Re: [PATCH v2] eal: add notes to SMP memory barrier APIs

["Mattias Rönnblom" <hofors@lysator.liu.se>]

On 2023-07-03 09:02, Ruifeng Wang wrote:
>> -----Original Message-----
>> From: Mattias Rönnblom <hofors@lysator.liu.se>
>> Sent: Friday, June 30, 2023 3:44 AM
>> To: Ruifeng Wang <Ruifeng.Wang@arm.com>; thomas@monjalon.net; david.marchand@redhat.com
>> Cc: dev@dpdk.org; roretzla@linux.microsoft.com; konstantin.v.ananyev@yandex.ru; Honnappa
>> Nagarahalli <Honnappa.Nagarahalli@arm.com>; nd <nd@arm.com>
>> Subject: Re: [PATCH v2] eal: add notes to SMP memory barrier APIs
>>
>> On 2023-06-26 09:12, Ruifeng Wang wrote:
>>> The rte_smp_xx() APIs are deprecated. But it is not mentioned in the
>>> function header.
>>> Added notes in function header for clarification.
>>>
>>> Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
>>> ---
>>> v2: Made the notes more specific.
>>>
>>>    lib/eal/include/generic/rte_atomic.h | 19 +++++++++++++++++++
>>>    1 file changed, 19 insertions(+)
>>>
>>> diff --git a/lib/eal/include/generic/rte_atomic.h
>>> b/lib/eal/include/generic/rte_atomic.h
>>> index 58df843c54..35e0041ce6 100644
>>> --- a/lib/eal/include/generic/rte_atomic.h
>>> +++ b/lib/eal/include/generic/rte_atomic.h
>>> @@ -55,6 +55,11 @@ static inline void rte_rmb(void);
>>>     * Guarantees that the LOAD and STORE operations that precede the
>>>     * rte_smp_mb() call are globally visible across the lcores
>>>     * before the LOAD and STORE operations that follows it.
>>> + *
>>> + * @note
>>> + *  This function is deprecated. It provides fence synchronization
>>> + *  primitive but doesn't take memory order parameter.
>>> + *  rte_atomic_thread_fence() should be used instead.
>>
>> I can't see why coding the memory model semantics into the name, rather than by
>> specification-by-means-of-a-parameter, could be the real issue.
>> Could you explain? Seems like just different syntax to me.
> 
> Yes, rte_smp_xx and rte_atomic_thread_fence have different syntaxes.
> 
> The compiler atomic builtins were accepted for memory ordering. It comprises atomic arithmetic,
> atomic load/store, and atomic fence. It is simpler and clearer to do memory ordering by using
> the atomic builtins whenever possible.
> rte_smp_xx has functionality overlap with atomic fence builtins but with different memory model
> semantics and different syntaxes. Because of the differences, it will make memory ordering a little
> more complex if rte_smp_xx is kept aside atomic builtins suite.
> 

I wasn't arguing for keeping Linux kernel-style barriers. It was just 
the rationale that seemed flawed to me.

If Linux kernel-style memory barriers took a memory model parameter, we 
would still prefer C11-style GCC barrier intrinsics (for this release).

>>
>> The old <rte_atomic.h> atomic arithmetic and atomic load/store operations suffered from
>> unspecified semantics in regards to any ordering they imposed on other memory accesses. I
>> guess that shortcoming could be described as a "missing parameter", although that too
>> would be misleading. Unclear semantics seems not be the case for the kernel-style barriers
>> though.
>>
>>>     */
>>>    static inline void rte_smp_mb(void);
>>>
>>> @@ -64,6 +69,11 @@ static inline void rte_smp_mb(void);
>>>     * Guarantees that the STORE operations that precede the
>>>     * rte_smp_wmb() call are globally visible across the lcores
>>>     * before the STORE operations that follows it.
>>> + *
>>> + * @note
>>> + *  This function is deprecated. It provides fence synchronization
>>> + *  primitive but doesn't take memory order parameter.
>>> + *  rte_atomic_thread_fence() should be used instead.
>>>     */
>>>    static inline void rte_smp_wmb(void);
>>>
>>> @@ -73,6 +83,11 @@ static inline void rte_smp_wmb(void);
>>>     * Guarantees that the LOAD operations that precede the
>>>     * rte_smp_rmb() call are globally visible across the lcores
>>>     * before the LOAD operations that follows it.
>>> + *
>>> + * @note
>>> + *  This function is deprecated. It provides fence synchronization
>>> + *  primitive but doesn't take memory order parameter.
>>> + *  rte_atomic_thread_fence() should be used instead.
>>>     */
>>>    static inline void rte_smp_rmb(void);
>>>    ///@}
>>> @@ -122,6 +137,10 @@ static inline void rte_io_rmb(void);
>>>
>>>    /**
>>>     * Synchronization fence between threads based on the specified memory order.
>>> + *
>>> + * @param memorder
>>> + *   The memory order defined by compiler atomic builtin at:
>>> + *   https://gcc.gnu.org/onlinedocs/gcc/_005f_005fatomic-Builtins.html
>>>     */
>>>    static inline void rte_atomic_thread_fence(int memorder);
>>>