Take Rx buffer start alignment and end-padding into account.
Reviewed-by: Andy Moreton <amoreton at solarflare.com>
Signed-off-by: Andrew Rybchenko <arybchenko at solarflare.com>
---
drivers/net/sfc/efx/sfc_rx.c | 110 +++++++++++++++++++++++++++++++++++++++++--
drivers/net/sfc/efx/sfc_rx.h | 1 +
2 files changed, 107 insertions(+), 4 deletions(-)
diff --git a/drivers/net/sfc/efx/sfc_rx.c b/drivers/net/sfc/efx/sfc_rx.c
index d4a1a02..bfe92bf 100644
--- a/drivers/net/sfc/efx/sfc_rx.c
+++ b/drivers/net/sfc/efx/sfc_rx.c
@@ -66,6 +66,106 @@ sfc_rx_qcheck_conf(struct sfc_adapter *sa,
return rc;
}
+static unsigned int
+sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
+{
+ int data_off;
+ int order;
+
+ /* The mbuf object itself is always cache line aligned */
+ order = rte_bsf32(RTE_CACHE_LINE_SIZE);
+
+ /* Data offset from mbuf object start */
+ data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
+ RTE_PKTMBUF_HEADROOM;
+
+ order = MIN(order, rte_bsf32(data_off));
+
+ return 1u << (order - 1);
+}
+
+static uint16_t
+sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
+{
+ const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
+ const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
+ const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
+ uint16_t buf_size;
+ unsigned int buf_aligned;
+ unsigned int start_alignment;
+ unsigned int end_padding_alignment;
+
+ /* Below it is assumed that both alignments are power of 2 */
+ SFC_ASSERT(rte_is_power_of_2(nic_align_start));
+ SFC_ASSERT(rte_is_power_of_2(nic_align_end));
+
+ /*
+ * mbuf is always cache line aligned, double-check
+ * that it meets rx buffer start alignment requirements.
+ */
+
+ /* Start from mbuf pool data room size */
+ buf_size = rte_pktmbuf_data_room_size(mb_pool);
+
+ /* Remove headroom */
+ if (buf_size <= RTE_PKTMBUF_HEADROOM) {
+ sfc_err(sa,
+ "RxQ mbuf pool %s object data room size %u is smaller
than headroom %u",
+ mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
+ return 0;
+ }
+ buf_size -= RTE_PKTMBUF_HEADROOM;
+
+ /* Calculate guaranteed data start alignment */
+ buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
+
+ /* Reserve space for start alignment */
+ if (buf_aligned < nic_align_start) {
+ start_alignment = nic_align_start - buf_aligned;
+ if (buf_size <= start_alignment) {
+ sfc_err(sa,
+ "RxQ mbuf pool %s object data room size %u is
insufficient for headroom %u and buffer start alignment %u required by NIC",
+ mb_pool->name,
+ rte_pktmbuf_data_room_size(mb_pool),
+ RTE_PKTMBUF_HEADROOM, start_alignment);
+ return 0;
+ }
+ buf_aligned = nic_align_start;
+ buf_size -= start_alignment;
+ } else {
+ start_alignment = 0;
+ }
+
+ /* Make sure that end padding does not write beyond the buffer */
+ if (buf_aligned < nic_align_end) {
+ /*
+ * Estimate space which can be lost. If guarnteed buffer
+ * size is odd, lost space is (nic_align_end - 1). More
+ * accurate formula is below.
+ */
+ end_padding_alignment = nic_align_end -
+ MIN(buf_aligned, 1 << (rte_bsf32(buf_size) - 1));
+ if (buf_size <= end_padding_alignment) {
+ sfc_err(sa,
+ "RxQ mbuf pool %s object data room size %u is
insufficient for headroom %u, buffer start alignment %u and end padding
alignment %u required by NIC",
+ mb_pool->name,
+ rte_pktmbuf_data_room_size(mb_pool),
+ RTE_PKTMBUF_HEADROOM, start_alignment,
+ end_padding_alignment);
+ return 0;
+ }
+ buf_size -= end_padding_alignment;
+ } else {
+ /*
+ * Start is aligned the same or better than end,
+ * just align length.
+ */
+ buf_size = P2ALIGN(buf_size, nic_align_end);
+ }
+
+ return buf_size;
+}
+
int
sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
uint16_t nb_rx_desc, unsigned int socket_id,
@@ -73,6 +173,7 @@ sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
struct rte_mempool *mb_pool)
{
int rc;
+ uint16_t buf_size;
struct sfc_rxq_info *rxq_info;
unsigned int evq_index;
struct sfc_evq *evq;
@@ -82,10 +183,10 @@ sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
if (rc != 0)
goto fail_bad_conf;
- if (rte_pktmbuf_data_room_size(mb_pool) <= RTE_PKTMBUF_HEADROOM) {
- sfc_err(sa, "RxQ %u mbuf is too small, %u vs headroom %u",
- sw_index, rte_pktmbuf_data_room_size(mb_pool),
- RTE_PKTMBUF_HEADROOM);
+ buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
+ if (buf_size == 0) {
+ sfc_err(sa, "RxQ %u mbuf pool object size is too small",
+ sw_index);
goto fail_bad_conf;
}
@@ -126,6 +227,7 @@ sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
rxq->evq = evq;
rxq->ptr_mask = rxq_info->entries - 1;
rxq->refill_mb_pool = mb_pool;
+ rxq->buf_size = buf_size;
rxq->hw_index = sw_index;
rxq->state = SFC_RXQ_INITIALIZED;
diff --git a/drivers/net/sfc/efx/sfc_rx.h b/drivers/net/sfc/efx/sfc_rx.h
index 503c3df..3e24434 100644
--- a/drivers/net/sfc/efx/sfc_rx.h
+++ b/drivers/net/sfc/efx/sfc_rx.h
@@ -71,6 +71,7 @@ struct sfc_rxq {
unsigned int ptr_mask;
/* Used on refill */
+ uint16_t buf_size;
struct rte_mempool *refill_mb_pool;
efx_rxq_t *common;
efsys_mem_t mem;
--
2.5.5
