xref: /openbmc/linux/drivers/infiniband/hw/mlx5/qp.c (revision d9fd5a71)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/module.h>
34 #include <rdma/ib_umem.h>
35 #include <rdma/ib_cache.h>
36 #include <rdma/ib_user_verbs.h>
37 #include <rdma/rdma_counter.h>
38 #include <linux/mlx5/fs.h>
39 #include "mlx5_ib.h"
40 #include "ib_rep.h"
41 #include "counters.h"
42 #include "cmd.h"
43 #include "qp.h"
44 #include "wr.h"
45 
46 enum {
47 	MLX5_IB_ACK_REQ_FREQ	= 8,
48 };
49 
50 enum {
51 	MLX5_IB_DEFAULT_SCHED_QUEUE	= 0x83,
52 	MLX5_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
53 	MLX5_IB_LINK_TYPE_IB		= 0,
54 	MLX5_IB_LINK_TYPE_ETH		= 1
55 };
56 
57 enum raw_qp_set_mask_map {
58 	MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID		= 1UL << 0,
59 	MLX5_RAW_QP_RATE_LIMIT			= 1UL << 1,
60 };
61 
62 struct mlx5_modify_raw_qp_param {
63 	u16 operation;
64 
65 	u32 set_mask; /* raw_qp_set_mask_map */
66 
67 	struct mlx5_rate_limit rl;
68 
69 	u8 rq_q_ctr_id;
70 	u16 port;
71 };
72 
73 static void get_cqs(enum ib_qp_type qp_type,
74 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
75 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
76 
77 static int is_qp0(enum ib_qp_type qp_type)
78 {
79 	return qp_type == IB_QPT_SMI;
80 }
81 
82 static int is_sqp(enum ib_qp_type qp_type)
83 {
84 	return is_qp0(qp_type) || is_qp1(qp_type);
85 }
86 
87 /**
88  * mlx5_ib_read_user_wqe_common() - Copy a WQE (or part of) from user WQ
89  * to kernel buffer
90  *
91  * @umem: User space memory where the WQ is
92  * @buffer: buffer to copy to
93  * @buflen: buffer length
94  * @wqe_index: index of WQE to copy from
95  * @wq_offset: offset to start of WQ
96  * @wq_wqe_cnt: number of WQEs in WQ
97  * @wq_wqe_shift: log2 of WQE size
98  * @bcnt: number of bytes to copy
99  * @bytes_copied: number of bytes to copy (return value)
100  *
101  * Copies from start of WQE bcnt or less bytes.
102  * Does not gurantee to copy the entire WQE.
103  *
104  * Return: zero on success, or an error code.
105  */
106 static int mlx5_ib_read_user_wqe_common(struct ib_umem *umem, void *buffer,
107 					size_t buflen, int wqe_index,
108 					int wq_offset, int wq_wqe_cnt,
109 					int wq_wqe_shift, int bcnt,
110 					size_t *bytes_copied)
111 {
112 	size_t offset = wq_offset + ((wqe_index % wq_wqe_cnt) << wq_wqe_shift);
113 	size_t wq_end = wq_offset + (wq_wqe_cnt << wq_wqe_shift);
114 	size_t copy_length;
115 	int ret;
116 
117 	/* don't copy more than requested, more than buffer length or
118 	 * beyond WQ end
119 	 */
120 	copy_length = min_t(u32, buflen, wq_end - offset);
121 	copy_length = min_t(u32, copy_length, bcnt);
122 
123 	ret = ib_umem_copy_from(buffer, umem, offset, copy_length);
124 	if (ret)
125 		return ret;
126 
127 	if (!ret && bytes_copied)
128 		*bytes_copied = copy_length;
129 
130 	return 0;
131 }
132 
133 static int mlx5_ib_read_kernel_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index,
134 				      void *buffer, size_t buflen, size_t *bc)
135 {
136 	struct mlx5_wqe_ctrl_seg *ctrl;
137 	size_t bytes_copied = 0;
138 	size_t wqe_length;
139 	void *p;
140 	int ds;
141 
142 	wqe_index = wqe_index & qp->sq.fbc.sz_m1;
143 
144 	/* read the control segment first */
145 	p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, wqe_index);
146 	ctrl = p;
147 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
148 	wqe_length = ds * MLX5_WQE_DS_UNITS;
149 
150 	/* read rest of WQE if it spreads over more than one stride */
151 	while (bytes_copied < wqe_length) {
152 		size_t copy_length =
153 			min_t(size_t, buflen - bytes_copied, MLX5_SEND_WQE_BB);
154 
155 		if (!copy_length)
156 			break;
157 
158 		memcpy(buffer + bytes_copied, p, copy_length);
159 		bytes_copied += copy_length;
160 
161 		wqe_index = (wqe_index + 1) & qp->sq.fbc.sz_m1;
162 		p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, wqe_index);
163 	}
164 	*bc = bytes_copied;
165 	return 0;
166 }
167 
168 static int mlx5_ib_read_user_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index,
169 				    void *buffer, size_t buflen, size_t *bc)
170 {
171 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
172 	struct ib_umem *umem = base->ubuffer.umem;
173 	struct mlx5_ib_wq *wq = &qp->sq;
174 	struct mlx5_wqe_ctrl_seg *ctrl;
175 	size_t bytes_copied;
176 	size_t bytes_copied2;
177 	size_t wqe_length;
178 	int ret;
179 	int ds;
180 
181 	/* at first read as much as possible */
182 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index,
183 					   wq->offset, wq->wqe_cnt,
184 					   wq->wqe_shift, buflen,
185 					   &bytes_copied);
186 	if (ret)
187 		return ret;
188 
189 	/* we need at least control segment size to proceed */
190 	if (bytes_copied < sizeof(*ctrl))
191 		return -EINVAL;
192 
193 	ctrl = buffer;
194 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
195 	wqe_length = ds * MLX5_WQE_DS_UNITS;
196 
197 	/* if we copied enough then we are done */
198 	if (bytes_copied >= wqe_length) {
199 		*bc = bytes_copied;
200 		return 0;
201 	}
202 
203 	/* otherwise this a wrapped around wqe
204 	 * so read the remaining bytes starting
205 	 * from  wqe_index 0
206 	 */
207 	ret = mlx5_ib_read_user_wqe_common(umem, buffer + bytes_copied,
208 					   buflen - bytes_copied, 0, wq->offset,
209 					   wq->wqe_cnt, wq->wqe_shift,
210 					   wqe_length - bytes_copied,
211 					   &bytes_copied2);
212 
213 	if (ret)
214 		return ret;
215 	*bc = bytes_copied + bytes_copied2;
216 	return 0;
217 }
218 
219 int mlx5_ib_read_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
220 			size_t buflen, size_t *bc)
221 {
222 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
223 	struct ib_umem *umem = base->ubuffer.umem;
224 
225 	if (buflen < sizeof(struct mlx5_wqe_ctrl_seg))
226 		return -EINVAL;
227 
228 	if (!umem)
229 		return mlx5_ib_read_kernel_wqe_sq(qp, wqe_index, buffer,
230 						  buflen, bc);
231 
232 	return mlx5_ib_read_user_wqe_sq(qp, wqe_index, buffer, buflen, bc);
233 }
234 
235 static int mlx5_ib_read_user_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index,
236 				    void *buffer, size_t buflen, size_t *bc)
237 {
238 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
239 	struct ib_umem *umem = base->ubuffer.umem;
240 	struct mlx5_ib_wq *wq = &qp->rq;
241 	size_t bytes_copied;
242 	int ret;
243 
244 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index,
245 					   wq->offset, wq->wqe_cnt,
246 					   wq->wqe_shift, buflen,
247 					   &bytes_copied);
248 
249 	if (ret)
250 		return ret;
251 	*bc = bytes_copied;
252 	return 0;
253 }
254 
255 int mlx5_ib_read_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
256 			size_t buflen, size_t *bc)
257 {
258 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
259 	struct ib_umem *umem = base->ubuffer.umem;
260 	struct mlx5_ib_wq *wq = &qp->rq;
261 	size_t wqe_size = 1 << wq->wqe_shift;
262 
263 	if (buflen < wqe_size)
264 		return -EINVAL;
265 
266 	if (!umem)
267 		return -EOPNOTSUPP;
268 
269 	return mlx5_ib_read_user_wqe_rq(qp, wqe_index, buffer, buflen, bc);
270 }
271 
272 static int mlx5_ib_read_user_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index,
273 				     void *buffer, size_t buflen, size_t *bc)
274 {
275 	struct ib_umem *umem = srq->umem;
276 	size_t bytes_copied;
277 	int ret;
278 
279 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index, 0,
280 					   srq->msrq.max, srq->msrq.wqe_shift,
281 					   buflen, &bytes_copied);
282 
283 	if (ret)
284 		return ret;
285 	*bc = bytes_copied;
286 	return 0;
287 }
288 
289 int mlx5_ib_read_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index, void *buffer,
290 			 size_t buflen, size_t *bc)
291 {
292 	struct ib_umem *umem = srq->umem;
293 	size_t wqe_size = 1 << srq->msrq.wqe_shift;
294 
295 	if (buflen < wqe_size)
296 		return -EINVAL;
297 
298 	if (!umem)
299 		return -EOPNOTSUPP;
300 
301 	return mlx5_ib_read_user_wqe_srq(srq, wqe_index, buffer, buflen, bc);
302 }
303 
304 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
305 {
306 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
307 	struct ib_event event;
308 
309 	if (type == MLX5_EVENT_TYPE_PATH_MIG) {
310 		/* This event is only valid for trans_qps */
311 		to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
312 	}
313 
314 	if (ibqp->event_handler) {
315 		event.device     = ibqp->device;
316 		event.element.qp = ibqp;
317 		switch (type) {
318 		case MLX5_EVENT_TYPE_PATH_MIG:
319 			event.event = IB_EVENT_PATH_MIG;
320 			break;
321 		case MLX5_EVENT_TYPE_COMM_EST:
322 			event.event = IB_EVENT_COMM_EST;
323 			break;
324 		case MLX5_EVENT_TYPE_SQ_DRAINED:
325 			event.event = IB_EVENT_SQ_DRAINED;
326 			break;
327 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
328 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
329 			break;
330 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
331 			event.event = IB_EVENT_QP_FATAL;
332 			break;
333 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
334 			event.event = IB_EVENT_PATH_MIG_ERR;
335 			break;
336 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
337 			event.event = IB_EVENT_QP_REQ_ERR;
338 			break;
339 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
340 			event.event = IB_EVENT_QP_ACCESS_ERR;
341 			break;
342 		default:
343 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
344 			return;
345 		}
346 
347 		ibqp->event_handler(&event, ibqp->qp_context);
348 	}
349 }
350 
351 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
352 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
353 {
354 	int wqe_size;
355 	int wq_size;
356 
357 	/* Sanity check RQ size before proceeding */
358 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
359 		return -EINVAL;
360 
361 	if (!has_rq) {
362 		qp->rq.max_gs = 0;
363 		qp->rq.wqe_cnt = 0;
364 		qp->rq.wqe_shift = 0;
365 		cap->max_recv_wr = 0;
366 		cap->max_recv_sge = 0;
367 	} else {
368 		int wq_sig = !!(qp->flags_en & MLX5_QP_FLAG_SIGNATURE);
369 
370 		if (ucmd) {
371 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
372 			if (ucmd->rq_wqe_shift > BITS_PER_BYTE * sizeof(ucmd->rq_wqe_shift))
373 				return -EINVAL;
374 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
375 			if ((1 << qp->rq.wqe_shift) /
376 				    sizeof(struct mlx5_wqe_data_seg) <
377 			    wq_sig)
378 				return -EINVAL;
379 			qp->rq.max_gs =
380 				(1 << qp->rq.wqe_shift) /
381 					sizeof(struct mlx5_wqe_data_seg) -
382 				wq_sig;
383 			qp->rq.max_post = qp->rq.wqe_cnt;
384 		} else {
385 			wqe_size =
386 				wq_sig ? sizeof(struct mlx5_wqe_signature_seg) :
387 					 0;
388 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
389 			wqe_size = roundup_pow_of_two(wqe_size);
390 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
391 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
392 			qp->rq.wqe_cnt = wq_size / wqe_size;
393 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
394 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
395 					    wqe_size,
396 					    MLX5_CAP_GEN(dev->mdev,
397 							 max_wqe_sz_rq));
398 				return -EINVAL;
399 			}
400 			qp->rq.wqe_shift = ilog2(wqe_size);
401 			qp->rq.max_gs =
402 				(1 << qp->rq.wqe_shift) /
403 					sizeof(struct mlx5_wqe_data_seg) -
404 				wq_sig;
405 			qp->rq.max_post = qp->rq.wqe_cnt;
406 		}
407 	}
408 
409 	return 0;
410 }
411 
412 static int sq_overhead(struct ib_qp_init_attr *attr)
413 {
414 	int size = 0;
415 
416 	switch (attr->qp_type) {
417 	case IB_QPT_XRC_INI:
418 		size += sizeof(struct mlx5_wqe_xrc_seg);
419 		fallthrough;
420 	case IB_QPT_RC:
421 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
422 			max(sizeof(struct mlx5_wqe_atomic_seg) +
423 			    sizeof(struct mlx5_wqe_raddr_seg),
424 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
425 			    sizeof(struct mlx5_mkey_seg) +
426 			    MLX5_IB_SQ_UMR_INLINE_THRESHOLD /
427 			    MLX5_IB_UMR_OCTOWORD);
428 		break;
429 
430 	case IB_QPT_XRC_TGT:
431 		return 0;
432 
433 	case IB_QPT_UC:
434 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
435 			max(sizeof(struct mlx5_wqe_raddr_seg),
436 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
437 			    sizeof(struct mlx5_mkey_seg));
438 		break;
439 
440 	case IB_QPT_UD:
441 		if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
442 			size += sizeof(struct mlx5_wqe_eth_pad) +
443 				sizeof(struct mlx5_wqe_eth_seg);
444 		fallthrough;
445 	case IB_QPT_SMI:
446 	case MLX5_IB_QPT_HW_GSI:
447 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
448 			sizeof(struct mlx5_wqe_datagram_seg);
449 		break;
450 
451 	case MLX5_IB_QPT_REG_UMR:
452 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
453 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
454 			sizeof(struct mlx5_mkey_seg);
455 		break;
456 
457 	default:
458 		return -EINVAL;
459 	}
460 
461 	return size;
462 }
463 
464 static int calc_send_wqe(struct ib_qp_init_attr *attr)
465 {
466 	int inl_size = 0;
467 	int size;
468 
469 	size = sq_overhead(attr);
470 	if (size < 0)
471 		return size;
472 
473 	if (attr->cap.max_inline_data) {
474 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
475 			attr->cap.max_inline_data;
476 	}
477 
478 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
479 	if (attr->create_flags & IB_QP_CREATE_INTEGRITY_EN &&
480 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
481 		return MLX5_SIG_WQE_SIZE;
482 	else
483 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
484 }
485 
486 static int get_send_sge(struct ib_qp_init_attr *attr, int wqe_size)
487 {
488 	int max_sge;
489 
490 	if (attr->qp_type == IB_QPT_RC)
491 		max_sge = (min_t(int, wqe_size, 512) -
492 			   sizeof(struct mlx5_wqe_ctrl_seg) -
493 			   sizeof(struct mlx5_wqe_raddr_seg)) /
494 			sizeof(struct mlx5_wqe_data_seg);
495 	else if (attr->qp_type == IB_QPT_XRC_INI)
496 		max_sge = (min_t(int, wqe_size, 512) -
497 			   sizeof(struct mlx5_wqe_ctrl_seg) -
498 			   sizeof(struct mlx5_wqe_xrc_seg) -
499 			   sizeof(struct mlx5_wqe_raddr_seg)) /
500 			sizeof(struct mlx5_wqe_data_seg);
501 	else
502 		max_sge = (wqe_size - sq_overhead(attr)) /
503 			sizeof(struct mlx5_wqe_data_seg);
504 
505 	return min_t(int, max_sge, wqe_size - sq_overhead(attr) /
506 		     sizeof(struct mlx5_wqe_data_seg));
507 }
508 
509 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
510 			struct mlx5_ib_qp *qp)
511 {
512 	int wqe_size;
513 	int wq_size;
514 
515 	if (!attr->cap.max_send_wr)
516 		return 0;
517 
518 	wqe_size = calc_send_wqe(attr);
519 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
520 	if (wqe_size < 0)
521 		return wqe_size;
522 
523 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
524 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
525 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
526 		return -EINVAL;
527 	}
528 
529 	qp->max_inline_data = wqe_size - sq_overhead(attr) -
530 			      sizeof(struct mlx5_wqe_inline_seg);
531 	attr->cap.max_inline_data = qp->max_inline_data;
532 
533 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
534 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
535 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
536 		mlx5_ib_dbg(dev, "send queue size (%d * %d / %d -> %d) exceeds limits(%d)\n",
537 			    attr->cap.max_send_wr, wqe_size, MLX5_SEND_WQE_BB,
538 			    qp->sq.wqe_cnt,
539 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
540 		return -ENOMEM;
541 	}
542 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
543 	qp->sq.max_gs = get_send_sge(attr, wqe_size);
544 	if (qp->sq.max_gs < attr->cap.max_send_sge)
545 		return -ENOMEM;
546 
547 	attr->cap.max_send_sge = qp->sq.max_gs;
548 	qp->sq.max_post = wq_size / wqe_size;
549 	attr->cap.max_send_wr = qp->sq.max_post;
550 
551 	return wq_size;
552 }
553 
554 static int set_user_buf_size(struct mlx5_ib_dev *dev,
555 			    struct mlx5_ib_qp *qp,
556 			    struct mlx5_ib_create_qp *ucmd,
557 			    struct mlx5_ib_qp_base *base,
558 			    struct ib_qp_init_attr *attr)
559 {
560 	int desc_sz = 1 << qp->sq.wqe_shift;
561 
562 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
563 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
564 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
565 		return -EINVAL;
566 	}
567 
568 	if (ucmd->sq_wqe_count && !is_power_of_2(ucmd->sq_wqe_count)) {
569 		mlx5_ib_warn(dev, "sq_wqe_count %d is not a power of two\n",
570 			     ucmd->sq_wqe_count);
571 		return -EINVAL;
572 	}
573 
574 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
575 
576 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
577 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
578 			     qp->sq.wqe_cnt,
579 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
580 		return -EINVAL;
581 	}
582 
583 	if (attr->qp_type == IB_QPT_RAW_PACKET ||
584 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
585 		base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
586 		qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
587 	} else {
588 		base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
589 					 (qp->sq.wqe_cnt << 6);
590 	}
591 
592 	return 0;
593 }
594 
595 static int qp_has_rq(struct ib_qp_init_attr *attr)
596 {
597 	if (attr->qp_type == IB_QPT_XRC_INI ||
598 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
599 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
600 	    !attr->cap.max_recv_wr)
601 		return 0;
602 
603 	return 1;
604 }
605 
606 enum {
607 	/* this is the first blue flame register in the array of bfregs assigned
608 	 * to a processes. Since we do not use it for blue flame but rather
609 	 * regular 64 bit doorbells, we do not need a lock for maintaiing
610 	 * "odd/even" order
611 	 */
612 	NUM_NON_BLUE_FLAME_BFREGS = 1,
613 };
614 
615 static int max_bfregs(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi)
616 {
617 	return get_num_static_uars(dev, bfregi) * MLX5_NON_FP_BFREGS_PER_UAR;
618 }
619 
620 static int num_med_bfreg(struct mlx5_ib_dev *dev,
621 			 struct mlx5_bfreg_info *bfregi)
622 {
623 	int n;
624 
625 	n = max_bfregs(dev, bfregi) - bfregi->num_low_latency_bfregs -
626 	    NUM_NON_BLUE_FLAME_BFREGS;
627 
628 	return n >= 0 ? n : 0;
629 }
630 
631 static int first_med_bfreg(struct mlx5_ib_dev *dev,
632 			   struct mlx5_bfreg_info *bfregi)
633 {
634 	return num_med_bfreg(dev, bfregi) ? 1 : -ENOMEM;
635 }
636 
637 static int first_hi_bfreg(struct mlx5_ib_dev *dev,
638 			  struct mlx5_bfreg_info *bfregi)
639 {
640 	int med;
641 
642 	med = num_med_bfreg(dev, bfregi);
643 	return ++med;
644 }
645 
646 static int alloc_high_class_bfreg(struct mlx5_ib_dev *dev,
647 				  struct mlx5_bfreg_info *bfregi)
648 {
649 	int i;
650 
651 	for (i = first_hi_bfreg(dev, bfregi); i < max_bfregs(dev, bfregi); i++) {
652 		if (!bfregi->count[i]) {
653 			bfregi->count[i]++;
654 			return i;
655 		}
656 	}
657 
658 	return -ENOMEM;
659 }
660 
661 static int alloc_med_class_bfreg(struct mlx5_ib_dev *dev,
662 				 struct mlx5_bfreg_info *bfregi)
663 {
664 	int minidx = first_med_bfreg(dev, bfregi);
665 	int i;
666 
667 	if (minidx < 0)
668 		return minidx;
669 
670 	for (i = minidx; i < first_hi_bfreg(dev, bfregi); i++) {
671 		if (bfregi->count[i] < bfregi->count[minidx])
672 			minidx = i;
673 		if (!bfregi->count[minidx])
674 			break;
675 	}
676 
677 	bfregi->count[minidx]++;
678 	return minidx;
679 }
680 
681 static int alloc_bfreg(struct mlx5_ib_dev *dev,
682 		       struct mlx5_bfreg_info *bfregi)
683 {
684 	int bfregn = -ENOMEM;
685 
686 	if (bfregi->lib_uar_dyn)
687 		return -EINVAL;
688 
689 	mutex_lock(&bfregi->lock);
690 	if (bfregi->ver >= 2) {
691 		bfregn = alloc_high_class_bfreg(dev, bfregi);
692 		if (bfregn < 0)
693 			bfregn = alloc_med_class_bfreg(dev, bfregi);
694 	}
695 
696 	if (bfregn < 0) {
697 		BUILD_BUG_ON(NUM_NON_BLUE_FLAME_BFREGS != 1);
698 		bfregn = 0;
699 		bfregi->count[bfregn]++;
700 	}
701 	mutex_unlock(&bfregi->lock);
702 
703 	return bfregn;
704 }
705 
706 void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
707 {
708 	mutex_lock(&bfregi->lock);
709 	bfregi->count[bfregn]--;
710 	mutex_unlock(&bfregi->lock);
711 }
712 
713 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
714 {
715 	switch (state) {
716 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
717 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
718 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
719 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
720 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
721 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
722 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
723 	default:		return -1;
724 	}
725 }
726 
727 static int to_mlx5_st(enum ib_qp_type type)
728 {
729 	switch (type) {
730 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
731 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
732 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
733 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
734 	case IB_QPT_XRC_INI:
735 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
736 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
737 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
738 	case MLX5_IB_QPT_DCI:		return MLX5_QP_ST_DCI;
739 	case IB_QPT_RAW_PACKET:		return MLX5_QP_ST_RAW_ETHERTYPE;
740 	default:		return -EINVAL;
741 	}
742 }
743 
744 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
745 			     struct mlx5_ib_cq *recv_cq);
746 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
747 			       struct mlx5_ib_cq *recv_cq);
748 
749 int bfregn_to_uar_index(struct mlx5_ib_dev *dev,
750 			struct mlx5_bfreg_info *bfregi, u32 bfregn,
751 			bool dyn_bfreg)
752 {
753 	unsigned int bfregs_per_sys_page;
754 	u32 index_of_sys_page;
755 	u32 offset;
756 
757 	if (bfregi->lib_uar_dyn)
758 		return -EINVAL;
759 
760 	bfregs_per_sys_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
761 				MLX5_NON_FP_BFREGS_PER_UAR;
762 	index_of_sys_page = bfregn / bfregs_per_sys_page;
763 
764 	if (dyn_bfreg) {
765 		index_of_sys_page += bfregi->num_static_sys_pages;
766 
767 		if (index_of_sys_page >= bfregi->num_sys_pages)
768 			return -EINVAL;
769 
770 		if (bfregn > bfregi->num_dyn_bfregs ||
771 		    bfregi->sys_pages[index_of_sys_page] == MLX5_IB_INVALID_UAR_INDEX) {
772 			mlx5_ib_dbg(dev, "Invalid dynamic uar index\n");
773 			return -EINVAL;
774 		}
775 	}
776 
777 	offset = bfregn % bfregs_per_sys_page / MLX5_NON_FP_BFREGS_PER_UAR;
778 	return bfregi->sys_pages[index_of_sys_page] + offset;
779 }
780 
781 static void destroy_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
782 			    struct mlx5_ib_rwq *rwq, struct ib_udata *udata)
783 {
784 	struct mlx5_ib_ucontext *context =
785 		rdma_udata_to_drv_context(
786 			udata,
787 			struct mlx5_ib_ucontext,
788 			ibucontext);
789 
790 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_DELAY_DROP)
791 		atomic_dec(&dev->delay_drop.rqs_cnt);
792 
793 	mlx5_ib_db_unmap_user(context, &rwq->db);
794 	ib_umem_release(rwq->umem);
795 }
796 
797 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
798 			  struct ib_udata *udata, struct mlx5_ib_rwq *rwq,
799 			  struct mlx5_ib_create_wq *ucmd)
800 {
801 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
802 		udata, struct mlx5_ib_ucontext, ibucontext);
803 	unsigned long page_size = 0;
804 	u32 offset = 0;
805 	int err;
806 
807 	if (!ucmd->buf_addr)
808 		return -EINVAL;
809 
810 	rwq->umem = ib_umem_get(&dev->ib_dev, ucmd->buf_addr, rwq->buf_size, 0);
811 	if (IS_ERR(rwq->umem)) {
812 		mlx5_ib_dbg(dev, "umem_get failed\n");
813 		err = PTR_ERR(rwq->umem);
814 		return err;
815 	}
816 
817 	page_size = mlx5_umem_find_best_quantized_pgoff(
818 		rwq->umem, wq, log_wq_pg_sz, MLX5_ADAPTER_PAGE_SHIFT,
819 		page_offset, 64, &rwq->rq_page_offset);
820 	if (!page_size) {
821 		mlx5_ib_warn(dev, "bad offset\n");
822 		err = -EINVAL;
823 		goto err_umem;
824 	}
825 
826 	rwq->rq_num_pas = ib_umem_num_dma_blocks(rwq->umem, page_size);
827 	rwq->page_shift = order_base_2(page_size);
828 	rwq->log_page_size =  rwq->page_shift - MLX5_ADAPTER_PAGE_SHIFT;
829 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
830 
831 	mlx5_ib_dbg(
832 		dev,
833 		"addr 0x%llx, size %zd, npages %zu, page_size %ld, ncont %d, offset %d\n",
834 		(unsigned long long)ucmd->buf_addr, rwq->buf_size,
835 		ib_umem_num_pages(rwq->umem), page_size, rwq->rq_num_pas,
836 		offset);
837 
838 	err = mlx5_ib_db_map_user(ucontext, udata, ucmd->db_addr, &rwq->db);
839 	if (err) {
840 		mlx5_ib_dbg(dev, "map failed\n");
841 		goto err_umem;
842 	}
843 
844 	return 0;
845 
846 err_umem:
847 	ib_umem_release(rwq->umem);
848 	return err;
849 }
850 
851 static int adjust_bfregn(struct mlx5_ib_dev *dev,
852 			 struct mlx5_bfreg_info *bfregi, int bfregn)
853 {
854 	return bfregn / MLX5_NON_FP_BFREGS_PER_UAR * MLX5_BFREGS_PER_UAR +
855 				bfregn % MLX5_NON_FP_BFREGS_PER_UAR;
856 }
857 
858 static int _create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
859 			   struct mlx5_ib_qp *qp, struct ib_udata *udata,
860 			   struct ib_qp_init_attr *attr, u32 **in,
861 			   struct mlx5_ib_create_qp_resp *resp, int *inlen,
862 			   struct mlx5_ib_qp_base *base,
863 			   struct mlx5_ib_create_qp *ucmd)
864 {
865 	struct mlx5_ib_ucontext *context;
866 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
867 	unsigned int page_offset_quantized = 0;
868 	unsigned long page_size = 0;
869 	int uar_index = 0;
870 	int bfregn;
871 	int ncont = 0;
872 	__be64 *pas;
873 	void *qpc;
874 	int err;
875 	u16 uid;
876 	u32 uar_flags;
877 
878 	context = rdma_udata_to_drv_context(udata, struct mlx5_ib_ucontext,
879 					    ibucontext);
880 	uar_flags = qp->flags_en &
881 		    (MLX5_QP_FLAG_UAR_PAGE_INDEX | MLX5_QP_FLAG_BFREG_INDEX);
882 	switch (uar_flags) {
883 	case MLX5_QP_FLAG_UAR_PAGE_INDEX:
884 		uar_index = ucmd->bfreg_index;
885 		bfregn = MLX5_IB_INVALID_BFREG;
886 		break;
887 	case MLX5_QP_FLAG_BFREG_INDEX:
888 		uar_index = bfregn_to_uar_index(dev, &context->bfregi,
889 						ucmd->bfreg_index, true);
890 		if (uar_index < 0)
891 			return uar_index;
892 		bfregn = MLX5_IB_INVALID_BFREG;
893 		break;
894 	case 0:
895 		if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
896 			return -EINVAL;
897 		bfregn = alloc_bfreg(dev, &context->bfregi);
898 		if (bfregn < 0)
899 			return bfregn;
900 		break;
901 	default:
902 		return -EINVAL;
903 	}
904 
905 	mlx5_ib_dbg(dev, "bfregn 0x%x, uar_index 0x%x\n", bfregn, uar_index);
906 	if (bfregn != MLX5_IB_INVALID_BFREG)
907 		uar_index = bfregn_to_uar_index(dev, &context->bfregi, bfregn,
908 						false);
909 
910 	qp->rq.offset = 0;
911 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
912 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
913 
914 	err = set_user_buf_size(dev, qp, ucmd, base, attr);
915 	if (err)
916 		goto err_bfreg;
917 
918 	if (ucmd->buf_addr && ubuffer->buf_size) {
919 		ubuffer->buf_addr = ucmd->buf_addr;
920 		ubuffer->umem = ib_umem_get(&dev->ib_dev, ubuffer->buf_addr,
921 					    ubuffer->buf_size, 0);
922 		if (IS_ERR(ubuffer->umem)) {
923 			err = PTR_ERR(ubuffer->umem);
924 			goto err_bfreg;
925 		}
926 		page_size = mlx5_umem_find_best_quantized_pgoff(
927 			ubuffer->umem, qpc, log_page_size,
928 			MLX5_ADAPTER_PAGE_SHIFT, page_offset, 64,
929 			&page_offset_quantized);
930 		if (!page_size) {
931 			err = -EINVAL;
932 			goto err_umem;
933 		}
934 		ncont = ib_umem_num_dma_blocks(ubuffer->umem, page_size);
935 	} else {
936 		ubuffer->umem = NULL;
937 	}
938 
939 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
940 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
941 	*in = kvzalloc(*inlen, GFP_KERNEL);
942 	if (!*in) {
943 		err = -ENOMEM;
944 		goto err_umem;
945 	}
946 
947 	uid = (attr->qp_type != IB_QPT_XRC_INI) ? to_mpd(pd)->uid : 0;
948 	MLX5_SET(create_qp_in, *in, uid, uid);
949 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
950 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
951 	if (ubuffer->umem) {
952 		mlx5_ib_populate_pas(ubuffer->umem, page_size, pas, 0);
953 		MLX5_SET(qpc, qpc, log_page_size,
954 			 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
955 		MLX5_SET(qpc, qpc, page_offset, page_offset_quantized);
956 	}
957 	MLX5_SET(qpc, qpc, uar_page, uar_index);
958 	if (bfregn != MLX5_IB_INVALID_BFREG)
959 		resp->bfreg_index = adjust_bfregn(dev, &context->bfregi, bfregn);
960 	else
961 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
962 	qp->bfregn = bfregn;
963 
964 	err = mlx5_ib_db_map_user(context, udata, ucmd->db_addr, &qp->db);
965 	if (err) {
966 		mlx5_ib_dbg(dev, "map failed\n");
967 		goto err_free;
968 	}
969 
970 	return 0;
971 
972 err_free:
973 	kvfree(*in);
974 
975 err_umem:
976 	ib_umem_release(ubuffer->umem);
977 
978 err_bfreg:
979 	if (bfregn != MLX5_IB_INVALID_BFREG)
980 		mlx5_ib_free_bfreg(dev, &context->bfregi, bfregn);
981 	return err;
982 }
983 
984 static void destroy_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
985 		       struct mlx5_ib_qp_base *base, struct ib_udata *udata)
986 {
987 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
988 		udata, struct mlx5_ib_ucontext, ibucontext);
989 
990 	if (udata) {
991 		/* User QP */
992 		mlx5_ib_db_unmap_user(context, &qp->db);
993 		ib_umem_release(base->ubuffer.umem);
994 
995 		/*
996 		 * Free only the BFREGs which are handled by the kernel.
997 		 * BFREGs of UARs allocated dynamically are handled by user.
998 		 */
999 		if (qp->bfregn != MLX5_IB_INVALID_BFREG)
1000 			mlx5_ib_free_bfreg(dev, &context->bfregi, qp->bfregn);
1001 		return;
1002 	}
1003 
1004 	/* Kernel QP */
1005 	kvfree(qp->sq.wqe_head);
1006 	kvfree(qp->sq.w_list);
1007 	kvfree(qp->sq.wrid);
1008 	kvfree(qp->sq.wr_data);
1009 	kvfree(qp->rq.wrid);
1010 	if (qp->db.db)
1011 		mlx5_db_free(dev->mdev, &qp->db);
1012 	if (qp->buf.frags)
1013 		mlx5_frag_buf_free(dev->mdev, &qp->buf);
1014 }
1015 
1016 static int _create_kernel_qp(struct mlx5_ib_dev *dev,
1017 			     struct ib_qp_init_attr *init_attr,
1018 			     struct mlx5_ib_qp *qp, u32 **in, int *inlen,
1019 			     struct mlx5_ib_qp_base *base)
1020 {
1021 	int uar_index;
1022 	void *qpc;
1023 	int err;
1024 
1025 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
1026 		qp->bf.bfreg = &dev->fp_bfreg;
1027 	else if (qp->flags & MLX5_IB_QP_CREATE_WC_TEST)
1028 		qp->bf.bfreg = &dev->wc_bfreg;
1029 	else
1030 		qp->bf.bfreg = &dev->bfreg;
1031 
1032 	/* We need to divide by two since each register is comprised of
1033 	 * two buffers of identical size, namely odd and even
1034 	 */
1035 	qp->bf.buf_size = (1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size)) / 2;
1036 	uar_index = qp->bf.bfreg->index;
1037 
1038 	err = calc_sq_size(dev, init_attr, qp);
1039 	if (err < 0) {
1040 		mlx5_ib_dbg(dev, "err %d\n", err);
1041 		return err;
1042 	}
1043 
1044 	qp->rq.offset = 0;
1045 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
1046 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
1047 
1048 	err = mlx5_frag_buf_alloc_node(dev->mdev, base->ubuffer.buf_size,
1049 				       &qp->buf, dev->mdev->priv.numa_node);
1050 	if (err) {
1051 		mlx5_ib_dbg(dev, "err %d\n", err);
1052 		return err;
1053 	}
1054 
1055 	if (qp->rq.wqe_cnt)
1056 		mlx5_init_fbc(qp->buf.frags, qp->rq.wqe_shift,
1057 			      ilog2(qp->rq.wqe_cnt), &qp->rq.fbc);
1058 
1059 	if (qp->sq.wqe_cnt) {
1060 		int sq_strides_offset = (qp->sq.offset  & (PAGE_SIZE - 1)) /
1061 					MLX5_SEND_WQE_BB;
1062 		mlx5_init_fbc_offset(qp->buf.frags +
1063 				     (qp->sq.offset / PAGE_SIZE),
1064 				     ilog2(MLX5_SEND_WQE_BB),
1065 				     ilog2(qp->sq.wqe_cnt),
1066 				     sq_strides_offset, &qp->sq.fbc);
1067 
1068 		qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
1069 	}
1070 
1071 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
1072 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
1073 	*in = kvzalloc(*inlen, GFP_KERNEL);
1074 	if (!*in) {
1075 		err = -ENOMEM;
1076 		goto err_buf;
1077 	}
1078 
1079 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
1080 	MLX5_SET(qpc, qpc, uar_page, uar_index);
1081 	MLX5_SET(qpc, qpc, ts_format, MLX5_QPC_TIMESTAMP_FORMAT_DEFAULT);
1082 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1083 
1084 	/* Set "fast registration enabled" for all kernel QPs */
1085 	MLX5_SET(qpc, qpc, fre, 1);
1086 	MLX5_SET(qpc, qpc, rlky, 1);
1087 
1088 	if (qp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
1089 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
1090 
1091 	mlx5_fill_page_frag_array(&qp->buf,
1092 				  (__be64 *)MLX5_ADDR_OF(create_qp_in,
1093 							 *in, pas));
1094 
1095 	err = mlx5_db_alloc(dev->mdev, &qp->db);
1096 	if (err) {
1097 		mlx5_ib_dbg(dev, "err %d\n", err);
1098 		goto err_free;
1099 	}
1100 
1101 	qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1102 				     sizeof(*qp->sq.wrid), GFP_KERNEL);
1103 	qp->sq.wr_data = kvmalloc_array(qp->sq.wqe_cnt,
1104 					sizeof(*qp->sq.wr_data), GFP_KERNEL);
1105 	qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1106 				     sizeof(*qp->rq.wrid), GFP_KERNEL);
1107 	qp->sq.w_list = kvmalloc_array(qp->sq.wqe_cnt,
1108 				       sizeof(*qp->sq.w_list), GFP_KERNEL);
1109 	qp->sq.wqe_head = kvmalloc_array(qp->sq.wqe_cnt,
1110 					 sizeof(*qp->sq.wqe_head), GFP_KERNEL);
1111 
1112 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
1113 	    !qp->sq.w_list || !qp->sq.wqe_head) {
1114 		err = -ENOMEM;
1115 		goto err_wrid;
1116 	}
1117 
1118 	return 0;
1119 
1120 err_wrid:
1121 	kvfree(qp->sq.wqe_head);
1122 	kvfree(qp->sq.w_list);
1123 	kvfree(qp->sq.wrid);
1124 	kvfree(qp->sq.wr_data);
1125 	kvfree(qp->rq.wrid);
1126 	mlx5_db_free(dev->mdev, &qp->db);
1127 
1128 err_free:
1129 	kvfree(*in);
1130 
1131 err_buf:
1132 	mlx5_frag_buf_free(dev->mdev, &qp->buf);
1133 	return err;
1134 }
1135 
1136 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
1137 {
1138 	if (attr->srq || (qp->type == IB_QPT_XRC_TGT) ||
1139 	    (qp->type == MLX5_IB_QPT_DCI) || (qp->type == IB_QPT_XRC_INI))
1140 		return MLX5_SRQ_RQ;
1141 	else if (!qp->has_rq)
1142 		return MLX5_ZERO_LEN_RQ;
1143 
1144 	return MLX5_NON_ZERO_RQ;
1145 }
1146 
1147 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1148 				    struct mlx5_ib_qp *qp,
1149 				    struct mlx5_ib_sq *sq, u32 tdn,
1150 				    struct ib_pd *pd)
1151 {
1152 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
1153 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1154 
1155 	MLX5_SET(create_tis_in, in, uid, to_mpd(pd)->uid);
1156 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1157 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
1158 		MLX5_SET(tisc, tisc, underlay_qpn, qp->underlay_qpn);
1159 
1160 	return mlx5_core_create_tis(dev->mdev, in, &sq->tisn);
1161 }
1162 
1163 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1164 				      struct mlx5_ib_sq *sq, struct ib_pd *pd)
1165 {
1166 	mlx5_cmd_destroy_tis(dev->mdev, sq->tisn, to_mpd(pd)->uid);
1167 }
1168 
1169 static void destroy_flow_rule_vport_sq(struct mlx5_ib_sq *sq)
1170 {
1171 	if (sq->flow_rule)
1172 		mlx5_del_flow_rules(sq->flow_rule);
1173 	sq->flow_rule = NULL;
1174 }
1175 
1176 static int get_rq_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq)
1177 {
1178 	bool fr_supported =
1179 		MLX5_CAP_GEN(dev->mdev, rq_ts_format) ==
1180 			MLX5_RQ_TIMESTAMP_FORMAT_CAP_FREE_RUNNING ||
1181 		MLX5_CAP_GEN(dev->mdev, rq_ts_format) ==
1182 			MLX5_RQ_TIMESTAMP_FORMAT_CAP_FREE_RUNNING_AND_REAL_TIME;
1183 
1184 	if (send_cq->create_flags & IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION) {
1185 		if (!fr_supported) {
1186 			mlx5_ib_dbg(dev, "Free running TS format is not supported\n");
1187 			return -EOPNOTSUPP;
1188 		}
1189 		return MLX5_RQC_TIMESTAMP_FORMAT_FREE_RUNNING;
1190 	}
1191 	return MLX5_RQC_TIMESTAMP_FORMAT_DEFAULT;
1192 }
1193 
1194 static int get_sq_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq)
1195 {
1196 	bool fr_supported =
1197 		MLX5_CAP_GEN(dev->mdev, sq_ts_format) ==
1198 			MLX5_SQ_TIMESTAMP_FORMAT_CAP_FREE_RUNNING ||
1199 		MLX5_CAP_GEN(dev->mdev, sq_ts_format) ==
1200 			MLX5_SQ_TIMESTAMP_FORMAT_CAP_FREE_RUNNING_AND_REAL_TIME;
1201 
1202 	if (send_cq->create_flags & IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION) {
1203 		if (!fr_supported) {
1204 			mlx5_ib_dbg(dev, "Free running TS format is not supported\n");
1205 			return -EOPNOTSUPP;
1206 		}
1207 		return MLX5_SQC_TIMESTAMP_FORMAT_FREE_RUNNING;
1208 	}
1209 	return MLX5_SQC_TIMESTAMP_FORMAT_DEFAULT;
1210 }
1211 
1212 static int get_qp_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq,
1213 			    struct mlx5_ib_cq *recv_cq)
1214 {
1215 	bool fr_supported =
1216 		MLX5_CAP_ROCE(dev->mdev, qp_ts_format) ==
1217 			MLX5_QP_TIMESTAMP_FORMAT_CAP_FREE_RUNNING ||
1218 		MLX5_CAP_ROCE(dev->mdev, qp_ts_format) ==
1219 			MLX5_QP_TIMESTAMP_FORMAT_CAP_FREE_RUNNING_AND_REAL_TIME;
1220 	int ts_format = MLX5_QPC_TIMESTAMP_FORMAT_DEFAULT;
1221 
1222 	if (recv_cq &&
1223 	    recv_cq->create_flags & IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION)
1224 		ts_format = MLX5_QPC_TIMESTAMP_FORMAT_FREE_RUNNING;
1225 
1226 	if (send_cq &&
1227 	    send_cq->create_flags & IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION)
1228 		ts_format = MLX5_QPC_TIMESTAMP_FORMAT_FREE_RUNNING;
1229 
1230 	if (ts_format == MLX5_QPC_TIMESTAMP_FORMAT_FREE_RUNNING &&
1231 	    !fr_supported) {
1232 		mlx5_ib_dbg(dev, "Free running TS format is not supported\n");
1233 		return -EOPNOTSUPP;
1234 	}
1235 	return ts_format;
1236 }
1237 
1238 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1239 				   struct ib_udata *udata,
1240 				   struct mlx5_ib_sq *sq, void *qpin,
1241 				   struct ib_pd *pd, struct mlx5_ib_cq *cq)
1242 {
1243 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1244 	__be64 *pas;
1245 	void *in;
1246 	void *sqc;
1247 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1248 	void *wq;
1249 	int inlen;
1250 	int err;
1251 	unsigned int page_offset_quantized;
1252 	unsigned long page_size;
1253 	int ts_format;
1254 
1255 	ts_format = get_sq_ts_format(dev, cq);
1256 	if (ts_format < 0)
1257 		return ts_format;
1258 
1259 	sq->ubuffer.umem = ib_umem_get(&dev->ib_dev, ubuffer->buf_addr,
1260 				       ubuffer->buf_size, 0);
1261 	if (IS_ERR(sq->ubuffer.umem))
1262 		return PTR_ERR(sq->ubuffer.umem);
1263 	page_size = mlx5_umem_find_best_quantized_pgoff(
1264 		ubuffer->umem, wq, log_wq_pg_sz, MLX5_ADAPTER_PAGE_SHIFT,
1265 		page_offset, 64, &page_offset_quantized);
1266 	if (!page_size) {
1267 		err = -EINVAL;
1268 		goto err_umem;
1269 	}
1270 
1271 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
1272 		sizeof(u64) *
1273 			ib_umem_num_dma_blocks(sq->ubuffer.umem, page_size);
1274 	in = kvzalloc(inlen, GFP_KERNEL);
1275 	if (!in) {
1276 		err = -ENOMEM;
1277 		goto err_umem;
1278 	}
1279 
1280 	MLX5_SET(create_sq_in, in, uid, to_mpd(pd)->uid);
1281 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1282 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1283 	if (MLX5_CAP_ETH(dev->mdev, multi_pkt_send_wqe))
1284 		MLX5_SET(sqc, sqc, allow_multi_pkt_send_wqe, 1);
1285 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1286 	MLX5_SET(sqc, sqc, ts_format, ts_format);
1287 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1288 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1289 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1290 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1291 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
1292 	    MLX5_CAP_ETH(dev->mdev, swp))
1293 		MLX5_SET(sqc, sqc, allow_swp, 1);
1294 
1295 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1296 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1297 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1298 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1299 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1300 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1301 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1302 	MLX5_SET(wq, wq, log_wq_pg_sz,
1303 		 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
1304 	MLX5_SET(wq, wq, page_offset, page_offset_quantized);
1305 
1306 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1307 	mlx5_ib_populate_pas(sq->ubuffer.umem, page_size, pas, 0);
1308 
1309 	err = mlx5_core_create_sq_tracked(dev, in, inlen, &sq->base.mqp);
1310 
1311 	kvfree(in);
1312 
1313 	if (err)
1314 		goto err_umem;
1315 
1316 	return 0;
1317 
1318 err_umem:
1319 	ib_umem_release(sq->ubuffer.umem);
1320 	sq->ubuffer.umem = NULL;
1321 
1322 	return err;
1323 }
1324 
1325 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1326 				     struct mlx5_ib_sq *sq)
1327 {
1328 	destroy_flow_rule_vport_sq(sq);
1329 	mlx5_core_destroy_sq_tracked(dev, &sq->base.mqp);
1330 	ib_umem_release(sq->ubuffer.umem);
1331 }
1332 
1333 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1334 				   struct mlx5_ib_rq *rq, void *qpin,
1335 				   struct ib_pd *pd, struct mlx5_ib_cq *cq)
1336 {
1337 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1338 	__be64 *pas;
1339 	void *in;
1340 	void *rqc;
1341 	void *wq;
1342 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1343 	struct ib_umem *umem = rq->base.ubuffer.umem;
1344 	unsigned int page_offset_quantized;
1345 	unsigned long page_size = 0;
1346 	int ts_format;
1347 	size_t inlen;
1348 	int err;
1349 
1350 	ts_format = get_rq_ts_format(dev, cq);
1351 	if (ts_format < 0)
1352 		return ts_format;
1353 
1354 	page_size = mlx5_umem_find_best_quantized_pgoff(umem, wq, log_wq_pg_sz,
1355 							MLX5_ADAPTER_PAGE_SHIFT,
1356 							page_offset, 64,
1357 							&page_offset_quantized);
1358 	if (!page_size)
1359 		return -EINVAL;
1360 
1361 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
1362 		sizeof(u64) * ib_umem_num_dma_blocks(umem, page_size);
1363 	in = kvzalloc(inlen, GFP_KERNEL);
1364 	if (!in)
1365 		return -ENOMEM;
1366 
1367 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
1368 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1369 	if (!(rq->flags & MLX5_IB_RQ_CVLAN_STRIPPING))
1370 		MLX5_SET(rqc, rqc, vsd, 1);
1371 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1372 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1373 	MLX5_SET(rqc, rqc, ts_format, ts_format);
1374 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1375 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1376 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1377 
1378 	if (mqp->flags & IB_QP_CREATE_SCATTER_FCS)
1379 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1380 
1381 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1382 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1383 	if (rq->flags & MLX5_IB_RQ_PCI_WRITE_END_PADDING)
1384 		MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1385 	MLX5_SET(wq, wq, page_offset, page_offset_quantized);
1386 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1387 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1388 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1389 	MLX5_SET(wq, wq, log_wq_pg_sz,
1390 		 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
1391 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1392 
1393 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1394 	mlx5_ib_populate_pas(umem, page_size, pas, 0);
1395 
1396 	err = mlx5_core_create_rq_tracked(dev, in, inlen, &rq->base.mqp);
1397 
1398 	kvfree(in);
1399 
1400 	return err;
1401 }
1402 
1403 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1404 				     struct mlx5_ib_rq *rq)
1405 {
1406 	mlx5_core_destroy_rq_tracked(dev, &rq->base.mqp);
1407 }
1408 
1409 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1410 				      struct mlx5_ib_rq *rq,
1411 				      u32 qp_flags_en,
1412 				      struct ib_pd *pd)
1413 {
1414 	if (qp_flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1415 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1416 		mlx5_ib_disable_lb(dev, false, true);
1417 	mlx5_cmd_destroy_tir(dev->mdev, rq->tirn, to_mpd(pd)->uid);
1418 }
1419 
1420 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1421 				    struct mlx5_ib_rq *rq, u32 tdn,
1422 				    u32 *qp_flags_en, struct ib_pd *pd,
1423 				    u32 *out)
1424 {
1425 	u8 lb_flag = 0;
1426 	u32 *in;
1427 	void *tirc;
1428 	int inlen;
1429 	int err;
1430 
1431 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1432 	in = kvzalloc(inlen, GFP_KERNEL);
1433 	if (!in)
1434 		return -ENOMEM;
1435 
1436 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1437 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1438 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1439 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1440 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1441 	if (*qp_flags_en & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1442 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1443 
1444 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1445 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1446 
1447 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1448 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1449 
1450 	if (dev->is_rep) {
1451 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1452 		*qp_flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1453 	}
1454 
1455 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1456 	MLX5_SET(create_tir_in, in, opcode, MLX5_CMD_OP_CREATE_TIR);
1457 	err = mlx5_cmd_exec_inout(dev->mdev, create_tir, in, out);
1458 	rq->tirn = MLX5_GET(create_tir_out, out, tirn);
1459 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1460 		err = mlx5_ib_enable_lb(dev, false, true);
1461 
1462 		if (err)
1463 			destroy_raw_packet_qp_tir(dev, rq, 0, pd);
1464 	}
1465 	kvfree(in);
1466 
1467 	return err;
1468 }
1469 
1470 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1471 				u32 *in, size_t inlen, struct ib_pd *pd,
1472 				struct ib_udata *udata,
1473 				struct mlx5_ib_create_qp_resp *resp,
1474 				struct ib_qp_init_attr *init_attr)
1475 {
1476 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1477 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1478 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1479 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1480 		udata, struct mlx5_ib_ucontext, ibucontext);
1481 	int err;
1482 	u32 tdn = mucontext->tdn;
1483 	u16 uid = to_mpd(pd)->uid;
1484 	u32 out[MLX5_ST_SZ_DW(create_tir_out)] = {};
1485 
1486 	if (!qp->sq.wqe_cnt && !qp->rq.wqe_cnt)
1487 		return -EINVAL;
1488 	if (qp->sq.wqe_cnt) {
1489 		err = create_raw_packet_qp_tis(dev, qp, sq, tdn, pd);
1490 		if (err)
1491 			return err;
1492 
1493 		err = create_raw_packet_qp_sq(dev, udata, sq, in, pd,
1494 					      to_mcq(init_attr->send_cq));
1495 		if (err)
1496 			goto err_destroy_tis;
1497 
1498 		if (uid) {
1499 			resp->tisn = sq->tisn;
1500 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TISN;
1501 			resp->sqn = sq->base.mqp.qpn;
1502 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_SQN;
1503 		}
1504 
1505 		sq->base.container_mibqp = qp;
1506 		sq->base.mqp.event = mlx5_ib_qp_event;
1507 	}
1508 
1509 	if (qp->rq.wqe_cnt) {
1510 		rq->base.container_mibqp = qp;
1511 
1512 		if (qp->flags & IB_QP_CREATE_CVLAN_STRIPPING)
1513 			rq->flags |= MLX5_IB_RQ_CVLAN_STRIPPING;
1514 		if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING)
1515 			rq->flags |= MLX5_IB_RQ_PCI_WRITE_END_PADDING;
1516 		err = create_raw_packet_qp_rq(dev, rq, in, pd,
1517 					      to_mcq(init_attr->recv_cq));
1518 		if (err)
1519 			goto err_destroy_sq;
1520 
1521 		err = create_raw_packet_qp_tir(dev, rq, tdn, &qp->flags_en, pd,
1522 					       out);
1523 		if (err)
1524 			goto err_destroy_rq;
1525 
1526 		if (uid) {
1527 			resp->rqn = rq->base.mqp.qpn;
1528 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_RQN;
1529 			resp->tirn = rq->tirn;
1530 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1531 			if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
1532 			    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2)) {
1533 				resp->tir_icm_addr = MLX5_GET(
1534 					create_tir_out, out, icm_address_31_0);
1535 				resp->tir_icm_addr |=
1536 					(u64)MLX5_GET(create_tir_out, out,
1537 						      icm_address_39_32)
1538 					<< 32;
1539 				resp->tir_icm_addr |=
1540 					(u64)MLX5_GET(create_tir_out, out,
1541 						      icm_address_63_40)
1542 					<< 40;
1543 				resp->comp_mask |=
1544 					MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1545 			}
1546 		}
1547 	}
1548 
1549 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1550 						     rq->base.mqp.qpn;
1551 	return 0;
1552 
1553 err_destroy_rq:
1554 	destroy_raw_packet_qp_rq(dev, rq);
1555 err_destroy_sq:
1556 	if (!qp->sq.wqe_cnt)
1557 		return err;
1558 	destroy_raw_packet_qp_sq(dev, sq);
1559 err_destroy_tis:
1560 	destroy_raw_packet_qp_tis(dev, sq, pd);
1561 
1562 	return err;
1563 }
1564 
1565 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1566 				  struct mlx5_ib_qp *qp)
1567 {
1568 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1569 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1570 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1571 
1572 	if (qp->rq.wqe_cnt) {
1573 		destroy_raw_packet_qp_tir(dev, rq, qp->flags_en, qp->ibqp.pd);
1574 		destroy_raw_packet_qp_rq(dev, rq);
1575 	}
1576 
1577 	if (qp->sq.wqe_cnt) {
1578 		destroy_raw_packet_qp_sq(dev, sq);
1579 		destroy_raw_packet_qp_tis(dev, sq, qp->ibqp.pd);
1580 	}
1581 }
1582 
1583 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1584 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1585 {
1586 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1587 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1588 
1589 	sq->sq = &qp->sq;
1590 	rq->rq = &qp->rq;
1591 	sq->doorbell = &qp->db;
1592 	rq->doorbell = &qp->db;
1593 }
1594 
1595 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1596 {
1597 	if (qp->flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1598 			    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1599 		mlx5_ib_disable_lb(dev, false, true);
1600 	mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1601 			     to_mpd(qp->ibqp.pd)->uid);
1602 }
1603 
1604 struct mlx5_create_qp_params {
1605 	struct ib_udata *udata;
1606 	size_t inlen;
1607 	size_t outlen;
1608 	size_t ucmd_size;
1609 	void *ucmd;
1610 	u8 is_rss_raw : 1;
1611 	struct ib_qp_init_attr *attr;
1612 	u32 uidx;
1613 	struct mlx5_ib_create_qp_resp resp;
1614 };
1615 
1616 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1617 				 struct mlx5_ib_qp *qp,
1618 				 struct mlx5_create_qp_params *params)
1619 {
1620 	struct ib_qp_init_attr *init_attr = params->attr;
1621 	struct mlx5_ib_create_qp_rss *ucmd = params->ucmd;
1622 	struct ib_udata *udata = params->udata;
1623 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1624 		udata, struct mlx5_ib_ucontext, ibucontext);
1625 	int inlen;
1626 	int outlen;
1627 	int err;
1628 	u32 *in;
1629 	u32 *out;
1630 	void *tirc;
1631 	void *hfso;
1632 	u32 selected_fields = 0;
1633 	u32 outer_l4;
1634 	u32 tdn = mucontext->tdn;
1635 	u8 lb_flag = 0;
1636 
1637 	if (ucmd->comp_mask) {
1638 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1639 		return -EOPNOTSUPP;
1640 	}
1641 
1642 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_INNER &&
1643 	    !(ucmd->flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)) {
1644 		mlx5_ib_dbg(dev, "Tunnel offloads must be set for inner RSS\n");
1645 		return -EOPNOTSUPP;
1646 	}
1647 
1648 	if (dev->is_rep)
1649 		qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1650 
1651 	if (qp->flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1652 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1653 
1654 	if (qp->flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1655 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1656 
1657 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1658 	outlen = MLX5_ST_SZ_BYTES(create_tir_out);
1659 	in = kvzalloc(inlen + outlen, GFP_KERNEL);
1660 	if (!in)
1661 		return -ENOMEM;
1662 
1663 	out = in + MLX5_ST_SZ_DW(create_tir_in);
1664 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1665 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1666 	MLX5_SET(tirc, tirc, disp_type,
1667 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1668 	MLX5_SET(tirc, tirc, indirect_table,
1669 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1670 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1671 
1672 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1673 
1674 	if (ucmd->flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1675 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1676 
1677 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1678 
1679 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_INNER)
1680 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner);
1681 	else
1682 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1683 
1684 	switch (ucmd->rx_hash_function) {
1685 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1686 	{
1687 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1688 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1689 
1690 		if (len != ucmd->rx_key_len) {
1691 			err = -EINVAL;
1692 			goto err;
1693 		}
1694 
1695 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1696 		memcpy(rss_key, ucmd->rx_hash_key, len);
1697 		break;
1698 	}
1699 	default:
1700 		err = -EOPNOTSUPP;
1701 		goto err;
1702 	}
1703 
1704 	if (!ucmd->rx_hash_fields_mask) {
1705 		/* special case when this TIR serves as steering entry without hashing */
1706 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1707 			goto create_tir;
1708 		err = -EINVAL;
1709 		goto err;
1710 	}
1711 
1712 	if (((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1713 	     (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1714 	     ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1715 	     (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1716 		err = -EINVAL;
1717 		goto err;
1718 	}
1719 
1720 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1721 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1722 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1723 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1724 			 MLX5_L3_PROT_TYPE_IPV4);
1725 	else if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1726 		 (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1727 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1728 			 MLX5_L3_PROT_TYPE_IPV6);
1729 
1730 	outer_l4 = ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1731 		    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1732 			   << 0 |
1733 		   ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1734 		    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1735 			   << 1 |
1736 		   (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI) << 2;
1737 
1738 	/* Check that only one l4 protocol is set */
1739 	if (outer_l4 & (outer_l4 - 1)) {
1740 		err = -EINVAL;
1741 		goto err;
1742 	}
1743 
1744 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1745 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1746 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1747 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1748 			 MLX5_L4_PROT_TYPE_TCP);
1749 	else if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1750 		 (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1751 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1752 			 MLX5_L4_PROT_TYPE_UDP);
1753 
1754 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1755 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1756 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1757 
1758 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1759 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1760 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1761 
1762 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1763 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1764 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1765 
1766 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1767 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1768 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1769 
1770 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI)
1771 		selected_fields |= MLX5_HASH_FIELD_SEL_IPSEC_SPI;
1772 
1773 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1774 
1775 create_tir:
1776 	MLX5_SET(create_tir_in, in, opcode, MLX5_CMD_OP_CREATE_TIR);
1777 	err = mlx5_cmd_exec_inout(dev->mdev, create_tir, in, out);
1778 
1779 	qp->rss_qp.tirn = MLX5_GET(create_tir_out, out, tirn);
1780 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1781 		err = mlx5_ib_enable_lb(dev, false, true);
1782 
1783 		if (err)
1784 			mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1785 					     to_mpd(pd)->uid);
1786 	}
1787 
1788 	if (err)
1789 		goto err;
1790 
1791 	if (mucontext->devx_uid) {
1792 		params->resp.comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1793 		params->resp.tirn = qp->rss_qp.tirn;
1794 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
1795 		    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2)) {
1796 			params->resp.tir_icm_addr =
1797 				MLX5_GET(create_tir_out, out, icm_address_31_0);
1798 			params->resp.tir_icm_addr |=
1799 				(u64)MLX5_GET(create_tir_out, out,
1800 					      icm_address_39_32)
1801 				<< 32;
1802 			params->resp.tir_icm_addr |=
1803 				(u64)MLX5_GET(create_tir_out, out,
1804 					      icm_address_63_40)
1805 				<< 40;
1806 			params->resp.comp_mask |=
1807 				MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1808 		}
1809 	}
1810 
1811 	kvfree(in);
1812 	/* qpn is reserved for that QP */
1813 	qp->trans_qp.base.mqp.qpn = 0;
1814 	qp->is_rss = true;
1815 	return 0;
1816 
1817 err:
1818 	kvfree(in);
1819 	return err;
1820 }
1821 
1822 static void configure_requester_scat_cqe(struct mlx5_ib_dev *dev,
1823 					 struct mlx5_ib_qp *qp,
1824 					 struct ib_qp_init_attr *init_attr,
1825 					 void *qpc)
1826 {
1827 	int scqe_sz;
1828 	bool allow_scat_cqe = false;
1829 
1830 	allow_scat_cqe = qp->flags_en & MLX5_QP_FLAG_ALLOW_SCATTER_CQE;
1831 
1832 	if (!allow_scat_cqe && init_attr->sq_sig_type != IB_SIGNAL_ALL_WR)
1833 		return;
1834 
1835 	scqe_sz = mlx5_ib_get_cqe_size(init_attr->send_cq);
1836 	if (scqe_sz == 128) {
1837 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1838 		return;
1839 	}
1840 
1841 	if (init_attr->qp_type != MLX5_IB_QPT_DCI ||
1842 	    MLX5_CAP_GEN(dev->mdev, dc_req_scat_data_cqe))
1843 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1844 }
1845 
1846 static int atomic_size_to_mode(int size_mask)
1847 {
1848 	/* driver does not support atomic_size > 256B
1849 	 * and does not know how to translate bigger sizes
1850 	 */
1851 	int supported_size_mask = size_mask & 0x1ff;
1852 	int log_max_size;
1853 
1854 	if (!supported_size_mask)
1855 		return -EOPNOTSUPP;
1856 
1857 	log_max_size = __fls(supported_size_mask);
1858 
1859 	if (log_max_size > 3)
1860 		return log_max_size;
1861 
1862 	return MLX5_ATOMIC_MODE_8B;
1863 }
1864 
1865 static int get_atomic_mode(struct mlx5_ib_dev *dev,
1866 			   enum ib_qp_type qp_type)
1867 {
1868 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
1869 	u8 atomic = MLX5_CAP_GEN(dev->mdev, atomic);
1870 	int atomic_mode = -EOPNOTSUPP;
1871 	int atomic_size_mask;
1872 
1873 	if (!atomic)
1874 		return -EOPNOTSUPP;
1875 
1876 	if (qp_type == MLX5_IB_QPT_DCT)
1877 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
1878 	else
1879 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
1880 
1881 	if ((atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_CMP_SWAP) ||
1882 	    (atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_FETCH_ADD))
1883 		atomic_mode = atomic_size_to_mode(atomic_size_mask);
1884 
1885 	if (atomic_mode <= 0 &&
1886 	    (atomic_operations & MLX5_ATOMIC_OPS_CMP_SWAP &&
1887 	     atomic_operations & MLX5_ATOMIC_OPS_FETCH_ADD))
1888 		atomic_mode = MLX5_ATOMIC_MODE_IB_COMP;
1889 
1890 	return atomic_mode;
1891 }
1892 
1893 static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1894 			     struct mlx5_create_qp_params *params)
1895 {
1896 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
1897 	struct ib_qp_init_attr *attr = params->attr;
1898 	u32 uidx = params->uidx;
1899 	struct mlx5_ib_resources *devr = &dev->devr;
1900 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
1901 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1902 	struct mlx5_core_dev *mdev = dev->mdev;
1903 	struct mlx5_ib_qp_base *base;
1904 	unsigned long flags;
1905 	void *qpc;
1906 	u32 *in;
1907 	int err;
1908 
1909 	if (attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1910 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1911 
1912 	in = kvzalloc(inlen, GFP_KERNEL);
1913 	if (!in)
1914 		return -ENOMEM;
1915 
1916 	if (MLX5_CAP_GEN(mdev, ece_support) && ucmd)
1917 		MLX5_SET(create_qp_in, in, ece, ucmd->ece_options);
1918 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1919 
1920 	MLX5_SET(qpc, qpc, st, MLX5_QP_ST_XRC);
1921 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
1922 	MLX5_SET(qpc, qpc, pd, to_mpd(devr->p0)->pdn);
1923 
1924 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1925 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
1926 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
1927 		MLX5_SET(qpc, qpc, cd_master, 1);
1928 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
1929 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
1930 	if (qp->flags & IB_QP_CREATE_MANAGED_RECV)
1931 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
1932 
1933 	MLX5_SET(qpc, qpc, rq_type, MLX5_SRQ_RQ);
1934 	MLX5_SET(qpc, qpc, no_sq, 1);
1935 	MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1936 	MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
1937 	MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
1938 	MLX5_SET(qpc, qpc, xrcd, to_mxrcd(attr->xrcd)->xrcdn);
1939 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
1940 
1941 	/* 0xffffff means we ask to work with cqe version 0 */
1942 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
1943 		MLX5_SET(qpc, qpc, user_index, uidx);
1944 
1945 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING) {
1946 		MLX5_SET(qpc, qpc, end_padding_mode,
1947 			 MLX5_WQ_END_PAD_MODE_ALIGN);
1948 		/* Special case to clean flag */
1949 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
1950 	}
1951 
1952 	base = &qp->trans_qp.base;
1953 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
1954 	kvfree(in);
1955 	if (err)
1956 		return err;
1957 
1958 	base->container_mibqp = qp;
1959 	base->mqp.event = mlx5_ib_qp_event;
1960 	if (MLX5_CAP_GEN(mdev, ece_support))
1961 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
1962 
1963 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1964 	list_add_tail(&qp->qps_list, &dev->qp_list);
1965 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1966 
1967 	qp->trans_qp.xrcdn = to_mxrcd(attr->xrcd)->xrcdn;
1968 	return 0;
1969 }
1970 
1971 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1972 			  struct mlx5_ib_qp *qp,
1973 			  struct mlx5_create_qp_params *params)
1974 {
1975 	struct ib_qp_init_attr *init_attr = params->attr;
1976 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
1977 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
1978 	struct ib_udata *udata = params->udata;
1979 	u32 uidx = params->uidx;
1980 	struct mlx5_ib_resources *devr = &dev->devr;
1981 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1982 	struct mlx5_core_dev *mdev = dev->mdev;
1983 	struct mlx5_ib_cq *send_cq;
1984 	struct mlx5_ib_cq *recv_cq;
1985 	unsigned long flags;
1986 	struct mlx5_ib_qp_base *base;
1987 	int ts_format;
1988 	int mlx5_st;
1989 	void *qpc;
1990 	u32 *in;
1991 	int err;
1992 
1993 	spin_lock_init(&qp->sq.lock);
1994 	spin_lock_init(&qp->rq.lock);
1995 
1996 	mlx5_st = to_mlx5_st(qp->type);
1997 	if (mlx5_st < 0)
1998 		return -EINVAL;
1999 
2000 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2001 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2002 
2003 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
2004 		qp->underlay_qpn = init_attr->source_qpn;
2005 
2006 	base = (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2007 		qp->flags & IB_QP_CREATE_SOURCE_QPN) ?
2008 	       &qp->raw_packet_qp.rq.base :
2009 	       &qp->trans_qp.base;
2010 
2011 	qp->has_rq = qp_has_rq(init_attr);
2012 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq, qp, ucmd);
2013 	if (err) {
2014 		mlx5_ib_dbg(dev, "err %d\n", err);
2015 		return err;
2016 	}
2017 
2018 	if (ucmd->rq_wqe_shift != qp->rq.wqe_shift ||
2019 	    ucmd->rq_wqe_count != qp->rq.wqe_cnt)
2020 		return -EINVAL;
2021 
2022 	if (ucmd->sq_wqe_count > (1 << MLX5_CAP_GEN(mdev, log_max_qp_sz)))
2023 		return -EINVAL;
2024 
2025 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2026 		ts_format = get_qp_ts_format(dev, to_mcq(init_attr->send_cq),
2027 					     to_mcq(init_attr->recv_cq));
2028 		if (ts_format < 0)
2029 			return ts_format;
2030 	}
2031 
2032 	err = _create_user_qp(dev, pd, qp, udata, init_attr, &in, &params->resp,
2033 			      &inlen, base, ucmd);
2034 	if (err)
2035 		return err;
2036 
2037 	if (is_sqp(init_attr->qp_type))
2038 		qp->port = init_attr->port_num;
2039 
2040 	if (MLX5_CAP_GEN(mdev, ece_support))
2041 		MLX5_SET(create_qp_in, in, ece, ucmd->ece_options);
2042 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2043 
2044 	MLX5_SET(qpc, qpc, st, mlx5_st);
2045 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2046 	MLX5_SET(qpc, qpc, pd, to_mpd(pd)->pdn);
2047 
2048 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE)
2049 		MLX5_SET(qpc, qpc, wq_signature, 1);
2050 
2051 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
2052 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2053 
2054 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
2055 		MLX5_SET(qpc, qpc, cd_master, 1);
2056 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
2057 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
2058 	if (qp->flags & IB_QP_CREATE_MANAGED_RECV)
2059 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
2060 	if (qp->flags_en & MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE)
2061 		MLX5_SET(qpc, qpc, req_e2e_credit_mode, 1);
2062 	if ((qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) &&
2063 	    (init_attr->qp_type == IB_QPT_RC ||
2064 	     init_attr->qp_type == IB_QPT_UC)) {
2065 		int rcqe_sz = mlx5_ib_get_cqe_size(init_attr->recv_cq);
2066 
2067 		MLX5_SET(qpc, qpc, cs_res,
2068 			 rcqe_sz == 128 ? MLX5_RES_SCAT_DATA64_CQE :
2069 					  MLX5_RES_SCAT_DATA32_CQE);
2070 	}
2071 	if ((qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) &&
2072 	    (qp->type == MLX5_IB_QPT_DCI || qp->type == IB_QPT_RC))
2073 		configure_requester_scat_cqe(dev, qp, init_attr, qpc);
2074 
2075 	if (qp->rq.wqe_cnt) {
2076 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2077 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2078 	}
2079 
2080 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
2081 		MLX5_SET(qpc, qpc, ts_format, ts_format);
2082 
2083 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
2084 
2085 	if (qp->sq.wqe_cnt) {
2086 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2087 	} else {
2088 		MLX5_SET(qpc, qpc, no_sq, 1);
2089 		if (init_attr->srq &&
2090 		    init_attr->srq->srq_type == IB_SRQT_TM)
2091 			MLX5_SET(qpc, qpc, offload_type,
2092 				 MLX5_QPC_OFFLOAD_TYPE_RNDV);
2093 	}
2094 
2095 	/* Set default resources */
2096 	switch (init_attr->qp_type) {
2097 	case IB_QPT_XRC_INI:
2098 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2099 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2100 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2101 		break;
2102 	default:
2103 		if (init_attr->srq) {
2104 			MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2105 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(init_attr->srq)->msrq.srqn);
2106 		} else {
2107 			MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2108 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s1)->msrq.srqn);
2109 		}
2110 	}
2111 
2112 	if (init_attr->send_cq)
2113 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
2114 
2115 	if (init_attr->recv_cq)
2116 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
2117 
2118 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2119 
2120 	/* 0xffffff means we ask to work with cqe version 0 */
2121 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2122 		MLX5_SET(qpc, qpc, user_index, uidx);
2123 
2124 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING &&
2125 	    init_attr->qp_type != IB_QPT_RAW_PACKET) {
2126 		MLX5_SET(qpc, qpc, end_padding_mode,
2127 			 MLX5_WQ_END_PAD_MODE_ALIGN);
2128 		/* Special case to clean flag */
2129 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
2130 	}
2131 
2132 	if (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2133 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
2134 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd->sq_buf_addr;
2135 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
2136 		err = create_raw_packet_qp(dev, qp, in, inlen, pd, udata,
2137 					   &params->resp, init_attr);
2138 	} else
2139 		err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2140 
2141 	kvfree(in);
2142 	if (err)
2143 		goto err_create;
2144 
2145 	base->container_mibqp = qp;
2146 	base->mqp.event = mlx5_ib_qp_event;
2147 	if (MLX5_CAP_GEN(mdev, ece_support))
2148 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
2149 
2150 	get_cqs(qp->type, init_attr->send_cq, init_attr->recv_cq,
2151 		&send_cq, &recv_cq);
2152 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2153 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2154 	/* Maintain device to QPs access, needed for further handling via reset
2155 	 * flow
2156 	 */
2157 	list_add_tail(&qp->qps_list, &dev->qp_list);
2158 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2159 	 */
2160 	if (send_cq)
2161 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2162 	if (recv_cq)
2163 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2164 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2165 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2166 
2167 	return 0;
2168 
2169 err_create:
2170 	destroy_qp(dev, qp, base, udata);
2171 	return err;
2172 }
2173 
2174 static int create_kernel_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2175 			    struct mlx5_ib_qp *qp,
2176 			    struct mlx5_create_qp_params *params)
2177 {
2178 	struct ib_qp_init_attr *attr = params->attr;
2179 	u32 uidx = params->uidx;
2180 	struct mlx5_ib_resources *devr = &dev->devr;
2181 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
2182 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
2183 	struct mlx5_core_dev *mdev = dev->mdev;
2184 	struct mlx5_ib_cq *send_cq;
2185 	struct mlx5_ib_cq *recv_cq;
2186 	unsigned long flags;
2187 	struct mlx5_ib_qp_base *base;
2188 	int mlx5_st;
2189 	void *qpc;
2190 	u32 *in;
2191 	int err;
2192 
2193 	spin_lock_init(&qp->sq.lock);
2194 	spin_lock_init(&qp->rq.lock);
2195 
2196 	mlx5_st = to_mlx5_st(qp->type);
2197 	if (mlx5_st < 0)
2198 		return -EINVAL;
2199 
2200 	if (attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2201 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2202 
2203 	base = &qp->trans_qp.base;
2204 
2205 	qp->has_rq = qp_has_rq(attr);
2206 	err = set_rq_size(dev, &attr->cap, qp->has_rq, qp, NULL);
2207 	if (err) {
2208 		mlx5_ib_dbg(dev, "err %d\n", err);
2209 		return err;
2210 	}
2211 
2212 	err = _create_kernel_qp(dev, attr, qp, &in, &inlen, base);
2213 	if (err)
2214 		return err;
2215 
2216 	if (is_sqp(attr->qp_type))
2217 		qp->port = attr->port_num;
2218 
2219 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2220 
2221 	MLX5_SET(qpc, qpc, st, mlx5_st);
2222 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2223 
2224 	if (attr->qp_type != MLX5_IB_QPT_REG_UMR)
2225 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
2226 	else
2227 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
2228 
2229 
2230 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
2231 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2232 
2233 	if (qp->rq.wqe_cnt) {
2234 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2235 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2236 	}
2237 
2238 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, attr));
2239 
2240 	if (qp->sq.wqe_cnt)
2241 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2242 	else
2243 		MLX5_SET(qpc, qpc, no_sq, 1);
2244 
2245 	if (attr->srq) {
2246 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2247 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2248 			 to_msrq(attr->srq)->msrq.srqn);
2249 	} else {
2250 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2251 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2252 			 to_msrq(devr->s1)->msrq.srqn);
2253 	}
2254 
2255 	if (attr->send_cq)
2256 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(attr->send_cq)->mcq.cqn);
2257 
2258 	if (attr->recv_cq)
2259 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(attr->recv_cq)->mcq.cqn);
2260 
2261 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2262 
2263 	/* 0xffffff means we ask to work with cqe version 0 */
2264 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2265 		MLX5_SET(qpc, qpc, user_index, uidx);
2266 
2267 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
2268 	if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO)
2269 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
2270 
2271 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2272 	kvfree(in);
2273 	if (err)
2274 		goto err_create;
2275 
2276 	base->container_mibqp = qp;
2277 	base->mqp.event = mlx5_ib_qp_event;
2278 
2279 	get_cqs(qp->type, attr->send_cq, attr->recv_cq,
2280 		&send_cq, &recv_cq);
2281 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2282 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2283 	/* Maintain device to QPs access, needed for further handling via reset
2284 	 * flow
2285 	 */
2286 	list_add_tail(&qp->qps_list, &dev->qp_list);
2287 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2288 	 */
2289 	if (send_cq)
2290 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2291 	if (recv_cq)
2292 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2293 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2294 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2295 
2296 	return 0;
2297 
2298 err_create:
2299 	destroy_qp(dev, qp, base, NULL);
2300 	return err;
2301 }
2302 
2303 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2304 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
2305 {
2306 	if (send_cq) {
2307 		if (recv_cq) {
2308 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2309 				spin_lock(&send_cq->lock);
2310 				spin_lock_nested(&recv_cq->lock,
2311 						 SINGLE_DEPTH_NESTING);
2312 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2313 				spin_lock(&send_cq->lock);
2314 				__acquire(&recv_cq->lock);
2315 			} else {
2316 				spin_lock(&recv_cq->lock);
2317 				spin_lock_nested(&send_cq->lock,
2318 						 SINGLE_DEPTH_NESTING);
2319 			}
2320 		} else {
2321 			spin_lock(&send_cq->lock);
2322 			__acquire(&recv_cq->lock);
2323 		}
2324 	} else if (recv_cq) {
2325 		spin_lock(&recv_cq->lock);
2326 		__acquire(&send_cq->lock);
2327 	} else {
2328 		__acquire(&send_cq->lock);
2329 		__acquire(&recv_cq->lock);
2330 	}
2331 }
2332 
2333 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2334 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
2335 {
2336 	if (send_cq) {
2337 		if (recv_cq) {
2338 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2339 				spin_unlock(&recv_cq->lock);
2340 				spin_unlock(&send_cq->lock);
2341 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2342 				__release(&recv_cq->lock);
2343 				spin_unlock(&send_cq->lock);
2344 			} else {
2345 				spin_unlock(&send_cq->lock);
2346 				spin_unlock(&recv_cq->lock);
2347 			}
2348 		} else {
2349 			__release(&recv_cq->lock);
2350 			spin_unlock(&send_cq->lock);
2351 		}
2352 	} else if (recv_cq) {
2353 		__release(&send_cq->lock);
2354 		spin_unlock(&recv_cq->lock);
2355 	} else {
2356 		__release(&recv_cq->lock);
2357 		__release(&send_cq->lock);
2358 	}
2359 }
2360 
2361 static void get_cqs(enum ib_qp_type qp_type,
2362 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
2363 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
2364 {
2365 	switch (qp_type) {
2366 	case IB_QPT_XRC_TGT:
2367 		*send_cq = NULL;
2368 		*recv_cq = NULL;
2369 		break;
2370 	case MLX5_IB_QPT_REG_UMR:
2371 	case IB_QPT_XRC_INI:
2372 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2373 		*recv_cq = NULL;
2374 		break;
2375 
2376 	case IB_QPT_SMI:
2377 	case MLX5_IB_QPT_HW_GSI:
2378 	case IB_QPT_RC:
2379 	case IB_QPT_UC:
2380 	case IB_QPT_UD:
2381 	case IB_QPT_RAW_PACKET:
2382 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2383 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
2384 		break;
2385 	default:
2386 		*send_cq = NULL;
2387 		*recv_cq = NULL;
2388 		break;
2389 	}
2390 }
2391 
2392 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2393 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2394 				u8 lag_tx_affinity);
2395 
2396 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2397 			      struct ib_udata *udata)
2398 {
2399 	struct mlx5_ib_cq *send_cq, *recv_cq;
2400 	struct mlx5_ib_qp_base *base;
2401 	unsigned long flags;
2402 	int err;
2403 
2404 	if (qp->is_rss) {
2405 		destroy_rss_raw_qp_tir(dev, qp);
2406 		return;
2407 	}
2408 
2409 	base = (qp->type == IB_QPT_RAW_PACKET ||
2410 		qp->flags & IB_QP_CREATE_SOURCE_QPN) ?
2411 		       &qp->raw_packet_qp.rq.base :
2412 		       &qp->trans_qp.base;
2413 
2414 	if (qp->state != IB_QPS_RESET) {
2415 		if (qp->type != IB_QPT_RAW_PACKET &&
2416 		    !(qp->flags & IB_QP_CREATE_SOURCE_QPN)) {
2417 			err = mlx5_core_qp_modify(dev, MLX5_CMD_OP_2RST_QP, 0,
2418 						  NULL, &base->mqp, NULL);
2419 		} else {
2420 			struct mlx5_modify_raw_qp_param raw_qp_param = {
2421 				.operation = MLX5_CMD_OP_2RST_QP
2422 			};
2423 
2424 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2425 		}
2426 		if (err)
2427 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
2428 				     base->mqp.qpn);
2429 	}
2430 
2431 	get_cqs(qp->type, qp->ibqp.send_cq, qp->ibqp.recv_cq, &send_cq,
2432 		&recv_cq);
2433 
2434 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2435 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2436 	/* del from lists under both locks above to protect reset flow paths */
2437 	list_del(&qp->qps_list);
2438 	if (send_cq)
2439 		list_del(&qp->cq_send_list);
2440 
2441 	if (recv_cq)
2442 		list_del(&qp->cq_recv_list);
2443 
2444 	if (!udata) {
2445 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2446 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
2447 		if (send_cq != recv_cq)
2448 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
2449 					   NULL);
2450 	}
2451 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2452 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2453 
2454 	if (qp->type == IB_QPT_RAW_PACKET ||
2455 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
2456 		destroy_raw_packet_qp(dev, qp);
2457 	} else {
2458 		err = mlx5_core_destroy_qp(dev, &base->mqp);
2459 		if (err)
2460 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
2461 				     base->mqp.qpn);
2462 	}
2463 
2464 	destroy_qp(dev, qp, base, udata);
2465 }
2466 
2467 static int create_dct(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2468 		      struct mlx5_ib_qp *qp,
2469 		      struct mlx5_create_qp_params *params)
2470 {
2471 	struct ib_qp_init_attr *attr = params->attr;
2472 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
2473 	u32 uidx = params->uidx;
2474 	void *dctc;
2475 
2476 	if (mlx5_lag_is_active(dev->mdev) && !MLX5_CAP_GEN(dev->mdev, lag_dct))
2477 		return -EOPNOTSUPP;
2478 
2479 	qp->dct.in = kzalloc(MLX5_ST_SZ_BYTES(create_dct_in), GFP_KERNEL);
2480 	if (!qp->dct.in)
2481 		return -ENOMEM;
2482 
2483 	MLX5_SET(create_dct_in, qp->dct.in, uid, to_mpd(pd)->uid);
2484 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
2485 	MLX5_SET(dctc, dctc, pd, to_mpd(pd)->pdn);
2486 	MLX5_SET(dctc, dctc, srqn_xrqn, to_msrq(attr->srq)->msrq.srqn);
2487 	MLX5_SET(dctc, dctc, cqn, to_mcq(attr->recv_cq)->mcq.cqn);
2488 	MLX5_SET64(dctc, dctc, dc_access_key, ucmd->access_key);
2489 	MLX5_SET(dctc, dctc, user_index, uidx);
2490 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
2491 		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
2492 
2493 	if (qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) {
2494 		int rcqe_sz = mlx5_ib_get_cqe_size(attr->recv_cq);
2495 
2496 		if (rcqe_sz == 128)
2497 			MLX5_SET(dctc, dctc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
2498 	}
2499 
2500 	qp->state = IB_QPS_RESET;
2501 	rdma_restrack_no_track(&qp->ibqp.res);
2502 	return 0;
2503 }
2504 
2505 static int check_qp_type(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
2506 			 enum ib_qp_type *type)
2507 {
2508 	if (attr->qp_type == IB_QPT_DRIVER && !MLX5_CAP_GEN(dev->mdev, dct))
2509 		goto out;
2510 
2511 	switch (attr->qp_type) {
2512 	case IB_QPT_XRC_TGT:
2513 	case IB_QPT_XRC_INI:
2514 		if (!MLX5_CAP_GEN(dev->mdev, xrc))
2515 			goto out;
2516 		fallthrough;
2517 	case IB_QPT_RC:
2518 	case IB_QPT_UC:
2519 	case IB_QPT_SMI:
2520 	case MLX5_IB_QPT_HW_GSI:
2521 	case IB_QPT_DRIVER:
2522 	case IB_QPT_GSI:
2523 	case IB_QPT_RAW_PACKET:
2524 	case IB_QPT_UD:
2525 	case MLX5_IB_QPT_REG_UMR:
2526 		break;
2527 	default:
2528 		goto out;
2529 	}
2530 
2531 	*type = attr->qp_type;
2532 	return 0;
2533 
2534 out:
2535 	mlx5_ib_dbg(dev, "Unsupported QP type %d\n", attr->qp_type);
2536 	return -EOPNOTSUPP;
2537 }
2538 
2539 static int check_valid_flow(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2540 			    struct ib_qp_init_attr *attr,
2541 			    struct ib_udata *udata)
2542 {
2543 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2544 		udata, struct mlx5_ib_ucontext, ibucontext);
2545 
2546 	if (!udata) {
2547 		/* Kernel create_qp callers */
2548 		if (attr->rwq_ind_tbl)
2549 			return -EOPNOTSUPP;
2550 
2551 		switch (attr->qp_type) {
2552 		case IB_QPT_RAW_PACKET:
2553 		case IB_QPT_DRIVER:
2554 			return -EOPNOTSUPP;
2555 		default:
2556 			return 0;
2557 		}
2558 	}
2559 
2560 	/* Userspace create_qp callers */
2561 	if (attr->qp_type == IB_QPT_RAW_PACKET && !ucontext->cqe_version) {
2562 		mlx5_ib_dbg(dev,
2563 			"Raw Packet QP is only supported for CQE version > 0\n");
2564 		return -EINVAL;
2565 	}
2566 
2567 	if (attr->qp_type != IB_QPT_RAW_PACKET && attr->rwq_ind_tbl) {
2568 		mlx5_ib_dbg(dev,
2569 			    "Wrong QP type %d for the RWQ indirect table\n",
2570 			    attr->qp_type);
2571 		return -EINVAL;
2572 	}
2573 
2574 	/*
2575 	 * We don't need to see this warning, it means that kernel code
2576 	 * missing ib_pd. Placed here to catch developer's mistakes.
2577 	 */
2578 	WARN_ONCE(!pd && attr->qp_type != IB_QPT_XRC_TGT,
2579 		  "There is a missing PD pointer assignment\n");
2580 	return 0;
2581 }
2582 
2583 static void process_vendor_flag(struct mlx5_ib_dev *dev, int *flags, int flag,
2584 				bool cond, struct mlx5_ib_qp *qp)
2585 {
2586 	if (!(*flags & flag))
2587 		return;
2588 
2589 	if (cond) {
2590 		qp->flags_en |= flag;
2591 		*flags &= ~flag;
2592 		return;
2593 	}
2594 
2595 	switch (flag) {
2596 	case MLX5_QP_FLAG_SCATTER_CQE:
2597 	case MLX5_QP_FLAG_ALLOW_SCATTER_CQE:
2598 		/*
2599 			 * We don't return error if these flags were provided,
2600 			 * and mlx5 doesn't have right capability.
2601 			 */
2602 		*flags &= ~(MLX5_QP_FLAG_SCATTER_CQE |
2603 			    MLX5_QP_FLAG_ALLOW_SCATTER_CQE);
2604 		return;
2605 	default:
2606 		break;
2607 	}
2608 	mlx5_ib_dbg(dev, "Vendor create QP flag 0x%X is not supported\n", flag);
2609 }
2610 
2611 static int process_vendor_flags(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2612 				void *ucmd, struct ib_qp_init_attr *attr)
2613 {
2614 	struct mlx5_core_dev *mdev = dev->mdev;
2615 	bool cond;
2616 	int flags;
2617 
2618 	if (attr->rwq_ind_tbl)
2619 		flags = ((struct mlx5_ib_create_qp_rss *)ucmd)->flags;
2620 	else
2621 		flags = ((struct mlx5_ib_create_qp *)ucmd)->flags;
2622 
2623 	switch (flags & (MLX5_QP_FLAG_TYPE_DCT | MLX5_QP_FLAG_TYPE_DCI)) {
2624 	case MLX5_QP_FLAG_TYPE_DCI:
2625 		qp->type = MLX5_IB_QPT_DCI;
2626 		break;
2627 	case MLX5_QP_FLAG_TYPE_DCT:
2628 		qp->type = MLX5_IB_QPT_DCT;
2629 		break;
2630 	default:
2631 		if (qp->type != IB_QPT_DRIVER)
2632 			break;
2633 		/*
2634 		 * It is IB_QPT_DRIVER and or no subtype or
2635 		 * wrong subtype were provided.
2636 		 */
2637 		return -EINVAL;
2638 	}
2639 
2640 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TYPE_DCI, true, qp);
2641 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TYPE_DCT, true, qp);
2642 
2643 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_SIGNATURE, true, qp);
2644 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_SCATTER_CQE,
2645 			    MLX5_CAP_GEN(mdev, sctr_data_cqe), qp);
2646 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_ALLOW_SCATTER_CQE,
2647 			    MLX5_CAP_GEN(mdev, sctr_data_cqe), qp);
2648 
2649 	if (qp->type == IB_QPT_RAW_PACKET) {
2650 		cond = MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan) ||
2651 		       MLX5_CAP_ETH(mdev, tunnel_stateless_gre) ||
2652 		       MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx);
2653 		process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TUNNEL_OFFLOADS,
2654 				    cond, qp);
2655 		process_vendor_flag(dev, &flags,
2656 				    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC, true,
2657 				    qp);
2658 		process_vendor_flag(dev, &flags,
2659 				    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC, true,
2660 				    qp);
2661 	}
2662 
2663 	if (qp->type == IB_QPT_RC)
2664 		process_vendor_flag(dev, &flags,
2665 				    MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE,
2666 				    MLX5_CAP_GEN(mdev, qp_packet_based), qp);
2667 
2668 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_BFREG_INDEX, true, qp);
2669 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_UAR_PAGE_INDEX, true, qp);
2670 
2671 	cond = qp->flags_en & ~(MLX5_QP_FLAG_TUNNEL_OFFLOADS |
2672 				MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
2673 				MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC);
2674 	if (attr->rwq_ind_tbl && cond) {
2675 		mlx5_ib_dbg(dev, "RSS RAW QP has unsupported flags 0x%X\n",
2676 			    cond);
2677 		return -EINVAL;
2678 	}
2679 
2680 	if (flags)
2681 		mlx5_ib_dbg(dev, "udata has unsupported flags 0x%X\n", flags);
2682 
2683 	return (flags) ? -EINVAL : 0;
2684 	}
2685 
2686 static void process_create_flag(struct mlx5_ib_dev *dev, int *flags, int flag,
2687 				bool cond, struct mlx5_ib_qp *qp)
2688 {
2689 	if (!(*flags & flag))
2690 		return;
2691 
2692 	if (cond) {
2693 		qp->flags |= flag;
2694 		*flags &= ~flag;
2695 		return;
2696 	}
2697 
2698 	if (flag == MLX5_IB_QP_CREATE_WC_TEST) {
2699 		/*
2700 		 * Special case, if condition didn't meet, it won't be error,
2701 		 * just different in-kernel flow.
2702 		 */
2703 		*flags &= ~MLX5_IB_QP_CREATE_WC_TEST;
2704 		return;
2705 	}
2706 	mlx5_ib_dbg(dev, "Verbs create QP flag 0x%X is not supported\n", flag);
2707 }
2708 
2709 static int process_create_flags(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2710 				struct ib_qp_init_attr *attr)
2711 {
2712 	enum ib_qp_type qp_type = qp->type;
2713 	struct mlx5_core_dev *mdev = dev->mdev;
2714 	int create_flags = attr->create_flags;
2715 	bool cond;
2716 
2717 	if (qp_type == MLX5_IB_QPT_DCT)
2718 		return (create_flags) ? -EINVAL : 0;
2719 
2720 	if (qp_type == IB_QPT_RAW_PACKET && attr->rwq_ind_tbl)
2721 		return (create_flags) ? -EINVAL : 0;
2722 
2723 	process_create_flag(dev, &create_flags, IB_QP_CREATE_NETIF_QP,
2724 			    mlx5_get_flow_namespace(dev->mdev,
2725 						    MLX5_FLOW_NAMESPACE_BYPASS),
2726 			    qp);
2727 	process_create_flag(dev, &create_flags,
2728 			    IB_QP_CREATE_INTEGRITY_EN,
2729 			    MLX5_CAP_GEN(mdev, sho), qp);
2730 	process_create_flag(dev, &create_flags,
2731 			    IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK,
2732 			    MLX5_CAP_GEN(mdev, block_lb_mc), qp);
2733 	process_create_flag(dev, &create_flags, IB_QP_CREATE_CROSS_CHANNEL,
2734 			    MLX5_CAP_GEN(mdev, cd), qp);
2735 	process_create_flag(dev, &create_flags, IB_QP_CREATE_MANAGED_SEND,
2736 			    MLX5_CAP_GEN(mdev, cd), qp);
2737 	process_create_flag(dev, &create_flags, IB_QP_CREATE_MANAGED_RECV,
2738 			    MLX5_CAP_GEN(mdev, cd), qp);
2739 
2740 	if (qp_type == IB_QPT_UD) {
2741 		process_create_flag(dev, &create_flags,
2742 				    IB_QP_CREATE_IPOIB_UD_LSO,
2743 				    MLX5_CAP_GEN(mdev, ipoib_basic_offloads),
2744 				    qp);
2745 		cond = MLX5_CAP_GEN(mdev, port_type) == MLX5_CAP_PORT_TYPE_IB;
2746 		process_create_flag(dev, &create_flags, IB_QP_CREATE_SOURCE_QPN,
2747 				    cond, qp);
2748 	}
2749 
2750 	if (qp_type == IB_QPT_RAW_PACKET) {
2751 		cond = MLX5_CAP_GEN(mdev, eth_net_offloads) &&
2752 		       MLX5_CAP_ETH(mdev, scatter_fcs);
2753 		process_create_flag(dev, &create_flags,
2754 				    IB_QP_CREATE_SCATTER_FCS, cond, qp);
2755 
2756 		cond = MLX5_CAP_GEN(mdev, eth_net_offloads) &&
2757 		       MLX5_CAP_ETH(mdev, vlan_cap);
2758 		process_create_flag(dev, &create_flags,
2759 				    IB_QP_CREATE_CVLAN_STRIPPING, cond, qp);
2760 	}
2761 
2762 	process_create_flag(dev, &create_flags,
2763 			    IB_QP_CREATE_PCI_WRITE_END_PADDING,
2764 			    MLX5_CAP_GEN(mdev, end_pad), qp);
2765 
2766 	process_create_flag(dev, &create_flags, MLX5_IB_QP_CREATE_WC_TEST,
2767 			    qp_type != MLX5_IB_QPT_REG_UMR, qp);
2768 	process_create_flag(dev, &create_flags, MLX5_IB_QP_CREATE_SQPN_QP1,
2769 			    true, qp);
2770 
2771 	if (create_flags) {
2772 		mlx5_ib_dbg(dev, "Create QP has unsupported flags 0x%X\n",
2773 			    create_flags);
2774 		return -EOPNOTSUPP;
2775 	}
2776 	return 0;
2777 }
2778 
2779 static int process_udata_size(struct mlx5_ib_dev *dev,
2780 			      struct mlx5_create_qp_params *params)
2781 {
2782 	size_t ucmd = sizeof(struct mlx5_ib_create_qp);
2783 	struct ib_udata *udata = params->udata;
2784 	size_t outlen = udata->outlen;
2785 	size_t inlen = udata->inlen;
2786 
2787 	params->outlen = min(outlen, sizeof(struct mlx5_ib_create_qp_resp));
2788 	params->ucmd_size = ucmd;
2789 	if (!params->is_rss_raw) {
2790 		/* User has old rdma-core, which doesn't support ECE */
2791 		size_t min_inlen =
2792 			offsetof(struct mlx5_ib_create_qp, ece_options);
2793 
2794 		/*
2795 		 * We will check in check_ucmd_data() that user
2796 		 * cleared everything after inlen.
2797 		 */
2798 		params->inlen = (inlen < min_inlen) ? 0 : min(inlen, ucmd);
2799 		goto out;
2800 	}
2801 
2802 	/* RSS RAW QP */
2803 	if (inlen < offsetofend(struct mlx5_ib_create_qp_rss, flags))
2804 		return -EINVAL;
2805 
2806 	if (outlen < offsetofend(struct mlx5_ib_create_qp_resp, bfreg_index))
2807 		return -EINVAL;
2808 
2809 	ucmd = sizeof(struct mlx5_ib_create_qp_rss);
2810 	params->ucmd_size = ucmd;
2811 	if (inlen > ucmd && !ib_is_udata_cleared(udata, ucmd, inlen - ucmd))
2812 		return -EINVAL;
2813 
2814 	params->inlen = min(ucmd, inlen);
2815 out:
2816 	if (!params->inlen)
2817 		mlx5_ib_dbg(dev, "udata is too small\n");
2818 
2819 	return (params->inlen) ? 0 : -EINVAL;
2820 }
2821 
2822 static int create_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2823 		     struct mlx5_ib_qp *qp,
2824 		     struct mlx5_create_qp_params *params)
2825 {
2826 	int err;
2827 
2828 	if (params->is_rss_raw) {
2829 		err = create_rss_raw_qp_tir(dev, pd, qp, params);
2830 		goto out;
2831 	}
2832 
2833 	switch (qp->type) {
2834 	case MLX5_IB_QPT_DCT:
2835 		err = create_dct(dev, pd, qp, params);
2836 		break;
2837 	case IB_QPT_XRC_TGT:
2838 		err = create_xrc_tgt_qp(dev, qp, params);
2839 		break;
2840 	case IB_QPT_GSI:
2841 		err = mlx5_ib_create_gsi(pd, qp, params->attr);
2842 		break;
2843 	default:
2844 		if (params->udata)
2845 			err = create_user_qp(dev, pd, qp, params);
2846 		else
2847 			err = create_kernel_qp(dev, pd, qp, params);
2848 	}
2849 
2850 out:
2851 	if (err) {
2852 		mlx5_ib_err(dev, "Create QP type %d failed\n", qp->type);
2853 		return err;
2854 	}
2855 
2856 	if (is_qp0(qp->type))
2857 		qp->ibqp.qp_num = 0;
2858 	else if (is_qp1(qp->type))
2859 		qp->ibqp.qp_num = 1;
2860 	else
2861 		qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
2862 
2863 	mlx5_ib_dbg(dev,
2864 		"QP type %d, ib qpn 0x%X, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x, ece 0x%x\n",
2865 		qp->type, qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2866 		params->attr->recv_cq ? to_mcq(params->attr->recv_cq)->mcq.cqn :
2867 					-1,
2868 		params->attr->send_cq ? to_mcq(params->attr->send_cq)->mcq.cqn :
2869 					-1,
2870 		params->resp.ece_options);
2871 
2872 	return 0;
2873 }
2874 
2875 static int check_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2876 			 struct ib_qp_init_attr *attr)
2877 {
2878 	int ret = 0;
2879 
2880 	switch (qp->type) {
2881 	case MLX5_IB_QPT_DCT:
2882 		ret = (!attr->srq || !attr->recv_cq) ? -EINVAL : 0;
2883 		break;
2884 	case MLX5_IB_QPT_DCI:
2885 		ret = (attr->cap.max_recv_wr || attr->cap.max_recv_sge) ?
2886 			      -EINVAL :
2887 			      0;
2888 		break;
2889 	case IB_QPT_RAW_PACKET:
2890 		ret = (attr->rwq_ind_tbl && attr->send_cq) ? -EINVAL : 0;
2891 		break;
2892 	default:
2893 		break;
2894 	}
2895 
2896 	if (ret)
2897 		mlx5_ib_dbg(dev, "QP type %d has wrong attributes\n", qp->type);
2898 
2899 	return ret;
2900 }
2901 
2902 static int get_qp_uidx(struct mlx5_ib_qp *qp,
2903 		       struct mlx5_create_qp_params *params)
2904 {
2905 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
2906 	struct ib_udata *udata = params->udata;
2907 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2908 		udata, struct mlx5_ib_ucontext, ibucontext);
2909 
2910 	if (params->is_rss_raw)
2911 		return 0;
2912 
2913 	return get_qp_user_index(ucontext, ucmd, sizeof(*ucmd), &params->uidx);
2914 }
2915 
2916 static int mlx5_ib_destroy_dct(struct mlx5_ib_qp *mqp)
2917 {
2918 	struct mlx5_ib_dev *dev = to_mdev(mqp->ibqp.device);
2919 
2920 	if (mqp->state == IB_QPS_RTR) {
2921 		int err;
2922 
2923 		err = mlx5_core_destroy_dct(dev, &mqp->dct.mdct);
2924 		if (err) {
2925 			mlx5_ib_warn(dev, "failed to destroy DCT %d\n", err);
2926 			return err;
2927 		}
2928 	}
2929 
2930 	kfree(mqp->dct.in);
2931 	kfree(mqp);
2932 	return 0;
2933 }
2934 
2935 static int check_ucmd_data(struct mlx5_ib_dev *dev,
2936 			   struct mlx5_create_qp_params *params)
2937 {
2938 	struct ib_udata *udata = params->udata;
2939 	size_t size, last;
2940 	int ret;
2941 
2942 	if (params->is_rss_raw)
2943 		/*
2944 		 * These QPs don't have "reserved" field in their
2945 		 * create_qp input struct, so their data is always valid.
2946 		 */
2947 		last = sizeof(struct mlx5_ib_create_qp_rss);
2948 	else
2949 		last = offsetof(struct mlx5_ib_create_qp, reserved);
2950 
2951 	if (udata->inlen <= last)
2952 		return 0;
2953 
2954 	/*
2955 	 * User provides different create_qp structures based on the
2956 	 * flow and we need to know if he cleared memory after our
2957 	 * struct create_qp ends.
2958 	 */
2959 	size = udata->inlen - last;
2960 	ret = ib_is_udata_cleared(params->udata, last, size);
2961 	if (!ret)
2962 		mlx5_ib_dbg(
2963 			dev,
2964 			"udata is not cleared, inlen = %zu, ucmd = %zu, last = %zu, size = %zu\n",
2965 			udata->inlen, params->ucmd_size, last, size);
2966 	return ret ? 0 : -EINVAL;
2967 }
2968 
2969 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attr,
2970 				struct ib_udata *udata)
2971 {
2972 	struct mlx5_create_qp_params params = {};
2973 	struct mlx5_ib_dev *dev;
2974 	struct mlx5_ib_qp *qp;
2975 	enum ib_qp_type type;
2976 	int err;
2977 
2978 	dev = pd ? to_mdev(pd->device) :
2979 		   to_mdev(to_mxrcd(attr->xrcd)->ibxrcd.device);
2980 
2981 	err = check_qp_type(dev, attr, &type);
2982 	if (err)
2983 		return ERR_PTR(err);
2984 
2985 	err = check_valid_flow(dev, pd, attr, udata);
2986 	if (err)
2987 		return ERR_PTR(err);
2988 
2989 	params.udata = udata;
2990 	params.uidx = MLX5_IB_DEFAULT_UIDX;
2991 	params.attr = attr;
2992 	params.is_rss_raw = !!attr->rwq_ind_tbl;
2993 
2994 	if (udata) {
2995 		err = process_udata_size(dev, &params);
2996 		if (err)
2997 			return ERR_PTR(err);
2998 
2999 		err = check_ucmd_data(dev, &params);
3000 		if (err)
3001 			return ERR_PTR(err);
3002 
3003 		params.ucmd = kzalloc(params.ucmd_size, GFP_KERNEL);
3004 		if (!params.ucmd)
3005 			return ERR_PTR(-ENOMEM);
3006 
3007 		err = ib_copy_from_udata(params.ucmd, udata, params.inlen);
3008 		if (err)
3009 			goto free_ucmd;
3010 	}
3011 
3012 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
3013 	if (!qp) {
3014 		err = -ENOMEM;
3015 		goto free_ucmd;
3016 	}
3017 
3018 	mutex_init(&qp->mutex);
3019 	qp->type = type;
3020 	if (udata) {
3021 		err = process_vendor_flags(dev, qp, params.ucmd, attr);
3022 		if (err)
3023 			goto free_qp;
3024 
3025 		err = get_qp_uidx(qp, &params);
3026 		if (err)
3027 			goto free_qp;
3028 	}
3029 	err = process_create_flags(dev, qp, attr);
3030 	if (err)
3031 		goto free_qp;
3032 
3033 	err = check_qp_attr(dev, qp, attr);
3034 	if (err)
3035 		goto free_qp;
3036 
3037 	err = create_qp(dev, pd, qp, &params);
3038 	if (err)
3039 		goto free_qp;
3040 
3041 	kfree(params.ucmd);
3042 	params.ucmd = NULL;
3043 
3044 	if (udata)
3045 		/*
3046 		 * It is safe to copy response for all user create QP flows,
3047 		 * including MLX5_IB_QPT_DCT, which doesn't need it.
3048 		 * In that case, resp will be filled with zeros.
3049 		 */
3050 		err = ib_copy_to_udata(udata, &params.resp, params.outlen);
3051 	if (err)
3052 		goto destroy_qp;
3053 
3054 	return &qp->ibqp;
3055 
3056 destroy_qp:
3057 	switch (qp->type) {
3058 	case MLX5_IB_QPT_DCT:
3059 		mlx5_ib_destroy_dct(qp);
3060 		break;
3061 	case IB_QPT_GSI:
3062 		mlx5_ib_destroy_gsi(qp);
3063 		break;
3064 	default:
3065 		/*
3066 		 * These lines below are temp solution till QP allocation
3067 		 * will be moved to be under IB/core responsiblity.
3068 		 */
3069 		qp->ibqp.send_cq = attr->send_cq;
3070 		qp->ibqp.recv_cq = attr->recv_cq;
3071 		qp->ibqp.pd = pd;
3072 		destroy_qp_common(dev, qp, udata);
3073 	}
3074 
3075 	qp = NULL;
3076 free_qp:
3077 	kfree(qp);
3078 free_ucmd:
3079 	kfree(params.ucmd);
3080 	return ERR_PTR(err);
3081 }
3082 
3083 int mlx5_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
3084 {
3085 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3086 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3087 
3088 	if (unlikely(qp->qp_type == IB_QPT_GSI))
3089 		return mlx5_ib_destroy_gsi(mqp);
3090 
3091 	if (mqp->type == MLX5_IB_QPT_DCT)
3092 		return mlx5_ib_destroy_dct(mqp);
3093 
3094 	destroy_qp_common(dev, mqp, udata);
3095 
3096 	kfree(mqp);
3097 
3098 	return 0;
3099 }
3100 
3101 static int set_qpc_atomic_flags(struct mlx5_ib_qp *qp,
3102 				const struct ib_qp_attr *attr, int attr_mask,
3103 				void *qpc)
3104 {
3105 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
3106 	u8 dest_rd_atomic;
3107 	u32 access_flags;
3108 
3109 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
3110 		dest_rd_atomic = attr->max_dest_rd_atomic;
3111 	else
3112 		dest_rd_atomic = qp->trans_qp.resp_depth;
3113 
3114 	if (attr_mask & IB_QP_ACCESS_FLAGS)
3115 		access_flags = attr->qp_access_flags;
3116 	else
3117 		access_flags = qp->trans_qp.atomic_rd_en;
3118 
3119 	if (!dest_rd_atomic)
3120 		access_flags &= IB_ACCESS_REMOTE_WRITE;
3121 
3122 	MLX5_SET(qpc, qpc, rre, !!(access_flags & IB_ACCESS_REMOTE_READ));
3123 
3124 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC) {
3125 		int atomic_mode;
3126 
3127 		atomic_mode = get_atomic_mode(dev, qp->ibqp.qp_type);
3128 		if (atomic_mode < 0)
3129 			return -EOPNOTSUPP;
3130 
3131 		MLX5_SET(qpc, qpc, rae, 1);
3132 		MLX5_SET(qpc, qpc, atomic_mode, atomic_mode);
3133 	}
3134 
3135 	MLX5_SET(qpc, qpc, rwe, !!(access_flags & IB_ACCESS_REMOTE_WRITE));
3136 	return 0;
3137 }
3138 
3139 enum {
3140 	MLX5_PATH_FLAG_FL	= 1 << 0,
3141 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
3142 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
3143 };
3144 
3145 static int ib_to_mlx5_rate_map(u8 rate)
3146 {
3147 	switch (rate) {
3148 	case IB_RATE_PORT_CURRENT:
3149 		return 0;
3150 	case IB_RATE_56_GBPS:
3151 		return 1;
3152 	case IB_RATE_25_GBPS:
3153 		return 2;
3154 	case IB_RATE_100_GBPS:
3155 		return 3;
3156 	case IB_RATE_200_GBPS:
3157 		return 4;
3158 	case IB_RATE_50_GBPS:
3159 		return 5;
3160 	case IB_RATE_400_GBPS:
3161 		return 6;
3162 	default:
3163 		return rate + MLX5_STAT_RATE_OFFSET;
3164 	}
3165 
3166 	return 0;
3167 }
3168 
3169 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
3170 {
3171 	u32 stat_rate_support;
3172 
3173 	if (rate == IB_RATE_PORT_CURRENT)
3174 		return 0;
3175 
3176 	if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_600_GBPS)
3177 		return -EINVAL;
3178 
3179 	stat_rate_support = MLX5_CAP_GEN(dev->mdev, stat_rate_support);
3180 	while (rate != IB_RATE_PORT_CURRENT &&
3181 	       !(1 << ib_to_mlx5_rate_map(rate) & stat_rate_support))
3182 		--rate;
3183 
3184 	return ib_to_mlx5_rate_map(rate);
3185 }
3186 
3187 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
3188 				      struct mlx5_ib_sq *sq, u8 sl,
3189 				      struct ib_pd *pd)
3190 {
3191 	void *in;
3192 	void *tisc;
3193 	int inlen;
3194 	int err;
3195 
3196 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
3197 	in = kvzalloc(inlen, GFP_KERNEL);
3198 	if (!in)
3199 		return -ENOMEM;
3200 
3201 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
3202 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
3203 
3204 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
3205 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
3206 
3207 	err = mlx5_core_modify_tis(dev, sq->tisn, in);
3208 
3209 	kvfree(in);
3210 
3211 	return err;
3212 }
3213 
3214 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
3215 					 struct mlx5_ib_sq *sq, u8 tx_affinity,
3216 					 struct ib_pd *pd)
3217 {
3218 	void *in;
3219 	void *tisc;
3220 	int inlen;
3221 	int err;
3222 
3223 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
3224 	in = kvzalloc(inlen, GFP_KERNEL);
3225 	if (!in)
3226 		return -ENOMEM;
3227 
3228 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
3229 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
3230 
3231 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
3232 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
3233 
3234 	err = mlx5_core_modify_tis(dev, sq->tisn, in);
3235 
3236 	kvfree(in);
3237 
3238 	return err;
3239 }
3240 
3241 static void mlx5_set_path_udp_sport(void *path, const struct rdma_ah_attr *ah,
3242 				    u32 lqpn, u32 rqpn)
3243 
3244 {
3245 	u32 fl = ah->grh.flow_label;
3246 
3247 	if (!fl)
3248 		fl = rdma_calc_flow_label(lqpn, rqpn);
3249 
3250 	MLX5_SET(ads, path, udp_sport, rdma_flow_label_to_udp_sport(fl));
3251 }
3252 
3253 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3254 			 const struct rdma_ah_attr *ah, void *path, u8 port,
3255 			 int attr_mask, u32 path_flags,
3256 			 const struct ib_qp_attr *attr, bool alt)
3257 {
3258 	const struct ib_global_route *grh = rdma_ah_read_grh(ah);
3259 	int err;
3260 	enum ib_gid_type gid_type;
3261 	u8 ah_flags = rdma_ah_get_ah_flags(ah);
3262 	u8 sl = rdma_ah_get_sl(ah);
3263 
3264 	if (attr_mask & IB_QP_PKEY_INDEX)
3265 		MLX5_SET(ads, path, pkey_index,
3266 			 alt ? attr->alt_pkey_index : attr->pkey_index);
3267 
3268 	if (ah_flags & IB_AH_GRH) {
3269 		const struct ib_port_immutable *immutable;
3270 
3271 		immutable = ib_port_immutable_read(&dev->ib_dev, port);
3272 		if (grh->sgid_index >= immutable->gid_tbl_len) {
3273 			pr_err("sgid_index (%u) too large. max is %d\n",
3274 			       grh->sgid_index,
3275 			       immutable->gid_tbl_len);
3276 			return -EINVAL;
3277 		}
3278 	}
3279 
3280 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
3281 		if (!(ah_flags & IB_AH_GRH))
3282 			return -EINVAL;
3283 
3284 		ether_addr_copy(MLX5_ADDR_OF(ads, path, rmac_47_32),
3285 				ah->roce.dmac);
3286 		if ((qp->ibqp.qp_type == IB_QPT_RC ||
3287 		     qp->ibqp.qp_type == IB_QPT_UC ||
3288 		     qp->ibqp.qp_type == IB_QPT_XRC_INI ||
3289 		     qp->ibqp.qp_type == IB_QPT_XRC_TGT) &&
3290 		    (grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) &&
3291 		    (attr_mask & IB_QP_DEST_QPN))
3292 			mlx5_set_path_udp_sport(path, ah,
3293 						qp->ibqp.qp_num,
3294 						attr->dest_qp_num);
3295 		MLX5_SET(ads, path, eth_prio, sl & 0x7);
3296 		gid_type = ah->grh.sgid_attr->gid_type;
3297 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
3298 			MLX5_SET(ads, path, dscp, grh->traffic_class >> 2);
3299 	} else {
3300 		MLX5_SET(ads, path, fl, !!(path_flags & MLX5_PATH_FLAG_FL));
3301 		MLX5_SET(ads, path, free_ar,
3302 			 !!(path_flags & MLX5_PATH_FLAG_FREE_AR));
3303 		MLX5_SET(ads, path, rlid, rdma_ah_get_dlid(ah));
3304 		MLX5_SET(ads, path, mlid, rdma_ah_get_path_bits(ah));
3305 		MLX5_SET(ads, path, grh, !!(ah_flags & IB_AH_GRH));
3306 		MLX5_SET(ads, path, sl, sl);
3307 	}
3308 
3309 	if (ah_flags & IB_AH_GRH) {
3310 		MLX5_SET(ads, path, src_addr_index, grh->sgid_index);
3311 		MLX5_SET(ads, path, hop_limit, grh->hop_limit);
3312 		MLX5_SET(ads, path, tclass, grh->traffic_class);
3313 		MLX5_SET(ads, path, flow_label, grh->flow_label);
3314 		memcpy(MLX5_ADDR_OF(ads, path, rgid_rip), grh->dgid.raw,
3315 		       sizeof(grh->dgid.raw));
3316 	}
3317 
3318 	err = ib_rate_to_mlx5(dev, rdma_ah_get_static_rate(ah));
3319 	if (err < 0)
3320 		return err;
3321 	MLX5_SET(ads, path, stat_rate, err);
3322 	MLX5_SET(ads, path, vhca_port_num, port);
3323 
3324 	if (attr_mask & IB_QP_TIMEOUT)
3325 		MLX5_SET(ads, path, ack_timeout,
3326 			 alt ? attr->alt_timeout : attr->timeout);
3327 
3328 	if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
3329 		return modify_raw_packet_eth_prio(dev->mdev,
3330 						  &qp->raw_packet_qp.sq,
3331 						  sl & 0xf, qp->ibqp.pd);
3332 
3333 	return 0;
3334 }
3335 
3336 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
3337 	[MLX5_QP_STATE_INIT] = {
3338 		[MLX5_QP_STATE_INIT] = {
3339 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3340 					  MLX5_QP_OPTPAR_RAE		|
3341 					  MLX5_QP_OPTPAR_RWE		|
3342 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3343 					  MLX5_QP_OPTPAR_PRI_PORT	|
3344 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3345 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3346 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3347 					  MLX5_QP_OPTPAR_PRI_PORT	|
3348 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3349 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3350 					  MLX5_QP_OPTPAR_Q_KEY		|
3351 					  MLX5_QP_OPTPAR_PRI_PORT,
3352 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3353 					  MLX5_QP_OPTPAR_RAE		|
3354 					  MLX5_QP_OPTPAR_RWE		|
3355 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3356 					  MLX5_QP_OPTPAR_PRI_PORT	|
3357 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3358 		},
3359 		[MLX5_QP_STATE_RTR] = {
3360 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3361 					  MLX5_QP_OPTPAR_RRE            |
3362 					  MLX5_QP_OPTPAR_RAE            |
3363 					  MLX5_QP_OPTPAR_RWE            |
3364 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3365 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3366 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3367 					  MLX5_QP_OPTPAR_RWE            |
3368 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3369 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3370 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
3371 					  MLX5_QP_OPTPAR_Q_KEY,
3372 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3373 					   MLX5_QP_OPTPAR_Q_KEY,
3374 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
3375 					  MLX5_QP_OPTPAR_RRE            |
3376 					  MLX5_QP_OPTPAR_RAE            |
3377 					  MLX5_QP_OPTPAR_RWE            |
3378 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3379 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3380 		},
3381 	},
3382 	[MLX5_QP_STATE_RTR] = {
3383 		[MLX5_QP_STATE_RTS] = {
3384 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3385 					  MLX5_QP_OPTPAR_RRE		|
3386 					  MLX5_QP_OPTPAR_RAE		|
3387 					  MLX5_QP_OPTPAR_RWE		|
3388 					  MLX5_QP_OPTPAR_PM_STATE	|
3389 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3390 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3391 					  MLX5_QP_OPTPAR_RWE		|
3392 					  MLX5_QP_OPTPAR_PM_STATE,
3393 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
3394 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3395 					  MLX5_QP_OPTPAR_RRE		|
3396 					  MLX5_QP_OPTPAR_RAE		|
3397 					  MLX5_QP_OPTPAR_RWE		|
3398 					  MLX5_QP_OPTPAR_PM_STATE	|
3399 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3400 		},
3401 	},
3402 	[MLX5_QP_STATE_RTS] = {
3403 		[MLX5_QP_STATE_RTS] = {
3404 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3405 					  MLX5_QP_OPTPAR_RAE		|
3406 					  MLX5_QP_OPTPAR_RWE		|
3407 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3408 					  MLX5_QP_OPTPAR_PM_STATE	|
3409 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3410 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3411 					  MLX5_QP_OPTPAR_PM_STATE	|
3412 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3413 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
3414 					  MLX5_QP_OPTPAR_SRQN		|
3415 					  MLX5_QP_OPTPAR_CQN_RCV,
3416 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3417 					  MLX5_QP_OPTPAR_RAE		|
3418 					  MLX5_QP_OPTPAR_RWE		|
3419 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3420 					  MLX5_QP_OPTPAR_PM_STATE	|
3421 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3422 		},
3423 	},
3424 	[MLX5_QP_STATE_SQER] = {
3425 		[MLX5_QP_STATE_RTS] = {
3426 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
3427 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
3428 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
3429 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3430 					   MLX5_QP_OPTPAR_RWE		|
3431 					   MLX5_QP_OPTPAR_RAE		|
3432 					   MLX5_QP_OPTPAR_RRE,
3433 			[MLX5_QP_ST_XRC]  = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3434 					   MLX5_QP_OPTPAR_RWE		|
3435 					   MLX5_QP_OPTPAR_RAE		|
3436 					   MLX5_QP_OPTPAR_RRE,
3437 		},
3438 	},
3439 };
3440 
3441 static int ib_nr_to_mlx5_nr(int ib_mask)
3442 {
3443 	switch (ib_mask) {
3444 	case IB_QP_STATE:
3445 		return 0;
3446 	case IB_QP_CUR_STATE:
3447 		return 0;
3448 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
3449 		return 0;
3450 	case IB_QP_ACCESS_FLAGS:
3451 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
3452 			MLX5_QP_OPTPAR_RAE;
3453 	case IB_QP_PKEY_INDEX:
3454 		return MLX5_QP_OPTPAR_PKEY_INDEX;
3455 	case IB_QP_PORT:
3456 		return MLX5_QP_OPTPAR_PRI_PORT;
3457 	case IB_QP_QKEY:
3458 		return MLX5_QP_OPTPAR_Q_KEY;
3459 	case IB_QP_AV:
3460 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
3461 			MLX5_QP_OPTPAR_PRI_PORT;
3462 	case IB_QP_PATH_MTU:
3463 		return 0;
3464 	case IB_QP_TIMEOUT:
3465 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
3466 	case IB_QP_RETRY_CNT:
3467 		return MLX5_QP_OPTPAR_RETRY_COUNT;
3468 	case IB_QP_RNR_RETRY:
3469 		return MLX5_QP_OPTPAR_RNR_RETRY;
3470 	case IB_QP_RQ_PSN:
3471 		return 0;
3472 	case IB_QP_MAX_QP_RD_ATOMIC:
3473 		return MLX5_QP_OPTPAR_SRA_MAX;
3474 	case IB_QP_ALT_PATH:
3475 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
3476 	case IB_QP_MIN_RNR_TIMER:
3477 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
3478 	case IB_QP_SQ_PSN:
3479 		return 0;
3480 	case IB_QP_MAX_DEST_RD_ATOMIC:
3481 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
3482 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
3483 	case IB_QP_PATH_MIG_STATE:
3484 		return MLX5_QP_OPTPAR_PM_STATE;
3485 	case IB_QP_CAP:
3486 		return 0;
3487 	case IB_QP_DEST_QPN:
3488 		return 0;
3489 	}
3490 	return 0;
3491 }
3492 
3493 static int ib_mask_to_mlx5_opt(int ib_mask)
3494 {
3495 	int result = 0;
3496 	int i;
3497 
3498 	for (i = 0; i < 8 * sizeof(int); i++) {
3499 		if ((1 << i) & ib_mask)
3500 			result |= ib_nr_to_mlx5_nr(1 << i);
3501 	}
3502 
3503 	return result;
3504 }
3505 
3506 static int modify_raw_packet_qp_rq(
3507 	struct mlx5_ib_dev *dev, struct mlx5_ib_rq *rq, int new_state,
3508 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3509 {
3510 	void *in;
3511 	void *rqc;
3512 	int inlen;
3513 	int err;
3514 
3515 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
3516 	in = kvzalloc(inlen, GFP_KERNEL);
3517 	if (!in)
3518 		return -ENOMEM;
3519 
3520 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
3521 	MLX5_SET(modify_rq_in, in, uid, to_mpd(pd)->uid);
3522 
3523 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
3524 	MLX5_SET(rqc, rqc, state, new_state);
3525 
3526 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
3527 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
3528 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
3529 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
3530 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
3531 		} else
3532 			dev_info_once(
3533 				&dev->ib_dev.dev,
3534 				"RAW PACKET QP counters are not supported on current FW\n");
3535 	}
3536 
3537 	err = mlx5_core_modify_rq(dev->mdev, rq->base.mqp.qpn, in);
3538 	if (err)
3539 		goto out;
3540 
3541 	rq->state = new_state;
3542 
3543 out:
3544 	kvfree(in);
3545 	return err;
3546 }
3547 
3548 static int modify_raw_packet_qp_sq(
3549 	struct mlx5_core_dev *dev, struct mlx5_ib_sq *sq, int new_state,
3550 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3551 {
3552 	struct mlx5_ib_qp *ibqp = sq->base.container_mibqp;
3553 	struct mlx5_rate_limit old_rl = ibqp->rl;
3554 	struct mlx5_rate_limit new_rl = old_rl;
3555 	bool new_rate_added = false;
3556 	u16 rl_index = 0;
3557 	void *in;
3558 	void *sqc;
3559 	int inlen;
3560 	int err;
3561 
3562 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
3563 	in = kvzalloc(inlen, GFP_KERNEL);
3564 	if (!in)
3565 		return -ENOMEM;
3566 
3567 	MLX5_SET(modify_sq_in, in, uid, to_mpd(pd)->uid);
3568 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
3569 
3570 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
3571 	MLX5_SET(sqc, sqc, state, new_state);
3572 
3573 	if (raw_qp_param->set_mask & MLX5_RAW_QP_RATE_LIMIT) {
3574 		if (new_state != MLX5_SQC_STATE_RDY)
3575 			pr_warn("%s: Rate limit can only be changed when SQ is moving to RDY\n",
3576 				__func__);
3577 		else
3578 			new_rl = raw_qp_param->rl;
3579 	}
3580 
3581 	if (!mlx5_rl_are_equal(&old_rl, &new_rl)) {
3582 		if (new_rl.rate) {
3583 			err = mlx5_rl_add_rate(dev, &rl_index, &new_rl);
3584 			if (err) {
3585 				pr_err("Failed configuring rate limit(err %d): \
3586 				       rate %u, max_burst_sz %u, typical_pkt_sz %u\n",
3587 				       err, new_rl.rate, new_rl.max_burst_sz,
3588 				       new_rl.typical_pkt_sz);
3589 
3590 				goto out;
3591 			}
3592 			new_rate_added = true;
3593 		}
3594 
3595 		MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
3596 		/* index 0 means no limit */
3597 		MLX5_SET(sqc, sqc, packet_pacing_rate_limit_index, rl_index);
3598 	}
3599 
3600 	err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in);
3601 	if (err) {
3602 		/* Remove new rate from table if failed */
3603 		if (new_rate_added)
3604 			mlx5_rl_remove_rate(dev, &new_rl);
3605 		goto out;
3606 	}
3607 
3608 	/* Only remove the old rate after new rate was set */
3609 	if ((old_rl.rate && !mlx5_rl_are_equal(&old_rl, &new_rl)) ||
3610 	    (new_state != MLX5_SQC_STATE_RDY)) {
3611 		mlx5_rl_remove_rate(dev, &old_rl);
3612 		if (new_state != MLX5_SQC_STATE_RDY)
3613 			memset(&new_rl, 0, sizeof(new_rl));
3614 	}
3615 
3616 	ibqp->rl = new_rl;
3617 	sq->state = new_state;
3618 
3619 out:
3620 	kvfree(in);
3621 	return err;
3622 }
3623 
3624 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3625 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
3626 				u8 tx_affinity)
3627 {
3628 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
3629 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
3630 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
3631 	int modify_rq = !!qp->rq.wqe_cnt;
3632 	int modify_sq = !!qp->sq.wqe_cnt;
3633 	int rq_state;
3634 	int sq_state;
3635 	int err;
3636 
3637 	switch (raw_qp_param->operation) {
3638 	case MLX5_CMD_OP_RST2INIT_QP:
3639 		rq_state = MLX5_RQC_STATE_RDY;
3640 		sq_state = MLX5_SQC_STATE_RST;
3641 		break;
3642 	case MLX5_CMD_OP_2ERR_QP:
3643 		rq_state = MLX5_RQC_STATE_ERR;
3644 		sq_state = MLX5_SQC_STATE_ERR;
3645 		break;
3646 	case MLX5_CMD_OP_2RST_QP:
3647 		rq_state = MLX5_RQC_STATE_RST;
3648 		sq_state = MLX5_SQC_STATE_RST;
3649 		break;
3650 	case MLX5_CMD_OP_RTR2RTS_QP:
3651 	case MLX5_CMD_OP_RTS2RTS_QP:
3652 		if (raw_qp_param->set_mask & ~MLX5_RAW_QP_RATE_LIMIT)
3653 			return -EINVAL;
3654 
3655 		modify_rq = 0;
3656 		sq_state = MLX5_SQC_STATE_RDY;
3657 		break;
3658 	case MLX5_CMD_OP_INIT2INIT_QP:
3659 	case MLX5_CMD_OP_INIT2RTR_QP:
3660 		if (raw_qp_param->set_mask)
3661 			return -EINVAL;
3662 		else
3663 			return 0;
3664 	default:
3665 		WARN_ON(1);
3666 		return -EINVAL;
3667 	}
3668 
3669 	if (modify_rq) {
3670 		err =  modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param,
3671 					       qp->ibqp.pd);
3672 		if (err)
3673 			return err;
3674 	}
3675 
3676 	if (modify_sq) {
3677 		struct mlx5_flow_handle *flow_rule;
3678 
3679 		if (tx_affinity) {
3680 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
3681 							    tx_affinity,
3682 							    qp->ibqp.pd);
3683 			if (err)
3684 				return err;
3685 		}
3686 
3687 		flow_rule = create_flow_rule_vport_sq(dev, sq,
3688 						      raw_qp_param->port);
3689 		if (IS_ERR(flow_rule))
3690 			return PTR_ERR(flow_rule);
3691 
3692 		err = modify_raw_packet_qp_sq(dev->mdev, sq, sq_state,
3693 					      raw_qp_param, qp->ibqp.pd);
3694 		if (err) {
3695 			if (flow_rule)
3696 				mlx5_del_flow_rules(flow_rule);
3697 			return err;
3698 		}
3699 
3700 		if (flow_rule) {
3701 			destroy_flow_rule_vport_sq(sq);
3702 			sq->flow_rule = flow_rule;
3703 		}
3704 
3705 		return err;
3706 	}
3707 
3708 	return 0;
3709 }
3710 
3711 static unsigned int get_tx_affinity_rr(struct mlx5_ib_dev *dev,
3712 				       struct ib_udata *udata)
3713 {
3714 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3715 		udata, struct mlx5_ib_ucontext, ibucontext);
3716 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3717 	atomic_t *tx_port_affinity;
3718 
3719 	if (ucontext)
3720 		tx_port_affinity = &ucontext->tx_port_affinity;
3721 	else
3722 		tx_port_affinity = &dev->port[port_num].roce.tx_port_affinity;
3723 
3724 	return (unsigned int)atomic_add_return(1, tx_port_affinity) %
3725 		MLX5_MAX_PORTS + 1;
3726 }
3727 
3728 static bool qp_supports_affinity(struct mlx5_ib_qp *qp)
3729 {
3730 	if ((qp->type == IB_QPT_RC) || (qp->type == IB_QPT_UD) ||
3731 	    (qp->type == IB_QPT_UC) || (qp->type == IB_QPT_RAW_PACKET) ||
3732 	    (qp->type == IB_QPT_XRC_INI) || (qp->type == IB_QPT_XRC_TGT) ||
3733 	    (qp->type == MLX5_IB_QPT_DCI))
3734 		return true;
3735 	return false;
3736 }
3737 
3738 static unsigned int get_tx_affinity(struct ib_qp *qp,
3739 				    const struct ib_qp_attr *attr,
3740 				    int attr_mask, u8 init,
3741 				    struct ib_udata *udata)
3742 {
3743 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3744 		udata, struct mlx5_ib_ucontext, ibucontext);
3745 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3746 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3747 	struct mlx5_ib_qp_base *qp_base;
3748 	unsigned int tx_affinity;
3749 
3750 	if (!(mlx5_ib_lag_should_assign_affinity(dev) &&
3751 	      qp_supports_affinity(mqp)))
3752 		return 0;
3753 
3754 	if (mqp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
3755 		tx_affinity = mqp->gsi_lag_port;
3756 	else if (init)
3757 		tx_affinity = get_tx_affinity_rr(dev, udata);
3758 	else if ((attr_mask & IB_QP_AV) && attr->xmit_slave)
3759 		tx_affinity =
3760 			mlx5_lag_get_slave_port(dev->mdev, attr->xmit_slave);
3761 	else
3762 		return 0;
3763 
3764 	qp_base = &mqp->trans_qp.base;
3765 	if (ucontext)
3766 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x ucontext %p\n",
3767 			    tx_affinity, qp_base->mqp.qpn, ucontext);
3768 	else
3769 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x\n",
3770 			    tx_affinity, qp_base->mqp.qpn);
3771 	return tx_affinity;
3772 }
3773 
3774 static int __mlx5_ib_qp_set_counter(struct ib_qp *qp,
3775 				    struct rdma_counter *counter)
3776 {
3777 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3778 	u32 in[MLX5_ST_SZ_DW(rts2rts_qp_in)] = {};
3779 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3780 	struct mlx5_ib_qp_base *base;
3781 	u32 set_id;
3782 	u32 *qpc;
3783 
3784 	if (counter)
3785 		set_id = counter->id;
3786 	else
3787 		set_id = mlx5_ib_get_counters_id(dev, mqp->port - 1);
3788 
3789 	base = &mqp->trans_qp.base;
3790 	MLX5_SET(rts2rts_qp_in, in, opcode, MLX5_CMD_OP_RTS2RTS_QP);
3791 	MLX5_SET(rts2rts_qp_in, in, qpn, base->mqp.qpn);
3792 	MLX5_SET(rts2rts_qp_in, in, uid, base->mqp.uid);
3793 	MLX5_SET(rts2rts_qp_in, in, opt_param_mask,
3794 		 MLX5_QP_OPTPAR_COUNTER_SET_ID);
3795 
3796 	qpc = MLX5_ADDR_OF(rts2rts_qp_in, in, qpc);
3797 	MLX5_SET(qpc, qpc, counter_set_id, set_id);
3798 	return mlx5_cmd_exec_in(dev->mdev, rts2rts_qp, in);
3799 }
3800 
3801 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
3802 			       const struct ib_qp_attr *attr, int attr_mask,
3803 			       enum ib_qp_state cur_state,
3804 			       enum ib_qp_state new_state,
3805 			       const struct mlx5_ib_modify_qp *ucmd,
3806 			       struct mlx5_ib_modify_qp_resp *resp,
3807 			       struct ib_udata *udata)
3808 {
3809 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
3810 		[MLX5_QP_STATE_RST] = {
3811 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3812 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3813 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
3814 		},
3815 		[MLX5_QP_STATE_INIT]  = {
3816 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3817 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3818 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
3819 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
3820 		},
3821 		[MLX5_QP_STATE_RTR]   = {
3822 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3823 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3824 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
3825 		},
3826 		[MLX5_QP_STATE_RTS]   = {
3827 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3828 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3829 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
3830 		},
3831 		[MLX5_QP_STATE_SQD] = {
3832 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3833 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3834 		},
3835 		[MLX5_QP_STATE_SQER] = {
3836 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3837 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3838 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
3839 		},
3840 		[MLX5_QP_STATE_ERR] = {
3841 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3842 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3843 		}
3844 	};
3845 
3846 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3847 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3848 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
3849 	struct mlx5_ib_cq *send_cq, *recv_cq;
3850 	struct mlx5_ib_pd *pd;
3851 	enum mlx5_qp_state mlx5_cur, mlx5_new;
3852 	void *qpc, *pri_path, *alt_path;
3853 	enum mlx5_qp_optpar optpar = 0;
3854 	u32 set_id = 0;
3855 	int mlx5_st;
3856 	int err;
3857 	u16 op;
3858 	u8 tx_affinity = 0;
3859 
3860 	mlx5_st = to_mlx5_st(qp->type);
3861 	if (mlx5_st < 0)
3862 		return -EINVAL;
3863 
3864 	qpc = kzalloc(MLX5_ST_SZ_BYTES(qpc), GFP_KERNEL);
3865 	if (!qpc)
3866 		return -ENOMEM;
3867 
3868 	pd = to_mpd(qp->ibqp.pd);
3869 	MLX5_SET(qpc, qpc, st, mlx5_st);
3870 
3871 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
3872 		MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
3873 	} else {
3874 		switch (attr->path_mig_state) {
3875 		case IB_MIG_MIGRATED:
3876 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
3877 			break;
3878 		case IB_MIG_REARM:
3879 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_REARM);
3880 			break;
3881 		case IB_MIG_ARMED:
3882 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_ARMED);
3883 			break;
3884 		}
3885 	}
3886 
3887 	tx_affinity = get_tx_affinity(ibqp, attr, attr_mask,
3888 				      cur_state == IB_QPS_RESET &&
3889 				      new_state == IB_QPS_INIT, udata);
3890 
3891 	MLX5_SET(qpc, qpc, lag_tx_port_affinity, tx_affinity);
3892 	if (tx_affinity && new_state == IB_QPS_RTR &&
3893 	    MLX5_CAP_GEN(dev->mdev, init2_lag_tx_port_affinity))
3894 		optpar |= MLX5_QP_OPTPAR_LAG_TX_AFF;
3895 
3896 	if (is_sqp(ibqp->qp_type)) {
3897 		MLX5_SET(qpc, qpc, mtu, IB_MTU_256);
3898 		MLX5_SET(qpc, qpc, log_msg_max, 8);
3899 	} else if ((ibqp->qp_type == IB_QPT_UD &&
3900 		    !(qp->flags & IB_QP_CREATE_SOURCE_QPN)) ||
3901 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
3902 		MLX5_SET(qpc, qpc, mtu, IB_MTU_4096);
3903 		MLX5_SET(qpc, qpc, log_msg_max, 12);
3904 	} else if (attr_mask & IB_QP_PATH_MTU) {
3905 		if (attr->path_mtu < IB_MTU_256 ||
3906 		    attr->path_mtu > IB_MTU_4096) {
3907 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
3908 			err = -EINVAL;
3909 			goto out;
3910 		}
3911 		MLX5_SET(qpc, qpc, mtu, attr->path_mtu);
3912 		MLX5_SET(qpc, qpc, log_msg_max,
3913 			 MLX5_CAP_GEN(dev->mdev, log_max_msg));
3914 	}
3915 
3916 	if (attr_mask & IB_QP_DEST_QPN)
3917 		MLX5_SET(qpc, qpc, remote_qpn, attr->dest_qp_num);
3918 
3919 	pri_path = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
3920 	alt_path = MLX5_ADDR_OF(qpc, qpc, secondary_address_path);
3921 
3922 	if (attr_mask & IB_QP_PKEY_INDEX)
3923 		MLX5_SET(ads, pri_path, pkey_index, attr->pkey_index);
3924 
3925 	/* todo implement counter_index functionality */
3926 
3927 	if (is_sqp(ibqp->qp_type))
3928 		MLX5_SET(ads, pri_path, vhca_port_num, qp->port);
3929 
3930 	if (attr_mask & IB_QP_PORT)
3931 		MLX5_SET(ads, pri_path, vhca_port_num, attr->port_num);
3932 
3933 	if (attr_mask & IB_QP_AV) {
3934 		err = mlx5_set_path(dev, qp, &attr->ah_attr, pri_path,
3935 				    attr_mask & IB_QP_PORT ? attr->port_num :
3936 							     qp->port,
3937 				    attr_mask, 0, attr, false);
3938 		if (err)
3939 			goto out;
3940 	}
3941 
3942 	if (attr_mask & IB_QP_TIMEOUT)
3943 		MLX5_SET(ads, pri_path, ack_timeout, attr->timeout);
3944 
3945 	if (attr_mask & IB_QP_ALT_PATH) {
3946 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr, alt_path,
3947 				    attr->alt_port_num,
3948 				    attr_mask | IB_QP_PKEY_INDEX |
3949 					    IB_QP_TIMEOUT,
3950 				    0, attr, true);
3951 		if (err)
3952 			goto out;
3953 	}
3954 
3955 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
3956 		&send_cq, &recv_cq);
3957 
3958 	MLX5_SET(qpc, qpc, pd, pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
3959 	if (send_cq)
3960 		MLX5_SET(qpc, qpc, cqn_snd, send_cq->mcq.cqn);
3961 	if (recv_cq)
3962 		MLX5_SET(qpc, qpc, cqn_rcv, recv_cq->mcq.cqn);
3963 
3964 	MLX5_SET(qpc, qpc, log_ack_req_freq, MLX5_IB_ACK_REQ_FREQ);
3965 
3966 	if (attr_mask & IB_QP_RNR_RETRY)
3967 		MLX5_SET(qpc, qpc, rnr_retry, attr->rnr_retry);
3968 
3969 	if (attr_mask & IB_QP_RETRY_CNT)
3970 		MLX5_SET(qpc, qpc, retry_count, attr->retry_cnt);
3971 
3972 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC && attr->max_rd_atomic)
3973 		MLX5_SET(qpc, qpc, log_sra_max, ilog2(attr->max_rd_atomic));
3974 
3975 	if (attr_mask & IB_QP_SQ_PSN)
3976 		MLX5_SET(qpc, qpc, next_send_psn, attr->sq_psn);
3977 
3978 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC && attr->max_dest_rd_atomic)
3979 		MLX5_SET(qpc, qpc, log_rra_max,
3980 			 ilog2(attr->max_dest_rd_atomic));
3981 
3982 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
3983 		err = set_qpc_atomic_flags(qp, attr, attr_mask, qpc);
3984 		if (err)
3985 			goto out;
3986 	}
3987 
3988 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
3989 		MLX5_SET(qpc, qpc, min_rnr_nak, attr->min_rnr_timer);
3990 
3991 	if (attr_mask & IB_QP_RQ_PSN)
3992 		MLX5_SET(qpc, qpc, next_rcv_psn, attr->rq_psn);
3993 
3994 	if (attr_mask & IB_QP_QKEY)
3995 		MLX5_SET(qpc, qpc, q_key, attr->qkey);
3996 
3997 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
3998 		MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
3999 
4000 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4001 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
4002 			       qp->port) - 1;
4003 
4004 		/* Underlay port should be used - index 0 function per port */
4005 		if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
4006 			port_num = 0;
4007 
4008 		if (ibqp->counter)
4009 			set_id = ibqp->counter->id;
4010 		else
4011 			set_id = mlx5_ib_get_counters_id(dev, port_num);
4012 		MLX5_SET(qpc, qpc, counter_set_id, set_id);
4013 	}
4014 
4015 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
4016 		MLX5_SET(qpc, qpc, rlky, 1);
4017 
4018 	if (qp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
4019 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
4020 
4021 	mlx5_cur = to_mlx5_state(cur_state);
4022 	mlx5_new = to_mlx5_state(new_state);
4023 
4024 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
4025 	    !optab[mlx5_cur][mlx5_new]) {
4026 		err = -EINVAL;
4027 		goto out;
4028 	}
4029 
4030 	op = optab[mlx5_cur][mlx5_new];
4031 	optpar |= ib_mask_to_mlx5_opt(attr_mask);
4032 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
4033 
4034 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
4035 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4036 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
4037 
4038 		raw_qp_param.operation = op;
4039 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4040 			raw_qp_param.rq_q_ctr_id = set_id;
4041 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
4042 		}
4043 
4044 		if (attr_mask & IB_QP_PORT)
4045 			raw_qp_param.port = attr->port_num;
4046 
4047 		if (attr_mask & IB_QP_RATE_LIMIT) {
4048 			raw_qp_param.rl.rate = attr->rate_limit;
4049 
4050 			if (ucmd->burst_info.max_burst_sz) {
4051 				if (attr->rate_limit &&
4052 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_burst_bound)) {
4053 					raw_qp_param.rl.max_burst_sz =
4054 						ucmd->burst_info.max_burst_sz;
4055 				} else {
4056 					err = -EINVAL;
4057 					goto out;
4058 				}
4059 			}
4060 
4061 			if (ucmd->burst_info.typical_pkt_sz) {
4062 				if (attr->rate_limit &&
4063 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_typical_size)) {
4064 					raw_qp_param.rl.typical_pkt_sz =
4065 						ucmd->burst_info.typical_pkt_sz;
4066 				} else {
4067 					err = -EINVAL;
4068 					goto out;
4069 				}
4070 			}
4071 
4072 			raw_qp_param.set_mask |= MLX5_RAW_QP_RATE_LIMIT;
4073 		}
4074 
4075 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, tx_affinity);
4076 	} else {
4077 		if (udata) {
4078 			/* For the kernel flows, the resp will stay zero */
4079 			resp->ece_options =
4080 				MLX5_CAP_GEN(dev->mdev, ece_support) ?
4081 					ucmd->ece_options : 0;
4082 			resp->response_length = sizeof(*resp);
4083 		}
4084 		err = mlx5_core_qp_modify(dev, op, optpar, qpc, &base->mqp,
4085 					  &resp->ece_options);
4086 	}
4087 
4088 	if (err)
4089 		goto out;
4090 
4091 	qp->state = new_state;
4092 
4093 	if (attr_mask & IB_QP_ACCESS_FLAGS)
4094 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
4095 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
4096 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
4097 	if (attr_mask & IB_QP_PORT)
4098 		qp->port = attr->port_num;
4099 	if (attr_mask & IB_QP_ALT_PATH)
4100 		qp->trans_qp.alt_port = attr->alt_port_num;
4101 
4102 	/*
4103 	 * If we moved a kernel QP to RESET, clean up all old CQ
4104 	 * entries and reinitialize the QP.
4105 	 */
4106 	if (new_state == IB_QPS_RESET &&
4107 	    !ibqp->uobject && ibqp->qp_type != IB_QPT_XRC_TGT) {
4108 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
4109 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
4110 		if (send_cq != recv_cq)
4111 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
4112 
4113 		qp->rq.head = 0;
4114 		qp->rq.tail = 0;
4115 		qp->sq.head = 0;
4116 		qp->sq.tail = 0;
4117 		qp->sq.cur_post = 0;
4118 		if (qp->sq.wqe_cnt)
4119 			qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
4120 		qp->sq.last_poll = 0;
4121 		qp->db.db[MLX5_RCV_DBR] = 0;
4122 		qp->db.db[MLX5_SND_DBR] = 0;
4123 	}
4124 
4125 	if ((new_state == IB_QPS_RTS) && qp->counter_pending) {
4126 		err = __mlx5_ib_qp_set_counter(ibqp, ibqp->counter);
4127 		if (!err)
4128 			qp->counter_pending = 0;
4129 	}
4130 
4131 out:
4132 	kfree(qpc);
4133 	return err;
4134 }
4135 
4136 static inline bool is_valid_mask(int mask, int req, int opt)
4137 {
4138 	if ((mask & req) != req)
4139 		return false;
4140 
4141 	if (mask & ~(req | opt))
4142 		return false;
4143 
4144 	return true;
4145 }
4146 
4147 /* check valid transition for driver QP types
4148  * for now the only QP type that this function supports is DCI
4149  */
4150 static bool modify_dci_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state new_state,
4151 				enum ib_qp_attr_mask attr_mask)
4152 {
4153 	int req = IB_QP_STATE;
4154 	int opt = 0;
4155 
4156 	if (new_state == IB_QPS_RESET) {
4157 		return is_valid_mask(attr_mask, req, opt);
4158 	} else if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4159 		req |= IB_QP_PKEY_INDEX | IB_QP_PORT;
4160 		return is_valid_mask(attr_mask, req, opt);
4161 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
4162 		opt = IB_QP_PKEY_INDEX | IB_QP_PORT;
4163 		return is_valid_mask(attr_mask, req, opt);
4164 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
4165 		req |= IB_QP_PATH_MTU;
4166 		opt = IB_QP_PKEY_INDEX | IB_QP_AV;
4167 		return is_valid_mask(attr_mask, req, opt);
4168 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
4169 		req |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4170 		       IB_QP_MAX_QP_RD_ATOMIC | IB_QP_SQ_PSN;
4171 		opt = IB_QP_MIN_RNR_TIMER;
4172 		return is_valid_mask(attr_mask, req, opt);
4173 	} else if (cur_state == IB_QPS_RTS && new_state == IB_QPS_RTS) {
4174 		opt = IB_QP_MIN_RNR_TIMER;
4175 		return is_valid_mask(attr_mask, req, opt);
4176 	} else if (cur_state != IB_QPS_RESET && new_state == IB_QPS_ERR) {
4177 		return is_valid_mask(attr_mask, req, opt);
4178 	}
4179 	return false;
4180 }
4181 
4182 /* mlx5_ib_modify_dct: modify a DCT QP
4183  * valid transitions are:
4184  * RESET to INIT: must set access_flags, pkey_index and port
4185  * INIT  to RTR : must set min_rnr_timer, tclass, flow_label,
4186  *			   mtu, gid_index and hop_limit
4187  * Other transitions and attributes are illegal
4188  */
4189 static int mlx5_ib_modify_dct(struct ib_qp *ibqp, struct ib_qp_attr *attr,
4190 			      int attr_mask, struct mlx5_ib_modify_qp *ucmd,
4191 			      struct ib_udata *udata)
4192 {
4193 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4194 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4195 	enum ib_qp_state cur_state, new_state;
4196 	int required = IB_QP_STATE;
4197 	void *dctc;
4198 	int err;
4199 
4200 	if (!(attr_mask & IB_QP_STATE))
4201 		return -EINVAL;
4202 
4203 	cur_state = qp->state;
4204 	new_state = attr->qp_state;
4205 
4206 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
4207 	if (MLX5_CAP_GEN(dev->mdev, ece_support) && ucmd->ece_options)
4208 		/*
4209 		 * DCT doesn't initialize QP till modify command is executed,
4210 		 * so we need to overwrite previously set ECE field if user
4211 		 * provided any value except zero, which means not set/not
4212 		 * valid.
4213 		 */
4214 		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
4215 
4216 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4217 		u16 set_id;
4218 
4219 		required |= IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
4220 		if (!is_valid_mask(attr_mask, required, 0))
4221 			return -EINVAL;
4222 
4223 		if (attr->port_num == 0 ||
4224 		    attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)) {
4225 			mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4226 				    attr->port_num, dev->num_ports);
4227 			return -EINVAL;
4228 		}
4229 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
4230 			MLX5_SET(dctc, dctc, rre, 1);
4231 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
4232 			MLX5_SET(dctc, dctc, rwe, 1);
4233 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC) {
4234 			int atomic_mode;
4235 
4236 			atomic_mode = get_atomic_mode(dev, MLX5_IB_QPT_DCT);
4237 			if (atomic_mode < 0)
4238 				return -EOPNOTSUPP;
4239 
4240 			MLX5_SET(dctc, dctc, atomic_mode, atomic_mode);
4241 			MLX5_SET(dctc, dctc, rae, 1);
4242 		}
4243 		MLX5_SET(dctc, dctc, pkey_index, attr->pkey_index);
4244 		if (mlx5_lag_is_active(dev->mdev))
4245 			MLX5_SET(dctc, dctc, port,
4246 				 get_tx_affinity_rr(dev, udata));
4247 		else
4248 			MLX5_SET(dctc, dctc, port, attr->port_num);
4249 
4250 		set_id = mlx5_ib_get_counters_id(dev, attr->port_num - 1);
4251 		MLX5_SET(dctc, dctc, counter_set_id, set_id);
4252 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
4253 		struct mlx5_ib_modify_qp_resp resp = {};
4254 		u32 out[MLX5_ST_SZ_DW(create_dct_out)] = {};
4255 		u32 min_resp_len = offsetofend(typeof(resp), dctn);
4256 
4257 		if (udata->outlen < min_resp_len)
4258 			return -EINVAL;
4259 		/*
4260 		 * If we don't have enough space for the ECE options,
4261 		 * simply indicate it with resp.response_length.
4262 		 */
4263 		resp.response_length = (udata->outlen < sizeof(resp)) ?
4264 					       min_resp_len :
4265 					       sizeof(resp);
4266 
4267 		required |= IB_QP_MIN_RNR_TIMER | IB_QP_AV | IB_QP_PATH_MTU;
4268 		if (!is_valid_mask(attr_mask, required, 0))
4269 			return -EINVAL;
4270 		MLX5_SET(dctc, dctc, min_rnr_nak, attr->min_rnr_timer);
4271 		MLX5_SET(dctc, dctc, tclass, attr->ah_attr.grh.traffic_class);
4272 		MLX5_SET(dctc, dctc, flow_label, attr->ah_attr.grh.flow_label);
4273 		MLX5_SET(dctc, dctc, mtu, attr->path_mtu);
4274 		MLX5_SET(dctc, dctc, my_addr_index, attr->ah_attr.grh.sgid_index);
4275 		MLX5_SET(dctc, dctc, hop_limit, attr->ah_attr.grh.hop_limit);
4276 
4277 		err = mlx5_core_create_dct(dev, &qp->dct.mdct, qp->dct.in,
4278 					   MLX5_ST_SZ_BYTES(create_dct_in), out,
4279 					   sizeof(out));
4280 		if (err)
4281 			return err;
4282 		resp.dctn = qp->dct.mdct.mqp.qpn;
4283 		if (MLX5_CAP_GEN(dev->mdev, ece_support))
4284 			resp.ece_options = MLX5_GET(create_dct_out, out, ece);
4285 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4286 		if (err) {
4287 			mlx5_core_destroy_dct(dev, &qp->dct.mdct);
4288 			return err;
4289 		}
4290 	} else {
4291 		mlx5_ib_warn(dev, "Modify DCT: Invalid transition from %d to %d\n", cur_state, new_state);
4292 		return -EINVAL;
4293 	}
4294 
4295 	qp->state = new_state;
4296 	return 0;
4297 }
4298 
4299 static bool mlx5_ib_modify_qp_allowed(struct mlx5_ib_dev *dev,
4300 				      struct mlx5_ib_qp *qp,
4301 				      enum ib_qp_type qp_type)
4302 {
4303 	if (dev->profile != &raw_eth_profile)
4304 		return true;
4305 
4306 	if (qp_type == IB_QPT_RAW_PACKET || qp_type == MLX5_IB_QPT_REG_UMR)
4307 		return true;
4308 
4309 	/* Internal QP used for wc testing, with NOPs in wq */
4310 	if (qp->flags & MLX5_IB_QP_CREATE_WC_TEST)
4311 		return true;
4312 
4313 	return false;
4314 }
4315 
4316 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
4317 		      int attr_mask, struct ib_udata *udata)
4318 {
4319 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4320 	struct mlx5_ib_modify_qp_resp resp = {};
4321 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4322 	struct mlx5_ib_modify_qp ucmd = {};
4323 	enum ib_qp_type qp_type;
4324 	enum ib_qp_state cur_state, new_state;
4325 	int err = -EINVAL;
4326 
4327 	if (!mlx5_ib_modify_qp_allowed(dev, qp, ibqp->qp_type))
4328 		return -EOPNOTSUPP;
4329 
4330 	if (attr_mask & ~(IB_QP_ATTR_STANDARD_BITS | IB_QP_RATE_LIMIT))
4331 		return -EOPNOTSUPP;
4332 
4333 	if (ibqp->rwq_ind_tbl)
4334 		return -ENOSYS;
4335 
4336 	if (udata && udata->inlen) {
4337 		if (udata->inlen < offsetofend(typeof(ucmd), ece_options))
4338 			return -EINVAL;
4339 
4340 		if (udata->inlen > sizeof(ucmd) &&
4341 		    !ib_is_udata_cleared(udata, sizeof(ucmd),
4342 					 udata->inlen - sizeof(ucmd)))
4343 			return -EOPNOTSUPP;
4344 
4345 		if (ib_copy_from_udata(&ucmd, udata,
4346 				       min(udata->inlen, sizeof(ucmd))))
4347 			return -EFAULT;
4348 
4349 		if (ucmd.comp_mask ||
4350 		    memchr_inv(&ucmd.burst_info.reserved, 0,
4351 			       sizeof(ucmd.burst_info.reserved)))
4352 			return -EOPNOTSUPP;
4353 
4354 	}
4355 
4356 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4357 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
4358 
4359 	qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ? IB_QPT_GSI :
4360 								    qp->type;
4361 
4362 	if (qp_type == MLX5_IB_QPT_DCT)
4363 		return mlx5_ib_modify_dct(ibqp, attr, attr_mask, &ucmd, udata);
4364 
4365 	mutex_lock(&qp->mutex);
4366 
4367 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
4368 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
4369 
4370 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4371 		if (attr_mask & ~(IB_QP_STATE | IB_QP_CUR_STATE)) {
4372 			mlx5_ib_dbg(dev, "invalid attr_mask 0x%x when underlay QP is used\n",
4373 				    attr_mask);
4374 			goto out;
4375 		}
4376 	} else if (qp_type != MLX5_IB_QPT_REG_UMR &&
4377 		   qp_type != MLX5_IB_QPT_DCI &&
4378 		   !ib_modify_qp_is_ok(cur_state, new_state, qp_type,
4379 				       attr_mask)) {
4380 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4381 			    cur_state, new_state, ibqp->qp_type, attr_mask);
4382 		goto out;
4383 	} else if (qp_type == MLX5_IB_QPT_DCI &&
4384 		   !modify_dci_qp_is_ok(cur_state, new_state, attr_mask)) {
4385 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4386 			    cur_state, new_state, qp_type, attr_mask);
4387 		goto out;
4388 	}
4389 
4390 	if ((attr_mask & IB_QP_PORT) &&
4391 	    (attr->port_num == 0 ||
4392 	     attr->port_num > dev->num_ports)) {
4393 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4394 			    attr->port_num, dev->num_ports);
4395 		goto out;
4396 	}
4397 
4398 	if ((attr_mask & IB_QP_PKEY_INDEX) &&
4399 	    attr->pkey_index >= dev->pkey_table_len) {
4400 		mlx5_ib_dbg(dev, "invalid pkey index %d\n", attr->pkey_index);
4401 		goto out;
4402 	}
4403 
4404 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
4405 	    attr->max_rd_atomic >
4406 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
4407 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
4408 			    attr->max_rd_atomic);
4409 		goto out;
4410 	}
4411 
4412 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
4413 	    attr->max_dest_rd_atomic >
4414 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
4415 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
4416 			    attr->max_dest_rd_atomic);
4417 		goto out;
4418 	}
4419 
4420 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
4421 		err = 0;
4422 		goto out;
4423 	}
4424 
4425 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state,
4426 				  new_state, &ucmd, &resp, udata);
4427 
4428 	/* resp.response_length is set in ECE supported flows only */
4429 	if (!err && resp.response_length &&
4430 	    udata->outlen >= resp.response_length)
4431 		/* Return -EFAULT to the user and expect him to destroy QP. */
4432 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4433 
4434 out:
4435 	mutex_unlock(&qp->mutex);
4436 	return err;
4437 }
4438 
4439 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4440 {
4441 	switch (mlx5_state) {
4442 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
4443 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
4444 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
4445 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
4446 	case MLX5_QP_STATE_SQ_DRAINING:
4447 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
4448 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
4449 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
4450 	default:		     return -1;
4451 	}
4452 }
4453 
4454 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4455 {
4456 	switch (mlx5_mig_state) {
4457 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
4458 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
4459 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4460 	default: return -1;
4461 	}
4462 }
4463 
4464 static void to_rdma_ah_attr(struct mlx5_ib_dev *ibdev,
4465 			    struct rdma_ah_attr *ah_attr, void *path)
4466 {
4467 	int port = MLX5_GET(ads, path, vhca_port_num);
4468 	int static_rate;
4469 
4470 	memset(ah_attr, 0, sizeof(*ah_attr));
4471 
4472 	if (!port || port > ibdev->num_ports)
4473 		return;
4474 
4475 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port);
4476 
4477 	rdma_ah_set_port_num(ah_attr, port);
4478 	rdma_ah_set_sl(ah_attr, MLX5_GET(ads, path, sl));
4479 
4480 	rdma_ah_set_dlid(ah_attr, MLX5_GET(ads, path, rlid));
4481 	rdma_ah_set_path_bits(ah_attr, MLX5_GET(ads, path, mlid));
4482 
4483 	static_rate = MLX5_GET(ads, path, stat_rate);
4484 	rdma_ah_set_static_rate(ah_attr, static_rate ? static_rate - 5 : 0);
4485 	if (MLX5_GET(ads, path, grh) ||
4486 	    ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
4487 		rdma_ah_set_grh(ah_attr, NULL, MLX5_GET(ads, path, flow_label),
4488 				MLX5_GET(ads, path, src_addr_index),
4489 				MLX5_GET(ads, path, hop_limit),
4490 				MLX5_GET(ads, path, tclass));
4491 		rdma_ah_set_dgid_raw(ah_attr, MLX5_ADDR_OF(ads, path, rgid_rip));
4492 	}
4493 }
4494 
4495 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4496 					struct mlx5_ib_sq *sq,
4497 					u8 *sq_state)
4498 {
4499 	int err;
4500 
4501 	err = mlx5_core_query_sq_state(dev->mdev, sq->base.mqp.qpn, sq_state);
4502 	if (err)
4503 		goto out;
4504 	sq->state = *sq_state;
4505 
4506 out:
4507 	return err;
4508 }
4509 
4510 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4511 					struct mlx5_ib_rq *rq,
4512 					u8 *rq_state)
4513 {
4514 	void *out;
4515 	void *rqc;
4516 	int inlen;
4517 	int err;
4518 
4519 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4520 	out = kvzalloc(inlen, GFP_KERNEL);
4521 	if (!out)
4522 		return -ENOMEM;
4523 
4524 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4525 	if (err)
4526 		goto out;
4527 
4528 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4529 	*rq_state = MLX5_GET(rqc, rqc, state);
4530 	rq->state = *rq_state;
4531 
4532 out:
4533 	kvfree(out);
4534 	return err;
4535 }
4536 
4537 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4538 				  struct mlx5_ib_qp *qp, u8 *qp_state)
4539 {
4540 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4541 		[MLX5_RQC_STATE_RST] = {
4542 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
4543 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4544 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
4545 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
4546 		},
4547 		[MLX5_RQC_STATE_RDY] = {
4548 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE,
4549 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4550 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
4551 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
4552 		},
4553 		[MLX5_RQC_STATE_ERR] = {
4554 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
4555 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4556 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
4557 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
4558 		},
4559 		[MLX5_RQ_STATE_NA] = {
4560 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE,
4561 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4562 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
4563 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
4564 		},
4565 	};
4566 
4567 	*qp_state = sqrq_trans[rq_state][sq_state];
4568 
4569 	if (*qp_state == MLX5_QP_STATE_BAD) {
4570 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4571 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4572 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4573 		return -EINVAL;
4574 	}
4575 
4576 	if (*qp_state == MLX5_QP_STATE)
4577 		*qp_state = qp->state;
4578 
4579 	return 0;
4580 }
4581 
4582 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4583 				     struct mlx5_ib_qp *qp,
4584 				     u8 *raw_packet_qp_state)
4585 {
4586 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4587 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4588 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4589 	int err;
4590 	u8 sq_state = MLX5_SQ_STATE_NA;
4591 	u8 rq_state = MLX5_RQ_STATE_NA;
4592 
4593 	if (qp->sq.wqe_cnt) {
4594 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4595 		if (err)
4596 			return err;
4597 	}
4598 
4599 	if (qp->rq.wqe_cnt) {
4600 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4601 		if (err)
4602 			return err;
4603 	}
4604 
4605 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4606 				      raw_packet_qp_state);
4607 }
4608 
4609 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4610 			 struct ib_qp_attr *qp_attr)
4611 {
4612 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
4613 	void *qpc, *pri_path, *alt_path;
4614 	u32 *outb;
4615 	int err;
4616 
4617 	outb = kzalloc(outlen, GFP_KERNEL);
4618 	if (!outb)
4619 		return -ENOMEM;
4620 
4621 	err = mlx5_core_qp_query(dev, &qp->trans_qp.base.mqp, outb, outlen);
4622 	if (err)
4623 		goto out;
4624 
4625 	qpc = MLX5_ADDR_OF(query_qp_out, outb, qpc);
4626 
4627 	qp->state = to_ib_qp_state(MLX5_GET(qpc, qpc, state));
4628 	if (MLX5_GET(qpc, qpc, state) == MLX5_QP_STATE_SQ_DRAINING)
4629 		qp_attr->sq_draining = 1;
4630 
4631 	qp_attr->path_mtu = MLX5_GET(qpc, qpc, mtu);
4632 	qp_attr->path_mig_state = to_ib_mig_state(MLX5_GET(qpc, qpc, pm_state));
4633 	qp_attr->qkey = MLX5_GET(qpc, qpc, q_key);
4634 	qp_attr->rq_psn = MLX5_GET(qpc, qpc, next_rcv_psn);
4635 	qp_attr->sq_psn = MLX5_GET(qpc, qpc, next_send_psn);
4636 	qp_attr->dest_qp_num = MLX5_GET(qpc, qpc, remote_qpn);
4637 
4638 	if (MLX5_GET(qpc, qpc, rre))
4639 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ;
4640 	if (MLX5_GET(qpc, qpc, rwe))
4641 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_WRITE;
4642 	if (MLX5_GET(qpc, qpc, rae))
4643 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_ATOMIC;
4644 
4645 	qp_attr->max_rd_atomic = 1 << MLX5_GET(qpc, qpc, log_sra_max);
4646 	qp_attr->max_dest_rd_atomic = 1 << MLX5_GET(qpc, qpc, log_rra_max);
4647 	qp_attr->min_rnr_timer = MLX5_GET(qpc, qpc, min_rnr_nak);
4648 	qp_attr->retry_cnt = MLX5_GET(qpc, qpc, retry_count);
4649 	qp_attr->rnr_retry = MLX5_GET(qpc, qpc, rnr_retry);
4650 
4651 	pri_path = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
4652 	alt_path = MLX5_ADDR_OF(qpc, qpc, secondary_address_path);
4653 
4654 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC ||
4655 	    qp->ibqp.qp_type == IB_QPT_XRC_INI ||
4656 	    qp->ibqp.qp_type == IB_QPT_XRC_TGT) {
4657 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, pri_path);
4658 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, alt_path);
4659 		qp_attr->alt_pkey_index = MLX5_GET(ads, alt_path, pkey_index);
4660 		qp_attr->alt_port_num = MLX5_GET(ads, alt_path, vhca_port_num);
4661 	}
4662 
4663 	qp_attr->pkey_index = MLX5_GET(ads, pri_path, pkey_index);
4664 	qp_attr->port_num = MLX5_GET(ads, pri_path, vhca_port_num);
4665 	qp_attr->timeout = MLX5_GET(ads, pri_path, ack_timeout);
4666 	qp_attr->alt_timeout = MLX5_GET(ads, alt_path, ack_timeout);
4667 
4668 out:
4669 	kfree(outb);
4670 	return err;
4671 }
4672 
4673 static int mlx5_ib_dct_query_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *mqp,
4674 				struct ib_qp_attr *qp_attr, int qp_attr_mask,
4675 				struct ib_qp_init_attr *qp_init_attr)
4676 {
4677 	struct mlx5_core_dct	*dct = &mqp->dct.mdct;
4678 	u32 *out;
4679 	u32 access_flags = 0;
4680 	int outlen = MLX5_ST_SZ_BYTES(query_dct_out);
4681 	void *dctc;
4682 	int err;
4683 	int supported_mask = IB_QP_STATE |
4684 			     IB_QP_ACCESS_FLAGS |
4685 			     IB_QP_PORT |
4686 			     IB_QP_MIN_RNR_TIMER |
4687 			     IB_QP_AV |
4688 			     IB_QP_PATH_MTU |
4689 			     IB_QP_PKEY_INDEX;
4690 
4691 	if (qp_attr_mask & ~supported_mask)
4692 		return -EINVAL;
4693 	if (mqp->state != IB_QPS_RTR)
4694 		return -EINVAL;
4695 
4696 	out = kzalloc(outlen, GFP_KERNEL);
4697 	if (!out)
4698 		return -ENOMEM;
4699 
4700 	err = mlx5_core_dct_query(dev, dct, out, outlen);
4701 	if (err)
4702 		goto out;
4703 
4704 	dctc = MLX5_ADDR_OF(query_dct_out, out, dct_context_entry);
4705 
4706 	if (qp_attr_mask & IB_QP_STATE)
4707 		qp_attr->qp_state = IB_QPS_RTR;
4708 
4709 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
4710 		if (MLX5_GET(dctc, dctc, rre))
4711 			access_flags |= IB_ACCESS_REMOTE_READ;
4712 		if (MLX5_GET(dctc, dctc, rwe))
4713 			access_flags |= IB_ACCESS_REMOTE_WRITE;
4714 		if (MLX5_GET(dctc, dctc, rae))
4715 			access_flags |= IB_ACCESS_REMOTE_ATOMIC;
4716 		qp_attr->qp_access_flags = access_flags;
4717 	}
4718 
4719 	if (qp_attr_mask & IB_QP_PORT)
4720 		qp_attr->port_num = MLX5_GET(dctc, dctc, port);
4721 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER)
4722 		qp_attr->min_rnr_timer = MLX5_GET(dctc, dctc, min_rnr_nak);
4723 	if (qp_attr_mask & IB_QP_AV) {
4724 		qp_attr->ah_attr.grh.traffic_class = MLX5_GET(dctc, dctc, tclass);
4725 		qp_attr->ah_attr.grh.flow_label = MLX5_GET(dctc, dctc, flow_label);
4726 		qp_attr->ah_attr.grh.sgid_index = MLX5_GET(dctc, dctc, my_addr_index);
4727 		qp_attr->ah_attr.grh.hop_limit = MLX5_GET(dctc, dctc, hop_limit);
4728 	}
4729 	if (qp_attr_mask & IB_QP_PATH_MTU)
4730 		qp_attr->path_mtu = MLX5_GET(dctc, dctc, mtu);
4731 	if (qp_attr_mask & IB_QP_PKEY_INDEX)
4732 		qp_attr->pkey_index = MLX5_GET(dctc, dctc, pkey_index);
4733 out:
4734 	kfree(out);
4735 	return err;
4736 }
4737 
4738 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4739 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4740 {
4741 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4742 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4743 	int err = 0;
4744 	u8 raw_packet_qp_state;
4745 
4746 	if (ibqp->rwq_ind_tbl)
4747 		return -ENOSYS;
4748 
4749 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4750 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4751 					    qp_init_attr);
4752 
4753 	/* Not all of output fields are applicable, make sure to zero them */
4754 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
4755 	memset(qp_attr, 0, sizeof(*qp_attr));
4756 
4757 	if (unlikely(qp->type == MLX5_IB_QPT_DCT))
4758 		return mlx5_ib_dct_query_qp(dev, qp, qp_attr,
4759 					    qp_attr_mask, qp_init_attr);
4760 
4761 	mutex_lock(&qp->mutex);
4762 
4763 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
4764 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4765 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4766 		if (err)
4767 			goto out;
4768 		qp->state = raw_packet_qp_state;
4769 		qp_attr->port_num = 1;
4770 	} else {
4771 		err = query_qp_attr(dev, qp, qp_attr);
4772 		if (err)
4773 			goto out;
4774 	}
4775 
4776 	qp_attr->qp_state	     = qp->state;
4777 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4778 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4779 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4780 
4781 	if (!ibqp->uobject) {
4782 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
4783 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4784 		qp_init_attr->qp_context = ibqp->qp_context;
4785 	} else {
4786 		qp_attr->cap.max_send_wr  = 0;
4787 		qp_attr->cap.max_send_sge = 0;
4788 	}
4789 
4790 	qp_init_attr->qp_type = ibqp->qp_type;
4791 	qp_init_attr->recv_cq = ibqp->recv_cq;
4792 	qp_init_attr->send_cq = ibqp->send_cq;
4793 	qp_init_attr->srq = ibqp->srq;
4794 	qp_attr->cap.max_inline_data = qp->max_inline_data;
4795 
4796 	qp_init_attr->cap	     = qp_attr->cap;
4797 
4798 	qp_init_attr->create_flags = qp->flags;
4799 
4800 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
4801 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4802 
4803 out:
4804 	mutex_unlock(&qp->mutex);
4805 	return err;
4806 }
4807 
4808 int mlx5_ib_alloc_xrcd(struct ib_xrcd *ibxrcd, struct ib_udata *udata)
4809 {
4810 	struct mlx5_ib_dev *dev = to_mdev(ibxrcd->device);
4811 	struct mlx5_ib_xrcd *xrcd = to_mxrcd(ibxrcd);
4812 
4813 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
4814 		return -EOPNOTSUPP;
4815 
4816 	return mlx5_cmd_xrcd_alloc(dev->mdev, &xrcd->xrcdn, 0);
4817 }
4818 
4819 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
4820 {
4821 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
4822 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
4823 
4824 	return mlx5_cmd_xrcd_dealloc(dev->mdev, xrcdn, 0);
4825 }
4826 
4827 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
4828 {
4829 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
4830 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
4831 	struct ib_event event;
4832 
4833 	if (rwq->ibwq.event_handler) {
4834 		event.device     = rwq->ibwq.device;
4835 		event.element.wq = &rwq->ibwq;
4836 		switch (type) {
4837 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
4838 			event.event = IB_EVENT_WQ_FATAL;
4839 			break;
4840 		default:
4841 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
4842 			return;
4843 		}
4844 
4845 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
4846 	}
4847 }
4848 
4849 static int set_delay_drop(struct mlx5_ib_dev *dev)
4850 {
4851 	int err = 0;
4852 
4853 	mutex_lock(&dev->delay_drop.lock);
4854 	if (dev->delay_drop.activate)
4855 		goto out;
4856 
4857 	err = mlx5_core_set_delay_drop(dev, dev->delay_drop.timeout);
4858 	if (err)
4859 		goto out;
4860 
4861 	dev->delay_drop.activate = true;
4862 out:
4863 	mutex_unlock(&dev->delay_drop.lock);
4864 
4865 	if (!err)
4866 		atomic_inc(&dev->delay_drop.rqs_cnt);
4867 	return err;
4868 }
4869 
4870 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
4871 		      struct ib_wq_init_attr *init_attr)
4872 {
4873 	struct mlx5_ib_dev *dev;
4874 	int has_net_offloads;
4875 	__be64 *rq_pas0;
4876 	void *in;
4877 	void *rqc;
4878 	void *wq;
4879 	int inlen;
4880 	int err;
4881 
4882 	dev = to_mdev(pd->device);
4883 
4884 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
4885 	in = kvzalloc(inlen, GFP_KERNEL);
4886 	if (!in)
4887 		return -ENOMEM;
4888 
4889 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
4890 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
4891 	MLX5_SET(rqc,  rqc, mem_rq_type,
4892 		 MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
4893 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
4894 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
4895 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
4896 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
4897 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
4898 	MLX5_SET(wq, wq, wq_type,
4899 		 rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ ?
4900 		 MLX5_WQ_TYPE_CYCLIC_STRIDING_RQ : MLX5_WQ_TYPE_CYCLIC);
4901 	if (init_attr->create_flags & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
4902 		if (!MLX5_CAP_GEN(dev->mdev, end_pad)) {
4903 			mlx5_ib_dbg(dev, "Scatter end padding is not supported\n");
4904 			err = -EOPNOTSUPP;
4905 			goto out;
4906 		} else {
4907 			MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
4908 		}
4909 	}
4910 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
4911 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ) {
4912 		/*
4913 		 * In Firmware number of strides in each WQE is:
4914 		 *   "512 * 2^single_wqe_log_num_of_strides"
4915 		 * Values 3 to 8 are accepted as 10 to 15, 9 to 18 are
4916 		 * accepted as 0 to 9
4917 		 */
4918 		static const u8 fw_map[] = { 10, 11, 12, 13, 14, 15, 0, 1,
4919 					     2,  3,  4,  5,  6,  7,  8, 9 };
4920 		MLX5_SET(wq, wq, two_byte_shift_en, rwq->two_byte_shift_en);
4921 		MLX5_SET(wq, wq, log_wqe_stride_size,
4922 			 rwq->single_stride_log_num_of_bytes -
4923 			 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES);
4924 		MLX5_SET(wq, wq, log_wqe_num_of_strides,
4925 			 fw_map[rwq->log_num_strides -
4926 				MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES]);
4927 	}
4928 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
4929 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
4930 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
4931 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
4932 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
4933 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
4934 	has_net_offloads = MLX5_CAP_GEN(dev->mdev, eth_net_offloads);
4935 	if (init_attr->create_flags & IB_WQ_FLAGS_CVLAN_STRIPPING) {
4936 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
4937 			mlx5_ib_dbg(dev, "VLAN offloads are not supported\n");
4938 			err = -EOPNOTSUPP;
4939 			goto out;
4940 		}
4941 	} else {
4942 		MLX5_SET(rqc, rqc, vsd, 1);
4943 	}
4944 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS) {
4945 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, scatter_fcs))) {
4946 			mlx5_ib_dbg(dev, "Scatter FCS is not supported\n");
4947 			err = -EOPNOTSUPP;
4948 			goto out;
4949 		}
4950 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
4951 	}
4952 	if (init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
4953 		if (!(dev->ib_dev.attrs.raw_packet_caps &
4954 		      IB_RAW_PACKET_CAP_DELAY_DROP)) {
4955 			mlx5_ib_dbg(dev, "Delay drop is not supported\n");
4956 			err = -EOPNOTSUPP;
4957 			goto out;
4958 		}
4959 		MLX5_SET(rqc, rqc, delay_drop_en, 1);
4960 	}
4961 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
4962 	mlx5_ib_populate_pas(rwq->umem, 1UL << rwq->page_shift, rq_pas0, 0);
4963 	err = mlx5_core_create_rq_tracked(dev, in, inlen, &rwq->core_qp);
4964 	if (!err && init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
4965 		err = set_delay_drop(dev);
4966 		if (err) {
4967 			mlx5_ib_warn(dev, "Failed to enable delay drop err=%d\n",
4968 				     err);
4969 			mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
4970 		} else {
4971 			rwq->create_flags |= MLX5_IB_WQ_FLAGS_DELAY_DROP;
4972 		}
4973 	}
4974 out:
4975 	kvfree(in);
4976 	return err;
4977 }
4978 
4979 static int set_user_rq_size(struct mlx5_ib_dev *dev,
4980 			    struct ib_wq_init_attr *wq_init_attr,
4981 			    struct mlx5_ib_create_wq *ucmd,
4982 			    struct mlx5_ib_rwq *rwq)
4983 {
4984 	/* Sanity check RQ size before proceeding */
4985 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
4986 		return -EINVAL;
4987 
4988 	if (!ucmd->rq_wqe_count)
4989 		return -EINVAL;
4990 
4991 	rwq->wqe_count = ucmd->rq_wqe_count;
4992 	rwq->wqe_shift = ucmd->rq_wqe_shift;
4993 	if (check_shl_overflow(rwq->wqe_count, rwq->wqe_shift, &rwq->buf_size))
4994 		return -EINVAL;
4995 
4996 	rwq->log_rq_stride = rwq->wqe_shift;
4997 	rwq->log_rq_size = ilog2(rwq->wqe_count);
4998 	return 0;
4999 }
5000 
5001 static bool log_of_strides_valid(struct mlx5_ib_dev *dev, u32 log_num_strides)
5002 {
5003 	if ((log_num_strides > MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES) ||
5004 	    (log_num_strides < MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
5005 		return false;
5006 
5007 	if (!MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) &&
5008 	    (log_num_strides < MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
5009 		return false;
5010 
5011 	return true;
5012 }
5013 
5014 static int prepare_user_rq(struct ib_pd *pd,
5015 			   struct ib_wq_init_attr *init_attr,
5016 			   struct ib_udata *udata,
5017 			   struct mlx5_ib_rwq *rwq)
5018 {
5019 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
5020 	struct mlx5_ib_create_wq ucmd = {};
5021 	int err;
5022 	size_t required_cmd_sz;
5023 
5024 	required_cmd_sz = offsetofend(struct mlx5_ib_create_wq,
5025 				      single_stride_log_num_of_bytes);
5026 	if (udata->inlen < required_cmd_sz) {
5027 		mlx5_ib_dbg(dev, "invalid inlen\n");
5028 		return -EINVAL;
5029 	}
5030 
5031 	if (udata->inlen > sizeof(ucmd) &&
5032 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
5033 				 udata->inlen - sizeof(ucmd))) {
5034 		mlx5_ib_dbg(dev, "inlen is not supported\n");
5035 		return -EOPNOTSUPP;
5036 	}
5037 
5038 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
5039 		mlx5_ib_dbg(dev, "copy failed\n");
5040 		return -EFAULT;
5041 	}
5042 
5043 	if (ucmd.comp_mask & (~MLX5_IB_CREATE_WQ_STRIDING_RQ)) {
5044 		mlx5_ib_dbg(dev, "invalid comp mask\n");
5045 		return -EOPNOTSUPP;
5046 	} else if (ucmd.comp_mask & MLX5_IB_CREATE_WQ_STRIDING_RQ) {
5047 		if (!MLX5_CAP_GEN(dev->mdev, striding_rq)) {
5048 			mlx5_ib_dbg(dev, "Striding RQ is not supported\n");
5049 			return -EOPNOTSUPP;
5050 		}
5051 		if ((ucmd.single_stride_log_num_of_bytes <
5052 		    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES) ||
5053 		    (ucmd.single_stride_log_num_of_bytes >
5054 		     MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES)) {
5055 			mlx5_ib_dbg(dev, "Invalid log stride size (%u. Range is %u - %u)\n",
5056 				    ucmd.single_stride_log_num_of_bytes,
5057 				    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES,
5058 				    MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES);
5059 			return -EINVAL;
5060 		}
5061 		if (!log_of_strides_valid(dev,
5062 					  ucmd.single_wqe_log_num_of_strides)) {
5063 			mlx5_ib_dbg(
5064 				dev,
5065 				"Invalid log num strides (%u. Range is %u - %u)\n",
5066 				ucmd.single_wqe_log_num_of_strides,
5067 				MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) ?
5068 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES :
5069 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES,
5070 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES);
5071 			return -EINVAL;
5072 		}
5073 		rwq->single_stride_log_num_of_bytes =
5074 			ucmd.single_stride_log_num_of_bytes;
5075 		rwq->log_num_strides = ucmd.single_wqe_log_num_of_strides;
5076 		rwq->two_byte_shift_en = !!ucmd.two_byte_shift_en;
5077 		rwq->create_flags |= MLX5_IB_WQ_FLAGS_STRIDING_RQ;
5078 	}
5079 
5080 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
5081 	if (err) {
5082 		mlx5_ib_dbg(dev, "err %d\n", err);
5083 		return err;
5084 	}
5085 
5086 	err = create_user_rq(dev, pd, udata, rwq, &ucmd);
5087 	if (err) {
5088 		mlx5_ib_dbg(dev, "err %d\n", err);
5089 		return err;
5090 	}
5091 
5092 	rwq->user_index = ucmd.user_index;
5093 	return 0;
5094 }
5095 
5096 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
5097 				struct ib_wq_init_attr *init_attr,
5098 				struct ib_udata *udata)
5099 {
5100 	struct mlx5_ib_dev *dev;
5101 	struct mlx5_ib_rwq *rwq;
5102 	struct mlx5_ib_create_wq_resp resp = {};
5103 	size_t min_resp_len;
5104 	int err;
5105 
5106 	if (!udata)
5107 		return ERR_PTR(-ENOSYS);
5108 
5109 	min_resp_len = offsetofend(struct mlx5_ib_create_wq_resp, reserved);
5110 	if (udata->outlen && udata->outlen < min_resp_len)
5111 		return ERR_PTR(-EINVAL);
5112 
5113 	if (!capable(CAP_SYS_RAWIO) &&
5114 	    init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP)
5115 		return ERR_PTR(-EPERM);
5116 
5117 	dev = to_mdev(pd->device);
5118 	switch (init_attr->wq_type) {
5119 	case IB_WQT_RQ:
5120 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
5121 		if (!rwq)
5122 			return ERR_PTR(-ENOMEM);
5123 		err = prepare_user_rq(pd, init_attr, udata, rwq);
5124 		if (err)
5125 			goto err;
5126 		err = create_rq(rwq, pd, init_attr);
5127 		if (err)
5128 			goto err_user_rq;
5129 		break;
5130 	default:
5131 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
5132 			    init_attr->wq_type);
5133 		return ERR_PTR(-EINVAL);
5134 	}
5135 
5136 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
5137 	rwq->ibwq.state = IB_WQS_RESET;
5138 	if (udata->outlen) {
5139 		resp.response_length = offsetofend(
5140 			struct mlx5_ib_create_wq_resp, response_length);
5141 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
5142 		if (err)
5143 			goto err_copy;
5144 	}
5145 
5146 	rwq->core_qp.event = mlx5_ib_wq_event;
5147 	rwq->ibwq.event_handler = init_attr->event_handler;
5148 	return &rwq->ibwq;
5149 
5150 err_copy:
5151 	mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5152 err_user_rq:
5153 	destroy_user_rq(dev, pd, rwq, udata);
5154 err:
5155 	kfree(rwq);
5156 	return ERR_PTR(err);
5157 }
5158 
5159 int mlx5_ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
5160 {
5161 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
5162 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
5163 	int ret;
5164 
5165 	ret = mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5166 	if (ret)
5167 		return ret;
5168 	destroy_user_rq(dev, wq->pd, rwq, udata);
5169 	kfree(rwq);
5170 	return 0;
5171 }
5172 
5173 int mlx5_ib_create_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_table,
5174 				 struct ib_rwq_ind_table_init_attr *init_attr,
5175 				 struct ib_udata *udata)
5176 {
5177 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl =
5178 		to_mrwq_ind_table(ib_rwq_ind_table);
5179 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_table->device);
5180 	int sz = 1 << init_attr->log_ind_tbl_size;
5181 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
5182 	size_t min_resp_len;
5183 	int inlen;
5184 	int err;
5185 	int i;
5186 	u32 *in;
5187 	void *rqtc;
5188 
5189 	if (udata->inlen > 0 &&
5190 	    !ib_is_udata_cleared(udata, 0,
5191 				 udata->inlen))
5192 		return -EOPNOTSUPP;
5193 
5194 	if (init_attr->log_ind_tbl_size >
5195 	    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
5196 		mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
5197 			    init_attr->log_ind_tbl_size,
5198 			    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
5199 		return -EINVAL;
5200 	}
5201 
5202 	min_resp_len =
5203 		offsetofend(struct mlx5_ib_create_rwq_ind_tbl_resp, reserved);
5204 	if (udata->outlen && udata->outlen < min_resp_len)
5205 		return -EINVAL;
5206 
5207 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
5208 	in = kvzalloc(inlen, GFP_KERNEL);
5209 	if (!in)
5210 		return -ENOMEM;
5211 
5212 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
5213 
5214 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
5215 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
5216 
5217 	for (i = 0; i < sz; i++)
5218 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
5219 
5220 	rwq_ind_tbl->uid = to_mpd(init_attr->ind_tbl[0]->pd)->uid;
5221 	MLX5_SET(create_rqt_in, in, uid, rwq_ind_tbl->uid);
5222 
5223 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
5224 	kvfree(in);
5225 	if (err)
5226 		return err;
5227 
5228 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
5229 	if (udata->outlen) {
5230 		resp.response_length =
5231 			offsetofend(struct mlx5_ib_create_rwq_ind_tbl_resp,
5232 				    response_length);
5233 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
5234 		if (err)
5235 			goto err_copy;
5236 	}
5237 
5238 	return 0;
5239 
5240 err_copy:
5241 	mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
5242 	return err;
5243 }
5244 
5245 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
5246 {
5247 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
5248 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
5249 
5250 	return mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
5251 }
5252 
5253 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
5254 		      u32 wq_attr_mask, struct ib_udata *udata)
5255 {
5256 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
5257 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
5258 	struct mlx5_ib_modify_wq ucmd = {};
5259 	size_t required_cmd_sz;
5260 	int curr_wq_state;
5261 	int wq_state;
5262 	int inlen;
5263 	int err;
5264 	void *rqc;
5265 	void *in;
5266 
5267 	required_cmd_sz = offsetofend(struct mlx5_ib_modify_wq, reserved);
5268 	if (udata->inlen < required_cmd_sz)
5269 		return -EINVAL;
5270 
5271 	if (udata->inlen > sizeof(ucmd) &&
5272 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
5273 				 udata->inlen - sizeof(ucmd)))
5274 		return -EOPNOTSUPP;
5275 
5276 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
5277 		return -EFAULT;
5278 
5279 	if (ucmd.comp_mask || ucmd.reserved)
5280 		return -EOPNOTSUPP;
5281 
5282 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
5283 	in = kvzalloc(inlen, GFP_KERNEL);
5284 	if (!in)
5285 		return -ENOMEM;
5286 
5287 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
5288 
5289 	curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
5290 		wq_attr->curr_wq_state : wq->state;
5291 	wq_state = (wq_attr_mask & IB_WQ_STATE) ?
5292 		wq_attr->wq_state : curr_wq_state;
5293 	if (curr_wq_state == IB_WQS_ERR)
5294 		curr_wq_state = MLX5_RQC_STATE_ERR;
5295 	if (wq_state == IB_WQS_ERR)
5296 		wq_state = MLX5_RQC_STATE_ERR;
5297 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
5298 	MLX5_SET(modify_rq_in, in, uid, to_mpd(wq->pd)->uid);
5299 	MLX5_SET(rqc, rqc, state, wq_state);
5300 
5301 	if (wq_attr_mask & IB_WQ_FLAGS) {
5302 		if (wq_attr->flags_mask & IB_WQ_FLAGS_CVLAN_STRIPPING) {
5303 			if (!(MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
5304 			      MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
5305 				mlx5_ib_dbg(dev, "VLAN offloads are not "
5306 					    "supported\n");
5307 				err = -EOPNOTSUPP;
5308 				goto out;
5309 			}
5310 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
5311 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
5312 			MLX5_SET(rqc, rqc, vsd,
5313 				 (wq_attr->flags & IB_WQ_FLAGS_CVLAN_STRIPPING) ? 0 : 1);
5314 		}
5315 
5316 		if (wq_attr->flags_mask & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
5317 			mlx5_ib_dbg(dev, "Modifying scatter end padding is not supported\n");
5318 			err = -EOPNOTSUPP;
5319 			goto out;
5320 		}
5321 	}
5322 
5323 	if (curr_wq_state == IB_WQS_RESET && wq_state == IB_WQS_RDY) {
5324 		u16 set_id;
5325 
5326 		set_id = mlx5_ib_get_counters_id(dev, 0);
5327 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
5328 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
5329 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
5330 			MLX5_SET(rqc, rqc, counter_set_id, set_id);
5331 		} else
5332 			dev_info_once(
5333 				&dev->ib_dev.dev,
5334 				"Receive WQ counters are not supported on current FW\n");
5335 	}
5336 
5337 	err = mlx5_core_modify_rq(dev->mdev, rwq->core_qp.qpn, in);
5338 	if (!err)
5339 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
5340 
5341 out:
5342 	kvfree(in);
5343 	return err;
5344 }
5345 
5346 struct mlx5_ib_drain_cqe {
5347 	struct ib_cqe cqe;
5348 	struct completion done;
5349 };
5350 
5351 static void mlx5_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
5352 {
5353 	struct mlx5_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
5354 						     struct mlx5_ib_drain_cqe,
5355 						     cqe);
5356 
5357 	complete(&cqe->done);
5358 }
5359 
5360 /* This function returns only once the drained WR was completed */
5361 static void handle_drain_completion(struct ib_cq *cq,
5362 				    struct mlx5_ib_drain_cqe *sdrain,
5363 				    struct mlx5_ib_dev *dev)
5364 {
5365 	struct mlx5_core_dev *mdev = dev->mdev;
5366 
5367 	if (cq->poll_ctx == IB_POLL_DIRECT) {
5368 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
5369 			ib_process_cq_direct(cq, -1);
5370 		return;
5371 	}
5372 
5373 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5374 		struct mlx5_ib_cq *mcq = to_mcq(cq);
5375 		bool triggered = false;
5376 		unsigned long flags;
5377 
5378 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
5379 		/* Make sure that the CQ handler won't run if wasn't run yet */
5380 		if (!mcq->mcq.reset_notify_added)
5381 			mcq->mcq.reset_notify_added = 1;
5382 		else
5383 			triggered = true;
5384 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
5385 
5386 		if (triggered) {
5387 			/* Wait for any scheduled/running task to be ended */
5388 			switch (cq->poll_ctx) {
5389 			case IB_POLL_SOFTIRQ:
5390 				irq_poll_disable(&cq->iop);
5391 				irq_poll_enable(&cq->iop);
5392 				break;
5393 			case IB_POLL_WORKQUEUE:
5394 				cancel_work_sync(&cq->work);
5395 				break;
5396 			default:
5397 				WARN_ON_ONCE(1);
5398 			}
5399 		}
5400 
5401 		/* Run the CQ handler - this makes sure that the drain WR will
5402 		 * be processed if wasn't processed yet.
5403 		 */
5404 		mcq->mcq.comp(&mcq->mcq, NULL);
5405 	}
5406 
5407 	wait_for_completion(&sdrain->done);
5408 }
5409 
5410 void mlx5_ib_drain_sq(struct ib_qp *qp)
5411 {
5412 	struct ib_cq *cq = qp->send_cq;
5413 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
5414 	struct mlx5_ib_drain_cqe sdrain;
5415 	const struct ib_send_wr *bad_swr;
5416 	struct ib_rdma_wr swr = {
5417 		.wr = {
5418 			.next = NULL,
5419 			{ .wr_cqe	= &sdrain.cqe, },
5420 			.opcode	= IB_WR_RDMA_WRITE,
5421 		},
5422 	};
5423 	int ret;
5424 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5425 	struct mlx5_core_dev *mdev = dev->mdev;
5426 
5427 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
5428 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5429 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
5430 		return;
5431 	}
5432 
5433 	sdrain.cqe.done = mlx5_ib_drain_qp_done;
5434 	init_completion(&sdrain.done);
5435 
5436 	ret = mlx5_ib_post_send_drain(qp, &swr.wr, &bad_swr);
5437 	if (ret) {
5438 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
5439 		return;
5440 	}
5441 
5442 	handle_drain_completion(cq, &sdrain, dev);
5443 }
5444 
5445 void mlx5_ib_drain_rq(struct ib_qp *qp)
5446 {
5447 	struct ib_cq *cq = qp->recv_cq;
5448 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
5449 	struct mlx5_ib_drain_cqe rdrain;
5450 	struct ib_recv_wr rwr = {};
5451 	const struct ib_recv_wr *bad_rwr;
5452 	int ret;
5453 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5454 	struct mlx5_core_dev *mdev = dev->mdev;
5455 
5456 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
5457 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5458 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
5459 		return;
5460 	}
5461 
5462 	rwr.wr_cqe = &rdrain.cqe;
5463 	rdrain.cqe.done = mlx5_ib_drain_qp_done;
5464 	init_completion(&rdrain.done);
5465 
5466 	ret = mlx5_ib_post_recv_drain(qp, &rwr, &bad_rwr);
5467 	if (ret) {
5468 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
5469 		return;
5470 	}
5471 
5472 	handle_drain_completion(cq, &rdrain, dev);
5473 }
5474 
5475 /*
5476  * Bind a qp to a counter. If @counter is NULL then bind the qp to
5477  * the default counter
5478  */
5479 int mlx5_ib_qp_set_counter(struct ib_qp *qp, struct rdma_counter *counter)
5480 {
5481 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5482 	struct mlx5_ib_qp *mqp = to_mqp(qp);
5483 	int err = 0;
5484 
5485 	mutex_lock(&mqp->mutex);
5486 	if (mqp->state == IB_QPS_RESET) {
5487 		qp->counter = counter;
5488 		goto out;
5489 	}
5490 
5491 	if (!MLX5_CAP_GEN(dev->mdev, rts2rts_qp_counters_set_id)) {
5492 		err = -EOPNOTSUPP;
5493 		goto out;
5494 	}
5495 
5496 	if (mqp->state == IB_QPS_RTS) {
5497 		err = __mlx5_ib_qp_set_counter(qp, counter);
5498 		if (!err)
5499 			qp->counter = counter;
5500 
5501 		goto out;
5502 	}
5503 
5504 	mqp->counter_pending = 1;
5505 	qp->counter = counter;
5506 
5507 out:
5508 	mutex_unlock(&mqp->mutex);
5509 	return err;
5510 }
5511