1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2014-2021 Broadcom
9 #include <rte_bitmap.h>
10 #include <rte_byteorder.h>
11 #include <rte_malloc.h>
12 #include <rte_memory.h>
15 #include "bnxt_reps.h"
16 #include "bnxt_ring.h"
19 #include "hsi_struct_def_dpdk.h"
20 #ifdef RTE_LIBRTE_IEEE1588
21 #include "bnxt_hwrm.h"
24 #include <bnxt_tf_common.h>
25 #include <ulp_mark_mgr.h>
31 static inline struct rte_mbuf *__bnxt_alloc_rx_data(struct rte_mempool *mb)
33 struct rte_mbuf *data;
35 data = rte_mbuf_raw_alloc(mb);
40 static inline int bnxt_alloc_rx_data(struct bnxt_rx_queue *rxq,
41 struct bnxt_rx_ring_info *rxr,
44 uint16_t prod = RING_IDX(rxr->rx_ring_struct, raw_prod);
45 struct rx_prod_pkt_bd *rxbd;
46 struct rte_mbuf **rx_buf;
47 struct rte_mbuf *mbuf;
49 rxbd = &rxr->rx_desc_ring[prod];
50 rx_buf = &rxr->rx_buf_ring[prod];
51 mbuf = __bnxt_alloc_rx_data(rxq->mb_pool);
53 rte_atomic64_inc(&rxq->rx_mbuf_alloc_fail);
58 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
60 rxbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
65 static inline int bnxt_alloc_ag_data(struct bnxt_rx_queue *rxq,
66 struct bnxt_rx_ring_info *rxr,
69 uint16_t prod = RING_IDX(rxr->ag_ring_struct, raw_prod);
70 struct rx_prod_pkt_bd *rxbd;
71 struct rte_mbuf **rx_buf;
72 struct rte_mbuf *mbuf;
74 rxbd = &rxr->ag_desc_ring[prod];
75 rx_buf = &rxr->ag_buf_ring[prod];
77 PMD_DRV_LOG(ERR, "Jumbo Frame. rxbd is NULL\n");
82 PMD_DRV_LOG(ERR, "Jumbo Frame. rx_buf is NULL\n");
86 mbuf = __bnxt_alloc_rx_data(rxq->mb_pool);
88 rte_atomic64_inc(&rxq->rx_mbuf_alloc_fail);
93 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
95 rxbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
100 static inline void bnxt_reuse_rx_mbuf(struct bnxt_rx_ring_info *rxr,
101 struct rte_mbuf *mbuf)
103 uint16_t prod, raw_prod = RING_NEXT(rxr->rx_raw_prod);
104 struct rte_mbuf **prod_rx_buf;
105 struct rx_prod_pkt_bd *prod_bd;
107 prod = RING_IDX(rxr->rx_ring_struct, raw_prod);
108 prod_rx_buf = &rxr->rx_buf_ring[prod];
110 RTE_ASSERT(*prod_rx_buf == NULL);
111 RTE_ASSERT(mbuf != NULL);
115 prod_bd = &rxr->rx_desc_ring[prod];
117 prod_bd->address = rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
119 rxr->rx_raw_prod = raw_prod;
123 struct rte_mbuf *bnxt_consume_rx_buf(struct bnxt_rx_ring_info *rxr,
126 struct rte_mbuf **cons_rx_buf;
127 struct rte_mbuf *mbuf;
129 cons_rx_buf = &rxr->rx_buf_ring[RING_IDX(rxr->rx_ring_struct, cons)];
130 RTE_ASSERT(*cons_rx_buf != NULL);
137 static void bnxt_tpa_get_metadata(struct bnxt *bp,
138 struct bnxt_tpa_info *tpa_info,
139 struct rx_tpa_start_cmpl *tpa_start,
140 struct rx_tpa_start_cmpl_hi *tpa_start1)
142 tpa_info->cfa_code_valid = 0;
143 tpa_info->vlan_valid = 0;
144 tpa_info->hash_valid = 0;
145 tpa_info->l4_csum_valid = 0;
147 if (likely(tpa_start->flags_type &
148 rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS_RSS_VALID))) {
149 tpa_info->hash_valid = 1;
150 tpa_info->rss_hash = rte_le_to_cpu_32(tpa_start->rss_hash);
153 if (bp->vnic_cap_flags & BNXT_VNIC_CAP_RX_CMPL_V2) {
154 struct rx_tpa_start_v2_cmpl *v2_tpa_start = (void *)tpa_start;
155 struct rx_tpa_start_v2_cmpl_hi *v2_tpa_start1 =
158 if (v2_tpa_start->agg_id &
159 RX_TPA_START_V2_CMPL_METADATA1_VALID) {
160 tpa_info->vlan_valid = 1;
162 rte_le_to_cpu_16(v2_tpa_start1->metadata0);
165 if (v2_tpa_start1->flags2 & RX_CMP_FLAGS2_L4_CSUM_ALL_OK_MASK)
166 tpa_info->l4_csum_valid = 1;
171 tpa_info->cfa_code_valid = 1;
172 tpa_info->cfa_code = rte_le_to_cpu_16(tpa_start1->cfa_code);
173 if (tpa_start1->flags2 &
174 rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS2_META_FORMAT_VLAN)) {
175 tpa_info->vlan_valid = 1;
176 tpa_info->vlan = rte_le_to_cpu_32(tpa_start1->metadata);
179 if (likely(tpa_start1->flags2 &
180 rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS2_L4_CS_CALC)))
181 tpa_info->l4_csum_valid = 1;
184 static void bnxt_tpa_start(struct bnxt_rx_queue *rxq,
185 struct rx_tpa_start_cmpl *tpa_start,
186 struct rx_tpa_start_cmpl_hi *tpa_start1)
188 struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
191 struct bnxt_tpa_info *tpa_info;
192 struct rte_mbuf *mbuf;
194 agg_id = bnxt_tpa_start_agg_id(rxq->bp, tpa_start);
196 data_cons = tpa_start->opaque;
197 tpa_info = &rxr->tpa_info[agg_id];
199 mbuf = bnxt_consume_rx_buf(rxr, data_cons);
201 bnxt_reuse_rx_mbuf(rxr, tpa_info->mbuf);
203 tpa_info->agg_count = 0;
204 tpa_info->mbuf = mbuf;
205 tpa_info->len = rte_le_to_cpu_32(tpa_start->len);
207 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
210 mbuf->pkt_len = rte_le_to_cpu_32(tpa_start->len);
211 mbuf->data_len = mbuf->pkt_len;
212 mbuf->port = rxq->port_id;
213 mbuf->ol_flags = PKT_RX_LRO;
215 bnxt_tpa_get_metadata(rxq->bp, tpa_info, tpa_start, tpa_start1);
217 if (likely(tpa_info->hash_valid)) {
218 mbuf->hash.rss = tpa_info->rss_hash;
219 mbuf->ol_flags |= PKT_RX_RSS_HASH;
220 } else if (tpa_info->cfa_code_valid) {
221 mbuf->hash.fdir.id = tpa_info->cfa_code;
222 mbuf->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
225 if (tpa_info->vlan_valid) {
226 mbuf->vlan_tci = tpa_info->vlan;
227 mbuf->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
230 if (likely(tpa_info->l4_csum_valid))
231 mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
233 /* recycle next mbuf */
234 data_cons = RING_NEXT(data_cons);
235 bnxt_reuse_rx_mbuf(rxr, bnxt_consume_rx_buf(rxr, data_cons));
238 static int bnxt_agg_bufs_valid(struct bnxt_cp_ring_info *cpr,
239 uint8_t agg_bufs, uint32_t raw_cp_cons)
241 uint16_t last_cp_cons;
242 struct rx_pkt_cmpl *agg_cmpl;
244 raw_cp_cons = ADV_RAW_CMP(raw_cp_cons, agg_bufs);
245 last_cp_cons = RING_CMP(cpr->cp_ring_struct, raw_cp_cons);
246 agg_cmpl = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[last_cp_cons];
247 cpr->valid = FLIP_VALID(raw_cp_cons,
248 cpr->cp_ring_struct->ring_mask,
250 return CMP_VALID(agg_cmpl, raw_cp_cons, cpr->cp_ring_struct);
253 /* TPA consume agg buffer out of order, allocate connected data only */
254 static int bnxt_prod_ag_mbuf(struct bnxt_rx_queue *rxq)
256 struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
257 uint16_t raw_next = RING_NEXT(rxr->ag_raw_prod);
258 uint16_t bmap_next = RING_IDX(rxr->ag_ring_struct, raw_next);
260 /* TODO batch allocation for better performance */
261 while (rte_bitmap_get(rxr->ag_bitmap, bmap_next)) {
262 if (unlikely(bnxt_alloc_ag_data(rxq, rxr, raw_next))) {
263 PMD_DRV_LOG(ERR, "agg mbuf alloc failed: prod=0x%x\n",
267 rte_bitmap_clear(rxr->ag_bitmap, bmap_next);
268 rxr->ag_raw_prod = raw_next;
269 raw_next = RING_NEXT(raw_next);
270 bmap_next = RING_IDX(rxr->ag_ring_struct, raw_next);
276 static int bnxt_rx_pages(struct bnxt_rx_queue *rxq,
277 struct rte_mbuf *mbuf, uint32_t *tmp_raw_cons,
278 uint8_t agg_buf, struct bnxt_tpa_info *tpa_info)
280 struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
281 struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
283 uint16_t cp_cons, ag_cons;
284 struct rx_pkt_cmpl *rxcmp;
285 struct rte_mbuf *last = mbuf;
286 bool is_p5_tpa = tpa_info && BNXT_CHIP_P5(rxq->bp);
288 for (i = 0; i < agg_buf; i++) {
289 struct rte_mbuf **ag_buf;
290 struct rte_mbuf *ag_mbuf;
293 rxcmp = (void *)&tpa_info->agg_arr[i];
295 *tmp_raw_cons = NEXT_RAW_CMP(*tmp_raw_cons);
296 cp_cons = RING_CMP(cpr->cp_ring_struct, *tmp_raw_cons);
297 rxcmp = (struct rx_pkt_cmpl *)
298 &cpr->cp_desc_ring[cp_cons];
302 bnxt_dump_cmpl(cp_cons, rxcmp);
305 ag_cons = rxcmp->opaque;
306 RTE_ASSERT(ag_cons <= rxr->ag_ring_struct->ring_mask);
307 ag_buf = &rxr->ag_buf_ring[ag_cons];
309 RTE_ASSERT(ag_mbuf != NULL);
311 ag_mbuf->data_len = rte_le_to_cpu_16(rxcmp->len);
314 mbuf->pkt_len += ag_mbuf->data_len;
316 last->next = ag_mbuf;
322 * As aggregation buffer consumed out of order in TPA module,
323 * use bitmap to track freed slots to be allocated and notified
326 rte_bitmap_set(rxr->ag_bitmap, ag_cons);
329 bnxt_prod_ag_mbuf(rxq);
333 static inline struct rte_mbuf *bnxt_tpa_end(
334 struct bnxt_rx_queue *rxq,
335 uint32_t *raw_cp_cons,
336 struct rx_tpa_end_cmpl *tpa_end,
337 struct rx_tpa_end_cmpl_hi *tpa_end1)
339 struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
340 struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
342 struct rte_mbuf *mbuf;
344 uint8_t payload_offset;
345 struct bnxt_tpa_info *tpa_info;
347 if (BNXT_CHIP_P5(rxq->bp)) {
348 struct rx_tpa_v2_end_cmpl *th_tpa_end;
349 struct rx_tpa_v2_end_cmpl_hi *th_tpa_end1;
351 th_tpa_end = (void *)tpa_end;
352 th_tpa_end1 = (void *)tpa_end1;
353 agg_id = BNXT_TPA_END_AGG_ID_TH(th_tpa_end);
354 agg_bufs = BNXT_TPA_END_AGG_BUFS_TH(th_tpa_end1);
355 payload_offset = th_tpa_end1->payload_offset;
357 agg_id = BNXT_TPA_END_AGG_ID(tpa_end);
358 agg_bufs = BNXT_TPA_END_AGG_BUFS(tpa_end);
359 if (!bnxt_agg_bufs_valid(cpr, agg_bufs, *raw_cp_cons))
361 payload_offset = tpa_end->payload_offset;
364 tpa_info = &rxr->tpa_info[agg_id];
365 mbuf = tpa_info->mbuf;
366 RTE_ASSERT(mbuf != NULL);
369 bnxt_rx_pages(rxq, mbuf, raw_cp_cons, agg_bufs, tpa_info);
371 mbuf->l4_len = payload_offset;
373 struct rte_mbuf *new_data = __bnxt_alloc_rx_data(rxq->mb_pool);
374 RTE_ASSERT(new_data != NULL);
376 rte_atomic64_inc(&rxq->rx_mbuf_alloc_fail);
379 tpa_info->mbuf = new_data;
384 uint32_t bnxt_ptype_table[BNXT_PTYPE_TBL_DIM] __rte_cache_aligned;
386 static void __rte_cold
387 bnxt_init_ptype_table(void)
389 uint32_t *pt = bnxt_ptype_table;
390 static bool initialized;
398 for (i = 0; i < BNXT_PTYPE_TBL_DIM; i++) {
399 if (i & (RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN >> 2))
400 pt[i] = RTE_PTYPE_L2_ETHER_VLAN;
402 pt[i] = RTE_PTYPE_L2_ETHER;
404 ip6 = i & (RX_PKT_CMPL_FLAGS2_IP_TYPE >> 7);
405 tun = i & (RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC >> 2);
406 type = (i & 0x78) << 9;
409 l3 = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
410 else if (!tun && ip6)
411 l3 = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
412 else if (tun && !ip6)
413 l3 = RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN;
415 l3 = RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN;
418 case RX_PKT_CMPL_FLAGS_ITYPE_ICMP:
420 pt[i] |= l3 | RTE_PTYPE_INNER_L4_ICMP;
422 pt[i] |= l3 | RTE_PTYPE_L4_ICMP;
424 case RX_PKT_CMPL_FLAGS_ITYPE_TCP:
426 pt[i] |= l3 | RTE_PTYPE_INNER_L4_TCP;
428 pt[i] |= l3 | RTE_PTYPE_L4_TCP;
430 case RX_PKT_CMPL_FLAGS_ITYPE_UDP:
432 pt[i] |= l3 | RTE_PTYPE_INNER_L4_UDP;
434 pt[i] |= l3 | RTE_PTYPE_L4_UDP;
436 case RX_PKT_CMPL_FLAGS_ITYPE_IP:
445 bnxt_parse_pkt_type(struct rx_pkt_cmpl *rxcmp, struct rx_pkt_cmpl_hi *rxcmp1)
447 uint32_t flags_type, flags2;
450 flags_type = rte_le_to_cpu_16(rxcmp->flags_type);
451 flags2 = rte_le_to_cpu_32(rxcmp1->flags2);
455 * bit 0: RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC
456 * bit 1: RX_CMPL_FLAGS2_IP_TYPE
457 * bit 2: RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN
458 * bits 3-6: RX_PKT_CMPL_FLAGS_ITYPE
460 index = ((flags_type & RX_PKT_CMPL_FLAGS_ITYPE_MASK) >> 9) |
461 ((flags2 & (RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN |
462 RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC)) >> 2) |
463 ((flags2 & RX_PKT_CMPL_FLAGS2_IP_TYPE) >> 7);
465 return bnxt_ptype_table[index];
468 static void __rte_cold
469 bnxt_init_ol_flags_tables(struct bnxt_rx_queue *rxq)
471 struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
472 struct rte_eth_conf *dev_conf;
473 bool outer_cksum_enabled;
478 dev_conf = &rxq->bp->eth_dev->data->dev_conf;
479 offloads = dev_conf->rxmode.offloads;
481 outer_cksum_enabled = !!(offloads & (DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM |
482 DEV_RX_OFFLOAD_OUTER_UDP_CKSUM));
484 /* Initialize ol_flags table. */
485 pt = rxr->ol_flags_table;
486 for (i = 0; i < BNXT_OL_FLAGS_TBL_DIM; i++) {
489 if (i & RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN)
490 pt[i] |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
492 if (i & (RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC << 3)) {
494 if (outer_cksum_enabled) {
495 if (i & RX_PKT_CMPL_FLAGS2_IP_CS_CALC)
496 pt[i] |= PKT_RX_IP_CKSUM_GOOD;
498 if (i & RX_PKT_CMPL_FLAGS2_L4_CS_CALC)
499 pt[i] |= PKT_RX_L4_CKSUM_GOOD;
501 if (i & RX_PKT_CMPL_FLAGS2_T_L4_CS_CALC)
502 pt[i] |= PKT_RX_OUTER_L4_CKSUM_GOOD;
504 if (i & RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC)
505 pt[i] |= PKT_RX_IP_CKSUM_GOOD;
507 if (i & RX_PKT_CMPL_FLAGS2_T_L4_CS_CALC)
508 pt[i] |= PKT_RX_L4_CKSUM_GOOD;
511 /* Non-tunnel case. */
512 if (i & RX_PKT_CMPL_FLAGS2_IP_CS_CALC)
513 pt[i] |= PKT_RX_IP_CKSUM_GOOD;
515 if (i & RX_PKT_CMPL_FLAGS2_L4_CS_CALC)
516 pt[i] |= PKT_RX_L4_CKSUM_GOOD;
520 /* Initialize checksum error table. */
521 pt = rxr->ol_flags_err_table;
522 for (i = 0; i < BNXT_OL_FLAGS_ERR_TBL_DIM; i++) {
525 if (i & (RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC << 2)) {
527 if (outer_cksum_enabled) {
528 if (i & (RX_PKT_CMPL_ERRORS_IP_CS_ERROR >> 4))
529 pt[i] |= PKT_RX_IP_CKSUM_BAD;
531 if (i & (RX_PKT_CMPL_ERRORS_T_IP_CS_ERROR >> 4))
532 pt[i] |= PKT_RX_EIP_CKSUM_BAD;
534 if (i & (RX_PKT_CMPL_ERRORS_L4_CS_ERROR >> 4))
535 pt[i] |= PKT_RX_L4_CKSUM_BAD;
537 if (i & (RX_PKT_CMPL_ERRORS_T_L4_CS_ERROR >> 4))
538 pt[i] |= PKT_RX_OUTER_L4_CKSUM_BAD;
540 if (i & (RX_PKT_CMPL_ERRORS_T_IP_CS_ERROR >> 4))
541 pt[i] |= PKT_RX_IP_CKSUM_BAD;
543 if (i & (RX_PKT_CMPL_ERRORS_T_L4_CS_ERROR >> 4))
544 pt[i] |= PKT_RX_L4_CKSUM_BAD;
547 /* Non-tunnel case. */
548 if (i & (RX_PKT_CMPL_ERRORS_IP_CS_ERROR >> 4))
549 pt[i] |= PKT_RX_IP_CKSUM_BAD;
551 if (i & (RX_PKT_CMPL_ERRORS_L4_CS_ERROR >> 4))
552 pt[i] |= PKT_RX_L4_CKSUM_BAD;
558 bnxt_set_ol_flags(struct bnxt_rx_ring_info *rxr, struct rx_pkt_cmpl *rxcmp,
559 struct rx_pkt_cmpl_hi *rxcmp1, struct rte_mbuf *mbuf)
561 uint16_t flags_type, errors, flags;
564 flags_type = rte_le_to_cpu_16(rxcmp->flags_type);
566 flags = rte_le_to_cpu_32(rxcmp1->flags2) &
567 (RX_PKT_CMPL_FLAGS2_IP_CS_CALC |
568 RX_PKT_CMPL_FLAGS2_L4_CS_CALC |
569 RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC |
570 RX_PKT_CMPL_FLAGS2_T_L4_CS_CALC |
571 RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN);
573 flags |= (flags & RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC) << 3;
574 errors = rte_le_to_cpu_16(rxcmp1->errors_v2) &
575 (RX_PKT_CMPL_ERRORS_IP_CS_ERROR |
576 RX_PKT_CMPL_ERRORS_L4_CS_ERROR |
577 RX_PKT_CMPL_ERRORS_T_IP_CS_ERROR |
578 RX_PKT_CMPL_ERRORS_T_L4_CS_ERROR);
579 errors = (errors >> 4) & flags;
581 ol_flags = rxr->ol_flags_table[flags & ~errors];
583 if (unlikely(errors)) {
584 errors |= (flags & RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC) << 2;
585 ol_flags |= rxr->ol_flags_err_table[errors];
588 if (flags_type & RX_PKT_CMPL_FLAGS_RSS_VALID) {
589 mbuf->hash.rss = rte_le_to_cpu_32(rxcmp->rss_hash);
590 ol_flags |= PKT_RX_RSS_HASH;
593 #ifdef RTE_LIBRTE_IEEE1588
594 if (unlikely((flags_type & RX_PKT_CMPL_FLAGS_MASK) ==
595 RX_PKT_CMPL_FLAGS_ITYPE_PTP_W_TIMESTAMP))
596 ol_flags |= PKT_RX_IEEE1588_PTP | PKT_RX_IEEE1588_TMST;
599 mbuf->ol_flags = ol_flags;
602 #ifdef RTE_LIBRTE_IEEE1588
604 bnxt_get_rx_ts_p5(struct bnxt *bp, uint32_t rx_ts_cmpl)
606 uint64_t systime_cycles = 0;
608 if (!BNXT_CHIP_P5(bp))
611 /* On Thor, Rx timestamps are provided directly in the
612 * Rx completion records to the driver. Only 32 bits of
613 * the timestamp is present in the completion. Driver needs
614 * to read the current 48 bit free running timer using the
615 * HWRM_PORT_TS_QUERY command and combine the upper 16 bits
616 * from the HWRM response with the lower 32 bits in the
617 * Rx completion to produce the 48 bit timestamp for the Rx packet
619 bnxt_hwrm_port_ts_query(bp, BNXT_PTP_FLAGS_CURRENT_TIME,
621 bp->ptp_cfg->rx_timestamp = (systime_cycles & 0xFFFF00000000);
622 bp->ptp_cfg->rx_timestamp |= rx_ts_cmpl;
627 bnxt_ulp_set_mark_in_mbuf(struct bnxt *bp, struct rx_pkt_cmpl_hi *rxcmp1,
628 struct rte_mbuf *mbuf, uint32_t *vfr_flag)
636 uint32_t gfid_support = 0;
639 if (BNXT_GFID_ENABLED(bp))
642 cfa_code = rte_le_to_cpu_16(rxcmp1->cfa_code);
643 flags2 = rte_le_to_cpu_32(rxcmp1->flags2);
644 meta = rte_le_to_cpu_32(rxcmp1->metadata);
647 * The flags field holds extra bits of info from [6:4]
648 * which indicate if the flow is in TCAM or EM or EEM
650 meta_fmt = (flags2 & BNXT_CFA_META_FMT_MASK) >>
651 BNXT_CFA_META_FMT_SHFT;
656 /* Not an LFID or GFID, a flush cmd. */
659 /* LFID mode, no vlan scenario */
667 * Assume that EM doesn't support Mark due to GFID
668 * collisions with EEM. Simply return without setting the mark
671 if (BNXT_CFA_META_EM_TEST(meta)) {
672 /*This is EM hit {EM(1), GFID[27:16], 19'd0 or vtag } */
674 meta >>= BNXT_RX_META_CFA_CODE_SHIFT;
675 cfa_code |= meta << BNXT_CFA_CODE_META_SHIFT;
678 * It is a TCAM entry, so it is an LFID.
679 * The TCAM IDX and Mode can also be determined
680 * by decoding the meta_data. We are not
681 * using these for now.
687 /* EEM Case, only using gfid in EEM for now. */
691 * For EEM flows, The first part of cfa_code is 16 bits.
692 * The second part is embedded in the
693 * metadata field from bit 19 onwards. The driver needs to
694 * ignore the first 19 bits of metadata and use the next 12
695 * bits as higher 12 bits of cfa_code.
697 meta >>= BNXT_RX_META_CFA_CODE_SHIFT;
698 cfa_code |= meta << BNXT_CFA_CODE_META_SHIFT;
701 /* For other values, the cfa_code is assumed to be an LFID. */
705 rc = ulp_mark_db_mark_get(bp->ulp_ctx, gfid,
706 cfa_code, vfr_flag, &mark_id);
708 /* VF to VFR Rx path. So, skip mark_id injection in mbuf */
709 if (vfr_flag && *vfr_flag)
711 /* Got the mark, write it to the mbuf and return */
712 mbuf->hash.fdir.hi = mark_id;
713 *bnxt_cfa_code_dynfield(mbuf) = cfa_code & 0xffffffffull;
714 mbuf->hash.fdir.id = rxcmp1->cfa_code;
715 mbuf->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
720 mbuf->hash.fdir.hi = 0;
721 mbuf->hash.fdir.id = 0;
726 void bnxt_set_mark_in_mbuf(struct bnxt *bp,
727 struct rx_pkt_cmpl_hi *rxcmp1,
728 struct rte_mbuf *mbuf)
730 uint32_t cfa_code = 0;
731 uint8_t meta_fmt = 0;
735 cfa_code = rte_le_to_cpu_16(rxcmp1->cfa_code);
739 if (cfa_code && !bp->mark_table[cfa_code].valid)
742 flags2 = rte_le_to_cpu_16(rxcmp1->flags2);
743 meta = rte_le_to_cpu_32(rxcmp1->metadata);
745 meta >>= BNXT_RX_META_CFA_CODE_SHIFT;
747 /* The flags field holds extra bits of info from [6:4]
748 * which indicate if the flow is in TCAM or EM or EEM
750 meta_fmt = (flags2 & BNXT_CFA_META_FMT_MASK) >>
751 BNXT_CFA_META_FMT_SHFT;
753 /* meta_fmt == 4 => 'b100 => 'b10x => EM.
754 * meta_fmt == 5 => 'b101 => 'b10x => EM + VLAN
755 * meta_fmt == 6 => 'b110 => 'b11x => EEM
756 * meta_fmt == 7 => 'b111 => 'b11x => EEM + VLAN.
758 meta_fmt >>= BNXT_CFA_META_FMT_EM_EEM_SHFT;
761 mbuf->hash.fdir.hi = bp->mark_table[cfa_code].mark_id;
762 mbuf->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
765 static int bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
766 struct bnxt_rx_queue *rxq, uint32_t *raw_cons)
768 struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
769 struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
770 struct rx_pkt_cmpl *rxcmp;
771 struct rx_pkt_cmpl_hi *rxcmp1;
772 uint32_t tmp_raw_cons = *raw_cons;
773 uint16_t cons, raw_prod, cp_cons =
774 RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
775 struct rte_mbuf *mbuf;
779 uint32_t vfr_flag = 0, mark_id = 0;
780 struct bnxt *bp = rxq->bp;
782 rxcmp = (struct rx_pkt_cmpl *)
783 &cpr->cp_desc_ring[cp_cons];
785 cmp_type = CMP_TYPE(rxcmp);
787 if (cmp_type == RX_TPA_V2_ABUF_CMPL_TYPE_RX_TPA_AGG) {
788 struct rx_tpa_v2_abuf_cmpl *rx_agg = (void *)rxcmp;
789 uint16_t agg_id = rte_cpu_to_le_16(rx_agg->agg_id);
790 struct bnxt_tpa_info *tpa_info;
792 tpa_info = &rxr->tpa_info[agg_id];
793 RTE_ASSERT(tpa_info->agg_count < 16);
794 tpa_info->agg_arr[tpa_info->agg_count++] = *rx_agg;
795 rc = -EINVAL; /* Continue w/o new mbuf */
799 tmp_raw_cons = NEXT_RAW_CMP(tmp_raw_cons);
800 cp_cons = RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
801 rxcmp1 = (struct rx_pkt_cmpl_hi *)&cpr->cp_desc_ring[cp_cons];
803 if (!CMP_VALID(rxcmp1, tmp_raw_cons, cpr->cp_ring_struct))
806 cpr->valid = FLIP_VALID(cp_cons,
807 cpr->cp_ring_struct->ring_mask,
810 if (cmp_type == RX_TPA_START_CMPL_TYPE_RX_TPA_START ||
811 cmp_type == RX_TPA_START_V2_CMPL_TYPE_RX_TPA_START_V2) {
812 bnxt_tpa_start(rxq, (struct rx_tpa_start_cmpl *)rxcmp,
813 (struct rx_tpa_start_cmpl_hi *)rxcmp1);
814 rc = -EINVAL; /* Continue w/o new mbuf */
816 } else if (cmp_type == RX_TPA_END_CMPL_TYPE_RX_TPA_END) {
817 mbuf = bnxt_tpa_end(rxq, &tmp_raw_cons,
818 (struct rx_tpa_end_cmpl *)rxcmp,
819 (struct rx_tpa_end_cmpl_hi *)rxcmp1);
824 } else if ((cmp_type != CMPL_BASE_TYPE_RX_L2) &&
825 (cmp_type != CMPL_BASE_TYPE_RX_L2_V2)) {
830 agg_buf = (rxcmp->agg_bufs_v1 & RX_PKT_CMPL_AGG_BUFS_MASK)
831 >> RX_PKT_CMPL_AGG_BUFS_SFT;
832 if (agg_buf && !bnxt_agg_bufs_valid(cpr, agg_buf, tmp_raw_cons))
835 raw_prod = rxr->rx_raw_prod;
837 cons = rxcmp->opaque;
838 mbuf = bnxt_consume_rx_buf(rxr, cons);
842 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
845 mbuf->pkt_len = rxcmp->len;
846 mbuf->data_len = mbuf->pkt_len;
847 mbuf->port = rxq->port_id;
849 #ifdef RTE_LIBRTE_IEEE1588
850 if (unlikely((rte_le_to_cpu_16(rxcmp->flags_type) &
851 RX_PKT_CMPL_FLAGS_MASK) ==
852 RX_PKT_CMPL_FLAGS_ITYPE_PTP_W_TIMESTAMP))
853 bnxt_get_rx_ts_p5(rxq->bp, rxcmp1->reorder);
856 if (cmp_type == CMPL_BASE_TYPE_RX_L2_V2) {
857 bnxt_parse_csum_v2(mbuf, rxcmp1);
858 bnxt_parse_pkt_type_v2(mbuf, rxcmp, rxcmp1);
859 bnxt_rx_vlan_v2(mbuf, rxcmp, rxcmp1);
860 /* TODO Add support for cfa_code parsing */
864 bnxt_set_ol_flags(rxr, rxcmp, rxcmp1, mbuf);
866 mbuf->packet_type = bnxt_parse_pkt_type(rxcmp, rxcmp1);
868 if (BNXT_TRUFLOW_EN(bp))
869 mark_id = bnxt_ulp_set_mark_in_mbuf(rxq->bp, rxcmp1, mbuf,
872 bnxt_set_mark_in_mbuf(rxq->bp, rxcmp1, mbuf);
876 bnxt_rx_pages(rxq, mbuf, &tmp_raw_cons, agg_buf, NULL);
879 if (rxcmp1->errors_v2 & RX_CMP_L2_ERRORS) {
880 /* Re-install the mbuf back to the rx ring */
881 bnxt_reuse_rx_mbuf(rxr, cons, mbuf);
888 * TODO: Redesign this....
889 * If the allocation fails, the packet does not get received.
890 * Simply returning this will result in slowly falling behind
891 * on the producer ring buffers.
892 * Instead, "filling up" the producer just before ringing the
893 * doorbell could be a better solution since it will let the
894 * producer ring starve until memory is available again pushing
895 * the drops into hardware and getting them out of the driver
896 * allowing recovery to a full producer ring.
898 * This could also help with cache usage by preventing per-packet
899 * calls in favour of a tight loop with the same function being called
902 raw_prod = RING_NEXT(raw_prod);
903 if (bnxt_alloc_rx_data(rxq, rxr, raw_prod)) {
904 PMD_DRV_LOG(ERR, "mbuf alloc failed with prod=0x%x\n",
909 rxr->rx_raw_prod = raw_prod;
911 if (BNXT_TRUFLOW_EN(bp) && (BNXT_VF_IS_TRUSTED(bp) || BNXT_PF(bp)) &&
913 bnxt_vfr_recv(mark_id, rxq->queue_id, mbuf);
914 /* Now return an error so that nb_rx_pkts is not
916 * This packet was meant to be given to the representor.
917 * So no need to account the packet and give it to
918 * parent Rx burst function.
924 * All MBUFs are allocated with the same size under DPDK,
925 * no optimization for rx_copy_thresh
932 *raw_cons = tmp_raw_cons;
937 uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
940 struct bnxt_rx_queue *rxq = rx_queue;
941 struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
942 struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
943 uint16_t rx_raw_prod = rxr->rx_raw_prod;
944 uint16_t ag_raw_prod = rxr->ag_raw_prod;
945 uint32_t raw_cons = cpr->cp_raw_cons;
946 bool alloc_failed = false;
949 int nb_rep_rx_pkts = 0;
950 struct rx_pkt_cmpl *rxcmp;
954 if (unlikely(is_bnxt_in_error(rxq->bp)))
957 /* If Rx Q was stopped return */
958 if (unlikely(!rxq->rx_started))
961 #if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
963 * Replenish buffers if needed when a transition has been made from
964 * vector- to non-vector- receive processing.
966 while (unlikely(rxq->rxrearm_nb)) {
967 if (!bnxt_alloc_rx_data(rxq, rxr, rxq->rxrearm_start)) {
968 rxr->rx_raw_prod = rxq->rxrearm_start;
969 bnxt_db_write(&rxr->rx_db, rxr->rx_raw_prod);
970 rxq->rxrearm_start++;
973 /* Retry allocation on next call. */
979 /* Handle RX burst request */
981 cons = RING_CMP(cpr->cp_ring_struct, raw_cons);
982 rxcmp = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[cons];
984 if (!CMP_VALID(rxcmp, raw_cons, cpr->cp_ring_struct))
986 cpr->valid = FLIP_VALID(cons,
987 cpr->cp_ring_struct->ring_mask,
990 if ((CMP_TYPE(rxcmp) >= CMPL_BASE_TYPE_RX_TPA_START_V2) &&
991 (CMP_TYPE(rxcmp) <= RX_TPA_V2_ABUF_CMPL_TYPE_RX_TPA_AGG)) {
992 rc = bnxt_rx_pkt(&rx_pkts[nb_rx_pkts], rxq, &raw_cons);
995 else if (rc == -EBUSY) /* partial completion */
997 else if (rc == -ENODEV) /* completion for representor */
999 else if (rc == -ENOMEM) {
1001 alloc_failed = true;
1003 } else if (!BNXT_NUM_ASYNC_CPR(rxq->bp)) {
1005 bnxt_event_hwrm_resp_handler(rxq->bp,
1006 (struct cmpl_base *)rxcmp);
1007 /* If the async event is Fatal error, return */
1008 if (unlikely(is_bnxt_in_error(rxq->bp)))
1012 raw_cons = NEXT_RAW_CMP(raw_cons);
1013 if (nb_rx_pkts == nb_pkts || nb_rep_rx_pkts == nb_pkts || evt)
1015 /* Post some Rx buf early in case of larger burst processing */
1016 if (nb_rx_pkts == BNXT_RX_POST_THRESH)
1017 bnxt_db_write(&rxr->rx_db, rxr->rx_raw_prod);
1020 cpr->cp_raw_cons = raw_cons;
1021 if (!nb_rx_pkts && !nb_rep_rx_pkts && !evt) {
1023 * For PMD, there is no need to keep on pushing to REARM
1024 * the doorbell if there are no new completions
1029 /* Ring the completion queue doorbell. */
1032 /* Ring the receive descriptor doorbell. */
1033 if (rx_raw_prod != rxr->rx_raw_prod)
1034 bnxt_db_write(&rxr->rx_db, rxr->rx_raw_prod);
1036 /* Ring the AGG ring DB */
1037 if (ag_raw_prod != rxr->ag_raw_prod)
1038 bnxt_db_write(&rxr->ag_db, rxr->ag_raw_prod);
1040 /* Attempt to alloc Rx buf in case of a previous allocation failure. */
1044 rx_raw_prod = RING_NEXT(rx_raw_prod);
1045 for (cnt = 0; cnt < nb_rx_pkts + nb_rep_rx_pkts; cnt++) {
1046 struct rte_mbuf **rx_buf;
1049 ndx = RING_IDX(rxr->rx_ring_struct, rx_raw_prod + cnt);
1050 rx_buf = &rxr->rx_buf_ring[ndx];
1052 /* Buffer already allocated for this index. */
1053 if (*rx_buf != NULL && *rx_buf != &rxq->fake_mbuf)
1056 /* This slot is empty. Alloc buffer for Rx */
1057 if (!bnxt_alloc_rx_data(rxq, rxr, rx_raw_prod + cnt)) {
1058 rxr->rx_raw_prod = rx_raw_prod + cnt;
1059 bnxt_db_write(&rxr->rx_db, rxr->rx_raw_prod);
1061 PMD_DRV_LOG(ERR, "Alloc mbuf failed\n");
1072 * Dummy DPDK callback for RX.
1074 * This function is used to temporarily replace the real callback during
1075 * unsafe control operations on the queue, or in case of error.
1078 bnxt_dummy_recv_pkts(void *rx_queue __rte_unused,
1079 struct rte_mbuf **rx_pkts __rte_unused,
1080 uint16_t nb_pkts __rte_unused)
1085 void bnxt_free_rx_rings(struct bnxt *bp)
1088 struct bnxt_rx_queue *rxq;
1093 for (i = 0; i < (int)bp->rx_nr_rings; i++) {
1094 rxq = bp->rx_queues[i];
1098 bnxt_free_ring(rxq->rx_ring->rx_ring_struct);
1099 rte_free(rxq->rx_ring->rx_ring_struct);
1101 /* Free the Aggregator ring */
1102 bnxt_free_ring(rxq->rx_ring->ag_ring_struct);
1103 rte_free(rxq->rx_ring->ag_ring_struct);
1104 rxq->rx_ring->ag_ring_struct = NULL;
1106 rte_free(rxq->rx_ring);
1108 bnxt_free_ring(rxq->cp_ring->cp_ring_struct);
1109 rte_free(rxq->cp_ring->cp_ring_struct);
1110 rte_free(rxq->cp_ring);
1113 bp->rx_queues[i] = NULL;
1117 int bnxt_init_rx_ring_struct(struct bnxt_rx_queue *rxq, unsigned int socket_id)
1119 struct bnxt_cp_ring_info *cpr;
1120 struct bnxt_rx_ring_info *rxr;
1121 struct bnxt_ring *ring;
1123 rxq->rx_buf_size = BNXT_MAX_PKT_LEN + sizeof(struct rte_mbuf);
1125 rxr = rte_zmalloc_socket("bnxt_rx_ring",
1126 sizeof(struct bnxt_rx_ring_info),
1127 RTE_CACHE_LINE_SIZE, socket_id);
1132 ring = rte_zmalloc_socket("bnxt_rx_ring_struct",
1133 sizeof(struct bnxt_ring),
1134 RTE_CACHE_LINE_SIZE, socket_id);
1137 rxr->rx_ring_struct = ring;
1138 ring->ring_size = rte_align32pow2(rxq->nb_rx_desc);
1139 ring->ring_mask = ring->ring_size - 1;
1140 ring->bd = (void *)rxr->rx_desc_ring;
1141 ring->bd_dma = rxr->rx_desc_mapping;
1143 /* Allocate extra rx ring entries for vector rx. */
1144 ring->vmem_size = sizeof(struct rte_mbuf *) *
1145 (ring->ring_size + RTE_BNXT_DESCS_PER_LOOP);
1147 ring->vmem = (void **)&rxr->rx_buf_ring;
1148 ring->fw_ring_id = INVALID_HW_RING_ID;
1150 cpr = rte_zmalloc_socket("bnxt_rx_ring",
1151 sizeof(struct bnxt_cp_ring_info),
1152 RTE_CACHE_LINE_SIZE, socket_id);
1157 ring = rte_zmalloc_socket("bnxt_rx_ring_struct",
1158 sizeof(struct bnxt_ring),
1159 RTE_CACHE_LINE_SIZE, socket_id);
1162 cpr->cp_ring_struct = ring;
1164 /* Allocate two completion slots per entry in desc ring. */
1165 ring->ring_size = rxr->rx_ring_struct->ring_size * 2;
1166 ring->ring_size *= AGG_RING_SIZE_FACTOR;
1168 ring->ring_size = rte_align32pow2(ring->ring_size);
1169 ring->ring_mask = ring->ring_size - 1;
1170 ring->bd = (void *)cpr->cp_desc_ring;
1171 ring->bd_dma = cpr->cp_desc_mapping;
1172 ring->vmem_size = 0;
1174 ring->fw_ring_id = INVALID_HW_RING_ID;
1176 /* Allocate Aggregator rings */
1177 ring = rte_zmalloc_socket("bnxt_rx_ring_struct",
1178 sizeof(struct bnxt_ring),
1179 RTE_CACHE_LINE_SIZE, socket_id);
1182 rxr->ag_ring_struct = ring;
1183 ring->ring_size = rte_align32pow2(rxq->nb_rx_desc *
1184 AGG_RING_SIZE_FACTOR);
1185 ring->ring_mask = ring->ring_size - 1;
1186 ring->bd = (void *)rxr->ag_desc_ring;
1187 ring->bd_dma = rxr->ag_desc_mapping;
1188 ring->vmem_size = ring->ring_size * sizeof(struct rte_mbuf *);
1189 ring->vmem = (void **)&rxr->ag_buf_ring;
1190 ring->fw_ring_id = INVALID_HW_RING_ID;
1195 static void bnxt_init_rxbds(struct bnxt_ring *ring, uint32_t type,
1199 struct rx_prod_pkt_bd *rx_bd_ring = (struct rx_prod_pkt_bd *)ring->bd;
1203 for (j = 0; j < ring->ring_size; j++) {
1204 rx_bd_ring[j].flags_type = rte_cpu_to_le_16(type);
1205 rx_bd_ring[j].len = rte_cpu_to_le_16(len);
1206 rx_bd_ring[j].opaque = j;
1210 int bnxt_init_one_rx_ring(struct bnxt_rx_queue *rxq)
1212 struct bnxt_rx_ring_info *rxr;
1213 struct bnxt_ring *ring;
1214 uint32_t raw_prod, type;
1218 /* Initialize packet type table. */
1219 bnxt_init_ptype_table();
1221 size = rte_pktmbuf_data_room_size(rxq->mb_pool) - RTE_PKTMBUF_HEADROOM;
1222 size = RTE_MIN(BNXT_MAX_PKT_LEN, size);
1224 type = RX_PROD_PKT_BD_TYPE_RX_PROD_PKT;
1227 ring = rxr->rx_ring_struct;
1228 bnxt_init_rxbds(ring, type, size);
1230 /* Initialize offload flags parsing table. */
1231 bnxt_init_ol_flags_tables(rxq);
1233 raw_prod = rxr->rx_raw_prod;
1234 for (i = 0; i < ring->ring_size; i++) {
1235 if (unlikely(!rxr->rx_buf_ring[i])) {
1236 if (bnxt_alloc_rx_data(rxq, rxr, raw_prod) != 0) {
1237 PMD_DRV_LOG(WARNING,
1238 "init'ed rx ring %d with %d/%d mbufs only\n",
1239 rxq->queue_id, i, ring->ring_size);
1243 rxr->rx_raw_prod = raw_prod;
1244 raw_prod = RING_NEXT(raw_prod);
1247 /* Initialize dummy mbuf pointers for vector mode rx. */
1248 for (i = ring->ring_size;
1249 i < ring->ring_size + RTE_BNXT_DESCS_PER_LOOP; i++) {
1250 rxr->rx_buf_ring[i] = &rxq->fake_mbuf;
1253 ring = rxr->ag_ring_struct;
1254 type = RX_PROD_AGG_BD_TYPE_RX_PROD_AGG;
1255 bnxt_init_rxbds(ring, type, size);
1256 raw_prod = rxr->ag_raw_prod;
1258 for (i = 0; i < ring->ring_size; i++) {
1259 if (unlikely(!rxr->ag_buf_ring[i])) {
1260 if (bnxt_alloc_ag_data(rxq, rxr, raw_prod) != 0) {
1261 PMD_DRV_LOG(WARNING,
1262 "init'ed AG ring %d with %d/%d mbufs only\n",
1263 rxq->queue_id, i, ring->ring_size);
1267 rxr->ag_raw_prod = raw_prod;
1268 raw_prod = RING_NEXT(raw_prod);
1270 PMD_DRV_LOG(DEBUG, "AGG Done!\n");
1272 if (rxr->tpa_info) {
1273 unsigned int max_aggs = BNXT_TPA_MAX_AGGS(rxq->bp);
1275 for (i = 0; i < max_aggs; i++) {
1276 if (unlikely(!rxr->tpa_info[i].mbuf)) {
1277 rxr->tpa_info[i].mbuf =
1278 __bnxt_alloc_rx_data(rxq->mb_pool);
1279 if (!rxr->tpa_info[i].mbuf) {
1280 rte_atomic64_inc(&rxq->rx_mbuf_alloc_fail);
1286 PMD_DRV_LOG(DEBUG, "TPA alloc Done!\n");