// ===========================================================================
// Overview (plain language)
// ===========================================================================
// Problem: when gbmcweb hits an uncaught C++ exception it crashes, but by the
// time the crash handler runs the stack has already unwound, so the usual
// backtrace points at framework code (asio / sdbusplus re-throwing a stashed
// exception) instead of the line that actually threw. That makes fleet crashes
// very hard to debug.
//
// Trick: every C++ `throw` funnels through one libstdc++ function, __cxa_throw.
// We define our own __cxa_throw in the main binary, so the linker makes the
// whole process (libstdc++ included) call ours. On each throw we snapshot the
// call stack *before* it unwinds, tag it with the exception object's address,
// and stash it in a small fixed ring. If the program later dies from an
// uncaught exception, our std::terminate handler finds that exception's
// snapshot and logs it -- the real throw site.
//
// How the pieces fit:
//   * Capture: backtrace() returns a list of raw code addresses ("here, called
//     from here, ..."), with no names.
//   * Store: a 64-slot circular buffer (the newest throw overwrites the oldest)
//     with no memory allocation. Each slot is tagged with the exception's
//     address. A lock-free seqlock lets the crash handler read a slot while
//     other threads keep throwing, without locks and without ever trusting a
//     half-written entry.
//   * Report at crash: look up the dying exception's snapshot and print each
//     address as "module + hex offset".
//   * Decode later: a developer runs the standard `addr2line` tool on the
//     matching debug binary to turn "module + offset" into file:line. We do not
//     resolve names at crash time -- that needs locks/allocation and the
//     process may be corrupt -- so at startup we snapshot once which address
//     ranges belong to which loaded binary.
//
// Why so defensive (no allocation, no locks, no re-entry): the handler runs
// when the process may already be corrupt; anything that could block, allocate,
// or crash again would lose the log entirely.
//
// Scope: built only in the gcc/yocto firmware image, since it relies on GNU
// libstdc++ exception internals; other builds compile it to "unavailable". It
// works on x86-64, ARM64 and ARM32 (it uses the standard stack unwinder, not a
// CPU-specific hack, so -fomit-frame-pointer and optimization levels do not
// affect it).
//
// The links below point to the specs and mechanisms this file relies on; the
// rationale now lives as comments next to the code each point describes.
// ===========================================================================

// References (formats and mechanisms this file depends on):
//  - Itanium C++ ABI, Exception Handling — __cxa_throw, the __cxa_exception
//    header layout, __cxa_current_exception_type, and the exception_ptr /
//    primary-object model this code keys on:
//    https://itanium-cxx-abi.github.io/cxx-abi/abi-eh.html
//  - folly exception_tracer — the __cxa_throw interposer this file reduces to a
//    single self-contained TU:
//    https://github.com/facebook/folly/blob/main/folly/debugging/exception_tracer/ExceptionTracerLib.cpp
//  - glibc backtrace(3) — the throw-site stack capture primitive:
//    https://man7.org/linux/man-pages/man3/backtrace.3.html
//  - dl_iterate_phdr(3) — source of the ELF program-header module table
//    snapshotted at install time:
//    https://man7.org/linux/man-pages/man3/dl_iterate_phdr.3.html
//  - Seqlock — the lock-free single-writer scheme each ring slot uses:
//    https://www.kernel.org/doc/html/latest/locking/seqlock.html
//  - Unwind tables consumed by backtrace():
//      DWARF .eh_frame (ARM64/x86-64) — LSB Core, Exception Frames:
//        https://refspecs.linuxfoundation.org/LSB_5.0.0/LSB-Core-generic/LSB-Core-generic/ehframechpt.html
//      ARM32 EHABI .ARM.exidx — Exception Handling ABI for the Arm Arch:
//        https://github.com/ARM-software/abi-aa/blob/main/ehabi32/ehabi32.rst
//      ARM64 frame-record chain (x29/x30), AAPCS64:
//        https://github.com/ARM-software/abi-aa/blob/main/aapcs64/aapcs64.rst

#include "exception_stack_tracer.hpp"

#include "exception_stack_tracer_internal.hpp"

#include <cxxabi.h>
#include <dlfcn.h>
#include <execinfo.h>
#include <link.h>

#include <algorithm>
#include <array>
#include <atomic>
#include <cstddef>
#include <cstdint>
#include <cstdlib>
#include <cstring>
#include <exception>
#include <typeinfo>

#include "absl/base/internal/raw_logging.h"

namespace {

// Frame 0 is the interposer itself, so ~9 caller frames survive; enough to
// clear the __throw_* helper and land in application code.
constexpr int kMaxFrames = 10;
constexpr uint32_t kRingSize = 64;
constexpr size_t kMaxModules = 64;
constexpr size_t kModuleNameSize = 40;

// A captured call stack as written by backtrace(): an array of up to
// kMaxFrames return addresses (code pointers / program counters), innermost
// frame first. Each entry is an opaque instruction address, later resolved to
// module + offset for offline addr2line; it is never dereferenced.
using FrameArray = std::array<void*, kMaxFrames>;

// A throw-site stack copied out of the ring for the terminate handler to print.
struct ThrowRecordSnapshot {
  const std::type_info* type = nullptr;
  int depth = 0;
  FrameArray frames{};
};

// ---------------------------------------------------------------------------
// ThrowSiteRing: a lock-free ring of throw-site stacks.
//
// Writers are the throwing threads (via __cxa_throw); the single reader is the
// terminate handler. Each slot is published as a seqlock: `object` is the
// sequence word, stored last with release order and cleared first by a writer,
// so a reader trusts a copy only if `object` matched before and after.
// ---------------------------------------------------------------------------
class ThrowSiteRing {
 public:
  void setEnabled(bool enable) {
    enabled_.store(enable, std::memory_order_relaxed);
  }

  // Writer, called from __cxa_throw before the unwinder runs: at catch time the
  // throw-site frames no longer exist.
  void record(void* object, const std::type_info* type) {
    if (!enabled_.load(std::memory_order_relaxed)) {
      return;
    }
    // backtrace() writes up to kMaxFrames return addresses into `frames` and
    // returns how many it wrote (`depth`), innermost frame first.
    FrameArray frames{};
    int depth = backtrace(frames.data(), kMaxFrames);
    nextSlot().publish(object, type, frames, depth);
  }

  // Reader: scan newest-first so that after wraparound a reused object address
  // resolves to the live exception's record, not an older evicted throw.
  bool lookup(void* object, ThrowRecordSnapshot& out) const {
    if (object == nullptr) {
      return false;
    }
    uint32_t end = next_.load(std::memory_order_acquire);
    for (uint32_t i = 0; i < kRingSize; ++i) {
      const Slot& slot = slots_[(end - 1 - i) % kRingSize];
      switch (slot.tryRead(object, out)) {
        case Slot::ReadResult::kNoMatch:
          continue;
        case Slot::ReadResult::kCopied:
          return true;
        case Slot::ReadResult::kEvicted:
          // The address matched but the slot was rewritten during the copy.
          // The caller's exception_ptr pins `object` alive, so no concurrent
          // throw can record this address; a rewrite therefore means this slot
          // was reused by an unrelated throw and the record is gone. Report it
          // evicted rather than matching a stale older same-address slot.
          return false;
      }
    }
    return false;
  }

  // Test-only: re-enable recording and drop all records.
  void reset() {
    enabled_.store(true, std::memory_order_relaxed);
    next_.store(0, std::memory_order_relaxed);
    for (Slot& slot : slots_) {
      slot.invalidate();
    }
  }

 private:
  class Slot {
   public:
    enum class ReadResult { kNoMatch, kCopied, kEvicted };

    // Writer half of the seqlock: invalidate, fill the payload, then publish
    // `object` last so a concurrent reader sees either the old value, a
    // no-match sentinel, or the fully written record — never a torn mix.
    void publish(void* object, const std::type_info* type,
                 const FrameArray& frames, int depth) {
      object_.store(nullptr, std::memory_order_relaxed);
      // A release *store* only keeps preceding writes above it; it does not stop
      // the relaxed payload writes below from being hoisted above the
      // invalidation. This release fence is the store-store barrier that keeps
      // the object_=nullptr invalidation ordered before the payload writes, so a
      // reader still seeing the old object during an eviction can never observe
      // a torn payload and revalidate it as kCopied.
      std::atomic_thread_fence(std::memory_order_release);
      type_.store(type, std::memory_order_relaxed);
      depth_.store(depth, std::memory_order_relaxed);
      for (size_t i = 0; i < kMaxFrames; ++i) {
        frames_[i].store(frames[i], std::memory_order_relaxed);
      }
      object_.store(object, std::memory_order_release);
    }

    // Reader half: copy the payload iff the slot holds `object`, then report
    // whether the sequence word still matches (kCopied) or was rewritten
    // mid-copy (kEvicted). Payload fields are atomic so the mid-copy read that
    // kEvicted discards is well defined rather than a data race.
    ReadResult tryRead(void* object, ThrowRecordSnapshot& out) const {
      if (object_.load(std::memory_order_acquire) != object) {
        return ReadResult::kNoMatch;
      }
      out.type = type_.load(std::memory_order_relaxed);
      out.depth = sanitizedDepth(depth_.load(std::memory_order_relaxed));
      for (size_t i = 0; i < kMaxFrames; ++i) {
        out.frames[i] = frames_[i].load(std::memory_order_relaxed);
      }
      // Symmetric to publish()'s release fence. An acquire *load* on object_
      // only orders operations that follow it; it would not stop the relaxed
      // payload loads above from being reordered *after* the revalidation, so a
      // slot rewritten mid-copy could be revalidated as kCopied over a torn
      // payload. This acquire fence is the load-load barrier that keeps the
      // payload reads ordered before the revalidation load, so a mid-copy
      // rewrite is reliably detected as kEvicted. With the fence in place the
      // revalidation load itself only needs to be relaxed.
      std::atomic_thread_fence(std::memory_order_acquire);
      return object_.load(std::memory_order_relaxed) == object
                 ? ReadResult::kCopied
                 : ReadResult::kEvicted;
    }

    void invalidate() { object_.store(nullptr, std::memory_order_relaxed); }

   private:
    // A valid frame count is 0..kMaxFrames; anything else can only come from a
    // torn read (later rejected by revalidation), so map it to 0 -- show no
    // frames rather than a bogus count.
    static int sanitizedDepth(int depth) {
      return (depth < 0 || depth > kMaxFrames) ? 0 : depth;
    }

    std::atomic<void*> object_{nullptr};
    std::atomic<const std::type_info*> type_{nullptr};
    std::atomic<int> depth_{0};
    std::array<std::atomic<void*>, kMaxFrames> frames_{};
  };

  Slot& nextSlot() {
    return slots_[next_.fetch_add(1, std::memory_order_relaxed) % kRingSize];
  }

  // Kill switch (--exception_stack_tracer=false): recording is a no-op and the
  // default terminate handler stays installed.
  std::atomic<bool> enabled_{true};
  std::atomic<uint32_t> next_{0};
  std::array<Slot, kRingSize> slots_{};
};

// ---------------------------------------------------------------------------
// ModuleTable: maps a runtime code address back to (module, in-file offset).
//
// backtrace() gives us raw runtime addresses. To symbolize them offline we need
// two things per address: which binary it came from, and its offset within that
// binary's file (that is what `addr2line -e <file> <offset>` consumes).
//
// Why we keep our own table instead of resolving at crash time:
//   The normal way to map an address to its module is dladdr(), but dladdr
//   takes the dynamic loader lock and may allocate -- both forbidden in the
//   terminate handler, which runs under suspected corruption and must never
//   block or allocate. So we walk the loaded modules once at install time,
//   while the process is healthy, and cache exactly what we need in a
//   fixed-size table the handler can scan with plain array reads: no locks, no
//   allocation, no loader calls. This stays correct because bmcweb does not
//   dlopen anything after startup, so the module map never changes.
//
// What the table holds:
//   One ModuleInfo row per loaded module (the executable and each .so). Each
//   row records that module's runtime address range (to test which module a
//   frame falls in), its load bias (to convert the frame to an in-file offset),
//   and its short name (for the log line) -- everything needed to turn a raw
//   address into "module + offset".
//
// ELF background (how a program is laid out in memory, i.e. where those numbers
// come from):
//   * A "module" is one ELF file mapped into the process: the main executable
//     plus every shared library (.so) it loaded.
//   * Each ELF file carries a "program header" table describing how the loader
//     should map it. Each entry is a segment; the ones with type PT_LOAD are
//     the chunks of code/data actually copied into memory (other types hold
//     dynamic-linking metadata, notes, etc., and take up no code space).
//   * Addresses inside the file are recorded relative to a link-time base of 0
//     (for position-independent code -- PIEs and all .so files). At load time
//     ASLR places the module at a random base address; the loader then shifts
//     every address by a fixed amount called the "load bias".
//       runtime address = load bias + in-file address
//   * So a PT_LOAD segment lives at runtime in
//       [ bias + p_vaddr , bias + p_vaddr + p_memsz )
//     where p_vaddr is the segment's in-file address and p_memsz its size in
//     memory. Unioning all PT_LOAD segments gives the module's whole runtime
//     address range (ModuleInfo::begin/end).
//   * To symbolize a frame we reverse the shift: in-file offset = addr - bias.
//
// The snapshot is taken with dl_iterate_phdr (see loadedRange below).
// ---------------------------------------------------------------------------
// One row of the table: a single loaded module (the executable or one .so), and
// everything the terminate handler needs to place and symbolize an address that
// falls inside it.
struct ModuleInfo {
  // Load bias (dl_phdr_info::dlpi_addr): the address a position-independent
  // object was mapped at. `frame - bias` is the in-file offset addr2line wants.
  uintptr_t bias = 0;
  // Half-open runtime address range [begin, end) the module occupies, used to
  // match a captured frame to its module.
  uintptr_t begin = 0;
  uintptr_t end = 0;
  // Module basename ("exe" for the main program), NUL-terminated.
  std::array<char, kModuleNameSize> name{};

  // Copy `n` into the fixed buffer, always leaving it NUL-terminated.
  void setName(const char* n) {
    std::strncpy(name.data(), n, kModuleNameSize - 1);
    name[kModuleNameSize - 1] = '\0';
  }
};

class ModuleTable {
 public:
  // Snapshot the currently loaded modules. Call once, at install time.
  void snapshot() { dl_iterate_phdr(&phdrCallback, this); }

  // Find the module containing `addr`, or nullptr if none does.
  const ModuleInfo* resolve(uintptr_t addr) const {
    size_t modules = count_.load(std::memory_order_acquire);
    for (size_t m = 0; m < modules; ++m) {
      if (addr >= modules_[m].begin && addr < modules_[m].end) {
        return &modules_[m];
      }
    }
    return nullptr;
  }

  void reset() { count_.store(0, std::memory_order_relaxed); }

 private:
  static int phdrCallback(dl_phdr_info* info, size_t /*size*/, void* self) {
    return static_cast<ModuleTable*>(self)->addModule(*info);
  }

  // A half-open [begin, end) address range, accumulated across PT_LOAD
  // segments. Starts empty (begin > end) so min/max fold correctly.
  struct AddressRange {
    uintptr_t begin = UINTPTR_MAX;
    uintptr_t end = 0;
    bool empty() const { return end == 0; }
  };

  // dl_iterate_phdr hands us one dl_phdr_info per loaded ELF object (the main
  // program and each shared library). The fields we use:
  //   dlpi_addr  - load bias: the base address the object was mapped at.
  //   dlpi_name  - the object's file path ("" for the main program).
  //   dlpi_phdr  - array of the object's ELF program headers (ElfW(Phdr),
  //   dlpi_phnum   which is Elf32/Elf64 chosen for the target).
  // Each PT_LOAD program header describes a segment actually mapped into
  // memory; its runtime span is [dlpi_addr + p_vaddr, + p_memsz). Unioning the
  // PT_LOAD segments yields the whole object's occupied address range.
  static AddressRange loadedRange(const dl_phdr_info& info) {
    AddressRange range;
    for (int i = 0; i < info.dlpi_phnum; ++i) {
      // NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
      const ElfW(Phdr)& phdr = info.dlpi_phdr[i];
      if (phdr.p_type != PT_LOAD) {
        continue;  // only PT_LOAD segments are mapped into memory
      }
      uintptr_t segmentBegin = info.dlpi_addr + phdr.p_vaddr;
      uintptr_t segmentEnd = segmentBegin + phdr.p_memsz;
      range.begin = std::min(range.begin, segmentBegin);
      range.end = std::max(range.end, segmentEnd);
    }
    return range;
  }

  // Reduce a module path to its basename for compact logging; the main
  // program's dlpi_name is empty, so report it as "exe".
  static const char* moduleBasename(const char* path) {
    if (path == nullptr || path[0] == '\0') {
      return "exe";
    }
    const char* slash = std::strrchr(path, '/');
    // NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
    return slash != nullptr ? slash + 1 : path;
  }

  // dl_iterate_phdr callback. Returns 1 to stop iteration once the table is
  // full, 0 to continue.
  int addModule(const dl_phdr_info& info) {
    size_t count = count_.load(std::memory_order_relaxed);
    if (count >= kMaxModules) {
      return 1;
    }
    AddressRange range = loadedRange(info);
    if (range.empty()) {
      return 0;  // no PT_LOAD segments (e.g. a linker placeholder); skip
    }
    ModuleInfo& m = modules_[count];
    m.bias = info.dlpi_addr;
    m.begin = range.begin;
    m.end = range.end;
    m.setName(moduleBasename(info.dlpi_name));
    count_.store(count + 1, std::memory_order_release);
    return 0;
  }

  // The snapshot: modules_[0, count_) are the loaded modules captured at
  // install time. Fixed capacity so no allocation is ever needed; count_ is
  // published with release so resolve() sees fully-written rows.
  std::array<ModuleInfo, kMaxModules> modules_{};
  std::atomic<size_t> count_{0};
};

// The two process-wide tables. constinit guarantees constant initialization,
// so they are ready before any static constructor (or early throw) runs.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
constinit ThrowSiteRing gThrowRing{};
// NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
constinit ModuleTable gModuleTable{};

// Returns the address that identifies the in-flight exception -- the same
// pointer __cxa_throw recorded -- which the handler uses to look up its stack.
// On libstdc++ (GNU), a std::exception_ptr is exactly one pointer to that
// object, so we read it out directly (layout checked by static_assert). On any
// other C++ runtime (e.g. clang/libc++) we don't know the layout, so we return
// nullptr; the handler then just reports "stack unavailable" and everything
// else keeps working.
#if defined(__GLIBCXX__)
void* primaryExceptionObject(const std::exception_ptr& p) {
  static_assert(sizeof(std::exception_ptr) == sizeof(void*),
                "libstdc++ exception_ptr layout assumption violated");
  void* object = nullptr;
  std::memcpy(&object, &p, sizeof(object));
  return object;
}
#else
void* primaryExceptionObject(const std::exception_ptr& /*p*/) {
  return nullptr;  // unknown exception_ptr layout: lookups report unavailable
}
#endif

void logResolvedFrames(const FrameArray& frames, int depth) {
  for (int i = 0; i < depth; ++i) {
    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
    const auto addr = reinterpret_cast<uintptr_t>(frames[static_cast<size_t>(i)]);
    const ModuleInfo* module = gModuleTable.resolve(addr);
    if (module != nullptr) {
      // Module-relative offset: symbolizable offline against the matching
      // .debug file (addr2line -e) even for a stripped PIE.
      ABSL_RAW_LOG(ERROR, "  #%02d %s+0x%lx", i, module->name.data(),
                   static_cast<unsigned long>(addr - module->bias));
    } else {
      ABSL_RAW_LOG(ERROR, "  #%02d [0x%lx]", i,
                   static_cast<unsigned long>(addr));
    }
  }
}

[[noreturn]] void terminateWithThrowStack() {
  // If anything below terminates again (a throwing what(), a second uncaught
  // exception on another thread), die immediately instead of recursing.
  // NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
  static std::atomic<bool> entered{false};
  if (entered.exchange(true, std::memory_order_relaxed)) {
    std::abort();
  }
  exception_stack_tracer_internal::logThrowSiteStackOfCurrentException();
  std::abort();
}

}  // namespace

namespace exception_stack_tracer_internal {

void logThrowSiteStackOfCurrentException() {
  const std::type_info* type = abi::__cxa_current_exception_type();
  if (type == nullptr) {
    ABSL_RAW_LOG(ERROR, "terminate: called without an active exception");
    return;
  }
  ABSL_RAW_LOG(ERROR, "terminate: uncaught exception of type %s", type->name());

  // __cxa_call_terminate (or a test's active catch) began the catch, so the
  // in-flight exception is rethrowable in place. Unlike std::rethrow_exception
  // this allocates nothing (same technique as libstdc++'s
  // __verbose_terminate_handler).
  try {
    throw;
  } catch (const std::exception& e) {
    // A user-defined what() override may return nullptr; passing that to %s is
    // undefined behaviour, so fall back to a literal (mirrors the type-name
    // guard below).
    const char* what = e.what();
    ABSL_RAW_LOG(ERROR, "  what(): %s", what != nullptr ? what : "(null)");
  } catch (...) {
  }

  // For a native exception this only bumps a refcount on the existing object
  // — no allocation. It also pins the object for the ring's revalidation.
  if (std::exception_ptr eptr = std::current_exception()) {
    ThrowRecordSnapshot snapshot;
    if (gThrowRing.lookup(primaryExceptionObject(eptr), snapshot)) {
      ABSL_RAW_LOG(ERROR, "throw-site stack (recorded as %s):",
                   snapshot.type != nullptr ? snapshot.type->name() : "?");
      logResolvedFrames(snapshot.frames, snapshot.depth);
    } else {
      ABSL_RAW_LOG(ERROR,
                   "throw-site stack unavailable (ring evicted, foreign "
                   "exception, or non-libstdc++ runtime)");
    }
  }
}

void resetForTest() {
  gThrowRing.reset();
  gModuleTable.reset();
}

}  // namespace exception_stack_tracer_internal

// Called once from run() at startup. Besides the one-time setup below, this
// explicit call is also what keeps this file in the binary: with no reference to
// it, the linker's --gc-sections would discard the whole translation unit -- and
// our __cxa_throw override along with it. enable=false (the
// --exception_stack_tracer kill switch) leaves the default terminate handler in
// place and turns recording into a no-op.
void installExceptionStackTracer(bool enable) {
  if (!enable) {
    gThrowRing.setEnabled(false);
    return;
  }
  // Snapshot the module table and force glibc's lazy libgcc unwinder load now
  // so the throw path and the terminate path never allocate, dlopen, or take
  // the loader lock.
  gModuleTable.snapshot();
  std::array<void*, 2> warm{};
  backtrace(warm.data(), 2);
  std::set_terminate(terminateWithThrowStack);
}

// Our replacement for the C++ runtime's throw entry point. Every `throw` in the
// process -- including libstdc++'s own __throw_* helpers and throws originating
// in other shared objects -- calls this function, because our definition is
// exported from the main executable (default visibility + a linker
// --dynamic-list entry for __cxa_throw; see meson.build) and the executable is
// searched first in the global symbol scope, so it interposes the library's.
// We record the throw site, then forward to the real __cxa_throw (found via
// dlsym(RTLD_NEXT, ...)) so throwing behaves exactly as normal.
//
// NOTE: exporting this symbol is what makes cross-module capture work. Without
// it (e.g. built LOCAL under -fvisibility=hidden), only throws whose __cxa_throw
// call site is compiled into the executable are recorded; throws executing
// inside libstdc++ or any other .so bind to the library's own __cxa_throw and
// are missed.
extern "C" [[gnu::visibility("default")]] void __cxa_throw(
    void* thrown_object, std::type_info* tinfo, void (*dest)(void*)) {
  gThrowRing.record(thrown_object, tinfo);
  using ThrowFn = void (*)(void*, std::type_info*, void (*)(void*));
  // NOLINTBEGIN(cppcoreguidelines-pro-type-reinterpret-cast)
  static const ThrowFn realCxaThrow =
      reinterpret_cast<ThrowFn>(dlsym(RTLD_NEXT, "__cxa_throw"));
  // NOLINTEND(cppcoreguidelines-pro-type-reinterpret-cast)
  realCxaThrow(thrown_object, tinfo, dest);
  __builtin_unreachable();
}