> On 27 Aug 2017, at 12:04, mrityunjay.kum...@wipro.com wrote:
>
> Dear Team
> I would like bring it to your kind notice of below code of vpp-1707-dpdk
> plunging.
>
> static_always_inline void dpdk_prefetch_buffer_by_index (vlib_main_t * vm,
> u32 bi)
> {
> vlib_buffer_t *b;
> struct rte_mbuf *mb;
> b = vlib_get_buffer (vm, bi);
> mb = rte_mbuf_from_vlib_buffer (b);
> CLIB_PREFETCH (mb, CLIB_CACHE_LINE_BYTES, LOAD);
> CLIB_PREFETCH (b, CLIB_CACHE_LINE_BYTES, LOAD);
> }
>
> #define CLIB_PREFETCH(addr,size,type) \
> do { \
> void * _addr = (addr); \
> \
> ASSERT ((size) <= 4*CLIB_CACHE_LINE_BYTES); \
> _CLIB_PREFETCH (0, size, type); \
> _CLIB_PREFETCH (1, size, type); \
> _CLIB_PREFETCH (2, size, type); \
> _CLIB_PREFETCH (3, size, type); \
> } while (0)
>
> <pastedImage.png>
>
>
> Here , Sizeof(rte_mbuf) = 128 and sizeof(vlib_buffer_t) = 128 +
> HEAD_ROOM(128)= 256.
>
> In above code part, vlib_buffer is ahead of 128 bytes from start of rte_mbuf
> structure. As i understood one CLIB_PREFETCH will load 256 bytes from
> memory. ,hence total pre-fetch is 512 bytes. As per the above code first
> CLIB_PREFETCH will load 256, which includes 128 of rte_mbuf + 128 of
> vlib_buffer as well . 2nd CLIB_PREFETCH will also load vlib_buffer whcih as
> ready has been loaded.
>
> I must say duplication in prefetching the memory. Please correct me if I am
> wrong.
Hi MJ,
Yes, you are wrong. Each invocation of CLIB_PREFETCH in the inline function you
listed above will prefetch one cacheline, so 64 bytes, not 256.
Please look at _CLIB_PREFETCH macro for details…
_______________________________________________
vpp-dev mailing list
vpp-dev@lists.fd.io
https://lists.fd.io/mailman/listinfo/vpp-dev
