xref: /openbmc/linux/drivers/infiniband/hw/hns/hns_roce_qp.c (revision 7b73a9c8e26ce5769c41d4b787767c10fe7269db)
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 = HNS_ROCE_ALIGN_UP((hr_qp->rq.wqe_cnt <<
397 					     hr_qp->rq.wqe_shift), PAGE_SIZE) +
398 				   HNS_ROCE_ALIGN_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 = HNS_ROCE_ALIGN_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 = HNS_ROCE_ALIGN_UP((hr_qp->rq.wqe_cnt <<
409 					     hr_qp->rq.wqe_shift), page_size) +
410 				   HNS_ROCE_ALIGN_UP((hr_qp->sge.sge_cnt <<
411 					     hr_qp->sge.sge_shift), page_size) +
412 				   HNS_ROCE_ALIGN_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 = HNS_ROCE_ALIGN_UP(
418 							(hr_qp->sq.wqe_cnt <<
419 							hr_qp->sq.wqe_shift),
420 							page_size);
421 			hr_qp->rq.offset = hr_qp->sge.offset +
422 					HNS_ROCE_ALIGN_UP((hr_qp->sge.sge_cnt <<
423 						hr_qp->sge.sge_shift),
424 						page_size);
425 		} else {
426 			hr_qp->rq.offset = HNS_ROCE_ALIGN_UP(
427 							(hr_qp->sq.wqe_cnt <<
428 							hr_qp->sq.wqe_shift),
429 							page_size);
430 		}
431 	}
432 
433 	return 0;
434 }
435 
436 static int split_wqe_buf_region(struct hns_roce_dev *hr_dev,
437 				struct hns_roce_qp *hr_qp,
438 				struct hns_roce_buf_region *regions,
439 				int region_max, int page_shift)
440 {
441 	int page_size = 1 << page_shift;
442 	bool is_extend_sge;
443 	int region_cnt = 0;
444 	int buf_size;
445 	int buf_cnt;
446 
447 	if (hr_qp->buff_size < 1 || region_max < 1)
448 		return region_cnt;
449 
450 	if (hr_qp->sge.sge_cnt > 0)
451 		is_extend_sge = true;
452 	else
453 		is_extend_sge = false;
454 
455 	/* sq region */
456 	if (is_extend_sge)
457 		buf_size = hr_qp->sge.offset - hr_qp->sq.offset;
458 	else
459 		buf_size = hr_qp->rq.offset - hr_qp->sq.offset;
460 
461 	if (buf_size > 0 && region_cnt < region_max) {
462 		buf_cnt = DIV_ROUND_UP(buf_size, page_size);
463 		hns_roce_init_buf_region(&regions[region_cnt],
464 					 hr_dev->caps.wqe_sq_hop_num,
465 					 hr_qp->sq.offset / page_size,
466 					 buf_cnt);
467 		region_cnt++;
468 	}
469 
470 	/* sge region */
471 	if (is_extend_sge) {
472 		buf_size = hr_qp->rq.offset - hr_qp->sge.offset;
473 		if (buf_size > 0 && region_cnt < region_max) {
474 			buf_cnt = DIV_ROUND_UP(buf_size, page_size);
475 			hns_roce_init_buf_region(&regions[region_cnt],
476 						 hr_dev->caps.wqe_sge_hop_num,
477 						 hr_qp->sge.offset / page_size,
478 						 buf_cnt);
479 			region_cnt++;
480 		}
481 	}
482 
483 	/* rq region */
484 	buf_size = hr_qp->buff_size - hr_qp->rq.offset;
485 	if (buf_size > 0) {
486 		buf_cnt = DIV_ROUND_UP(buf_size, page_size);
487 		hns_roce_init_buf_region(&regions[region_cnt],
488 					 hr_dev->caps.wqe_rq_hop_num,
489 					 hr_qp->rq.offset / page_size,
490 					 buf_cnt);
491 		region_cnt++;
492 	}
493 
494 	return region_cnt;
495 }
496 
497 static int calc_wqe_bt_page_shift(struct hns_roce_dev *hr_dev,
498 				  struct hns_roce_buf_region *regions,
499 				  int region_cnt)
500 {
501 	int bt_pg_shift;
502 	int ba_num;
503 	int ret;
504 
505 	bt_pg_shift = PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz;
506 
507 	/* all root ba entries must in one bt page */
508 	do {
509 		ba_num = (1 << bt_pg_shift) / BA_BYTE_LEN;
510 		ret = hns_roce_hem_list_calc_root_ba(regions, region_cnt,
511 						     ba_num);
512 		if (ret <= ba_num)
513 			break;
514 
515 		bt_pg_shift++;
516 	} while (ret > ba_num);
517 
518 	return bt_pg_shift - PAGE_SHIFT;
519 }
520 
521 static int set_extend_sge_param(struct hns_roce_dev *hr_dev,
522 				struct hns_roce_qp *hr_qp)
523 {
524 	struct device *dev = hr_dev->dev;
525 
526 	if (hr_qp->sq.max_gs > 2) {
527 		hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
528 				     (hr_qp->sq.max_gs - 2));
529 		hr_qp->sge.sge_shift = 4;
530 	}
531 
532 	/* ud sqwqe's sge use extend sge */
533 	if (hr_dev->caps.max_sq_sg > 2 && hr_qp->ibqp.qp_type == IB_QPT_GSI) {
534 		hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
535 				     hr_qp->sq.max_gs);
536 		hr_qp->sge.sge_shift = 4;
537 	}
538 
539 	if ((hr_qp->sq.max_gs > 2) && hr_dev->pci_dev->revision == 0x20) {
540 		if (hr_qp->sge.sge_cnt > hr_dev->caps.max_extend_sg) {
541 			dev_err(dev, "The extended sge cnt error! sge_cnt=%d\n",
542 				hr_qp->sge.sge_cnt);
543 			return -EINVAL;
544 		}
545 	}
546 
547 	return 0;
548 }
549 
550 static int hns_roce_set_kernel_sq_size(struct hns_roce_dev *hr_dev,
551 				       struct ib_qp_cap *cap,
552 				       struct hns_roce_qp *hr_qp)
553 {
554 	struct device *dev = hr_dev->dev;
555 	u32 page_size;
556 	u32 max_cnt;
557 	int size;
558 	int ret;
559 
560 	if (cap->max_send_wr  > hr_dev->caps.max_wqes  ||
561 	    cap->max_send_sge > hr_dev->caps.max_sq_sg ||
562 	    cap->max_inline_data > hr_dev->caps.max_sq_inline) {
563 		dev_err(dev, "SQ WR or sge or inline data error!\n");
564 		return -EINVAL;
565 	}
566 
567 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
568 
569 	if (hr_dev->caps.min_wqes)
570 		max_cnt = max(cap->max_send_wr, hr_dev->caps.min_wqes);
571 	else
572 		max_cnt = cap->max_send_wr;
573 
574 	hr_qp->sq.wqe_cnt = roundup_pow_of_two(max_cnt);
575 	if ((u32)hr_qp->sq.wqe_cnt > hr_dev->caps.max_wqes) {
576 		dev_err(dev, "while setting kernel sq size, sq.wqe_cnt too large\n");
577 		return -EINVAL;
578 	}
579 
580 	/* Get data_seg numbers */
581 	max_cnt = max(1U, cap->max_send_sge);
582 	if (hr_dev->caps.max_sq_sg <= 2)
583 		hr_qp->sq.max_gs = roundup_pow_of_two(max_cnt);
584 	else
585 		hr_qp->sq.max_gs = max_cnt;
586 
587 	ret = set_extend_sge_param(hr_dev, hr_qp);
588 	if (ret) {
589 		dev_err(dev, "set extend sge parameters fail\n");
590 		return ret;
591 	}
592 
593 	/* Get buf size, SQ and RQ are aligned to PAGE_SIZE */
594 	page_size = 1 << (hr_dev->caps.mtt_buf_pg_sz + PAGE_SHIFT);
595 	hr_qp->sq.offset = 0;
596 	size = HNS_ROCE_ALIGN_UP(hr_qp->sq.wqe_cnt << hr_qp->sq.wqe_shift,
597 				 page_size);
598 
599 	if (hr_dev->caps.max_sq_sg > 2 && hr_qp->sge.sge_cnt) {
600 		hr_qp->sge.sge_cnt = max(page_size/(1 << hr_qp->sge.sge_shift),
601 					(u32)hr_qp->sge.sge_cnt);
602 		hr_qp->sge.offset = size;
603 		size += HNS_ROCE_ALIGN_UP(hr_qp->sge.sge_cnt <<
604 					  hr_qp->sge.sge_shift, page_size);
605 	}
606 
607 	hr_qp->rq.offset = size;
608 	size += HNS_ROCE_ALIGN_UP((hr_qp->rq.wqe_cnt << hr_qp->rq.wqe_shift),
609 				  page_size);
610 	hr_qp->buff_size = size;
611 
612 	/* Get wr and sge number which send */
613 	cap->max_send_wr = hr_qp->sq.wqe_cnt;
614 	cap->max_send_sge = hr_qp->sq.max_gs;
615 
616 	/* We don't support inline sends for kernel QPs (yet) */
617 	cap->max_inline_data = 0;
618 
619 	return 0;
620 }
621 
622 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
623 {
624 	if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
625 		return 0;
626 
627 	return 1;
628 }
629 
630 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
631 {
632 	if (attr->qp_type == IB_QPT_XRC_INI ||
633 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
634 	    !attr->cap.max_recv_wr)
635 		return 0;
636 
637 	return 1;
638 }
639 
640 static int alloc_rq_inline_buf(struct hns_roce_qp *hr_qp,
641 			       struct ib_qp_init_attr *init_attr)
642 {
643 	u32 max_recv_sge = init_attr->cap.max_recv_sge;
644 	struct hns_roce_rinl_wqe *wqe_list;
645 	u32 wqe_cnt = hr_qp->rq.wqe_cnt;
646 	int i;
647 
648 	/* allocate recv inline buf */
649 	wqe_list = kcalloc(wqe_cnt, sizeof(struct hns_roce_rinl_wqe),
650 			   GFP_KERNEL);
651 
652 	if (!wqe_list)
653 		goto err;
654 
655 	/* Allocate a continuous buffer for all inline sge we need */
656 	wqe_list[0].sg_list = kcalloc(wqe_cnt, (max_recv_sge *
657 				      sizeof(struct hns_roce_rinl_sge)),
658 				      GFP_KERNEL);
659 	if (!wqe_list[0].sg_list)
660 		goto err_wqe_list;
661 
662 	/* Assign buffers of sg_list to each inline wqe */
663 	for (i = 1; i < wqe_cnt; i++)
664 		wqe_list[i].sg_list = &wqe_list[0].sg_list[i * max_recv_sge];
665 
666 	hr_qp->rq_inl_buf.wqe_list = wqe_list;
667 	hr_qp->rq_inl_buf.wqe_cnt = wqe_cnt;
668 
669 	return 0;
670 
671 err_wqe_list:
672 	kfree(wqe_list);
673 
674 err:
675 	return -ENOMEM;
676 }
677 
678 static void free_rq_inline_buf(struct hns_roce_qp *hr_qp)
679 {
680 	kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
681 	kfree(hr_qp->rq_inl_buf.wqe_list);
682 }
683 
684 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
685 				     struct ib_pd *ib_pd,
686 				     struct ib_qp_init_attr *init_attr,
687 				     struct ib_udata *udata, unsigned long sqpn,
688 				     struct hns_roce_qp *hr_qp)
689 {
690 	dma_addr_t *buf_list[ARRAY_SIZE(hr_qp->regions)] = { NULL };
691 	struct device *dev = hr_dev->dev;
692 	struct hns_roce_ib_create_qp ucmd;
693 	struct hns_roce_ib_create_qp_resp resp = {};
694 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
695 		udata, struct hns_roce_ucontext, ibucontext);
696 	struct hns_roce_buf_region *r;
697 	unsigned long qpn = 0;
698 	u32 page_shift;
699 	int buf_count;
700 	int ret;
701 	int i;
702 
703 	mutex_init(&hr_qp->mutex);
704 	spin_lock_init(&hr_qp->sq.lock);
705 	spin_lock_init(&hr_qp->rq.lock);
706 
707 	hr_qp->state = IB_QPS_RESET;
708 
709 	hr_qp->ibqp.qp_type = init_attr->qp_type;
710 
711 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
712 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
713 	else
714 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
715 
716 	ret = hns_roce_set_rq_size(hr_dev, &init_attr->cap, udata,
717 				   hns_roce_qp_has_rq(init_attr), hr_qp);
718 	if (ret) {
719 		dev_err(dev, "hns_roce_set_rq_size failed\n");
720 		goto err_out;
721 	}
722 
723 	if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) &&
724 	    hns_roce_qp_has_rq(init_attr)) {
725 		ret = alloc_rq_inline_buf(hr_qp, init_attr);
726 		if (ret) {
727 			dev_err(dev, "allocate receive inline buffer failed\n");
728 			goto err_out;
729 		}
730 	}
731 
732 	page_shift = PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
733 	if (udata) {
734 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
735 			dev_err(dev, "ib_copy_from_udata error for create qp\n");
736 			ret = -EFAULT;
737 			goto err_alloc_rq_inline_buf;
738 		}
739 
740 		ret = hns_roce_set_user_sq_size(hr_dev, &init_attr->cap, hr_qp,
741 						&ucmd);
742 		if (ret) {
743 			dev_err(dev, "hns_roce_set_user_sq_size error for create qp\n");
744 			goto err_alloc_rq_inline_buf;
745 		}
746 
747 		hr_qp->umem = ib_umem_get(udata, ucmd.buf_addr,
748 					  hr_qp->buff_size, 0);
749 		if (IS_ERR(hr_qp->umem)) {
750 			dev_err(dev, "ib_umem_get error for create qp\n");
751 			ret = PTR_ERR(hr_qp->umem);
752 			goto err_alloc_rq_inline_buf;
753 		}
754 		hr_qp->region_cnt = split_wqe_buf_region(hr_dev, hr_qp,
755 				hr_qp->regions, ARRAY_SIZE(hr_qp->regions),
756 				page_shift);
757 		ret = hns_roce_alloc_buf_list(hr_qp->regions, buf_list,
758 					      hr_qp->region_cnt);
759 		if (ret) {
760 			dev_err(dev, "alloc buf_list error for create qp\n");
761 			goto err_alloc_list;
762 		}
763 
764 		for (i = 0; i < hr_qp->region_cnt; i++) {
765 			r = &hr_qp->regions[i];
766 			buf_count = hns_roce_get_umem_bufs(hr_dev,
767 					buf_list[i], r->count, r->offset,
768 					hr_qp->umem, page_shift);
769 			if (buf_count != r->count) {
770 				dev_err(dev,
771 					"get umem buf err, expect %d,ret %d.\n",
772 					r->count, buf_count);
773 				ret = -ENOBUFS;
774 				goto err_get_bufs;
775 			}
776 		}
777 
778 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SQ_RECORD_DB) &&
779 		    (udata->inlen >= sizeof(ucmd)) &&
780 		    (udata->outlen >= sizeof(resp)) &&
781 		    hns_roce_qp_has_sq(init_attr)) {
782 			ret = hns_roce_db_map_user(uctx, udata, ucmd.sdb_addr,
783 						   &hr_qp->sdb);
784 			if (ret) {
785 				dev_err(dev, "sq record doorbell map failed!\n");
786 				goto err_get_bufs;
787 			}
788 
789 			/* indicate kernel supports sq record db */
790 			resp.cap_flags |= HNS_ROCE_SUPPORT_SQ_RECORD_DB;
791 			hr_qp->sdb_en = 1;
792 		}
793 
794 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
795 		    (udata->outlen >= sizeof(resp)) &&
796 		    hns_roce_qp_has_rq(init_attr)) {
797 			ret = hns_roce_db_map_user(uctx, udata, ucmd.db_addr,
798 						   &hr_qp->rdb);
799 			if (ret) {
800 				dev_err(dev, "rq record doorbell map failed!\n");
801 				goto err_sq_dbmap;
802 			}
803 
804 			/* indicate kernel supports rq record db */
805 			resp.cap_flags |= HNS_ROCE_SUPPORT_RQ_RECORD_DB;
806 			hr_qp->rdb_en = 1;
807 		}
808 	} else {
809 		if (init_attr->create_flags &
810 		    IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
811 			dev_err(dev, "init_attr->create_flags error!\n");
812 			ret = -EINVAL;
813 			goto err_alloc_rq_inline_buf;
814 		}
815 
816 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO) {
817 			dev_err(dev, "init_attr->create_flags error!\n");
818 			ret = -EINVAL;
819 			goto err_alloc_rq_inline_buf;
820 		}
821 
822 		/* Set SQ size */
823 		ret = hns_roce_set_kernel_sq_size(hr_dev, &init_attr->cap,
824 						  hr_qp);
825 		if (ret) {
826 			dev_err(dev, "hns_roce_set_kernel_sq_size error!\n");
827 			goto err_alloc_rq_inline_buf;
828 		}
829 
830 		/* QP doorbell register address */
831 		hr_qp->sq.db_reg_l = hr_dev->reg_base + hr_dev->sdb_offset +
832 				     DB_REG_OFFSET * hr_dev->priv_uar.index;
833 		hr_qp->rq.db_reg_l = hr_dev->reg_base + hr_dev->odb_offset +
834 				     DB_REG_OFFSET * hr_dev->priv_uar.index;
835 
836 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
837 		    hns_roce_qp_has_rq(init_attr)) {
838 			ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
839 			if (ret) {
840 				dev_err(dev, "rq record doorbell alloc failed!\n");
841 				goto err_alloc_rq_inline_buf;
842 			}
843 			*hr_qp->rdb.db_record = 0;
844 			hr_qp->rdb_en = 1;
845 		}
846 
847 		/* Allocate QP buf */
848 		if (hns_roce_buf_alloc(hr_dev, hr_qp->buff_size,
849 				       (1 << page_shift) * 2,
850 				       &hr_qp->hr_buf, page_shift)) {
851 			dev_err(dev, "hns_roce_buf_alloc error!\n");
852 			ret = -ENOMEM;
853 			goto err_db;
854 		}
855 		hr_qp->region_cnt = split_wqe_buf_region(hr_dev, hr_qp,
856 				hr_qp->regions, ARRAY_SIZE(hr_qp->regions),
857 				page_shift);
858 		ret = hns_roce_alloc_buf_list(hr_qp->regions, buf_list,
859 					      hr_qp->region_cnt);
860 		if (ret) {
861 			dev_err(dev, "alloc buf_list error for create qp!\n");
862 			goto err_alloc_list;
863 		}
864 
865 		for (i = 0; i < hr_qp->region_cnt; i++) {
866 			r = &hr_qp->regions[i];
867 			buf_count = hns_roce_get_kmem_bufs(hr_dev,
868 					buf_list[i], r->count, r->offset,
869 					&hr_qp->hr_buf);
870 			if (buf_count != r->count) {
871 				dev_err(dev,
872 					"get kmem buf err, expect %d,ret %d.\n",
873 					r->count, buf_count);
874 				ret = -ENOBUFS;
875 				goto err_get_bufs;
876 			}
877 		}
878 
879 		hr_qp->sq.wrid = kcalloc(hr_qp->sq.wqe_cnt, sizeof(u64),
880 					 GFP_KERNEL);
881 		if (ZERO_OR_NULL_PTR(hr_qp->sq.wrid)) {
882 			ret = -ENOMEM;
883 			goto err_get_bufs;
884 		}
885 
886 		if (hr_qp->rq.wqe_cnt) {
887 			hr_qp->rq.wrid = kcalloc(hr_qp->rq.wqe_cnt, sizeof(u64),
888 						 GFP_KERNEL);
889 			if (ZERO_OR_NULL_PTR(hr_qp->rq.wrid)) {
890 				ret = -ENOMEM;
891 				goto err_sq_wrid;
892 			}
893 		}
894 	}
895 
896 	if (sqpn) {
897 		qpn = sqpn;
898 	} else {
899 		/* Get QPN */
900 		ret = hns_roce_reserve_range_qp(hr_dev, 1, 1, &qpn);
901 		if (ret) {
902 			dev_err(dev, "hns_roce_reserve_range_qp alloc qpn error\n");
903 			goto err_wrid;
904 		}
905 	}
906 
907 	hr_qp->wqe_bt_pg_shift = calc_wqe_bt_page_shift(hr_dev, hr_qp->regions,
908 							hr_qp->region_cnt);
909 	hns_roce_mtr_init(&hr_qp->mtr, PAGE_SHIFT + hr_qp->wqe_bt_pg_shift,
910 			  page_shift);
911 	ret = hns_roce_mtr_attach(hr_dev, &hr_qp->mtr, buf_list,
912 				  hr_qp->regions, hr_qp->region_cnt);
913 	if (ret) {
914 		dev_err(dev, "mtr attach error for create qp\n");
915 		goto err_mtr;
916 	}
917 
918 	if (init_attr->qp_type == IB_QPT_GSI &&
919 	    hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
920 		/* In v1 engine, GSI QP context in RoCE engine's register */
921 		ret = hns_roce_gsi_qp_alloc(hr_dev, qpn, hr_qp);
922 		if (ret) {
923 			dev_err(dev, "hns_roce_qp_alloc failed!\n");
924 			goto err_qpn;
925 		}
926 	} else {
927 		ret = hns_roce_qp_alloc(hr_dev, qpn, hr_qp);
928 		if (ret) {
929 			dev_err(dev, "hns_roce_qp_alloc failed!\n");
930 			goto err_qpn;
931 		}
932 	}
933 
934 	if (sqpn)
935 		hr_qp->doorbell_qpn = 1;
936 	else
937 		hr_qp->doorbell_qpn = (u32)hr_qp->qpn;
938 
939 	if (udata) {
940 		ret = ib_copy_to_udata(udata, &resp,
941 				       min(udata->outlen, sizeof(resp)));
942 		if (ret)
943 			goto err_qp;
944 	}
945 
946 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
947 		ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
948 		if (ret)
949 			goto err_qp;
950 	}
951 
952 	hr_qp->event = hns_roce_ib_qp_event;
953 	hns_roce_free_buf_list(buf_list, hr_qp->region_cnt);
954 
955 	return 0;
956 
957 err_qp:
958 	if (init_attr->qp_type == IB_QPT_GSI &&
959 		hr_dev->hw_rev == HNS_ROCE_HW_VER1)
960 		hns_roce_qp_remove(hr_dev, hr_qp);
961 	else
962 		hns_roce_qp_free(hr_dev, hr_qp);
963 
964 err_qpn:
965 	if (!sqpn)
966 		hns_roce_release_range_qp(hr_dev, qpn, 1);
967 
968 err_mtr:
969 	hns_roce_mtr_cleanup(hr_dev, &hr_qp->mtr);
970 
971 err_wrid:
972 	if (udata) {
973 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
974 		    (udata->outlen >= sizeof(resp)) &&
975 		    hns_roce_qp_has_rq(init_attr))
976 			hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
977 	} else {
978 		if (hr_qp->rq.wqe_cnt)
979 			kfree(hr_qp->rq.wrid);
980 	}
981 
982 err_sq_dbmap:
983 	if (udata)
984 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SQ_RECORD_DB) &&
985 		    (udata->inlen >= sizeof(ucmd)) &&
986 		    (udata->outlen >= sizeof(resp)) &&
987 		    hns_roce_qp_has_sq(init_attr))
988 			hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
989 
990 err_sq_wrid:
991 	if (!udata)
992 		kfree(hr_qp->sq.wrid);
993 
994 err_get_bufs:
995 	hns_roce_free_buf_list(buf_list, hr_qp->region_cnt);
996 
997 err_alloc_list:
998 	if (!hr_qp->umem)
999 		hns_roce_buf_free(hr_dev, hr_qp->buff_size, &hr_qp->hr_buf);
1000 	ib_umem_release(hr_qp->umem);
1001 
1002 err_db:
1003 	if (!udata && hns_roce_qp_has_rq(init_attr) &&
1004 	    (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB))
1005 		hns_roce_free_db(hr_dev, &hr_qp->rdb);
1006 
1007 err_alloc_rq_inline_buf:
1008 	if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) &&
1009 	     hns_roce_qp_has_rq(init_attr))
1010 		free_rq_inline_buf(hr_qp);
1011 
1012 err_out:
1013 	return ret;
1014 }
1015 
1016 struct ib_qp *hns_roce_create_qp(struct ib_pd *pd,
1017 				 struct ib_qp_init_attr *init_attr,
1018 				 struct ib_udata *udata)
1019 {
1020 	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
1021 	struct ib_device *ibdev = &hr_dev->ib_dev;
1022 	struct hns_roce_qp *hr_qp;
1023 	int ret;
1024 
1025 	switch (init_attr->qp_type) {
1026 	case IB_QPT_RC: {
1027 		hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
1028 		if (!hr_qp)
1029 			return ERR_PTR(-ENOMEM);
1030 
1031 		ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, 0,
1032 						hr_qp);
1033 		if (ret) {
1034 			ibdev_err(ibdev, "Create QP 0x%06lx failed(%d)\n",
1035 				  hr_qp->qpn, ret);
1036 			kfree(hr_qp);
1037 			return ERR_PTR(ret);
1038 		}
1039 
1040 		hr_qp->ibqp.qp_num = hr_qp->qpn;
1041 
1042 		break;
1043 	}
1044 	case IB_QPT_GSI: {
1045 		/* Userspace is not allowed to create special QPs: */
1046 		if (udata) {
1047 			ibdev_err(ibdev, "not support usr space GSI\n");
1048 			return ERR_PTR(-EINVAL);
1049 		}
1050 
1051 		hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
1052 		if (!hr_qp)
1053 			return ERR_PTR(-ENOMEM);
1054 
1055 		hr_qp->port = init_attr->port_num - 1;
1056 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
1057 
1058 		/* when hw version is v1, the sqpn is allocated */
1059 		if (hr_dev->caps.max_sq_sg <= 2)
1060 			hr_qp->ibqp.qp_num = HNS_ROCE_MAX_PORTS +
1061 					     hr_dev->iboe.phy_port[hr_qp->port];
1062 		else
1063 			hr_qp->ibqp.qp_num = 1;
1064 
1065 		ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata,
1066 						hr_qp->ibqp.qp_num, hr_qp);
1067 		if (ret) {
1068 			ibdev_err(ibdev, "Create GSI QP failed!\n");
1069 			kfree(hr_qp);
1070 			return ERR_PTR(ret);
1071 		}
1072 
1073 		break;
1074 	}
1075 	default:{
1076 		ibdev_err(ibdev, "not support QP type %d\n",
1077 			  init_attr->qp_type);
1078 		return ERR_PTR(-EINVAL);
1079 	}
1080 	}
1081 
1082 	return &hr_qp->ibqp;
1083 }
1084 
1085 int to_hr_qp_type(int qp_type)
1086 {
1087 	int transport_type;
1088 
1089 	if (qp_type == IB_QPT_RC)
1090 		transport_type = SERV_TYPE_RC;
1091 	else if (qp_type == IB_QPT_UC)
1092 		transport_type = SERV_TYPE_UC;
1093 	else if (qp_type == IB_QPT_UD)
1094 		transport_type = SERV_TYPE_UD;
1095 	else if (qp_type == IB_QPT_GSI)
1096 		transport_type = SERV_TYPE_UD;
1097 	else
1098 		transport_type = -1;
1099 
1100 	return transport_type;
1101 }
1102 
1103 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1104 			      struct hns_roce_qp *hr_qp,
1105 			      struct ib_qp_attr *attr, int attr_mask)
1106 {
1107 	enum ib_mtu active_mtu;
1108 	int p;
1109 
1110 	p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1111 	active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1112 
1113 	if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1114 	    attr->path_mtu > hr_dev->caps.max_mtu) ||
1115 	    attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1116 		ibdev_err(&hr_dev->ib_dev,
1117 			"attr path_mtu(%d)invalid while modify qp",
1118 			attr->path_mtu);
1119 		return -EINVAL;
1120 	}
1121 
1122 	return 0;
1123 }
1124 
1125 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1126 				  int attr_mask)
1127 {
1128 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1129 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1130 	int p;
1131 
1132 	if ((attr_mask & IB_QP_PORT) &&
1133 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1134 		ibdev_err(&hr_dev->ib_dev,
1135 			"attr port_num invalid.attr->port_num=%d\n",
1136 			attr->port_num);
1137 		return -EINVAL;
1138 	}
1139 
1140 	if (attr_mask & IB_QP_PKEY_INDEX) {
1141 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1142 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1143 			ibdev_err(&hr_dev->ib_dev,
1144 				"attr pkey_index invalid.attr->pkey_index=%d\n",
1145 				attr->pkey_index);
1146 			return -EINVAL;
1147 		}
1148 	}
1149 
1150 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1151 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1152 		ibdev_err(&hr_dev->ib_dev,
1153 			"attr max_rd_atomic invalid.attr->max_rd_atomic=%d\n",
1154 			attr->max_rd_atomic);
1155 		return -EINVAL;
1156 	}
1157 
1158 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1159 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1160 		ibdev_err(&hr_dev->ib_dev,
1161 			"attr max_dest_rd_atomic invalid.attr->max_dest_rd_atomic=%d\n",
1162 			attr->max_dest_rd_atomic);
1163 		return -EINVAL;
1164 	}
1165 
1166 	if (attr_mask & IB_QP_PATH_MTU)
1167 		return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1168 
1169 	return 0;
1170 }
1171 
1172 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1173 		       int attr_mask, struct ib_udata *udata)
1174 {
1175 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1176 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1177 	enum ib_qp_state cur_state, new_state;
1178 	int ret = -EINVAL;
1179 
1180 	mutex_lock(&hr_qp->mutex);
1181 
1182 	cur_state = attr_mask & IB_QP_CUR_STATE ?
1183 		    attr->cur_qp_state : (enum ib_qp_state)hr_qp->state;
1184 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1185 
1186 	if (ibqp->uobject &&
1187 	    (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1188 		if (hr_qp->sdb_en == 1) {
1189 			hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1190 
1191 			if (hr_qp->rdb_en == 1)
1192 				hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1193 		} else {
1194 			ibdev_warn(&hr_dev->ib_dev,
1195 				  "flush cqe is not supported in userspace!\n");
1196 			goto out;
1197 		}
1198 	}
1199 
1200 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1201 				attr_mask)) {
1202 		ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1203 		goto out;
1204 	}
1205 
1206 	ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1207 	if (ret)
1208 		goto out;
1209 
1210 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1211 		if (hr_dev->caps.min_wqes) {
1212 			ret = -EPERM;
1213 			ibdev_err(&hr_dev->ib_dev,
1214 				"cur_state=%d new_state=%d\n", cur_state,
1215 				new_state);
1216 		} else {
1217 			ret = 0;
1218 		}
1219 
1220 		goto out;
1221 	}
1222 
1223 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1224 				    new_state);
1225 
1226 out:
1227 	mutex_unlock(&hr_qp->mutex);
1228 
1229 	return ret;
1230 }
1231 
1232 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1233 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1234 {
1235 	if (send_cq == recv_cq) {
1236 		spin_lock_irq(&send_cq->lock);
1237 		__acquire(&recv_cq->lock);
1238 	} else if (send_cq->cqn < recv_cq->cqn) {
1239 		spin_lock_irq(&send_cq->lock);
1240 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1241 	} else {
1242 		spin_lock_irq(&recv_cq->lock);
1243 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1244 	}
1245 }
1246 
1247 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1248 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1249 			 __releases(&recv_cq->lock)
1250 {
1251 	if (send_cq == recv_cq) {
1252 		__release(&recv_cq->lock);
1253 		spin_unlock_irq(&send_cq->lock);
1254 	} else if (send_cq->cqn < recv_cq->cqn) {
1255 		spin_unlock(&recv_cq->lock);
1256 		spin_unlock_irq(&send_cq->lock);
1257 	} else {
1258 		spin_unlock(&send_cq->lock);
1259 		spin_unlock_irq(&recv_cq->lock);
1260 	}
1261 }
1262 
1263 static void *get_wqe(struct hns_roce_qp *hr_qp, int offset)
1264 {
1265 
1266 	return hns_roce_buf_offset(&hr_qp->hr_buf, offset);
1267 }
1268 
1269 void *get_recv_wqe(struct hns_roce_qp *hr_qp, int n)
1270 {
1271 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1272 }
1273 
1274 void *get_send_wqe(struct hns_roce_qp *hr_qp, int n)
1275 {
1276 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1277 }
1278 
1279 void *get_send_extend_sge(struct hns_roce_qp *hr_qp, int n)
1280 {
1281 	return hns_roce_buf_offset(&hr_qp->hr_buf, hr_qp->sge.offset +
1282 					(n << hr_qp->sge.sge_shift));
1283 }
1284 
1285 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, int nreq,
1286 			  struct ib_cq *ib_cq)
1287 {
1288 	struct hns_roce_cq *hr_cq;
1289 	u32 cur;
1290 
1291 	cur = hr_wq->head - hr_wq->tail;
1292 	if (likely(cur + nreq < hr_wq->wqe_cnt))
1293 		return false;
1294 
1295 	hr_cq = to_hr_cq(ib_cq);
1296 	spin_lock(&hr_cq->lock);
1297 	cur = hr_wq->head - hr_wq->tail;
1298 	spin_unlock(&hr_cq->lock);
1299 
1300 	return cur + nreq >= hr_wq->wqe_cnt;
1301 }
1302 
1303 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1304 {
1305 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1306 	int reserved_from_top = 0;
1307 	int reserved_from_bot;
1308 	int ret;
1309 
1310 	mutex_init(&qp_table->scc_mutex);
1311 	xa_init(&hr_dev->qp_table_xa);
1312 
1313 	reserved_from_bot = hr_dev->caps.reserved_qps;
1314 
1315 	ret = hns_roce_bitmap_init(&qp_table->bitmap, hr_dev->caps.num_qps,
1316 				   hr_dev->caps.num_qps - 1, reserved_from_bot,
1317 				   reserved_from_top);
1318 	if (ret) {
1319 		dev_err(hr_dev->dev, "qp bitmap init failed!error=%d\n",
1320 			ret);
1321 		return ret;
1322 	}
1323 
1324 	return 0;
1325 }
1326 
1327 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1328 {
1329 	hns_roce_bitmap_cleanup(&hr_dev->qp_table.bitmap);
1330 }
1331