net/sfc: implement EF10 native Rx datapath
authorAndrew Rybchenko <arybchenko@solarflare.com>
Mon, 20 Mar 2017 10:15:13 +0000 (10:15 +0000)
committerFerruh Yigit <ferruh.yigit@intel.com>
Tue, 4 Apr 2017 16:59:42 +0000 (18:59 +0200)
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: David Riddoch <driddoch@solarflare.com>
doc/guides/nics/sfc_efx.rst
drivers/net/sfc/Makefile
drivers/net/sfc/sfc_dp.h
drivers/net/sfc/sfc_dp_rx.h
drivers/net/sfc/sfc_ef10.h [new file with mode: 0644]
drivers/net/sfc/sfc_ef10_rx.c [new file with mode: 0644]
drivers/net/sfc/sfc_ethdev.c
drivers/net/sfc/sfc_ev.c
drivers/net/sfc/sfc_kvargs.h
drivers/net/sfc/sfc_rx.c

index 2d7a241..43bd9f1 100644 (file)
@@ -226,12 +226,15 @@ whitelist option like "-w 02:00.0,arg1=value1,...".
 Case-insensitive 1/y/yes/on or 0/n/no/off may be used to specify
 boolean parameters value.
 
-- ``rx_datapath`` [auto|efx] (default **auto**)
+- ``rx_datapath`` [auto|efx|ef10] (default **auto**)
 
   Choose receive datapath implementation.
   **auto** allows the driver itself to make a choice based on firmware
   features available and required by the datapath implementation.
   **efx** chooses libefx-based datapath which supports Rx scatter.
+  **ef10** chooses EF10 (SFN7xxx, SFN8xxx) native datapath which is
+  more efficient than libefx-based and provides richer packet type
+  classification, but lacks Rx scatter support.
 
 - ``perf_profile`` [auto|throughput|low-latency] (default **throughput**)
 
index 088d5d8..c1fa38c 100644 (file)
@@ -95,6 +95,7 @@ SRCS-$(CONFIG_RTE_LIBRTE_SFC_EFX_PMD) += sfc_tso.c
 SRCS-$(CONFIG_RTE_LIBRTE_SFC_EFX_PMD) += sfc_filter.c
 SRCS-$(CONFIG_RTE_LIBRTE_SFC_EFX_PMD) += sfc_flow.c
 SRCS-$(CONFIG_RTE_LIBRTE_SFC_EFX_PMD) += sfc_dp.c
+SRCS-$(CONFIG_RTE_LIBRTE_SFC_EFX_PMD) += sfc_ef10_rx.c
 
 VPATH += $(SRCDIR)/base
 
index c44c44d..f83ea58 100644 (file)
@@ -98,6 +98,7 @@ struct sfc_dp {
        enum sfc_dp_type                type;
        /* Mask of required hardware/firmware capabilities */
        unsigned int                    hw_fw_caps;
+#define SFC_DP_HW_FW_CAP_EF10          0x1
 };
 
 /** List of datapath variants */
index b3c6a6c..9d05a4b 100644 (file)
@@ -83,6 +83,21 @@ struct sfc_dp_rx_qcreate_info {
 
        /** Rx queue size */
        unsigned int            rxq_entries;
+       /** DMA-mapped Rx descriptors ring */
+       void                    *rxq_hw_ring;
+
+       /** Associated event queue size */
+       unsigned int            evq_entries;
+       /** Hardware event ring */
+       void                    *evq_hw_ring;
+
+       /** The queue index in hardware (required to push right doorbell) */
+       unsigned int            hw_index;
+       /**
+        * Virtual address of the memory-mapped BAR to push Rx refill
+        * doorbell
+        */
+       volatile void           *mem_bar;
 };
 
 /**
@@ -122,6 +137,11 @@ typedef int (sfc_dp_rx_qstart_t)(struct sfc_dp_rxq *dp_rxq,
 typedef void (sfc_dp_rx_qstop_t)(struct sfc_dp_rxq *dp_rxq,
                                 unsigned int *evq_read_ptr);
 
+/**
+ * Receive event handler used during queue flush only.
+ */
+typedef bool (sfc_dp_rx_qrx_ev_t)(struct sfc_dp_rxq *dp_rxq, unsigned int id);
+
 /**
  * Receive queue purge function called after queue flush.
  *
@@ -145,6 +165,7 @@ struct sfc_dp_rx {
        sfc_dp_rx_qdestroy_t                    *qdestroy;
        sfc_dp_rx_qstart_t                      *qstart;
        sfc_dp_rx_qstop_t                       *qstop;
+       sfc_dp_rx_qrx_ev_t                      *qrx_ev;
        sfc_dp_rx_qpurge_t                      *qpurge;
        sfc_dp_rx_supported_ptypes_get_t        *supported_ptypes_get;
        sfc_dp_rx_qdesc_npending_t              *qdesc_npending;
@@ -168,6 +189,7 @@ sfc_dp_find_rx_by_caps(struct sfc_dp_list *head, unsigned int avail_caps)
 }
 
 extern struct sfc_dp_rx sfc_efx_rx;
+extern struct sfc_dp_rx sfc_ef10_rx;
 
 #ifdef __cplusplus
 }
diff --git a/drivers/net/sfc/sfc_ef10.h b/drivers/net/sfc/sfc_ef10.h
new file mode 100644 (file)
index 0000000..060d8fe
--- /dev/null
@@ -0,0 +1,107 @@
+/*-
+ *   BSD LICENSE
+ *
+ * Copyright (c) 2017 Solarflare Communications Inc.
+ * All rights reserved.
+ *
+ * This software was jointly developed between OKTET Labs (under contract
+ * for Solarflare) and Solarflare Communications, Inc.
+ *
+ * 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 OWNER 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.
+ */
+
+#ifndef _SFC_EF10_H
+#define _SFC_EF10_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Number of events in one cache line */
+#define SFC_EF10_EV_PER_CACHE_LINE \
+       (RTE_CACHE_LINE_SIZE / sizeof(efx_qword_t))
+
+#define SFC_EF10_EV_QCLEAR_MASK                (~(SFC_EF10_EV_PER_CACHE_LINE - 1))
+
+#if defined(SFC_EF10_EV_QCLEAR_USE_EFX)
+static inline void
+sfc_ef10_ev_qclear_cache_line(void *ptr)
+{
+       efx_qword_t *entry = ptr;
+       unsigned int i;
+
+       for (i = 0; i < SFC_EF10_EV_PER_CACHE_LINE; ++i)
+               EFX_SET_QWORD(entry[i]);
+}
+#else
+/*
+ * It is possible to do it using AVX2 and AVX512F, but it shows less
+ * performance.
+ */
+static inline void
+sfc_ef10_ev_qclear_cache_line(void *ptr)
+{
+       const __m128i val = _mm_set1_epi64x(UINT64_MAX);
+       __m128i *addr = ptr;
+       unsigned int i;
+
+       RTE_BUILD_BUG_ON(sizeof(val) > RTE_CACHE_LINE_SIZE);
+       RTE_BUILD_BUG_ON(RTE_CACHE_LINE_SIZE % sizeof(val) != 0);
+
+       for (i = 0; i < RTE_CACHE_LINE_SIZE / sizeof(val); ++i)
+               _mm_store_si128(&addr[i], val);
+}
+#endif
+
+static inline void
+sfc_ef10_ev_qclear(efx_qword_t *hw_ring, unsigned int ptr_mask,
+                  unsigned int old_read_ptr, unsigned int read_ptr)
+{
+       const unsigned int clear_ptr = read_ptr & SFC_EF10_EV_QCLEAR_MASK;
+       unsigned int old_clear_ptr = old_read_ptr & SFC_EF10_EV_QCLEAR_MASK;
+
+       while (old_clear_ptr != clear_ptr) {
+               sfc_ef10_ev_qclear_cache_line(
+                       &hw_ring[old_clear_ptr & ptr_mask]);
+               old_clear_ptr += SFC_EF10_EV_PER_CACHE_LINE;
+       }
+
+       /*
+        * No barriers here.
+        * Functions which push doorbell should care about correct
+        * ordering: store instructions which fill in EvQ ring should be
+        * retired from CPU and DMA sync before doorbell which will allow
+        * to use these event entries.
+        */
+}
+
+static inline bool
+sfc_ef10_ev_present(const efx_qword_t ev)
+{
+       return ~EFX_QWORD_FIELD(ev, EFX_DWORD_0) |
+              ~EFX_QWORD_FIELD(ev, EFX_DWORD_1);
+}
+
+#ifdef __cplusplus
+}
+#endif
+#endif /* _SFC_EF10_H */
diff --git a/drivers/net/sfc/sfc_ef10_rx.c b/drivers/net/sfc/sfc_ef10_rx.c
new file mode 100644 (file)
index 0000000..2a3bf89
--- /dev/null
@@ -0,0 +1,712 @@
+/*-
+ *   BSD LICENSE
+ *
+ * Copyright (c) 2016 Solarflare Communications Inc.
+ * All rights reserved.
+ *
+ * This software was jointly developed between OKTET Labs (under contract
+ * for Solarflare) and Solarflare Communications, Inc.
+ *
+ * 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 OWNER 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.
+ */
+
+/* EF10 native datapath implementation */
+
+#include <stdbool.h>
+
+#include <rte_byteorder.h>
+#include <rte_mbuf_ptype.h>
+#include <rte_mbuf.h>
+#include <rte_io.h>
+
+#include "efx.h"
+#include "efx_types.h"
+#include "efx_regs.h"
+#include "efx_regs_ef10.h"
+
+#include "sfc_tweak.h"
+#include "sfc_dp_rx.h"
+#include "sfc_kvargs.h"
+#include "sfc_ef10.h"
+
+#define sfc_ef10_rx_err(dpq, ...) \
+       SFC_DP_LOG(SFC_KVARG_DATAPATH_EF10, ERR, dpq, __VA_ARGS__)
+
+/**
+ * Alignment requirement for value written to RX WPTR:
+ * the WPTR must be aligned to an 8 descriptor boundary.
+ */
+#define SFC_EF10_RX_WPTR_ALIGN 8
+
+/**
+ * Maximum number of descriptors/buffers in the Rx ring.
+ * It should guarantee that corresponding event queue never overfill.
+ * EF10 native datapath uses event queue of the same size as Rx queue.
+ * Maximum number of events on datapath can be estimated as number of
+ * Rx queue entries (one event per Rx buffer in the worst case) plus
+ * Rx error and flush events.
+ */
+#define SFC_EF10_RXQ_LIMIT(_ndesc) \
+       ((_ndesc) - 1 /* head must not step on tail */ - \
+        (SFC_EF10_EV_PER_CACHE_LINE - 1) /* max unused EvQ entries */ - \
+        1 /* Rx error */ - 1 /* flush */)
+
+struct sfc_ef10_rx_sw_desc {
+       struct rte_mbuf                 *mbuf;
+};
+
+struct sfc_ef10_rxq {
+       /* Used on data path */
+       unsigned int                    flags;
+#define SFC_EF10_RXQ_STARTED           0x1
+#define SFC_EF10_RXQ_NOT_RUNNING       0x2
+#define SFC_EF10_RXQ_EXCEPTION         0x4
+#define SFC_EF10_RXQ_RSS_HASH          0x8
+       unsigned int                    ptr_mask;
+       unsigned int                    prepared;
+       unsigned int                    completed;
+       unsigned int                    evq_read_ptr;
+       efx_qword_t                     *evq_hw_ring;
+       struct sfc_ef10_rx_sw_desc      *sw_ring;
+       uint64_t                        rearm_data;
+       uint16_t                        prefix_size;
+
+       /* Used on refill */
+       uint16_t                        buf_size;
+       unsigned int                    added;
+       unsigned int                    refill_threshold;
+       struct rte_mempool              *refill_mb_pool;
+       efx_qword_t                     *rxq_hw_ring;
+       volatile void                   *doorbell;
+
+       /* Datapath receive queue anchor */
+       struct sfc_dp_rxq               dp;
+};
+
+static inline struct sfc_ef10_rxq *
+sfc_ef10_rxq_by_dp_rxq(struct sfc_dp_rxq *dp_rxq)
+{
+       return container_of(dp_rxq, struct sfc_ef10_rxq, dp);
+}
+
+static void
+sfc_ef10_rx_qpush(struct sfc_ef10_rxq *rxq)
+{
+       efx_dword_t dword;
+
+       /* Hardware has alignment restriction for WPTR */
+       RTE_BUILD_BUG_ON(SFC_RX_REFILL_BULK % SFC_EF10_RX_WPTR_ALIGN != 0);
+       SFC_ASSERT(RTE_ALIGN(rxq->added, SFC_EF10_RX_WPTR_ALIGN) == rxq->added);
+
+       EFX_POPULATE_DWORD_1(dword, ERF_DZ_RX_DESC_WPTR,
+                            rxq->added & rxq->ptr_mask);
+
+       /* DMA sync to device is not required */
+
+       /*
+        * rte_write32() has rte_io_wmb() which guarantees that the STORE
+        * operations (i.e. Rx and event descriptor updates) that precede
+        * the rte_io_wmb() call are visible to NIC before the STORE
+        * operations that follow it (i.e. doorbell write).
+        */
+       rte_write32(dword.ed_u32[0], rxq->doorbell);
+}
+
+static void
+sfc_ef10_rx_qrefill(struct sfc_ef10_rxq *rxq)
+{
+       const unsigned int ptr_mask = rxq->ptr_mask;
+       const uint32_t buf_size = rxq->buf_size;
+       unsigned int free_space;
+       unsigned int bulks;
+       void *objs[SFC_RX_REFILL_BULK];
+       unsigned int added = rxq->added;
+
+       free_space = SFC_EF10_RXQ_LIMIT(ptr_mask + 1) -
+               (added - rxq->completed);
+
+       if (free_space < rxq->refill_threshold)
+               return;
+
+       bulks = free_space / RTE_DIM(objs);
+       /* refill_threshold guarantees that bulks is positive */
+       SFC_ASSERT(bulks > 0);
+
+       do {
+               unsigned int id;
+               unsigned int i;
+
+               if (unlikely(rte_mempool_get_bulk(rxq->refill_mb_pool, objs,
+                                                 RTE_DIM(objs)) < 0)) {
+                       struct rte_eth_dev_data *dev_data =
+                               rte_eth_devices[rxq->dp.dpq.port_id].data;
+
+                       /*
+                        * It is hardly a safe way to increment counter
+                        * from different contexts, but all PMDs do it.
+                        */
+                       dev_data->rx_mbuf_alloc_failed += RTE_DIM(objs);
+                       /* Return if we have posted nothing yet */
+                       if (added == rxq->added)
+                               return;
+                       /* Push posted */
+                       break;
+               }
+
+               for (i = 0, id = added & ptr_mask;
+                    i < RTE_DIM(objs);
+                    ++i, ++id) {
+                       struct rte_mbuf *m = objs[i];
+                       struct sfc_ef10_rx_sw_desc *rxd;
+                       phys_addr_t phys_addr;
+
+                       SFC_ASSERT((id & ~ptr_mask) == 0);
+                       rxd = &rxq->sw_ring[id];
+                       rxd->mbuf = m;
+
+                       /*
+                        * Avoid writing to mbuf. It is cheaper to do it
+                        * when we receive packet and fill in nearby
+                        * structure members.
+                        */
+
+                       phys_addr = rte_mbuf_data_dma_addr_default(m);
+                       EFX_POPULATE_QWORD_2(rxq->rxq_hw_ring[id],
+                           ESF_DZ_RX_KER_BYTE_CNT, buf_size,
+                           ESF_DZ_RX_KER_BUF_ADDR, phys_addr);
+               }
+
+               added += RTE_DIM(objs);
+       } while (--bulks > 0);
+
+       SFC_ASSERT(rxq->added != added);
+       rxq->added = added;
+       sfc_ef10_rx_qpush(rxq);
+}
+
+static void
+sfc_ef10_rx_prefetch_next(struct sfc_ef10_rxq *rxq, unsigned int next_id)
+{
+       struct rte_mbuf *next_mbuf;
+
+       /* Prefetch next bunch of software descriptors */
+       if ((next_id % (RTE_CACHE_LINE_SIZE / sizeof(rxq->sw_ring[0]))) == 0)
+               rte_prefetch0(&rxq->sw_ring[next_id]);
+
+       /*
+        * It looks strange to prefetch depending on previous prefetch
+        * data, but measurements show that it is really efficient and
+        * increases packet rate.
+        */
+       next_mbuf = rxq->sw_ring[next_id].mbuf;
+       if (likely(next_mbuf != NULL)) {
+               /* Prefetch the next mbuf structure */
+               rte_mbuf_prefetch_part1(next_mbuf);
+
+               /* Prefetch pseudo header of the next packet */
+               /* data_off is not filled in yet */
+               /* Yes, data could be not ready yet, but we hope */
+               rte_prefetch0((uint8_t *)next_mbuf->buf_addr +
+                             RTE_PKTMBUF_HEADROOM);
+       }
+}
+
+static uint16_t
+sfc_ef10_rx_prepared(struct sfc_ef10_rxq *rxq, struct rte_mbuf **rx_pkts,
+                    uint16_t nb_pkts)
+{
+       uint16_t n_rx_pkts = RTE_MIN(nb_pkts, rxq->prepared);
+       unsigned int completed = rxq->completed;
+       unsigned int i;
+
+       rxq->prepared -= n_rx_pkts;
+       rxq->completed = completed + n_rx_pkts;
+
+       for (i = 0; i < n_rx_pkts; ++i, ++completed)
+               rx_pkts[i] = rxq->sw_ring[completed & rxq->ptr_mask].mbuf;
+
+       return n_rx_pkts;
+}
+
+static void
+sfc_ef10_rx_ev_to_offloads(struct sfc_ef10_rxq *rxq, const efx_qword_t rx_ev,
+                          struct rte_mbuf *m)
+{
+       uint32_t l2_ptype = 0;
+       uint32_t l3_ptype = 0;
+       uint32_t l4_ptype = 0;
+       uint64_t ol_flags = 0;
+
+       if (unlikely(EFX_TEST_QWORD_BIT(rx_ev, ESF_DZ_RX_PARSE_INCOMPLETE_LBN)))
+               goto done;
+
+       switch (EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_ETH_TAG_CLASS)) {
+       case ESE_DZ_ETH_TAG_CLASS_NONE:
+               l2_ptype = RTE_PTYPE_L2_ETHER;
+               break;
+       case ESE_DZ_ETH_TAG_CLASS_VLAN1:
+               l2_ptype = RTE_PTYPE_L2_ETHER_VLAN;
+               break;
+       case ESE_DZ_ETH_TAG_CLASS_VLAN2:
+               l2_ptype = RTE_PTYPE_L2_ETHER_QINQ;
+               break;
+       default:
+               /* Unexpected Eth tag class */
+               SFC_ASSERT(false);
+       }
+
+       switch (EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_L3_CLASS)) {
+       case ESE_DZ_L3_CLASS_IP4_FRAG:
+               l4_ptype = RTE_PTYPE_L4_FRAG;
+               /* FALLTHROUGH */
+       case ESE_DZ_L3_CLASS_IP4:
+               l3_ptype = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
+               ol_flags |= PKT_RX_RSS_HASH |
+                       ((EFX_TEST_QWORD_BIT(rx_ev,
+                                            ESF_DZ_RX_IPCKSUM_ERR_LBN)) ?
+                        PKT_RX_IP_CKSUM_BAD : PKT_RX_IP_CKSUM_GOOD);
+               break;
+       case ESE_DZ_L3_CLASS_IP6_FRAG:
+               l4_ptype |= RTE_PTYPE_L4_FRAG;
+               /* FALLTHROUGH */
+       case ESE_DZ_L3_CLASS_IP6:
+               l3_ptype |= RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
+               ol_flags |= PKT_RX_RSS_HASH;
+               break;
+       case ESE_DZ_L3_CLASS_ARP:
+               /* Override Layer 2 packet type */
+               l2_ptype = RTE_PTYPE_L2_ETHER_ARP;
+               break;
+       default:
+               /* Unexpected Layer 3 class */
+               SFC_ASSERT(false);
+       }
+
+       switch (EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_L4_CLASS)) {
+       case ESE_DZ_L4_CLASS_TCP:
+               l4_ptype = RTE_PTYPE_L4_TCP;
+               ol_flags |=
+                       (EFX_TEST_QWORD_BIT(rx_ev,
+                                           ESF_DZ_RX_TCPUDP_CKSUM_ERR_LBN)) ?
+                       PKT_RX_L4_CKSUM_BAD : PKT_RX_L4_CKSUM_GOOD;
+               break;
+       case ESE_DZ_L4_CLASS_UDP:
+               l4_ptype = RTE_PTYPE_L4_UDP;
+               ol_flags |=
+                       (EFX_TEST_QWORD_BIT(rx_ev,
+                                           ESF_DZ_RX_TCPUDP_CKSUM_ERR_LBN)) ?
+                       PKT_RX_L4_CKSUM_BAD : PKT_RX_L4_CKSUM_GOOD;
+               break;
+       case ESE_DZ_L4_CLASS_UNKNOWN:
+               break;
+       default:
+               /* Unexpected Layer 4 class */
+               SFC_ASSERT(false);
+       }
+
+       /* Remove RSS hash offload flag if RSS is not enabled */
+       if (~rxq->flags & SFC_EF10_RXQ_RSS_HASH)
+               ol_flags &= ~PKT_RX_RSS_HASH;
+
+done:
+       m->ol_flags = ol_flags;
+       m->packet_type = l2_ptype | l3_ptype | l4_ptype;
+}
+
+static uint16_t
+sfc_ef10_rx_pseudo_hdr_get_len(const uint8_t *pseudo_hdr)
+{
+       return rte_le_to_cpu_16(*(const uint16_t *)&pseudo_hdr[8]);
+}
+
+static uint32_t
+sfc_ef10_rx_pseudo_hdr_get_hash(const uint8_t *pseudo_hdr)
+{
+       return rte_le_to_cpu_32(*(const uint32_t *)pseudo_hdr);
+}
+
+static uint16_t
+sfc_ef10_rx_process_event(struct sfc_ef10_rxq *rxq, efx_qword_t rx_ev,
+                         struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
+{
+       const unsigned int ptr_mask = rxq->ptr_mask;
+       unsigned int completed = rxq->completed;
+       unsigned int ready;
+       struct sfc_ef10_rx_sw_desc *rxd;
+       struct rte_mbuf *m;
+       struct rte_mbuf *m0;
+       uint16_t n_rx_pkts;
+       const uint8_t *pseudo_hdr;
+       uint16_t pkt_len;
+
+       ready = (EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_DSC_PTR_LBITS) - completed) &
+               EFX_MASK32(ESF_DZ_RX_DSC_PTR_LBITS);
+       SFC_ASSERT(ready > 0);
+
+       if (rx_ev.eq_u64[0] &
+           rte_cpu_to_le_64((1ull << ESF_DZ_RX_ECC_ERR_LBN) |
+                            (1ull << ESF_DZ_RX_ECRC_ERR_LBN))) {
+               SFC_ASSERT(rxq->prepared == 0);
+               rxq->completed += ready;
+               while (ready-- > 0) {
+                       rxd = &rxq->sw_ring[completed++ & ptr_mask];
+                       rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
+               }
+               return 0;
+       }
+
+       n_rx_pkts = RTE_MIN(ready, nb_pkts);
+       rxq->prepared = ready - n_rx_pkts;
+       rxq->completed += n_rx_pkts;
+
+       rxd = &rxq->sw_ring[completed++ & ptr_mask];
+
+       sfc_ef10_rx_prefetch_next(rxq, completed & ptr_mask);
+
+       m = rxd->mbuf;
+
+       *rx_pkts++ = m;
+
+       *(uint64_t *)(&m->rearm_data) = rxq->rearm_data;
+       /* rearm_data rewrites ol_flags which is updated below */
+       rte_compiler_barrier();
+
+       /* Classify packet based on Rx event */
+       sfc_ef10_rx_ev_to_offloads(rxq, rx_ev, m);
+
+       /* data_off already moved past pseudo header */
+       pseudo_hdr = (uint8_t *)m->buf_addr + RTE_PKTMBUF_HEADROOM;
+
+       /*
+        * Always get RSS hash from pseudo header to avoid
+        * condition/branching. If it is valid or not depends on
+        * PKT_RX_RSS_HASH in m->ol_flags.
+        */
+       m->hash.rss = sfc_ef10_rx_pseudo_hdr_get_hash(pseudo_hdr);
+
+       if (ready == 1)
+               pkt_len = EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_BYTES) -
+                       rxq->prefix_size;
+       else
+               pkt_len = sfc_ef10_rx_pseudo_hdr_get_len(pseudo_hdr);
+       SFC_ASSERT(pkt_len > 0);
+       rte_pktmbuf_data_len(m) = pkt_len;
+       rte_pktmbuf_pkt_len(m) = pkt_len;
+
+       m->next = NULL;
+
+       /* Remember mbuf to copy offload flags and packet type from */
+       m0 = m;
+       for (--ready; ready > 0; --ready) {
+               rxd = &rxq->sw_ring[completed++ & ptr_mask];
+
+               sfc_ef10_rx_prefetch_next(rxq, completed & ptr_mask);
+
+               m = rxd->mbuf;
+
+               if (ready > rxq->prepared)
+                       *rx_pkts++ = m;
+
+               *(uint64_t *)(&m->rearm_data) = rxq->rearm_data;
+               /* rearm_data rewrites ol_flags which is updated below */
+               rte_compiler_barrier();
+
+               /* Event-dependent information is the same */
+               m->ol_flags = m0->ol_flags;
+               m->packet_type = m0->packet_type;
+
+               /* data_off already moved past pseudo header */
+               pseudo_hdr = (uint8_t *)m->buf_addr + RTE_PKTMBUF_HEADROOM;
+
+               /*
+                * Always get RSS hash from pseudo header to avoid
+                * condition/branching. If it is valid or not depends on
+                * PKT_RX_RSS_HASH in m->ol_flags.
+                */
+               m->hash.rss = sfc_ef10_rx_pseudo_hdr_get_hash(pseudo_hdr);
+
+               pkt_len = sfc_ef10_rx_pseudo_hdr_get_len(pseudo_hdr);
+               SFC_ASSERT(pkt_len > 0);
+               rte_pktmbuf_data_len(m) = pkt_len;
+               rte_pktmbuf_pkt_len(m) = pkt_len;
+
+               m->next = NULL;
+       }
+
+       return n_rx_pkts;
+}
+
+static bool
+sfc_ef10_rx_get_event(struct sfc_ef10_rxq *rxq, efx_qword_t *rx_ev)
+{
+       *rx_ev = rxq->evq_hw_ring[rxq->evq_read_ptr & rxq->ptr_mask];
+
+       if (!sfc_ef10_ev_present(*rx_ev))
+               return false;
+
+       if (unlikely(EFX_QWORD_FIELD(*rx_ev, FSF_AZ_EV_CODE) !=
+                    FSE_AZ_EV_CODE_RX_EV)) {
+               /*
+                * Do not move read_ptr to keep the event for exception
+                * handling by the control path.
+                */
+               rxq->flags |= SFC_EF10_RXQ_EXCEPTION;
+               sfc_ef10_rx_err(&rxq->dp.dpq,
+                               "RxQ exception at EvQ read ptr %#x",
+                               rxq->evq_read_ptr);
+               return false;
+       }
+
+       rxq->evq_read_ptr++;
+       return true;
+}
+
+static uint16_t
+sfc_ef10_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
+{
+       struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(rx_queue);
+       unsigned int evq_old_read_ptr;
+       uint16_t n_rx_pkts;
+       efx_qword_t rx_ev;
+
+       if (unlikely(rxq->flags &
+                    (SFC_EF10_RXQ_NOT_RUNNING | SFC_EF10_RXQ_EXCEPTION)))
+               return 0;
+
+       n_rx_pkts = sfc_ef10_rx_prepared(rxq, rx_pkts, nb_pkts);
+
+       evq_old_read_ptr = rxq->evq_read_ptr;
+       while (n_rx_pkts != nb_pkts && sfc_ef10_rx_get_event(rxq, &rx_ev)) {
+               /*
+                * DROP_EVENT is an internal to the NIC, software should
+                * never see it and, therefore, may ignore it.
+                */
+
+               n_rx_pkts += sfc_ef10_rx_process_event(rxq, rx_ev,
+                                                      rx_pkts + n_rx_pkts,
+                                                      nb_pkts - n_rx_pkts);
+       }
+
+       sfc_ef10_ev_qclear(rxq->evq_hw_ring, rxq->ptr_mask, evq_old_read_ptr,
+                          rxq->evq_read_ptr);
+
+       /* It is not a problem if we refill in the case of exception */
+       sfc_ef10_rx_qrefill(rxq);
+
+       return n_rx_pkts;
+}
+
+static const uint32_t *
+sfc_ef10_supported_ptypes_get(void)
+{
+       static const uint32_t ef10_native_ptypes[] = {
+               RTE_PTYPE_L2_ETHER,
+               RTE_PTYPE_L2_ETHER_ARP,
+               RTE_PTYPE_L2_ETHER_VLAN,
+               RTE_PTYPE_L2_ETHER_QINQ,
+               RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
+               RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
+               RTE_PTYPE_L4_FRAG,
+               RTE_PTYPE_L4_TCP,
+               RTE_PTYPE_L4_UDP,
+               RTE_PTYPE_UNKNOWN
+       };
+
+       return ef10_native_ptypes;
+}
+
+static sfc_dp_rx_qdesc_npending_t sfc_ef10_rx_qdesc_npending;
+static unsigned int
+sfc_ef10_rx_qdesc_npending(__rte_unused struct sfc_dp_rxq *dp_rxq)
+{
+       /*
+        * Correct implementation requires EvQ polling and events
+        * processing (keeping all ready mbufs in prepared).
+        */
+       return -ENOTSUP;
+}
+
+
+static uint64_t
+sfc_ef10_mk_mbuf_rearm_data(uint16_t port_id, uint16_t prefix_size)
+{
+       struct rte_mbuf m;
+
+       memset(&m, 0, sizeof(m));
+
+       rte_mbuf_refcnt_set(&m, 1);
+       m.data_off = RTE_PKTMBUF_HEADROOM + prefix_size;
+       m.nb_segs = 1;
+       m.port = port_id;
+
+       /* rearm_data covers structure members filled in above */
+       rte_compiler_barrier();
+       return *(uint64_t *)(&m.rearm_data);
+}
+
+static sfc_dp_rx_qcreate_t sfc_ef10_rx_qcreate;
+static int
+sfc_ef10_rx_qcreate(uint16_t port_id, uint16_t queue_id,
+                   const struct rte_pci_addr *pci_addr, int socket_id,
+                   const struct sfc_dp_rx_qcreate_info *info,
+                   struct sfc_dp_rxq **dp_rxqp)
+{
+       struct sfc_ef10_rxq *rxq;
+       int rc;
+
+       rc = EINVAL;
+       if (info->rxq_entries != info->evq_entries)
+               goto fail_rxq_args;
+
+       rc = ENOMEM;
+       rxq = rte_zmalloc_socket("sfc-ef10-rxq", sizeof(*rxq),
+                                RTE_CACHE_LINE_SIZE, socket_id);
+       if (rxq == NULL)
+               goto fail_rxq_alloc;
+
+       sfc_dp_queue_init(&rxq->dp.dpq, port_id, queue_id, pci_addr);
+
+       rc = ENOMEM;
+       rxq->sw_ring = rte_calloc_socket("sfc-ef10-rxq-sw_ring",
+                                        info->rxq_entries,
+                                        sizeof(*rxq->sw_ring),
+                                        RTE_CACHE_LINE_SIZE, socket_id);
+       if (rxq->sw_ring == NULL)
+               goto fail_desc_alloc;
+
+       rxq->flags |= SFC_EF10_RXQ_NOT_RUNNING;
+       if (info->flags & SFC_RXQ_FLAG_RSS_HASH)
+               rxq->flags |= SFC_EF10_RXQ_RSS_HASH;
+       rxq->ptr_mask = info->rxq_entries - 1;
+       rxq->evq_hw_ring = info->evq_hw_ring;
+       rxq->refill_threshold = info->refill_threshold;
+       rxq->rearm_data =
+               sfc_ef10_mk_mbuf_rearm_data(port_id, info->prefix_size);
+       rxq->prefix_size = info->prefix_size;
+       rxq->buf_size = info->buf_size;
+       rxq->refill_mb_pool = info->refill_mb_pool;
+       rxq->rxq_hw_ring = info->rxq_hw_ring;
+       rxq->doorbell = (volatile uint8_t *)info->mem_bar +
+                       ER_DZ_RX_DESC_UPD_REG_OFST +
+                       info->hw_index * ER_DZ_RX_DESC_UPD_REG_STEP;
+
+       *dp_rxqp = &rxq->dp;
+       return 0;
+
+fail_desc_alloc:
+       rte_free(rxq);
+
+fail_rxq_alloc:
+fail_rxq_args:
+       return rc;
+}
+
+static sfc_dp_rx_qdestroy_t sfc_ef10_rx_qdestroy;
+static void
+sfc_ef10_rx_qdestroy(struct sfc_dp_rxq *dp_rxq)
+{
+       struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(dp_rxq);
+
+       rte_free(rxq->sw_ring);
+       rte_free(rxq);
+}
+
+static sfc_dp_rx_qstart_t sfc_ef10_rx_qstart;
+static int
+sfc_ef10_rx_qstart(struct sfc_dp_rxq *dp_rxq, unsigned int evq_read_ptr)
+{
+       struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(dp_rxq);
+
+       rxq->prepared = 0;
+       rxq->completed = rxq->added = 0;
+
+       sfc_ef10_rx_qrefill(rxq);
+
+       rxq->evq_read_ptr = evq_read_ptr;
+
+       rxq->flags |= SFC_EF10_RXQ_STARTED;
+       rxq->flags &= ~(SFC_EF10_RXQ_NOT_RUNNING | SFC_EF10_RXQ_EXCEPTION);
+
+       return 0;
+}
+
+static sfc_dp_rx_qstop_t sfc_ef10_rx_qstop;
+static void
+sfc_ef10_rx_qstop(struct sfc_dp_rxq *dp_rxq, unsigned int *evq_read_ptr)
+{
+       struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(dp_rxq);
+
+       rxq->flags |= SFC_EF10_RXQ_NOT_RUNNING;
+
+       *evq_read_ptr = rxq->evq_read_ptr;
+}
+
+static sfc_dp_rx_qrx_ev_t sfc_ef10_rx_qrx_ev;
+static bool
+sfc_ef10_rx_qrx_ev(struct sfc_dp_rxq *dp_rxq, __rte_unused unsigned int id)
+{
+       __rte_unused struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(dp_rxq);
+
+       SFC_ASSERT(rxq->flags & SFC_EF10_RXQ_NOT_RUNNING);
+
+       /*
+        * It is safe to ignore Rx event since we free all mbufs on
+        * queue purge anyway.
+        */
+
+       return false;
+}
+
+static sfc_dp_rx_qpurge_t sfc_ef10_rx_qpurge;
+static void
+sfc_ef10_rx_qpurge(struct sfc_dp_rxq *dp_rxq)
+{
+       struct sfc_ef10_rxq *rxq = sfc_ef10_rxq_by_dp_rxq(dp_rxq);
+       unsigned int i;
+       struct sfc_ef10_rx_sw_desc *rxd;
+
+       for (i = rxq->completed; i != rxq->added; ++i) {
+               rxd = &rxq->sw_ring[i & rxq->ptr_mask];
+               rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
+               rxd->mbuf = NULL;
+       }
+
+       rxq->flags &= ~SFC_EF10_RXQ_STARTED;
+}
+
+struct sfc_dp_rx sfc_ef10_rx = {
+       .dp = {
+               .name           = SFC_KVARG_DATAPATH_EF10,
+               .type           = SFC_DP_RX,
+               .hw_fw_caps     = SFC_DP_HW_FW_CAP_EF10,
+       },
+       .features               = 0,
+       .qcreate                = sfc_ef10_rx_qcreate,
+       .qdestroy               = sfc_ef10_rx_qdestroy,
+       .qstart                 = sfc_ef10_rx_qstart,
+       .qstop                  = sfc_ef10_rx_qstop,
+       .qrx_ev                 = sfc_ef10_rx_qrx_ev,
+       .qpurge                 = sfc_ef10_rx_qpurge,
+       .supported_ptypes_get   = sfc_ef10_supported_ptypes_get,
+       .qdesc_npending         = sfc_ef10_rx_qdesc_npending,
+       .pkt_burst              = sfc_ef10_recv_pkts,
+};
index 6dffc1c..010641b 100644 (file)
@@ -1366,6 +1366,15 @@ sfc_eth_dev_set_ops(struct rte_eth_dev *dev)
        if (sa == NULL || sa->state == SFC_ADAPTER_UNINITIALIZED)
                return -E_RTE_SECONDARY;
 
+       switch (sa->family) {
+       case EFX_FAMILY_HUNTINGTON:
+       case EFX_FAMILY_MEDFORD:
+               avail_caps |= SFC_DP_HW_FW_CAP_EF10;
+               break;
+       default:
+               break;
+       }
+
        rc = sfc_kvargs_process(sa, SFC_KVARG_RX_DATAPATH,
                                sfc_kvarg_string_handler, &rx_name);
        if (rc != 0)
@@ -1414,8 +1423,11 @@ static void
 sfc_register_dp(void)
 {
        /* Register once */
-       if (TAILQ_EMPTY(&sfc_dp_head))
+       if (TAILQ_EMPTY(&sfc_dp_head)) {
+               /* Prefer EF10 datapath */
+               sfc_dp_register(&sfc_dp_head, &sfc_ef10_rx.dp);
                sfc_dp_register(&sfc_dp_head, &sfc_efx_rx.dp);
+       }
 }
 
 static int
index c6b02f2..84b2fd1 100644 (file)
@@ -155,6 +155,20 @@ done:
        return B_FALSE;
 }
 
+static boolean_t
+sfc_ev_dp_rx(void *arg, __rte_unused uint32_t label, uint32_t id,
+            __rte_unused uint32_t size, __rte_unused uint16_t flags)
+{
+       struct sfc_evq *evq = arg;
+       struct sfc_dp_rxq *dp_rxq;
+
+       dp_rxq = evq->dp_rxq;
+       SFC_ASSERT(dp_rxq != NULL);
+
+       SFC_ASSERT(evq->sa->dp_rx->qrx_ev != NULL);
+       return evq->sa->dp_rx->qrx_ev(dp_rxq, id);
+}
+
 static boolean_t
 sfc_ev_nop_tx(void *arg, uint32_t label, uint32_t id)
 {
@@ -414,7 +428,7 @@ static const efx_ev_callbacks_t sfc_ev_callbacks_efx_rx = {
 
 static const efx_ev_callbacks_t sfc_ev_callbacks_dp_rx = {
        .eec_initialized        = sfc_ev_initialized,
-       .eec_rx                 = sfc_ev_nop_rx,
+       .eec_rx                 = sfc_ev_dp_rx,
        .eec_tx                 = sfc_ev_nop_tx,
        .eec_exception          = sfc_ev_exception,
        .eec_rxq_flush_done     = sfc_ev_rxq_flush_done,
index e86a7e9..38d17e0 100644 (file)
@@ -57,10 +57,12 @@ extern "C" {
 #define SFC_KVARG_STATS_UPDATE_PERIOD_MS       "stats_update_period_ms"
 
 #define SFC_KVARG_DATAPATH_EFX         "efx"
+#define SFC_KVARG_DATAPATH_EF10                "ef10"
 
 #define SFC_KVARG_RX_DATAPATH          "rx_datapath"
 #define SFC_KVARG_VALUES_RX_DATAPATH \
-       "[" SFC_KVARG_DATAPATH_EFX "]"
+       "[" SFC_KVARG_DATAPATH_EFX "|" \
+           SFC_KVARG_DATAPATH_EF10 "]"
 
 struct sfc_adapter;
 
index f412376..eef4ce0 100644 (file)
@@ -940,6 +940,11 @@ sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
 #endif
 
        info.rxq_entries = rxq_info->entries;
+       info.rxq_hw_ring = rxq->mem.esm_base;
+       info.evq_entries = rxq_info->entries;
+       info.evq_hw_ring = evq->mem.esm_base;
+       info.hw_index = rxq->hw_index;
+       info.mem_bar = sa->mem_bar.esb_base;
 
        rc = sa->dp_rx->qcreate(sa->eth_dev->data->port_id, sw_index,
                                &SFC_DEV_TO_PCI(sa->eth_dev)->addr,