xref: /dpdk/drivers/net/intel/ice/ice_rxtx_vec_common.h (revision 5cc9919fd443fbd3fce77a257601890a0ee6a247)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2019 Intel Corporation
3  */
4 
5 #ifndef _ICE_RXTX_VEC_COMMON_H_
6 #define _ICE_RXTX_VEC_COMMON_H_
7 
8 #include "../common/rx.h"
9 #include "ice_rxtx.h"
10 
11 static inline uint16_t
12 ice_rx_reassemble_packets(struct ice_rx_queue *rxq, struct rte_mbuf **rx_bufs,
13 			  uint16_t nb_bufs, uint8_t *split_flags)
14 {
15 	struct rte_mbuf *pkts[ICE_VPMD_RX_BURST] = {0}; /*finished pkts*/
16 	struct rte_mbuf *start = rxq->pkt_first_seg;
17 	struct rte_mbuf *end =  rxq->pkt_last_seg;
18 	unsigned int pkt_idx, buf_idx;
19 
20 	for (buf_idx = 0, pkt_idx = 0; buf_idx < nb_bufs; buf_idx++) {
21 		if (end) {
22 			/* processing a split packet */
23 			end->next = rx_bufs[buf_idx];
24 			rx_bufs[buf_idx]->data_len += rxq->crc_len;
25 
26 			start->nb_segs++;
27 			start->pkt_len += rx_bufs[buf_idx]->data_len;
28 			end = end->next;
29 
30 			if (!split_flags[buf_idx]) {
31 				/* it's the last packet of the set */
32 				start->hash = end->hash;
33 				start->vlan_tci = end->vlan_tci;
34 				start->ol_flags = end->ol_flags;
35 				/* we need to strip crc for the whole packet */
36 				start->pkt_len -= rxq->crc_len;
37 				if (end->data_len > rxq->crc_len) {
38 					end->data_len -= rxq->crc_len;
39 				} else {
40 					/* free up last mbuf */
41 					struct rte_mbuf *secondlast = start;
42 
43 					start->nb_segs--;
44 					while (secondlast->next != end)
45 						secondlast = secondlast->next;
46 					secondlast->data_len -= (rxq->crc_len -
47 							end->data_len);
48 					secondlast->next = NULL;
49 					rte_pktmbuf_free_seg(end);
50 				}
51 				pkts[pkt_idx++] = start;
52 				start = NULL;
53 				end = NULL;
54 			}
55 		} else {
56 			/* not processing a split packet */
57 			if (!split_flags[buf_idx]) {
58 				/* not a split packet, save and skip */
59 				pkts[pkt_idx++] = rx_bufs[buf_idx];
60 				continue;
61 			}
62 			start = rx_bufs[buf_idx];
63 			end = start;
64 			rx_bufs[buf_idx]->data_len += rxq->crc_len;
65 			rx_bufs[buf_idx]->pkt_len += rxq->crc_len;
66 		}
67 	}
68 
69 	/* save the partial packet for next time */
70 	rxq->pkt_first_seg = start;
71 	rxq->pkt_last_seg = end;
72 	memcpy(rx_bufs, pkts, pkt_idx * (sizeof(*pkts)));
73 	return pkt_idx;
74 }
75 
76 static __rte_always_inline int
77 ice_tx_free_bufs_vec(struct ice_tx_queue *txq)
78 {
79 	struct ci_tx_entry *txep;
80 	uint32_t n;
81 	uint32_t i;
82 	int nb_free = 0;
83 	struct rte_mbuf *m, *free[ICE_TX_MAX_FREE_BUF_SZ];
84 
85 	/* check DD bits on threshold descriptor */
86 	if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
87 			rte_cpu_to_le_64(ICE_TXD_QW1_DTYPE_M)) !=
88 			rte_cpu_to_le_64(ICE_TX_DESC_DTYPE_DESC_DONE))
89 		return 0;
90 
91 	n = txq->tx_rs_thresh;
92 
93 	 /* first buffer to free from S/W ring is at index
94 	  * tx_next_dd - (tx_rs_thresh-1)
95 	  */
96 	txep = &txq->sw_ring[txq->tx_next_dd - (n - 1)];
97 	m = rte_pktmbuf_prefree_seg(txep[0].mbuf);
98 	if (likely(m)) {
99 		free[0] = m;
100 		nb_free = 1;
101 		for (i = 1; i < n; i++) {
102 			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
103 			if (likely(m)) {
104 				if (likely(m->pool == free[0]->pool)) {
105 					free[nb_free++] = m;
106 				} else {
107 					rte_mempool_put_bulk(free[0]->pool,
108 							     (void *)free,
109 							     nb_free);
110 					free[0] = m;
111 					nb_free = 1;
112 				}
113 			}
114 		}
115 		rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
116 	} else {
117 		for (i = 1; i < n; i++) {
118 			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
119 			if (m)
120 				rte_mempool_put(m->pool, m);
121 		}
122 	}
123 
124 	/* buffers were freed, update counters */
125 	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
126 	txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
127 	if (txq->tx_next_dd >= txq->nb_tx_desc)
128 		txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
129 
130 	return txq->tx_rs_thresh;
131 }
132 
133 static __rte_always_inline void
134 ice_tx_backlog_entry(struct ci_tx_entry *txep,
135 		     struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
136 {
137 	int i;
138 
139 	for (i = 0; i < (int)nb_pkts; ++i)
140 		txep[i].mbuf = tx_pkts[i];
141 }
142 
143 static inline void
144 _ice_rx_queue_release_mbufs_vec(struct ice_rx_queue *rxq)
145 {
146 	const unsigned int mask = rxq->nb_rx_desc - 1;
147 	unsigned int i;
148 
149 	if (unlikely(!rxq->sw_ring)) {
150 		PMD_DRV_LOG(DEBUG, "sw_ring is NULL");
151 		return;
152 	}
153 
154 	if (rxq->rxrearm_nb >= rxq->nb_rx_desc)
155 		return;
156 
157 	/* free all mbufs that are valid in the ring */
158 	if (rxq->rxrearm_nb == 0) {
159 		for (i = 0; i < rxq->nb_rx_desc; i++) {
160 			if (rxq->sw_ring[i].mbuf)
161 				rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
162 		}
163 	} else {
164 		for (i = rxq->rx_tail;
165 		     i != rxq->rxrearm_start;
166 		     i = (i + 1) & mask) {
167 			if (rxq->sw_ring[i].mbuf)
168 				rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
169 		}
170 	}
171 
172 	rxq->rxrearm_nb = rxq->nb_rx_desc;
173 
174 	/* set all entries to NULL */
175 	memset(rxq->sw_ring, 0, sizeof(rxq->sw_ring[0]) * rxq->nb_rx_desc);
176 }
177 
178 static inline void
179 _ice_tx_queue_release_mbufs_vec(struct ice_tx_queue *txq)
180 {
181 	uint16_t i;
182 
183 	if (unlikely(!txq || !txq->sw_ring)) {
184 		PMD_DRV_LOG(DEBUG, "Pointer to rxq or sw_ring is NULL");
185 		return;
186 	}
187 
188 	/**
189 	 *  vPMD tx will not set sw_ring's mbuf to NULL after free,
190 	 *  so need to free remains more carefully.
191 	 */
192 	i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
193 
194 #ifdef __AVX512VL__
195 	struct rte_eth_dev *dev = &rte_eth_devices[txq->vsi->adapter->pf.dev_data->port_id];
196 
197 	if (dev->tx_pkt_burst == ice_xmit_pkts_vec_avx512 ||
198 	    dev->tx_pkt_burst == ice_xmit_pkts_vec_avx512_offload) {
199 		struct ci_tx_entry_vec *swr = (void *)txq->sw_ring;
200 
201 		if (txq->tx_tail < i) {
202 			for (; i < txq->nb_tx_desc; i++) {
203 				rte_pktmbuf_free_seg(swr[i].mbuf);
204 				swr[i].mbuf = NULL;
205 			}
206 			i = 0;
207 		}
208 		for (; i < txq->tx_tail; i++) {
209 			rte_pktmbuf_free_seg(swr[i].mbuf);
210 			swr[i].mbuf = NULL;
211 		}
212 	} else
213 #endif
214 	{
215 		if (txq->tx_tail < i) {
216 			for (; i < txq->nb_tx_desc; i++) {
217 				rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
218 				txq->sw_ring[i].mbuf = NULL;
219 			}
220 			i = 0;
221 		}
222 		for (; i < txq->tx_tail; i++) {
223 			rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
224 			txq->sw_ring[i].mbuf = NULL;
225 		}
226 	}
227 }
228 
229 static inline int
230 ice_rxq_vec_setup_default(struct ice_rx_queue *rxq)
231 {
232 	uintptr_t p;
233 	struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
234 
235 	mb_def.nb_segs = 1;
236 	mb_def.data_off = RTE_PKTMBUF_HEADROOM;
237 	mb_def.port = rxq->port_id;
238 	rte_mbuf_refcnt_set(&mb_def, 1);
239 
240 	/* prevent compiler reordering: rearm_data covers previous fields */
241 	rte_compiler_barrier();
242 	p = (uintptr_t)&mb_def.rearm_data;
243 	rxq->mbuf_initializer = *(uint64_t *)p;
244 	return 0;
245 }
246 
247 #define ICE_TX_NO_VECTOR_FLAGS (			\
248 		RTE_ETH_TX_OFFLOAD_MULTI_SEGS |		\
249 		RTE_ETH_TX_OFFLOAD_OUTER_IPV4_CKSUM |	\
250 		RTE_ETH_TX_OFFLOAD_TCP_TSO |	\
251 		RTE_ETH_TX_OFFLOAD_VXLAN_TNL_TSO |    \
252 		RTE_ETH_TX_OFFLOAD_GRE_TNL_TSO |    \
253 		RTE_ETH_TX_OFFLOAD_IPIP_TNL_TSO |    \
254 		RTE_ETH_TX_OFFLOAD_GENEVE_TNL_TSO |    \
255 		RTE_ETH_TX_OFFLOAD_OUTER_UDP_CKSUM)
256 
257 #define ICE_TX_VECTOR_OFFLOAD (				\
258 		RTE_ETH_TX_OFFLOAD_VLAN_INSERT |		\
259 		RTE_ETH_TX_OFFLOAD_QINQ_INSERT |		\
260 		RTE_ETH_TX_OFFLOAD_IPV4_CKSUM |		\
261 		RTE_ETH_TX_OFFLOAD_SCTP_CKSUM |		\
262 		RTE_ETH_TX_OFFLOAD_UDP_CKSUM |		\
263 		RTE_ETH_TX_OFFLOAD_TCP_CKSUM)
264 
265 #define ICE_RX_VECTOR_OFFLOAD (				\
266 		RTE_ETH_RX_OFFLOAD_CHECKSUM |		\
267 		RTE_ETH_RX_OFFLOAD_SCTP_CKSUM |		\
268 		RTE_ETH_RX_OFFLOAD_VLAN |			\
269 		RTE_ETH_RX_OFFLOAD_RSS_HASH)
270 
271 #define ICE_VECTOR_PATH		0
272 #define ICE_VECTOR_OFFLOAD_PATH	1
273 
274 static inline int
275 ice_rx_vec_queue_default(struct ice_rx_queue *rxq)
276 {
277 	if (!rxq)
278 		return -1;
279 
280 	if (!rte_is_power_of_2(rxq->nb_rx_desc))
281 		return -1;
282 
283 	if (rxq->rx_free_thresh < ICE_VPMD_RX_BURST)
284 		return -1;
285 
286 	if (rxq->nb_rx_desc % rxq->rx_free_thresh)
287 		return -1;
288 
289 	if (rxq->proto_xtr != PROTO_XTR_NONE)
290 		return -1;
291 
292 	if (rxq->offloads & RTE_ETH_RX_OFFLOAD_TIMESTAMP)
293 		return -1;
294 
295 	if (rxq->offloads & RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT)
296 		return -1;
297 
298 	if (rxq->offloads & ICE_RX_VECTOR_OFFLOAD)
299 		return ICE_VECTOR_OFFLOAD_PATH;
300 
301 	return ICE_VECTOR_PATH;
302 }
303 
304 static inline int
305 ice_tx_vec_queue_default(struct ice_tx_queue *txq)
306 {
307 	if (!txq)
308 		return -1;
309 
310 	if (txq->tx_rs_thresh < ICE_VPMD_TX_BURST ||
311 	    txq->tx_rs_thresh > ICE_TX_MAX_FREE_BUF_SZ)
312 		return -1;
313 
314 	if (txq->offloads & ICE_TX_NO_VECTOR_FLAGS)
315 		return -1;
316 
317 	if (txq->offloads & ICE_TX_VECTOR_OFFLOAD)
318 		return ICE_VECTOR_OFFLOAD_PATH;
319 
320 	return ICE_VECTOR_PATH;
321 }
322 
323 static inline int
324 ice_rx_vec_dev_check_default(struct rte_eth_dev *dev)
325 {
326 	int i;
327 	struct ice_rx_queue *rxq;
328 	int ret = 0;
329 	int result = 0;
330 
331 	for (i = 0; i < dev->data->nb_rx_queues; i++) {
332 		rxq = dev->data->rx_queues[i];
333 		ret = (ice_rx_vec_queue_default(rxq));
334 		if (ret < 0)
335 			return -1;
336 		if (ret == ICE_VECTOR_OFFLOAD_PATH)
337 			result = ret;
338 	}
339 
340 	return result;
341 }
342 
343 static inline int
344 ice_tx_vec_dev_check_default(struct rte_eth_dev *dev)
345 {
346 	int i;
347 	struct ice_tx_queue *txq;
348 	int ret = 0;
349 	int result = 0;
350 
351 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
352 		txq = dev->data->tx_queues[i];
353 		ret = ice_tx_vec_queue_default(txq);
354 		if (ret < 0)
355 			return -1;
356 		if (ret == ICE_VECTOR_OFFLOAD_PATH)
357 			result = ret;
358 	}
359 
360 	return result;
361 }
362 
363 static inline void
364 ice_txd_enable_offload(struct rte_mbuf *tx_pkt,
365 		       uint64_t *txd_hi)
366 {
367 	uint64_t ol_flags = tx_pkt->ol_flags;
368 	uint32_t td_cmd = 0;
369 	uint32_t td_offset = 0;
370 
371 	/* Tx Checksum Offload */
372 	/* SET MACLEN */
373 	td_offset |= (tx_pkt->l2_len >> 1) <<
374 		ICE_TX_DESC_LEN_MACLEN_S;
375 
376 	/* Enable L3 checksum offload */
377 	if (ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
378 		td_cmd |= ICE_TX_DESC_CMD_IIPT_IPV4_CSUM;
379 		td_offset |= (tx_pkt->l3_len >> 2) <<
380 			ICE_TX_DESC_LEN_IPLEN_S;
381 	} else if (ol_flags & RTE_MBUF_F_TX_IPV4) {
382 		td_cmd |= ICE_TX_DESC_CMD_IIPT_IPV4;
383 		td_offset |= (tx_pkt->l3_len >> 2) <<
384 			ICE_TX_DESC_LEN_IPLEN_S;
385 	} else if (ol_flags & RTE_MBUF_F_TX_IPV6) {
386 		td_cmd |= ICE_TX_DESC_CMD_IIPT_IPV6;
387 		td_offset |= (tx_pkt->l3_len >> 2) <<
388 			ICE_TX_DESC_LEN_IPLEN_S;
389 	}
390 
391 	/* Enable L4 checksum offloads */
392 	switch (ol_flags & RTE_MBUF_F_TX_L4_MASK) {
393 	case RTE_MBUF_F_TX_TCP_CKSUM:
394 		td_cmd |= ICE_TX_DESC_CMD_L4T_EOFT_TCP;
395 		td_offset |= (sizeof(struct rte_tcp_hdr) >> 2) <<
396 			ICE_TX_DESC_LEN_L4_LEN_S;
397 		break;
398 	case RTE_MBUF_F_TX_SCTP_CKSUM:
399 		td_cmd |= ICE_TX_DESC_CMD_L4T_EOFT_SCTP;
400 		td_offset |= (sizeof(struct rte_sctp_hdr) >> 2) <<
401 			ICE_TX_DESC_LEN_L4_LEN_S;
402 		break;
403 	case RTE_MBUF_F_TX_UDP_CKSUM:
404 		td_cmd |= ICE_TX_DESC_CMD_L4T_EOFT_UDP;
405 		td_offset |= (sizeof(struct rte_udp_hdr) >> 2) <<
406 			ICE_TX_DESC_LEN_L4_LEN_S;
407 		break;
408 	default:
409 		break;
410 	}
411 
412 	*txd_hi |= ((uint64_t)td_offset) << ICE_TXD_QW1_OFFSET_S;
413 
414 	/* Tx VLAN/QINQ insertion Offload */
415 	if (ol_flags & (RTE_MBUF_F_TX_VLAN | RTE_MBUF_F_TX_QINQ)) {
416 		td_cmd |= ICE_TX_DESC_CMD_IL2TAG1;
417 		*txd_hi |= ((uint64_t)tx_pkt->vlan_tci <<
418 				ICE_TXD_QW1_L2TAG1_S);
419 	}
420 
421 	*txd_hi |= ((uint64_t)td_cmd) << ICE_TXD_QW1_CMD_S;
422 }
423 #endif
424