xref: /dpdk/drivers/net/mlx5/windows/mlx5_os.c (revision c9902a15bd005b6d4fe072cf7b60fe4ee679155f)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2020 Mellanox Technologies, Ltd
3  */
4 
5 #include <errno.h>
6 #include <stdalign.h>
7 #include <stddef.h>
8 #include <stdint.h>
9 #include <stdlib.h>
10 
11 #include <rte_windows.h>
12 #include <ethdev_pci.h>
13 
14 #include <mlx5_glue.h>
15 #include <mlx5_devx_cmds.h>
16 #include <mlx5_common.h>
17 #include <mlx5_common_mp.h>
18 #include <mlx5_common_mr.h>
19 #include <mlx5_malloc.h>
20 
21 #include "mlx5_defs.h"
22 #include "mlx5.h"
23 #include "mlx5_common_os.h"
24 #include "mlx5_utils.h"
25 #include "mlx5_rxtx.h"
26 #include "mlx5_rx.h"
27 #include "mlx5_tx.h"
28 #include "mlx5_autoconf.h"
29 #include "mlx5_mr.h"
30 #include "mlx5_flow.h"
31 #include "mlx5_devx.h"
32 
33 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
34 
35 /* Spinlock for mlx5_shared_data allocation. */
36 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
37 
38 /* rte flow indexed pool configuration. */
39 static struct mlx5_indexed_pool_config icfg[] = {
40 	{
41 		.size = sizeof(struct rte_flow),
42 		.trunk_size = 64,
43 		.need_lock = 1,
44 		.release_mem_en = 0,
45 		.malloc = mlx5_malloc,
46 		.free = mlx5_free,
47 		.per_core_cache = 0,
48 		.type = "ctl_flow_ipool",
49 	},
50 	{
51 		.size = sizeof(struct rte_flow),
52 		.trunk_size = 64,
53 		.grow_trunk = 3,
54 		.grow_shift = 2,
55 		.need_lock = 1,
56 		.release_mem_en = 0,
57 		.malloc = mlx5_malloc,
58 		.free = mlx5_free,
59 		.per_core_cache = 1 << 14,
60 		.type = "rte_flow_ipool",
61 	},
62 	{
63 		.size = sizeof(struct rte_flow),
64 		.trunk_size = 64,
65 		.grow_trunk = 3,
66 		.grow_shift = 2,
67 		.need_lock = 1,
68 		.release_mem_en = 0,
69 		.malloc = mlx5_malloc,
70 		.free = mlx5_free,
71 		.per_core_cache = 0,
72 		.type = "mcp_flow_ipool",
73 	},
74 };
75 
76 /**
77  * Initialize shared data between primary and secondary process.
78  *
79  * A memzone is reserved by primary process and secondary processes attach to
80  * the memzone.
81  *
82  * @return
83  *   0 on success, a negative errno value otherwise and rte_errno is set.
84  */
85 static int
86 mlx5_init_shared_data(void)
87 {
88 	const struct rte_memzone *mz;
89 	int ret = 0;
90 
91 	rte_spinlock_lock(&mlx5_shared_data_lock);
92 	if (mlx5_shared_data == NULL) {
93 		/* Allocate shared memory. */
94 		mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
95 					 sizeof(*mlx5_shared_data),
96 					 SOCKET_ID_ANY, 0);
97 		if (mz == NULL) {
98 			DRV_LOG(ERR,
99 				"Cannot allocate mlx5 shared data");
100 			ret = -rte_errno;
101 			goto error;
102 		}
103 		mlx5_shared_data = mz->addr;
104 		memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
105 		rte_spinlock_init(&mlx5_shared_data->lock);
106 	}
107 error:
108 	rte_spinlock_unlock(&mlx5_shared_data_lock);
109 	return ret;
110 }
111 
112 /**
113  * PMD global initialization.
114  *
115  * Independent from individual device, this function initializes global
116  * per-PMD data structures distinguishing primary and secondary processes.
117  * Hence, each initialization is called once per a process.
118  *
119  * @return
120  *   0 on success, a negative errno value otherwise and rte_errno is set.
121  */
122 static int
123 mlx5_init_once(void)
124 {
125 	if (mlx5_init_shared_data())
126 		return -rte_errno;
127 	return 0;
128 }
129 
130 /**
131  * Get mlx5 device attributes.
132  *
133  * @param ctx
134  *   Pointer to device context.
135  *
136  * @param device_attr
137  *   Pointer to mlx5 device attributes.
138  *
139  * @return
140  *   0 on success, non zero error number otherwise
141  */
142 int
143 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
144 {
145 	struct mlx5_context *mlx5_ctx;
146 	struct mlx5_hca_attr hca_attr;
147 	void *pv_iseg = NULL;
148 	u32 cb_iseg = 0;
149 	int err = 0;
150 
151 	if (!ctx)
152 		return -EINVAL;
153 	mlx5_ctx = (struct mlx5_context *)ctx;
154 	memset(device_attr, 0, sizeof(*device_attr));
155 	err = mlx5_devx_cmd_query_hca_attr(mlx5_ctx, &hca_attr);
156 	if (err) {
157 		DRV_LOG(ERR, "Failed to get device hca_cap");
158 		return err;
159 	}
160 	device_attr->max_cq = 1 << hca_attr.log_max_cq;
161 	device_attr->max_qp = 1 << hca_attr.log_max_qp;
162 	device_attr->max_qp_wr = 1 << hca_attr.log_max_qp_sz;
163 	device_attr->max_cqe = 1 << hca_attr.log_max_cq_sz;
164 	device_attr->max_mr = 1 << hca_attr.log_max_mrw_sz;
165 	device_attr->max_pd = 1 << hca_attr.log_max_pd;
166 	device_attr->max_srq = 1 << hca_attr.log_max_srq;
167 	device_attr->max_srq_wr = 1 << hca_attr.log_max_srq_sz;
168 	if (hca_attr.rss_ind_tbl_cap) {
169 		device_attr->max_rwq_indirection_table_size =
170 			1 << hca_attr.rss_ind_tbl_cap;
171 	}
172 	pv_iseg = mlx5_glue->query_hca_iseg(mlx5_ctx, &cb_iseg);
173 	if (pv_iseg == NULL) {
174 		DRV_LOG(ERR, "Failed to get device hca_iseg");
175 		return errno;
176 	}
177 	if (!err) {
178 		snprintf(device_attr->fw_ver, 64, "%x.%x.%04x",
179 			MLX5_GET(initial_seg, pv_iseg, fw_rev_major),
180 			MLX5_GET(initial_seg, pv_iseg, fw_rev_minor),
181 			MLX5_GET(initial_seg, pv_iseg, fw_rev_subminor));
182 	}
183 	return err;
184 }
185 
186 /**
187  * Initialize DR related data within private structure.
188  * Routine checks the reference counter and does actual
189  * resources creation/initialization only if counter is zero.
190  *
191  * @param[in] priv
192  *   Pointer to the private device data structure.
193  *
194  * @return
195  *   Zero on success, positive error code otherwise.
196  */
197 static int
198 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
199 {
200 	struct mlx5_dev_ctx_shared *sh = priv->sh;
201 	int err = 0;
202 
203 	if (!sh->flow_tbls)
204 		err = mlx5_alloc_table_hash_list(priv);
205 	else
206 		DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse",
207 			(void *)sh->flow_tbls);
208 	return err;
209 }
210 /**
211  * Destroy DR related data within private structure.
212  *
213  * @param[in] priv
214  *   Pointer to the private device data structure.
215  */
216 void
217 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
218 {
219 	mlx5_free_table_hash_list(priv);
220 }
221 
222 /**
223  * Set the completion channel file descriptor interrupt as non-blocking.
224  * Currently it has no support under Windows.
225  *
226  * @param[in] rxq_obj
227  *   Pointer to RQ channel object, which includes the channel fd
228  *
229  * @param[out] fd
230  *   The file descriptor (representing the intetrrupt) used in this channel.
231  *
232  * @return
233  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
234  */
235 int
236 mlx5_os_set_nonblock_channel_fd(int fd)
237 {
238 	(void)fd;
239 	DRV_LOG(WARNING, "%s: is not supported", __func__);
240 	return -ENOTSUP;
241 }
242 
243 /**
244  * Function API open device under Windows
245  *
246  * This function calls the Windows glue APIs to open a device.
247  *
248  * @param[in] spawn
249  *   Pointer to the device attributes (name, port, etc).
250  * @param[out] config
251  *   Pointer to device configuration structure.
252  * @param[out] sh
253  *   Pointer to shared context structure.
254  *
255  * @return
256  *   0 on success, a positive error value otherwise.
257  */
258 int
259 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
260 		 const struct mlx5_dev_config *config,
261 		 struct mlx5_dev_ctx_shared *sh)
262 {
263 	RTE_SET_USED(config);
264 	int err = 0;
265 	struct mlx5_context *mlx5_ctx;
266 
267 	pthread_mutex_init(&sh->txpp.mutex, NULL);
268 	/* Set numa node from pci probe */
269 	sh->numa_node = spawn->pci_dev->device.numa_node;
270 
271 	/* Try to open device with DevX */
272 	rte_errno = 0;
273 	sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
274 	if (!sh->ctx) {
275 		DRV_LOG(ERR, "open_device failed");
276 		err = errno;
277 		return err;
278 	}
279 	sh->devx = 1;
280 	mlx5_ctx = (struct mlx5_context *)sh->ctx;
281 	err = mlx5_glue->query_device(spawn->phys_dev, &mlx5_ctx->mlx5_dev);
282 	if (err)
283 		DRV_LOG(ERR, "Failed to query device context fields.");
284 	return err;
285 }
286 
287 /**
288  * DV flow counter mode detect and config.
289  *
290  * @param dev
291  *   Pointer to rte_eth_dev structure.
292  *
293  */
294 static void
295 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
296 {
297 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
298 	struct mlx5_priv *priv = dev->data->dev_private;
299 	struct mlx5_dev_ctx_shared *sh = priv->sh;
300 	bool fallback;
301 
302 #ifndef HAVE_IBV_DEVX_ASYNC
303 	fallback = true;
304 #else
305 	fallback = false;
306 	if (!priv->config.devx || !priv->config.dv_flow_en ||
307 	    !priv->config.hca_attr.flow_counters_dump ||
308 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
309 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
310 		fallback = true;
311 #endif
312 	if (fallback)
313 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
314 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
315 			priv->config.hca_attr.flow_counters_dump,
316 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
317 	/* Initialize fallback mode only on the port initializes sh. */
318 	if (sh->refcnt == 1)
319 		sh->cmng.counter_fallback = fallback;
320 	else if (fallback != sh->cmng.counter_fallback)
321 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
322 			"with others:%d.", PORT_ID(priv), fallback);
323 #endif
324 }
325 
326 /**
327  * Spawn an Ethernet device from Verbs information.
328  *
329  * @param dpdk_dev
330  *   Backing DPDK device.
331  * @param spawn
332  *   Verbs device parameters (name, port, switch_info) to spawn.
333  * @param config
334  *   Device configuration parameters.
335  *
336  * @return
337  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
338  *   is set. The following errors are defined:
339  *
340  *   EEXIST: device is already spawned
341  */
342 static struct rte_eth_dev *
343 mlx5_dev_spawn(struct rte_device *dpdk_dev,
344 	       struct mlx5_dev_spawn_data *spawn,
345 	       struct mlx5_dev_config *config)
346 {
347 	const struct mlx5_switch_info *switch_info = &spawn->info;
348 	struct mlx5_dev_ctx_shared *sh = NULL;
349 	struct mlx5_dev_attr device_attr;
350 	struct rte_eth_dev *eth_dev = NULL;
351 	struct mlx5_priv *priv = NULL;
352 	int err = 0;
353 	unsigned int cqe_comp;
354 	struct rte_ether_addr mac;
355 	char name[RTE_ETH_NAME_MAX_LEN];
356 	int own_domain_id = 0;
357 	uint16_t port_id;
358 	int i;
359 
360 	/* Build device name. */
361 	strlcpy(name, dpdk_dev->name, sizeof(name));
362 	/* check if the device is already spawned */
363 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
364 		rte_errno = EEXIST;
365 		return NULL;
366 	}
367 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
368 	/*
369 	 * Some parameters are needed in advance to create device context. We
370 	 * process the devargs here to get ones, and later process devargs
371 	 * again to override some hardware settings.
372 	 */
373 	err = mlx5_args(config, dpdk_dev->devargs);
374 	if (err) {
375 		err = rte_errno;
376 		DRV_LOG(ERR, "failed to process device arguments: %s",
377 			strerror(rte_errno));
378 		goto error;
379 	}
380 	mlx5_malloc_mem_select(config->sys_mem_en);
381 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
382 	if (!sh)
383 		return NULL;
384 	config->devx = sh->devx;
385 	/* Initialize the shutdown event in mlx5_dev_spawn to
386 	 * support mlx5_is_removed for Windows.
387 	 */
388 	err = mlx5_glue->devx_init_showdown_event(sh->ctx);
389 	if (err) {
390 		DRV_LOG(ERR, "failed to init showdown event: %s",
391 			strerror(errno));
392 		goto error;
393 	}
394 	DRV_LOG(DEBUG, "MPW isn't supported");
395 	mlx5_os_get_dev_attr(sh->ctx, &device_attr);
396 	config->swp = 0;
397 	config->ind_table_max_size =
398 		sh->device_attr.max_rwq_indirection_table_size;
399 	cqe_comp = 0;
400 	config->cqe_comp = cqe_comp;
401 	DRV_LOG(DEBUG, "tunnel offloading is not supported");
402 	config->tunnel_en = 0;
403 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is no supported");
404 	config->mpls_en = 0;
405 	/* Allocate private eth device data. */
406 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
407 			   sizeof(*priv),
408 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
409 	if (priv == NULL) {
410 		DRV_LOG(ERR, "priv allocation failure");
411 		err = ENOMEM;
412 		goto error;
413 	}
414 	priv->sh = sh;
415 	priv->dev_port = spawn->phys_port;
416 	priv->pci_dev = spawn->pci_dev;
417 	priv->mtu = RTE_ETHER_MTU;
418 	priv->mp_id.port_id = port_id;
419 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
420 	priv->representor = !!switch_info->representor;
421 	priv->master = !!switch_info->master;
422 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
423 	priv->vport_meta_tag = 0;
424 	priv->vport_meta_mask = 0;
425 	priv->pf_bond = spawn->pf_bond;
426 	priv->vport_id = -1;
427 	/* representor_id field keeps the unmodified VF index. */
428 	priv->representor_id = -1;
429 	/*
430 	 * Look for sibling devices in order to reuse their switch domain
431 	 * if any, otherwise allocate one.
432 	 */
433 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
434 		const struct mlx5_priv *opriv =
435 			rte_eth_devices[port_id].data->dev_private;
436 
437 		if (!opriv ||
438 		    opriv->sh != priv->sh ||
439 			opriv->domain_id ==
440 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
441 			continue;
442 		priv->domain_id = opriv->domain_id;
443 		break;
444 	}
445 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
446 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
447 		if (err) {
448 			err = rte_errno;
449 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
450 				strerror(rte_errno));
451 			goto error;
452 		}
453 		own_domain_id = 1;
454 	}
455 	/* Override some values set by hardware configuration. */
456 	mlx5_args(config, dpdk_dev->devargs);
457 	err = mlx5_dev_check_sibling_config(priv, config, dpdk_dev);
458 	if (err)
459 		goto error;
460 	DRV_LOG(DEBUG, "counters are not supported");
461 	config->ind_table_max_size =
462 		sh->device_attr.max_rwq_indirection_table_size;
463 	/*
464 	 * Remove this check once DPDK supports larger/variable
465 	 * indirection tables.
466 	 */
467 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
468 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
469 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
470 		config->ind_table_max_size);
471 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
472 		(config->hw_vlan_strip ? "" : "not "));
473 	if (config->hw_padding) {
474 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
475 		config->hw_padding = 0;
476 	}
477 	if (config->tso)
478 		config->tso_max_payload_sz = sh->device_attr.max_tso;
479 	DRV_LOG(DEBUG, "%sMPS is %s.",
480 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
481 		config->mps == MLX5_MPW ? "legacy " : "",
482 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
483 	if (config->cqe_comp && !cqe_comp) {
484 		DRV_LOG(WARNING, "Rx CQE compression isn't supported.");
485 		config->cqe_comp = 0;
486 	}
487 	if (config->devx) {
488 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
489 		if (err) {
490 			err = -err;
491 			goto error;
492 		}
493 		/* Check relax ordering support. */
494 		sh->cmng.relaxed_ordering_read = 0;
495 		sh->cmng.relaxed_ordering_write = 0;
496 		if (!haswell_broadwell_cpu) {
497 			sh->cmng.relaxed_ordering_write =
498 				config->hca_attr.relaxed_ordering_write;
499 			sh->cmng.relaxed_ordering_read =
500 				config->hca_attr.relaxed_ordering_read;
501 		}
502 		config->hw_csum = config->hca_attr.csum_cap;
503 		DRV_LOG(DEBUG, "checksum offloading is %ssupported",
504 		    (config->hw_csum ? "" : "not "));
505 	}
506 	if (config->devx) {
507 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
508 
509 		err = config->hca_attr.access_register_user ?
510 			mlx5_devx_cmd_register_read
511 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
512 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
513 		if (!err) {
514 			uint32_t ts_mode;
515 
516 			/* MTUTC register is read successfully. */
517 			ts_mode = MLX5_GET(register_mtutc, reg,
518 					   time_stamp_mode);
519 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
520 				config->rt_timestamp = 1;
521 		} else {
522 			/* Kernel does not support register reading. */
523 			if (config->hca_attr.dev_freq_khz ==
524 						 (NS_PER_S / MS_PER_S))
525 				config->rt_timestamp = 1;
526 		}
527 		sh->rq_ts_format = config->hca_attr.rq_ts_format;
528 		sh->sq_ts_format = config->hca_attr.sq_ts_format;
529 		sh->qp_ts_format = config->hca_attr.qp_ts_format;
530 	}
531 	if (config->mprq.enabled) {
532 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
533 		config->mprq.enabled = 0;
534 	}
535 	if (config->max_dump_files_num == 0)
536 		config->max_dump_files_num = 128;
537 	eth_dev = rte_eth_dev_allocate(name);
538 	if (eth_dev == NULL) {
539 		DRV_LOG(ERR, "can not allocate rte ethdev");
540 		err = ENOMEM;
541 		goto error;
542 	}
543 	if (priv->representor) {
544 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
545 		eth_dev->data->representor_id = priv->representor_id;
546 	}
547 	/*
548 	 * Store associated network device interface index. This index
549 	 * is permanent throughout the lifetime of device. So, we may store
550 	 * the ifindex here and use the cached value further.
551 	 */
552 	MLX5_ASSERT(spawn->ifindex);
553 	priv->if_index = spawn->ifindex;
554 	eth_dev->data->dev_private = priv;
555 	priv->dev_data = eth_dev->data;
556 	eth_dev->data->mac_addrs = priv->mac;
557 	eth_dev->device = dpdk_dev;
558 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
559 	/* Configure the first MAC address by default. */
560 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
561 		DRV_LOG(ERR,
562 			"port %u cannot get MAC address, is mlx5_en"
563 			" loaded? (errno: %s).",
564 			eth_dev->data->port_id, strerror(rte_errno));
565 		err = ENODEV;
566 		goto error;
567 	}
568 	DRV_LOG(INFO,
569 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
570 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
571 #ifdef RTE_LIBRTE_MLX5_DEBUG
572 	{
573 		char ifname[MLX5_NAMESIZE];
574 
575 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
576 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
577 				eth_dev->data->port_id, ifname);
578 		else
579 			DRV_LOG(DEBUG, "port %u ifname is unknown.",
580 				eth_dev->data->port_id);
581 	}
582 #endif
583 	/* Get actual MTU if possible. */
584 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
585 	if (err) {
586 		err = rte_errno;
587 		goto error;
588 	}
589 	DRV_LOG(DEBUG, "port %u MTU is %u.", eth_dev->data->port_id,
590 		priv->mtu);
591 	/* Initialize burst functions to prevent crashes before link-up. */
592 	eth_dev->rx_pkt_burst = removed_rx_burst;
593 	eth_dev->tx_pkt_burst = removed_tx_burst;
594 	eth_dev->dev_ops = &mlx5_dev_ops;
595 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
596 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
597 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
598 	/* Register MAC address. */
599 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
600 	priv->ctrl_flows = 0;
601 	TAILQ_INIT(&priv->flow_meters);
602 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
603 	if (!priv->mtr_profile_tbl)
604 		goto error;
605 	/* Bring Ethernet device up. */
606 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up.",
607 		eth_dev->data->port_id);
608 	/* nl calls are unsupported - set to -1 not to fail on release */
609 	priv->nl_socket_rdma = -1;
610 	priv->nl_socket_route = -1;
611 	mlx5_set_link_up(eth_dev);
612 	/*
613 	 * Even though the interrupt handler is not installed yet,
614 	 * interrupts will still trigger on the async_fd from
615 	 * Verbs context returned by ibv_open_device().
616 	 */
617 	mlx5_link_update(eth_dev, 0);
618 	config->dv_esw_en = 0;
619 	/* Detect minimal data bytes to inline. */
620 	mlx5_set_min_inline(spawn, config);
621 	/* Store device configuration on private structure. */
622 	priv->config = *config;
623 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
624 		icfg[i].release_mem_en = !!config->reclaim_mode;
625 		if (config->reclaim_mode)
626 			icfg[i].per_core_cache = 0;
627 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
628 		if (!priv->flows[i])
629 			goto error;
630 	}
631 	/* Create context for virtual machine VLAN workaround. */
632 	priv->vmwa_context = NULL;
633 	if (config->dv_flow_en) {
634 		err = mlx5_alloc_shared_dr(priv);
635 		if (err)
636 			goto error;
637 	}
638 	/* No supported flow priority number detection. */
639 	priv->config.flow_prio = -1;
640 	if (!priv->config.dv_esw_en &&
641 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
642 		DRV_LOG(WARNING, "metadata mode %u is not supported "
643 				 "(no E-Switch)", priv->config.dv_xmeta_en);
644 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
645 	}
646 	mlx5_set_metadata_mask(eth_dev);
647 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
648 	    !priv->sh->dv_regc0_mask) {
649 		DRV_LOG(ERR, "metadata mode %u is not supported "
650 			     "(no metadata reg_c[0] is available).",
651 			     priv->config.dv_xmeta_en);
652 			err = ENOTSUP;
653 			goto error;
654 	}
655 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
656 		mlx5_hrxq_create_cb, mlx5_hrxq_match_cb,
657 		mlx5_hrxq_remove_cb, mlx5_hrxq_clone_cb,
658 		mlx5_hrxq_clone_free_cb);
659 	/* Query availability of metadata reg_c's. */
660 	err = mlx5_flow_discover_mreg_c(eth_dev);
661 	if (err < 0) {
662 		err = -err;
663 		goto error;
664 	}
665 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
666 		DRV_LOG(DEBUG,
667 			"port %u extensive metadata register is not supported.",
668 			eth_dev->data->port_id);
669 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
670 			DRV_LOG(ERR, "metadata mode %u is not supported "
671 				     "(no metadata registers available).",
672 				     priv->config.dv_xmeta_en);
673 			err = ENOTSUP;
674 			goto error;
675 		}
676 	}
677 	if (config->devx && config->dv_flow_en) {
678 		priv->obj_ops = devx_obj_ops;
679 	} else {
680 		DRV_LOG(ERR, "Flow mode %u is not supported "
681 				"(Windows flow must be DevX with DV flow enabled).",
682 				priv->config.dv_flow_en);
683 		err = ENOTSUP;
684 		goto error;
685 	}
686 	mlx5_flow_counter_mode_config(eth_dev);
687 	return eth_dev;
688 error:
689 	if (priv) {
690 		if (priv->mtr_profile_tbl)
691 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
692 		if (own_domain_id)
693 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
694 		mlx5_free(priv);
695 		if (eth_dev != NULL)
696 			eth_dev->data->dev_private = NULL;
697 	}
698 	if (eth_dev != NULL) {
699 		/* mac_addrs must not be freed alone because part of
700 		 * dev_private
701 		 **/
702 		eth_dev->data->mac_addrs = NULL;
703 		rte_eth_dev_release_port(eth_dev);
704 	}
705 	if (sh)
706 		mlx5_free_shared_dev_ctx(sh);
707 	MLX5_ASSERT(err > 0);
708 	rte_errno = err;
709 	return NULL;
710 }
711 
712 /**
713  * This function should share events between multiple ports of single IB
714  * device.  Currently it has no support under Windows.
715  *
716  * @param sh
717  *   Pointer to mlx5_dev_ctx_shared object.
718  */
719 void
720 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
721 {
722 	(void)sh;
723 	DRV_LOG(WARNING, "%s: is not supported", __func__);
724 }
725 
726 /**
727  * This function should share events between multiple ports of single IB
728  * device.  Currently it has no support under Windows.
729  *
730  * @param dev
731  *   Pointer to mlx5_dev_ctx_shared object.
732  */
733 void
734 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
735 {
736 	(void)sh;
737 	DRV_LOG(WARNING, "%s: is not supported", __func__);
738 }
739 
740 /**
741  * Read statistics by a named counter.
742  *
743  * @param[in] priv
744  *   Pointer to the private device data structure.
745  * @param[in] ctr_name
746  *   Pointer to the name of the statistic counter to read
747  * @param[out] stat
748  *   Pointer to read statistic value.
749  * @return
750  *   0 on success and stat is valud, 1 if failed to read the value
751  *   rte_errno is set.
752  *
753  */
754 int
755 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
756 		      uint64_t *stat)
757 {
758 	RTE_SET_USED(priv);
759 	RTE_SET_USED(ctr_name);
760 	RTE_SET_USED(stat);
761 	DRV_LOG(WARNING, "%s: is not supported", __func__);
762 	return -ENOTSUP;
763 }
764 
765 /**
766  * Flush device MAC addresses
767  * Currently it has no support under Windows.
768  *
769  * @param dev
770  *   Pointer to Ethernet device structure.
771  *
772  */
773 void
774 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
775 {
776 	(void)dev;
777 	DRV_LOG(WARNING, "%s: is not supported", __func__);
778 }
779 
780 /**
781  * Remove a MAC address from device
782  * Currently it has no support under Windows.
783  *
784  * @param dev
785  *   Pointer to Ethernet device structure.
786  * @param index
787  *   MAC address index.
788  */
789 void
790 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
791 {
792 	(void)dev;
793 	(void)(index);
794 	DRV_LOG(WARNING, "%s: is not supported", __func__);
795 }
796 
797 /**
798  * Adds a MAC address to the device
799  * Currently it has no support under Windows.
800  *
801  * @param dev
802  *   Pointer to Ethernet device structure.
803  * @param mac_addr
804  *   MAC address to register.
805  * @param index
806  *   MAC address index.
807  *
808  * @return
809  *   0 on success, a negative errno value otherwise
810  */
811 int
812 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
813 		     uint32_t index)
814 {
815 	(void)index;
816 	struct rte_ether_addr lmac;
817 
818 	if (mlx5_get_mac(dev, &lmac.addr_bytes)) {
819 		DRV_LOG(ERR,
820 			"port %u cannot get MAC address, is mlx5_en"
821 			" loaded? (errno: %s)",
822 			dev->data->port_id, strerror(rte_errno));
823 		return rte_errno;
824 	}
825 	if (!rte_is_same_ether_addr(&lmac, mac)) {
826 		DRV_LOG(ERR,
827 			"adding new mac address to device is unsupported");
828 		return -ENOTSUP;
829 	}
830 	return 0;
831 }
832 
833 /**
834  * Modify a VF MAC address
835  * Currently it has no support under Windows.
836  *
837  * @param priv
838  *   Pointer to device private data.
839  * @param mac_addr
840  *   MAC address to modify into.
841  * @param iface_idx
842  *   Net device interface index
843  * @param vf_index
844  *   VF index
845  *
846  * @return
847  *   0 on success, a negative errno value otherwise
848  */
849 int
850 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
851 			   unsigned int iface_idx,
852 			   struct rte_ether_addr *mac_addr,
853 			   int vf_index)
854 {
855 	(void)priv;
856 	(void)iface_idx;
857 	(void)mac_addr;
858 	(void)vf_index;
859 	DRV_LOG(WARNING, "%s: is not supported", __func__);
860 	return -ENOTSUP;
861 }
862 
863 /**
864  * Set device promiscuous mode
865  * Currently it has no support under Windows.
866  *
867  * @param dev
868  *   Pointer to Ethernet device structure.
869  * @param enable
870  *   0 - promiscuous is disabled, otherwise - enabled
871  *
872  * @return
873  *   0 on success, a negative error value otherwise
874  */
875 int
876 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
877 {
878 	(void)dev;
879 	(void)enable;
880 	DRV_LOG(WARNING, "%s: is not supported", __func__);
881 	return -ENOTSUP;
882 }
883 
884 /**
885  * Set device allmulti mode
886  *
887  * @param dev
888  *   Pointer to Ethernet device structure.
889  * @param enable
890  *   0 - all multicase is disabled, otherwise - enabled
891  *
892  * @return
893  *   0 on success, a negative error value otherwise
894  */
895 int
896 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
897 {
898 	(void)dev;
899 	(void)enable;
900 	DRV_LOG(WARNING, "%s: is not supported", __func__);
901 	return -ENOTSUP;
902 }
903 
904 /**
905  * Detect if a devx_device_bdf object has identical DBDF values to the
906  * rte_pci_addr found in bus/pci probing
907  *
908  * @param[in] devx_bdf
909  *   Pointer to the devx_device_bdf structure.
910  * @param[in] addr
911  *   Pointer to the rte_pci_addr structure.
912  *
913  * @return
914  *   1 on Device match, 0 on mismatch.
915  */
916 static int
917 mlx5_match_devx_bdf_to_addr(struct devx_device_bdf *devx_bdf,
918 			    struct rte_pci_addr *addr)
919 {
920 	if (addr->domain != (devx_bdf->bus_id >> 8) ||
921 	    addr->bus != (devx_bdf->bus_id & 0xff) ||
922 	    addr->devid != devx_bdf->dev_id ||
923 	    addr->function != devx_bdf->fnc_id) {
924 		return 0;
925 	}
926 	return 1;
927 }
928 
929 /**
930  * Detect if a devx_device_bdf object matches the rte_pci_addr
931  * found in bus/pci probing
932  * Compare both the Native/PF BDF and the raw_bdf representing a VF BDF.
933  *
934  * @param[in] devx_bdf
935  *   Pointer to the devx_device_bdf structure.
936  * @param[in] addr
937  *   Pointer to the rte_pci_addr structure.
938  *
939  * @return
940  *   1 on Device match, 0 on mismatch, rte_errno code on failure.
941  */
942 static int
943 mlx5_match_devx_devices_to_addr(struct devx_device_bdf *devx_bdf,
944 				struct rte_pci_addr *addr)
945 {
946 	int err;
947 	struct devx_device mlx5_dev;
948 
949 	if (mlx5_match_devx_bdf_to_addr(devx_bdf, addr))
950 		return 1;
951 	/**
952 	 * Didn't match on Native/PF BDF, could still
953 	 * Match a VF BDF, check it next
954 	 */
955 	err = mlx5_glue->query_device(devx_bdf, &mlx5_dev);
956 	if (err) {
957 		DRV_LOG(ERR, "query_device failed");
958 		rte_errno = err;
959 		return rte_errno;
960 	}
961 	if (mlx5_match_devx_bdf_to_addr(&mlx5_dev.raw_bdf, addr))
962 		return 1;
963 	return 0;
964 }
965 
966 /**
967  * DPDK callback to register a PCI device.
968  *
969  * This function spawns Ethernet devices out of a given device.
970  *
971  * @param[in] dev
972  *   Pointer to the generic device.
973  *
974  * @return
975  *   0 on success, a negative errno value otherwise and rte_errno is set.
976  */
977 int
978 mlx5_os_net_probe(struct rte_device *dev)
979 {
980 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev);
981 	struct devx_device_bdf *devx_bdf_devs, *orig_devx_bdf_devs;
982 	/*
983 	 * Number of found IB Devices matching with requested PCI BDF.
984 	 * nd != 1 means there are multiple IB devices over the same
985 	 * PCI device and we have representors and master.
986 	 */
987 	unsigned int nd = 0;
988 	/*
989 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
990 	 * we have the single multiport IB device, and there may be
991 	 * representors attached to some of found ports.
992 	 * Currently not supported.
993 	 * unsigned int np = 0;
994 	 */
995 
996 	/*
997 	 * Number of DPDK ethernet devices to Spawn - either over
998 	 * multiple IB devices or multiple ports of single IB device.
999 	 * Actually this is the number of iterations to spawn.
1000 	 */
1001 	unsigned int ns = 0;
1002 	/*
1003 	 * Bonding device
1004 	 *   < 0 - no bonding device (single one)
1005 	 *  >= 0 - bonding device (value is slave PF index)
1006 	 */
1007 	int bd = -1;
1008 	struct mlx5_dev_spawn_data *list = NULL;
1009 	struct mlx5_dev_config dev_config;
1010 	unsigned int dev_config_vf;
1011 	int ret, err;
1012 	uint32_t restore;
1013 
1014 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1015 		DRV_LOG(ERR, "Secondary process is not supported on Windows.");
1016 		return -ENOTSUP;
1017 	}
1018 	ret = mlx5_init_once();
1019 	if (ret) {
1020 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1021 			strerror(rte_errno));
1022 		return -rte_errno;
1023 	}
1024 	errno = 0;
1025 	devx_bdf_devs = mlx5_glue->get_device_list(&ret);
1026 	orig_devx_bdf_devs = devx_bdf_devs;
1027 	if (!devx_bdf_devs) {
1028 		rte_errno = errno ? errno : ENOSYS;
1029 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1030 		return -rte_errno;
1031 	}
1032 	/*
1033 	 * First scan the list of all Infiniband devices to find
1034 	 * matching ones, gathering into the list.
1035 	 */
1036 	struct devx_device_bdf *devx_bdf_match[ret + 1];
1037 
1038 	while (ret-- > 0) {
1039 		err = mlx5_match_devx_devices_to_addr(devx_bdf_devs,
1040 		    &pci_dev->addr);
1041 		if (!err) {
1042 			devx_bdf_devs++;
1043 			continue;
1044 		}
1045 		if (err != 1) {
1046 			ret = -err;
1047 			goto exit;
1048 		}
1049 		devx_bdf_match[nd++] = devx_bdf_devs;
1050 	}
1051 	devx_bdf_match[nd] = NULL;
1052 	if (!nd) {
1053 		/* No device matches, just complain and bail out. */
1054 		DRV_LOG(WARNING,
1055 			"no DevX device matches PCI device " PCI_PRI_FMT ","
1056 			" is DevX Configured?",
1057 			pci_dev->addr.domain, pci_dev->addr.bus,
1058 			pci_dev->addr.devid, pci_dev->addr.function);
1059 		rte_errno = ENOENT;
1060 		ret = -rte_errno;
1061 		goto exit;
1062 	}
1063 	/*
1064 	 * Now we can determine the maximal
1065 	 * amount of devices to be spawned.
1066 	 */
1067 	list = mlx5_malloc(MLX5_MEM_ZERO,
1068 			   sizeof(struct mlx5_dev_spawn_data),
1069 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1070 	if (!list) {
1071 		DRV_LOG(ERR, "spawn data array allocation failure");
1072 		rte_errno = ENOMEM;
1073 		ret = -rte_errno;
1074 		goto exit;
1075 	}
1076 	memset(&list[ns].info, 0, sizeof(list[ns].info));
1077 	list[ns].max_port = 1;
1078 	list[ns].phys_port = 1;
1079 	list[ns].phys_dev = devx_bdf_match[ns];
1080 	list[ns].eth_dev = NULL;
1081 	list[ns].pci_dev = pci_dev;
1082 	list[ns].pf_bond = bd;
1083 	list[ns].ifindex = -1; /* Spawn will assign */
1084 	list[ns].info =
1085 		(struct mlx5_switch_info){
1086 			.master = 0,
1087 			.representor = 0,
1088 			.name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK,
1089 			.port_name = 0,
1090 			.switch_id = 0,
1091 		};
1092 	/* Device specific configuration. */
1093 	switch (pci_dev->id.device_id) {
1094 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1095 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1096 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1097 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1098 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1099 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1100 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
1101 		dev_config_vf = 1;
1102 		break;
1103 	default:
1104 		dev_config_vf = 0;
1105 		break;
1106 	}
1107 	/* Default configuration. */
1108 	memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
1109 	dev_config.vf = dev_config_vf;
1110 	dev_config.mps = 0;
1111 	dev_config.dbnc = MLX5_ARG_UNSET;
1112 	dev_config.rx_vec_en = 1;
1113 	dev_config.txq_inline_max = MLX5_ARG_UNSET;
1114 	dev_config.txq_inline_min = MLX5_ARG_UNSET;
1115 	dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
1116 	dev_config.txqs_inline = MLX5_ARG_UNSET;
1117 	dev_config.vf_nl_en = 0;
1118 	dev_config.mr_ext_memseg_en = 1;
1119 	dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
1120 	dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
1121 	dev_config.dv_esw_en = 0;
1122 	dev_config.dv_flow_en = 1;
1123 	dev_config.decap_en = 0;
1124 	dev_config.log_hp_size = MLX5_ARG_UNSET;
1125 	list[ns].numa_node = pci_dev->device.numa_node;
1126 	list[ns].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1127 					  &list[ns],
1128 					  &dev_config);
1129 	if (!list[ns].eth_dev)
1130 		goto exit;
1131 	restore = list[ns].eth_dev->data->dev_flags;
1132 	rte_eth_copy_pci_info(list[ns].eth_dev, pci_dev);
1133 	/* Restore non-PCI flags cleared by the above call. */
1134 	list[ns].eth_dev->data->dev_flags |= restore;
1135 	rte_eth_dev_probing_finish(list[ns].eth_dev);
1136 	ret = 0;
1137 exit:
1138 	/*
1139 	 * Do the routine cleanup:
1140 	 * - free allocated spawn data array
1141 	 * - free the device list
1142 	 */
1143 	if (list)
1144 		mlx5_free(list);
1145 	MLX5_ASSERT(orig_devx_bdf_devs);
1146 	mlx5_glue->free_device_list(orig_devx_bdf_devs);
1147 	return ret;
1148 }
1149 
1150 /**
1151  * Set the reg_mr and dereg_mr call backs
1152  *
1153  * @param reg_mr_cb[out]
1154  *   Pointer to reg_mr func
1155  * @param dereg_mr_cb[out]
1156  *   Pointer to dereg_mr func
1157  *
1158  */
1159 void
1160 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
1161 		      mlx5_dereg_mr_t *dereg_mr_cb)
1162 {
1163 	*reg_mr_cb = mlx5_os_reg_mr;
1164 	*dereg_mr_cb = mlx5_os_dereg_mr;
1165 }
1166 
1167 /**
1168  * Extract pdn of PD object using DevX
1169  *
1170  * @param[in] pd
1171  *   Pointer to the DevX PD object.
1172  * @param[out] pdn
1173  *   Pointer to the PD object number variable.
1174  *
1175  * @return
1176  *   0 on success, error value otherwise.
1177  */
1178 int
1179 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1180 {
1181 	if (!pd)
1182 		return -EINVAL;
1183 
1184 	*pdn = ((struct mlx5_pd *)pd)->pdn;
1185 	return 0;
1186 }
1187 
1188 const struct mlx5_flow_driver_ops mlx5_flow_verbs_drv_ops = {0};
1189