[dpdk-dev] [PATCH v5 1/2] mem: balanced allocation of hugepages

[Ilya Maximets]

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.

libnuma added as a general dependency for EAL.

Fixes: 77988fc08dc5 ("mem: fix allocating all free hugepages")

Signed-off-by: Ilya Maximets <i.maximets at samsung.com>
---
 lib/librte_eal/linuxapp/eal/Makefile     |  1 +
 lib/librte_eal/linuxapp/eal/eal_memory.c | 94 ++++++++++++++++++++++++++++++--
 mk/rte.app.mk                            |  3 +
 3 files changed, 94 insertions(+), 4 deletions(-)

diff --git a/lib/librte_eal/linuxapp/eal/Makefile b/lib/librte_eal/linuxapp/eal/Makefile
index 640afd0..1440fc5 100644
--- a/lib/librte_eal/linuxapp/eal/Makefile
+++ b/lib/librte_eal/linuxapp/eal/Makefile
@@ -50,6 +50,7 @@ LDLIBS += -ldl
 LDLIBS += -lpthread
 LDLIBS += -lgcc_s
 LDLIBS += -lrt
+LDLIBS += -lnuma
 
 # 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 9c9baf6..5947434 100644
--- a/lib/librte_eal/linuxapp/eal/eal_memory.c
+++ b/lib/librte_eal/linuxapp/eal/eal_memory.c
@@ -54,6 +54,7 @@
 #include <sys/time.h>
 #include <signal.h>
 #include <setjmp.h>
+#include <numaif.h>
 
 #include <rte_log.h>
 #include <rte_memory.h>
@@ -358,6 +359,19 @@ static int huge_wrap_sigsetjmp(void)
 	return sigsetjmp(huge_jmpenv, 1);
 }
 
+#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
+
 /*
  * 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
@@ -366,18 +380,78 @@ 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, int orig)
 {
 	int fd;
 	unsigned i;
 	void *virtaddr;
 	void *vma_addr = NULL;
 	size_t vma_len = 0;
+	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;
+	}
 
 	for (i = 0; i < hpi->num_pages[0]; i++) {
 		uint64_t hugepage_sz = hpi->hugepage_sz;
 
+		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;
+		}
+
 		if (orig) {
 			hugepg_tbl[i].file_id = i;
 			hugepg_tbl[i].size = hugepage_sz;
@@ -488,6 +562,9 @@ map_all_hugepages(struct hugepage_file *hugepg_tbl,
 		vma_len -= hugepage_sz;
 	}
 
+	if (maxnode && set_mempolicy(MPOL_DEFAULT, NULL, 0) < 0)
+		RTE_LOG(ERR, EAL, "Failed to set mempolicy MPOL_DEFAULT\n");
+
 	return i;
 }
 
@@ -572,6 +649,9 @@ 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++;
+				RTE_LOG(DEBUG, EAL,
+					"Hugepage %s is on socket %d\n",
+					hugepg_tbl[i].filepath, socket_id);
 			}
 		}
 	}
@@ -1010,6 +1090,11 @@ 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;
@@ -1027,7 +1112,8 @@ 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",
@@ -1070,7 +1156,7 @@ 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..5f370c9 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 ($(CONFIG_RTE_EXEC_ENV_LINUXAPP),y)
+_LDLIBS-$(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