| 1 | // | 
|---|
| 2 | // Cforall Version 1.0.0 Copyright (C) 2018 University of Waterloo | 
|---|
| 3 | // | 
|---|
| 4 | // The contents of this file are covered under the licence agreement in the | 
|---|
| 5 | // file "LICENCE" distributed with Cforall. | 
|---|
| 6 | // | 
|---|
| 7 | // Heap.cpp -- | 
|---|
| 8 | // | 
|---|
| 9 | // Author           : Thierry Delisle | 
|---|
| 10 | // Created On       : Thu May  3 16:16:10 2018 | 
|---|
| 11 | // Last Modified By : Peter A. Buhr | 
|---|
| 12 | // Last Modified On : Fri May  4 17:27:31 2018 | 
|---|
| 13 | // Update Count     : 28 | 
|---|
| 14 | // | 
|---|
| 15 |  | 
|---|
| 16 | #include <cassert> | 
|---|
| 17 | #include <cmath> | 
|---|
| 18 | #include <cstddef> | 
|---|
| 19 | #include <cstring> | 
|---|
| 20 | #include <iomanip> | 
|---|
| 21 | #include <iostream> | 
|---|
| 22 |  | 
|---|
| 23 | // Most of the other statistics features are deactivated only by defining | 
|---|
| 24 | // NO_STATISTICS (or their NO_%_STATISTICS macro). However the heap has some | 
|---|
| 25 | // other compatability concerns and will disable itself in some cases. | 
|---|
| 26 | // | 
|---|
| 27 | // I do not claim to understand these cases. But TCMALLOC is often defined by | 
|---|
| 28 | // default and you can pass --disable-gprofiler to configure to remove it. | 
|---|
| 29 |  | 
|---|
| 30 | #if defined(NO_STATISTICS) || defined(TCMALLOC) || defined(__SANITIZE_ADDRESS__) | 
|---|
| 31 | #define NO_HEAP_STATISTICS | 
|---|
| 32 | #elif defined(__has_feature) | 
|---|
| 33 | #if __has_feature(address_sanitizer) | 
|---|
| 34 | #define NO_HEAP_STATISTICS | 
|---|
| 35 | #endif | 
|---|
| 36 | #endif | 
|---|
| 37 |  | 
|---|
| 38 | namespace Stats { | 
|---|
| 39 | namespace Heap { | 
|---|
| 40 | #if defined( NO_HEAP_STATISTICS ) | 
|---|
| 41 | void newPass( const char * const ) {} | 
|---|
| 42 |  | 
|---|
| 43 | void print() {} | 
|---|
| 44 | #else | 
|---|
| 45 | extern bool enabled; | 
|---|
| 46 |  | 
|---|
| 47 | struct StatBlock { | 
|---|
| 48 | const char * name  = nullptr;   ///< Name of this pass | 
|---|
| 49 | size_t mallocs     = 0;                 ///< Allocations in this pass | 
|---|
| 50 | size_t frees       = 0;                 ///< Frees in this pass | 
|---|
| 51 | size_t n_allocs    = 0;                 ///< Current number of live allocations | 
|---|
| 52 | size_t peak_allocs = 0;                 ///< Peak number of live allocations this pass | 
|---|
| 53 | }; | 
|---|
| 54 |  | 
|---|
| 55 | StatBlock    passes[100] = {{ "Pre-Parse", 0, 0, 0, 0 }}; | 
|---|
| 56 | const size_t passes_size = sizeof(passes) / sizeof(passes[0]); | 
|---|
| 57 | size_t       passes_cnt = 1; | 
|---|
| 58 |  | 
|---|
| 59 | StatBlock    stacktrace_stats[100]; | 
|---|
| 60 | size_t       stacktrace_stats_count = 0; | 
|---|
| 61 | bool         stacktrace_stats_enabled = true; | 
|---|
| 62 |  | 
|---|
| 63 | size_t       trace[1000]; | 
|---|
| 64 | const size_t stacktrace_max_depth = sizeof(trace) / sizeof(size_t); | 
|---|
| 65 | size_t       stacktrace_depth; | 
|---|
| 66 |  | 
|---|
| 67 | size_t new_stacktrace_id(const char * const name) { | 
|---|
| 68 | stacktrace_stats[stacktrace_stats_count].name = name; | 
|---|
| 69 | return stacktrace_stats_count++; | 
|---|
| 70 | } | 
|---|
| 71 |  | 
|---|
| 72 | void stacktrace_push(size_t id) { | 
|---|
| 73 | ++stacktrace_depth; | 
|---|
| 74 | assertf(stacktrace_depth < stacktrace_max_depth, "Stack trace too deep: increase size of array in Heap.cpp"); | 
|---|
| 75 | trace[stacktrace_depth] = id; | 
|---|
| 76 | } | 
|---|
| 77 |  | 
|---|
| 78 | void stacktrace_pop() { | 
|---|
| 79 | assertf(stacktrace_depth > 0, "Invalid stack tracing operation: trace is empty"); | 
|---|
| 80 | --stacktrace_depth; | 
|---|
| 81 | } | 
|---|
| 82 |  | 
|---|
| 83 | void newPass( const char * const name ) { | 
|---|
| 84 | passes[passes_cnt].name    = name; | 
|---|
| 85 | passes[passes_cnt].mallocs = 0; | 
|---|
| 86 | passes[passes_cnt].frees   = 0; | 
|---|
| 87 | passes[passes_cnt].n_allocs | 
|---|
| 88 | = passes[passes_cnt].peak_allocs | 
|---|
| 89 | = passes[passes_cnt-1].n_allocs; | 
|---|
| 90 | passes_cnt++; | 
|---|
| 91 |  | 
|---|
| 92 | assertf(passes_cnt < passes_size, "Too many passes for Stats::Heap, increase the size of the array in Heap.cpp"); | 
|---|
| 93 | } | 
|---|
| 94 |  | 
|---|
| 95 | void print(size_t value, size_t total) { | 
|---|
| 96 | std::cerr << std::setw(12) << value; | 
|---|
| 97 | std::cerr << "(" << std::setw(3); | 
|---|
| 98 | std::cerr << (value == 0 ? 0 : value * 100 / total); | 
|---|
| 99 | std::cerr << "%) | "; | 
|---|
| 100 | } | 
|---|
| 101 |  | 
|---|
| 102 | void print(const StatBlock& stat, size_t nc, size_t total_mallocs, size_t total_frees, size_t overall_peak) { | 
|---|
| 103 | std::cerr << std::setw(nc) << stat.name; | 
|---|
| 104 | std::cerr << " | "; | 
|---|
| 105 |  | 
|---|
| 106 | print(stat.mallocs,     total_mallocs); | 
|---|
| 107 | print(stat.frees,       total_frees  ); | 
|---|
| 108 | print(stat.peak_allocs, overall_peak ); | 
|---|
| 109 | std::cerr << "\n"; | 
|---|
| 110 | } | 
|---|
| 111 |  | 
|---|
| 112 | void print(char c, size_t nc) { | 
|---|
| 113 | for(size_t i = 0; i < nc; i++) { | 
|---|
| 114 | std::cerr << c; | 
|---|
| 115 | } | 
|---|
| 116 | std::cerr << '\n'; | 
|---|
| 117 | } | 
|---|
| 118 |  | 
|---|
| 119 | void print() { | 
|---|
| 120 | if(!enabled) return; | 
|---|
| 121 |  | 
|---|
| 122 | size_t nc = 0; | 
|---|
| 123 | size_t total_mallocs = 0; | 
|---|
| 124 | size_t total_frees   = 0; | 
|---|
| 125 | size_t overall_peak  = 0; | 
|---|
| 126 | for(size_t i = 0; i < passes_cnt; i++) { | 
|---|
| 127 | nc = std::max(nc, std::strlen(passes[i].name)); | 
|---|
| 128 | total_mallocs += passes[i].mallocs; | 
|---|
| 129 | total_frees   += passes[i].frees; | 
|---|
| 130 | overall_peak = std::max(overall_peak, passes[i].peak_allocs); | 
|---|
| 131 | } | 
|---|
| 132 | size_t nct = nc + 65; | 
|---|
| 133 |  | 
|---|
| 134 | const char * const title = "Heap Usage Statistic"; | 
|---|
| 135 | print('=', nct); | 
|---|
| 136 | for(size_t i = 0; i < (nct - std::strlen(title)) / 2; i++) std::cerr << ' '; | 
|---|
| 137 | std::cerr << title << std::endl; | 
|---|
| 138 | print('-', nct); | 
|---|
| 139 | std::cerr << std::setw(nc) << "Pass"; | 
|---|
| 140 | std::cerr << " |       Malloc Count |         Free Count |        Peak Allocs |" << std::endl; | 
|---|
| 141 |  | 
|---|
| 142 | print('-', nct); | 
|---|
| 143 | for(size_t i = 0; i < passes_cnt; i++) { | 
|---|
| 144 | print(passes[i], nc, total_mallocs, total_frees, overall_peak); | 
|---|
| 145 | } | 
|---|
| 146 |  | 
|---|
| 147 | print('-', nct); | 
|---|
| 148 | std::cerr << std::setw(nc) << "Trace"; | 
|---|
| 149 | std::cerr << " |       Malloc Count |         Free Count |        Peak Allocs |" << std::endl; | 
|---|
| 150 |  | 
|---|
| 151 | print('-', nct); | 
|---|
| 152 | for (size_t i = 0; i < stacktrace_stats_count; i++) { | 
|---|
| 153 | print(stacktrace_stats[i], nc, total_mallocs, total_frees, overall_peak); | 
|---|
| 154 | } | 
|---|
| 155 | print('-', nct); | 
|---|
| 156 | print({"Sum", total_mallocs, total_frees, 0, overall_peak}, | 
|---|
| 157 | nc, total_mallocs, total_frees, overall_peak); | 
|---|
| 158 |  | 
|---|
| 159 | } | 
|---|
| 160 |  | 
|---|
| 161 | #include <stdarg.h> | 
|---|
| 162 | #include <stddef.h> | 
|---|
| 163 | #include <stdio.h> | 
|---|
| 164 | #include <string.h> | 
|---|
| 165 | #include <unistd.h> | 
|---|
| 166 | #include <signal.h> | 
|---|
| 167 | extern "C" { | 
|---|
| 168 | #include <dlfcn.h> | 
|---|
| 169 | #include <execinfo.h> | 
|---|
| 170 | } | 
|---|
| 171 |  | 
|---|
| 172 | //============================================================================================= | 
|---|
| 173 | // Interposing helpers | 
|---|
| 174 | //============================================================================================= | 
|---|
| 175 |  | 
|---|
| 176 | typedef void (* generic_fptr_t)(void); | 
|---|
| 177 | generic_fptr_t interpose_symbol( const char * symbol, const char * version ) { | 
|---|
| 178 | const char * error; | 
|---|
| 179 |  | 
|---|
| 180 | static void * library; | 
|---|
| 181 | if ( ! library ) { | 
|---|
| 182 | #                               if defined( RTLD_NEXT ) | 
|---|
| 183 | library = RTLD_NEXT; | 
|---|
| 184 | #                               else | 
|---|
| 185 | // missing RTLD_NEXT => must hard-code library name, assuming libstdc++ | 
|---|
| 186 | library = dlopen( "libc.so.6", RTLD_LAZY ); | 
|---|
| 187 | error = dlerror(); | 
|---|
| 188 | if ( error ) { | 
|---|
| 189 | std::cerr << "interpose_symbol : failed to open libc, " << error << std::endl; | 
|---|
| 190 | abort(); | 
|---|
| 191 | } | 
|---|
| 192 | #                               endif // RTLD_NEXT | 
|---|
| 193 | } // if | 
|---|
| 194 |  | 
|---|
| 195 | generic_fptr_t fptr; | 
|---|
| 196 |  | 
|---|
| 197 | #                       if defined( _GNU_SOURCE ) | 
|---|
| 198 | if ( version ) { | 
|---|
| 199 | fptr = (generic_fptr_t)dlvsym( library, symbol, version ); | 
|---|
| 200 | } else { | 
|---|
| 201 | fptr = (generic_fptr_t)dlsym( library, symbol ); | 
|---|
| 202 | } | 
|---|
| 203 | #                       else | 
|---|
| 204 | fptr = (generic_fptr_t)dlsym( library, symbol ); | 
|---|
| 205 | #                       endif // _GNU_SOURCE | 
|---|
| 206 |  | 
|---|
| 207 | error = dlerror(); | 
|---|
| 208 | if ( error ) { | 
|---|
| 209 | std::cerr << "interpose_symbol : internal error, " << error << std::endl; | 
|---|
| 210 | abort(); | 
|---|
| 211 | } | 
|---|
| 212 |  | 
|---|
| 213 | return fptr; | 
|---|
| 214 | } | 
|---|
| 215 |  | 
|---|
| 216 | extern "C" { | 
|---|
| 217 | void * malloc( size_t size ) __attribute__((malloc)); | 
|---|
| 218 | void * malloc( size_t size ) { | 
|---|
| 219 | static auto __malloc = reinterpret_cast<void * (*)(size_t)>(interpose_symbol( "malloc", nullptr )); | 
|---|
| 220 | if( enabled && passes_cnt > 0 ) { | 
|---|
| 221 | passes[passes_cnt - 1].mallocs++; | 
|---|
| 222 | passes[passes_cnt - 1].n_allocs++; | 
|---|
| 223 | passes[passes_cnt - 1].peak_allocs | 
|---|
| 224 | = std::max(passes[passes_cnt - 1].peak_allocs, passes[passes_cnt - 1].n_allocs); | 
|---|
| 225 | } | 
|---|
| 226 |  | 
|---|
| 227 | if ( stacktrace_stats_enabled && stacktrace_depth > 0) { | 
|---|
| 228 | stacktrace_stats[trace[stacktrace_depth]].mallocs++; | 
|---|
| 229 | } | 
|---|
| 230 | return __malloc( size ); | 
|---|
| 231 | } | 
|---|
| 232 |  | 
|---|
| 233 | void free( void * ptr ) { | 
|---|
| 234 | static auto __free = reinterpret_cast<void   (*)(void *)>(interpose_symbol( "free", nullptr )); | 
|---|
| 235 | if( enabled && passes_cnt > 0 ) { | 
|---|
| 236 | passes[passes_cnt - 1].frees++; | 
|---|
| 237 | passes[passes_cnt - 1].n_allocs--; | 
|---|
| 238 | } | 
|---|
| 239 | if ( stacktrace_stats_enabled && stacktrace_depth > 0) { | 
|---|
| 240 | stacktrace_stats[trace[stacktrace_depth]].frees++; | 
|---|
| 241 | } | 
|---|
| 242 | return __free( ptr ); | 
|---|
| 243 | } | 
|---|
| 244 |  | 
|---|
| 245 | void * calloc( size_t nelem, size_t size ) { | 
|---|
| 246 | static auto __calloc = reinterpret_cast<void * (*)(size_t, size_t)>(interpose_symbol( "calloc", nullptr )); | 
|---|
| 247 | if( enabled && passes_cnt > 0 ) { | 
|---|
| 248 | passes[passes_cnt - 1].mallocs++; | 
|---|
| 249 | passes[passes_cnt - 1].n_allocs++; | 
|---|
| 250 | passes[passes_cnt - 1].peak_allocs | 
|---|
| 251 | = std::max(passes[passes_cnt - 1].peak_allocs, passes[passes_cnt - 1].n_allocs); | 
|---|
| 252 | } | 
|---|
| 253 | if ( stacktrace_stats_enabled && stacktrace_depth > 0) { | 
|---|
| 254 | stacktrace_stats[trace[stacktrace_depth]].mallocs++; | 
|---|
| 255 | } | 
|---|
| 256 | return __calloc( nelem, size ); | 
|---|
| 257 | } | 
|---|
| 258 |  | 
|---|
| 259 | void * realloc( void * ptr, size_t size ) { | 
|---|
| 260 | static auto __realloc = reinterpret_cast<void * (*)(void *, size_t)>(interpose_symbol( "realloc", nullptr )); | 
|---|
| 261 | void * s = __realloc( ptr, size ); | 
|---|
| 262 | if ( enabled && s != ptr && passes_cnt > 0 ) {                  // did realloc get new storage ? | 
|---|
| 263 | passes[passes_cnt - 1].mallocs++; | 
|---|
| 264 | passes[passes_cnt - 1].frees++; | 
|---|
| 265 | } // if | 
|---|
| 266 | if ( stacktrace_stats_enabled && stacktrace_depth > 0) { | 
|---|
| 267 | stacktrace_stats[trace[stacktrace_depth]].mallocs++; | 
|---|
| 268 | stacktrace_stats[trace[stacktrace_depth]].frees++; | 
|---|
| 269 | } | 
|---|
| 270 | return s; | 
|---|
| 271 | } | 
|---|
| 272 | } | 
|---|
| 273 | #endif | 
|---|
| 274 | } | 
|---|
| 275 | } | 
|---|