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/platform_device.h>
35 #include <rdma/ib_addr.h>
36 #include <rdma/ib_umem.h>
37 #include "hns_roce_common.h"
38 #include "hns_roce_device.h"
39 #include "hns_roce_hem.h"
40 #include <rdma/hns-abi.h>
41 
42 #define SQP_NUM				(2 * HNS_ROCE_MAX_PORTS)
43 
44 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
45 {
46 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
47 	struct device *dev = hr_dev->dev;
48 	struct hns_roce_qp *qp;
49 
50 	spin_lock(&qp_table->lock);
51 
52 	qp = __hns_roce_qp_lookup(hr_dev, qpn);
53 	if (qp)
54 		atomic_inc(&qp->refcount);
55 
56 	spin_unlock(&qp_table->lock);
57 
58 	if (!qp) {
59 		dev_warn(dev, "Async event for bogus QP %08x\n", qpn);
60 		return;
61 	}
62 
63 	qp->event(qp, (enum hns_roce_event)event_type);
64 
65 	if (atomic_dec_and_test(&qp->refcount))
66 		complete(&qp->free);
67 }
68 EXPORT_SYMBOL_GPL(hns_roce_qp_event);
69 
70 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
71 				 enum hns_roce_event type)
72 {
73 	struct ib_event event;
74 	struct ib_qp *ibqp = &hr_qp->ibqp;
75 
76 	if (ibqp->event_handler) {
77 		event.device = ibqp->device;
78 		event.element.qp = ibqp;
79 		switch (type) {
80 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
81 			event.event = IB_EVENT_PATH_MIG;
82 			break;
83 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
84 			event.event = IB_EVENT_COMM_EST;
85 			break;
86 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
87 			event.event = IB_EVENT_SQ_DRAINED;
88 			break;
89 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
90 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
91 			break;
92 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
93 			event.event = IB_EVENT_QP_FATAL;
94 			break;
95 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
96 			event.event = IB_EVENT_PATH_MIG_ERR;
97 			break;
98 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
99 			event.event = IB_EVENT_QP_REQ_ERR;
100 			break;
101 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
102 			event.event = IB_EVENT_QP_ACCESS_ERR;
103 			break;
104 		default:
105 			dev_dbg(ibqp->device->dev.parent, "roce_ib: Unexpected event type %d on QP %06lx\n",
106 				type, hr_qp->qpn);
107 			return;
108 		}
109 		ibqp->event_handler(&event, ibqp->qp_context);
110 	}
111 }
112 
113 static int hns_roce_reserve_range_qp(struct hns_roce_dev *hr_dev, int cnt,
114 				     int align, unsigned long *base)
115 {
116 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
117 
118 	return hns_roce_bitmap_alloc_range(&qp_table->bitmap, cnt, align, base);
119 }
120 
121 enum hns_roce_qp_state to_hns_roce_state(enum ib_qp_state state)
122 {
123 	switch (state) {
124 	case IB_QPS_RESET:
125 		return HNS_ROCE_QP_STATE_RST;
126 	case IB_QPS_INIT:
127 		return HNS_ROCE_QP_STATE_INIT;
128 	case IB_QPS_RTR:
129 		return HNS_ROCE_QP_STATE_RTR;
130 	case IB_QPS_RTS:
131 		return HNS_ROCE_QP_STATE_RTS;
132 	case IB_QPS_SQD:
133 		return HNS_ROCE_QP_STATE_SQD;
134 	case IB_QPS_ERR:
135 		return HNS_ROCE_QP_STATE_ERR;
136 	default:
137 		return HNS_ROCE_QP_NUM_STATE;
138 	}
139 }
140 EXPORT_SYMBOL_GPL(to_hns_roce_state);
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 hns_roce_qp_table *qp_table = &hr_dev->qp_table;
146 	int ret;
147 
148 	if (!qpn)
149 		return -EINVAL;
150 
151 	hr_qp->qpn = qpn;
152 
153 	spin_lock_irq(&qp_table->lock);
154 	ret = radix_tree_insert(&hr_dev->qp_table_tree,
155 				hr_qp->qpn & (hr_dev->caps.num_qps - 1), hr_qp);
156 	spin_unlock_irq(&qp_table->lock);
157 	if (ret) {
158 		dev_err(hr_dev->dev, "QPC radix_tree_insert failed\n");
159 		goto err_put_irrl;
160 	}
161 
162 	atomic_set(&hr_qp->refcount, 1);
163 	init_completion(&hr_qp->free);
164 
165 	return 0;
166 
167 err_put_irrl:
168 
169 	return ret;
170 }
171 
172 static int hns_roce_qp_alloc(struct hns_roce_dev *hr_dev, unsigned long qpn,
173 			     struct hns_roce_qp *hr_qp)
174 {
175 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
176 	struct device *dev = hr_dev->dev;
177 	int ret;
178 
179 	if (!qpn)
180 		return -EINVAL;
181 
182 	hr_qp->qpn = qpn;
183 
184 	/* Alloc memory for QPC */
185 	ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
186 	if (ret) {
187 		dev_err(dev, "QPC table get failed\n");
188 		goto err_out;
189 	}
190 
191 	/* Alloc memory for IRRL */
192 	ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
193 	if (ret) {
194 		dev_err(dev, "IRRL table get failed\n");
195 		goto err_put_qp;
196 	}
197 
198 	if (hr_dev->caps.trrl_entry_sz) {
199 		/* Alloc memory for TRRL */
200 		ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
201 					 hr_qp->qpn);
202 		if (ret) {
203 			dev_err(dev, "TRRL table get failed\n");
204 			goto err_put_irrl;
205 		}
206 	}
207 
208 	spin_lock_irq(&qp_table->lock);
209 	ret = radix_tree_insert(&hr_dev->qp_table_tree,
210 				hr_qp->qpn & (hr_dev->caps.num_qps - 1), hr_qp);
211 	spin_unlock_irq(&qp_table->lock);
212 	if (ret) {
213 		dev_err(dev, "QPC radix_tree_insert failed\n");
214 		goto err_put_trrl;
215 	}
216 
217 	atomic_set(&hr_qp->refcount, 1);
218 	init_completion(&hr_qp->free);
219 
220 	return 0;
221 
222 err_put_trrl:
223 	if (hr_dev->caps.trrl_entry_sz)
224 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
225 
226 err_put_irrl:
227 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
228 
229 err_put_qp:
230 	hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
231 
232 err_out:
233 	return ret;
234 }
235 
236 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
237 {
238 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
239 	unsigned long flags;
240 
241 	spin_lock_irqsave(&qp_table->lock, flags);
242 	radix_tree_delete(&hr_dev->qp_table_tree,
243 			  hr_qp->qpn & (hr_dev->caps.num_qps - 1));
244 	spin_unlock_irqrestore(&qp_table->lock, flags);
245 }
246 EXPORT_SYMBOL_GPL(hns_roce_qp_remove);
247 
248 void hns_roce_qp_free(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
249 {
250 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
251 
252 	if (atomic_dec_and_test(&hr_qp->refcount))
253 		complete(&hr_qp->free);
254 	wait_for_completion(&hr_qp->free);
255 
256 	if ((hr_qp->ibqp.qp_type) != IB_QPT_GSI) {
257 		if (hr_dev->caps.trrl_entry_sz)
258 			hns_roce_table_put(hr_dev, &qp_table->trrl_table,
259 					   hr_qp->qpn);
260 		hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
261 		hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
262 	}
263 }
264 EXPORT_SYMBOL_GPL(hns_roce_qp_free);
265 
266 void hns_roce_release_range_qp(struct hns_roce_dev *hr_dev, int base_qpn,
267 			       int cnt)
268 {
269 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
270 
271 	if (base_qpn < SQP_NUM)
272 		return;
273 
274 	hns_roce_bitmap_free_range(&qp_table->bitmap, base_qpn, cnt, BITMAP_RR);
275 }
276 EXPORT_SYMBOL_GPL(hns_roce_release_range_qp);
277 
278 static int hns_roce_set_rq_size(struct hns_roce_dev *hr_dev,
279 				struct ib_qp_cap *cap, int is_user, int has_srq,
280 				struct hns_roce_qp *hr_qp)
281 {
282 	struct device *dev = hr_dev->dev;
283 	u32 max_cnt;
284 
285 	/* Check the validity of QP support capacity */
286 	if (cap->max_recv_wr > hr_dev->caps.max_wqes ||
287 	    cap->max_recv_sge > hr_dev->caps.max_rq_sg) {
288 		dev_err(dev, "RQ WR or sge error!max_recv_wr=%d max_recv_sge=%d\n",
289 			cap->max_recv_wr, cap->max_recv_sge);
290 		return -EINVAL;
291 	}
292 
293 	/* If srq exit, set zero for relative number of rq */
294 	if (has_srq) {
295 		if (cap->max_recv_wr) {
296 			dev_dbg(dev, "srq no need config max_recv_wr\n");
297 			return -EINVAL;
298 		}
299 
300 		hr_qp->rq.wqe_cnt = hr_qp->rq.max_gs = 0;
301 	} else {
302 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge)) {
303 			dev_err(dev, "user space no need config max_recv_wr max_recv_sge\n");
304 			return -EINVAL;
305 		}
306 
307 		if (hr_dev->caps.min_wqes)
308 			max_cnt = max(cap->max_recv_wr, hr_dev->caps.min_wqes);
309 		else
310 			max_cnt = cap->max_recv_wr;
311 
312 		hr_qp->rq.wqe_cnt = roundup_pow_of_two(max_cnt);
313 
314 		if ((u32)hr_qp->rq.wqe_cnt > hr_dev->caps.max_wqes) {
315 			dev_err(dev, "while setting rq size, rq.wqe_cnt too large\n");
316 			return -EINVAL;
317 		}
318 
319 		max_cnt = max(1U, cap->max_recv_sge);
320 		hr_qp->rq.max_gs = roundup_pow_of_two(max_cnt);
321 		if (hr_dev->caps.max_rq_sg <= 2)
322 			hr_qp->rq.wqe_shift =
323 					ilog2(hr_dev->caps.max_rq_desc_sz);
324 		else
325 			hr_qp->rq.wqe_shift =
326 					ilog2(hr_dev->caps.max_rq_desc_sz
327 					      * hr_qp->rq.max_gs);
328 	}
329 
330 	cap->max_recv_wr = hr_qp->rq.max_post = hr_qp->rq.wqe_cnt;
331 	cap->max_recv_sge = hr_qp->rq.max_gs;
332 
333 	return 0;
334 }
335 
336 static int hns_roce_set_user_sq_size(struct hns_roce_dev *hr_dev,
337 				     struct ib_qp_cap *cap,
338 				     struct hns_roce_qp *hr_qp,
339 				     struct hns_roce_ib_create_qp *ucmd)
340 {
341 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
342 	u8 max_sq_stride = ilog2(roundup_sq_stride);
343 	u32 page_size;
344 	u32 max_cnt;
345 
346 	/* Sanity check SQ size before proceeding */
347 	if ((u32)(1 << ucmd->log_sq_bb_count) > hr_dev->caps.max_wqes ||
348 	     ucmd->log_sq_stride > max_sq_stride ||
349 	     ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
350 		dev_err(hr_dev->dev, "check SQ size error!\n");
351 		return -EINVAL;
352 	}
353 
354 	if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
355 		dev_err(hr_dev->dev, "SQ sge error! max_send_sge=%d\n",
356 			cap->max_send_sge);
357 		return -EINVAL;
358 	}
359 
360 	hr_qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count;
361 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
362 
363 	max_cnt = max(1U, cap->max_send_sge);
364 	if (hr_dev->caps.max_sq_sg <= 2)
365 		hr_qp->sq.max_gs = roundup_pow_of_two(max_cnt);
366 	else
367 		hr_qp->sq.max_gs = max_cnt;
368 
369 	if (hr_qp->sq.max_gs > 2)
370 		hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
371 							(hr_qp->sq.max_gs - 2));
372 	hr_qp->sge.sge_shift = 4;
373 
374 	/* Get buf size, SQ and RQ  are aligned to page_szie */
375 	if (hr_dev->caps.max_sq_sg <= 2) {
376 		hr_qp->buff_size = HNS_ROCE_ALOGN_UP((hr_qp->rq.wqe_cnt <<
377 					     hr_qp->rq.wqe_shift), PAGE_SIZE) +
378 				   HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
379 					     hr_qp->sq.wqe_shift), PAGE_SIZE);
380 
381 		hr_qp->sq.offset = 0;
382 		hr_qp->rq.offset = HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
383 					     hr_qp->sq.wqe_shift), PAGE_SIZE);
384 	} else {
385 		page_size = 1 << (hr_dev->caps.mtt_buf_pg_sz + PAGE_SHIFT);
386 		hr_qp->buff_size = HNS_ROCE_ALOGN_UP((hr_qp->rq.wqe_cnt <<
387 					     hr_qp->rq.wqe_shift), page_size) +
388 				   HNS_ROCE_ALOGN_UP((hr_qp->sge.sge_cnt <<
389 					     hr_qp->sge.sge_shift), page_size) +
390 				   HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
391 					     hr_qp->sq.wqe_shift), page_size);
392 
393 		hr_qp->sq.offset = 0;
394 		if (hr_qp->sge.sge_cnt) {
395 			hr_qp->sge.offset = HNS_ROCE_ALOGN_UP(
396 							(hr_qp->sq.wqe_cnt <<
397 							hr_qp->sq.wqe_shift),
398 							page_size);
399 			hr_qp->rq.offset = hr_qp->sge.offset +
400 					HNS_ROCE_ALOGN_UP((hr_qp->sge.sge_cnt <<
401 						hr_qp->sge.sge_shift),
402 						page_size);
403 		} else {
404 			hr_qp->rq.offset = HNS_ROCE_ALOGN_UP(
405 							(hr_qp->sq.wqe_cnt <<
406 							hr_qp->sq.wqe_shift),
407 							page_size);
408 		}
409 	}
410 
411 	return 0;
412 }
413 
414 static int hns_roce_set_kernel_sq_size(struct hns_roce_dev *hr_dev,
415 				       struct ib_qp_cap *cap,
416 				       struct hns_roce_qp *hr_qp)
417 {
418 	struct device *dev = hr_dev->dev;
419 	u32 page_size;
420 	u32 max_cnt;
421 	int size;
422 
423 	if (cap->max_send_wr  > hr_dev->caps.max_wqes  ||
424 	    cap->max_send_sge > hr_dev->caps.max_sq_sg ||
425 	    cap->max_inline_data > hr_dev->caps.max_sq_inline) {
426 		dev_err(dev, "SQ WR or sge or inline data error!\n");
427 		return -EINVAL;
428 	}
429 
430 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
431 	hr_qp->sq_max_wqes_per_wr = 1;
432 	hr_qp->sq_spare_wqes = 0;
433 
434 	if (hr_dev->caps.min_wqes)
435 		max_cnt = max(cap->max_send_wr, hr_dev->caps.min_wqes);
436 	else
437 		max_cnt = cap->max_send_wr;
438 
439 	hr_qp->sq.wqe_cnt = roundup_pow_of_two(max_cnt);
440 	if ((u32)hr_qp->sq.wqe_cnt > hr_dev->caps.max_wqes) {
441 		dev_err(dev, "while setting kernel sq size, sq.wqe_cnt too large\n");
442 		return -EINVAL;
443 	}
444 
445 	/* Get data_seg numbers */
446 	max_cnt = max(1U, cap->max_send_sge);
447 	if (hr_dev->caps.max_sq_sg <= 2)
448 		hr_qp->sq.max_gs = roundup_pow_of_two(max_cnt);
449 	else
450 		hr_qp->sq.max_gs = max_cnt;
451 
452 	if (hr_qp->sq.max_gs > 2) {
453 		hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
454 				     (hr_qp->sq.max_gs - 2));
455 		hr_qp->sge.sge_shift = 4;
456 	}
457 
458 	/* ud sqwqe's sge use extend sge */
459 	if (hr_dev->caps.max_sq_sg > 2 && hr_qp->ibqp.qp_type == IB_QPT_GSI) {
460 		hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
461 				     hr_qp->sq.max_gs);
462 		hr_qp->sge.sge_shift = 4;
463 	}
464 
465 	/* Get buf size, SQ and RQ are aligned to PAGE_SIZE */
466 	page_size = 1 << (hr_dev->caps.mtt_buf_pg_sz + PAGE_SHIFT);
467 	hr_qp->sq.offset = 0;
468 	size = HNS_ROCE_ALOGN_UP(hr_qp->sq.wqe_cnt << hr_qp->sq.wqe_shift,
469 				 page_size);
470 
471 	if (hr_dev->caps.max_sq_sg > 2 && hr_qp->sge.sge_cnt) {
472 		hr_qp->sge.offset = size;
473 		size += HNS_ROCE_ALOGN_UP(hr_qp->sge.sge_cnt <<
474 					  hr_qp->sge.sge_shift, page_size);
475 	}
476 
477 	hr_qp->rq.offset = size;
478 	size += HNS_ROCE_ALOGN_UP((hr_qp->rq.wqe_cnt << hr_qp->rq.wqe_shift),
479 				  page_size);
480 	hr_qp->buff_size = size;
481 
482 	/* Get wr and sge number which send */
483 	cap->max_send_wr = hr_qp->sq.max_post = hr_qp->sq.wqe_cnt;
484 	cap->max_send_sge = hr_qp->sq.max_gs;
485 
486 	/* We don't support inline sends for kernel QPs (yet) */
487 	cap->max_inline_data = 0;
488 
489 	return 0;
490 }
491 
492 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
493 {
494 	if (attr->qp_type == IB_QPT_XRC_INI ||
495 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq)
496 		return 0;
497 
498 	return 1;
499 }
500 
501 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
502 				     struct ib_pd *ib_pd,
503 				     struct ib_qp_init_attr *init_attr,
504 				     struct ib_udata *udata, unsigned long sqpn,
505 				     struct hns_roce_qp *hr_qp)
506 {
507 	struct device *dev = hr_dev->dev;
508 	struct hns_roce_ib_create_qp ucmd;
509 	struct hns_roce_ib_create_qp_resp resp = {};
510 	unsigned long qpn = 0;
511 	int ret = 0;
512 	u32 page_shift;
513 	u32 npages;
514 	int i;
515 
516 	mutex_init(&hr_qp->mutex);
517 	spin_lock_init(&hr_qp->sq.lock);
518 	spin_lock_init(&hr_qp->rq.lock);
519 
520 	hr_qp->state = IB_QPS_RESET;
521 
522 	hr_qp->ibqp.qp_type = init_attr->qp_type;
523 
524 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
525 		hr_qp->sq_signal_bits = cpu_to_le32(IB_SIGNAL_ALL_WR);
526 	else
527 		hr_qp->sq_signal_bits = cpu_to_le32(IB_SIGNAL_REQ_WR);
528 
529 	ret = hns_roce_set_rq_size(hr_dev, &init_attr->cap, !!ib_pd->uobject,
530 				   !!init_attr->srq, hr_qp);
531 	if (ret) {
532 		dev_err(dev, "hns_roce_set_rq_size failed\n");
533 		goto err_out;
534 	}
535 
536 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) {
537 		/* allocate recv inline buf */
538 		hr_qp->rq_inl_buf.wqe_list = kcalloc(hr_qp->rq.wqe_cnt,
539 					       sizeof(struct hns_roce_rinl_wqe),
540 					       GFP_KERNEL);
541 		if (!hr_qp->rq_inl_buf.wqe_list) {
542 			ret = -ENOMEM;
543 			goto err_out;
544 		}
545 
546 		hr_qp->rq_inl_buf.wqe_cnt = hr_qp->rq.wqe_cnt;
547 
548 		/* Firstly, allocate a list of sge space buffer */
549 		hr_qp->rq_inl_buf.wqe_list[0].sg_list =
550 					kcalloc(hr_qp->rq_inl_buf.wqe_cnt,
551 					       init_attr->cap.max_recv_sge *
552 					       sizeof(struct hns_roce_rinl_sge),
553 					       GFP_KERNEL);
554 		if (!hr_qp->rq_inl_buf.wqe_list[0].sg_list) {
555 			ret = -ENOMEM;
556 			goto err_wqe_list;
557 		}
558 
559 		for (i = 1; i < hr_qp->rq_inl_buf.wqe_cnt; i++)
560 			/* Secondly, reallocate the buffer */
561 			hr_qp->rq_inl_buf.wqe_list[i].sg_list =
562 				&hr_qp->rq_inl_buf.wqe_list[0].sg_list[i *
563 				init_attr->cap.max_recv_sge];
564 	}
565 
566 	if (ib_pd->uobject) {
567 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
568 			dev_err(dev, "ib_copy_from_udata error for create qp\n");
569 			ret = -EFAULT;
570 			goto err_rq_sge_list;
571 		}
572 
573 		ret = hns_roce_set_user_sq_size(hr_dev, &init_attr->cap, hr_qp,
574 						&ucmd);
575 		if (ret) {
576 			dev_err(dev, "hns_roce_set_user_sq_size error for create qp\n");
577 			goto err_rq_sge_list;
578 		}
579 
580 		hr_qp->umem = ib_umem_get(ib_pd->uobject->context,
581 					  ucmd.buf_addr, hr_qp->buff_size, 0,
582 					  0);
583 		if (IS_ERR(hr_qp->umem)) {
584 			dev_err(dev, "ib_umem_get error for create qp\n");
585 			ret = PTR_ERR(hr_qp->umem);
586 			goto err_rq_sge_list;
587 		}
588 
589 		hr_qp->mtt.mtt_type = MTT_TYPE_WQE;
590 		if (hr_dev->caps.mtt_buf_pg_sz) {
591 			npages = (ib_umem_page_count(hr_qp->umem) +
592 				  (1 << hr_dev->caps.mtt_buf_pg_sz) - 1) /
593 				  (1 << hr_dev->caps.mtt_buf_pg_sz);
594 			page_shift = PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
595 			ret = hns_roce_mtt_init(hr_dev, npages,
596 				    page_shift,
597 				    &hr_qp->mtt);
598 		} else {
599 			ret = hns_roce_mtt_init(hr_dev,
600 				    ib_umem_page_count(hr_qp->umem),
601 				    hr_qp->umem->page_shift,
602 				    &hr_qp->mtt);
603 		}
604 		if (ret) {
605 			dev_err(dev, "hns_roce_mtt_init error for create qp\n");
606 			goto err_buf;
607 		}
608 
609 		ret = hns_roce_ib_umem_write_mtt(hr_dev, &hr_qp->mtt,
610 						 hr_qp->umem);
611 		if (ret) {
612 			dev_err(dev, "hns_roce_ib_umem_write_mtt error for create qp\n");
613 			goto err_mtt;
614 		}
615 
616 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
617 		    (udata->outlen >= sizeof(resp)) &&
618 		    hns_roce_qp_has_rq(init_attr)) {
619 			ret = hns_roce_db_map_user(
620 					to_hr_ucontext(ib_pd->uobject->context),
621 					ucmd.db_addr, &hr_qp->rdb);
622 			if (ret) {
623 				dev_err(dev, "rp record doorbell map failed!\n");
624 				goto err_mtt;
625 			}
626 		}
627 	} else {
628 		if (init_attr->create_flags &
629 		    IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
630 			dev_err(dev, "init_attr->create_flags error!\n");
631 			ret = -EINVAL;
632 			goto err_rq_sge_list;
633 		}
634 
635 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO) {
636 			dev_err(dev, "init_attr->create_flags error!\n");
637 			ret = -EINVAL;
638 			goto err_rq_sge_list;
639 		}
640 
641 		/* Set SQ size */
642 		ret = hns_roce_set_kernel_sq_size(hr_dev, &init_attr->cap,
643 						  hr_qp);
644 		if (ret) {
645 			dev_err(dev, "hns_roce_set_kernel_sq_size error!\n");
646 			goto err_rq_sge_list;
647 		}
648 
649 		/* QP doorbell register address */
650 		hr_qp->sq.db_reg_l = hr_dev->reg_base + hr_dev->sdb_offset +
651 				     DB_REG_OFFSET * hr_dev->priv_uar.index;
652 		hr_qp->rq.db_reg_l = hr_dev->reg_base + hr_dev->odb_offset +
653 				     DB_REG_OFFSET * hr_dev->priv_uar.index;
654 
655 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
656 		    hns_roce_qp_has_rq(init_attr)) {
657 			ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
658 			if (ret) {
659 				dev_err(dev, "rq record doorbell alloc failed!\n");
660 				goto err_rq_sge_list;
661 			}
662 			*hr_qp->rdb.db_record = 0;
663 		}
664 
665 		/* Allocate QP buf */
666 		page_shift = PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
667 		if (hns_roce_buf_alloc(hr_dev, hr_qp->buff_size,
668 				       (1 << page_shift) * 2,
669 				       &hr_qp->hr_buf, page_shift)) {
670 			dev_err(dev, "hns_roce_buf_alloc error!\n");
671 			ret = -ENOMEM;
672 			goto err_db;
673 		}
674 
675 		hr_qp->mtt.mtt_type = MTT_TYPE_WQE;
676 		/* Write MTT */
677 		ret = hns_roce_mtt_init(hr_dev, hr_qp->hr_buf.npages,
678 					hr_qp->hr_buf.page_shift, &hr_qp->mtt);
679 		if (ret) {
680 			dev_err(dev, "hns_roce_mtt_init error for kernel create qp\n");
681 			goto err_buf;
682 		}
683 
684 		ret = hns_roce_buf_write_mtt(hr_dev, &hr_qp->mtt,
685 					     &hr_qp->hr_buf);
686 		if (ret) {
687 			dev_err(dev, "hns_roce_buf_write_mtt error for kernel create qp\n");
688 			goto err_mtt;
689 		}
690 
691 		hr_qp->sq.wrid = kmalloc_array(hr_qp->sq.wqe_cnt, sizeof(u64),
692 					       GFP_KERNEL);
693 		hr_qp->rq.wrid = kmalloc_array(hr_qp->rq.wqe_cnt, sizeof(u64),
694 					       GFP_KERNEL);
695 		if (!hr_qp->sq.wrid || !hr_qp->rq.wrid) {
696 			ret = -ENOMEM;
697 			goto err_wrid;
698 		}
699 	}
700 
701 	if (sqpn) {
702 		qpn = sqpn;
703 	} else {
704 		/* Get QPN */
705 		ret = hns_roce_reserve_range_qp(hr_dev, 1, 1, &qpn);
706 		if (ret) {
707 			dev_err(dev, "hns_roce_reserve_range_qp alloc qpn error\n");
708 			goto err_wrid;
709 		}
710 	}
711 
712 	if (init_attr->qp_type == IB_QPT_GSI &&
713 	    hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
714 		/* In v1 engine, GSI QP context in RoCE engine's register */
715 		ret = hns_roce_gsi_qp_alloc(hr_dev, qpn, hr_qp);
716 		if (ret) {
717 			dev_err(dev, "hns_roce_qp_alloc failed!\n");
718 			goto err_qpn;
719 		}
720 	} else {
721 		ret = hns_roce_qp_alloc(hr_dev, qpn, hr_qp);
722 		if (ret) {
723 			dev_err(dev, "hns_roce_qp_alloc failed!\n");
724 			goto err_qpn;
725 		}
726 	}
727 
728 	if (sqpn)
729 		hr_qp->doorbell_qpn = 1;
730 	else
731 		hr_qp->doorbell_qpn = cpu_to_le64(hr_qp->qpn);
732 
733 	if (ib_pd->uobject && (udata->outlen >= sizeof(resp)) &&
734 		(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB)) {
735 
736 		/* indicate kernel supports record db */
737 		resp.cap_flags |= HNS_ROCE_SUPPORT_RQ_RECORD_DB;
738 		ret = ib_copy_to_udata(udata, &resp, sizeof(resp));
739 		if (ret)
740 			goto err_qp;
741 
742 		hr_qp->rdb_en = 1;
743 	}
744 	hr_qp->event = hns_roce_ib_qp_event;
745 
746 	return 0;
747 
748 err_qp:
749 	if (init_attr->qp_type == IB_QPT_GSI &&
750 		hr_dev->hw_rev == HNS_ROCE_HW_VER1)
751 		hns_roce_qp_remove(hr_dev, hr_qp);
752 	else
753 		hns_roce_qp_free(hr_dev, hr_qp);
754 
755 err_qpn:
756 	if (!sqpn)
757 		hns_roce_release_range_qp(hr_dev, qpn, 1);
758 
759 err_wrid:
760 	if (ib_pd->uobject) {
761 		if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
762 		    (udata->outlen >= sizeof(resp)) &&
763 		    hns_roce_qp_has_rq(init_attr))
764 			hns_roce_db_unmap_user(
765 					to_hr_ucontext(ib_pd->uobject->context),
766 					&hr_qp->rdb);
767 	} else {
768 		kfree(hr_qp->sq.wrid);
769 		kfree(hr_qp->rq.wrid);
770 	}
771 
772 err_mtt:
773 	hns_roce_mtt_cleanup(hr_dev, &hr_qp->mtt);
774 
775 err_buf:
776 	if (ib_pd->uobject)
777 		ib_umem_release(hr_qp->umem);
778 	else
779 		hns_roce_buf_free(hr_dev, hr_qp->buff_size, &hr_qp->hr_buf);
780 
781 err_db:
782 	if (!ib_pd->uobject && hns_roce_qp_has_rq(init_attr) &&
783 	    (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB))
784 		hns_roce_free_db(hr_dev, &hr_qp->rdb);
785 
786 err_rq_sge_list:
787 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE)
788 		kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
789 
790 err_wqe_list:
791 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE)
792 		kfree(hr_qp->rq_inl_buf.wqe_list);
793 
794 err_out:
795 	return ret;
796 }
797 
798 struct ib_qp *hns_roce_create_qp(struct ib_pd *pd,
799 				 struct ib_qp_init_attr *init_attr,
800 				 struct ib_udata *udata)
801 {
802 	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
803 	struct device *dev = hr_dev->dev;
804 	struct hns_roce_sqp *hr_sqp;
805 	struct hns_roce_qp *hr_qp;
806 	int ret;
807 
808 	switch (init_attr->qp_type) {
809 	case IB_QPT_RC: {
810 		hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
811 		if (!hr_qp)
812 			return ERR_PTR(-ENOMEM);
813 
814 		ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, 0,
815 						hr_qp);
816 		if (ret) {
817 			dev_err(dev, "Create RC QP failed\n");
818 			kfree(hr_qp);
819 			return ERR_PTR(ret);
820 		}
821 
822 		hr_qp->ibqp.qp_num = hr_qp->qpn;
823 
824 		break;
825 	}
826 	case IB_QPT_GSI: {
827 		/* Userspace is not allowed to create special QPs: */
828 		if (pd->uobject) {
829 			dev_err(dev, "not support usr space GSI\n");
830 			return ERR_PTR(-EINVAL);
831 		}
832 
833 		hr_sqp = kzalloc(sizeof(*hr_sqp), GFP_KERNEL);
834 		if (!hr_sqp)
835 			return ERR_PTR(-ENOMEM);
836 
837 		hr_qp = &hr_sqp->hr_qp;
838 		hr_qp->port = init_attr->port_num - 1;
839 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
840 
841 		/* when hw version is v1, the sqpn is allocated */
842 		if (hr_dev->caps.max_sq_sg <= 2)
843 			hr_qp->ibqp.qp_num = HNS_ROCE_MAX_PORTS +
844 					     hr_dev->iboe.phy_port[hr_qp->port];
845 		else
846 			hr_qp->ibqp.qp_num = 1;
847 
848 		ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata,
849 						hr_qp->ibqp.qp_num, hr_qp);
850 		if (ret) {
851 			dev_err(dev, "Create GSI QP failed!\n");
852 			kfree(hr_sqp);
853 			return ERR_PTR(ret);
854 		}
855 
856 		break;
857 	}
858 	default:{
859 		dev_err(dev, "not support QP type %d\n", init_attr->qp_type);
860 		return ERR_PTR(-EINVAL);
861 	}
862 	}
863 
864 	return &hr_qp->ibqp;
865 }
866 EXPORT_SYMBOL_GPL(hns_roce_create_qp);
867 
868 int to_hr_qp_type(int qp_type)
869 {
870 	int transport_type;
871 
872 	if (qp_type == IB_QPT_RC)
873 		transport_type = SERV_TYPE_RC;
874 	else if (qp_type == IB_QPT_UC)
875 		transport_type = SERV_TYPE_UC;
876 	else if (qp_type == IB_QPT_UD)
877 		transport_type = SERV_TYPE_UD;
878 	else if (qp_type == IB_QPT_GSI)
879 		transport_type = SERV_TYPE_UD;
880 	else
881 		transport_type = -1;
882 
883 	return transport_type;
884 }
885 EXPORT_SYMBOL_GPL(to_hr_qp_type);
886 
887 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
888 		       int attr_mask, struct ib_udata *udata)
889 {
890 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
891 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
892 	enum ib_qp_state cur_state, new_state;
893 	struct device *dev = hr_dev->dev;
894 	int ret = -EINVAL;
895 	int p;
896 	enum ib_mtu active_mtu;
897 
898 	mutex_lock(&hr_qp->mutex);
899 
900 	cur_state = attr_mask & IB_QP_CUR_STATE ?
901 		    attr->cur_qp_state : (enum ib_qp_state)hr_qp->state;
902 	new_state = attr_mask & IB_QP_STATE ?
903 		    attr->qp_state : cur_state;
904 
905 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask,
906 				IB_LINK_LAYER_ETHERNET)) {
907 		dev_err(dev, "ib_modify_qp_is_ok failed\n");
908 		goto out;
909 	}
910 
911 	if ((attr_mask & IB_QP_PORT) &&
912 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
913 		dev_err(dev, "attr port_num invalid.attr->port_num=%d\n",
914 			attr->port_num);
915 		goto out;
916 	}
917 
918 	if (attr_mask & IB_QP_PKEY_INDEX) {
919 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
920 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
921 			dev_err(dev, "attr pkey_index invalid.attr->pkey_index=%d\n",
922 				attr->pkey_index);
923 			goto out;
924 		}
925 	}
926 
927 	if (attr_mask & IB_QP_PATH_MTU) {
928 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
929 		active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
930 
931 		if ((hr_dev->caps.max_mtu == IB_MTU_4096 &&
932 		    attr->path_mtu > IB_MTU_4096) ||
933 		    (hr_dev->caps.max_mtu == IB_MTU_2048 &&
934 		    attr->path_mtu > IB_MTU_2048) ||
935 		    attr->path_mtu < IB_MTU_256 ||
936 		    attr->path_mtu > active_mtu) {
937 			dev_err(dev, "attr path_mtu(%d)invalid while modify qp",
938 				attr->path_mtu);
939 			goto out;
940 		}
941 	}
942 
943 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
944 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
945 		dev_err(dev, "attr max_rd_atomic invalid.attr->max_rd_atomic=%d\n",
946 			attr->max_rd_atomic);
947 		goto out;
948 	}
949 
950 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
951 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
952 		dev_err(dev, "attr max_dest_rd_atomic invalid.attr->max_dest_rd_atomic=%d\n",
953 			attr->max_dest_rd_atomic);
954 		goto out;
955 	}
956 
957 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
958 		ret = 0;
959 		goto out;
960 	}
961 
962 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
963 				    new_state);
964 
965 out:
966 	mutex_unlock(&hr_qp->mutex);
967 
968 	return ret;
969 }
970 
971 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
972 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
973 {
974 	if (send_cq == recv_cq) {
975 		spin_lock_irq(&send_cq->lock);
976 		__acquire(&recv_cq->lock);
977 	} else if (send_cq->cqn < recv_cq->cqn) {
978 		spin_lock_irq(&send_cq->lock);
979 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
980 	} else {
981 		spin_lock_irq(&recv_cq->lock);
982 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
983 	}
984 }
985 EXPORT_SYMBOL_GPL(hns_roce_lock_cqs);
986 
987 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
988 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
989 			 __releases(&recv_cq->lock)
990 {
991 	if (send_cq == recv_cq) {
992 		__release(&recv_cq->lock);
993 		spin_unlock_irq(&send_cq->lock);
994 	} else if (send_cq->cqn < recv_cq->cqn) {
995 		spin_unlock(&recv_cq->lock);
996 		spin_unlock_irq(&send_cq->lock);
997 	} else {
998 		spin_unlock(&send_cq->lock);
999 		spin_unlock_irq(&recv_cq->lock);
1000 	}
1001 }
1002 EXPORT_SYMBOL_GPL(hns_roce_unlock_cqs);
1003 
1004 static void *get_wqe(struct hns_roce_qp *hr_qp, int offset)
1005 {
1006 
1007 	return hns_roce_buf_offset(&hr_qp->hr_buf, offset);
1008 }
1009 
1010 void *get_recv_wqe(struct hns_roce_qp *hr_qp, int n)
1011 {
1012 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1013 }
1014 EXPORT_SYMBOL_GPL(get_recv_wqe);
1015 
1016 void *get_send_wqe(struct hns_roce_qp *hr_qp, int n)
1017 {
1018 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1019 }
1020 EXPORT_SYMBOL_GPL(get_send_wqe);
1021 
1022 void *get_send_extend_sge(struct hns_roce_qp *hr_qp, int n)
1023 {
1024 	return hns_roce_buf_offset(&hr_qp->hr_buf, hr_qp->sge.offset +
1025 					(n << hr_qp->sge.sge_shift));
1026 }
1027 EXPORT_SYMBOL_GPL(get_send_extend_sge);
1028 
1029 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, int nreq,
1030 			  struct ib_cq *ib_cq)
1031 {
1032 	struct hns_roce_cq *hr_cq;
1033 	u32 cur;
1034 
1035 	cur = hr_wq->head - hr_wq->tail;
1036 	if (likely(cur + nreq < hr_wq->max_post))
1037 		return false;
1038 
1039 	hr_cq = to_hr_cq(ib_cq);
1040 	spin_lock(&hr_cq->lock);
1041 	cur = hr_wq->head - hr_wq->tail;
1042 	spin_unlock(&hr_cq->lock);
1043 
1044 	return cur + nreq >= hr_wq->max_post;
1045 }
1046 EXPORT_SYMBOL_GPL(hns_roce_wq_overflow);
1047 
1048 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1049 {
1050 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1051 	int reserved_from_top = 0;
1052 	int ret;
1053 
1054 	spin_lock_init(&qp_table->lock);
1055 	INIT_RADIX_TREE(&hr_dev->qp_table_tree, GFP_ATOMIC);
1056 
1057 	/* A port include two SQP, six port total 12 */
1058 	ret = hns_roce_bitmap_init(&qp_table->bitmap, hr_dev->caps.num_qps,
1059 				   hr_dev->caps.num_qps - 1, SQP_NUM,
1060 				   reserved_from_top);
1061 	if (ret) {
1062 		dev_err(hr_dev->dev, "qp bitmap init failed!error=%d\n",
1063 			ret);
1064 		return ret;
1065 	}
1066 
1067 	return 0;
1068 }
1069 
1070 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1071 {
1072 	hns_roce_bitmap_cleanup(&hr_dev->qp_table.bitmap);
1073 }
1074