15821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Copyright (c) 2005, Google Inc. 25821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// All rights reserved. 35821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 45821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Redistribution and use in source and binary forms, with or without 55821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// modification, are permitted provided that the following conditions are 65821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// met: 75821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 85821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// * Redistributions of source code must retain the above copyright 95821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// notice, this list of conditions and the following disclaimer. 105821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// * Redistributions in binary form must reproduce the above 115821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// copyright notice, this list of conditions and the following disclaimer 125821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// in the documentation and/or other materials provided with the 135821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// distribution. 145821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// * Neither the name of Google Inc. nor the names of its 155821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// contributors may be used to endorse or promote products derived from 165821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// this software without specific prior written permission. 175821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 185821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 195821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 205821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 215821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 225821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 235821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 245821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 255821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 265821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 275821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 285821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 295821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 305821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// --- 315821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Author: Craig Silverstein 325821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 335821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// This is an internal header file used by profiler.cc. It defines 345821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// the single (inline) function GetPC. GetPC is used in a signal 355821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// handler to figure out the instruction that was being executed when 365821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// the signal-handler was triggered. 375821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 385821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// To get this, we use the ucontext_t argument to the signal-handler 395821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// callback, which holds the full context of what was going on when 405821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// the signal triggered. How to get from a ucontext_t to a Program 415821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Counter is OS-dependent. 425821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 435821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#ifndef BASE_GETPC_H_ 445821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#define BASE_GETPC_H_ 455821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 465821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "config.h" 475821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 485821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// On many linux systems, we may need _GNU_SOURCE to get access to 495821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// the defined constants that define the register we want to see (eg 505821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// REG_EIP). Note this #define must come first! 515821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#define _GNU_SOURCE 1 525821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// If #define _GNU_SOURCE causes problems, this might work instead. 535821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// It will cause problems for FreeBSD though!, because it turns off 545821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// the needed __BSD_VISIBLE. 555821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//#define _XOPEN_SOURCE 500 565821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 575821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include <string.h> // for memcmp 585821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#if defined(HAVE_SYS_UCONTEXT_H) 595821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include <sys/ucontext.h> 605821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#elif defined(HAVE_UCONTEXT_H) 615821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include <ucontext.h> // for ucontext_t (and also mcontext_t) 625821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#elif defined(HAVE_CYGWIN_SIGNAL_H) 635821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include <cygwin/signal.h> 645821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)typedef ucontext ucontext_t; 655821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#endif 665821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 675821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 685821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Take the example where function Foo() calls function Bar(). For 695821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// many architectures, Bar() is responsible for setting up and tearing 705821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// down its own stack frame. In that case, it's possible for the 715821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// interrupt to happen when execution is in Bar(), but the stack frame 725821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// is not properly set up (either before it's done being set up, or 735821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// after it's been torn down but before Bar() returns). In those 745821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// cases, the stack trace cannot see the caller function anymore. 755821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 765821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// GetPC can try to identify this situation, on architectures where it 775821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// might occur, and unwind the current function call in that case to 785821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// avoid false edges in the profile graph (that is, edges that appear 795821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// to show a call skipping over a function). To do this, we hard-code 805821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// in the asm instructions we might see when setting up or tearing 815821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// down a stack frame. 825821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 835821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// This is difficult to get right: the instructions depend on the 845821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// processor, the compiler ABI, and even the optimization level. This 855821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// is a best effort patch -- if we fail to detect such a situation, or 865821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// mess up the PC, nothing happens; the returned PC is not used for 875821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// any further processing. 885821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)struct CallUnrollInfo { 895821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // Offset from (e)ip register where this instruction sequence 905821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // should be matched. Interpreted as bytes. Offset 0 is the next 915821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // instruction to execute. Be extra careful with negative offsets in 925821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // architectures of variable instruction length (like x86) - it is 935821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // not that easy as taking an offset to step one instruction back! 945821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) int pc_offset; 955821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // The actual instruction bytes. Feel free to make it larger if you 965821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // need a longer sequence. 975821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) char ins[16]; 985821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // How many bytes to match from ins array? 995821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) int ins_size; 1005821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // The offset from the stack pointer (e)sp where to look for the 1015821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // call return address. Interpreted as bytes. 1025821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) int return_sp_offset; 1035821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)}; 1045821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 1055821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 1065821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// The dereferences needed to get the PC from a struct ucontext were 1075821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// determined at configure time, and stored in the macro 1085821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// PC_FROM_UCONTEXT in config.h. The only thing we need to do here, 1095821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// then, is to do the magic call-unrolling for systems that support it. 1105821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 1115821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// -- Special case 1: linux x86, for which we have CallUnrollInfo 1125821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#if defined(__linux) && defined(__i386) && defined(__GNUC__) 1135821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)static const CallUnrollInfo callunrollinfo[] = { 1145821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // Entry to a function: push %ebp; mov %esp,%ebp 1155821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // Top-of-stack contains the caller IP. 1165821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) { 0, 1175821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) {0x55, 0x89, 0xe5}, 3, 1185821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 0 1195821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) }, 1205821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // Entry to a function, second instruction: push %ebp; mov %esp,%ebp 1215821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // Top-of-stack contains the old frame, caller IP is +4. 1225821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) { -1, 1235821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) {0x55, 0x89, 0xe5}, 3, 1245821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 4 1255821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) }, 1265821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // Return from a function: RET. 1275821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // Top-of-stack contains the caller IP. 1285821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) { 0, 1295821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) {0xc3}, 1, 1305821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 0 1315821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) } 1325821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)}; 1335821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 1345821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)inline void* GetPC(const ucontext_t& signal_ucontext) { 1355821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // See comment above struct CallUnrollInfo. Only try instruction 1365821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // flow matching if both eip and esp looks reasonable. 1375821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) const int eip = signal_ucontext.uc_mcontext.gregs[REG_EIP]; 1385821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) const int esp = signal_ucontext.uc_mcontext.gregs[REG_ESP]; 1395821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) if ((eip & 0xffff0000) != 0 && (~eip & 0xffff0000) != 0 && 1405821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) (esp & 0xffff0000) != 0) { 1415821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) char* eip_char = reinterpret_cast<char*>(eip); 1425821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) for (int i = 0; i < sizeof(callunrollinfo)/sizeof(*callunrollinfo); ++i) { 1435821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) if (!memcmp(eip_char + callunrollinfo[i].pc_offset, 1445821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) callunrollinfo[i].ins, callunrollinfo[i].ins_size)) { 1455821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // We have a match. 1465821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) void **retaddr = (void**)(esp + callunrollinfo[i].return_sp_offset); 1475821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) return *retaddr; 1485821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) } 1495821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) } 1505821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) } 1515821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) return (void*)eip; 1525821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)} 1535821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 1545821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Special case #2: Windows, which has to do something totally different. 1555821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#elif defined(_WIN32) || defined(__CYGWIN__) || defined(__CYGWIN32__) || defined(__MINGW32__) 1565821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// If this is ever implemented, probably the way to do it is to have 1575821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// profiler.cc use a high-precision timer via timeSetEvent: 1585821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// http://msdn2.microsoft.com/en-us/library/ms712713.aspx 1595821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// We'd use it in mode TIME_CALLBACK_FUNCTION/TIME_PERIODIC. 1605821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// The callback function would be something like prof_handler, but 1615821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// alas the arguments are different: no ucontext_t! I don't know 1625821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// how we'd get the PC (using StackWalk64?) 1635821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// http://msdn2.microsoft.com/en-us/library/ms680650.aspx 1645821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 1655821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "base/logging.h" // for RAW_LOG 1665821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#ifndef HAVE_CYGWIN_SIGNAL_H 1675821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)typedef int ucontext_t; 1685821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#endif 1695821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 1705821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)inline void* GetPC(const struct ucontext_t& signal_ucontext) { 1715821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) RAW_LOG(ERROR, "GetPC is not yet implemented on Windows\n"); 1725821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) return NULL; 1735821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)} 1745821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 1755821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Normal cases. If this doesn't compile, it's probably because 1765821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// PC_FROM_UCONTEXT is the empty string. You need to figure out 1775821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// the right value for your system, and add it to the list in 1785821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// configure.ac (or set it manually in your config.h). 1795821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#else 1805821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)inline void* GetPC(const ucontext_t& signal_ucontext) { 1815821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) return (void*)signal_ucontext.PC_FROM_UCONTEXT; // defined in config.h 1825821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)} 1835821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 1845821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#endif 1855821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 1865821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#endif // BASE_GETPC_H_ 187