xref: /linux/drivers/infiniband/hw/hns/hns_roce_qp.c (revision 64b14a184e83eb62ea0615e31a409956049d40e7)
1 /*
2  * Copyright (c) 2016 Hisilicon Limited.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/pci.h>
35 #include <rdma/ib_addr.h>
36 #include <rdma/ib_umem.h>
37 #include <rdma/uverbs_ioctl.h>
38 #include "hns_roce_common.h"
39 #include "hns_roce_device.h"
40 #include "hns_roce_hem.h"
41 
42 static void flush_work_handle(struct work_struct *work)
43 {
44 	struct hns_roce_work *flush_work = container_of(work,
45 					struct hns_roce_work, work);
46 	struct hns_roce_qp *hr_qp = container_of(flush_work,
47 					struct hns_roce_qp, flush_work);
48 	struct device *dev = flush_work->hr_dev->dev;
49 	struct ib_qp_attr attr;
50 	int attr_mask;
51 	int ret;
52 
53 	attr_mask = IB_QP_STATE;
54 	attr.qp_state = IB_QPS_ERR;
55 
56 	if (test_and_clear_bit(HNS_ROCE_FLUSH_FLAG, &hr_qp->flush_flag)) {
57 		ret = hns_roce_modify_qp(&hr_qp->ibqp, &attr, attr_mask, NULL);
58 		if (ret)
59 			dev_err(dev, "Modify QP to error state failed(%d) during CQE flush\n",
60 				ret);
61 	}
62 
63 	/*
64 	 * make sure we signal QP destroy leg that flush QP was completed
65 	 * so that it can safely proceed ahead now and destroy QP
66 	 */
67 	if (refcount_dec_and_test(&hr_qp->refcount))
68 		complete(&hr_qp->free);
69 }
70 
71 void init_flush_work(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
72 {
73 	struct hns_roce_work *flush_work = &hr_qp->flush_work;
74 
75 	flush_work->hr_dev = hr_dev;
76 	INIT_WORK(&flush_work->work, flush_work_handle);
77 	refcount_inc(&hr_qp->refcount);
78 	queue_work(hr_dev->irq_workq, &flush_work->work);
79 }
80 
81 void flush_cqe(struct hns_roce_dev *dev, struct hns_roce_qp *qp)
82 {
83 	/*
84 	 * Hip08 hardware cannot flush the WQEs in SQ/RQ if the QP state
85 	 * gets into errored mode. Hence, as a workaround to this
86 	 * hardware limitation, driver needs to assist in flushing. But
87 	 * the flushing operation uses mailbox to convey the QP state to
88 	 * the hardware and which can sleep due to the mutex protection
89 	 * around the mailbox calls. Hence, use the deferred flush for
90 	 * now.
91 	 */
92 	if (!test_and_set_bit(HNS_ROCE_FLUSH_FLAG, &qp->flush_flag))
93 		init_flush_work(dev, qp);
94 }
95 
96 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
97 {
98 	struct device *dev = hr_dev->dev;
99 	struct hns_roce_qp *qp;
100 
101 	xa_lock(&hr_dev->qp_table_xa);
102 	qp = __hns_roce_qp_lookup(hr_dev, qpn);
103 	if (qp)
104 		refcount_inc(&qp->refcount);
105 	xa_unlock(&hr_dev->qp_table_xa);
106 
107 	if (!qp) {
108 		dev_warn(dev, "Async event for bogus QP %08x\n", qpn);
109 		return;
110 	}
111 
112 	if (event_type == HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR ||
113 	    event_type == HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR ||
114 	    event_type == HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR ||
115 	    event_type == HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION ||
116 	    event_type == HNS_ROCE_EVENT_TYPE_INVALID_XRCETH) {
117 		qp->state = IB_QPS_ERR;
118 
119 		flush_cqe(hr_dev, qp);
120 	}
121 
122 	qp->event(qp, (enum hns_roce_event)event_type);
123 
124 	if (refcount_dec_and_test(&qp->refcount))
125 		complete(&qp->free);
126 }
127 
128 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
129 				 enum hns_roce_event type)
130 {
131 	struct ib_qp *ibqp = &hr_qp->ibqp;
132 	struct ib_event event;
133 
134 	if (ibqp->event_handler) {
135 		event.device = ibqp->device;
136 		event.element.qp = ibqp;
137 		switch (type) {
138 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
139 			event.event = IB_EVENT_PATH_MIG;
140 			break;
141 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
142 			event.event = IB_EVENT_COMM_EST;
143 			break;
144 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
145 			event.event = IB_EVENT_SQ_DRAINED;
146 			break;
147 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
148 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
149 			break;
150 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
151 			event.event = IB_EVENT_QP_FATAL;
152 			break;
153 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
154 			event.event = IB_EVENT_PATH_MIG_ERR;
155 			break;
156 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
157 			event.event = IB_EVENT_QP_REQ_ERR;
158 			break;
159 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
160 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
161 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
162 			event.event = IB_EVENT_QP_ACCESS_ERR;
163 			break;
164 		default:
165 			dev_dbg(ibqp->device->dev.parent, "roce_ib: Unexpected event type %d on QP %06lx\n",
166 				type, hr_qp->qpn);
167 			return;
168 		}
169 		ibqp->event_handler(&event, ibqp->qp_context);
170 	}
171 }
172 
173 static u8 get_least_load_bankid_for_qp(struct hns_roce_bank *bank)
174 {
175 	u32 least_load = bank[0].inuse;
176 	u8 bankid = 0;
177 	u32 bankcnt;
178 	u8 i;
179 
180 	for (i = 1; i < HNS_ROCE_QP_BANK_NUM; i++) {
181 		bankcnt = bank[i].inuse;
182 		if (bankcnt < least_load) {
183 			least_load = bankcnt;
184 			bankid = i;
185 		}
186 	}
187 
188 	return bankid;
189 }
190 
191 static int alloc_qpn_with_bankid(struct hns_roce_bank *bank, u8 bankid,
192 				 unsigned long *qpn)
193 {
194 	int id;
195 
196 	id = ida_alloc_range(&bank->ida, bank->next, bank->max, GFP_KERNEL);
197 	if (id < 0) {
198 		id = ida_alloc_range(&bank->ida, bank->min, bank->max,
199 				     GFP_KERNEL);
200 		if (id < 0)
201 			return id;
202 	}
203 
204 	/* the QPN should keep increasing until the max value is reached. */
205 	bank->next = (id + 1) > bank->max ? bank->min : id + 1;
206 
207 	/* the lower 3 bits is bankid */
208 	*qpn = (id << 3) | bankid;
209 
210 	return 0;
211 }
212 static int alloc_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
213 {
214 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
215 	unsigned long num = 0;
216 	u8 bankid;
217 	int ret;
218 
219 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI) {
220 		num = 1;
221 		hr_qp->doorbell_qpn = 1;
222 	} else {
223 		mutex_lock(&qp_table->bank_mutex);
224 		bankid = get_least_load_bankid_for_qp(qp_table->bank);
225 
226 		ret = alloc_qpn_with_bankid(&qp_table->bank[bankid], bankid,
227 					    &num);
228 		if (ret) {
229 			ibdev_err(&hr_dev->ib_dev,
230 				  "failed to alloc QPN, ret = %d\n", ret);
231 			mutex_unlock(&qp_table->bank_mutex);
232 			return ret;
233 		}
234 
235 		qp_table->bank[bankid].inuse++;
236 		mutex_unlock(&qp_table->bank_mutex);
237 
238 		hr_qp->doorbell_qpn = (u32)num;
239 	}
240 
241 	hr_qp->qpn = num;
242 
243 	return 0;
244 }
245 
246 enum hns_roce_qp_state to_hns_roce_state(enum ib_qp_state state)
247 {
248 	switch (state) {
249 	case IB_QPS_RESET:
250 		return HNS_ROCE_QP_STATE_RST;
251 	case IB_QPS_INIT:
252 		return HNS_ROCE_QP_STATE_INIT;
253 	case IB_QPS_RTR:
254 		return HNS_ROCE_QP_STATE_RTR;
255 	case IB_QPS_RTS:
256 		return HNS_ROCE_QP_STATE_RTS;
257 	case IB_QPS_SQD:
258 		return HNS_ROCE_QP_STATE_SQD;
259 	case IB_QPS_ERR:
260 		return HNS_ROCE_QP_STATE_ERR;
261 	default:
262 		return HNS_ROCE_QP_NUM_STATE;
263 	}
264 }
265 
266 static void add_qp_to_list(struct hns_roce_dev *hr_dev,
267 			   struct hns_roce_qp *hr_qp,
268 			   struct ib_cq *send_cq, struct ib_cq *recv_cq)
269 {
270 	struct hns_roce_cq *hr_send_cq, *hr_recv_cq;
271 	unsigned long flags;
272 
273 	hr_send_cq = send_cq ? to_hr_cq(send_cq) : NULL;
274 	hr_recv_cq = recv_cq ? to_hr_cq(recv_cq) : NULL;
275 
276 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
277 	hns_roce_lock_cqs(hr_send_cq, hr_recv_cq);
278 
279 	list_add_tail(&hr_qp->node, &hr_dev->qp_list);
280 	if (hr_send_cq)
281 		list_add_tail(&hr_qp->sq_node, &hr_send_cq->sq_list);
282 	if (hr_recv_cq)
283 		list_add_tail(&hr_qp->rq_node, &hr_recv_cq->rq_list);
284 
285 	hns_roce_unlock_cqs(hr_send_cq, hr_recv_cq);
286 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
287 }
288 
289 static int hns_roce_qp_store(struct hns_roce_dev *hr_dev,
290 			     struct hns_roce_qp *hr_qp,
291 			     struct ib_qp_init_attr *init_attr)
292 {
293 	struct xarray *xa = &hr_dev->qp_table_xa;
294 	int ret;
295 
296 	if (!hr_qp->qpn)
297 		return -EINVAL;
298 
299 	ret = xa_err(xa_store_irq(xa, hr_qp->qpn, hr_qp, GFP_KERNEL));
300 	if (ret)
301 		dev_err(hr_dev->dev, "Failed to xa store for QPC\n");
302 	else
303 		/* add QP to device's QP list for softwc */
304 		add_qp_to_list(hr_dev, hr_qp, init_attr->send_cq,
305 			       init_attr->recv_cq);
306 
307 	return ret;
308 }
309 
310 static int alloc_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
311 {
312 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
313 	struct device *dev = hr_dev->dev;
314 	int ret;
315 
316 	if (!hr_qp->qpn)
317 		return -EINVAL;
318 
319 	/* Alloc memory for QPC */
320 	ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
321 	if (ret) {
322 		dev_err(dev, "Failed to get QPC table\n");
323 		goto err_out;
324 	}
325 
326 	/* Alloc memory for IRRL */
327 	ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
328 	if (ret) {
329 		dev_err(dev, "Failed to get IRRL table\n");
330 		goto err_put_qp;
331 	}
332 
333 	if (hr_dev->caps.trrl_entry_sz) {
334 		/* Alloc memory for TRRL */
335 		ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
336 					 hr_qp->qpn);
337 		if (ret) {
338 			dev_err(dev, "Failed to get TRRL table\n");
339 			goto err_put_irrl;
340 		}
341 	}
342 
343 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
344 		/* Alloc memory for SCC CTX */
345 		ret = hns_roce_table_get(hr_dev, &qp_table->sccc_table,
346 					 hr_qp->qpn);
347 		if (ret) {
348 			dev_err(dev, "Failed to get SCC CTX table\n");
349 			goto err_put_trrl;
350 		}
351 	}
352 
353 	return 0;
354 
355 err_put_trrl:
356 	if (hr_dev->caps.trrl_entry_sz)
357 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
358 
359 err_put_irrl:
360 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
361 
362 err_put_qp:
363 	hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
364 
365 err_out:
366 	return ret;
367 }
368 
369 static void qp_user_mmap_entry_remove(struct hns_roce_qp *hr_qp)
370 {
371 	rdma_user_mmap_entry_remove(&hr_qp->dwqe_mmap_entry->rdma_entry);
372 }
373 
374 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
375 {
376 	struct xarray *xa = &hr_dev->qp_table_xa;
377 	unsigned long flags;
378 
379 	list_del(&hr_qp->node);
380 
381 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
382 		list_del(&hr_qp->sq_node);
383 
384 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_INI &&
385 	    hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
386 		list_del(&hr_qp->rq_node);
387 
388 	xa_lock_irqsave(xa, flags);
389 	__xa_erase(xa, hr_qp->qpn);
390 	xa_unlock_irqrestore(xa, flags);
391 }
392 
393 static void free_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
394 {
395 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
396 
397 	if (hr_dev->caps.trrl_entry_sz)
398 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
399 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
400 }
401 
402 static inline u8 get_qp_bankid(unsigned long qpn)
403 {
404 	/* The lower 3 bits of QPN are used to hash to different banks */
405 	return (u8)(qpn & GENMASK(2, 0));
406 }
407 
408 static void free_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
409 {
410 	u8 bankid;
411 
412 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI)
413 		return;
414 
415 	if (hr_qp->qpn < hr_dev->caps.reserved_qps)
416 		return;
417 
418 	bankid = get_qp_bankid(hr_qp->qpn);
419 
420 	ida_free(&hr_dev->qp_table.bank[bankid].ida, hr_qp->qpn >> 3);
421 
422 	mutex_lock(&hr_dev->qp_table.bank_mutex);
423 	hr_dev->qp_table.bank[bankid].inuse--;
424 	mutex_unlock(&hr_dev->qp_table.bank_mutex);
425 }
426 
427 static u32 proc_rq_sge(struct hns_roce_dev *dev, struct hns_roce_qp *hr_qp,
428 		       bool user)
429 {
430 	u32 max_sge = dev->caps.max_rq_sg;
431 
432 	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
433 		return max_sge;
434 
435 	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
436 	 * calculate number of max_sge with reserved SGEs when allocating wqe
437 	 * buf, so there is no need to do this again in kernel. But the number
438 	 * may exceed the capacity of SGEs recorded in the firmware, so the
439 	 * kernel driver should just adapt the value accordingly.
440 	 */
441 	if (user)
442 		max_sge = roundup_pow_of_two(max_sge + 1);
443 	else
444 		hr_qp->rq.rsv_sge = 1;
445 
446 	return max_sge;
447 }
448 
449 static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
450 		       struct hns_roce_qp *hr_qp, int has_rq, bool user)
451 {
452 	u32 max_sge = proc_rq_sge(hr_dev, hr_qp, user);
453 	u32 cnt;
454 
455 	/* If srq exist, set zero for relative number of rq */
456 	if (!has_rq) {
457 		hr_qp->rq.wqe_cnt = 0;
458 		hr_qp->rq.max_gs = 0;
459 		hr_qp->rq_inl_buf.wqe_cnt = 0;
460 		cap->max_recv_wr = 0;
461 		cap->max_recv_sge = 0;
462 
463 		return 0;
464 	}
465 
466 	/* Check the validity of QP support capacity */
467 	if (!cap->max_recv_wr || cap->max_recv_wr > hr_dev->caps.max_wqes ||
468 	    cap->max_recv_sge > max_sge) {
469 		ibdev_err(&hr_dev->ib_dev,
470 			  "RQ config error, depth = %u, sge = %u\n",
471 			  cap->max_recv_wr, cap->max_recv_sge);
472 		return -EINVAL;
473 	}
474 
475 	cnt = roundup_pow_of_two(max(cap->max_recv_wr, hr_dev->caps.min_wqes));
476 	if (cnt > hr_dev->caps.max_wqes) {
477 		ibdev_err(&hr_dev->ib_dev, "rq depth %u too large\n",
478 			  cap->max_recv_wr);
479 		return -EINVAL;
480 	}
481 
482 	hr_qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge) +
483 					      hr_qp->rq.rsv_sge);
484 
485 	if (hr_dev->caps.max_rq_sg <= HNS_ROCE_SGE_IN_WQE)
486 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz);
487 	else
488 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz *
489 					    hr_qp->rq.max_gs);
490 
491 	hr_qp->rq.wqe_cnt = cnt;
492 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE &&
493 	    hr_qp->ibqp.qp_type != IB_QPT_UD &&
494 	    hr_qp->ibqp.qp_type != IB_QPT_GSI)
495 		hr_qp->rq_inl_buf.wqe_cnt = cnt;
496 	else
497 		hr_qp->rq_inl_buf.wqe_cnt = 0;
498 
499 	cap->max_recv_wr = cnt;
500 	cap->max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
501 
502 	return 0;
503 }
504 
505 static u32 get_wqe_ext_sge_cnt(struct hns_roce_qp *qp)
506 {
507 	/* GSI/UD QP only has extended sge */
508 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_UD)
509 		return qp->sq.max_gs;
510 
511 	if (qp->sq.max_gs > HNS_ROCE_SGE_IN_WQE)
512 		return qp->sq.max_gs - HNS_ROCE_SGE_IN_WQE;
513 
514 	return 0;
515 }
516 
517 static void set_ext_sge_param(struct hns_roce_dev *hr_dev, u32 sq_wqe_cnt,
518 			      struct hns_roce_qp *hr_qp, struct ib_qp_cap *cap)
519 {
520 	u32 total_sge_cnt;
521 	u32 wqe_sge_cnt;
522 
523 	hr_qp->sge.sge_shift = HNS_ROCE_SGE_SHIFT;
524 
525 	hr_qp->sq.max_gs = max(1U, cap->max_send_sge);
526 
527 	wqe_sge_cnt = get_wqe_ext_sge_cnt(hr_qp);
528 
529 	/* If the number of extended sge is not zero, they MUST use the
530 	 * space of HNS_HW_PAGE_SIZE at least.
531 	 */
532 	if (wqe_sge_cnt) {
533 		total_sge_cnt = roundup_pow_of_two(sq_wqe_cnt * wqe_sge_cnt);
534 		hr_qp->sge.sge_cnt = max(total_sge_cnt,
535 				(u32)HNS_HW_PAGE_SIZE / HNS_ROCE_SGE_SIZE);
536 	}
537 }
538 
539 static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
540 					struct ib_qp_cap *cap,
541 					struct hns_roce_ib_create_qp *ucmd)
542 {
543 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
544 	u8 max_sq_stride = ilog2(roundup_sq_stride);
545 
546 	/* Sanity check SQ size before proceeding */
547 	if (ucmd->log_sq_stride > max_sq_stride ||
548 	    ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
549 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ stride size.\n");
550 		return -EINVAL;
551 	}
552 
553 	if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
554 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ SGE size %u.\n",
555 			  cap->max_send_sge);
556 		return -EINVAL;
557 	}
558 
559 	return 0;
560 }
561 
562 static int set_user_sq_size(struct hns_roce_dev *hr_dev,
563 			    struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp,
564 			    struct hns_roce_ib_create_qp *ucmd)
565 {
566 	struct ib_device *ibdev = &hr_dev->ib_dev;
567 	u32 cnt = 0;
568 	int ret;
569 
570 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
571 	    cnt > hr_dev->caps.max_wqes)
572 		return -EINVAL;
573 
574 	ret = check_sq_size_with_integrity(hr_dev, cap, ucmd);
575 	if (ret) {
576 		ibdev_err(ibdev, "failed to check user SQ size, ret = %d.\n",
577 			  ret);
578 		return ret;
579 	}
580 
581 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
582 
583 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
584 	hr_qp->sq.wqe_cnt = cnt;
585 
586 	return 0;
587 }
588 
589 static int set_wqe_buf_attr(struct hns_roce_dev *hr_dev,
590 			    struct hns_roce_qp *hr_qp,
591 			    struct hns_roce_buf_attr *buf_attr)
592 {
593 	int buf_size;
594 	int idx = 0;
595 
596 	hr_qp->buff_size = 0;
597 
598 	/* SQ WQE */
599 	hr_qp->sq.offset = 0;
600 	buf_size = to_hr_hem_entries_size(hr_qp->sq.wqe_cnt,
601 					  hr_qp->sq.wqe_shift);
602 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
603 		buf_attr->region[idx].size = buf_size;
604 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sq_hop_num;
605 		idx++;
606 		hr_qp->buff_size += buf_size;
607 	}
608 
609 	/* extend SGE WQE in SQ */
610 	hr_qp->sge.offset = hr_qp->buff_size;
611 	buf_size = to_hr_hem_entries_size(hr_qp->sge.sge_cnt,
612 					  hr_qp->sge.sge_shift);
613 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
614 		buf_attr->region[idx].size = buf_size;
615 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sge_hop_num;
616 		idx++;
617 		hr_qp->buff_size += buf_size;
618 	}
619 
620 	/* RQ WQE */
621 	hr_qp->rq.offset = hr_qp->buff_size;
622 	buf_size = to_hr_hem_entries_size(hr_qp->rq.wqe_cnt,
623 					  hr_qp->rq.wqe_shift);
624 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
625 		buf_attr->region[idx].size = buf_size;
626 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_rq_hop_num;
627 		idx++;
628 		hr_qp->buff_size += buf_size;
629 	}
630 
631 	if (hr_qp->buff_size < 1)
632 		return -EINVAL;
633 
634 	buf_attr->page_shift = HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
635 	buf_attr->region_count = idx;
636 
637 	return 0;
638 }
639 
640 static int set_kernel_sq_size(struct hns_roce_dev *hr_dev,
641 			      struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp)
642 {
643 	struct ib_device *ibdev = &hr_dev->ib_dev;
644 	u32 cnt;
645 
646 	if (!cap->max_send_wr || cap->max_send_wr > hr_dev->caps.max_wqes ||
647 	    cap->max_send_sge > hr_dev->caps.max_sq_sg) {
648 		ibdev_err(ibdev, "failed to check SQ WR or SGE num.\n");
649 		return -EINVAL;
650 	}
651 
652 	cnt = roundup_pow_of_two(max(cap->max_send_wr, hr_dev->caps.min_wqes));
653 	if (cnt > hr_dev->caps.max_wqes) {
654 		ibdev_err(ibdev, "failed to check WQE num, WQE num = %u.\n",
655 			  cnt);
656 		return -EINVAL;
657 	}
658 
659 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
660 	hr_qp->sq.wqe_cnt = cnt;
661 
662 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
663 
664 	/* sync the parameters of kernel QP to user's configuration */
665 	cap->max_send_wr = cnt;
666 	cap->max_send_sge = hr_qp->sq.max_gs;
667 
668 	return 0;
669 }
670 
671 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
672 {
673 	if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
674 		return 0;
675 
676 	return 1;
677 }
678 
679 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
680 {
681 	if (attr->qp_type == IB_QPT_XRC_INI ||
682 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
683 	    !attr->cap.max_recv_wr)
684 		return 0;
685 
686 	return 1;
687 }
688 
689 static int alloc_rq_inline_buf(struct hns_roce_qp *hr_qp,
690 			       struct ib_qp_init_attr *init_attr)
691 {
692 	u32 max_recv_sge = init_attr->cap.max_recv_sge;
693 	u32 wqe_cnt = hr_qp->rq_inl_buf.wqe_cnt;
694 	struct hns_roce_rinl_wqe *wqe_list;
695 	int i;
696 
697 	/* allocate recv inline buf */
698 	wqe_list = kcalloc(wqe_cnt, sizeof(struct hns_roce_rinl_wqe),
699 			   GFP_KERNEL);
700 	if (!wqe_list)
701 		goto err;
702 
703 	/* Allocate a continuous buffer for all inline sge we need */
704 	wqe_list[0].sg_list = kcalloc(wqe_cnt, (max_recv_sge *
705 				      sizeof(struct hns_roce_rinl_sge)),
706 				      GFP_KERNEL);
707 	if (!wqe_list[0].sg_list)
708 		goto err_wqe_list;
709 
710 	/* Assign buffers of sg_list to each inline wqe */
711 	for (i = 1; i < wqe_cnt; i++)
712 		wqe_list[i].sg_list = &wqe_list[0].sg_list[i * max_recv_sge];
713 
714 	hr_qp->rq_inl_buf.wqe_list = wqe_list;
715 
716 	return 0;
717 
718 err_wqe_list:
719 	kfree(wqe_list);
720 
721 err:
722 	return -ENOMEM;
723 }
724 
725 static void free_rq_inline_buf(struct hns_roce_qp *hr_qp)
726 {
727 	if (hr_qp->rq_inl_buf.wqe_list)
728 		kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
729 	kfree(hr_qp->rq_inl_buf.wqe_list);
730 }
731 
732 static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
733 			struct ib_qp_init_attr *init_attr,
734 			struct ib_udata *udata, unsigned long addr)
735 {
736 	struct ib_device *ibdev = &hr_dev->ib_dev;
737 	struct hns_roce_buf_attr buf_attr = {};
738 	int ret;
739 
740 	if (!udata && hr_qp->rq_inl_buf.wqe_cnt) {
741 		ret = alloc_rq_inline_buf(hr_qp, init_attr);
742 		if (ret) {
743 			ibdev_err(ibdev,
744 				  "failed to alloc inline buf, ret = %d.\n",
745 				  ret);
746 			return ret;
747 		}
748 	} else {
749 		hr_qp->rq_inl_buf.wqe_list = NULL;
750 	}
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 	free_rq_inline_buf(hr_qp);
772 
773 	return ret;
774 }
775 
776 static void free_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
777 {
778 	hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
779 	free_rq_inline_buf(hr_qp);
780 }
781 
782 static inline bool user_qp_has_sdb(struct hns_roce_dev *hr_dev,
783 				   struct ib_qp_init_attr *init_attr,
784 				   struct ib_udata *udata,
785 				   struct hns_roce_ib_create_qp_resp *resp,
786 				   struct hns_roce_ib_create_qp *ucmd)
787 {
788 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
789 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
790 		hns_roce_qp_has_sq(init_attr) &&
791 		udata->inlen >= offsetofend(typeof(*ucmd), sdb_addr));
792 }
793 
794 static inline bool user_qp_has_rdb(struct hns_roce_dev *hr_dev,
795 				   struct ib_qp_init_attr *init_attr,
796 				   struct ib_udata *udata,
797 				   struct hns_roce_ib_create_qp_resp *resp)
798 {
799 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
800 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
801 		hns_roce_qp_has_rq(init_attr));
802 }
803 
804 static inline bool kernel_qp_has_rdb(struct hns_roce_dev *hr_dev,
805 				     struct ib_qp_init_attr *init_attr)
806 {
807 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
808 		hns_roce_qp_has_rq(init_attr));
809 }
810 
811 static int qp_mmap_entry(struct hns_roce_qp *hr_qp,
812 			 struct hns_roce_dev *hr_dev,
813 			 struct ib_udata *udata,
814 			 struct hns_roce_ib_create_qp_resp *resp)
815 {
816 	struct hns_roce_ucontext *uctx =
817 		rdma_udata_to_drv_context(udata,
818 			struct hns_roce_ucontext, ibucontext);
819 	struct rdma_user_mmap_entry *rdma_entry;
820 	u64 address;
821 
822 	address = hr_dev->dwqe_page + hr_qp->qpn * HNS_ROCE_DWQE_SIZE;
823 
824 	hr_qp->dwqe_mmap_entry =
825 		hns_roce_user_mmap_entry_insert(&uctx->ibucontext, address,
826 						HNS_ROCE_DWQE_SIZE,
827 						HNS_ROCE_MMAP_TYPE_DWQE);
828 
829 	if (!hr_qp->dwqe_mmap_entry) {
830 		ibdev_err(&hr_dev->ib_dev, "failed to get dwqe mmap entry.\n");
831 		return -ENOMEM;
832 	}
833 
834 	rdma_entry = &hr_qp->dwqe_mmap_entry->rdma_entry;
835 	resp->dwqe_mmap_key = rdma_user_mmap_get_offset(rdma_entry);
836 
837 	return 0;
838 }
839 
840 static int alloc_user_qp_db(struct hns_roce_dev *hr_dev,
841 			    struct hns_roce_qp *hr_qp,
842 			    struct ib_qp_init_attr *init_attr,
843 			    struct ib_udata *udata,
844 			    struct hns_roce_ib_create_qp *ucmd,
845 			    struct hns_roce_ib_create_qp_resp *resp)
846 {
847 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(udata,
848 		struct hns_roce_ucontext, ibucontext);
849 	struct ib_device *ibdev = &hr_dev->ib_dev;
850 	int ret;
851 
852 	if (user_qp_has_sdb(hr_dev, init_attr, udata, resp, ucmd)) {
853 		ret = hns_roce_db_map_user(uctx, ucmd->sdb_addr, &hr_qp->sdb);
854 		if (ret) {
855 			ibdev_err(ibdev,
856 				  "failed to map user SQ doorbell, ret = %d.\n",
857 				  ret);
858 			goto err_out;
859 		}
860 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
861 	}
862 
863 	if (user_qp_has_rdb(hr_dev, init_attr, udata, resp)) {
864 		ret = hns_roce_db_map_user(uctx, ucmd->db_addr, &hr_qp->rdb);
865 		if (ret) {
866 			ibdev_err(ibdev,
867 				  "failed to map user RQ doorbell, ret = %d.\n",
868 				  ret);
869 			goto err_sdb;
870 		}
871 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
872 	}
873 
874 	return 0;
875 
876 err_sdb:
877 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
878 		hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
879 err_out:
880 	return ret;
881 }
882 
883 static int alloc_kernel_qp_db(struct hns_roce_dev *hr_dev,
884 			      struct hns_roce_qp *hr_qp,
885 			      struct ib_qp_init_attr *init_attr)
886 {
887 	struct ib_device *ibdev = &hr_dev->ib_dev;
888 	int ret;
889 
890 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
891 		hr_qp->sq.db_reg = hr_dev->mem_base +
892 				   HNS_ROCE_DWQE_SIZE * hr_qp->qpn;
893 	else
894 		hr_qp->sq.db_reg = hr_dev->reg_base + hr_dev->sdb_offset +
895 				   DB_REG_OFFSET * hr_dev->priv_uar.index;
896 
897 	hr_qp->rq.db_reg = hr_dev->reg_base + hr_dev->odb_offset +
898 			   DB_REG_OFFSET * hr_dev->priv_uar.index;
899 
900 	if (kernel_qp_has_rdb(hr_dev, init_attr)) {
901 		ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
902 		if (ret) {
903 			ibdev_err(ibdev,
904 				  "failed to alloc kernel RQ doorbell, ret = %d.\n",
905 				  ret);
906 			return ret;
907 		}
908 		*hr_qp->rdb.db_record = 0;
909 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
910 	}
911 
912 	return 0;
913 }
914 
915 static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
916 		       struct ib_qp_init_attr *init_attr,
917 		       struct ib_udata *udata,
918 		       struct hns_roce_ib_create_qp *ucmd,
919 		       struct hns_roce_ib_create_qp_resp *resp)
920 {
921 	int ret;
922 
923 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SDI_MODE)
924 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_OWNER_DB;
925 
926 	if (udata) {
927 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE) {
928 			ret = qp_mmap_entry(hr_qp, hr_dev, udata, resp);
929 			if (ret)
930 				return ret;
931 		}
932 
933 		ret = alloc_user_qp_db(hr_dev, hr_qp, init_attr, udata, ucmd,
934 				       resp);
935 		if (ret)
936 			goto err_remove_qp;
937 	} else {
938 		ret = alloc_kernel_qp_db(hr_dev, hr_qp, init_attr);
939 		if (ret)
940 			return ret;
941 	}
942 
943 	return 0;
944 
945 err_remove_qp:
946 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
947 		qp_user_mmap_entry_remove(hr_qp);
948 
949 	return ret;
950 }
951 
952 static void free_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
953 		       struct ib_udata *udata)
954 {
955 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
956 		udata, struct hns_roce_ucontext, ibucontext);
957 
958 	if (udata) {
959 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
960 			hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
961 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
962 			hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
963 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
964 			qp_user_mmap_entry_remove(hr_qp);
965 	} else {
966 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
967 			hns_roce_free_db(hr_dev, &hr_qp->rdb);
968 	}
969 }
970 
971 static int alloc_kernel_wrid(struct hns_roce_dev *hr_dev,
972 			     struct hns_roce_qp *hr_qp)
973 {
974 	struct ib_device *ibdev = &hr_dev->ib_dev;
975 	u64 *sq_wrid = NULL;
976 	u64 *rq_wrid = NULL;
977 	int ret;
978 
979 	sq_wrid = kcalloc(hr_qp->sq.wqe_cnt, sizeof(u64), GFP_KERNEL);
980 	if (ZERO_OR_NULL_PTR(sq_wrid)) {
981 		ibdev_err(ibdev, "failed to alloc SQ wrid.\n");
982 		return -ENOMEM;
983 	}
984 
985 	if (hr_qp->rq.wqe_cnt) {
986 		rq_wrid = kcalloc(hr_qp->rq.wqe_cnt, sizeof(u64), GFP_KERNEL);
987 		if (ZERO_OR_NULL_PTR(rq_wrid)) {
988 			ibdev_err(ibdev, "failed to alloc RQ wrid.\n");
989 			ret = -ENOMEM;
990 			goto err_sq;
991 		}
992 	}
993 
994 	hr_qp->sq.wrid = sq_wrid;
995 	hr_qp->rq.wrid = rq_wrid;
996 	return 0;
997 err_sq:
998 	kfree(sq_wrid);
999 
1000 	return ret;
1001 }
1002 
1003 static void free_kernel_wrid(struct hns_roce_qp *hr_qp)
1004 {
1005 	kfree(hr_qp->rq.wrid);
1006 	kfree(hr_qp->sq.wrid);
1007 }
1008 
1009 static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1010 			struct ib_qp_init_attr *init_attr,
1011 			struct ib_udata *udata,
1012 			struct hns_roce_ib_create_qp *ucmd)
1013 {
1014 	struct ib_device *ibdev = &hr_dev->ib_dev;
1015 	int ret;
1016 
1017 	if (init_attr->cap.max_inline_data > hr_dev->caps.max_sq_inline)
1018 		init_attr->cap.max_inline_data = hr_dev->caps.max_sq_inline;
1019 
1020 	hr_qp->max_inline_data = init_attr->cap.max_inline_data;
1021 
1022 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1023 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
1024 	else
1025 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
1026 
1027 	ret = set_rq_size(hr_dev, &init_attr->cap, hr_qp,
1028 			  hns_roce_qp_has_rq(init_attr), !!udata);
1029 	if (ret) {
1030 		ibdev_err(ibdev, "failed to set user RQ size, ret = %d.\n",
1031 			  ret);
1032 		return ret;
1033 	}
1034 
1035 	if (udata) {
1036 		ret = ib_copy_from_udata(ucmd, udata,
1037 					 min(udata->inlen, sizeof(*ucmd)));
1038 		if (ret) {
1039 			ibdev_err(ibdev,
1040 				  "failed to copy QP ucmd, ret = %d\n", ret);
1041 			return ret;
1042 		}
1043 
1044 		ret = set_user_sq_size(hr_dev, &init_attr->cap, hr_qp, ucmd);
1045 		if (ret)
1046 			ibdev_err(ibdev,
1047 				  "failed to set user SQ size, ret = %d.\n",
1048 				  ret);
1049 	} else {
1050 		ret = set_kernel_sq_size(hr_dev, &init_attr->cap, hr_qp);
1051 		if (ret)
1052 			ibdev_err(ibdev,
1053 				  "failed to set kernel SQ size, ret = %d.\n",
1054 				  ret);
1055 	}
1056 
1057 	return ret;
1058 }
1059 
1060 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
1061 				     struct ib_pd *ib_pd,
1062 				     struct ib_qp_init_attr *init_attr,
1063 				     struct ib_udata *udata,
1064 				     struct hns_roce_qp *hr_qp)
1065 {
1066 	struct hns_roce_ib_create_qp_resp resp = {};
1067 	struct ib_device *ibdev = &hr_dev->ib_dev;
1068 	struct hns_roce_ib_create_qp ucmd;
1069 	int ret;
1070 
1071 	mutex_init(&hr_qp->mutex);
1072 	spin_lock_init(&hr_qp->sq.lock);
1073 	spin_lock_init(&hr_qp->rq.lock);
1074 
1075 	hr_qp->state = IB_QPS_RESET;
1076 	hr_qp->flush_flag = 0;
1077 
1078 	if (init_attr->create_flags)
1079 		return -EOPNOTSUPP;
1080 
1081 	ret = set_qp_param(hr_dev, hr_qp, init_attr, udata, &ucmd);
1082 	if (ret) {
1083 		ibdev_err(ibdev, "failed to set QP param, ret = %d.\n", ret);
1084 		return ret;
1085 	}
1086 
1087 	if (!udata) {
1088 		ret = alloc_kernel_wrid(hr_dev, hr_qp);
1089 		if (ret) {
1090 			ibdev_err(ibdev, "failed to alloc wrid, ret = %d.\n",
1091 				  ret);
1092 			return ret;
1093 		}
1094 	}
1095 
1096 	ret = alloc_qp_buf(hr_dev, hr_qp, init_attr, udata, ucmd.buf_addr);
1097 	if (ret) {
1098 		ibdev_err(ibdev, "failed to alloc QP buffer, ret = %d.\n", ret);
1099 		goto err_buf;
1100 	}
1101 
1102 	ret = alloc_qpn(hr_dev, hr_qp);
1103 	if (ret) {
1104 		ibdev_err(ibdev, "failed to alloc QPN, ret = %d.\n", ret);
1105 		goto err_qpn;
1106 	}
1107 
1108 	ret = alloc_qp_db(hr_dev, hr_qp, init_attr, udata, &ucmd, &resp);
1109 	if (ret) {
1110 		ibdev_err(ibdev, "failed to alloc QP doorbell, ret = %d.\n",
1111 			  ret);
1112 		goto err_db;
1113 	}
1114 
1115 	ret = alloc_qpc(hr_dev, hr_qp);
1116 	if (ret) {
1117 		ibdev_err(ibdev, "failed to alloc QP context, ret = %d.\n",
1118 			  ret);
1119 		goto err_qpc;
1120 	}
1121 
1122 	ret = hns_roce_qp_store(hr_dev, hr_qp, init_attr);
1123 	if (ret) {
1124 		ibdev_err(ibdev, "failed to store QP, ret = %d.\n", ret);
1125 		goto err_store;
1126 	}
1127 
1128 	if (udata) {
1129 		resp.cap_flags = hr_qp->en_flags;
1130 		ret = ib_copy_to_udata(udata, &resp,
1131 				       min(udata->outlen, sizeof(resp)));
1132 		if (ret) {
1133 			ibdev_err(ibdev, "copy qp resp failed!\n");
1134 			goto err_store;
1135 		}
1136 	}
1137 
1138 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
1139 		ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
1140 		if (ret)
1141 			goto err_flow_ctrl;
1142 	}
1143 
1144 	hr_qp->ibqp.qp_num = hr_qp->qpn;
1145 	hr_qp->event = hns_roce_ib_qp_event;
1146 	refcount_set(&hr_qp->refcount, 1);
1147 	init_completion(&hr_qp->free);
1148 
1149 	return 0;
1150 
1151 err_flow_ctrl:
1152 	hns_roce_qp_remove(hr_dev, hr_qp);
1153 err_store:
1154 	free_qpc(hr_dev, hr_qp);
1155 err_qpc:
1156 	free_qp_db(hr_dev, hr_qp, udata);
1157 err_db:
1158 	free_qpn(hr_dev, hr_qp);
1159 err_qpn:
1160 	free_qp_buf(hr_dev, hr_qp);
1161 err_buf:
1162 	free_kernel_wrid(hr_qp);
1163 	return ret;
1164 }
1165 
1166 void hns_roce_qp_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1167 			 struct ib_udata *udata)
1168 {
1169 	if (refcount_dec_and_test(&hr_qp->refcount))
1170 		complete(&hr_qp->free);
1171 	wait_for_completion(&hr_qp->free);
1172 
1173 	free_qpc(hr_dev, hr_qp);
1174 	free_qpn(hr_dev, hr_qp);
1175 	free_qp_buf(hr_dev, hr_qp);
1176 	free_kernel_wrid(hr_qp);
1177 	free_qp_db(hr_dev, hr_qp, udata);
1178 }
1179 
1180 static int check_qp_type(struct hns_roce_dev *hr_dev, enum ib_qp_type type,
1181 			 bool is_user)
1182 {
1183 	switch (type) {
1184 	case IB_QPT_XRC_INI:
1185 	case IB_QPT_XRC_TGT:
1186 		if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC))
1187 			goto out;
1188 		break;
1189 	case IB_QPT_UD:
1190 		if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 &&
1191 		    is_user)
1192 			goto out;
1193 		break;
1194 	case IB_QPT_RC:
1195 	case IB_QPT_GSI:
1196 		break;
1197 	default:
1198 		goto out;
1199 	}
1200 
1201 	return 0;
1202 
1203 out:
1204 	ibdev_err(&hr_dev->ib_dev, "not support QP type %d\n", type);
1205 
1206 	return -EOPNOTSUPP;
1207 }
1208 
1209 int hns_roce_create_qp(struct ib_qp *qp, struct ib_qp_init_attr *init_attr,
1210 		       struct ib_udata *udata)
1211 {
1212 	struct ib_device *ibdev = qp->device;
1213 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
1214 	struct hns_roce_qp *hr_qp = to_hr_qp(qp);
1215 	struct ib_pd *pd = qp->pd;
1216 	int ret;
1217 
1218 	ret = check_qp_type(hr_dev, init_attr->qp_type, !!udata);
1219 	if (ret)
1220 		return ret;
1221 
1222 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1223 		hr_qp->xrcdn = to_hr_xrcd(init_attr->xrcd)->xrcdn;
1224 
1225 	if (init_attr->qp_type == IB_QPT_GSI) {
1226 		hr_qp->port = init_attr->port_num - 1;
1227 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
1228 	}
1229 
1230 	ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, hr_qp);
1231 	if (ret)
1232 		ibdev_err(ibdev, "Create QP type 0x%x failed(%d)\n",
1233 			  init_attr->qp_type, ret);
1234 
1235 	return ret;
1236 }
1237 
1238 int to_hr_qp_type(int qp_type)
1239 {
1240 	switch (qp_type) {
1241 	case IB_QPT_RC:
1242 		return SERV_TYPE_RC;
1243 	case IB_QPT_UD:
1244 	case IB_QPT_GSI:
1245 		return SERV_TYPE_UD;
1246 	case IB_QPT_XRC_INI:
1247 	case IB_QPT_XRC_TGT:
1248 		return SERV_TYPE_XRC;
1249 	default:
1250 		return -1;
1251 	}
1252 }
1253 
1254 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1255 			      struct hns_roce_qp *hr_qp,
1256 			      struct ib_qp_attr *attr, int attr_mask)
1257 {
1258 	enum ib_mtu active_mtu;
1259 	int p;
1260 
1261 	p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1262 	active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1263 
1264 	if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1265 	    attr->path_mtu > hr_dev->caps.max_mtu) ||
1266 	    attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1267 		ibdev_err(&hr_dev->ib_dev,
1268 			"attr path_mtu(%d)invalid while modify qp",
1269 			attr->path_mtu);
1270 		return -EINVAL;
1271 	}
1272 
1273 	return 0;
1274 }
1275 
1276 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1277 				  int attr_mask)
1278 {
1279 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1280 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1281 	int p;
1282 
1283 	if ((attr_mask & IB_QP_PORT) &&
1284 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1285 		ibdev_err(&hr_dev->ib_dev, "invalid attr, port_num = %u.\n",
1286 			  attr->port_num);
1287 		return -EINVAL;
1288 	}
1289 
1290 	if (attr_mask & IB_QP_PKEY_INDEX) {
1291 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1292 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1293 			ibdev_err(&hr_dev->ib_dev,
1294 				  "invalid attr, pkey_index = %u.\n",
1295 				  attr->pkey_index);
1296 			return -EINVAL;
1297 		}
1298 	}
1299 
1300 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1301 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1302 		ibdev_err(&hr_dev->ib_dev,
1303 			  "invalid attr, max_rd_atomic = %u.\n",
1304 			  attr->max_rd_atomic);
1305 		return -EINVAL;
1306 	}
1307 
1308 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1309 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1310 		ibdev_err(&hr_dev->ib_dev,
1311 			  "invalid attr, max_dest_rd_atomic = %u.\n",
1312 			  attr->max_dest_rd_atomic);
1313 		return -EINVAL;
1314 	}
1315 
1316 	if (attr_mask & IB_QP_PATH_MTU)
1317 		return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1318 
1319 	return 0;
1320 }
1321 
1322 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1323 		       int attr_mask, struct ib_udata *udata)
1324 {
1325 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1326 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1327 	enum ib_qp_state cur_state, new_state;
1328 	int ret = -EINVAL;
1329 
1330 	mutex_lock(&hr_qp->mutex);
1331 
1332 	if (attr_mask & IB_QP_CUR_STATE && attr->cur_qp_state != hr_qp->state)
1333 		goto out;
1334 
1335 	cur_state = hr_qp->state;
1336 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1337 
1338 	if (ibqp->uobject &&
1339 	    (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1340 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB) {
1341 			hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1342 
1343 			if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
1344 				hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1345 		} else {
1346 			ibdev_warn(&hr_dev->ib_dev,
1347 				  "flush cqe is not supported in userspace!\n");
1348 			goto out;
1349 		}
1350 	}
1351 
1352 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1353 				attr_mask)) {
1354 		ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1355 		goto out;
1356 	}
1357 
1358 	ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1359 	if (ret)
1360 		goto out;
1361 
1362 	if (cur_state == new_state && cur_state == IB_QPS_RESET)
1363 		goto out;
1364 
1365 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1366 				    new_state);
1367 
1368 out:
1369 	mutex_unlock(&hr_qp->mutex);
1370 
1371 	return ret;
1372 }
1373 
1374 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1375 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1376 {
1377 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1378 		__acquire(&send_cq->lock);
1379 		__acquire(&recv_cq->lock);
1380 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1381 		spin_lock_irq(&send_cq->lock);
1382 		__acquire(&recv_cq->lock);
1383 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1384 		spin_lock_irq(&recv_cq->lock);
1385 		__acquire(&send_cq->lock);
1386 	} else if (send_cq == recv_cq) {
1387 		spin_lock_irq(&send_cq->lock);
1388 		__acquire(&recv_cq->lock);
1389 	} else if (send_cq->cqn < recv_cq->cqn) {
1390 		spin_lock_irq(&send_cq->lock);
1391 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1392 	} else {
1393 		spin_lock_irq(&recv_cq->lock);
1394 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1395 	}
1396 }
1397 
1398 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1399 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1400 			 __releases(&recv_cq->lock)
1401 {
1402 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1403 		__release(&recv_cq->lock);
1404 		__release(&send_cq->lock);
1405 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1406 		__release(&recv_cq->lock);
1407 		spin_unlock(&send_cq->lock);
1408 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1409 		__release(&send_cq->lock);
1410 		spin_unlock(&recv_cq->lock);
1411 	} else if (send_cq == recv_cq) {
1412 		__release(&recv_cq->lock);
1413 		spin_unlock_irq(&send_cq->lock);
1414 	} else if (send_cq->cqn < recv_cq->cqn) {
1415 		spin_unlock(&recv_cq->lock);
1416 		spin_unlock_irq(&send_cq->lock);
1417 	} else {
1418 		spin_unlock(&send_cq->lock);
1419 		spin_unlock_irq(&recv_cq->lock);
1420 	}
1421 }
1422 
1423 static inline void *get_wqe(struct hns_roce_qp *hr_qp, u32 offset)
1424 {
1425 	return hns_roce_buf_offset(hr_qp->mtr.kmem, offset);
1426 }
1427 
1428 void *hns_roce_get_recv_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1429 {
1430 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1431 }
1432 
1433 void *hns_roce_get_send_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1434 {
1435 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1436 }
1437 
1438 void *hns_roce_get_extend_sge(struct hns_roce_qp *hr_qp, unsigned int n)
1439 {
1440 	return get_wqe(hr_qp, hr_qp->sge.offset + (n << hr_qp->sge.sge_shift));
1441 }
1442 
1443 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, u32 nreq,
1444 			  struct ib_cq *ib_cq)
1445 {
1446 	struct hns_roce_cq *hr_cq;
1447 	u32 cur;
1448 
1449 	cur = hr_wq->head - hr_wq->tail;
1450 	if (likely(cur + nreq < hr_wq->wqe_cnt))
1451 		return false;
1452 
1453 	hr_cq = to_hr_cq(ib_cq);
1454 	spin_lock(&hr_cq->lock);
1455 	cur = hr_wq->head - hr_wq->tail;
1456 	spin_unlock(&hr_cq->lock);
1457 
1458 	return cur + nreq >= hr_wq->wqe_cnt;
1459 }
1460 
1461 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1462 {
1463 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1464 	unsigned int reserved_from_bot;
1465 	unsigned int i;
1466 
1467 	qp_table->idx_table.spare_idx = kcalloc(hr_dev->caps.num_qps,
1468 					sizeof(u32), GFP_KERNEL);
1469 	if (!qp_table->idx_table.spare_idx)
1470 		return -ENOMEM;
1471 
1472 	mutex_init(&qp_table->scc_mutex);
1473 	mutex_init(&qp_table->bank_mutex);
1474 	xa_init(&hr_dev->qp_table_xa);
1475 
1476 	reserved_from_bot = hr_dev->caps.reserved_qps;
1477 
1478 	for (i = 0; i < reserved_from_bot; i++) {
1479 		hr_dev->qp_table.bank[get_qp_bankid(i)].inuse++;
1480 		hr_dev->qp_table.bank[get_qp_bankid(i)].min++;
1481 	}
1482 
1483 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) {
1484 		ida_init(&hr_dev->qp_table.bank[i].ida);
1485 		hr_dev->qp_table.bank[i].max = hr_dev->caps.num_qps /
1486 					       HNS_ROCE_QP_BANK_NUM - 1;
1487 		hr_dev->qp_table.bank[i].next = hr_dev->qp_table.bank[i].min;
1488 	}
1489 
1490 	return 0;
1491 }
1492 
1493 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1494 {
1495 	int i;
1496 
1497 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++)
1498 		ida_destroy(&hr_dev->qp_table.bank[i].ida);
1499 	kfree(hr_dev->qp_table.idx_table.spare_idx);
1500 }
1501