
|
2019-06-20 02:26:12 +03:00
|
|
|
_mi_warning_message("munmap failed: %s, addr 0x%8li, size %lu\n", strerror(errno), (size_t)addr, size);
|
2019-06-24 09:15:42 +03:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
return true;
|
2019-06-20 02:26:12 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-09-06 19:19:26 +03:00
|
|
|
static void* mi_os_get_aligned_hint(size_t try_alignment, size_t size);
|
|
|
|
|
2019-07-04 00:52:32 +03:00
|
|
|
#ifdef _WIN32
|
|
|
|
static void* mi_win_virtual_allocx(void* addr, size_t size, size_t try_alignment, DWORD flags) {
|
2019-11-02 08:01:52 +03:00
|
|
|
#if (MI_INTPTR_SIZE >= 8)
|
2019-09-06 19:19:26 +03:00
|
|
|
// on 64-bit systems, try to use the virtual address area after 4TiB for 4MiB aligned allocations
|
|
|
|
void* hint;
|
2019-11-21 01:55:12 +03:00
|
|
|
if (addr == NULL && (hint = mi_os_get_aligned_hint(try_alignment,size)) != NULL) {
|
2020-03-07 03:43:39 +03:00
|
|
|
void* p = VirtualAlloc(hint, size, flags, PAGE_READWRITE);
|
|
|
|
if (p != NULL) return p;
|
|
|
|
DWORD err = GetLastError();
|
|
|
|
if (err != ERROR_INVALID_ADDRESS) { // if linked with multiple instances, we may have tried to allocate at an already allocated area
|
|
|
|
return NULL;
|
|
|
|
}
|
2019-08-23 21:33:06 +03:00
|
|
|
}
|
|
|
|
#endif
|
2019-11-02 08:01:52 +03:00
|
|
|
#if defined(MEM_EXTENDED_PARAMETER_TYPE_BITS)
|
2019-08-20 04:16:12 +03:00
|
|
|
// on modern Windows try use VirtualAlloc2 for aligned allocation
|
2019-07-04 00:52:32 +03:00
|
|
|
if (try_alignment > 0 && (try_alignment % _mi_os_page_size()) == 0 && pVirtualAlloc2 != NULL) {
|
2020-02-03 02:43:13 +03:00
|
|
|
MEM_ADDRESS_REQUIREMENTS reqs = { 0, 0, 0 };
|
2019-07-04 00:52:32 +03:00
|
|
|
reqs.Alignment = try_alignment;
|
2020-02-03 02:43:13 +03:00
|
|
|
MEM_EXTENDED_PARAMETER param = { {0, 0}, {0} };
|
2019-07-04 00:52:32 +03:00
|
|
|
param.Type = MemExtendedParameterAddressRequirements;
|
|
|
|
param.Pointer = &reqs;
|
2019-08-20 04:16:12 +03:00
|
|
|
return (*pVirtualAlloc2)(GetCurrentProcess(), addr, size, flags, PAGE_READWRITE, ¶m, 1);
|
2019-07-04 00:52:32 +03:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
return VirtualAlloc(addr, size, flags, PAGE_READWRITE);
|
|
|
|
}
|
|
|
|
|
2019-08-27 08:45:26 +03:00
|
|
|
static void* mi_win_virtual_alloc(void* addr, size_t size, size_t try_alignment, DWORD flags, bool large_only, bool allow_large, bool* is_large) {
|
|
|
|
mi_assert_internal(!(large_only && !allow_large));
|
2019-08-26 09:06:18 +03:00
|
|
|
static volatile _Atomic(uintptr_t) large_page_try_ok; // = 0;
|
2019-06-29 08:35:57 +03:00
|
|
|
void* p = NULL;
|
2019-11-02 08:01:52 +03:00
|
|
|
if ((large_only || use_large_os_page(size, try_alignment))
|
2019-11-21 01:55:12 +03:00
|
|
|
&& allow_large && (flags&MEM_COMMIT)!=0 && (flags&MEM_RESERVE)!=0) {
|
2019-07-22 03:13:36 +03:00
|
|
|
uintptr_t try_ok = mi_atomic_read(&large_page_try_ok);
|
2019-08-20 04:16:12 +03:00
|
|
|
if (!large_only && try_ok > 0) {
|
2019-07-19 11:23:14 +03:00
|
|
|
// if a large page allocation fails, it seems the calls to VirtualAlloc get very expensive.
|
2019-07-15 19:46:58 +03:00
|
|
|
// therefore, once a large page allocation failed, we don't try again for `large_page_try_ok` times.
|
2019-08-26 08:59:12 +03:00
|
|
|
mi_atomic_cas_weak(&large_page_try_ok, try_ok - 1, try_ok);
|
2019-07-15 19:46:58 +03:00
|
|
|
}
|
|
|
|
else {
|
|
|
|
// large OS pages must always reserve and commit.
|
2019-08-27 08:45:26 +03:00
|
|
|
*is_large = true;
|
|
|
|
p = mi_win_virtual_allocx(addr, size, try_alignment, flags | MEM_LARGE_PAGES);
|
2019-08-20 04:16:12 +03:00
|
|
|
if (large_only) return p;
|
2019-07-15 19:46:58 +03:00
|
|
|
// fall back to non-large page allocation on error (`p == NULL`).
|
|
|
|
if (p == NULL) {
|
2019-11-21 01:55:12 +03:00
|
|
|
mi_atomic_write(&large_page_try_ok,10); // on error, don't try again for the next N allocations
|
2019-07-15 19:46:58 +03:00
|
|
|
}
|
|
|
|
}
|
2019-06-29 08:35:57 +03:00
|
|
|
}
|
|
|
|
if (p == NULL) {
|
2019-11-21 01:55:12 +03:00
|
|
|
*is_large = ((flags&MEM_LARGE_PAGES) != 0);
|
2019-07-04 00:52:32 +03:00
|
|
|
p = mi_win_virtual_allocx(addr, size, try_alignment, flags);
|
2019-06-29 08:35:57 +03:00
|
|
|
}
|
2019-08-21 21:11:36 +03:00
|
|
|
if (p == NULL) {
|
2019-10-28 22:26:57 +03:00
|
|
|
_mi_warning_message("unable to allocate memory: error code: %i, addr: %p, size: 0x%x, large only: %d, allow_large: %d\n", GetLastError(), addr, size, large_only, allow_large);
|
2019-08-21 21:11:36 +03:00
|
|
|
}
|
2019-07-04 00:52:32 +03:00
|
|
|
return p;
|
|
|
|
}
|
|
|
|
|
2019-07-10 21:23:20 +03:00
|
|
|
#elif defined(__wasi__)
|
|
|
|
static void* mi_wasm_heap_grow(size_t size, size_t try_alignment) {
|
2019-08-11 18:06:17 +03:00
|
|
|
uintptr_t base = __builtin_wasm_memory_size(0) * _mi_os_page_size();
|
2019-11-21 01:55:12 +03:00
|
|
|
uintptr_t aligned_base = _mi_align_up(base, (uintptr_t) try_alignment);
|
|
|
|
size_t alloc_size = _mi_align_up( aligned_base - base + size, _mi_os_page_size());
|
2019-08-11 18:06:17 +03:00
|
|
|
mi_assert(alloc_size >= size && (alloc_size % _mi_os_page_size()) == 0);
|
2019-07-10 21:23:20 +03:00
|
|
|
if (alloc_size < size) return NULL;
|
2019-08-11 18:06:17 +03:00
|
|
|
if (__builtin_wasm_memory_grow(0, alloc_size / _mi_os_page_size()) == SIZE_MAX) {
|
2019-07-10 21:23:20 +03:00
|
|
|
errno = ENOMEM;
|
|
|
|
return NULL;
|
|
|
|
}
|
2019-08-11 18:06:17 +03:00
|
|
|
return (void*)aligned_base;
|
2019-07-10 21:23:20 +03:00
|
|
|
}
|
2019-06-20 02:26:12 +03:00
|
|
|
#else
|
2019-08-19 21:10:06 +03:00
|
|
|
#define MI_OS_USE_MMAP
|
|
|
|
static void* mi_unix_mmapx(void* addr, size_t size, size_t try_alignment, int protect_flags, int flags, int fd) {
|
2019-07-22 03:08:09 +03:00
|
|
|
void* p = NULL;
|
2019-11-21 01:55:12 +03:00
|
|
|
#if (MI_INTPTR_SIZE >= 8) && !defined(MAP_ALIGNED)
|
2019-08-11 18:00:10 +03:00
|
|
|
// on 64-bit systems, use the virtual address area after 4TiB for 4MiB aligned allocations
|
2019-09-06 19:19:26 +03:00
|
|
|
void* hint;
|
|
|
|
if (addr == NULL && (hint = mi_os_get_aligned_hint(try_alignment, size)) != NULL) {
|
2019-11-21 01:55:12 +03:00
|
|
|
p = mmap(hint,size,protect_flags,flags,fd,0);
|
|
|
|
if (p==MAP_FAILED) p = NULL; // fall back to regular mmap
|
2019-07-22 03:08:09 +03:00
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
#else
|
2019-08-28 10:17:58 +03:00
|
|
|
UNUSED(try_alignment);
|
2020-02-02 20:25:39 +03:00
|
|
|
UNUSED(mi_os_get_aligned_hint);
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
|
|
|
if (p==NULL) {
|
|
|
|
p = mmap(addr,size,protect_flags,flags,fd,0);
|
|
|
|
if (p==MAP_FAILED) p = NULL;
|
2019-07-22 03:08:09 +03:00
|
|
|
}
|
|
|
|
return p;
|
|
|
|
}
|
|
|
|
|
2019-08-27 08:45:26 +03:00
|
|
|
static void* mi_unix_mmap(void* addr, size_t size, size_t try_alignment, int protect_flags, bool large_only, bool allow_large, bool* is_large) {
|
2019-07-04 00:52:32 +03:00
|
|
|
void* p = NULL;
|
2019-11-21 01:55:12 +03:00
|
|
|
#if !defined(MAP_ANONYMOUS)
|
|
|
|
#define MAP_ANONYMOUS MAP_ANON
|
|
|
|
#endif
|
|
|
|
#if !defined(MAP_NORESERVE)
|
|
|
|
#define MAP_NORESERVE 0
|
|
|
|
#endif
|
2019-11-21 01:13:02 +03:00
|
|
|
int flags = MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE;
|
2019-08-10 20:26:46 +03:00
|
|
|
int fd = -1;
|
2019-11-21 01:55:12 +03:00
|
|
|
#if defined(MAP_ALIGNED) // BSD
|
2019-07-04 00:52:32 +03:00
|
|
|
if (try_alignment > 0) {
|
|
|
|
size_t n = _mi_bsr(try_alignment);
|
|
|
|
if (((size_t)1 << n) == try_alignment && n >= 12 && n <= 30) { // alignment is a power of 2 and 4096 <= alignment <= 1GiB
|
|
|
|
flags |= MAP_ALIGNED(n);
|
|
|
|
}
|
2019-06-20 02:26:12 +03:00
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
|
|
|
#if defined(PROT_MAX)
|
2019-07-04 00:52:32 +03:00
|
|
|
protect_flags |= PROT_MAX(PROT_READ | PROT_WRITE); // BSD
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
|
|
|
#if defined(VM_MAKE_TAG)
|
|
|
|
// macOS: tracking anonymous page with a specific ID. (All up to 98 are taken officially but LLVM sanitizers had taken 99)
|
2019-08-27 20:43:50 +03:00
|
|
|
int os_tag = (int)mi_option_get(mi_option_os_tag);
|
|
|
|
if (os_tag < 100 || os_tag > 255) os_tag = 100;
|
|
|
|
fd = VM_MAKE_TAG(os_tag);
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
2019-08-27 08:45:26 +03:00
|
|
|
if ((large_only || use_large_os_page(size, try_alignment)) && allow_large) {
|
2019-08-26 09:06:18 +03:00
|
|
|
static volatile _Atomic(uintptr_t) large_page_try_ok; // = 0;
|
2019-08-12 04:23:31 +03:00
|
|
|
uintptr_t try_ok = mi_atomic_read(&large_page_try_ok);
|
2019-08-20 04:16:12 +03:00
|
|
|
if (!large_only && try_ok > 0) {
|
2019-08-12 04:23:31 +03:00
|
|
|
// If the OS is not configured for large OS pages, or the user does not have
|
|
|
|
// enough permission, the `mmap` will always fail (but it might also fail for other reasons).
|
|
|
|
// Therefore, once a large page allocation failed, we don't try again for `large_page_try_ok` times
|
|
|
|
// to avoid too many failing calls to mmap.
|
2019-08-26 08:59:12 +03:00
|
|
|
mi_atomic_cas_weak(&large_page_try_ok, try_ok - 1, try_ok);
|
2019-08-12 04:23:31 +03:00
|
|
|
}
|
|
|
|
else {
|
2020-01-06 09:07:16 +03:00
|
|
|
int lflags = flags & ~MAP_NORESERVE; // using NORESERVE on huge pages seems to fail on Linux
|
2019-08-12 04:23:31 +03:00
|
|
|
int lfd = fd;
|
2019-11-21 01:55:12 +03:00
|
|
|
#ifdef MAP_ALIGNED_SUPER
|
2019-08-12 04:23:31 +03:00
|
|
|
lflags |= MAP_ALIGNED_SUPER;
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
|
|
|
#ifdef MAP_HUGETLB
|
2019-08-12 04:23:31 +03:00
|
|
|
lflags |= MAP_HUGETLB;
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
|
|
|
#ifdef MAP_HUGE_1GB
|
2019-11-04 00:16:07 +03:00
|
|
|
static bool mi_huge_pages_available = true;
|
|
|
|
if ((size % GiB) == 0 && mi_huge_pages_available) {
|
2019-08-20 04:16:12 +03:00
|
|
|
lflags |= MAP_HUGE_1GB;
|
|
|
|
}
|
|
|
|
else
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
2019-08-20 04:16:12 +03:00
|
|
|
{
|
2019-11-21 01:55:12 +03:00
|
|
|
#ifdef MAP_HUGE_2MB
|
2019-08-20 04:16:12 +03:00
|
|
|
lflags |= MAP_HUGE_2MB;
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
2019-08-20 04:16:12 +03:00
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
#ifdef VM_FLAGS_SUPERPAGE_SIZE_2MB
|
2019-08-12 04:23:31 +03:00
|
|
|
lfd |= VM_FLAGS_SUPERPAGE_SIZE_2MB;
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
2019-08-20 04:16:12 +03:00
|
|
|
if (large_only || lflags != flags) {
|
2019-08-12 04:23:31 +03:00
|
|
|
// try large OS page allocation
|
2019-08-27 08:45:26 +03:00
|
|
|
*is_large = true;
|
2019-08-20 04:16:12 +03:00
|
|
|
p = mi_unix_mmapx(addr, size, try_alignment, protect_flags, lflags, lfd);
|
2019-11-21 01:55:12 +03:00
|
|
|
#ifdef MAP_HUGE_1GB
|
2019-08-26 23:47:52 +03:00
|
|
|
if (p == NULL && (lflags & MAP_HUGE_1GB) != 0) {
|
2019-11-04 00:16:07 +03:00
|
|
|
mi_huge_pages_available = false; // don't try huge 1GiB pages again
|
2019-08-26 23:47:52 +03:00
|
|
|
_mi_warning_message("unable to allocate huge (1GiB) page, trying large (2MiB) pages instead (error %i)\n", errno);
|
|
|
|
lflags = ((lflags & ~MAP_HUGE_1GB) | MAP_HUGE_2MB);
|
|
|
|
p = mi_unix_mmapx(addr, size, try_alignment, protect_flags, lflags, lfd);
|
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
2019-08-20 04:16:12 +03:00
|
|
|
if (large_only) return p;
|
2019-08-19 21:10:06 +03:00
|
|
|
if (p == NULL) {
|
2019-08-12 18:58:22 +03:00
|
|
|
mi_atomic_write(&large_page_try_ok, 10); // on error, don't try again for the next N allocations
|
2019-08-12 04:23:31 +03:00
|
|
|
}
|
|
|
|
}
|
2019-07-02 17:23:24 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
if (p == NULL) {
|
2019-08-27 08:45:26 +03:00
|
|
|
*is_large = false;
|
2019-11-02 08:01:52 +03:00
|
|
|
p = mi_unix_mmapx(addr, size, try_alignment, protect_flags, flags, fd);
|
2019-11-21 01:55:12 +03:00
|
|
|
#if defined(MADV_HUGEPAGE)
|
2019-08-12 04:23:31 +03:00
|
|
|
// Many Linux systems don't allow MAP_HUGETLB but they support instead
|
2019-11-02 05:53:07 +03:00
|
|
|
// transparent huge pages (THP). It is not required to call `madvise` with MADV_HUGE
|
2019-08-12 04:23:31 +03:00
|
|
|
// though since properly aligned allocations will already use large pages if available
|
|
|
|
// in that case -- in particular for our large regions (in `memory.c`).
|
2019-11-02 05:53:07 +03:00
|
|
|
// However, some systems only allow THP if called with explicit `madvise`, so
|
2019-08-12 04:23:31 +03:00
|
|
|
// when large OS pages are enabled for mimalloc, we call `madvice` anyways.
|
2019-08-27 08:45:26 +03:00
|
|
|
if (allow_large && use_large_os_page(size, try_alignment)) {
|
|
|
|
if (madvise(p, size, MADV_HUGEPAGE) == 0) {
|
|
|
|
*is_large = true; // possibly
|
|
|
|
};
|
2019-08-12 04:23:31 +03:00
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
2019-07-02 17:23:24 +03:00
|
|
|
}
|
2019-07-04 00:52:32 +03:00
|
|
|
return p;
|
|
|
|
}
|
2019-06-20 02:26:12 +03:00
|
|
|
#endif
|
2019-07-04 00:52:32 +03:00
|
|
|
|
2019-11-02 08:01:52 +03:00
|
|
|
// On 64-bit systems, we can do efficient aligned allocation by using
|
2019-09-06 19:19:26 +03:00
|
|
|
// the 4TiB to 30TiB area to allocate them.
|
|
|
|
#if (MI_INTPTR_SIZE >= 8) && (defined(_WIN32) || (defined(MI_OS_USE_MMAP) && !defined(MAP_ALIGNED)))
|
2020-01-25 06:02:13 +03:00
|
|
|
static volatile mi_decl_cache_align _Atomic(uintptr_t) aligned_base;
|
2019-09-06 19:19:26 +03:00
|
|
|
|
|
|
|
// Return a 4MiB aligned address that is probably available
|
|
|
|
static void* mi_os_get_aligned_hint(size_t try_alignment, size_t size) {
|
|
|
|
if (try_alignment == 0 || try_alignment > MI_SEGMENT_SIZE) return NULL;
|
2019-11-21 01:55:12 +03:00
|
|
|
if ((size%MI_SEGMENT_SIZE) != 0) return NULL;
|
2020-01-23 06:44:54 +03:00
|
|
|
uintptr_t hint = mi_atomic_add(&aligned_base, size);
|
2019-11-21 01:55:12 +03:00
|
|
|
if (hint == 0 || hint > ((intptr_t)30<<40)) { // try to wrap around after 30TiB (area after 32TiB is used for huge OS pages)
|
2020-01-23 06:44:54 +03:00
|
|
|
uintptr_t init = ((uintptr_t)4 << 40); // start at 4TiB area
|
2019-11-21 01:55:12 +03:00
|
|
|
#if (MI_SECURE>0 || MI_DEBUG==0) // security: randomize start of aligned allocations unless in debug mode
|
2019-12-23 04:07:01 +03:00
|
|
|
uintptr_t r = _mi_heap_random_next(mi_get_default_heap());
|
|
|
|
init = init + (MI_SEGMENT_SIZE * ((r>>17) & 0xFFFFF)); // (randomly 20 bits)*4MiB == 0 to 4TiB
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
2020-01-23 06:44:54 +03:00
|
|
|
mi_atomic_cas_strong(&aligned_base, init, hint + size);
|
2019-09-06 19:19:26 +03:00
|
|
|
hint = mi_atomic_add(&aligned_base, size); // this may still give 0 or > 30TiB but that is ok, it is a hint after all
|
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
if (hint%try_alignment != 0) return NULL;
|
2019-09-06 19:19:26 +03:00
|
|
|
return (void*)hint;
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
static void* mi_os_get_aligned_hint(size_t try_alignment, size_t size) {
|
|
|
|
UNUSED(try_alignment); UNUSED(size);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
2019-07-04 00:52:32 +03:00
|
|
|
// Primitive allocation from the OS.
|
2019-08-11 18:00:10 +03:00
|
|
|
// Note: the `try_alignment` is just a hint and the returned pointer is not guaranteed to be aligned.
|
2019-08-27 08:45:26 +03:00
|
|
|
static void* mi_os_mem_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, mi_stats_t* stats) {
|
2019-07-04 00:52:32 +03:00
|
|
|
mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
|
|
|
|
if (size == 0) return NULL;
|
2019-08-27 08:45:26 +03:00
|
|
|
if (!commit) allow_large = false;
|
2019-07-04 00:52:32 +03:00
|
|
|
|
2019-08-27 08:45:26 +03:00
|
|
|
void* p = NULL;
|
2019-10-17 19:24:57 +03:00
|
|
|
/*
|
2019-08-28 21:58:30 +03:00
|
|
|
if (commit && allow_large) {
|
|
|
|
p = _mi_os_try_alloc_from_huge_reserved(size, try_alignment);
|
2019-08-27 08:45:26 +03:00
|
|
|
if (p != NULL) {
|
|
|
|
*is_large = true;
|
|
|
|
return p;
|
|
|
|
}
|
|
|
|
}
|
2019-10-17 19:24:57 +03:00
|
|
|
*/
|
2019-08-20 04:16:12 +03:00
|
|
|
|
2019-11-21 01:55:12 +03:00
|
|
|
#if defined(_WIN32)
|
|
|
|
int flags = MEM_RESERVE;
|
|
|
|
if (commit) flags |= MEM_COMMIT;
|
|
|
|
p = mi_win_virtual_alloc(NULL, size, try_alignment, flags, false, allow_large, is_large);
|
|
|
|
#elif defined(__wasi__)
|
|
|
|
*is_large = false;
|
|
|
|
p = mi_wasm_heap_grow(size, try_alignment);
|
|
|
|
#else
|
|
|
|
int protect_flags = (commit ? (PROT_WRITE | PROT_READ) : PROT_NONE);
|
|
|
|
p = mi_unix_mmap(NULL, size, try_alignment, protect_flags, false, allow_large, is_large);
|
|
|
|
#endif
|
2019-10-28 23:43:42 +03:00
|
|
|
mi_stat_counter_increase(stats->mmap_calls, 1);
|
2019-07-02 17:23:24 +03:00
|
|
|
if (p != NULL) {
|
2019-07-03 10:16:27 +03:00
|
|
|
_mi_stat_increase(&stats->reserved, size);
|
2019-08-27 08:45:26 +03:00
|
|
|
if (commit) { _mi_stat_increase(&stats->committed, size); }
|
2019-07-02 17:23:24 +03:00
|
|
|
}
|
2019-06-20 02:26:12 +03:00
|
|
|
return p;
|
|
|
|
}
|
|
|
|
|
2019-07-04 00:52:32 +03:00
|
|
|
|
|
|
|
// Primitive aligned allocation from the OS.
|
|
|
|
// This function guarantees the allocated memory is aligned.
|
2019-08-27 08:45:26 +03:00
|
|
|
static void* mi_os_mem_alloc_aligned(size_t size, size_t alignment, bool commit, bool allow_large, bool* is_large, mi_stats_t* stats) {
|
2019-07-04 00:52:32 +03:00
|
|
|
mi_assert_internal(alignment >= _mi_os_page_size() && ((alignment & (alignment - 1)) == 0));
|
|
|
|
mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
|
2019-08-27 08:45:26 +03:00
|
|
|
if (!commit) allow_large = false;
|
2019-07-04 00:52:32 +03:00
|
|
|
if (!(alignment >= _mi_os_page_size() && ((alignment & (alignment - 1)) == 0))) return NULL;
|
|
|
|
size = _mi_align_up(size, _mi_os_page_size());
|
2019-07-04 04:12:55 +03:00
|
|
|
|
2019-07-04 00:52:32 +03:00
|
|
|
// try first with a hint (this will be aligned directly on Win 10+ or BSD)
|
2019-08-27 08:45:26 +03:00
|
|
|
void* p = mi_os_mem_alloc(size, alignment, commit, allow_large, is_large, stats);
|
2019-07-04 00:52:32 +03:00
|
|
|
if (p == NULL) return NULL;
|
|
|
|
|
|
|
|
// if not aligned, free it, overallocate, and unmap around it
|
|
|
|
if (((uintptr_t)p % alignment != 0)) {
|
2019-08-28 04:08:03 +03:00
|
|
|
mi_os_mem_free(p, size, commit, stats);
|
2019-07-04 00:52:32 +03:00
|
|
|
if (size >= (SIZE_MAX - alignment)) return NULL; // overflow
|
|
|
|
size_t over_size = size + alignment;
|
|
|
|
|
|
|
|
#if _WIN32
|
|
|
|
// over-allocate and than re-allocate exactly at an aligned address in there.
|
|
|
|
// this may fail due to threads allocating at the same time so we
|
2019-07-04 04:12:55 +03:00
|
|
|
// retry this at most 3 times before giving up.
|
2019-07-04 00:52:32 +03:00
|
|
|
// (we can not decommit around the overallocation on Windows, because we can only
|
|
|
|
// free the original pointer, not one pointing inside the area)
|
|
|
|
int flags = MEM_RESERVE;
|
2019-07-02 17:23:24 +03:00
|
|
|
if (commit) flags |= MEM_COMMIT;
|
2019-07-04 00:52:32 +03:00
|
|
|
for (int tries = 0; tries < 3; tries++) {
|
|
|
|
// over-allocate to determine a virtual memory range
|
2019-08-27 08:45:26 +03:00
|
|
|
p = mi_os_mem_alloc(over_size, alignment, commit, false, is_large, stats);
|
2019-07-04 00:52:32 +03:00
|
|
|
if (p == NULL) return NULL; // error
|
|
|
|
if (((uintptr_t)p % alignment) == 0) {
|
|
|
|
// if p happens to be aligned, just decommit the left-over area
|
|
|
|
_mi_os_decommit((uint8_t*)p + size, over_size - size, stats);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
// otherwise free and allocate at an aligned address in there
|
2019-08-28 04:08:03 +03:00
|
|
|
mi_os_mem_free(p, over_size, commit, stats);
|
2019-07-04 00:52:32 +03:00
|
|
|
void* aligned_p = mi_align_up_ptr(p, alignment);
|
2019-08-27 08:45:26 +03:00
|
|
|
p = mi_win_virtual_alloc(aligned_p, size, alignment, flags, false, allow_large, is_large);
|
2019-07-04 00:52:32 +03:00
|
|
|
if (p == aligned_p) break; // success!
|
|
|
|
if (p != NULL) { // should not happen?
|
2019-08-28 04:08:03 +03:00
|
|
|
mi_os_mem_free(p, size, commit, stats);
|
2019-07-04 00:52:32 +03:00
|
|
|
p = NULL;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
// overallocate...
|
2019-08-27 08:45:26 +03:00
|
|
|
p = mi_os_mem_alloc(over_size, alignment, commit, false, is_large, stats);
|
2019-07-04 00:52:32 +03:00
|
|
|
if (p == NULL) return NULL;
|
|
|
|
// and selectively unmap parts around the over-allocated area.
|
|
|
|
void* aligned_p = mi_align_up_ptr(p, alignment);
|
|
|
|
size_t pre_size = (uint8_t*)aligned_p - (uint8_t*)p;
|
|
|
|
size_t mid_size = _mi_align_up(size, _mi_os_page_size());
|
|
|
|
size_t post_size = over_size - pre_size - mid_size;
|
|
|
|
mi_assert_internal(pre_size < over_size && post_size < over_size && mid_size >= size);
|
2019-08-28 04:08:03 +03:00
|
|
|
if (pre_size > 0) mi_os_mem_free(p, pre_size, commit, stats);
|
|
|
|
if (post_size > 0) mi_os_mem_free((uint8_t*)aligned_p + mid_size, post_size, commit, stats);
|
2019-07-04 00:52:32 +03:00
|
|
|
// we can return the aligned pointer on `mmap` systems
|
|
|
|
p = aligned_p;
|
|
|
|
#endif
|
2019-07-02 17:23:24 +03:00
|
|
|
}
|
2019-07-04 00:52:32 +03:00
|
|
|
|
|
|
|
mi_assert_internal(p == NULL || (p != NULL && ((uintptr_t)p % alignment) == 0));
|
2019-06-29 05:48:30 +03:00
|
|
|
return p;
|
|
|
|
}
|
|
|
|
|
2019-07-04 00:52:32 +03:00
|
|
|
/* -----------------------------------------------------------
|
|
|
|
OS API: alloc, free, alloc_aligned
|
|
|
|
----------------------------------------------------------- */
|
|
|
|
|
|
|
|
void* _mi_os_alloc(size_t size, mi_stats_t* stats) {
|
|
|
|
if (size == 0) return NULL;
|
2019-09-10 23:26:51 +03:00
|
|
|
size = _mi_os_good_alloc_size(size);
|
2019-08-27 08:45:26 +03:00
|
|
|
bool is_large = false;
|
|
|
|
return mi_os_mem_alloc(size, 0, true, false, &is_large, stats);
|
2019-07-04 00:52:32 +03:00
|
|
|
}
|
|
|
|
|
2019-08-28 04:08:03 +03:00
|
|
|
void _mi_os_free_ex(void* p, size_t size, bool was_committed, mi_stats_t* stats) {
|
2019-07-04 00:52:32 +03:00
|
|
|
if (size == 0 || p == NULL) return;
|
2019-09-10 23:26:51 +03:00
|
|
|
size = _mi_os_good_alloc_size(size);
|
2019-08-28 04:08:03 +03:00
|
|
|
mi_os_mem_free(p, size, was_committed, stats);
|
|
|
|
}
|
|
|
|
|
|
|
|
void _mi_os_free(void* p, size_t size, mi_stats_t* stats) {
|
|
|
|
_mi_os_free_ex(p, size, true, stats);
|
2019-07-04 00:52:32 +03:00
|
|
|
}
|
|
|
|
|
2019-08-27 08:45:26 +03:00
|
|
|
void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool* large, mi_os_tld_t* tld)
|
2019-07-04 00:52:32 +03:00
|
|
|
{
|
|
|
|
if (size == 0) return NULL;
|
2019-09-10 23:26:51 +03:00
|
|
|
size = _mi_os_good_alloc_size(size);
|
2019-07-04 00:52:32 +03:00
|
|
|
alignment = _mi_align_up(alignment, _mi_os_page_size());
|
2019-08-27 08:45:26 +03:00
|
|
|
bool allow_large = false;
|
|
|
|
if (large != NULL) {
|
|
|
|
allow_large = *large;
|
|
|
|
*large = false;
|
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
return mi_os_mem_alloc_aligned(size, alignment, commit, allow_large, (large!=NULL?large:&allow_large), tld->stats);
|
2019-07-04 00:52:32 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/* -----------------------------------------------------------
|
|
|
|
OS memory API: reset, commit, decommit, protect, unprotect.
|
|
|
|
----------------------------------------------------------- */
|
|
|
|
|
|
|
|
|
|
|
|
// OS page align within a given area, either conservative (pages inside the area only),
|
2019-07-03 08:49:12 +03:00
|
|
|
// or not (straddling pages outside the area is possible)
|
|
|
|
static void* mi_os_page_align_areax(bool conservative, void* addr, size_t size, size_t* newsize) {
|
2019-06-20 02:26:12 +03:00
|
|
|
mi_assert(addr != NULL && size > 0);
|
|
|
|
if (newsize != NULL) *newsize = 0;
|
|
|
|
if (size == 0 || addr == NULL) return NULL;
|
|
|
|
|
|
|
|
// page align conservatively within the range
|
2019-07-03 08:49:12 +03:00
|
|
|
void* start = (conservative ? mi_align_up_ptr(addr, _mi_os_page_size())
|
2019-07-04 00:52:32 +03:00
|
|
|
: mi_align_down_ptr(addr, _mi_os_page_size()));
|
2019-07-03 08:49:12 +03:00
|
|
|
void* end = (conservative ? mi_align_down_ptr((uint8_t*)addr + size, _mi_os_page_size())
|
2019-07-04 00:52:32 +03:00
|
|
|
: mi_align_up_ptr((uint8_t*)addr + size, _mi_os_page_size()));
|
2019-06-20 02:26:12 +03:00
|
|
|
ptrdiff_t diff = (uint8_t*)end - (uint8_t*)start;
|
|
|
|
if (diff <= 0) return NULL;
|
|
|
|
|
2019-07-10 09:34:17 +03:00
|
|
|
mi_assert_internal((conservative && (size_t)diff <= size) || (!conservative && (size_t)diff >= size));
|
2019-06-24 06:41:34 +03:00
|
|
|
if (newsize != NULL) *newsize = (size_t)diff;
|
2019-06-20 02:26:12 +03:00
|
|
|
return start;
|
|
|
|
}
|
|
|
|
|
2019-07-03 08:49:12 +03:00
|
|
|
static void* mi_os_page_align_area_conservative(void* addr, size_t size, size_t* newsize) {
|
2019-07-04 00:52:32 +03:00
|
|
|
return mi_os_page_align_areax(true, addr, size, newsize);
|
2019-07-03 08:49:12 +03:00
|
|
|
}
|
|
|
|
|
2020-01-10 04:52:28 +03:00
|
|
|
static void mi_mprotect_hint(int err) {
|
|
|
|
#if defined(MI_OS_USE_MMAP) && (MI_SECURE>=2) // guard page around every mimalloc page
|
|
|
|
if (err == ENOMEM) {
|
|
|
|
_mi_warning_message("the previous warning may have been caused by a low memory map limit.\n"
|
|
|
|
" On Linux this is controlled by the vm.max_map_count. For example:\n"
|
|
|
|
" > sudo sysctl -w vm.max_map_count=262144\n");
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
UNUSED(err);
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
2019-07-22 03:08:09 +03:00
|
|
|
// Commit/Decommit memory.
|
2019-07-10 17:17:21 +03:00
|
|
|
// Usuelly commit is aligned liberal, while decommit is aligned conservative.
|
|
|
|
// (but not for the reset version where we want commit to be conservative as well)
|
2019-09-02 23:16:52 +03:00
|
|
|
static bool mi_os_commitx(void* addr, size_t size, bool commit, bool conservative, bool* is_zero, mi_stats_t* stats) {
|
2019-07-10 06:24:00 +03:00
|
|
|
// page align in the range, commit liberally, decommit conservative
|
2019-11-25 09:00:11 +03:00
|
|
|
if (is_zero != NULL) { *is_zero = false; }
|
2019-07-10 06:24:00 +03:00
|
|
|
size_t csize;
|
2019-07-10 17:50:04 +03:00
|
|
|
void* start = mi_os_page_align_areax(conservative, addr, size, &csize);
|
2019-11-01 05:39:49 +03:00
|
|
|
if (csize == 0) return true; // || _mi_os_is_huge_reserved(addr))
|
2019-07-10 06:24:00 +03:00
|
|
|
int err = 0;
|
|
|
|
if (commit) {
|
|
|
|
_mi_stat_increase(&stats->committed, csize);
|
2019-10-28 23:43:42 +03:00
|
|
|
_mi_stat_counter_increase(&stats->commit_calls, 1);
|
2019-07-10 06:24:00 +03:00
|
|
|
}
|
|
|
|
else {
|
|
|
|
_mi_stat_decrease(&stats->committed, csize);
|
|
|
|
}
|
|
|
|
|
2019-11-21 01:55:12 +03:00
|
|
|
#if defined(_WIN32)
|
2019-07-10 06:24:00 +03:00
|
|
|
if (commit) {
|
2019-09-03 00:53:22 +03:00
|
|
|
// if the memory was already committed, the call succeeds but it is not zero'd
|
|
|
|
// *is_zero = true;
|
2019-07-10 06:24:00 +03:00
|
|
|
void* p = VirtualAlloc(start, csize, MEM_COMMIT, PAGE_READWRITE);
|
|
|
|
err = (p == start ? 0 : GetLastError());
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
BOOL ok = VirtualFree(start, csize, MEM_DECOMMIT);
|
|
|
|
err = (ok ? 0 : GetLastError());
|
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
#elif defined(__wasi__)
|
2019-07-10 21:23:20 +03:00
|
|
|
// WebAssembly guests can't control memory protection
|
2019-11-21 01:55:12 +03:00
|
|
|
#elif defined(MAP_FIXED)
|
2019-11-21 01:13:02 +03:00
|
|
|
if (!commit) {
|
|
|
|
// use mmap with MAP_FIXED to discard the existing memory (and reduce commit charge)
|
2019-11-25 21:11:29 +03:00
|
|
|
void* p = mmap(start, csize, PROT_NONE, (MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE), -1, 0);
|
2019-11-21 01:13:02 +03:00
|
|
|
if (p != start) { err = errno; }
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
// for commit, just change the protection
|
|
|
|
err = mprotect(start, csize, (PROT_READ | PROT_WRITE));
|
|
|
|
if (err != 0) { err = errno; }
|
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
#else
|
2019-07-10 06:24:00 +03:00
|
|
|
err = mprotect(start, csize, (commit ? (PROT_READ | PROT_WRITE) : PROT_NONE));
|
2019-09-18 03:23:20 +03:00
|
|
|
if (err != 0) { err = errno; }
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
2019-07-10 06:24:00 +03:00
|
|
|
if (err != 0) {
|
2020-01-18 06:59:55 +03:00
|
|
|
_mi_warning_message("%s error: start: %p, csize: 0x%x, err: %i\n", commit ? "commit" : "decommit", start, csize, err);
|
2020-01-10 04:52:28 +03:00
|
|
|
mi_mprotect_hint(err);
|
2019-07-10 06:24:00 +03:00
|
|
|
}
|
|
|
|
mi_assert_internal(err == 0);
|
|
|
|
return (err == 0);
|
|
|
|
}
|
2019-07-03 08:49:12 +03:00
|
|
|
|
2019-09-02 23:16:52 +03:00
|
|
|
bool _mi_os_commit(void* addr, size_t size, bool* is_zero, mi_stats_t* stats) {
|
2019-11-21 01:13:02 +03:00
|
|
|
return mi_os_commitx(addr, size, true, false /* liberal */, is_zero, stats);
|
2019-07-10 06:24:00 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* stats) {
|
2019-09-02 23:16:52 +03:00
|
|
|
bool is_zero;
|
2019-11-21 01:13:02 +03:00
|
|
|
return mi_os_commitx(addr, size, false, true /* conservative */, &is_zero, stats);
|
2019-07-10 06:24:00 +03:00
|
|
|
}
|
2019-07-03 08:49:12 +03:00
|
|
|
|
2019-11-21 01:13:02 +03:00
|
|
|
bool _mi_os_commit_unreset(void* addr, size_t size, bool* is_zero, mi_stats_t* stats) {
|
|
|
|
return mi_os_commitx(addr, size, true, true /* conservative */, is_zero, stats);
|
|
|
|
}
|
2019-07-10 17:17:21 +03:00
|
|
|
|
2019-06-20 02:26:12 +03:00
|
|
|
// Signal to the OS that the address range is no longer in use
|
|
|
|
// but may be used later again. This will release physical memory
|
|
|
|
// pages and reduce swapping while keeping the memory committed.
|
|
|
|
// We page align to a conservative area inside the range to reset.
|
2019-07-10 06:24:00 +03:00
|
|
|
static bool mi_os_resetx(void* addr, size_t size, bool reset, mi_stats_t* stats) {
|
2019-06-20 02:26:12 +03:00
|
|
|
// page align conservatively within the range
|
|
|
|
size_t csize;
|
2019-07-04 00:52:32 +03:00
|
|
|
void* start = mi_os_page_align_area_conservative(addr, size, &csize);
|
2019-11-01 05:39:49 +03:00
|
|
|
if (csize == 0) return true; // || _mi_os_is_huge_reserved(addr)
|
2019-07-10 06:24:00 +03:00
|
|
|
if (reset) _mi_stat_increase(&stats->reset, csize);
|
2019-11-21 01:55:12 +03:00
|
|
|
else _mi_stat_decrease(&stats->reset, csize);
|
2019-07-10 06:24:00 +03:00
|
|
|
if (!reset) return true; // nothing to do on unreset!
|
2019-06-20 02:26:12 +03:00
|
|
|
|
2019-11-21 01:55:12 +03:00
|
|
|
#if (MI_DEBUG>1)
|
|
|
|
if (MI_SECURE==0) {
|
2019-07-19 19:27:32 +03:00
|
|
|
memset(start, 0, csize); // pretend it is eagerly reset
|
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
2019-07-10 17:17:21 +03:00
|
|
|
|
2019-06-20 02:26:12 +03:00
|
|
|
#if defined(_WIN32)
|
2019-07-02 17:23:24 +03:00
|
|
|
// Testing shows that for us (on `malloc-large`) MEM_RESET is 2x faster than DiscardVirtualMemory
|
2019-08-20 04:16:12 +03:00
|
|
|
void* p = VirtualAlloc(start, csize, MEM_RESET, PAGE_READWRITE);
|
|
|
|
mi_assert_internal(p == start);
|
2019-11-21 01:55:12 +03:00
|
|
|
#if 1
|
2019-10-18 02:56:57 +03:00
|
|
|
if (p == start && start != NULL) {
|
2019-11-21 01:55:12 +03:00
|
|
|
VirtualUnlock(start,csize); // VirtualUnlock after MEM_RESET removes the memory from the working set
|
2019-08-27 08:45:26 +03:00
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
2019-08-20 17:06:11 +03:00
|
|
|
if (p != start) return false;
|
2019-06-20 02:26:12 +03:00
|
|
|
#else
|
2019-07-04 00:52:32 +03:00
|
|
|
#if defined(MADV_FREE)
|
|
|
|
static int advice = MADV_FREE;
|
|
|
|
int err = madvise(start, csize, advice);
|
|
|
|
if (err != 0 && errno == EINVAL && advice == MADV_FREE) {
|
|
|
|
// if MADV_FREE is not supported, fall back to MADV_DONTNEED from now on
|
|
|
|
advice = MADV_DONTNEED;
|
|
|
|
err = madvise(start, csize, advice);
|
|
|
|
}
|
2019-07-10 21:23:20 +03:00
|
|
|
#elif defined(__wasi__)
|
|
|
|
int err = 0;
|
2019-07-04 00:52:32 +03:00
|
|
|
#else
|
|
|
|
int err = madvise(start, csize, MADV_DONTNEED);
|
|
|
|
#endif
|
2019-06-20 02:26:12 +03:00
|
|
|
if (err != 0) {
|
2020-01-18 06:59:55 +03:00
|
|
|
_mi_warning_message("madvise reset error: start: %p, csize: 0x%x, errno: %i\n", start, csize, errno);
|
2019-06-20 02:26:12 +03:00
|
|
|
}
|
|
|
|
//mi_assert(err == 0);
|
2019-07-10 17:17:21 +03:00
|
|
|
if (err != 0) return false;
|
2019-06-20 02:26:12 +03:00
|
|
|
#endif
|
2019-07-10 17:17:21 +03:00
|
|
|
return true;
|
2019-06-20 02:26:12 +03:00
|
|
|
}
|
|
|
|
|
2019-07-10 06:24:00 +03:00
|
|
|
// Signal to the OS that the address range is no longer in use
|
|
|
|
// but may be used later again. This will release physical memory
|
|
|
|
// pages and reduce swapping while keeping the memory committed.
|
|
|
|
// We page align to a conservative area inside the range to reset.
|
|
|
|
bool _mi_os_reset(void* addr, size_t size, mi_stats_t* stats) {
|
2019-11-21 01:13:02 +03:00
|
|
|
if (mi_option_is_enabled(mi_option_reset_decommits)) {
|
|
|
|
return _mi_os_decommit(addr, size, stats);
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
return mi_os_resetx(addr, size, true, stats);
|
|
|
|
}
|
2019-07-10 06:24:00 +03:00
|
|
|
}
|
|
|
|
|
2019-09-02 23:16:52 +03:00
|
|
|
bool _mi_os_unreset(void* addr, size_t size, bool* is_zero, mi_stats_t* stats) {
|
2019-11-21 01:13:02 +03:00
|
|
|
if (mi_option_is_enabled(mi_option_reset_decommits)) {
|
|
|
|
return _mi_os_commit_unreset(addr, size, is_zero, stats); // re-commit it (conservatively!)
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
*is_zero = false;
|
|
|
|
return mi_os_resetx(addr, size, false, stats);
|
|
|
|
}
|
2019-07-10 06:24:00 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2019-06-20 02:26:12 +03:00
|
|
|
// Protect a region in memory to be not accessible.
|
|
|
|
static bool mi_os_protectx(void* addr, size_t size, bool protect) {
|
|
|
|
// page align conservatively within the range
|
|
|
|
size_t csize = 0;
|
2019-07-03 08:49:12 +03:00
|
|
|
void* start = mi_os_page_align_area_conservative(addr, size, &csize);
|
2019-07-04 00:52:32 +03:00
|
|
|
if (csize == 0) return false;
|
2019-11-01 05:39:49 +03:00
|
|
|
/*
|
2019-08-27 08:45:26 +03:00
|
|
|
if (_mi_os_is_huge_reserved(addr)) {
|
2019-11-21 01:55:12 +03:00
|
|
|
_mi_warning_message("cannot mprotect memory allocated in huge OS pages\n");
|
2019-08-23 21:22:35 +03:00
|
|
|
}
|
2019-11-01 05:39:49 +03:00
|
|
|
*/
|
2019-06-20 02:26:12 +03:00
|
|
|
int err = 0;
|
|
|
|
#ifdef _WIN32
|
|
|
|
DWORD oldprotect = 0;
|
2019-07-04 00:52:32 +03:00
|
|
|
BOOL ok = VirtualProtect(start, csize, protect ? PAGE_NOACCESS : PAGE_READWRITE, &oldprotect);
|
2019-07-02 17:23:24 +03:00
|
|
|
err = (ok ? 0 : GetLastError());
|
2019-07-10 21:23:20 +03:00
|
|
|
#elif defined(__wasi__)
|
|
|
|
err = 0;
|
2019-06-20 02:26:12 +03:00
|
|
|
#else
|
2019-07-04 00:52:32 +03:00
|
|
|
err = mprotect(start, csize, protect ? PROT_NONE : (PROT_READ | PROT_WRITE));
|
2019-09-18 03:23:20 +03:00
|
|
|
if (err != 0) { err = errno; }
|
2019-06-20 02:26:12 +03:00
|
|
|
#endif
|
|
|
|
if (err != 0) {
|
2020-01-18 06:59:55 +03:00
|
|
|
_mi_warning_message("mprotect error: start: %p, csize: 0x%x, err: %i\n", start, csize, err);
|
2020-01-10 04:52:28 +03:00
|
|
|
mi_mprotect_hint(err);
|
2019-06-20 02:26:12 +03:00
|
|
|
}
|
2019-07-04 00:52:32 +03:00
|
|
|
return (err == 0);
|
2019-06-20 02:26:12 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
bool _mi_os_protect(void* addr, size_t size) {
|
2019-07-04 00:52:32 +03:00
|
|
|
return mi_os_protectx(addr, size, true);
|
2019-06-20 02:26:12 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
bool _mi_os_unprotect(void* addr, size_t size) {
|
|
|
|
return mi_os_protectx(addr, size, false);
|
|
|
|
}
|
|
|
|
|
2019-07-02 17:23:24 +03:00
|
|
|
|
2019-06-24 09:15:42 +03:00
|
|
|
|
2019-07-03 03:17:03 +03:00
|
|
|
bool _mi_os_shrink(void* p, size_t oldsize, size_t newsize, mi_stats_t* stats) {
|
2019-06-24 09:15:42 +03:00
|
|
|
// page align conservatively within the range
|
2019-11-21 01:55:12 +03:00
|
|
|
mi_assert_internal(oldsize > newsize && p != NULL);
|
2019-07-04 00:52:32 +03:00
|
|
|
if (oldsize < newsize || p == NULL) return false;
|
2019-06-24 09:15:42 +03:00
|
|
|
if (oldsize == newsize) return true;
|
|
|
|
|
|
|
|
// oldsize and newsize should be page aligned or we cannot shrink precisely
|
|
|
|
void* addr = (uint8_t*)p + newsize;
|
|
|
|
size_t size = 0;
|
2019-07-03 08:49:12 +03:00
|
|
|
void* start = mi_os_page_align_area_conservative(addr, oldsize - newsize, &size);
|
2019-07-04 00:52:32 +03:00
|
|
|
if (size == 0 || start != addr) return false;
|
2019-06-20 02:26:12 +03:00
|
|
|
|
2019-07-04 00:52:32 +03:00
|
|
|
#ifdef _WIN32
|
2019-07-03 08:49:12 +03:00
|
|
|
// we cannot shrink on windows, but we can decommit
|
2019-07-03 10:16:27 +03:00
|
|
|
return _mi_os_decommit(start, size, stats);
|
2019-07-04 00:52:32 +03:00
|
|
|
#else
|
2019-08-28 04:08:03 +03:00
|
|
|
return mi_os_mem_free(start, size, true, stats);
|
2019-06-20 02:26:12 +03:00
|
|
|
#endif
|
|
|
|
}
|
2019-08-19 21:10:06 +03:00
|
|
|
|
|
|
|
|
|
|
|
/* ----------------------------------------------------------------------------
|
2019-11-02 08:01:52 +03:00
|
|
|
Support for allocating huge OS pages (1Gib) that are reserved up-front
|
2019-11-02 05:53:07 +03:00
|
|
|
and possibly associated with a specific NUMA node. (use `numa_node>=0`)
|
2019-08-19 21:10:06 +03:00
|
|
|
-----------------------------------------------------------------------------*/
|
2019-11-02 08:01:52 +03:00
|
|
|
#define MI_HUGE_OS_PAGE_SIZE (GiB)
|
2019-11-02 05:53:07 +03:00
|
|
|
|
|
|
|
#if defined(WIN32) && (MI_INTPTR_SIZE >= 8)
|
2019-11-03 03:49:34 +03:00
|
|
|
static void* mi_os_alloc_huge_os_pagesx(void* addr, size_t size, int numa_node)
|
2019-11-02 08:01:52 +03:00
|
|
|
{
|
2019-11-21 01:55:12 +03:00
|
|
|
mi_assert_internal(size%GiB == 0);
|
2019-11-03 03:49:34 +03:00
|
|
|
mi_assert_internal(addr != NULL);
|
|
|
|
const DWORD flags = MEM_LARGE_PAGES | MEM_COMMIT | MEM_RESERVE;
|
2019-11-02 05:53:07 +03:00
|
|
|
|
2019-11-03 03:49:34 +03:00
|
|
|
mi_win_enable_large_os_pages();
|
2019-11-12 21:16:59 +03:00
|
|
|
|
2019-11-21 01:55:12 +03:00
|
|
|
#if defined(MEM_EXTENDED_PARAMETER_TYPE_BITS)
|
2020-02-03 02:43:13 +03:00
|
|
|
MEM_EXTENDED_PARAMETER params[3] = { {{0,0},{0}},{{0,0},{0}},{{0,0},{0}} };
|
2019-11-02 08:01:52 +03:00
|
|
|
// on modern Windows try use NtAllocateVirtualMemoryEx for 1GiB huge pages
|
2019-11-04 00:16:07 +03:00
|
|
|
static bool mi_huge_pages_available = true;
|
|
|
|
if (pNtAllocateVirtualMemoryEx != NULL && mi_huge_pages_available) {
|
2019-11-21 01:55:12 +03:00
|
|
|
#ifndef MEM_EXTENDED_PARAMETER_NONPAGED_HUGE
|
|
|
|
#define MEM_EXTENDED_PARAMETER_NONPAGED_HUGE (0x10)
|
|
|
|
#endif
|
2019-11-03 03:49:34 +03:00
|
|
|
params[0].Type = 5; // == MemExtendedParameterAttributeFlags;
|
|
|
|
params[0].ULong64 = MEM_EXTENDED_PARAMETER_NONPAGED_HUGE;
|
|
|
|
ULONG param_count = 1;
|
2019-11-02 05:53:07 +03:00
|
|
|
if (numa_node >= 0) {
|
|
|
|
param_count++;
|
2019-11-03 03:49:34 +03:00
|
|
|
params[1].Type = MemExtendedParameterNumaNode;
|
|
|
|
params[1].ULong = (unsigned)numa_node;
|
2019-08-19 21:10:06 +03:00
|
|
|
}
|
2019-11-02 05:53:07 +03:00
|
|
|
SIZE_T psize = size;
|
2019-11-03 03:49:34 +03:00
|
|
|
void* base = addr;
|
2019-11-02 05:53:07 +03:00
|
|
|
NTSTATUS err = (*pNtAllocateVirtualMemoryEx)(GetCurrentProcess(), &base, &psize, flags, PAGE_READWRITE, params, param_count);
|
2019-11-03 03:49:34 +03:00
|
|
|
if (err == 0 && base != NULL) {
|
2019-11-02 05:53:07 +03:00
|
|
|
return base;
|
2019-08-19 21:10:06 +03:00
|
|
|
}
|
2019-11-02 05:53:07 +03:00
|
|
|
else {
|
2019-11-04 00:16:07 +03:00
|
|
|
// fall back to regular large pages
|
|
|
|
mi_huge_pages_available = false; // don't try further huge pages
|
2020-02-01 04:28:26 +03:00
|
|
|
_mi_warning_message("unable to allocate using huge (1gb) pages, trying large (2mb) pages instead (status 0x%lx)\n", err);
|
2019-08-19 21:10:06 +03:00
|
|
|
}
|
2019-11-02 08:01:52 +03:00
|
|
|
}
|
2019-11-03 03:49:34 +03:00
|
|
|
// on modern Windows try use VirtualAlloc2 for numa aware large OS page allocation
|
|
|
|
if (pVirtualAlloc2 != NULL && numa_node >= 0) {
|
|
|
|
params[0].Type = MemExtendedParameterNumaNode;
|
2019-11-12 21:16:59 +03:00
|
|
|
params[0].ULong = (unsigned)numa_node;
|
2019-11-04 00:16:07 +03:00
|
|
|
return (*pVirtualAlloc2)(GetCurrentProcess(), addr, size, flags, PAGE_READWRITE, params, 1);
|
2019-11-02 05:53:07 +03:00
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
#endif
|
2019-11-04 00:16:07 +03:00
|
|
|
// otherwise use regular virtual alloc on older windows
|
|
|
|
return VirtualAlloc(addr, size, flags, PAGE_READWRITE);
|
2019-11-02 05:53:07 +03:00
|
|
|
}
|
2019-11-03 03:49:34 +03:00
|
|
|
|
2019-11-02 05:53:07 +03:00
|
|
|
#elif defined(MI_OS_USE_MMAP) && (MI_INTPTR_SIZE >= 8)
|
2019-11-12 21:37:15 +03:00
|
|
|
#include <sys/syscall.h>
|
|
|
|
#ifndef MPOL_PREFERRED
|
|
|
|
#define MPOL_PREFERRED 1
|
|
|
|
#endif
|
|
|
|
#if defined(SYS_mbind)
|
|
|
|
static long mi_os_mbind(void* start, unsigned long len, unsigned long mode, const unsigned long* nmask, unsigned long maxnode, unsigned flags) {
|
|
|
|
return syscall(SYS_mbind, start, len, mode, nmask, maxnode, flags);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
static long mi_os_mbind(void* start, unsigned long len, unsigned long mode, const unsigned long* nmask, unsigned long maxnode, unsigned flags) {
|
|
|
|
UNUSED(start); UNUSED(len); UNUSED(mode); UNUSED(nmask); UNUSED(maxnode); UNUSED(flags);
|
|
|
|
return 0;
|
|
|
|
}
|
2019-11-02 05:53:07 +03:00
|
|
|
#endif
|
2019-11-03 03:49:34 +03:00
|
|
|
static void* mi_os_alloc_huge_os_pagesx(void* addr, size_t size, int numa_node) {
|
2019-11-21 01:55:12 +03:00
|
|
|
mi_assert_internal(size%GiB == 0);
|
2019-11-02 05:53:07 +03:00
|
|
|
bool is_large = true;
|
2019-11-03 03:49:34 +03:00
|
|
|
void* p = mi_unix_mmap(addr, size, MI_SEGMENT_SIZE, PROT_READ | PROT_WRITE, true, true, &is_large);
|
2019-11-02 05:53:07 +03:00
|
|
|
if (p == NULL) return NULL;
|
2019-11-21 01:55:12 +03:00
|
|
|
if (numa_node >= 0 && numa_node < 8*MI_INTPTR_SIZE) { // at most 64 nodes
|
2019-11-02 05:53:07 +03:00
|
|
|
uintptr_t numa_mask = (1UL << numa_node);
|
2019-11-12 21:16:59 +03:00
|
|
|
// TODO: does `mbind` work correctly for huge OS pages? should we
|
2019-11-02 20:30:16 +03:00
|
|
|
// use `set_mempolicy` before calling mmap instead?
|
|
|
|
// see: <https://lkml.org/lkml/2017/2/9/875>
|
2019-11-21 01:55:12 +03:00
|
|
|
long err = mi_os_mbind(p, size, MPOL_PREFERRED, &numa_mask, 8*MI_INTPTR_SIZE, 0);
|
2019-11-02 05:53:07 +03:00
|
|
|
if (err != 0) {
|
2020-02-01 04:28:26 +03:00
|
|
|
_mi_warning_message("failed to bind huge (1gb) pages to numa node %d: %s\n", numa_node, strerror(errno));
|
2019-11-02 05:53:07 +03:00
|
|
|
}
|
|
|
|
}
|
|
|
|
return p;
|
|
|
|
}
|
2019-11-02 08:01:52 +03:00
|
|
|
#else
|
2019-11-03 03:49:34 +03:00
|
|
|
static void* mi_os_alloc_huge_os_pagesx(void* addr, size_t size, int numa_node) {
|
2019-11-02 05:53:07 +03:00
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2019-11-12 21:16:59 +03:00
|
|
|
#if (MI_INTPTR_SIZE >= 8)
|
2019-11-03 03:49:34 +03:00
|
|
|
// To ensure proper alignment, use our own area for huge OS pages
|
2020-01-25 06:02:13 +03:00
|
|
|
static mi_decl_cache_align _Atomic(uintptr_t) mi_huge_start; // = 0
|
2019-11-03 03:49:34 +03:00
|
|
|
|
2019-11-04 00:16:07 +03:00
|
|
|
// Claim an aligned address range for huge pages
|
|
|
|
static uint8_t* mi_os_claim_huge_pages(size_t pages, size_t* total_size) {
|
|
|
|
if (total_size != NULL) *total_size = 0;
|
2019-11-03 03:49:34 +03:00
|
|
|
const size_t size = pages * MI_HUGE_OS_PAGE_SIZE;
|
|
|
|
|
|
|
|
uintptr_t start = 0;
|
|
|
|
uintptr_t end = 0;
|
|
|
|
uintptr_t expected;
|
|
|
|
do {
|
2019-11-04 00:16:07 +03:00
|
|
|
start = expected = mi_atomic_read_relaxed(&mi_huge_start);
|
2019-11-03 03:49:34 +03:00
|
|
|
if (start == 0) {
|
|
|
|
// Initialize the start address after the 32TiB area
|
2019-11-04 00:16:07 +03:00
|
|
|
start = ((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
|
2019-12-23 04:07:01 +03:00
|
|
|
uintptr_t r = _mi_heap_random_next(mi_get_default_heap());
|
|
|
|
start = start + ((uintptr_t)MI_HUGE_OS_PAGE_SIZE * ((r>>17) & 0x0FFF)); // (randomly 12bits)*1GiB == between 0 to 4TiB
|
2019-11-04 00:16:07 +03:00
|
|
|
#endif
|
2019-11-03 03:49:34 +03:00
|
|
|
}
|
|
|
|
end = start + size;
|
|
|
|
mi_assert_internal(end % MI_SEGMENT_SIZE == 0);
|
|
|
|
} while (!mi_atomic_cas_strong(&mi_huge_start, end, expected));
|
|
|
|
|
2019-11-04 00:16:07 +03:00
|
|
|
if (total_size != NULL) *total_size = size;
|
|
|
|
return (uint8_t*)start;
|
|
|
|
}
|
2019-11-04 19:44:40 +03:00
|
|
|
#else
|
|
|
|
static uint8_t* mi_os_claim_huge_pages(size_t pages, size_t* total_size) {
|
|
|
|
if (total_size != NULL) *total_size = 0;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
#endif
|
2019-11-04 00:16:07 +03:00
|
|
|
|
|
|
|
// Allocate MI_SEGMENT_SIZE aligned huge pages
|
2019-11-04 19:44:40 +03:00
|
|
|
void* _mi_os_alloc_huge_os_pages(size_t pages, int numa_node, mi_msecs_t max_msecs, size_t* pages_reserved, size_t* psize) {
|
2019-11-04 00:16:07 +03:00
|
|
|
if (psize != NULL) *psize = 0;
|
|
|
|
if (pages_reserved != NULL) *pages_reserved = 0;
|
|
|
|
size_t size = 0;
|
|
|
|
uint8_t* start = mi_os_claim_huge_pages(pages, &size);
|
2019-11-04 19:44:40 +03:00
|
|
|
if (start == NULL) return NULL; // or 32-bit systems
|
2019-11-12 21:16:59 +03:00
|
|
|
|
2019-11-04 00:16:07 +03:00
|
|
|
// 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
|
|
|
|
// or to at least allocate as many as available on the system.
|
2019-11-04 19:44:40 +03:00
|
|
|
mi_msecs_t start_t = _mi_clock_start();
|
2019-11-04 00:16:07 +03:00
|
|
|
size_t page;
|
|
|
|
for (page = 0; page < pages; page++) {
|
|
|
|
// allocate a page
|
|
|
|
void* addr = start + (page * MI_HUGE_OS_PAGE_SIZE);
|
|
|
|
void* p = mi_os_alloc_huge_os_pagesx(addr, MI_HUGE_OS_PAGE_SIZE, numa_node);
|
|
|
|
|
|
|
|
// Did we succeed at a contiguous address?
|
|
|
|
if (p != addr) {
|
|
|
|
// no success, issue a warning and break
|
|
|
|
if (p != NULL) {
|
2020-01-18 06:59:55 +03:00
|
|
|
_mi_warning_message("could not allocate contiguous huge page %zu at %p\n", page, addr);
|
2019-11-04 00:16:07 +03:00
|
|
|
_mi_os_free(p, MI_HUGE_OS_PAGE_SIZE, &_mi_stats_main);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
2019-11-12 21:16:59 +03:00
|
|
|
|
2019-11-04 00:16:07 +03:00
|
|
|
// success, record it
|
|
|
|
_mi_stat_increase(&_mi_stats_main.committed, MI_HUGE_OS_PAGE_SIZE);
|
|
|
|
_mi_stat_increase(&_mi_stats_main.reserved, MI_HUGE_OS_PAGE_SIZE);
|
2019-11-12 21:16:59 +03:00
|
|
|
|
2019-11-04 00:16:07 +03:00
|
|
|
// check for timeout
|
2019-11-07 09:49:01 +03:00
|
|
|
if (max_msecs > 0) {
|
|
|
|
mi_msecs_t elapsed = _mi_clock_end(start_t);
|
|
|
|
if (page >= 1) {
|
2019-11-21 01:55:12 +03:00
|
|
|
mi_msecs_t estimate = ((elapsed / (page+1)) * pages);
|
|
|
|
if (estimate > 2*max_msecs) { // seems like we are going to timeout, break
|
2019-11-07 09:49:01 +03:00
|
|
|
elapsed = max_msecs + 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (elapsed > max_msecs) {
|
|
|
|
_mi_warning_message("huge page allocation timed out\n");
|
|
|
|
break;
|
2019-11-04 00:16:07 +03:00
|
|
|
}
|
|
|
|
}
|
2019-11-03 03:49:34 +03:00
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
mi_assert_internal(page*MI_HUGE_OS_PAGE_SIZE <= size);
|
2019-11-04 00:16:07 +03:00
|
|
|
if (pages_reserved != NULL) *pages_reserved = page;
|
|
|
|
if (psize != NULL) *psize = page * MI_HUGE_OS_PAGE_SIZE;
|
|
|
|
return (page == 0 ? NULL : start);
|
|
|
|
}
|
|
|
|
|
|
|
|
// free every huge page in a range individually (as we allocated per page)
|
|
|
|
// note: needed with VirtualAlloc but could potentially be done in one go on mmap'd systems.
|
|
|
|
void _mi_os_free_huge_pages(void* p, size_t size, mi_stats_t* stats) {
|
2019-11-21 01:55:12 +03:00
|
|
|
if (p==NULL || size==0) return;
|
2019-11-04 00:16:07 +03:00
|
|
|
uint8_t* base = (uint8_t*)p;
|
|
|
|
while (size >= MI_HUGE_OS_PAGE_SIZE) {
|
|
|
|
_mi_os_free(base, MI_HUGE_OS_PAGE_SIZE, stats);
|
|
|
|
size -= MI_HUGE_OS_PAGE_SIZE;
|
2019-11-03 03:49:34 +03:00
|
|
|
}
|
2019-11-02 05:53:07 +03:00
|
|
|
}
|
|
|
|
|
2019-11-02 20:30:16 +03:00
|
|
|
/* ----------------------------------------------------------------------------
|
2019-11-12 21:16:59 +03:00
|
|
|
Support NUMA aware allocation
|
2019-11-02 20:30:16 +03:00
|
|
|
-----------------------------------------------------------------------------*/
|
2019-11-02 05:53:07 +03:00
|
|
|
#ifdef WIN32
|
2019-11-14 00:35:50 +03:00
|
|
|
static size_t mi_os_numa_nodex() {
|
2019-11-02 05:53:07 +03:00
|
|
|
PROCESSOR_NUMBER pnum;
|
|
|
|
USHORT numa_node = 0;
|
|
|
|
GetCurrentProcessorNumberEx(&pnum);
|
2019-11-21 01:55:12 +03:00
|
|
|
GetNumaProcessorNodeEx(&pnum,&numa_node);
|
2019-11-14 00:35:50 +03:00
|
|
|
return numa_node;
|
2019-11-02 05:53:07 +03:00
|
|
|
}
|
|
|
|
|
2019-11-14 00:35:50 +03:00
|
|
|
static size_t mi_os_numa_node_countx(void) {
|
2019-11-02 05:53:07 +03:00
|
|
|
ULONG numa_max = 0;
|
|
|
|
GetNumaHighestNodeNumber(&numa_max);
|
2019-11-14 00:35:50 +03:00
|
|
|
return (numa_max + 1);
|
2019-11-02 05:53:07 +03:00
|
|
|
}
|
2019-11-02 21:55:03 +03:00
|
|
|
#elif defined(__linux__)
|
2019-11-12 21:16:59 +03:00
|
|
|
#include <sys/syscall.h> // getcpu
|
|
|
|
#include <stdio.h> // access
|
2019-11-02 21:55:03 +03:00
|
|
|
|
2019-11-14 00:35:50 +03:00
|
|
|
static size_t mi_os_numa_nodex(void) {
|
2019-11-02 21:55:03 +03:00
|
|
|
#ifdef SYS_getcpu
|
2019-11-14 00:35:50 +03:00
|
|
|
unsigned long node = 0;
|
|
|
|
unsigned long ncpu = 0;
|
|
|
|
long err = syscall(SYS_getcpu, &ncpu, &node, NULL);
|
2019-11-02 08:01:52 +03:00
|
|
|
if (err != 0) return 0;
|
2019-11-14 00:35:50 +03:00
|
|
|
return node;
|
2019-11-02 21:55:03 +03:00
|
|
|
#else
|
|
|
|
return 0;
|
|
|
|
#endif
|
2019-11-02 05:53:07 +03:00
|
|
|
}
|
2019-11-14 00:35:50 +03:00
|
|
|
static size_t mi_os_numa_node_countx(void) {
|
2019-11-12 21:16:59 +03:00
|
|
|
char buf[128];
|
2019-11-14 00:35:50 +03:00
|
|
|
unsigned node = 0;
|
2019-11-21 01:55:12 +03:00
|
|
|
for(node = 0; node < 256; node++) {
|
2019-11-12 23:04:43 +03:00
|
|
|
// enumerate node entries -- todo: it there a more efficient way to do this? (but ensure there is no allocation)
|
2019-11-14 00:35:50 +03:00
|
|
|
snprintf(buf, 127, "/sys/devices/system/node/node%u", node + 1);
|
2019-11-21 01:55:12 +03:00
|
|
|
if (access(buf,R_OK) != 0) break;
|
2019-11-02 08:01:52 +03:00
|
|
|
}
|
2019-11-21 01:55:12 +03:00
|
|
|
return (node+1);
|
2019-11-02 05:53:07 +03:00
|
|
|
}
|
|
|
|
#else
|
2019-11-14 00:35:50 +03:00
|
|
|
static size_t mi_os_numa_nodex(void) {
|
2019-08-19 21:10:06 +03:00
|
|
|
return 0;
|
|
|
|
}
|
2019-11-14 00:35:50 +03:00
|
|
|
static size_t mi_os_numa_node_countx(void) {
|
2019-11-02 05:53:07 +03:00
|
|
|
return 1;
|
|
|
|
}
|
2019-08-19 21:10:06 +03:00
|
|
|
#endif
|
|
|
|
|
2019-11-14 00:35:50 +03:00
|
|
|
size_t _mi_numa_node_count = 0; // cache the node count
|
2019-11-12 21:16:59 +03:00
|
|
|
|
2019-11-14 00:35:50 +03:00
|
|
|
size_t _mi_os_numa_node_count_get(void) {
|
2019-11-12 21:16:59 +03:00
|
|
|
if (mi_unlikely(_mi_numa_node_count <= 0)) {
|
2019-11-14 00:35:50 +03:00
|
|
|
long ncount = mi_option_get(mi_option_use_numa_nodes); // given explicitly?
|
|
|
|
if (ncount <= 0) ncount = (long)mi_os_numa_node_countx(); // or detect dynamically
|
|
|
|
_mi_numa_node_count = (size_t)(ncount <= 0 ? 1 : ncount);
|
|
|
|
_mi_verbose_message("using %zd numa regions\n", _mi_numa_node_count);
|
2019-11-02 08:01:52 +03:00
|
|
|
}
|
2019-11-12 21:16:59 +03:00
|
|
|
mi_assert_internal(_mi_numa_node_count >= 1);
|
|
|
|
return _mi_numa_node_count;
|
2019-11-02 08:01:52 +03:00
|
|
|
}
|
|
|
|
|
2019-11-12 21:16:59 +03:00
|
|
|
int _mi_os_numa_node_get(mi_os_tld_t* tld) {
|
2019-11-02 20:30:16 +03:00
|
|
|
UNUSED(tld);
|
2019-11-14 00:35:50 +03:00
|
|
|
size_t numa_count = _mi_os_numa_node_count();
|
2019-11-21 01:55:12 +03:00
|
|
|
if (numa_count<=1) return 0; // optimize on single numa node systems: always node 0
|
2019-11-07 01:17:36 +03:00
|
|
|
// never more than the node count and >= 0
|
2019-11-14 00:35:50 +03:00
|
|
|
size_t numa_node = mi_os_numa_nodex();
|
2019-11-02 20:30:16 +03:00
|
|
|
if (numa_node >= numa_count) { numa_node = numa_node % numa_count; }
|
2019-11-14 00:35:50 +03:00
|
|
|
return (int)numa_node;
|
2019-11-02 05:53:07 +03:00
|
|
|
}
|