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 (refcount_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 	refcount_inc(&hr_qp->refcount);
79 	queue_work(hr_dev->irq_workq, &flush_work->work);
80 }
81 
82 void flush_cqe(struct hns_roce_dev *dev, struct hns_roce_qp *qp)
83 {
84 	/*
85 	 * Hip08 hardware cannot flush the WQEs in SQ/RQ if the QP state
86 	 * gets into errored mode. Hence, as a workaround to this
87 	 * hardware limitation, driver needs to assist in flushing. But
88 	 * the flushing operation uses mailbox to convey the QP state to
89 	 * the hardware and which can sleep due to the mutex protection
90 	 * around the mailbox calls. Hence, use the deferred flush for
91 	 * now.
92 	 */
93 	if (!test_and_set_bit(HNS_ROCE_FLUSH_FLAG, &qp->flush_flag))
94 		init_flush_work(dev, qp);
95 }
96 
97 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
98 {
99 	struct device *dev = hr_dev->dev;
100 	struct hns_roce_qp *qp;
101 
102 	xa_lock(&hr_dev->qp_table_xa);
103 	qp = __hns_roce_qp_lookup(hr_dev, qpn);
104 	if (qp)
105 		refcount_inc(&qp->refcount);
106 	xa_unlock(&hr_dev->qp_table_xa);
107 
108 	if (!qp) {
109 		dev_warn(dev, "Async event for bogus QP %08x\n", qpn);
110 		return;
111 	}
112 
113 	if (hr_dev->hw_rev != HNS_ROCE_HW_VER1 &&
114 	    (event_type == HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR ||
115 	     event_type == HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR ||
116 	     event_type == HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR ||
117 	     event_type == HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION ||
118 	     event_type == HNS_ROCE_EVENT_TYPE_INVALID_XRCETH)) {
119 		qp->state = IB_QPS_ERR;
120 
121 		flush_cqe(hr_dev, qp);
122 	}
123 
124 	qp->event(qp, (enum hns_roce_event)event_type);
125 
126 	if (refcount_dec_and_test(&qp->refcount))
127 		complete(&qp->free);
128 }
129 
130 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
131 				 enum hns_roce_event type)
132 {
133 	struct ib_qp *ibqp = &hr_qp->ibqp;
134 	struct ib_event event;
135 
136 	if (ibqp->event_handler) {
137 		event.device = ibqp->device;
138 		event.element.qp = ibqp;
139 		switch (type) {
140 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
141 			event.event = IB_EVENT_PATH_MIG;
142 			break;
143 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
144 			event.event = IB_EVENT_COMM_EST;
145 			break;
146 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
147 			event.event = IB_EVENT_SQ_DRAINED;
148 			break;
149 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
150 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
151 			break;
152 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
153 			event.event = IB_EVENT_QP_FATAL;
154 			break;
155 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
156 			event.event = IB_EVENT_PATH_MIG_ERR;
157 			break;
158 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
159 			event.event = IB_EVENT_QP_REQ_ERR;
160 			break;
161 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
162 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
163 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
164 			event.event = IB_EVENT_QP_ACCESS_ERR;
165 			break;
166 		default:
167 			dev_dbg(ibqp->device->dev.parent, "roce_ib: Unexpected event type %d on QP %06lx\n",
168 				type, hr_qp->qpn);
169 			return;
170 		}
171 		ibqp->event_handler(&event, ibqp->qp_context);
172 	}
173 }
174 
175 static u8 get_least_load_bankid_for_qp(struct hns_roce_bank *bank)
176 {
177 	u32 least_load = bank[0].inuse;
178 	u8 bankid = 0;
179 	u32 bankcnt;
180 	u8 i;
181 
182 	for (i = 1; i < HNS_ROCE_QP_BANK_NUM; i++) {
183 		bankcnt = bank[i].inuse;
184 		if (bankcnt < least_load) {
185 			least_load = bankcnt;
186 			bankid = i;
187 		}
188 	}
189 
190 	return bankid;
191 }
192 
193 static int alloc_qpn_with_bankid(struct hns_roce_bank *bank, u8 bankid,
194 				 unsigned long *qpn)
195 {
196 	int id;
197 
198 	id = ida_alloc_range(&bank->ida, bank->next, bank->max, GFP_KERNEL);
199 	if (id < 0) {
200 		id = ida_alloc_range(&bank->ida, bank->min, bank->max,
201 				     GFP_KERNEL);
202 		if (id < 0)
203 			return id;
204 	}
205 
206 	/* the QPN should keep increasing until the max value is reached. */
207 	bank->next = (id + 1) > bank->max ? bank->min : id + 1;
208 
209 	/* the lower 3 bits is bankid */
210 	*qpn = (id << 3) | bankid;
211 
212 	return 0;
213 }
214 static int alloc_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
215 {
216 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
217 	unsigned long num = 0;
218 	u8 bankid;
219 	int ret;
220 
221 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI) {
222 		/* when hw version is v1, the sqpn is allocated */
223 		if (hr_dev->hw_rev == HNS_ROCE_HW_VER1)
224 			num = HNS_ROCE_MAX_PORTS +
225 			      hr_dev->iboe.phy_port[hr_qp->port];
226 		else
227 			num = 1;
228 
229 		hr_qp->doorbell_qpn = 1;
230 	} else {
231 		mutex_lock(&qp_table->bank_mutex);
232 		bankid = get_least_load_bankid_for_qp(qp_table->bank);
233 
234 		ret = alloc_qpn_with_bankid(&qp_table->bank[bankid], bankid,
235 					    &num);
236 		if (ret) {
237 			ibdev_err(&hr_dev->ib_dev,
238 				  "failed to alloc QPN, ret = %d\n", ret);
239 			mutex_unlock(&qp_table->bank_mutex);
240 			return ret;
241 		}
242 
243 		qp_table->bank[bankid].inuse++;
244 		mutex_unlock(&qp_table->bank_mutex);
245 
246 		hr_qp->doorbell_qpn = (u32)num;
247 	}
248 
249 	hr_qp->qpn = num;
250 
251 	return 0;
252 }
253 
254 enum hns_roce_qp_state to_hns_roce_state(enum ib_qp_state state)
255 {
256 	switch (state) {
257 	case IB_QPS_RESET:
258 		return HNS_ROCE_QP_STATE_RST;
259 	case IB_QPS_INIT:
260 		return HNS_ROCE_QP_STATE_INIT;
261 	case IB_QPS_RTR:
262 		return HNS_ROCE_QP_STATE_RTR;
263 	case IB_QPS_RTS:
264 		return HNS_ROCE_QP_STATE_RTS;
265 	case IB_QPS_SQD:
266 		return HNS_ROCE_QP_STATE_SQD;
267 	case IB_QPS_ERR:
268 		return HNS_ROCE_QP_STATE_ERR;
269 	default:
270 		return HNS_ROCE_QP_NUM_STATE;
271 	}
272 }
273 
274 static void add_qp_to_list(struct hns_roce_dev *hr_dev,
275 			   struct hns_roce_qp *hr_qp,
276 			   struct ib_cq *send_cq, struct ib_cq *recv_cq)
277 {
278 	struct hns_roce_cq *hr_send_cq, *hr_recv_cq;
279 	unsigned long flags;
280 
281 	hr_send_cq = send_cq ? to_hr_cq(send_cq) : NULL;
282 	hr_recv_cq = recv_cq ? to_hr_cq(recv_cq) : NULL;
283 
284 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
285 	hns_roce_lock_cqs(hr_send_cq, hr_recv_cq);
286 
287 	list_add_tail(&hr_qp->node, &hr_dev->qp_list);
288 	if (hr_send_cq)
289 		list_add_tail(&hr_qp->sq_node, &hr_send_cq->sq_list);
290 	if (hr_recv_cq)
291 		list_add_tail(&hr_qp->rq_node, &hr_recv_cq->rq_list);
292 
293 	hns_roce_unlock_cqs(hr_send_cq, hr_recv_cq);
294 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
295 }
296 
297 static int hns_roce_qp_store(struct hns_roce_dev *hr_dev,
298 			     struct hns_roce_qp *hr_qp,
299 			     struct ib_qp_init_attr *init_attr)
300 {
301 	struct xarray *xa = &hr_dev->qp_table_xa;
302 	int ret;
303 
304 	if (!hr_qp->qpn)
305 		return -EINVAL;
306 
307 	ret = xa_err(xa_store_irq(xa, hr_qp->qpn, hr_qp, GFP_KERNEL));
308 	if (ret)
309 		dev_err(hr_dev->dev, "Failed to xa store for QPC\n");
310 	else
311 		/* add QP to device's QP list for softwc */
312 		add_qp_to_list(hr_dev, hr_qp, init_attr->send_cq,
313 			       init_attr->recv_cq);
314 
315 	return ret;
316 }
317 
318 static int alloc_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
319 {
320 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
321 	struct device *dev = hr_dev->dev;
322 	int ret;
323 
324 	if (!hr_qp->qpn)
325 		return -EINVAL;
326 
327 	/* In v1 engine, GSI QP context is saved in the RoCE hw's register */
328 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI &&
329 	    hr_dev->hw_rev == HNS_ROCE_HW_VER1)
330 		return 0;
331 
332 	/* Alloc memory for QPC */
333 	ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
334 	if (ret) {
335 		dev_err(dev, "Failed to get QPC table\n");
336 		goto err_out;
337 	}
338 
339 	/* Alloc memory for IRRL */
340 	ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
341 	if (ret) {
342 		dev_err(dev, "Failed to get IRRL table\n");
343 		goto err_put_qp;
344 	}
345 
346 	if (hr_dev->caps.trrl_entry_sz) {
347 		/* Alloc memory for TRRL */
348 		ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
349 					 hr_qp->qpn);
350 		if (ret) {
351 			dev_err(dev, "Failed to get TRRL table\n");
352 			goto err_put_irrl;
353 		}
354 	}
355 
356 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
357 		/* Alloc memory for SCC CTX */
358 		ret = hns_roce_table_get(hr_dev, &qp_table->sccc_table,
359 					 hr_qp->qpn);
360 		if (ret) {
361 			dev_err(dev, "Failed to get SCC CTX table\n");
362 			goto err_put_trrl;
363 		}
364 	}
365 
366 	return 0;
367 
368 err_put_trrl:
369 	if (hr_dev->caps.trrl_entry_sz)
370 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
371 
372 err_put_irrl:
373 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
374 
375 err_put_qp:
376 	hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
377 
378 err_out:
379 	return ret;
380 }
381 
382 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
383 {
384 	struct xarray *xa = &hr_dev->qp_table_xa;
385 	unsigned long flags;
386 
387 	list_del(&hr_qp->node);
388 
389 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
390 		list_del(&hr_qp->sq_node);
391 
392 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_INI &&
393 	    hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
394 		list_del(&hr_qp->rq_node);
395 
396 	xa_lock_irqsave(xa, flags);
397 	__xa_erase(xa, hr_qp->qpn);
398 	xa_unlock_irqrestore(xa, flags);
399 }
400 
401 static void free_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
402 {
403 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
404 
405 	/* In v1 engine, GSI QP context is saved in the RoCE hw's register */
406 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI &&
407 	    hr_dev->hw_rev == HNS_ROCE_HW_VER1)
408 		return;
409 
410 	if (hr_dev->caps.trrl_entry_sz)
411 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
412 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
413 }
414 
415 static inline u8 get_qp_bankid(unsigned long qpn)
416 {
417 	/* The lower 3 bits of QPN are used to hash to different banks */
418 	return (u8)(qpn & GENMASK(2, 0));
419 }
420 
421 static void free_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
422 {
423 	u8 bankid;
424 
425 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI)
426 		return;
427 
428 	if (hr_qp->qpn < hr_dev->caps.reserved_qps)
429 		return;
430 
431 	bankid = get_qp_bankid(hr_qp->qpn);
432 
433 	ida_free(&hr_dev->qp_table.bank[bankid].ida, hr_qp->qpn >> 3);
434 
435 	mutex_lock(&hr_dev->qp_table.bank_mutex);
436 	hr_dev->qp_table.bank[bankid].inuse--;
437 	mutex_unlock(&hr_dev->qp_table.bank_mutex);
438 }
439 
440 static u32 proc_rq_sge(struct hns_roce_dev *dev, struct hns_roce_qp *hr_qp,
441 		       bool user)
442 {
443 	u32 max_sge = dev->caps.max_rq_sg;
444 
445 	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
446 		return max_sge;
447 
448 	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
449 	 * calculate number of max_sge with reserved SGEs when allocating wqe
450 	 * buf, so there is no need to do this again in kernel. But the number
451 	 * may exceed the capacity of SGEs recorded in the firmware, so the
452 	 * kernel driver should just adapt the value accordingly.
453 	 */
454 	if (user)
455 		max_sge = roundup_pow_of_two(max_sge + 1);
456 	else
457 		hr_qp->rq.rsv_sge = 1;
458 
459 	return max_sge;
460 }
461 
462 static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
463 		       struct hns_roce_qp *hr_qp, int has_rq, bool user)
464 {
465 	u32 max_sge = proc_rq_sge(hr_dev, hr_qp, user);
466 	u32 cnt;
467 
468 	/* If srq exist, set zero for relative number of rq */
469 	if (!has_rq) {
470 		hr_qp->rq.wqe_cnt = 0;
471 		hr_qp->rq.max_gs = 0;
472 		hr_qp->rq_inl_buf.wqe_cnt = 0;
473 		cap->max_recv_wr = 0;
474 		cap->max_recv_sge = 0;
475 
476 		return 0;
477 	}
478 
479 	/* Check the validity of QP support capacity */
480 	if (!cap->max_recv_wr || cap->max_recv_wr > hr_dev->caps.max_wqes ||
481 	    cap->max_recv_sge > max_sge) {
482 		ibdev_err(&hr_dev->ib_dev,
483 			  "RQ config error, depth = %u, sge = %u\n",
484 			  cap->max_recv_wr, cap->max_recv_sge);
485 		return -EINVAL;
486 	}
487 
488 	cnt = roundup_pow_of_two(max(cap->max_recv_wr, hr_dev->caps.min_wqes));
489 	if (cnt > hr_dev->caps.max_wqes) {
490 		ibdev_err(&hr_dev->ib_dev, "rq depth %u too large\n",
491 			  cap->max_recv_wr);
492 		return -EINVAL;
493 	}
494 
495 	hr_qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge) +
496 					      hr_qp->rq.rsv_sge);
497 
498 	if (hr_dev->caps.max_rq_sg <= HNS_ROCE_SGE_IN_WQE)
499 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz);
500 	else
501 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz *
502 					    hr_qp->rq.max_gs);
503 
504 	hr_qp->rq.wqe_cnt = cnt;
505 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE &&
506 	    hr_qp->ibqp.qp_type != IB_QPT_UD &&
507 	    hr_qp->ibqp.qp_type != IB_QPT_GSI)
508 		hr_qp->rq_inl_buf.wqe_cnt = cnt;
509 	else
510 		hr_qp->rq_inl_buf.wqe_cnt = 0;
511 
512 	cap->max_recv_wr = cnt;
513 	cap->max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
514 
515 	return 0;
516 }
517 
518 static u32 get_wqe_ext_sge_cnt(struct hns_roce_qp *qp)
519 {
520 	/* GSI/UD QP only has extended sge */
521 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_UD)
522 		return qp->sq.max_gs;
523 
524 	if (qp->sq.max_gs > HNS_ROCE_SGE_IN_WQE)
525 		return qp->sq.max_gs - HNS_ROCE_SGE_IN_WQE;
526 
527 	return 0;
528 }
529 
530 static void set_ext_sge_param(struct hns_roce_dev *hr_dev, u32 sq_wqe_cnt,
531 			      struct hns_roce_qp *hr_qp, struct ib_qp_cap *cap)
532 {
533 	u32 total_sge_cnt;
534 	u32 wqe_sge_cnt;
535 
536 	hr_qp->sge.sge_shift = HNS_ROCE_SGE_SHIFT;
537 
538 	if (hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
539 		hr_qp->sq.max_gs = HNS_ROCE_SGE_IN_WQE;
540 		return;
541 	}
542 
543 	hr_qp->sq.max_gs = max(1U, cap->max_send_sge);
544 
545 	wqe_sge_cnt = get_wqe_ext_sge_cnt(hr_qp);
546 
547 	/* If the number of extended sge is not zero, they MUST use the
548 	 * space of HNS_HW_PAGE_SIZE at least.
549 	 */
550 	if (wqe_sge_cnt) {
551 		total_sge_cnt = roundup_pow_of_two(sq_wqe_cnt * wqe_sge_cnt);
552 		hr_qp->sge.sge_cnt = max(total_sge_cnt,
553 				(u32)HNS_HW_PAGE_SIZE / HNS_ROCE_SGE_SIZE);
554 	}
555 }
556 
557 static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
558 					struct ib_qp_cap *cap,
559 					struct hns_roce_ib_create_qp *ucmd)
560 {
561 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
562 	u8 max_sq_stride = ilog2(roundup_sq_stride);
563 
564 	/* Sanity check SQ size before proceeding */
565 	if (ucmd->log_sq_stride > max_sq_stride ||
566 	    ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
567 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ stride size.\n");
568 		return -EINVAL;
569 	}
570 
571 	if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
572 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ SGE size %u.\n",
573 			  cap->max_send_sge);
574 		return -EINVAL;
575 	}
576 
577 	return 0;
578 }
579 
580 static int set_user_sq_size(struct hns_roce_dev *hr_dev,
581 			    struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp,
582 			    struct hns_roce_ib_create_qp *ucmd)
583 {
584 	struct ib_device *ibdev = &hr_dev->ib_dev;
585 	u32 cnt = 0;
586 	int ret;
587 
588 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
589 	    cnt > hr_dev->caps.max_wqes)
590 		return -EINVAL;
591 
592 	ret = check_sq_size_with_integrity(hr_dev, cap, ucmd);
593 	if (ret) {
594 		ibdev_err(ibdev, "failed to check user SQ size, ret = %d.\n",
595 			  ret);
596 		return ret;
597 	}
598 
599 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
600 
601 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
602 	hr_qp->sq.wqe_cnt = cnt;
603 
604 	return 0;
605 }
606 
607 static int set_wqe_buf_attr(struct hns_roce_dev *hr_dev,
608 			    struct hns_roce_qp *hr_qp,
609 			    struct hns_roce_buf_attr *buf_attr)
610 {
611 	int buf_size;
612 	int idx = 0;
613 
614 	hr_qp->buff_size = 0;
615 
616 	/* SQ WQE */
617 	hr_qp->sq.offset = 0;
618 	buf_size = to_hr_hem_entries_size(hr_qp->sq.wqe_cnt,
619 					  hr_qp->sq.wqe_shift);
620 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
621 		buf_attr->region[idx].size = buf_size;
622 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sq_hop_num;
623 		idx++;
624 		hr_qp->buff_size += buf_size;
625 	}
626 
627 	/* extend SGE WQE in SQ */
628 	hr_qp->sge.offset = hr_qp->buff_size;
629 	buf_size = to_hr_hem_entries_size(hr_qp->sge.sge_cnt,
630 					  hr_qp->sge.sge_shift);
631 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
632 		buf_attr->region[idx].size = buf_size;
633 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sge_hop_num;
634 		idx++;
635 		hr_qp->buff_size += buf_size;
636 	}
637 
638 	/* RQ WQE */
639 	hr_qp->rq.offset = hr_qp->buff_size;
640 	buf_size = to_hr_hem_entries_size(hr_qp->rq.wqe_cnt,
641 					  hr_qp->rq.wqe_shift);
642 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
643 		buf_attr->region[idx].size = buf_size;
644 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_rq_hop_num;
645 		idx++;
646 		hr_qp->buff_size += buf_size;
647 	}
648 
649 	if (hr_qp->buff_size < 1)
650 		return -EINVAL;
651 
652 	buf_attr->page_shift = HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
653 	buf_attr->region_count = idx;
654 
655 	return 0;
656 }
657 
658 static int set_kernel_sq_size(struct hns_roce_dev *hr_dev,
659 			      struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp)
660 {
661 	struct ib_device *ibdev = &hr_dev->ib_dev;
662 	u32 cnt;
663 
664 	if (!cap->max_send_wr || cap->max_send_wr > hr_dev->caps.max_wqes ||
665 	    cap->max_send_sge > hr_dev->caps.max_sq_sg) {
666 		ibdev_err(ibdev, "failed to check SQ WR or SGE num.\n");
667 		return -EINVAL;
668 	}
669 
670 	cnt = roundup_pow_of_two(max(cap->max_send_wr, hr_dev->caps.min_wqes));
671 	if (cnt > hr_dev->caps.max_wqes) {
672 		ibdev_err(ibdev, "failed to check WQE num, WQE num = %u.\n",
673 			  cnt);
674 		return -EINVAL;
675 	}
676 
677 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
678 	hr_qp->sq.wqe_cnt = cnt;
679 
680 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
681 
682 	/* sync the parameters of kernel QP to user's configuration */
683 	cap->max_send_wr = cnt;
684 	cap->max_send_sge = hr_qp->sq.max_gs;
685 
686 	return 0;
687 }
688 
689 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
690 {
691 	if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
692 		return 0;
693 
694 	return 1;
695 }
696 
697 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
698 {
699 	if (attr->qp_type == IB_QPT_XRC_INI ||
700 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
701 	    !attr->cap.max_recv_wr)
702 		return 0;
703 
704 	return 1;
705 }
706 
707 static int alloc_rq_inline_buf(struct hns_roce_qp *hr_qp,
708 			       struct ib_qp_init_attr *init_attr)
709 {
710 	u32 max_recv_sge = init_attr->cap.max_recv_sge;
711 	u32 wqe_cnt = hr_qp->rq_inl_buf.wqe_cnt;
712 	struct hns_roce_rinl_wqe *wqe_list;
713 	int i;
714 
715 	/* allocate recv inline buf */
716 	wqe_list = kcalloc(wqe_cnt, sizeof(struct hns_roce_rinl_wqe),
717 			   GFP_KERNEL);
718 	if (!wqe_list)
719 		goto err;
720 
721 	/* Allocate a continuous buffer for all inline sge we need */
722 	wqe_list[0].sg_list = kcalloc(wqe_cnt, (max_recv_sge *
723 				      sizeof(struct hns_roce_rinl_sge)),
724 				      GFP_KERNEL);
725 	if (!wqe_list[0].sg_list)
726 		goto err_wqe_list;
727 
728 	/* Assign buffers of sg_list to each inline wqe */
729 	for (i = 1; i < wqe_cnt; i++)
730 		wqe_list[i].sg_list = &wqe_list[0].sg_list[i * max_recv_sge];
731 
732 	hr_qp->rq_inl_buf.wqe_list = wqe_list;
733 
734 	return 0;
735 
736 err_wqe_list:
737 	kfree(wqe_list);
738 
739 err:
740 	return -ENOMEM;
741 }
742 
743 static void free_rq_inline_buf(struct hns_roce_qp *hr_qp)
744 {
745 	if (hr_qp->rq_inl_buf.wqe_list)
746 		kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
747 	kfree(hr_qp->rq_inl_buf.wqe_list);
748 }
749 
750 static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
751 			struct ib_qp_init_attr *init_attr,
752 			struct ib_udata *udata, unsigned long addr)
753 {
754 	struct ib_device *ibdev = &hr_dev->ib_dev;
755 	struct hns_roce_buf_attr buf_attr = {};
756 	int ret;
757 
758 	if (!udata && hr_qp->rq_inl_buf.wqe_cnt) {
759 		ret = alloc_rq_inline_buf(hr_qp, init_attr);
760 		if (ret) {
761 			ibdev_err(ibdev,
762 				  "failed to alloc inline buf, ret = %d.\n",
763 				  ret);
764 			return ret;
765 		}
766 	} else {
767 		hr_qp->rq_inl_buf.wqe_list = NULL;
768 	}
769 
770 	ret = set_wqe_buf_attr(hr_dev, hr_qp, &buf_attr);
771 	if (ret) {
772 		ibdev_err(ibdev, "failed to split WQE buf, ret = %d.\n", ret);
773 		goto err_inline;
774 	}
775 	ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, &buf_attr,
776 				  PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
777 				  udata, addr);
778 	if (ret) {
779 		ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret);
780 		goto err_inline;
781 	}
782 
783 	return 0;
784 err_inline:
785 	free_rq_inline_buf(hr_qp);
786 
787 	return ret;
788 }
789 
790 static void free_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
791 {
792 	hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
793 	free_rq_inline_buf(hr_qp);
794 }
795 
796 static inline bool user_qp_has_sdb(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 				   struct hns_roce_ib_create_qp *ucmd)
801 {
802 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
803 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
804 		hns_roce_qp_has_sq(init_attr) &&
805 		udata->inlen >= offsetofend(typeof(*ucmd), sdb_addr));
806 }
807 
808 static inline bool user_qp_has_rdb(struct hns_roce_dev *hr_dev,
809 				   struct ib_qp_init_attr *init_attr,
810 				   struct ib_udata *udata,
811 				   struct hns_roce_ib_create_qp_resp *resp)
812 {
813 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
814 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
815 		hns_roce_qp_has_rq(init_attr));
816 }
817 
818 static inline bool kernel_qp_has_rdb(struct hns_roce_dev *hr_dev,
819 				     struct ib_qp_init_attr *init_attr)
820 {
821 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
822 		hns_roce_qp_has_rq(init_attr));
823 }
824 
825 static int alloc_user_qp_db(struct hns_roce_dev *hr_dev,
826 			    struct hns_roce_qp *hr_qp,
827 			    struct ib_qp_init_attr *init_attr,
828 			    struct ib_udata *udata,
829 			    struct hns_roce_ib_create_qp *ucmd,
830 			    struct hns_roce_ib_create_qp_resp *resp)
831 {
832 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(udata,
833 		struct hns_roce_ucontext, ibucontext);
834 	struct ib_device *ibdev = &hr_dev->ib_dev;
835 	int ret;
836 
837 	if (user_qp_has_sdb(hr_dev, init_attr, udata, resp, ucmd)) {
838 		ret = hns_roce_db_map_user(uctx, ucmd->sdb_addr, &hr_qp->sdb);
839 		if (ret) {
840 			ibdev_err(ibdev,
841 				  "failed to map user SQ doorbell, ret = %d.\n",
842 				  ret);
843 			goto err_out;
844 		}
845 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
846 	}
847 
848 	if (user_qp_has_rdb(hr_dev, init_attr, udata, resp)) {
849 		ret = hns_roce_db_map_user(uctx, ucmd->db_addr, &hr_qp->rdb);
850 		if (ret) {
851 			ibdev_err(ibdev,
852 				  "failed to map user RQ doorbell, ret = %d.\n",
853 				  ret);
854 			goto err_sdb;
855 		}
856 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
857 	}
858 
859 	return 0;
860 
861 err_sdb:
862 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
863 		hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
864 err_out:
865 	return ret;
866 }
867 
868 static int alloc_kernel_qp_db(struct hns_roce_dev *hr_dev,
869 			      struct hns_roce_qp *hr_qp,
870 			      struct ib_qp_init_attr *init_attr)
871 {
872 	struct ib_device *ibdev = &hr_dev->ib_dev;
873 	int ret;
874 
875 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
876 		hr_qp->sq.db_reg = hr_dev->mem_base +
877 				   HNS_ROCE_DWQE_SIZE * hr_qp->qpn;
878 	else
879 		hr_qp->sq.db_reg = hr_dev->reg_base + hr_dev->sdb_offset +
880 				   DB_REG_OFFSET * hr_dev->priv_uar.index;
881 
882 	hr_qp->rq.db_reg = hr_dev->reg_base + hr_dev->odb_offset +
883 			   DB_REG_OFFSET * hr_dev->priv_uar.index;
884 
885 	if (kernel_qp_has_rdb(hr_dev, init_attr)) {
886 		ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
887 		if (ret) {
888 			ibdev_err(ibdev,
889 				  "failed to alloc kernel RQ doorbell, ret = %d.\n",
890 				  ret);
891 			return ret;
892 		}
893 		*hr_qp->rdb.db_record = 0;
894 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
895 	}
896 
897 	return 0;
898 }
899 
900 static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
901 		       struct ib_qp_init_attr *init_attr,
902 		       struct ib_udata *udata,
903 		       struct hns_roce_ib_create_qp *ucmd,
904 		       struct hns_roce_ib_create_qp_resp *resp)
905 {
906 	int ret;
907 
908 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SDI_MODE)
909 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_OWNER_DB;
910 
911 	if (udata) {
912 		ret = alloc_user_qp_db(hr_dev, hr_qp, init_attr, udata, ucmd,
913 				       resp);
914 		if (ret)
915 			return ret;
916 	} else {
917 		ret = alloc_kernel_qp_db(hr_dev, hr_qp, init_attr);
918 		if (ret)
919 			return ret;
920 	}
921 
922 	return 0;
923 }
924 
925 static void free_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
926 		       struct ib_udata *udata)
927 {
928 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
929 		udata, struct hns_roce_ucontext, ibucontext);
930 
931 	if (udata) {
932 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
933 			hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
934 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
935 			hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
936 	} else {
937 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
938 			hns_roce_free_db(hr_dev, &hr_qp->rdb);
939 	}
940 }
941 
942 static int alloc_kernel_wrid(struct hns_roce_dev *hr_dev,
943 			     struct hns_roce_qp *hr_qp)
944 {
945 	struct ib_device *ibdev = &hr_dev->ib_dev;
946 	u64 *sq_wrid = NULL;
947 	u64 *rq_wrid = NULL;
948 	int ret;
949 
950 	sq_wrid = kcalloc(hr_qp->sq.wqe_cnt, sizeof(u64), GFP_KERNEL);
951 	if (ZERO_OR_NULL_PTR(sq_wrid)) {
952 		ibdev_err(ibdev, "failed to alloc SQ wrid.\n");
953 		return -ENOMEM;
954 	}
955 
956 	if (hr_qp->rq.wqe_cnt) {
957 		rq_wrid = kcalloc(hr_qp->rq.wqe_cnt, sizeof(u64), GFP_KERNEL);
958 		if (ZERO_OR_NULL_PTR(rq_wrid)) {
959 			ibdev_err(ibdev, "failed to alloc RQ wrid.\n");
960 			ret = -ENOMEM;
961 			goto err_sq;
962 		}
963 	}
964 
965 	hr_qp->sq.wrid = sq_wrid;
966 	hr_qp->rq.wrid = rq_wrid;
967 	return 0;
968 err_sq:
969 	kfree(sq_wrid);
970 
971 	return ret;
972 }
973 
974 static void free_kernel_wrid(struct hns_roce_qp *hr_qp)
975 {
976 	kfree(hr_qp->rq.wrid);
977 	kfree(hr_qp->sq.wrid);
978 }
979 
980 static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
981 			struct ib_qp_init_attr *init_attr,
982 			struct ib_udata *udata,
983 			struct hns_roce_ib_create_qp *ucmd)
984 {
985 	struct ib_device *ibdev = &hr_dev->ib_dev;
986 	int ret;
987 
988 	if (init_attr->cap.max_inline_data > hr_dev->caps.max_sq_inline)
989 		init_attr->cap.max_inline_data = hr_dev->caps.max_sq_inline;
990 
991 	hr_qp->max_inline_data = init_attr->cap.max_inline_data;
992 
993 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
994 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
995 	else
996 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
997 
998 	ret = set_rq_size(hr_dev, &init_attr->cap, hr_qp,
999 			  hns_roce_qp_has_rq(init_attr), !!udata);
1000 	if (ret) {
1001 		ibdev_err(ibdev, "failed to set user RQ size, ret = %d.\n",
1002 			  ret);
1003 		return ret;
1004 	}
1005 
1006 	if (udata) {
1007 		ret = ib_copy_from_udata(ucmd, udata,
1008 					 min(udata->inlen, sizeof(*ucmd)));
1009 		if (ret) {
1010 			ibdev_err(ibdev,
1011 				  "failed to copy QP ucmd, ret = %d\n", ret);
1012 			return ret;
1013 		}
1014 
1015 		ret = set_user_sq_size(hr_dev, &init_attr->cap, hr_qp, ucmd);
1016 		if (ret)
1017 			ibdev_err(ibdev,
1018 				  "failed to set user SQ size, ret = %d.\n",
1019 				  ret);
1020 	} else {
1021 		ret = set_kernel_sq_size(hr_dev, &init_attr->cap, hr_qp);
1022 		if (ret)
1023 			ibdev_err(ibdev,
1024 				  "failed to set kernel SQ size, ret = %d.\n",
1025 				  ret);
1026 	}
1027 
1028 	return ret;
1029 }
1030 
1031 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
1032 				     struct ib_pd *ib_pd,
1033 				     struct ib_qp_init_attr *init_attr,
1034 				     struct ib_udata *udata,
1035 				     struct hns_roce_qp *hr_qp)
1036 {
1037 	struct hns_roce_ib_create_qp_resp resp = {};
1038 	struct ib_device *ibdev = &hr_dev->ib_dev;
1039 	struct hns_roce_ib_create_qp ucmd;
1040 	int ret;
1041 
1042 	mutex_init(&hr_qp->mutex);
1043 	spin_lock_init(&hr_qp->sq.lock);
1044 	spin_lock_init(&hr_qp->rq.lock);
1045 
1046 	hr_qp->state = IB_QPS_RESET;
1047 	hr_qp->flush_flag = 0;
1048 
1049 	if (init_attr->create_flags)
1050 		return -EOPNOTSUPP;
1051 
1052 	ret = set_qp_param(hr_dev, hr_qp, init_attr, udata, &ucmd);
1053 	if (ret) {
1054 		ibdev_err(ibdev, "failed to set QP param, ret = %d.\n", ret);
1055 		return ret;
1056 	}
1057 
1058 	if (!udata) {
1059 		ret = alloc_kernel_wrid(hr_dev, hr_qp);
1060 		if (ret) {
1061 			ibdev_err(ibdev, "failed to alloc wrid, ret = %d.\n",
1062 				  ret);
1063 			return ret;
1064 		}
1065 	}
1066 
1067 	ret = alloc_qp_buf(hr_dev, hr_qp, init_attr, udata, ucmd.buf_addr);
1068 	if (ret) {
1069 		ibdev_err(ibdev, "failed to alloc QP buffer, ret = %d.\n", ret);
1070 		goto err_buf;
1071 	}
1072 
1073 	ret = alloc_qpn(hr_dev, hr_qp);
1074 	if (ret) {
1075 		ibdev_err(ibdev, "failed to alloc QPN, ret = %d.\n", ret);
1076 		goto err_qpn;
1077 	}
1078 
1079 	ret = alloc_qp_db(hr_dev, hr_qp, init_attr, udata, &ucmd, &resp);
1080 	if (ret) {
1081 		ibdev_err(ibdev, "failed to alloc QP doorbell, ret = %d.\n",
1082 			  ret);
1083 		goto err_db;
1084 	}
1085 
1086 	ret = alloc_qpc(hr_dev, hr_qp);
1087 	if (ret) {
1088 		ibdev_err(ibdev, "failed to alloc QP context, ret = %d.\n",
1089 			  ret);
1090 		goto err_qpc;
1091 	}
1092 
1093 	ret = hns_roce_qp_store(hr_dev, hr_qp, init_attr);
1094 	if (ret) {
1095 		ibdev_err(ibdev, "failed to store QP, ret = %d.\n", ret);
1096 		goto err_store;
1097 	}
1098 
1099 	if (udata) {
1100 		resp.cap_flags = hr_qp->en_flags;
1101 		ret = ib_copy_to_udata(udata, &resp,
1102 				       min(udata->outlen, sizeof(resp)));
1103 		if (ret) {
1104 			ibdev_err(ibdev, "copy qp resp failed!\n");
1105 			goto err_store;
1106 		}
1107 	}
1108 
1109 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
1110 		ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
1111 		if (ret)
1112 			goto err_flow_ctrl;
1113 	}
1114 
1115 	hr_qp->ibqp.qp_num = hr_qp->qpn;
1116 	hr_qp->event = hns_roce_ib_qp_event;
1117 	refcount_set(&hr_qp->refcount, 1);
1118 	init_completion(&hr_qp->free);
1119 
1120 	return 0;
1121 
1122 err_flow_ctrl:
1123 	hns_roce_qp_remove(hr_dev, hr_qp);
1124 err_store:
1125 	free_qpc(hr_dev, hr_qp);
1126 err_qpc:
1127 	free_qp_db(hr_dev, hr_qp, udata);
1128 err_db:
1129 	free_qpn(hr_dev, hr_qp);
1130 err_qpn:
1131 	free_qp_buf(hr_dev, hr_qp);
1132 err_buf:
1133 	free_kernel_wrid(hr_qp);
1134 	return ret;
1135 }
1136 
1137 void hns_roce_qp_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1138 			 struct ib_udata *udata)
1139 {
1140 	if (refcount_dec_and_test(&hr_qp->refcount))
1141 		complete(&hr_qp->free);
1142 	wait_for_completion(&hr_qp->free);
1143 
1144 	free_qpc(hr_dev, hr_qp);
1145 	free_qpn(hr_dev, hr_qp);
1146 	free_qp_buf(hr_dev, hr_qp);
1147 	free_kernel_wrid(hr_qp);
1148 	free_qp_db(hr_dev, hr_qp, udata);
1149 }
1150 
1151 static int check_qp_type(struct hns_roce_dev *hr_dev, enum ib_qp_type type,
1152 			 bool is_user)
1153 {
1154 	switch (type) {
1155 	case IB_QPT_XRC_INI:
1156 	case IB_QPT_XRC_TGT:
1157 		if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC))
1158 			goto out;
1159 		break;
1160 	case IB_QPT_UD:
1161 		if (hr_dev->pci_dev->revision <= PCI_REVISION_ID_HIP08 &&
1162 		    is_user)
1163 			goto out;
1164 		break;
1165 	case IB_QPT_RC:
1166 	case IB_QPT_GSI:
1167 		break;
1168 	default:
1169 		goto out;
1170 	}
1171 
1172 	return 0;
1173 
1174 out:
1175 	ibdev_err(&hr_dev->ib_dev, "not support QP type %d\n", type);
1176 
1177 	return -EOPNOTSUPP;
1178 }
1179 
1180 int hns_roce_create_qp(struct ib_qp *qp, struct ib_qp_init_attr *init_attr,
1181 		       struct ib_udata *udata)
1182 {
1183 	struct ib_device *ibdev = qp->device;
1184 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
1185 	struct hns_roce_qp *hr_qp = to_hr_qp(qp);
1186 	struct ib_pd *pd = qp->pd;
1187 	int ret;
1188 
1189 	ret = check_qp_type(hr_dev, init_attr->qp_type, !!udata);
1190 	if (ret)
1191 		return ret;
1192 
1193 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1194 		hr_qp->xrcdn = to_hr_xrcd(init_attr->xrcd)->xrcdn;
1195 
1196 	if (init_attr->qp_type == IB_QPT_GSI) {
1197 		hr_qp->port = init_attr->port_num - 1;
1198 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
1199 	}
1200 
1201 	ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, hr_qp);
1202 	if (ret)
1203 		ibdev_err(ibdev, "Create QP type 0x%x failed(%d)\n",
1204 			  init_attr->qp_type, ret);
1205 
1206 	return ret;
1207 }
1208 
1209 int to_hr_qp_type(int qp_type)
1210 {
1211 	switch (qp_type) {
1212 	case IB_QPT_RC:
1213 		return SERV_TYPE_RC;
1214 	case IB_QPT_UD:
1215 	case IB_QPT_GSI:
1216 		return SERV_TYPE_UD;
1217 	case IB_QPT_XRC_INI:
1218 	case IB_QPT_XRC_TGT:
1219 		return SERV_TYPE_XRC;
1220 	default:
1221 		return -1;
1222 	}
1223 }
1224 
1225 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1226 			      struct hns_roce_qp *hr_qp,
1227 			      struct ib_qp_attr *attr, int attr_mask)
1228 {
1229 	enum ib_mtu active_mtu;
1230 	int p;
1231 
1232 	p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1233 	active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1234 
1235 	if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1236 	    attr->path_mtu > hr_dev->caps.max_mtu) ||
1237 	    attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1238 		ibdev_err(&hr_dev->ib_dev,
1239 			"attr path_mtu(%d)invalid while modify qp",
1240 			attr->path_mtu);
1241 		return -EINVAL;
1242 	}
1243 
1244 	return 0;
1245 }
1246 
1247 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1248 				  int attr_mask)
1249 {
1250 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1251 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1252 	int p;
1253 
1254 	if ((attr_mask & IB_QP_PORT) &&
1255 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1256 		ibdev_err(&hr_dev->ib_dev, "invalid attr, port_num = %u.\n",
1257 			  attr->port_num);
1258 		return -EINVAL;
1259 	}
1260 
1261 	if (attr_mask & IB_QP_PKEY_INDEX) {
1262 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1263 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1264 			ibdev_err(&hr_dev->ib_dev,
1265 				  "invalid attr, pkey_index = %u.\n",
1266 				  attr->pkey_index);
1267 			return -EINVAL;
1268 		}
1269 	}
1270 
1271 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1272 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1273 		ibdev_err(&hr_dev->ib_dev,
1274 			  "invalid attr, max_rd_atomic = %u.\n",
1275 			  attr->max_rd_atomic);
1276 		return -EINVAL;
1277 	}
1278 
1279 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1280 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1281 		ibdev_err(&hr_dev->ib_dev,
1282 			  "invalid attr, max_dest_rd_atomic = %u.\n",
1283 			  attr->max_dest_rd_atomic);
1284 		return -EINVAL;
1285 	}
1286 
1287 	if (attr_mask & IB_QP_PATH_MTU)
1288 		return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1289 
1290 	return 0;
1291 }
1292 
1293 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1294 		       int attr_mask, struct ib_udata *udata)
1295 {
1296 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1297 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1298 	enum ib_qp_state cur_state, new_state;
1299 	int ret = -EINVAL;
1300 
1301 	mutex_lock(&hr_qp->mutex);
1302 
1303 	if (attr_mask & IB_QP_CUR_STATE && attr->cur_qp_state != hr_qp->state)
1304 		goto out;
1305 
1306 	cur_state = hr_qp->state;
1307 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1308 
1309 	if (ibqp->uobject &&
1310 	    (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1311 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB) {
1312 			hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1313 
1314 			if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
1315 				hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1316 		} else {
1317 			ibdev_warn(&hr_dev->ib_dev,
1318 				  "flush cqe is not supported in userspace!\n");
1319 			goto out;
1320 		}
1321 	}
1322 
1323 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1324 				attr_mask)) {
1325 		ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1326 		goto out;
1327 	}
1328 
1329 	ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1330 	if (ret)
1331 		goto out;
1332 
1333 	if (cur_state == new_state && cur_state == IB_QPS_RESET)
1334 		goto out;
1335 
1336 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1337 				    new_state);
1338 
1339 out:
1340 	mutex_unlock(&hr_qp->mutex);
1341 
1342 	return ret;
1343 }
1344 
1345 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1346 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1347 {
1348 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1349 		__acquire(&send_cq->lock);
1350 		__acquire(&recv_cq->lock);
1351 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1352 		spin_lock_irq(&send_cq->lock);
1353 		__acquire(&recv_cq->lock);
1354 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1355 		spin_lock_irq(&recv_cq->lock);
1356 		__acquire(&send_cq->lock);
1357 	} else if (send_cq == recv_cq) {
1358 		spin_lock_irq(&send_cq->lock);
1359 		__acquire(&recv_cq->lock);
1360 	} else if (send_cq->cqn < recv_cq->cqn) {
1361 		spin_lock_irq(&send_cq->lock);
1362 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1363 	} else {
1364 		spin_lock_irq(&recv_cq->lock);
1365 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1366 	}
1367 }
1368 
1369 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1370 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1371 			 __releases(&recv_cq->lock)
1372 {
1373 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1374 		__release(&recv_cq->lock);
1375 		__release(&send_cq->lock);
1376 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1377 		__release(&recv_cq->lock);
1378 		spin_unlock(&send_cq->lock);
1379 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1380 		__release(&send_cq->lock);
1381 		spin_unlock(&recv_cq->lock);
1382 	} else if (send_cq == recv_cq) {
1383 		__release(&recv_cq->lock);
1384 		spin_unlock_irq(&send_cq->lock);
1385 	} else if (send_cq->cqn < recv_cq->cqn) {
1386 		spin_unlock(&recv_cq->lock);
1387 		spin_unlock_irq(&send_cq->lock);
1388 	} else {
1389 		spin_unlock(&send_cq->lock);
1390 		spin_unlock_irq(&recv_cq->lock);
1391 	}
1392 }
1393 
1394 static inline void *get_wqe(struct hns_roce_qp *hr_qp, int offset)
1395 {
1396 	return hns_roce_buf_offset(hr_qp->mtr.kmem, offset);
1397 }
1398 
1399 void *hns_roce_get_recv_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1400 {
1401 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1402 }
1403 
1404 void *hns_roce_get_send_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1405 {
1406 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1407 }
1408 
1409 void *hns_roce_get_extend_sge(struct hns_roce_qp *hr_qp, unsigned int n)
1410 {
1411 	return get_wqe(hr_qp, hr_qp->sge.offset + (n << hr_qp->sge.sge_shift));
1412 }
1413 
1414 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, u32 nreq,
1415 			  struct ib_cq *ib_cq)
1416 {
1417 	struct hns_roce_cq *hr_cq;
1418 	u32 cur;
1419 
1420 	cur = hr_wq->head - hr_wq->tail;
1421 	if (likely(cur + nreq < hr_wq->wqe_cnt))
1422 		return false;
1423 
1424 	hr_cq = to_hr_cq(ib_cq);
1425 	spin_lock(&hr_cq->lock);
1426 	cur = hr_wq->head - hr_wq->tail;
1427 	spin_unlock(&hr_cq->lock);
1428 
1429 	return cur + nreq >= hr_wq->wqe_cnt;
1430 }
1431 
1432 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1433 {
1434 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1435 	unsigned int reserved_from_bot;
1436 	unsigned int i;
1437 
1438 	qp_table->idx_table.spare_idx = kcalloc(hr_dev->caps.num_qps,
1439 					sizeof(u32), GFP_KERNEL);
1440 	if (!qp_table->idx_table.spare_idx)
1441 		return -ENOMEM;
1442 
1443 	mutex_init(&qp_table->scc_mutex);
1444 	mutex_init(&qp_table->bank_mutex);
1445 	xa_init(&hr_dev->qp_table_xa);
1446 
1447 	reserved_from_bot = hr_dev->caps.reserved_qps;
1448 
1449 	for (i = 0; i < reserved_from_bot; i++) {
1450 		hr_dev->qp_table.bank[get_qp_bankid(i)].inuse++;
1451 		hr_dev->qp_table.bank[get_qp_bankid(i)].min++;
1452 	}
1453 
1454 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) {
1455 		ida_init(&hr_dev->qp_table.bank[i].ida);
1456 		hr_dev->qp_table.bank[i].max = hr_dev->caps.num_qps /
1457 					       HNS_ROCE_QP_BANK_NUM - 1;
1458 		hr_dev->qp_table.bank[i].next = hr_dev->qp_table.bank[i].min;
1459 	}
1460 
1461 	return 0;
1462 }
1463 
1464 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1465 {
1466 	int i;
1467 
1468 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++)
1469 		ida_destroy(&hr_dev->qp_table.bank[i].ida);
1470 	kfree(hr_dev->qp_table.idx_table.spare_idx);
1471 }
1472