Re: [PATCH 4/4] perf: Optimize perf_output_begin() -- weaker memory barrier

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

* peterz@infradead.org (peterz@infradead.org) wrote:
> Apply the fancy new smp_load_acquire() and smp_store_release() to
> potentially avoid the full memory barrier in perf_output_begin().
> 
> On x86 (and other TSO like architectures) this removes all explicit
> memory fences, on weakly ordered systems this often allows the use of
> weaker barriers; in particular on powerpc we demote from a full sync
> to a cheaper lwsync.
> 
> Cc: Tony Luck <tony.luck@intel.com>
> Cc: Oleg Nesterov <oleg@redhat.com>
> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
> Cc: Frederic Weisbecker <fweisbec@gmail.com>
> Cc: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
> Cc: Michael Ellerman <michael@ellerman.id.au>
> Cc: Michael Neuling <mikey@neuling.org>
> Cc: Russell King <linux@arm.linux.org.uk>
> Cc: Geert Uytterhoeven <geert@linux-m68k.org>
> Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
> Cc: Linus Torvalds <torvalds@linux-foundation.org>
> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
> Cc: Victor Kaplansky <VICTORK@il.ibm.com>
> Suggested-by: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
> Signed-off-by: Peter Zijlstra <peterz@infradead.org>
> ---
>  kernel/events/ring_buffer.c |   62 +++++++++++++++++++++++++-------------------
>  1 file changed, 36 insertions(+), 26 deletions(-)
> 
> --- a/kernel/events/ring_buffer.c
> +++ b/kernel/events/ring_buffer.c
> @@ -41,6 +41,32 @@ static void perf_output_get_handle(struc
>  	handle->wakeup = local_read(&rb->wakeup);
>  }
>  
> +/*
> + * Our user visible data structure (struct perf_event_mmap_page) uses
> + * u64 values for ->data_head and ->data_tail to avoid size variance
> + * across 32/64 bit.
> + *
> + * Since you cannot mmap() a buffer larger than your memory address space
> + * we're naturally limited to unsigned long and can avoid writing the
> + * high word on 32bit systems (its always 0)
> + *
> + * This allows us to always use a single load/store.
> + */
> +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
> +static inline unsigned long *low_word(u64 *ptr)
> +{
> +	return (unsigned long *)ptr;
> +}
> +#else /* __ORDER_BIG_ENDIAN__ */
> +static inline unsigned long *low_word(u64 *ptr)
> +{
> +	void *_ptr = ptr;
> +	_ptr += sizeof(u64);
> +	_ptr -= sizeof(unsigned long);
> +	return (unsigned long *)_ptr;
> +}
> +#endif
> +
>  static void perf_output_put_handle(struct perf_output_handle *handle)
>  {
>  	struct ring_buffer *rb = handle->rb;
> @@ -61,28 +87,15 @@ static void perf_output_put_handle(struc
>  	 *
>  	 *   kernel				user
>  	 *
> -	 *   READ ->data_tail			READ ->data_head
> -	 *   smp_mb()	(A)			smp_rmb()	(C)
> +	 *   READ.acq ->data_tail  (A)		READ.acq ->data_head  (C)

I don't get how the barrier() in the TSO implementation of
smp_load_acquire (A) orders the following writes to $data after the
READ.acq of data_tail. I'm probably missing something.

Also, I don't get how the smp_load_acquire (C) with the barrier() (x86
TSO) orders READ $data after the READ.acq of data_head.

I don't have the TSO model fresh in memory however.

>  	 *   WRITE $data			READ $data
> -	 *   smp_wmb()	(B)			smp_mb()	(D)
> -	 *   STORE ->data_head			WRITE ->data_tail
> +	 *   STORE.rel ->data_head (B)		WRITE.rel ->data_tail (D)

You might want to choose either STORE or WRITE for consistency.

Thanks,

Mathieu

>  	 *
>  	 * Where A pairs with D, and B pairs with C.
>  	 *
> -	 * I don't think A needs to be a full barrier because we won't in fact
> -	 * write data until we see the store from userspace. So we simply don't
> -	 * issue the data WRITE until we observe it. Be conservative for now.
> -	 *
> -	 * OTOH, D needs to be a full barrier since it separates the data READ
> -	 * from the tail WRITE.
> -	 *
> -	 * For B a WMB is sufficient since it separates two WRITEs, and for C
> -	 * an RMB is sufficient since it separates two READs.
> -	 *
>  	 * See perf_output_begin().
>  	 */
> -	smp_wmb();
> -	rb->user_page->data_head = head;
> +	smp_store_release(low_word(&rb->user_page->data_head), head);
>  
>  	/*
>  	 * Now check if we missed an update -- rely on previous implied
> @@ -139,7 +152,13 @@ int perf_output_begin(struct perf_output
>  	perf_output_get_handle(handle);
>  
>  	do {
> -		tail = ACCESS_ONCE(rb->user_page->data_tail);
> +		tail = smp_load_acquire(low_word(&rb->user_page->data_tail));
> +		/*
> +		 * STORES of the data below cannot pass the ACQUIRE barrier.
> +		 *
> +		 * Matches with an smp_mb() or smp_store_release() in userspace
> +		 * as described in perf_output_put_handle().
> +		 */
>  		offset = head = local_read(&rb->head);
>  		if (!rb->overwrite &&
>  		    unlikely(CIRC_SPACE(head, tail, perf_data_size(rb)) < size))
> @@ -147,15 +166,6 @@ int perf_output_begin(struct perf_output
>  		head += size;
>  	} while (local_cmpxchg(&rb->head, offset, head) != offset);
>  
> -	/*
> -	 * Separate the userpage->tail read from the data stores below.
> -	 * Matches the MB userspace SHOULD issue after reading the data
> -	 * and before storing the new tail position.
> -	 *
> -	 * See perf_output_put_handle().
> -	 */
> -	smp_mb();
> -
>  	if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
>  		local_add(rb->watermark, &rb->wakeup);
>  
> 
> 

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com