On Thu, Nov 26, 2020 at 09:36:34AM +0800, Guo Ren wrote:
> On Wed, Nov 25, 2020 at 10:31 PM Will Deacon <will@xxxxxxxxxx> wrote:
> > On Wed, Nov 25, 2020 at 03:16:45PM +0100, Peter Zijlstra wrote:
> > > @@ -207,6 +187,32 @@ static __always_inline void clear_pending_set_locked(struct qspinlock *lock)
> > > atomic_add(-_Q_PENDING_VAL + _Q_LOCKED_VAL, &lock->val);
> > > }
> > >
> > > +#endif /* _Q_PENDING_BITS == 8 */
> > > +
> > > +#if _Q_PENDING_BITS == 8 && ARCH_HAS_XCHG16
> > > +
> > > +/*
> > > + * xchg_tail - Put in the new queue tail code word & retrieve previous one
> > > + * @lock : Pointer to queued spinlock structure
> > > + * @tail : The new queue tail code word
> > > + * Return: The previous queue tail code word
> > > + *
> > > + * xchg(lock, tail), which heads an address dependency
> > > + *
> > > + * p,*,* -> n,*,* ; prev = xchg(lock, node)
> > > + */
> > > +static __always_inline u32 xchg_tail(struct qspinlock *lock, u32 tail)
> > > +{
> > > + /*
> > > + * We can use relaxed semantics since the caller ensures that the
> > > + * MCS node is properly initialized before updating the tail.
> > > + */
> > > + return (u32)xchg_relaxed(&lock->tail,
> > > + tail >> _Q_TAIL_OFFSET) << _Q_TAIL_OFFSET;
> > > +}
> > > +
> > > +#else /* !(_Q_PENDING_BITS == 8 && ARCH_HAS_XCHG16) */
> >
> > Why can't architectures just implement this with a 32-bit xchg instruction
> > if they don't have one that operates on 16 bits? Sure, they'll store more
> > data, but it's atomic so you shouldn't be able to tell... (ignoring parisc
> > crazy).
> >
> > Also, I'm surprised qspinlock benefits riscv. On arm64, there's nothing in
> > it over tickets for <= 16 CPUs.
> NUMA is on the way:
> https://lore.kernel.org/linux-riscv/20201119003829.1282810-1-atish.patra@xxxxxxx/
Well, they're patches not hardware (and I only see mention of QEMU over
there for the RISCV platform) ;)
> With your advice, I think we could using tickets lock when <= 16 CPUs
> and using qspinlock when > 16 CPUs.
> Is that right?
No, when I say "there's nothing in it", it means they're interchangeable.
It's just that qspinlock introduces a lot of complexity and I'm not keen
massively keen on changing the core code (which is used by many other
architectures) just because you don't have a 16-bit xchg() implementation.
So if you need qspinlock on riscv (not sure you do), then go ahead and make
that your one true lock implementation, but implement 16-bit xchg() at the
same time. Bonus points if you implement that in terms of 32-bit xchg() in
generic code (might need to watch out for endianness when shifting the new
value after aligning the base pointer).
Will
