Re: [PATCH 1/3] [POC] test implementaion of rt-signals

[Richard Weinberger <richard.weinberger@gmail.com>]

On Wed, Aug 16, 2023 at 12:19 PM Johannes Kirchmair
<johannes.kirchmair@sigmatek.at> wrote:
> diff --git a/kernel/cobalt/dovetail/kevents.c b/kernel/cobalt/dovetail/kevents.c
> index 4da4f51b7..61417717b 100644
> --- a/kernel/cobalt/dovetail/kevents.c
> +++ b/kernel/cobalt/dovetail/kevents.c
> @@ -57,6 +57,9 @@ void handle_oob_trap_entry(unsigned int trapnr, struct pt_regs *regs)
>                 xnsched_run();
>         }
>
> +       if (xnthread_handle_rt_signals(trapnr, regs) == 0)
> +               return;
> +

While continuing Johannes' work I noticed a problem with this code.
If we happen to trigger another exception while setting up the signal
frame a domain
switch could have happened and the subsequently call to
xnthread_relax() will crash the kernel
because it works only when we're in the primary domain.

e.g. if regs->sp points to a bogus memory location in userspace,
xnthread_handle_rt_signals() will
not only fail but also switch domains due to a page fault.
We'll return here relaxed.

My current workaround is a check like:
if (is_secondary_domain())
    return;

>         /*
>          * If we experienced a trap on behalf of a shadow thread
>          * running in primary mode, move it to the Linux domain,
> @@ -88,6 +91,8 @@ void handle_oob_trap_entry(unsigned int trapnr, struct pt_regs *regs)
>                 xnstat_counter_inc(&thread->stat.pf);
>
>         xnthread_relax(xnarch_fault_notify(trapnr), SIGDEBUG_MIGRATE_FAULT);
> +
> +       return;
>  }



-- 
Thanks,
//richard