2018-02-27 02:25:27

by Andrea Parri

[permalink] [raw]
Subject: [PATCH] riscv/barrier: Define __smp_{store_release,load_acquire}

Introduce __smp_{store_release,load_acquire}, and rely on the generic
definitions for smp_{store_release,load_acquire}. This avoids the use
of full ("rw,rw") fences on SMP.

Signed-off-by: Andrea Parri <[email protected]>
---
arch/riscv/include/asm/barrier.h | 15 +++++++++++++++
1 file changed, 15 insertions(+)

diff --git a/arch/riscv/include/asm/barrier.h b/arch/riscv/include/asm/barrier.h
index 5510366d169ae..d4628e4b3a5ea 100644
--- a/arch/riscv/include/asm/barrier.h
+++ b/arch/riscv/include/asm/barrier.h
@@ -38,6 +38,21 @@
#define __smp_rmb() RISCV_FENCE(r,r)
#define __smp_wmb() RISCV_FENCE(w,w)

+#define __smp_store_release(p, v) \
+do { \
+ compiletime_assert_atomic_type(*p); \
+ RISCV_FENCE(rw,w); \
+ WRITE_ONCE(*p, v); \
+} while (0)
+
+#define __smp_load_acquire(p) \
+({ \
+ typeof(*p) ___p1 = READ_ONCE(*p); \
+ compiletime_assert_atomic_type(*p); \
+ RISCV_FENCE(r,rw); \
+ ___p1; \
+})
+
/*
* This is a very specific barrier: it's currently only used in two places in
* the kernel, both in the scheduler. See include/linux/spinlock.h for the two
--
2.7.4



2018-02-27 18:22:40

by Palmer Dabbelt

[permalink] [raw]
Subject: Re: [PATCH] riscv/barrier: Define __smp_{store_release,load_acquire}

On Mon, 26 Feb 2018 18:24:11 PST (-0800), [email protected] wrote:
> Introduce __smp_{store_release,load_acquire}, and rely on the generic
> definitions for smp_{store_release,load_acquire}. This avoids the use
> of full ("rw,rw") fences on SMP.
>
> Signed-off-by: Andrea Parri <[email protected]>
> ---
> arch/riscv/include/asm/barrier.h | 15 +++++++++++++++
> 1 file changed, 15 insertions(+)
>
> diff --git a/arch/riscv/include/asm/barrier.h b/arch/riscv/include/asm/barrier.h
> index 5510366d169ae..d4628e4b3a5ea 100644
> --- a/arch/riscv/include/asm/barrier.h
> +++ b/arch/riscv/include/asm/barrier.h
> @@ -38,6 +38,21 @@
> #define __smp_rmb() RISCV_FENCE(r,r)
> #define __smp_wmb() RISCV_FENCE(w,w)
>
> +#define __smp_store_release(p, v) \
> +do { \
> + compiletime_assert_atomic_type(*p); \
> + RISCV_FENCE(rw,w); \
> + WRITE_ONCE(*p, v); \
> +} while (0)
> +
> +#define __smp_load_acquire(p) \
> +({ \
> + typeof(*p) ___p1 = READ_ONCE(*p); \
> + compiletime_assert_atomic_type(*p); \
> + RISCV_FENCE(r,rw); \
> + ___p1; \
> +})
> +
> /*
> * This is a very specific barrier: it's currently only used in two places in
> * the kernel, both in the scheduler. See include/linux/spinlock.h for the two

I'm adding Daniel just in case I misunderstood what's going on here, but these
look good to me. As this is a non-trivial memory model change I'm going to let
it bake in linux-next for a bit just so it gets some visibility.

Thanks!

2018-02-27 22:22:10

by Daniel Lustig

[permalink] [raw]
Subject: Re: [PATCH] riscv/barrier: Define __smp_{store_release,load_acquire}

On 2/27/2018 10:21 AM, Palmer Dabbelt wrote:
> On Mon, 26 Feb 2018 18:24:11 PST (-0800), [email protected] wrote:
>> Introduce __smp_{store_release,load_acquire}, and rely on the generic
>> definitions for smp_{store_release,load_acquire}. This avoids the use
>> of full ("rw,rw") fences on SMP.
>>
>> Signed-off-by: Andrea Parri <[email protected]>
>> ---
>>  arch/riscv/include/asm/barrier.h | 15 +++++++++++++++
>>  1 file changed, 15 insertions(+)
>>
>> diff --git a/arch/riscv/include/asm/barrier.h b/arch/riscv/include/asm/barrier.h
>> index 5510366d169ae..d4628e4b3a5ea 100644
>> --- a/arch/riscv/include/asm/barrier.h
>> +++ b/arch/riscv/include/asm/barrier.h
>> @@ -38,6 +38,21 @@
>>  #define __smp_rmb()    RISCV_FENCE(r,r)
>>  #define __smp_wmb()    RISCV_FENCE(w,w)
>>
>> +#define __smp_store_release(p, v)                    \
>> +do {                                    \
>> +    compiletime_assert_atomic_type(*p);                \
>> +    RISCV_FENCE(rw,w);                        \
>> +    WRITE_ONCE(*p, v);                        \
>> +} while (0)
>> +
>> +#define __smp_load_acquire(p)                        \
>> +({                                    \
>> +    typeof(*p) ___p1 = READ_ONCE(*p);                \
>> +    compiletime_assert_atomic_type(*p);                \
>> +    RISCV_FENCE(r,rw);                        \
>> +    ___p1;                                \
>> +})
>> +
>>  /*
>>   * This is a very specific barrier: it's currently only used in two places in
>>   * the kernel, both in the scheduler.  See include/linux/spinlock.h for the two
>
> I'm adding Daniel just in case I misunderstood what's going on here,
> but these look good to me. As this is a non-trivial memory model
> change I'm going to let it bake in linux-next for a bit just so it
> gets some visibility.

Looks good to me too. In particular, it also covers the
Write->release(p)->acquire(p)->Write ordering that we were debating
in the broader LKMM thread, which is good.

Dan

>
> Thanks

2018-02-28 00:17:49

by Andrea Parri

[permalink] [raw]
Subject: Re: [PATCH] riscv/barrier: Define __smp_{store_release,load_acquire}

On Tue, Feb 27, 2018 at 02:20:37PM -0800, Daniel Lustig wrote:
> On 2/27/2018 10:21 AM, Palmer Dabbelt wrote:
> > On Mon, 26 Feb 2018 18:24:11 PST (-0800), [email protected] wrote:
> >> Introduce __smp_{store_release,load_acquire}, and rely on the generic
> >> definitions for smp_{store_release,load_acquire}. This avoids the use
> >> of full ("rw,rw") fences on SMP.
> >>
> >> Signed-off-by: Andrea Parri <[email protected]>
> >> ---
> >> ?arch/riscv/include/asm/barrier.h | 15 +++++++++++++++
> >> ?1 file changed, 15 insertions(+)
> >>
> >> diff --git a/arch/riscv/include/asm/barrier.h b/arch/riscv/include/asm/barrier.h
> >> index 5510366d169ae..d4628e4b3a5ea 100644
> >> --- a/arch/riscv/include/asm/barrier.h
> >> +++ b/arch/riscv/include/asm/barrier.h
> >> @@ -38,6 +38,21 @@
> >> ?#define __smp_rmb()??? RISCV_FENCE(r,r)
> >> ?#define __smp_wmb()??? RISCV_FENCE(w,w)
> >>
> >> +#define __smp_store_release(p, v)??????????????????? \
> >> +do {??????????????????????????????????? \
> >> +??? compiletime_assert_atomic_type(*p);??????????????? \
> >> +??? RISCV_FENCE(rw,w);??????????????????????? \
> >> +??? WRITE_ONCE(*p, v);??????????????????????? \
> >> +} while (0)
> >> +
> >> +#define __smp_load_acquire(p)??????????????????????? \
> >> +({??????????????????????????????????? \
> >> +??? typeof(*p) ___p1 = READ_ONCE(*p);??????????????? \
> >> +??? compiletime_assert_atomic_type(*p);??????????????? \
> >> +??? RISCV_FENCE(r,rw);??????????????????????? \
> >> +??? ___p1;??????????????????????????????? \
> >> +})
> >> +
> >> ?/*
> >> ? * This is a very specific barrier: it's currently only used in two places in
> >> ? * the kernel, both in the scheduler.? See include/linux/spinlock.h for the two
> >
> > I'm adding Daniel just in case I misunderstood what's going on here,
> > but these look good to me. As this is a non-trivial memory model
> > change I'm going to let it bake in linux-next for a bit just so it
> > gets some visibility.
>
> Looks good to me too. In particular, it also covers the
> Write->release(p)->acquire(p)->Write ordering that we were debating
> in the broader LKMM thread, which is good.

Yeah, I think that other changes would be required to completely cover
the issues debated in that thread: I plan to prepare and to post a new
series/RFC to address those (unless someone precedes me of course ;-).

Andrea


>
> Dan
>
> >
> > Thanks