[dpdk-dev] Fwd: [PATCH v3 2/2] eal/x86: Use lock-prefixed instructions to reduce cost of rte_smp_mb()

[Michael S. Tsirkin]

On Mon, Jan 15, 2018 at 04:15:00PM +0100, Maxime Coquelin wrote:
> Hi Michael,
> 
> FYI:
> 
> -------- Forwarded Message --------
> Subject: [dpdk-dev] [PATCH v3 2/2] eal/x86: Use lock-prefixed instructions
> to reduce cost of rte_smp_mb()
> Date: Mon, 15 Jan 2018 15:09:31 +0000
> From: Konstantin Ananyev <konstantin.ananyev at intel.com>
> To: dev at dpdk.org
> CC: Konstantin Ananyev <konstantin.ananyev at intel.com>
> 
> On x86 it  is possible to use lock-prefixed instructions to get
> the similar effect as mfence.
> As pointed by Java guys, on most modern HW that gives a better
> performance than using mfence:
> https://shipilev.net/blog/2014/on-the-fence-with-dependencies/
> That patch adopts that technique for rte_smp_mb() implementation.
> On BDW 2.2 mb_autotest on single lcore reports 2X cycle reduction,
> i.e. from ~110 to ~55 cycles per operation.
> 
> Signed-off-by: Konstantin Ananyev <konstantin.ananyev at intel.com>
> Acked-by: Bruce Richardson <bruce.richardson at intel.com>
> ---
>  .../common/include/arch/x86/rte_atomic.h           | 44
> +++++++++++++++++++++-
>  1 file changed, 42 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/librte_eal/common/include/arch/x86/rte_atomic.h
> b/lib/librte_eal/common/include/arch/x86/rte_atomic.h
> index 8469f97e1..9d466d94a 100644
> --- a/lib/librte_eal/common/include/arch/x86/rte_atomic.h
> +++ b/lib/librte_eal/common/include/arch/x86/rte_atomic.h
> @@ -26,12 +26,52 @@ extern "C" {
>   #define	rte_rmb() _mm_lfence()
>  -#define rte_smp_mb() rte_mb()
> -
>  #define rte_smp_wmb() rte_compiler_barrier()
>   #define rte_smp_rmb() rte_compiler_barrier()
>  +/*
> + * From Intel Software Development Manual; Vol 3;
> + * 8.2.2 Memory Ordering in P6 and More Recent Processor Families:
> + * ...
> + * . Reads are not reordered with other reads.
> + * . Writes are not reordered with older reads.
> + * . Writes to memory are not reordered with other writes,
> + *   with the following exceptions:
> + *   . streaming stores (writes) executed with the non-temporal move
> + *     instructions (MOVNTI, MOVNTQ, MOVNTDQ, MOVNTPS, and MOVNTPD); and
> + *   . string operations (see Section 8.2.4.1).
> + *  ...
> + * . Reads may be reordered with older writes to different locations but
> not
> + * with older writes to the same location.
> + * . Reads or writes cannot be reordered with I/O instructions,
> + * locked instructions, or serializing instructions.
> + * . Reads cannot pass earlier LFENCE and MFENCE instructions.
> + * . Writes ... cannot pass earlier LFENCE, SFENCE, and MFENCE
> instructions.
> + * . LFENCE instructions cannot pass earlier reads.
> + * . SFENCE instructions cannot pass earlier writes ...
> + * . MFENCE instructions cannot pass earlier reads, writes ...
> + *
> + * As pointed by Java guys, that makes possible to use lock-prefixed
> + * instructions to get the same effect as mfence and on most modern HW
> + * that gives a better perfomance then using mfence:
> + * https://shipilev.net/blog/2014/on-the-fence-with-dependencies/
> + * Basic idea is to use lock prefixed add with some dummy memory location
> + * as the destination. From their experiments 128B(2 cache lines) below
> + * current stack pointer looks like a good candidate.
> + * So below we use that techinque for rte_smp_mb() implementation.
> + */
> +
> +static __rte_always_inline void
> +rte_smp_mb(void)
> +{
> +#ifdef RTE_ARCH_I686
> +	asm volatile("lock addl $0, -128(%%esp); " ::: "memory");
> +#else
> +	asm volatile("lock addl $0, -128(%%rsp); " ::: "memory");
> +#endif
> +}
> +
>  #define rte_io_mb() rte_mb()
>   #define rte_io_wmb() rte_compiler_barrier()

In my testing this appears to be suboptimal when the calling
function is large. The following seems to work better:

+static __rte_always_inline void
+rte_smp_mb(void)
+{
+#ifdef RTE_ARCH_I686
+	asm volatile("lock addl $0, -132(%%esp); " ::: "memory");
+#else
+	asm volatile("lock addl $0, -132(%%rsp); " ::: "memory");
+#endif
+}
+

The reason most likely is that 128 still overlaps the x86
red zone by 4 bytes.

Feel free to reuse, and add
Signed-off-by: Michael S. Tsirkin <mst at redhat.com>


> -- 
> 2.13.6