ClickHouse/src/Common/StackTrace.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

466 lines
15 KiB
C++
Raw Normal View History

2023-03-02 22:27:54 +00:00
#include "StackTrace.h"
#include <base/FnTraits.h>
#include <base/constexpr_helpers.h>
#include <base/demangle.h>
2019-07-29 22:26:44 +00:00
#include <Common/Dwarf.h>
#include <Common/Elf.h>
2020-06-01 18:10:19 +00:00
#include <Common/MemorySanitizer.h>
2023-03-02 22:27:54 +00:00
#include <Common/SymbolIndex.h>
2023-05-06 16:14:01 +00:00
#include <IO/WriteBufferFromString.h>
#include <IO/WriteHelpers.h>
#include <IO/Operators.h>
#include <atomic>
2019-07-29 22:26:44 +00:00
#include <filesystem>
2023-03-02 22:27:54 +00:00
#include <map>
#include <mutex>
#include <sstream>
2019-07-29 22:26:44 +00:00
#include <unordered_map>
2023-03-02 22:27:54 +00:00
#include <fmt/format.h>
2019-06-28 18:06:38 +00:00
#include "config.h"
#if USE_UNWIND
# include <libunwind.h>
#endif
2019-06-28 18:06:38 +00:00
namespace
{
2023-03-02 22:27:54 +00:00
/// Currently this variable is set up once on server startup.
/// But we use atomic just in case, so it is possible to be modified at runtime.
std::atomic<bool> show_addresses = true;
2022-07-29 23:45:06 +00:00
2023-03-02 22:27:54 +00:00
bool shouldShowAddress(const void * addr)
{
/// If the address is less than 4096, most likely it is a nullptr dereference with offset,
/// and showing this offset is secure nevertheless.
/// NOTE: 4096 is the page size on x86 and it can be different on other systems,
/// but for the purpose of this branch, it does not matter.
if (reinterpret_cast<uintptr_t>(addr) < 4096)
return true;
return show_addresses.load(std::memory_order_relaxed);
}
}
void StackTrace::setShowAddresses(bool show)
{
show_addresses.store(show, std::memory_order_relaxed);
}
2023-03-02 22:27:54 +00:00
std::string SigsegvErrorString(const siginfo_t & info, [[maybe_unused]] const ucontext_t & context)
2019-06-28 18:06:38 +00:00
{
2023-03-02 22:27:54 +00:00
using namespace std::string_literals;
std::string address
= info.si_addr == nullptr ? "NULL pointer"s : (shouldShowAddress(info.si_addr) ? fmt::format("{}", info.si_addr) : ""s);
2019-06-28 18:06:38 +00:00
2023-03-02 22:27:54 +00:00
const std::string_view access =
#if defined(__x86_64__) && !defined(OS_FREEBSD) && !defined(OS_DARWIN) && !defined(__arm__) && !defined(__powerpc__)
2023-03-02 22:27:54 +00:00
(context.uc_mcontext.gregs[REG_ERR] & 0x02) ? "write" : "read";
#else
"";
2019-06-28 18:06:38 +00:00
#endif
2023-03-02 22:27:54 +00:00
std::string_view message;
switch (info.si_code)
{
case SEGV_ACCERR:
message = "Attempted access has violated the permissions assigned to the memory area";
2019-06-28 18:06:38 +00:00
break;
2023-03-02 22:27:54 +00:00
case SEGV_MAPERR:
message = "Address not mapped to object";
break;
default:
message = "Unknown si_code";
break;
}
2019-06-28 18:06:38 +00:00
2023-03-02 22:27:54 +00:00
return fmt::format("Address: {}. Access: {}. {}.", std::move(address), access, message);
}
constexpr std::string_view SigbusErrorString(int si_code)
{
switch (si_code)
{
case BUS_ADRALN:
return "Invalid address alignment.";
case BUS_ADRERR:
return "Non-existent physical address.";
case BUS_OBJERR:
return "Object specific hardware error.";
// Linux specific
2019-06-28 18:06:38 +00:00
#if defined(BUS_MCEERR_AR)
2023-03-02 22:27:54 +00:00
case BUS_MCEERR_AR:
return "Hardware memory error: action required.";
2019-06-28 18:06:38 +00:00
#endif
#if defined(BUS_MCEERR_AO)
2023-03-02 22:27:54 +00:00
case BUS_MCEERR_AO:
return "Hardware memory error: action optional.";
2019-06-28 18:06:38 +00:00
#endif
2023-03-02 22:27:54 +00:00
default:
return "Unknown si_code.";
}
}
2019-06-28 18:06:38 +00:00
2023-03-02 22:27:54 +00:00
constexpr std::string_view SigfpeErrorString(int si_code)
{
switch (si_code)
{
case FPE_INTDIV:
return "Integer divide by zero.";
case FPE_INTOVF:
return "Integer overflow.";
case FPE_FLTDIV:
return "Floating point divide by zero.";
case FPE_FLTOVF:
return "Floating point overflow.";
case FPE_FLTUND:
return "Floating point underflow.";
case FPE_FLTRES:
return "Floating point inexact result.";
case FPE_FLTINV:
return "Floating point invalid operation.";
case FPE_FLTSUB:
return "Subscript out of range.";
default:
return "Unknown si_code.";
}
}
2019-06-28 18:06:38 +00:00
2023-03-02 22:27:54 +00:00
constexpr std::string_view SigillErrorString(int si_code)
{
switch (si_code)
{
case ILL_ILLOPC:
return "Illegal opcode.";
case ILL_ILLOPN:
return "Illegal operand.";
case ILL_ILLADR:
return "Illegal addressing mode.";
case ILL_ILLTRP:
return "Illegal trap.";
case ILL_PRVOPC:
return "Privileged opcode.";
case ILL_PRVREG:
return "Privileged register.";
case ILL_COPROC:
return "Coprocessor error.";
case ILL_BADSTK:
return "Internal stack error.";
default:
return "Unknown si_code.";
}
}
2019-06-28 18:06:38 +00:00
2023-03-02 22:27:54 +00:00
std::string signalToErrorMessage(int sig, const siginfo_t & info, [[maybe_unused]] const ucontext_t & context)
{
switch (sig)
{
case SIGSEGV:
return SigsegvErrorString(info, context);
case SIGBUS:
return std::string{SigbusErrorString(info.si_code)};
case SIGILL:
return std::string{SigillErrorString(info.si_code)};
2020-05-30 08:01:15 +00:00
case SIGFPE:
2023-03-02 22:27:54 +00:00
return std::string{SigfpeErrorString(info.si_code)};
2020-05-30 08:01:15 +00:00
case SIGTSTP:
2023-03-02 22:27:54 +00:00
return "This is a signal used for debugging purposes by the user.";
default:
return "";
2019-12-02 17:29:19 +00:00
}
2019-06-28 18:06:38 +00:00
}
static void * getCallerAddress(const ucontext_t & context)
2019-06-28 18:06:38 +00:00
{
#if defined(__x86_64__)
/// Get the address at the time the signal was raised from the RIP (x86-64)
# if defined(OS_FREEBSD)
2019-06-28 18:06:38 +00:00
return reinterpret_cast<void *>(context.uc_mcontext.mc_rip);
# elif defined(OS_DARWIN)
2019-06-28 18:06:38 +00:00
return reinterpret_cast<void *>(context.uc_mcontext->__ss.__rip);
2020-05-29 21:36:47 +00:00
# else
2019-06-28 18:06:38 +00:00
return reinterpret_cast<void *>(context.uc_mcontext.gregs[REG_RIP]);
2020-05-29 21:36:47 +00:00
# endif
#elif defined(OS_DARWIN) && defined(__aarch64__)
2021-03-11 17:12:52 +00:00
return reinterpret_cast<void *>(context.uc_mcontext->__ss.__pc);
#elif defined(OS_FREEBSD) && defined(__aarch64__)
2021-07-22 12:06:45 +00:00
return reinterpret_cast<void *>(context.uc_mcontext.mc_gpregs.gp_elr);
2019-06-28 18:06:38 +00:00
#elif defined(__aarch64__)
return reinterpret_cast<void *>(context.uc_mcontext.pc);
2022-08-19 18:45:51 +00:00
#elif defined(__powerpc64__) && defined(__linux__)
2021-04-01 07:44:26 +00:00
return reinterpret_cast<void *>(context.uc_mcontext.gp_regs[PT_NIP]);
2022-08-19 18:45:51 +00:00
#elif defined(__powerpc64__) && defined(__FreeBSD__)
return reinterpret_cast<void *>(context.uc_mcontext.mc_srr0);
#elif defined(__riscv)
return reinterpret_cast<void *>(context.uc_mcontext.__gregs[REG_PC]);
2023-01-20 21:16:55 +00:00
#elif defined(__s390x__)
return reinterpret_cast<void *>(context.uc_mcontext.psw.addr);
2019-07-29 22:26:44 +00:00
#else
2019-06-28 18:06:38 +00:00
return nullptr;
2019-07-29 22:26:44 +00:00
#endif
2019-06-28 18:06:38 +00:00
}
2023-03-02 22:27:54 +00:00
// FIXME: looks like this is used only for Sentry but duplicates the whole algo, maybe replace?
void StackTrace::symbolize(
2023-03-02 22:27:54 +00:00
const StackTrace::FramePointers & frame_pointers, [[maybe_unused]] size_t offset, size_t size, StackTrace::Frames & frames)
{
#if defined(__ELF__) && !defined(OS_FREEBSD)
2020-11-30 14:30:55 +00:00
auto symbol_index_ptr = DB::SymbolIndex::instance();
const DB::SymbolIndex & symbol_index = *symbol_index_ptr;
std::unordered_map<std::string, DB::Dwarf> dwarfs;
2020-05-29 21:36:47 +00:00
for (size_t i = 0; i < offset; ++i)
2020-06-10 13:30:12 +00:00
frames[i].virtual_addr = frame_pointers[i];
for (size_t i = offset; i < size; ++i)
{
2020-06-10 13:30:12 +00:00
StackTrace::Frame & current_frame = frames[i];
current_frame.virtual_addr = frame_pointers[i];
const auto * object = symbol_index.findObject(current_frame.virtual_addr);
uintptr_t virtual_offset = object ? uintptr_t(object->address_begin) : 0;
current_frame.physical_addr = reinterpret_cast<void *>(uintptr_t(current_frame.virtual_addr) - virtual_offset);
if (object)
{
current_frame.object = object->name;
2023-03-02 22:27:54 +00:00
if (std::error_code ec; std::filesystem::exists(current_frame.object.value(), ec) && !ec)
{
auto dwarf_it = dwarfs.try_emplace(object->name, object->elf).first;
DB::Dwarf::LocationInfo location;
std::vector<DB::Dwarf::SymbolizedFrame> inline_frames;
if (dwarf_it->second.findAddress(
uintptr_t(current_frame.physical_addr), location, DB::Dwarf::LocationInfoMode::FAST, inline_frames))
2020-05-29 21:36:47 +00:00
{
current_frame.file = location.file.toString();
current_frame.line = location.line;
}
}
}
else
current_frame.object = "?";
2023-03-02 22:27:54 +00:00
if (const auto * symbol = symbol_index.findSymbol(current_frame.virtual_addr))
current_frame.symbol = demangle(symbol->name);
else
current_frame.symbol = "?";
}
2020-05-29 21:36:47 +00:00
#else
for (size_t i = 0; i < size; ++i)
2020-06-10 13:30:12 +00:00
frames[i].virtual_addr = frame_pointers[i];
#endif
}
2019-07-01 22:11:11 +00:00
StackTrace::StackTrace(const ucontext_t & signal_context)
2019-06-28 18:06:38 +00:00
{
2019-07-01 22:11:11 +00:00
tryCapture();
2019-06-28 18:06:38 +00:00
/// This variable from signal handler is not instrumented by Memory Sanitizer.
__msan_unpoison(&signal_context, sizeof(signal_context));
void * caller_address = getCallerAddress(signal_context);
if (size == 0 && caller_address)
2019-07-01 22:11:11 +00:00
{
frame_pointers[0] = caller_address;
size = 1;
}
else
{
/// Skip excessive stack frames that we have created while finding stack trace.
for (size_t i = 0; i < size; ++i)
{
if (frame_pointers[i] == caller_address)
{
offset = i;
break;
}
}
2019-06-28 18:06:38 +00:00
}
}
2019-07-01 22:11:11 +00:00
void StackTrace::tryCapture()
{
#if USE_UNWIND
size = unw_backtrace(frame_pointers.data(), capacity);
__msan_unpoison(frame_pointers.data(), size * sizeof(frame_pointers[0]));
2023-03-02 22:27:54 +00:00
#else
size = 0;
2019-07-01 22:11:11 +00:00
#endif
}
2023-03-06 22:30:44 +00:00
/// ClickHouse uses bundled libc++ so type names will be the same on every system thus it's safe to hardcode them
2023-03-02 22:27:54 +00:00
constexpr std::pair<std::string_view, std::string_view> replacements[]
= {{"::__1", ""}, {"std::basic_string<char, std::char_traits<char>, std::allocator<char>>", "String"}};
String collapseNames(String && haystack)
2019-07-01 22:11:11 +00:00
{
2023-03-02 22:27:54 +00:00
// TODO: surely there is a written version already for better in place search&replace
for (auto [needle, to] : replacements)
{
size_t pos = 0;
while ((pos = haystack.find(needle, pos)) != std::string::npos)
{
haystack.replace(pos, needle.length(), to);
pos += to.length();
}
}
return haystack;
2019-06-28 18:06:38 +00:00
}
2023-03-02 22:27:54 +00:00
struct StackTraceRefTriple
2019-07-01 22:11:11 +00:00
{
2023-03-02 22:27:54 +00:00
const StackTrace::FramePointers & pointers;
size_t offset;
size_t size;
};
struct StackTraceTriple
{
StackTrace::FramePointers pointers;
size_t offset;
size_t size;
};
2019-06-28 18:06:38 +00:00
2023-03-02 22:27:54 +00:00
template <class T>
concept MaybeRef = std::is_same_v<T, StackTraceTriple> || std::is_same_v<T, StackTraceRefTriple>;
constexpr bool operator<(const MaybeRef auto & left, const MaybeRef auto & right)
{
2023-03-02 22:27:54 +00:00
return std::tuple{left.pointers, left.size, left.offset} < std::tuple{right.pointers, right.size, right.offset};
}
2023-03-02 22:27:54 +00:00
static void
toStringEveryLineImpl([[maybe_unused]] bool fatal, const StackTraceRefTriple & stack_trace, Fn<void(std::string_view)> auto && callback)
2019-06-28 18:06:38 +00:00
{
2023-03-02 22:27:54 +00:00
if (stack_trace.size == 0)
return callback("<Empty trace>");
2019-06-28 18:06:38 +00:00
#if defined(__ELF__) && !defined(OS_FREEBSD)
2023-03-02 22:27:54 +00:00
using enum DB::Dwarf::LocationInfoMode;
const auto mode = fatal ? FULL_WITH_INLINE : FAST;
2020-11-30 14:30:55 +00:00
auto symbol_index_ptr = DB::SymbolIndex::instance();
const DB::SymbolIndex & symbol_index = *symbol_index_ptr;
2023-03-02 22:27:54 +00:00
std::unordered_map<String, DB::Dwarf> dwarfs;
2020-06-22 07:22:13 +00:00
2023-03-02 22:27:54 +00:00
for (size_t i = stack_trace.offset; i < stack_trace.size; ++i)
2019-06-28 18:06:38 +00:00
{
std::vector<DB::Dwarf::SymbolizedFrame> inline_frames;
2023-03-02 22:27:54 +00:00
const void * virtual_addr = stack_trace.pointers[i];
2020-06-22 07:22:13 +00:00
const auto * object = symbol_index.findObject(virtual_addr);
uintptr_t virtual_offset = object ? uintptr_t(object->address_begin) : 0;
const void * physical_addr = reinterpret_cast<const void *>(uintptr_t(virtual_addr) - virtual_offset);
2023-05-06 18:54:52 +00:00
DB::WriteBufferFromOwnString out;
2020-01-28 14:39:24 +00:00
out << i << ". ";
2019-06-28 18:06:38 +00:00
2023-03-02 22:27:54 +00:00
if (std::error_code ec; object && std::filesystem::exists(object->name, ec) && !ec)
2019-07-29 22:26:44 +00:00
{
2023-03-02 22:27:54 +00:00
auto dwarf_it = dwarfs.try_emplace(object->name, object->elf).first;
2020-01-28 14:39:24 +00:00
2023-03-02 22:27:54 +00:00
DB::Dwarf::LocationInfo location;
2019-07-29 22:26:44 +00:00
2023-03-02 22:27:54 +00:00
if (dwarf_it->second.findAddress(uintptr_t(physical_addr), location, mode, inline_frames))
out << location.file.toString() << ":" << location.line << ": ";
}
2023-03-02 22:27:54 +00:00
if (const auto * const symbol = symbol_index.findSymbol(virtual_addr))
out << collapseNames(demangle(symbol->name));
2020-06-22 07:22:13 +00:00
else
out << "?";
2022-07-29 23:45:06 +00:00
if (shouldShowAddress(physical_addr))
2023-05-06 16:14:01 +00:00
{
out << " @ ";
DB::writePointerHex(physical_addr, out);
}
2020-06-22 07:22:13 +00:00
out << " in " << (object ? object->name : "?");
2019-08-21 00:48:34 +00:00
for (size_t j = 0; j < inline_frames.size(); ++j)
{
const auto & frame = inline_frames[j];
2023-03-02 22:27:54 +00:00
callback(fmt::format(
"{}.{}. inlined from {}:{}: {}",
i,
j + 1,
frame.location.file.toString(),
frame.location.line,
collapseNames(demangle(frame.name))));
}
2019-08-21 00:48:34 +00:00
callback(out.str());
}
2020-06-22 07:22:13 +00:00
#else
2023-03-02 22:27:54 +00:00
for (size_t i = stack_trace.offset; i < stack_trace.size; ++i)
if (const void * const addr = stack_trace.pointers[i]; shouldShowAddress(addr))
callback(fmt::format("{}. {}", i, addr));
2020-06-22 07:22:13 +00:00
#endif
}
2019-06-28 18:06:38 +00:00
2023-03-02 22:27:54 +00:00
void StackTrace::toStringEveryLine(std::function<void(std::string_view)> callback) const
{
2023-03-02 22:27:54 +00:00
toStringEveryLineImpl(true, {frame_pointers, offset, size}, std::move(callback));
2021-08-25 13:05:48 +00:00
}
2023-03-02 22:27:54 +00:00
using StackTraceCache = std::map<StackTraceTriple, String, std::less<>>;
static StackTraceCache & cacheInstance()
2021-08-28 03:43:43 +00:00
{
static StackTraceCache cache;
2021-08-28 03:43:43 +00:00
return cache;
}
static std::mutex stacktrace_cache_mutex;
2023-03-02 22:27:54 +00:00
String toStringCached(const StackTrace::FramePointers & pointers, size_t offset, size_t size)
2021-08-25 13:05:48 +00:00
{
/// Calculation of stack trace text is extremely slow.
/// We use simple cache because otherwise the server could be overloaded by trash queries.
/// Note that this cache can grow unconditionally, but practically it should be small.
std::lock_guard lock{stacktrace_cache_mutex};
2023-03-02 22:27:54 +00:00
StackTraceCache & cache = cacheInstance();
2023-03-04 18:37:28 +00:00
const StackTraceRefTriple key{pointers, offset, size};
2023-03-02 22:27:54 +00:00
2023-03-04 18:37:28 +00:00
if (auto it = cache.find(key); it != cache.end())
2023-03-02 22:27:54 +00:00
return it->second;
else
{
2023-05-06 16:14:01 +00:00
DB::WriteBufferFromOwnString out;
2023-03-04 18:37:28 +00:00
toStringEveryLineImpl(false, key, [&](std::string_view str) { out << str << '\n'; });
2023-03-02 22:27:54 +00:00
return cache.emplace(StackTraceTriple{pointers, offset, size}, out.str()).first->second;
}
}
std::string StackTrace::toString() const
{
return toStringCached(frame_pointers, offset, size);
}
std::string StackTrace::toString(void ** frame_pointers_raw, size_t offset, size_t size)
{
__msan_unpoison(frame_pointers_raw, size * sizeof(*frame_pointers_raw));
2023-03-04 18:37:28 +00:00
StackTrace::FramePointers frame_pointers{};
2023-03-02 22:27:54 +00:00
std::copy_n(frame_pointers_raw, size, frame_pointers.begin());
2023-03-02 22:27:54 +00:00
return toStringCached(frame_pointers, offset, size);
2021-08-28 03:43:43 +00:00
}
void StackTrace::dropCache()
{
std::lock_guard lock{stacktrace_cache_mutex};
cacheInstance().clear();
}