xref: /dpdk/drivers/net/mlx5/windows/mlx5_os.c (revision 3bb3ebb51b789d4ecb417cbdb1dce5c7211f6f18)
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_autoconf.h"
27 #include "mlx5_mr.h"
28 #include "mlx5_flow.h"
29 #include "mlx5_devx.h"
30 
31 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
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 /**
39  * Initialize shared data between primary and secondary process.
40  *
41  * A memzone is reserved by primary process and secondary processes attach to
42  * the memzone.
43  *
44  * @return
45  *   0 on success, a negative errno value otherwise and rte_errno is set.
46  */
47 static int
48 mlx5_init_shared_data(void)
49 {
50 	const struct rte_memzone *mz;
51 	int ret = 0;
52 
53 	rte_spinlock_lock(&mlx5_shared_data_lock);
54 	if (mlx5_shared_data == NULL) {
55 		/* Allocate shared memory. */
56 		mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
57 					 sizeof(*mlx5_shared_data),
58 					 SOCKET_ID_ANY, 0);
59 		if (mz == NULL) {
60 			DRV_LOG(ERR,
61 				"Cannot allocate mlx5 shared data");
62 			ret = -rte_errno;
63 			goto error;
64 		}
65 		mlx5_shared_data = mz->addr;
66 		memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
67 		rte_spinlock_init(&mlx5_shared_data->lock);
68 	}
69 error:
70 	rte_spinlock_unlock(&mlx5_shared_data_lock);
71 	return ret;
72 }
73 
74 /**
75  * PMD global initialization.
76  *
77  * Independent from individual device, this function initializes global
78  * per-PMD data structures distinguishing primary and secondary processes.
79  * Hence, each initialization is called once per a process.
80  *
81  * @return
82  *   0 on success, a negative errno value otherwise and rte_errno is set.
83  */
84 static int
85 mlx5_init_once(void)
86 {
87 	if (mlx5_init_shared_data())
88 		return -rte_errno;
89 	return 0;
90 }
91 
92 /**
93  * Get mlx5 device attributes.
94  *
95  * @param ctx
96  *   Pointer to device context.
97  *
98  * @param device_attr
99  *   Pointer to mlx5 device attributes.
100  *
101  * @return
102  *   0 on success, non zero error number otherwise
103  */
104 int
105 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
106 {
107 	struct mlx5_context *mlx5_ctx;
108 	struct mlx5_hca_attr hca_attr;
109 	void *pv_iseg = NULL;
110 	u32 cb_iseg = 0;
111 	int err = 0;
112 
113 	if (!ctx)
114 		return -EINVAL;
115 	mlx5_ctx = (struct mlx5_context *)ctx;
116 	memset(device_attr, 0, sizeof(*device_attr));
117 	err = mlx5_devx_cmd_query_hca_attr(mlx5_ctx, &hca_attr);
118 	if (err) {
119 		DRV_LOG(ERR, "Failed to get device hca_cap");
120 		return err;
121 	}
122 	device_attr->max_cq = 1 << hca_attr.log_max_cq;
123 	device_attr->max_qp = 1 << hca_attr.log_max_qp;
124 	device_attr->max_qp_wr = 1 << hca_attr.log_max_qp_sz;
125 	device_attr->max_cqe = 1 << hca_attr.log_max_cq_sz;
126 	device_attr->max_mr = 1 << hca_attr.log_max_mrw_sz;
127 	device_attr->max_pd = 1 << hca_attr.log_max_pd;
128 	device_attr->max_srq = 1 << hca_attr.log_max_srq;
129 	device_attr->max_srq_wr = 1 << hca_attr.log_max_srq_sz;
130 	if (hca_attr.rss_ind_tbl_cap) {
131 		device_attr->max_rwq_indirection_table_size =
132 			1 << hca_attr.rss_ind_tbl_cap;
133 	}
134 	pv_iseg = mlx5_glue->query_hca_iseg(mlx5_ctx, &cb_iseg);
135 	if (pv_iseg == NULL) {
136 		DRV_LOG(ERR, "Failed to get device hca_iseg");
137 		return errno;
138 	}
139 	if (!err) {
140 		snprintf(device_attr->fw_ver, 64, "%x.%x.%04x",
141 			MLX5_GET(initial_seg, pv_iseg, fw_rev_major),
142 			MLX5_GET(initial_seg, pv_iseg, fw_rev_minor),
143 			MLX5_GET(initial_seg, pv_iseg, fw_rev_subminor));
144 	}
145 	return err;
146 }
147 
148 /**
149  * Initialize DR related data within private structure.
150  * Routine checks the reference counter and does actual
151  * resources creation/initialization only if counter is zero.
152  *
153  * @param[in] priv
154  *   Pointer to the private device data structure.
155  *
156  * @return
157  *   Zero on success, positive error code otherwise.
158  */
159 static int
160 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
161 {
162 	struct mlx5_dev_ctx_shared *sh = priv->sh;
163 	int err = 0;
164 
165 	if (!sh->flow_tbls)
166 		err = mlx5_alloc_table_hash_list(priv);
167 	else
168 		DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse",
169 			(void *)sh->flow_tbls);
170 	return err;
171 }
172 /**
173  * Destroy DR related data within private structure.
174  *
175  * @param[in] priv
176  *   Pointer to the private device data structure.
177  */
178 void
179 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
180 {
181 	mlx5_free_table_hash_list(priv);
182 }
183 
184 /**
185  * Set the completion channel file descriptor interrupt as non-blocking.
186  * Currently it has no support under Windows.
187  *
188  * @param[in] rxq_obj
189  *   Pointer to RQ channel object, which includes the channel fd
190  *
191  * @param[out] fd
192  *   The file descriptor (representing the intetrrupt) used in this channel.
193  *
194  * @return
195  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
196  */
197 int
198 mlx5_os_set_nonblock_channel_fd(int fd)
199 {
200 	(void)fd;
201 	DRV_LOG(WARNING, "%s: is not supported", __func__);
202 	return -ENOTSUP;
203 }
204 
205 /**
206  * Function API open device under Windows
207  *
208  * This function calls the Windows glue APIs to open a device.
209  *
210  * @param[in] spawn
211  *   Pointer to the device attributes (name, port, etc).
212  * @param[out] config
213  *   Pointer to device configuration structure.
214  * @param[out] sh
215  *   Pointer to shared context structure.
216  *
217  * @return
218  *   0 on success, a positive error value otherwise.
219  */
220 int
221 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
222 		 const struct mlx5_dev_config *config,
223 		 struct mlx5_dev_ctx_shared *sh)
224 {
225 	RTE_SET_USED(config);
226 	int err = 0;
227 	struct mlx5_context *mlx5_ctx;
228 
229 	pthread_mutex_init(&sh->txpp.mutex, NULL);
230 	/* Set numa node from pci probe */
231 	sh->numa_node = spawn->pci_dev->device.numa_node;
232 
233 	/* Try to open device with DevX */
234 	rte_errno = 0;
235 	sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
236 	if (!sh->ctx) {
237 		DRV_LOG(ERR, "open_device failed");
238 		err = errno;
239 		return err;
240 	}
241 	sh->devx = 1;
242 	mlx5_ctx = (struct mlx5_context *)sh->ctx;
243 	err = mlx5_glue->query_device(spawn->phys_dev, &mlx5_ctx->mlx5_dev);
244 	if (err)
245 		DRV_LOG(ERR, "Failed to query device context fields.");
246 	return err;
247 }
248 
249 /**
250  * DV flow counter mode detect and config.
251  *
252  * @param dev
253  *   Pointer to rte_eth_dev structure.
254  *
255  */
256 static void
257 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
258 {
259 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
260 	struct mlx5_priv *priv = dev->data->dev_private;
261 	struct mlx5_dev_ctx_shared *sh = priv->sh;
262 	bool fallback;
263 
264 #ifndef HAVE_IBV_DEVX_ASYNC
265 	fallback = true;
266 #else
267 	fallback = false;
268 	if (!priv->config.devx || !priv->config.dv_flow_en ||
269 	    !priv->config.hca_attr.flow_counters_dump ||
270 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
271 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
272 		fallback = true;
273 #endif
274 	if (fallback)
275 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
276 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
277 			priv->config.hca_attr.flow_counters_dump,
278 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
279 	/* Initialize fallback mode only on the port initializes sh. */
280 	if (sh->refcnt == 1)
281 		sh->cmng.counter_fallback = fallback;
282 	else if (fallback != sh->cmng.counter_fallback)
283 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
284 			"with others:%d.", PORT_ID(priv), fallback);
285 #endif
286 }
287 
288 /**
289  * Spawn an Ethernet device from Verbs information.
290  *
291  * @param dpdk_dev
292  *   Backing DPDK device.
293  * @param spawn
294  *   Verbs device parameters (name, port, switch_info) to spawn.
295  * @param config
296  *   Device configuration parameters.
297  *
298  * @return
299  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
300  *   is set. The following errors are defined:
301  *
302  *   EEXIST: device is already spawned
303  */
304 static struct rte_eth_dev *
305 mlx5_dev_spawn(struct rte_device *dpdk_dev,
306 	       struct mlx5_dev_spawn_data *spawn,
307 	       struct mlx5_dev_config *config)
308 {
309 	const struct mlx5_switch_info *switch_info = &spawn->info;
310 	struct mlx5_dev_ctx_shared *sh = NULL;
311 	struct mlx5_dev_attr device_attr;
312 	struct rte_eth_dev *eth_dev = NULL;
313 	struct mlx5_priv *priv = NULL;
314 	int err = 0;
315 	unsigned int cqe_comp;
316 	struct rte_ether_addr mac;
317 	char name[RTE_ETH_NAME_MAX_LEN];
318 	int own_domain_id = 0;
319 	uint16_t port_id;
320 
321 	/* Build device name. */
322 	strlcpy(name, dpdk_dev->name, sizeof(name));
323 	/* check if the device is already spawned */
324 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
325 		rte_errno = EEXIST;
326 		return NULL;
327 	}
328 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
329 	/*
330 	 * Some parameters are needed in advance to create device context. We
331 	 * process the devargs here to get ones, and later process devargs
332 	 * again to override some hardware settings.
333 	 */
334 	err = mlx5_args(config, dpdk_dev->devargs);
335 	if (err) {
336 		err = rte_errno;
337 		DRV_LOG(ERR, "failed to process device arguments: %s",
338 			strerror(rte_errno));
339 		goto error;
340 	}
341 	mlx5_malloc_mem_select(config->sys_mem_en);
342 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
343 	if (!sh)
344 		return NULL;
345 	config->devx = sh->devx;
346 	/* Initialize the shutdown event in mlx5_dev_spawn to
347 	 * support mlx5_is_removed for Windows.
348 	 */
349 	err = mlx5_glue->devx_init_showdown_event(sh->ctx);
350 	if (err) {
351 		DRV_LOG(ERR, "failed to init showdown event: %s",
352 			strerror(errno));
353 		goto error;
354 	}
355 	DRV_LOG(DEBUG, "MPW isn't supported");
356 	mlx5_os_get_dev_attr(sh->ctx, &device_attr);
357 	config->swp = 0;
358 	config->ind_table_max_size =
359 		sh->device_attr.max_rwq_indirection_table_size;
360 	if (RTE_CACHE_LINE_SIZE == 128 &&
361 	    !(device_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
362 		cqe_comp = 0;
363 	else
364 		cqe_comp = 1;
365 	config->cqe_comp = cqe_comp;
366 	DRV_LOG(DEBUG, "tunnel offloading is not supported");
367 	config->tunnel_en = 0;
368 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is no supported");
369 	config->mpls_en = 0;
370 	/* Allocate private eth device data. */
371 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
372 			   sizeof(*priv),
373 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
374 	if (priv == NULL) {
375 		DRV_LOG(ERR, "priv allocation failure");
376 		err = ENOMEM;
377 		goto error;
378 	}
379 	priv->sh = sh;
380 	priv->dev_port = spawn->phys_port;
381 	priv->pci_dev = spawn->pci_dev;
382 	priv->mtu = RTE_ETHER_MTU;
383 	priv->mp_id.port_id = port_id;
384 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
385 	priv->representor = !!switch_info->representor;
386 	priv->master = !!switch_info->master;
387 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
388 	priv->vport_meta_tag = 0;
389 	priv->vport_meta_mask = 0;
390 	priv->pf_bond = spawn->pf_bond;
391 	priv->vport_id = -1;
392 	/* representor_id field keeps the unmodified VF index. */
393 	priv->representor_id = -1;
394 	/*
395 	 * Look for sibling devices in order to reuse their switch domain
396 	 * if any, otherwise allocate one.
397 	 */
398 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
399 		const struct mlx5_priv *opriv =
400 			rte_eth_devices[port_id].data->dev_private;
401 
402 		if (!opriv ||
403 		    opriv->sh != priv->sh ||
404 			opriv->domain_id ==
405 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
406 			continue;
407 		priv->domain_id = opriv->domain_id;
408 		break;
409 	}
410 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
411 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
412 		if (err) {
413 			err = rte_errno;
414 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
415 				strerror(rte_errno));
416 			goto error;
417 		}
418 		own_domain_id = 1;
419 	}
420 	/* Override some values set by hardware configuration. */
421 	mlx5_args(config, dpdk_dev->devargs);
422 	err = mlx5_dev_check_sibling_config(priv, config);
423 	if (err)
424 		goto error;
425 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
426 			    IBV_DEVICE_RAW_IP_CSUM);
427 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
428 		(config->hw_csum ? "" : "not "));
429 	DRV_LOG(DEBUG, "counters are not supported");
430 	config->ind_table_max_size =
431 		sh->device_attr.max_rwq_indirection_table_size;
432 	/*
433 	 * Remove this check once DPDK supports larger/variable
434 	 * indirection tables.
435 	 */
436 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
437 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
438 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
439 		config->ind_table_max_size);
440 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
441 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
442 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
443 		(config->hw_vlan_strip ? "" : "not "));
444 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
445 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
446 	if (config->hw_padding) {
447 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
448 		config->hw_padding = 0;
449 	}
450 	config->tso = (sh->device_attr.max_tso > 0 &&
451 		      (sh->device_attr.tso_supported_qpts &
452 		       (1 << IBV_QPT_RAW_PACKET)));
453 	if (config->tso)
454 		config->tso_max_payload_sz = sh->device_attr.max_tso;
455 	DRV_LOG(DEBUG, "%sMPS is %s.",
456 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
457 		config->mps == MLX5_MPW ? "legacy " : "",
458 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
459 	if (config->cqe_comp && !cqe_comp) {
460 		DRV_LOG(WARNING, "Rx CQE compression isn't supported.");
461 		config->cqe_comp = 0;
462 	}
463 	if (config->devx) {
464 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
465 		if (err) {
466 			err = -err;
467 			goto error;
468 		}
469 		/* Check relax ordering support. */
470 		sh->cmng.relaxed_ordering_read = 0;
471 		sh->cmng.relaxed_ordering_write = 0;
472 		if (!haswell_broadwell_cpu) {
473 			sh->cmng.relaxed_ordering_write =
474 				config->hca_attr.relaxed_ordering_write;
475 			sh->cmng.relaxed_ordering_read =
476 				config->hca_attr.relaxed_ordering_read;
477 		}
478 	}
479 	if (config->devx) {
480 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
481 
482 		err = config->hca_attr.access_register_user ?
483 			mlx5_devx_cmd_register_read
484 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
485 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
486 		if (!err) {
487 			uint32_t ts_mode;
488 
489 			/* MTUTC register is read successfully. */
490 			ts_mode = MLX5_GET(register_mtutc, reg,
491 					   time_stamp_mode);
492 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
493 				config->rt_timestamp = 1;
494 		} else {
495 			/* Kernel does not support register reading. */
496 			if (config->hca_attr.dev_freq_khz ==
497 						 (NS_PER_S / MS_PER_S))
498 				config->rt_timestamp = 1;
499 		}
500 		sh->rq_ts_format = config->hca_attr.rq_ts_format;
501 		sh->sq_ts_format = config->hca_attr.sq_ts_format;
502 		sh->qp_ts_format = config->hca_attr.qp_ts_format;
503 	}
504 	if (config->mprq.enabled) {
505 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
506 		config->mprq.enabled = 0;
507 	}
508 	if (config->max_dump_files_num == 0)
509 		config->max_dump_files_num = 128;
510 	eth_dev = rte_eth_dev_allocate(name);
511 	if (eth_dev == NULL) {
512 		DRV_LOG(ERR, "can not allocate rte ethdev");
513 		err = ENOMEM;
514 		goto error;
515 	}
516 	if (priv->representor) {
517 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
518 		eth_dev->data->representor_id = priv->representor_id;
519 	}
520 	/*
521 	 * Store associated network device interface index. This index
522 	 * is permanent throughout the lifetime of device. So, we may store
523 	 * the ifindex here and use the cached value further.
524 	 */
525 	MLX5_ASSERT(spawn->ifindex);
526 	priv->if_index = spawn->ifindex;
527 	eth_dev->data->dev_private = priv;
528 	priv->dev_data = eth_dev->data;
529 	eth_dev->data->mac_addrs = priv->mac;
530 	eth_dev->device = dpdk_dev;
531 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
532 	/* Configure the first MAC address by default. */
533 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
534 		DRV_LOG(ERR,
535 			"port %u cannot get MAC address, is mlx5_en"
536 			" loaded? (errno: %s).",
537 			eth_dev->data->port_id, strerror(rte_errno));
538 		err = ENODEV;
539 		goto error;
540 	}
541 	DRV_LOG(INFO,
542 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
543 		eth_dev->data->port_id,
544 		mac.addr_bytes[0], mac.addr_bytes[1],
545 		mac.addr_bytes[2], mac.addr_bytes[3],
546 		mac.addr_bytes[4], mac.addr_bytes[5]);
547 #ifdef RTE_LIBRTE_MLX5_DEBUG
548 	{
549 		char ifname[MLX5_NAMESIZE];
550 
551 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
552 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
553 				eth_dev->data->port_id, ifname);
554 		else
555 			DRV_LOG(DEBUG, "port %u ifname is unknown.",
556 				eth_dev->data->port_id);
557 	}
558 #endif
559 	/* Get actual MTU if possible. */
560 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
561 	if (err) {
562 		err = rte_errno;
563 		goto error;
564 	}
565 	DRV_LOG(DEBUG, "port %u MTU is %u.", eth_dev->data->port_id,
566 		priv->mtu);
567 	/* Initialize burst functions to prevent crashes before link-up. */
568 	eth_dev->rx_pkt_burst = removed_rx_burst;
569 	eth_dev->tx_pkt_burst = removed_tx_burst;
570 	eth_dev->dev_ops = &mlx5_dev_ops;
571 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
572 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
573 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
574 	/* Register MAC address. */
575 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
576 	priv->flows = 0;
577 	priv->ctrl_flows = 0;
578 	TAILQ_INIT(&priv->flow_meters);
579 	TAILQ_INIT(&priv->flow_meter_profiles);
580 	/* Bring Ethernet device up. */
581 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up.",
582 		eth_dev->data->port_id);
583 	/* nl calls are unsupported - set to -1 not to fail on release */
584 	priv->nl_socket_rdma = -1;
585 	priv->nl_socket_route = -1;
586 	mlx5_set_link_up(eth_dev);
587 	/*
588 	 * Even though the interrupt handler is not installed yet,
589 	 * interrupts will still trigger on the async_fd from
590 	 * Verbs context returned by ibv_open_device().
591 	 */
592 	mlx5_link_update(eth_dev, 0);
593 	config->dv_esw_en = 0;
594 	/* Detect minimal data bytes to inline. */
595 	mlx5_set_min_inline(spawn, config);
596 	/* Store device configuration on private structure. */
597 	priv->config = *config;
598 	/* Create context for virtual machine VLAN workaround. */
599 	priv->vmwa_context = NULL;
600 	if (config->dv_flow_en) {
601 		err = mlx5_alloc_shared_dr(priv);
602 		if (err)
603 			goto error;
604 	}
605 	/* No supported flow priority number detection. */
606 	priv->config.flow_prio = -1;
607 	if (!priv->config.dv_esw_en &&
608 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
609 		DRV_LOG(WARNING, "metadata mode %u is not supported "
610 				 "(no E-Switch)", priv->config.dv_xmeta_en);
611 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
612 	}
613 	mlx5_set_metadata_mask(eth_dev);
614 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
615 	    !priv->sh->dv_regc0_mask) {
616 		DRV_LOG(ERR, "metadata mode %u is not supported "
617 			     "(no metadata reg_c[0] is available).",
618 			     priv->config.dv_xmeta_en);
619 			err = ENOTSUP;
620 			goto error;
621 	}
622 	mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
623 			     mlx5_hrxq_create_cb,
624 			     mlx5_hrxq_match_cb,
625 			     mlx5_hrxq_remove_cb);
626 	/* Query availability of metadata reg_c's. */
627 	err = mlx5_flow_discover_mreg_c(eth_dev);
628 	if (err < 0) {
629 		err = -err;
630 		goto error;
631 	}
632 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
633 		DRV_LOG(DEBUG,
634 			"port %u extensive metadata register is not supported.",
635 			eth_dev->data->port_id);
636 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
637 			DRV_LOG(ERR, "metadata mode %u is not supported "
638 				     "(no metadata registers available).",
639 				     priv->config.dv_xmeta_en);
640 			err = ENOTSUP;
641 			goto error;
642 		}
643 	}
644 	if (config->devx && config->dv_flow_en) {
645 		priv->obj_ops = devx_obj_ops;
646 	} else {
647 		DRV_LOG(ERR, "Flow mode %u is not supported "
648 				"(Windows flow must be DevX with DV flow enabled).",
649 				priv->config.dv_flow_en);
650 		err = ENOTSUP;
651 		goto error;
652 	}
653 	mlx5_flow_counter_mode_config(eth_dev);
654 	return eth_dev;
655 error:
656 	if (priv) {
657 		if (own_domain_id)
658 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
659 		mlx5_free(priv);
660 		if (eth_dev != NULL)
661 			eth_dev->data->dev_private = NULL;
662 	}
663 	if (eth_dev != NULL) {
664 		/* mac_addrs must not be freed alone because part of
665 		 * dev_private
666 		 **/
667 		eth_dev->data->mac_addrs = NULL;
668 		rte_eth_dev_release_port(eth_dev);
669 	}
670 	if (sh)
671 		mlx5_free_shared_dev_ctx(sh);
672 	MLX5_ASSERT(err > 0);
673 	rte_errno = err;
674 	return NULL;
675 }
676 
677 /**
678  * This function should share events between multiple ports of single IB
679  * device.  Currently it has no support under Windows.
680  *
681  * @param sh
682  *   Pointer to mlx5_dev_ctx_shared object.
683  */
684 void
685 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
686 {
687 	(void)sh;
688 	DRV_LOG(WARNING, "%s: is not supported", __func__);
689 }
690 
691 /**
692  * This function should share events between multiple ports of single IB
693  * device.  Currently it has no support under Windows.
694  *
695  * @param dev
696  *   Pointer to mlx5_dev_ctx_shared object.
697  */
698 void
699 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
700 {
701 	(void)sh;
702 	DRV_LOG(WARNING, "%s: is not supported", __func__);
703 }
704 
705 /**
706  * Read statistics by a named counter.
707  *
708  * @param[in] priv
709  *   Pointer to the private device data structure.
710  * @param[in] ctr_name
711  *   Pointer to the name of the statistic counter to read
712  * @param[out] stat
713  *   Pointer to read statistic value.
714  * @return
715  *   0 on success and stat is valud, 1 if failed to read the value
716  *   rte_errno is set.
717  *
718  */
719 int
720 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
721 		      uint64_t *stat)
722 {
723 	RTE_SET_USED(priv);
724 	RTE_SET_USED(ctr_name);
725 	RTE_SET_USED(stat);
726 	DRV_LOG(WARNING, "%s: is not supported", __func__);
727 	return -ENOTSUP;
728 }
729 
730 /**
731  * Flush device MAC addresses
732  * Currently it has no support under Windows.
733  *
734  * @param dev
735  *   Pointer to Ethernet device structure.
736  *
737  */
738 void
739 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
740 {
741 	(void)dev;
742 	DRV_LOG(WARNING, "%s: is not supported", __func__);
743 }
744 
745 /**
746  * Remove a MAC address from device
747  * Currently it has no support under Windows.
748  *
749  * @param dev
750  *   Pointer to Ethernet device structure.
751  * @param index
752  *   MAC address index.
753  */
754 void
755 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
756 {
757 	(void)dev;
758 	(void)(index);
759 	DRV_LOG(WARNING, "%s: is not supported", __func__);
760 }
761 
762 /**
763  * Adds a MAC address to the device
764  * Currently it has no support under Windows.
765  *
766  * @param dev
767  *   Pointer to Ethernet device structure.
768  * @param mac_addr
769  *   MAC address to register.
770  * @param index
771  *   MAC address index.
772  *
773  * @return
774  *   0 on success, a negative errno value otherwise
775  */
776 int
777 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
778 		     uint32_t index)
779 {
780 	(void)index;
781 	struct rte_ether_addr lmac;
782 
783 	if (mlx5_get_mac(dev, &lmac.addr_bytes)) {
784 		DRV_LOG(ERR,
785 			"port %u cannot get MAC address, is mlx5_en"
786 			" loaded? (errno: %s)",
787 			dev->data->port_id, strerror(rte_errno));
788 		return rte_errno;
789 	}
790 	if (!rte_is_same_ether_addr(&lmac, mac)) {
791 		DRV_LOG(ERR,
792 			"adding new mac address to device is unsupported");
793 		return -ENOTSUP;
794 	}
795 	return 0;
796 }
797 
798 /**
799  * Modify a VF MAC address
800  * Currently it has no support under Windows.
801  *
802  * @param priv
803  *   Pointer to device private data.
804  * @param mac_addr
805  *   MAC address to modify into.
806  * @param iface_idx
807  *   Net device interface index
808  * @param vf_index
809  *   VF index
810  *
811  * @return
812  *   0 on success, a negative errno value otherwise
813  */
814 int
815 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
816 			   unsigned int iface_idx,
817 			   struct rte_ether_addr *mac_addr,
818 			   int vf_index)
819 {
820 	(void)priv;
821 	(void)iface_idx;
822 	(void)mac_addr;
823 	(void)vf_index;
824 	DRV_LOG(WARNING, "%s: is not supported", __func__);
825 	return -ENOTSUP;
826 }
827 
828 /**
829  * Set device promiscuous mode
830  * Currently it has no support under Windows.
831  *
832  * @param dev
833  *   Pointer to Ethernet device structure.
834  * @param enable
835  *   0 - promiscuous is disabled, otherwise - enabled
836  *
837  * @return
838  *   0 on success, a negative error value otherwise
839  */
840 int
841 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
842 {
843 	(void)dev;
844 	(void)enable;
845 	DRV_LOG(WARNING, "%s: is not supported", __func__);
846 	return -ENOTSUP;
847 }
848 
849 /**
850  * Set device allmulti mode
851  *
852  * @param dev
853  *   Pointer to Ethernet device structure.
854  * @param enable
855  *   0 - all multicase is disabled, otherwise - enabled
856  *
857  * @return
858  *   0 on success, a negative error value otherwise
859  */
860 int
861 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
862 {
863 	(void)dev;
864 	(void)enable;
865 	DRV_LOG(WARNING, "%s: is not supported", __func__);
866 	return -ENOTSUP;
867 }
868 
869 /**
870  * Detect if a devx_device_bdf object has identical DBDF values to the
871  * rte_pci_addr found in bus/pci probing
872  *
873  * @param[in] devx_bdf
874  *   Pointer to the devx_device_bdf structure.
875  * @param[in] addr
876  *   Pointer to the rte_pci_addr structure.
877  *
878  * @return
879  *   1 on Device match, 0 on mismatch.
880  */
881 static int
882 mlx5_match_devx_bdf_to_addr(struct devx_device_bdf *devx_bdf,
883 			    struct rte_pci_addr *addr)
884 {
885 	if (addr->domain != (devx_bdf->bus_id >> 8) ||
886 	    addr->bus != (devx_bdf->bus_id & 0xff) ||
887 	    addr->devid != devx_bdf->dev_id ||
888 	    addr->function != devx_bdf->fnc_id) {
889 		return 0;
890 	}
891 	return 1;
892 }
893 
894 /**
895  * Detect if a devx_device_bdf object matches the rte_pci_addr
896  * found in bus/pci probing
897  * Compare both the Native/PF BDF and the raw_bdf representing a VF BDF.
898  *
899  * @param[in] devx_bdf
900  *   Pointer to the devx_device_bdf structure.
901  * @param[in] addr
902  *   Pointer to the rte_pci_addr structure.
903  *
904  * @return
905  *   1 on Device match, 0 on mismatch, rte_errno code on failure.
906  */
907 static int
908 mlx5_match_devx_devices_to_addr(struct devx_device_bdf *devx_bdf,
909 				struct rte_pci_addr *addr)
910 {
911 	int err;
912 	struct devx_device mlx5_dev;
913 
914 	if (mlx5_match_devx_bdf_to_addr(devx_bdf, addr))
915 		return 1;
916 	/**
917 	 * Didn't match on Native/PF BDF, could still
918 	 * Match a VF BDF, check it next
919 	 */
920 	err = mlx5_glue->query_device(devx_bdf, &mlx5_dev);
921 	if (err) {
922 		DRV_LOG(ERR, "query_device failed");
923 		rte_errno = err;
924 		return rte_errno;
925 	}
926 	if (mlx5_match_devx_bdf_to_addr(&mlx5_dev.raw_bdf, addr))
927 		return 1;
928 	return 0;
929 }
930 
931 /**
932  * DPDK callback to register a PCI device.
933  *
934  * This function spawns Ethernet devices out of a given PCI device.
935  *
936  * @param[in] pci_drv
937  *   PCI driver structure (mlx5_driver).
938  * @param[in] pci_dev
939  *   PCI device information.
940  *
941  * @return
942  *   0 on success, a negative errno value otherwise and rte_errno is set.
943  */
944 int
945 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
946 		  struct rte_pci_device *pci_dev)
947 {
948 	struct devx_device_bdf *devx_bdf_devs, *orig_devx_bdf_devs;
949 	/*
950 	 * Number of found IB Devices matching with requested PCI BDF.
951 	 * nd != 1 means there are multiple IB devices over the same
952 	 * PCI device and we have representors and master.
953 	 */
954 	unsigned int nd = 0;
955 	/*
956 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
957 	 * we have the single multiport IB device, and there may be
958 	 * representors attached to some of found ports.
959 	 * Currently not supported.
960 	 * unsigned int np = 0;
961 	 */
962 
963 	/*
964 	 * Number of DPDK ethernet devices to Spawn - either over
965 	 * multiple IB devices or multiple ports of single IB device.
966 	 * Actually this is the number of iterations to spawn.
967 	 */
968 	unsigned int ns = 0;
969 	/*
970 	 * Bonding device
971 	 *   < 0 - no bonding device (single one)
972 	 *  >= 0 - bonding device (value is slave PF index)
973 	 */
974 	int bd = -1;
975 	struct mlx5_dev_spawn_data *list = NULL;
976 	struct mlx5_dev_config dev_config;
977 	unsigned int dev_config_vf;
978 	int ret, err;
979 	uint32_t restore;
980 
981 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
982 		DRV_LOG(ERR, "Secondary process is not supported on Windows.");
983 		return -ENOTSUP;
984 	}
985 	ret = mlx5_init_once();
986 	if (ret) {
987 		DRV_LOG(ERR, "unable to init PMD global data: %s",
988 			strerror(rte_errno));
989 		return -rte_errno;
990 	}
991 	errno = 0;
992 	devx_bdf_devs = mlx5_glue->get_device_list(&ret);
993 	orig_devx_bdf_devs = devx_bdf_devs;
994 	if (!devx_bdf_devs) {
995 		rte_errno = errno ? errno : ENOSYS;
996 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
997 		return -rte_errno;
998 	}
999 	/*
1000 	 * First scan the list of all Infiniband devices to find
1001 	 * matching ones, gathering into the list.
1002 	 */
1003 	struct devx_device_bdf *devx_bdf_match[ret + 1];
1004 
1005 	while (ret-- > 0) {
1006 		err = mlx5_match_devx_devices_to_addr(devx_bdf_devs,
1007 		    &pci_dev->addr);
1008 		if (!err) {
1009 			devx_bdf_devs++;
1010 			continue;
1011 		}
1012 		if (err != 1) {
1013 			ret = -err;
1014 			goto exit;
1015 		}
1016 		devx_bdf_match[nd++] = devx_bdf_devs;
1017 	}
1018 	devx_bdf_match[nd] = NULL;
1019 	if (!nd) {
1020 		/* No device matches, just complain and bail out. */
1021 		DRV_LOG(WARNING,
1022 			"no DevX device matches PCI device " PCI_PRI_FMT ","
1023 			" is DevX Configured?",
1024 			pci_dev->addr.domain, pci_dev->addr.bus,
1025 			pci_dev->addr.devid, pci_dev->addr.function);
1026 		rte_errno = ENOENT;
1027 		ret = -rte_errno;
1028 		goto exit;
1029 	}
1030 	/*
1031 	 * Now we can determine the maximal
1032 	 * amount of devices to be spawned.
1033 	 */
1034 	list = mlx5_malloc(MLX5_MEM_ZERO,
1035 			   sizeof(struct mlx5_dev_spawn_data),
1036 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1037 	if (!list) {
1038 		DRV_LOG(ERR, "spawn data array allocation failure");
1039 		rte_errno = ENOMEM;
1040 		ret = -rte_errno;
1041 		goto exit;
1042 	}
1043 	memset(&list[ns].info, 0, sizeof(list[ns].info));
1044 	list[ns].max_port = 1;
1045 	list[ns].phys_port = 1;
1046 	list[ns].phys_dev = devx_bdf_match[ns];
1047 	list[ns].eth_dev = NULL;
1048 	list[ns].pci_dev = pci_dev;
1049 	list[ns].pf_bond = bd;
1050 	list[ns].ifindex = -1; /* Spawn will assign */
1051 	list[ns].info =
1052 		(struct mlx5_switch_info){
1053 			.master = 0,
1054 			.representor = 0,
1055 			.name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK,
1056 			.port_name = 0,
1057 			.switch_id = 0,
1058 		};
1059 	/* Device specific configuration. */
1060 	switch (pci_dev->id.device_id) {
1061 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1062 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1063 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1064 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1065 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1066 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1067 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
1068 		dev_config_vf = 1;
1069 		break;
1070 	default:
1071 		dev_config_vf = 0;
1072 		break;
1073 	}
1074 	/* Default configuration. */
1075 	memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
1076 	dev_config.vf = dev_config_vf;
1077 	dev_config.mps = 0;
1078 	dev_config.dbnc = MLX5_ARG_UNSET;
1079 	dev_config.rx_vec_en = 1;
1080 	dev_config.txq_inline_max = MLX5_ARG_UNSET;
1081 	dev_config.txq_inline_min = MLX5_ARG_UNSET;
1082 	dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
1083 	dev_config.txqs_inline = MLX5_ARG_UNSET;
1084 	dev_config.vf_nl_en = 0;
1085 	dev_config.mr_ext_memseg_en = 1;
1086 	dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
1087 	dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
1088 	dev_config.dv_esw_en = 0;
1089 	dev_config.dv_flow_en = 1;
1090 	dev_config.decap_en = 0;
1091 	dev_config.log_hp_size = MLX5_ARG_UNSET;
1092 	list[ns].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1093 					  &list[ns],
1094 					  &dev_config);
1095 	if (!list[ns].eth_dev)
1096 		goto exit;
1097 	restore = list[ns].eth_dev->data->dev_flags;
1098 	rte_eth_copy_pci_info(list[ns].eth_dev, pci_dev);
1099 	/* Restore non-PCI flags cleared by the above call. */
1100 	list[ns].eth_dev->data->dev_flags |= restore;
1101 	rte_eth_dev_probing_finish(list[ns].eth_dev);
1102 	ret = 0;
1103 exit:
1104 	/*
1105 	 * Do the routine cleanup:
1106 	 * - free allocated spawn data array
1107 	 * - free the device list
1108 	 */
1109 	if (list)
1110 		mlx5_free(list);
1111 	MLX5_ASSERT(orig_devx_bdf_devs);
1112 	mlx5_glue->free_device_list(orig_devx_bdf_devs);
1113 	return ret;
1114 }
1115 
1116 /**
1117  * Set the reg_mr and dereg_mr call backs
1118  *
1119  * @param reg_mr_cb[out]
1120  *   Pointer to reg_mr func
1121  * @param dereg_mr_cb[out]
1122  *   Pointer to dereg_mr func
1123  *
1124  */
1125 void
1126 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
1127 		      mlx5_dereg_mr_t *dereg_mr_cb)
1128 {
1129 	*reg_mr_cb = mlx5_os_reg_mr;
1130 	*dereg_mr_cb = mlx5_os_dereg_mr;
1131 }
1132 
1133 /**
1134  * Extract pdn of PD object using DevX
1135  *
1136  * @param[in] pd
1137  *   Pointer to the DevX PD object.
1138  * @param[out] pdn
1139  *   Pointer to the PD object number variable.
1140  *
1141  * @return
1142  *   0 on success, error value otherwise.
1143  */
1144 int
1145 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1146 {
1147 	if (!pd)
1148 		return -EINVAL;
1149 
1150 	*pdn = ((struct mlx5_pd *)pd)->pdn;
1151 	return 0;
1152 }
1153 
1154 const struct mlx5_flow_driver_ops mlx5_flow_verbs_drv_ops = {0};
1155