1 /* SPDX-License-Identifier: BSD-3-Clause
3 * Copyright (c) 2016-2018 Solarflare Communications Inc.
6 * This software was jointly developed between OKTET Labs (under contract
7 * for Solarflare) and Solarflare Communications, Inc.
11 #include "sfc_debug.h"
15 #include "sfc_tweak.h"
16 #include "sfc_kvargs.h"
19 * Maximum number of TX queue flush attempts in case of
20 * failure or flush timeout
22 #define SFC_TX_QFLUSH_ATTEMPTS (3)
25 * Time to wait between event queue polling attempts when waiting for TX
26 * queue flush done or flush failed events
28 #define SFC_TX_QFLUSH_POLL_WAIT_MS (1)
31 * Maximum number of event queue polling attempts when waiting for TX queue
32 * flush done or flush failed events; it defines TX queue flush attempt timeout
33 * together with SFC_TX_QFLUSH_POLL_WAIT_MS
35 #define SFC_TX_QFLUSH_POLL_ATTEMPTS (2000)
38 sfc_tx_get_dev_offload_caps(struct sfc_adapter *sa)
40 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
43 if ((sa->dp_tx->features & SFC_DP_TX_FEAT_VLAN_INSERT) &&
44 encp->enc_hw_tx_insert_vlan_enabled)
45 caps |= DEV_TX_OFFLOAD_VLAN_INSERT;
47 if (sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_SEG)
48 caps |= DEV_TX_OFFLOAD_MULTI_SEGS;
50 if ((~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_POOL) &&
51 (~sa->dp_tx->features & SFC_DP_TX_FEAT_REFCNT))
52 caps |= DEV_TX_OFFLOAD_MBUF_FAST_FREE;
58 sfc_tx_get_queue_offload_caps(struct sfc_adapter *sa)
60 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
63 caps |= DEV_TX_OFFLOAD_IPV4_CKSUM;
64 caps |= DEV_TX_OFFLOAD_UDP_CKSUM;
65 caps |= DEV_TX_OFFLOAD_TCP_CKSUM;
67 if (encp->enc_tunnel_encapsulations_supported)
68 caps |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM;
71 caps |= DEV_TX_OFFLOAD_TCP_TSO;
77 sfc_tx_qcheck_conf(struct sfc_adapter *sa, unsigned int txq_max_fill_level,
78 const struct rte_eth_txconf *tx_conf,
83 if (tx_conf->tx_rs_thresh != 0) {
84 sfc_err(sa, "RS bit in transmit descriptor is not supported");
88 if (tx_conf->tx_free_thresh > txq_max_fill_level) {
90 "TxQ free threshold too large: %u vs maximum %u",
91 tx_conf->tx_free_thresh, txq_max_fill_level);
95 if (tx_conf->tx_thresh.pthresh != 0 ||
96 tx_conf->tx_thresh.hthresh != 0 ||
97 tx_conf->tx_thresh.wthresh != 0) {
99 "prefetch/host/writeback thresholds are not supported");
102 /* We either perform both TCP and UDP offload, or no offload at all */
103 if (((offloads & DEV_TX_OFFLOAD_TCP_CKSUM) == 0) !=
104 ((offloads & DEV_TX_OFFLOAD_UDP_CKSUM) == 0)) {
105 sfc_err(sa, "TCP and UDP offloads can't be set independently");
113 sfc_tx_qflush_done(struct sfc_txq *txq)
115 txq->state |= SFC_TXQ_FLUSHED;
116 txq->state &= ~SFC_TXQ_FLUSHING;
120 sfc_tx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
121 uint16_t nb_tx_desc, unsigned int socket_id,
122 const struct rte_eth_txconf *tx_conf)
124 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
125 unsigned int txq_entries;
126 unsigned int evq_entries;
127 unsigned int txq_max_fill_level;
128 struct sfc_txq_info *txq_info;
132 struct sfc_dp_tx_qcreate_info info;
135 sfc_log_init(sa, "TxQ = %u", sw_index);
137 rc = sa->dp_tx->qsize_up_rings(nb_tx_desc, &txq_entries, &evq_entries,
138 &txq_max_fill_level);
140 goto fail_size_up_rings;
141 SFC_ASSERT(txq_entries >= EFX_TXQ_MINNDESCS);
142 SFC_ASSERT(txq_entries <= sa->txq_max_entries);
143 SFC_ASSERT(txq_entries >= nb_tx_desc);
144 SFC_ASSERT(txq_max_fill_level <= nb_tx_desc);
146 offloads = tx_conf->offloads |
147 sa->eth_dev->data->dev_conf.txmode.offloads;
148 rc = sfc_tx_qcheck_conf(sa, txq_max_fill_level, tx_conf, offloads);
152 SFC_ASSERT(sw_index < sa->txq_count);
153 txq_info = &sa->txq_info[sw_index];
155 txq_info->entries = txq_entries;
157 rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_TX, sw_index,
158 evq_entries, socket_id, &evq);
163 txq = rte_zmalloc_socket("sfc-txq", sizeof(*txq), 0, socket_id);
169 txq->hw_index = sw_index;
172 (tx_conf->tx_free_thresh) ? tx_conf->tx_free_thresh :
173 SFC_TX_DEFAULT_FREE_THRESH;
174 txq->offloads = offloads;
176 rc = sfc_dma_alloc(sa, "txq", sw_index, EFX_TXQ_SIZE(txq_info->entries),
177 socket_id, &txq->mem);
181 memset(&info, 0, sizeof(info));
182 info.max_fill_level = txq_max_fill_level;
183 info.free_thresh = txq->free_thresh;
184 info.offloads = offloads;
185 info.txq_entries = txq_info->entries;
186 info.dma_desc_size_max = encp->enc_tx_dma_desc_size_max;
187 info.txq_hw_ring = txq->mem.esm_base;
188 info.evq_entries = evq_entries;
189 info.evq_hw_ring = evq->mem.esm_base;
190 info.hw_index = txq->hw_index;
191 info.mem_bar = sa->mem_bar.esb_base;
192 info.vi_window_shift = encp->enc_vi_window_shift;
194 rc = sa->dp_tx->qcreate(sa->eth_dev->data->port_id, sw_index,
195 &RTE_ETH_DEV_TO_PCI(sa->eth_dev)->addr,
196 socket_id, &info, &txq->dp);
198 goto fail_dp_tx_qinit;
200 evq->dp_txq = txq->dp;
202 txq->state = SFC_TXQ_INITIALIZED;
204 txq_info->deferred_start = (tx_conf->tx_deferred_start != 0);
209 sfc_dma_free(sa, &txq->mem);
212 txq_info->txq = NULL;
219 txq_info->entries = 0;
223 sfc_log_init(sa, "failed (TxQ = %u, rc = %d)", sw_index, rc);
228 sfc_tx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
230 struct sfc_txq_info *txq_info;
233 sfc_log_init(sa, "TxQ = %u", sw_index);
235 SFC_ASSERT(sw_index < sa->txq_count);
236 txq_info = &sa->txq_info[sw_index];
239 SFC_ASSERT(txq != NULL);
240 SFC_ASSERT(txq->state == SFC_TXQ_INITIALIZED);
242 sa->dp_tx->qdestroy(txq->dp);
245 txq_info->txq = NULL;
246 txq_info->entries = 0;
248 sfc_dma_free(sa, &txq->mem);
250 sfc_ev_qfini(txq->evq);
257 sfc_tx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
259 sfc_log_init(sa, "TxQ = %u", sw_index);
265 sfc_tx_check_mode(struct sfc_adapter *sa, const struct rte_eth_txmode *txmode)
269 switch (txmode->mq_mode) {
273 sfc_err(sa, "Tx multi-queue mode %u not supported",
279 * These features are claimed to be i40e-specific,
280 * but it does make sense to double-check their absence
282 if (txmode->hw_vlan_reject_tagged) {
283 sfc_err(sa, "Rejecting tagged packets not supported");
287 if (txmode->hw_vlan_reject_untagged) {
288 sfc_err(sa, "Rejecting untagged packets not supported");
292 if (txmode->hw_vlan_insert_pvid) {
293 sfc_err(sa, "Port-based VLAN insertion not supported");
301 * Destroy excess queues that are no longer needed after reconfiguration
305 sfc_tx_fini_queues(struct sfc_adapter *sa, unsigned int nb_tx_queues)
309 SFC_ASSERT(nb_tx_queues <= sa->txq_count);
311 sw_index = sa->txq_count;
312 while (--sw_index >= (int)nb_tx_queues) {
313 if (sa->txq_info[sw_index].txq != NULL)
314 sfc_tx_qfini(sa, sw_index);
317 sa->txq_count = nb_tx_queues;
321 sfc_tx_configure(struct sfc_adapter *sa)
323 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
324 const struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
325 const unsigned int nb_tx_queues = sa->eth_dev->data->nb_tx_queues;
328 sfc_log_init(sa, "nb_tx_queues=%u (old %u)",
329 nb_tx_queues, sa->txq_count);
332 * The datapath implementation assumes absence of boundary
333 * limits on Tx DMA descriptors. Addition of these checks on
334 * datapath would simply make the datapath slower.
336 if (encp->enc_tx_dma_desc_boundary != 0) {
338 goto fail_tx_dma_desc_boundary;
341 rc = sfc_tx_check_mode(sa, &dev_conf->txmode);
343 goto fail_check_mode;
345 if (nb_tx_queues == sa->txq_count)
348 if (sa->txq_info == NULL) {
349 sa->txq_info = rte_calloc_socket("sfc-txqs", nb_tx_queues,
350 sizeof(sa->txq_info[0]), 0,
352 if (sa->txq_info == NULL)
353 goto fail_txqs_alloc;
355 struct sfc_txq_info *new_txq_info;
357 if (nb_tx_queues < sa->txq_count)
358 sfc_tx_fini_queues(sa, nb_tx_queues);
361 rte_realloc(sa->txq_info,
362 nb_tx_queues * sizeof(sa->txq_info[0]), 0);
363 if (new_txq_info == NULL && nb_tx_queues > 0)
364 goto fail_txqs_realloc;
366 sa->txq_info = new_txq_info;
367 if (nb_tx_queues > sa->txq_count)
368 memset(&sa->txq_info[sa->txq_count], 0,
369 (nb_tx_queues - sa->txq_count) *
370 sizeof(sa->txq_info[0]));
373 while (sa->txq_count < nb_tx_queues) {
374 rc = sfc_tx_qinit_info(sa, sa->txq_count);
376 goto fail_tx_qinit_info;
390 fail_tx_dma_desc_boundary:
391 sfc_log_init(sa, "failed (rc = %d)", rc);
396 sfc_tx_close(struct sfc_adapter *sa)
398 sfc_tx_fini_queues(sa, 0);
400 rte_free(sa->txq_info);
405 sfc_tx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
407 uint64_t offloads_supported = sfc_tx_get_dev_offload_caps(sa) |
408 sfc_tx_get_queue_offload_caps(sa);
409 struct rte_eth_dev_data *dev_data;
410 struct sfc_txq_info *txq_info;
414 unsigned int desc_index;
417 sfc_log_init(sa, "TxQ = %u", sw_index);
419 SFC_ASSERT(sw_index < sa->txq_count);
420 txq_info = &sa->txq_info[sw_index];
424 SFC_ASSERT(txq != NULL);
425 SFC_ASSERT(txq->state == SFC_TXQ_INITIALIZED);
429 rc = sfc_ev_qstart(evq, sfc_evq_index_by_txq_sw_index(sa, sw_index));
433 if (txq->offloads & DEV_TX_OFFLOAD_IPV4_CKSUM)
434 flags |= EFX_TXQ_CKSUM_IPV4;
436 if (txq->offloads & DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
437 flags |= EFX_TXQ_CKSUM_INNER_IPV4;
439 if ((txq->offloads & DEV_TX_OFFLOAD_TCP_CKSUM) ||
440 (txq->offloads & DEV_TX_OFFLOAD_UDP_CKSUM)) {
441 flags |= EFX_TXQ_CKSUM_TCPUDP;
443 if (offloads_supported & DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
444 flags |= EFX_TXQ_CKSUM_INNER_TCPUDP;
447 if (txq->offloads & DEV_TX_OFFLOAD_TCP_TSO)
448 flags |= EFX_TXQ_FATSOV2;
450 rc = efx_tx_qcreate(sa->nic, sw_index, 0, &txq->mem,
451 txq_info->entries, 0 /* not used on EF10 */,
453 &txq->common, &desc_index);
455 if (sa->tso && (rc == ENOSPC))
456 sfc_err(sa, "ran out of TSO contexts");
458 goto fail_tx_qcreate;
461 efx_tx_qenable(txq->common);
463 txq->state |= SFC_TXQ_STARTED;
465 rc = sa->dp_tx->qstart(txq->dp, evq->read_ptr, desc_index);
470 * It seems to be used by DPDK for debug purposes only ('rte_ether')
472 dev_data = sa->eth_dev->data;
473 dev_data->tx_queue_state[sw_index] = RTE_ETH_QUEUE_STATE_STARTED;
478 txq->state = SFC_TXQ_INITIALIZED;
479 efx_tx_qdestroy(txq->common);
489 sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
491 struct rte_eth_dev_data *dev_data;
492 struct sfc_txq_info *txq_info;
494 unsigned int retry_count;
495 unsigned int wait_count;
498 sfc_log_init(sa, "TxQ = %u", sw_index);
500 SFC_ASSERT(sw_index < sa->txq_count);
501 txq_info = &sa->txq_info[sw_index];
505 if (txq == NULL || txq->state == SFC_TXQ_INITIALIZED)
508 SFC_ASSERT(txq->state & SFC_TXQ_STARTED);
510 sa->dp_tx->qstop(txq->dp, &txq->evq->read_ptr);
513 * Retry TX queue flushing in case of flush failed or
514 * timeout; in the worst case it can delay for 6 seconds
516 for (retry_count = 0;
517 ((txq->state & SFC_TXQ_FLUSHED) == 0) &&
518 (retry_count < SFC_TX_QFLUSH_ATTEMPTS);
520 rc = efx_tx_qflush(txq->common);
522 txq->state |= (rc == EALREADY) ?
523 SFC_TXQ_FLUSHED : SFC_TXQ_FLUSH_FAILED;
528 * Wait for TX queue flush done or flush failed event at least
529 * SFC_TX_QFLUSH_POLL_WAIT_MS milliseconds and not more
530 * than 2 seconds (SFC_TX_QFLUSH_POLL_WAIT_MS multiplied
531 * by SFC_TX_QFLUSH_POLL_ATTEMPTS)
535 rte_delay_ms(SFC_TX_QFLUSH_POLL_WAIT_MS);
536 sfc_ev_qpoll(txq->evq);
537 } while ((txq->state & SFC_TXQ_FLUSHING) &&
538 wait_count++ < SFC_TX_QFLUSH_POLL_ATTEMPTS);
540 if (txq->state & SFC_TXQ_FLUSHING)
541 sfc_err(sa, "TxQ %u flush timed out", sw_index);
543 if (txq->state & SFC_TXQ_FLUSHED)
544 sfc_notice(sa, "TxQ %u flushed", sw_index);
547 sa->dp_tx->qreap(txq->dp);
549 txq->state = SFC_TXQ_INITIALIZED;
551 efx_tx_qdestroy(txq->common);
553 sfc_ev_qstop(txq->evq);
556 * It seems to be used by DPDK for debug purposes only ('rte_ether')
558 dev_data = sa->eth_dev->data;
559 dev_data->tx_queue_state[sw_index] = RTE_ETH_QUEUE_STATE_STOPPED;
563 sfc_tx_start(struct sfc_adapter *sa)
565 unsigned int sw_index;
568 sfc_log_init(sa, "txq_count = %u", sa->txq_count);
571 if (!efx_nic_cfg_get(sa->nic)->enc_fw_assisted_tso_v2_enabled) {
572 sfc_warn(sa, "TSO support was unable to be restored");
577 rc = efx_tx_init(sa->nic);
579 goto fail_efx_tx_init;
581 for (sw_index = 0; sw_index < sa->txq_count; ++sw_index) {
582 if (sa->txq_info[sw_index].txq != NULL &&
583 (!(sa->txq_info[sw_index].deferred_start) ||
584 sa->txq_info[sw_index].deferred_started)) {
585 rc = sfc_tx_qstart(sa, sw_index);
594 while (sw_index-- > 0)
595 sfc_tx_qstop(sa, sw_index);
597 efx_tx_fini(sa->nic);
600 sfc_log_init(sa, "failed (rc = %d)", rc);
605 sfc_tx_stop(struct sfc_adapter *sa)
607 unsigned int sw_index;
609 sfc_log_init(sa, "txq_count = %u", sa->txq_count);
611 sw_index = sa->txq_count;
612 while (sw_index-- > 0) {
613 if (sa->txq_info[sw_index].txq != NULL)
614 sfc_tx_qstop(sa, sw_index);
617 efx_tx_fini(sa->nic);
621 sfc_efx_tx_reap(struct sfc_efx_txq *txq)
623 unsigned int completed;
625 sfc_ev_qpoll(txq->evq);
627 for (completed = txq->completed;
628 completed != txq->pending; completed++) {
629 struct sfc_efx_tx_sw_desc *txd;
631 txd = &txq->sw_ring[completed & txq->ptr_mask];
633 if (txd->mbuf != NULL) {
634 rte_pktmbuf_free(txd->mbuf);
639 txq->completed = completed;
643 * The function is used to insert or update VLAN tag;
644 * the firmware has state of the firmware tag to insert per TxQ
645 * (controlled by option descriptors), hence, if the tag of the
646 * packet to be sent is different from one remembered by the firmware,
647 * the function will update it
650 sfc_efx_tx_maybe_insert_tag(struct sfc_efx_txq *txq, struct rte_mbuf *m,
653 uint16_t this_tag = ((m->ol_flags & PKT_TX_VLAN_PKT) ?
656 if (this_tag == txq->hw_vlan_tci)
660 * The expression inside SFC_ASSERT() is not desired to be checked in
661 * a non-debug build because it might be too expensive on the data path
663 SFC_ASSERT(efx_nic_cfg_get(txq->evq->sa->nic)->enc_hw_tx_insert_vlan_enabled);
665 efx_tx_qdesc_vlantci_create(txq->common, rte_cpu_to_be_16(this_tag),
668 txq->hw_vlan_tci = this_tag;
674 sfc_efx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
676 struct sfc_dp_txq *dp_txq = (struct sfc_dp_txq *)tx_queue;
677 struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
678 unsigned int added = txq->added;
679 unsigned int pushed = added;
680 unsigned int pkts_sent = 0;
681 efx_desc_t *pend = &txq->pend_desc[0];
682 const unsigned int hard_max_fill = txq->max_fill_level;
683 const unsigned int soft_max_fill = hard_max_fill - txq->free_thresh;
684 unsigned int fill_level = added - txq->completed;
687 struct rte_mbuf **pktp;
689 if (unlikely((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) == 0))
693 * If insufficient space for a single packet is present,
694 * we should reap; otherwise, we shouldn't do that all the time
695 * to avoid latency increase
697 reap_done = (fill_level > soft_max_fill);
700 sfc_efx_tx_reap(txq);
702 * Recalculate fill level since 'txq->completed'
703 * might have changed on reap
705 fill_level = added - txq->completed;
708 for (pkts_sent = 0, pktp = &tx_pkts[0];
709 (pkts_sent < nb_pkts) && (fill_level <= soft_max_fill);
710 pkts_sent++, pktp++) {
711 struct rte_mbuf *m_seg = *pktp;
712 size_t pkt_len = m_seg->pkt_len;
713 unsigned int pkt_descs = 0;
717 * Here VLAN TCI is expected to be zero in case if no
718 * DEV_TX_OFFLOAD_VLAN_INSERT capability is advertised;
719 * if the calling app ignores the absence of
720 * DEV_TX_OFFLOAD_VLAN_INSERT and pushes VLAN TCI, then
721 * TX_ERROR will occur
723 pkt_descs += sfc_efx_tx_maybe_insert_tag(txq, m_seg, &pend);
725 if (m_seg->ol_flags & PKT_TX_TCP_SEG) {
727 * We expect correct 'pkt->l[2, 3, 4]_len' values
728 * to be set correctly by the caller
730 if (sfc_efx_tso_do(txq, added, &m_seg, &in_off, &pend,
731 &pkt_descs, &pkt_len) != 0) {
732 /* We may have reached this place for
733 * one of the following reasons:
735 * 1) Packet header length is greater
736 * than SFC_TSOH_STD_LEN
737 * 2) TCP header starts at more then
738 * 208 bytes into the frame
740 * We will deceive RTE saying that we have sent
741 * the packet, but we will actually drop it.
742 * Hence, we should revert 'pend' to the
743 * previous state (in case we have added
744 * VLAN descriptor) and start processing
745 * another one packet. But the original
746 * mbuf shouldn't be orphaned
750 rte_pktmbuf_free(*pktp);
756 * We've only added 2 FATSOv2 option descriptors
757 * and 1 descriptor for the linearized packet header.
758 * The outstanding work will be done in the same manner
759 * as for the usual non-TSO path
763 for (; m_seg != NULL; m_seg = m_seg->next) {
764 efsys_dma_addr_t next_frag;
767 seg_len = m_seg->data_len;
768 next_frag = rte_mbuf_data_iova(m_seg);
771 * If we've started TSO transaction few steps earlier,
772 * we'll skip packet header using an offset in the
773 * current segment (which has been set to the
774 * first one containing payload)
781 efsys_dma_addr_t frag_addr = next_frag;
785 * It is assumed here that there is no
786 * limitation on address boundary
787 * crossing by DMA descriptor.
789 frag_len = MIN(seg_len, txq->dma_desc_size_max);
790 next_frag += frag_len;
794 efx_tx_qdesc_dma_create(txq->common,
800 } while (seg_len != 0);
805 fill_level += pkt_descs;
806 if (unlikely(fill_level > hard_max_fill)) {
808 * Our estimation for maximum number of descriptors
809 * required to send a packet seems to be wrong.
810 * Try to reap (if we haven't yet).
813 sfc_efx_tx_reap(txq);
815 fill_level = added - txq->completed;
816 if (fill_level > hard_max_fill) {
826 /* Assign mbuf to the last used desc */
827 txq->sw_ring[(added - 1) & txq->ptr_mask].mbuf = *pktp;
830 if (likely(pkts_sent > 0)) {
831 rc = efx_tx_qdesc_post(txq->common, txq->pend_desc,
832 pend - &txq->pend_desc[0],
833 txq->completed, &txq->added);
836 if (likely(pushed != txq->added))
837 efx_tx_qpush(txq->common, txq->added, pushed);
840 #if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
842 sfc_efx_tx_reap(txq);
850 sfc_txq_by_dp_txq(const struct sfc_dp_txq *dp_txq)
852 const struct sfc_dp_queue *dpq = &dp_txq->dpq;
853 struct rte_eth_dev *eth_dev;
854 struct sfc_adapter *sa;
857 SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
858 eth_dev = &rte_eth_devices[dpq->port_id];
860 sa = eth_dev->data->dev_private;
862 SFC_ASSERT(dpq->queue_id < sa->txq_count);
863 txq = sa->txq_info[dpq->queue_id].txq;
865 SFC_ASSERT(txq != NULL);
869 static sfc_dp_tx_qsize_up_rings_t sfc_efx_tx_qsize_up_rings;
871 sfc_efx_tx_qsize_up_rings(uint16_t nb_tx_desc,
872 unsigned int *txq_entries,
873 unsigned int *evq_entries,
874 unsigned int *txq_max_fill_level)
876 *txq_entries = nb_tx_desc;
877 *evq_entries = nb_tx_desc;
878 *txq_max_fill_level = EFX_TXQ_LIMIT(*txq_entries);
882 static sfc_dp_tx_qcreate_t sfc_efx_tx_qcreate;
884 sfc_efx_tx_qcreate(uint16_t port_id, uint16_t queue_id,
885 const struct rte_pci_addr *pci_addr,
887 const struct sfc_dp_tx_qcreate_info *info,
888 struct sfc_dp_txq **dp_txqp)
890 struct sfc_efx_txq *txq;
891 struct sfc_txq *ctrl_txq;
895 txq = rte_zmalloc_socket("sfc-efx-txq", sizeof(*txq),
896 RTE_CACHE_LINE_SIZE, socket_id);
900 sfc_dp_queue_init(&txq->dp.dpq, port_id, queue_id, pci_addr);
903 txq->pend_desc = rte_calloc_socket("sfc-efx-txq-pend-desc",
904 EFX_TXQ_LIMIT(info->txq_entries),
905 sizeof(*txq->pend_desc), 0,
907 if (txq->pend_desc == NULL)
908 goto fail_pend_desc_alloc;
911 txq->sw_ring = rte_calloc_socket("sfc-efx-txq-sw_ring",
913 sizeof(*txq->sw_ring),
914 RTE_CACHE_LINE_SIZE, socket_id);
915 if (txq->sw_ring == NULL)
916 goto fail_sw_ring_alloc;
918 ctrl_txq = sfc_txq_by_dp_txq(&txq->dp);
919 if (ctrl_txq->evq->sa->tso) {
920 rc = sfc_efx_tso_alloc_tsoh_objs(txq->sw_ring,
921 info->txq_entries, socket_id);
923 goto fail_alloc_tsoh_objs;
926 txq->evq = ctrl_txq->evq;
927 txq->ptr_mask = info->txq_entries - 1;
928 txq->max_fill_level = info->max_fill_level;
929 txq->free_thresh = info->free_thresh;
930 txq->dma_desc_size_max = info->dma_desc_size_max;
935 fail_alloc_tsoh_objs:
936 rte_free(txq->sw_ring);
939 rte_free(txq->pend_desc);
941 fail_pend_desc_alloc:
948 static sfc_dp_tx_qdestroy_t sfc_efx_tx_qdestroy;
950 sfc_efx_tx_qdestroy(struct sfc_dp_txq *dp_txq)
952 struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
954 sfc_efx_tso_free_tsoh_objs(txq->sw_ring, txq->ptr_mask + 1);
955 rte_free(txq->sw_ring);
956 rte_free(txq->pend_desc);
960 static sfc_dp_tx_qstart_t sfc_efx_tx_qstart;
962 sfc_efx_tx_qstart(struct sfc_dp_txq *dp_txq,
963 __rte_unused unsigned int evq_read_ptr,
964 unsigned int txq_desc_index)
966 /* libefx-based datapath is specific to libefx-based PMD */
967 struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
968 struct sfc_txq *ctrl_txq = sfc_txq_by_dp_txq(dp_txq);
970 txq->common = ctrl_txq->common;
972 txq->pending = txq->completed = txq->added = txq_desc_index;
973 txq->hw_vlan_tci = 0;
975 txq->flags |= (SFC_EFX_TXQ_FLAG_STARTED | SFC_EFX_TXQ_FLAG_RUNNING);
980 static sfc_dp_tx_qstop_t sfc_efx_tx_qstop;
982 sfc_efx_tx_qstop(struct sfc_dp_txq *dp_txq,
983 __rte_unused unsigned int *evq_read_ptr)
985 struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
987 txq->flags &= ~SFC_EFX_TXQ_FLAG_RUNNING;
990 static sfc_dp_tx_qreap_t sfc_efx_tx_qreap;
992 sfc_efx_tx_qreap(struct sfc_dp_txq *dp_txq)
994 struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
997 sfc_efx_tx_reap(txq);
999 for (txds = 0; txds <= txq->ptr_mask; txds++) {
1000 if (txq->sw_ring[txds].mbuf != NULL) {
1001 rte_pktmbuf_free(txq->sw_ring[txds].mbuf);
1002 txq->sw_ring[txds].mbuf = NULL;
1006 txq->flags &= ~SFC_EFX_TXQ_FLAG_STARTED;
1009 static sfc_dp_tx_qdesc_status_t sfc_efx_tx_qdesc_status;
1011 sfc_efx_tx_qdesc_status(struct sfc_dp_txq *dp_txq, uint16_t offset)
1013 struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1015 if (unlikely(offset > txq->ptr_mask))
1018 if (unlikely(offset >= txq->max_fill_level))
1019 return RTE_ETH_TX_DESC_UNAVAIL;
1022 * Poll EvQ to derive up-to-date 'txq->pending' figure;
1023 * it is required for the queue to be running, but the
1024 * check is omitted because API design assumes that it
1025 * is the duty of the caller to satisfy all conditions
1027 SFC_ASSERT((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) ==
1028 SFC_EFX_TXQ_FLAG_RUNNING);
1029 sfc_ev_qpoll(txq->evq);
1032 * Ring tail is 'txq->pending', and although descriptors
1033 * between 'txq->completed' and 'txq->pending' are still
1034 * in use by the driver, they should be reported as DONE
1036 if (unlikely(offset < (txq->added - txq->pending)))
1037 return RTE_ETH_TX_DESC_FULL;
1040 * There is no separate return value for unused descriptors;
1041 * the latter will be reported as DONE because genuine DONE
1042 * descriptors will be freed anyway in SW on the next burst
1044 return RTE_ETH_TX_DESC_DONE;
1047 struct sfc_dp_tx sfc_efx_tx = {
1049 .name = SFC_KVARG_DATAPATH_EFX,
1053 .features = SFC_DP_TX_FEAT_VLAN_INSERT |
1054 SFC_DP_TX_FEAT_TSO |
1055 SFC_DP_TX_FEAT_MULTI_POOL |
1056 SFC_DP_TX_FEAT_REFCNT |
1057 SFC_DP_TX_FEAT_MULTI_SEG,
1058 .qsize_up_rings = sfc_efx_tx_qsize_up_rings,
1059 .qcreate = sfc_efx_tx_qcreate,
1060 .qdestroy = sfc_efx_tx_qdestroy,
1061 .qstart = sfc_efx_tx_qstart,
1062 .qstop = sfc_efx_tx_qstop,
1063 .qreap = sfc_efx_tx_qreap,
1064 .qdesc_status = sfc_efx_tx_qdesc_status,
1065 .pkt_burst = sfc_efx_xmit_pkts,