xref: /dpdk/drivers/net/mlx5/mlx5_txpp.c (revision 551c94c83efbce7e860e36963daae8f62f6e87af)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2020 Mellanox Technologies, Ltd
3  */
4 #include <rte_ether.h>
5 #include <rte_ethdev_driver.h>
6 #include <rte_interrupts.h>
7 #include <rte_alarm.h>
8 #include <rte_malloc.h>
9 
10 #include "mlx5.h"
11 #include "mlx5_rxtx.h"
12 #include "mlx5_common_os.h"
13 
14 /* Destroy Event Queue Notification Channel. */
15 static void
16 mlx5_txpp_destroy_eqn(struct mlx5_dev_ctx_shared *sh)
17 {
18 	if (sh->txpp.echan) {
19 		mlx5_glue->devx_destroy_event_channel(sh->txpp.echan);
20 		sh->txpp.echan = NULL;
21 	}
22 	sh->txpp.eqn = 0;
23 }
24 
25 /* Create Event Queue Notification Channel. */
26 static int
27 mlx5_txpp_create_eqn(struct mlx5_dev_ctx_shared *sh)
28 {
29 	uint32_t lcore;
30 
31 	MLX5_ASSERT(!sh->txpp.echan);
32 	lcore = (uint32_t)rte_lcore_to_cpu_id(-1);
33 	if (mlx5_glue->devx_query_eqn(sh->ctx, lcore, &sh->txpp.eqn)) {
34 		rte_errno = errno;
35 		DRV_LOG(ERR, "Failed to query EQ number %d.", rte_errno);
36 		sh->txpp.eqn = 0;
37 		return -rte_errno;
38 	}
39 	sh->txpp.echan = mlx5_glue->devx_create_event_channel(sh->ctx,
40 			MLX5DV_DEVX_CREATE_EVENT_CHANNEL_FLAGS_OMIT_EV_DATA);
41 	if (!sh->txpp.echan) {
42 		sh->txpp.eqn = 0;
43 		rte_errno = errno;
44 		DRV_LOG(ERR, "Failed to create event channel %d.",
45 			rte_errno);
46 		return -rte_errno;
47 	}
48 	return 0;
49 }
50 
51 static void
52 mlx5_txpp_destroy_send_queue(struct mlx5_txpp_wq *wq)
53 {
54 	if (wq->sq)
55 		claim_zero(mlx5_devx_cmd_destroy(wq->sq));
56 	if (wq->sq_umem)
57 		claim_zero(mlx5_glue->devx_umem_dereg(wq->sq_umem));
58 	if (wq->sq_buf)
59 		rte_free((void *)(uintptr_t)wq->sq_buf);
60 	if (wq->cq)
61 		claim_zero(mlx5_devx_cmd_destroy(wq->cq));
62 	if (wq->cq_umem)
63 		claim_zero(mlx5_glue->devx_umem_dereg(wq->cq_umem));
64 	if (wq->cq_buf)
65 		rte_free((void *)(uintptr_t)wq->cq_buf);
66 	memset(wq, 0, sizeof(*wq));
67 }
68 
69 static void
70 mlx5_txpp_destroy_rearm_queue(struct mlx5_dev_ctx_shared *sh)
71 {
72 	struct mlx5_txpp_wq *wq = &sh->txpp.rearm_queue;
73 
74 	mlx5_txpp_destroy_send_queue(wq);
75 }
76 
77 static void
78 mlx5_txpp_destroy_clock_queue(struct mlx5_dev_ctx_shared *sh)
79 {
80 	struct mlx5_txpp_wq *wq = &sh->txpp.clock_queue;
81 
82 	mlx5_txpp_destroy_send_queue(wq);
83 }
84 
85 static void
86 mlx5_txpp_fill_cqe_rearm_queue(struct mlx5_dev_ctx_shared *sh)
87 {
88 	struct mlx5_txpp_wq *wq = &sh->txpp.rearm_queue;
89 	struct mlx5_cqe *cqe = (struct mlx5_cqe *)(uintptr_t)wq->cqes;
90 	uint32_t i;
91 
92 	for (i = 0; i < MLX5_TXPP_REARM_CQ_SIZE; i++) {
93 		cqe->op_own = (MLX5_CQE_INVALID << 4) | MLX5_CQE_OWNER_MASK;
94 		++cqe;
95 	}
96 }
97 
98 static void
99 mlx5_txpp_fill_wqe_rearm_queue(struct mlx5_dev_ctx_shared *sh)
100 {
101 	struct mlx5_txpp_wq *wq = &sh->txpp.rearm_queue;
102 	struct mlx5_wqe *wqe = (struct mlx5_wqe *)(uintptr_t)wq->wqes;
103 	uint32_t i;
104 
105 	for (i = 0; i < wq->sq_size; i += 2) {
106 		struct mlx5_wqe_cseg *cs;
107 		struct mlx5_wqe_qseg *qs;
108 		uint32_t index;
109 
110 		/* Build SEND_EN request with slave WQE index. */
111 		cs = &wqe[i + 0].cseg;
112 		cs->opcode = RTE_BE32(MLX5_OPCODE_SEND_EN | 0);
113 		cs->sq_ds = rte_cpu_to_be_32((wq->sq->id << 8) | 2);
114 		cs->flags = RTE_BE32(MLX5_COMP_ALWAYS <<
115 				     MLX5_COMP_MODE_OFFSET);
116 		cs->misc = RTE_BE32(0);
117 		qs = RTE_PTR_ADD(cs, sizeof(struct mlx5_wqe_cseg));
118 		index = (i * MLX5_TXPP_REARM / 2 + MLX5_TXPP_REARM) &
119 			((1 << MLX5_WQ_INDEX_WIDTH) - 1);
120 		qs->max_index = rte_cpu_to_be_32(index);
121 		qs->qpn_cqn = rte_cpu_to_be_32(sh->txpp.clock_queue.sq->id);
122 		/* Build WAIT request with slave CQE index. */
123 		cs = &wqe[i + 1].cseg;
124 		cs->opcode = RTE_BE32(MLX5_OPCODE_WAIT | 0);
125 		cs->sq_ds = rte_cpu_to_be_32((wq->sq->id << 8) | 2);
126 		cs->flags = RTE_BE32(MLX5_COMP_ONLY_ERR <<
127 				     MLX5_COMP_MODE_OFFSET);
128 		cs->misc = RTE_BE32(0);
129 		qs = RTE_PTR_ADD(cs, sizeof(struct mlx5_wqe_cseg));
130 		index = (i * MLX5_TXPP_REARM / 2 + MLX5_TXPP_REARM / 2) &
131 			((1 << MLX5_CQ_INDEX_WIDTH) - 1);
132 		qs->max_index = rte_cpu_to_be_32(index);
133 		qs->qpn_cqn = rte_cpu_to_be_32(sh->txpp.clock_queue.cq->id);
134 	}
135 }
136 
137 /* Creates the Rearm Queue to fire the requests to Clock Queue in realtime. */
138 static int
139 mlx5_txpp_create_rearm_queue(struct mlx5_dev_ctx_shared *sh)
140 {
141 	struct mlx5_devx_create_sq_attr sq_attr = { 0 };
142 	struct mlx5_devx_modify_sq_attr msq_attr = { 0 };
143 	struct mlx5_devx_cq_attr cq_attr = { 0 };
144 	struct mlx5_txpp_wq *wq = &sh->txpp.rearm_queue;
145 	size_t page_size = sysconf(_SC_PAGESIZE);
146 	uint32_t umem_size, umem_dbrec;
147 	int ret;
148 
149 	/* Allocate memory buffer for CQEs and doorbell record. */
150 	umem_size = sizeof(struct mlx5_cqe) * MLX5_TXPP_REARM_CQ_SIZE;
151 	umem_dbrec = RTE_ALIGN(umem_size, MLX5_DBR_SIZE);
152 	umem_size += MLX5_DBR_SIZE;
153 	wq->cq_buf = rte_zmalloc_socket(__func__, umem_size,
154 					page_size, sh->numa_node);
155 	if (!wq->cq_buf) {
156 		DRV_LOG(ERR, "Failed to allocate memory for Rearm Queue.");
157 		return -ENOMEM;
158 	}
159 	/* Register allocated buffer in user space with DevX. */
160 	wq->cq_umem = mlx5_glue->devx_umem_reg(sh->ctx,
161 					       (void *)(uintptr_t)wq->cq_buf,
162 					       umem_size,
163 					       IBV_ACCESS_LOCAL_WRITE);
164 	if (!wq->cq_umem) {
165 		rte_errno = errno;
166 		DRV_LOG(ERR, "Failed to register umem for Rearm Queue.");
167 		goto error;
168 	}
169 	/* Create completion queue object for Rearm Queue. */
170 	cq_attr.cqe_size = (sizeof(struct mlx5_cqe) == 128) ?
171 			    MLX5_CQE_SIZE_128B : MLX5_CQE_SIZE_64B;
172 	cq_attr.uar_page_id = sh->tx_uar->page_id;
173 	cq_attr.eqn = sh->txpp.eqn;
174 	cq_attr.q_umem_valid = 1;
175 	cq_attr.q_umem_offset = 0;
176 	cq_attr.q_umem_id = mlx5_os_get_umem_id(wq->cq_umem);
177 	cq_attr.db_umem_valid = 1;
178 	cq_attr.db_umem_offset = umem_dbrec;
179 	cq_attr.db_umem_id = mlx5_os_get_umem_id(wq->cq_umem);
180 	cq_attr.log_cq_size = rte_log2_u32(MLX5_TXPP_REARM_CQ_SIZE);
181 	cq_attr.log_page_size = rte_log2_u32(page_size);
182 	wq->cq = mlx5_devx_cmd_create_cq(sh->ctx, &cq_attr);
183 	if (!wq->cq) {
184 		rte_errno = errno;
185 		DRV_LOG(ERR, "Failed to create CQ for Rearm Queue.");
186 		goto error;
187 	}
188 	wq->cq_dbrec = RTE_PTR_ADD(wq->cq_buf, umem_dbrec);
189 	wq->cq_ci = 0;
190 	wq->arm_sn = 0;
191 	/* Mark all CQEs initially as invalid. */
192 	mlx5_txpp_fill_cqe_rearm_queue(sh);
193 	/*
194 	 * Allocate memory buffer for Send Queue WQEs.
195 	 * There should be no WQE leftovers in the cyclic queue.
196 	 */
197 	wq->sq_size = MLX5_TXPP_REARM_SQ_SIZE;
198 	MLX5_ASSERT(wq->sq_size == (1 << log2above(wq->sq_size)));
199 	umem_size =  MLX5_WQE_SIZE * wq->sq_size;
200 	umem_dbrec = RTE_ALIGN(umem_size, MLX5_DBR_SIZE);
201 	umem_size += MLX5_DBR_SIZE;
202 	wq->sq_buf = rte_zmalloc_socket(__func__, umem_size,
203 					page_size, sh->numa_node);
204 	if (!wq->sq_buf) {
205 		DRV_LOG(ERR, "Failed to allocate memory for Rearm Queue.");
206 		rte_errno = ENOMEM;
207 		goto error;
208 	}
209 	/* Register allocated buffer in user space with DevX. */
210 	wq->sq_umem = mlx5_glue->devx_umem_reg(sh->ctx,
211 					       (void *)(uintptr_t)wq->sq_buf,
212 					       umem_size,
213 					       IBV_ACCESS_LOCAL_WRITE);
214 	if (!wq->sq_umem) {
215 		rte_errno = errno;
216 		DRV_LOG(ERR, "Failed to register umem for Rearm Queue.");
217 		goto error;
218 	}
219 	/* Create send queue object for Rearm Queue. */
220 	sq_attr.state = MLX5_SQC_STATE_RST;
221 	sq_attr.tis_lst_sz = 1;
222 	sq_attr.tis_num = sh->tis->id;
223 	sq_attr.cqn = wq->cq->id;
224 	sq_attr.cd_master = 1;
225 	sq_attr.wq_attr.uar_page = sh->tx_uar->page_id;
226 	sq_attr.wq_attr.wq_type = MLX5_WQ_TYPE_CYCLIC;
227 	sq_attr.wq_attr.pd = sh->pdn;
228 	sq_attr.wq_attr.log_wq_stride = rte_log2_u32(MLX5_WQE_SIZE);
229 	sq_attr.wq_attr.log_wq_sz = rte_log2_u32(wq->sq_size);
230 	sq_attr.wq_attr.dbr_umem_valid = 1;
231 	sq_attr.wq_attr.dbr_addr = umem_dbrec;
232 	sq_attr.wq_attr.dbr_umem_id = mlx5_os_get_umem_id(wq->sq_umem);
233 	sq_attr.wq_attr.wq_umem_valid = 1;
234 	sq_attr.wq_attr.wq_umem_id = mlx5_os_get_umem_id(wq->sq_umem);
235 	sq_attr.wq_attr.wq_umem_offset = 0;
236 	wq->sq = mlx5_devx_cmd_create_sq(sh->ctx, &sq_attr);
237 	if (!wq->sq) {
238 		rte_errno = errno;
239 		DRV_LOG(ERR, "Failed to create SQ for Rearm Queue.");
240 		goto error;
241 	}
242 	wq->sq_dbrec = RTE_PTR_ADD(wq->sq_buf, umem_dbrec +
243 				   MLX5_SND_DBR * sizeof(uint32_t));
244 	/* Build the WQEs in the Send Queue before goto Ready state. */
245 	mlx5_txpp_fill_wqe_rearm_queue(sh);
246 	/* Change queue state to ready. */
247 	msq_attr.sq_state = MLX5_SQC_STATE_RST;
248 	msq_attr.state = MLX5_SQC_STATE_RDY;
249 	ret = mlx5_devx_cmd_modify_sq(wq->sq, &msq_attr);
250 	if (ret) {
251 		DRV_LOG(ERR, "Failed to set SQ ready state Rearm Queue.");
252 		goto error;
253 	}
254 	return 0;
255 error:
256 	ret = -rte_errno;
257 	mlx5_txpp_destroy_rearm_queue(sh);
258 	rte_errno = -ret;
259 	return ret;
260 }
261 
262 static void
263 mlx5_txpp_fill_wqe_clock_queue(struct mlx5_dev_ctx_shared *sh)
264 {
265 	struct mlx5_txpp_wq *wq = &sh->txpp.clock_queue;
266 	struct mlx5_wqe *wqe = (struct mlx5_wqe *)(uintptr_t)wq->wqes;
267 	struct mlx5_wqe_cseg *cs = &wqe->cseg;
268 	uint32_t wqe_size, opcode, i;
269 	uint8_t *dst;
270 
271 	/* For test purposes fill the WQ with SEND inline packet. */
272 	if (sh->txpp.test) {
273 		wqe_size = RTE_ALIGN(MLX5_TXPP_TEST_PKT_SIZE +
274 				     MLX5_WQE_CSEG_SIZE +
275 				     2 * MLX5_WQE_ESEG_SIZE -
276 				     MLX5_ESEG_MIN_INLINE_SIZE,
277 				     MLX5_WSEG_SIZE);
278 		opcode = MLX5_OPCODE_SEND;
279 	} else {
280 		wqe_size = MLX5_WSEG_SIZE;
281 		opcode = MLX5_OPCODE_NOP;
282 	}
283 	cs->opcode = rte_cpu_to_be_32(opcode | 0); /* Index is ignored. */
284 	cs->sq_ds = rte_cpu_to_be_32((wq->sq->id << 8) |
285 				     (wqe_size / MLX5_WSEG_SIZE));
286 	cs->flags = RTE_BE32(MLX5_COMP_ALWAYS << MLX5_COMP_MODE_OFFSET);
287 	cs->misc = RTE_BE32(0);
288 	wqe_size = RTE_ALIGN(wqe_size, MLX5_WQE_SIZE);
289 	if (sh->txpp.test) {
290 		struct mlx5_wqe_eseg *es = &wqe->eseg;
291 		struct rte_ether_hdr *eth_hdr;
292 		struct rte_ipv4_hdr *ip_hdr;
293 		struct rte_udp_hdr *udp_hdr;
294 
295 		/* Build the inline test packet pattern. */
296 		MLX5_ASSERT(wqe_size <= MLX5_WQE_SIZE_MAX);
297 		MLX5_ASSERT(MLX5_TXPP_TEST_PKT_SIZE >=
298 				(sizeof(struct rte_ether_hdr) +
299 				 sizeof(struct rte_ipv4_hdr)));
300 		es->flags = 0;
301 		es->cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
302 		es->swp_offs = 0;
303 		es->metadata = 0;
304 		es->swp_flags = 0;
305 		es->mss = 0;
306 		es->inline_hdr_sz = RTE_BE16(MLX5_TXPP_TEST_PKT_SIZE);
307 		/* Build test packet L2 header (Ethernet). */
308 		dst = (uint8_t *)&es->inline_data;
309 		eth_hdr = (struct rte_ether_hdr *)dst;
310 		rte_eth_random_addr(&eth_hdr->d_addr.addr_bytes[0]);
311 		rte_eth_random_addr(&eth_hdr->s_addr.addr_bytes[0]);
312 		eth_hdr->ether_type = rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV4);
313 		/* Build test packet L3 header (IP v4). */
314 		dst += sizeof(struct rte_ether_hdr);
315 		ip_hdr = (struct rte_ipv4_hdr *)dst;
316 		ip_hdr->version_ihl = RTE_IPV4_VHL_DEF;
317 		ip_hdr->type_of_service = 0;
318 		ip_hdr->fragment_offset = 0;
319 		ip_hdr->time_to_live = 64;
320 		ip_hdr->next_proto_id = IPPROTO_UDP;
321 		ip_hdr->packet_id = 0;
322 		ip_hdr->total_length = RTE_BE16(MLX5_TXPP_TEST_PKT_SIZE -
323 						sizeof(struct rte_ether_hdr));
324 		/* use RFC5735 / RFC2544 reserved network test addresses */
325 		ip_hdr->src_addr = RTE_BE32((198U << 24) | (18 << 16) |
326 					    (0 << 8) | 1);
327 		ip_hdr->dst_addr = RTE_BE32((198U << 24) | (18 << 16) |
328 					    (0 << 8) | 2);
329 		if (MLX5_TXPP_TEST_PKT_SIZE <
330 					(sizeof(struct rte_ether_hdr) +
331 					 sizeof(struct rte_ipv4_hdr) +
332 					 sizeof(struct rte_udp_hdr)))
333 			goto wcopy;
334 		/* Build test packet L4 header (UDP). */
335 		dst += sizeof(struct rte_ipv4_hdr);
336 		udp_hdr = (struct rte_udp_hdr *)dst;
337 		udp_hdr->src_port = RTE_BE16(9); /* RFC863 Discard. */
338 		udp_hdr->dst_port = RTE_BE16(9);
339 		udp_hdr->dgram_len = RTE_BE16(MLX5_TXPP_TEST_PKT_SIZE -
340 					      sizeof(struct rte_ether_hdr) -
341 					      sizeof(struct rte_ipv4_hdr));
342 		udp_hdr->dgram_cksum = 0;
343 		/* Fill the test packet data. */
344 		dst += sizeof(struct rte_udp_hdr);
345 		for (i = sizeof(struct rte_ether_hdr) +
346 			sizeof(struct rte_ipv4_hdr) +
347 			sizeof(struct rte_udp_hdr);
348 				i < MLX5_TXPP_TEST_PKT_SIZE; i++)
349 			*dst++ = (uint8_t)(i & 0xFF);
350 	}
351 wcopy:
352 	/* Duplicate the pattern to the next WQEs. */
353 	dst = (uint8_t *)(uintptr_t)wq->sq_buf;
354 	for (i = 1; i < MLX5_TXPP_CLKQ_SIZE; i++) {
355 		dst += wqe_size;
356 		rte_memcpy(dst, (void *)(uintptr_t)wq->sq_buf, wqe_size);
357 	}
358 }
359 
360 /* Creates the Clock Queue for packet pacing, returns zero on success. */
361 static int
362 mlx5_txpp_create_clock_queue(struct mlx5_dev_ctx_shared *sh)
363 {
364 	struct mlx5_devx_create_sq_attr sq_attr = { 0 };
365 	struct mlx5_devx_modify_sq_attr msq_attr = { 0 };
366 	struct mlx5_devx_cq_attr cq_attr = { 0 };
367 	struct mlx5_txpp_wq *wq = &sh->txpp.clock_queue;
368 	size_t page_size = sysconf(_SC_PAGESIZE);
369 	uint32_t umem_size, umem_dbrec;
370 	int ret;
371 
372 	/* Allocate memory buffer for CQEs and doorbell record. */
373 	umem_size = sizeof(struct mlx5_cqe) * MLX5_TXPP_CLKQ_SIZE;
374 	umem_dbrec = RTE_ALIGN(umem_size, MLX5_DBR_SIZE);
375 	umem_size += MLX5_DBR_SIZE;
376 	wq->cq_buf = rte_zmalloc_socket(__func__, umem_size,
377 					page_size, sh->numa_node);
378 	if (!wq->cq_buf) {
379 		DRV_LOG(ERR, "Failed to allocate memory for Clock Queue.");
380 		return -ENOMEM;
381 	}
382 	/* Register allocated buffer in user space with DevX. */
383 	wq->cq_umem = mlx5_glue->devx_umem_reg(sh->ctx,
384 					       (void *)(uintptr_t)wq->cq_buf,
385 					       umem_size,
386 					       IBV_ACCESS_LOCAL_WRITE);
387 	if (!wq->cq_umem) {
388 		rte_errno = errno;
389 		DRV_LOG(ERR, "Failed to register umem for Clock Queue.");
390 		goto error;
391 	}
392 	/* Create completion queue object for Clock Queue. */
393 	cq_attr.cqe_size = (sizeof(struct mlx5_cqe) == 128) ?
394 			    MLX5_CQE_SIZE_128B : MLX5_CQE_SIZE_64B;
395 	cq_attr.use_first_only = 1;
396 	cq_attr.overrun_ignore = 1;
397 	cq_attr.uar_page_id = sh->tx_uar->page_id;
398 	cq_attr.eqn = sh->txpp.eqn;
399 	cq_attr.q_umem_valid = 1;
400 	cq_attr.q_umem_offset = 0;
401 	cq_attr.q_umem_id = wq->cq_umem->umem_id;
402 	cq_attr.db_umem_valid = 1;
403 	cq_attr.db_umem_offset = umem_dbrec;
404 	cq_attr.db_umem_id = wq->cq_umem->umem_id;
405 	cq_attr.log_cq_size = rte_log2_u32(MLX5_TXPP_CLKQ_SIZE);
406 	cq_attr.log_page_size = rte_log2_u32(page_size);
407 	wq->cq = mlx5_devx_cmd_create_cq(sh->ctx, &cq_attr);
408 	if (!wq->cq) {
409 		rte_errno = errno;
410 		DRV_LOG(ERR, "Failed to create CQ for Clock Queue.");
411 		goto error;
412 	}
413 	wq->cq_dbrec = RTE_PTR_ADD(wq->cq_buf, umem_dbrec);
414 	wq->cq_ci = 0;
415 	/* Allocate memory buffer for Send Queue WQEs. */
416 	if (sh->txpp.test) {
417 		wq->sq_size = RTE_ALIGN(MLX5_TXPP_TEST_PKT_SIZE +
418 					MLX5_WQE_CSEG_SIZE +
419 					2 * MLX5_WQE_ESEG_SIZE -
420 					MLX5_ESEG_MIN_INLINE_SIZE,
421 					MLX5_WQE_SIZE) / MLX5_WQE_SIZE;
422 		wq->sq_size *= MLX5_TXPP_CLKQ_SIZE;
423 	} else {
424 		wq->sq_size = MLX5_TXPP_CLKQ_SIZE;
425 	}
426 	/* There should not be WQE leftovers in the cyclic queue. */
427 	MLX5_ASSERT(wq->sq_size == (1 << log2above(wq->sq_size)));
428 	umem_size =  MLX5_WQE_SIZE * wq->sq_size;
429 	umem_dbrec = RTE_ALIGN(umem_size, MLX5_DBR_SIZE);
430 	umem_size += MLX5_DBR_SIZE;
431 	wq->sq_buf = rte_zmalloc_socket(__func__, umem_size,
432 					page_size, sh->numa_node);
433 	if (!wq->sq_buf) {
434 		DRV_LOG(ERR, "Failed to allocate memory for Clock Queue.");
435 		rte_errno = ENOMEM;
436 		goto error;
437 	}
438 	/* Register allocated buffer in user space with DevX. */
439 	wq->sq_umem = mlx5_glue->devx_umem_reg(sh->ctx,
440 					       (void *)(uintptr_t)wq->sq_buf,
441 					       umem_size,
442 					       IBV_ACCESS_LOCAL_WRITE);
443 	if (!wq->sq_umem) {
444 		rte_errno = errno;
445 		DRV_LOG(ERR, "Failed to register umem for Clock Queue.");
446 		goto error;
447 	}
448 	/* Create send queue object for Clock Queue. */
449 	if (sh->txpp.test) {
450 		sq_attr.tis_lst_sz = 1;
451 		sq_attr.tis_num = sh->tis->id;
452 		sq_attr.non_wire = 0;
453 		sq_attr.static_sq_wq = 1;
454 	} else {
455 		sq_attr.non_wire = 1;
456 		sq_attr.static_sq_wq = 1;
457 	}
458 	sq_attr.state = MLX5_SQC_STATE_RST;
459 	sq_attr.cqn = wq->cq->id;
460 	sq_attr.wq_attr.cd_slave = 1;
461 	sq_attr.wq_attr.uar_page = sh->tx_uar->page_id;
462 	sq_attr.wq_attr.wq_type = MLX5_WQ_TYPE_CYCLIC;
463 	sq_attr.wq_attr.pd = sh->pdn;
464 	sq_attr.wq_attr.log_wq_stride = rte_log2_u32(MLX5_WQE_SIZE);
465 	sq_attr.wq_attr.log_wq_sz = rte_log2_u32(wq->sq_size);
466 	sq_attr.wq_attr.dbr_umem_valid = 1;
467 	sq_attr.wq_attr.dbr_addr = umem_dbrec;
468 	sq_attr.wq_attr.dbr_umem_id = wq->sq_umem->umem_id;
469 	sq_attr.wq_attr.wq_umem_valid = 1;
470 	sq_attr.wq_attr.wq_umem_id = wq->sq_umem->umem_id;
471 	/* umem_offset must be zero for static_sq_wq queue. */
472 	sq_attr.wq_attr.wq_umem_offset = 0;
473 	wq->sq = mlx5_devx_cmd_create_sq(sh->ctx, &sq_attr);
474 	if (!wq->sq) {
475 		rte_errno = errno;
476 		DRV_LOG(ERR, "Failed to create SQ for Clock Queue.");
477 		goto error;
478 	}
479 	wq->sq_dbrec = RTE_PTR_ADD(wq->sq_buf, umem_dbrec +
480 				   MLX5_SND_DBR * sizeof(uint32_t));
481 	/* Build the WQEs in the Send Queue before goto Ready state. */
482 	mlx5_txpp_fill_wqe_clock_queue(sh);
483 	/* Change queue state to ready. */
484 	msq_attr.sq_state = MLX5_SQC_STATE_RST;
485 	msq_attr.state = MLX5_SQC_STATE_RDY;
486 	wq->sq_ci = 0;
487 	ret = mlx5_devx_cmd_modify_sq(wq->sq, &msq_attr);
488 	if (ret) {
489 		DRV_LOG(ERR, "Failed to set SQ ready state Clock Queue.");
490 		goto error;
491 	}
492 	return 0;
493 error:
494 	ret = -rte_errno;
495 	mlx5_txpp_destroy_clock_queue(sh);
496 	rte_errno = -ret;
497 	return ret;
498 }
499 
500 /*
501  * The routine initializes the packet pacing infrastructure:
502  * - allocates PP context
503  * - Clock CQ/SQ
504  * - Rearm CQ/SQ
505  * - attaches rearm interrupt handler
506  *
507  * Returns 0 on success, negative otherwise
508  */
509 static int
510 mlx5_txpp_create(struct mlx5_dev_ctx_shared *sh, struct mlx5_priv *priv)
511 {
512 	int tx_pp = priv->config.tx_pp;
513 	int ret;
514 
515 	/* Store the requested pacing parameters. */
516 	sh->txpp.tick = tx_pp >= 0 ? tx_pp : -tx_pp;
517 	sh->txpp.test = !!(tx_pp < 0);
518 	sh->txpp.skew = priv->config.tx_skew;
519 	sh->txpp.freq = priv->config.hca_attr.dev_freq_khz;
520 	ret = mlx5_txpp_create_eqn(sh);
521 	if (ret)
522 		goto exit;
523 	ret = mlx5_txpp_create_clock_queue(sh);
524 	if (ret)
525 		goto exit;
526 	ret = mlx5_txpp_create_rearm_queue(sh);
527 	if (ret)
528 		goto exit;
529 exit:
530 	if (ret) {
531 		mlx5_txpp_destroy_rearm_queue(sh);
532 		mlx5_txpp_destroy_clock_queue(sh);
533 		mlx5_txpp_destroy_eqn(sh);
534 		sh->txpp.tick = 0;
535 		sh->txpp.test = 0;
536 		sh->txpp.skew = 0;
537 	}
538 	return ret;
539 }
540 
541 /*
542  * The routine destroys the packet pacing infrastructure:
543  * - detaches rearm interrupt handler
544  * - Rearm CQ/SQ
545  * - Clock CQ/SQ
546  * - PP context
547  */
548 static void
549 mlx5_txpp_destroy(struct mlx5_dev_ctx_shared *sh)
550 {
551 	mlx5_txpp_destroy_rearm_queue(sh);
552 	mlx5_txpp_destroy_clock_queue(sh);
553 	mlx5_txpp_destroy_eqn(sh);
554 	sh->txpp.tick = 0;
555 	sh->txpp.test = 0;
556 	sh->txpp.skew = 0;
557 }
558 
559 /**
560  * Creates and starts packet pacing infrastructure on specified device.
561  *
562  * @param dev
563  *   Pointer to Ethernet device structure.
564  *
565  * @return
566  *   0 on success, a negative errno value otherwise and rte_errno is set.
567  */
568 int
569 mlx5_txpp_start(struct rte_eth_dev *dev)
570 {
571 	struct mlx5_priv *priv = dev->data->dev_private;
572 	struct mlx5_dev_ctx_shared *sh = priv->sh;
573 	int err = 0;
574 	int ret;
575 
576 	if (!priv->config.tx_pp) {
577 		/* Packet pacing is not requested for the device. */
578 		MLX5_ASSERT(priv->txpp_en == 0);
579 		return 0;
580 	}
581 	if (priv->txpp_en) {
582 		/* Packet pacing is already enabled for the device. */
583 		MLX5_ASSERT(sh->txpp.refcnt);
584 		return 0;
585 	}
586 	if (priv->config.tx_pp > 0) {
587 		ret = rte_mbuf_dynflag_lookup
588 				(RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME, NULL);
589 		if (ret < 0)
590 			return 0;
591 	}
592 	ret = pthread_mutex_lock(&sh->txpp.mutex);
593 	MLX5_ASSERT(!ret);
594 	RTE_SET_USED(ret);
595 	if (sh->txpp.refcnt) {
596 		priv->txpp_en = 1;
597 		++sh->txpp.refcnt;
598 	} else {
599 		err = mlx5_txpp_create(sh, priv);
600 		if (!err) {
601 			MLX5_ASSERT(sh->txpp.tick);
602 			priv->txpp_en = 1;
603 			sh->txpp.refcnt = 1;
604 		} else {
605 			rte_errno = -err;
606 		}
607 	}
608 	ret = pthread_mutex_unlock(&sh->txpp.mutex);
609 	MLX5_ASSERT(!ret);
610 	RTE_SET_USED(ret);
611 	return err;
612 }
613 
614 /**
615  * Stops and destroys packet pacing infrastructure on specified device.
616  *
617  * @param dev
618  *   Pointer to Ethernet device structure.
619  *
620  * @return
621  *   0 on success, a negative errno value otherwise and rte_errno is set.
622  */
623 void
624 mlx5_txpp_stop(struct rte_eth_dev *dev)
625 {
626 	struct mlx5_priv *priv = dev->data->dev_private;
627 	struct mlx5_dev_ctx_shared *sh = priv->sh;
628 	int ret;
629 
630 	if (!priv->txpp_en) {
631 		/* Packet pacing is already disabled for the device. */
632 		return;
633 	}
634 	priv->txpp_en = 0;
635 	ret = pthread_mutex_lock(&sh->txpp.mutex);
636 	MLX5_ASSERT(!ret);
637 	RTE_SET_USED(ret);
638 	MLX5_ASSERT(sh->txpp.refcnt);
639 	if (!sh->txpp.refcnt || --sh->txpp.refcnt)
640 		return;
641 	/* No references any more, do actual destroy. */
642 	mlx5_txpp_destroy(sh);
643 	ret = pthread_mutex_unlock(&sh->txpp.mutex);
644 	MLX5_ASSERT(!ret);
645 	RTE_SET_USED(ret);
646 }
647