xref: /linux/drivers/infiniband/hw/hns/hns_roce_qp.c (revision f86fd32d)
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 
34 #include <linux/pci.h>
35 #include <linux/platform_device.h>
36 #include <rdma/ib_addr.h>
37 #include <rdma/ib_umem.h>
38 #include <rdma/uverbs_ioctl.h>
39 #include "hns_roce_common.h"
40 #include "hns_roce_device.h"
41 #include "hns_roce_hem.h"
42 #include <rdma/hns-abi.h>
43 
44 #define SQP_NUM				(2 * HNS_ROCE_MAX_PORTS)
45 
46 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
47 {
48 	struct device *dev = hr_dev->dev;
49 	struct hns_roce_qp *qp;
50 
51 	xa_lock(&hr_dev->qp_table_xa);
52 	qp = __hns_roce_qp_lookup(hr_dev, qpn);
53 	if (qp)
54 		atomic_inc(&qp->refcount);
55 	xa_unlock(&hr_dev->qp_table_xa);
56 
57 	if (!qp) {
58 		dev_warn(dev, "Async event for bogus QP %08x\n", qpn);
59 		return;
60 	}
61 
62 	qp->event(qp, (enum hns_roce_event)event_type);
63 
64 	if (atomic_dec_and_test(&qp->refcount))
65 		complete(&qp->free);
66 }
67 
68 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
69 				 enum hns_roce_event type)
70 {
71 	struct ib_event event;
72 	struct ib_qp *ibqp = &hr_qp->ibqp;
73 
74 	if (ibqp->event_handler) {
75 		event.device = ibqp->device;
76 		event.element.qp = ibqp;
77 		switch (type) {
78 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
79 			event.event = IB_EVENT_PATH_MIG;
80 			break;
81 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
82 			event.event = IB_EVENT_COMM_EST;
83 			break;
84 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
85 			event.event = IB_EVENT_SQ_DRAINED;
86 			break;
87 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
88 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
89 			break;
90 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
91 			event.event = IB_EVENT_QP_FATAL;
92 			break;
93 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
94 			event.event = IB_EVENT_PATH_MIG_ERR;
95 			break;
96 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
97 			event.event = IB_EVENT_QP_REQ_ERR;
98 			break;
99 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
100 			event.event = IB_EVENT_QP_ACCESS_ERR;
101 			break;
102 		default:
103 			dev_dbg(ibqp->device->dev.parent, "roce_ib: Unexpected event type %d on QP %06lx\n",
104 				type, hr_qp->qpn);
105 			return;
106 		}
107 		ibqp->event_handler(&event, ibqp->qp_context);
108 	}
109 }
110 
111 static int hns_roce_reserve_range_qp(struct hns_roce_dev *hr_dev, int cnt,
112 				     int align, unsigned long *base)
113 {
114 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
115 
116 	return hns_roce_bitmap_alloc_range(&qp_table->bitmap, cnt, align,
117 					   base) ?
118 		       -ENOMEM :
119 		       0;
120 }
121 
122 enum hns_roce_qp_state to_hns_roce_state(enum ib_qp_state state)
123 {
124 	switch (state) {
125 	case IB_QPS_RESET:
126 		return HNS_ROCE_QP_STATE_RST;
127 	case IB_QPS_INIT:
128 		return HNS_ROCE_QP_STATE_INIT;
129 	case IB_QPS_RTR:
130 		return HNS_ROCE_QP_STATE_RTR;
131 	case IB_QPS_RTS:
132 		return HNS_ROCE_QP_STATE_RTS;
133 	case IB_QPS_SQD:
134 		return HNS_ROCE_QP_STATE_SQD;
135 	case IB_QPS_ERR:
136 		return HNS_ROCE_QP_STATE_ERR;
137 	default:
138 		return HNS_ROCE_QP_NUM_STATE;
139 	}
140 }
141 
142 static int hns_roce_gsi_qp_alloc(struct hns_roce_dev *hr_dev, unsigned long qpn,
143 				 struct hns_roce_qp *hr_qp)
144 {
145 	struct xarray *xa = &hr_dev->qp_table_xa;
146 	int ret;
147 
148 	if (!qpn)
149 		return -EINVAL;
150 
151 	hr_qp->qpn = qpn;
152 	atomic_set(&hr_qp->refcount, 1);
153 	init_completion(&hr_qp->free);
154 
155 	ret = xa_err(xa_store_irq(xa, hr_qp->qpn & (hr_dev->caps.num_qps - 1),
156 				hr_qp, GFP_KERNEL));
157 	if (ret)
158 		dev_err(hr_dev->dev, "QPC xa_store failed\n");
159 
160 	return ret;
161 }
162 
163 static int hns_roce_qp_alloc(struct hns_roce_dev *hr_dev, unsigned long qpn,
164 			     struct hns_roce_qp *hr_qp)
165 {
166 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
167 	struct device *dev = hr_dev->dev;
168 	int ret;
169 
170 	if (!qpn)
171 		return -EINVAL;
172 
173 	hr_qp->qpn = qpn;
174 
175 	/* Alloc memory for QPC */
176 	ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
177 	if (ret) {
178 		dev_err(dev, "QPC table get failed\n");
179 		goto err_out;
180 	}
181 
182 	/* Alloc memory for IRRL */
183 	ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
184 	if (ret) {
185 		dev_err(dev, "IRRL table get failed\n");
186 		goto err_put_qp;
187 	}
188 
189 	if (hr_dev->caps.trrl_entry_sz) {
190 		/* Alloc memory for TRRL */
191 		ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
192 					 hr_qp->qpn);
193 		if (ret) {
194 			dev_err(dev, "TRRL table get failed\n");
195 			goto err_put_irrl;
196 		}
197 	}
198 
199 	if (hr_dev->caps.sccc_entry_sz) {
200 		/* Alloc memory for SCC CTX */
201 		ret = hns_roce_table_get(hr_dev, &qp_table->sccc_table,
202 					 hr_qp->qpn);
203 		if (ret) {
204 			dev_err(dev, "SCC CTX table get failed\n");
205 			goto err_put_trrl;
206 		}
207 	}
208 
209 	ret = hns_roce_gsi_qp_alloc(hr_dev, qpn, hr_qp);
210 	if (ret)
211 		goto err_put_sccc;
212 
213 	return 0;
214 
215 err_put_sccc:
216 	if (hr_dev->caps.sccc_entry_sz)
217 		hns_roce_table_put(hr_dev, &qp_table->sccc_table,
218 				   hr_qp->qpn);
219 
220 err_put_trrl:
221 	if (hr_dev->caps.trrl_entry_sz)
222 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
223 
224 err_put_irrl:
225 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
226 
227 err_put_qp:
228 	hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
229 
230 err_out:
231 	return ret;
232 }
233 
234 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
235 {
236 	struct xarray *xa = &hr_dev->qp_table_xa;
237 	unsigned long flags;
238 
239 	xa_lock_irqsave(xa, flags);
240 	__xa_erase(xa, hr_qp->qpn & (hr_dev->caps.num_qps - 1));
241 	xa_unlock_irqrestore(xa, flags);
242 }
243 
244 void hns_roce_qp_free(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
245 {
246 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
247 
248 	if (atomic_dec_and_test(&hr_qp->refcount))
249 		complete(&hr_qp->free);
250 	wait_for_completion(&hr_qp->free);
251 
252 	if ((hr_qp->ibqp.qp_type) != IB_QPT_GSI) {
253 		if (hr_dev->caps.trrl_entry_sz)
254 			hns_roce_table_put(hr_dev, &qp_table->trrl_table,
255 					   hr_qp->qpn);
256 		hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
257 	}
258 }
259 
260 void hns_roce_release_range_qp(struct hns_roce_dev *hr_dev, int base_qpn,
261 			       int cnt)
262 {
263 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
264 
265 	if (base_qpn < hr_dev->caps.reserved_qps)
266 		return;
267 
268 	hns_roce_bitmap_free_range(&qp_table->bitmap, base_qpn, cnt, BITMAP_RR);
269 }
270 
271 static int hns_roce_set_rq_size(struct hns_roce_dev *hr_dev,
272 				struct ib_qp_cap *cap, bool is_user, int has_rq,
273 				struct hns_roce_qp *hr_qp)
274 {
275 	struct device *dev = hr_dev->dev;
276 	u32 max_cnt;
277 
278 	/* Check the validity of QP support capacity */
279 	if (cap->max_recv_wr > hr_dev->caps.max_wqes ||
280 	    cap->max_recv_sge > hr_dev->caps.max_rq_sg) {
281 		dev_err(dev, "RQ WR or sge error!max_recv_wr=%d max_recv_sge=%d\n",
282 			cap->max_recv_wr, cap->max_recv_sge);
283 		return -EINVAL;
284 	}
285 
286 	/* If srq exist, set zero for relative number of rq */
287 	if (!has_rq) {
288 		hr_qp->rq.wqe_cnt = 0;
289 		hr_qp->rq.max_gs = 0;
290 		cap->max_recv_wr = 0;
291 		cap->max_recv_sge = 0;
292 	} else {
293 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge)) {
294 			dev_err(dev, "user space no need config max_recv_wr max_recv_sge\n");
295 			return -EINVAL;
296 		}
297 
298 		if (hr_dev->caps.min_wqes)
299 			max_cnt = max(cap->max_recv_wr, hr_dev->caps.min_wqes);
300 		else
301 			max_cnt = cap->max_recv_wr;
302 
303 		hr_qp->rq.wqe_cnt = roundup_pow_of_two(max_cnt);
304 
305 		if ((u32)hr_qp->rq.wqe_cnt > hr_dev->caps.max_wqes) {
306 			dev_err(dev, "while setting rq size, rq.wqe_cnt too large\n");
307 			return -EINVAL;
308 		}
309 
310 		max_cnt = max(1U, cap->max_recv_sge);
311 		hr_qp->rq.max_gs = roundup_pow_of_two(max_cnt);
312 		if (hr_dev->caps.max_rq_sg <= 2)
313 			hr_qp->rq.wqe_shift =
314 					ilog2(hr_dev->caps.max_rq_desc_sz);
315 		else
316 			hr_qp->rq.wqe_shift =
317 					ilog2(hr_dev->caps.max_rq_desc_sz
318 					      * hr_qp->rq.max_gs);
319 	}
320 
321 	cap->max_recv_wr = hr_qp->rq.wqe_cnt;
322 	cap->max_recv_sge = hr_qp->rq.max_gs;
323 
324 	return 0;
325 }
326 
327 static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
328 					struct ib_qp_cap *cap,
329 					struct hns_roce_ib_create_qp *ucmd)
330 {
331 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
332 	u8 max_sq_stride = ilog2(roundup_sq_stride);
333 
334 	/* Sanity check SQ size before proceeding */
335 	if (ucmd->log_sq_stride > max_sq_stride ||
336 	    ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
337 		ibdev_err(&hr_dev->ib_dev, "check SQ size error!\n");
338 		return -EINVAL;
339 	}
340 
341 	if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
342 		ibdev_err(&hr_dev->ib_dev, "SQ sge error! max_send_sge=%d\n",
343 			  cap->max_send_sge);
344 		return -EINVAL;
345 	}
346 
347 	return 0;
348 }
349 
350 static int hns_roce_set_user_sq_size(struct hns_roce_dev *hr_dev,
351 				     struct ib_qp_cap *cap,
352 				     struct hns_roce_qp *hr_qp,
353 				     struct hns_roce_ib_create_qp *ucmd)
354 {
355 	u32 ex_sge_num;
356 	u32 page_size;
357 	u32 max_cnt;
358 	int ret;
359 
360 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &hr_qp->sq.wqe_cnt) ||
361 	    hr_qp->sq.wqe_cnt > hr_dev->caps.max_wqes)
362 		return -EINVAL;
363 
364 	ret = check_sq_size_with_integrity(hr_dev, cap, ucmd);
365 	if (ret) {
366 		ibdev_err(&hr_dev->ib_dev, "Sanity check sq size failed\n");
367 		return ret;
368 	}
369 
370 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
371 
372 	max_cnt = max(1U, cap->max_send_sge);
373 	if (hr_dev->caps.max_sq_sg <= 2)
374 		hr_qp->sq.max_gs = roundup_pow_of_two(max_cnt);
375 	else
376 		hr_qp->sq.max_gs = max_cnt;
377 
378 	if (hr_qp->sq.max_gs > 2)
379 		hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
380 							(hr_qp->sq.max_gs - 2));
381 
382 	if ((hr_qp->sq.max_gs > 2) && (hr_dev->pci_dev->revision == 0x20)) {
383 		if (hr_qp->sge.sge_cnt > hr_dev->caps.max_extend_sg) {
384 			dev_err(hr_dev->dev,
385 				"The extended sge cnt error! sge_cnt=%d\n",
386 				hr_qp->sge.sge_cnt);
387 			return -EINVAL;
388 		}
389 	}
390 
391 	hr_qp->sge.sge_shift = 4;
392 	ex_sge_num = hr_qp->sge.sge_cnt;
393 
394 	/* Get buf size, SQ and RQ  are aligned to page_szie */
395 	if (hr_dev->caps.max_sq_sg <= 2) {
396 		hr_qp->buff_size = round_up((hr_qp->rq.wqe_cnt <<
397 					     hr_qp->rq.wqe_shift), PAGE_SIZE) +
398 				   round_up((hr_qp->sq.wqe_cnt <<
399 					     hr_qp->sq.wqe_shift), PAGE_SIZE);
400 
401 		hr_qp->sq.offset = 0;
402 		hr_qp->rq.offset = round_up((hr_qp->sq.wqe_cnt <<
403 					     hr_qp->sq.wqe_shift), PAGE_SIZE);
404 	} else {
405 		page_size = 1 << (hr_dev->caps.mtt_buf_pg_sz + PAGE_SHIFT);
406 		hr_qp->sge.sge_cnt = ex_sge_num ?
407 		   max(page_size / (1 << hr_qp->sge.sge_shift), ex_sge_num) : 0;
408 		hr_qp->buff_size = round_up((hr_qp->rq.wqe_cnt <<
409 					     hr_qp->rq.wqe_shift), page_size) +
410 				   round_up((hr_qp->sge.sge_cnt <<
411 					     hr_qp->sge.sge_shift), page_size) +
412 				   round_up((hr_qp->sq.wqe_cnt <<
413 					     hr_qp->sq.wqe_shift), page_size);
414 
415 		hr_qp->sq.offset = 0;
416 		if (ex_sge_num) {
417 			hr_qp->sge.offset = round_up((hr_qp->sq.wqe_cnt <<
418 						      hr_qp->sq.wqe_shift),
419 						     page_size);
420 			hr_qp->rq.offset = hr_qp->sge.offset +
421 					   round_up((hr_qp->sge.sge_cnt <<
422 						     hr_qp->sge.sge_shift),
423 						    page_size);
424 		} else {
425 			hr_qp->rq.offset = round_up((hr_qp->sq.wqe_cnt <<
426 						     hr_qp->sq.wqe_shift),
427 						    page_size);
428 		}
429 	}
430 
431 	return 0;
432 }
433 
434 static int split_wqe_buf_region(struct hns_roce_dev *hr_dev,
435 				struct hns_roce_qp *hr_qp,
436 				struct hns_roce_buf_region *regions,
437 				int region_max, int page_shift)
438 {
439 	int page_size = 1 << page_shift;
440 	bool is_extend_sge;
441 	int region_cnt = 0;
442 	int buf_size;
443 	int buf_cnt;
444 
445 	if (hr_qp->buff_size < 1 || region_max < 1)
446 		return region_cnt;
447 
448 	if (hr_qp->sge.sge_cnt > 0)
449 		is_extend_sge = true;
450 	else
451 		is_extend_sge = false;
452 
453 	/* sq region */
454 	if (is_extend_sge)
455 		buf_size = hr_qp->sge.offset - hr_qp->sq.offset;
456 	else
457 		buf_size = hr_qp->rq.offset - hr_qp->sq.offset;
458 
459 	if (buf_size > 0 && region_cnt < region_max) {
460 		buf_cnt = DIV_ROUND_UP(buf_size, page_size);
461 		hns_roce_init_buf_region(&regions[region_cnt],
462 					 hr_dev->caps.wqe_sq_hop_num,
463 					 hr_qp->sq.offset / page_size,
464 					 buf_cnt);
465 		region_cnt++;
466 	}
467 
468 	/* sge region */
469 	if (is_extend_sge) {
470 		buf_size = hr_qp->rq.offset - hr_qp->sge.offset;
471 		if (buf_size > 0 && region_cnt < region_max) {
472 			buf_cnt = DIV_ROUND_UP(buf_size, page_size);
473 			hns_roce_init_buf_region(&regions[region_cnt],
474 						 hr_dev->caps.wqe_sge_hop_num,
475 						 hr_qp->sge.offset / page_size,
476 						 buf_cnt);
477 			region_cnt++;
478 		}
479 	}
480 
481 	/* rq region */
482 	buf_size = hr_qp->buff_size - hr_qp->rq.offset;
483 	if (buf_size > 0) {
484 		buf_cnt = DIV_ROUND_UP(buf_size, page_size);
485 		hns_roce_init_buf_region(&regions[region_cnt],
486 					 hr_dev->caps.wqe_rq_hop_num,
487 					 hr_qp->rq.offset / page_size,
488 					 buf_cnt);
489 		region_cnt++;
490 	}
491 
492 	return region_cnt;
493 }
494 
495 static int calc_wqe_bt_page_shift(struct hns_roce_dev *hr_dev,
496 				  struct hns_roce_buf_region *regions,
497 				  int region_cnt)
498 {
499 	int bt_pg_shift;
500 	int ba_num;
501 	int ret;
502 
503 	bt_pg_shift = PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz;
504 
505 	/* all root ba entries must in one bt page */
506 	do {
507 		ba_num = (1 << bt_pg_shift) / BA_BYTE_LEN;
508 		ret = hns_roce_hem_list_calc_root_ba(regions, region_cnt,
509 						     ba_num);
510 		if (ret <= ba_num)
511 			break;
512 
513 		bt_pg_shift++;
514 	} while (ret > ba_num);
515 
516 	return bt_pg_shift - PAGE_SHIFT;
517 }
518 
519 static int set_extend_sge_param(struct hns_roce_dev *hr_dev,
520 				struct hns_roce_qp *hr_qp)
521 {
522 	struct device *dev = hr_dev->dev;
523 
524 	if (hr_qp->sq.max_gs > 2) {
525 		hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
526 				     (hr_qp->sq.max_gs - 2));
527 		hr_qp->sge.sge_shift = 4;
528 	}
529 
530 	/* ud sqwqe's sge use extend sge */
531 	if (hr_dev->caps.max_sq_sg > 2 && hr_qp->ibqp.qp_type == IB_QPT_GSI) {
532 		hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
533 				     hr_qp->sq.max_gs);
534 		hr_qp->sge.sge_shift = 4;
535 	}
536 
537 	if ((hr_qp->sq.max_gs > 2) && hr_dev->pci_dev->revision == 0x20) {
538 		if (hr_qp->sge.sge_cnt > hr_dev->caps.max_extend_sg) {
539 			dev_err(dev, "The extended sge cnt error! sge_cnt=%d\n",
540 				hr_qp->sge.sge_cnt);
541 			return -EINVAL;
542 		}
543 	}
544 
545 	return 0;
546 }
547 
548 static int hns_roce_set_kernel_sq_size(struct hns_roce_dev *hr_dev,
549 				       struct ib_qp_cap *cap,
550 				       struct hns_roce_qp *hr_qp)
551 {
552 	struct device *dev = hr_dev->dev;
553 	u32 page_size;
554 	u32 max_cnt;
555 	int size;
556 	int ret;
557 
558 	if (cap->max_send_wr  > hr_dev->caps.max_wqes  ||
559 	    cap->max_send_sge > hr_dev->caps.max_sq_sg ||
560 	    cap->max_inline_data > hr_dev->caps.max_sq_inline) {
561 		dev_err(dev, "SQ WR or sge or inline data error!\n");
562 		return -EINVAL;
563 	}
564 
565 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
566 
567 	if (hr_dev->caps.min_wqes)
568 		max_cnt = max(cap->max_send_wr, hr_dev->caps.min_wqes);
569 	else
570 		max_cnt = cap->max_send_wr;
571 
572 	hr_qp->sq.wqe_cnt = roundup_pow_of_two(max_cnt);
573 	if ((u32)hr_qp->sq.wqe_cnt > hr_dev->caps.max_wqes) {
574 		dev_err(dev, "while setting kernel sq size, sq.wqe_cnt too large\n");
575 		return -EINVAL;
576 	}
577 
578 	/* Get data_seg numbers */
579 	max_cnt = max(1U, cap->max_send_sge);
580 	if (hr_dev->caps.max_sq_sg <= 2)
581 		hr_qp->sq.max_gs = roundup_pow_of_two(max_cnt);
582 	else
583 		hr_qp->sq.max_gs = max_cnt;
584 
585 	ret = set_extend_sge_param(hr_dev, hr_qp);
586 	if (ret) {
587 		dev_err(dev, "set extend sge parameters fail\n");
588 		return ret;
589 	}
590 
591 	/* Get buf size, SQ and RQ are aligned to PAGE_SIZE */
592 	page_size = 1 << (hr_dev->caps.mtt_buf_pg_sz + PAGE_SHIFT);
593 	hr_qp->sq.offset = 0;
594 	size = round_up(hr_qp->sq.wqe_cnt << hr_qp->sq.wqe_shift, page_size);
595 
596 	if (hr_dev->caps.max_sq_sg > 2 && hr_qp->sge.sge_cnt) {
597 		hr_qp->sge.sge_cnt = max(page_size/(1 << hr_qp->sge.sge_shift),
598 					 (u32)hr_qp->sge.sge_cnt);
599 		hr_qp->sge.offset = size;
600 		size += round_up(hr_qp->sge.sge_cnt << hr_qp->sge.sge_shift,
601 				 page_size);
602 	}
603 
604 	hr_qp->rq.offset = size;
605 	size += round_up((hr_qp->rq.wqe_cnt << hr_qp->rq.wqe_shift), page_size);
606 	hr_qp->buff_size = size;
607 
608 	/* Get wr and sge number which send */
609 	cap->max_send_wr = hr_qp->sq.wqe_cnt;
610 	cap->max_send_sge = hr_qp->sq.max_gs;
611 
612 	/* We don't support inline sends for kernel QPs (yet) */
613 	cap->max_inline_data = 0;
614 
615 	return 0;
616 }
617 
618 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
619 {
620 	if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
621 		return 0;
622 
623 	return 1;
624 }
625 
626 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
627 {
628 	if (attr->qp_type == IB_QPT_XRC_INI ||
629 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
630 	    !attr->cap.max_recv_wr)
631 		return 0;
632 
633 	return 1;
634 }
635 
636 static int alloc_rq_inline_buf(struct hns_roce_qp *hr_qp,
637 			       struct ib_qp_init_attr *init_attr)
638 {
639 	u32 max_recv_sge = init_attr->cap.max_recv_sge;
640 	struct hns_roce_rinl_wqe *wqe_list;
641 	u32 wqe_cnt = hr_qp->rq.wqe_cnt;
642 	int i;
643 
644 	/* allocate recv inline buf */
645 	wqe_list = kcalloc(wqe_cnt, sizeof(struct hns_roce_rinl_wqe),
646 			   GFP_KERNEL);
647 
648 	if (!wqe_list)
649 		goto err;
650 
651 	/* Allocate a continuous buffer for all inline sge we need */
652 	wqe_list[0].sg_list = kcalloc(wqe_cnt, (max_recv_sge *
653 				      sizeof(struct hns_roce_rinl_sge)),
654 				      GFP_KERNEL);
655 	if (!wqe_list[0].sg_list)
656 		goto err_wqe_list;
657 
658 	/* Assign buffers of sg_list to each inline wqe */
659 	for (i = 1; i < wqe_cnt; i++)
660 		wqe_list[i].sg_list = &wqe_list[0].sg_list[i * max_recv_sge];
661 
662 	hr_qp->rq_inl_buf.wqe_list = wqe_list;
663 	hr_qp->rq_inl_buf.wqe_cnt = wqe_cnt;
664 
665 	return 0;
666 
667 err_wqe_list:
668 	kfree(wqe_list);
669 
670 err:
671 	return -ENOMEM;
672 }
673 
674 static void free_rq_inline_buf(struct hns_roce_qp *hr_qp)
675 {
676 	kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
677 	kfree(hr_qp->rq_inl_buf.wqe_list);
678 }
679 
680 static void add_qp_to_list(struct hns_roce_dev *hr_dev,
681 			   struct hns_roce_qp *hr_qp,
682 			   struct ib_cq *send_cq, struct ib_cq *recv_cq)
683 {
684 	struct hns_roce_cq *hr_send_cq, *hr_recv_cq;
685 	unsigned long flags;
686 
687 	hr_send_cq = send_cq ? to_hr_cq(send_cq) : NULL;
688 	hr_recv_cq = recv_cq ? to_hr_cq(recv_cq) : NULL;
689 
690 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
691 	hns_roce_lock_cqs(hr_send_cq, hr_recv_cq);
692 
693 	list_add_tail(&hr_qp->node, &hr_dev->qp_list);
694 	if (hr_send_cq)
695 		list_add_tail(&hr_qp->sq_node, &hr_send_cq->sq_list);
696 	if (hr_recv_cq)
697 		list_add_tail(&hr_qp->rq_node, &hr_recv_cq->rq_list);
698 
699 	hns_roce_unlock_cqs(hr_send_cq, hr_recv_cq);
700 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
701 }
702 
703 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
704 				     struct ib_pd *ib_pd,
705 				     struct ib_qp_init_attr *init_attr,
706 				     struct ib_udata *udata, unsigned long sqpn,
707 				     struct hns_roce_qp *hr_qp)
708 {
709 	dma_addr_t *buf_list[ARRAY_SIZE(hr_qp->regions)] = { NULL };
710 	struct device *dev = hr_dev->dev;
711 	struct hns_roce_ib_create_qp ucmd;
712 	struct hns_roce_ib_create_qp_resp resp = {};
713 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
714 		udata, struct hns_roce_ucontext, ibucontext);
715 	struct hns_roce_buf_region *r;
716 	unsigned long qpn = 0;
717 	u32 page_shift;
718 	int buf_count;
719 	int ret;
720 	int i;
721 
722 	mutex_init(&hr_qp->mutex);
723 	spin_lock_init(&hr_qp->sq.lock);
724 	spin_lock_init(&hr_qp->rq.lock);
725 
726 	hr_qp->state = IB_QPS_RESET;
727 
728 	hr_qp->ibqp.qp_type = init_attr->qp_type;
729 
730 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
731 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
732 	else
733 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
734 
735 	ret = hns_roce_set_rq_size(hr_dev, &init_attr->cap, udata,
736 				   hns_roce_qp_has_rq(init_attr), hr_qp);
737 	if (ret) {
738 		dev_err(dev, "hns_roce_set_rq_size failed\n");
739 		goto err_out;
740 	}
741 
742 	if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) &&
743 	    hns_roce_qp_has_rq(init_attr)) {
744 		ret = alloc_rq_inline_buf(hr_qp, init_attr);
745 		if (ret) {
746 			dev_err(dev, "allocate receive inline buffer failed\n");
747 			goto err_out;
748 		}
749 	}
750 
751 	page_shift = PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
752 	if (udata) {
753 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
754 			dev_err(dev, "ib_copy_from_udata error for create qp\n");
755 			ret = -EFAULT;
756 			goto err_alloc_rq_inline_buf;
757 		}
758 
759 		ret = hns_roce_set_user_sq_size(hr_dev, &init_attr->cap, hr_qp,
760 						&ucmd);
761 		if (ret) {
762 			dev_err(dev, "hns_roce_set_user_sq_size error for create qp\n");
763 			goto err_alloc_rq_inline_buf;
764 		}
765 
766 		hr_qp->umem = ib_umem_get(ib_pd->device, ucmd.buf_addr,
767 					  hr_qp->buff_size, 0);
768 		if (IS_ERR(hr_qp->umem)) {
769 			dev_err(dev, "ib_umem_get error for create qp\n");
770 			ret = PTR_ERR(hr_qp->umem);
771 			goto err_alloc_rq_inline_buf;
772 		}
773 		hr_qp->region_cnt = split_wqe_buf_region(hr_dev, hr_qp,
774 				hr_qp->regions, ARRAY_SIZE(hr_qp->regions),
775 				page_shift);
776 		ret = hns_roce_alloc_buf_list(hr_qp->regions, buf_list,
777 					      hr_qp->region_cnt);
778 		if (ret) {
779 			dev_err(dev, "alloc buf_list error for create qp\n");
780 			goto err_alloc_list;
781 		}
782 
783 		for (i = 0; i < hr_qp->region_cnt; i++) {
784 			r = &hr_qp->regions[i];
785 			buf_count = hns_roce_get_umem_bufs(hr_dev,
786 					buf_list[i], r->count, r->offset,
787 					hr_qp->umem, page_shift);
788 			if (buf_count != r->count) {
789 				dev_err(dev,
790 					"get umem buf err, expect %d,ret %d.\n",
791 					r->count, buf_count);
792 				ret = -ENOBUFS;
793 				goto err_get_bufs;
794 			}
795 		}
796 
797 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SQ_RECORD_DB) &&
798 		    (udata->inlen >= sizeof(ucmd)) &&
799 		    (udata->outlen >= sizeof(resp)) &&
800 		    hns_roce_qp_has_sq(init_attr)) {
801 			ret = hns_roce_db_map_user(uctx, udata, ucmd.sdb_addr,
802 						   &hr_qp->sdb);
803 			if (ret) {
804 				dev_err(dev, "sq record doorbell map failed!\n");
805 				goto err_get_bufs;
806 			}
807 
808 			/* indicate kernel supports sq record db */
809 			resp.cap_flags |= HNS_ROCE_SUPPORT_SQ_RECORD_DB;
810 			hr_qp->sdb_en = 1;
811 		}
812 
813 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
814 		    (udata->outlen >= sizeof(resp)) &&
815 		    hns_roce_qp_has_rq(init_attr)) {
816 			ret = hns_roce_db_map_user(uctx, udata, ucmd.db_addr,
817 						   &hr_qp->rdb);
818 			if (ret) {
819 				dev_err(dev, "rq record doorbell map failed!\n");
820 				goto err_sq_dbmap;
821 			}
822 
823 			/* indicate kernel supports rq record db */
824 			resp.cap_flags |= HNS_ROCE_SUPPORT_RQ_RECORD_DB;
825 			hr_qp->rdb_en = 1;
826 		}
827 	} else {
828 		if (init_attr->create_flags &
829 		    IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
830 			dev_err(dev, "init_attr->create_flags error!\n");
831 			ret = -EINVAL;
832 			goto err_alloc_rq_inline_buf;
833 		}
834 
835 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO) {
836 			dev_err(dev, "init_attr->create_flags error!\n");
837 			ret = -EINVAL;
838 			goto err_alloc_rq_inline_buf;
839 		}
840 
841 		/* Set SQ size */
842 		ret = hns_roce_set_kernel_sq_size(hr_dev, &init_attr->cap,
843 						  hr_qp);
844 		if (ret) {
845 			dev_err(dev, "hns_roce_set_kernel_sq_size error!\n");
846 			goto err_alloc_rq_inline_buf;
847 		}
848 
849 		/* QP doorbell register address */
850 		hr_qp->sq.db_reg_l = hr_dev->reg_base + hr_dev->sdb_offset +
851 				     DB_REG_OFFSET * hr_dev->priv_uar.index;
852 		hr_qp->rq.db_reg_l = hr_dev->reg_base + hr_dev->odb_offset +
853 				     DB_REG_OFFSET * hr_dev->priv_uar.index;
854 
855 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
856 		    hns_roce_qp_has_rq(init_attr)) {
857 			ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
858 			if (ret) {
859 				dev_err(dev, "rq record doorbell alloc failed!\n");
860 				goto err_alloc_rq_inline_buf;
861 			}
862 			*hr_qp->rdb.db_record = 0;
863 			hr_qp->rdb_en = 1;
864 		}
865 
866 		/* Allocate QP buf */
867 		if (hns_roce_buf_alloc(hr_dev, hr_qp->buff_size,
868 				       (1 << page_shift) * 2,
869 				       &hr_qp->hr_buf, page_shift)) {
870 			dev_err(dev, "hns_roce_buf_alloc error!\n");
871 			ret = -ENOMEM;
872 			goto err_db;
873 		}
874 		hr_qp->region_cnt = split_wqe_buf_region(hr_dev, hr_qp,
875 				hr_qp->regions, ARRAY_SIZE(hr_qp->regions),
876 				page_shift);
877 		ret = hns_roce_alloc_buf_list(hr_qp->regions, buf_list,
878 					      hr_qp->region_cnt);
879 		if (ret) {
880 			dev_err(dev, "alloc buf_list error for create qp!\n");
881 			goto err_alloc_list;
882 		}
883 
884 		for (i = 0; i < hr_qp->region_cnt; i++) {
885 			r = &hr_qp->regions[i];
886 			buf_count = hns_roce_get_kmem_bufs(hr_dev,
887 					buf_list[i], r->count, r->offset,
888 					&hr_qp->hr_buf);
889 			if (buf_count != r->count) {
890 				dev_err(dev,
891 					"get kmem buf err, expect %d,ret %d.\n",
892 					r->count, buf_count);
893 				ret = -ENOBUFS;
894 				goto err_get_bufs;
895 			}
896 		}
897 
898 		hr_qp->sq.wrid = kcalloc(hr_qp->sq.wqe_cnt, sizeof(u64),
899 					 GFP_KERNEL);
900 		if (ZERO_OR_NULL_PTR(hr_qp->sq.wrid)) {
901 			ret = -ENOMEM;
902 			goto err_get_bufs;
903 		}
904 
905 		if (hr_qp->rq.wqe_cnt) {
906 			hr_qp->rq.wrid = kcalloc(hr_qp->rq.wqe_cnt, sizeof(u64),
907 						 GFP_KERNEL);
908 			if (ZERO_OR_NULL_PTR(hr_qp->rq.wrid)) {
909 				ret = -ENOMEM;
910 				goto err_sq_wrid;
911 			}
912 		}
913 	}
914 
915 	if (sqpn) {
916 		qpn = sqpn;
917 	} else {
918 		/* Get QPN */
919 		ret = hns_roce_reserve_range_qp(hr_dev, 1, 1, &qpn);
920 		if (ret) {
921 			dev_err(dev, "hns_roce_reserve_range_qp alloc qpn error\n");
922 			goto err_wrid;
923 		}
924 	}
925 
926 	hr_qp->wqe_bt_pg_shift = calc_wqe_bt_page_shift(hr_dev, hr_qp->regions,
927 							hr_qp->region_cnt);
928 	hns_roce_mtr_init(&hr_qp->mtr, PAGE_SHIFT + hr_qp->wqe_bt_pg_shift,
929 			  page_shift);
930 	ret = hns_roce_mtr_attach(hr_dev, &hr_qp->mtr, buf_list,
931 				  hr_qp->regions, hr_qp->region_cnt);
932 	if (ret) {
933 		dev_err(dev, "mtr attach error for create qp\n");
934 		goto err_mtr;
935 	}
936 
937 	if (init_attr->qp_type == IB_QPT_GSI &&
938 	    hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
939 		/* In v1 engine, GSI QP context in RoCE engine's register */
940 		ret = hns_roce_gsi_qp_alloc(hr_dev, qpn, hr_qp);
941 		if (ret) {
942 			dev_err(dev, "hns_roce_qp_alloc failed!\n");
943 			goto err_qpn;
944 		}
945 	} else {
946 		ret = hns_roce_qp_alloc(hr_dev, qpn, hr_qp);
947 		if (ret) {
948 			dev_err(dev, "hns_roce_qp_alloc failed!\n");
949 			goto err_qpn;
950 		}
951 	}
952 
953 	if (sqpn)
954 		hr_qp->doorbell_qpn = 1;
955 	else
956 		hr_qp->doorbell_qpn = (u32)hr_qp->qpn;
957 
958 	if (udata) {
959 		ret = ib_copy_to_udata(udata, &resp,
960 				       min(udata->outlen, sizeof(resp)));
961 		if (ret)
962 			goto err_qp;
963 	}
964 
965 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
966 		ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
967 		if (ret)
968 			goto err_qp;
969 	}
970 
971 	hr_qp->event = hns_roce_ib_qp_event;
972 
973 	add_qp_to_list(hr_dev, hr_qp, init_attr->send_cq, init_attr->recv_cq);
974 
975 	hns_roce_free_buf_list(buf_list, hr_qp->region_cnt);
976 
977 	return 0;
978 
979 err_qp:
980 	if (init_attr->qp_type == IB_QPT_GSI &&
981 		hr_dev->hw_rev == HNS_ROCE_HW_VER1)
982 		hns_roce_qp_remove(hr_dev, hr_qp);
983 	else
984 		hns_roce_qp_free(hr_dev, hr_qp);
985 
986 err_qpn:
987 	if (!sqpn)
988 		hns_roce_release_range_qp(hr_dev, qpn, 1);
989 
990 err_mtr:
991 	hns_roce_mtr_cleanup(hr_dev, &hr_qp->mtr);
992 
993 err_wrid:
994 	if (udata) {
995 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
996 		    (udata->outlen >= sizeof(resp)) &&
997 		    hns_roce_qp_has_rq(init_attr))
998 			hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
999 	} else {
1000 		if (hr_qp->rq.wqe_cnt)
1001 			kfree(hr_qp->rq.wrid);
1002 	}
1003 
1004 err_sq_dbmap:
1005 	if (udata)
1006 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SQ_RECORD_DB) &&
1007 		    (udata->inlen >= sizeof(ucmd)) &&
1008 		    (udata->outlen >= sizeof(resp)) &&
1009 		    hns_roce_qp_has_sq(init_attr))
1010 			hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
1011 
1012 err_sq_wrid:
1013 	if (!udata)
1014 		kfree(hr_qp->sq.wrid);
1015 
1016 err_get_bufs:
1017 	hns_roce_free_buf_list(buf_list, hr_qp->region_cnt);
1018 
1019 err_alloc_list:
1020 	if (!hr_qp->umem)
1021 		hns_roce_buf_free(hr_dev, hr_qp->buff_size, &hr_qp->hr_buf);
1022 	ib_umem_release(hr_qp->umem);
1023 
1024 err_db:
1025 	if (!udata && hns_roce_qp_has_rq(init_attr) &&
1026 	    (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB))
1027 		hns_roce_free_db(hr_dev, &hr_qp->rdb);
1028 
1029 err_alloc_rq_inline_buf:
1030 	if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) &&
1031 	     hns_roce_qp_has_rq(init_attr))
1032 		free_rq_inline_buf(hr_qp);
1033 
1034 err_out:
1035 	return ret;
1036 }
1037 
1038 struct ib_qp *hns_roce_create_qp(struct ib_pd *pd,
1039 				 struct ib_qp_init_attr *init_attr,
1040 				 struct ib_udata *udata)
1041 {
1042 	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
1043 	struct ib_device *ibdev = &hr_dev->ib_dev;
1044 	struct hns_roce_qp *hr_qp;
1045 	int ret;
1046 
1047 	switch (init_attr->qp_type) {
1048 	case IB_QPT_RC: {
1049 		hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
1050 		if (!hr_qp)
1051 			return ERR_PTR(-ENOMEM);
1052 
1053 		ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, 0,
1054 						hr_qp);
1055 		if (ret) {
1056 			ibdev_err(ibdev, "Create QP 0x%06lx failed(%d)\n",
1057 				  hr_qp->qpn, ret);
1058 			kfree(hr_qp);
1059 			return ERR_PTR(ret);
1060 		}
1061 
1062 		hr_qp->ibqp.qp_num = hr_qp->qpn;
1063 
1064 		break;
1065 	}
1066 	case IB_QPT_GSI: {
1067 		/* Userspace is not allowed to create special QPs: */
1068 		if (udata) {
1069 			ibdev_err(ibdev, "not support usr space GSI\n");
1070 			return ERR_PTR(-EINVAL);
1071 		}
1072 
1073 		hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
1074 		if (!hr_qp)
1075 			return ERR_PTR(-ENOMEM);
1076 
1077 		hr_qp->port = init_attr->port_num - 1;
1078 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
1079 
1080 		/* when hw version is v1, the sqpn is allocated */
1081 		if (hr_dev->caps.max_sq_sg <= 2)
1082 			hr_qp->ibqp.qp_num = HNS_ROCE_MAX_PORTS +
1083 					     hr_dev->iboe.phy_port[hr_qp->port];
1084 		else
1085 			hr_qp->ibqp.qp_num = 1;
1086 
1087 		ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata,
1088 						hr_qp->ibqp.qp_num, hr_qp);
1089 		if (ret) {
1090 			ibdev_err(ibdev, "Create GSI QP failed!\n");
1091 			kfree(hr_qp);
1092 			return ERR_PTR(ret);
1093 		}
1094 
1095 		break;
1096 	}
1097 	default:{
1098 		ibdev_err(ibdev, "not support QP type %d\n",
1099 			  init_attr->qp_type);
1100 		return ERR_PTR(-EINVAL);
1101 	}
1102 	}
1103 
1104 	return &hr_qp->ibqp;
1105 }
1106 
1107 int to_hr_qp_type(int qp_type)
1108 {
1109 	int transport_type;
1110 
1111 	if (qp_type == IB_QPT_RC)
1112 		transport_type = SERV_TYPE_RC;
1113 	else if (qp_type == IB_QPT_UC)
1114 		transport_type = SERV_TYPE_UC;
1115 	else if (qp_type == IB_QPT_UD)
1116 		transport_type = SERV_TYPE_UD;
1117 	else if (qp_type == IB_QPT_GSI)
1118 		transport_type = SERV_TYPE_UD;
1119 	else
1120 		transport_type = -1;
1121 
1122 	return transport_type;
1123 }
1124 
1125 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1126 			      struct hns_roce_qp *hr_qp,
1127 			      struct ib_qp_attr *attr, int attr_mask)
1128 {
1129 	enum ib_mtu active_mtu;
1130 	int p;
1131 
1132 	p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1133 	active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1134 
1135 	if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1136 	    attr->path_mtu > hr_dev->caps.max_mtu) ||
1137 	    attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1138 		ibdev_err(&hr_dev->ib_dev,
1139 			"attr path_mtu(%d)invalid while modify qp",
1140 			attr->path_mtu);
1141 		return -EINVAL;
1142 	}
1143 
1144 	return 0;
1145 }
1146 
1147 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1148 				  int attr_mask)
1149 {
1150 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1151 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1152 	int p;
1153 
1154 	if ((attr_mask & IB_QP_PORT) &&
1155 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1156 		ibdev_err(&hr_dev->ib_dev,
1157 			"attr port_num invalid.attr->port_num=%d\n",
1158 			attr->port_num);
1159 		return -EINVAL;
1160 	}
1161 
1162 	if (attr_mask & IB_QP_PKEY_INDEX) {
1163 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1164 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1165 			ibdev_err(&hr_dev->ib_dev,
1166 				"attr pkey_index invalid.attr->pkey_index=%d\n",
1167 				attr->pkey_index);
1168 			return -EINVAL;
1169 		}
1170 	}
1171 
1172 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1173 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1174 		ibdev_err(&hr_dev->ib_dev,
1175 			"attr max_rd_atomic invalid.attr->max_rd_atomic=%d\n",
1176 			attr->max_rd_atomic);
1177 		return -EINVAL;
1178 	}
1179 
1180 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1181 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1182 		ibdev_err(&hr_dev->ib_dev,
1183 			"attr max_dest_rd_atomic invalid.attr->max_dest_rd_atomic=%d\n",
1184 			attr->max_dest_rd_atomic);
1185 		return -EINVAL;
1186 	}
1187 
1188 	if (attr_mask & IB_QP_PATH_MTU)
1189 		return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1190 
1191 	return 0;
1192 }
1193 
1194 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1195 		       int attr_mask, struct ib_udata *udata)
1196 {
1197 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1198 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1199 	enum ib_qp_state cur_state, new_state;
1200 	int ret = -EINVAL;
1201 
1202 	mutex_lock(&hr_qp->mutex);
1203 
1204 	cur_state = attr_mask & IB_QP_CUR_STATE ?
1205 		    attr->cur_qp_state : (enum ib_qp_state)hr_qp->state;
1206 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1207 
1208 	if (ibqp->uobject &&
1209 	    (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1210 		if (hr_qp->sdb_en == 1) {
1211 			hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1212 
1213 			if (hr_qp->rdb_en == 1)
1214 				hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1215 		} else {
1216 			ibdev_warn(&hr_dev->ib_dev,
1217 				  "flush cqe is not supported in userspace!\n");
1218 			goto out;
1219 		}
1220 	}
1221 
1222 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1223 				attr_mask)) {
1224 		ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1225 		goto out;
1226 	}
1227 
1228 	ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1229 	if (ret)
1230 		goto out;
1231 
1232 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1233 		if (hr_dev->caps.min_wqes) {
1234 			ret = -EPERM;
1235 			ibdev_err(&hr_dev->ib_dev,
1236 				"cur_state=%d new_state=%d\n", cur_state,
1237 				new_state);
1238 		} else {
1239 			ret = 0;
1240 		}
1241 
1242 		goto out;
1243 	}
1244 
1245 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1246 				    new_state);
1247 
1248 out:
1249 	mutex_unlock(&hr_qp->mutex);
1250 
1251 	return ret;
1252 }
1253 
1254 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1255 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1256 {
1257 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1258 		__acquire(&send_cq->lock);
1259 		__acquire(&recv_cq->lock);
1260 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1261 		spin_lock_irq(&send_cq->lock);
1262 		__acquire(&recv_cq->lock);
1263 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1264 		spin_lock_irq(&recv_cq->lock);
1265 		__acquire(&send_cq->lock);
1266 	} else if (send_cq == recv_cq) {
1267 		spin_lock_irq(&send_cq->lock);
1268 		__acquire(&recv_cq->lock);
1269 	} else if (send_cq->cqn < recv_cq->cqn) {
1270 		spin_lock_irq(&send_cq->lock);
1271 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1272 	} else {
1273 		spin_lock_irq(&recv_cq->lock);
1274 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1275 	}
1276 }
1277 
1278 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1279 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1280 			 __releases(&recv_cq->lock)
1281 {
1282 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1283 		__release(&recv_cq->lock);
1284 		__release(&send_cq->lock);
1285 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1286 		__release(&recv_cq->lock);
1287 		spin_unlock(&send_cq->lock);
1288 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1289 		__release(&send_cq->lock);
1290 		spin_unlock(&recv_cq->lock);
1291 	} else if (send_cq == recv_cq) {
1292 		__release(&recv_cq->lock);
1293 		spin_unlock_irq(&send_cq->lock);
1294 	} else if (send_cq->cqn < recv_cq->cqn) {
1295 		spin_unlock(&recv_cq->lock);
1296 		spin_unlock_irq(&send_cq->lock);
1297 	} else {
1298 		spin_unlock(&send_cq->lock);
1299 		spin_unlock_irq(&recv_cq->lock);
1300 	}
1301 }
1302 
1303 static void *get_wqe(struct hns_roce_qp *hr_qp, int offset)
1304 {
1305 
1306 	return hns_roce_buf_offset(&hr_qp->hr_buf, offset);
1307 }
1308 
1309 void *get_recv_wqe(struct hns_roce_qp *hr_qp, int n)
1310 {
1311 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1312 }
1313 
1314 void *get_send_wqe(struct hns_roce_qp *hr_qp, int n)
1315 {
1316 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1317 }
1318 
1319 void *get_send_extend_sge(struct hns_roce_qp *hr_qp, int n)
1320 {
1321 	return hns_roce_buf_offset(&hr_qp->hr_buf, hr_qp->sge.offset +
1322 					(n << hr_qp->sge.sge_shift));
1323 }
1324 
1325 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, int nreq,
1326 			  struct ib_cq *ib_cq)
1327 {
1328 	struct hns_roce_cq *hr_cq;
1329 	u32 cur;
1330 
1331 	cur = hr_wq->head - hr_wq->tail;
1332 	if (likely(cur + nreq < hr_wq->wqe_cnt))
1333 		return false;
1334 
1335 	hr_cq = to_hr_cq(ib_cq);
1336 	spin_lock(&hr_cq->lock);
1337 	cur = hr_wq->head - hr_wq->tail;
1338 	spin_unlock(&hr_cq->lock);
1339 
1340 	return cur + nreq >= hr_wq->wqe_cnt;
1341 }
1342 
1343 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1344 {
1345 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1346 	int reserved_from_top = 0;
1347 	int reserved_from_bot;
1348 	int ret;
1349 
1350 	mutex_init(&qp_table->scc_mutex);
1351 	xa_init(&hr_dev->qp_table_xa);
1352 
1353 	reserved_from_bot = hr_dev->caps.reserved_qps;
1354 
1355 	ret = hns_roce_bitmap_init(&qp_table->bitmap, hr_dev->caps.num_qps,
1356 				   hr_dev->caps.num_qps - 1, reserved_from_bot,
1357 				   reserved_from_top);
1358 	if (ret) {
1359 		dev_err(hr_dev->dev, "qp bitmap init failed!error=%d\n",
1360 			ret);
1361 		return ret;
1362 	}
1363 
1364 	return 0;
1365 }
1366 
1367 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1368 {
1369 	hns_roce_bitmap_cleanup(&hr_dev->qp_table.bitmap);
1370 }
1371