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 <linux/platform_device.h>
36 #include <rdma/ib_addr.h>
37 #include <rdma/ib_umem.h>
38 #include <rdma/uverbs_ioctl.h>
39 #include "hns_roce_common.h"
40 #include "hns_roce_device.h"
41 #include "hns_roce_hem.h"
42 
43 static void flush_work_handle(struct work_struct *work)
44 {
45 	struct hns_roce_work *flush_work = container_of(work,
46 					struct hns_roce_work, work);
47 	struct hns_roce_qp *hr_qp = container_of(flush_work,
48 					struct hns_roce_qp, flush_work);
49 	struct device *dev = flush_work->hr_dev->dev;
50 	struct ib_qp_attr attr;
51 	int attr_mask;
52 	int ret;
53 
54 	attr_mask = IB_QP_STATE;
55 	attr.qp_state = IB_QPS_ERR;
56 
57 	if (test_and_clear_bit(HNS_ROCE_FLUSH_FLAG, &hr_qp->flush_flag)) {
58 		ret = hns_roce_modify_qp(&hr_qp->ibqp, &attr, attr_mask, NULL);
59 		if (ret)
60 			dev_err(dev, "Modify QP to error state failed(%d) during CQE flush\n",
61 				ret);
62 	}
63 
64 	/*
65 	 * make sure we signal QP destroy leg that flush QP was completed
66 	 * so that it can safely proceed ahead now and destroy QP
67 	 */
68 	if (refcount_dec_and_test(&hr_qp->refcount))
69 		complete(&hr_qp->free);
70 }
71 
72 void init_flush_work(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
73 {
74 	struct hns_roce_work *flush_work = &hr_qp->flush_work;
75 
76 	flush_work->hr_dev = hr_dev;
77 	INIT_WORK(&flush_work->work, flush_work_handle);
78 	refcount_inc(&hr_qp->refcount);
79 	queue_work(hr_dev->irq_workq, &flush_work->work);
80 }
81 
82 void flush_cqe(struct hns_roce_dev *dev, struct hns_roce_qp *qp)
83 {
84 	/*
85 	 * Hip08 hardware cannot flush the WQEs in SQ/RQ if the QP state
86 	 * gets into errored mode. Hence, as a workaround to this
87 	 * hardware limitation, driver needs to assist in flushing. But
88 	 * the flushing operation uses mailbox to convey the QP state to
89 	 * the hardware and which can sleep due to the mutex protection
90 	 * around the mailbox calls. Hence, use the deferred flush for
91 	 * now.
92 	 */
93 	if (!test_and_set_bit(HNS_ROCE_FLUSH_FLAG, &qp->flush_flag))
94 		init_flush_work(dev, qp);
95 }
96 
97 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
98 {
99 	struct device *dev = hr_dev->dev;
100 	struct hns_roce_qp *qp;
101 
102 	xa_lock(&hr_dev->qp_table_xa);
103 	qp = __hns_roce_qp_lookup(hr_dev, qpn);
104 	if (qp)
105 		refcount_inc(&qp->refcount);
106 	xa_unlock(&hr_dev->qp_table_xa);
107 
108 	if (!qp) {
109 		dev_warn(dev, "Async event for bogus QP %08x\n", qpn);
110 		return;
111 	}
112 
113 	if (hr_dev->hw_rev != HNS_ROCE_HW_VER1 &&
114 	    (event_type == HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR ||
115 	     event_type == HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR ||
116 	     event_type == HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR ||
117 	     event_type == HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION ||
118 	     event_type == HNS_ROCE_EVENT_TYPE_INVALID_XRCETH)) {
119 		qp->state = IB_QPS_ERR;
120 
121 		flush_cqe(hr_dev, qp);
122 	}
123 
124 	qp->event(qp, (enum hns_roce_event)event_type);
125 
126 	if (refcount_dec_and_test(&qp->refcount))
127 		complete(&qp->free);
128 }
129 
130 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
131 				 enum hns_roce_event type)
132 {
133 	struct ib_qp *ibqp = &hr_qp->ibqp;
134 	struct ib_event event;
135 
136 	if (ibqp->event_handler) {
137 		event.device = ibqp->device;
138 		event.element.qp = ibqp;
139 		switch (type) {
140 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
141 			event.event = IB_EVENT_PATH_MIG;
142 			break;
143 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
144 			event.event = IB_EVENT_COMM_EST;
145 			break;
146 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
147 			event.event = IB_EVENT_SQ_DRAINED;
148 			break;
149 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
150 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
151 			break;
152 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
153 			event.event = IB_EVENT_QP_FATAL;
154 			break;
155 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
156 			event.event = IB_EVENT_PATH_MIG_ERR;
157 			break;
158 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
159 			event.event = IB_EVENT_QP_REQ_ERR;
160 			break;
161 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
162 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
163 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
164 			event.event = IB_EVENT_QP_ACCESS_ERR;
165 			break;
166 		default:
167 			dev_dbg(ibqp->device->dev.parent, "roce_ib: Unexpected event type %d on QP %06lx\n",
168 				type, hr_qp->qpn);
169 			return;
170 		}
171 		ibqp->event_handler(&event, ibqp->qp_context);
172 	}
173 }
174 
175 static u8 get_least_load_bankid_for_qp(struct hns_roce_bank *bank)
176 {
177 	u32 least_load = bank[0].inuse;
178 	u8 bankid = 0;
179 	u32 bankcnt;
180 	u8 i;
181 
182 	for (i = 1; i < HNS_ROCE_QP_BANK_NUM; i++) {
183 		bankcnt = bank[i].inuse;
184 		if (bankcnt < least_load) {
185 			least_load = bankcnt;
186 			bankid = i;
187 		}
188 	}
189 
190 	return bankid;
191 }
192 
193 static int alloc_qpn_with_bankid(struct hns_roce_bank *bank, u8 bankid,
194 				 unsigned long *qpn)
195 {
196 	int id;
197 
198 	id = ida_alloc_range(&bank->ida, bank->next, bank->max, GFP_KERNEL);
199 	if (id < 0) {
200 		id = ida_alloc_range(&bank->ida, bank->min, bank->max,
201 				     GFP_KERNEL);
202 		if (id < 0)
203 			return id;
204 	}
205 
206 	/* the QPN should keep increasing until the max value is reached. */
207 	bank->next = (id + 1) > bank->max ? bank->min : id + 1;
208 
209 	/* the lower 3 bits is bankid */
210 	*qpn = (id << 3) | bankid;
211 
212 	return 0;
213 }
214 static int alloc_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
215 {
216 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
217 	unsigned long num = 0;
218 	u8 bankid;
219 	int ret;
220 
221 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI) {
222 		/* when hw version is v1, the sqpn is allocated */
223 		if (hr_dev->hw_rev == HNS_ROCE_HW_VER1)
224 			num = HNS_ROCE_MAX_PORTS +
225 			      hr_dev->iboe.phy_port[hr_qp->port];
226 		else
227 			num = 1;
228 
229 		hr_qp->doorbell_qpn = 1;
230 	} else {
231 		mutex_lock(&qp_table->bank_mutex);
232 		bankid = get_least_load_bankid_for_qp(qp_table->bank);
233 
234 		ret = alloc_qpn_with_bankid(&qp_table->bank[bankid], bankid,
235 					    &num);
236 		if (ret) {
237 			ibdev_err(&hr_dev->ib_dev,
238 				  "failed to alloc QPN, ret = %d\n", ret);
239 			mutex_unlock(&qp_table->bank_mutex);
240 			return ret;
241 		}
242 
243 		qp_table->bank[bankid].inuse++;
244 		mutex_unlock(&qp_table->bank_mutex);
245 
246 		hr_qp->doorbell_qpn = (u32)num;
247 	}
248 
249 	hr_qp->qpn = num;
250 
251 	return 0;
252 }
253 
254 enum hns_roce_qp_state to_hns_roce_state(enum ib_qp_state state)
255 {
256 	switch (state) {
257 	case IB_QPS_RESET:
258 		return HNS_ROCE_QP_STATE_RST;
259 	case IB_QPS_INIT:
260 		return HNS_ROCE_QP_STATE_INIT;
261 	case IB_QPS_RTR:
262 		return HNS_ROCE_QP_STATE_RTR;
263 	case IB_QPS_RTS:
264 		return HNS_ROCE_QP_STATE_RTS;
265 	case IB_QPS_SQD:
266 		return HNS_ROCE_QP_STATE_SQD;
267 	case IB_QPS_ERR:
268 		return HNS_ROCE_QP_STATE_ERR;
269 	default:
270 		return HNS_ROCE_QP_NUM_STATE;
271 	}
272 }
273 
274 static void add_qp_to_list(struct hns_roce_dev *hr_dev,
275 			   struct hns_roce_qp *hr_qp,
276 			   struct ib_cq *send_cq, struct ib_cq *recv_cq)
277 {
278 	struct hns_roce_cq *hr_send_cq, *hr_recv_cq;
279 	unsigned long flags;
280 
281 	hr_send_cq = send_cq ? to_hr_cq(send_cq) : NULL;
282 	hr_recv_cq = recv_cq ? to_hr_cq(recv_cq) : NULL;
283 
284 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
285 	hns_roce_lock_cqs(hr_send_cq, hr_recv_cq);
286 
287 	list_add_tail(&hr_qp->node, &hr_dev->qp_list);
288 	if (hr_send_cq)
289 		list_add_tail(&hr_qp->sq_node, &hr_send_cq->sq_list);
290 	if (hr_recv_cq)
291 		list_add_tail(&hr_qp->rq_node, &hr_recv_cq->rq_list);
292 
293 	hns_roce_unlock_cqs(hr_send_cq, hr_recv_cq);
294 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
295 }
296 
297 static int hns_roce_qp_store(struct hns_roce_dev *hr_dev,
298 			     struct hns_roce_qp *hr_qp,
299 			     struct ib_qp_init_attr *init_attr)
300 {
301 	struct xarray *xa = &hr_dev->qp_table_xa;
302 	int ret;
303 
304 	if (!hr_qp->qpn)
305 		return -EINVAL;
306 
307 	ret = xa_err(xa_store_irq(xa, hr_qp->qpn, hr_qp, GFP_KERNEL));
308 	if (ret)
309 		dev_err(hr_dev->dev, "Failed to xa store for QPC\n");
310 	else
311 		/* add QP to device's QP list for softwc */
312 		add_qp_to_list(hr_dev, hr_qp, init_attr->send_cq,
313 			       init_attr->recv_cq);
314 
315 	return ret;
316 }
317 
318 static int alloc_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
319 {
320 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
321 	struct device *dev = hr_dev->dev;
322 	int ret;
323 
324 	if (!hr_qp->qpn)
325 		return -EINVAL;
326 
327 	/* In v1 engine, GSI QP context is saved in the RoCE hw's register */
328 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI &&
329 	    hr_dev->hw_rev == HNS_ROCE_HW_VER1)
330 		return 0;
331 
332 	/* Alloc memory for QPC */
333 	ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
334 	if (ret) {
335 		dev_err(dev, "Failed to get QPC table\n");
336 		goto err_out;
337 	}
338 
339 	/* Alloc memory for IRRL */
340 	ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
341 	if (ret) {
342 		dev_err(dev, "Failed to get IRRL table\n");
343 		goto err_put_qp;
344 	}
345 
346 	if (hr_dev->caps.trrl_entry_sz) {
347 		/* Alloc memory for TRRL */
348 		ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
349 					 hr_qp->qpn);
350 		if (ret) {
351 			dev_err(dev, "Failed to get TRRL table\n");
352 			goto err_put_irrl;
353 		}
354 	}
355 
356 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
357 		/* Alloc memory for SCC CTX */
358 		ret = hns_roce_table_get(hr_dev, &qp_table->sccc_table,
359 					 hr_qp->qpn);
360 		if (ret) {
361 			dev_err(dev, "Failed to get SCC CTX table\n");
362 			goto err_put_trrl;
363 		}
364 	}
365 
366 	return 0;
367 
368 err_put_trrl:
369 	if (hr_dev->caps.trrl_entry_sz)
370 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
371 
372 err_put_irrl:
373 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
374 
375 err_put_qp:
376 	hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
377 
378 err_out:
379 	return ret;
380 }
381 
382 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
383 {
384 	struct xarray *xa = &hr_dev->qp_table_xa;
385 	unsigned long flags;
386 
387 	list_del(&hr_qp->node);
388 
389 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
390 		list_del(&hr_qp->sq_node);
391 
392 	if (hr_qp->ibqp.qp_type != IB_QPT_XRC_INI &&
393 	    hr_qp->ibqp.qp_type != IB_QPT_XRC_TGT)
394 		list_del(&hr_qp->rq_node);
395 
396 	xa_lock_irqsave(xa, flags);
397 	__xa_erase(xa, hr_qp->qpn);
398 	xa_unlock_irqrestore(xa, flags);
399 }
400 
401 static void free_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
402 {
403 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
404 
405 	/* In v1 engine, GSI QP context is saved in the RoCE hw's register */
406 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI &&
407 	    hr_dev->hw_rev == HNS_ROCE_HW_VER1)
408 		return;
409 
410 	if (hr_dev->caps.trrl_entry_sz)
411 		hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
412 	hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
413 }
414 
415 static inline u8 get_qp_bankid(unsigned long qpn)
416 {
417 	/* The lower 3 bits of QPN are used to hash to different banks */
418 	return (u8)(qpn & GENMASK(2, 0));
419 }
420 
421 static void free_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
422 {
423 	u8 bankid;
424 
425 	if (hr_qp->ibqp.qp_type == IB_QPT_GSI)
426 		return;
427 
428 	if (hr_qp->qpn < hr_dev->caps.reserved_qps)
429 		return;
430 
431 	bankid = get_qp_bankid(hr_qp->qpn);
432 
433 	ida_free(&hr_dev->qp_table.bank[bankid].ida, hr_qp->qpn >> 3);
434 
435 	mutex_lock(&hr_dev->qp_table.bank_mutex);
436 	hr_dev->qp_table.bank[bankid].inuse--;
437 	mutex_unlock(&hr_dev->qp_table.bank_mutex);
438 }
439 
440 static u32 proc_rq_sge(struct hns_roce_dev *dev, struct hns_roce_qp *hr_qp,
441 		       bool user)
442 {
443 	u32 max_sge = dev->caps.max_rq_sg;
444 
445 	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
446 		return max_sge;
447 
448 	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
449 	 * calculate number of max_sge with reserved SGEs when allocating wqe
450 	 * buf, so there is no need to do this again in kernel. But the number
451 	 * may exceed the capacity of SGEs recorded in the firmware, so the
452 	 * kernel driver should just adapt the value accordingly.
453 	 */
454 	if (user)
455 		max_sge = roundup_pow_of_two(max_sge + 1);
456 	else
457 		hr_qp->rq.rsv_sge = 1;
458 
459 	return max_sge;
460 }
461 
462 static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
463 		       struct hns_roce_qp *hr_qp, int has_rq, bool user)
464 {
465 	u32 max_sge = proc_rq_sge(hr_dev, hr_qp, user);
466 	u32 cnt;
467 
468 	/* If srq exist, set zero for relative number of rq */
469 	if (!has_rq) {
470 		hr_qp->rq.wqe_cnt = 0;
471 		hr_qp->rq.max_gs = 0;
472 		hr_qp->rq_inl_buf.wqe_cnt = 0;
473 		cap->max_recv_wr = 0;
474 		cap->max_recv_sge = 0;
475 
476 		return 0;
477 	}
478 
479 	/* Check the validity of QP support capacity */
480 	if (!cap->max_recv_wr || cap->max_recv_wr > hr_dev->caps.max_wqes ||
481 	    cap->max_recv_sge > max_sge) {
482 		ibdev_err(&hr_dev->ib_dev,
483 			  "RQ config error, depth = %u, sge = %u\n",
484 			  cap->max_recv_wr, cap->max_recv_sge);
485 		return -EINVAL;
486 	}
487 
488 	cnt = roundup_pow_of_two(max(cap->max_recv_wr, hr_dev->caps.min_wqes));
489 	if (cnt > hr_dev->caps.max_wqes) {
490 		ibdev_err(&hr_dev->ib_dev, "rq depth %u too large\n",
491 			  cap->max_recv_wr);
492 		return -EINVAL;
493 	}
494 
495 	hr_qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge) +
496 					      hr_qp->rq.rsv_sge);
497 
498 	if (hr_dev->caps.max_rq_sg <= HNS_ROCE_SGE_IN_WQE)
499 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz);
500 	else
501 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz *
502 					    hr_qp->rq.max_gs);
503 
504 	hr_qp->rq.wqe_cnt = cnt;
505 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE &&
506 	    hr_qp->ibqp.qp_type != IB_QPT_UD &&
507 	    hr_qp->ibqp.qp_type != IB_QPT_GSI)
508 		hr_qp->rq_inl_buf.wqe_cnt = cnt;
509 	else
510 		hr_qp->rq_inl_buf.wqe_cnt = 0;
511 
512 	cap->max_recv_wr = cnt;
513 	cap->max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
514 
515 	return 0;
516 }
517 
518 static u32 get_wqe_ext_sge_cnt(struct hns_roce_qp *qp)
519 {
520 	/* GSI/UD QP only has extended sge */
521 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_UD)
522 		return qp->sq.max_gs;
523 
524 	if (qp->sq.max_gs > HNS_ROCE_SGE_IN_WQE)
525 		return qp->sq.max_gs - HNS_ROCE_SGE_IN_WQE;
526 
527 	return 0;
528 }
529 
530 static void set_ext_sge_param(struct hns_roce_dev *hr_dev, u32 sq_wqe_cnt,
531 			      struct hns_roce_qp *hr_qp, struct ib_qp_cap *cap)
532 {
533 	u32 total_sge_cnt;
534 	u32 wqe_sge_cnt;
535 
536 	hr_qp->sge.sge_shift = HNS_ROCE_SGE_SHIFT;
537 
538 	if (hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
539 		hr_qp->sq.max_gs = HNS_ROCE_SGE_IN_WQE;
540 		return;
541 	}
542 
543 	hr_qp->sq.max_gs = max(1U, cap->max_send_sge);
544 
545 	wqe_sge_cnt = get_wqe_ext_sge_cnt(hr_qp);
546 
547 	/* If the number of extended sge is not zero, they MUST use the
548 	 * space of HNS_HW_PAGE_SIZE at least.
549 	 */
550 	if (wqe_sge_cnt) {
551 		total_sge_cnt = roundup_pow_of_two(sq_wqe_cnt * wqe_sge_cnt);
552 		hr_qp->sge.sge_cnt = max(total_sge_cnt,
553 				(u32)HNS_HW_PAGE_SIZE / HNS_ROCE_SGE_SIZE);
554 	}
555 }
556 
557 static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
558 					struct ib_qp_cap *cap,
559 					struct hns_roce_ib_create_qp *ucmd)
560 {
561 	u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
562 	u8 max_sq_stride = ilog2(roundup_sq_stride);
563 
564 	/* Sanity check SQ size before proceeding */
565 	if (ucmd->log_sq_stride > max_sq_stride ||
566 	    ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
567 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ stride size.\n");
568 		return -EINVAL;
569 	}
570 
571 	if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
572 		ibdev_err(&hr_dev->ib_dev, "failed to check SQ SGE size %u.\n",
573 			  cap->max_send_sge);
574 		return -EINVAL;
575 	}
576 
577 	return 0;
578 }
579 
580 static int set_user_sq_size(struct hns_roce_dev *hr_dev,
581 			    struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp,
582 			    struct hns_roce_ib_create_qp *ucmd)
583 {
584 	struct ib_device *ibdev = &hr_dev->ib_dev;
585 	u32 cnt = 0;
586 	int ret;
587 
588 	if (check_shl_overflow(1, ucmd->log_sq_bb_count, &cnt) ||
589 	    cnt > hr_dev->caps.max_wqes)
590 		return -EINVAL;
591 
592 	ret = check_sq_size_with_integrity(hr_dev, cap, ucmd);
593 	if (ret) {
594 		ibdev_err(ibdev, "failed to check user SQ size, ret = %d.\n",
595 			  ret);
596 		return ret;
597 	}
598 
599 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
600 
601 	hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
602 	hr_qp->sq.wqe_cnt = cnt;
603 
604 	return 0;
605 }
606 
607 static int set_wqe_buf_attr(struct hns_roce_dev *hr_dev,
608 			    struct hns_roce_qp *hr_qp,
609 			    struct hns_roce_buf_attr *buf_attr)
610 {
611 	int buf_size;
612 	int idx = 0;
613 
614 	hr_qp->buff_size = 0;
615 
616 	/* SQ WQE */
617 	hr_qp->sq.offset = 0;
618 	buf_size = to_hr_hem_entries_size(hr_qp->sq.wqe_cnt,
619 					  hr_qp->sq.wqe_shift);
620 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
621 		buf_attr->region[idx].size = buf_size;
622 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sq_hop_num;
623 		idx++;
624 		hr_qp->buff_size += buf_size;
625 	}
626 
627 	/* extend SGE WQE in SQ */
628 	hr_qp->sge.offset = hr_qp->buff_size;
629 	buf_size = to_hr_hem_entries_size(hr_qp->sge.sge_cnt,
630 					  hr_qp->sge.sge_shift);
631 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
632 		buf_attr->region[idx].size = buf_size;
633 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_sge_hop_num;
634 		idx++;
635 		hr_qp->buff_size += buf_size;
636 	}
637 
638 	/* RQ WQE */
639 	hr_qp->rq.offset = hr_qp->buff_size;
640 	buf_size = to_hr_hem_entries_size(hr_qp->rq.wqe_cnt,
641 					  hr_qp->rq.wqe_shift);
642 	if (buf_size > 0 && idx < ARRAY_SIZE(buf_attr->region)) {
643 		buf_attr->region[idx].size = buf_size;
644 		buf_attr->region[idx].hopnum = hr_dev->caps.wqe_rq_hop_num;
645 		idx++;
646 		hr_qp->buff_size += buf_size;
647 	}
648 
649 	if (hr_qp->buff_size < 1)
650 		return -EINVAL;
651 
652 	buf_attr->page_shift = HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
653 	buf_attr->region_count = idx;
654 
655 	return 0;
656 }
657 
658 static int set_kernel_sq_size(struct hns_roce_dev *hr_dev,
659 			      struct ib_qp_cap *cap, struct hns_roce_qp *hr_qp)
660 {
661 	struct ib_device *ibdev = &hr_dev->ib_dev;
662 	u32 cnt;
663 
664 	if (!cap->max_send_wr || cap->max_send_wr > hr_dev->caps.max_wqes ||
665 	    cap->max_send_sge > hr_dev->caps.max_sq_sg) {
666 		ibdev_err(ibdev, "failed to check SQ WR or SGE num.\n");
667 		return -EINVAL;
668 	}
669 
670 	cnt = roundup_pow_of_two(max(cap->max_send_wr, hr_dev->caps.min_wqes));
671 	if (cnt > hr_dev->caps.max_wqes) {
672 		ibdev_err(ibdev, "failed to check WQE num, WQE num = %u.\n",
673 			  cnt);
674 		return -EINVAL;
675 	}
676 
677 	hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
678 	hr_qp->sq.wqe_cnt = cnt;
679 
680 	set_ext_sge_param(hr_dev, cnt, hr_qp, cap);
681 
682 	/* sync the parameters of kernel QP to user's configuration */
683 	cap->max_send_wr = cnt;
684 	cap->max_send_sge = hr_qp->sq.max_gs;
685 
686 	return 0;
687 }
688 
689 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
690 {
691 	if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
692 		return 0;
693 
694 	return 1;
695 }
696 
697 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
698 {
699 	if (attr->qp_type == IB_QPT_XRC_INI ||
700 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
701 	    !attr->cap.max_recv_wr)
702 		return 0;
703 
704 	return 1;
705 }
706 
707 static int alloc_rq_inline_buf(struct hns_roce_qp *hr_qp,
708 			       struct ib_qp_init_attr *init_attr)
709 {
710 	u32 max_recv_sge = init_attr->cap.max_recv_sge;
711 	u32 wqe_cnt = hr_qp->rq_inl_buf.wqe_cnt;
712 	struct hns_roce_rinl_wqe *wqe_list;
713 	int i;
714 
715 	/* allocate recv inline buf */
716 	wqe_list = kcalloc(wqe_cnt, sizeof(struct hns_roce_rinl_wqe),
717 			   GFP_KERNEL);
718 
719 	if (!wqe_list)
720 		goto err;
721 
722 	/* Allocate a continuous buffer for all inline sge we need */
723 	wqe_list[0].sg_list = kcalloc(wqe_cnt, (max_recv_sge *
724 				      sizeof(struct hns_roce_rinl_sge)),
725 				      GFP_KERNEL);
726 	if (!wqe_list[0].sg_list)
727 		goto err_wqe_list;
728 
729 	/* Assign buffers of sg_list to each inline wqe */
730 	for (i = 1; i < wqe_cnt; i++)
731 		wqe_list[i].sg_list = &wqe_list[0].sg_list[i * max_recv_sge];
732 
733 	hr_qp->rq_inl_buf.wqe_list = wqe_list;
734 
735 	return 0;
736 
737 err_wqe_list:
738 	kfree(wqe_list);
739 
740 err:
741 	return -ENOMEM;
742 }
743 
744 static void free_rq_inline_buf(struct hns_roce_qp *hr_qp)
745 {
746 	if (hr_qp->rq_inl_buf.wqe_list)
747 		kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
748 	kfree(hr_qp->rq_inl_buf.wqe_list);
749 }
750 
751 static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
752 			struct ib_qp_init_attr *init_attr,
753 			struct ib_udata *udata, unsigned long addr)
754 {
755 	struct ib_device *ibdev = &hr_dev->ib_dev;
756 	struct hns_roce_buf_attr buf_attr = {};
757 	int ret;
758 
759 	if (!udata && hr_qp->rq_inl_buf.wqe_cnt) {
760 		ret = alloc_rq_inline_buf(hr_qp, init_attr);
761 		if (ret) {
762 			ibdev_err(ibdev,
763 				  "failed to alloc inline buf, ret = %d.\n",
764 				  ret);
765 			return ret;
766 		}
767 	} else {
768 		hr_qp->rq_inl_buf.wqe_list = NULL;
769 	}
770 
771 	ret = set_wqe_buf_attr(hr_dev, hr_qp, &buf_attr);
772 	if (ret) {
773 		ibdev_err(ibdev, "failed to split WQE buf, ret = %d.\n", ret);
774 		goto err_inline;
775 	}
776 	ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, &buf_attr,
777 				  PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
778 				  udata, addr);
779 	if (ret) {
780 		ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret);
781 		goto err_inline;
782 	}
783 
784 	return 0;
785 err_inline:
786 	free_rq_inline_buf(hr_qp);
787 
788 	return ret;
789 }
790 
791 static void free_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
792 {
793 	hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
794 	free_rq_inline_buf(hr_qp);
795 }
796 
797 static inline bool user_qp_has_sdb(struct hns_roce_dev *hr_dev,
798 				   struct ib_qp_init_attr *init_attr,
799 				   struct ib_udata *udata,
800 				   struct hns_roce_ib_create_qp_resp *resp,
801 				   struct hns_roce_ib_create_qp *ucmd)
802 {
803 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
804 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
805 		hns_roce_qp_has_sq(init_attr) &&
806 		udata->inlen >= offsetofend(typeof(*ucmd), sdb_addr));
807 }
808 
809 static inline bool user_qp_has_rdb(struct hns_roce_dev *hr_dev,
810 				   struct ib_qp_init_attr *init_attr,
811 				   struct ib_udata *udata,
812 				   struct hns_roce_ib_create_qp_resp *resp)
813 {
814 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
815 		udata->outlen >= offsetofend(typeof(*resp), cap_flags) &&
816 		hns_roce_qp_has_rq(init_attr));
817 }
818 
819 static inline bool kernel_qp_has_rdb(struct hns_roce_dev *hr_dev,
820 				     struct ib_qp_init_attr *init_attr)
821 {
822 	return ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) &&
823 		hns_roce_qp_has_rq(init_attr));
824 }
825 
826 static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
827 		       struct ib_qp_init_attr *init_attr,
828 		       struct ib_udata *udata,
829 		       struct hns_roce_ib_create_qp *ucmd,
830 		       struct hns_roce_ib_create_qp_resp *resp)
831 {
832 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
833 		udata, struct hns_roce_ucontext, ibucontext);
834 	struct ib_device *ibdev = &hr_dev->ib_dev;
835 	int ret;
836 
837 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SDI_MODE)
838 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_OWNER_DB;
839 
840 	if (udata) {
841 		if (user_qp_has_sdb(hr_dev, init_attr, udata, resp, ucmd)) {
842 			ret = hns_roce_db_map_user(uctx, ucmd->sdb_addr,
843 						   &hr_qp->sdb);
844 			if (ret) {
845 				ibdev_err(ibdev,
846 					  "failed to map user SQ doorbell, ret = %d.\n",
847 					  ret);
848 				goto err_out;
849 			}
850 			hr_qp->en_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
851 			resp->cap_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
852 		}
853 
854 		if (user_qp_has_rdb(hr_dev, init_attr, udata, resp)) {
855 			ret = hns_roce_db_map_user(uctx, ucmd->db_addr,
856 						   &hr_qp->rdb);
857 			if (ret) {
858 				ibdev_err(ibdev,
859 					  "failed to map user RQ doorbell, ret = %d.\n",
860 					  ret);
861 				goto err_sdb;
862 			}
863 			hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
864 			resp->cap_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
865 		}
866 	} else {
867 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
868 			hr_qp->sq.db_reg = hr_dev->mem_base +
869 					   HNS_ROCE_DWQE_SIZE * hr_qp->qpn;
870 		else
871 			hr_qp->sq.db_reg =
872 				hr_dev->reg_base + hr_dev->sdb_offset +
873 				DB_REG_OFFSET * hr_dev->priv_uar.index;
874 
875 		hr_qp->rq.db_reg = hr_dev->reg_base + hr_dev->odb_offset +
876 				   DB_REG_OFFSET * hr_dev->priv_uar.index;
877 
878 		if (kernel_qp_has_rdb(hr_dev, init_attr)) {
879 			ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
880 			if (ret) {
881 				ibdev_err(ibdev,
882 					  "failed to alloc kernel RQ doorbell, ret = %d.\n",
883 					  ret);
884 				goto err_out;
885 			}
886 			*hr_qp->rdb.db_record = 0;
887 			hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
888 		}
889 	}
890 
891 	return 0;
892 err_sdb:
893 	if (udata && hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
894 		hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
895 err_out:
896 	return ret;
897 }
898 
899 static void free_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
900 		       struct ib_udata *udata)
901 {
902 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
903 		udata, struct hns_roce_ucontext, ibucontext);
904 
905 	if (udata) {
906 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
907 			hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
908 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
909 			hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
910 	} else {
911 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
912 			hns_roce_free_db(hr_dev, &hr_qp->rdb);
913 	}
914 }
915 
916 static int alloc_kernel_wrid(struct hns_roce_dev *hr_dev,
917 			     struct hns_roce_qp *hr_qp)
918 {
919 	struct ib_device *ibdev = &hr_dev->ib_dev;
920 	u64 *sq_wrid = NULL;
921 	u64 *rq_wrid = NULL;
922 	int ret;
923 
924 	sq_wrid = kcalloc(hr_qp->sq.wqe_cnt, sizeof(u64), GFP_KERNEL);
925 	if (ZERO_OR_NULL_PTR(sq_wrid)) {
926 		ibdev_err(ibdev, "failed to alloc SQ wrid.\n");
927 		return -ENOMEM;
928 	}
929 
930 	if (hr_qp->rq.wqe_cnt) {
931 		rq_wrid = kcalloc(hr_qp->rq.wqe_cnt, sizeof(u64), GFP_KERNEL);
932 		if (ZERO_OR_NULL_PTR(rq_wrid)) {
933 			ibdev_err(ibdev, "failed to alloc RQ wrid.\n");
934 			ret = -ENOMEM;
935 			goto err_sq;
936 		}
937 	}
938 
939 	hr_qp->sq.wrid = sq_wrid;
940 	hr_qp->rq.wrid = rq_wrid;
941 	return 0;
942 err_sq:
943 	kfree(sq_wrid);
944 
945 	return ret;
946 }
947 
948 static void free_kernel_wrid(struct hns_roce_qp *hr_qp)
949 {
950 	kfree(hr_qp->rq.wrid);
951 	kfree(hr_qp->sq.wrid);
952 }
953 
954 static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
955 			struct ib_qp_init_attr *init_attr,
956 			struct ib_udata *udata,
957 			struct hns_roce_ib_create_qp *ucmd)
958 {
959 	struct ib_device *ibdev = &hr_dev->ib_dev;
960 	int ret;
961 
962 	hr_qp->ibqp.qp_type = init_attr->qp_type;
963 
964 	if (init_attr->cap.max_inline_data > hr_dev->caps.max_sq_inline)
965 		init_attr->cap.max_inline_data = hr_dev->caps.max_sq_inline;
966 
967 	hr_qp->max_inline_data = init_attr->cap.max_inline_data;
968 
969 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
970 		hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
971 	else
972 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
973 
974 	ret = set_rq_size(hr_dev, &init_attr->cap, hr_qp,
975 			  hns_roce_qp_has_rq(init_attr), !!udata);
976 	if (ret) {
977 		ibdev_err(ibdev, "failed to set user RQ size, ret = %d.\n",
978 			  ret);
979 		return ret;
980 	}
981 
982 	if (udata) {
983 		ret = ib_copy_from_udata(ucmd, udata,
984 					 min(udata->inlen, sizeof(*ucmd)));
985 		if (ret) {
986 			ibdev_err(ibdev,
987 				  "failed to copy QP ucmd, ret = %d\n", ret);
988 			return ret;
989 		}
990 
991 		ret = set_user_sq_size(hr_dev, &init_attr->cap, hr_qp, ucmd);
992 		if (ret)
993 			ibdev_err(ibdev,
994 				  "failed to set user SQ size, ret = %d.\n",
995 				  ret);
996 	} else {
997 		ret = set_kernel_sq_size(hr_dev, &init_attr->cap, hr_qp);
998 		if (ret)
999 			ibdev_err(ibdev,
1000 				  "failed to set kernel SQ size, ret = %d.\n",
1001 				  ret);
1002 	}
1003 
1004 	return ret;
1005 }
1006 
1007 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
1008 				     struct ib_pd *ib_pd,
1009 				     struct ib_qp_init_attr *init_attr,
1010 				     struct ib_udata *udata,
1011 				     struct hns_roce_qp *hr_qp)
1012 {
1013 	struct hns_roce_ib_create_qp_resp resp = {};
1014 	struct ib_device *ibdev = &hr_dev->ib_dev;
1015 	struct hns_roce_ib_create_qp ucmd;
1016 	int ret;
1017 
1018 	mutex_init(&hr_qp->mutex);
1019 	spin_lock_init(&hr_qp->sq.lock);
1020 	spin_lock_init(&hr_qp->rq.lock);
1021 
1022 	hr_qp->state = IB_QPS_RESET;
1023 	hr_qp->flush_flag = 0;
1024 
1025 	if (init_attr->create_flags)
1026 		return -EOPNOTSUPP;
1027 
1028 	ret = set_qp_param(hr_dev, hr_qp, init_attr, udata, &ucmd);
1029 	if (ret) {
1030 		ibdev_err(ibdev, "failed to set QP param, ret = %d.\n", ret);
1031 		return ret;
1032 	}
1033 
1034 	if (!udata) {
1035 		ret = alloc_kernel_wrid(hr_dev, hr_qp);
1036 		if (ret) {
1037 			ibdev_err(ibdev, "failed to alloc wrid, ret = %d.\n",
1038 				  ret);
1039 			return ret;
1040 		}
1041 	}
1042 
1043 	ret = alloc_qp_buf(hr_dev, hr_qp, init_attr, udata, ucmd.buf_addr);
1044 	if (ret) {
1045 		ibdev_err(ibdev, "failed to alloc QP buffer, ret = %d.\n", ret);
1046 		goto err_buf;
1047 	}
1048 
1049 	ret = alloc_qpn(hr_dev, hr_qp);
1050 	if (ret) {
1051 		ibdev_err(ibdev, "failed to alloc QPN, ret = %d.\n", ret);
1052 		goto err_qpn;
1053 	}
1054 
1055 	ret = alloc_qp_db(hr_dev, hr_qp, init_attr, udata, &ucmd, &resp);
1056 	if (ret) {
1057 		ibdev_err(ibdev, "failed to alloc QP doorbell, ret = %d.\n",
1058 			  ret);
1059 		goto err_db;
1060 	}
1061 
1062 	ret = alloc_qpc(hr_dev, hr_qp);
1063 	if (ret) {
1064 		ibdev_err(ibdev, "failed to alloc QP context, ret = %d.\n",
1065 			  ret);
1066 		goto err_qpc;
1067 	}
1068 
1069 	ret = hns_roce_qp_store(hr_dev, hr_qp, init_attr);
1070 	if (ret) {
1071 		ibdev_err(ibdev, "failed to store QP, ret = %d.\n", ret);
1072 		goto err_store;
1073 	}
1074 
1075 	if (udata) {
1076 		ret = ib_copy_to_udata(udata, &resp,
1077 				       min(udata->outlen, sizeof(resp)));
1078 		if (ret) {
1079 			ibdev_err(ibdev, "copy qp resp failed!\n");
1080 			goto err_store;
1081 		}
1082 	}
1083 
1084 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
1085 		ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
1086 		if (ret)
1087 			goto err_flow_ctrl;
1088 	}
1089 
1090 	hr_qp->ibqp.qp_num = hr_qp->qpn;
1091 	hr_qp->event = hns_roce_ib_qp_event;
1092 	refcount_set(&hr_qp->refcount, 1);
1093 	init_completion(&hr_qp->free);
1094 
1095 	return 0;
1096 
1097 err_flow_ctrl:
1098 	hns_roce_qp_remove(hr_dev, hr_qp);
1099 err_store:
1100 	free_qpc(hr_dev, hr_qp);
1101 err_qpc:
1102 	free_qp_db(hr_dev, hr_qp, udata);
1103 err_db:
1104 	free_qpn(hr_dev, hr_qp);
1105 err_qpn:
1106 	free_qp_buf(hr_dev, hr_qp);
1107 err_buf:
1108 	free_kernel_wrid(hr_qp);
1109 	return ret;
1110 }
1111 
1112 void hns_roce_qp_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
1113 			 struct ib_udata *udata)
1114 {
1115 	if (refcount_dec_and_test(&hr_qp->refcount))
1116 		complete(&hr_qp->free);
1117 	wait_for_completion(&hr_qp->free);
1118 
1119 	free_qpc(hr_dev, hr_qp);
1120 	free_qpn(hr_dev, hr_qp);
1121 	free_qp_buf(hr_dev, hr_qp);
1122 	free_kernel_wrid(hr_qp);
1123 	free_qp_db(hr_dev, hr_qp, udata);
1124 
1125 	kfree(hr_qp);
1126 }
1127 
1128 static int check_qp_type(struct hns_roce_dev *hr_dev, enum ib_qp_type type,
1129 			 bool is_user)
1130 {
1131 	switch (type) {
1132 	case IB_QPT_XRC_INI:
1133 	case IB_QPT_XRC_TGT:
1134 		if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC))
1135 			goto out;
1136 		break;
1137 	case IB_QPT_UD:
1138 		if (hr_dev->pci_dev->revision <= PCI_REVISION_ID_HIP08 &&
1139 		    is_user)
1140 			goto out;
1141 		break;
1142 	case IB_QPT_RC:
1143 	case IB_QPT_GSI:
1144 		break;
1145 	default:
1146 		goto out;
1147 	}
1148 
1149 	return 0;
1150 
1151 out:
1152 	ibdev_err(&hr_dev->ib_dev, "not support QP type %d\n", type);
1153 
1154 	return -EOPNOTSUPP;
1155 }
1156 
1157 struct ib_qp *hns_roce_create_qp(struct ib_pd *pd,
1158 				 struct ib_qp_init_attr *init_attr,
1159 				 struct ib_udata *udata)
1160 {
1161 	struct ib_device *ibdev = pd ? pd->device : init_attr->xrcd->device;
1162 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
1163 	struct hns_roce_qp *hr_qp;
1164 	int ret;
1165 
1166 	ret = check_qp_type(hr_dev, init_attr->qp_type, !!udata);
1167 	if (ret)
1168 		return ERR_PTR(ret);
1169 
1170 	hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
1171 	if (!hr_qp)
1172 		return ERR_PTR(-ENOMEM);
1173 
1174 	if (init_attr->qp_type == IB_QPT_XRC_INI)
1175 		init_attr->recv_cq = NULL;
1176 
1177 	if (init_attr->qp_type == IB_QPT_XRC_TGT) {
1178 		hr_qp->xrcdn = to_hr_xrcd(init_attr->xrcd)->xrcdn;
1179 		init_attr->recv_cq = NULL;
1180 		init_attr->send_cq = NULL;
1181 	}
1182 
1183 	if (init_attr->qp_type == IB_QPT_GSI) {
1184 		hr_qp->port = init_attr->port_num - 1;
1185 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
1186 	}
1187 
1188 	ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, hr_qp);
1189 	if (ret) {
1190 		ibdev_err(ibdev, "Create QP type 0x%x failed(%d)\n",
1191 			  init_attr->qp_type, ret);
1192 
1193 		kfree(hr_qp);
1194 		return ERR_PTR(ret);
1195 	}
1196 
1197 	return &hr_qp->ibqp;
1198 }
1199 
1200 int to_hr_qp_type(int qp_type)
1201 {
1202 	switch (qp_type) {
1203 	case IB_QPT_RC:
1204 		return SERV_TYPE_RC;
1205 	case IB_QPT_UD:
1206 	case IB_QPT_GSI:
1207 		return SERV_TYPE_UD;
1208 	case IB_QPT_XRC_INI:
1209 	case IB_QPT_XRC_TGT:
1210 		return SERV_TYPE_XRC;
1211 	default:
1212 		return -1;
1213 	}
1214 }
1215 
1216 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1217 			      struct hns_roce_qp *hr_qp,
1218 			      struct ib_qp_attr *attr, int attr_mask)
1219 {
1220 	enum ib_mtu active_mtu;
1221 	int p;
1222 
1223 	p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1224 	active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1225 
1226 	if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1227 	    attr->path_mtu > hr_dev->caps.max_mtu) ||
1228 	    attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1229 		ibdev_err(&hr_dev->ib_dev,
1230 			"attr path_mtu(%d)invalid while modify qp",
1231 			attr->path_mtu);
1232 		return -EINVAL;
1233 	}
1234 
1235 	return 0;
1236 }
1237 
1238 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1239 				  int attr_mask)
1240 {
1241 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1242 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1243 	int p;
1244 
1245 	if ((attr_mask & IB_QP_PORT) &&
1246 	    (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1247 		ibdev_err(&hr_dev->ib_dev, "invalid attr, port_num = %u.\n",
1248 			  attr->port_num);
1249 		return -EINVAL;
1250 	}
1251 
1252 	if (attr_mask & IB_QP_PKEY_INDEX) {
1253 		p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1254 		if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1255 			ibdev_err(&hr_dev->ib_dev,
1256 				  "invalid attr, pkey_index = %u.\n",
1257 				  attr->pkey_index);
1258 			return -EINVAL;
1259 		}
1260 	}
1261 
1262 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1263 	    attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1264 		ibdev_err(&hr_dev->ib_dev,
1265 			  "invalid attr, max_rd_atomic = %u.\n",
1266 			  attr->max_rd_atomic);
1267 		return -EINVAL;
1268 	}
1269 
1270 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1271 	    attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1272 		ibdev_err(&hr_dev->ib_dev,
1273 			  "invalid attr, max_dest_rd_atomic = %u.\n",
1274 			  attr->max_dest_rd_atomic);
1275 		return -EINVAL;
1276 	}
1277 
1278 	if (attr_mask & IB_QP_PATH_MTU)
1279 		return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1280 
1281 	return 0;
1282 }
1283 
1284 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1285 		       int attr_mask, struct ib_udata *udata)
1286 {
1287 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1288 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1289 	enum ib_qp_state cur_state, new_state;
1290 	int ret = -EINVAL;
1291 
1292 	mutex_lock(&hr_qp->mutex);
1293 
1294 	if (attr_mask & IB_QP_CUR_STATE && attr->cur_qp_state != hr_qp->state)
1295 		goto out;
1296 
1297 	cur_state = hr_qp->state;
1298 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1299 
1300 	if (ibqp->uobject &&
1301 	    (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1302 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB) {
1303 			hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1304 
1305 			if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
1306 				hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1307 		} else {
1308 			ibdev_warn(&hr_dev->ib_dev,
1309 				  "flush cqe is not supported in userspace!\n");
1310 			goto out;
1311 		}
1312 	}
1313 
1314 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1315 				attr_mask)) {
1316 		ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1317 		goto out;
1318 	}
1319 
1320 	ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1321 	if (ret)
1322 		goto out;
1323 
1324 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1325 		if (hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
1326 			ret = -EPERM;
1327 			ibdev_err(&hr_dev->ib_dev,
1328 				  "RST2RST state is not supported\n");
1329 		} else {
1330 			ret = 0;
1331 		}
1332 
1333 		goto out;
1334 	}
1335 
1336 	ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1337 				    new_state);
1338 
1339 out:
1340 	mutex_unlock(&hr_qp->mutex);
1341 
1342 	return ret;
1343 }
1344 
1345 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1346 		       __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1347 {
1348 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1349 		__acquire(&send_cq->lock);
1350 		__acquire(&recv_cq->lock);
1351 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1352 		spin_lock_irq(&send_cq->lock);
1353 		__acquire(&recv_cq->lock);
1354 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1355 		spin_lock_irq(&recv_cq->lock);
1356 		__acquire(&send_cq->lock);
1357 	} else if (send_cq == recv_cq) {
1358 		spin_lock_irq(&send_cq->lock);
1359 		__acquire(&recv_cq->lock);
1360 	} else if (send_cq->cqn < recv_cq->cqn) {
1361 		spin_lock_irq(&send_cq->lock);
1362 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1363 	} else {
1364 		spin_lock_irq(&recv_cq->lock);
1365 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1366 	}
1367 }
1368 
1369 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1370 			 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1371 			 __releases(&recv_cq->lock)
1372 {
1373 	if (unlikely(send_cq == NULL && recv_cq == NULL)) {
1374 		__release(&recv_cq->lock);
1375 		__release(&send_cq->lock);
1376 	} else if (unlikely(send_cq != NULL && recv_cq == NULL)) {
1377 		__release(&recv_cq->lock);
1378 		spin_unlock(&send_cq->lock);
1379 	} else if (unlikely(send_cq == NULL && recv_cq != NULL)) {
1380 		__release(&send_cq->lock);
1381 		spin_unlock(&recv_cq->lock);
1382 	} else if (send_cq == recv_cq) {
1383 		__release(&recv_cq->lock);
1384 		spin_unlock_irq(&send_cq->lock);
1385 	} else if (send_cq->cqn < recv_cq->cqn) {
1386 		spin_unlock(&recv_cq->lock);
1387 		spin_unlock_irq(&send_cq->lock);
1388 	} else {
1389 		spin_unlock(&send_cq->lock);
1390 		spin_unlock_irq(&recv_cq->lock);
1391 	}
1392 }
1393 
1394 static inline void *get_wqe(struct hns_roce_qp *hr_qp, int offset)
1395 {
1396 	return hns_roce_buf_offset(hr_qp->mtr.kmem, offset);
1397 }
1398 
1399 void *hns_roce_get_recv_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1400 {
1401 	return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1402 }
1403 
1404 void *hns_roce_get_send_wqe(struct hns_roce_qp *hr_qp, unsigned int n)
1405 {
1406 	return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1407 }
1408 
1409 void *hns_roce_get_extend_sge(struct hns_roce_qp *hr_qp, unsigned int n)
1410 {
1411 	return get_wqe(hr_qp, hr_qp->sge.offset + (n << hr_qp->sge.sge_shift));
1412 }
1413 
1414 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, u32 nreq,
1415 			  struct ib_cq *ib_cq)
1416 {
1417 	struct hns_roce_cq *hr_cq;
1418 	u32 cur;
1419 
1420 	cur = hr_wq->head - hr_wq->tail;
1421 	if (likely(cur + nreq < hr_wq->wqe_cnt))
1422 		return false;
1423 
1424 	hr_cq = to_hr_cq(ib_cq);
1425 	spin_lock(&hr_cq->lock);
1426 	cur = hr_wq->head - hr_wq->tail;
1427 	spin_unlock(&hr_cq->lock);
1428 
1429 	return cur + nreq >= hr_wq->wqe_cnt;
1430 }
1431 
1432 void hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1433 {
1434 	struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1435 	unsigned int reserved_from_bot;
1436 	unsigned int i;
1437 
1438 	mutex_init(&qp_table->scc_mutex);
1439 	mutex_init(&qp_table->bank_mutex);
1440 	xa_init(&hr_dev->qp_table_xa);
1441 
1442 	reserved_from_bot = hr_dev->caps.reserved_qps;
1443 
1444 	for (i = 0; i < reserved_from_bot; i++) {
1445 		hr_dev->qp_table.bank[get_qp_bankid(i)].inuse++;
1446 		hr_dev->qp_table.bank[get_qp_bankid(i)].min++;
1447 	}
1448 
1449 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) {
1450 		ida_init(&hr_dev->qp_table.bank[i].ida);
1451 		hr_dev->qp_table.bank[i].max = hr_dev->caps.num_qps /
1452 					       HNS_ROCE_QP_BANK_NUM - 1;
1453 		hr_dev->qp_table.bank[i].next = hr_dev->qp_table.bank[i].min;
1454 	}
1455 }
1456 
1457 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1458 {
1459 	int i;
1460 
1461 	for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++)
1462 		ida_destroy(&hr_dev->qp_table.bank[i].ida);
1463 }
1464