xref: /dpdk/drivers/net/sfc/sfc_dp_tx.h (revision 68a03efeed657e6e05f281479b33b51102797e15)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  *
3  * Copyright(c) 2019-2021 Xilinx, Inc.
4  * Copyright(c) 2016-2019 Solarflare Communications Inc.
5  *
6  * This software was jointly developed between OKTET Labs (under contract
7  * for Solarflare) and Solarflare Communications, Inc.
8  */
9 
10 #ifndef _SFC_DP_TX_H
11 #define _SFC_DP_TX_H
12 
13 #include <ethdev_driver.h>
14 
15 #include "sfc_dp.h"
16 #include "sfc_debug.h"
17 #include "sfc_tso.h"
18 
19 #ifdef __cplusplus
20 extern "C" {
21 #endif
22 
23 /**
24  * Generic transmit queue information used on data path.
25  * It must be kept as small as it is possible since it is built into
26  * the structure used on datapath.
27  */
28 struct sfc_dp_txq {
29 	struct sfc_dp_queue	dpq;
30 };
31 
32 /** Datapath transmit queue descriptor number limitations */
33 struct sfc_dp_tx_hw_limits {
34 	unsigned int txq_max_entries;
35 	unsigned int txq_min_entries;
36 };
37 
38 /**
39  * Datapath transmit queue creation information.
40  *
41  * The structure is used just to pass information from control path to
42  * datapath. It could be just function arguments, but it would be hardly
43  * readable.
44  */
45 struct sfc_dp_tx_qcreate_info {
46 	/** Maximum number of pushed Tx descriptors */
47 	unsigned int		max_fill_level;
48 	/** Minimum number of unused Tx descriptors to do reap */
49 	unsigned int		free_thresh;
50 	/** Offloads enabled on the transmit queue */
51 	uint64_t		offloads;
52 	/** Tx queue size */
53 	unsigned int		txq_entries;
54 	/** Maximum size of data in the DMA descriptor */
55 	uint16_t		dma_desc_size_max;
56 	/** DMA-mapped Tx descriptors ring */
57 	void			*txq_hw_ring;
58 	/** Associated event queue size */
59 	unsigned int		evq_entries;
60 	/** Hardware event ring */
61 	void			*evq_hw_ring;
62 	/** The queue index in hardware (required to push right doorbell) */
63 	unsigned int		hw_index;
64 	/** Virtual address of the memory-mapped BAR to push Tx doorbell */
65 	volatile void		*mem_bar;
66 	/** VI window size shift */
67 	unsigned int		vi_window_shift;
68 	/**
69 	 * Maximum number of bytes into the packet the TCP header can start for
70 	 * the hardware to apply TSO packet edits.
71 	 */
72 	uint16_t		tso_tcp_header_offset_limit;
73 	/** Maximum number of header DMA descriptors per TSOv3 transaction */
74 	uint16_t		tso_max_nb_header_descs;
75 	/** Maximum header length acceptable by TSOv3 transaction */
76 	uint16_t		tso_max_header_len;
77 	/** Maximum number of payload DMA descriptors per TSOv3 transaction */
78 	uint16_t		tso_max_nb_payload_descs;
79 	/** Maximum payload length per TSOv3 transaction */
80 	uint32_t		tso_max_payload_len;
81 	/** Maximum number of frames to be generated per TSOv3 transaction */
82 	uint32_t		tso_max_nb_outgoing_frames;
83 };
84 
85 /**
86  * Get Tx datapath specific device info.
87  *
88  * @param dev_info		Device info to be adjusted
89  */
90 typedef void (sfc_dp_tx_get_dev_info_t)(struct rte_eth_dev_info *dev_info);
91 
92 /**
93  * Get size of transmit and event queue rings by the number of Tx
94  * descriptors.
95  *
96  * @param nb_tx_desc		Number of Tx descriptors
97  * @param txq_entries		Location for number of Tx ring entries
98  * @param evq_entries		Location for number of event ring entries
99  * @param txq_max_fill_level	Location for maximum Tx ring fill level
100  *
101  * @return 0 or positive errno.
102  */
103 typedef int (sfc_dp_tx_qsize_up_rings_t)(uint16_t nb_tx_desc,
104 					 struct sfc_dp_tx_hw_limits *limits,
105 					 unsigned int *txq_entries,
106 					 unsigned int *evq_entries,
107 					 unsigned int *txq_max_fill_level);
108 
109 /**
110  * Allocate and initialize datapath transmit queue.
111  *
112  * @param port_id	The port identifier
113  * @param queue_id	The queue identifier
114  * @param pci_addr	PCI function address
115  * @param socket_id	Socket identifier to allocate memory
116  * @param info		Tx queue details wrapped in structure
117  * @param dp_txqp	Location for generic datapath transmit queue pointer
118  *
119  * @return 0 or positive errno.
120  */
121 typedef int (sfc_dp_tx_qcreate_t)(uint16_t port_id, uint16_t queue_id,
122 				  const struct rte_pci_addr *pci_addr,
123 				  int socket_id,
124 				  const struct sfc_dp_tx_qcreate_info *info,
125 				  struct sfc_dp_txq **dp_txqp);
126 
127 /**
128  * Free resources allocated for datapath transmit queue.
129  */
130 typedef void (sfc_dp_tx_qdestroy_t)(struct sfc_dp_txq *dp_txq);
131 
132 /**
133  * Transmit queue start callback.
134  *
135  * It handovers EvQ to the datapath.
136  */
137 typedef int (sfc_dp_tx_qstart_t)(struct sfc_dp_txq *dp_txq,
138 				 unsigned int evq_read_ptr,
139 				 unsigned int txq_desc_index);
140 
141 /**
142  * Transmit queue stop function called before the queue flush.
143  *
144  * It returns EvQ to the control path.
145  */
146 typedef void (sfc_dp_tx_qstop_t)(struct sfc_dp_txq *dp_txq,
147 				 unsigned int *evq_read_ptr);
148 
149 /**
150  * Transmit event handler used during queue flush only.
151  */
152 typedef bool (sfc_dp_tx_qtx_ev_t)(struct sfc_dp_txq *dp_txq, unsigned int id);
153 
154 /**
155  * Transmit queue function called after the queue flush.
156  */
157 typedef void (sfc_dp_tx_qreap_t)(struct sfc_dp_txq *dp_txq);
158 
159 /**
160  * Check Tx descriptor status
161  */
162 typedef int (sfc_dp_tx_qdesc_status_t)(struct sfc_dp_txq *dp_txq,
163 				       uint16_t offset);
164 
165 /** Transmit datapath definition */
166 struct sfc_dp_tx {
167 	struct sfc_dp			dp;
168 
169 	unsigned int			features;
170 #define SFC_DP_TX_FEAT_MULTI_PROCESS	0x1
171 	/**
172 	 * Tx offload capabilities supported by the datapath on device
173 	 * level only if HW/FW supports it.
174 	 */
175 	uint64_t			dev_offload_capa;
176 	/**
177 	 * Tx offload capabilities supported by the datapath per-queue
178 	 * if HW/FW supports it.
179 	 */
180 	uint64_t			queue_offload_capa;
181 	sfc_dp_tx_get_dev_info_t	*get_dev_info;
182 	sfc_dp_tx_qsize_up_rings_t	*qsize_up_rings;
183 	sfc_dp_tx_qcreate_t		*qcreate;
184 	sfc_dp_tx_qdestroy_t		*qdestroy;
185 	sfc_dp_tx_qstart_t		*qstart;
186 	sfc_dp_tx_qstop_t		*qstop;
187 	sfc_dp_tx_qtx_ev_t		*qtx_ev;
188 	sfc_dp_tx_qreap_t		*qreap;
189 	sfc_dp_tx_qdesc_status_t	*qdesc_status;
190 	eth_tx_prep_t			pkt_prepare;
191 	eth_tx_burst_t			pkt_burst;
192 };
193 
194 static inline struct sfc_dp_tx *
195 sfc_dp_find_tx_by_name(struct sfc_dp_list *head, const char *name)
196 {
197 	struct sfc_dp *p = sfc_dp_find_by_name(head, SFC_DP_TX, name);
198 
199 	return (p == NULL) ? NULL : container_of(p, struct sfc_dp_tx, dp);
200 }
201 
202 static inline struct sfc_dp_tx *
203 sfc_dp_find_tx_by_caps(struct sfc_dp_list *head, unsigned int avail_caps)
204 {
205 	struct sfc_dp *p = sfc_dp_find_by_caps(head, SFC_DP_TX, avail_caps);
206 
207 	return (p == NULL) ? NULL : container_of(p, struct sfc_dp_tx, dp);
208 }
209 
210 /** Get Tx datapath ops by the datapath TxQ handle */
211 const struct sfc_dp_tx *sfc_dp_tx_by_dp_txq(const struct sfc_dp_txq *dp_txq);
212 
213 static inline uint64_t
214 sfc_dp_tx_offload_capa(const struct sfc_dp_tx *dp_tx)
215 {
216 	return dp_tx->dev_offload_capa | dp_tx->queue_offload_capa;
217 }
218 
219 static inline unsigned int
220 sfc_dp_tx_pkt_extra_hdr_segs(struct rte_mbuf **m_seg,
221 			     unsigned int *header_len_remaining)
222 {
223 	unsigned int nb_extra_header_segs = 0;
224 
225 	while (rte_pktmbuf_data_len(*m_seg) < *header_len_remaining) {
226 		*header_len_remaining -= rte_pktmbuf_data_len(*m_seg);
227 		*m_seg = (*m_seg)->next;
228 		++nb_extra_header_segs;
229 	}
230 
231 	return nb_extra_header_segs;
232 }
233 
234 static inline int
235 sfc_dp_tx_prepare_pkt(struct rte_mbuf *m,
236 			   unsigned int max_nb_header_segs,
237 			   unsigned int tso_bounce_buffer_len,
238 			   uint32_t tso_tcp_header_offset_limit,
239 			   unsigned int max_fill_level,
240 			   unsigned int nb_tso_descs,
241 			   unsigned int nb_vlan_descs)
242 {
243 	unsigned int descs_required = m->nb_segs;
244 	unsigned int tcph_off = ((m->ol_flags & PKT_TX_TUNNEL_MASK) ?
245 				 m->outer_l2_len + m->outer_l3_len : 0) +
246 				m->l2_len + m->l3_len;
247 	unsigned int header_len = tcph_off + m->l4_len;
248 	unsigned int header_len_remaining = header_len;
249 	unsigned int nb_header_segs = 1;
250 	struct rte_mbuf *m_seg = m;
251 
252 #ifdef RTE_LIBRTE_SFC_EFX_DEBUG
253 	int ret;
254 
255 	ret = rte_validate_tx_offload(m);
256 	if (ret != 0) {
257 		/*
258 		 * Negative error code is returned by rte_validate_tx_offload(),
259 		 * but positive are used inside net/sfc PMD.
260 		 */
261 		SFC_ASSERT(ret < 0);
262 		return -ret;
263 	}
264 #endif
265 
266 	if (max_nb_header_segs != 0) {
267 		/* There is a limit on the number of header segments. */
268 
269 		nb_header_segs +=
270 		    sfc_dp_tx_pkt_extra_hdr_segs(&m_seg,
271 						 &header_len_remaining);
272 
273 		if (unlikely(nb_header_segs > max_nb_header_segs)) {
274 			/*
275 			 * The number of header segments is too large.
276 			 *
277 			 * If TSO is requested and if the datapath supports
278 			 * linearisation of TSO headers, allow the packet
279 			 * to proceed with additional checks below.
280 			 * Otherwise, throw an error.
281 			 */
282 			if ((m->ol_flags & PKT_TX_TCP_SEG) == 0 ||
283 			    tso_bounce_buffer_len == 0)
284 				return EINVAL;
285 		}
286 	}
287 
288 	if (m->ol_flags & PKT_TX_TCP_SEG) {
289 		switch (m->ol_flags & PKT_TX_TUNNEL_MASK) {
290 		case 0:
291 			break;
292 		case PKT_TX_TUNNEL_VXLAN:
293 			/* FALLTHROUGH */
294 		case PKT_TX_TUNNEL_GENEVE:
295 			if (!(m->ol_flags &
296 			      (PKT_TX_OUTER_IPV4 | PKT_TX_OUTER_IPV6)))
297 				return EINVAL;
298 		}
299 
300 		if (unlikely(tcph_off > tso_tcp_header_offset_limit))
301 			return EINVAL;
302 
303 		descs_required += nb_tso_descs;
304 
305 		/*
306 		 * If headers segments are already counted above, here
307 		 * nothing is done since remaining length is smaller
308 		 * then current segment size.
309 		 */
310 		nb_header_segs +=
311 		    sfc_dp_tx_pkt_extra_hdr_segs(&m_seg,
312 						 &header_len_remaining);
313 
314 		/*
315 		 * Extra descriptor which is required when (a part of) payload
316 		 * shares the same segment with (a part of) the header.
317 		 */
318 		if (rte_pktmbuf_data_len(m_seg) > header_len_remaining)
319 			descs_required++;
320 
321 		if (tso_bounce_buffer_len != 0) {
322 			if (nb_header_segs > 1 &&
323 			    unlikely(header_len > tso_bounce_buffer_len)) {
324 				/*
325 				 * Header linearization is required and
326 				 * the header is too big to be linearized
327 				 */
328 				return EINVAL;
329 			}
330 		}
331 	}
332 
333 	/*
334 	 * The number of VLAN descriptors is added regardless of requested
335 	 * VLAN offload since VLAN is sticky and sending packet without VLAN
336 	 * insertion may require VLAN descriptor to reset the sticky to 0.
337 	 */
338 	descs_required += nb_vlan_descs;
339 
340 	/*
341 	 * Max fill level must be sufficient to hold all required descriptors
342 	 * to send the packet entirely.
343 	 */
344 	if (descs_required > max_fill_level)
345 		return ENOBUFS;
346 
347 	return 0;
348 }
349 
350 extern struct sfc_dp_tx sfc_efx_tx;
351 extern struct sfc_dp_tx sfc_ef10_tx;
352 extern struct sfc_dp_tx sfc_ef10_simple_tx;
353 extern struct sfc_dp_tx sfc_ef100_tx;
354 
355 #ifdef __cplusplus
356 }
357 #endif
358 #endif /* _SFC_DP_TX_H */
359