• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
2 /* Copyright (c) 2019 Mellanox Technologies. */
3 
4 #ifndef __MLX5_EN_TXRX_H___
5 #define __MLX5_EN_TXRX_H___
6 
7 #include "en.h"
8 #include <linux/indirect_call_wrapper.h>
9 
10 #define MLX5E_TX_WQE_EMPTY_DS_COUNT (sizeof(struct mlx5e_tx_wqe) / MLX5_SEND_WQE_DS)
11 
12 /* The mult of MLX5_SEND_WQE_MAX_WQEBBS * MLX5_SEND_WQEBB_NUM_DS
13  * (16 * 4 == 64) does not fit in the 6-bit DS field of Ctrl Segment.
14  * We use a bound lower that MLX5_SEND_WQE_MAX_WQEBBS to let a
15  * full-session WQE be cache-aligned.
16  */
17 #if L1_CACHE_BYTES < 128
18 #define MLX5E_TX_MPW_MAX_WQEBBS (MLX5_SEND_WQE_MAX_WQEBBS - 1)
19 #else
20 #define MLX5E_TX_MPW_MAX_WQEBBS (MLX5_SEND_WQE_MAX_WQEBBS - 2)
21 #endif
22 
23 #define MLX5E_TX_MPW_MAX_NUM_DS (MLX5E_TX_MPW_MAX_WQEBBS * MLX5_SEND_WQEBB_NUM_DS)
24 
25 #define INL_HDR_START_SZ (sizeof(((struct mlx5_wqe_eth_seg *)NULL)->inline_hdr.start))
26 
27 #define MLX5E_RX_ERR_CQE(cqe) (get_cqe_opcode(cqe) != MLX5_CQE_RESP_SEND)
28 
29 static inline
mlx5e_cqe_ts_to_ns(cqe_ts_to_ns func,struct mlx5_clock * clock,u64 cqe_ts)30 ktime_t mlx5e_cqe_ts_to_ns(cqe_ts_to_ns func, struct mlx5_clock *clock, u64 cqe_ts)
31 {
32 	return INDIRECT_CALL_2(func, mlx5_real_time_cyc2time, mlx5_timecounter_cyc2time,
33 			       clock, cqe_ts);
34 }
35 
36 enum mlx5e_icosq_wqe_type {
37 	MLX5E_ICOSQ_WQE_NOP,
38 	MLX5E_ICOSQ_WQE_UMR_RX,
39 #ifdef CONFIG_MLX5_EN_TLS
40 	MLX5E_ICOSQ_WQE_UMR_TLS,
41 	MLX5E_ICOSQ_WQE_SET_PSV_TLS,
42 	MLX5E_ICOSQ_WQE_GET_PSV_TLS,
43 #endif
44 };
45 
46 /* General */
mlx5e_skb_is_multicast(struct sk_buff * skb)47 static inline bool mlx5e_skb_is_multicast(struct sk_buff *skb)
48 {
49 	return skb->pkt_type == PACKET_MULTICAST || skb->pkt_type == PACKET_BROADCAST;
50 }
51 
52 void mlx5e_trigger_irq(struct mlx5e_icosq *sq);
53 void mlx5e_completion_event(struct mlx5_core_cq *mcq, struct mlx5_eqe *eqe);
54 void mlx5e_cq_error_event(struct mlx5_core_cq *mcq, enum mlx5_event event);
55 int mlx5e_napi_poll(struct napi_struct *napi, int budget);
56 int mlx5e_poll_ico_cq(struct mlx5e_cq *cq);
57 
58 /* RX */
59 void mlx5e_page_dma_unmap(struct mlx5e_rq *rq, struct mlx5e_dma_info *dma_info);
60 void mlx5e_page_release_dynamic(struct mlx5e_rq *rq,
61 				struct mlx5e_dma_info *dma_info,
62 				bool recycle);
63 INDIRECT_CALLABLE_DECLARE(bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq));
64 INDIRECT_CALLABLE_DECLARE(bool mlx5e_post_rx_mpwqes(struct mlx5e_rq *rq));
65 int mlx5e_poll_rx_cq(struct mlx5e_cq *cq, int budget);
66 void mlx5e_free_rx_descs(struct mlx5e_rq *rq);
67 void mlx5e_free_rx_in_progress_descs(struct mlx5e_rq *rq);
68 
69 /* TX */
70 u16 mlx5e_select_queue(struct net_device *dev, struct sk_buff *skb,
71 		       struct net_device *sb_dev);
72 netdev_tx_t mlx5e_xmit(struct sk_buff *skb, struct net_device *dev);
73 bool mlx5e_poll_tx_cq(struct mlx5e_cq *cq, int napi_budget);
74 void mlx5e_free_txqsq_descs(struct mlx5e_txqsq *sq);
75 
76 static inline bool
mlx5e_skb_fifo_has_room(struct mlx5e_skb_fifo * fifo)77 mlx5e_skb_fifo_has_room(struct mlx5e_skb_fifo *fifo)
78 {
79 	return (*fifo->pc - *fifo->cc) < fifo->mask;
80 }
81 
82 static inline bool
mlx5e_wqc_has_room_for(struct mlx5_wq_cyc * wq,u16 cc,u16 pc,u16 n)83 mlx5e_wqc_has_room_for(struct mlx5_wq_cyc *wq, u16 cc, u16 pc, u16 n)
84 {
85 	return (mlx5_wq_cyc_ctr2ix(wq, cc - pc) >= n) || (cc == pc);
86 }
87 
mlx5e_fetch_wqe(struct mlx5_wq_cyc * wq,u16 pi,size_t wqe_size)88 static inline void *mlx5e_fetch_wqe(struct mlx5_wq_cyc *wq, u16 pi, size_t wqe_size)
89 {
90 	void *wqe;
91 
92 	wqe = mlx5_wq_cyc_get_wqe(wq, pi);
93 	memset(wqe, 0, wqe_size);
94 
95 	return wqe;
96 }
97 
98 #define MLX5E_TX_FETCH_WQE(sq, pi) \
99 	((struct mlx5e_tx_wqe *)mlx5e_fetch_wqe(&(sq)->wq, pi, sizeof(struct mlx5e_tx_wqe)))
100 
101 static inline struct mlx5e_tx_wqe *
mlx5e_post_nop(struct mlx5_wq_cyc * wq,u32 sqn,u16 * pc)102 mlx5e_post_nop(struct mlx5_wq_cyc *wq, u32 sqn, u16 *pc)
103 {
104 	u16                         pi   = mlx5_wq_cyc_ctr2ix(wq, *pc);
105 	struct mlx5e_tx_wqe        *wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
106 	struct mlx5_wqe_ctrl_seg   *cseg = &wqe->ctrl;
107 
108 	memset(cseg, 0, sizeof(*cseg));
109 
110 	cseg->opmod_idx_opcode = cpu_to_be32((*pc << 8) | MLX5_OPCODE_NOP);
111 	cseg->qpn_ds           = cpu_to_be32((sqn << 8) | 0x01);
112 
113 	(*pc)++;
114 
115 	return wqe;
116 }
117 
118 static inline struct mlx5e_tx_wqe *
mlx5e_post_nop_fence(struct mlx5_wq_cyc * wq,u32 sqn,u16 * pc)119 mlx5e_post_nop_fence(struct mlx5_wq_cyc *wq, u32 sqn, u16 *pc)
120 {
121 	u16                         pi   = mlx5_wq_cyc_ctr2ix(wq, *pc);
122 	struct mlx5e_tx_wqe        *wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
123 	struct mlx5_wqe_ctrl_seg   *cseg = &wqe->ctrl;
124 
125 	memset(cseg, 0, sizeof(*cseg));
126 
127 	cseg->opmod_idx_opcode = cpu_to_be32((*pc << 8) | MLX5_OPCODE_NOP);
128 	cseg->qpn_ds           = cpu_to_be32((sqn << 8) | 0x01);
129 	cseg->fm_ce_se         = MLX5_FENCE_MODE_INITIATOR_SMALL;
130 
131 	(*pc)++;
132 
133 	return wqe;
134 }
135 
136 struct mlx5e_tx_wqe_info {
137 	struct sk_buff *skb;
138 	u32 num_bytes;
139 	u8 num_wqebbs;
140 	u8 num_dma;
141 	u8 num_fifo_pkts;
142 #ifdef CONFIG_MLX5_EN_TLS
143 	struct page *resync_dump_frag_page;
144 #endif
145 };
146 
mlx5e_txqsq_get_next_pi(struct mlx5e_txqsq * sq,u16 size)147 static inline u16 mlx5e_txqsq_get_next_pi(struct mlx5e_txqsq *sq, u16 size)
148 {
149 	struct mlx5_wq_cyc *wq = &sq->wq;
150 	u16 pi, contig_wqebbs;
151 
152 	pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
153 	contig_wqebbs = mlx5_wq_cyc_get_contig_wqebbs(wq, pi);
154 	if (unlikely(contig_wqebbs < size)) {
155 		struct mlx5e_tx_wqe_info *wi, *edge_wi;
156 
157 		wi = &sq->db.wqe_info[pi];
158 		edge_wi = wi + contig_wqebbs;
159 
160 		/* Fill SQ frag edge with NOPs to avoid WQE wrapping two pages. */
161 		for (; wi < edge_wi; wi++) {
162 			*wi = (struct mlx5e_tx_wqe_info) {
163 				.num_wqebbs = 1,
164 			};
165 			mlx5e_post_nop(wq, sq->sqn, &sq->pc);
166 		}
167 		sq->stats->nop += contig_wqebbs;
168 
169 		pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
170 	}
171 
172 	return pi;
173 }
174 
175 void mlx5e_txqsq_wake(struct mlx5e_txqsq *sq);
176 
177 struct mlx5e_icosq_wqe_info {
178 	u8 wqe_type;
179 	u8 num_wqebbs;
180 
181 	/* Auxiliary data for different wqe types. */
182 	union {
183 		struct {
184 			struct mlx5e_rq *rq;
185 		} umr;
186 #ifdef CONFIG_MLX5_EN_TLS
187 		struct {
188 			struct mlx5e_ktls_offload_context_rx *priv_rx;
189 		} tls_set_params;
190 		struct {
191 			struct mlx5e_ktls_rx_resync_buf *buf;
192 		} tls_get_params;
193 #endif
194 	};
195 };
196 
197 void mlx5e_free_icosq_descs(struct mlx5e_icosq *sq);
198 
mlx5e_icosq_get_next_pi(struct mlx5e_icosq * sq,u16 size)199 static inline u16 mlx5e_icosq_get_next_pi(struct mlx5e_icosq *sq, u16 size)
200 {
201 	struct mlx5_wq_cyc *wq = &sq->wq;
202 	u16 pi, contig_wqebbs;
203 
204 	pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
205 	contig_wqebbs = mlx5_wq_cyc_get_contig_wqebbs(wq, pi);
206 	if (unlikely(contig_wqebbs < size)) {
207 		struct mlx5e_icosq_wqe_info *wi, *edge_wi;
208 
209 		wi = &sq->db.wqe_info[pi];
210 		edge_wi = wi + contig_wqebbs;
211 
212 		/* Fill SQ frag edge with NOPs to avoid WQE wrapping two pages. */
213 		for (; wi < edge_wi; wi++) {
214 			*wi = (struct mlx5e_icosq_wqe_info) {
215 				.wqe_type   = MLX5E_ICOSQ_WQE_NOP,
216 				.num_wqebbs = 1,
217 			};
218 			mlx5e_post_nop(wq, sq->sqn, &sq->pc);
219 		}
220 
221 		pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
222 	}
223 
224 	return pi;
225 }
226 
227 static inline void
mlx5e_notify_hw(struct mlx5_wq_cyc * wq,u16 pc,void __iomem * uar_map,struct mlx5_wqe_ctrl_seg * ctrl)228 mlx5e_notify_hw(struct mlx5_wq_cyc *wq, u16 pc, void __iomem *uar_map,
229 		struct mlx5_wqe_ctrl_seg *ctrl)
230 {
231 	ctrl->fm_ce_se |= MLX5_WQE_CTRL_CQ_UPDATE;
232 	/* ensure wqe is visible to device before updating doorbell record */
233 	dma_wmb();
234 
235 	*wq->db = cpu_to_be32(pc);
236 
237 	/* ensure doorbell record is visible to device before ringing the
238 	 * doorbell
239 	 */
240 	wmb();
241 
242 	mlx5_write64((__be32 *)ctrl, uar_map);
243 }
244 
mlx5e_cq_arm(struct mlx5e_cq * cq)245 static inline void mlx5e_cq_arm(struct mlx5e_cq *cq)
246 {
247 	struct mlx5_core_cq *mcq;
248 
249 	mcq = &cq->mcq;
250 	mlx5_cq_arm(mcq, MLX5_CQ_DB_REQ_NOT, mcq->uar->map, cq->wq.cc);
251 }
252 
253 static inline struct mlx5e_sq_dma *
mlx5e_dma_get(struct mlx5e_txqsq * sq,u32 i)254 mlx5e_dma_get(struct mlx5e_txqsq *sq, u32 i)
255 {
256 	return &sq->db.dma_fifo[i & sq->dma_fifo_mask];
257 }
258 
259 static inline void
mlx5e_dma_push(struct mlx5e_txqsq * sq,dma_addr_t addr,u32 size,enum mlx5e_dma_map_type map_type)260 mlx5e_dma_push(struct mlx5e_txqsq *sq, dma_addr_t addr, u32 size,
261 	       enum mlx5e_dma_map_type map_type)
262 {
263 	struct mlx5e_sq_dma *dma = mlx5e_dma_get(sq, sq->dma_fifo_pc++);
264 
265 	dma->addr = addr;
266 	dma->size = size;
267 	dma->type = map_type;
268 }
269 
270 static inline
mlx5e_skb_fifo_get(struct mlx5e_skb_fifo * fifo,u16 i)271 struct sk_buff **mlx5e_skb_fifo_get(struct mlx5e_skb_fifo *fifo, u16 i)
272 {
273 	return &fifo->fifo[i & fifo->mask];
274 }
275 
276 static inline
mlx5e_skb_fifo_push(struct mlx5e_skb_fifo * fifo,struct sk_buff * skb)277 void mlx5e_skb_fifo_push(struct mlx5e_skb_fifo *fifo, struct sk_buff *skb)
278 {
279 	struct sk_buff **skb_item = mlx5e_skb_fifo_get(fifo, (*fifo->pc)++);
280 
281 	*skb_item = skb;
282 }
283 
284 static inline
mlx5e_skb_fifo_pop(struct mlx5e_skb_fifo * fifo)285 struct sk_buff *mlx5e_skb_fifo_pop(struct mlx5e_skb_fifo *fifo)
286 {
287 	return *mlx5e_skb_fifo_get(fifo, (*fifo->cc)++);
288 }
289 
290 static inline void
mlx5e_tx_dma_unmap(struct device * pdev,struct mlx5e_sq_dma * dma)291 mlx5e_tx_dma_unmap(struct device *pdev, struct mlx5e_sq_dma *dma)
292 {
293 	switch (dma->type) {
294 	case MLX5E_DMA_MAP_SINGLE:
295 		dma_unmap_single(pdev, dma->addr, dma->size, DMA_TO_DEVICE);
296 		break;
297 	case MLX5E_DMA_MAP_PAGE:
298 		dma_unmap_page(pdev, dma->addr, dma->size, DMA_TO_DEVICE);
299 		break;
300 	default:
301 		WARN_ONCE(true, "mlx5e_tx_dma_unmap unknown DMA type!\n");
302 	}
303 }
304 
305 void mlx5e_sq_xmit_simple(struct mlx5e_txqsq *sq, struct sk_buff *skb, bool xmit_more);
306 void mlx5e_tx_mpwqe_ensure_complete(struct mlx5e_txqsq *sq);
307 
mlx5e_tx_mpwqe_is_full(struct mlx5e_tx_mpwqe * session)308 static inline bool mlx5e_tx_mpwqe_is_full(struct mlx5e_tx_mpwqe *session)
309 {
310 	return session->ds_count == MLX5E_TX_MPW_MAX_NUM_DS;
311 }
312 
mlx5e_rqwq_reset(struct mlx5e_rq * rq)313 static inline void mlx5e_rqwq_reset(struct mlx5e_rq *rq)
314 {
315 	if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
316 		mlx5_wq_ll_reset(&rq->mpwqe.wq);
317 		rq->mpwqe.actual_wq_head = 0;
318 	} else {
319 		mlx5_wq_cyc_reset(&rq->wqe.wq);
320 	}
321 }
322 
mlx5e_dump_error_cqe(struct mlx5e_cq * cq,u32 qn,struct mlx5_err_cqe * err_cqe)323 static inline void mlx5e_dump_error_cqe(struct mlx5e_cq *cq, u32 qn,
324 					struct mlx5_err_cqe *err_cqe)
325 {
326 	struct mlx5_cqwq *wq = &cq->wq;
327 	u32 ci;
328 
329 	ci = mlx5_cqwq_ctr2ix(wq, wq->cc - 1);
330 
331 	netdev_err(cq->netdev,
332 		   "Error cqe on cqn 0x%x, ci 0x%x, qn 0x%x, opcode 0x%x, syndrome 0x%x, vendor syndrome 0x%x\n",
333 		   cq->mcq.cqn, ci, qn,
334 		   get_cqe_opcode((struct mlx5_cqe64 *)err_cqe),
335 		   err_cqe->syndrome, err_cqe->vendor_err_synd);
336 	mlx5_dump_err_cqe(cq->mdev, err_cqe);
337 }
338 
mlx5e_rqwq_get_size(struct mlx5e_rq * rq)339 static inline u32 mlx5e_rqwq_get_size(struct mlx5e_rq *rq)
340 {
341 	switch (rq->wq_type) {
342 	case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
343 		return mlx5_wq_ll_get_size(&rq->mpwqe.wq);
344 	default:
345 		return mlx5_wq_cyc_get_size(&rq->wqe.wq);
346 	}
347 }
348 
mlx5e_rqwq_get_cur_sz(struct mlx5e_rq * rq)349 static inline u32 mlx5e_rqwq_get_cur_sz(struct mlx5e_rq *rq)
350 {
351 	switch (rq->wq_type) {
352 	case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
353 		return rq->mpwqe.wq.cur_sz;
354 	default:
355 		return rq->wqe.wq.cur_sz;
356 	}
357 }
358 
mlx5e_rqwq_get_head(struct mlx5e_rq * rq)359 static inline u16 mlx5e_rqwq_get_head(struct mlx5e_rq *rq)
360 {
361 	switch (rq->wq_type) {
362 	case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
363 		return mlx5_wq_ll_get_head(&rq->mpwqe.wq);
364 	default:
365 		return mlx5_wq_cyc_get_head(&rq->wqe.wq);
366 	}
367 }
368 
mlx5e_rqwq_get_wqe_counter(struct mlx5e_rq * rq)369 static inline u16 mlx5e_rqwq_get_wqe_counter(struct mlx5e_rq *rq)
370 {
371 	switch (rq->wq_type) {
372 	case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
373 		return mlx5_wq_ll_get_counter(&rq->mpwqe.wq);
374 	default:
375 		return mlx5_wq_cyc_get_counter(&rq->wqe.wq);
376 	}
377 }
378 
379 /* SW parser related functions */
380 
381 struct mlx5e_swp_spec {
382 	__be16 l3_proto;
383 	u8 l4_proto;
384 	u8 is_tun;
385 	__be16 tun_l3_proto;
386 	u8 tun_l4_proto;
387 };
388 
mlx5e_eseg_swp_offsets_add_vlan(struct mlx5_wqe_eth_seg * eseg)389 static inline void mlx5e_eseg_swp_offsets_add_vlan(struct mlx5_wqe_eth_seg *eseg)
390 {
391 	/* SWP offsets are in 2-bytes words */
392 	eseg->swp_outer_l3_offset += VLAN_HLEN / 2;
393 	eseg->swp_outer_l4_offset += VLAN_HLEN / 2;
394 	eseg->swp_inner_l3_offset += VLAN_HLEN / 2;
395 	eseg->swp_inner_l4_offset += VLAN_HLEN / 2;
396 }
397 
398 static inline void
mlx5e_set_eseg_swp(struct sk_buff * skb,struct mlx5_wqe_eth_seg * eseg,struct mlx5e_swp_spec * swp_spec)399 mlx5e_set_eseg_swp(struct sk_buff *skb, struct mlx5_wqe_eth_seg *eseg,
400 		   struct mlx5e_swp_spec *swp_spec)
401 {
402 	/* SWP offsets are in 2-bytes words */
403 	eseg->swp_outer_l3_offset = skb_network_offset(skb) / 2;
404 	if (swp_spec->l3_proto == htons(ETH_P_IPV6))
405 		eseg->swp_flags |= MLX5_ETH_WQE_SWP_OUTER_L3_IPV6;
406 	if (swp_spec->l4_proto) {
407 		eseg->swp_outer_l4_offset = skb_transport_offset(skb) / 2;
408 		if (swp_spec->l4_proto == IPPROTO_UDP)
409 			eseg->swp_flags |= MLX5_ETH_WQE_SWP_OUTER_L4_UDP;
410 	}
411 
412 	if (swp_spec->is_tun) {
413 		eseg->swp_inner_l3_offset = skb_inner_network_offset(skb) / 2;
414 		if (swp_spec->tun_l3_proto == htons(ETH_P_IPV6))
415 			eseg->swp_flags |= MLX5_ETH_WQE_SWP_INNER_L3_IPV6;
416 	} else { /* typically for ipsec when xfrm mode != XFRM_MODE_TUNNEL */
417 		eseg->swp_inner_l3_offset = skb_network_offset(skb) / 2;
418 		if (swp_spec->l3_proto == htons(ETH_P_IPV6))
419 			eseg->swp_flags |= MLX5_ETH_WQE_SWP_INNER_L3_IPV6;
420 	}
421 	switch (swp_spec->tun_l4_proto) {
422 	case IPPROTO_UDP:
423 		eseg->swp_flags |= MLX5_ETH_WQE_SWP_INNER_L4_UDP;
424 		fallthrough;
425 	case IPPROTO_TCP:
426 		eseg->swp_inner_l4_offset = skb_inner_transport_offset(skb) / 2;
427 		break;
428 	}
429 }
430 
mlx5e_stop_room_for_wqe(u16 wqe_size)431 static inline u16 mlx5e_stop_room_for_wqe(u16 wqe_size)
432 {
433 	BUILD_BUG_ON(PAGE_SIZE / MLX5_SEND_WQE_BB < MLX5_SEND_WQE_MAX_WQEBBS);
434 
435 	/* A WQE must not cross the page boundary, hence two conditions:
436 	 * 1. Its size must not exceed the page size.
437 	 * 2. If the WQE size is X, and the space remaining in a page is less
438 	 *    than X, this space needs to be padded with NOPs. So, one WQE of
439 	 *    size X may require up to X-1 WQEBBs of padding, which makes the
440 	 *    stop room of X-1 + X.
441 	 * WQE size is also limited by the hardware limit.
442 	 */
443 
444 	if (__builtin_constant_p(wqe_size))
445 		BUILD_BUG_ON(wqe_size > MLX5_SEND_WQE_MAX_WQEBBS);
446 	else
447 		WARN_ON_ONCE(wqe_size > MLX5_SEND_WQE_MAX_WQEBBS);
448 
449 	return wqe_size * 2 - 1;
450 }
451 
mlx5e_icosq_can_post_wqe(struct mlx5e_icosq * sq,u16 wqe_size)452 static inline bool mlx5e_icosq_can_post_wqe(struct mlx5e_icosq *sq, u16 wqe_size)
453 {
454 	u16 room = sq->reserved_room + mlx5e_stop_room_for_wqe(wqe_size);
455 
456 	return mlx5e_wqc_has_room_for(&sq->wq, sq->cc, sq->pc, room);
457 }
458 #endif
459