1 /*
2  * Copyright (c) 2016 Hisilicon Limited.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/pci.h>
35 #include <rdma/ib_addr.h>
36 #include <rdma/ib_umem.h>
37 #include <rdma/uverbs_ioctl.h>
38 #include "hns_roce_common.h"
39 #include "hns_roce_device.h"
40 #include "hns_roce_hem.h"
41 
42 static void flush_work_handle(struct work_struct *work)
43 {
44 	struct hns_roce_work *flush_work = container_of(work,
45 					struct hns_roce_work, work);
46 	struct hns_roce_qp *hr_qp = container_of(flush_work,
47 					struct hns_roce_qp, flush_work);
48 	struct device *dev = flush_work->hr_dev->dev;
49 	struct ib_qp_attr attr;
50 	int attr_mask;
51 	int ret;
52 
53 	attr_mask = IB_QP_STATE;
54 	attr.qp_state = IB_QPS_ERR;
55 
56 	if (test_and_clear_bit(HNS_ROCE_FLUSH_FLAG, &hr_qp->flush_flag)) {
57 		ret = hns_roce_modify_qp(&hr_qp->ibqp, &attr, attr_mask, NULL);
58 		if (ret)
59 			dev_err(dev, "modify QP to error state failed(%d) during CQE flush\n",
60 				ret);
61 	}
62 
63 	/*
64 	 * make sure we signal QP destroy leg that flush QP was completed
65 	 * so that it can safely proceed ahead now and destroy QP
66 	 */
67 	if (refcount_dec_and_test(&hr_qp->refcount))
68 		complete(&hr_qp->free);
69 }
70 
71 void init_flush_work(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
72 {
73 	struct hns_roce_work *flush_work = &hr_qp->flush_work;
74 
75 	flush_work->hr_dev = hr_dev;
76 	INIT_WORK(&flush_work->work, flush_work_handle);
77 	refcount_inc(&hr_qp->refcount);
78 	queue_work(hr_dev->irq_workq, &flush_work->work);
79 }
80 
81 void flush_cqe(struct hns_roce_dev *dev, struct hns_roce_qp *qp)
82 {
83 	/*
84 	 * Hip08 hardware cannot flush the WQEs in SQ/RQ if the QP state
85 	 * gets into errored mode. Hence, as a workaround to this
86 	 * hardware limitation, driver needs to assist in flushing. But
87 	 * the flushing operation uses mailbox to convey the QP state to
88 	 * the hardware and which can sleep due to the mutex protection
89 	 * around the mailbox calls. Hence, use the deferred flush for
90 	 * now.
91 	 */
92 	if (!test_and_set_bit(HNS_ROCE_FLUSH_FLAG, &qp->flush_flag))
93 		init_flush_work(dev, qp);
94 }
95 
96 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
97 {
98 	struct device *dev = hr_dev->dev;
99 	struct hns_roce_qp *qp;
100 
101 	xa_lock(&hr_dev->qp_table_xa);
102 	qp = __hns_roce_qp_lookup(hr_dev, qpn);
103 	if (qp)
104 		refcount_inc(&qp->refcount);
105 	xa_unlock(&hr_dev->qp_table_xa);
106 
107 	if (!qp) {
108 		dev_warn(dev, "async event for bogus QP %08x\n", qpn);
109 		return;
110 	}
111 
112 	if (event_type == HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR ||
113 	    event_type == HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR ||
114 	    event_type == HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR ||
115 	    event_type == HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION ||
116 	    event_type == HNS_ROCE_EVENT_TYPE_INVALID_XRCETH) {
117 		qp->state = IB_QPS_ERR;
118 
119 		flush_cqe(hr_dev, qp);
120 	}
121 
122 	qp->event(qp, (enum hns_roce_event)event_type);
123 
124 	if (refcount_dec_and_test(&qp->refcount))
125 		complete(&qp->free);
126 }
127 
128 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
129 				 enum hns_roce_event type)
130 {
131 	struct ib_qp *ibqp = &hr_qp->ibqp;
132 	struct ib_event event;
133 
134 	if (ibqp->event_handler) {
135 		event.device = ibqp->device;
136 		event.element.qp = ibqp;
137 		switch (type) {
138 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
139 			event.event = IB_EVENT_PATH_MIG;
140 			break;
141 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
142 			event.event = IB_EVENT_COMM_EST;
143 			break;
144 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
145 			event.event = IB_EVENT_SQ_DRAINED;
146 			break;
147 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
148 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
149 			break;
150 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
151 			event.event = IB_EVENT_QP_FATAL;
152 			break;
153 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
154 			event.event = IB_EVENT_PATH_MIG_ERR;
155 			break;
156 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
157 			event.event = IB_EVENT_QP_REQ_ERR;
158 			break;
159 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
160 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
161 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
162 			event.event = IB_EVENT_QP_ACCESS_ERR;
163 			break;
164 		default:
165 			dev_dbg(ibqp->device->dev.parent, "roce_ib: Unexpected event type %d on QP %06lx\n",
166 				type, hr_qp->qpn);
167 			return;
168 		}
169 		ibqp->event_handler(&event, ibqp->qp_context);
170 	}
171 }
172 
173 static u8 get_affinity_cq_bank(u8 qp_bank)
174 {
175 	return (qp_bank >> 1) & CQ_BANKID_MASK;
176 }
177 
178 static u8 get_least_load_bankid_for_qp(struct ib_qp_init_attr *init_attr,
179 					struct hns_roce_bank *bank)
180 {
181 #define INVALID_LOAD_QPNUM 0xFFFFFFFF
182 	struct ib_cq *scq = init_attr->send_cq;
183 	u32 least_load = INVALID_LOAD_QPNUM;
184 	unsigned long cqn = 0;
185 	u8 bankid = 0;
186 	u32 bankcnt;
187 	u8 i;
188 
189 	if (scq)
190 		cqn = to_hr_cq(scq)->cqn;
191 
192 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) {
193 		if (scq && (get_affinity_cq_bank(i) != (cqn & CQ_BANKID_MASK)))
194 			continue;
195 
196 		bankcnt = bank[i].inuse;
197 		if (bankcnt < least_load) {
198 			least_load = bankcnt;
199 			bankid = i;
200 		}
201 	}
202 
203 	return bankid;
204 }
205 
206 static int alloc_qpn_with_bankid(struct hns_roce_bank *bank, u8 bankid,
207 				 unsigned long *qpn)
208 {
209 	int id;
210 
211 	id = ida_alloc_range(&bank->ida, bank->next, bank->max, GFP_KERNEL);
212 	if (id < 0) {
213 		id = ida_alloc_range(&bank->ida, bank->min, bank->max,
214 				     GFP_KERNEL);
215 		if (id < 0)
216 			return id;
217 	}
218 
219 	/* the QPN should keep increasing until the max value is reached. */
220 	bank->next = (id + 1) > bank->max ? bank->min : id + 1;
221 
222 	/* the lower 3 bits is bankid */
223 	*qpn = (id << 3) | bankid;
224 
225 	return 0;
226 }
227 static int alloc_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
228 		     struct ib_qp_init_attr *init_attr)
229 {
230 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
231 	unsigned long num = 0;
232 	u8 bankid;
233 	int ret;
234 
235 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI) {
236 		num = 1;
237 	} else {
238 		mutex_lock(&qp_table->bank_mutex);
239 		bankid = get_least_load_bankid_for_qp(init_attr, qp_table->bank);
240 
241 		ret = alloc_qpn_with_bankid(&qp_table->bank[bankid], bankid,
242 					    &num);
243 		if (ret) {
244 			ibdev_err(&hr_dev->ib_dev,
245 				  "failed to alloc QPN, ret = %d\n", ret);
246 			mutex_unlock(&qp_table->bank_mutex);
247 			return ret;
248 		}
249 
250 		qp_table->bank[bankid].inuse++;
251 		mutex_unlock(&qp_table->bank_mutex);
252 	}
253 
254 	hr_qp->qpn = num;
255 
256 	return 0;
257 }
258 
259 static void add_qp_to_list(struct hns_roce_dev *hr_dev,
260 			   struct hns_roce_qp *hr_qp,
261 			   struct ib_cq *send_cq, struct ib_cq *recv_cq)
262 {
263 	struct hns_roce_cq *hr_send_cq, *hr_recv_cq;
264 	unsigned long flags;
265 
266 	hr_send_cq = send_cq ? to_hr_cq(send_cq) : NULL;
267 	hr_recv_cq = recv_cq ? to_hr_cq(recv_cq) : NULL;
268 
269 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
270 	hns_roce_lock_cqs(hr_send_cq, hr_recv_cq);
271 
272 	list_add_tail(&hr_qp->node, &hr_dev->qp_list);
273 	if (hr_send_cq)
274 		list_add_tail(&hr_qp->sq_node, &hr_send_cq->sq_list);
275 	if (hr_recv_cq)
276 		list_add_tail(&hr_qp->rq_node, &hr_recv_cq->rq_list);
277 
278 	hns_roce_unlock_cqs(hr_send_cq, hr_recv_cq);
279 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
280 }
281 
282 static int hns_roce_qp_store(struct hns_roce_dev *hr_dev,
283 			     struct hns_roce_qp *hr_qp,
284 			     struct ib_qp_init_attr *init_attr)
285 {
286 	struct xarray *xa = &hr_dev->qp_table_xa;
287 	int ret;
288 
289 	if (!hr_qp->qpn)
290 		return -EINVAL;
291 
292 	ret = xa_err(xa_store_irq(xa, hr_qp->qpn, hr_qp, GFP_KERNEL));
293 	if (ret)
294 		dev_err(hr_dev->dev, "failed to xa store for QPC\n");
295 	else
296 		/* add QP to device's QP list for softwc */
297 		add_qp_to_list(hr_dev, hr_qp, init_attr->send_cq,
298 			       init_attr->recv_cq);
299 
300 	return ret;
301 }
302 
303 static int alloc_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
304 {
305 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
306 	struct device *dev = hr_dev->dev;
307 	int ret;
308 
309 	if (!hr_qp->qpn)
310 		return -EINVAL;
311 
312 	/* Alloc memory for QPC */
313 	ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
314 	if (ret) {
315 		dev_err(dev, "failed to get QPC table\n");
316 		goto err_out;
317 	}
318 
319 	/* Alloc memory for IRRL */
320 	ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
321 	if (ret) {
322 		dev_err(dev, "failed to get IRRL table\n");
323 		goto err_put_qp;
324 	}
325 
326 	if (hr_dev->caps.trrl_entry_sz) {
327 		/* Alloc memory for TRRL */
328 		ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
329 					 hr_qp->qpn);
330 		if (ret) {
331 			dev_err(dev, "failed to get TRRL table\n");
332 			goto err_put_irrl;
333 		}
334 	}
335 
336 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
337 		/* Alloc memory for SCC CTX */
338 		ret = hns_roce_table_get(hr_dev, &qp_table->sccc_table,
339 					 hr_qp->qpn);
340 		if (ret) {
341 			dev_err(dev, "failed to get SCC CTX table\n");
342 			goto err_put_trrl;
343 		}
344 	}
345 
346 	return 0;
347 
348 err_put_trrl:
349 	if (hr_dev->caps.trrl_entry_sz)
350 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
351 
352 err_put_irrl:
353 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
354 
355 err_put_qp:
356 	hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
357 
358 err_out:
359 	return ret;
360 }
361 
362 static void qp_user_mmap_entry_remove(struct hns_roce_qp *hr_qp)
363 {
364 	rdma_user_mmap_entry_remove(&hr_qp->dwqe_mmap_entry->rdma_entry);
365 }
366 
367 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
368 {
369 	struct xarray *xa = &hr_dev->qp_table_xa;
370 	unsigned long flags;
371 
372 	list_del(&hr_qp->node);
373 
374 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
375 		list_del(&hr_qp->sq_node);
376 
377 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_INI &&
378 	    hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
379 		list_del(&hr_qp->rq_node);
380 
381 	xa_lock_irqsave(xa, flags);
382 	__xa_erase(xa, hr_qp->qpn);
383 	xa_unlock_irqrestore(xa, flags);
384 }
385 
386 static void free_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
387 {
388 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
389 
390 	if (hr_dev->caps.trrl_entry_sz)
391 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
392 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
393 }
394 
395 static inline u8 get_qp_bankid(unsigned long qpn)
396 {
397 	/* The lower 3 bits of QPN are used to hash to different banks */
398 	return (u8)(qpn & GENMASK(2, 0));
399 }
400 
401 static void free_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
402 {
403 	u8 bankid;
404 
405 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI)
406 		return;
407 
408 	if (hr_qp->qpn < hr_dev->caps.reserved_qps)
409 		return;
410 
411 	bankid = get_qp_bankid(hr_qp->qpn);
412 
413 	ida_free(&hr_dev->qp_table.bank[bankid].ida, hr_qp->qpn >> 3);
414 
415 	mutex_lock(&hr_dev->qp_table.bank_mutex);
416 	hr_dev->qp_table.bank[bankid].inuse--;
417 	mutex_unlock(&hr_dev->qp_table.bank_mutex);
418 }
419 
420 static u32 proc_rq_sge(struct hns_roce_dev *dev, struct hns_roce_qp *hr_qp,
421 		       bool user)
422 {
423 	u32 max_sge = dev->caps.max_rq_sg;
424 
425 	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
426 		return max_sge;
427 
428 	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
429 	 * calculate number of max_sge with reserved SGEs when allocating wqe
430 	 * buf, so there is no need to do this again in kernel. But the number
431 	 * may exceed the capacity of SGEs recorded in the firmware, so the
432 	 * kernel driver should just adapt the value accordingly.
433 	 */
434 	if (user)
435 		max_sge = roundup_pow_of_two(max_sge + 1);
436 	else
437 		hr_qp->rq.rsv_sge = 1;
438 
439 	return max_sge;
440 }
441 
442 static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
443 		       struct hns_roce_qp *hr_qp, int has_rq, bool user)
444 {
445 	u32 max_sge = proc_rq_sge(hr_dev, hr_qp, user);
446 	u32 cnt;
447 
448 	/* If srq exist, set zero for relative number of rq */
449 	if (!has_rq) {
450 		hr_qp->rq.wqe_cnt = 0;
451 		hr_qp->rq.max_gs = 0;
452 		cap->max_recv_wr = 0;
453 		cap->max_recv_sge = 0;
454 
455 		return 0;
456 	}
457 
458 	/* Check the validity of QP support capacity */
459 	if (!cap->max_recv_wr || cap->max_recv_wr > hr_dev->caps.max_wqes ||
460 	    cap->max_recv_sge > max_sge) {
461 		ibdev_err(&hr_dev->ib_dev,
462 			  "RQ config error, depth = %u, sge = %u\n",
463 			  cap->max_recv_wr, cap->max_recv_sge);
464 		return -EINVAL;
465 	}
466 
467 	cnt = roundup_pow_of_two(max(cap->max_recv_wr, hr_dev->caps.min_wqes));
468 	if (cnt > hr_dev->caps.max_wqes) {
469 		ibdev_err(&hr_dev->ib_dev, "rq depth %u too large\n",
470 			  cap->max_recv_wr);
471 		return -EINVAL;
472 	}
473 
474 	hr_qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge) +
475 					      hr_qp->rq.rsv_sge);
476 
477 	hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz *
478 				    hr_qp->rq.max_gs);
479 
480 	hr_qp->rq.wqe_cnt = cnt;
481 
482 	cap->max_recv_wr = cnt;
483 	cap->max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
484 
485 	return 0;
486 }
487 
488 static u32 get_max_inline_data(struct hns_roce_dev *hr_dev,
489 			       struct ib_qp_cap *cap)
490 {
491 	if (cap->max_inline_data) {
492 		cap->max_inline_data = roundup_pow_of_two(cap->max_inline_data);
493 		return min(cap->max_inline_data,
494 			   hr_dev->caps.max_sq_inline);
495 	}
496 
497 	return 0;
498 }
499 
500 static void update_inline_data(struct hns_roce_qp *hr_qp,
501 			       struct ib_qp_cap *cap)
502 {
503 	u32 sge_num = hr_qp->sq.ext_sge_cnt;
504 
505 	if (hr_qp->config & HNS_ROCE_EXSGE_FLAGS) {
506 		if (!(hr_qp->ibqp.qp_type == IB_QPT_GSI ||
507 		      hr_qp->ibqp.qp_type == IB_QPT_UD))
508 			sge_num = max((u32)HNS_ROCE_SGE_IN_WQE, sge_num);
509 
510 		cap->max_inline_data = max(cap->max_inline_data,
511 					   sge_num * HNS_ROCE_SGE_SIZE);
512 	}
513 
514 	hr_qp->max_inline_data = cap->max_inline_data;
515 }
516 
517 static u32 get_sge_num_from_max_send_sge(bool is_ud_or_gsi,
518 					 u32 max_send_sge)
519 {
520 	unsigned int std_sge_num;
521 	unsigned int min_sge;
522 
523 	std_sge_num = is_ud_or_gsi ? 0 : HNS_ROCE_SGE_IN_WQE;
524 	min_sge = is_ud_or_gsi ? 1 : 0;
525 	return max_send_sge > std_sge_num ? (max_send_sge - std_sge_num) :
526 				min_sge;
527 }
528 
529 static unsigned int get_sge_num_from_max_inl_data(bool is_ud_or_gsi,
530 						  u32 max_inline_data)
531 {
532 	unsigned int inline_sge;
533 
534 	if (!max_inline_data)
535 		return 0;
536 
537 	/*
538 	 * if max_inline_data less than
539 	 * HNS_ROCE_SGE_IN_WQE * HNS_ROCE_SGE_SIZE,
540 	 * In addition to ud's mode, no need to extend sge.
541 	 */
542 	inline_sge = roundup_pow_of_two(max_inline_data) / HNS_ROCE_SGE_SIZE;
543 	if (!is_ud_or_gsi && inline_sge <= HNS_ROCE_SGE_IN_WQE)
544 		inline_sge = 0;
545 
546 	return inline_sge;
547 }
548 
549 static void set_ext_sge_param(struct hns_roce_dev *hr_dev, u32 sq_wqe_cnt,
550 			      struct hns_roce_qp *hr_qp, struct ib_qp_cap *cap)
551 {
552 	bool is_ud_or_gsi = (hr_qp->ibqp.qp_type == IB_QPT_GSI ||
553 				hr_qp->ibqp.qp_type == IB_QPT_UD);
554 	unsigned int std_sge_num;
555 	u32 inline_ext_sge = 0;
556 	u32 ext_wqe_sge_cnt;
557 	u32 total_sge_cnt;
558 
559 	cap->max_inline_data = get_max_inline_data(hr_dev, cap);
560 
561 	hr_qp->sge.sge_shift = HNS_ROCE_SGE_SHIFT;
562 	std_sge_num = is_ud_or_gsi ? 0 : HNS_ROCE_SGE_IN_WQE;
563 	ext_wqe_sge_cnt = get_sge_num_from_max_send_sge(is_ud_or_gsi,
564 							cap->max_send_sge);
565 
566 	if (hr_qp->config & HNS_ROCE_EXSGE_FLAGS) {
567 		inline_ext_sge = max(ext_wqe_sge_cnt,
568 				     get_sge_num_from_max_inl_data(is_ud_or_gsi,
569 							 cap->max_inline_data));
570 		hr_qp->sq.ext_sge_cnt = inline_ext_sge ?
571 					roundup_pow_of_two(inline_ext_sge) : 0;
572 
573 		hr_qp->sq.max_gs = max(1U, (hr_qp->sq.ext_sge_cnt + std_sge_num));
574 		hr_qp->sq.max_gs = min(hr_qp->sq.max_gs, hr_dev->caps.max_sq_sg);
575 
576 		ext_wqe_sge_cnt = hr_qp->sq.ext_sge_cnt;
577 	} else {
578 		hr_qp->sq.max_gs = max(1U, cap->max_send_sge);
579 		hr_qp->sq.max_gs = min(hr_qp->sq.max_gs, hr_dev->caps.max_sq_sg);
580 		hr_qp->sq.ext_sge_cnt = hr_qp->sq.max_gs;
581 	}
582 
583 	/* If the number of extended sge is not zero, they MUST use the
584 	 * space of HNS_HW_PAGE_SIZE at least.
585 	 */
586 	if (ext_wqe_sge_cnt) {
587 		total_sge_cnt = roundup_pow_of_two(sq_wqe_cnt * ext_wqe_sge_cnt);
588 		hr_qp->sge.sge_cnt = max(total_sge_cnt,
589 				(u32)HNS_HW_PAGE_SIZE / HNS_ROCE_SGE_SIZE);
590 	}
591 
592 	update_inline_data(hr_qp, cap);
593 }
594 
595 static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
596 					struct ib_qp_cap *cap,
597 					struct hns_roce_ib_create_qp *ucmd)
598 {
599 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
600 	u8 max_sq_stride = ilog2(roundup_sq_stride);
601 
602 	/* Sanity check SQ size before proceeding */
603 	if (ucmd->log_sq_stride > max_sq_stride ||
604 	    ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
605 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ stride size.\n");
606 		return -EINVAL;
607 	}
608 
609 	if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
610 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ SGE size %u.\n",
611 			  cap->max_send_sge);
612 		return -EINVAL;
613 	}
614 
615 	return 0;
616 }
617 
618 static int set_user_sq_size(struct hns_roce_dev *hr_dev,
619 			    struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp,
620 			    struct hns_roce_ib_create_qp *ucmd)
621 {
622 	struct ib_device *ibdev = &hr_dev->ib_dev;
623 	u32 cnt = 0;
624 	int ret;
625 
626 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
627 	    cnt > hr_dev->caps.max_wqes)
628 		return -EINVAL;
629 
630 	ret = check_sq_size_with_integrity(hr_dev, cap, ucmd);
631 	if (ret) {
632 		ibdev_err(ibdev, "failed to check user SQ size, ret = %d.\n",
633 			  ret);
634 		return ret;
635 	}
636 
637 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
638 
639 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
640 	hr_qp->sq.wqe_cnt = cnt;
641 	cap->max_send_sge = hr_qp->sq.max_gs;
642 
643 	return 0;
644 }
645 
646 static int set_wqe_buf_attr(struct hns_roce_dev *hr_dev,
647 			    struct hns_roce_qp *hr_qp,
648 			    struct hns_roce_buf_attr *buf_attr)
649 {
650 	int buf_size;
651 	int idx = 0;
652 
653 	hr_qp->buff_size = 0;
654 
655 	/* SQ WQE */
656 	hr_qp->sq.offset = 0;
657 	buf_size = to_hr_hem_entries_size(hr_qp->sq.wqe_cnt,
658 					  hr_qp->sq.wqe_shift);
659 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
660 		buf_attr->region[idx].size = buf_size;
661 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sq_hop_num;
662 		idx++;
663 		hr_qp->buff_size += buf_size;
664 	}
665 
666 	/* extend SGE WQE in SQ */
667 	hr_qp->sge.offset = hr_qp->buff_size;
668 	buf_size = to_hr_hem_entries_size(hr_qp->sge.sge_cnt,
669 					  hr_qp->sge.sge_shift);
670 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
671 		buf_attr->region[idx].size = buf_size;
672 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sge_hop_num;
673 		idx++;
674 		hr_qp->buff_size += buf_size;
675 	}
676 
677 	/* RQ WQE */
678 	hr_qp->rq.offset = hr_qp->buff_size;
679 	buf_size = to_hr_hem_entries_size(hr_qp->rq.wqe_cnt,
680 					  hr_qp->rq.wqe_shift);
681 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
682 		buf_attr->region[idx].size = buf_size;
683 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_rq_hop_num;
684 		idx++;
685 		hr_qp->buff_size += buf_size;
686 	}
687 
688 	if (hr_qp->buff_size < 1)
689 		return -EINVAL;
690 
691 	buf_attr->page_shift = HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
692 	buf_attr->region_count = idx;
693 
694 	return 0;
695 }
696 
697 static int set_kernel_sq_size(struct hns_roce_dev *hr_dev,
698 			      struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp)
699 {
700 	struct ib_device *ibdev = &hr_dev->ib_dev;
701 	u32 cnt;
702 
703 	if (!cap->max_send_wr || cap->max_send_wr > hr_dev->caps.max_wqes ||
704 	    cap->max_send_sge > hr_dev->caps.max_sq_sg) {
705 		ibdev_err(ibdev, "failed to check SQ WR or SGE num.\n");
706 		return -EINVAL;
707 	}
708 
709 	cnt = roundup_pow_of_two(max(cap->max_send_wr, hr_dev->caps.min_wqes));
710 	if (cnt > hr_dev->caps.max_wqes) {
711 		ibdev_err(ibdev, "failed to check WQE num, WQE num = %u.\n",
712 			  cnt);
713 		return -EINVAL;
714 	}
715 
716 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
717 	hr_qp->sq.wqe_cnt = cnt;
718 
719 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
720 
721 	/* sync the parameters of kernel QP to user's configuration */
722 	cap->max_send_wr = cnt;
723 	cap->max_send_sge = hr_qp->sq.max_gs;
724 
725 	return 0;
726 }
727 
728 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
729 {
730 	if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
731 		return 0;
732 
733 	return 1;
734 }
735 
736 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
737 {
738 	if (attr->qp_type == IB_QPT_XRC_INI ||
739 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
740 	    !attr->cap.max_recv_wr)
741 		return 0;
742 
743 	return 1;
744 }
745 
746 static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
747 			struct ib_qp_init_attr *init_attr,
748 			struct ib_udata *udata, unsigned long addr)
749 {
750 	struct ib_device *ibdev = &hr_dev->ib_dev;
751 	struct hns_roce_buf_attr buf_attr = {};
752 	int ret;
753 
754 	ret = set_wqe_buf_attr(hr_dev, hr_qp, &buf_attr);
755 	if (ret) {
756 		ibdev_err(ibdev, "failed to split WQE buf, ret = %d.\n", ret);
757 		goto err_inline;
758 	}
759 	ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, &buf_attr,
760 				  PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
761 				  udata, addr);
762 	if (ret) {
763 		ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret);
764 		goto err_inline;
765 	}
766 
767 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_DIRECT_WQE)
768 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_DIRECT_WQE;
769 
770 	return 0;
771 
772 err_inline:
773 
774 	return ret;
775 }
776 
777 static void free_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
778 {
779 	hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
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 	struct hns_roce_ucontext *uctx;
1016 	int ret;
1017 
1018 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1019 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
1020 	else
1021 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
1022 
1023 	ret = set_rq_size(hr_dev, &init_attr->cap, hr_qp,
1024 			  hns_roce_qp_has_rq(init_attr), !!udata);
1025 	if (ret) {
1026 		ibdev_err(ibdev, "failed to set user RQ size, ret = %d.\n",
1027 			  ret);
1028 		return ret;
1029 	}
1030 
1031 	if (udata) {
1032 		ret = ib_copy_from_udata(ucmd, udata,
1033 					 min(udata->inlen, sizeof(*ucmd)));
1034 		if (ret) {
1035 			ibdev_err(ibdev,
1036 				  "failed to copy QP ucmd, ret = %d\n", ret);
1037 			return ret;
1038 		}
1039 
1040 		uctx = rdma_udata_to_drv_context(udata, struct hns_roce_ucontext,
1041 						 ibucontext);
1042 		hr_qp->config = uctx->config;
1043 		ret = set_user_sq_size(hr_dev, &init_attr->cap, hr_qp, ucmd);
1044 		if (ret)
1045 			ibdev_err(ibdev,
1046 				  "failed to set user SQ size, ret = %d.\n",
1047 				  ret);
1048 	} else {
1049 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
1050 			hr_qp->config = HNS_ROCE_EXSGE_FLAGS;
1051 		ret = set_kernel_sq_size(hr_dev, &init_attr->cap, hr_qp);
1052 		if (ret)
1053 			ibdev_err(ibdev,
1054 				  "failed to set kernel SQ size, ret = %d.\n",
1055 				  ret);
1056 	}
1057 
1058 	return ret;
1059 }
1060 
1061 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
1062 				     struct ib_pd *ib_pd,
1063 				     struct ib_qp_init_attr *init_attr,
1064 				     struct ib_udata *udata,
1065 				     struct hns_roce_qp *hr_qp)
1066 {
1067 	struct hns_roce_ib_create_qp_resp resp = {};
1068 	struct ib_device *ibdev = &hr_dev->ib_dev;
1069 	struct hns_roce_ib_create_qp ucmd = {};
1070 	int ret;
1071 
1072 	mutex_init(&hr_qp->mutex);
1073 	spin_lock_init(&hr_qp->sq.lock);
1074 	spin_lock_init(&hr_qp->rq.lock);
1075 
1076 	hr_qp->state = IB_QPS_RESET;
1077 	hr_qp->flush_flag = 0;
1078 
1079 	if (init_attr->create_flags)
1080 		return -EOPNOTSUPP;
1081 
1082 	ret = set_qp_param(hr_dev, hr_qp, init_attr, udata, &ucmd);
1083 	if (ret) {
1084 		ibdev_err(ibdev, "failed to set QP param, ret = %d.\n", ret);
1085 		return ret;
1086 	}
1087 
1088 	if (!udata) {
1089 		ret = alloc_kernel_wrid(hr_dev, hr_qp);
1090 		if (ret) {
1091 			ibdev_err(ibdev, "failed to alloc wrid, ret = %d.\n",
1092 				  ret);
1093 			return ret;
1094 		}
1095 	}
1096 
1097 	ret = alloc_qp_buf(hr_dev, hr_qp, init_attr, udata, ucmd.buf_addr);
1098 	if (ret) {
1099 		ibdev_err(ibdev, "failed to alloc QP buffer, ret = %d.\n", ret);
1100 		goto err_buf;
1101 	}
1102 
1103 	ret = alloc_qpn(hr_dev, hr_qp, init_attr);
1104 	if (ret) {
1105 		ibdev_err(ibdev, "failed to alloc QPN, ret = %d.\n", ret);
1106 		goto err_qpn;
1107 	}
1108 
1109 	ret = alloc_qp_db(hr_dev, hr_qp, init_attr, udata, &ucmd, &resp);
1110 	if (ret) {
1111 		ibdev_err(ibdev, "failed to alloc QP doorbell, ret = %d.\n",
1112 			  ret);
1113 		goto err_db;
1114 	}
1115 
1116 	ret = alloc_qpc(hr_dev, hr_qp);
1117 	if (ret) {
1118 		ibdev_err(ibdev, "failed to alloc QP context, ret = %d.\n",
1119 			  ret);
1120 		goto err_qpc;
1121 	}
1122 
1123 	ret = hns_roce_qp_store(hr_dev, hr_qp, init_attr);
1124 	if (ret) {
1125 		ibdev_err(ibdev, "failed to store QP, ret = %d.\n", ret);
1126 		goto err_store;
1127 	}
1128 
1129 	if (udata) {
1130 		resp.cap_flags = hr_qp->en_flags;
1131 		ret = ib_copy_to_udata(udata, &resp,
1132 				       min(udata->outlen, sizeof(resp)));
1133 		if (ret) {
1134 			ibdev_err(ibdev, "copy qp resp failed!\n");
1135 			goto err_store;
1136 		}
1137 	}
1138 
1139 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
1140 		ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
1141 		if (ret)
1142 			goto err_flow_ctrl;
1143 	}
1144 
1145 	hr_qp->ibqp.qp_num = hr_qp->qpn;
1146 	hr_qp->event = hns_roce_ib_qp_event;
1147 	refcount_set(&hr_qp->refcount, 1);
1148 	init_completion(&hr_qp->free);
1149 
1150 	return 0;
1151 
1152 err_flow_ctrl:
1153 	hns_roce_qp_remove(hr_dev, hr_qp);
1154 err_store:
1155 	free_qpc(hr_dev, hr_qp);
1156 err_qpc:
1157 	free_qp_db(hr_dev, hr_qp, udata);
1158 err_db:
1159 	free_qpn(hr_dev, hr_qp);
1160 err_qpn:
1161 	free_qp_buf(hr_dev, hr_qp);
1162 err_buf:
1163 	free_kernel_wrid(hr_qp);
1164 	return ret;
1165 }
1166 
1167 void hns_roce_qp_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1168 			 struct ib_udata *udata)
1169 {
1170 	if (refcount_dec_and_test(&hr_qp->refcount))
1171 		complete(&hr_qp->free);
1172 	wait_for_completion(&hr_qp->free);
1173 
1174 	free_qpc(hr_dev, hr_qp);
1175 	free_qpn(hr_dev, hr_qp);
1176 	free_qp_buf(hr_dev, hr_qp);
1177 	free_kernel_wrid(hr_qp);
1178 	free_qp_db(hr_dev, hr_qp, udata);
1179 }
1180 
1181 static int check_qp_type(struct hns_roce_dev *hr_dev, enum ib_qp_type type,
1182 			 bool is_user)
1183 {
1184 	switch (type) {
1185 	case IB_QPT_XRC_INI:
1186 	case IB_QPT_XRC_TGT:
1187 		if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC))
1188 			goto out;
1189 		break;
1190 	case IB_QPT_UD:
1191 		if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 &&
1192 		    is_user)
1193 			goto out;
1194 		break;
1195 	case IB_QPT_RC:
1196 	case IB_QPT_GSI:
1197 		break;
1198 	default:
1199 		goto out;
1200 	}
1201 
1202 	return 0;
1203 
1204 out:
1205 	ibdev_err(&hr_dev->ib_dev, "not support QP type %d\n", type);
1206 
1207 	return -EOPNOTSUPP;
1208 }
1209 
1210 int hns_roce_create_qp(struct ib_qp *qp, struct ib_qp_init_attr *init_attr,
1211 		       struct ib_udata *udata)
1212 {
1213 	struct ib_device *ibdev = qp->device;
1214 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
1215 	struct hns_roce_qp *hr_qp = to_hr_qp(qp);
1216 	struct ib_pd *pd = qp->pd;
1217 	int ret;
1218 
1219 	ret = check_qp_type(hr_dev, init_attr->qp_type, !!udata);
1220 	if (ret)
1221 		return ret;
1222 
1223 	if (init_attr->qp_type == IB_QPT_XRC_TGT)
1224 		hr_qp->xrcdn = to_hr_xrcd(init_attr->xrcd)->xrcdn;
1225 
1226 	if (init_attr->qp_type == IB_QPT_GSI) {
1227 		hr_qp->port = init_attr->port_num - 1;
1228 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
1229 	}
1230 
1231 	ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, hr_qp);
1232 	if (ret)
1233 		ibdev_err(ibdev, "create QP type 0x%x failed(%d)\n",
1234 			  init_attr->qp_type, ret);
1235 
1236 	return ret;
1237 }
1238 
1239 int to_hr_qp_type(int qp_type)
1240 {
1241 	switch (qp_type) {
1242 	case IB_QPT_RC:
1243 		return SERV_TYPE_RC;
1244 	case IB_QPT_UD:
1245 	case IB_QPT_GSI:
1246 		return SERV_TYPE_UD;
1247 	case IB_QPT_XRC_INI:
1248 	case IB_QPT_XRC_TGT:
1249 		return SERV_TYPE_XRC;
1250 	default:
1251 		return -1;
1252 	}
1253 }
1254 
1255 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1256 			      struct hns_roce_qp *hr_qp,
1257 			      struct ib_qp_attr *attr, int attr_mask)
1258 {
1259 	enum ib_mtu active_mtu;
1260 	int p;
1261 
1262 	p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1263 	active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1264 
1265 	if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1266 	    attr->path_mtu > hr_dev->caps.max_mtu) ||
1267 	    attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1268 		ibdev_err(&hr_dev->ib_dev,
1269 			"attr path_mtu(%d)invalid while modify qp",
1270 			attr->path_mtu);
1271 		return -EINVAL;
1272 	}
1273 
1274 	return 0;
1275 }
1276 
1277 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1278 				  int attr_mask)
1279 {
1280 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1281 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1282 	int p;
1283 
1284 	if ((attr_mask & IB_QP_PORT) &&
1285 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1286 		ibdev_err(&hr_dev->ib_dev, "invalid attr, port_num = %u.\n",
1287 			  attr->port_num);
1288 		return -EINVAL;
1289 	}
1290 
1291 	if (attr_mask & IB_QP_PKEY_INDEX) {
1292 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1293 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1294 			ibdev_err(&hr_dev->ib_dev,
1295 				  "invalid attr, pkey_index = %u.\n",
1296 				  attr->pkey_index);
1297 			return -EINVAL;
1298 		}
1299 	}
1300 
1301 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1302 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1303 		ibdev_err(&hr_dev->ib_dev,
1304 			  "invalid attr, max_rd_atomic = %u.\n",
1305 			  attr->max_rd_atomic);
1306 		return -EINVAL;
1307 	}
1308 
1309 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1310 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1311 		ibdev_err(&hr_dev->ib_dev,
1312 			  "invalid attr, max_dest_rd_atomic = %u.\n",
1313 			  attr->max_dest_rd_atomic);
1314 		return -EINVAL;
1315 	}
1316 
1317 	if (attr_mask & IB_QP_PATH_MTU)
1318 		return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1319 
1320 	return 0;
1321 }
1322 
1323 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1324 		       int attr_mask, struct ib_udata *udata)
1325 {
1326 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1327 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1328 	enum ib_qp_state cur_state, new_state;
1329 	int ret = -EINVAL;
1330 
1331 	mutex_lock(&hr_qp->mutex);
1332 
1333 	if (attr_mask & IB_QP_CUR_STATE && attr->cur_qp_state != hr_qp->state)
1334 		goto out;
1335 
1336 	cur_state = hr_qp->state;
1337 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1338 
1339 	if (ibqp->uobject &&
1340 	    (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1341 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB) {
1342 			hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1343 
1344 			if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
1345 				hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1346 		} else {
1347 			ibdev_warn(&hr_dev->ib_dev,
1348 				  "flush cqe is not supported in userspace!\n");
1349 			goto out;
1350 		}
1351 	}
1352 
1353 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1354 				attr_mask)) {
1355 		ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1356 		goto out;
1357 	}
1358 
1359 	ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1360 	if (ret)
1361 		goto out;
1362 
1363 	if (cur_state == new_state && cur_state == IB_QPS_RESET)
1364 		goto out;
1365 
1366 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1367 				    new_state, udata);
1368 
1369 out:
1370 	mutex_unlock(&hr_qp->mutex);
1371 
1372 	return ret;
1373 }
1374 
1375 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1376 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1377 {
1378 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1379 		__acquire(&send_cq->lock);
1380 		__acquire(&recv_cq->lock);
1381 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1382 		spin_lock(&send_cq->lock);
1383 		__acquire(&recv_cq->lock);
1384 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1385 		spin_lock(&recv_cq->lock);
1386 		__acquire(&send_cq->lock);
1387 	} else if (send_cq == recv_cq) {
1388 		spin_lock(&send_cq->lock);
1389 		__acquire(&recv_cq->lock);
1390 	} else if (send_cq->cqn < recv_cq->cqn) {
1391 		spin_lock(&send_cq->lock);
1392 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1393 	} else {
1394 		spin_lock(&recv_cq->lock);
1395 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1396 	}
1397 }
1398 
1399 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1400 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1401 			 __releases(&recv_cq->lock)
1402 {
1403 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1404 		__release(&recv_cq->lock);
1405 		__release(&send_cq->lock);
1406 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1407 		__release(&recv_cq->lock);
1408 		spin_unlock(&send_cq->lock);
1409 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1410 		__release(&send_cq->lock);
1411 		spin_unlock(&recv_cq->lock);
1412 	} else if (send_cq == recv_cq) {
1413 		__release(&recv_cq->lock);
1414 		spin_unlock(&send_cq->lock);
1415 	} else if (send_cq->cqn < recv_cq->cqn) {
1416 		spin_unlock(&recv_cq->lock);
1417 		spin_unlock(&send_cq->lock);
1418 	} else {
1419 		spin_unlock(&send_cq->lock);
1420 		spin_unlock(&recv_cq->lock);
1421 	}
1422 }
1423 
1424 static inline void *get_wqe(struct hns_roce_qp *hr_qp, u32 offset)
1425 {
1426 	return hns_roce_buf_offset(hr_qp->mtr.kmem, offset);
1427 }
1428 
1429 void *hns_roce_get_recv_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1430 {
1431 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1432 }
1433 
1434 void *hns_roce_get_send_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1435 {
1436 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1437 }
1438 
1439 void *hns_roce_get_extend_sge(struct hns_roce_qp *hr_qp, unsigned int n)
1440 {
1441 	return get_wqe(hr_qp, hr_qp->sge.offset + (n << hr_qp->sge.sge_shift));
1442 }
1443 
1444 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, u32 nreq,
1445 			  struct ib_cq *ib_cq)
1446 {
1447 	struct hns_roce_cq *hr_cq;
1448 	u32 cur;
1449 
1450 	cur = hr_wq->head - hr_wq->tail;
1451 	if (likely(cur + nreq < hr_wq->wqe_cnt))
1452 		return false;
1453 
1454 	hr_cq = to_hr_cq(ib_cq);
1455 	spin_lock(&hr_cq->lock);
1456 	cur = hr_wq->head - hr_wq->tail;
1457 	spin_unlock(&hr_cq->lock);
1458 
1459 	return cur + nreq >= hr_wq->wqe_cnt;
1460 }
1461 
1462 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1463 {
1464 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1465 	unsigned int reserved_from_bot;
1466 	unsigned int i;
1467 
1468 	qp_table->idx_table.spare_idx = kcalloc(hr_dev->caps.num_qps,
1469 					sizeof(u32), GFP_KERNEL);
1470 	if (!qp_table->idx_table.spare_idx)
1471 		return -ENOMEM;
1472 
1473 	mutex_init(&qp_table->scc_mutex);
1474 	mutex_init(&qp_table->bank_mutex);
1475 	xa_init(&hr_dev->qp_table_xa);
1476 
1477 	reserved_from_bot = hr_dev->caps.reserved_qps;
1478 
1479 	for (i = 0; i < reserved_from_bot; i++) {
1480 		hr_dev->qp_table.bank[get_qp_bankid(i)].inuse++;
1481 		hr_dev->qp_table.bank[get_qp_bankid(i)].min++;
1482 	}
1483 
1484 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) {
1485 		ida_init(&hr_dev->qp_table.bank[i].ida);
1486 		hr_dev->qp_table.bank[i].max = hr_dev->caps.num_qps /
1487 					       HNS_ROCE_QP_BANK_NUM - 1;
1488 		hr_dev->qp_table.bank[i].next = hr_dev->qp_table.bank[i].min;
1489 	}
1490 
1491 	return 0;
1492 }
1493 
1494 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1495 {
1496 	int i;
1497 
1498 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++)
1499 		ida_destroy(&hr_dev->qp_table.bank[i].ida);
1500 	kfree(hr_dev->qp_table.idx_table.spare_idx);
1501 }
1502