xref: /dpdk/drivers/net/mlx5/windows/mlx5_os.c (revision d38febb08d57fec29fed27a2d12a507fc6fcdfa1)
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 %02x:%02x:%02x:%02x:%02x:%02x",
570 		eth_dev->data->port_id,
571 		mac.addr_bytes[0], mac.addr_bytes[1],
572 		mac.addr_bytes[2], mac.addr_bytes[3],
573 		mac.addr_bytes[4], mac.addr_bytes[5]);
574 #ifdef RTE_LIBRTE_MLX5_DEBUG
575 	{
576 		char ifname[MLX5_NAMESIZE];
577 
578 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
579 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
580 				eth_dev->data->port_id, ifname);
581 		else
582 			DRV_LOG(DEBUG, "port %u ifname is unknown.",
583 				eth_dev->data->port_id);
584 	}
585 #endif
586 	/* Get actual MTU if possible. */
587 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
588 	if (err) {
589 		err = rte_errno;
590 		goto error;
591 	}
592 	DRV_LOG(DEBUG, "port %u MTU is %u.", eth_dev->data->port_id,
593 		priv->mtu);
594 	/* Initialize burst functions to prevent crashes before link-up. */
595 	eth_dev->rx_pkt_burst = removed_rx_burst;
596 	eth_dev->tx_pkt_burst = removed_tx_burst;
597 	eth_dev->dev_ops = &mlx5_dev_ops;
598 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
599 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
600 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
601 	/* Register MAC address. */
602 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
603 	priv->ctrl_flows = 0;
604 	TAILQ_INIT(&priv->flow_meters);
605 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
606 	if (!priv->mtr_profile_tbl)
607 		goto error;
608 	/* Bring Ethernet device up. */
609 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up.",
610 		eth_dev->data->port_id);
611 	/* nl calls are unsupported - set to -1 not to fail on release */
612 	priv->nl_socket_rdma = -1;
613 	priv->nl_socket_route = -1;
614 	mlx5_set_link_up(eth_dev);
615 	/*
616 	 * Even though the interrupt handler is not installed yet,
617 	 * interrupts will still trigger on the async_fd from
618 	 * Verbs context returned by ibv_open_device().
619 	 */
620 	mlx5_link_update(eth_dev, 0);
621 	config->dv_esw_en = 0;
622 	/* Detect minimal data bytes to inline. */
623 	mlx5_set_min_inline(spawn, config);
624 	/* Store device configuration on private structure. */
625 	priv->config = *config;
626 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
627 		icfg[i].release_mem_en = !!config->reclaim_mode;
628 		if (config->reclaim_mode)
629 			icfg[i].per_core_cache = 0;
630 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
631 		if (!priv->flows[i])
632 			goto error;
633 	}
634 	/* Create context for virtual machine VLAN workaround. */
635 	priv->vmwa_context = NULL;
636 	if (config->dv_flow_en) {
637 		err = mlx5_alloc_shared_dr(priv);
638 		if (err)
639 			goto error;
640 	}
641 	/* No supported flow priority number detection. */
642 	priv->config.flow_prio = -1;
643 	if (!priv->config.dv_esw_en &&
644 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
645 		DRV_LOG(WARNING, "metadata mode %u is not supported "
646 				 "(no E-Switch)", priv->config.dv_xmeta_en);
647 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
648 	}
649 	mlx5_set_metadata_mask(eth_dev);
650 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
651 	    !priv->sh->dv_regc0_mask) {
652 		DRV_LOG(ERR, "metadata mode %u is not supported "
653 			     "(no metadata reg_c[0] is available).",
654 			     priv->config.dv_xmeta_en);
655 			err = ENOTSUP;
656 			goto error;
657 	}
658 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
659 		mlx5_hrxq_create_cb, mlx5_hrxq_match_cb,
660 		mlx5_hrxq_remove_cb, mlx5_hrxq_clone_cb,
661 		mlx5_hrxq_clone_free_cb);
662 	/* Query availability of metadata reg_c's. */
663 	err = mlx5_flow_discover_mreg_c(eth_dev);
664 	if (err < 0) {
665 		err = -err;
666 		goto error;
667 	}
668 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
669 		DRV_LOG(DEBUG,
670 			"port %u extensive metadata register is not supported.",
671 			eth_dev->data->port_id);
672 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
673 			DRV_LOG(ERR, "metadata mode %u is not supported "
674 				     "(no metadata registers available).",
675 				     priv->config.dv_xmeta_en);
676 			err = ENOTSUP;
677 			goto error;
678 		}
679 	}
680 	if (config->devx && config->dv_flow_en) {
681 		priv->obj_ops = devx_obj_ops;
682 	} else {
683 		DRV_LOG(ERR, "Flow mode %u is not supported "
684 				"(Windows flow must be DevX with DV flow enabled).",
685 				priv->config.dv_flow_en);
686 		err = ENOTSUP;
687 		goto error;
688 	}
689 	mlx5_flow_counter_mode_config(eth_dev);
690 	return eth_dev;
691 error:
692 	if (priv) {
693 		if (priv->mtr_profile_tbl)
694 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
695 		if (own_domain_id)
696 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
697 		mlx5_free(priv);
698 		if (eth_dev != NULL)
699 			eth_dev->data->dev_private = NULL;
700 	}
701 	if (eth_dev != NULL) {
702 		/* mac_addrs must not be freed alone because part of
703 		 * dev_private
704 		 **/
705 		eth_dev->data->mac_addrs = NULL;
706 		rte_eth_dev_release_port(eth_dev);
707 	}
708 	if (sh)
709 		mlx5_free_shared_dev_ctx(sh);
710 	MLX5_ASSERT(err > 0);
711 	rte_errno = err;
712 	return NULL;
713 }
714 
715 /**
716  * This function should share events between multiple ports of single IB
717  * device.  Currently it has no support under Windows.
718  *
719  * @param sh
720  *   Pointer to mlx5_dev_ctx_shared object.
721  */
722 void
723 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
724 {
725 	(void)sh;
726 	DRV_LOG(WARNING, "%s: is not supported", __func__);
727 }
728 
729 /**
730  * This function should share events between multiple ports of single IB
731  * device.  Currently it has no support under Windows.
732  *
733  * @param dev
734  *   Pointer to mlx5_dev_ctx_shared object.
735  */
736 void
737 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
738 {
739 	(void)sh;
740 	DRV_LOG(WARNING, "%s: is not supported", __func__);
741 }
742 
743 /**
744  * Read statistics by a named counter.
745  *
746  * @param[in] priv
747  *   Pointer to the private device data structure.
748  * @param[in] ctr_name
749  *   Pointer to the name of the statistic counter to read
750  * @param[out] stat
751  *   Pointer to read statistic value.
752  * @return
753  *   0 on success and stat is valud, 1 if failed to read the value
754  *   rte_errno is set.
755  *
756  */
757 int
758 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
759 		      uint64_t *stat)
760 {
761 	RTE_SET_USED(priv);
762 	RTE_SET_USED(ctr_name);
763 	RTE_SET_USED(stat);
764 	DRV_LOG(WARNING, "%s: is not supported", __func__);
765 	return -ENOTSUP;
766 }
767 
768 /**
769  * Flush device MAC addresses
770  * Currently it has no support under Windows.
771  *
772  * @param dev
773  *   Pointer to Ethernet device structure.
774  *
775  */
776 void
777 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
778 {
779 	(void)dev;
780 	DRV_LOG(WARNING, "%s: is not supported", __func__);
781 }
782 
783 /**
784  * Remove a MAC address from device
785  * Currently it has no support under Windows.
786  *
787  * @param dev
788  *   Pointer to Ethernet device structure.
789  * @param index
790  *   MAC address index.
791  */
792 void
793 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
794 {
795 	(void)dev;
796 	(void)(index);
797 	DRV_LOG(WARNING, "%s: is not supported", __func__);
798 }
799 
800 /**
801  * Adds a MAC address to the device
802  * Currently it has no support under Windows.
803  *
804  * @param dev
805  *   Pointer to Ethernet device structure.
806  * @param mac_addr
807  *   MAC address to register.
808  * @param index
809  *   MAC address index.
810  *
811  * @return
812  *   0 on success, a negative errno value otherwise
813  */
814 int
815 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
816 		     uint32_t index)
817 {
818 	(void)index;
819 	struct rte_ether_addr lmac;
820 
821 	if (mlx5_get_mac(dev, &lmac.addr_bytes)) {
822 		DRV_LOG(ERR,
823 			"port %u cannot get MAC address, is mlx5_en"
824 			" loaded? (errno: %s)",
825 			dev->data->port_id, strerror(rte_errno));
826 		return rte_errno;
827 	}
828 	if (!rte_is_same_ether_addr(&lmac, mac)) {
829 		DRV_LOG(ERR,
830 			"adding new mac address to device is unsupported");
831 		return -ENOTSUP;
832 	}
833 	return 0;
834 }
835 
836 /**
837  * Modify a VF MAC address
838  * Currently it has no support under Windows.
839  *
840  * @param priv
841  *   Pointer to device private data.
842  * @param mac_addr
843  *   MAC address to modify into.
844  * @param iface_idx
845  *   Net device interface index
846  * @param vf_index
847  *   VF index
848  *
849  * @return
850  *   0 on success, a negative errno value otherwise
851  */
852 int
853 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
854 			   unsigned int iface_idx,
855 			   struct rte_ether_addr *mac_addr,
856 			   int vf_index)
857 {
858 	(void)priv;
859 	(void)iface_idx;
860 	(void)mac_addr;
861 	(void)vf_index;
862 	DRV_LOG(WARNING, "%s: is not supported", __func__);
863 	return -ENOTSUP;
864 }
865 
866 /**
867  * Set device promiscuous mode
868  * Currently it has no support under Windows.
869  *
870  * @param dev
871  *   Pointer to Ethernet device structure.
872  * @param enable
873  *   0 - promiscuous is disabled, otherwise - enabled
874  *
875  * @return
876  *   0 on success, a negative error value otherwise
877  */
878 int
879 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
880 {
881 	(void)dev;
882 	(void)enable;
883 	DRV_LOG(WARNING, "%s: is not supported", __func__);
884 	return -ENOTSUP;
885 }
886 
887 /**
888  * Set device allmulti mode
889  *
890  * @param dev
891  *   Pointer to Ethernet device structure.
892  * @param enable
893  *   0 - all multicase is disabled, otherwise - enabled
894  *
895  * @return
896  *   0 on success, a negative error value otherwise
897  */
898 int
899 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
900 {
901 	(void)dev;
902 	(void)enable;
903 	DRV_LOG(WARNING, "%s: is not supported", __func__);
904 	return -ENOTSUP;
905 }
906 
907 /**
908  * Detect if a devx_device_bdf object has identical DBDF values to the
909  * rte_pci_addr found in bus/pci probing
910  *
911  * @param[in] devx_bdf
912  *   Pointer to the devx_device_bdf structure.
913  * @param[in] addr
914  *   Pointer to the rte_pci_addr structure.
915  *
916  * @return
917  *   1 on Device match, 0 on mismatch.
918  */
919 static int
920 mlx5_match_devx_bdf_to_addr(struct devx_device_bdf *devx_bdf,
921 			    struct rte_pci_addr *addr)
922 {
923 	if (addr->domain != (devx_bdf->bus_id >> 8) ||
924 	    addr->bus != (devx_bdf->bus_id & 0xff) ||
925 	    addr->devid != devx_bdf->dev_id ||
926 	    addr->function != devx_bdf->fnc_id) {
927 		return 0;
928 	}
929 	return 1;
930 }
931 
932 /**
933  * Detect if a devx_device_bdf object matches the rte_pci_addr
934  * found in bus/pci probing
935  * Compare both the Native/PF BDF and the raw_bdf representing a VF BDF.
936  *
937  * @param[in] devx_bdf
938  *   Pointer to the devx_device_bdf structure.
939  * @param[in] addr
940  *   Pointer to the rte_pci_addr structure.
941  *
942  * @return
943  *   1 on Device match, 0 on mismatch, rte_errno code on failure.
944  */
945 static int
946 mlx5_match_devx_devices_to_addr(struct devx_device_bdf *devx_bdf,
947 				struct rte_pci_addr *addr)
948 {
949 	int err;
950 	struct devx_device mlx5_dev;
951 
952 	if (mlx5_match_devx_bdf_to_addr(devx_bdf, addr))
953 		return 1;
954 	/**
955 	 * Didn't match on Native/PF BDF, could still
956 	 * Match a VF BDF, check it next
957 	 */
958 	err = mlx5_glue->query_device(devx_bdf, &mlx5_dev);
959 	if (err) {
960 		DRV_LOG(ERR, "query_device failed");
961 		rte_errno = err;
962 		return rte_errno;
963 	}
964 	if (mlx5_match_devx_bdf_to_addr(&mlx5_dev.raw_bdf, addr))
965 		return 1;
966 	return 0;
967 }
968 
969 /**
970  * DPDK callback to register a PCI device.
971  *
972  * This function spawns Ethernet devices out of a given device.
973  *
974  * @param[in] dev
975  *   Pointer to the generic device.
976  *
977  * @return
978  *   0 on success, a negative errno value otherwise and rte_errno is set.
979  */
980 int
981 mlx5_os_net_probe(struct rte_device *dev)
982 {
983 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev);
984 	struct devx_device_bdf *devx_bdf_devs, *orig_devx_bdf_devs;
985 	/*
986 	 * Number of found IB Devices matching with requested PCI BDF.
987 	 * nd != 1 means there are multiple IB devices over the same
988 	 * PCI device and we have representors and master.
989 	 */
990 	unsigned int nd = 0;
991 	/*
992 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
993 	 * we have the single multiport IB device, and there may be
994 	 * representors attached to some of found ports.
995 	 * Currently not supported.
996 	 * unsigned int np = 0;
997 	 */
998 
999 	/*
1000 	 * Number of DPDK ethernet devices to Spawn - either over
1001 	 * multiple IB devices or multiple ports of single IB device.
1002 	 * Actually this is the number of iterations to spawn.
1003 	 */
1004 	unsigned int ns = 0;
1005 	/*
1006 	 * Bonding device
1007 	 *   < 0 - no bonding device (single one)
1008 	 *  >= 0 - bonding device (value is slave PF index)
1009 	 */
1010 	int bd = -1;
1011 	struct mlx5_dev_spawn_data *list = NULL;
1012 	struct mlx5_dev_config dev_config;
1013 	unsigned int dev_config_vf;
1014 	int ret, err;
1015 	uint32_t restore;
1016 
1017 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1018 		DRV_LOG(ERR, "Secondary process is not supported on Windows.");
1019 		return -ENOTSUP;
1020 	}
1021 	ret = mlx5_init_once();
1022 	if (ret) {
1023 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1024 			strerror(rte_errno));
1025 		return -rte_errno;
1026 	}
1027 	errno = 0;
1028 	devx_bdf_devs = mlx5_glue->get_device_list(&ret);
1029 	orig_devx_bdf_devs = devx_bdf_devs;
1030 	if (!devx_bdf_devs) {
1031 		rte_errno = errno ? errno : ENOSYS;
1032 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1033 		return -rte_errno;
1034 	}
1035 	/*
1036 	 * First scan the list of all Infiniband devices to find
1037 	 * matching ones, gathering into the list.
1038 	 */
1039 	struct devx_device_bdf *devx_bdf_match[ret + 1];
1040 
1041 	while (ret-- > 0) {
1042 		err = mlx5_match_devx_devices_to_addr(devx_bdf_devs,
1043 		    &pci_dev->addr);
1044 		if (!err) {
1045 			devx_bdf_devs++;
1046 			continue;
1047 		}
1048 		if (err != 1) {
1049 			ret = -err;
1050 			goto exit;
1051 		}
1052 		devx_bdf_match[nd++] = devx_bdf_devs;
1053 	}
1054 	devx_bdf_match[nd] = NULL;
1055 	if (!nd) {
1056 		/* No device matches, just complain and bail out. */
1057 		DRV_LOG(WARNING,
1058 			"no DevX device matches PCI device " PCI_PRI_FMT ","
1059 			" is DevX Configured?",
1060 			pci_dev->addr.domain, pci_dev->addr.bus,
1061 			pci_dev->addr.devid, pci_dev->addr.function);
1062 		rte_errno = ENOENT;
1063 		ret = -rte_errno;
1064 		goto exit;
1065 	}
1066 	/*
1067 	 * Now we can determine the maximal
1068 	 * amount of devices to be spawned.
1069 	 */
1070 	list = mlx5_malloc(MLX5_MEM_ZERO,
1071 			   sizeof(struct mlx5_dev_spawn_data),
1072 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1073 	if (!list) {
1074 		DRV_LOG(ERR, "spawn data array allocation failure");
1075 		rte_errno = ENOMEM;
1076 		ret = -rte_errno;
1077 		goto exit;
1078 	}
1079 	memset(&list[ns].info, 0, sizeof(list[ns].info));
1080 	list[ns].max_port = 1;
1081 	list[ns].phys_port = 1;
1082 	list[ns].phys_dev = devx_bdf_match[ns];
1083 	list[ns].eth_dev = NULL;
1084 	list[ns].pci_dev = pci_dev;
1085 	list[ns].pf_bond = bd;
1086 	list[ns].ifindex = -1; /* Spawn will assign */
1087 	list[ns].info =
1088 		(struct mlx5_switch_info){
1089 			.master = 0,
1090 			.representor = 0,
1091 			.name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK,
1092 			.port_name = 0,
1093 			.switch_id = 0,
1094 		};
1095 	/* Device specific configuration. */
1096 	switch (pci_dev->id.device_id) {
1097 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1098 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1099 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1100 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1101 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1102 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1103 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
1104 		dev_config_vf = 1;
1105 		break;
1106 	default:
1107 		dev_config_vf = 0;
1108 		break;
1109 	}
1110 	/* Default configuration. */
1111 	memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
1112 	dev_config.vf = dev_config_vf;
1113 	dev_config.mps = 0;
1114 	dev_config.dbnc = MLX5_ARG_UNSET;
1115 	dev_config.rx_vec_en = 1;
1116 	dev_config.txq_inline_max = MLX5_ARG_UNSET;
1117 	dev_config.txq_inline_min = MLX5_ARG_UNSET;
1118 	dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
1119 	dev_config.txqs_inline = MLX5_ARG_UNSET;
1120 	dev_config.vf_nl_en = 0;
1121 	dev_config.mr_ext_memseg_en = 1;
1122 	dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
1123 	dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
1124 	dev_config.dv_esw_en = 0;
1125 	dev_config.dv_flow_en = 1;
1126 	dev_config.decap_en = 0;
1127 	dev_config.log_hp_size = MLX5_ARG_UNSET;
1128 	list[ns].numa_node = pci_dev->device.numa_node;
1129 	list[ns].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1130 					  &list[ns],
1131 					  &dev_config);
1132 	if (!list[ns].eth_dev)
1133 		goto exit;
1134 	restore = list[ns].eth_dev->data->dev_flags;
1135 	rte_eth_copy_pci_info(list[ns].eth_dev, pci_dev);
1136 	/* Restore non-PCI flags cleared by the above call. */
1137 	list[ns].eth_dev->data->dev_flags |= restore;
1138 	rte_eth_dev_probing_finish(list[ns].eth_dev);
1139 	ret = 0;
1140 exit:
1141 	/*
1142 	 * Do the routine cleanup:
1143 	 * - free allocated spawn data array
1144 	 * - free the device list
1145 	 */
1146 	if (list)
1147 		mlx5_free(list);
1148 	MLX5_ASSERT(orig_devx_bdf_devs);
1149 	mlx5_glue->free_device_list(orig_devx_bdf_devs);
1150 	return ret;
1151 }
1152 
1153 /**
1154  * Set the reg_mr and dereg_mr call backs
1155  *
1156  * @param reg_mr_cb[out]
1157  *   Pointer to reg_mr func
1158  * @param dereg_mr_cb[out]
1159  *   Pointer to dereg_mr func
1160  *
1161  */
1162 void
1163 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
1164 		      mlx5_dereg_mr_t *dereg_mr_cb)
1165 {
1166 	*reg_mr_cb = mlx5_os_reg_mr;
1167 	*dereg_mr_cb = mlx5_os_dereg_mr;
1168 }
1169 
1170 /**
1171  * Extract pdn of PD object using DevX
1172  *
1173  * @param[in] pd
1174  *   Pointer to the DevX PD object.
1175  * @param[out] pdn
1176  *   Pointer to the PD object number variable.
1177  *
1178  * @return
1179  *   0 on success, error value otherwise.
1180  */
1181 int
1182 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1183 {
1184 	if (!pd)
1185 		return -EINVAL;
1186 
1187 	*pdn = ((struct mlx5_pd *)pd)->pdn;
1188 	return 0;
1189 }
1190 
1191 const struct mlx5_flow_driver_ops mlx5_flow_verbs_drv_ops = {0};
1192