[dpdk-dev] [PATCH] test/bpf: use hton instead of __builtin_bswap

Malvika Gupta Malvika.Gupta at arm.com
Thu Oct 25 18:59:55 CEST 2018


Hi Konstantin,

At time of upstreaming the patch, I had compiled it on Cortex-A72 and it had worked with the -I parameter. But when tried building it on x86-64, I encountered the error that you were facing. From my understanding of this problem, this is a portability issue. As you pointed out, gcc and clang implement __builtin_constant_p in a different way on x86 but that may not be true for ARM. 

In such a case, to enable compilation on x86, we may want to write specific hton and ntoh functions. But this would become a tedious task because, in future, to enable compilation on other architectures besides x86 and ARM, we may have to repeat this process of writing analogs for each architecture. I think the best solution in the current scenario would be to use the -U parameter to enable compilation on x86 and later decide if we want to write a new patch with analogs for builtin hton() and ntoh().   

Best,
Malvika 

-----Original Message-----
From: Ananyev, Konstantin <konstantin.ananyev at intel.com> 
Sent: Friday, October 19, 2018 7:17 AM
To: Malvika Gupta <Malvika.Gupta at arm.com>
Cc: dev at dpdk.org; Gavin Hu (Arm Technology China) <Gavin.Hu at arm.com>; Honnappa Nagarahalli <Honnappa.Nagarahalli at arm.com>; nd <nd at arm.com>
Subject: RE: [PATCH] test/bpf: use hton instead of __builtin_bswap


Hi Malvika,

> 
> Hi Konstantin,
> 
> I installed the clang version 4.0.1 to check for the issue you were 
> facing with -O2 compilation. I was able to compile with -O2 and -O0 optimization without any errors. Please see the exact command I used and the following output for your reference:
> 
> $ clang -O2 -target bpf -I /usr/include/aarch64-linux-gnu/ -c t1.c $ 
> clang -O0 -target bpf -I /usr/include/aarch64-linux-gnu/ -c t1.c $ 
> clang -v clang version 4.0.1-10 (tags/RELEASE_401/final)
> Target: aarch64-unknown-linux-gnu
> Thread model: posix
> InstalledDir: /usr/bin
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/6
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/6.4.0
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/7
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/7.3.0
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/8
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/8.0.1
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/6 
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/6.4.0
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/7 
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/7.3.0
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/8 
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/8.0.1
> Selected GCC installation: /usr/bin/../lib/gcc/aarch64-linux-gnu/7.3.0
> Candidate multilib: .;@m64
> Selected multilib: .;@m64
> 
> I hope this was helpful to you.

Sorry for delay.
I tried with clang 6.0, but still seeing same issue.
Digging a bit more, I think the culprit is here:

#ifdef __GNUC__
# if __GNUC__ >= 2
#  define __bswap_16(x) \
     (__extension__                                                           \
      ({ unsigned short int __v, __x = (unsigned short int) (x);              \
         if (__builtin_constant_p (__x))                                      \     <------ evaluates to FALSE !!!
           __v = __bswap_constant_16 (__x);                                   \
         else                                                                 \
           __asm__ ("rorw $8, %w0"                                            \
                    : "=r" (__v)                                              \
                    : "0" (__x)                                               \
                    : "cc");                                                  \
         __v; }))
# else
/* This is better than nothing.  */
#  define __bswap_16(x) \
     (__extension__                                                           \
      ({ unsigned short int __x = (unsigned short int) (x);                   \
         __bswap_constant_16 (__x); }))
# endif
#else
static __inline unsigned short int
__bswap_16 (unsigned short int __bsx)
{
  return __bswap_constant_16 (__bsx);
}
#endif

Seems that gcc and clang implements __builtin_constant_p in a different way.
Let say, the following function:
int bbb1(void) {uint16_t x=0; return__bswap_constant_16(x);} when compiled with gcc would return 1, while when compiled with clang - 0.
Adding 'const' to definition of 'x' cures the problem for clang.
I don't know why there is no 'const' in original __bswap_16() macro. 
To overcome that problem, I used:
clang -O2 -U__GNUC__ -target bpf -c t1.c

Another possible option - define our own analogs for buitin hton*().
Konstantin


> Best Regards,
> Malvika Gupta
> 
> 
> -----Original Message-----
> From: Malvika Gupta
> Sent: Wednesday, September 5, 2018 4:43 PM
> To: Ananyev, Konstantin <konstantin.ananyev at intel.com>
> Cc: dev at dpdk.org; Gavin Hu (Arm Technology China) <Gavin.Hu at arm.com>; 
> Honnappa Nagarahalli <Honnappa.Nagarahalli at arm.com>; Brian Brooks 
> <Brian.Brooks at arm.com>; nd <nd at arm.com>
> Subject: RE: [PATCH] test/bpf: use hton instead of __builtin_bswap
> 
> Hi Ananyev,
> 
> I used clang version 6.0.0. Please see the following output for your reference.
> 
> $ clang -v
> clang version 6.0.0-1ubuntu2 (tags/RELEASE_600/final)
> Target: aarch64-unknown-linux-gnu
> Thread model: posix
> InstalledDir: /usr/bin
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/6
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/6.4.0
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/7
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/7.3.0
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/8
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/aarch64-linux-gnu/8.0.1
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/6 
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/6.4.0
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/7 
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/7.3.0
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/8 
> Found candidate GCC installation: /usr/lib/gcc/aarch64-linux-gnu/8.0.1
> Selected GCC installation: /usr/bin/../lib/gcc/aarch64-linux-gnu/7.3.0
> Candidate multilib: .;@m64
> Selected multilib: .;@m64
> 
> Also, the code compiles with both -O2 and -O0 for me.
> 
> I hope this was helpful
> Best,
> Malvika
> 
> -----Original Message-----
> From: Ananyev, Konstantin <konstantin.ananyev at intel.com>
> Sent: Tuesday, September 4, 2018 8:56 AM
> To: Malvika Gupta <Malvika.Gupta at arm.com>
> Cc: dev at dpdk.org; Gavin Hu (Arm Technology China) <Gavin.Hu at arm.com>; 
> Honnappa Nagarahalli <Honnappa.Nagarahalli at arm.com>; Brian Brooks 
> <Brian.Brooks at arm.com>; nd <nd at arm.com>
> Subject: RE: [PATCH] test/bpf: use hton instead of __builtin_bswap
> 
> Hi,
> 
> >
> > Convert host machine endianness to networking endianness for 
> > comparison of incoming packets with BPF filter
> >
> >
> > Signed-off-by: Malvika Gupta <malvika.gupta at arm.com>
> > Reviewed-by: Gavin Hu <gavin.hu at arm.com>
> > Reviewed-by: Brian Brooks <brian.brooks at arm.com>
> > Suggested-by: Brian Brooks <brian.brooks at arm.com>
> > ---
> >  test/bpf/t1.c | 7 ++++---
> >  test/bpf/t3.c | 3 ++-
> >  2 files changed, 6 insertions(+), 4 deletions(-)
> >
> > diff --git a/test/bpf/t1.c b/test/bpf/t1.c index 
> > 60f9434ab..7943fcf34
> > 100644
> > --- a/test/bpf/t1.c
> > +++ b/test/bpf/t1.c
> > @@ -28,24 +28,25 @@
> >  #include <net/ethernet.h>
> >  #include <netinet/ip.h>
> >  #include <netinet/udp.h>
> > +#include <arpa/inet.h>
> >
> >  uint64_t
> >  entry(void *pkt)
> >  {
> >  	struct ether_header *ether_header = (void *)pkt;
> >
> > -	if (ether_header->ether_type != __builtin_bswap16(0x0800))
> > +	if (ether_header->ether_type != htons(0x0800))
> 
> Which version of clang do you use?
> With my one I get:
> $ clang -O2 -target bpf -c t1.c
> t1.c:37:34: error: couldn't allocate output register for constraint 'r'
>         if (ether_header->ether_type != ntohs(0x0800))
>                                         ^
> /usr/include/netinet/in.h:402:21: note: expanded from macro 'ntohs'
> #   define ntohs(x)     __bswap_16 (x)
>                         ^
> /usr/include/bits/byteswap-16.h:31:14: note: expanded from macro '__bswap_16'
>            __asm__ ("rorw $8, %w0"
> 
> With '-O0' it compiles ok.
> 
> $ clang -v
> clang version 4.0.1 (tags/RELEASE_401/final)
> Target: x86_64-unknown-linux-gnu
> Thread model: posix
> InstalledDir: /usr/bin
> Found candidate GCC installation: 
> /usr/bin/../lib/gcc/x86_64-redhat-linux/7
> Found candidate GCC installation: /usr/lib/gcc/x86_64-redhat-linux/7
> Selected GCC installation: /usr/bin/../lib/gcc/x86_64-redhat-linux/7
> Candidate multilib: .;@m64
> Candidate multilib: 32;@m32
> Selected multilib: .;@m64
> 
> Konstantin
> 
> >  		return 0;
> >
> >  	struct iphdr *iphdr = (void *)(ether_header + 1);
> >  	if (iphdr->protocol != 17 || (iphdr->frag_off & 0x1ffff) != 0 ||
> > -			iphdr->daddr != __builtin_bswap32(0x1020304))
> > +			iphdr->daddr != htonl(0x1020304))
> >  		return 0;
> >
> >  	int hlen = iphdr->ihl * 4;
> >  	struct udphdr *udphdr = (void *)iphdr + hlen;
> >
> > -	if (udphdr->dest !=  __builtin_bswap16(5000))
> > +	if (udphdr->dest != htons(5000))
> >  		return 0;
> >
> >  	return 1;
> > diff --git a/test/bpf/t3.c b/test/bpf/t3.c index 
> > 531b9cb8c..24298b7c7
> > 100644
> > --- a/test/bpf/t3.c
> > +++ b/test/bpf/t3.c
> > @@ -17,6 +17,7 @@
> >  #include <net/ethernet.h>
> >  #include <rte_config.h>
> >  #include "mbuf.h"
> > +#include <arpa/inet.h>
> >
> >  extern void rte_pktmbuf_dump(FILE *, const struct rte_mbuf *, 
> > unsigned int);
> >
> > @@ -29,7 +30,7 @@ entry(const void *pkt)
> >  	mb = pkt;
> >  	eth = rte_pktmbuf_mtod(mb, const struct ether_header *);
> >
> > -	if (eth->ether_type == __builtin_bswap16(ETHERTYPE_ARP))
> > +	if (eth->ether_type == htons(ETHERTYPE_ARP))
> >  		rte_pktmbuf_dump(stdout, mb, 64);
> >
> >  	return 1;
> > --
> > 2.17.1



More information about the dev mailing list