Re: Win32 CTRL-C handling (Was: interrupt handing and the GUI (bug #37672))

[Michael Goffioul]

On Tue, Nov 19, 2013 at 1:10 AM, John W. Eaton <address@hidden> wrote:

On 11/18/2013 05:58 PM, Michael Goffioul wrote:

On Mon, Nov 18, 2013 at 10:50 AM, Michael Goffioul

<address@hidden <mailto:michael.goffioul@gmail.com>> wrote:

    One main issue I'd like to raise and discuss is the interaction with
    readline. Basically when you combine octave (using gnulib), readline
    (not using gnulib) and the fact that Win32 does not have siglongjmp,
    you start having conflicting situations.

    Typically, readline installs its own SIGINT handler, but also store
    the previous handler for later use. During SIGINT handling, it does
    some internal clean-up, restore the previous handler, than
    "raise(SIGINT)" to execute the previous handler.

    OTOH, octave uses gnulib. The gnulib's signal handling is along
    those lines: block the signal, execute the handler, unblock the
    signal. If the signal handler performs a longjmp, the signal will
    stay in "blocked" state (in gnulib, this translates into having a
    dummy handler that does nothing). Now the problem is that on the
    next run of readline, the stored handler will be the gnulib's dummy
    one, and CTRL-C will end up being ignored while being in readline
    loop. This does not happen on platforms having sigsetjmp/siglongjmp,
    or more generally on POSIX platforms.

    I've worked around that issue by installing octave_interrupt_hook
    and make sure SIGINT is handled after an interrupt. Another way,
    which I tested, is to disable signal handling in readline by setting
    rl_catch_signals to 0, but I'm not sure to what extend this is a
    problem in readline itself.

Scrap that, it didn't work. It appears that gnulib is keeping
internal/static states about signal handlers, which is not really
compatible with readline making direct calls to the "signal" function
from windows runtime DLL.

I ended up disabling signals in readline. The attached path works for me
in GUI and CLI. Although as explained earlier, CTRL-C in CLI mode
doesn't work as expected when in readline, because the key sequence is
reported in the input buffer, not as a signal.

It would nice if someone can test it in MinGW.

Your change works for me with MinGW.

I forgot to mention one thing about that patch. Most of it is not MSVC specific, but there's one bit that is MSVC-specific: I need to call _fpreset before doing the longjmp. The reason is I noticed weird floating-point results after interrupting some LAPACK routine (typically, I used "inv" for testing interrupt); the first floating-point comparison was always invalid (like "0 == 0" returning 0). I tracked down the issue until an assembly-call FLD that was loading an valid value on the FPU stack, although the memory it was loaded from appeared to contain the correct bytes. I figured it could be due to the sudden interruption of the LAPACK routine, leaving the FPU in some undefined state; hence I added the _fpreset call.

I don't know whether the problem also occurs in MinGW, but if it does, I suppose there's a MinGW-equivalent to _fpreset.

Michael.