net/virtio: support Tx checksum offload
[dpdk.git] / drivers / net / virtio / virtio_rxtx.c
index a27208e..675dc43 100644 (file)
 #include <rte_string_fns.h>
 #include <rte_errno.h>
 #include <rte_byteorder.h>
+#include <rte_cpuflags.h>
+#include <rte_net.h>
+#include <rte_ip.h>
+#include <rte_udp.h>
+#include <rte_tcp.h>
 
 #include "virtio_logs.h"
 #include "virtio_ethdev.h"
 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
        ETH_TXQ_FLAGS_NOOFFLOADS)
 
-#ifdef RTE_MACHINE_CPUFLAG_SSSE3
-static int use_simple_rxtx;
-#endif
-
 static void
 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
 {
@@ -193,7 +194,8 @@ virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 
        start_dp = vq->vq_ring.desc;
        start_dp[idx].addr =
-               MBUF_DATA_DMA_ADDR(cookie, vq->offset) - hw->vtnet_hdr_size;
+               VIRTIO_MBUF_ADDR(cookie, vq) +
+               RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
        start_dp[idx].len =
                cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
        start_dp[idx].flags =  VRING_DESC_F_WRITE;
@@ -207,18 +209,27 @@ virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
        return 0;
 }
 
+static inline int
+tx_offload_enabled(struct virtio_hw *hw)
+{
+       return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM);
+}
+
 static inline void
 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
                       uint16_t needed, int use_indirect, int can_push)
 {
+       struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
        struct vq_desc_extra *dxp;
        struct virtqueue *vq = txvq->vq;
        struct vring_desc *start_dp;
        uint16_t seg_num = cookie->nb_segs;
        uint16_t head_idx, idx;
        uint16_t head_size = vq->hw->vtnet_hdr_size;
-       unsigned long offs;
+       struct virtio_net_hdr *hdr;
+       int offload;
 
+       offload = tx_offload_enabled(vq->hw);
        head_idx = vq->vq_desc_head_idx;
        idx = head_idx;
        dxp = &vq->vq_descx[idx];
@@ -228,10 +239,12 @@ virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
        start_dp = vq->vq_ring.desc;
 
        if (can_push) {
-               /* put on zero'd transmit header (no offloads) */
-               void *hdr = rte_pktmbuf_prepend(cookie, head_size);
-
-               memset(hdr, 0, head_size);
+               /* prepend cannot fail, checked by caller */
+               hdr = (struct virtio_net_hdr *)
+                       rte_pktmbuf_prepend(cookie, head_size);
+               /* if offload disabled, it is not zeroed below, do it now */
+               if (offload == 0)
+                       memset(hdr, 0, head_size);
        } else if (use_indirect) {
                /* setup tx ring slot to point to indirect
                 * descriptor list stored in reserved region.
@@ -239,14 +252,11 @@ virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
                 * the first slot in indirect ring is already preset
                 * to point to the header in reserved region
                 */
-               struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
-
-               offs = idx * sizeof(struct virtio_tx_region)
-                       + offsetof(struct virtio_tx_region, tx_indir);
-
-               start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
+               start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
+                       RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
                start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
                start_dp[idx].flags = VRING_DESC_F_INDIRECT;
+               hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 
                /* loop below will fill in rest of the indirect elements */
                start_dp = txr[idx].tx_indir;
@@ -255,17 +265,45 @@ virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
                /* setup first tx ring slot to point to header
                 * stored in reserved region.
                 */
-               offs = idx * sizeof(struct virtio_tx_region)
-                       + offsetof(struct virtio_tx_region, tx_hdr);
-
-               start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
+               start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
+                       RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
                start_dp[idx].len   = vq->hw->vtnet_hdr_size;
                start_dp[idx].flags = VRING_DESC_F_NEXT;
+               hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
+
                idx = start_dp[idx].next;
        }
 
+       if (offload) {
+               /* Checksum Offload */
+               switch (cookie->ol_flags & PKT_TX_L4_MASK) {
+               case PKT_TX_UDP_CKSUM:
+                       hdr->csum_start = cookie->l2_len + cookie->l3_len;
+                       hdr->csum_offset = offsetof(struct udp_hdr,
+                               dgram_cksum);
+                       hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
+                       break;
+
+               case PKT_TX_TCP_CKSUM:
+                       hdr->csum_start = cookie->l2_len + cookie->l3_len;
+                       hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
+                       hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
+                       break;
+
+               default:
+                       hdr->csum_start = 0;
+                       hdr->csum_offset = 0;
+                       hdr->flags = 0;
+                       break;
+               }
+
+               hdr->gso_type = 0;
+               hdr->gso_size = 0;
+               hdr->hdr_len = 0;
+       }
+
        do {
-               start_dp[idx].addr  = MBUF_DATA_DMA_ADDR(cookie, vq->offset);
+               start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
                start_dp[idx].len   = cookie->data_len;
                start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
                idx = start_dp[idx].next;
@@ -333,6 +371,7 @@ virtio_dev_rxtx_start(struct rte_eth_dev *dev)
         */
        uint16_t i;
        uint16_t desc_idx;
+       struct virtio_hw *hw = dev->data->dev_private;
 
        PMD_INIT_FUNC_TRACE();
 
@@ -353,8 +392,7 @@ virtio_dev_rxtx_start(struct rte_eth_dev *dev)
                nbufs = 0;
                error = ENOSPC;
 
-#ifdef RTE_MACHINE_CPUFLAG_SSSE3
-               if (use_simple_rxtx) {
+               if (hw->use_simple_rxtx) {
                        for (desc_idx = 0; desc_idx < vq->vq_nentries;
                             desc_idx++) {
                                vq->vq_ring.avail->ring[desc_idx] = desc_idx;
@@ -362,7 +400,7 @@ virtio_dev_rxtx_start(struct rte_eth_dev *dev)
                                        VRING_DESC_F_WRITE;
                        }
                }
-#endif
+
                memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
                for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
                     desc_idx++) {
@@ -378,12 +416,11 @@ virtio_dev_rxtx_start(struct rte_eth_dev *dev)
                        /******************************************
                        *         Enqueue allocated buffers        *
                        *******************************************/
-#ifdef RTE_MACHINE_CPUFLAG_SSSE3
-                       if (use_simple_rxtx)
+                       if (hw->use_simple_rxtx)
                                error = virtqueue_enqueue_recv_refill_simple(vq, m);
                        else
-#endif
                                error = virtqueue_enqueue_recv_refill(vq, m);
+
                        if (error) {
                                rte_pktmbuf_free(m);
                                break;
@@ -404,8 +441,7 @@ virtio_dev_rxtx_start(struct rte_eth_dev *dev)
                struct virtqueue *vq = txvq->vq;
 
                virtio_dev_vring_start(vq);
-#ifdef RTE_MACHINE_CPUFLAG_SSSE3
-               if (use_simple_rxtx) {
+               if (hw->use_simple_rxtx) {
                        uint16_t mid_idx  = vq->vq_nentries >> 1;
 
                        for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
@@ -426,7 +462,7 @@ virtio_dev_rxtx_start(struct rte_eth_dev *dev)
                             desc_idx++)
                                vq->vq_ring.avail->ring[desc_idx] = desc_idx;
                }
-#endif
+
                VIRTQUEUE_DUMP(vq);
        }
 }
@@ -456,9 +492,7 @@ virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 
        dev->data->rx_queues[queue_idx] = rxvq;
 
-#ifdef RTE_MACHINE_CPUFLAG_SSSE3
        virtio_rxq_vec_setup(rxvq);
-#endif
 
        return 0;
 }
@@ -467,17 +501,48 @@ void
 virtio_dev_rx_queue_release(void *rxq)
 {
        struct virtnet_rx *rxvq = rxq;
-       struct virtqueue *vq = rxvq->vq;
-       /* rxvq is freed when vq is freed, and as mz should be freed after the
+       struct virtqueue *vq;
+       const struct rte_memzone *mz;
+
+       if (rxvq == NULL)
+               return;
+
+       /*
+        * rxvq is freed when vq is freed, and as mz should be freed after the
         * del_queue, so we reserve the mz pointer first.
         */
-       const struct rte_memzone *mz = rxvq->mz;
+       vq = rxvq->vq;
+       mz = rxvq->mz;
 
-       /* no need to free rxq as vq and rxq are allocated together */
        virtio_dev_queue_release(vq);
        rte_memzone_free(mz);
 }
 
+static void
+virtio_update_rxtx_handler(struct rte_eth_dev *dev,
+                          const struct rte_eth_txconf *tx_conf)
+{
+       uint8_t use_simple_rxtx = 0;
+       struct virtio_hw *hw = dev->data->dev_private;
+
+#if defined RTE_ARCH_X86
+       if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE3))
+               use_simple_rxtx = 1;
+#elif defined RTE_ARCH_ARM64 || defined CONFIG_RTE_ARCH_ARM
+       if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_NEON))
+               use_simple_rxtx = 1;
+#endif
+       /* Use simple rx/tx func if single segment and no offloads */
+       if (use_simple_rxtx &&
+           (tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
+           !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
+               PMD_INIT_LOG(INFO, "Using simple rx/tx path");
+               dev->tx_pkt_burst = virtio_xmit_pkts_simple;
+               dev->rx_pkt_burst = virtio_recv_pkts_vec;
+               hw->use_simple_rxtx = use_simple_rxtx;
+       }
+}
+
 /*
  * struct rte_eth_dev *dev: Used to update dev
  * uint16_t nb_desc: Defaults to values read from config space
@@ -493,10 +558,6 @@ virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
                        const struct rte_eth_txconf *tx_conf)
 {
        uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
-
-#ifdef RTE_MACHINE_CPUFLAG_SSSE3
-       struct virtio_hw *hw = dev->data->dev_private;
-#endif
        struct virtnet_tx *txvq;
        struct virtqueue *vq;
        uint16_t tx_free_thresh;
@@ -504,22 +565,8 @@ virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 
        PMD_INIT_FUNC_TRACE();
 
-       if ((tx_conf->txq_flags & ETH_TXQ_FLAGS_NOXSUMS)
-           != ETH_TXQ_FLAGS_NOXSUMS) {
-               PMD_INIT_LOG(ERR, "TX checksum offload not supported\n");
-               return -EINVAL;
-       }
 
-#ifdef RTE_MACHINE_CPUFLAG_SSSE3
-       /* Use simple rx/tx func if single segment and no offloads */
-       if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
-            !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
-               PMD_INIT_LOG(INFO, "Using simple rx/tx path");
-               dev->tx_pkt_burst = virtio_xmit_pkts_simple;
-               dev->rx_pkt_burst = virtio_recv_pkts_vec;
-               use_simple_rxtx = 1;
-       }
-#endif
+       virtio_update_rxtx_handler(dev, tx_conf);
 
        ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
                        nb_desc, socket_id, (void **)&txvq);
@@ -553,12 +600,20 @@ void
 virtio_dev_tx_queue_release(void *txq)
 {
        struct virtnet_tx *txvq = txq;
-       struct virtqueue *vq = txvq->vq;
-       /* txvq is freed when vq is freed, and as mz should be freed after the
+       struct virtqueue *vq;
+       const struct rte_memzone *mz;
+       const struct rte_memzone *hdr_mz;
+
+       if (txvq == NULL)
+               return;
+
+       /*
+        * txvq is freed when vq is freed, and as mz should be freed after the
         * del_queue, so we reserve the mz pointer first.
         */
-       const struct rte_memzone *hdr_mz = txvq->virtio_net_hdr_mz;
-       const struct rte_memzone *mz = txvq->mz;
+       vq = txvq->vq;
+       mz = txvq->mz;
+       hdr_mz = txvq->virtio_net_hdr_mz;
 
        virtio_dev_queue_release(vq);
        rte_memzone_free(mz);
@@ -612,6 +667,63 @@ virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
        }
 }
 
+/* Optionally fill offload information in structure */
+static int
+virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
+{
+       struct rte_net_hdr_lens hdr_lens;
+       uint32_t hdrlen, ptype;
+       int l4_supported = 0;
+
+       /* nothing to do */
+       if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
+               return 0;
+
+       m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
+
+       ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
+       m->packet_type = ptype;
+       if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
+           (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
+           (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
+               l4_supported = 1;
+
+       if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
+               hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
+               if (hdr->csum_start <= hdrlen && l4_supported) {
+                       m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
+               } else {
+                       /* Unknown proto or tunnel, do sw cksum. We can assume
+                        * the cksum field is in the first segment since the
+                        * buffers we provided to the host are large enough.
+                        * In case of SCTP, this will be wrong since it's a CRC
+                        * but there's nothing we can do.
+                        */
+                       uint16_t csum, off;
+
+                       rte_raw_cksum_mbuf(m, hdr->csum_start,
+                               rte_pktmbuf_pkt_len(m) - hdr->csum_start,
+                               &csum);
+                       if (likely(csum != 0xffff))
+                               csum = ~csum;
+                       off = hdr->csum_offset + hdr->csum_start;
+                       if (rte_pktmbuf_data_len(m) >= off + 1)
+                               *rte_pktmbuf_mtod_offset(m, uint16_t *,
+                                       off) = csum;
+               }
+       } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
+               m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
+       }
+
+       return 0;
+}
+
+static inline int
+rx_offload_enabled(struct virtio_hw *hw)
+{
+       return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM);
+}
+
 #define VIRTIO_MBUF_BURST_SZ 64
 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 uint16_t
@@ -627,6 +739,8 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
        int error;
        uint32_t i, nb_enqueued;
        uint32_t hdr_size;
+       int offload;
+       struct virtio_net_hdr *hdr;
 
        nb_used = VIRTQUEUE_NUSED(vq);
 
@@ -644,6 +758,7 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
        nb_rx = 0;
        nb_enqueued = 0;
        hdr_size = hw->vtnet_hdr_size;
+       offload = rx_offload_enabled(hw);
 
        for (i = 0; i < num ; i++) {
                rxm = rcv_pkts[i];
@@ -668,9 +783,18 @@ virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
                rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
                rxm->data_len = (uint16_t)(len[i] - hdr_size);
 
+               hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
+                       RTE_PKTMBUF_HEADROOM - hdr_size);
+
                if (hw->vlan_strip)
                        rte_vlan_strip(rxm);
 
+               if (offload && virtio_rx_offload(rxm, hdr) < 0) {
+                       virtio_discard_rxbuf(vq, rxm);
+                       rxvq->stats.errors++;
+                       continue;
+               }
+
                VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 
                rx_pkts[nb_rx++] = rxm;
@@ -730,6 +854,7 @@ virtio_recv_mergeable_pkts(void *rx_queue,
        uint16_t extra_idx;
        uint32_t seg_res;
        uint32_t hdr_size;
+       int offload;
 
        nb_used = VIRTQUEUE_NUSED(vq);
 
@@ -745,6 +870,7 @@ virtio_recv_mergeable_pkts(void *rx_queue,
        extra_idx = 0;
        seg_res = 0;
        hdr_size = hw->vtnet_hdr_size;
+       offload = rx_offload_enabled(hw);
 
        while (i < nb_used) {
                struct virtio_net_hdr_mrg_rxbuf *header;
@@ -790,6 +916,12 @@ virtio_recv_mergeable_pkts(void *rx_queue,
                rx_pkts[nb_rx] = rxm;
                prev = rxm;
 
+               if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
+                       virtio_discard_rxbuf(vq, rxm);
+                       rxvq->stats.errors++;
+                       continue;
+               }
+
                seg_res = seg_num - 1;
 
                while (seg_res != 0) {