net/bnxt: support LRO
[dpdk.git] / drivers / net / bnxt / bnxt_rxr.c
index 8489a1c..5f30e02 100644 (file)
@@ -34,6 +34,7 @@
 #include <inttypes.h>
 #include <stdbool.h>
 
+#include <rte_bitmap.h>
 #include <rte_byteorder.h>
 #include <rte_malloc.h>
 #include <rte_memory.h>
@@ -102,24 +103,28 @@ static inline int bnxt_alloc_ag_data(struct bnxt_rx_queue *rxq,
        return 0;
 }
 
-#ifdef BNXT_DEBUG
-static void bnxt_reuse_rx_mbuf(struct bnxt_rx_ring_info *rxr, uint16_t cons,
+static inline void bnxt_reuse_rx_mbuf(struct bnxt_rx_ring_info *rxr,
                               struct rte_mbuf *mbuf)
 {
-       uint16_t prod = rxr->rx_prod;
+       uint16_t prod = RING_NEXT(rxr->rx_ring_struct, rxr->rx_prod);
        struct bnxt_sw_rx_bd *prod_rx_buf;
-       struct rx_prod_pkt_bd *prod_bd, *cons_bd;
+       struct rx_prod_pkt_bd *prod_bd;
 
        prod_rx_buf = &rxr->rx_buf_ring[prod];
 
+       RTE_ASSERT(prod_rx_buf->mbuf == NULL);
+       RTE_ASSERT(mbuf != NULL);
+
        prod_rx_buf->mbuf = mbuf;
 
        prod_bd = &rxr->rx_desc_ring[prod];
-       cons_bd = &rxr->rx_desc_ring[cons];
 
-       prod_bd->addr = cons_bd->addr;
+       prod_bd->addr = rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR(mbuf));
+
+       rxr->rx_prod = prod;
 }
 
+#ifdef BNXT_DEBUG
 static void bnxt_reuse_ag_mbuf(struct bnxt_rx_ring_info *rxr, uint16_t cons,
                               struct rte_mbuf *mbuf)
 {
@@ -138,7 +143,190 @@ static void bnxt_reuse_ag_mbuf(struct bnxt_rx_ring_info *rxr, uint16_t cons,
 }
 #endif
 
-static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
+static inline
+struct rte_mbuf *bnxt_consume_rx_buf(struct bnxt_rx_ring_info *rxr,
+                                    uint16_t cons)
+{
+       struct bnxt_sw_rx_bd *cons_rx_buf;
+       struct rte_mbuf *mbuf;
+
+       cons_rx_buf = &rxr->rx_buf_ring[cons];
+       RTE_ASSERT(cons_rx_buf->mbuf != NULL);
+       mbuf = cons_rx_buf->mbuf;
+       cons_rx_buf->mbuf = NULL;
+       return mbuf;
+}
+
+static void bnxt_tpa_start(struct bnxt_rx_queue *rxq,
+                          struct rx_tpa_start_cmpl *tpa_start,
+                          struct rx_tpa_start_cmpl_hi *tpa_start1)
+{
+       struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+       uint8_t agg_id = rte_le_to_cpu_32(tpa_start->agg_id &
+               RX_TPA_START_CMPL_AGG_ID_MASK) >> RX_TPA_START_CMPL_AGG_ID_SFT;
+       uint16_t data_cons;
+       struct bnxt_tpa_info *tpa_info;
+       struct rte_mbuf *mbuf;
+
+       data_cons = tpa_start->opaque;
+       tpa_info = &rxr->tpa_info[agg_id];
+
+       mbuf = bnxt_consume_rx_buf(rxr, data_cons);
+
+       bnxt_reuse_rx_mbuf(rxr, tpa_info->mbuf);
+
+       tpa_info->mbuf = mbuf;
+       tpa_info->len = rte_le_to_cpu_32(tpa_start->len);
+
+       mbuf->nb_segs = 1;
+       mbuf->next = NULL;
+       mbuf->pkt_len = rte_le_to_cpu_32(tpa_start->len);
+       mbuf->data_len = mbuf->pkt_len;
+       mbuf->port = rxq->port_id;
+       mbuf->ol_flags = PKT_RX_LRO;
+       if (likely(tpa_start->flags_type &
+                  rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS_RSS_VALID))) {
+               mbuf->hash.rss = rte_le_to_cpu_32(tpa_start->rss_hash);
+               mbuf->ol_flags |= PKT_RX_RSS_HASH;
+       } else {
+               mbuf->hash.fdir.id = rte_le_to_cpu_16(tpa_start1->cfa_code);
+               mbuf->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
+       }
+       if (tpa_start1->flags2 &
+           rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS2_META_FORMAT_VLAN)) {
+               mbuf->vlan_tci = rte_le_to_cpu_32(tpa_start1->metadata);
+               mbuf->ol_flags |= PKT_RX_VLAN_PKT;
+       }
+       if (likely(tpa_start1->flags2 &
+                  rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS2_L4_CS_CALC)))
+               mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
+
+       /* recycle next mbuf */
+       data_cons = RING_NEXT(rxr->rx_ring_struct, data_cons);
+       bnxt_reuse_rx_mbuf(rxr, bnxt_consume_rx_buf(rxr, data_cons));
+}
+
+static int bnxt_agg_bufs_valid(struct bnxt_cp_ring_info *cpr,
+               uint8_t agg_bufs, uint32_t raw_cp_cons)
+{
+       uint16_t last_cp_cons;
+       struct rx_pkt_cmpl *agg_cmpl;
+
+       raw_cp_cons = ADV_RAW_CMP(raw_cp_cons, agg_bufs);
+       last_cp_cons = RING_CMP(cpr->cp_ring_struct, raw_cp_cons);
+       agg_cmpl = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[last_cp_cons];
+       return CMP_VALID(agg_cmpl, raw_cp_cons, cpr->cp_ring_struct);
+}
+
+/* TPA consume agg buffer out of order, allocate connected data only */
+static int bnxt_prod_ag_mbuf(struct bnxt_rx_queue *rxq)
+{
+       struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+       uint16_t next = RING_NEXT(rxr->ag_ring_struct, rxr->ag_prod);
+
+       /* TODO batch allocation for better performance */
+       while (rte_bitmap_get(rxr->ag_bitmap, next)) {
+               if (unlikely(bnxt_alloc_ag_data(rxq, rxr, next))) {
+                       RTE_LOG(ERR, PMD,
+                               "agg mbuf alloc failed: prod=0x%x\n", next);
+                       break;
+               }
+               rte_bitmap_clear(rxr->ag_bitmap, next);
+               rxr->ag_prod = next;
+               next = RING_NEXT(rxr->ag_ring_struct, next);
+       }
+
+       return 0;
+}
+
+static int bnxt_rx_pages(struct bnxt_rx_queue *rxq,
+                        struct rte_mbuf *mbuf, uint32_t *tmp_raw_cons,
+                        uint8_t agg_buf)
+{
+       struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
+       struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+       int i;
+       uint16_t cp_cons, ag_cons;
+       struct rx_pkt_cmpl *rxcmp;
+       struct rte_mbuf *last = mbuf;
+
+       for (i = 0; i < agg_buf; i++) {
+               struct bnxt_sw_rx_bd *ag_buf;
+               struct rte_mbuf *ag_mbuf;
+               *tmp_raw_cons = NEXT_RAW_CMP(*tmp_raw_cons);
+               cp_cons = RING_CMP(cpr->cp_ring_struct, *tmp_raw_cons);
+               rxcmp = (struct rx_pkt_cmpl *)
+                                       &cpr->cp_desc_ring[cp_cons];
+
+#ifdef BNXT_DEBUG
+               bnxt_dump_cmpl(cp_cons, rxcmp);
+#endif
+
+               ag_cons = rxcmp->opaque;
+               RTE_ASSERT(ag_cons <= rxr->ag_ring_struct->ring_mask);
+               ag_buf = &rxr->ag_buf_ring[ag_cons];
+               ag_mbuf = ag_buf->mbuf;
+               RTE_ASSERT(ag_mbuf != NULL);
+
+               ag_mbuf->data_len = rte_le_to_cpu_16(rxcmp->len);
+
+               mbuf->nb_segs++;
+               mbuf->pkt_len += ag_mbuf->data_len;
+
+               last->next = ag_mbuf;
+               last = ag_mbuf;
+
+               ag_buf->mbuf = NULL;
+
+               /*
+                * As aggregation buffer consumed out of order in TPA module,
+                * use bitmap to track freed slots to be allocated and notified
+                * to NIC
+                */
+               rte_bitmap_set(rxr->ag_bitmap, ag_cons);
+       }
+       bnxt_prod_ag_mbuf(rxq);
+       return 0;
+}
+
+static inline struct rte_mbuf *bnxt_tpa_end(
+               struct bnxt_rx_queue *rxq,
+               uint32_t *raw_cp_cons,
+               struct rx_tpa_end_cmpl *tpa_end,
+               struct rx_tpa_end_cmpl_hi *tpa_end1 __rte_unused)
+{
+       struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
+       struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+       uint8_t agg_id = (tpa_end->agg_id & RX_TPA_END_CMPL_AGG_ID_MASK)
+                       >> RX_TPA_END_CMPL_AGG_ID_SFT;
+       struct rte_mbuf *mbuf;
+       uint8_t agg_bufs;
+       struct bnxt_tpa_info *tpa_info;
+
+       tpa_info = &rxr->tpa_info[agg_id];
+       mbuf = tpa_info->mbuf;
+       RTE_ASSERT(mbuf != NULL);
+
+       rte_prefetch0(mbuf);
+       agg_bufs = (rte_le_to_cpu_32(tpa_end->agg_bufs_v1) &
+               RX_TPA_END_CMPL_AGG_BUFS_MASK) >> RX_TPA_END_CMPL_AGG_BUFS_SFT;
+       if (agg_bufs) {
+               if (!bnxt_agg_bufs_valid(cpr, agg_bufs, *raw_cp_cons))
+                       return NULL;
+               bnxt_rx_pages(rxq, mbuf, raw_cp_cons, agg_bufs);
+       }
+       mbuf->l4_len = tpa_end->payload_offset;
+
+       struct rte_mbuf *new_data = __bnxt_alloc_rx_data(rxq->mb_pool);
+       RTE_ASSERT(new_data != NULL);
+       if (!new_data)
+               return NULL;
+       tpa_info->mbuf = new_data;
+
+       return mbuf;
+}
+
+static int bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
                            struct bnxt_rx_queue *rxq, uint32_t *raw_cons)
 {
        struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
@@ -148,12 +336,13 @@ static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
        uint32_t tmp_raw_cons = *raw_cons;
        uint16_t cons, prod, cp_cons =
            RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
-       uint16_t ag_cons, ag_prod = rxr->ag_prod;
-       struct bnxt_sw_rx_bd *rx_buf;
+#ifdef BNXT_DEBUG
+       uint16_t ag_cons;
+#endif
        struct rte_mbuf *mbuf;
        int rc = 0;
-       uint8_t i;
        uint8_t agg_buf = 0;
+       uint16_t cmp_type;
 
        rxcmp = (struct rx_pkt_cmpl *)
            &cpr->cp_desc_ring[cp_cons];
@@ -165,12 +354,34 @@ static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
        if (!CMP_VALID(rxcmp1, tmp_raw_cons, cpr->cp_ring_struct))
                return -EBUSY;
 
+       cmp_type = CMP_TYPE(rxcmp);
+       if (cmp_type == RX_PKT_CMPL_TYPE_RX_L2_TPA_START) {
+               bnxt_tpa_start(rxq, (struct rx_tpa_start_cmpl *)rxcmp,
+                              (struct rx_tpa_start_cmpl_hi *)rxcmp1);
+               rc = -EINVAL; /* Continue w/o new mbuf */
+               goto next_rx;
+       } else if (cmp_type == RX_PKT_CMPL_TYPE_RX_L2_TPA_END) {
+               mbuf = bnxt_tpa_end(rxq, &tmp_raw_cons,
+                                  (struct rx_tpa_end_cmpl *)rxcmp,
+                                  (struct rx_tpa_end_cmpl_hi *)rxcmp1);
+               if (unlikely(!mbuf))
+                       return -EBUSY;
+               *rx_pkt = mbuf;
+               goto next_rx;
+       } else if (cmp_type != 0x11) {
+               rc = -EINVAL;
+               goto next_rx;
+       }
+
+       agg_buf = (rxcmp->agg_bufs_v1 & RX_PKT_CMPL_AGG_BUFS_MASK)
+                       >> RX_PKT_CMPL_AGG_BUFS_SFT;
+       if (agg_buf && !bnxt_agg_bufs_valid(cpr, agg_buf, tmp_raw_cons))
+               return -EBUSY;
+
        prod = rxr->rx_prod;
 
-       /* EW - GRO deferred to phase 3 */
        cons = rxcmp->opaque;
-       rx_buf = &rxr->rx_buf_ring[cons];
-       mbuf = rx_buf->mbuf;
+       mbuf = bnxt_consume_rx_buf(rxr, cons);
        rte_prefetch0(mbuf);
 
        if (mbuf == NULL)
@@ -190,61 +401,8 @@ static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
                mbuf->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
        }
 
-       agg_buf = (rxcmp->agg_bufs_v1 & RX_PKT_CMPL_AGG_BUFS_MASK)
-                       >> RX_PKT_CMPL_AGG_BUFS_SFT;
-       if (agg_buf) {
-               cp_cons = RING_CMP(cpr->cp_ring_struct, tmp_raw_cons + agg_buf);
-               rxcmp = (struct rx_pkt_cmpl *)
-                                       &cpr->cp_desc_ring[cp_cons];
-               if (!CMP_VALID(rxcmp, tmp_raw_cons + agg_buf,
-                              cpr->cp_ring_struct))
-                       return -EBUSY;
-               RTE_LOG(DEBUG, PMD, "JUMBO Frame %d. %x, agg_buf %x,\n",
-                       mbuf->pkt_len, rxcmp->agg_bufs_v1,  agg_buf);
-       }
-
-       for (i = 0; i < agg_buf; i++) {
-               struct bnxt_sw_rx_bd *ag_buf;
-               struct rte_mbuf *ag_mbuf;
-               tmp_raw_cons = NEXT_RAW_CMP(tmp_raw_cons);
-               cp_cons = RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
-               rxcmp = (struct rx_pkt_cmpl *)
-                                       &cpr->cp_desc_ring[cp_cons];
-               ag_cons = rxcmp->opaque;
-               ag_buf = &rxr->ag_buf_ring[ag_cons];
-               ag_mbuf = ag_buf->mbuf;
-               ag_mbuf->nb_segs = 1;
-               ag_mbuf->data_len = rxcmp->len;
-
-               mbuf->nb_segs++;
-               mbuf->pkt_len += ag_mbuf->data_len;
-               if (mbuf->next == NULL) {
-                       mbuf->next = ag_mbuf;
-               } else {
-                       struct rte_mbuf *temp_mbuf = mbuf;
-
-                       while (temp_mbuf->next != NULL)
-                               temp_mbuf = temp_mbuf->next;
-                       temp_mbuf->next = ag_mbuf;
-               }
-               ag_buf->mbuf = NULL;
-
-               ag_prod = RING_NEXT(rxr->ag_ring_struct, ag_prod);
-               if (bnxt_alloc_ag_data(rxq, rxr, ag_prod)) {
-                       RTE_LOG(ERR, PMD,
-                               "agg mbuf alloc failed: prod=0x%x\n",
-                               ag_prod);
-                       rc = -ENOMEM;
-               }
-               rxr->ag_prod = ag_prod;
-
-#ifdef BNXT_DEBUG
-               if (!CMP_VALID((struct cmpl_base *)
-                       &cpr->cp_desc_ring[cp_cons], tmp_raw_cons,
-                       cpr->cp_ring_struct))
-                       return -EBUSY;
-#endif
-       }
+       if (agg_buf)
+               bnxt_rx_pages(rxq, mbuf, &tmp_raw_cons, agg_buf);
 
        if (rxcmp1->flags2 & RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN) {
                mbuf->vlan_tci = rxcmp1->metadata &
@@ -254,7 +412,6 @@ static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
                mbuf->ol_flags |= PKT_RX_VLAN_PKT;
        }
 
-       rx_buf->mbuf = NULL;
 #ifdef BNXT_DEBUG
        if (rxcmp1->errors_v2 & RX_CMP_L2_ERRORS) {
                /* Re-install the mbuf back to the rx ring */
@@ -293,9 +450,9 @@ static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
         */
 
        *rx_pkt = mbuf;
-#ifdef BNXT_DEBUG
+
 next_rx:
-#endif
+
        *raw_cons = tmp_raw_cons;
 
        return rc;
@@ -330,7 +487,7 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
                        rc = bnxt_rx_pkt(&rx_pkts[nb_rx_pkts], rxq, &raw_cons);
                        if (likely(!rc))
                                nb_rx_pkts++;
-                       else if (rc == -EBUSY)  /* partial completion */
+                       if (rc == -EBUSY)       /* partial completion */
                                break;
                }
                raw_cons = NEXT_RAW_CMP(raw_cons);
@@ -338,6 +495,7 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
                        break;
        }
 
+       cpr->cp_raw_cons = raw_cons;
        if (prod == rxr->rx_prod && ag_prod == rxr->ag_prod) {
                /*
                 * For PMD, there is no need to keep on pushing to REARM
@@ -345,7 +503,6 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
                 */
                return nb_rx_pkts;
        }
-       cpr->cp_raw_cons = raw_cons;
 
        B_CP_DIS_DB(cpr, cpr->cp_raw_cons);
        B_RX_DB(rxr->rx_doorbell, rxr->rx_prod);
@@ -515,5 +672,15 @@ int bnxt_init_one_rx_ring(struct bnxt_rx_queue *rxq)
        }
        RTE_LOG(DEBUG, PMD, "%s AGG Done!\n", __func__);
 
+       if (rxr->tpa_info) {
+               for (i = 0; i < BNXT_TPA_MAX; i++) {
+                       rxr->tpa_info[i].mbuf =
+                               __bnxt_alloc_rx_data(rxq->mb_pool);
+                       if (!rxr->tpa_info[i].mbuf)
+                               return -ENOMEM;
+               }
+       }
+       RTE_LOG(DEBUG, PMD, "%s TPA alloc Done!\n", __func__);
+
        return 0;
 }