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 
43 static void flush_work_handle(struct work_struct *work)
44 {
45 	struct hns_roce_work *flush_work = container_of(work,
46 					struct hns_roce_work, work);
47 	struct hns_roce_qp *hr_qp = container_of(flush_work,
48 					struct hns_roce_qp, flush_work);
49 	struct device *dev = flush_work->hr_dev->dev;
50 	struct ib_qp_attr attr;
51 	int attr_mask;
52 	int ret;
53 
54 	attr_mask = IB_QP_STATE;
55 	attr.qp_state = IB_QPS_ERR;
56 
57 	if (test_and_clear_bit(HNS_ROCE_FLUSH_FLAG, &hr_qp->flush_flag)) {
58 		ret = hns_roce_modify_qp(&hr_qp->ibqp, &attr, attr_mask, NULL);
59 		if (ret)
60 			dev_err(dev, "Modify QP to error state failed(%d) during CQE flush\n",
61 				ret);
62 	}
63 
64 	/*
65 	 * make sure we signal QP destroy leg that flush QP was completed
66 	 * so that it can safely proceed ahead now and destroy QP
67 	 */
68 	if (atomic_dec_and_test(&hr_qp->refcount))
69 		complete(&hr_qp->free);
70 }
71 
72 void init_flush_work(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
73 {
74 	struct hns_roce_work *flush_work = &hr_qp->flush_work;
75 
76 	flush_work->hr_dev = hr_dev;
77 	INIT_WORK(&flush_work->work, flush_work_handle);
78 	atomic_inc(&hr_qp->refcount);
79 	queue_work(hr_dev->irq_workq, &flush_work->work);
80 }
81 
82 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
83 {
84 	struct device *dev = hr_dev->dev;
85 	struct hns_roce_qp *qp;
86 
87 	xa_lock(&hr_dev->qp_table_xa);
88 	qp = __hns_roce_qp_lookup(hr_dev, qpn);
89 	if (qp)
90 		atomic_inc(&qp->refcount);
91 	xa_unlock(&hr_dev->qp_table_xa);
92 
93 	if (!qp) {
94 		dev_warn(dev, "Async event for bogus QP %08x\n", qpn);
95 		return;
96 	}
97 
98 	if (hr_dev->hw_rev != HNS_ROCE_HW_VER1 &&
99 	    (event_type == HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR ||
100 	     event_type == HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR ||
101 	     event_type == HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR ||
102 	     event_type == HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION ||
103 	     event_type == HNS_ROCE_EVENT_TYPE_INVALID_XRCETH)) {
104 		qp->state = IB_QPS_ERR;
105 		if (!test_and_set_bit(HNS_ROCE_FLUSH_FLAG, &qp->flush_flag))
106 			init_flush_work(hr_dev, qp);
107 	}
108 
109 	qp->event(qp, (enum hns_roce_event)event_type);
110 
111 	if (atomic_dec_and_test(&qp->refcount))
112 		complete(&qp->free);
113 }
114 
115 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
116 				 enum hns_roce_event type)
117 {
118 	struct ib_qp *ibqp = &hr_qp->ibqp;
119 	struct ib_event event;
120 
121 	if (ibqp->event_handler) {
122 		event.device = ibqp->device;
123 		event.element.qp = ibqp;
124 		switch (type) {
125 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
126 			event.event = IB_EVENT_PATH_MIG;
127 			break;
128 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
129 			event.event = IB_EVENT_COMM_EST;
130 			break;
131 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
132 			event.event = IB_EVENT_SQ_DRAINED;
133 			break;
134 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
135 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
136 			break;
137 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
138 			event.event = IB_EVENT_QP_FATAL;
139 			break;
140 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
141 			event.event = IB_EVENT_PATH_MIG_ERR;
142 			break;
143 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
144 			event.event = IB_EVENT_QP_REQ_ERR;
145 			break;
146 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
147 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
148 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
149 			event.event = IB_EVENT_QP_ACCESS_ERR;
150 			break;
151 		default:
152 			dev_dbg(ibqp->device->dev.parent, "roce_ib: Unexpected event type %d on QP %06lx\n",
153 				type, hr_qp->qpn);
154 			return;
155 		}
156 		ibqp->event_handler(&event, ibqp->qp_context);
157 	}
158 }
159 
160 static u8 get_least_load_bankid_for_qp(struct hns_roce_bank *bank)
161 {
162 	u32 least_load = bank[0].inuse;
163 	u8 bankid = 0;
164 	u32 bankcnt;
165 	u8 i;
166 
167 	for (i = 1; i < HNS_ROCE_QP_BANK_NUM; i++) {
168 		bankcnt = bank[i].inuse;
169 		if (bankcnt < least_load) {
170 			least_load = bankcnt;
171 			bankid = i;
172 		}
173 	}
174 
175 	return bankid;
176 }
177 
178 static int alloc_qpn_with_bankid(struct hns_roce_bank *bank, u8 bankid,
179 				 unsigned long *qpn)
180 {
181 	int id;
182 
183 	id = ida_alloc_range(&bank->ida, bank->next, bank->max, GFP_KERNEL);
184 	if (id < 0) {
185 		id = ida_alloc_range(&bank->ida, bank->min, bank->max,
186 				     GFP_KERNEL);
187 		if (id < 0)
188 			return id;
189 	}
190 
191 	/* the QPN should keep increasing until the max value is reached. */
192 	bank->next = (id + 1) > bank->max ? bank->min : id + 1;
193 
194 	/* the lower 3 bits is bankid */
195 	*qpn = (id << 3) | bankid;
196 
197 	return 0;
198 }
199 static int alloc_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
200 {
201 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
202 	unsigned long num = 0;
203 	u8 bankid;
204 	int ret;
205 
206 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI) {
207 		/* when hw version is v1, the sqpn is allocated */
208 		if (hr_dev->hw_rev == HNS_ROCE_HW_VER1)
209 			num = HNS_ROCE_MAX_PORTS +
210 			      hr_dev->iboe.phy_port[hr_qp->port];
211 		else
212 			num = 1;
213 
214 		hr_qp->doorbell_qpn = 1;
215 	} else {
216 		mutex_lock(&qp_table->bank_mutex);
217 		bankid = get_least_load_bankid_for_qp(qp_table->bank);
218 
219 		ret = alloc_qpn_with_bankid(&qp_table->bank[bankid], bankid,
220 					    &num);
221 		if (ret) {
222 			ibdev_err(&hr_dev->ib_dev,
223 				  "failed to alloc QPN, ret = %d\n", ret);
224 			mutex_unlock(&qp_table->bank_mutex);
225 			return ret;
226 		}
227 
228 		qp_table->bank[bankid].inuse++;
229 		mutex_unlock(&qp_table->bank_mutex);
230 
231 		hr_qp->doorbell_qpn = (u32)num;
232 	}
233 
234 	hr_qp->qpn = num;
235 
236 	return 0;
237 }
238 
239 enum hns_roce_qp_state to_hns_roce_state(enum ib_qp_state state)
240 {
241 	switch (state) {
242 	case IB_QPS_RESET:
243 		return HNS_ROCE_QP_STATE_RST;
244 	case IB_QPS_INIT:
245 		return HNS_ROCE_QP_STATE_INIT;
246 	case IB_QPS_RTR:
247 		return HNS_ROCE_QP_STATE_RTR;
248 	case IB_QPS_RTS:
249 		return HNS_ROCE_QP_STATE_RTS;
250 	case IB_QPS_SQD:
251 		return HNS_ROCE_QP_STATE_SQD;
252 	case IB_QPS_ERR:
253 		return HNS_ROCE_QP_STATE_ERR;
254 	default:
255 		return HNS_ROCE_QP_NUM_STATE;
256 	}
257 }
258 
259 static void add_qp_to_list(struct hns_roce_dev *hr_dev,
260 			   struct hns_roce_qp *hr_qp,
261 			   struct ib_cq *send_cq, struct ib_cq *recv_cq)
262 {
263 	struct hns_roce_cq *hr_send_cq, *hr_recv_cq;
264 	unsigned long flags;
265 
266 	hr_send_cq = send_cq ? to_hr_cq(send_cq) : NULL;
267 	hr_recv_cq = recv_cq ? to_hr_cq(recv_cq) : NULL;
268 
269 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
270 	hns_roce_lock_cqs(hr_send_cq, hr_recv_cq);
271 
272 	list_add_tail(&hr_qp->node, &hr_dev->qp_list);
273 	if (hr_send_cq)
274 		list_add_tail(&hr_qp->sq_node, &hr_send_cq->sq_list);
275 	if (hr_recv_cq)
276 		list_add_tail(&hr_qp->rq_node, &hr_recv_cq->rq_list);
277 
278 	hns_roce_unlock_cqs(hr_send_cq, hr_recv_cq);
279 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
280 }
281 
282 static int hns_roce_qp_store(struct hns_roce_dev *hr_dev,
283 			     struct hns_roce_qp *hr_qp,
284 			     struct ib_qp_init_attr *init_attr)
285 {
286 	struct xarray *xa = &hr_dev->qp_table_xa;
287 	int ret;
288 
289 	if (!hr_qp->qpn)
290 		return -EINVAL;
291 
292 	ret = xa_err(xa_store_irq(xa, hr_qp->qpn, hr_qp, GFP_KERNEL));
293 	if (ret)
294 		dev_err(hr_dev->dev, "Failed to xa store for QPC\n");
295 	else
296 		/* add QP to device's QP list for softwc */
297 		add_qp_to_list(hr_dev, hr_qp, init_attr->send_cq,
298 			       init_attr->recv_cq);
299 
300 	return ret;
301 }
302 
303 static int alloc_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
304 {
305 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
306 	struct device *dev = hr_dev->dev;
307 	int ret;
308 
309 	if (!hr_qp->qpn)
310 		return -EINVAL;
311 
312 	/* In v1 engine, GSI QP context is saved in the RoCE hw's register */
313 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI &&
314 	    hr_dev->hw_rev == HNS_ROCE_HW_VER1)
315 		return 0;
316 
317 	/* Alloc memory for QPC */
318 	ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
319 	if (ret) {
320 		dev_err(dev, "Failed to get QPC table\n");
321 		goto err_out;
322 	}
323 
324 	/* Alloc memory for IRRL */
325 	ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
326 	if (ret) {
327 		dev_err(dev, "Failed to get IRRL table\n");
328 		goto err_put_qp;
329 	}
330 
331 	if (hr_dev->caps.trrl_entry_sz) {
332 		/* Alloc memory for TRRL */
333 		ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
334 					 hr_qp->qpn);
335 		if (ret) {
336 			dev_err(dev, "Failed to get TRRL table\n");
337 			goto err_put_irrl;
338 		}
339 	}
340 
341 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
342 		/* Alloc memory for SCC CTX */
343 		ret = hns_roce_table_get(hr_dev, &qp_table->sccc_table,
344 					 hr_qp->qpn);
345 		if (ret) {
346 			dev_err(dev, "Failed to get SCC CTX table\n");
347 			goto err_put_trrl;
348 		}
349 	}
350 
351 	return 0;
352 
353 err_put_trrl:
354 	if (hr_dev->caps.trrl_entry_sz)
355 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
356 
357 err_put_irrl:
358 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
359 
360 err_put_qp:
361 	hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
362 
363 err_out:
364 	return ret;
365 }
366 
367 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
368 {
369 	struct xarray *xa = &hr_dev->qp_table_xa;
370 	unsigned long flags;
371 
372 	list_del(&hr_qp->node);
373 
374 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
375 		list_del(&hr_qp->sq_node);
376 
377 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_INI &&
378 	    hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
379 		list_del(&hr_qp->rq_node);
380 
381 	xa_lock_irqsave(xa, flags);
382 	__xa_erase(xa, hr_qp->qpn & (hr_dev->caps.num_qps - 1));
383 	xa_unlock_irqrestore(xa, flags);
384 }
385 
386 static void free_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
387 {
388 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
389 
390 	/* In v1 engine, GSI QP context is saved in the RoCE hw's register */
391 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI &&
392 	    hr_dev->hw_rev == HNS_ROCE_HW_VER1)
393 		return;
394 
395 	if (hr_dev->caps.trrl_entry_sz)
396 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
397 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
398 }
399 
400 static inline u8 get_qp_bankid(unsigned long qpn)
401 {
402 	/* The lower 3 bits of QPN are used to hash to different banks */
403 	return (u8)(qpn & GENMASK(2, 0));
404 }
405 
406 static void free_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
407 {
408 	u8 bankid;
409 
410 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI)
411 		return;
412 
413 	if (hr_qp->qpn < hr_dev->caps.reserved_qps)
414 		return;
415 
416 	bankid = get_qp_bankid(hr_qp->qpn);
417 
418 	ida_free(&hr_dev->qp_table.bank[bankid].ida, hr_qp->qpn >> 3);
419 
420 	mutex_lock(&hr_dev->qp_table.bank_mutex);
421 	hr_dev->qp_table.bank[bankid].inuse--;
422 	mutex_unlock(&hr_dev->qp_table.bank_mutex);
423 }
424 
425 static u32 proc_rq_sge(struct hns_roce_dev *dev, struct hns_roce_qp *hr_qp,
426 		       bool user)
427 {
428 	u32 max_sge = dev->caps.max_rq_sg;
429 
430 	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
431 		return max_sge;
432 
433 	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
434 	 * calculate number of max_sge with reserved SGEs when allocating wqe
435 	 * buf, so there is no need to do this again in kernel. But the number
436 	 * may exceed the capacity of SGEs recorded in the firmware, so the
437 	 * kernel driver should just adapt the value accordingly.
438 	 */
439 	if (user)
440 		max_sge = roundup_pow_of_two(max_sge + 1);
441 	else
442 		hr_qp->rq.rsv_sge = 1;
443 
444 	return max_sge;
445 }
446 
447 static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
448 		       struct hns_roce_qp *hr_qp, int has_rq, bool user)
449 {
450 	u32 max_sge = proc_rq_sge(hr_dev, hr_qp, user);
451 	u32 cnt;
452 
453 	/* If srq exist, set zero for relative number of rq */
454 	if (!has_rq) {
455 		hr_qp->rq.wqe_cnt = 0;
456 		hr_qp->rq.max_gs = 0;
457 		hr_qp->rq_inl_buf.wqe_cnt = 0;
458 		cap->max_recv_wr = 0;
459 		cap->max_recv_sge = 0;
460 
461 		return 0;
462 	}
463 
464 	/* Check the validity of QP support capacity */
465 	if (!cap->max_recv_wr || cap->max_recv_wr > hr_dev->caps.max_wqes ||
466 	    cap->max_recv_sge > max_sge) {
467 		ibdev_err(&hr_dev->ib_dev,
468 			  "RQ config error, depth = %u, sge = %u\n",
469 			  cap->max_recv_wr, cap->max_recv_sge);
470 		return -EINVAL;
471 	}
472 
473 	cnt = roundup_pow_of_two(max(cap->max_recv_wr, hr_dev->caps.min_wqes));
474 	if (cnt > hr_dev->caps.max_wqes) {
475 		ibdev_err(&hr_dev->ib_dev, "rq depth %u too large\n",
476 			  cap->max_recv_wr);
477 		return -EINVAL;
478 	}
479 
480 	hr_qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge) +
481 					      hr_qp->rq.rsv_sge);
482 
483 	if (hr_dev->caps.max_rq_sg <= HNS_ROCE_SGE_IN_WQE)
484 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz);
485 	else
486 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz *
487 					    hr_qp->rq.max_gs);
488 
489 	hr_qp->rq.wqe_cnt = cnt;
490 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE &&
491 	    hr_qp->ibqp.qp_type != IB_QPT_UD &&
492 	    hr_qp->ibqp.qp_type != IB_QPT_GSI)
493 		hr_qp->rq_inl_buf.wqe_cnt = cnt;
494 	else
495 		hr_qp->rq_inl_buf.wqe_cnt = 0;
496 
497 	cap->max_recv_wr = cnt;
498 	cap->max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
499 
500 	return 0;
501 }
502 
503 static u32 get_wqe_ext_sge_cnt(struct hns_roce_qp *qp)
504 {
505 	/* GSI/UD QP only has extended sge */
506 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_UD)
507 		return qp->sq.max_gs;
508 
509 	if (qp->sq.max_gs > HNS_ROCE_SGE_IN_WQE)
510 		return qp->sq.max_gs - HNS_ROCE_SGE_IN_WQE;
511 
512 	return 0;
513 }
514 
515 static void set_ext_sge_param(struct hns_roce_dev *hr_dev, u32 sq_wqe_cnt,
516 			      struct hns_roce_qp *hr_qp, struct ib_qp_cap *cap)
517 {
518 	u32 total_sge_cnt;
519 	u32 wqe_sge_cnt;
520 
521 	hr_qp->sge.sge_shift = HNS_ROCE_SGE_SHIFT;
522 
523 	if (hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
524 		hr_qp->sq.max_gs = HNS_ROCE_SGE_IN_WQE;
525 		return;
526 	}
527 
528 	hr_qp->sq.max_gs = max(1U, cap->max_send_sge);
529 
530 	wqe_sge_cnt = get_wqe_ext_sge_cnt(hr_qp);
531 
532 	/* If the number of extended sge is not zero, they MUST use the
533 	 * space of HNS_HW_PAGE_SIZE at least.
534 	 */
535 	if (wqe_sge_cnt) {
536 		total_sge_cnt = roundup_pow_of_two(sq_wqe_cnt * wqe_sge_cnt);
537 		hr_qp->sge.sge_cnt = max(total_sge_cnt,
538 				(u32)HNS_HW_PAGE_SIZE / HNS_ROCE_SGE_SIZE);
539 	}
540 }
541 
542 static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
543 					struct ib_qp_cap *cap,
544 					struct hns_roce_ib_create_qp *ucmd)
545 {
546 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
547 	u8 max_sq_stride = ilog2(roundup_sq_stride);
548 
549 	/* Sanity check SQ size before proceeding */
550 	if (ucmd->log_sq_stride > max_sq_stride ||
551 	    ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
552 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ stride size.\n");
553 		return -EINVAL;
554 	}
555 
556 	if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
557 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ SGE size %u.\n",
558 			  cap->max_send_sge);
559 		return -EINVAL;
560 	}
561 
562 	return 0;
563 }
564 
565 static int set_user_sq_size(struct hns_roce_dev *hr_dev,
566 			    struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp,
567 			    struct hns_roce_ib_create_qp *ucmd)
568 {
569 	struct ib_device *ibdev = &hr_dev->ib_dev;
570 	u32 cnt = 0;
571 	int ret;
572 
573 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
574 	    cnt > hr_dev->caps.max_wqes)
575 		return -EINVAL;
576 
577 	ret = check_sq_size_with_integrity(hr_dev, cap, ucmd);
578 	if (ret) {
579 		ibdev_err(ibdev, "failed to check user SQ size, ret = %d.\n",
580 			  ret);
581 		return ret;
582 	}
583 
584 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
585 
586 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
587 	hr_qp->sq.wqe_cnt = cnt;
588 
589 	return 0;
590 }
591 
592 static int set_wqe_buf_attr(struct hns_roce_dev *hr_dev,
593 			    struct hns_roce_qp *hr_qp,
594 			    struct hns_roce_buf_attr *buf_attr)
595 {
596 	int buf_size;
597 	int idx = 0;
598 
599 	hr_qp->buff_size = 0;
600 
601 	/* SQ WQE */
602 	hr_qp->sq.offset = 0;
603 	buf_size = to_hr_hem_entries_size(hr_qp->sq.wqe_cnt,
604 					  hr_qp->sq.wqe_shift);
605 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
606 		buf_attr->region[idx].size = buf_size;
607 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sq_hop_num;
608 		idx++;
609 		hr_qp->buff_size += buf_size;
610 	}
611 
612 	/* extend SGE WQE in SQ */
613 	hr_qp->sge.offset = hr_qp->buff_size;
614 	buf_size = to_hr_hem_entries_size(hr_qp->sge.sge_cnt,
615 					  hr_qp->sge.sge_shift);
616 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
617 		buf_attr->region[idx].size = buf_size;
618 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sge_hop_num;
619 		idx++;
620 		hr_qp->buff_size += buf_size;
621 	}
622 
623 	/* RQ WQE */
624 	hr_qp->rq.offset = hr_qp->buff_size;
625 	buf_size = to_hr_hem_entries_size(hr_qp->rq.wqe_cnt,
626 					  hr_qp->rq.wqe_shift);
627 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
628 		buf_attr->region[idx].size = buf_size;
629 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_rq_hop_num;
630 		idx++;
631 		hr_qp->buff_size += buf_size;
632 	}
633 
634 	if (hr_qp->buff_size < 1)
635 		return -EINVAL;
636 
637 	buf_attr->page_shift = HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
638 	buf_attr->region_count = idx;
639 
640 	return 0;
641 }
642 
643 static int set_kernel_sq_size(struct hns_roce_dev *hr_dev,
644 			      struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp)
645 {
646 	struct ib_device *ibdev = &hr_dev->ib_dev;
647 	u32 cnt;
648 
649 	if (!cap->max_send_wr || cap->max_send_wr > hr_dev->caps.max_wqes ||
650 	    cap->max_send_sge > hr_dev->caps.max_sq_sg) {
651 		ibdev_err(ibdev,
652 			  "failed to check SQ WR or SGE num, ret = %d.\n",
653 			  -EINVAL);
654 		return -EINVAL;
655 	}
656 
657 	cnt = roundup_pow_of_two(max(cap->max_send_wr, hr_dev->caps.min_wqes));
658 	if (cnt > hr_dev->caps.max_wqes) {
659 		ibdev_err(ibdev, "failed to check WQE num, WQE num = %u.\n",
660 			  cnt);
661 		return -EINVAL;
662 	}
663 
664 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
665 	hr_qp->sq.wqe_cnt = cnt;
666 
667 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
668 
669 	/* sync the parameters of kernel QP to user's configuration */
670 	cap->max_send_wr = cnt;
671 	cap->max_send_sge = hr_qp->sq.max_gs;
672 
673 	return 0;
674 }
675 
676 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
677 {
678 	if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
679 		return 0;
680 
681 	return 1;
682 }
683 
684 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
685 {
686 	if (attr->qp_type == IB_QPT_XRC_INI ||
687 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
688 	    !attr->cap.max_recv_wr)
689 		return 0;
690 
691 	return 1;
692 }
693 
694 static int alloc_rq_inline_buf(struct hns_roce_qp *hr_qp,
695 			       struct ib_qp_init_attr *init_attr)
696 {
697 	u32 max_recv_sge = init_attr->cap.max_recv_sge;
698 	u32 wqe_cnt = hr_qp->rq_inl_buf.wqe_cnt;
699 	struct hns_roce_rinl_wqe *wqe_list;
700 	int i;
701 
702 	/* allocate recv inline buf */
703 	wqe_list = kcalloc(wqe_cnt, sizeof(struct hns_roce_rinl_wqe),
704 			   GFP_KERNEL);
705 
706 	if (!wqe_list)
707 		goto err;
708 
709 	/* Allocate a continuous buffer for all inline sge we need */
710 	wqe_list[0].sg_list = kcalloc(wqe_cnt, (max_recv_sge *
711 				      sizeof(struct hns_roce_rinl_sge)),
712 				      GFP_KERNEL);
713 	if (!wqe_list[0].sg_list)
714 		goto err_wqe_list;
715 
716 	/* Assign buffers of sg_list to each inline wqe */
717 	for (i = 1; i < wqe_cnt; i++)
718 		wqe_list[i].sg_list = &wqe_list[0].sg_list[i * max_recv_sge];
719 
720 	hr_qp->rq_inl_buf.wqe_list = wqe_list;
721 
722 	return 0;
723 
724 err_wqe_list:
725 	kfree(wqe_list);
726 
727 err:
728 	return -ENOMEM;
729 }
730 
731 static void free_rq_inline_buf(struct hns_roce_qp *hr_qp)
732 {
733 	if (hr_qp->rq_inl_buf.wqe_list)
734 		kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
735 	kfree(hr_qp->rq_inl_buf.wqe_list);
736 }
737 
738 static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
739 			struct ib_qp_init_attr *init_attr,
740 			struct ib_udata *udata, unsigned long addr)
741 {
742 	struct ib_device *ibdev = &hr_dev->ib_dev;
743 	struct hns_roce_buf_attr buf_attr = {};
744 	int ret;
745 
746 	if (!udata && hr_qp->rq_inl_buf.wqe_cnt) {
747 		ret = alloc_rq_inline_buf(hr_qp, init_attr);
748 		if (ret) {
749 			ibdev_err(ibdev,
750 				  "failed to alloc inline buf, ret = %d.\n",
751 				  ret);
752 			return ret;
753 		}
754 	} else {
755 		hr_qp->rq_inl_buf.wqe_list = NULL;
756 	}
757 
758 	ret = set_wqe_buf_attr(hr_dev, hr_qp, &buf_attr);
759 	if (ret) {
760 		ibdev_err(ibdev, "failed to split WQE buf, ret = %d.\n", ret);
761 		goto err_inline;
762 	}
763 	ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, &buf_attr,
764 				  HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
765 				  udata, addr);
766 	if (ret) {
767 		ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret);
768 		goto err_inline;
769 	}
770 
771 	return 0;
772 err_inline:
773 	free_rq_inline_buf(hr_qp);
774 
775 	return ret;
776 }
777 
778 static void free_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
779 {
780 	hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
781 	free_rq_inline_buf(hr_qp);
782 }
783 
784 static inline bool user_qp_has_sdb(struct hns_roce_dev *hr_dev,
785 				   struct ib_qp_init_attr *init_attr,
786 				   struct ib_udata *udata,
787 				   struct hns_roce_ib_create_qp_resp *resp,
788 				   struct hns_roce_ib_create_qp *ucmd)
789 {
790 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
791 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
792 		hns_roce_qp_has_sq(init_attr) &&
793 		udata->inlen >= offsetofend(typeof(*ucmd), sdb_addr));
794 }
795 
796 static inline bool user_qp_has_rdb(struct hns_roce_dev *hr_dev,
797 				   struct ib_qp_init_attr *init_attr,
798 				   struct ib_udata *udata,
799 				   struct hns_roce_ib_create_qp_resp *resp)
800 {
801 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
802 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
803 		hns_roce_qp_has_rq(init_attr));
804 }
805 
806 static inline bool kernel_qp_has_rdb(struct hns_roce_dev *hr_dev,
807 				     struct ib_qp_init_attr *init_attr)
808 {
809 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
810 		hns_roce_qp_has_rq(init_attr));
811 }
812 
813 static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
814 		       struct ib_qp_init_attr *init_attr,
815 		       struct ib_udata *udata,
816 		       struct hns_roce_ib_create_qp *ucmd,
817 		       struct hns_roce_ib_create_qp_resp *resp)
818 {
819 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
820 		udata, struct hns_roce_ucontext, ibucontext);
821 	struct ib_device *ibdev = &hr_dev->ib_dev;
822 	int ret;
823 
824 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SDI_MODE)
825 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_OWNER_DB;
826 
827 	if (udata) {
828 		if (user_qp_has_sdb(hr_dev, init_attr, udata, resp, ucmd)) {
829 			ret = hns_roce_db_map_user(uctx, udata, ucmd->sdb_addr,
830 						   &hr_qp->sdb);
831 			if (ret) {
832 				ibdev_err(ibdev,
833 					  "failed to map user SQ doorbell, ret = %d.\n",
834 					  ret);
835 				goto err_out;
836 			}
837 			hr_qp->en_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
838 			resp->cap_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
839 		}
840 
841 		if (user_qp_has_rdb(hr_dev, init_attr, udata, resp)) {
842 			ret = hns_roce_db_map_user(uctx, udata, ucmd->db_addr,
843 						   &hr_qp->rdb);
844 			if (ret) {
845 				ibdev_err(ibdev,
846 					  "failed to map user RQ doorbell, ret = %d.\n",
847 					  ret);
848 				goto err_sdb;
849 			}
850 			hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
851 			resp->cap_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
852 		}
853 	} else {
854 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
855 			hr_qp->sq.db_reg = hr_dev->mem_base +
856 					   HNS_ROCE_DWQE_SIZE * hr_qp->qpn;
857 		else
858 			hr_qp->sq.db_reg =
859 				hr_dev->reg_base + hr_dev->sdb_offset +
860 				DB_REG_OFFSET * hr_dev->priv_uar.index;
861 
862 		hr_qp->rq.db_reg = hr_dev->reg_base + hr_dev->odb_offset +
863 				   DB_REG_OFFSET * hr_dev->priv_uar.index;
864 
865 		if (kernel_qp_has_rdb(hr_dev, init_attr)) {
866 			ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
867 			if (ret) {
868 				ibdev_err(ibdev,
869 					  "failed to alloc kernel RQ doorbell, ret = %d.\n",
870 					  ret);
871 				goto err_out;
872 			}
873 			*hr_qp->rdb.db_record = 0;
874 			hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
875 		}
876 	}
877 
878 	return 0;
879 err_sdb:
880 	if (udata && hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
881 		hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
882 err_out:
883 	return ret;
884 }
885 
886 static void free_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
887 		       struct ib_udata *udata)
888 {
889 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
890 		udata, struct hns_roce_ucontext, ibucontext);
891 
892 	if (udata) {
893 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
894 			hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
895 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
896 			hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
897 	} else {
898 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
899 			hns_roce_free_db(hr_dev, &hr_qp->rdb);
900 	}
901 }
902 
903 static int alloc_kernel_wrid(struct hns_roce_dev *hr_dev,
904 			     struct hns_roce_qp *hr_qp)
905 {
906 	struct ib_device *ibdev = &hr_dev->ib_dev;
907 	u64 *sq_wrid = NULL;
908 	u64 *rq_wrid = NULL;
909 	int ret;
910 
911 	sq_wrid = kcalloc(hr_qp->sq.wqe_cnt, sizeof(u64), GFP_KERNEL);
912 	if (ZERO_OR_NULL_PTR(sq_wrid)) {
913 		ibdev_err(ibdev, "failed to alloc SQ wrid.\n");
914 		return -ENOMEM;
915 	}
916 
917 	if (hr_qp->rq.wqe_cnt) {
918 		rq_wrid = kcalloc(hr_qp->rq.wqe_cnt, sizeof(u64), GFP_KERNEL);
919 		if (ZERO_OR_NULL_PTR(rq_wrid)) {
920 			ibdev_err(ibdev, "failed to alloc RQ wrid.\n");
921 			ret = -ENOMEM;
922 			goto err_sq;
923 		}
924 	}
925 
926 	hr_qp->sq.wrid = sq_wrid;
927 	hr_qp->rq.wrid = rq_wrid;
928 	return 0;
929 err_sq:
930 	kfree(sq_wrid);
931 
932 	return ret;
933 }
934 
935 static void free_kernel_wrid(struct hns_roce_qp *hr_qp)
936 {
937 	kfree(hr_qp->rq.wrid);
938 	kfree(hr_qp->sq.wrid);
939 }
940 
941 static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
942 			struct ib_qp_init_attr *init_attr,
943 			struct ib_udata *udata,
944 			struct hns_roce_ib_create_qp *ucmd)
945 {
946 	struct ib_device *ibdev = &hr_dev->ib_dev;
947 	int ret;
948 
949 	hr_qp->ibqp.qp_type = init_attr->qp_type;
950 
951 	if (init_attr->cap.max_inline_data > hr_dev->caps.max_sq_inline)
952 		init_attr->cap.max_inline_data = hr_dev->caps.max_sq_inline;
953 
954 	hr_qp->max_inline_data = init_attr->cap.max_inline_data;
955 
956 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
957 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
958 	else
959 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
960 
961 	ret = set_rq_size(hr_dev, &init_attr->cap, hr_qp,
962 			  hns_roce_qp_has_rq(init_attr), !!udata);
963 	if (ret) {
964 		ibdev_err(ibdev, "failed to set user RQ size, ret = %d.\n",
965 			  ret);
966 		return ret;
967 	}
968 
969 	if (udata) {
970 		ret = ib_copy_from_udata(ucmd, udata,
971 					 min(udata->inlen, sizeof(*ucmd)));
972 		if (ret) {
973 			ibdev_err(ibdev,
974 				  "failed to copy QP ucmd, ret = %d\n", ret);
975 			return ret;
976 		}
977 
978 		ret = set_user_sq_size(hr_dev, &init_attr->cap, hr_qp, ucmd);
979 		if (ret)
980 			ibdev_err(ibdev,
981 				  "failed to set user SQ size, ret = %d.\n",
982 				  ret);
983 	} else {
984 		ret = set_kernel_sq_size(hr_dev, &init_attr->cap, hr_qp);
985 		if (ret)
986 			ibdev_err(ibdev,
987 				  "failed to set kernel SQ size, ret = %d.\n",
988 				  ret);
989 	}
990 
991 	return ret;
992 }
993 
994 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
995 				     struct ib_pd *ib_pd,
996 				     struct ib_qp_init_attr *init_attr,
997 				     struct ib_udata *udata,
998 				     struct hns_roce_qp *hr_qp)
999 {
1000 	struct hns_roce_ib_create_qp_resp resp = {};
1001 	struct ib_device *ibdev = &hr_dev->ib_dev;
1002 	struct hns_roce_ib_create_qp ucmd;
1003 	int ret;
1004 
1005 	mutex_init(&hr_qp->mutex);
1006 	spin_lock_init(&hr_qp->sq.lock);
1007 	spin_lock_init(&hr_qp->rq.lock);
1008 
1009 	hr_qp->state = IB_QPS_RESET;
1010 	hr_qp->flush_flag = 0;
1011 
1012 	if (init_attr->create_flags)
1013 		return -EOPNOTSUPP;
1014 
1015 	ret = set_qp_param(hr_dev, hr_qp, init_attr, udata, &ucmd);
1016 	if (ret) {
1017 		ibdev_err(ibdev, "failed to set QP param, ret = %d.\n", ret);
1018 		return ret;
1019 	}
1020 
1021 	if (!udata) {
1022 		ret = alloc_kernel_wrid(hr_dev, hr_qp);
1023 		if (ret) {
1024 			ibdev_err(ibdev, "failed to alloc wrid, ret = %d.\n",
1025 				  ret);
1026 			return ret;
1027 		}
1028 	}
1029 
1030 	ret = alloc_qp_buf(hr_dev, hr_qp, init_attr, udata, ucmd.buf_addr);
1031 	if (ret) {
1032 		ibdev_err(ibdev, "failed to alloc QP buffer, ret = %d.\n", ret);
1033 		goto err_buf;
1034 	}
1035 
1036 	ret = alloc_qpn(hr_dev, hr_qp);
1037 	if (ret) {
1038 		ibdev_err(ibdev, "failed to alloc QPN, ret = %d.\n", ret);
1039 		goto err_qpn;
1040 	}
1041 
1042 	ret = alloc_qp_db(hr_dev, hr_qp, init_attr, udata, &ucmd, &resp);
1043 	if (ret) {
1044 		ibdev_err(ibdev, "failed to alloc QP doorbell, ret = %d.\n",
1045 			  ret);
1046 		goto err_db;
1047 	}
1048 
1049 	ret = alloc_qpc(hr_dev, hr_qp);
1050 	if (ret) {
1051 		ibdev_err(ibdev, "failed to alloc QP context, ret = %d.\n",
1052 			  ret);
1053 		goto err_qpc;
1054 	}
1055 
1056 	ret = hns_roce_qp_store(hr_dev, hr_qp, init_attr);
1057 	if (ret) {
1058 		ibdev_err(ibdev, "failed to store QP, ret = %d.\n", ret);
1059 		goto err_store;
1060 	}
1061 
1062 	if (udata) {
1063 		ret = ib_copy_to_udata(udata, &resp,
1064 				       min(udata->outlen, sizeof(resp)));
1065 		if (ret) {
1066 			ibdev_err(ibdev, "copy qp resp failed!\n");
1067 			goto err_store;
1068 		}
1069 	}
1070 
1071 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
1072 		ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
1073 		if (ret)
1074 			goto err_flow_ctrl;
1075 	}
1076 
1077 	hr_qp->ibqp.qp_num = hr_qp->qpn;
1078 	hr_qp->event = hns_roce_ib_qp_event;
1079 	atomic_set(&hr_qp->refcount, 1);
1080 	init_completion(&hr_qp->free);
1081 
1082 	return 0;
1083 
1084 err_flow_ctrl:
1085 	hns_roce_qp_remove(hr_dev, hr_qp);
1086 err_store:
1087 	free_qpc(hr_dev, hr_qp);
1088 err_qpc:
1089 	free_qp_db(hr_dev, hr_qp, udata);
1090 err_db:
1091 	free_qpn(hr_dev, hr_qp);
1092 err_qpn:
1093 	free_qp_buf(hr_dev, hr_qp);
1094 err_buf:
1095 	free_kernel_wrid(hr_qp);
1096 	return ret;
1097 }
1098 
1099 void hns_roce_qp_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1100 			 struct ib_udata *udata)
1101 {
1102 	if (atomic_dec_and_test(&hr_qp->refcount))
1103 		complete(&hr_qp->free);
1104 	wait_for_completion(&hr_qp->free);
1105 
1106 	free_qpc(hr_dev, hr_qp);
1107 	free_qpn(hr_dev, hr_qp);
1108 	free_qp_buf(hr_dev, hr_qp);
1109 	free_kernel_wrid(hr_qp);
1110 	free_qp_db(hr_dev, hr_qp, udata);
1111 
1112 	kfree(hr_qp);
1113 }
1114 
1115 static int check_qp_type(struct hns_roce_dev *hr_dev, enum ib_qp_type type,
1116 			 bool is_user)
1117 {
1118 	switch (type) {
1119 	case IB_QPT_XRC_INI:
1120 	case IB_QPT_XRC_TGT:
1121 		if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC))
1122 			goto out;
1123 		break;
1124 	case IB_QPT_UD:
1125 		if (hr_dev->pci_dev->revision <= PCI_REVISION_ID_HIP08 &&
1126 		    is_user)
1127 			goto out;
1128 		break;
1129 	case IB_QPT_RC:
1130 	case IB_QPT_GSI:
1131 		break;
1132 	default:
1133 		goto out;
1134 	}
1135 
1136 	return 0;
1137 
1138 out:
1139 	ibdev_err(&hr_dev->ib_dev, "not support QP type %d\n", type);
1140 
1141 	return -EOPNOTSUPP;
1142 }
1143 
1144 struct ib_qp *hns_roce_create_qp(struct ib_pd *pd,
1145 				 struct ib_qp_init_attr *init_attr,
1146 				 struct ib_udata *udata)
1147 {
1148 	struct ib_device *ibdev = pd ? pd->device : init_attr->xrcd->device;
1149 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
1150 	struct hns_roce_qp *hr_qp;
1151 	int ret;
1152 
1153 	ret = check_qp_type(hr_dev, init_attr->qp_type, !!udata);
1154 	if (ret)
1155 		return ERR_PTR(ret);
1156 
1157 	hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
1158 	if (!hr_qp)
1159 		return ERR_PTR(-ENOMEM);
1160 
1161 	if (init_attr->qp_type == IB_QPT_XRC_INI)
1162 		init_attr->recv_cq = NULL;
1163 
1164 	if (init_attr->qp_type == IB_QPT_XRC_TGT) {
1165 		hr_qp->xrcdn = to_hr_xrcd(init_attr->xrcd)->xrcdn;
1166 		init_attr->recv_cq = NULL;
1167 		init_attr->send_cq = NULL;
1168 	}
1169 
1170 	if (init_attr->qp_type == IB_QPT_GSI) {
1171 		hr_qp->port = init_attr->port_num - 1;
1172 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
1173 	}
1174 
1175 	ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, hr_qp);
1176 	if (ret) {
1177 		ibdev_err(ibdev, "Create QP type 0x%x failed(%d)\n",
1178 			  init_attr->qp_type, ret);
1179 
1180 		kfree(hr_qp);
1181 		return ERR_PTR(ret);
1182 	}
1183 
1184 	return &hr_qp->ibqp;
1185 }
1186 
1187 int to_hr_qp_type(int qp_type)
1188 {
1189 	switch (qp_type) {
1190 	case IB_QPT_RC:
1191 		return SERV_TYPE_RC;
1192 	case IB_QPT_UD:
1193 	case IB_QPT_GSI:
1194 		return SERV_TYPE_UD;
1195 	case IB_QPT_XRC_INI:
1196 	case IB_QPT_XRC_TGT:
1197 		return SERV_TYPE_XRC;
1198 	default:
1199 		return -1;
1200 	}
1201 }
1202 
1203 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1204 			      struct hns_roce_qp *hr_qp,
1205 			      struct ib_qp_attr *attr, int attr_mask)
1206 {
1207 	enum ib_mtu active_mtu;
1208 	int p;
1209 
1210 	p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1211 	active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1212 
1213 	if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1214 	    attr->path_mtu > hr_dev->caps.max_mtu) ||
1215 	    attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1216 		ibdev_err(&hr_dev->ib_dev,
1217 			"attr path_mtu(%d)invalid while modify qp",
1218 			attr->path_mtu);
1219 		return -EINVAL;
1220 	}
1221 
1222 	return 0;
1223 }
1224 
1225 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1226 				  int attr_mask)
1227 {
1228 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1229 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1230 	int p;
1231 
1232 	if ((attr_mask & IB_QP_PORT) &&
1233 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1234 		ibdev_err(&hr_dev->ib_dev, "invalid attr, port_num = %u.\n",
1235 			  attr->port_num);
1236 		return -EINVAL;
1237 	}
1238 
1239 	if (attr_mask & IB_QP_PKEY_INDEX) {
1240 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1241 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1242 			ibdev_err(&hr_dev->ib_dev,
1243 				  "invalid attr, pkey_index = %u.\n",
1244 				  attr->pkey_index);
1245 			return -EINVAL;
1246 		}
1247 	}
1248 
1249 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1250 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1251 		ibdev_err(&hr_dev->ib_dev,
1252 			  "invalid attr, max_rd_atomic = %u.\n",
1253 			  attr->max_rd_atomic);
1254 		return -EINVAL;
1255 	}
1256 
1257 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1258 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1259 		ibdev_err(&hr_dev->ib_dev,
1260 			  "invalid attr, max_dest_rd_atomic = %u.\n",
1261 			  attr->max_dest_rd_atomic);
1262 		return -EINVAL;
1263 	}
1264 
1265 	if (attr_mask & IB_QP_PATH_MTU)
1266 		return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1267 
1268 	return 0;
1269 }
1270 
1271 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1272 		       int attr_mask, struct ib_udata *udata)
1273 {
1274 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1275 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1276 	enum ib_qp_state cur_state, new_state;
1277 	int ret = -EINVAL;
1278 
1279 	mutex_lock(&hr_qp->mutex);
1280 
1281 	if (attr_mask & IB_QP_CUR_STATE && attr->cur_qp_state != hr_qp->state)
1282 		goto out;
1283 
1284 	cur_state = hr_qp->state;
1285 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1286 
1287 	if (ibqp->uobject &&
1288 	    (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1289 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB) {
1290 			hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1291 
1292 			if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
1293 				hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1294 		} else {
1295 			ibdev_warn(&hr_dev->ib_dev,
1296 				  "flush cqe is not supported in userspace!\n");
1297 			goto out;
1298 		}
1299 	}
1300 
1301 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1302 				attr_mask)) {
1303 		ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1304 		goto out;
1305 	}
1306 
1307 	ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1308 	if (ret)
1309 		goto out;
1310 
1311 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1312 		if (hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
1313 			ret = -EPERM;
1314 			ibdev_err(&hr_dev->ib_dev,
1315 				  "RST2RST state is not supported\n");
1316 		} else {
1317 			ret = 0;
1318 		}
1319 
1320 		goto out;
1321 	}
1322 
1323 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1324 				    new_state);
1325 
1326 out:
1327 	mutex_unlock(&hr_qp->mutex);
1328 
1329 	return ret;
1330 }
1331 
1332 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1333 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1334 {
1335 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1336 		__acquire(&send_cq->lock);
1337 		__acquire(&recv_cq->lock);
1338 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1339 		spin_lock_irq(&send_cq->lock);
1340 		__acquire(&recv_cq->lock);
1341 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1342 		spin_lock_irq(&recv_cq->lock);
1343 		__acquire(&send_cq->lock);
1344 	} else if (send_cq == recv_cq) {
1345 		spin_lock_irq(&send_cq->lock);
1346 		__acquire(&recv_cq->lock);
1347 	} else if (send_cq->cqn < recv_cq->cqn) {
1348 		spin_lock_irq(&send_cq->lock);
1349 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1350 	} else {
1351 		spin_lock_irq(&recv_cq->lock);
1352 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1353 	}
1354 }
1355 
1356 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1357 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1358 			 __releases(&recv_cq->lock)
1359 {
1360 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1361 		__release(&recv_cq->lock);
1362 		__release(&send_cq->lock);
1363 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1364 		__release(&recv_cq->lock);
1365 		spin_unlock(&send_cq->lock);
1366 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1367 		__release(&send_cq->lock);
1368 		spin_unlock(&recv_cq->lock);
1369 	} else if (send_cq == recv_cq) {
1370 		__release(&recv_cq->lock);
1371 		spin_unlock_irq(&send_cq->lock);
1372 	} else if (send_cq->cqn < recv_cq->cqn) {
1373 		spin_unlock(&recv_cq->lock);
1374 		spin_unlock_irq(&send_cq->lock);
1375 	} else {
1376 		spin_unlock(&send_cq->lock);
1377 		spin_unlock_irq(&recv_cq->lock);
1378 	}
1379 }
1380 
1381 static inline void *get_wqe(struct hns_roce_qp *hr_qp, int offset)
1382 {
1383 	return hns_roce_buf_offset(hr_qp->mtr.kmem, offset);
1384 }
1385 
1386 void *hns_roce_get_recv_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1387 {
1388 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1389 }
1390 
1391 void *hns_roce_get_send_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1392 {
1393 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1394 }
1395 
1396 void *hns_roce_get_extend_sge(struct hns_roce_qp *hr_qp, unsigned int n)
1397 {
1398 	return get_wqe(hr_qp, hr_qp->sge.offset + (n << hr_qp->sge.sge_shift));
1399 }
1400 
1401 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, u32 nreq,
1402 			  struct ib_cq *ib_cq)
1403 {
1404 	struct hns_roce_cq *hr_cq;
1405 	u32 cur;
1406 
1407 	cur = hr_wq->head - hr_wq->tail;
1408 	if (likely(cur + nreq < hr_wq->wqe_cnt))
1409 		return false;
1410 
1411 	hr_cq = to_hr_cq(ib_cq);
1412 	spin_lock(&hr_cq->lock);
1413 	cur = hr_wq->head - hr_wq->tail;
1414 	spin_unlock(&hr_cq->lock);
1415 
1416 	return cur + nreq >= hr_wq->wqe_cnt;
1417 }
1418 
1419 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1420 {
1421 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1422 	unsigned int reserved_from_bot;
1423 	unsigned int i;
1424 
1425 	mutex_init(&qp_table->scc_mutex);
1426 	mutex_init(&qp_table->bank_mutex);
1427 	xa_init(&hr_dev->qp_table_xa);
1428 
1429 	reserved_from_bot = hr_dev->caps.reserved_qps;
1430 
1431 	for (i = 0; i < reserved_from_bot; i++) {
1432 		hr_dev->qp_table.bank[get_qp_bankid(i)].inuse++;
1433 		hr_dev->qp_table.bank[get_qp_bankid(i)].min++;
1434 	}
1435 
1436 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) {
1437 		ida_init(&hr_dev->qp_table.bank[i].ida);
1438 		hr_dev->qp_table.bank[i].max = hr_dev->caps.num_qps /
1439 					       HNS_ROCE_QP_BANK_NUM - 1;
1440 		hr_dev->qp_table.bank[i].next = hr_dev->qp_table.bank[i].min;
1441 	}
1442 
1443 	return 0;
1444 }
1445 
1446 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1447 {
1448 	int i;
1449 
1450 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++)
1451 		ida_destroy(&hr_dev->qp_table.bank[i].ida);
1452 }
1453