From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ilya Maximets Subject: [PATCH v7 1/2] mem: balanced allocation of hugepages Date: Wed, 21 Jun 2017 13:08:30 +0300 Message-ID: <1498039711-26570-2-git-send-email-i.maximets@samsung.com> References: <1498032250-24924-1-git-send-email-i.maximets@samsung.com> <1498039711-26570-1-git-send-email-i.maximets@samsung.com> Cc: Heetae Ahn , Yuanhan Liu , Jianfeng Tan , Neil Horman , Yulong Pei , Bruce Richardson , Jerin Jacob , Ilya Maximets To: dev@dpdk.org, David Marchand , Sergio Gonzalez Monroy , Thomas Monjalon Return-path: Received: from mailout3.w1.samsung.com (mailout3.w1.samsung.com [210.118.77.13]) by dpdk.org (Postfix) with ESMTP id 043CA5A98 for ; Wed, 21 Jun 2017 12:08:44 +0200 (CEST) Received: from eucas1p2.samsung.com (unknown [182.198.249.207]) by mailout3.w1.samsung.com (Oracle Communications Messaging Server 7.0.5.31.0 64bit (built May 5 2014)) with ESMTP id <0ORW008586UJHV80@mailout3.w1.samsung.com> for dev@dpdk.org; Wed, 21 Jun 2017 11:08:43 +0100 (BST) In-reply-to: <1498039711-26570-1-git-send-email-i.maximets@samsung.com> List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org Sender: "dev" Currently EAL allocates hugepages one by one not paying attention from which NUMA node allocation was done. Such behaviour leads to allocation failure if number of available hugepages for application limited by cgroups or hugetlbfs and memory requested not only from the first socket. Example: # 90 x 1GB hugepages availavle in a system cgcreate -g hugetlb:/test # Limit to 32GB of hugepages cgset -r hugetlb.1GB.limit_in_bytes=34359738368 test # Request 4GB from each of 2 sockets cgexec -g hugetlb:test testpmd --socket-mem=4096,4096 ... EAL: SIGBUS: Cannot mmap more hugepages of size 1024 MB EAL: 32 not 90 hugepages of size 1024 MB allocated EAL: Not enough memory available on socket 1! Requested: 4096MB, available: 0MB PANIC in rte_eal_init(): Cannot init memory This happens beacause all allocated pages are on socket 0. Fix this issue by setting mempolicy MPOL_PREFERRED for each hugepage to one of requested nodes using following schema: 1) Allocate essential hugepages: 1.1) Allocate as many hugepages from numa N to only fit requested memory for this numa. 1.2) repeat 1.1 for all numa nodes. 2) Try to map all remaining free hugepages in a round-robin fashion. 3) Sort pages and choose the most suitable. In this case all essential memory will be allocated and all remaining pages will be fairly distributed between all requested nodes. New config option RTE_EAL_NUMA_AWARE_HUGEPAGES introduced and enabled by default for linuxapp on x86, ppc and thunderx. Enabling of this option adds libnuma as a dependency for EAL. Fixes: 77988fc08dc5 ("mem: fix allocating all free hugepages") Signed-off-by: Ilya Maximets --- config/common_base | 1 + config/common_linuxapp | 2 + config/defconfig_arm-armv7a-linuxapp-gcc | 3 + config/defconfig_arm64-armv8a-linuxapp-gcc | 3 + config/defconfig_arm64-thunderx-linuxapp-gcc | 3 + lib/librte_eal/linuxapp/eal/Makefile | 3 + lib/librte_eal/linuxapp/eal/eal_memory.c | 105 ++++++++++++++++++++++++++- mk/rte.app.mk | 3 + 8 files changed, 119 insertions(+), 4 deletions(-) diff --git a/config/common_base b/config/common_base index f6aafd1..660588a 100644 --- a/config/common_base +++ b/config/common_base @@ -103,6 +103,7 @@ CONFIG_RTE_EAL_ALWAYS_PANIC_ON_ERROR=n CONFIG_RTE_EAL_IGB_UIO=n CONFIG_RTE_EAL_VFIO=n CONFIG_RTE_MALLOC_DEBUG=n +CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES=n # # Recognize/ignore the AVX/AVX512 CPU flags for performance/power testing. diff --git a/config/common_linuxapp b/config/common_linuxapp index b3cf41b..050526f 100644 --- a/config/common_linuxapp +++ b/config/common_linuxapp @@ -35,6 +35,8 @@ CONFIG_RTE_EXEC_ENV="linuxapp" CONFIG_RTE_EXEC_ENV_LINUXAPP=y +CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES=y + CONFIG_RTE_EAL_IGB_UIO=y CONFIG_RTE_EAL_VFIO=y CONFIG_RTE_KNI_KMOD=y diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc index 19607eb..e06b1d4 100644 --- a/config/defconfig_arm-armv7a-linuxapp-gcc +++ b/config/defconfig_arm-armv7a-linuxapp-gcc @@ -47,6 +47,9 @@ CONFIG_RTE_ARCH_STRICT_ALIGN=y CONFIG_RTE_TOOLCHAIN="gcc" CONFIG_RTE_TOOLCHAIN_GCC=y +# NUMA is not supported on ARM +CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES=n + # ARM doesn't have support for vmware TSC map CONFIG_RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT=n diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc index 9f32766..2c67cdc 100644 --- a/config/defconfig_arm64-armv8a-linuxapp-gcc +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc @@ -47,6 +47,9 @@ CONFIG_RTE_TOOLCHAIN_GCC=y # to address minimum DMA alignment across all arm64 implementations. CONFIG_RTE_CACHE_LINE_SIZE=128 +# Most ARMv8 systems doesn't support NUMA +CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES=n + CONFIG_RTE_EAL_IGB_UIO=n CONFIG_RTE_LIBRTE_FM10K_PMD=n diff --git a/config/defconfig_arm64-thunderx-linuxapp-gcc b/config/defconfig_arm64-thunderx-linuxapp-gcc index f64da4c..3e79fa8 100644 --- a/config/defconfig_arm64-thunderx-linuxapp-gcc +++ b/config/defconfig_arm64-thunderx-linuxapp-gcc @@ -37,6 +37,9 @@ CONFIG_RTE_CACHE_LINE_SIZE=128 CONFIG_RTE_MAX_NUMA_NODES=2 CONFIG_RTE_MAX_LCORE=96 +# ThunderX supports NUMA +CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES=y + # # Compile PMD for octeontx sso event device # diff --git a/lib/librte_eal/linuxapp/eal/Makefile b/lib/librte_eal/linuxapp/eal/Makefile index 640afd0..8651e27 100644 --- a/lib/librte_eal/linuxapp/eal/Makefile +++ b/lib/librte_eal/linuxapp/eal/Makefile @@ -50,6 +50,9 @@ LDLIBS += -ldl LDLIBS += -lpthread LDLIBS += -lgcc_s LDLIBS += -lrt +ifeq ($(CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES),y) +LDLIBS += -lnuma +endif # specific to linuxapp exec-env SRCS-$(CONFIG_RTE_EXEC_ENV_LINUXAPP) := eal.c diff --git a/lib/librte_eal/linuxapp/eal/eal_memory.c b/lib/librte_eal/linuxapp/eal/eal_memory.c index e17c9cb..ceadca7 100644 --- a/lib/librte_eal/linuxapp/eal/eal_memory.c +++ b/lib/librte_eal/linuxapp/eal/eal_memory.c @@ -54,6 +54,9 @@ #include #include #include +#ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES +#include +#endif #include #include @@ -348,6 +351,21 @@ static int huge_wrap_sigsetjmp(void) return sigsetjmp(huge_jmpenv, 1); } +#ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES +#ifndef ULONG_SIZE +#define ULONG_SIZE sizeof(unsigned long) +#endif +#ifndef ULONG_BITS +#define ULONG_BITS (ULONG_SIZE * CHAR_BIT) +#endif +#ifndef DIV_ROUND_UP +#define DIV_ROUND_UP(n, d) (((n) + (d) - 1) / (d)) +#endif +#ifndef BITS_TO_LONGS +#define BITS_TO_LONGS(nr) DIV_ROUND_UP(nr, ULONG_SIZE) +#endif +#endif + /* * Mmap all hugepages of hugepage table: it first open a file in * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the @@ -356,18 +374,82 @@ static int huge_wrap_sigsetjmp(void) * map continguous physical blocks in contiguous virtual blocks. */ static unsigned -map_all_hugepages(struct hugepage_file *hugepg_tbl, - struct hugepage_info *hpi, int orig) +map_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi, + uint64_t *essential_memory __rte_unused, int orig) { int fd; unsigned i; void *virtaddr; void *vma_addr = NULL; size_t vma_len = 0; +#ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES + unsigned long nodemask[BITS_TO_LONGS(RTE_MAX_NUMA_NODES)] = {0UL}; + unsigned long maxnode = 0; + int node_id = -1; + bool numa_available = true; + + /* Check if kernel supports NUMA. */ + if (get_mempolicy(NULL, NULL, 0, 0, 0) < 0 && errno == ENOSYS) { + RTE_LOG(DEBUG, EAL, "NUMA is not supported.\n"); + numa_available = false; + } + + if (orig && numa_available) { + for (i = 0; i < RTE_MAX_NUMA_NODES; i++) + if (internal_config.socket_mem[i]) + maxnode = i + 1; + } +#endif for (i = 0; i < hpi->num_pages[0]; i++) { uint64_t hugepage_sz = hpi->hugepage_sz; +#ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES + if (maxnode) { + unsigned int j; + + for (j = 0; j < RTE_MAX_NUMA_NODES; j++) + if (essential_memory[j]) + break; + + if (j == RTE_MAX_NUMA_NODES) { + node_id = (node_id + 1) % RTE_MAX_NUMA_NODES; + while (!internal_config.socket_mem[node_id]) { + node_id++; + node_id %= RTE_MAX_NUMA_NODES; + } + } else { + node_id = j; + if (essential_memory[j] < hugepage_sz) + essential_memory[j] = 0; + else + essential_memory[j] -= hugepage_sz; + } + + nodemask[node_id / ULONG_BITS] = + 1UL << (node_id % ULONG_BITS); + + RTE_LOG(DEBUG, EAL, + "Setting policy MPOL_PREFERRED for socket %d\n", + node_id); + /* + * Due to old linux kernel bug (feature?) we have to + * increase maxnode by 1. It will be unconditionally + * decreased back to normal value inside the syscall + * handler. + */ + if (set_mempolicy(MPOL_PREFERRED, + nodemask, maxnode + 1) < 0) { + RTE_LOG(ERR, EAL, + "Failed to set policy MPOL_PREFERRED: " + "%s\n", strerror(errno)); + return i; + } + + nodemask[node_id / ULONG_BITS] = 0UL; + } +#endif + if (orig) { hugepg_tbl[i].file_id = i; hugepg_tbl[i].size = hugepage_sz; @@ -478,6 +560,10 @@ map_all_hugepages(struct hugepage_file *hugepg_tbl, vma_len -= hugepage_sz; } +#ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES + if (maxnode && set_mempolicy(MPOL_DEFAULT, NULL, 0) < 0) + RTE_LOG(ERR, EAL, "Failed to set mempolicy MPOL_DEFAULT\n"); +#endif return i; } @@ -562,6 +648,11 @@ find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi) if (hugepg_tbl[i].orig_va == va) { hugepg_tbl[i].socket_id = socket_id; hp_count++; +#ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES + RTE_LOG(DEBUG, EAL, + "Hugepage %s is on socket %d\n", + hugepg_tbl[i].filepath, socket_id); +#endif } } } @@ -1000,6 +1091,11 @@ rte_eal_hugepage_init(void) huge_register_sigbus(); + /* make a copy of socket_mem, needed for balanced allocation. */ + for (i = 0; i < RTE_MAX_NUMA_NODES; i++) + memory[i] = internal_config.socket_mem[i]; + + /* map all hugepages and sort them */ for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){ unsigned pages_old, pages_new; @@ -1017,7 +1113,8 @@ rte_eal_hugepage_init(void) /* map all hugepages available */ pages_old = hpi->num_pages[0]; - pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, 1); + pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, + memory, 1); if (pages_new < pages_old) { RTE_LOG(DEBUG, EAL, "%d not %d hugepages of size %u MB allocated\n", @@ -1060,7 +1157,7 @@ rte_eal_hugepage_init(void) sizeof(struct hugepage_file), cmp_physaddr); /* remap all hugepages */ - if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) != + if (map_all_hugepages(&tmp_hp[hp_offset], hpi, NULL, 0) != hpi->num_pages[0]) { RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n", (unsigned)(hpi->hugepage_sz / 0x100000)); diff --git a/mk/rte.app.mk b/mk/rte.app.mk index bcaf1b3..4fe22d1 100644 --- a/mk/rte.app.mk +++ b/mk/rte.app.mk @@ -186,6 +186,9 @@ ifeq ($(CONFIG_RTE_BUILD_SHARED_LIB),n) # The static libraries do not know their dependencies. # So linking with static library requires explicit dependencies. _LDLIBS-$(CONFIG_RTE_LIBRTE_EAL) += -lrt +ifeq ($(CONFIG_RTE_EXEC_ENV_LINUXAPP)$(CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES),yy) +_LDLIBS-$(CONFIG_RTE_LIBRTE_EAL) += -lnuma +endif _LDLIBS-$(CONFIG_RTE_LIBRTE_SCHED) += -lm _LDLIBS-$(CONFIG_RTE_LIBRTE_SCHED) += -lrt _LDLIBS-$(CONFIG_RTE_LIBRTE_METER) += -lm -- 2.7.4