On Sat, Feb 27, 2016 at 05:03:46PM +0100, Patrick Wildt wrote:
> On Fri, Feb 26, 2016 at 08:59:41PM -0800, Philip Guenther wrote:
> > On Sat, Feb 20, 2016 at 1:49 PM, Patrick Wildt <patr...@blueri.se> wrote:
> > > since ARMv6 the coprocessor provides special registers to store software
> > > defined values. Those registers are:
> > >
> > > * TPIDRURW -> kernel RW, user RW
> > > * TPIDRURO -> kernel RW, user RO
> > > * TPIDRPRW -> kernel RW
> > >
> > > TPIDRPRW is typically used to store the pointer to the curcpu struct,
> > > while TPIDRURO is used to point to the TCB.
> > >
> > > The following diff implements using TPIDRPRW to store and retrieve the
> > > curcpu struct pointer. This will especially be helpful in future MP
> > > efforts. I have guarded it for ARMv7 only, as that's the only hardware
> > > I have that supports it and I was able to test on.
> >
> > I don't know arm, but of those choices that makes sense.
> >
> >
> > > If TCB_GET()'s only user is libpthread, then it could make sense to
> > > store the TCB pointer in TPIDRURO. If it's possible that it's also
> > > used in ports, then the arm packages won't be compatible to
> > > armish/zaurus anymore. But the TCB stuff is not part of this diff.
> >
> > The userland bits in <machine/tcb.h> should match the platform's ABI
> > for the thread register. The last time I tried to figure this out
> > from the arm ABI docs I wanted to stab someone, but I was hopefully
> > just looking at the wrong docs. Looking at the bits in FreeBSD I see
> >
> > static inline void
> > set_tls(void *tls)
> > {
> > __asm __volatile("mcr p15, 0, %0, c13, c0, 3" : : "r" (tls));
> > }
> >
> > so, uh, whichever register that sets is presumably the correct one for
> > userland.
>
> Yep, that's TPIDRURO. It's not a lot of work to use this register, but
> it might just not work out as zaurus and armv7 use the same packages.
Time to revisit this and switch to the armv7 path in
lib/librthread/arch/arm/_atomic_lock.c ?
