Re: [PATCH] LoongArch: Add qspinlock support

[hev <r@hev.cc>]

Hello,

On Sat, Jun 18, 2022 at 8:59 PM WANG Xuerui <kernel@xen0n.name> wrote:
>
> On 6/18/22 01:45, Guo Ren wrote:
> >
> >> I see that the qspinlock() code actually calls a 'relaxed' version of xchg16(),
> >> but you only implement the one with the full barrier. Is it possible to
> >> directly provide a relaxed version that has something less than the
> >> __WEAK_LLSC_MB?
> > I am also curious that __WEAK_LLSC_MB is very magic. How does it
> > prevent preceded accesses from happening after sc for a strong
> > cmpxchg?
> >
> > #define __cmpxchg_asm(ld, st, m, old, new)                              \
> > ({                                                                      \
> >          __typeof(old) __ret;                                            \
> >                                                                          \
> >          __asm__ __volatile__(                                           \
> >          "1:     " ld "  %0, %2          # __cmpxchg_asm \n"             \
> >          "       bne     %0, %z3, 2f                     \n"             \
> >          "       or      $t0, %z4, $zero                 \n"             \
> >          "       " st "  $t0, %1                         \n"             \
> >          "       beq     $zero, $t0, 1b                  \n"             \
> >          "2:                                             \n"             \
> >          __WEAK_LLSC_MB                                                  \
> >
> > And its __smp_mb__xxx are just defined as a compiler barrier()?
> > #define __smp_mb__before_atomic()       barrier()
> > #define __smp_mb__after_atomic()        barrier()
> I know this one. There is only one type of barrier defined in the v1.00
> of LoongArch, that is the full barrier, but this is going to change.
> Huacai hinted in the bringup patchset that 3A6000 and later models would
> have finer-grained barriers. So these indeed could be relaxed in the
> future, just that Huacai has to wait for their embargo to expire.
>

IIRC, The Loongson LL/SC behaves differently than others:

Loongson:
LL: Full barrier + Load exclusive
SC: Store conditional + Full barrier

Others:
LL: Load exclusive + Acquire barrier
SC: Release barrier + Store conditional

So we just need to prevent compiler reorder before/after atomic.
And this is why we need __WEAK_LLSC_MB to prevent runtime reorder for
loads after LL.

hev