On Wed, Oct 12, 2016 at 03:55:18PM +0200, Claudio Imbrenda wrote:
> On 12/10/16 15:15, David Hildenbrand wrote:
> >>> + for (cx = 0; ccpus && ccpus[cx]; cx++) {
> >>> + cpu_single_step(cpu, 0);
> >
> > This looks suspicious
>
> why? we set all cpus to single step, since that is the default, and then
> we clear the single-step property from all CPUs that should be restarted
> in normal mode, then we restart all CPUs. Those in single-step will
> indeed only perform one single step, the others will run freely (at
> least until the first single-step CPU stops again).
actually I was more concerned about calling it on "cpu" in a loop.
GDB will:
- single step one thread only (stopping all other)
- use vCont
as default. So this means quite some ioctls on every step with some VCPUs.
I doubt that it will really be a problem (e.g. for GDB single stepping
instead of setting breakpoints when returning froma function), but still I
want to have it said. (we actually only need 1 ioctl but call quite a lot).
>
> >>> + }
> >>> + CPU_FOREACH(cpu) {
> >>> + cpu_resume(cpu);
> >>> + }
> >
> > Claudio, did you have a look at how s->c_cpu is used later on? I remember
> > that we
> > have to take care of some query reply packages.
>
> yes, that's set by the H packet and used by the c,s,m,etc packets. vCont
> ignores it and doesn't change it
> (see here https://sourceware.org/gdb/onlinedocs/gdb/Packets.html )
I remember something different (also having to do with clients detaching and
re-attaching). Will have a look at the code when I have time.
