[v2] net/nfp: fix issue of data len exceeds descriptor limitation
Checks
Commit Message
From: Long Wu <long.wu@corigine.com>
If dma_len is larger than NFDK_DESC_TX_DMA_LEN_HEAD, the value of
dma_len bitwise and NFDK_DESC_TX_DMA_LEN_HEAD maybe less than packet
head length and the packet will be dropped. Fill maximum dma_len in
first tx descriptor to make sure the whole head is included in the
first descriptor. In addition, add comments to better explain the
code flow.
Fixes: c73dced48c8c ("net/nfp: add NFDk Tx")
Cc: jin.liu@corigine.com
Cc: stable@dpdk.org
Signed-off-by: Long Wu <long.wu@corigine.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@corigine.com>
---
v2:
* Rewrite the commit message.
* Revise some logic according to the advice.
---
drivers/net/nfp/nfp_rxtx.c | 22 +++++++++++++++++++++-
1 file changed, 21 insertions(+), 1 deletion(-)
Comments
On 11/29/2022 1:21 AM, Chaoyong He wrote:
> From: Long Wu <long.wu@corigine.com>
>
> If dma_len is larger than NFDK_DESC_TX_DMA_LEN_HEAD, the value of
> dma_len bitwise and NFDK_DESC_TX_DMA_LEN_HEAD maybe less than packet
> head length and the packet will be dropped. Fill maximum dma_len in
> first tx descriptor to make sure the whole head is included in the
> first descriptor. In addition, add comments to better explain the
> code flow.
>
updated patch title as:
net/nfp: fix Tx packet drop for large data length
> Fixes: c73dced48c8c ("net/nfp: add NFDk Tx")
> Cc: jin.liu@corigine.com
> Cc: stable@dpdk.org
>
> Signed-off-by: Long Wu <long.wu@corigine.com>
> Reviewed-by: Niklas Söderlund <niklas.soderlund@corigine.com>
Applied to dpdk-next-net/main, thanks.
Below changes done while merging.
> + /*
> + * We will do our best to pass as much data as we can in descriptor
> + * and we need to make sure the first descriptor includes whole
> + * head since there is limitation in firmware side. Sometimes the
> + * value of dma_len bitwise & NFDK_DESC_TX_DMA_LEN_HEAD will less
> + * than packet head len.
> + */
Updated last line as:
" * value of 'dma_len & NFDK_DESC_TX_DMA_LEN_HEAD' will be less "
<...>
>
> + /*
> + * The rest of the data (if any) will be in larger dma descritors
s/descritors/descriptors/
s/dma/DMA/
@@ -1063,6 +1063,7 @@ nfp_net_nfdk_tx_maybe_close_block(struct nfp_net_txq *txq, struct rte_mbuf *pkt)
if (unlikely(n_descs > NFDK_TX_DESC_GATHER_MAX))
return -EINVAL;
+ /* Under count by 1 (don't count meta) for the round down to work out */
n_descs += !!(pkt->ol_flags & RTE_MBUF_F_TX_TCP_SEG);
if (round_down(txq->wr_p, NFDK_TX_DESC_BLOCK_CNT) !=
@@ -1219,8 +1220,19 @@ nfp_net_nfdk_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pk
} else {
type = NFDK_DESC_TX_TYPE_GATHER;
}
+
+ /* Implicitly truncates to chunk in below logic */
dma_len -= 1;
- dlen_type = (NFDK_DESC_TX_DMA_LEN_HEAD & dma_len) |
+
+ /*
+ * We will do our best to pass as much data as we can in descriptor
+ * and we need to make sure the first descriptor includes whole
+ * head since there is limitation in firmware side. Sometimes the
+ * value of dma_len bitwise & NFDK_DESC_TX_DMA_LEN_HEAD will less
+ * than packet head len.
+ */
+ dlen_type = (dma_len > NFDK_DESC_TX_DMA_LEN_HEAD ?
+ NFDK_DESC_TX_DMA_LEN_HEAD : dma_len) |
(NFDK_DESC_TX_TYPE_HEAD & (type << 12));
ktxds->dma_len_type = rte_cpu_to_le_16(dlen_type);
dma_addr = rte_mbuf_data_iova(pkt);
@@ -1230,10 +1242,18 @@ nfp_net_nfdk_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pk
ktxds->dma_addr_lo = rte_cpu_to_le_32(dma_addr & 0xffffffff);
ktxds++;
+ /*
+ * Preserve the original dlen_type, this way below the EOP logic
+ * can use dlen_type.
+ */
tmp_dlen = dlen_type & NFDK_DESC_TX_DMA_LEN_HEAD;
dma_len -= tmp_dlen;
dma_addr += tmp_dlen + 1;
+ /*
+ * The rest of the data (if any) will be in larger dma descritors
+ * and is handled with the dma_len loop.
+ */
while (pkt) {
if (*lmbuf)
rte_pktmbuf_free_seg(*lmbuf);