net/sfc: factor out libefx-based Rx datapath
[dpdk.git] / drivers / net / sfc / sfc_ev.c
index 2bcc2be..c6b02f2 100644 (file)
@@ -1,5 +1,7 @@
 /*-
- * Copyright (c) 2016 Solarflare Communications Inc.
+ *   BSD LICENSE
+ *
+ * Copyright (c) 2016-2017 Solarflare Communications Inc.
  * All rights reserved.
  *
  * This software was jointly developed between OKTET Labs (under contract
@@ -30,6 +32,7 @@
 #include <rte_debug.h>
 #include <rte_cycles.h>
 #include <rte_alarm.h>
+#include <rte_branch_prediction.h>
 
 #include "efx.h"
 
@@ -38,6 +41,8 @@
 #include "sfc_log.h"
 #include "sfc_ev.h"
 #include "sfc_rx.h"
+#include "sfc_tx.h"
+#include "sfc_kvargs.h"
 
 
 /* Initial delay when waiting for event queue init complete event */
@@ -66,24 +71,127 @@ sfc_ev_initialized(void *arg)
 }
 
 static boolean_t
-sfc_ev_rx(void *arg, __rte_unused uint32_t label, __rte_unused uint32_t id,
-         __rte_unused uint32_t size, __rte_unused uint16_t flags)
+sfc_ev_nop_rx(void *arg, uint32_t label, uint32_t id,
+             uint32_t size, uint16_t flags)
 {
        struct sfc_evq *evq = arg;
 
-       sfc_err(evq->sa, "EVQ %u unexpected Rx event", evq->evq_index);
+       sfc_err(evq->sa,
+               "EVQ %u unexpected Rx event label=%u id=%#x size=%u flags=%#x",
+               evq->evq_index, label, id, size, flags);
        return B_TRUE;
 }
 
 static boolean_t
-sfc_ev_tx(void *arg, __rte_unused uint32_t label, __rte_unused uint32_t id)
+sfc_ev_efx_rx(void *arg, __rte_unused uint32_t label, uint32_t id,
+             uint32_t size, uint16_t flags)
 {
        struct sfc_evq *evq = arg;
+       struct sfc_efx_rxq *rxq;
+       unsigned int stop;
+       unsigned int pending_id;
+       unsigned int delta;
+       unsigned int i;
+       struct sfc_efx_rx_sw_desc *rxd;
+
+       if (unlikely(evq->exception))
+               goto done;
+
+       rxq = sfc_efx_rxq_by_dp_rxq(evq->dp_rxq);
+
+       SFC_ASSERT(rxq != NULL);
+       SFC_ASSERT(rxq->evq == evq);
+       SFC_ASSERT(rxq->flags & SFC_EFX_RXQ_FLAG_STARTED);
+
+       stop = (id + 1) & rxq->ptr_mask;
+       pending_id = rxq->pending & rxq->ptr_mask;
+       delta = (stop >= pending_id) ? (stop - pending_id) :
+               (rxq->ptr_mask + 1 - pending_id + stop);
+
+       if (delta == 0) {
+               /*
+                * Rx event with no new descriptors done and zero length
+                * is used to abort scattered packet when there is no room
+                * for the tail.
+                */
+               if (unlikely(size != 0)) {
+                       evq->exception = B_TRUE;
+                       sfc_err(evq->sa,
+                               "EVQ %u RxQ %u invalid RX abort "
+                               "(id=%#x size=%u flags=%#x); needs restart",
+                               evq->evq_index, rxq->dp.dpq.queue_id,
+                               id, size, flags);
+                       goto done;
+               }
+
+               /* Add discard flag to the first fragment */
+               rxq->sw_desc[pending_id].flags |= EFX_DISCARD;
+               /* Remove continue flag from the last fragment */
+               rxq->sw_desc[id].flags &= ~EFX_PKT_CONT;
+       } else if (unlikely(delta > rxq->batch_max)) {
+               evq->exception = B_TRUE;
+
+               sfc_err(evq->sa,
+                       "EVQ %u RxQ %u completion out of order "
+                       "(id=%#x delta=%u flags=%#x); needs restart",
+                       evq->evq_index, rxq->dp.dpq.queue_id,
+                       id, delta, flags);
+
+               goto done;
+       }
+
+       for (i = pending_id; i != stop; i = (i + 1) & rxq->ptr_mask) {
+               rxd = &rxq->sw_desc[i];
+
+               rxd->flags = flags;
+
+               SFC_ASSERT(size < (1 << 16));
+               rxd->size = (uint16_t)size;
+       }
+
+       rxq->pending += delta;
 
-       sfc_err(evq->sa, "EVQ %u unexpected Tx event", evq->evq_index);
+done:
+       return B_FALSE;
+}
+
+static boolean_t
+sfc_ev_nop_tx(void *arg, uint32_t label, uint32_t id)
+{
+       struct sfc_evq *evq = arg;
+
+       sfc_err(evq->sa, "EVQ %u unexpected Tx event label=%u id=%#x",
+               evq->evq_index, label, id);
        return B_TRUE;
 }
 
+static boolean_t
+sfc_ev_tx(void *arg, __rte_unused uint32_t label, uint32_t id)
+{
+       struct sfc_evq *evq = arg;
+       struct sfc_txq *txq;
+       unsigned int stop;
+       unsigned int delta;
+
+       txq = evq->txq;
+
+       SFC_ASSERT(txq != NULL);
+       SFC_ASSERT(txq->evq == evq);
+
+       if (unlikely((txq->state & SFC_TXQ_STARTED) == 0))
+               goto done;
+
+       stop = (id + 1) & txq->ptr_mask;
+       id = txq->pending & txq->ptr_mask;
+
+       delta = (stop >= id) ? (stop - id) : (txq->ptr_mask + 1 - id + stop);
+
+       txq->pending += delta;
+
+done:
+       return B_FALSE;
+}
+
 static boolean_t
 sfc_ev_exception(void *arg, __rte_unused uint32_t code,
                 __rte_unused uint32_t data)
@@ -111,13 +219,27 @@ sfc_ev_exception(void *arg, __rte_unused uint32_t code,
        return B_TRUE;
 }
 
+static boolean_t
+sfc_ev_nop_rxq_flush_done(void *arg, uint32_t rxq_hw_index)
+{
+       struct sfc_evq *evq = arg;
+
+       sfc_err(evq->sa, "EVQ %u unexpected RxQ %u flush done",
+               evq->evq_index, rxq_hw_index);
+       return B_TRUE;
+}
+
 static boolean_t
 sfc_ev_rxq_flush_done(void *arg, __rte_unused uint32_t rxq_hw_index)
 {
        struct sfc_evq *evq = arg;
+       struct sfc_dp_rxq *dp_rxq;
        struct sfc_rxq *rxq;
 
-       rxq = evq->rxq;
+       dp_rxq = evq->dp_rxq;
+       SFC_ASSERT(dp_rxq != NULL);
+
+       rxq = sfc_rxq_by_dp_rxq(dp_rxq);
        SFC_ASSERT(rxq != NULL);
        SFC_ASSERT(rxq->hw_index == rxq_hw_index);
        SFC_ASSERT(rxq->evq == evq);
@@ -126,13 +248,27 @@ sfc_ev_rxq_flush_done(void *arg, __rte_unused uint32_t rxq_hw_index)
        return B_FALSE;
 }
 
+static boolean_t
+sfc_ev_nop_rxq_flush_failed(void *arg, uint32_t rxq_hw_index)
+{
+       struct sfc_evq *evq = arg;
+
+       sfc_err(evq->sa, "EVQ %u unexpected RxQ %u flush failed",
+               evq->evq_index, rxq_hw_index);
+       return B_TRUE;
+}
+
 static boolean_t
 sfc_ev_rxq_flush_failed(void *arg, __rte_unused uint32_t rxq_hw_index)
 {
        struct sfc_evq *evq = arg;
+       struct sfc_dp_rxq *dp_rxq;
        struct sfc_rxq *rxq;
 
-       rxq = evq->rxq;
+       dp_rxq = evq->dp_rxq;
+       SFC_ASSERT(dp_rxq != NULL);
+
+       rxq = sfc_rxq_by_dp_rxq(dp_rxq);
        SFC_ASSERT(rxq != NULL);
        SFC_ASSERT(rxq->hw_index == rxq_hw_index);
        SFC_ASSERT(rxq->evq == evq);
@@ -142,15 +278,30 @@ sfc_ev_rxq_flush_failed(void *arg, __rte_unused uint32_t rxq_hw_index)
 }
 
 static boolean_t
-sfc_ev_txq_flush_done(void *arg, __rte_unused uint32_t txq_hw_index)
+sfc_ev_nop_txq_flush_done(void *arg, uint32_t txq_hw_index)
 {
        struct sfc_evq *evq = arg;
 
-       sfc_err(evq->sa, "EVQ %u unexpected Tx flush done event",
-               evq->evq_index);
+       sfc_err(evq->sa, "EVQ %u unexpected TxQ %u flush done",
+               evq->evq_index, txq_hw_index);
        return B_TRUE;
 }
 
+static boolean_t
+sfc_ev_txq_flush_done(void *arg, __rte_unused uint32_t txq_hw_index)
+{
+       struct sfc_evq *evq = arg;
+       struct sfc_txq *txq;
+
+       txq = evq->txq;
+       SFC_ASSERT(txq != NULL);
+       SFC_ASSERT(txq->hw_index == txq_hw_index);
+       SFC_ASSERT(txq->evq == evq);
+       sfc_tx_qflush_done(txq);
+
+       return B_FALSE;
+}
+
 static boolean_t
 sfc_ev_software(void *arg, uint16_t magic)
 {
@@ -191,6 +342,16 @@ sfc_ev_timer(void *arg, uint32_t index)
        return B_TRUE;
 }
 
+static boolean_t
+sfc_ev_nop_link_change(void *arg, __rte_unused efx_link_mode_t link_mode)
+{
+       struct sfc_evq *evq = arg;
+
+       sfc_err(evq->sa, "EVQ %u unexpected link change event",
+               evq->evq_index);
+       return B_TRUE;
+}
+
 static boolean_t
 sfc_ev_link_change(void *arg, efx_link_mode_t link_mode)
 {
@@ -198,28 +359,87 @@ sfc_ev_link_change(void *arg, efx_link_mode_t link_mode)
        struct sfc_adapter *sa = evq->sa;
        struct rte_eth_link *dev_link = &sa->eth_dev->data->dev_link;
        struct rte_eth_link new_link;
+       uint64_t new_link_u64;
+       uint64_t old_link_u64;
 
        EFX_STATIC_ASSERT(sizeof(*dev_link) == sizeof(rte_atomic64_t));
 
        sfc_port_link_mode_to_info(link_mode, &new_link);
-       rte_atomic64_set((rte_atomic64_t *)dev_link, *(uint64_t *)&new_link);
+
+       new_link_u64 = *(uint64_t *)&new_link;
+       do {
+               old_link_u64 = rte_atomic64_read((rte_atomic64_t *)dev_link);
+               if (old_link_u64 == new_link_u64)
+                       break;
+
+               if (rte_atomic64_cmpset((volatile uint64_t *)dev_link,
+                                       old_link_u64, new_link_u64)) {
+                       evq->sa->port.lsc_seq++;
+                       break;
+               }
+       } while (B_TRUE);
 
        return B_FALSE;
 }
 
 static const efx_ev_callbacks_t sfc_ev_callbacks = {
        .eec_initialized        = sfc_ev_initialized,
-       .eec_rx                 = sfc_ev_rx,
-       .eec_tx                 = sfc_ev_tx,
+       .eec_rx                 = sfc_ev_nop_rx,
+       .eec_tx                 = sfc_ev_nop_tx,
+       .eec_exception          = sfc_ev_exception,
+       .eec_rxq_flush_done     = sfc_ev_nop_rxq_flush_done,
+       .eec_rxq_flush_failed   = sfc_ev_nop_rxq_flush_failed,
+       .eec_txq_flush_done     = sfc_ev_nop_txq_flush_done,
+       .eec_software           = sfc_ev_software,
+       .eec_sram               = sfc_ev_sram,
+       .eec_wake_up            = sfc_ev_wake_up,
+       .eec_timer              = sfc_ev_timer,
+       .eec_link_change        = sfc_ev_link_change,
+};
+
+static const efx_ev_callbacks_t sfc_ev_callbacks_efx_rx = {
+       .eec_initialized        = sfc_ev_initialized,
+       .eec_rx                 = sfc_ev_efx_rx,
+       .eec_tx                 = sfc_ev_nop_tx,
        .eec_exception          = sfc_ev_exception,
        .eec_rxq_flush_done     = sfc_ev_rxq_flush_done,
        .eec_rxq_flush_failed   = sfc_ev_rxq_flush_failed,
+       .eec_txq_flush_done     = sfc_ev_nop_txq_flush_done,
+       .eec_software           = sfc_ev_software,
+       .eec_sram               = sfc_ev_sram,
+       .eec_wake_up            = sfc_ev_wake_up,
+       .eec_timer              = sfc_ev_timer,
+       .eec_link_change        = sfc_ev_nop_link_change,
+};
+
+static const efx_ev_callbacks_t sfc_ev_callbacks_dp_rx = {
+       .eec_initialized        = sfc_ev_initialized,
+       .eec_rx                 = sfc_ev_nop_rx,
+       .eec_tx                 = sfc_ev_nop_tx,
+       .eec_exception          = sfc_ev_exception,
+       .eec_rxq_flush_done     = sfc_ev_rxq_flush_done,
+       .eec_rxq_flush_failed   = sfc_ev_rxq_flush_failed,
+       .eec_txq_flush_done     = sfc_ev_nop_txq_flush_done,
+       .eec_software           = sfc_ev_software,
+       .eec_sram               = sfc_ev_sram,
+       .eec_wake_up            = sfc_ev_wake_up,
+       .eec_timer              = sfc_ev_timer,
+       .eec_link_change        = sfc_ev_nop_link_change,
+};
+
+static const efx_ev_callbacks_t sfc_ev_callbacks_tx = {
+       .eec_initialized        = sfc_ev_initialized,
+       .eec_rx                 = sfc_ev_nop_rx,
+       .eec_tx                 = sfc_ev_tx,
+       .eec_exception          = sfc_ev_exception,
+       .eec_rxq_flush_done     = sfc_ev_nop_rxq_flush_done,
+       .eec_rxq_flush_failed   = sfc_ev_nop_rxq_flush_failed,
        .eec_txq_flush_done     = sfc_ev_txq_flush_done,
        .eec_software           = sfc_ev_software,
        .eec_sram               = sfc_ev_sram,
        .eec_wake_up            = sfc_ev_wake_up,
        .eec_timer              = sfc_ev_timer,
-       .eec_link_change        = sfc_ev_link_change,
+       .eec_link_change        = sfc_ev_nop_link_change,
 };
 
 
@@ -231,7 +451,48 @@ sfc_ev_qpoll(struct sfc_evq *evq)
 
        /* Synchronize the DMA memory for reading not required */
 
-       efx_ev_qpoll(evq->common, &evq->read_ptr, &sfc_ev_callbacks, evq);
+       efx_ev_qpoll(evq->common, &evq->read_ptr, evq->callbacks, evq);
+
+       if (unlikely(evq->exception) && sfc_adapter_trylock(evq->sa)) {
+               struct sfc_adapter *sa = evq->sa;
+               int rc;
+
+               if (evq->dp_rxq != NULL) {
+                       unsigned int rxq_sw_index;
+
+                       rxq_sw_index = evq->dp_rxq->dpq.queue_id;
+
+                       sfc_warn(sa,
+                                "restart RxQ %u because of exception on its EvQ %u",
+                                rxq_sw_index, evq->evq_index);
+
+                       sfc_rx_qstop(sa, rxq_sw_index);
+                       rc = sfc_rx_qstart(sa, rxq_sw_index);
+                       if (rc != 0)
+                               sfc_err(sa, "cannot restart RxQ %u",
+                                       rxq_sw_index);
+               }
+
+               if (evq->txq != NULL) {
+                       unsigned int txq_sw_index = sfc_txq_sw_index(evq->txq);
+
+                       sfc_warn(sa,
+                                "restart TxQ %u because of exception on its EvQ %u",
+                                txq_sw_index, evq->evq_index);
+
+                       sfc_tx_qstop(sa, txq_sw_index);
+                       rc = sfc_tx_qstart(sa, txq_sw_index);
+                       if (rc != 0)
+                               sfc_err(sa, "cannot restart TxQ %u",
+                                       txq_sw_index);
+               }
+
+               if (evq->exception)
+                       sfc_panic(sa, "unrecoverable exception on EvQ %u",
+                                 evq->evq_index);
+
+               sfc_adapter_unlock(sa);
+       }
 
        /* Poll-mode driver does not re-prime the event queue for interrupts */
 }
@@ -278,13 +539,23 @@ sfc_ev_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 
        /* Create the common code event queue */
        rc = efx_ev_qcreate(sa->nic, sw_index, esmp, evq_info->entries,
-                           0 /* unused on EF10 */, 0,
-                           EFX_EVQ_FLAGS_TYPE_THROUGHPUT |
-                           EFX_EVQ_FLAGS_NOTIFY_DISABLED,
+                           0 /* unused on EF10 */, 0, evq_info->flags,
                            &evq->common);
        if (rc != 0)
                goto fail_ev_qcreate;
 
+       SFC_ASSERT(evq->dp_rxq == NULL || evq->txq == NULL);
+       if (evq->dp_rxq != 0) {
+               if (strcmp(sa->dp_rx->dp.name, SFC_KVARG_DATAPATH_EFX) == 0)
+                       evq->callbacks = &sfc_ev_callbacks_efx_rx;
+               else
+                       evq->callbacks = &sfc_ev_callbacks_dp_rx;
+       } else if (evq->txq != 0) {
+               evq->callbacks = &sfc_ev_callbacks_tx;
+       } else {
+               evq->callbacks = &sfc_ev_callbacks;
+       }
+
        evq->init_state = SFC_EVQ_STARTING;
 
        /* Wait for the initialization event */
@@ -343,6 +614,7 @@ sfc_ev_qstop(struct sfc_adapter *sa, unsigned int sw_index)
                return;
 
        evq->init_state = SFC_EVQ_INITIALIZED;
+       evq->callbacks = NULL;
        evq->read_ptr = 0;
        evq->exception = B_FALSE;
 
@@ -359,10 +631,14 @@ sfc_ev_mgmt_periodic_qpoll(void *arg)
 
        rc = rte_eal_alarm_set(SFC_MGMT_EV_QPOLL_PERIOD_US,
                               sfc_ev_mgmt_periodic_qpoll, sa);
-       if (rc != 0)
-               sfc_panic(sa,
-                         "cannot rearm management EVQ polling alarm (rc=%d)",
-                         rc);
+       if (rc == -ENOTSUP) {
+               sfc_warn(sa, "alarms are not supported");
+               sfc_warn(sa, "management EVQ must be polled indirectly using no-wait link status update");
+       } else if (rc != 0) {
+               sfc_err(sa,
+                       "cannot rearm management EVQ polling alarm (rc=%d)",
+                       rc);
+       }
 }
 
 static void
@@ -395,6 +671,12 @@ sfc_ev_start(struct sfc_adapter *sa)
        if (rc != 0)
                goto fail_mgmt_evq_start;
 
+       if (sa->intr.lsc_intr) {
+               rc = sfc_ev_qprime(sa->evq_info[sa->mgmt_evq_index].evq);
+               if (rc != 0)
+                       goto fail_evq0_prime;
+       }
+
        rte_spinlock_unlock(&sa->mgmt_evq_lock);
 
        /*
@@ -412,6 +694,9 @@ sfc_ev_start(struct sfc_adapter *sa)
 
        return 0;
 
+fail_evq0_prime:
+       sfc_ev_qstop(sa, 0);
+
 fail_mgmt_evq_start:
        rte_spinlock_unlock(&sa->mgmt_evq_lock);
        efx_ev_fini(sa->nic);
@@ -513,6 +798,28 @@ sfc_ev_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
        SFC_ASSERT(rte_is_power_of_2(max_entries));
 
        evq_info->max_entries = max_entries;
+       evq_info->flags = sa->evq_flags |
+               ((sa->intr.lsc_intr && sw_index == sa->mgmt_evq_index) ?
+                       EFX_EVQ_FLAGS_NOTIFY_INTERRUPT :
+                       EFX_EVQ_FLAGS_NOTIFY_DISABLED);
+
+       return 0;
+}
+
+static int
+sfc_kvarg_perf_profile_handler(__rte_unused const char *key,
+                              const char *value_str, void *opaque)
+{
+       uint64_t *value = opaque;
+
+       if (strcasecmp(value_str, SFC_KVARG_PERF_PROFILE_THROUGHPUT) == 0)
+               *value = EFX_EVQ_FLAGS_TYPE_THROUGHPUT;
+       else if (strcasecmp(value_str, SFC_KVARG_PERF_PROFILE_LOW_LATENCY) == 0)
+               *value = EFX_EVQ_FLAGS_TYPE_LOW_LATENCY;
+       else if (strcasecmp(value_str, SFC_KVARG_PERF_PROFILE_AUTO) == 0)
+               *value = EFX_EVQ_FLAGS_TYPE_AUTO;
+       else
+               return -EINVAL;
 
        return 0;
 }
@@ -533,6 +840,16 @@ sfc_ev_init(struct sfc_adapter *sa)
 
        sfc_log_init(sa, "entry");
 
+       sa->evq_flags = EFX_EVQ_FLAGS_TYPE_THROUGHPUT;
+       rc = sfc_kvargs_process(sa, SFC_KVARG_PERF_PROFILE,
+                               sfc_kvarg_perf_profile_handler,
+                               &sa->evq_flags);
+       if (rc != 0) {
+               sfc_err(sa, "invalid %s parameter value",
+                       SFC_KVARG_PERF_PROFILE);
+               goto fail_kvarg_perf_profile;
+       }
+
        sa->evq_count = sfc_ev_qcount(sa);
        sa->mgmt_evq_index = 0;
        rte_spinlock_init(&sa->mgmt_evq_lock);
@@ -573,6 +890,8 @@ fail_ev_qinit_info:
 
 fail_evqs_alloc:
        sa->evq_count = 0;
+
+fail_kvarg_perf_profile:
        sfc_log_init(sa, "failed %d", rc);
        return rc;
 }