[PATCH] mempool/cnxk: limit usage of async allocs

[PATCH] mempool/cnxk: limit usage of async allocs

[Ashwin Sekhar T K]

Currently mempool_cnxk driver uses asynchronous allocation
for all pools. This asynchronous allocation can result in
local caching of additional 512 objects on a single core
even when cache is disabled. This will eventually lead to
starvation on pools where the number of objects is very less.

This commit changes this logic to use asynchronous allocation on
only those pools which have local cache enabled. Also the async buffer
size will be RTE_ALIGN_CEIL(rte_mempool->cache_size, 16). This
means that when cache is disabled, async alloc will be completely
disabled and when cache is enabled, the additional caching due
to asynchronous allocation will be limited.

A limitation has been added to cnxk documentation warning the users
to adjust the local cache sizes accordingly.

Signed-off-by: Ashwin Sekhar T K <asekhar@marvell.com>
---
 doc/guides/nics/cnxk.rst                 |  18 ++++
 drivers/mempool/cnxk/cn10k_mempool_ops.c | 103 ++++++++++++++++++++---
 2 files changed, 110 insertions(+), 11 deletions(-)

diff --git a/doc/guides/nics/cnxk.rst b/doc/guides/nics/cnxk.rst
index 9229056f6f..5d90d22e2b 100644
--- a/doc/guides/nics/cnxk.rst
+++ b/doc/guides/nics/cnxk.rst
@@ -433,6 +433,24 @@ The OCTEON CN9K/CN10K SoC family NIC has inbuilt HW assisted external mempool ma
 as it is performance wise most effective way for packet allocation and Tx buffer
 recycling on OCTEON 9 SoC platform.
 
+``mempool_cnxk`` rte_mempool cache sizes for CN10K
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The OCTEON CN10K SoC Family supports asynchronous batch allocation of
+objects from an NPA pool. In the CNXK mempool driver, asynchronous batch
+allocation is enabled when local caches are enabled. This asynchronous
+batch allocation will be using an additional local async buffer whose size
+will be equal to ``RTE_ALIGN_CEIL(rte_mempool->cache_size, 16)``. This can
+result in additional objects being cached locally. While creating an
+rte_mempool using ``mempool_cnxk`` driver for OCTEON CN10K, this must be
+taken into consideration and the local cache sizes should be adjusted
+accordingly so that starvation does not happen.
+
+For Eg: If the ``cache_size`` passed into ``rte_mempool_create`` is ``8``,
+then the max objects than can get cached locally on a core would be the
+sum of max objects in the local cache + max objects in the async buffer
+i.e ``8 + RTE_ALIGN_CEIL(8, 16) = 24``.
+
 CRC stripping
 ~~~~~~~~~~~~~
 
diff --git a/drivers/mempool/cnxk/cn10k_mempool_ops.c b/drivers/mempool/cnxk/cn10k_mempool_ops.c
index ff0015d8de..41b755b52b 100644
--- a/drivers/mempool/cnxk/cn10k_mempool_ops.c
+++ b/drivers/mempool/cnxk/cn10k_mempool_ops.c
@@ -10,6 +10,7 @@
 #define BATCH_ALLOC_SZ              ROC_CN10K_NPA_BATCH_ALLOC_MAX_PTRS
 #define BATCH_OP_DATA_TABLE_MZ_NAME "batch_op_data_table_mz"
 #define BATCH_ALLOC_WAIT_US         5
+#define BATCH_ALLOC_RETRIES         4
 
 enum batch_op_status {
 	BATCH_ALLOC_OP_NOT_ISSUED = 0,
@@ -25,6 +26,7 @@ struct batch_op_mem {
 
 struct batch_op_data {
 	uint64_t lmt_addr;
+	uint32_t max_async_batch;
 	struct batch_op_mem mem[RTE_MAX_LCORE] __rte_aligned(ROC_ALIGN);
 };
 
@@ -97,6 +99,10 @@ batch_op_init(struct rte_mempool *mp)
 	}
 
 	op_data->lmt_addr = roc_idev_lmt_base_addr_get();
+	op_data->max_async_batch =
+		RTE_MIN((unsigned int)BATCH_ALLOC_SZ,
+			RTE_ALIGN_CEIL(mp->cache_size, ROC_ALIGN / 8));
+
 	batch_op_data_set(mp->pool_id, op_data);
 	rte_wmb();
 
@@ -117,13 +123,17 @@ batch_op_fini(struct rte_mempool *mp)
 		return;
 	}
 
+	/* If max_async_batch == 0, then batch mem will be empty */
+	if (op_data->max_async_batch == 0)
+		goto free_op_data;
+
 	rte_wmb();
 	for (i = 0; i < RTE_MAX_LCORE; i++) {
 		struct batch_op_mem *mem = &op_data->mem[i];
 
 		if (mem->status == BATCH_ALLOC_OP_ISSUED) {
 			mem->sz = roc_npa_aura_batch_alloc_extract(
-				mem->objs, mem->objs, BATCH_ALLOC_SZ);
+				mem->objs, mem->objs, op_data->max_async_batch);
 			mem->status = BATCH_ALLOC_OP_DONE;
 		}
 		if (mem->status == BATCH_ALLOC_OP_DONE) {
@@ -133,6 +143,7 @@ batch_op_fini(struct rte_mempool *mp)
 		}
 	}
 
+free_op_data:
 	rte_free(op_data);
 	batch_op_data_set(mp->pool_id, NULL);
 	rte_wmb();
@@ -172,6 +183,9 @@ cn10k_mempool_get_count(const struct rte_mempool *mp)
 	int i;
 
 	op_data = batch_op_data_get(mp->pool_id);
+	/* If max_async_batch == 0, then batch alloc mem will be empty */
+	if (op_data->max_async_batch == 0)
+		goto npa_pool_count;
 
 	rte_wmb();
 	for (i = 0; i < RTE_MAX_LCORE; i++) {
@@ -179,19 +193,27 @@ cn10k_mempool_get_count(const struct rte_mempool *mp)
 
 		if (mem->status == BATCH_ALLOC_OP_ISSUED)
 			count += roc_npa_aura_batch_alloc_count(
-				mem->objs, BATCH_ALLOC_SZ, BATCH_ALLOC_WAIT_US);
+				mem->objs, op_data->max_async_batch,
+				BATCH_ALLOC_WAIT_US);
 
 		if (mem->status == BATCH_ALLOC_OP_DONE)
 			count += mem->sz;
 	}
 
+npa_pool_count:
 	count += cnxk_mempool_get_count(mp);
 
 	return count;
 }
 
-static int __rte_hot
-cn10k_mempool_deq(struct rte_mempool *mp, void **obj_table, unsigned int n)
+static inline unsigned int __rte_hot
+mempool_deq(struct rte_mempool *mp, void **obj_table, unsigned int n)
+{
+	return cnxk_mempool_deq(mp, obj_table, n) ? 0 : n;
+}
+
+static inline unsigned int __rte_hot
+mempool_deq_batch_async(struct rte_mempool *mp, void **obj_table, unsigned int n)
 {
 	struct batch_op_data *op_data;
 	struct batch_op_mem *mem;
@@ -205,24 +227,24 @@ cn10k_mempool_deq(struct rte_mempool *mp, void **obj_table, unsigned int n)
 
 	/* Issue batch alloc */
 	if (mem->status == BATCH_ALLOC_OP_NOT_ISSUED) {
-		rc = roc_npa_aura_batch_alloc_issue(mp->pool_id, mem->objs,
-						    BATCH_ALLOC_SZ, 0, 1);
+		rc = roc_npa_aura_batch_alloc_issue(
+			mp->pool_id, mem->objs, op_data->max_async_batch, 0, 1);
 		/* If issue fails, try falling back to default alloc */
 		if (unlikely(rc))
-			return cnxk_mempool_deq(mp, obj_table, n);
+			return mempool_deq(mp, obj_table, n);
 		mem->status = BATCH_ALLOC_OP_ISSUED;
 	}
 
-	retry = 4;
+	retry = BATCH_ALLOC_RETRIES;
 	while (loop) {
 		unsigned int cur_sz;
 
 		if (mem->status == BATCH_ALLOC_OP_ISSUED) {
 			mem->sz = roc_npa_aura_batch_alloc_extract(
-				mem->objs, mem->objs, BATCH_ALLOC_SZ);
+				mem->objs, mem->objs, op_data->max_async_batch);
 
 			/* If partial alloc reduce the retry count */
-			retry -= (mem->sz != BATCH_ALLOC_SZ);
+			retry -= (mem->sz != op_data->max_async_batch);
 			/* Break the loop if retry count exhausted */
 			loop = !!retry;
 			mem->status = BATCH_ALLOC_OP_DONE;
@@ -244,13 +266,72 @@ cn10k_mempool_deq(struct rte_mempool *mp, void **obj_table, unsigned int n)
 		/* Issue next batch alloc if pointers are exhausted */
 		if (mem->sz == 0) {
 			rc = roc_npa_aura_batch_alloc_issue(
-				mp->pool_id, mem->objs, BATCH_ALLOC_SZ, 0, 1);
+				mp->pool_id, mem->objs,
+				op_data->max_async_batch, 0, 1);
 			/* Break loop if issue failed and set status */
 			loop &= !rc;
 			mem->status = !rc;
 		}
 	}
 
+	return count;
+}
+
+static inline unsigned int __rte_hot
+mempool_deq_batch_sync(struct rte_mempool *mp, void **obj_table, unsigned int n)
+{
+	struct batch_op_data *op_data;
+	struct batch_op_mem *mem;
+	unsigned int count = 0;
+	int tid, retry, rc;
+
+	op_data = batch_op_data_get(mp->pool_id);
+	tid = rte_lcore_id();
+	mem = &op_data->mem[tid];
+
+	retry = BATCH_ALLOC_RETRIES;
+	while (count != n && retry) {
+		unsigned int cur_sz, batch_sz;
+
+		cur_sz = n - count;
+		batch_sz = RTE_MIN(BATCH_ALLOC_SZ, (int)cur_sz);
+
+		/* Issue batch alloc */
+		rc = roc_npa_aura_batch_alloc_issue(mp->pool_id, mem->objs,
+						    batch_sz, 0, 1);
+
+		/* If issue fails, try falling back to default alloc */
+		if (unlikely(rc))
+			return count +
+			       mempool_deq(mp, obj_table + count, n - count);
+
+		cur_sz = roc_npa_aura_batch_alloc_extract(mem->objs, mem->objs,
+							  batch_sz);
+
+		/* Dequeue the pointers */
+		memcpy(&obj_table[count], mem->objs,
+		       cur_sz * sizeof(uintptr_t));
+		count += cur_sz;
+
+		/* If partial alloc reduce the retry count */
+		retry -= (batch_sz != cur_sz);
+	}
+
+	return count;
+}
+
+static int __rte_hot
+cn10k_mempool_deq(struct rte_mempool *mp, void **obj_table, unsigned int n)
+{
+	struct batch_op_data *op_data;
+	unsigned int count = 0;
+
+	op_data = batch_op_data_get(mp->pool_id);
+	if (op_data->max_async_batch)
+		count = mempool_deq_batch_async(mp, obj_table, n);
+	else
+		count = mempool_deq_batch_sync(mp, obj_table, n);
+
 	if (unlikely(count != n)) {
 		/* No partial alloc allowed. Free up allocated pointers */
 		cn10k_mempool_enq(mp, obj_table, count);
-- 
2.25.1