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 rte_mbuf *);
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;
108 bnxt_xmit_need_long_bd(struct rte_mbuf *tx_pkt, struct bnxt_tx_queue *txq)
110 if (tx_pkt->ol_flags & (PKT_TX_TCP_SEG | PKT_TX_TCP_CKSUM |
111 PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM |
112 PKT_TX_VLAN_PKT | PKT_TX_OUTER_IP_CKSUM |
113 PKT_TX_TUNNEL_GRE | PKT_TX_TUNNEL_VXLAN |
114 PKT_TX_TUNNEL_GENEVE | PKT_TX_IEEE1588_TMST |
116 (BNXT_TRUFLOW_EN(txq->bp) &&
117 (txq->bp->tx_cfa_action || txq->vfr_tx_cfa_action)))
122 static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
123 struct bnxt_tx_queue *txq,
125 struct tx_bd_long **last_txbd)
127 struct bnxt_tx_ring_info *txr = txq->tx_ring;
128 struct bnxt_ring *ring = txr->tx_ring_struct;
129 uint32_t outer_tpid_bd = 0;
130 struct tx_bd_long *txbd;
131 struct tx_bd_long_hi *txbd1 = NULL;
132 uint32_t vlan_tag_flags;
133 bool long_bd = false;
134 unsigned short nr_bds;
136 struct rte_mbuf *m_seg;
137 struct rte_mbuf **tx_buf;
138 static const uint32_t lhint_arr[4] = {
139 TX_BD_LONG_FLAGS_LHINT_LT512,
140 TX_BD_LONG_FLAGS_LHINT_LT1K,
141 TX_BD_LONG_FLAGS_LHINT_LT2K,
142 TX_BD_LONG_FLAGS_LHINT_LT2K
145 if (unlikely(is_bnxt_in_error(txq->bp)))
148 long_bd = bnxt_xmit_need_long_bd(tx_pkt, txq);
149 nr_bds = long_bd + tx_pkt->nb_segs;
151 if (unlikely(bnxt_tx_avail(txq) < nr_bds))
154 /* Check if number of Tx descriptors is above HW limit */
155 if (unlikely(nr_bds > BNXT_MAX_TSO_SEGS)) {
157 "Num descriptors %d exceeds HW limit\n", nr_bds);
161 /* If packet length is less than minimum packet size, pad it */
162 if (unlikely(rte_pktmbuf_pkt_len(tx_pkt) < BNXT_MIN_PKT_SIZE)) {
163 uint8_t pad = BNXT_MIN_PKT_SIZE - rte_pktmbuf_pkt_len(tx_pkt);
164 char *seg = rte_pktmbuf_append(tx_pkt, pad);
168 "Failed to pad mbuf by %d bytes\n",
173 /* Note: data_len, pkt len are updated in rte_pktmbuf_append */
177 /* Check non zero data_len */
178 RTE_VERIFY(tx_pkt->data_len);
180 prod = RING_IDX(ring, txr->tx_raw_prod);
181 tx_buf = &txr->tx_buf_ring[prod];
184 txbd = &txr->tx_desc_ring[prod];
185 txbd->opaque = *coal_pkts;
186 txbd->flags_type = nr_bds << TX_BD_LONG_FLAGS_BD_CNT_SFT;
187 txbd->flags_type |= TX_BD_SHORT_FLAGS_COAL_NOW;
188 txbd->flags_type |= TX_BD_LONG_FLAGS_NO_CMPL;
189 txbd->len = tx_pkt->data_len;
190 if (tx_pkt->pkt_len >= 2048)
191 txbd->flags_type |= TX_BD_LONG_FLAGS_LHINT_GTE2K;
193 txbd->flags_type |= lhint_arr[tx_pkt->pkt_len >> 9];
194 txbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova(tx_pkt));
198 txbd->flags_type |= TX_BD_LONG_TYPE_TX_BD_LONG;
201 /* HW can accelerate only outer vlan in QinQ mode */
202 if (tx_pkt->ol_flags & PKT_TX_QINQ_PKT) {
203 vlan_tag_flags = TX_BD_LONG_CFA_META_KEY_VLAN_TAG |
204 tx_pkt->vlan_tci_outer;
205 outer_tpid_bd = txq->bp->outer_tpid_bd &
206 BNXT_OUTER_TPID_BD_MASK;
207 vlan_tag_flags |= outer_tpid_bd;
208 } else if (tx_pkt->ol_flags & PKT_TX_VLAN_PKT) {
209 /* shurd: Should this mask at
210 * TX_BD_LONG_CFA_META_VLAN_VID_MASK?
212 vlan_tag_flags = TX_BD_LONG_CFA_META_KEY_VLAN_TAG |
214 /* Currently supports 8021Q, 8021AD vlan offloads
215 * QINQ1, QINQ2, QINQ3 vlan headers are deprecated
217 /* DPDK only supports 802.11q VLAN packets */
219 TX_BD_LONG_CFA_META_VLAN_TPID_TPID8100;
222 txr->tx_raw_prod = RING_NEXT(txr->tx_raw_prod);
224 prod = RING_IDX(ring, txr->tx_raw_prod);
225 txbd1 = (struct tx_bd_long_hi *)&txr->tx_desc_ring[prod];
227 txbd1->cfa_meta = vlan_tag_flags;
228 /* Legacy tx_bd_long_hi->mss =
229 * tx_bd_long_hi->kid_or_ts_high_mss
231 txbd1->kid_or_ts_high_mss = 0;
233 if (txq->vfr_tx_cfa_action)
234 txbd1->cfa_action = txq->vfr_tx_cfa_action;
236 txbd1->cfa_action = txq->bp->tx_cfa_action;
238 if (tx_pkt->ol_flags & PKT_TX_TCP_SEG) {
242 txbd1->lflags |= TX_BD_LONG_LFLAGS_LSO |
243 TX_BD_LONG_LFLAGS_T_IPID;
244 hdr_size = tx_pkt->l2_len + tx_pkt->l3_len +
246 hdr_size += (tx_pkt->ol_flags & PKT_TX_TUNNEL_MASK) ?
247 tx_pkt->outer_l2_len +
248 tx_pkt->outer_l3_len : 0;
249 /* The hdr_size is multiple of 16bit units not 8bit.
251 * Also legacy hdr_size = kid_or_ts_low_hdr_size.
253 txbd1->kid_or_ts_low_hdr_size = hdr_size >> 1;
254 txbd1->kid_or_ts_high_mss = tx_pkt->tso_segsz;
255 RTE_VERIFY(txbd1->kid_or_ts_high_mss);
257 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_TCP_UDP_CKSUM) ==
258 PKT_TX_OIP_IIP_TCP_UDP_CKSUM) {
259 /* Outer IP, Inner IP, Inner TCP/UDP CSO */
260 txbd1->lflags |= TX_BD_FLG_TIP_IP_TCP_UDP_CHKSUM;
261 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_TCP_CKSUM) ==
262 PKT_TX_OIP_IIP_TCP_CKSUM) {
263 /* Outer IP, Inner IP, Inner TCP/UDP CSO */
264 txbd1->lflags |= TX_BD_FLG_TIP_IP_TCP_UDP_CHKSUM;
265 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_UDP_CKSUM) ==
266 PKT_TX_OIP_IIP_UDP_CKSUM) {
267 /* Outer IP, Inner IP, Inner TCP/UDP CSO */
268 txbd1->lflags |= TX_BD_FLG_TIP_IP_TCP_UDP_CHKSUM;
269 } else if ((tx_pkt->ol_flags & PKT_TX_IIP_TCP_UDP_CKSUM) ==
270 PKT_TX_IIP_TCP_UDP_CKSUM) {
271 /* (Inner) IP, (Inner) TCP/UDP CSO */
272 txbd1->lflags |= TX_BD_FLG_IP_TCP_UDP_CHKSUM;
273 } else if ((tx_pkt->ol_flags & PKT_TX_IIP_UDP_CKSUM) ==
274 PKT_TX_IIP_UDP_CKSUM) {
275 /* (Inner) IP, (Inner) TCP/UDP CSO */
276 txbd1->lflags |= TX_BD_FLG_IP_TCP_UDP_CHKSUM;
277 } else if ((tx_pkt->ol_flags & PKT_TX_IIP_TCP_CKSUM) ==
278 PKT_TX_IIP_TCP_CKSUM) {
279 /* (Inner) IP, (Inner) TCP/UDP CSO */
280 txbd1->lflags |= TX_BD_FLG_IP_TCP_UDP_CHKSUM;
281 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_TCP_UDP_CKSUM) ==
282 PKT_TX_OIP_TCP_UDP_CKSUM) {
283 /* Outer IP, (Inner) TCP/UDP CSO */
284 txbd1->lflags |= TX_BD_FLG_TIP_TCP_UDP_CHKSUM;
285 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_UDP_CKSUM) ==
286 PKT_TX_OIP_UDP_CKSUM) {
287 /* Outer IP, (Inner) TCP/UDP CSO */
288 txbd1->lflags |= TX_BD_FLG_TIP_TCP_UDP_CHKSUM;
289 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_TCP_CKSUM) ==
290 PKT_TX_OIP_TCP_CKSUM) {
291 /* Outer IP, (Inner) TCP/UDP CSO */
292 txbd1->lflags |= TX_BD_FLG_TIP_TCP_UDP_CHKSUM;
293 } else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_CKSUM) ==
294 PKT_TX_OIP_IIP_CKSUM) {
295 /* Outer IP, Inner IP CSO */
296 txbd1->lflags |= TX_BD_FLG_TIP_IP_CHKSUM;
297 } else if ((tx_pkt->ol_flags & PKT_TX_TCP_UDP_CKSUM) ==
298 PKT_TX_TCP_UDP_CKSUM) {
300 txbd1->lflags |= TX_BD_LONG_LFLAGS_TCP_UDP_CHKSUM;
301 } else if ((tx_pkt->ol_flags & PKT_TX_TCP_CKSUM) ==
304 txbd1->lflags |= TX_BD_LONG_LFLAGS_TCP_UDP_CHKSUM;
305 } else if ((tx_pkt->ol_flags & PKT_TX_UDP_CKSUM) ==
308 txbd1->lflags |= TX_BD_LONG_LFLAGS_TCP_UDP_CHKSUM;
309 } else if ((tx_pkt->ol_flags & PKT_TX_IP_CKSUM) ==
312 txbd1->lflags |= TX_BD_LONG_LFLAGS_IP_CHKSUM;
313 } else if ((tx_pkt->ol_flags & PKT_TX_OUTER_IP_CKSUM) ==
314 PKT_TX_OUTER_IP_CKSUM) {
316 txbd1->lflags |= TX_BD_LONG_LFLAGS_T_IP_CHKSUM;
317 } else if ((tx_pkt->ol_flags & PKT_TX_IEEE1588_TMST) ==
318 PKT_TX_IEEE1588_TMST) {
320 txbd1->lflags |= TX_BD_LONG_LFLAGS_STAMP;
323 txbd->flags_type |= TX_BD_SHORT_TYPE_TX_BD_SHORT;
326 m_seg = tx_pkt->next;
328 /* Check non zero data_len */
329 RTE_VERIFY(m_seg->data_len);
330 txr->tx_raw_prod = RING_NEXT(txr->tx_raw_prod);
332 prod = RING_IDX(ring, txr->tx_raw_prod);
333 tx_buf = &txr->tx_buf_ring[prod];
336 txbd = &txr->tx_desc_ring[prod];
337 txbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova(m_seg));
338 txbd->flags_type = TX_BD_SHORT_TYPE_TX_BD_SHORT;
339 txbd->len = m_seg->data_len;
344 txbd->flags_type |= TX_BD_LONG_FLAGS_PACKET_END;
346 txr->tx_raw_prod = RING_NEXT(txr->tx_raw_prod);
352 * Transmit completion function for use when DEV_TX_OFFLOAD_MBUF_FAST_FREE
355 static void bnxt_tx_cmp_fast(struct bnxt_tx_queue *txq, int nr_pkts)
357 struct bnxt_tx_ring_info *txr = txq->tx_ring;
358 struct bnxt_ring *ring = txr->tx_ring_struct;
359 struct rte_mbuf **free = txq->free;
360 uint16_t raw_cons = txr->tx_raw_cons;
361 unsigned int blk = 0;
364 for (i = 0; i < nr_pkts; i++) {
365 struct rte_mbuf **tx_buf;
366 unsigned short nr_bds;
368 tx_buf = &txr->tx_buf_ring[RING_IDX(ring, raw_cons)];
369 nr_bds = (*tx_buf)->nb_segs +
370 bnxt_xmit_need_long_bd(*tx_buf, txq);
371 for (j = 0; j < nr_bds; j++) {
373 /* Add mbuf to the bulk free array */
374 free[blk++] = *tx_buf;
377 raw_cons = RING_NEXT(raw_cons);
378 tx_buf = &txr->tx_buf_ring[RING_IDX(ring, raw_cons)];
382 rte_mempool_put_bulk(free[0]->pool, (void *)free, blk);
384 txr->tx_raw_cons = raw_cons;
387 static void bnxt_tx_cmp(struct bnxt_tx_queue *txq, int nr_pkts)
389 struct bnxt_tx_ring_info *txr = txq->tx_ring;
390 struct bnxt_ring *ring = txr->tx_ring_struct;
391 struct rte_mempool *pool = NULL;
392 struct rte_mbuf **free = txq->free;
393 uint16_t raw_cons = txr->tx_raw_cons;
394 unsigned int blk = 0;
397 for (i = 0; i < nr_pkts; i++) {
398 struct rte_mbuf *mbuf;
399 struct rte_mbuf **tx_buf;
400 unsigned short nr_bds;
402 tx_buf = &txr->tx_buf_ring[RING_IDX(ring, raw_cons)];
403 nr_bds = (*tx_buf)->nb_segs +
404 bnxt_xmit_need_long_bd(*tx_buf, txq);
405 for (j = 0; j < nr_bds; j++) {
408 raw_cons = RING_NEXT(raw_cons);
409 tx_buf = &txr->tx_buf_ring[RING_IDX(ring, raw_cons)];
410 if (!mbuf) /* long_bd's tx_buf ? */
413 mbuf = rte_pktmbuf_prefree_seg(mbuf);
417 /* EW - no need to unmap DMA memory? */
419 if (likely(mbuf->pool == pool)) {
420 /* Add mbuf to the bulk free array */
423 /* Found an mbuf from a different pool. Free
424 * mbufs accumulated so far to the previous
427 if (likely(pool != NULL))
428 rte_mempool_put_bulk(pool,
432 /* Start accumulating mbufs in a new pool */
440 rte_mempool_put_bulk(pool, (void *)free, blk);
442 txr->tx_raw_cons = raw_cons;
445 static int bnxt_handle_tx_cp(struct bnxt_tx_queue *txq)
447 uint32_t nb_tx_pkts = 0, cons, ring_mask, opaque;
448 struct bnxt_cp_ring_info *cpr = txq->cp_ring;
449 uint32_t raw_cons = cpr->cp_raw_cons;
450 struct bnxt_ring *cp_ring_struct;
451 struct tx_cmpl *txcmp;
453 if (bnxt_tx_bds_in_hw(txq) < txq->tx_free_thresh)
456 cp_ring_struct = cpr->cp_ring_struct;
457 ring_mask = cp_ring_struct->ring_mask;
460 cons = RING_CMPL(ring_mask, raw_cons);
461 txcmp = (struct tx_cmpl *)&cpr->cp_desc_ring[cons];
463 if (!CMP_VALID(txcmp, raw_cons, cp_ring_struct))
466 opaque = rte_le_to_cpu_32(txcmp->opaque);
468 if (CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2)
469 nb_tx_pkts += opaque;
472 "Unhandled CMP type %02x\n",
474 raw_cons = NEXT_RAW_CMP(raw_cons);
475 } while (nb_tx_pkts < ring_mask);
477 cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);
480 if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
481 bnxt_tx_cmp_fast(txq, nb_tx_pkts);
483 bnxt_tx_cmp(txq, nb_tx_pkts);
484 cpr->cp_raw_cons = raw_cons;
491 uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
495 uint16_t nb_tx_pkts = 0;
496 uint16_t coal_pkts = 0;
497 struct bnxt_tx_queue *txq = tx_queue;
498 struct tx_bd_long *last_txbd = NULL;
500 /* Handle TX completions */
501 bnxt_handle_tx_cp(txq);
503 /* Tx queue was stopped; wait for it to be restarted */
504 if (unlikely(!txq->tx_started)) {
505 PMD_DRV_LOG(DEBUG, "Tx q stopped;return\n");
509 /* Handle TX burst request */
510 for (nb_tx_pkts = 0; nb_tx_pkts < nb_pkts; nb_tx_pkts++) {
512 rc = bnxt_start_xmit(tx_pkts[nb_tx_pkts], txq,
513 &coal_pkts, &last_txbd);
519 if (likely(nb_tx_pkts)) {
520 /* Request a completion on the last packet */
521 last_txbd->flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
522 bnxt_db_write(&txq->tx_ring->tx_db, txq->tx_ring->tx_raw_prod);
529 * Dummy DPDK callback for TX.
531 * This function is used to temporarily replace the real callback during
532 * unsafe control operations on the queue, or in case of error.
535 bnxt_dummy_xmit_pkts(void *tx_queue __rte_unused,
536 struct rte_mbuf **tx_pkts __rte_unused,
537 uint16_t nb_pkts __rte_unused)
542 int bnxt_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
544 struct bnxt *bp = dev->data->dev_private;
545 struct bnxt_tx_queue *txq = bp->tx_queues[tx_queue_id];
548 rc = is_bnxt_in_error(bp);
552 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
553 txq->tx_started = true;
554 PMD_DRV_LOG(DEBUG, "Tx queue started\n");
559 int bnxt_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
561 struct bnxt *bp = dev->data->dev_private;
562 struct bnxt_tx_queue *txq = bp->tx_queues[tx_queue_id];
565 rc = is_bnxt_in_error(bp);
569 /* Handle TX completions */
570 bnxt_handle_tx_cp(txq);
572 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
573 txq->tx_started = false;
574 PMD_DRV_LOG(DEBUG, "Tx queue stopped\n");
579 /* Sweep the Tx completion queue till HWRM_DONE for ring flush is received.
580 * The mbufs will not be freed in this call.
581 * They will be freed during ring free as a part of mem cleanup.
583 int bnxt_flush_tx_cmp(struct bnxt_cp_ring_info *cpr)
585 uint32_t raw_cons = cpr->cp_raw_cons;
587 uint32_t nb_tx_pkts = 0;
588 struct tx_cmpl *txcmp;
589 struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
590 struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
591 uint32_t ring_mask = cp_ring_struct->ring_mask;
595 cons = RING_CMPL(ring_mask, raw_cons);
596 txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
598 opaque = rte_cpu_to_le_32(txcmp->opaque);
599 raw_cons = NEXT_RAW_CMP(raw_cons);
601 if (CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2)
602 nb_tx_pkts += opaque;
603 else if (CMP_TYPE(txcmp) == HWRM_CMPL_TYPE_HWRM_DONE)
605 } while (nb_tx_pkts < ring_mask);
608 cpr->cp_raw_cons = raw_cons;