net/mlx5: support connection tracking between two ports
[dpdk.git] / drivers / net / octeontx_ep / otx_ep_rxtx.c
index 401bbe5..28ff74d 100644 (file)
@@ -7,6 +7,8 @@
 #include <rte_eal.h>
 #include <rte_mempool.h>
 #include <rte_mbuf.h>
+#include <rte_io.h>
+#include <rte_net.h>
 #include <ethdev_pci.h>
 
 #include "otx_ep_common.h"
 #include "otx2_ep_vf.h"
 #include "otx_ep_rxtx.h"
 
+/* SDP_LENGTH_S specifies packet length and is of 8-byte size */
+#define INFO_SIZE 8
+#define DROQ_REFILL_THRESHOLD 16
+
 static void
 otx_ep_dmazone_free(const struct rte_memzone *mz)
 {
@@ -36,6 +42,139 @@ otx_ep_dmazone_free(const struct rte_memzone *mz)
                otx_ep_err("Memzone free failed : ret = %d\n", ret);
 }
 
+/* Free IQ resources */
+int
+otx_ep_delete_iqs(struct otx_ep_device *otx_ep, uint32_t iq_no)
+{
+       struct otx_ep_instr_queue *iq;
+
+       iq = otx_ep->instr_queue[iq_no];
+       if (iq == NULL) {
+               otx_ep_err("Invalid IQ[%d]\n", iq_no);
+               return -EINVAL;
+       }
+
+       rte_free(iq->req_list);
+       iq->req_list = NULL;
+
+       if (iq->iq_mz) {
+               otx_ep_dmazone_free(iq->iq_mz);
+               iq->iq_mz = NULL;
+       }
+
+       rte_free(otx_ep->instr_queue[iq_no]);
+       otx_ep->instr_queue[iq_no] = NULL;
+
+       otx_ep->nb_tx_queues--;
+
+       otx_ep_info("IQ[%d] is deleted\n", iq_no);
+
+       return 0;
+}
+
+/* IQ initialization */
+static int
+otx_ep_init_instr_queue(struct otx_ep_device *otx_ep, int iq_no, int num_descs,
+                    unsigned int socket_id)
+{
+       const struct otx_ep_config *conf;
+       struct otx_ep_instr_queue *iq;
+       uint32_t q_size;
+
+       conf = otx_ep->conf;
+       iq = otx_ep->instr_queue[iq_no];
+       q_size = conf->iq.instr_type * num_descs;
+
+       /* IQ memory creation for Instruction submission to OCTEON TX2 */
+       iq->iq_mz = rte_eth_dma_zone_reserve(otx_ep->eth_dev,
+                                            "instr_queue", iq_no, q_size,
+                                            OTX_EP_PCI_RING_ALIGN,
+                                            socket_id);
+       if (iq->iq_mz == NULL) {
+               otx_ep_err("IQ[%d] memzone alloc failed\n", iq_no);
+               goto iq_init_fail;
+       }
+
+       iq->base_addr_dma = iq->iq_mz->iova;
+       iq->base_addr = (uint8_t *)iq->iq_mz->addr;
+
+       if (num_descs & (num_descs - 1)) {
+               otx_ep_err("IQ[%d] descs not in power of 2\n", iq_no);
+               goto iq_init_fail;
+       }
+
+       iq->nb_desc = num_descs;
+
+       /* Create a IQ request list to hold requests that have been
+        * posted to OCTEON TX2. This list will be used for freeing the IQ
+        * data buffer(s) later once the OCTEON TX2 fetched the requests.
+        */
+       iq->req_list = rte_zmalloc_socket("request_list",
+                       (iq->nb_desc * OTX_EP_IQREQ_LIST_SIZE),
+                       RTE_CACHE_LINE_SIZE,
+                       rte_socket_id());
+       if (iq->req_list == NULL) {
+               otx_ep_err("IQ[%d] req_list alloc failed\n", iq_no);
+               goto iq_init_fail;
+       }
+
+       otx_ep_info("IQ[%d]: base: %p basedma: %lx count: %d\n",
+                    iq_no, iq->base_addr, (unsigned long)iq->base_addr_dma,
+                    iq->nb_desc);
+
+       iq->otx_ep_dev = otx_ep;
+       iq->q_no = iq_no;
+       iq->fill_cnt = 0;
+       iq->host_write_index = 0;
+       iq->otx_read_index = 0;
+       iq->flush_index = 0;
+       iq->instr_pending = 0;
+
+       otx_ep->io_qmask.iq |= (1ull << iq_no);
+
+       /* Set 32B/64B mode for each input queue */
+       if (conf->iq.instr_type == 64)
+               otx_ep->io_qmask.iq64B |= (1ull << iq_no);
+
+       iq->iqcmd_64B = (conf->iq.instr_type == 64);
+
+       /* Set up IQ registers */
+       otx_ep->fn_list.setup_iq_regs(otx_ep, iq_no);
+
+       return 0;
+
+iq_init_fail:
+       return -ENOMEM;
+}
+
+int
+otx_ep_setup_iqs(struct otx_ep_device *otx_ep, uint32_t iq_no, int num_descs,
+                unsigned int socket_id)
+{
+       struct otx_ep_instr_queue *iq;
+
+       iq = (struct otx_ep_instr_queue *)rte_zmalloc("otx_ep_IQ", sizeof(*iq),
+                                               RTE_CACHE_LINE_SIZE);
+       if (iq == NULL)
+               return -ENOMEM;
+
+       otx_ep->instr_queue[iq_no] = iq;
+
+       if (otx_ep_init_instr_queue(otx_ep, iq_no, num_descs, socket_id)) {
+               otx_ep_err("IQ init is failed\n");
+               goto delete_IQ;
+       }
+       otx_ep->nb_tx_queues++;
+
+       otx_ep_info("IQ[%d] is created.\n", iq_no);
+
+       return 0;
+
+delete_IQ:
+       otx_ep_delete_iqs(otx_ep, iq_no);
+       return -ENOMEM;
+}
+
 static void
 otx_ep_droq_reset_indices(struct otx_ep_droq *droq)
 {
@@ -220,3 +359,706 @@ delete_OQ:
        otx_ep_delete_oqs(otx_ep, oq_no);
        return -ENOMEM;
 }
+
+static inline void
+otx_ep_iqreq_delete(struct otx_ep_instr_queue *iq, uint32_t idx)
+{
+       uint32_t reqtype;
+       void *buf;
+       struct otx_ep_buf_free_info *finfo;
+
+       buf     = iq->req_list[idx].buf;
+       reqtype = iq->req_list[idx].reqtype;
+
+       switch (reqtype) {
+       case OTX_EP_REQTYPE_NORESP_NET:
+               rte_pktmbuf_free((struct rte_mbuf *)buf);
+               otx_ep_dbg("IQ buffer freed at idx[%d]\n", idx);
+               break;
+
+       case OTX_EP_REQTYPE_NORESP_GATHER:
+               finfo = (struct  otx_ep_buf_free_info *)buf;
+               /* This will take care of multiple segments also */
+               rte_pktmbuf_free(finfo->mbuf);
+               rte_free(finfo->g.sg);
+               rte_free(finfo);
+               break;
+
+       case OTX_EP_REQTYPE_NONE:
+       default:
+               otx_ep_info("This iqreq mode is not supported:%d\n", reqtype);
+       }
+
+       /* Reset the request list at this index */
+       iq->req_list[idx].buf = NULL;
+       iq->req_list[idx].reqtype = 0;
+}
+
+static inline void
+otx_ep_iqreq_add(struct otx_ep_instr_queue *iq, void *buf,
+               uint32_t reqtype, int index)
+{
+       iq->req_list[index].buf = buf;
+       iq->req_list[index].reqtype = reqtype;
+}
+
+static uint32_t
+otx_vf_update_read_index(struct otx_ep_instr_queue *iq)
+{
+       uint32_t new_idx = rte_read32(iq->inst_cnt_reg);
+       if (unlikely(new_idx == 0xFFFFFFFFU))
+               rte_write32(new_idx, iq->inst_cnt_reg);
+       /* Modulo of the new index with the IQ size will give us
+        * the new index.
+        */
+       new_idx &= (iq->nb_desc - 1);
+
+       return new_idx;
+}
+
+static void
+otx_ep_flush_iq(struct otx_ep_instr_queue *iq)
+{
+       uint32_t instr_processed = 0;
+
+       iq->otx_read_index = otx_vf_update_read_index(iq);
+       while (iq->flush_index != iq->otx_read_index) {
+               /* Free the IQ data buffer to the pool */
+               otx_ep_iqreq_delete(iq, iq->flush_index);
+               iq->flush_index =
+                       otx_ep_incr_index(iq->flush_index, 1, iq->nb_desc);
+
+               instr_processed++;
+       }
+
+       iq->stats.instr_processed = instr_processed;
+       iq->instr_pending -= instr_processed;
+}
+
+static inline void
+otx_ep_ring_doorbell(struct otx_ep_device *otx_ep __rte_unused,
+               struct otx_ep_instr_queue *iq)
+{
+       rte_wmb();
+       rte_write64(iq->fill_cnt, iq->doorbell_reg);
+       iq->fill_cnt = 0;
+}
+
+static inline int
+post_iqcmd(struct otx_ep_instr_queue *iq, uint8_t *iqcmd)
+{
+       uint8_t *iqptr, cmdsize;
+
+       /* This ensures that the read index does not wrap around to
+        * the same position if queue gets full before OCTEON TX2 could
+        * fetch any instr.
+        */
+       if (iq->instr_pending > (iq->nb_desc - 1))
+               return OTX_EP_IQ_SEND_FAILED;
+
+       /* Copy cmd into iq */
+       cmdsize = 64;
+       iqptr   = iq->base_addr + (iq->host_write_index << 6);
+
+       rte_memcpy(iqptr, iqcmd, cmdsize);
+
+       /* Increment the host write index */
+       iq->host_write_index =
+               otx_ep_incr_index(iq->host_write_index, 1, iq->nb_desc);
+
+       iq->fill_cnt++;
+
+       /* Flush the command into memory. We need to be sure the data
+        * is in memory before indicating that the instruction is
+        * pending.
+        */
+       iq->instr_pending++;
+       /* OTX_EP_IQ_SEND_SUCCESS */
+       return 0;
+}
+
+
+static int
+otx_ep_send_data(struct otx_ep_device *otx_ep, struct otx_ep_instr_queue *iq,
+                void *cmd, int dbell)
+{
+       uint32_t ret;
+
+       /* Submit IQ command */
+       ret = post_iqcmd(iq, cmd);
+
+       if (ret == OTX_EP_IQ_SEND_SUCCESS) {
+               if (dbell)
+                       otx_ep_ring_doorbell(otx_ep, iq);
+               iq->stats.instr_posted++;
+
+       } else {
+               iq->stats.instr_dropped++;
+               if (iq->fill_cnt)
+                       otx_ep_ring_doorbell(otx_ep, iq);
+       }
+       return ret;
+}
+
+static inline void
+set_sg_size(struct otx_ep_sg_entry *sg_entry, uint16_t size, uint32_t pos)
+{
+#if RTE_BYTE_ORDER == RTE_BIG_ENDIAN
+       sg_entry->u.size[pos] = size;
+#elif RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
+       sg_entry->u.size[3 - pos] = size;
+#endif
+}
+
+/* Enqueue requests/packets to OTX_EP IQ queue.
+ * returns number of requests enqueued successfully
+ */
+uint16_t
+otx_ep_xmit_pkts(void *tx_queue, struct rte_mbuf **pkts, uint16_t nb_pkts)
+{
+       struct otx_ep_instr_64B iqcmd;
+       struct otx_ep_instr_queue *iq;
+       struct otx_ep_device *otx_ep;
+       struct rte_mbuf *m;
+
+       uint32_t iqreq_type, sgbuf_sz;
+       int dbell, index, count = 0;
+       unsigned int pkt_len, i;
+       int gather, gsz;
+       void *iqreq_buf;
+       uint64_t dptr;
+
+       iq = (struct otx_ep_instr_queue *)tx_queue;
+       otx_ep = iq->otx_ep_dev;
+
+       iqcmd.ih.u64 = 0;
+       iqcmd.pki_ih3.u64 = 0;
+       iqcmd.irh.u64 = 0;
+
+       /* ih invars */
+       iqcmd.ih.s.fsz = OTX_EP_FSZ;
+       iqcmd.ih.s.pkind = otx_ep->pkind; /* The SDK decided PKIND value */
+
+       /* pki ih3 invars */
+       iqcmd.pki_ih3.s.w = 1;
+       iqcmd.pki_ih3.s.utt = 1;
+       iqcmd.pki_ih3.s.tagtype = ORDERED_TAG;
+       /* sl will be sizeof(pki_ih3) */
+       iqcmd.pki_ih3.s.sl = OTX_EP_FSZ + OTX_CUST_DATA_LEN;
+
+       /* irh invars */
+       iqcmd.irh.s.opcode = OTX_EP_NW_PKT_OP;
+
+       for (i = 0; i < nb_pkts; i++) {
+               m = pkts[i];
+               if (m->nb_segs == 1) {
+                       /* dptr */
+                       dptr = rte_mbuf_data_iova(m);
+                       pkt_len = rte_pktmbuf_data_len(m);
+                       iqreq_buf = m;
+                       iqreq_type = OTX_EP_REQTYPE_NORESP_NET;
+                       gather = 0;
+                       gsz = 0;
+               } else {
+                       struct otx_ep_buf_free_info *finfo;
+                       int j, frags, num_sg;
+
+                       if (!(otx_ep->tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS))
+                               goto xmit_fail;
+
+                       finfo = (struct otx_ep_buf_free_info *)rte_malloc(NULL,
+                                                       sizeof(*finfo), 0);
+                       if (finfo == NULL) {
+                               otx_ep_err("free buffer alloc failed\n");
+                               goto xmit_fail;
+                       }
+                       num_sg = (m->nb_segs + 3) / 4;
+                       sgbuf_sz = sizeof(struct otx_ep_sg_entry) * num_sg;
+                       finfo->g.sg =
+                               rte_zmalloc(NULL, sgbuf_sz, OTX_EP_SG_ALIGN);
+                       if (finfo->g.sg == NULL) {
+                               rte_free(finfo);
+                               otx_ep_err("sg entry alloc failed\n");
+                               goto xmit_fail;
+                       }
+                       gather = 1;
+                       gsz = m->nb_segs;
+                       finfo->g.num_sg = num_sg;
+                       finfo->g.sg[0].ptr[0] = rte_mbuf_data_iova(m);
+                       set_sg_size(&finfo->g.sg[0], m->data_len, 0);
+                       pkt_len = m->data_len;
+                       finfo->mbuf = m;
+
+                       frags = m->nb_segs - 1;
+                       j = 1;
+                       m = m->next;
+                       while (frags--) {
+                               finfo->g.sg[(j >> 2)].ptr[(j & 3)] =
+                                               rte_mbuf_data_iova(m);
+                               set_sg_size(&finfo->g.sg[(j >> 2)],
+                                               m->data_len, (j & 3));
+                               pkt_len += m->data_len;
+                               j++;
+                               m = m->next;
+                       }
+                       dptr = rte_mem_virt2iova(finfo->g.sg);
+                       iqreq_buf = finfo;
+                       iqreq_type = OTX_EP_REQTYPE_NORESP_GATHER;
+                       if (pkt_len > OTX_EP_MAX_PKT_SZ) {
+                               rte_free(finfo->g.sg);
+                               rte_free(finfo);
+                               otx_ep_err("failed\n");
+                               goto xmit_fail;
+                       }
+               }
+               /* ih vars */
+               iqcmd.ih.s.tlen = pkt_len + iqcmd.ih.s.fsz;
+               iqcmd.ih.s.gather = gather;
+               iqcmd.ih.s.gsz = gsz;
+
+               iqcmd.dptr = dptr;
+               otx_ep_swap_8B_data(&iqcmd.irh.u64, 1);
+
+#ifdef OTX_EP_IO_DEBUG
+               otx_ep_dbg("After swapping\n");
+               otx_ep_dbg("Word0 [dptr]: 0x%016lx\n",
+                          (unsigned long)iqcmd.dptr);
+               otx_ep_dbg("Word1 [ihtx]: 0x%016lx\n", (unsigned long)iqcmd.ih);
+               otx_ep_dbg("Word2 [pki_ih3]: 0x%016lx\n",
+                          (unsigned long)iqcmd.pki_ih3);
+               otx_ep_dbg("Word3 [rptr]: 0x%016lx\n",
+                          (unsigned long)iqcmd.rptr);
+               otx_ep_dbg("Word4 [irh]: 0x%016lx\n", (unsigned long)iqcmd.irh);
+               otx_ep_dbg("Word5 [exhdr[0]]: 0x%016lx\n",
+                               (unsigned long)iqcmd.exhdr[0]);
+               rte_pktmbuf_dump(stdout, m, rte_pktmbuf_pkt_len(m));
+#endif
+               dbell = (i == (unsigned int)(nb_pkts - 1)) ? 1 : 0;
+               index = iq->host_write_index;
+               if (otx_ep_send_data(otx_ep, iq, &iqcmd, dbell))
+                       goto xmit_fail;
+               otx_ep_iqreq_add(iq, iqreq_buf, iqreq_type, index);
+               iq->stats.tx_pkts++;
+               iq->stats.tx_bytes += pkt_len;
+               count++;
+       }
+
+xmit_fail:
+       if (iq->instr_pending >= OTX_EP_MAX_INSTR)
+               otx_ep_flush_iq(iq);
+
+       /* Return no# of instructions posted successfully. */
+       return count;
+}
+
+/* Enqueue requests/packets to OTX_EP IQ queue.
+ * returns number of requests enqueued successfully
+ */
+uint16_t
+otx2_ep_xmit_pkts(void *tx_queue, struct rte_mbuf **pkts, uint16_t nb_pkts)
+{
+       struct otx2_ep_instr_64B iqcmd2;
+       struct otx_ep_instr_queue *iq;
+       struct otx_ep_device *otx_ep;
+       uint64_t dptr;
+       int count = 0;
+       unsigned int i;
+       struct rte_mbuf *m;
+       unsigned int pkt_len;
+       void *iqreq_buf;
+       uint32_t iqreq_type, sgbuf_sz;
+       int gather, gsz;
+       int dbell;
+       int index;
+
+       iq = (struct otx_ep_instr_queue *)tx_queue;
+       otx_ep = iq->otx_ep_dev;
+
+       iqcmd2.ih.u64 = 0;
+       iqcmd2.irh.u64 = 0;
+
+       /* ih invars */
+       iqcmd2.ih.s.fsz = OTX2_EP_FSZ;
+       iqcmd2.ih.s.pkind = otx_ep->pkind; /* The SDK decided PKIND value */
+       /* irh invars */
+       iqcmd2.irh.s.opcode = OTX_EP_NW_PKT_OP;
+
+       for (i = 0; i < nb_pkts; i++) {
+               m = pkts[i];
+               if (m->nb_segs == 1) {
+                       /* dptr */
+                       dptr = rte_mbuf_data_iova(m);
+                       pkt_len = rte_pktmbuf_data_len(m);
+                       iqreq_buf = m;
+                       iqreq_type = OTX_EP_REQTYPE_NORESP_NET;
+                       gather = 0;
+                       gsz = 0;
+               } else {
+                       struct otx_ep_buf_free_info *finfo;
+                       int j, frags, num_sg;
+
+                       if (!(otx_ep->tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS))
+                               goto xmit_fail;
+
+                       finfo = (struct otx_ep_buf_free_info *)
+                                       rte_malloc(NULL, sizeof(*finfo), 0);
+                       if (finfo == NULL) {
+                               otx_ep_err("free buffer alloc failed\n");
+                               goto xmit_fail;
+                       }
+                       num_sg = (m->nb_segs + 3) / 4;
+                       sgbuf_sz = sizeof(struct otx_ep_sg_entry) * num_sg;
+                       finfo->g.sg =
+                               rte_zmalloc(NULL, sgbuf_sz, OTX_EP_SG_ALIGN);
+                       if (finfo->g.sg == NULL) {
+                               rte_free(finfo);
+                               otx_ep_err("sg entry alloc failed\n");
+                               goto xmit_fail;
+                       }
+                       gather = 1;
+                       gsz = m->nb_segs;
+                       finfo->g.num_sg = num_sg;
+                       finfo->g.sg[0].ptr[0] = rte_mbuf_data_iova(m);
+                       set_sg_size(&finfo->g.sg[0], m->data_len, 0);
+                       pkt_len = m->data_len;
+                       finfo->mbuf = m;
+
+                       frags = m->nb_segs - 1;
+                       j = 1;
+                       m = m->next;
+                       while (frags--) {
+                               finfo->g.sg[(j >> 2)].ptr[(j & 3)] =
+                                               rte_mbuf_data_iova(m);
+                               set_sg_size(&finfo->g.sg[(j >> 2)],
+                                               m->data_len, (j & 3));
+                               pkt_len += m->data_len;
+                               j++;
+                               m = m->next;
+                       }
+                       dptr = rte_mem_virt2iova(finfo->g.sg);
+                       iqreq_buf = finfo;
+                       iqreq_type = OTX_EP_REQTYPE_NORESP_GATHER;
+                       if (pkt_len > OTX_EP_MAX_PKT_SZ) {
+                               rte_free(finfo->g.sg);
+                               rte_free(finfo);
+                               otx_ep_err("failed\n");
+                               goto xmit_fail;
+                       }
+               }
+               /* ih vars */
+               iqcmd2.ih.s.tlen = pkt_len + iqcmd2.ih.s.fsz;
+               iqcmd2.ih.s.gather = gather;
+               iqcmd2.ih.s.gsz = gsz;
+               iqcmd2.dptr = dptr;
+               otx_ep_swap_8B_data(&iqcmd2.irh.u64, 1);
+
+#ifdef OTX_EP_IO_DEBUG
+               otx_ep_dbg("After swapping\n");
+               otx_ep_dbg("Word0 [dptr]: 0x%016lx\n",
+                          (unsigned long)iqcmd.dptr);
+               otx_ep_dbg("Word1 [ihtx]: 0x%016lx\n", (unsigned long)iqcmd.ih);
+               otx_ep_dbg("Word2 [pki_ih3]: 0x%016lx\n",
+                          (unsigned long)iqcmd.pki_ih3);
+               otx_ep_dbg("Word3 [rptr]: 0x%016lx\n",
+                          (unsigned long)iqcmd.rptr);
+               otx_ep_dbg("Word4 [irh]: 0x%016lx\n", (unsigned long)iqcmd.irh);
+               otx_ep_dbg("Word5 [exhdr[0]]: 0x%016lx\n",
+                          (unsigned long)iqcmd.exhdr[0]);
+#endif
+               index = iq->host_write_index;
+               dbell = (i == (unsigned int)(nb_pkts - 1)) ? 1 : 0;
+               if (otx_ep_send_data(otx_ep, iq, &iqcmd2, dbell))
+                       goto xmit_fail;
+               otx_ep_iqreq_add(iq, iqreq_buf, iqreq_type, index);
+               iq->stats.tx_pkts++;
+               iq->stats.tx_bytes += pkt_len;
+               count++;
+       }
+
+xmit_fail:
+       if (iq->instr_pending >= OTX_EP_MAX_INSTR)
+               otx_ep_flush_iq(iq);
+
+       /* Return no# of instructions posted successfully. */
+       return count;
+}
+
+static uint32_t
+otx_ep_droq_refill(struct otx_ep_droq *droq)
+{
+       struct otx_ep_droq_desc *desc_ring;
+       struct otx_ep_droq_info *info;
+       struct rte_mbuf *buf = NULL;
+       uint32_t desc_refilled = 0;
+
+       desc_ring = droq->desc_ring;
+
+       while (droq->refill_count && (desc_refilled < droq->nb_desc)) {
+               /* If a valid buffer exists (happens if there is no dispatch),
+                * reuse the buffer, else allocate.
+                */
+               if (droq->recv_buf_list[droq->refill_idx] != NULL)
+                       break;
+
+               buf = rte_pktmbuf_alloc(droq->mpool);
+               /* If a buffer could not be allocated, no point in
+                * continuing
+                */
+               if (buf == NULL) {
+                       droq->stats.rx_alloc_failure++;
+                       break;
+               }
+               info = rte_pktmbuf_mtod(buf, struct otx_ep_droq_info *);
+               memset(info, 0, sizeof(*info));
+
+               droq->recv_buf_list[droq->refill_idx] = buf;
+               desc_ring[droq->refill_idx].buffer_ptr =
+                                       rte_mbuf_data_iova_default(buf);
+
+
+               droq->refill_idx = otx_ep_incr_index(droq->refill_idx, 1,
+                               droq->nb_desc);
+
+               desc_refilled++;
+               droq->refill_count--;
+       }
+
+       return desc_refilled;
+}
+
+static struct rte_mbuf *
+otx_ep_droq_read_packet(struct otx_ep_device *otx_ep,
+                       struct otx_ep_droq *droq, int next_fetch)
+{
+       volatile struct otx_ep_droq_info *info;
+       struct rte_mbuf *droq_pkt2 = NULL;
+       struct rte_mbuf *droq_pkt = NULL;
+       struct rte_net_hdr_lens hdr_lens;
+       struct otx_ep_droq_info *info2;
+       uint64_t total_pkt_len;
+       uint32_t pkt_len = 0;
+       int next_idx;
+
+       droq_pkt  = droq->recv_buf_list[droq->read_idx];
+       droq_pkt2  = droq->recv_buf_list[droq->read_idx];
+       info = rte_pktmbuf_mtod(droq_pkt, struct otx_ep_droq_info *);
+       /* make sure info is available */
+       rte_rmb();
+       if (unlikely(!info->length)) {
+               int retry = OTX_EP_MAX_DELAYED_PKT_RETRIES;
+               /* otx_ep_dbg("OCTEON DROQ[%d]: read_idx: %d; Data not ready "
+                * "yet, Retry; pending=%lu\n", droq->q_no, droq->read_idx,
+                * droq->pkts_pending);
+                */
+               droq->stats.pkts_delayed_data++;
+               while (retry && !info->length)
+                       retry--;
+               if (!retry && !info->length) {
+                       otx_ep_err("OCTEON DROQ[%d]: read_idx: %d; Retry failed !!\n",
+                                  droq->q_no, droq->read_idx);
+                       /* May be zero length packet; drop it */
+                       rte_pktmbuf_free(droq_pkt);
+                       droq->recv_buf_list[droq->read_idx] = NULL;
+                       droq->read_idx = otx_ep_incr_index(droq->read_idx, 1,
+                                                          droq->nb_desc);
+                       droq->stats.dropped_zlp++;
+                       droq->refill_count++;
+                       goto oq_read_fail;
+               }
+       }
+       if (next_fetch) {
+               next_idx = otx_ep_incr_index(droq->read_idx, 1, droq->nb_desc);
+               droq_pkt2  = droq->recv_buf_list[next_idx];
+               info2 = rte_pktmbuf_mtod(droq_pkt2, struct otx_ep_droq_info *);
+               rte_prefetch_non_temporal((const void *)info2);
+       }
+
+       info->length = rte_bswap64(info->length);
+       /* Deduce the actual data size */
+       total_pkt_len = info->length + INFO_SIZE;
+       if (total_pkt_len <= droq->buffer_size) {
+               info->length -=  OTX_EP_RH_SIZE;
+               droq_pkt  = droq->recv_buf_list[droq->read_idx];
+               if (likely(droq_pkt != NULL)) {
+                       droq_pkt->data_off += OTX_EP_DROQ_INFO_SIZE;
+                       /* otx_ep_dbg("OQ: pkt_len[%ld], buffer_size %d\n",
+                        * (long)info->length, droq->buffer_size);
+                        */
+                       pkt_len = (uint32_t)info->length;
+                       droq_pkt->pkt_len  = pkt_len;
+                       droq_pkt->data_len  = pkt_len;
+                       droq_pkt->port = otx_ep->port_id;
+                       droq->recv_buf_list[droq->read_idx] = NULL;
+                       droq->read_idx = otx_ep_incr_index(droq->read_idx, 1,
+                                                          droq->nb_desc);
+                       droq->refill_count++;
+               }
+       } else {
+               struct rte_mbuf *first_buf = NULL;
+               struct rte_mbuf *last_buf = NULL;
+
+               while (pkt_len < total_pkt_len) {
+                       int cpy_len = 0;
+
+                       cpy_len = ((pkt_len + droq->buffer_size) >
+                                       total_pkt_len)
+                                       ? ((uint32_t)total_pkt_len -
+                                               pkt_len)
+                                       : droq->buffer_size;
+
+                       droq_pkt = droq->recv_buf_list[droq->read_idx];
+                       droq->recv_buf_list[droq->read_idx] = NULL;
+
+                       if (likely(droq_pkt != NULL)) {
+                               /* Note the first seg */
+                               if (!pkt_len)
+                                       first_buf = droq_pkt;
+
+                               droq_pkt->port = otx_ep->port_id;
+                               if (!pkt_len) {
+                                       droq_pkt->data_off +=
+                                               OTX_EP_DROQ_INFO_SIZE;
+                                       droq_pkt->pkt_len =
+                                               cpy_len - OTX_EP_DROQ_INFO_SIZE;
+                                       droq_pkt->data_len =
+                                               cpy_len - OTX_EP_DROQ_INFO_SIZE;
+                               } else {
+                                       droq_pkt->pkt_len = cpy_len;
+                                       droq_pkt->data_len = cpy_len;
+                               }
+
+                               if (pkt_len) {
+                                       first_buf->nb_segs++;
+                                       first_buf->pkt_len += droq_pkt->pkt_len;
+                               }
+
+                               if (last_buf)
+                                       last_buf->next = droq_pkt;
+
+                               last_buf = droq_pkt;
+                       } else {
+                               otx_ep_err("no buf\n");
+                       }
+
+                       pkt_len += cpy_len;
+                       droq->read_idx = otx_ep_incr_index(droq->read_idx, 1,
+                                                          droq->nb_desc);
+                       droq->refill_count++;
+               }
+               droq_pkt = first_buf;
+       }
+       droq_pkt->packet_type = rte_net_get_ptype(droq_pkt, &hdr_lens,
+                                       RTE_PTYPE_ALL_MASK);
+       droq_pkt->l2_len = hdr_lens.l2_len;
+       droq_pkt->l3_len = hdr_lens.l3_len;
+       droq_pkt->l4_len = hdr_lens.l4_len;
+
+       if ((droq_pkt->pkt_len > (RTE_ETHER_MAX_LEN + OTX_CUST_DATA_LEN)) &&
+           !(otx_ep->rx_offloads & DEV_RX_OFFLOAD_JUMBO_FRAME)) {
+               rte_pktmbuf_free(droq_pkt);
+               goto oq_read_fail;
+       }
+
+       if (droq_pkt->nb_segs > 1 &&
+           !(otx_ep->rx_offloads & DEV_RX_OFFLOAD_SCATTER)) {
+               rte_pktmbuf_free(droq_pkt);
+               goto oq_read_fail;
+       }
+
+       return droq_pkt;
+
+oq_read_fail:
+       return NULL;
+}
+
+static inline uint32_t
+otx_ep_check_droq_pkts(struct otx_ep_droq *droq)
+{
+       volatile uint64_t pkt_count;
+       uint32_t new_pkts;
+
+       /* Latest available OQ packets */
+       pkt_count = rte_read32(droq->pkts_sent_reg);
+       rte_write32(pkt_count, droq->pkts_sent_reg);
+       new_pkts = pkt_count;
+       droq->pkts_pending += new_pkts;
+       return new_pkts;
+}
+
+/* Check for response arrival from OCTEON TX2
+ * returns number of requests completed
+ */
+uint16_t
+otx_ep_recv_pkts(void *rx_queue,
+                 struct rte_mbuf **rx_pkts,
+                 uint16_t budget)
+{
+       struct otx_ep_droq *droq = rx_queue;
+       struct otx_ep_device *otx_ep;
+       struct rte_mbuf *oq_pkt;
+
+       uint32_t pkts = 0;
+       uint32_t new_pkts = 0;
+       int next_fetch;
+
+       otx_ep = droq->otx_ep_dev;
+
+       if (droq->pkts_pending > budget) {
+               new_pkts = budget;
+       } else {
+               new_pkts = droq->pkts_pending;
+               new_pkts += otx_ep_check_droq_pkts(droq);
+               if (new_pkts > budget)
+                       new_pkts = budget;
+       }
+
+       if (!new_pkts)
+               goto update_credit; /* No pkts at this moment */
+
+       for (pkts = 0; pkts < new_pkts; pkts++) {
+               /* Push the received pkt to application */
+               next_fetch = (pkts == new_pkts - 1) ? 0 : 1;
+               oq_pkt = otx_ep_droq_read_packet(otx_ep, droq, next_fetch);
+               if (!oq_pkt) {
+                       RTE_LOG_DP(ERR, PMD,
+                                  "DROQ read pkt failed pending %" PRIu64
+                                   "last_pkt_count %" PRIu64 "new_pkts %d.\n",
+                                  droq->pkts_pending, droq->last_pkt_count,
+                                  new_pkts);
+                       droq->pkts_pending -= pkts;
+                       droq->stats.rx_err++;
+                       goto finish;
+               }
+               rx_pkts[pkts] = oq_pkt;
+               /* Stats */
+               droq->stats.pkts_received++;
+               droq->stats.bytes_received += oq_pkt->pkt_len;
+       }
+       droq->pkts_pending -= pkts;
+
+       /* Refill DROQ buffers */
+update_credit:
+       if (droq->refill_count >= DROQ_REFILL_THRESHOLD) {
+               int desc_refilled = otx_ep_droq_refill(droq);
+
+               /* Flush the droq descriptor data to memory to be sure
+                * that when we update the credits the data in memory is
+                * accurate.
+                */
+               rte_wmb();
+               rte_write32(desc_refilled, droq->pkts_credit_reg);
+       } else {
+               /*
+                * SDP output goes into DROP state when output doorbell count
+                * goes below drop count. When door bell count is written with
+                * a value greater than drop count SDP output should come out
+                * of DROP state. Due to a race condition this is not happening.
+                * Writing doorbell register with 0 again may make SDP output
+                * come out of this state.
+                */
+
+               rte_write32(0, droq->pkts_credit_reg);
+       }
+finish:
+       return pkts;
+}