On Mon, Oct 25, 2021 at 03:06:24PM +0200, Arnd Bergmann wrote:
> On Mon, Oct 25, 2021 at 11:57 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > On Sat, Oct 23, 2021 at 06:04:57PM +0200, Arnd Bergmann wrote:
> > > On Sat, Oct 23, 2021 at 3:37 AM Waiman Long <longman@xxxxxxxxxx> wrote:
> > > >> On 10/22/21 7:59 AM, Arnd Bergmann wrote:
> > > > > From: Arnd Bergmann <arnd@xxxxxxxx>
> > > > >
> > > > > As this is all dead code, just remove it and the helper functions built
> > > > > around it. For arch/ia64, the inline asm could be cleaned up, but
> > > > > it seems safer to leave it untouched.
> > > > >
> > > > > Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
> > > >
> > > > Does that mean we can also remove the GENERIC_LOCKBREAK config option
> > > > from the Kconfig files as well?
> > >
> > > I couldn't figure this out.
> > >
> > > What I see is that the only architectures setting GENERIC_LOCKBREAK are
> > > nds32, parisc, powerpc, s390, sh and sparc64, while the only architectures
> > > implementing arch_spin_is_contended() are arm32, csky and ia64.
> > >
> > > The part I don't understand is whether the option actually does anything
> > > useful any more after commit d89c70356acf ("locking/core: Remove break_lock
> > > field when CONFIG_GENERIC_LOCKBREAK=y").
> >
> > Urgh, what a mess.. AFAICT there's still code in
> > kernel/locking/spinlock.c that relies on it. Specifically when
> > GENERIC_LOCKBREAK=y we seem to create _lock*() variants that are
> > basically TaS locks which drop preempt/irq disable while spinning.
> >
> > Anybody having this on and not having native TaS locks is in for a rude
> > surprise I suppose... sparc64 being the obvious candidate there :/
>
> Is this a problem on s390 and powerpc, those two being the ones
> that matter in practice?
>
> On s390, we pick between the cmpxchg() based directed-yield when
> running on virtualized CPUs, and a normal qspinlock when running on a
> dedicated CPU.
>
> On PowerPC, we pick at compile-time between either the qspinlock
> (default-enabled on Book3S-64, i.e. all server chips) or a ll/sc based
> spinlock plus vm_yield() (default on embedded and 32-bit mac).
Urgh, yeah, so this crud undermines the whole point of having a fair
lock. I'm thinking s390 and Power want to have this fixed.
