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 <rdma/ib_addr.h>
36 #include <rdma/ib_umem.h>
37 #include <rdma/uverbs_ioctl.h>
38 #include "hns_roce_common.h"
39 #include "hns_roce_device.h"
40 #include "hns_roce_hem.h"
41 
42 static void flush_work_handle(struct work_struct *work)
43 {
44 	struct hns_roce_work *flush_work = container_of(work,
45 					struct hns_roce_work, work);
46 	struct hns_roce_qp *hr_qp = container_of(flush_work,
47 					struct hns_roce_qp, flush_work);
48 	struct device *dev = flush_work->hr_dev->dev;
49 	struct ib_qp_attr attr;
50 	int attr_mask;
51 	int ret;
52 
53 	attr_mask = IB_QP_STATE;
54 	attr.qp_state = IB_QPS_ERR;
55 
56 	if (test_and_clear_bit(HNS_ROCE_FLUSH_FLAG, &hr_qp->flush_flag)) {
57 		ret = hns_roce_modify_qp(&hr_qp->ibqp, &attr, attr_mask, NULL);
58 		if (ret)
59 			dev_err(dev, "Modify QP to error state failed(%d) during CQE flush\n",
60 				ret);
61 	}
62 
63 	/*
64 	 * make sure we signal QP destroy leg that flush QP was completed
65 	 * so that it can safely proceed ahead now and destroy QP
66 	 */
67 	if (refcount_dec_and_test(&hr_qp->refcount))
68 		complete(&hr_qp->free);
69 }
70 
71 void init_flush_work(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
72 {
73 	struct hns_roce_work *flush_work = &hr_qp->flush_work;
74 
75 	flush_work->hr_dev = hr_dev;
76 	INIT_WORK(&flush_work->work, flush_work_handle);
77 	refcount_inc(&hr_qp->refcount);
78 	queue_work(hr_dev->irq_workq, &flush_work->work);
79 }
80 
81 void flush_cqe(struct hns_roce_dev *dev, struct hns_roce_qp *qp)
82 {
83 	/*
84 	 * Hip08 hardware cannot flush the WQEs in SQ/RQ if the QP state
85 	 * gets into errored mode. Hence, as a workaround to this
86 	 * hardware limitation, driver needs to assist in flushing. But
87 	 * the flushing operation uses mailbox to convey the QP state to
88 	 * the hardware and which can sleep due to the mutex protection
89 	 * around the mailbox calls. Hence, use the deferred flush for
90 	 * now.
91 	 */
92 	if (!test_and_set_bit(HNS_ROCE_FLUSH_FLAG, &qp->flush_flag))
93 		init_flush_work(dev, qp);
94 }
95 
96 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
97 {
98 	struct device *dev = hr_dev->dev;
99 	struct hns_roce_qp *qp;
100 
101 	xa_lock(&hr_dev->qp_table_xa);
102 	qp = __hns_roce_qp_lookup(hr_dev, qpn);
103 	if (qp)
104 		refcount_inc(&qp->refcount);
105 	xa_unlock(&hr_dev->qp_table_xa);
106 
107 	if (!qp) {
108 		dev_warn(dev, "Async event for bogus QP %08x\n", qpn);
109 		return;
110 	}
111 
112 	if (event_type == HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR ||
113 	    event_type == HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR ||
114 	    event_type == HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR ||
115 	    event_type == HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION ||
116 	    event_type == HNS_ROCE_EVENT_TYPE_INVALID_XRCETH) {
117 		qp->state = IB_QPS_ERR;
118 
119 		flush_cqe(hr_dev, qp);
120 	}
121 
122 	qp->event(qp, (enum hns_roce_event)event_type);
123 
124 	if (refcount_dec_and_test(&qp->refcount))
125 		complete(&qp->free);
126 }
127 
128 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
129 				 enum hns_roce_event type)
130 {
131 	struct ib_qp *ibqp = &hr_qp->ibqp;
132 	struct ib_event event;
133 
134 	if (ibqp->event_handler) {
135 		event.device = ibqp->device;
136 		event.element.qp = ibqp;
137 		switch (type) {
138 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
139 			event.event = IB_EVENT_PATH_MIG;
140 			break;
141 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
142 			event.event = IB_EVENT_COMM_EST;
143 			break;
144 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
145 			event.event = IB_EVENT_SQ_DRAINED;
146 			break;
147 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
148 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
149 			break;
150 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
151 			event.event = IB_EVENT_QP_FATAL;
152 			break;
153 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
154 			event.event = IB_EVENT_PATH_MIG_ERR;
155 			break;
156 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
157 			event.event = IB_EVENT_QP_REQ_ERR;
158 			break;
159 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
160 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
161 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
162 			event.event = IB_EVENT_QP_ACCESS_ERR;
163 			break;
164 		default:
165 			dev_dbg(ibqp->device->dev.parent, "roce_ib: Unexpected event type %d on QP %06lx\n",
166 				type, hr_qp->qpn);
167 			return;
168 		}
169 		ibqp->event_handler(&event, ibqp->qp_context);
170 	}
171 }
172 
173 static u8 get_least_load_bankid_for_qp(struct hns_roce_bank *bank)
174 {
175 	u32 least_load = bank[0].inuse;
176 	u8 bankid = 0;
177 	u32 bankcnt;
178 	u8 i;
179 
180 	for (i = 1; i < HNS_ROCE_QP_BANK_NUM; i++) {
181 		bankcnt = bank[i].inuse;
182 		if (bankcnt < least_load) {
183 			least_load = bankcnt;
184 			bankid = i;
185 		}
186 	}
187 
188 	return bankid;
189 }
190 
191 static int alloc_qpn_with_bankid(struct hns_roce_bank *bank, u8 bankid,
192 				 unsigned long *qpn)
193 {
194 	int id;
195 
196 	id = ida_alloc_range(&bank->ida, bank->next, bank->max, GFP_KERNEL);
197 	if (id < 0) {
198 		id = ida_alloc_range(&bank->ida, bank->min, bank->max,
199 				     GFP_KERNEL);
200 		if (id < 0)
201 			return id;
202 	}
203 
204 	/* the QPN should keep increasing until the max value is reached. */
205 	bank->next = (id + 1) > bank->max ? bank->min : id + 1;
206 
207 	/* the lower 3 bits is bankid */
208 	*qpn = (id << 3) | bankid;
209 
210 	return 0;
211 }
212 static int alloc_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
213 {
214 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
215 	unsigned long num = 0;
216 	u8 bankid;
217 	int ret;
218 
219 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI) {
220 		num = 1;
221 		hr_qp->doorbell_qpn = 1;
222 	} else {
223 		mutex_lock(&qp_table->bank_mutex);
224 		bankid = get_least_load_bankid_for_qp(qp_table->bank);
225 
226 		ret = alloc_qpn_with_bankid(&qp_table->bank[bankid], bankid,
227 					    &num);
228 		if (ret) {
229 			ibdev_err(&hr_dev->ib_dev,
230 				  "failed to alloc QPN, ret = %d\n", ret);
231 			mutex_unlock(&qp_table->bank_mutex);
232 			return ret;
233 		}
234 
235 		qp_table->bank[bankid].inuse++;
236 		mutex_unlock(&qp_table->bank_mutex);
237 
238 		hr_qp->doorbell_qpn = (u32)num;
239 	}
240 
241 	hr_qp->qpn = num;
242 
243 	return 0;
244 }
245 
246 static void add_qp_to_list(struct hns_roce_dev *hr_dev,
247 			   struct hns_roce_qp *hr_qp,
248 			   struct ib_cq *send_cq, struct ib_cq *recv_cq)
249 {
250 	struct hns_roce_cq *hr_send_cq, *hr_recv_cq;
251 	unsigned long flags;
252 
253 	hr_send_cq = send_cq ? to_hr_cq(send_cq) : NULL;
254 	hr_recv_cq = recv_cq ? to_hr_cq(recv_cq) : NULL;
255 
256 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
257 	hns_roce_lock_cqs(hr_send_cq, hr_recv_cq);
258 
259 	list_add_tail(&hr_qp->node, &hr_dev->qp_list);
260 	if (hr_send_cq)
261 		list_add_tail(&hr_qp->sq_node, &hr_send_cq->sq_list);
262 	if (hr_recv_cq)
263 		list_add_tail(&hr_qp->rq_node, &hr_recv_cq->rq_list);
264 
265 	hns_roce_unlock_cqs(hr_send_cq, hr_recv_cq);
266 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
267 }
268 
269 static int hns_roce_qp_store(struct hns_roce_dev *hr_dev,
270 			     struct hns_roce_qp *hr_qp,
271 			     struct ib_qp_init_attr *init_attr)
272 {
273 	struct xarray *xa = &hr_dev->qp_table_xa;
274 	int ret;
275 
276 	if (!hr_qp->qpn)
277 		return -EINVAL;
278 
279 	ret = xa_err(xa_store_irq(xa, hr_qp->qpn, hr_qp, GFP_KERNEL));
280 	if (ret)
281 		dev_err(hr_dev->dev, "Failed to xa store for QPC\n");
282 	else
283 		/* add QP to device's QP list for softwc */
284 		add_qp_to_list(hr_dev, hr_qp, init_attr->send_cq,
285 			       init_attr->recv_cq);
286 
287 	return ret;
288 }
289 
290 static int alloc_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
291 {
292 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
293 	struct device *dev = hr_dev->dev;
294 	int ret;
295 
296 	if (!hr_qp->qpn)
297 		return -EINVAL;
298 
299 	/* Alloc memory for QPC */
300 	ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
301 	if (ret) {
302 		dev_err(dev, "Failed to get QPC table\n");
303 		goto err_out;
304 	}
305 
306 	/* Alloc memory for IRRL */
307 	ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
308 	if (ret) {
309 		dev_err(dev, "Failed to get IRRL table\n");
310 		goto err_put_qp;
311 	}
312 
313 	if (hr_dev->caps.trrl_entry_sz) {
314 		/* Alloc memory for TRRL */
315 		ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
316 					 hr_qp->qpn);
317 		if (ret) {
318 			dev_err(dev, "Failed to get TRRL table\n");
319 			goto err_put_irrl;
320 		}
321 	}
322 
323 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
324 		/* Alloc memory for SCC CTX */
325 		ret = hns_roce_table_get(hr_dev, &qp_table->sccc_table,
326 					 hr_qp->qpn);
327 		if (ret) {
328 			dev_err(dev, "Failed to get SCC CTX table\n");
329 			goto err_put_trrl;
330 		}
331 	}
332 
333 	return 0;
334 
335 err_put_trrl:
336 	if (hr_dev->caps.trrl_entry_sz)
337 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
338 
339 err_put_irrl:
340 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
341 
342 err_put_qp:
343 	hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
344 
345 err_out:
346 	return ret;
347 }
348 
349 static void qp_user_mmap_entry_remove(struct hns_roce_qp *hr_qp)
350 {
351 	rdma_user_mmap_entry_remove(&hr_qp->dwqe_mmap_entry->rdma_entry);
352 }
353 
354 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
355 {
356 	struct xarray *xa = &hr_dev->qp_table_xa;
357 	unsigned long flags;
358 
359 	list_del(&hr_qp->node);
360 
361 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
362 		list_del(&hr_qp->sq_node);
363 
364 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_INI &&
365 	    hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
366 		list_del(&hr_qp->rq_node);
367 
368 	xa_lock_irqsave(xa, flags);
369 	__xa_erase(xa, hr_qp->qpn);
370 	xa_unlock_irqrestore(xa, flags);
371 }
372 
373 static void free_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
374 {
375 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
376 
377 	if (hr_dev->caps.trrl_entry_sz)
378 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
379 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
380 }
381 
382 static inline u8 get_qp_bankid(unsigned long qpn)
383 {
384 	/* The lower 3 bits of QPN are used to hash to different banks */
385 	return (u8)(qpn & GENMASK(2, 0));
386 }
387 
388 static void free_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
389 {
390 	u8 bankid;
391 
392 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI)
393 		return;
394 
395 	if (hr_qp->qpn < hr_dev->caps.reserved_qps)
396 		return;
397 
398 	bankid = get_qp_bankid(hr_qp->qpn);
399 
400 	ida_free(&hr_dev->qp_table.bank[bankid].ida, hr_qp->qpn >> 3);
401 
402 	mutex_lock(&hr_dev->qp_table.bank_mutex);
403 	hr_dev->qp_table.bank[bankid].inuse--;
404 	mutex_unlock(&hr_dev->qp_table.bank_mutex);
405 }
406 
407 static u32 proc_rq_sge(struct hns_roce_dev *dev, struct hns_roce_qp *hr_qp,
408 		       bool user)
409 {
410 	u32 max_sge = dev->caps.max_rq_sg;
411 
412 	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
413 		return max_sge;
414 
415 	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
416 	 * calculate number of max_sge with reserved SGEs when allocating wqe
417 	 * buf, so there is no need to do this again in kernel. But the number
418 	 * may exceed the capacity of SGEs recorded in the firmware, so the
419 	 * kernel driver should just adapt the value accordingly.
420 	 */
421 	if (user)
422 		max_sge = roundup_pow_of_two(max_sge + 1);
423 	else
424 		hr_qp->rq.rsv_sge = 1;
425 
426 	return max_sge;
427 }
428 
429 static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
430 		       struct hns_roce_qp *hr_qp, int has_rq, bool user)
431 {
432 	u32 max_sge = proc_rq_sge(hr_dev, hr_qp, user);
433 	u32 cnt;
434 
435 	/* If srq exist, set zero for relative number of rq */
436 	if (!has_rq) {
437 		hr_qp->rq.wqe_cnt = 0;
438 		hr_qp->rq.max_gs = 0;
439 		hr_qp->rq_inl_buf.wqe_cnt = 0;
440 		cap->max_recv_wr = 0;
441 		cap->max_recv_sge = 0;
442 
443 		return 0;
444 	}
445 
446 	/* Check the validity of QP support capacity */
447 	if (!cap->max_recv_wr || cap->max_recv_wr > hr_dev->caps.max_wqes ||
448 	    cap->max_recv_sge > max_sge) {
449 		ibdev_err(&hr_dev->ib_dev,
450 			  "RQ config error, depth = %u, sge = %u\n",
451 			  cap->max_recv_wr, cap->max_recv_sge);
452 		return -EINVAL;
453 	}
454 
455 	cnt = roundup_pow_of_two(max(cap->max_recv_wr, hr_dev->caps.min_wqes));
456 	if (cnt > hr_dev->caps.max_wqes) {
457 		ibdev_err(&hr_dev->ib_dev, "rq depth %u too large\n",
458 			  cap->max_recv_wr);
459 		return -EINVAL;
460 	}
461 
462 	hr_qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge) +
463 					      hr_qp->rq.rsv_sge);
464 
465 	if (hr_dev->caps.max_rq_sg <= HNS_ROCE_SGE_IN_WQE)
466 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz);
467 	else
468 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz *
469 					    hr_qp->rq.max_gs);
470 
471 	hr_qp->rq.wqe_cnt = cnt;
472 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE &&
473 	    hr_qp->ibqp.qp_type != IB_QPT_UD &&
474 	    hr_qp->ibqp.qp_type != IB_QPT_GSI)
475 		hr_qp->rq_inl_buf.wqe_cnt = cnt;
476 	else
477 		hr_qp->rq_inl_buf.wqe_cnt = 0;
478 
479 	cap->max_recv_wr = cnt;
480 	cap->max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
481 
482 	return 0;
483 }
484 
485 static u32 get_wqe_ext_sge_cnt(struct hns_roce_qp *qp)
486 {
487 	/* GSI/UD QP only has extended sge */
488 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_UD)
489 		return qp->sq.max_gs;
490 
491 	if (qp->sq.max_gs > HNS_ROCE_SGE_IN_WQE)
492 		return qp->sq.max_gs - HNS_ROCE_SGE_IN_WQE;
493 
494 	return 0;
495 }
496 
497 static void set_ext_sge_param(struct hns_roce_dev *hr_dev, u32 sq_wqe_cnt,
498 			      struct hns_roce_qp *hr_qp, struct ib_qp_cap *cap)
499 {
500 	u32 total_sge_cnt;
501 	u32 wqe_sge_cnt;
502 
503 	hr_qp->sge.sge_shift = HNS_ROCE_SGE_SHIFT;
504 
505 	hr_qp->sq.max_gs = max(1U, cap->max_send_sge);
506 
507 	wqe_sge_cnt = get_wqe_ext_sge_cnt(hr_qp);
508 
509 	/* If the number of extended sge is not zero, they MUST use the
510 	 * space of HNS_HW_PAGE_SIZE at least.
511 	 */
512 	if (wqe_sge_cnt) {
513 		total_sge_cnt = roundup_pow_of_two(sq_wqe_cnt * wqe_sge_cnt);
514 		hr_qp->sge.sge_cnt = max(total_sge_cnt,
515 				(u32)HNS_HW_PAGE_SIZE / HNS_ROCE_SGE_SIZE);
516 	}
517 }
518 
519 static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
520 					struct ib_qp_cap *cap,
521 					struct hns_roce_ib_create_qp *ucmd)
522 {
523 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
524 	u8 max_sq_stride = ilog2(roundup_sq_stride);
525 
526 	/* Sanity check SQ size before proceeding */
527 	if (ucmd->log_sq_stride > max_sq_stride ||
528 	    ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
529 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ stride size.\n");
530 		return -EINVAL;
531 	}
532 
533 	if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
534 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ SGE size %u.\n",
535 			  cap->max_send_sge);
536 		return -EINVAL;
537 	}
538 
539 	return 0;
540 }
541 
542 static int set_user_sq_size(struct hns_roce_dev *hr_dev,
543 			    struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp,
544 			    struct hns_roce_ib_create_qp *ucmd)
545 {
546 	struct ib_device *ibdev = &hr_dev->ib_dev;
547 	u32 cnt = 0;
548 	int ret;
549 
550 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
551 	    cnt > hr_dev->caps.max_wqes)
552 		return -EINVAL;
553 
554 	ret = check_sq_size_with_integrity(hr_dev, cap, ucmd);
555 	if (ret) {
556 		ibdev_err(ibdev, "failed to check user SQ size, ret = %d.\n",
557 			  ret);
558 		return ret;
559 	}
560 
561 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
562 
563 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
564 	hr_qp->sq.wqe_cnt = cnt;
565 
566 	return 0;
567 }
568 
569 static int set_wqe_buf_attr(struct hns_roce_dev *hr_dev,
570 			    struct hns_roce_qp *hr_qp,
571 			    struct hns_roce_buf_attr *buf_attr)
572 {
573 	int buf_size;
574 	int idx = 0;
575 
576 	hr_qp->buff_size = 0;
577 
578 	/* SQ WQE */
579 	hr_qp->sq.offset = 0;
580 	buf_size = to_hr_hem_entries_size(hr_qp->sq.wqe_cnt,
581 					  hr_qp->sq.wqe_shift);
582 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
583 		buf_attr->region[idx].size = buf_size;
584 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sq_hop_num;
585 		idx++;
586 		hr_qp->buff_size += buf_size;
587 	}
588 
589 	/* extend SGE WQE in SQ */
590 	hr_qp->sge.offset = hr_qp->buff_size;
591 	buf_size = to_hr_hem_entries_size(hr_qp->sge.sge_cnt,
592 					  hr_qp->sge.sge_shift);
593 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
594 		buf_attr->region[idx].size = buf_size;
595 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sge_hop_num;
596 		idx++;
597 		hr_qp->buff_size += buf_size;
598 	}
599 
600 	/* RQ WQE */
601 	hr_qp->rq.offset = hr_qp->buff_size;
602 	buf_size = to_hr_hem_entries_size(hr_qp->rq.wqe_cnt,
603 					  hr_qp->rq.wqe_shift);
604 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
605 		buf_attr->region[idx].size = buf_size;
606 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_rq_hop_num;
607 		idx++;
608 		hr_qp->buff_size += buf_size;
609 	}
610 
611 	if (hr_qp->buff_size < 1)
612 		return -EINVAL;
613 
614 	buf_attr->page_shift = HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
615 	buf_attr->region_count = idx;
616 
617 	return 0;
618 }
619 
620 static int set_kernel_sq_size(struct hns_roce_dev *hr_dev,
621 			      struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp)
622 {
623 	struct ib_device *ibdev = &hr_dev->ib_dev;
624 	u32 cnt;
625 
626 	if (!cap->max_send_wr || cap->max_send_wr > hr_dev->caps.max_wqes ||
627 	    cap->max_send_sge > hr_dev->caps.max_sq_sg) {
628 		ibdev_err(ibdev, "failed to check SQ WR or SGE num.\n");
629 		return -EINVAL;
630 	}
631 
632 	cnt = roundup_pow_of_two(max(cap->max_send_wr, hr_dev->caps.min_wqes));
633 	if (cnt > hr_dev->caps.max_wqes) {
634 		ibdev_err(ibdev, "failed to check WQE num, WQE num = %u.\n",
635 			  cnt);
636 		return -EINVAL;
637 	}
638 
639 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
640 	hr_qp->sq.wqe_cnt = cnt;
641 
642 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
643 
644 	/* sync the parameters of kernel QP to user's configuration */
645 	cap->max_send_wr = cnt;
646 	cap->max_send_sge = hr_qp->sq.max_gs;
647 
648 	return 0;
649 }
650 
651 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
652 {
653 	if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
654 		return 0;
655 
656 	return 1;
657 }
658 
659 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
660 {
661 	if (attr->qp_type == IB_QPT_XRC_INI ||
662 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
663 	    !attr->cap.max_recv_wr)
664 		return 0;
665 
666 	return 1;
667 }
668 
669 static int alloc_rq_inline_buf(struct hns_roce_qp *hr_qp,
670 			       struct ib_qp_init_attr *init_attr)
671 {
672 	u32 max_recv_sge = init_attr->cap.max_recv_sge;
673 	u32 wqe_cnt = hr_qp->rq_inl_buf.wqe_cnt;
674 	struct hns_roce_rinl_wqe *wqe_list;
675 	int i;
676 
677 	/* allocate recv inline buf */
678 	wqe_list = kcalloc(wqe_cnt, sizeof(struct hns_roce_rinl_wqe),
679 			   GFP_KERNEL);
680 	if (!wqe_list)
681 		goto err;
682 
683 	/* Allocate a continuous buffer for all inline sge we need */
684 	wqe_list[0].sg_list = kcalloc(wqe_cnt, (max_recv_sge *
685 				      sizeof(struct hns_roce_rinl_sge)),
686 				      GFP_KERNEL);
687 	if (!wqe_list[0].sg_list)
688 		goto err_wqe_list;
689 
690 	/* Assign buffers of sg_list to each inline wqe */
691 	for (i = 1; i < wqe_cnt; i++)
692 		wqe_list[i].sg_list = &wqe_list[0].sg_list[i * max_recv_sge];
693 
694 	hr_qp->rq_inl_buf.wqe_list = wqe_list;
695 
696 	return 0;
697 
698 err_wqe_list:
699 	kfree(wqe_list);
700 
701 err:
702 	return -ENOMEM;
703 }
704 
705 static void free_rq_inline_buf(struct hns_roce_qp *hr_qp)
706 {
707 	if (hr_qp->rq_inl_buf.wqe_list)
708 		kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
709 	kfree(hr_qp->rq_inl_buf.wqe_list);
710 }
711 
712 static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
713 			struct ib_qp_init_attr *init_attr,
714 			struct ib_udata *udata, unsigned long addr)
715 {
716 	struct ib_device *ibdev = &hr_dev->ib_dev;
717 	struct hns_roce_buf_attr buf_attr = {};
718 	int ret;
719 
720 	if (!udata && hr_qp->rq_inl_buf.wqe_cnt) {
721 		ret = alloc_rq_inline_buf(hr_qp, init_attr);
722 		if (ret) {
723 			ibdev_err(ibdev,
724 				  "failed to alloc inline buf, ret = %d.\n",
725 				  ret);
726 			return ret;
727 		}
728 	} else {
729 		hr_qp->rq_inl_buf.wqe_list = NULL;
730 	}
731 
732 	ret = set_wqe_buf_attr(hr_dev, hr_qp, &buf_attr);
733 	if (ret) {
734 		ibdev_err(ibdev, "failed to split WQE buf, ret = %d.\n", ret);
735 		goto err_inline;
736 	}
737 	ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, &buf_attr,
738 				  PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
739 				  udata, addr);
740 	if (ret) {
741 		ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret);
742 		goto err_inline;
743 	}
744 
745 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_DIRECT_WQE)
746 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_DIRECT_WQE;
747 
748 	return 0;
749 
750 err_inline:
751 	free_rq_inline_buf(hr_qp);
752 
753 	return ret;
754 }
755 
756 static void free_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
757 {
758 	hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
759 	free_rq_inline_buf(hr_qp);
760 }
761 
762 static inline bool user_qp_has_sdb(struct hns_roce_dev *hr_dev,
763 				   struct ib_qp_init_attr *init_attr,
764 				   struct ib_udata *udata,
765 				   struct hns_roce_ib_create_qp_resp *resp,
766 				   struct hns_roce_ib_create_qp *ucmd)
767 {
768 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
769 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
770 		hns_roce_qp_has_sq(init_attr) &&
771 		udata->inlen >= offsetofend(typeof(*ucmd), sdb_addr));
772 }
773 
774 static inline bool user_qp_has_rdb(struct hns_roce_dev *hr_dev,
775 				   struct ib_qp_init_attr *init_attr,
776 				   struct ib_udata *udata,
777 				   struct hns_roce_ib_create_qp_resp *resp)
778 {
779 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
780 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
781 		hns_roce_qp_has_rq(init_attr));
782 }
783 
784 static inline bool kernel_qp_has_rdb(struct hns_roce_dev *hr_dev,
785 				     struct ib_qp_init_attr *init_attr)
786 {
787 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
788 		hns_roce_qp_has_rq(init_attr));
789 }
790 
791 static int qp_mmap_entry(struct hns_roce_qp *hr_qp,
792 			 struct hns_roce_dev *hr_dev,
793 			 struct ib_udata *udata,
794 			 struct hns_roce_ib_create_qp_resp *resp)
795 {
796 	struct hns_roce_ucontext *uctx =
797 		rdma_udata_to_drv_context(udata,
798 			struct hns_roce_ucontext, ibucontext);
799 	struct rdma_user_mmap_entry *rdma_entry;
800 	u64 address;
801 
802 	address = hr_dev->dwqe_page + hr_qp->qpn * HNS_ROCE_DWQE_SIZE;
803 
804 	hr_qp->dwqe_mmap_entry =
805 		hns_roce_user_mmap_entry_insert(&uctx->ibucontext, address,
806 						HNS_ROCE_DWQE_SIZE,
807 						HNS_ROCE_MMAP_TYPE_DWQE);
808 
809 	if (!hr_qp->dwqe_mmap_entry) {
810 		ibdev_err(&hr_dev->ib_dev, "failed to get dwqe mmap entry.\n");
811 		return -ENOMEM;
812 	}
813 
814 	rdma_entry = &hr_qp->dwqe_mmap_entry->rdma_entry;
815 	resp->dwqe_mmap_key = rdma_user_mmap_get_offset(rdma_entry);
816 
817 	return 0;
818 }
819 
820 static int alloc_user_qp_db(struct hns_roce_dev *hr_dev,
821 			    struct hns_roce_qp *hr_qp,
822 			    struct ib_qp_init_attr *init_attr,
823 			    struct ib_udata *udata,
824 			    struct hns_roce_ib_create_qp *ucmd,
825 			    struct hns_roce_ib_create_qp_resp *resp)
826 {
827 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(udata,
828 		struct hns_roce_ucontext, ibucontext);
829 	struct ib_device *ibdev = &hr_dev->ib_dev;
830 	int ret;
831 
832 	if (user_qp_has_sdb(hr_dev, init_attr, udata, resp, ucmd)) {
833 		ret = hns_roce_db_map_user(uctx, ucmd->sdb_addr, &hr_qp->sdb);
834 		if (ret) {
835 			ibdev_err(ibdev,
836 				  "failed to map user SQ doorbell, ret = %d.\n",
837 				  ret);
838 			goto err_out;
839 		}
840 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
841 	}
842 
843 	if (user_qp_has_rdb(hr_dev, init_attr, udata, resp)) {
844 		ret = hns_roce_db_map_user(uctx, ucmd->db_addr, &hr_qp->rdb);
845 		if (ret) {
846 			ibdev_err(ibdev,
847 				  "failed to map user RQ doorbell, ret = %d.\n",
848 				  ret);
849 			goto err_sdb;
850 		}
851 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
852 	}
853 
854 	return 0;
855 
856 err_sdb:
857 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
858 		hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
859 err_out:
860 	return ret;
861 }
862 
863 static int alloc_kernel_qp_db(struct hns_roce_dev *hr_dev,
864 			      struct hns_roce_qp *hr_qp,
865 			      struct ib_qp_init_attr *init_attr)
866 {
867 	struct ib_device *ibdev = &hr_dev->ib_dev;
868 	int ret;
869 
870 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
871 		hr_qp->sq.db_reg = hr_dev->mem_base +
872 				   HNS_ROCE_DWQE_SIZE * hr_qp->qpn;
873 	else
874 		hr_qp->sq.db_reg = hr_dev->reg_base + hr_dev->sdb_offset +
875 				   DB_REG_OFFSET * hr_dev->priv_uar.index;
876 
877 	hr_qp->rq.db_reg = hr_dev->reg_base + hr_dev->odb_offset +
878 			   DB_REG_OFFSET * hr_dev->priv_uar.index;
879 
880 	if (kernel_qp_has_rdb(hr_dev, init_attr)) {
881 		ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
882 		if (ret) {
883 			ibdev_err(ibdev,
884 				  "failed to alloc kernel RQ doorbell, ret = %d.\n",
885 				  ret);
886 			return ret;
887 		}
888 		*hr_qp->rdb.db_record = 0;
889 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
890 	}
891 
892 	return 0;
893 }
894 
895 static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
896 		       struct ib_qp_init_attr *init_attr,
897 		       struct ib_udata *udata,
898 		       struct hns_roce_ib_create_qp *ucmd,
899 		       struct hns_roce_ib_create_qp_resp *resp)
900 {
901 	int ret;
902 
903 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SDI_MODE)
904 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_OWNER_DB;
905 
906 	if (udata) {
907 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE) {
908 			ret = qp_mmap_entry(hr_qp, hr_dev, udata, resp);
909 			if (ret)
910 				return ret;
911 		}
912 
913 		ret = alloc_user_qp_db(hr_dev, hr_qp, init_attr, udata, ucmd,
914 				       resp);
915 		if (ret)
916 			goto err_remove_qp;
917 	} else {
918 		ret = alloc_kernel_qp_db(hr_dev, hr_qp, init_attr);
919 		if (ret)
920 			return ret;
921 	}
922 
923 	return 0;
924 
925 err_remove_qp:
926 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
927 		qp_user_mmap_entry_remove(hr_qp);
928 
929 	return ret;
930 }
931 
932 static void free_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
933 		       struct ib_udata *udata)
934 {
935 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
936 		udata, struct hns_roce_ucontext, ibucontext);
937 
938 	if (udata) {
939 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
940 			hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
941 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
942 			hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
943 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
944 			qp_user_mmap_entry_remove(hr_qp);
945 	} else {
946 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
947 			hns_roce_free_db(hr_dev, &hr_qp->rdb);
948 	}
949 }
950 
951 static int alloc_kernel_wrid(struct hns_roce_dev *hr_dev,
952 			     struct hns_roce_qp *hr_qp)
953 {
954 	struct ib_device *ibdev = &hr_dev->ib_dev;
955 	u64 *sq_wrid = NULL;
956 	u64 *rq_wrid = NULL;
957 	int ret;
958 
959 	sq_wrid = kcalloc(hr_qp->sq.wqe_cnt, sizeof(u64), GFP_KERNEL);
960 	if (ZERO_OR_NULL_PTR(sq_wrid)) {
961 		ibdev_err(ibdev, "failed to alloc SQ wrid.\n");
962 		return -ENOMEM;
963 	}
964 
965 	if (hr_qp->rq.wqe_cnt) {
966 		rq_wrid = kcalloc(hr_qp->rq.wqe_cnt, sizeof(u64), GFP_KERNEL);
967 		if (ZERO_OR_NULL_PTR(rq_wrid)) {
968 			ibdev_err(ibdev, "failed to alloc RQ wrid.\n");
969 			ret = -ENOMEM;
970 			goto err_sq;
971 		}
972 	}
973 
974 	hr_qp->sq.wrid = sq_wrid;
975 	hr_qp->rq.wrid = rq_wrid;
976 	return 0;
977 err_sq:
978 	kfree(sq_wrid);
979 
980 	return ret;
981 }
982 
983 static void free_kernel_wrid(struct hns_roce_qp *hr_qp)
984 {
985 	kfree(hr_qp->rq.wrid);
986 	kfree(hr_qp->sq.wrid);
987 }
988 
989 static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
990 			struct ib_qp_init_attr *init_attr,
991 			struct ib_udata *udata,
992 			struct hns_roce_ib_create_qp *ucmd)
993 {
994 	struct ib_device *ibdev = &hr_dev->ib_dev;
995 	int ret;
996 
997 	if (init_attr->cap.max_inline_data > hr_dev->caps.max_sq_inline)
998 		init_attr->cap.max_inline_data = hr_dev->caps.max_sq_inline;
999 
1000 	hr_qp->max_inline_data = init_attr->cap.max_inline_data;
1001 
1002 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1003 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
1004 	else
1005 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
1006 
1007 	ret = set_rq_size(hr_dev, &init_attr->cap, hr_qp,
1008 			  hns_roce_qp_has_rq(init_attr), !!udata);
1009 	if (ret) {
1010 		ibdev_err(ibdev, "failed to set user RQ size, ret = %d.\n",
1011 			  ret);
1012 		return ret;
1013 	}
1014 
1015 	if (udata) {
1016 		ret = ib_copy_from_udata(ucmd, udata,
1017 					 min(udata->inlen, sizeof(*ucmd)));
1018 		if (ret) {
1019 			ibdev_err(ibdev,
1020 				  "failed to copy QP ucmd, ret = %d\n", ret);
1021 			return ret;
1022 		}
1023 
1024 		ret = set_user_sq_size(hr_dev, &init_attr->cap, hr_qp, ucmd);
1025 		if (ret)
1026 			ibdev_err(ibdev,
1027 				  "failed to set user SQ size, ret = %d.\n",
1028 				  ret);
1029 	} else {
1030 		ret = set_kernel_sq_size(hr_dev, &init_attr->cap, hr_qp);
1031 		if (ret)
1032 			ibdev_err(ibdev,
1033 				  "failed to set kernel SQ size, ret = %d.\n",
1034 				  ret);
1035 	}
1036 
1037 	return ret;
1038 }
1039 
1040 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
1041 				     struct ib_pd *ib_pd,
1042 				     struct ib_qp_init_attr *init_attr,
1043 				     struct ib_udata *udata,
1044 				     struct hns_roce_qp *hr_qp)
1045 {
1046 	struct hns_roce_ib_create_qp_resp resp = {};
1047 	struct ib_device *ibdev = &hr_dev->ib_dev;
1048 	struct hns_roce_ib_create_qp ucmd;
1049 	int ret;
1050 
1051 	mutex_init(&hr_qp->mutex);
1052 	spin_lock_init(&hr_qp->sq.lock);
1053 	spin_lock_init(&hr_qp->rq.lock);
1054 
1055 	hr_qp->state = IB_QPS_RESET;
1056 	hr_qp->flush_flag = 0;
1057 
1058 	if (init_attr->create_flags)
1059 		return -EOPNOTSUPP;
1060 
1061 	ret = set_qp_param(hr_dev, hr_qp, init_attr, udata, &ucmd);
1062 	if (ret) {
1063 		ibdev_err(ibdev, "failed to set QP param, ret = %d.\n", ret);
1064 		return ret;
1065 	}
1066 
1067 	if (!udata) {
1068 		ret = alloc_kernel_wrid(hr_dev, hr_qp);
1069 		if (ret) {
1070 			ibdev_err(ibdev, "failed to alloc wrid, ret = %d.\n",
1071 				  ret);
1072 			return ret;
1073 		}
1074 	}
1075 
1076 	ret = alloc_qp_buf(hr_dev, hr_qp, init_attr, udata, ucmd.buf_addr);
1077 	if (ret) {
1078 		ibdev_err(ibdev, "failed to alloc QP buffer, ret = %d.\n", ret);
1079 		goto err_buf;
1080 	}
1081 
1082 	ret = alloc_qpn(hr_dev, hr_qp);
1083 	if (ret) {
1084 		ibdev_err(ibdev, "failed to alloc QPN, ret = %d.\n", ret);
1085 		goto err_qpn;
1086 	}
1087 
1088 	ret = alloc_qp_db(hr_dev, hr_qp, init_attr, udata, &ucmd, &resp);
1089 	if (ret) {
1090 		ibdev_err(ibdev, "failed to alloc QP doorbell, ret = %d.\n",
1091 			  ret);
1092 		goto err_db;
1093 	}
1094 
1095 	ret = alloc_qpc(hr_dev, hr_qp);
1096 	if (ret) {
1097 		ibdev_err(ibdev, "failed to alloc QP context, ret = %d.\n",
1098 			  ret);
1099 		goto err_qpc;
1100 	}
1101 
1102 	ret = hns_roce_qp_store(hr_dev, hr_qp, init_attr);
1103 	if (ret) {
1104 		ibdev_err(ibdev, "failed to store QP, ret = %d.\n", ret);
1105 		goto err_store;
1106 	}
1107 
1108 	if (udata) {
1109 		resp.cap_flags = hr_qp->en_flags;
1110 		ret = ib_copy_to_udata(udata, &resp,
1111 				       min(udata->outlen, sizeof(resp)));
1112 		if (ret) {
1113 			ibdev_err(ibdev, "copy qp resp failed!\n");
1114 			goto err_store;
1115 		}
1116 	}
1117 
1118 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
1119 		ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
1120 		if (ret)
1121 			goto err_flow_ctrl;
1122 	}
1123 
1124 	hr_qp->ibqp.qp_num = hr_qp->qpn;
1125 	hr_qp->event = hns_roce_ib_qp_event;
1126 	refcount_set(&hr_qp->refcount, 1);
1127 	init_completion(&hr_qp->free);
1128 
1129 	return 0;
1130 
1131 err_flow_ctrl:
1132 	hns_roce_qp_remove(hr_dev, hr_qp);
1133 err_store:
1134 	free_qpc(hr_dev, hr_qp);
1135 err_qpc:
1136 	free_qp_db(hr_dev, hr_qp, udata);
1137 err_db:
1138 	free_qpn(hr_dev, hr_qp);
1139 err_qpn:
1140 	free_qp_buf(hr_dev, hr_qp);
1141 err_buf:
1142 	free_kernel_wrid(hr_qp);
1143 	return ret;
1144 }
1145 
1146 void hns_roce_qp_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1147 			 struct ib_udata *udata)
1148 {
1149 	if (refcount_dec_and_test(&hr_qp->refcount))
1150 		complete(&hr_qp->free);
1151 	wait_for_completion(&hr_qp->free);
1152 
1153 	free_qpc(hr_dev, hr_qp);
1154 	free_qpn(hr_dev, hr_qp);
1155 	free_qp_buf(hr_dev, hr_qp);
1156 	free_kernel_wrid(hr_qp);
1157 	free_qp_db(hr_dev, hr_qp, udata);
1158 }
1159 
1160 static int check_qp_type(struct hns_roce_dev *hr_dev, enum ib_qp_type type,
1161 			 bool is_user)
1162 {
1163 	switch (type) {
1164 	case IB_QPT_XRC_INI:
1165 	case IB_QPT_XRC_TGT:
1166 		if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC))
1167 			goto out;
1168 		break;
1169 	case IB_QPT_UD:
1170 		if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 &&
1171 		    is_user)
1172 			goto out;
1173 		break;
1174 	case IB_QPT_RC:
1175 	case IB_QPT_GSI:
1176 		break;
1177 	default:
1178 		goto out;
1179 	}
1180 
1181 	return 0;
1182 
1183 out:
1184 	ibdev_err(&hr_dev->ib_dev, "not support QP type %d\n", type);
1185 
1186 	return -EOPNOTSUPP;
1187 }
1188 
1189 int hns_roce_create_qp(struct ib_qp *qp, struct ib_qp_init_attr *init_attr,
1190 		       struct ib_udata *udata)
1191 {
1192 	struct ib_device *ibdev = qp->device;
1193 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
1194 	struct hns_roce_qp *hr_qp = to_hr_qp(qp);
1195 	struct ib_pd *pd = qp->pd;
1196 	int ret;
1197 
1198 	ret = check_qp_type(hr_dev, init_attr->qp_type, !!udata);
1199 	if (ret)
1200 		return ret;
1201 
1202 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1203 		hr_qp->xrcdn = to_hr_xrcd(init_attr->xrcd)->xrcdn;
1204 
1205 	if (init_attr->qp_type == IB_QPT_GSI) {
1206 		hr_qp->port = init_attr->port_num - 1;
1207 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
1208 	}
1209 
1210 	ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, hr_qp);
1211 	if (ret)
1212 		ibdev_err(ibdev, "Create QP type 0x%x failed(%d)\n",
1213 			  init_attr->qp_type, ret);
1214 
1215 	return ret;
1216 }
1217 
1218 int to_hr_qp_type(int qp_type)
1219 {
1220 	switch (qp_type) {
1221 	case IB_QPT_RC:
1222 		return SERV_TYPE_RC;
1223 	case IB_QPT_UD:
1224 	case IB_QPT_GSI:
1225 		return SERV_TYPE_UD;
1226 	case IB_QPT_XRC_INI:
1227 	case IB_QPT_XRC_TGT:
1228 		return SERV_TYPE_XRC;
1229 	default:
1230 		return -1;
1231 	}
1232 }
1233 
1234 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1235 			      struct hns_roce_qp *hr_qp,
1236 			      struct ib_qp_attr *attr, int attr_mask)
1237 {
1238 	enum ib_mtu active_mtu;
1239 	int p;
1240 
1241 	p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1242 	active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1243 
1244 	if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1245 	    attr->path_mtu > hr_dev->caps.max_mtu) ||
1246 	    attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1247 		ibdev_err(&hr_dev->ib_dev,
1248 			"attr path_mtu(%d)invalid while modify qp",
1249 			attr->path_mtu);
1250 		return -EINVAL;
1251 	}
1252 
1253 	return 0;
1254 }
1255 
1256 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1257 				  int attr_mask)
1258 {
1259 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1260 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1261 	int p;
1262 
1263 	if ((attr_mask & IB_QP_PORT) &&
1264 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1265 		ibdev_err(&hr_dev->ib_dev, "invalid attr, port_num = %u.\n",
1266 			  attr->port_num);
1267 		return -EINVAL;
1268 	}
1269 
1270 	if (attr_mask & IB_QP_PKEY_INDEX) {
1271 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1272 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1273 			ibdev_err(&hr_dev->ib_dev,
1274 				  "invalid attr, pkey_index = %u.\n",
1275 				  attr->pkey_index);
1276 			return -EINVAL;
1277 		}
1278 	}
1279 
1280 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1281 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1282 		ibdev_err(&hr_dev->ib_dev,
1283 			  "invalid attr, max_rd_atomic = %u.\n",
1284 			  attr->max_rd_atomic);
1285 		return -EINVAL;
1286 	}
1287 
1288 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1289 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1290 		ibdev_err(&hr_dev->ib_dev,
1291 			  "invalid attr, max_dest_rd_atomic = %u.\n",
1292 			  attr->max_dest_rd_atomic);
1293 		return -EINVAL;
1294 	}
1295 
1296 	if (attr_mask & IB_QP_PATH_MTU)
1297 		return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1298 
1299 	return 0;
1300 }
1301 
1302 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1303 		       int attr_mask, struct ib_udata *udata)
1304 {
1305 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1306 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1307 	enum ib_qp_state cur_state, new_state;
1308 	int ret = -EINVAL;
1309 
1310 	mutex_lock(&hr_qp->mutex);
1311 
1312 	if (attr_mask & IB_QP_CUR_STATE && attr->cur_qp_state != hr_qp->state)
1313 		goto out;
1314 
1315 	cur_state = hr_qp->state;
1316 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1317 
1318 	if (ibqp->uobject &&
1319 	    (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1320 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB) {
1321 			hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1322 
1323 			if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
1324 				hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1325 		} else {
1326 			ibdev_warn(&hr_dev->ib_dev,
1327 				  "flush cqe is not supported in userspace!\n");
1328 			goto out;
1329 		}
1330 	}
1331 
1332 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1333 				attr_mask)) {
1334 		ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1335 		goto out;
1336 	}
1337 
1338 	ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1339 	if (ret)
1340 		goto out;
1341 
1342 	if (cur_state == new_state && cur_state == IB_QPS_RESET)
1343 		goto out;
1344 
1345 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1346 				    new_state);
1347 
1348 out:
1349 	mutex_unlock(&hr_qp->mutex);
1350 
1351 	return ret;
1352 }
1353 
1354 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1355 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1356 {
1357 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1358 		__acquire(&send_cq->lock);
1359 		__acquire(&recv_cq->lock);
1360 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1361 		spin_lock_irq(&send_cq->lock);
1362 		__acquire(&recv_cq->lock);
1363 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1364 		spin_lock_irq(&recv_cq->lock);
1365 		__acquire(&send_cq->lock);
1366 	} else if (send_cq == recv_cq) {
1367 		spin_lock_irq(&send_cq->lock);
1368 		__acquire(&recv_cq->lock);
1369 	} else if (send_cq->cqn < recv_cq->cqn) {
1370 		spin_lock_irq(&send_cq->lock);
1371 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1372 	} else {
1373 		spin_lock_irq(&recv_cq->lock);
1374 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1375 	}
1376 }
1377 
1378 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1379 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1380 			 __releases(&recv_cq->lock)
1381 {
1382 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1383 		__release(&recv_cq->lock);
1384 		__release(&send_cq->lock);
1385 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1386 		__release(&recv_cq->lock);
1387 		spin_unlock(&send_cq->lock);
1388 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1389 		__release(&send_cq->lock);
1390 		spin_unlock(&recv_cq->lock);
1391 	} else if (send_cq == recv_cq) {
1392 		__release(&recv_cq->lock);
1393 		spin_unlock_irq(&send_cq->lock);
1394 	} else if (send_cq->cqn < recv_cq->cqn) {
1395 		spin_unlock(&recv_cq->lock);
1396 		spin_unlock_irq(&send_cq->lock);
1397 	} else {
1398 		spin_unlock(&send_cq->lock);
1399 		spin_unlock_irq(&recv_cq->lock);
1400 	}
1401 }
1402 
1403 static inline void *get_wqe(struct hns_roce_qp *hr_qp, u32 offset)
1404 {
1405 	return hns_roce_buf_offset(hr_qp->mtr.kmem, offset);
1406 }
1407 
1408 void *hns_roce_get_recv_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1409 {
1410 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1411 }
1412 
1413 void *hns_roce_get_send_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1414 {
1415 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1416 }
1417 
1418 void *hns_roce_get_extend_sge(struct hns_roce_qp *hr_qp, unsigned int n)
1419 {
1420 	return get_wqe(hr_qp, hr_qp->sge.offset + (n << hr_qp->sge.sge_shift));
1421 }
1422 
1423 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, u32 nreq,
1424 			  struct ib_cq *ib_cq)
1425 {
1426 	struct hns_roce_cq *hr_cq;
1427 	u32 cur;
1428 
1429 	cur = hr_wq->head - hr_wq->tail;
1430 	if (likely(cur + nreq < hr_wq->wqe_cnt))
1431 		return false;
1432 
1433 	hr_cq = to_hr_cq(ib_cq);
1434 	spin_lock(&hr_cq->lock);
1435 	cur = hr_wq->head - hr_wq->tail;
1436 	spin_unlock(&hr_cq->lock);
1437 
1438 	return cur + nreq >= hr_wq->wqe_cnt;
1439 }
1440 
1441 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1442 {
1443 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1444 	unsigned int reserved_from_bot;
1445 	unsigned int i;
1446 
1447 	qp_table->idx_table.spare_idx = kcalloc(hr_dev->caps.num_qps,
1448 					sizeof(u32), GFP_KERNEL);
1449 	if (!qp_table->idx_table.spare_idx)
1450 		return -ENOMEM;
1451 
1452 	mutex_init(&qp_table->scc_mutex);
1453 	mutex_init(&qp_table->bank_mutex);
1454 	xa_init(&hr_dev->qp_table_xa);
1455 
1456 	reserved_from_bot = hr_dev->caps.reserved_qps;
1457 
1458 	for (i = 0; i < reserved_from_bot; i++) {
1459 		hr_dev->qp_table.bank[get_qp_bankid(i)].inuse++;
1460 		hr_dev->qp_table.bank[get_qp_bankid(i)].min++;
1461 	}
1462 
1463 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) {
1464 		ida_init(&hr_dev->qp_table.bank[i].ida);
1465 		hr_dev->qp_table.bank[i].max = hr_dev->caps.num_qps /
1466 					       HNS_ROCE_QP_BANK_NUM - 1;
1467 		hr_dev->qp_table.bank[i].next = hr_dev->qp_table.bank[i].min;
1468 	}
1469 
1470 	return 0;
1471 }
1472 
1473 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1474 {
1475 	int i;
1476 
1477 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++)
1478 		ida_destroy(&hr_dev->qp_table.bank[i].ida);
1479 	kfree(hr_dev->qp_table.idx_table.spare_idx);
1480 }
1481