net/cnxk: add Rx burst for CN10K
authorJerin Jacob <jerinj@marvell.com>
Wed, 23 Jun 2021 04:46:25 +0000 (10:16 +0530)
committerJerin Jacob <jerinj@marvell.com>
Tue, 29 Jun 2021 21:10:12 +0000 (23:10 +0200)
Add Rx burst support for CN10K SoC.

Signed-off-by: Jerin Jacob <jerinj@marvell.com>
Signed-off-by: Nithin Dabilpuram <ndabilpuram@marvell.com>
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Signed-off-by: Harman Kalra <hkalra@marvell.com>
drivers/net/cnxk/cn10k_ethdev.h
drivers/net/cnxk/cn10k_rx.c [new file with mode: 0644]
drivers/net/cnxk/cn10k_rx.h
drivers/net/cnxk/meson.build

index 18deb95..596985f 100644 (file)
@@ -33,4 +33,7 @@ struct cn10k_eth_rxq {
        uint16_t rq;
 } __plt_cache_aligned;
 
+/* Rx and Tx routines */
+void cn10k_eth_set_rx_function(struct rte_eth_dev *eth_dev);
+
 #endif /* __CN10K_ETHDEV_H__ */
diff --git a/drivers/net/cnxk/cn10k_rx.c b/drivers/net/cnxk/cn10k_rx.c
new file mode 100644 (file)
index 0000000..8b422d0
--- /dev/null
@@ -0,0 +1,45 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(C) 2021 Marvell.
+ */
+
+#include "cn10k_ethdev.h"
+#include "cn10k_rx.h"
+
+#define R(name, f3, f2, f1, f0, flags)                                        \
+       uint16_t __rte_noinline __rte_hot cn10k_nix_recv_pkts_##name(          \
+               void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t pkts)      \
+       {                                                                      \
+               return cn10k_nix_recv_pkts(rx_queue, rx_pkts, pkts, (flags));  \
+       }
+
+NIX_RX_FASTPATH_MODES
+#undef R
+
+static inline void
+pick_rx_func(struct rte_eth_dev *eth_dev,
+            const eth_rx_burst_t rx_burst[2][2][2][2])
+{
+       struct cnxk_eth_dev *dev = cnxk_eth_pmd_priv(eth_dev);
+
+       /* [MARK] [CKSUM] [PTYPE] [RSS] */
+       eth_dev->rx_pkt_burst = rx_burst
+               [!!(dev->rx_offload_flags & NIX_RX_OFFLOAD_MARK_UPDATE_F)]
+               [!!(dev->rx_offload_flags & NIX_RX_OFFLOAD_CHECKSUM_F)]
+               [!!(dev->rx_offload_flags & NIX_RX_OFFLOAD_PTYPE_F)]
+               [!!(dev->rx_offload_flags & NIX_RX_OFFLOAD_RSS_F)];
+}
+
+void
+cn10k_eth_set_rx_function(struct rte_eth_dev *eth_dev)
+{
+       const eth_rx_burst_t nix_eth_rx_burst[2][2][2][2] = {
+#define R(name, f3, f2, f1, f0, flags)                                       \
+       [f3][f2][f1][f0] = cn10k_nix_recv_pkts_##name,
+
+               NIX_RX_FASTPATH_MODES
+#undef R
+       };
+
+       pick_rx_func(eth_dev, nix_eth_rx_burst);
+       rte_mb();
+}
index d3d1661..01c9d29 100644 (file)
@@ -6,6 +6,242 @@
 
 #include <rte_ether.h>
 
+#define NIX_RX_OFFLOAD_NONE         (0)
+#define NIX_RX_OFFLOAD_RSS_F        BIT(0)
 #define NIX_RX_OFFLOAD_PTYPE_F      BIT(1)
+#define NIX_RX_OFFLOAD_CHECKSUM_F    BIT(2)
+#define NIX_RX_OFFLOAD_MARK_UPDATE_F BIT(3)
+
+/* Flags to control cqe_to_mbuf conversion function.
+ * Defining it from backwards to denote its been
+ * not used as offload flags to pick function
+ */
+#define NIX_RX_MULTI_SEG_F BIT(15)
+
+#define CNXK_NIX_CQ_ENTRY_SZ 128
+#define NIX_DESCS_PER_LOOP   4
+#define CQE_CAST(x)         ((struct nix_cqe_hdr_s *)(x))
+#define CQE_SZ(x)           ((x) * CNXK_NIX_CQ_ENTRY_SZ)
+
+union mbuf_initializer {
+       struct {
+               uint16_t data_off;
+               uint16_t refcnt;
+               uint16_t nb_segs;
+               uint16_t port;
+       } fields;
+       uint64_t value;
+};
+
+static __rte_always_inline uint64_t
+nix_clear_data_off(uint64_t oldval)
+{
+       union mbuf_initializer mbuf_init = {.value = oldval};
+
+       mbuf_init.fields.data_off = 0;
+       return mbuf_init.value;
+}
+
+static __rte_always_inline struct rte_mbuf *
+nix_get_mbuf_from_cqe(void *cq, const uint64_t data_off)
+{
+       rte_iova_t buff;
+
+       /* Skip CQE, NIX_RX_PARSE_S and SG HDR(9 DWORDs) and peek buff addr */
+       buff = *((rte_iova_t *)((uint64_t *)cq + 9));
+       return (struct rte_mbuf *)(buff - data_off);
+}
+
+static __rte_always_inline uint32_t
+nix_ptype_get(const void *const lookup_mem, const uint64_t in)
+{
+       const uint16_t *const ptype = lookup_mem;
+       const uint16_t lh_lg_lf = (in & 0xFFF0000000000000) >> 52;
+       const uint16_t tu_l2 = ptype[(in & 0x000FFFF000000000) >> 36];
+       const uint16_t il4_tu = ptype[PTYPE_NON_TUNNEL_ARRAY_SZ + lh_lg_lf];
+
+       return (il4_tu << PTYPE_NON_TUNNEL_WIDTH) | tu_l2;
+}
+
+static __rte_always_inline uint32_t
+nix_rx_olflags_get(const void *const lookup_mem, const uint64_t in)
+{
+       const uint32_t *const ol_flags =
+               (const uint32_t *)((const uint8_t *)lookup_mem +
+                                  PTYPE_ARRAY_SZ);
+
+       return ol_flags[(in & 0xfff00000) >> 20];
+}
+
+static inline uint64_t
+nix_update_match_id(const uint16_t match_id, uint64_t ol_flags,
+                   struct rte_mbuf *mbuf)
+{
+       /* There is no separate bit to check match_id
+        * is valid or not? and no flag to identify it is an
+        * RTE_FLOW_ACTION_TYPE_FLAG vs RTE_FLOW_ACTION_TYPE_MARK
+        * action. The former case addressed through 0 being invalid
+        * value and inc/dec match_id pair when MARK is activated.
+        * The later case addressed through defining
+        * CNXK_FLOW_MARK_DEFAULT as value for
+        * RTE_FLOW_ACTION_TYPE_MARK.
+        * This would translate to not use
+        * CNXK_FLOW_ACTION_FLAG_DEFAULT - 1 and
+        * CNXK_FLOW_ACTION_FLAG_DEFAULT for match_id.
+        * i.e valid mark_id's are from
+        * 0 to CNXK_FLOW_ACTION_FLAG_DEFAULT - 2
+        */
+       if (likely(match_id)) {
+               ol_flags |= PKT_RX_FDIR;
+               if (match_id != CNXK_FLOW_ACTION_FLAG_DEFAULT) {
+                       ol_flags |= PKT_RX_FDIR_ID;
+                       mbuf->hash.fdir.hi = match_id - 1;
+               }
+       }
+
+       return ol_flags;
+}
+
+static __rte_always_inline void
+cn10k_nix_cqe_to_mbuf(const struct nix_cqe_hdr_s *cq, const uint32_t tag,
+                     struct rte_mbuf *mbuf, const void *lookup_mem,
+                     const uint64_t val, const uint16_t flag)
+{
+       const union nix_rx_parse_u *rx =
+               (const union nix_rx_parse_u *)((const uint64_t *)cq + 1);
+       const uint16_t len = rx->pkt_lenm1 + 1;
+       const uint64_t w1 = *(const uint64_t *)rx;
+       uint64_t ol_flags = 0;
+
+       /* Mark mempool obj as "get" as it is alloc'ed by NIX */
+       __mempool_check_cookies(mbuf->pool, (void **)&mbuf, 1, 1);
+
+       if (flag & NIX_RX_OFFLOAD_PTYPE_F)
+               mbuf->packet_type = nix_ptype_get(lookup_mem, w1);
+       else
+               mbuf->packet_type = 0;
+
+       if (flag & NIX_RX_OFFLOAD_RSS_F) {
+               mbuf->hash.rss = tag;
+               ol_flags |= PKT_RX_RSS_HASH;
+       }
+
+       if (flag & NIX_RX_OFFLOAD_CHECKSUM_F)
+               ol_flags |= nix_rx_olflags_get(lookup_mem, w1);
+
+       if (flag & NIX_RX_OFFLOAD_MARK_UPDATE_F)
+               ol_flags = nix_update_match_id(rx->match_id, ol_flags, mbuf);
+
+       mbuf->ol_flags = ol_flags;
+       *(uint64_t *)(&mbuf->rearm_data) = val;
+       mbuf->pkt_len = len;
+
+       mbuf->data_len = len;
+       mbuf->next = NULL;
+}
+
+static inline uint16_t
+nix_rx_nb_pkts(struct cn10k_eth_rxq *rxq, const uint64_t wdata,
+              const uint16_t pkts, const uint32_t qmask)
+{
+       uint32_t available = rxq->available;
+
+       /* Update the available count if cached value is not enough */
+       if (unlikely(available < pkts)) {
+               uint64_t reg, head, tail;
+
+               /* Use LDADDA version to avoid reorder */
+               reg = roc_atomic64_add_sync(wdata, rxq->cq_status);
+               /* CQ_OP_STATUS operation error */
+               if (reg & BIT_ULL(NIX_CQ_OP_STAT_OP_ERR) ||
+                   reg & BIT_ULL(NIX_CQ_OP_STAT_CQ_ERR))
+                       return 0;
+
+               tail = reg & 0xFFFFF;
+               head = (reg >> 20) & 0xFFFFF;
+               if (tail < head)
+                       available = tail - head + qmask + 1;
+               else
+                       available = tail - head;
+
+               rxq->available = available;
+       }
+
+       return RTE_MIN(pkts, available);
+}
+
+static __rte_always_inline uint16_t
+cn10k_nix_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t pkts,
+                   const uint16_t flags)
+{
+       struct cn10k_eth_rxq *rxq = rx_queue;
+       const uint64_t mbuf_init = rxq->mbuf_initializer;
+       const void *lookup_mem = rxq->lookup_mem;
+       const uint64_t data_off = rxq->data_off;
+       const uintptr_t desc = rxq->desc;
+       const uint64_t wdata = rxq->wdata;
+       const uint32_t qmask = rxq->qmask;
+       uint16_t packets = 0, nb_pkts;
+       uint32_t head = rxq->head;
+       struct nix_cqe_hdr_s *cq;
+       struct rte_mbuf *mbuf;
+
+       nb_pkts = nix_rx_nb_pkts(rxq, wdata, pkts, qmask);
+
+       while (packets < nb_pkts) {
+               /* Prefetch N desc ahead */
+               rte_prefetch_non_temporal(
+                       (void *)(desc + (CQE_SZ((head + 2) & qmask))));
+               cq = (struct nix_cqe_hdr_s *)(desc + CQE_SZ(head));
+
+               mbuf = nix_get_mbuf_from_cqe(cq, data_off);
+
+               cn10k_nix_cqe_to_mbuf(cq, cq->tag, mbuf, lookup_mem, mbuf_init,
+                                     flags);
+               rx_pkts[packets++] = mbuf;
+               roc_prefetch_store_keep(mbuf);
+               head++;
+               head &= qmask;
+       }
+
+       rxq->head = head;
+       rxq->available -= nb_pkts;
+
+       /* Free all the CQs that we've processed */
+       plt_write64((wdata | nb_pkts), rxq->cq_door);
+
+       return nb_pkts;
+}
+
+#define RSS_F    NIX_RX_OFFLOAD_RSS_F
+#define PTYPE_F          NIX_RX_OFFLOAD_PTYPE_F
+#define CKSUM_F          NIX_RX_OFFLOAD_CHECKSUM_F
+#define MARK_F   NIX_RX_OFFLOAD_MARK_UPDATE_F
+
+/* [MARK] [CKSUM] [PTYPE] [RSS] */
+#define NIX_RX_FASTPATH_MODES                                         \
+R(no_offload,                  0, 0, 0, 0, NIX_RX_OFFLOAD_NONE)       \
+R(rss,                         0, 0, 0, 1, RSS_F)                     \
+R(ptype,                       0, 0, 1, 0, PTYPE_F)                   \
+R(ptype_rss,                   0, 0, 1, 1, PTYPE_F | RSS_F)           \
+R(cksum,                       0, 1, 0, 0, CKSUM_F)                   \
+R(cksum_rss,                   0, 1, 0, 1, CKSUM_F | RSS_F)           \
+R(cksum_ptype,                 0, 1, 1, 0, CKSUM_F | PTYPE_F)         \
+R(cksum_ptype_rss,             0, 1, 1, 1, CKSUM_F | PTYPE_F | RSS_F) \
+R(mark,                                1, 0, 0, 0, MARK_F)                    \
+R(mark_rss,                    1, 0, 0, 1, MARK_F | RSS_F)            \
+R(mark_ptype,                  1, 0, 1, 0, MARK_F | PTYPE_F)          \
+R(mark_ptype_rss,              1, 0, 1, 1, MARK_F | PTYPE_F | RSS_F)  \
+R(mark_cksum,                  1, 1, 0, 0, MARK_F | CKSUM_F)          \
+R(mark_cksum_rss,              1, 1, 0, 1, MARK_F | CKSUM_F | RSS_F)  \
+R(mark_cksum_ptype,            1, 1, 1, 0, MARK_F | CKSUM_F | PTYPE_F)\
+R(mark_cksum_ptype_rss,                1, 1, 1, 1, MARK_F | CKSUM_F | PTYPE_F | RSS_F)
+
+#define R(name, f3, f2, f1, f0, flags)                                         \
+       uint16_t __rte_noinline __rte_hot cn10k_nix_recv_pkts_##name(          \
+               void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t pkts);
+
+NIX_RX_FASTPATH_MODES
+#undef R
 
 #endif /* __CN10K_RX_H__ */
index f6b95cd..c831e76 100644 (file)
@@ -29,6 +29,7 @@ sources += files(
 # CN10K
 sources += files(
         'cn10k_ethdev.c',
+        'cn10k_rx.c',
 )
 
 deps += ['bus_pci', 'cryptodev', 'eventdev', 'security']