[1/4] net/ena: add fast mbuf free support
Checks
Commit Message
From: Dawid Gorecki <dgr@semihalf.com>
Add support for RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE offload. It can be
enabled if all the mbufs for a given queue belong to the same mempool
and their reference count is equal to 1.
Signed-off-by: Dawid Gorecki <dgr@semihalf.com>
Reviewed-by: Michal Krawczyk <mk@semihalf.com>
Reviewed-by: Shai Brandes <shaibran@amazon.com>
Reviewed-by: Amit Bernstein <amitbern@amazon.com>
---
doc/guides/nics/features/ena.ini | 1 +
doc/guides/rel_notes/release_22_07.rst | 7 ++++
drivers/net/ena/ena_ethdev.c | 49 ++++++++++++++++++++++++--
3 files changed, 55 insertions(+), 2 deletions(-)
@@ -7,6 +7,7 @@
Link status = Y
Link status event = Y
Rx interrupt = Y
+Fast mbuf free = Y
Free Tx mbuf on demand = Y
MTU update = Y
Scattered Rx = Y
@@ -87,6 +87,13 @@ New Features
Added an API which can get the device type of vDPA device.
+
+* **Updated Amazon ena driver.**
+
+ The new driver version (v2.7.0) includes:
+
+ * Added fast mbuf free feature support.
+
* **Updated Intel iavf driver.**
* Added Tx QoS queue rate limitation support.
@@ -36,6 +36,12 @@
#define ENA_MIN_RING_DESC 128
+/*
+ * We should try to keep ENA_CLEANUP_BUF_SIZE lower than
+ * RTE_MEMPOOL_CACHE_MAX_SIZE, so we can fit this in mempool local cache.
+ */
+#define ENA_CLEANUP_BUF_SIZE 256
+
#define ENA_PTYPE_HAS_HASH (RTE_PTYPE_L4_TCP | RTE_PTYPE_L4_UDP)
struct ena_stats {
@@ -2402,6 +2408,8 @@ static uint64_t ena_get_tx_port_offloads(struct ena_adapter *adapter)
port_offloads |= RTE_ETH_TX_OFFLOAD_MULTI_SEGS;
+ port_offloads |= RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE;
+
return port_offloads;
}
@@ -2414,9 +2422,12 @@ static uint64_t ena_get_rx_queue_offloads(struct ena_adapter *adapter)
static uint64_t ena_get_tx_queue_offloads(struct ena_adapter *adapter)
{
+ uint64_t queue_offloads = 0;
RTE_SET_USED(adapter);
- return 0;
+ queue_offloads |= RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE;
+
+ return queue_offloads;
}
static int ena_infos_get(struct rte_eth_dev *dev,
@@ -3001,13 +3012,38 @@ static int ena_xmit_mbuf(struct ena_ring *tx_ring, struct rte_mbuf *mbuf)
return 0;
}
+static __rte_always_inline size_t
+ena_tx_cleanup_mbuf_fast(struct rte_mbuf **mbufs_to_clean,
+ struct rte_mbuf *mbuf,
+ size_t mbuf_cnt,
+ size_t buf_size)
+{
+ struct rte_mbuf *m_next;
+
+ while (mbuf != NULL) {
+ m_next = mbuf->next;
+ mbufs_to_clean[mbuf_cnt++] = mbuf;
+ if (mbuf_cnt == buf_size) {
+ rte_mempool_put_bulk(mbufs_to_clean[0]->pool, (void **)mbufs_to_clean,
+ (unsigned int)mbuf_cnt);
+ mbuf_cnt = 0;
+ }
+ mbuf = m_next;
+ }
+
+ return mbuf_cnt;
+}
+
static int ena_tx_cleanup(void *txp, uint32_t free_pkt_cnt)
{
+ struct rte_mbuf *mbufs_to_clean[ENA_CLEANUP_BUF_SIZE];
struct ena_ring *tx_ring = (struct ena_ring *)txp;
+ size_t mbuf_cnt = 0;
unsigned int total_tx_descs = 0;
unsigned int total_tx_pkts = 0;
uint16_t cleanup_budget;
uint16_t next_to_clean = tx_ring->next_to_clean;
+ bool fast_free = tx_ring->offloads & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE;
/*
* If free_pkt_cnt is equal to 0, it means that the user requested
@@ -3032,7 +3068,12 @@ static int ena_tx_cleanup(void *txp, uint32_t free_pkt_cnt)
tx_info->timestamp = 0;
mbuf = tx_info->mbuf;
- rte_pktmbuf_free(mbuf);
+ if (fast_free) {
+ mbuf_cnt = ena_tx_cleanup_mbuf_fast(mbufs_to_clean, mbuf, mbuf_cnt,
+ ENA_CLEANUP_BUF_SIZE);
+ } else {
+ rte_pktmbuf_free(mbuf);
+ }
tx_info->mbuf = NULL;
tx_ring->empty_tx_reqs[next_to_clean] = req_id;
@@ -3052,6 +3093,10 @@ static int ena_tx_cleanup(void *txp, uint32_t free_pkt_cnt)
ena_com_update_dev_comp_head(tx_ring->ena_com_io_cq);
}
+ if (mbuf_cnt != 0)
+ rte_mempool_put_bulk(mbufs_to_clean[0]->pool,
+ (void **)mbufs_to_clean, mbuf_cnt);
+
/* Notify completion handler that full cleanup was performed */
if (free_pkt_cnt == 0 || total_tx_pkts < cleanup_budget)
tx_ring->last_cleanup_ticks = rte_get_timer_cycles();