summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rwxr-xr-xdpdk/dpdk-16.04_patches/0015-enic-counter_improvement.patch148
-rwxr-xr-xdpdk/dpdk-16.04_patches/0016-enic-scatter-rx.patch650
-rw-r--r--vnet/vnet/devices/dpdk/dpdk_priv.h1
-rw-r--r--vnet/vnet/devices/dpdk/init.c3
4 files changed, 800 insertions, 2 deletions
diff --git a/dpdk/dpdk-16.04_patches/0015-enic-counter_improvement.patch b/dpdk/dpdk-16.04_patches/0015-enic-counter_improvement.patch
new file mode 100755
index 00000000000..6c8986bbb1b
--- /dev/null
+++ b/dpdk/dpdk-16.04_patches/0015-enic-counter_improvement.patch
@@ -0,0 +1,148 @@
+diff -ur dpdk-16.04.orig/drivers/net/enic/enic.h dpdk-16.04/drivers/net/enic/enic.h
+--- dpdk-16.04.orig/drivers/net/enic/enic.h 2016-05-26 16:59:16.531326660 -0700
++++ dpdk-16.04/drivers/net/enic/enic.h 2016-05-26 16:59:52.689262489 -0700
+@@ -91,6 +91,11 @@
+ struct enic_fdir_node *nodes[ENICPMD_FDIR_MAX];
+ };
+
++struct enic_soft_stats {
++ rte_atomic64_t rx_nombuf;
++ rte_atomic64_t rx_packet_errors;
++};
++
+ /* Per-instance private data structure */
+ struct enic {
+ struct enic *next;
+@@ -133,6 +138,8 @@
+ /* interrupt resource */
+ struct vnic_intr intr;
+ unsigned int intr_count;
++
++ struct enic_soft_stats soft_stats;
+ };
+
+ static inline unsigned int enic_cq_rq(__rte_unused struct enic *enic, unsigned int rq)
+diff -ur dpdk-16.04.orig/drivers/net/enic/enic_main.c dpdk-16.04/drivers/net/enic/enic_main.c
+--- dpdk-16.04.orig/drivers/net/enic/enic_main.c 2016-05-26 16:59:16.533326822 -0700
++++ dpdk-16.04/drivers/net/enic/enic_main.c 2016-05-26 17:08:11.768801926 -0700
+@@ -142,22 +142,51 @@
+ }
+
+
++static void enic_clear_soft_stats(struct enic *enic)
++{
++ struct enic_soft_stats *soft_stats = &enic->soft_stats;
++ rte_atomic64_clear(&soft_stats->rx_nombuf);
++ rte_atomic64_clear(&soft_stats->rx_packet_errors);
++}
++
++static void enic_init_soft_stats(struct enic *enic)
++{
++ struct enic_soft_stats *soft_stats = &enic->soft_stats;
++ rte_atomic64_init(&soft_stats->rx_nombuf);
++ rte_atomic64_init(&soft_stats->rx_packet_errors);
++ enic_clear_soft_stats(enic);
++}
++
+ void enic_dev_stats_clear(struct enic *enic)
+ {
+ if (vnic_dev_stats_clear(enic->vdev))
+ dev_err(enic, "Error in clearing stats\n");
++ enic_clear_soft_stats(enic);
+ }
+
+ void enic_dev_stats_get(struct enic *enic, struct rte_eth_stats *r_stats)
+ {
+ struct vnic_stats *stats;
++ struct enic_soft_stats *soft_stats = &enic->soft_stats;
++ int64_t rx_truncated;
++ uint64_t rx_packet_errors;
+
+ if (vnic_dev_stats_dump(enic->vdev, &stats)) {
+ dev_err(enic, "Error in getting stats\n");
+ return;
+ }
+
+- r_stats->ipackets = stats->rx.rx_frames_ok;
++ /* The number of truncated packets can only be calculated by
++ * subtracting a hardware counter from error packets received by
++ * the driver. Note: this causes transient inaccuracies in the
++ * ipackets count. Also, the length of truncated packets are
++ * counted in ibytes even though truncated packets are dropped
++ * which can make ibytes be slightly higher than it should be.
++ */
++ rx_packet_errors = rte_atomic64_read(&soft_stats->rx_packet_errors);
++ rx_truncated = rx_packet_errors - stats->rx.rx_errors;
++
++ r_stats->ipackets = stats->rx.rx_frames_ok - rx_truncated;
+ r_stats->opackets = stats->tx.tx_frames_ok;
+
+ r_stats->ibytes = stats->rx.rx_bytes_ok;
+@@ -166,10 +195,9 @@
+ r_stats->ierrors = stats->rx.rx_errors + stats->rx.rx_drop;
+ r_stats->oerrors = stats->tx.tx_errors;
+
+- r_stats->imissed = stats->rx.rx_no_bufs;
++ r_stats->imissed = stats->rx.rx_no_bufs + rx_truncated;
+
+- r_stats->imcasts = stats->rx.rx_multicast_frames_ok;
+- r_stats->rx_nombuf = stats->rx.rx_no_bufs;
++ r_stats->rx_nombuf = rte_atomic64_read(&soft_stats->rx_nombuf);
+ }
+
+ void enic_del_mac_address(struct enic *enic)
+@@ -755,6 +783,8 @@
+ {
+ int ret;
+
++ enic_init_soft_stats(enic);
++
+ ret = enic_set_rss_nic_cfg(enic);
+ if (ret) {
+ dev_err(enic, "Failed to config nic, aborting.\n");
+diff -ur dpdk-16.04.orig/drivers/net/enic/enic_rxtx.c dpdk-16.04/drivers/net/enic/enic_rxtx.c
+--- dpdk-16.04.orig/drivers/net/enic/enic_rxtx.c 2016-05-26 16:59:16.522325929 -0700
++++ dpdk-16.04/drivers/net/enic/enic_rxtx.c 2016-05-26 16:59:52.694262896 -0700
+@@ -251,6 +251,7 @@
+ struct vnic_cq *cq;
+ volatile struct cq_desc *cqd_ptr;
+ uint8_t color;
++ uint16_t nb_err = 0;
+
+ cq = &enic->cq[enic_cq_rq(enic, rq->index)];
+ rx_id = cq->to_clean; /* index of cqd, rqd, mbuf_table */
+@@ -278,10 +279,7 @@
+ /* 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++;
++ rte_atomic64_inc(&enic->soft_stats.rx_nombuf);
+ break;
+ }
+
+@@ -323,9 +321,10 @@
+ rxmb->packet_type = enic_cq_rx_flags_to_pkt_type(&cqd);
+ enic_cq_rx_to_pkt_flags(&cqd, rxmb);
+ } else {
+- rxmb->pkt_len = 0;
+- rxmb->packet_type = 0;
+- rxmb->ol_flags = 0;
++ rte_pktmbuf_free(rxmb);
++ rte_atomic64_inc(&enic->soft_stats.rx_packet_errors);
++ nb_err++;
++ continue;
+ }
+ rxmb->data_len = rxmb->pkt_len;
+
+@@ -337,7 +336,7 @@
+ rx_pkts[nb_rx++] = rxmb;
+ }
+
+- nb_hold += nb_rx;
++ nb_hold += nb_rx + nb_err;
+ cq->to_clean = rx_id;
+
+ if (nb_hold > rq->rx_free_thresh) {
diff --git a/dpdk/dpdk-16.04_patches/0016-enic-scatter-rx.patch b/dpdk/dpdk-16.04_patches/0016-enic-scatter-rx.patch
new file mode 100755
index 00000000000..0f1f3161bed
--- /dev/null
+++ b/dpdk/dpdk-16.04_patches/0016-enic-scatter-rx.patch
@@ -0,0 +1,650 @@
+diff -r -u dpdk-16.04.orig2/drivers/net/enic/base/rq_enet_desc.h dpdk-16.04/drivers/net/enic/base/rq_enet_desc.h
+--- dpdk-16.04.orig2/drivers/net/enic/base/rq_enet_desc.h 2016-05-13 18:09:07.523938072 -0700
++++ dpdk-16.04/drivers/net/enic/base/rq_enet_desc.h 2016-05-13 18:09:54.359743075 -0700
+@@ -55,7 +55,7 @@
+ #define RQ_ENET_TYPE_BITS 2
+ #define RQ_ENET_TYPE_MASK ((1 << RQ_ENET_TYPE_BITS) - 1)
+
+-static inline void rq_enet_desc_enc(struct rq_enet_desc *desc,
++static inline void rq_enet_desc_enc(volatile struct rq_enet_desc *desc,
+ u64 address, u8 type, u16 length)
+ {
+ desc->address = cpu_to_le64(address);
+diff -r -u dpdk-16.04.orig2/drivers/net/enic/base/vnic_rq.c dpdk-16.04/drivers/net/enic/base/vnic_rq.c
+--- dpdk-16.04.orig2/drivers/net/enic/base/vnic_rq.c 2016-05-13 18:09:07.533938883 -0700
++++ dpdk-16.04/drivers/net/enic/base/vnic_rq.c 2016-05-13 18:09:54.360743158 -0700
+@@ -84,11 +84,16 @@
+ iowrite32(cq_index, &rq->ctrl->cq_index);
+ iowrite32(error_interrupt_enable, &rq->ctrl->error_interrupt_enable);
+ iowrite32(error_interrupt_offset, &rq->ctrl->error_interrupt_offset);
+- iowrite32(0, &rq->ctrl->dropped_packet_count);
+ iowrite32(0, &rq->ctrl->error_status);
+ iowrite32(fetch_index, &rq->ctrl->fetch_index);
+ iowrite32(posted_index, &rq->ctrl->posted_index);
+-
++ if (rq->is_sop) {
++// printf("Writing 0x%x to %s rq\n",
++// ((rq->is_sop << 10) | rq->data_queue_idx),
++// rq->is_sop ? "sop":"data");
++ iowrite32(((rq->is_sop << 10) | rq->data_queue_idx),
++ &rq->ctrl->data_ring);
++ }
+ }
+
+ void vnic_rq_init(struct vnic_rq *rq, unsigned int cq_index,
+@@ -96,6 +101,7 @@
+ unsigned int error_interrupt_offset)
+ {
+ u32 fetch_index = 0;
++
+ /* Use current fetch_index as the ring starting point */
+ fetch_index = ioread32(&rq->ctrl->fetch_index);
+
+@@ -110,6 +116,8 @@
+ error_interrupt_offset);
+ rq->rxst_idx = 0;
+ rq->tot_pkts = 0;
++ rq->pkt_first_seg = NULL;
++ rq->pkt_last_seg = NULL;
+ }
+
+ void vnic_rq_error_out(struct vnic_rq *rq, unsigned int error)
+diff -r -u dpdk-16.04.orig2/drivers/net/enic/base/vnic_rq.h dpdk-16.04/drivers/net/enic/base/vnic_rq.h
+--- dpdk-16.04.orig2/drivers/net/enic/base/vnic_rq.h 2016-05-13 18:09:07.540939452 -0700
++++ dpdk-16.04/drivers/net/enic/base/vnic_rq.h 2016-05-13 18:09:54.362743322 -0700
+@@ -60,10 +60,18 @@
+ u32 pad7;
+ u32 error_status; /* 0x48 */
+ u32 pad8;
+- u32 dropped_packet_count; /* 0x50 */
++ u32 tcp_sn; /* 0x50 */
+ u32 pad9;
+- u32 dropped_packet_count_rc; /* 0x58 */
++ u32 unused; /* 0x58 */
+ u32 pad10;
++ u32 dca_select; /* 0x60 */
++ u32 pad11;
++ u32 dca_value; /* 0x68 */
++ u32 pad12;
++ u32 data_ring; /* 0x70 */
++ u32 pad13;
++ u32 header_split; /* 0x78 */
++ u32 pad14;
+ };
+
+ struct vnic_rq {
+@@ -82,6 +90,12 @@
+ struct rte_mempool *mp;
+ uint16_t rxst_idx;
+ uint32_t tot_pkts;
++ uint16_t data_queue_idx;
++ uint8_t is_sop;
++ uint8_t in_use;
++ struct rte_mbuf *pkt_first_seg;
++ struct rte_mbuf *pkt_last_seg;
++ unsigned int max_mbufs_per_pkt;
+ };
+
+ static inline unsigned int vnic_rq_desc_avail(struct vnic_rq *rq)
+diff -r -u dpdk-16.04.orig2/drivers/net/enic/enic.h dpdk-16.04/drivers/net/enic/enic.h
+--- dpdk-16.04.orig2/drivers/net/enic/enic.h 2016-05-13 18:09:07.553940507 -0700
++++ dpdk-16.04/drivers/net/enic/enic.h 2016-05-13 18:09:54.365743565 -0700
+@@ -142,6 +142,16 @@
+ struct enic_soft_stats soft_stats;
+ };
+
++static inline unsigned int enic_sop_rq(__rte_unused struct enic *enic, unsigned int rq)
++{
++ return rq * 2;
++}
++
++static inline unsigned int enic_data_rq(__rte_unused struct enic *enic, unsigned int rq)
++{
++ return rq * 2 + 1;
++}
++
+ static inline unsigned int enic_cq_rq(__rte_unused struct enic *enic, unsigned int rq)
+ {
+ return rq;
+diff -r -u dpdk-16.04.orig2/drivers/net/enic/enic_main.c dpdk-16.04/drivers/net/enic/enic_main.c
+--- dpdk-16.04.orig2/drivers/net/enic/enic_main.c 2016-05-13 18:09:07.557940834 -0700
++++ dpdk-16.04/drivers/net/enic/enic_main.c 2016-05-13 18:10:40.099459001 -0700
+@@ -248,15 +248,23 @@
+ unsigned int error_interrupt_offset = 0;
+ unsigned int index = 0;
+ unsigned int cq_idx;
++ struct vnic_rq *data_rq;
+
+ vnic_dev_stats_clear(enic->vdev);
+
+ for (index = 0; index < enic->rq_count; index++) {
+- vnic_rq_init(&enic->rq[index],
++ vnic_rq_init(&enic->rq[enic_sop_rq(enic, index)],
+ enic_cq_rq(enic, index),
+ error_interrupt_enable,
+ error_interrupt_offset);
+
++ data_rq = &enic->rq[enic_data_rq(enic, index)];
++ if (data_rq->in_use)
++ vnic_rq_init(data_rq,
++ enic_cq_rq(enic, index),
++ error_interrupt_enable,
++ error_interrupt_offset);
++
+ cq_idx = enic_cq_rq(enic, index);
+ vnic_cq_init(&enic->cq[cq_idx],
+ 0 /* flow_control_enable */,
+@@ -306,6 +314,9 @@
+ unsigned i;
+ dma_addr_t dma_addr;
+
++ if (!rq->in_use)
++ return 0;
++
+ dev_debug(enic, "queue %u, allocating %u rx queue mbufs\n", rq->index,
+ rq->ring.desc_count);
+
+@@ -317,20 +328,20 @@
+ return -ENOMEM;
+ }
+
+- dma_addr = (dma_addr_t)(mb->buf_physaddr
+- + RTE_PKTMBUF_HEADROOM);
+-
+- rq_enet_desc_enc(rqd, dma_addr, RQ_ENET_TYPE_ONLY_SOP,
+- mb->buf_len - RTE_PKTMBUF_HEADROOM);
++ dma_addr = (dma_addr_t)(mb->buf_physaddr + RTE_PKTMBUF_HEADROOM);
++ rq_enet_desc_enc(rqd, dma_addr,
++ (rq->is_sop ? RQ_ENET_TYPE_ONLY_SOP
++ : RQ_ENET_TYPE_NOT_SOP),
++ mb->buf_len - RTE_PKTMBUF_HEADROOM);
+ rq->mbuf_ring[i] = mb;
+ }
+
+ /* make sure all prior writes are complete before doing the PIO write */
+ rte_rmb();
+
+- /* 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));
++ /* Post all but the last buffer to VIC. */
++ rq->posted_index = rq->ring.desc_count - 1;
++
+ rq->rx_nb_hold = 0;
+
+ dev_debug(enic, "port=%u, qidx=%u, Write %u posted idx, %u sw held\n",
+@@ -338,6 +349,8 @@
+ iowrite32(rq->posted_index, &rq->ctrl->posted_index);
+ rte_rmb();
+
++// printf("posted %d buffers to %s rq\n", rq->ring.desc_count,
++// rq->is_sop ? "sop" : "data");
+ return 0;
+
+ }
+@@ -399,17 +412,25 @@
+ "Flow director feature will not work\n");
+
+ for (index = 0; index < enic->rq_count; index++) {
+- err = enic_alloc_rx_queue_mbufs(enic, &enic->rq[index]);
++ err = enic_alloc_rx_queue_mbufs(enic, &enic->rq[enic_sop_rq(enic, index)]);
+ if (err) {
+- dev_err(enic, "Failed to alloc RX queue mbufs\n");
++ dev_err(enic, "Failed to alloc sop RX queue mbufs\n");
++ return err;
++ }
++ err = enic_alloc_rx_queue_mbufs(enic, &enic->rq[enic_data_rq(enic, index)]);
++ if (err) {
++ /* release the previously allocated mbufs for the sop rq */
++ enic_rxmbuf_queue_release(enic, &enic->rq[enic_sop_rq(enic, index)]);
++
++ dev_err(enic, "Failed to alloc data RX queue mbufs\n");
+ return err;
+ }
+ }
+
+ for (index = 0; index < enic->wq_count; index++)
+- vnic_wq_enable(&enic->wq[index]);
++ enic_start_wq(enic, index);
+ for (index = 0; index < enic->rq_count; index++)
+- vnic_rq_enable(&enic->rq[index]);
++ enic_start_rq(enic, index);
+
+ vnic_dev_enable_wait(enic->vdev);
+
+@@ -441,14 +462,26 @@
+
+ void enic_free_rq(void *rxq)
+ {
+- struct vnic_rq *rq = (struct vnic_rq *)rxq;
+- struct enic *enic = vnic_dev_priv(rq->vdev);
++ struct vnic_rq *rq_sop = (struct vnic_rq *)rxq;
++ struct enic *enic = vnic_dev_priv(rq_sop->vdev);
++ struct vnic_rq *rq_data = &enic->rq[rq_sop->data_queue_idx];
++
++ enic_rxmbuf_queue_release(enic, rq_sop);
++ if (rq_data->in_use)
++ enic_rxmbuf_queue_release(enic, rq_data);
++
++ rte_free(rq_sop->mbuf_ring);
++ if (rq_data->in_use)
++ rte_free(rq_data->mbuf_ring);
++
++ rq_sop->mbuf_ring = NULL;
++ rq_data->mbuf_ring = NULL;
++
++ vnic_rq_free(rq_sop);
++ if (rq_data->in_use)
++ vnic_rq_free(rq_data);
+
+- 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]);
++ vnic_cq_free(&enic->cq[rq_sop->index]);
+ }
+
+ void enic_start_wq(struct enic *enic, uint16_t queue_idx)
+@@ -463,12 +496,32 @@
+
+ void enic_start_rq(struct enic *enic, uint16_t queue_idx)
+ {
+- vnic_rq_enable(&enic->rq[queue_idx]);
++ struct vnic_rq *rq_sop = &enic->rq[enic_sop_rq(enic, queue_idx)];
++ struct vnic_rq *rq_data = &enic->rq[rq_sop->data_queue_idx];
++
++ if (rq_data->in_use)
++ vnic_rq_enable(rq_data);
++ rte_mb();
++ vnic_rq_enable(rq_sop);
++
+ }
+
+ int enic_stop_rq(struct enic *enic, uint16_t queue_idx)
+ {
+- return vnic_rq_disable(&enic->rq[queue_idx]);
++ int ret1 = 0, ret2 = 0;
++
++ struct vnic_rq *rq_sop = &enic->rq[enic_sop_rq(enic, queue_idx)];
++ struct vnic_rq *rq_data = &enic->rq[rq_sop->data_queue_idx];
++
++ ret2 = vnic_rq_disable(rq_sop);
++ rte_mb();
++ if (rq_data->in_use)
++ ret1 = vnic_rq_disable(rq_data);
++
++ if (ret2)
++ return ret2;
++ else
++ return ret1;
+ }
+
+ int enic_alloc_rq(struct enic *enic, uint16_t queue_idx,
+@@ -476,53 +529,128 @@
+ uint16_t nb_desc)
+ {
+ int rc;
+- struct vnic_rq *rq = &enic->rq[queue_idx];
+-
+- rq->socket_id = socket_id;
+- rq->mp = mp;
++ uint16_t sop_queue_idx = enic_sop_rq(enic, queue_idx);
++ uint16_t data_queue_idx = enic_data_rq(enic, queue_idx);
++ struct vnic_rq *rq_sop = &enic->rq[sop_queue_idx];
++ struct vnic_rq *rq_data = &enic->rq[data_queue_idx];
++ unsigned int mbuf_size, mbufs_per_pkt;
++ unsigned int nb_sop_desc, nb_data_desc;
++ uint16_t min_sop, max_sop, min_data, max_data;
++
++ rq_sop->is_sop = 1;
++ rq_sop->data_queue_idx = data_queue_idx;
++ rq_data->is_sop = 0;
++ rq_data->data_queue_idx = 0;
++ rq_sop->socket_id = socket_id;
++ rq_sop->mp = mp;
++ rq_data->socket_id = socket_id;
++ rq_data->mp = mp;
++ rq_sop->in_use = 1;
++
++ mbuf_size = (uint16_t)(rte_pktmbuf_data_room_size(mp) - RTE_PKTMBUF_HEADROOM);
++
++ /* ceil(mtu/mbuf_size) */
++ mbufs_per_pkt = (enic->config.mtu + (mbuf_size - 1)) / mbuf_size;
++
++ if (mbufs_per_pkt > 1)
++ rq_data->in_use = 1;
++ else
++ rq_data->in_use = 0;
++
++ /* number of descriptors have to be a multiple of 32 */
++ nb_sop_desc = (nb_desc / mbufs_per_pkt) & ~0x1F;
++ nb_data_desc = (nb_desc - nb_sop_desc) & ~0x1F;
++
++ rq_sop->max_mbufs_per_pkt = mbufs_per_pkt;
++ rq_data->max_mbufs_per_pkt = mbufs_per_pkt;
++
++ //printf("mtu = %u, mbuf_size = %u, mbuf_per_pkt = %u\n",
++ // enic->config.mtu, mbuf_size, mbufs_per_pkt);
++
++ if (mbufs_per_pkt > 1) {
++ min_sop = 64;
++ max_sop = ((enic->config.rq_desc_count / (mbufs_per_pkt - 1)) & ~0x1F);
++ min_data = min_sop * (mbufs_per_pkt - 1);
++ max_data = enic->config.rq_desc_count;
++ } else {
++ min_sop = 64;
++ max_sop = enic->config.rq_desc_count;
++ min_data = 0;
++ max_data = 0;
++ }
+
+- if (nb_desc) {
+- if (nb_desc > enic->config.rq_desc_count) {
+- dev_warning(enic,
+- "RQ %d - number of rx desc in cmd line (%d)"\
+- "is greater than that in the UCSM/CIMC adapter"\
+- "policy. Applying the value in the adapter "\
+- "policy (%d).\n",
+- queue_idx, nb_desc, enic->config.rq_desc_count);
+- nb_desc = enic->config.rq_desc_count;
+- }
+- dev_info(enic, "RX Queues - effective number of descs:%d\n",
+- nb_desc);
++ if (nb_desc < (min_sop + min_data)) {
++ dev_warning(enic,
++ "Number of rx descs too low, adjusting to minimum\n");
++ nb_sop_desc = min_sop;
++ nb_data_desc = min_data;
++ } else if (nb_desc > (max_sop + max_data)){
++ dev_warning(enic,
++ "Number of rx_descs too high, adjusting to maximum\n");
++ nb_sop_desc = max_sop;
++ nb_data_desc = max_data;
+ }
++ dev_info(enic, "For mtu %d and mbuf size %d valid rx descriptor range is %d to %d\n",
++ enic->config.mtu, mbuf_size, min_sop + min_data, max_sop + max_data);
++
++ dev_info(enic, "Using %d rx descriptors (sop %d, data %d)\n",
++ nb_sop_desc + nb_data_desc, nb_sop_desc, nb_data_desc);
+
+- /* Allocate queue resources */
+- rc = vnic_rq_alloc(enic->vdev, rq, queue_idx,
+- nb_desc, sizeof(struct rq_enet_desc));
++ /* Allocate sop queue resources */
++ rc = vnic_rq_alloc(enic->vdev, rq_sop, sop_queue_idx,
++ nb_sop_desc, sizeof(struct rq_enet_desc));
+ if (rc) {
+- dev_err(enic, "error in allocation of rq\n");
++ dev_err(enic, "error in allocation of sop rq\n");
+ goto err_exit;
+ }
++ nb_sop_desc = rq_sop->ring.desc_count;
+
++ if (rq_data->in_use) {
++ /* Allocate data queue resources */
++ rc = vnic_rq_alloc(enic->vdev, rq_data, data_queue_idx,
++ nb_data_desc,
++ sizeof(struct rq_enet_desc));
++ if (rc) {
++ dev_err(enic, "error in allocation of data rq\n");
++ goto err_free_rq_sop;
++ }
++ nb_data_desc = rq_data->ring.desc_count;
++ }
+ rc = vnic_cq_alloc(enic->vdev, &enic->cq[queue_idx], queue_idx,
+- socket_id, nb_desc,
+- sizeof(struct cq_enet_rq_desc));
++ socket_id, nb_sop_desc + nb_data_desc,
++ sizeof(struct cq_enet_rq_desc));
+ if (rc) {
+ dev_err(enic, "error in allocation of cq for rq\n");
+- goto err_free_rq_exit;
++ goto err_free_rq_data;
+ }
+
+- /* Allocate the mbuf ring */
+- rq->mbuf_ring = (struct rte_mbuf **)rte_zmalloc_socket("rq->mbuf_ring",
+- sizeof(struct rte_mbuf *) * nb_desc,
+- RTE_CACHE_LINE_SIZE, rq->socket_id);
++ /* Allocate the mbuf rings */
++ rq_sop->mbuf_ring = (struct rte_mbuf **)rte_zmalloc_socket("rq->mbuf_ring",
++ sizeof(struct rte_mbuf *) * nb_sop_desc,
++ RTE_CACHE_LINE_SIZE, rq_sop->socket_id);
++ if (rq_sop->mbuf_ring == NULL)
++ goto err_free_cq;
++
++ if (rq_data->in_use) {
++ rq_data->mbuf_ring = (struct rte_mbuf **)rte_zmalloc_socket("rq->mbuf_ring",
++ sizeof(struct rte_mbuf *) * nb_data_desc,
++ RTE_CACHE_LINE_SIZE, rq_sop->socket_id);
++ if (rq_data->mbuf_ring == NULL)
++ goto err_free_sop_mbuf;
++ }
+
+- if (rq->mbuf_ring != NULL)
+- return 0;
++ return 0;
+
++err_free_sop_mbuf:
++ rte_free(rq_sop->mbuf_ring);
++err_free_cq:
+ /* cleanup on error */
+ vnic_cq_free(&enic->cq[queue_idx]);
+-err_free_rq_exit:
+- vnic_rq_free(rq);
++err_free_rq_data:
++ if (rq_data->in_use)
++ vnic_rq_free(rq_data);
++err_free_rq_sop:
++ vnic_rq_free(rq_sop);
+ err_exit:
+ return -ENOMEM;
+ }
+
+diff -r -u dpdk-16.04.orig2/drivers/net/enic/enic_rxtx.c dpdk-16.04/drivers/net/enic/enic_rxtx.c
+--- dpdk-16.04.orig2/drivers/net/enic/enic_rxtx.c 2016-05-13 18:09:07.556940752 -0700
++++ dpdk-16.04/drivers/net/enic/enic_rxtx.c 2016-05-13 18:12:22.225755674 -0700
+@@ -242,22 +242,27 @@
+ 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 vnic_rq *sop_rq = rx_queue;
++ struct vnic_rq *data_rq;
++ struct vnic_rq *rq;
++ struct enic *enic = vnic_dev_priv(sop_rq->vdev);
++ uint16_t cq_idx;
++ uint16_t rq_idx;
++ uint16_t rq_num;
+ 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;
+- uint16_t nb_err = 0;
++ uint16_t seg_length;
++ struct rte_mbuf *first_seg = sop_rq->pkt_first_seg;
++ struct rte_mbuf *last_seg = sop_rq->pkt_last_seg;
++
++ cq = &enic->cq[enic_cq_rq(enic, sop_rq->index)];
++ cq_idx = cq->to_clean; /* index of cqd, rqd, mbuf_table */
++ cqd_ptr = (struct cq_desc *)(cq->ring.descs) + cq_idx;
+
+- 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 */
++ data_rq = &enic->rq[sop_rq->data_queue_idx];
+
+ while (nb_rx < nb_pkts) {
+ volatile struct rq_enet_desc *rqd_ptr;
+@@ -265,6 +270,7 @@
+ struct cq_desc cqd;
+ uint64_t ol_err_flags;
+ uint8_t packet_error;
++ uint16_t ciflags;
+
+ /* Check for pkts available */
+ color = (cqd_ptr->type_color >> CQ_DESC_COLOR_SHIFT)
+@@ -272,9 +278,13 @@
+ if (color == cq->last_color)
+ break;
+
+- /* Get the cq descriptor and rq pointer */
++ /* Get the cq descriptor and extract rq info from it */
+ cqd = *cqd_ptr;
+- rqd_ptr = (struct rq_enet_desc *)(rq->ring.descs) + rx_id;
++ rq_num = cqd.q_number & CQ_DESC_Q_NUM_MASK;
++ rq_idx = cqd.completed_index & CQ_DESC_COMP_NDX_MASK;
++
++ rq = &enic->rq[rq_num];
++ rqd_ptr = ((struct rq_enet_desc *)rq->ring.descs) + rq_idx;
+
+ /* allocate a new mbuf */
+ nmb = rte_rxmbuf_alloc(rq->mp);
+@@ -287,67 +297,106 @@
+ packet_error = enic_cq_rx_to_pkt_err_flags(&cqd, &ol_err_flags);
+
+ /* Get the mbuf to return and replace with one just allocated */
+- rxmb = rq->mbuf_ring[rx_id];
+- rq->mbuf_ring[rx_id] = nmb;
++ rxmb = rq->mbuf_ring[rq_idx];
++ rq->mbuf_ring[rq_idx] = nmb;
+
+ /* Increment cqd, rqd, mbuf_table index */
+- rx_id++;
+- if (unlikely(rx_id == rq->ring.desc_count)) {
+- rx_id = 0;
++ cq_idx++;
++ if (unlikely(cq_idx == cq->ring.desc_count)) {
++ cq_idx = 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;
++ cqd_ptr = (struct cq_desc *)(cq->ring.descs) + cq_idx;
+ 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);
++// rte_enic_prefetch(rq->mbuf_ring[rx_id]);
++// rte_enic_prefetch((struct rq_enet_desc *)(rq->ring.descs)
++// + rx_id);
++
++ ciflags = enic_cq_rx_desc_ciflags((struct cq_enet_rq_desc *) &cqd);
+
+ /* Push descriptor for newly allocated mbuf */
+- dma_addr = (dma_addr_t)(nmb->buf_physaddr
+- + RTE_PKTMBUF_HEADROOM);
+- rqd_ptr->address = rte_cpu_to_le_64(dma_addr);
+- rqd_ptr->length_type = cpu_to_le16(nmb->buf_len
+- - RTE_PKTMBUF_HEADROOM);
++
++ dma_addr = (dma_addr_t)(nmb->buf_physaddr + RTE_PKTMBUF_HEADROOM);
++ rq_enet_desc_enc(rqd_ptr, dma_addr,
++ (rq->is_sop ? RQ_ENET_TYPE_ONLY_SOP
++ : RQ_ENET_TYPE_NOT_SOP),
++ nmb->buf_len - RTE_PKTMBUF_HEADROOM);
+
+ /* Fill in the rest of the mbuf */
+- rxmb->data_off = RTE_PKTMBUF_HEADROOM;
+- rxmb->nb_segs = 1;
++ seg_length = enic_cq_rx_desc_n_bytes(&cqd);
++ rxmb->packet_type = enic_cq_rx_flags_to_pkt_type(&cqd);
++ enic_cq_rx_to_pkt_flags(&cqd, rxmb);
++ if (rq->is_sop) {
++ first_seg = rxmb;
++ first_seg->nb_segs = 1;
++ first_seg->pkt_len = seg_length;
++ } else {
++ first_seg->pkt_len = (uint16_t)(first_seg->pkt_len
++ + seg_length);
++ first_seg->nb_segs++;
++ last_seg->next = rxmb;
++ }
++
+ rxmb->next = NULL;
+ rxmb->port = enic->port_id;
+- if (!packet_error) {
+- rxmb->pkt_len = enic_cq_rx_desc_n_bytes(&cqd);
+- rxmb->packet_type = enic_cq_rx_flags_to_pkt_type(&cqd);
+- enic_cq_rx_to_pkt_flags(&cqd, rxmb);
+- } else {
+- rte_pktmbuf_free(rxmb);
++ rxmb->data_len = seg_length;
++
++ rq->rx_nb_hold++;
++
++ if (!(enic_cq_rx_desc_eop(ciflags))) {
++ last_seg = rxmb;
++ continue;
++ }
++
++ if (unlikely(packet_error)) {
++ rte_pktmbuf_free(first_seg);
+ rte_atomic64_inc(&enic->soft_stats.rx_packet_errors);
+- nb_err++;
++
+ continue;
+ }
+- rxmb->data_len = rxmb->pkt_len;
++
++
++// printf("EOP: final packet length is %d\n", first_seg->pkt_len);
++// rte_pktmbuf_dump(stdout, first_seg, 64);
+
+ /* prefetch mbuf data for caller */
+- rte_packet_prefetch(RTE_PTR_ADD(rxmb->buf_addr,
++ rte_packet_prefetch(RTE_PTR_ADD(first_seg->buf_addr,
+ RTE_PKTMBUF_HEADROOM));
+
+ /* store the mbuf address into the next entry of the array */
+- rx_pkts[nb_rx++] = rxmb;
++ rx_pkts[nb_rx++] = first_seg;
+ }
+
+- nb_hold += nb_rx + nb_err;
+- cq->to_clean = rx_id;
++ sop_rq->pkt_first_seg = first_seg;
++ sop_rq->pkt_last_seg = last_seg;
++
++ cq->to_clean = cq_idx;
++
++ if ((sop_rq->rx_nb_hold + data_rq->rx_nb_hold) > sop_rq->rx_free_thresh) {
++ if (data_rq->in_use) {
++ data_rq->posted_index = enic_ring_add(data_rq->ring.desc_count,
++ data_rq->posted_index,
++ data_rq->rx_nb_hold);
++ //printf("Processed %d data descs. Posted index now %d\n",
++ // data_rq->rx_nb_hold, data_rq->posted_index);
++ data_rq->rx_nb_hold = 0;
++ }
++ sop_rq->posted_index = enic_ring_add(sop_rq->ring.desc_count,
++ sop_rq->posted_index,
++ sop_rq->rx_nb_hold);
++ //printf("Processed %d sop descs. Posted index now %d\n",
++ // sop_rq->rx_nb_hold, sop_rq->posted_index);
++ sop_rq->rx_nb_hold = 0;
+
+- 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);
++ if (data_rq->in_use)
++ iowrite32(data_rq->posted_index, &data_rq->ctrl->posted_index);
++ rte_compiler_barrier();
++ iowrite32(sop_rq->posted_index, &sop_rq->ctrl->posted_index);
+ }
+
+- rq->rx_nb_hold = nb_hold;
+
+ return nb_rx;
+ }
diff --git a/vnet/vnet/devices/dpdk/dpdk_priv.h b/vnet/vnet/devices/dpdk/dpdk_priv.h
index 9f17348435c..e5a67978a07 100644
--- a/vnet/vnet/devices/dpdk/dpdk_priv.h
+++ b/vnet/vnet/devices/dpdk/dpdk_priv.h
@@ -21,6 +21,7 @@
#define DPDK_NB_TX_DESC_10GE 2048
#define DPDK_NB_RX_DESC_40GE (4096-128)
#define DPDK_NB_TX_DESC_40GE 2048
+#define DPDK_NB_RX_DESC_ENIC (4096+1024)
/* These args appear by themselves */
#define foreach_eal_double_hyphen_predicate_arg \
diff --git a/vnet/vnet/devices/dpdk/init.c b/vnet/vnet/devices/dpdk/init.c
index fac7e9acfeb..28659a9d136 100644
--- a/vnet/vnet/devices/dpdk/init.c
+++ b/vnet/vnet/devices/dpdk/init.c
@@ -385,16 +385,15 @@ dpdk_lib_init (dpdk_main_t * dm)
case VNET_DPDK_PMD_VICE:
case VNET_DPDK_PMD_ENIC:
rte_eth_link_get_nowait(i, &l);
+ xd->nb_rx_desc = DPDK_NB_RX_DESC_ENIC;
if (l.link_speed == 40000)
{
xd->port_type = VNET_DPDK_PORT_TYPE_ETH_40G;
- xd->nb_rx_desc = DPDK_NB_RX_DESC_40GE;
xd->nb_tx_desc = DPDK_NB_TX_DESC_40GE;
}
else
{
xd->port_type = VNET_DPDK_PORT_TYPE_ETH_10G;
- xd->nb_rx_desc = DPDK_NB_RX_DESC_10GE;
xd->nb_tx_desc = DPDK_NB_TX_DESC_10GE;
}
break;