xref: /linux/drivers/infiniband/hw/hns/hns_roce_qp.c (revision faabed295cccc2aba2b67f2e7b309f2892d55004)
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 static void flush_work_handle(struct work_struct *work)
47 {
48 	struct hns_roce_work *flush_work = container_of(work,
49 					struct hns_roce_work, work);
50 	struct hns_roce_qp *hr_qp = container_of(flush_work,
51 					struct hns_roce_qp, flush_work);
52 	struct device *dev = flush_work->hr_dev->dev;
53 	struct ib_qp_attr attr;
54 	int attr_mask;
55 	int ret;
56 
57 	attr_mask = IB_QP_STATE;
58 	attr.qp_state = IB_QPS_ERR;
59 
60 	if (test_and_clear_bit(HNS_ROCE_FLUSH_FLAG, &hr_qp->flush_flag)) {
61 		ret = hns_roce_modify_qp(&hr_qp->ibqp, &attr, attr_mask, NULL);
62 		if (ret)
63 			dev_err(dev, "Modify QP to error state failed(%d) during CQE flush\n",
64 				ret);
65 	}
66 
67 	/*
68 	 * make sure we signal QP destroy leg that flush QP was completed
69 	 * so that it can safely proceed ahead now and destroy QP
70 	 */
71 	if (atomic_dec_and_test(&hr_qp->refcount))
72 		complete(&hr_qp->free);
73 }
74 
75 void init_flush_work(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
76 {
77 	struct hns_roce_work *flush_work = &hr_qp->flush_work;
78 
79 	flush_work->hr_dev = hr_dev;
80 	INIT_WORK(&flush_work->work, flush_work_handle);
81 	atomic_inc(&hr_qp->refcount);
82 	queue_work(hr_dev->irq_workq, &flush_work->work);
83 }
84 
85 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
86 {
87 	struct device *dev = hr_dev->dev;
88 	struct hns_roce_qp *qp;
89 
90 	xa_lock(&hr_dev->qp_table_xa);
91 	qp = __hns_roce_qp_lookup(hr_dev, qpn);
92 	if (qp)
93 		atomic_inc(&qp->refcount);
94 	xa_unlock(&hr_dev->qp_table_xa);
95 
96 	if (!qp) {
97 		dev_warn(dev, "Async event for bogus QP %08x\n", qpn);
98 		return;
99 	}
100 
101 	if (hr_dev->hw_rev != HNS_ROCE_HW_VER1 &&
102 	    (event_type == HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR ||
103 	     event_type == HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR ||
104 	     event_type == HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR)) {
105 		qp->state = IB_QPS_ERR;
106 		if (!test_and_set_bit(HNS_ROCE_FLUSH_FLAG, &qp->flush_flag))
107 			init_flush_work(hr_dev, qp);
108 	}
109 
110 	qp->event(qp, (enum hns_roce_event)event_type);
111 
112 	if (atomic_dec_and_test(&qp->refcount))
113 		complete(&qp->free);
114 }
115 
116 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
117 				 enum hns_roce_event type)
118 {
119 	struct ib_event event;
120 	struct ib_qp *ibqp = &hr_qp->ibqp;
121 
122 	if (ibqp->event_handler) {
123 		event.device = ibqp->device;
124 		event.element.qp = ibqp;
125 		switch (type) {
126 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
127 			event.event = IB_EVENT_PATH_MIG;
128 			break;
129 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
130 			event.event = IB_EVENT_COMM_EST;
131 			break;
132 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
133 			event.event = IB_EVENT_SQ_DRAINED;
134 			break;
135 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
136 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
137 			break;
138 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
139 			event.event = IB_EVENT_QP_FATAL;
140 			break;
141 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
142 			event.event = IB_EVENT_PATH_MIG_ERR;
143 			break;
144 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
145 			event.event = IB_EVENT_QP_REQ_ERR;
146 			break;
147 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
148 			event.event = IB_EVENT_QP_ACCESS_ERR;
149 			break;
150 		default:
151 			dev_dbg(ibqp->device->dev.parent, "roce_ib: Unexpected event type %d on QP %06lx\n",
152 				type, hr_qp->qpn);
153 			return;
154 		}
155 		ibqp->event_handler(&event, ibqp->qp_context);
156 	}
157 }
158 
159 static int alloc_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
160 {
161 	unsigned long num = 0;
162 	int ret;
163 
164 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI) {
165 		/* when hw version is v1, the sqpn is allocated */
166 		if (hr_dev->hw_rev == HNS_ROCE_HW_VER1)
167 			num = HNS_ROCE_MAX_PORTS +
168 			      hr_dev->iboe.phy_port[hr_qp->port];
169 		else
170 			num = 1;
171 
172 		hr_qp->doorbell_qpn = 1;
173 	} else {
174 		ret = hns_roce_bitmap_alloc_range(&hr_dev->qp_table.bitmap,
175 						  1, 1, &num);
176 		if (ret) {
177 			ibdev_err(&hr_dev->ib_dev, "Failed to alloc bitmap\n");
178 			return -ENOMEM;
179 		}
180 
181 		hr_qp->doorbell_qpn = (u32)num;
182 	}
183 
184 	hr_qp->qpn = num;
185 
186 	return 0;
187 }
188 
189 enum hns_roce_qp_state to_hns_roce_state(enum ib_qp_state state)
190 {
191 	switch (state) {
192 	case IB_QPS_RESET:
193 		return HNS_ROCE_QP_STATE_RST;
194 	case IB_QPS_INIT:
195 		return HNS_ROCE_QP_STATE_INIT;
196 	case IB_QPS_RTR:
197 		return HNS_ROCE_QP_STATE_RTR;
198 	case IB_QPS_RTS:
199 		return HNS_ROCE_QP_STATE_RTS;
200 	case IB_QPS_SQD:
201 		return HNS_ROCE_QP_STATE_SQD;
202 	case IB_QPS_ERR:
203 		return HNS_ROCE_QP_STATE_ERR;
204 	default:
205 		return HNS_ROCE_QP_NUM_STATE;
206 	}
207 }
208 
209 static void add_qp_to_list(struct hns_roce_dev *hr_dev,
210 			   struct hns_roce_qp *hr_qp,
211 			   struct ib_cq *send_cq, struct ib_cq *recv_cq)
212 {
213 	struct hns_roce_cq *hr_send_cq, *hr_recv_cq;
214 	unsigned long flags;
215 
216 	hr_send_cq = send_cq ? to_hr_cq(send_cq) : NULL;
217 	hr_recv_cq = recv_cq ? to_hr_cq(recv_cq) : NULL;
218 
219 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
220 	hns_roce_lock_cqs(hr_send_cq, hr_recv_cq);
221 
222 	list_add_tail(&hr_qp->node, &hr_dev->qp_list);
223 	if (hr_send_cq)
224 		list_add_tail(&hr_qp->sq_node, &hr_send_cq->sq_list);
225 	if (hr_recv_cq)
226 		list_add_tail(&hr_qp->rq_node, &hr_recv_cq->rq_list);
227 
228 	hns_roce_unlock_cqs(hr_send_cq, hr_recv_cq);
229 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
230 }
231 
232 static int hns_roce_qp_store(struct hns_roce_dev *hr_dev,
233 			     struct hns_roce_qp *hr_qp,
234 			     struct ib_qp_init_attr *init_attr)
235 {
236 	struct xarray *xa = &hr_dev->qp_table_xa;
237 	int ret;
238 
239 	if (!hr_qp->qpn)
240 		return -EINVAL;
241 
242 	ret = xa_err(xa_store_irq(xa, hr_qp->qpn, hr_qp, GFP_KERNEL));
243 	if (ret)
244 		dev_err(hr_dev->dev, "Failed to xa store for QPC\n");
245 	else
246 		/* add QP to device's QP list for softwc */
247 		add_qp_to_list(hr_dev, hr_qp, init_attr->send_cq,
248 			       init_attr->recv_cq);
249 
250 	return ret;
251 }
252 
253 static int alloc_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
254 {
255 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
256 	struct device *dev = hr_dev->dev;
257 	int ret;
258 
259 	if (!hr_qp->qpn)
260 		return -EINVAL;
261 
262 	/* In v1 engine, GSI QP context is saved in the RoCE hw's register */
263 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI &&
264 	    hr_dev->hw_rev == HNS_ROCE_HW_VER1)
265 		return 0;
266 
267 	/* Alloc memory for QPC */
268 	ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
269 	if (ret) {
270 		dev_err(dev, "Failed to get QPC table\n");
271 		goto err_out;
272 	}
273 
274 	/* Alloc memory for IRRL */
275 	ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
276 	if (ret) {
277 		dev_err(dev, "Failed to get IRRL table\n");
278 		goto err_put_qp;
279 	}
280 
281 	if (hr_dev->caps.trrl_entry_sz) {
282 		/* Alloc memory for TRRL */
283 		ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
284 					 hr_qp->qpn);
285 		if (ret) {
286 			dev_err(dev, "Failed to get TRRL table\n");
287 			goto err_put_irrl;
288 		}
289 	}
290 
291 	if (hr_dev->caps.sccc_entry_sz) {
292 		/* Alloc memory for SCC CTX */
293 		ret = hns_roce_table_get(hr_dev, &qp_table->sccc_table,
294 					 hr_qp->qpn);
295 		if (ret) {
296 			dev_err(dev, "Failed to get SCC CTX table\n");
297 			goto err_put_trrl;
298 		}
299 	}
300 
301 	return 0;
302 
303 err_put_trrl:
304 	if (hr_dev->caps.trrl_entry_sz)
305 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
306 
307 err_put_irrl:
308 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
309 
310 err_put_qp:
311 	hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
312 
313 err_out:
314 	return ret;
315 }
316 
317 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
318 {
319 	struct xarray *xa = &hr_dev->qp_table_xa;
320 	unsigned long flags;
321 
322 	list_del(&hr_qp->node);
323 	list_del(&hr_qp->sq_node);
324 	list_del(&hr_qp->rq_node);
325 
326 	xa_lock_irqsave(xa, flags);
327 	__xa_erase(xa, hr_qp->qpn & (hr_dev->caps.num_qps - 1));
328 	xa_unlock_irqrestore(xa, flags);
329 }
330 
331 static void free_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
332 {
333 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
334 
335 	/* In v1 engine, GSI QP context is saved in the RoCE hw's register */
336 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI &&
337 	    hr_dev->hw_rev == HNS_ROCE_HW_VER1)
338 		return;
339 
340 	if (hr_dev->caps.trrl_entry_sz)
341 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
342 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
343 }
344 
345 static void free_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
346 {
347 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
348 
349 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI)
350 		return;
351 
352 	if (hr_qp->qpn < hr_dev->caps.reserved_qps)
353 		return;
354 
355 	hns_roce_bitmap_free_range(&qp_table->bitmap, hr_qp->qpn, 1, BITMAP_RR);
356 }
357 
358 static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
359 		       struct hns_roce_qp *hr_qp, int has_rq)
360 {
361 	u32 cnt;
362 
363 	/* If srq exist, set zero for relative number of rq */
364 	if (!has_rq) {
365 		hr_qp->rq.wqe_cnt = 0;
366 		hr_qp->rq.max_gs = 0;
367 		hr_qp->rq_inl_buf.wqe_cnt = 0;
368 		cap->max_recv_wr = 0;
369 		cap->max_recv_sge = 0;
370 
371 		return 0;
372 	}
373 
374 	/* Check the validity of QP support capacity */
375 	if (!cap->max_recv_wr || cap->max_recv_wr > hr_dev->caps.max_wqes ||
376 	    cap->max_recv_sge > hr_dev->caps.max_rq_sg) {
377 		ibdev_err(&hr_dev->ib_dev, "RQ config error, depth=%u, sge=%d\n",
378 			  cap->max_recv_wr, cap->max_recv_sge);
379 		return -EINVAL;
380 	}
381 
382 	cnt = roundup_pow_of_two(max(cap->max_recv_wr, hr_dev->caps.min_wqes));
383 	if (cnt > hr_dev->caps.max_wqes) {
384 		ibdev_err(&hr_dev->ib_dev, "rq depth %u too large\n",
385 			  cap->max_recv_wr);
386 		return -EINVAL;
387 	}
388 
389 	hr_qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge) +
390 					      HNS_ROCE_RESERVED_SGE);
391 
392 	if (hr_dev->caps.max_rq_sg <= HNS_ROCE_SGE_IN_WQE)
393 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz);
394 	else
395 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz *
396 					    hr_qp->rq.max_gs);
397 
398 	hr_qp->rq.wqe_cnt = cnt;
399 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE)
400 		hr_qp->rq_inl_buf.wqe_cnt = cnt;
401 	else
402 		hr_qp->rq_inl_buf.wqe_cnt = 0;
403 
404 	cap->max_recv_wr = cnt;
405 	cap->max_recv_sge = hr_qp->rq.max_gs - HNS_ROCE_RESERVED_SGE;
406 
407 	return 0;
408 }
409 
410 static int set_extend_sge_param(struct hns_roce_dev *hr_dev, u32 sq_wqe_cnt,
411 				struct hns_roce_qp *hr_qp,
412 				struct ib_qp_cap *cap)
413 {
414 	struct ib_device *ibdev = &hr_dev->ib_dev;
415 	u32 cnt;
416 
417 	cnt = max(1U, cap->max_send_sge);
418 	if (hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
419 		hr_qp->sq.max_gs = roundup_pow_of_two(cnt);
420 		hr_qp->sge.sge_cnt = 0;
421 
422 		return 0;
423 	}
424 
425 	hr_qp->sq.max_gs = cnt;
426 
427 	/* UD sqwqe's sge use extend sge */
428 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI ||
429 	    hr_qp->ibqp.qp_type == IB_QPT_UD) {
430 		cnt = roundup_pow_of_two(sq_wqe_cnt * hr_qp->sq.max_gs);
431 	} else if (hr_qp->sq.max_gs > HNS_ROCE_SGE_IN_WQE) {
432 		cnt = roundup_pow_of_two(sq_wqe_cnt *
433 				     (hr_qp->sq.max_gs - HNS_ROCE_SGE_IN_WQE));
434 
435 		if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08_A) {
436 			if (cnt > hr_dev->caps.max_extend_sg) {
437 				ibdev_err(ibdev,
438 					  "failed to check exSGE num, exSGE num = %d.\n",
439 					  cnt);
440 				return -EINVAL;
441 			}
442 		}
443 	} else {
444 		cnt = 0;
445 	}
446 
447 	hr_qp->sge.sge_shift = HNS_ROCE_SGE_SHIFT;
448 	hr_qp->sge.sge_cnt = cnt;
449 
450 	return 0;
451 }
452 
453 static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
454 					struct ib_qp_cap *cap,
455 					struct hns_roce_ib_create_qp *ucmd)
456 {
457 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
458 	u8 max_sq_stride = ilog2(roundup_sq_stride);
459 
460 	/* Sanity check SQ size before proceeding */
461 	if (ucmd->log_sq_stride > max_sq_stride ||
462 	    ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
463 		ibdev_err(&hr_dev->ib_dev, "Failed to check SQ stride size\n");
464 		return -EINVAL;
465 	}
466 
467 	if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
468 		ibdev_err(&hr_dev->ib_dev, "Failed to check SQ SGE size %d\n",
469 			  cap->max_send_sge);
470 		return -EINVAL;
471 	}
472 
473 	return 0;
474 }
475 
476 static int set_user_sq_size(struct hns_roce_dev *hr_dev,
477 			    struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp,
478 			    struct hns_roce_ib_create_qp *ucmd)
479 {
480 	struct ib_device *ibdev = &hr_dev->ib_dev;
481 	u32 cnt = 0;
482 	int ret;
483 
484 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
485 	    cnt > hr_dev->caps.max_wqes)
486 		return -EINVAL;
487 
488 	ret = check_sq_size_with_integrity(hr_dev, cap, ucmd);
489 	if (ret) {
490 		ibdev_err(ibdev, "failed to check user SQ size, ret = %d.\n",
491 			  ret);
492 		return ret;
493 	}
494 
495 	ret = set_extend_sge_param(hr_dev, cnt, hr_qp, cap);
496 	if (ret)
497 		return ret;
498 
499 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
500 	hr_qp->sq.wqe_cnt = cnt;
501 
502 	return 0;
503 }
504 
505 static int set_wqe_buf_attr(struct hns_roce_dev *hr_dev,
506 			    struct hns_roce_qp *hr_qp,
507 			    struct hns_roce_buf_attr *buf_attr)
508 {
509 	int buf_size;
510 	int idx = 0;
511 
512 	hr_qp->buff_size = 0;
513 
514 	/* SQ WQE */
515 	hr_qp->sq.offset = 0;
516 	buf_size = to_hr_hem_entries_size(hr_qp->sq.wqe_cnt,
517 					  hr_qp->sq.wqe_shift);
518 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
519 		buf_attr->region[idx].size = buf_size;
520 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sq_hop_num;
521 		idx++;
522 		hr_qp->buff_size += buf_size;
523 	}
524 
525 	/* extend SGE WQE in SQ */
526 	hr_qp->sge.offset = hr_qp->buff_size;
527 	buf_size = to_hr_hem_entries_size(hr_qp->sge.sge_cnt,
528 					  hr_qp->sge.sge_shift);
529 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
530 		buf_attr->region[idx].size = buf_size;
531 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sge_hop_num;
532 		idx++;
533 		hr_qp->buff_size += buf_size;
534 	}
535 
536 	/* RQ WQE */
537 	hr_qp->rq.offset = hr_qp->buff_size;
538 	buf_size = to_hr_hem_entries_size(hr_qp->rq.wqe_cnt,
539 					  hr_qp->rq.wqe_shift);
540 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
541 		buf_attr->region[idx].size = buf_size;
542 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_rq_hop_num;
543 		idx++;
544 		hr_qp->buff_size += buf_size;
545 	}
546 
547 	if (hr_qp->buff_size < 1)
548 		return -EINVAL;
549 
550 	buf_attr->page_shift = HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
551 	buf_attr->fixed_page = true;
552 	buf_attr->region_count = idx;
553 
554 	return 0;
555 }
556 
557 static int set_kernel_sq_size(struct hns_roce_dev *hr_dev,
558 			      struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp)
559 {
560 	struct ib_device *ibdev = &hr_dev->ib_dev;
561 	u32 cnt;
562 	int ret;
563 
564 	if (!cap->max_send_wr || cap->max_send_wr > hr_dev->caps.max_wqes ||
565 	    cap->max_send_sge > hr_dev->caps.max_sq_sg ||
566 	    cap->max_inline_data > hr_dev->caps.max_sq_inline) {
567 		ibdev_err(ibdev,
568 			  "failed to check SQ WR, SGE or inline num, ret = %d.\n",
569 			  -EINVAL);
570 		return -EINVAL;
571 	}
572 
573 	cnt = roundup_pow_of_two(max(cap->max_send_wr, hr_dev->caps.min_wqes));
574 	if (cnt > hr_dev->caps.max_wqes) {
575 		ibdev_err(ibdev, "failed to check WQE num, WQE num = %d.\n",
576 			  cnt);
577 		return -EINVAL;
578 	}
579 
580 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
581 	hr_qp->sq.wqe_cnt = cnt;
582 
583 	ret = set_extend_sge_param(hr_dev, cnt, hr_qp, cap);
584 	if (ret)
585 		return ret;
586 
587 	/* sync the parameters of kernel QP to user's configuration */
588 	cap->max_send_wr = cnt;
589 	cap->max_send_sge = hr_qp->sq.max_gs;
590 
591 	/* We don't support inline sends for kernel QPs (yet) */
592 	cap->max_inline_data = 0;
593 
594 	return 0;
595 }
596 
597 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
598 {
599 	if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
600 		return 0;
601 
602 	return 1;
603 }
604 
605 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
606 {
607 	if (attr->qp_type == IB_QPT_XRC_INI ||
608 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
609 	    !attr->cap.max_recv_wr)
610 		return 0;
611 
612 	return 1;
613 }
614 
615 static int alloc_rq_inline_buf(struct hns_roce_qp *hr_qp,
616 			       struct ib_qp_init_attr *init_attr)
617 {
618 	u32 max_recv_sge = init_attr->cap.max_recv_sge;
619 	u32 wqe_cnt = hr_qp->rq_inl_buf.wqe_cnt;
620 	struct hns_roce_rinl_wqe *wqe_list;
621 	int i;
622 
623 	/* allocate recv inline buf */
624 	wqe_list = kcalloc(wqe_cnt, sizeof(struct hns_roce_rinl_wqe),
625 			   GFP_KERNEL);
626 
627 	if (!wqe_list)
628 		goto err;
629 
630 	/* Allocate a continuous buffer for all inline sge we need */
631 	wqe_list[0].sg_list = kcalloc(wqe_cnt, (max_recv_sge *
632 				      sizeof(struct hns_roce_rinl_sge)),
633 				      GFP_KERNEL);
634 	if (!wqe_list[0].sg_list)
635 		goto err_wqe_list;
636 
637 	/* Assign buffers of sg_list to each inline wqe */
638 	for (i = 1; i < wqe_cnt; i++)
639 		wqe_list[i].sg_list = &wqe_list[0].sg_list[i * max_recv_sge];
640 
641 	hr_qp->rq_inl_buf.wqe_list = wqe_list;
642 
643 	return 0;
644 
645 err_wqe_list:
646 	kfree(wqe_list);
647 
648 err:
649 	return -ENOMEM;
650 }
651 
652 static void free_rq_inline_buf(struct hns_roce_qp *hr_qp)
653 {
654 	if (hr_qp->rq_inl_buf.wqe_list)
655 		kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
656 	kfree(hr_qp->rq_inl_buf.wqe_list);
657 }
658 
659 static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
660 			struct ib_qp_init_attr *init_attr,
661 			struct ib_udata *udata, unsigned long addr)
662 {
663 	struct ib_device *ibdev = &hr_dev->ib_dev;
664 	struct hns_roce_buf_attr buf_attr = {};
665 	int ret;
666 
667 	if (!udata && hr_qp->rq_inl_buf.wqe_cnt) {
668 		ret = alloc_rq_inline_buf(hr_qp, init_attr);
669 		if (ret) {
670 			ibdev_err(ibdev,
671 				  "failed to alloc inline buf, ret = %d.\n",
672 				  ret);
673 			return ret;
674 		}
675 	} else {
676 		hr_qp->rq_inl_buf.wqe_list = NULL;
677 	}
678 
679 	ret = set_wqe_buf_attr(hr_dev, hr_qp, &buf_attr);
680 	if (ret) {
681 		ibdev_err(ibdev, "failed to split WQE buf, ret = %d.\n", ret);
682 		goto err_inline;
683 	}
684 	ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, &buf_attr,
685 				  HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
686 				  udata, addr);
687 	if (ret) {
688 		ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret);
689 		goto err_inline;
690 	}
691 
692 	return 0;
693 err_inline:
694 	free_rq_inline_buf(hr_qp);
695 
696 	return ret;
697 }
698 
699 static void free_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
700 {
701 	hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
702 	free_rq_inline_buf(hr_qp);
703 }
704 
705 static inline bool user_qp_has_sdb(struct hns_roce_dev *hr_dev,
706 				   struct ib_qp_init_attr *init_attr,
707 				   struct ib_udata *udata,
708 				   struct hns_roce_ib_create_qp_resp *resp,
709 				   struct hns_roce_ib_create_qp *ucmd)
710 {
711 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SQ_RECORD_DB) &&
712 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
713 		hns_roce_qp_has_sq(init_attr) &&
714 		udata->inlen >= offsetofend(typeof(*ucmd), sdb_addr));
715 }
716 
717 static inline bool user_qp_has_rdb(struct hns_roce_dev *hr_dev,
718 				   struct ib_qp_init_attr *init_attr,
719 				   struct ib_udata *udata,
720 				   struct hns_roce_ib_create_qp_resp *resp)
721 {
722 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
723 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
724 		hns_roce_qp_has_rq(init_attr));
725 }
726 
727 static inline bool kernel_qp_has_rdb(struct hns_roce_dev *hr_dev,
728 				     struct ib_qp_init_attr *init_attr)
729 {
730 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
731 		hns_roce_qp_has_rq(init_attr));
732 }
733 
734 static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
735 		       struct ib_qp_init_attr *init_attr,
736 		       struct ib_udata *udata,
737 		       struct hns_roce_ib_create_qp *ucmd,
738 		       struct hns_roce_ib_create_qp_resp *resp)
739 {
740 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
741 		udata, struct hns_roce_ucontext, ibucontext);
742 	struct ib_device *ibdev = &hr_dev->ib_dev;
743 	int ret;
744 
745 	if (udata) {
746 		if (user_qp_has_sdb(hr_dev, init_attr, udata, resp, ucmd)) {
747 			ret = hns_roce_db_map_user(uctx, udata, ucmd->sdb_addr,
748 						   &hr_qp->sdb);
749 			if (ret) {
750 				ibdev_err(ibdev,
751 					  "Failed to map user SQ doorbell\n");
752 				goto err_out;
753 			}
754 			hr_qp->en_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
755 			resp->cap_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
756 		}
757 
758 		if (user_qp_has_rdb(hr_dev, init_attr, udata, resp)) {
759 			ret = hns_roce_db_map_user(uctx, udata, ucmd->db_addr,
760 						   &hr_qp->rdb);
761 			if (ret) {
762 				ibdev_err(ibdev,
763 					  "Failed to map user RQ doorbell\n");
764 				goto err_sdb;
765 			}
766 			hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
767 			resp->cap_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
768 		}
769 	} else {
770 		/* QP doorbell register address */
771 		hr_qp->sq.db_reg_l = hr_dev->reg_base + hr_dev->sdb_offset +
772 				     DB_REG_OFFSET * hr_dev->priv_uar.index;
773 		hr_qp->rq.db_reg_l = hr_dev->reg_base + hr_dev->odb_offset +
774 				     DB_REG_OFFSET * hr_dev->priv_uar.index;
775 
776 		if (kernel_qp_has_rdb(hr_dev, init_attr)) {
777 			ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
778 			if (ret) {
779 				ibdev_err(ibdev,
780 					  "Failed to alloc kernel RQ doorbell\n");
781 				goto err_out;
782 			}
783 			*hr_qp->rdb.db_record = 0;
784 			hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
785 		}
786 	}
787 
788 	return 0;
789 err_sdb:
790 	if (udata && hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
791 		hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
792 err_out:
793 	return ret;
794 }
795 
796 static void free_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
797 		       struct ib_udata *udata)
798 {
799 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
800 		udata, struct hns_roce_ucontext, ibucontext);
801 
802 	if (udata) {
803 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
804 			hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
805 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
806 			hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
807 	} else {
808 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
809 			hns_roce_free_db(hr_dev, &hr_qp->rdb);
810 	}
811 }
812 
813 static int alloc_kernel_wrid(struct hns_roce_dev *hr_dev,
814 			     struct hns_roce_qp *hr_qp)
815 {
816 	struct ib_device *ibdev = &hr_dev->ib_dev;
817 	u64 *sq_wrid = NULL;
818 	u64 *rq_wrid = NULL;
819 	int ret;
820 
821 	sq_wrid = kcalloc(hr_qp->sq.wqe_cnt, sizeof(u64), GFP_KERNEL);
822 	if (ZERO_OR_NULL_PTR(sq_wrid)) {
823 		ibdev_err(ibdev, "Failed to alloc SQ wrid\n");
824 		return -ENOMEM;
825 	}
826 
827 	if (hr_qp->rq.wqe_cnt) {
828 		rq_wrid = kcalloc(hr_qp->rq.wqe_cnt, sizeof(u64), GFP_KERNEL);
829 		if (ZERO_OR_NULL_PTR(rq_wrid)) {
830 			ibdev_err(ibdev, "Failed to alloc RQ wrid\n");
831 			ret = -ENOMEM;
832 			goto err_sq;
833 		}
834 	}
835 
836 	hr_qp->sq.wrid = sq_wrid;
837 	hr_qp->rq.wrid = rq_wrid;
838 	return 0;
839 err_sq:
840 	kfree(sq_wrid);
841 
842 	return ret;
843 }
844 
845 static void free_kernel_wrid(struct hns_roce_qp *hr_qp)
846 {
847 	kfree(hr_qp->rq.wrid);
848 	kfree(hr_qp->sq.wrid);
849 }
850 
851 static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
852 			struct ib_qp_init_attr *init_attr,
853 			struct ib_udata *udata,
854 			struct hns_roce_ib_create_qp *ucmd)
855 {
856 	struct ib_device *ibdev = &hr_dev->ib_dev;
857 	int ret;
858 
859 	hr_qp->ibqp.qp_type = init_attr->qp_type;
860 
861 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
862 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
863 	else
864 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
865 
866 	ret = set_rq_size(hr_dev, &init_attr->cap, hr_qp,
867 			  hns_roce_qp_has_rq(init_attr));
868 	if (ret) {
869 		ibdev_err(ibdev, "failed to set user RQ size, ret = %d.\n",
870 			  ret);
871 		return ret;
872 	}
873 
874 	if (udata) {
875 		if (ib_copy_from_udata(ucmd, udata, sizeof(*ucmd))) {
876 			ibdev_err(ibdev, "Failed to copy QP ucmd\n");
877 			return -EFAULT;
878 		}
879 
880 		ret = set_user_sq_size(hr_dev, &init_attr->cap, hr_qp, ucmd);
881 		if (ret)
882 			ibdev_err(ibdev, "Failed to set user SQ size\n");
883 	} else {
884 		if (init_attr->create_flags &
885 		    IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
886 			ibdev_err(ibdev, "Failed to check multicast loopback\n");
887 			return -EINVAL;
888 		}
889 
890 		if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO) {
891 			ibdev_err(ibdev, "Failed to check ipoib ud lso\n");
892 			return -EINVAL;
893 		}
894 
895 		ret = set_kernel_sq_size(hr_dev, &init_attr->cap, hr_qp);
896 		if (ret)
897 			ibdev_err(ibdev, "Failed to set kernel SQ size\n");
898 	}
899 
900 	return ret;
901 }
902 
903 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
904 				     struct ib_pd *ib_pd,
905 				     struct ib_qp_init_attr *init_attr,
906 				     struct ib_udata *udata,
907 				     struct hns_roce_qp *hr_qp)
908 {
909 	struct hns_roce_ib_create_qp_resp resp = {};
910 	struct ib_device *ibdev = &hr_dev->ib_dev;
911 	struct hns_roce_ib_create_qp ucmd;
912 	int ret;
913 
914 	mutex_init(&hr_qp->mutex);
915 	spin_lock_init(&hr_qp->sq.lock);
916 	spin_lock_init(&hr_qp->rq.lock);
917 
918 	hr_qp->state = IB_QPS_RESET;
919 	hr_qp->flush_flag = 0;
920 
921 	ret = set_qp_param(hr_dev, hr_qp, init_attr, udata, &ucmd);
922 	if (ret) {
923 		ibdev_err(ibdev, "Failed to set QP param\n");
924 		return ret;
925 	}
926 
927 	if (!udata) {
928 		ret = alloc_kernel_wrid(hr_dev, hr_qp);
929 		if (ret) {
930 			ibdev_err(ibdev, "Failed to alloc wrid\n");
931 			return ret;
932 		}
933 	}
934 
935 	ret = alloc_qp_db(hr_dev, hr_qp, init_attr, udata, &ucmd, &resp);
936 	if (ret) {
937 		ibdev_err(ibdev, "Failed to alloc QP doorbell\n");
938 		goto err_wrid;
939 	}
940 
941 	ret = alloc_qp_buf(hr_dev, hr_qp, init_attr, udata, ucmd.buf_addr);
942 	if (ret) {
943 		ibdev_err(ibdev, "Failed to alloc QP buffer\n");
944 		goto err_db;
945 	}
946 
947 	ret = alloc_qpn(hr_dev, hr_qp);
948 	if (ret) {
949 		ibdev_err(ibdev, "Failed to alloc QPN\n");
950 		goto err_buf;
951 	}
952 
953 	ret = alloc_qpc(hr_dev, hr_qp);
954 	if (ret) {
955 		ibdev_err(ibdev, "Failed to alloc QP context\n");
956 		goto err_qpn;
957 	}
958 
959 	ret = hns_roce_qp_store(hr_dev, hr_qp, init_attr);
960 	if (ret) {
961 		ibdev_err(ibdev, "Failed to store QP\n");
962 		goto err_qpc;
963 	}
964 
965 	if (udata) {
966 		ret = ib_copy_to_udata(udata, &resp,
967 				       min(udata->outlen, sizeof(resp)));
968 		if (ret) {
969 			ibdev_err(ibdev, "copy qp resp failed!\n");
970 			goto err_store;
971 		}
972 	}
973 
974 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
975 		ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
976 		if (ret)
977 			goto err_store;
978 	}
979 
980 	hr_qp->ibqp.qp_num = hr_qp->qpn;
981 	hr_qp->event = hns_roce_ib_qp_event;
982 	atomic_set(&hr_qp->refcount, 1);
983 	init_completion(&hr_qp->free);
984 
985 	return 0;
986 
987 err_store:
988 	hns_roce_qp_remove(hr_dev, hr_qp);
989 err_qpc:
990 	free_qpc(hr_dev, hr_qp);
991 err_qpn:
992 	free_qpn(hr_dev, hr_qp);
993 err_buf:
994 	free_qp_buf(hr_dev, hr_qp);
995 err_db:
996 	free_qp_db(hr_dev, hr_qp, udata);
997 err_wrid:
998 	free_kernel_wrid(hr_qp);
999 	return ret;
1000 }
1001 
1002 void hns_roce_qp_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1003 			 struct ib_udata *udata)
1004 {
1005 	if (atomic_dec_and_test(&hr_qp->refcount))
1006 		complete(&hr_qp->free);
1007 	wait_for_completion(&hr_qp->free);
1008 
1009 	free_qpc(hr_dev, hr_qp);
1010 	free_qpn(hr_dev, hr_qp);
1011 	free_qp_buf(hr_dev, hr_qp);
1012 	free_kernel_wrid(hr_qp);
1013 	free_qp_db(hr_dev, hr_qp, udata);
1014 
1015 	kfree(hr_qp);
1016 }
1017 
1018 struct ib_qp *hns_roce_create_qp(struct ib_pd *pd,
1019 				 struct ib_qp_init_attr *init_attr,
1020 				 struct ib_udata *udata)
1021 {
1022 	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
1023 	struct ib_device *ibdev = &hr_dev->ib_dev;
1024 	struct hns_roce_qp *hr_qp;
1025 	int ret;
1026 
1027 	switch (init_attr->qp_type) {
1028 	case IB_QPT_RC: {
1029 		hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
1030 		if (!hr_qp)
1031 			return ERR_PTR(-ENOMEM);
1032 
1033 		ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata,
1034 						hr_qp);
1035 		if (ret) {
1036 			ibdev_err(ibdev, "Create QP 0x%06lx failed(%d)\n",
1037 				  hr_qp->qpn, ret);
1038 			kfree(hr_qp);
1039 			return ERR_PTR(ret);
1040 		}
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 		ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata,
1059 						hr_qp);
1060 		if (ret) {
1061 			ibdev_err(ibdev, "Create GSI QP failed!\n");
1062 			kfree(hr_qp);
1063 			return ERR_PTR(ret);
1064 		}
1065 
1066 		break;
1067 	}
1068 	default:{
1069 		ibdev_err(ibdev, "not support QP type %d\n",
1070 			  init_attr->qp_type);
1071 		return ERR_PTR(-EOPNOTSUPP);
1072 	}
1073 	}
1074 
1075 	return &hr_qp->ibqp;
1076 }
1077 
1078 int to_hr_qp_type(int qp_type)
1079 {
1080 	int transport_type;
1081 
1082 	if (qp_type == IB_QPT_RC)
1083 		transport_type = SERV_TYPE_RC;
1084 	else if (qp_type == IB_QPT_UC)
1085 		transport_type = SERV_TYPE_UC;
1086 	else if (qp_type == IB_QPT_UD)
1087 		transport_type = SERV_TYPE_UD;
1088 	else if (qp_type == IB_QPT_GSI)
1089 		transport_type = SERV_TYPE_UD;
1090 	else
1091 		transport_type = -1;
1092 
1093 	return transport_type;
1094 }
1095 
1096 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1097 			      struct hns_roce_qp *hr_qp,
1098 			      struct ib_qp_attr *attr, int attr_mask)
1099 {
1100 	enum ib_mtu active_mtu;
1101 	int p;
1102 
1103 	p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1104 	active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1105 
1106 	if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1107 	    attr->path_mtu > hr_dev->caps.max_mtu) ||
1108 	    attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1109 		ibdev_err(&hr_dev->ib_dev,
1110 			"attr path_mtu(%d)invalid while modify qp",
1111 			attr->path_mtu);
1112 		return -EINVAL;
1113 	}
1114 
1115 	return 0;
1116 }
1117 
1118 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1119 				  int attr_mask)
1120 {
1121 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1122 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1123 	int p;
1124 
1125 	if ((attr_mask & IB_QP_PORT) &&
1126 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1127 		ibdev_err(&hr_dev->ib_dev,
1128 			"attr port_num invalid.attr->port_num=%d\n",
1129 			attr->port_num);
1130 		return -EINVAL;
1131 	}
1132 
1133 	if (attr_mask & IB_QP_PKEY_INDEX) {
1134 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1135 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1136 			ibdev_err(&hr_dev->ib_dev,
1137 				"attr pkey_index invalid.attr->pkey_index=%d\n",
1138 				attr->pkey_index);
1139 			return -EINVAL;
1140 		}
1141 	}
1142 
1143 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1144 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1145 		ibdev_err(&hr_dev->ib_dev,
1146 			"attr max_rd_atomic invalid.attr->max_rd_atomic=%d\n",
1147 			attr->max_rd_atomic);
1148 		return -EINVAL;
1149 	}
1150 
1151 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1152 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1153 		ibdev_err(&hr_dev->ib_dev,
1154 			"attr max_dest_rd_atomic invalid.attr->max_dest_rd_atomic=%d\n",
1155 			attr->max_dest_rd_atomic);
1156 		return -EINVAL;
1157 	}
1158 
1159 	if (attr_mask & IB_QP_PATH_MTU)
1160 		return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1161 
1162 	return 0;
1163 }
1164 
1165 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1166 		       int attr_mask, struct ib_udata *udata)
1167 {
1168 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1169 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1170 	enum ib_qp_state cur_state, new_state;
1171 	int ret = -EINVAL;
1172 
1173 	mutex_lock(&hr_qp->mutex);
1174 
1175 	cur_state = attr_mask & IB_QP_CUR_STATE ?
1176 		    attr->cur_qp_state : (enum ib_qp_state)hr_qp->state;
1177 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1178 
1179 	if (ibqp->uobject &&
1180 	    (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1181 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB) {
1182 			hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1183 
1184 			if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
1185 				hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1186 		} else {
1187 			ibdev_warn(&hr_dev->ib_dev,
1188 				  "flush cqe is not supported in userspace!\n");
1189 			goto out;
1190 		}
1191 	}
1192 
1193 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1194 				attr_mask)) {
1195 		ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1196 		goto out;
1197 	}
1198 
1199 	ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1200 	if (ret)
1201 		goto out;
1202 
1203 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1204 		if (hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
1205 			ret = -EPERM;
1206 			ibdev_err(&hr_dev->ib_dev,
1207 				  "RST2RST state is not supported\n");
1208 		} else {
1209 			ret = 0;
1210 		}
1211 
1212 		goto out;
1213 	}
1214 
1215 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1216 				    new_state);
1217 
1218 out:
1219 	mutex_unlock(&hr_qp->mutex);
1220 
1221 	return ret;
1222 }
1223 
1224 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1225 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1226 {
1227 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1228 		__acquire(&send_cq->lock);
1229 		__acquire(&recv_cq->lock);
1230 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1231 		spin_lock_irq(&send_cq->lock);
1232 		__acquire(&recv_cq->lock);
1233 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1234 		spin_lock_irq(&recv_cq->lock);
1235 		__acquire(&send_cq->lock);
1236 	} else if (send_cq == recv_cq) {
1237 		spin_lock_irq(&send_cq->lock);
1238 		__acquire(&recv_cq->lock);
1239 	} else if (send_cq->cqn < recv_cq->cqn) {
1240 		spin_lock_irq(&send_cq->lock);
1241 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1242 	} else {
1243 		spin_lock_irq(&recv_cq->lock);
1244 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1245 	}
1246 }
1247 
1248 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1249 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1250 			 __releases(&recv_cq->lock)
1251 {
1252 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1253 		__release(&recv_cq->lock);
1254 		__release(&send_cq->lock);
1255 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1256 		__release(&recv_cq->lock);
1257 		spin_unlock(&send_cq->lock);
1258 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1259 		__release(&send_cq->lock);
1260 		spin_unlock(&recv_cq->lock);
1261 	} else if (send_cq == recv_cq) {
1262 		__release(&recv_cq->lock);
1263 		spin_unlock_irq(&send_cq->lock);
1264 	} else if (send_cq->cqn < recv_cq->cqn) {
1265 		spin_unlock(&recv_cq->lock);
1266 		spin_unlock_irq(&send_cq->lock);
1267 	} else {
1268 		spin_unlock(&send_cq->lock);
1269 		spin_unlock_irq(&recv_cq->lock);
1270 	}
1271 }
1272 
1273 static inline void *get_wqe(struct hns_roce_qp *hr_qp, int offset)
1274 {
1275 	return hns_roce_buf_offset(hr_qp->mtr.kmem, offset);
1276 }
1277 
1278 void *hns_roce_get_recv_wqe(struct hns_roce_qp *hr_qp, int n)
1279 {
1280 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1281 }
1282 
1283 void *hns_roce_get_send_wqe(struct hns_roce_qp *hr_qp, int n)
1284 {
1285 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1286 }
1287 
1288 void *hns_roce_get_extend_sge(struct hns_roce_qp *hr_qp, int n)
1289 {
1290 	return get_wqe(hr_qp, hr_qp->sge.offset + (n << hr_qp->sge.sge_shift));
1291 }
1292 
1293 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, int nreq,
1294 			  struct ib_cq *ib_cq)
1295 {
1296 	struct hns_roce_cq *hr_cq;
1297 	u32 cur;
1298 
1299 	cur = hr_wq->head - hr_wq->tail;
1300 	if (likely(cur + nreq < hr_wq->wqe_cnt))
1301 		return false;
1302 
1303 	hr_cq = to_hr_cq(ib_cq);
1304 	spin_lock(&hr_cq->lock);
1305 	cur = hr_wq->head - hr_wq->tail;
1306 	spin_unlock(&hr_cq->lock);
1307 
1308 	return cur + nreq >= hr_wq->wqe_cnt;
1309 }
1310 
1311 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1312 {
1313 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1314 	int reserved_from_top = 0;
1315 	int reserved_from_bot;
1316 	int ret;
1317 
1318 	mutex_init(&qp_table->scc_mutex);
1319 	xa_init(&hr_dev->qp_table_xa);
1320 
1321 	reserved_from_bot = hr_dev->caps.reserved_qps;
1322 
1323 	ret = hns_roce_bitmap_init(&qp_table->bitmap, hr_dev->caps.num_qps,
1324 				   hr_dev->caps.num_qps - 1, reserved_from_bot,
1325 				   reserved_from_top);
1326 	if (ret) {
1327 		dev_err(hr_dev->dev, "qp bitmap init failed!error=%d\n",
1328 			ret);
1329 		return ret;
1330 	}
1331 
1332 	return 0;
1333 }
1334 
1335 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1336 {
1337 	hns_roce_bitmap_cleanup(&hr_dev->qp_table.bitmap);
1338 }
1339