net: fec: recycle pages for transmitted XDP frames
authorWei Fang <wei.fang@nxp.com>
Thu, 6 Jul 2023 08:10:10 +0000 (16:10 +0800)
committerPaolo Abeni <pabeni@redhat.com>
Tue, 11 Jul 2023 08:00:49 +0000 (10:00 +0200)
Once the XDP frames have been successfully transmitted through the
ndo_xdp_xmit() interface, it's the driver responsibility to free
the frames so that the page_pool can recycle the pages and reuse
them. However, this action is not implemented in the fec driver.
This leads to a user-visible problem that the console will print
the following warning log.

[  157.568851] page_pool_release_retry() stalled pool shutdown 1389 inflight 60 sec
[  217.983446] page_pool_release_retry() stalled pool shutdown 1389 inflight 120 sec
[  278.399006] page_pool_release_retry() stalled pool shutdown 1389 inflight 181 sec
[  338.812885] page_pool_release_retry() stalled pool shutdown 1389 inflight 241 sec
[  399.226946] page_pool_release_retry() stalled pool shutdown 1389 inflight 302 sec

Therefore, to solve this issue, we free XDP frames via xdp_return_frame()
while cleaning the tx BD ring.

Fixes: 6d6b39f180b8 ("net: fec: add initial XDP support")
Signed-off-by: Wei Fang <wei.fang@nxp.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
drivers/net/ethernet/freescale/fec.h
drivers/net/ethernet/freescale/fec_main.c

index 9939ccafb556609f8b6931ef42bb49f267aadea8..8c0226d061fecc66b4d5edf9b39f39eafe610d11 100644 (file)
@@ -544,10 +544,23 @@ enum {
        XDP_STATS_TOTAL,
 };
 
+enum fec_txbuf_type {
+       FEC_TXBUF_T_SKB,
+       FEC_TXBUF_T_XDP_NDO,
+};
+
+struct fec_tx_buffer {
+       union {
+               struct sk_buff *skb;
+               struct xdp_frame *xdp;
+       };
+       enum fec_txbuf_type type;
+};
+
 struct fec_enet_priv_tx_q {
        struct bufdesc_prop bd;
        unsigned char *tx_bounce[TX_RING_SIZE];
-       struct  sk_buff *tx_skbuff[TX_RING_SIZE];
+       struct fec_tx_buffer tx_buf[TX_RING_SIZE];
 
        unsigned short tx_stop_threshold;
        unsigned short tx_wake_threshold;
index 9ce0319b33c342293998a7cc54d556f1fb45e0bc..940d3afe1d247dbace1e23012532ee173f9441ac 100644 (file)
@@ -397,7 +397,7 @@ static void fec_dump(struct net_device *ndev)
                        fec16_to_cpu(bdp->cbd_sc),
                        fec32_to_cpu(bdp->cbd_bufaddr),
                        fec16_to_cpu(bdp->cbd_datlen),
-                       txq->tx_skbuff[index]);
+                       txq->tx_buf[index].skb);
                bdp = fec_enet_get_nextdesc(bdp, &txq->bd);
                index++;
        } while (bdp != txq->bd.base);
@@ -654,7 +654,7 @@ static int fec_enet_txq_submit_skb(struct fec_enet_priv_tx_q *txq,
 
        index = fec_enet_get_bd_index(last_bdp, &txq->bd);
        /* Save skb pointer */
-       txq->tx_skbuff[index] = skb;
+       txq->tx_buf[index].skb = skb;
 
        /* Make sure the updates to rest of the descriptor are performed before
         * transferring ownership.
@@ -672,9 +672,7 @@ static int fec_enet_txq_submit_skb(struct fec_enet_priv_tx_q *txq,
 
        skb_tx_timestamp(skb);
 
-       /* Make sure the update to bdp and tx_skbuff are performed before
-        * txq->bd.cur.
-        */
+       /* Make sure the update to bdp is performed before txq->bd.cur. */
        wmb();
        txq->bd.cur = bdp;
 
@@ -862,7 +860,7 @@ static int fec_enet_txq_submit_tso(struct fec_enet_priv_tx_q *txq,
        }
 
        /* Save skb pointer */
-       txq->tx_skbuff[index] = skb;
+       txq->tx_buf[index].skb = skb;
 
        skb_tx_timestamp(skb);
        txq->bd.cur = bdp;
@@ -952,16 +950,33 @@ static void fec_enet_bd_init(struct net_device *dev)
                for (i = 0; i < txq->bd.ring_size; i++) {
                        /* Initialize the BD for every fragment in the page. */
                        bdp->cbd_sc = cpu_to_fec16(0);
-                       if (bdp->cbd_bufaddr &&
-                           !IS_TSO_HEADER(txq, fec32_to_cpu(bdp->cbd_bufaddr)))
-                               dma_unmap_single(&fep->pdev->dev,
-                                                fec32_to_cpu(bdp->cbd_bufaddr),
-                                                fec16_to_cpu(bdp->cbd_datlen),
-                                                DMA_TO_DEVICE);
-                       if (txq->tx_skbuff[i]) {
-                               dev_kfree_skb_any(txq->tx_skbuff[i]);
-                               txq->tx_skbuff[i] = NULL;
+                       if (txq->tx_buf[i].type == FEC_TXBUF_T_SKB) {
+                               if (bdp->cbd_bufaddr &&
+                                   !IS_TSO_HEADER(txq, fec32_to_cpu(bdp->cbd_bufaddr)))
+                                       dma_unmap_single(&fep->pdev->dev,
+                                                        fec32_to_cpu(bdp->cbd_bufaddr),
+                                                        fec16_to_cpu(bdp->cbd_datlen),
+                                                        DMA_TO_DEVICE);
+                               if (txq->tx_buf[i].skb) {
+                                       dev_kfree_skb_any(txq->tx_buf[i].skb);
+                                       txq->tx_buf[i].skb = NULL;
+                               }
+                       } else {
+                               if (bdp->cbd_bufaddr)
+                                       dma_unmap_single(&fep->pdev->dev,
+                                                        fec32_to_cpu(bdp->cbd_bufaddr),
+                                                        fec16_to_cpu(bdp->cbd_datlen),
+                                                        DMA_TO_DEVICE);
+
+                               if (txq->tx_buf[i].xdp) {
+                                       xdp_return_frame(txq->tx_buf[i].xdp);
+                                       txq->tx_buf[i].xdp = NULL;
+                               }
+
+                               /* restore default tx buffer type: FEC_TXBUF_T_SKB */
+                               txq->tx_buf[i].type = FEC_TXBUF_T_SKB;
                        }
+
                        bdp->cbd_bufaddr = cpu_to_fec32(0);
                        bdp = fec_enet_get_nextdesc(bdp, &txq->bd);
                }
@@ -1360,6 +1375,7 @@ static void
 fec_enet_tx_queue(struct net_device *ndev, u16 queue_id)
 {
        struct  fec_enet_private *fep;
+       struct xdp_frame *xdpf;
        struct bufdesc *bdp;
        unsigned short status;
        struct  sk_buff *skb;
@@ -1387,16 +1403,31 @@ fec_enet_tx_queue(struct net_device *ndev, u16 queue_id)
 
                index = fec_enet_get_bd_index(bdp, &txq->bd);
 
-               skb = txq->tx_skbuff[index];
-               txq->tx_skbuff[index] = NULL;
-               if (!IS_TSO_HEADER(txq, fec32_to_cpu(bdp->cbd_bufaddr)))
-                       dma_unmap_single(&fep->pdev->dev,
-                                        fec32_to_cpu(bdp->cbd_bufaddr),
-                                        fec16_to_cpu(bdp->cbd_datlen),
-                                        DMA_TO_DEVICE);
-               bdp->cbd_bufaddr = cpu_to_fec32(0);
-               if (!skb)
-                       goto skb_done;
+               if (txq->tx_buf[index].type == FEC_TXBUF_T_SKB) {
+                       skb = txq->tx_buf[index].skb;
+                       txq->tx_buf[index].skb = NULL;
+                       if (bdp->cbd_bufaddr &&
+                           !IS_TSO_HEADER(txq, fec32_to_cpu(bdp->cbd_bufaddr)))
+                               dma_unmap_single(&fep->pdev->dev,
+                                                fec32_to_cpu(bdp->cbd_bufaddr),
+                                                fec16_to_cpu(bdp->cbd_datlen),
+                                                DMA_TO_DEVICE);
+                       bdp->cbd_bufaddr = cpu_to_fec32(0);
+                       if (!skb)
+                               goto tx_buf_done;
+               } else {
+                       xdpf = txq->tx_buf[index].xdp;
+                       if (bdp->cbd_bufaddr)
+                               dma_unmap_single(&fep->pdev->dev,
+                                                fec32_to_cpu(bdp->cbd_bufaddr),
+                                                fec16_to_cpu(bdp->cbd_datlen),
+                                                DMA_TO_DEVICE);
+                       bdp->cbd_bufaddr = cpu_to_fec32(0);
+                       if (!xdpf) {
+                               txq->tx_buf[index].type = FEC_TXBUF_T_SKB;
+                               goto tx_buf_done;
+                       }
+               }
 
                /* Check for errors. */
                if (status & (BD_ENET_TX_HB | BD_ENET_TX_LC |
@@ -1415,21 +1446,11 @@ fec_enet_tx_queue(struct net_device *ndev, u16 queue_id)
                                ndev->stats.tx_carrier_errors++;
                } else {
                        ndev->stats.tx_packets++;
-                       ndev->stats.tx_bytes += skb->len;
-               }
 
-               /* NOTE: SKBTX_IN_PROGRESS being set does not imply it's we who
-                * are to time stamp the packet, so we still need to check time
-                * stamping enabled flag.
-                */
-               if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_IN_PROGRESS &&
-                            fep->hwts_tx_en) &&
-                   fep->bufdesc_ex) {
-                       struct skb_shared_hwtstamps shhwtstamps;
-                       struct bufdesc_ex *ebdp = (struct bufdesc_ex *)bdp;
-
-                       fec_enet_hwtstamp(fep, fec32_to_cpu(ebdp->ts), &shhwtstamps);
-                       skb_tstamp_tx(skb, &shhwtstamps);
+                       if (txq->tx_buf[index].type == FEC_TXBUF_T_SKB)
+                               ndev->stats.tx_bytes += skb->len;
+                       else
+                               ndev->stats.tx_bytes += xdpf->len;
                }
 
                /* Deferred means some collisions occurred during transmit,
@@ -1438,10 +1459,32 @@ fec_enet_tx_queue(struct net_device *ndev, u16 queue_id)
                if (status & BD_ENET_TX_DEF)
                        ndev->stats.collisions++;
 
-               /* Free the sk buffer associated with this last transmit */
-               dev_kfree_skb_any(skb);
-skb_done:
-               /* Make sure the update to bdp and tx_skbuff are performed
+               if (txq->tx_buf[index].type == FEC_TXBUF_T_SKB) {
+                       /* NOTE: SKBTX_IN_PROGRESS being set does not imply it's we who
+                        * are to time stamp the packet, so we still need to check time
+                        * stamping enabled flag.
+                        */
+                       if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_IN_PROGRESS &&
+                                    fep->hwts_tx_en) && fep->bufdesc_ex) {
+                               struct skb_shared_hwtstamps shhwtstamps;
+                               struct bufdesc_ex *ebdp = (struct bufdesc_ex *)bdp;
+
+                               fec_enet_hwtstamp(fep, fec32_to_cpu(ebdp->ts), &shhwtstamps);
+                               skb_tstamp_tx(skb, &shhwtstamps);
+                       }
+
+                       /* Free the sk buffer associated with this last transmit */
+                       dev_kfree_skb_any(skb);
+               } else {
+                       xdp_return_frame(xdpf);
+
+                       txq->tx_buf[index].xdp = NULL;
+                       /* restore default tx buffer type: FEC_TXBUF_T_SKB */
+                       txq->tx_buf[index].type = FEC_TXBUF_T_SKB;
+               }
+
+tx_buf_done:
+               /* Make sure the update to bdp and tx_buf are performed
                 * before dirty_tx
                 */
                wmb();
@@ -3249,9 +3292,19 @@ static void fec_enet_free_buffers(struct net_device *ndev)
                for (i = 0; i < txq->bd.ring_size; i++) {
                        kfree(txq->tx_bounce[i]);
                        txq->tx_bounce[i] = NULL;
-                       skb = txq->tx_skbuff[i];
-                       txq->tx_skbuff[i] = NULL;
-                       dev_kfree_skb(skb);
+
+                       if (txq->tx_buf[i].type == FEC_TXBUF_T_SKB) {
+                               skb = txq->tx_buf[i].skb;
+                               txq->tx_buf[i].skb = NULL;
+                               dev_kfree_skb(skb);
+                       } else {
+                               if (txq->tx_buf[i].xdp) {
+                                       xdp_return_frame(txq->tx_buf[i].xdp);
+                                       txq->tx_buf[i].xdp = NULL;
+                               }
+
+                               txq->tx_buf[i].type = FEC_TXBUF_T_SKB;
+                       }
                }
        }
 }
@@ -3817,7 +3870,8 @@ static int fec_enet_txq_xmit_frame(struct fec_enet_private *fep,
                ebdp->cbd_esc = cpu_to_fec32(estatus);
        }
 
-       txq->tx_skbuff[index] = NULL;
+       txq->tx_buf[index].type = FEC_TXBUF_T_XDP_NDO;
+       txq->tx_buf[index].xdp = frame;
 
        /* Make sure the updates to rest of the descriptor are performed before
         * transferring ownership.