[dpdk-dev] [PATCH v9 10/21] malloc: add function to check if socket is external

[Anatoly Burakov <anatoly.burakov@intel.com>]

An API is needed to check whether a particular socket ID belongs
to an internal or external heap. Prime user of this would be
mempool allocator, because normal assumptions of IOVA
contiguousness in IOVA as VA mode do not hold in case of
externally allocated memory.

Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
---
 lib/librte_eal/common/include/rte_malloc.h | 15 +++++++++++++
 lib/librte_eal/common/rte_malloc.c         | 25 ++++++++++++++++++++++
 lib/librte_eal/rte_eal_version.map         |  1 +
 lib/librte_mempool/rte_mempool.c           | 22 ++++++++++++++++---
 4 files changed, 60 insertions(+), 3 deletions(-)

diff --git a/lib/librte_eal/common/include/rte_malloc.h b/lib/librte_eal/common/include/rte_malloc.h
index 8870732a6..403271ddc 100644
--- a/lib/librte_eal/common/include/rte_malloc.h
+++ b/lib/librte_eal/common/include/rte_malloc.h
@@ -277,6 +277,21 @@ rte_malloc_get_socket_stats(int socket,
 int __rte_experimental
 rte_malloc_heap_get_socket(const char *name);
 
+/**
+ * Check if a given socket ID refers to externally allocated memory.
+ *
+ * @note Passing SOCKET_ID_ANY will return 0.
+ *
+ * @param socket_id
+ *   Socket ID to check
+ * @return
+ *   1 if socket ID refers to externally allocated memory
+ *   0 if socket ID refers to internal DPDK memory
+ *   -1 if socket ID is invalid
+ */
+int __rte_experimental
+rte_malloc_heap_socket_is_external(int socket_id);
+
 /**
  * Dump statistics.
  *
diff --git a/lib/librte_eal/common/rte_malloc.c b/lib/librte_eal/common/rte_malloc.c
index b807dfe09..fa81d7862 100644
--- a/lib/librte_eal/common/rte_malloc.c
+++ b/lib/librte_eal/common/rte_malloc.c
@@ -220,6 +220,31 @@ rte_malloc_heap_get_socket(const char *name)
 	return ret;
 }
 
+int
+rte_malloc_heap_socket_is_external(int socket_id)
+{
+	struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
+	unsigned int idx;
+	int ret = -1;
+
+	if (socket_id == SOCKET_ID_ANY)
+		return 0;
+
+	rte_rwlock_read_lock(&mcfg->memory_hotplug_lock);
+	for (idx = 0; idx < RTE_MAX_HEAPS; idx++) {
+		struct malloc_heap *tmp = &mcfg->malloc_heaps[idx];
+
+		if ((int)tmp->socket_id == socket_id) {
+			/* external memory always has large socket ID's */
+			ret = tmp->socket_id >= RTE_MAX_NUMA_NODES;
+			break;
+		}
+	}
+	rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
+
+	return ret;
+}
+
 /*
  * Print stats on memory type. If type is NULL, info on all types is printed
  */
diff --git a/lib/librte_eal/rte_eal_version.map b/lib/librte_eal/rte_eal_version.map
index d8f9665b8..bd60506af 100644
--- a/lib/librte_eal/rte_eal_version.map
+++ b/lib/librte_eal/rte_eal_version.map
@@ -319,6 +319,7 @@ EXPERIMENTAL {
 	rte_log_register_type_and_pick_level;
 	rte_malloc_dump_heaps;
 	rte_malloc_heap_get_socket;
+	rte_malloc_heap_socket_is_external;
 	rte_mem_alloc_validator_register;
 	rte_mem_alloc_validator_unregister;
 	rte_mem_event_callback_register;
diff --git a/lib/librte_mempool/rte_mempool.c b/lib/librte_mempool/rte_mempool.c
index 2ed539f01..683b216f9 100644
--- a/lib/librte_mempool/rte_mempool.c
+++ b/lib/librte_mempool/rte_mempool.c
@@ -428,12 +428,18 @@ rte_mempool_populate_default(struct rte_mempool *mp)
 	rte_iova_t iova;
 	unsigned mz_id, n;
 	int ret;
-	bool no_contig, try_contig, no_pageshift;
+	bool no_contig, try_contig, no_pageshift, external;
 
 	ret = mempool_ops_alloc_once(mp);
 	if (ret != 0)
 		return ret;
 
+	/* check if we can retrieve a valid socket ID */
+	ret = rte_malloc_heap_socket_is_external(mp->socket_id);
+	if (ret < 0)
+		return -EINVAL;
+	external = ret;
+
 	/* mempool must not be populated */
 	if (mp->nb_mem_chunks != 0)
 		return -EEXIST;
@@ -481,9 +487,19 @@ rte_mempool_populate_default(struct rte_mempool *mp)
 	 * in one contiguous chunk as well (otherwise we might end up wasting a
 	 * 1G page on a 10MB memzone). If we fail to get enough contiguous
 	 * memory, then we'll go and reserve space page-by-page.
+	 *
+	 * We also have to take into account the fact that memory that we're
+	 * going to allocate from can belong to an externally allocated memory
+	 * area, in which case the assumption of IOVA as VA mode being
+	 * synonymous with IOVA contiguousness will not hold. We should also try
+	 * to go for contiguous memory even if we're in no-huge mode, because
+	 * external memory may in fact be IOVA-contiguous.
 	 */
-	no_pageshift = no_contig || rte_eal_iova_mode() == RTE_IOVA_VA;
-	try_contig = !no_contig && !no_pageshift && rte_eal_has_hugepages();
+	external = rte_malloc_heap_socket_is_external(mp->socket_id) == 1;
+	no_pageshift = no_contig ||
+			(!external && rte_eal_iova_mode() == RTE_IOVA_VA);
+	try_contig = !no_contig && !no_pageshift &&
+			(rte_eal_has_hugepages() || external);
 
 	if (no_pageshift) {
 		pg_sz = 0;
-- 
2.17.1