xref: /linux/drivers/infiniband/hw/hns/hns_roce_main.c (revision 79d2e1919a2728ef49d938eb20ebd5903c14dfb0)
1 /*
2  * Copyright (c) 2016 Hisilicon Limited.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 #include <linux/acpi.h>
34 #include <linux/module.h>
35 #include <linux/pci.h>
36 #include <rdma/ib_addr.h>
37 #include <rdma/ib_smi.h>
38 #include <rdma/ib_user_verbs.h>
39 #include <rdma/ib_cache.h>
40 #include "hnae3.h"
41 #include "hns_roce_common.h"
42 #include "hns_roce_device.h"
43 #include "hns_roce_hem.h"
44 #include "hns_roce_hw_v2.h"
45 
46 static int hns_roce_set_mac(struct hns_roce_dev *hr_dev, u32 port,
47 			    const u8 *addr)
48 {
49 	u8 phy_port;
50 	u32 i;
51 
52 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
53 		return 0;
54 
55 	if (!memcmp(hr_dev->dev_addr[port], addr, ETH_ALEN))
56 		return 0;
57 
58 	for (i = 0; i < ETH_ALEN; i++)
59 		hr_dev->dev_addr[port][i] = addr[i];
60 
61 	phy_port = hr_dev->iboe.phy_port[port];
62 	return hr_dev->hw->set_mac(hr_dev, phy_port, addr);
63 }
64 
65 static int hns_roce_add_gid(const struct ib_gid_attr *attr, void **context)
66 {
67 	struct hns_roce_dev *hr_dev = to_hr_dev(attr->device);
68 	u32 port = attr->port_num - 1;
69 	int ret;
70 
71 	if (port >= hr_dev->caps.num_ports)
72 		return -EINVAL;
73 
74 	ret = hr_dev->hw->set_gid(hr_dev, attr->index, &attr->gid, attr);
75 
76 	return ret;
77 }
78 
79 static int hns_roce_del_gid(const struct ib_gid_attr *attr, void **context)
80 {
81 	struct hns_roce_dev *hr_dev = to_hr_dev(attr->device);
82 	u32 port = attr->port_num - 1;
83 	int ret;
84 
85 	if (port >= hr_dev->caps.num_ports)
86 		return -EINVAL;
87 
88 	ret = hr_dev->hw->set_gid(hr_dev, attr->index, NULL, NULL);
89 
90 	return ret;
91 }
92 
93 static int handle_en_event(struct hns_roce_dev *hr_dev, u32 port,
94 			   unsigned long event)
95 {
96 	struct device *dev = hr_dev->dev;
97 	struct net_device *netdev;
98 	int ret = 0;
99 
100 	netdev = hr_dev->iboe.netdevs[port];
101 	if (!netdev) {
102 		dev_err(dev, "can't find netdev on port(%u)!\n", port);
103 		return -ENODEV;
104 	}
105 
106 	switch (event) {
107 	case NETDEV_UP:
108 	case NETDEV_CHANGE:
109 	case NETDEV_REGISTER:
110 	case NETDEV_CHANGEADDR:
111 		ret = hns_roce_set_mac(hr_dev, port, netdev->dev_addr);
112 		break;
113 	case NETDEV_DOWN:
114 		/*
115 		 * In v1 engine, only support all ports closed together.
116 		 */
117 		break;
118 	default:
119 		dev_dbg(dev, "NETDEV event = 0x%x!\n", (u32)(event));
120 		break;
121 	}
122 
123 	return ret;
124 }
125 
126 static int hns_roce_netdev_event(struct notifier_block *self,
127 				 unsigned long event, void *ptr)
128 {
129 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
130 	struct hns_roce_ib_iboe *iboe = NULL;
131 	struct hns_roce_dev *hr_dev = NULL;
132 	int ret;
133 	u32 port;
134 
135 	hr_dev = container_of(self, struct hns_roce_dev, iboe.nb);
136 	iboe = &hr_dev->iboe;
137 
138 	for (port = 0; port < hr_dev->caps.num_ports; port++) {
139 		if (dev == iboe->netdevs[port]) {
140 			ret = handle_en_event(hr_dev, port, event);
141 			if (ret)
142 				return NOTIFY_DONE;
143 			break;
144 		}
145 	}
146 
147 	return NOTIFY_DONE;
148 }
149 
150 static int hns_roce_setup_mtu_mac(struct hns_roce_dev *hr_dev)
151 {
152 	int ret;
153 	u8 i;
154 
155 	for (i = 0; i < hr_dev->caps.num_ports; i++) {
156 		ret = hns_roce_set_mac(hr_dev, i,
157 				       hr_dev->iboe.netdevs[i]->dev_addr);
158 		if (ret)
159 			return ret;
160 	}
161 
162 	return 0;
163 }
164 
165 static int hns_roce_query_device(struct ib_device *ib_dev,
166 				 struct ib_device_attr *props,
167 				 struct ib_udata *uhw)
168 {
169 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev);
170 
171 	memset(props, 0, sizeof(*props));
172 
173 	props->fw_ver = hr_dev->caps.fw_ver;
174 	props->sys_image_guid = cpu_to_be64(hr_dev->sys_image_guid);
175 	props->max_mr_size = (u64)(~(0ULL));
176 	props->page_size_cap = hr_dev->caps.page_size_cap;
177 	props->vendor_id = hr_dev->vendor_id;
178 	props->vendor_part_id = hr_dev->vendor_part_id;
179 	props->hw_ver = hr_dev->hw_rev;
180 	props->max_qp = hr_dev->caps.num_qps;
181 	props->max_qp_wr = hr_dev->caps.max_wqes;
182 	props->device_cap_flags = IB_DEVICE_PORT_ACTIVE_EVENT |
183 				  IB_DEVICE_RC_RNR_NAK_GEN;
184 	props->max_send_sge = hr_dev->caps.max_sq_sg;
185 	props->max_recv_sge = hr_dev->caps.max_rq_sg;
186 	props->max_sge_rd = 1;
187 	props->max_cq = hr_dev->caps.num_cqs;
188 	props->max_cqe = hr_dev->caps.max_cqes;
189 	props->max_mr = hr_dev->caps.num_mtpts;
190 	props->max_pd = hr_dev->caps.num_pds;
191 	props->max_qp_rd_atom = hr_dev->caps.max_qp_dest_rdma;
192 	props->max_qp_init_rd_atom = hr_dev->caps.max_qp_init_rdma;
193 	props->atomic_cap = hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_ATOMIC ?
194 			    IB_ATOMIC_HCA : IB_ATOMIC_NONE;
195 	props->max_pkeys = 1;
196 	props->local_ca_ack_delay = hr_dev->caps.local_ca_ack_delay;
197 	props->max_ah = INT_MAX;
198 	props->cq_caps.max_cq_moderation_period = HNS_ROCE_MAX_CQ_PERIOD;
199 	props->cq_caps.max_cq_moderation_count = HNS_ROCE_MAX_CQ_COUNT;
200 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
201 		props->cq_caps.max_cq_moderation_period = HNS_ROCE_MAX_CQ_PERIOD_HIP08;
202 
203 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ) {
204 		props->max_srq = hr_dev->caps.num_srqs;
205 		props->max_srq_wr = hr_dev->caps.max_srq_wrs;
206 		props->max_srq_sge = hr_dev->caps.max_srq_sges;
207 	}
208 
209 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_FRMR &&
210 	    hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
211 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
212 		props->max_fast_reg_page_list_len = HNS_ROCE_FRMR_MAX_PA;
213 	}
214 
215 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC)
216 		props->device_cap_flags |= IB_DEVICE_XRC;
217 
218 	return 0;
219 }
220 
221 static int hns_roce_query_port(struct ib_device *ib_dev, u32 port_num,
222 			       struct ib_port_attr *props)
223 {
224 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev);
225 	struct device *dev = hr_dev->dev;
226 	struct net_device *net_dev;
227 	unsigned long flags;
228 	enum ib_mtu mtu;
229 	u32 port;
230 	int ret;
231 
232 	port = port_num - 1;
233 
234 	/* props being zeroed by the caller, avoid zeroing it here */
235 
236 	props->max_mtu = hr_dev->caps.max_mtu;
237 	props->gid_tbl_len = hr_dev->caps.gid_table_len[port];
238 	props->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
239 				IB_PORT_VENDOR_CLASS_SUP |
240 				IB_PORT_BOOT_MGMT_SUP;
241 	props->max_msg_sz = HNS_ROCE_MAX_MSG_LEN;
242 	props->pkey_tbl_len = 1;
243 	ret = ib_get_eth_speed(ib_dev, port_num, &props->active_speed,
244 			       &props->active_width);
245 	if (ret)
246 		ibdev_warn(ib_dev, "failed to get speed, ret = %d.\n", ret);
247 
248 	spin_lock_irqsave(&hr_dev->iboe.lock, flags);
249 
250 	net_dev = hr_dev->iboe.netdevs[port];
251 	if (!net_dev) {
252 		spin_unlock_irqrestore(&hr_dev->iboe.lock, flags);
253 		dev_err(dev, "find netdev %u failed!\n", port);
254 		return -EINVAL;
255 	}
256 
257 	mtu = iboe_get_mtu(net_dev->mtu);
258 	props->active_mtu = mtu ? min(props->max_mtu, mtu) : IB_MTU_256;
259 	props->state = netif_running(net_dev) && netif_carrier_ok(net_dev) ?
260 			       IB_PORT_ACTIVE :
261 			       IB_PORT_DOWN;
262 	props->phys_state = props->state == IB_PORT_ACTIVE ?
263 				    IB_PORT_PHYS_STATE_LINK_UP :
264 				    IB_PORT_PHYS_STATE_DISABLED;
265 
266 	spin_unlock_irqrestore(&hr_dev->iboe.lock, flags);
267 
268 	return 0;
269 }
270 
271 static enum rdma_link_layer hns_roce_get_link_layer(struct ib_device *device,
272 						    u32 port_num)
273 {
274 	return IB_LINK_LAYER_ETHERNET;
275 }
276 
277 static int hns_roce_query_pkey(struct ib_device *ib_dev, u32 port, u16 index,
278 			       u16 *pkey)
279 {
280 	if (index > 0)
281 		return -EINVAL;
282 
283 	*pkey = PKEY_ID;
284 
285 	return 0;
286 }
287 
288 static int hns_roce_modify_device(struct ib_device *ib_dev, int mask,
289 				  struct ib_device_modify *props)
290 {
291 	unsigned long flags;
292 
293 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
294 		return -EOPNOTSUPP;
295 
296 	if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
297 		spin_lock_irqsave(&to_hr_dev(ib_dev)->sm_lock, flags);
298 		memcpy(ib_dev->node_desc, props->node_desc, NODE_DESC_SIZE);
299 		spin_unlock_irqrestore(&to_hr_dev(ib_dev)->sm_lock, flags);
300 	}
301 
302 	return 0;
303 }
304 
305 struct hns_user_mmap_entry *
306 hns_roce_user_mmap_entry_insert(struct ib_ucontext *ucontext, u64 address,
307 				size_t length,
308 				enum hns_roce_mmap_type mmap_type)
309 {
310 	struct hns_user_mmap_entry *entry;
311 	int ret;
312 
313 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
314 	if (!entry)
315 		return NULL;
316 
317 	entry->address = address;
318 	entry->mmap_type = mmap_type;
319 
320 	switch (mmap_type) {
321 	/* pgoff 0 must be used by DB for compatibility */
322 	case HNS_ROCE_MMAP_TYPE_DB:
323 		ret = rdma_user_mmap_entry_insert_exact(
324 				ucontext, &entry->rdma_entry, length, 0);
325 		break;
326 	case HNS_ROCE_MMAP_TYPE_DWQE:
327 		ret = rdma_user_mmap_entry_insert_range(
328 				ucontext, &entry->rdma_entry, length, 1,
329 				U32_MAX);
330 		break;
331 	default:
332 		ret = -EINVAL;
333 		break;
334 	}
335 
336 	if (ret) {
337 		kfree(entry);
338 		return NULL;
339 	}
340 
341 	return entry;
342 }
343 
344 static void hns_roce_dealloc_uar_entry(struct hns_roce_ucontext *context)
345 {
346 	if (context->db_mmap_entry)
347 		rdma_user_mmap_entry_remove(
348 			&context->db_mmap_entry->rdma_entry);
349 }
350 
351 static int hns_roce_alloc_uar_entry(struct ib_ucontext *uctx)
352 {
353 	struct hns_roce_ucontext *context = to_hr_ucontext(uctx);
354 	u64 address;
355 
356 	address = context->uar.pfn << PAGE_SHIFT;
357 	context->db_mmap_entry = hns_roce_user_mmap_entry_insert(
358 		uctx, address, PAGE_SIZE, HNS_ROCE_MMAP_TYPE_DB);
359 	if (!context->db_mmap_entry)
360 		return -ENOMEM;
361 
362 	return 0;
363 }
364 
365 static int hns_roce_alloc_ucontext(struct ib_ucontext *uctx,
366 				   struct ib_udata *udata)
367 {
368 	struct hns_roce_ucontext *context = to_hr_ucontext(uctx);
369 	struct hns_roce_dev *hr_dev = to_hr_dev(uctx->device);
370 	struct hns_roce_ib_alloc_ucontext_resp resp = {};
371 	struct hns_roce_ib_alloc_ucontext ucmd = {};
372 	int ret = -EAGAIN;
373 
374 	if (!hr_dev->active)
375 		goto error_out;
376 
377 	resp.qp_tab_size = hr_dev->caps.num_qps;
378 	resp.srq_tab_size = hr_dev->caps.num_srqs;
379 
380 	ret = ib_copy_from_udata(&ucmd, udata,
381 				 min(udata->inlen, sizeof(ucmd)));
382 	if (ret)
383 		goto error_out;
384 
385 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
386 		context->config = ucmd.config & HNS_ROCE_EXSGE_FLAGS;
387 
388 	if (context->config & HNS_ROCE_EXSGE_FLAGS) {
389 		resp.config |= HNS_ROCE_RSP_EXSGE_FLAGS;
390 		resp.max_inline_data = hr_dev->caps.max_sq_inline;
391 	}
392 
393 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) {
394 		context->config |= ucmd.config & HNS_ROCE_RQ_INLINE_FLAGS;
395 		if (context->config & HNS_ROCE_RQ_INLINE_FLAGS)
396 			resp.config |= HNS_ROCE_RSP_RQ_INLINE_FLAGS;
397 	}
398 
399 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQE_INLINE) {
400 		context->config |= ucmd.config & HNS_ROCE_CQE_INLINE_FLAGS;
401 		if (context->config & HNS_ROCE_CQE_INLINE_FLAGS)
402 			resp.config |= HNS_ROCE_RSP_CQE_INLINE_FLAGS;
403 	}
404 
405 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
406 		resp.congest_type = hr_dev->caps.cong_cap;
407 
408 	ret = hns_roce_uar_alloc(hr_dev, &context->uar);
409 	if (ret)
410 		goto error_out;
411 
412 	ret = hns_roce_alloc_uar_entry(uctx);
413 	if (ret)
414 		goto error_fail_uar_entry;
415 
416 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB ||
417 	    hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) {
418 		INIT_LIST_HEAD(&context->page_list);
419 		mutex_init(&context->page_mutex);
420 	}
421 
422 	resp.cqe_size = hr_dev->caps.cqe_sz;
423 
424 	ret = ib_copy_to_udata(udata, &resp,
425 			       min(udata->outlen, sizeof(resp)));
426 	if (ret)
427 		goto error_fail_copy_to_udata;
428 
429 	return 0;
430 
431 error_fail_copy_to_udata:
432 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB ||
433 	    hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB)
434 		mutex_destroy(&context->page_mutex);
435 	hns_roce_dealloc_uar_entry(context);
436 
437 error_fail_uar_entry:
438 	ida_free(&hr_dev->uar_ida.ida, (int)context->uar.logic_idx);
439 
440 error_out:
441 	atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_UCTX_ALLOC_ERR_CNT]);
442 
443 	return ret;
444 }
445 
446 static void hns_roce_dealloc_ucontext(struct ib_ucontext *ibcontext)
447 {
448 	struct hns_roce_ucontext *context = to_hr_ucontext(ibcontext);
449 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcontext->device);
450 
451 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB ||
452 	    hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB)
453 		mutex_destroy(&context->page_mutex);
454 
455 	hns_roce_dealloc_uar_entry(context);
456 
457 	ida_free(&hr_dev->uar_ida.ida, (int)context->uar.logic_idx);
458 }
459 
460 static int hns_roce_mmap(struct ib_ucontext *uctx, struct vm_area_struct *vma)
461 {
462 	struct hns_roce_dev *hr_dev = to_hr_dev(uctx->device);
463 	struct rdma_user_mmap_entry *rdma_entry;
464 	struct hns_user_mmap_entry *entry;
465 	phys_addr_t pfn;
466 	pgprot_t prot;
467 	int ret;
468 
469 	if (hr_dev->dis_db) {
470 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_MMAP_ERR_CNT]);
471 		return -EPERM;
472 	}
473 
474 	rdma_entry = rdma_user_mmap_entry_get_pgoff(uctx, vma->vm_pgoff);
475 	if (!rdma_entry) {
476 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_MMAP_ERR_CNT]);
477 		return -EINVAL;
478 	}
479 
480 	entry = to_hns_mmap(rdma_entry);
481 	pfn = entry->address >> PAGE_SHIFT;
482 
483 	switch (entry->mmap_type) {
484 	case HNS_ROCE_MMAP_TYPE_DB:
485 	case HNS_ROCE_MMAP_TYPE_DWQE:
486 		prot = pgprot_device(vma->vm_page_prot);
487 		break;
488 	default:
489 		ret = -EINVAL;
490 		goto out;
491 	}
492 
493 	ret = rdma_user_mmap_io(uctx, vma, pfn, rdma_entry->npages * PAGE_SIZE,
494 				prot, rdma_entry);
495 
496 out:
497 	rdma_user_mmap_entry_put(rdma_entry);
498 	if (ret)
499 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_MMAP_ERR_CNT]);
500 
501 	return ret;
502 }
503 
504 static void hns_roce_free_mmap(struct rdma_user_mmap_entry *rdma_entry)
505 {
506 	struct hns_user_mmap_entry *entry = to_hns_mmap(rdma_entry);
507 
508 	kfree(entry);
509 }
510 
511 static int hns_roce_port_immutable(struct ib_device *ib_dev, u32 port_num,
512 				   struct ib_port_immutable *immutable)
513 {
514 	struct ib_port_attr attr;
515 	int ret;
516 
517 	ret = ib_query_port(ib_dev, port_num, &attr);
518 	if (ret)
519 		return ret;
520 
521 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
522 	immutable->gid_tbl_len = attr.gid_tbl_len;
523 
524 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
525 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
526 	if (to_hr_dev(ib_dev)->caps.flags & HNS_ROCE_CAP_FLAG_ROCE_V1_V2)
527 		immutable->core_cap_flags |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
528 
529 	return 0;
530 }
531 
532 static void hns_roce_disassociate_ucontext(struct ib_ucontext *ibcontext)
533 {
534 }
535 
536 static void hns_roce_get_fw_ver(struct ib_device *device, char *str)
537 {
538 	u64 fw_ver = to_hr_dev(device)->caps.fw_ver;
539 	unsigned int major, minor, sub_minor;
540 
541 	major = upper_32_bits(fw_ver);
542 	minor = high_16_bits(lower_32_bits(fw_ver));
543 	sub_minor = low_16_bits(fw_ver);
544 
545 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u.%04u", major, minor,
546 		 sub_minor);
547 }
548 
549 #define HNS_ROCE_HW_CNT(ename, cname) \
550 	[HNS_ROCE_HW_##ename##_CNT].name = cname
551 
552 static const struct rdma_stat_desc hns_roce_port_stats_descs[] = {
553 	HNS_ROCE_HW_CNT(RX_RC_PKT, "rx_rc_pkt"),
554 	HNS_ROCE_HW_CNT(RX_UC_PKT, "rx_uc_pkt"),
555 	HNS_ROCE_HW_CNT(RX_UD_PKT, "rx_ud_pkt"),
556 	HNS_ROCE_HW_CNT(RX_XRC_PKT, "rx_xrc_pkt"),
557 	HNS_ROCE_HW_CNT(RX_PKT, "rx_pkt"),
558 	HNS_ROCE_HW_CNT(RX_ERR_PKT, "rx_err_pkt"),
559 	HNS_ROCE_HW_CNT(RX_CNP_PKT, "rx_cnp_pkt"),
560 	HNS_ROCE_HW_CNT(TX_RC_PKT, "tx_rc_pkt"),
561 	HNS_ROCE_HW_CNT(TX_UC_PKT, "tx_uc_pkt"),
562 	HNS_ROCE_HW_CNT(TX_UD_PKT, "tx_ud_pkt"),
563 	HNS_ROCE_HW_CNT(TX_XRC_PKT, "tx_xrc_pkt"),
564 	HNS_ROCE_HW_CNT(TX_PKT, "tx_pkt"),
565 	HNS_ROCE_HW_CNT(TX_ERR_PKT, "tx_err_pkt"),
566 	HNS_ROCE_HW_CNT(TX_CNP_PKT, "tx_cnp_pkt"),
567 	HNS_ROCE_HW_CNT(TRP_GET_MPT_ERR_PKT, "trp_get_mpt_err_pkt"),
568 	HNS_ROCE_HW_CNT(TRP_GET_IRRL_ERR_PKT, "trp_get_irrl_err_pkt"),
569 	HNS_ROCE_HW_CNT(ECN_DB, "ecn_doorbell"),
570 	HNS_ROCE_HW_CNT(RX_BUF, "rx_buffer"),
571 	HNS_ROCE_HW_CNT(TRP_RX_SOF, "trp_rx_sof"),
572 	HNS_ROCE_HW_CNT(CQ_CQE, "cq_cqe"),
573 	HNS_ROCE_HW_CNT(CQ_POE, "cq_poe"),
574 	HNS_ROCE_HW_CNT(CQ_NOTIFY, "cq_notify"),
575 };
576 
577 static struct rdma_hw_stats *hns_roce_alloc_hw_port_stats(
578 				struct ib_device *device, u32 port_num)
579 {
580 	struct hns_roce_dev *hr_dev = to_hr_dev(device);
581 
582 	if (port_num > hr_dev->caps.num_ports) {
583 		ibdev_err(device, "invalid port num.\n");
584 		return NULL;
585 	}
586 
587 	return rdma_alloc_hw_stats_struct(hns_roce_port_stats_descs,
588 					  ARRAY_SIZE(hns_roce_port_stats_descs),
589 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
590 }
591 
592 static int hns_roce_get_hw_stats(struct ib_device *device,
593 				 struct rdma_hw_stats *stats,
594 				 u32 port, int index)
595 {
596 	struct hns_roce_dev *hr_dev = to_hr_dev(device);
597 	int num_counters = HNS_ROCE_HW_CNT_TOTAL;
598 	int ret;
599 
600 	if (port == 0)
601 		return 0;
602 
603 	if (port > hr_dev->caps.num_ports)
604 		return -EINVAL;
605 
606 	ret = hr_dev->hw->query_hw_counter(hr_dev, stats->value, port,
607 					   &num_counters);
608 	if (ret) {
609 		ibdev_err(device, "failed to query hw counter, ret = %d\n",
610 			  ret);
611 		return ret;
612 	}
613 
614 	return num_counters;
615 }
616 
617 static void hns_roce_unregister_device(struct hns_roce_dev *hr_dev)
618 {
619 	struct hns_roce_ib_iboe *iboe = &hr_dev->iboe;
620 
621 	hr_dev->active = false;
622 	unregister_netdevice_notifier(&iboe->nb);
623 	ib_unregister_device(&hr_dev->ib_dev);
624 }
625 
626 static const struct ib_device_ops hns_roce_dev_ops = {
627 	.owner = THIS_MODULE,
628 	.driver_id = RDMA_DRIVER_HNS,
629 	.uverbs_abi_ver = 1,
630 	.uverbs_no_driver_id_binding = 1,
631 
632 	.get_dev_fw_str = hns_roce_get_fw_ver,
633 	.add_gid = hns_roce_add_gid,
634 	.alloc_pd = hns_roce_alloc_pd,
635 	.alloc_ucontext = hns_roce_alloc_ucontext,
636 	.create_ah = hns_roce_create_ah,
637 	.create_user_ah = hns_roce_create_ah,
638 	.create_cq = hns_roce_create_cq,
639 	.create_qp = hns_roce_create_qp,
640 	.dealloc_pd = hns_roce_dealloc_pd,
641 	.dealloc_ucontext = hns_roce_dealloc_ucontext,
642 	.del_gid = hns_roce_del_gid,
643 	.dereg_mr = hns_roce_dereg_mr,
644 	.destroy_ah = hns_roce_destroy_ah,
645 	.destroy_cq = hns_roce_destroy_cq,
646 	.disassociate_ucontext = hns_roce_disassociate_ucontext,
647 	.get_dma_mr = hns_roce_get_dma_mr,
648 	.get_link_layer = hns_roce_get_link_layer,
649 	.get_port_immutable = hns_roce_port_immutable,
650 	.mmap = hns_roce_mmap,
651 	.mmap_free = hns_roce_free_mmap,
652 	.modify_device = hns_roce_modify_device,
653 	.modify_qp = hns_roce_modify_qp,
654 	.query_ah = hns_roce_query_ah,
655 	.query_device = hns_roce_query_device,
656 	.query_pkey = hns_roce_query_pkey,
657 	.query_port = hns_roce_query_port,
658 	.reg_user_mr = hns_roce_reg_user_mr,
659 
660 	INIT_RDMA_OBJ_SIZE(ib_ah, hns_roce_ah, ibah),
661 	INIT_RDMA_OBJ_SIZE(ib_cq, hns_roce_cq, ib_cq),
662 	INIT_RDMA_OBJ_SIZE(ib_pd, hns_roce_pd, ibpd),
663 	INIT_RDMA_OBJ_SIZE(ib_qp, hns_roce_qp, ibqp),
664 	INIT_RDMA_OBJ_SIZE(ib_ucontext, hns_roce_ucontext, ibucontext),
665 };
666 
667 static const struct ib_device_ops hns_roce_dev_hw_stats_ops = {
668 	.alloc_hw_port_stats = hns_roce_alloc_hw_port_stats,
669 	.get_hw_stats = hns_roce_get_hw_stats,
670 };
671 
672 static const struct ib_device_ops hns_roce_dev_mr_ops = {
673 	.rereg_user_mr = hns_roce_rereg_user_mr,
674 };
675 
676 static const struct ib_device_ops hns_roce_dev_mw_ops = {
677 	.alloc_mw = hns_roce_alloc_mw,
678 	.dealloc_mw = hns_roce_dealloc_mw,
679 
680 	INIT_RDMA_OBJ_SIZE(ib_mw, hns_roce_mw, ibmw),
681 };
682 
683 static const struct ib_device_ops hns_roce_dev_frmr_ops = {
684 	.alloc_mr = hns_roce_alloc_mr,
685 	.map_mr_sg = hns_roce_map_mr_sg,
686 };
687 
688 static const struct ib_device_ops hns_roce_dev_srq_ops = {
689 	.create_srq = hns_roce_create_srq,
690 	.destroy_srq = hns_roce_destroy_srq,
691 
692 	INIT_RDMA_OBJ_SIZE(ib_srq, hns_roce_srq, ibsrq),
693 };
694 
695 static const struct ib_device_ops hns_roce_dev_xrcd_ops = {
696 	.alloc_xrcd = hns_roce_alloc_xrcd,
697 	.dealloc_xrcd = hns_roce_dealloc_xrcd,
698 
699 	INIT_RDMA_OBJ_SIZE(ib_xrcd, hns_roce_xrcd, ibxrcd),
700 };
701 
702 static const struct ib_device_ops hns_roce_dev_restrack_ops = {
703 	.fill_res_cq_entry = hns_roce_fill_res_cq_entry,
704 	.fill_res_cq_entry_raw = hns_roce_fill_res_cq_entry_raw,
705 	.fill_res_qp_entry = hns_roce_fill_res_qp_entry,
706 	.fill_res_qp_entry_raw = hns_roce_fill_res_qp_entry_raw,
707 	.fill_res_mr_entry = hns_roce_fill_res_mr_entry,
708 	.fill_res_mr_entry_raw = hns_roce_fill_res_mr_entry_raw,
709 	.fill_res_srq_entry = hns_roce_fill_res_srq_entry,
710 	.fill_res_srq_entry_raw = hns_roce_fill_res_srq_entry_raw,
711 };
712 
713 static int hns_roce_register_device(struct hns_roce_dev *hr_dev)
714 {
715 	int ret;
716 	struct hns_roce_ib_iboe *iboe = NULL;
717 	struct ib_device *ib_dev = NULL;
718 	struct device *dev = hr_dev->dev;
719 	unsigned int i;
720 
721 	iboe = &hr_dev->iboe;
722 	spin_lock_init(&iboe->lock);
723 
724 	ib_dev = &hr_dev->ib_dev;
725 
726 	ib_dev->node_type = RDMA_NODE_IB_CA;
727 	ib_dev->dev.parent = dev;
728 
729 	ib_dev->phys_port_cnt = hr_dev->caps.num_ports;
730 	ib_dev->local_dma_lkey = hr_dev->caps.reserved_lkey;
731 	ib_dev->num_comp_vectors = hr_dev->caps.num_comp_vectors;
732 
733 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_REREG_MR)
734 		ib_set_device_ops(ib_dev, &hns_roce_dev_mr_ops);
735 
736 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_MW)
737 		ib_set_device_ops(ib_dev, &hns_roce_dev_mw_ops);
738 
739 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_FRMR)
740 		ib_set_device_ops(ib_dev, &hns_roce_dev_frmr_ops);
741 
742 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ) {
743 		ib_set_device_ops(ib_dev, &hns_roce_dev_srq_ops);
744 		ib_set_device_ops(ib_dev, hr_dev->hw->hns_roce_dev_srq_ops);
745 	}
746 
747 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC)
748 		ib_set_device_ops(ib_dev, &hns_roce_dev_xrcd_ops);
749 
750 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09 &&
751 	    !hr_dev->is_vf)
752 		ib_set_device_ops(ib_dev, &hns_roce_dev_hw_stats_ops);
753 
754 	ib_set_device_ops(ib_dev, hr_dev->hw->hns_roce_dev_ops);
755 	ib_set_device_ops(ib_dev, &hns_roce_dev_ops);
756 	ib_set_device_ops(ib_dev, &hns_roce_dev_restrack_ops);
757 	for (i = 0; i < hr_dev->caps.num_ports; i++) {
758 		if (!hr_dev->iboe.netdevs[i])
759 			continue;
760 
761 		ret = ib_device_set_netdev(ib_dev, hr_dev->iboe.netdevs[i],
762 					   i + 1);
763 		if (ret)
764 			return ret;
765 	}
766 	dma_set_max_seg_size(dev, UINT_MAX);
767 	ret = ib_register_device(ib_dev, "hns_%d", dev);
768 	if (ret) {
769 		dev_err(dev, "ib_register_device failed!\n");
770 		return ret;
771 	}
772 
773 	ret = hns_roce_setup_mtu_mac(hr_dev);
774 	if (ret) {
775 		dev_err(dev, "setup_mtu_mac failed!\n");
776 		goto error_failed_setup_mtu_mac;
777 	}
778 
779 	iboe->nb.notifier_call = hns_roce_netdev_event;
780 	ret = register_netdevice_notifier(&iboe->nb);
781 	if (ret) {
782 		dev_err(dev, "register_netdevice_notifier failed!\n");
783 		goto error_failed_setup_mtu_mac;
784 	}
785 
786 	hr_dev->active = true;
787 	return 0;
788 
789 error_failed_setup_mtu_mac:
790 	ib_unregister_device(ib_dev);
791 
792 	return ret;
793 }
794 
795 static int hns_roce_init_hem(struct hns_roce_dev *hr_dev)
796 {
797 	struct device *dev = hr_dev->dev;
798 	int ret;
799 
800 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table,
801 				      HEM_TYPE_MTPT, hr_dev->caps.mtpt_entry_sz,
802 				      hr_dev->caps.num_mtpts);
803 	if (ret) {
804 		dev_err(dev, "failed to init MTPT context memory, aborting.\n");
805 		return ret;
806 	}
807 
808 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->qp_table.qp_table,
809 				      HEM_TYPE_QPC, hr_dev->caps.qpc_sz,
810 				      hr_dev->caps.num_qps);
811 	if (ret) {
812 		dev_err(dev, "failed to init QP context memory, aborting.\n");
813 		goto err_unmap_dmpt;
814 	}
815 
816 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->qp_table.irrl_table,
817 				      HEM_TYPE_IRRL,
818 				      hr_dev->caps.irrl_entry_sz *
819 				      hr_dev->caps.max_qp_init_rdma,
820 				      hr_dev->caps.num_qps);
821 	if (ret) {
822 		dev_err(dev, "failed to init irrl_table memory, aborting.\n");
823 		goto err_unmap_qp;
824 	}
825 
826 	if (hr_dev->caps.trrl_entry_sz) {
827 		ret = hns_roce_init_hem_table(hr_dev,
828 					      &hr_dev->qp_table.trrl_table,
829 					      HEM_TYPE_TRRL,
830 					      hr_dev->caps.trrl_entry_sz *
831 					      hr_dev->caps.max_qp_dest_rdma,
832 					      hr_dev->caps.num_qps);
833 		if (ret) {
834 			dev_err(dev,
835 				"failed to init trrl_table memory, aborting.\n");
836 			goto err_unmap_irrl;
837 		}
838 	}
839 
840 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->cq_table.table,
841 				      HEM_TYPE_CQC, hr_dev->caps.cqc_entry_sz,
842 				      hr_dev->caps.num_cqs);
843 	if (ret) {
844 		dev_err(dev, "failed to init CQ context memory, aborting.\n");
845 		goto err_unmap_trrl;
846 	}
847 
848 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ) {
849 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->srq_table.table,
850 					      HEM_TYPE_SRQC,
851 					      hr_dev->caps.srqc_entry_sz,
852 					      hr_dev->caps.num_srqs);
853 		if (ret) {
854 			dev_err(dev,
855 				"failed to init SRQ context memory, aborting.\n");
856 			goto err_unmap_cq;
857 		}
858 	}
859 
860 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
861 		ret = hns_roce_init_hem_table(hr_dev,
862 					      &hr_dev->qp_table.sccc_table,
863 					      HEM_TYPE_SCCC,
864 					      hr_dev->caps.sccc_sz,
865 					      hr_dev->caps.num_qps);
866 		if (ret) {
867 			dev_err(dev,
868 				"failed to init SCC context memory, aborting.\n");
869 			goto err_unmap_srq;
870 		}
871 	}
872 
873 	if (hr_dev->caps.qpc_timer_entry_sz) {
874 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->qpc_timer_table,
875 					      HEM_TYPE_QPC_TIMER,
876 					      hr_dev->caps.qpc_timer_entry_sz,
877 					      hr_dev->caps.qpc_timer_bt_num);
878 		if (ret) {
879 			dev_err(dev,
880 				"failed to init QPC timer memory, aborting.\n");
881 			goto err_unmap_ctx;
882 		}
883 	}
884 
885 	if (hr_dev->caps.cqc_timer_entry_sz) {
886 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->cqc_timer_table,
887 					      HEM_TYPE_CQC_TIMER,
888 					      hr_dev->caps.cqc_timer_entry_sz,
889 					      hr_dev->caps.cqc_timer_bt_num);
890 		if (ret) {
891 			dev_err(dev,
892 				"failed to init CQC timer memory, aborting.\n");
893 			goto err_unmap_qpc_timer;
894 		}
895 	}
896 
897 	if (hr_dev->caps.gmv_entry_sz) {
898 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->gmv_table,
899 					      HEM_TYPE_GMV,
900 					      hr_dev->caps.gmv_entry_sz,
901 					      hr_dev->caps.gmv_entry_num);
902 		if (ret) {
903 			dev_err(dev,
904 				"failed to init gmv table memory, ret = %d\n",
905 				ret);
906 			goto err_unmap_cqc_timer;
907 		}
908 	}
909 
910 	return 0;
911 
912 err_unmap_cqc_timer:
913 	if (hr_dev->caps.cqc_timer_entry_sz)
914 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cqc_timer_table);
915 
916 err_unmap_qpc_timer:
917 	if (hr_dev->caps.qpc_timer_entry_sz)
918 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qpc_timer_table);
919 
920 err_unmap_ctx:
921 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
922 		hns_roce_cleanup_hem_table(hr_dev,
923 					   &hr_dev->qp_table.sccc_table);
924 err_unmap_srq:
925 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
926 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->srq_table.table);
927 
928 err_unmap_cq:
929 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cq_table.table);
930 
931 err_unmap_trrl:
932 	if (hr_dev->caps.trrl_entry_sz)
933 		hns_roce_cleanup_hem_table(hr_dev,
934 					   &hr_dev->qp_table.trrl_table);
935 
936 err_unmap_irrl:
937 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.irrl_table);
938 
939 err_unmap_qp:
940 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.qp_table);
941 
942 err_unmap_dmpt:
943 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table);
944 
945 	return ret;
946 }
947 
948 static void hns_roce_teardown_hca(struct hns_roce_dev *hr_dev)
949 {
950 	hns_roce_cleanup_bitmap(hr_dev);
951 
952 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB ||
953 	    hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB)
954 		mutex_destroy(&hr_dev->pgdir_mutex);
955 }
956 
957 /**
958  * hns_roce_setup_hca - setup host channel adapter
959  * @hr_dev: pointer to hns roce device
960  * Return : int
961  */
962 static int hns_roce_setup_hca(struct hns_roce_dev *hr_dev)
963 {
964 	struct device *dev = hr_dev->dev;
965 	int ret;
966 
967 	spin_lock_init(&hr_dev->sm_lock);
968 
969 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB ||
970 	    hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) {
971 		INIT_LIST_HEAD(&hr_dev->pgdir_list);
972 		mutex_init(&hr_dev->pgdir_mutex);
973 	}
974 
975 	hns_roce_init_uar_table(hr_dev);
976 
977 	ret = hns_roce_uar_alloc(hr_dev, &hr_dev->priv_uar);
978 	if (ret) {
979 		dev_err(dev, "failed to allocate priv_uar.\n");
980 		goto err_uar_table_free;
981 	}
982 
983 	ret = hns_roce_init_qp_table(hr_dev);
984 	if (ret) {
985 		dev_err(dev, "failed to init qp_table.\n");
986 		goto err_uar_table_free;
987 	}
988 
989 	hns_roce_init_pd_table(hr_dev);
990 
991 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC)
992 		hns_roce_init_xrcd_table(hr_dev);
993 
994 	hns_roce_init_mr_table(hr_dev);
995 
996 	hns_roce_init_cq_table(hr_dev);
997 
998 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
999 		hns_roce_init_srq_table(hr_dev);
1000 
1001 	return 0;
1002 
1003 err_uar_table_free:
1004 	ida_destroy(&hr_dev->uar_ida.ida);
1005 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB ||
1006 	    hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB)
1007 		mutex_destroy(&hr_dev->pgdir_mutex);
1008 
1009 	return ret;
1010 }
1011 
1012 static void check_and_get_armed_cq(struct list_head *cq_list, struct ib_cq *cq)
1013 {
1014 	struct hns_roce_cq *hr_cq = to_hr_cq(cq);
1015 	unsigned long flags;
1016 
1017 	spin_lock_irqsave(&hr_cq->lock, flags);
1018 	if (cq->comp_handler) {
1019 		if (!hr_cq->is_armed) {
1020 			hr_cq->is_armed = 1;
1021 			list_add_tail(&hr_cq->node, cq_list);
1022 		}
1023 	}
1024 	spin_unlock_irqrestore(&hr_cq->lock, flags);
1025 }
1026 
1027 void hns_roce_handle_device_err(struct hns_roce_dev *hr_dev)
1028 {
1029 	struct hns_roce_qp *hr_qp;
1030 	struct hns_roce_cq *hr_cq;
1031 	struct list_head cq_list;
1032 	unsigned long flags_qp;
1033 	unsigned long flags;
1034 
1035 	INIT_LIST_HEAD(&cq_list);
1036 
1037 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
1038 	list_for_each_entry(hr_qp, &hr_dev->qp_list, node) {
1039 		spin_lock_irqsave(&hr_qp->sq.lock, flags_qp);
1040 		if (hr_qp->sq.tail != hr_qp->sq.head)
1041 			check_and_get_armed_cq(&cq_list, hr_qp->ibqp.send_cq);
1042 		spin_unlock_irqrestore(&hr_qp->sq.lock, flags_qp);
1043 
1044 		spin_lock_irqsave(&hr_qp->rq.lock, flags_qp);
1045 		if ((!hr_qp->ibqp.srq) && (hr_qp->rq.tail != hr_qp->rq.head))
1046 			check_and_get_armed_cq(&cq_list, hr_qp->ibqp.recv_cq);
1047 		spin_unlock_irqrestore(&hr_qp->rq.lock, flags_qp);
1048 	}
1049 
1050 	list_for_each_entry(hr_cq, &cq_list, node)
1051 		hns_roce_cq_completion(hr_dev, hr_cq->cqn);
1052 
1053 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
1054 }
1055 
1056 static int hns_roce_alloc_dfx_cnt(struct hns_roce_dev *hr_dev)
1057 {
1058 	hr_dev->dfx_cnt = kvcalloc(HNS_ROCE_DFX_CNT_TOTAL, sizeof(atomic64_t),
1059 				   GFP_KERNEL);
1060 	if (!hr_dev->dfx_cnt)
1061 		return -ENOMEM;
1062 
1063 	return 0;
1064 }
1065 
1066 static void hns_roce_dealloc_dfx_cnt(struct hns_roce_dev *hr_dev)
1067 {
1068 	kvfree(hr_dev->dfx_cnt);
1069 }
1070 
1071 int hns_roce_init(struct hns_roce_dev *hr_dev)
1072 {
1073 	struct device *dev = hr_dev->dev;
1074 	int ret;
1075 
1076 	hr_dev->is_reset = false;
1077 
1078 	ret = hns_roce_alloc_dfx_cnt(hr_dev);
1079 	if (ret)
1080 		return ret;
1081 
1082 	if (hr_dev->hw->cmq_init) {
1083 		ret = hr_dev->hw->cmq_init(hr_dev);
1084 		if (ret) {
1085 			dev_err(dev, "init RoCE Command Queue failed!\n");
1086 			goto error_failed_alloc_dfx_cnt;
1087 		}
1088 	}
1089 
1090 	ret = hr_dev->hw->hw_profile(hr_dev);
1091 	if (ret) {
1092 		dev_err(dev, "get RoCE engine profile failed!\n");
1093 		goto error_failed_cmd_init;
1094 	}
1095 
1096 	ret = hns_roce_cmd_init(hr_dev);
1097 	if (ret) {
1098 		dev_err(dev, "cmd init failed!\n");
1099 		goto error_failed_cmd_init;
1100 	}
1101 
1102 	/* EQ depends on poll mode, event mode depends on EQ */
1103 	ret = hr_dev->hw->init_eq(hr_dev);
1104 	if (ret) {
1105 		dev_err(dev, "eq init failed!\n");
1106 		goto error_failed_eq_table;
1107 	}
1108 
1109 	if (hr_dev->cmd_mod) {
1110 		ret = hns_roce_cmd_use_events(hr_dev);
1111 		if (ret)
1112 			dev_warn(dev,
1113 				 "Cmd event  mode failed, set back to poll!\n");
1114 	}
1115 
1116 	ret = hns_roce_init_hem(hr_dev);
1117 	if (ret) {
1118 		dev_err(dev, "init HEM(Hardware Entry Memory) failed!\n");
1119 		goto error_failed_init_hem;
1120 	}
1121 
1122 	ret = hns_roce_setup_hca(hr_dev);
1123 	if (ret) {
1124 		dev_err(dev, "setup hca failed!\n");
1125 		goto error_failed_setup_hca;
1126 	}
1127 
1128 	if (hr_dev->hw->hw_init) {
1129 		ret = hr_dev->hw->hw_init(hr_dev);
1130 		if (ret) {
1131 			dev_err(dev, "hw_init failed!\n");
1132 			goto error_failed_engine_init;
1133 		}
1134 	}
1135 
1136 	INIT_LIST_HEAD(&hr_dev->qp_list);
1137 	spin_lock_init(&hr_dev->qp_list_lock);
1138 
1139 	ret = hns_roce_register_device(hr_dev);
1140 	if (ret)
1141 		goto error_failed_register_device;
1142 
1143 	hns_roce_register_debugfs(hr_dev);
1144 
1145 	return 0;
1146 
1147 error_failed_register_device:
1148 	if (hr_dev->hw->hw_exit)
1149 		hr_dev->hw->hw_exit(hr_dev);
1150 
1151 error_failed_engine_init:
1152 	hns_roce_teardown_hca(hr_dev);
1153 
1154 error_failed_setup_hca:
1155 	hns_roce_cleanup_hem(hr_dev);
1156 
1157 error_failed_init_hem:
1158 	if (hr_dev->cmd_mod)
1159 		hns_roce_cmd_use_polling(hr_dev);
1160 	hr_dev->hw->cleanup_eq(hr_dev);
1161 
1162 error_failed_eq_table:
1163 	hns_roce_cmd_cleanup(hr_dev);
1164 
1165 error_failed_cmd_init:
1166 	if (hr_dev->hw->cmq_exit)
1167 		hr_dev->hw->cmq_exit(hr_dev);
1168 
1169 error_failed_alloc_dfx_cnt:
1170 	hns_roce_dealloc_dfx_cnt(hr_dev);
1171 
1172 	return ret;
1173 }
1174 
1175 void hns_roce_exit(struct hns_roce_dev *hr_dev)
1176 {
1177 	hns_roce_unregister_debugfs(hr_dev);
1178 	hns_roce_unregister_device(hr_dev);
1179 
1180 	if (hr_dev->hw->hw_exit)
1181 		hr_dev->hw->hw_exit(hr_dev);
1182 	hns_roce_teardown_hca(hr_dev);
1183 	hns_roce_cleanup_hem(hr_dev);
1184 
1185 	if (hr_dev->cmd_mod)
1186 		hns_roce_cmd_use_polling(hr_dev);
1187 
1188 	hr_dev->hw->cleanup_eq(hr_dev);
1189 	hns_roce_cmd_cleanup(hr_dev);
1190 	if (hr_dev->hw->cmq_exit)
1191 		hr_dev->hw->cmq_exit(hr_dev);
1192 	hns_roce_dealloc_dfx_cnt(hr_dev);
1193 }
1194 
1195 MODULE_LICENSE("Dual BSD/GPL");
1196 MODULE_AUTHOR("Wei Hu <xavier.huwei@huawei.com>");
1197 MODULE_AUTHOR("Nenglong Zhao <zhaonenglong@hisilicon.com>");
1198 MODULE_AUTHOR("Lijun Ou <oulijun@huawei.com>");
1199 MODULE_DESCRIPTION("HNS RoCE Driver");
1200