Re: [PATCH -fixes] riscv: Fix fully ordered LR/SC xchg[8|16]() implementations
From: Alexandre Ghiti
Date: Thu May 30 2024 - 08:05:54 EST
Andrea,
On Thu, May 30, 2024 at 1:54 PM Andrea Parri <parri.andrea@xxxxxxxxx> wrote:
>
> > -#define _arch_xchg(ptr, new, sfx, prepend, append) \
> > +#define _arch_xchg(ptr, new, sc_sfx, swap_sfx, prepend, append) \
> > ({ \
> > __typeof__(ptr) __ptr = (ptr); \
> > __typeof__(*(__ptr)) __new = (new); \
> > @@ -55,15 +55,15 @@
> > switch (sizeof(*__ptr)) { \
> > case 1: \
> > case 2: \
> > - __arch_xchg_masked(prepend, append, \
> > + __arch_xchg_masked(sc_sfx, prepend, append, \
> > __ret, __ptr, __new); \
> > break; \
> > case 4: \
> > - __arch_xchg(".w" sfx, prepend, append, \
> > + __arch_xchg(".w" swap_sfx, prepend, append, \
> > __ret, __ptr, __new); \
> > break; \
> > case 8: \
> > - __arch_xchg(".d" sfx, prepend, append, \
> > + __arch_xchg(".d" swap_sfx, prepend, append, \
> > __ret, __ptr, __new); \
> > break; \
> > default: \
> > @@ -73,16 +73,16 @@
> > })
> >
> > #define arch_xchg_relaxed(ptr, x) \
> > - _arch_xchg(ptr, x, "", "", "")
> > + _arch_xchg(ptr, x, "", "", "", "")
> >
> > #define arch_xchg_acquire(ptr, x) \
> > - _arch_xchg(ptr, x, "", "", RISCV_ACQUIRE_BARRIER)
> > + _arch_xchg(ptr, x, "", "", "", RISCV_ACQUIRE_BARRIER)
> >
> > #define arch_xchg_release(ptr, x) \
> > - _arch_xchg(ptr, x, "", RISCV_RELEASE_BARRIER, "")
> > + _arch_xchg(ptr, x, "", "", RISCV_RELEASE_BARRIER, "")
> >
> > #define arch_xchg(ptr, x) \
> > - _arch_xchg(ptr, x, ".aqrl", "", "")
> > + _arch_xchg(ptr, x, ".rl", ".aqrl", "", " fence rw, rw\n")
>
> This does indeed fix the fully-ordered variant of xchg8/16(). But this
> also changes the fully-ordered xchg32() to
>
> amoswap.w.aqrl a4,a5,(s1)
> fence rw,rw
>
> (and similarly for xchg64()); we should be able to restore the original
> mapping with the diff below on top of this patch.
And you already told me that privately...Sorry, my mind has been
elsewhere lately...I'll fix that right now.
Sorry again and thanks,
Alex
>
> Andrea
>
> P.S. Perhaps expand the width of the macros to avoid newlines (I didn't
> do it keep the diff smaller).
>
> P.S. With Zabha, we'd probably like to pass swap_sfx and swap_append as
> well to __arch_xchg_masked().
>
>
> diff --git a/arch/riscv/include/asm/cmpxchg.h b/arch/riscv/include/asm/cmpxchg.h
> index e1e564f5dc7ba..88c8bb7ec1c34 100644
> --- a/arch/riscv/include/asm/cmpxchg.h
> +++ b/arch/riscv/include/asm/cmpxchg.h
> @@ -46,7 +46,8 @@
> : "memory"); \
> })
>
> -#define _arch_xchg(ptr, new, sc_sfx, swap_sfx, prepend, append) \
> +#define _arch_xchg(ptr, new, sc_sfx, swap_sfx, prepend, \
> + sc_append, swap_append) \
> ({ \
> __typeof__(ptr) __ptr = (ptr); \
> __typeof__(*(__ptr)) __new = (new); \
> @@ -55,15 +56,15 @@
> switch (sizeof(*__ptr)) { \
> case 1: \
> case 2: \
> - __arch_xchg_masked(sc_sfx, prepend, append, \
> + __arch_xchg_masked(sc_sfx, prepend, sc_append, \
> __ret, __ptr, __new); \
> break; \
> case 4: \
> - __arch_xchg(".w" swap_sfx, prepend, append, \
> + __arch_xchg(".w" swap_sfx, prepend, swap_append, \
> __ret, __ptr, __new); \
> break; \
> case 8: \
> - __arch_xchg(".d" swap_sfx, prepend, append, \
> + __arch_xchg(".d" swap_sfx, prepend, swap_append, \
> __ret, __ptr, __new); \
> break; \
> default: \
> @@ -73,16 +74,16 @@
> })
>
> #define arch_xchg_relaxed(ptr, x) \
> - _arch_xchg(ptr, x, "", "", "", "")
> + _arch_xchg(ptr, x, "", "", "", "", "")
>
> #define arch_xchg_acquire(ptr, x) \
> - _arch_xchg(ptr, x, "", "", "", RISCV_ACQUIRE_BARRIER)
> + _arch_xchg(ptr, x, "", "", "", RISCV_ACQUIRE_BARRIER, RISCV_ACQUIRE_BARRIER)
>
> #define arch_xchg_release(ptr, x) \
> - _arch_xchg(ptr, x, "", "", RISCV_RELEASE_BARRIER, "")
> + _arch_xchg(ptr, x, "", "", RISCV_RELEASE_BARRIER, "", "")
>
> #define arch_xchg(ptr, x) \
> - _arch_xchg(ptr, x, ".rl", ".aqrl", "", " fence rw, rw\n")
> + _arch_xchg(ptr, x, ".rl", ".aqrl", "", " fence rw, rw\n", "")
>
> #define xchg32(ptr, x) \
> ({ \
>