net/memif: replace memcpy() with rte_memcpy() to improve perf
Checks
Commit Message
Replacing memcpy() with rte_memcpy() improved 64 byte packet
performance by 33% on a POWER9 system and by 10% on an x86_64
system.
Signed-off-by: David Christensen <drc@linux.vnet.ibm.com>
---
drivers/net/memif/rte_eth_memif.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
Comments
On Wed, 23 Jun 2021 09:49:35 -0700
David Christensen <drc@linux.vnet.ibm.com> wrote:
> Replacing memcpy() with rte_memcpy() improved 64 byte packet
> performance by 33% on a POWER9 system and by 10% on an x86_64
> system.
I see rte_memcpy was already used in the patch diff
> Signed-off-by: David Christensen <drc@linux.vnet.ibm.com>
> ---
> drivers/net/memif/rte_eth_memif.c | 7 ++++---
> 1 file changed, 4 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
> index d17222c612..330c9c2fd6 100644
> --- a/drivers/net/memif/rte_eth_memif.c
> +++ b/drivers/net/memif/rte_eth_memif.c
> @@ -384,8 +384,8 @@ eth_memif_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
> rte_pktmbuf_pkt_len(mbuf_head) += cp_len;
>
> rte_memcpy(rte_pktmbuf_mtod_offset(mbuf, void *, dst_off),
> - (uint8_t *)memif_get_buffer(proc_private, d0) + src_off,
> - cp_len);
> + (uint8_t *)memif_get_buffer(proc_private, d0)
> + + src_off, cp_len);
This just changes line break for no good reason.
>
> src_off += cp_len;
> dst_off += cp_len;
> @@ -644,7 +644,8 @@ eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
> }
> cp_len = RTE_MIN(dst_len, src_len);
>
> - rte_memcpy((uint8_t *)memif_get_buffer(proc_private, d0) + dst_off,
> + rte_memcpy((uint8_t *)
> + memif_get_buffer(proc_private, d0) + dst_off,
> rte_pktmbuf_mtod_offset(mbuf, void *, src_off),
> cp_len);
>
ditto.
Look like this patch is confused, the description does not match the code.
On 6/23/21 8:03 PM, Stephen Hemminger wrote:
> On Wed, 23 Jun 2021 09:49:35 -0700
> David Christensen <drc@linux.vnet.ibm.com> wrote:
>
>> Replacing memcpy() with rte_memcpy() improved 64 byte packet
>> performance by 33% on a POWER9 system and by 10% on an x86_64
>> system.
>
> I see rte_memcpy was already used in the patch diff
>
>> Signed-off-by: David Christensen <drc@linux.vnet.ibm.com>
>> ---
>> drivers/net/memif/rte_eth_memif.c | 7 ++++---
>> 1 file changed, 4 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/net/memif/rte_eth_memif.c b/drivers/net/memif/rte_eth_memif.c
>> index d17222c612..330c9c2fd6 100644
>> --- a/drivers/net/memif/rte_eth_memif.c
>> +++ b/drivers/net/memif/rte_eth_memif.c
>> @@ -384,8 +384,8 @@ eth_memif_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
>> rte_pktmbuf_pkt_len(mbuf_head) += cp_len;
>>
>> rte_memcpy(rte_pktmbuf_mtod_offset(mbuf, void *, dst_off),
>> - (uint8_t *)memif_get_buffer(proc_private, d0) + src_off,
>> - cp_len);
>> + (uint8_t *)memif_get_buffer(proc_private, d0)
>> + + src_off, cp_len);
>
> This just changes line break for no good reason.
>>
>> src_off += cp_len;
>> dst_off += cp_len;
>> @@ -644,7 +644,8 @@ eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
>> }
>> cp_len = RTE_MIN(dst_len, src_len);
>>
>> - rte_memcpy((uint8_t *)memif_get_buffer(proc_private, d0) + dst_off,
>> + rte_memcpy((uint8_t *)
>> + memif_get_buffer(proc_private, d0) + dst_off,
>> rte_pktmbuf_mtod_offset(mbuf, void *, src_off),
>> cp_len);
>>
>
> ditto.
>
> Look like this patch is confused, the description does not match the code.
>
I guess it was a patch on top of [1]. Marking this one as rejected since
I've applied style fixes before applying it.
[1]
https://patches.dpdk.org/project/dpdk/patch/20210623203154.72409-1-drc@linux.vnet.ibm.com/
@@ -384,8 +384,8 @@ eth_memif_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
rte_pktmbuf_pkt_len(mbuf_head) += cp_len;
rte_memcpy(rte_pktmbuf_mtod_offset(mbuf, void *, dst_off),
- (uint8_t *)memif_get_buffer(proc_private, d0) + src_off,
- cp_len);
+ (uint8_t *)memif_get_buffer(proc_private, d0)
+ + src_off, cp_len);
src_off += cp_len;
dst_off += cp_len;
@@ -644,7 +644,8 @@ eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
}
cp_len = RTE_MIN(dst_len, src_len);
- rte_memcpy((uint8_t *)memif_get_buffer(proc_private, d0) + dst_off,
+ rte_memcpy((uint8_t *)
+ memif_get_buffer(proc_private, d0) + dst_off,
rte_pktmbuf_mtod_offset(mbuf, void *, src_off),
cp_len);