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