[dpdk-dev] librte_lpm: use field access instead of type conversion.

  On Wed, Feb 11, 2015 at 02:12:59PM +0800, xuelin.shi@freescale.com wrote:
> From: Xuelin Shi <xuelin.shi@freescale.com>
> 
> struct tbl_entry{
>         uint8_t next_hop;
>         uint8_t valid :1;
>         uint8_t valid_group :1;
>         uint8_t depth :6
> }
> uint16_t tbl = (uint16_t)tbl_entry;
> next_hop = (uint8_t)tbl;
> 
> next_hop cannot get the correct value of the field
> if the cpu arch is BIG_ENDIAN.
> 
> change it to field access.
> 
> Signed-off-by: Xuelin Shi <xuelin.shi@freescale.com>
> ---
>  lib/librte_lpm/rte_lpm.h | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h
> index 586300b..1af150c 100644
> --- a/lib/librte_lpm/rte_lpm.h
> +++ b/lib/librte_lpm/rte_lpm.h
> @@ -273,6 +273,7 @@ rte_lpm_lookup(struct rte_lpm *lpm, uint32_t ip, uint8_t *next_hop)
>  {
>  	unsigned tbl24_index = (ip >> 8);
>  	uint16_t tbl_entry;
> +	struct rte_lpm_tbl8_entry *entry;
>  
>  	/* DEBUG: Check user input arguments. */
>  	RTE_LPM_RETURN_IF_TRUE(((lpm == NULL) || (next_hop == NULL)), -EINVAL);
> @@ -290,8 +291,10 @@ rte_lpm_lookup(struct rte_lpm *lpm, uint32_t ip, uint8_t *next_hop)
>  		tbl_entry = *(const uint16_t *)&lpm->tbl8[tbl8_index];
>  	}
>  
> -	*next_hop = (uint8_t)tbl_entry;
> -	return (tbl_entry & RTE_LPM_LOOKUP_SUCCESS) ? 0 : -ENOENT;
> +	entry = (struct rte_lpm_tbl8_entry *)&tbl_entry;
> +	*next_hop = entry->next_hop;
> +
> +	return (entry->valid) ? 0 : -ENOENT;
>  }
>  
>  /**
> -- 
> 1.9.1
> 
I've run a quick test using "lpm_autotest" inside the test app, and on my (Intel)
platform, this patch has a small (but none-the-less significant) performance
regression. How about the below as an alternative fix?

/Bruce
  

Hi,

Needs more consideration.

RTE_LPM_VALID_EXT_ENTRY_BITMASK is defined as 0x0030, also little endian assumed.
Seems like the struct bit field also need some position conversion.

I would like to send v2 patch to fix that.

Thanks,
Shi Xuelin

> -----Original Message-----
> From: Bruce Richardson [mailto:bruce.richardson@intel.com]
> Sent: Thursday, February 12, 2015 19:18
> To: Shi Xuelin-B29237
> Cc: thomas.monjalon@6wind.com; dev@dpdk.org
> Subject: Re: [dpdk-dev] [PATCH] librte_lpm: use field access instead of
> type conversion.
> 
> On Wed, Feb 11, 2015 at 02:12:59PM +0800, xuelin.shi@freescale.com wrote:
> > From: Xuelin Shi <xuelin.shi@freescale.com>
> >
> > struct tbl_entry{
> >         uint8_t next_hop;
> >         uint8_t valid :1;
> >         uint8_t valid_group :1;
> >         uint8_t depth :6
> > }
> > uint16_t tbl = (uint16_t)tbl_entry;
> > next_hop = (uint8_t)tbl;
> >
> > next_hop cannot get the correct value of the field if the cpu arch is
> > BIG_ENDIAN.
> >
> > change it to field access.
> >
> > Signed-off-by: Xuelin Shi <xuelin.shi@freescale.com>
> > ---
> >  lib/librte_lpm/rte_lpm.h | 7 +++++--
> >  1 file changed, 5 insertions(+), 2 deletions(-)
> >
> > diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h index
> > 586300b..1af150c 100644
> > --- a/lib/librte_lpm/rte_lpm.h
> > +++ b/lib/librte_lpm/rte_lpm.h
> > @@ -273,6 +273,7 @@ rte_lpm_lookup(struct rte_lpm *lpm, uint32_t ip,
> > uint8_t *next_hop)  {
> >  	unsigned tbl24_index = (ip >> 8);
> >  	uint16_t tbl_entry;
> > +	struct rte_lpm_tbl8_entry *entry;
> >
> >  	/* DEBUG: Check user input arguments. */
> >  	RTE_LPM_RETURN_IF_TRUE(((lpm == NULL) || (next_hop == NULL)),
> > -EINVAL); @@ -290,8 +291,10 @@ rte_lpm_lookup(struct rte_lpm *lpm,
> uint32_t ip, uint8_t *next_hop)
> >  		tbl_entry = *(const uint16_t *)&lpm->tbl8[tbl8_index];
> >  	}
> >
> > -	*next_hop = (uint8_t)tbl_entry;
> > -	return (tbl_entry & RTE_LPM_LOOKUP_SUCCESS) ? 0 : -ENOENT;
> > +	entry = (struct rte_lpm_tbl8_entry *)&tbl_entry;
> > +	*next_hop = entry->next_hop;
> > +
> > +	return (entry->valid) ? 0 : -ENOENT;
> >  }
> >
> >  /**
> > --
> > 1.9.1
> >
> I've run a quick test using "lpm_autotest" inside the test app, and on my
> (Intel) platform, this patch has a small (but none-the-less significant)
> performance regression. How about the below as an alternative fix?
> 
> /Bruce
> 
> diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h index
> 586300b..de6f1cb 100644
> --- a/lib/librte_lpm/rte_lpm.h
> +++ b/lib/librte_lpm/rte_lpm.h
> @@ -44,6 +44,7 @@
>  #include <stdint.h>
>  #include <stdlib.h>
>  #include <rte_branch_prediction.h>
> +#include <rte_byteorder.h>
>  #include <rte_memory.h>
>  #include <rte_common.h>
>  #include <rte_common_vect.h>
> @@ -287,7 +288,8 @@ rte_lpm_lookup(struct rte_lpm *lpm, uint32_t ip,
> uint8_t *next_hop)
>                 unsigned tbl8_index = (uint8_t)ip +
>                                 ((uint8_t)tbl_entry *
> RTE_LPM_TBL8_GROUP_NUM_ENTRIES);
> 
> -               tbl_entry = *(const uint16_t *)&lpm->tbl8[tbl8_index];
> +               tbl_entry = rte_cpu_to_le_16(
> +                               *(const uint16_t
> + *)&lpm->tbl8[tbl8_index]);
>         }
> 
>         *next_hop = (uint8_t)tbl_entry;