rocksdb/port/stack_trace.cc

337 lines
11 KiB
C++
Raw Normal View History

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
#include "port/stack_trace.h"
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
#if !(defined(ROCKSDB_BACKTRACE) || defined(OS_MACOSX)) || defined(CYGWIN) || \
defined(OS_SOLARIS) || defined(OS_WIN)
// noop
namespace ROCKSDB_NAMESPACE {
namespace port {
void InstallStackTraceHandler() {}
void PrintStack(int /*first_frames_to_skip*/) {}
void PrintAndFreeStack(void* /*callstack*/, int /*num_frames*/) {}
void* SaveStack(int* /*num_frames*/, int /*first_frames_to_skip*/) {
return nullptr;
}
} // namespace port
} // namespace ROCKSDB_NAMESPACE
#else
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
#include <cxxabi.h>
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
#include <execinfo.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#ifdef OS_FREEBSD
#include <sys/sysctl.h>
#endif // OS_FREEBSD
#ifdef OS_LINUX
#include <sys/prctl.h>
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
#include <sys/types.h>
#include <sys/wait.h>
#if __GLIBC__ < 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ < 30)
#include <sys/syscall.h>
#define gettid() syscall(SYS_gettid)
#endif // GLIBC version
#endif // OS_LINUX
#include "port/lang.h"
namespace ROCKSDB_NAMESPACE {
namespace port {
namespace {
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
#if defined(OS_LINUX) || defined(OS_FREEBSD) || defined(OS_GNU_KFREEBSD)
const char* GetExecutableName() {
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
static char name[1024];
#if !defined(OS_FREEBSD)
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
char link[1024];
snprintf(link, sizeof(link), "/proc/%d/exe", getpid());
auto read = readlink(link, name, sizeof(name) - 1);
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
if (-1 == read) {
return nullptr;
} else {
name[read] = 0;
return name;
}
#else
int mib[4] = {CTL_KERN, KERN_PROC, KERN_PROC_PATHNAME, -1};
size_t namesz = sizeof(name);
auto ret = sysctl(mib, 4, name, &namesz, nullptr, 0);
if (-1 == ret) {
return nullptr;
} else {
return name;
}
#endif
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
}
void PrintStackTraceLine(const char* symbol, void* frame) {
static const char* executable = GetExecutableName();
if (symbol) {
fprintf(stderr, "%s ", symbol);
}
if (executable) {
// out source to addr2line, for the address translation
const int kLineMax = 256;
char cmd[kLineMax];
snprintf(cmd, kLineMax, "addr2line %p -e %s -f -C 2>&1", frame, executable);
auto f = popen(cmd, "r");
if (f) {
char line[kLineMax];
while (fgets(line, sizeof(line), f)) {
line[strlen(line) - 1] = 0; // remove newline
fprintf(stderr, "%s\t", line);
}
pclose(f);
}
} else {
fprintf(stderr, " %p", frame);
}
fprintf(stderr, "\n");
}
#elif defined(OS_MACOSX)
void PrintStackTraceLine(const char* symbol, void* frame) {
static int pid = getpid();
// out source to atos, for the address translation
const int kLineMax = 256;
char cmd[kLineMax];
2015-02-05 00:24:02 +00:00
snprintf(cmd, kLineMax, "xcrun atos %p -p %d 2>&1", frame, pid);
auto f = popen(cmd, "r");
if (f) {
char line[kLineMax];
while (fgets(line, sizeof(line), f)) {
line[strlen(line) - 1] = 0; // remove newline
fprintf(stderr, "%s\t", line);
}
pclose(f);
} else if (symbol) {
fprintf(stderr, "%s ", symbol);
}
fprintf(stderr, "\n");
}
#endif
const char* GetLldbScriptSelectThread(long long tid) {
// NOTE: called from a signal handler, so no heap allocation
static char script[80];
snprintf(script, sizeof(script),
"script -l python -- lldb.process.SetSelectedThreadByID(%lld)", tid);
return script;
}
} // namespace
void PrintStack(void* frames[], int num_frames) {
auto symbols = backtrace_symbols(frames, num_frames);
for (int i = 0; i < num_frames; ++i) {
fprintf(stderr, "#%-2d ", i);
PrintStackTraceLine((symbols != nullptr) ? symbols[i] : nullptr, frames[i]);
}
free(symbols);
}
void PrintStack(int first_frames_to_skip) {
// Default to getting stack traces with GDB, at least on Linux where we
// know how to attach to a particular thread.
//
// * Address space layout randomization (ASLR) interferes with getting good
// stack information from backtrace+addr2line. This is more likely to show
// up with LIB_MODE=shared builds (when kernel.randomize_va_space >= 1)
// but can also show up with LIB_MODE=static builds ((when
// kernel.randomize_va_space == 2).
// * It doesn't appear easy to detect when ASLR is in use.
// * With DEBUG_LEVEL < 2, backtrace() can skip frames that are not skipped
// in GDB.
//
// LLDB also available as an option
bool lldb_stack_trace = getenv("ROCKSDB_LLDB_STACK") != nullptr;
#if defined(OS_LINUX)
// Default true, override with ROCKSDB_BACKTRACE_STACK=1
bool gdb_stack_trace =
!lldb_stack_trace && getenv("ROCKSDB_BACKTRACE_STACK") == nullptr;
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
#else
// Default false, override with ROCKSDB_GDB_STACK=1
bool gdb_stack_trace = getenv("ROCKSDB_GDB_STACK") != nullptr;
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
#endif
// Also support invoking interactive debugger on stack trace, with this
// envvar set to non-empty
char* debug_env = getenv("ROCKSDB_DEBUG");
bool debug = debug_env != nullptr && strlen(debug_env) > 0;
if (lldb_stack_trace || gdb_stack_trace || debug) {
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
// Allow ouside debugger to attach, even with Yama security restrictions
#ifdef PR_SET_PTRACER_ANY
(void)prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY, 0, 0, 0);
#endif
// Try to invoke GDB, either for stack trace or debugging.
long long attach_pid = getpid();
// NOTE: we're in a signal handler, so no heap allocation
char attach_pid_str[20];
snprintf(attach_pid_str, sizeof(attach_pid_str), "%lld", attach_pid);
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
// `gdb -p PID` seems to always attach to main thread, but `gdb -p TID`
// seems to be able to attach to a particular thread in a process, which
// makes sense as the main thread TID == PID of the process.
// But I haven't found that gdb capability documented anywhere, so leave
// a back door to attach to main thread.
long long gdb_attach_id = attach_pid;
// Save current thread id before fork
long long attach_tid = 0;
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
#ifdef OS_LINUX
attach_tid = gettid();
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
if (getenv("ROCKSDB_DEBUG_USE_PID") == nullptr) {
gdb_attach_id = attach_tid;
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
}
#endif
char gdb_attach_id_str[20];
snprintf(gdb_attach_id_str, sizeof(gdb_attach_id_str), "%lld",
gdb_attach_id);
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
pid_t child_pid = fork();
if (child_pid == 0) {
// child process
if (debug) {
if (strcmp(debug_env, "lldb") == 0) {
fprintf(stderr, "Invoking LLDB for debugging (ROCKSDB_DEBUG=%s)...\n",
debug_env);
execlp(/*cmd in PATH*/ "lldb", /*arg0*/ "lldb", "-p", attach_pid_str,
/*"-Q",*/ "-o", GetLldbScriptSelectThread(attach_tid),
(char*)nullptr);
return;
} else {
fprintf(stderr, "Invoking GDB for debugging (ROCKSDB_DEBUG=%s)...\n",
debug_env);
execlp(/*cmd in PATH*/ "gdb", /*arg0*/ "gdb", "-p", gdb_attach_id_str,
(char*)nullptr);
return;
}
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
} else {
// Redirect child stdout to original stderr
dup2(2, 1);
// No child stdin (don't use pager)
close(0);
if (lldb_stack_trace) {
fprintf(stderr, "Invoking LLDB for stack trace...\n");
// Skip top ~8 frames here in PrintStack
auto bt_in_lldb =
"script -l python -- for f in lldb.thread.frames[8:]: print(f)";
execlp(/*cmd in PATH*/ "lldb", /*arg0*/ "lldb", "-p", attach_pid_str,
"-b", "-Q", "-o", GetLldbScriptSelectThread(attach_tid), "-o",
bt_in_lldb, (char*)nullptr);
} else {
// gdb_stack_trace
fprintf(stderr, "Invoking GDB for stack trace...\n");
// Skip top ~4 frames here in PrintStack
// See https://stackoverflow.com/q/40991943/454544
auto bt_in_gdb =
"frame apply level 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 "
"21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 "
"42 43 44 -q frame";
// -n : Loading config files can apparently cause failures with the
// other options here.
// -batch : non-interactive; suppress banners as much as possible
execlp(/*cmd in PATH*/ "gdb", /*arg0*/ "gdb", "-n", "-batch", "-p",
gdb_attach_id_str, "-ex", bt_in_gdb, (char*)nullptr);
}
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
return;
}
} else {
// parent process; wait for child
int wstatus;
waitpid(child_pid, &wstatus, 0);
if (WIFEXITED(wstatus) && WEXITSTATUS(wstatus) == EXIT_SUCCESS) {
// Good
return;
}
}
fprintf(stderr, "GDB failed; falling back on backtrace+addr2line...\n");
}
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
const int kMaxFrames = 100;
void* frames[kMaxFrames];
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
auto num_frames = backtrace(frames, kMaxFrames);
PrintStack(&frames[first_frames_to_skip], num_frames - first_frames_to_skip);
}
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
void PrintAndFreeStack(void* callstack, int num_frames) {
PrintStack(static_cast<void**>(callstack), num_frames);
free(callstack);
}
void* SaveStack(int* num_frames, int first_frames_to_skip) {
const int kMaxFrames = 100;
void* frames[kMaxFrames];
auto count = backtrace(frames, kMaxFrames);
*num_frames = count - first_frames_to_skip;
void* callstack = malloc(sizeof(void*) * *num_frames);
memcpy(callstack, &frames[first_frames_to_skip], sizeof(void*) * *num_frames);
return callstack;
}
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
static void StackTraceHandler(int sig) {
// reset to default handler
signal(sig, SIG_DFL);
fprintf(stderr, "Received signal %d (%s)\n", sig, strsignal(sig));
// skip the top three signal handler related frames
PrintStack(3);
// Efforts to fix or suppress TSAN warnings "signal-unsafe call inside of
// a signal" have failed, so just warn the user about them.
#ifdef __SANITIZE_THREAD__
fprintf(stderr,
"==> NOTE: any above warnings about \"signal-unsafe call\" are\n"
"==> ignorable, as they are expected when generating a stack\n"
"==> trace because of a signal under TSAN. Consider why the\n"
"==> signal was generated to begin with, and the stack trace\n"
"==> in the TSAN warning can be useful for that. (The stack\n"
"==> trace printed by the signal handler is likely obscured\n"
"==> by TSAN output.)\n");
#endif
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
// re-signal to default handler (so we still get core dump if needed...)
raise(sig);
}
void InstallStackTraceHandler() {
// just use the plain old signal as it's simple and sufficient
// for this use case
signal(SIGILL, StackTraceHandler);
signal(SIGSEGV, StackTraceHandler);
signal(SIGBUS, StackTraceHandler);
signal(SIGABRT, StackTraceHandler);
Support stack traces with gdb (and debugger invocation) (#11150) Summary: LIB_MODE=shared is much more efficient for building all the unit tests but comes with the downside of ugly stack traces, generally missing name demangling and source line info. Searching the internet suggests the reliable way to get stack traces with dynamic loading is with gdb. This change automatically tries to use gdb to get a stack trace if built with LIB_MODE=shared, and only on Linux because that's where we have the capability to attach to the proper thread. (We could revise the exact conditions in the future.) If there's a failure invoking gdb, it falls back on the old method. Obscure details of making the output reasonable / pretty are in the source code comments. Based on this, it was easy to make it so that running a test command with ROCKSDB_DEBUG=1 would invoke gdb whenever the stack trace handler was invoked, so I included that. Intended follow-up: make LIB_MODE=shared the new default `make` build config Pull Request resolved: https://github.com/facebook/rocksdb/pull/11150 Test Plan: manual, mostly by injecting an "assert(false)" into a unit test and trying different build modes etc. Although gdb is slower to start showing stack trace output, it seems overall faster in many if not most cases, presumably because it doesn't reload the symbol table for each stack entry. At least with parallel test runs, having many tests dumping stacks with the old method can take so long it appears to hang the test run. Reviewed By: cbi42 Differential Revision: D42894064 Pulled By: pdillinger fbshipit-source-id: 608143309d8c69c40049c9a4abcde4f22e87b4d8
2023-02-03 21:21:03 +00:00
// Allow ouside debugger to attach, even with Yama security restrictions.
// This is needed even outside of PrintStack() so that external mechanisms
// can dump stacks if they suspect that a test has hung.
#ifdef PR_SET_PTRACER_ANY
(void)prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY, 0, 0, 0);
#endif
[RocksDB] Add stacktrace signal handler Summary: This diff provides the ability to print out a stacktrace when the process receives certain signals. Currently, we enable this for the following signals (program error related): SIGILL SIGSEGV SIGBUS SIGABRT Application simply #include "util/stack_trace.h" and call leveldb::InstallStackTraceHandler() during initialization, if signal handler is needed. It's not done automatically when openning db, because it's the application(process)'s responsibility to install signal handler and some applications might already have their own (like fbcode). Sample output: Received signal 11 (Segmentation fault) #0 0x408ff0 ./signal_test() [0x408ff0] /home/haobo/rocksdb/util/signal_test.cc:4 #1 0x40827d ./signal_test() [0x40827d] /home/haobo/rocksdb/util/signal_test.cc:24 #2 0x7f8bb183172e /usr/local/fbcode/gcc-4.7.1-glibc-2.14.1/lib/libc.so.6(__libc_start_main+0x10e) [0x7f8bb183172e] ??:0 #3 0x408ebc ./signal_test() [0x408ebc] /home/engshare/third-party/src/glibc/glibc-2.14.1/glibc-2.14.1/csu/../sysdeps/x86_64/elf/start.S:113 Segmentation fault (core dumped) For each frame, we print the raw pointer, the symbol provided by backtrace_symbols (still not good enough), and the source file/line. Note that address translation is done by directly shell out to addr2line. ??:0 means addr2line fails to do the translation. Hacky, but I think it's good for now. Test Plan: signal_test.cc Reviewers: dhruba, MarkCallaghan Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D10173
2013-04-11 17:54:35 +00:00
}
} // namespace port
} // namespace ROCKSDB_NAMESPACE
#endif