xref: /openbmc/linux/drivers/infiniband/hw/hns/hns_roce_main.c (revision 2bb185c68bf4c147f43d932e8a34fa150d148940)
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 
223 	port = port_num - 1;
224 
225 	/* props being zeroed by the caller, avoid zeroing it here */
226 
227 	props->max_mtu = hr_dev->caps.max_mtu;
228 	props->gid_tbl_len = hr_dev->caps.gid_table_len[port];
229 	props->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
230 				IB_PORT_VENDOR_CLASS_SUP |
231 				IB_PORT_BOOT_MGMT_SUP;
232 	props->max_msg_sz = HNS_ROCE_MAX_MSG_LEN;
233 	props->pkey_tbl_len = 1;
234 	props->active_width = IB_WIDTH_4X;
235 	props->active_speed = 1;
236 
237 	spin_lock_irqsave(&hr_dev->iboe.lock, flags);
238 
239 	net_dev = hr_dev->iboe.netdevs[port];
240 	if (!net_dev) {
241 		spin_unlock_irqrestore(&hr_dev->iboe.lock, flags);
242 		dev_err(dev, "find netdev %u failed!\n", port);
243 		return -EINVAL;
244 	}
245 
246 	mtu = iboe_get_mtu(net_dev->mtu);
247 	props->active_mtu = mtu ? min(props->max_mtu, mtu) : IB_MTU_256;
248 	props->state = netif_running(net_dev) && netif_carrier_ok(net_dev) ?
249 			       IB_PORT_ACTIVE :
250 			       IB_PORT_DOWN;
251 	props->phys_state = props->state == IB_PORT_ACTIVE ?
252 				    IB_PORT_PHYS_STATE_LINK_UP :
253 				    IB_PORT_PHYS_STATE_DISABLED;
254 
255 	spin_unlock_irqrestore(&hr_dev->iboe.lock, flags);
256 
257 	return 0;
258 }
259 
260 static enum rdma_link_layer hns_roce_get_link_layer(struct ib_device *device,
261 						    u32 port_num)
262 {
263 	return IB_LINK_LAYER_ETHERNET;
264 }
265 
266 static int hns_roce_query_pkey(struct ib_device *ib_dev, u32 port, u16 index,
267 			       u16 *pkey)
268 {
269 	if (index > 0)
270 		return -EINVAL;
271 
272 	*pkey = PKEY_ID;
273 
274 	return 0;
275 }
276 
277 static int hns_roce_modify_device(struct ib_device *ib_dev, int mask,
278 				  struct ib_device_modify *props)
279 {
280 	unsigned long flags;
281 
282 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
283 		return -EOPNOTSUPP;
284 
285 	if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
286 		spin_lock_irqsave(&to_hr_dev(ib_dev)->sm_lock, flags);
287 		memcpy(ib_dev->node_desc, props->node_desc, NODE_DESC_SIZE);
288 		spin_unlock_irqrestore(&to_hr_dev(ib_dev)->sm_lock, flags);
289 	}
290 
291 	return 0;
292 }
293 
294 struct hns_user_mmap_entry *
295 hns_roce_user_mmap_entry_insert(struct ib_ucontext *ucontext, u64 address,
296 				size_t length,
297 				enum hns_roce_mmap_type mmap_type)
298 {
299 	struct hns_user_mmap_entry *entry;
300 	int ret;
301 
302 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
303 	if (!entry)
304 		return NULL;
305 
306 	entry->address = address;
307 	entry->mmap_type = mmap_type;
308 
309 	switch (mmap_type) {
310 	/* pgoff 0 must be used by DB for compatibility */
311 	case HNS_ROCE_MMAP_TYPE_DB:
312 		ret = rdma_user_mmap_entry_insert_exact(
313 				ucontext, &entry->rdma_entry, length, 0);
314 		break;
315 	case HNS_ROCE_MMAP_TYPE_DWQE:
316 		ret = rdma_user_mmap_entry_insert_range(
317 				ucontext, &entry->rdma_entry, length, 1,
318 				U32_MAX);
319 		break;
320 	default:
321 		ret = -EINVAL;
322 		break;
323 	}
324 
325 	if (ret) {
326 		kfree(entry);
327 		return NULL;
328 	}
329 
330 	return entry;
331 }
332 
333 static void hns_roce_dealloc_uar_entry(struct hns_roce_ucontext *context)
334 {
335 	if (context->db_mmap_entry)
336 		rdma_user_mmap_entry_remove(
337 			&context->db_mmap_entry->rdma_entry);
338 }
339 
340 static int hns_roce_alloc_uar_entry(struct ib_ucontext *uctx)
341 {
342 	struct hns_roce_ucontext *context = to_hr_ucontext(uctx);
343 	u64 address;
344 
345 	address = context->uar.pfn << PAGE_SHIFT;
346 	context->db_mmap_entry = hns_roce_user_mmap_entry_insert(
347 		uctx, address, PAGE_SIZE, HNS_ROCE_MMAP_TYPE_DB);
348 	if (!context->db_mmap_entry)
349 		return -ENOMEM;
350 
351 	return 0;
352 }
353 
354 static int hns_roce_alloc_ucontext(struct ib_ucontext *uctx,
355 				   struct ib_udata *udata)
356 {
357 	struct hns_roce_ucontext *context = to_hr_ucontext(uctx);
358 	struct hns_roce_dev *hr_dev = to_hr_dev(uctx->device);
359 	struct hns_roce_ib_alloc_ucontext_resp resp = {};
360 	struct hns_roce_ib_alloc_ucontext ucmd = {};
361 	int ret;
362 
363 	if (!hr_dev->active)
364 		return -EAGAIN;
365 
366 	resp.qp_tab_size = hr_dev->caps.num_qps;
367 	resp.srq_tab_size = hr_dev->caps.num_srqs;
368 
369 	ret = ib_copy_from_udata(&ucmd, udata,
370 				 min(udata->inlen, sizeof(ucmd)));
371 	if (ret)
372 		return ret;
373 
374 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
375 		context->config = ucmd.config & HNS_ROCE_EXSGE_FLAGS;
376 
377 	if (context->config & HNS_ROCE_EXSGE_FLAGS) {
378 		resp.config |= HNS_ROCE_RSP_EXSGE_FLAGS;
379 		resp.max_inline_data = hr_dev->caps.max_sq_inline;
380 	}
381 
382 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) {
383 		context->config |= ucmd.config & HNS_ROCE_RQ_INLINE_FLAGS;
384 		if (context->config & HNS_ROCE_RQ_INLINE_FLAGS)
385 			resp.config |= HNS_ROCE_RSP_RQ_INLINE_FLAGS;
386 	}
387 
388 	ret = hns_roce_uar_alloc(hr_dev, &context->uar);
389 	if (ret)
390 		goto error_fail_uar_alloc;
391 
392 	ret = hns_roce_alloc_uar_entry(uctx);
393 	if (ret)
394 		goto error_fail_uar_entry;
395 
396 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB ||
397 	    hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) {
398 		INIT_LIST_HEAD(&context->page_list);
399 		mutex_init(&context->page_mutex);
400 	}
401 
402 	resp.cqe_size = hr_dev->caps.cqe_sz;
403 
404 	ret = ib_copy_to_udata(udata, &resp,
405 			       min(udata->outlen, sizeof(resp)));
406 	if (ret)
407 		goto error_fail_copy_to_udata;
408 
409 	return 0;
410 
411 error_fail_copy_to_udata:
412 	hns_roce_dealloc_uar_entry(context);
413 
414 error_fail_uar_entry:
415 	ida_free(&hr_dev->uar_ida.ida, (int)context->uar.logic_idx);
416 
417 error_fail_uar_alloc:
418 	return ret;
419 }
420 
421 static void hns_roce_dealloc_ucontext(struct ib_ucontext *ibcontext)
422 {
423 	struct hns_roce_ucontext *context = to_hr_ucontext(ibcontext);
424 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcontext->device);
425 
426 	hns_roce_dealloc_uar_entry(context);
427 
428 	ida_free(&hr_dev->uar_ida.ida, (int)context->uar.logic_idx);
429 }
430 
431 static int hns_roce_mmap(struct ib_ucontext *uctx, struct vm_area_struct *vma)
432 {
433 	struct rdma_user_mmap_entry *rdma_entry;
434 	struct hns_user_mmap_entry *entry;
435 	phys_addr_t pfn;
436 	pgprot_t prot;
437 	int ret;
438 
439 	rdma_entry = rdma_user_mmap_entry_get_pgoff(uctx, vma->vm_pgoff);
440 	if (!rdma_entry)
441 		return -EINVAL;
442 
443 	entry = to_hns_mmap(rdma_entry);
444 	pfn = entry->address >> PAGE_SHIFT;
445 
446 	switch (entry->mmap_type) {
447 	case HNS_ROCE_MMAP_TYPE_DB:
448 	case HNS_ROCE_MMAP_TYPE_DWQE:
449 		prot = pgprot_device(vma->vm_page_prot);
450 		break;
451 	default:
452 		ret = -EINVAL;
453 		goto out;
454 	}
455 
456 	ret = rdma_user_mmap_io(uctx, vma, pfn, rdma_entry->npages * PAGE_SIZE,
457 				prot, rdma_entry);
458 
459 out:
460 	rdma_user_mmap_entry_put(rdma_entry);
461 	return ret;
462 }
463 
464 static void hns_roce_free_mmap(struct rdma_user_mmap_entry *rdma_entry)
465 {
466 	struct hns_user_mmap_entry *entry = to_hns_mmap(rdma_entry);
467 
468 	kfree(entry);
469 }
470 
471 static int hns_roce_port_immutable(struct ib_device *ib_dev, u32 port_num,
472 				   struct ib_port_immutable *immutable)
473 {
474 	struct ib_port_attr attr;
475 	int ret;
476 
477 	ret = ib_query_port(ib_dev, port_num, &attr);
478 	if (ret)
479 		return ret;
480 
481 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
482 	immutable->gid_tbl_len = attr.gid_tbl_len;
483 
484 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
485 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
486 	if (to_hr_dev(ib_dev)->caps.flags & HNS_ROCE_CAP_FLAG_ROCE_V1_V2)
487 		immutable->core_cap_flags |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
488 
489 	return 0;
490 }
491 
492 static void hns_roce_disassociate_ucontext(struct ib_ucontext *ibcontext)
493 {
494 }
495 
496 static void hns_roce_get_fw_ver(struct ib_device *device, char *str)
497 {
498 	u64 fw_ver = to_hr_dev(device)->caps.fw_ver;
499 	unsigned int major, minor, sub_minor;
500 
501 	major = upper_32_bits(fw_ver);
502 	minor = high_16_bits(lower_32_bits(fw_ver));
503 	sub_minor = low_16_bits(fw_ver);
504 
505 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u.%04u", major, minor,
506 		 sub_minor);
507 }
508 
509 static void hns_roce_unregister_device(struct hns_roce_dev *hr_dev)
510 {
511 	struct hns_roce_ib_iboe *iboe = &hr_dev->iboe;
512 
513 	hr_dev->active = false;
514 	unregister_netdevice_notifier(&iboe->nb);
515 	ib_unregister_device(&hr_dev->ib_dev);
516 }
517 
518 static const struct ib_device_ops hns_roce_dev_ops = {
519 	.owner = THIS_MODULE,
520 	.driver_id = RDMA_DRIVER_HNS,
521 	.uverbs_abi_ver = 1,
522 	.uverbs_no_driver_id_binding = 1,
523 
524 	.get_dev_fw_str = hns_roce_get_fw_ver,
525 	.add_gid = hns_roce_add_gid,
526 	.alloc_pd = hns_roce_alloc_pd,
527 	.alloc_ucontext = hns_roce_alloc_ucontext,
528 	.create_ah = hns_roce_create_ah,
529 	.create_user_ah = hns_roce_create_ah,
530 	.create_cq = hns_roce_create_cq,
531 	.create_qp = hns_roce_create_qp,
532 	.dealloc_pd = hns_roce_dealloc_pd,
533 	.dealloc_ucontext = hns_roce_dealloc_ucontext,
534 	.del_gid = hns_roce_del_gid,
535 	.dereg_mr = hns_roce_dereg_mr,
536 	.destroy_ah = hns_roce_destroy_ah,
537 	.destroy_cq = hns_roce_destroy_cq,
538 	.disassociate_ucontext = hns_roce_disassociate_ucontext,
539 	.get_dma_mr = hns_roce_get_dma_mr,
540 	.get_link_layer = hns_roce_get_link_layer,
541 	.get_port_immutable = hns_roce_port_immutable,
542 	.mmap = hns_roce_mmap,
543 	.mmap_free = hns_roce_free_mmap,
544 	.modify_device = hns_roce_modify_device,
545 	.modify_qp = hns_roce_modify_qp,
546 	.query_ah = hns_roce_query_ah,
547 	.query_device = hns_roce_query_device,
548 	.query_pkey = hns_roce_query_pkey,
549 	.query_port = hns_roce_query_port,
550 	.reg_user_mr = hns_roce_reg_user_mr,
551 
552 	INIT_RDMA_OBJ_SIZE(ib_ah, hns_roce_ah, ibah),
553 	INIT_RDMA_OBJ_SIZE(ib_cq, hns_roce_cq, ib_cq),
554 	INIT_RDMA_OBJ_SIZE(ib_pd, hns_roce_pd, ibpd),
555 	INIT_RDMA_OBJ_SIZE(ib_qp, hns_roce_qp, ibqp),
556 	INIT_RDMA_OBJ_SIZE(ib_ucontext, hns_roce_ucontext, ibucontext),
557 };
558 
559 static const struct ib_device_ops hns_roce_dev_mr_ops = {
560 	.rereg_user_mr = hns_roce_rereg_user_mr,
561 };
562 
563 static const struct ib_device_ops hns_roce_dev_mw_ops = {
564 	.alloc_mw = hns_roce_alloc_mw,
565 	.dealloc_mw = hns_roce_dealloc_mw,
566 
567 	INIT_RDMA_OBJ_SIZE(ib_mw, hns_roce_mw, ibmw),
568 };
569 
570 static const struct ib_device_ops hns_roce_dev_frmr_ops = {
571 	.alloc_mr = hns_roce_alloc_mr,
572 	.map_mr_sg = hns_roce_map_mr_sg,
573 };
574 
575 static const struct ib_device_ops hns_roce_dev_srq_ops = {
576 	.create_srq = hns_roce_create_srq,
577 	.destroy_srq = hns_roce_destroy_srq,
578 
579 	INIT_RDMA_OBJ_SIZE(ib_srq, hns_roce_srq, ibsrq),
580 };
581 
582 static const struct ib_device_ops hns_roce_dev_xrcd_ops = {
583 	.alloc_xrcd = hns_roce_alloc_xrcd,
584 	.dealloc_xrcd = hns_roce_dealloc_xrcd,
585 
586 	INIT_RDMA_OBJ_SIZE(ib_xrcd, hns_roce_xrcd, ibxrcd),
587 };
588 
589 static const struct ib_device_ops hns_roce_dev_restrack_ops = {
590 	.fill_res_cq_entry = hns_roce_fill_res_cq_entry,
591 	.fill_res_cq_entry_raw = hns_roce_fill_res_cq_entry_raw,
592 	.fill_res_qp_entry = hns_roce_fill_res_qp_entry,
593 	.fill_res_qp_entry_raw = hns_roce_fill_res_qp_entry_raw,
594 	.fill_res_mr_entry = hns_roce_fill_res_mr_entry,
595 	.fill_res_mr_entry_raw = hns_roce_fill_res_mr_entry_raw,
596 };
597 
598 static int hns_roce_register_device(struct hns_roce_dev *hr_dev)
599 {
600 	int ret;
601 	struct hns_roce_ib_iboe *iboe = NULL;
602 	struct ib_device *ib_dev = NULL;
603 	struct device *dev = hr_dev->dev;
604 	unsigned int i;
605 
606 	iboe = &hr_dev->iboe;
607 	spin_lock_init(&iboe->lock);
608 
609 	ib_dev = &hr_dev->ib_dev;
610 
611 	ib_dev->node_type = RDMA_NODE_IB_CA;
612 	ib_dev->dev.parent = dev;
613 
614 	ib_dev->phys_port_cnt = hr_dev->caps.num_ports;
615 	ib_dev->local_dma_lkey = hr_dev->caps.reserved_lkey;
616 	ib_dev->num_comp_vectors = hr_dev->caps.num_comp_vectors;
617 
618 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_REREG_MR)
619 		ib_set_device_ops(ib_dev, &hns_roce_dev_mr_ops);
620 
621 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_MW)
622 		ib_set_device_ops(ib_dev, &hns_roce_dev_mw_ops);
623 
624 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_FRMR)
625 		ib_set_device_ops(ib_dev, &hns_roce_dev_frmr_ops);
626 
627 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ) {
628 		ib_set_device_ops(ib_dev, &hns_roce_dev_srq_ops);
629 		ib_set_device_ops(ib_dev, hr_dev->hw->hns_roce_dev_srq_ops);
630 	}
631 
632 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC)
633 		ib_set_device_ops(ib_dev, &hns_roce_dev_xrcd_ops);
634 
635 	ib_set_device_ops(ib_dev, hr_dev->hw->hns_roce_dev_ops);
636 	ib_set_device_ops(ib_dev, &hns_roce_dev_ops);
637 	ib_set_device_ops(ib_dev, &hns_roce_dev_restrack_ops);
638 	for (i = 0; i < hr_dev->caps.num_ports; i++) {
639 		if (!hr_dev->iboe.netdevs[i])
640 			continue;
641 
642 		ret = ib_device_set_netdev(ib_dev, hr_dev->iboe.netdevs[i],
643 					   i + 1);
644 		if (ret)
645 			return ret;
646 	}
647 	dma_set_max_seg_size(dev, UINT_MAX);
648 	ret = ib_register_device(ib_dev, "hns_%d", dev);
649 	if (ret) {
650 		dev_err(dev, "ib_register_device failed!\n");
651 		return ret;
652 	}
653 
654 	ret = hns_roce_setup_mtu_mac(hr_dev);
655 	if (ret) {
656 		dev_err(dev, "setup_mtu_mac failed!\n");
657 		goto error_failed_setup_mtu_mac;
658 	}
659 
660 	iboe->nb.notifier_call = hns_roce_netdev_event;
661 	ret = register_netdevice_notifier(&iboe->nb);
662 	if (ret) {
663 		dev_err(dev, "register_netdevice_notifier failed!\n");
664 		goto error_failed_setup_mtu_mac;
665 	}
666 
667 	hr_dev->active = true;
668 	return 0;
669 
670 error_failed_setup_mtu_mac:
671 	ib_unregister_device(ib_dev);
672 
673 	return ret;
674 }
675 
676 static int hns_roce_init_hem(struct hns_roce_dev *hr_dev)
677 {
678 	struct device *dev = hr_dev->dev;
679 	int ret;
680 
681 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table,
682 				      HEM_TYPE_MTPT, hr_dev->caps.mtpt_entry_sz,
683 				      hr_dev->caps.num_mtpts);
684 	if (ret) {
685 		dev_err(dev, "failed to init MTPT context memory, aborting.\n");
686 		return ret;
687 	}
688 
689 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->qp_table.qp_table,
690 				      HEM_TYPE_QPC, hr_dev->caps.qpc_sz,
691 				      hr_dev->caps.num_qps);
692 	if (ret) {
693 		dev_err(dev, "failed to init QP context memory, aborting.\n");
694 		goto err_unmap_dmpt;
695 	}
696 
697 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->qp_table.irrl_table,
698 				      HEM_TYPE_IRRL,
699 				      hr_dev->caps.irrl_entry_sz *
700 				      hr_dev->caps.max_qp_init_rdma,
701 				      hr_dev->caps.num_qps);
702 	if (ret) {
703 		dev_err(dev, "failed to init irrl_table memory, aborting.\n");
704 		goto err_unmap_qp;
705 	}
706 
707 	if (hr_dev->caps.trrl_entry_sz) {
708 		ret = hns_roce_init_hem_table(hr_dev,
709 					      &hr_dev->qp_table.trrl_table,
710 					      HEM_TYPE_TRRL,
711 					      hr_dev->caps.trrl_entry_sz *
712 					      hr_dev->caps.max_qp_dest_rdma,
713 					      hr_dev->caps.num_qps);
714 		if (ret) {
715 			dev_err(dev,
716 				"failed to init trrl_table memory, aborting.\n");
717 			goto err_unmap_irrl;
718 		}
719 	}
720 
721 	ret = hns_roce_init_hem_table(hr_dev, &hr_dev->cq_table.table,
722 				      HEM_TYPE_CQC, hr_dev->caps.cqc_entry_sz,
723 				      hr_dev->caps.num_cqs);
724 	if (ret) {
725 		dev_err(dev, "failed to init CQ context memory, aborting.\n");
726 		goto err_unmap_trrl;
727 	}
728 
729 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ) {
730 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->srq_table.table,
731 					      HEM_TYPE_SRQC,
732 					      hr_dev->caps.srqc_entry_sz,
733 					      hr_dev->caps.num_srqs);
734 		if (ret) {
735 			dev_err(dev,
736 				"failed to init SRQ context memory, aborting.\n");
737 			goto err_unmap_cq;
738 		}
739 	}
740 
741 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
742 		ret = hns_roce_init_hem_table(hr_dev,
743 					      &hr_dev->qp_table.sccc_table,
744 					      HEM_TYPE_SCCC,
745 					      hr_dev->caps.sccc_sz,
746 					      hr_dev->caps.num_qps);
747 		if (ret) {
748 			dev_err(dev,
749 				"failed to init SCC context memory, aborting.\n");
750 			goto err_unmap_srq;
751 		}
752 	}
753 
754 	if (hr_dev->caps.qpc_timer_entry_sz) {
755 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->qpc_timer_table,
756 					      HEM_TYPE_QPC_TIMER,
757 					      hr_dev->caps.qpc_timer_entry_sz,
758 					      hr_dev->caps.qpc_timer_bt_num);
759 		if (ret) {
760 			dev_err(dev,
761 				"failed to init QPC timer memory, aborting.\n");
762 			goto err_unmap_ctx;
763 		}
764 	}
765 
766 	if (hr_dev->caps.cqc_timer_entry_sz) {
767 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->cqc_timer_table,
768 					      HEM_TYPE_CQC_TIMER,
769 					      hr_dev->caps.cqc_timer_entry_sz,
770 					      hr_dev->caps.cqc_timer_bt_num);
771 		if (ret) {
772 			dev_err(dev,
773 				"failed to init CQC timer memory, aborting.\n");
774 			goto err_unmap_qpc_timer;
775 		}
776 	}
777 
778 	if (hr_dev->caps.gmv_entry_sz) {
779 		ret = hns_roce_init_hem_table(hr_dev, &hr_dev->gmv_table,
780 					      HEM_TYPE_GMV,
781 					      hr_dev->caps.gmv_entry_sz,
782 					      hr_dev->caps.gmv_entry_num);
783 		if (ret) {
784 			dev_err(dev,
785 				"failed to init gmv table memory, ret = %d\n",
786 				ret);
787 			goto err_unmap_cqc_timer;
788 		}
789 	}
790 
791 	return 0;
792 
793 err_unmap_cqc_timer:
794 	if (hr_dev->caps.cqc_timer_entry_sz)
795 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cqc_timer_table);
796 
797 err_unmap_qpc_timer:
798 	if (hr_dev->caps.qpc_timer_entry_sz)
799 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qpc_timer_table);
800 
801 err_unmap_ctx:
802 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
803 		hns_roce_cleanup_hem_table(hr_dev,
804 					   &hr_dev->qp_table.sccc_table);
805 err_unmap_srq:
806 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
807 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->srq_table.table);
808 
809 err_unmap_cq:
810 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cq_table.table);
811 
812 err_unmap_trrl:
813 	if (hr_dev->caps.trrl_entry_sz)
814 		hns_roce_cleanup_hem_table(hr_dev,
815 					   &hr_dev->qp_table.trrl_table);
816 
817 err_unmap_irrl:
818 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.irrl_table);
819 
820 err_unmap_qp:
821 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.qp_table);
822 
823 err_unmap_dmpt:
824 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table);
825 
826 	return ret;
827 }
828 
829 /**
830  * hns_roce_setup_hca - setup host channel adapter
831  * @hr_dev: pointer to hns roce device
832  * Return : int
833  */
834 static int hns_roce_setup_hca(struct hns_roce_dev *hr_dev)
835 {
836 	struct device *dev = hr_dev->dev;
837 	int ret;
838 
839 	spin_lock_init(&hr_dev->sm_lock);
840 
841 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB ||
842 	    hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) {
843 		INIT_LIST_HEAD(&hr_dev->pgdir_list);
844 		mutex_init(&hr_dev->pgdir_mutex);
845 	}
846 
847 	hns_roce_init_uar_table(hr_dev);
848 
849 	ret = hns_roce_uar_alloc(hr_dev, &hr_dev->priv_uar);
850 	if (ret) {
851 		dev_err(dev, "failed to allocate priv_uar.\n");
852 		goto err_uar_table_free;
853 	}
854 
855 	ret = hns_roce_init_qp_table(hr_dev);
856 	if (ret) {
857 		dev_err(dev, "failed to init qp_table.\n");
858 		goto err_uar_table_free;
859 	}
860 
861 	hns_roce_init_pd_table(hr_dev);
862 
863 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC)
864 		hns_roce_init_xrcd_table(hr_dev);
865 
866 	hns_roce_init_mr_table(hr_dev);
867 
868 	hns_roce_init_cq_table(hr_dev);
869 
870 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
871 		hns_roce_init_srq_table(hr_dev);
872 
873 	return 0;
874 
875 err_uar_table_free:
876 	ida_destroy(&hr_dev->uar_ida.ida);
877 	return ret;
878 }
879 
880 static void check_and_get_armed_cq(struct list_head *cq_list, struct ib_cq *cq)
881 {
882 	struct hns_roce_cq *hr_cq = to_hr_cq(cq);
883 	unsigned long flags;
884 
885 	spin_lock_irqsave(&hr_cq->lock, flags);
886 	if (cq->comp_handler) {
887 		if (!hr_cq->is_armed) {
888 			hr_cq->is_armed = 1;
889 			list_add_tail(&hr_cq->node, cq_list);
890 		}
891 	}
892 	spin_unlock_irqrestore(&hr_cq->lock, flags);
893 }
894 
895 void hns_roce_handle_device_err(struct hns_roce_dev *hr_dev)
896 {
897 	struct hns_roce_qp *hr_qp;
898 	struct hns_roce_cq *hr_cq;
899 	struct list_head cq_list;
900 	unsigned long flags_qp;
901 	unsigned long flags;
902 
903 	INIT_LIST_HEAD(&cq_list);
904 
905 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
906 	list_for_each_entry(hr_qp, &hr_dev->qp_list, node) {
907 		spin_lock_irqsave(&hr_qp->sq.lock, flags_qp);
908 		if (hr_qp->sq.tail != hr_qp->sq.head)
909 			check_and_get_armed_cq(&cq_list, hr_qp->ibqp.send_cq);
910 		spin_unlock_irqrestore(&hr_qp->sq.lock, flags_qp);
911 
912 		spin_lock_irqsave(&hr_qp->rq.lock, flags_qp);
913 		if ((!hr_qp->ibqp.srq) && (hr_qp->rq.tail != hr_qp->rq.head))
914 			check_and_get_armed_cq(&cq_list, hr_qp->ibqp.recv_cq);
915 		spin_unlock_irqrestore(&hr_qp->rq.lock, flags_qp);
916 	}
917 
918 	list_for_each_entry(hr_cq, &cq_list, node)
919 		hns_roce_cq_completion(hr_dev, hr_cq->cqn);
920 
921 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
922 }
923 
924 int hns_roce_init(struct hns_roce_dev *hr_dev)
925 {
926 	struct device *dev = hr_dev->dev;
927 	int ret;
928 
929 	hr_dev->is_reset = false;
930 
931 	if (hr_dev->hw->cmq_init) {
932 		ret = hr_dev->hw->cmq_init(hr_dev);
933 		if (ret) {
934 			dev_err(dev, "init RoCE Command Queue failed!\n");
935 			return ret;
936 		}
937 	}
938 
939 	ret = hr_dev->hw->hw_profile(hr_dev);
940 	if (ret) {
941 		dev_err(dev, "get RoCE engine profile failed!\n");
942 		goto error_failed_cmd_init;
943 	}
944 
945 	ret = hns_roce_cmd_init(hr_dev);
946 	if (ret) {
947 		dev_err(dev, "cmd init failed!\n");
948 		goto error_failed_cmd_init;
949 	}
950 
951 	/* EQ depends on poll mode, event mode depends on EQ */
952 	ret = hr_dev->hw->init_eq(hr_dev);
953 	if (ret) {
954 		dev_err(dev, "eq init failed!\n");
955 		goto error_failed_eq_table;
956 	}
957 
958 	if (hr_dev->cmd_mod) {
959 		ret = hns_roce_cmd_use_events(hr_dev);
960 		if (ret)
961 			dev_warn(dev,
962 				 "Cmd event  mode failed, set back to poll!\n");
963 	}
964 
965 	ret = hns_roce_init_hem(hr_dev);
966 	if (ret) {
967 		dev_err(dev, "init HEM(Hardware Entry Memory) failed!\n");
968 		goto error_failed_init_hem;
969 	}
970 
971 	ret = hns_roce_setup_hca(hr_dev);
972 	if (ret) {
973 		dev_err(dev, "setup hca failed!\n");
974 		goto error_failed_setup_hca;
975 	}
976 
977 	if (hr_dev->hw->hw_init) {
978 		ret = hr_dev->hw->hw_init(hr_dev);
979 		if (ret) {
980 			dev_err(dev, "hw_init failed!\n");
981 			goto error_failed_engine_init;
982 		}
983 	}
984 
985 	INIT_LIST_HEAD(&hr_dev->qp_list);
986 	spin_lock_init(&hr_dev->qp_list_lock);
987 	INIT_LIST_HEAD(&hr_dev->dip_list);
988 	spin_lock_init(&hr_dev->dip_list_lock);
989 
990 	ret = hns_roce_register_device(hr_dev);
991 	if (ret)
992 		goto error_failed_register_device;
993 
994 	return 0;
995 
996 error_failed_register_device:
997 	if (hr_dev->hw->hw_exit)
998 		hr_dev->hw->hw_exit(hr_dev);
999 
1000 error_failed_engine_init:
1001 	hns_roce_cleanup_bitmap(hr_dev);
1002 
1003 error_failed_setup_hca:
1004 	hns_roce_cleanup_hem(hr_dev);
1005 
1006 error_failed_init_hem:
1007 	if (hr_dev->cmd_mod)
1008 		hns_roce_cmd_use_polling(hr_dev);
1009 	hr_dev->hw->cleanup_eq(hr_dev);
1010 
1011 error_failed_eq_table:
1012 	hns_roce_cmd_cleanup(hr_dev);
1013 
1014 error_failed_cmd_init:
1015 	if (hr_dev->hw->cmq_exit)
1016 		hr_dev->hw->cmq_exit(hr_dev);
1017 
1018 	return ret;
1019 }
1020 
1021 void hns_roce_exit(struct hns_roce_dev *hr_dev)
1022 {
1023 	hns_roce_unregister_device(hr_dev);
1024 
1025 	if (hr_dev->hw->hw_exit)
1026 		hr_dev->hw->hw_exit(hr_dev);
1027 	hns_roce_cleanup_bitmap(hr_dev);
1028 	hns_roce_cleanup_hem(hr_dev);
1029 
1030 	if (hr_dev->cmd_mod)
1031 		hns_roce_cmd_use_polling(hr_dev);
1032 
1033 	hr_dev->hw->cleanup_eq(hr_dev);
1034 	hns_roce_cmd_cleanup(hr_dev);
1035 	if (hr_dev->hw->cmq_exit)
1036 		hr_dev->hw->cmq_exit(hr_dev);
1037 }
1038 
1039 MODULE_LICENSE("Dual BSD/GPL");
1040 MODULE_AUTHOR("Wei Hu <xavier.huwei@huawei.com>");
1041 MODULE_AUTHOR("Nenglong Zhao <zhaonenglong@hisilicon.com>");
1042 MODULE_AUTHOR("Lijun Ou <oulijun@huawei.com>");
1043 MODULE_DESCRIPTION("HNS RoCE Driver");
1044