X-Git-Url: http://git.droids-corp.org/?a=blobdiff_plain;f=drivers%2Fnet%2Fsfc%2Fsfc_rx.c;h=eae41a0167d03bde0081c1a38b18635d8f972fab;hb=28944ac098aa;hp=88e33198f8709ff766b66849a747ffe2e8828dcf;hpb=976f2e5a0cebc1c60d9ff5b5608e9ad8e547510c;p=dpdk.git

diff --git a/drivers/net/sfc/sfc_rx.c b/drivers/net/sfc/sfc_rx.c
index 88e33198f8..eae41a0167 100644
--- a/drivers/net/sfc/sfc_rx.c
+++ b/drivers/net/sfc/sfc_rx.c
@@ -27,11 +27,563 @@
  * 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,
+		   const struct rte_eth_rxconf *rx_conf)
+{
+	int rc = 0;
+
+	if (rx_conf->rx_thresh.pthresh != 0 ||
+	    rx_conf->rx_thresh.hthresh != 0 ||
+	    rx_conf->rx_thresh.wthresh != 0) {
+		sfc_err(sa,
+			"RxQ prefetch/host/writeback thresholds are not supported");
+		rc = EINVAL;
+	}
+
+	if (rx_conf->rx_free_thresh != 0) {
+		sfc_err(sa, "RxQ free threshold is not supported");
+		rc = EINVAL;
+	}
+
+	if (rx_conf->rx_drop_en == 0) {
+		sfc_err(sa, "RxQ drop disable is not supported");
+		rc = EINVAL;
+	}
+
+	if (rx_conf->rx_deferred_start != 0) {
+		sfc_err(sa, "RxQ deferred start is not supported");
+		rc = EINVAL;
+	}
+
+	return rc;
+}
+
+static unsigned int
+sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
+{
+	uint32_t data_off;
+	uint32_t order;
+
+	/* The mbuf object itself is always cache line aligned */
+	order = rte_bsf32(RTE_CACHE_LINE_SIZE);
+
+	/* Data offset from mbuf object start */
+	data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
+		RTE_PKTMBUF_HEADROOM;
+
+	order = MIN(order, rte_bsf32(data_off));
+
+	return 1u << (order - 1);
+}
+
+static uint16_t
+sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
+{
+	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
+	const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
+	const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
+	uint16_t buf_size;
+	unsigned int buf_aligned;
+	unsigned int start_alignment;
+	unsigned int end_padding_alignment;
+
+	/* Below it is assumed that both alignments are power of 2 */
+	SFC_ASSERT(rte_is_power_of_2(nic_align_start));
+	SFC_ASSERT(rte_is_power_of_2(nic_align_end));
+
+	/*
+	 * mbuf is always cache line aligned, double-check
+	 * that it meets rx buffer start alignment requirements.
+	 */
+
+	/* Start from mbuf pool data room size */
+	buf_size = rte_pktmbuf_data_room_size(mb_pool);
+
+	/* Remove headroom */
+	if (buf_size <= RTE_PKTMBUF_HEADROOM) {
+		sfc_err(sa,
+			"RxQ mbuf pool %s object data room size %u is smaller than headroom %u",
+			mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
+		return 0;
+	}
+	buf_size -= RTE_PKTMBUF_HEADROOM;
+
+	/* Calculate guaranteed data start alignment */
+	buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
+
+	/* Reserve space for start alignment */
+	if (buf_aligned < nic_align_start) {
+		start_alignment = nic_align_start - buf_aligned;
+		if (buf_size <= start_alignment) {
+			sfc_err(sa,
+				"RxQ mbuf pool %s object data room size %u is insufficient for headroom %u and buffer start alignment %u required by NIC",
+				mb_pool->name,
+				rte_pktmbuf_data_room_size(mb_pool),
+				RTE_PKTMBUF_HEADROOM, start_alignment);
+			return 0;
+		}
+		buf_aligned = nic_align_start;
+		buf_size -= start_alignment;
+	} else {
+		start_alignment = 0;
+	}
+
+	/* Make sure that end padding does not write beyond the buffer */
+	if (buf_aligned < nic_align_end) {
+		/*
+		 * Estimate space which can be lost. If guarnteed buffer
+		 * size is odd, lost space is (nic_align_end - 1). More
+		 * accurate formula is below.
+		 */
+		end_padding_alignment = nic_align_end -
+			MIN(buf_aligned, 1u << (rte_bsf32(buf_size) - 1));
+		if (buf_size <= end_padding_alignment) {
+			sfc_err(sa,
+				"RxQ mbuf pool %s object data room size %u is insufficient for headroom %u, buffer start alignment %u and end padding alignment %u required by NIC",
+				mb_pool->name,
+				rte_pktmbuf_data_room_size(mb_pool),
+				RTE_PKTMBUF_HEADROOM, start_alignment,
+				end_padding_alignment);
+			return 0;
+		}
+		buf_size -= end_padding_alignment;
+	} else {
+		/*
+		 * Start is aligned the same or better than end,
+		 * just align length.
+		 */
+		buf_size = P2ALIGN(buf_size, nic_align_end);
+	}
+
+	return buf_size;
+}
+
+int
+sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
+	     uint16_t nb_rx_desc, unsigned int socket_id,
+	     const struct rte_eth_rxconf *rx_conf,
+	     struct rte_mempool *mb_pool)
+{
+	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
+	int rc;
+	uint16_t buf_size;
+	struct sfc_rxq_info *rxq_info;
+	unsigned int evq_index;
+	struct sfc_evq *evq;
+	struct sfc_rxq *rxq;
+
+	rc = sfc_rx_qcheck_conf(sa, rx_conf);
+	if (rc != 0)
+		goto fail_bad_conf;
+
+	buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
+	if (buf_size == 0) {
+		sfc_err(sa, "RxQ %u mbuf pool object size is too small",
+			sw_index);
+		rc = EINVAL;
+		goto fail_bad_conf;
+	}
+
+	if ((buf_size < sa->port.pdu + encp->enc_rx_prefix_size) &&
+	    !sa->eth_dev->data->dev_conf.rxmode.enable_scatter) {
+		sfc_err(sa, "Rx scatter is disabled and RxQ %u mbuf pool "
+			"object size is too small", sw_index);
+		sfc_err(sa, "RxQ %u calculated Rx buffer size is %u vs "
+			"PDU size %u plus Rx prefix %u bytes",
+			sw_index, buf_size, (unsigned int)sa->port.pdu,
+			encp->enc_rx_prefix_size);
+		rc = EINVAL;
+		goto fail_bad_conf;
+	}
+
+	SFC_ASSERT(sw_index < sa->rxq_count);
+	rxq_info = &sa->rxq_info[sw_index];
+
+	SFC_ASSERT(nb_rx_desc <= rxq_info->max_entries);
+	rxq_info->entries = nb_rx_desc;
+	rxq_info->type = EFX_RXQ_TYPE_DEFAULT;
+
+	evq_index = sfc_evq_index_by_rxq_sw_index(sa, sw_index);
+
+	rc = sfc_ev_qinit(sa, evq_index, rxq_info->entries, socket_id);
+	if (rc != 0)
+		goto fail_ev_qinit;
+
+	evq = sa->evq_info[evq_index].evq;
+
+	rc = ENOMEM;
+	rxq = rte_zmalloc_socket("sfc-rxq", sizeof(*rxq), RTE_CACHE_LINE_SIZE,
+				 socket_id);
+	if (rxq == NULL)
+		goto fail_rxq_alloc;
+
+	rc = sfc_dma_alloc(sa, "rxq", sw_index, EFX_RXQ_SIZE(rxq_info->entries),
+			   socket_id, &rxq->mem);
+	if (rc != 0)
+		goto fail_dma_alloc;
+
+	rc = ENOMEM;
+	rxq->sw_desc = rte_calloc_socket("sfc-rxq-sw_desc", rxq_info->entries,
+					 sizeof(*rxq->sw_desc),
+					 RTE_CACHE_LINE_SIZE, socket_id);
+	if (rxq->sw_desc == NULL)
+		goto fail_desc_alloc;
+
+	evq->rxq = rxq;
+	rxq->evq = evq;
+	rxq->ptr_mask = rxq_info->entries - 1;
+	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;
+
+	rxq_info->rxq = rxq;
+
+	return 0;
+
+fail_desc_alloc:
+	sfc_dma_free(sa, &rxq->mem);
+
+fail_dma_alloc:
+	rte_free(rxq);
+
+fail_rxq_alloc:
+	sfc_ev_qfini(sa, evq_index);
+
+fail_ev_qinit:
+	rxq_info->entries = 0;
+
+fail_bad_conf:
+	sfc_log_init(sa, "failed %d", rc);
+	return rc;
+}
+
+void
+sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
+{
+	struct sfc_rxq_info *rxq_info;
+	struct sfc_rxq *rxq;
+
+	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);
+
+	rxq_info->rxq = NULL;
+	rxq_info->entries = 0;
+
+	rte_free(rxq->sw_desc);
+	sfc_dma_free(sa, &rxq->mem);
+	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)
@@ -158,6 +710,14 @@ fail_check_mode:
 void
 sfc_rx_fini(struct sfc_adapter *sa)
 {
+	unsigned int sw_index;
+
+	sw_index = sa->rxq_count;
+	while (sw_index-- > 0) {
+		if (sa->rxq_info[sw_index].rxq != NULL)
+			sfc_rx_qfini(sa, sw_index);
+	}
+
 	rte_free(sa->rxq_info);
 	sa->rxq_info = NULL;
 	sa->rxq_count = 0;