On 20/09/2021 12:24, Conor Walsh wrote:
From: Konstantin Ananyev <konstantin.anan...@intel.com>
Few changes in ioat sample behaviour:
- Always do SW copy for packet metadata (mbuf fields)
- Always use same lcore for both DMA requests enqueue and dequeue
Main reasons for that:
a) it is safer, as idxd PMD doesn't support MT safe enqueue/dequeue
(yet).
b) sort of more apples to apples comparison with sw copy.
c) from my testing things are faster that way.
Signed-off-by: Konstantin Ananyev <konstantin.anan...@intel.com>
---
examples/ioat/ioatfwd.c | 185 ++++++++++++++++++++++------------------
1 file changed, 101 insertions(+), 84 deletions(-)
diff --git a/examples/ioat/ioatfwd.c b/examples/ioat/ioatfwd.c
index b3977a8be5..1498343492 100644
--- a/examples/ioat/ioatfwd.c
+++ b/examples/ioat/ioatfwd.c
@@ -331,43 +331,36 @@ update_mac_addrs(struct rte_mbuf *m, uint32_t
dest_portid)
/* Perform packet copy there is a user-defined function. 8< */
static inline void
-pktmbuf_sw_copy(struct rte_mbuf *src, struct rte_mbuf *dst)
+pktmbuf_metadata_copy(const struct rte_mbuf *src, struct rte_mbuf *dst)
{
- /* Copy packet metadata */
- rte_memcpy(&dst->rearm_data,
- &src->rearm_data,
- offsetof(struct rte_mbuf, cacheline1)
- - offsetof(struct rte_mbuf, rearm_data));
+ dst->data_off = src->data_off;
+ memcpy(&dst->rx_descriptor_fields1, &src->rx_descriptor_fields1,
+ offsetof(struct rte_mbuf, buf_len) -
+ offsetof(struct rte_mbuf, rx_descriptor_fields1));
+}
- /* Copy packet data */
+/* Copy packet data */
+static inline void
+pktmbuf_sw_copy(struct rte_mbuf *src, struct rte_mbuf *dst)
+{
rte_memcpy(rte_pktmbuf_mtod(dst, char *),
rte_pktmbuf_mtod(src, char *), src->data_len);
}
/* >8 End of perform packet copy there is a user-defined function. */
Might need to redo these snippet markers as the function is now split
in two.
Will rework this with the overall documentation update after moving to
dmafwd.
<snip>
+static inline uint32_t
+ioat_dequeue(struct rte_mbuf *src[], struct rte_mbuf *dst[],
uint32_t num,
+ uint16_t dev_id)
+{
+ int32_t rc;
rc should be uint32_t, but this is removed in patch 4 of this set
during the change from raw to dma so it shouldn't really matter.
I belive int32_t is correct here, since the return value from
rte_ioat_completed_ops() is "int". Otherwise we would not be able to
error check the return.
If rc is negative, we set it to 0 before returning. This ensure that we
have a positive value, which can be safely typecast to uint32_t before
returning.
Thanks for the review, Conor!
+ /* Dequeue the mbufs from IOAT device. Since all memory
+ * is DPDK pinned memory and therefore all addresses should
+ * be valid, we don't check for copy errors
+ */
+ rc = rte_ioat_completed_ops(dev_id, num, NULL, NULL,
+ (void *)src, (void *)dst);
+ if (rc < 0) {
+ RTE_LOG(CRIT, IOAT,
+ "rte_ioat_completed_ops(%hu) failedi, error: %d\n",
+ dev_id, rte_errno);
+ rc = 0;
+ }
+ return rc;
+}
Reviewed-by: Conor Walsh <conor.wa...@intel.com>
