net/sfc: implement Rx queue start and stop operations
authorAndrew Rybchenko <arybchenko@solarflare.com>
Tue, 29 Nov 2016 16:19:20 +0000 (16:19 +0000)
committerFerruh Yigit <ferruh.yigit@intel.com>
Tue, 17 Jan 2017 18:39:26 +0000 (19:39 +0100)
These functions should set the queue state in dev->data->rx_queue_state
array.

Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Andy Moreton <amoreton@solarflare.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
drivers/net/sfc/sfc.c
drivers/net/sfc/sfc_ev.c
drivers/net/sfc/sfc_rx.c
drivers/net/sfc/sfc_rx.h
drivers/net/sfc/sfc_tweak.h [new file with mode: 0644]

index a473b64..4245fb8 100644 (file)
@@ -271,10 +271,17 @@ sfc_start(struct sfc_adapter *sa)
        if (rc != 0)
                goto fail_port_start;
 
+       rc = sfc_rx_start(sa);
+       if (rc != 0)
+               goto fail_rx_start;
+
        sa->state = SFC_ADAPTER_STARTED;
        sfc_log_init(sa, "done");
        return 0;
 
+fail_rx_start:
+       sfc_port_stop(sa);
+
 fail_port_start:
        sfc_ev_stop(sa);
 
@@ -313,6 +320,7 @@ sfc_stop(struct sfc_adapter *sa)
 
        sa->state = SFC_ADAPTER_STOPPING;
 
+       sfc_rx_stop(sa);
        sfc_port_stop(sa);
        sfc_ev_stop(sa);
        sfc_intr_stop(sa);
index ee4cd70..2bcc2be 100644 (file)
@@ -37,6 +37,7 @@
 #include "sfc_debug.h"
 #include "sfc_log.h"
 #include "sfc_ev.h"
+#include "sfc_rx.h"
 
 
 /* Initial delay when waiting for event queue init complete event */
@@ -114,20 +115,30 @@ static boolean_t
 sfc_ev_rxq_flush_done(void *arg, __rte_unused uint32_t rxq_hw_index)
 {
        struct sfc_evq *evq = arg;
+       struct sfc_rxq *rxq;
 
-       sfc_err(evq->sa, "EVQ %u unexpected Rx flush done event",
-               evq->evq_index);
-       return B_TRUE;
+       rxq = evq->rxq;
+       SFC_ASSERT(rxq != NULL);
+       SFC_ASSERT(rxq->hw_index == rxq_hw_index);
+       SFC_ASSERT(rxq->evq == evq);
+       sfc_rx_qflush_done(rxq);
+
+       return B_FALSE;
 }
 
 static boolean_t
 sfc_ev_rxq_flush_failed(void *arg, __rte_unused uint32_t rxq_hw_index)
 {
        struct sfc_evq *evq = arg;
+       struct sfc_rxq *rxq;
 
-       sfc_err(evq->sa, "EVQ %u unexpected Rx flush failed event",
-               evq->evq_index);
-       return B_TRUE;
+       rxq = evq->rxq;
+       SFC_ASSERT(rxq != NULL);
+       SFC_ASSERT(rxq->hw_index == rxq_hw_index);
+       SFC_ASSERT(rxq->evq == evq);
+       sfc_rx_qflush_failed(rxq);
+
+       return B_FALSE;
 }
 
 static boolean_t
index a6b9aea..eae41a0 100644 (file)
  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
+#include <rte_mempool.h>
+
 #include "efx.h"
 
 #include "sfc.h"
 #include "sfc_log.h"
 #include "sfc_ev.h"
 #include "sfc_rx.h"
+#include "sfc_tweak.h"
+
+/*
+ * Maximum number of Rx queue flush attempt in the case of failure or
+ * flush timeout
+ */
+#define SFC_RX_QFLUSH_ATTEMPTS         (3)
+
+/*
+ * Time to wait between event queue polling attempts when waiting for Rx
+ * queue flush done or failed events.
+ */
+#define SFC_RX_QFLUSH_POLL_WAIT_MS     (1)
+
+/*
+ * Maximum number of event queue polling attempts when waiting for Rx queue
+ * flush done or failed events. It defines Rx queue flush attempt timeout
+ * together with SFC_RX_QFLUSH_POLL_WAIT_MS.
+ */
+#define SFC_RX_QFLUSH_POLL_ATTEMPTS    (2000)
+
+void
+sfc_rx_qflush_done(struct sfc_rxq *rxq)
+{
+       rxq->state |= SFC_RXQ_FLUSHED;
+       rxq->state &= ~SFC_RXQ_FLUSHING;
+}
+
+void
+sfc_rx_qflush_failed(struct sfc_rxq *rxq)
+{
+       rxq->state |= SFC_RXQ_FLUSH_FAILED;
+       rxq->state &= ~SFC_RXQ_FLUSHING;
+}
+
+static void
+sfc_rx_qrefill(struct sfc_rxq *rxq)
+{
+       unsigned int free_space;
+       unsigned int bulks;
+       void *objs[SFC_RX_REFILL_BULK];
+       efsys_dma_addr_t addr[RTE_DIM(objs)];
+       unsigned int added = rxq->added;
+       unsigned int id;
+       unsigned int i;
+       struct sfc_rx_sw_desc *rxd;
+       struct rte_mbuf *m;
+       uint8_t port_id = rxq->port_id;
+
+       free_space = EFX_RXQ_LIMIT(rxq->ptr_mask + 1) -
+               (added - rxq->completed);
+       bulks = free_space / RTE_DIM(objs);
+
+       id = added & rxq->ptr_mask;
+       while (bulks-- > 0) {
+               if (rte_mempool_get_bulk(rxq->refill_mb_pool, objs,
+                                        RTE_DIM(objs)) < 0) {
+                       /*
+                        * It is hardly a safe way to increment counter
+                        * from different contexts, but all PMDs do it.
+                        */
+                       rxq->evq->sa->eth_dev->data->rx_mbuf_alloc_failed +=
+                               RTE_DIM(objs);
+                       break;
+               }
+
+               for (i = 0; i < RTE_DIM(objs);
+                    ++i, id = (id + 1) & rxq->ptr_mask) {
+                       m = objs[i];
+
+                       rxd = &rxq->sw_desc[id];
+                       rxd->mbuf = m;
+
+                       rte_mbuf_refcnt_set(m, 1);
+                       m->data_off = RTE_PKTMBUF_HEADROOM;
+                       m->next = NULL;
+                       m->nb_segs = 1;
+                       m->port = port_id;
+
+                       addr[i] = rte_pktmbuf_mtophys(m);
+               }
+
+               efx_rx_qpost(rxq->common, addr, rxq->buf_size,
+                            RTE_DIM(objs), rxq->completed, added);
+               added += RTE_DIM(objs);
+       }
+
+       /* Push doorbell if something is posted */
+       if (rxq->added != added) {
+               rxq->added = added;
+               efx_rx_qpush(rxq->common, added, &rxq->pushed);
+       }
+}
+
+static void
+sfc_rx_qpurge(struct sfc_rxq *rxq)
+{
+       unsigned int i;
+       struct sfc_rx_sw_desc *rxd;
+
+       for (i = rxq->completed; i != rxq->added; ++i) {
+               rxd = &rxq->sw_desc[i & rxq->ptr_mask];
+               rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
+               rxd->mbuf = NULL;
+       }
+}
+
+static void
+sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
+{
+       struct sfc_rxq *rxq;
+       unsigned int retry_count;
+       unsigned int wait_count;
+
+       rxq = sa->rxq_info[sw_index].rxq;
+       SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
+
+       /*
+        * Retry Rx queue flushing in the case of flush failed or
+        * timeout. In the worst case it can delay for 6 seconds.
+        */
+       for (retry_count = 0;
+            ((rxq->state & SFC_RXQ_FLUSHED) == 0) &&
+            (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
+            ++retry_count) {
+               if (efx_rx_qflush(rxq->common) != 0) {
+                       rxq->state |= SFC_RXQ_FLUSH_FAILED;
+                       break;
+               }
+               rxq->state &= ~SFC_RXQ_FLUSH_FAILED;
+               rxq->state |= SFC_RXQ_FLUSHING;
+
+               /*
+                * Wait for Rx queue flush done or failed event at least
+                * SFC_RX_QFLUSH_POLL_WAIT_MS milliseconds and not more
+                * than 2 seconds (SFC_RX_QFLUSH_POLL_WAIT_MS multiplied
+                * by SFC_RX_QFLUSH_POLL_ATTEMPTS).
+                */
+               wait_count = 0;
+               do {
+                       rte_delay_ms(SFC_RX_QFLUSH_POLL_WAIT_MS);
+                       sfc_ev_qpoll(rxq->evq);
+               } while ((rxq->state & SFC_RXQ_FLUSHING) &&
+                        (wait_count++ < SFC_RX_QFLUSH_POLL_ATTEMPTS));
+
+               if (rxq->state & SFC_RXQ_FLUSHING)
+                       sfc_err(sa, "RxQ %u flush timed out", sw_index);
+
+               if (rxq->state & SFC_RXQ_FLUSH_FAILED)
+                       sfc_err(sa, "RxQ %u flush failed", sw_index);
+
+               if (rxq->state & SFC_RXQ_FLUSHED)
+                       sfc_info(sa, "RxQ %u flushed", sw_index);
+       }
+
+       sfc_rx_qpurge(rxq);
+}
+
+int
+sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
+{
+       struct sfc_rxq_info *rxq_info;
+       struct sfc_rxq *rxq;
+       struct sfc_evq *evq;
+       int rc;
+
+       sfc_log_init(sa, "sw_index=%u", sw_index);
+
+       SFC_ASSERT(sw_index < sa->rxq_count);
+
+       rxq_info = &sa->rxq_info[sw_index];
+       rxq = rxq_info->rxq;
+       SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
+
+       evq = rxq->evq;
+
+       rc = sfc_ev_qstart(sa, evq->evq_index);
+       if (rc != 0)
+               goto fail_ev_qstart;
+
+       rc = efx_rx_qcreate(sa->nic, rxq->hw_index, 0, rxq_info->type,
+                           &rxq->mem, rxq_info->entries,
+                           0 /* not used on EF10 */, evq->common,
+                           &rxq->common);
+       if (rc != 0)
+               goto fail_rx_qcreate;
+
+       efx_rx_qenable(rxq->common);
+
+       rxq->pending = rxq->completed = rxq->added = rxq->pushed = 0;
+
+       rxq->state |= SFC_RXQ_STARTED;
+
+       sfc_rx_qrefill(rxq);
+
+       if (sw_index == 0) {
+               rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common,
+                                                   B_FALSE);
+               if (rc != 0)
+                       goto fail_mac_filter_default_rxq_set;
+       }
+
+       /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
+       sa->eth_dev->data->rx_queue_state[sw_index] =
+               RTE_ETH_QUEUE_STATE_STARTED;
+
+       return 0;
+
+fail_mac_filter_default_rxq_set:
+       sfc_rx_qflush(sa, sw_index);
+
+fail_rx_qcreate:
+       sfc_ev_qstop(sa, evq->evq_index);
+
+fail_ev_qstart:
+       return rc;
+}
+
+void
+sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
+{
+       struct sfc_rxq_info *rxq_info;
+       struct sfc_rxq *rxq;
+
+       sfc_log_init(sa, "sw_index=%u", sw_index);
+
+       SFC_ASSERT(sw_index < sa->rxq_count);
+
+       rxq_info = &sa->rxq_info[sw_index];
+       rxq = rxq_info->rxq;
+       SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
+
+       /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
+       sa->eth_dev->data->rx_queue_state[sw_index] =
+               RTE_ETH_QUEUE_STATE_STOPPED;
+
+       if (sw_index == 0)
+               efx_mac_filter_default_rxq_clear(sa->nic);
+
+       sfc_rx_qflush(sa, sw_index);
+
+       rxq->state = SFC_RXQ_INITIALIZED;
+
+       efx_rx_qdestroy(rxq->common);
+
+       sfc_ev_qstop(sa, rxq->evq->evq_index);
+}
 
 static int
 sfc_rx_qcheck_conf(struct sfc_adapter *sa,
@@ -243,6 +492,7 @@ sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
        rxq->refill_mb_pool = mb_pool;
        rxq->buf_size = buf_size;
        rxq->hw_index = sw_index;
+       rxq->port_id = sa->eth_dev->data->port_id;
 
        rxq->state = SFC_RXQ_INITIALIZED;
 
@@ -288,6 +538,53 @@ sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
        rte_free(rxq);
 }
 
+int
+sfc_rx_start(struct sfc_adapter *sa)
+{
+       unsigned int sw_index;
+       int rc;
+
+       sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
+
+       rc = efx_rx_init(sa->nic);
+       if (rc != 0)
+               goto fail_rx_init;
+
+       for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
+               rc = sfc_rx_qstart(sa, sw_index);
+               if (rc != 0)
+                       goto fail_rx_qstart;
+       }
+
+       return 0;
+
+fail_rx_qstart:
+       while (sw_index-- > 0)
+               sfc_rx_qstop(sa, sw_index);
+
+       efx_rx_fini(sa->nic);
+
+fail_rx_init:
+       sfc_log_init(sa, "failed %d", rc);
+       return rc;
+}
+
+void
+sfc_rx_stop(struct sfc_adapter *sa)
+{
+       unsigned int sw_index;
+
+       sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
+
+       sw_index = sa->rxq_count;
+       while (sw_index-- > 0) {
+               if (sa->rxq_info[sw_index].rxq != NULL)
+                       sfc_rx_qstop(sa, sw_index);
+       }
+
+       efx_rx_fini(sa->nic);
+}
+
 static int
 sfc_rx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
 {
index 9719eae..537da9f 100644 (file)
@@ -57,6 +57,14 @@ struct sfc_rx_sw_desc {
 enum sfc_rxq_state_bit {
        SFC_RXQ_INITIALIZED_BIT = 0,
 #define SFC_RXQ_INITIALIZED    (1 << SFC_RXQ_INITIALIZED_BIT)
+       SFC_RXQ_STARTED_BIT,
+#define SFC_RXQ_STARTED                (1 << SFC_RXQ_STARTED_BIT)
+       SFC_RXQ_FLUSHING_BIT,
+#define SFC_RXQ_FLUSHING       (1 << SFC_RXQ_FLUSHING_BIT)
+       SFC_RXQ_FLUSHED_BIT,
+#define SFC_RXQ_FLUSHED                (1 << SFC_RXQ_FLUSHED_BIT)
+       SFC_RXQ_FLUSH_FAILED_BIT,
+#define SFC_RXQ_FLUSH_FAILED   (1 << SFC_RXQ_FLUSH_FAILED_BIT)
 };
 
 /**
@@ -69,8 +77,13 @@ struct sfc_rxq {
        struct sfc_rx_sw_desc   *sw_desc;
        unsigned int            state;
        unsigned int            ptr_mask;
+       unsigned int            pending;
+       unsigned int            completed;
 
        /* Used on refill */
+       unsigned int            added;
+       unsigned int            pushed;
+       uint8_t                 port_id;
        uint16_t                buf_size;
        struct rte_mempool      *refill_mb_pool;
        efx_rxq_t               *common;
@@ -105,12 +118,19 @@ struct sfc_rxq_info {
 
 int sfc_rx_init(struct sfc_adapter *sa);
 void sfc_rx_fini(struct sfc_adapter *sa);
+int sfc_rx_start(struct sfc_adapter *sa);
+void sfc_rx_stop(struct sfc_adapter *sa);
 
 int sfc_rx_qinit(struct sfc_adapter *sa, unsigned int rx_queue_id,
                 uint16_t nb_rx_desc, unsigned int socket_id,
                 const struct rte_eth_rxconf *rx_conf,
                 struct rte_mempool *mb_pool);
 void sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index);
+int sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index);
+void sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index);
+
+void sfc_rx_qflush_done(struct sfc_rxq *rxq);
+void sfc_rx_qflush_failed(struct sfc_rxq *rxq);
 
 #ifdef __cplusplus
 }
diff --git a/drivers/net/sfc/sfc_tweak.h b/drivers/net/sfc/sfc_tweak.h
new file mode 100644 (file)
index 0000000..8538d56
--- /dev/null
@@ -0,0 +1,44 @@
+/*-
+ * 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.
+ */
+
+#ifndef _SFC_TWEAK_H_
+#define _SFC_TWEAK_H_
+
+/*
+ * The header is intended to collect defines/constants which could be
+ * tweaked to improve the PMD performance characteristics depending on
+ * the usecase or requirements (CPU load, packet rate, latency).
+ */
+
+/**
+ * Number of Rx descriptors in the bulk submitted on Rx ring refill.
+ */
+#define SFC_RX_REFILL_BULK     (RTE_CACHE_LINE_SIZE / sizeof(efx_qword_t))
+
+#endif /* _SFC_TWEAK_H_ */