net/sfc: allow control threads for counter queue polling
authorViacheslav Galaktionov <viacheslav.galaktionov@oktetlabs.ru>
Wed, 3 Nov 2021 13:13:31 +0000 (16:13 +0300)
committerFerruh Yigit <ferruh.yigit@intel.com>
Thu, 4 Nov 2021 16:57:00 +0000 (17:57 +0100)
MAE counters can be polled from a control thread if no service core is
allocated for this.

Signed-off-by: Viacheslav Galaktionov <viacheslav.galaktionov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
Reviewed-by: Andy Moreton <amoreton@xilinx.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
doc/guides/rel_notes/release_21_11.rst
drivers/net/sfc/sfc_mae.h
drivers/net/sfc/sfc_mae_counter.c

index f6d07a5..9a27dfe 100644 (file)
@@ -205,6 +205,7 @@ New Features
   * Added port representors support on SN1000 SmartNICs
   * Added flow API transfer proxy support
   * Added SN1000 virtual functions (VF) support
+  * Added support for flow counters without service cores
 
 * **Added power monitor API in vhost library.**
 
index 23dcf1e..2750fbb 100644 (file)
@@ -127,6 +127,13 @@ struct sfc_mae_counters {
        unsigned int                    n_mae_counters;
 };
 
+/** Options for MAE counter polling mode */
+enum sfc_mae_counter_polling_mode {
+       SFC_MAE_COUNTER_POLLING_OFF = 0,
+       SFC_MAE_COUNTER_POLLING_SERVICE,
+       SFC_MAE_COUNTER_POLLING_THREAD,
+};
+
 struct sfc_mae_counter_registry {
        /* Common counter information */
        /** Counters collection */
@@ -143,10 +150,21 @@ struct sfc_mae_counter_registry {
        bool                            use_credits;
 
        /* Information used by configuration routines */
-       /** Counter service core ID */
-       uint32_t                        service_core_id;
-       /** Counter service ID */
-       uint32_t                        service_id;
+       enum sfc_mae_counter_polling_mode polling_mode;
+       union {
+               struct {
+                       /** Counter service core ID */
+                       uint32_t                        core_id;
+                       /** Counter service ID */
+                       uint32_t                        id;
+               } service;
+               struct {
+                       /** Counter thread ID */
+                       pthread_t                       id;
+                       /** The thread should keep running */
+                       bool                            run;
+               } thread;
+       } polling;
 };
 
 /**
index 418caff..e208e97 100644 (file)
 #include "sfc_mae_counter.h"
 #include "sfc_service.h"
 
+/**
+ * Approximate maximum number of counters per packet.
+ * In fact maximum depends on per-counter data offset which is specified
+ * in counter packet header.
+ */
+#define SFC_MAE_COUNTERS_PER_PACKET_MAX \
+       ((SFC_MAE_COUNTER_STREAM_PACKET_SIZE - \
+         ER_RX_SL_PACKETISER_HEADER_WORD_SIZE) / \
+         ER_RX_SL_PACKETISER_PAYLOAD_WORD_SIZE)
+
+/**
+ * Minimum number of Rx buffers in counters only Rx queue.
+ */
+#define SFC_MAE_COUNTER_RXQ_BUFS_MIN \
+       (SFC_COUNTER_RXQ_RX_DESC_COUNT - SFC_COUNTER_RXQ_REFILL_LEVEL)
+
+/**
+ * Approximate number of counter updates fit in counters only Rx queue.
+ * The number is inaccurate since SFC_MAE_COUNTERS_PER_PACKET_MAX is
+ * inaccurate (see above). However, it provides the gist for a number of
+ * counter updates which can fit in an Rx queue after empty poll.
+ *
+ * The define is not actually used, but provides calculations details.
+ */
+#define SFC_MAE_COUNTERS_RXQ_SPACE \
+       (SFC_MAE_COUNTER_RXQ_BUFS_MIN * SFC_MAE_COUNTERS_PER_PACKET_MAX)
+
 static uint32_t
 sfc_mae_counter_get_service_lcore(struct sfc_adapter *sa)
 {
@@ -45,9 +72,6 @@ sfc_mae_counter_rxq_required(struct sfc_adapter *sa)
        if (encp->enc_mae_supported == B_FALSE)
                return false;
 
-       if (sfc_mae_counter_get_service_lcore(sa) == RTE_MAX_LCORE)
-               return false;
-
        return true;
 }
 
@@ -358,9 +382,8 @@ sfc_mae_parse_counter_packet(struct sfc_adapter *sa,
 }
 
 static int32_t
-sfc_mae_counter_routine(void *arg)
+sfc_mae_counter_poll_packets(struct sfc_adapter *sa)
 {
-       struct sfc_adapter *sa = arg;
        struct sfc_mae_counter_registry *counter_registry =
                &sa->mae.counter_registry;
        struct rte_mbuf *mbufs[SFC_MAE_COUNTER_RX_BURST];
@@ -379,7 +402,7 @@ sfc_mae_counter_routine(void *arg)
        rte_pktmbuf_free_bulk(mbufs, n);
 
        if (!counter_registry->use_credits)
-               return 0;
+               return n;
 
        pushed = sfc_rx_get_pushed(sa, counter_registry->rx_dp);
        pushed_diff = pushed - counter_registry->pushed_n_buffers;
@@ -399,9 +422,57 @@ sfc_mae_counter_routine(void *arg)
                }
        }
 
+       return n;
+}
+
+static int32_t
+sfc_mae_counter_service_routine(void *arg)
+{
+       struct sfc_adapter *sa = arg;
+
+       /*
+        * We cannot propagate any errors and we don't need to know
+        * the number of packets we've received.
+        */
+       (void)sfc_mae_counter_poll_packets(sa);
+
        return 0;
 }
 
+static void *
+sfc_mae_counter_thread(void *data)
+{
+       struct sfc_adapter *sa = data;
+       struct sfc_mae_counter_registry *counter_registry =
+               &sa->mae.counter_registry;
+       int32_t rc;
+
+       while (__atomic_load_n(&counter_registry->polling.thread.run,
+                              __ATOMIC_ACQUIRE)) {
+               rc = sfc_mae_counter_poll_packets(sa);
+               if (rc == 0) {
+                       /*
+                        * The queue is empty. Do not burn CPU.
+                        * An empty queue has just enough space for about
+                        * SFC_MAE_COUNTERS_RXQ_SPACE counter updates which is
+                        * more than 100K, so we can sleep a bit. The queue uses
+                        * a credit-based flow control anyway, so firmware will
+                        * not enqueue more counter updates until the host
+                        * supplies it with additional credits. The counters are
+                        * 48bits wide, so the timeout need only be short enough
+                        * to ensure that the counter values do not overflow
+                        * before the next counter update. Also we should not
+                        * delay counter updates for a long time, otherwise
+                        * application may decide that flow is idle and should
+                        * be removed.
+                        */
+                       rte_delay_ms(1);
+               }
+       }
+
+       return NULL;
+}
+
 static void
 sfc_mae_counter_service_unregister(struct sfc_adapter *sa)
 {
@@ -410,15 +481,15 @@ sfc_mae_counter_service_unregister(struct sfc_adapter *sa)
        const unsigned int wait_ms = 10000;
        unsigned int i;
 
-       rte_service_runstate_set(registry->service_id, 0);
-       rte_service_component_runstate_set(registry->service_id, 0);
+       rte_service_runstate_set(registry->polling.service.id, 0);
+       rte_service_component_runstate_set(registry->polling.service.id, 0);
 
        /*
         * Wait for the counter routine to finish the last iteration.
         * Give up on timeout.
         */
        for (i = 0; i < wait_ms; i++) {
-               if (rte_service_may_be_active(registry->service_id) == 0)
+               if (rte_service_may_be_active(registry->polling.service.id) == 0)
                        break;
 
                rte_delay_ms(1);
@@ -426,10 +497,10 @@ sfc_mae_counter_service_unregister(struct sfc_adapter *sa)
        if (i == wait_ms)
                sfc_warn(sa, "failed to wait for counter service to stop");
 
-       rte_service_map_lcore_set(registry->service_id,
-                                 registry->service_core_id, 0);
+       rte_service_map_lcore_set(registry->polling.service.id,
+                                 registry->polling.service.core_id, 0);
 
-       rte_service_component_unregister(registry->service_id);
+       rte_service_component_unregister(registry->polling.service.id);
 }
 
 static struct sfc_rxq_info *
@@ -438,6 +509,18 @@ sfc_counter_rxq_info_get(struct sfc_adapter *sa)
        return &sfc_sa2shared(sa)->rxq_info[sa->counter_rxq.sw_index];
 }
 
+static void
+sfc_mae_counter_registry_prepare(struct sfc_mae_counter_registry *registry,
+                                struct sfc_adapter *sa,
+                                uint32_t counter_stream_flags)
+{
+       registry->rx_pkt_burst = sa->eth_dev->rx_pkt_burst;
+       registry->rx_dp = sfc_counter_rxq_info_get(sa)->dp;
+       registry->pushed_n_buffers = 0;
+       registry->use_credits = counter_stream_flags &
+               EFX_MAE_COUNTERS_STREAM_OUT_USES_CREDITS;
+}
+
 static int
 sfc_mae_counter_service_register(struct sfc_adapter *sa,
                                 uint32_t counter_stream_flags)
@@ -456,13 +539,10 @@ sfc_mae_counter_service_register(struct sfc_adapter *sa,
        memset(&service, 0, sizeof(service));
        rte_strscpy(service.name, counter_service_name, sizeof(service.name));
        service.socket_id = sa->socket_id;
-       service.callback = sfc_mae_counter_routine;
+       service.callback = sfc_mae_counter_service_routine;
        service.callback_userdata = sa;
-       counter_registry->rx_pkt_burst = sa->eth_dev->rx_pkt_burst;
-       counter_registry->rx_dp = sfc_counter_rxq_info_get(sa)->dp;
-       counter_registry->pushed_n_buffers = 0;
-       counter_registry->use_credits = counter_stream_flags &
-               EFX_MAE_COUNTERS_STREAM_OUT_USES_CREDITS;
+       sfc_mae_counter_registry_prepare(counter_registry, sa,
+                                        counter_stream_flags);
 
        cid = sfc_get_service_lcore(sa->socket_id);
        if (cid == RTE_MAX_LCORE && sa->socket_id != SOCKET_ID_ANY) {
@@ -520,8 +600,9 @@ sfc_mae_counter_service_register(struct sfc_adapter *sa,
                goto fail_runstate_set;
        }
 
-       counter_registry->service_core_id = cid;
-       counter_registry->service_id = sid;
+       counter_registry->polling_mode = SFC_MAE_COUNTER_POLLING_SERVICE;
+       counter_registry->polling.service.core_id = cid;
+       counter_registry->polling.service.id = sid;
 
        sfc_log_init(sa, "done");
 
@@ -544,6 +625,47 @@ fail_get_service_lcore:
        return rc;
 }
 
+static void
+sfc_mae_counter_thread_stop(struct sfc_adapter *sa)
+{
+       struct sfc_mae_counter_registry *counter_registry =
+               &sa->mae.counter_registry;
+       int rc;
+
+       /* Ensure that flag is set before attempting to join thread */
+       __atomic_store_n(&counter_registry->polling.thread.run, false,
+                        __ATOMIC_RELEASE);
+
+       rc = pthread_join(counter_registry->polling.thread.id, NULL);
+       if (rc != 0)
+               sfc_err(sa, "failed to join the MAE counter polling thread");
+
+       counter_registry->polling_mode = SFC_MAE_COUNTER_POLLING_OFF;
+}
+
+static int
+sfc_mae_counter_thread_spawn(struct sfc_adapter *sa,
+                            uint32_t counter_stream_flags)
+{
+       struct sfc_mae_counter_registry *counter_registry =
+               &sa->mae.counter_registry;
+       int rc;
+
+       sfc_log_init(sa, "entry");
+
+       sfc_mae_counter_registry_prepare(counter_registry, sa,
+                                        counter_stream_flags);
+
+       counter_registry->polling_mode = SFC_MAE_COUNTER_POLLING_THREAD;
+       counter_registry->polling.thread.run = true;
+
+       rc = rte_ctrl_thread_create(&sa->mae.counter_registry.polling.thread.id,
+                                   "mae_counter_thread", NULL,
+                                   sfc_mae_counter_thread, sa);
+
+       return rc;
+}
+
 int
 sfc_mae_counters_init(struct sfc_mae_counters *counters,
                      uint32_t nb_counters_max)
@@ -754,7 +876,15 @@ sfc_mae_counter_stop(struct sfc_adapter *sa)
                return;
        }
 
-       sfc_mae_counter_service_unregister(sa);
+       SFC_ASSERT(mae->counter_registry.polling_mode !=
+                       SFC_MAE_COUNTER_POLLING_OFF);
+
+       if (mae->counter_registry.polling_mode ==
+                       SFC_MAE_COUNTER_POLLING_SERVICE)
+               sfc_mae_counter_service_unregister(sa);
+       else
+               sfc_mae_counter_thread_stop(sa);
+
        efx_mae_counters_stream_stop(sa->nic, sa->counter_rxq.sw_index, NULL);
 
        mae->counter_rxq_running = false;
@@ -787,15 +917,22 @@ sfc_mae_counter_start(struct sfc_adapter *sa)
 
        sfc_log_init(sa, "stream start flags: 0x%x", flags);
 
-       rc = sfc_mae_counter_service_register(sa, flags);
-       if (rc != 0)
-               goto fail_service_register;
+       if (sfc_mae_counter_get_service_lcore(sa) != RTE_MAX_LCORE) {
+               rc = sfc_mae_counter_service_register(sa, flags);
+               if (rc != 0)
+                       goto fail_service_register;
+       } else {
+               rc = sfc_mae_counter_thread_spawn(sa, flags);
+               if (rc != 0)
+                       goto fail_thread_spawn;
+       }
 
        mae->counter_rxq_running = true;
 
        return 0;
 
 fail_service_register:
+fail_thread_spawn:
        efx_mae_counters_stream_stop(sa->nic, sa->counter_rxq.sw_index, NULL);
 
 fail_counter_stream: