xref: /openbmc/linux/drivers/infiniband/hw/mlx5/qp.c (revision 11a163f2)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/module.h>
34 #include <rdma/ib_umem.h>
35 #include <rdma/ib_cache.h>
36 #include <rdma/ib_user_verbs.h>
37 #include <rdma/rdma_counter.h>
38 #include <linux/mlx5/fs.h>
39 #include "mlx5_ib.h"
40 #include "ib_rep.h"
41 #include "counters.h"
42 #include "cmd.h"
43 #include "qp.h"
44 #include "wr.h"
45 
46 enum {
47 	MLX5_IB_ACK_REQ_FREQ	= 8,
48 };
49 
50 enum {
51 	MLX5_IB_DEFAULT_SCHED_QUEUE	= 0x83,
52 	MLX5_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
53 	MLX5_IB_LINK_TYPE_IB		= 0,
54 	MLX5_IB_LINK_TYPE_ETH		= 1
55 };
56 
57 enum raw_qp_set_mask_map {
58 	MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID		= 1UL << 0,
59 	MLX5_RAW_QP_RATE_LIMIT			= 1UL << 1,
60 };
61 
62 struct mlx5_modify_raw_qp_param {
63 	u16 operation;
64 
65 	u32 set_mask; /* raw_qp_set_mask_map */
66 
67 	struct mlx5_rate_limit rl;
68 
69 	u8 rq_q_ctr_id;
70 	u16 port;
71 };
72 
73 static void get_cqs(enum ib_qp_type qp_type,
74 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
75 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
76 
77 static int is_qp0(enum ib_qp_type qp_type)
78 {
79 	return qp_type == IB_QPT_SMI;
80 }
81 
82 static int is_sqp(enum ib_qp_type qp_type)
83 {
84 	return is_qp0(qp_type) || is_qp1(qp_type);
85 }
86 
87 /**
88  * mlx5_ib_read_user_wqe_common() - Copy a WQE (or part of) from user WQ
89  * to kernel buffer
90  *
91  * @umem: User space memory where the WQ is
92  * @buffer: buffer to copy to
93  * @buflen: buffer length
94  * @wqe_index: index of WQE to copy from
95  * @wq_offset: offset to start of WQ
96  * @wq_wqe_cnt: number of WQEs in WQ
97  * @wq_wqe_shift: log2 of WQE size
98  * @bcnt: number of bytes to copy
99  * @bytes_copied: number of bytes to copy (return value)
100  *
101  * Copies from start of WQE bcnt or less bytes.
102  * Does not gurantee to copy the entire WQE.
103  *
104  * Return: zero on success, or an error code.
105  */
106 static int mlx5_ib_read_user_wqe_common(struct ib_umem *umem, void *buffer,
107 					size_t buflen, int wqe_index,
108 					int wq_offset, int wq_wqe_cnt,
109 					int wq_wqe_shift, int bcnt,
110 					size_t *bytes_copied)
111 {
112 	size_t offset = wq_offset + ((wqe_index % wq_wqe_cnt) << wq_wqe_shift);
113 	size_t wq_end = wq_offset + (wq_wqe_cnt << wq_wqe_shift);
114 	size_t copy_length;
115 	int ret;
116 
117 	/* don't copy more than requested, more than buffer length or
118 	 * beyond WQ end
119 	 */
120 	copy_length = min_t(u32, buflen, wq_end - offset);
121 	copy_length = min_t(u32, copy_length, bcnt);
122 
123 	ret = ib_umem_copy_from(buffer, umem, offset, copy_length);
124 	if (ret)
125 		return ret;
126 
127 	if (!ret && bytes_copied)
128 		*bytes_copied = copy_length;
129 
130 	return 0;
131 }
132 
133 static int mlx5_ib_read_kernel_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index,
134 				      void *buffer, size_t buflen, size_t *bc)
135 {
136 	struct mlx5_wqe_ctrl_seg *ctrl;
137 	size_t bytes_copied = 0;
138 	size_t wqe_length;
139 	void *p;
140 	int ds;
141 
142 	wqe_index = wqe_index & qp->sq.fbc.sz_m1;
143 
144 	/* read the control segment first */
145 	p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, wqe_index);
146 	ctrl = p;
147 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
148 	wqe_length = ds * MLX5_WQE_DS_UNITS;
149 
150 	/* read rest of WQE if it spreads over more than one stride */
151 	while (bytes_copied < wqe_length) {
152 		size_t copy_length =
153 			min_t(size_t, buflen - bytes_copied, MLX5_SEND_WQE_BB);
154 
155 		if (!copy_length)
156 			break;
157 
158 		memcpy(buffer + bytes_copied, p, copy_length);
159 		bytes_copied += copy_length;
160 
161 		wqe_index = (wqe_index + 1) & qp->sq.fbc.sz_m1;
162 		p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, wqe_index);
163 	}
164 	*bc = bytes_copied;
165 	return 0;
166 }
167 
168 static int mlx5_ib_read_user_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index,
169 				    void *buffer, size_t buflen, size_t *bc)
170 {
171 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
172 	struct ib_umem *umem = base->ubuffer.umem;
173 	struct mlx5_ib_wq *wq = &qp->sq;
174 	struct mlx5_wqe_ctrl_seg *ctrl;
175 	size_t bytes_copied;
176 	size_t bytes_copied2;
177 	size_t wqe_length;
178 	int ret;
179 	int ds;
180 
181 	/* at first read as much as possible */
182 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index,
183 					   wq->offset, wq->wqe_cnt,
184 					   wq->wqe_shift, buflen,
185 					   &bytes_copied);
186 	if (ret)
187 		return ret;
188 
189 	/* we need at least control segment size to proceed */
190 	if (bytes_copied < sizeof(*ctrl))
191 		return -EINVAL;
192 
193 	ctrl = buffer;
194 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
195 	wqe_length = ds * MLX5_WQE_DS_UNITS;
196 
197 	/* if we copied enough then we are done */
198 	if (bytes_copied >= wqe_length) {
199 		*bc = bytes_copied;
200 		return 0;
201 	}
202 
203 	/* otherwise this a wrapped around wqe
204 	 * so read the remaining bytes starting
205 	 * from  wqe_index 0
206 	 */
207 	ret = mlx5_ib_read_user_wqe_common(umem, buffer + bytes_copied,
208 					   buflen - bytes_copied, 0, wq->offset,
209 					   wq->wqe_cnt, wq->wqe_shift,
210 					   wqe_length - bytes_copied,
211 					   &bytes_copied2);
212 
213 	if (ret)
214 		return ret;
215 	*bc = bytes_copied + bytes_copied2;
216 	return 0;
217 }
218 
219 int mlx5_ib_read_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
220 			size_t buflen, size_t *bc)
221 {
222 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
223 	struct ib_umem *umem = base->ubuffer.umem;
224 
225 	if (buflen < sizeof(struct mlx5_wqe_ctrl_seg))
226 		return -EINVAL;
227 
228 	if (!umem)
229 		return mlx5_ib_read_kernel_wqe_sq(qp, wqe_index, buffer,
230 						  buflen, bc);
231 
232 	return mlx5_ib_read_user_wqe_sq(qp, wqe_index, buffer, buflen, bc);
233 }
234 
235 static int mlx5_ib_read_user_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index,
236 				    void *buffer, size_t buflen, size_t *bc)
237 {
238 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
239 	struct ib_umem *umem = base->ubuffer.umem;
240 	struct mlx5_ib_wq *wq = &qp->rq;
241 	size_t bytes_copied;
242 	int ret;
243 
244 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index,
245 					   wq->offset, wq->wqe_cnt,
246 					   wq->wqe_shift, buflen,
247 					   &bytes_copied);
248 
249 	if (ret)
250 		return ret;
251 	*bc = bytes_copied;
252 	return 0;
253 }
254 
255 int mlx5_ib_read_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
256 			size_t buflen, size_t *bc)
257 {
258 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
259 	struct ib_umem *umem = base->ubuffer.umem;
260 	struct mlx5_ib_wq *wq = &qp->rq;
261 	size_t wqe_size = 1 << wq->wqe_shift;
262 
263 	if (buflen < wqe_size)
264 		return -EINVAL;
265 
266 	if (!umem)
267 		return -EOPNOTSUPP;
268 
269 	return mlx5_ib_read_user_wqe_rq(qp, wqe_index, buffer, buflen, bc);
270 }
271 
272 static int mlx5_ib_read_user_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index,
273 				     void *buffer, size_t buflen, size_t *bc)
274 {
275 	struct ib_umem *umem = srq->umem;
276 	size_t bytes_copied;
277 	int ret;
278 
279 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index, 0,
280 					   srq->msrq.max, srq->msrq.wqe_shift,
281 					   buflen, &bytes_copied);
282 
283 	if (ret)
284 		return ret;
285 	*bc = bytes_copied;
286 	return 0;
287 }
288 
289 int mlx5_ib_read_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index, void *buffer,
290 			 size_t buflen, size_t *bc)
291 {
292 	struct ib_umem *umem = srq->umem;
293 	size_t wqe_size = 1 << srq->msrq.wqe_shift;
294 
295 	if (buflen < wqe_size)
296 		return -EINVAL;
297 
298 	if (!umem)
299 		return -EOPNOTSUPP;
300 
301 	return mlx5_ib_read_user_wqe_srq(srq, wqe_index, buffer, buflen, bc);
302 }
303 
304 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
305 {
306 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
307 	struct ib_event event;
308 
309 	if (type == MLX5_EVENT_TYPE_PATH_MIG) {
310 		/* This event is only valid for trans_qps */
311 		to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
312 	}
313 
314 	if (ibqp->event_handler) {
315 		event.device     = ibqp->device;
316 		event.element.qp = ibqp;
317 		switch (type) {
318 		case MLX5_EVENT_TYPE_PATH_MIG:
319 			event.event = IB_EVENT_PATH_MIG;
320 			break;
321 		case MLX5_EVENT_TYPE_COMM_EST:
322 			event.event = IB_EVENT_COMM_EST;
323 			break;
324 		case MLX5_EVENT_TYPE_SQ_DRAINED:
325 			event.event = IB_EVENT_SQ_DRAINED;
326 			break;
327 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
328 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
329 			break;
330 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
331 			event.event = IB_EVENT_QP_FATAL;
332 			break;
333 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
334 			event.event = IB_EVENT_PATH_MIG_ERR;
335 			break;
336 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
337 			event.event = IB_EVENT_QP_REQ_ERR;
338 			break;
339 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
340 			event.event = IB_EVENT_QP_ACCESS_ERR;
341 			break;
342 		default:
343 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
344 			return;
345 		}
346 
347 		ibqp->event_handler(&event, ibqp->qp_context);
348 	}
349 }
350 
351 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
352 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
353 {
354 	int wqe_size;
355 	int wq_size;
356 
357 	/* Sanity check RQ size before proceeding */
358 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
359 		return -EINVAL;
360 
361 	if (!has_rq) {
362 		qp->rq.max_gs = 0;
363 		qp->rq.wqe_cnt = 0;
364 		qp->rq.wqe_shift = 0;
365 		cap->max_recv_wr = 0;
366 		cap->max_recv_sge = 0;
367 	} else {
368 		int wq_sig = !!(qp->flags_en & MLX5_QP_FLAG_SIGNATURE);
369 
370 		if (ucmd) {
371 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
372 			if (ucmd->rq_wqe_shift > BITS_PER_BYTE * sizeof(ucmd->rq_wqe_shift))
373 				return -EINVAL;
374 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
375 			if ((1 << qp->rq.wqe_shift) /
376 				    sizeof(struct mlx5_wqe_data_seg) <
377 			    wq_sig)
378 				return -EINVAL;
379 			qp->rq.max_gs =
380 				(1 << qp->rq.wqe_shift) /
381 					sizeof(struct mlx5_wqe_data_seg) -
382 				wq_sig;
383 			qp->rq.max_post = qp->rq.wqe_cnt;
384 		} else {
385 			wqe_size =
386 				wq_sig ? sizeof(struct mlx5_wqe_signature_seg) :
387 					 0;
388 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
389 			wqe_size = roundup_pow_of_two(wqe_size);
390 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
391 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
392 			qp->rq.wqe_cnt = wq_size / wqe_size;
393 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
394 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
395 					    wqe_size,
396 					    MLX5_CAP_GEN(dev->mdev,
397 							 max_wqe_sz_rq));
398 				return -EINVAL;
399 			}
400 			qp->rq.wqe_shift = ilog2(wqe_size);
401 			qp->rq.max_gs =
402 				(1 << qp->rq.wqe_shift) /
403 					sizeof(struct mlx5_wqe_data_seg) -
404 				wq_sig;
405 			qp->rq.max_post = qp->rq.wqe_cnt;
406 		}
407 	}
408 
409 	return 0;
410 }
411 
412 static int sq_overhead(struct ib_qp_init_attr *attr)
413 {
414 	int size = 0;
415 
416 	switch (attr->qp_type) {
417 	case IB_QPT_XRC_INI:
418 		size += sizeof(struct mlx5_wqe_xrc_seg);
419 		fallthrough;
420 	case IB_QPT_RC:
421 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
422 			max(sizeof(struct mlx5_wqe_atomic_seg) +
423 			    sizeof(struct mlx5_wqe_raddr_seg),
424 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
425 			    sizeof(struct mlx5_mkey_seg) +
426 			    MLX5_IB_SQ_UMR_INLINE_THRESHOLD /
427 			    MLX5_IB_UMR_OCTOWORD);
428 		break;
429 
430 	case IB_QPT_XRC_TGT:
431 		return 0;
432 
433 	case IB_QPT_UC:
434 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
435 			max(sizeof(struct mlx5_wqe_raddr_seg),
436 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
437 			    sizeof(struct mlx5_mkey_seg));
438 		break;
439 
440 	case IB_QPT_UD:
441 		if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
442 			size += sizeof(struct mlx5_wqe_eth_pad) +
443 				sizeof(struct mlx5_wqe_eth_seg);
444 		fallthrough;
445 	case IB_QPT_SMI:
446 	case MLX5_IB_QPT_HW_GSI:
447 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
448 			sizeof(struct mlx5_wqe_datagram_seg);
449 		break;
450 
451 	case MLX5_IB_QPT_REG_UMR:
452 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
453 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
454 			sizeof(struct mlx5_mkey_seg);
455 		break;
456 
457 	default:
458 		return -EINVAL;
459 	}
460 
461 	return size;
462 }
463 
464 static int calc_send_wqe(struct ib_qp_init_attr *attr)
465 {
466 	int inl_size = 0;
467 	int size;
468 
469 	size = sq_overhead(attr);
470 	if (size < 0)
471 		return size;
472 
473 	if (attr->cap.max_inline_data) {
474 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
475 			attr->cap.max_inline_data;
476 	}
477 
478 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
479 	if (attr->create_flags & IB_QP_CREATE_INTEGRITY_EN &&
480 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
481 		return MLX5_SIG_WQE_SIZE;
482 	else
483 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
484 }
485 
486 static int get_send_sge(struct ib_qp_init_attr *attr, int wqe_size)
487 {
488 	int max_sge;
489 
490 	if (attr->qp_type == IB_QPT_RC)
491 		max_sge = (min_t(int, wqe_size, 512) -
492 			   sizeof(struct mlx5_wqe_ctrl_seg) -
493 			   sizeof(struct mlx5_wqe_raddr_seg)) /
494 			sizeof(struct mlx5_wqe_data_seg);
495 	else if (attr->qp_type == IB_QPT_XRC_INI)
496 		max_sge = (min_t(int, wqe_size, 512) -
497 			   sizeof(struct mlx5_wqe_ctrl_seg) -
498 			   sizeof(struct mlx5_wqe_xrc_seg) -
499 			   sizeof(struct mlx5_wqe_raddr_seg)) /
500 			sizeof(struct mlx5_wqe_data_seg);
501 	else
502 		max_sge = (wqe_size - sq_overhead(attr)) /
503 			sizeof(struct mlx5_wqe_data_seg);
504 
505 	return min_t(int, max_sge, wqe_size - sq_overhead(attr) /
506 		     sizeof(struct mlx5_wqe_data_seg));
507 }
508 
509 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
510 			struct mlx5_ib_qp *qp)
511 {
512 	int wqe_size;
513 	int wq_size;
514 
515 	if (!attr->cap.max_send_wr)
516 		return 0;
517 
518 	wqe_size = calc_send_wqe(attr);
519 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
520 	if (wqe_size < 0)
521 		return wqe_size;
522 
523 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
524 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
525 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
526 		return -EINVAL;
527 	}
528 
529 	qp->max_inline_data = wqe_size - sq_overhead(attr) -
530 			      sizeof(struct mlx5_wqe_inline_seg);
531 	attr->cap.max_inline_data = qp->max_inline_data;
532 
533 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
534 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
535 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
536 		mlx5_ib_dbg(dev, "send queue size (%d * %d / %d -> %d) exceeds limits(%d)\n",
537 			    attr->cap.max_send_wr, wqe_size, MLX5_SEND_WQE_BB,
538 			    qp->sq.wqe_cnt,
539 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
540 		return -ENOMEM;
541 	}
542 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
543 	qp->sq.max_gs = get_send_sge(attr, wqe_size);
544 	if (qp->sq.max_gs < attr->cap.max_send_sge)
545 		return -ENOMEM;
546 
547 	attr->cap.max_send_sge = qp->sq.max_gs;
548 	qp->sq.max_post = wq_size / wqe_size;
549 	attr->cap.max_send_wr = qp->sq.max_post;
550 
551 	return wq_size;
552 }
553 
554 static int set_user_buf_size(struct mlx5_ib_dev *dev,
555 			    struct mlx5_ib_qp *qp,
556 			    struct mlx5_ib_create_qp *ucmd,
557 			    struct mlx5_ib_qp_base *base,
558 			    struct ib_qp_init_attr *attr)
559 {
560 	int desc_sz = 1 << qp->sq.wqe_shift;
561 
562 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
563 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
564 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
565 		return -EINVAL;
566 	}
567 
568 	if (ucmd->sq_wqe_count && !is_power_of_2(ucmd->sq_wqe_count)) {
569 		mlx5_ib_warn(dev, "sq_wqe_count %d is not a power of two\n",
570 			     ucmd->sq_wqe_count);
571 		return -EINVAL;
572 	}
573 
574 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
575 
576 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
577 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
578 			     qp->sq.wqe_cnt,
579 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
580 		return -EINVAL;
581 	}
582 
583 	if (attr->qp_type == IB_QPT_RAW_PACKET ||
584 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
585 		base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
586 		qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
587 	} else {
588 		base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
589 					 (qp->sq.wqe_cnt << 6);
590 	}
591 
592 	return 0;
593 }
594 
595 static int qp_has_rq(struct ib_qp_init_attr *attr)
596 {
597 	if (attr->qp_type == IB_QPT_XRC_INI ||
598 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
599 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
600 	    !attr->cap.max_recv_wr)
601 		return 0;
602 
603 	return 1;
604 }
605 
606 enum {
607 	/* this is the first blue flame register in the array of bfregs assigned
608 	 * to a processes. Since we do not use it for blue flame but rather
609 	 * regular 64 bit doorbells, we do not need a lock for maintaiing
610 	 * "odd/even" order
611 	 */
612 	NUM_NON_BLUE_FLAME_BFREGS = 1,
613 };
614 
615 static int max_bfregs(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi)
616 {
617 	return get_num_static_uars(dev, bfregi) * MLX5_NON_FP_BFREGS_PER_UAR;
618 }
619 
620 static int num_med_bfreg(struct mlx5_ib_dev *dev,
621 			 struct mlx5_bfreg_info *bfregi)
622 {
623 	int n;
624 
625 	n = max_bfregs(dev, bfregi) - bfregi->num_low_latency_bfregs -
626 	    NUM_NON_BLUE_FLAME_BFREGS;
627 
628 	return n >= 0 ? n : 0;
629 }
630 
631 static int first_med_bfreg(struct mlx5_ib_dev *dev,
632 			   struct mlx5_bfreg_info *bfregi)
633 {
634 	return num_med_bfreg(dev, bfregi) ? 1 : -ENOMEM;
635 }
636 
637 static int first_hi_bfreg(struct mlx5_ib_dev *dev,
638 			  struct mlx5_bfreg_info *bfregi)
639 {
640 	int med;
641 
642 	med = num_med_bfreg(dev, bfregi);
643 	return ++med;
644 }
645 
646 static int alloc_high_class_bfreg(struct mlx5_ib_dev *dev,
647 				  struct mlx5_bfreg_info *bfregi)
648 {
649 	int i;
650 
651 	for (i = first_hi_bfreg(dev, bfregi); i < max_bfregs(dev, bfregi); i++) {
652 		if (!bfregi->count[i]) {
653 			bfregi->count[i]++;
654 			return i;
655 		}
656 	}
657 
658 	return -ENOMEM;
659 }
660 
661 static int alloc_med_class_bfreg(struct mlx5_ib_dev *dev,
662 				 struct mlx5_bfreg_info *bfregi)
663 {
664 	int minidx = first_med_bfreg(dev, bfregi);
665 	int i;
666 
667 	if (minidx < 0)
668 		return minidx;
669 
670 	for (i = minidx; i < first_hi_bfreg(dev, bfregi); i++) {
671 		if (bfregi->count[i] < bfregi->count[minidx])
672 			minidx = i;
673 		if (!bfregi->count[minidx])
674 			break;
675 	}
676 
677 	bfregi->count[minidx]++;
678 	return minidx;
679 }
680 
681 static int alloc_bfreg(struct mlx5_ib_dev *dev,
682 		       struct mlx5_bfreg_info *bfregi)
683 {
684 	int bfregn = -ENOMEM;
685 
686 	if (bfregi->lib_uar_dyn)
687 		return -EINVAL;
688 
689 	mutex_lock(&bfregi->lock);
690 	if (bfregi->ver >= 2) {
691 		bfregn = alloc_high_class_bfreg(dev, bfregi);
692 		if (bfregn < 0)
693 			bfregn = alloc_med_class_bfreg(dev, bfregi);
694 	}
695 
696 	if (bfregn < 0) {
697 		BUILD_BUG_ON(NUM_NON_BLUE_FLAME_BFREGS != 1);
698 		bfregn = 0;
699 		bfregi->count[bfregn]++;
700 	}
701 	mutex_unlock(&bfregi->lock);
702 
703 	return bfregn;
704 }
705 
706 void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
707 {
708 	mutex_lock(&bfregi->lock);
709 	bfregi->count[bfregn]--;
710 	mutex_unlock(&bfregi->lock);
711 }
712 
713 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
714 {
715 	switch (state) {
716 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
717 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
718 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
719 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
720 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
721 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
722 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
723 	default:		return -1;
724 	}
725 }
726 
727 static int to_mlx5_st(enum ib_qp_type type)
728 {
729 	switch (type) {
730 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
731 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
732 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
733 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
734 	case IB_QPT_XRC_INI:
735 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
736 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
737 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
738 	case MLX5_IB_QPT_DCI:		return MLX5_QP_ST_DCI;
739 	case IB_QPT_RAW_PACKET:		return MLX5_QP_ST_RAW_ETHERTYPE;
740 	default:		return -EINVAL;
741 	}
742 }
743 
744 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
745 			     struct mlx5_ib_cq *recv_cq);
746 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
747 			       struct mlx5_ib_cq *recv_cq);
748 
749 int bfregn_to_uar_index(struct mlx5_ib_dev *dev,
750 			struct mlx5_bfreg_info *bfregi, u32 bfregn,
751 			bool dyn_bfreg)
752 {
753 	unsigned int bfregs_per_sys_page;
754 	u32 index_of_sys_page;
755 	u32 offset;
756 
757 	if (bfregi->lib_uar_dyn)
758 		return -EINVAL;
759 
760 	bfregs_per_sys_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
761 				MLX5_NON_FP_BFREGS_PER_UAR;
762 	index_of_sys_page = bfregn / bfregs_per_sys_page;
763 
764 	if (dyn_bfreg) {
765 		index_of_sys_page += bfregi->num_static_sys_pages;
766 
767 		if (index_of_sys_page >= bfregi->num_sys_pages)
768 			return -EINVAL;
769 
770 		if (bfregn > bfregi->num_dyn_bfregs ||
771 		    bfregi->sys_pages[index_of_sys_page] == MLX5_IB_INVALID_UAR_INDEX) {
772 			mlx5_ib_dbg(dev, "Invalid dynamic uar index\n");
773 			return -EINVAL;
774 		}
775 	}
776 
777 	offset = bfregn % bfregs_per_sys_page / MLX5_NON_FP_BFREGS_PER_UAR;
778 	return bfregi->sys_pages[index_of_sys_page] + offset;
779 }
780 
781 static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev, struct ib_udata *udata,
782 			    unsigned long addr, size_t size,
783 			    struct ib_umem **umem, int *npages, int *page_shift,
784 			    int *ncont, u32 *offset)
785 {
786 	int err;
787 
788 	*umem = ib_umem_get(&dev->ib_dev, addr, size, 0);
789 	if (IS_ERR(*umem)) {
790 		mlx5_ib_dbg(dev, "umem_get failed\n");
791 		return PTR_ERR(*umem);
792 	}
793 
794 	mlx5_ib_cont_pages(*umem, addr, 0, npages, page_shift, ncont, NULL);
795 
796 	err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
797 	if (err) {
798 		mlx5_ib_warn(dev, "bad offset\n");
799 		goto err_umem;
800 	}
801 
802 	mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
803 		    addr, size, *npages, *page_shift, *ncont, *offset);
804 
805 	return 0;
806 
807 err_umem:
808 	ib_umem_release(*umem);
809 	*umem = NULL;
810 
811 	return err;
812 }
813 
814 static void destroy_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
815 			    struct mlx5_ib_rwq *rwq, struct ib_udata *udata)
816 {
817 	struct mlx5_ib_ucontext *context =
818 		rdma_udata_to_drv_context(
819 			udata,
820 			struct mlx5_ib_ucontext,
821 			ibucontext);
822 
823 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_DELAY_DROP)
824 		atomic_dec(&dev->delay_drop.rqs_cnt);
825 
826 	mlx5_ib_db_unmap_user(context, &rwq->db);
827 	ib_umem_release(rwq->umem);
828 }
829 
830 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
831 			  struct ib_udata *udata, struct mlx5_ib_rwq *rwq,
832 			  struct mlx5_ib_create_wq *ucmd)
833 {
834 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
835 		udata, struct mlx5_ib_ucontext, ibucontext);
836 	int page_shift = 0;
837 	int npages;
838 	u32 offset = 0;
839 	int ncont = 0;
840 	int err;
841 
842 	if (!ucmd->buf_addr)
843 		return -EINVAL;
844 
845 	rwq->umem = ib_umem_get(&dev->ib_dev, ucmd->buf_addr, rwq->buf_size, 0);
846 	if (IS_ERR(rwq->umem)) {
847 		mlx5_ib_dbg(dev, "umem_get failed\n");
848 		err = PTR_ERR(rwq->umem);
849 		return err;
850 	}
851 
852 	mlx5_ib_cont_pages(rwq->umem, ucmd->buf_addr, 0, &npages, &page_shift,
853 			   &ncont, NULL);
854 	err = mlx5_ib_get_buf_offset(ucmd->buf_addr, page_shift,
855 				     &rwq->rq_page_offset);
856 	if (err) {
857 		mlx5_ib_warn(dev, "bad offset\n");
858 		goto err_umem;
859 	}
860 
861 	rwq->rq_num_pas = ncont;
862 	rwq->page_shift = page_shift;
863 	rwq->log_page_size =  page_shift - MLX5_ADAPTER_PAGE_SHIFT;
864 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
865 
866 	mlx5_ib_dbg(dev, "addr 0x%llx, size %zd, npages %d, page_shift %d, ncont %d, offset %d\n",
867 		    (unsigned long long)ucmd->buf_addr, rwq->buf_size,
868 		    npages, page_shift, ncont, offset);
869 
870 	err = mlx5_ib_db_map_user(ucontext, udata, ucmd->db_addr, &rwq->db);
871 	if (err) {
872 		mlx5_ib_dbg(dev, "map failed\n");
873 		goto err_umem;
874 	}
875 
876 	return 0;
877 
878 err_umem:
879 	ib_umem_release(rwq->umem);
880 	return err;
881 }
882 
883 static int adjust_bfregn(struct mlx5_ib_dev *dev,
884 			 struct mlx5_bfreg_info *bfregi, int bfregn)
885 {
886 	return bfregn / MLX5_NON_FP_BFREGS_PER_UAR * MLX5_BFREGS_PER_UAR +
887 				bfregn % MLX5_NON_FP_BFREGS_PER_UAR;
888 }
889 
890 static int _create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
891 			   struct mlx5_ib_qp *qp, struct ib_udata *udata,
892 			   struct ib_qp_init_attr *attr, u32 **in,
893 			   struct mlx5_ib_create_qp_resp *resp, int *inlen,
894 			   struct mlx5_ib_qp_base *base,
895 			   struct mlx5_ib_create_qp *ucmd)
896 {
897 	struct mlx5_ib_ucontext *context;
898 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
899 	int page_shift = 0;
900 	int uar_index = 0;
901 	int npages;
902 	u32 offset = 0;
903 	int bfregn;
904 	int ncont = 0;
905 	__be64 *pas;
906 	void *qpc;
907 	int err;
908 	u16 uid;
909 	u32 uar_flags;
910 
911 	context = rdma_udata_to_drv_context(udata, struct mlx5_ib_ucontext,
912 					    ibucontext);
913 	uar_flags = qp->flags_en &
914 		    (MLX5_QP_FLAG_UAR_PAGE_INDEX | MLX5_QP_FLAG_BFREG_INDEX);
915 	switch (uar_flags) {
916 	case MLX5_QP_FLAG_UAR_PAGE_INDEX:
917 		uar_index = ucmd->bfreg_index;
918 		bfregn = MLX5_IB_INVALID_BFREG;
919 		break;
920 	case MLX5_QP_FLAG_BFREG_INDEX:
921 		uar_index = bfregn_to_uar_index(dev, &context->bfregi,
922 						ucmd->bfreg_index, true);
923 		if (uar_index < 0)
924 			return uar_index;
925 		bfregn = MLX5_IB_INVALID_BFREG;
926 		break;
927 	case 0:
928 		if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
929 			return -EINVAL;
930 		bfregn = alloc_bfreg(dev, &context->bfregi);
931 		if (bfregn < 0)
932 			return bfregn;
933 		break;
934 	default:
935 		return -EINVAL;
936 	}
937 
938 	mlx5_ib_dbg(dev, "bfregn 0x%x, uar_index 0x%x\n", bfregn, uar_index);
939 	if (bfregn != MLX5_IB_INVALID_BFREG)
940 		uar_index = bfregn_to_uar_index(dev, &context->bfregi, bfregn,
941 						false);
942 
943 	qp->rq.offset = 0;
944 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
945 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
946 
947 	err = set_user_buf_size(dev, qp, ucmd, base, attr);
948 	if (err)
949 		goto err_bfreg;
950 
951 	if (ucmd->buf_addr && ubuffer->buf_size) {
952 		ubuffer->buf_addr = ucmd->buf_addr;
953 		err = mlx5_ib_umem_get(dev, udata, ubuffer->buf_addr,
954 				       ubuffer->buf_size, &ubuffer->umem,
955 				       &npages, &page_shift, &ncont, &offset);
956 		if (err)
957 			goto err_bfreg;
958 	} else {
959 		ubuffer->umem = NULL;
960 	}
961 
962 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
963 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
964 	*in = kvzalloc(*inlen, GFP_KERNEL);
965 	if (!*in) {
966 		err = -ENOMEM;
967 		goto err_umem;
968 	}
969 
970 	uid = (attr->qp_type != IB_QPT_XRC_INI) ? to_mpd(pd)->uid : 0;
971 	MLX5_SET(create_qp_in, *in, uid, uid);
972 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
973 	if (ubuffer->umem)
974 		mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift, pas, 0);
975 
976 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
977 
978 	MLX5_SET(qpc, qpc, log_page_size, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
979 	MLX5_SET(qpc, qpc, page_offset, offset);
980 
981 	MLX5_SET(qpc, qpc, uar_page, uar_index);
982 	if (bfregn != MLX5_IB_INVALID_BFREG)
983 		resp->bfreg_index = adjust_bfregn(dev, &context->bfregi, bfregn);
984 	else
985 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
986 	qp->bfregn = bfregn;
987 
988 	err = mlx5_ib_db_map_user(context, udata, ucmd->db_addr, &qp->db);
989 	if (err) {
990 		mlx5_ib_dbg(dev, "map failed\n");
991 		goto err_free;
992 	}
993 
994 	return 0;
995 
996 err_free:
997 	kvfree(*in);
998 
999 err_umem:
1000 	ib_umem_release(ubuffer->umem);
1001 
1002 err_bfreg:
1003 	if (bfregn != MLX5_IB_INVALID_BFREG)
1004 		mlx5_ib_free_bfreg(dev, &context->bfregi, bfregn);
1005 	return err;
1006 }
1007 
1008 static void destroy_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1009 		       struct mlx5_ib_qp_base *base, struct ib_udata *udata)
1010 {
1011 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
1012 		udata, struct mlx5_ib_ucontext, ibucontext);
1013 
1014 	if (udata) {
1015 		/* User QP */
1016 		mlx5_ib_db_unmap_user(context, &qp->db);
1017 		ib_umem_release(base->ubuffer.umem);
1018 
1019 		/*
1020 		 * Free only the BFREGs which are handled by the kernel.
1021 		 * BFREGs of UARs allocated dynamically are handled by user.
1022 		 */
1023 		if (qp->bfregn != MLX5_IB_INVALID_BFREG)
1024 			mlx5_ib_free_bfreg(dev, &context->bfregi, qp->bfregn);
1025 		return;
1026 	}
1027 
1028 	/* Kernel QP */
1029 	kvfree(qp->sq.wqe_head);
1030 	kvfree(qp->sq.w_list);
1031 	kvfree(qp->sq.wrid);
1032 	kvfree(qp->sq.wr_data);
1033 	kvfree(qp->rq.wrid);
1034 	if (qp->db.db)
1035 		mlx5_db_free(dev->mdev, &qp->db);
1036 	if (qp->buf.frags)
1037 		mlx5_frag_buf_free(dev->mdev, &qp->buf);
1038 }
1039 
1040 static int _create_kernel_qp(struct mlx5_ib_dev *dev,
1041 			     struct ib_qp_init_attr *init_attr,
1042 			     struct mlx5_ib_qp *qp, u32 **in, int *inlen,
1043 			     struct mlx5_ib_qp_base *base)
1044 {
1045 	int uar_index;
1046 	void *qpc;
1047 	int err;
1048 
1049 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
1050 		qp->bf.bfreg = &dev->fp_bfreg;
1051 	else if (qp->flags & MLX5_IB_QP_CREATE_WC_TEST)
1052 		qp->bf.bfreg = &dev->wc_bfreg;
1053 	else
1054 		qp->bf.bfreg = &dev->bfreg;
1055 
1056 	/* We need to divide by two since each register is comprised of
1057 	 * two buffers of identical size, namely odd and even
1058 	 */
1059 	qp->bf.buf_size = (1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size)) / 2;
1060 	uar_index = qp->bf.bfreg->index;
1061 
1062 	err = calc_sq_size(dev, init_attr, qp);
1063 	if (err < 0) {
1064 		mlx5_ib_dbg(dev, "err %d\n", err);
1065 		return err;
1066 	}
1067 
1068 	qp->rq.offset = 0;
1069 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
1070 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
1071 
1072 	err = mlx5_frag_buf_alloc_node(dev->mdev, base->ubuffer.buf_size,
1073 				       &qp->buf, dev->mdev->priv.numa_node);
1074 	if (err) {
1075 		mlx5_ib_dbg(dev, "err %d\n", err);
1076 		return err;
1077 	}
1078 
1079 	if (qp->rq.wqe_cnt)
1080 		mlx5_init_fbc(qp->buf.frags, qp->rq.wqe_shift,
1081 			      ilog2(qp->rq.wqe_cnt), &qp->rq.fbc);
1082 
1083 	if (qp->sq.wqe_cnt) {
1084 		int sq_strides_offset = (qp->sq.offset  & (PAGE_SIZE - 1)) /
1085 					MLX5_SEND_WQE_BB;
1086 		mlx5_init_fbc_offset(qp->buf.frags +
1087 				     (qp->sq.offset / PAGE_SIZE),
1088 				     ilog2(MLX5_SEND_WQE_BB),
1089 				     ilog2(qp->sq.wqe_cnt),
1090 				     sq_strides_offset, &qp->sq.fbc);
1091 
1092 		qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
1093 	}
1094 
1095 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
1096 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
1097 	*in = kvzalloc(*inlen, GFP_KERNEL);
1098 	if (!*in) {
1099 		err = -ENOMEM;
1100 		goto err_buf;
1101 	}
1102 
1103 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
1104 	MLX5_SET(qpc, qpc, uar_page, uar_index);
1105 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1106 
1107 	/* Set "fast registration enabled" for all kernel QPs */
1108 	MLX5_SET(qpc, qpc, fre, 1);
1109 	MLX5_SET(qpc, qpc, rlky, 1);
1110 
1111 	if (qp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
1112 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
1113 
1114 	mlx5_fill_page_frag_array(&qp->buf,
1115 				  (__be64 *)MLX5_ADDR_OF(create_qp_in,
1116 							 *in, pas));
1117 
1118 	err = mlx5_db_alloc(dev->mdev, &qp->db);
1119 	if (err) {
1120 		mlx5_ib_dbg(dev, "err %d\n", err);
1121 		goto err_free;
1122 	}
1123 
1124 	qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1125 				     sizeof(*qp->sq.wrid), GFP_KERNEL);
1126 	qp->sq.wr_data = kvmalloc_array(qp->sq.wqe_cnt,
1127 					sizeof(*qp->sq.wr_data), GFP_KERNEL);
1128 	qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1129 				     sizeof(*qp->rq.wrid), GFP_KERNEL);
1130 	qp->sq.w_list = kvmalloc_array(qp->sq.wqe_cnt,
1131 				       sizeof(*qp->sq.w_list), GFP_KERNEL);
1132 	qp->sq.wqe_head = kvmalloc_array(qp->sq.wqe_cnt,
1133 					 sizeof(*qp->sq.wqe_head), GFP_KERNEL);
1134 
1135 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
1136 	    !qp->sq.w_list || !qp->sq.wqe_head) {
1137 		err = -ENOMEM;
1138 		goto err_wrid;
1139 	}
1140 
1141 	return 0;
1142 
1143 err_wrid:
1144 	kvfree(qp->sq.wqe_head);
1145 	kvfree(qp->sq.w_list);
1146 	kvfree(qp->sq.wrid);
1147 	kvfree(qp->sq.wr_data);
1148 	kvfree(qp->rq.wrid);
1149 	mlx5_db_free(dev->mdev, &qp->db);
1150 
1151 err_free:
1152 	kvfree(*in);
1153 
1154 err_buf:
1155 	mlx5_frag_buf_free(dev->mdev, &qp->buf);
1156 	return err;
1157 }
1158 
1159 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
1160 {
1161 	if (attr->srq || (qp->type == IB_QPT_XRC_TGT) ||
1162 	    (qp->type == MLX5_IB_QPT_DCI) || (qp->type == IB_QPT_XRC_INI))
1163 		return MLX5_SRQ_RQ;
1164 	else if (!qp->has_rq)
1165 		return MLX5_ZERO_LEN_RQ;
1166 
1167 	return MLX5_NON_ZERO_RQ;
1168 }
1169 
1170 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1171 				    struct mlx5_ib_qp *qp,
1172 				    struct mlx5_ib_sq *sq, u32 tdn,
1173 				    struct ib_pd *pd)
1174 {
1175 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
1176 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1177 
1178 	MLX5_SET(create_tis_in, in, uid, to_mpd(pd)->uid);
1179 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1180 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
1181 		MLX5_SET(tisc, tisc, underlay_qpn, qp->underlay_qpn);
1182 
1183 	return mlx5_core_create_tis(dev->mdev, in, &sq->tisn);
1184 }
1185 
1186 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1187 				      struct mlx5_ib_sq *sq, struct ib_pd *pd)
1188 {
1189 	mlx5_cmd_destroy_tis(dev->mdev, sq->tisn, to_mpd(pd)->uid);
1190 }
1191 
1192 static void destroy_flow_rule_vport_sq(struct mlx5_ib_sq *sq)
1193 {
1194 	if (sq->flow_rule)
1195 		mlx5_del_flow_rules(sq->flow_rule);
1196 	sq->flow_rule = NULL;
1197 }
1198 
1199 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1200 				   struct ib_udata *udata,
1201 				   struct mlx5_ib_sq *sq, void *qpin,
1202 				   struct ib_pd *pd)
1203 {
1204 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1205 	__be64 *pas;
1206 	void *in;
1207 	void *sqc;
1208 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1209 	void *wq;
1210 	int inlen;
1211 	int err;
1212 	int page_shift = 0;
1213 	int npages;
1214 	int ncont = 0;
1215 	u32 offset = 0;
1216 
1217 	err = mlx5_ib_umem_get(dev, udata, ubuffer->buf_addr, ubuffer->buf_size,
1218 			       &sq->ubuffer.umem, &npages, &page_shift, &ncont,
1219 			       &offset);
1220 	if (err)
1221 		return err;
1222 
1223 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
1224 	in = kvzalloc(inlen, GFP_KERNEL);
1225 	if (!in) {
1226 		err = -ENOMEM;
1227 		goto err_umem;
1228 	}
1229 
1230 	MLX5_SET(create_sq_in, in, uid, to_mpd(pd)->uid);
1231 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1232 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1233 	if (MLX5_CAP_ETH(dev->mdev, multi_pkt_send_wqe))
1234 		MLX5_SET(sqc, sqc, allow_multi_pkt_send_wqe, 1);
1235 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1236 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1237 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1238 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1239 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1240 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
1241 	    MLX5_CAP_ETH(dev->mdev, swp))
1242 		MLX5_SET(sqc, sqc, allow_swp, 1);
1243 
1244 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1245 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1246 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1247 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1248 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1249 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1250 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1251 	MLX5_SET(wq, wq, log_wq_pg_sz,  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1252 	MLX5_SET(wq, wq, page_offset, offset);
1253 
1254 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1255 	mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
1256 
1257 	err = mlx5_core_create_sq_tracked(dev, in, inlen, &sq->base.mqp);
1258 
1259 	kvfree(in);
1260 
1261 	if (err)
1262 		goto err_umem;
1263 
1264 	return 0;
1265 
1266 err_umem:
1267 	ib_umem_release(sq->ubuffer.umem);
1268 	sq->ubuffer.umem = NULL;
1269 
1270 	return err;
1271 }
1272 
1273 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1274 				     struct mlx5_ib_sq *sq)
1275 {
1276 	destroy_flow_rule_vport_sq(sq);
1277 	mlx5_core_destroy_sq_tracked(dev, &sq->base.mqp);
1278 	ib_umem_release(sq->ubuffer.umem);
1279 }
1280 
1281 static size_t get_rq_pas_size(void *qpc)
1282 {
1283 	u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1284 	u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1285 	u32 log_rq_size   = MLX5_GET(qpc, qpc, log_rq_size);
1286 	u32 page_offset   = MLX5_GET(qpc, qpc, page_offset);
1287 	u32 po_quanta	  = 1 << (log_page_size - 6);
1288 	u32 rq_sz	  = 1 << (log_rq_size + 4 + log_rq_stride);
1289 	u32 page_size	  = 1 << log_page_size;
1290 	u32 rq_sz_po      = rq_sz + (page_offset * po_quanta);
1291 	u32 rq_num_pas	  = (rq_sz_po + page_size - 1) / page_size;
1292 
1293 	return rq_num_pas * sizeof(u64);
1294 }
1295 
1296 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1297 				   struct mlx5_ib_rq *rq, void *qpin,
1298 				   size_t qpinlen, struct ib_pd *pd)
1299 {
1300 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1301 	__be64 *pas;
1302 	__be64 *qp_pas;
1303 	void *in;
1304 	void *rqc;
1305 	void *wq;
1306 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1307 	size_t rq_pas_size = get_rq_pas_size(qpc);
1308 	size_t inlen;
1309 	int err;
1310 
1311 	if (qpinlen < rq_pas_size + MLX5_BYTE_OFF(create_qp_in, pas))
1312 		return -EINVAL;
1313 
1314 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1315 	in = kvzalloc(inlen, GFP_KERNEL);
1316 	if (!in)
1317 		return -ENOMEM;
1318 
1319 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
1320 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1321 	if (!(rq->flags & MLX5_IB_RQ_CVLAN_STRIPPING))
1322 		MLX5_SET(rqc, rqc, vsd, 1);
1323 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1324 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1325 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1326 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1327 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1328 
1329 	if (mqp->flags & IB_QP_CREATE_SCATTER_FCS)
1330 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1331 
1332 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1333 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1334 	if (rq->flags & MLX5_IB_RQ_PCI_WRITE_END_PADDING)
1335 		MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1336 	MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1337 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1338 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1339 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1340 	MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1341 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1342 
1343 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1344 	qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1345 	memcpy(pas, qp_pas, rq_pas_size);
1346 
1347 	err = mlx5_core_create_rq_tracked(dev, in, inlen, &rq->base.mqp);
1348 
1349 	kvfree(in);
1350 
1351 	return err;
1352 }
1353 
1354 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1355 				     struct mlx5_ib_rq *rq)
1356 {
1357 	mlx5_core_destroy_rq_tracked(dev, &rq->base.mqp);
1358 }
1359 
1360 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1361 				      struct mlx5_ib_rq *rq,
1362 				      u32 qp_flags_en,
1363 				      struct ib_pd *pd)
1364 {
1365 	if (qp_flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1366 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1367 		mlx5_ib_disable_lb(dev, false, true);
1368 	mlx5_cmd_destroy_tir(dev->mdev, rq->tirn, to_mpd(pd)->uid);
1369 }
1370 
1371 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1372 				    struct mlx5_ib_rq *rq, u32 tdn,
1373 				    u32 *qp_flags_en, struct ib_pd *pd,
1374 				    u32 *out)
1375 {
1376 	u8 lb_flag = 0;
1377 	u32 *in;
1378 	void *tirc;
1379 	int inlen;
1380 	int err;
1381 
1382 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1383 	in = kvzalloc(inlen, GFP_KERNEL);
1384 	if (!in)
1385 		return -ENOMEM;
1386 
1387 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1388 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1389 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1390 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1391 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1392 	if (*qp_flags_en & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1393 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1394 
1395 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1396 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1397 
1398 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1399 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1400 
1401 	if (dev->is_rep) {
1402 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1403 		*qp_flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1404 	}
1405 
1406 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1407 	MLX5_SET(create_tir_in, in, opcode, MLX5_CMD_OP_CREATE_TIR);
1408 	err = mlx5_cmd_exec_inout(dev->mdev, create_tir, in, out);
1409 	rq->tirn = MLX5_GET(create_tir_out, out, tirn);
1410 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1411 		err = mlx5_ib_enable_lb(dev, false, true);
1412 
1413 		if (err)
1414 			destroy_raw_packet_qp_tir(dev, rq, 0, pd);
1415 	}
1416 	kvfree(in);
1417 
1418 	return err;
1419 }
1420 
1421 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1422 				u32 *in, size_t inlen,
1423 				struct ib_pd *pd,
1424 				struct ib_udata *udata,
1425 				struct mlx5_ib_create_qp_resp *resp)
1426 {
1427 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1428 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1429 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1430 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1431 		udata, struct mlx5_ib_ucontext, ibucontext);
1432 	int err;
1433 	u32 tdn = mucontext->tdn;
1434 	u16 uid = to_mpd(pd)->uid;
1435 	u32 out[MLX5_ST_SZ_DW(create_tir_out)] = {};
1436 
1437 	if (!qp->sq.wqe_cnt && !qp->rq.wqe_cnt)
1438 		return -EINVAL;
1439 	if (qp->sq.wqe_cnt) {
1440 		err = create_raw_packet_qp_tis(dev, qp, sq, tdn, pd);
1441 		if (err)
1442 			return err;
1443 
1444 		err = create_raw_packet_qp_sq(dev, udata, sq, in, pd);
1445 		if (err)
1446 			goto err_destroy_tis;
1447 
1448 		if (uid) {
1449 			resp->tisn = sq->tisn;
1450 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TISN;
1451 			resp->sqn = sq->base.mqp.qpn;
1452 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_SQN;
1453 		}
1454 
1455 		sq->base.container_mibqp = qp;
1456 		sq->base.mqp.event = mlx5_ib_qp_event;
1457 	}
1458 
1459 	if (qp->rq.wqe_cnt) {
1460 		rq->base.container_mibqp = qp;
1461 
1462 		if (qp->flags & IB_QP_CREATE_CVLAN_STRIPPING)
1463 			rq->flags |= MLX5_IB_RQ_CVLAN_STRIPPING;
1464 		if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING)
1465 			rq->flags |= MLX5_IB_RQ_PCI_WRITE_END_PADDING;
1466 		err = create_raw_packet_qp_rq(dev, rq, in, inlen, pd);
1467 		if (err)
1468 			goto err_destroy_sq;
1469 
1470 		err = create_raw_packet_qp_tir(dev, rq, tdn, &qp->flags_en, pd,
1471 					       out);
1472 		if (err)
1473 			goto err_destroy_rq;
1474 
1475 		if (uid) {
1476 			resp->rqn = rq->base.mqp.qpn;
1477 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_RQN;
1478 			resp->tirn = rq->tirn;
1479 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1480 			if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
1481 			    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2)) {
1482 				resp->tir_icm_addr = MLX5_GET(
1483 					create_tir_out, out, icm_address_31_0);
1484 				resp->tir_icm_addr |=
1485 					(u64)MLX5_GET(create_tir_out, out,
1486 						      icm_address_39_32)
1487 					<< 32;
1488 				resp->tir_icm_addr |=
1489 					(u64)MLX5_GET(create_tir_out, out,
1490 						      icm_address_63_40)
1491 					<< 40;
1492 				resp->comp_mask |=
1493 					MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1494 			}
1495 		}
1496 	}
1497 
1498 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1499 						     rq->base.mqp.qpn;
1500 	return 0;
1501 
1502 err_destroy_rq:
1503 	destroy_raw_packet_qp_rq(dev, rq);
1504 err_destroy_sq:
1505 	if (!qp->sq.wqe_cnt)
1506 		return err;
1507 	destroy_raw_packet_qp_sq(dev, sq);
1508 err_destroy_tis:
1509 	destroy_raw_packet_qp_tis(dev, sq, pd);
1510 
1511 	return err;
1512 }
1513 
1514 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1515 				  struct mlx5_ib_qp *qp)
1516 {
1517 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1518 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1519 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1520 
1521 	if (qp->rq.wqe_cnt) {
1522 		destroy_raw_packet_qp_tir(dev, rq, qp->flags_en, qp->ibqp.pd);
1523 		destroy_raw_packet_qp_rq(dev, rq);
1524 	}
1525 
1526 	if (qp->sq.wqe_cnt) {
1527 		destroy_raw_packet_qp_sq(dev, sq);
1528 		destroy_raw_packet_qp_tis(dev, sq, qp->ibqp.pd);
1529 	}
1530 }
1531 
1532 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1533 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1534 {
1535 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1536 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1537 
1538 	sq->sq = &qp->sq;
1539 	rq->rq = &qp->rq;
1540 	sq->doorbell = &qp->db;
1541 	rq->doorbell = &qp->db;
1542 }
1543 
1544 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1545 {
1546 	if (qp->flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1547 			    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1548 		mlx5_ib_disable_lb(dev, false, true);
1549 	mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1550 			     to_mpd(qp->ibqp.pd)->uid);
1551 }
1552 
1553 struct mlx5_create_qp_params {
1554 	struct ib_udata *udata;
1555 	size_t inlen;
1556 	size_t outlen;
1557 	size_t ucmd_size;
1558 	void *ucmd;
1559 	u8 is_rss_raw : 1;
1560 	struct ib_qp_init_attr *attr;
1561 	u32 uidx;
1562 	struct mlx5_ib_create_qp_resp resp;
1563 };
1564 
1565 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1566 				 struct mlx5_ib_qp *qp,
1567 				 struct mlx5_create_qp_params *params)
1568 {
1569 	struct ib_qp_init_attr *init_attr = params->attr;
1570 	struct mlx5_ib_create_qp_rss *ucmd = params->ucmd;
1571 	struct ib_udata *udata = params->udata;
1572 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1573 		udata, struct mlx5_ib_ucontext, ibucontext);
1574 	int inlen;
1575 	int outlen;
1576 	int err;
1577 	u32 *in;
1578 	u32 *out;
1579 	void *tirc;
1580 	void *hfso;
1581 	u32 selected_fields = 0;
1582 	u32 outer_l4;
1583 	u32 tdn = mucontext->tdn;
1584 	u8 lb_flag = 0;
1585 
1586 	if (ucmd->comp_mask) {
1587 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1588 		return -EOPNOTSUPP;
1589 	}
1590 
1591 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_INNER &&
1592 	    !(ucmd->flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)) {
1593 		mlx5_ib_dbg(dev, "Tunnel offloads must be set for inner RSS\n");
1594 		return -EOPNOTSUPP;
1595 	}
1596 
1597 	if (dev->is_rep)
1598 		qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1599 
1600 	if (qp->flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1601 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1602 
1603 	if (qp->flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1604 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1605 
1606 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1607 	outlen = MLX5_ST_SZ_BYTES(create_tir_out);
1608 	in = kvzalloc(inlen + outlen, GFP_KERNEL);
1609 	if (!in)
1610 		return -ENOMEM;
1611 
1612 	out = in + MLX5_ST_SZ_DW(create_tir_in);
1613 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1614 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1615 	MLX5_SET(tirc, tirc, disp_type,
1616 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1617 	MLX5_SET(tirc, tirc, indirect_table,
1618 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1619 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1620 
1621 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1622 
1623 	if (ucmd->flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1624 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1625 
1626 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1627 
1628 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_INNER)
1629 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner);
1630 	else
1631 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1632 
1633 	switch (ucmd->rx_hash_function) {
1634 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1635 	{
1636 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1637 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1638 
1639 		if (len != ucmd->rx_key_len) {
1640 			err = -EINVAL;
1641 			goto err;
1642 		}
1643 
1644 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1645 		memcpy(rss_key, ucmd->rx_hash_key, len);
1646 		break;
1647 	}
1648 	default:
1649 		err = -EOPNOTSUPP;
1650 		goto err;
1651 	}
1652 
1653 	if (!ucmd->rx_hash_fields_mask) {
1654 		/* special case when this TIR serves as steering entry without hashing */
1655 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1656 			goto create_tir;
1657 		err = -EINVAL;
1658 		goto err;
1659 	}
1660 
1661 	if (((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1662 	     (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1663 	     ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1664 	     (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1665 		err = -EINVAL;
1666 		goto err;
1667 	}
1668 
1669 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1670 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1671 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1672 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1673 			 MLX5_L3_PROT_TYPE_IPV4);
1674 	else if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1675 		 (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1676 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1677 			 MLX5_L3_PROT_TYPE_IPV6);
1678 
1679 	outer_l4 = ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1680 		    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1681 			   << 0 |
1682 		   ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1683 		    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1684 			   << 1 |
1685 		   (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI) << 2;
1686 
1687 	/* Check that only one l4 protocol is set */
1688 	if (outer_l4 & (outer_l4 - 1)) {
1689 		err = -EINVAL;
1690 		goto err;
1691 	}
1692 
1693 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1694 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1695 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1696 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1697 			 MLX5_L4_PROT_TYPE_TCP);
1698 	else if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1699 		 (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1700 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1701 			 MLX5_L4_PROT_TYPE_UDP);
1702 
1703 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1704 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1705 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1706 
1707 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1708 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1709 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1710 
1711 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1712 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1713 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1714 
1715 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1716 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1717 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1718 
1719 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI)
1720 		selected_fields |= MLX5_HASH_FIELD_SEL_IPSEC_SPI;
1721 
1722 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1723 
1724 create_tir:
1725 	MLX5_SET(create_tir_in, in, opcode, MLX5_CMD_OP_CREATE_TIR);
1726 	err = mlx5_cmd_exec_inout(dev->mdev, create_tir, in, out);
1727 
1728 	qp->rss_qp.tirn = MLX5_GET(create_tir_out, out, tirn);
1729 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1730 		err = mlx5_ib_enable_lb(dev, false, true);
1731 
1732 		if (err)
1733 			mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1734 					     to_mpd(pd)->uid);
1735 	}
1736 
1737 	if (err)
1738 		goto err;
1739 
1740 	if (mucontext->devx_uid) {
1741 		params->resp.comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1742 		params->resp.tirn = qp->rss_qp.tirn;
1743 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
1744 		    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2)) {
1745 			params->resp.tir_icm_addr =
1746 				MLX5_GET(create_tir_out, out, icm_address_31_0);
1747 			params->resp.tir_icm_addr |=
1748 				(u64)MLX5_GET(create_tir_out, out,
1749 					      icm_address_39_32)
1750 				<< 32;
1751 			params->resp.tir_icm_addr |=
1752 				(u64)MLX5_GET(create_tir_out, out,
1753 					      icm_address_63_40)
1754 				<< 40;
1755 			params->resp.comp_mask |=
1756 				MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1757 		}
1758 	}
1759 
1760 	kvfree(in);
1761 	/* qpn is reserved for that QP */
1762 	qp->trans_qp.base.mqp.qpn = 0;
1763 	qp->is_rss = true;
1764 	return 0;
1765 
1766 err:
1767 	kvfree(in);
1768 	return err;
1769 }
1770 
1771 static void configure_requester_scat_cqe(struct mlx5_ib_dev *dev,
1772 					 struct mlx5_ib_qp *qp,
1773 					 struct ib_qp_init_attr *init_attr,
1774 					 void *qpc)
1775 {
1776 	int scqe_sz;
1777 	bool allow_scat_cqe = false;
1778 
1779 	allow_scat_cqe = qp->flags_en & MLX5_QP_FLAG_ALLOW_SCATTER_CQE;
1780 
1781 	if (!allow_scat_cqe && init_attr->sq_sig_type != IB_SIGNAL_ALL_WR)
1782 		return;
1783 
1784 	scqe_sz = mlx5_ib_get_cqe_size(init_attr->send_cq);
1785 	if (scqe_sz == 128) {
1786 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1787 		return;
1788 	}
1789 
1790 	if (init_attr->qp_type != MLX5_IB_QPT_DCI ||
1791 	    MLX5_CAP_GEN(dev->mdev, dc_req_scat_data_cqe))
1792 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1793 }
1794 
1795 static int atomic_size_to_mode(int size_mask)
1796 {
1797 	/* driver does not support atomic_size > 256B
1798 	 * and does not know how to translate bigger sizes
1799 	 */
1800 	int supported_size_mask = size_mask & 0x1ff;
1801 	int log_max_size;
1802 
1803 	if (!supported_size_mask)
1804 		return -EOPNOTSUPP;
1805 
1806 	log_max_size = __fls(supported_size_mask);
1807 
1808 	if (log_max_size > 3)
1809 		return log_max_size;
1810 
1811 	return MLX5_ATOMIC_MODE_8B;
1812 }
1813 
1814 static int get_atomic_mode(struct mlx5_ib_dev *dev,
1815 			   enum ib_qp_type qp_type)
1816 {
1817 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
1818 	u8 atomic = MLX5_CAP_GEN(dev->mdev, atomic);
1819 	int atomic_mode = -EOPNOTSUPP;
1820 	int atomic_size_mask;
1821 
1822 	if (!atomic)
1823 		return -EOPNOTSUPP;
1824 
1825 	if (qp_type == MLX5_IB_QPT_DCT)
1826 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
1827 	else
1828 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
1829 
1830 	if ((atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_CMP_SWAP) ||
1831 	    (atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_FETCH_ADD))
1832 		atomic_mode = atomic_size_to_mode(atomic_size_mask);
1833 
1834 	if (atomic_mode <= 0 &&
1835 	    (atomic_operations & MLX5_ATOMIC_OPS_CMP_SWAP &&
1836 	     atomic_operations & MLX5_ATOMIC_OPS_FETCH_ADD))
1837 		atomic_mode = MLX5_ATOMIC_MODE_IB_COMP;
1838 
1839 	return atomic_mode;
1840 }
1841 
1842 static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1843 			     struct mlx5_create_qp_params *params)
1844 {
1845 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
1846 	struct ib_qp_init_attr *attr = params->attr;
1847 	u32 uidx = params->uidx;
1848 	struct mlx5_ib_resources *devr = &dev->devr;
1849 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
1850 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1851 	struct mlx5_core_dev *mdev = dev->mdev;
1852 	struct mlx5_ib_qp_base *base;
1853 	unsigned long flags;
1854 	void *qpc;
1855 	u32 *in;
1856 	int err;
1857 
1858 	if (attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1859 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1860 
1861 	in = kvzalloc(inlen, GFP_KERNEL);
1862 	if (!in)
1863 		return -ENOMEM;
1864 
1865 	if (MLX5_CAP_GEN(mdev, ece_support) && ucmd)
1866 		MLX5_SET(create_qp_in, in, ece, ucmd->ece_options);
1867 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1868 
1869 	MLX5_SET(qpc, qpc, st, MLX5_QP_ST_XRC);
1870 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
1871 	MLX5_SET(qpc, qpc, pd, to_mpd(devr->p0)->pdn);
1872 
1873 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1874 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
1875 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
1876 		MLX5_SET(qpc, qpc, cd_master, 1);
1877 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
1878 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
1879 	if (qp->flags & IB_QP_CREATE_MANAGED_RECV)
1880 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
1881 
1882 	MLX5_SET(qpc, qpc, rq_type, MLX5_SRQ_RQ);
1883 	MLX5_SET(qpc, qpc, no_sq, 1);
1884 	MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1885 	MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
1886 	MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
1887 	MLX5_SET(qpc, qpc, xrcd, to_mxrcd(attr->xrcd)->xrcdn);
1888 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
1889 
1890 	/* 0xffffff means we ask to work with cqe version 0 */
1891 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
1892 		MLX5_SET(qpc, qpc, user_index, uidx);
1893 
1894 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING) {
1895 		MLX5_SET(qpc, qpc, end_padding_mode,
1896 			 MLX5_WQ_END_PAD_MODE_ALIGN);
1897 		/* Special case to clean flag */
1898 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
1899 	}
1900 
1901 	base = &qp->trans_qp.base;
1902 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
1903 	kvfree(in);
1904 	if (err)
1905 		return err;
1906 
1907 	base->container_mibqp = qp;
1908 	base->mqp.event = mlx5_ib_qp_event;
1909 	if (MLX5_CAP_GEN(mdev, ece_support))
1910 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
1911 
1912 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1913 	list_add_tail(&qp->qps_list, &dev->qp_list);
1914 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1915 
1916 	qp->trans_qp.xrcdn = to_mxrcd(attr->xrcd)->xrcdn;
1917 	return 0;
1918 }
1919 
1920 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1921 			  struct mlx5_ib_qp *qp,
1922 			  struct mlx5_create_qp_params *params)
1923 {
1924 	struct ib_qp_init_attr *init_attr = params->attr;
1925 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
1926 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
1927 	struct ib_udata *udata = params->udata;
1928 	u32 uidx = params->uidx;
1929 	struct mlx5_ib_resources *devr = &dev->devr;
1930 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1931 	struct mlx5_core_dev *mdev = dev->mdev;
1932 	struct mlx5_ib_cq *send_cq;
1933 	struct mlx5_ib_cq *recv_cq;
1934 	unsigned long flags;
1935 	struct mlx5_ib_qp_base *base;
1936 	int mlx5_st;
1937 	void *qpc;
1938 	u32 *in;
1939 	int err;
1940 
1941 	spin_lock_init(&qp->sq.lock);
1942 	spin_lock_init(&qp->rq.lock);
1943 
1944 	mlx5_st = to_mlx5_st(qp->type);
1945 	if (mlx5_st < 0)
1946 		return -EINVAL;
1947 
1948 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1949 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1950 
1951 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
1952 		qp->underlay_qpn = init_attr->source_qpn;
1953 
1954 	base = (init_attr->qp_type == IB_QPT_RAW_PACKET ||
1955 		qp->flags & IB_QP_CREATE_SOURCE_QPN) ?
1956 	       &qp->raw_packet_qp.rq.base :
1957 	       &qp->trans_qp.base;
1958 
1959 	qp->has_rq = qp_has_rq(init_attr);
1960 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq, qp, ucmd);
1961 	if (err) {
1962 		mlx5_ib_dbg(dev, "err %d\n", err);
1963 		return err;
1964 	}
1965 
1966 	if (ucmd->rq_wqe_shift != qp->rq.wqe_shift ||
1967 	    ucmd->rq_wqe_count != qp->rq.wqe_cnt)
1968 		return -EINVAL;
1969 
1970 	if (ucmd->sq_wqe_count > (1 << MLX5_CAP_GEN(mdev, log_max_qp_sz)))
1971 		return -EINVAL;
1972 
1973 	err = _create_user_qp(dev, pd, qp, udata, init_attr, &in, &params->resp,
1974 			      &inlen, base, ucmd);
1975 	if (err)
1976 		return err;
1977 
1978 	if (is_sqp(init_attr->qp_type))
1979 		qp->port = init_attr->port_num;
1980 
1981 	if (MLX5_CAP_GEN(mdev, ece_support))
1982 		MLX5_SET(create_qp_in, in, ece, ucmd->ece_options);
1983 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1984 
1985 	MLX5_SET(qpc, qpc, st, mlx5_st);
1986 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
1987 	MLX5_SET(qpc, qpc, pd, to_mpd(pd)->pdn);
1988 
1989 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE)
1990 		MLX5_SET(qpc, qpc, wq_signature, 1);
1991 
1992 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1993 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
1994 
1995 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
1996 		MLX5_SET(qpc, qpc, cd_master, 1);
1997 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
1998 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
1999 	if (qp->flags & IB_QP_CREATE_MANAGED_RECV)
2000 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
2001 	if (qp->flags_en & MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE)
2002 		MLX5_SET(qpc, qpc, req_e2e_credit_mode, 1);
2003 	if ((qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) &&
2004 	    (init_attr->qp_type == IB_QPT_RC ||
2005 	     init_attr->qp_type == IB_QPT_UC)) {
2006 		int rcqe_sz = mlx5_ib_get_cqe_size(init_attr->recv_cq);
2007 
2008 		MLX5_SET(qpc, qpc, cs_res,
2009 			 rcqe_sz == 128 ? MLX5_RES_SCAT_DATA64_CQE :
2010 					  MLX5_RES_SCAT_DATA32_CQE);
2011 	}
2012 	if ((qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) &&
2013 	    (qp->type == MLX5_IB_QPT_DCI || qp->type == IB_QPT_RC))
2014 		configure_requester_scat_cqe(dev, qp, init_attr, qpc);
2015 
2016 	if (qp->rq.wqe_cnt) {
2017 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2018 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2019 	}
2020 
2021 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
2022 
2023 	if (qp->sq.wqe_cnt) {
2024 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2025 	} else {
2026 		MLX5_SET(qpc, qpc, no_sq, 1);
2027 		if (init_attr->srq &&
2028 		    init_attr->srq->srq_type == IB_SRQT_TM)
2029 			MLX5_SET(qpc, qpc, offload_type,
2030 				 MLX5_QPC_OFFLOAD_TYPE_RNDV);
2031 	}
2032 
2033 	/* Set default resources */
2034 	switch (init_attr->qp_type) {
2035 	case IB_QPT_XRC_INI:
2036 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2037 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2038 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2039 		break;
2040 	default:
2041 		if (init_attr->srq) {
2042 			MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2043 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(init_attr->srq)->msrq.srqn);
2044 		} else {
2045 			MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2046 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s1)->msrq.srqn);
2047 		}
2048 	}
2049 
2050 	if (init_attr->send_cq)
2051 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
2052 
2053 	if (init_attr->recv_cq)
2054 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
2055 
2056 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2057 
2058 	/* 0xffffff means we ask to work with cqe version 0 */
2059 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2060 		MLX5_SET(qpc, qpc, user_index, uidx);
2061 
2062 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING &&
2063 	    init_attr->qp_type != IB_QPT_RAW_PACKET) {
2064 		MLX5_SET(qpc, qpc, end_padding_mode,
2065 			 MLX5_WQ_END_PAD_MODE_ALIGN);
2066 		/* Special case to clean flag */
2067 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
2068 	}
2069 
2070 	if (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2071 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
2072 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd->sq_buf_addr;
2073 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
2074 		err = create_raw_packet_qp(dev, qp, in, inlen, pd, udata,
2075 					   &params->resp);
2076 	} else
2077 		err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2078 
2079 	kvfree(in);
2080 	if (err)
2081 		goto err_create;
2082 
2083 	base->container_mibqp = qp;
2084 	base->mqp.event = mlx5_ib_qp_event;
2085 	if (MLX5_CAP_GEN(mdev, ece_support))
2086 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
2087 
2088 	get_cqs(qp->type, init_attr->send_cq, init_attr->recv_cq,
2089 		&send_cq, &recv_cq);
2090 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2091 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2092 	/* Maintain device to QPs access, needed for further handling via reset
2093 	 * flow
2094 	 */
2095 	list_add_tail(&qp->qps_list, &dev->qp_list);
2096 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2097 	 */
2098 	if (send_cq)
2099 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2100 	if (recv_cq)
2101 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2102 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2103 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2104 
2105 	return 0;
2106 
2107 err_create:
2108 	destroy_qp(dev, qp, base, udata);
2109 	return err;
2110 }
2111 
2112 static int create_kernel_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2113 			    struct mlx5_ib_qp *qp,
2114 			    struct mlx5_create_qp_params *params)
2115 {
2116 	struct ib_qp_init_attr *attr = params->attr;
2117 	u32 uidx = params->uidx;
2118 	struct mlx5_ib_resources *devr = &dev->devr;
2119 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
2120 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
2121 	struct mlx5_core_dev *mdev = dev->mdev;
2122 	struct mlx5_ib_cq *send_cq;
2123 	struct mlx5_ib_cq *recv_cq;
2124 	unsigned long flags;
2125 	struct mlx5_ib_qp_base *base;
2126 	int mlx5_st;
2127 	void *qpc;
2128 	u32 *in;
2129 	int err;
2130 
2131 	spin_lock_init(&qp->sq.lock);
2132 	spin_lock_init(&qp->rq.lock);
2133 
2134 	mlx5_st = to_mlx5_st(qp->type);
2135 	if (mlx5_st < 0)
2136 		return -EINVAL;
2137 
2138 	if (attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2139 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2140 
2141 	base = &qp->trans_qp.base;
2142 
2143 	qp->has_rq = qp_has_rq(attr);
2144 	err = set_rq_size(dev, &attr->cap, qp->has_rq, qp, NULL);
2145 	if (err) {
2146 		mlx5_ib_dbg(dev, "err %d\n", err);
2147 		return err;
2148 	}
2149 
2150 	err = _create_kernel_qp(dev, attr, qp, &in, &inlen, base);
2151 	if (err)
2152 		return err;
2153 
2154 	if (is_sqp(attr->qp_type))
2155 		qp->port = attr->port_num;
2156 
2157 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2158 
2159 	MLX5_SET(qpc, qpc, st, mlx5_st);
2160 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2161 
2162 	if (attr->qp_type != MLX5_IB_QPT_REG_UMR)
2163 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
2164 	else
2165 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
2166 
2167 
2168 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
2169 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2170 
2171 	if (qp->rq.wqe_cnt) {
2172 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2173 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2174 	}
2175 
2176 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, attr));
2177 
2178 	if (qp->sq.wqe_cnt)
2179 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2180 	else
2181 		MLX5_SET(qpc, qpc, no_sq, 1);
2182 
2183 	if (attr->srq) {
2184 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2185 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2186 			 to_msrq(attr->srq)->msrq.srqn);
2187 	} else {
2188 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2189 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2190 			 to_msrq(devr->s1)->msrq.srqn);
2191 	}
2192 
2193 	if (attr->send_cq)
2194 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(attr->send_cq)->mcq.cqn);
2195 
2196 	if (attr->recv_cq)
2197 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(attr->recv_cq)->mcq.cqn);
2198 
2199 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2200 
2201 	/* 0xffffff means we ask to work with cqe version 0 */
2202 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2203 		MLX5_SET(qpc, qpc, user_index, uidx);
2204 
2205 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
2206 	if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO)
2207 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
2208 
2209 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2210 	kvfree(in);
2211 	if (err)
2212 		goto err_create;
2213 
2214 	base->container_mibqp = qp;
2215 	base->mqp.event = mlx5_ib_qp_event;
2216 
2217 	get_cqs(qp->type, attr->send_cq, attr->recv_cq,
2218 		&send_cq, &recv_cq);
2219 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2220 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2221 	/* Maintain device to QPs access, needed for further handling via reset
2222 	 * flow
2223 	 */
2224 	list_add_tail(&qp->qps_list, &dev->qp_list);
2225 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2226 	 */
2227 	if (send_cq)
2228 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2229 	if (recv_cq)
2230 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2231 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2232 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2233 
2234 	return 0;
2235 
2236 err_create:
2237 	destroy_qp(dev, qp, base, NULL);
2238 	return err;
2239 }
2240 
2241 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2242 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
2243 {
2244 	if (send_cq) {
2245 		if (recv_cq) {
2246 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2247 				spin_lock(&send_cq->lock);
2248 				spin_lock_nested(&recv_cq->lock,
2249 						 SINGLE_DEPTH_NESTING);
2250 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2251 				spin_lock(&send_cq->lock);
2252 				__acquire(&recv_cq->lock);
2253 			} else {
2254 				spin_lock(&recv_cq->lock);
2255 				spin_lock_nested(&send_cq->lock,
2256 						 SINGLE_DEPTH_NESTING);
2257 			}
2258 		} else {
2259 			spin_lock(&send_cq->lock);
2260 			__acquire(&recv_cq->lock);
2261 		}
2262 	} else if (recv_cq) {
2263 		spin_lock(&recv_cq->lock);
2264 		__acquire(&send_cq->lock);
2265 	} else {
2266 		__acquire(&send_cq->lock);
2267 		__acquire(&recv_cq->lock);
2268 	}
2269 }
2270 
2271 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2272 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
2273 {
2274 	if (send_cq) {
2275 		if (recv_cq) {
2276 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2277 				spin_unlock(&recv_cq->lock);
2278 				spin_unlock(&send_cq->lock);
2279 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2280 				__release(&recv_cq->lock);
2281 				spin_unlock(&send_cq->lock);
2282 			} else {
2283 				spin_unlock(&send_cq->lock);
2284 				spin_unlock(&recv_cq->lock);
2285 			}
2286 		} else {
2287 			__release(&recv_cq->lock);
2288 			spin_unlock(&send_cq->lock);
2289 		}
2290 	} else if (recv_cq) {
2291 		__release(&send_cq->lock);
2292 		spin_unlock(&recv_cq->lock);
2293 	} else {
2294 		__release(&recv_cq->lock);
2295 		__release(&send_cq->lock);
2296 	}
2297 }
2298 
2299 static void get_cqs(enum ib_qp_type qp_type,
2300 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
2301 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
2302 {
2303 	switch (qp_type) {
2304 	case IB_QPT_XRC_TGT:
2305 		*send_cq = NULL;
2306 		*recv_cq = NULL;
2307 		break;
2308 	case MLX5_IB_QPT_REG_UMR:
2309 	case IB_QPT_XRC_INI:
2310 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2311 		*recv_cq = NULL;
2312 		break;
2313 
2314 	case IB_QPT_SMI:
2315 	case MLX5_IB_QPT_HW_GSI:
2316 	case IB_QPT_RC:
2317 	case IB_QPT_UC:
2318 	case IB_QPT_UD:
2319 	case IB_QPT_RAW_PACKET:
2320 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2321 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
2322 		break;
2323 	default:
2324 		*send_cq = NULL;
2325 		*recv_cq = NULL;
2326 		break;
2327 	}
2328 }
2329 
2330 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2331 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2332 				u8 lag_tx_affinity);
2333 
2334 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2335 			      struct ib_udata *udata)
2336 {
2337 	struct mlx5_ib_cq *send_cq, *recv_cq;
2338 	struct mlx5_ib_qp_base *base;
2339 	unsigned long flags;
2340 	int err;
2341 
2342 	if (qp->is_rss) {
2343 		destroy_rss_raw_qp_tir(dev, qp);
2344 		return;
2345 	}
2346 
2347 	base = (qp->type == IB_QPT_RAW_PACKET ||
2348 		qp->flags & IB_QP_CREATE_SOURCE_QPN) ?
2349 		       &qp->raw_packet_qp.rq.base :
2350 		       &qp->trans_qp.base;
2351 
2352 	if (qp->state != IB_QPS_RESET) {
2353 		if (qp->type != IB_QPT_RAW_PACKET &&
2354 		    !(qp->flags & IB_QP_CREATE_SOURCE_QPN)) {
2355 			err = mlx5_core_qp_modify(dev, MLX5_CMD_OP_2RST_QP, 0,
2356 						  NULL, &base->mqp, NULL);
2357 		} else {
2358 			struct mlx5_modify_raw_qp_param raw_qp_param = {
2359 				.operation = MLX5_CMD_OP_2RST_QP
2360 			};
2361 
2362 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2363 		}
2364 		if (err)
2365 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
2366 				     base->mqp.qpn);
2367 	}
2368 
2369 	get_cqs(qp->type, qp->ibqp.send_cq, qp->ibqp.recv_cq, &send_cq,
2370 		&recv_cq);
2371 
2372 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2373 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2374 	/* del from lists under both locks above to protect reset flow paths */
2375 	list_del(&qp->qps_list);
2376 	if (send_cq)
2377 		list_del(&qp->cq_send_list);
2378 
2379 	if (recv_cq)
2380 		list_del(&qp->cq_recv_list);
2381 
2382 	if (!udata) {
2383 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2384 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
2385 		if (send_cq != recv_cq)
2386 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
2387 					   NULL);
2388 	}
2389 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2390 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2391 
2392 	if (qp->type == IB_QPT_RAW_PACKET ||
2393 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
2394 		destroy_raw_packet_qp(dev, qp);
2395 	} else {
2396 		err = mlx5_core_destroy_qp(dev, &base->mqp);
2397 		if (err)
2398 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
2399 				     base->mqp.qpn);
2400 	}
2401 
2402 	destroy_qp(dev, qp, base, udata);
2403 }
2404 
2405 static int create_dct(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2406 		      struct mlx5_ib_qp *qp,
2407 		      struct mlx5_create_qp_params *params)
2408 {
2409 	struct ib_qp_init_attr *attr = params->attr;
2410 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
2411 	u32 uidx = params->uidx;
2412 	void *dctc;
2413 
2414 	if (mlx5_lag_is_active(dev->mdev) && !MLX5_CAP_GEN(dev->mdev, lag_dct))
2415 		return -EOPNOTSUPP;
2416 
2417 	qp->dct.in = kzalloc(MLX5_ST_SZ_BYTES(create_dct_in), GFP_KERNEL);
2418 	if (!qp->dct.in)
2419 		return -ENOMEM;
2420 
2421 	MLX5_SET(create_dct_in, qp->dct.in, uid, to_mpd(pd)->uid);
2422 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
2423 	MLX5_SET(dctc, dctc, pd, to_mpd(pd)->pdn);
2424 	MLX5_SET(dctc, dctc, srqn_xrqn, to_msrq(attr->srq)->msrq.srqn);
2425 	MLX5_SET(dctc, dctc, cqn, to_mcq(attr->recv_cq)->mcq.cqn);
2426 	MLX5_SET64(dctc, dctc, dc_access_key, ucmd->access_key);
2427 	MLX5_SET(dctc, dctc, user_index, uidx);
2428 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
2429 		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
2430 
2431 	if (qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) {
2432 		int rcqe_sz = mlx5_ib_get_cqe_size(attr->recv_cq);
2433 
2434 		if (rcqe_sz == 128)
2435 			MLX5_SET(dctc, dctc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
2436 	}
2437 
2438 	qp->state = IB_QPS_RESET;
2439 
2440 	return 0;
2441 }
2442 
2443 static int check_qp_type(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
2444 			 enum ib_qp_type *type)
2445 {
2446 	if (attr->qp_type == IB_QPT_DRIVER && !MLX5_CAP_GEN(dev->mdev, dct))
2447 		goto out;
2448 
2449 	switch (attr->qp_type) {
2450 	case IB_QPT_XRC_TGT:
2451 	case IB_QPT_XRC_INI:
2452 		if (!MLX5_CAP_GEN(dev->mdev, xrc))
2453 			goto out;
2454 		fallthrough;
2455 	case IB_QPT_RC:
2456 	case IB_QPT_UC:
2457 	case IB_QPT_SMI:
2458 	case MLX5_IB_QPT_HW_GSI:
2459 	case IB_QPT_DRIVER:
2460 	case IB_QPT_GSI:
2461 		if (dev->profile == &raw_eth_profile)
2462 			goto out;
2463 	case IB_QPT_RAW_PACKET:
2464 	case IB_QPT_UD:
2465 	case MLX5_IB_QPT_REG_UMR:
2466 		break;
2467 	default:
2468 		goto out;
2469 	}
2470 
2471 	*type = attr->qp_type;
2472 	return 0;
2473 
2474 out:
2475 	mlx5_ib_dbg(dev, "Unsupported QP type %d\n", attr->qp_type);
2476 	return -EOPNOTSUPP;
2477 }
2478 
2479 static int check_valid_flow(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2480 			    struct ib_qp_init_attr *attr,
2481 			    struct ib_udata *udata)
2482 {
2483 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2484 		udata, struct mlx5_ib_ucontext, ibucontext);
2485 
2486 	if (!udata) {
2487 		/* Kernel create_qp callers */
2488 		if (attr->rwq_ind_tbl)
2489 			return -EOPNOTSUPP;
2490 
2491 		switch (attr->qp_type) {
2492 		case IB_QPT_RAW_PACKET:
2493 		case IB_QPT_DRIVER:
2494 			return -EOPNOTSUPP;
2495 		default:
2496 			return 0;
2497 		}
2498 	}
2499 
2500 	/* Userspace create_qp callers */
2501 	if (attr->qp_type == IB_QPT_RAW_PACKET && !ucontext->cqe_version) {
2502 		mlx5_ib_dbg(dev,
2503 			"Raw Packet QP is only supported for CQE version > 0\n");
2504 		return -EINVAL;
2505 	}
2506 
2507 	if (attr->qp_type != IB_QPT_RAW_PACKET && attr->rwq_ind_tbl) {
2508 		mlx5_ib_dbg(dev,
2509 			    "Wrong QP type %d for the RWQ indirect table\n",
2510 			    attr->qp_type);
2511 		return -EINVAL;
2512 	}
2513 
2514 	/*
2515 	 * We don't need to see this warning, it means that kernel code
2516 	 * missing ib_pd. Placed here to catch developer's mistakes.
2517 	 */
2518 	WARN_ONCE(!pd && attr->qp_type != IB_QPT_XRC_TGT,
2519 		  "There is a missing PD pointer assignment\n");
2520 	return 0;
2521 }
2522 
2523 static void process_vendor_flag(struct mlx5_ib_dev *dev, int *flags, int flag,
2524 				bool cond, struct mlx5_ib_qp *qp)
2525 {
2526 	if (!(*flags & flag))
2527 		return;
2528 
2529 	if (cond) {
2530 		qp->flags_en |= flag;
2531 		*flags &= ~flag;
2532 		return;
2533 	}
2534 
2535 	switch (flag) {
2536 	case MLX5_QP_FLAG_SCATTER_CQE:
2537 	case MLX5_QP_FLAG_ALLOW_SCATTER_CQE:
2538 		/*
2539 			 * We don't return error if these flags were provided,
2540 			 * and mlx5 doesn't have right capability.
2541 			 */
2542 		*flags &= ~(MLX5_QP_FLAG_SCATTER_CQE |
2543 			    MLX5_QP_FLAG_ALLOW_SCATTER_CQE);
2544 		return;
2545 	default:
2546 		break;
2547 	}
2548 	mlx5_ib_dbg(dev, "Vendor create QP flag 0x%X is not supported\n", flag);
2549 }
2550 
2551 static int process_vendor_flags(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2552 				void *ucmd, struct ib_qp_init_attr *attr)
2553 {
2554 	struct mlx5_core_dev *mdev = dev->mdev;
2555 	bool cond;
2556 	int flags;
2557 
2558 	if (attr->rwq_ind_tbl)
2559 		flags = ((struct mlx5_ib_create_qp_rss *)ucmd)->flags;
2560 	else
2561 		flags = ((struct mlx5_ib_create_qp *)ucmd)->flags;
2562 
2563 	switch (flags & (MLX5_QP_FLAG_TYPE_DCT | MLX5_QP_FLAG_TYPE_DCI)) {
2564 	case MLX5_QP_FLAG_TYPE_DCI:
2565 		qp->type = MLX5_IB_QPT_DCI;
2566 		break;
2567 	case MLX5_QP_FLAG_TYPE_DCT:
2568 		qp->type = MLX5_IB_QPT_DCT;
2569 		break;
2570 	default:
2571 		if (qp->type != IB_QPT_DRIVER)
2572 			break;
2573 		/*
2574 		 * It is IB_QPT_DRIVER and or no subtype or
2575 		 * wrong subtype were provided.
2576 		 */
2577 		return -EINVAL;
2578 	}
2579 
2580 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TYPE_DCI, true, qp);
2581 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TYPE_DCT, true, qp);
2582 
2583 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_SIGNATURE, true, qp);
2584 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_SCATTER_CQE,
2585 			    MLX5_CAP_GEN(mdev, sctr_data_cqe), qp);
2586 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_ALLOW_SCATTER_CQE,
2587 			    MLX5_CAP_GEN(mdev, sctr_data_cqe), qp);
2588 
2589 	if (qp->type == IB_QPT_RAW_PACKET) {
2590 		cond = MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan) ||
2591 		       MLX5_CAP_ETH(mdev, tunnel_stateless_gre) ||
2592 		       MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx);
2593 		process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TUNNEL_OFFLOADS,
2594 				    cond, qp);
2595 		process_vendor_flag(dev, &flags,
2596 				    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC, true,
2597 				    qp);
2598 		process_vendor_flag(dev, &flags,
2599 				    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC, true,
2600 				    qp);
2601 	}
2602 
2603 	if (qp->type == IB_QPT_RC)
2604 		process_vendor_flag(dev, &flags,
2605 				    MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE,
2606 				    MLX5_CAP_GEN(mdev, qp_packet_based), qp);
2607 
2608 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_BFREG_INDEX, true, qp);
2609 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_UAR_PAGE_INDEX, true, qp);
2610 
2611 	cond = qp->flags_en & ~(MLX5_QP_FLAG_TUNNEL_OFFLOADS |
2612 				MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
2613 				MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC);
2614 	if (attr->rwq_ind_tbl && cond) {
2615 		mlx5_ib_dbg(dev, "RSS RAW QP has unsupported flags 0x%X\n",
2616 			    cond);
2617 		return -EINVAL;
2618 	}
2619 
2620 	if (flags)
2621 		mlx5_ib_dbg(dev, "udata has unsupported flags 0x%X\n", flags);
2622 
2623 	return (flags) ? -EINVAL : 0;
2624 	}
2625 
2626 static void process_create_flag(struct mlx5_ib_dev *dev, int *flags, int flag,
2627 				bool cond, struct mlx5_ib_qp *qp)
2628 {
2629 	if (!(*flags & flag))
2630 		return;
2631 
2632 	if (cond) {
2633 		qp->flags |= flag;
2634 		*flags &= ~flag;
2635 		return;
2636 	}
2637 
2638 	if (flag == MLX5_IB_QP_CREATE_WC_TEST) {
2639 		/*
2640 		 * Special case, if condition didn't meet, it won't be error,
2641 		 * just different in-kernel flow.
2642 		 */
2643 		*flags &= ~MLX5_IB_QP_CREATE_WC_TEST;
2644 		return;
2645 	}
2646 	mlx5_ib_dbg(dev, "Verbs create QP flag 0x%X is not supported\n", flag);
2647 }
2648 
2649 static int process_create_flags(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2650 				struct ib_qp_init_attr *attr)
2651 {
2652 	enum ib_qp_type qp_type = qp->type;
2653 	struct mlx5_core_dev *mdev = dev->mdev;
2654 	int create_flags = attr->create_flags;
2655 	bool cond;
2656 
2657 	if (qp->type == IB_QPT_UD && dev->profile == &raw_eth_profile)
2658 		if (create_flags & ~MLX5_IB_QP_CREATE_WC_TEST)
2659 			return -EINVAL;
2660 
2661 	if (qp_type == MLX5_IB_QPT_DCT)
2662 		return (create_flags) ? -EINVAL : 0;
2663 
2664 	if (qp_type == IB_QPT_RAW_PACKET && attr->rwq_ind_tbl)
2665 		return (create_flags) ? -EINVAL : 0;
2666 
2667 	process_create_flag(dev, &create_flags, IB_QP_CREATE_NETIF_QP,
2668 			    mlx5_get_flow_namespace(dev->mdev,
2669 						    MLX5_FLOW_NAMESPACE_BYPASS),
2670 			    qp);
2671 	process_create_flag(dev, &create_flags,
2672 			    IB_QP_CREATE_INTEGRITY_EN,
2673 			    MLX5_CAP_GEN(mdev, sho), qp);
2674 	process_create_flag(dev, &create_flags,
2675 			    IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK,
2676 			    MLX5_CAP_GEN(mdev, block_lb_mc), qp);
2677 	process_create_flag(dev, &create_flags, IB_QP_CREATE_CROSS_CHANNEL,
2678 			    MLX5_CAP_GEN(mdev, cd), qp);
2679 	process_create_flag(dev, &create_flags, IB_QP_CREATE_MANAGED_SEND,
2680 			    MLX5_CAP_GEN(mdev, cd), qp);
2681 	process_create_flag(dev, &create_flags, IB_QP_CREATE_MANAGED_RECV,
2682 			    MLX5_CAP_GEN(mdev, cd), qp);
2683 
2684 	if (qp_type == IB_QPT_UD) {
2685 		process_create_flag(dev, &create_flags,
2686 				    IB_QP_CREATE_IPOIB_UD_LSO,
2687 				    MLX5_CAP_GEN(mdev, ipoib_basic_offloads),
2688 				    qp);
2689 		cond = MLX5_CAP_GEN(mdev, port_type) == MLX5_CAP_PORT_TYPE_IB;
2690 		process_create_flag(dev, &create_flags, IB_QP_CREATE_SOURCE_QPN,
2691 				    cond, qp);
2692 	}
2693 
2694 	if (qp_type == IB_QPT_RAW_PACKET) {
2695 		cond = MLX5_CAP_GEN(mdev, eth_net_offloads) &&
2696 		       MLX5_CAP_ETH(mdev, scatter_fcs);
2697 		process_create_flag(dev, &create_flags,
2698 				    IB_QP_CREATE_SCATTER_FCS, cond, qp);
2699 
2700 		cond = MLX5_CAP_GEN(mdev, eth_net_offloads) &&
2701 		       MLX5_CAP_ETH(mdev, vlan_cap);
2702 		process_create_flag(dev, &create_flags,
2703 				    IB_QP_CREATE_CVLAN_STRIPPING, cond, qp);
2704 	}
2705 
2706 	process_create_flag(dev, &create_flags,
2707 			    IB_QP_CREATE_PCI_WRITE_END_PADDING,
2708 			    MLX5_CAP_GEN(mdev, end_pad), qp);
2709 
2710 	process_create_flag(dev, &create_flags, MLX5_IB_QP_CREATE_WC_TEST,
2711 			    qp_type != MLX5_IB_QPT_REG_UMR, qp);
2712 	process_create_flag(dev, &create_flags, MLX5_IB_QP_CREATE_SQPN_QP1,
2713 			    true, qp);
2714 
2715 	if (create_flags)
2716 		mlx5_ib_dbg(dev, "Create QP has unsupported flags 0x%X\n",
2717 			    create_flags);
2718 
2719 	return (create_flags) ? -EINVAL : 0;
2720 }
2721 
2722 static int process_udata_size(struct mlx5_ib_dev *dev,
2723 			      struct mlx5_create_qp_params *params)
2724 {
2725 	size_t ucmd = sizeof(struct mlx5_ib_create_qp);
2726 	struct ib_udata *udata = params->udata;
2727 	size_t outlen = udata->outlen;
2728 	size_t inlen = udata->inlen;
2729 
2730 	params->outlen = min(outlen, sizeof(struct mlx5_ib_create_qp_resp));
2731 	params->ucmd_size = ucmd;
2732 	if (!params->is_rss_raw) {
2733 		/* User has old rdma-core, which doesn't support ECE */
2734 		size_t min_inlen =
2735 			offsetof(struct mlx5_ib_create_qp, ece_options);
2736 
2737 		/*
2738 		 * We will check in check_ucmd_data() that user
2739 		 * cleared everything after inlen.
2740 		 */
2741 		params->inlen = (inlen < min_inlen) ? 0 : min(inlen, ucmd);
2742 		goto out;
2743 	}
2744 
2745 	/* RSS RAW QP */
2746 	if (inlen < offsetofend(struct mlx5_ib_create_qp_rss, flags))
2747 		return -EINVAL;
2748 
2749 	if (outlen < offsetofend(struct mlx5_ib_create_qp_resp, bfreg_index))
2750 		return -EINVAL;
2751 
2752 	ucmd = sizeof(struct mlx5_ib_create_qp_rss);
2753 	params->ucmd_size = ucmd;
2754 	if (inlen > ucmd && !ib_is_udata_cleared(udata, ucmd, inlen - ucmd))
2755 		return -EINVAL;
2756 
2757 	params->inlen = min(ucmd, inlen);
2758 out:
2759 	if (!params->inlen)
2760 		mlx5_ib_dbg(dev, "udata is too small\n");
2761 
2762 	return (params->inlen) ? 0 : -EINVAL;
2763 }
2764 
2765 static int create_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2766 		     struct mlx5_ib_qp *qp,
2767 		     struct mlx5_create_qp_params *params)
2768 {
2769 	int err;
2770 
2771 	if (params->is_rss_raw) {
2772 		err = create_rss_raw_qp_tir(dev, pd, qp, params);
2773 		goto out;
2774 	}
2775 
2776 	switch (qp->type) {
2777 	case MLX5_IB_QPT_DCT:
2778 		err = create_dct(dev, pd, qp, params);
2779 		break;
2780 	case IB_QPT_XRC_TGT:
2781 		err = create_xrc_tgt_qp(dev, qp, params);
2782 		break;
2783 	case IB_QPT_GSI:
2784 		err = mlx5_ib_create_gsi(pd, qp, params->attr);
2785 		break;
2786 	default:
2787 		if (params->udata)
2788 			err = create_user_qp(dev, pd, qp, params);
2789 		else
2790 			err = create_kernel_qp(dev, pd, qp, params);
2791 	}
2792 
2793 out:
2794 	if (err) {
2795 		mlx5_ib_err(dev, "Create QP type %d failed\n", qp->type);
2796 		return err;
2797 	}
2798 
2799 	if (is_qp0(qp->type))
2800 		qp->ibqp.qp_num = 0;
2801 	else if (is_qp1(qp->type))
2802 		qp->ibqp.qp_num = 1;
2803 	else
2804 		qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
2805 
2806 	mlx5_ib_dbg(dev,
2807 		"QP type %d, ib qpn 0x%X, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x, ece 0x%x\n",
2808 		qp->type, qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2809 		params->attr->recv_cq ? to_mcq(params->attr->recv_cq)->mcq.cqn :
2810 					-1,
2811 		params->attr->send_cq ? to_mcq(params->attr->send_cq)->mcq.cqn :
2812 					-1,
2813 		params->resp.ece_options);
2814 
2815 	return 0;
2816 }
2817 
2818 static int check_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2819 			 struct ib_qp_init_attr *attr)
2820 {
2821 	int ret = 0;
2822 
2823 	switch (qp->type) {
2824 	case MLX5_IB_QPT_DCT:
2825 		ret = (!attr->srq || !attr->recv_cq) ? -EINVAL : 0;
2826 		break;
2827 	case MLX5_IB_QPT_DCI:
2828 		ret = (attr->cap.max_recv_wr || attr->cap.max_recv_sge) ?
2829 			      -EINVAL :
2830 			      0;
2831 		break;
2832 	case IB_QPT_RAW_PACKET:
2833 		ret = (attr->rwq_ind_tbl && attr->send_cq) ? -EINVAL : 0;
2834 		break;
2835 	default:
2836 		break;
2837 	}
2838 
2839 	if (ret)
2840 		mlx5_ib_dbg(dev, "QP type %d has wrong attributes\n", qp->type);
2841 
2842 	return ret;
2843 }
2844 
2845 static int get_qp_uidx(struct mlx5_ib_qp *qp,
2846 		       struct mlx5_create_qp_params *params)
2847 {
2848 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
2849 	struct ib_udata *udata = params->udata;
2850 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2851 		udata, struct mlx5_ib_ucontext, ibucontext);
2852 
2853 	if (params->is_rss_raw)
2854 		return 0;
2855 
2856 	return get_qp_user_index(ucontext, ucmd, sizeof(*ucmd), &params->uidx);
2857 }
2858 
2859 static int mlx5_ib_destroy_dct(struct mlx5_ib_qp *mqp)
2860 {
2861 	struct mlx5_ib_dev *dev = to_mdev(mqp->ibqp.device);
2862 
2863 	if (mqp->state == IB_QPS_RTR) {
2864 		int err;
2865 
2866 		err = mlx5_core_destroy_dct(dev, &mqp->dct.mdct);
2867 		if (err) {
2868 			mlx5_ib_warn(dev, "failed to destroy DCT %d\n", err);
2869 			return err;
2870 		}
2871 	}
2872 
2873 	kfree(mqp->dct.in);
2874 	kfree(mqp);
2875 	return 0;
2876 }
2877 
2878 static int check_ucmd_data(struct mlx5_ib_dev *dev,
2879 			   struct mlx5_create_qp_params *params)
2880 {
2881 	struct ib_udata *udata = params->udata;
2882 	size_t size, last;
2883 	int ret;
2884 
2885 	if (params->is_rss_raw)
2886 		/*
2887 		 * These QPs don't have "reserved" field in their
2888 		 * create_qp input struct, so their data is always valid.
2889 		 */
2890 		last = sizeof(struct mlx5_ib_create_qp_rss);
2891 	else
2892 		last = offsetof(struct mlx5_ib_create_qp, reserved);
2893 
2894 	if (udata->inlen <= last)
2895 		return 0;
2896 
2897 	/*
2898 	 * User provides different create_qp structures based on the
2899 	 * flow and we need to know if he cleared memory after our
2900 	 * struct create_qp ends.
2901 	 */
2902 	size = udata->inlen - last;
2903 	ret = ib_is_udata_cleared(params->udata, last, size);
2904 	if (!ret)
2905 		mlx5_ib_dbg(
2906 			dev,
2907 			"udata is not cleared, inlen = %zu, ucmd = %zu, last = %zu, size = %zu\n",
2908 			udata->inlen, params->ucmd_size, last, size);
2909 	return ret ? 0 : -EINVAL;
2910 }
2911 
2912 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attr,
2913 				struct ib_udata *udata)
2914 {
2915 	struct mlx5_create_qp_params params = {};
2916 	struct mlx5_ib_dev *dev;
2917 	struct mlx5_ib_qp *qp;
2918 	enum ib_qp_type type;
2919 	int err;
2920 
2921 	dev = pd ? to_mdev(pd->device) :
2922 		   to_mdev(to_mxrcd(attr->xrcd)->ibxrcd.device);
2923 
2924 	err = check_qp_type(dev, attr, &type);
2925 	if (err)
2926 		return ERR_PTR(err);
2927 
2928 	err = check_valid_flow(dev, pd, attr, udata);
2929 	if (err)
2930 		return ERR_PTR(err);
2931 
2932 	params.udata = udata;
2933 	params.uidx = MLX5_IB_DEFAULT_UIDX;
2934 	params.attr = attr;
2935 	params.is_rss_raw = !!attr->rwq_ind_tbl;
2936 
2937 	if (udata) {
2938 		err = process_udata_size(dev, &params);
2939 		if (err)
2940 			return ERR_PTR(err);
2941 
2942 		err = check_ucmd_data(dev, &params);
2943 		if (err)
2944 			return ERR_PTR(err);
2945 
2946 		params.ucmd = kzalloc(params.ucmd_size, GFP_KERNEL);
2947 		if (!params.ucmd)
2948 			return ERR_PTR(-ENOMEM);
2949 
2950 		err = ib_copy_from_udata(params.ucmd, udata, params.inlen);
2951 		if (err)
2952 			goto free_ucmd;
2953 	}
2954 
2955 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2956 	if (!qp) {
2957 		err = -ENOMEM;
2958 		goto free_ucmd;
2959 	}
2960 
2961 	mutex_init(&qp->mutex);
2962 	qp->type = type;
2963 	if (udata) {
2964 		err = process_vendor_flags(dev, qp, params.ucmd, attr);
2965 		if (err)
2966 			goto free_qp;
2967 
2968 		err = get_qp_uidx(qp, &params);
2969 		if (err)
2970 			goto free_qp;
2971 	}
2972 	err = process_create_flags(dev, qp, attr);
2973 	if (err)
2974 		goto free_qp;
2975 
2976 	err = check_qp_attr(dev, qp, attr);
2977 	if (err)
2978 		goto free_qp;
2979 
2980 	err = create_qp(dev, pd, qp, &params);
2981 	if (err)
2982 		goto free_qp;
2983 
2984 	kfree(params.ucmd);
2985 	params.ucmd = NULL;
2986 
2987 	if (udata)
2988 		/*
2989 		 * It is safe to copy response for all user create QP flows,
2990 		 * including MLX5_IB_QPT_DCT, which doesn't need it.
2991 		 * In that case, resp will be filled with zeros.
2992 		 */
2993 		err = ib_copy_to_udata(udata, &params.resp, params.outlen);
2994 	if (err)
2995 		goto destroy_qp;
2996 
2997 	return &qp->ibqp;
2998 
2999 destroy_qp:
3000 	switch (qp->type) {
3001 	case MLX5_IB_QPT_DCT:
3002 		mlx5_ib_destroy_dct(qp);
3003 		break;
3004 	case IB_QPT_GSI:
3005 		mlx5_ib_destroy_gsi(qp);
3006 		break;
3007 	default:
3008 		/*
3009 		 * These lines below are temp solution till QP allocation
3010 		 * will be moved to be under IB/core responsiblity.
3011 		 */
3012 		qp->ibqp.send_cq = attr->send_cq;
3013 		qp->ibqp.recv_cq = attr->recv_cq;
3014 		qp->ibqp.pd = pd;
3015 		destroy_qp_common(dev, qp, udata);
3016 	}
3017 
3018 	qp = NULL;
3019 free_qp:
3020 	kfree(qp);
3021 free_ucmd:
3022 	kfree(params.ucmd);
3023 	return ERR_PTR(err);
3024 }
3025 
3026 int mlx5_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
3027 {
3028 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3029 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3030 
3031 	if (unlikely(qp->qp_type == IB_QPT_GSI))
3032 		return mlx5_ib_destroy_gsi(mqp);
3033 
3034 	if (mqp->type == MLX5_IB_QPT_DCT)
3035 		return mlx5_ib_destroy_dct(mqp);
3036 
3037 	destroy_qp_common(dev, mqp, udata);
3038 
3039 	kfree(mqp);
3040 
3041 	return 0;
3042 }
3043 
3044 static int set_qpc_atomic_flags(struct mlx5_ib_qp *qp,
3045 				const struct ib_qp_attr *attr, int attr_mask,
3046 				void *qpc)
3047 {
3048 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
3049 	u8 dest_rd_atomic;
3050 	u32 access_flags;
3051 
3052 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
3053 		dest_rd_atomic = attr->max_dest_rd_atomic;
3054 	else
3055 		dest_rd_atomic = qp->trans_qp.resp_depth;
3056 
3057 	if (attr_mask & IB_QP_ACCESS_FLAGS)
3058 		access_flags = attr->qp_access_flags;
3059 	else
3060 		access_flags = qp->trans_qp.atomic_rd_en;
3061 
3062 	if (!dest_rd_atomic)
3063 		access_flags &= IB_ACCESS_REMOTE_WRITE;
3064 
3065 	MLX5_SET(qpc, qpc, rre, !!(access_flags & IB_ACCESS_REMOTE_READ));
3066 
3067 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC) {
3068 		int atomic_mode;
3069 
3070 		atomic_mode = get_atomic_mode(dev, qp->ibqp.qp_type);
3071 		if (atomic_mode < 0)
3072 			return -EOPNOTSUPP;
3073 
3074 		MLX5_SET(qpc, qpc, rae, 1);
3075 		MLX5_SET(qpc, qpc, atomic_mode, atomic_mode);
3076 	}
3077 
3078 	MLX5_SET(qpc, qpc, rwe, !!(access_flags & IB_ACCESS_REMOTE_WRITE));
3079 	return 0;
3080 }
3081 
3082 enum {
3083 	MLX5_PATH_FLAG_FL	= 1 << 0,
3084 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
3085 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
3086 };
3087 
3088 static int ib_to_mlx5_rate_map(u8 rate)
3089 {
3090 	switch (rate) {
3091 	case IB_RATE_PORT_CURRENT:
3092 		return 0;
3093 	case IB_RATE_56_GBPS:
3094 		return 1;
3095 	case IB_RATE_25_GBPS:
3096 		return 2;
3097 	case IB_RATE_100_GBPS:
3098 		return 3;
3099 	case IB_RATE_200_GBPS:
3100 		return 4;
3101 	case IB_RATE_50_GBPS:
3102 		return 5;
3103 	default:
3104 		return rate + MLX5_STAT_RATE_OFFSET;
3105 	};
3106 
3107 	return 0;
3108 }
3109 
3110 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
3111 {
3112 	u32 stat_rate_support;
3113 
3114 	if (rate == IB_RATE_PORT_CURRENT)
3115 		return 0;
3116 
3117 	if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_600_GBPS)
3118 		return -EINVAL;
3119 
3120 	stat_rate_support = MLX5_CAP_GEN(dev->mdev, stat_rate_support);
3121 	while (rate != IB_RATE_PORT_CURRENT &&
3122 	       !(1 << ib_to_mlx5_rate_map(rate) & stat_rate_support))
3123 		--rate;
3124 
3125 	return ib_to_mlx5_rate_map(rate);
3126 }
3127 
3128 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
3129 				      struct mlx5_ib_sq *sq, u8 sl,
3130 				      struct ib_pd *pd)
3131 {
3132 	void *in;
3133 	void *tisc;
3134 	int inlen;
3135 	int err;
3136 
3137 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
3138 	in = kvzalloc(inlen, GFP_KERNEL);
3139 	if (!in)
3140 		return -ENOMEM;
3141 
3142 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
3143 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
3144 
3145 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
3146 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
3147 
3148 	err = mlx5_core_modify_tis(dev, sq->tisn, in);
3149 
3150 	kvfree(in);
3151 
3152 	return err;
3153 }
3154 
3155 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
3156 					 struct mlx5_ib_sq *sq, u8 tx_affinity,
3157 					 struct ib_pd *pd)
3158 {
3159 	void *in;
3160 	void *tisc;
3161 	int inlen;
3162 	int err;
3163 
3164 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
3165 	in = kvzalloc(inlen, GFP_KERNEL);
3166 	if (!in)
3167 		return -ENOMEM;
3168 
3169 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
3170 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
3171 
3172 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
3173 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
3174 
3175 	err = mlx5_core_modify_tis(dev, sq->tisn, in);
3176 
3177 	kvfree(in);
3178 
3179 	return err;
3180 }
3181 
3182 static void mlx5_set_path_udp_sport(void *path, const struct rdma_ah_attr *ah,
3183 				    u32 lqpn, u32 rqpn)
3184 
3185 {
3186 	u32 fl = ah->grh.flow_label;
3187 
3188 	if (!fl)
3189 		fl = rdma_calc_flow_label(lqpn, rqpn);
3190 
3191 	MLX5_SET(ads, path, udp_sport, rdma_flow_label_to_udp_sport(fl));
3192 }
3193 
3194 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3195 			 const struct rdma_ah_attr *ah, void *path, u8 port,
3196 			 int attr_mask, u32 path_flags,
3197 			 const struct ib_qp_attr *attr, bool alt)
3198 {
3199 	const struct ib_global_route *grh = rdma_ah_read_grh(ah);
3200 	int err;
3201 	enum ib_gid_type gid_type;
3202 	u8 ah_flags = rdma_ah_get_ah_flags(ah);
3203 	u8 sl = rdma_ah_get_sl(ah);
3204 
3205 	if (attr_mask & IB_QP_PKEY_INDEX)
3206 		MLX5_SET(ads, path, pkey_index,
3207 			 alt ? attr->alt_pkey_index : attr->pkey_index);
3208 
3209 	if (ah_flags & IB_AH_GRH) {
3210 		if (grh->sgid_index >=
3211 		    dev->mdev->port_caps[port - 1].gid_table_len) {
3212 			pr_err("sgid_index (%u) too large. max is %d\n",
3213 			       grh->sgid_index,
3214 			       dev->mdev->port_caps[port - 1].gid_table_len);
3215 			return -EINVAL;
3216 		}
3217 	}
3218 
3219 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
3220 		if (!(ah_flags & IB_AH_GRH))
3221 			return -EINVAL;
3222 
3223 		ether_addr_copy(MLX5_ADDR_OF(ads, path, rmac_47_32),
3224 				ah->roce.dmac);
3225 		if ((qp->ibqp.qp_type == IB_QPT_RC ||
3226 		     qp->ibqp.qp_type == IB_QPT_UC ||
3227 		     qp->ibqp.qp_type == IB_QPT_XRC_INI ||
3228 		     qp->ibqp.qp_type == IB_QPT_XRC_TGT) &&
3229 		    (grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) &&
3230 		    (attr_mask & IB_QP_DEST_QPN))
3231 			mlx5_set_path_udp_sport(path, ah,
3232 						qp->ibqp.qp_num,
3233 						attr->dest_qp_num);
3234 		MLX5_SET(ads, path, eth_prio, sl & 0x7);
3235 		gid_type = ah->grh.sgid_attr->gid_type;
3236 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
3237 			MLX5_SET(ads, path, dscp, grh->traffic_class >> 2);
3238 	} else {
3239 		MLX5_SET(ads, path, fl, !!(path_flags & MLX5_PATH_FLAG_FL));
3240 		MLX5_SET(ads, path, free_ar,
3241 			 !!(path_flags & MLX5_PATH_FLAG_FREE_AR));
3242 		MLX5_SET(ads, path, rlid, rdma_ah_get_dlid(ah));
3243 		MLX5_SET(ads, path, mlid, rdma_ah_get_path_bits(ah));
3244 		MLX5_SET(ads, path, grh, !!(ah_flags & IB_AH_GRH));
3245 		MLX5_SET(ads, path, sl, sl);
3246 	}
3247 
3248 	if (ah_flags & IB_AH_GRH) {
3249 		MLX5_SET(ads, path, src_addr_index, grh->sgid_index);
3250 		MLX5_SET(ads, path, hop_limit, grh->hop_limit);
3251 		MLX5_SET(ads, path, tclass, grh->traffic_class);
3252 		MLX5_SET(ads, path, flow_label, grh->flow_label);
3253 		memcpy(MLX5_ADDR_OF(ads, path, rgid_rip), grh->dgid.raw,
3254 		       sizeof(grh->dgid.raw));
3255 	}
3256 
3257 	err = ib_rate_to_mlx5(dev, rdma_ah_get_static_rate(ah));
3258 	if (err < 0)
3259 		return err;
3260 	MLX5_SET(ads, path, stat_rate, err);
3261 	MLX5_SET(ads, path, vhca_port_num, port);
3262 
3263 	if (attr_mask & IB_QP_TIMEOUT)
3264 		MLX5_SET(ads, path, ack_timeout,
3265 			 alt ? attr->alt_timeout : attr->timeout);
3266 
3267 	if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
3268 		return modify_raw_packet_eth_prio(dev->mdev,
3269 						  &qp->raw_packet_qp.sq,
3270 						  sl & 0xf, qp->ibqp.pd);
3271 
3272 	return 0;
3273 }
3274 
3275 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
3276 	[MLX5_QP_STATE_INIT] = {
3277 		[MLX5_QP_STATE_INIT] = {
3278 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3279 					  MLX5_QP_OPTPAR_RAE		|
3280 					  MLX5_QP_OPTPAR_RWE		|
3281 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3282 					  MLX5_QP_OPTPAR_PRI_PORT	|
3283 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3284 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3285 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3286 					  MLX5_QP_OPTPAR_PRI_PORT	|
3287 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3288 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3289 					  MLX5_QP_OPTPAR_Q_KEY		|
3290 					  MLX5_QP_OPTPAR_PRI_PORT,
3291 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3292 					  MLX5_QP_OPTPAR_RAE		|
3293 					  MLX5_QP_OPTPAR_RWE		|
3294 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3295 					  MLX5_QP_OPTPAR_PRI_PORT	|
3296 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3297 		},
3298 		[MLX5_QP_STATE_RTR] = {
3299 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3300 					  MLX5_QP_OPTPAR_RRE            |
3301 					  MLX5_QP_OPTPAR_RAE            |
3302 					  MLX5_QP_OPTPAR_RWE            |
3303 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3304 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3305 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3306 					  MLX5_QP_OPTPAR_RWE            |
3307 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3308 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3309 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
3310 					  MLX5_QP_OPTPAR_Q_KEY,
3311 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3312 					   MLX5_QP_OPTPAR_Q_KEY,
3313 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
3314 					  MLX5_QP_OPTPAR_RRE            |
3315 					  MLX5_QP_OPTPAR_RAE            |
3316 					  MLX5_QP_OPTPAR_RWE            |
3317 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3318 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3319 		},
3320 	},
3321 	[MLX5_QP_STATE_RTR] = {
3322 		[MLX5_QP_STATE_RTS] = {
3323 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3324 					  MLX5_QP_OPTPAR_RRE		|
3325 					  MLX5_QP_OPTPAR_RAE		|
3326 					  MLX5_QP_OPTPAR_RWE		|
3327 					  MLX5_QP_OPTPAR_PM_STATE	|
3328 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3329 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3330 					  MLX5_QP_OPTPAR_RWE		|
3331 					  MLX5_QP_OPTPAR_PM_STATE,
3332 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
3333 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3334 					  MLX5_QP_OPTPAR_RRE		|
3335 					  MLX5_QP_OPTPAR_RAE		|
3336 					  MLX5_QP_OPTPAR_RWE		|
3337 					  MLX5_QP_OPTPAR_PM_STATE	|
3338 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3339 		},
3340 	},
3341 	[MLX5_QP_STATE_RTS] = {
3342 		[MLX5_QP_STATE_RTS] = {
3343 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3344 					  MLX5_QP_OPTPAR_RAE		|
3345 					  MLX5_QP_OPTPAR_RWE		|
3346 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3347 					  MLX5_QP_OPTPAR_PM_STATE	|
3348 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3349 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3350 					  MLX5_QP_OPTPAR_PM_STATE	|
3351 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3352 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
3353 					  MLX5_QP_OPTPAR_SRQN		|
3354 					  MLX5_QP_OPTPAR_CQN_RCV,
3355 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3356 					  MLX5_QP_OPTPAR_RAE		|
3357 					  MLX5_QP_OPTPAR_RWE		|
3358 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3359 					  MLX5_QP_OPTPAR_PM_STATE	|
3360 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3361 		},
3362 	},
3363 	[MLX5_QP_STATE_SQER] = {
3364 		[MLX5_QP_STATE_RTS] = {
3365 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
3366 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
3367 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
3368 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3369 					   MLX5_QP_OPTPAR_RWE		|
3370 					   MLX5_QP_OPTPAR_RAE		|
3371 					   MLX5_QP_OPTPAR_RRE,
3372 			[MLX5_QP_ST_XRC]  = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3373 					   MLX5_QP_OPTPAR_RWE		|
3374 					   MLX5_QP_OPTPAR_RAE		|
3375 					   MLX5_QP_OPTPAR_RRE,
3376 		},
3377 	},
3378 };
3379 
3380 static int ib_nr_to_mlx5_nr(int ib_mask)
3381 {
3382 	switch (ib_mask) {
3383 	case IB_QP_STATE:
3384 		return 0;
3385 	case IB_QP_CUR_STATE:
3386 		return 0;
3387 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
3388 		return 0;
3389 	case IB_QP_ACCESS_FLAGS:
3390 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
3391 			MLX5_QP_OPTPAR_RAE;
3392 	case IB_QP_PKEY_INDEX:
3393 		return MLX5_QP_OPTPAR_PKEY_INDEX;
3394 	case IB_QP_PORT:
3395 		return MLX5_QP_OPTPAR_PRI_PORT;
3396 	case IB_QP_QKEY:
3397 		return MLX5_QP_OPTPAR_Q_KEY;
3398 	case IB_QP_AV:
3399 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
3400 			MLX5_QP_OPTPAR_PRI_PORT;
3401 	case IB_QP_PATH_MTU:
3402 		return 0;
3403 	case IB_QP_TIMEOUT:
3404 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
3405 	case IB_QP_RETRY_CNT:
3406 		return MLX5_QP_OPTPAR_RETRY_COUNT;
3407 	case IB_QP_RNR_RETRY:
3408 		return MLX5_QP_OPTPAR_RNR_RETRY;
3409 	case IB_QP_RQ_PSN:
3410 		return 0;
3411 	case IB_QP_MAX_QP_RD_ATOMIC:
3412 		return MLX5_QP_OPTPAR_SRA_MAX;
3413 	case IB_QP_ALT_PATH:
3414 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
3415 	case IB_QP_MIN_RNR_TIMER:
3416 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
3417 	case IB_QP_SQ_PSN:
3418 		return 0;
3419 	case IB_QP_MAX_DEST_RD_ATOMIC:
3420 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
3421 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
3422 	case IB_QP_PATH_MIG_STATE:
3423 		return MLX5_QP_OPTPAR_PM_STATE;
3424 	case IB_QP_CAP:
3425 		return 0;
3426 	case IB_QP_DEST_QPN:
3427 		return 0;
3428 	}
3429 	return 0;
3430 }
3431 
3432 static int ib_mask_to_mlx5_opt(int ib_mask)
3433 {
3434 	int result = 0;
3435 	int i;
3436 
3437 	for (i = 0; i < 8 * sizeof(int); i++) {
3438 		if ((1 << i) & ib_mask)
3439 			result |= ib_nr_to_mlx5_nr(1 << i);
3440 	}
3441 
3442 	return result;
3443 }
3444 
3445 static int modify_raw_packet_qp_rq(
3446 	struct mlx5_ib_dev *dev, struct mlx5_ib_rq *rq, int new_state,
3447 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3448 {
3449 	void *in;
3450 	void *rqc;
3451 	int inlen;
3452 	int err;
3453 
3454 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
3455 	in = kvzalloc(inlen, GFP_KERNEL);
3456 	if (!in)
3457 		return -ENOMEM;
3458 
3459 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
3460 	MLX5_SET(modify_rq_in, in, uid, to_mpd(pd)->uid);
3461 
3462 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
3463 	MLX5_SET(rqc, rqc, state, new_state);
3464 
3465 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
3466 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
3467 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
3468 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
3469 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
3470 		} else
3471 			dev_info_once(
3472 				&dev->ib_dev.dev,
3473 				"RAW PACKET QP counters are not supported on current FW\n");
3474 	}
3475 
3476 	err = mlx5_core_modify_rq(dev->mdev, rq->base.mqp.qpn, in);
3477 	if (err)
3478 		goto out;
3479 
3480 	rq->state = new_state;
3481 
3482 out:
3483 	kvfree(in);
3484 	return err;
3485 }
3486 
3487 static int modify_raw_packet_qp_sq(
3488 	struct mlx5_core_dev *dev, struct mlx5_ib_sq *sq, int new_state,
3489 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3490 {
3491 	struct mlx5_ib_qp *ibqp = sq->base.container_mibqp;
3492 	struct mlx5_rate_limit old_rl = ibqp->rl;
3493 	struct mlx5_rate_limit new_rl = old_rl;
3494 	bool new_rate_added = false;
3495 	u16 rl_index = 0;
3496 	void *in;
3497 	void *sqc;
3498 	int inlen;
3499 	int err;
3500 
3501 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
3502 	in = kvzalloc(inlen, GFP_KERNEL);
3503 	if (!in)
3504 		return -ENOMEM;
3505 
3506 	MLX5_SET(modify_sq_in, in, uid, to_mpd(pd)->uid);
3507 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
3508 
3509 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
3510 	MLX5_SET(sqc, sqc, state, new_state);
3511 
3512 	if (raw_qp_param->set_mask & MLX5_RAW_QP_RATE_LIMIT) {
3513 		if (new_state != MLX5_SQC_STATE_RDY)
3514 			pr_warn("%s: Rate limit can only be changed when SQ is moving to RDY\n",
3515 				__func__);
3516 		else
3517 			new_rl = raw_qp_param->rl;
3518 	}
3519 
3520 	if (!mlx5_rl_are_equal(&old_rl, &new_rl)) {
3521 		if (new_rl.rate) {
3522 			err = mlx5_rl_add_rate(dev, &rl_index, &new_rl);
3523 			if (err) {
3524 				pr_err("Failed configuring rate limit(err %d): \
3525 				       rate %u, max_burst_sz %u, typical_pkt_sz %u\n",
3526 				       err, new_rl.rate, new_rl.max_burst_sz,
3527 				       new_rl.typical_pkt_sz);
3528 
3529 				goto out;
3530 			}
3531 			new_rate_added = true;
3532 		}
3533 
3534 		MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
3535 		/* index 0 means no limit */
3536 		MLX5_SET(sqc, sqc, packet_pacing_rate_limit_index, rl_index);
3537 	}
3538 
3539 	err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in);
3540 	if (err) {
3541 		/* Remove new rate from table if failed */
3542 		if (new_rate_added)
3543 			mlx5_rl_remove_rate(dev, &new_rl);
3544 		goto out;
3545 	}
3546 
3547 	/* Only remove the old rate after new rate was set */
3548 	if ((old_rl.rate && !mlx5_rl_are_equal(&old_rl, &new_rl)) ||
3549 	    (new_state != MLX5_SQC_STATE_RDY)) {
3550 		mlx5_rl_remove_rate(dev, &old_rl);
3551 		if (new_state != MLX5_SQC_STATE_RDY)
3552 			memset(&new_rl, 0, sizeof(new_rl));
3553 	}
3554 
3555 	ibqp->rl = new_rl;
3556 	sq->state = new_state;
3557 
3558 out:
3559 	kvfree(in);
3560 	return err;
3561 }
3562 
3563 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3564 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
3565 				u8 tx_affinity)
3566 {
3567 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
3568 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
3569 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
3570 	int modify_rq = !!qp->rq.wqe_cnt;
3571 	int modify_sq = !!qp->sq.wqe_cnt;
3572 	int rq_state;
3573 	int sq_state;
3574 	int err;
3575 
3576 	switch (raw_qp_param->operation) {
3577 	case MLX5_CMD_OP_RST2INIT_QP:
3578 		rq_state = MLX5_RQC_STATE_RDY;
3579 		sq_state = MLX5_SQC_STATE_RST;
3580 		break;
3581 	case MLX5_CMD_OP_2ERR_QP:
3582 		rq_state = MLX5_RQC_STATE_ERR;
3583 		sq_state = MLX5_SQC_STATE_ERR;
3584 		break;
3585 	case MLX5_CMD_OP_2RST_QP:
3586 		rq_state = MLX5_RQC_STATE_RST;
3587 		sq_state = MLX5_SQC_STATE_RST;
3588 		break;
3589 	case MLX5_CMD_OP_RTR2RTS_QP:
3590 	case MLX5_CMD_OP_RTS2RTS_QP:
3591 		if (raw_qp_param->set_mask & ~MLX5_RAW_QP_RATE_LIMIT)
3592 			return -EINVAL;
3593 
3594 		modify_rq = 0;
3595 		sq_state = MLX5_SQC_STATE_RDY;
3596 		break;
3597 	case MLX5_CMD_OP_INIT2INIT_QP:
3598 	case MLX5_CMD_OP_INIT2RTR_QP:
3599 		if (raw_qp_param->set_mask)
3600 			return -EINVAL;
3601 		else
3602 			return 0;
3603 	default:
3604 		WARN_ON(1);
3605 		return -EINVAL;
3606 	}
3607 
3608 	if (modify_rq) {
3609 		err =  modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param,
3610 					       qp->ibqp.pd);
3611 		if (err)
3612 			return err;
3613 	}
3614 
3615 	if (modify_sq) {
3616 		struct mlx5_flow_handle *flow_rule;
3617 
3618 		if (tx_affinity) {
3619 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
3620 							    tx_affinity,
3621 							    qp->ibqp.pd);
3622 			if (err)
3623 				return err;
3624 		}
3625 
3626 		flow_rule = create_flow_rule_vport_sq(dev, sq,
3627 						      raw_qp_param->port);
3628 		if (IS_ERR(flow_rule))
3629 			return PTR_ERR(flow_rule);
3630 
3631 		err = modify_raw_packet_qp_sq(dev->mdev, sq, sq_state,
3632 					      raw_qp_param, qp->ibqp.pd);
3633 		if (err) {
3634 			if (flow_rule)
3635 				mlx5_del_flow_rules(flow_rule);
3636 			return err;
3637 		}
3638 
3639 		if (flow_rule) {
3640 			destroy_flow_rule_vport_sq(sq);
3641 			sq->flow_rule = flow_rule;
3642 		}
3643 
3644 		return err;
3645 	}
3646 
3647 	return 0;
3648 }
3649 
3650 static unsigned int get_tx_affinity_rr(struct mlx5_ib_dev *dev,
3651 				       struct ib_udata *udata)
3652 {
3653 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3654 		udata, struct mlx5_ib_ucontext, ibucontext);
3655 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3656 	atomic_t *tx_port_affinity;
3657 
3658 	if (ucontext)
3659 		tx_port_affinity = &ucontext->tx_port_affinity;
3660 	else
3661 		tx_port_affinity = &dev->port[port_num].roce.tx_port_affinity;
3662 
3663 	return (unsigned int)atomic_add_return(1, tx_port_affinity) %
3664 		MLX5_MAX_PORTS + 1;
3665 }
3666 
3667 static bool qp_supports_affinity(struct mlx5_ib_qp *qp)
3668 {
3669 	if ((qp->type == IB_QPT_RC) || (qp->type == IB_QPT_UD) ||
3670 	    (qp->type == IB_QPT_UC) || (qp->type == IB_QPT_RAW_PACKET) ||
3671 	    (qp->type == IB_QPT_XRC_INI) || (qp->type == IB_QPT_XRC_TGT) ||
3672 	    (qp->type == MLX5_IB_QPT_DCI))
3673 		return true;
3674 	return false;
3675 }
3676 
3677 static unsigned int get_tx_affinity(struct ib_qp *qp,
3678 				    const struct ib_qp_attr *attr,
3679 				    int attr_mask, u8 init,
3680 				    struct ib_udata *udata)
3681 {
3682 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3683 		udata, struct mlx5_ib_ucontext, ibucontext);
3684 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3685 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3686 	struct mlx5_ib_qp_base *qp_base;
3687 	unsigned int tx_affinity;
3688 
3689 	if (!(mlx5_ib_lag_should_assign_affinity(dev) &&
3690 	      qp_supports_affinity(mqp)))
3691 		return 0;
3692 
3693 	if (mqp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
3694 		tx_affinity = mqp->gsi_lag_port;
3695 	else if (init)
3696 		tx_affinity = get_tx_affinity_rr(dev, udata);
3697 	else if ((attr_mask & IB_QP_AV) && attr->xmit_slave)
3698 		tx_affinity =
3699 			mlx5_lag_get_slave_port(dev->mdev, attr->xmit_slave);
3700 	else
3701 		return 0;
3702 
3703 	qp_base = &mqp->trans_qp.base;
3704 	if (ucontext)
3705 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x ucontext %p\n",
3706 			    tx_affinity, qp_base->mqp.qpn, ucontext);
3707 	else
3708 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x\n",
3709 			    tx_affinity, qp_base->mqp.qpn);
3710 	return tx_affinity;
3711 }
3712 
3713 static int __mlx5_ib_qp_set_counter(struct ib_qp *qp,
3714 				    struct rdma_counter *counter)
3715 {
3716 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3717 	u32 in[MLX5_ST_SZ_DW(rts2rts_qp_in)] = {};
3718 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3719 	struct mlx5_ib_qp_base *base;
3720 	u32 set_id;
3721 	u32 *qpc;
3722 
3723 	if (counter)
3724 		set_id = counter->id;
3725 	else
3726 		set_id = mlx5_ib_get_counters_id(dev, mqp->port - 1);
3727 
3728 	base = &mqp->trans_qp.base;
3729 	MLX5_SET(rts2rts_qp_in, in, opcode, MLX5_CMD_OP_RTS2RTS_QP);
3730 	MLX5_SET(rts2rts_qp_in, in, qpn, base->mqp.qpn);
3731 	MLX5_SET(rts2rts_qp_in, in, uid, base->mqp.uid);
3732 	MLX5_SET(rts2rts_qp_in, in, opt_param_mask,
3733 		 MLX5_QP_OPTPAR_COUNTER_SET_ID);
3734 
3735 	qpc = MLX5_ADDR_OF(rts2rts_qp_in, in, qpc);
3736 	MLX5_SET(qpc, qpc, counter_set_id, set_id);
3737 	return mlx5_cmd_exec_in(dev->mdev, rts2rts_qp, in);
3738 }
3739 
3740 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
3741 			       const struct ib_qp_attr *attr, int attr_mask,
3742 			       enum ib_qp_state cur_state,
3743 			       enum ib_qp_state new_state,
3744 			       const struct mlx5_ib_modify_qp *ucmd,
3745 			       struct mlx5_ib_modify_qp_resp *resp,
3746 			       struct ib_udata *udata)
3747 {
3748 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
3749 		[MLX5_QP_STATE_RST] = {
3750 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3751 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3752 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
3753 		},
3754 		[MLX5_QP_STATE_INIT]  = {
3755 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3756 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3757 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
3758 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
3759 		},
3760 		[MLX5_QP_STATE_RTR]   = {
3761 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3762 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3763 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
3764 		},
3765 		[MLX5_QP_STATE_RTS]   = {
3766 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3767 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3768 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
3769 		},
3770 		[MLX5_QP_STATE_SQD] = {
3771 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3772 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3773 		},
3774 		[MLX5_QP_STATE_SQER] = {
3775 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3776 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3777 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
3778 		},
3779 		[MLX5_QP_STATE_ERR] = {
3780 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3781 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3782 		}
3783 	};
3784 
3785 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3786 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3787 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
3788 	struct mlx5_ib_cq *send_cq, *recv_cq;
3789 	struct mlx5_ib_pd *pd;
3790 	enum mlx5_qp_state mlx5_cur, mlx5_new;
3791 	void *qpc, *pri_path, *alt_path;
3792 	enum mlx5_qp_optpar optpar = 0;
3793 	u32 set_id = 0;
3794 	int mlx5_st;
3795 	int err;
3796 	u16 op;
3797 	u8 tx_affinity = 0;
3798 
3799 	mlx5_st = to_mlx5_st(qp->type);
3800 	if (mlx5_st < 0)
3801 		return -EINVAL;
3802 
3803 	qpc = kzalloc(MLX5_ST_SZ_BYTES(qpc), GFP_KERNEL);
3804 	if (!qpc)
3805 		return -ENOMEM;
3806 
3807 	pd = to_mpd(qp->ibqp.pd);
3808 	MLX5_SET(qpc, qpc, st, mlx5_st);
3809 
3810 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
3811 		MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
3812 	} else {
3813 		switch (attr->path_mig_state) {
3814 		case IB_MIG_MIGRATED:
3815 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
3816 			break;
3817 		case IB_MIG_REARM:
3818 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_REARM);
3819 			break;
3820 		case IB_MIG_ARMED:
3821 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_ARMED);
3822 			break;
3823 		}
3824 	}
3825 
3826 	tx_affinity = get_tx_affinity(ibqp, attr, attr_mask,
3827 				      cur_state == IB_QPS_RESET &&
3828 				      new_state == IB_QPS_INIT, udata);
3829 
3830 	MLX5_SET(qpc, qpc, lag_tx_port_affinity, tx_affinity);
3831 	if (tx_affinity && new_state == IB_QPS_RTR &&
3832 	    MLX5_CAP_GEN(dev->mdev, init2_lag_tx_port_affinity))
3833 		optpar |= MLX5_QP_OPTPAR_LAG_TX_AFF;
3834 
3835 	if (is_sqp(ibqp->qp_type)) {
3836 		MLX5_SET(qpc, qpc, mtu, IB_MTU_256);
3837 		MLX5_SET(qpc, qpc, log_msg_max, 8);
3838 	} else if ((ibqp->qp_type == IB_QPT_UD &&
3839 		    !(qp->flags & IB_QP_CREATE_SOURCE_QPN)) ||
3840 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
3841 		MLX5_SET(qpc, qpc, mtu, IB_MTU_4096);
3842 		MLX5_SET(qpc, qpc, log_msg_max, 12);
3843 	} else if (attr_mask & IB_QP_PATH_MTU) {
3844 		if (attr->path_mtu < IB_MTU_256 ||
3845 		    attr->path_mtu > IB_MTU_4096) {
3846 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
3847 			err = -EINVAL;
3848 			goto out;
3849 		}
3850 		MLX5_SET(qpc, qpc, mtu, attr->path_mtu);
3851 		MLX5_SET(qpc, qpc, log_msg_max,
3852 			 MLX5_CAP_GEN(dev->mdev, log_max_msg));
3853 	}
3854 
3855 	if (attr_mask & IB_QP_DEST_QPN)
3856 		MLX5_SET(qpc, qpc, remote_qpn, attr->dest_qp_num);
3857 
3858 	pri_path = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
3859 	alt_path = MLX5_ADDR_OF(qpc, qpc, secondary_address_path);
3860 
3861 	if (attr_mask & IB_QP_PKEY_INDEX)
3862 		MLX5_SET(ads, pri_path, pkey_index, attr->pkey_index);
3863 
3864 	/* todo implement counter_index functionality */
3865 
3866 	if (is_sqp(ibqp->qp_type))
3867 		MLX5_SET(ads, pri_path, vhca_port_num, qp->port);
3868 
3869 	if (attr_mask & IB_QP_PORT)
3870 		MLX5_SET(ads, pri_path, vhca_port_num, attr->port_num);
3871 
3872 	if (attr_mask & IB_QP_AV) {
3873 		err = mlx5_set_path(dev, qp, &attr->ah_attr, pri_path,
3874 				    attr_mask & IB_QP_PORT ? attr->port_num :
3875 							     qp->port,
3876 				    attr_mask, 0, attr, false);
3877 		if (err)
3878 			goto out;
3879 	}
3880 
3881 	if (attr_mask & IB_QP_TIMEOUT)
3882 		MLX5_SET(ads, pri_path, ack_timeout, attr->timeout);
3883 
3884 	if (attr_mask & IB_QP_ALT_PATH) {
3885 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr, alt_path,
3886 				    attr->alt_port_num,
3887 				    attr_mask | IB_QP_PKEY_INDEX |
3888 					    IB_QP_TIMEOUT,
3889 				    0, attr, true);
3890 		if (err)
3891 			goto out;
3892 	}
3893 
3894 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
3895 		&send_cq, &recv_cq);
3896 
3897 	MLX5_SET(qpc, qpc, pd, pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
3898 	if (send_cq)
3899 		MLX5_SET(qpc, qpc, cqn_snd, send_cq->mcq.cqn);
3900 	if (recv_cq)
3901 		MLX5_SET(qpc, qpc, cqn_rcv, recv_cq->mcq.cqn);
3902 
3903 	MLX5_SET(qpc, qpc, log_ack_req_freq, MLX5_IB_ACK_REQ_FREQ);
3904 
3905 	if (attr_mask & IB_QP_RNR_RETRY)
3906 		MLX5_SET(qpc, qpc, rnr_retry, attr->rnr_retry);
3907 
3908 	if (attr_mask & IB_QP_RETRY_CNT)
3909 		MLX5_SET(qpc, qpc, retry_count, attr->retry_cnt);
3910 
3911 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC && attr->max_rd_atomic)
3912 		MLX5_SET(qpc, qpc, log_sra_max, ilog2(attr->max_rd_atomic));
3913 
3914 	if (attr_mask & IB_QP_SQ_PSN)
3915 		MLX5_SET(qpc, qpc, next_send_psn, attr->sq_psn);
3916 
3917 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC && attr->max_dest_rd_atomic)
3918 		MLX5_SET(qpc, qpc, log_rra_max,
3919 			 ilog2(attr->max_dest_rd_atomic));
3920 
3921 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
3922 		err = set_qpc_atomic_flags(qp, attr, attr_mask, qpc);
3923 		if (err)
3924 			goto out;
3925 	}
3926 
3927 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
3928 		MLX5_SET(qpc, qpc, min_rnr_nak, attr->min_rnr_timer);
3929 
3930 	if (attr_mask & IB_QP_RQ_PSN)
3931 		MLX5_SET(qpc, qpc, next_rcv_psn, attr->rq_psn);
3932 
3933 	if (attr_mask & IB_QP_QKEY)
3934 		MLX5_SET(qpc, qpc, q_key, attr->qkey);
3935 
3936 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
3937 		MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
3938 
3939 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3940 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
3941 			       qp->port) - 1;
3942 
3943 		/* Underlay port should be used - index 0 function per port */
3944 		if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
3945 			port_num = 0;
3946 
3947 		if (ibqp->counter)
3948 			set_id = ibqp->counter->id;
3949 		else
3950 			set_id = mlx5_ib_get_counters_id(dev, port_num);
3951 		MLX5_SET(qpc, qpc, counter_set_id, set_id);
3952 	}
3953 
3954 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
3955 		MLX5_SET(qpc, qpc, rlky, 1);
3956 
3957 	if (qp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
3958 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
3959 
3960 	mlx5_cur = to_mlx5_state(cur_state);
3961 	mlx5_new = to_mlx5_state(new_state);
3962 
3963 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
3964 	    !optab[mlx5_cur][mlx5_new]) {
3965 		err = -EINVAL;
3966 		goto out;
3967 	}
3968 
3969 	op = optab[mlx5_cur][mlx5_new];
3970 	optpar |= ib_mask_to_mlx5_opt(attr_mask);
3971 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
3972 
3973 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
3974 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
3975 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
3976 
3977 		raw_qp_param.operation = op;
3978 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3979 			raw_qp_param.rq_q_ctr_id = set_id;
3980 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
3981 		}
3982 
3983 		if (attr_mask & IB_QP_PORT)
3984 			raw_qp_param.port = attr->port_num;
3985 
3986 		if (attr_mask & IB_QP_RATE_LIMIT) {
3987 			raw_qp_param.rl.rate = attr->rate_limit;
3988 
3989 			if (ucmd->burst_info.max_burst_sz) {
3990 				if (attr->rate_limit &&
3991 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_burst_bound)) {
3992 					raw_qp_param.rl.max_burst_sz =
3993 						ucmd->burst_info.max_burst_sz;
3994 				} else {
3995 					err = -EINVAL;
3996 					goto out;
3997 				}
3998 			}
3999 
4000 			if (ucmd->burst_info.typical_pkt_sz) {
4001 				if (attr->rate_limit &&
4002 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_typical_size)) {
4003 					raw_qp_param.rl.typical_pkt_sz =
4004 						ucmd->burst_info.typical_pkt_sz;
4005 				} else {
4006 					err = -EINVAL;
4007 					goto out;
4008 				}
4009 			}
4010 
4011 			raw_qp_param.set_mask |= MLX5_RAW_QP_RATE_LIMIT;
4012 		}
4013 
4014 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, tx_affinity);
4015 	} else {
4016 		if (udata) {
4017 			/* For the kernel flows, the resp will stay zero */
4018 			resp->ece_options =
4019 				MLX5_CAP_GEN(dev->mdev, ece_support) ?
4020 					ucmd->ece_options : 0;
4021 			resp->response_length = sizeof(*resp);
4022 		}
4023 		err = mlx5_core_qp_modify(dev, op, optpar, qpc, &base->mqp,
4024 					  &resp->ece_options);
4025 	}
4026 
4027 	if (err)
4028 		goto out;
4029 
4030 	qp->state = new_state;
4031 
4032 	if (attr_mask & IB_QP_ACCESS_FLAGS)
4033 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
4034 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
4035 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
4036 	if (attr_mask & IB_QP_PORT)
4037 		qp->port = attr->port_num;
4038 	if (attr_mask & IB_QP_ALT_PATH)
4039 		qp->trans_qp.alt_port = attr->alt_port_num;
4040 
4041 	/*
4042 	 * If we moved a kernel QP to RESET, clean up all old CQ
4043 	 * entries and reinitialize the QP.
4044 	 */
4045 	if (new_state == IB_QPS_RESET &&
4046 	    !ibqp->uobject && ibqp->qp_type != IB_QPT_XRC_TGT) {
4047 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
4048 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
4049 		if (send_cq != recv_cq)
4050 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
4051 
4052 		qp->rq.head = 0;
4053 		qp->rq.tail = 0;
4054 		qp->sq.head = 0;
4055 		qp->sq.tail = 0;
4056 		qp->sq.cur_post = 0;
4057 		if (qp->sq.wqe_cnt)
4058 			qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
4059 		qp->sq.last_poll = 0;
4060 		qp->db.db[MLX5_RCV_DBR] = 0;
4061 		qp->db.db[MLX5_SND_DBR] = 0;
4062 	}
4063 
4064 	if ((new_state == IB_QPS_RTS) && qp->counter_pending) {
4065 		err = __mlx5_ib_qp_set_counter(ibqp, ibqp->counter);
4066 		if (!err)
4067 			qp->counter_pending = 0;
4068 	}
4069 
4070 out:
4071 	kfree(qpc);
4072 	return err;
4073 }
4074 
4075 static inline bool is_valid_mask(int mask, int req, int opt)
4076 {
4077 	if ((mask & req) != req)
4078 		return false;
4079 
4080 	if (mask & ~(req | opt))
4081 		return false;
4082 
4083 	return true;
4084 }
4085 
4086 /* check valid transition for driver QP types
4087  * for now the only QP type that this function supports is DCI
4088  */
4089 static bool modify_dci_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state new_state,
4090 				enum ib_qp_attr_mask attr_mask)
4091 {
4092 	int req = IB_QP_STATE;
4093 	int opt = 0;
4094 
4095 	if (new_state == IB_QPS_RESET) {
4096 		return is_valid_mask(attr_mask, req, opt);
4097 	} else if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4098 		req |= IB_QP_PKEY_INDEX | IB_QP_PORT;
4099 		return is_valid_mask(attr_mask, req, opt);
4100 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
4101 		opt = IB_QP_PKEY_INDEX | IB_QP_PORT;
4102 		return is_valid_mask(attr_mask, req, opt);
4103 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
4104 		req |= IB_QP_PATH_MTU;
4105 		opt = IB_QP_PKEY_INDEX | IB_QP_AV;
4106 		return is_valid_mask(attr_mask, req, opt);
4107 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
4108 		req |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4109 		       IB_QP_MAX_QP_RD_ATOMIC | IB_QP_SQ_PSN;
4110 		opt = IB_QP_MIN_RNR_TIMER;
4111 		return is_valid_mask(attr_mask, req, opt);
4112 	} else if (cur_state == IB_QPS_RTS && new_state == IB_QPS_RTS) {
4113 		opt = IB_QP_MIN_RNR_TIMER;
4114 		return is_valid_mask(attr_mask, req, opt);
4115 	} else if (cur_state != IB_QPS_RESET && new_state == IB_QPS_ERR) {
4116 		return is_valid_mask(attr_mask, req, opt);
4117 	}
4118 	return false;
4119 }
4120 
4121 /* mlx5_ib_modify_dct: modify a DCT QP
4122  * valid transitions are:
4123  * RESET to INIT: must set access_flags, pkey_index and port
4124  * INIT  to RTR : must set min_rnr_timer, tclass, flow_label,
4125  *			   mtu, gid_index and hop_limit
4126  * Other transitions and attributes are illegal
4127  */
4128 static int mlx5_ib_modify_dct(struct ib_qp *ibqp, struct ib_qp_attr *attr,
4129 			      int attr_mask, struct mlx5_ib_modify_qp *ucmd,
4130 			      struct ib_udata *udata)
4131 {
4132 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4133 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4134 	enum ib_qp_state cur_state, new_state;
4135 	int required = IB_QP_STATE;
4136 	void *dctc;
4137 	int err;
4138 
4139 	if (!(attr_mask & IB_QP_STATE))
4140 		return -EINVAL;
4141 
4142 	cur_state = qp->state;
4143 	new_state = attr->qp_state;
4144 
4145 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
4146 	if (MLX5_CAP_GEN(dev->mdev, ece_support) && ucmd->ece_options)
4147 		/*
4148 		 * DCT doesn't initialize QP till modify command is executed,
4149 		 * so we need to overwrite previously set ECE field if user
4150 		 * provided any value except zero, which means not set/not
4151 		 * valid.
4152 		 */
4153 		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
4154 
4155 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4156 		u16 set_id;
4157 
4158 		required |= IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
4159 		if (!is_valid_mask(attr_mask, required, 0))
4160 			return -EINVAL;
4161 
4162 		if (attr->port_num == 0 ||
4163 		    attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)) {
4164 			mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4165 				    attr->port_num, dev->num_ports);
4166 			return -EINVAL;
4167 		}
4168 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
4169 			MLX5_SET(dctc, dctc, rre, 1);
4170 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
4171 			MLX5_SET(dctc, dctc, rwe, 1);
4172 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC) {
4173 			int atomic_mode;
4174 
4175 			atomic_mode = get_atomic_mode(dev, MLX5_IB_QPT_DCT);
4176 			if (atomic_mode < 0)
4177 				return -EOPNOTSUPP;
4178 
4179 			MLX5_SET(dctc, dctc, atomic_mode, atomic_mode);
4180 			MLX5_SET(dctc, dctc, rae, 1);
4181 		}
4182 		MLX5_SET(dctc, dctc, pkey_index, attr->pkey_index);
4183 		if (mlx5_lag_is_active(dev->mdev))
4184 			MLX5_SET(dctc, dctc, port,
4185 				 get_tx_affinity_rr(dev, udata));
4186 		else
4187 			MLX5_SET(dctc, dctc, port, attr->port_num);
4188 
4189 		set_id = mlx5_ib_get_counters_id(dev, attr->port_num - 1);
4190 		MLX5_SET(dctc, dctc, counter_set_id, set_id);
4191 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
4192 		struct mlx5_ib_modify_qp_resp resp = {};
4193 		u32 out[MLX5_ST_SZ_DW(create_dct_out)] = {};
4194 		u32 min_resp_len = offsetofend(typeof(resp), dctn);
4195 
4196 		if (udata->outlen < min_resp_len)
4197 			return -EINVAL;
4198 		/*
4199 		 * If we don't have enough space for the ECE options,
4200 		 * simply indicate it with resp.response_length.
4201 		 */
4202 		resp.response_length = (udata->outlen < sizeof(resp)) ?
4203 					       min_resp_len :
4204 					       sizeof(resp);
4205 
4206 		required |= IB_QP_MIN_RNR_TIMER | IB_QP_AV | IB_QP_PATH_MTU;
4207 		if (!is_valid_mask(attr_mask, required, 0))
4208 			return -EINVAL;
4209 		MLX5_SET(dctc, dctc, min_rnr_nak, attr->min_rnr_timer);
4210 		MLX5_SET(dctc, dctc, tclass, attr->ah_attr.grh.traffic_class);
4211 		MLX5_SET(dctc, dctc, flow_label, attr->ah_attr.grh.flow_label);
4212 		MLX5_SET(dctc, dctc, mtu, attr->path_mtu);
4213 		MLX5_SET(dctc, dctc, my_addr_index, attr->ah_attr.grh.sgid_index);
4214 		MLX5_SET(dctc, dctc, hop_limit, attr->ah_attr.grh.hop_limit);
4215 
4216 		err = mlx5_core_create_dct(dev, &qp->dct.mdct, qp->dct.in,
4217 					   MLX5_ST_SZ_BYTES(create_dct_in), out,
4218 					   sizeof(out));
4219 		if (err)
4220 			return err;
4221 		resp.dctn = qp->dct.mdct.mqp.qpn;
4222 		if (MLX5_CAP_GEN(dev->mdev, ece_support))
4223 			resp.ece_options = MLX5_GET(create_dct_out, out, ece);
4224 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4225 		if (err) {
4226 			mlx5_core_destroy_dct(dev, &qp->dct.mdct);
4227 			return err;
4228 		}
4229 	} else {
4230 		mlx5_ib_warn(dev, "Modify DCT: Invalid transition from %d to %d\n", cur_state, new_state);
4231 		return -EINVAL;
4232 	}
4233 
4234 	qp->state = new_state;
4235 	return 0;
4236 }
4237 
4238 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
4239 		      int attr_mask, struct ib_udata *udata)
4240 {
4241 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4242 	struct mlx5_ib_modify_qp_resp resp = {};
4243 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4244 	struct mlx5_ib_modify_qp ucmd = {};
4245 	enum ib_qp_type qp_type;
4246 	enum ib_qp_state cur_state, new_state;
4247 	int err = -EINVAL;
4248 	int port;
4249 
4250 	if (ibqp->rwq_ind_tbl)
4251 		return -ENOSYS;
4252 
4253 	if (udata && udata->inlen) {
4254 		if (udata->inlen < offsetofend(typeof(ucmd), ece_options))
4255 			return -EINVAL;
4256 
4257 		if (udata->inlen > sizeof(ucmd) &&
4258 		    !ib_is_udata_cleared(udata, sizeof(ucmd),
4259 					 udata->inlen - sizeof(ucmd)))
4260 			return -EOPNOTSUPP;
4261 
4262 		if (ib_copy_from_udata(&ucmd, udata,
4263 				       min(udata->inlen, sizeof(ucmd))))
4264 			return -EFAULT;
4265 
4266 		if (ucmd.comp_mask ||
4267 		    memchr_inv(&ucmd.burst_info.reserved, 0,
4268 			       sizeof(ucmd.burst_info.reserved)))
4269 			return -EOPNOTSUPP;
4270 
4271 	}
4272 
4273 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4274 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
4275 
4276 	qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ? IB_QPT_GSI :
4277 								    qp->type;
4278 
4279 	if (qp_type == MLX5_IB_QPT_DCT)
4280 		return mlx5_ib_modify_dct(ibqp, attr, attr_mask, &ucmd, udata);
4281 
4282 	mutex_lock(&qp->mutex);
4283 
4284 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
4285 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
4286 
4287 	if (!(cur_state == new_state && cur_state == IB_QPS_RESET)) {
4288 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
4289 	}
4290 
4291 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4292 		if (attr_mask & ~(IB_QP_STATE | IB_QP_CUR_STATE)) {
4293 			mlx5_ib_dbg(dev, "invalid attr_mask 0x%x when underlay QP is used\n",
4294 				    attr_mask);
4295 			goto out;
4296 		}
4297 	} else if (qp_type != MLX5_IB_QPT_REG_UMR &&
4298 		   qp_type != MLX5_IB_QPT_DCI &&
4299 		   !ib_modify_qp_is_ok(cur_state, new_state, qp_type,
4300 				       attr_mask)) {
4301 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4302 			    cur_state, new_state, ibqp->qp_type, attr_mask);
4303 		goto out;
4304 	} else if (qp_type == MLX5_IB_QPT_DCI &&
4305 		   !modify_dci_qp_is_ok(cur_state, new_state, attr_mask)) {
4306 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4307 			    cur_state, new_state, qp_type, attr_mask);
4308 		goto out;
4309 	}
4310 
4311 	if ((attr_mask & IB_QP_PORT) &&
4312 	    (attr->port_num == 0 ||
4313 	     attr->port_num > dev->num_ports)) {
4314 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4315 			    attr->port_num, dev->num_ports);
4316 		goto out;
4317 	}
4318 
4319 	if (attr_mask & IB_QP_PKEY_INDEX) {
4320 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
4321 		if (attr->pkey_index >=
4322 		    dev->mdev->port_caps[port - 1].pkey_table_len) {
4323 			mlx5_ib_dbg(dev, "invalid pkey index %d\n",
4324 				    attr->pkey_index);
4325 			goto out;
4326 		}
4327 	}
4328 
4329 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
4330 	    attr->max_rd_atomic >
4331 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
4332 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
4333 			    attr->max_rd_atomic);
4334 		goto out;
4335 	}
4336 
4337 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
4338 	    attr->max_dest_rd_atomic >
4339 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
4340 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
4341 			    attr->max_dest_rd_atomic);
4342 		goto out;
4343 	}
4344 
4345 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
4346 		err = 0;
4347 		goto out;
4348 	}
4349 
4350 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state,
4351 				  new_state, &ucmd, &resp, udata);
4352 
4353 	/* resp.response_length is set in ECE supported flows only */
4354 	if (!err && resp.response_length &&
4355 	    udata->outlen >= resp.response_length)
4356 		/* Return -EFAULT to the user and expect him to destroy QP. */
4357 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4358 
4359 out:
4360 	mutex_unlock(&qp->mutex);
4361 	return err;
4362 }
4363 
4364 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4365 {
4366 	switch (mlx5_state) {
4367 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
4368 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
4369 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
4370 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
4371 	case MLX5_QP_STATE_SQ_DRAINING:
4372 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
4373 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
4374 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
4375 	default:		     return -1;
4376 	}
4377 }
4378 
4379 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4380 {
4381 	switch (mlx5_mig_state) {
4382 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
4383 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
4384 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4385 	default: return -1;
4386 	}
4387 }
4388 
4389 static void to_rdma_ah_attr(struct mlx5_ib_dev *ibdev,
4390 			    struct rdma_ah_attr *ah_attr, void *path)
4391 {
4392 	int port = MLX5_GET(ads, path, vhca_port_num);
4393 	int static_rate;
4394 
4395 	memset(ah_attr, 0, sizeof(*ah_attr));
4396 
4397 	if (!port || port > ibdev->num_ports)
4398 		return;
4399 
4400 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port);
4401 
4402 	rdma_ah_set_port_num(ah_attr, port);
4403 	rdma_ah_set_sl(ah_attr, MLX5_GET(ads, path, sl));
4404 
4405 	rdma_ah_set_dlid(ah_attr, MLX5_GET(ads, path, rlid));
4406 	rdma_ah_set_path_bits(ah_attr, MLX5_GET(ads, path, mlid));
4407 
4408 	static_rate = MLX5_GET(ads, path, stat_rate);
4409 	rdma_ah_set_static_rate(ah_attr, static_rate ? static_rate - 5 : 0);
4410 	if (MLX5_GET(ads, path, grh) ||
4411 	    ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
4412 		rdma_ah_set_grh(ah_attr, NULL, MLX5_GET(ads, path, flow_label),
4413 				MLX5_GET(ads, path, src_addr_index),
4414 				MLX5_GET(ads, path, hop_limit),
4415 				MLX5_GET(ads, path, tclass));
4416 		rdma_ah_set_dgid_raw(ah_attr, MLX5_ADDR_OF(ads, path, rgid_rip));
4417 	}
4418 }
4419 
4420 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4421 					struct mlx5_ib_sq *sq,
4422 					u8 *sq_state)
4423 {
4424 	int err;
4425 
4426 	err = mlx5_core_query_sq_state(dev->mdev, sq->base.mqp.qpn, sq_state);
4427 	if (err)
4428 		goto out;
4429 	sq->state = *sq_state;
4430 
4431 out:
4432 	return err;
4433 }
4434 
4435 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4436 					struct mlx5_ib_rq *rq,
4437 					u8 *rq_state)
4438 {
4439 	void *out;
4440 	void *rqc;
4441 	int inlen;
4442 	int err;
4443 
4444 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4445 	out = kvzalloc(inlen, GFP_KERNEL);
4446 	if (!out)
4447 		return -ENOMEM;
4448 
4449 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4450 	if (err)
4451 		goto out;
4452 
4453 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4454 	*rq_state = MLX5_GET(rqc, rqc, state);
4455 	rq->state = *rq_state;
4456 
4457 out:
4458 	kvfree(out);
4459 	return err;
4460 }
4461 
4462 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4463 				  struct mlx5_ib_qp *qp, u8 *qp_state)
4464 {
4465 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4466 		[MLX5_RQC_STATE_RST] = {
4467 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
4468 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4469 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
4470 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
4471 		},
4472 		[MLX5_RQC_STATE_RDY] = {
4473 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE,
4474 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4475 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
4476 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
4477 		},
4478 		[MLX5_RQC_STATE_ERR] = {
4479 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
4480 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4481 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
4482 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
4483 		},
4484 		[MLX5_RQ_STATE_NA] = {
4485 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE,
4486 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4487 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
4488 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
4489 		},
4490 	};
4491 
4492 	*qp_state = sqrq_trans[rq_state][sq_state];
4493 
4494 	if (*qp_state == MLX5_QP_STATE_BAD) {
4495 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4496 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4497 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4498 		return -EINVAL;
4499 	}
4500 
4501 	if (*qp_state == MLX5_QP_STATE)
4502 		*qp_state = qp->state;
4503 
4504 	return 0;
4505 }
4506 
4507 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4508 				     struct mlx5_ib_qp *qp,
4509 				     u8 *raw_packet_qp_state)
4510 {
4511 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4512 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4513 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4514 	int err;
4515 	u8 sq_state = MLX5_SQ_STATE_NA;
4516 	u8 rq_state = MLX5_RQ_STATE_NA;
4517 
4518 	if (qp->sq.wqe_cnt) {
4519 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4520 		if (err)
4521 			return err;
4522 	}
4523 
4524 	if (qp->rq.wqe_cnt) {
4525 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4526 		if (err)
4527 			return err;
4528 	}
4529 
4530 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4531 				      raw_packet_qp_state);
4532 }
4533 
4534 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4535 			 struct ib_qp_attr *qp_attr)
4536 {
4537 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
4538 	void *qpc, *pri_path, *alt_path;
4539 	u32 *outb;
4540 	int err;
4541 
4542 	outb = kzalloc(outlen, GFP_KERNEL);
4543 	if (!outb)
4544 		return -ENOMEM;
4545 
4546 	err = mlx5_core_qp_query(dev, &qp->trans_qp.base.mqp, outb, outlen);
4547 	if (err)
4548 		goto out;
4549 
4550 	qpc = MLX5_ADDR_OF(query_qp_out, outb, qpc);
4551 
4552 	qp->state = to_ib_qp_state(MLX5_GET(qpc, qpc, state));
4553 	if (MLX5_GET(qpc, qpc, state) == MLX5_QP_STATE_SQ_DRAINING)
4554 		qp_attr->sq_draining = 1;
4555 
4556 	qp_attr->path_mtu = MLX5_GET(qpc, qpc, mtu);
4557 	qp_attr->path_mig_state = to_ib_mig_state(MLX5_GET(qpc, qpc, pm_state));
4558 	qp_attr->qkey = MLX5_GET(qpc, qpc, q_key);
4559 	qp_attr->rq_psn = MLX5_GET(qpc, qpc, next_rcv_psn);
4560 	qp_attr->sq_psn = MLX5_GET(qpc, qpc, next_send_psn);
4561 	qp_attr->dest_qp_num = MLX5_GET(qpc, qpc, remote_qpn);
4562 
4563 	if (MLX5_GET(qpc, qpc, rre))
4564 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ;
4565 	if (MLX5_GET(qpc, qpc, rwe))
4566 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_WRITE;
4567 	if (MLX5_GET(qpc, qpc, rae))
4568 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_ATOMIC;
4569 
4570 	qp_attr->max_rd_atomic = 1 << MLX5_GET(qpc, qpc, log_sra_max);
4571 	qp_attr->max_dest_rd_atomic = 1 << MLX5_GET(qpc, qpc, log_rra_max);
4572 	qp_attr->min_rnr_timer = MLX5_GET(qpc, qpc, min_rnr_nak);
4573 	qp_attr->retry_cnt = MLX5_GET(qpc, qpc, retry_count);
4574 	qp_attr->rnr_retry = MLX5_GET(qpc, qpc, rnr_retry);
4575 
4576 	pri_path = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
4577 	alt_path = MLX5_ADDR_OF(qpc, qpc, secondary_address_path);
4578 
4579 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4580 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, pri_path);
4581 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, alt_path);
4582 		qp_attr->alt_pkey_index = MLX5_GET(ads, alt_path, pkey_index);
4583 		qp_attr->alt_port_num = MLX5_GET(ads, alt_path, vhca_port_num);
4584 	}
4585 
4586 	qp_attr->pkey_index = MLX5_GET(ads, pri_path, pkey_index);
4587 	qp_attr->port_num = MLX5_GET(ads, pri_path, vhca_port_num);
4588 	qp_attr->timeout = MLX5_GET(ads, pri_path, ack_timeout);
4589 	qp_attr->alt_timeout = MLX5_GET(ads, alt_path, ack_timeout);
4590 
4591 out:
4592 	kfree(outb);
4593 	return err;
4594 }
4595 
4596 static int mlx5_ib_dct_query_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *mqp,
4597 				struct ib_qp_attr *qp_attr, int qp_attr_mask,
4598 				struct ib_qp_init_attr *qp_init_attr)
4599 {
4600 	struct mlx5_core_dct	*dct = &mqp->dct.mdct;
4601 	u32 *out;
4602 	u32 access_flags = 0;
4603 	int outlen = MLX5_ST_SZ_BYTES(query_dct_out);
4604 	void *dctc;
4605 	int err;
4606 	int supported_mask = IB_QP_STATE |
4607 			     IB_QP_ACCESS_FLAGS |
4608 			     IB_QP_PORT |
4609 			     IB_QP_MIN_RNR_TIMER |
4610 			     IB_QP_AV |
4611 			     IB_QP_PATH_MTU |
4612 			     IB_QP_PKEY_INDEX;
4613 
4614 	if (qp_attr_mask & ~supported_mask)
4615 		return -EINVAL;
4616 	if (mqp->state != IB_QPS_RTR)
4617 		return -EINVAL;
4618 
4619 	out = kzalloc(outlen, GFP_KERNEL);
4620 	if (!out)
4621 		return -ENOMEM;
4622 
4623 	err = mlx5_core_dct_query(dev, dct, out, outlen);
4624 	if (err)
4625 		goto out;
4626 
4627 	dctc = MLX5_ADDR_OF(query_dct_out, out, dct_context_entry);
4628 
4629 	if (qp_attr_mask & IB_QP_STATE)
4630 		qp_attr->qp_state = IB_QPS_RTR;
4631 
4632 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
4633 		if (MLX5_GET(dctc, dctc, rre))
4634 			access_flags |= IB_ACCESS_REMOTE_READ;
4635 		if (MLX5_GET(dctc, dctc, rwe))
4636 			access_flags |= IB_ACCESS_REMOTE_WRITE;
4637 		if (MLX5_GET(dctc, dctc, rae))
4638 			access_flags |= IB_ACCESS_REMOTE_ATOMIC;
4639 		qp_attr->qp_access_flags = access_flags;
4640 	}
4641 
4642 	if (qp_attr_mask & IB_QP_PORT)
4643 		qp_attr->port_num = MLX5_GET(dctc, dctc, port);
4644 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER)
4645 		qp_attr->min_rnr_timer = MLX5_GET(dctc, dctc, min_rnr_nak);
4646 	if (qp_attr_mask & IB_QP_AV) {
4647 		qp_attr->ah_attr.grh.traffic_class = MLX5_GET(dctc, dctc, tclass);
4648 		qp_attr->ah_attr.grh.flow_label = MLX5_GET(dctc, dctc, flow_label);
4649 		qp_attr->ah_attr.grh.sgid_index = MLX5_GET(dctc, dctc, my_addr_index);
4650 		qp_attr->ah_attr.grh.hop_limit = MLX5_GET(dctc, dctc, hop_limit);
4651 	}
4652 	if (qp_attr_mask & IB_QP_PATH_MTU)
4653 		qp_attr->path_mtu = MLX5_GET(dctc, dctc, mtu);
4654 	if (qp_attr_mask & IB_QP_PKEY_INDEX)
4655 		qp_attr->pkey_index = MLX5_GET(dctc, dctc, pkey_index);
4656 out:
4657 	kfree(out);
4658 	return err;
4659 }
4660 
4661 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4662 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4663 {
4664 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4665 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4666 	int err = 0;
4667 	u8 raw_packet_qp_state;
4668 
4669 	if (ibqp->rwq_ind_tbl)
4670 		return -ENOSYS;
4671 
4672 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4673 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4674 					    qp_init_attr);
4675 
4676 	/* Not all of output fields are applicable, make sure to zero them */
4677 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
4678 	memset(qp_attr, 0, sizeof(*qp_attr));
4679 
4680 	if (unlikely(qp->type == MLX5_IB_QPT_DCT))
4681 		return mlx5_ib_dct_query_qp(dev, qp, qp_attr,
4682 					    qp_attr_mask, qp_init_attr);
4683 
4684 	mutex_lock(&qp->mutex);
4685 
4686 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
4687 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4688 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4689 		if (err)
4690 			goto out;
4691 		qp->state = raw_packet_qp_state;
4692 		qp_attr->port_num = 1;
4693 	} else {
4694 		err = query_qp_attr(dev, qp, qp_attr);
4695 		if (err)
4696 			goto out;
4697 	}
4698 
4699 	qp_attr->qp_state	     = qp->state;
4700 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4701 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4702 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4703 
4704 	if (!ibqp->uobject) {
4705 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
4706 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4707 		qp_init_attr->qp_context = ibqp->qp_context;
4708 	} else {
4709 		qp_attr->cap.max_send_wr  = 0;
4710 		qp_attr->cap.max_send_sge = 0;
4711 	}
4712 
4713 	qp_init_attr->qp_type = ibqp->qp_type;
4714 	qp_init_attr->recv_cq = ibqp->recv_cq;
4715 	qp_init_attr->send_cq = ibqp->send_cq;
4716 	qp_init_attr->srq = ibqp->srq;
4717 	qp_attr->cap.max_inline_data = qp->max_inline_data;
4718 
4719 	qp_init_attr->cap	     = qp_attr->cap;
4720 
4721 	qp_init_attr->create_flags = qp->flags;
4722 
4723 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
4724 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4725 
4726 out:
4727 	mutex_unlock(&qp->mutex);
4728 	return err;
4729 }
4730 
4731 int mlx5_ib_alloc_xrcd(struct ib_xrcd *ibxrcd, struct ib_udata *udata)
4732 {
4733 	struct mlx5_ib_dev *dev = to_mdev(ibxrcd->device);
4734 	struct mlx5_ib_xrcd *xrcd = to_mxrcd(ibxrcd);
4735 
4736 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
4737 		return -EOPNOTSUPP;
4738 
4739 	return mlx5_cmd_xrcd_alloc(dev->mdev, &xrcd->xrcdn, 0);
4740 }
4741 
4742 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
4743 {
4744 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
4745 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
4746 
4747 	return mlx5_cmd_xrcd_dealloc(dev->mdev, xrcdn, 0);
4748 }
4749 
4750 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
4751 {
4752 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
4753 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
4754 	struct ib_event event;
4755 
4756 	if (rwq->ibwq.event_handler) {
4757 		event.device     = rwq->ibwq.device;
4758 		event.element.wq = &rwq->ibwq;
4759 		switch (type) {
4760 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
4761 			event.event = IB_EVENT_WQ_FATAL;
4762 			break;
4763 		default:
4764 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
4765 			return;
4766 		}
4767 
4768 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
4769 	}
4770 }
4771 
4772 static int set_delay_drop(struct mlx5_ib_dev *dev)
4773 {
4774 	int err = 0;
4775 
4776 	mutex_lock(&dev->delay_drop.lock);
4777 	if (dev->delay_drop.activate)
4778 		goto out;
4779 
4780 	err = mlx5_core_set_delay_drop(dev, dev->delay_drop.timeout);
4781 	if (err)
4782 		goto out;
4783 
4784 	dev->delay_drop.activate = true;
4785 out:
4786 	mutex_unlock(&dev->delay_drop.lock);
4787 
4788 	if (!err)
4789 		atomic_inc(&dev->delay_drop.rqs_cnt);
4790 	return err;
4791 }
4792 
4793 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
4794 		      struct ib_wq_init_attr *init_attr)
4795 {
4796 	struct mlx5_ib_dev *dev;
4797 	int has_net_offloads;
4798 	__be64 *rq_pas0;
4799 	void *in;
4800 	void *rqc;
4801 	void *wq;
4802 	int inlen;
4803 	int err;
4804 
4805 	dev = to_mdev(pd->device);
4806 
4807 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
4808 	in = kvzalloc(inlen, GFP_KERNEL);
4809 	if (!in)
4810 		return -ENOMEM;
4811 
4812 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
4813 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
4814 	MLX5_SET(rqc,  rqc, mem_rq_type,
4815 		 MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
4816 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
4817 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
4818 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
4819 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
4820 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
4821 	MLX5_SET(wq, wq, wq_type,
4822 		 rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ ?
4823 		 MLX5_WQ_TYPE_CYCLIC_STRIDING_RQ : MLX5_WQ_TYPE_CYCLIC);
4824 	if (init_attr->create_flags & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
4825 		if (!MLX5_CAP_GEN(dev->mdev, end_pad)) {
4826 			mlx5_ib_dbg(dev, "Scatter end padding is not supported\n");
4827 			err = -EOPNOTSUPP;
4828 			goto out;
4829 		} else {
4830 			MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
4831 		}
4832 	}
4833 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
4834 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ) {
4835 		/*
4836 		 * In Firmware number of strides in each WQE is:
4837 		 *   "512 * 2^single_wqe_log_num_of_strides"
4838 		 * Values 3 to 8 are accepted as 10 to 15, 9 to 18 are
4839 		 * accepted as 0 to 9
4840 		 */
4841 		static const u8 fw_map[] = { 10, 11, 12, 13, 14, 15, 0, 1,
4842 					     2,  3,  4,  5,  6,  7,  8, 9 };
4843 		MLX5_SET(wq, wq, two_byte_shift_en, rwq->two_byte_shift_en);
4844 		MLX5_SET(wq, wq, log_wqe_stride_size,
4845 			 rwq->single_stride_log_num_of_bytes -
4846 			 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES);
4847 		MLX5_SET(wq, wq, log_wqe_num_of_strides,
4848 			 fw_map[rwq->log_num_strides -
4849 				MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES]);
4850 	}
4851 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
4852 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
4853 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
4854 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
4855 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
4856 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
4857 	has_net_offloads = MLX5_CAP_GEN(dev->mdev, eth_net_offloads);
4858 	if (init_attr->create_flags & IB_WQ_FLAGS_CVLAN_STRIPPING) {
4859 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
4860 			mlx5_ib_dbg(dev, "VLAN offloads are not supported\n");
4861 			err = -EOPNOTSUPP;
4862 			goto out;
4863 		}
4864 	} else {
4865 		MLX5_SET(rqc, rqc, vsd, 1);
4866 	}
4867 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS) {
4868 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, scatter_fcs))) {
4869 			mlx5_ib_dbg(dev, "Scatter FCS is not supported\n");
4870 			err = -EOPNOTSUPP;
4871 			goto out;
4872 		}
4873 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
4874 	}
4875 	if (init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
4876 		if (!(dev->ib_dev.attrs.raw_packet_caps &
4877 		      IB_RAW_PACKET_CAP_DELAY_DROP)) {
4878 			mlx5_ib_dbg(dev, "Delay drop is not supported\n");
4879 			err = -EOPNOTSUPP;
4880 			goto out;
4881 		}
4882 		MLX5_SET(rqc, rqc, delay_drop_en, 1);
4883 	}
4884 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
4885 	mlx5_ib_populate_pas(dev, rwq->umem, rwq->page_shift, rq_pas0, 0);
4886 	err = mlx5_core_create_rq_tracked(dev, in, inlen, &rwq->core_qp);
4887 	if (!err && init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
4888 		err = set_delay_drop(dev);
4889 		if (err) {
4890 			mlx5_ib_warn(dev, "Failed to enable delay drop err=%d\n",
4891 				     err);
4892 			mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
4893 		} else {
4894 			rwq->create_flags |= MLX5_IB_WQ_FLAGS_DELAY_DROP;
4895 		}
4896 	}
4897 out:
4898 	kvfree(in);
4899 	return err;
4900 }
4901 
4902 static int set_user_rq_size(struct mlx5_ib_dev *dev,
4903 			    struct ib_wq_init_attr *wq_init_attr,
4904 			    struct mlx5_ib_create_wq *ucmd,
4905 			    struct mlx5_ib_rwq *rwq)
4906 {
4907 	/* Sanity check RQ size before proceeding */
4908 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
4909 		return -EINVAL;
4910 
4911 	if (!ucmd->rq_wqe_count)
4912 		return -EINVAL;
4913 
4914 	rwq->wqe_count = ucmd->rq_wqe_count;
4915 	rwq->wqe_shift = ucmd->rq_wqe_shift;
4916 	if (check_shl_overflow(rwq->wqe_count, rwq->wqe_shift, &rwq->buf_size))
4917 		return -EINVAL;
4918 
4919 	rwq->log_rq_stride = rwq->wqe_shift;
4920 	rwq->log_rq_size = ilog2(rwq->wqe_count);
4921 	return 0;
4922 }
4923 
4924 static bool log_of_strides_valid(struct mlx5_ib_dev *dev, u32 log_num_strides)
4925 {
4926 	if ((log_num_strides > MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES) ||
4927 	    (log_num_strides < MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
4928 		return false;
4929 
4930 	if (!MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) &&
4931 	    (log_num_strides < MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
4932 		return false;
4933 
4934 	return true;
4935 }
4936 
4937 static int prepare_user_rq(struct ib_pd *pd,
4938 			   struct ib_wq_init_attr *init_attr,
4939 			   struct ib_udata *udata,
4940 			   struct mlx5_ib_rwq *rwq)
4941 {
4942 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
4943 	struct mlx5_ib_create_wq ucmd = {};
4944 	int err;
4945 	size_t required_cmd_sz;
4946 
4947 	required_cmd_sz = offsetofend(struct mlx5_ib_create_wq,
4948 				      single_stride_log_num_of_bytes);
4949 	if (udata->inlen < required_cmd_sz) {
4950 		mlx5_ib_dbg(dev, "invalid inlen\n");
4951 		return -EINVAL;
4952 	}
4953 
4954 	if (udata->inlen > sizeof(ucmd) &&
4955 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
4956 				 udata->inlen - sizeof(ucmd))) {
4957 		mlx5_ib_dbg(dev, "inlen is not supported\n");
4958 		return -EOPNOTSUPP;
4959 	}
4960 
4961 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
4962 		mlx5_ib_dbg(dev, "copy failed\n");
4963 		return -EFAULT;
4964 	}
4965 
4966 	if (ucmd.comp_mask & (~MLX5_IB_CREATE_WQ_STRIDING_RQ)) {
4967 		mlx5_ib_dbg(dev, "invalid comp mask\n");
4968 		return -EOPNOTSUPP;
4969 	} else if (ucmd.comp_mask & MLX5_IB_CREATE_WQ_STRIDING_RQ) {
4970 		if (!MLX5_CAP_GEN(dev->mdev, striding_rq)) {
4971 			mlx5_ib_dbg(dev, "Striding RQ is not supported\n");
4972 			return -EOPNOTSUPP;
4973 		}
4974 		if ((ucmd.single_stride_log_num_of_bytes <
4975 		    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES) ||
4976 		    (ucmd.single_stride_log_num_of_bytes >
4977 		     MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES)) {
4978 			mlx5_ib_dbg(dev, "Invalid log stride size (%u. Range is %u - %u)\n",
4979 				    ucmd.single_stride_log_num_of_bytes,
4980 				    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES,
4981 				    MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES);
4982 			return -EINVAL;
4983 		}
4984 		if (!log_of_strides_valid(dev,
4985 					  ucmd.single_wqe_log_num_of_strides)) {
4986 			mlx5_ib_dbg(
4987 				dev,
4988 				"Invalid log num strides (%u. Range is %u - %u)\n",
4989 				ucmd.single_wqe_log_num_of_strides,
4990 				MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) ?
4991 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES :
4992 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES,
4993 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES);
4994 			return -EINVAL;
4995 		}
4996 		rwq->single_stride_log_num_of_bytes =
4997 			ucmd.single_stride_log_num_of_bytes;
4998 		rwq->log_num_strides = ucmd.single_wqe_log_num_of_strides;
4999 		rwq->two_byte_shift_en = !!ucmd.two_byte_shift_en;
5000 		rwq->create_flags |= MLX5_IB_WQ_FLAGS_STRIDING_RQ;
5001 	}
5002 
5003 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
5004 	if (err) {
5005 		mlx5_ib_dbg(dev, "err %d\n", err);
5006 		return err;
5007 	}
5008 
5009 	err = create_user_rq(dev, pd, udata, rwq, &ucmd);
5010 	if (err) {
5011 		mlx5_ib_dbg(dev, "err %d\n", err);
5012 		return err;
5013 	}
5014 
5015 	rwq->user_index = ucmd.user_index;
5016 	return 0;
5017 }
5018 
5019 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
5020 				struct ib_wq_init_attr *init_attr,
5021 				struct ib_udata *udata)
5022 {
5023 	struct mlx5_ib_dev *dev;
5024 	struct mlx5_ib_rwq *rwq;
5025 	struct mlx5_ib_create_wq_resp resp = {};
5026 	size_t min_resp_len;
5027 	int err;
5028 
5029 	if (!udata)
5030 		return ERR_PTR(-ENOSYS);
5031 
5032 	min_resp_len = offsetofend(struct mlx5_ib_create_wq_resp, reserved);
5033 	if (udata->outlen && udata->outlen < min_resp_len)
5034 		return ERR_PTR(-EINVAL);
5035 
5036 	if (!capable(CAP_SYS_RAWIO) &&
5037 	    init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP)
5038 		return ERR_PTR(-EPERM);
5039 
5040 	dev = to_mdev(pd->device);
5041 	switch (init_attr->wq_type) {
5042 	case IB_WQT_RQ:
5043 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
5044 		if (!rwq)
5045 			return ERR_PTR(-ENOMEM);
5046 		err = prepare_user_rq(pd, init_attr, udata, rwq);
5047 		if (err)
5048 			goto err;
5049 		err = create_rq(rwq, pd, init_attr);
5050 		if (err)
5051 			goto err_user_rq;
5052 		break;
5053 	default:
5054 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
5055 			    init_attr->wq_type);
5056 		return ERR_PTR(-EINVAL);
5057 	}
5058 
5059 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
5060 	rwq->ibwq.state = IB_WQS_RESET;
5061 	if (udata->outlen) {
5062 		resp.response_length = offsetofend(
5063 			struct mlx5_ib_create_wq_resp, response_length);
5064 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
5065 		if (err)
5066 			goto err_copy;
5067 	}
5068 
5069 	rwq->core_qp.event = mlx5_ib_wq_event;
5070 	rwq->ibwq.event_handler = init_attr->event_handler;
5071 	return &rwq->ibwq;
5072 
5073 err_copy:
5074 	mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5075 err_user_rq:
5076 	destroy_user_rq(dev, pd, rwq, udata);
5077 err:
5078 	kfree(rwq);
5079 	return ERR_PTR(err);
5080 }
5081 
5082 int mlx5_ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
5083 {
5084 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
5085 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
5086 	int ret;
5087 
5088 	ret = mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5089 	if (ret)
5090 		return ret;
5091 	destroy_user_rq(dev, wq->pd, rwq, udata);
5092 	kfree(rwq);
5093 	return 0;
5094 }
5095 
5096 int mlx5_ib_create_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_table,
5097 				 struct ib_rwq_ind_table_init_attr *init_attr,
5098 				 struct ib_udata *udata)
5099 {
5100 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl =
5101 		to_mrwq_ind_table(ib_rwq_ind_table);
5102 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_table->device);
5103 	int sz = 1 << init_attr->log_ind_tbl_size;
5104 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
5105 	size_t min_resp_len;
5106 	int inlen;
5107 	int err;
5108 	int i;
5109 	u32 *in;
5110 	void *rqtc;
5111 
5112 	if (udata->inlen > 0 &&
5113 	    !ib_is_udata_cleared(udata, 0,
5114 				 udata->inlen))
5115 		return -EOPNOTSUPP;
5116 
5117 	if (init_attr->log_ind_tbl_size >
5118 	    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
5119 		mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
5120 			    init_attr->log_ind_tbl_size,
5121 			    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
5122 		return -EINVAL;
5123 	}
5124 
5125 	min_resp_len =
5126 		offsetofend(struct mlx5_ib_create_rwq_ind_tbl_resp, reserved);
5127 	if (udata->outlen && udata->outlen < min_resp_len)
5128 		return -EINVAL;
5129 
5130 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
5131 	in = kvzalloc(inlen, GFP_KERNEL);
5132 	if (!in)
5133 		return -ENOMEM;
5134 
5135 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
5136 
5137 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
5138 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
5139 
5140 	for (i = 0; i < sz; i++)
5141 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
5142 
5143 	rwq_ind_tbl->uid = to_mpd(init_attr->ind_tbl[0]->pd)->uid;
5144 	MLX5_SET(create_rqt_in, in, uid, rwq_ind_tbl->uid);
5145 
5146 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
5147 	kvfree(in);
5148 	if (err)
5149 		return err;
5150 
5151 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
5152 	if (udata->outlen) {
5153 		resp.response_length =
5154 			offsetofend(struct mlx5_ib_create_rwq_ind_tbl_resp,
5155 				    response_length);
5156 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
5157 		if (err)
5158 			goto err_copy;
5159 	}
5160 
5161 	return 0;
5162 
5163 err_copy:
5164 	mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
5165 	return err;
5166 }
5167 
5168 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
5169 {
5170 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
5171 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
5172 
5173 	return mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
5174 }
5175 
5176 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
5177 		      u32 wq_attr_mask, struct ib_udata *udata)
5178 {
5179 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
5180 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
5181 	struct mlx5_ib_modify_wq ucmd = {};
5182 	size_t required_cmd_sz;
5183 	int curr_wq_state;
5184 	int wq_state;
5185 	int inlen;
5186 	int err;
5187 	void *rqc;
5188 	void *in;
5189 
5190 	required_cmd_sz = offsetofend(struct mlx5_ib_modify_wq, reserved);
5191 	if (udata->inlen < required_cmd_sz)
5192 		return -EINVAL;
5193 
5194 	if (udata->inlen > sizeof(ucmd) &&
5195 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
5196 				 udata->inlen - sizeof(ucmd)))
5197 		return -EOPNOTSUPP;
5198 
5199 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
5200 		return -EFAULT;
5201 
5202 	if (ucmd.comp_mask || ucmd.reserved)
5203 		return -EOPNOTSUPP;
5204 
5205 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
5206 	in = kvzalloc(inlen, GFP_KERNEL);
5207 	if (!in)
5208 		return -ENOMEM;
5209 
5210 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
5211 
5212 	curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
5213 		wq_attr->curr_wq_state : wq->state;
5214 	wq_state = (wq_attr_mask & IB_WQ_STATE) ?
5215 		wq_attr->wq_state : curr_wq_state;
5216 	if (curr_wq_state == IB_WQS_ERR)
5217 		curr_wq_state = MLX5_RQC_STATE_ERR;
5218 	if (wq_state == IB_WQS_ERR)
5219 		wq_state = MLX5_RQC_STATE_ERR;
5220 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
5221 	MLX5_SET(modify_rq_in, in, uid, to_mpd(wq->pd)->uid);
5222 	MLX5_SET(rqc, rqc, state, wq_state);
5223 
5224 	if (wq_attr_mask & IB_WQ_FLAGS) {
5225 		if (wq_attr->flags_mask & IB_WQ_FLAGS_CVLAN_STRIPPING) {
5226 			if (!(MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
5227 			      MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
5228 				mlx5_ib_dbg(dev, "VLAN offloads are not "
5229 					    "supported\n");
5230 				err = -EOPNOTSUPP;
5231 				goto out;
5232 			}
5233 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
5234 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
5235 			MLX5_SET(rqc, rqc, vsd,
5236 				 (wq_attr->flags & IB_WQ_FLAGS_CVLAN_STRIPPING) ? 0 : 1);
5237 		}
5238 
5239 		if (wq_attr->flags_mask & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
5240 			mlx5_ib_dbg(dev, "Modifying scatter end padding is not supported\n");
5241 			err = -EOPNOTSUPP;
5242 			goto out;
5243 		}
5244 	}
5245 
5246 	if (curr_wq_state == IB_WQS_RESET && wq_state == IB_WQS_RDY) {
5247 		u16 set_id;
5248 
5249 		set_id = mlx5_ib_get_counters_id(dev, 0);
5250 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
5251 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
5252 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
5253 			MLX5_SET(rqc, rqc, counter_set_id, set_id);
5254 		} else
5255 			dev_info_once(
5256 				&dev->ib_dev.dev,
5257 				"Receive WQ counters are not supported on current FW\n");
5258 	}
5259 
5260 	err = mlx5_core_modify_rq(dev->mdev, rwq->core_qp.qpn, in);
5261 	if (!err)
5262 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
5263 
5264 out:
5265 	kvfree(in);
5266 	return err;
5267 }
5268 
5269 struct mlx5_ib_drain_cqe {
5270 	struct ib_cqe cqe;
5271 	struct completion done;
5272 };
5273 
5274 static void mlx5_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
5275 {
5276 	struct mlx5_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
5277 						     struct mlx5_ib_drain_cqe,
5278 						     cqe);
5279 
5280 	complete(&cqe->done);
5281 }
5282 
5283 /* This function returns only once the drained WR was completed */
5284 static void handle_drain_completion(struct ib_cq *cq,
5285 				    struct mlx5_ib_drain_cqe *sdrain,
5286 				    struct mlx5_ib_dev *dev)
5287 {
5288 	struct mlx5_core_dev *mdev = dev->mdev;
5289 
5290 	if (cq->poll_ctx == IB_POLL_DIRECT) {
5291 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
5292 			ib_process_cq_direct(cq, -1);
5293 		return;
5294 	}
5295 
5296 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5297 		struct mlx5_ib_cq *mcq = to_mcq(cq);
5298 		bool triggered = false;
5299 		unsigned long flags;
5300 
5301 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
5302 		/* Make sure that the CQ handler won't run if wasn't run yet */
5303 		if (!mcq->mcq.reset_notify_added)
5304 			mcq->mcq.reset_notify_added = 1;
5305 		else
5306 			triggered = true;
5307 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
5308 
5309 		if (triggered) {
5310 			/* Wait for any scheduled/running task to be ended */
5311 			switch (cq->poll_ctx) {
5312 			case IB_POLL_SOFTIRQ:
5313 				irq_poll_disable(&cq->iop);
5314 				irq_poll_enable(&cq->iop);
5315 				break;
5316 			case IB_POLL_WORKQUEUE:
5317 				cancel_work_sync(&cq->work);
5318 				break;
5319 			default:
5320 				WARN_ON_ONCE(1);
5321 			}
5322 		}
5323 
5324 		/* Run the CQ handler - this makes sure that the drain WR will
5325 		 * be processed if wasn't processed yet.
5326 		 */
5327 		mcq->mcq.comp(&mcq->mcq, NULL);
5328 	}
5329 
5330 	wait_for_completion(&sdrain->done);
5331 }
5332 
5333 void mlx5_ib_drain_sq(struct ib_qp *qp)
5334 {
5335 	struct ib_cq *cq = qp->send_cq;
5336 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
5337 	struct mlx5_ib_drain_cqe sdrain;
5338 	const struct ib_send_wr *bad_swr;
5339 	struct ib_rdma_wr swr = {
5340 		.wr = {
5341 			.next = NULL,
5342 			{ .wr_cqe	= &sdrain.cqe, },
5343 			.opcode	= IB_WR_RDMA_WRITE,
5344 		},
5345 	};
5346 	int ret;
5347 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5348 	struct mlx5_core_dev *mdev = dev->mdev;
5349 
5350 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
5351 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5352 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
5353 		return;
5354 	}
5355 
5356 	sdrain.cqe.done = mlx5_ib_drain_qp_done;
5357 	init_completion(&sdrain.done);
5358 
5359 	ret = mlx5_ib_post_send_drain(qp, &swr.wr, &bad_swr);
5360 	if (ret) {
5361 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
5362 		return;
5363 	}
5364 
5365 	handle_drain_completion(cq, &sdrain, dev);
5366 }
5367 
5368 void mlx5_ib_drain_rq(struct ib_qp *qp)
5369 {
5370 	struct ib_cq *cq = qp->recv_cq;
5371 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
5372 	struct mlx5_ib_drain_cqe rdrain;
5373 	struct ib_recv_wr rwr = {};
5374 	const struct ib_recv_wr *bad_rwr;
5375 	int ret;
5376 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5377 	struct mlx5_core_dev *mdev = dev->mdev;
5378 
5379 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
5380 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5381 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
5382 		return;
5383 	}
5384 
5385 	rwr.wr_cqe = &rdrain.cqe;
5386 	rdrain.cqe.done = mlx5_ib_drain_qp_done;
5387 	init_completion(&rdrain.done);
5388 
5389 	ret = mlx5_ib_post_recv_drain(qp, &rwr, &bad_rwr);
5390 	if (ret) {
5391 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
5392 		return;
5393 	}
5394 
5395 	handle_drain_completion(cq, &rdrain, dev);
5396 }
5397 
5398 /**
5399  * Bind a qp to a counter. If @counter is NULL then bind the qp to
5400  * the default counter
5401  */
5402 int mlx5_ib_qp_set_counter(struct ib_qp *qp, struct rdma_counter *counter)
5403 {
5404 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5405 	struct mlx5_ib_qp *mqp = to_mqp(qp);
5406 	int err = 0;
5407 
5408 	mutex_lock(&mqp->mutex);
5409 	if (mqp->state == IB_QPS_RESET) {
5410 		qp->counter = counter;
5411 		goto out;
5412 	}
5413 
5414 	if (!MLX5_CAP_GEN(dev->mdev, rts2rts_qp_counters_set_id)) {
5415 		err = -EOPNOTSUPP;
5416 		goto out;
5417 	}
5418 
5419 	if (mqp->state == IB_QPS_RTS) {
5420 		err = __mlx5_ib_qp_set_counter(qp, counter);
5421 		if (!err)
5422 			qp->counter = counter;
5423 
5424 		goto out;
5425 	}
5426 
5427 	mqp->counter_pending = 1;
5428 	qp->counter = counter;
5429 
5430 out:
5431 	mutex_unlock(&mqp->mutex);
5432 	return err;
5433 }
5434