xref: /openbmc/linux/drivers/infiniband/hw/mlx5/qp.c (revision 7a2f6f61)
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/etherdevice.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 	u32 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_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
618 	       bfregi->num_static_sys_pages * MLX5_NON_FP_BFREGS_PER_UAR;
619 }
620 
621 static int num_med_bfreg(struct mlx5_ib_dev *dev,
622 			 struct mlx5_bfreg_info *bfregi)
623 {
624 	int n;
625 
626 	n = max_bfregs(dev, bfregi) - bfregi->num_low_latency_bfregs -
627 	    NUM_NON_BLUE_FLAME_BFREGS;
628 
629 	return n >= 0 ? n : 0;
630 }
631 
632 static int first_med_bfreg(struct mlx5_ib_dev *dev,
633 			   struct mlx5_bfreg_info *bfregi)
634 {
635 	return num_med_bfreg(dev, bfregi) ? 1 : -ENOMEM;
636 }
637 
638 static int first_hi_bfreg(struct mlx5_ib_dev *dev,
639 			  struct mlx5_bfreg_info *bfregi)
640 {
641 	int med;
642 
643 	med = num_med_bfreg(dev, bfregi);
644 	return ++med;
645 }
646 
647 static int alloc_high_class_bfreg(struct mlx5_ib_dev *dev,
648 				  struct mlx5_bfreg_info *bfregi)
649 {
650 	int i;
651 
652 	for (i = first_hi_bfreg(dev, bfregi); i < max_bfregs(dev, bfregi); i++) {
653 		if (!bfregi->count[i]) {
654 			bfregi->count[i]++;
655 			return i;
656 		}
657 	}
658 
659 	return -ENOMEM;
660 }
661 
662 static int alloc_med_class_bfreg(struct mlx5_ib_dev *dev,
663 				 struct mlx5_bfreg_info *bfregi)
664 {
665 	int minidx = first_med_bfreg(dev, bfregi);
666 	int i;
667 
668 	if (minidx < 0)
669 		return minidx;
670 
671 	for (i = minidx; i < first_hi_bfreg(dev, bfregi); i++) {
672 		if (bfregi->count[i] < bfregi->count[minidx])
673 			minidx = i;
674 		if (!bfregi->count[minidx])
675 			break;
676 	}
677 
678 	bfregi->count[minidx]++;
679 	return minidx;
680 }
681 
682 static int alloc_bfreg(struct mlx5_ib_dev *dev,
683 		       struct mlx5_bfreg_info *bfregi)
684 {
685 	int bfregn = -ENOMEM;
686 
687 	if (bfregi->lib_uar_dyn)
688 		return -EINVAL;
689 
690 	mutex_lock(&bfregi->lock);
691 	if (bfregi->ver >= 2) {
692 		bfregn = alloc_high_class_bfreg(dev, bfregi);
693 		if (bfregn < 0)
694 			bfregn = alloc_med_class_bfreg(dev, bfregi);
695 	}
696 
697 	if (bfregn < 0) {
698 		BUILD_BUG_ON(NUM_NON_BLUE_FLAME_BFREGS != 1);
699 		bfregn = 0;
700 		bfregi->count[bfregn]++;
701 	}
702 	mutex_unlock(&bfregi->lock);
703 
704 	return bfregn;
705 }
706 
707 void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
708 {
709 	mutex_lock(&bfregi->lock);
710 	bfregi->count[bfregn]--;
711 	mutex_unlock(&bfregi->lock);
712 }
713 
714 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
715 {
716 	switch (state) {
717 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
718 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
719 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
720 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
721 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
722 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
723 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
724 	default:		return -1;
725 	}
726 }
727 
728 static int to_mlx5_st(enum ib_qp_type type)
729 {
730 	switch (type) {
731 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
732 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
733 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
734 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
735 	case IB_QPT_XRC_INI:
736 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
737 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
738 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
739 	case MLX5_IB_QPT_DCI:		return MLX5_QP_ST_DCI;
740 	case IB_QPT_RAW_PACKET:		return MLX5_QP_ST_RAW_ETHERTYPE;
741 	default:		return -EINVAL;
742 	}
743 }
744 
745 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
746 			     struct mlx5_ib_cq *recv_cq);
747 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
748 			       struct mlx5_ib_cq *recv_cq);
749 
750 int bfregn_to_uar_index(struct mlx5_ib_dev *dev,
751 			struct mlx5_bfreg_info *bfregi, u32 bfregn,
752 			bool dyn_bfreg)
753 {
754 	unsigned int bfregs_per_sys_page;
755 	u32 index_of_sys_page;
756 	u32 offset;
757 
758 	if (bfregi->lib_uar_dyn)
759 		return -EINVAL;
760 
761 	bfregs_per_sys_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
762 				MLX5_NON_FP_BFREGS_PER_UAR;
763 	index_of_sys_page = bfregn / bfregs_per_sys_page;
764 
765 	if (dyn_bfreg) {
766 		index_of_sys_page += bfregi->num_static_sys_pages;
767 
768 		if (index_of_sys_page >= bfregi->num_sys_pages)
769 			return -EINVAL;
770 
771 		if (bfregn > bfregi->num_dyn_bfregs ||
772 		    bfregi->sys_pages[index_of_sys_page] == MLX5_IB_INVALID_UAR_INDEX) {
773 			mlx5_ib_dbg(dev, "Invalid dynamic uar index\n");
774 			return -EINVAL;
775 		}
776 	}
777 
778 	offset = bfregn % bfregs_per_sys_page / MLX5_NON_FP_BFREGS_PER_UAR;
779 	return bfregi->sys_pages[index_of_sys_page] + offset;
780 }
781 
782 static void destroy_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
783 			    struct mlx5_ib_rwq *rwq, struct ib_udata *udata)
784 {
785 	struct mlx5_ib_ucontext *context =
786 		rdma_udata_to_drv_context(
787 			udata,
788 			struct mlx5_ib_ucontext,
789 			ibucontext);
790 
791 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_DELAY_DROP)
792 		atomic_dec(&dev->delay_drop.rqs_cnt);
793 
794 	mlx5_ib_db_unmap_user(context, &rwq->db);
795 	ib_umem_release(rwq->umem);
796 }
797 
798 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
799 			  struct ib_udata *udata, struct mlx5_ib_rwq *rwq,
800 			  struct mlx5_ib_create_wq *ucmd)
801 {
802 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
803 		udata, struct mlx5_ib_ucontext, ibucontext);
804 	unsigned long page_size = 0;
805 	u32 offset = 0;
806 	int err;
807 
808 	if (!ucmd->buf_addr)
809 		return -EINVAL;
810 
811 	rwq->umem = ib_umem_get(&dev->ib_dev, ucmd->buf_addr, rwq->buf_size, 0);
812 	if (IS_ERR(rwq->umem)) {
813 		mlx5_ib_dbg(dev, "umem_get failed\n");
814 		err = PTR_ERR(rwq->umem);
815 		return err;
816 	}
817 
818 	page_size = mlx5_umem_find_best_quantized_pgoff(
819 		rwq->umem, wq, log_wq_pg_sz, MLX5_ADAPTER_PAGE_SHIFT,
820 		page_offset, 64, &rwq->rq_page_offset);
821 	if (!page_size) {
822 		mlx5_ib_warn(dev, "bad offset\n");
823 		err = -EINVAL;
824 		goto err_umem;
825 	}
826 
827 	rwq->rq_num_pas = ib_umem_num_dma_blocks(rwq->umem, page_size);
828 	rwq->page_shift = order_base_2(page_size);
829 	rwq->log_page_size =  rwq->page_shift - MLX5_ADAPTER_PAGE_SHIFT;
830 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
831 
832 	mlx5_ib_dbg(
833 		dev,
834 		"addr 0x%llx, size %zd, npages %zu, page_size %ld, ncont %d, offset %d\n",
835 		(unsigned long long)ucmd->buf_addr, rwq->buf_size,
836 		ib_umem_num_pages(rwq->umem), page_size, rwq->rq_num_pas,
837 		offset);
838 
839 	err = mlx5_ib_db_map_user(ucontext, ucmd->db_addr, &rwq->db);
840 	if (err) {
841 		mlx5_ib_dbg(dev, "map failed\n");
842 		goto err_umem;
843 	}
844 
845 	return 0;
846 
847 err_umem:
848 	ib_umem_release(rwq->umem);
849 	return err;
850 }
851 
852 static int adjust_bfregn(struct mlx5_ib_dev *dev,
853 			 struct mlx5_bfreg_info *bfregi, int bfregn)
854 {
855 	return bfregn / MLX5_NON_FP_BFREGS_PER_UAR * MLX5_BFREGS_PER_UAR +
856 				bfregn % MLX5_NON_FP_BFREGS_PER_UAR;
857 }
858 
859 static int _create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
860 			   struct mlx5_ib_qp *qp, struct ib_udata *udata,
861 			   struct ib_qp_init_attr *attr, u32 **in,
862 			   struct mlx5_ib_create_qp_resp *resp, int *inlen,
863 			   struct mlx5_ib_qp_base *base,
864 			   struct mlx5_ib_create_qp *ucmd)
865 {
866 	struct mlx5_ib_ucontext *context;
867 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
868 	unsigned int page_offset_quantized = 0;
869 	unsigned long page_size = 0;
870 	int uar_index = 0;
871 	int bfregn;
872 	int ncont = 0;
873 	__be64 *pas;
874 	void *qpc;
875 	int err;
876 	u16 uid;
877 	u32 uar_flags;
878 
879 	context = rdma_udata_to_drv_context(udata, struct mlx5_ib_ucontext,
880 					    ibucontext);
881 	uar_flags = qp->flags_en &
882 		    (MLX5_QP_FLAG_UAR_PAGE_INDEX | MLX5_QP_FLAG_BFREG_INDEX);
883 	switch (uar_flags) {
884 	case MLX5_QP_FLAG_UAR_PAGE_INDEX:
885 		uar_index = ucmd->bfreg_index;
886 		bfregn = MLX5_IB_INVALID_BFREG;
887 		break;
888 	case MLX5_QP_FLAG_BFREG_INDEX:
889 		uar_index = bfregn_to_uar_index(dev, &context->bfregi,
890 						ucmd->bfreg_index, true);
891 		if (uar_index < 0)
892 			return uar_index;
893 		bfregn = MLX5_IB_INVALID_BFREG;
894 		break;
895 	case 0:
896 		if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
897 			return -EINVAL;
898 		bfregn = alloc_bfreg(dev, &context->bfregi);
899 		if (bfregn < 0)
900 			return bfregn;
901 		break;
902 	default:
903 		return -EINVAL;
904 	}
905 
906 	mlx5_ib_dbg(dev, "bfregn 0x%x, uar_index 0x%x\n", bfregn, uar_index);
907 	if (bfregn != MLX5_IB_INVALID_BFREG)
908 		uar_index = bfregn_to_uar_index(dev, &context->bfregi, bfregn,
909 						false);
910 
911 	qp->rq.offset = 0;
912 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
913 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
914 
915 	err = set_user_buf_size(dev, qp, ucmd, base, attr);
916 	if (err)
917 		goto err_bfreg;
918 
919 	if (ucmd->buf_addr && ubuffer->buf_size) {
920 		ubuffer->buf_addr = ucmd->buf_addr;
921 		ubuffer->umem = ib_umem_get(&dev->ib_dev, ubuffer->buf_addr,
922 					    ubuffer->buf_size, 0);
923 		if (IS_ERR(ubuffer->umem)) {
924 			err = PTR_ERR(ubuffer->umem);
925 			goto err_bfreg;
926 		}
927 		page_size = mlx5_umem_find_best_quantized_pgoff(
928 			ubuffer->umem, qpc, log_page_size,
929 			MLX5_ADAPTER_PAGE_SHIFT, page_offset, 64,
930 			&page_offset_quantized);
931 		if (!page_size) {
932 			err = -EINVAL;
933 			goto err_umem;
934 		}
935 		ncont = ib_umem_num_dma_blocks(ubuffer->umem, page_size);
936 	} else {
937 		ubuffer->umem = NULL;
938 	}
939 
940 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
941 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
942 	*in = kvzalloc(*inlen, GFP_KERNEL);
943 	if (!*in) {
944 		err = -ENOMEM;
945 		goto err_umem;
946 	}
947 
948 	uid = (attr->qp_type != IB_QPT_XRC_INI) ? to_mpd(pd)->uid : 0;
949 	MLX5_SET(create_qp_in, *in, uid, uid);
950 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
951 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
952 	if (ubuffer->umem) {
953 		mlx5_ib_populate_pas(ubuffer->umem, page_size, pas, 0);
954 		MLX5_SET(qpc, qpc, log_page_size,
955 			 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
956 		MLX5_SET(qpc, qpc, page_offset, page_offset_quantized);
957 	}
958 	MLX5_SET(qpc, qpc, uar_page, uar_index);
959 	if (bfregn != MLX5_IB_INVALID_BFREG)
960 		resp->bfreg_index = adjust_bfregn(dev, &context->bfregi, bfregn);
961 	else
962 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
963 	qp->bfregn = bfregn;
964 
965 	err = mlx5_ib_db_map_user(context, ucmd->db_addr, &qp->db);
966 	if (err) {
967 		mlx5_ib_dbg(dev, "map failed\n");
968 		goto err_free;
969 	}
970 
971 	return 0;
972 
973 err_free:
974 	kvfree(*in);
975 
976 err_umem:
977 	ib_umem_release(ubuffer->umem);
978 
979 err_bfreg:
980 	if (bfregn != MLX5_IB_INVALID_BFREG)
981 		mlx5_ib_free_bfreg(dev, &context->bfregi, bfregn);
982 	return err;
983 }
984 
985 static void destroy_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
986 		       struct mlx5_ib_qp_base *base, struct ib_udata *udata)
987 {
988 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
989 		udata, struct mlx5_ib_ucontext, ibucontext);
990 
991 	if (udata) {
992 		/* User QP */
993 		mlx5_ib_db_unmap_user(context, &qp->db);
994 		ib_umem_release(base->ubuffer.umem);
995 
996 		/*
997 		 * Free only the BFREGs which are handled by the kernel.
998 		 * BFREGs of UARs allocated dynamically are handled by user.
999 		 */
1000 		if (qp->bfregn != MLX5_IB_INVALID_BFREG)
1001 			mlx5_ib_free_bfreg(dev, &context->bfregi, qp->bfregn);
1002 		return;
1003 	}
1004 
1005 	/* Kernel QP */
1006 	kvfree(qp->sq.wqe_head);
1007 	kvfree(qp->sq.w_list);
1008 	kvfree(qp->sq.wrid);
1009 	kvfree(qp->sq.wr_data);
1010 	kvfree(qp->rq.wrid);
1011 	if (qp->db.db)
1012 		mlx5_db_free(dev->mdev, &qp->db);
1013 	if (qp->buf.frags)
1014 		mlx5_frag_buf_free(dev->mdev, &qp->buf);
1015 }
1016 
1017 static int _create_kernel_qp(struct mlx5_ib_dev *dev,
1018 			     struct ib_qp_init_attr *init_attr,
1019 			     struct mlx5_ib_qp *qp, u32 **in, int *inlen,
1020 			     struct mlx5_ib_qp_base *base)
1021 {
1022 	int uar_index;
1023 	void *qpc;
1024 	int err;
1025 
1026 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
1027 		qp->bf.bfreg = &dev->fp_bfreg;
1028 	else if (qp->flags & MLX5_IB_QP_CREATE_WC_TEST)
1029 		qp->bf.bfreg = &dev->wc_bfreg;
1030 	else
1031 		qp->bf.bfreg = &dev->bfreg;
1032 
1033 	/* We need to divide by two since each register is comprised of
1034 	 * two buffers of identical size, namely odd and even
1035 	 */
1036 	qp->bf.buf_size = (1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size)) / 2;
1037 	uar_index = qp->bf.bfreg->index;
1038 
1039 	err = calc_sq_size(dev, init_attr, qp);
1040 	if (err < 0) {
1041 		mlx5_ib_dbg(dev, "err %d\n", err);
1042 		return err;
1043 	}
1044 
1045 	qp->rq.offset = 0;
1046 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
1047 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
1048 
1049 	err = mlx5_frag_buf_alloc_node(dev->mdev, base->ubuffer.buf_size,
1050 				       &qp->buf, dev->mdev->priv.numa_node);
1051 	if (err) {
1052 		mlx5_ib_dbg(dev, "err %d\n", err);
1053 		return err;
1054 	}
1055 
1056 	if (qp->rq.wqe_cnt)
1057 		mlx5_init_fbc(qp->buf.frags, qp->rq.wqe_shift,
1058 			      ilog2(qp->rq.wqe_cnt), &qp->rq.fbc);
1059 
1060 	if (qp->sq.wqe_cnt) {
1061 		int sq_strides_offset = (qp->sq.offset  & (PAGE_SIZE - 1)) /
1062 					MLX5_SEND_WQE_BB;
1063 		mlx5_init_fbc_offset(qp->buf.frags +
1064 				     (qp->sq.offset / PAGE_SIZE),
1065 				     ilog2(MLX5_SEND_WQE_BB),
1066 				     ilog2(qp->sq.wqe_cnt),
1067 				     sq_strides_offset, &qp->sq.fbc);
1068 
1069 		qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
1070 	}
1071 
1072 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
1073 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
1074 	*in = kvzalloc(*inlen, GFP_KERNEL);
1075 	if (!*in) {
1076 		err = -ENOMEM;
1077 		goto err_buf;
1078 	}
1079 
1080 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
1081 	MLX5_SET(qpc, qpc, uar_page, uar_index);
1082 	MLX5_SET(qpc, qpc, ts_format, mlx5_get_qp_default_ts(dev->mdev));
1083 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1084 
1085 	/* Set "fast registration enabled" for all kernel QPs */
1086 	MLX5_SET(qpc, qpc, fre, 1);
1087 	MLX5_SET(qpc, qpc, rlky, 1);
1088 
1089 	if (qp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
1090 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
1091 
1092 	mlx5_fill_page_frag_array(&qp->buf,
1093 				  (__be64 *)MLX5_ADDR_OF(create_qp_in,
1094 							 *in, pas));
1095 
1096 	err = mlx5_db_alloc(dev->mdev, &qp->db);
1097 	if (err) {
1098 		mlx5_ib_dbg(dev, "err %d\n", err);
1099 		goto err_free;
1100 	}
1101 
1102 	qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1103 				     sizeof(*qp->sq.wrid), GFP_KERNEL);
1104 	qp->sq.wr_data = kvmalloc_array(qp->sq.wqe_cnt,
1105 					sizeof(*qp->sq.wr_data), GFP_KERNEL);
1106 	qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1107 				     sizeof(*qp->rq.wrid), GFP_KERNEL);
1108 	qp->sq.w_list = kvmalloc_array(qp->sq.wqe_cnt,
1109 				       sizeof(*qp->sq.w_list), GFP_KERNEL);
1110 	qp->sq.wqe_head = kvmalloc_array(qp->sq.wqe_cnt,
1111 					 sizeof(*qp->sq.wqe_head), GFP_KERNEL);
1112 
1113 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
1114 	    !qp->sq.w_list || !qp->sq.wqe_head) {
1115 		err = -ENOMEM;
1116 		goto err_wrid;
1117 	}
1118 
1119 	return 0;
1120 
1121 err_wrid:
1122 	kvfree(qp->sq.wqe_head);
1123 	kvfree(qp->sq.w_list);
1124 	kvfree(qp->sq.wrid);
1125 	kvfree(qp->sq.wr_data);
1126 	kvfree(qp->rq.wrid);
1127 	mlx5_db_free(dev->mdev, &qp->db);
1128 
1129 err_free:
1130 	kvfree(*in);
1131 
1132 err_buf:
1133 	mlx5_frag_buf_free(dev->mdev, &qp->buf);
1134 	return err;
1135 }
1136 
1137 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
1138 {
1139 	if (attr->srq || (qp->type == IB_QPT_XRC_TGT) ||
1140 	    (qp->type == MLX5_IB_QPT_DCI) || (qp->type == IB_QPT_XRC_INI))
1141 		return MLX5_SRQ_RQ;
1142 	else if (!qp->has_rq)
1143 		return MLX5_ZERO_LEN_RQ;
1144 
1145 	return MLX5_NON_ZERO_RQ;
1146 }
1147 
1148 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1149 				    struct mlx5_ib_qp *qp,
1150 				    struct mlx5_ib_sq *sq, u32 tdn,
1151 				    struct ib_pd *pd)
1152 {
1153 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
1154 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1155 
1156 	MLX5_SET(create_tis_in, in, uid, to_mpd(pd)->uid);
1157 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1158 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
1159 		MLX5_SET(tisc, tisc, underlay_qpn, qp->underlay_qpn);
1160 
1161 	return mlx5_core_create_tis(dev->mdev, in, &sq->tisn);
1162 }
1163 
1164 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1165 				      struct mlx5_ib_sq *sq, struct ib_pd *pd)
1166 {
1167 	mlx5_cmd_destroy_tis(dev->mdev, sq->tisn, to_mpd(pd)->uid);
1168 }
1169 
1170 static void destroy_flow_rule_vport_sq(struct mlx5_ib_sq *sq)
1171 {
1172 	if (sq->flow_rule)
1173 		mlx5_del_flow_rules(sq->flow_rule);
1174 	sq->flow_rule = NULL;
1175 }
1176 
1177 static bool fr_supported(int ts_cap)
1178 {
1179 	return ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_FREE_RUNNING ||
1180 	       ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_FREE_RUNNING_AND_REAL_TIME;
1181 }
1182 
1183 static int get_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1184 			 bool fr_sup, bool rt_sup)
1185 {
1186 	if (cq->private_flags & MLX5_IB_CQ_PR_FLAGS_REAL_TIME_TS) {
1187 		if (!rt_sup) {
1188 			mlx5_ib_dbg(dev,
1189 				    "Real time TS format is not supported\n");
1190 			return -EOPNOTSUPP;
1191 		}
1192 		return MLX5_TIMESTAMP_FORMAT_REAL_TIME;
1193 	}
1194 	if (cq->create_flags & IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION) {
1195 		if (!fr_sup) {
1196 			mlx5_ib_dbg(dev,
1197 				    "Free running TS format is not supported\n");
1198 			return -EOPNOTSUPP;
1199 		}
1200 		return MLX5_TIMESTAMP_FORMAT_FREE_RUNNING;
1201 	}
1202 	return fr_sup ? MLX5_TIMESTAMP_FORMAT_FREE_RUNNING :
1203 			MLX5_TIMESTAMP_FORMAT_DEFAULT;
1204 }
1205 
1206 static int get_rq_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *recv_cq)
1207 {
1208 	u8 ts_cap = MLX5_CAP_GEN(dev->mdev, rq_ts_format);
1209 
1210 	return get_ts_format(dev, recv_cq, fr_supported(ts_cap),
1211 			     rt_supported(ts_cap));
1212 }
1213 
1214 static int get_sq_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq)
1215 {
1216 	u8 ts_cap = MLX5_CAP_GEN(dev->mdev, sq_ts_format);
1217 
1218 	return get_ts_format(dev, send_cq, fr_supported(ts_cap),
1219 			     rt_supported(ts_cap));
1220 }
1221 
1222 static int get_qp_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq,
1223 			    struct mlx5_ib_cq *recv_cq)
1224 {
1225 	u8 ts_cap = MLX5_CAP_ROCE(dev->mdev, qp_ts_format);
1226 	bool fr_sup = fr_supported(ts_cap);
1227 	bool rt_sup = rt_supported(ts_cap);
1228 	u8 default_ts = fr_sup ? MLX5_TIMESTAMP_FORMAT_FREE_RUNNING :
1229 				 MLX5_TIMESTAMP_FORMAT_DEFAULT;
1230 	int send_ts_format =
1231 		send_cq ? get_ts_format(dev, send_cq, fr_sup, rt_sup) :
1232 			  default_ts;
1233 	int recv_ts_format =
1234 		recv_cq ? get_ts_format(dev, recv_cq, fr_sup, rt_sup) :
1235 			  default_ts;
1236 
1237 	if (send_ts_format < 0 || recv_ts_format < 0)
1238 		return -EOPNOTSUPP;
1239 
1240 	if (send_ts_format != MLX5_TIMESTAMP_FORMAT_DEFAULT &&
1241 	    recv_ts_format != MLX5_TIMESTAMP_FORMAT_DEFAULT &&
1242 	    send_ts_format != recv_ts_format) {
1243 		mlx5_ib_dbg(
1244 			dev,
1245 			"The send ts_format does not match the receive ts_format\n");
1246 		return -EOPNOTSUPP;
1247 	}
1248 
1249 	return send_ts_format == default_ts ? recv_ts_format : send_ts_format;
1250 }
1251 
1252 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1253 				   struct ib_udata *udata,
1254 				   struct mlx5_ib_sq *sq, void *qpin,
1255 				   struct ib_pd *pd, struct mlx5_ib_cq *cq)
1256 {
1257 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1258 	__be64 *pas;
1259 	void *in;
1260 	void *sqc;
1261 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1262 	void *wq;
1263 	int inlen;
1264 	int err;
1265 	unsigned int page_offset_quantized;
1266 	unsigned long page_size;
1267 	int ts_format;
1268 
1269 	ts_format = get_sq_ts_format(dev, cq);
1270 	if (ts_format < 0)
1271 		return ts_format;
1272 
1273 	sq->ubuffer.umem = ib_umem_get(&dev->ib_dev, ubuffer->buf_addr,
1274 				       ubuffer->buf_size, 0);
1275 	if (IS_ERR(sq->ubuffer.umem))
1276 		return PTR_ERR(sq->ubuffer.umem);
1277 	page_size = mlx5_umem_find_best_quantized_pgoff(
1278 		ubuffer->umem, wq, log_wq_pg_sz, MLX5_ADAPTER_PAGE_SHIFT,
1279 		page_offset, 64, &page_offset_quantized);
1280 	if (!page_size) {
1281 		err = -EINVAL;
1282 		goto err_umem;
1283 	}
1284 
1285 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
1286 		sizeof(u64) *
1287 			ib_umem_num_dma_blocks(sq->ubuffer.umem, page_size);
1288 	in = kvzalloc(inlen, GFP_KERNEL);
1289 	if (!in) {
1290 		err = -ENOMEM;
1291 		goto err_umem;
1292 	}
1293 
1294 	MLX5_SET(create_sq_in, in, uid, to_mpd(pd)->uid);
1295 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1296 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1297 	if (MLX5_CAP_ETH(dev->mdev, multi_pkt_send_wqe))
1298 		MLX5_SET(sqc, sqc, allow_multi_pkt_send_wqe, 1);
1299 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1300 	MLX5_SET(sqc, sqc, ts_format, ts_format);
1301 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1302 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1303 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1304 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1305 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
1306 	    MLX5_CAP_ETH(dev->mdev, swp))
1307 		MLX5_SET(sqc, sqc, allow_swp, 1);
1308 
1309 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1310 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1311 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1312 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1313 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1314 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1315 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1316 	MLX5_SET(wq, wq, log_wq_pg_sz,
1317 		 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
1318 	MLX5_SET(wq, wq, page_offset, page_offset_quantized);
1319 
1320 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1321 	mlx5_ib_populate_pas(sq->ubuffer.umem, page_size, pas, 0);
1322 
1323 	err = mlx5_core_create_sq_tracked(dev, in, inlen, &sq->base.mqp);
1324 
1325 	kvfree(in);
1326 
1327 	if (err)
1328 		goto err_umem;
1329 
1330 	return 0;
1331 
1332 err_umem:
1333 	ib_umem_release(sq->ubuffer.umem);
1334 	sq->ubuffer.umem = NULL;
1335 
1336 	return err;
1337 }
1338 
1339 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1340 				     struct mlx5_ib_sq *sq)
1341 {
1342 	destroy_flow_rule_vport_sq(sq);
1343 	mlx5_core_destroy_sq_tracked(dev, &sq->base.mqp);
1344 	ib_umem_release(sq->ubuffer.umem);
1345 }
1346 
1347 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1348 				   struct mlx5_ib_rq *rq, void *qpin,
1349 				   struct ib_pd *pd, struct mlx5_ib_cq *cq)
1350 {
1351 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1352 	__be64 *pas;
1353 	void *in;
1354 	void *rqc;
1355 	void *wq;
1356 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1357 	struct ib_umem *umem = rq->base.ubuffer.umem;
1358 	unsigned int page_offset_quantized;
1359 	unsigned long page_size = 0;
1360 	int ts_format;
1361 	size_t inlen;
1362 	int err;
1363 
1364 	ts_format = get_rq_ts_format(dev, cq);
1365 	if (ts_format < 0)
1366 		return ts_format;
1367 
1368 	page_size = mlx5_umem_find_best_quantized_pgoff(umem, wq, log_wq_pg_sz,
1369 							MLX5_ADAPTER_PAGE_SHIFT,
1370 							page_offset, 64,
1371 							&page_offset_quantized);
1372 	if (!page_size)
1373 		return -EINVAL;
1374 
1375 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
1376 		sizeof(u64) * ib_umem_num_dma_blocks(umem, page_size);
1377 	in = kvzalloc(inlen, GFP_KERNEL);
1378 	if (!in)
1379 		return -ENOMEM;
1380 
1381 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
1382 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1383 	if (!(rq->flags & MLX5_IB_RQ_CVLAN_STRIPPING))
1384 		MLX5_SET(rqc, rqc, vsd, 1);
1385 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1386 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1387 	MLX5_SET(rqc, rqc, ts_format, ts_format);
1388 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1389 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1390 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1391 
1392 	if (mqp->flags & IB_QP_CREATE_SCATTER_FCS)
1393 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1394 
1395 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1396 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1397 	if (rq->flags & MLX5_IB_RQ_PCI_WRITE_END_PADDING)
1398 		MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1399 	MLX5_SET(wq, wq, page_offset, page_offset_quantized);
1400 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1401 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1402 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1403 	MLX5_SET(wq, wq, log_wq_pg_sz,
1404 		 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
1405 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1406 
1407 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1408 	mlx5_ib_populate_pas(umem, page_size, pas, 0);
1409 
1410 	err = mlx5_core_create_rq_tracked(dev, in, inlen, &rq->base.mqp);
1411 
1412 	kvfree(in);
1413 
1414 	return err;
1415 }
1416 
1417 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1418 				     struct mlx5_ib_rq *rq)
1419 {
1420 	mlx5_core_destroy_rq_tracked(dev, &rq->base.mqp);
1421 }
1422 
1423 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1424 				      struct mlx5_ib_rq *rq,
1425 				      u32 qp_flags_en,
1426 				      struct ib_pd *pd)
1427 {
1428 	if (qp_flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1429 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1430 		mlx5_ib_disable_lb(dev, false, true);
1431 	mlx5_cmd_destroy_tir(dev->mdev, rq->tirn, to_mpd(pd)->uid);
1432 }
1433 
1434 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1435 				    struct mlx5_ib_rq *rq, u32 tdn,
1436 				    u32 *qp_flags_en, struct ib_pd *pd,
1437 				    u32 *out)
1438 {
1439 	u8 lb_flag = 0;
1440 	u32 *in;
1441 	void *tirc;
1442 	int inlen;
1443 	int err;
1444 
1445 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1446 	in = kvzalloc(inlen, GFP_KERNEL);
1447 	if (!in)
1448 		return -ENOMEM;
1449 
1450 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1451 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1452 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1453 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1454 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1455 	if (*qp_flags_en & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1456 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1457 
1458 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1459 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1460 
1461 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1462 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1463 
1464 	if (dev->is_rep) {
1465 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1466 		*qp_flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1467 	}
1468 
1469 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1470 	MLX5_SET(create_tir_in, in, opcode, MLX5_CMD_OP_CREATE_TIR);
1471 	err = mlx5_cmd_exec_inout(dev->mdev, create_tir, in, out);
1472 	rq->tirn = MLX5_GET(create_tir_out, out, tirn);
1473 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1474 		err = mlx5_ib_enable_lb(dev, false, true);
1475 
1476 		if (err)
1477 			destroy_raw_packet_qp_tir(dev, rq, 0, pd);
1478 	}
1479 	kvfree(in);
1480 
1481 	return err;
1482 }
1483 
1484 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1485 				u32 *in, size_t inlen, struct ib_pd *pd,
1486 				struct ib_udata *udata,
1487 				struct mlx5_ib_create_qp_resp *resp,
1488 				struct ib_qp_init_attr *init_attr)
1489 {
1490 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1491 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1492 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1493 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1494 		udata, struct mlx5_ib_ucontext, ibucontext);
1495 	int err;
1496 	u32 tdn = mucontext->tdn;
1497 	u16 uid = to_mpd(pd)->uid;
1498 	u32 out[MLX5_ST_SZ_DW(create_tir_out)] = {};
1499 
1500 	if (!qp->sq.wqe_cnt && !qp->rq.wqe_cnt)
1501 		return -EINVAL;
1502 	if (qp->sq.wqe_cnt) {
1503 		err = create_raw_packet_qp_tis(dev, qp, sq, tdn, pd);
1504 		if (err)
1505 			return err;
1506 
1507 		err = create_raw_packet_qp_sq(dev, udata, sq, in, pd,
1508 					      to_mcq(init_attr->send_cq));
1509 		if (err)
1510 			goto err_destroy_tis;
1511 
1512 		if (uid) {
1513 			resp->tisn = sq->tisn;
1514 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TISN;
1515 			resp->sqn = sq->base.mqp.qpn;
1516 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_SQN;
1517 		}
1518 
1519 		sq->base.container_mibqp = qp;
1520 		sq->base.mqp.event = mlx5_ib_qp_event;
1521 	}
1522 
1523 	if (qp->rq.wqe_cnt) {
1524 		rq->base.container_mibqp = qp;
1525 
1526 		if (qp->flags & IB_QP_CREATE_CVLAN_STRIPPING)
1527 			rq->flags |= MLX5_IB_RQ_CVLAN_STRIPPING;
1528 		if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING)
1529 			rq->flags |= MLX5_IB_RQ_PCI_WRITE_END_PADDING;
1530 		err = create_raw_packet_qp_rq(dev, rq, in, pd,
1531 					      to_mcq(init_attr->recv_cq));
1532 		if (err)
1533 			goto err_destroy_sq;
1534 
1535 		err = create_raw_packet_qp_tir(dev, rq, tdn, &qp->flags_en, pd,
1536 					       out);
1537 		if (err)
1538 			goto err_destroy_rq;
1539 
1540 		if (uid) {
1541 			resp->rqn = rq->base.mqp.qpn;
1542 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_RQN;
1543 			resp->tirn = rq->tirn;
1544 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1545 			if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
1546 			    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2)) {
1547 				resp->tir_icm_addr = MLX5_GET(
1548 					create_tir_out, out, icm_address_31_0);
1549 				resp->tir_icm_addr |=
1550 					(u64)MLX5_GET(create_tir_out, out,
1551 						      icm_address_39_32)
1552 					<< 32;
1553 				resp->tir_icm_addr |=
1554 					(u64)MLX5_GET(create_tir_out, out,
1555 						      icm_address_63_40)
1556 					<< 40;
1557 				resp->comp_mask |=
1558 					MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1559 			}
1560 		}
1561 	}
1562 
1563 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1564 						     rq->base.mqp.qpn;
1565 	return 0;
1566 
1567 err_destroy_rq:
1568 	destroy_raw_packet_qp_rq(dev, rq);
1569 err_destroy_sq:
1570 	if (!qp->sq.wqe_cnt)
1571 		return err;
1572 	destroy_raw_packet_qp_sq(dev, sq);
1573 err_destroy_tis:
1574 	destroy_raw_packet_qp_tis(dev, sq, pd);
1575 
1576 	return err;
1577 }
1578 
1579 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1580 				  struct mlx5_ib_qp *qp)
1581 {
1582 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1583 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1584 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1585 
1586 	if (qp->rq.wqe_cnt) {
1587 		destroy_raw_packet_qp_tir(dev, rq, qp->flags_en, qp->ibqp.pd);
1588 		destroy_raw_packet_qp_rq(dev, rq);
1589 	}
1590 
1591 	if (qp->sq.wqe_cnt) {
1592 		destroy_raw_packet_qp_sq(dev, sq);
1593 		destroy_raw_packet_qp_tis(dev, sq, qp->ibqp.pd);
1594 	}
1595 }
1596 
1597 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1598 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1599 {
1600 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1601 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1602 
1603 	sq->sq = &qp->sq;
1604 	rq->rq = &qp->rq;
1605 	sq->doorbell = &qp->db;
1606 	rq->doorbell = &qp->db;
1607 }
1608 
1609 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1610 {
1611 	if (qp->flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1612 			    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1613 		mlx5_ib_disable_lb(dev, false, true);
1614 	mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1615 			     to_mpd(qp->ibqp.pd)->uid);
1616 }
1617 
1618 struct mlx5_create_qp_params {
1619 	struct ib_udata *udata;
1620 	size_t inlen;
1621 	size_t outlen;
1622 	size_t ucmd_size;
1623 	void *ucmd;
1624 	u8 is_rss_raw : 1;
1625 	struct ib_qp_init_attr *attr;
1626 	u32 uidx;
1627 	struct mlx5_ib_create_qp_resp resp;
1628 };
1629 
1630 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1631 				 struct mlx5_ib_qp *qp,
1632 				 struct mlx5_create_qp_params *params)
1633 {
1634 	struct ib_qp_init_attr *init_attr = params->attr;
1635 	struct mlx5_ib_create_qp_rss *ucmd = params->ucmd;
1636 	struct ib_udata *udata = params->udata;
1637 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1638 		udata, struct mlx5_ib_ucontext, ibucontext);
1639 	int inlen;
1640 	int outlen;
1641 	int err;
1642 	u32 *in;
1643 	u32 *out;
1644 	void *tirc;
1645 	void *hfso;
1646 	u32 selected_fields = 0;
1647 	u32 outer_l4;
1648 	u32 tdn = mucontext->tdn;
1649 	u8 lb_flag = 0;
1650 
1651 	if (ucmd->comp_mask) {
1652 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1653 		return -EOPNOTSUPP;
1654 	}
1655 
1656 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_INNER &&
1657 	    !(ucmd->flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)) {
1658 		mlx5_ib_dbg(dev, "Tunnel offloads must be set for inner RSS\n");
1659 		return -EOPNOTSUPP;
1660 	}
1661 
1662 	if (dev->is_rep)
1663 		qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1664 
1665 	if (qp->flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1666 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1667 
1668 	if (qp->flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1669 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1670 
1671 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1672 	outlen = MLX5_ST_SZ_BYTES(create_tir_out);
1673 	in = kvzalloc(inlen + outlen, GFP_KERNEL);
1674 	if (!in)
1675 		return -ENOMEM;
1676 
1677 	out = in + MLX5_ST_SZ_DW(create_tir_in);
1678 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1679 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1680 	MLX5_SET(tirc, tirc, disp_type,
1681 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1682 	MLX5_SET(tirc, tirc, indirect_table,
1683 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1684 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1685 
1686 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1687 
1688 	if (ucmd->flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1689 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1690 
1691 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1692 
1693 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_INNER)
1694 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner);
1695 	else
1696 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1697 
1698 	switch (ucmd->rx_hash_function) {
1699 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1700 	{
1701 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1702 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1703 
1704 		if (len != ucmd->rx_key_len) {
1705 			err = -EINVAL;
1706 			goto err;
1707 		}
1708 
1709 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1710 		memcpy(rss_key, ucmd->rx_hash_key, len);
1711 		break;
1712 	}
1713 	default:
1714 		err = -EOPNOTSUPP;
1715 		goto err;
1716 	}
1717 
1718 	if (!ucmd->rx_hash_fields_mask) {
1719 		/* special case when this TIR serves as steering entry without hashing */
1720 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1721 			goto create_tir;
1722 		err = -EINVAL;
1723 		goto err;
1724 	}
1725 
1726 	if (((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1727 	     (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1728 	     ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1729 	     (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1730 		err = -EINVAL;
1731 		goto err;
1732 	}
1733 
1734 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1735 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1736 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1737 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1738 			 MLX5_L3_PROT_TYPE_IPV4);
1739 	else if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1740 		 (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1741 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1742 			 MLX5_L3_PROT_TYPE_IPV6);
1743 
1744 	outer_l4 = ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1745 		    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1746 			   << 0 |
1747 		   ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1748 		    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1749 			   << 1 |
1750 		   (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI) << 2;
1751 
1752 	/* Check that only one l4 protocol is set */
1753 	if (outer_l4 & (outer_l4 - 1)) {
1754 		err = -EINVAL;
1755 		goto err;
1756 	}
1757 
1758 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1759 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1760 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1761 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1762 			 MLX5_L4_PROT_TYPE_TCP);
1763 	else if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1764 		 (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1765 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1766 			 MLX5_L4_PROT_TYPE_UDP);
1767 
1768 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1769 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1770 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1771 
1772 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1773 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1774 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1775 
1776 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1777 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1778 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1779 
1780 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1781 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1782 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1783 
1784 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI)
1785 		selected_fields |= MLX5_HASH_FIELD_SEL_IPSEC_SPI;
1786 
1787 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1788 
1789 create_tir:
1790 	MLX5_SET(create_tir_in, in, opcode, MLX5_CMD_OP_CREATE_TIR);
1791 	err = mlx5_cmd_exec_inout(dev->mdev, create_tir, in, out);
1792 
1793 	qp->rss_qp.tirn = MLX5_GET(create_tir_out, out, tirn);
1794 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1795 		err = mlx5_ib_enable_lb(dev, false, true);
1796 
1797 		if (err)
1798 			mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1799 					     to_mpd(pd)->uid);
1800 	}
1801 
1802 	if (err)
1803 		goto err;
1804 
1805 	if (mucontext->devx_uid) {
1806 		params->resp.comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1807 		params->resp.tirn = qp->rss_qp.tirn;
1808 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
1809 		    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2)) {
1810 			params->resp.tir_icm_addr =
1811 				MLX5_GET(create_tir_out, out, icm_address_31_0);
1812 			params->resp.tir_icm_addr |=
1813 				(u64)MLX5_GET(create_tir_out, out,
1814 					      icm_address_39_32)
1815 				<< 32;
1816 			params->resp.tir_icm_addr |=
1817 				(u64)MLX5_GET(create_tir_out, out,
1818 					      icm_address_63_40)
1819 				<< 40;
1820 			params->resp.comp_mask |=
1821 				MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1822 		}
1823 	}
1824 
1825 	kvfree(in);
1826 	/* qpn is reserved for that QP */
1827 	qp->trans_qp.base.mqp.qpn = 0;
1828 	qp->is_rss = true;
1829 	return 0;
1830 
1831 err:
1832 	kvfree(in);
1833 	return err;
1834 }
1835 
1836 static void configure_requester_scat_cqe(struct mlx5_ib_dev *dev,
1837 					 struct mlx5_ib_qp *qp,
1838 					 struct ib_qp_init_attr *init_attr,
1839 					 void *qpc)
1840 {
1841 	int scqe_sz;
1842 	bool allow_scat_cqe = false;
1843 
1844 	allow_scat_cqe = qp->flags_en & MLX5_QP_FLAG_ALLOW_SCATTER_CQE;
1845 
1846 	if (!allow_scat_cqe && init_attr->sq_sig_type != IB_SIGNAL_ALL_WR)
1847 		return;
1848 
1849 	scqe_sz = mlx5_ib_get_cqe_size(init_attr->send_cq);
1850 	if (scqe_sz == 128) {
1851 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1852 		return;
1853 	}
1854 
1855 	if (init_attr->qp_type != MLX5_IB_QPT_DCI ||
1856 	    MLX5_CAP_GEN(dev->mdev, dc_req_scat_data_cqe))
1857 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1858 }
1859 
1860 static int atomic_size_to_mode(int size_mask)
1861 {
1862 	/* driver does not support atomic_size > 256B
1863 	 * and does not know how to translate bigger sizes
1864 	 */
1865 	int supported_size_mask = size_mask & 0x1ff;
1866 	int log_max_size;
1867 
1868 	if (!supported_size_mask)
1869 		return -EOPNOTSUPP;
1870 
1871 	log_max_size = __fls(supported_size_mask);
1872 
1873 	if (log_max_size > 3)
1874 		return log_max_size;
1875 
1876 	return MLX5_ATOMIC_MODE_8B;
1877 }
1878 
1879 static int get_atomic_mode(struct mlx5_ib_dev *dev,
1880 			   enum ib_qp_type qp_type)
1881 {
1882 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
1883 	u8 atomic = MLX5_CAP_GEN(dev->mdev, atomic);
1884 	int atomic_mode = -EOPNOTSUPP;
1885 	int atomic_size_mask;
1886 
1887 	if (!atomic)
1888 		return -EOPNOTSUPP;
1889 
1890 	if (qp_type == MLX5_IB_QPT_DCT)
1891 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
1892 	else
1893 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
1894 
1895 	if ((atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_CMP_SWAP) ||
1896 	    (atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_FETCH_ADD))
1897 		atomic_mode = atomic_size_to_mode(atomic_size_mask);
1898 
1899 	if (atomic_mode <= 0 &&
1900 	    (atomic_operations & MLX5_ATOMIC_OPS_CMP_SWAP &&
1901 	     atomic_operations & MLX5_ATOMIC_OPS_FETCH_ADD))
1902 		atomic_mode = MLX5_ATOMIC_MODE_IB_COMP;
1903 
1904 	return atomic_mode;
1905 }
1906 
1907 static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1908 			     struct mlx5_create_qp_params *params)
1909 {
1910 	struct ib_qp_init_attr *attr = params->attr;
1911 	u32 uidx = params->uidx;
1912 	struct mlx5_ib_resources *devr = &dev->devr;
1913 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
1914 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1915 	struct mlx5_core_dev *mdev = dev->mdev;
1916 	struct mlx5_ib_qp_base *base;
1917 	unsigned long flags;
1918 	void *qpc;
1919 	u32 *in;
1920 	int err;
1921 
1922 	if (attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1923 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1924 
1925 	in = kvzalloc(inlen, GFP_KERNEL);
1926 	if (!in)
1927 		return -ENOMEM;
1928 
1929 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1930 
1931 	MLX5_SET(qpc, qpc, st, MLX5_QP_ST_XRC);
1932 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
1933 	MLX5_SET(qpc, qpc, pd, to_mpd(devr->p0)->pdn);
1934 
1935 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1936 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
1937 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
1938 		MLX5_SET(qpc, qpc, cd_master, 1);
1939 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
1940 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
1941 	if (qp->flags & IB_QP_CREATE_MANAGED_RECV)
1942 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
1943 
1944 	MLX5_SET(qpc, qpc, ts_format, mlx5_get_qp_default_ts(dev->mdev));
1945 	MLX5_SET(qpc, qpc, rq_type, MLX5_SRQ_RQ);
1946 	MLX5_SET(qpc, qpc, no_sq, 1);
1947 	MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1948 	MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
1949 	MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
1950 	MLX5_SET(qpc, qpc, xrcd, to_mxrcd(attr->xrcd)->xrcdn);
1951 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
1952 
1953 	/* 0xffffff means we ask to work with cqe version 0 */
1954 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
1955 		MLX5_SET(qpc, qpc, user_index, uidx);
1956 
1957 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING) {
1958 		MLX5_SET(qpc, qpc, end_padding_mode,
1959 			 MLX5_WQ_END_PAD_MODE_ALIGN);
1960 		/* Special case to clean flag */
1961 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
1962 	}
1963 
1964 	base = &qp->trans_qp.base;
1965 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
1966 	kvfree(in);
1967 	if (err)
1968 		return err;
1969 
1970 	base->container_mibqp = qp;
1971 	base->mqp.event = mlx5_ib_qp_event;
1972 	if (MLX5_CAP_GEN(mdev, ece_support))
1973 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
1974 
1975 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1976 	list_add_tail(&qp->qps_list, &dev->qp_list);
1977 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1978 
1979 	qp->trans_qp.xrcdn = to_mxrcd(attr->xrcd)->xrcdn;
1980 	return 0;
1981 }
1982 
1983 static int create_dci(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1984 		      struct mlx5_ib_qp *qp,
1985 		      struct mlx5_create_qp_params *params)
1986 {
1987 	struct ib_qp_init_attr *init_attr = params->attr;
1988 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
1989 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
1990 	struct ib_udata *udata = params->udata;
1991 	u32 uidx = params->uidx;
1992 	struct mlx5_ib_resources *devr = &dev->devr;
1993 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1994 	struct mlx5_core_dev *mdev = dev->mdev;
1995 	struct mlx5_ib_cq *send_cq;
1996 	struct mlx5_ib_cq *recv_cq;
1997 	unsigned long flags;
1998 	struct mlx5_ib_qp_base *base;
1999 	int ts_format;
2000 	int mlx5_st;
2001 	void *qpc;
2002 	u32 *in;
2003 	int err;
2004 
2005 	spin_lock_init(&qp->sq.lock);
2006 	spin_lock_init(&qp->rq.lock);
2007 
2008 	mlx5_st = to_mlx5_st(qp->type);
2009 	if (mlx5_st < 0)
2010 		return -EINVAL;
2011 
2012 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2013 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2014 
2015 	base = &qp->trans_qp.base;
2016 
2017 	qp->has_rq = qp_has_rq(init_attr);
2018 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq, qp, ucmd);
2019 	if (err) {
2020 		mlx5_ib_dbg(dev, "err %d\n", err);
2021 		return err;
2022 	}
2023 
2024 	if (ucmd->rq_wqe_shift != qp->rq.wqe_shift ||
2025 	    ucmd->rq_wqe_count != qp->rq.wqe_cnt)
2026 		return -EINVAL;
2027 
2028 	if (ucmd->sq_wqe_count > (1 << MLX5_CAP_GEN(mdev, log_max_qp_sz)))
2029 		return -EINVAL;
2030 
2031 	ts_format = get_qp_ts_format(dev, to_mcq(init_attr->send_cq),
2032 				     to_mcq(init_attr->recv_cq));
2033 
2034 	if (ts_format < 0)
2035 		return ts_format;
2036 
2037 	err = _create_user_qp(dev, pd, qp, udata, init_attr, &in, &params->resp,
2038 			      &inlen, base, ucmd);
2039 	if (err)
2040 		return err;
2041 
2042 	if (MLX5_CAP_GEN(mdev, ece_support))
2043 		MLX5_SET(create_qp_in, in, ece, ucmd->ece_options);
2044 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2045 
2046 	MLX5_SET(qpc, qpc, st, mlx5_st);
2047 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2048 	MLX5_SET(qpc, qpc, pd, to_mpd(pd)->pdn);
2049 
2050 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE)
2051 		MLX5_SET(qpc, qpc, wq_signature, 1);
2052 
2053 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
2054 		MLX5_SET(qpc, qpc, cd_master, 1);
2055 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
2056 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
2057 	if (qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE)
2058 		configure_requester_scat_cqe(dev, qp, init_attr, qpc);
2059 
2060 	if (qp->rq.wqe_cnt) {
2061 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2062 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2063 	}
2064 
2065 	if (qp->flags_en & MLX5_QP_FLAG_DCI_STREAM) {
2066 		MLX5_SET(qpc, qpc, log_num_dci_stream_channels,
2067 			 ucmd->dci_streams.log_num_concurent);
2068 		MLX5_SET(qpc, qpc, log_num_dci_errored_streams,
2069 			 ucmd->dci_streams.log_num_errored);
2070 	}
2071 
2072 	MLX5_SET(qpc, qpc, ts_format, ts_format);
2073 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
2074 
2075 	MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2076 
2077 	/* Set default resources */
2078 	if (init_attr->srq) {
2079 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2080 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2081 			 to_msrq(init_attr->srq)->msrq.srqn);
2082 	} else {
2083 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2084 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2085 			 to_msrq(devr->s1)->msrq.srqn);
2086 	}
2087 
2088 	if (init_attr->send_cq)
2089 		MLX5_SET(qpc, qpc, cqn_snd,
2090 			 to_mcq(init_attr->send_cq)->mcq.cqn);
2091 
2092 	if (init_attr->recv_cq)
2093 		MLX5_SET(qpc, qpc, cqn_rcv,
2094 			 to_mcq(init_attr->recv_cq)->mcq.cqn);
2095 
2096 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2097 
2098 	/* 0xffffff means we ask to work with cqe version 0 */
2099 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2100 		MLX5_SET(qpc, qpc, user_index, uidx);
2101 
2102 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING) {
2103 		MLX5_SET(qpc, qpc, end_padding_mode,
2104 			 MLX5_WQ_END_PAD_MODE_ALIGN);
2105 		/* Special case to clean flag */
2106 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
2107 	}
2108 
2109 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2110 
2111 	kvfree(in);
2112 	if (err)
2113 		goto err_create;
2114 
2115 	base->container_mibqp = qp;
2116 	base->mqp.event = mlx5_ib_qp_event;
2117 	if (MLX5_CAP_GEN(mdev, ece_support))
2118 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
2119 
2120 	get_cqs(qp->type, init_attr->send_cq, init_attr->recv_cq,
2121 		&send_cq, &recv_cq);
2122 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2123 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2124 	/* Maintain device to QPs access, needed for further handling via reset
2125 	 * flow
2126 	 */
2127 	list_add_tail(&qp->qps_list, &dev->qp_list);
2128 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2129 	 */
2130 	if (send_cq)
2131 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2132 	if (recv_cq)
2133 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2134 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2135 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2136 
2137 	return 0;
2138 
2139 err_create:
2140 	destroy_qp(dev, qp, base, udata);
2141 	return err;
2142 }
2143 
2144 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2145 			  struct mlx5_ib_qp *qp,
2146 			  struct mlx5_create_qp_params *params)
2147 {
2148 	struct ib_qp_init_attr *init_attr = params->attr;
2149 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
2150 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
2151 	struct ib_udata *udata = params->udata;
2152 	u32 uidx = params->uidx;
2153 	struct mlx5_ib_resources *devr = &dev->devr;
2154 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
2155 	struct mlx5_core_dev *mdev = dev->mdev;
2156 	struct mlx5_ib_cq *send_cq;
2157 	struct mlx5_ib_cq *recv_cq;
2158 	unsigned long flags;
2159 	struct mlx5_ib_qp_base *base;
2160 	int ts_format;
2161 	int mlx5_st;
2162 	void *qpc;
2163 	u32 *in;
2164 	int err;
2165 
2166 	spin_lock_init(&qp->sq.lock);
2167 	spin_lock_init(&qp->rq.lock);
2168 
2169 	mlx5_st = to_mlx5_st(qp->type);
2170 	if (mlx5_st < 0)
2171 		return -EINVAL;
2172 
2173 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2174 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2175 
2176 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
2177 		qp->underlay_qpn = init_attr->source_qpn;
2178 
2179 	base = (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2180 		qp->flags & IB_QP_CREATE_SOURCE_QPN) ?
2181 	       &qp->raw_packet_qp.rq.base :
2182 	       &qp->trans_qp.base;
2183 
2184 	qp->has_rq = qp_has_rq(init_attr);
2185 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq, qp, ucmd);
2186 	if (err) {
2187 		mlx5_ib_dbg(dev, "err %d\n", err);
2188 		return err;
2189 	}
2190 
2191 	if (ucmd->rq_wqe_shift != qp->rq.wqe_shift ||
2192 	    ucmd->rq_wqe_count != qp->rq.wqe_cnt)
2193 		return -EINVAL;
2194 
2195 	if (ucmd->sq_wqe_count > (1 << MLX5_CAP_GEN(mdev, log_max_qp_sz)))
2196 		return -EINVAL;
2197 
2198 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2199 		ts_format = get_qp_ts_format(dev, to_mcq(init_attr->send_cq),
2200 					     to_mcq(init_attr->recv_cq));
2201 		if (ts_format < 0)
2202 			return ts_format;
2203 	}
2204 
2205 	err = _create_user_qp(dev, pd, qp, udata, init_attr, &in, &params->resp,
2206 			      &inlen, base, ucmd);
2207 	if (err)
2208 		return err;
2209 
2210 	if (is_sqp(init_attr->qp_type))
2211 		qp->port = init_attr->port_num;
2212 
2213 	if (MLX5_CAP_GEN(mdev, ece_support))
2214 		MLX5_SET(create_qp_in, in, ece, ucmd->ece_options);
2215 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2216 
2217 	MLX5_SET(qpc, qpc, st, mlx5_st);
2218 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2219 	MLX5_SET(qpc, qpc, pd, to_mpd(pd)->pdn);
2220 
2221 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE)
2222 		MLX5_SET(qpc, qpc, wq_signature, 1);
2223 
2224 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
2225 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2226 
2227 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
2228 		MLX5_SET(qpc, qpc, cd_master, 1);
2229 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
2230 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
2231 	if (qp->flags & IB_QP_CREATE_MANAGED_RECV)
2232 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
2233 	if (qp->flags_en & MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE)
2234 		MLX5_SET(qpc, qpc, req_e2e_credit_mode, 1);
2235 	if ((qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) &&
2236 	    (init_attr->qp_type == IB_QPT_RC ||
2237 	     init_attr->qp_type == IB_QPT_UC)) {
2238 		int rcqe_sz = mlx5_ib_get_cqe_size(init_attr->recv_cq);
2239 
2240 		MLX5_SET(qpc, qpc, cs_res,
2241 			 rcqe_sz == 128 ? MLX5_RES_SCAT_DATA64_CQE :
2242 					  MLX5_RES_SCAT_DATA32_CQE);
2243 	}
2244 	if ((qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) &&
2245 	    (qp->type == MLX5_IB_QPT_DCI || qp->type == IB_QPT_RC))
2246 		configure_requester_scat_cqe(dev, qp, init_attr, qpc);
2247 
2248 	if (qp->rq.wqe_cnt) {
2249 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2250 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2251 	}
2252 
2253 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
2254 		MLX5_SET(qpc, qpc, ts_format, ts_format);
2255 
2256 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
2257 
2258 	if (qp->sq.wqe_cnt) {
2259 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2260 	} else {
2261 		MLX5_SET(qpc, qpc, no_sq, 1);
2262 		if (init_attr->srq &&
2263 		    init_attr->srq->srq_type == IB_SRQT_TM)
2264 			MLX5_SET(qpc, qpc, offload_type,
2265 				 MLX5_QPC_OFFLOAD_TYPE_RNDV);
2266 	}
2267 
2268 	/* Set default resources */
2269 	switch (init_attr->qp_type) {
2270 	case IB_QPT_XRC_INI:
2271 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2272 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2273 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2274 		break;
2275 	default:
2276 		if (init_attr->srq) {
2277 			MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2278 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(init_attr->srq)->msrq.srqn);
2279 		} else {
2280 			MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2281 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s1)->msrq.srqn);
2282 		}
2283 	}
2284 
2285 	if (init_attr->send_cq)
2286 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
2287 
2288 	if (init_attr->recv_cq)
2289 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
2290 
2291 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2292 
2293 	/* 0xffffff means we ask to work with cqe version 0 */
2294 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2295 		MLX5_SET(qpc, qpc, user_index, uidx);
2296 
2297 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING &&
2298 	    init_attr->qp_type != IB_QPT_RAW_PACKET) {
2299 		MLX5_SET(qpc, qpc, end_padding_mode,
2300 			 MLX5_WQ_END_PAD_MODE_ALIGN);
2301 		/* Special case to clean flag */
2302 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
2303 	}
2304 
2305 	if (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2306 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
2307 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd->sq_buf_addr;
2308 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
2309 		err = create_raw_packet_qp(dev, qp, in, inlen, pd, udata,
2310 					   &params->resp, init_attr);
2311 	} else
2312 		err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2313 
2314 	kvfree(in);
2315 	if (err)
2316 		goto err_create;
2317 
2318 	base->container_mibqp = qp;
2319 	base->mqp.event = mlx5_ib_qp_event;
2320 	if (MLX5_CAP_GEN(mdev, ece_support))
2321 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
2322 
2323 	get_cqs(qp->type, init_attr->send_cq, init_attr->recv_cq,
2324 		&send_cq, &recv_cq);
2325 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2326 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2327 	/* Maintain device to QPs access, needed for further handling via reset
2328 	 * flow
2329 	 */
2330 	list_add_tail(&qp->qps_list, &dev->qp_list);
2331 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2332 	 */
2333 	if (send_cq)
2334 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2335 	if (recv_cq)
2336 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2337 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2338 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2339 
2340 	return 0;
2341 
2342 err_create:
2343 	destroy_qp(dev, qp, base, udata);
2344 	return err;
2345 }
2346 
2347 static int create_kernel_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2348 			    struct mlx5_ib_qp *qp,
2349 			    struct mlx5_create_qp_params *params)
2350 {
2351 	struct ib_qp_init_attr *attr = params->attr;
2352 	u32 uidx = params->uidx;
2353 	struct mlx5_ib_resources *devr = &dev->devr;
2354 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
2355 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
2356 	struct mlx5_core_dev *mdev = dev->mdev;
2357 	struct mlx5_ib_cq *send_cq;
2358 	struct mlx5_ib_cq *recv_cq;
2359 	unsigned long flags;
2360 	struct mlx5_ib_qp_base *base;
2361 	int mlx5_st;
2362 	void *qpc;
2363 	u32 *in;
2364 	int err;
2365 
2366 	spin_lock_init(&qp->sq.lock);
2367 	spin_lock_init(&qp->rq.lock);
2368 
2369 	mlx5_st = to_mlx5_st(qp->type);
2370 	if (mlx5_st < 0)
2371 		return -EINVAL;
2372 
2373 	if (attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2374 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2375 
2376 	base = &qp->trans_qp.base;
2377 
2378 	qp->has_rq = qp_has_rq(attr);
2379 	err = set_rq_size(dev, &attr->cap, qp->has_rq, qp, NULL);
2380 	if (err) {
2381 		mlx5_ib_dbg(dev, "err %d\n", err);
2382 		return err;
2383 	}
2384 
2385 	err = _create_kernel_qp(dev, attr, qp, &in, &inlen, base);
2386 	if (err)
2387 		return err;
2388 
2389 	if (is_sqp(attr->qp_type))
2390 		qp->port = attr->port_num;
2391 
2392 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2393 
2394 	MLX5_SET(qpc, qpc, st, mlx5_st);
2395 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2396 
2397 	if (attr->qp_type != MLX5_IB_QPT_REG_UMR)
2398 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
2399 	else
2400 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
2401 
2402 
2403 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
2404 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2405 
2406 	if (qp->rq.wqe_cnt) {
2407 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2408 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2409 	}
2410 
2411 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, attr));
2412 
2413 	if (qp->sq.wqe_cnt)
2414 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2415 	else
2416 		MLX5_SET(qpc, qpc, no_sq, 1);
2417 
2418 	if (attr->srq) {
2419 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2420 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2421 			 to_msrq(attr->srq)->msrq.srqn);
2422 	} else {
2423 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2424 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2425 			 to_msrq(devr->s1)->msrq.srqn);
2426 	}
2427 
2428 	if (attr->send_cq)
2429 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(attr->send_cq)->mcq.cqn);
2430 
2431 	if (attr->recv_cq)
2432 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(attr->recv_cq)->mcq.cqn);
2433 
2434 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2435 
2436 	/* 0xffffff means we ask to work with cqe version 0 */
2437 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2438 		MLX5_SET(qpc, qpc, user_index, uidx);
2439 
2440 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
2441 	if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO)
2442 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
2443 
2444 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2445 	kvfree(in);
2446 	if (err)
2447 		goto err_create;
2448 
2449 	base->container_mibqp = qp;
2450 	base->mqp.event = mlx5_ib_qp_event;
2451 
2452 	get_cqs(qp->type, attr->send_cq, attr->recv_cq,
2453 		&send_cq, &recv_cq);
2454 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2455 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2456 	/* Maintain device to QPs access, needed for further handling via reset
2457 	 * flow
2458 	 */
2459 	list_add_tail(&qp->qps_list, &dev->qp_list);
2460 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2461 	 */
2462 	if (send_cq)
2463 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2464 	if (recv_cq)
2465 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2466 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2467 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2468 
2469 	return 0;
2470 
2471 err_create:
2472 	destroy_qp(dev, qp, base, NULL);
2473 	return err;
2474 }
2475 
2476 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2477 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
2478 {
2479 	if (send_cq) {
2480 		if (recv_cq) {
2481 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2482 				spin_lock(&send_cq->lock);
2483 				spin_lock_nested(&recv_cq->lock,
2484 						 SINGLE_DEPTH_NESTING);
2485 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2486 				spin_lock(&send_cq->lock);
2487 				__acquire(&recv_cq->lock);
2488 			} else {
2489 				spin_lock(&recv_cq->lock);
2490 				spin_lock_nested(&send_cq->lock,
2491 						 SINGLE_DEPTH_NESTING);
2492 			}
2493 		} else {
2494 			spin_lock(&send_cq->lock);
2495 			__acquire(&recv_cq->lock);
2496 		}
2497 	} else if (recv_cq) {
2498 		spin_lock(&recv_cq->lock);
2499 		__acquire(&send_cq->lock);
2500 	} else {
2501 		__acquire(&send_cq->lock);
2502 		__acquire(&recv_cq->lock);
2503 	}
2504 }
2505 
2506 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2507 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
2508 {
2509 	if (send_cq) {
2510 		if (recv_cq) {
2511 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2512 				spin_unlock(&recv_cq->lock);
2513 				spin_unlock(&send_cq->lock);
2514 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2515 				__release(&recv_cq->lock);
2516 				spin_unlock(&send_cq->lock);
2517 			} else {
2518 				spin_unlock(&send_cq->lock);
2519 				spin_unlock(&recv_cq->lock);
2520 			}
2521 		} else {
2522 			__release(&recv_cq->lock);
2523 			spin_unlock(&send_cq->lock);
2524 		}
2525 	} else if (recv_cq) {
2526 		__release(&send_cq->lock);
2527 		spin_unlock(&recv_cq->lock);
2528 	} else {
2529 		__release(&recv_cq->lock);
2530 		__release(&send_cq->lock);
2531 	}
2532 }
2533 
2534 static void get_cqs(enum ib_qp_type qp_type,
2535 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
2536 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
2537 {
2538 	switch (qp_type) {
2539 	case IB_QPT_XRC_TGT:
2540 		*send_cq = NULL;
2541 		*recv_cq = NULL;
2542 		break;
2543 	case MLX5_IB_QPT_REG_UMR:
2544 	case IB_QPT_XRC_INI:
2545 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2546 		*recv_cq = NULL;
2547 		break;
2548 
2549 	case IB_QPT_SMI:
2550 	case MLX5_IB_QPT_HW_GSI:
2551 	case IB_QPT_RC:
2552 	case IB_QPT_UC:
2553 	case IB_QPT_UD:
2554 	case IB_QPT_RAW_PACKET:
2555 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2556 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
2557 		break;
2558 	default:
2559 		*send_cq = NULL;
2560 		*recv_cq = NULL;
2561 		break;
2562 	}
2563 }
2564 
2565 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2566 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2567 				u8 lag_tx_affinity);
2568 
2569 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2570 			      struct ib_udata *udata)
2571 {
2572 	struct mlx5_ib_cq *send_cq, *recv_cq;
2573 	struct mlx5_ib_qp_base *base;
2574 	unsigned long flags;
2575 	int err;
2576 
2577 	if (qp->is_rss) {
2578 		destroy_rss_raw_qp_tir(dev, qp);
2579 		return;
2580 	}
2581 
2582 	base = (qp->type == IB_QPT_RAW_PACKET ||
2583 		qp->flags & IB_QP_CREATE_SOURCE_QPN) ?
2584 		       &qp->raw_packet_qp.rq.base :
2585 		       &qp->trans_qp.base;
2586 
2587 	if (qp->state != IB_QPS_RESET) {
2588 		if (qp->type != IB_QPT_RAW_PACKET &&
2589 		    !(qp->flags & IB_QP_CREATE_SOURCE_QPN)) {
2590 			err = mlx5_core_qp_modify(dev, MLX5_CMD_OP_2RST_QP, 0,
2591 						  NULL, &base->mqp, NULL);
2592 		} else {
2593 			struct mlx5_modify_raw_qp_param raw_qp_param = {
2594 				.operation = MLX5_CMD_OP_2RST_QP
2595 			};
2596 
2597 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2598 		}
2599 		if (err)
2600 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
2601 				     base->mqp.qpn);
2602 	}
2603 
2604 	get_cqs(qp->type, qp->ibqp.send_cq, qp->ibqp.recv_cq, &send_cq,
2605 		&recv_cq);
2606 
2607 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2608 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2609 	/* del from lists under both locks above to protect reset flow paths */
2610 	list_del(&qp->qps_list);
2611 	if (send_cq)
2612 		list_del(&qp->cq_send_list);
2613 
2614 	if (recv_cq)
2615 		list_del(&qp->cq_recv_list);
2616 
2617 	if (!udata) {
2618 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2619 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
2620 		if (send_cq != recv_cq)
2621 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
2622 					   NULL);
2623 	}
2624 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2625 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2626 
2627 	if (qp->type == IB_QPT_RAW_PACKET ||
2628 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
2629 		destroy_raw_packet_qp(dev, qp);
2630 	} else {
2631 		err = mlx5_core_destroy_qp(dev, &base->mqp);
2632 		if (err)
2633 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
2634 				     base->mqp.qpn);
2635 	}
2636 
2637 	destroy_qp(dev, qp, base, udata);
2638 }
2639 
2640 static int create_dct(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2641 		      struct mlx5_ib_qp *qp,
2642 		      struct mlx5_create_qp_params *params)
2643 {
2644 	struct ib_qp_init_attr *attr = params->attr;
2645 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
2646 	u32 uidx = params->uidx;
2647 	void *dctc;
2648 
2649 	if (mlx5_lag_is_active(dev->mdev) && !MLX5_CAP_GEN(dev->mdev, lag_dct))
2650 		return -EOPNOTSUPP;
2651 
2652 	qp->dct.in = kzalloc(MLX5_ST_SZ_BYTES(create_dct_in), GFP_KERNEL);
2653 	if (!qp->dct.in)
2654 		return -ENOMEM;
2655 
2656 	MLX5_SET(create_dct_in, qp->dct.in, uid, to_mpd(pd)->uid);
2657 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
2658 	MLX5_SET(dctc, dctc, pd, to_mpd(pd)->pdn);
2659 	MLX5_SET(dctc, dctc, srqn_xrqn, to_msrq(attr->srq)->msrq.srqn);
2660 	MLX5_SET(dctc, dctc, cqn, to_mcq(attr->recv_cq)->mcq.cqn);
2661 	MLX5_SET64(dctc, dctc, dc_access_key, ucmd->access_key);
2662 	MLX5_SET(dctc, dctc, user_index, uidx);
2663 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
2664 		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
2665 
2666 	if (qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) {
2667 		int rcqe_sz = mlx5_ib_get_cqe_size(attr->recv_cq);
2668 
2669 		if (rcqe_sz == 128)
2670 			MLX5_SET(dctc, dctc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
2671 	}
2672 
2673 	qp->state = IB_QPS_RESET;
2674 	return 0;
2675 }
2676 
2677 static int check_qp_type(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
2678 			 enum ib_qp_type *type)
2679 {
2680 	if (attr->qp_type == IB_QPT_DRIVER && !MLX5_CAP_GEN(dev->mdev, dct))
2681 		goto out;
2682 
2683 	switch (attr->qp_type) {
2684 	case IB_QPT_XRC_TGT:
2685 	case IB_QPT_XRC_INI:
2686 		if (!MLX5_CAP_GEN(dev->mdev, xrc))
2687 			goto out;
2688 		fallthrough;
2689 	case IB_QPT_RC:
2690 	case IB_QPT_UC:
2691 	case IB_QPT_SMI:
2692 	case MLX5_IB_QPT_HW_GSI:
2693 	case IB_QPT_DRIVER:
2694 	case IB_QPT_GSI:
2695 	case IB_QPT_RAW_PACKET:
2696 	case IB_QPT_UD:
2697 	case MLX5_IB_QPT_REG_UMR:
2698 		break;
2699 	default:
2700 		goto out;
2701 	}
2702 
2703 	*type = attr->qp_type;
2704 	return 0;
2705 
2706 out:
2707 	mlx5_ib_dbg(dev, "Unsupported QP type %d\n", attr->qp_type);
2708 	return -EOPNOTSUPP;
2709 }
2710 
2711 static int check_valid_flow(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2712 			    struct ib_qp_init_attr *attr,
2713 			    struct ib_udata *udata)
2714 {
2715 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2716 		udata, struct mlx5_ib_ucontext, ibucontext);
2717 
2718 	if (!udata) {
2719 		/* Kernel create_qp callers */
2720 		if (attr->rwq_ind_tbl)
2721 			return -EOPNOTSUPP;
2722 
2723 		switch (attr->qp_type) {
2724 		case IB_QPT_RAW_PACKET:
2725 		case IB_QPT_DRIVER:
2726 			return -EOPNOTSUPP;
2727 		default:
2728 			return 0;
2729 		}
2730 	}
2731 
2732 	/* Userspace create_qp callers */
2733 	if (attr->qp_type == IB_QPT_RAW_PACKET && !ucontext->cqe_version) {
2734 		mlx5_ib_dbg(dev,
2735 			"Raw Packet QP is only supported for CQE version > 0\n");
2736 		return -EINVAL;
2737 	}
2738 
2739 	if (attr->qp_type != IB_QPT_RAW_PACKET && attr->rwq_ind_tbl) {
2740 		mlx5_ib_dbg(dev,
2741 			    "Wrong QP type %d for the RWQ indirect table\n",
2742 			    attr->qp_type);
2743 		return -EINVAL;
2744 	}
2745 
2746 	/*
2747 	 * We don't need to see this warning, it means that kernel code
2748 	 * missing ib_pd. Placed here to catch developer's mistakes.
2749 	 */
2750 	WARN_ONCE(!pd && attr->qp_type != IB_QPT_XRC_TGT,
2751 		  "There is a missing PD pointer assignment\n");
2752 	return 0;
2753 }
2754 
2755 static void process_vendor_flag(struct mlx5_ib_dev *dev, int *flags, int flag,
2756 				bool cond, struct mlx5_ib_qp *qp)
2757 {
2758 	if (!(*flags & flag))
2759 		return;
2760 
2761 	if (cond) {
2762 		qp->flags_en |= flag;
2763 		*flags &= ~flag;
2764 		return;
2765 	}
2766 
2767 	switch (flag) {
2768 	case MLX5_QP_FLAG_SCATTER_CQE:
2769 	case MLX5_QP_FLAG_ALLOW_SCATTER_CQE:
2770 		/*
2771 			 * We don't return error if these flags were provided,
2772 			 * and mlx5 doesn't have right capability.
2773 			 */
2774 		*flags &= ~(MLX5_QP_FLAG_SCATTER_CQE |
2775 			    MLX5_QP_FLAG_ALLOW_SCATTER_CQE);
2776 		return;
2777 	default:
2778 		break;
2779 	}
2780 	mlx5_ib_dbg(dev, "Vendor create QP flag 0x%X is not supported\n", flag);
2781 }
2782 
2783 static int process_vendor_flags(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2784 				void *ucmd, struct ib_qp_init_attr *attr)
2785 {
2786 	struct mlx5_core_dev *mdev = dev->mdev;
2787 	bool cond;
2788 	int flags;
2789 
2790 	if (attr->rwq_ind_tbl)
2791 		flags = ((struct mlx5_ib_create_qp_rss *)ucmd)->flags;
2792 	else
2793 		flags = ((struct mlx5_ib_create_qp *)ucmd)->flags;
2794 
2795 	switch (flags & (MLX5_QP_FLAG_TYPE_DCT | MLX5_QP_FLAG_TYPE_DCI)) {
2796 	case MLX5_QP_FLAG_TYPE_DCI:
2797 		qp->type = MLX5_IB_QPT_DCI;
2798 		break;
2799 	case MLX5_QP_FLAG_TYPE_DCT:
2800 		qp->type = MLX5_IB_QPT_DCT;
2801 		break;
2802 	default:
2803 		if (qp->type != IB_QPT_DRIVER)
2804 			break;
2805 		/*
2806 		 * It is IB_QPT_DRIVER and or no subtype or
2807 		 * wrong subtype were provided.
2808 		 */
2809 		return -EINVAL;
2810 	}
2811 
2812 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TYPE_DCI, true, qp);
2813 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TYPE_DCT, true, qp);
2814 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_DCI_STREAM,
2815 			    MLX5_CAP_GEN(mdev, log_max_dci_stream_channels),
2816 			    qp);
2817 
2818 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_SIGNATURE, true, qp);
2819 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_SCATTER_CQE,
2820 			    MLX5_CAP_GEN(mdev, sctr_data_cqe), qp);
2821 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_ALLOW_SCATTER_CQE,
2822 			    MLX5_CAP_GEN(mdev, sctr_data_cqe), qp);
2823 
2824 	if (qp->type == IB_QPT_RAW_PACKET) {
2825 		cond = MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan) ||
2826 		       MLX5_CAP_ETH(mdev, tunnel_stateless_gre) ||
2827 		       MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx);
2828 		process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TUNNEL_OFFLOADS,
2829 				    cond, qp);
2830 		process_vendor_flag(dev, &flags,
2831 				    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC, true,
2832 				    qp);
2833 		process_vendor_flag(dev, &flags,
2834 				    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC, true,
2835 				    qp);
2836 	}
2837 
2838 	if (qp->type == IB_QPT_RC)
2839 		process_vendor_flag(dev, &flags,
2840 				    MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE,
2841 				    MLX5_CAP_GEN(mdev, qp_packet_based), qp);
2842 
2843 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_BFREG_INDEX, true, qp);
2844 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_UAR_PAGE_INDEX, true, qp);
2845 
2846 	cond = qp->flags_en & ~(MLX5_QP_FLAG_TUNNEL_OFFLOADS |
2847 				MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
2848 				MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC);
2849 	if (attr->rwq_ind_tbl && cond) {
2850 		mlx5_ib_dbg(dev, "RSS RAW QP has unsupported flags 0x%X\n",
2851 			    cond);
2852 		return -EINVAL;
2853 	}
2854 
2855 	if (flags)
2856 		mlx5_ib_dbg(dev, "udata has unsupported flags 0x%X\n", flags);
2857 
2858 	return (flags) ? -EINVAL : 0;
2859 	}
2860 
2861 static void process_create_flag(struct mlx5_ib_dev *dev, int *flags, int flag,
2862 				bool cond, struct mlx5_ib_qp *qp)
2863 {
2864 	if (!(*flags & flag))
2865 		return;
2866 
2867 	if (cond) {
2868 		qp->flags |= flag;
2869 		*flags &= ~flag;
2870 		return;
2871 	}
2872 
2873 	if (flag == MLX5_IB_QP_CREATE_WC_TEST) {
2874 		/*
2875 		 * Special case, if condition didn't meet, it won't be error,
2876 		 * just different in-kernel flow.
2877 		 */
2878 		*flags &= ~MLX5_IB_QP_CREATE_WC_TEST;
2879 		return;
2880 	}
2881 	mlx5_ib_dbg(dev, "Verbs create QP flag 0x%X is not supported\n", flag);
2882 }
2883 
2884 static int process_create_flags(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2885 				struct ib_qp_init_attr *attr)
2886 {
2887 	enum ib_qp_type qp_type = qp->type;
2888 	struct mlx5_core_dev *mdev = dev->mdev;
2889 	int create_flags = attr->create_flags;
2890 	bool cond;
2891 
2892 	if (qp_type == MLX5_IB_QPT_DCT)
2893 		return (create_flags) ? -EINVAL : 0;
2894 
2895 	if (qp_type == IB_QPT_RAW_PACKET && attr->rwq_ind_tbl)
2896 		return (create_flags) ? -EINVAL : 0;
2897 
2898 	process_create_flag(dev, &create_flags, IB_QP_CREATE_NETIF_QP,
2899 			    mlx5_get_flow_namespace(dev->mdev,
2900 						    MLX5_FLOW_NAMESPACE_BYPASS),
2901 			    qp);
2902 	process_create_flag(dev, &create_flags,
2903 			    IB_QP_CREATE_INTEGRITY_EN,
2904 			    MLX5_CAP_GEN(mdev, sho), qp);
2905 	process_create_flag(dev, &create_flags,
2906 			    IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK,
2907 			    MLX5_CAP_GEN(mdev, block_lb_mc), qp);
2908 	process_create_flag(dev, &create_flags, IB_QP_CREATE_CROSS_CHANNEL,
2909 			    MLX5_CAP_GEN(mdev, cd), qp);
2910 	process_create_flag(dev, &create_flags, IB_QP_CREATE_MANAGED_SEND,
2911 			    MLX5_CAP_GEN(mdev, cd), qp);
2912 	process_create_flag(dev, &create_flags, IB_QP_CREATE_MANAGED_RECV,
2913 			    MLX5_CAP_GEN(mdev, cd), qp);
2914 
2915 	if (qp_type == IB_QPT_UD) {
2916 		process_create_flag(dev, &create_flags,
2917 				    IB_QP_CREATE_IPOIB_UD_LSO,
2918 				    MLX5_CAP_GEN(mdev, ipoib_basic_offloads),
2919 				    qp);
2920 		cond = MLX5_CAP_GEN(mdev, port_type) == MLX5_CAP_PORT_TYPE_IB;
2921 		process_create_flag(dev, &create_flags, IB_QP_CREATE_SOURCE_QPN,
2922 				    cond, qp);
2923 	}
2924 
2925 	if (qp_type == IB_QPT_RAW_PACKET) {
2926 		cond = MLX5_CAP_GEN(mdev, eth_net_offloads) &&
2927 		       MLX5_CAP_ETH(mdev, scatter_fcs);
2928 		process_create_flag(dev, &create_flags,
2929 				    IB_QP_CREATE_SCATTER_FCS, cond, qp);
2930 
2931 		cond = MLX5_CAP_GEN(mdev, eth_net_offloads) &&
2932 		       MLX5_CAP_ETH(mdev, vlan_cap);
2933 		process_create_flag(dev, &create_flags,
2934 				    IB_QP_CREATE_CVLAN_STRIPPING, cond, qp);
2935 	}
2936 
2937 	process_create_flag(dev, &create_flags,
2938 			    IB_QP_CREATE_PCI_WRITE_END_PADDING,
2939 			    MLX5_CAP_GEN(mdev, end_pad), qp);
2940 
2941 	process_create_flag(dev, &create_flags, MLX5_IB_QP_CREATE_WC_TEST,
2942 			    qp_type != MLX5_IB_QPT_REG_UMR, qp);
2943 	process_create_flag(dev, &create_flags, MLX5_IB_QP_CREATE_SQPN_QP1,
2944 			    true, qp);
2945 
2946 	if (create_flags) {
2947 		mlx5_ib_dbg(dev, "Create QP has unsupported flags 0x%X\n",
2948 			    create_flags);
2949 		return -EOPNOTSUPP;
2950 	}
2951 	return 0;
2952 }
2953 
2954 static int process_udata_size(struct mlx5_ib_dev *dev,
2955 			      struct mlx5_create_qp_params *params)
2956 {
2957 	size_t ucmd = sizeof(struct mlx5_ib_create_qp);
2958 	struct ib_udata *udata = params->udata;
2959 	size_t outlen = udata->outlen;
2960 	size_t inlen = udata->inlen;
2961 
2962 	params->outlen = min(outlen, sizeof(struct mlx5_ib_create_qp_resp));
2963 	params->ucmd_size = ucmd;
2964 	if (!params->is_rss_raw) {
2965 		/* User has old rdma-core, which doesn't support ECE */
2966 		size_t min_inlen =
2967 			offsetof(struct mlx5_ib_create_qp, ece_options);
2968 
2969 		/*
2970 		 * We will check in check_ucmd_data() that user
2971 		 * cleared everything after inlen.
2972 		 */
2973 		params->inlen = (inlen < min_inlen) ? 0 : min(inlen, ucmd);
2974 		goto out;
2975 	}
2976 
2977 	/* RSS RAW QP */
2978 	if (inlen < offsetofend(struct mlx5_ib_create_qp_rss, flags))
2979 		return -EINVAL;
2980 
2981 	if (outlen < offsetofend(struct mlx5_ib_create_qp_resp, bfreg_index))
2982 		return -EINVAL;
2983 
2984 	ucmd = sizeof(struct mlx5_ib_create_qp_rss);
2985 	params->ucmd_size = ucmd;
2986 	if (inlen > ucmd && !ib_is_udata_cleared(udata, ucmd, inlen - ucmd))
2987 		return -EINVAL;
2988 
2989 	params->inlen = min(ucmd, inlen);
2990 out:
2991 	if (!params->inlen)
2992 		mlx5_ib_dbg(dev, "udata is too small\n");
2993 
2994 	return (params->inlen) ? 0 : -EINVAL;
2995 }
2996 
2997 static int create_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2998 		     struct mlx5_ib_qp *qp,
2999 		     struct mlx5_create_qp_params *params)
3000 {
3001 	int err;
3002 
3003 	if (params->is_rss_raw) {
3004 		err = create_rss_raw_qp_tir(dev, pd, qp, params);
3005 		goto out;
3006 	}
3007 
3008 	switch (qp->type) {
3009 	case MLX5_IB_QPT_DCT:
3010 		err = create_dct(dev, pd, qp, params);
3011 		rdma_restrack_no_track(&qp->ibqp.res);
3012 		break;
3013 	case MLX5_IB_QPT_DCI:
3014 		err = create_dci(dev, pd, qp, params);
3015 		break;
3016 	case IB_QPT_XRC_TGT:
3017 		err = create_xrc_tgt_qp(dev, qp, params);
3018 		break;
3019 	case IB_QPT_GSI:
3020 		err = mlx5_ib_create_gsi(pd, qp, params->attr);
3021 		break;
3022 	case MLX5_IB_QPT_HW_GSI:
3023 	case MLX5_IB_QPT_REG_UMR:
3024 		rdma_restrack_no_track(&qp->ibqp.res);
3025 		fallthrough;
3026 	default:
3027 		if (params->udata)
3028 			err = create_user_qp(dev, pd, qp, params);
3029 		else
3030 			err = create_kernel_qp(dev, pd, qp, params);
3031 	}
3032 
3033 out:
3034 	if (err) {
3035 		mlx5_ib_err(dev, "Create QP type %d failed\n", qp->type);
3036 		return err;
3037 	}
3038 
3039 	if (is_qp0(qp->type))
3040 		qp->ibqp.qp_num = 0;
3041 	else if (is_qp1(qp->type))
3042 		qp->ibqp.qp_num = 1;
3043 	else
3044 		qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
3045 
3046 	mlx5_ib_dbg(dev,
3047 		"QP type %d, ib qpn 0x%X, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x, ece 0x%x\n",
3048 		qp->type, qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
3049 		params->attr->recv_cq ? to_mcq(params->attr->recv_cq)->mcq.cqn :
3050 					-1,
3051 		params->attr->send_cq ? to_mcq(params->attr->send_cq)->mcq.cqn :
3052 					-1,
3053 		params->resp.ece_options);
3054 
3055 	return 0;
3056 }
3057 
3058 static int check_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3059 			 struct ib_qp_init_attr *attr)
3060 {
3061 	int ret = 0;
3062 
3063 	switch (qp->type) {
3064 	case MLX5_IB_QPT_DCT:
3065 		ret = (!attr->srq || !attr->recv_cq) ? -EINVAL : 0;
3066 		break;
3067 	case MLX5_IB_QPT_DCI:
3068 		ret = (attr->cap.max_recv_wr || attr->cap.max_recv_sge) ?
3069 			      -EINVAL :
3070 			      0;
3071 		break;
3072 	case IB_QPT_RAW_PACKET:
3073 		ret = (attr->rwq_ind_tbl && attr->send_cq) ? -EINVAL : 0;
3074 		break;
3075 	default:
3076 		break;
3077 	}
3078 
3079 	if (ret)
3080 		mlx5_ib_dbg(dev, "QP type %d has wrong attributes\n", qp->type);
3081 
3082 	return ret;
3083 }
3084 
3085 static int get_qp_uidx(struct mlx5_ib_qp *qp,
3086 		       struct mlx5_create_qp_params *params)
3087 {
3088 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
3089 	struct ib_udata *udata = params->udata;
3090 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3091 		udata, struct mlx5_ib_ucontext, ibucontext);
3092 
3093 	if (params->is_rss_raw)
3094 		return 0;
3095 
3096 	return get_qp_user_index(ucontext, ucmd, sizeof(*ucmd), &params->uidx);
3097 }
3098 
3099 static int mlx5_ib_destroy_dct(struct mlx5_ib_qp *mqp)
3100 {
3101 	struct mlx5_ib_dev *dev = to_mdev(mqp->ibqp.device);
3102 
3103 	if (mqp->state == IB_QPS_RTR) {
3104 		int err;
3105 
3106 		err = mlx5_core_destroy_dct(dev, &mqp->dct.mdct);
3107 		if (err) {
3108 			mlx5_ib_warn(dev, "failed to destroy DCT %d\n", err);
3109 			return err;
3110 		}
3111 	}
3112 
3113 	kfree(mqp->dct.in);
3114 	return 0;
3115 }
3116 
3117 static int check_ucmd_data(struct mlx5_ib_dev *dev,
3118 			   struct mlx5_create_qp_params *params)
3119 {
3120 	struct ib_udata *udata = params->udata;
3121 	size_t size, last;
3122 	int ret;
3123 
3124 	if (params->is_rss_raw)
3125 		/*
3126 		 * These QPs don't have "reserved" field in their
3127 		 * create_qp input struct, so their data is always valid.
3128 		 */
3129 		last = sizeof(struct mlx5_ib_create_qp_rss);
3130 	else
3131 		last = offsetof(struct mlx5_ib_create_qp, reserved);
3132 
3133 	if (udata->inlen <= last)
3134 		return 0;
3135 
3136 	/*
3137 	 * User provides different create_qp structures based on the
3138 	 * flow and we need to know if he cleared memory after our
3139 	 * struct create_qp ends.
3140 	 */
3141 	size = udata->inlen - last;
3142 	ret = ib_is_udata_cleared(params->udata, last, size);
3143 	if (!ret)
3144 		mlx5_ib_dbg(
3145 			dev,
3146 			"udata is not cleared, inlen = %zu, ucmd = %zu, last = %zu, size = %zu\n",
3147 			udata->inlen, params->ucmd_size, last, size);
3148 	return ret ? 0 : -EINVAL;
3149 }
3150 
3151 int mlx5_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
3152 		      struct ib_udata *udata)
3153 {
3154 	struct mlx5_create_qp_params params = {};
3155 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3156 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3157 	struct ib_pd *pd = ibqp->pd;
3158 	enum ib_qp_type type;
3159 	int err;
3160 
3161 	err = check_qp_type(dev, attr, &type);
3162 	if (err)
3163 		return err;
3164 
3165 	err = check_valid_flow(dev, pd, attr, udata);
3166 	if (err)
3167 		return err;
3168 
3169 	params.udata = udata;
3170 	params.uidx = MLX5_IB_DEFAULT_UIDX;
3171 	params.attr = attr;
3172 	params.is_rss_raw = !!attr->rwq_ind_tbl;
3173 
3174 	if (udata) {
3175 		err = process_udata_size(dev, &params);
3176 		if (err)
3177 			return err;
3178 
3179 		err = check_ucmd_data(dev, &params);
3180 		if (err)
3181 			return err;
3182 
3183 		params.ucmd = kzalloc(params.ucmd_size, GFP_KERNEL);
3184 		if (!params.ucmd)
3185 			return -ENOMEM;
3186 
3187 		err = ib_copy_from_udata(params.ucmd, udata, params.inlen);
3188 		if (err)
3189 			goto free_ucmd;
3190 	}
3191 
3192 	mutex_init(&qp->mutex);
3193 	qp->type = type;
3194 	if (udata) {
3195 		err = process_vendor_flags(dev, qp, params.ucmd, attr);
3196 		if (err)
3197 			goto free_ucmd;
3198 
3199 		err = get_qp_uidx(qp, &params);
3200 		if (err)
3201 			goto free_ucmd;
3202 	}
3203 	err = process_create_flags(dev, qp, attr);
3204 	if (err)
3205 		goto free_ucmd;
3206 
3207 	err = check_qp_attr(dev, qp, attr);
3208 	if (err)
3209 		goto free_ucmd;
3210 
3211 	err = create_qp(dev, pd, qp, &params);
3212 	if (err)
3213 		goto free_ucmd;
3214 
3215 	kfree(params.ucmd);
3216 	params.ucmd = NULL;
3217 
3218 	if (udata)
3219 		/*
3220 		 * It is safe to copy response for all user create QP flows,
3221 		 * including MLX5_IB_QPT_DCT, which doesn't need it.
3222 		 * In that case, resp will be filled with zeros.
3223 		 */
3224 		err = ib_copy_to_udata(udata, &params.resp, params.outlen);
3225 	if (err)
3226 		goto destroy_qp;
3227 
3228 	return 0;
3229 
3230 destroy_qp:
3231 	switch (qp->type) {
3232 	case MLX5_IB_QPT_DCT:
3233 		mlx5_ib_destroy_dct(qp);
3234 		break;
3235 	case IB_QPT_GSI:
3236 		mlx5_ib_destroy_gsi(qp);
3237 		break;
3238 	default:
3239 		destroy_qp_common(dev, qp, udata);
3240 	}
3241 
3242 free_ucmd:
3243 	kfree(params.ucmd);
3244 	return err;
3245 }
3246 
3247 int mlx5_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
3248 {
3249 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3250 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3251 
3252 	if (mqp->type == IB_QPT_GSI)
3253 		return mlx5_ib_destroy_gsi(mqp);
3254 
3255 	if (mqp->type == MLX5_IB_QPT_DCT)
3256 		return mlx5_ib_destroy_dct(mqp);
3257 
3258 	destroy_qp_common(dev, mqp, udata);
3259 	return 0;
3260 }
3261 
3262 static int set_qpc_atomic_flags(struct mlx5_ib_qp *qp,
3263 				const struct ib_qp_attr *attr, int attr_mask,
3264 				void *qpc)
3265 {
3266 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
3267 	u8 dest_rd_atomic;
3268 	u32 access_flags;
3269 
3270 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
3271 		dest_rd_atomic = attr->max_dest_rd_atomic;
3272 	else
3273 		dest_rd_atomic = qp->trans_qp.resp_depth;
3274 
3275 	if (attr_mask & IB_QP_ACCESS_FLAGS)
3276 		access_flags = attr->qp_access_flags;
3277 	else
3278 		access_flags = qp->trans_qp.atomic_rd_en;
3279 
3280 	if (!dest_rd_atomic)
3281 		access_flags &= IB_ACCESS_REMOTE_WRITE;
3282 
3283 	MLX5_SET(qpc, qpc, rre, !!(access_flags & IB_ACCESS_REMOTE_READ));
3284 
3285 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC) {
3286 		int atomic_mode;
3287 
3288 		atomic_mode = get_atomic_mode(dev, qp->type);
3289 		if (atomic_mode < 0)
3290 			return -EOPNOTSUPP;
3291 
3292 		MLX5_SET(qpc, qpc, rae, 1);
3293 		MLX5_SET(qpc, qpc, atomic_mode, atomic_mode);
3294 	}
3295 
3296 	MLX5_SET(qpc, qpc, rwe, !!(access_flags & IB_ACCESS_REMOTE_WRITE));
3297 	return 0;
3298 }
3299 
3300 enum {
3301 	MLX5_PATH_FLAG_FL	= 1 << 0,
3302 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
3303 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
3304 };
3305 
3306 static int mlx5_to_ib_rate_map(u8 rate)
3307 {
3308 	static const int rates[] = { IB_RATE_PORT_CURRENT, IB_RATE_56_GBPS,
3309 				     IB_RATE_25_GBPS,	   IB_RATE_100_GBPS,
3310 				     IB_RATE_200_GBPS,	   IB_RATE_50_GBPS,
3311 				     IB_RATE_400_GBPS };
3312 
3313 	if (rate < ARRAY_SIZE(rates))
3314 		return rates[rate];
3315 
3316 	return rate - MLX5_STAT_RATE_OFFSET;
3317 }
3318 
3319 static int ib_to_mlx5_rate_map(u8 rate)
3320 {
3321 	switch (rate) {
3322 	case IB_RATE_PORT_CURRENT:
3323 		return 0;
3324 	case IB_RATE_56_GBPS:
3325 		return 1;
3326 	case IB_RATE_25_GBPS:
3327 		return 2;
3328 	case IB_RATE_100_GBPS:
3329 		return 3;
3330 	case IB_RATE_200_GBPS:
3331 		return 4;
3332 	case IB_RATE_50_GBPS:
3333 		return 5;
3334 	case IB_RATE_400_GBPS:
3335 		return 6;
3336 	default:
3337 		return rate + MLX5_STAT_RATE_OFFSET;
3338 	}
3339 
3340 	return 0;
3341 }
3342 
3343 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
3344 {
3345 	u32 stat_rate_support;
3346 
3347 	if (rate == IB_RATE_PORT_CURRENT)
3348 		return 0;
3349 
3350 	if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_600_GBPS)
3351 		return -EINVAL;
3352 
3353 	stat_rate_support = MLX5_CAP_GEN(dev->mdev, stat_rate_support);
3354 	while (rate != IB_RATE_PORT_CURRENT &&
3355 	       !(1 << ib_to_mlx5_rate_map(rate) & stat_rate_support))
3356 		--rate;
3357 
3358 	return ib_to_mlx5_rate_map(rate);
3359 }
3360 
3361 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
3362 				      struct mlx5_ib_sq *sq, u8 sl,
3363 				      struct ib_pd *pd)
3364 {
3365 	void *in;
3366 	void *tisc;
3367 	int inlen;
3368 	int err;
3369 
3370 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
3371 	in = kvzalloc(inlen, GFP_KERNEL);
3372 	if (!in)
3373 		return -ENOMEM;
3374 
3375 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
3376 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
3377 
3378 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
3379 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
3380 
3381 	err = mlx5_core_modify_tis(dev, sq->tisn, in);
3382 
3383 	kvfree(in);
3384 
3385 	return err;
3386 }
3387 
3388 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
3389 					 struct mlx5_ib_sq *sq, u8 tx_affinity,
3390 					 struct ib_pd *pd)
3391 {
3392 	void *in;
3393 	void *tisc;
3394 	int inlen;
3395 	int err;
3396 
3397 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
3398 	in = kvzalloc(inlen, GFP_KERNEL);
3399 	if (!in)
3400 		return -ENOMEM;
3401 
3402 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
3403 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
3404 
3405 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
3406 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
3407 
3408 	err = mlx5_core_modify_tis(dev, sq->tisn, in);
3409 
3410 	kvfree(in);
3411 
3412 	return err;
3413 }
3414 
3415 static void mlx5_set_path_udp_sport(void *path, const struct rdma_ah_attr *ah,
3416 				    u32 lqpn, u32 rqpn)
3417 
3418 {
3419 	u32 fl = ah->grh.flow_label;
3420 
3421 	if (!fl)
3422 		fl = rdma_calc_flow_label(lqpn, rqpn);
3423 
3424 	MLX5_SET(ads, path, udp_sport, rdma_flow_label_to_udp_sport(fl));
3425 }
3426 
3427 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3428 			 const struct rdma_ah_attr *ah, void *path, u8 port,
3429 			 int attr_mask, u32 path_flags,
3430 			 const struct ib_qp_attr *attr, bool alt)
3431 {
3432 	const struct ib_global_route *grh = rdma_ah_read_grh(ah);
3433 	int err;
3434 	enum ib_gid_type gid_type;
3435 	u8 ah_flags = rdma_ah_get_ah_flags(ah);
3436 	u8 sl = rdma_ah_get_sl(ah);
3437 
3438 	if (attr_mask & IB_QP_PKEY_INDEX)
3439 		MLX5_SET(ads, path, pkey_index,
3440 			 alt ? attr->alt_pkey_index : attr->pkey_index);
3441 
3442 	if (ah_flags & IB_AH_GRH) {
3443 		const struct ib_port_immutable *immutable;
3444 
3445 		immutable = ib_port_immutable_read(&dev->ib_dev, port);
3446 		if (grh->sgid_index >= immutable->gid_tbl_len) {
3447 			pr_err("sgid_index (%u) too large. max is %d\n",
3448 			       grh->sgid_index,
3449 			       immutable->gid_tbl_len);
3450 			return -EINVAL;
3451 		}
3452 	}
3453 
3454 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
3455 		if (!(ah_flags & IB_AH_GRH))
3456 			return -EINVAL;
3457 
3458 		ether_addr_copy(MLX5_ADDR_OF(ads, path, rmac_47_32),
3459 				ah->roce.dmac);
3460 		if ((qp->type == IB_QPT_RC ||
3461 		     qp->type == IB_QPT_UC ||
3462 		     qp->type == IB_QPT_XRC_INI ||
3463 		     qp->type == IB_QPT_XRC_TGT) &&
3464 		    (grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) &&
3465 		    (attr_mask & IB_QP_DEST_QPN))
3466 			mlx5_set_path_udp_sport(path, ah,
3467 						qp->ibqp.qp_num,
3468 						attr->dest_qp_num);
3469 		MLX5_SET(ads, path, eth_prio, sl & 0x7);
3470 		gid_type = ah->grh.sgid_attr->gid_type;
3471 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
3472 			MLX5_SET(ads, path, dscp, grh->traffic_class >> 2);
3473 	} else {
3474 		MLX5_SET(ads, path, fl, !!(path_flags & MLX5_PATH_FLAG_FL));
3475 		MLX5_SET(ads, path, free_ar,
3476 			 !!(path_flags & MLX5_PATH_FLAG_FREE_AR));
3477 		MLX5_SET(ads, path, rlid, rdma_ah_get_dlid(ah));
3478 		MLX5_SET(ads, path, mlid, rdma_ah_get_path_bits(ah));
3479 		MLX5_SET(ads, path, grh, !!(ah_flags & IB_AH_GRH));
3480 		MLX5_SET(ads, path, sl, sl);
3481 	}
3482 
3483 	if (ah_flags & IB_AH_GRH) {
3484 		MLX5_SET(ads, path, src_addr_index, grh->sgid_index);
3485 		MLX5_SET(ads, path, hop_limit, grh->hop_limit);
3486 		MLX5_SET(ads, path, tclass, grh->traffic_class);
3487 		MLX5_SET(ads, path, flow_label, grh->flow_label);
3488 		memcpy(MLX5_ADDR_OF(ads, path, rgid_rip), grh->dgid.raw,
3489 		       sizeof(grh->dgid.raw));
3490 	}
3491 
3492 	err = ib_rate_to_mlx5(dev, rdma_ah_get_static_rate(ah));
3493 	if (err < 0)
3494 		return err;
3495 	MLX5_SET(ads, path, stat_rate, err);
3496 	MLX5_SET(ads, path, vhca_port_num, port);
3497 
3498 	if (attr_mask & IB_QP_TIMEOUT)
3499 		MLX5_SET(ads, path, ack_timeout,
3500 			 alt ? attr->alt_timeout : attr->timeout);
3501 
3502 	if ((qp->type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
3503 		return modify_raw_packet_eth_prio(dev->mdev,
3504 						  &qp->raw_packet_qp.sq,
3505 						  sl & 0xf, qp->ibqp.pd);
3506 
3507 	return 0;
3508 }
3509 
3510 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
3511 	[MLX5_QP_STATE_INIT] = {
3512 		[MLX5_QP_STATE_INIT] = {
3513 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3514 					  MLX5_QP_OPTPAR_RAE		|
3515 					  MLX5_QP_OPTPAR_RWE		|
3516 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3517 					  MLX5_QP_OPTPAR_PRI_PORT	|
3518 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3519 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3520 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3521 					  MLX5_QP_OPTPAR_PRI_PORT	|
3522 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3523 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3524 					  MLX5_QP_OPTPAR_Q_KEY		|
3525 					  MLX5_QP_OPTPAR_PRI_PORT,
3526 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3527 					  MLX5_QP_OPTPAR_RAE		|
3528 					  MLX5_QP_OPTPAR_RWE		|
3529 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3530 					  MLX5_QP_OPTPAR_PRI_PORT	|
3531 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3532 		},
3533 		[MLX5_QP_STATE_RTR] = {
3534 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3535 					  MLX5_QP_OPTPAR_RRE            |
3536 					  MLX5_QP_OPTPAR_RAE            |
3537 					  MLX5_QP_OPTPAR_RWE            |
3538 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3539 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3540 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3541 					  MLX5_QP_OPTPAR_RWE            |
3542 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3543 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3544 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
3545 					  MLX5_QP_OPTPAR_Q_KEY,
3546 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3547 					   MLX5_QP_OPTPAR_Q_KEY,
3548 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
3549 					  MLX5_QP_OPTPAR_RRE            |
3550 					  MLX5_QP_OPTPAR_RAE            |
3551 					  MLX5_QP_OPTPAR_RWE            |
3552 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3553 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3554 		},
3555 	},
3556 	[MLX5_QP_STATE_RTR] = {
3557 		[MLX5_QP_STATE_RTS] = {
3558 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3559 					  MLX5_QP_OPTPAR_RRE		|
3560 					  MLX5_QP_OPTPAR_RAE		|
3561 					  MLX5_QP_OPTPAR_RWE		|
3562 					  MLX5_QP_OPTPAR_PM_STATE	|
3563 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3564 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3565 					  MLX5_QP_OPTPAR_RWE		|
3566 					  MLX5_QP_OPTPAR_PM_STATE,
3567 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
3568 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3569 					  MLX5_QP_OPTPAR_RRE		|
3570 					  MLX5_QP_OPTPAR_RAE		|
3571 					  MLX5_QP_OPTPAR_RWE		|
3572 					  MLX5_QP_OPTPAR_PM_STATE	|
3573 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3574 		},
3575 	},
3576 	[MLX5_QP_STATE_RTS] = {
3577 		[MLX5_QP_STATE_RTS] = {
3578 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3579 					  MLX5_QP_OPTPAR_RAE		|
3580 					  MLX5_QP_OPTPAR_RWE		|
3581 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3582 					  MLX5_QP_OPTPAR_PM_STATE	|
3583 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3584 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3585 					  MLX5_QP_OPTPAR_PM_STATE	|
3586 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3587 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
3588 					  MLX5_QP_OPTPAR_SRQN		|
3589 					  MLX5_QP_OPTPAR_CQN_RCV,
3590 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3591 					  MLX5_QP_OPTPAR_RAE		|
3592 					  MLX5_QP_OPTPAR_RWE		|
3593 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3594 					  MLX5_QP_OPTPAR_PM_STATE	|
3595 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3596 		},
3597 	},
3598 	[MLX5_QP_STATE_SQER] = {
3599 		[MLX5_QP_STATE_RTS] = {
3600 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
3601 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
3602 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
3603 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3604 					   MLX5_QP_OPTPAR_RWE		|
3605 					   MLX5_QP_OPTPAR_RAE		|
3606 					   MLX5_QP_OPTPAR_RRE,
3607 			[MLX5_QP_ST_XRC]  = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3608 					   MLX5_QP_OPTPAR_RWE		|
3609 					   MLX5_QP_OPTPAR_RAE		|
3610 					   MLX5_QP_OPTPAR_RRE,
3611 		},
3612 	},
3613 	[MLX5_QP_STATE_SQD] = {
3614 		[MLX5_QP_STATE_RTS] = {
3615 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
3616 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
3617 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE,
3618 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3619 					  MLX5_QP_OPTPAR_RWE		|
3620 					  MLX5_QP_OPTPAR_RAE		|
3621 					  MLX5_QP_OPTPAR_RRE,
3622 		},
3623 	},
3624 };
3625 
3626 static int ib_nr_to_mlx5_nr(int ib_mask)
3627 {
3628 	switch (ib_mask) {
3629 	case IB_QP_STATE:
3630 		return 0;
3631 	case IB_QP_CUR_STATE:
3632 		return 0;
3633 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
3634 		return 0;
3635 	case IB_QP_ACCESS_FLAGS:
3636 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
3637 			MLX5_QP_OPTPAR_RAE;
3638 	case IB_QP_PKEY_INDEX:
3639 		return MLX5_QP_OPTPAR_PKEY_INDEX;
3640 	case IB_QP_PORT:
3641 		return MLX5_QP_OPTPAR_PRI_PORT;
3642 	case IB_QP_QKEY:
3643 		return MLX5_QP_OPTPAR_Q_KEY;
3644 	case IB_QP_AV:
3645 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
3646 			MLX5_QP_OPTPAR_PRI_PORT;
3647 	case IB_QP_PATH_MTU:
3648 		return 0;
3649 	case IB_QP_TIMEOUT:
3650 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
3651 	case IB_QP_RETRY_CNT:
3652 		return MLX5_QP_OPTPAR_RETRY_COUNT;
3653 	case IB_QP_RNR_RETRY:
3654 		return MLX5_QP_OPTPAR_RNR_RETRY;
3655 	case IB_QP_RQ_PSN:
3656 		return 0;
3657 	case IB_QP_MAX_QP_RD_ATOMIC:
3658 		return MLX5_QP_OPTPAR_SRA_MAX;
3659 	case IB_QP_ALT_PATH:
3660 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
3661 	case IB_QP_MIN_RNR_TIMER:
3662 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
3663 	case IB_QP_SQ_PSN:
3664 		return 0;
3665 	case IB_QP_MAX_DEST_RD_ATOMIC:
3666 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
3667 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
3668 	case IB_QP_PATH_MIG_STATE:
3669 		return MLX5_QP_OPTPAR_PM_STATE;
3670 	case IB_QP_CAP:
3671 		return 0;
3672 	case IB_QP_DEST_QPN:
3673 		return 0;
3674 	}
3675 	return 0;
3676 }
3677 
3678 static int ib_mask_to_mlx5_opt(int ib_mask)
3679 {
3680 	int result = 0;
3681 	int i;
3682 
3683 	for (i = 0; i < 8 * sizeof(int); i++) {
3684 		if ((1 << i) & ib_mask)
3685 			result |= ib_nr_to_mlx5_nr(1 << i);
3686 	}
3687 
3688 	return result;
3689 }
3690 
3691 static int modify_raw_packet_qp_rq(
3692 	struct mlx5_ib_dev *dev, struct mlx5_ib_rq *rq, int new_state,
3693 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3694 {
3695 	void *in;
3696 	void *rqc;
3697 	int inlen;
3698 	int err;
3699 
3700 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
3701 	in = kvzalloc(inlen, GFP_KERNEL);
3702 	if (!in)
3703 		return -ENOMEM;
3704 
3705 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
3706 	MLX5_SET(modify_rq_in, in, uid, to_mpd(pd)->uid);
3707 
3708 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
3709 	MLX5_SET(rqc, rqc, state, new_state);
3710 
3711 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
3712 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
3713 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
3714 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
3715 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
3716 		} else
3717 			dev_info_once(
3718 				&dev->ib_dev.dev,
3719 				"RAW PACKET QP counters are not supported on current FW\n");
3720 	}
3721 
3722 	err = mlx5_core_modify_rq(dev->mdev, rq->base.mqp.qpn, in);
3723 	if (err)
3724 		goto out;
3725 
3726 	rq->state = new_state;
3727 
3728 out:
3729 	kvfree(in);
3730 	return err;
3731 }
3732 
3733 static int modify_raw_packet_qp_sq(
3734 	struct mlx5_core_dev *dev, struct mlx5_ib_sq *sq, int new_state,
3735 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3736 {
3737 	struct mlx5_ib_qp *ibqp = sq->base.container_mibqp;
3738 	struct mlx5_rate_limit old_rl = ibqp->rl;
3739 	struct mlx5_rate_limit new_rl = old_rl;
3740 	bool new_rate_added = false;
3741 	u16 rl_index = 0;
3742 	void *in;
3743 	void *sqc;
3744 	int inlen;
3745 	int err;
3746 
3747 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
3748 	in = kvzalloc(inlen, GFP_KERNEL);
3749 	if (!in)
3750 		return -ENOMEM;
3751 
3752 	MLX5_SET(modify_sq_in, in, uid, to_mpd(pd)->uid);
3753 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
3754 
3755 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
3756 	MLX5_SET(sqc, sqc, state, new_state);
3757 
3758 	if (raw_qp_param->set_mask & MLX5_RAW_QP_RATE_LIMIT) {
3759 		if (new_state != MLX5_SQC_STATE_RDY)
3760 			pr_warn("%s: Rate limit can only be changed when SQ is moving to RDY\n",
3761 				__func__);
3762 		else
3763 			new_rl = raw_qp_param->rl;
3764 	}
3765 
3766 	if (!mlx5_rl_are_equal(&old_rl, &new_rl)) {
3767 		if (new_rl.rate) {
3768 			err = mlx5_rl_add_rate(dev, &rl_index, &new_rl);
3769 			if (err) {
3770 				pr_err("Failed configuring rate limit(err %d): \
3771 				       rate %u, max_burst_sz %u, typical_pkt_sz %u\n",
3772 				       err, new_rl.rate, new_rl.max_burst_sz,
3773 				       new_rl.typical_pkt_sz);
3774 
3775 				goto out;
3776 			}
3777 			new_rate_added = true;
3778 		}
3779 
3780 		MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
3781 		/* index 0 means no limit */
3782 		MLX5_SET(sqc, sqc, packet_pacing_rate_limit_index, rl_index);
3783 	}
3784 
3785 	err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in);
3786 	if (err) {
3787 		/* Remove new rate from table if failed */
3788 		if (new_rate_added)
3789 			mlx5_rl_remove_rate(dev, &new_rl);
3790 		goto out;
3791 	}
3792 
3793 	/* Only remove the old rate after new rate was set */
3794 	if ((old_rl.rate && !mlx5_rl_are_equal(&old_rl, &new_rl)) ||
3795 	    (new_state != MLX5_SQC_STATE_RDY)) {
3796 		mlx5_rl_remove_rate(dev, &old_rl);
3797 		if (new_state != MLX5_SQC_STATE_RDY)
3798 			memset(&new_rl, 0, sizeof(new_rl));
3799 	}
3800 
3801 	ibqp->rl = new_rl;
3802 	sq->state = new_state;
3803 
3804 out:
3805 	kvfree(in);
3806 	return err;
3807 }
3808 
3809 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3810 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
3811 				u8 tx_affinity)
3812 {
3813 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
3814 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
3815 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
3816 	int modify_rq = !!qp->rq.wqe_cnt;
3817 	int modify_sq = !!qp->sq.wqe_cnt;
3818 	int rq_state;
3819 	int sq_state;
3820 	int err;
3821 
3822 	switch (raw_qp_param->operation) {
3823 	case MLX5_CMD_OP_RST2INIT_QP:
3824 		rq_state = MLX5_RQC_STATE_RDY;
3825 		sq_state = MLX5_SQC_STATE_RST;
3826 		break;
3827 	case MLX5_CMD_OP_2ERR_QP:
3828 		rq_state = MLX5_RQC_STATE_ERR;
3829 		sq_state = MLX5_SQC_STATE_ERR;
3830 		break;
3831 	case MLX5_CMD_OP_2RST_QP:
3832 		rq_state = MLX5_RQC_STATE_RST;
3833 		sq_state = MLX5_SQC_STATE_RST;
3834 		break;
3835 	case MLX5_CMD_OP_RTR2RTS_QP:
3836 	case MLX5_CMD_OP_RTS2RTS_QP:
3837 		if (raw_qp_param->set_mask & ~MLX5_RAW_QP_RATE_LIMIT)
3838 			return -EINVAL;
3839 
3840 		modify_rq = 0;
3841 		sq_state = MLX5_SQC_STATE_RDY;
3842 		break;
3843 	case MLX5_CMD_OP_INIT2INIT_QP:
3844 	case MLX5_CMD_OP_INIT2RTR_QP:
3845 		if (raw_qp_param->set_mask)
3846 			return -EINVAL;
3847 		else
3848 			return 0;
3849 	default:
3850 		WARN_ON(1);
3851 		return -EINVAL;
3852 	}
3853 
3854 	if (modify_rq) {
3855 		err =  modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param,
3856 					       qp->ibqp.pd);
3857 		if (err)
3858 			return err;
3859 	}
3860 
3861 	if (modify_sq) {
3862 		struct mlx5_flow_handle *flow_rule;
3863 
3864 		if (tx_affinity) {
3865 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
3866 							    tx_affinity,
3867 							    qp->ibqp.pd);
3868 			if (err)
3869 				return err;
3870 		}
3871 
3872 		flow_rule = create_flow_rule_vport_sq(dev, sq,
3873 						      raw_qp_param->port);
3874 		if (IS_ERR(flow_rule))
3875 			return PTR_ERR(flow_rule);
3876 
3877 		err = modify_raw_packet_qp_sq(dev->mdev, sq, sq_state,
3878 					      raw_qp_param, qp->ibqp.pd);
3879 		if (err) {
3880 			if (flow_rule)
3881 				mlx5_del_flow_rules(flow_rule);
3882 			return err;
3883 		}
3884 
3885 		if (flow_rule) {
3886 			destroy_flow_rule_vport_sq(sq);
3887 			sq->flow_rule = flow_rule;
3888 		}
3889 
3890 		return err;
3891 	}
3892 
3893 	return 0;
3894 }
3895 
3896 static unsigned int get_tx_affinity_rr(struct mlx5_ib_dev *dev,
3897 				       struct ib_udata *udata)
3898 {
3899 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3900 		udata, struct mlx5_ib_ucontext, ibucontext);
3901 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3902 	atomic_t *tx_port_affinity;
3903 
3904 	if (ucontext)
3905 		tx_port_affinity = &ucontext->tx_port_affinity;
3906 	else
3907 		tx_port_affinity = &dev->port[port_num].roce.tx_port_affinity;
3908 
3909 	return (unsigned int)atomic_add_return(1, tx_port_affinity) %
3910 		MLX5_MAX_PORTS + 1;
3911 }
3912 
3913 static bool qp_supports_affinity(struct mlx5_ib_qp *qp)
3914 {
3915 	if ((qp->type == IB_QPT_RC) || (qp->type == IB_QPT_UD) ||
3916 	    (qp->type == IB_QPT_UC) || (qp->type == IB_QPT_RAW_PACKET) ||
3917 	    (qp->type == IB_QPT_XRC_INI) || (qp->type == IB_QPT_XRC_TGT) ||
3918 	    (qp->type == MLX5_IB_QPT_DCI))
3919 		return true;
3920 	return false;
3921 }
3922 
3923 static unsigned int get_tx_affinity(struct ib_qp *qp,
3924 				    const struct ib_qp_attr *attr,
3925 				    int attr_mask, u8 init,
3926 				    struct ib_udata *udata)
3927 {
3928 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3929 		udata, struct mlx5_ib_ucontext, ibucontext);
3930 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3931 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3932 	struct mlx5_ib_qp_base *qp_base;
3933 	unsigned int tx_affinity;
3934 
3935 	if (!(mlx5_ib_lag_should_assign_affinity(dev) &&
3936 	      qp_supports_affinity(mqp)))
3937 		return 0;
3938 
3939 	if (mqp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
3940 		tx_affinity = mqp->gsi_lag_port;
3941 	else if (init)
3942 		tx_affinity = get_tx_affinity_rr(dev, udata);
3943 	else if ((attr_mask & IB_QP_AV) && attr->xmit_slave)
3944 		tx_affinity =
3945 			mlx5_lag_get_slave_port(dev->mdev, attr->xmit_slave);
3946 	else
3947 		return 0;
3948 
3949 	qp_base = &mqp->trans_qp.base;
3950 	if (ucontext)
3951 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x ucontext %p\n",
3952 			    tx_affinity, qp_base->mqp.qpn, ucontext);
3953 	else
3954 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x\n",
3955 			    tx_affinity, qp_base->mqp.qpn);
3956 	return tx_affinity;
3957 }
3958 
3959 static int __mlx5_ib_qp_set_counter(struct ib_qp *qp,
3960 				    struct rdma_counter *counter)
3961 {
3962 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3963 	u32 in[MLX5_ST_SZ_DW(rts2rts_qp_in)] = {};
3964 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3965 	struct mlx5_ib_qp_base *base;
3966 	u32 set_id;
3967 	u32 *qpc;
3968 
3969 	if (counter)
3970 		set_id = counter->id;
3971 	else
3972 		set_id = mlx5_ib_get_counters_id(dev, mqp->port - 1);
3973 
3974 	base = &mqp->trans_qp.base;
3975 	MLX5_SET(rts2rts_qp_in, in, opcode, MLX5_CMD_OP_RTS2RTS_QP);
3976 	MLX5_SET(rts2rts_qp_in, in, qpn, base->mqp.qpn);
3977 	MLX5_SET(rts2rts_qp_in, in, uid, base->mqp.uid);
3978 	MLX5_SET(rts2rts_qp_in, in, opt_param_mask,
3979 		 MLX5_QP_OPTPAR_COUNTER_SET_ID);
3980 
3981 	qpc = MLX5_ADDR_OF(rts2rts_qp_in, in, qpc);
3982 	MLX5_SET(qpc, qpc, counter_set_id, set_id);
3983 	return mlx5_cmd_exec_in(dev->mdev, rts2rts_qp, in);
3984 }
3985 
3986 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
3987 			       const struct ib_qp_attr *attr, int attr_mask,
3988 			       enum ib_qp_state cur_state,
3989 			       enum ib_qp_state new_state,
3990 			       const struct mlx5_ib_modify_qp *ucmd,
3991 			       struct mlx5_ib_modify_qp_resp *resp,
3992 			       struct ib_udata *udata)
3993 {
3994 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
3995 		[MLX5_QP_STATE_RST] = {
3996 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3997 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3998 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
3999 		},
4000 		[MLX5_QP_STATE_INIT]  = {
4001 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4002 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4003 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
4004 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
4005 		},
4006 		[MLX5_QP_STATE_RTR]   = {
4007 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4008 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4009 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
4010 		},
4011 		[MLX5_QP_STATE_RTS]   = {
4012 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4013 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4014 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
4015 		},
4016 		[MLX5_QP_STATE_SQD] = {
4017 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4018 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4019 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQD_RTS_QP,
4020 		},
4021 		[MLX5_QP_STATE_SQER] = {
4022 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4023 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4024 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
4025 		},
4026 		[MLX5_QP_STATE_ERR] = {
4027 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4028 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4029 		}
4030 	};
4031 
4032 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4033 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4034 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
4035 	struct mlx5_ib_cq *send_cq, *recv_cq;
4036 	struct mlx5_ib_pd *pd;
4037 	enum mlx5_qp_state mlx5_cur, mlx5_new;
4038 	void *qpc, *pri_path, *alt_path;
4039 	enum mlx5_qp_optpar optpar = 0;
4040 	u32 set_id = 0;
4041 	int mlx5_st;
4042 	int err;
4043 	u16 op;
4044 	u8 tx_affinity = 0;
4045 
4046 	mlx5_st = to_mlx5_st(qp->type);
4047 	if (mlx5_st < 0)
4048 		return -EINVAL;
4049 
4050 	qpc = kzalloc(MLX5_ST_SZ_BYTES(qpc), GFP_KERNEL);
4051 	if (!qpc)
4052 		return -ENOMEM;
4053 
4054 	pd = to_mpd(qp->ibqp.pd);
4055 	MLX5_SET(qpc, qpc, st, mlx5_st);
4056 
4057 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
4058 		MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
4059 	} else {
4060 		switch (attr->path_mig_state) {
4061 		case IB_MIG_MIGRATED:
4062 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
4063 			break;
4064 		case IB_MIG_REARM:
4065 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_REARM);
4066 			break;
4067 		case IB_MIG_ARMED:
4068 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_ARMED);
4069 			break;
4070 		}
4071 	}
4072 
4073 	tx_affinity = get_tx_affinity(ibqp, attr, attr_mask,
4074 				      cur_state == IB_QPS_RESET &&
4075 				      new_state == IB_QPS_INIT, udata);
4076 
4077 	MLX5_SET(qpc, qpc, lag_tx_port_affinity, tx_affinity);
4078 	if (tx_affinity && new_state == IB_QPS_RTR &&
4079 	    MLX5_CAP_GEN(dev->mdev, init2_lag_tx_port_affinity))
4080 		optpar |= MLX5_QP_OPTPAR_LAG_TX_AFF;
4081 
4082 	if (is_sqp(qp->type)) {
4083 		MLX5_SET(qpc, qpc, mtu, IB_MTU_256);
4084 		MLX5_SET(qpc, qpc, log_msg_max, 8);
4085 	} else if ((qp->type == IB_QPT_UD &&
4086 		    !(qp->flags & IB_QP_CREATE_SOURCE_QPN)) ||
4087 		   qp->type == MLX5_IB_QPT_REG_UMR) {
4088 		MLX5_SET(qpc, qpc, mtu, IB_MTU_4096);
4089 		MLX5_SET(qpc, qpc, log_msg_max, 12);
4090 	} else if (attr_mask & IB_QP_PATH_MTU) {
4091 		if (attr->path_mtu < IB_MTU_256 ||
4092 		    attr->path_mtu > IB_MTU_4096) {
4093 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
4094 			err = -EINVAL;
4095 			goto out;
4096 		}
4097 		MLX5_SET(qpc, qpc, mtu, attr->path_mtu);
4098 		MLX5_SET(qpc, qpc, log_msg_max,
4099 			 MLX5_CAP_GEN(dev->mdev, log_max_msg));
4100 	}
4101 
4102 	if (attr_mask & IB_QP_DEST_QPN)
4103 		MLX5_SET(qpc, qpc, remote_qpn, attr->dest_qp_num);
4104 
4105 	pri_path = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
4106 	alt_path = MLX5_ADDR_OF(qpc, qpc, secondary_address_path);
4107 
4108 	if (attr_mask & IB_QP_PKEY_INDEX)
4109 		MLX5_SET(ads, pri_path, pkey_index, attr->pkey_index);
4110 
4111 	/* todo implement counter_index functionality */
4112 
4113 	if (is_sqp(qp->type))
4114 		MLX5_SET(ads, pri_path, vhca_port_num, qp->port);
4115 
4116 	if (attr_mask & IB_QP_PORT)
4117 		MLX5_SET(ads, pri_path, vhca_port_num, attr->port_num);
4118 
4119 	if (attr_mask & IB_QP_AV) {
4120 		err = mlx5_set_path(dev, qp, &attr->ah_attr, pri_path,
4121 				    attr_mask & IB_QP_PORT ? attr->port_num :
4122 							     qp->port,
4123 				    attr_mask, 0, attr, false);
4124 		if (err)
4125 			goto out;
4126 	}
4127 
4128 	if (attr_mask & IB_QP_TIMEOUT)
4129 		MLX5_SET(ads, pri_path, ack_timeout, attr->timeout);
4130 
4131 	if (attr_mask & IB_QP_ALT_PATH) {
4132 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr, alt_path,
4133 				    attr->alt_port_num,
4134 				    attr_mask | IB_QP_PKEY_INDEX |
4135 					    IB_QP_TIMEOUT,
4136 				    0, attr, true);
4137 		if (err)
4138 			goto out;
4139 	}
4140 
4141 	get_cqs(qp->type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
4142 		&send_cq, &recv_cq);
4143 
4144 	MLX5_SET(qpc, qpc, pd, pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
4145 	if (send_cq)
4146 		MLX5_SET(qpc, qpc, cqn_snd, send_cq->mcq.cqn);
4147 	if (recv_cq)
4148 		MLX5_SET(qpc, qpc, cqn_rcv, recv_cq->mcq.cqn);
4149 
4150 	MLX5_SET(qpc, qpc, log_ack_req_freq, MLX5_IB_ACK_REQ_FREQ);
4151 
4152 	if (attr_mask & IB_QP_RNR_RETRY)
4153 		MLX5_SET(qpc, qpc, rnr_retry, attr->rnr_retry);
4154 
4155 	if (attr_mask & IB_QP_RETRY_CNT)
4156 		MLX5_SET(qpc, qpc, retry_count, attr->retry_cnt);
4157 
4158 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC && attr->max_rd_atomic)
4159 		MLX5_SET(qpc, qpc, log_sra_max, ilog2(attr->max_rd_atomic));
4160 
4161 	if (attr_mask & IB_QP_SQ_PSN)
4162 		MLX5_SET(qpc, qpc, next_send_psn, attr->sq_psn);
4163 
4164 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC && attr->max_dest_rd_atomic)
4165 		MLX5_SET(qpc, qpc, log_rra_max,
4166 			 ilog2(attr->max_dest_rd_atomic));
4167 
4168 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
4169 		err = set_qpc_atomic_flags(qp, attr, attr_mask, qpc);
4170 		if (err)
4171 			goto out;
4172 	}
4173 
4174 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
4175 		MLX5_SET(qpc, qpc, min_rnr_nak, attr->min_rnr_timer);
4176 
4177 	if (attr_mask & IB_QP_RQ_PSN)
4178 		MLX5_SET(qpc, qpc, next_rcv_psn, attr->rq_psn);
4179 
4180 	if (attr_mask & IB_QP_QKEY)
4181 		MLX5_SET(qpc, qpc, q_key, attr->qkey);
4182 
4183 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
4184 		MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
4185 
4186 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4187 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
4188 			       qp->port) - 1;
4189 
4190 		/* Underlay port should be used - index 0 function per port */
4191 		if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
4192 			port_num = 0;
4193 
4194 		if (ibqp->counter)
4195 			set_id = ibqp->counter->id;
4196 		else
4197 			set_id = mlx5_ib_get_counters_id(dev, port_num);
4198 		MLX5_SET(qpc, qpc, counter_set_id, set_id);
4199 	}
4200 
4201 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
4202 		MLX5_SET(qpc, qpc, rlky, 1);
4203 
4204 	if (qp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
4205 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
4206 
4207 	mlx5_cur = to_mlx5_state(cur_state);
4208 	mlx5_new = to_mlx5_state(new_state);
4209 
4210 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
4211 	    !optab[mlx5_cur][mlx5_new]) {
4212 		err = -EINVAL;
4213 		goto out;
4214 	}
4215 
4216 	op = optab[mlx5_cur][mlx5_new];
4217 	optpar |= ib_mask_to_mlx5_opt(attr_mask);
4218 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
4219 
4220 	if (qp->type == IB_QPT_RAW_PACKET ||
4221 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4222 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
4223 
4224 		raw_qp_param.operation = op;
4225 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4226 			raw_qp_param.rq_q_ctr_id = set_id;
4227 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
4228 		}
4229 
4230 		if (attr_mask & IB_QP_PORT)
4231 			raw_qp_param.port = attr->port_num;
4232 
4233 		if (attr_mask & IB_QP_RATE_LIMIT) {
4234 			raw_qp_param.rl.rate = attr->rate_limit;
4235 
4236 			if (ucmd->burst_info.max_burst_sz) {
4237 				if (attr->rate_limit &&
4238 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_burst_bound)) {
4239 					raw_qp_param.rl.max_burst_sz =
4240 						ucmd->burst_info.max_burst_sz;
4241 				} else {
4242 					err = -EINVAL;
4243 					goto out;
4244 				}
4245 			}
4246 
4247 			if (ucmd->burst_info.typical_pkt_sz) {
4248 				if (attr->rate_limit &&
4249 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_typical_size)) {
4250 					raw_qp_param.rl.typical_pkt_sz =
4251 						ucmd->burst_info.typical_pkt_sz;
4252 				} else {
4253 					err = -EINVAL;
4254 					goto out;
4255 				}
4256 			}
4257 
4258 			raw_qp_param.set_mask |= MLX5_RAW_QP_RATE_LIMIT;
4259 		}
4260 
4261 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, tx_affinity);
4262 	} else {
4263 		if (udata) {
4264 			/* For the kernel flows, the resp will stay zero */
4265 			resp->ece_options =
4266 				MLX5_CAP_GEN(dev->mdev, ece_support) ?
4267 					ucmd->ece_options : 0;
4268 			resp->response_length = sizeof(*resp);
4269 		}
4270 		err = mlx5_core_qp_modify(dev, op, optpar, qpc, &base->mqp,
4271 					  &resp->ece_options);
4272 	}
4273 
4274 	if (err)
4275 		goto out;
4276 
4277 	qp->state = new_state;
4278 
4279 	if (attr_mask & IB_QP_ACCESS_FLAGS)
4280 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
4281 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
4282 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
4283 	if (attr_mask & IB_QP_PORT)
4284 		qp->port = attr->port_num;
4285 	if (attr_mask & IB_QP_ALT_PATH)
4286 		qp->trans_qp.alt_port = attr->alt_port_num;
4287 
4288 	/*
4289 	 * If we moved a kernel QP to RESET, clean up all old CQ
4290 	 * entries and reinitialize the QP.
4291 	 */
4292 	if (new_state == IB_QPS_RESET &&
4293 	    !ibqp->uobject && qp->type != IB_QPT_XRC_TGT) {
4294 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
4295 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
4296 		if (send_cq != recv_cq)
4297 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
4298 
4299 		qp->rq.head = 0;
4300 		qp->rq.tail = 0;
4301 		qp->sq.head = 0;
4302 		qp->sq.tail = 0;
4303 		qp->sq.cur_post = 0;
4304 		if (qp->sq.wqe_cnt)
4305 			qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
4306 		qp->sq.last_poll = 0;
4307 		qp->db.db[MLX5_RCV_DBR] = 0;
4308 		qp->db.db[MLX5_SND_DBR] = 0;
4309 	}
4310 
4311 	if ((new_state == IB_QPS_RTS) && qp->counter_pending) {
4312 		err = __mlx5_ib_qp_set_counter(ibqp, ibqp->counter);
4313 		if (!err)
4314 			qp->counter_pending = 0;
4315 	}
4316 
4317 out:
4318 	kfree(qpc);
4319 	return err;
4320 }
4321 
4322 static inline bool is_valid_mask(int mask, int req, int opt)
4323 {
4324 	if ((mask & req) != req)
4325 		return false;
4326 
4327 	if (mask & ~(req | opt))
4328 		return false;
4329 
4330 	return true;
4331 }
4332 
4333 /* check valid transition for driver QP types
4334  * for now the only QP type that this function supports is DCI
4335  */
4336 static bool modify_dci_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state new_state,
4337 				enum ib_qp_attr_mask attr_mask)
4338 {
4339 	int req = IB_QP_STATE;
4340 	int opt = 0;
4341 
4342 	if (new_state == IB_QPS_RESET) {
4343 		return is_valid_mask(attr_mask, req, opt);
4344 	} else if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4345 		req |= IB_QP_PKEY_INDEX | IB_QP_PORT;
4346 		return is_valid_mask(attr_mask, req, opt);
4347 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
4348 		opt = IB_QP_PKEY_INDEX | IB_QP_PORT;
4349 		return is_valid_mask(attr_mask, req, opt);
4350 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
4351 		req |= IB_QP_PATH_MTU;
4352 		opt = IB_QP_PKEY_INDEX | IB_QP_AV;
4353 		return is_valid_mask(attr_mask, req, opt);
4354 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
4355 		req |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4356 		       IB_QP_MAX_QP_RD_ATOMIC | IB_QP_SQ_PSN;
4357 		opt = IB_QP_MIN_RNR_TIMER;
4358 		return is_valid_mask(attr_mask, req, opt);
4359 	} else if (cur_state == IB_QPS_RTS && new_state == IB_QPS_RTS) {
4360 		opt = IB_QP_MIN_RNR_TIMER;
4361 		return is_valid_mask(attr_mask, req, opt);
4362 	} else if (cur_state != IB_QPS_RESET && new_state == IB_QPS_ERR) {
4363 		return is_valid_mask(attr_mask, req, opt);
4364 	}
4365 	return false;
4366 }
4367 
4368 /* mlx5_ib_modify_dct: modify a DCT QP
4369  * valid transitions are:
4370  * RESET to INIT: must set access_flags, pkey_index and port
4371  * INIT  to RTR : must set min_rnr_timer, tclass, flow_label,
4372  *			   mtu, gid_index and hop_limit
4373  * Other transitions and attributes are illegal
4374  */
4375 static int mlx5_ib_modify_dct(struct ib_qp *ibqp, struct ib_qp_attr *attr,
4376 			      int attr_mask, struct mlx5_ib_modify_qp *ucmd,
4377 			      struct ib_udata *udata)
4378 {
4379 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4380 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4381 	enum ib_qp_state cur_state, new_state;
4382 	int required = IB_QP_STATE;
4383 	void *dctc;
4384 	int err;
4385 
4386 	if (!(attr_mask & IB_QP_STATE))
4387 		return -EINVAL;
4388 
4389 	cur_state = qp->state;
4390 	new_state = attr->qp_state;
4391 
4392 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
4393 	if (MLX5_CAP_GEN(dev->mdev, ece_support) && ucmd->ece_options)
4394 		/*
4395 		 * DCT doesn't initialize QP till modify command is executed,
4396 		 * so we need to overwrite previously set ECE field if user
4397 		 * provided any value except zero, which means not set/not
4398 		 * valid.
4399 		 */
4400 		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
4401 
4402 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4403 		u16 set_id;
4404 
4405 		required |= IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
4406 		if (!is_valid_mask(attr_mask, required, 0))
4407 			return -EINVAL;
4408 
4409 		if (attr->port_num == 0 ||
4410 		    attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)) {
4411 			mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4412 				    attr->port_num, dev->num_ports);
4413 			return -EINVAL;
4414 		}
4415 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
4416 			MLX5_SET(dctc, dctc, rre, 1);
4417 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
4418 			MLX5_SET(dctc, dctc, rwe, 1);
4419 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC) {
4420 			int atomic_mode;
4421 
4422 			atomic_mode = get_atomic_mode(dev, MLX5_IB_QPT_DCT);
4423 			if (atomic_mode < 0)
4424 				return -EOPNOTSUPP;
4425 
4426 			MLX5_SET(dctc, dctc, atomic_mode, atomic_mode);
4427 			MLX5_SET(dctc, dctc, rae, 1);
4428 		}
4429 		MLX5_SET(dctc, dctc, pkey_index, attr->pkey_index);
4430 		if (mlx5_lag_is_active(dev->mdev))
4431 			MLX5_SET(dctc, dctc, port,
4432 				 get_tx_affinity_rr(dev, udata));
4433 		else
4434 			MLX5_SET(dctc, dctc, port, attr->port_num);
4435 
4436 		set_id = mlx5_ib_get_counters_id(dev, attr->port_num - 1);
4437 		MLX5_SET(dctc, dctc, counter_set_id, set_id);
4438 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
4439 		struct mlx5_ib_modify_qp_resp resp = {};
4440 		u32 out[MLX5_ST_SZ_DW(create_dct_out)] = {};
4441 		u32 min_resp_len = offsetofend(typeof(resp), dctn);
4442 
4443 		if (udata->outlen < min_resp_len)
4444 			return -EINVAL;
4445 		/*
4446 		 * If we don't have enough space for the ECE options,
4447 		 * simply indicate it with resp.response_length.
4448 		 */
4449 		resp.response_length = (udata->outlen < sizeof(resp)) ?
4450 					       min_resp_len :
4451 					       sizeof(resp);
4452 
4453 		required |= IB_QP_MIN_RNR_TIMER | IB_QP_AV | IB_QP_PATH_MTU;
4454 		if (!is_valid_mask(attr_mask, required, 0))
4455 			return -EINVAL;
4456 		MLX5_SET(dctc, dctc, min_rnr_nak, attr->min_rnr_timer);
4457 		MLX5_SET(dctc, dctc, tclass, attr->ah_attr.grh.traffic_class);
4458 		MLX5_SET(dctc, dctc, flow_label, attr->ah_attr.grh.flow_label);
4459 		MLX5_SET(dctc, dctc, mtu, attr->path_mtu);
4460 		MLX5_SET(dctc, dctc, my_addr_index, attr->ah_attr.grh.sgid_index);
4461 		MLX5_SET(dctc, dctc, hop_limit, attr->ah_attr.grh.hop_limit);
4462 		if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
4463 			MLX5_SET(dctc, dctc, eth_prio, attr->ah_attr.sl & 0x7);
4464 
4465 		err = mlx5_core_create_dct(dev, &qp->dct.mdct, qp->dct.in,
4466 					   MLX5_ST_SZ_BYTES(create_dct_in), out,
4467 					   sizeof(out));
4468 		err = mlx5_cmd_check(dev->mdev, err, qp->dct.in, out);
4469 		if (err)
4470 			return err;
4471 		resp.dctn = qp->dct.mdct.mqp.qpn;
4472 		if (MLX5_CAP_GEN(dev->mdev, ece_support))
4473 			resp.ece_options = MLX5_GET(create_dct_out, out, ece);
4474 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4475 		if (err) {
4476 			mlx5_core_destroy_dct(dev, &qp->dct.mdct);
4477 			return err;
4478 		}
4479 	} else {
4480 		mlx5_ib_warn(dev, "Modify DCT: Invalid transition from %d to %d\n", cur_state, new_state);
4481 		return -EINVAL;
4482 	}
4483 
4484 	qp->state = new_state;
4485 	return 0;
4486 }
4487 
4488 static bool mlx5_ib_modify_qp_allowed(struct mlx5_ib_dev *dev,
4489 				      struct mlx5_ib_qp *qp)
4490 {
4491 	if (dev->profile != &raw_eth_profile)
4492 		return true;
4493 
4494 	if (qp->type == IB_QPT_RAW_PACKET || qp->type == MLX5_IB_QPT_REG_UMR)
4495 		return true;
4496 
4497 	/* Internal QP used for wc testing, with NOPs in wq */
4498 	if (qp->flags & MLX5_IB_QP_CREATE_WC_TEST)
4499 		return true;
4500 
4501 	return false;
4502 }
4503 
4504 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
4505 		      int attr_mask, struct ib_udata *udata)
4506 {
4507 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4508 	struct mlx5_ib_modify_qp_resp resp = {};
4509 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4510 	struct mlx5_ib_modify_qp ucmd = {};
4511 	enum ib_qp_type qp_type;
4512 	enum ib_qp_state cur_state, new_state;
4513 	int err = -EINVAL;
4514 
4515 	if (!mlx5_ib_modify_qp_allowed(dev, qp))
4516 		return -EOPNOTSUPP;
4517 
4518 	if (attr_mask & ~(IB_QP_ATTR_STANDARD_BITS | IB_QP_RATE_LIMIT))
4519 		return -EOPNOTSUPP;
4520 
4521 	if (ibqp->rwq_ind_tbl)
4522 		return -ENOSYS;
4523 
4524 	if (udata && udata->inlen) {
4525 		if (udata->inlen < offsetofend(typeof(ucmd), ece_options))
4526 			return -EINVAL;
4527 
4528 		if (udata->inlen > sizeof(ucmd) &&
4529 		    !ib_is_udata_cleared(udata, sizeof(ucmd),
4530 					 udata->inlen - sizeof(ucmd)))
4531 			return -EOPNOTSUPP;
4532 
4533 		if (ib_copy_from_udata(&ucmd, udata,
4534 				       min(udata->inlen, sizeof(ucmd))))
4535 			return -EFAULT;
4536 
4537 		if (ucmd.comp_mask ||
4538 		    memchr_inv(&ucmd.burst_info.reserved, 0,
4539 			       sizeof(ucmd.burst_info.reserved)))
4540 			return -EOPNOTSUPP;
4541 
4542 	}
4543 
4544 	if (qp->type == IB_QPT_GSI)
4545 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
4546 
4547 	qp_type = (qp->type == MLX5_IB_QPT_HW_GSI) ? IB_QPT_GSI : qp->type;
4548 
4549 	if (qp_type == MLX5_IB_QPT_DCT)
4550 		return mlx5_ib_modify_dct(ibqp, attr, attr_mask, &ucmd, udata);
4551 
4552 	mutex_lock(&qp->mutex);
4553 
4554 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
4555 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
4556 
4557 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4558 		if (attr_mask & ~(IB_QP_STATE | IB_QP_CUR_STATE)) {
4559 			mlx5_ib_dbg(dev, "invalid attr_mask 0x%x when underlay QP is used\n",
4560 				    attr_mask);
4561 			goto out;
4562 		}
4563 	} else if (qp_type != MLX5_IB_QPT_REG_UMR &&
4564 		   qp_type != MLX5_IB_QPT_DCI &&
4565 		   !ib_modify_qp_is_ok(cur_state, new_state, qp_type,
4566 				       attr_mask)) {
4567 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4568 			    cur_state, new_state, qp->type, attr_mask);
4569 		goto out;
4570 	} else if (qp_type == MLX5_IB_QPT_DCI &&
4571 		   !modify_dci_qp_is_ok(cur_state, new_state, attr_mask)) {
4572 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4573 			    cur_state, new_state, qp_type, attr_mask);
4574 		goto out;
4575 	}
4576 
4577 	if ((attr_mask & IB_QP_PORT) &&
4578 	    (attr->port_num == 0 ||
4579 	     attr->port_num > dev->num_ports)) {
4580 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4581 			    attr->port_num, dev->num_ports);
4582 		goto out;
4583 	}
4584 
4585 	if ((attr_mask & IB_QP_PKEY_INDEX) &&
4586 	    attr->pkey_index >= dev->pkey_table_len) {
4587 		mlx5_ib_dbg(dev, "invalid pkey index %d\n", attr->pkey_index);
4588 		goto out;
4589 	}
4590 
4591 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
4592 	    attr->max_rd_atomic >
4593 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
4594 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
4595 			    attr->max_rd_atomic);
4596 		goto out;
4597 	}
4598 
4599 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
4600 	    attr->max_dest_rd_atomic >
4601 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
4602 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
4603 			    attr->max_dest_rd_atomic);
4604 		goto out;
4605 	}
4606 
4607 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
4608 		err = 0;
4609 		goto out;
4610 	}
4611 
4612 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state,
4613 				  new_state, &ucmd, &resp, udata);
4614 
4615 	/* resp.response_length is set in ECE supported flows only */
4616 	if (!err && resp.response_length &&
4617 	    udata->outlen >= resp.response_length)
4618 		/* Return -EFAULT to the user and expect him to destroy QP. */
4619 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4620 
4621 out:
4622 	mutex_unlock(&qp->mutex);
4623 	return err;
4624 }
4625 
4626 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4627 {
4628 	switch (mlx5_state) {
4629 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
4630 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
4631 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
4632 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
4633 	case MLX5_QP_STATE_SQ_DRAINING:
4634 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
4635 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
4636 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
4637 	default:		     return -1;
4638 	}
4639 }
4640 
4641 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4642 {
4643 	switch (mlx5_mig_state) {
4644 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
4645 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
4646 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4647 	default: return -1;
4648 	}
4649 }
4650 
4651 static void to_rdma_ah_attr(struct mlx5_ib_dev *ibdev,
4652 			    struct rdma_ah_attr *ah_attr, void *path)
4653 {
4654 	int port = MLX5_GET(ads, path, vhca_port_num);
4655 	int static_rate;
4656 
4657 	memset(ah_attr, 0, sizeof(*ah_attr));
4658 
4659 	if (!port || port > ibdev->num_ports)
4660 		return;
4661 
4662 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port);
4663 
4664 	rdma_ah_set_port_num(ah_attr, port);
4665 	rdma_ah_set_sl(ah_attr, MLX5_GET(ads, path, sl));
4666 
4667 	rdma_ah_set_dlid(ah_attr, MLX5_GET(ads, path, rlid));
4668 	rdma_ah_set_path_bits(ah_attr, MLX5_GET(ads, path, mlid));
4669 
4670 	static_rate = MLX5_GET(ads, path, stat_rate);
4671 	rdma_ah_set_static_rate(ah_attr, mlx5_to_ib_rate_map(static_rate));
4672 	if (MLX5_GET(ads, path, grh) ||
4673 	    ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
4674 		rdma_ah_set_grh(ah_attr, NULL, MLX5_GET(ads, path, flow_label),
4675 				MLX5_GET(ads, path, src_addr_index),
4676 				MLX5_GET(ads, path, hop_limit),
4677 				MLX5_GET(ads, path, tclass));
4678 		rdma_ah_set_dgid_raw(ah_attr, MLX5_ADDR_OF(ads, path, rgid_rip));
4679 	}
4680 }
4681 
4682 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4683 					struct mlx5_ib_sq *sq,
4684 					u8 *sq_state)
4685 {
4686 	int err;
4687 
4688 	err = mlx5_core_query_sq_state(dev->mdev, sq->base.mqp.qpn, sq_state);
4689 	if (err)
4690 		goto out;
4691 	sq->state = *sq_state;
4692 
4693 out:
4694 	return err;
4695 }
4696 
4697 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4698 					struct mlx5_ib_rq *rq,
4699 					u8 *rq_state)
4700 {
4701 	void *out;
4702 	void *rqc;
4703 	int inlen;
4704 	int err;
4705 
4706 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4707 	out = kvzalloc(inlen, GFP_KERNEL);
4708 	if (!out)
4709 		return -ENOMEM;
4710 
4711 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4712 	if (err)
4713 		goto out;
4714 
4715 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4716 	*rq_state = MLX5_GET(rqc, rqc, state);
4717 	rq->state = *rq_state;
4718 
4719 out:
4720 	kvfree(out);
4721 	return err;
4722 }
4723 
4724 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4725 				  struct mlx5_ib_qp *qp, u8 *qp_state)
4726 {
4727 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4728 		[MLX5_RQC_STATE_RST] = {
4729 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
4730 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4731 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
4732 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
4733 		},
4734 		[MLX5_RQC_STATE_RDY] = {
4735 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE,
4736 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4737 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
4738 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
4739 		},
4740 		[MLX5_RQC_STATE_ERR] = {
4741 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
4742 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4743 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
4744 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
4745 		},
4746 		[MLX5_RQ_STATE_NA] = {
4747 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE,
4748 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4749 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
4750 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
4751 		},
4752 	};
4753 
4754 	*qp_state = sqrq_trans[rq_state][sq_state];
4755 
4756 	if (*qp_state == MLX5_QP_STATE_BAD) {
4757 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4758 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4759 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4760 		return -EINVAL;
4761 	}
4762 
4763 	if (*qp_state == MLX5_QP_STATE)
4764 		*qp_state = qp->state;
4765 
4766 	return 0;
4767 }
4768 
4769 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4770 				     struct mlx5_ib_qp *qp,
4771 				     u8 *raw_packet_qp_state)
4772 {
4773 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4774 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4775 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4776 	int err;
4777 	u8 sq_state = MLX5_SQ_STATE_NA;
4778 	u8 rq_state = MLX5_RQ_STATE_NA;
4779 
4780 	if (qp->sq.wqe_cnt) {
4781 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4782 		if (err)
4783 			return err;
4784 	}
4785 
4786 	if (qp->rq.wqe_cnt) {
4787 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4788 		if (err)
4789 			return err;
4790 	}
4791 
4792 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4793 				      raw_packet_qp_state);
4794 }
4795 
4796 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4797 			 struct ib_qp_attr *qp_attr)
4798 {
4799 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
4800 	void *qpc, *pri_path, *alt_path;
4801 	u32 *outb;
4802 	int err;
4803 
4804 	outb = kzalloc(outlen, GFP_KERNEL);
4805 	if (!outb)
4806 		return -ENOMEM;
4807 
4808 	err = mlx5_core_qp_query(dev, &qp->trans_qp.base.mqp, outb, outlen);
4809 	if (err)
4810 		goto out;
4811 
4812 	qpc = MLX5_ADDR_OF(query_qp_out, outb, qpc);
4813 
4814 	qp->state = to_ib_qp_state(MLX5_GET(qpc, qpc, state));
4815 	if (MLX5_GET(qpc, qpc, state) == MLX5_QP_STATE_SQ_DRAINING)
4816 		qp_attr->sq_draining = 1;
4817 
4818 	qp_attr->path_mtu = MLX5_GET(qpc, qpc, mtu);
4819 	qp_attr->path_mig_state = to_ib_mig_state(MLX5_GET(qpc, qpc, pm_state));
4820 	qp_attr->qkey = MLX5_GET(qpc, qpc, q_key);
4821 	qp_attr->rq_psn = MLX5_GET(qpc, qpc, next_rcv_psn);
4822 	qp_attr->sq_psn = MLX5_GET(qpc, qpc, next_send_psn);
4823 	qp_attr->dest_qp_num = MLX5_GET(qpc, qpc, remote_qpn);
4824 
4825 	if (MLX5_GET(qpc, qpc, rre))
4826 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ;
4827 	if (MLX5_GET(qpc, qpc, rwe))
4828 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_WRITE;
4829 	if (MLX5_GET(qpc, qpc, rae))
4830 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_ATOMIC;
4831 
4832 	qp_attr->max_rd_atomic = 1 << MLX5_GET(qpc, qpc, log_sra_max);
4833 	qp_attr->max_dest_rd_atomic = 1 << MLX5_GET(qpc, qpc, log_rra_max);
4834 	qp_attr->min_rnr_timer = MLX5_GET(qpc, qpc, min_rnr_nak);
4835 	qp_attr->retry_cnt = MLX5_GET(qpc, qpc, retry_count);
4836 	qp_attr->rnr_retry = MLX5_GET(qpc, qpc, rnr_retry);
4837 
4838 	pri_path = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
4839 	alt_path = MLX5_ADDR_OF(qpc, qpc, secondary_address_path);
4840 
4841 	if (qp->type == IB_QPT_RC || qp->type == IB_QPT_UC ||
4842 	    qp->type == IB_QPT_XRC_INI || qp->type == IB_QPT_XRC_TGT) {
4843 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, pri_path);
4844 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, alt_path);
4845 		qp_attr->alt_pkey_index = MLX5_GET(ads, alt_path, pkey_index);
4846 		qp_attr->alt_port_num = MLX5_GET(ads, alt_path, vhca_port_num);
4847 	}
4848 
4849 	qp_attr->pkey_index = MLX5_GET(ads, pri_path, pkey_index);
4850 	qp_attr->port_num = MLX5_GET(ads, pri_path, vhca_port_num);
4851 	qp_attr->timeout = MLX5_GET(ads, pri_path, ack_timeout);
4852 	qp_attr->alt_timeout = MLX5_GET(ads, alt_path, ack_timeout);
4853 
4854 out:
4855 	kfree(outb);
4856 	return err;
4857 }
4858 
4859 static int mlx5_ib_dct_query_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *mqp,
4860 				struct ib_qp_attr *qp_attr, int qp_attr_mask,
4861 				struct ib_qp_init_attr *qp_init_attr)
4862 {
4863 	struct mlx5_core_dct	*dct = &mqp->dct.mdct;
4864 	u32 *out;
4865 	u32 access_flags = 0;
4866 	int outlen = MLX5_ST_SZ_BYTES(query_dct_out);
4867 	void *dctc;
4868 	int err;
4869 	int supported_mask = IB_QP_STATE |
4870 			     IB_QP_ACCESS_FLAGS |
4871 			     IB_QP_PORT |
4872 			     IB_QP_MIN_RNR_TIMER |
4873 			     IB_QP_AV |
4874 			     IB_QP_PATH_MTU |
4875 			     IB_QP_PKEY_INDEX;
4876 
4877 	if (qp_attr_mask & ~supported_mask)
4878 		return -EINVAL;
4879 	if (mqp->state != IB_QPS_RTR)
4880 		return -EINVAL;
4881 
4882 	out = kzalloc(outlen, GFP_KERNEL);
4883 	if (!out)
4884 		return -ENOMEM;
4885 
4886 	err = mlx5_core_dct_query(dev, dct, out, outlen);
4887 	if (err)
4888 		goto out;
4889 
4890 	dctc = MLX5_ADDR_OF(query_dct_out, out, dct_context_entry);
4891 
4892 	if (qp_attr_mask & IB_QP_STATE)
4893 		qp_attr->qp_state = IB_QPS_RTR;
4894 
4895 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
4896 		if (MLX5_GET(dctc, dctc, rre))
4897 			access_flags |= IB_ACCESS_REMOTE_READ;
4898 		if (MLX5_GET(dctc, dctc, rwe))
4899 			access_flags |= IB_ACCESS_REMOTE_WRITE;
4900 		if (MLX5_GET(dctc, dctc, rae))
4901 			access_flags |= IB_ACCESS_REMOTE_ATOMIC;
4902 		qp_attr->qp_access_flags = access_flags;
4903 	}
4904 
4905 	if (qp_attr_mask & IB_QP_PORT)
4906 		qp_attr->port_num = MLX5_GET(dctc, dctc, port);
4907 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER)
4908 		qp_attr->min_rnr_timer = MLX5_GET(dctc, dctc, min_rnr_nak);
4909 	if (qp_attr_mask & IB_QP_AV) {
4910 		qp_attr->ah_attr.grh.traffic_class = MLX5_GET(dctc, dctc, tclass);
4911 		qp_attr->ah_attr.grh.flow_label = MLX5_GET(dctc, dctc, flow_label);
4912 		qp_attr->ah_attr.grh.sgid_index = MLX5_GET(dctc, dctc, my_addr_index);
4913 		qp_attr->ah_attr.grh.hop_limit = MLX5_GET(dctc, dctc, hop_limit);
4914 	}
4915 	if (qp_attr_mask & IB_QP_PATH_MTU)
4916 		qp_attr->path_mtu = MLX5_GET(dctc, dctc, mtu);
4917 	if (qp_attr_mask & IB_QP_PKEY_INDEX)
4918 		qp_attr->pkey_index = MLX5_GET(dctc, dctc, pkey_index);
4919 out:
4920 	kfree(out);
4921 	return err;
4922 }
4923 
4924 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4925 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4926 {
4927 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4928 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4929 	int err = 0;
4930 	u8 raw_packet_qp_state;
4931 
4932 	if (ibqp->rwq_ind_tbl)
4933 		return -ENOSYS;
4934 
4935 	if (qp->type == IB_QPT_GSI)
4936 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4937 					    qp_init_attr);
4938 
4939 	/* Not all of output fields are applicable, make sure to zero them */
4940 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
4941 	memset(qp_attr, 0, sizeof(*qp_attr));
4942 
4943 	if (unlikely(qp->type == MLX5_IB_QPT_DCT))
4944 		return mlx5_ib_dct_query_qp(dev, qp, qp_attr,
4945 					    qp_attr_mask, qp_init_attr);
4946 
4947 	mutex_lock(&qp->mutex);
4948 
4949 	if (qp->type == IB_QPT_RAW_PACKET ||
4950 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4951 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4952 		if (err)
4953 			goto out;
4954 		qp->state = raw_packet_qp_state;
4955 		qp_attr->port_num = 1;
4956 	} else {
4957 		err = query_qp_attr(dev, qp, qp_attr);
4958 		if (err)
4959 			goto out;
4960 	}
4961 
4962 	qp_attr->qp_state	     = qp->state;
4963 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4964 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4965 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4966 
4967 	if (!ibqp->uobject) {
4968 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
4969 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4970 		qp_init_attr->qp_context = ibqp->qp_context;
4971 	} else {
4972 		qp_attr->cap.max_send_wr  = 0;
4973 		qp_attr->cap.max_send_sge = 0;
4974 	}
4975 
4976 	qp_init_attr->qp_type = qp->type;
4977 	qp_init_attr->recv_cq = ibqp->recv_cq;
4978 	qp_init_attr->send_cq = ibqp->send_cq;
4979 	qp_init_attr->srq = ibqp->srq;
4980 	qp_attr->cap.max_inline_data = qp->max_inline_data;
4981 
4982 	qp_init_attr->cap	     = qp_attr->cap;
4983 
4984 	qp_init_attr->create_flags = qp->flags;
4985 
4986 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
4987 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4988 
4989 out:
4990 	mutex_unlock(&qp->mutex);
4991 	return err;
4992 }
4993 
4994 int mlx5_ib_alloc_xrcd(struct ib_xrcd *ibxrcd, struct ib_udata *udata)
4995 {
4996 	struct mlx5_ib_dev *dev = to_mdev(ibxrcd->device);
4997 	struct mlx5_ib_xrcd *xrcd = to_mxrcd(ibxrcd);
4998 
4999 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
5000 		return -EOPNOTSUPP;
5001 
5002 	return mlx5_cmd_xrcd_alloc(dev->mdev, &xrcd->xrcdn, 0);
5003 }
5004 
5005 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
5006 {
5007 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
5008 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
5009 
5010 	return mlx5_cmd_xrcd_dealloc(dev->mdev, xrcdn, 0);
5011 }
5012 
5013 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
5014 {
5015 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
5016 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
5017 	struct ib_event event;
5018 
5019 	if (rwq->ibwq.event_handler) {
5020 		event.device     = rwq->ibwq.device;
5021 		event.element.wq = &rwq->ibwq;
5022 		switch (type) {
5023 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
5024 			event.event = IB_EVENT_WQ_FATAL;
5025 			break;
5026 		default:
5027 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
5028 			return;
5029 		}
5030 
5031 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
5032 	}
5033 }
5034 
5035 static int set_delay_drop(struct mlx5_ib_dev *dev)
5036 {
5037 	int err = 0;
5038 
5039 	mutex_lock(&dev->delay_drop.lock);
5040 	if (dev->delay_drop.activate)
5041 		goto out;
5042 
5043 	err = mlx5_core_set_delay_drop(dev, dev->delay_drop.timeout);
5044 	if (err)
5045 		goto out;
5046 
5047 	dev->delay_drop.activate = true;
5048 out:
5049 	mutex_unlock(&dev->delay_drop.lock);
5050 
5051 	if (!err)
5052 		atomic_inc(&dev->delay_drop.rqs_cnt);
5053 	return err;
5054 }
5055 
5056 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
5057 		      struct ib_wq_init_attr *init_attr)
5058 {
5059 	struct mlx5_ib_dev *dev;
5060 	int has_net_offloads;
5061 	__be64 *rq_pas0;
5062 	int ts_format;
5063 	void *in;
5064 	void *rqc;
5065 	void *wq;
5066 	int inlen;
5067 	int err;
5068 
5069 	dev = to_mdev(pd->device);
5070 
5071 	ts_format = get_rq_ts_format(dev, to_mcq(init_attr->cq));
5072 	if (ts_format < 0)
5073 		return ts_format;
5074 
5075 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
5076 	in = kvzalloc(inlen, GFP_KERNEL);
5077 	if (!in)
5078 		return -ENOMEM;
5079 
5080 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
5081 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
5082 	MLX5_SET(rqc,  rqc, mem_rq_type,
5083 		 MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
5084 	MLX5_SET(rqc, rqc, ts_format, ts_format);
5085 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
5086 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
5087 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
5088 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
5089 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
5090 	MLX5_SET(wq, wq, wq_type,
5091 		 rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ ?
5092 		 MLX5_WQ_TYPE_CYCLIC_STRIDING_RQ : MLX5_WQ_TYPE_CYCLIC);
5093 	if (init_attr->create_flags & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
5094 		if (!MLX5_CAP_GEN(dev->mdev, end_pad)) {
5095 			mlx5_ib_dbg(dev, "Scatter end padding is not supported\n");
5096 			err = -EOPNOTSUPP;
5097 			goto out;
5098 		} else {
5099 			MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
5100 		}
5101 	}
5102 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
5103 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ) {
5104 		/*
5105 		 * In Firmware number of strides in each WQE is:
5106 		 *   "512 * 2^single_wqe_log_num_of_strides"
5107 		 * Values 3 to 8 are accepted as 10 to 15, 9 to 18 are
5108 		 * accepted as 0 to 9
5109 		 */
5110 		static const u8 fw_map[] = { 10, 11, 12, 13, 14, 15, 0, 1,
5111 					     2,  3,  4,  5,  6,  7,  8, 9 };
5112 		MLX5_SET(wq, wq, two_byte_shift_en, rwq->two_byte_shift_en);
5113 		MLX5_SET(wq, wq, log_wqe_stride_size,
5114 			 rwq->single_stride_log_num_of_bytes -
5115 			 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES);
5116 		MLX5_SET(wq, wq, log_wqe_num_of_strides,
5117 			 fw_map[rwq->log_num_strides -
5118 				MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES]);
5119 	}
5120 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
5121 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
5122 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
5123 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
5124 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
5125 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
5126 	has_net_offloads = MLX5_CAP_GEN(dev->mdev, eth_net_offloads);
5127 	if (init_attr->create_flags & IB_WQ_FLAGS_CVLAN_STRIPPING) {
5128 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
5129 			mlx5_ib_dbg(dev, "VLAN offloads are not supported\n");
5130 			err = -EOPNOTSUPP;
5131 			goto out;
5132 		}
5133 	} else {
5134 		MLX5_SET(rqc, rqc, vsd, 1);
5135 	}
5136 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS) {
5137 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, scatter_fcs))) {
5138 			mlx5_ib_dbg(dev, "Scatter FCS is not supported\n");
5139 			err = -EOPNOTSUPP;
5140 			goto out;
5141 		}
5142 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
5143 	}
5144 	if (init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
5145 		if (!(dev->ib_dev.attrs.raw_packet_caps &
5146 		      IB_RAW_PACKET_CAP_DELAY_DROP)) {
5147 			mlx5_ib_dbg(dev, "Delay drop is not supported\n");
5148 			err = -EOPNOTSUPP;
5149 			goto out;
5150 		}
5151 		MLX5_SET(rqc, rqc, delay_drop_en, 1);
5152 	}
5153 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
5154 	mlx5_ib_populate_pas(rwq->umem, 1UL << rwq->page_shift, rq_pas0, 0);
5155 	err = mlx5_core_create_rq_tracked(dev, in, inlen, &rwq->core_qp);
5156 	if (!err && init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
5157 		err = set_delay_drop(dev);
5158 		if (err) {
5159 			mlx5_ib_warn(dev, "Failed to enable delay drop err=%d\n",
5160 				     err);
5161 			mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5162 		} else {
5163 			rwq->create_flags |= MLX5_IB_WQ_FLAGS_DELAY_DROP;
5164 		}
5165 	}
5166 out:
5167 	kvfree(in);
5168 	return err;
5169 }
5170 
5171 static int set_user_rq_size(struct mlx5_ib_dev *dev,
5172 			    struct ib_wq_init_attr *wq_init_attr,
5173 			    struct mlx5_ib_create_wq *ucmd,
5174 			    struct mlx5_ib_rwq *rwq)
5175 {
5176 	/* Sanity check RQ size before proceeding */
5177 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
5178 		return -EINVAL;
5179 
5180 	if (!ucmd->rq_wqe_count)
5181 		return -EINVAL;
5182 
5183 	rwq->wqe_count = ucmd->rq_wqe_count;
5184 	rwq->wqe_shift = ucmd->rq_wqe_shift;
5185 	if (check_shl_overflow(rwq->wqe_count, rwq->wqe_shift, &rwq->buf_size))
5186 		return -EINVAL;
5187 
5188 	rwq->log_rq_stride = rwq->wqe_shift;
5189 	rwq->log_rq_size = ilog2(rwq->wqe_count);
5190 	return 0;
5191 }
5192 
5193 static bool log_of_strides_valid(struct mlx5_ib_dev *dev, u32 log_num_strides)
5194 {
5195 	if ((log_num_strides > MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES) ||
5196 	    (log_num_strides < MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
5197 		return false;
5198 
5199 	if (!MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) &&
5200 	    (log_num_strides < MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
5201 		return false;
5202 
5203 	return true;
5204 }
5205 
5206 static int prepare_user_rq(struct ib_pd *pd,
5207 			   struct ib_wq_init_attr *init_attr,
5208 			   struct ib_udata *udata,
5209 			   struct mlx5_ib_rwq *rwq)
5210 {
5211 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
5212 	struct mlx5_ib_create_wq ucmd = {};
5213 	int err;
5214 	size_t required_cmd_sz;
5215 
5216 	required_cmd_sz = offsetofend(struct mlx5_ib_create_wq,
5217 				      single_stride_log_num_of_bytes);
5218 	if (udata->inlen < required_cmd_sz) {
5219 		mlx5_ib_dbg(dev, "invalid inlen\n");
5220 		return -EINVAL;
5221 	}
5222 
5223 	if (udata->inlen > sizeof(ucmd) &&
5224 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
5225 				 udata->inlen - sizeof(ucmd))) {
5226 		mlx5_ib_dbg(dev, "inlen is not supported\n");
5227 		return -EOPNOTSUPP;
5228 	}
5229 
5230 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
5231 		mlx5_ib_dbg(dev, "copy failed\n");
5232 		return -EFAULT;
5233 	}
5234 
5235 	if (ucmd.comp_mask & (~MLX5_IB_CREATE_WQ_STRIDING_RQ)) {
5236 		mlx5_ib_dbg(dev, "invalid comp mask\n");
5237 		return -EOPNOTSUPP;
5238 	} else if (ucmd.comp_mask & MLX5_IB_CREATE_WQ_STRIDING_RQ) {
5239 		if (!MLX5_CAP_GEN(dev->mdev, striding_rq)) {
5240 			mlx5_ib_dbg(dev, "Striding RQ is not supported\n");
5241 			return -EOPNOTSUPP;
5242 		}
5243 		if ((ucmd.single_stride_log_num_of_bytes <
5244 		    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES) ||
5245 		    (ucmd.single_stride_log_num_of_bytes >
5246 		     MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES)) {
5247 			mlx5_ib_dbg(dev, "Invalid log stride size (%u. Range is %u - %u)\n",
5248 				    ucmd.single_stride_log_num_of_bytes,
5249 				    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES,
5250 				    MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES);
5251 			return -EINVAL;
5252 		}
5253 		if (!log_of_strides_valid(dev,
5254 					  ucmd.single_wqe_log_num_of_strides)) {
5255 			mlx5_ib_dbg(
5256 				dev,
5257 				"Invalid log num strides (%u. Range is %u - %u)\n",
5258 				ucmd.single_wqe_log_num_of_strides,
5259 				MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) ?
5260 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES :
5261 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES,
5262 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES);
5263 			return -EINVAL;
5264 		}
5265 		rwq->single_stride_log_num_of_bytes =
5266 			ucmd.single_stride_log_num_of_bytes;
5267 		rwq->log_num_strides = ucmd.single_wqe_log_num_of_strides;
5268 		rwq->two_byte_shift_en = !!ucmd.two_byte_shift_en;
5269 		rwq->create_flags |= MLX5_IB_WQ_FLAGS_STRIDING_RQ;
5270 	}
5271 
5272 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
5273 	if (err) {
5274 		mlx5_ib_dbg(dev, "err %d\n", err);
5275 		return err;
5276 	}
5277 
5278 	err = create_user_rq(dev, pd, udata, rwq, &ucmd);
5279 	if (err) {
5280 		mlx5_ib_dbg(dev, "err %d\n", err);
5281 		return err;
5282 	}
5283 
5284 	rwq->user_index = ucmd.user_index;
5285 	return 0;
5286 }
5287 
5288 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
5289 				struct ib_wq_init_attr *init_attr,
5290 				struct ib_udata *udata)
5291 {
5292 	struct mlx5_ib_dev *dev;
5293 	struct mlx5_ib_rwq *rwq;
5294 	struct mlx5_ib_create_wq_resp resp = {};
5295 	size_t min_resp_len;
5296 	int err;
5297 
5298 	if (!udata)
5299 		return ERR_PTR(-ENOSYS);
5300 
5301 	min_resp_len = offsetofend(struct mlx5_ib_create_wq_resp, reserved);
5302 	if (udata->outlen && udata->outlen < min_resp_len)
5303 		return ERR_PTR(-EINVAL);
5304 
5305 	if (!capable(CAP_SYS_RAWIO) &&
5306 	    init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP)
5307 		return ERR_PTR(-EPERM);
5308 
5309 	dev = to_mdev(pd->device);
5310 	switch (init_attr->wq_type) {
5311 	case IB_WQT_RQ:
5312 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
5313 		if (!rwq)
5314 			return ERR_PTR(-ENOMEM);
5315 		err = prepare_user_rq(pd, init_attr, udata, rwq);
5316 		if (err)
5317 			goto err;
5318 		err = create_rq(rwq, pd, init_attr);
5319 		if (err)
5320 			goto err_user_rq;
5321 		break;
5322 	default:
5323 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
5324 			    init_attr->wq_type);
5325 		return ERR_PTR(-EINVAL);
5326 	}
5327 
5328 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
5329 	rwq->ibwq.state = IB_WQS_RESET;
5330 	if (udata->outlen) {
5331 		resp.response_length = offsetofend(
5332 			struct mlx5_ib_create_wq_resp, response_length);
5333 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
5334 		if (err)
5335 			goto err_copy;
5336 	}
5337 
5338 	rwq->core_qp.event = mlx5_ib_wq_event;
5339 	rwq->ibwq.event_handler = init_attr->event_handler;
5340 	return &rwq->ibwq;
5341 
5342 err_copy:
5343 	mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5344 err_user_rq:
5345 	destroy_user_rq(dev, pd, rwq, udata);
5346 err:
5347 	kfree(rwq);
5348 	return ERR_PTR(err);
5349 }
5350 
5351 int mlx5_ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
5352 {
5353 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
5354 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
5355 	int ret;
5356 
5357 	ret = mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5358 	if (ret)
5359 		return ret;
5360 	destroy_user_rq(dev, wq->pd, rwq, udata);
5361 	kfree(rwq);
5362 	return 0;
5363 }
5364 
5365 int mlx5_ib_create_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_table,
5366 				 struct ib_rwq_ind_table_init_attr *init_attr,
5367 				 struct ib_udata *udata)
5368 {
5369 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl =
5370 		to_mrwq_ind_table(ib_rwq_ind_table);
5371 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_table->device);
5372 	int sz = 1 << init_attr->log_ind_tbl_size;
5373 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
5374 	size_t min_resp_len;
5375 	int inlen;
5376 	int err;
5377 	int i;
5378 	u32 *in;
5379 	void *rqtc;
5380 
5381 	if (udata->inlen > 0 &&
5382 	    !ib_is_udata_cleared(udata, 0,
5383 				 udata->inlen))
5384 		return -EOPNOTSUPP;
5385 
5386 	if (init_attr->log_ind_tbl_size >
5387 	    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
5388 		mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
5389 			    init_attr->log_ind_tbl_size,
5390 			    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
5391 		return -EINVAL;
5392 	}
5393 
5394 	min_resp_len =
5395 		offsetofend(struct mlx5_ib_create_rwq_ind_tbl_resp, reserved);
5396 	if (udata->outlen && udata->outlen < min_resp_len)
5397 		return -EINVAL;
5398 
5399 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
5400 	in = kvzalloc(inlen, GFP_KERNEL);
5401 	if (!in)
5402 		return -ENOMEM;
5403 
5404 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
5405 
5406 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
5407 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
5408 
5409 	for (i = 0; i < sz; i++)
5410 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
5411 
5412 	rwq_ind_tbl->uid = to_mpd(init_attr->ind_tbl[0]->pd)->uid;
5413 	MLX5_SET(create_rqt_in, in, uid, rwq_ind_tbl->uid);
5414 
5415 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
5416 	kvfree(in);
5417 	if (err)
5418 		return err;
5419 
5420 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
5421 	if (udata->outlen) {
5422 		resp.response_length =
5423 			offsetofend(struct mlx5_ib_create_rwq_ind_tbl_resp,
5424 				    response_length);
5425 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
5426 		if (err)
5427 			goto err_copy;
5428 	}
5429 
5430 	return 0;
5431 
5432 err_copy:
5433 	mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
5434 	return err;
5435 }
5436 
5437 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
5438 {
5439 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
5440 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
5441 
5442 	return mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
5443 }
5444 
5445 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
5446 		      u32 wq_attr_mask, struct ib_udata *udata)
5447 {
5448 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
5449 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
5450 	struct mlx5_ib_modify_wq ucmd = {};
5451 	size_t required_cmd_sz;
5452 	int curr_wq_state;
5453 	int wq_state;
5454 	int inlen;
5455 	int err;
5456 	void *rqc;
5457 	void *in;
5458 
5459 	required_cmd_sz = offsetofend(struct mlx5_ib_modify_wq, reserved);
5460 	if (udata->inlen < required_cmd_sz)
5461 		return -EINVAL;
5462 
5463 	if (udata->inlen > sizeof(ucmd) &&
5464 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
5465 				 udata->inlen - sizeof(ucmd)))
5466 		return -EOPNOTSUPP;
5467 
5468 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
5469 		return -EFAULT;
5470 
5471 	if (ucmd.comp_mask || ucmd.reserved)
5472 		return -EOPNOTSUPP;
5473 
5474 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
5475 	in = kvzalloc(inlen, GFP_KERNEL);
5476 	if (!in)
5477 		return -ENOMEM;
5478 
5479 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
5480 
5481 	curr_wq_state = wq_attr->curr_wq_state;
5482 	wq_state = wq_attr->wq_state;
5483 	if (curr_wq_state == IB_WQS_ERR)
5484 		curr_wq_state = MLX5_RQC_STATE_ERR;
5485 	if (wq_state == IB_WQS_ERR)
5486 		wq_state = MLX5_RQC_STATE_ERR;
5487 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
5488 	MLX5_SET(modify_rq_in, in, uid, to_mpd(wq->pd)->uid);
5489 	MLX5_SET(rqc, rqc, state, wq_state);
5490 
5491 	if (wq_attr_mask & IB_WQ_FLAGS) {
5492 		if (wq_attr->flags_mask & IB_WQ_FLAGS_CVLAN_STRIPPING) {
5493 			if (!(MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
5494 			      MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
5495 				mlx5_ib_dbg(dev, "VLAN offloads are not "
5496 					    "supported\n");
5497 				err = -EOPNOTSUPP;
5498 				goto out;
5499 			}
5500 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
5501 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
5502 			MLX5_SET(rqc, rqc, vsd,
5503 				 (wq_attr->flags & IB_WQ_FLAGS_CVLAN_STRIPPING) ? 0 : 1);
5504 		}
5505 
5506 		if (wq_attr->flags_mask & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
5507 			mlx5_ib_dbg(dev, "Modifying scatter end padding is not supported\n");
5508 			err = -EOPNOTSUPP;
5509 			goto out;
5510 		}
5511 	}
5512 
5513 	if (curr_wq_state == IB_WQS_RESET && wq_state == IB_WQS_RDY) {
5514 		u16 set_id;
5515 
5516 		set_id = mlx5_ib_get_counters_id(dev, 0);
5517 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
5518 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
5519 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
5520 			MLX5_SET(rqc, rqc, counter_set_id, set_id);
5521 		} else
5522 			dev_info_once(
5523 				&dev->ib_dev.dev,
5524 				"Receive WQ counters are not supported on current FW\n");
5525 	}
5526 
5527 	err = mlx5_core_modify_rq(dev->mdev, rwq->core_qp.qpn, in);
5528 	if (!err)
5529 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
5530 
5531 out:
5532 	kvfree(in);
5533 	return err;
5534 }
5535 
5536 struct mlx5_ib_drain_cqe {
5537 	struct ib_cqe cqe;
5538 	struct completion done;
5539 };
5540 
5541 static void mlx5_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
5542 {
5543 	struct mlx5_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
5544 						     struct mlx5_ib_drain_cqe,
5545 						     cqe);
5546 
5547 	complete(&cqe->done);
5548 }
5549 
5550 /* This function returns only once the drained WR was completed */
5551 static void handle_drain_completion(struct ib_cq *cq,
5552 				    struct mlx5_ib_drain_cqe *sdrain,
5553 				    struct mlx5_ib_dev *dev)
5554 {
5555 	struct mlx5_core_dev *mdev = dev->mdev;
5556 
5557 	if (cq->poll_ctx == IB_POLL_DIRECT) {
5558 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
5559 			ib_process_cq_direct(cq, -1);
5560 		return;
5561 	}
5562 
5563 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5564 		struct mlx5_ib_cq *mcq = to_mcq(cq);
5565 		bool triggered = false;
5566 		unsigned long flags;
5567 
5568 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
5569 		/* Make sure that the CQ handler won't run if wasn't run yet */
5570 		if (!mcq->mcq.reset_notify_added)
5571 			mcq->mcq.reset_notify_added = 1;
5572 		else
5573 			triggered = true;
5574 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
5575 
5576 		if (triggered) {
5577 			/* Wait for any scheduled/running task to be ended */
5578 			switch (cq->poll_ctx) {
5579 			case IB_POLL_SOFTIRQ:
5580 				irq_poll_disable(&cq->iop);
5581 				irq_poll_enable(&cq->iop);
5582 				break;
5583 			case IB_POLL_WORKQUEUE:
5584 				cancel_work_sync(&cq->work);
5585 				break;
5586 			default:
5587 				WARN_ON_ONCE(1);
5588 			}
5589 		}
5590 
5591 		/* Run the CQ handler - this makes sure that the drain WR will
5592 		 * be processed if wasn't processed yet.
5593 		 */
5594 		mcq->mcq.comp(&mcq->mcq, NULL);
5595 	}
5596 
5597 	wait_for_completion(&sdrain->done);
5598 }
5599 
5600 void mlx5_ib_drain_sq(struct ib_qp *qp)
5601 {
5602 	struct ib_cq *cq = qp->send_cq;
5603 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
5604 	struct mlx5_ib_drain_cqe sdrain;
5605 	const struct ib_send_wr *bad_swr;
5606 	struct ib_rdma_wr swr = {
5607 		.wr = {
5608 			.next = NULL,
5609 			{ .wr_cqe	= &sdrain.cqe, },
5610 			.opcode	= IB_WR_RDMA_WRITE,
5611 		},
5612 	};
5613 	int ret;
5614 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5615 	struct mlx5_core_dev *mdev = dev->mdev;
5616 
5617 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
5618 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5619 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
5620 		return;
5621 	}
5622 
5623 	sdrain.cqe.done = mlx5_ib_drain_qp_done;
5624 	init_completion(&sdrain.done);
5625 
5626 	ret = mlx5_ib_post_send_drain(qp, &swr.wr, &bad_swr);
5627 	if (ret) {
5628 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
5629 		return;
5630 	}
5631 
5632 	handle_drain_completion(cq, &sdrain, dev);
5633 }
5634 
5635 void mlx5_ib_drain_rq(struct ib_qp *qp)
5636 {
5637 	struct ib_cq *cq = qp->recv_cq;
5638 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
5639 	struct mlx5_ib_drain_cqe rdrain;
5640 	struct ib_recv_wr rwr = {};
5641 	const struct ib_recv_wr *bad_rwr;
5642 	int ret;
5643 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5644 	struct mlx5_core_dev *mdev = dev->mdev;
5645 
5646 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
5647 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5648 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
5649 		return;
5650 	}
5651 
5652 	rwr.wr_cqe = &rdrain.cqe;
5653 	rdrain.cqe.done = mlx5_ib_drain_qp_done;
5654 	init_completion(&rdrain.done);
5655 
5656 	ret = mlx5_ib_post_recv_drain(qp, &rwr, &bad_rwr);
5657 	if (ret) {
5658 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
5659 		return;
5660 	}
5661 
5662 	handle_drain_completion(cq, &rdrain, dev);
5663 }
5664 
5665 /*
5666  * Bind a qp to a counter. If @counter is NULL then bind the qp to
5667  * the default counter
5668  */
5669 int mlx5_ib_qp_set_counter(struct ib_qp *qp, struct rdma_counter *counter)
5670 {
5671 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5672 	struct mlx5_ib_qp *mqp = to_mqp(qp);
5673 	int err = 0;
5674 
5675 	mutex_lock(&mqp->mutex);
5676 	if (mqp->state == IB_QPS_RESET) {
5677 		qp->counter = counter;
5678 		goto out;
5679 	}
5680 
5681 	if (!MLX5_CAP_GEN(dev->mdev, rts2rts_qp_counters_set_id)) {
5682 		err = -EOPNOTSUPP;
5683 		goto out;
5684 	}
5685 
5686 	if (mqp->state == IB_QPS_RTS) {
5687 		err = __mlx5_ib_qp_set_counter(qp, counter);
5688 		if (!err)
5689 			qp->counter = counter;
5690 
5691 		goto out;
5692 	}
5693 
5694 	mqp->counter_pending = 1;
5695 	qp->counter = counter;
5696 
5697 out:
5698 	mutex_unlock(&mqp->mutex);
5699 	return err;
5700 }
5701