xref: /openbmc/linux/drivers/infiniband/hw/hns/hns_roce_main.c (revision df1bcf90a66a10967a3a43510b42cb3566208011)
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 "hns_roce_common.h"
41 #include "hns_roce_device.h"
42 #include "hns_roce_hem.h"
43 
44 static int hns_roce_set_mac(struct hns_roce_dev *hr_dev, u32 port,
45 			    const u8 *addr)
46 {
47 	u8 phy_port;
48 	u32 i;
49 
50 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
51 		return 0;
52 
53 	if (!memcmp(hr_dev->dev_addr[port], addr, ETH_ALEN))
54 		return 0;
55 
56 	for (i = 0; i < ETH_ALEN; i++)
57 		hr_dev->dev_addr[port][i] = addr[i];
58 
59 	phy_port = hr_dev->iboe.phy_port[port];
60 	return hr_dev->hw->set_mac(hr_dev, phy_port, addr);
61 }
62 
63 static int hns_roce_add_gid(const struct ib_gid_attr *attr, void **context)
64 {
65 	struct hns_roce_dev *hr_dev = to_hr_dev(attr->device);
66 	u32 port = attr->port_num - 1;
67 	int ret;
68 
69 	if (port >= hr_dev->caps.num_ports)
70 		return -EINVAL;
71 
72 	ret = hr_dev->hw->set_gid(hr_dev, attr->index, &attr->gid, attr);
73 
74 	return ret;
75 }
76 
77 static int hns_roce_del_gid(const struct ib_gid_attr *attr, void **context)
78 {
79 	struct hns_roce_dev *hr_dev = to_hr_dev(attr->device);
80 	u32 port = attr->port_num - 1;
81 	int ret;
82 
83 	if (port >= hr_dev->caps.num_ports)
84 		return -EINVAL;
85 
86 	ret = hr_dev->hw->set_gid(hr_dev, attr->index, NULL, NULL);
87 
88 	return ret;
89 }
90 
91 static int handle_en_event(struct hns_roce_dev *hr_dev, u32 port,
92 			   unsigned long event)
93 {
94 	struct device *dev = hr_dev->dev;
95 	struct net_device *netdev;
96 	int ret = 0;
97 
98 	netdev = hr_dev->iboe.netdevs[port];
99 	if (!netdev) {
100 		dev_err(dev, "can't find netdev on port(%u)!\n", port);
101 		return -ENODEV;
102 	}
103 
104 	switch (event) {
105 	case NETDEV_UP:
106 	case NETDEV_CHANGE:
107 	case NETDEV_REGISTER:
108 	case NETDEV_CHANGEADDR:
109 		ret = hns_roce_set_mac(hr_dev, port, netdev->dev_addr);
110 		break;
111 	case NETDEV_DOWN:
112 		/*
113 		 * In v1 engine, only support all ports closed together.
114 		 */
115 		break;
116 	default:
117 		dev_dbg(dev, "NETDEV event = 0x%x!\n", (u32)(event));
118 		break;
119 	}
120 
121 	return ret;
122 }
123 
124 static int hns_roce_netdev_event(struct notifier_block *self,
125 				 unsigned long event, void *ptr)
126 {
127 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
128 	struct hns_roce_ib_iboe *iboe = NULL;
129 	struct hns_roce_dev *hr_dev = NULL;
130 	int ret;
131 	u32 port;
132 
133 	hr_dev = container_of(self, struct hns_roce_dev, iboe.nb);
134 	iboe = &hr_dev->iboe;
135 
136 	for (port = 0; port < hr_dev->caps.num_ports; port++) {
137 		if (dev == iboe->netdevs[port]) {
138 			ret = handle_en_event(hr_dev, port, event);
139 			if (ret)
140 				return NOTIFY_DONE;
141 			break;
142 		}
143 	}
144 
145 	return NOTIFY_DONE;
146 }
147 
148 static int hns_roce_setup_mtu_mac(struct hns_roce_dev *hr_dev)
149 {
150 	int ret;
151 	u8 i;
152 
153 	for (i = 0; i < hr_dev->caps.num_ports; i++) {
154 		ret = hns_roce_set_mac(hr_dev, i,
155 				       hr_dev->iboe.netdevs[i]->dev_addr);
156 		if (ret)
157 			return ret;
158 	}
159 
160 	return 0;
161 }
162 
163 static int hns_roce_query_device(struct ib_device *ib_dev,
164 				 struct ib_device_attr *props,
165 				 struct ib_udata *uhw)
166 {
167 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev);
168 
169 	memset(props, 0, sizeof(*props));
170 
171 	props->fw_ver = hr_dev->caps.fw_ver;
172 	props->sys_image_guid = cpu_to_be64(hr_dev->sys_image_guid);
173 	props->max_mr_size = (u64)(~(0ULL));
174 	props->page_size_cap = hr_dev->caps.page_size_cap;
175 	props->vendor_id = hr_dev->vendor_id;
176 	props->vendor_part_id = hr_dev->vendor_part_id;
177 	props->hw_ver = hr_dev->hw_rev;
178 	props->max_qp = hr_dev->caps.num_qps;
179 	props->max_qp_wr = hr_dev->caps.max_wqes;
180 	props->device_cap_flags = IB_DEVICE_PORT_ACTIVE_EVENT |
181 				  IB_DEVICE_RC_RNR_NAK_GEN;
182 	props->max_send_sge = hr_dev->caps.max_sq_sg;
183 	props->max_recv_sge = hr_dev->caps.max_rq_sg;
184 	props->max_sge_rd = 1;
185 	props->max_cq = hr_dev->caps.num_cqs;
186 	props->max_cqe = hr_dev->caps.max_cqes;
187 	props->max_mr = hr_dev->caps.num_mtpts;
188 	props->max_pd = hr_dev->caps.num_pds;
189 	props->max_qp_rd_atom = hr_dev->caps.max_qp_dest_rdma;
190 	props->max_qp_init_rd_atom = hr_dev->caps.max_qp_init_rdma;
191 	props->atomic_cap = hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_ATOMIC ?
192 			    IB_ATOMIC_HCA : IB_ATOMIC_NONE;
193 	props->max_pkeys = 1;
194 	props->local_ca_ack_delay = hr_dev->caps.local_ca_ack_delay;
195 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ) {
196 		props->max_srq = hr_dev->caps.num_srqs;
197 		props->max_srq_wr = hr_dev->caps.max_srq_wrs;
198 		props->max_srq_sge = hr_dev->caps.max_srq_sges;
199 	}
200 
201 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_FRMR &&
202 	    hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
203 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
204 		props->max_fast_reg_page_list_len = HNS_ROCE_FRMR_MAX_PA;
205 	}
206 
207 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC)
208 		props->device_cap_flags |= IB_DEVICE_XRC;
209 
210 	return 0;
211 }
212 
213 static int hns_roce_query_port(struct ib_device *ib_dev, u32 port_num,
214 			       struct ib_port_attr *props)
215 {
216 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev);
217 	struct device *dev = hr_dev->dev;
218 	struct net_device *net_dev;
219 	unsigned long flags;
220 	enum ib_mtu mtu;
221 	u32 port;
222 	int ret;
223 
224 	port = port_num - 1;
225 
226 	/* props being zeroed by the caller, avoid zeroing it here */
227 
228 	props->max_mtu = hr_dev->caps.max_mtu;
229 	props->gid_tbl_len = hr_dev->caps.gid_table_len[port];
230 	props->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
231 				IB_PORT_VENDOR_CLASS_SUP |
232 				IB_PORT_BOOT_MGMT_SUP;
233 	props->max_msg_sz = HNS_ROCE_MAX_MSG_LEN;
234 	props->pkey_tbl_len = 1;
235 	ret = ib_get_eth_speed(ib_dev, port_num, &props->active_speed,
236 			       &props->active_width);
237 	if (ret)
238 		ibdev_warn(ib_dev, "failed to get speed, ret = %d.\n", ret);
239 
240 	spin_lock_irqsave(&hr_dev->iboe.lock, flags);
241 
242 	net_dev = hr_dev->iboe.netdevs[port];
243 	if (!net_dev) {
244 		spin_unlock_irqrestore(&hr_dev->iboe.lock, flags);
245 		dev_err(dev, "find netdev %u failed!\n", port);
246 		return -EINVAL;
247 	}
248 
249 	mtu = iboe_get_mtu(net_dev->mtu);
250 	props->active_mtu = mtu ? min(props->max_mtu, mtu) : IB_MTU_256;
251 	props->state = netif_running(net_dev) && netif_carrier_ok(net_dev) ?
252 			       IB_PORT_ACTIVE :
253 			       IB_PORT_DOWN;
254 	props->phys_state = props->state == IB_PORT_ACTIVE ?
255 				    IB_PORT_PHYS_STATE_LINK_UP :
256 				    IB_PORT_PHYS_STATE_DISABLED;
257 
258 	spin_unlock_irqrestore(&hr_dev->iboe.lock, flags);
259 
260 	return 0;
261 }
262 
263 static enum rdma_link_layer hns_roce_get_link_layer(struct ib_device *device,
264 						    u32 port_num)
265 {
266 	return IB_LINK_LAYER_ETHERNET;
267 }
268 
269 static int hns_roce_query_pkey(struct ib_device *ib_dev, u32 port, u16 index,
270 			       u16 *pkey)
271 {
272 	if (index > 0)
273 		return -EINVAL;
274 
275 	*pkey = PKEY_ID;
276 
277 	return 0;
278 }
279 
280 static int hns_roce_modify_device(struct ib_device *ib_dev, int mask,
281 				  struct ib_device_modify *props)
282 {
283 	unsigned long flags;
284 
285 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
286 		return -EOPNOTSUPP;
287 
288 	if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
289 		spin_lock_irqsave(&to_hr_dev(ib_dev)->sm_lock, flags);
290 		memcpy(ib_dev->node_desc, props->node_desc, NODE_DESC_SIZE);
291 		spin_unlock_irqrestore(&to_hr_dev(ib_dev)->sm_lock, flags);
292 	}
293 
294 	return 0;
295 }
296 
297 struct hns_user_mmap_entry *
298 hns_roce_user_mmap_entry_insert(struct ib_ucontext *ucontext, u64 address,
299 				size_t length,
300 				enum hns_roce_mmap_type mmap_type)
301 {
302 	struct hns_user_mmap_entry *entry;
303 	int ret;
304 
305 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
306 	if (!entry)
307 		return NULL;
308 
309 	entry->address = address;
310 	entry->mmap_type = mmap_type;
311 
312 	switch (mmap_type) {
313 	/* pgoff 0 must be used by DB for compatibility */
314 	case HNS_ROCE_MMAP_TYPE_DB:
315 		ret = rdma_user_mmap_entry_insert_exact(
316 				ucontext, &entry->rdma_entry, length, 0);
317 		break;
318 	case HNS_ROCE_MMAP_TYPE_DWQE:
319 		ret = rdma_user_mmap_entry_insert_range(
320 				ucontext, &entry->rdma_entry, length, 1,
321 				U32_MAX);
322 		break;
323 	default:
324 		ret = -EINVAL;
325 		break;
326 	}
327 
328 	if (ret) {
329 		kfree(entry);
330 		return NULL;
331 	}
332 
333 	return entry;
334 }
335 
336 static void hns_roce_dealloc_uar_entry(struct hns_roce_ucontext *context)
337 {
338 	if (context->db_mmap_entry)
339 		rdma_user_mmap_entry_remove(
340 			&context->db_mmap_entry->rdma_entry);
341 }
342 
343 static int hns_roce_alloc_uar_entry(struct ib_ucontext *uctx)
344 {
345 	struct hns_roce_ucontext *context = to_hr_ucontext(uctx);
346 	u64 address;
347 
348 	address = context->uar.pfn << PAGE_SHIFT;
349 	context->db_mmap_entry = hns_roce_user_mmap_entry_insert(
350 		uctx, address, PAGE_SIZE, HNS_ROCE_MMAP_TYPE_DB);
351 	if (!context->db_mmap_entry)
352 		return -ENOMEM;
353 
354 	return 0;
355 }
356 
357 static int hns_roce_alloc_ucontext(struct ib_ucontext *uctx,
358 				   struct ib_udata *udata)
359 {
360 	struct hns_roce_ucontext *context = to_hr_ucontext(uctx);
361 	struct hns_roce_dev *hr_dev = to_hr_dev(uctx->device);
362 	struct hns_roce_ib_alloc_ucontext_resp resp = {};
363 	struct hns_roce_ib_alloc_ucontext ucmd = {};
364 	int ret;
365 
366 	if (!hr_dev->active)
367 		return -EAGAIN;
368 
369 	resp.qp_tab_size = hr_dev->caps.num_qps;
370 	resp.srq_tab_size = hr_dev->caps.num_srqs;
371 
372 	ret = ib_copy_from_udata(&ucmd, udata,
373 				 min(udata->inlen, sizeof(ucmd)));
374 	if (ret)
375 		return ret;
376 
377 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
378 		context->config = ucmd.config & HNS_ROCE_EXSGE_FLAGS;
379 
380 	if (context->config & HNS_ROCE_EXSGE_FLAGS) {
381 		resp.config |= HNS_ROCE_RSP_EXSGE_FLAGS;
382 		resp.max_inline_data = hr_dev->caps.max_sq_inline;
383 	}
384 
385 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) {
386 		context->config |= ucmd.config & HNS_ROCE_RQ_INLINE_FLAGS;
387 		if (context->config & HNS_ROCE_RQ_INLINE_FLAGS)
388 			resp.config |= HNS_ROCE_RSP_RQ_INLINE_FLAGS;
389 	}
390 
391 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQE_INLINE) {
392 		context->config |= ucmd.config & HNS_ROCE_CQE_INLINE_FLAGS;
393 		if (context->config & HNS_ROCE_CQE_INLINE_FLAGS)
394 			resp.config |= HNS_ROCE_RSP_CQE_INLINE_FLAGS;
395 	}
396 
397 	ret = hns_roce_uar_alloc(hr_dev, &context->uar);
398 	if (ret)
399 		goto error_fail_uar_alloc;
400 
401 	ret = hns_roce_alloc_uar_entry(uctx);
402 	if (ret)
403 		goto error_fail_uar_entry;
404 
405 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB ||
406 	    hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) {
407 		INIT_LIST_HEAD(&context->page_list);
408 		mutex_init(&context->page_mutex);
409 	}
410 
411 	resp.cqe_size = hr_dev->caps.cqe_sz;
412 
413 	ret = ib_copy_to_udata(udata, &resp,
414 			       min(udata->outlen, sizeof(resp)));
415 	if (ret)
416 		goto error_fail_copy_to_udata;
417 
418 	return 0;
419 
420 error_fail_copy_to_udata:
421 	hns_roce_dealloc_uar_entry(context);
422 
423 error_fail_uar_entry:
424 	ida_free(&hr_dev->uar_ida.ida, (int)context->uar.logic_idx);
425 
426 error_fail_uar_alloc:
427 	return ret;
428 }
429 
430 static void hns_roce_dealloc_ucontext(struct ib_ucontext *ibcontext)
431 {
432 	struct hns_roce_ucontext *context = to_hr_ucontext(ibcontext);
433 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcontext->device);
434 
435 	hns_roce_dealloc_uar_entry(context);
436 
437 	ida_free(&hr_dev->uar_ida.ida, (int)context->uar.logic_idx);
438 }
439 
440 static int hns_roce_mmap(struct ib_ucontext *uctx, struct vm_area_struct *vma)
441 {
442 	struct rdma_user_mmap_entry *rdma_entry;
443 	struct hns_user_mmap_entry *entry;
444 	phys_addr_t pfn;
445 	pgprot_t prot;
446 	int ret;
447 
448 	rdma_entry = rdma_user_mmap_entry_get_pgoff(uctx, vma->vm_pgoff);
449 	if (!rdma_entry)
450 		return -EINVAL;
451 
452 	entry = to_hns_mmap(rdma_entry);
453 	pfn = entry->address >> PAGE_SHIFT;
454 
455 	switch (entry->mmap_type) {
456 	case HNS_ROCE_MMAP_TYPE_DB:
457 	case HNS_ROCE_MMAP_TYPE_DWQE:
458 		prot = pgprot_device(vma->vm_page_prot);
459 		break;
460 	default:
461 		ret = -EINVAL;
462 		goto out;
463 	}
464 
465 	ret = rdma_user_mmap_io(uctx, vma, pfn, rdma_entry->npages * PAGE_SIZE,
466 				prot, rdma_entry);
467 
468 out:
469 	rdma_user_mmap_entry_put(rdma_entry);
470 	return ret;
471 }
472 
473 static void hns_roce_free_mmap(struct rdma_user_mmap_entry *rdma_entry)
474 {
475 	struct hns_user_mmap_entry *entry = to_hns_mmap(rdma_entry);
476 
477 	kfree(entry);
478 }
479 
480 static int hns_roce_port_immutable(struct ib_device *ib_dev, u32 port_num,
481 				   struct ib_port_immutable *immutable)
482 {
483 	struct ib_port_attr attr;
484 	int ret;
485 
486 	ret = ib_query_port(ib_dev, port_num, &attr);
487 	if (ret)
488 		return ret;
489 
490 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
491 	immutable->gid_tbl_len = attr.gid_tbl_len;
492 
493 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
494 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
495 	if (to_hr_dev(ib_dev)->caps.flags & HNS_ROCE_CAP_FLAG_ROCE_V1_V2)
496 		immutable->core_cap_flags |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
497 
498 	return 0;
499 }
500 
501 static void hns_roce_disassociate_ucontext(struct ib_ucontext *ibcontext)
502 {
503 }
504 
505 static void hns_roce_get_fw_ver(struct ib_device *device, char *str)
506 {
507 	u64 fw_ver = to_hr_dev(device)->caps.fw_ver;
508 	unsigned int major, minor, sub_minor;
509 
510 	major = upper_32_bits(fw_ver);
511 	minor = high_16_bits(lower_32_bits(fw_ver));
512 	sub_minor = low_16_bits(fw_ver);
513 
514 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u.%04u", major, minor,
515 		 sub_minor);
516 }
517 
518 static void hns_roce_unregister_device(struct hns_roce_dev *hr_dev)
519 {
520 	struct hns_roce_ib_iboe *iboe = &hr_dev->iboe;
521 
522 	hr_dev->active = false;
523 	unregister_netdevice_notifier(&iboe->nb);
524 	ib_unregister_device(&hr_dev->ib_dev);
525 }
526 
527 static const struct ib_device_ops hns_roce_dev_ops = {
528 	.owner = THIS_MODULE,
529 	.driver_id = RDMA_DRIVER_HNS,
530 	.uverbs_abi_ver = 1,
531 	.uverbs_no_driver_id_binding = 1,
532 
533 	.get_dev_fw_str = hns_roce_get_fw_ver,
534 	.add_gid = hns_roce_add_gid,
535 	.alloc_pd = hns_roce_alloc_pd,
536 	.alloc_ucontext = hns_roce_alloc_ucontext,
537 	.create_ah = hns_roce_create_ah,
538 	.create_user_ah = hns_roce_create_ah,
539 	.create_cq = hns_roce_create_cq,
540 	.create_qp = hns_roce_create_qp,
541 	.dealloc_pd = hns_roce_dealloc_pd,
542 	.dealloc_ucontext = hns_roce_dealloc_ucontext,
543 	.del_gid = hns_roce_del_gid,
544 	.dereg_mr = hns_roce_dereg_mr,
545 	.destroy_ah = hns_roce_destroy_ah,
546 	.destroy_cq = hns_roce_destroy_cq,
547 	.disassociate_ucontext = hns_roce_disassociate_ucontext,
548 	.get_dma_mr = hns_roce_get_dma_mr,
549 	.get_link_layer = hns_roce_get_link_layer,
550 	.get_port_immutable = hns_roce_port_immutable,
551 	.mmap = hns_roce_mmap,
552 	.mmap_free = hns_roce_free_mmap,
553 	.modify_device = hns_roce_modify_device,
554 	.modify_qp = hns_roce_modify_qp,
555 	.query_ah = hns_roce_query_ah,
556 	.query_device = hns_roce_query_device,
557 	.query_pkey = hns_roce_query_pkey,
558 	.query_port = hns_roce_query_port,
559 	.reg_user_mr = hns_roce_reg_user_mr,
560 
561 	INIT_RDMA_OBJ_SIZE(ib_ah, hns_roce_ah, ibah),
562 	INIT_RDMA_OBJ_SIZE(ib_cq, hns_roce_cq, ib_cq),
563 	INIT_RDMA_OBJ_SIZE(ib_pd, hns_roce_pd, ibpd),
564 	INIT_RDMA_OBJ_SIZE(ib_qp, hns_roce_qp, ibqp),
565 	INIT_RDMA_OBJ_SIZE(ib_ucontext, hns_roce_ucontext, ibucontext),
566 };
567 
568 static const struct ib_device_ops hns_roce_dev_mr_ops = {
569 	.rereg_user_mr = hns_roce_rereg_user_mr,
570 };
571 
572 static const struct ib_device_ops hns_roce_dev_mw_ops = {
573 	.alloc_mw = hns_roce_alloc_mw,
574 	.dealloc_mw = hns_roce_dealloc_mw,
575 
576 	INIT_RDMA_OBJ_SIZE(ib_mw, hns_roce_mw, ibmw),
577 };
578 
579 static const struct ib_device_ops hns_roce_dev_frmr_ops = {
580 	.alloc_mr = hns_roce_alloc_mr,
581 	.map_mr_sg = hns_roce_map_mr_sg,
582 };
583 
584 static const struct ib_device_ops hns_roce_dev_srq_ops = {
585 	.create_srq = hns_roce_create_srq,
586 	.destroy_srq = hns_roce_destroy_srq,
587 
588 	INIT_RDMA_OBJ_SIZE(ib_srq, hns_roce_srq, ibsrq),
589 };
590 
591 static const struct ib_device_ops hns_roce_dev_xrcd_ops = {
592 	.alloc_xrcd = hns_roce_alloc_xrcd,
593 	.dealloc_xrcd = hns_roce_dealloc_xrcd,
594 
595 	INIT_RDMA_OBJ_SIZE(ib_xrcd, hns_roce_xrcd, ibxrcd),
596 };
597 
598 static const struct ib_device_ops hns_roce_dev_restrack_ops = {
599 	.fill_res_cq_entry = hns_roce_fill_res_cq_entry,
600 	.fill_res_cq_entry_raw = hns_roce_fill_res_cq_entry_raw,
601 	.fill_res_qp_entry = hns_roce_fill_res_qp_entry,
602 	.fill_res_qp_entry_raw = hns_roce_fill_res_qp_entry_raw,
603 	.fill_res_mr_entry = hns_roce_fill_res_mr_entry,
604 	.fill_res_mr_entry_raw = hns_roce_fill_res_mr_entry_raw,
605 };
606 
607 static int hns_roce_register_device(struct hns_roce_dev *hr_dev)
608 {
609 	int ret;
610 	struct hns_roce_ib_iboe *iboe = NULL;
611 	struct ib_device *ib_dev = NULL;
612 	struct device *dev = hr_dev->dev;
613 	unsigned int i;
614 
615 	iboe = &hr_dev->iboe;
616 	spin_lock_init(&iboe->lock);
617 
618 	ib_dev = &hr_dev->ib_dev;
619 
620 	ib_dev->node_type = RDMA_NODE_IB_CA;
621 	ib_dev->dev.parent = dev;
622 
623 	ib_dev->phys_port_cnt = hr_dev->caps.num_ports;
624 	ib_dev->local_dma_lkey = hr_dev->caps.reserved_lkey;
625 	ib_dev->num_comp_vectors = hr_dev->caps.num_comp_vectors;
626 
627 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_REREG_MR)
628 		ib_set_device_ops(ib_dev, &hns_roce_dev_mr_ops);
629 
630 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_MW)
631 		ib_set_device_ops(ib_dev, &hns_roce_dev_mw_ops);
632 
633 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_FRMR)
634 		ib_set_device_ops(ib_dev, &hns_roce_dev_frmr_ops);
635 
636 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ) {
637 		ib_set_device_ops(ib_dev, &hns_roce_dev_srq_ops);
638 		ib_set_device_ops(ib_dev, hr_dev->hw->hns_roce_dev_srq_ops);
639 	}
640 
641 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC)
642 		ib_set_device_ops(ib_dev, &hns_roce_dev_xrcd_ops);
643 
644 	ib_set_device_ops(ib_dev, hr_dev->hw->hns_roce_dev_ops);
645 	ib_set_device_ops(ib_dev, &hns_roce_dev_ops);
646 	ib_set_device_ops(ib_dev, &hns_roce_dev_restrack_ops);
647 	for (i = 0; i < hr_dev->caps.num_ports; i++) {
648 		if (!hr_dev->iboe.netdevs[i])
649 			continue;
650 
651 		ret = ib_device_set_netdev(ib_dev, hr_dev->iboe.netdevs[i],
652 					   i + 1);
653 		if (ret)
654 			return ret;
655 	}
656 	dma_set_max_seg_size(dev, UINT_MAX);
657 	ret = ib_register_device(ib_dev, "hns_%d", dev);
658 	if (ret) {
659 		dev_err(dev, "ib_register_device failed!\n");
660 		return ret;
661 	}
662 
663 	ret = hns_roce_setup_mtu_mac(hr_dev);
664 	if (ret) {
665 		dev_err(dev, "setup_mtu_mac failed!\n");
666 		goto error_failed_setup_mtu_mac;
667 	}
668 
669 	iboe->nb.notifier_call = hns_roce_netdev_event;
670 	ret = register_netdevice_notifier(&iboe->nb);
671 	if (ret) {
672 		dev_err(dev, "register_netdevice_notifier failed!\n");
673 		goto error_failed_setup_mtu_mac;
674 	}
675 
676 	hr_dev->active = true;
677 	return 0;
678 
679 error_failed_setup_mtu_mac:
680 	ib_unregister_device(ib_dev);
681 
682 	return ret;
683 }
684 
685 static int hns_roce_init_hem(struct hns_roce_dev *hr_dev)
686 {
687 	struct device *dev = hr_dev->dev;
688 	int ret;
689 
690 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table,
691 				      HEM_TYPE_MTPT, hr_dev->caps.mtpt_entry_sz,
692 				      hr_dev->caps.num_mtpts);
693 	if (ret) {
694 		dev_err(dev, "failed to init MTPT context memory, aborting.\n");
695 		return ret;
696 	}
697 
698 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->qp_table.qp_table,
699 				      HEM_TYPE_QPC, hr_dev->caps.qpc_sz,
700 				      hr_dev->caps.num_qps);
701 	if (ret) {
702 		dev_err(dev, "failed to init QP context memory, aborting.\n");
703 		goto err_unmap_dmpt;
704 	}
705 
706 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->qp_table.irrl_table,
707 				      HEM_TYPE_IRRL,
708 				      hr_dev->caps.irrl_entry_sz *
709 				      hr_dev->caps.max_qp_init_rdma,
710 				      hr_dev->caps.num_qps);
711 	if (ret) {
712 		dev_err(dev, "failed to init irrl_table memory, aborting.\n");
713 		goto err_unmap_qp;
714 	}
715 
716 	if (hr_dev->caps.trrl_entry_sz) {
717 		ret = hns_roce_init_hem_table(hr_dev,
718 					      &hr_dev->qp_table.trrl_table,
719 					      HEM_TYPE_TRRL,
720 					      hr_dev->caps.trrl_entry_sz *
721 					      hr_dev->caps.max_qp_dest_rdma,
722 					      hr_dev->caps.num_qps);
723 		if (ret) {
724 			dev_err(dev,
725 				"failed to init trrl_table memory, aborting.\n");
726 			goto err_unmap_irrl;
727 		}
728 	}
729 
730 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->cq_table.table,
731 				      HEM_TYPE_CQC, hr_dev->caps.cqc_entry_sz,
732 				      hr_dev->caps.num_cqs);
733 	if (ret) {
734 		dev_err(dev, "failed to init CQ context memory, aborting.\n");
735 		goto err_unmap_trrl;
736 	}
737 
738 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ) {
739 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->srq_table.table,
740 					      HEM_TYPE_SRQC,
741 					      hr_dev->caps.srqc_entry_sz,
742 					      hr_dev->caps.num_srqs);
743 		if (ret) {
744 			dev_err(dev,
745 				"failed to init SRQ context memory, aborting.\n");
746 			goto err_unmap_cq;
747 		}
748 	}
749 
750 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
751 		ret = hns_roce_init_hem_table(hr_dev,
752 					      &hr_dev->qp_table.sccc_table,
753 					      HEM_TYPE_SCCC,
754 					      hr_dev->caps.sccc_sz,
755 					      hr_dev->caps.num_qps);
756 		if (ret) {
757 			dev_err(dev,
758 				"failed to init SCC context memory, aborting.\n");
759 			goto err_unmap_srq;
760 		}
761 	}
762 
763 	if (hr_dev->caps.qpc_timer_entry_sz) {
764 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->qpc_timer_table,
765 					      HEM_TYPE_QPC_TIMER,
766 					      hr_dev->caps.qpc_timer_entry_sz,
767 					      hr_dev->caps.qpc_timer_bt_num);
768 		if (ret) {
769 			dev_err(dev,
770 				"failed to init QPC timer memory, aborting.\n");
771 			goto err_unmap_ctx;
772 		}
773 	}
774 
775 	if (hr_dev->caps.cqc_timer_entry_sz) {
776 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->cqc_timer_table,
777 					      HEM_TYPE_CQC_TIMER,
778 					      hr_dev->caps.cqc_timer_entry_sz,
779 					      hr_dev->caps.cqc_timer_bt_num);
780 		if (ret) {
781 			dev_err(dev,
782 				"failed to init CQC timer memory, aborting.\n");
783 			goto err_unmap_qpc_timer;
784 		}
785 	}
786 
787 	if (hr_dev->caps.gmv_entry_sz) {
788 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->gmv_table,
789 					      HEM_TYPE_GMV,
790 					      hr_dev->caps.gmv_entry_sz,
791 					      hr_dev->caps.gmv_entry_num);
792 		if (ret) {
793 			dev_err(dev,
794 				"failed to init gmv table memory, ret = %d\n",
795 				ret);
796 			goto err_unmap_cqc_timer;
797 		}
798 	}
799 
800 	return 0;
801 
802 err_unmap_cqc_timer:
803 	if (hr_dev->caps.cqc_timer_entry_sz)
804 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cqc_timer_table);
805 
806 err_unmap_qpc_timer:
807 	if (hr_dev->caps.qpc_timer_entry_sz)
808 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qpc_timer_table);
809 
810 err_unmap_ctx:
811 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
812 		hns_roce_cleanup_hem_table(hr_dev,
813 					   &hr_dev->qp_table.sccc_table);
814 err_unmap_srq:
815 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
816 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->srq_table.table);
817 
818 err_unmap_cq:
819 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cq_table.table);
820 
821 err_unmap_trrl:
822 	if (hr_dev->caps.trrl_entry_sz)
823 		hns_roce_cleanup_hem_table(hr_dev,
824 					   &hr_dev->qp_table.trrl_table);
825 
826 err_unmap_irrl:
827 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.irrl_table);
828 
829 err_unmap_qp:
830 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.qp_table);
831 
832 err_unmap_dmpt:
833 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table);
834 
835 	return ret;
836 }
837 
838 /**
839  * hns_roce_setup_hca - setup host channel adapter
840  * @hr_dev: pointer to hns roce device
841  * Return : int
842  */
843 static int hns_roce_setup_hca(struct hns_roce_dev *hr_dev)
844 {
845 	struct device *dev = hr_dev->dev;
846 	int ret;
847 
848 	spin_lock_init(&hr_dev->sm_lock);
849 
850 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB ||
851 	    hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) {
852 		INIT_LIST_HEAD(&hr_dev->pgdir_list);
853 		mutex_init(&hr_dev->pgdir_mutex);
854 	}
855 
856 	hns_roce_init_uar_table(hr_dev);
857 
858 	ret = hns_roce_uar_alloc(hr_dev, &hr_dev->priv_uar);
859 	if (ret) {
860 		dev_err(dev, "failed to allocate priv_uar.\n");
861 		goto err_uar_table_free;
862 	}
863 
864 	ret = hns_roce_init_qp_table(hr_dev);
865 	if (ret) {
866 		dev_err(dev, "failed to init qp_table.\n");
867 		goto err_uar_table_free;
868 	}
869 
870 	hns_roce_init_pd_table(hr_dev);
871 
872 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC)
873 		hns_roce_init_xrcd_table(hr_dev);
874 
875 	hns_roce_init_mr_table(hr_dev);
876 
877 	hns_roce_init_cq_table(hr_dev);
878 
879 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
880 		hns_roce_init_srq_table(hr_dev);
881 
882 	return 0;
883 
884 err_uar_table_free:
885 	ida_destroy(&hr_dev->uar_ida.ida);
886 	return ret;
887 }
888 
889 static void check_and_get_armed_cq(struct list_head *cq_list, struct ib_cq *cq)
890 {
891 	struct hns_roce_cq *hr_cq = to_hr_cq(cq);
892 	unsigned long flags;
893 
894 	spin_lock_irqsave(&hr_cq->lock, flags);
895 	if (cq->comp_handler) {
896 		if (!hr_cq->is_armed) {
897 			hr_cq->is_armed = 1;
898 			list_add_tail(&hr_cq->node, cq_list);
899 		}
900 	}
901 	spin_unlock_irqrestore(&hr_cq->lock, flags);
902 }
903 
904 void hns_roce_handle_device_err(struct hns_roce_dev *hr_dev)
905 {
906 	struct hns_roce_qp *hr_qp;
907 	struct hns_roce_cq *hr_cq;
908 	struct list_head cq_list;
909 	unsigned long flags_qp;
910 	unsigned long flags;
911 
912 	INIT_LIST_HEAD(&cq_list);
913 
914 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
915 	list_for_each_entry(hr_qp, &hr_dev->qp_list, node) {
916 		spin_lock_irqsave(&hr_qp->sq.lock, flags_qp);
917 		if (hr_qp->sq.tail != hr_qp->sq.head)
918 			check_and_get_armed_cq(&cq_list, hr_qp->ibqp.send_cq);
919 		spin_unlock_irqrestore(&hr_qp->sq.lock, flags_qp);
920 
921 		spin_lock_irqsave(&hr_qp->rq.lock, flags_qp);
922 		if ((!hr_qp->ibqp.srq) && (hr_qp->rq.tail != hr_qp->rq.head))
923 			check_and_get_armed_cq(&cq_list, hr_qp->ibqp.recv_cq);
924 		spin_unlock_irqrestore(&hr_qp->rq.lock, flags_qp);
925 	}
926 
927 	list_for_each_entry(hr_cq, &cq_list, node)
928 		hns_roce_cq_completion(hr_dev, hr_cq->cqn);
929 
930 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
931 }
932 
933 int hns_roce_init(struct hns_roce_dev *hr_dev)
934 {
935 	struct device *dev = hr_dev->dev;
936 	int ret;
937 
938 	hr_dev->is_reset = false;
939 
940 	if (hr_dev->hw->cmq_init) {
941 		ret = hr_dev->hw->cmq_init(hr_dev);
942 		if (ret) {
943 			dev_err(dev, "init RoCE Command Queue failed!\n");
944 			return ret;
945 		}
946 	}
947 
948 	ret = hr_dev->hw->hw_profile(hr_dev);
949 	if (ret) {
950 		dev_err(dev, "get RoCE engine profile failed!\n");
951 		goto error_failed_cmd_init;
952 	}
953 
954 	ret = hns_roce_cmd_init(hr_dev);
955 	if (ret) {
956 		dev_err(dev, "cmd init failed!\n");
957 		goto error_failed_cmd_init;
958 	}
959 
960 	/* EQ depends on poll mode, event mode depends on EQ */
961 	ret = hr_dev->hw->init_eq(hr_dev);
962 	if (ret) {
963 		dev_err(dev, "eq init failed!\n");
964 		goto error_failed_eq_table;
965 	}
966 
967 	if (hr_dev->cmd_mod) {
968 		ret = hns_roce_cmd_use_events(hr_dev);
969 		if (ret)
970 			dev_warn(dev,
971 				 "Cmd event  mode failed, set back to poll!\n");
972 	}
973 
974 	ret = hns_roce_init_hem(hr_dev);
975 	if (ret) {
976 		dev_err(dev, "init HEM(Hardware Entry Memory) failed!\n");
977 		goto error_failed_init_hem;
978 	}
979 
980 	ret = hns_roce_setup_hca(hr_dev);
981 	if (ret) {
982 		dev_err(dev, "setup hca failed!\n");
983 		goto error_failed_setup_hca;
984 	}
985 
986 	if (hr_dev->hw->hw_init) {
987 		ret = hr_dev->hw->hw_init(hr_dev);
988 		if (ret) {
989 			dev_err(dev, "hw_init failed!\n");
990 			goto error_failed_engine_init;
991 		}
992 	}
993 
994 	INIT_LIST_HEAD(&hr_dev->qp_list);
995 	spin_lock_init(&hr_dev->qp_list_lock);
996 	INIT_LIST_HEAD(&hr_dev->dip_list);
997 	spin_lock_init(&hr_dev->dip_list_lock);
998 
999 	ret = hns_roce_register_device(hr_dev);
1000 	if (ret)
1001 		goto error_failed_register_device;
1002 
1003 	return 0;
1004 
1005 error_failed_register_device:
1006 	if (hr_dev->hw->hw_exit)
1007 		hr_dev->hw->hw_exit(hr_dev);
1008 
1009 error_failed_engine_init:
1010 	hns_roce_cleanup_bitmap(hr_dev);
1011 
1012 error_failed_setup_hca:
1013 	hns_roce_cleanup_hem(hr_dev);
1014 
1015 error_failed_init_hem:
1016 	if (hr_dev->cmd_mod)
1017 		hns_roce_cmd_use_polling(hr_dev);
1018 	hr_dev->hw->cleanup_eq(hr_dev);
1019 
1020 error_failed_eq_table:
1021 	hns_roce_cmd_cleanup(hr_dev);
1022 
1023 error_failed_cmd_init:
1024 	if (hr_dev->hw->cmq_exit)
1025 		hr_dev->hw->cmq_exit(hr_dev);
1026 
1027 	return ret;
1028 }
1029 
1030 void hns_roce_exit(struct hns_roce_dev *hr_dev)
1031 {
1032 	hns_roce_unregister_device(hr_dev);
1033 
1034 	if (hr_dev->hw->hw_exit)
1035 		hr_dev->hw->hw_exit(hr_dev);
1036 	hns_roce_cleanup_bitmap(hr_dev);
1037 	hns_roce_cleanup_hem(hr_dev);
1038 
1039 	if (hr_dev->cmd_mod)
1040 		hns_roce_cmd_use_polling(hr_dev);
1041 
1042 	hr_dev->hw->cleanup_eq(hr_dev);
1043 	hns_roce_cmd_cleanup(hr_dev);
1044 	if (hr_dev->hw->cmq_exit)
1045 		hr_dev->hw->cmq_exit(hr_dev);
1046 }
1047 
1048 MODULE_LICENSE("Dual BSD/GPL");
1049 MODULE_AUTHOR("Wei Hu <xavier.huwei@huawei.com>");
1050 MODULE_AUTHOR("Nenglong Zhao <zhaonenglong@hisilicon.com>");
1051 MODULE_AUTHOR("Lijun Ou <oulijun@huawei.com>");
1052 MODULE_DESCRIPTION("HNS RoCE Driver");
1053