xref: /linux/drivers/infiniband/hw/hns/hns_roce_qp.c (revision be239684b18e1cdcafcf8c7face4a2f562c745ad)
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_affinity_cq_bank(u8 qp_bank)
174 {
175 	return (qp_bank >> 1) & CQ_BANKID_MASK;
176 }
177 
178 static u8 get_least_load_bankid_for_qp(struct ib_qp_init_attr *init_attr,
179 					struct hns_roce_bank *bank)
180 {
181 #define INVALID_LOAD_QPNUM 0xFFFFFFFF
182 	struct ib_cq *scq = init_attr->send_cq;
183 	u32 least_load = INVALID_LOAD_QPNUM;
184 	unsigned long cqn = 0;
185 	u8 bankid = 0;
186 	u32 bankcnt;
187 	u8 i;
188 
189 	if (scq)
190 		cqn = to_hr_cq(scq)->cqn;
191 
192 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) {
193 		if (scq && (get_affinity_cq_bank(i) != (cqn & CQ_BANKID_MASK)))
194 			continue;
195 
196 		bankcnt = bank[i].inuse;
197 		if (bankcnt < least_load) {
198 			least_load = bankcnt;
199 			bankid = i;
200 		}
201 	}
202 
203 	return bankid;
204 }
205 
206 static int alloc_qpn_with_bankid(struct hns_roce_bank *bank, u8 bankid,
207 				 unsigned long *qpn)
208 {
209 	int id;
210 
211 	id = ida_alloc_range(&bank->ida, bank->next, bank->max, GFP_KERNEL);
212 	if (id < 0) {
213 		id = ida_alloc_range(&bank->ida, bank->min, bank->max,
214 				     GFP_KERNEL);
215 		if (id < 0)
216 			return id;
217 	}
218 
219 	/* the QPN should keep increasing until the max value is reached. */
220 	bank->next = (id + 1) > bank->max ? bank->min : id + 1;
221 
222 	/* the lower 3 bits is bankid */
223 	*qpn = (id << 3) | bankid;
224 
225 	return 0;
226 }
227 static int alloc_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
228 		     struct ib_qp_init_attr *init_attr)
229 {
230 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
231 	unsigned long num = 0;
232 	u8 bankid;
233 	int ret;
234 
235 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI) {
236 		num = 1;
237 	} else {
238 		mutex_lock(&qp_table->bank_mutex);
239 		bankid = get_least_load_bankid_for_qp(init_attr, qp_table->bank);
240 
241 		ret = alloc_qpn_with_bankid(&qp_table->bank[bankid], bankid,
242 					    &num);
243 		if (ret) {
244 			ibdev_err(&hr_dev->ib_dev,
245 				  "failed to alloc QPN, ret = %d\n", ret);
246 			mutex_unlock(&qp_table->bank_mutex);
247 			return ret;
248 		}
249 
250 		qp_table->bank[bankid].inuse++;
251 		mutex_unlock(&qp_table->bank_mutex);
252 	}
253 
254 	hr_qp->qpn = num;
255 
256 	return 0;
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 	/* Alloc memory for QPC */
313 	ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
314 	if (ret) {
315 		dev_err(dev, "failed to get QPC table\n");
316 		goto err_out;
317 	}
318 
319 	/* Alloc memory for IRRL */
320 	ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
321 	if (ret) {
322 		dev_err(dev, "failed to get IRRL table\n");
323 		goto err_put_qp;
324 	}
325 
326 	if (hr_dev->caps.trrl_entry_sz) {
327 		/* Alloc memory for TRRL */
328 		ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
329 					 hr_qp->qpn);
330 		if (ret) {
331 			dev_err(dev, "failed to get TRRL table\n");
332 			goto err_put_irrl;
333 		}
334 	}
335 
336 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
337 		/* Alloc memory for SCC CTX */
338 		ret = hns_roce_table_get(hr_dev, &qp_table->sccc_table,
339 					 hr_qp->qpn);
340 		if (ret) {
341 			dev_err(dev, "failed to get SCC CTX table\n");
342 			goto err_put_trrl;
343 		}
344 	}
345 
346 	return 0;
347 
348 err_put_trrl:
349 	if (hr_dev->caps.trrl_entry_sz)
350 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
351 
352 err_put_irrl:
353 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
354 
355 err_put_qp:
356 	hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
357 
358 err_out:
359 	return ret;
360 }
361 
362 static void qp_user_mmap_entry_remove(struct hns_roce_qp *hr_qp)
363 {
364 	rdma_user_mmap_entry_remove(&hr_qp->dwqe_mmap_entry->rdma_entry);
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);
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 	if (hr_dev->caps.trrl_entry_sz)
391 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
392 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
393 }
394 
395 static inline u8 get_qp_bankid(unsigned long qpn)
396 {
397 	/* The lower 3 bits of QPN are used to hash to different banks */
398 	return (u8)(qpn & GENMASK(2, 0));
399 }
400 
401 static void free_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
402 {
403 	u8 bankid;
404 
405 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI)
406 		return;
407 
408 	if (hr_qp->qpn < hr_dev->caps.reserved_qps)
409 		return;
410 
411 	bankid = get_qp_bankid(hr_qp->qpn);
412 
413 	ida_free(&hr_dev->qp_table.bank[bankid].ida, hr_qp->qpn >> 3);
414 
415 	mutex_lock(&hr_dev->qp_table.bank_mutex);
416 	hr_dev->qp_table.bank[bankid].inuse--;
417 	mutex_unlock(&hr_dev->qp_table.bank_mutex);
418 }
419 
420 static u32 proc_rq_sge(struct hns_roce_dev *dev, struct hns_roce_qp *hr_qp,
421 		       bool user)
422 {
423 	u32 max_sge = dev->caps.max_rq_sg;
424 
425 	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
426 		return max_sge;
427 
428 	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
429 	 * calculate number of max_sge with reserved SGEs when allocating wqe
430 	 * buf, so there is no need to do this again in kernel. But the number
431 	 * may exceed the capacity of SGEs recorded in the firmware, so the
432 	 * kernel driver should just adapt the value accordingly.
433 	 */
434 	if (user)
435 		max_sge = roundup_pow_of_two(max_sge + 1);
436 	else
437 		hr_qp->rq.rsv_sge = 1;
438 
439 	return max_sge;
440 }
441 
442 static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
443 		       struct hns_roce_qp *hr_qp, int has_rq, bool user)
444 {
445 	u32 max_sge = proc_rq_sge(hr_dev, hr_qp, user);
446 	u32 cnt;
447 
448 	/* If srq exist, set zero for relative number of rq */
449 	if (!has_rq) {
450 		hr_qp->rq.wqe_cnt = 0;
451 		hr_qp->rq.max_gs = 0;
452 		cap->max_recv_wr = 0;
453 		cap->max_recv_sge = 0;
454 
455 		return 0;
456 	}
457 
458 	/* Check the validity of QP support capacity */
459 	if (!cap->max_recv_wr || cap->max_recv_wr > hr_dev->caps.max_wqes ||
460 	    cap->max_recv_sge > max_sge) {
461 		ibdev_err(&hr_dev->ib_dev,
462 			  "RQ config error, depth = %u, sge = %u\n",
463 			  cap->max_recv_wr, cap->max_recv_sge);
464 		return -EINVAL;
465 	}
466 
467 	cnt = roundup_pow_of_two(max(cap->max_recv_wr, hr_dev->caps.min_wqes));
468 	if (cnt > hr_dev->caps.max_wqes) {
469 		ibdev_err(&hr_dev->ib_dev, "rq depth %u too large\n",
470 			  cap->max_recv_wr);
471 		return -EINVAL;
472 	}
473 
474 	hr_qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge) +
475 					      hr_qp->rq.rsv_sge);
476 
477 	hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz *
478 				    hr_qp->rq.max_gs);
479 
480 	hr_qp->rq.wqe_cnt = cnt;
481 
482 	cap->max_recv_wr = cnt;
483 	cap->max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
484 
485 	return 0;
486 }
487 
488 static u32 get_max_inline_data(struct hns_roce_dev *hr_dev,
489 			       struct ib_qp_cap *cap)
490 {
491 	if (cap->max_inline_data) {
492 		cap->max_inline_data = roundup_pow_of_two(cap->max_inline_data);
493 		return min(cap->max_inline_data,
494 			   hr_dev->caps.max_sq_inline);
495 	}
496 
497 	return 0;
498 }
499 
500 static void update_inline_data(struct hns_roce_qp *hr_qp,
501 			       struct ib_qp_cap *cap)
502 {
503 	u32 sge_num = hr_qp->sq.ext_sge_cnt;
504 
505 	if (hr_qp->config & HNS_ROCE_EXSGE_FLAGS) {
506 		if (!(hr_qp->ibqp.qp_type == IB_QPT_GSI ||
507 		      hr_qp->ibqp.qp_type == IB_QPT_UD))
508 			sge_num = max((u32)HNS_ROCE_SGE_IN_WQE, sge_num);
509 
510 		cap->max_inline_data = max(cap->max_inline_data,
511 					   sge_num * HNS_ROCE_SGE_SIZE);
512 	}
513 
514 	hr_qp->max_inline_data = cap->max_inline_data;
515 }
516 
517 static u32 get_sge_num_from_max_send_sge(bool is_ud_or_gsi,
518 					 u32 max_send_sge)
519 {
520 	unsigned int std_sge_num;
521 	unsigned int min_sge;
522 
523 	std_sge_num = is_ud_or_gsi ? 0 : HNS_ROCE_SGE_IN_WQE;
524 	min_sge = is_ud_or_gsi ? 1 : 0;
525 	return max_send_sge > std_sge_num ? (max_send_sge - std_sge_num) :
526 				min_sge;
527 }
528 
529 static unsigned int get_sge_num_from_max_inl_data(bool is_ud_or_gsi,
530 						  u32 max_inline_data)
531 {
532 	unsigned int inline_sge;
533 
534 	inline_sge = roundup_pow_of_two(max_inline_data) / HNS_ROCE_SGE_SIZE;
535 
536 	/*
537 	 * if max_inline_data less than
538 	 * HNS_ROCE_SGE_IN_WQE * HNS_ROCE_SGE_SIZE,
539 	 * In addition to ud's mode, no need to extend sge.
540 	 */
541 	if (!is_ud_or_gsi && inline_sge <= HNS_ROCE_SGE_IN_WQE)
542 		inline_sge = 0;
543 
544 	return inline_sge;
545 }
546 
547 static void set_ext_sge_param(struct hns_roce_dev *hr_dev, u32 sq_wqe_cnt,
548 			      struct hns_roce_qp *hr_qp, struct ib_qp_cap *cap)
549 {
550 	bool is_ud_or_gsi = (hr_qp->ibqp.qp_type == IB_QPT_GSI ||
551 				hr_qp->ibqp.qp_type == IB_QPT_UD);
552 	unsigned int std_sge_num;
553 	u32 inline_ext_sge = 0;
554 	u32 ext_wqe_sge_cnt;
555 	u32 total_sge_cnt;
556 
557 	cap->max_inline_data = get_max_inline_data(hr_dev, cap);
558 
559 	hr_qp->sge.sge_shift = HNS_ROCE_SGE_SHIFT;
560 	std_sge_num = is_ud_or_gsi ? 0 : HNS_ROCE_SGE_IN_WQE;
561 	ext_wqe_sge_cnt = get_sge_num_from_max_send_sge(is_ud_or_gsi,
562 							cap->max_send_sge);
563 
564 	if (hr_qp->config & HNS_ROCE_EXSGE_FLAGS) {
565 		inline_ext_sge = max(ext_wqe_sge_cnt,
566 				     get_sge_num_from_max_inl_data(is_ud_or_gsi,
567 							 cap->max_inline_data));
568 		hr_qp->sq.ext_sge_cnt = inline_ext_sge ?
569 					roundup_pow_of_two(inline_ext_sge) : 0;
570 
571 		hr_qp->sq.max_gs = max(1U, (hr_qp->sq.ext_sge_cnt + std_sge_num));
572 		hr_qp->sq.max_gs = min(hr_qp->sq.max_gs, hr_dev->caps.max_sq_sg);
573 
574 		ext_wqe_sge_cnt = hr_qp->sq.ext_sge_cnt;
575 	} else {
576 		hr_qp->sq.max_gs = max(1U, cap->max_send_sge);
577 		hr_qp->sq.max_gs = min(hr_qp->sq.max_gs, hr_dev->caps.max_sq_sg);
578 		hr_qp->sq.ext_sge_cnt = hr_qp->sq.max_gs;
579 	}
580 
581 	/* If the number of extended sge is not zero, they MUST use the
582 	 * space of HNS_HW_PAGE_SIZE at least.
583 	 */
584 	if (ext_wqe_sge_cnt) {
585 		total_sge_cnt = roundup_pow_of_two(sq_wqe_cnt * ext_wqe_sge_cnt);
586 		hr_qp->sge.sge_cnt = max(total_sge_cnt,
587 				(u32)HNS_HW_PAGE_SIZE / HNS_ROCE_SGE_SIZE);
588 	}
589 
590 	update_inline_data(hr_qp, cap);
591 }
592 
593 static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
594 					struct ib_qp_cap *cap,
595 					struct hns_roce_ib_create_qp *ucmd)
596 {
597 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
598 	u8 max_sq_stride = ilog2(roundup_sq_stride);
599 
600 	/* Sanity check SQ size before proceeding */
601 	if (ucmd->log_sq_stride > max_sq_stride ||
602 	    ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
603 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ stride size.\n");
604 		return -EINVAL;
605 	}
606 
607 	if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
608 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ SGE size %u.\n",
609 			  cap->max_send_sge);
610 		return -EINVAL;
611 	}
612 
613 	return 0;
614 }
615 
616 static int set_user_sq_size(struct hns_roce_dev *hr_dev,
617 			    struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp,
618 			    struct hns_roce_ib_create_qp *ucmd)
619 {
620 	struct ib_device *ibdev = &hr_dev->ib_dev;
621 	u32 cnt = 0;
622 	int ret;
623 
624 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
625 	    cnt > hr_dev->caps.max_wqes)
626 		return -EINVAL;
627 
628 	ret = check_sq_size_with_integrity(hr_dev, cap, ucmd);
629 	if (ret) {
630 		ibdev_err(ibdev, "failed to check user SQ size, ret = %d.\n",
631 			  ret);
632 		return ret;
633 	}
634 
635 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
636 
637 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
638 	hr_qp->sq.wqe_cnt = cnt;
639 	cap->max_send_sge = hr_qp->sq.max_gs;
640 
641 	return 0;
642 }
643 
644 static int set_wqe_buf_attr(struct hns_roce_dev *hr_dev,
645 			    struct hns_roce_qp *hr_qp,
646 			    struct hns_roce_buf_attr *buf_attr)
647 {
648 	int buf_size;
649 	int idx = 0;
650 
651 	hr_qp->buff_size = 0;
652 
653 	/* SQ WQE */
654 	hr_qp->sq.offset = 0;
655 	buf_size = to_hr_hem_entries_size(hr_qp->sq.wqe_cnt,
656 					  hr_qp->sq.wqe_shift);
657 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
658 		buf_attr->region[idx].size = buf_size;
659 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sq_hop_num;
660 		idx++;
661 		hr_qp->buff_size += buf_size;
662 	}
663 
664 	/* extend SGE WQE in SQ */
665 	hr_qp->sge.offset = hr_qp->buff_size;
666 	buf_size = to_hr_hem_entries_size(hr_qp->sge.sge_cnt,
667 					  hr_qp->sge.sge_shift);
668 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
669 		buf_attr->region[idx].size = buf_size;
670 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sge_hop_num;
671 		idx++;
672 		hr_qp->buff_size += buf_size;
673 	}
674 
675 	/* RQ WQE */
676 	hr_qp->rq.offset = hr_qp->buff_size;
677 	buf_size = to_hr_hem_entries_size(hr_qp->rq.wqe_cnt,
678 					  hr_qp->rq.wqe_shift);
679 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
680 		buf_attr->region[idx].size = buf_size;
681 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_rq_hop_num;
682 		idx++;
683 		hr_qp->buff_size += buf_size;
684 	}
685 
686 	if (hr_qp->buff_size < 1)
687 		return -EINVAL;
688 
689 	buf_attr->page_shift = HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
690 	buf_attr->region_count = idx;
691 
692 	return 0;
693 }
694 
695 static int set_kernel_sq_size(struct hns_roce_dev *hr_dev,
696 			      struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp)
697 {
698 	struct ib_device *ibdev = &hr_dev->ib_dev;
699 	u32 cnt;
700 
701 	if (!cap->max_send_wr || cap->max_send_wr > hr_dev->caps.max_wqes ||
702 	    cap->max_send_sge > hr_dev->caps.max_sq_sg) {
703 		ibdev_err(ibdev, "failed to check SQ WR or SGE num.\n");
704 		return -EINVAL;
705 	}
706 
707 	cnt = roundup_pow_of_two(max(cap->max_send_wr, hr_dev->caps.min_wqes));
708 	if (cnt > hr_dev->caps.max_wqes) {
709 		ibdev_err(ibdev, "failed to check WQE num, WQE num = %u.\n",
710 			  cnt);
711 		return -EINVAL;
712 	}
713 
714 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
715 	hr_qp->sq.wqe_cnt = cnt;
716 
717 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
718 
719 	/* sync the parameters of kernel QP to user's configuration */
720 	cap->max_send_wr = cnt;
721 	cap->max_send_sge = hr_qp->sq.max_gs;
722 
723 	return 0;
724 }
725 
726 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
727 {
728 	if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
729 		return 0;
730 
731 	return 1;
732 }
733 
734 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
735 {
736 	if (attr->qp_type == IB_QPT_XRC_INI ||
737 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
738 	    !attr->cap.max_recv_wr)
739 		return 0;
740 
741 	return 1;
742 }
743 
744 static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
745 			struct ib_qp_init_attr *init_attr,
746 			struct ib_udata *udata, unsigned long addr)
747 {
748 	struct ib_device *ibdev = &hr_dev->ib_dev;
749 	struct hns_roce_buf_attr buf_attr = {};
750 	int ret;
751 
752 	ret = set_wqe_buf_attr(hr_dev, hr_qp, &buf_attr);
753 	if (ret) {
754 		ibdev_err(ibdev, "failed to split WQE buf, ret = %d.\n", ret);
755 		goto err_inline;
756 	}
757 	ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, &buf_attr,
758 				  PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
759 				  udata, addr);
760 	if (ret) {
761 		ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret);
762 		goto err_inline;
763 	}
764 
765 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_DIRECT_WQE)
766 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_DIRECT_WQE;
767 
768 	return 0;
769 
770 err_inline:
771 
772 	return ret;
773 }
774 
775 static void free_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
776 {
777 	hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
778 }
779 
780 static inline bool user_qp_has_sdb(struct hns_roce_dev *hr_dev,
781 				   struct ib_qp_init_attr *init_attr,
782 				   struct ib_udata *udata,
783 				   struct hns_roce_ib_create_qp_resp *resp,
784 				   struct hns_roce_ib_create_qp *ucmd)
785 {
786 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
787 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
788 		hns_roce_qp_has_sq(init_attr) &&
789 		udata->inlen >= offsetofend(typeof(*ucmd), sdb_addr));
790 }
791 
792 static inline bool user_qp_has_rdb(struct hns_roce_dev *hr_dev,
793 				   struct ib_qp_init_attr *init_attr,
794 				   struct ib_udata *udata,
795 				   struct hns_roce_ib_create_qp_resp *resp)
796 {
797 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
798 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
799 		hns_roce_qp_has_rq(init_attr));
800 }
801 
802 static inline bool kernel_qp_has_rdb(struct hns_roce_dev *hr_dev,
803 				     struct ib_qp_init_attr *init_attr)
804 {
805 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
806 		hns_roce_qp_has_rq(init_attr));
807 }
808 
809 static int qp_mmap_entry(struct hns_roce_qp *hr_qp,
810 			 struct hns_roce_dev *hr_dev,
811 			 struct ib_udata *udata,
812 			 struct hns_roce_ib_create_qp_resp *resp)
813 {
814 	struct hns_roce_ucontext *uctx =
815 		rdma_udata_to_drv_context(udata,
816 			struct hns_roce_ucontext, ibucontext);
817 	struct rdma_user_mmap_entry *rdma_entry;
818 	u64 address;
819 
820 	address = hr_dev->dwqe_page + hr_qp->qpn * HNS_ROCE_DWQE_SIZE;
821 
822 	hr_qp->dwqe_mmap_entry =
823 		hns_roce_user_mmap_entry_insert(&uctx->ibucontext, address,
824 						HNS_ROCE_DWQE_SIZE,
825 						HNS_ROCE_MMAP_TYPE_DWQE);
826 
827 	if (!hr_qp->dwqe_mmap_entry) {
828 		ibdev_err(&hr_dev->ib_dev, "failed to get dwqe mmap entry.\n");
829 		return -ENOMEM;
830 	}
831 
832 	rdma_entry = &hr_qp->dwqe_mmap_entry->rdma_entry;
833 	resp->dwqe_mmap_key = rdma_user_mmap_get_offset(rdma_entry);
834 
835 	return 0;
836 }
837 
838 static int alloc_user_qp_db(struct hns_roce_dev *hr_dev,
839 			    struct hns_roce_qp *hr_qp,
840 			    struct ib_qp_init_attr *init_attr,
841 			    struct ib_udata *udata,
842 			    struct hns_roce_ib_create_qp *ucmd,
843 			    struct hns_roce_ib_create_qp_resp *resp)
844 {
845 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(udata,
846 		struct hns_roce_ucontext, ibucontext);
847 	struct ib_device *ibdev = &hr_dev->ib_dev;
848 	int ret;
849 
850 	if (user_qp_has_sdb(hr_dev, init_attr, udata, resp, ucmd)) {
851 		ret = hns_roce_db_map_user(uctx, ucmd->sdb_addr, &hr_qp->sdb);
852 		if (ret) {
853 			ibdev_err(ibdev,
854 				  "failed to map user SQ doorbell, ret = %d.\n",
855 				  ret);
856 			goto err_out;
857 		}
858 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
859 	}
860 
861 	if (user_qp_has_rdb(hr_dev, init_attr, udata, resp)) {
862 		ret = hns_roce_db_map_user(uctx, ucmd->db_addr, &hr_qp->rdb);
863 		if (ret) {
864 			ibdev_err(ibdev,
865 				  "failed to map user RQ doorbell, ret = %d.\n",
866 				  ret);
867 			goto err_sdb;
868 		}
869 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
870 	}
871 
872 	return 0;
873 
874 err_sdb:
875 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
876 		hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
877 err_out:
878 	return ret;
879 }
880 
881 static int alloc_kernel_qp_db(struct hns_roce_dev *hr_dev,
882 			      struct hns_roce_qp *hr_qp,
883 			      struct ib_qp_init_attr *init_attr)
884 {
885 	struct ib_device *ibdev = &hr_dev->ib_dev;
886 	int ret;
887 
888 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
889 		hr_qp->sq.db_reg = hr_dev->mem_base +
890 				   HNS_ROCE_DWQE_SIZE * hr_qp->qpn;
891 	else
892 		hr_qp->sq.db_reg = hr_dev->reg_base + hr_dev->sdb_offset +
893 				   DB_REG_OFFSET * hr_dev->priv_uar.index;
894 
895 	hr_qp->rq.db_reg = hr_dev->reg_base + hr_dev->odb_offset +
896 			   DB_REG_OFFSET * hr_dev->priv_uar.index;
897 
898 	if (kernel_qp_has_rdb(hr_dev, init_attr)) {
899 		ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
900 		if (ret) {
901 			ibdev_err(ibdev,
902 				  "failed to alloc kernel RQ doorbell, ret = %d.\n",
903 				  ret);
904 			return ret;
905 		}
906 		*hr_qp->rdb.db_record = 0;
907 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
908 	}
909 
910 	return 0;
911 }
912 
913 static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
914 		       struct ib_qp_init_attr *init_attr,
915 		       struct ib_udata *udata,
916 		       struct hns_roce_ib_create_qp *ucmd,
917 		       struct hns_roce_ib_create_qp_resp *resp)
918 {
919 	int ret;
920 
921 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SDI_MODE)
922 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_OWNER_DB;
923 
924 	if (udata) {
925 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE) {
926 			ret = qp_mmap_entry(hr_qp, hr_dev, udata, resp);
927 			if (ret)
928 				return ret;
929 		}
930 
931 		ret = alloc_user_qp_db(hr_dev, hr_qp, init_attr, udata, ucmd,
932 				       resp);
933 		if (ret)
934 			goto err_remove_qp;
935 	} else {
936 		ret = alloc_kernel_qp_db(hr_dev, hr_qp, init_attr);
937 		if (ret)
938 			return ret;
939 	}
940 
941 	return 0;
942 
943 err_remove_qp:
944 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
945 		qp_user_mmap_entry_remove(hr_qp);
946 
947 	return ret;
948 }
949 
950 static void free_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
951 		       struct ib_udata *udata)
952 {
953 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
954 		udata, struct hns_roce_ucontext, ibucontext);
955 
956 	if (udata) {
957 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
958 			hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
959 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
960 			hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
961 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
962 			qp_user_mmap_entry_remove(hr_qp);
963 	} else {
964 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
965 			hns_roce_free_db(hr_dev, &hr_qp->rdb);
966 	}
967 }
968 
969 static int alloc_kernel_wrid(struct hns_roce_dev *hr_dev,
970 			     struct hns_roce_qp *hr_qp)
971 {
972 	struct ib_device *ibdev = &hr_dev->ib_dev;
973 	u64 *sq_wrid = NULL;
974 	u64 *rq_wrid = NULL;
975 	int ret;
976 
977 	sq_wrid = kcalloc(hr_qp->sq.wqe_cnt, sizeof(u64), GFP_KERNEL);
978 	if (ZERO_OR_NULL_PTR(sq_wrid)) {
979 		ibdev_err(ibdev, "failed to alloc SQ wrid.\n");
980 		return -ENOMEM;
981 	}
982 
983 	if (hr_qp->rq.wqe_cnt) {
984 		rq_wrid = kcalloc(hr_qp->rq.wqe_cnt, sizeof(u64), GFP_KERNEL);
985 		if (ZERO_OR_NULL_PTR(rq_wrid)) {
986 			ibdev_err(ibdev, "failed to alloc RQ wrid.\n");
987 			ret = -ENOMEM;
988 			goto err_sq;
989 		}
990 	}
991 
992 	hr_qp->sq.wrid = sq_wrid;
993 	hr_qp->rq.wrid = rq_wrid;
994 	return 0;
995 err_sq:
996 	kfree(sq_wrid);
997 
998 	return ret;
999 }
1000 
1001 static void free_kernel_wrid(struct hns_roce_qp *hr_qp)
1002 {
1003 	kfree(hr_qp->rq.wrid);
1004 	kfree(hr_qp->sq.wrid);
1005 }
1006 
1007 static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1008 			struct ib_qp_init_attr *init_attr,
1009 			struct ib_udata *udata,
1010 			struct hns_roce_ib_create_qp *ucmd)
1011 {
1012 	struct ib_device *ibdev = &hr_dev->ib_dev;
1013 	struct hns_roce_ucontext *uctx;
1014 	int ret;
1015 
1016 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1017 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
1018 	else
1019 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
1020 
1021 	ret = set_rq_size(hr_dev, &init_attr->cap, hr_qp,
1022 			  hns_roce_qp_has_rq(init_attr), !!udata);
1023 	if (ret) {
1024 		ibdev_err(ibdev, "failed to set user RQ size, ret = %d.\n",
1025 			  ret);
1026 		return ret;
1027 	}
1028 
1029 	if (udata) {
1030 		ret = ib_copy_from_udata(ucmd, udata,
1031 					 min(udata->inlen, sizeof(*ucmd)));
1032 		if (ret) {
1033 			ibdev_err(ibdev,
1034 				  "failed to copy QP ucmd, ret = %d\n", ret);
1035 			return ret;
1036 		}
1037 
1038 		uctx = rdma_udata_to_drv_context(udata, struct hns_roce_ucontext,
1039 						 ibucontext);
1040 		hr_qp->config = uctx->config;
1041 		ret = set_user_sq_size(hr_dev, &init_attr->cap, hr_qp, ucmd);
1042 		if (ret)
1043 			ibdev_err(ibdev,
1044 				  "failed to set user SQ size, ret = %d.\n",
1045 				  ret);
1046 	} else {
1047 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
1048 			hr_qp->config = HNS_ROCE_EXSGE_FLAGS;
1049 		ret = set_kernel_sq_size(hr_dev, &init_attr->cap, hr_qp);
1050 		if (ret)
1051 			ibdev_err(ibdev,
1052 				  "failed to set kernel SQ size, ret = %d.\n",
1053 				  ret);
1054 	}
1055 
1056 	return ret;
1057 }
1058 
1059 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
1060 				     struct ib_pd *ib_pd,
1061 				     struct ib_qp_init_attr *init_attr,
1062 				     struct ib_udata *udata,
1063 				     struct hns_roce_qp *hr_qp)
1064 {
1065 	struct hns_roce_ib_create_qp_resp resp = {};
1066 	struct ib_device *ibdev = &hr_dev->ib_dev;
1067 	struct hns_roce_ib_create_qp ucmd = {};
1068 	int ret;
1069 
1070 	mutex_init(&hr_qp->mutex);
1071 	spin_lock_init(&hr_qp->sq.lock);
1072 	spin_lock_init(&hr_qp->rq.lock);
1073 
1074 	hr_qp->state = IB_QPS_RESET;
1075 	hr_qp->flush_flag = 0;
1076 
1077 	if (init_attr->create_flags)
1078 		return -EOPNOTSUPP;
1079 
1080 	ret = set_qp_param(hr_dev, hr_qp, init_attr, udata, &ucmd);
1081 	if (ret) {
1082 		ibdev_err(ibdev, "failed to set QP param, ret = %d.\n", ret);
1083 		return ret;
1084 	}
1085 
1086 	if (!udata) {
1087 		ret = alloc_kernel_wrid(hr_dev, hr_qp);
1088 		if (ret) {
1089 			ibdev_err(ibdev, "failed to alloc wrid, ret = %d.\n",
1090 				  ret);
1091 			return ret;
1092 		}
1093 	}
1094 
1095 	ret = alloc_qp_buf(hr_dev, hr_qp, init_attr, udata, ucmd.buf_addr);
1096 	if (ret) {
1097 		ibdev_err(ibdev, "failed to alloc QP buffer, ret = %d.\n", ret);
1098 		goto err_buf;
1099 	}
1100 
1101 	ret = alloc_qpn(hr_dev, hr_qp, init_attr);
1102 	if (ret) {
1103 		ibdev_err(ibdev, "failed to alloc QPN, ret = %d.\n", ret);
1104 		goto err_qpn;
1105 	}
1106 
1107 	ret = alloc_qp_db(hr_dev, hr_qp, init_attr, udata, &ucmd, &resp);
1108 	if (ret) {
1109 		ibdev_err(ibdev, "failed to alloc QP doorbell, ret = %d.\n",
1110 			  ret);
1111 		goto err_db;
1112 	}
1113 
1114 	ret = alloc_qpc(hr_dev, hr_qp);
1115 	if (ret) {
1116 		ibdev_err(ibdev, "failed to alloc QP context, ret = %d.\n",
1117 			  ret);
1118 		goto err_qpc;
1119 	}
1120 
1121 	ret = hns_roce_qp_store(hr_dev, hr_qp, init_attr);
1122 	if (ret) {
1123 		ibdev_err(ibdev, "failed to store QP, ret = %d.\n", ret);
1124 		goto err_store;
1125 	}
1126 
1127 	if (udata) {
1128 		resp.cap_flags = hr_qp->en_flags;
1129 		ret = ib_copy_to_udata(udata, &resp,
1130 				       min(udata->outlen, sizeof(resp)));
1131 		if (ret) {
1132 			ibdev_err(ibdev, "copy qp resp failed!\n");
1133 			goto err_store;
1134 		}
1135 	}
1136 
1137 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
1138 		ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
1139 		if (ret)
1140 			goto err_flow_ctrl;
1141 	}
1142 
1143 	hr_qp->ibqp.qp_num = hr_qp->qpn;
1144 	hr_qp->event = hns_roce_ib_qp_event;
1145 	refcount_set(&hr_qp->refcount, 1);
1146 	init_completion(&hr_qp->free);
1147 
1148 	return 0;
1149 
1150 err_flow_ctrl:
1151 	hns_roce_qp_remove(hr_dev, hr_qp);
1152 err_store:
1153 	free_qpc(hr_dev, hr_qp);
1154 err_qpc:
1155 	free_qp_db(hr_dev, hr_qp, udata);
1156 err_db:
1157 	free_qpn(hr_dev, hr_qp);
1158 err_qpn:
1159 	free_qp_buf(hr_dev, hr_qp);
1160 err_buf:
1161 	free_kernel_wrid(hr_qp);
1162 	return ret;
1163 }
1164 
1165 void hns_roce_qp_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1166 			 struct ib_udata *udata)
1167 {
1168 	if (refcount_dec_and_test(&hr_qp->refcount))
1169 		complete(&hr_qp->free);
1170 	wait_for_completion(&hr_qp->free);
1171 
1172 	free_qpc(hr_dev, hr_qp);
1173 	free_qpn(hr_dev, hr_qp);
1174 	free_qp_buf(hr_dev, hr_qp);
1175 	free_kernel_wrid(hr_qp);
1176 	free_qp_db(hr_dev, hr_qp, udata);
1177 }
1178 
1179 static int check_qp_type(struct hns_roce_dev *hr_dev, enum ib_qp_type type,
1180 			 bool is_user)
1181 {
1182 	switch (type) {
1183 	case IB_QPT_XRC_INI:
1184 	case IB_QPT_XRC_TGT:
1185 		if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC))
1186 			goto out;
1187 		break;
1188 	case IB_QPT_UD:
1189 		if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 &&
1190 		    is_user)
1191 			goto out;
1192 		break;
1193 	case IB_QPT_RC:
1194 	case IB_QPT_GSI:
1195 		break;
1196 	default:
1197 		goto out;
1198 	}
1199 
1200 	return 0;
1201 
1202 out:
1203 	ibdev_err(&hr_dev->ib_dev, "not support QP type %d\n", type);
1204 
1205 	return -EOPNOTSUPP;
1206 }
1207 
1208 int hns_roce_create_qp(struct ib_qp *qp, struct ib_qp_init_attr *init_attr,
1209 		       struct ib_udata *udata)
1210 {
1211 	struct ib_device *ibdev = qp->device;
1212 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
1213 	struct hns_roce_qp *hr_qp = to_hr_qp(qp);
1214 	struct ib_pd *pd = qp->pd;
1215 	int ret;
1216 
1217 	ret = check_qp_type(hr_dev, init_attr->qp_type, !!udata);
1218 	if (ret)
1219 		goto err_out;
1220 
1221 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1222 		hr_qp->xrcdn = to_hr_xrcd(init_attr->xrcd)->xrcdn;
1223 
1224 	if (init_attr->qp_type == IB_QPT_GSI) {
1225 		hr_qp->port = init_attr->port_num - 1;
1226 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
1227 	}
1228 
1229 	ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, hr_qp);
1230 	if (ret)
1231 		ibdev_err(ibdev, "create QP type 0x%x failed(%d)\n",
1232 			  init_attr->qp_type, ret);
1233 
1234 err_out:
1235 	if (ret)
1236 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_QP_CREATE_ERR_CNT]);
1237 
1238 	return ret;
1239 }
1240 
1241 int to_hr_qp_type(int qp_type)
1242 {
1243 	switch (qp_type) {
1244 	case IB_QPT_RC:
1245 		return SERV_TYPE_RC;
1246 	case IB_QPT_UD:
1247 	case IB_QPT_GSI:
1248 		return SERV_TYPE_UD;
1249 	case IB_QPT_XRC_INI:
1250 	case IB_QPT_XRC_TGT:
1251 		return SERV_TYPE_XRC;
1252 	default:
1253 		return -1;
1254 	}
1255 }
1256 
1257 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1258 			      struct hns_roce_qp *hr_qp,
1259 			      struct ib_qp_attr *attr, int attr_mask)
1260 {
1261 	enum ib_mtu active_mtu;
1262 	int p;
1263 
1264 	p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1265 	active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1266 
1267 	if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1268 	    attr->path_mtu > hr_dev->caps.max_mtu) ||
1269 	    attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1270 		ibdev_err(&hr_dev->ib_dev,
1271 			"attr path_mtu(%d)invalid while modify qp",
1272 			attr->path_mtu);
1273 		return -EINVAL;
1274 	}
1275 
1276 	return 0;
1277 }
1278 
1279 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1280 				  int attr_mask)
1281 {
1282 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1283 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1284 	int p;
1285 
1286 	if ((attr_mask & IB_QP_PORT) &&
1287 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1288 		ibdev_err(&hr_dev->ib_dev, "invalid attr, port_num = %u.\n",
1289 			  attr->port_num);
1290 		return -EINVAL;
1291 	}
1292 
1293 	if (attr_mask & IB_QP_PKEY_INDEX) {
1294 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1295 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1296 			ibdev_err(&hr_dev->ib_dev,
1297 				  "invalid attr, pkey_index = %u.\n",
1298 				  attr->pkey_index);
1299 			return -EINVAL;
1300 		}
1301 	}
1302 
1303 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1304 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1305 		ibdev_err(&hr_dev->ib_dev,
1306 			  "invalid attr, max_rd_atomic = %u.\n",
1307 			  attr->max_rd_atomic);
1308 		return -EINVAL;
1309 	}
1310 
1311 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1312 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1313 		ibdev_err(&hr_dev->ib_dev,
1314 			  "invalid attr, max_dest_rd_atomic = %u.\n",
1315 			  attr->max_dest_rd_atomic);
1316 		return -EINVAL;
1317 	}
1318 
1319 	if (attr_mask & IB_QP_PATH_MTU)
1320 		return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1321 
1322 	return 0;
1323 }
1324 
1325 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1326 		       int attr_mask, struct ib_udata *udata)
1327 {
1328 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1329 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1330 	enum ib_qp_state cur_state, new_state;
1331 	int ret = -EINVAL;
1332 
1333 	mutex_lock(&hr_qp->mutex);
1334 
1335 	if (attr_mask & IB_QP_CUR_STATE && attr->cur_qp_state != hr_qp->state)
1336 		goto out;
1337 
1338 	cur_state = hr_qp->state;
1339 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1340 
1341 	if (ibqp->uobject &&
1342 	    (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1343 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB) {
1344 			hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1345 
1346 			if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
1347 				hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1348 		} else {
1349 			ibdev_warn(&hr_dev->ib_dev,
1350 				  "flush cqe is not supported in userspace!\n");
1351 			goto out;
1352 		}
1353 	}
1354 
1355 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1356 				attr_mask)) {
1357 		ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1358 		goto out;
1359 	}
1360 
1361 	ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1362 	if (ret)
1363 		goto out;
1364 
1365 	if (cur_state == new_state && cur_state == IB_QPS_RESET)
1366 		goto out;
1367 
1368 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1369 				    new_state, udata);
1370 
1371 out:
1372 	mutex_unlock(&hr_qp->mutex);
1373 	if (ret)
1374 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_QP_MODIFY_ERR_CNT]);
1375 
1376 	return ret;
1377 }
1378 
1379 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1380 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1381 {
1382 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1383 		__acquire(&send_cq->lock);
1384 		__acquire(&recv_cq->lock);
1385 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1386 		spin_lock_irq(&send_cq->lock);
1387 		__acquire(&recv_cq->lock);
1388 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1389 		spin_lock_irq(&recv_cq->lock);
1390 		__acquire(&send_cq->lock);
1391 	} else if (send_cq == recv_cq) {
1392 		spin_lock_irq(&send_cq->lock);
1393 		__acquire(&recv_cq->lock);
1394 	} else if (send_cq->cqn < recv_cq->cqn) {
1395 		spin_lock_irq(&send_cq->lock);
1396 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1397 	} else {
1398 		spin_lock_irq(&recv_cq->lock);
1399 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1400 	}
1401 }
1402 
1403 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1404 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1405 			 __releases(&recv_cq->lock)
1406 {
1407 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1408 		__release(&recv_cq->lock);
1409 		__release(&send_cq->lock);
1410 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1411 		__release(&recv_cq->lock);
1412 		spin_unlock(&send_cq->lock);
1413 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1414 		__release(&send_cq->lock);
1415 		spin_unlock(&recv_cq->lock);
1416 	} else if (send_cq == recv_cq) {
1417 		__release(&recv_cq->lock);
1418 		spin_unlock_irq(&send_cq->lock);
1419 	} else if (send_cq->cqn < recv_cq->cqn) {
1420 		spin_unlock(&recv_cq->lock);
1421 		spin_unlock_irq(&send_cq->lock);
1422 	} else {
1423 		spin_unlock(&send_cq->lock);
1424 		spin_unlock_irq(&recv_cq->lock);
1425 	}
1426 }
1427 
1428 static inline void *get_wqe(struct hns_roce_qp *hr_qp, u32 offset)
1429 {
1430 	return hns_roce_buf_offset(hr_qp->mtr.kmem, offset);
1431 }
1432 
1433 void *hns_roce_get_recv_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1434 {
1435 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1436 }
1437 
1438 void *hns_roce_get_send_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1439 {
1440 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1441 }
1442 
1443 void *hns_roce_get_extend_sge(struct hns_roce_qp *hr_qp, unsigned int n)
1444 {
1445 	return get_wqe(hr_qp, hr_qp->sge.offset + (n << hr_qp->sge.sge_shift));
1446 }
1447 
1448 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, u32 nreq,
1449 			  struct ib_cq *ib_cq)
1450 {
1451 	struct hns_roce_cq *hr_cq;
1452 	u32 cur;
1453 
1454 	cur = hr_wq->head - hr_wq->tail;
1455 	if (likely(cur + nreq < hr_wq->wqe_cnt))
1456 		return false;
1457 
1458 	hr_cq = to_hr_cq(ib_cq);
1459 	spin_lock(&hr_cq->lock);
1460 	cur = hr_wq->head - hr_wq->tail;
1461 	spin_unlock(&hr_cq->lock);
1462 
1463 	return cur + nreq >= hr_wq->wqe_cnt;
1464 }
1465 
1466 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1467 {
1468 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1469 	unsigned int reserved_from_bot;
1470 	unsigned int i;
1471 
1472 	qp_table->idx_table.spare_idx = kcalloc(hr_dev->caps.num_qps,
1473 					sizeof(u32), GFP_KERNEL);
1474 	if (!qp_table->idx_table.spare_idx)
1475 		return -ENOMEM;
1476 
1477 	mutex_init(&qp_table->scc_mutex);
1478 	mutex_init(&qp_table->bank_mutex);
1479 	xa_init(&hr_dev->qp_table_xa);
1480 
1481 	reserved_from_bot = hr_dev->caps.reserved_qps;
1482 
1483 	for (i = 0; i < reserved_from_bot; i++) {
1484 		hr_dev->qp_table.bank[get_qp_bankid(i)].inuse++;
1485 		hr_dev->qp_table.bank[get_qp_bankid(i)].min++;
1486 	}
1487 
1488 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) {
1489 		ida_init(&hr_dev->qp_table.bank[i].ida);
1490 		hr_dev->qp_table.bank[i].max = hr_dev->caps.num_qps /
1491 					       HNS_ROCE_QP_BANK_NUM - 1;
1492 		hr_dev->qp_table.bank[i].next = hr_dev->qp_table.bank[i].min;
1493 	}
1494 
1495 	return 0;
1496 }
1497 
1498 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1499 {
1500 	int i;
1501 
1502 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++)
1503 		ida_destroy(&hr_dev->qp_table.bank[i].ida);
1504 	kfree(hr_dev->qp_table.idx_table.spare_idx);
1505 }
1506