1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2014-2021 Broadcom
8 #include <rte_byteorder.h>
9 #include <rte_malloc.h>
12 #include "bnxt_ring.h"
15 #include "hsi_struct_def_dpdk.h"
22 void bnxt_free_tx_rings(struct bnxt *bp)
26 for (i = 0; i < (int)bp->tx_nr_rings; i++) {
27 struct bnxt_tx_queue *txq = bp->tx_queues[i];
32 bnxt_free_ring(txq->tx_ring->tx_ring_struct);
33 rte_free(txq->tx_ring->tx_ring_struct);
34 rte_free(txq->tx_ring);
36 bnxt_free_ring(txq->cp_ring->cp_ring_struct);
37 rte_free(txq->cp_ring->cp_ring_struct);
38 rte_free(txq->cp_ring);
41 bp->tx_queues[i] = NULL;
45 int bnxt_init_one_tx_ring(struct bnxt_tx_queue *txq)
47 struct bnxt_tx_ring_info *txr = txq->tx_ring;
48 struct bnxt_ring *ring = txr->tx_ring_struct;
50 txq->tx_wake_thresh = ring->ring_size / 2;
51 ring->fw_ring_id = INVALID_HW_RING_ID;
56 int bnxt_init_tx_ring_struct(struct bnxt_tx_queue *txq, unsigned int socket_id)
58 struct bnxt_cp_ring_info *cpr;
59 struct bnxt_tx_ring_info *txr;
60 struct bnxt_ring *ring;
62 txr = rte_zmalloc_socket("bnxt_tx_ring",
63 sizeof(struct bnxt_tx_ring_info),
64 RTE_CACHE_LINE_SIZE, socket_id);
69 ring = rte_zmalloc_socket("bnxt_tx_ring_struct",
70 sizeof(struct bnxt_ring),
71 RTE_CACHE_LINE_SIZE, socket_id);
74 txr->tx_ring_struct = ring;
75 ring->ring_size = rte_align32pow2(txq->nb_tx_desc);
76 ring->ring_mask = ring->ring_size - 1;
77 ring->bd = (void *)txr->tx_desc_ring;
78 ring->bd_dma = txr->tx_desc_mapping;
79 ring->vmem_size = ring->ring_size * sizeof(struct bnxt_sw_tx_bd);
80 ring->vmem = (void **)&txr->tx_buf_ring;
81 ring->fw_ring_id = INVALID_HW_RING_ID;
83 cpr = rte_zmalloc_socket("bnxt_tx_ring",
84 sizeof(struct bnxt_cp_ring_info),
85 RTE_CACHE_LINE_SIZE, socket_id);
90 ring = rte_zmalloc_socket("bnxt_tx_ring_struct",
91 sizeof(struct bnxt_ring),
92 RTE_CACHE_LINE_SIZE, socket_id);
95 cpr->cp_ring_struct = ring;
96 ring->ring_size = txr->tx_ring_struct->ring_size;
97 ring->ring_mask = ring->ring_size - 1;
98 ring->bd = (void *)cpr->cp_desc_ring;
99 ring->bd_dma = cpr->cp_desc_mapping;
102 ring->fw_ring_id = INVALID_HW_RING_ID;
107 static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
108 struct bnxt_tx_queue *txq,
110 struct tx_bd_long **last_txbd)
112 struct bnxt_tx_ring_info *txr = txq->tx_ring;
113 struct bnxt_ring *ring = txr->tx_ring_struct;
114 uint32_t outer_tpid_bd = 0;
115 struct tx_bd_long *txbd;
116 struct tx_bd_long_hi *txbd1 = NULL;
117 uint32_t vlan_tag_flags;
118 bool long_bd = false;
119 unsigned short nr_bds = 0;
121 struct rte_mbuf *m_seg;
122 struct bnxt_sw_tx_bd *tx_buf;
123 static const uint32_t lhint_arr[4] = {
124 TX_BD_LONG_FLAGS_LHINT_LT512,
125 TX_BD_LONG_FLAGS_LHINT_LT1K,
126 TX_BD_LONG_FLAGS_LHINT_LT2K,
127 TX_BD_LONG_FLAGS_LHINT_LT2K
130 if (unlikely(is_bnxt_in_error(txq->bp)))
133 if (tx_pkt->ol_flags & (PKT_TX_TCP_SEG | PKT_TX_TCP_CKSUM |
134 PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM |
135 PKT_TX_VLAN_PKT | PKT_TX_OUTER_IP_CKSUM |
136 PKT_TX_TUNNEL_GRE | PKT_TX_TUNNEL_VXLAN |
137 PKT_TX_TUNNEL_GENEVE | PKT_TX_IEEE1588_TMST |
139 (BNXT_TRUFLOW_EN(txq->bp) &&
140 (txq->bp->tx_cfa_action || txq->vfr_tx_cfa_action)))
143 nr_bds = long_bd + tx_pkt->nb_segs;
144 if (unlikely(bnxt_tx_avail(txq) < nr_bds))
147 /* Check if number of Tx descriptors is above HW limit */
148 if (unlikely(nr_bds > BNXT_MAX_TSO_SEGS)) {
150 "Num descriptors %d exceeds HW limit\n", nr_bds);
154 /* If packet length is less than minimum packet size, pad it */
155 if (unlikely(rte_pktmbuf_pkt_len(tx_pkt) < BNXT_MIN_PKT_SIZE)) {
156 uint8_t pad = BNXT_MIN_PKT_SIZE - rte_pktmbuf_pkt_len(tx_pkt);
157 char *seg = rte_pktmbuf_append(tx_pkt, pad);
161 "Failed to pad mbuf by %d bytes\n",
166 /* Note: data_len, pkt len are updated in rte_pktmbuf_append */
170 /* Check non zero data_len */
171 RTE_VERIFY(tx_pkt->data_len);
173 prod = RING_IDX(ring, txr->tx_raw_prod);
174 tx_buf = &txr->tx_buf_ring[prod];
175 tx_buf->mbuf = tx_pkt;
176 tx_buf->nr_bds = nr_bds;
178 txbd = &txr->tx_desc_ring[prod];
179 txbd->opaque = *coal_pkts;
180 txbd->flags_type = nr_bds << TX_BD_LONG_FLAGS_BD_CNT_SFT;
181 txbd->flags_type |= TX_BD_SHORT_FLAGS_COAL_NOW;
182 txbd->flags_type |= TX_BD_LONG_FLAGS_NO_CMPL;
183 txbd->len = tx_pkt->data_len;
184 if (tx_pkt->pkt_len >= 2014)
185 txbd->flags_type |= TX_BD_LONG_FLAGS_LHINT_GTE2K;
187 txbd->flags_type |= lhint_arr[tx_pkt->pkt_len >> 9];
188 txbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova(tx_buf->mbuf));
192 txbd->flags_type |= TX_BD_LONG_TYPE_TX_BD_LONG;
195 /* HW can accelerate only outer vlan in QinQ mode */
196 if (tx_buf->mbuf->ol_flags & PKT_TX_QINQ_PKT) {
197 vlan_tag_flags = TX_BD_LONG_CFA_META_KEY_VLAN_TAG |
198 tx_buf->mbuf->vlan_tci_outer;
199 outer_tpid_bd = txq->bp->outer_tpid_bd &
200 BNXT_OUTER_TPID_BD_MASK;
201 vlan_tag_flags |= outer_tpid_bd;
202 } else if (tx_buf->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
203 /* shurd: Should this mask at
204 * TX_BD_LONG_CFA_META_VLAN_VID_MASK?
206 vlan_tag_flags = TX_BD_LONG_CFA_META_KEY_VLAN_TAG |
207 tx_buf->mbuf->vlan_tci;
208 /* Currently supports 8021Q, 8021AD vlan offloads
209 * QINQ1, QINQ2, QINQ3 vlan headers are deprecated
211 /* DPDK only supports 802.11q VLAN packets */
213 TX_BD_LONG_CFA_META_VLAN_TPID_TPID8100;
216 txr->tx_raw_prod = RING_NEXT(txr->tx_raw_prod);
218 prod = RING_IDX(ring, txr->tx_raw_prod);
219 txbd1 = (struct tx_bd_long_hi *)&txr->tx_desc_ring[prod];
221 txbd1->cfa_meta = vlan_tag_flags;
222 /* Legacy tx_bd_long_hi->mss =
223 * tx_bd_long_hi->kid_or_ts_high_mss
225 txbd1->kid_or_ts_high_mss = 0;
227 if (txq->vfr_tx_cfa_action)
228 txbd1->cfa_action = txq->vfr_tx_cfa_action;
230 txbd1->cfa_action = txq->bp->tx_cfa_action;
232 if (tx_pkt->ol_flags & PKT_TX_TCP_SEG) {
236 txbd1->lflags |= TX_BD_LONG_LFLAGS_LSO |
237 TX_BD_LONG_LFLAGS_T_IPID;
238 hdr_size = tx_pkt->l2_len + tx_pkt->l3_len +
240 hdr_size += (tx_pkt->ol_flags & PKT_TX_TUNNEL_MASK) ?
241 tx_pkt->outer_l2_len +
242 tx_pkt->outer_l3_len : 0;
243 /* The hdr_size is multiple of 16bit units not 8bit.
245 * Also legacy hdr_size = kid_or_ts_low_hdr_size.
247 txbd1->kid_or_ts_low_hdr_size = hdr_size >> 1;
248 txbd1->kid_or_ts_high_mss = tx_pkt->tso_segsz;
249 RTE_VERIFY(txbd1->kid_or_ts_high_mss);
251 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_TCP_UDP_CKSUM) ==
252 PKT_TX_OIP_IIP_TCP_UDP_CKSUM) {
253 /* Outer IP, Inner IP, Inner TCP/UDP CSO */
254 txbd1->lflags |= TX_BD_FLG_TIP_IP_TCP_UDP_CHKSUM;
255 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_TCP_CKSUM) ==
256 PKT_TX_OIP_IIP_TCP_CKSUM) {
257 /* Outer IP, Inner IP, Inner TCP/UDP CSO */
258 txbd1->lflags |= TX_BD_FLG_TIP_IP_TCP_UDP_CHKSUM;
259 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_UDP_CKSUM) ==
260 PKT_TX_OIP_IIP_UDP_CKSUM) {
261 /* Outer IP, Inner IP, Inner TCP/UDP CSO */
262 txbd1->lflags |= TX_BD_FLG_TIP_IP_TCP_UDP_CHKSUM;
263 } else if ((tx_pkt->ol_flags & PKT_TX_IIP_TCP_UDP_CKSUM) ==
264 PKT_TX_IIP_TCP_UDP_CKSUM) {
265 /* (Inner) IP, (Inner) TCP/UDP CSO */
266 txbd1->lflags |= TX_BD_FLG_IP_TCP_UDP_CHKSUM;
267 } else if ((tx_pkt->ol_flags & PKT_TX_IIP_UDP_CKSUM) ==
268 PKT_TX_IIP_UDP_CKSUM) {
269 /* (Inner) IP, (Inner) TCP/UDP CSO */
270 txbd1->lflags |= TX_BD_FLG_IP_TCP_UDP_CHKSUM;
271 } else if ((tx_pkt->ol_flags & PKT_TX_IIP_TCP_CKSUM) ==
272 PKT_TX_IIP_TCP_CKSUM) {
273 /* (Inner) IP, (Inner) TCP/UDP CSO */
274 txbd1->lflags |= TX_BD_FLG_IP_TCP_UDP_CHKSUM;
275 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_TCP_UDP_CKSUM) ==
276 PKT_TX_OIP_TCP_UDP_CKSUM) {
277 /* Outer IP, (Inner) TCP/UDP CSO */
278 txbd1->lflags |= TX_BD_FLG_TIP_TCP_UDP_CHKSUM;
279 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_UDP_CKSUM) ==
280 PKT_TX_OIP_UDP_CKSUM) {
281 /* Outer IP, (Inner) TCP/UDP CSO */
282 txbd1->lflags |= TX_BD_FLG_TIP_TCP_UDP_CHKSUM;
283 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_TCP_CKSUM) ==
284 PKT_TX_OIP_TCP_CKSUM) {
285 /* Outer IP, (Inner) TCP/UDP CSO */
286 txbd1->lflags |= TX_BD_FLG_TIP_TCP_UDP_CHKSUM;
287 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_CKSUM) ==
288 PKT_TX_OIP_IIP_CKSUM) {
289 /* Outer IP, Inner IP CSO */
290 txbd1->lflags |= TX_BD_FLG_TIP_IP_CHKSUM;
291 } else if ((tx_pkt->ol_flags & PKT_TX_TCP_UDP_CKSUM) ==
292 PKT_TX_TCP_UDP_CKSUM) {
294 txbd1->lflags |= TX_BD_LONG_LFLAGS_TCP_UDP_CHKSUM;
295 } else if ((tx_pkt->ol_flags & PKT_TX_TCP_CKSUM) ==
298 txbd1->lflags |= TX_BD_LONG_LFLAGS_TCP_UDP_CHKSUM;
299 } else if ((tx_pkt->ol_flags & PKT_TX_UDP_CKSUM) ==
302 txbd1->lflags |= TX_BD_LONG_LFLAGS_TCP_UDP_CHKSUM;
303 } else if ((tx_pkt->ol_flags & PKT_TX_IP_CKSUM) ==
306 txbd1->lflags |= TX_BD_LONG_LFLAGS_IP_CHKSUM;
307 } else if ((tx_pkt->ol_flags & PKT_TX_OUTER_IP_CKSUM) ==
308 PKT_TX_OUTER_IP_CKSUM) {
310 txbd1->lflags |= TX_BD_LONG_LFLAGS_T_IP_CHKSUM;
311 } else if ((tx_pkt->ol_flags & PKT_TX_IEEE1588_TMST) ==
312 PKT_TX_IEEE1588_TMST) {
314 txbd1->lflags |= TX_BD_LONG_LFLAGS_STAMP;
317 txbd->flags_type |= TX_BD_SHORT_TYPE_TX_BD_SHORT;
320 m_seg = tx_pkt->next;
322 /* Check non zero data_len */
323 RTE_VERIFY(m_seg->data_len);
324 txr->tx_raw_prod = RING_NEXT(txr->tx_raw_prod);
326 prod = RING_IDX(ring, txr->tx_raw_prod);
327 tx_buf = &txr->tx_buf_ring[prod];
328 tx_buf->mbuf = m_seg;
330 txbd = &txr->tx_desc_ring[prod];
331 txbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova(m_seg));
332 txbd->flags_type = TX_BD_SHORT_TYPE_TX_BD_SHORT;
333 txbd->len = m_seg->data_len;
338 txbd->flags_type |= TX_BD_LONG_FLAGS_PACKET_END;
340 txr->tx_raw_prod = RING_NEXT(txr->tx_raw_prod);
346 * Transmit completion function for use when DEV_TX_OFFLOAD_MBUF_FAST_FREE
349 static void bnxt_tx_cmp_fast(struct bnxt_tx_queue *txq, int nr_pkts)
351 struct bnxt_tx_ring_info *txr = txq->tx_ring;
352 struct bnxt_ring *ring = txr->tx_ring_struct;
353 struct rte_mbuf **free = txq->free;
354 uint16_t raw_cons = txr->tx_raw_cons;
355 unsigned int blk = 0;
358 for (i = 0; i < nr_pkts; i++) {
359 struct bnxt_sw_tx_bd *tx_buf;
360 unsigned short nr_bds;
362 tx_buf = &txr->tx_buf_ring[RING_IDX(ring, raw_cons)];
363 nr_bds = tx_buf->nr_bds;
364 for (j = 0; j < nr_bds; j++) {
366 /* Add mbuf to the bulk free array */
367 free[blk++] = tx_buf->mbuf;
370 raw_cons = RING_NEXT(raw_cons);
371 tx_buf = &txr->tx_buf_ring[RING_IDX(ring, raw_cons)];
375 rte_mempool_put_bulk(free[0]->pool, (void *)free, blk);
377 txr->tx_raw_cons = raw_cons;
380 static void bnxt_tx_cmp(struct bnxt_tx_queue *txq, int nr_pkts)
382 struct bnxt_tx_ring_info *txr = txq->tx_ring;
383 struct bnxt_ring *ring = txr->tx_ring_struct;
384 struct rte_mempool *pool = NULL;
385 struct rte_mbuf **free = txq->free;
386 uint16_t raw_cons = txr->tx_raw_cons;
387 unsigned int blk = 0;
390 for (i = 0; i < nr_pkts; i++) {
391 struct rte_mbuf *mbuf;
392 struct bnxt_sw_tx_bd *tx_buf;
393 unsigned short nr_bds;
395 tx_buf = &txr->tx_buf_ring[RING_IDX(ring, raw_cons)];
396 nr_bds = tx_buf->nr_bds;
397 for (j = 0; j < nr_bds; j++) {
400 raw_cons = RING_NEXT(raw_cons);
401 tx_buf = &txr->tx_buf_ring[RING_IDX(ring, raw_cons)];
402 if (!mbuf) /* long_bd's tx_buf ? */
405 mbuf = rte_pktmbuf_prefree_seg(mbuf);
409 /* EW - no need to unmap DMA memory? */
411 if (likely(mbuf->pool == pool)) {
412 /* Add mbuf to the bulk free array */
415 /* Found an mbuf from a different pool. Free
416 * mbufs accumulated so far to the previous
419 if (likely(pool != NULL))
420 rte_mempool_put_bulk(pool,
424 /* Start accumulating mbufs in a new pool */
432 rte_mempool_put_bulk(pool, (void *)free, blk);
434 txr->tx_raw_cons = raw_cons;
437 static int bnxt_handle_tx_cp(struct bnxt_tx_queue *txq)
439 struct bnxt_cp_ring_info *cpr = txq->cp_ring;
440 uint32_t raw_cons = cpr->cp_raw_cons;
442 uint32_t nb_tx_pkts = 0;
443 struct tx_cmpl *txcmp;
444 struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
445 struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
446 uint32_t ring_mask = cp_ring_struct->ring_mask;
449 if (bnxt_tx_bds_in_hw(txq) < txq->tx_free_thresh)
453 cons = RING_CMPL(ring_mask, raw_cons);
454 txcmp = (struct tx_cmpl *)&cpr->cp_desc_ring[cons];
455 rte_prefetch_non_temporal(&cp_desc_ring[(cons + 2) &
458 if (!CMPL_VALID(txcmp, cpr->valid))
460 opaque = rte_cpu_to_le_32(txcmp->opaque);
461 NEXT_CMPL(cpr, cons, cpr->valid, 1);
462 rte_prefetch0(&cp_desc_ring[cons]);
464 if (CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2)
465 nb_tx_pkts += opaque;
468 "Unhandled CMP type %02x\n",
471 } while (nb_tx_pkts < ring_mask);
474 if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
475 bnxt_tx_cmp_fast(txq, nb_tx_pkts);
477 bnxt_tx_cmp(txq, nb_tx_pkts);
478 cpr->cp_raw_cons = raw_cons;
485 uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
489 uint16_t nb_tx_pkts = 0;
490 uint16_t coal_pkts = 0;
491 struct bnxt_tx_queue *txq = tx_queue;
492 struct tx_bd_long *last_txbd = NULL;
494 /* Handle TX completions */
495 bnxt_handle_tx_cp(txq);
497 /* Tx queue was stopped; wait for it to be restarted */
498 if (unlikely(!txq->tx_started)) {
499 PMD_DRV_LOG(DEBUG, "Tx q stopped;return\n");
503 /* Handle TX burst request */
504 for (nb_tx_pkts = 0; nb_tx_pkts < nb_pkts; nb_tx_pkts++) {
506 rc = bnxt_start_xmit(tx_pkts[nb_tx_pkts], txq,
507 &coal_pkts, &last_txbd);
513 if (likely(nb_tx_pkts)) {
514 /* Request a completion on the last packet */
515 last_txbd->flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
516 bnxt_db_write(&txq->tx_ring->tx_db, txq->tx_ring->tx_raw_prod);
523 * Dummy DPDK callback for TX.
525 * This function is used to temporarily replace the real callback during
526 * unsafe control operations on the queue, or in case of error.
529 bnxt_dummy_xmit_pkts(void *tx_queue __rte_unused,
530 struct rte_mbuf **tx_pkts __rte_unused,
531 uint16_t nb_pkts __rte_unused)
536 int bnxt_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
538 struct bnxt *bp = dev->data->dev_private;
539 struct bnxt_tx_queue *txq = bp->tx_queues[tx_queue_id];
542 rc = is_bnxt_in_error(bp);
546 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
547 txq->tx_started = true;
548 PMD_DRV_LOG(DEBUG, "Tx queue started\n");
553 int bnxt_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
555 struct bnxt *bp = dev->data->dev_private;
556 struct bnxt_tx_queue *txq = bp->tx_queues[tx_queue_id];
559 rc = is_bnxt_in_error(bp);
563 /* Handle TX completions */
564 bnxt_handle_tx_cp(txq);
566 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
567 txq->tx_started = false;
568 PMD_DRV_LOG(DEBUG, "Tx queue stopped\n");
573 /* Sweep the Tx completion queue till HWRM_DONE for ring flush is received.
574 * The mbufs will not be freed in this call.
575 * They will be freed during ring free as a part of mem cleanup.
577 int bnxt_flush_tx_cmp(struct bnxt_cp_ring_info *cpr)
579 uint32_t raw_cons = cpr->cp_raw_cons;
581 uint32_t nb_tx_pkts = 0;
582 struct tx_cmpl *txcmp;
583 struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
584 struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
585 uint32_t ring_mask = cp_ring_struct->ring_mask;
589 cons = RING_CMPL(ring_mask, raw_cons);
590 txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
592 opaque = rte_cpu_to_le_32(txcmp->opaque);
593 raw_cons = NEXT_RAW_CMP(raw_cons);
595 if (CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2)
596 nb_tx_pkts += opaque;
597 else if (CMP_TYPE(txcmp) == HWRM_CMPL_TYPE_HWRM_DONE)
599 } while (nb_tx_pkts < ring_mask);
602 cpr->cp_raw_cons = raw_cons;