2018-12-01 00:04:07

by Andrea Parri

Subject: [PATCH] riscv, atomic: Add #define's for the atomic_{cmp,}xchg_*() variants

If an architecture does not define the atomic_{cmp,}xchg_*() variants,
the generic implementation defaults them to the fully-ordered version.

riscv's had its own variants since "the beginning", but it never told
(#define-d these for) the generic implementation: it is time to do so.

Signed-off-by: Andrea Parri <[email protected]>
Cc: Palmer Dabbelt <[email protected]>
Cc: Albert Ou <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Boqun Feng <[email protected]>
---
TBH, the delay was not intentional: I've just become aware of it while
working on moving riscv over to queued rwlocks. There's currently one
callsite for the non-fully-ordered variants mentioned above for riscv:
for atomic_cmpxchg_acquire() in kernel/sched/rt.c:rto_start_trylock(),

[before]

51a: 100726af lr.w a3,(a4)
51e: 00069763 bnez a3,52c <.L17>
522: 1af7262f sc.w.rl a2,a5,(a4)
526: fa75 bnez a2,51a <.L1>
528: 0330000f fence rw,rw

[after]

51a: 100726af lr.w a3,(a4)
51e: 00069763 bnez a3,52c <.L17>
522: 18f7262f sc.w a2,a5,(a4)
526: fa75 bnez a2,51a <.L1>
528: 0230000f fence r,rw

---
arch/riscv/include/asm/atomic.h | 9 +++++++++
1 file changed, 9 insertions(+)

diff --git a/arch/riscv/include/asm/atomic.h b/arch/riscv/include/asm/atomic.h
index c452359c9cb8a..93826771b616a 100644
--- a/arch/riscv/include/asm/atomic.h
+++ b/arch/riscv/include/asm/atomic.h
@@ -303,6 +303,15 @@ c_t atomic##prefix##_cmpxchg(atomic##prefix##_t *v, c_t o, c_t n) \

ATOMIC_OPS()

+#define atomic_xchg_relaxed atomic_xchg_relaxed
+#define atomic_xchg_acquire atomic_xchg_acquire
+#define atomic_xchg_release atomic_xchg_release
+#define atomic_xchg atomic_xchg
+#define atomic_cmpxchg_relaxed atomic_cmpxchg_relaxed
+#define atomic_cmpxchg_acquire atomic_cmpxchg_acquire
+#define atomic_cmpxchg_release atomic_cmpxchg_release
+#define atomic_cmpxchg atomic_cmpxchg
+
#undef ATOMIC_OPS
#undef ATOMIC_OP

--
2.17.1

2018-12-01 10:25:34

by Peter Zijlstra

Subject: Re: [PATCH] riscv, atomic: Add #define's for the atomic_{cmp,}xchg_*() variants

On Sat, Dec 01, 2018 at 01:01:56AM +0100, Andrea Parri wrote:
> If an architecture does not define the atomic_{cmp,}xchg_*() variants,
> the generic implementation defaults them to the fully-ordered version.
>
> riscv's had its own variants since "the beginning", but it never told
> (#define-d these for) the generic implementation: it is time to do so.
>
> Signed-off-by: Andrea Parri <[email protected]>
> Cc: Palmer Dabbelt <[email protected]>
> Cc: Albert Ou <[email protected]>
> Cc: Will Deacon <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> Cc: Boqun Feng <[email protected]>

Acked-by: Peter Zijlstra (Intel) <[email protected]>

> ---
> arch/riscv/include/asm/atomic.h | 9 +++++++++
> 1 file changed, 9 insertions(+)
>
> diff --git a/arch/riscv/include/asm/atomic.h b/arch/riscv/include/asm/atomic.h
> index c452359c9cb8a..93826771b616a 100644
> --- a/arch/riscv/include/asm/atomic.h
> +++ b/arch/riscv/include/asm/atomic.h
> @@ -303,6 +303,15 @@ c_t atomic##prefix##_cmpxchg(atomic##prefix##_t *v, c_t o, c_t n) \
>
> ATOMIC_OPS()
>
> +#define atomic_xchg_relaxed atomic_xchg_relaxed
> +#define atomic_xchg_acquire atomic_xchg_acquire
> +#define atomic_xchg_release atomic_xchg_release
> +#define atomic_xchg atomic_xchg
> +#define atomic_cmpxchg_relaxed atomic_cmpxchg_relaxed
> +#define atomic_cmpxchg_acquire atomic_cmpxchg_acquire
> +#define atomic_cmpxchg_release atomic_cmpxchg_release
> +#define atomic_cmpxchg atomic_cmpxchg
> +
> #undef ATOMIC_OPS
> #undef ATOMIC_OP
>
> --
> 2.17.1
>