// Copyright (c) 2009, Google Inc.
// All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// * Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
// * Redistributions in binary form must reproduce the above
// copyright notice, this list of conditions and the following disclaimer
// in the documentation and/or other materials provided with the
// distribution.
// * Neither the name of Google Inc. nor the names of its
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
// ---
// Author: Craig Silverstein
//
// This forks out to pprof to do the actual symbolizing. We might
// be better off writing our own in C++.
#include "config.h"
#include "symbolize.h"
#include <stdlib.h>
#ifdef HAVE_UNISTD_H
#include <unistd.h> // for write()
#endif
#ifdef HAVE_SYS_SOCKET_H
#include <sys/socket.h> // for socketpair() -- needed by Symbolize
#endif
#ifdef HAVE_SYS_WAIT_H
#include <sys/wait.h> // for wait() -- needed by Symbolize
#endif
#ifdef HAVE_POLL_H
#include <poll.h>
#endif
#ifdef __MACH__
#include <mach-o/dyld.h> // for GetProgramInvocationName()
#include <limits.h> // for PATH_MAX
#endif
#if defined(__CYGWIN__) || defined(__CYGWIN32__)
#include <io.h> // for get_osfhandle()
#endif
#include <string>
#include "base/commandlineflags.h"
#include "base/logging.h"
#include "base/sysinfo.h"
using std::string;
using tcmalloc::DumpProcSelfMaps; // from sysinfo.h
DEFINE_string(symbolize_pprof,
EnvToString("PPROF_PATH", "pprof"),
"Path to pprof to call for reporting function names.");
// heap_profile_table_pprof may be referenced after destructors are
// called (since that's when leak-checking is done), so we make
// a more-permanent copy that won't ever get destroyed.
static string* g_pprof_path = new string(FLAGS_symbolize_pprof);
// Returns NULL if we're on an OS where we can't get the invocation name.
// Using a static var is ok because we're not called from a thread.
static char* GetProgramInvocationName() {
#if defined(HAVE_PROGRAM_INVOCATION_NAME)
extern char* program_invocation_name; // gcc provides this
return program_invocation_name;
#elif defined(__MACH__)
// We don't want to allocate memory for this since we may be
// calculating it when memory is corrupted.
static char program_invocation_name[PATH_MAX];
if (program_invocation_name[0] == '\0') { // first time calculating
uint32_t length = sizeof(program_invocation_name);
if (_NSGetExecutablePath(program_invocation_name, &length))
return NULL;
}
return program_invocation_name;
#else
return NULL; // figure out a way to get argv[0]
#endif
}
// Prints an error message when you can't run Symbolize().
static void PrintError(const char* reason) {
RAW_LOG(ERROR,
"*** WARNING: Cannot convert addresses to symbols in output below.\n"
"*** Reason: %s\n"
"*** If you cannot fix this, try running pprof directly.\n",
reason);
}
void SymbolTable::Add(const void* addr) {
symbolization_table_[addr] = "";
}
const char* SymbolTable::GetSymbol(const void* addr) {
return symbolization_table_[addr];
}
// Updates symbolization_table with the pointers to symbol names corresponding
// to its keys. The symbol names are stored in out, which is allocated and
// freed by the caller of this routine.
// Note that the forking/etc is not thread-safe or re-entrant. That's
// ok for the purpose we need -- reporting leaks detected by heap-checker
// -- but be careful if you decide to use this routine for other purposes.
// Returns number of symbols read on error. If can't symbolize, returns 0
// and emits an error message about why.
int SymbolTable::Symbolize() {
#if !defined(HAVE_UNISTD_H) || !defined(HAVE_SYS_SOCKET_H) || !defined(HAVE_SYS_WAIT_H)
PrintError("Perftools does not know how to call a sub-process on this O/S");
return 0;
#else
const char* argv0 = GetProgramInvocationName();
if (argv0 == NULL) { // can't call symbolize if we can't figure out our name
PrintError("Cannot figure out the name of this executable (argv0)");
return 0;
}
if (access(g_pprof_path->c_str(), R_OK) != 0) {
PrintError("Cannot find 'pprof' (is PPROF_PATH set correctly?)");
return 0;
}
// All this work is to do two-way communication. ugh.
int *child_in = NULL; // file descriptors
int *child_out = NULL; // for now, we don't worry about child_err
int child_fds[5][2]; // socketpair may be called up to five times below
// The client program may close its stdin and/or stdout and/or stderr
// thus allowing socketpair to reuse file descriptors 0, 1 or 2.
// In this case the communication between the forked processes may be broken
// if either the parent or the child tries to close or duplicate these
// descriptors. The loop below produces two pairs of file descriptors, each
// greater than 2 (stderr).
for (int i = 0; i < 5; i++) {
if (socketpair(AF_UNIX, SOCK_STREAM, 0, child_fds[i]) == -1) {
for (int j = 0; j < i; j++) {
close(child_fds[j][0]);
close(child_fds[j][1]);
PrintError("Cannot create a socket pair");
return 0;
}
} else {
if ((child_fds[i][0] > 2) && (child_fds[i][1] > 2)) {
if (child_in == NULL) {
child_in = child_fds[i];
} else {
child_out = child_fds[i];
for (int j = 0; j < i; j++) {
if (child_fds[j] == child_in) continue;
close(child_fds[j][0]);
close(child_fds[j][1]);
}
break;
}
}
}
}
switch (fork()) {
case -1: { // error
close(child_in[0]);
close(child_in[1]);
close(child_out[0]);
close(child_out[1]);
PrintError("Unknown error calling fork()");
return 0;
}
case 0: { // child
close(child_in[1]); // child uses the 0's, parent uses the 1's
close(child_out[1]); // child uses the 0's, parent uses the 1's
close(0);
close(1);
if (dup2(child_in[0], 0) == -1) _exit(1);
if (dup2(child_out[0], 1) == -1) _exit(2);
// Unset vars that might cause trouble when we fork
unsetenv("CPUPROFILE");
unsetenv("HEAPPROFILE");
unsetenv("HEAPCHECK");
unsetenv("PERFTOOLS_VERBOSE");
execlp(g_pprof_path->c_str(), g_pprof_path->c_str(),
"--symbols", argv0, NULL);
_exit(3); // if execvp fails, it's bad news for us
}
default: { // parent
close(child_in[0]); // child uses the 0's, parent uses the 1's
close(child_out[0]); // child uses the 0's, parent uses the 1's
#ifdef HAVE_POLL_H
// Waiting for 1ms seems to give the OS time to notice any errors.
poll(0, 0, 1);
// For maximum safety, we check to make sure the execlp
// succeeded before trying to write. (Otherwise we'll get a
// SIGPIPE.) For systems without poll.h, we'll just skip this
// check, and trust that the user set PPROF_PATH correctly!
struct pollfd pfd = { child_in[1], POLLOUT, 0 };
if (!poll(&pfd, 1, 0) || !(pfd.revents & POLLOUT) ||
(pfd.revents & (POLLHUP|POLLERR))) {
PrintError("Cannot run 'pprof' (is PPROF_PATH set correctly?)");
return 0;
}
#endif
#if defined(__CYGWIN__) || defined(__CYGWIN32__)
// On cygwin, DumpProcSelfMaps() takes a HANDLE, not an fd. Convert.
const HANDLE symbols_handle = (HANDLE) get_osfhandle(child_in[1]);
DumpProcSelfMaps(symbols_handle);
#else
DumpProcSelfMaps(child_in[1]); // what pprof expects on stdin
#endif
// Allocate 24 bytes = ("0x" + 8 bytes + "\n" + overhead) for each
// address to feed to pprof.
const int kOutBufSize = 24 * symbolization_table_.size();
char *pprof_buffer = new char[kOutBufSize];
int written = 0;
for (SymbolMap::const_iterator iter = symbolization_table_.begin();
iter != symbolization_table_.end(); ++iter) {
written += snprintf(pprof_buffer + written, kOutBufSize - written,
// pprof expects format to be 0xXXXXXX
"0x%" PRIxPTR "\n", reinterpret_cast<uintptr_t>(iter->first));
}
write(child_in[1], pprof_buffer, strlen(pprof_buffer));
close(child_in[1]); // that's all we need to write
const int kSymbolBufferSize = kSymbolSize * symbolization_table_.size();
int total_bytes_read = 0;
delete[] symbol_buffer_;
symbol_buffer_ = new char[kSymbolBufferSize];
memset(symbol_buffer_, '\0', kSymbolBufferSize);
while (1) {
int bytes_read = read(child_out[1], symbol_buffer_ + total_bytes_read,
kSymbolBufferSize - total_bytes_read);
if (bytes_read < 0) {
close(child_out[1]);
PrintError("Cannot read data from pprof");
return 0;
} else if (bytes_read == 0) {
close(child_out[1]);
wait(NULL);
break;
} else {
total_bytes_read += bytes_read;
}
}
// We have successfully read the output of pprof into out. Make sure
// the last symbol is full (we can tell because it ends with a \n).
if (total_bytes_read == 0 || symbol_buffer_[total_bytes_read - 1] != '\n')
return 0;
// make the symbolization_table_ values point to the output vector
SymbolMap::iterator fill = symbolization_table_.begin();
int num_symbols = 0;
const char *current_name = symbol_buffer_;
for (int i = 0; i < total_bytes_read; i++) {
if (symbol_buffer_[i] == '\n') {
fill->second = current_name;
symbol_buffer_[i] = '\0';
current_name = symbol_buffer_ + i + 1;
fill++;
num_symbols++;
}
}
return num_symbols;
}
}
PrintError("Unkown error (should never occur!)");
return 0; // shouldn't be reachable
#endif
}