[176] | 1 | /** |
---|
| 2 | * @file userealtls.cpp |
---|
| 3 | * @author Emery Berger <http://www.cs.umass.edu/~emery> |
---|
| 4 | * |
---|
| 5 | * This file leverages compiler support for thread-local variables for |
---|
| 6 | * access to thread-local heaps. It also intercepts thread completions |
---|
| 7 | * to flush these local heaps, returning any unused memory to the |
---|
| 8 | * global Hoard heap. On Windows, this happens in DllMain. On Unix |
---|
| 9 | * platforms, we interpose our own versions of pthread_create and |
---|
| 10 | * pthread_exit. |
---|
| 11 | */ |
---|
| 12 | |
---|
| 13 | #include "VERSION.h" |
---|
| 14 | |
---|
| 15 | #if defined(_WIN32) |
---|
| 16 | DWORD LocalTLABIndex; |
---|
| 17 | #endif |
---|
| 18 | |
---|
| 19 | #include <new> |
---|
| 20 | |
---|
| 21 | // For now, we only use thread-local variables (__thread) |
---|
| 22 | // (a) for Linux x86-32 platforms with gcc version > 3.3.0, and |
---|
| 23 | // (b) when compiling with the SunPro compilers. |
---|
| 24 | |
---|
| 25 | #if ((GCC_VERSION >= 30300) && !defined(__x86_64) && !defined(__SVR4) && !defined(__MACH__)) \ |
---|
| 26 | || defined(__SUNPRO_CC) |
---|
| 27 | #define USE_THREAD_KEYWORD 1 |
---|
| 28 | #endif |
---|
| 29 | |
---|
| 30 | |
---|
| 31 | #if defined(USE_THREAD_KEYWORD) |
---|
| 32 | |
---|
| 33 | static __thread double tlabBuffer[sizeof(TheCustomHeapType) / sizeof(double) + 1]; |
---|
| 34 | |
---|
| 35 | static __thread TheCustomHeapType * theTLAB = NULL; |
---|
| 36 | |
---|
| 37 | static TheCustomHeapType * initializeCustomHeap (void) { |
---|
| 38 | new ((char *) &tlabBuffer) TheCustomHeapType (getMainHoardHeap()); |
---|
| 39 | return (theTLAB = (TheCustomHeapType *) &tlabBuffer); |
---|
| 40 | } |
---|
| 41 | |
---|
| 42 | inline TheCustomHeapType * getCustomHeap (void) { |
---|
| 43 | // The pointer to the TLAB itself. |
---|
| 44 | theTLAB = (theTLAB ? theTLAB : initializeCustomHeap()); |
---|
| 45 | return theTLAB; |
---|
| 46 | } |
---|
| 47 | |
---|
| 48 | #else |
---|
| 49 | |
---|
| 50 | #if !defined(_WIN32) |
---|
| 51 | static pthread_key_t theHeapKey; |
---|
| 52 | static pthread_once_t key_once = PTHREAD_ONCE_INIT; |
---|
| 53 | |
---|
| 54 | static void deleteThatHeap (void * p) { |
---|
| 55 | // Called when the thread goes away - reclaims the heap. |
---|
| 56 | TheCustomHeapType * heap = (TheCustomHeapType *) p; |
---|
| 57 | heap->~TheCustomHeapType(); |
---|
| 58 | getMainHoardHeap()->free ((void *) heap); |
---|
| 59 | } |
---|
| 60 | |
---|
| 61 | static void make_heap_key (void) |
---|
| 62 | { |
---|
| 63 | if (pthread_key_create (&theHeapKey, deleteThatHeap) != 0) { |
---|
| 64 | // This should never happen. |
---|
| 65 | } |
---|
| 66 | } |
---|
| 67 | |
---|
| 68 | static void createKey (void) { |
---|
| 69 | // Ensure that the key is initialized. |
---|
| 70 | pthread_once (&key_once, make_heap_key); |
---|
| 71 | } |
---|
| 72 | |
---|
| 73 | #endif |
---|
| 74 | |
---|
| 75 | |
---|
| 76 | static TheCustomHeapType * initializeCustomHeap (void) |
---|
| 77 | { |
---|
| 78 | #if !defined(_WIN32) |
---|
| 79 | createKey(); |
---|
| 80 | assert (pthread_getspecific (theHeapKey) == NULL); |
---|
| 81 | #endif |
---|
| 82 | // Allocate a per-thread heap. |
---|
| 83 | TheCustomHeapType * heap; |
---|
| 84 | size_t sz = sizeof(TheCustomHeapType) + sizeof(double); |
---|
| 85 | void * mh = getMainHoardHeap()->malloc(sz); |
---|
| 86 | heap = new ((char *) mh) TheCustomHeapType (getMainHoardHeap()); |
---|
| 87 | |
---|
| 88 | // Store it in the appropriate thread-local area. |
---|
| 89 | #if defined(_WIN32) |
---|
| 90 | TlsSetValue (LocalTLABIndex, heap); |
---|
| 91 | #else |
---|
| 92 | int r = pthread_setspecific (theHeapKey, (void *) heap); |
---|
| 93 | assert (!r); |
---|
| 94 | #endif |
---|
| 95 | |
---|
| 96 | return heap; |
---|
| 97 | } |
---|
| 98 | |
---|
| 99 | |
---|
| 100 | inline TheCustomHeapType * getCustomHeap (void) { |
---|
| 101 | TheCustomHeapType * heap; |
---|
| 102 | #if defined(_WIN32) |
---|
| 103 | heap = (TheCustomHeapType *) TlsGetValue (LocalTLABIndex); |
---|
| 104 | #else |
---|
| 105 | createKey(); |
---|
| 106 | heap = (TheCustomHeapType *) pthread_getspecific (theHeapKey); |
---|
| 107 | #endif |
---|
| 108 | if (heap == NULL) { |
---|
| 109 | heap = initializeCustomHeap(); |
---|
| 110 | } |
---|
| 111 | return heap; |
---|
| 112 | } |
---|
| 113 | |
---|
| 114 | |
---|
| 115 | |
---|
| 116 | #endif |
---|
| 117 | |
---|
| 118 | // |
---|
| 119 | // Intercept thread creation and destruction to flush the TLABs. |
---|
| 120 | // |
---|
| 121 | |
---|
| 122 | #if defined(_WIN32) |
---|
| 123 | |
---|
| 124 | #ifndef HOARD_PRE_ACTION |
---|
| 125 | #define HOARD_PRE_ACTION |
---|
| 126 | #endif |
---|
| 127 | |
---|
| 128 | #ifndef HOARD_POST_ACTION |
---|
| 129 | #define HOARD_POST_ACTION |
---|
| 130 | #endif |
---|
| 131 | |
---|
| 132 | #include <stdio.h> |
---|
| 133 | |
---|
| 134 | #ifndef CUSTOM_DLLNAME |
---|
| 135 | #define CUSTOM_DLLNAME DllMain |
---|
| 136 | #endif |
---|
| 137 | |
---|
| 138 | extern "C" { |
---|
| 139 | |
---|
| 140 | BOOL WINAPI CUSTOM_DLLNAME (HANDLE hinstDLL, DWORD fdwReason, LPVOID lpreserved) |
---|
| 141 | { |
---|
| 142 | int i; |
---|
| 143 | int tid; |
---|
| 144 | static int np = HL::CPUInfo::computeNumProcessors(); |
---|
| 145 | |
---|
| 146 | switch (fdwReason) { |
---|
| 147 | |
---|
| 148 | case DLL_PROCESS_ATTACH: |
---|
| 149 | { |
---|
| 150 | LocalTLABIndex = TlsAlloc(); |
---|
| 151 | if (LocalTLABIndex == TLS_OUT_OF_INDEXES) { |
---|
| 152 | // Not sure what to do here! |
---|
| 153 | } |
---|
| 154 | HOARD_PRE_ACTION; |
---|
| 155 | getCustomHeap(); |
---|
| 156 | } |
---|
| 157 | break; |
---|
| 158 | |
---|
| 159 | case DLL_THREAD_ATTACH: |
---|
| 160 | if (np == 1) { |
---|
| 161 | // We have exactly one processor - just assign the thread to |
---|
| 162 | // heap 0. |
---|
| 163 | getMainHoardHeap()->chooseZero(); |
---|
| 164 | } else { |
---|
| 165 | getMainHoardHeap()->findUnusedHeap(); |
---|
| 166 | } |
---|
| 167 | getCustomHeap(); |
---|
| 168 | break; |
---|
| 169 | |
---|
| 170 | case DLL_THREAD_DETACH: |
---|
| 171 | { |
---|
| 172 | // Dump the memory from the TLAB. |
---|
| 173 | getCustomHeap()->clear(); |
---|
| 174 | |
---|
| 175 | TheCustomHeapType *heap |
---|
| 176 | = (TheCustomHeapType *) TlsGetValue(LocalTLABIndex); |
---|
| 177 | |
---|
| 178 | if (np != 1) { |
---|
| 179 | // If we're on a multiprocessor box, relinquish the heap |
---|
| 180 | // assigned to this thread. |
---|
| 181 | getMainHoardHeap()->releaseHeap(); |
---|
| 182 | } |
---|
| 183 | |
---|
| 184 | if (heap != 0) { |
---|
| 185 | TlsSetValue (LocalTLABIndex, 0); |
---|
| 186 | } |
---|
| 187 | } |
---|
| 188 | break; |
---|
| 189 | |
---|
| 190 | case DLL_PROCESS_DETACH: |
---|
| 191 | HOARD_POST_ACTION; |
---|
| 192 | break; |
---|
| 193 | |
---|
| 194 | default: |
---|
| 195 | return TRUE; |
---|
| 196 | } |
---|
| 197 | |
---|
| 198 | return TRUE; |
---|
| 199 | } |
---|
| 200 | |
---|
| 201 | } |
---|
| 202 | |
---|
| 203 | #else |
---|
| 204 | |
---|
| 205 | extern "C" { |
---|
| 206 | |
---|
| 207 | typedef void * (*threadFunctionType) (void *); |
---|
| 208 | |
---|
| 209 | typedef |
---|
| 210 | int (*pthread_create_function) (pthread_t *thread, |
---|
| 211 | const pthread_attr_t *attr, |
---|
| 212 | threadFunctionType start_routine, |
---|
| 213 | void *arg); |
---|
| 214 | } |
---|
| 215 | |
---|
| 216 | // A special routine we call on thread exits to free up some resources. |
---|
| 217 | static void exitRoutine (void) { |
---|
| 218 | // Clear the TLAB's buffer. |
---|
| 219 | getCustomHeap()->clear(); |
---|
| 220 | |
---|
| 221 | // Relinquish the assigned heap. |
---|
| 222 | getMainHoardHeap()->releaseHeap(); |
---|
| 223 | } |
---|
| 224 | |
---|
| 225 | |
---|
| 226 | extern "C" { |
---|
| 227 | |
---|
| 228 | static void * startMeUp (void * a) |
---|
| 229 | { |
---|
| 230 | getCustomHeap(); |
---|
| 231 | getMainHoardHeap()->findUnusedHeap(); |
---|
| 232 | |
---|
| 233 | pair<threadFunctionType, void *> * z |
---|
| 234 | = (pair<threadFunctionType, void *> *) a; |
---|
| 235 | |
---|
| 236 | threadFunctionType f = z->first; |
---|
| 237 | void * arg = z->second; |
---|
| 238 | void * result = (*f)(arg); |
---|
| 239 | exitRoutine(); |
---|
| 240 | delete z; |
---|
| 241 | return result; |
---|
| 242 | } |
---|
| 243 | |
---|
| 244 | } |
---|
| 245 | |
---|
| 246 | #include <dlfcn.h> |
---|
| 247 | |
---|
| 248 | |
---|
| 249 | extern "C" int pthread_create (pthread_t *thread, |
---|
| 250 | const pthread_attr_t *attr, |
---|
| 251 | void * (*start_routine) (void *), |
---|
| 252 | void * arg) |
---|
| 253 | #if !defined(__SUNPRO_CC) |
---|
| 254 | throw () |
---|
| 255 | #endif |
---|
| 256 | { |
---|
| 257 | // A pointer to the library version of pthread_create. |
---|
| 258 | static pthread_create_function real_pthread_create = NULL; |
---|
| 259 | |
---|
| 260 | // Force initialization of the TLAB before our first thread is created. |
---|
| 261 | volatile static TheCustomHeapType * t = getCustomHeap(); |
---|
| 262 | |
---|
| 263 | #if defined(linux) |
---|
| 264 | char fname[] = "pthread_create"; |
---|
| 265 | #else |
---|
| 266 | char fname[] = "_pthread_create"; |
---|
| 267 | #endif |
---|
| 268 | |
---|
| 269 | // Instantiate the pointer to pthread_create, if it hasn't been |
---|
| 270 | // instantiated yet. |
---|
| 271 | |
---|
| 272 | if (real_pthread_create == NULL) { |
---|
| 273 | real_pthread_create = (pthread_create_function) dlsym (RTLD_NEXT, fname); |
---|
| 274 | if (real_pthread_create == NULL) { |
---|
| 275 | fprintf (stderr, "Could not find the pthread_create function!\n"); |
---|
| 276 | fprintf (stderr, "Please report this problem to emery@cs.umass.edu.\n"); |
---|
| 277 | abort(); |
---|
| 278 | } |
---|
| 279 | } |
---|
| 280 | |
---|
| 281 | anyThreadCreated = 1; |
---|
| 282 | |
---|
| 283 | pair<threadFunctionType, void *> * args = |
---|
| 284 | new pair<threadFunctionType, void *> (start_routine, arg); |
---|
| 285 | |
---|
| 286 | int result = (*real_pthread_create)(thread, attr, startMeUp, args); |
---|
| 287 | |
---|
| 288 | return result; |
---|
| 289 | } |
---|
| 290 | |
---|
| 291 | #endif |
---|