enic: improve Rx performance
authorJohn Daley <johndale@cisco.com>
Fri, 4 Mar 2016 21:09:00 +0000 (13:09 -0800)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Wed, 16 Mar 2016 15:57:12 +0000 (16:57 +0100)
This is a wholesale replacement of the Enic PMD receive path in order
to improve performance and code clarity. The changes are:
- Simplify and reduce code path length of receive function.
- Put most of the fast-path receive functions in one file.
- Reduce the number of posted_index updates (pay attention to
  rx_free_thresh)
- Remove the unneeded container structure around the RQ mbuf ring
- Prefetch next Mbuf and descriptors while processing the current one
- Use a lookup table for converting CQ flags to mbuf flags.

Signed-off-by: John Daley <johndale@cisco.com>
drivers/net/enic/Makefile
drivers/net/enic/base/vnic_rq.c
drivers/net/enic/base/vnic_rq.h
drivers/net/enic/enic.h
drivers/net/enic/enic_ethdev.c
drivers/net/enic/enic_main.c
drivers/net/enic/enic_res.h
drivers/net/enic/enic_rx.c [new file with mode: 0644]

index f0ee093..f316274 100644 (file)
@@ -53,6 +53,7 @@ VPATH += $(SRCDIR)/src
 #
 SRCS-$(CONFIG_RTE_LIBRTE_ENIC_PMD) += enic_ethdev.c
 SRCS-$(CONFIG_RTE_LIBRTE_ENIC_PMD) += enic_main.c
+SRCS-$(CONFIG_RTE_LIBRTE_ENIC_PMD) += enic_rx.c
 SRCS-$(CONFIG_RTE_LIBRTE_ENIC_PMD) += enic_clsf.c
 SRCS-$(CONFIG_RTE_LIBRTE_ENIC_PMD) += enic_res.c
 SRCS-$(CONFIG_RTE_LIBRTE_ENIC_PMD) += base/vnic_cq.c
index 1441604..cb62c5e 100644 (file)
 #include "vnic_dev.h"
 #include "vnic_rq.h"
 
-static int vnic_rq_alloc_bufs(struct vnic_rq *rq)
-{
-       struct vnic_rq_buf *buf;
-       unsigned int i, j, count = rq->ring.desc_count;
-       unsigned int blks = VNIC_RQ_BUF_BLKS_NEEDED(count);
-
-       for (i = 0; i < blks; i++) {
-               rq->bufs[i] = kzalloc(VNIC_RQ_BUF_BLK_SZ(count), GFP_ATOMIC);
-               if (!rq->bufs[i])
-                       return -ENOMEM;
-       }
-
-       for (i = 0; i < blks; i++) {
-               buf = rq->bufs[i];
-               for (j = 0; j < VNIC_RQ_BUF_BLK_ENTRIES(count); j++) {
-                       buf->index = i * VNIC_RQ_BUF_BLK_ENTRIES(count) + j;
-                       buf->desc = (u8 *)rq->ring.descs +
-                               rq->ring.desc_size * buf->index;
-                       if (buf->index + 1 == count) {
-                               buf->next = rq->bufs[0];
-                               break;
-                       } else if (j + 1 == VNIC_RQ_BUF_BLK_ENTRIES(count)) {
-                               buf->next = rq->bufs[i + 1];
-                       } else {
-                               buf->next = buf + 1;
-                               buf++;
-                       }
-               }
-       }
-
-       rq->to_use = rq->to_clean = rq->bufs[0];
-
-       return 0;
-}
-
-int vnic_rq_mem_size(struct vnic_rq *rq, unsigned int desc_count,
-       unsigned int desc_size)
-{
-       int mem_size = 0;
-
-       mem_size += vnic_dev_desc_ring_size(&rq->ring, desc_count, desc_size);
-
-       mem_size += VNIC_RQ_BUF_BLKS_NEEDED(rq->ring.desc_count) *
-               VNIC_RQ_BUF_BLK_SZ(rq->ring.desc_count);
-
-       return mem_size;
-}
-
 void vnic_rq_free(struct vnic_rq *rq)
 {
        struct vnic_dev *vdev;
-       unsigned int i;
 
        vdev = rq->vdev;
 
        vnic_dev_free_desc_ring(vdev, &rq->ring);
 
-       for (i = 0; i < VNIC_RQ_BUF_BLKS_MAX; i++) {
-               if (rq->bufs[i]) {
-                       kfree(rq->bufs[i]);
-                       rq->bufs[i] = NULL;
-               }
-       }
-
        rq->ctrl = NULL;
 }
 
 int vnic_rq_alloc(struct vnic_dev *vdev, struct vnic_rq *rq, unsigned int index,
        unsigned int desc_count, unsigned int desc_size)
 {
-       int err;
+       int rc;
        char res_name[NAME_MAX];
        static int instance;
 
@@ -121,18 +65,9 @@ int vnic_rq_alloc(struct vnic_dev *vdev, struct vnic_rq *rq, unsigned int index,
        vnic_rq_disable(rq);
 
        snprintf(res_name, sizeof(res_name), "%d-rq-%d", instance++, index);
-       err = vnic_dev_alloc_desc_ring(vdev, &rq->ring, desc_count, desc_size,
+       rc = vnic_dev_alloc_desc_ring(vdev, &rq->ring, desc_count, desc_size,
                rq->socket_id, res_name);
-       if (err)
-               return err;
-
-       err = vnic_rq_alloc_bufs(rq);
-       if (err) {
-               vnic_rq_free(rq);
-               return err;
-       }
-
-       return 0;
+       return rc;
 }
 
 void vnic_rq_init_start(struct vnic_rq *rq, unsigned int cq_index,
@@ -154,9 +89,6 @@ void vnic_rq_init_start(struct vnic_rq *rq, unsigned int cq_index,
        iowrite32(fetch_index, &rq->ctrl->fetch_index);
        iowrite32(posted_index, &rq->ctrl->posted_index);
 
-       rq->to_use = rq->to_clean =
-               &rq->bufs[fetch_index / VNIC_RQ_BUF_BLK_ENTRIES(count)]
-                       [fetch_index % VNIC_RQ_BUF_BLK_ENTRIES(count)];
 }
 
 void vnic_rq_init(struct vnic_rq *rq, unsigned int cq_index,
@@ -176,6 +108,8 @@ void vnic_rq_init(struct vnic_rq *rq, unsigned int cq_index,
                fetch_index, fetch_index,
                error_interrupt_enable,
                error_interrupt_offset);
+       rq->rxst_idx = 0;
+       rq->tot_pkts = 0;
 }
 
 void vnic_rq_error_out(struct vnic_rq *rq, unsigned int error)
@@ -212,21 +146,20 @@ int vnic_rq_disable(struct vnic_rq *rq)
 }
 
 void vnic_rq_clean(struct vnic_rq *rq,
-       void (*buf_clean)(struct vnic_rq *rq, struct vnic_rq_buf *buf))
+       void (*buf_clean)(struct rte_mbuf **buf))
 {
-       struct vnic_rq_buf *buf;
-       u32 fetch_index;
+       struct rte_mbuf **buf;
+       u32 fetch_index, i;
        unsigned int count = rq->ring.desc_count;
 
-       buf = rq->to_clean;
-
-       while (vnic_rq_desc_used(rq) > 0) {
+       buf = &rq->mbuf_ring[0];
 
-               (*buf_clean)(rq, buf);
-
-               buf = rq->to_clean = buf->next;
-               rq->ring.desc_avail++;
+       for (i = 0; i < count; i++) {
+               (*buf_clean)(buf);
+               buf++;
        }
+       rq->ring.desc_avail = count - 1;
+       rq->rx_nb_hold = 0;
 
        /* Use current fetch_index as the ring starting point */
        fetch_index = ioread32(&rq->ctrl->fetch_index);
@@ -235,9 +168,7 @@ void vnic_rq_clean(struct vnic_rq *rq,
                /* Hardware surprise removal: reset fetch_index */
                fetch_index = 0;
        }
-       rq->to_use = rq->to_clean =
-               &rq->bufs[fetch_index / VNIC_RQ_BUF_BLK_ENTRIES(count)]
-                       [fetch_index % VNIC_RQ_BUF_BLK_ENTRIES(count)];
+
        iowrite32(fetch_index, &rq->ctrl->posted_index);
 
        vnic_dev_clear_desc_ring(&rq->ring);
index 0f5c3c1..e083ccc 100644 (file)
@@ -66,42 +66,22 @@ struct vnic_rq_ctrl {
        u32 pad10;
 };
 
-/* Break the vnic_rq_buf allocations into blocks of 32/64 entries */
-#define VNIC_RQ_BUF_MIN_BLK_ENTRIES 32
-#define VNIC_RQ_BUF_DFLT_BLK_ENTRIES 64
-#define VNIC_RQ_BUF_BLK_ENTRIES(entries) \
-       ((unsigned int)((entries < VNIC_RQ_BUF_DFLT_BLK_ENTRIES) ? \
-       VNIC_RQ_BUF_MIN_BLK_ENTRIES : VNIC_RQ_BUF_DFLT_BLK_ENTRIES))
-#define VNIC_RQ_BUF_BLK_SZ(entries) \
-       (VNIC_RQ_BUF_BLK_ENTRIES(entries) * sizeof(struct vnic_rq_buf))
-#define VNIC_RQ_BUF_BLKS_NEEDED(entries) \
-       DIV_ROUND_UP(entries, VNIC_RQ_BUF_BLK_ENTRIES(entries))
-#define VNIC_RQ_BUF_BLKS_MAX VNIC_RQ_BUF_BLKS_NEEDED(4096)
-
-struct vnic_rq_buf {
-       struct vnic_rq_buf *next;
-       dma_addr_t dma_addr;
-       void *os_buf;
-       unsigned int os_buf_index;
-       unsigned int len;
-       unsigned int index;
-       void *desc;
-       uint64_t wr_id;
-};
-
 struct vnic_rq {
        unsigned int index;
+       unsigned int posted_index;
        struct vnic_dev *vdev;
-       struct vnic_rq_ctrl __iomem *ctrl;              /* memory-mapped */
+       struct vnic_rq_ctrl __iomem *ctrl;      /* memory-mapped */
        struct vnic_dev_ring ring;
-       struct vnic_rq_buf *bufs[VNIC_RQ_BUF_BLKS_MAX];
-       struct vnic_rq_buf *to_use;
-       struct vnic_rq_buf *to_clean;
+       struct rte_mbuf **mbuf_ring;            /* array of allocated mbufs */
+       unsigned int mbuf_next_idx;             /* next mb to consume */
        void *os_buf_head;
        unsigned int pkts_outstanding;
-
+       uint16_t rx_nb_hold;
+       uint16_t rx_free_thresh;
        unsigned int socket_id;
        struct rte_mempool *mp;
+       uint16_t rxst_idx;
+       uint32_t tot_pkts;
 };
 
 static inline unsigned int vnic_rq_desc_avail(struct vnic_rq *rq)
@@ -116,119 +96,13 @@ static inline unsigned int vnic_rq_desc_used(struct vnic_rq *rq)
        return rq->ring.desc_count - rq->ring.desc_avail - 1;
 }
 
-static inline void *vnic_rq_next_desc(struct vnic_rq *rq)
-{
-       return rq->to_use->desc;
-}
-
-static inline unsigned int vnic_rq_next_index(struct vnic_rq *rq)
-{
-       return rq->to_use->index;
-}
-
-static inline void vnic_rq_post(struct vnic_rq *rq,
-       void *os_buf, unsigned int os_buf_index,
-       dma_addr_t dma_addr, unsigned int len,
-       uint64_t wrid)
-{
-       struct vnic_rq_buf *buf = rq->to_use;
-
-       buf->os_buf = os_buf;
-       buf->os_buf_index = os_buf_index;
-       buf->dma_addr = dma_addr;
-       buf->len = len;
-       buf->wr_id = wrid;
-
-       buf = buf->next;
-       rq->to_use = buf;
-       rq->ring.desc_avail--;
-
-       /* Move the posted_index every nth descriptor
-        */
-
-#ifndef VNIC_RQ_RETURN_RATE
-#define VNIC_RQ_RETURN_RATE            0xf     /* keep 2^n - 1 */
-#endif
-
-       if ((buf->index & VNIC_RQ_RETURN_RATE) == 0) {
-               /* Adding write memory barrier prevents compiler and/or CPU
-                * reordering, thus avoiding descriptor posting before
-                * descriptor is initialized. Otherwise, hardware can read
-                * stale descriptor fields.
-                */
-               wmb();
-               iowrite32(buf->index, &rq->ctrl->posted_index);
-       }
-}
-
-static inline void vnic_rq_post_commit(struct vnic_rq *rq,
-       void *os_buf, unsigned int os_buf_index,
-       dma_addr_t dma_addr, unsigned int len)
-{
-       struct vnic_rq_buf *buf = rq->to_use;
-
-       buf->os_buf = os_buf;
-       buf->os_buf_index = os_buf_index;
-       buf->dma_addr = dma_addr;
-       buf->len = len;
-
-       buf = buf->next;
-       rq->to_use = buf;
-       rq->ring.desc_avail--;
-
-       /* Move the posted_index every descriptor
-        */
-
-       /* Adding write memory barrier prevents compiler and/or CPU
-        * reordering, thus avoiding descriptor posting before
-        * descriptor is initialized. Otherwise, hardware can read
-        * stale descriptor fields.
-        */
-       wmb();
-       iowrite32(buf->index, &rq->ctrl->posted_index);
-}
 
-static inline void vnic_rq_return_descs(struct vnic_rq *rq, unsigned int count)
-{
-       rq->ring.desc_avail += count;
-}
 
 enum desc_return_options {
        VNIC_RQ_RETURN_DESC,
        VNIC_RQ_DEFER_RETURN_DESC,
 };
 
-static inline int vnic_rq_service(struct vnic_rq *rq,
-       struct cq_desc *cq_desc, u16 completed_index,
-       int desc_return, int (*buf_service)(struct vnic_rq *rq,
-       struct cq_desc *cq_desc, struct vnic_rq_buf *buf,
-       int skipped, void *opaque), void *opaque)
-{
-       struct vnic_rq_buf *buf;
-       int skipped;
-       int eop = 0;
-
-       buf = rq->to_clean;
-       while (1) {
-
-               skipped = (buf->index != completed_index);
-
-               if ((*buf_service)(rq, cq_desc, buf, skipped, opaque))
-                       eop++;
-
-               if (desc_return == VNIC_RQ_RETURN_DESC)
-                       rq->ring.desc_avail++;
-
-               rq->to_clean = buf->next;
-
-               if (!skipped)
-                       break;
-
-               buf = rq->to_clean;
-       }
-       return eop;
-}
-
 static inline int vnic_rq_fill(struct vnic_rq *rq,
        int (*buf_fill)(struct vnic_rq *rq))
 {
@@ -274,8 +148,5 @@ unsigned int vnic_rq_error_status(struct vnic_rq *rq);
 void vnic_rq_enable(struct vnic_rq *rq);
 int vnic_rq_disable(struct vnic_rq *rq);
 void vnic_rq_clean(struct vnic_rq *rq,
-       void (*buf_clean)(struct vnic_rq *rq, struct vnic_rq_buf *buf));
-int vnic_rq_mem_size(struct vnic_rq *rq, unsigned int desc_count,
-       unsigned int desc_size);
-
+       void (*buf_clean)(struct rte_mbuf **buf));
 #endif /* _VNIC_RQ_H_ */
index 9e78305..8c914f5 100644 (file)
@@ -45,6 +45,7 @@
 #include "vnic_nic.h"
 #include "vnic_rss.h"
 #include "enic_res.h"
+#include "cq_enet_desc.h"
 
 #define DRV_NAME               "enic_pmd"
 #define DRV_DESCRIPTION                "Cisco VIC Ethernet NIC Poll-mode Driver"
@@ -154,6 +155,16 @@ static inline struct enic *pmd_priv(struct rte_eth_dev *eth_dev)
        return (struct enic *)eth_dev->data->dev_private;
 }
 
+#define RTE_LIBRTE_ENIC_ASSERT_ENABLE
+#ifdef RTE_LIBRTE_ENIC_ASSERT_ENABLE
+#define ASSERT(x) do {                 \
+       if (!(x))                       \
+               rte_panic("ENIC: x");   \
+} while (0)
+#else
+#define ASSERT(x)
+#endif
+
 extern void enic_fdir_stats_get(struct enic *enic,
        struct rte_eth_fdir_stats *stats);
 extern int enic_fdir_add_fltr(struct enic *enic,
@@ -193,9 +204,10 @@ extern void enic_send_pkt(struct enic *enic, struct vnic_wq *wq,
                          uint16_t ol_flags, uint16_t vlan_tag);
 
 extern void enic_post_wq_index(struct vnic_wq *wq);
-extern int enic_poll(struct vnic_rq *rq, struct rte_mbuf **rx_pkts,
-       unsigned int budget, unsigned int *work_done);
 extern int enic_probe(struct enic *enic);
 extern int enic_clsf_init(struct enic *enic);
 extern void enic_clsf_destroy(struct enic *enic);
+uint16_t enic_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
+                       uint16_t nb_pkts);
+
 #endif /* _ENIC_H_ */
index 2a88043..6f2ada5 100644 (file)
@@ -255,7 +255,7 @@ static int enicpmd_dev_rx_queue_setup(struct rte_eth_dev *eth_dev,
        uint16_t queue_idx,
        uint16_t nb_desc,
        unsigned int socket_id,
-       __rte_unused const struct rte_eth_rxconf *rx_conf,
+       const struct rte_eth_rxconf *rx_conf,
        struct rte_mempool *mp)
 {
        int ret;
@@ -270,6 +270,10 @@ static int enicpmd_dev_rx_queue_setup(struct rte_eth_dev *eth_dev,
                return ret;
        }
 
+       enic->rq[queue_idx].rx_free_thresh = rx_conf->rx_free_thresh;
+       dev_debug(enic, "Set queue_id:%u free thresh:%u\n", queue_idx,
+                       enic->rq[queue_idx].rx_free_thresh);
+
        return enicpmd_dev_setup_intr(enic);
 }
 
@@ -429,6 +433,9 @@ static void enicpmd_dev_info_get(struct rte_eth_dev *eth_dev,
                DEV_TX_OFFLOAD_IPV4_CKSUM  |
                DEV_TX_OFFLOAD_UDP_CKSUM   |
                DEV_TX_OFFLOAD_TCP_CKSUM;
+       device_info->default_rxconf = (struct rte_eth_rxconf) {
+               .rx_free_thresh = ENIC_DEFAULT_RX_FREE_THRESH
+       };
 }
 
 static void enicpmd_dev_promiscuous_enable(struct rte_eth_dev *eth_dev)
@@ -538,18 +545,6 @@ static uint16_t enicpmd_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
        return index;
 }
 
-static uint16_t enicpmd_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
-       uint16_t nb_pkts)
-{
-       struct vnic_rq *rq = (struct vnic_rq *)rx_queue;
-       unsigned int work_done;
-
-       if (enic_poll(rq, rx_pkts, (unsigned int)nb_pkts, &work_done))
-               dev_err(enic, "error in enicpmd poll\n");
-
-       return work_done;
-}
-
 static const struct eth_dev_ops enicpmd_eth_dev_ops = {
        .dev_configure        = enicpmd_dev_configure,
        .dev_start            = enicpmd_dev_start,
@@ -606,7 +601,7 @@ static int eth_enicpmd_dev_init(struct rte_eth_dev *eth_dev)
        enic->port_id = eth_dev->data->port_id;
        enic->rte_dev = eth_dev;
        eth_dev->dev_ops = &enicpmd_eth_dev_ops;
-       eth_dev->rx_pkt_burst = &enicpmd_recv_pkts;
+       eth_dev->rx_pkt_burst = &enic_recv_pkts;
        eth_dev->tx_pkt_burst = &enicpmd_xmit_pkts;
 
        pdev = eth_dev->pci_dev;
@@ -635,8 +630,8 @@ static struct eth_driver rte_enic_pmd = {
  * Register as the [Poll Mode] Driver of Cisco ENIC device.
  */
 static int
-rte_enic_pmd_init(const char *name __rte_unused,
-       const char *params __rte_unused)
+rte_enic_pmd_init(__rte_unused const char *name,
+        __rte_unused const char *params)
 {
        ENICPMD_FUNC_TRACE();
 
index f818c32..9fff020 100644 (file)
 #include "vnic_nic.h"
 #include "enic_vnic_wq.h"
 
+static inline struct rte_mbuf *
+rte_rxmbuf_alloc(struct rte_mempool *mp)
+{
+       struct rte_mbuf *m;
+
+       m = __rte_mbuf_raw_alloc(mp);
+       __rte_mbuf_sanity_check_raw(m, 0);
+       return m;
+}
+
+
 static inline int enic_is_sriov_vf(struct enic *enic)
 {
        return enic->pdev->id.device_id == PCI_DEVICE_ID_CISCO_VIC_ENET_VF;
@@ -80,16 +91,25 @@ static int is_eth_addr_valid(uint8_t *addr)
        return !is_mcast_addr(addr) && !is_zero_addr(addr);
 }
 
-static inline struct rte_mbuf *
-enic_rxmbuf_alloc(struct rte_mempool *mp)
+static void
+enic_rxmbuf_queue_release(struct enic *enic, struct vnic_rq *rq)
 {
-       struct rte_mbuf *m;
+       uint16_t i;
 
-       m = __rte_mbuf_raw_alloc(mp);
-       __rte_mbuf_sanity_check_raw(m, 0);
-       return m;
+       if (!rq || !rq->mbuf_ring) {
+               dev_debug(enic, "Pointer to rq or mbuf_ring is NULL");
+               return;
+       }
+
+       for (i = 0; i < enic->config.rq_desc_count; i++) {
+               if (rq->mbuf_ring[i]) {
+                       rte_pktmbuf_free_seg(rq->mbuf_ring[i]);
+                       rq->mbuf_ring[i] = NULL;
+               }
+       }
 }
 
+
 void enic_set_hdr_split_size(struct enic *enic, u16 split_hdr_size)
 {
        vnic_set_hdr_split_size(enic->vdev, split_hdr_size);
@@ -262,13 +282,13 @@ void enic_set_mac_address(struct enic *enic, uint8_t *mac_addr)
 }
 
 static void
-enic_free_rq_buf(__rte_unused struct vnic_rq *rq, struct vnic_rq_buf *buf)
+enic_free_rq_buf(struct rte_mbuf **mbuf)
 {
-       if (!buf->os_buf)
+       if (*mbuf == NULL)
                return;
 
-       rte_pktmbuf_free((struct rte_mbuf *)buf->os_buf);
-       buf->os_buf = NULL;
+       rte_pktmbuf_free(*mbuf);
+       mbuf = NULL;
 }
 
 void enic_init_vnic_resources(struct enic *enic)
@@ -314,221 +334,47 @@ void enic_init_vnic_resources(struct enic *enic)
 }
 
 
-static int enic_rq_alloc_buf(struct vnic_rq *rq)
+static int
+enic_alloc_rx_queue_mbufs(struct enic *enic, struct vnic_rq *rq)
 {
-       struct enic *enic = vnic_dev_priv(rq->vdev);
+       struct rte_mbuf *mb;
+       struct rq_enet_desc *rqd = rq->ring.descs;
+       unsigned i;
        dma_addr_t dma_addr;
-       struct rq_enet_desc *desc = vnic_rq_next_desc(rq);
-       uint8_t type = RQ_ENET_TYPE_ONLY_SOP;
-       u16 split_hdr_size = vnic_get_hdr_split_size(enic->vdev);
-       struct rte_mbuf *mbuf = enic_rxmbuf_alloc(rq->mp);
-       struct rte_mbuf *hdr_mbuf = NULL;
-
-       if (!mbuf) {
-               dev_err(enic, "mbuf alloc in enic_rq_alloc_buf failed\n");
-               return -1;
-       }
-
-       if (unlikely(split_hdr_size)) {
-               if (vnic_rq_desc_avail(rq) < 2) {
-                       rte_mempool_put(mbuf->pool, mbuf);
-                       return -1;
-               }
-               hdr_mbuf = enic_rxmbuf_alloc(rq->mp);
-               if (!hdr_mbuf) {
-                       rte_mempool_put(mbuf->pool, mbuf);
-                       dev_err(enic,
-                               "hdr_mbuf alloc in enic_rq_alloc_buf failed\n");
-                       return -1;
-               }
-
-               hdr_mbuf->data_off = RTE_PKTMBUF_HEADROOM;
-
-               hdr_mbuf->nb_segs = 2;
-               hdr_mbuf->port = enic->port_id;
-               hdr_mbuf->next = mbuf;
-
-               dma_addr = (dma_addr_t)
-                   (hdr_mbuf->buf_physaddr + hdr_mbuf->data_off);
-
-               rq_enet_desc_enc(desc, dma_addr, type, split_hdr_size);
 
-               vnic_rq_post(rq, (void *)hdr_mbuf, 0 /*os_buf_index*/, dma_addr,
-                       (unsigned int)split_hdr_size, 0 /*wrid*/);
+       dev_debug(enic, "queue %u, allocating %u rx queue mbufs", rq->index,
+                 rq->ring.desc_count);
 
-               desc = vnic_rq_next_desc(rq);
-               type = RQ_ENET_TYPE_NOT_SOP;
-       } else {
-               mbuf->nb_segs = 1;
-               mbuf->port = enic->port_id;
-       }
-
-       mbuf->data_off = RTE_PKTMBUF_HEADROOM;
-       mbuf->next = NULL;
-
-       dma_addr = (dma_addr_t)
-           (mbuf->buf_physaddr + mbuf->data_off);
-
-       rq_enet_desc_enc(desc, dma_addr, type, mbuf->buf_len);
-
-       vnic_rq_post(rq, (void *)mbuf, 0 /*os_buf_index*/, dma_addr,
-               (unsigned int)mbuf->buf_len, 0 /*wrid*/);
-
-       return 0;
-}
-
-static int enic_rq_indicate_buf(struct vnic_rq *rq,
-       struct cq_desc *cq_desc, struct vnic_rq_buf *buf,
-       int skipped, void *opaque)
-{
-       struct enic *enic = vnic_dev_priv(rq->vdev);
-       struct rte_mbuf **rx_pkt_bucket = (struct rte_mbuf **)opaque;
-       struct rte_mbuf *rx_pkt = NULL;
-       struct rte_mbuf *hdr_rx_pkt = NULL;
-
-       u8 type, color, eop, sop, ingress_port, vlan_stripped;
-       u8 fcoe, fcoe_sof, fcoe_fc_crc_ok, fcoe_enc_error, fcoe_eof;
-       u8 tcp_udp_csum_ok, udp, tcp, ipv4_csum_ok;
-       u8 ipv6, ipv4, ipv4_fragment, fcs_ok, rss_type, csum_not_calc;
-       u8 packet_error;
-       u16 q_number, completed_index, bytes_written, vlan_tci, checksum;
-       u32 rss_hash;
-
-       cq_enet_rq_desc_dec((struct cq_enet_rq_desc *)cq_desc,
-               &type, &color, &q_number, &completed_index,
-               &ingress_port, &fcoe, &eop, &sop, &rss_type,
-               &csum_not_calc, &rss_hash, &bytes_written,
-               &packet_error, &vlan_stripped, &vlan_tci, &checksum,
-               &fcoe_sof, &fcoe_fc_crc_ok, &fcoe_enc_error,
-               &fcoe_eof, &tcp_udp_csum_ok, &udp, &tcp,
-               &ipv4_csum_ok, &ipv6, &ipv4, &ipv4_fragment,
-               &fcs_ok);
-
-       rx_pkt = (struct rte_mbuf *)buf->os_buf;
-       buf->os_buf = NULL;
-
-       if (unlikely(packet_error)) {
-               dev_err(enic, "packet error\n");
-               rx_pkt->data_len = 0;
-               return 0;
-       }
-
-       if (unlikely(skipped)) {
-               rx_pkt->data_len = 0;
-               return 0;
-       }
-
-       if (likely(!vnic_get_hdr_split_size(enic->vdev))) {
-               /* No header split configured */
-               *rx_pkt_bucket = rx_pkt;
-               rx_pkt->pkt_len = bytes_written;
-
-               if (ipv4) {
-                       rx_pkt->packet_type = RTE_PTYPE_L3_IPV4;
-                       if (!csum_not_calc) {
-                               if (unlikely(!ipv4_csum_ok))
-                                       rx_pkt->ol_flags |= PKT_RX_IP_CKSUM_BAD;
-
-                               if ((tcp || udp) && (!tcp_udp_csum_ok))
-                                       rx_pkt->ol_flags |= PKT_RX_L4_CKSUM_BAD;
-                       }
-               } else if (ipv6)
-                       rx_pkt->packet_type = RTE_PTYPE_L3_IPV6;
-       } else {
-               /* Header split */
-               if (sop && !eop) {
-                       /* This piece is header */
-                       *rx_pkt_bucket = rx_pkt;
-                       rx_pkt->pkt_len = bytes_written;
-               } else {
-                       if (sop && eop) {
-                               /* The packet is smaller than split_hdr_size */
-                               *rx_pkt_bucket = rx_pkt;
-                               rx_pkt->pkt_len = bytes_written;
-                               if (ipv4) {
-                                       rx_pkt->packet_type = RTE_PTYPE_L3_IPV4;
-                                       if (!csum_not_calc) {
-                                               if (unlikely(!ipv4_csum_ok))
-                                                       rx_pkt->ol_flags |=
-                                                           PKT_RX_IP_CKSUM_BAD;
-
-                                               if ((tcp || udp) &&
-                                                   (!tcp_udp_csum_ok))
-                                                       rx_pkt->ol_flags |=
-                                                           PKT_RX_L4_CKSUM_BAD;
-                                       }
-                               } else if (ipv6)
-                                       rx_pkt->packet_type = RTE_PTYPE_L3_IPV6;
-                       } else {
-                               /* Payload */
-                               hdr_rx_pkt = *rx_pkt_bucket;
-                               hdr_rx_pkt->pkt_len += bytes_written;
-                               if (ipv4) {
-                                       hdr_rx_pkt->packet_type =
-                                               RTE_PTYPE_L3_IPV4;
-                                       if (!csum_not_calc) {
-                                               if (unlikely(!ipv4_csum_ok))
-                                                       hdr_rx_pkt->ol_flags |=
-                                                           PKT_RX_IP_CKSUM_BAD;
-
-                                               if ((tcp || udp) &&
-                                                   (!tcp_udp_csum_ok))
-                                                       hdr_rx_pkt->ol_flags |=
-                                                           PKT_RX_L4_CKSUM_BAD;
-                                       }
-                               } else if (ipv6)
-                                       hdr_rx_pkt->packet_type =
-                                               RTE_PTYPE_L3_IPV6;
-                       }
+       for (i = 0; i < rq->ring.desc_count; i++, rqd++) {
+               mb = rte_rxmbuf_alloc(rq->mp);
+               if (mb == NULL) {
+                       dev_err(enic, "RX mbuf alloc failed queue_id=%u",
+                       (unsigned)rq->index);
+                       return -ENOMEM;
                }
-       }
 
-       rx_pkt->data_len = bytes_written;
+               dma_addr = (dma_addr_t)(mb->buf_physaddr + mb->data_off);
 
-       if (rss_hash) {
-               rx_pkt->ol_flags |= PKT_RX_RSS_HASH;
-               rx_pkt->hash.rss = rss_hash;
+               rq_enet_desc_enc(rqd, dma_addr, RQ_ENET_TYPE_ONLY_SOP,
+                                mb->buf_len);
+               rq->mbuf_ring[i] = mb;
        }
 
-       if (vlan_tci) {
-               rx_pkt->ol_flags |= PKT_RX_VLAN_PKT;
-               rx_pkt->vlan_tci = vlan_tci;
-       }
+       /* make sure all prior writes are complete before doing the PIO write */
+       rte_rmb();
 
-       return eop;
-}
+       /* Post all but the last 2 cache lines' worth of descriptors */
+       rq->posted_index = rq->ring.desc_count - (2 * RTE_CACHE_LINE_SIZE
+                       / sizeof(struct rq_enet_desc));
+       rq->rx_nb_hold = 0;
 
-static int enic_rq_service(struct vnic_dev *vdev, struct cq_desc *cq_desc,
-       __rte_unused u8 type, u16 q_number, u16 completed_index, void *opaque)
-{
-       struct enic *enic = vnic_dev_priv(vdev);
-
-       return vnic_rq_service(&enic->rq[q_number], cq_desc,
-               completed_index, VNIC_RQ_RETURN_DESC,
-               enic_rq_indicate_buf, opaque);
-
-}
+       dev_debug(enic, "port=%u, qidx=%u, Write %u posted idx, %u sw held\n",
+               enic->port_id, rq->index, rq->posted_index, rq->rx_nb_hold);
+       iowrite32(rq->posted_index, &rq->ctrl->posted_index);
+       rte_rmb();
 
-int enic_poll(struct vnic_rq *rq, struct rte_mbuf **rx_pkts,
-       unsigned int budget, unsigned int *work_done)
-{
-       struct enic *enic = vnic_dev_priv(rq->vdev);
-       unsigned int cq = enic_cq_rq(enic, rq->index);
-       int err = 0;
-
-       *work_done = vnic_cq_service(&enic->cq[cq],
-               budget, enic_rq_service, (void *)rx_pkts);
-
-       if (*work_done) {
-               vnic_rq_fill(rq, enic_rq_alloc_buf);
+       return 0;
 
-               /* Need at least one buffer on ring to get going */
-               if (vnic_rq_desc_used(rq) == 0) {
-                       dev_err(enic, "Unable to alloc receive buffers\n");
-                       err = -1;
-               }
-       }
-       return err;
 }
 
 static void *
@@ -576,6 +422,7 @@ enic_intr_handler(__rte_unused struct rte_intr_handle *handle,
 int enic_enable(struct enic *enic)
 {
        unsigned int index;
+       int err;
        struct rte_eth_dev *eth_dev = enic->rte_dev;
 
        eth_dev->data->dev_link.link_speed = vnic_dev_port_speed(enic->vdev);
@@ -586,15 +433,11 @@ int enic_enable(struct enic *enic)
                dev_warning(enic, "Init of hash table for clsf failed."\
                        "Flow director feature will not work\n");
 
-       /* Fill RQ bufs */
        for (index = 0; index < enic->rq_count; index++) {
-               vnic_rq_fill(&enic->rq[index], enic_rq_alloc_buf);
-
-               /* Need at least one buffer on ring to get going
-               */
-               if (vnic_rq_desc_used(&enic->rq[index]) == 0) {
-                       dev_err(enic, "Unable to alloc receive buffers\n");
-                       return -1;
+               err = enic_alloc_rx_queue_mbufs(enic, &enic->rq[index]);
+               if (err) {
+                       dev_err(enic, "Failed to alloc RX queue mbufs\n");
+                       return err;
                }
        }
 
@@ -636,6 +479,9 @@ void enic_free_rq(void *rxq)
        struct vnic_rq *rq = (struct vnic_rq *)rxq;
        struct enic *enic = vnic_dev_priv(rq->vdev);
 
+       enic_rxmbuf_queue_release(enic, rq);
+       rte_free(rq->mbuf_ring);
+       rq->mbuf_ring = NULL;
        vnic_rq_free(rq);
        vnic_cq_free(&enic->cq[rq->index]);
 }
@@ -664,7 +510,7 @@ int enic_alloc_rq(struct enic *enic, uint16_t queue_idx,
        unsigned int socket_id, struct rte_mempool *mp,
        uint16_t nb_desc)
 {
-       int err;
+       int rc;
        struct vnic_rq *rq = &enic->rq[queue_idx];
 
        rq->socket_id = socket_id;
@@ -687,23 +533,35 @@ int enic_alloc_rq(struct enic *enic, uint16_t queue_idx,
        }
 
        /* Allocate queue resources */
-       err = vnic_rq_alloc(enic->vdev, &enic->rq[queue_idx], queue_idx,
-               enic->config.rq_desc_count,
-               sizeof(struct rq_enet_desc));
-       if (err) {
+       rc = vnic_rq_alloc(enic->vdev, rq, queue_idx,
+               enic->config.rq_desc_count, sizeof(struct rq_enet_desc));
+       if (rc) {
                dev_err(enic, "error in allocation of rq\n");
-               return err;
+               goto err_exit;
        }
 
-       err = vnic_cq_alloc(enic->vdev, &enic->cq[queue_idx], queue_idx,
+       rc = vnic_cq_alloc(enic->vdev, &enic->cq[queue_idx], queue_idx,
                socket_id, enic->config.rq_desc_count,
                sizeof(struct cq_enet_rq_desc));
-       if (err) {
-               vnic_rq_free(rq);
+       if (rc) {
                dev_err(enic, "error in allocation of cq for rq\n");
+               goto err_free_rq_exit;
        }
 
-       return err;
+       /* Allocate the mbuf ring */
+       rq->mbuf_ring = (struct rte_mbuf **)rte_zmalloc_socket("rq->mbuf_ring",
+                       sizeof(struct rte_mbuf *) * enic->config.rq_desc_count,
+                       RTE_CACHE_LINE_SIZE, rq->socket_id);
+
+       if (rq->mbuf_ring != NULL)
+               return 0;
+
+       /* cleanup on error */
+       vnic_cq_free(&enic->cq[queue_idx]);
+err_free_rq_exit:
+       vnic_rq_free(rq);
+err_exit:
+       return -ENOMEM;
 }
 
 void enic_free_wq(void *txq)
@@ -790,6 +648,7 @@ int enic_disable(struct enic *enic)
 
        for (i = 0; i < enic->wq_count; i++)
                vnic_wq_clean(&enic->wq[i], enic_free_wq_buf);
+
        for (i = 0; i < enic->rq_count; i++)
                vnic_rq_clean(&enic->rq[i], enic_free_rq_buf);
        for (i = 0; i < enic->cq_count; i++)
@@ -1074,7 +933,7 @@ int enic_probe(struct enic *enic)
 
        /* Set ingress vlan rewrite mode before vnic initialization */
        err = vnic_dev_set_ig_vlan_rewrite_mode(enic->vdev,
-               IG_VLAN_REWRITE_MODE_PRIORITY_TAG_DEFAULT_VLAN);
+               IG_VLAN_REWRITE_MODE_PASS_THRU);
        if (err) {
                dev_err(enic,
                        "Failed to set ingress vlan rewrite mode, aborting.\n");
index 49f7e22..33f2e84 100644 (file)
@@ -52,6 +52,7 @@
 #define ENIC_UNICAST_PERFECT_FILTERS   32
 
 #define ENIC_NON_TSO_MAX_DESC          16
+#define ENIC_DEFAULT_RX_FREE_THRESH    32
 
 #define ENIC_SETTING(enic, f) ((enic->config.flags & VENETF_##f) ? 1 : 0)
 
@@ -133,21 +134,6 @@ static inline void enic_queue_wq_desc_tso(struct vnic_wq *wq,
                WQ_ENET_OFFLOAD_MODE_TSO,
                eop, 1 /* SOP */, eop, loopback);
 }
-static inline void enic_queue_rq_desc(struct vnic_rq *rq,
-       void *os_buf, unsigned int os_buf_index,
-       dma_addr_t dma_addr, unsigned int len)
-{
-       struct rq_enet_desc *desc = vnic_rq_next_desc(rq);
-       u64 wrid = 0;
-       u8 type = os_buf_index ?
-               RQ_ENET_TYPE_NOT_SOP : RQ_ENET_TYPE_ONLY_SOP;
-
-       rq_enet_desc_enc(desc,
-               (u64)dma_addr | VNIC_PADDR_TARGET,
-               type, (u16)len);
-
-       vnic_rq_post(rq, os_buf, os_buf_index, dma_addr, len, wrid);
-}
 
 struct enic;
 
diff --git a/drivers/net/enic/enic_rx.c b/drivers/net/enic/enic_rx.c
new file mode 100644 (file)
index 0000000..945a60f
--- /dev/null
@@ -0,0 +1,370 @@
+/*
+ * Copyright 2008-2014 Cisco Systems, Inc.  All rights reserved.
+ * Copyright 2007 Nuova Systems, Inc.  All rights reserved.
+ *
+ * Copyright (c) 2014, Cisco Systems, Inc.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in
+ * the documentation and/or other materials provided with the
+ * distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+ * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
+ * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+ * POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+#include <rte_mbuf.h>
+#include <rte_ethdev.h>
+#include <rte_prefetch.h>
+
+#include "enic_compat.h"
+#include "rq_enet_desc.h"
+#include "enic.h"
+
+#define RTE_PMD_USE_PREFETCH
+
+#ifdef RTE_PMD_USE_PREFETCH
+/*
+ * Prefetch a cache line into all cache levels.
+ */
+#define rte_enic_prefetch(p) rte_prefetch0(p)
+#else
+#define rte_enic_prefetch(p) do {} while (0)
+#endif
+
+#ifdef RTE_PMD_PACKET_PREFETCH
+#define rte_packet_prefetch(p) rte_prefetch1(p)
+#else
+#define rte_packet_prefetch(p) do {} while (0)
+#endif
+
+static inline struct rte_mbuf *
+rte_rxmbuf_alloc(struct rte_mempool *mp)
+{
+       struct rte_mbuf *m;
+
+       m = __rte_mbuf_raw_alloc(mp);
+       __rte_mbuf_sanity_check_raw(m, 0);
+       return m;
+}
+
+static inline uint16_t
+enic_cq_rx_desc_ciflags(struct cq_enet_rq_desc *crd)
+{
+       return le16_to_cpu(crd->completed_index_flags) & ~CQ_DESC_COMP_NDX_MASK;
+}
+
+static inline uint16_t
+enic_cq_rx_desc_bwflags(struct cq_enet_rq_desc *crd)
+{
+       return(le16_to_cpu(crd->bytes_written_flags) &
+               ~CQ_ENET_RQ_DESC_BYTES_WRITTEN_MASK);
+}
+
+static inline uint8_t
+enic_cq_rx_desc_packet_error(uint16_t bwflags)
+{
+       return((bwflags & CQ_ENET_RQ_DESC_FLAGS_TRUNCATED) ==
+               CQ_ENET_RQ_DESC_FLAGS_TRUNCATED);
+}
+
+static inline uint8_t
+enic_cq_rx_desc_eop(uint16_t ciflags)
+{
+       return (ciflags & CQ_ENET_RQ_DESC_FLAGS_EOP)
+               == CQ_ENET_RQ_DESC_FLAGS_EOP;
+}
+
+static inline uint8_t
+enic_cq_rx_desc_csum_not_calc(struct cq_enet_rq_desc *cqrd)
+{
+       return ((le16_to_cpu(cqrd->q_number_rss_type_flags) &
+               CQ_ENET_RQ_DESC_FLAGS_CSUM_NOT_CALC) ==
+               CQ_ENET_RQ_DESC_FLAGS_CSUM_NOT_CALC);
+}
+
+static inline uint8_t
+enic_cq_rx_desc_ipv4_csum_ok(struct cq_enet_rq_desc *cqrd)
+{
+       return ((cqrd->flags & CQ_ENET_RQ_DESC_FLAGS_IPV4_CSUM_OK) ==
+               CQ_ENET_RQ_DESC_FLAGS_IPV4_CSUM_OK);
+}
+
+static inline uint8_t
+enic_cq_rx_desc_tcp_udp_csum_ok(struct cq_enet_rq_desc *cqrd)
+{
+       return((cqrd->flags & CQ_ENET_RQ_DESC_FLAGS_TCP_UDP_CSUM_OK) ==
+               CQ_ENET_RQ_DESC_FLAGS_TCP_UDP_CSUM_OK);
+}
+
+static inline uint8_t
+enic_cq_rx_desc_rss_type(struct cq_enet_rq_desc *cqrd)
+{
+       return (uint8_t)((le16_to_cpu(cqrd->q_number_rss_type_flags) >>
+               CQ_DESC_Q_NUM_BITS) & CQ_ENET_RQ_DESC_RSS_TYPE_MASK);
+}
+
+static inline uint32_t
+enic_cq_rx_desc_rss_hash(struct cq_enet_rq_desc *cqrd)
+{
+       return le32_to_cpu(cqrd->rss_hash);
+}
+
+static inline uint8_t
+enic_cq_rx_desc_fcs_ok(struct cq_enet_rq_desc *cqrd)
+{
+       return ((cqrd->flags & CQ_ENET_RQ_DESC_FLAGS_FCS_OK) ==
+               CQ_ENET_RQ_DESC_FLAGS_FCS_OK);
+}
+
+static inline uint16_t
+enic_cq_rx_desc_vlan(struct cq_enet_rq_desc *cqrd)
+{
+       return le16_to_cpu(cqrd->vlan);
+}
+
+static inline uint16_t
+enic_cq_rx_desc_n_bytes(struct cq_desc *cqd)
+{
+       struct cq_enet_rq_desc *cqrd = (struct cq_enet_rq_desc *)cqd;
+       return le16_to_cpu(cqrd->bytes_written_flags) &
+               CQ_ENET_RQ_DESC_BYTES_WRITTEN_MASK;
+}
+
+static inline uint64_t
+enic_cq_rx_to_pkt_err_flags(struct cq_desc *cqd)
+{
+       struct cq_enet_rq_desc *cqrd = (struct cq_enet_rq_desc *)cqd;
+       uint16_t bwflags;
+       uint64_t pkt_err_flags = 0;
+
+       bwflags = enic_cq_rx_desc_bwflags(cqrd);
+
+       /* Check for packet error. Can't be more specific than MAC error */
+       if (enic_cq_rx_desc_packet_error(bwflags)) {
+               pkt_err_flags |= PKT_RX_MAC_ERR;
+       }
+
+       /* Check for bad FCS. MAC error isn't quite, but no other choice */
+       if (!enic_cq_rx_desc_fcs_ok(cqrd)) {
+               pkt_err_flags |= PKT_RX_MAC_ERR;
+       }
+       return pkt_err_flags;
+}
+
+/*
+ * Lookup table to translate RX CQ flags to mbuf flags.
+ */
+static inline uint32_t
+enic_cq_rx_flags_to_pkt_type(struct cq_desc *cqd)
+{
+       struct cq_enet_rq_desc *cqrd = (struct cq_enet_rq_desc *)cqd;
+       uint8_t cqrd_flags = cqrd->flags;
+       static const uint32_t cq_type_table[128] __rte_cache_aligned = {
+               [32] =  RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4,
+               [34] =  RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4
+                       | RTE_PTYPE_L4_UDP,
+               [36] =  RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4
+                       | RTE_PTYPE_L4_TCP,
+               [96] =  RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4
+                       | RTE_PTYPE_L4_FRAG,
+               [16] =  RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6,
+               [18] =  RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6
+                       | RTE_PTYPE_L4_UDP,
+               [20] =  RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6
+                       | RTE_PTYPE_L4_TCP,
+               [80] =  RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6
+                       | RTE_PTYPE_L4_FRAG,
+               /* All others reserved */
+       };
+       cqrd_flags &= CQ_ENET_RQ_DESC_FLAGS_IPV4_FRAGMENT
+               | CQ_ENET_RQ_DESC_FLAGS_IPV4 | CQ_ENET_RQ_DESC_FLAGS_IPV6
+               | CQ_ENET_RQ_DESC_FLAGS_TCP | CQ_ENET_RQ_DESC_FLAGS_UDP;
+       return cq_type_table[cqrd_flags];
+}
+
+static inline void
+enic_cq_rx_to_pkt_flags(struct cq_desc *cqd, struct rte_mbuf *mbuf)
+{
+       struct cq_enet_rq_desc *cqrd = (struct cq_enet_rq_desc *)cqd;
+       uint16_t ciflags, bwflags, pkt_flags = 0;
+       ciflags = enic_cq_rx_desc_ciflags(cqrd);
+       bwflags = enic_cq_rx_desc_bwflags(cqrd);
+
+       ASSERT(mbuf->ol_flags == 0);
+
+       /* flags are meaningless if !EOP */
+       if (unlikely(!enic_cq_rx_desc_eop(ciflags)))
+               goto mbuf_flags_done;
+
+       /* VLAN stripping */
+       if (bwflags & CQ_ENET_RQ_DESC_FLAGS_VLAN_STRIPPED) {
+               pkt_flags |= PKT_RX_VLAN_PKT;
+               mbuf->vlan_tci = enic_cq_rx_desc_vlan(cqrd);
+       } else {
+               mbuf->vlan_tci = 0;
+       }
+
+       /* RSS flag */
+       if (enic_cq_rx_desc_rss_type(cqrd)) {
+               pkt_flags |= PKT_RX_RSS_HASH;
+               mbuf->hash.rss = enic_cq_rx_desc_rss_hash(cqrd);
+       }
+
+       /* checksum flags */
+       if (!enic_cq_rx_desc_csum_not_calc(cqrd) &&
+               (mbuf->packet_type & RTE_PTYPE_L3_IPV4)) {
+               if (unlikely(!enic_cq_rx_desc_ipv4_csum_ok(cqrd)))
+                       pkt_flags |= PKT_RX_IP_CKSUM_BAD;
+               if (mbuf->packet_type & (RTE_PTYPE_L4_UDP | RTE_PTYPE_L4_TCP)) {
+                       if (unlikely(!enic_cq_rx_desc_tcp_udp_csum_ok(cqrd)))
+                               pkt_flags |= PKT_RX_L4_CKSUM_BAD;
+               }
+       }
+
+ mbuf_flags_done:
+       mbuf->ol_flags = pkt_flags;
+}
+
+static inline uint32_t
+enic_ring_add(uint32_t n_descriptors, uint32_t i0, uint32_t i1)
+{
+       uint32_t d = i0 + i1;
+       ASSERT(i0 < n_descriptors);
+       ASSERT(i1 < n_descriptors);
+       d -= (d >= n_descriptors) ? n_descriptors : 0;
+       return d;
+}
+
+
+uint16_t
+enic_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
+              uint16_t nb_pkts)
+{
+       struct vnic_rq *rq = rx_queue;
+       struct enic *enic = vnic_dev_priv(rq->vdev);
+       unsigned int rx_id;
+       struct rte_mbuf *nmb, *rxmb;
+       uint16_t nb_rx = 0;
+       uint16_t nb_hold;
+       struct vnic_cq *cq;
+       volatile struct cq_desc *cqd_ptr;
+       uint8_t color;
+
+       cq = &enic->cq[enic_cq_rq(enic, rq->index)];
+       rx_id = cq->to_clean;           /* index of cqd, rqd, mbuf_table */
+       cqd_ptr = (struct cq_desc *)(cq->ring.descs) + rx_id;
+
+       nb_hold = rq->rx_nb_hold;       /* mbufs held by software */
+
+       while (nb_rx < nb_pkts) {
+               uint16_t rx_pkt_len;
+               volatile struct rq_enet_desc *rqd_ptr;
+               dma_addr_t dma_addr;
+               struct cq_desc cqd;
+               uint64_t ol_err_flags;
+
+               /* Check for pkts available */
+               color = (cqd_ptr->type_color >> CQ_DESC_COLOR_SHIFT)
+                       & CQ_DESC_COLOR_MASK;
+               if (color == cq->last_color)
+                       break;
+
+               /* Get the cq descriptor and rq pointer */
+               cqd = *cqd_ptr;
+               rqd_ptr = (struct rq_enet_desc *)(rq->ring.descs) + rx_id;
+
+               /* allocate a new mbuf */
+               nmb = rte_rxmbuf_alloc(rq->mp);
+               if (nmb == NULL) {
+                       dev_err(enic, "RX mbuf alloc failed port=%u qid=%u",
+                       enic->port_id, (unsigned)rq->index);
+                       rte_eth_devices[enic->port_id].
+                                       data->rx_mbuf_alloc_failed++;
+                       break;
+               }
+
+               /* Check for FCS or packet errors */
+               ol_err_flags = enic_cq_rx_to_pkt_err_flags(&cqd);
+               if (ol_err_flags == 0)
+                       rx_pkt_len = enic_cq_rx_desc_n_bytes(&cqd);
+               else
+                       rx_pkt_len = 0;
+
+               /* Get the mbuf to return and replace with one just allocated */
+               rxmb = rq->mbuf_ring[rx_id];
+               rq->mbuf_ring[rx_id] = nmb;
+
+               /* Increment cqd, rqd, mbuf_table index */
+               rx_id++;
+               if (unlikely(rx_id == rq->ring.desc_count)) {
+                       rx_id = 0;
+                       cq->last_color = cq->last_color ? 0 : 1;
+               }
+
+               /* Prefetch next mbuf & desc while processing current one */
+               cqd_ptr = (struct cq_desc *)(cq->ring.descs) + rx_id;
+               rte_enic_prefetch(cqd_ptr);
+               rte_enic_prefetch(rq->mbuf_ring[rx_id]);
+               rte_enic_prefetch((struct rq_enet_desc *)(rq->ring.descs)
+                                + rx_id);
+
+               /* Push descriptor for newly allocated mbuf */
+               dma_addr = (dma_addr_t)(nmb->buf_physaddr + nmb->data_off);
+               rqd_ptr->address = rte_cpu_to_le_64(dma_addr);
+               rqd_ptr->length_type = cpu_to_le16(nmb->buf_len);
+
+               /* Fill in the rest of the mbuf */
+               rxmb->data_off = RTE_PKTMBUF_HEADROOM;
+               rxmb->nb_segs = 1;
+               rxmb->next = NULL;
+               rxmb->pkt_len = rx_pkt_len;
+               rxmb->data_len = rx_pkt_len;
+               rxmb->port = enic->port_id;
+               rxmb->ol_flags = ol_err_flags;
+               if (!ol_err_flags)
+                       enic_cq_rx_to_pkt_flags(&cqd, rxmb);
+               rxmb->packet_type = enic_cq_rx_flags_to_pkt_type(&cqd);
+
+               /* prefetch mbuf data for caller */
+               rte_packet_prefetch(RTE_PTR_ADD(rxmb->buf_addr,
+                                   RTE_PKTMBUF_HEADROOM));
+
+               /* store the mbuf address into the next entry of the array */
+               rx_pkts[nb_rx++] = rxmb;
+       }
+
+       nb_hold += nb_rx;
+       cq->to_clean = rx_id;
+
+       if (nb_hold > rq->rx_free_thresh) {
+               rq->posted_index = enic_ring_add(rq->ring.desc_count,
+                               rq->posted_index, nb_hold);
+               nb_hold = 0;
+               rte_mb();
+               iowrite32(rq->posted_index, &rq->ctrl->posted_index);
+       }
+
+       rq->rx_nb_hold = nb_hold;
+
+       return nb_rx;
+}