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.
18 #include "efx_types.h"
20 #include "efx_regs_ef10.h"
22 #include "sfc_dp_tx.h"
23 #include "sfc_tweak.h"
24 #include "sfc_kvargs.h"
28 #define sfc_ef10_tx_err(dpq, ...) \
29 SFC_DP_LOG(SFC_KVARG_DATAPATH_EF10, ERR, dpq, __VA_ARGS__)
31 /** Maximum length of the DMA descriptor data */
32 #define SFC_EF10_TX_DMA_DESC_LEN_MAX \
33 ((1u << ESF_DZ_TX_KER_BYTE_CNT_WIDTH) - 1)
36 * Maximum number of descriptors/buffers in the Tx ring.
37 * It should guarantee that corresponding event queue never overfill.
38 * EF10 native datapath uses event queue of the same size as Tx queue.
39 * Maximum number of events on datapath can be estimated as number of
40 * Tx queue entries (one event per Tx buffer in the worst case) plus
41 * Tx error and flush events.
43 #define SFC_EF10_TXQ_LIMIT(_ndesc) \
44 ((_ndesc) - 1 /* head must not step on tail */ - \
45 (SFC_EF10_EV_PER_CACHE_LINE - 1) /* max unused EvQ entries */ - \
46 1 /* Rx error */ - 1 /* flush */)
48 struct sfc_ef10_tx_sw_desc {
49 struct rte_mbuf *mbuf;
54 #define SFC_EF10_TXQ_STARTED 0x1
55 #define SFC_EF10_TXQ_NOT_RUNNING 0x2
56 #define SFC_EF10_TXQ_EXCEPTION 0x4
58 unsigned int ptr_mask;
60 unsigned int completed;
61 unsigned int max_fill_level;
62 unsigned int free_thresh;
63 unsigned int evq_read_ptr;
64 struct sfc_ef10_tx_sw_desc *sw_ring;
65 efx_qword_t *txq_hw_ring;
66 volatile void *doorbell;
67 efx_qword_t *evq_hw_ring;
70 uint16_t tso_tcp_header_offset_limit;
72 /* Datapath transmit queue anchor */
76 static inline struct sfc_ef10_txq *
77 sfc_ef10_txq_by_dp_txq(struct sfc_dp_txq *dp_txq)
79 return container_of(dp_txq, struct sfc_ef10_txq, dp);
83 sfc_ef10_tx_get_event(struct sfc_ef10_txq *txq, efx_qword_t *tx_ev)
85 volatile efx_qword_t *evq_hw_ring = txq->evq_hw_ring;
88 * Exception flag is set when reap is done.
89 * It is never done twice per packet burst get and absence of
90 * the flag is checked on burst get entry.
92 SFC_ASSERT((txq->flags & SFC_EF10_TXQ_EXCEPTION) == 0);
94 *tx_ev = evq_hw_ring[txq->evq_read_ptr & txq->ptr_mask];
96 if (!sfc_ef10_ev_present(*tx_ev))
99 if (unlikely(EFX_QWORD_FIELD(*tx_ev, FSF_AZ_EV_CODE) !=
100 FSE_AZ_EV_CODE_TX_EV)) {
102 * Do not move read_ptr to keep the event for exception
103 * handling by the control path.
105 txq->flags |= SFC_EF10_TXQ_EXCEPTION;
106 sfc_ef10_tx_err(&txq->dp.dpq,
107 "TxQ exception at EvQ read ptr %#x",
117 sfc_ef10_tx_process_events(struct sfc_ef10_txq *txq)
119 const unsigned int curr_done = txq->completed - 1;
120 unsigned int anew_done = curr_done;
123 while (sfc_ef10_tx_get_event(txq, &tx_ev)) {
125 * DROP_EVENT is an internal to the NIC, software should
126 * never see it and, therefore, may ignore it.
129 /* Update the latest done descriptor */
130 anew_done = EFX_QWORD_FIELD(tx_ev, ESF_DZ_TX_DESCR_INDX);
132 return (anew_done - curr_done) & txq->ptr_mask;
136 sfc_ef10_tx_reap(struct sfc_ef10_txq *txq)
138 const unsigned int old_read_ptr = txq->evq_read_ptr;
139 const unsigned int ptr_mask = txq->ptr_mask;
140 unsigned int completed = txq->completed;
141 unsigned int pending = completed;
143 pending += sfc_ef10_tx_process_events(txq);
145 if (pending != completed) {
146 struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
150 struct sfc_ef10_tx_sw_desc *txd;
153 txd = &txq->sw_ring[completed & ptr_mask];
154 if (txd->mbuf == NULL)
157 m = rte_pktmbuf_prefree_seg(txd->mbuf);
162 if ((nb == RTE_DIM(bulk)) ||
163 ((nb != 0) && (m->pool != bulk[0]->pool))) {
164 rte_mempool_put_bulk(bulk[0]->pool,
170 } while (++completed != pending);
173 rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
175 txq->completed = completed;
178 sfc_ef10_ev_qclear(txq->evq_hw_ring, ptr_mask, old_read_ptr,
183 sfc_ef10_tx_qdesc_dma_create(rte_iova_t addr, uint16_t size, bool eop,
186 EFX_POPULATE_QWORD_4(*edp,
187 ESF_DZ_TX_KER_TYPE, 0,
188 ESF_DZ_TX_KER_CONT, !eop,
189 ESF_DZ_TX_KER_BYTE_CNT, size,
190 ESF_DZ_TX_KER_BUF_ADDR, addr);
194 sfc_ef10_tx_qdesc_tso2_create(struct sfc_ef10_txq * const txq,
195 unsigned int added, uint16_t ipv4_id,
196 uint16_t outer_ipv4_id, uint32_t tcp_seq,
199 EFX_POPULATE_QWORD_5(txq->txq_hw_ring[added & txq->ptr_mask],
200 ESF_DZ_TX_DESC_IS_OPT, 1,
201 ESF_DZ_TX_OPTION_TYPE,
202 ESE_DZ_TX_OPTION_DESC_TSO,
203 ESF_DZ_TX_TSO_OPTION_TYPE,
204 ESE_DZ_TX_TSO_OPTION_DESC_FATSO2A,
205 ESF_DZ_TX_TSO_IP_ID, ipv4_id,
206 ESF_DZ_TX_TSO_TCP_SEQNO, tcp_seq);
207 EFX_POPULATE_QWORD_5(txq->txq_hw_ring[(added + 1) & txq->ptr_mask],
208 ESF_DZ_TX_DESC_IS_OPT, 1,
209 ESF_DZ_TX_OPTION_TYPE,
210 ESE_DZ_TX_OPTION_DESC_TSO,
211 ESF_DZ_TX_TSO_OPTION_TYPE,
212 ESE_DZ_TX_TSO_OPTION_DESC_FATSO2B,
213 ESF_DZ_TX_TSO_TCP_MSS, tcp_mss,
214 ESF_DZ_TX_TSO_OUTER_IPID, outer_ipv4_id);
218 sfc_ef10_tx_qpush(struct sfc_ef10_txq *txq, unsigned int added,
225 * This improves performance by pushing a TX descriptor at the same
226 * time as the doorbell. The descriptor must be added to the TXQ,
227 * so that can be used if the hardware decides not to use the pushed
230 desc.eq_u64[0] = txq->txq_hw_ring[pushed & txq->ptr_mask].eq_u64[0];
231 EFX_POPULATE_OWORD_3(oword,
232 ERF_DZ_TX_DESC_WPTR, added & txq->ptr_mask,
233 ERF_DZ_TX_DESC_HWORD, EFX_QWORD_FIELD(desc, EFX_DWORD_1),
234 ERF_DZ_TX_DESC_LWORD, EFX_QWORD_FIELD(desc, EFX_DWORD_0));
236 /* DMA sync to device is not required */
239 * rte_io_wmb() which guarantees that the STORE operations
240 * (i.e. Tx and event descriptor updates) that precede
241 * the rte_io_wmb() call are visible to NIC before the STORE
242 * operations that follow it (i.e. doorbell write).
246 *(volatile __m128i *)txq->doorbell = oword.eo_u128[0];
250 sfc_ef10_tx_pkt_descs_max(const struct rte_mbuf *m)
252 unsigned int extra_descs_per_seg;
253 unsigned int extra_descs_per_pkt;
256 * VLAN offload is not supported yet, so no extra descriptors
257 * are required for VLAN option descriptor.
260 /** Maximum length of the mbuf segment data */
261 #define SFC_MBUF_SEG_LEN_MAX UINT16_MAX
262 RTE_BUILD_BUG_ON(sizeof(m->data_len) != 2);
265 * Each segment is already counted once below. So, calculate
266 * how many extra DMA descriptors may be required per segment in
267 * the worst case because of maximum DMA descriptor length limit.
268 * If maximum segment length is less or equal to maximum DMA
269 * descriptor length, no extra DMA descriptors are required.
271 extra_descs_per_seg =
272 (SFC_MBUF_SEG_LEN_MAX - 1) / SFC_EF10_TX_DMA_DESC_LEN_MAX;
274 /** Maximum length of the packet */
275 #define SFC_MBUF_PKT_LEN_MAX UINT32_MAX
276 RTE_BUILD_BUG_ON(sizeof(m->pkt_len) != 4);
279 * One more limitation on maximum number of extra DMA descriptors
280 * comes from slicing entire packet because of DMA descriptor length
281 * limit taking into account that there is at least one segment
282 * which is already counted below (so division of the maximum
283 * packet length minus one with round down).
284 * TSO is not supported yet, so packet length is limited by
287 extra_descs_per_pkt =
288 (RTE_MIN((unsigned int)EFX_MAC_PDU_MAX,
289 SFC_MBUF_PKT_LEN_MAX) - 1) /
290 SFC_EF10_TX_DMA_DESC_LEN_MAX;
292 return m->nb_segs + RTE_MIN(m->nb_segs * extra_descs_per_seg,
293 extra_descs_per_pkt);
297 sfc_ef10_try_reap(struct sfc_ef10_txq * const txq, unsigned int added,
298 unsigned int needed_desc, unsigned int *dma_desc_space,
304 if (added != txq->added) {
305 sfc_ef10_tx_qpush(txq, added, txq->added);
309 sfc_ef10_tx_reap(txq);
313 * Recalculate DMA descriptor space since Tx reap may change
314 * the number of completed descriptors
316 *dma_desc_space = txq->max_fill_level -
317 (added - txq->completed);
319 return (needed_desc <= *dma_desc_space);
323 sfc_ef10_xmit_tso_pkt(struct sfc_ef10_txq * const txq, struct rte_mbuf *m_seg,
324 unsigned int *added, unsigned int *dma_desc_space,
327 size_t iph_off = m_seg->l2_len;
328 size_t tcph_off = m_seg->l2_len + m_seg->l3_len;
329 size_t header_len = m_seg->l2_len + m_seg->l3_len + m_seg->l4_len;
330 /* Offset of the payload in the last segment that contains the header */
332 const struct tcp_hdr *th;
337 struct rte_mbuf *first_m_seg = m_seg;
338 unsigned int pkt_start = *added;
339 unsigned int needed_desc;
340 struct rte_mbuf *m_seg_to_free_up_to = first_m_seg;
343 /* Both checks may be done, so use bit OR to have only one branching */
344 if (unlikely((header_len > SFC_TSOH_STD_LEN) |
345 (tcph_off > txq->tso_tcp_header_offset_limit)))
349 * Preliminary estimation of required DMA descriptors, including extra
350 * descriptor for TSO header that is needed when the header is
351 * separated from payload in one segment. It does not include
352 * extra descriptors that may appear when a big segment is split across
353 * several descriptors.
355 needed_desc = m_seg->nb_segs +
356 (unsigned int)SFC_TSO_OPT_DESCS_NUM +
357 (unsigned int)SFC_TSO_HDR_DESCS_NUM;
359 if (needed_desc > *dma_desc_space &&
360 !sfc_ef10_try_reap(txq, pkt_start, needed_desc,
361 dma_desc_space, reap_done)) {
363 * If a future Tx reap may increase available DMA descriptor
364 * space, do not try to send the packet.
366 if (txq->completed != pkt_start)
369 * Do not allow to send packet if the maximum DMA
370 * descriptor space is not sufficient to hold TSO
371 * descriptors, header descriptor and at least 1
372 * segment descriptor.
374 if (*dma_desc_space < SFC_TSO_OPT_DESCS_NUM +
375 SFC_TSO_HDR_DESCS_NUM + 1)
379 /* Check if the header is not fragmented */
380 if (rte_pktmbuf_data_len(m_seg) >= header_len) {
381 hdr_addr = rte_pktmbuf_mtod(m_seg, uint8_t *);
382 hdr_iova = rte_mbuf_data_iova(m_seg);
383 if (rte_pktmbuf_data_len(m_seg) == header_len) {
385 * Associate header mbuf with header descriptor
386 * which is located after TSO descriptors.
388 txq->sw_ring[(pkt_start + SFC_TSO_OPT_DESCS_NUM) &
389 txq->ptr_mask].mbuf = m_seg;
394 * If there is no payload offset (payload starts at the
395 * beginning of a segment) then an extra descriptor for
396 * separated header is not needed.
403 unsigned int copied_segs;
404 unsigned int hdr_addr_off = (*added & txq->ptr_mask) *
407 hdr_addr = txq->tsoh + hdr_addr_off;
408 hdr_iova = txq->tsoh_iova + hdr_addr_off;
409 copied_segs = sfc_tso_prepare_header(hdr_addr, header_len,
412 m_seg_to_free_up_to = m_seg;
414 * Reduce the number of needed descriptors by the number of
415 * segments that entirely consist of header data.
417 needed_desc -= copied_segs;
419 /* Extra descriptor for separated header is not needed */
424 switch (first_m_seg->ol_flags & (PKT_TX_IPV4 | PKT_TX_IPV6)) {
426 const struct ipv4_hdr *iphe4;
428 iphe4 = (const struct ipv4_hdr *)(hdr_addr + iph_off);
429 rte_memcpy(&packet_id, &iphe4->packet_id, sizeof(uint16_t));
430 packet_id = rte_be_to_cpu_16(packet_id);
440 th = (const struct tcp_hdr *)(hdr_addr + tcph_off);
441 rte_memcpy(&sent_seq, &th->sent_seq, sizeof(uint32_t));
442 sent_seq = rte_be_to_cpu_32(sent_seq);
444 sfc_ef10_tx_qdesc_tso2_create(txq, *added, packet_id, 0, sent_seq,
445 first_m_seg->tso_segsz);
446 (*added) += SFC_TSO_OPT_DESCS_NUM;
448 sfc_ef10_tx_qdesc_dma_create(hdr_iova, header_len, false,
449 &txq->txq_hw_ring[(*added) & txq->ptr_mask]);
453 rte_iova_t next_frag = rte_mbuf_data_iova(m_seg);
454 unsigned int seg_len = rte_pktmbuf_data_len(m_seg);
462 rte_iova_t frag_addr = next_frag;
465 frag_len = RTE_MIN(seg_len,
466 SFC_EF10_TX_DMA_DESC_LEN_MAX);
468 next_frag += frag_len;
471 eop = (seg_len == 0 && m_seg->next == NULL);
473 id = (*added) & txq->ptr_mask;
477 * Initially we assume that one DMA descriptor is needed
478 * for every segment. When the segment is split across
479 * several DMA descriptors, increase the estimation.
481 needed_desc += (seg_len != 0);
484 * When no more descriptors can be added, but not all
485 * segments are processed.
487 if (*added - pkt_start == *dma_desc_space &&
489 !sfc_ef10_try_reap(txq, pkt_start, needed_desc,
490 dma_desc_space, reap_done)) {
492 struct rte_mbuf *m_next;
494 if (txq->completed != pkt_start) {
498 * Reset mbuf associations with added
501 for (i = pkt_start; i != *added; i++) {
502 id = i & txq->ptr_mask;
503 txq->sw_ring[id].mbuf = NULL;
508 /* Free the segments that cannot be sent */
509 for (m = m_seg->next; m != NULL; m = m_next) {
511 rte_pktmbuf_free_seg(m);
514 /* Ignore the rest of the segment */
518 sfc_ef10_tx_qdesc_dma_create(frag_addr, frag_len,
519 eop, &txq->txq_hw_ring[id]);
521 } while (seg_len != 0);
523 txq->sw_ring[id].mbuf = m_seg;
529 * Free segments which content was entirely copied to the TSO header
530 * memory space of Tx queue
532 for (m_seg = first_m_seg; m_seg != m_seg_to_free_up_to;) {
533 struct rte_mbuf *seg_to_free = m_seg;
536 rte_pktmbuf_free_seg(seg_to_free);
543 sfc_ef10_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
545 struct sfc_ef10_txq * const txq = sfc_ef10_txq_by_dp_txq(tx_queue);
547 unsigned int dma_desc_space;
549 struct rte_mbuf **pktp;
550 struct rte_mbuf **pktp_end;
552 if (unlikely(txq->flags &
553 (SFC_EF10_TXQ_NOT_RUNNING | SFC_EF10_TXQ_EXCEPTION)))
557 dma_desc_space = txq->max_fill_level - (added - txq->completed);
559 reap_done = (dma_desc_space < txq->free_thresh);
561 sfc_ef10_tx_reap(txq);
562 dma_desc_space = txq->max_fill_level - (added - txq->completed);
565 for (pktp = &tx_pkts[0], pktp_end = &tx_pkts[nb_pkts];
568 struct rte_mbuf *m_seg = *pktp;
569 unsigned int pkt_start = added;
572 if (likely(pktp + 1 != pktp_end))
573 rte_mbuf_prefetch_part1(pktp[1]);
575 if (m_seg->ol_flags & PKT_TX_TCP_SEG) {
578 rc = sfc_ef10_xmit_tso_pkt(txq, m_seg, &added,
579 &dma_desc_space, &reap_done);
583 /* Packet can be sent in following xmit calls */
584 if (likely(rc == ENOSPC))
588 * Packet cannot be sent, tell RTE that
589 * it is sent, but actually drop it and
590 * continue with another packet
592 rte_pktmbuf_free(*pktp);
596 goto dma_desc_space_update;
599 if (sfc_ef10_tx_pkt_descs_max(m_seg) > dma_desc_space) {
603 /* Push already prepared descriptors before polling */
604 if (added != txq->added) {
605 sfc_ef10_tx_qpush(txq, added, txq->added);
609 sfc_ef10_tx_reap(txq);
611 dma_desc_space = txq->max_fill_level -
612 (added - txq->completed);
613 if (sfc_ef10_tx_pkt_descs_max(m_seg) > dma_desc_space)
617 pkt_len = m_seg->pkt_len;
619 rte_iova_t seg_addr = rte_mbuf_data_iova(m_seg);
620 unsigned int seg_len = rte_pktmbuf_data_len(m_seg);
621 unsigned int id = added & txq->ptr_mask;
623 SFC_ASSERT(seg_len <= SFC_EF10_TX_DMA_DESC_LEN_MAX);
627 sfc_ef10_tx_qdesc_dma_create(seg_addr,
628 seg_len, (pkt_len == 0),
629 &txq->txq_hw_ring[id]);
632 * rte_pktmbuf_free() is commonly used in DPDK for
633 * recycling packets - the function checks every
634 * segment's reference counter and returns the
635 * buffer to its pool whenever possible;
636 * nevertheless, freeing mbuf segments one by one
637 * may entail some performance decline;
638 * from this point, sfc_efx_tx_reap() does the same job
639 * on its own and frees buffers in bulks (all mbufs
640 * within a bulk belong to the same pool);
641 * from this perspective, individual segment pointers
642 * must be associated with the corresponding SW
643 * descriptors independently so that only one loop
644 * is sufficient on reap to inspect all the buffers
646 txq->sw_ring[id].mbuf = m_seg;
650 } while ((m_seg = m_seg->next) != 0);
652 dma_desc_space_update:
653 dma_desc_space -= (added - pkt_start);
656 if (likely(added != txq->added)) {
657 sfc_ef10_tx_qpush(txq, added, txq->added);
661 #if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
663 sfc_ef10_tx_reap(txq);
666 return pktp - &tx_pkts[0];
670 sfc_ef10_simple_tx_reap(struct sfc_ef10_txq *txq)
672 const unsigned int old_read_ptr = txq->evq_read_ptr;
673 const unsigned int ptr_mask = txq->ptr_mask;
674 unsigned int completed = txq->completed;
675 unsigned int pending = completed;
677 pending += sfc_ef10_tx_process_events(txq);
679 if (pending != completed) {
680 struct rte_mbuf *bulk[SFC_TX_REAP_BULK_SIZE];
684 struct sfc_ef10_tx_sw_desc *txd;
686 txd = &txq->sw_ring[completed & ptr_mask];
688 if (nb == RTE_DIM(bulk)) {
689 rte_mempool_put_bulk(bulk[0]->pool,
694 bulk[nb++] = txd->mbuf;
695 } while (++completed != pending);
697 rte_mempool_put_bulk(bulk[0]->pool, (void *)bulk, nb);
699 txq->completed = completed;
702 sfc_ef10_ev_qclear(txq->evq_hw_ring, ptr_mask, old_read_ptr,
708 sfc_ef10_simple_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
711 struct sfc_ef10_txq * const txq = sfc_ef10_txq_by_dp_txq(tx_queue);
712 unsigned int ptr_mask;
714 unsigned int dma_desc_space;
716 struct rte_mbuf **pktp;
717 struct rte_mbuf **pktp_end;
719 if (unlikely(txq->flags &
720 (SFC_EF10_TXQ_NOT_RUNNING | SFC_EF10_TXQ_EXCEPTION)))
723 ptr_mask = txq->ptr_mask;
725 dma_desc_space = txq->max_fill_level - (added - txq->completed);
727 reap_done = (dma_desc_space < RTE_MAX(txq->free_thresh, nb_pkts));
729 sfc_ef10_simple_tx_reap(txq);
730 dma_desc_space = txq->max_fill_level - (added - txq->completed);
733 pktp_end = &tx_pkts[MIN(nb_pkts, dma_desc_space)];
734 for (pktp = &tx_pkts[0]; pktp != pktp_end; ++pktp) {
735 struct rte_mbuf *pkt = *pktp;
736 unsigned int id = added & ptr_mask;
738 SFC_ASSERT(rte_pktmbuf_data_len(pkt) <=
739 SFC_EF10_TX_DMA_DESC_LEN_MAX);
741 sfc_ef10_tx_qdesc_dma_create(rte_mbuf_data_iova(pkt),
742 rte_pktmbuf_data_len(pkt),
743 true, &txq->txq_hw_ring[id]);
745 txq->sw_ring[id].mbuf = pkt;
750 if (likely(added != txq->added)) {
751 sfc_ef10_tx_qpush(txq, added, txq->added);
755 #if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
757 sfc_ef10_simple_tx_reap(txq);
760 return pktp - &tx_pkts[0];
763 static sfc_dp_tx_get_dev_info_t sfc_ef10_get_dev_info;
765 sfc_ef10_get_dev_info(struct rte_eth_dev_info *dev_info)
768 * Number of descriptors just defines maximum number of pushed
769 * descriptors (fill level).
771 dev_info->tx_desc_lim.nb_min = 1;
772 dev_info->tx_desc_lim.nb_align = 1;
775 static sfc_dp_tx_qsize_up_rings_t sfc_ef10_tx_qsize_up_rings;
777 sfc_ef10_tx_qsize_up_rings(uint16_t nb_tx_desc,
778 unsigned int *txq_entries,
779 unsigned int *evq_entries,
780 unsigned int *txq_max_fill_level)
783 * rte_ethdev API guarantees that the number meets min, max and
784 * alignment requirements.
786 if (nb_tx_desc <= EFX_TXQ_MINNDESCS)
787 *txq_entries = EFX_TXQ_MINNDESCS;
789 *txq_entries = rte_align32pow2(nb_tx_desc);
791 *evq_entries = *txq_entries;
793 *txq_max_fill_level = RTE_MIN(nb_tx_desc,
794 SFC_EF10_TXQ_LIMIT(*evq_entries));
798 static sfc_dp_tx_qcreate_t sfc_ef10_tx_qcreate;
800 sfc_ef10_tx_qcreate(uint16_t port_id, uint16_t queue_id,
801 const struct rte_pci_addr *pci_addr, int socket_id,
802 const struct sfc_dp_tx_qcreate_info *info,
803 struct sfc_dp_txq **dp_txqp)
805 struct sfc_ef10_txq *txq;
809 if (info->txq_entries != info->evq_entries)
813 txq = rte_zmalloc_socket("sfc-ef10-txq", sizeof(*txq),
814 RTE_CACHE_LINE_SIZE, socket_id);
818 sfc_dp_queue_init(&txq->dp.dpq, port_id, queue_id, pci_addr);
821 txq->sw_ring = rte_calloc_socket("sfc-ef10-txq-sw_ring",
823 sizeof(*txq->sw_ring),
824 RTE_CACHE_LINE_SIZE, socket_id);
825 if (txq->sw_ring == NULL)
826 goto fail_sw_ring_alloc;
828 if (info->offloads & DEV_TX_OFFLOAD_TCP_TSO) {
829 txq->tsoh = rte_calloc_socket("sfc-ef10-txq-tsoh",
834 if (txq->tsoh == NULL)
835 goto fail_tsoh_alloc;
837 txq->tsoh_iova = rte_malloc_virt2iova(txq->tsoh);
840 txq->flags = SFC_EF10_TXQ_NOT_RUNNING;
841 txq->ptr_mask = info->txq_entries - 1;
842 txq->max_fill_level = info->max_fill_level;
843 txq->free_thresh = info->free_thresh;
844 txq->txq_hw_ring = info->txq_hw_ring;
845 txq->doorbell = (volatile uint8_t *)info->mem_bar +
846 ER_DZ_TX_DESC_UPD_REG_OFST +
847 (info->hw_index << info->vi_window_shift);
848 txq->evq_hw_ring = info->evq_hw_ring;
849 txq->tso_tcp_header_offset_limit = info->tso_tcp_header_offset_limit;
855 rte_free(txq->sw_ring);
865 static sfc_dp_tx_qdestroy_t sfc_ef10_tx_qdestroy;
867 sfc_ef10_tx_qdestroy(struct sfc_dp_txq *dp_txq)
869 struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
872 rte_free(txq->sw_ring);
876 static sfc_dp_tx_qstart_t sfc_ef10_tx_qstart;
878 sfc_ef10_tx_qstart(struct sfc_dp_txq *dp_txq, unsigned int evq_read_ptr,
879 unsigned int txq_desc_index)
881 struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
883 txq->evq_read_ptr = evq_read_ptr;
884 txq->added = txq->completed = txq_desc_index;
886 txq->flags |= SFC_EF10_TXQ_STARTED;
887 txq->flags &= ~(SFC_EF10_TXQ_NOT_RUNNING | SFC_EF10_TXQ_EXCEPTION);
892 static sfc_dp_tx_qstop_t sfc_ef10_tx_qstop;
894 sfc_ef10_tx_qstop(struct sfc_dp_txq *dp_txq, unsigned int *evq_read_ptr)
896 struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
898 txq->flags |= SFC_EF10_TXQ_NOT_RUNNING;
900 *evq_read_ptr = txq->evq_read_ptr;
903 static sfc_dp_tx_qtx_ev_t sfc_ef10_tx_qtx_ev;
905 sfc_ef10_tx_qtx_ev(struct sfc_dp_txq *dp_txq, __rte_unused unsigned int id)
907 __rte_unused struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
909 SFC_ASSERT(txq->flags & SFC_EF10_TXQ_NOT_RUNNING);
912 * It is safe to ignore Tx event since we reap all mbufs on
913 * queue purge anyway.
919 static sfc_dp_tx_qreap_t sfc_ef10_tx_qreap;
921 sfc_ef10_tx_qreap(struct sfc_dp_txq *dp_txq)
923 struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
924 unsigned int completed;
926 for (completed = txq->completed; completed != txq->added; ++completed) {
927 struct sfc_ef10_tx_sw_desc *txd;
929 txd = &txq->sw_ring[completed & txq->ptr_mask];
930 if (txd->mbuf != NULL) {
931 rte_pktmbuf_free_seg(txd->mbuf);
936 txq->flags &= ~SFC_EF10_TXQ_STARTED;
940 sfc_ef10_tx_qdesc_npending(struct sfc_ef10_txq *txq)
942 const unsigned int curr_done = txq->completed - 1;
943 unsigned int anew_done = curr_done;
945 const unsigned int evq_old_read_ptr = txq->evq_read_ptr;
947 if (unlikely(txq->flags &
948 (SFC_EF10_TXQ_NOT_RUNNING | SFC_EF10_TXQ_EXCEPTION)))
951 while (sfc_ef10_tx_get_event(txq, &tx_ev))
952 anew_done = EFX_QWORD_FIELD(tx_ev, ESF_DZ_TX_DESCR_INDX);
955 * The function does not process events, so return event queue read
956 * pointer to the original position to allow the events that were
957 * read to be processed later
959 txq->evq_read_ptr = evq_old_read_ptr;
961 return (anew_done - curr_done) & txq->ptr_mask;
964 static sfc_dp_tx_qdesc_status_t sfc_ef10_tx_qdesc_status;
966 sfc_ef10_tx_qdesc_status(struct sfc_dp_txq *dp_txq,
969 struct sfc_ef10_txq *txq = sfc_ef10_txq_by_dp_txq(dp_txq);
970 unsigned int npending = sfc_ef10_tx_qdesc_npending(txq);
972 if (unlikely(offset > txq->ptr_mask))
975 if (unlikely(offset >= txq->max_fill_level))
976 return RTE_ETH_TX_DESC_UNAVAIL;
978 if (unlikely(offset < npending))
979 return RTE_ETH_TX_DESC_FULL;
981 return RTE_ETH_TX_DESC_DONE;
984 struct sfc_dp_tx sfc_ef10_tx = {
986 .name = SFC_KVARG_DATAPATH_EF10,
988 .hw_fw_caps = SFC_DP_HW_FW_CAP_EF10,
990 .features = SFC_DP_TX_FEAT_TSO |
991 SFC_DP_TX_FEAT_MULTI_SEG |
992 SFC_DP_TX_FEAT_MULTI_POOL |
993 SFC_DP_TX_FEAT_REFCNT |
994 SFC_DP_TX_FEAT_MULTI_PROCESS,
995 .get_dev_info = sfc_ef10_get_dev_info,
996 .qsize_up_rings = sfc_ef10_tx_qsize_up_rings,
997 .qcreate = sfc_ef10_tx_qcreate,
998 .qdestroy = sfc_ef10_tx_qdestroy,
999 .qstart = sfc_ef10_tx_qstart,
1000 .qtx_ev = sfc_ef10_tx_qtx_ev,
1001 .qstop = sfc_ef10_tx_qstop,
1002 .qreap = sfc_ef10_tx_qreap,
1003 .qdesc_status = sfc_ef10_tx_qdesc_status,
1004 .pkt_burst = sfc_ef10_xmit_pkts,
1007 struct sfc_dp_tx sfc_ef10_simple_tx = {
1009 .name = SFC_KVARG_DATAPATH_EF10_SIMPLE,
1012 .features = SFC_DP_TX_FEAT_MULTI_PROCESS,
1013 .get_dev_info = sfc_ef10_get_dev_info,
1014 .qsize_up_rings = sfc_ef10_tx_qsize_up_rings,
1015 .qcreate = sfc_ef10_tx_qcreate,
1016 .qdestroy = sfc_ef10_tx_qdestroy,
1017 .qstart = sfc_ef10_tx_qstart,
1018 .qtx_ev = sfc_ef10_tx_qtx_ev,
1019 .qstop = sfc_ef10_tx_qstop,
1020 .qreap = sfc_ef10_tx_qreap,
1021 .qdesc_status = sfc_ef10_tx_qdesc_status,
1022 .pkt_burst = sfc_ef10_simple_xmit_pkts,