Re: [PATCH] riscv: locks: introduce ticket-based spinlock implementation

From: "Christoph Müllner" <christophm30@gmail.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: Palmer Dabbelt <palmer@dabbelt.com>,
	Anup Patel <anup@brainfault.org>, Guo Ren <guoren@kernel.org>,
	linux-riscv <linux-riscv@lists.infradead.org>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	Guo Ren <guoren@linux.alibaba.com>,
	catalin.marinas@arm.com, will.deacon@arm.com,
	Arnd Bergmann <arnd@arndb.de>
Subject: Re: [PATCH] riscv: locks: introduce ticket-based spinlock implementation
Date: Tue, 13 Apr 2021 12:25:00 +0200	[thread overview]
Message-ID: <CAHB2gtTzEuD7j-+5ztui0eV6UNiEisBTgoK+2Sr=Z0b4PPXRyA@mail.gmail.com> (raw)
In-Reply-To: <YHVl8fFfdQDn+QYW@hirez.programming.kicks-ass.net>

On Tue, Apr 13, 2021 at 11:37 AM Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Tue, Apr 13, 2021 at 11:22:40AM +0200, Christoph Müllner wrote:
>
> > > For ticket locks you really only needs atomic_fetch_add() and
> > > smp_store_release() and an architectural guarantees that the
> > > atomic_fetch_add() has fwd progress under contention and that a sub-word
> > > store (through smp_store_release()) will fail the SC.
> > >
> > > Then you can do something like:
> > >
> > > void lock(atomic_t *lock)
> > > {
> > >         u32 val = atomic_fetch_add(1<<16, lock); /* SC, gives us RCsc */
> > >         u16 ticket = val >> 16;
> > >
> > >         for (;;) {
> > >                 if (ticket == (u16)val)
> > >                         break;
> > >                 cpu_relax();
> > >                 val = atomic_read_acquire(lock);
> > >         }
> > > }
> > >
> > > void unlock(atomic_t *lock)
> > > {
> > >         u16 *ptr = (u16 *)lock + (!!__BIG_ENDIAN__);
> > >         u32 val = atomic_read(lock);
> > >
> > >         smp_store_release(ptr, (u16)val + 1);
> > > }
> > >
> > > That's _almost_ as simple as a test-and-set :-) It isn't quite optimal
> > > on x86 for not being allowed to use a memop on unlock, since its being
> > > forced into a load-store because of all the volatile, but whatever.
> >
> > What about trylock()?
> > I.e. one could implement trylock() without a loop, by letting
> > trylock() fail if the SC fails.
> > That looks safe on first view, but nobody does this right now.
>
> Generic code has to use cmpxchg(), and then you get something like this:
>
> bool trylock(atomic_t *lock)
> {
>         u32 old = atomic_read(lock);
>
>         if ((old >> 16) != (old & 0xffff))
>                 return false;
>
>         return atomic_try_cmpxchg(lock, &old, old + (1<<16)); /* SC, for RCsc */
> }

This approach requires two loads (atomic_read() and cmpxchg()), which
is not required.
Detecting this pattern and optimizing it in a compiler is quite unlikely.

A bit less generic solution would be to wrap the LL/SC (would be
mandatory in this case)
instructions and do something like this:

uint32_t __smp_load_acquire_reserved(void*);
int __smp_store_release_conditional(void*, uint32_t);

typedef union {
    uint32_t v32;
    struct {
        uint16_t owner;
        uint16_t next;
    };
} arch_spinlock_t;

int trylock(arch_spinlock_t *lock)
{
    arch_spinlock_t l;
    int success;
    do {
        l.v32 = __smp_load_acquire_reserved(lock);
        if (l.owner != l.next)
            return 0;
        l.next++;
        success = __smp_store_release_conditional(lock, l.v32);
    } while (!success);
    return success;
}

But here we can't tell the compiler to optimize the code between LL and SC...

>
> That will try and do the full LL/SC loop, because it wants to complete
> the cmpxchg, but in generic code we have no other option.
>
> (Is this what C11's weak cmpxchg is for?)