xref: /linux/drivers/infiniband/hw/mlx5/qp.c (revision 07f0148aafe8c95a3a76cd59e9e75b4d78d1d31d)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/etherdevice.h>
34 #include <rdma/ib_umem.h>
35 #include <rdma/ib_cache.h>
36 #include <rdma/ib_user_verbs.h>
37 #include <rdma/rdma_counter.h>
38 #include <linux/mlx5/fs.h>
39 #include "mlx5_ib.h"
40 #include "ib_rep.h"
41 #include "counters.h"
42 #include "cmd.h"
43 #include "umr.h"
44 #include "qp.h"
45 #include "wr.h"
46 
47 enum {
48 	MLX5_IB_ACK_REQ_FREQ	= 8,
49 };
50 
51 enum {
52 	MLX5_IB_DEFAULT_SCHED_QUEUE	= 0x83,
53 	MLX5_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
54 	MLX5_IB_LINK_TYPE_IB		= 0,
55 	MLX5_IB_LINK_TYPE_ETH		= 1
56 };
57 
58 enum raw_qp_set_mask_map {
59 	MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID		= 1UL << 0,
60 	MLX5_RAW_QP_RATE_LIMIT			= 1UL << 1,
61 };
62 
63 struct mlx5_modify_raw_qp_param {
64 	u16 operation;
65 
66 	u32 set_mask; /* raw_qp_set_mask_map */
67 
68 	struct mlx5_rate_limit rl;
69 
70 	u8 rq_q_ctr_id;
71 	u32 port;
72 };
73 
74 static void get_cqs(enum ib_qp_type qp_type,
75 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
76 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
77 
78 static int is_qp0(enum ib_qp_type qp_type)
79 {
80 	return qp_type == IB_QPT_SMI;
81 }
82 
83 static int is_sqp(enum ib_qp_type qp_type)
84 {
85 	return is_qp0(qp_type) || is_qp1(qp_type);
86 }
87 
88 /**
89  * mlx5_ib_read_user_wqe_common() - Copy a WQE (or part of) from user WQ
90  * to kernel buffer
91  *
92  * @umem: User space memory where the WQ is
93  * @buffer: buffer to copy to
94  * @buflen: buffer length
95  * @wqe_index: index of WQE to copy from
96  * @wq_offset: offset to start of WQ
97  * @wq_wqe_cnt: number of WQEs in WQ
98  * @wq_wqe_shift: log2 of WQE size
99  * @bcnt: number of bytes to copy
100  * @bytes_copied: number of bytes to copy (return value)
101  *
102  * Copies from start of WQE bcnt or less bytes.
103  * Does not gurantee to copy the entire WQE.
104  *
105  * Return: zero on success, or an error code.
106  */
107 static int mlx5_ib_read_user_wqe_common(struct ib_umem *umem, void *buffer,
108 					size_t buflen, int wqe_index,
109 					int wq_offset, int wq_wqe_cnt,
110 					int wq_wqe_shift, int bcnt,
111 					size_t *bytes_copied)
112 {
113 	size_t offset = wq_offset + ((wqe_index % wq_wqe_cnt) << wq_wqe_shift);
114 	size_t wq_end = wq_offset + (wq_wqe_cnt << wq_wqe_shift);
115 	size_t copy_length;
116 	int ret;
117 
118 	/* don't copy more than requested, more than buffer length or
119 	 * beyond WQ end
120 	 */
121 	copy_length = min_t(u32, buflen, wq_end - offset);
122 	copy_length = min_t(u32, copy_length, bcnt);
123 
124 	ret = ib_umem_copy_from(buffer, umem, offset, copy_length);
125 	if (ret)
126 		return ret;
127 
128 	if (!ret && bytes_copied)
129 		*bytes_copied = copy_length;
130 
131 	return 0;
132 }
133 
134 static int mlx5_ib_read_kernel_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index,
135 				      void *buffer, size_t buflen, size_t *bc)
136 {
137 	struct mlx5_wqe_ctrl_seg *ctrl;
138 	size_t bytes_copied = 0;
139 	size_t wqe_length;
140 	void *p;
141 	int ds;
142 
143 	wqe_index = wqe_index & qp->sq.fbc.sz_m1;
144 
145 	/* read the control segment first */
146 	p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, wqe_index);
147 	ctrl = p;
148 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
149 	wqe_length = ds * MLX5_WQE_DS_UNITS;
150 
151 	/* read rest of WQE if it spreads over more than one stride */
152 	while (bytes_copied < wqe_length) {
153 		size_t copy_length =
154 			min_t(size_t, buflen - bytes_copied, MLX5_SEND_WQE_BB);
155 
156 		if (!copy_length)
157 			break;
158 
159 		memcpy(buffer + bytes_copied, p, copy_length);
160 		bytes_copied += copy_length;
161 
162 		wqe_index = (wqe_index + 1) & qp->sq.fbc.sz_m1;
163 		p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, wqe_index);
164 	}
165 	*bc = bytes_copied;
166 	return 0;
167 }
168 
169 static int mlx5_ib_read_user_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index,
170 				    void *buffer, size_t buflen, size_t *bc)
171 {
172 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
173 	struct ib_umem *umem = base->ubuffer.umem;
174 	struct mlx5_ib_wq *wq = &qp->sq;
175 	struct mlx5_wqe_ctrl_seg *ctrl;
176 	size_t bytes_copied;
177 	size_t bytes_copied2;
178 	size_t wqe_length;
179 	int ret;
180 	int ds;
181 
182 	/* at first read as much as possible */
183 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index,
184 					   wq->offset, wq->wqe_cnt,
185 					   wq->wqe_shift, buflen,
186 					   &bytes_copied);
187 	if (ret)
188 		return ret;
189 
190 	/* we need at least control segment size to proceed */
191 	if (bytes_copied < sizeof(*ctrl))
192 		return -EINVAL;
193 
194 	ctrl = buffer;
195 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
196 	wqe_length = ds * MLX5_WQE_DS_UNITS;
197 
198 	/* if we copied enough then we are done */
199 	if (bytes_copied >= wqe_length) {
200 		*bc = bytes_copied;
201 		return 0;
202 	}
203 
204 	/* otherwise this a wrapped around wqe
205 	 * so read the remaining bytes starting
206 	 * from  wqe_index 0
207 	 */
208 	ret = mlx5_ib_read_user_wqe_common(umem, buffer + bytes_copied,
209 					   buflen - bytes_copied, 0, wq->offset,
210 					   wq->wqe_cnt, wq->wqe_shift,
211 					   wqe_length - bytes_copied,
212 					   &bytes_copied2);
213 
214 	if (ret)
215 		return ret;
216 	*bc = bytes_copied + bytes_copied2;
217 	return 0;
218 }
219 
220 int mlx5_ib_read_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
221 			size_t buflen, size_t *bc)
222 {
223 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
224 	struct ib_umem *umem = base->ubuffer.umem;
225 
226 	if (buflen < sizeof(struct mlx5_wqe_ctrl_seg))
227 		return -EINVAL;
228 
229 	if (!umem)
230 		return mlx5_ib_read_kernel_wqe_sq(qp, wqe_index, buffer,
231 						  buflen, bc);
232 
233 	return mlx5_ib_read_user_wqe_sq(qp, wqe_index, buffer, buflen, bc);
234 }
235 
236 static int mlx5_ib_read_user_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index,
237 				    void *buffer, size_t buflen, size_t *bc)
238 {
239 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
240 	struct ib_umem *umem = base->ubuffer.umem;
241 	struct mlx5_ib_wq *wq = &qp->rq;
242 	size_t bytes_copied;
243 	int ret;
244 
245 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index,
246 					   wq->offset, wq->wqe_cnt,
247 					   wq->wqe_shift, buflen,
248 					   &bytes_copied);
249 
250 	if (ret)
251 		return ret;
252 	*bc = bytes_copied;
253 	return 0;
254 }
255 
256 int mlx5_ib_read_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
257 			size_t buflen, size_t *bc)
258 {
259 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
260 	struct ib_umem *umem = base->ubuffer.umem;
261 	struct mlx5_ib_wq *wq = &qp->rq;
262 	size_t wqe_size = 1 << wq->wqe_shift;
263 
264 	if (buflen < wqe_size)
265 		return -EINVAL;
266 
267 	if (!umem)
268 		return -EOPNOTSUPP;
269 
270 	return mlx5_ib_read_user_wqe_rq(qp, wqe_index, buffer, buflen, bc);
271 }
272 
273 static int mlx5_ib_read_user_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index,
274 				     void *buffer, size_t buflen, size_t *bc)
275 {
276 	struct ib_umem *umem = srq->umem;
277 	size_t bytes_copied;
278 	int ret;
279 
280 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index, 0,
281 					   srq->msrq.max, srq->msrq.wqe_shift,
282 					   buflen, &bytes_copied);
283 
284 	if (ret)
285 		return ret;
286 	*bc = bytes_copied;
287 	return 0;
288 }
289 
290 int mlx5_ib_read_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index, void *buffer,
291 			 size_t buflen, size_t *bc)
292 {
293 	struct ib_umem *umem = srq->umem;
294 	size_t wqe_size = 1 << srq->msrq.wqe_shift;
295 
296 	if (buflen < wqe_size)
297 		return -EINVAL;
298 
299 	if (!umem)
300 		return -EOPNOTSUPP;
301 
302 	return mlx5_ib_read_user_wqe_srq(srq, wqe_index, buffer, buflen, bc);
303 }
304 
305 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
306 {
307 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
308 	struct ib_event event;
309 
310 	if (type == MLX5_EVENT_TYPE_PATH_MIG) {
311 		/* This event is only valid for trans_qps */
312 		to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
313 	}
314 
315 	if (ibqp->event_handler) {
316 		event.device     = ibqp->device;
317 		event.element.qp = ibqp;
318 		switch (type) {
319 		case MLX5_EVENT_TYPE_PATH_MIG:
320 			event.event = IB_EVENT_PATH_MIG;
321 			break;
322 		case MLX5_EVENT_TYPE_COMM_EST:
323 			event.event = IB_EVENT_COMM_EST;
324 			break;
325 		case MLX5_EVENT_TYPE_SQ_DRAINED:
326 			event.event = IB_EVENT_SQ_DRAINED;
327 			break;
328 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
329 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
330 			break;
331 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
332 			event.event = IB_EVENT_QP_FATAL;
333 			break;
334 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
335 			event.event = IB_EVENT_PATH_MIG_ERR;
336 			break;
337 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
338 			event.event = IB_EVENT_QP_REQ_ERR;
339 			break;
340 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
341 			event.event = IB_EVENT_QP_ACCESS_ERR;
342 			break;
343 		default:
344 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
345 			return;
346 		}
347 
348 		ibqp->event_handler(&event, ibqp->qp_context);
349 	}
350 }
351 
352 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
353 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
354 {
355 	int wqe_size;
356 	int wq_size;
357 
358 	/* Sanity check RQ size before proceeding */
359 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
360 		return -EINVAL;
361 
362 	if (!has_rq) {
363 		qp->rq.max_gs = 0;
364 		qp->rq.wqe_cnt = 0;
365 		qp->rq.wqe_shift = 0;
366 		cap->max_recv_wr = 0;
367 		cap->max_recv_sge = 0;
368 	} else {
369 		int wq_sig = !!(qp->flags_en & MLX5_QP_FLAG_SIGNATURE);
370 
371 		if (ucmd) {
372 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
373 			if (ucmd->rq_wqe_shift > BITS_PER_BYTE * sizeof(ucmd->rq_wqe_shift))
374 				return -EINVAL;
375 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
376 			if ((1 << qp->rq.wqe_shift) /
377 				    sizeof(struct mlx5_wqe_data_seg) <
378 			    wq_sig)
379 				return -EINVAL;
380 			qp->rq.max_gs =
381 				(1 << qp->rq.wqe_shift) /
382 					sizeof(struct mlx5_wqe_data_seg) -
383 				wq_sig;
384 			qp->rq.max_post = qp->rq.wqe_cnt;
385 		} else {
386 			wqe_size =
387 				wq_sig ? sizeof(struct mlx5_wqe_signature_seg) :
388 					 0;
389 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
390 			wqe_size = roundup_pow_of_two(wqe_size);
391 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
392 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
393 			qp->rq.wqe_cnt = wq_size / wqe_size;
394 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
395 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
396 					    wqe_size,
397 					    MLX5_CAP_GEN(dev->mdev,
398 							 max_wqe_sz_rq));
399 				return -EINVAL;
400 			}
401 			qp->rq.wqe_shift = ilog2(wqe_size);
402 			qp->rq.max_gs =
403 				(1 << qp->rq.wqe_shift) /
404 					sizeof(struct mlx5_wqe_data_seg) -
405 				wq_sig;
406 			qp->rq.max_post = qp->rq.wqe_cnt;
407 		}
408 	}
409 
410 	return 0;
411 }
412 
413 static int sq_overhead(struct ib_qp_init_attr *attr)
414 {
415 	int size = 0;
416 
417 	switch (attr->qp_type) {
418 	case IB_QPT_XRC_INI:
419 		size += sizeof(struct mlx5_wqe_xrc_seg);
420 		fallthrough;
421 	case IB_QPT_RC:
422 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
423 			max(sizeof(struct mlx5_wqe_atomic_seg) +
424 			    sizeof(struct mlx5_wqe_raddr_seg),
425 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
426 			    sizeof(struct mlx5_mkey_seg) +
427 			    MLX5_IB_SQ_UMR_INLINE_THRESHOLD /
428 			    MLX5_IB_UMR_OCTOWORD);
429 		break;
430 
431 	case IB_QPT_XRC_TGT:
432 		return 0;
433 
434 	case IB_QPT_UC:
435 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
436 			max(sizeof(struct mlx5_wqe_raddr_seg),
437 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
438 			    sizeof(struct mlx5_mkey_seg));
439 		break;
440 
441 	case IB_QPT_UD:
442 		if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
443 			size += sizeof(struct mlx5_wqe_eth_pad) +
444 				sizeof(struct mlx5_wqe_eth_seg);
445 		fallthrough;
446 	case IB_QPT_SMI:
447 	case MLX5_IB_QPT_HW_GSI:
448 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
449 			sizeof(struct mlx5_wqe_datagram_seg);
450 		break;
451 
452 	case MLX5_IB_QPT_REG_UMR:
453 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
454 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
455 			sizeof(struct mlx5_mkey_seg);
456 		break;
457 
458 	default:
459 		return -EINVAL;
460 	}
461 
462 	return size;
463 }
464 
465 static int calc_send_wqe(struct ib_qp_init_attr *attr)
466 {
467 	int inl_size = 0;
468 	int size;
469 
470 	size = sq_overhead(attr);
471 	if (size < 0)
472 		return size;
473 
474 	if (attr->cap.max_inline_data) {
475 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
476 			attr->cap.max_inline_data;
477 	}
478 
479 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
480 	if (attr->create_flags & IB_QP_CREATE_INTEGRITY_EN &&
481 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
482 		return MLX5_SIG_WQE_SIZE;
483 	else
484 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
485 }
486 
487 static int get_send_sge(struct ib_qp_init_attr *attr, int wqe_size)
488 {
489 	int max_sge;
490 
491 	if (attr->qp_type == IB_QPT_RC)
492 		max_sge = (min_t(int, wqe_size, 512) -
493 			   sizeof(struct mlx5_wqe_ctrl_seg) -
494 			   sizeof(struct mlx5_wqe_raddr_seg)) /
495 			sizeof(struct mlx5_wqe_data_seg);
496 	else if (attr->qp_type == IB_QPT_XRC_INI)
497 		max_sge = (min_t(int, wqe_size, 512) -
498 			   sizeof(struct mlx5_wqe_ctrl_seg) -
499 			   sizeof(struct mlx5_wqe_xrc_seg) -
500 			   sizeof(struct mlx5_wqe_raddr_seg)) /
501 			sizeof(struct mlx5_wqe_data_seg);
502 	else
503 		max_sge = (wqe_size - sq_overhead(attr)) /
504 			sizeof(struct mlx5_wqe_data_seg);
505 
506 	return min_t(int, max_sge, wqe_size - sq_overhead(attr) /
507 		     sizeof(struct mlx5_wqe_data_seg));
508 }
509 
510 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
511 			struct mlx5_ib_qp *qp)
512 {
513 	int wqe_size;
514 	int wq_size;
515 
516 	if (!attr->cap.max_send_wr)
517 		return 0;
518 
519 	wqe_size = calc_send_wqe(attr);
520 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
521 	if (wqe_size < 0)
522 		return wqe_size;
523 
524 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
525 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
526 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
527 		return -EINVAL;
528 	}
529 
530 	qp->max_inline_data = wqe_size - sq_overhead(attr) -
531 			      sizeof(struct mlx5_wqe_inline_seg);
532 	attr->cap.max_inline_data = qp->max_inline_data;
533 
534 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
535 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
536 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
537 		mlx5_ib_dbg(dev, "send queue size (%d * %d / %d -> %d) exceeds limits(%d)\n",
538 			    attr->cap.max_send_wr, wqe_size, MLX5_SEND_WQE_BB,
539 			    qp->sq.wqe_cnt,
540 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
541 		return -ENOMEM;
542 	}
543 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
544 	qp->sq.max_gs = get_send_sge(attr, wqe_size);
545 	if (qp->sq.max_gs < attr->cap.max_send_sge)
546 		return -ENOMEM;
547 
548 	attr->cap.max_send_sge = qp->sq.max_gs;
549 	qp->sq.max_post = wq_size / wqe_size;
550 	attr->cap.max_send_wr = qp->sq.max_post;
551 
552 	return wq_size;
553 }
554 
555 static int set_user_buf_size(struct mlx5_ib_dev *dev,
556 			    struct mlx5_ib_qp *qp,
557 			    struct mlx5_ib_create_qp *ucmd,
558 			    struct mlx5_ib_qp_base *base,
559 			    struct ib_qp_init_attr *attr)
560 {
561 	int desc_sz = 1 << qp->sq.wqe_shift;
562 
563 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
564 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
565 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
566 		return -EINVAL;
567 	}
568 
569 	if (ucmd->sq_wqe_count && !is_power_of_2(ucmd->sq_wqe_count)) {
570 		mlx5_ib_warn(dev, "sq_wqe_count %d is not a power of two\n",
571 			     ucmd->sq_wqe_count);
572 		return -EINVAL;
573 	}
574 
575 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
576 
577 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
578 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
579 			     qp->sq.wqe_cnt,
580 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
581 		return -EINVAL;
582 	}
583 
584 	if (attr->qp_type == IB_QPT_RAW_PACKET ||
585 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
586 		base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
587 		qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
588 	} else {
589 		base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
590 					 (qp->sq.wqe_cnt << 6);
591 	}
592 
593 	return 0;
594 }
595 
596 static int qp_has_rq(struct ib_qp_init_attr *attr)
597 {
598 	if (attr->qp_type == IB_QPT_XRC_INI ||
599 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
600 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
601 	    !attr->cap.max_recv_wr)
602 		return 0;
603 
604 	return 1;
605 }
606 
607 enum {
608 	/* this is the first blue flame register in the array of bfregs assigned
609 	 * to a processes. Since we do not use it for blue flame but rather
610 	 * regular 64 bit doorbells, we do not need a lock for maintaiing
611 	 * "odd/even" order
612 	 */
613 	NUM_NON_BLUE_FLAME_BFREGS = 1,
614 };
615 
616 static int max_bfregs(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi)
617 {
618 	return get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
619 	       bfregi->num_static_sys_pages * MLX5_NON_FP_BFREGS_PER_UAR;
620 }
621 
622 static int num_med_bfreg(struct mlx5_ib_dev *dev,
623 			 struct mlx5_bfreg_info *bfregi)
624 {
625 	int n;
626 
627 	n = max_bfregs(dev, bfregi) - bfregi->num_low_latency_bfregs -
628 	    NUM_NON_BLUE_FLAME_BFREGS;
629 
630 	return n >= 0 ? n : 0;
631 }
632 
633 static int first_med_bfreg(struct mlx5_ib_dev *dev,
634 			   struct mlx5_bfreg_info *bfregi)
635 {
636 	return num_med_bfreg(dev, bfregi) ? 1 : -ENOMEM;
637 }
638 
639 static int first_hi_bfreg(struct mlx5_ib_dev *dev,
640 			  struct mlx5_bfreg_info *bfregi)
641 {
642 	int med;
643 
644 	med = num_med_bfreg(dev, bfregi);
645 	return ++med;
646 }
647 
648 static int alloc_high_class_bfreg(struct mlx5_ib_dev *dev,
649 				  struct mlx5_bfreg_info *bfregi)
650 {
651 	int i;
652 
653 	for (i = first_hi_bfreg(dev, bfregi); i < max_bfregs(dev, bfregi); i++) {
654 		if (!bfregi->count[i]) {
655 			bfregi->count[i]++;
656 			return i;
657 		}
658 	}
659 
660 	return -ENOMEM;
661 }
662 
663 static int alloc_med_class_bfreg(struct mlx5_ib_dev *dev,
664 				 struct mlx5_bfreg_info *bfregi)
665 {
666 	int minidx = first_med_bfreg(dev, bfregi);
667 	int i;
668 
669 	if (minidx < 0)
670 		return minidx;
671 
672 	for (i = minidx; i < first_hi_bfreg(dev, bfregi); i++) {
673 		if (bfregi->count[i] < bfregi->count[minidx])
674 			minidx = i;
675 		if (!bfregi->count[minidx])
676 			break;
677 	}
678 
679 	bfregi->count[minidx]++;
680 	return minidx;
681 }
682 
683 static int alloc_bfreg(struct mlx5_ib_dev *dev,
684 		       struct mlx5_bfreg_info *bfregi)
685 {
686 	int bfregn = -ENOMEM;
687 
688 	if (bfregi->lib_uar_dyn)
689 		return -EINVAL;
690 
691 	mutex_lock(&bfregi->lock);
692 	if (bfregi->ver >= 2) {
693 		bfregn = alloc_high_class_bfreg(dev, bfregi);
694 		if (bfregn < 0)
695 			bfregn = alloc_med_class_bfreg(dev, bfregi);
696 	}
697 
698 	if (bfregn < 0) {
699 		BUILD_BUG_ON(NUM_NON_BLUE_FLAME_BFREGS != 1);
700 		bfregn = 0;
701 		bfregi->count[bfregn]++;
702 	}
703 	mutex_unlock(&bfregi->lock);
704 
705 	return bfregn;
706 }
707 
708 void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
709 {
710 	mutex_lock(&bfregi->lock);
711 	bfregi->count[bfregn]--;
712 	mutex_unlock(&bfregi->lock);
713 }
714 
715 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
716 {
717 	switch (state) {
718 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
719 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
720 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
721 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
722 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
723 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
724 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
725 	default:		return -1;
726 	}
727 }
728 
729 static int to_mlx5_st(enum ib_qp_type type)
730 {
731 	switch (type) {
732 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
733 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
734 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
735 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
736 	case IB_QPT_XRC_INI:
737 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
738 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
739 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
740 	case MLX5_IB_QPT_DCI:		return MLX5_QP_ST_DCI;
741 	case IB_QPT_RAW_PACKET:		return MLX5_QP_ST_RAW_ETHERTYPE;
742 	default:		return -EINVAL;
743 	}
744 }
745 
746 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
747 			     struct mlx5_ib_cq *recv_cq);
748 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
749 			       struct mlx5_ib_cq *recv_cq);
750 
751 int bfregn_to_uar_index(struct mlx5_ib_dev *dev,
752 			struct mlx5_bfreg_info *bfregi, u32 bfregn,
753 			bool dyn_bfreg)
754 {
755 	unsigned int bfregs_per_sys_page;
756 	u32 index_of_sys_page;
757 	u32 offset;
758 
759 	if (bfregi->lib_uar_dyn)
760 		return -EINVAL;
761 
762 	bfregs_per_sys_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
763 				MLX5_NON_FP_BFREGS_PER_UAR;
764 	index_of_sys_page = bfregn / bfregs_per_sys_page;
765 
766 	if (dyn_bfreg) {
767 		index_of_sys_page += bfregi->num_static_sys_pages;
768 
769 		if (index_of_sys_page >= bfregi->num_sys_pages)
770 			return -EINVAL;
771 
772 		if (bfregn > bfregi->num_dyn_bfregs ||
773 		    bfregi->sys_pages[index_of_sys_page] == MLX5_IB_INVALID_UAR_INDEX) {
774 			mlx5_ib_dbg(dev, "Invalid dynamic uar index\n");
775 			return -EINVAL;
776 		}
777 	}
778 
779 	offset = bfregn % bfregs_per_sys_page / MLX5_NON_FP_BFREGS_PER_UAR;
780 	return bfregi->sys_pages[index_of_sys_page] + offset;
781 }
782 
783 static void destroy_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
784 			    struct mlx5_ib_rwq *rwq, struct ib_udata *udata)
785 {
786 	struct mlx5_ib_ucontext *context =
787 		rdma_udata_to_drv_context(
788 			udata,
789 			struct mlx5_ib_ucontext,
790 			ibucontext);
791 
792 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_DELAY_DROP)
793 		atomic_dec(&dev->delay_drop.rqs_cnt);
794 
795 	mlx5_ib_db_unmap_user(context, &rwq->db);
796 	ib_umem_release(rwq->umem);
797 }
798 
799 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
800 			  struct ib_udata *udata, struct mlx5_ib_rwq *rwq,
801 			  struct mlx5_ib_create_wq *ucmd)
802 {
803 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
804 		udata, struct mlx5_ib_ucontext, ibucontext);
805 	unsigned long page_size = 0;
806 	u32 offset = 0;
807 	int err;
808 
809 	if (!ucmd->buf_addr)
810 		return -EINVAL;
811 
812 	rwq->umem = ib_umem_get(&dev->ib_dev, ucmd->buf_addr, rwq->buf_size, 0);
813 	if (IS_ERR(rwq->umem)) {
814 		mlx5_ib_dbg(dev, "umem_get failed\n");
815 		err = PTR_ERR(rwq->umem);
816 		return err;
817 	}
818 
819 	page_size = mlx5_umem_find_best_quantized_pgoff(
820 		rwq->umem, wq, log_wq_pg_sz, MLX5_ADAPTER_PAGE_SHIFT,
821 		page_offset, 64, &rwq->rq_page_offset);
822 	if (!page_size) {
823 		mlx5_ib_warn(dev, "bad offset\n");
824 		err = -EINVAL;
825 		goto err_umem;
826 	}
827 
828 	rwq->rq_num_pas = ib_umem_num_dma_blocks(rwq->umem, page_size);
829 	rwq->page_shift = order_base_2(page_size);
830 	rwq->log_page_size =  rwq->page_shift - MLX5_ADAPTER_PAGE_SHIFT;
831 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
832 
833 	mlx5_ib_dbg(
834 		dev,
835 		"addr 0x%llx, size %zd, npages %zu, page_size %ld, ncont %d, offset %d\n",
836 		(unsigned long long)ucmd->buf_addr, rwq->buf_size,
837 		ib_umem_num_pages(rwq->umem), page_size, rwq->rq_num_pas,
838 		offset);
839 
840 	err = mlx5_ib_db_map_user(ucontext, ucmd->db_addr, &rwq->db);
841 	if (err) {
842 		mlx5_ib_dbg(dev, "map failed\n");
843 		goto err_umem;
844 	}
845 
846 	return 0;
847 
848 err_umem:
849 	ib_umem_release(rwq->umem);
850 	return err;
851 }
852 
853 static int adjust_bfregn(struct mlx5_ib_dev *dev,
854 			 struct mlx5_bfreg_info *bfregi, int bfregn)
855 {
856 	return bfregn / MLX5_NON_FP_BFREGS_PER_UAR * MLX5_BFREGS_PER_UAR +
857 				bfregn % MLX5_NON_FP_BFREGS_PER_UAR;
858 }
859 
860 static int _create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
861 			   struct mlx5_ib_qp *qp, struct ib_udata *udata,
862 			   struct ib_qp_init_attr *attr, u32 **in,
863 			   struct mlx5_ib_create_qp_resp *resp, int *inlen,
864 			   struct mlx5_ib_qp_base *base,
865 			   struct mlx5_ib_create_qp *ucmd)
866 {
867 	struct mlx5_ib_ucontext *context;
868 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
869 	unsigned int page_offset_quantized = 0;
870 	unsigned long page_size = 0;
871 	int uar_index = 0;
872 	int bfregn;
873 	int ncont = 0;
874 	__be64 *pas;
875 	void *qpc;
876 	int err;
877 	u16 uid;
878 	u32 uar_flags;
879 
880 	context = rdma_udata_to_drv_context(udata, struct mlx5_ib_ucontext,
881 					    ibucontext);
882 	uar_flags = qp->flags_en &
883 		    (MLX5_QP_FLAG_UAR_PAGE_INDEX | MLX5_QP_FLAG_BFREG_INDEX);
884 	switch (uar_flags) {
885 	case MLX5_QP_FLAG_UAR_PAGE_INDEX:
886 		uar_index = ucmd->bfreg_index;
887 		bfregn = MLX5_IB_INVALID_BFREG;
888 		break;
889 	case MLX5_QP_FLAG_BFREG_INDEX:
890 		uar_index = bfregn_to_uar_index(dev, &context->bfregi,
891 						ucmd->bfreg_index, true);
892 		if (uar_index < 0)
893 			return uar_index;
894 		bfregn = MLX5_IB_INVALID_BFREG;
895 		break;
896 	case 0:
897 		if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
898 			return -EINVAL;
899 		bfregn = alloc_bfreg(dev, &context->bfregi);
900 		if (bfregn < 0)
901 			return bfregn;
902 		break;
903 	default:
904 		return -EINVAL;
905 	}
906 
907 	mlx5_ib_dbg(dev, "bfregn 0x%x, uar_index 0x%x\n", bfregn, uar_index);
908 	if (bfregn != MLX5_IB_INVALID_BFREG)
909 		uar_index = bfregn_to_uar_index(dev, &context->bfregi, bfregn,
910 						false);
911 
912 	qp->rq.offset = 0;
913 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
914 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
915 
916 	err = set_user_buf_size(dev, qp, ucmd, base, attr);
917 	if (err)
918 		goto err_bfreg;
919 
920 	if (ucmd->buf_addr && ubuffer->buf_size) {
921 		ubuffer->buf_addr = ucmd->buf_addr;
922 		ubuffer->umem = ib_umem_get(&dev->ib_dev, ubuffer->buf_addr,
923 					    ubuffer->buf_size, 0);
924 		if (IS_ERR(ubuffer->umem)) {
925 			err = PTR_ERR(ubuffer->umem);
926 			goto err_bfreg;
927 		}
928 		page_size = mlx5_umem_find_best_quantized_pgoff(
929 			ubuffer->umem, qpc, log_page_size,
930 			MLX5_ADAPTER_PAGE_SHIFT, page_offset, 64,
931 			&page_offset_quantized);
932 		if (!page_size) {
933 			err = -EINVAL;
934 			goto err_umem;
935 		}
936 		ncont = ib_umem_num_dma_blocks(ubuffer->umem, page_size);
937 	} else {
938 		ubuffer->umem = NULL;
939 	}
940 
941 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
942 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
943 	*in = kvzalloc(*inlen, GFP_KERNEL);
944 	if (!*in) {
945 		err = -ENOMEM;
946 		goto err_umem;
947 	}
948 
949 	uid = (attr->qp_type != IB_QPT_XRC_INI) ? to_mpd(pd)->uid : 0;
950 	MLX5_SET(create_qp_in, *in, uid, uid);
951 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
952 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
953 	if (ubuffer->umem) {
954 		mlx5_ib_populate_pas(ubuffer->umem, page_size, pas, 0);
955 		MLX5_SET(qpc, qpc, log_page_size,
956 			 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
957 		MLX5_SET(qpc, qpc, page_offset, page_offset_quantized);
958 	}
959 	MLX5_SET(qpc, qpc, uar_page, uar_index);
960 	if (bfregn != MLX5_IB_INVALID_BFREG)
961 		resp->bfreg_index = adjust_bfregn(dev, &context->bfregi, bfregn);
962 	else
963 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
964 	qp->bfregn = bfregn;
965 
966 	err = mlx5_ib_db_map_user(context, ucmd->db_addr, &qp->db);
967 	if (err) {
968 		mlx5_ib_dbg(dev, "map failed\n");
969 		goto err_free;
970 	}
971 
972 	return 0;
973 
974 err_free:
975 	kvfree(*in);
976 
977 err_umem:
978 	ib_umem_release(ubuffer->umem);
979 
980 err_bfreg:
981 	if (bfregn != MLX5_IB_INVALID_BFREG)
982 		mlx5_ib_free_bfreg(dev, &context->bfregi, bfregn);
983 	return err;
984 }
985 
986 static void destroy_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
987 		       struct mlx5_ib_qp_base *base, struct ib_udata *udata)
988 {
989 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
990 		udata, struct mlx5_ib_ucontext, ibucontext);
991 
992 	if (udata) {
993 		/* User QP */
994 		mlx5_ib_db_unmap_user(context, &qp->db);
995 		ib_umem_release(base->ubuffer.umem);
996 
997 		/*
998 		 * Free only the BFREGs which are handled by the kernel.
999 		 * BFREGs of UARs allocated dynamically are handled by user.
1000 		 */
1001 		if (qp->bfregn != MLX5_IB_INVALID_BFREG)
1002 			mlx5_ib_free_bfreg(dev, &context->bfregi, qp->bfregn);
1003 		return;
1004 	}
1005 
1006 	/* Kernel QP */
1007 	kvfree(qp->sq.wqe_head);
1008 	kvfree(qp->sq.w_list);
1009 	kvfree(qp->sq.wrid);
1010 	kvfree(qp->sq.wr_data);
1011 	kvfree(qp->rq.wrid);
1012 	if (qp->db.db)
1013 		mlx5_db_free(dev->mdev, &qp->db);
1014 	if (qp->buf.frags)
1015 		mlx5_frag_buf_free(dev->mdev, &qp->buf);
1016 }
1017 
1018 static int _create_kernel_qp(struct mlx5_ib_dev *dev,
1019 			     struct ib_qp_init_attr *init_attr,
1020 			     struct mlx5_ib_qp *qp, u32 **in, int *inlen,
1021 			     struct mlx5_ib_qp_base *base)
1022 {
1023 	int uar_index;
1024 	void *qpc;
1025 	int err;
1026 
1027 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
1028 		qp->bf.bfreg = &dev->fp_bfreg;
1029 	else if (qp->flags & MLX5_IB_QP_CREATE_WC_TEST)
1030 		qp->bf.bfreg = &dev->wc_bfreg;
1031 	else
1032 		qp->bf.bfreg = &dev->bfreg;
1033 
1034 	/* We need to divide by two since each register is comprised of
1035 	 * two buffers of identical size, namely odd and even
1036 	 */
1037 	qp->bf.buf_size = (1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size)) / 2;
1038 	uar_index = qp->bf.bfreg->index;
1039 
1040 	err = calc_sq_size(dev, init_attr, qp);
1041 	if (err < 0) {
1042 		mlx5_ib_dbg(dev, "err %d\n", err);
1043 		return err;
1044 	}
1045 
1046 	qp->rq.offset = 0;
1047 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
1048 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
1049 
1050 	err = mlx5_frag_buf_alloc_node(dev->mdev, base->ubuffer.buf_size,
1051 				       &qp->buf, dev->mdev->priv.numa_node);
1052 	if (err) {
1053 		mlx5_ib_dbg(dev, "err %d\n", err);
1054 		return err;
1055 	}
1056 
1057 	if (qp->rq.wqe_cnt)
1058 		mlx5_init_fbc(qp->buf.frags, qp->rq.wqe_shift,
1059 			      ilog2(qp->rq.wqe_cnt), &qp->rq.fbc);
1060 
1061 	if (qp->sq.wqe_cnt) {
1062 		int sq_strides_offset = (qp->sq.offset  & (PAGE_SIZE - 1)) /
1063 					MLX5_SEND_WQE_BB;
1064 		mlx5_init_fbc_offset(qp->buf.frags +
1065 				     (qp->sq.offset / PAGE_SIZE),
1066 				     ilog2(MLX5_SEND_WQE_BB),
1067 				     ilog2(qp->sq.wqe_cnt),
1068 				     sq_strides_offset, &qp->sq.fbc);
1069 
1070 		qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
1071 	}
1072 
1073 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
1074 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
1075 	*in = kvzalloc(*inlen, GFP_KERNEL);
1076 	if (!*in) {
1077 		err = -ENOMEM;
1078 		goto err_buf;
1079 	}
1080 
1081 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
1082 	MLX5_SET(qpc, qpc, uar_page, uar_index);
1083 	MLX5_SET(qpc, qpc, ts_format, mlx5_get_qp_default_ts(dev->mdev));
1084 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1085 
1086 	/* Set "fast registration enabled" for all kernel QPs */
1087 	MLX5_SET(qpc, qpc, fre, 1);
1088 	MLX5_SET(qpc, qpc, rlky, 1);
1089 
1090 	if (qp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
1091 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
1092 
1093 	mlx5_fill_page_frag_array(&qp->buf,
1094 				  (__be64 *)MLX5_ADDR_OF(create_qp_in,
1095 							 *in, pas));
1096 
1097 	err = mlx5_db_alloc(dev->mdev, &qp->db);
1098 	if (err) {
1099 		mlx5_ib_dbg(dev, "err %d\n", err);
1100 		goto err_free;
1101 	}
1102 
1103 	qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1104 				     sizeof(*qp->sq.wrid), GFP_KERNEL);
1105 	qp->sq.wr_data = kvmalloc_array(qp->sq.wqe_cnt,
1106 					sizeof(*qp->sq.wr_data), GFP_KERNEL);
1107 	qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1108 				     sizeof(*qp->rq.wrid), GFP_KERNEL);
1109 	qp->sq.w_list = kvmalloc_array(qp->sq.wqe_cnt,
1110 				       sizeof(*qp->sq.w_list), GFP_KERNEL);
1111 	qp->sq.wqe_head = kvmalloc_array(qp->sq.wqe_cnt,
1112 					 sizeof(*qp->sq.wqe_head), GFP_KERNEL);
1113 
1114 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
1115 	    !qp->sq.w_list || !qp->sq.wqe_head) {
1116 		err = -ENOMEM;
1117 		goto err_wrid;
1118 	}
1119 
1120 	return 0;
1121 
1122 err_wrid:
1123 	kvfree(qp->sq.wqe_head);
1124 	kvfree(qp->sq.w_list);
1125 	kvfree(qp->sq.wrid);
1126 	kvfree(qp->sq.wr_data);
1127 	kvfree(qp->rq.wrid);
1128 	mlx5_db_free(dev->mdev, &qp->db);
1129 
1130 err_free:
1131 	kvfree(*in);
1132 
1133 err_buf:
1134 	mlx5_frag_buf_free(dev->mdev, &qp->buf);
1135 	return err;
1136 }
1137 
1138 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
1139 {
1140 	if (attr->srq || (qp->type == IB_QPT_XRC_TGT) ||
1141 	    (qp->type == MLX5_IB_QPT_DCI) || (qp->type == IB_QPT_XRC_INI))
1142 		return MLX5_SRQ_RQ;
1143 	else if (!qp->has_rq)
1144 		return MLX5_ZERO_LEN_RQ;
1145 
1146 	return MLX5_NON_ZERO_RQ;
1147 }
1148 
1149 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1150 				    struct mlx5_ib_qp *qp,
1151 				    struct mlx5_ib_sq *sq, u32 tdn,
1152 				    struct ib_pd *pd)
1153 {
1154 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
1155 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1156 
1157 	MLX5_SET(create_tis_in, in, uid, to_mpd(pd)->uid);
1158 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1159 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
1160 		MLX5_SET(tisc, tisc, underlay_qpn, qp->underlay_qpn);
1161 
1162 	return mlx5_core_create_tis(dev->mdev, in, &sq->tisn);
1163 }
1164 
1165 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1166 				      struct mlx5_ib_sq *sq, struct ib_pd *pd)
1167 {
1168 	mlx5_cmd_destroy_tis(dev->mdev, sq->tisn, to_mpd(pd)->uid);
1169 }
1170 
1171 static void destroy_flow_rule_vport_sq(struct mlx5_ib_sq *sq)
1172 {
1173 	if (sq->flow_rule)
1174 		mlx5_del_flow_rules(sq->flow_rule);
1175 	sq->flow_rule = NULL;
1176 }
1177 
1178 static bool fr_supported(int ts_cap)
1179 {
1180 	return ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_FREE_RUNNING ||
1181 	       ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_FREE_RUNNING_AND_REAL_TIME;
1182 }
1183 
1184 static int get_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1185 			 bool fr_sup, bool rt_sup)
1186 {
1187 	if (cq->private_flags & MLX5_IB_CQ_PR_FLAGS_REAL_TIME_TS) {
1188 		if (!rt_sup) {
1189 			mlx5_ib_dbg(dev,
1190 				    "Real time TS format is not supported\n");
1191 			return -EOPNOTSUPP;
1192 		}
1193 		return MLX5_TIMESTAMP_FORMAT_REAL_TIME;
1194 	}
1195 	if (cq->create_flags & IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION) {
1196 		if (!fr_sup) {
1197 			mlx5_ib_dbg(dev,
1198 				    "Free running TS format is not supported\n");
1199 			return -EOPNOTSUPP;
1200 		}
1201 		return MLX5_TIMESTAMP_FORMAT_FREE_RUNNING;
1202 	}
1203 	return fr_sup ? MLX5_TIMESTAMP_FORMAT_FREE_RUNNING :
1204 			MLX5_TIMESTAMP_FORMAT_DEFAULT;
1205 }
1206 
1207 static int get_rq_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *recv_cq)
1208 {
1209 	u8 ts_cap = MLX5_CAP_GEN(dev->mdev, rq_ts_format);
1210 
1211 	return get_ts_format(dev, recv_cq, fr_supported(ts_cap),
1212 			     rt_supported(ts_cap));
1213 }
1214 
1215 static int get_sq_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq)
1216 {
1217 	u8 ts_cap = MLX5_CAP_GEN(dev->mdev, sq_ts_format);
1218 
1219 	return get_ts_format(dev, send_cq, fr_supported(ts_cap),
1220 			     rt_supported(ts_cap));
1221 }
1222 
1223 static int get_qp_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq,
1224 			    struct mlx5_ib_cq *recv_cq)
1225 {
1226 	u8 ts_cap = MLX5_CAP_ROCE(dev->mdev, qp_ts_format);
1227 	bool fr_sup = fr_supported(ts_cap);
1228 	bool rt_sup = rt_supported(ts_cap);
1229 	u8 default_ts = fr_sup ? MLX5_TIMESTAMP_FORMAT_FREE_RUNNING :
1230 				 MLX5_TIMESTAMP_FORMAT_DEFAULT;
1231 	int send_ts_format =
1232 		send_cq ? get_ts_format(dev, send_cq, fr_sup, rt_sup) :
1233 			  default_ts;
1234 	int recv_ts_format =
1235 		recv_cq ? get_ts_format(dev, recv_cq, fr_sup, rt_sup) :
1236 			  default_ts;
1237 
1238 	if (send_ts_format < 0 || recv_ts_format < 0)
1239 		return -EOPNOTSUPP;
1240 
1241 	if (send_ts_format != MLX5_TIMESTAMP_FORMAT_DEFAULT &&
1242 	    recv_ts_format != MLX5_TIMESTAMP_FORMAT_DEFAULT &&
1243 	    send_ts_format != recv_ts_format) {
1244 		mlx5_ib_dbg(
1245 			dev,
1246 			"The send ts_format does not match the receive ts_format\n");
1247 		return -EOPNOTSUPP;
1248 	}
1249 
1250 	return send_ts_format == default_ts ? recv_ts_format : send_ts_format;
1251 }
1252 
1253 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1254 				   struct ib_udata *udata,
1255 				   struct mlx5_ib_sq *sq, void *qpin,
1256 				   struct ib_pd *pd, struct mlx5_ib_cq *cq)
1257 {
1258 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1259 	__be64 *pas;
1260 	void *in;
1261 	void *sqc;
1262 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1263 	void *wq;
1264 	int inlen;
1265 	int err;
1266 	unsigned int page_offset_quantized;
1267 	unsigned long page_size;
1268 	int ts_format;
1269 
1270 	ts_format = get_sq_ts_format(dev, cq);
1271 	if (ts_format < 0)
1272 		return ts_format;
1273 
1274 	sq->ubuffer.umem = ib_umem_get(&dev->ib_dev, ubuffer->buf_addr,
1275 				       ubuffer->buf_size, 0);
1276 	if (IS_ERR(sq->ubuffer.umem))
1277 		return PTR_ERR(sq->ubuffer.umem);
1278 	page_size = mlx5_umem_find_best_quantized_pgoff(
1279 		ubuffer->umem, wq, log_wq_pg_sz, MLX5_ADAPTER_PAGE_SHIFT,
1280 		page_offset, 64, &page_offset_quantized);
1281 	if (!page_size) {
1282 		err = -EINVAL;
1283 		goto err_umem;
1284 	}
1285 
1286 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
1287 		sizeof(u64) *
1288 			ib_umem_num_dma_blocks(sq->ubuffer.umem, page_size);
1289 	in = kvzalloc(inlen, GFP_KERNEL);
1290 	if (!in) {
1291 		err = -ENOMEM;
1292 		goto err_umem;
1293 	}
1294 
1295 	MLX5_SET(create_sq_in, in, uid, to_mpd(pd)->uid);
1296 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1297 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1298 	if (MLX5_CAP_ETH(dev->mdev, multi_pkt_send_wqe))
1299 		MLX5_SET(sqc, sqc, allow_multi_pkt_send_wqe, 1);
1300 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1301 	MLX5_SET(sqc, sqc, ts_format, ts_format);
1302 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1303 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1304 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1305 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1306 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
1307 	    MLX5_CAP_ETH(dev->mdev, swp))
1308 		MLX5_SET(sqc, sqc, allow_swp, 1);
1309 
1310 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1311 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1312 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1313 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1314 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1315 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1316 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1317 	MLX5_SET(wq, wq, log_wq_pg_sz,
1318 		 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
1319 	MLX5_SET(wq, wq, page_offset, page_offset_quantized);
1320 
1321 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1322 	mlx5_ib_populate_pas(sq->ubuffer.umem, page_size, pas, 0);
1323 
1324 	err = mlx5_core_create_sq_tracked(dev, in, inlen, &sq->base.mqp);
1325 
1326 	kvfree(in);
1327 
1328 	if (err)
1329 		goto err_umem;
1330 
1331 	return 0;
1332 
1333 err_umem:
1334 	ib_umem_release(sq->ubuffer.umem);
1335 	sq->ubuffer.umem = NULL;
1336 
1337 	return err;
1338 }
1339 
1340 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1341 				     struct mlx5_ib_sq *sq)
1342 {
1343 	destroy_flow_rule_vport_sq(sq);
1344 	mlx5_core_destroy_sq_tracked(dev, &sq->base.mqp);
1345 	ib_umem_release(sq->ubuffer.umem);
1346 }
1347 
1348 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1349 				   struct mlx5_ib_rq *rq, void *qpin,
1350 				   struct ib_pd *pd, struct mlx5_ib_cq *cq)
1351 {
1352 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1353 	__be64 *pas;
1354 	void *in;
1355 	void *rqc;
1356 	void *wq;
1357 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1358 	struct ib_umem *umem = rq->base.ubuffer.umem;
1359 	unsigned int page_offset_quantized;
1360 	unsigned long page_size = 0;
1361 	int ts_format;
1362 	size_t inlen;
1363 	int err;
1364 
1365 	ts_format = get_rq_ts_format(dev, cq);
1366 	if (ts_format < 0)
1367 		return ts_format;
1368 
1369 	page_size = mlx5_umem_find_best_quantized_pgoff(umem, wq, log_wq_pg_sz,
1370 							MLX5_ADAPTER_PAGE_SHIFT,
1371 							page_offset, 64,
1372 							&page_offset_quantized);
1373 	if (!page_size)
1374 		return -EINVAL;
1375 
1376 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
1377 		sizeof(u64) * ib_umem_num_dma_blocks(umem, page_size);
1378 	in = kvzalloc(inlen, GFP_KERNEL);
1379 	if (!in)
1380 		return -ENOMEM;
1381 
1382 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
1383 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1384 	if (!(rq->flags & MLX5_IB_RQ_CVLAN_STRIPPING))
1385 		MLX5_SET(rqc, rqc, vsd, 1);
1386 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1387 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1388 	MLX5_SET(rqc, rqc, ts_format, ts_format);
1389 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1390 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1391 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1392 
1393 	if (mqp->flags & IB_QP_CREATE_SCATTER_FCS)
1394 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1395 
1396 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1397 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1398 	if (rq->flags & MLX5_IB_RQ_PCI_WRITE_END_PADDING)
1399 		MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1400 	MLX5_SET(wq, wq, page_offset, page_offset_quantized);
1401 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1402 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1403 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1404 	MLX5_SET(wq, wq, log_wq_pg_sz,
1405 		 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
1406 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1407 
1408 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1409 	mlx5_ib_populate_pas(umem, page_size, pas, 0);
1410 
1411 	err = mlx5_core_create_rq_tracked(dev, in, inlen, &rq->base.mqp);
1412 
1413 	kvfree(in);
1414 
1415 	return err;
1416 }
1417 
1418 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1419 				     struct mlx5_ib_rq *rq)
1420 {
1421 	mlx5_core_destroy_rq_tracked(dev, &rq->base.mqp);
1422 }
1423 
1424 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1425 				      struct mlx5_ib_rq *rq,
1426 				      u32 qp_flags_en,
1427 				      struct ib_pd *pd)
1428 {
1429 	if (qp_flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1430 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1431 		mlx5_ib_disable_lb(dev, false, true);
1432 	mlx5_cmd_destroy_tir(dev->mdev, rq->tirn, to_mpd(pd)->uid);
1433 }
1434 
1435 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1436 				    struct mlx5_ib_rq *rq, u32 tdn,
1437 				    u32 *qp_flags_en, struct ib_pd *pd,
1438 				    u32 *out)
1439 {
1440 	u8 lb_flag = 0;
1441 	u32 *in;
1442 	void *tirc;
1443 	int inlen;
1444 	int err;
1445 
1446 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1447 	in = kvzalloc(inlen, GFP_KERNEL);
1448 	if (!in)
1449 		return -ENOMEM;
1450 
1451 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1452 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1453 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1454 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1455 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1456 	if (*qp_flags_en & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1457 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1458 
1459 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1460 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1461 
1462 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1463 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1464 
1465 	if (dev->is_rep) {
1466 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1467 		*qp_flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1468 	}
1469 
1470 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1471 	MLX5_SET(create_tir_in, in, opcode, MLX5_CMD_OP_CREATE_TIR);
1472 	err = mlx5_cmd_exec_inout(dev->mdev, create_tir, in, out);
1473 	rq->tirn = MLX5_GET(create_tir_out, out, tirn);
1474 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1475 		err = mlx5_ib_enable_lb(dev, false, true);
1476 
1477 		if (err)
1478 			destroy_raw_packet_qp_tir(dev, rq, 0, pd);
1479 	}
1480 	kvfree(in);
1481 
1482 	return err;
1483 }
1484 
1485 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1486 				u32 *in, size_t inlen, struct ib_pd *pd,
1487 				struct ib_udata *udata,
1488 				struct mlx5_ib_create_qp_resp *resp,
1489 				struct ib_qp_init_attr *init_attr)
1490 {
1491 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1492 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1493 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1494 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1495 		udata, struct mlx5_ib_ucontext, ibucontext);
1496 	int err;
1497 	u32 tdn = mucontext->tdn;
1498 	u16 uid = to_mpd(pd)->uid;
1499 	u32 out[MLX5_ST_SZ_DW(create_tir_out)] = {};
1500 
1501 	if (!qp->sq.wqe_cnt && !qp->rq.wqe_cnt)
1502 		return -EINVAL;
1503 	if (qp->sq.wqe_cnt) {
1504 		err = create_raw_packet_qp_tis(dev, qp, sq, tdn, pd);
1505 		if (err)
1506 			return err;
1507 
1508 		err = create_raw_packet_qp_sq(dev, udata, sq, in, pd,
1509 					      to_mcq(init_attr->send_cq));
1510 		if (err)
1511 			goto err_destroy_tis;
1512 
1513 		if (uid) {
1514 			resp->tisn = sq->tisn;
1515 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TISN;
1516 			resp->sqn = sq->base.mqp.qpn;
1517 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_SQN;
1518 		}
1519 
1520 		sq->base.container_mibqp = qp;
1521 		sq->base.mqp.event = mlx5_ib_qp_event;
1522 	}
1523 
1524 	if (qp->rq.wqe_cnt) {
1525 		rq->base.container_mibqp = qp;
1526 
1527 		if (qp->flags & IB_QP_CREATE_CVLAN_STRIPPING)
1528 			rq->flags |= MLX5_IB_RQ_CVLAN_STRIPPING;
1529 		if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING)
1530 			rq->flags |= MLX5_IB_RQ_PCI_WRITE_END_PADDING;
1531 		err = create_raw_packet_qp_rq(dev, rq, in, pd,
1532 					      to_mcq(init_attr->recv_cq));
1533 		if (err)
1534 			goto err_destroy_sq;
1535 
1536 		err = create_raw_packet_qp_tir(dev, rq, tdn, &qp->flags_en, pd,
1537 					       out);
1538 		if (err)
1539 			goto err_destroy_rq;
1540 
1541 		if (uid) {
1542 			resp->rqn = rq->base.mqp.qpn;
1543 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_RQN;
1544 			resp->tirn = rq->tirn;
1545 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1546 			if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
1547 			    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2)) {
1548 				resp->tir_icm_addr = MLX5_GET(
1549 					create_tir_out, out, icm_address_31_0);
1550 				resp->tir_icm_addr |=
1551 					(u64)MLX5_GET(create_tir_out, out,
1552 						      icm_address_39_32)
1553 					<< 32;
1554 				resp->tir_icm_addr |=
1555 					(u64)MLX5_GET(create_tir_out, out,
1556 						      icm_address_63_40)
1557 					<< 40;
1558 				resp->comp_mask |=
1559 					MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1560 			}
1561 		}
1562 	}
1563 
1564 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1565 						     rq->base.mqp.qpn;
1566 	return 0;
1567 
1568 err_destroy_rq:
1569 	destroy_raw_packet_qp_rq(dev, rq);
1570 err_destroy_sq:
1571 	if (!qp->sq.wqe_cnt)
1572 		return err;
1573 	destroy_raw_packet_qp_sq(dev, sq);
1574 err_destroy_tis:
1575 	destroy_raw_packet_qp_tis(dev, sq, pd);
1576 
1577 	return err;
1578 }
1579 
1580 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1581 				  struct mlx5_ib_qp *qp)
1582 {
1583 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1584 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1585 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1586 
1587 	if (qp->rq.wqe_cnt) {
1588 		destroy_raw_packet_qp_tir(dev, rq, qp->flags_en, qp->ibqp.pd);
1589 		destroy_raw_packet_qp_rq(dev, rq);
1590 	}
1591 
1592 	if (qp->sq.wqe_cnt) {
1593 		destroy_raw_packet_qp_sq(dev, sq);
1594 		destroy_raw_packet_qp_tis(dev, sq, qp->ibqp.pd);
1595 	}
1596 }
1597 
1598 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1599 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1600 {
1601 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1602 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1603 
1604 	sq->sq = &qp->sq;
1605 	rq->rq = &qp->rq;
1606 	sq->doorbell = &qp->db;
1607 	rq->doorbell = &qp->db;
1608 }
1609 
1610 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1611 {
1612 	if (qp->flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1613 			    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1614 		mlx5_ib_disable_lb(dev, false, true);
1615 	mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1616 			     to_mpd(qp->ibqp.pd)->uid);
1617 }
1618 
1619 struct mlx5_create_qp_params {
1620 	struct ib_udata *udata;
1621 	size_t inlen;
1622 	size_t outlen;
1623 	size_t ucmd_size;
1624 	void *ucmd;
1625 	u8 is_rss_raw : 1;
1626 	struct ib_qp_init_attr *attr;
1627 	u32 uidx;
1628 	struct mlx5_ib_create_qp_resp resp;
1629 };
1630 
1631 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1632 				 struct mlx5_ib_qp *qp,
1633 				 struct mlx5_create_qp_params *params)
1634 {
1635 	struct ib_qp_init_attr *init_attr = params->attr;
1636 	struct mlx5_ib_create_qp_rss *ucmd = params->ucmd;
1637 	struct ib_udata *udata = params->udata;
1638 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1639 		udata, struct mlx5_ib_ucontext, ibucontext);
1640 	int inlen;
1641 	int outlen;
1642 	int err;
1643 	u32 *in;
1644 	u32 *out;
1645 	void *tirc;
1646 	void *hfso;
1647 	u32 selected_fields = 0;
1648 	u32 outer_l4;
1649 	u32 tdn = mucontext->tdn;
1650 	u8 lb_flag = 0;
1651 
1652 	if (ucmd->comp_mask) {
1653 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1654 		return -EOPNOTSUPP;
1655 	}
1656 
1657 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_INNER &&
1658 	    !(ucmd->flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)) {
1659 		mlx5_ib_dbg(dev, "Tunnel offloads must be set for inner RSS\n");
1660 		return -EOPNOTSUPP;
1661 	}
1662 
1663 	if (dev->is_rep)
1664 		qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1665 
1666 	if (qp->flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1667 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1668 
1669 	if (qp->flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1670 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1671 
1672 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1673 	outlen = MLX5_ST_SZ_BYTES(create_tir_out);
1674 	in = kvzalloc(inlen + outlen, GFP_KERNEL);
1675 	if (!in)
1676 		return -ENOMEM;
1677 
1678 	out = in + MLX5_ST_SZ_DW(create_tir_in);
1679 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1680 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1681 	MLX5_SET(tirc, tirc, disp_type,
1682 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1683 	MLX5_SET(tirc, tirc, indirect_table,
1684 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1685 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1686 
1687 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1688 
1689 	if (ucmd->flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1690 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1691 
1692 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1693 
1694 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_INNER)
1695 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner);
1696 	else
1697 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1698 
1699 	switch (ucmd->rx_hash_function) {
1700 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1701 	{
1702 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1703 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1704 
1705 		if (len != ucmd->rx_key_len) {
1706 			err = -EINVAL;
1707 			goto err;
1708 		}
1709 
1710 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1711 		memcpy(rss_key, ucmd->rx_hash_key, len);
1712 		break;
1713 	}
1714 	default:
1715 		err = -EOPNOTSUPP;
1716 		goto err;
1717 	}
1718 
1719 	if (!ucmd->rx_hash_fields_mask) {
1720 		/* special case when this TIR serves as steering entry without hashing */
1721 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1722 			goto create_tir;
1723 		err = -EINVAL;
1724 		goto err;
1725 	}
1726 
1727 	if (((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1728 	     (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1729 	     ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1730 	     (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1731 		err = -EINVAL;
1732 		goto err;
1733 	}
1734 
1735 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1736 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1737 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1738 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1739 			 MLX5_L3_PROT_TYPE_IPV4);
1740 	else if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1741 		 (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1742 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1743 			 MLX5_L3_PROT_TYPE_IPV6);
1744 
1745 	outer_l4 = ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1746 		    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1747 			   << 0 |
1748 		   ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1749 		    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1750 			   << 1 |
1751 		   (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI) << 2;
1752 
1753 	/* Check that only one l4 protocol is set */
1754 	if (outer_l4 & (outer_l4 - 1)) {
1755 		err = -EINVAL;
1756 		goto err;
1757 	}
1758 
1759 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1760 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1761 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1762 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1763 			 MLX5_L4_PROT_TYPE_TCP);
1764 	else if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1765 		 (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1766 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1767 			 MLX5_L4_PROT_TYPE_UDP);
1768 
1769 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1770 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1771 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1772 
1773 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1774 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1775 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1776 
1777 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1778 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1779 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1780 
1781 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1782 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1783 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1784 
1785 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI)
1786 		selected_fields |= MLX5_HASH_FIELD_SEL_IPSEC_SPI;
1787 
1788 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1789 
1790 create_tir:
1791 	MLX5_SET(create_tir_in, in, opcode, MLX5_CMD_OP_CREATE_TIR);
1792 	err = mlx5_cmd_exec_inout(dev->mdev, create_tir, in, out);
1793 
1794 	qp->rss_qp.tirn = MLX5_GET(create_tir_out, out, tirn);
1795 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1796 		err = mlx5_ib_enable_lb(dev, false, true);
1797 
1798 		if (err)
1799 			mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1800 					     to_mpd(pd)->uid);
1801 	}
1802 
1803 	if (err)
1804 		goto err;
1805 
1806 	if (mucontext->devx_uid) {
1807 		params->resp.comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1808 		params->resp.tirn = qp->rss_qp.tirn;
1809 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
1810 		    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2)) {
1811 			params->resp.tir_icm_addr =
1812 				MLX5_GET(create_tir_out, out, icm_address_31_0);
1813 			params->resp.tir_icm_addr |=
1814 				(u64)MLX5_GET(create_tir_out, out,
1815 					      icm_address_39_32)
1816 				<< 32;
1817 			params->resp.tir_icm_addr |=
1818 				(u64)MLX5_GET(create_tir_out, out,
1819 					      icm_address_63_40)
1820 				<< 40;
1821 			params->resp.comp_mask |=
1822 				MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1823 		}
1824 	}
1825 
1826 	kvfree(in);
1827 	/* qpn is reserved for that QP */
1828 	qp->trans_qp.base.mqp.qpn = 0;
1829 	qp->is_rss = true;
1830 	return 0;
1831 
1832 err:
1833 	kvfree(in);
1834 	return err;
1835 }
1836 
1837 static void configure_requester_scat_cqe(struct mlx5_ib_dev *dev,
1838 					 struct mlx5_ib_qp *qp,
1839 					 struct ib_qp_init_attr *init_attr,
1840 					 void *qpc)
1841 {
1842 	int scqe_sz;
1843 	bool allow_scat_cqe = false;
1844 
1845 	allow_scat_cqe = qp->flags_en & MLX5_QP_FLAG_ALLOW_SCATTER_CQE;
1846 
1847 	if (!allow_scat_cqe && init_attr->sq_sig_type != IB_SIGNAL_ALL_WR)
1848 		return;
1849 
1850 	scqe_sz = mlx5_ib_get_cqe_size(init_attr->send_cq);
1851 	if (scqe_sz == 128) {
1852 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1853 		return;
1854 	}
1855 
1856 	if (init_attr->qp_type != MLX5_IB_QPT_DCI ||
1857 	    MLX5_CAP_GEN(dev->mdev, dc_req_scat_data_cqe))
1858 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1859 }
1860 
1861 static int atomic_size_to_mode(int size_mask)
1862 {
1863 	/* driver does not support atomic_size > 256B
1864 	 * and does not know how to translate bigger sizes
1865 	 */
1866 	int supported_size_mask = size_mask & 0x1ff;
1867 	int log_max_size;
1868 
1869 	if (!supported_size_mask)
1870 		return -EOPNOTSUPP;
1871 
1872 	log_max_size = __fls(supported_size_mask);
1873 
1874 	if (log_max_size > 3)
1875 		return log_max_size;
1876 
1877 	return MLX5_ATOMIC_MODE_8B;
1878 }
1879 
1880 static int get_atomic_mode(struct mlx5_ib_dev *dev,
1881 			   enum ib_qp_type qp_type)
1882 {
1883 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
1884 	u8 atomic = MLX5_CAP_GEN(dev->mdev, atomic);
1885 	int atomic_mode = -EOPNOTSUPP;
1886 	int atomic_size_mask;
1887 
1888 	if (!atomic)
1889 		return -EOPNOTSUPP;
1890 
1891 	if (qp_type == MLX5_IB_QPT_DCT)
1892 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
1893 	else
1894 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
1895 
1896 	if ((atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_CMP_SWAP) ||
1897 	    (atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_FETCH_ADD))
1898 		atomic_mode = atomic_size_to_mode(atomic_size_mask);
1899 
1900 	if (atomic_mode <= 0 &&
1901 	    (atomic_operations & MLX5_ATOMIC_OPS_CMP_SWAP &&
1902 	     atomic_operations & MLX5_ATOMIC_OPS_FETCH_ADD))
1903 		atomic_mode = MLX5_ATOMIC_MODE_IB_COMP;
1904 
1905 	return atomic_mode;
1906 }
1907 
1908 static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1909 			     struct mlx5_create_qp_params *params)
1910 {
1911 	struct ib_qp_init_attr *attr = params->attr;
1912 	u32 uidx = params->uidx;
1913 	struct mlx5_ib_resources *devr = &dev->devr;
1914 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
1915 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1916 	struct mlx5_core_dev *mdev = dev->mdev;
1917 	struct mlx5_ib_qp_base *base;
1918 	unsigned long flags;
1919 	void *qpc;
1920 	u32 *in;
1921 	int err;
1922 
1923 	if (attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1924 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1925 
1926 	in = kvzalloc(inlen, GFP_KERNEL);
1927 	if (!in)
1928 		return -ENOMEM;
1929 
1930 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1931 
1932 	MLX5_SET(qpc, qpc, st, MLX5_QP_ST_XRC);
1933 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
1934 	MLX5_SET(qpc, qpc, pd, to_mpd(devr->p0)->pdn);
1935 
1936 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1937 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
1938 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
1939 		MLX5_SET(qpc, qpc, cd_master, 1);
1940 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
1941 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
1942 	if (qp->flags & IB_QP_CREATE_MANAGED_RECV)
1943 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
1944 
1945 	MLX5_SET(qpc, qpc, ts_format, mlx5_get_qp_default_ts(dev->mdev));
1946 	MLX5_SET(qpc, qpc, rq_type, MLX5_SRQ_RQ);
1947 	MLX5_SET(qpc, qpc, no_sq, 1);
1948 	MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1949 	MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
1950 	MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
1951 	MLX5_SET(qpc, qpc, xrcd, to_mxrcd(attr->xrcd)->xrcdn);
1952 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
1953 
1954 	/* 0xffffff means we ask to work with cqe version 0 */
1955 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
1956 		MLX5_SET(qpc, qpc, user_index, uidx);
1957 
1958 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING) {
1959 		MLX5_SET(qpc, qpc, end_padding_mode,
1960 			 MLX5_WQ_END_PAD_MODE_ALIGN);
1961 		/* Special case to clean flag */
1962 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
1963 	}
1964 
1965 	base = &qp->trans_qp.base;
1966 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
1967 	kvfree(in);
1968 	if (err)
1969 		return err;
1970 
1971 	base->container_mibqp = qp;
1972 	base->mqp.event = mlx5_ib_qp_event;
1973 	if (MLX5_CAP_GEN(mdev, ece_support))
1974 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
1975 
1976 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1977 	list_add_tail(&qp->qps_list, &dev->qp_list);
1978 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1979 
1980 	qp->trans_qp.xrcdn = to_mxrcd(attr->xrcd)->xrcdn;
1981 	return 0;
1982 }
1983 
1984 static int create_dci(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1985 		      struct mlx5_ib_qp *qp,
1986 		      struct mlx5_create_qp_params *params)
1987 {
1988 	struct ib_qp_init_attr *init_attr = params->attr;
1989 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
1990 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
1991 	struct ib_udata *udata = params->udata;
1992 	u32 uidx = params->uidx;
1993 	struct mlx5_ib_resources *devr = &dev->devr;
1994 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1995 	struct mlx5_core_dev *mdev = dev->mdev;
1996 	struct mlx5_ib_cq *send_cq;
1997 	struct mlx5_ib_cq *recv_cq;
1998 	unsigned long flags;
1999 	struct mlx5_ib_qp_base *base;
2000 	int ts_format;
2001 	int mlx5_st;
2002 	void *qpc;
2003 	u32 *in;
2004 	int err;
2005 
2006 	spin_lock_init(&qp->sq.lock);
2007 	spin_lock_init(&qp->rq.lock);
2008 
2009 	mlx5_st = to_mlx5_st(qp->type);
2010 	if (mlx5_st < 0)
2011 		return -EINVAL;
2012 
2013 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2014 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2015 
2016 	base = &qp->trans_qp.base;
2017 
2018 	qp->has_rq = qp_has_rq(init_attr);
2019 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq, qp, ucmd);
2020 	if (err) {
2021 		mlx5_ib_dbg(dev, "err %d\n", err);
2022 		return err;
2023 	}
2024 
2025 	if (ucmd->rq_wqe_shift != qp->rq.wqe_shift ||
2026 	    ucmd->rq_wqe_count != qp->rq.wqe_cnt)
2027 		return -EINVAL;
2028 
2029 	if (ucmd->sq_wqe_count > (1 << MLX5_CAP_GEN(mdev, log_max_qp_sz)))
2030 		return -EINVAL;
2031 
2032 	ts_format = get_qp_ts_format(dev, to_mcq(init_attr->send_cq),
2033 				     to_mcq(init_attr->recv_cq));
2034 
2035 	if (ts_format < 0)
2036 		return ts_format;
2037 
2038 	err = _create_user_qp(dev, pd, qp, udata, init_attr, &in, &params->resp,
2039 			      &inlen, base, ucmd);
2040 	if (err)
2041 		return err;
2042 
2043 	if (MLX5_CAP_GEN(mdev, ece_support))
2044 		MLX5_SET(create_qp_in, in, ece, ucmd->ece_options);
2045 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2046 
2047 	MLX5_SET(qpc, qpc, st, mlx5_st);
2048 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2049 	MLX5_SET(qpc, qpc, pd, to_mpd(pd)->pdn);
2050 
2051 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE)
2052 		MLX5_SET(qpc, qpc, wq_signature, 1);
2053 
2054 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
2055 		MLX5_SET(qpc, qpc, cd_master, 1);
2056 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
2057 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
2058 	if (qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE)
2059 		configure_requester_scat_cqe(dev, qp, init_attr, qpc);
2060 
2061 	if (qp->rq.wqe_cnt) {
2062 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2063 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2064 	}
2065 
2066 	if (qp->flags_en & MLX5_QP_FLAG_DCI_STREAM) {
2067 		MLX5_SET(qpc, qpc, log_num_dci_stream_channels,
2068 			 ucmd->dci_streams.log_num_concurent);
2069 		MLX5_SET(qpc, qpc, log_num_dci_errored_streams,
2070 			 ucmd->dci_streams.log_num_errored);
2071 	}
2072 
2073 	MLX5_SET(qpc, qpc, ts_format, ts_format);
2074 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
2075 
2076 	MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2077 
2078 	/* Set default resources */
2079 	if (init_attr->srq) {
2080 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2081 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2082 			 to_msrq(init_attr->srq)->msrq.srqn);
2083 	} else {
2084 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2085 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2086 			 to_msrq(devr->s1)->msrq.srqn);
2087 	}
2088 
2089 	if (init_attr->send_cq)
2090 		MLX5_SET(qpc, qpc, cqn_snd,
2091 			 to_mcq(init_attr->send_cq)->mcq.cqn);
2092 
2093 	if (init_attr->recv_cq)
2094 		MLX5_SET(qpc, qpc, cqn_rcv,
2095 			 to_mcq(init_attr->recv_cq)->mcq.cqn);
2096 
2097 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2098 
2099 	/* 0xffffff means we ask to work with cqe version 0 */
2100 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2101 		MLX5_SET(qpc, qpc, user_index, uidx);
2102 
2103 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING) {
2104 		MLX5_SET(qpc, qpc, end_padding_mode,
2105 			 MLX5_WQ_END_PAD_MODE_ALIGN);
2106 		/* Special case to clean flag */
2107 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
2108 	}
2109 
2110 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2111 
2112 	kvfree(in);
2113 	if (err)
2114 		goto err_create;
2115 
2116 	base->container_mibqp = qp;
2117 	base->mqp.event = mlx5_ib_qp_event;
2118 	if (MLX5_CAP_GEN(mdev, ece_support))
2119 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
2120 
2121 	get_cqs(qp->type, init_attr->send_cq, init_attr->recv_cq,
2122 		&send_cq, &recv_cq);
2123 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2124 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2125 	/* Maintain device to QPs access, needed for further handling via reset
2126 	 * flow
2127 	 */
2128 	list_add_tail(&qp->qps_list, &dev->qp_list);
2129 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2130 	 */
2131 	if (send_cq)
2132 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2133 	if (recv_cq)
2134 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2135 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2136 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2137 
2138 	return 0;
2139 
2140 err_create:
2141 	destroy_qp(dev, qp, base, udata);
2142 	return err;
2143 }
2144 
2145 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2146 			  struct mlx5_ib_qp *qp,
2147 			  struct mlx5_create_qp_params *params)
2148 {
2149 	struct ib_qp_init_attr *init_attr = params->attr;
2150 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
2151 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
2152 	struct ib_udata *udata = params->udata;
2153 	u32 uidx = params->uidx;
2154 	struct mlx5_ib_resources *devr = &dev->devr;
2155 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
2156 	struct mlx5_core_dev *mdev = dev->mdev;
2157 	struct mlx5_ib_cq *send_cq;
2158 	struct mlx5_ib_cq *recv_cq;
2159 	unsigned long flags;
2160 	struct mlx5_ib_qp_base *base;
2161 	int ts_format;
2162 	int mlx5_st;
2163 	void *qpc;
2164 	u32 *in;
2165 	int err;
2166 
2167 	spin_lock_init(&qp->sq.lock);
2168 	spin_lock_init(&qp->rq.lock);
2169 
2170 	mlx5_st = to_mlx5_st(qp->type);
2171 	if (mlx5_st < 0)
2172 		return -EINVAL;
2173 
2174 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2175 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2176 
2177 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
2178 		qp->underlay_qpn = init_attr->source_qpn;
2179 
2180 	base = (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2181 		qp->flags & IB_QP_CREATE_SOURCE_QPN) ?
2182 	       &qp->raw_packet_qp.rq.base :
2183 	       &qp->trans_qp.base;
2184 
2185 	qp->has_rq = qp_has_rq(init_attr);
2186 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq, qp, ucmd);
2187 	if (err) {
2188 		mlx5_ib_dbg(dev, "err %d\n", err);
2189 		return err;
2190 	}
2191 
2192 	if (ucmd->rq_wqe_shift != qp->rq.wqe_shift ||
2193 	    ucmd->rq_wqe_count != qp->rq.wqe_cnt)
2194 		return -EINVAL;
2195 
2196 	if (ucmd->sq_wqe_count > (1 << MLX5_CAP_GEN(mdev, log_max_qp_sz)))
2197 		return -EINVAL;
2198 
2199 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2200 		ts_format = get_qp_ts_format(dev, to_mcq(init_attr->send_cq),
2201 					     to_mcq(init_attr->recv_cq));
2202 		if (ts_format < 0)
2203 			return ts_format;
2204 	}
2205 
2206 	err = _create_user_qp(dev, pd, qp, udata, init_attr, &in, &params->resp,
2207 			      &inlen, base, ucmd);
2208 	if (err)
2209 		return err;
2210 
2211 	if (is_sqp(init_attr->qp_type))
2212 		qp->port = init_attr->port_num;
2213 
2214 	if (MLX5_CAP_GEN(mdev, ece_support))
2215 		MLX5_SET(create_qp_in, in, ece, ucmd->ece_options);
2216 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2217 
2218 	MLX5_SET(qpc, qpc, st, mlx5_st);
2219 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2220 	MLX5_SET(qpc, qpc, pd, to_mpd(pd)->pdn);
2221 
2222 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE)
2223 		MLX5_SET(qpc, qpc, wq_signature, 1);
2224 
2225 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
2226 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2227 
2228 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
2229 		MLX5_SET(qpc, qpc, cd_master, 1);
2230 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
2231 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
2232 	if (qp->flags & IB_QP_CREATE_MANAGED_RECV)
2233 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
2234 	if (qp->flags_en & MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE)
2235 		MLX5_SET(qpc, qpc, req_e2e_credit_mode, 1);
2236 	if ((qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) &&
2237 	    (init_attr->qp_type == IB_QPT_RC ||
2238 	     init_attr->qp_type == IB_QPT_UC)) {
2239 		int rcqe_sz = mlx5_ib_get_cqe_size(init_attr->recv_cq);
2240 
2241 		MLX5_SET(qpc, qpc, cs_res,
2242 			 rcqe_sz == 128 ? MLX5_RES_SCAT_DATA64_CQE :
2243 					  MLX5_RES_SCAT_DATA32_CQE);
2244 	}
2245 	if ((qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) &&
2246 	    (qp->type == MLX5_IB_QPT_DCI || qp->type == IB_QPT_RC))
2247 		configure_requester_scat_cqe(dev, qp, init_attr, qpc);
2248 
2249 	if (qp->rq.wqe_cnt) {
2250 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2251 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2252 	}
2253 
2254 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
2255 		MLX5_SET(qpc, qpc, ts_format, ts_format);
2256 
2257 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
2258 
2259 	if (qp->sq.wqe_cnt) {
2260 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2261 	} else {
2262 		MLX5_SET(qpc, qpc, no_sq, 1);
2263 		if (init_attr->srq &&
2264 		    init_attr->srq->srq_type == IB_SRQT_TM)
2265 			MLX5_SET(qpc, qpc, offload_type,
2266 				 MLX5_QPC_OFFLOAD_TYPE_RNDV);
2267 	}
2268 
2269 	/* Set default resources */
2270 	switch (init_attr->qp_type) {
2271 	case IB_QPT_XRC_INI:
2272 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2273 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2274 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2275 		break;
2276 	default:
2277 		if (init_attr->srq) {
2278 			MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2279 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(init_attr->srq)->msrq.srqn);
2280 		} else {
2281 			MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2282 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s1)->msrq.srqn);
2283 		}
2284 	}
2285 
2286 	if (init_attr->send_cq)
2287 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
2288 
2289 	if (init_attr->recv_cq)
2290 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
2291 
2292 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2293 
2294 	/* 0xffffff means we ask to work with cqe version 0 */
2295 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2296 		MLX5_SET(qpc, qpc, user_index, uidx);
2297 
2298 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING &&
2299 	    init_attr->qp_type != IB_QPT_RAW_PACKET) {
2300 		MLX5_SET(qpc, qpc, end_padding_mode,
2301 			 MLX5_WQ_END_PAD_MODE_ALIGN);
2302 		/* Special case to clean flag */
2303 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
2304 	}
2305 
2306 	if (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2307 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
2308 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd->sq_buf_addr;
2309 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
2310 		err = create_raw_packet_qp(dev, qp, in, inlen, pd, udata,
2311 					   &params->resp, init_attr);
2312 	} else
2313 		err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2314 
2315 	kvfree(in);
2316 	if (err)
2317 		goto err_create;
2318 
2319 	base->container_mibqp = qp;
2320 	base->mqp.event = mlx5_ib_qp_event;
2321 	if (MLX5_CAP_GEN(mdev, ece_support))
2322 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
2323 
2324 	get_cqs(qp->type, init_attr->send_cq, init_attr->recv_cq,
2325 		&send_cq, &recv_cq);
2326 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2327 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2328 	/* Maintain device to QPs access, needed for further handling via reset
2329 	 * flow
2330 	 */
2331 	list_add_tail(&qp->qps_list, &dev->qp_list);
2332 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2333 	 */
2334 	if (send_cq)
2335 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2336 	if (recv_cq)
2337 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2338 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2339 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2340 
2341 	return 0;
2342 
2343 err_create:
2344 	destroy_qp(dev, qp, base, udata);
2345 	return err;
2346 }
2347 
2348 static int create_kernel_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2349 			    struct mlx5_ib_qp *qp,
2350 			    struct mlx5_create_qp_params *params)
2351 {
2352 	struct ib_qp_init_attr *attr = params->attr;
2353 	u32 uidx = params->uidx;
2354 	struct mlx5_ib_resources *devr = &dev->devr;
2355 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
2356 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
2357 	struct mlx5_core_dev *mdev = dev->mdev;
2358 	struct mlx5_ib_cq *send_cq;
2359 	struct mlx5_ib_cq *recv_cq;
2360 	unsigned long flags;
2361 	struct mlx5_ib_qp_base *base;
2362 	int mlx5_st;
2363 	void *qpc;
2364 	u32 *in;
2365 	int err;
2366 
2367 	spin_lock_init(&qp->sq.lock);
2368 	spin_lock_init(&qp->rq.lock);
2369 
2370 	mlx5_st = to_mlx5_st(qp->type);
2371 	if (mlx5_st < 0)
2372 		return -EINVAL;
2373 
2374 	if (attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2375 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2376 
2377 	base = &qp->trans_qp.base;
2378 
2379 	qp->has_rq = qp_has_rq(attr);
2380 	err = set_rq_size(dev, &attr->cap, qp->has_rq, qp, NULL);
2381 	if (err) {
2382 		mlx5_ib_dbg(dev, "err %d\n", err);
2383 		return err;
2384 	}
2385 
2386 	err = _create_kernel_qp(dev, attr, qp, &in, &inlen, base);
2387 	if (err)
2388 		return err;
2389 
2390 	if (is_sqp(attr->qp_type))
2391 		qp->port = attr->port_num;
2392 
2393 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2394 
2395 	MLX5_SET(qpc, qpc, st, mlx5_st);
2396 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2397 
2398 	if (attr->qp_type != MLX5_IB_QPT_REG_UMR)
2399 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
2400 	else
2401 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
2402 
2403 
2404 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
2405 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2406 
2407 	if (qp->rq.wqe_cnt) {
2408 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2409 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2410 	}
2411 
2412 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, attr));
2413 
2414 	if (qp->sq.wqe_cnt)
2415 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2416 	else
2417 		MLX5_SET(qpc, qpc, no_sq, 1);
2418 
2419 	if (attr->srq) {
2420 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2421 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2422 			 to_msrq(attr->srq)->msrq.srqn);
2423 	} else {
2424 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2425 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2426 			 to_msrq(devr->s1)->msrq.srqn);
2427 	}
2428 
2429 	if (attr->send_cq)
2430 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(attr->send_cq)->mcq.cqn);
2431 
2432 	if (attr->recv_cq)
2433 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(attr->recv_cq)->mcq.cqn);
2434 
2435 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2436 
2437 	/* 0xffffff means we ask to work with cqe version 0 */
2438 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2439 		MLX5_SET(qpc, qpc, user_index, uidx);
2440 
2441 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
2442 	if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO)
2443 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
2444 
2445 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2446 	kvfree(in);
2447 	if (err)
2448 		goto err_create;
2449 
2450 	base->container_mibqp = qp;
2451 	base->mqp.event = mlx5_ib_qp_event;
2452 
2453 	get_cqs(qp->type, attr->send_cq, attr->recv_cq,
2454 		&send_cq, &recv_cq);
2455 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2456 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2457 	/* Maintain device to QPs access, needed for further handling via reset
2458 	 * flow
2459 	 */
2460 	list_add_tail(&qp->qps_list, &dev->qp_list);
2461 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2462 	 */
2463 	if (send_cq)
2464 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2465 	if (recv_cq)
2466 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2467 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2468 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2469 
2470 	return 0;
2471 
2472 err_create:
2473 	destroy_qp(dev, qp, base, NULL);
2474 	return err;
2475 }
2476 
2477 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2478 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
2479 {
2480 	if (send_cq) {
2481 		if (recv_cq) {
2482 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2483 				spin_lock(&send_cq->lock);
2484 				spin_lock_nested(&recv_cq->lock,
2485 						 SINGLE_DEPTH_NESTING);
2486 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2487 				spin_lock(&send_cq->lock);
2488 				__acquire(&recv_cq->lock);
2489 			} else {
2490 				spin_lock(&recv_cq->lock);
2491 				spin_lock_nested(&send_cq->lock,
2492 						 SINGLE_DEPTH_NESTING);
2493 			}
2494 		} else {
2495 			spin_lock(&send_cq->lock);
2496 			__acquire(&recv_cq->lock);
2497 		}
2498 	} else if (recv_cq) {
2499 		spin_lock(&recv_cq->lock);
2500 		__acquire(&send_cq->lock);
2501 	} else {
2502 		__acquire(&send_cq->lock);
2503 		__acquire(&recv_cq->lock);
2504 	}
2505 }
2506 
2507 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2508 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
2509 {
2510 	if (send_cq) {
2511 		if (recv_cq) {
2512 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2513 				spin_unlock(&recv_cq->lock);
2514 				spin_unlock(&send_cq->lock);
2515 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2516 				__release(&recv_cq->lock);
2517 				spin_unlock(&send_cq->lock);
2518 			} else {
2519 				spin_unlock(&send_cq->lock);
2520 				spin_unlock(&recv_cq->lock);
2521 			}
2522 		} else {
2523 			__release(&recv_cq->lock);
2524 			spin_unlock(&send_cq->lock);
2525 		}
2526 	} else if (recv_cq) {
2527 		__release(&send_cq->lock);
2528 		spin_unlock(&recv_cq->lock);
2529 	} else {
2530 		__release(&recv_cq->lock);
2531 		__release(&send_cq->lock);
2532 	}
2533 }
2534 
2535 static void get_cqs(enum ib_qp_type qp_type,
2536 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
2537 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
2538 {
2539 	switch (qp_type) {
2540 	case IB_QPT_XRC_TGT:
2541 		*send_cq = NULL;
2542 		*recv_cq = NULL;
2543 		break;
2544 	case MLX5_IB_QPT_REG_UMR:
2545 	case IB_QPT_XRC_INI:
2546 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2547 		*recv_cq = NULL;
2548 		break;
2549 
2550 	case IB_QPT_SMI:
2551 	case MLX5_IB_QPT_HW_GSI:
2552 	case IB_QPT_RC:
2553 	case IB_QPT_UC:
2554 	case IB_QPT_UD:
2555 	case IB_QPT_RAW_PACKET:
2556 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2557 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
2558 		break;
2559 	default:
2560 		*send_cq = NULL;
2561 		*recv_cq = NULL;
2562 		break;
2563 	}
2564 }
2565 
2566 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2567 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2568 				u8 lag_tx_affinity);
2569 
2570 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2571 			      struct ib_udata *udata)
2572 {
2573 	struct mlx5_ib_cq *send_cq, *recv_cq;
2574 	struct mlx5_ib_qp_base *base;
2575 	unsigned long flags;
2576 	int err;
2577 
2578 	if (qp->is_rss) {
2579 		destroy_rss_raw_qp_tir(dev, qp);
2580 		return;
2581 	}
2582 
2583 	base = (qp->type == IB_QPT_RAW_PACKET ||
2584 		qp->flags & IB_QP_CREATE_SOURCE_QPN) ?
2585 		       &qp->raw_packet_qp.rq.base :
2586 		       &qp->trans_qp.base;
2587 
2588 	if (qp->state != IB_QPS_RESET) {
2589 		if (qp->type != IB_QPT_RAW_PACKET &&
2590 		    !(qp->flags & IB_QP_CREATE_SOURCE_QPN)) {
2591 			err = mlx5_core_qp_modify(dev, MLX5_CMD_OP_2RST_QP, 0,
2592 						  NULL, &base->mqp, NULL);
2593 		} else {
2594 			struct mlx5_modify_raw_qp_param raw_qp_param = {
2595 				.operation = MLX5_CMD_OP_2RST_QP
2596 			};
2597 
2598 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2599 		}
2600 		if (err)
2601 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
2602 				     base->mqp.qpn);
2603 	}
2604 
2605 	get_cqs(qp->type, qp->ibqp.send_cq, qp->ibqp.recv_cq, &send_cq,
2606 		&recv_cq);
2607 
2608 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2609 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2610 	/* del from lists under both locks above to protect reset flow paths */
2611 	list_del(&qp->qps_list);
2612 	if (send_cq)
2613 		list_del(&qp->cq_send_list);
2614 
2615 	if (recv_cq)
2616 		list_del(&qp->cq_recv_list);
2617 
2618 	if (!udata) {
2619 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2620 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
2621 		if (send_cq != recv_cq)
2622 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
2623 					   NULL);
2624 	}
2625 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2626 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2627 
2628 	if (qp->type == IB_QPT_RAW_PACKET ||
2629 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
2630 		destroy_raw_packet_qp(dev, qp);
2631 	} else {
2632 		err = mlx5_core_destroy_qp(dev, &base->mqp);
2633 		if (err)
2634 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
2635 				     base->mqp.qpn);
2636 	}
2637 
2638 	destroy_qp(dev, qp, base, udata);
2639 }
2640 
2641 static int create_dct(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2642 		      struct mlx5_ib_qp *qp,
2643 		      struct mlx5_create_qp_params *params)
2644 {
2645 	struct ib_qp_init_attr *attr = params->attr;
2646 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
2647 	u32 uidx = params->uidx;
2648 	void *dctc;
2649 
2650 	if (mlx5_lag_is_active(dev->mdev) && !MLX5_CAP_GEN(dev->mdev, lag_dct))
2651 		return -EOPNOTSUPP;
2652 
2653 	qp->dct.in = kzalloc(MLX5_ST_SZ_BYTES(create_dct_in), GFP_KERNEL);
2654 	if (!qp->dct.in)
2655 		return -ENOMEM;
2656 
2657 	MLX5_SET(create_dct_in, qp->dct.in, uid, to_mpd(pd)->uid);
2658 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
2659 	MLX5_SET(dctc, dctc, pd, to_mpd(pd)->pdn);
2660 	MLX5_SET(dctc, dctc, srqn_xrqn, to_msrq(attr->srq)->msrq.srqn);
2661 	MLX5_SET(dctc, dctc, cqn, to_mcq(attr->recv_cq)->mcq.cqn);
2662 	MLX5_SET64(dctc, dctc, dc_access_key, ucmd->access_key);
2663 	MLX5_SET(dctc, dctc, user_index, uidx);
2664 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
2665 		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
2666 
2667 	if (qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) {
2668 		int rcqe_sz = mlx5_ib_get_cqe_size(attr->recv_cq);
2669 
2670 		if (rcqe_sz == 128)
2671 			MLX5_SET(dctc, dctc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
2672 	}
2673 
2674 	qp->state = IB_QPS_RESET;
2675 	return 0;
2676 }
2677 
2678 static int check_qp_type(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
2679 			 enum ib_qp_type *type)
2680 {
2681 	if (attr->qp_type == IB_QPT_DRIVER && !MLX5_CAP_GEN(dev->mdev, dct))
2682 		goto out;
2683 
2684 	switch (attr->qp_type) {
2685 	case IB_QPT_XRC_TGT:
2686 	case IB_QPT_XRC_INI:
2687 		if (!MLX5_CAP_GEN(dev->mdev, xrc))
2688 			goto out;
2689 		fallthrough;
2690 	case IB_QPT_RC:
2691 	case IB_QPT_UC:
2692 	case IB_QPT_SMI:
2693 	case MLX5_IB_QPT_HW_GSI:
2694 	case IB_QPT_DRIVER:
2695 	case IB_QPT_GSI:
2696 	case IB_QPT_RAW_PACKET:
2697 	case IB_QPT_UD:
2698 	case MLX5_IB_QPT_REG_UMR:
2699 		break;
2700 	default:
2701 		goto out;
2702 	}
2703 
2704 	*type = attr->qp_type;
2705 	return 0;
2706 
2707 out:
2708 	mlx5_ib_dbg(dev, "Unsupported QP type %d\n", attr->qp_type);
2709 	return -EOPNOTSUPP;
2710 }
2711 
2712 static int check_valid_flow(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2713 			    struct ib_qp_init_attr *attr,
2714 			    struct ib_udata *udata)
2715 {
2716 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2717 		udata, struct mlx5_ib_ucontext, ibucontext);
2718 
2719 	if (!udata) {
2720 		/* Kernel create_qp callers */
2721 		if (attr->rwq_ind_tbl)
2722 			return -EOPNOTSUPP;
2723 
2724 		switch (attr->qp_type) {
2725 		case IB_QPT_RAW_PACKET:
2726 		case IB_QPT_DRIVER:
2727 			return -EOPNOTSUPP;
2728 		default:
2729 			return 0;
2730 		}
2731 	}
2732 
2733 	/* Userspace create_qp callers */
2734 	if (attr->qp_type == IB_QPT_RAW_PACKET && !ucontext->cqe_version) {
2735 		mlx5_ib_dbg(dev,
2736 			"Raw Packet QP is only supported for CQE version > 0\n");
2737 		return -EINVAL;
2738 	}
2739 
2740 	if (attr->qp_type != IB_QPT_RAW_PACKET && attr->rwq_ind_tbl) {
2741 		mlx5_ib_dbg(dev,
2742 			    "Wrong QP type %d for the RWQ indirect table\n",
2743 			    attr->qp_type);
2744 		return -EINVAL;
2745 	}
2746 
2747 	/*
2748 	 * We don't need to see this warning, it means that kernel code
2749 	 * missing ib_pd. Placed here to catch developer's mistakes.
2750 	 */
2751 	WARN_ONCE(!pd && attr->qp_type != IB_QPT_XRC_TGT,
2752 		  "There is a missing PD pointer assignment\n");
2753 	return 0;
2754 }
2755 
2756 static void process_vendor_flag(struct mlx5_ib_dev *dev, int *flags, int flag,
2757 				bool cond, struct mlx5_ib_qp *qp)
2758 {
2759 	if (!(*flags & flag))
2760 		return;
2761 
2762 	if (cond) {
2763 		qp->flags_en |= flag;
2764 		*flags &= ~flag;
2765 		return;
2766 	}
2767 
2768 	switch (flag) {
2769 	case MLX5_QP_FLAG_SCATTER_CQE:
2770 	case MLX5_QP_FLAG_ALLOW_SCATTER_CQE:
2771 		/*
2772 			 * We don't return error if these flags were provided,
2773 			 * and mlx5 doesn't have right capability.
2774 			 */
2775 		*flags &= ~(MLX5_QP_FLAG_SCATTER_CQE |
2776 			    MLX5_QP_FLAG_ALLOW_SCATTER_CQE);
2777 		return;
2778 	default:
2779 		break;
2780 	}
2781 	mlx5_ib_dbg(dev, "Vendor create QP flag 0x%X is not supported\n", flag);
2782 }
2783 
2784 static int process_vendor_flags(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2785 				void *ucmd, struct ib_qp_init_attr *attr)
2786 {
2787 	struct mlx5_core_dev *mdev = dev->mdev;
2788 	bool cond;
2789 	int flags;
2790 
2791 	if (attr->rwq_ind_tbl)
2792 		flags = ((struct mlx5_ib_create_qp_rss *)ucmd)->flags;
2793 	else
2794 		flags = ((struct mlx5_ib_create_qp *)ucmd)->flags;
2795 
2796 	switch (flags & (MLX5_QP_FLAG_TYPE_DCT | MLX5_QP_FLAG_TYPE_DCI)) {
2797 	case MLX5_QP_FLAG_TYPE_DCI:
2798 		qp->type = MLX5_IB_QPT_DCI;
2799 		break;
2800 	case MLX5_QP_FLAG_TYPE_DCT:
2801 		qp->type = MLX5_IB_QPT_DCT;
2802 		break;
2803 	default:
2804 		if (qp->type != IB_QPT_DRIVER)
2805 			break;
2806 		/*
2807 		 * It is IB_QPT_DRIVER and or no subtype or
2808 		 * wrong subtype were provided.
2809 		 */
2810 		return -EINVAL;
2811 	}
2812 
2813 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TYPE_DCI, true, qp);
2814 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TYPE_DCT, true, qp);
2815 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_DCI_STREAM,
2816 			    MLX5_CAP_GEN(mdev, log_max_dci_stream_channels),
2817 			    qp);
2818 
2819 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_SIGNATURE, true, qp);
2820 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_SCATTER_CQE,
2821 			    MLX5_CAP_GEN(mdev, sctr_data_cqe), qp);
2822 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_ALLOW_SCATTER_CQE,
2823 			    MLX5_CAP_GEN(mdev, sctr_data_cqe), qp);
2824 
2825 	if (qp->type == IB_QPT_RAW_PACKET) {
2826 		cond = MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan) ||
2827 		       MLX5_CAP_ETH(mdev, tunnel_stateless_gre) ||
2828 		       MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx);
2829 		process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TUNNEL_OFFLOADS,
2830 				    cond, qp);
2831 		process_vendor_flag(dev, &flags,
2832 				    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC, true,
2833 				    qp);
2834 		process_vendor_flag(dev, &flags,
2835 				    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC, true,
2836 				    qp);
2837 	}
2838 
2839 	if (qp->type == IB_QPT_RC)
2840 		process_vendor_flag(dev, &flags,
2841 				    MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE,
2842 				    MLX5_CAP_GEN(mdev, qp_packet_based), qp);
2843 
2844 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_BFREG_INDEX, true, qp);
2845 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_UAR_PAGE_INDEX, true, qp);
2846 
2847 	cond = qp->flags_en & ~(MLX5_QP_FLAG_TUNNEL_OFFLOADS |
2848 				MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
2849 				MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC);
2850 	if (attr->rwq_ind_tbl && cond) {
2851 		mlx5_ib_dbg(dev, "RSS RAW QP has unsupported flags 0x%X\n",
2852 			    cond);
2853 		return -EINVAL;
2854 	}
2855 
2856 	if (flags)
2857 		mlx5_ib_dbg(dev, "udata has unsupported flags 0x%X\n", flags);
2858 
2859 	return (flags) ? -EINVAL : 0;
2860 	}
2861 
2862 static void process_create_flag(struct mlx5_ib_dev *dev, int *flags, int flag,
2863 				bool cond, struct mlx5_ib_qp *qp)
2864 {
2865 	if (!(*flags & flag))
2866 		return;
2867 
2868 	if (cond) {
2869 		qp->flags |= flag;
2870 		*flags &= ~flag;
2871 		return;
2872 	}
2873 
2874 	if (flag == MLX5_IB_QP_CREATE_WC_TEST) {
2875 		/*
2876 		 * Special case, if condition didn't meet, it won't be error,
2877 		 * just different in-kernel flow.
2878 		 */
2879 		*flags &= ~MLX5_IB_QP_CREATE_WC_TEST;
2880 		return;
2881 	}
2882 	mlx5_ib_dbg(dev, "Verbs create QP flag 0x%X is not supported\n", flag);
2883 }
2884 
2885 static int process_create_flags(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2886 				struct ib_qp_init_attr *attr)
2887 {
2888 	enum ib_qp_type qp_type = qp->type;
2889 	struct mlx5_core_dev *mdev = dev->mdev;
2890 	int create_flags = attr->create_flags;
2891 	bool cond;
2892 
2893 	if (qp_type == MLX5_IB_QPT_DCT)
2894 		return (create_flags) ? -EINVAL : 0;
2895 
2896 	if (qp_type == IB_QPT_RAW_PACKET && attr->rwq_ind_tbl)
2897 		return (create_flags) ? -EINVAL : 0;
2898 
2899 	process_create_flag(dev, &create_flags, IB_QP_CREATE_NETIF_QP,
2900 			    mlx5_get_flow_namespace(dev->mdev,
2901 						    MLX5_FLOW_NAMESPACE_BYPASS),
2902 			    qp);
2903 	process_create_flag(dev, &create_flags,
2904 			    IB_QP_CREATE_INTEGRITY_EN,
2905 			    MLX5_CAP_GEN(mdev, sho), qp);
2906 	process_create_flag(dev, &create_flags,
2907 			    IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK,
2908 			    MLX5_CAP_GEN(mdev, block_lb_mc), qp);
2909 	process_create_flag(dev, &create_flags, IB_QP_CREATE_CROSS_CHANNEL,
2910 			    MLX5_CAP_GEN(mdev, cd), qp);
2911 	process_create_flag(dev, &create_flags, IB_QP_CREATE_MANAGED_SEND,
2912 			    MLX5_CAP_GEN(mdev, cd), qp);
2913 	process_create_flag(dev, &create_flags, IB_QP_CREATE_MANAGED_RECV,
2914 			    MLX5_CAP_GEN(mdev, cd), qp);
2915 
2916 	if (qp_type == IB_QPT_UD) {
2917 		process_create_flag(dev, &create_flags,
2918 				    IB_QP_CREATE_IPOIB_UD_LSO,
2919 				    MLX5_CAP_GEN(mdev, ipoib_basic_offloads),
2920 				    qp);
2921 		cond = MLX5_CAP_GEN(mdev, port_type) == MLX5_CAP_PORT_TYPE_IB;
2922 		process_create_flag(dev, &create_flags, IB_QP_CREATE_SOURCE_QPN,
2923 				    cond, qp);
2924 	}
2925 
2926 	if (qp_type == IB_QPT_RAW_PACKET) {
2927 		cond = MLX5_CAP_GEN(mdev, eth_net_offloads) &&
2928 		       MLX5_CAP_ETH(mdev, scatter_fcs);
2929 		process_create_flag(dev, &create_flags,
2930 				    IB_QP_CREATE_SCATTER_FCS, cond, qp);
2931 
2932 		cond = MLX5_CAP_GEN(mdev, eth_net_offloads) &&
2933 		       MLX5_CAP_ETH(mdev, vlan_cap);
2934 		process_create_flag(dev, &create_flags,
2935 				    IB_QP_CREATE_CVLAN_STRIPPING, cond, qp);
2936 	}
2937 
2938 	process_create_flag(dev, &create_flags,
2939 			    IB_QP_CREATE_PCI_WRITE_END_PADDING,
2940 			    MLX5_CAP_GEN(mdev, end_pad), qp);
2941 
2942 	process_create_flag(dev, &create_flags, MLX5_IB_QP_CREATE_WC_TEST,
2943 			    qp_type != MLX5_IB_QPT_REG_UMR, qp);
2944 	process_create_flag(dev, &create_flags, MLX5_IB_QP_CREATE_SQPN_QP1,
2945 			    true, qp);
2946 
2947 	if (create_flags) {
2948 		mlx5_ib_dbg(dev, "Create QP has unsupported flags 0x%X\n",
2949 			    create_flags);
2950 		return -EOPNOTSUPP;
2951 	}
2952 	return 0;
2953 }
2954 
2955 static int process_udata_size(struct mlx5_ib_dev *dev,
2956 			      struct mlx5_create_qp_params *params)
2957 {
2958 	size_t ucmd = sizeof(struct mlx5_ib_create_qp);
2959 	struct ib_udata *udata = params->udata;
2960 	size_t outlen = udata->outlen;
2961 	size_t inlen = udata->inlen;
2962 
2963 	params->outlen = min(outlen, sizeof(struct mlx5_ib_create_qp_resp));
2964 	params->ucmd_size = ucmd;
2965 	if (!params->is_rss_raw) {
2966 		/* User has old rdma-core, which doesn't support ECE */
2967 		size_t min_inlen =
2968 			offsetof(struct mlx5_ib_create_qp, ece_options);
2969 
2970 		/*
2971 		 * We will check in check_ucmd_data() that user
2972 		 * cleared everything after inlen.
2973 		 */
2974 		params->inlen = (inlen < min_inlen) ? 0 : min(inlen, ucmd);
2975 		goto out;
2976 	}
2977 
2978 	/* RSS RAW QP */
2979 	if (inlen < offsetofend(struct mlx5_ib_create_qp_rss, flags))
2980 		return -EINVAL;
2981 
2982 	if (outlen < offsetofend(struct mlx5_ib_create_qp_resp, bfreg_index))
2983 		return -EINVAL;
2984 
2985 	ucmd = sizeof(struct mlx5_ib_create_qp_rss);
2986 	params->ucmd_size = ucmd;
2987 	if (inlen > ucmd && !ib_is_udata_cleared(udata, ucmd, inlen - ucmd))
2988 		return -EINVAL;
2989 
2990 	params->inlen = min(ucmd, inlen);
2991 out:
2992 	if (!params->inlen)
2993 		mlx5_ib_dbg(dev, "udata is too small\n");
2994 
2995 	return (params->inlen) ? 0 : -EINVAL;
2996 }
2997 
2998 static int create_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2999 		     struct mlx5_ib_qp *qp,
3000 		     struct mlx5_create_qp_params *params)
3001 {
3002 	int err;
3003 
3004 	if (params->is_rss_raw) {
3005 		err = create_rss_raw_qp_tir(dev, pd, qp, params);
3006 		goto out;
3007 	}
3008 
3009 	switch (qp->type) {
3010 	case MLX5_IB_QPT_DCT:
3011 		err = create_dct(dev, pd, qp, params);
3012 		rdma_restrack_no_track(&qp->ibqp.res);
3013 		break;
3014 	case MLX5_IB_QPT_DCI:
3015 		err = create_dci(dev, pd, qp, params);
3016 		break;
3017 	case IB_QPT_XRC_TGT:
3018 		err = create_xrc_tgt_qp(dev, qp, params);
3019 		break;
3020 	case IB_QPT_GSI:
3021 		err = mlx5_ib_create_gsi(pd, qp, params->attr);
3022 		break;
3023 	case MLX5_IB_QPT_HW_GSI:
3024 	case MLX5_IB_QPT_REG_UMR:
3025 		rdma_restrack_no_track(&qp->ibqp.res);
3026 		fallthrough;
3027 	default:
3028 		if (params->udata)
3029 			err = create_user_qp(dev, pd, qp, params);
3030 		else
3031 			err = create_kernel_qp(dev, pd, qp, params);
3032 	}
3033 
3034 out:
3035 	if (err) {
3036 		mlx5_ib_err(dev, "Create QP type %d failed\n", qp->type);
3037 		return err;
3038 	}
3039 
3040 	if (is_qp0(qp->type))
3041 		qp->ibqp.qp_num = 0;
3042 	else if (is_qp1(qp->type))
3043 		qp->ibqp.qp_num = 1;
3044 	else
3045 		qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
3046 
3047 	mlx5_ib_dbg(dev,
3048 		"QP type %d, ib qpn 0x%X, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x, ece 0x%x\n",
3049 		qp->type, qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
3050 		params->attr->recv_cq ? to_mcq(params->attr->recv_cq)->mcq.cqn :
3051 					-1,
3052 		params->attr->send_cq ? to_mcq(params->attr->send_cq)->mcq.cqn :
3053 					-1,
3054 		params->resp.ece_options);
3055 
3056 	return 0;
3057 }
3058 
3059 static int check_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3060 			 struct ib_qp_init_attr *attr)
3061 {
3062 	int ret = 0;
3063 
3064 	switch (qp->type) {
3065 	case MLX5_IB_QPT_DCT:
3066 		ret = (!attr->srq || !attr->recv_cq) ? -EINVAL : 0;
3067 		break;
3068 	case MLX5_IB_QPT_DCI:
3069 		ret = (attr->cap.max_recv_wr || attr->cap.max_recv_sge) ?
3070 			      -EINVAL :
3071 			      0;
3072 		break;
3073 	case IB_QPT_RAW_PACKET:
3074 		ret = (attr->rwq_ind_tbl && attr->send_cq) ? -EINVAL : 0;
3075 		break;
3076 	default:
3077 		break;
3078 	}
3079 
3080 	if (ret)
3081 		mlx5_ib_dbg(dev, "QP type %d has wrong attributes\n", qp->type);
3082 
3083 	return ret;
3084 }
3085 
3086 static int get_qp_uidx(struct mlx5_ib_qp *qp,
3087 		       struct mlx5_create_qp_params *params)
3088 {
3089 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
3090 	struct ib_udata *udata = params->udata;
3091 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3092 		udata, struct mlx5_ib_ucontext, ibucontext);
3093 
3094 	if (params->is_rss_raw)
3095 		return 0;
3096 
3097 	return get_qp_user_index(ucontext, ucmd, sizeof(*ucmd), &params->uidx);
3098 }
3099 
3100 static int mlx5_ib_destroy_dct(struct mlx5_ib_qp *mqp)
3101 {
3102 	struct mlx5_ib_dev *dev = to_mdev(mqp->ibqp.device);
3103 
3104 	if (mqp->state == IB_QPS_RTR) {
3105 		int err;
3106 
3107 		err = mlx5_core_destroy_dct(dev, &mqp->dct.mdct);
3108 		if (err) {
3109 			mlx5_ib_warn(dev, "failed to destroy DCT %d\n", err);
3110 			return err;
3111 		}
3112 	}
3113 
3114 	kfree(mqp->dct.in);
3115 	return 0;
3116 }
3117 
3118 static int check_ucmd_data(struct mlx5_ib_dev *dev,
3119 			   struct mlx5_create_qp_params *params)
3120 {
3121 	struct ib_udata *udata = params->udata;
3122 	size_t size, last;
3123 	int ret;
3124 
3125 	if (params->is_rss_raw)
3126 		/*
3127 		 * These QPs don't have "reserved" field in their
3128 		 * create_qp input struct, so their data is always valid.
3129 		 */
3130 		last = sizeof(struct mlx5_ib_create_qp_rss);
3131 	else
3132 		last = offsetof(struct mlx5_ib_create_qp, reserved);
3133 
3134 	if (udata->inlen <= last)
3135 		return 0;
3136 
3137 	/*
3138 	 * User provides different create_qp structures based on the
3139 	 * flow and we need to know if he cleared memory after our
3140 	 * struct create_qp ends.
3141 	 */
3142 	size = udata->inlen - last;
3143 	ret = ib_is_udata_cleared(params->udata, last, size);
3144 	if (!ret)
3145 		mlx5_ib_dbg(
3146 			dev,
3147 			"udata is not cleared, inlen = %zu, ucmd = %zu, last = %zu, size = %zu\n",
3148 			udata->inlen, params->ucmd_size, last, size);
3149 	return ret ? 0 : -EINVAL;
3150 }
3151 
3152 int mlx5_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
3153 		      struct ib_udata *udata)
3154 {
3155 	struct mlx5_create_qp_params params = {};
3156 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3157 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3158 	struct ib_pd *pd = ibqp->pd;
3159 	enum ib_qp_type type;
3160 	int err;
3161 
3162 	err = check_qp_type(dev, attr, &type);
3163 	if (err)
3164 		return err;
3165 
3166 	err = check_valid_flow(dev, pd, attr, udata);
3167 	if (err)
3168 		return err;
3169 
3170 	params.udata = udata;
3171 	params.uidx = MLX5_IB_DEFAULT_UIDX;
3172 	params.attr = attr;
3173 	params.is_rss_raw = !!attr->rwq_ind_tbl;
3174 
3175 	if (udata) {
3176 		err = process_udata_size(dev, &params);
3177 		if (err)
3178 			return err;
3179 
3180 		err = check_ucmd_data(dev, &params);
3181 		if (err)
3182 			return err;
3183 
3184 		params.ucmd = kzalloc(params.ucmd_size, GFP_KERNEL);
3185 		if (!params.ucmd)
3186 			return -ENOMEM;
3187 
3188 		err = ib_copy_from_udata(params.ucmd, udata, params.inlen);
3189 		if (err)
3190 			goto free_ucmd;
3191 	}
3192 
3193 	mutex_init(&qp->mutex);
3194 	qp->type = type;
3195 	if (udata) {
3196 		err = process_vendor_flags(dev, qp, params.ucmd, attr);
3197 		if (err)
3198 			goto free_ucmd;
3199 
3200 		err = get_qp_uidx(qp, &params);
3201 		if (err)
3202 			goto free_ucmd;
3203 	}
3204 	err = process_create_flags(dev, qp, attr);
3205 	if (err)
3206 		goto free_ucmd;
3207 
3208 	err = check_qp_attr(dev, qp, attr);
3209 	if (err)
3210 		goto free_ucmd;
3211 
3212 	err = create_qp(dev, pd, qp, &params);
3213 	if (err)
3214 		goto free_ucmd;
3215 
3216 	kfree(params.ucmd);
3217 	params.ucmd = NULL;
3218 
3219 	if (udata)
3220 		/*
3221 		 * It is safe to copy response for all user create QP flows,
3222 		 * including MLX5_IB_QPT_DCT, which doesn't need it.
3223 		 * In that case, resp will be filled with zeros.
3224 		 */
3225 		err = ib_copy_to_udata(udata, &params.resp, params.outlen);
3226 	if (err)
3227 		goto destroy_qp;
3228 
3229 	return 0;
3230 
3231 destroy_qp:
3232 	switch (qp->type) {
3233 	case MLX5_IB_QPT_DCT:
3234 		mlx5_ib_destroy_dct(qp);
3235 		break;
3236 	case IB_QPT_GSI:
3237 		mlx5_ib_destroy_gsi(qp);
3238 		break;
3239 	default:
3240 		destroy_qp_common(dev, qp, udata);
3241 	}
3242 
3243 free_ucmd:
3244 	kfree(params.ucmd);
3245 	return err;
3246 }
3247 
3248 int mlx5_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
3249 {
3250 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3251 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3252 
3253 	if (mqp->type == IB_QPT_GSI)
3254 		return mlx5_ib_destroy_gsi(mqp);
3255 
3256 	if (mqp->type == MLX5_IB_QPT_DCT)
3257 		return mlx5_ib_destroy_dct(mqp);
3258 
3259 	destroy_qp_common(dev, mqp, udata);
3260 	return 0;
3261 }
3262 
3263 static int set_qpc_atomic_flags(struct mlx5_ib_qp *qp,
3264 				const struct ib_qp_attr *attr, int attr_mask,
3265 				void *qpc)
3266 {
3267 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
3268 	u8 dest_rd_atomic;
3269 	u32 access_flags;
3270 
3271 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
3272 		dest_rd_atomic = attr->max_dest_rd_atomic;
3273 	else
3274 		dest_rd_atomic = qp->trans_qp.resp_depth;
3275 
3276 	if (attr_mask & IB_QP_ACCESS_FLAGS)
3277 		access_flags = attr->qp_access_flags;
3278 	else
3279 		access_flags = qp->trans_qp.atomic_rd_en;
3280 
3281 	if (!dest_rd_atomic)
3282 		access_flags &= IB_ACCESS_REMOTE_WRITE;
3283 
3284 	MLX5_SET(qpc, qpc, rre, !!(access_flags & IB_ACCESS_REMOTE_READ));
3285 
3286 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC) {
3287 		int atomic_mode;
3288 
3289 		atomic_mode = get_atomic_mode(dev, qp->type);
3290 		if (atomic_mode < 0)
3291 			return -EOPNOTSUPP;
3292 
3293 		MLX5_SET(qpc, qpc, rae, 1);
3294 		MLX5_SET(qpc, qpc, atomic_mode, atomic_mode);
3295 	}
3296 
3297 	MLX5_SET(qpc, qpc, rwe, !!(access_flags & IB_ACCESS_REMOTE_WRITE));
3298 	return 0;
3299 }
3300 
3301 enum {
3302 	MLX5_PATH_FLAG_FL	= 1 << 0,
3303 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
3304 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
3305 };
3306 
3307 static int mlx5_to_ib_rate_map(u8 rate)
3308 {
3309 	static const int rates[] = { IB_RATE_PORT_CURRENT, IB_RATE_56_GBPS,
3310 				     IB_RATE_25_GBPS,	   IB_RATE_100_GBPS,
3311 				     IB_RATE_200_GBPS,	   IB_RATE_50_GBPS,
3312 				     IB_RATE_400_GBPS };
3313 
3314 	if (rate < ARRAY_SIZE(rates))
3315 		return rates[rate];
3316 
3317 	return rate - MLX5_STAT_RATE_OFFSET;
3318 }
3319 
3320 static int ib_to_mlx5_rate_map(u8 rate)
3321 {
3322 	switch (rate) {
3323 	case IB_RATE_PORT_CURRENT:
3324 		return 0;
3325 	case IB_RATE_56_GBPS:
3326 		return 1;
3327 	case IB_RATE_25_GBPS:
3328 		return 2;
3329 	case IB_RATE_100_GBPS:
3330 		return 3;
3331 	case IB_RATE_200_GBPS:
3332 		return 4;
3333 	case IB_RATE_50_GBPS:
3334 		return 5;
3335 	case IB_RATE_400_GBPS:
3336 		return 6;
3337 	default:
3338 		return rate + MLX5_STAT_RATE_OFFSET;
3339 	}
3340 
3341 	return 0;
3342 }
3343 
3344 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
3345 {
3346 	u32 stat_rate_support;
3347 
3348 	if (rate == IB_RATE_PORT_CURRENT)
3349 		return 0;
3350 
3351 	if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_600_GBPS)
3352 		return -EINVAL;
3353 
3354 	stat_rate_support = MLX5_CAP_GEN(dev->mdev, stat_rate_support);
3355 	while (rate != IB_RATE_PORT_CURRENT &&
3356 	       !(1 << ib_to_mlx5_rate_map(rate) & stat_rate_support))
3357 		--rate;
3358 
3359 	return ib_to_mlx5_rate_map(rate);
3360 }
3361 
3362 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
3363 				      struct mlx5_ib_sq *sq, u8 sl,
3364 				      struct ib_pd *pd)
3365 {
3366 	void *in;
3367 	void *tisc;
3368 	int inlen;
3369 	int err;
3370 
3371 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
3372 	in = kvzalloc(inlen, GFP_KERNEL);
3373 	if (!in)
3374 		return -ENOMEM;
3375 
3376 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
3377 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
3378 
3379 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
3380 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
3381 
3382 	err = mlx5_core_modify_tis(dev, sq->tisn, in);
3383 
3384 	kvfree(in);
3385 
3386 	return err;
3387 }
3388 
3389 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
3390 					 struct mlx5_ib_sq *sq, u8 tx_affinity,
3391 					 struct ib_pd *pd)
3392 {
3393 	void *in;
3394 	void *tisc;
3395 	int inlen;
3396 	int err;
3397 
3398 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
3399 	in = kvzalloc(inlen, GFP_KERNEL);
3400 	if (!in)
3401 		return -ENOMEM;
3402 
3403 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
3404 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
3405 
3406 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
3407 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
3408 
3409 	err = mlx5_core_modify_tis(dev, sq->tisn, in);
3410 
3411 	kvfree(in);
3412 
3413 	return err;
3414 }
3415 
3416 static void mlx5_set_path_udp_sport(void *path, const struct rdma_ah_attr *ah,
3417 				    u32 lqpn, u32 rqpn)
3418 
3419 {
3420 	u32 fl = ah->grh.flow_label;
3421 
3422 	if (!fl)
3423 		fl = rdma_calc_flow_label(lqpn, rqpn);
3424 
3425 	MLX5_SET(ads, path, udp_sport, rdma_flow_label_to_udp_sport(fl));
3426 }
3427 
3428 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3429 			 const struct rdma_ah_attr *ah, void *path, u8 port,
3430 			 int attr_mask, u32 path_flags,
3431 			 const struct ib_qp_attr *attr, bool alt)
3432 {
3433 	const struct ib_global_route *grh = rdma_ah_read_grh(ah);
3434 	int err;
3435 	enum ib_gid_type gid_type;
3436 	u8 ah_flags = rdma_ah_get_ah_flags(ah);
3437 	u8 sl = rdma_ah_get_sl(ah);
3438 
3439 	if (attr_mask & IB_QP_PKEY_INDEX)
3440 		MLX5_SET(ads, path, pkey_index,
3441 			 alt ? attr->alt_pkey_index : attr->pkey_index);
3442 
3443 	if (ah_flags & IB_AH_GRH) {
3444 		const struct ib_port_immutable *immutable;
3445 
3446 		immutable = ib_port_immutable_read(&dev->ib_dev, port);
3447 		if (grh->sgid_index >= immutable->gid_tbl_len) {
3448 			pr_err("sgid_index (%u) too large. max is %d\n",
3449 			       grh->sgid_index,
3450 			       immutable->gid_tbl_len);
3451 			return -EINVAL;
3452 		}
3453 	}
3454 
3455 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
3456 		if (!(ah_flags & IB_AH_GRH))
3457 			return -EINVAL;
3458 
3459 		ether_addr_copy(MLX5_ADDR_OF(ads, path, rmac_47_32),
3460 				ah->roce.dmac);
3461 		if ((qp->type == IB_QPT_RC ||
3462 		     qp->type == IB_QPT_UC ||
3463 		     qp->type == IB_QPT_XRC_INI ||
3464 		     qp->type == IB_QPT_XRC_TGT) &&
3465 		    (grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) &&
3466 		    (attr_mask & IB_QP_DEST_QPN))
3467 			mlx5_set_path_udp_sport(path, ah,
3468 						qp->ibqp.qp_num,
3469 						attr->dest_qp_num);
3470 		MLX5_SET(ads, path, eth_prio, sl & 0x7);
3471 		gid_type = ah->grh.sgid_attr->gid_type;
3472 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
3473 			MLX5_SET(ads, path, dscp, grh->traffic_class >> 2);
3474 	} else {
3475 		MLX5_SET(ads, path, fl, !!(path_flags & MLX5_PATH_FLAG_FL));
3476 		MLX5_SET(ads, path, free_ar,
3477 			 !!(path_flags & MLX5_PATH_FLAG_FREE_AR));
3478 		MLX5_SET(ads, path, rlid, rdma_ah_get_dlid(ah));
3479 		MLX5_SET(ads, path, mlid, rdma_ah_get_path_bits(ah));
3480 		MLX5_SET(ads, path, grh, !!(ah_flags & IB_AH_GRH));
3481 		MLX5_SET(ads, path, sl, sl);
3482 	}
3483 
3484 	if (ah_flags & IB_AH_GRH) {
3485 		MLX5_SET(ads, path, src_addr_index, grh->sgid_index);
3486 		MLX5_SET(ads, path, hop_limit, grh->hop_limit);
3487 		MLX5_SET(ads, path, tclass, grh->traffic_class);
3488 		MLX5_SET(ads, path, flow_label, grh->flow_label);
3489 		memcpy(MLX5_ADDR_OF(ads, path, rgid_rip), grh->dgid.raw,
3490 		       sizeof(grh->dgid.raw));
3491 	}
3492 
3493 	err = ib_rate_to_mlx5(dev, rdma_ah_get_static_rate(ah));
3494 	if (err < 0)
3495 		return err;
3496 	MLX5_SET(ads, path, stat_rate, err);
3497 	MLX5_SET(ads, path, vhca_port_num, port);
3498 
3499 	if (attr_mask & IB_QP_TIMEOUT)
3500 		MLX5_SET(ads, path, ack_timeout,
3501 			 alt ? attr->alt_timeout : attr->timeout);
3502 
3503 	if ((qp->type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
3504 		return modify_raw_packet_eth_prio(dev->mdev,
3505 						  &qp->raw_packet_qp.sq,
3506 						  sl & 0xf, qp->ibqp.pd);
3507 
3508 	return 0;
3509 }
3510 
3511 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
3512 	[MLX5_QP_STATE_INIT] = {
3513 		[MLX5_QP_STATE_INIT] = {
3514 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3515 					  MLX5_QP_OPTPAR_RAE		|
3516 					  MLX5_QP_OPTPAR_RWE		|
3517 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3518 					  MLX5_QP_OPTPAR_PRI_PORT	|
3519 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3520 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3521 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3522 					  MLX5_QP_OPTPAR_PRI_PORT	|
3523 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3524 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3525 					  MLX5_QP_OPTPAR_Q_KEY		|
3526 					  MLX5_QP_OPTPAR_PRI_PORT,
3527 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3528 					  MLX5_QP_OPTPAR_RAE		|
3529 					  MLX5_QP_OPTPAR_RWE		|
3530 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3531 					  MLX5_QP_OPTPAR_PRI_PORT	|
3532 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3533 		},
3534 		[MLX5_QP_STATE_RTR] = {
3535 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3536 					  MLX5_QP_OPTPAR_RRE            |
3537 					  MLX5_QP_OPTPAR_RAE            |
3538 					  MLX5_QP_OPTPAR_RWE            |
3539 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3540 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3541 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3542 					  MLX5_QP_OPTPAR_RWE            |
3543 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3544 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3545 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
3546 					  MLX5_QP_OPTPAR_Q_KEY,
3547 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3548 					   MLX5_QP_OPTPAR_Q_KEY,
3549 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
3550 					  MLX5_QP_OPTPAR_RRE            |
3551 					  MLX5_QP_OPTPAR_RAE            |
3552 					  MLX5_QP_OPTPAR_RWE            |
3553 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3554 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3555 		},
3556 	},
3557 	[MLX5_QP_STATE_RTR] = {
3558 		[MLX5_QP_STATE_RTS] = {
3559 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3560 					  MLX5_QP_OPTPAR_RRE		|
3561 					  MLX5_QP_OPTPAR_RAE		|
3562 					  MLX5_QP_OPTPAR_RWE		|
3563 					  MLX5_QP_OPTPAR_PM_STATE	|
3564 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3565 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3566 					  MLX5_QP_OPTPAR_RWE		|
3567 					  MLX5_QP_OPTPAR_PM_STATE,
3568 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
3569 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3570 					  MLX5_QP_OPTPAR_RRE		|
3571 					  MLX5_QP_OPTPAR_RAE		|
3572 					  MLX5_QP_OPTPAR_RWE		|
3573 					  MLX5_QP_OPTPAR_PM_STATE	|
3574 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3575 		},
3576 	},
3577 	[MLX5_QP_STATE_RTS] = {
3578 		[MLX5_QP_STATE_RTS] = {
3579 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3580 					  MLX5_QP_OPTPAR_RAE		|
3581 					  MLX5_QP_OPTPAR_RWE		|
3582 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3583 					  MLX5_QP_OPTPAR_PM_STATE	|
3584 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3585 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3586 					  MLX5_QP_OPTPAR_PM_STATE	|
3587 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3588 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
3589 					  MLX5_QP_OPTPAR_SRQN		|
3590 					  MLX5_QP_OPTPAR_CQN_RCV,
3591 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3592 					  MLX5_QP_OPTPAR_RAE		|
3593 					  MLX5_QP_OPTPAR_RWE		|
3594 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3595 					  MLX5_QP_OPTPAR_PM_STATE	|
3596 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3597 		},
3598 	},
3599 	[MLX5_QP_STATE_SQER] = {
3600 		[MLX5_QP_STATE_RTS] = {
3601 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
3602 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
3603 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
3604 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3605 					   MLX5_QP_OPTPAR_RWE		|
3606 					   MLX5_QP_OPTPAR_RAE		|
3607 					   MLX5_QP_OPTPAR_RRE,
3608 			[MLX5_QP_ST_XRC]  = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3609 					   MLX5_QP_OPTPAR_RWE		|
3610 					   MLX5_QP_OPTPAR_RAE		|
3611 					   MLX5_QP_OPTPAR_RRE,
3612 		},
3613 	},
3614 	[MLX5_QP_STATE_SQD] = {
3615 		[MLX5_QP_STATE_RTS] = {
3616 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
3617 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
3618 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE,
3619 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3620 					  MLX5_QP_OPTPAR_RWE		|
3621 					  MLX5_QP_OPTPAR_RAE		|
3622 					  MLX5_QP_OPTPAR_RRE,
3623 		},
3624 	},
3625 };
3626 
3627 static int ib_nr_to_mlx5_nr(int ib_mask)
3628 {
3629 	switch (ib_mask) {
3630 	case IB_QP_STATE:
3631 		return 0;
3632 	case IB_QP_CUR_STATE:
3633 		return 0;
3634 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
3635 		return 0;
3636 	case IB_QP_ACCESS_FLAGS:
3637 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
3638 			MLX5_QP_OPTPAR_RAE;
3639 	case IB_QP_PKEY_INDEX:
3640 		return MLX5_QP_OPTPAR_PKEY_INDEX;
3641 	case IB_QP_PORT:
3642 		return MLX5_QP_OPTPAR_PRI_PORT;
3643 	case IB_QP_QKEY:
3644 		return MLX5_QP_OPTPAR_Q_KEY;
3645 	case IB_QP_AV:
3646 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
3647 			MLX5_QP_OPTPAR_PRI_PORT;
3648 	case IB_QP_PATH_MTU:
3649 		return 0;
3650 	case IB_QP_TIMEOUT:
3651 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
3652 	case IB_QP_RETRY_CNT:
3653 		return MLX5_QP_OPTPAR_RETRY_COUNT;
3654 	case IB_QP_RNR_RETRY:
3655 		return MLX5_QP_OPTPAR_RNR_RETRY;
3656 	case IB_QP_RQ_PSN:
3657 		return 0;
3658 	case IB_QP_MAX_QP_RD_ATOMIC:
3659 		return MLX5_QP_OPTPAR_SRA_MAX;
3660 	case IB_QP_ALT_PATH:
3661 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
3662 	case IB_QP_MIN_RNR_TIMER:
3663 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
3664 	case IB_QP_SQ_PSN:
3665 		return 0;
3666 	case IB_QP_MAX_DEST_RD_ATOMIC:
3667 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
3668 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
3669 	case IB_QP_PATH_MIG_STATE:
3670 		return MLX5_QP_OPTPAR_PM_STATE;
3671 	case IB_QP_CAP:
3672 		return 0;
3673 	case IB_QP_DEST_QPN:
3674 		return 0;
3675 	}
3676 	return 0;
3677 }
3678 
3679 static int ib_mask_to_mlx5_opt(int ib_mask)
3680 {
3681 	int result = 0;
3682 	int i;
3683 
3684 	for (i = 0; i < 8 * sizeof(int); i++) {
3685 		if ((1 << i) & ib_mask)
3686 			result |= ib_nr_to_mlx5_nr(1 << i);
3687 	}
3688 
3689 	return result;
3690 }
3691 
3692 static int modify_raw_packet_qp_rq(
3693 	struct mlx5_ib_dev *dev, struct mlx5_ib_rq *rq, int new_state,
3694 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3695 {
3696 	void *in;
3697 	void *rqc;
3698 	int inlen;
3699 	int err;
3700 
3701 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
3702 	in = kvzalloc(inlen, GFP_KERNEL);
3703 	if (!in)
3704 		return -ENOMEM;
3705 
3706 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
3707 	MLX5_SET(modify_rq_in, in, uid, to_mpd(pd)->uid);
3708 
3709 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
3710 	MLX5_SET(rqc, rqc, state, new_state);
3711 
3712 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
3713 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
3714 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
3715 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
3716 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
3717 		} else
3718 			dev_info_once(
3719 				&dev->ib_dev.dev,
3720 				"RAW PACKET QP counters are not supported on current FW\n");
3721 	}
3722 
3723 	err = mlx5_core_modify_rq(dev->mdev, rq->base.mqp.qpn, in);
3724 	if (err)
3725 		goto out;
3726 
3727 	rq->state = new_state;
3728 
3729 out:
3730 	kvfree(in);
3731 	return err;
3732 }
3733 
3734 static int modify_raw_packet_qp_sq(
3735 	struct mlx5_core_dev *dev, struct mlx5_ib_sq *sq, int new_state,
3736 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3737 {
3738 	struct mlx5_ib_qp *ibqp = sq->base.container_mibqp;
3739 	struct mlx5_rate_limit old_rl = ibqp->rl;
3740 	struct mlx5_rate_limit new_rl = old_rl;
3741 	bool new_rate_added = false;
3742 	u16 rl_index = 0;
3743 	void *in;
3744 	void *sqc;
3745 	int inlen;
3746 	int err;
3747 
3748 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
3749 	in = kvzalloc(inlen, GFP_KERNEL);
3750 	if (!in)
3751 		return -ENOMEM;
3752 
3753 	MLX5_SET(modify_sq_in, in, uid, to_mpd(pd)->uid);
3754 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
3755 
3756 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
3757 	MLX5_SET(sqc, sqc, state, new_state);
3758 
3759 	if (raw_qp_param->set_mask & MLX5_RAW_QP_RATE_LIMIT) {
3760 		if (new_state != MLX5_SQC_STATE_RDY)
3761 			pr_warn("%s: Rate limit can only be changed when SQ is moving to RDY\n",
3762 				__func__);
3763 		else
3764 			new_rl = raw_qp_param->rl;
3765 	}
3766 
3767 	if (!mlx5_rl_are_equal(&old_rl, &new_rl)) {
3768 		if (new_rl.rate) {
3769 			err = mlx5_rl_add_rate(dev, &rl_index, &new_rl);
3770 			if (err) {
3771 				pr_err("Failed configuring rate limit(err %d): \
3772 				       rate %u, max_burst_sz %u, typical_pkt_sz %u\n",
3773 				       err, new_rl.rate, new_rl.max_burst_sz,
3774 				       new_rl.typical_pkt_sz);
3775 
3776 				goto out;
3777 			}
3778 			new_rate_added = true;
3779 		}
3780 
3781 		MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
3782 		/* index 0 means no limit */
3783 		MLX5_SET(sqc, sqc, packet_pacing_rate_limit_index, rl_index);
3784 	}
3785 
3786 	err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in);
3787 	if (err) {
3788 		/* Remove new rate from table if failed */
3789 		if (new_rate_added)
3790 			mlx5_rl_remove_rate(dev, &new_rl);
3791 		goto out;
3792 	}
3793 
3794 	/* Only remove the old rate after new rate was set */
3795 	if ((old_rl.rate && !mlx5_rl_are_equal(&old_rl, &new_rl)) ||
3796 	    (new_state != MLX5_SQC_STATE_RDY)) {
3797 		mlx5_rl_remove_rate(dev, &old_rl);
3798 		if (new_state != MLX5_SQC_STATE_RDY)
3799 			memset(&new_rl, 0, sizeof(new_rl));
3800 	}
3801 
3802 	ibqp->rl = new_rl;
3803 	sq->state = new_state;
3804 
3805 out:
3806 	kvfree(in);
3807 	return err;
3808 }
3809 
3810 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3811 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
3812 				u8 tx_affinity)
3813 {
3814 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
3815 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
3816 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
3817 	int modify_rq = !!qp->rq.wqe_cnt;
3818 	int modify_sq = !!qp->sq.wqe_cnt;
3819 	int rq_state;
3820 	int sq_state;
3821 	int err;
3822 
3823 	switch (raw_qp_param->operation) {
3824 	case MLX5_CMD_OP_RST2INIT_QP:
3825 		rq_state = MLX5_RQC_STATE_RDY;
3826 		sq_state = MLX5_SQC_STATE_RST;
3827 		break;
3828 	case MLX5_CMD_OP_2ERR_QP:
3829 		rq_state = MLX5_RQC_STATE_ERR;
3830 		sq_state = MLX5_SQC_STATE_ERR;
3831 		break;
3832 	case MLX5_CMD_OP_2RST_QP:
3833 		rq_state = MLX5_RQC_STATE_RST;
3834 		sq_state = MLX5_SQC_STATE_RST;
3835 		break;
3836 	case MLX5_CMD_OP_RTR2RTS_QP:
3837 	case MLX5_CMD_OP_RTS2RTS_QP:
3838 		if (raw_qp_param->set_mask & ~MLX5_RAW_QP_RATE_LIMIT)
3839 			return -EINVAL;
3840 
3841 		modify_rq = 0;
3842 		sq_state = MLX5_SQC_STATE_RDY;
3843 		break;
3844 	case MLX5_CMD_OP_INIT2INIT_QP:
3845 	case MLX5_CMD_OP_INIT2RTR_QP:
3846 		if (raw_qp_param->set_mask)
3847 			return -EINVAL;
3848 		else
3849 			return 0;
3850 	default:
3851 		WARN_ON(1);
3852 		return -EINVAL;
3853 	}
3854 
3855 	if (modify_rq) {
3856 		err =  modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param,
3857 					       qp->ibqp.pd);
3858 		if (err)
3859 			return err;
3860 	}
3861 
3862 	if (modify_sq) {
3863 		struct mlx5_flow_handle *flow_rule;
3864 
3865 		if (tx_affinity) {
3866 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
3867 							    tx_affinity,
3868 							    qp->ibqp.pd);
3869 			if (err)
3870 				return err;
3871 		}
3872 
3873 		flow_rule = create_flow_rule_vport_sq(dev, sq,
3874 						      raw_qp_param->port);
3875 		if (IS_ERR(flow_rule))
3876 			return PTR_ERR(flow_rule);
3877 
3878 		err = modify_raw_packet_qp_sq(dev->mdev, sq, sq_state,
3879 					      raw_qp_param, qp->ibqp.pd);
3880 		if (err) {
3881 			if (flow_rule)
3882 				mlx5_del_flow_rules(flow_rule);
3883 			return err;
3884 		}
3885 
3886 		if (flow_rule) {
3887 			destroy_flow_rule_vport_sq(sq);
3888 			sq->flow_rule = flow_rule;
3889 		}
3890 
3891 		return err;
3892 	}
3893 
3894 	return 0;
3895 }
3896 
3897 static unsigned int get_tx_affinity_rr(struct mlx5_ib_dev *dev,
3898 				       struct ib_udata *udata)
3899 {
3900 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3901 		udata, struct mlx5_ib_ucontext, ibucontext);
3902 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3903 	atomic_t *tx_port_affinity;
3904 
3905 	if (ucontext)
3906 		tx_port_affinity = &ucontext->tx_port_affinity;
3907 	else
3908 		tx_port_affinity = &dev->port[port_num].roce.tx_port_affinity;
3909 
3910 	return (unsigned int)atomic_add_return(1, tx_port_affinity) %
3911 		(dev->lag_active ? dev->lag_ports : MLX5_CAP_GEN(dev->mdev, num_lag_ports)) + 1;
3912 }
3913 
3914 static bool qp_supports_affinity(struct mlx5_ib_qp *qp)
3915 {
3916 	if ((qp->type == IB_QPT_RC) || (qp->type == IB_QPT_UD) ||
3917 	    (qp->type == IB_QPT_UC) || (qp->type == IB_QPT_RAW_PACKET) ||
3918 	    (qp->type == IB_QPT_XRC_INI) || (qp->type == IB_QPT_XRC_TGT) ||
3919 	    (qp->type == MLX5_IB_QPT_DCI))
3920 		return true;
3921 	return false;
3922 }
3923 
3924 static unsigned int get_tx_affinity(struct ib_qp *qp,
3925 				    const struct ib_qp_attr *attr,
3926 				    int attr_mask, u8 init,
3927 				    struct ib_udata *udata)
3928 {
3929 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3930 		udata, struct mlx5_ib_ucontext, ibucontext);
3931 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3932 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3933 	struct mlx5_ib_qp_base *qp_base;
3934 	unsigned int tx_affinity;
3935 
3936 	if (!(mlx5_ib_lag_should_assign_affinity(dev) &&
3937 	      qp_supports_affinity(mqp)))
3938 		return 0;
3939 
3940 	if (mqp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
3941 		tx_affinity = mqp->gsi_lag_port;
3942 	else if (init)
3943 		tx_affinity = get_tx_affinity_rr(dev, udata);
3944 	else if ((attr_mask & IB_QP_AV) && attr->xmit_slave)
3945 		tx_affinity =
3946 			mlx5_lag_get_slave_port(dev->mdev, attr->xmit_slave);
3947 	else
3948 		return 0;
3949 
3950 	qp_base = &mqp->trans_qp.base;
3951 	if (ucontext)
3952 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x ucontext %p\n",
3953 			    tx_affinity, qp_base->mqp.qpn, ucontext);
3954 	else
3955 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x\n",
3956 			    tx_affinity, qp_base->mqp.qpn);
3957 	return tx_affinity;
3958 }
3959 
3960 static int __mlx5_ib_qp_set_counter(struct ib_qp *qp,
3961 				    struct rdma_counter *counter)
3962 {
3963 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3964 	u32 in[MLX5_ST_SZ_DW(rts2rts_qp_in)] = {};
3965 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3966 	struct mlx5_ib_qp_base *base;
3967 	u32 set_id;
3968 	u32 *qpc;
3969 
3970 	if (counter)
3971 		set_id = counter->id;
3972 	else
3973 		set_id = mlx5_ib_get_counters_id(dev, mqp->port - 1);
3974 
3975 	base = &mqp->trans_qp.base;
3976 	MLX5_SET(rts2rts_qp_in, in, opcode, MLX5_CMD_OP_RTS2RTS_QP);
3977 	MLX5_SET(rts2rts_qp_in, in, qpn, base->mqp.qpn);
3978 	MLX5_SET(rts2rts_qp_in, in, uid, base->mqp.uid);
3979 	MLX5_SET(rts2rts_qp_in, in, opt_param_mask,
3980 		 MLX5_QP_OPTPAR_COUNTER_SET_ID);
3981 
3982 	qpc = MLX5_ADDR_OF(rts2rts_qp_in, in, qpc);
3983 	MLX5_SET(qpc, qpc, counter_set_id, set_id);
3984 	return mlx5_cmd_exec_in(dev->mdev, rts2rts_qp, in);
3985 }
3986 
3987 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
3988 			       const struct ib_qp_attr *attr, int attr_mask,
3989 			       enum ib_qp_state cur_state,
3990 			       enum ib_qp_state new_state,
3991 			       const struct mlx5_ib_modify_qp *ucmd,
3992 			       struct mlx5_ib_modify_qp_resp *resp,
3993 			       struct ib_udata *udata)
3994 {
3995 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
3996 		[MLX5_QP_STATE_RST] = {
3997 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3998 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3999 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
4000 		},
4001 		[MLX5_QP_STATE_INIT]  = {
4002 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4003 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4004 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
4005 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
4006 		},
4007 		[MLX5_QP_STATE_RTR]   = {
4008 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4009 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4010 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
4011 		},
4012 		[MLX5_QP_STATE_RTS]   = {
4013 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4014 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4015 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
4016 		},
4017 		[MLX5_QP_STATE_SQD] = {
4018 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4019 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4020 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQD_RTS_QP,
4021 		},
4022 		[MLX5_QP_STATE_SQER] = {
4023 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4024 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4025 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
4026 		},
4027 		[MLX5_QP_STATE_ERR] = {
4028 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4029 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4030 		}
4031 	};
4032 
4033 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4034 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4035 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
4036 	struct mlx5_ib_cq *send_cq, *recv_cq;
4037 	struct mlx5_ib_pd *pd;
4038 	enum mlx5_qp_state mlx5_cur, mlx5_new;
4039 	void *qpc, *pri_path, *alt_path;
4040 	enum mlx5_qp_optpar optpar = 0;
4041 	u32 set_id = 0;
4042 	int mlx5_st;
4043 	int err;
4044 	u16 op;
4045 	u8 tx_affinity = 0;
4046 
4047 	mlx5_st = to_mlx5_st(qp->type);
4048 	if (mlx5_st < 0)
4049 		return -EINVAL;
4050 
4051 	qpc = kzalloc(MLX5_ST_SZ_BYTES(qpc), GFP_KERNEL);
4052 	if (!qpc)
4053 		return -ENOMEM;
4054 
4055 	pd = to_mpd(qp->ibqp.pd);
4056 	MLX5_SET(qpc, qpc, st, mlx5_st);
4057 
4058 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
4059 		MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
4060 	} else {
4061 		switch (attr->path_mig_state) {
4062 		case IB_MIG_MIGRATED:
4063 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
4064 			break;
4065 		case IB_MIG_REARM:
4066 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_REARM);
4067 			break;
4068 		case IB_MIG_ARMED:
4069 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_ARMED);
4070 			break;
4071 		}
4072 	}
4073 
4074 	tx_affinity = get_tx_affinity(ibqp, attr, attr_mask,
4075 				      cur_state == IB_QPS_RESET &&
4076 				      new_state == IB_QPS_INIT, udata);
4077 
4078 	MLX5_SET(qpc, qpc, lag_tx_port_affinity, tx_affinity);
4079 	if (tx_affinity && new_state == IB_QPS_RTR &&
4080 	    MLX5_CAP_GEN(dev->mdev, init2_lag_tx_port_affinity))
4081 		optpar |= MLX5_QP_OPTPAR_LAG_TX_AFF;
4082 
4083 	if (is_sqp(qp->type)) {
4084 		MLX5_SET(qpc, qpc, mtu, IB_MTU_256);
4085 		MLX5_SET(qpc, qpc, log_msg_max, 8);
4086 	} else if ((qp->type == IB_QPT_UD &&
4087 		    !(qp->flags & IB_QP_CREATE_SOURCE_QPN)) ||
4088 		   qp->type == MLX5_IB_QPT_REG_UMR) {
4089 		MLX5_SET(qpc, qpc, mtu, IB_MTU_4096);
4090 		MLX5_SET(qpc, qpc, log_msg_max, 12);
4091 	} else if (attr_mask & IB_QP_PATH_MTU) {
4092 		if (attr->path_mtu < IB_MTU_256 ||
4093 		    attr->path_mtu > IB_MTU_4096) {
4094 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
4095 			err = -EINVAL;
4096 			goto out;
4097 		}
4098 		MLX5_SET(qpc, qpc, mtu, attr->path_mtu);
4099 		MLX5_SET(qpc, qpc, log_msg_max,
4100 			 MLX5_CAP_GEN(dev->mdev, log_max_msg));
4101 	}
4102 
4103 	if (attr_mask & IB_QP_DEST_QPN)
4104 		MLX5_SET(qpc, qpc, remote_qpn, attr->dest_qp_num);
4105 
4106 	pri_path = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
4107 	alt_path = MLX5_ADDR_OF(qpc, qpc, secondary_address_path);
4108 
4109 	if (attr_mask & IB_QP_PKEY_INDEX)
4110 		MLX5_SET(ads, pri_path, pkey_index, attr->pkey_index);
4111 
4112 	/* todo implement counter_index functionality */
4113 
4114 	if (is_sqp(qp->type))
4115 		MLX5_SET(ads, pri_path, vhca_port_num, qp->port);
4116 
4117 	if (attr_mask & IB_QP_PORT)
4118 		MLX5_SET(ads, pri_path, vhca_port_num, attr->port_num);
4119 
4120 	if (attr_mask & IB_QP_AV) {
4121 		err = mlx5_set_path(dev, qp, &attr->ah_attr, pri_path,
4122 				    attr_mask & IB_QP_PORT ? attr->port_num :
4123 							     qp->port,
4124 				    attr_mask, 0, attr, false);
4125 		if (err)
4126 			goto out;
4127 	}
4128 
4129 	if (attr_mask & IB_QP_TIMEOUT)
4130 		MLX5_SET(ads, pri_path, ack_timeout, attr->timeout);
4131 
4132 	if (attr_mask & IB_QP_ALT_PATH) {
4133 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr, alt_path,
4134 				    attr->alt_port_num,
4135 				    attr_mask | IB_QP_PKEY_INDEX |
4136 					    IB_QP_TIMEOUT,
4137 				    0, attr, true);
4138 		if (err)
4139 			goto out;
4140 	}
4141 
4142 	get_cqs(qp->type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
4143 		&send_cq, &recv_cq);
4144 
4145 	MLX5_SET(qpc, qpc, pd, pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
4146 	if (send_cq)
4147 		MLX5_SET(qpc, qpc, cqn_snd, send_cq->mcq.cqn);
4148 	if (recv_cq)
4149 		MLX5_SET(qpc, qpc, cqn_rcv, recv_cq->mcq.cqn);
4150 
4151 	MLX5_SET(qpc, qpc, log_ack_req_freq, MLX5_IB_ACK_REQ_FREQ);
4152 
4153 	if (attr_mask & IB_QP_RNR_RETRY)
4154 		MLX5_SET(qpc, qpc, rnr_retry, attr->rnr_retry);
4155 
4156 	if (attr_mask & IB_QP_RETRY_CNT)
4157 		MLX5_SET(qpc, qpc, retry_count, attr->retry_cnt);
4158 
4159 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC && attr->max_rd_atomic)
4160 		MLX5_SET(qpc, qpc, log_sra_max, ilog2(attr->max_rd_atomic));
4161 
4162 	if (attr_mask & IB_QP_SQ_PSN)
4163 		MLX5_SET(qpc, qpc, next_send_psn, attr->sq_psn);
4164 
4165 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC && attr->max_dest_rd_atomic)
4166 		MLX5_SET(qpc, qpc, log_rra_max,
4167 			 ilog2(attr->max_dest_rd_atomic));
4168 
4169 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
4170 		err = set_qpc_atomic_flags(qp, attr, attr_mask, qpc);
4171 		if (err)
4172 			goto out;
4173 	}
4174 
4175 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
4176 		MLX5_SET(qpc, qpc, min_rnr_nak, attr->min_rnr_timer);
4177 
4178 	if (attr_mask & IB_QP_RQ_PSN)
4179 		MLX5_SET(qpc, qpc, next_rcv_psn, attr->rq_psn);
4180 
4181 	if (attr_mask & IB_QP_QKEY)
4182 		MLX5_SET(qpc, qpc, q_key, attr->qkey);
4183 
4184 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
4185 		MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
4186 
4187 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4188 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
4189 			       qp->port) - 1;
4190 
4191 		/* Underlay port should be used - index 0 function per port */
4192 		if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
4193 			port_num = 0;
4194 
4195 		if (ibqp->counter)
4196 			set_id = ibqp->counter->id;
4197 		else
4198 			set_id = mlx5_ib_get_counters_id(dev, port_num);
4199 		MLX5_SET(qpc, qpc, counter_set_id, set_id);
4200 	}
4201 
4202 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
4203 		MLX5_SET(qpc, qpc, rlky, 1);
4204 
4205 	if (qp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
4206 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
4207 
4208 	mlx5_cur = to_mlx5_state(cur_state);
4209 	mlx5_new = to_mlx5_state(new_state);
4210 
4211 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
4212 	    !optab[mlx5_cur][mlx5_new]) {
4213 		err = -EINVAL;
4214 		goto out;
4215 	}
4216 
4217 	op = optab[mlx5_cur][mlx5_new];
4218 	optpar |= ib_mask_to_mlx5_opt(attr_mask);
4219 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
4220 
4221 	if (qp->type == IB_QPT_RAW_PACKET ||
4222 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4223 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
4224 
4225 		raw_qp_param.operation = op;
4226 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4227 			raw_qp_param.rq_q_ctr_id = set_id;
4228 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
4229 		}
4230 
4231 		if (attr_mask & IB_QP_PORT)
4232 			raw_qp_param.port = attr->port_num;
4233 
4234 		if (attr_mask & IB_QP_RATE_LIMIT) {
4235 			raw_qp_param.rl.rate = attr->rate_limit;
4236 
4237 			if (ucmd->burst_info.max_burst_sz) {
4238 				if (attr->rate_limit &&
4239 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_burst_bound)) {
4240 					raw_qp_param.rl.max_burst_sz =
4241 						ucmd->burst_info.max_burst_sz;
4242 				} else {
4243 					err = -EINVAL;
4244 					goto out;
4245 				}
4246 			}
4247 
4248 			if (ucmd->burst_info.typical_pkt_sz) {
4249 				if (attr->rate_limit &&
4250 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_typical_size)) {
4251 					raw_qp_param.rl.typical_pkt_sz =
4252 						ucmd->burst_info.typical_pkt_sz;
4253 				} else {
4254 					err = -EINVAL;
4255 					goto out;
4256 				}
4257 			}
4258 
4259 			raw_qp_param.set_mask |= MLX5_RAW_QP_RATE_LIMIT;
4260 		}
4261 
4262 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, tx_affinity);
4263 	} else {
4264 		if (udata) {
4265 			/* For the kernel flows, the resp will stay zero */
4266 			resp->ece_options =
4267 				MLX5_CAP_GEN(dev->mdev, ece_support) ?
4268 					ucmd->ece_options : 0;
4269 			resp->response_length = sizeof(*resp);
4270 		}
4271 		err = mlx5_core_qp_modify(dev, op, optpar, qpc, &base->mqp,
4272 					  &resp->ece_options);
4273 	}
4274 
4275 	if (err)
4276 		goto out;
4277 
4278 	qp->state = new_state;
4279 
4280 	if (attr_mask & IB_QP_ACCESS_FLAGS)
4281 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
4282 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
4283 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
4284 	if (attr_mask & IB_QP_PORT)
4285 		qp->port = attr->port_num;
4286 	if (attr_mask & IB_QP_ALT_PATH)
4287 		qp->trans_qp.alt_port = attr->alt_port_num;
4288 
4289 	/*
4290 	 * If we moved a kernel QP to RESET, clean up all old CQ
4291 	 * entries and reinitialize the QP.
4292 	 */
4293 	if (new_state == IB_QPS_RESET &&
4294 	    !ibqp->uobject && qp->type != IB_QPT_XRC_TGT) {
4295 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
4296 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
4297 		if (send_cq != recv_cq)
4298 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
4299 
4300 		qp->rq.head = 0;
4301 		qp->rq.tail = 0;
4302 		qp->sq.head = 0;
4303 		qp->sq.tail = 0;
4304 		qp->sq.cur_post = 0;
4305 		if (qp->sq.wqe_cnt)
4306 			qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
4307 		qp->sq.last_poll = 0;
4308 		qp->db.db[MLX5_RCV_DBR] = 0;
4309 		qp->db.db[MLX5_SND_DBR] = 0;
4310 	}
4311 
4312 	if ((new_state == IB_QPS_RTS) && qp->counter_pending) {
4313 		err = __mlx5_ib_qp_set_counter(ibqp, ibqp->counter);
4314 		if (!err)
4315 			qp->counter_pending = 0;
4316 	}
4317 
4318 out:
4319 	kfree(qpc);
4320 	return err;
4321 }
4322 
4323 static inline bool is_valid_mask(int mask, int req, int opt)
4324 {
4325 	if ((mask & req) != req)
4326 		return false;
4327 
4328 	if (mask & ~(req | opt))
4329 		return false;
4330 
4331 	return true;
4332 }
4333 
4334 /* check valid transition for driver QP types
4335  * for now the only QP type that this function supports is DCI
4336  */
4337 static bool modify_dci_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state new_state,
4338 				enum ib_qp_attr_mask attr_mask)
4339 {
4340 	int req = IB_QP_STATE;
4341 	int opt = 0;
4342 
4343 	if (new_state == IB_QPS_RESET) {
4344 		return is_valid_mask(attr_mask, req, opt);
4345 	} else if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4346 		req |= IB_QP_PKEY_INDEX | IB_QP_PORT;
4347 		return is_valid_mask(attr_mask, req, opt);
4348 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
4349 		opt = IB_QP_PKEY_INDEX | IB_QP_PORT;
4350 		return is_valid_mask(attr_mask, req, opt);
4351 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
4352 		req |= IB_QP_PATH_MTU;
4353 		opt = IB_QP_PKEY_INDEX | IB_QP_AV;
4354 		return is_valid_mask(attr_mask, req, opt);
4355 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
4356 		req |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4357 		       IB_QP_MAX_QP_RD_ATOMIC | IB_QP_SQ_PSN;
4358 		opt = IB_QP_MIN_RNR_TIMER;
4359 		return is_valid_mask(attr_mask, req, opt);
4360 	} else if (cur_state == IB_QPS_RTS && new_state == IB_QPS_RTS) {
4361 		opt = IB_QP_MIN_RNR_TIMER;
4362 		return is_valid_mask(attr_mask, req, opt);
4363 	} else if (cur_state != IB_QPS_RESET && new_state == IB_QPS_ERR) {
4364 		return is_valid_mask(attr_mask, req, opt);
4365 	}
4366 	return false;
4367 }
4368 
4369 /* mlx5_ib_modify_dct: modify a DCT QP
4370  * valid transitions are:
4371  * RESET to INIT: must set access_flags, pkey_index and port
4372  * INIT  to RTR : must set min_rnr_timer, tclass, flow_label,
4373  *			   mtu, gid_index and hop_limit
4374  * Other transitions and attributes are illegal
4375  */
4376 static int mlx5_ib_modify_dct(struct ib_qp *ibqp, struct ib_qp_attr *attr,
4377 			      int attr_mask, struct mlx5_ib_modify_qp *ucmd,
4378 			      struct ib_udata *udata)
4379 {
4380 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4381 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4382 	enum ib_qp_state cur_state, new_state;
4383 	int required = IB_QP_STATE;
4384 	void *dctc;
4385 	int err;
4386 
4387 	if (!(attr_mask & IB_QP_STATE))
4388 		return -EINVAL;
4389 
4390 	cur_state = qp->state;
4391 	new_state = attr->qp_state;
4392 
4393 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
4394 	if (MLX5_CAP_GEN(dev->mdev, ece_support) && ucmd->ece_options)
4395 		/*
4396 		 * DCT doesn't initialize QP till modify command is executed,
4397 		 * so we need to overwrite previously set ECE field if user
4398 		 * provided any value except zero, which means not set/not
4399 		 * valid.
4400 		 */
4401 		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
4402 
4403 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4404 		u16 set_id;
4405 
4406 		required |= IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
4407 		if (!is_valid_mask(attr_mask, required, 0))
4408 			return -EINVAL;
4409 
4410 		if (attr->port_num == 0 ||
4411 		    attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)) {
4412 			mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4413 				    attr->port_num, dev->num_ports);
4414 			return -EINVAL;
4415 		}
4416 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
4417 			MLX5_SET(dctc, dctc, rre, 1);
4418 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
4419 			MLX5_SET(dctc, dctc, rwe, 1);
4420 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC) {
4421 			int atomic_mode;
4422 
4423 			atomic_mode = get_atomic_mode(dev, MLX5_IB_QPT_DCT);
4424 			if (atomic_mode < 0)
4425 				return -EOPNOTSUPP;
4426 
4427 			MLX5_SET(dctc, dctc, atomic_mode, atomic_mode);
4428 			MLX5_SET(dctc, dctc, rae, 1);
4429 		}
4430 		MLX5_SET(dctc, dctc, pkey_index, attr->pkey_index);
4431 		if (mlx5_lag_is_active(dev->mdev))
4432 			MLX5_SET(dctc, dctc, port,
4433 				 get_tx_affinity_rr(dev, udata));
4434 		else
4435 			MLX5_SET(dctc, dctc, port, attr->port_num);
4436 
4437 		set_id = mlx5_ib_get_counters_id(dev, attr->port_num - 1);
4438 		MLX5_SET(dctc, dctc, counter_set_id, set_id);
4439 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
4440 		struct mlx5_ib_modify_qp_resp resp = {};
4441 		u32 out[MLX5_ST_SZ_DW(create_dct_out)] = {};
4442 		u32 min_resp_len = offsetofend(typeof(resp), dctn);
4443 
4444 		if (udata->outlen < min_resp_len)
4445 			return -EINVAL;
4446 		/*
4447 		 * If we don't have enough space for the ECE options,
4448 		 * simply indicate it with resp.response_length.
4449 		 */
4450 		resp.response_length = (udata->outlen < sizeof(resp)) ?
4451 					       min_resp_len :
4452 					       sizeof(resp);
4453 
4454 		required |= IB_QP_MIN_RNR_TIMER | IB_QP_AV | IB_QP_PATH_MTU;
4455 		if (!is_valid_mask(attr_mask, required, 0))
4456 			return -EINVAL;
4457 		MLX5_SET(dctc, dctc, min_rnr_nak, attr->min_rnr_timer);
4458 		MLX5_SET(dctc, dctc, tclass, attr->ah_attr.grh.traffic_class);
4459 		MLX5_SET(dctc, dctc, flow_label, attr->ah_attr.grh.flow_label);
4460 		MLX5_SET(dctc, dctc, mtu, attr->path_mtu);
4461 		MLX5_SET(dctc, dctc, my_addr_index, attr->ah_attr.grh.sgid_index);
4462 		MLX5_SET(dctc, dctc, hop_limit, attr->ah_attr.grh.hop_limit);
4463 		if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
4464 			MLX5_SET(dctc, dctc, eth_prio, attr->ah_attr.sl & 0x7);
4465 
4466 		err = mlx5_core_create_dct(dev, &qp->dct.mdct, qp->dct.in,
4467 					   MLX5_ST_SZ_BYTES(create_dct_in), out,
4468 					   sizeof(out));
4469 		err = mlx5_cmd_check(dev->mdev, err, qp->dct.in, out);
4470 		if (err)
4471 			return err;
4472 		resp.dctn = qp->dct.mdct.mqp.qpn;
4473 		if (MLX5_CAP_GEN(dev->mdev, ece_support))
4474 			resp.ece_options = MLX5_GET(create_dct_out, out, ece);
4475 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4476 		if (err) {
4477 			mlx5_core_destroy_dct(dev, &qp->dct.mdct);
4478 			return err;
4479 		}
4480 	} else {
4481 		mlx5_ib_warn(dev, "Modify DCT: Invalid transition from %d to %d\n", cur_state, new_state);
4482 		return -EINVAL;
4483 	}
4484 
4485 	qp->state = new_state;
4486 	return 0;
4487 }
4488 
4489 static bool mlx5_ib_modify_qp_allowed(struct mlx5_ib_dev *dev,
4490 				      struct mlx5_ib_qp *qp)
4491 {
4492 	if (dev->profile != &raw_eth_profile)
4493 		return true;
4494 
4495 	if (qp->type == IB_QPT_RAW_PACKET || qp->type == MLX5_IB_QPT_REG_UMR)
4496 		return true;
4497 
4498 	/* Internal QP used for wc testing, with NOPs in wq */
4499 	if (qp->flags & MLX5_IB_QP_CREATE_WC_TEST)
4500 		return true;
4501 
4502 	return false;
4503 }
4504 
4505 static int validate_rd_atomic(struct mlx5_ib_dev *dev, struct ib_qp_attr *attr,
4506 			      int attr_mask, enum ib_qp_type qp_type)
4507 {
4508 	int log_max_ra_res;
4509 	int log_max_ra_req;
4510 
4511 	if (qp_type == MLX5_IB_QPT_DCI) {
4512 		log_max_ra_res = 1 << MLX5_CAP_GEN(dev->mdev,
4513 						   log_max_ra_res_dc);
4514 		log_max_ra_req = 1 << MLX5_CAP_GEN(dev->mdev,
4515 						   log_max_ra_req_dc);
4516 	} else {
4517 		log_max_ra_res = 1 << MLX5_CAP_GEN(dev->mdev,
4518 						   log_max_ra_res_qp);
4519 		log_max_ra_req = 1 << MLX5_CAP_GEN(dev->mdev,
4520 						   log_max_ra_req_qp);
4521 	}
4522 
4523 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
4524 	    attr->max_rd_atomic > log_max_ra_res) {
4525 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
4526 			    attr->max_rd_atomic);
4527 		return false;
4528 	}
4529 
4530 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
4531 	    attr->max_dest_rd_atomic > log_max_ra_req) {
4532 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
4533 			    attr->max_dest_rd_atomic);
4534 		return false;
4535 	}
4536 	return true;
4537 }
4538 
4539 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
4540 		      int attr_mask, struct ib_udata *udata)
4541 {
4542 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4543 	struct mlx5_ib_modify_qp_resp resp = {};
4544 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4545 	struct mlx5_ib_modify_qp ucmd = {};
4546 	enum ib_qp_type qp_type;
4547 	enum ib_qp_state cur_state, new_state;
4548 	int err = -EINVAL;
4549 
4550 	if (!mlx5_ib_modify_qp_allowed(dev, qp))
4551 		return -EOPNOTSUPP;
4552 
4553 	if (attr_mask & ~(IB_QP_ATTR_STANDARD_BITS | IB_QP_RATE_LIMIT))
4554 		return -EOPNOTSUPP;
4555 
4556 	if (ibqp->rwq_ind_tbl)
4557 		return -ENOSYS;
4558 
4559 	if (udata && udata->inlen) {
4560 		if (udata->inlen < offsetofend(typeof(ucmd), ece_options))
4561 			return -EINVAL;
4562 
4563 		if (udata->inlen > sizeof(ucmd) &&
4564 		    !ib_is_udata_cleared(udata, sizeof(ucmd),
4565 					 udata->inlen - sizeof(ucmd)))
4566 			return -EOPNOTSUPP;
4567 
4568 		if (ib_copy_from_udata(&ucmd, udata,
4569 				       min(udata->inlen, sizeof(ucmd))))
4570 			return -EFAULT;
4571 
4572 		if (ucmd.comp_mask ||
4573 		    memchr_inv(&ucmd.burst_info.reserved, 0,
4574 			       sizeof(ucmd.burst_info.reserved)))
4575 			return -EOPNOTSUPP;
4576 
4577 	}
4578 
4579 	if (qp->type == IB_QPT_GSI)
4580 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
4581 
4582 	qp_type = (qp->type == MLX5_IB_QPT_HW_GSI) ? IB_QPT_GSI : qp->type;
4583 
4584 	if (qp_type == MLX5_IB_QPT_DCT)
4585 		return mlx5_ib_modify_dct(ibqp, attr, attr_mask, &ucmd, udata);
4586 
4587 	mutex_lock(&qp->mutex);
4588 
4589 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
4590 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
4591 
4592 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4593 		if (attr_mask & ~(IB_QP_STATE | IB_QP_CUR_STATE)) {
4594 			mlx5_ib_dbg(dev, "invalid attr_mask 0x%x when underlay QP is used\n",
4595 				    attr_mask);
4596 			goto out;
4597 		}
4598 	} else if (qp_type != MLX5_IB_QPT_REG_UMR &&
4599 		   qp_type != MLX5_IB_QPT_DCI &&
4600 		   !ib_modify_qp_is_ok(cur_state, new_state, qp_type,
4601 				       attr_mask)) {
4602 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4603 			    cur_state, new_state, qp->type, attr_mask);
4604 		goto out;
4605 	} else if (qp_type == MLX5_IB_QPT_DCI &&
4606 		   !modify_dci_qp_is_ok(cur_state, new_state, attr_mask)) {
4607 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4608 			    cur_state, new_state, qp_type, attr_mask);
4609 		goto out;
4610 	}
4611 
4612 	if ((attr_mask & IB_QP_PORT) &&
4613 	    (attr->port_num == 0 ||
4614 	     attr->port_num > dev->num_ports)) {
4615 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4616 			    attr->port_num, dev->num_ports);
4617 		goto out;
4618 	}
4619 
4620 	if ((attr_mask & IB_QP_PKEY_INDEX) &&
4621 	    attr->pkey_index >= dev->pkey_table_len) {
4622 		mlx5_ib_dbg(dev, "invalid pkey index %d\n", attr->pkey_index);
4623 		goto out;
4624 	}
4625 
4626 	if (!validate_rd_atomic(dev, attr, attr_mask, qp_type))
4627 		goto out;
4628 
4629 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
4630 		err = 0;
4631 		goto out;
4632 	}
4633 
4634 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state,
4635 				  new_state, &ucmd, &resp, udata);
4636 
4637 	/* resp.response_length is set in ECE supported flows only */
4638 	if (!err && resp.response_length &&
4639 	    udata->outlen >= resp.response_length)
4640 		/* Return -EFAULT to the user and expect him to destroy QP. */
4641 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4642 
4643 out:
4644 	mutex_unlock(&qp->mutex);
4645 	return err;
4646 }
4647 
4648 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4649 {
4650 	switch (mlx5_state) {
4651 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
4652 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
4653 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
4654 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
4655 	case MLX5_QP_STATE_SQ_DRAINING:
4656 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
4657 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
4658 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
4659 	default:		     return -1;
4660 	}
4661 }
4662 
4663 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4664 {
4665 	switch (mlx5_mig_state) {
4666 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
4667 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
4668 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4669 	default: return -1;
4670 	}
4671 }
4672 
4673 static void to_rdma_ah_attr(struct mlx5_ib_dev *ibdev,
4674 			    struct rdma_ah_attr *ah_attr, void *path)
4675 {
4676 	int port = MLX5_GET(ads, path, vhca_port_num);
4677 	int static_rate;
4678 
4679 	memset(ah_attr, 0, sizeof(*ah_attr));
4680 
4681 	if (!port || port > ibdev->num_ports)
4682 		return;
4683 
4684 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port);
4685 
4686 	rdma_ah_set_port_num(ah_attr, port);
4687 	rdma_ah_set_sl(ah_attr, MLX5_GET(ads, path, sl));
4688 
4689 	rdma_ah_set_dlid(ah_attr, MLX5_GET(ads, path, rlid));
4690 	rdma_ah_set_path_bits(ah_attr, MLX5_GET(ads, path, mlid));
4691 
4692 	static_rate = MLX5_GET(ads, path, stat_rate);
4693 	rdma_ah_set_static_rate(ah_attr, mlx5_to_ib_rate_map(static_rate));
4694 	if (MLX5_GET(ads, path, grh) ||
4695 	    ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
4696 		rdma_ah_set_grh(ah_attr, NULL, MLX5_GET(ads, path, flow_label),
4697 				MLX5_GET(ads, path, src_addr_index),
4698 				MLX5_GET(ads, path, hop_limit),
4699 				MLX5_GET(ads, path, tclass));
4700 		rdma_ah_set_dgid_raw(ah_attr, MLX5_ADDR_OF(ads, path, rgid_rip));
4701 	}
4702 }
4703 
4704 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4705 					struct mlx5_ib_sq *sq,
4706 					u8 *sq_state)
4707 {
4708 	int err;
4709 
4710 	err = mlx5_core_query_sq_state(dev->mdev, sq->base.mqp.qpn, sq_state);
4711 	if (err)
4712 		goto out;
4713 	sq->state = *sq_state;
4714 
4715 out:
4716 	return err;
4717 }
4718 
4719 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4720 					struct mlx5_ib_rq *rq,
4721 					u8 *rq_state)
4722 {
4723 	void *out;
4724 	void *rqc;
4725 	int inlen;
4726 	int err;
4727 
4728 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4729 	out = kvzalloc(inlen, GFP_KERNEL);
4730 	if (!out)
4731 		return -ENOMEM;
4732 
4733 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4734 	if (err)
4735 		goto out;
4736 
4737 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4738 	*rq_state = MLX5_GET(rqc, rqc, state);
4739 	rq->state = *rq_state;
4740 
4741 out:
4742 	kvfree(out);
4743 	return err;
4744 }
4745 
4746 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4747 				  struct mlx5_ib_qp *qp, u8 *qp_state)
4748 {
4749 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4750 		[MLX5_RQC_STATE_RST] = {
4751 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
4752 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4753 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
4754 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
4755 		},
4756 		[MLX5_RQC_STATE_RDY] = {
4757 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE,
4758 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4759 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
4760 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
4761 		},
4762 		[MLX5_RQC_STATE_ERR] = {
4763 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
4764 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4765 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
4766 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
4767 		},
4768 		[MLX5_RQ_STATE_NA] = {
4769 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE,
4770 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4771 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
4772 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
4773 		},
4774 	};
4775 
4776 	*qp_state = sqrq_trans[rq_state][sq_state];
4777 
4778 	if (*qp_state == MLX5_QP_STATE_BAD) {
4779 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4780 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4781 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4782 		return -EINVAL;
4783 	}
4784 
4785 	if (*qp_state == MLX5_QP_STATE)
4786 		*qp_state = qp->state;
4787 
4788 	return 0;
4789 }
4790 
4791 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4792 				     struct mlx5_ib_qp *qp,
4793 				     u8 *raw_packet_qp_state)
4794 {
4795 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4796 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4797 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4798 	int err;
4799 	u8 sq_state = MLX5_SQ_STATE_NA;
4800 	u8 rq_state = MLX5_RQ_STATE_NA;
4801 
4802 	if (qp->sq.wqe_cnt) {
4803 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4804 		if (err)
4805 			return err;
4806 	}
4807 
4808 	if (qp->rq.wqe_cnt) {
4809 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4810 		if (err)
4811 			return err;
4812 	}
4813 
4814 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4815 				      raw_packet_qp_state);
4816 }
4817 
4818 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4819 			 struct ib_qp_attr *qp_attr)
4820 {
4821 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
4822 	void *qpc, *pri_path, *alt_path;
4823 	u32 *outb;
4824 	int err;
4825 
4826 	outb = kzalloc(outlen, GFP_KERNEL);
4827 	if (!outb)
4828 		return -ENOMEM;
4829 
4830 	err = mlx5_core_qp_query(dev, &qp->trans_qp.base.mqp, outb, outlen);
4831 	if (err)
4832 		goto out;
4833 
4834 	qpc = MLX5_ADDR_OF(query_qp_out, outb, qpc);
4835 
4836 	qp->state = to_ib_qp_state(MLX5_GET(qpc, qpc, state));
4837 	if (MLX5_GET(qpc, qpc, state) == MLX5_QP_STATE_SQ_DRAINING)
4838 		qp_attr->sq_draining = 1;
4839 
4840 	qp_attr->path_mtu = MLX5_GET(qpc, qpc, mtu);
4841 	qp_attr->path_mig_state = to_ib_mig_state(MLX5_GET(qpc, qpc, pm_state));
4842 	qp_attr->qkey = MLX5_GET(qpc, qpc, q_key);
4843 	qp_attr->rq_psn = MLX5_GET(qpc, qpc, next_rcv_psn);
4844 	qp_attr->sq_psn = MLX5_GET(qpc, qpc, next_send_psn);
4845 	qp_attr->dest_qp_num = MLX5_GET(qpc, qpc, remote_qpn);
4846 
4847 	if (MLX5_GET(qpc, qpc, rre))
4848 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ;
4849 	if (MLX5_GET(qpc, qpc, rwe))
4850 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_WRITE;
4851 	if (MLX5_GET(qpc, qpc, rae))
4852 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_ATOMIC;
4853 
4854 	qp_attr->max_rd_atomic = 1 << MLX5_GET(qpc, qpc, log_sra_max);
4855 	qp_attr->max_dest_rd_atomic = 1 << MLX5_GET(qpc, qpc, log_rra_max);
4856 	qp_attr->min_rnr_timer = MLX5_GET(qpc, qpc, min_rnr_nak);
4857 	qp_attr->retry_cnt = MLX5_GET(qpc, qpc, retry_count);
4858 	qp_attr->rnr_retry = MLX5_GET(qpc, qpc, rnr_retry);
4859 
4860 	pri_path = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
4861 	alt_path = MLX5_ADDR_OF(qpc, qpc, secondary_address_path);
4862 
4863 	if (qp->type == IB_QPT_RC || qp->type == IB_QPT_UC ||
4864 	    qp->type == IB_QPT_XRC_INI || qp->type == IB_QPT_XRC_TGT) {
4865 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, pri_path);
4866 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, alt_path);
4867 		qp_attr->alt_pkey_index = MLX5_GET(ads, alt_path, pkey_index);
4868 		qp_attr->alt_port_num = MLX5_GET(ads, alt_path, vhca_port_num);
4869 	}
4870 
4871 	qp_attr->pkey_index = MLX5_GET(ads, pri_path, pkey_index);
4872 	qp_attr->port_num = MLX5_GET(ads, pri_path, vhca_port_num);
4873 	qp_attr->timeout = MLX5_GET(ads, pri_path, ack_timeout);
4874 	qp_attr->alt_timeout = MLX5_GET(ads, alt_path, ack_timeout);
4875 
4876 out:
4877 	kfree(outb);
4878 	return err;
4879 }
4880 
4881 static int mlx5_ib_dct_query_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *mqp,
4882 				struct ib_qp_attr *qp_attr, int qp_attr_mask,
4883 				struct ib_qp_init_attr *qp_init_attr)
4884 {
4885 	struct mlx5_core_dct	*dct = &mqp->dct.mdct;
4886 	u32 *out;
4887 	u32 access_flags = 0;
4888 	int outlen = MLX5_ST_SZ_BYTES(query_dct_out);
4889 	void *dctc;
4890 	int err;
4891 	int supported_mask = IB_QP_STATE |
4892 			     IB_QP_ACCESS_FLAGS |
4893 			     IB_QP_PORT |
4894 			     IB_QP_MIN_RNR_TIMER |
4895 			     IB_QP_AV |
4896 			     IB_QP_PATH_MTU |
4897 			     IB_QP_PKEY_INDEX;
4898 
4899 	if (qp_attr_mask & ~supported_mask)
4900 		return -EINVAL;
4901 	if (mqp->state != IB_QPS_RTR)
4902 		return -EINVAL;
4903 
4904 	out = kzalloc(outlen, GFP_KERNEL);
4905 	if (!out)
4906 		return -ENOMEM;
4907 
4908 	err = mlx5_core_dct_query(dev, dct, out, outlen);
4909 	if (err)
4910 		goto out;
4911 
4912 	dctc = MLX5_ADDR_OF(query_dct_out, out, dct_context_entry);
4913 
4914 	if (qp_attr_mask & IB_QP_STATE)
4915 		qp_attr->qp_state = IB_QPS_RTR;
4916 
4917 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
4918 		if (MLX5_GET(dctc, dctc, rre))
4919 			access_flags |= IB_ACCESS_REMOTE_READ;
4920 		if (MLX5_GET(dctc, dctc, rwe))
4921 			access_flags |= IB_ACCESS_REMOTE_WRITE;
4922 		if (MLX5_GET(dctc, dctc, rae))
4923 			access_flags |= IB_ACCESS_REMOTE_ATOMIC;
4924 		qp_attr->qp_access_flags = access_flags;
4925 	}
4926 
4927 	if (qp_attr_mask & IB_QP_PORT)
4928 		qp_attr->port_num = MLX5_GET(dctc, dctc, port);
4929 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER)
4930 		qp_attr->min_rnr_timer = MLX5_GET(dctc, dctc, min_rnr_nak);
4931 	if (qp_attr_mask & IB_QP_AV) {
4932 		qp_attr->ah_attr.grh.traffic_class = MLX5_GET(dctc, dctc, tclass);
4933 		qp_attr->ah_attr.grh.flow_label = MLX5_GET(dctc, dctc, flow_label);
4934 		qp_attr->ah_attr.grh.sgid_index = MLX5_GET(dctc, dctc, my_addr_index);
4935 		qp_attr->ah_attr.grh.hop_limit = MLX5_GET(dctc, dctc, hop_limit);
4936 	}
4937 	if (qp_attr_mask & IB_QP_PATH_MTU)
4938 		qp_attr->path_mtu = MLX5_GET(dctc, dctc, mtu);
4939 	if (qp_attr_mask & IB_QP_PKEY_INDEX)
4940 		qp_attr->pkey_index = MLX5_GET(dctc, dctc, pkey_index);
4941 out:
4942 	kfree(out);
4943 	return err;
4944 }
4945 
4946 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4947 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4948 {
4949 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4950 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4951 	int err = 0;
4952 	u8 raw_packet_qp_state;
4953 
4954 	if (ibqp->rwq_ind_tbl)
4955 		return -ENOSYS;
4956 
4957 	if (qp->type == IB_QPT_GSI)
4958 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4959 					    qp_init_attr);
4960 
4961 	/* Not all of output fields are applicable, make sure to zero them */
4962 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
4963 	memset(qp_attr, 0, sizeof(*qp_attr));
4964 
4965 	if (unlikely(qp->type == MLX5_IB_QPT_DCT))
4966 		return mlx5_ib_dct_query_qp(dev, qp, qp_attr,
4967 					    qp_attr_mask, qp_init_attr);
4968 
4969 	mutex_lock(&qp->mutex);
4970 
4971 	if (qp->type == IB_QPT_RAW_PACKET ||
4972 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4973 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4974 		if (err)
4975 			goto out;
4976 		qp->state = raw_packet_qp_state;
4977 		qp_attr->port_num = 1;
4978 	} else {
4979 		err = query_qp_attr(dev, qp, qp_attr);
4980 		if (err)
4981 			goto out;
4982 	}
4983 
4984 	qp_attr->qp_state	     = qp->state;
4985 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
4986 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
4987 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
4988 
4989 	if (!ibqp->uobject) {
4990 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
4991 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
4992 		qp_init_attr->qp_context = ibqp->qp_context;
4993 	} else {
4994 		qp_attr->cap.max_send_wr  = 0;
4995 		qp_attr->cap.max_send_sge = 0;
4996 	}
4997 
4998 	qp_init_attr->qp_type = qp->type;
4999 	qp_init_attr->recv_cq = ibqp->recv_cq;
5000 	qp_init_attr->send_cq = ibqp->send_cq;
5001 	qp_init_attr->srq = ibqp->srq;
5002 	qp_attr->cap.max_inline_data = qp->max_inline_data;
5003 
5004 	qp_init_attr->cap	     = qp_attr->cap;
5005 
5006 	qp_init_attr->create_flags = qp->flags;
5007 
5008 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
5009 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
5010 
5011 out:
5012 	mutex_unlock(&qp->mutex);
5013 	return err;
5014 }
5015 
5016 int mlx5_ib_alloc_xrcd(struct ib_xrcd *ibxrcd, struct ib_udata *udata)
5017 {
5018 	struct mlx5_ib_dev *dev = to_mdev(ibxrcd->device);
5019 	struct mlx5_ib_xrcd *xrcd = to_mxrcd(ibxrcd);
5020 
5021 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
5022 		return -EOPNOTSUPP;
5023 
5024 	return mlx5_cmd_xrcd_alloc(dev->mdev, &xrcd->xrcdn, 0);
5025 }
5026 
5027 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
5028 {
5029 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
5030 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
5031 
5032 	return mlx5_cmd_xrcd_dealloc(dev->mdev, xrcdn, 0);
5033 }
5034 
5035 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
5036 {
5037 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
5038 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
5039 	struct ib_event event;
5040 
5041 	if (rwq->ibwq.event_handler) {
5042 		event.device     = rwq->ibwq.device;
5043 		event.element.wq = &rwq->ibwq;
5044 		switch (type) {
5045 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
5046 			event.event = IB_EVENT_WQ_FATAL;
5047 			break;
5048 		default:
5049 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
5050 			return;
5051 		}
5052 
5053 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
5054 	}
5055 }
5056 
5057 static int set_delay_drop(struct mlx5_ib_dev *dev)
5058 {
5059 	int err = 0;
5060 
5061 	mutex_lock(&dev->delay_drop.lock);
5062 	if (dev->delay_drop.activate)
5063 		goto out;
5064 
5065 	err = mlx5_core_set_delay_drop(dev, dev->delay_drop.timeout);
5066 	if (err)
5067 		goto out;
5068 
5069 	dev->delay_drop.activate = true;
5070 out:
5071 	mutex_unlock(&dev->delay_drop.lock);
5072 
5073 	if (!err)
5074 		atomic_inc(&dev->delay_drop.rqs_cnt);
5075 	return err;
5076 }
5077 
5078 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
5079 		      struct ib_wq_init_attr *init_attr)
5080 {
5081 	struct mlx5_ib_dev *dev;
5082 	int has_net_offloads;
5083 	__be64 *rq_pas0;
5084 	int ts_format;
5085 	void *in;
5086 	void *rqc;
5087 	void *wq;
5088 	int inlen;
5089 	int err;
5090 
5091 	dev = to_mdev(pd->device);
5092 
5093 	ts_format = get_rq_ts_format(dev, to_mcq(init_attr->cq));
5094 	if (ts_format < 0)
5095 		return ts_format;
5096 
5097 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
5098 	in = kvzalloc(inlen, GFP_KERNEL);
5099 	if (!in)
5100 		return -ENOMEM;
5101 
5102 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
5103 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
5104 	MLX5_SET(rqc,  rqc, mem_rq_type,
5105 		 MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
5106 	MLX5_SET(rqc, rqc, ts_format, ts_format);
5107 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
5108 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
5109 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
5110 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
5111 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
5112 	MLX5_SET(wq, wq, wq_type,
5113 		 rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ ?
5114 		 MLX5_WQ_TYPE_CYCLIC_STRIDING_RQ : MLX5_WQ_TYPE_CYCLIC);
5115 	if (init_attr->create_flags & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
5116 		if (!MLX5_CAP_GEN(dev->mdev, end_pad)) {
5117 			mlx5_ib_dbg(dev, "Scatter end padding is not supported\n");
5118 			err = -EOPNOTSUPP;
5119 			goto out;
5120 		} else {
5121 			MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
5122 		}
5123 	}
5124 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
5125 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ) {
5126 		/*
5127 		 * In Firmware number of strides in each WQE is:
5128 		 *   "512 * 2^single_wqe_log_num_of_strides"
5129 		 * Values 3 to 8 are accepted as 10 to 15, 9 to 18 are
5130 		 * accepted as 0 to 9
5131 		 */
5132 		static const u8 fw_map[] = { 10, 11, 12, 13, 14, 15, 0, 1,
5133 					     2,  3,  4,  5,  6,  7,  8, 9 };
5134 		MLX5_SET(wq, wq, two_byte_shift_en, rwq->two_byte_shift_en);
5135 		MLX5_SET(wq, wq, log_wqe_stride_size,
5136 			 rwq->single_stride_log_num_of_bytes -
5137 			 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES);
5138 		MLX5_SET(wq, wq, log_wqe_num_of_strides,
5139 			 fw_map[rwq->log_num_strides -
5140 				MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES]);
5141 	}
5142 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
5143 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
5144 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
5145 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
5146 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
5147 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
5148 	has_net_offloads = MLX5_CAP_GEN(dev->mdev, eth_net_offloads);
5149 	if (init_attr->create_flags & IB_WQ_FLAGS_CVLAN_STRIPPING) {
5150 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
5151 			mlx5_ib_dbg(dev, "VLAN offloads are not supported\n");
5152 			err = -EOPNOTSUPP;
5153 			goto out;
5154 		}
5155 	} else {
5156 		MLX5_SET(rqc, rqc, vsd, 1);
5157 	}
5158 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS) {
5159 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, scatter_fcs))) {
5160 			mlx5_ib_dbg(dev, "Scatter FCS is not supported\n");
5161 			err = -EOPNOTSUPP;
5162 			goto out;
5163 		}
5164 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
5165 	}
5166 	if (init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
5167 		if (!(dev->ib_dev.attrs.raw_packet_caps &
5168 		      IB_RAW_PACKET_CAP_DELAY_DROP)) {
5169 			mlx5_ib_dbg(dev, "Delay drop is not supported\n");
5170 			err = -EOPNOTSUPP;
5171 			goto out;
5172 		}
5173 		MLX5_SET(rqc, rqc, delay_drop_en, 1);
5174 	}
5175 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
5176 	mlx5_ib_populate_pas(rwq->umem, 1UL << rwq->page_shift, rq_pas0, 0);
5177 	err = mlx5_core_create_rq_tracked(dev, in, inlen, &rwq->core_qp);
5178 	if (!err && init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
5179 		err = set_delay_drop(dev);
5180 		if (err) {
5181 			mlx5_ib_warn(dev, "Failed to enable delay drop err=%d\n",
5182 				     err);
5183 			mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5184 		} else {
5185 			rwq->create_flags |= MLX5_IB_WQ_FLAGS_DELAY_DROP;
5186 		}
5187 	}
5188 out:
5189 	kvfree(in);
5190 	return err;
5191 }
5192 
5193 static int set_user_rq_size(struct mlx5_ib_dev *dev,
5194 			    struct ib_wq_init_attr *wq_init_attr,
5195 			    struct mlx5_ib_create_wq *ucmd,
5196 			    struct mlx5_ib_rwq *rwq)
5197 {
5198 	/* Sanity check RQ size before proceeding */
5199 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
5200 		return -EINVAL;
5201 
5202 	if (!ucmd->rq_wqe_count)
5203 		return -EINVAL;
5204 
5205 	rwq->wqe_count = ucmd->rq_wqe_count;
5206 	rwq->wqe_shift = ucmd->rq_wqe_shift;
5207 	if (check_shl_overflow(rwq->wqe_count, rwq->wqe_shift, &rwq->buf_size))
5208 		return -EINVAL;
5209 
5210 	rwq->log_rq_stride = rwq->wqe_shift;
5211 	rwq->log_rq_size = ilog2(rwq->wqe_count);
5212 	return 0;
5213 }
5214 
5215 static bool log_of_strides_valid(struct mlx5_ib_dev *dev, u32 log_num_strides)
5216 {
5217 	if ((log_num_strides > MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES) ||
5218 	    (log_num_strides < MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
5219 		return false;
5220 
5221 	if (!MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) &&
5222 	    (log_num_strides < MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
5223 		return false;
5224 
5225 	return true;
5226 }
5227 
5228 static int prepare_user_rq(struct ib_pd *pd,
5229 			   struct ib_wq_init_attr *init_attr,
5230 			   struct ib_udata *udata,
5231 			   struct mlx5_ib_rwq *rwq)
5232 {
5233 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
5234 	struct mlx5_ib_create_wq ucmd = {};
5235 	int err;
5236 	size_t required_cmd_sz;
5237 
5238 	required_cmd_sz = offsetofend(struct mlx5_ib_create_wq,
5239 				      single_stride_log_num_of_bytes);
5240 	if (udata->inlen < required_cmd_sz) {
5241 		mlx5_ib_dbg(dev, "invalid inlen\n");
5242 		return -EINVAL;
5243 	}
5244 
5245 	if (udata->inlen > sizeof(ucmd) &&
5246 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
5247 				 udata->inlen - sizeof(ucmd))) {
5248 		mlx5_ib_dbg(dev, "inlen is not supported\n");
5249 		return -EOPNOTSUPP;
5250 	}
5251 
5252 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
5253 		mlx5_ib_dbg(dev, "copy failed\n");
5254 		return -EFAULT;
5255 	}
5256 
5257 	if (ucmd.comp_mask & (~MLX5_IB_CREATE_WQ_STRIDING_RQ)) {
5258 		mlx5_ib_dbg(dev, "invalid comp mask\n");
5259 		return -EOPNOTSUPP;
5260 	} else if (ucmd.comp_mask & MLX5_IB_CREATE_WQ_STRIDING_RQ) {
5261 		if (!MLX5_CAP_GEN(dev->mdev, striding_rq)) {
5262 			mlx5_ib_dbg(dev, "Striding RQ is not supported\n");
5263 			return -EOPNOTSUPP;
5264 		}
5265 		if ((ucmd.single_stride_log_num_of_bytes <
5266 		    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES) ||
5267 		    (ucmd.single_stride_log_num_of_bytes >
5268 		     MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES)) {
5269 			mlx5_ib_dbg(dev, "Invalid log stride size (%u. Range is %u - %u)\n",
5270 				    ucmd.single_stride_log_num_of_bytes,
5271 				    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES,
5272 				    MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES);
5273 			return -EINVAL;
5274 		}
5275 		if (!log_of_strides_valid(dev,
5276 					  ucmd.single_wqe_log_num_of_strides)) {
5277 			mlx5_ib_dbg(
5278 				dev,
5279 				"Invalid log num strides (%u. Range is %u - %u)\n",
5280 				ucmd.single_wqe_log_num_of_strides,
5281 				MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) ?
5282 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES :
5283 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES,
5284 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES);
5285 			return -EINVAL;
5286 		}
5287 		rwq->single_stride_log_num_of_bytes =
5288 			ucmd.single_stride_log_num_of_bytes;
5289 		rwq->log_num_strides = ucmd.single_wqe_log_num_of_strides;
5290 		rwq->two_byte_shift_en = !!ucmd.two_byte_shift_en;
5291 		rwq->create_flags |= MLX5_IB_WQ_FLAGS_STRIDING_RQ;
5292 	}
5293 
5294 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
5295 	if (err) {
5296 		mlx5_ib_dbg(dev, "err %d\n", err);
5297 		return err;
5298 	}
5299 
5300 	err = create_user_rq(dev, pd, udata, rwq, &ucmd);
5301 	if (err) {
5302 		mlx5_ib_dbg(dev, "err %d\n", err);
5303 		return err;
5304 	}
5305 
5306 	rwq->user_index = ucmd.user_index;
5307 	return 0;
5308 }
5309 
5310 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
5311 				struct ib_wq_init_attr *init_attr,
5312 				struct ib_udata *udata)
5313 {
5314 	struct mlx5_ib_dev *dev;
5315 	struct mlx5_ib_rwq *rwq;
5316 	struct mlx5_ib_create_wq_resp resp = {};
5317 	size_t min_resp_len;
5318 	int err;
5319 
5320 	if (!udata)
5321 		return ERR_PTR(-ENOSYS);
5322 
5323 	min_resp_len = offsetofend(struct mlx5_ib_create_wq_resp, reserved);
5324 	if (udata->outlen && udata->outlen < min_resp_len)
5325 		return ERR_PTR(-EINVAL);
5326 
5327 	if (!capable(CAP_SYS_RAWIO) &&
5328 	    init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP)
5329 		return ERR_PTR(-EPERM);
5330 
5331 	dev = to_mdev(pd->device);
5332 	switch (init_attr->wq_type) {
5333 	case IB_WQT_RQ:
5334 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
5335 		if (!rwq)
5336 			return ERR_PTR(-ENOMEM);
5337 		err = prepare_user_rq(pd, init_attr, udata, rwq);
5338 		if (err)
5339 			goto err;
5340 		err = create_rq(rwq, pd, init_attr);
5341 		if (err)
5342 			goto err_user_rq;
5343 		break;
5344 	default:
5345 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
5346 			    init_attr->wq_type);
5347 		return ERR_PTR(-EINVAL);
5348 	}
5349 
5350 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
5351 	rwq->ibwq.state = IB_WQS_RESET;
5352 	if (udata->outlen) {
5353 		resp.response_length = offsetofend(
5354 			struct mlx5_ib_create_wq_resp, response_length);
5355 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
5356 		if (err)
5357 			goto err_copy;
5358 	}
5359 
5360 	rwq->core_qp.event = mlx5_ib_wq_event;
5361 	rwq->ibwq.event_handler = init_attr->event_handler;
5362 	return &rwq->ibwq;
5363 
5364 err_copy:
5365 	mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5366 err_user_rq:
5367 	destroy_user_rq(dev, pd, rwq, udata);
5368 err:
5369 	kfree(rwq);
5370 	return ERR_PTR(err);
5371 }
5372 
5373 int mlx5_ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
5374 {
5375 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
5376 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
5377 	int ret;
5378 
5379 	ret = mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5380 	if (ret)
5381 		return ret;
5382 	destroy_user_rq(dev, wq->pd, rwq, udata);
5383 	kfree(rwq);
5384 	return 0;
5385 }
5386 
5387 int mlx5_ib_create_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_table,
5388 				 struct ib_rwq_ind_table_init_attr *init_attr,
5389 				 struct ib_udata *udata)
5390 {
5391 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl =
5392 		to_mrwq_ind_table(ib_rwq_ind_table);
5393 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_table->device);
5394 	int sz = 1 << init_attr->log_ind_tbl_size;
5395 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
5396 	size_t min_resp_len;
5397 	int inlen;
5398 	int err;
5399 	int i;
5400 	u32 *in;
5401 	void *rqtc;
5402 
5403 	if (udata->inlen > 0 &&
5404 	    !ib_is_udata_cleared(udata, 0,
5405 				 udata->inlen))
5406 		return -EOPNOTSUPP;
5407 
5408 	if (init_attr->log_ind_tbl_size >
5409 	    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
5410 		mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
5411 			    init_attr->log_ind_tbl_size,
5412 			    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
5413 		return -EINVAL;
5414 	}
5415 
5416 	min_resp_len =
5417 		offsetofend(struct mlx5_ib_create_rwq_ind_tbl_resp, reserved);
5418 	if (udata->outlen && udata->outlen < min_resp_len)
5419 		return -EINVAL;
5420 
5421 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
5422 	in = kvzalloc(inlen, GFP_KERNEL);
5423 	if (!in)
5424 		return -ENOMEM;
5425 
5426 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
5427 
5428 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
5429 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
5430 
5431 	for (i = 0; i < sz; i++)
5432 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
5433 
5434 	rwq_ind_tbl->uid = to_mpd(init_attr->ind_tbl[0]->pd)->uid;
5435 	MLX5_SET(create_rqt_in, in, uid, rwq_ind_tbl->uid);
5436 
5437 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
5438 	kvfree(in);
5439 	if (err)
5440 		return err;
5441 
5442 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
5443 	if (udata->outlen) {
5444 		resp.response_length =
5445 			offsetofend(struct mlx5_ib_create_rwq_ind_tbl_resp,
5446 				    response_length);
5447 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
5448 		if (err)
5449 			goto err_copy;
5450 	}
5451 
5452 	return 0;
5453 
5454 err_copy:
5455 	mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
5456 	return err;
5457 }
5458 
5459 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
5460 {
5461 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
5462 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
5463 
5464 	return mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
5465 }
5466 
5467 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
5468 		      u32 wq_attr_mask, struct ib_udata *udata)
5469 {
5470 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
5471 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
5472 	struct mlx5_ib_modify_wq ucmd = {};
5473 	size_t required_cmd_sz;
5474 	int curr_wq_state;
5475 	int wq_state;
5476 	int inlen;
5477 	int err;
5478 	void *rqc;
5479 	void *in;
5480 
5481 	required_cmd_sz = offsetofend(struct mlx5_ib_modify_wq, reserved);
5482 	if (udata->inlen < required_cmd_sz)
5483 		return -EINVAL;
5484 
5485 	if (udata->inlen > sizeof(ucmd) &&
5486 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
5487 				 udata->inlen - sizeof(ucmd)))
5488 		return -EOPNOTSUPP;
5489 
5490 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
5491 		return -EFAULT;
5492 
5493 	if (ucmd.comp_mask || ucmd.reserved)
5494 		return -EOPNOTSUPP;
5495 
5496 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
5497 	in = kvzalloc(inlen, GFP_KERNEL);
5498 	if (!in)
5499 		return -ENOMEM;
5500 
5501 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
5502 
5503 	curr_wq_state = wq_attr->curr_wq_state;
5504 	wq_state = wq_attr->wq_state;
5505 	if (curr_wq_state == IB_WQS_ERR)
5506 		curr_wq_state = MLX5_RQC_STATE_ERR;
5507 	if (wq_state == IB_WQS_ERR)
5508 		wq_state = MLX5_RQC_STATE_ERR;
5509 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
5510 	MLX5_SET(modify_rq_in, in, uid, to_mpd(wq->pd)->uid);
5511 	MLX5_SET(rqc, rqc, state, wq_state);
5512 
5513 	if (wq_attr_mask & IB_WQ_FLAGS) {
5514 		if (wq_attr->flags_mask & IB_WQ_FLAGS_CVLAN_STRIPPING) {
5515 			if (!(MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
5516 			      MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
5517 				mlx5_ib_dbg(dev, "VLAN offloads are not "
5518 					    "supported\n");
5519 				err = -EOPNOTSUPP;
5520 				goto out;
5521 			}
5522 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
5523 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
5524 			MLX5_SET(rqc, rqc, vsd,
5525 				 (wq_attr->flags & IB_WQ_FLAGS_CVLAN_STRIPPING) ? 0 : 1);
5526 		}
5527 
5528 		if (wq_attr->flags_mask & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
5529 			mlx5_ib_dbg(dev, "Modifying scatter end padding is not supported\n");
5530 			err = -EOPNOTSUPP;
5531 			goto out;
5532 		}
5533 	}
5534 
5535 	if (curr_wq_state == IB_WQS_RESET && wq_state == IB_WQS_RDY) {
5536 		u16 set_id;
5537 
5538 		set_id = mlx5_ib_get_counters_id(dev, 0);
5539 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
5540 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
5541 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
5542 			MLX5_SET(rqc, rqc, counter_set_id, set_id);
5543 		} else
5544 			dev_info_once(
5545 				&dev->ib_dev.dev,
5546 				"Receive WQ counters are not supported on current FW\n");
5547 	}
5548 
5549 	err = mlx5_core_modify_rq(dev->mdev, rwq->core_qp.qpn, in);
5550 	if (!err)
5551 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
5552 
5553 out:
5554 	kvfree(in);
5555 	return err;
5556 }
5557 
5558 struct mlx5_ib_drain_cqe {
5559 	struct ib_cqe cqe;
5560 	struct completion done;
5561 };
5562 
5563 static void mlx5_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
5564 {
5565 	struct mlx5_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
5566 						     struct mlx5_ib_drain_cqe,
5567 						     cqe);
5568 
5569 	complete(&cqe->done);
5570 }
5571 
5572 /* This function returns only once the drained WR was completed */
5573 static void handle_drain_completion(struct ib_cq *cq,
5574 				    struct mlx5_ib_drain_cqe *sdrain,
5575 				    struct mlx5_ib_dev *dev)
5576 {
5577 	struct mlx5_core_dev *mdev = dev->mdev;
5578 
5579 	if (cq->poll_ctx == IB_POLL_DIRECT) {
5580 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
5581 			ib_process_cq_direct(cq, -1);
5582 		return;
5583 	}
5584 
5585 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5586 		struct mlx5_ib_cq *mcq = to_mcq(cq);
5587 		bool triggered = false;
5588 		unsigned long flags;
5589 
5590 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
5591 		/* Make sure that the CQ handler won't run if wasn't run yet */
5592 		if (!mcq->mcq.reset_notify_added)
5593 			mcq->mcq.reset_notify_added = 1;
5594 		else
5595 			triggered = true;
5596 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
5597 
5598 		if (triggered) {
5599 			/* Wait for any scheduled/running task to be ended */
5600 			switch (cq->poll_ctx) {
5601 			case IB_POLL_SOFTIRQ:
5602 				irq_poll_disable(&cq->iop);
5603 				irq_poll_enable(&cq->iop);
5604 				break;
5605 			case IB_POLL_WORKQUEUE:
5606 				cancel_work_sync(&cq->work);
5607 				break;
5608 			default:
5609 				WARN_ON_ONCE(1);
5610 			}
5611 		}
5612 
5613 		/* Run the CQ handler - this makes sure that the drain WR will
5614 		 * be processed if wasn't processed yet.
5615 		 */
5616 		mcq->mcq.comp(&mcq->mcq, NULL);
5617 	}
5618 
5619 	wait_for_completion(&sdrain->done);
5620 }
5621 
5622 void mlx5_ib_drain_sq(struct ib_qp *qp)
5623 {
5624 	struct ib_cq *cq = qp->send_cq;
5625 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
5626 	struct mlx5_ib_drain_cqe sdrain;
5627 	const struct ib_send_wr *bad_swr;
5628 	struct ib_rdma_wr swr = {
5629 		.wr = {
5630 			.next = NULL,
5631 			{ .wr_cqe	= &sdrain.cqe, },
5632 			.opcode	= IB_WR_RDMA_WRITE,
5633 		},
5634 	};
5635 	int ret;
5636 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5637 	struct mlx5_core_dev *mdev = dev->mdev;
5638 
5639 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
5640 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5641 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
5642 		return;
5643 	}
5644 
5645 	sdrain.cqe.done = mlx5_ib_drain_qp_done;
5646 	init_completion(&sdrain.done);
5647 
5648 	ret = mlx5_ib_post_send_drain(qp, &swr.wr, &bad_swr);
5649 	if (ret) {
5650 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
5651 		return;
5652 	}
5653 
5654 	handle_drain_completion(cq, &sdrain, dev);
5655 }
5656 
5657 void mlx5_ib_drain_rq(struct ib_qp *qp)
5658 {
5659 	struct ib_cq *cq = qp->recv_cq;
5660 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
5661 	struct mlx5_ib_drain_cqe rdrain;
5662 	struct ib_recv_wr rwr = {};
5663 	const struct ib_recv_wr *bad_rwr;
5664 	int ret;
5665 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5666 	struct mlx5_core_dev *mdev = dev->mdev;
5667 
5668 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
5669 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5670 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
5671 		return;
5672 	}
5673 
5674 	rwr.wr_cqe = &rdrain.cqe;
5675 	rdrain.cqe.done = mlx5_ib_drain_qp_done;
5676 	init_completion(&rdrain.done);
5677 
5678 	ret = mlx5_ib_post_recv_drain(qp, &rwr, &bad_rwr);
5679 	if (ret) {
5680 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
5681 		return;
5682 	}
5683 
5684 	handle_drain_completion(cq, &rdrain, dev);
5685 }
5686 
5687 /*
5688  * Bind a qp to a counter. If @counter is NULL then bind the qp to
5689  * the default counter
5690  */
5691 int mlx5_ib_qp_set_counter(struct ib_qp *qp, struct rdma_counter *counter)
5692 {
5693 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5694 	struct mlx5_ib_qp *mqp = to_mqp(qp);
5695 	int err = 0;
5696 
5697 	mutex_lock(&mqp->mutex);
5698 	if (mqp->state == IB_QPS_RESET) {
5699 		qp->counter = counter;
5700 		goto out;
5701 	}
5702 
5703 	if (!MLX5_CAP_GEN(dev->mdev, rts2rts_qp_counters_set_id)) {
5704 		err = -EOPNOTSUPP;
5705 		goto out;
5706 	}
5707 
5708 	if (mqp->state == IB_QPS_RTS) {
5709 		err = __mlx5_ib_qp_set_counter(qp, counter);
5710 		if (!err)
5711 			qp->counter = counter;
5712 
5713 		goto out;
5714 	}
5715 
5716 	mqp->counter_pending = 1;
5717 	qp->counter = counter;
5718 
5719 out:
5720 	mutex_unlock(&mqp->mutex);
5721 	return err;
5722 }
5723