On Mon, 20 Jun 2022 at 16:10, Luc Michel <[email protected]> wrote:
>
> On 15:35 Mon 20 Jun     , Peter Maydell wrote:
> > On Mon, 20 Jun 2022 at 15:25, Luc Michel <[email protected]> wrote:
> > > This series implements a clean way for semihosted exit syscalls to
> > > terminate QEMU with a given return code.
> > >
> > > Until now, exit syscalls implementations consisted in calling exit()
> > > with the wanted return code. The problem with this approach is that
> > > other CPUs are not properly stopped, leading to possible crashes in
> > > MTTCG mode, especially when at_exit callbacks have been registered. This
> > > can be the case e.g., when plugins are in use. Plugins can register
> > > at_exit callbacks. Those will be called on the CPU thread the exit
> > > syscall is comming from, while other CPUs can continue to run and thus
> > > call other plugin callbacks.
> >
> > The other option would be to say "if you register an atexit
> > callback in your plugin that's your problem to sort out" :-)
> > There's lots of situations where code inside QEMU might just
> > call exit(), not just this one. (Mostly these are "we detected
> > an error and decided to just bail out" codepaths.)
>
> Sorry I was a bit unclear. I meant plugins using the
> qemu_plugin_register_atexit_cb() register function, not directly calling
> atexit(). This function documentation stats:
>
> "Plugins should be able to free all their resources at this point much like
> after a reset/uninstall callback is called."
>
> If other CPUs are still running, this is not possible. I guess it's
> reasonable to assume CPUs have reached a quiescent state when those
> callbacks are called.

Ah, I see. I wonder if we should be handling the system-mode
exit() more along the lines of what we're currently doing
for usermode exit(). In general I don't think it's safe for
the plugins/core.c code to assume that when its atexit() handler
runs that all the TCG CPUs have stopped. Semihosting exit is just
the easiest reliably-reproducible situation when that's false.

thanks
-- PMM

Reply via email to