4 * Copyright (c) 2016-2017 Solarflare Communications Inc.
7 * This software was jointly developed between OKTET Labs (under contract
8 * for Solarflare) and Solarflare Communications, Inc.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions are met:
13 * 1. Redistributions of source code must retain the above copyright notice,
14 * this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright notice,
16 * this list of conditions and the following disclaimer in the documentation
17 * and/or other materials provided with the distribution.
19 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
20 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
21 * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
23 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
24 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
25 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
26 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
27 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
28 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
29 * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33 #include "sfc_debug.h"
37 #include "sfc_tweak.h"
38 #include "sfc_kvargs.h"
41 * Maximum number of TX queue flush attempts in case of
42 * failure or flush timeout
44 #define SFC_TX_QFLUSH_ATTEMPTS (3)
47 * Time to wait between event queue polling attempts when waiting for TX
48 * queue flush done or flush failed events
50 #define SFC_TX_QFLUSH_POLL_WAIT_MS (1)
53 * Maximum number of event queue polling attempts when waiting for TX queue
54 * flush done or flush failed events; it defines TX queue flush attempt timeout
55 * together with SFC_TX_QFLUSH_POLL_WAIT_MS
57 #define SFC_TX_QFLUSH_POLL_ATTEMPTS (2000)
60 sfc_tx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_tx_desc,
61 const struct rte_eth_txconf *tx_conf)
63 unsigned int flags = tx_conf->txq_flags;
64 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
67 if (tx_conf->tx_rs_thresh != 0) {
68 sfc_err(sa, "RS bit in transmit descriptor is not supported");
72 if (tx_conf->tx_free_thresh > EFX_TXQ_LIMIT(nb_tx_desc)) {
74 "TxQ free threshold too large: %u vs maximum %u",
75 tx_conf->tx_free_thresh, EFX_TXQ_LIMIT(nb_tx_desc));
79 if (tx_conf->tx_thresh.pthresh != 0 ||
80 tx_conf->tx_thresh.hthresh != 0 ||
81 tx_conf->tx_thresh.wthresh != 0) {
83 "prefetch/host/writeback thresholds are not supported");
87 if (((flags & ETH_TXQ_FLAGS_NOMULTSEGS) == 0) &&
88 (~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_SEG)) {
89 sfc_err(sa, "Multi-segment is not supported by %s datapath",
94 if (((flags & ETH_TXQ_FLAGS_NOMULTMEMP) == 0) &&
95 (~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_POOL)) {
96 sfc_err(sa, "multi-mempool is not supported by %s datapath",
101 if (((flags & ETH_TXQ_FLAGS_NOREFCOUNT) == 0) &&
102 (~sa->dp_tx->features & SFC_DP_TX_FEAT_REFCNT)) {
104 "mbuf reference counters are neglected by %s datapath",
109 if ((flags & ETH_TXQ_FLAGS_NOVLANOFFL) == 0) {
110 if (!encp->enc_hw_tx_insert_vlan_enabled) {
111 sfc_err(sa, "VLAN offload is not supported");
113 } else if (~sa->dp_tx->features & SFC_DP_TX_FEAT_VLAN_INSERT) {
115 "VLAN offload is not supported by %s datapath",
121 if ((flags & ETH_TXQ_FLAGS_NOXSUMSCTP) == 0) {
122 sfc_err(sa, "SCTP offload is not supported");
126 /* We either perform both TCP and UDP offload, or no offload at all */
127 if (((flags & ETH_TXQ_FLAGS_NOXSUMTCP) == 0) !=
128 ((flags & ETH_TXQ_FLAGS_NOXSUMUDP) == 0)) {
129 sfc_err(sa, "TCP and UDP offloads can't be set independently");
137 sfc_tx_qflush_done(struct sfc_txq *txq)
139 txq->state |= SFC_TXQ_FLUSHED;
140 txq->state &= ~SFC_TXQ_FLUSHING;
144 sfc_tx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
145 uint16_t nb_tx_desc, unsigned int socket_id,
146 const struct rte_eth_txconf *tx_conf)
148 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
149 struct sfc_txq_info *txq_info;
153 struct sfc_dp_tx_qcreate_info info;
155 sfc_log_init(sa, "TxQ = %u", sw_index);
157 rc = sfc_tx_qcheck_conf(sa, nb_tx_desc, tx_conf);
161 SFC_ASSERT(sw_index < sa->txq_count);
162 txq_info = &sa->txq_info[sw_index];
164 SFC_ASSERT(nb_tx_desc <= sa->txq_max_entries);
165 txq_info->entries = nb_tx_desc;
167 rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_TX, sw_index,
168 txq_info->entries, socket_id, &evq);
173 txq = rte_zmalloc_socket("sfc-txq", sizeof(*txq), 0, socket_id);
179 txq->hw_index = sw_index;
182 (tx_conf->tx_free_thresh) ? tx_conf->tx_free_thresh :
183 SFC_TX_DEFAULT_FREE_THRESH;
184 txq->flags = tx_conf->txq_flags;
186 rc = sfc_dma_alloc(sa, "txq", sw_index, EFX_TXQ_SIZE(txq_info->entries),
187 socket_id, &txq->mem);
191 memset(&info, 0, sizeof(info));
192 info.free_thresh = txq->free_thresh;
193 info.flags = tx_conf->txq_flags;
194 info.txq_entries = txq_info->entries;
195 info.dma_desc_size_max = encp->enc_tx_dma_desc_size_max;
196 info.txq_hw_ring = txq->mem.esm_base;
197 info.evq_entries = txq_info->entries;
198 info.evq_hw_ring = evq->mem.esm_base;
199 info.hw_index = txq->hw_index;
200 info.mem_bar = sa->mem_bar.esb_base;
202 rc = sa->dp_tx->qcreate(sa->eth_dev->data->port_id, sw_index,
203 &RTE_ETH_DEV_TO_PCI(sa->eth_dev)->addr,
204 socket_id, &info, &txq->dp);
206 goto fail_dp_tx_qinit;
208 evq->dp_txq = txq->dp;
210 txq->state = SFC_TXQ_INITIALIZED;
212 txq_info->deferred_start = (tx_conf->tx_deferred_start != 0);
217 sfc_dma_free(sa, &txq->mem);
220 txq_info->txq = NULL;
227 txq_info->entries = 0;
230 sfc_log_init(sa, "failed (TxQ = %u, rc = %d)", sw_index, rc);
235 sfc_tx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
237 struct sfc_txq_info *txq_info;
240 sfc_log_init(sa, "TxQ = %u", sw_index);
242 SFC_ASSERT(sw_index < sa->txq_count);
243 txq_info = &sa->txq_info[sw_index];
246 SFC_ASSERT(txq != NULL);
247 SFC_ASSERT(txq->state == SFC_TXQ_INITIALIZED);
249 sa->dp_tx->qdestroy(txq->dp);
252 txq_info->txq = NULL;
253 txq_info->entries = 0;
255 sfc_dma_free(sa, &txq->mem);
257 sfc_ev_qfini(txq->evq);
264 sfc_tx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
266 sfc_log_init(sa, "TxQ = %u", sw_index);
272 sfc_tx_check_mode(struct sfc_adapter *sa, const struct rte_eth_txmode *txmode)
276 switch (txmode->mq_mode) {
280 sfc_err(sa, "Tx multi-queue mode %u not supported",
286 * These features are claimed to be i40e-specific,
287 * but it does make sense to double-check their absence
289 if (txmode->hw_vlan_reject_tagged) {
290 sfc_err(sa, "Rejecting tagged packets not supported");
294 if (txmode->hw_vlan_reject_untagged) {
295 sfc_err(sa, "Rejecting untagged packets not supported");
299 if (txmode->hw_vlan_insert_pvid) {
300 sfc_err(sa, "Port-based VLAN insertion not supported");
308 * Destroy excess queues that are no longer needed after reconfiguration
312 sfc_tx_fini_queues(struct sfc_adapter *sa, unsigned int nb_tx_queues)
316 SFC_ASSERT(nb_tx_queues <= sa->txq_count);
318 sw_index = sa->txq_count;
319 while (--sw_index >= (int)nb_tx_queues) {
320 if (sa->txq_info[sw_index].txq != NULL)
321 sfc_tx_qfini(sa, sw_index);
324 sa->txq_count = nb_tx_queues;
328 sfc_tx_configure(struct sfc_adapter *sa)
330 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
331 const struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
332 const unsigned int nb_tx_queues = sa->eth_dev->data->nb_tx_queues;
335 sfc_log_init(sa, "nb_tx_queues=%u (old %u)",
336 nb_tx_queues, sa->txq_count);
339 * The datapath implementation assumes absence of boundary
340 * limits on Tx DMA descriptors. Addition of these checks on
341 * datapath would simply make the datapath slower.
343 if (encp->enc_tx_dma_desc_boundary != 0) {
345 goto fail_tx_dma_desc_boundary;
348 rc = sfc_tx_check_mode(sa, &dev_conf->txmode);
350 goto fail_check_mode;
352 if (nb_tx_queues == sa->txq_count)
355 if (sa->txq_info == NULL) {
356 sa->txq_info = rte_calloc_socket("sfc-txqs", nb_tx_queues,
357 sizeof(sa->txq_info[0]), 0,
359 if (sa->txq_info == NULL)
360 goto fail_txqs_alloc;
362 struct sfc_txq_info *new_txq_info;
364 if (nb_tx_queues < sa->txq_count)
365 sfc_tx_fini_queues(sa, nb_tx_queues);
368 rte_realloc(sa->txq_info,
369 nb_tx_queues * sizeof(sa->txq_info[0]), 0);
370 if (new_txq_info == NULL && nb_tx_queues > 0)
371 goto fail_txqs_realloc;
373 sa->txq_info = new_txq_info;
374 if (nb_tx_queues > sa->txq_count)
375 memset(&sa->txq_info[sa->txq_count], 0,
376 (nb_tx_queues - sa->txq_count) *
377 sizeof(sa->txq_info[0]));
380 while (sa->txq_count < nb_tx_queues) {
381 rc = sfc_tx_qinit_info(sa, sa->txq_count);
383 goto fail_tx_qinit_info;
397 fail_tx_dma_desc_boundary:
398 sfc_log_init(sa, "failed (rc = %d)", rc);
403 sfc_tx_close(struct sfc_adapter *sa)
405 sfc_tx_fini_queues(sa, 0);
407 rte_free(sa->txq_info);
412 sfc_tx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
414 struct rte_eth_dev_data *dev_data;
415 struct sfc_txq_info *txq_info;
419 unsigned int desc_index;
422 sfc_log_init(sa, "TxQ = %u", sw_index);
424 SFC_ASSERT(sw_index < sa->txq_count);
425 txq_info = &sa->txq_info[sw_index];
429 SFC_ASSERT(txq->state == SFC_TXQ_INITIALIZED);
433 rc = sfc_ev_qstart(evq, sfc_evq_index_by_txq_sw_index(sa, sw_index));
438 * It seems that DPDK has no controls regarding IPv4 offloads,
439 * hence, we always enable it here
441 if ((txq->flags & ETH_TXQ_FLAGS_NOXSUMTCP) ||
442 (txq->flags & ETH_TXQ_FLAGS_NOXSUMUDP)) {
443 flags = EFX_TXQ_CKSUM_IPV4;
445 flags = EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP;
448 flags |= EFX_TXQ_FATSOV2;
451 rc = efx_tx_qcreate(sa->nic, sw_index, 0, &txq->mem,
452 txq_info->entries, 0 /* not used on EF10 */,
454 &txq->common, &desc_index);
456 if (sa->tso && (rc == ENOSPC))
457 sfc_err(sa, "ran out of TSO contexts");
459 goto fail_tx_qcreate;
462 efx_tx_qenable(txq->common);
464 txq->state |= SFC_TXQ_STARTED;
466 rc = sa->dp_tx->qstart(txq->dp, evq->read_ptr, desc_index);
471 * It seems to be used by DPDK for debug purposes only ('rte_ether')
473 dev_data = sa->eth_dev->data;
474 dev_data->tx_queue_state[sw_index] = RTE_ETH_QUEUE_STATE_STARTED;
479 txq->state = SFC_TXQ_INITIALIZED;
480 efx_tx_qdestroy(txq->common);
490 sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
492 struct rte_eth_dev_data *dev_data;
493 struct sfc_txq_info *txq_info;
495 unsigned int retry_count;
496 unsigned int wait_count;
499 sfc_log_init(sa, "TxQ = %u", sw_index);
501 SFC_ASSERT(sw_index < sa->txq_count);
502 txq_info = &sa->txq_info[sw_index];
506 if (txq->state == SFC_TXQ_INITIALIZED)
509 SFC_ASSERT(txq->state & SFC_TXQ_STARTED);
511 sa->dp_tx->qstop(txq->dp, &txq->evq->read_ptr);
514 * Retry TX queue flushing in case of flush failed or
515 * timeout; in the worst case it can delay for 6 seconds
517 for (retry_count = 0;
518 ((txq->state & SFC_TXQ_FLUSHED) == 0) &&
519 (retry_count < SFC_TX_QFLUSH_ATTEMPTS);
521 rc = efx_tx_qflush(txq->common);
523 txq->state |= (rc == EALREADY) ?
524 SFC_TXQ_FLUSHED : SFC_TXQ_FLUSH_FAILED;
529 * Wait for TX queue flush done or flush failed event at least
530 * SFC_TX_QFLUSH_POLL_WAIT_MS milliseconds and not more
531 * than 2 seconds (SFC_TX_QFLUSH_POLL_WAIT_MS multiplied
532 * by SFC_TX_QFLUSH_POLL_ATTEMPTS)
536 rte_delay_ms(SFC_TX_QFLUSH_POLL_WAIT_MS);
537 sfc_ev_qpoll(txq->evq);
538 } while ((txq->state & SFC_TXQ_FLUSHING) &&
539 wait_count++ < SFC_TX_QFLUSH_POLL_ATTEMPTS);
541 if (txq->state & SFC_TXQ_FLUSHING)
542 sfc_err(sa, "TxQ %u flush timed out", sw_index);
544 if (txq->state & SFC_TXQ_FLUSHED)
545 sfc_info(sa, "TxQ %u flushed", sw_index);
548 sa->dp_tx->qreap(txq->dp);
550 txq->state = SFC_TXQ_INITIALIZED;
552 efx_tx_qdestroy(txq->common);
554 sfc_ev_qstop(txq->evq);
557 * It seems to be used by DPDK for debug purposes only ('rte_ether')
559 dev_data = sa->eth_dev->data;
560 dev_data->tx_queue_state[sw_index] = RTE_ETH_QUEUE_STATE_STOPPED;
564 sfc_tx_start(struct sfc_adapter *sa)
566 unsigned int sw_index;
569 sfc_log_init(sa, "txq_count = %u", sa->txq_count);
572 if (!efx_nic_cfg_get(sa->nic)->enc_fw_assisted_tso_v2_enabled) {
573 sfc_warn(sa, "TSO support was unable to be restored");
578 rc = efx_tx_init(sa->nic);
580 goto fail_efx_tx_init;
582 for (sw_index = 0; sw_index < sa->txq_count; ++sw_index) {
583 if (!(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 = EFX_TXQ_LIMIT(txq->ptr_mask + 1);
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_VLAN_OFFLOAD capability is advertised;
719 * if the calling app ignores the absence of
720 * DEV_TX_VLAN_OFFLOAD 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_dma_addr(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_qcreate_t sfc_efx_tx_qcreate;
871 sfc_efx_tx_qcreate(uint16_t port_id, uint16_t queue_id,
872 const struct rte_pci_addr *pci_addr,
874 const struct sfc_dp_tx_qcreate_info *info,
875 struct sfc_dp_txq **dp_txqp)
877 struct sfc_efx_txq *txq;
878 struct sfc_txq *ctrl_txq;
882 txq = rte_zmalloc_socket("sfc-efx-txq", sizeof(*txq),
883 RTE_CACHE_LINE_SIZE, socket_id);
887 sfc_dp_queue_init(&txq->dp.dpq, port_id, queue_id, pci_addr);
890 txq->pend_desc = rte_calloc_socket("sfc-efx-txq-pend-desc",
891 EFX_TXQ_LIMIT(info->txq_entries),
892 sizeof(*txq->pend_desc), 0,
894 if (txq->pend_desc == NULL)
895 goto fail_pend_desc_alloc;
898 txq->sw_ring = rte_calloc_socket("sfc-efx-txq-sw_ring",
900 sizeof(*txq->sw_ring),
901 RTE_CACHE_LINE_SIZE, socket_id);
902 if (txq->sw_ring == NULL)
903 goto fail_sw_ring_alloc;
905 ctrl_txq = sfc_txq_by_dp_txq(&txq->dp);
906 if (ctrl_txq->evq->sa->tso) {
907 rc = sfc_efx_tso_alloc_tsoh_objs(txq->sw_ring,
908 info->txq_entries, socket_id);
910 goto fail_alloc_tsoh_objs;
913 txq->evq = ctrl_txq->evq;
914 txq->ptr_mask = info->txq_entries - 1;
915 txq->free_thresh = info->free_thresh;
916 txq->dma_desc_size_max = info->dma_desc_size_max;
921 fail_alloc_tsoh_objs:
922 rte_free(txq->sw_ring);
925 rte_free(txq->pend_desc);
927 fail_pend_desc_alloc:
934 static sfc_dp_tx_qdestroy_t sfc_efx_tx_qdestroy;
936 sfc_efx_tx_qdestroy(struct sfc_dp_txq *dp_txq)
938 struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
940 sfc_efx_tso_free_tsoh_objs(txq->sw_ring, txq->ptr_mask + 1);
941 rte_free(txq->sw_ring);
942 rte_free(txq->pend_desc);
946 static sfc_dp_tx_qstart_t sfc_efx_tx_qstart;
948 sfc_efx_tx_qstart(struct sfc_dp_txq *dp_txq,
949 __rte_unused unsigned int evq_read_ptr,
950 unsigned int txq_desc_index)
952 /* libefx-based datapath is specific to libefx-based PMD */
953 struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
954 struct sfc_txq *ctrl_txq = sfc_txq_by_dp_txq(dp_txq);
956 txq->common = ctrl_txq->common;
958 txq->pending = txq->completed = txq->added = txq_desc_index;
959 txq->hw_vlan_tci = 0;
961 txq->flags |= (SFC_EFX_TXQ_FLAG_STARTED | SFC_EFX_TXQ_FLAG_RUNNING);
966 static sfc_dp_tx_qstop_t sfc_efx_tx_qstop;
968 sfc_efx_tx_qstop(struct sfc_dp_txq *dp_txq,
969 __rte_unused unsigned int *evq_read_ptr)
971 struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
973 txq->flags &= ~SFC_EFX_TXQ_FLAG_RUNNING;
976 static sfc_dp_tx_qreap_t sfc_efx_tx_qreap;
978 sfc_efx_tx_qreap(struct sfc_dp_txq *dp_txq)
980 struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
983 sfc_efx_tx_reap(txq);
985 for (txds = 0; txds <= txq->ptr_mask; txds++) {
986 if (txq->sw_ring[txds].mbuf != NULL) {
987 rte_pktmbuf_free(txq->sw_ring[txds].mbuf);
988 txq->sw_ring[txds].mbuf = NULL;
992 txq->flags &= ~SFC_EFX_TXQ_FLAG_STARTED;
995 static sfc_dp_tx_qdesc_status_t sfc_efx_tx_qdesc_status;
997 sfc_efx_tx_qdesc_status(struct sfc_dp_txq *dp_txq, uint16_t offset)
999 struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
1001 if (unlikely(offset > txq->ptr_mask))
1004 if (unlikely(offset >= EFX_TXQ_LIMIT(txq->ptr_mask + 1)))
1005 return RTE_ETH_TX_DESC_UNAVAIL;
1008 * Poll EvQ to derive up-to-date 'txq->pending' figure;
1009 * it is required for the queue to be running, but the
1010 * check is omitted because API design assumes that it
1011 * is the duty of the caller to satisfy all conditions
1013 SFC_ASSERT((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) ==
1014 SFC_EFX_TXQ_FLAG_RUNNING);
1015 sfc_ev_qpoll(txq->evq);
1018 * Ring tail is 'txq->pending', and although descriptors
1019 * between 'txq->completed' and 'txq->pending' are still
1020 * in use by the driver, they should be reported as DONE
1022 if (unlikely(offset < (txq->added - txq->pending)))
1023 return RTE_ETH_TX_DESC_FULL;
1026 * There is no separate return value for unused descriptors;
1027 * the latter will be reported as DONE because genuine DONE
1028 * descriptors will be freed anyway in SW on the next burst
1030 return RTE_ETH_TX_DESC_DONE;
1033 struct sfc_dp_tx sfc_efx_tx = {
1035 .name = SFC_KVARG_DATAPATH_EFX,
1039 .features = SFC_DP_TX_FEAT_VLAN_INSERT |
1040 SFC_DP_TX_FEAT_TSO |
1041 SFC_DP_TX_FEAT_MULTI_POOL |
1042 SFC_DP_TX_FEAT_REFCNT |
1043 SFC_DP_TX_FEAT_MULTI_SEG,
1044 .qcreate = sfc_efx_tx_qcreate,
1045 .qdestroy = sfc_efx_tx_qdestroy,
1046 .qstart = sfc_efx_tx_qstart,
1047 .qstop = sfc_efx_tx_qstop,
1048 .qreap = sfc_efx_tx_qreap,
1049 .qdesc_status = sfc_efx_tx_qdesc_status,
1050 .pkt_burst = sfc_efx_xmit_pkts,