xref: /linux/drivers/infiniband/hw/hns/hns_roce_qp.c (revision 903a7d37d9ea03cfed21040467d3d345d1e6fc76)
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,
414 		 hr_qp->qpn / HNS_ROCE_QP_BANK_NUM);
415 
416 	mutex_lock(&hr_dev->qp_table.bank_mutex);
417 	hr_dev->qp_table.bank[bankid].inuse--;
418 	mutex_unlock(&hr_dev->qp_table.bank_mutex);
419 }
420 
421 static u32 proc_rq_sge(struct hns_roce_dev *dev, struct hns_roce_qp *hr_qp,
422 		       bool user)
423 {
424 	u32 max_sge = dev->caps.max_rq_sg;
425 
426 	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
427 		return max_sge;
428 
429 	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
430 	 * calculate number of max_sge with reserved SGEs when allocating wqe
431 	 * buf, so there is no need to do this again in kernel. But the number
432 	 * may exceed the capacity of SGEs recorded in the firmware, so the
433 	 * kernel driver should just adapt the value accordingly.
434 	 */
435 	if (user)
436 		max_sge = roundup_pow_of_two(max_sge + 1);
437 	else
438 		hr_qp->rq.rsv_sge = 1;
439 
440 	return max_sge;
441 }
442 
443 static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
444 		       struct hns_roce_qp *hr_qp, int has_rq, bool user)
445 {
446 	u32 max_sge = proc_rq_sge(hr_dev, hr_qp, user);
447 	u32 cnt;
448 
449 	/* If srq exist, set zero for relative number of rq */
450 	if (!has_rq) {
451 		hr_qp->rq.wqe_cnt = 0;
452 		hr_qp->rq.max_gs = 0;
453 		cap->max_recv_wr = 0;
454 		cap->max_recv_sge = 0;
455 
456 		return 0;
457 	}
458 
459 	/* Check the validity of QP support capacity */
460 	if (!cap->max_recv_wr || cap->max_recv_wr > hr_dev->caps.max_wqes ||
461 	    cap->max_recv_sge > max_sge) {
462 		ibdev_err(&hr_dev->ib_dev,
463 			  "RQ config error, depth = %u, sge = %u\n",
464 			  cap->max_recv_wr, cap->max_recv_sge);
465 		return -EINVAL;
466 	}
467 
468 	cnt = roundup_pow_of_two(max(cap->max_recv_wr, hr_dev->caps.min_wqes));
469 	if (cnt > hr_dev->caps.max_wqes) {
470 		ibdev_err(&hr_dev->ib_dev, "rq depth %u too large\n",
471 			  cap->max_recv_wr);
472 		return -EINVAL;
473 	}
474 
475 	hr_qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge) +
476 					      hr_qp->rq.rsv_sge);
477 
478 	hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz *
479 				    hr_qp->rq.max_gs);
480 
481 	hr_qp->rq.wqe_cnt = cnt;
482 
483 	cap->max_recv_wr = cnt;
484 	cap->max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
485 
486 	return 0;
487 }
488 
489 static u32 get_max_inline_data(struct hns_roce_dev *hr_dev,
490 			       struct ib_qp_cap *cap)
491 {
492 	if (cap->max_inline_data) {
493 		cap->max_inline_data = roundup_pow_of_two(cap->max_inline_data);
494 		return min(cap->max_inline_data,
495 			   hr_dev->caps.max_sq_inline);
496 	}
497 
498 	return 0;
499 }
500 
501 static void update_inline_data(struct hns_roce_qp *hr_qp,
502 			       struct ib_qp_cap *cap)
503 {
504 	u32 sge_num = hr_qp->sq.ext_sge_cnt;
505 
506 	if (hr_qp->config & HNS_ROCE_EXSGE_FLAGS) {
507 		if (!(hr_qp->ibqp.qp_type == IB_QPT_GSI ||
508 		      hr_qp->ibqp.qp_type == IB_QPT_UD))
509 			sge_num = max((u32)HNS_ROCE_SGE_IN_WQE, sge_num);
510 
511 		cap->max_inline_data = max(cap->max_inline_data,
512 					   sge_num * HNS_ROCE_SGE_SIZE);
513 	}
514 
515 	hr_qp->max_inline_data = cap->max_inline_data;
516 }
517 
518 static u32 get_sge_num_from_max_send_sge(bool is_ud_or_gsi,
519 					 u32 max_send_sge)
520 {
521 	unsigned int std_sge_num;
522 	unsigned int min_sge;
523 
524 	std_sge_num = is_ud_or_gsi ? 0 : HNS_ROCE_SGE_IN_WQE;
525 	min_sge = is_ud_or_gsi ? 1 : 0;
526 	return max_send_sge > std_sge_num ? (max_send_sge - std_sge_num) :
527 				min_sge;
528 }
529 
530 static unsigned int get_sge_num_from_max_inl_data(bool is_ud_or_gsi,
531 						  u32 max_inline_data)
532 {
533 	unsigned int inline_sge;
534 
535 	if (!max_inline_data)
536 		return 0;
537 
538 	/*
539 	 * if max_inline_data less than
540 	 * HNS_ROCE_SGE_IN_WQE * HNS_ROCE_SGE_SIZE,
541 	 * In addition to ud's mode, no need to extend sge.
542 	 */
543 	inline_sge = roundup_pow_of_two(max_inline_data) / HNS_ROCE_SGE_SIZE;
544 	if (!is_ud_or_gsi && inline_sge <= HNS_ROCE_SGE_IN_WQE)
545 		inline_sge = 0;
546 
547 	return inline_sge;
548 }
549 
550 static void set_ext_sge_param(struct hns_roce_dev *hr_dev, u32 sq_wqe_cnt,
551 			      struct hns_roce_qp *hr_qp, struct ib_qp_cap *cap)
552 {
553 	bool is_ud_or_gsi = (hr_qp->ibqp.qp_type == IB_QPT_GSI ||
554 				hr_qp->ibqp.qp_type == IB_QPT_UD);
555 	unsigned int std_sge_num;
556 	u32 inline_ext_sge = 0;
557 	u32 ext_wqe_sge_cnt;
558 	u32 total_sge_cnt;
559 
560 	cap->max_inline_data = get_max_inline_data(hr_dev, cap);
561 
562 	hr_qp->sge.sge_shift = HNS_ROCE_SGE_SHIFT;
563 	std_sge_num = is_ud_or_gsi ? 0 : HNS_ROCE_SGE_IN_WQE;
564 	ext_wqe_sge_cnt = get_sge_num_from_max_send_sge(is_ud_or_gsi,
565 							cap->max_send_sge);
566 
567 	if (hr_qp->config & HNS_ROCE_EXSGE_FLAGS) {
568 		inline_ext_sge = max(ext_wqe_sge_cnt,
569 				     get_sge_num_from_max_inl_data(is_ud_or_gsi,
570 							 cap->max_inline_data));
571 		hr_qp->sq.ext_sge_cnt = inline_ext_sge ?
572 					roundup_pow_of_two(inline_ext_sge) : 0;
573 
574 		hr_qp->sq.max_gs = max(1U, (hr_qp->sq.ext_sge_cnt + std_sge_num));
575 		hr_qp->sq.max_gs = min(hr_qp->sq.max_gs, hr_dev->caps.max_sq_sg);
576 
577 		ext_wqe_sge_cnt = hr_qp->sq.ext_sge_cnt;
578 	} else {
579 		hr_qp->sq.max_gs = max(1U, cap->max_send_sge);
580 		hr_qp->sq.max_gs = min(hr_qp->sq.max_gs, hr_dev->caps.max_sq_sg);
581 		hr_qp->sq.ext_sge_cnt = hr_qp->sq.max_gs;
582 	}
583 
584 	/* If the number of extended sge is not zero, they MUST use the
585 	 * space of HNS_HW_PAGE_SIZE at least.
586 	 */
587 	if (ext_wqe_sge_cnt) {
588 		total_sge_cnt = roundup_pow_of_two(sq_wqe_cnt * ext_wqe_sge_cnt);
589 		hr_qp->sge.sge_cnt = max(total_sge_cnt,
590 				(u32)HNS_HW_PAGE_SIZE / HNS_ROCE_SGE_SIZE);
591 	}
592 
593 	update_inline_data(hr_qp, cap);
594 }
595 
596 static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
597 					struct ib_qp_cap *cap,
598 					struct hns_roce_ib_create_qp *ucmd)
599 {
600 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
601 	u8 max_sq_stride = ilog2(roundup_sq_stride);
602 
603 	/* Sanity check SQ size before proceeding */
604 	if (ucmd->log_sq_stride > max_sq_stride ||
605 	    ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
606 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ stride size.\n");
607 		return -EINVAL;
608 	}
609 
610 	if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
611 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ SGE size %u.\n",
612 			  cap->max_send_sge);
613 		return -EINVAL;
614 	}
615 
616 	return 0;
617 }
618 
619 static int set_user_sq_size(struct hns_roce_dev *hr_dev,
620 			    struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp,
621 			    struct hns_roce_ib_create_qp *ucmd)
622 {
623 	struct ib_device *ibdev = &hr_dev->ib_dev;
624 	u32 cnt = 0;
625 	int ret;
626 
627 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
628 	    cnt > hr_dev->caps.max_wqes)
629 		return -EINVAL;
630 
631 	ret = check_sq_size_with_integrity(hr_dev, cap, ucmd);
632 	if (ret) {
633 		ibdev_err(ibdev, "failed to check user SQ size, ret = %d.\n",
634 			  ret);
635 		return ret;
636 	}
637 
638 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
639 
640 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
641 	hr_qp->sq.wqe_cnt = cnt;
642 	cap->max_send_sge = hr_qp->sq.max_gs;
643 
644 	return 0;
645 }
646 
647 static int set_wqe_buf_attr(struct hns_roce_dev *hr_dev,
648 			    struct hns_roce_qp *hr_qp,
649 			    struct hns_roce_buf_attr *buf_attr)
650 {
651 	int buf_size;
652 	int idx = 0;
653 
654 	hr_qp->buff_size = 0;
655 
656 	/* SQ WQE */
657 	hr_qp->sq.offset = 0;
658 	buf_size = to_hr_hem_entries_size(hr_qp->sq.wqe_cnt,
659 					  hr_qp->sq.wqe_shift);
660 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
661 		buf_attr->region[idx].size = buf_size;
662 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sq_hop_num;
663 		idx++;
664 		hr_qp->buff_size += buf_size;
665 	}
666 
667 	/* extend SGE WQE in SQ */
668 	hr_qp->sge.offset = hr_qp->buff_size;
669 	buf_size = to_hr_hem_entries_size(hr_qp->sge.sge_cnt,
670 					  hr_qp->sge.sge_shift);
671 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
672 		buf_attr->region[idx].size = buf_size;
673 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sge_hop_num;
674 		idx++;
675 		hr_qp->buff_size += buf_size;
676 	}
677 
678 	/* RQ WQE */
679 	hr_qp->rq.offset = hr_qp->buff_size;
680 	buf_size = to_hr_hem_entries_size(hr_qp->rq.wqe_cnt,
681 					  hr_qp->rq.wqe_shift);
682 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
683 		buf_attr->region[idx].size = buf_size;
684 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_rq_hop_num;
685 		idx++;
686 		hr_qp->buff_size += buf_size;
687 	}
688 
689 	if (hr_qp->buff_size < 1)
690 		return -EINVAL;
691 
692 	buf_attr->page_shift = HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
693 	buf_attr->region_count = idx;
694 
695 	return 0;
696 }
697 
698 static int set_kernel_sq_size(struct hns_roce_dev *hr_dev,
699 			      struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp)
700 {
701 	struct ib_device *ibdev = &hr_dev->ib_dev;
702 	u32 cnt;
703 
704 	if (!cap->max_send_wr || cap->max_send_wr > hr_dev->caps.max_wqes ||
705 	    cap->max_send_sge > hr_dev->caps.max_sq_sg) {
706 		ibdev_err(ibdev, "failed to check SQ WR or SGE num.\n");
707 		return -EINVAL;
708 	}
709 
710 	cnt = roundup_pow_of_two(max(cap->max_send_wr, hr_dev->caps.min_wqes));
711 	if (cnt > hr_dev->caps.max_wqes) {
712 		ibdev_err(ibdev, "failed to check WQE num, WQE num = %u.\n",
713 			  cnt);
714 		return -EINVAL;
715 	}
716 
717 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
718 	hr_qp->sq.wqe_cnt = cnt;
719 
720 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
721 
722 	/* sync the parameters of kernel QP to user's configuration */
723 	cap->max_send_wr = cnt;
724 	cap->max_send_sge = hr_qp->sq.max_gs;
725 
726 	return 0;
727 }
728 
729 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
730 {
731 	if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
732 		return 0;
733 
734 	return 1;
735 }
736 
737 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
738 {
739 	if (attr->qp_type == IB_QPT_XRC_INI ||
740 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
741 	    !attr->cap.max_recv_wr)
742 		return 0;
743 
744 	return 1;
745 }
746 
747 static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
748 			struct ib_qp_init_attr *init_attr,
749 			struct ib_udata *udata, unsigned long addr)
750 {
751 	struct ib_device *ibdev = &hr_dev->ib_dev;
752 	struct hns_roce_buf_attr buf_attr = {};
753 	int ret;
754 
755 	ret = set_wqe_buf_attr(hr_dev, hr_qp, &buf_attr);
756 	if (ret) {
757 		ibdev_err(ibdev, "failed to split WQE buf, ret = %d.\n", ret);
758 		goto err_inline;
759 	}
760 	ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, &buf_attr,
761 				  PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
762 				  udata, addr);
763 	if (ret) {
764 		ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret);
765 		goto err_inline;
766 	}
767 
768 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_DIRECT_WQE)
769 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_DIRECT_WQE;
770 
771 	return 0;
772 
773 err_inline:
774 
775 	return ret;
776 }
777 
778 static void free_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
779 {
780 	hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
781 }
782 
783 static inline bool user_qp_has_sdb(struct hns_roce_dev *hr_dev,
784 				   struct ib_qp_init_attr *init_attr,
785 				   struct ib_udata *udata,
786 				   struct hns_roce_ib_create_qp_resp *resp,
787 				   struct hns_roce_ib_create_qp *ucmd)
788 {
789 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
790 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
791 		hns_roce_qp_has_sq(init_attr) &&
792 		udata->inlen >= offsetofend(typeof(*ucmd), sdb_addr));
793 }
794 
795 static inline bool user_qp_has_rdb(struct hns_roce_dev *hr_dev,
796 				   struct ib_qp_init_attr *init_attr,
797 				   struct ib_udata *udata,
798 				   struct hns_roce_ib_create_qp_resp *resp)
799 {
800 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
801 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
802 		hns_roce_qp_has_rq(init_attr));
803 }
804 
805 static inline bool kernel_qp_has_rdb(struct hns_roce_dev *hr_dev,
806 				     struct ib_qp_init_attr *init_attr)
807 {
808 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
809 		hns_roce_qp_has_rq(init_attr));
810 }
811 
812 static int qp_mmap_entry(struct hns_roce_qp *hr_qp,
813 			 struct hns_roce_dev *hr_dev,
814 			 struct ib_udata *udata,
815 			 struct hns_roce_ib_create_qp_resp *resp)
816 {
817 	struct hns_roce_ucontext *uctx =
818 		rdma_udata_to_drv_context(udata,
819 			struct hns_roce_ucontext, ibucontext);
820 	struct rdma_user_mmap_entry *rdma_entry;
821 	u64 address;
822 
823 	address = hr_dev->dwqe_page + hr_qp->qpn * HNS_ROCE_DWQE_SIZE;
824 
825 	hr_qp->dwqe_mmap_entry =
826 		hns_roce_user_mmap_entry_insert(&uctx->ibucontext, address,
827 						HNS_ROCE_DWQE_SIZE,
828 						HNS_ROCE_MMAP_TYPE_DWQE);
829 
830 	if (!hr_qp->dwqe_mmap_entry) {
831 		ibdev_err(&hr_dev->ib_dev, "failed to get dwqe mmap entry.\n");
832 		return -ENOMEM;
833 	}
834 
835 	rdma_entry = &hr_qp->dwqe_mmap_entry->rdma_entry;
836 	resp->dwqe_mmap_key = rdma_user_mmap_get_offset(rdma_entry);
837 
838 	return 0;
839 }
840 
841 static int alloc_user_qp_db(struct hns_roce_dev *hr_dev,
842 			    struct hns_roce_qp *hr_qp,
843 			    struct ib_qp_init_attr *init_attr,
844 			    struct ib_udata *udata,
845 			    struct hns_roce_ib_create_qp *ucmd,
846 			    struct hns_roce_ib_create_qp_resp *resp)
847 {
848 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(udata,
849 		struct hns_roce_ucontext, ibucontext);
850 	struct ib_device *ibdev = &hr_dev->ib_dev;
851 	int ret;
852 
853 	if (user_qp_has_sdb(hr_dev, init_attr, udata, resp, ucmd)) {
854 		ret = hns_roce_db_map_user(uctx, ucmd->sdb_addr, &hr_qp->sdb);
855 		if (ret) {
856 			ibdev_err(ibdev,
857 				  "failed to map user SQ doorbell, ret = %d.\n",
858 				  ret);
859 			goto err_out;
860 		}
861 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
862 	}
863 
864 	if (user_qp_has_rdb(hr_dev, init_attr, udata, resp)) {
865 		ret = hns_roce_db_map_user(uctx, ucmd->db_addr, &hr_qp->rdb);
866 		if (ret) {
867 			ibdev_err(ibdev,
868 				  "failed to map user RQ doorbell, ret = %d.\n",
869 				  ret);
870 			goto err_sdb;
871 		}
872 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
873 	}
874 
875 	return 0;
876 
877 err_sdb:
878 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
879 		hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
880 err_out:
881 	return ret;
882 }
883 
884 static int alloc_kernel_qp_db(struct hns_roce_dev *hr_dev,
885 			      struct hns_roce_qp *hr_qp,
886 			      struct ib_qp_init_attr *init_attr)
887 {
888 	struct ib_device *ibdev = &hr_dev->ib_dev;
889 	int ret;
890 
891 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
892 		hr_qp->sq.db_reg = hr_dev->mem_base +
893 				   HNS_ROCE_DWQE_SIZE * hr_qp->qpn;
894 	else
895 		hr_qp->sq.db_reg = hr_dev->reg_base + hr_dev->sdb_offset +
896 				   DB_REG_OFFSET * hr_dev->priv_uar.index;
897 
898 	hr_qp->rq.db_reg = hr_dev->reg_base + hr_dev->odb_offset +
899 			   DB_REG_OFFSET * hr_dev->priv_uar.index;
900 
901 	if (kernel_qp_has_rdb(hr_dev, init_attr)) {
902 		ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
903 		if (ret) {
904 			ibdev_err(ibdev,
905 				  "failed to alloc kernel RQ doorbell, ret = %d.\n",
906 				  ret);
907 			return ret;
908 		}
909 		*hr_qp->rdb.db_record = 0;
910 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
911 	}
912 
913 	return 0;
914 }
915 
916 static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
917 		       struct ib_qp_init_attr *init_attr,
918 		       struct ib_udata *udata,
919 		       struct hns_roce_ib_create_qp *ucmd,
920 		       struct hns_roce_ib_create_qp_resp *resp)
921 {
922 	int ret;
923 
924 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SDI_MODE)
925 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_OWNER_DB;
926 
927 	if (udata) {
928 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE) {
929 			ret = qp_mmap_entry(hr_qp, hr_dev, udata, resp);
930 			if (ret)
931 				return ret;
932 		}
933 
934 		ret = alloc_user_qp_db(hr_dev, hr_qp, init_attr, udata, ucmd,
935 				       resp);
936 		if (ret)
937 			goto err_remove_qp;
938 	} else {
939 		ret = alloc_kernel_qp_db(hr_dev, hr_qp, init_attr);
940 		if (ret)
941 			return ret;
942 	}
943 
944 	return 0;
945 
946 err_remove_qp:
947 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
948 		qp_user_mmap_entry_remove(hr_qp);
949 
950 	return ret;
951 }
952 
953 static void free_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
954 		       struct ib_udata *udata)
955 {
956 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
957 		udata, struct hns_roce_ucontext, ibucontext);
958 
959 	if (udata) {
960 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
961 			hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
962 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
963 			hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
964 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
965 			qp_user_mmap_entry_remove(hr_qp);
966 	} else {
967 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
968 			hns_roce_free_db(hr_dev, &hr_qp->rdb);
969 	}
970 }
971 
972 static int alloc_kernel_wrid(struct hns_roce_dev *hr_dev,
973 			     struct hns_roce_qp *hr_qp)
974 {
975 	struct ib_device *ibdev = &hr_dev->ib_dev;
976 	u64 *sq_wrid = NULL;
977 	u64 *rq_wrid = NULL;
978 	int ret;
979 
980 	sq_wrid = kcalloc(hr_qp->sq.wqe_cnt, sizeof(u64), GFP_KERNEL);
981 	if (ZERO_OR_NULL_PTR(sq_wrid)) {
982 		ibdev_err(ibdev, "failed to alloc SQ wrid.\n");
983 		return -ENOMEM;
984 	}
985 
986 	if (hr_qp->rq.wqe_cnt) {
987 		rq_wrid = kcalloc(hr_qp->rq.wqe_cnt, sizeof(u64), GFP_KERNEL);
988 		if (ZERO_OR_NULL_PTR(rq_wrid)) {
989 			ibdev_err(ibdev, "failed to alloc RQ wrid.\n");
990 			ret = -ENOMEM;
991 			goto err_sq;
992 		}
993 	}
994 
995 	hr_qp->sq.wrid = sq_wrid;
996 	hr_qp->rq.wrid = rq_wrid;
997 	return 0;
998 err_sq:
999 	kfree(sq_wrid);
1000 
1001 	return ret;
1002 }
1003 
1004 static void free_kernel_wrid(struct hns_roce_qp *hr_qp)
1005 {
1006 	kfree(hr_qp->rq.wrid);
1007 	kfree(hr_qp->sq.wrid);
1008 }
1009 
1010 static void default_congest_type(struct hns_roce_dev *hr_dev,
1011 				 struct hns_roce_qp *hr_qp)
1012 {
1013 	if (hr_qp->ibqp.qp_type == IB_QPT_UD ||
1014 	    hr_qp->ibqp.qp_type == IB_QPT_GSI)
1015 		hr_qp->cong_type = CONG_TYPE_DCQCN;
1016 	else
1017 		hr_qp->cong_type = hr_dev->caps.default_cong_type;
1018 }
1019 
1020 static int set_congest_type(struct hns_roce_qp *hr_qp,
1021 			    struct hns_roce_ib_create_qp *ucmd)
1022 {
1023 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_qp->ibqp.device);
1024 
1025 	switch (ucmd->cong_type_flags) {
1026 	case HNS_ROCE_CREATE_QP_FLAGS_DCQCN:
1027 		hr_qp->cong_type = CONG_TYPE_DCQCN;
1028 		break;
1029 	case HNS_ROCE_CREATE_QP_FLAGS_LDCP:
1030 		hr_qp->cong_type = CONG_TYPE_LDCP;
1031 		break;
1032 	case HNS_ROCE_CREATE_QP_FLAGS_HC3:
1033 		hr_qp->cong_type = CONG_TYPE_HC3;
1034 		break;
1035 	case HNS_ROCE_CREATE_QP_FLAGS_DIP:
1036 		hr_qp->cong_type = CONG_TYPE_DIP;
1037 		break;
1038 	default:
1039 		return -EINVAL;
1040 	}
1041 
1042 	if (!test_bit(hr_qp->cong_type, (unsigned long *)&hr_dev->caps.cong_cap))
1043 		return -EOPNOTSUPP;
1044 
1045 	if (hr_qp->ibqp.qp_type == IB_QPT_UD &&
1046 	    hr_qp->cong_type != CONG_TYPE_DCQCN)
1047 		return -EOPNOTSUPP;
1048 
1049 	return 0;
1050 }
1051 
1052 static int set_congest_param(struct hns_roce_dev *hr_dev,
1053 			     struct hns_roce_qp *hr_qp,
1054 			     struct hns_roce_ib_create_qp *ucmd)
1055 {
1056 	if (ucmd->comp_mask & HNS_ROCE_CREATE_QP_MASK_CONGEST_TYPE)
1057 		return set_congest_type(hr_qp, ucmd);
1058 
1059 	default_congest_type(hr_dev, hr_qp);
1060 
1061 	return 0;
1062 }
1063 
1064 static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1065 			struct ib_qp_init_attr *init_attr,
1066 			struct ib_udata *udata,
1067 			struct hns_roce_ib_create_qp *ucmd)
1068 {
1069 	struct ib_device *ibdev = &hr_dev->ib_dev;
1070 	struct hns_roce_ucontext *uctx;
1071 	int ret;
1072 
1073 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1074 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
1075 	else
1076 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
1077 
1078 	ret = set_rq_size(hr_dev, &init_attr->cap, hr_qp,
1079 			  hns_roce_qp_has_rq(init_attr), !!udata);
1080 	if (ret) {
1081 		ibdev_err(ibdev, "failed to set user RQ size, ret = %d.\n",
1082 			  ret);
1083 		return ret;
1084 	}
1085 
1086 	if (udata) {
1087 		ret = ib_copy_from_udata(ucmd, udata,
1088 					 min(udata->inlen, sizeof(*ucmd)));
1089 		if (ret) {
1090 			ibdev_err(ibdev,
1091 				  "failed to copy QP ucmd, ret = %d\n", ret);
1092 			return ret;
1093 		}
1094 
1095 		uctx = rdma_udata_to_drv_context(udata, struct hns_roce_ucontext,
1096 						 ibucontext);
1097 		hr_qp->config = uctx->config;
1098 		ret = set_user_sq_size(hr_dev, &init_attr->cap, hr_qp, ucmd);
1099 		if (ret)
1100 			ibdev_err(ibdev,
1101 				  "failed to set user SQ size, ret = %d.\n",
1102 				  ret);
1103 
1104 		ret = set_congest_param(hr_dev, hr_qp, ucmd);
1105 		if (ret)
1106 			return ret;
1107 	} else {
1108 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
1109 			hr_qp->config = HNS_ROCE_EXSGE_FLAGS;
1110 		ret = set_kernel_sq_size(hr_dev, &init_attr->cap, hr_qp);
1111 		if (ret)
1112 			ibdev_err(ibdev,
1113 				  "failed to set kernel SQ size, ret = %d.\n",
1114 				  ret);
1115 
1116 		default_congest_type(hr_dev, hr_qp);
1117 	}
1118 
1119 	return ret;
1120 }
1121 
1122 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
1123 				     struct ib_qp_init_attr *init_attr,
1124 				     struct ib_udata *udata,
1125 				     struct hns_roce_qp *hr_qp)
1126 {
1127 	struct hns_roce_ib_create_qp_resp resp = {};
1128 	struct ib_device *ibdev = &hr_dev->ib_dev;
1129 	struct hns_roce_ib_create_qp ucmd = {};
1130 	int ret;
1131 
1132 	mutex_init(&hr_qp->mutex);
1133 	spin_lock_init(&hr_qp->sq.lock);
1134 	spin_lock_init(&hr_qp->rq.lock);
1135 
1136 	hr_qp->state = IB_QPS_RESET;
1137 	hr_qp->flush_flag = 0;
1138 
1139 	if (init_attr->create_flags)
1140 		return -EOPNOTSUPP;
1141 
1142 	ret = set_qp_param(hr_dev, hr_qp, init_attr, udata, &ucmd);
1143 	if (ret) {
1144 		ibdev_err(ibdev, "failed to set QP param, ret = %d.\n", ret);
1145 		goto err_out;
1146 	}
1147 
1148 	if (!udata) {
1149 		ret = alloc_kernel_wrid(hr_dev, hr_qp);
1150 		if (ret) {
1151 			ibdev_err(ibdev, "failed to alloc wrid, ret = %d.\n",
1152 				  ret);
1153 			goto err_out;
1154 		}
1155 	}
1156 
1157 	ret = alloc_qp_buf(hr_dev, hr_qp, init_attr, udata, ucmd.buf_addr);
1158 	if (ret) {
1159 		ibdev_err(ibdev, "failed to alloc QP buffer, ret = %d.\n", ret);
1160 		goto err_buf;
1161 	}
1162 
1163 	ret = alloc_qpn(hr_dev, hr_qp, init_attr);
1164 	if (ret) {
1165 		ibdev_err(ibdev, "failed to alloc QPN, ret = %d.\n", ret);
1166 		goto err_qpn;
1167 	}
1168 
1169 	ret = alloc_qp_db(hr_dev, hr_qp, init_attr, udata, &ucmd, &resp);
1170 	if (ret) {
1171 		ibdev_err(ibdev, "failed to alloc QP doorbell, ret = %d.\n",
1172 			  ret);
1173 		goto err_db;
1174 	}
1175 
1176 	ret = alloc_qpc(hr_dev, hr_qp);
1177 	if (ret) {
1178 		ibdev_err(ibdev, "failed to alloc QP context, ret = %d.\n",
1179 			  ret);
1180 		goto err_qpc;
1181 	}
1182 
1183 	ret = hns_roce_qp_store(hr_dev, hr_qp, init_attr);
1184 	if (ret) {
1185 		ibdev_err(ibdev, "failed to store QP, ret = %d.\n", ret);
1186 		goto err_store;
1187 	}
1188 
1189 	if (udata) {
1190 		resp.cap_flags = hr_qp->en_flags;
1191 		ret = ib_copy_to_udata(udata, &resp,
1192 				       min(udata->outlen, sizeof(resp)));
1193 		if (ret) {
1194 			ibdev_err(ibdev, "copy qp resp failed!\n");
1195 			goto err_store;
1196 		}
1197 	}
1198 
1199 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
1200 		ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
1201 		if (ret)
1202 			goto err_flow_ctrl;
1203 	}
1204 
1205 	hr_qp->ibqp.qp_num = hr_qp->qpn;
1206 	hr_qp->event = hns_roce_ib_qp_event;
1207 	refcount_set(&hr_qp->refcount, 1);
1208 	init_completion(&hr_qp->free);
1209 
1210 	return 0;
1211 
1212 err_flow_ctrl:
1213 	hns_roce_qp_remove(hr_dev, hr_qp);
1214 err_store:
1215 	free_qpc(hr_dev, hr_qp);
1216 err_qpc:
1217 	free_qp_db(hr_dev, hr_qp, udata);
1218 err_db:
1219 	free_qpn(hr_dev, hr_qp);
1220 err_qpn:
1221 	free_qp_buf(hr_dev, hr_qp);
1222 err_buf:
1223 	free_kernel_wrid(hr_qp);
1224 err_out:
1225 	mutex_destroy(&hr_qp->mutex);
1226 	return ret;
1227 }
1228 
1229 void hns_roce_qp_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1230 			 struct ib_udata *udata)
1231 {
1232 	if (refcount_dec_and_test(&hr_qp->refcount))
1233 		complete(&hr_qp->free);
1234 	wait_for_completion(&hr_qp->free);
1235 
1236 	free_qpc(hr_dev, hr_qp);
1237 	free_qpn(hr_dev, hr_qp);
1238 	free_qp_buf(hr_dev, hr_qp);
1239 	free_kernel_wrid(hr_qp);
1240 	free_qp_db(hr_dev, hr_qp, udata);
1241 	mutex_destroy(&hr_qp->mutex);
1242 }
1243 
1244 static int check_qp_type(struct hns_roce_dev *hr_dev, enum ib_qp_type type,
1245 			 bool is_user)
1246 {
1247 	switch (type) {
1248 	case IB_QPT_XRC_INI:
1249 	case IB_QPT_XRC_TGT:
1250 		if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC))
1251 			goto out;
1252 		break;
1253 	case IB_QPT_UD:
1254 		if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 &&
1255 		    is_user)
1256 			goto out;
1257 		break;
1258 	case IB_QPT_RC:
1259 	case IB_QPT_GSI:
1260 		break;
1261 	default:
1262 		goto out;
1263 	}
1264 
1265 	return 0;
1266 
1267 out:
1268 	ibdev_err(&hr_dev->ib_dev, "not support QP type %d\n", type);
1269 
1270 	return -EOPNOTSUPP;
1271 }
1272 
1273 int hns_roce_create_qp(struct ib_qp *qp, struct ib_qp_init_attr *init_attr,
1274 		       struct ib_udata *udata)
1275 {
1276 	struct ib_device *ibdev = qp->device;
1277 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
1278 	struct hns_roce_qp *hr_qp = to_hr_qp(qp);
1279 	int ret;
1280 
1281 	ret = check_qp_type(hr_dev, init_attr->qp_type, !!udata);
1282 	if (ret)
1283 		goto err_out;
1284 
1285 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1286 		hr_qp->xrcdn = to_hr_xrcd(init_attr->xrcd)->xrcdn;
1287 
1288 	if (init_attr->qp_type == IB_QPT_GSI) {
1289 		hr_qp->port = init_attr->port_num - 1;
1290 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
1291 	}
1292 
1293 	ret = hns_roce_create_qp_common(hr_dev, init_attr, udata, hr_qp);
1294 	if (ret)
1295 		ibdev_err(ibdev, "create QP type 0x%x failed(%d)\n",
1296 			  init_attr->qp_type, ret);
1297 
1298 err_out:
1299 	if (ret)
1300 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_QP_CREATE_ERR_CNT]);
1301 
1302 	return ret;
1303 }
1304 
1305 int to_hr_qp_type(int qp_type)
1306 {
1307 	switch (qp_type) {
1308 	case IB_QPT_RC:
1309 		return SERV_TYPE_RC;
1310 	case IB_QPT_UD:
1311 	case IB_QPT_GSI:
1312 		return SERV_TYPE_UD;
1313 	case IB_QPT_XRC_INI:
1314 	case IB_QPT_XRC_TGT:
1315 		return SERV_TYPE_XRC;
1316 	default:
1317 		return -1;
1318 	}
1319 }
1320 
1321 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1322 			      struct hns_roce_qp *hr_qp,
1323 			      struct ib_qp_attr *attr, int attr_mask)
1324 {
1325 	enum ib_mtu active_mtu;
1326 	int p;
1327 
1328 	p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1329 	active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1330 
1331 	if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1332 	    attr->path_mtu > hr_dev->caps.max_mtu) ||
1333 	    attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1334 		ibdev_err(&hr_dev->ib_dev,
1335 			"attr path_mtu(%d)invalid while modify qp",
1336 			attr->path_mtu);
1337 		return -EINVAL;
1338 	}
1339 
1340 	return 0;
1341 }
1342 
1343 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1344 				  int attr_mask)
1345 {
1346 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1347 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1348 	int p;
1349 
1350 	if ((attr_mask & IB_QP_PORT) &&
1351 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1352 		ibdev_err(&hr_dev->ib_dev, "invalid attr, port_num = %u.\n",
1353 			  attr->port_num);
1354 		return -EINVAL;
1355 	}
1356 
1357 	if (attr_mask & IB_QP_PKEY_INDEX) {
1358 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1359 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1360 			ibdev_err(&hr_dev->ib_dev,
1361 				  "invalid attr, pkey_index = %u.\n",
1362 				  attr->pkey_index);
1363 			return -EINVAL;
1364 		}
1365 	}
1366 
1367 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1368 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1369 		ibdev_err(&hr_dev->ib_dev,
1370 			  "invalid attr, max_rd_atomic = %u.\n",
1371 			  attr->max_rd_atomic);
1372 		return -EINVAL;
1373 	}
1374 
1375 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1376 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1377 		ibdev_err(&hr_dev->ib_dev,
1378 			  "invalid attr, max_dest_rd_atomic = %u.\n",
1379 			  attr->max_dest_rd_atomic);
1380 		return -EINVAL;
1381 	}
1382 
1383 	if (attr_mask & IB_QP_PATH_MTU)
1384 		return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1385 
1386 	return 0;
1387 }
1388 
1389 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1390 		       int attr_mask, struct ib_udata *udata)
1391 {
1392 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1393 	struct hns_roce_ib_modify_qp_resp resp = {};
1394 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1395 	enum ib_qp_state cur_state, new_state;
1396 	int ret = -EINVAL;
1397 
1398 	mutex_lock(&hr_qp->mutex);
1399 
1400 	if (attr_mask & IB_QP_CUR_STATE && attr->cur_qp_state != hr_qp->state)
1401 		goto out;
1402 
1403 	cur_state = hr_qp->state;
1404 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1405 
1406 	if (ibqp->uobject &&
1407 	    (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1408 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB) {
1409 			hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1410 
1411 			if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
1412 				hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1413 		} else {
1414 			ibdev_warn(&hr_dev->ib_dev,
1415 				  "flush cqe is not supported in userspace!\n");
1416 			goto out;
1417 		}
1418 	}
1419 
1420 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1421 				attr_mask)) {
1422 		ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1423 		goto out;
1424 	}
1425 
1426 	ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1427 	if (ret)
1428 		goto out;
1429 
1430 	if (cur_state == new_state && cur_state == IB_QPS_RESET)
1431 		goto out;
1432 
1433 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1434 				    new_state, udata);
1435 	if (ret)
1436 		goto out;
1437 
1438 	if (udata && udata->outlen) {
1439 		resp.tc_mode = hr_qp->tc_mode;
1440 		resp.priority = hr_qp->sl;
1441 		ret = ib_copy_to_udata(udata, &resp,
1442 				       min(udata->outlen, sizeof(resp)));
1443 		if (ret)
1444 			ibdev_err_ratelimited(&hr_dev->ib_dev,
1445 					      "failed to copy modify qp resp.\n");
1446 	}
1447 
1448 out:
1449 	mutex_unlock(&hr_qp->mutex);
1450 	if (ret)
1451 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_QP_MODIFY_ERR_CNT]);
1452 
1453 	return ret;
1454 }
1455 
1456 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1457 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1458 {
1459 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1460 		__acquire(&send_cq->lock);
1461 		__acquire(&recv_cq->lock);
1462 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1463 		spin_lock(&send_cq->lock);
1464 		__acquire(&recv_cq->lock);
1465 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1466 		spin_lock(&recv_cq->lock);
1467 		__acquire(&send_cq->lock);
1468 	} else if (send_cq == recv_cq) {
1469 		spin_lock(&send_cq->lock);
1470 		__acquire(&recv_cq->lock);
1471 	} else if (send_cq->cqn < recv_cq->cqn) {
1472 		spin_lock(&send_cq->lock);
1473 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1474 	} else {
1475 		spin_lock(&recv_cq->lock);
1476 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1477 	}
1478 }
1479 
1480 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1481 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1482 			 __releases(&recv_cq->lock)
1483 {
1484 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1485 		__release(&recv_cq->lock);
1486 		__release(&send_cq->lock);
1487 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1488 		__release(&recv_cq->lock);
1489 		spin_unlock(&send_cq->lock);
1490 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1491 		__release(&send_cq->lock);
1492 		spin_unlock(&recv_cq->lock);
1493 	} else if (send_cq == recv_cq) {
1494 		__release(&recv_cq->lock);
1495 		spin_unlock(&send_cq->lock);
1496 	} else if (send_cq->cqn < recv_cq->cqn) {
1497 		spin_unlock(&recv_cq->lock);
1498 		spin_unlock(&send_cq->lock);
1499 	} else {
1500 		spin_unlock(&send_cq->lock);
1501 		spin_unlock(&recv_cq->lock);
1502 	}
1503 }
1504 
1505 static inline void *get_wqe(struct hns_roce_qp *hr_qp, u32 offset)
1506 {
1507 	return hns_roce_buf_offset(hr_qp->mtr.kmem, offset);
1508 }
1509 
1510 void *hns_roce_get_recv_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1511 {
1512 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1513 }
1514 
1515 void *hns_roce_get_send_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1516 {
1517 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1518 }
1519 
1520 void *hns_roce_get_extend_sge(struct hns_roce_qp *hr_qp, unsigned int n)
1521 {
1522 	return get_wqe(hr_qp, hr_qp->sge.offset + (n << hr_qp->sge.sge_shift));
1523 }
1524 
1525 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, u32 nreq,
1526 			  struct ib_cq *ib_cq)
1527 {
1528 	struct hns_roce_cq *hr_cq;
1529 	u32 cur;
1530 
1531 	cur = hr_wq->head - hr_wq->tail;
1532 	if (likely(cur + nreq < hr_wq->wqe_cnt))
1533 		return false;
1534 
1535 	hr_cq = to_hr_cq(ib_cq);
1536 	spin_lock(&hr_cq->lock);
1537 	cur = hr_wq->head - hr_wq->tail;
1538 	spin_unlock(&hr_cq->lock);
1539 
1540 	return cur + nreq >= hr_wq->wqe_cnt;
1541 }
1542 
1543 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1544 {
1545 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1546 	unsigned int reserved_from_bot;
1547 	unsigned int i;
1548 
1549 	qp_table->idx_table.spare_idx = kcalloc(hr_dev->caps.num_qps,
1550 					sizeof(u32), GFP_KERNEL);
1551 	if (!qp_table->idx_table.spare_idx)
1552 		return -ENOMEM;
1553 
1554 	mutex_init(&qp_table->scc_mutex);
1555 	mutex_init(&qp_table->bank_mutex);
1556 	xa_init(&hr_dev->qp_table_xa);
1557 
1558 	reserved_from_bot = hr_dev->caps.reserved_qps;
1559 
1560 	for (i = 0; i < reserved_from_bot; i++) {
1561 		hr_dev->qp_table.bank[get_qp_bankid(i)].inuse++;
1562 		hr_dev->qp_table.bank[get_qp_bankid(i)].min++;
1563 	}
1564 
1565 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) {
1566 		ida_init(&hr_dev->qp_table.bank[i].ida);
1567 		hr_dev->qp_table.bank[i].max = hr_dev->caps.num_qps /
1568 					       HNS_ROCE_QP_BANK_NUM - 1;
1569 		hr_dev->qp_table.bank[i].next = hr_dev->qp_table.bank[i].min;
1570 	}
1571 
1572 	return 0;
1573 }
1574 
1575 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1576 {
1577 	int i;
1578 
1579 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++)
1580 		ida_destroy(&hr_dev->qp_table.bank[i].ida);
1581 	mutex_destroy(&hr_dev->qp_table.bank_mutex);
1582 	mutex_destroy(&hr_dev->qp_table.scc_mutex);
1583 	kfree(hr_dev->qp_table.idx_table.spare_idx);
1584 }
1585