Re: [dpdk-dev] 1.7.0 release failed to compile on linux

[Neil Horman <nhorman@tuxdriver.com>]

On Thu, Aug 14, 2014 at 06:15:24PM -0400, Aws Ismail wrote:
> 
> On 08/14/2014 05:44 PM, Qiaobing Xie wrote:
> >Hi,
> >
> >I got the following error when I tried to compile 1.7.0 release tarball on
> >a Linux box (Ubuntu 12.04/kernel=3.13.0-32-generic, gcc=4.6.3):
> >
> >============
> >In file included from /home/qxie/dpdk-1.7.0/lib/librte_eal/linuxapp/kni/ethtool/igb/e1000_osdep.h:41:0,
> >from
> >/home/qxie/dpdk-1.7.0/lib/librte_eal/linuxapp/kni/ethtool/igb/e1000_hw.h:31,
> >from /home/qxie/dpdk-1.7.0/lib/librte_eal/linuxapp/kni/ethtool/igb/e1000_api.h:31,
> >from /home/qxie/dpdk-1.7.0/build/build/lib/librte_eal/linuxapp/kni/e1000_82575.c:38:
> >/home/qxie/dpdk-1.7.0/lib/librte_eal/linuxapp/kni/ethtool/igb/kcompat.h:3853:1:
> >error: conflicting types for ‘skb_set_hash’
> >/usr/src/linux-headers-3.13.0-32-generic/include/linux/skbuff.h:740:1:
> >note: previous definition of ‘skb_set_hash’ was here
> >make[8]: *** [/home/qxie/dpdk-1.7.0/build/build/lib/librte_eal/linuxapp/kni/e1000_82575.o]
> >Error 1
> >============
> >
> >Apparently the version check around 'skb_set_hash' in kcompat.h failed for
> >some reason. To work around it, I temporarily commented out that
> >'skb_set_hash' definition code and rte_kni complied fine. But it failed
> >again in librte_pmd_ixgbe:
> >
> >=============
> >In file included from
> >/home/qxie/dpdk-1.7.0/lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c:41:0:
> >/usr/lib/gcc/x86_64-linux-gnu/4.6/include/nmmintrin.h:31:3: error: #error
> >"SSE4.2 instruction set not enabled"
> From the error, nmmintrin.h contains built-in CPU calls that depend on the
> CPU having SSE4.2 extensions. You need to hunt down the call that is being
> executed and see which file is including nmmintrin.h. For example, I have
> encountered similar error on a CPU that does not have SSE4.1 and 4.2
> extension and it was trying to call the built-in HW crc hashing call.
> instead I switched with rte_jhash.h, the alternative SW implementation of
> the hashing call.
> 
> Looking below, nmmintrin.h is being included by the
> lib/librte_pmd_ixgbe/ixgbe_rxtx_vec.c, from the looks of it. This means that
> your defconfig (did you use the default one?) has
> CONFIG_RTE_IXGBE_INC_VECTOR value set to "y". try setting it to "n" and see
> where the build would stop next looking for another SSE4.x based builtin
> call.
> 
This is fixed at the head of the dpdk git tree.  The vector rx code has been
converted to use builtin intrinsics that can work without sse support in the
cpu.  The ACL library has a simmilar problem, but we are working at enabling it
for all cpus, with runtime updates to faster paths when the executing cpu
supports it.

> My ultimate recommendation is to update you CPU to a newer one or perform
> the build/run on a separate newer CPU.
> 
Building on a newer cpu won't help if you intend to run it on an older cpu.

Neil