This allows us to disable Unix sockets cleanly on Windows. Replaces some
read/write calls with recv/send counterparts, as Winsock does not like
read/writes.
Instead of a pipe we use a TCP socketpair (can't select() a _pipe()), and
Windsock2 send/recv functions instead of read/write.
Currently supported (and required) are file descriptors provided by Winsock
only; we might use a separate mechanism for traditional file handles if
required (or switch to Windows events and WaitForMultipleObjects) for a future
version.
While DbgHelp provides a convenient API to create backtraces, any executable
linking against DbgHelp gets a more than a significant slow down. Further, it
can only lookup global symbols, as it expects PDB files we can't produce with a
MinGW build.
With some core Kernel32.dll functionality, we can capture stack traces much
faster. Together with the optional libbfd, we can print very fine backtraces.
When --enable-bfd-backtraces is used on Windows, a libbfd.dll is required for
the build. Such a DLL can be created from the binutils sources using:
# build binutils with mingw...
# extract archive members from binutils libraries
x86_64-w64-mingw32-ar x $BINUTILS/bfd/.libs/libbfd.a
x86_64-w64-mingw32-ar x $BINUTILS/intl/libintl.a
x86_64-w64-mingw32-ar x $BINUTILS/libiberty/libiberty.a
# create self-contained libbfd.a, with index
x86_64-w64-mingw32-ar qs libbfd.a *.o
# create DLL from static library
x86_64-w64-mingw32-dlltool -e libbfd.o -l libbfd.lib libbfd.a
x86_64-w64-mingw32-gcc -shared libbfd.a libbfd.o -o libbfd.dll
Because CURLOPT_FAILONERROR is enabled in the curl plugin an error code
will often (not always) cause the client to close the TCP connection
before the server has written the complete response.
In some cases the main thread is not ready to immediately call siglongjmp(),
e.g. if it currently holds a mutex that is later required during
shutdown.
Therefore, we delay handling errors in worker threads until the main
thread performs the next check itself (or the test function ends).
The same issue remains with SIGALRM.
When running the tests in GDB the working directory apparently is
different. With the relative path used previously the plugins would not
be found and those installed on the system would get used.
Handling the result for enum_from_name() is difficult, as checking for
negative return values requires a cast if the enum type is unsigned. The new
signature clearly differentiates lookup result from lookup value.
Further, this actually allows to convert real -1 enum values, which could not
be distinguished from "not-found" and the -1 return value.
This also fixes several clang warnings where enums are unsigned.
These definitions are directly derived from the RFC, so it should be safe
to cast them. clang complains about the different types, so cast them
explicitly.
Martin Willi