The POSIX pipe read wrappers stored the return of ::read() into an
unsigned result, so a negative count (e.g. EBADF from a concurrent
close on another thread) became SIZE_MAX. ContentReader::buffer()
would then try to grow its deque by ~18 exabytes and crash with
either std::length_error (glibc) or a stack smash (libc++). Treat
any non-positive ::read() return as EOF/error, close the pipe, and
return 0 so the peer's SessionThread observes a clean empty payload.
cmDebuggerPipeClient is test-only infrastructure; production cmake
is always the pipe server. Production close() assumes sequential
access, which holds in normal use (the adapter destructor joins
SessionThread before closing the connection). The new abrupt-
disconnect test however needs to wake a sibling thread blocked in
read() on the same fd, which Linux ::close() does not do.
Add a ShutdownForTesting() method that calls shutdown(SHUT_RDWR)
without freeing the fd, so any concurrent blocking read wakes
with a clean zero-length return. On Windows, CloseHandle already
cancels pending overlapped I/O, so the helper just forwards to
close(). Use it from testProtocolWithPipesAbruptDisconnect in
place of close().
Adding an integer to a C string does pointer math, rather than
converting to string. Instead convert the result of `GetLastError` to
string before adding it to the error message.
This problem was accidentally introduced by commit 8b1257e7bf (Debugger:
Replace libuv with platform-specific connection code, 2023-07-29).
Signed-off-by: William R. Dieter <william.r.dieter@intel.com>
Remove libuv usage from CMake debugger. Libuv has an async io model
and cppdap uses a sync model, so an extra thread and a buffer copy
were necessary to match semantics. In order to eliminate those
costs this commit implements the IO using platform specific APIs.
Christopher Wellons
Kitware Robot
William R. Dieter
Paul Maybee