1// Copyright (c) 2005, Google Inc. 2// All rights reserved. 3// 4// Redistribution and use in source and binary forms, with or without 5// modification, are permitted provided that the following conditions are 6// met: 7// 8// * Redistributions of source code must retain the above copyright 9// notice, this list of conditions and the following disclaimer. 10// * Redistributions in binary form must reproduce the above 11// copyright notice, this list of conditions and the following disclaimer 12// in the documentation and/or other materials provided with the 13// distribution. 14// * Neither the name of Google Inc. nor the names of its 15// contributors may be used to endorse or promote products derived from 16// this software without specific prior written permission. 17// 18// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 19// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 20// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 21// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 22// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 23// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 24// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 25// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 26// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 27// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 28// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 29 30// --- 31// Author: Sanjay Ghemawat 32// 33// Produce stack trace 34 35#ifndef BASE_STACKTRACE_X86_INL_H_ 36#define BASE_STACKTRACE_X86_INL_H_ 37// Note: this file is included into stacktrace.cc more than once. 38// Anything that should only be defined once should be here: 39 40#include "config.h" 41#include <stdlib.h> // for NULL 42#include <assert.h> 43#if defined(HAVE_SYS_UCONTEXT_H) 44#include <sys/ucontext.h> 45#elif defined(HAVE_UCONTEXT_H) 46#include <ucontext.h> // for ucontext_t 47#elif defined(HAVE_CYGWIN_SIGNAL_H) 48// cygwin/signal.h has a buglet where it uses pthread_attr_t without 49// #including <pthread.h> itself. So we have to do it. 50# ifdef HAVE_PTHREAD 51# include <pthread.h> 52# endif 53#include <cygwin/signal.h> 54typedef ucontext ucontext_t; 55#endif 56#ifdef HAVE_STDINT_H 57#include <stdint.h> // for uintptr_t 58#endif 59#ifdef HAVE_UNISTD_H 60#include <unistd.h> 61#endif 62#ifdef HAVE_MMAP 63#include <sys/mman.h> // for msync 64#include "base/vdso_support.h" 65#endif 66 67#include "gperftools/stacktrace.h" 68 69#if defined(__linux__) && defined(__i386__) && defined(__ELF__) && defined(HAVE_MMAP) 70// Count "push %reg" instructions in VDSO __kernel_vsyscall(), 71// preceeding "syscall" or "sysenter". 72// If __kernel_vsyscall uses frame pointer, answer 0. 73// 74// kMaxBytes tells how many instruction bytes of __kernel_vsyscall 75// to analyze before giving up. Up to kMaxBytes+1 bytes of 76// instructions could be accessed. 77// 78// Here are known __kernel_vsyscall instruction sequences: 79// 80// SYSENTER (linux-2.6.26/arch/x86/vdso/vdso32/sysenter.S). 81// Used on Intel. 82// 0xffffe400 <__kernel_vsyscall+0>: push %ecx 83// 0xffffe401 <__kernel_vsyscall+1>: push %edx 84// 0xffffe402 <__kernel_vsyscall+2>: push %ebp 85// 0xffffe403 <__kernel_vsyscall+3>: mov %esp,%ebp 86// 0xffffe405 <__kernel_vsyscall+5>: sysenter 87// 88// SYSCALL (see linux-2.6.26/arch/x86/vdso/vdso32/syscall.S). 89// Used on AMD. 90// 0xffffe400 <__kernel_vsyscall+0>: push %ebp 91// 0xffffe401 <__kernel_vsyscall+1>: mov %ecx,%ebp 92// 0xffffe403 <__kernel_vsyscall+3>: syscall 93// 94// i386 (see linux-2.6.26/arch/x86/vdso/vdso32/int80.S) 95// 0xffffe400 <__kernel_vsyscall+0>: int $0x80 96// 0xffffe401 <__kernel_vsyscall+1>: ret 97// 98static const int kMaxBytes = 10; 99 100// We use assert()s instead of DCHECK()s -- this is too low level 101// for DCHECK(). 102 103static int CountPushInstructions(const unsigned char *const addr) { 104 int result = 0; 105 for (int i = 0; i < kMaxBytes; ++i) { 106 if (addr[i] == 0x89) { 107 // "mov reg,reg" 108 if (addr[i + 1] == 0xE5) { 109 // Found "mov %esp,%ebp". 110 return 0; 111 } 112 ++i; // Skip register encoding byte. 113 } else if (addr[i] == 0x0F && 114 (addr[i + 1] == 0x34 || addr[i + 1] == 0x05)) { 115 // Found "sysenter" or "syscall". 116 return result; 117 } else if ((addr[i] & 0xF0) == 0x50) { 118 // Found "push %reg". 119 ++result; 120 } else if (addr[i] == 0xCD && addr[i + 1] == 0x80) { 121 // Found "int $0x80" 122 assert(result == 0); 123 return 0; 124 } else { 125 // Unexpected instruction. 126 assert(0 == "unexpected instruction in __kernel_vsyscall"); 127 return 0; 128 } 129 } 130 // Unexpected: didn't find SYSENTER or SYSCALL in 131 // [__kernel_vsyscall, __kernel_vsyscall + kMaxBytes) interval. 132 assert(0 == "did not find SYSENTER or SYSCALL in __kernel_vsyscall"); 133 return 0; 134} 135#endif 136 137// Given a pointer to a stack frame, locate and return the calling 138// stackframe, or return NULL if no stackframe can be found. Perform sanity 139// checks (the strictness of which is controlled by the boolean parameter 140// "STRICT_UNWINDING") to reduce the chance that a bad pointer is returned. 141template<bool STRICT_UNWINDING, bool WITH_CONTEXT> 142static void **NextStackFrame(void **old_sp, const void *uc) { 143 void **new_sp = (void **) *old_sp; 144 145#if defined(__linux__) && defined(__i386__) && defined(HAVE_VDSO_SUPPORT) 146 if (WITH_CONTEXT && uc != NULL) { 147 // How many "push %reg" instructions are there at __kernel_vsyscall? 148 // This is constant for a given kernel and processor, so compute 149 // it only once. 150 static int num_push_instructions = -1; // Sentinel: not computed yet. 151 // Initialize with sentinel value: __kernel_rt_sigreturn can not possibly 152 // be there. 153 static const unsigned char *kernel_rt_sigreturn_address = NULL; 154 static const unsigned char *kernel_vsyscall_address = NULL; 155 if (num_push_instructions == -1) { 156 base::VDSOSupport vdso; 157 if (vdso.IsPresent()) { 158 base::VDSOSupport::SymbolInfo rt_sigreturn_symbol_info; 159 base::VDSOSupport::SymbolInfo vsyscall_symbol_info; 160 if (!vdso.LookupSymbol("__kernel_rt_sigreturn", "LINUX_2.5", 161 STT_FUNC, &rt_sigreturn_symbol_info) || 162 !vdso.LookupSymbol("__kernel_vsyscall", "LINUX_2.5", 163 STT_FUNC, &vsyscall_symbol_info) || 164 rt_sigreturn_symbol_info.address == NULL || 165 vsyscall_symbol_info.address == NULL) { 166 // Unexpected: 32-bit VDSO is present, yet one of the expected 167 // symbols is missing or NULL. 168 assert(0 == "VDSO is present, but doesn't have expected symbols"); 169 num_push_instructions = 0; 170 } else { 171 kernel_rt_sigreturn_address = 172 reinterpret_cast<const unsigned char *>( 173 rt_sigreturn_symbol_info.address); 174 kernel_vsyscall_address = 175 reinterpret_cast<const unsigned char *>( 176 vsyscall_symbol_info.address); 177 num_push_instructions = 178 CountPushInstructions(kernel_vsyscall_address); 179 } 180 } else { 181 num_push_instructions = 0; 182 } 183 } 184 if (num_push_instructions != 0 && kernel_rt_sigreturn_address != NULL && 185 old_sp[1] == kernel_rt_sigreturn_address) { 186 const ucontext_t *ucv = static_cast<const ucontext_t *>(uc); 187 // This kernel does not use frame pointer in its VDSO code, 188 // and so %ebp is not suitable for unwinding. 189 void **const reg_ebp = 190 reinterpret_cast<void **>(ucv->uc_mcontext.gregs[REG_EBP]); 191 const unsigned char *const reg_eip = 192 reinterpret_cast<unsigned char *>(ucv->uc_mcontext.gregs[REG_EIP]); 193 if (new_sp == reg_ebp && 194 kernel_vsyscall_address <= reg_eip && 195 reg_eip - kernel_vsyscall_address < kMaxBytes) { 196 // We "stepped up" to __kernel_vsyscall, but %ebp is not usable. 197 // Restore from 'ucv' instead. 198 void **const reg_esp = 199 reinterpret_cast<void **>(ucv->uc_mcontext.gregs[REG_ESP]); 200 // Check that alleged %esp is not NULL and is reasonably aligned. 201 if (reg_esp && 202 ((uintptr_t)reg_esp & (sizeof(reg_esp) - 1)) == 0) { 203 // Check that alleged %esp is actually readable. This is to prevent 204 // "double fault" in case we hit the first fault due to e.g. stack 205 // corruption. 206 // 207 // page_size is linker-initalized to avoid async-unsafe locking 208 // that GCC would otherwise insert (__cxa_guard_acquire etc). 209 static int page_size; 210 if (page_size == 0) { 211 // First time through. 212 page_size = getpagesize(); 213 } 214 void *const reg_esp_aligned = 215 reinterpret_cast<void *>( 216 (uintptr_t)(reg_esp + num_push_instructions - 1) & 217 ~(page_size - 1)); 218 if (msync(reg_esp_aligned, page_size, MS_ASYNC) == 0) { 219 // Alleged %esp is readable, use it for further unwinding. 220 new_sp = reinterpret_cast<void **>( 221 reg_esp[num_push_instructions - 1]); 222 } 223 } 224 } 225 } 226 } 227#endif 228 229 // Check that the transition from frame pointer old_sp to frame 230 // pointer new_sp isn't clearly bogus 231 if (STRICT_UNWINDING) { 232 // With the stack growing downwards, older stack frame must be 233 // at a greater address that the current one. 234 if (new_sp <= old_sp) return NULL; 235 // Assume stack frames larger than 100,000 bytes are bogus. 236 if ((uintptr_t)new_sp - (uintptr_t)old_sp > 100000) return NULL; 237 } else { 238 // In the non-strict mode, allow discontiguous stack frames. 239 // (alternate-signal-stacks for example). 240 if (new_sp == old_sp) return NULL; 241 if (new_sp > old_sp) { 242 // And allow frames upto about 1MB. 243 const uintptr_t delta = (uintptr_t)new_sp - (uintptr_t)old_sp; 244 const uintptr_t acceptable_delta = 1000000; 245 if (delta > acceptable_delta) { 246 return NULL; 247 } 248 } 249 } 250 if ((uintptr_t)new_sp & (sizeof(void *) - 1)) return NULL; 251#ifdef __i386__ 252 // On 64-bit machines, the stack pointer can be very close to 253 // 0xffffffff, so we explicitly check for a pointer into the 254 // last two pages in the address space 255 if ((uintptr_t)new_sp >= 0xffffe000) return NULL; 256#endif 257#ifdef HAVE_MMAP 258 if (!STRICT_UNWINDING) { 259 // Lax sanity checks cause a crash on AMD-based machines with 260 // VDSO-enabled kernels. 261 // Make an extra sanity check to insure new_sp is readable. 262 // Note: NextStackFrame<false>() is only called while the program 263 // is already on its last leg, so it's ok to be slow here. 264 static int page_size = getpagesize(); 265 void *new_sp_aligned = (void *)((uintptr_t)new_sp & ~(page_size - 1)); 266 if (msync(new_sp_aligned, page_size, MS_ASYNC) == -1) 267 return NULL; 268 } 269#endif 270 return new_sp; 271} 272 273#endif // BASE_STACKTRACE_X86_INL_H_ 274 275// Note: this part of the file is included several times. 276// Do not put globals below. 277 278// The following 4 functions are generated from the code below: 279// GetStack{Trace,Frames}() 280// GetStack{Trace,Frames}WithContext() 281// 282// These functions take the following args: 283// void** result: the stack-trace, as an array 284// int* sizes: the size of each stack frame, as an array 285// (GetStackFrames* only) 286// int max_depth: the size of the result (and sizes) array(s) 287// int skip_count: how many stack pointers to skip before storing in result 288// void* ucp: a ucontext_t* (GetStack{Trace,Frames}WithContext only) 289 290int GET_STACK_TRACE_OR_FRAMES { 291 void **sp; 292#if (__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 2) || __llvm__ 293 // __builtin_frame_address(0) can return the wrong address on gcc-4.1.0-k8. 294 // It's always correct on llvm, and the techniques below aren't (in 295 // particular, llvm-gcc will make a copy of pcs, so it's not in sp[2]), 296 // so we also prefer __builtin_frame_address when running under llvm. 297 sp = reinterpret_cast<void**>(__builtin_frame_address(0)); 298#elif defined(__i386__) 299 // Stack frame format: 300 // sp[0] pointer to previous frame 301 // sp[1] caller address 302 // sp[2] first argument 303 // ... 304 // NOTE: This will break under llvm, since result is a copy and not in sp[2] 305 sp = (void **)&result - 2; 306#elif defined(__x86_64__) 307 unsigned long rbp; 308 // Move the value of the register %rbp into the local variable rbp. 309 // We need 'volatile' to prevent this instruction from getting moved 310 // around during optimization to before function prologue is done. 311 // An alternative way to achieve this 312 // would be (before this __asm__ instruction) to call Noop() defined as 313 // static void Noop() __attribute__ ((noinline)); // prevent inlining 314 // static void Noop() { asm(""); } // prevent optimizing-away 315 __asm__ volatile ("mov %%rbp, %0" : "=r" (rbp)); 316 // Arguments are passed in registers on x86-64, so we can't just 317 // offset from &result 318 sp = (void **) rbp; 319#else 320# error Using stacktrace_x86-inl.h on a non x86 architecture! 321#endif 322 323 int n = 0; 324 while (sp && n < max_depth) { 325 if (*(sp+1) == reinterpret_cast<void *>(0)) { 326 // In 64-bit code, we often see a frame that 327 // points to itself and has a return address of 0. 328 break; 329 } 330#if !IS_WITH_CONTEXT 331 const void *const ucp = NULL; 332#endif 333 void **next_sp = NextStackFrame<!IS_STACK_FRAMES, IS_WITH_CONTEXT>(sp, ucp); 334 if (skip_count > 0) { 335 skip_count--; 336 } else { 337 result[n] = *(sp+1); 338#if IS_STACK_FRAMES 339 if (next_sp > sp) { 340 sizes[n] = (uintptr_t)next_sp - (uintptr_t)sp; 341 } else { 342 // A frame-size of 0 is used to indicate unknown frame size. 343 sizes[n] = 0; 344 } 345#endif 346 n++; 347 } 348 sp = next_sp; 349 } 350 return n; 351} 352