2 * Copyright (c) 2016 Solarflare Communications Inc.
5 * This software was jointly developed between OKTET Labs (under contract
6 * for Solarflare) and Solarflare Communications, Inc.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions are met:
11 * 1. Redistributions of source code must retain the above copyright notice,
12 * this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright notice,
14 * this list of conditions and the following disclaimer in the documentation
15 * and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
18 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
19 * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
24 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
25 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
26 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
27 * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30 #include <rte_mempool.h>
35 #include "sfc_debug.h"
39 #include "sfc_tweak.h"
42 * Maximum number of Rx queue flush attempt in the case of failure or
45 #define SFC_RX_QFLUSH_ATTEMPTS (3)
48 * Time to wait between event queue polling attempts when waiting for Rx
49 * queue flush done or failed events.
51 #define SFC_RX_QFLUSH_POLL_WAIT_MS (1)
54 * Maximum number of event queue polling attempts when waiting for Rx queue
55 * flush done or failed events. It defines Rx queue flush attempt timeout
56 * together with SFC_RX_QFLUSH_POLL_WAIT_MS.
58 #define SFC_RX_QFLUSH_POLL_ATTEMPTS (2000)
61 sfc_rx_qflush_done(struct sfc_rxq *rxq)
63 rxq->state |= SFC_RXQ_FLUSHED;
64 rxq->state &= ~SFC_RXQ_FLUSHING;
68 sfc_rx_qflush_failed(struct sfc_rxq *rxq)
70 rxq->state |= SFC_RXQ_FLUSH_FAILED;
71 rxq->state &= ~SFC_RXQ_FLUSHING;
75 sfc_rx_qrefill(struct sfc_rxq *rxq)
77 unsigned int free_space;
79 void *objs[SFC_RX_REFILL_BULK];
80 efsys_dma_addr_t addr[RTE_DIM(objs)];
81 unsigned int added = rxq->added;
84 struct sfc_rx_sw_desc *rxd;
86 uint8_t port_id = rxq->port_id;
88 free_space = EFX_RXQ_LIMIT(rxq->ptr_mask + 1) -
89 (added - rxq->completed);
91 if (free_space < rxq->refill_threshold)
94 bulks = free_space / RTE_DIM(objs);
96 id = added & rxq->ptr_mask;
98 if (rte_mempool_get_bulk(rxq->refill_mb_pool, objs,
101 * It is hardly a safe way to increment counter
102 * from different contexts, but all PMDs do it.
104 rxq->evq->sa->eth_dev->data->rx_mbuf_alloc_failed +=
109 for (i = 0; i < RTE_DIM(objs);
110 ++i, id = (id + 1) & rxq->ptr_mask) {
113 rxd = &rxq->sw_desc[id];
116 rte_mbuf_refcnt_set(m, 1);
117 m->data_off = RTE_PKTMBUF_HEADROOM;
122 addr[i] = rte_pktmbuf_mtophys(m);
125 efx_rx_qpost(rxq->common, addr, rxq->buf_size,
126 RTE_DIM(objs), rxq->completed, added);
127 added += RTE_DIM(objs);
130 /* Push doorbell if something is posted */
131 if (rxq->added != added) {
133 efx_rx_qpush(rxq->common, added, &rxq->pushed);
138 sfc_rx_desc_flags_to_offload_flags(const unsigned int desc_flags)
140 uint64_t mbuf_flags = 0;
142 switch (desc_flags & (EFX_PKT_IPV4 | EFX_CKSUM_IPV4)) {
143 case (EFX_PKT_IPV4 | EFX_CKSUM_IPV4):
144 mbuf_flags |= PKT_RX_IP_CKSUM_GOOD;
147 mbuf_flags |= PKT_RX_IP_CKSUM_BAD;
150 RTE_BUILD_BUG_ON(PKT_RX_IP_CKSUM_UNKNOWN != 0);
151 SFC_ASSERT((mbuf_flags & PKT_RX_IP_CKSUM_MASK) ==
152 PKT_RX_IP_CKSUM_UNKNOWN);
156 switch ((desc_flags &
157 (EFX_PKT_TCP | EFX_PKT_UDP | EFX_CKSUM_TCPUDP))) {
158 case (EFX_PKT_TCP | EFX_CKSUM_TCPUDP):
159 case (EFX_PKT_UDP | EFX_CKSUM_TCPUDP):
160 mbuf_flags |= PKT_RX_L4_CKSUM_GOOD;
164 mbuf_flags |= PKT_RX_L4_CKSUM_BAD;
167 RTE_BUILD_BUG_ON(PKT_RX_L4_CKSUM_UNKNOWN != 0);
168 SFC_ASSERT((mbuf_flags & PKT_RX_L4_CKSUM_MASK) ==
169 PKT_RX_L4_CKSUM_UNKNOWN);
177 sfc_rx_desc_flags_to_packet_type(const unsigned int desc_flags)
179 return RTE_PTYPE_L2_ETHER |
180 ((desc_flags & EFX_PKT_IPV4) ?
181 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN : 0) |
182 ((desc_flags & EFX_PKT_IPV6) ?
183 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN : 0) |
184 ((desc_flags & EFX_PKT_TCP) ? RTE_PTYPE_L4_TCP : 0) |
185 ((desc_flags & EFX_PKT_UDP) ? RTE_PTYPE_L4_UDP : 0);
189 sfc_rx_set_rss_hash(struct sfc_rxq *rxq, unsigned int flags, struct rte_mbuf *m)
191 #if EFSYS_OPT_RX_SCALE
195 if ((rxq->flags & SFC_RXQ_RSS_HASH) == 0)
198 mbuf_data = rte_pktmbuf_mtod(m, uint8_t *);
200 if (flags & (EFX_PKT_IPV4 | EFX_PKT_IPV6)) {
201 m->hash.rss = efx_pseudo_hdr_hash_get(rxq->common,
202 EFX_RX_HASHALG_TOEPLITZ,
205 m->ol_flags |= PKT_RX_RSS_HASH;
211 sfc_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
213 struct sfc_rxq *rxq = rx_queue;
214 unsigned int completed;
215 unsigned int prefix_size = rxq->prefix_size;
216 unsigned int done_pkts = 0;
217 boolean_t discard_next = B_FALSE;
218 struct rte_mbuf *scatter_pkt = NULL;
220 if (unlikely((rxq->state & SFC_RXQ_RUNNING) == 0))
223 sfc_ev_qpoll(rxq->evq);
225 completed = rxq->completed;
226 while (completed != rxq->pending && done_pkts < nb_pkts) {
228 struct sfc_rx_sw_desc *rxd;
230 unsigned int seg_len;
231 unsigned int desc_flags;
233 id = completed++ & rxq->ptr_mask;
234 rxd = &rxq->sw_desc[id];
236 desc_flags = rxd->flags;
241 if (desc_flags & (EFX_ADDR_MISMATCH | EFX_DISCARD))
244 if (desc_flags & EFX_PKT_PREFIX_LEN) {
248 rc = efx_pseudo_hdr_pkt_length_get(rxq->common,
249 rte_pktmbuf_mtod(m, uint8_t *), &tmp_size);
253 seg_len = rxd->size - prefix_size;
256 rte_pktmbuf_data_len(m) = seg_len;
257 rte_pktmbuf_pkt_len(m) = seg_len;
259 if (scatter_pkt != NULL) {
260 if (rte_pktmbuf_chain(scatter_pkt, m) != 0) {
261 rte_mempool_put(rxq->refill_mb_pool,
265 /* The packet to deliver */
269 if (desc_flags & EFX_PKT_CONT) {
270 /* The packet is scattered, more fragments to come */
272 /* Futher fragments have no prefix */
277 /* Scattered packet is done */
279 /* The first fragment of the packet has prefix */
280 prefix_size = rxq->prefix_size;
282 m->ol_flags = sfc_rx_desc_flags_to_offload_flags(desc_flags);
283 m->packet_type = sfc_rx_desc_flags_to_packet_type(desc_flags);
286 * Extract RSS hash from the packet prefix and
287 * set the corresponding field (if needed and possible)
289 sfc_rx_set_rss_hash(rxq, desc_flags, m);
291 m->data_off += prefix_size;
298 discard_next = ((desc_flags & EFX_PKT_CONT) != 0);
299 rte_mempool_put(rxq->refill_mb_pool, m);
303 /* pending is only moved when entire packet is received */
304 SFC_ASSERT(scatter_pkt == NULL);
306 rxq->completed = completed;
314 sfc_rx_qdesc_npending(struct sfc_adapter *sa, unsigned int sw_index)
318 SFC_ASSERT(sw_index < sa->rxq_count);
319 rxq = sa->rxq_info[sw_index].rxq;
321 if (rxq == NULL || (rxq->state & SFC_RXQ_RUNNING) == 0)
324 sfc_ev_qpoll(rxq->evq);
326 return rxq->pending - rxq->completed;
330 sfc_rx_qdesc_done(struct sfc_rxq *rxq, unsigned int offset)
332 if ((rxq->state & SFC_RXQ_RUNNING) == 0)
335 sfc_ev_qpoll(rxq->evq);
337 return offset < (rxq->pending - rxq->completed);
341 sfc_rx_qpurge(struct sfc_rxq *rxq)
344 struct sfc_rx_sw_desc *rxd;
346 for (i = rxq->completed; i != rxq->added; ++i) {
347 rxd = &rxq->sw_desc[i & rxq->ptr_mask];
348 rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
354 sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
357 unsigned int retry_count;
358 unsigned int wait_count;
360 rxq = sa->rxq_info[sw_index].rxq;
361 SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
364 * Retry Rx queue flushing in the case of flush failed or
365 * timeout. In the worst case it can delay for 6 seconds.
367 for (retry_count = 0;
368 ((rxq->state & SFC_RXQ_FLUSHED) == 0) &&
369 (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
371 if (efx_rx_qflush(rxq->common) != 0) {
372 rxq->state |= SFC_RXQ_FLUSH_FAILED;
375 rxq->state &= ~SFC_RXQ_FLUSH_FAILED;
376 rxq->state |= SFC_RXQ_FLUSHING;
379 * Wait for Rx queue flush done or failed event at least
380 * SFC_RX_QFLUSH_POLL_WAIT_MS milliseconds and not more
381 * than 2 seconds (SFC_RX_QFLUSH_POLL_WAIT_MS multiplied
382 * by SFC_RX_QFLUSH_POLL_ATTEMPTS).
386 rte_delay_ms(SFC_RX_QFLUSH_POLL_WAIT_MS);
387 sfc_ev_qpoll(rxq->evq);
388 } while ((rxq->state & SFC_RXQ_FLUSHING) &&
389 (wait_count++ < SFC_RX_QFLUSH_POLL_ATTEMPTS));
391 if (rxq->state & SFC_RXQ_FLUSHING)
392 sfc_err(sa, "RxQ %u flush timed out", sw_index);
394 if (rxq->state & SFC_RXQ_FLUSH_FAILED)
395 sfc_err(sa, "RxQ %u flush failed", sw_index);
397 if (rxq->state & SFC_RXQ_FLUSHED)
398 sfc_info(sa, "RxQ %u flushed", sw_index);
405 sfc_rx_default_rxq_set_filter(struct sfc_adapter *sa, struct sfc_rxq *rxq)
407 boolean_t rss = (sa->rss_channels > 1) ? B_TRUE : B_FALSE;
408 struct sfc_port *port = &sa->port;
412 * If promiscuous or all-multicast mode has been requested, setting
413 * filter for the default Rx queue might fail, in particular, while
414 * running over PCI function which is not a member of corresponding
415 * privilege groups; if this occurs, few iterations will be made to
416 * repeat this step without promiscuous and all-multicast flags set
419 rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common, rss);
422 else if (rc != EOPNOTSUPP)
426 sfc_warn(sa, "promiscuous mode has been requested, "
427 "but the HW rejects it");
428 sfc_warn(sa, "promiscuous mode will be disabled");
430 port->promisc = B_FALSE;
431 rc = sfc_set_rx_mode(sa);
438 if (port->allmulti) {
439 sfc_warn(sa, "all-multicast mode has been requested, "
440 "but the HW rejects it");
441 sfc_warn(sa, "all-multicast mode will be disabled");
443 port->allmulti = B_FALSE;
444 rc = sfc_set_rx_mode(sa);
455 sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
457 struct sfc_rxq_info *rxq_info;
462 sfc_log_init(sa, "sw_index=%u", sw_index);
464 SFC_ASSERT(sw_index < sa->rxq_count);
466 rxq_info = &sa->rxq_info[sw_index];
468 SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
472 rc = sfc_ev_qstart(sa, evq->evq_index);
476 rc = efx_rx_qcreate(sa->nic, rxq->hw_index, 0, rxq_info->type,
477 &rxq->mem, rxq_info->entries,
478 0 /* not used on EF10 */, evq->common,
481 goto fail_rx_qcreate;
483 efx_rx_qenable(rxq->common);
485 rxq->pending = rxq->completed = rxq->added = rxq->pushed = 0;
487 rxq->state |= (SFC_RXQ_STARTED | SFC_RXQ_RUNNING);
492 rc = sfc_rx_default_rxq_set_filter(sa, rxq);
494 goto fail_mac_filter_default_rxq_set;
497 /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
498 sa->eth_dev->data->rx_queue_state[sw_index] =
499 RTE_ETH_QUEUE_STATE_STARTED;
503 fail_mac_filter_default_rxq_set:
504 sfc_rx_qflush(sa, sw_index);
507 sfc_ev_qstop(sa, evq->evq_index);
514 sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
516 struct sfc_rxq_info *rxq_info;
519 sfc_log_init(sa, "sw_index=%u", sw_index);
521 SFC_ASSERT(sw_index < sa->rxq_count);
523 rxq_info = &sa->rxq_info[sw_index];
526 if (rxq->state == SFC_RXQ_INITIALIZED)
528 SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
530 /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
531 sa->eth_dev->data->rx_queue_state[sw_index] =
532 RTE_ETH_QUEUE_STATE_STOPPED;
534 rxq->state &= ~SFC_RXQ_RUNNING;
537 efx_mac_filter_default_rxq_clear(sa->nic);
539 sfc_rx_qflush(sa, sw_index);
541 rxq->state = SFC_RXQ_INITIALIZED;
543 efx_rx_qdestroy(rxq->common);
545 sfc_ev_qstop(sa, rxq->evq->evq_index);
549 sfc_rx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_rx_desc,
550 const struct rte_eth_rxconf *rx_conf)
552 const uint16_t rx_free_thresh_max = EFX_RXQ_LIMIT(nb_rx_desc);
555 if (rx_conf->rx_thresh.pthresh != 0 ||
556 rx_conf->rx_thresh.hthresh != 0 ||
557 rx_conf->rx_thresh.wthresh != 0) {
559 "RxQ prefetch/host/writeback thresholds are not supported");
563 if (rx_conf->rx_free_thresh > rx_free_thresh_max) {
565 "RxQ free threshold too large: %u vs maximum %u",
566 rx_conf->rx_free_thresh, rx_free_thresh_max);
570 if (rx_conf->rx_drop_en == 0) {
571 sfc_err(sa, "RxQ drop disable is not supported");
579 sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
584 /* The mbuf object itself is always cache line aligned */
585 order = rte_bsf32(RTE_CACHE_LINE_SIZE);
587 /* Data offset from mbuf object start */
588 data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
589 RTE_PKTMBUF_HEADROOM;
591 order = MIN(order, rte_bsf32(data_off));
593 return 1u << (order - 1);
597 sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
599 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
600 const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
601 const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
603 unsigned int buf_aligned;
604 unsigned int start_alignment;
605 unsigned int end_padding_alignment;
607 /* Below it is assumed that both alignments are power of 2 */
608 SFC_ASSERT(rte_is_power_of_2(nic_align_start));
609 SFC_ASSERT(rte_is_power_of_2(nic_align_end));
612 * mbuf is always cache line aligned, double-check
613 * that it meets rx buffer start alignment requirements.
616 /* Start from mbuf pool data room size */
617 buf_size = rte_pktmbuf_data_room_size(mb_pool);
619 /* Remove headroom */
620 if (buf_size <= RTE_PKTMBUF_HEADROOM) {
622 "RxQ mbuf pool %s object data room size %u is smaller than headroom %u",
623 mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
626 buf_size -= RTE_PKTMBUF_HEADROOM;
628 /* Calculate guaranteed data start alignment */
629 buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
631 /* Reserve space for start alignment */
632 if (buf_aligned < nic_align_start) {
633 start_alignment = nic_align_start - buf_aligned;
634 if (buf_size <= start_alignment) {
636 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u and buffer start alignment %u required by NIC",
638 rte_pktmbuf_data_room_size(mb_pool),
639 RTE_PKTMBUF_HEADROOM, start_alignment);
642 buf_aligned = nic_align_start;
643 buf_size -= start_alignment;
648 /* Make sure that end padding does not write beyond the buffer */
649 if (buf_aligned < nic_align_end) {
651 * Estimate space which can be lost. If guarnteed buffer
652 * size is odd, lost space is (nic_align_end - 1). More
653 * accurate formula is below.
655 end_padding_alignment = nic_align_end -
656 MIN(buf_aligned, 1u << (rte_bsf32(buf_size) - 1));
657 if (buf_size <= end_padding_alignment) {
659 "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",
661 rte_pktmbuf_data_room_size(mb_pool),
662 RTE_PKTMBUF_HEADROOM, start_alignment,
663 end_padding_alignment);
666 buf_size -= end_padding_alignment;
669 * Start is aligned the same or better than end,
672 buf_size = P2ALIGN(buf_size, nic_align_end);
679 sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
680 uint16_t nb_rx_desc, unsigned int socket_id,
681 const struct rte_eth_rxconf *rx_conf,
682 struct rte_mempool *mb_pool)
684 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
687 struct sfc_rxq_info *rxq_info;
688 unsigned int evq_index;
692 rc = sfc_rx_qcheck_conf(sa, nb_rx_desc, rx_conf);
696 buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
698 sfc_err(sa, "RxQ %u mbuf pool object size is too small",
704 if ((buf_size < sa->port.pdu + encp->enc_rx_prefix_size) &&
705 !sa->eth_dev->data->dev_conf.rxmode.enable_scatter) {
706 sfc_err(sa, "Rx scatter is disabled and RxQ %u mbuf pool "
707 "object size is too small", sw_index);
708 sfc_err(sa, "RxQ %u calculated Rx buffer size is %u vs "
709 "PDU size %u plus Rx prefix %u bytes",
710 sw_index, buf_size, (unsigned int)sa->port.pdu,
711 encp->enc_rx_prefix_size);
716 SFC_ASSERT(sw_index < sa->rxq_count);
717 rxq_info = &sa->rxq_info[sw_index];
719 SFC_ASSERT(nb_rx_desc <= rxq_info->max_entries);
720 rxq_info->entries = nb_rx_desc;
722 sa->eth_dev->data->dev_conf.rxmode.enable_scatter ?
723 EFX_RXQ_TYPE_SCATTER : EFX_RXQ_TYPE_DEFAULT;
725 evq_index = sfc_evq_index_by_rxq_sw_index(sa, sw_index);
727 rc = sfc_ev_qinit(sa, evq_index, rxq_info->entries, socket_id);
731 evq = sa->evq_info[evq_index].evq;
734 rxq = rte_zmalloc_socket("sfc-rxq", sizeof(*rxq), RTE_CACHE_LINE_SIZE,
739 rc = sfc_dma_alloc(sa, "rxq", sw_index, EFX_RXQ_SIZE(rxq_info->entries),
740 socket_id, &rxq->mem);
745 rxq->sw_desc = rte_calloc_socket("sfc-rxq-sw_desc", rxq_info->entries,
746 sizeof(*rxq->sw_desc),
747 RTE_CACHE_LINE_SIZE, socket_id);
748 if (rxq->sw_desc == NULL)
749 goto fail_desc_alloc;
753 rxq->ptr_mask = rxq_info->entries - 1;
754 rxq->refill_threshold = rx_conf->rx_free_thresh;
755 rxq->refill_mb_pool = mb_pool;
756 rxq->buf_size = buf_size;
757 rxq->hw_index = sw_index;
758 rxq->port_id = sa->eth_dev->data->port_id;
760 /* Cache limits required on datapath in RxQ structure */
761 rxq->batch_max = encp->enc_rx_batch_max;
762 rxq->prefix_size = encp->enc_rx_prefix_size;
764 #if EFSYS_OPT_RX_SCALE
765 if (sa->hash_support == EFX_RX_HASH_AVAILABLE)
766 rxq->flags |= SFC_RXQ_RSS_HASH;
769 rxq->state = SFC_RXQ_INITIALIZED;
772 rxq_info->deferred_start = (rx_conf->rx_deferred_start != 0);
777 sfc_dma_free(sa, &rxq->mem);
783 sfc_ev_qfini(sa, evq_index);
786 rxq_info->entries = 0;
789 sfc_log_init(sa, "failed %d", rc);
794 sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
796 struct sfc_rxq_info *rxq_info;
799 SFC_ASSERT(sw_index < sa->rxq_count);
801 rxq_info = &sa->rxq_info[sw_index];
804 SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
806 rxq_info->rxq = NULL;
807 rxq_info->entries = 0;
809 rte_free(rxq->sw_desc);
810 sfc_dma_free(sa, &rxq->mem);
814 #if EFSYS_OPT_RX_SCALE
816 sfc_rte_to_efx_hash_type(uint64_t rss_hf)
818 efx_rx_hash_type_t efx_hash_types = 0;
820 if ((rss_hf & (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
821 ETH_RSS_NONFRAG_IPV4_OTHER)) != 0)
822 efx_hash_types |= EFX_RX_HASH_IPV4;
824 if ((rss_hf & ETH_RSS_NONFRAG_IPV4_TCP) != 0)
825 efx_hash_types |= EFX_RX_HASH_TCPIPV4;
827 if ((rss_hf & (ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
828 ETH_RSS_NONFRAG_IPV6_OTHER | ETH_RSS_IPV6_EX)) != 0)
829 efx_hash_types |= EFX_RX_HASH_IPV6;
831 if ((rss_hf & (ETH_RSS_NONFRAG_IPV6_TCP | ETH_RSS_IPV6_TCP_EX)) != 0)
832 efx_hash_types |= EFX_RX_HASH_TCPIPV6;
834 return efx_hash_types;
838 sfc_efx_to_rte_hash_type(efx_rx_hash_type_t efx_hash_types)
842 if ((efx_hash_types & EFX_RX_HASH_IPV4) != 0)
843 rss_hf |= (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
844 ETH_RSS_NONFRAG_IPV4_OTHER);
846 if ((efx_hash_types & EFX_RX_HASH_TCPIPV4) != 0)
847 rss_hf |= ETH_RSS_NONFRAG_IPV4_TCP;
849 if ((efx_hash_types & EFX_RX_HASH_IPV6) != 0)
850 rss_hf |= (ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
851 ETH_RSS_NONFRAG_IPV6_OTHER | ETH_RSS_IPV6_EX);
853 if ((efx_hash_types & EFX_RX_HASH_TCPIPV6) != 0)
854 rss_hf |= (ETH_RSS_NONFRAG_IPV6_TCP | ETH_RSS_IPV6_TCP_EX);
861 sfc_rx_rss_config(struct sfc_adapter *sa)
865 #if EFSYS_OPT_RX_SCALE
866 if (sa->rss_channels > 1) {
867 rc = efx_rx_scale_mode_set(sa->nic, EFX_RX_HASHALG_TOEPLITZ,
868 sa->rss_hash_types, B_TRUE);
872 rc = efx_rx_scale_key_set(sa->nic, sa->rss_key,
873 sizeof(sa->rss_key));
877 rc = efx_rx_scale_tbl_set(sa->nic, sa->rss_tbl,
878 sizeof(sa->rss_tbl));
887 sfc_rx_start(struct sfc_adapter *sa)
889 unsigned int sw_index;
892 sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
894 rc = efx_rx_init(sa->nic);
898 rc = sfc_rx_rss_config(sa);
900 goto fail_rss_config;
902 for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
903 if ((!sa->rxq_info[sw_index].deferred_start ||
904 sa->rxq_info[sw_index].deferred_started)) {
905 rc = sfc_rx_qstart(sa, sw_index);
914 while (sw_index-- > 0)
915 sfc_rx_qstop(sa, sw_index);
918 efx_rx_fini(sa->nic);
921 sfc_log_init(sa, "failed %d", rc);
926 sfc_rx_stop(struct sfc_adapter *sa)
928 unsigned int sw_index;
930 sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
932 sw_index = sa->rxq_count;
933 while (sw_index-- > 0) {
934 if (sa->rxq_info[sw_index].rxq != NULL)
935 sfc_rx_qstop(sa, sw_index);
938 efx_rx_fini(sa->nic);
942 sfc_rx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
944 struct sfc_rxq_info *rxq_info = &sa->rxq_info[sw_index];
945 unsigned int max_entries;
947 max_entries = EFX_RXQ_MAXNDESCS;
948 SFC_ASSERT(rte_is_power_of_2(max_entries));
950 rxq_info->max_entries = max_entries;
956 sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
960 switch (rxmode->mq_mode) {
962 /* No special checks are required */
964 #if EFSYS_OPT_RX_SCALE
966 if (sa->rss_support == EFX_RX_SCALE_UNAVAILABLE) {
967 sfc_err(sa, "RSS is not available");
973 sfc_err(sa, "Rx multi-queue mode %u not supported",
978 if (rxmode->header_split) {
979 sfc_err(sa, "Header split on Rx not supported");
983 if (rxmode->hw_vlan_filter) {
984 sfc_err(sa, "HW VLAN filtering not supported");
988 if (rxmode->hw_vlan_strip) {
989 sfc_err(sa, "HW VLAN stripping not supported");
993 if (rxmode->hw_vlan_extend) {
995 "Q-in-Q HW VLAN stripping not supported");
999 if (!rxmode->hw_strip_crc) {
1001 "FCS stripping control not supported - always stripped");
1002 rxmode->hw_strip_crc = 1;
1005 if (rxmode->enable_lro) {
1006 sfc_err(sa, "LRO not supported");
1014 * Initialize Rx subsystem.
1016 * Called at device configuration stage when number of receive queues is
1017 * specified together with other device level receive configuration.
1019 * It should be used to allocate NUMA-unaware resources.
1022 sfc_rx_init(struct sfc_adapter *sa)
1024 struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
1025 unsigned int sw_index;
1028 rc = sfc_rx_check_mode(sa, &dev_conf->rxmode);
1030 goto fail_check_mode;
1032 sa->rxq_count = sa->eth_dev->data->nb_rx_queues;
1035 sa->rxq_info = rte_calloc_socket("sfc-rxqs", sa->rxq_count,
1036 sizeof(struct sfc_rxq_info), 0,
1038 if (sa->rxq_info == NULL)
1039 goto fail_rxqs_alloc;
1041 for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
1042 rc = sfc_rx_qinit_info(sa, sw_index);
1044 goto fail_rx_qinit_info;
1047 #if EFSYS_OPT_RX_SCALE
1048 sa->rss_channels = (dev_conf->rxmode.mq_mode == ETH_MQ_RX_RSS) ?
1049 MIN(sa->rxq_count, EFX_MAXRSS) : 1;
1051 if (sa->rss_channels > 1) {
1052 for (sw_index = 0; sw_index < EFX_RSS_TBL_SIZE; ++sw_index)
1053 sa->rss_tbl[sw_index] = sw_index % sa->rss_channels;
1060 rte_free(sa->rxq_info);
1061 sa->rxq_info = NULL;
1066 sfc_log_init(sa, "failed %d", rc);
1071 * Shutdown Rx subsystem.
1073 * Called at device close stage, for example, before device
1074 * reconfiguration or shutdown.
1077 sfc_rx_fini(struct sfc_adapter *sa)
1079 unsigned int sw_index;
1081 sw_index = sa->rxq_count;
1082 while (sw_index-- > 0) {
1083 if (sa->rxq_info[sw_index].rxq != NULL)
1084 sfc_rx_qfini(sa, sw_index);
1087 rte_free(sa->rxq_info);
1088 sa->rxq_info = NULL;