| 1 | // Copyright 2009-2021 Intel Corporation |
| 2 | // SPDX-License-Identifier: Apache-2.0 |
| 3 | |
| 4 | #include "alloc.h" |
| 5 | #include "intrinsics.h" |
| 6 | #include "sysinfo.h" |
| 7 | #include "mutex.h" |
| 8 | |
| 9 | //////////////////////////////////////////////////////////////////////////////// |
| 10 | /// All Platforms |
| 11 | //////////////////////////////////////////////////////////////////////////////// |
| 12 | |
| 13 | namespace embree |
| 14 | { |
| 15 | void* alignedMalloc(size_t size, size_t align) |
| 16 | { |
| 17 | if (size == 0) |
| 18 | return nullptr; |
| 19 | |
| 20 | assert((align & (align-1)) == 0); |
| 21 | void* ptr = _mm_malloc(size,align); |
| 22 | |
| 23 | if (size != 0 && ptr == nullptr) |
| 24 | // -- GODOT start -- |
| 25 | // throw std::bad_alloc(); |
| 26 | abort(); |
| 27 | // -- GODOT end -- |
| 28 | |
| 29 | return ptr; |
| 30 | } |
| 31 | |
| 32 | void alignedFree(void* ptr) |
| 33 | { |
| 34 | if (ptr) |
| 35 | _mm_free(ptr); |
| 36 | } |
| 37 | |
| 38 | static bool huge_pages_enabled = false; |
| 39 | static MutexSys os_init_mutex; |
| 40 | |
| 41 | __forceinline bool isHugePageCandidate(const size_t bytes) |
| 42 | { |
| 43 | if (!huge_pages_enabled) |
| 44 | return false; |
| 45 | |
| 46 | /* use huge pages only when memory overhead is low */ |
| 47 | const size_t hbytes = (bytes+PAGE_SIZE_2M-1) & ~size_t(PAGE_SIZE_2M-1); |
| 48 | return 66*(hbytes-bytes) < bytes; // at most 1.5% overhead |
| 49 | } |
| 50 | } |
| 51 | |
| 52 | //////////////////////////////////////////////////////////////////////////////// |
| 53 | /// Windows Platform |
| 54 | //////////////////////////////////////////////////////////////////////////////// |
| 55 | |
| 56 | #ifdef _WIN32 |
| 57 | |
| 58 | #define WIN32_LEAN_AND_MEAN |
| 59 | #include <windows.h> |
| 60 | #include <malloc.h> |
| 61 | |
| 62 | namespace embree |
| 63 | { |
| 64 | bool win_enable_selockmemoryprivilege (bool verbose) |
| 65 | { |
| 66 | HANDLE hToken; |
| 67 | if (!OpenProcessToken(GetCurrentProcess(), TOKEN_QUERY | TOKEN_ADJUST_PRIVILEGES, &hToken)) { |
| 68 | if (verbose) std::cout << "WARNING: OpenProcessToken failed while trying to enable SeLockMemoryPrivilege: " << GetLastError() << std::endl; |
| 69 | return false; |
| 70 | } |
| 71 | |
| 72 | TOKEN_PRIVILEGES tp; |
| 73 | tp.PrivilegeCount = 1; |
| 74 | tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED; |
| 75 | |
| 76 | if (!LookupPrivilegeValueW(nullptr, L"SeLockMemoryPrivilege" , &tp.Privileges[0].Luid)) { |
| 77 | if (verbose) std::cout << "WARNING: LookupPrivilegeValue failed while trying to enable SeLockMemoryPrivilege: " << GetLastError() << std::endl; |
| 78 | return false; |
| 79 | } |
| 80 | |
| 81 | SetLastError(ERROR_SUCCESS); |
| 82 | if (!AdjustTokenPrivileges(hToken, FALSE, &tp, sizeof(tp), nullptr, 0)) { |
| 83 | if (verbose) std::cout << "WARNING: AdjustTokenPrivileges failed while trying to enable SeLockMemoryPrivilege" << std::endl; |
| 84 | return false; |
| 85 | } |
| 86 | |
| 87 | if (GetLastError() == ERROR_NOT_ALL_ASSIGNED) { |
| 88 | if (verbose) std::cout << "WARNING: AdjustTokenPrivileges failed to enable SeLockMemoryPrivilege: Add SeLockMemoryPrivilege for current user and run process in elevated mode (Run as administrator)." << std::endl; |
| 89 | return false; |
| 90 | } |
| 91 | |
| 92 | return true; |
| 93 | } |
| 94 | |
| 95 | bool os_init(bool hugepages, bool verbose) |
| 96 | { |
| 97 | Lock<MutexSys> lock(os_init_mutex); |
| 98 | |
| 99 | if (!hugepages) { |
| 100 | huge_pages_enabled = false; |
| 101 | return true; |
| 102 | } |
| 103 | |
| 104 | if (GetLargePageMinimum() != PAGE_SIZE_2M) { |
| 105 | huge_pages_enabled = false; |
| 106 | return false; |
| 107 | } |
| 108 | |
| 109 | huge_pages_enabled = true; |
| 110 | return true; |
| 111 | } |
| 112 | |
| 113 | void* os_malloc(size_t bytes, bool& hugepages) |
| 114 | { |
| 115 | if (bytes == 0) { |
| 116 | hugepages = false; |
| 117 | return nullptr; |
| 118 | } |
| 119 | |
| 120 | /* try direct huge page allocation first */ |
| 121 | if (isHugePageCandidate(bytes)) |
| 122 | { |
| 123 | int flags = MEM_COMMIT | MEM_RESERVE | MEM_LARGE_PAGES; |
| 124 | char* ptr = (char*) VirtualAlloc(nullptr,bytes,flags,PAGE_READWRITE); |
| 125 | if (ptr != nullptr) { |
| 126 | hugepages = true; |
| 127 | return ptr; |
| 128 | } |
| 129 | } |
| 130 | |
| 131 | /* fall back to 4k pages */ |
| 132 | int flags = MEM_COMMIT | MEM_RESERVE; |
| 133 | char* ptr = (char*) VirtualAlloc(nullptr,bytes,flags,PAGE_READWRITE); |
| 134 | // -- GODOT start -- |
| 135 | // if (ptr == nullptr) throw std::bad_alloc(); |
| 136 | if (ptr == nullptr) abort(); |
| 137 | // -- GODOT end -- |
| 138 | hugepages = false; |
| 139 | return ptr; |
| 140 | } |
| 141 | |
| 142 | size_t os_shrink(void* ptr, size_t bytesNew, size_t bytesOld, bool hugepages) |
| 143 | { |
| 144 | if (hugepages) // decommitting huge pages seems not to work under Windows |
| 145 | return bytesOld; |
| 146 | |
| 147 | const size_t pageSize = hugepages ? PAGE_SIZE_2M : PAGE_SIZE_4K; |
| 148 | bytesNew = (bytesNew+pageSize-1) & ~(pageSize-1); |
| 149 | bytesOld = (bytesOld+pageSize-1) & ~(pageSize-1); |
| 150 | if (bytesNew >= bytesOld) |
| 151 | return bytesOld; |
| 152 | |
| 153 | if (!VirtualFree((char*)ptr+bytesNew,bytesOld-bytesNew,MEM_DECOMMIT)) |
| 154 | // -- GODOT start -- |
| 155 | // throw std::bad_alloc(); |
| 156 | abort(); |
| 157 | // -- GODOT end -- |
| 158 | |
| 159 | return bytesNew; |
| 160 | } |
| 161 | |
| 162 | void os_free(void* ptr, size_t bytes, bool hugepages) |
| 163 | { |
| 164 | if (bytes == 0) |
| 165 | return; |
| 166 | |
| 167 | if (!VirtualFree(ptr,0,MEM_RELEASE)) |
| 168 | // -- GODOT start -- |
| 169 | // throw std::bad_alloc(); |
| 170 | abort(); |
| 171 | // -- GODOT end -- |
| 172 | } |
| 173 | |
| 174 | void os_advise(void *ptr, size_t bytes) |
| 175 | { |
| 176 | } |
| 177 | } |
| 178 | |
| 179 | #endif |
| 180 | |
| 181 | //////////////////////////////////////////////////////////////////////////////// |
| 182 | /// Unix Platform |
| 183 | //////////////////////////////////////////////////////////////////////////////// |
| 184 | |
| 185 | #if defined(__UNIX__) |
| 186 | |
| 187 | #include <sys/mman.h> |
| 188 | #include <errno.h> |
| 189 | #include <stdlib.h> |
| 190 | #include <string.h> |
| 191 | #include <sstream> |
| 192 | |
| 193 | #if defined(__MACOSX__) |
| 194 | #include <mach/vm_statistics.h> |
| 195 | #endif |
| 196 | |
| 197 | namespace embree |
| 198 | { |
| 199 | bool os_init(bool hugepages, bool verbose) |
| 200 | { |
| 201 | Lock<MutexSys> lock(os_init_mutex); |
| 202 | |
| 203 | if (!hugepages) { |
| 204 | huge_pages_enabled = false; |
| 205 | return true; |
| 206 | } |
| 207 | |
| 208 | #if defined(__LINUX__) |
| 209 | |
| 210 | int hugepagesize = 0; |
| 211 | |
| 212 | std::ifstream file; |
| 213 | file.open("/proc/meminfo" ,std::ios::in); |
| 214 | if (!file.is_open()) { |
| 215 | if (verbose) std::cout << "WARNING: Could not open /proc/meminfo. Huge page support cannot get enabled!" << std::endl; |
| 216 | huge_pages_enabled = false; |
| 217 | return false; |
| 218 | } |
| 219 | |
| 220 | std::string line; |
| 221 | while (getline(file,line)) |
| 222 | { |
| 223 | std::stringstream sline(line); |
| 224 | while (!sline.eof() && sline.peek() == ' ') sline.ignore(); |
| 225 | std::string tag; getline(sline,tag,' '); |
| 226 | while (!sline.eof() && sline.peek() == ' ') sline.ignore(); |
| 227 | std::string val; getline(sline,val,' '); |
| 228 | while (!sline.eof() && sline.peek() == ' ') sline.ignore(); |
| 229 | std::string unit; getline(sline,unit,' '); |
| 230 | if (tag == "Hugepagesize:" && unit == "kB" ) { |
| 231 | hugepagesize = std::stoi(val)*1024; |
| 232 | break; |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | if (hugepagesize != PAGE_SIZE_2M) |
| 237 | { |
| 238 | if (verbose) std::cout << "WARNING: Only 2MB huge pages supported. Huge page support cannot get enabled!" << std::endl; |
| 239 | huge_pages_enabled = false; |
| 240 | return false; |
| 241 | } |
| 242 | #endif |
| 243 | |
| 244 | huge_pages_enabled = true; |
| 245 | return true; |
| 246 | } |
| 247 | |
| 248 | void* os_malloc(size_t bytes, bool& hugepages) |
| 249 | { |
| 250 | if (bytes == 0) { |
| 251 | hugepages = false; |
| 252 | return nullptr; |
| 253 | } |
| 254 | |
| 255 | /* try direct huge page allocation first */ |
| 256 | if (isHugePageCandidate(bytes)) |
| 257 | { |
| 258 | #if defined(__MACOSX__) |
| 259 | void* ptr = mmap(0, bytes, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON, VM_FLAGS_SUPERPAGE_SIZE_2MB, 0); |
| 260 | if (ptr != MAP_FAILED) { |
| 261 | hugepages = true; |
| 262 | return ptr; |
| 263 | } |
| 264 | #elif defined(MAP_HUGETLB) |
| 265 | void* ptr = mmap(0, bytes, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON | MAP_HUGETLB, -1, 0); |
| 266 | if (ptr != MAP_FAILED) { |
| 267 | hugepages = true; |
| 268 | return ptr; |
| 269 | } |
| 270 | #endif |
| 271 | } |
| 272 | |
| 273 | /* fallback to 4k pages */ |
| 274 | void* ptr = (char*) mmap(0, bytes, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON, -1, 0); |
| 275 | // -- GODOT start -- |
| 276 | // if (ptr == MAP_FAILED) throw std::bad_alloc(); |
| 277 | if (ptr == MAP_FAILED) abort(); |
| 278 | // -- GODOT end -- |
| 279 | hugepages = false; |
| 280 | |
| 281 | /* advise huge page hint for THP */ |
| 282 | os_advise(ptr,bytes); |
| 283 | return ptr; |
| 284 | } |
| 285 | |
| 286 | size_t os_shrink(void* ptr, size_t bytesNew, size_t bytesOld, bool hugepages) |
| 287 | { |
| 288 | const size_t pageSize = hugepages ? PAGE_SIZE_2M : PAGE_SIZE_4K; |
| 289 | bytesNew = (bytesNew+pageSize-1) & ~(pageSize-1); |
| 290 | bytesOld = (bytesOld+pageSize-1) & ~(pageSize-1); |
| 291 | if (bytesNew >= bytesOld) |
| 292 | return bytesOld; |
| 293 | |
| 294 | if (munmap((char*)ptr+bytesNew,bytesOld-bytesNew) == -1) |
| 295 | // -- GODOT start -- |
| 296 | // throw std::bad_alloc(); |
| 297 | abort(); |
| 298 | // -- GODOT end -- |
| 299 | |
| 300 | return bytesNew; |
| 301 | } |
| 302 | |
| 303 | void os_free(void* ptr, size_t bytes, bool hugepages) |
| 304 | { |
| 305 | if (bytes == 0) |
| 306 | return; |
| 307 | |
| 308 | /* for hugepages we need to also align the size */ |
| 309 | const size_t pageSize = hugepages ? PAGE_SIZE_2M : PAGE_SIZE_4K; |
| 310 | bytes = (bytes+pageSize-1) & ~(pageSize-1); |
| 311 | if (munmap(ptr,bytes) == -1) |
| 312 | // -- GODOT start -- |
| 313 | // throw std::bad_alloc(); |
| 314 | abort(); |
| 315 | // -- GODOT end -- |
| 316 | } |
| 317 | |
| 318 | /* hint for transparent huge pages (THP) */ |
| 319 | void os_advise(void* pptr, size_t bytes) |
| 320 | { |
| 321 | #if defined(MADV_HUGEPAGE) |
| 322 | madvise(pptr,bytes,MADV_HUGEPAGE); |
| 323 | #endif |
| 324 | } |
| 325 | } |
| 326 | |
| 327 | #endif |
| 328 | |