Hi all,
I need to install DPDK on 64-bit Linux machines whose processors might or
might not support AVX instructions. In my current configuration file,
defconfig_x86-64-native-linuxapp-gcc, the parameter CONFIG_RTE_MACHINE is
set to native. If I compile DPDK on a machine that supports AVX and then I
install it (along with openvswitch) on a machine that does *not* support
AVX, openvswitch won't start and will output an "illegal instruction"
error.
I'm proposing the following patch to fix this issue:
diff -r ac0f646132b8 src/lib/librte_net/rte_net_crc.c
--- a/src/lib/librte_net/rte_net_crc.c Thu Apr 26 17:25:01 2018 +0200
+++ b/src/lib/librte_net/rte_net_crc.c Thu Apr 26 17:25:04 2018 +0200
@@ -212,8 +212,10 @@
rte_net_crc_scalar_init();
#ifdef X86_64_SSE42_PCLMULQDQ
- alg = RTE_NET_CRC_SSE42;
- rte_net_crc_sse42_init();
+ if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX)) {
+ alg = RTE_NET_CRC_SSE42;
+ rte_net_crc_sse42_init();
+ }
#elif defined ARM64_NEON_PMULL
if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_PMULL)) {
alg = RTE_NET_CRC_NEON;
The original check above is for SSE4, not for AVX, but the chosen function
rte_net_crc_sse42_init does run AVX instructions. So I suggest making the
check for AVX explicit, as outlined above.
I tested this on a machine with an Intel Atom C3000 processor and
openvswitch seems to function correctly.
Are there any potential side effects?
Thanks!
Riccardo
