enable lpm, acl and other missing libraries in ppc64le
v6 cover enable lpm, acl and other missing libraries in ppc64le
This patchset enables LPM, ACL and other few missing libs in ppc64le and also address few patches in related examples (ip_pipeline and l3fwd).
Test report:
1. LPM and ACL unit tests passed.
Steps:
compile test app and run (with any needed params) lpm_autotest
acl_autotest
table_autotest
test_table_lpm_combined fails same as in intel in current master.
2. Example ip_pipeline application verified for port forwarding. compile examples/ip_pipeline (adjust app_init_core_map param for ht to 1 in case of ppc64le in SMT=off mode). modify config/l3fwd.cfg as per enabled PMD ports. run ip_pipeline with config file option and check packets fwd.
v6 changes:
- added cache alignment fix for rte hash table structs.
v5 changes:
- no change in lpm lib enablement
- no change in acl lib enablement
- config file changes individually for sched,table,port,pipeline lib enablement
- ip_pipeline patch description and changes flagged only for ppc64le. app_init_core_map changes removed (due to bug found and under investigation only on ppc64le/smt=off case).
v4 changes:
- fix transition4 in acl_run_altivec.h for gcc strict-aliasing error. Thanks to Chao Zhu for bringing up.
v3 changes:
- rebase over master to fix conflict in examples/l3fwd/l3fwd_em.c
v2 changes:
- enabling libs in config included as part of lib changes itself.
Gowrishankar Muthukrishnan (9):
lpm: add altivec intrinsics for dpdk lpm on ppc_64
acl: add altivec intrinsics for dpdk acl on ppc_64
l3fwd: add altivec support for em_hash_key
table: enable table library for ppc64le
sched: enable sched library for ppc64le
port: enable port library for ppc64le
pipeline: enable pipeline library for ppc64le
ip_pipeline: fix lcore mapping for varying SMT threads as in ppc64
table: align rte table hash structs for cache line size
app/test-acl/main.c | 4 + app/test/test_xmmt_ops.h | 16 + config/defconfig_ppc_64-power8-linuxapp-gcc | 7 - examples/ip_pipeline/cpu_core_map.c | 8 + examples/l3fwd/l3fwd_em.c | 10 +- lib/librte_acl/Makefile | 2 + lib/librte_acl/acl.h | 4 + lib/librte_acl/acl_run.h | 2 + lib/librte_acl/acl_run_altivec.c | 47 +++ lib/librte_acl/acl_run_altivec.h | 329 +++++++++++++++++++++ lib/librte_acl/rte_acl.c | 13 + lib/librte_acl/rte_acl.h | 1 + .../common/include/arch/ppc_64/rte_vect.h | 60 ++++ lib/librte_lpm/Makefile | 2 + lib/librte_lpm/rte_lpm.h | 2 + lib/librte_lpm/rte_lpm_altivec.h | 154 ++++++++++ lib/librte_table/rte_table_hash_key16.c | 4 +- lib/librte_table/rte_table_hash_key32.c | 4 +- lib/librte_table/rte_table_hash_key8.c | 2 +- 19 files changed, 658 insertions(+), 13 deletions(-) create mode 100644 lib/librte_acl/acl_run_altivec.c create mode 100644 lib/librte_acl/acl_run_altivec.h create mode 100644 lib/librte_eal/common/include/arch/ppc_64/rte_vect.h create mode 100644 lib/librte_lpm/rte_lpm_altivec.h
Signature · 1 line
1.9.1
PATCH v6 1/9 lpm: add altivec intrinsics for dpdk lpm on ppc_64
This patch adds ppc64le port for LPM library in DPDK.
--- app/test/test_xmmt_ops.h | 16 +++ config/defconfig_ppc_64-power8-linuxapp-gcc | 1 - .../common/include/arch/ppc_64/rte_vect.h | 60 ++++++++ lib/librte_lpm/Makefile | 2 + lib/librte_lpm/rte_lpm.h | 2 + lib/librte_lpm/rte_lpm_altivec.h | 154 +++++++++++++++++++++ 6 files changed, 234 insertions(+), 1 deletion(-) create mode 100644 lib/librte_eal/common/include/arch/ppc_64/rte_vect.h create mode 100644 lib/librte_lpm/rte_lpm_altivec.h
app/test/test_xmmt_ops.h +16 -0
@@ -62,6 +62,22 @@ vect_set_epi32(int i3, int i2, int i1, int i0)/* sets the 4 signed 32-bit integer values and returns the xmm_t variable */ #define vect_set_epi32 (i3, i2, i1, i0) _mm_set_epi32(i3, i2, i1, i0) +#elif defined(RTE_ARCH_PPC_64 ) + +/* vect_* abstraction implementation using ALTIVEC */ + +/* loads the xmm_t value from address p(does not need to be 16-byte aligned)*/ +#define vect_loadu_sil128 (p) vec_ld(0, p) + +/* sets the 4 signed 32-bit integer values and returns the xmm_t variable */ +static inline xmm_t __attribute__((always_inline)) +vect_set_epi32 (int i3,int i2,int i1,int i0) +{ +xmm_t data = (xmm_t ){i0, i1, i2, i3}; + +return data; +} +#endif #endif /* _TEST_XMMT_OPS_H_ */
config/defconfig_ppc_64-power8-linuxapp-gcc +0 -1
@@ -57,7 +57,6 @@ CONFIG_RTE_LIBRTE_ENIC_PMD=n CONFIG_RTE_LIBRTE_FM10K_PMD=n # This following libraries are not available on Power. So they're turned off. -CONFIG_RTE_LIBRTE_LPM=n CONFIG_RTE_LIBRTE_ACL=n CONFIG_RTE_LIBRTE_SCHED=n CONFIG_RTE_LIBRTE_PORT=n
lib/librte_eal/common/include/arch/ppc_64/rte_vect.h new file +60 -0
60 lines not on this page · raw message
lib/librte_lpm/Makefile +2 -0
@@ -51,6 +51,8 @@ ifneq ($(filter y,$(CONFIG_RTE_ARCH_ARM) $(CONFIG_RTE_ARCH_ARM64)),) SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_neon.h else ifeq ($(CONFIG_RTE_ARCH_X86),y) SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_sse.h +else ifeq ($(CONFIG_RTE_ARCH_PPC_64),y) +SYMLINK-$(CONFIG_RTE_LIBRTE_LPM)-include += rte_lpm_altivec.h endif # this lib needs eal
lib/librte_lpm/rte_lpm.h +2 -0
@@ -480,6 +480,8 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, xmm_t ip, uint32_t hop[4],#if defined(RTE_ARCH_ARM ) || defined(RTE_ARCH_ARM64 )#include "rte_lpm_neon.h" +#elif defined(RTE_ARCH_PPC_64 ) +#include "rte_lpm_altivec.h" #else #include "rte_lpm_sse.h" #endif
lib/librte_lpm/rte_lpm_altivec.h new file +154 -0
154 lines not on this page · raw message
Signature · 1 line
1.9.1
Re: [PATCH v6 1/9] lpm: add altivec intrinsics for dpdk lpm on ppc_64
Quoted 14 lines · Gowrishankar Muthukrishnan
This patch adds ppc64le port for LPM library in DPDK.
Signed-off-by Gowrishankar Muthukrishnan <[email protected]>--- app/test/test_xmmt_ops.h | 16 +++ config/defconfig_ppc_64-power8-linuxapp-gcc | 1 - .../common/include/arch/ppc_64/rte_vect.h | 60 ++++++++ lib/librte_lpm/Makefile | 2 + lib/librte_lpm/rte_lpm.h | 2 + lib/librte_lpm/rte_lpm_altivec.h | 154 +++++++++++++++++++++ 6 files changed, 234 insertions(+), 1 deletion(-) create mode 100644 lib/librte_eal/common/include/arch/ppc_64/rte_vect.h create mode 100644 lib/librte_lpm/rte_lpm_altivec.h
Compilation and unit tests all still work fine on x86_64 systems, and patch looks ok to me, so
PATCH v6 2/9 acl: add altivec intrinsics for dpdk acl on ppc_64
In reply to Gowrishankar MuthukrishnanThis patch adds port for ACL library in ppc64le.
--- app/test-acl/main.c | 4 + config/defconfig_ppc_64-power8-linuxapp-gcc | 1 - lib/librte_acl/Makefile | 2 + lib/librte_acl/acl.h | 4 + lib/librte_acl/acl_run.h | 2 + lib/librte_acl/acl_run_altivec.c | 47 ++++ lib/librte_acl/acl_run_altivec.h | 329 ++++++++++++++++++++++++++++ lib/librte_acl/rte_acl.c | 13 ++ lib/librte_acl/rte_acl.h | 1 + 9 files changed, 402 insertions(+), 1 deletion(-) create mode 100644 lib/librte_acl/acl_run_altivec.c create mode 100644 lib/librte_acl/acl_run_altivec.h
app/test-acl/main.c +4 -0
@@ -105,6 +105,10 @@ static const struct acl_alg acl_alg[] = { .name = "neon", .alg = RTE_ACL_CLASSIFY_NEON, }, + { + .name = "altivec", + .alg = RTE_ACL_CLASSIFY_ALTIVEC, + }, }; static struct {
config/defconfig_ppc_64-power8-linuxapp-gcc +0 -1
@@ -57,7 +57,6 @@ CONFIG_RTE_LIBRTE_ENIC_PMD=n CONFIG_RTE_LIBRTE_FM10K_PMD=n # This following libraries are not available on Power. So they're turned off. -CONFIG_RTE_LIBRTE_ACL=n CONFIG_RTE_LIBRTE_SCHED=n CONFIG_RTE_LIBRTE_PORT=n CONFIG_RTE_LIBRTE_TABLE=n
lib/librte_acl/Makefile +2 -0
@@ -52,6 +52,8 @@ SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_scalar.c ifneq ($(filter y,$(CONFIG_RTE_ARCH_ARM) $(CONFIG_RTE_ARCH_ARM64)),) SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_neon.c CFLAGS_acl_run_neon.o += -flax-vector-conversions -Wno-maybe-uninitialized +else ifeq ($(CONFIG_RTE_ARCH_PPC_64),y) +SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_altivec.c else SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_sse.c #check if flag for SSE4.1 is already on, if not set it up manually
lib/librte_acl/acl.h +4 -0
@@ -234,6 +234,10 @@ int rte_acl_classify_neon(const struct rte_acl_ctx *ctx, const uint8_t **data, uint32_t *results, uint32_t num, uint32_t categories); +int +rte_acl_classify_altivec (const struct rte_acl_ctx *ctx,const uint8_t **data, +uint32_t *results,uint32_t num,uint32_t categories); + #ifdef __cplusplus }#endif /* __cplusplus */
lib/librte_acl/acl_run.h +2 -0
@@ -39,7 +39,9 @@#define MAX_SEARCHES_AVX16 16#define MAX_SEARCHES_SSE8 8 +#define MAX_SEARCHES_ALTIVEC8 8#define MAX_SEARCHES_SSE4 4 +#define MAX_SEARCHES_ALTIVEC4 4#define MAX_SEARCHES_SCALAR 2#define GET_NEXT_4BYTES (prm, idx) \
lib/librte_acl/acl_run_altivec.c new file +47 -0
47 lines not on this page · raw message
lib/librte_acl/acl_run_altivec.h new file +329 -0
329 lines not on this page · raw message
lib/librte_acl/rte_acl.c +13 -0
@@ -75,12 +75,23 @@ rte_acl_classify_neon(__rte_unused const struct rte_acl_ctx *ctx,return -ENOTSUP ; } +int __attribute__ ((weak)) +rte_acl_classify_altivec(__rte_unused const struct rte_acl_ctx *ctx, + __rte_unused const uint8_t **data, + __rte_unused uint32_t *results, + __rte_unused uint32_t num, + __rte_unused uint32_t categories) +{ +return -ENOTSUP ; +} +static const rte_acl_classify_t classify_fns[] = { [RTE_ACL_CLASSIFY_DEFAULT ] = rte_acl_classify_scalar, [RTE_ACL_CLASSIFY_SCALAR ] = rte_acl_classify_scalar, [RTE_ACL_CLASSIFY_SSE ] = rte_acl_classify_sse, [RTE_ACL_CLASSIFY_AVX2 ] = rte_acl_classify_avx2, [RTE_ACL_CLASSIFY_NEON ] = rte_acl_classify_neon, + [RTE_ACL_CLASSIFY_ALTIVEC ] = rte_acl_classify_altivec, };/* by default, use always available scalar code path. */
@@ -119,6 +130,8 @@ rte_acl_init(void)#elif defined(RTE_ARCH_ARM )if (rte_cpu_get_flag_enabled (RTE_CPUFLAG_NEON )) alg =RTE_ACL_CLASSIFY_NEON ; +#elif defined(RTE_ARCH_PPC_64 ) + alg =RTE_ACL_CLASSIFY_ALTIVEC ;#else #ifdef CC_AVX2_SUPPORT if (rte_cpu_get_flag_enabled (RTE_CPUFLAG_AVX2 ))
lib/librte_acl/rte_acl.h +1 -0
@@ -271,6 +271,7 @@ enum rte_acl_classify_alg {RTE_ACL_CLASSIFY_SSE =2 ,/**< requires SSE4.1 support. */ RTE_ACL_CLASSIFY_AVX2 =3 ,/**< requires AVX2 support. */ RTE_ACL_CLASSIFY_NEON =4 ,/**< requires NEON support. */ +RTE_ACL_CLASSIFY_ALTIVEC =5 ,/**< requires ALTIVEC support. */ RTE_ACL_CLASSIFY_NUM /* should always be the last one. */ };
Signature · 1 line
1.9.1
Re: [PATCH v6 2/9] acl: add altivec intrinsics for dpdk acl on ppc_64
Quoted 15 lines
This patch adds port for ACL library in ppc64le.
Signed-off-by Gowrishankar Muthukrishnan <[email protected]>--- app/test-acl/main.c | 4 + config/defconfig_ppc_64-power8-linuxapp-gcc | 1 - lib/librte_acl/Makefile | 2 + lib/librte_acl/acl.h | 4 + lib/librte_acl/acl_run.h | 2 + lib/librte_acl/acl_run_altivec.c | 47 ++++ lib/librte_acl/acl_run_altivec.h | 329 ++++++++++++++++++++++++++++ lib/librte_acl/rte_acl.c | 13 ++ lib/librte_acl/rte_acl.h | 1 + 9 files changed, 402 insertions(+), 1 deletion(-) create mode 100644 lib/librte_acl/acl_run_altivec.c create mode 100644
Quoted 2 lines
lib/librte_acl/acl_run_altivec.h
I am not a ppc expert, and I don't have ppc HW to try, But from IA perspective - all looks good.
PATCH v6 3/9 l3fwd: add altivec support for em_hash_key
In reply to Gowrishankar MuthukrishnanThis patch adds ppc64le port for em_mask_key function.
--- examples/l3fwd/l3fwd_em.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-)
examples/l3fwd/l3fwd_em.c +9 -1
@@ -259,8 +259,16 @@ em_mask_key(void *key, xmm_t mask)return vandq_s32 (data, mask); } +#elif defined(RTE_MACHINE_CPUFLAG_ALTIVEC ) +static inline xmm_t +em_mask_key (void *key,xmm_t mask) +{ +xmm_t data =vec_ld (0 , (xmm_t *)(key)); + +return vec_and (data, mask); +}#else -#error No vector engine (SSE, NEON) available, check your toolchain +#error No vector engine (SSE, NEON, ALTIVEC) available, check your toolchain#endif static inline uint8_t
Signature · 1 line
1.9.1
PATCH v6 4/9 table: enable table library for ppc64le
In reply to Gowrishankar MuthukrishnanThis patch enables librte_table in ppc64le.
--- config/defconfig_ppc_64-power8-linuxapp-gcc | 1 - 1 file changed, 1 deletion(-)
config/defconfig_ppc_64-power8-linuxapp-gcc +0 -1
@@ -59,5 +59,4 @@ CONFIG_RTE_LIBRTE_FM10K_PMD=n # This following libraries are not available on Power. So they're turned off. CONFIG_RTE_LIBRTE_SCHED=n CONFIG_RTE_LIBRTE_PORT=n -CONFIG_RTE_LIBRTE_TABLE=n CONFIG_RTE_LIBRTE_PIPELINE=n
Signature · 1 line
1.9.1
PATCH v6 5/9 sched: enable sched library for ppc64le
In reply to Gowrishankar MuthukrishnanThis patch enables librte_sched in ppc64le.
--- config/defconfig_ppc_64-power8-linuxapp-gcc | 1 - 1 file changed, 1 deletion(-)
config/defconfig_ppc_64-power8-linuxapp-gcc +0 -1
@@ -57,6 +57,5 @@ CONFIG_RTE_LIBRTE_ENIC_PMD=n CONFIG_RTE_LIBRTE_FM10K_PMD=n # This following libraries are not available on Power. So they're turned off. -CONFIG_RTE_LIBRTE_SCHED=n CONFIG_RTE_LIBRTE_PORT=n CONFIG_RTE_LIBRTE_PIPELINE=n
Signature · 1 line
1.9.1
PATCH v6 6/9 port: enable port library for ppc64le
In reply to Gowrishankar MuthukrishnanThis patch enables librte_port in ppc64le.
--- config/defconfig_ppc_64-power8-linuxapp-gcc | 1 - 1 file changed, 1 deletion(-)
config/defconfig_ppc_64-power8-linuxapp-gcc +0 -1
@@ -57,5 +57,4 @@ CONFIG_RTE_LIBRTE_ENIC_PMD=n CONFIG_RTE_LIBRTE_FM10K_PMD=n # This following libraries are not available on Power. So they're turned off. -CONFIG_RTE_LIBRTE_PORT=n CONFIG_RTE_LIBRTE_PIPELINE=n
Signature · 1 line
1.9.1
PATCH v6 7/9 pipeline: enable pipeline library for ppc64le
In reply to Gowrishankar MuthukrishnanThis patch enables librte_pipeline for ppc64le.
--- config/defconfig_ppc_64-power8-linuxapp-gcc | 2 -- 1 file changed, 2 deletions(-)
config/defconfig_ppc_64-power8-linuxapp-gcc +0 -2
@@ -56,5 +56,3 @@ CONFIG_RTE_LIBRTE_PMD_BOND=n CONFIG_RTE_LIBRTE_ENIC_PMD=n CONFIG_RTE_LIBRTE_FM10K_PMD=n -# This following libraries are not available on Power. So they're turned off. -CONFIG_RTE_LIBRTE_PIPELINE=n
Signature · 1 line
1.9.1
PATCH v6 8/9 ip_pipeline: fix lcore mapping for varying SMT threads as in ppc64
In reply to Gowrishankar MuthukrishnanThis patch fixes ip_pipeline panic in app_init_core_map while preparing cpu core map in powerpc with SMT off. cpu_core_map_compute_linux currently prepares core mapping based on file existence in sysfs ie.
/sys/devices/system/cpu/cpu<LCORE_NUM>/topology/physical_package_id
/sys/devices/system/cpu/cpu<LCORE_NUM>/topology/core_id
These files do not exist for lcores which are offline for any reason (as in powerpc, while SMT is off). In this situation, this function should further continue preparing map for other online lcores instead of returning with -1 for a first unavailable lcore.
Also, in SMT=off scenario for powerpc, lcore ids can not be always indexed from 0 upto 'number of cores present' (/sys/devices/system/cpu/present). For eg, for an online lcore 32, core_id returned in sysfs is 112 where online lcores are 10 (as in one configuration), hence sysfs lcore id can not be checked with indexing lcore number before positioning lcore map array.
--- examples/ip_pipeline/cpu_core_map.c | 8 ++++++++ 1 file changed, 8 insertions(+)
examples/ip_pipeline/cpu_core_map.c +8 -0
@@ -351,8 +351,10 @@ cpu_core_map_compute_linux(struct cpu_core_map *map)int lcore_socket_id =cpu_core_map_get_socket_id_linux (lcore_id); +#if !defined(RTE_ARCH_PPC_64 )if (lcore_socket_id <0 )return -1 ; +#endif if (((uint32_t ) lcore_socket_id) == socket_id) n_detected++;
@@ -368,6 +370,7 @@ cpu_core_map_compute_linux(struct cpu_core_map *map)cpu_core_map_get_socket_id_linux ( lcore_id); +#if !defined(RTE_ARCH_PPC_64 )if (lcore_socket_id <0 )return -1 ;
@@ -377,9 +380,14 @@ cpu_core_map_compute_linux(struct cpu_core_map *map)if (lcore_core_id <0 )return -1 ; +#endif +#if !defined(RTE_ARCH_PPC_64 )if (((uint32_t ) lcore_socket_id == socket_id) && ((uint32_t ) lcore_core_id == core_id)) { +#else +if (((uint32_t ) lcore_socket_id == socket_id)) { +#endif uint32_t pos = cpu_core_map_pos(map, socket_id, core_id_contig,
Signature · 1 line
1.9.1
Re: [PATCH v6 8/9] ip_pipeline: fix lcore mapping for varying SMT threads as in ppc64
Quoted 90 lines
-----Original Message-----
From: Gowrishankar Muthukrishnan [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 11:28 AM
To: [email protected]
Cc: Chao Zhu <[email protected]>; Richardson, Bruce <[email protected]>; Ananyev, Konstantin <[email protected]>; Thomas Monjalon <[email protected]>; Dumitrescu, Cristian <[email protected]>; Pradeep <[email protected]>
Subject: [PATCH v6 8/9]ip_pipeline: fix lcore mapping for varying SMT threads as in ppc64This patch fixes
ip_pipelinepanic inapp_init_core_mapwhile preparing cpu core map in powerpc with SMT off.cpu_core_map_compute_linuxcurrently prepares
core mapping based on file existence in sysfs ie./sys/devices/system/cpu/cpu<LCORE_NUM>/topology/physical_package_id
/sys/devices/system/cpu/cpu<LCORE_NUM>/topology/core_idThese files do not exist for lcores which are offline for any reason (as in powerpc, while SMT is off). In this situation, this function should further continue preparing map for other online lcores instead of returning with -1 for a first unavailable lcore.
Also, in SMT=off scenario for powerpc, lcore ids can not be always indexed from
0 upto 'number of cores present' (/sys/devices/system/cpu/present). For eg, for
an online lcore 32,core_idreturned in sysfs is 112 where online lcores are 10 (as in one configuration), hence sysfs lcore id can not be checked with indexing lcore number before positioning lcore map array.Signed-off-by: Gowrishankar Muthukrishnan
<[email protected]>--- examples/ip_pipeline/cpu_core_map.c | 8 ++++++++ 1 file changed, 8 insertions(+)examples/ip_pipeline/cpu_core_map.c +2
-0@@ -351,8 +351,10 @@ cpu_core_map_compute_linux(struct cpu_core_map *map)int lcore_socket_id =cpu_core_map_get_socket_id_linux (lcore_id); +#if !defined(RTE_ARCH_PPC_64 )if (lcore_socket_id <0 )return -1 ; +#endif examples/ip_pipeline/cpu_core_map.c +3
-0if (((uint32_t ) lcore_socket_id) == socket_id) n_detected++;@@ -368,6 +370,7 @@ cpu_core_map_compute_linux(struct cpu_core_map *map)cpu_core_map_get_socket_id_linux ( lcore_id); +#if !defined(RTE_ARCH_PPC_64 )if (lcore_socket_id <0 )return -1 ;@@ -377,9 +380,14 @@ cpu_core_map_compute_linux(struct cpu_core_map *map)if (lcore_core_id <0 )return -1 ; +#endif +#if !defined(RTE_ARCH_PPC_64 ) if (((uint32_t) lcore_socket_id == socket_id)&&
((uint32_t)lcore_core_id==core_id)) {examples/ip_pipeline/cpu_core_map.c +2
-0+#else +if (((uint32_t ) lcore_socket_id == socket_id)){
examples/ip_pipeline/cpu_core_map.c +1
-0+#endif uint32_t pos =cpu_core_map_pos(map,
examples/ip_pipeline/cpu_core_map.c +0
-1socket_id, core_id_contig, --
Quoted 1 line
1.9.1
This patch only changes the code for PPC CPUs, I don't have the hardware to check it myself, but I will take Gowrishankar's and Chao's word it is the right thing to do for PPC CPUs, so ...
Acked by: Cristian Dumitrescu <[email protected]>
PATCH v6 9/9 table: align rte table hash structs for cache line size
In reply to Gowrishankar Muthukrishnanrte table hash structs rte_bucket_4_8, rte_bucket_4_16 and rte_bucket_4_32 have to be cache aligned as required by their corresponding hash create functions rte_table_hash_create_key8_lru etc.
--- lib/librte_table/rte_table_hash_key16.c | 4 ++-- lib/librte_table/rte_table_hash_key32.c | 4 ++-- lib/librte_table/rte_table_hash_key8.c | 2 +- 3 files changed, 5 insertions(+), 5 deletions(-)
lib/librte_table/rte_table_hash_key16.c +2 -2
@@ -68,10 +68,10 @@ struct rte_bucket_4_16 {uint64_t next_valid;/* Cache line 1 */ -uint64_t key[4 ][2 ]; + uint64_t key[4][2] __rte_cache_aligned;/* Cache line 2 */ -uint8_t data[0 ]; + uint8_t data[0] __rte_cache_aligned; };struct rte_table_hash {
lib/librte_table/rte_table_hash_key32.c +2 -2
@@ -68,10 +68,10 @@ struct rte_bucket_4_32 {uint64_t next_valid;/* Cache lines 1 and 2 */ -uint64_t key[4 ][4 ]; + uint64_t key[4][4] __rte_cache_aligned;/* Cache line 3 */ -uint8_t data[0 ]; + uint8_t data[0] __rte_cache_aligned; };struct rte_table_hash {
lib/librte_table/rte_table_hash_key8.c +1 -1
@@ -68,7 +68,7 @@ struct rte_bucket_4_8 {uint64_t key[4 ];/* Cache line 1 */ -uint8_t data[0 ]; + uint8_t data[0] __rte_cache_aligned; };struct rte_table_hash {
Signature · 1 line
1.9.1
Re: [PATCH v6 9/9] table: align rte table hash structs for cache line size
Quoted 78 lines
-----Original Message-----
From: Gowrishankar Muthukrishnan [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 11:28 AM
To: [email protected]
Cc: Chao Zhu <[email protected]>; Richardson, Bruce <[email protected]>; Ananyev, Konstantin <[email protected]>; Thomas Monjalon <[email protected]>; Dumitrescu, Cristian <[email protected]>; Pradeep <[email protected]>
Subject: [PATCH v6 9/9] table: align rte table hash structs for cache line sizerte table hash structs
rte_bucket_4_8,rte_bucket_4_16andrte_bucket_4_32have
to be cache aligned as required by their corresponding hash create functionsrte_table_hash_create_key8_lruetc.Signed-off-by: Gowrishankar Muthukrishnan
<[email protected]>--- lib/librte_table/rte_table_hash_key16.c | 4 ++-- lib/librte_table/rte_table_hash_key32.c | 4 ++-- lib/librte_table/rte_table_hash_key8.c | 2 +- 3 files changed, 5 insertions(+), 5 deletions(-)lib/librte_table/rte_table_hash_key16.c +2
-2@@ -68,10 +68,10 @@ struct rte_bucket_4_16 {uint64_t next_valid;/* Cache line 1 */ -uint64_t key[4 ][2 ]; + uint64_t key[4][2] __rte_cache_aligned;/* Cache line 2 */ -uint8_t data[0 ]; + uint8_t data[0] __rte_cache_aligned; };struct rte_table_hash {lib/librte_table/rte_table_hash_key32.c +2
-2@@ -68,10 +68,10 @@ struct rte_bucket_4_32 {uint64_t next_valid;/* Cache lines 1 and 2 */ -uint64_t key[4 ][4 ]; + uint64_t key[4][4] __rte_cache_aligned;/* Cache line 3 */ -uint8_t data[0 ]; + uint8_t data[0] __rte_cache_aligned; };struct rte_table_hash {lib/librte_table/rte_table_hash_key8.c +1
-2@@ -68,7 +68,7 @@ struct rte_bucket_4_8 {uint64_t key[4 ];/* Cache line 1 */ -uint8_t data[0 ]; + uint8_t data[0] __rte_cache_aligned; };struct rte_table_hash { --
Quoted 1 line
1.9.1
Hi Gowrishankar,
My understanding is you are trying to work around the check invoked by the hash table create functions that verifies the size of the bucket header structure is a multiple of the cache line, right?
Given that the size of this structure is 1x, 2x or 3x times 64 bytes, the check passes on IA CPUs (cache line of 64 bytes; explicit alignment to cache line size is not needed in order to make the size of the structure a multiple of cache line), but not on PPC CPUs (cache line of 128 bytes), correct?
I don't think your proposal provides the best way to fix this issue, since your code leads to a considerable increase in the memory consumption used per bucket in most cases:
- 100% more memory for 8-byte key hash table
- 0% more for 16-byte key hash table (code does not fix anything, explicit alignment is not needed)
- 50% more for 32-byte key hash table
I suggest you simply change the check: instead of validating this data structure is a multiple of cache line size, validate it is a multiple of 64 bytes.
Regards,
Cristian
Re: [PATCH v6 9/9] table: align rte table hash structs for cache line size
2016-08-31 17:29, Dumitrescu, Cristian:
Quoted 10 lines
From: Gowrishankar Muthukrishnan
Quoted 4 lines
rte table hash structs
rte_bucket_4_8,rte_bucket_4_16andrte_bucket_4_32have
to be cache aligned as required by their corresponding hash create functionsrte_table_hash_create_key8_lruetc.Hi Gowrishankar,
My understanding is you are trying to work around the check invoked by the hash table create functions that verifies the size of the bucket header structure is a multiple of the cache line, right?
Quoted 8 lines
Given that the size of this structure is 1x, 2x or 3x times 64 bytes, the check passes on IA CPUs (cache line of 64 bytes; explicit alignment to cache line size is not needed in order to make the size of the structure a multiple of cache line), but not on PPC CPUs (cache line of 128 bytes), correct?
I don't think your proposal provides the best way to fix this issue, since your code leads to a considerable increase in the memory consumption used per bucket in most cases:
- 100% more memory for 8-byte key hash table
- 0% more for 16-byte key hash table (code does not fix anything, explicit alignment is not needed)
- 50% more for 32-byte key hash tableI suggest you simply change the check: instead of validating this data structure is a multiple of cache line size, validate it is a multiple of 64 bytes.
Any news please?
The whole series is blocked for this patch.
Should we expect a v7?
Re: [PATCH v6 9/9] table: align rte table hash structs for cache line size
Quoted 38 lines
-----Original Message-----
From: Thomas Monjalon [mailto:[email protected]]
Sent: Thursday, September 8, 2016 10:36 AM
To: Gowrishankar Muthukrishnan <[email protected]>
Cc: [email protected]; Dumitrescu, Cristian <[email protected]>; Chao Zhu <[email protected]>; Richardson, Bruce <[email protected]>; Ananyev, Konstantin <[email protected]>; Pradeep <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v6 9/9] table: align rte table hash structs for cache line size2016-08-31 17:29, Dumitrescu, Cristian:
Quoted 4 lines
From: Gowrishankar Muthukrishnan
Quoted 3 lines
rte table hash structs
rte_bucket_4_8,rte_bucket_4_16andrte_bucket_4_32have
to be cache aligned as required by their corresponding hash createfunctions
Quoted 5 lines
Quoted 1 line
rte_table_hash_create_key8_lruetc.Hi Gowrishankar,
My understanding is you are trying to work around the check invoked by
the hash table create functions that verifies the size of the bucket header structure is a multiple of the cache line, right?
Quoted 2 lines
Given that the size of this structure is 1x, 2x or 3x times 64 bytes, the check
passes on IA CPUs (cache line of 64 bytes; explicit alignment to cache line size is not needed in order to make the size of the structure a multiple of cache line), but not on PPC CPUs (cache line of 128 bytes), correct?
Quoted 2 lines
I don't think your proposal provides the best way to fix this issue, since
your code leads to a considerable increase in the memory consumption used per bucket in most cases:
Quoted 2 lines
- 100% more memory for 8-byte key hash table
- 0% more for 16-byte key hash table (code does not fix anything,explicit alignment is not needed)
Quoted 8 lines
Quoted 3 lines
- 50% more for 32-byte key hash table
I suggest you simply change the check: instead of validating this data
structure is a multiple of cache line size, validate it is a multiple of 64 bytes.
Any news please?
The whole series is blocked for this patch.
Should we expect a v7?
Yes, I think we should. Small fix for a considerable benefit.
Re: [PATCH v6 9/9] table: align rte table hash structs for cache line size
Thanks Cristian and Thomas for your feedback. I have taken your suggestions and sent out v7. Please check if the new patch is fine.
Thanks,
Gowrishankar
Quoted 45 lines · Cristian
Quoted 34 lines
-----Original Message-----
From: Thomas Monjalon [mailto:[email protected]]
Sent: Thursday, September 8, 2016 10:36 AM
To: Gowrishankar Muthukrishnan <[email protected]>
Cc: [email protected]; Dumitrescu, Cristian <[email protected]>; Chao Zhu <[email protected]>; Richardson, Bruce <[email protected]>; Ananyev, Konstantin <[email protected]>; Pradeep <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v6 9/9] table: align rte table hash structs for cache line size2016-08-31 17:29, Dumitrescu, Cristian:
Quoted 4 lines
From: Gowrishankar Muthukrishnan
Quoted 3 lines
rte table hash structs
rte_bucket_4_8,rte_bucket_4_16andrte_bucket_4_32have
to be cache aligned as required by their corresponding hash createfunctions
Quoted 4 lines
Quoted 1 line
rte_table_hash_create_key8_lruetc.Hi Gowrishankar,
My understanding is you are trying to work around the check invoked by
the hash table create functions that verifies the size of the bucket header structure is a multiple of the cache line, right?
Quoted 1 line
Given that the size of this structure is 1x, 2x or 3x times 64 bytes, the check
passes on IA CPUs (cache line of 64 bytes; explicit alignment to cache line size is not needed in order to make the size of the structure a multiple of cache line), but not on PPC CPUs (cache line of 128 bytes), correct?
Quoted 1 line
I don't think your proposal provides the best way to fix this issue, since
your code leads to a considerable increase in the memory consumption used per bucket in most cases:
Quoted 2 lines
- 100% more memory for 8-byte key hash table
- 0% more for 16-byte key hash table (code does not fix anything,explicit alignment is not needed)
Quoted 8 lines
Quoted 3 lines
- 50% more for 32-byte key hash table
I suggest you simply change the check: instead of validating this data
structure is a multiple of cache line size, validate it is a multiple of 64 bytes.
Any news please?
The whole series is blocked for this patch.
Should we expect a v7?Yes, I think we should. Small fix for a considerable benefit.
-----Original Message-----
From: Gowrishankar Muthukrishnan [mailto:[email protected]]
Sent: 2016年8月16日 18:28
To: [email protected]
Cc: Chao Zhu <[email protected]>; Bruce Richardson <[email protected]>; Konstantin Ananyev <[email protected]>; Thomas Monjalon <[email protected]>; Cristian Dumitrescu <[email protected]>; Pradeep <[email protected]>
Subject: [PATCH v6 0/9] enable lpm, acl and other missing libraries in ppc64le
This patchset enables LPM, ACL and other few missing libs in ppc64le and also address few patches in related examples (ip_pipeline and l3fwd).
Test report:
1. LPM and ACL unit tests passed.
Steps:
compile test app and run (with any needed params) lpm_autotest
acl_autotest
table_autotest
test_table_lpm_combined fails same as in intel in current master.
2. Example ip_pipeline application verified for port forwarding. compile examples/ip_pipeline (adjust app_init_core_map param for ht to 1 in case of ppc64le in SMT=off mode). modify config/l3fwd.cfg as per enabled PMD ports. run ip_pipeline with config file option and check packets fwd.
v6 changes:
- added cache alignment fix for rte hash table structs.
v5 changes:
- no change in lpm lib enablement
- no change in acl lib enablement
- config file changes individually for sched,table,port,pipeline lib enablement
- ip_pipeline patch description and changes flagged only for ppc64le. app_init_core_map changes removed (due to bug found and under investigation only on ppc64le/smt=off case).
v4 changes:
- fix transition4 in acl_run_altivec.h for gcc strict-aliasing error. Thanks to Chao Zhu for bringing up.
v3 changes:
- rebase over master to fix conflict in examples/l3fwd/l3fwd_em.c
v2 changes:
- enabling libs in config included as part of lib changes itself.
Gowrishankar Muthukrishnan (9):
lpm: add altivec intrinsics for dpdk lpm on ppc_64
acl: add altivec intrinsics for dpdk acl on ppc_64
l3fwd: add altivec support for em_hash_key
table: enable table library for ppc64le
sched: enable sched library for ppc64le
port: enable port library for ppc64le
pipeline: enable pipeline library for ppc64le
ip_pipeline: fix lcore mapping for varying SMT threads as in ppc64
table: align rte table hash structs for cache line size
app/test-acl/main.c | 4 + app/test/test_xmmt_ops.h | 16 + config/defconfig_ppc_64-power8-linuxapp-gcc | 7 - examples/ip_pipeline/cpu_core_map.c | 8 + examples/l3fwd/l3fwd_em.c | 10 +- lib/librte_acl/Makefile | 2 + lib/librte_acl/acl.h | 4 + lib/librte_acl/acl_run.h | 2 + lib/librte_acl/acl_run_altivec.c | 47 +++ lib/librte_acl/acl_run_altivec.h | 329 +++++++++++++++++++++ lib/librte_acl/rte_acl.c | 13 + lib/librte_acl/rte_acl.h | 1 + .../common/include/arch/ppc_64/rte_vect.h | 60 ++++ lib/librte_lpm/Makefile | 2 + lib/librte_lpm/rte_lpm.h | 2 + lib/librte_lpm/rte_lpm_altivec.h | 154 ++++++++++ lib/librte_table/rte_table_hash_key16.c | 4 +- lib/librte_table/rte_table_hash_key32.c | 4 +- lib/librte_table/rte_table_hash_key8.c | 2 +- 19 files changed, 658 insertions(+), 13 deletions(-) create mode 100644 lib/librte_acl/acl_run_altivec.c create mode 100644 lib/librte_acl/acl_run_altivec.h create mode 100644 lib/librte_eal/common/include/arch/ppc_64/rte_vect.h create mode 100644 lib/librte_lpm/rte_lpm_altivec.h
--
1.9.1
I think this patch set is good enough.
v6 cover enable lpm, acl and other missing libraries in ppc64le
Thomas,
Any comments of this patch set? Are we waiting for more acks?
Thank you!
-----Original Message-----
From: dev [mailto:[email protected]] On Behalf Of Chao Zhu
Sent: 2016年8月17日 16:49
To: 'Gowrishankar Muthukrishnan' <[email protected]>; [email protected]
Cc: 'Bruce Richardson' <[email protected]>; 'Konstantin Ananyev' <[email protected]>; 'Thomas Monjalon' <[email protected]>; 'Cristian Dumitrescu' <[email protected]>; 'Pradeep' <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v6 0/9] enable lpm, acl and other missing libraries in ppc64le
-----Original Message-----
From: Gowrishankar Muthukrishnan [mailto:[email protected]]
Sent: 2016年8月16日 18:28
To: [email protected]
Cc: Chao Zhu <[email protected]>; Bruce Richardson <[email protected]>; Konstantin Ananyev <[email protected]>; Thomas Monjalon <[email protected]>; Cristian Dumitrescu <[email protected]>; Pradeep <[email protected]>
Subject: [PATCH v6 0/9] enable lpm, acl and other missing libraries in ppc64le
This patchset enables LPM, ACL and other few missing libs in ppc64le and also address few patches in related examples (ip_pipeline and l3fwd).
Test report:
1. LPM and ACL unit tests passed.
Steps:
compile test app and run (with any needed params) lpm_autotest
acl_autotest
table_autotest
test_table_lpm_combined fails same as in intel in current master.
2. Example ip_pipeline application verified for port forwarding. compile examples/ip_pipeline (adjust app_init_core_map param for ht to 1 in case of ppc64le in SMT=off mode). modify config/l3fwd.cfg as per enabled PMD ports. run ip_pipeline with config file option and check packets fwd.
v6 changes:
- added cache alignment fix for rte hash table structs.
v5 changes:
- no change in lpm lib enablement
- no change in acl lib enablement
- config file changes individually for sched,table,port,pipeline lib enablement
- ip_pipeline patch description and changes flagged only for ppc64le. app_init_core_map changes removed (due to bug found and under investigation only on ppc64le/smt=off case).
v4 changes:
- fix transition4 in acl_run_altivec.h for gcc strict-aliasing error. Thanks to Chao Zhu for bringing up.
v3 changes:
- rebase over master to fix conflict in examples/l3fwd/l3fwd_em.c
v2 changes:
- enabling libs in config included as part of lib changes itself.
Gowrishankar Muthukrishnan (9):
lpm: add altivec intrinsics for dpdk lpm on ppc_64
acl: add altivec intrinsics for dpdk acl on ppc_64
l3fwd: add altivec support for em_hash_key
table: enable table library for ppc64le
sched: enable sched library for ppc64le
port: enable port library for ppc64le
pipeline: enable pipeline library for ppc64le
ip_pipeline: fix lcore mapping for varying SMT threads as in ppc64
table: align rte table hash structs for cache line size
app/test-acl/main.c | 4 + app/test/test_xmmt_ops.h | 16 + config/defconfig_ppc_64-power8-linuxapp-gcc | 7 - examples/ip_pipeline/cpu_core_map.c | 8 + examples/l3fwd/l3fwd_em.c | 10 +- lib/librte_acl/Makefile | 2 + lib/librte_acl/acl.h | 4 + lib/librte_acl/acl_run.h | 2 + lib/librte_acl/acl_run_altivec.c | 47 +++ lib/librte_acl/acl_run_altivec.h | 329 +++++++++++++++++++++ lib/librte_acl/rte_acl.c | 13 + lib/librte_acl/rte_acl.h | 1 + .../common/include/arch/ppc_64/rte_vect.h | 60 ++++ lib/librte_lpm/Makefile | 2 + lib/librte_lpm/rte_lpm.h | 2 + lib/librte_lpm/rte_lpm_altivec.h | 154 ++++++++++ lib/librte_table/rte_table_hash_key16.c | 4 +- lib/librte_table/rte_table_hash_key32.c | 4 +- lib/librte_table/rte_table_hash_key8.c | 2 +- 19 files changed, 658 insertions(+), 13 deletions(-) create mode 100644 lib/librte_acl/acl_run_altivec.c create mode 100644 lib/librte_acl/acl_run_altivec.h create mode 100644 lib/librte_eal/common/include/arch/ppc_64/rte_vect.h create mode 100644 lib/librte_lpm/rte_lpm_altivec.h
--
1.9.1
I think this patch set is good enough.
2016-08-26 18:55, Chao Zhu:
Quoted 3 lines
Thomas,
Any comments of this patch set? Are we waiting for more acks?
Yes, especially for lpm and acl, a review from respective maintainers may be needed.
Konstantin? Bruce?
If no review is done in 1 week, it will be applied as-is.
Gowrishankar, using --in-reply when sending would help to check the history of the patchset. Please remind for next time.