Re: [PATCH 3/6] linux-user: Add code for PR_GET/SET_UNALIGN

[Richard Henderson]

On 1/9/24 10:21, Philippe Mathieu-Daudé wrote:
> On 8/1/24 22:13, Richard Henderson wrote:
> > On 1/9/24 04:15, Philippe Mathieu-Daudé wrote:
> > > > +/*
> > > > + * This can't go in hw/core/cpu.c because that file is compiled only
> > > > + * once for both user-mode and system builds.
> > > > + */
> > > >   static Property cpu_common_props[] = {
> > > > -#ifndef CONFIG_USER_ONLY
> > > > +#ifdef CONFIG_USER_ONLY
> > > >       /*
> > > > -     * Create a memory property for softmmu CPU object,
> > > > -     * so users can wire up its memory. (This can't go in hw/core/cpu.c
> > > > -     * because that file is compiled only once for both user-mode
> > > > -     * and system builds.) The default if no link is set up is to use
> > > > +     * Create a property for the user-only object, so users can
> > > > +     * adjust prctl(PR_SET_UNALIGN) from the command-line.
> > >
> > > How can I test this per-thread property?
> >
> > -cpu foo,prctl-unalign-sigbus=true
> >
> >
> > > Shouldn't this be an accel (TCG/user) property, for all threads?
> >
> > There is always one cpu at user-only startup, and it is copied on clone.
> >
> > Logically it would be a kernel property, since it's something the kernel does, not the 
> > cpu.  But cpu vs accel makes no difference to me; it was just easy here.
>
> Can a process started with prctl(PR_SET_UNALIGN) unset it before
> forking?

Yes.

> "kernel property" as "accel property" works for me.

Does it really matter though, especially now that the cpu property has been present for 
two years now.


r~