[7abee38] | 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.cc -- |
---|
| 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 | |
---|
[120a28c3] | 23 | #if defined(__has_feature) |
---|
| 24 | #if __has_feature(address_sanitizer) |
---|
| 25 | #define NO_HEAP_STATISTICS |
---|
| 26 | # endif |
---|
| 27 | #endif |
---|
| 28 | |
---|
| 29 | #if defined( NO_STATISTICS ) || defined( TCMALLOC ) || defined(__SANITIZE_ADDRESS__) |
---|
| 30 | #if !defined(NO_HEAP_STATISTICS) |
---|
| 31 | #define NO_HEAP_STATISTICS |
---|
| 32 | #endif |
---|
[1cb7fab2] | 33 | #endif |
---|
| 34 | |
---|
[7abee38] | 35 | namespace Stats { |
---|
| 36 | namespace Heap { |
---|
[8f74a6a] | 37 | #if defined( NO_HEAP_STATISTICS ) |
---|
[7abee38] | 38 | void newPass( const char * const ) {} |
---|
| 39 | |
---|
[8f74a6a] | 40 | void print() {} |
---|
[7abee38] | 41 | #else |
---|
[1cb7fab2] | 42 | extern bool enabled; |
---|
| 43 | |
---|
[7abee38] | 44 | struct StatBlock { |
---|
| 45 | const char * name = nullptr; ///< Name of this pass |
---|
| 46 | size_t mallocs = 0; ///< Allocations in this pass |
---|
| 47 | size_t frees = 0; ///< Frees in this pass |
---|
| 48 | size_t n_allocs = 0; ///< Current number of live allocations |
---|
| 49 | size_t peak_allocs = 0; ///< Peak number of live allocations this pass |
---|
| 50 | }; |
---|
| 51 | |
---|
| 52 | StatBlock passes[100] = {{ "Pre-Parse", 0, 0, 0, 0 }}; |
---|
| 53 | const size_t passes_size = sizeof(passes) / sizeof(passes[0]); |
---|
| 54 | size_t passes_cnt = 1; |
---|
| 55 | |
---|
[c15085d] | 56 | StatBlock stacktrace_stats[100]; |
---|
| 57 | size_t stacktrace_stats_count = 0; |
---|
| 58 | bool stacktrace_stats_enabled = true; |
---|
| 59 | |
---|
| 60 | size_t trace[1000]; |
---|
| 61 | const size_t stacktrace_max_depth = sizeof(trace) / sizeof(size_t); |
---|
| 62 | size_t stacktrace_depth; |
---|
| 63 | |
---|
| 64 | size_t new_stacktrace_id(const char * const name) { |
---|
| 65 | stacktrace_stats[stacktrace_stats_count].name = name; |
---|
| 66 | return stacktrace_stats_count++; |
---|
| 67 | } |
---|
| 68 | |
---|
| 69 | void stacktrace_push(size_t id) { |
---|
| 70 | ++stacktrace_depth; |
---|
| 71 | assertf(stacktrace_depth < stacktrace_max_depth, "Stack trace too deep: increase size of array in Heap.cc"); |
---|
| 72 | trace[stacktrace_depth] = id; |
---|
| 73 | } |
---|
| 74 | |
---|
| 75 | void stacktrace_pop() { |
---|
| 76 | assertf(stacktrace_depth > 0, "Invalid stack tracing operation: trace is empty"); |
---|
| 77 | --stacktrace_depth; |
---|
| 78 | } |
---|
| 79 | |
---|
[7abee38] | 80 | void newPass( const char * const name ) { |
---|
| 81 | passes[passes_cnt].name = name; |
---|
| 82 | passes[passes_cnt].mallocs = 0; |
---|
| 83 | passes[passes_cnt].frees = 0; |
---|
| 84 | passes[passes_cnt].n_allocs |
---|
| 85 | = passes[passes_cnt].peak_allocs |
---|
| 86 | = passes[passes_cnt-1].n_allocs; |
---|
| 87 | passes_cnt++; |
---|
| 88 | |
---|
| 89 | assertf(passes_cnt < passes_size, "Too many passes for Stats::Heap, increase the size of the array in Heap.cc"); |
---|
| 90 | } |
---|
| 91 | |
---|
| 92 | void print(size_t value, size_t total) { |
---|
| 93 | std::cerr << std::setw(12) << value; |
---|
| 94 | std::cerr << "(" << std::setw(3); |
---|
| 95 | std::cerr << (value == 0 ? 0 : value * 100 / total); |
---|
| 96 | std::cerr << "%) | "; |
---|
| 97 | } |
---|
| 98 | |
---|
| 99 | void print(const StatBlock& stat, size_t nc, size_t total_mallocs, size_t total_frees, size_t overall_peak) { |
---|
| 100 | std::cerr << std::setw(nc) << stat.name; |
---|
| 101 | std::cerr << " | "; |
---|
| 102 | |
---|
| 103 | print(stat.mallocs, total_mallocs); |
---|
| 104 | print(stat.frees, total_frees ); |
---|
| 105 | print(stat.peak_allocs, overall_peak ); |
---|
| 106 | std::cerr << "\n"; |
---|
| 107 | } |
---|
| 108 | |
---|
| 109 | void print(char c, size_t nc) { |
---|
| 110 | for(size_t i = 0; i < nc; i++) { |
---|
| 111 | std::cerr << c; |
---|
| 112 | } |
---|
| 113 | std::cerr << '\n'; |
---|
| 114 | } |
---|
| 115 | |
---|
[8f74a6a] | 116 | void print() { |
---|
[1cb7fab2] | 117 | if(!enabled) return; |
---|
| 118 | |
---|
[7abee38] | 119 | size_t nc = 0; |
---|
| 120 | size_t total_mallocs = 0; |
---|
| 121 | size_t total_frees = 0; |
---|
| 122 | size_t overall_peak = 0; |
---|
| 123 | for(size_t i = 0; i < passes_cnt; i++) { |
---|
| 124 | nc = std::max(nc, std::strlen(passes[i].name)); |
---|
| 125 | total_mallocs += passes[i].mallocs; |
---|
| 126 | total_frees += passes[i].frees; |
---|
| 127 | overall_peak = std::max(overall_peak, passes[i].peak_allocs); |
---|
| 128 | } |
---|
| 129 | size_t nct = nc + 65; |
---|
| 130 | |
---|
| 131 | const char * const title = "Heap Usage Statistic"; |
---|
| 132 | print('=', nct); |
---|
| 133 | for(size_t i = 0; i < (nct - std::strlen(title)) / 2; i++) std::cerr << ' '; |
---|
| 134 | std::cerr << title << std::endl; |
---|
| 135 | print('-', nct); |
---|
| 136 | std::cerr << std::setw(nc) << "Pass"; |
---|
| 137 | std::cerr << " | Malloc Count | Free Count | Peak Allocs |" << std::endl; |
---|
| 138 | |
---|
| 139 | print('-', nct); |
---|
| 140 | for(size_t i = 0; i < passes_cnt; i++) { |
---|
| 141 | print(passes[i], nc, total_mallocs, total_frees, overall_peak); |
---|
| 142 | } |
---|
[c15085d] | 143 | |
---|
| 144 | print('-', nct); |
---|
| 145 | std::cerr << std::setw(nc) << "Trace"; |
---|
| 146 | std::cerr << " | Malloc Count | Free Count | Peak Allocs |" << std::endl; |
---|
| 147 | |
---|
| 148 | print('-', nct); |
---|
| 149 | for (size_t i = 0; i < stacktrace_stats_count; i++) { |
---|
| 150 | print(stacktrace_stats[i], nc, total_mallocs, total_frees, overall_peak); |
---|
| 151 | } |
---|
[7abee38] | 152 | print('-', nct); |
---|
| 153 | print({"Sum", total_mallocs, total_frees, 0, overall_peak}, |
---|
| 154 | nc, total_mallocs, total_frees, overall_peak); |
---|
| 155 | |
---|
| 156 | } |
---|
| 157 | |
---|
| 158 | #include <stdarg.h> |
---|
| 159 | #include <stddef.h> |
---|
| 160 | #include <stdio.h> |
---|
| 161 | #include <string.h> |
---|
| 162 | #include <unistd.h> |
---|
| 163 | #include <signal.h> |
---|
| 164 | extern "C" { |
---|
| 165 | #include <dlfcn.h> |
---|
| 166 | #include <execinfo.h> |
---|
| 167 | } |
---|
| 168 | |
---|
| 169 | //============================================================================================= |
---|
| 170 | // Interposing helpers |
---|
| 171 | //============================================================================================= |
---|
| 172 | |
---|
| 173 | typedef void (* generic_fptr_t)(void); |
---|
| 174 | generic_fptr_t interpose_symbol( const char * symbol, const char * version ) { |
---|
| 175 | const char * error; |
---|
| 176 | |
---|
| 177 | static void * library; |
---|
| 178 | if ( ! library ) { |
---|
| 179 | # if defined( RTLD_NEXT ) |
---|
| 180 | library = RTLD_NEXT; |
---|
| 181 | # else |
---|
| 182 | // missing RTLD_NEXT => must hard-code library name, assuming libstdc++ |
---|
| 183 | library = dlopen( "libc.so.6", RTLD_LAZY ); |
---|
| 184 | error = dlerror(); |
---|
| 185 | if ( error ) { |
---|
| 186 | std::cerr << "interpose_symbol : failed to open libc, " << error << std::endl; |
---|
| 187 | abort(); |
---|
| 188 | } |
---|
| 189 | # endif // RTLD_NEXT |
---|
| 190 | } // if |
---|
| 191 | |
---|
| 192 | generic_fptr_t fptr; |
---|
| 193 | |
---|
| 194 | # if defined( _GNU_SOURCE ) |
---|
| 195 | if ( version ) { |
---|
| 196 | fptr = (generic_fptr_t)dlvsym( library, symbol, version ); |
---|
| 197 | } else { |
---|
| 198 | fptr = (generic_fptr_t)dlsym( library, symbol ); |
---|
| 199 | } |
---|
| 200 | # else |
---|
| 201 | fptr = (generic_fptr_t)dlsym( library, symbol ); |
---|
| 202 | # endif // _GNU_SOURCE |
---|
| 203 | |
---|
| 204 | error = dlerror(); |
---|
| 205 | if ( error ) { |
---|
| 206 | std::cerr << "interpose_symbol : internal error, " << error << std::endl; |
---|
| 207 | abort(); |
---|
| 208 | } |
---|
| 209 | |
---|
| 210 | return fptr; |
---|
| 211 | } |
---|
| 212 | |
---|
| 213 | extern "C" { |
---|
| 214 | void * malloc( size_t size ) __attribute__((malloc)); |
---|
| 215 | void * malloc( size_t size ) { |
---|
| 216 | static auto __malloc = reinterpret_cast<void * (*)(size_t)>(interpose_symbol( "malloc", nullptr )); |
---|
[1cb7fab2] | 217 | if( enabled && passes_cnt > 0 ) { |
---|
[7abee38] | 218 | passes[passes_cnt - 1].mallocs++; |
---|
| 219 | passes[passes_cnt - 1].n_allocs++; |
---|
| 220 | passes[passes_cnt - 1].peak_allocs |
---|
| 221 | = std::max(passes[passes_cnt - 1].peak_allocs, passes[passes_cnt - 1].n_allocs); |
---|
| 222 | } |
---|
[c15085d] | 223 | |
---|
| 224 | if ( stacktrace_stats_enabled && stacktrace_depth > 0) { |
---|
| 225 | stacktrace_stats[trace[stacktrace_depth]].mallocs++; |
---|
| 226 | } |
---|
[7abee38] | 227 | return __malloc( size ); |
---|
| 228 | } |
---|
| 229 | |
---|
| 230 | void free( void * ptr ) { |
---|
| 231 | static auto __free = reinterpret_cast<void (*)(void *)>(interpose_symbol( "free", nullptr )); |
---|
[1cb7fab2] | 232 | if( enabled && passes_cnt > 0 ) { |
---|
[7abee38] | 233 | passes[passes_cnt - 1].frees++; |
---|
| 234 | passes[passes_cnt - 1].n_allocs--; |
---|
| 235 | } |
---|
[c15085d] | 236 | if ( stacktrace_stats_enabled && stacktrace_depth > 0) { |
---|
| 237 | stacktrace_stats[trace[stacktrace_depth]].frees++; |
---|
| 238 | } |
---|
[7abee38] | 239 | return __free( ptr ); |
---|
| 240 | } |
---|
| 241 | |
---|
| 242 | void * calloc( size_t nelem, size_t size ) { |
---|
| 243 | static auto __calloc = reinterpret_cast<void * (*)(size_t, size_t)>(interpose_symbol( "calloc", nullptr )); |
---|
[1cb7fab2] | 244 | if( enabled && passes_cnt > 0 ) { |
---|
[7abee38] | 245 | passes[passes_cnt - 1].mallocs++; |
---|
| 246 | passes[passes_cnt - 1].n_allocs++; |
---|
| 247 | passes[passes_cnt - 1].peak_allocs |
---|
| 248 | = std::max(passes[passes_cnt - 1].peak_allocs, passes[passes_cnt - 1].n_allocs); |
---|
| 249 | } |
---|
[c15085d] | 250 | if ( stacktrace_stats_enabled && stacktrace_depth > 0) { |
---|
| 251 | stacktrace_stats[trace[stacktrace_depth]].mallocs++; |
---|
| 252 | } |
---|
[7abee38] | 253 | return __calloc( nelem, size ); |
---|
| 254 | } |
---|
| 255 | |
---|
| 256 | void * realloc( void * ptr, size_t size ) { |
---|
| 257 | static auto __realloc = reinterpret_cast<void * (*)(void *, size_t)>(interpose_symbol( "realloc", nullptr )); |
---|
| 258 | void * s = __realloc( ptr, size ); |
---|
[1cb7fab2] | 259 | if ( enabled && s != ptr && passes_cnt > 0 ) { // did realloc get new storage ? |
---|
[7abee38] | 260 | passes[passes_cnt - 1].mallocs++; |
---|
| 261 | passes[passes_cnt - 1].frees++; |
---|
| 262 | } // if |
---|
[c15085d] | 263 | if ( stacktrace_stats_enabled && stacktrace_depth > 0) { |
---|
| 264 | stacktrace_stats[trace[stacktrace_depth]].mallocs++; |
---|
| 265 | stacktrace_stats[trace[stacktrace_depth]].frees++; |
---|
| 266 | } |
---|
[7abee38] | 267 | return s; |
---|
| 268 | } |
---|
| 269 | } |
---|
| 270 | #endif |
---|
| 271 | } |
---|
| 272 | } |
---|