microsoft/qdk
Publicmirrored fromhttps://github.com/microsoft/qdkAvailable
allocator/mimalloc-sys/mimalloc/src/prim/windows/prim.c
622lines · modecode
| 1 | /* ---------------------------------------------------------------------------- |
| 2 | Copyright (c) 2018-2023, Microsoft Research, Daan Leijen |
| 3 | This is free software; you can redistribute it and/or modify it under the |
| 4 | terms of the MIT license. A copy of the license can be found in the file |
| 5 | "LICENSE" at the root of this distribution. |
| 6 | -----------------------------------------------------------------------------*/ |
| 7 | |
| 8 | // This file is included in `src/prim/prim.c` |
| 9 | |
| 10 | #include "mimalloc.h" |
| 11 | #include "mimalloc/internal.h" |
| 12 | #include "mimalloc/atomic.h" |
| 13 | #include "mimalloc/prim.h" |
| 14 | #include <stdio.h> // fputs, stderr |
| 15 | |
| 16 | |
| 17 | //--------------------------------------------- |
| 18 | // Dynamically bind Windows API points for portability |
| 19 | //--------------------------------------------- |
| 20 | |
| 21 | // We use VirtualAlloc2 for aligned allocation, but it is only supported on Windows 10 and Windows Server 2016. |
| 22 | // So, we need to look it up dynamically to run on older systems. (use __stdcall for 32-bit compatibility) |
| 23 | // NtAllocateVirtualAllocEx is used for huge OS page allocation (1GiB) |
| 24 | // We define a minimal MEM_EXTENDED_PARAMETER ourselves in order to be able to compile with older SDK's. |
| 25 | typedef enum MI_MEM_EXTENDED_PARAMETER_TYPE_E { |
| 26 | MiMemExtendedParameterInvalidType = 0, |
| 27 | MiMemExtendedParameterAddressRequirements, |
| 28 | MiMemExtendedParameterNumaNode, |
| 29 | MiMemExtendedParameterPartitionHandle, |
| 30 | MiMemExtendedParameterUserPhysicalHandle, |
| 31 | MiMemExtendedParameterAttributeFlags, |
| 32 | MiMemExtendedParameterMax |
| 33 | } MI_MEM_EXTENDED_PARAMETER_TYPE; |
| 34 | |
| 35 | typedef struct DECLSPEC_ALIGN(8) MI_MEM_EXTENDED_PARAMETER_S { |
| 36 | struct { DWORD64 Type : 8; DWORD64 Reserved : 56; } Type; |
| 37 | union { DWORD64 ULong64; PVOID Pointer; SIZE_T Size; HANDLE Handle; DWORD ULong; } Arg; |
| 38 | } MI_MEM_EXTENDED_PARAMETER; |
| 39 | |
| 40 | typedef struct MI_MEM_ADDRESS_REQUIREMENTS_S { |
| 41 | PVOID LowestStartingAddress; |
| 42 | PVOID HighestEndingAddress; |
| 43 | SIZE_T Alignment; |
| 44 | } MI_MEM_ADDRESS_REQUIREMENTS; |
| 45 | |
| 46 | #define MI_MEM_EXTENDED_PARAMETER_NONPAGED_HUGE 0x00000010 |
| 47 | |
| 48 | #include <winternl.h> |
| 49 | typedef PVOID (__stdcall *PVirtualAlloc2)(HANDLE, PVOID, SIZE_T, ULONG, ULONG, MI_MEM_EXTENDED_PARAMETER*, ULONG); |
| 50 | typedef NTSTATUS (__stdcall *PNtAllocateVirtualMemoryEx)(HANDLE, PVOID*, SIZE_T*, ULONG, ULONG, MI_MEM_EXTENDED_PARAMETER*, ULONG); |
| 51 | static PVirtualAlloc2 pVirtualAlloc2 = NULL; |
| 52 | static PNtAllocateVirtualMemoryEx pNtAllocateVirtualMemoryEx = NULL; |
| 53 | |
| 54 | // Similarly, GetNumaProcesorNodeEx is only supported since Windows 7 |
| 55 | typedef struct MI_PROCESSOR_NUMBER_S { WORD Group; BYTE Number; BYTE Reserved; } MI_PROCESSOR_NUMBER; |
| 56 | |
| 57 | typedef VOID (__stdcall *PGetCurrentProcessorNumberEx)(MI_PROCESSOR_NUMBER* ProcNumber); |
| 58 | typedef BOOL (__stdcall *PGetNumaProcessorNodeEx)(MI_PROCESSOR_NUMBER* Processor, PUSHORT NodeNumber); |
| 59 | typedef BOOL (__stdcall* PGetNumaNodeProcessorMaskEx)(USHORT Node, PGROUP_AFFINITY ProcessorMask); |
| 60 | typedef BOOL (__stdcall *PGetNumaProcessorNode)(UCHAR Processor, PUCHAR NodeNumber); |
| 61 | static PGetCurrentProcessorNumberEx pGetCurrentProcessorNumberEx = NULL; |
| 62 | static PGetNumaProcessorNodeEx pGetNumaProcessorNodeEx = NULL; |
| 63 | static PGetNumaNodeProcessorMaskEx pGetNumaNodeProcessorMaskEx = NULL; |
| 64 | static PGetNumaProcessorNode pGetNumaProcessorNode = NULL; |
| 65 | |
| 66 | //--------------------------------------------- |
| 67 | // Enable large page support dynamically (if possible) |
| 68 | //--------------------------------------------- |
| 69 | |
| 70 | static bool win_enable_large_os_pages(size_t* large_page_size) |
| 71 | { |
| 72 | static bool large_initialized = false; |
| 73 | if (large_initialized) return (_mi_os_large_page_size() > 0); |
| 74 | large_initialized = true; |
| 75 | |
| 76 | // Try to see if large OS pages are supported |
| 77 | // To use large pages on Windows, we first need access permission |
| 78 | // Set "Lock pages in memory" permission in the group policy editor |
| 79 | // <https://devblogs.microsoft.com/oldnewthing/20110128-00/?p=11643> |
| 80 | unsigned long err = 0; |
| 81 | HANDLE token = NULL; |
| 82 | BOOL ok = OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &token); |
| 83 | if (ok) { |
| 84 | TOKEN_PRIVILEGES tp; |
| 85 | ok = LookupPrivilegeValue(NULL, TEXT("SeLockMemoryPrivilege"), &tp.Privileges[0].Luid); |
| 86 | if (ok) { |
| 87 | tp.PrivilegeCount = 1; |
| 88 | tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED; |
| 89 | ok = AdjustTokenPrivileges(token, FALSE, &tp, 0, (PTOKEN_PRIVILEGES)NULL, 0); |
| 90 | if (ok) { |
| 91 | err = GetLastError(); |
| 92 | ok = (err == ERROR_SUCCESS); |
| 93 | if (ok && large_page_size != NULL) { |
| 94 | *large_page_size = GetLargePageMinimum(); |
| 95 | } |
| 96 | } |
| 97 | } |
| 98 | CloseHandle(token); |
| 99 | } |
| 100 | if (!ok) { |
| 101 | if (err == 0) err = GetLastError(); |
| 102 | _mi_warning_message("cannot enable large OS page support, error %lu\n", err); |
| 103 | } |
| 104 | return (ok!=0); |
| 105 | } |
| 106 | |
| 107 | |
| 108 | //--------------------------------------------- |
| 109 | // Initialize |
| 110 | //--------------------------------------------- |
| 111 | |
| 112 | void _mi_prim_mem_init( mi_os_mem_config_t* config ) |
| 113 | { |
| 114 | config->has_overcommit = false; |
| 115 | config->must_free_whole = true; |
| 116 | config->has_virtual_reserve = true; |
| 117 | // get the page size |
| 118 | SYSTEM_INFO si; |
| 119 | GetSystemInfo(&si); |
| 120 | if (si.dwPageSize > 0) { config->page_size = si.dwPageSize; } |
| 121 | if (si.dwAllocationGranularity > 0) { config->alloc_granularity = si.dwAllocationGranularity; } |
| 122 | // get the VirtualAlloc2 function |
| 123 | HINSTANCE hDll; |
| 124 | hDll = LoadLibrary(TEXT("kernelbase.dll")); |
| 125 | if (hDll != NULL) { |
| 126 | // use VirtualAlloc2FromApp if possible as it is available to Windows store apps |
| 127 | pVirtualAlloc2 = (PVirtualAlloc2)(void (*)(void))GetProcAddress(hDll, "VirtualAlloc2FromApp"); |
| 128 | if (pVirtualAlloc2==NULL) pVirtualAlloc2 = (PVirtualAlloc2)(void (*)(void))GetProcAddress(hDll, "VirtualAlloc2"); |
| 129 | FreeLibrary(hDll); |
| 130 | } |
| 131 | // NtAllocateVirtualMemoryEx is used for huge page allocation |
| 132 | hDll = LoadLibrary(TEXT("ntdll.dll")); |
| 133 | if (hDll != NULL) { |
| 134 | pNtAllocateVirtualMemoryEx = (PNtAllocateVirtualMemoryEx)(void (*)(void))GetProcAddress(hDll, "NtAllocateVirtualMemoryEx"); |
| 135 | FreeLibrary(hDll); |
| 136 | } |
| 137 | // Try to use Win7+ numa API |
| 138 | hDll = LoadLibrary(TEXT("kernel32.dll")); |
| 139 | if (hDll != NULL) { |
| 140 | pGetCurrentProcessorNumberEx = (PGetCurrentProcessorNumberEx)(void (*)(void))GetProcAddress(hDll, "GetCurrentProcessorNumberEx"); |
| 141 | pGetNumaProcessorNodeEx = (PGetNumaProcessorNodeEx)(void (*)(void))GetProcAddress(hDll, "GetNumaProcessorNodeEx"); |
| 142 | pGetNumaNodeProcessorMaskEx = (PGetNumaNodeProcessorMaskEx)(void (*)(void))GetProcAddress(hDll, "GetNumaNodeProcessorMaskEx"); |
| 143 | pGetNumaProcessorNode = (PGetNumaProcessorNode)(void (*)(void))GetProcAddress(hDll, "GetNumaProcessorNode"); |
| 144 | FreeLibrary(hDll); |
| 145 | } |
| 146 | if (mi_option_is_enabled(mi_option_allow_large_os_pages) || mi_option_is_enabled(mi_option_reserve_huge_os_pages)) { |
| 147 | win_enable_large_os_pages(&config->large_page_size); |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | |
| 152 | //--------------------------------------------- |
| 153 | // Free |
| 154 | //--------------------------------------------- |
| 155 | |
| 156 | int _mi_prim_free(void* addr, size_t size ) { |
| 157 | MI_UNUSED(size); |
| 158 | DWORD errcode = 0; |
| 159 | bool err = (VirtualFree(addr, 0, MEM_RELEASE) == 0); |
| 160 | if (err) { errcode = GetLastError(); } |
| 161 | if (errcode == ERROR_INVALID_ADDRESS) { |
| 162 | // In mi_os_mem_alloc_aligned the fallback path may have returned a pointer inside |
| 163 | // the memory region returned by VirtualAlloc; in that case we need to free using |
| 164 | // the start of the region. |
| 165 | MEMORY_BASIC_INFORMATION info = { 0 }; |
| 166 | VirtualQuery(addr, &info, sizeof(info)); |
| 167 | if (info.AllocationBase < addr && ((uint8_t*)addr - (uint8_t*)info.AllocationBase) < (ptrdiff_t)MI_SEGMENT_SIZE) { |
| 168 | errcode = 0; |
| 169 | err = (VirtualFree(info.AllocationBase, 0, MEM_RELEASE) == 0); |
| 170 | if (err) { errcode = GetLastError(); } |
| 171 | } |
| 172 | } |
| 173 | return (int)errcode; |
| 174 | } |
| 175 | |
| 176 | |
| 177 | //--------------------------------------------- |
| 178 | // VirtualAlloc |
| 179 | //--------------------------------------------- |
| 180 | |
| 181 | static void* win_virtual_alloc_prim(void* addr, size_t size, size_t try_alignment, DWORD flags) { |
| 182 | #if (MI_INTPTR_SIZE >= 8) |
| 183 | // on 64-bit systems, try to use the virtual address area after 2TiB for 4MiB aligned allocations |
| 184 | if (addr == NULL) { |
| 185 | void* hint = _mi_os_get_aligned_hint(try_alignment,size); |
| 186 | if (hint != NULL) { |
| 187 | void* p = VirtualAlloc(hint, size, flags, PAGE_READWRITE); |
| 188 | if (p != NULL) return p; |
| 189 | _mi_verbose_message("warning: unable to allocate hinted aligned OS memory (%zu bytes, error code: 0x%x, address: %p, alignment: %zu, flags: 0x%x)\n", size, GetLastError(), hint, try_alignment, flags); |
| 190 | // fall through on error |
| 191 | } |
| 192 | } |
| 193 | #endif |
| 194 | // on modern Windows try use VirtualAlloc2 for aligned allocation |
| 195 | if (try_alignment > 1 && (try_alignment % _mi_os_page_size()) == 0 && pVirtualAlloc2 != NULL) { |
| 196 | MI_MEM_ADDRESS_REQUIREMENTS reqs = { 0, 0, 0 }; |
| 197 | reqs.Alignment = try_alignment; |
| 198 | MI_MEM_EXTENDED_PARAMETER param = { {0, 0}, {0} }; |
| 199 | param.Type.Type = MiMemExtendedParameterAddressRequirements; |
| 200 | param.Arg.Pointer = &reqs; |
| 201 | void* p = (*pVirtualAlloc2)(GetCurrentProcess(), addr, size, flags, PAGE_READWRITE, ¶m, 1); |
| 202 | if (p != NULL) return p; |
| 203 | _mi_warning_message("unable to allocate aligned OS memory (%zu bytes, error code: 0x%x, address: %p, alignment: %zu, flags: 0x%x)\n", size, GetLastError(), addr, try_alignment, flags); |
| 204 | // fall through on error |
| 205 | } |
| 206 | // last resort |
| 207 | return VirtualAlloc(addr, size, flags, PAGE_READWRITE); |
| 208 | } |
| 209 | |
| 210 | static void* win_virtual_alloc(void* addr, size_t size, size_t try_alignment, DWORD flags, bool large_only, bool allow_large, bool* is_large) { |
| 211 | mi_assert_internal(!(large_only && !allow_large)); |
| 212 | static _Atomic(size_t) large_page_try_ok; // = 0; |
| 213 | void* p = NULL; |
| 214 | // Try to allocate large OS pages (2MiB) if allowed or required. |
| 215 | if ((large_only || _mi_os_use_large_page(size, try_alignment)) |
| 216 | && allow_large && (flags&MEM_COMMIT)!=0 && (flags&MEM_RESERVE)!=0) { |
| 217 | size_t try_ok = mi_atomic_load_acquire(&large_page_try_ok); |
| 218 | if (!large_only && try_ok > 0) { |
| 219 | // if a large page allocation fails, it seems the calls to VirtualAlloc get very expensive. |
| 220 | // therefore, once a large page allocation failed, we don't try again for `large_page_try_ok` times. |
| 221 | mi_atomic_cas_strong_acq_rel(&large_page_try_ok, &try_ok, try_ok - 1); |
| 222 | } |
| 223 | else { |
| 224 | // large OS pages must always reserve and commit. |
| 225 | *is_large = true; |
| 226 | p = win_virtual_alloc_prim(addr, size, try_alignment, flags | MEM_LARGE_PAGES); |
| 227 | if (large_only) return p; |
| 228 | // fall back to non-large page allocation on error (`p == NULL`). |
| 229 | if (p == NULL) { |
| 230 | mi_atomic_store_release(&large_page_try_ok,10UL); // on error, don't try again for the next N allocations |
| 231 | } |
| 232 | } |
| 233 | } |
| 234 | // Fall back to regular page allocation |
| 235 | if (p == NULL) { |
| 236 | *is_large = ((flags&MEM_LARGE_PAGES) != 0); |
| 237 | p = win_virtual_alloc_prim(addr, size, try_alignment, flags); |
| 238 | } |
| 239 | //if (p == NULL) { _mi_warning_message("unable to allocate OS memory (%zu bytes, error code: 0x%x, address: %p, alignment: %zu, flags: 0x%x, large only: %d, allow large: %d)\n", size, GetLastError(), addr, try_alignment, flags, large_only, allow_large); } |
| 240 | return p; |
| 241 | } |
| 242 | |
| 243 | int _mi_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, void** addr) { |
| 244 | mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0); |
| 245 | mi_assert_internal(commit || !allow_large); |
| 246 | mi_assert_internal(try_alignment > 0); |
| 247 | *is_zero = true; |
| 248 | int flags = MEM_RESERVE; |
| 249 | if (commit) { flags |= MEM_COMMIT; } |
| 250 | *addr = win_virtual_alloc(NULL, size, try_alignment, flags, false, allow_large, is_large); |
| 251 | return (*addr != NULL ? 0 : (int)GetLastError()); |
| 252 | } |
| 253 | |
| 254 | |
| 255 | //--------------------------------------------- |
| 256 | // Commit/Reset/Protect |
| 257 | //--------------------------------------------- |
| 258 | #ifdef _MSC_VER |
| 259 | #pragma warning(disable:6250) // suppress warning calling VirtualFree without MEM_RELEASE (for decommit) |
| 260 | #endif |
| 261 | |
| 262 | int _mi_prim_commit(void* addr, size_t size, bool* is_zero) { |
| 263 | *is_zero = false; |
| 264 | /* |
| 265 | // zero'ing only happens on an initial commit... but checking upfront seems expensive.. |
| 266 | _MEMORY_BASIC_INFORMATION meminfo; _mi_memzero_var(meminfo); |
| 267 | if (VirtualQuery(addr, &meminfo, size) > 0) { |
| 268 | if ((meminfo.State & MEM_COMMIT) == 0) { |
| 269 | *is_zero = true; |
| 270 | } |
| 271 | } |
| 272 | */ |
| 273 | // commit |
| 274 | void* p = VirtualAlloc(addr, size, MEM_COMMIT, PAGE_READWRITE); |
| 275 | if (p == NULL) return (int)GetLastError(); |
| 276 | return 0; |
| 277 | } |
| 278 | |
| 279 | int _mi_prim_decommit(void* addr, size_t size, bool* needs_recommit) { |
| 280 | BOOL ok = VirtualFree(addr, size, MEM_DECOMMIT); |
| 281 | *needs_recommit = true; // for safety, assume always decommitted even in the case of an error. |
| 282 | return (ok ? 0 : (int)GetLastError()); |
| 283 | } |
| 284 | |
| 285 | int _mi_prim_reset(void* addr, size_t size) { |
| 286 | void* p = VirtualAlloc(addr, size, MEM_RESET, PAGE_READWRITE); |
| 287 | mi_assert_internal(p == addr); |
| 288 | #if 0 |
| 289 | if (p != NULL) { |
| 290 | VirtualUnlock(addr,size); // VirtualUnlock after MEM_RESET removes the memory directly from the working set |
| 291 | } |
| 292 | #endif |
| 293 | return (p != NULL ? 0 : (int)GetLastError()); |
| 294 | } |
| 295 | |
| 296 | int _mi_prim_protect(void* addr, size_t size, bool protect) { |
| 297 | DWORD oldprotect = 0; |
| 298 | BOOL ok = VirtualProtect(addr, size, protect ? PAGE_NOACCESS : PAGE_READWRITE, &oldprotect); |
| 299 | return (ok ? 0 : (int)GetLastError()); |
| 300 | } |
| 301 | |
| 302 | |
| 303 | //--------------------------------------------- |
| 304 | // Huge page allocation |
| 305 | //--------------------------------------------- |
| 306 | |
| 307 | static void* _mi_prim_alloc_huge_os_pagesx(void* hint_addr, size_t size, int numa_node) |
| 308 | { |
| 309 | const DWORD flags = MEM_LARGE_PAGES | MEM_COMMIT | MEM_RESERVE; |
| 310 | |
| 311 | win_enable_large_os_pages(NULL); |
| 312 | |
| 313 | MI_MEM_EXTENDED_PARAMETER params[3] = { {{0,0},{0}},{{0,0},{0}},{{0,0},{0}} }; |
| 314 | // on modern Windows try use NtAllocateVirtualMemoryEx for 1GiB huge pages |
| 315 | static bool mi_huge_pages_available = true; |
| 316 | if (pNtAllocateVirtualMemoryEx != NULL && mi_huge_pages_available) { |
| 317 | params[0].Type.Type = MiMemExtendedParameterAttributeFlags; |
| 318 | params[0].Arg.ULong64 = MI_MEM_EXTENDED_PARAMETER_NONPAGED_HUGE; |
| 319 | ULONG param_count = 1; |
| 320 | if (numa_node >= 0) { |
| 321 | param_count++; |
| 322 | params[1].Type.Type = MiMemExtendedParameterNumaNode; |
| 323 | params[1].Arg.ULong = (unsigned)numa_node; |
| 324 | } |
| 325 | SIZE_T psize = size; |
| 326 | void* base = hint_addr; |
| 327 | NTSTATUS err = (*pNtAllocateVirtualMemoryEx)(GetCurrentProcess(), &base, &psize, flags, PAGE_READWRITE, params, param_count); |
| 328 | if (err == 0 && base != NULL) { |
| 329 | return base; |
| 330 | } |
| 331 | else { |
| 332 | // fall back to regular large pages |
| 333 | mi_huge_pages_available = false; // don't try further huge pages |
| 334 | _mi_warning_message("unable to allocate using huge (1GiB) pages, trying large (2MiB) pages instead (status 0x%lx)\n", err); |
| 335 | } |
| 336 | } |
| 337 | // on modern Windows try use VirtualAlloc2 for numa aware large OS page allocation |
| 338 | if (pVirtualAlloc2 != NULL && numa_node >= 0) { |
| 339 | params[0].Type.Type = MiMemExtendedParameterNumaNode; |
| 340 | params[0].Arg.ULong = (unsigned)numa_node; |
| 341 | return (*pVirtualAlloc2)(GetCurrentProcess(), hint_addr, size, flags, PAGE_READWRITE, params, 1); |
| 342 | } |
| 343 | |
| 344 | // otherwise use regular virtual alloc on older windows |
| 345 | return VirtualAlloc(hint_addr, size, flags, PAGE_READWRITE); |
| 346 | } |
| 347 | |
| 348 | int _mi_prim_alloc_huge_os_pages(void* hint_addr, size_t size, int numa_node, bool* is_zero, void** addr) { |
| 349 | *is_zero = true; |
| 350 | *addr = _mi_prim_alloc_huge_os_pagesx(hint_addr,size,numa_node); |
| 351 | return (*addr != NULL ? 0 : (int)GetLastError()); |
| 352 | } |
| 353 | |
| 354 | |
| 355 | //--------------------------------------------- |
| 356 | // Numa nodes |
| 357 | //--------------------------------------------- |
| 358 | |
| 359 | size_t _mi_prim_numa_node(void) { |
| 360 | USHORT numa_node = 0; |
| 361 | if (pGetCurrentProcessorNumberEx != NULL && pGetNumaProcessorNodeEx != NULL) { |
| 362 | // Extended API is supported |
| 363 | MI_PROCESSOR_NUMBER pnum; |
| 364 | (*pGetCurrentProcessorNumberEx)(&pnum); |
| 365 | USHORT nnode = 0; |
| 366 | BOOL ok = (*pGetNumaProcessorNodeEx)(&pnum, &nnode); |
| 367 | if (ok) { numa_node = nnode; } |
| 368 | } |
| 369 | else if (pGetNumaProcessorNode != NULL) { |
| 370 | // Vista or earlier, use older API that is limited to 64 processors. Issue #277 |
| 371 | DWORD pnum = GetCurrentProcessorNumber(); |
| 372 | UCHAR nnode = 0; |
| 373 | BOOL ok = pGetNumaProcessorNode((UCHAR)pnum, &nnode); |
| 374 | if (ok) { numa_node = nnode; } |
| 375 | } |
| 376 | return numa_node; |
| 377 | } |
| 378 | |
| 379 | size_t _mi_prim_numa_node_count(void) { |
| 380 | ULONG numa_max = 0; |
| 381 | GetNumaHighestNodeNumber(&numa_max); |
| 382 | // find the highest node number that has actual processors assigned to it. Issue #282 |
| 383 | while(numa_max > 0) { |
| 384 | if (pGetNumaNodeProcessorMaskEx != NULL) { |
| 385 | // Extended API is supported |
| 386 | GROUP_AFFINITY affinity; |
| 387 | if ((*pGetNumaNodeProcessorMaskEx)((USHORT)numa_max, &affinity)) { |
| 388 | if (affinity.Mask != 0) break; // found the maximum non-empty node |
| 389 | } |
| 390 | } |
| 391 | else { |
| 392 | // Vista or earlier, use older API that is limited to 64 processors. |
| 393 | ULONGLONG mask; |
| 394 | if (GetNumaNodeProcessorMask((UCHAR)numa_max, &mask)) { |
| 395 | if (mask != 0) break; // found the maximum non-empty node |
| 396 | }; |
| 397 | } |
| 398 | // max node was invalid or had no processor assigned, try again |
| 399 | numa_max--; |
| 400 | } |
| 401 | return ((size_t)numa_max + 1); |
| 402 | } |
| 403 | |
| 404 | |
| 405 | //---------------------------------------------------------------- |
| 406 | // Clock |
| 407 | //---------------------------------------------------------------- |
| 408 | |
| 409 | static mi_msecs_t mi_to_msecs(LARGE_INTEGER t) { |
| 410 | static LARGE_INTEGER mfreq; // = 0 |
| 411 | if (mfreq.QuadPart == 0LL) { |
| 412 | LARGE_INTEGER f; |
| 413 | QueryPerformanceFrequency(&f); |
| 414 | mfreq.QuadPart = f.QuadPart/1000LL; |
| 415 | if (mfreq.QuadPart == 0) mfreq.QuadPart = 1; |
| 416 | } |
| 417 | return (mi_msecs_t)(t.QuadPart / mfreq.QuadPart); |
| 418 | } |
| 419 | |
| 420 | mi_msecs_t _mi_prim_clock_now(void) { |
| 421 | LARGE_INTEGER t; |
| 422 | QueryPerformanceCounter(&t); |
| 423 | return mi_to_msecs(t); |
| 424 | } |
| 425 | |
| 426 | |
| 427 | //---------------------------------------------------------------- |
| 428 | // Process Info |
| 429 | //---------------------------------------------------------------- |
| 430 | |
| 431 | #include <windows.h> |
| 432 | #include <psapi.h> |
| 433 | |
| 434 | static mi_msecs_t filetime_msecs(const FILETIME* ftime) { |
| 435 | ULARGE_INTEGER i; |
| 436 | i.LowPart = ftime->dwLowDateTime; |
| 437 | i.HighPart = ftime->dwHighDateTime; |
| 438 | mi_msecs_t msecs = (i.QuadPart / 10000); // FILETIME is in 100 nano seconds |
| 439 | return msecs; |
| 440 | } |
| 441 | |
| 442 | typedef BOOL (WINAPI *PGetProcessMemoryInfo)(HANDLE, PPROCESS_MEMORY_COUNTERS, DWORD); |
| 443 | static PGetProcessMemoryInfo pGetProcessMemoryInfo = NULL; |
| 444 | |
| 445 | void _mi_prim_process_info(mi_process_info_t* pinfo) |
| 446 | { |
| 447 | FILETIME ct; |
| 448 | FILETIME ut; |
| 449 | FILETIME st; |
| 450 | FILETIME et; |
| 451 | GetProcessTimes(GetCurrentProcess(), &ct, &et, &st, &ut); |
| 452 | pinfo->utime = filetime_msecs(&ut); |
| 453 | pinfo->stime = filetime_msecs(&st); |
| 454 | |
| 455 | // load psapi on demand |
| 456 | if (pGetProcessMemoryInfo == NULL) { |
| 457 | HINSTANCE hDll = LoadLibrary(TEXT("psapi.dll")); |
| 458 | if (hDll != NULL) { |
| 459 | pGetProcessMemoryInfo = (PGetProcessMemoryInfo)(void (*)(void))GetProcAddress(hDll, "GetProcessMemoryInfo"); |
| 460 | } |
| 461 | } |
| 462 | |
| 463 | // get process info |
| 464 | PROCESS_MEMORY_COUNTERS info; |
| 465 | memset(&info, 0, sizeof(info)); |
| 466 | if (pGetProcessMemoryInfo != NULL) { |
| 467 | pGetProcessMemoryInfo(GetCurrentProcess(), &info, sizeof(info)); |
| 468 | } |
| 469 | pinfo->current_rss = (size_t)info.WorkingSetSize; |
| 470 | pinfo->peak_rss = (size_t)info.PeakWorkingSetSize; |
| 471 | pinfo->current_commit = (size_t)info.PagefileUsage; |
| 472 | pinfo->peak_commit = (size_t)info.PeakPagefileUsage; |
| 473 | pinfo->page_faults = (size_t)info.PageFaultCount; |
| 474 | } |
| 475 | |
| 476 | //---------------------------------------------------------------- |
| 477 | // Output |
| 478 | //---------------------------------------------------------------- |
| 479 | |
| 480 | void _mi_prim_out_stderr( const char* msg ) |
| 481 | { |
| 482 | // on windows with redirection, the C runtime cannot handle locale dependent output |
| 483 | // after the main thread closes so we use direct console output. |
| 484 | if (!_mi_preloading()) { |
| 485 | // _cputs(msg); // _cputs cannot be used at is aborts if it fails to lock the console |
| 486 | static HANDLE hcon = INVALID_HANDLE_VALUE; |
| 487 | static bool hconIsConsole; |
| 488 | if (hcon == INVALID_HANDLE_VALUE) { |
| 489 | CONSOLE_SCREEN_BUFFER_INFO sbi; |
| 490 | hcon = GetStdHandle(STD_ERROR_HANDLE); |
| 491 | hconIsConsole = ((hcon != INVALID_HANDLE_VALUE) && GetConsoleScreenBufferInfo(hcon, &sbi)); |
| 492 | } |
| 493 | const size_t len = _mi_strlen(msg); |
| 494 | if (len > 0 && len < UINT32_MAX) { |
| 495 | DWORD written = 0; |
| 496 | if (hconIsConsole) { |
| 497 | WriteConsoleA(hcon, msg, (DWORD)len, &written, NULL); |
| 498 | } |
| 499 | else if (hcon != INVALID_HANDLE_VALUE) { |
| 500 | // use direct write if stderr was redirected |
| 501 | WriteFile(hcon, msg, (DWORD)len, &written, NULL); |
| 502 | } |
| 503 | else { |
| 504 | // finally fall back to fputs after all |
| 505 | fputs(msg, stderr); |
| 506 | } |
| 507 | } |
| 508 | } |
| 509 | } |
| 510 | |
| 511 | |
| 512 | //---------------------------------------------------------------- |
| 513 | // Environment |
| 514 | //---------------------------------------------------------------- |
| 515 | |
| 516 | // On Windows use GetEnvironmentVariable instead of getenv to work |
| 517 | // reliably even when this is invoked before the C runtime is initialized. |
| 518 | // i.e. when `_mi_preloading() == true`. |
| 519 | // Note: on windows, environment names are not case sensitive. |
| 520 | bool _mi_prim_getenv(const char* name, char* result, size_t result_size) { |
| 521 | result[0] = 0; |
| 522 | size_t len = GetEnvironmentVariableA(name, result, (DWORD)result_size); |
| 523 | return (len > 0 && len < result_size); |
| 524 | } |
| 525 | |
| 526 | |
| 527 | |
| 528 | //---------------------------------------------------------------- |
| 529 | // Random |
| 530 | //---------------------------------------------------------------- |
| 531 | |
| 532 | #if defined(MI_USE_RTLGENRANDOM) // || defined(__cplusplus) |
| 533 | // We prefer to use BCryptGenRandom instead of (the unofficial) RtlGenRandom but when using |
| 534 | // dynamic overriding, we observed it can raise an exception when compiled with C++, and |
| 535 | // sometimes deadlocks when also running under the VS debugger. |
| 536 | // In contrast, issue #623 implies that on Windows Server 2019 we need to use BCryptGenRandom. |
| 537 | // To be continued.. |
| 538 | #pragma comment (lib,"advapi32.lib") |
| 539 | #define RtlGenRandom SystemFunction036 |
| 540 | mi_decl_externc BOOLEAN NTAPI RtlGenRandom(PVOID RandomBuffer, ULONG RandomBufferLength); |
| 541 | |
| 542 | bool _mi_prim_random_buf(void* buf, size_t buf_len) { |
| 543 | return (RtlGenRandom(buf, (ULONG)buf_len) != 0); |
| 544 | } |
| 545 | |
| 546 | #else |
| 547 | |
| 548 | #ifndef BCRYPT_USE_SYSTEM_PREFERRED_RNG |
| 549 | #define BCRYPT_USE_SYSTEM_PREFERRED_RNG 0x00000002 |
| 550 | #endif |
| 551 | |
| 552 | typedef LONG (NTAPI *PBCryptGenRandom)(HANDLE, PUCHAR, ULONG, ULONG); |
| 553 | static PBCryptGenRandom pBCryptGenRandom = NULL; |
| 554 | |
| 555 | bool _mi_prim_random_buf(void* buf, size_t buf_len) { |
| 556 | if (pBCryptGenRandom == NULL) { |
| 557 | HINSTANCE hDll = LoadLibrary(TEXT("bcrypt.dll")); |
| 558 | if (hDll != NULL) { |
| 559 | pBCryptGenRandom = (PBCryptGenRandom)(void (*)(void))GetProcAddress(hDll, "BCryptGenRandom"); |
| 560 | } |
| 561 | if (pBCryptGenRandom == NULL) return false; |
| 562 | } |
| 563 | return (pBCryptGenRandom(NULL, (PUCHAR)buf, (ULONG)buf_len, BCRYPT_USE_SYSTEM_PREFERRED_RNG) >= 0); |
| 564 | } |
| 565 | |
| 566 | #endif // MI_USE_RTLGENRANDOM |
| 567 | |
| 568 | //---------------------------------------------------------------- |
| 569 | // Thread init/done |
| 570 | //---------------------------------------------------------------- |
| 571 | |
| 572 | #if !defined(MI_SHARED_LIB) |
| 573 | |
| 574 | // use thread local storage keys to detect thread ending |
| 575 | #include <fibersapi.h> |
| 576 | #if (_WIN32_WINNT < 0x600) // before Windows Vista |
| 577 | WINBASEAPI DWORD WINAPI FlsAlloc( _In_opt_ PFLS_CALLBACK_FUNCTION lpCallback ); |
| 578 | WINBASEAPI PVOID WINAPI FlsGetValue( _In_ DWORD dwFlsIndex ); |
| 579 | WINBASEAPI BOOL WINAPI FlsSetValue( _In_ DWORD dwFlsIndex, _In_opt_ PVOID lpFlsData ); |
| 580 | WINBASEAPI BOOL WINAPI FlsFree(_In_ DWORD dwFlsIndex); |
| 581 | #endif |
| 582 | |
| 583 | static DWORD mi_fls_key = (DWORD)(-1); |
| 584 | |
| 585 | static void NTAPI mi_fls_done(PVOID value) { |
| 586 | mi_heap_t* heap = (mi_heap_t*)value; |
| 587 | if (heap != NULL) { |
| 588 | _mi_thread_done(heap); |
| 589 | FlsSetValue(mi_fls_key, NULL); // prevent recursion as _mi_thread_done may set it back to the main heap, issue #672 |
| 590 | } |
| 591 | } |
| 592 | |
| 593 | void _mi_prim_thread_init_auto_done(void) { |
| 594 | mi_fls_key = FlsAlloc(&mi_fls_done); |
| 595 | } |
| 596 | |
| 597 | void _mi_prim_thread_done_auto_done(void) { |
| 598 | // call thread-done on all threads (except the main thread) to prevent |
| 599 | // dangling callback pointer if statically linked with a DLL; Issue #208 |
| 600 | FlsFree(mi_fls_key); |
| 601 | } |
| 602 | |
| 603 | void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) { |
| 604 | mi_assert_internal(mi_fls_key != (DWORD)(-1)); |
| 605 | FlsSetValue(mi_fls_key, heap); |
| 606 | } |
| 607 | |
| 608 | #else |
| 609 | |
| 610 | // Dll; nothing to do as in that case thread_done is handled through the DLL_THREAD_DETACH event. |
| 611 | |
| 612 | void _mi_prim_thread_init_auto_done(void) { |
| 613 | } |
| 614 | |
| 615 | void _mi_prim_thread_done_auto_done(void) { |
| 616 | } |
| 617 | |
| 618 | void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) { |
| 619 | MI_UNUSED(heap); |
| 620 | } |
| 621 | |
| 622 | #endif |
| 623 | |