Re: [PATCH 2/4] arch: Add lightweight memory barriers fast_rmb() and fast_wmb()

[Alexander Duyck <alexander.h.duyck@redhat.com>]

On 11/17/2014 03:17 PM, Paul E. McKenney wrote:
> On Mon, Nov 17, 2014 at 01:11:57PM -0800, Alexander Duyck wrote:
>> On 11/17/2014 12:18 PM, Paul E. McKenney wrote:
>>> On Mon, Nov 17, 2014 at 09:18:13AM -0800, Alexander Duyck wrote:
>>>> There are a number of situations where the mandatory barriers rmb() and
>>>> wmb() are used to order memory/memory operations in the device drivers
>>>> and those barriers are much heavier than they actually need to be.  For
>>>> example in the case of PowerPC wmb() calls the heavy-weight sync
>>>> instruction when for memory/memory operations all that is really needed is
>>>> an lsync or eieio instruction.
>>> Is this still the case if one of the memory operations is MMIO?  Last
>>> I knew, it was not.
>> This barrier is not meant for use in MMIO operations, for that you
>> still need a full barrier as I call out in the documentation
>> section. What the barrier does is allow for a lightweight barrier
>> for accesses to coherent system memory. So for example many device
>> drivers have to perform a read of the descriptor to see if the
>> device is done with it. We need an rmb() following that check to
>> prevent any other accesses.
>>
>> Right now on x86 that rmb() becomes an lfence instruction and is
>> quite expensive, and as it turns out we don't need it since the x86
>> doesn't reorder reads. The same kind of thing applies to PowerPC,
>> only in that case we use a sync when what we really need is a
>> lwsync.
> Would it make sense to have a memory barrier that enforced the
> non-store-buffer orderings, that is prior reads before later
> reads and writes and prior writes before later writes?  This was
> discussed earlier this year ((http://lwn.net/Articles/586838/,
> https://lwn.net/Articles/588300/).  If I recall correctly, one of
> the biggest obstacles was the name.  ;-)

You''re talking bout acquire and release barriers, or something else?  
For most devices the two barriers I have defined should do the job, I 
had tried doing load_acquire/store_release type functions in the 
previous path set and that was shot down as the preference seemed to be 
for barriers instead to remove some of the abstraction as to what was 
occurring.

>>>> This commit adds a fast (and loose) version of the mandatory memory
>>>> barriers rmb() and wmb().  The prefix to the name is actually based on the
>>>> version of the functions that already exist in the mips and tile trees.
>>>> However I thought it applicable since it gets at what we are trying to
>>>> accomplish with these barriers and somethat implies their risky nature.
>>>>
>>>> These new barriers are not as safe as the standard rmb() and wmb().
>>>> Specifically they do not guarantee ordering between cache-enabled and
>>>> cache-inhibited memories.  The primary use case for these would be to
>>>> enforce ordering of memory reads/writes when accessing cache-enabled memory
>>>> that is shared between the CPU and a device.
>>>>
>>>> It may also be noted that there is no fast_mb().  This is due to the fact
>>>> that most architectures didn't seem to have a good way to do a full memory
>>>> barrier quickly and so they usually resorted to an mb() for their smp_mb
>>>> call.  As such there no point in adding a fast_mb() function if it is going
>>>> to map to mb() for all architectures anyway.
>>> I must confess that I still don't entirely understand the motivation.
>> The motivation is to provide finer grained barriers. So this
>> provides an in-between that allows us to "choose the right hammer".
>> In the case of PowerPC it is the difference between sync/lwsync, on
>> ARM it is dsb()/dmb(), and on x86 it is lfence/barrier().
> Ah, so ARM will motivate a fast_wmb(), given its instruction set.

Actually it was x86 that started this,  lfence or sfence is much more 
expensive then just barrier().  From there I realized we had issues in 
PowerPC as well with sync vs lwsync, and ARM with dsb() vs dmb().

>> <snip>
>>
>>>> diff --git a/arch/powerpc/include/asm/barrier.h b/arch/powerpc/include/asm/barrier.h
>>>> index cb6d66c..f480097 100644
>>>> --- a/arch/powerpc/include/asm/barrier.h
>>>> +++ b/arch/powerpc/include/asm/barrier.h
>>>> @@ -36,22 +36,20 @@
>>>>
>>>>   #define set_mb(var, value)	do { var = value; mb(); } while (0)
>>>>
>>>> -#ifdef CONFIG_SMP
>>>> -
>>>>   #ifdef __SUBARCH_HAS_LWSYNC
>>>>   #    define SMPWMB      LWSYNC
>>>>   #else
>>>>   #    define SMPWMB      eieio
>>>>   #endif
>>>>
>>>> -#define __lwsync()	__asm__ __volatile__ (stringify_in_c(LWSYNC) : : :"memory")
>>>> +#define fast_rmb()	__asm__ __volatile__ (stringify_in_c(LWSYNC) : : :"memory")
>>>> +#define fast_wmb()	__asm__ __volatile__ (stringify_in_c(SMPWMB) : : :"memory")
>>>>
>>>> +#ifdef CONFIG_SMP
>>>>   #define smp_mb()	mb()
>>>> -#define smp_rmb()	__lwsync()
>>>> -#define smp_wmb()	__asm__ __volatile__ (stringify_in_c(SMPWMB) : : :"memory")
>>>> +#define smp_rmb()	fast_rmb()
>>>> +#define smp_wmb()	fast_wmb()
>>>>   #else
>>>> -#define __lwsync()	barrier()
>>>> -
>>>>   #define smp_mb()	barrier()
>>>>   #define smp_rmb()	barrier()
>>>>   #define smp_wmb()	barrier()
>>>> @@ -69,10 +67,16 @@
>>>>   #define data_barrier(x)	\
>>>>   	asm volatile("twi 0,%0,0; isync" : : "r" (x) : "memory");
>>>>
>>>> +/*
>>>> + * The use of smp_rmb() in these functions are actually meant to map from
>>>> + * smp_rmb()->fast_rmb()->LWSYNC.  This way if smp is disabled then
>>>> + * smp_rmb()->barrier(), or if the platform doesn't support lwsync it will
>>>> + * map to the more heavy-weight sync.
>>>> + */
>>>>   #define smp_store_release(p, v)						\
>>>>   do {									\
>>>>   	compiletime_assert_atomic_type(*p);				\
>>>> -	__lwsync();							\
>>>> +	smp_rmb();							\
>>> This is not good at all.  For smp_store_release(), we absolutely
>>> must order prior loads and stores against the assignment on the following
>>> line.  This is not something that smp_rmb() does, nor is it something
>>> that smp_wmb() does.  Yes, it might happen to now, but this could easily
>>> break in the future -- plus this change is extremely misleading.
>>>
>>> The original __lwsync() is much more clear.
>> The problem I had with __lwsync is that it really wasn't all that
>> clear. It was the lwsync instruction if SMP was enabled, otherwise
>> it was just a barrier call. What I did is move the definition of
>> __lwsync in the SMP case into fast_rmb, which in turn is accessed by
>> smp_rmb. I tried to make this clear in the comment just above the
>> two calls. The resultant assembly code should be exactly the same.
>>
>> What I could do is have it added back as a smp_lwsync if that works
>> for you. That way there is something there to give you a hint that
>> it becomes a barrier() call as soon as SMP is disabled.
> An smp_lwsync() would be a great improvement!
>
> 							Thanx, Paul
>

Okay, that will be in the next patch then.

Thanks,

Alex