mbuf: make segment prefree function public
[dpdk.git] / drivers / net / fm10k / fm10k_rxtx.c
index 7d5e32c..c9bb04a 100644 (file)
@@ -1,7 +1,7 @@
 /*-
  *   BSD LICENSE
  *
- *   Copyright(c) 2013-2015 Intel Corporation. All rights reserved.
+ *   Copyright(c) 2013-2016 Intel Corporation. All rights reserved.
  *   All rights reserved.
  *
  *   Redistribution and use in source and binary forms, with or without
@@ -35,6 +35,7 @@
 
 #include <rte_ethdev.h>
 #include <rte_common.h>
+#include <rte_net.h>
 #include "fm10k.h"
 #include "base/fm10k_type.h"
 
@@ -65,16 +66,44 @@ static inline void dump_rxd(union fm10k_rx_desc *rxd)
 }
 #endif
 
+#define FM10K_TX_OFFLOAD_MASK (  \
+               PKT_TX_VLAN_PKT |        \
+               PKT_TX_IP_CKSUM |        \
+               PKT_TX_L4_MASK |         \
+               PKT_TX_TCP_SEG)
+
+#define FM10K_TX_OFFLOAD_NOTSUP_MASK \
+               (PKT_TX_OFFLOAD_MASK ^ FM10K_TX_OFFLOAD_MASK)
+
+/* @note: When this function is changed, make corresponding change to
+ * fm10k_dev_supported_ptypes_get()
+ */
 static inline void
 rx_desc_to_ol_flags(struct rte_mbuf *m, const union fm10k_rx_desc *d)
 {
-       uint16_t ptype;
-       static const uint16_t pt_lut[] = { 0,
-               PKT_RX_IPV4_HDR, PKT_RX_IPV4_HDR_EXT,
-               PKT_RX_IPV6_HDR, PKT_RX_IPV6_HDR_EXT,
-               0, 0, 0
+       static const uint32_t
+               ptype_table[FM10K_RXD_PKTTYPE_MASK >> FM10K_RXD_PKTTYPE_SHIFT]
+                       __rte_cache_aligned = {
+               [FM10K_PKTTYPE_OTHER] = RTE_PTYPE_L2_ETHER,
+               [FM10K_PKTTYPE_IPV4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4,
+               [FM10K_PKTTYPE_IPV4_EX] = RTE_PTYPE_L2_ETHER |
+                       RTE_PTYPE_L3_IPV4_EXT,
+               [FM10K_PKTTYPE_IPV6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6,
+               [FM10K_PKTTYPE_IPV6_EX] = RTE_PTYPE_L2_ETHER |
+                       RTE_PTYPE_L3_IPV6_EXT,
+               [FM10K_PKTTYPE_IPV4 | FM10K_PKTTYPE_TCP] = RTE_PTYPE_L2_ETHER |
+                       RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_TCP,
+               [FM10K_PKTTYPE_IPV6 | FM10K_PKTTYPE_TCP] = RTE_PTYPE_L2_ETHER |
+                       RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_TCP,
+               [FM10K_PKTTYPE_IPV4 | FM10K_PKTTYPE_UDP] = RTE_PTYPE_L2_ETHER |
+                       RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_UDP,
+               [FM10K_PKTTYPE_IPV6 | FM10K_PKTTYPE_UDP] = RTE_PTYPE_L2_ETHER |
+                       RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_UDP,
        };
 
+       m->packet_type = ptype_table[(d->w.pkt_info & FM10K_RXD_PKTTYPE_MASK)
+                                               >> FM10K_RXD_PKTTYPE_SHIFT];
+
        if (d->w.pkt_info & FM10K_RXD_RSSTYPE_MASK)
                m->ol_flags |= PKT_RX_RSS_HASH;
 
@@ -82,24 +111,15 @@ rx_desc_to_ol_flags(struct rte_mbuf *m, const union fm10k_rx_desc *d)
                (FM10K_RXD_STATUS_IPCS | FM10K_RXD_STATUS_IPE)) ==
                (FM10K_RXD_STATUS_IPCS | FM10K_RXD_STATUS_IPE)))
                m->ol_flags |= PKT_RX_IP_CKSUM_BAD;
+       else
+               m->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
 
        if (unlikely((d->d.staterr &
                (FM10K_RXD_STATUS_L4CS | FM10K_RXD_STATUS_L4E)) ==
                (FM10K_RXD_STATUS_L4CS | FM10K_RXD_STATUS_L4E)))
                m->ol_flags |= PKT_RX_L4_CKSUM_BAD;
-
-       if (d->d.staterr & FM10K_RXD_STATUS_VEXT)
-               m->ol_flags |= PKT_RX_VLAN_PKT;
-
-       if (unlikely(d->d.staterr & FM10K_RXD_STATUS_HBO))
-               m->ol_flags |= PKT_RX_HBUF_OVERFLOW;
-
-       if (unlikely(d->d.staterr & FM10K_RXD_STATUS_RXE))
-               m->ol_flags |= PKT_RX_RECIP_ERR;
-
-       ptype = (d->d.data & FM10K_RXD_PKTTYPE_MASK_L3) >>
-                                               FM10K_RXD_PKTTYPE_SHIFT;
-       m->ol_flags |= pt_lut[(uint8_t)ptype];
+       else
+               m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 }
 
 uint16_t
@@ -118,10 +138,10 @@ fm10k_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 
        nb_pkts = RTE_MIN(nb_pkts, q->alloc_thresh);
        for (count = 0; count < nb_pkts; ++count) {
+               if (!(q->hw_ring[next_dd].d.staterr & FM10K_RXD_STATUS_DD))
+                       break;
                mbuf = q->sw_ring[next_dd];
                desc = q->hw_ring[next_dd];
-               if (!(desc.d.staterr & FM10K_RXD_STATUS_DD))
-                       break;
 #ifdef RTE_LIBRTE_FM10K_DEBUG_RX
                dump_rxd(&desc);
 #endif
@@ -134,6 +154,21 @@ fm10k_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 #endif
 
                mbuf->hash.rss = desc.d.rss;
+               /**
+                * Packets in fm10k device always carry at least one VLAN tag.
+                * For those packets coming in without VLAN tag,
+                * the port default VLAN tag will be used.
+                * So, always PKT_RX_VLAN_PKT flag is set and vlan_tci
+                * is valid for each RX packet's mbuf.
+                */
+               mbuf->ol_flags |= PKT_RX_VLAN_PKT;
+               mbuf->vlan_tci = desc.w.vlan;
+               /**
+                * mbuf->vlan_tci_outer is an idle field in fm10k driver,
+                * so it can be selected to store sglort value.
+                */
+               if (q->rx_ftag_en)
+                       mbuf->vlan_tci_outer = rte_le_to_cpu_16(desc.w.sglort);
 
                rx_pkts[count] = mbuf;
                if (++next_dd == q->nb_desc) {
@@ -217,10 +252,10 @@ fm10k_recv_scattered_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 
        nb_seg = RTE_MIN(nb_pkts, q->alloc_thresh);
        for (count = 0; count < nb_seg; count++) {
+               if (!(q->hw_ring[next_dd].d.staterr & FM10K_RXD_STATUS_DD))
+                       break;
                mbuf = q->sw_ring[next_dd];
                desc = q->hw_ring[next_dd];
-               if (!(desc.d.staterr & FM10K_RXD_STATUS_DD))
-                       break;
 #ifdef RTE_LIBRTE_FM10K_DEBUG_RX
                dump_rxd(&desc);
 #endif
@@ -280,6 +315,22 @@ fm10k_recv_scattered_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
                rx_desc_to_ol_flags(first_seg, &desc);
 #endif
                first_seg->hash.rss = desc.d.rss;
+               /**
+                * Packets in fm10k device always carry at least one VLAN tag.
+                * For those packets coming in without VLAN tag,
+                * the port default VLAN tag will be used.
+                * So, always PKT_RX_VLAN_PKT flag is set and vlan_tci
+                * is valid for each RX packet's mbuf.
+                */
+               first_seg->ol_flags |= PKT_RX_VLAN_PKT;
+               first_seg->vlan_tci = desc.w.vlan;
+               /**
+                * mbuf->vlan_tci_outer is an idle field in fm10k driver,
+                * so it can be selected to store sglort value.
+                */
+               if (q->rx_ftag_en)
+                       first_seg->vlan_tci_outer =
+                               rte_le_to_cpu_16(desc.w.sglort);
 
                /* Prefetch data of first segment, if configured to do so. */
                rte_packet_prefetch((char *)first_seg->buf_addr +
@@ -342,6 +393,76 @@ fm10k_recv_scattered_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
        return nb_rcv;
 }
 
+int
+fm10k_dev_rx_descriptor_done(void *rx_queue, uint16_t offset)
+{
+       volatile union fm10k_rx_desc *rxdp;
+       struct fm10k_rx_queue *rxq = rx_queue;
+       uint16_t desc;
+       int ret;
+
+       if (unlikely(offset >= rxq->nb_desc)) {
+               PMD_DRV_LOG(ERR, "Invalid RX descriptor offset %u", offset);
+               return 0;
+       }
+
+       desc = rxq->next_dd + offset;
+       if (desc >= rxq->nb_desc)
+               desc -= rxq->nb_desc;
+
+       rxdp = &rxq->hw_ring[desc];
+
+       ret = !!(rxdp->w.status &
+                       rte_cpu_to_le_16(FM10K_RXD_STATUS_DD));
+
+       return ret;
+}
+
+/*
+ * Free multiple TX mbuf at a time if they are in the same pool
+ *
+ * @txep: software desc ring index that starts to free
+ * @num: number of descs to free
+ *
+ */
+static inline void tx_free_bulk_mbuf(struct rte_mbuf **txep, int num)
+{
+       struct rte_mbuf *m, *free[RTE_FM10K_TX_MAX_FREE_BUF_SZ];
+       int i;
+       int nb_free = 0;
+
+       if (unlikely(num == 0))
+               return;
+
+       m = rte_pktmbuf_prefree_seg(txep[0]);
+       if (likely(m != NULL)) {
+               free[0] = m;
+               nb_free = 1;
+               for (i = 1; i < num; i++) {
+                       m = rte_pktmbuf_prefree_seg(txep[i]);
+                       if (likely(m != NULL)) {
+                               if (likely(m->pool == free[0]->pool))
+                                       free[nb_free++] = m;
+                               else {
+                                       rte_mempool_put_bulk(free[0]->pool,
+                                                       (void *)free, nb_free);
+                                       free[0] = m;
+                                       nb_free = 1;
+                               }
+                       }
+                       txep[i] = NULL;
+               }
+               rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
+       } else {
+               for (i = 1; i < num; i++) {
+                       m = rte_pktmbuf_prefree_seg(txep[i]);
+                       if (m != NULL)
+                               rte_mempool_put(m->pool, m);
+                       txep[i] = NULL;
+               }
+       }
+}
+
 static inline void tx_free_descriptors(struct fm10k_tx_queue *q)
 {
        uint16_t next_rs, count = 0;
@@ -358,11 +479,7 @@ static inline void tx_free_descriptors(struct fm10k_tx_queue *q)
         * including nb_desc */
        if (q->last_free > next_rs) {
                count = q->nb_desc - q->last_free;
-               while (q->last_free < q->nb_desc) {
-                       rte_pktmbuf_free_seg(q->sw_ring[q->last_free]);
-                       q->sw_ring[q->last_free] = NULL;
-                       ++q->last_free;
-               }
+               tx_free_bulk_mbuf(&q->sw_ring[q->last_free], count);
                q->last_free = 0;
        }
 
@@ -370,10 +487,10 @@ static inline void tx_free_descriptors(struct fm10k_tx_queue *q)
        q->nb_free += count + (next_rs + 1 - q->last_free);
 
        /* free buffers from last_free, up to and including next_rs */
-       while (q->last_free <= next_rs) {
-               rte_pktmbuf_free_seg(q->sw_ring[q->last_free]);
-               q->sw_ring[q->last_free] = NULL;
-               ++q->last_free;
+       if (q->last_free <= next_rs) {
+               count = next_rs - q->last_free + 1;
+               tx_free_bulk_mbuf(&q->sw_ring[q->last_free], count);
+               q->last_free += count;
        }
 
        if (q->last_free == q->nb_desc)
@@ -383,7 +500,7 @@ static inline void tx_free_descriptors(struct fm10k_tx_queue *q)
 static inline void tx_xmit_pkt(struct fm10k_tx_queue *q, struct rte_mbuf *mb)
 {
        uint16_t last_id;
-       uint8_t flags;
+       uint8_t flags, hdrlen;
 
        /* always set the LAST flag on the last descriptor used to
         * transmit the packet */
@@ -405,10 +522,12 @@ static inline void tx_xmit_pkt(struct fm10k_tx_queue *q, struct rte_mbuf *mb)
        q->nb_free -= mb->nb_segs;
 
        q->hw_ring[q->next_free].flags = 0;
+       if (q->tx_ftag_en)
+               q->hw_ring[q->next_free].flags |= FM10K_TXD_FLAG_FTAG;
        /* set checksum flags on first descriptor of packet. SCTP checksum
         * offload is not supported, but we do not explicitly check for this
         * case in favor of greatly simplified processing. */
-       if (mb->ol_flags & (PKT_TX_IP_CKSUM | PKT_TX_L4_MASK))
+       if (mb->ol_flags & (PKT_TX_IP_CKSUM | PKT_TX_L4_MASK | PKT_TX_TCP_SEG))
                q->hw_ring[q->next_free].flags |= FM10K_TXD_FLAG_CSUM;
 
        /* set vlan if requested */
@@ -420,6 +539,21 @@ static inline void tx_xmit_pkt(struct fm10k_tx_queue *q, struct rte_mbuf *mb)
                        rte_cpu_to_le_64(MBUF_DMA_ADDR(mb));
        q->hw_ring[q->next_free].buflen =
                        rte_cpu_to_le_16(rte_pktmbuf_data_len(mb));
+
+       if (mb->ol_flags & PKT_TX_TCP_SEG) {
+               hdrlen = mb->outer_l2_len + mb->outer_l3_len + mb->l2_len +
+                       mb->l3_len + mb->l4_len;
+               if (q->hw_ring[q->next_free].flags & FM10K_TXD_FLAG_FTAG)
+                       hdrlen += sizeof(struct fm10k_ftag);
+
+               if (likely((hdrlen >= FM10K_TSO_MIN_HEADERLEN) &&
+                               (hdrlen <= FM10K_TSO_MAX_HEADERLEN) &&
+                               (mb->tso_segsz >= FM10K_TSO_MINMSS))) {
+                       q->hw_ring[q->next_free].mss = mb->tso_segsz;
+                       q->hw_ring[q->next_free].hdrlen = hdrlen;
+               }
+       }
+
        if (++q->next_free == q->nb_desc)
                q->next_free = 0;
 
@@ -435,7 +569,7 @@ static inline void tx_xmit_pkt(struct fm10k_tx_queue *q, struct rte_mbuf *mb)
                        q->next_free = 0;
        }
 
-       q->hw_ring[last_id].flags = flags;
+       q->hw_ring[last_id].flags |= flags;
 }
 
 uint16_t
@@ -473,3 +607,41 @@ fm10k_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 
        return count;
 }
+
+uint16_t
+fm10k_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
+               uint16_t nb_pkts)
+{
+       int i, ret;
+       struct rte_mbuf *m;
+
+       for (i = 0; i < nb_pkts; i++) {
+               m = tx_pkts[i];
+
+               if ((m->ol_flags & PKT_TX_TCP_SEG) &&
+                               (m->tso_segsz < FM10K_TSO_MINMSS)) {
+                       rte_errno = -EINVAL;
+                       return i;
+               }
+
+               if (m->ol_flags & FM10K_TX_OFFLOAD_NOTSUP_MASK) {
+                       rte_errno = -ENOTSUP;
+                       return i;
+               }
+
+#ifdef RTE_LIBRTE_ETHDEV_DEBUG
+               ret = rte_validate_tx_offload(m);
+               if (ret != 0) {
+                       rte_errno = ret;
+                       return i;
+               }
+#endif
+               ret = rte_net_intel_cksum_prepare(m);
+               if (ret != 0) {
+                       rte_errno = ret;
+                       return i;
+               }
+       }
+
+       return i;
+}