xref: /linux/drivers/infiniband/hw/mlx5/qp.c (revision 69bfec7548f4c1595bac0e3ddfc0458a5af31f4c)
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 struct mlx5_ib_qp_event_work {
75 	struct work_struct work;
76 	struct mlx5_core_qp *qp;
77 	int type;
78 };
79 
80 static struct workqueue_struct *mlx5_ib_qp_event_wq;
81 
82 static void get_cqs(enum ib_qp_type qp_type,
83 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
84 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
85 
86 static int is_qp0(enum ib_qp_type qp_type)
87 {
88 	return qp_type == IB_QPT_SMI;
89 }
90 
91 static int is_sqp(enum ib_qp_type qp_type)
92 {
93 	return is_qp0(qp_type) || is_qp1(qp_type);
94 }
95 
96 /**
97  * mlx5_ib_read_user_wqe_common() - Copy a WQE (or part of) from user WQ
98  * to kernel buffer
99  *
100  * @umem: User space memory where the WQ is
101  * @buffer: buffer to copy to
102  * @buflen: buffer length
103  * @wqe_index: index of WQE to copy from
104  * @wq_offset: offset to start of WQ
105  * @wq_wqe_cnt: number of WQEs in WQ
106  * @wq_wqe_shift: log2 of WQE size
107  * @bcnt: number of bytes to copy
108  * @bytes_copied: number of bytes to copy (return value)
109  *
110  * Copies from start of WQE bcnt or less bytes.
111  * Does not gurantee to copy the entire WQE.
112  *
113  * Return: zero on success, or an error code.
114  */
115 static int mlx5_ib_read_user_wqe_common(struct ib_umem *umem, void *buffer,
116 					size_t buflen, int wqe_index,
117 					int wq_offset, int wq_wqe_cnt,
118 					int wq_wqe_shift, int bcnt,
119 					size_t *bytes_copied)
120 {
121 	size_t offset = wq_offset + ((wqe_index % wq_wqe_cnt) << wq_wqe_shift);
122 	size_t wq_end = wq_offset + (wq_wqe_cnt << wq_wqe_shift);
123 	size_t copy_length;
124 	int ret;
125 
126 	/* don't copy more than requested, more than buffer length or
127 	 * beyond WQ end
128 	 */
129 	copy_length = min_t(u32, buflen, wq_end - offset);
130 	copy_length = min_t(u32, copy_length, bcnt);
131 
132 	ret = ib_umem_copy_from(buffer, umem, offset, copy_length);
133 	if (ret)
134 		return ret;
135 
136 	if (!ret && bytes_copied)
137 		*bytes_copied = copy_length;
138 
139 	return 0;
140 }
141 
142 static int mlx5_ib_read_kernel_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index,
143 				      void *buffer, size_t buflen, size_t *bc)
144 {
145 	struct mlx5_wqe_ctrl_seg *ctrl;
146 	size_t bytes_copied = 0;
147 	size_t wqe_length;
148 	void *p;
149 	int ds;
150 
151 	wqe_index = wqe_index & qp->sq.fbc.sz_m1;
152 
153 	/* read the control segment first */
154 	p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, wqe_index);
155 	ctrl = p;
156 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
157 	wqe_length = ds * MLX5_WQE_DS_UNITS;
158 
159 	/* read rest of WQE if it spreads over more than one stride */
160 	while (bytes_copied < wqe_length) {
161 		size_t copy_length =
162 			min_t(size_t, buflen - bytes_copied, MLX5_SEND_WQE_BB);
163 
164 		if (!copy_length)
165 			break;
166 
167 		memcpy(buffer + bytes_copied, p, copy_length);
168 		bytes_copied += copy_length;
169 
170 		wqe_index = (wqe_index + 1) & qp->sq.fbc.sz_m1;
171 		p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, wqe_index);
172 	}
173 	*bc = bytes_copied;
174 	return 0;
175 }
176 
177 static int mlx5_ib_read_user_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index,
178 				    void *buffer, size_t buflen, size_t *bc)
179 {
180 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
181 	struct ib_umem *umem = base->ubuffer.umem;
182 	struct mlx5_ib_wq *wq = &qp->sq;
183 	struct mlx5_wqe_ctrl_seg *ctrl;
184 	size_t bytes_copied;
185 	size_t bytes_copied2;
186 	size_t wqe_length;
187 	int ret;
188 	int ds;
189 
190 	/* at first read as much as possible */
191 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index,
192 					   wq->offset, wq->wqe_cnt,
193 					   wq->wqe_shift, buflen,
194 					   &bytes_copied);
195 	if (ret)
196 		return ret;
197 
198 	/* we need at least control segment size to proceed */
199 	if (bytes_copied < sizeof(*ctrl))
200 		return -EINVAL;
201 
202 	ctrl = buffer;
203 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
204 	wqe_length = ds * MLX5_WQE_DS_UNITS;
205 
206 	/* if we copied enough then we are done */
207 	if (bytes_copied >= wqe_length) {
208 		*bc = bytes_copied;
209 		return 0;
210 	}
211 
212 	/* otherwise this a wrapped around wqe
213 	 * so read the remaining bytes starting
214 	 * from  wqe_index 0
215 	 */
216 	ret = mlx5_ib_read_user_wqe_common(umem, buffer + bytes_copied,
217 					   buflen - bytes_copied, 0, wq->offset,
218 					   wq->wqe_cnt, wq->wqe_shift,
219 					   wqe_length - bytes_copied,
220 					   &bytes_copied2);
221 
222 	if (ret)
223 		return ret;
224 	*bc = bytes_copied + bytes_copied2;
225 	return 0;
226 }
227 
228 int mlx5_ib_read_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
229 			size_t buflen, size_t *bc)
230 {
231 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
232 	struct ib_umem *umem = base->ubuffer.umem;
233 
234 	if (buflen < sizeof(struct mlx5_wqe_ctrl_seg))
235 		return -EINVAL;
236 
237 	if (!umem)
238 		return mlx5_ib_read_kernel_wqe_sq(qp, wqe_index, buffer,
239 						  buflen, bc);
240 
241 	return mlx5_ib_read_user_wqe_sq(qp, wqe_index, buffer, buflen, bc);
242 }
243 
244 static int mlx5_ib_read_user_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index,
245 				    void *buffer, size_t buflen, size_t *bc)
246 {
247 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
248 	struct ib_umem *umem = base->ubuffer.umem;
249 	struct mlx5_ib_wq *wq = &qp->rq;
250 	size_t bytes_copied;
251 	int ret;
252 
253 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index,
254 					   wq->offset, wq->wqe_cnt,
255 					   wq->wqe_shift, buflen,
256 					   &bytes_copied);
257 
258 	if (ret)
259 		return ret;
260 	*bc = bytes_copied;
261 	return 0;
262 }
263 
264 int mlx5_ib_read_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
265 			size_t buflen, size_t *bc)
266 {
267 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
268 	struct ib_umem *umem = base->ubuffer.umem;
269 	struct mlx5_ib_wq *wq = &qp->rq;
270 	size_t wqe_size = 1 << wq->wqe_shift;
271 
272 	if (buflen < wqe_size)
273 		return -EINVAL;
274 
275 	if (!umem)
276 		return -EOPNOTSUPP;
277 
278 	return mlx5_ib_read_user_wqe_rq(qp, wqe_index, buffer, buflen, bc);
279 }
280 
281 static int mlx5_ib_read_user_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index,
282 				     void *buffer, size_t buflen, size_t *bc)
283 {
284 	struct ib_umem *umem = srq->umem;
285 	size_t bytes_copied;
286 	int ret;
287 
288 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index, 0,
289 					   srq->msrq.max, srq->msrq.wqe_shift,
290 					   buflen, &bytes_copied);
291 
292 	if (ret)
293 		return ret;
294 	*bc = bytes_copied;
295 	return 0;
296 }
297 
298 int mlx5_ib_read_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index, void *buffer,
299 			 size_t buflen, size_t *bc)
300 {
301 	struct ib_umem *umem = srq->umem;
302 	size_t wqe_size = 1 << srq->msrq.wqe_shift;
303 
304 	if (buflen < wqe_size)
305 		return -EINVAL;
306 
307 	if (!umem)
308 		return -EOPNOTSUPP;
309 
310 	return mlx5_ib_read_user_wqe_srq(srq, wqe_index, buffer, buflen, bc);
311 }
312 
313 static void mlx5_ib_qp_err_syndrome(struct ib_qp *ibqp)
314 {
315 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
316 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
317 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
318 	void *pas_ext_union, *err_syn;
319 	u32 *outb;
320 	int err;
321 
322 	if (!MLX5_CAP_GEN(dev->mdev, qpc_extension) ||
323 	    !MLX5_CAP_GEN(dev->mdev, qp_error_syndrome))
324 		return;
325 
326 	outb = kzalloc(outlen, GFP_KERNEL);
327 	if (!outb)
328 		return;
329 
330 	err = mlx5_core_qp_query(dev, &qp->trans_qp.base.mqp, outb, outlen,
331 				 true);
332 	if (err)
333 		goto out;
334 
335 	pas_ext_union =
336 		MLX5_ADDR_OF(query_qp_out, outb, qp_pas_or_qpc_ext_and_pas);
337 	err_syn = MLX5_ADDR_OF(qpc_extension_and_pas_list_in, pas_ext_union,
338 			       qpc_data_extension.error_syndrome);
339 
340 	pr_err("%s/%d: QP %d error: %s (0x%x 0x%x 0x%x)\n",
341 	       ibqp->device->name, ibqp->port, ibqp->qp_num,
342 	       ib_wc_status_msg(
343 		       MLX5_GET(cqe_error_syndrome, err_syn, syndrome)),
344 	       MLX5_GET(cqe_error_syndrome, err_syn, vendor_error_syndrome),
345 	       MLX5_GET(cqe_error_syndrome, err_syn, hw_syndrome_type),
346 	       MLX5_GET(cqe_error_syndrome, err_syn, hw_error_syndrome));
347 out:
348 	kfree(outb);
349 }
350 
351 static void mlx5_ib_handle_qp_event(struct work_struct *_work)
352 {
353 	struct mlx5_ib_qp_event_work *qpe_work =
354 		container_of(_work, struct mlx5_ib_qp_event_work, work);
355 	struct ib_qp *ibqp = &to_mibqp(qpe_work->qp)->ibqp;
356 	struct ib_event event = {};
357 
358 	event.device = ibqp->device;
359 	event.element.qp = ibqp;
360 	switch (qpe_work->type) {
361 	case MLX5_EVENT_TYPE_PATH_MIG:
362 		event.event = IB_EVENT_PATH_MIG;
363 		break;
364 	case MLX5_EVENT_TYPE_COMM_EST:
365 		event.event = IB_EVENT_COMM_EST;
366 		break;
367 	case MLX5_EVENT_TYPE_SQ_DRAINED:
368 		event.event = IB_EVENT_SQ_DRAINED;
369 		break;
370 	case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
371 		event.event = IB_EVENT_QP_LAST_WQE_REACHED;
372 		break;
373 	case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
374 		event.event = IB_EVENT_QP_FATAL;
375 		break;
376 	case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
377 		event.event = IB_EVENT_PATH_MIG_ERR;
378 		break;
379 	case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
380 		event.event = IB_EVENT_QP_REQ_ERR;
381 		break;
382 	case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
383 		event.event = IB_EVENT_QP_ACCESS_ERR;
384 		break;
385 	default:
386 		pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n",
387 			qpe_work->type, qpe_work->qp->qpn);
388 		goto out;
389 	}
390 
391 	if ((event.event == IB_EVENT_QP_FATAL) ||
392 	    (event.event == IB_EVENT_QP_ACCESS_ERR))
393 		mlx5_ib_qp_err_syndrome(ibqp);
394 
395 	ibqp->event_handler(&event, ibqp->qp_context);
396 
397 out:
398 	mlx5_core_res_put(&qpe_work->qp->common);
399 	kfree(qpe_work);
400 }
401 
402 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
403 {
404 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
405 	struct mlx5_ib_qp_event_work *qpe_work;
406 
407 	if (type == MLX5_EVENT_TYPE_PATH_MIG) {
408 		/* This event is only valid for trans_qps */
409 		to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
410 	}
411 
412 	if (!ibqp->event_handler)
413 		goto out_no_handler;
414 
415 	qpe_work = kzalloc(sizeof(*qpe_work), GFP_ATOMIC);
416 	if (!qpe_work)
417 		goto out_no_handler;
418 
419 	qpe_work->qp = qp;
420 	qpe_work->type = type;
421 	INIT_WORK(&qpe_work->work, mlx5_ib_handle_qp_event);
422 	queue_work(mlx5_ib_qp_event_wq, &qpe_work->work);
423 	return;
424 
425 out_no_handler:
426 	mlx5_core_res_put(&qp->common);
427 }
428 
429 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
430 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
431 {
432 	int wqe_size;
433 	int wq_size;
434 
435 	/* Sanity check RQ size before proceeding */
436 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
437 		return -EINVAL;
438 
439 	if (!has_rq) {
440 		qp->rq.max_gs = 0;
441 		qp->rq.wqe_cnt = 0;
442 		qp->rq.wqe_shift = 0;
443 		cap->max_recv_wr = 0;
444 		cap->max_recv_sge = 0;
445 	} else {
446 		int wq_sig = !!(qp->flags_en & MLX5_QP_FLAG_SIGNATURE);
447 
448 		if (ucmd) {
449 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
450 			if (ucmd->rq_wqe_shift > BITS_PER_BYTE * sizeof(ucmd->rq_wqe_shift))
451 				return -EINVAL;
452 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
453 			if ((1 << qp->rq.wqe_shift) /
454 				    sizeof(struct mlx5_wqe_data_seg) <
455 			    wq_sig)
456 				return -EINVAL;
457 			qp->rq.max_gs =
458 				(1 << qp->rq.wqe_shift) /
459 					sizeof(struct mlx5_wqe_data_seg) -
460 				wq_sig;
461 			qp->rq.max_post = qp->rq.wqe_cnt;
462 		} else {
463 			wqe_size =
464 				wq_sig ? sizeof(struct mlx5_wqe_signature_seg) :
465 					 0;
466 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
467 			wqe_size = roundup_pow_of_two(wqe_size);
468 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
469 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
470 			qp->rq.wqe_cnt = wq_size / wqe_size;
471 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
472 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
473 					    wqe_size,
474 					    MLX5_CAP_GEN(dev->mdev,
475 							 max_wqe_sz_rq));
476 				return -EINVAL;
477 			}
478 			qp->rq.wqe_shift = ilog2(wqe_size);
479 			qp->rq.max_gs =
480 				(1 << qp->rq.wqe_shift) /
481 					sizeof(struct mlx5_wqe_data_seg) -
482 				wq_sig;
483 			qp->rq.max_post = qp->rq.wqe_cnt;
484 		}
485 	}
486 
487 	return 0;
488 }
489 
490 static int sq_overhead(struct ib_qp_init_attr *attr)
491 {
492 	int size = 0;
493 
494 	switch (attr->qp_type) {
495 	case IB_QPT_XRC_INI:
496 		size += sizeof(struct mlx5_wqe_xrc_seg);
497 		fallthrough;
498 	case IB_QPT_RC:
499 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
500 			max(sizeof(struct mlx5_wqe_atomic_seg) +
501 			    sizeof(struct mlx5_wqe_raddr_seg),
502 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
503 			    sizeof(struct mlx5_mkey_seg) +
504 			    MLX5_IB_SQ_UMR_INLINE_THRESHOLD /
505 			    MLX5_IB_UMR_OCTOWORD);
506 		break;
507 
508 	case IB_QPT_XRC_TGT:
509 		return 0;
510 
511 	case IB_QPT_UC:
512 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
513 			max(sizeof(struct mlx5_wqe_raddr_seg),
514 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
515 			    sizeof(struct mlx5_mkey_seg));
516 		break;
517 
518 	case IB_QPT_UD:
519 		if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
520 			size += sizeof(struct mlx5_wqe_eth_pad) +
521 				sizeof(struct mlx5_wqe_eth_seg);
522 		fallthrough;
523 	case IB_QPT_SMI:
524 	case MLX5_IB_QPT_HW_GSI:
525 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
526 			sizeof(struct mlx5_wqe_datagram_seg);
527 		break;
528 
529 	case MLX5_IB_QPT_REG_UMR:
530 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
531 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
532 			sizeof(struct mlx5_mkey_seg);
533 		break;
534 
535 	default:
536 		return -EINVAL;
537 	}
538 
539 	return size;
540 }
541 
542 static int calc_send_wqe(struct ib_qp_init_attr *attr)
543 {
544 	int inl_size = 0;
545 	int size;
546 
547 	size = sq_overhead(attr);
548 	if (size < 0)
549 		return size;
550 
551 	if (attr->cap.max_inline_data) {
552 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
553 			attr->cap.max_inline_data;
554 	}
555 
556 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
557 	if (attr->create_flags & IB_QP_CREATE_INTEGRITY_EN &&
558 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
559 		return MLX5_SIG_WQE_SIZE;
560 	else
561 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
562 }
563 
564 static int get_send_sge(struct ib_qp_init_attr *attr, int wqe_size)
565 {
566 	int max_sge;
567 
568 	if (attr->qp_type == IB_QPT_RC)
569 		max_sge = (min_t(int, wqe_size, 512) -
570 			   sizeof(struct mlx5_wqe_ctrl_seg) -
571 			   sizeof(struct mlx5_wqe_raddr_seg)) /
572 			sizeof(struct mlx5_wqe_data_seg);
573 	else if (attr->qp_type == IB_QPT_XRC_INI)
574 		max_sge = (min_t(int, wqe_size, 512) -
575 			   sizeof(struct mlx5_wqe_ctrl_seg) -
576 			   sizeof(struct mlx5_wqe_xrc_seg) -
577 			   sizeof(struct mlx5_wqe_raddr_seg)) /
578 			sizeof(struct mlx5_wqe_data_seg);
579 	else
580 		max_sge = (wqe_size - sq_overhead(attr)) /
581 			sizeof(struct mlx5_wqe_data_seg);
582 
583 	return min_t(int, max_sge, wqe_size - sq_overhead(attr) /
584 		     sizeof(struct mlx5_wqe_data_seg));
585 }
586 
587 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
588 			struct mlx5_ib_qp *qp)
589 {
590 	int wqe_size;
591 	int wq_size;
592 
593 	if (!attr->cap.max_send_wr)
594 		return 0;
595 
596 	wqe_size = calc_send_wqe(attr);
597 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
598 	if (wqe_size < 0)
599 		return wqe_size;
600 
601 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
602 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
603 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
604 		return -EINVAL;
605 	}
606 
607 	qp->max_inline_data = wqe_size - sq_overhead(attr) -
608 			      sizeof(struct mlx5_wqe_inline_seg);
609 	attr->cap.max_inline_data = qp->max_inline_data;
610 
611 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
612 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
613 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
614 		mlx5_ib_dbg(dev, "send queue size (%d * %d / %d -> %d) exceeds limits(%d)\n",
615 			    attr->cap.max_send_wr, wqe_size, MLX5_SEND_WQE_BB,
616 			    qp->sq.wqe_cnt,
617 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
618 		return -ENOMEM;
619 	}
620 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
621 	qp->sq.max_gs = get_send_sge(attr, wqe_size);
622 	if (qp->sq.max_gs < attr->cap.max_send_sge)
623 		return -ENOMEM;
624 
625 	attr->cap.max_send_sge = qp->sq.max_gs;
626 	qp->sq.max_post = wq_size / wqe_size;
627 	attr->cap.max_send_wr = qp->sq.max_post;
628 
629 	return wq_size;
630 }
631 
632 static int set_user_buf_size(struct mlx5_ib_dev *dev,
633 			    struct mlx5_ib_qp *qp,
634 			    struct mlx5_ib_create_qp *ucmd,
635 			    struct mlx5_ib_qp_base *base,
636 			    struct ib_qp_init_attr *attr)
637 {
638 	int desc_sz = 1 << qp->sq.wqe_shift;
639 
640 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
641 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
642 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
643 		return -EINVAL;
644 	}
645 
646 	if (ucmd->sq_wqe_count && !is_power_of_2(ucmd->sq_wqe_count)) {
647 		mlx5_ib_warn(dev, "sq_wqe_count %d is not a power of two\n",
648 			     ucmd->sq_wqe_count);
649 		return -EINVAL;
650 	}
651 
652 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
653 
654 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
655 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
656 			     qp->sq.wqe_cnt,
657 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
658 		return -EINVAL;
659 	}
660 
661 	if (attr->qp_type == IB_QPT_RAW_PACKET ||
662 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
663 		base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
664 		qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
665 	} else {
666 		base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
667 					 (qp->sq.wqe_cnt << 6);
668 	}
669 
670 	return 0;
671 }
672 
673 static int qp_has_rq(struct ib_qp_init_attr *attr)
674 {
675 	if (attr->qp_type == IB_QPT_XRC_INI ||
676 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
677 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
678 	    !attr->cap.max_recv_wr)
679 		return 0;
680 
681 	return 1;
682 }
683 
684 enum {
685 	/* this is the first blue flame register in the array of bfregs assigned
686 	 * to a processes. Since we do not use it for blue flame but rather
687 	 * regular 64 bit doorbells, we do not need a lock for maintaiing
688 	 * "odd/even" order
689 	 */
690 	NUM_NON_BLUE_FLAME_BFREGS = 1,
691 };
692 
693 static int max_bfregs(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi)
694 {
695 	return get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
696 	       bfregi->num_static_sys_pages * MLX5_NON_FP_BFREGS_PER_UAR;
697 }
698 
699 static int num_med_bfreg(struct mlx5_ib_dev *dev,
700 			 struct mlx5_bfreg_info *bfregi)
701 {
702 	int n;
703 
704 	n = max_bfregs(dev, bfregi) - bfregi->num_low_latency_bfregs -
705 	    NUM_NON_BLUE_FLAME_BFREGS;
706 
707 	return n >= 0 ? n : 0;
708 }
709 
710 static int first_med_bfreg(struct mlx5_ib_dev *dev,
711 			   struct mlx5_bfreg_info *bfregi)
712 {
713 	return num_med_bfreg(dev, bfregi) ? 1 : -ENOMEM;
714 }
715 
716 static int first_hi_bfreg(struct mlx5_ib_dev *dev,
717 			  struct mlx5_bfreg_info *bfregi)
718 {
719 	int med;
720 
721 	med = num_med_bfreg(dev, bfregi);
722 	return ++med;
723 }
724 
725 static int alloc_high_class_bfreg(struct mlx5_ib_dev *dev,
726 				  struct mlx5_bfreg_info *bfregi)
727 {
728 	int i;
729 
730 	for (i = first_hi_bfreg(dev, bfregi); i < max_bfregs(dev, bfregi); i++) {
731 		if (!bfregi->count[i]) {
732 			bfregi->count[i]++;
733 			return i;
734 		}
735 	}
736 
737 	return -ENOMEM;
738 }
739 
740 static int alloc_med_class_bfreg(struct mlx5_ib_dev *dev,
741 				 struct mlx5_bfreg_info *bfregi)
742 {
743 	int minidx = first_med_bfreg(dev, bfregi);
744 	int i;
745 
746 	if (minidx < 0)
747 		return minidx;
748 
749 	for (i = minidx; i < first_hi_bfreg(dev, bfregi); i++) {
750 		if (bfregi->count[i] < bfregi->count[minidx])
751 			minidx = i;
752 		if (!bfregi->count[minidx])
753 			break;
754 	}
755 
756 	bfregi->count[minidx]++;
757 	return minidx;
758 }
759 
760 static int alloc_bfreg(struct mlx5_ib_dev *dev,
761 		       struct mlx5_bfreg_info *bfregi)
762 {
763 	int bfregn = -ENOMEM;
764 
765 	if (bfregi->lib_uar_dyn)
766 		return -EINVAL;
767 
768 	mutex_lock(&bfregi->lock);
769 	if (bfregi->ver >= 2) {
770 		bfregn = alloc_high_class_bfreg(dev, bfregi);
771 		if (bfregn < 0)
772 			bfregn = alloc_med_class_bfreg(dev, bfregi);
773 	}
774 
775 	if (bfregn < 0) {
776 		BUILD_BUG_ON(NUM_NON_BLUE_FLAME_BFREGS != 1);
777 		bfregn = 0;
778 		bfregi->count[bfregn]++;
779 	}
780 	mutex_unlock(&bfregi->lock);
781 
782 	return bfregn;
783 }
784 
785 void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
786 {
787 	mutex_lock(&bfregi->lock);
788 	bfregi->count[bfregn]--;
789 	mutex_unlock(&bfregi->lock);
790 }
791 
792 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
793 {
794 	switch (state) {
795 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
796 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
797 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
798 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
799 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
800 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
801 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
802 	default:		return -1;
803 	}
804 }
805 
806 static int to_mlx5_st(enum ib_qp_type type)
807 {
808 	switch (type) {
809 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
810 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
811 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
812 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
813 	case IB_QPT_XRC_INI:
814 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
815 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
816 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
817 	case MLX5_IB_QPT_DCI:		return MLX5_QP_ST_DCI;
818 	case IB_QPT_RAW_PACKET:		return MLX5_QP_ST_RAW_ETHERTYPE;
819 	default:		return -EINVAL;
820 	}
821 }
822 
823 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
824 			     struct mlx5_ib_cq *recv_cq);
825 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
826 			       struct mlx5_ib_cq *recv_cq);
827 
828 int bfregn_to_uar_index(struct mlx5_ib_dev *dev,
829 			struct mlx5_bfreg_info *bfregi, u32 bfregn,
830 			bool dyn_bfreg)
831 {
832 	unsigned int bfregs_per_sys_page;
833 	u32 index_of_sys_page;
834 	u32 offset;
835 
836 	if (bfregi->lib_uar_dyn)
837 		return -EINVAL;
838 
839 	bfregs_per_sys_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
840 				MLX5_NON_FP_BFREGS_PER_UAR;
841 	index_of_sys_page = bfregn / bfregs_per_sys_page;
842 
843 	if (dyn_bfreg) {
844 		index_of_sys_page += bfregi->num_static_sys_pages;
845 
846 		if (index_of_sys_page >= bfregi->num_sys_pages)
847 			return -EINVAL;
848 
849 		if (bfregn > bfregi->num_dyn_bfregs ||
850 		    bfregi->sys_pages[index_of_sys_page] == MLX5_IB_INVALID_UAR_INDEX) {
851 			mlx5_ib_dbg(dev, "Invalid dynamic uar index\n");
852 			return -EINVAL;
853 		}
854 	}
855 
856 	offset = bfregn % bfregs_per_sys_page / MLX5_NON_FP_BFREGS_PER_UAR;
857 	return bfregi->sys_pages[index_of_sys_page] + offset;
858 }
859 
860 static void destroy_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
861 			    struct mlx5_ib_rwq *rwq, struct ib_udata *udata)
862 {
863 	struct mlx5_ib_ucontext *context =
864 		rdma_udata_to_drv_context(
865 			udata,
866 			struct mlx5_ib_ucontext,
867 			ibucontext);
868 
869 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_DELAY_DROP)
870 		atomic_dec(&dev->delay_drop.rqs_cnt);
871 
872 	mlx5_ib_db_unmap_user(context, &rwq->db);
873 	ib_umem_release(rwq->umem);
874 }
875 
876 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
877 			  struct ib_udata *udata, struct mlx5_ib_rwq *rwq,
878 			  struct mlx5_ib_create_wq *ucmd)
879 {
880 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
881 		udata, struct mlx5_ib_ucontext, ibucontext);
882 	unsigned long page_size = 0;
883 	u32 offset = 0;
884 	int err;
885 
886 	if (!ucmd->buf_addr)
887 		return -EINVAL;
888 
889 	rwq->umem = ib_umem_get(&dev->ib_dev, ucmd->buf_addr, rwq->buf_size, 0);
890 	if (IS_ERR(rwq->umem)) {
891 		mlx5_ib_dbg(dev, "umem_get failed\n");
892 		err = PTR_ERR(rwq->umem);
893 		return err;
894 	}
895 
896 	page_size = mlx5_umem_find_best_quantized_pgoff(
897 		rwq->umem, wq, log_wq_pg_sz, MLX5_ADAPTER_PAGE_SHIFT,
898 		page_offset, 64, &rwq->rq_page_offset);
899 	if (!page_size) {
900 		mlx5_ib_warn(dev, "bad offset\n");
901 		err = -EINVAL;
902 		goto err_umem;
903 	}
904 
905 	rwq->rq_num_pas = ib_umem_num_dma_blocks(rwq->umem, page_size);
906 	rwq->page_shift = order_base_2(page_size);
907 	rwq->log_page_size =  rwq->page_shift - MLX5_ADAPTER_PAGE_SHIFT;
908 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
909 
910 	mlx5_ib_dbg(
911 		dev,
912 		"addr 0x%llx, size %zd, npages %zu, page_size %ld, ncont %d, offset %d\n",
913 		(unsigned long long)ucmd->buf_addr, rwq->buf_size,
914 		ib_umem_num_pages(rwq->umem), page_size, rwq->rq_num_pas,
915 		offset);
916 
917 	err = mlx5_ib_db_map_user(ucontext, ucmd->db_addr, &rwq->db);
918 	if (err) {
919 		mlx5_ib_dbg(dev, "map failed\n");
920 		goto err_umem;
921 	}
922 
923 	return 0;
924 
925 err_umem:
926 	ib_umem_release(rwq->umem);
927 	return err;
928 }
929 
930 static int adjust_bfregn(struct mlx5_ib_dev *dev,
931 			 struct mlx5_bfreg_info *bfregi, int bfregn)
932 {
933 	return bfregn / MLX5_NON_FP_BFREGS_PER_UAR * MLX5_BFREGS_PER_UAR +
934 				bfregn % MLX5_NON_FP_BFREGS_PER_UAR;
935 }
936 
937 static int _create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
938 			   struct mlx5_ib_qp *qp, struct ib_udata *udata,
939 			   struct ib_qp_init_attr *attr, u32 **in,
940 			   struct mlx5_ib_create_qp_resp *resp, int *inlen,
941 			   struct mlx5_ib_qp_base *base,
942 			   struct mlx5_ib_create_qp *ucmd)
943 {
944 	struct mlx5_ib_ucontext *context;
945 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
946 	unsigned int page_offset_quantized = 0;
947 	unsigned long page_size = 0;
948 	int uar_index = 0;
949 	int bfregn;
950 	int ncont = 0;
951 	__be64 *pas;
952 	void *qpc;
953 	int err;
954 	u16 uid;
955 	u32 uar_flags;
956 
957 	context = rdma_udata_to_drv_context(udata, struct mlx5_ib_ucontext,
958 					    ibucontext);
959 	uar_flags = qp->flags_en &
960 		    (MLX5_QP_FLAG_UAR_PAGE_INDEX | MLX5_QP_FLAG_BFREG_INDEX);
961 	switch (uar_flags) {
962 	case MLX5_QP_FLAG_UAR_PAGE_INDEX:
963 		uar_index = ucmd->bfreg_index;
964 		bfregn = MLX5_IB_INVALID_BFREG;
965 		break;
966 	case MLX5_QP_FLAG_BFREG_INDEX:
967 		uar_index = bfregn_to_uar_index(dev, &context->bfregi,
968 						ucmd->bfreg_index, true);
969 		if (uar_index < 0)
970 			return uar_index;
971 		bfregn = MLX5_IB_INVALID_BFREG;
972 		break;
973 	case 0:
974 		if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
975 			return -EINVAL;
976 		bfregn = alloc_bfreg(dev, &context->bfregi);
977 		if (bfregn < 0)
978 			return bfregn;
979 		break;
980 	default:
981 		return -EINVAL;
982 	}
983 
984 	mlx5_ib_dbg(dev, "bfregn 0x%x, uar_index 0x%x\n", bfregn, uar_index);
985 	if (bfregn != MLX5_IB_INVALID_BFREG)
986 		uar_index = bfregn_to_uar_index(dev, &context->bfregi, bfregn,
987 						false);
988 
989 	qp->rq.offset = 0;
990 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
991 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
992 
993 	err = set_user_buf_size(dev, qp, ucmd, base, attr);
994 	if (err)
995 		goto err_bfreg;
996 
997 	if (ucmd->buf_addr && ubuffer->buf_size) {
998 		ubuffer->buf_addr = ucmd->buf_addr;
999 		ubuffer->umem = ib_umem_get(&dev->ib_dev, ubuffer->buf_addr,
1000 					    ubuffer->buf_size, 0);
1001 		if (IS_ERR(ubuffer->umem)) {
1002 			err = PTR_ERR(ubuffer->umem);
1003 			goto err_bfreg;
1004 		}
1005 		page_size = mlx5_umem_find_best_quantized_pgoff(
1006 			ubuffer->umem, qpc, log_page_size,
1007 			MLX5_ADAPTER_PAGE_SHIFT, page_offset, 64,
1008 			&page_offset_quantized);
1009 		if (!page_size) {
1010 			err = -EINVAL;
1011 			goto err_umem;
1012 		}
1013 		ncont = ib_umem_num_dma_blocks(ubuffer->umem, page_size);
1014 	} else {
1015 		ubuffer->umem = NULL;
1016 	}
1017 
1018 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
1019 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
1020 	*in = kvzalloc(*inlen, GFP_KERNEL);
1021 	if (!*in) {
1022 		err = -ENOMEM;
1023 		goto err_umem;
1024 	}
1025 
1026 	uid = (attr->qp_type != IB_QPT_XRC_INI) ? to_mpd(pd)->uid : 0;
1027 	MLX5_SET(create_qp_in, *in, uid, uid);
1028 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
1029 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
1030 	if (ubuffer->umem) {
1031 		mlx5_ib_populate_pas(ubuffer->umem, page_size, pas, 0);
1032 		MLX5_SET(qpc, qpc, log_page_size,
1033 			 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
1034 		MLX5_SET(qpc, qpc, page_offset, page_offset_quantized);
1035 	}
1036 	MLX5_SET(qpc, qpc, uar_page, uar_index);
1037 	if (bfregn != MLX5_IB_INVALID_BFREG)
1038 		resp->bfreg_index = adjust_bfregn(dev, &context->bfregi, bfregn);
1039 	else
1040 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
1041 	qp->bfregn = bfregn;
1042 
1043 	err = mlx5_ib_db_map_user(context, ucmd->db_addr, &qp->db);
1044 	if (err) {
1045 		mlx5_ib_dbg(dev, "map failed\n");
1046 		goto err_free;
1047 	}
1048 
1049 	return 0;
1050 
1051 err_free:
1052 	kvfree(*in);
1053 
1054 err_umem:
1055 	ib_umem_release(ubuffer->umem);
1056 
1057 err_bfreg:
1058 	if (bfregn != MLX5_IB_INVALID_BFREG)
1059 		mlx5_ib_free_bfreg(dev, &context->bfregi, bfregn);
1060 	return err;
1061 }
1062 
1063 static void destroy_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1064 		       struct mlx5_ib_qp_base *base, struct ib_udata *udata)
1065 {
1066 	struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context(
1067 		udata, struct mlx5_ib_ucontext, ibucontext);
1068 
1069 	if (udata) {
1070 		/* User QP */
1071 		mlx5_ib_db_unmap_user(context, &qp->db);
1072 		ib_umem_release(base->ubuffer.umem);
1073 
1074 		/*
1075 		 * Free only the BFREGs which are handled by the kernel.
1076 		 * BFREGs of UARs allocated dynamically are handled by user.
1077 		 */
1078 		if (qp->bfregn != MLX5_IB_INVALID_BFREG)
1079 			mlx5_ib_free_bfreg(dev, &context->bfregi, qp->bfregn);
1080 		return;
1081 	}
1082 
1083 	/* Kernel QP */
1084 	kvfree(qp->sq.wqe_head);
1085 	kvfree(qp->sq.w_list);
1086 	kvfree(qp->sq.wrid);
1087 	kvfree(qp->sq.wr_data);
1088 	kvfree(qp->rq.wrid);
1089 	if (qp->db.db)
1090 		mlx5_db_free(dev->mdev, &qp->db);
1091 	if (qp->buf.frags)
1092 		mlx5_frag_buf_free(dev->mdev, &qp->buf);
1093 }
1094 
1095 static int _create_kernel_qp(struct mlx5_ib_dev *dev,
1096 			     struct ib_qp_init_attr *init_attr,
1097 			     struct mlx5_ib_qp *qp, u32 **in, int *inlen,
1098 			     struct mlx5_ib_qp_base *base)
1099 {
1100 	int uar_index;
1101 	void *qpc;
1102 	int err;
1103 
1104 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
1105 		qp->bf.bfreg = &dev->fp_bfreg;
1106 	else if (qp->flags & MLX5_IB_QP_CREATE_WC_TEST)
1107 		qp->bf.bfreg = &dev->wc_bfreg;
1108 	else
1109 		qp->bf.bfreg = &dev->bfreg;
1110 
1111 	/* We need to divide by two since each register is comprised of
1112 	 * two buffers of identical size, namely odd and even
1113 	 */
1114 	qp->bf.buf_size = (1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size)) / 2;
1115 	uar_index = qp->bf.bfreg->index;
1116 
1117 	err = calc_sq_size(dev, init_attr, qp);
1118 	if (err < 0) {
1119 		mlx5_ib_dbg(dev, "err %d\n", err);
1120 		return err;
1121 	}
1122 
1123 	qp->rq.offset = 0;
1124 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
1125 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
1126 
1127 	err = mlx5_frag_buf_alloc_node(dev->mdev, base->ubuffer.buf_size,
1128 				       &qp->buf, dev->mdev->priv.numa_node);
1129 	if (err) {
1130 		mlx5_ib_dbg(dev, "err %d\n", err);
1131 		return err;
1132 	}
1133 
1134 	if (qp->rq.wqe_cnt)
1135 		mlx5_init_fbc(qp->buf.frags, qp->rq.wqe_shift,
1136 			      ilog2(qp->rq.wqe_cnt), &qp->rq.fbc);
1137 
1138 	if (qp->sq.wqe_cnt) {
1139 		int sq_strides_offset = (qp->sq.offset  & (PAGE_SIZE - 1)) /
1140 					MLX5_SEND_WQE_BB;
1141 		mlx5_init_fbc_offset(qp->buf.frags +
1142 				     (qp->sq.offset / PAGE_SIZE),
1143 				     ilog2(MLX5_SEND_WQE_BB),
1144 				     ilog2(qp->sq.wqe_cnt),
1145 				     sq_strides_offset, &qp->sq.fbc);
1146 
1147 		qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
1148 	}
1149 
1150 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
1151 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
1152 	*in = kvzalloc(*inlen, GFP_KERNEL);
1153 	if (!*in) {
1154 		err = -ENOMEM;
1155 		goto err_buf;
1156 	}
1157 
1158 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
1159 	MLX5_SET(qpc, qpc, uar_page, uar_index);
1160 	MLX5_SET(qpc, qpc, ts_format, mlx5_get_qp_default_ts(dev->mdev));
1161 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1162 
1163 	/* Set "fast registration enabled" for all kernel QPs */
1164 	MLX5_SET(qpc, qpc, fre, 1);
1165 	MLX5_SET(qpc, qpc, rlky, 1);
1166 
1167 	if (qp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
1168 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
1169 
1170 	mlx5_fill_page_frag_array(&qp->buf,
1171 				  (__be64 *)MLX5_ADDR_OF(create_qp_in,
1172 							 *in, pas));
1173 
1174 	err = mlx5_db_alloc(dev->mdev, &qp->db);
1175 	if (err) {
1176 		mlx5_ib_dbg(dev, "err %d\n", err);
1177 		goto err_free;
1178 	}
1179 
1180 	qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1181 				     sizeof(*qp->sq.wrid), GFP_KERNEL);
1182 	qp->sq.wr_data = kvmalloc_array(qp->sq.wqe_cnt,
1183 					sizeof(*qp->sq.wr_data), GFP_KERNEL);
1184 	qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1185 				     sizeof(*qp->rq.wrid), GFP_KERNEL);
1186 	qp->sq.w_list = kvmalloc_array(qp->sq.wqe_cnt,
1187 				       sizeof(*qp->sq.w_list), GFP_KERNEL);
1188 	qp->sq.wqe_head = kvmalloc_array(qp->sq.wqe_cnt,
1189 					 sizeof(*qp->sq.wqe_head), GFP_KERNEL);
1190 
1191 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
1192 	    !qp->sq.w_list || !qp->sq.wqe_head) {
1193 		err = -ENOMEM;
1194 		goto err_wrid;
1195 	}
1196 
1197 	return 0;
1198 
1199 err_wrid:
1200 	kvfree(qp->sq.wqe_head);
1201 	kvfree(qp->sq.w_list);
1202 	kvfree(qp->sq.wrid);
1203 	kvfree(qp->sq.wr_data);
1204 	kvfree(qp->rq.wrid);
1205 	mlx5_db_free(dev->mdev, &qp->db);
1206 
1207 err_free:
1208 	kvfree(*in);
1209 
1210 err_buf:
1211 	mlx5_frag_buf_free(dev->mdev, &qp->buf);
1212 	return err;
1213 }
1214 
1215 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
1216 {
1217 	if (attr->srq || (qp->type == IB_QPT_XRC_TGT) ||
1218 	    (qp->type == MLX5_IB_QPT_DCI) || (qp->type == IB_QPT_XRC_INI))
1219 		return MLX5_SRQ_RQ;
1220 	else if (!qp->has_rq)
1221 		return MLX5_ZERO_LEN_RQ;
1222 
1223 	return MLX5_NON_ZERO_RQ;
1224 }
1225 
1226 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1227 				    struct mlx5_ib_qp *qp,
1228 				    struct mlx5_ib_sq *sq, u32 tdn,
1229 				    struct ib_pd *pd)
1230 {
1231 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
1232 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1233 
1234 	MLX5_SET(create_tis_in, in, uid, to_mpd(pd)->uid);
1235 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1236 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
1237 		MLX5_SET(tisc, tisc, underlay_qpn, qp->underlay_qpn);
1238 
1239 	return mlx5_core_create_tis(dev->mdev, in, &sq->tisn);
1240 }
1241 
1242 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1243 				      struct mlx5_ib_sq *sq, struct ib_pd *pd)
1244 {
1245 	mlx5_cmd_destroy_tis(dev->mdev, sq->tisn, to_mpd(pd)->uid);
1246 }
1247 
1248 static void destroy_flow_rule_vport_sq(struct mlx5_ib_sq *sq)
1249 {
1250 	if (sq->flow_rule)
1251 		mlx5_del_flow_rules(sq->flow_rule);
1252 	sq->flow_rule = NULL;
1253 }
1254 
1255 static bool fr_supported(int ts_cap)
1256 {
1257 	return ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_FREE_RUNNING ||
1258 	       ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_FREE_RUNNING_AND_REAL_TIME;
1259 }
1260 
1261 static int get_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
1262 			 bool fr_sup, bool rt_sup)
1263 {
1264 	if (cq->private_flags & MLX5_IB_CQ_PR_FLAGS_REAL_TIME_TS) {
1265 		if (!rt_sup) {
1266 			mlx5_ib_dbg(dev,
1267 				    "Real time TS format is not supported\n");
1268 			return -EOPNOTSUPP;
1269 		}
1270 		return MLX5_TIMESTAMP_FORMAT_REAL_TIME;
1271 	}
1272 	if (cq->create_flags & IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION) {
1273 		if (!fr_sup) {
1274 			mlx5_ib_dbg(dev,
1275 				    "Free running TS format is not supported\n");
1276 			return -EOPNOTSUPP;
1277 		}
1278 		return MLX5_TIMESTAMP_FORMAT_FREE_RUNNING;
1279 	}
1280 	return fr_sup ? MLX5_TIMESTAMP_FORMAT_FREE_RUNNING :
1281 			MLX5_TIMESTAMP_FORMAT_DEFAULT;
1282 }
1283 
1284 static int get_rq_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *recv_cq)
1285 {
1286 	u8 ts_cap = MLX5_CAP_GEN(dev->mdev, rq_ts_format);
1287 
1288 	return get_ts_format(dev, recv_cq, fr_supported(ts_cap),
1289 			     rt_supported(ts_cap));
1290 }
1291 
1292 static int get_sq_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq)
1293 {
1294 	u8 ts_cap = MLX5_CAP_GEN(dev->mdev, sq_ts_format);
1295 
1296 	return get_ts_format(dev, send_cq, fr_supported(ts_cap),
1297 			     rt_supported(ts_cap));
1298 }
1299 
1300 static int get_qp_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq,
1301 			    struct mlx5_ib_cq *recv_cq)
1302 {
1303 	u8 ts_cap = MLX5_CAP_ROCE(dev->mdev, qp_ts_format);
1304 	bool fr_sup = fr_supported(ts_cap);
1305 	bool rt_sup = rt_supported(ts_cap);
1306 	u8 default_ts = fr_sup ? MLX5_TIMESTAMP_FORMAT_FREE_RUNNING :
1307 				 MLX5_TIMESTAMP_FORMAT_DEFAULT;
1308 	int send_ts_format =
1309 		send_cq ? get_ts_format(dev, send_cq, fr_sup, rt_sup) :
1310 			  default_ts;
1311 	int recv_ts_format =
1312 		recv_cq ? get_ts_format(dev, recv_cq, fr_sup, rt_sup) :
1313 			  default_ts;
1314 
1315 	if (send_ts_format < 0 || recv_ts_format < 0)
1316 		return -EOPNOTSUPP;
1317 
1318 	if (send_ts_format != MLX5_TIMESTAMP_FORMAT_DEFAULT &&
1319 	    recv_ts_format != MLX5_TIMESTAMP_FORMAT_DEFAULT &&
1320 	    send_ts_format != recv_ts_format) {
1321 		mlx5_ib_dbg(
1322 			dev,
1323 			"The send ts_format does not match the receive ts_format\n");
1324 		return -EOPNOTSUPP;
1325 	}
1326 
1327 	return send_ts_format == default_ts ? recv_ts_format : send_ts_format;
1328 }
1329 
1330 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1331 				   struct ib_udata *udata,
1332 				   struct mlx5_ib_sq *sq, void *qpin,
1333 				   struct ib_pd *pd, struct mlx5_ib_cq *cq)
1334 {
1335 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1336 	__be64 *pas;
1337 	void *in;
1338 	void *sqc;
1339 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1340 	void *wq;
1341 	int inlen;
1342 	int err;
1343 	unsigned int page_offset_quantized;
1344 	unsigned long page_size;
1345 	int ts_format;
1346 
1347 	ts_format = get_sq_ts_format(dev, cq);
1348 	if (ts_format < 0)
1349 		return ts_format;
1350 
1351 	sq->ubuffer.umem = ib_umem_get(&dev->ib_dev, ubuffer->buf_addr,
1352 				       ubuffer->buf_size, 0);
1353 	if (IS_ERR(sq->ubuffer.umem))
1354 		return PTR_ERR(sq->ubuffer.umem);
1355 	page_size = mlx5_umem_find_best_quantized_pgoff(
1356 		ubuffer->umem, wq, log_wq_pg_sz, MLX5_ADAPTER_PAGE_SHIFT,
1357 		page_offset, 64, &page_offset_quantized);
1358 	if (!page_size) {
1359 		err = -EINVAL;
1360 		goto err_umem;
1361 	}
1362 
1363 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
1364 		sizeof(u64) *
1365 			ib_umem_num_dma_blocks(sq->ubuffer.umem, page_size);
1366 	in = kvzalloc(inlen, GFP_KERNEL);
1367 	if (!in) {
1368 		err = -ENOMEM;
1369 		goto err_umem;
1370 	}
1371 
1372 	MLX5_SET(create_sq_in, in, uid, to_mpd(pd)->uid);
1373 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1374 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1375 	if (MLX5_CAP_ETH(dev->mdev, multi_pkt_send_wqe))
1376 		MLX5_SET(sqc, sqc, allow_multi_pkt_send_wqe, 1);
1377 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1378 	MLX5_SET(sqc, sqc, ts_format, ts_format);
1379 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1380 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1381 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1382 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1383 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
1384 	    MLX5_CAP_ETH(dev->mdev, swp))
1385 		MLX5_SET(sqc, sqc, allow_swp, 1);
1386 
1387 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1388 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1389 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1390 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1391 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1392 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1393 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1394 	MLX5_SET(wq, wq, log_wq_pg_sz,
1395 		 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
1396 	MLX5_SET(wq, wq, page_offset, page_offset_quantized);
1397 
1398 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1399 	mlx5_ib_populate_pas(sq->ubuffer.umem, page_size, pas, 0);
1400 
1401 	err = mlx5_core_create_sq_tracked(dev, in, inlen, &sq->base.mqp);
1402 
1403 	kvfree(in);
1404 
1405 	if (err)
1406 		goto err_umem;
1407 
1408 	return 0;
1409 
1410 err_umem:
1411 	ib_umem_release(sq->ubuffer.umem);
1412 	sq->ubuffer.umem = NULL;
1413 
1414 	return err;
1415 }
1416 
1417 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1418 				     struct mlx5_ib_sq *sq)
1419 {
1420 	destroy_flow_rule_vport_sq(sq);
1421 	mlx5_core_destroy_sq_tracked(dev, &sq->base.mqp);
1422 	ib_umem_release(sq->ubuffer.umem);
1423 }
1424 
1425 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1426 				   struct mlx5_ib_rq *rq, void *qpin,
1427 				   struct ib_pd *pd, struct mlx5_ib_cq *cq)
1428 {
1429 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1430 	__be64 *pas;
1431 	void *in;
1432 	void *rqc;
1433 	void *wq;
1434 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1435 	struct ib_umem *umem = rq->base.ubuffer.umem;
1436 	unsigned int page_offset_quantized;
1437 	unsigned long page_size = 0;
1438 	int ts_format;
1439 	size_t inlen;
1440 	int err;
1441 
1442 	ts_format = get_rq_ts_format(dev, cq);
1443 	if (ts_format < 0)
1444 		return ts_format;
1445 
1446 	page_size = mlx5_umem_find_best_quantized_pgoff(umem, wq, log_wq_pg_sz,
1447 							MLX5_ADAPTER_PAGE_SHIFT,
1448 							page_offset, 64,
1449 							&page_offset_quantized);
1450 	if (!page_size)
1451 		return -EINVAL;
1452 
1453 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
1454 		sizeof(u64) * ib_umem_num_dma_blocks(umem, page_size);
1455 	in = kvzalloc(inlen, GFP_KERNEL);
1456 	if (!in)
1457 		return -ENOMEM;
1458 
1459 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
1460 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1461 	if (!(rq->flags & MLX5_IB_RQ_CVLAN_STRIPPING))
1462 		MLX5_SET(rqc, rqc, vsd, 1);
1463 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1464 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1465 	MLX5_SET(rqc, rqc, ts_format, ts_format);
1466 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1467 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1468 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1469 
1470 	if (mqp->flags & IB_QP_CREATE_SCATTER_FCS)
1471 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1472 
1473 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1474 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1475 	if (rq->flags & MLX5_IB_RQ_PCI_WRITE_END_PADDING)
1476 		MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1477 	MLX5_SET(wq, wq, page_offset, page_offset_quantized);
1478 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1479 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1480 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1481 	MLX5_SET(wq, wq, log_wq_pg_sz,
1482 		 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT);
1483 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1484 
1485 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1486 	mlx5_ib_populate_pas(umem, page_size, pas, 0);
1487 
1488 	err = mlx5_core_create_rq_tracked(dev, in, inlen, &rq->base.mqp);
1489 
1490 	kvfree(in);
1491 
1492 	return err;
1493 }
1494 
1495 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1496 				     struct mlx5_ib_rq *rq)
1497 {
1498 	mlx5_core_destroy_rq_tracked(dev, &rq->base.mqp);
1499 }
1500 
1501 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1502 				      struct mlx5_ib_rq *rq,
1503 				      u32 qp_flags_en,
1504 				      struct ib_pd *pd)
1505 {
1506 	if (qp_flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1507 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1508 		mlx5_ib_disable_lb(dev, false, true);
1509 	mlx5_cmd_destroy_tir(dev->mdev, rq->tirn, to_mpd(pd)->uid);
1510 }
1511 
1512 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1513 				    struct mlx5_ib_rq *rq, u32 tdn,
1514 				    u32 *qp_flags_en, struct ib_pd *pd,
1515 				    u32 *out)
1516 {
1517 	u8 lb_flag = 0;
1518 	u32 *in;
1519 	void *tirc;
1520 	int inlen;
1521 	int err;
1522 
1523 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1524 	in = kvzalloc(inlen, GFP_KERNEL);
1525 	if (!in)
1526 		return -ENOMEM;
1527 
1528 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1529 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1530 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1531 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1532 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1533 	if (*qp_flags_en & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1534 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1535 
1536 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1537 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1538 
1539 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1540 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1541 
1542 	if (dev->is_rep) {
1543 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1544 		*qp_flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1545 	}
1546 
1547 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1548 	MLX5_SET(create_tir_in, in, opcode, MLX5_CMD_OP_CREATE_TIR);
1549 	err = mlx5_cmd_exec_inout(dev->mdev, create_tir, in, out);
1550 	rq->tirn = MLX5_GET(create_tir_out, out, tirn);
1551 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1552 		err = mlx5_ib_enable_lb(dev, false, true);
1553 
1554 		if (err)
1555 			destroy_raw_packet_qp_tir(dev, rq, 0, pd);
1556 	}
1557 	kvfree(in);
1558 
1559 	return err;
1560 }
1561 
1562 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1563 				u32 *in, size_t inlen, struct ib_pd *pd,
1564 				struct ib_udata *udata,
1565 				struct mlx5_ib_create_qp_resp *resp,
1566 				struct ib_qp_init_attr *init_attr)
1567 {
1568 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1569 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1570 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1571 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1572 		udata, struct mlx5_ib_ucontext, ibucontext);
1573 	int err;
1574 	u32 tdn = mucontext->tdn;
1575 	u16 uid = to_mpd(pd)->uid;
1576 	u32 out[MLX5_ST_SZ_DW(create_tir_out)] = {};
1577 
1578 	if (!qp->sq.wqe_cnt && !qp->rq.wqe_cnt)
1579 		return -EINVAL;
1580 	if (qp->sq.wqe_cnt) {
1581 		err = create_raw_packet_qp_tis(dev, qp, sq, tdn, pd);
1582 		if (err)
1583 			return err;
1584 
1585 		err = create_raw_packet_qp_sq(dev, udata, sq, in, pd,
1586 					      to_mcq(init_attr->send_cq));
1587 		if (err)
1588 			goto err_destroy_tis;
1589 
1590 		if (uid) {
1591 			resp->tisn = sq->tisn;
1592 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TISN;
1593 			resp->sqn = sq->base.mqp.qpn;
1594 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_SQN;
1595 		}
1596 
1597 		sq->base.container_mibqp = qp;
1598 		sq->base.mqp.event = mlx5_ib_qp_event;
1599 	}
1600 
1601 	if (qp->rq.wqe_cnt) {
1602 		rq->base.container_mibqp = qp;
1603 
1604 		if (qp->flags & IB_QP_CREATE_CVLAN_STRIPPING)
1605 			rq->flags |= MLX5_IB_RQ_CVLAN_STRIPPING;
1606 		if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING)
1607 			rq->flags |= MLX5_IB_RQ_PCI_WRITE_END_PADDING;
1608 		err = create_raw_packet_qp_rq(dev, rq, in, pd,
1609 					      to_mcq(init_attr->recv_cq));
1610 		if (err)
1611 			goto err_destroy_sq;
1612 
1613 		err = create_raw_packet_qp_tir(dev, rq, tdn, &qp->flags_en, pd,
1614 					       out);
1615 		if (err)
1616 			goto err_destroy_rq;
1617 
1618 		if (uid) {
1619 			resp->rqn = rq->base.mqp.qpn;
1620 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_RQN;
1621 			resp->tirn = rq->tirn;
1622 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1623 			if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
1624 			    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2)) {
1625 				resp->tir_icm_addr = MLX5_GET(
1626 					create_tir_out, out, icm_address_31_0);
1627 				resp->tir_icm_addr |=
1628 					(u64)MLX5_GET(create_tir_out, out,
1629 						      icm_address_39_32)
1630 					<< 32;
1631 				resp->tir_icm_addr |=
1632 					(u64)MLX5_GET(create_tir_out, out,
1633 						      icm_address_63_40)
1634 					<< 40;
1635 				resp->comp_mask |=
1636 					MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1637 			}
1638 		}
1639 	}
1640 
1641 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1642 						     rq->base.mqp.qpn;
1643 	return 0;
1644 
1645 err_destroy_rq:
1646 	destroy_raw_packet_qp_rq(dev, rq);
1647 err_destroy_sq:
1648 	if (!qp->sq.wqe_cnt)
1649 		return err;
1650 	destroy_raw_packet_qp_sq(dev, sq);
1651 err_destroy_tis:
1652 	destroy_raw_packet_qp_tis(dev, sq, pd);
1653 
1654 	return err;
1655 }
1656 
1657 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1658 				  struct mlx5_ib_qp *qp)
1659 {
1660 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1661 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1662 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1663 
1664 	if (qp->rq.wqe_cnt) {
1665 		destroy_raw_packet_qp_tir(dev, rq, qp->flags_en, qp->ibqp.pd);
1666 		destroy_raw_packet_qp_rq(dev, rq);
1667 	}
1668 
1669 	if (qp->sq.wqe_cnt) {
1670 		destroy_raw_packet_qp_sq(dev, sq);
1671 		destroy_raw_packet_qp_tis(dev, sq, qp->ibqp.pd);
1672 	}
1673 }
1674 
1675 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1676 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1677 {
1678 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1679 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1680 
1681 	sq->sq = &qp->sq;
1682 	rq->rq = &qp->rq;
1683 	sq->doorbell = &qp->db;
1684 	rq->doorbell = &qp->db;
1685 }
1686 
1687 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1688 {
1689 	if (qp->flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1690 			    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1691 		mlx5_ib_disable_lb(dev, false, true);
1692 	mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1693 			     to_mpd(qp->ibqp.pd)->uid);
1694 }
1695 
1696 struct mlx5_create_qp_params {
1697 	struct ib_udata *udata;
1698 	size_t inlen;
1699 	size_t outlen;
1700 	size_t ucmd_size;
1701 	void *ucmd;
1702 	u8 is_rss_raw : 1;
1703 	struct ib_qp_init_attr *attr;
1704 	u32 uidx;
1705 	struct mlx5_ib_create_qp_resp resp;
1706 };
1707 
1708 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1709 				 struct mlx5_ib_qp *qp,
1710 				 struct mlx5_create_qp_params *params)
1711 {
1712 	struct ib_qp_init_attr *init_attr = params->attr;
1713 	struct mlx5_ib_create_qp_rss *ucmd = params->ucmd;
1714 	struct ib_udata *udata = params->udata;
1715 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1716 		udata, struct mlx5_ib_ucontext, ibucontext);
1717 	int inlen;
1718 	int outlen;
1719 	int err;
1720 	u32 *in;
1721 	u32 *out;
1722 	void *tirc;
1723 	void *hfso;
1724 	u32 selected_fields = 0;
1725 	u32 outer_l4;
1726 	u32 tdn = mucontext->tdn;
1727 	u8 lb_flag = 0;
1728 
1729 	if (ucmd->comp_mask) {
1730 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1731 		return -EOPNOTSUPP;
1732 	}
1733 
1734 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_INNER &&
1735 	    !(ucmd->flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)) {
1736 		mlx5_ib_dbg(dev, "Tunnel offloads must be set for inner RSS\n");
1737 		return -EOPNOTSUPP;
1738 	}
1739 
1740 	if (dev->is_rep)
1741 		qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1742 
1743 	if (qp->flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1744 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1745 
1746 	if (qp->flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1747 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1748 
1749 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1750 	outlen = MLX5_ST_SZ_BYTES(create_tir_out);
1751 	in = kvzalloc(inlen + outlen, GFP_KERNEL);
1752 	if (!in)
1753 		return -ENOMEM;
1754 
1755 	out = in + MLX5_ST_SZ_DW(create_tir_in);
1756 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1757 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1758 	MLX5_SET(tirc, tirc, disp_type,
1759 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1760 	MLX5_SET(tirc, tirc, indirect_table,
1761 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1762 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1763 
1764 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1765 
1766 	if (ucmd->flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1767 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1768 
1769 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1770 
1771 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_INNER)
1772 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner);
1773 	else
1774 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1775 
1776 	switch (ucmd->rx_hash_function) {
1777 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1778 	{
1779 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1780 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1781 
1782 		if (len != ucmd->rx_key_len) {
1783 			err = -EINVAL;
1784 			goto err;
1785 		}
1786 
1787 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1788 		memcpy(rss_key, ucmd->rx_hash_key, len);
1789 		break;
1790 	}
1791 	default:
1792 		err = -EOPNOTSUPP;
1793 		goto err;
1794 	}
1795 
1796 	if (!ucmd->rx_hash_fields_mask) {
1797 		/* special case when this TIR serves as steering entry without hashing */
1798 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1799 			goto create_tir;
1800 		err = -EINVAL;
1801 		goto err;
1802 	}
1803 
1804 	if (((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1805 	     (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1806 	     ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1807 	     (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1808 		err = -EINVAL;
1809 		goto err;
1810 	}
1811 
1812 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1813 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1814 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1815 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1816 			 MLX5_L3_PROT_TYPE_IPV4);
1817 	else if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1818 		 (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1819 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1820 			 MLX5_L3_PROT_TYPE_IPV6);
1821 
1822 	outer_l4 = ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1823 		    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1824 			   << 0 |
1825 		   ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1826 		    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1827 			   << 1 |
1828 		   (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI) << 2;
1829 
1830 	/* Check that only one l4 protocol is set */
1831 	if (outer_l4 & (outer_l4 - 1)) {
1832 		err = -EINVAL;
1833 		goto err;
1834 	}
1835 
1836 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1837 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1838 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1839 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1840 			 MLX5_L4_PROT_TYPE_TCP);
1841 	else if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1842 		 (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1843 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1844 			 MLX5_L4_PROT_TYPE_UDP);
1845 
1846 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1847 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1848 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1849 
1850 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1851 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1852 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1853 
1854 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1855 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1856 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1857 
1858 	if ((ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1859 	    (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1860 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1861 
1862 	if (ucmd->rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI)
1863 		selected_fields |= MLX5_HASH_FIELD_SEL_IPSEC_SPI;
1864 
1865 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1866 
1867 create_tir:
1868 	MLX5_SET(create_tir_in, in, opcode, MLX5_CMD_OP_CREATE_TIR);
1869 	err = mlx5_cmd_exec_inout(dev->mdev, create_tir, in, out);
1870 
1871 	qp->rss_qp.tirn = MLX5_GET(create_tir_out, out, tirn);
1872 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1873 		err = mlx5_ib_enable_lb(dev, false, true);
1874 
1875 		if (err)
1876 			mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1877 					     to_mpd(pd)->uid);
1878 	}
1879 
1880 	if (err)
1881 		goto err;
1882 
1883 	if (mucontext->devx_uid) {
1884 		params->resp.comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1885 		params->resp.tirn = qp->rss_qp.tirn;
1886 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner) ||
1887 		    MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner_v2)) {
1888 			params->resp.tir_icm_addr =
1889 				MLX5_GET(create_tir_out, out, icm_address_31_0);
1890 			params->resp.tir_icm_addr |=
1891 				(u64)MLX5_GET(create_tir_out, out,
1892 					      icm_address_39_32)
1893 				<< 32;
1894 			params->resp.tir_icm_addr |=
1895 				(u64)MLX5_GET(create_tir_out, out,
1896 					      icm_address_63_40)
1897 				<< 40;
1898 			params->resp.comp_mask |=
1899 				MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1900 		}
1901 	}
1902 
1903 	kvfree(in);
1904 	/* qpn is reserved for that QP */
1905 	qp->trans_qp.base.mqp.qpn = 0;
1906 	qp->is_rss = true;
1907 	return 0;
1908 
1909 err:
1910 	kvfree(in);
1911 	return err;
1912 }
1913 
1914 static void configure_requester_scat_cqe(struct mlx5_ib_dev *dev,
1915 					 struct mlx5_ib_qp *qp,
1916 					 struct ib_qp_init_attr *init_attr,
1917 					 void *qpc)
1918 {
1919 	int scqe_sz;
1920 	bool allow_scat_cqe = false;
1921 
1922 	allow_scat_cqe = qp->flags_en & MLX5_QP_FLAG_ALLOW_SCATTER_CQE;
1923 
1924 	if (!allow_scat_cqe && init_attr->sq_sig_type != IB_SIGNAL_ALL_WR)
1925 		return;
1926 
1927 	scqe_sz = mlx5_ib_get_cqe_size(init_attr->send_cq);
1928 	if (scqe_sz == 128) {
1929 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1930 		return;
1931 	}
1932 
1933 	if (init_attr->qp_type != MLX5_IB_QPT_DCI ||
1934 	    MLX5_CAP_GEN(dev->mdev, dc_req_scat_data_cqe))
1935 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1936 }
1937 
1938 static int atomic_size_to_mode(int size_mask)
1939 {
1940 	/* driver does not support atomic_size > 256B
1941 	 * and does not know how to translate bigger sizes
1942 	 */
1943 	int supported_size_mask = size_mask & 0x1ff;
1944 	int log_max_size;
1945 
1946 	if (!supported_size_mask)
1947 		return -EOPNOTSUPP;
1948 
1949 	log_max_size = __fls(supported_size_mask);
1950 
1951 	if (log_max_size > 3)
1952 		return log_max_size;
1953 
1954 	return MLX5_ATOMIC_MODE_8B;
1955 }
1956 
1957 static int get_atomic_mode(struct mlx5_ib_dev *dev,
1958 			   enum ib_qp_type qp_type)
1959 {
1960 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
1961 	u8 atomic = MLX5_CAP_GEN(dev->mdev, atomic);
1962 	int atomic_mode = -EOPNOTSUPP;
1963 	int atomic_size_mask;
1964 
1965 	if (!atomic)
1966 		return -EOPNOTSUPP;
1967 
1968 	if (qp_type == MLX5_IB_QPT_DCT)
1969 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
1970 	else
1971 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
1972 
1973 	if ((atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_CMP_SWAP) ||
1974 	    (atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_FETCH_ADD))
1975 		atomic_mode = atomic_size_to_mode(atomic_size_mask);
1976 
1977 	if (atomic_mode <= 0 &&
1978 	    (atomic_operations & MLX5_ATOMIC_OPS_CMP_SWAP &&
1979 	     atomic_operations & MLX5_ATOMIC_OPS_FETCH_ADD))
1980 		atomic_mode = MLX5_ATOMIC_MODE_IB_COMP;
1981 
1982 	return atomic_mode;
1983 }
1984 
1985 static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1986 			     struct mlx5_create_qp_params *params)
1987 {
1988 	struct ib_qp_init_attr *attr = params->attr;
1989 	u32 uidx = params->uidx;
1990 	struct mlx5_ib_resources *devr = &dev->devr;
1991 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
1992 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1993 	struct mlx5_core_dev *mdev = dev->mdev;
1994 	struct mlx5_ib_qp_base *base;
1995 	unsigned long flags;
1996 	void *qpc;
1997 	u32 *in;
1998 	int err;
1999 
2000 	if (attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2001 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2002 
2003 	in = kvzalloc(inlen, GFP_KERNEL);
2004 	if (!in)
2005 		return -ENOMEM;
2006 
2007 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2008 
2009 	MLX5_SET(qpc, qpc, st, MLX5_QP_ST_XRC);
2010 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2011 	MLX5_SET(qpc, qpc, pd, to_mpd(devr->p0)->pdn);
2012 
2013 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
2014 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2015 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
2016 		MLX5_SET(qpc, qpc, cd_master, 1);
2017 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
2018 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
2019 	if (qp->flags & IB_QP_CREATE_MANAGED_RECV)
2020 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
2021 
2022 	MLX5_SET(qpc, qpc, ts_format, mlx5_get_qp_default_ts(dev->mdev));
2023 	MLX5_SET(qpc, qpc, rq_type, MLX5_SRQ_RQ);
2024 	MLX5_SET(qpc, qpc, no_sq, 1);
2025 	MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2026 	MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
2027 	MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2028 	MLX5_SET(qpc, qpc, xrcd, to_mxrcd(attr->xrcd)->xrcdn);
2029 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2030 
2031 	/* 0xffffff means we ask to work with cqe version 0 */
2032 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2033 		MLX5_SET(qpc, qpc, user_index, uidx);
2034 
2035 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING) {
2036 		MLX5_SET(qpc, qpc, end_padding_mode,
2037 			 MLX5_WQ_END_PAD_MODE_ALIGN);
2038 		/* Special case to clean flag */
2039 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
2040 	}
2041 
2042 	base = &qp->trans_qp.base;
2043 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2044 	kvfree(in);
2045 	if (err)
2046 		return err;
2047 
2048 	base->container_mibqp = qp;
2049 	base->mqp.event = mlx5_ib_qp_event;
2050 	if (MLX5_CAP_GEN(mdev, ece_support))
2051 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
2052 
2053 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2054 	list_add_tail(&qp->qps_list, &dev->qp_list);
2055 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2056 
2057 	qp->trans_qp.xrcdn = to_mxrcd(attr->xrcd)->xrcdn;
2058 	return 0;
2059 }
2060 
2061 static int create_dci(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2062 		      struct mlx5_ib_qp *qp,
2063 		      struct mlx5_create_qp_params *params)
2064 {
2065 	struct ib_qp_init_attr *init_attr = params->attr;
2066 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
2067 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
2068 	struct ib_udata *udata = params->udata;
2069 	u32 uidx = params->uidx;
2070 	struct mlx5_ib_resources *devr = &dev->devr;
2071 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
2072 	struct mlx5_core_dev *mdev = dev->mdev;
2073 	struct mlx5_ib_cq *send_cq;
2074 	struct mlx5_ib_cq *recv_cq;
2075 	unsigned long flags;
2076 	struct mlx5_ib_qp_base *base;
2077 	int ts_format;
2078 	int mlx5_st;
2079 	void *qpc;
2080 	u32 *in;
2081 	int err;
2082 
2083 	spin_lock_init(&qp->sq.lock);
2084 	spin_lock_init(&qp->rq.lock);
2085 
2086 	mlx5_st = to_mlx5_st(qp->type);
2087 	if (mlx5_st < 0)
2088 		return -EINVAL;
2089 
2090 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2091 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2092 
2093 	base = &qp->trans_qp.base;
2094 
2095 	qp->has_rq = qp_has_rq(init_attr);
2096 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq, qp, ucmd);
2097 	if (err) {
2098 		mlx5_ib_dbg(dev, "err %d\n", err);
2099 		return err;
2100 	}
2101 
2102 	if (ucmd->rq_wqe_shift != qp->rq.wqe_shift ||
2103 	    ucmd->rq_wqe_count != qp->rq.wqe_cnt)
2104 		return -EINVAL;
2105 
2106 	if (ucmd->sq_wqe_count > (1 << MLX5_CAP_GEN(mdev, log_max_qp_sz)))
2107 		return -EINVAL;
2108 
2109 	ts_format = get_qp_ts_format(dev, to_mcq(init_attr->send_cq),
2110 				     to_mcq(init_attr->recv_cq));
2111 
2112 	if (ts_format < 0)
2113 		return ts_format;
2114 
2115 	err = _create_user_qp(dev, pd, qp, udata, init_attr, &in, &params->resp,
2116 			      &inlen, base, ucmd);
2117 	if (err)
2118 		return err;
2119 
2120 	if (MLX5_CAP_GEN(mdev, ece_support))
2121 		MLX5_SET(create_qp_in, in, ece, ucmd->ece_options);
2122 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2123 
2124 	MLX5_SET(qpc, qpc, st, mlx5_st);
2125 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2126 	MLX5_SET(qpc, qpc, pd, to_mpd(pd)->pdn);
2127 
2128 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE)
2129 		MLX5_SET(qpc, qpc, wq_signature, 1);
2130 
2131 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
2132 		MLX5_SET(qpc, qpc, cd_master, 1);
2133 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
2134 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
2135 	if (qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE)
2136 		configure_requester_scat_cqe(dev, qp, init_attr, qpc);
2137 
2138 	if (qp->rq.wqe_cnt) {
2139 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2140 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2141 	}
2142 
2143 	if (qp->flags_en & MLX5_QP_FLAG_DCI_STREAM) {
2144 		MLX5_SET(qpc, qpc, log_num_dci_stream_channels,
2145 			 ucmd->dci_streams.log_num_concurent);
2146 		MLX5_SET(qpc, qpc, log_num_dci_errored_streams,
2147 			 ucmd->dci_streams.log_num_errored);
2148 	}
2149 
2150 	MLX5_SET(qpc, qpc, ts_format, ts_format);
2151 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
2152 
2153 	MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2154 
2155 	/* Set default resources */
2156 	if (init_attr->srq) {
2157 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2158 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2159 			 to_msrq(init_attr->srq)->msrq.srqn);
2160 	} else {
2161 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2162 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2163 			 to_msrq(devr->s1)->msrq.srqn);
2164 	}
2165 
2166 	if (init_attr->send_cq)
2167 		MLX5_SET(qpc, qpc, cqn_snd,
2168 			 to_mcq(init_attr->send_cq)->mcq.cqn);
2169 
2170 	if (init_attr->recv_cq)
2171 		MLX5_SET(qpc, qpc, cqn_rcv,
2172 			 to_mcq(init_attr->recv_cq)->mcq.cqn);
2173 
2174 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2175 
2176 	/* 0xffffff means we ask to work with cqe version 0 */
2177 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2178 		MLX5_SET(qpc, qpc, user_index, uidx);
2179 
2180 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING) {
2181 		MLX5_SET(qpc, qpc, end_padding_mode,
2182 			 MLX5_WQ_END_PAD_MODE_ALIGN);
2183 		/* Special case to clean flag */
2184 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
2185 	}
2186 
2187 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2188 
2189 	kvfree(in);
2190 	if (err)
2191 		goto err_create;
2192 
2193 	base->container_mibqp = qp;
2194 	base->mqp.event = mlx5_ib_qp_event;
2195 	if (MLX5_CAP_GEN(mdev, ece_support))
2196 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
2197 
2198 	get_cqs(qp->type, init_attr->send_cq, init_attr->recv_cq,
2199 		&send_cq, &recv_cq);
2200 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2201 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2202 	/* Maintain device to QPs access, needed for further handling via reset
2203 	 * flow
2204 	 */
2205 	list_add_tail(&qp->qps_list, &dev->qp_list);
2206 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2207 	 */
2208 	if (send_cq)
2209 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2210 	if (recv_cq)
2211 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2212 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2213 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2214 
2215 	return 0;
2216 
2217 err_create:
2218 	destroy_qp(dev, qp, base, udata);
2219 	return err;
2220 }
2221 
2222 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2223 			  struct mlx5_ib_qp *qp,
2224 			  struct mlx5_create_qp_params *params)
2225 {
2226 	struct ib_qp_init_attr *init_attr = params->attr;
2227 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
2228 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
2229 	struct ib_udata *udata = params->udata;
2230 	u32 uidx = params->uidx;
2231 	struct mlx5_ib_resources *devr = &dev->devr;
2232 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
2233 	struct mlx5_core_dev *mdev = dev->mdev;
2234 	struct mlx5_ib_cq *send_cq;
2235 	struct mlx5_ib_cq *recv_cq;
2236 	unsigned long flags;
2237 	struct mlx5_ib_qp_base *base;
2238 	int ts_format;
2239 	int mlx5_st;
2240 	void *qpc;
2241 	u32 *in;
2242 	int err;
2243 
2244 	spin_lock_init(&qp->sq.lock);
2245 	spin_lock_init(&qp->rq.lock);
2246 
2247 	mlx5_st = to_mlx5_st(qp->type);
2248 	if (mlx5_st < 0)
2249 		return -EINVAL;
2250 
2251 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2252 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2253 
2254 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
2255 		qp->underlay_qpn = init_attr->source_qpn;
2256 
2257 	base = (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2258 		qp->flags & IB_QP_CREATE_SOURCE_QPN) ?
2259 	       &qp->raw_packet_qp.rq.base :
2260 	       &qp->trans_qp.base;
2261 
2262 	qp->has_rq = qp_has_rq(init_attr);
2263 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq, qp, ucmd);
2264 	if (err) {
2265 		mlx5_ib_dbg(dev, "err %d\n", err);
2266 		return err;
2267 	}
2268 
2269 	if (ucmd->rq_wqe_shift != qp->rq.wqe_shift ||
2270 	    ucmd->rq_wqe_count != qp->rq.wqe_cnt)
2271 		return -EINVAL;
2272 
2273 	if (ucmd->sq_wqe_count > (1 << MLX5_CAP_GEN(mdev, log_max_qp_sz)))
2274 		return -EINVAL;
2275 
2276 	if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2277 		ts_format = get_qp_ts_format(dev, to_mcq(init_attr->send_cq),
2278 					     to_mcq(init_attr->recv_cq));
2279 		if (ts_format < 0)
2280 			return ts_format;
2281 	}
2282 
2283 	err = _create_user_qp(dev, pd, qp, udata, init_attr, &in, &params->resp,
2284 			      &inlen, base, ucmd);
2285 	if (err)
2286 		return err;
2287 
2288 	if (is_sqp(init_attr->qp_type))
2289 		qp->port = init_attr->port_num;
2290 
2291 	if (MLX5_CAP_GEN(mdev, ece_support))
2292 		MLX5_SET(create_qp_in, in, ece, ucmd->ece_options);
2293 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2294 
2295 	MLX5_SET(qpc, qpc, st, mlx5_st);
2296 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2297 	MLX5_SET(qpc, qpc, pd, to_mpd(pd)->pdn);
2298 
2299 	if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE)
2300 		MLX5_SET(qpc, qpc, wq_signature, 1);
2301 
2302 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
2303 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2304 
2305 	if (qp->flags & IB_QP_CREATE_CROSS_CHANNEL)
2306 		MLX5_SET(qpc, qpc, cd_master, 1);
2307 	if (qp->flags & IB_QP_CREATE_MANAGED_SEND)
2308 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
2309 	if (qp->flags & IB_QP_CREATE_MANAGED_RECV)
2310 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
2311 	if (qp->flags_en & MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE)
2312 		MLX5_SET(qpc, qpc, req_e2e_credit_mode, 1);
2313 	if ((qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) &&
2314 	    (init_attr->qp_type == IB_QPT_RC ||
2315 	     init_attr->qp_type == IB_QPT_UC)) {
2316 		int rcqe_sz = mlx5_ib_get_cqe_size(init_attr->recv_cq);
2317 
2318 		MLX5_SET(qpc, qpc, cs_res,
2319 			 rcqe_sz == 128 ? MLX5_RES_SCAT_DATA64_CQE :
2320 					  MLX5_RES_SCAT_DATA32_CQE);
2321 	}
2322 	if ((qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) &&
2323 	    (qp->type == MLX5_IB_QPT_DCI || qp->type == IB_QPT_RC))
2324 		configure_requester_scat_cqe(dev, qp, init_attr, qpc);
2325 
2326 	if (qp->rq.wqe_cnt) {
2327 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2328 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2329 	}
2330 
2331 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
2332 		MLX5_SET(qpc, qpc, ts_format, ts_format);
2333 
2334 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
2335 
2336 	if (qp->sq.wqe_cnt) {
2337 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2338 	} else {
2339 		MLX5_SET(qpc, qpc, no_sq, 1);
2340 		if (init_attr->srq &&
2341 		    init_attr->srq->srq_type == IB_SRQT_TM)
2342 			MLX5_SET(qpc, qpc, offload_type,
2343 				 MLX5_QPC_OFFLOAD_TYPE_RNDV);
2344 	}
2345 
2346 	/* Set default resources */
2347 	switch (init_attr->qp_type) {
2348 	case IB_QPT_XRC_INI:
2349 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2350 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2351 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2352 		break;
2353 	default:
2354 		if (init_attr->srq) {
2355 			MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2356 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(init_attr->srq)->msrq.srqn);
2357 		} else {
2358 			MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2359 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s1)->msrq.srqn);
2360 		}
2361 	}
2362 
2363 	if (init_attr->send_cq)
2364 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
2365 
2366 	if (init_attr->recv_cq)
2367 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
2368 
2369 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2370 
2371 	/* 0xffffff means we ask to work with cqe version 0 */
2372 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2373 		MLX5_SET(qpc, qpc, user_index, uidx);
2374 
2375 	if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING &&
2376 	    init_attr->qp_type != IB_QPT_RAW_PACKET) {
2377 		MLX5_SET(qpc, qpc, end_padding_mode,
2378 			 MLX5_WQ_END_PAD_MODE_ALIGN);
2379 		/* Special case to clean flag */
2380 		qp->flags &= ~IB_QP_CREATE_PCI_WRITE_END_PADDING;
2381 	}
2382 
2383 	if (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2384 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
2385 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd->sq_buf_addr;
2386 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
2387 		err = create_raw_packet_qp(dev, qp, in, inlen, pd, udata,
2388 					   &params->resp, init_attr);
2389 	} else
2390 		err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2391 
2392 	kvfree(in);
2393 	if (err)
2394 		goto err_create;
2395 
2396 	base->container_mibqp = qp;
2397 	base->mqp.event = mlx5_ib_qp_event;
2398 	if (MLX5_CAP_GEN(mdev, ece_support))
2399 		params->resp.ece_options = MLX5_GET(create_qp_out, out, ece);
2400 
2401 	get_cqs(qp->type, init_attr->send_cq, init_attr->recv_cq,
2402 		&send_cq, &recv_cq);
2403 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2404 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2405 	/* Maintain device to QPs access, needed for further handling via reset
2406 	 * flow
2407 	 */
2408 	list_add_tail(&qp->qps_list, &dev->qp_list);
2409 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2410 	 */
2411 	if (send_cq)
2412 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2413 	if (recv_cq)
2414 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2415 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2416 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2417 
2418 	return 0;
2419 
2420 err_create:
2421 	destroy_qp(dev, qp, base, udata);
2422 	return err;
2423 }
2424 
2425 static int create_kernel_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2426 			    struct mlx5_ib_qp *qp,
2427 			    struct mlx5_create_qp_params *params)
2428 {
2429 	struct ib_qp_init_attr *attr = params->attr;
2430 	u32 uidx = params->uidx;
2431 	struct mlx5_ib_resources *devr = &dev->devr;
2432 	u32 out[MLX5_ST_SZ_DW(create_qp_out)] = {};
2433 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
2434 	struct mlx5_core_dev *mdev = dev->mdev;
2435 	struct mlx5_ib_cq *send_cq;
2436 	struct mlx5_ib_cq *recv_cq;
2437 	unsigned long flags;
2438 	struct mlx5_ib_qp_base *base;
2439 	int mlx5_st;
2440 	void *qpc;
2441 	u32 *in;
2442 	int err;
2443 
2444 	spin_lock_init(&qp->sq.lock);
2445 	spin_lock_init(&qp->rq.lock);
2446 
2447 	mlx5_st = to_mlx5_st(qp->type);
2448 	if (mlx5_st < 0)
2449 		return -EINVAL;
2450 
2451 	if (attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2452 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2453 
2454 	base = &qp->trans_qp.base;
2455 
2456 	qp->has_rq = qp_has_rq(attr);
2457 	err = set_rq_size(dev, &attr->cap, qp->has_rq, qp, NULL);
2458 	if (err) {
2459 		mlx5_ib_dbg(dev, "err %d\n", err);
2460 		return err;
2461 	}
2462 
2463 	err = _create_kernel_qp(dev, attr, qp, &in, &inlen, base);
2464 	if (err)
2465 		return err;
2466 
2467 	if (is_sqp(attr->qp_type))
2468 		qp->port = attr->port_num;
2469 
2470 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2471 
2472 	MLX5_SET(qpc, qpc, st, mlx5_st);
2473 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2474 
2475 	if (attr->qp_type != MLX5_IB_QPT_REG_UMR)
2476 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
2477 	else
2478 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
2479 
2480 
2481 	if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
2482 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2483 
2484 	if (qp->rq.wqe_cnt) {
2485 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2486 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2487 	}
2488 
2489 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, attr));
2490 
2491 	if (qp->sq.wqe_cnt)
2492 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2493 	else
2494 		MLX5_SET(qpc, qpc, no_sq, 1);
2495 
2496 	if (attr->srq) {
2497 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn0);
2498 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2499 			 to_msrq(attr->srq)->msrq.srqn);
2500 	} else {
2501 		MLX5_SET(qpc, qpc, xrcd, devr->xrcdn1);
2502 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
2503 			 to_msrq(devr->s1)->msrq.srqn);
2504 	}
2505 
2506 	if (attr->send_cq)
2507 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(attr->send_cq)->mcq.cqn);
2508 
2509 	if (attr->recv_cq)
2510 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(attr->recv_cq)->mcq.cqn);
2511 
2512 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2513 
2514 	/* 0xffffff means we ask to work with cqe version 0 */
2515 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2516 		MLX5_SET(qpc, qpc, user_index, uidx);
2517 
2518 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
2519 	if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO)
2520 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
2521 
2522 	err = mlx5_qpc_create_qp(dev, &base->mqp, in, inlen, out);
2523 	kvfree(in);
2524 	if (err)
2525 		goto err_create;
2526 
2527 	base->container_mibqp = qp;
2528 	base->mqp.event = mlx5_ib_qp_event;
2529 
2530 	get_cqs(qp->type, attr->send_cq, attr->recv_cq,
2531 		&send_cq, &recv_cq);
2532 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2533 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2534 	/* Maintain device to QPs access, needed for further handling via reset
2535 	 * flow
2536 	 */
2537 	list_add_tail(&qp->qps_list, &dev->qp_list);
2538 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2539 	 */
2540 	if (send_cq)
2541 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2542 	if (recv_cq)
2543 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2544 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2545 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2546 
2547 	return 0;
2548 
2549 err_create:
2550 	destroy_qp(dev, qp, base, NULL);
2551 	return err;
2552 }
2553 
2554 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2555 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
2556 {
2557 	if (send_cq) {
2558 		if (recv_cq) {
2559 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2560 				spin_lock(&send_cq->lock);
2561 				spin_lock_nested(&recv_cq->lock,
2562 						 SINGLE_DEPTH_NESTING);
2563 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2564 				spin_lock(&send_cq->lock);
2565 				__acquire(&recv_cq->lock);
2566 			} else {
2567 				spin_lock(&recv_cq->lock);
2568 				spin_lock_nested(&send_cq->lock,
2569 						 SINGLE_DEPTH_NESTING);
2570 			}
2571 		} else {
2572 			spin_lock(&send_cq->lock);
2573 			__acquire(&recv_cq->lock);
2574 		}
2575 	} else if (recv_cq) {
2576 		spin_lock(&recv_cq->lock);
2577 		__acquire(&send_cq->lock);
2578 	} else {
2579 		__acquire(&send_cq->lock);
2580 		__acquire(&recv_cq->lock);
2581 	}
2582 }
2583 
2584 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2585 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
2586 {
2587 	if (send_cq) {
2588 		if (recv_cq) {
2589 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2590 				spin_unlock(&recv_cq->lock);
2591 				spin_unlock(&send_cq->lock);
2592 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2593 				__release(&recv_cq->lock);
2594 				spin_unlock(&send_cq->lock);
2595 			} else {
2596 				spin_unlock(&send_cq->lock);
2597 				spin_unlock(&recv_cq->lock);
2598 			}
2599 		} else {
2600 			__release(&recv_cq->lock);
2601 			spin_unlock(&send_cq->lock);
2602 		}
2603 	} else if (recv_cq) {
2604 		__release(&send_cq->lock);
2605 		spin_unlock(&recv_cq->lock);
2606 	} else {
2607 		__release(&recv_cq->lock);
2608 		__release(&send_cq->lock);
2609 	}
2610 }
2611 
2612 static void get_cqs(enum ib_qp_type qp_type,
2613 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
2614 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
2615 {
2616 	switch (qp_type) {
2617 	case IB_QPT_XRC_TGT:
2618 		*send_cq = NULL;
2619 		*recv_cq = NULL;
2620 		break;
2621 	case MLX5_IB_QPT_REG_UMR:
2622 	case IB_QPT_XRC_INI:
2623 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2624 		*recv_cq = NULL;
2625 		break;
2626 
2627 	case IB_QPT_SMI:
2628 	case MLX5_IB_QPT_HW_GSI:
2629 	case IB_QPT_RC:
2630 	case IB_QPT_UC:
2631 	case IB_QPT_UD:
2632 	case IB_QPT_RAW_PACKET:
2633 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2634 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
2635 		break;
2636 	default:
2637 		*send_cq = NULL;
2638 		*recv_cq = NULL;
2639 		break;
2640 	}
2641 }
2642 
2643 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2644 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2645 				u8 lag_tx_affinity);
2646 
2647 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2648 			      struct ib_udata *udata)
2649 {
2650 	struct mlx5_ib_cq *send_cq, *recv_cq;
2651 	struct mlx5_ib_qp_base *base;
2652 	unsigned long flags;
2653 	int err;
2654 
2655 	if (qp->is_rss) {
2656 		destroy_rss_raw_qp_tir(dev, qp);
2657 		return;
2658 	}
2659 
2660 	base = (qp->type == IB_QPT_RAW_PACKET ||
2661 		qp->flags & IB_QP_CREATE_SOURCE_QPN) ?
2662 		       &qp->raw_packet_qp.rq.base :
2663 		       &qp->trans_qp.base;
2664 
2665 	if (qp->state != IB_QPS_RESET) {
2666 		if (qp->type != IB_QPT_RAW_PACKET &&
2667 		    !(qp->flags & IB_QP_CREATE_SOURCE_QPN)) {
2668 			err = mlx5_core_qp_modify(dev, MLX5_CMD_OP_2RST_QP, 0,
2669 						  NULL, &base->mqp, NULL);
2670 		} else {
2671 			struct mlx5_modify_raw_qp_param raw_qp_param = {
2672 				.operation = MLX5_CMD_OP_2RST_QP
2673 			};
2674 
2675 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2676 		}
2677 		if (err)
2678 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
2679 				     base->mqp.qpn);
2680 	}
2681 
2682 	get_cqs(qp->type, qp->ibqp.send_cq, qp->ibqp.recv_cq, &send_cq,
2683 		&recv_cq);
2684 
2685 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2686 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2687 	/* del from lists under both locks above to protect reset flow paths */
2688 	list_del(&qp->qps_list);
2689 	if (send_cq)
2690 		list_del(&qp->cq_send_list);
2691 
2692 	if (recv_cq)
2693 		list_del(&qp->cq_recv_list);
2694 
2695 	if (!udata) {
2696 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2697 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
2698 		if (send_cq != recv_cq)
2699 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
2700 					   NULL);
2701 	}
2702 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2703 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2704 
2705 	if (qp->type == IB_QPT_RAW_PACKET ||
2706 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
2707 		destroy_raw_packet_qp(dev, qp);
2708 	} else {
2709 		err = mlx5_core_destroy_qp(dev, &base->mqp);
2710 		if (err)
2711 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
2712 				     base->mqp.qpn);
2713 	}
2714 
2715 	destroy_qp(dev, qp, base, udata);
2716 }
2717 
2718 static int create_dct(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2719 		      struct mlx5_ib_qp *qp,
2720 		      struct mlx5_create_qp_params *params)
2721 {
2722 	struct ib_qp_init_attr *attr = params->attr;
2723 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
2724 	u32 uidx = params->uidx;
2725 	void *dctc;
2726 
2727 	if (mlx5_lag_is_active(dev->mdev) && !MLX5_CAP_GEN(dev->mdev, lag_dct))
2728 		return -EOPNOTSUPP;
2729 
2730 	qp->dct.in = kzalloc(MLX5_ST_SZ_BYTES(create_dct_in), GFP_KERNEL);
2731 	if (!qp->dct.in)
2732 		return -ENOMEM;
2733 
2734 	MLX5_SET(create_dct_in, qp->dct.in, uid, to_mpd(pd)->uid);
2735 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
2736 	MLX5_SET(dctc, dctc, pd, to_mpd(pd)->pdn);
2737 	MLX5_SET(dctc, dctc, srqn_xrqn, to_msrq(attr->srq)->msrq.srqn);
2738 	MLX5_SET(dctc, dctc, cqn, to_mcq(attr->recv_cq)->mcq.cqn);
2739 	MLX5_SET64(dctc, dctc, dc_access_key, ucmd->access_key);
2740 	MLX5_SET(dctc, dctc, user_index, uidx);
2741 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
2742 		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
2743 
2744 	if (qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) {
2745 		int rcqe_sz = mlx5_ib_get_cqe_size(attr->recv_cq);
2746 
2747 		if (rcqe_sz == 128)
2748 			MLX5_SET(dctc, dctc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
2749 	}
2750 
2751 	qp->state = IB_QPS_RESET;
2752 	return 0;
2753 }
2754 
2755 static int check_qp_type(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
2756 			 enum ib_qp_type *type)
2757 {
2758 	if (attr->qp_type == IB_QPT_DRIVER && !MLX5_CAP_GEN(dev->mdev, dct))
2759 		goto out;
2760 
2761 	switch (attr->qp_type) {
2762 	case IB_QPT_XRC_TGT:
2763 	case IB_QPT_XRC_INI:
2764 		if (!MLX5_CAP_GEN(dev->mdev, xrc))
2765 			goto out;
2766 		fallthrough;
2767 	case IB_QPT_RC:
2768 	case IB_QPT_UC:
2769 	case IB_QPT_SMI:
2770 	case MLX5_IB_QPT_HW_GSI:
2771 	case IB_QPT_DRIVER:
2772 	case IB_QPT_GSI:
2773 	case IB_QPT_RAW_PACKET:
2774 	case IB_QPT_UD:
2775 	case MLX5_IB_QPT_REG_UMR:
2776 		break;
2777 	default:
2778 		goto out;
2779 	}
2780 
2781 	*type = attr->qp_type;
2782 	return 0;
2783 
2784 out:
2785 	mlx5_ib_dbg(dev, "Unsupported QP type %d\n", attr->qp_type);
2786 	return -EOPNOTSUPP;
2787 }
2788 
2789 static int check_valid_flow(struct mlx5_ib_dev *dev, struct ib_pd *pd,
2790 			    struct ib_qp_init_attr *attr,
2791 			    struct ib_udata *udata)
2792 {
2793 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2794 		udata, struct mlx5_ib_ucontext, ibucontext);
2795 
2796 	if (!udata) {
2797 		/* Kernel create_qp callers */
2798 		if (attr->rwq_ind_tbl)
2799 			return -EOPNOTSUPP;
2800 
2801 		switch (attr->qp_type) {
2802 		case IB_QPT_RAW_PACKET:
2803 		case IB_QPT_DRIVER:
2804 			return -EOPNOTSUPP;
2805 		default:
2806 			return 0;
2807 		}
2808 	}
2809 
2810 	/* Userspace create_qp callers */
2811 	if (attr->qp_type == IB_QPT_RAW_PACKET && !ucontext->cqe_version) {
2812 		mlx5_ib_dbg(dev,
2813 			"Raw Packet QP is only supported for CQE version > 0\n");
2814 		return -EINVAL;
2815 	}
2816 
2817 	if (attr->qp_type != IB_QPT_RAW_PACKET && attr->rwq_ind_tbl) {
2818 		mlx5_ib_dbg(dev,
2819 			    "Wrong QP type %d for the RWQ indirect table\n",
2820 			    attr->qp_type);
2821 		return -EINVAL;
2822 	}
2823 
2824 	/*
2825 	 * We don't need to see this warning, it means that kernel code
2826 	 * missing ib_pd. Placed here to catch developer's mistakes.
2827 	 */
2828 	WARN_ONCE(!pd && attr->qp_type != IB_QPT_XRC_TGT,
2829 		  "There is a missing PD pointer assignment\n");
2830 	return 0;
2831 }
2832 
2833 static void process_vendor_flag(struct mlx5_ib_dev *dev, int *flags, int flag,
2834 				bool cond, struct mlx5_ib_qp *qp)
2835 {
2836 	if (!(*flags & flag))
2837 		return;
2838 
2839 	if (cond) {
2840 		qp->flags_en |= flag;
2841 		*flags &= ~flag;
2842 		return;
2843 	}
2844 
2845 	switch (flag) {
2846 	case MLX5_QP_FLAG_SCATTER_CQE:
2847 	case MLX5_QP_FLAG_ALLOW_SCATTER_CQE:
2848 		/*
2849 			 * We don't return error if these flags were provided,
2850 			 * and mlx5 doesn't have right capability.
2851 			 */
2852 		*flags &= ~(MLX5_QP_FLAG_SCATTER_CQE |
2853 			    MLX5_QP_FLAG_ALLOW_SCATTER_CQE);
2854 		return;
2855 	default:
2856 		break;
2857 	}
2858 	mlx5_ib_dbg(dev, "Vendor create QP flag 0x%X is not supported\n", flag);
2859 }
2860 
2861 static int process_vendor_flags(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2862 				void *ucmd, struct ib_qp_init_attr *attr)
2863 {
2864 	struct mlx5_core_dev *mdev = dev->mdev;
2865 	bool cond;
2866 	int flags;
2867 
2868 	if (attr->rwq_ind_tbl)
2869 		flags = ((struct mlx5_ib_create_qp_rss *)ucmd)->flags;
2870 	else
2871 		flags = ((struct mlx5_ib_create_qp *)ucmd)->flags;
2872 
2873 	switch (flags & (MLX5_QP_FLAG_TYPE_DCT | MLX5_QP_FLAG_TYPE_DCI)) {
2874 	case MLX5_QP_FLAG_TYPE_DCI:
2875 		qp->type = MLX5_IB_QPT_DCI;
2876 		break;
2877 	case MLX5_QP_FLAG_TYPE_DCT:
2878 		qp->type = MLX5_IB_QPT_DCT;
2879 		break;
2880 	default:
2881 		if (qp->type != IB_QPT_DRIVER)
2882 			break;
2883 		/*
2884 		 * It is IB_QPT_DRIVER and or no subtype or
2885 		 * wrong subtype were provided.
2886 		 */
2887 		return -EINVAL;
2888 	}
2889 
2890 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TYPE_DCI, true, qp);
2891 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TYPE_DCT, true, qp);
2892 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_DCI_STREAM,
2893 			    MLX5_CAP_GEN(mdev, log_max_dci_stream_channels),
2894 			    qp);
2895 
2896 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_SIGNATURE, true, qp);
2897 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_SCATTER_CQE,
2898 			    MLX5_CAP_GEN(mdev, sctr_data_cqe), qp);
2899 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_ALLOW_SCATTER_CQE,
2900 			    MLX5_CAP_GEN(mdev, sctr_data_cqe), qp);
2901 
2902 	if (qp->type == IB_QPT_RAW_PACKET) {
2903 		cond = MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan) ||
2904 		       MLX5_CAP_ETH(mdev, tunnel_stateless_gre) ||
2905 		       MLX5_CAP_ETH(mdev, tunnel_stateless_geneve_rx);
2906 		process_vendor_flag(dev, &flags, MLX5_QP_FLAG_TUNNEL_OFFLOADS,
2907 				    cond, qp);
2908 		process_vendor_flag(dev, &flags,
2909 				    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC, true,
2910 				    qp);
2911 		process_vendor_flag(dev, &flags,
2912 				    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC, true,
2913 				    qp);
2914 	}
2915 
2916 	if (qp->type == IB_QPT_RC)
2917 		process_vendor_flag(dev, &flags,
2918 				    MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE,
2919 				    MLX5_CAP_GEN(mdev, qp_packet_based), qp);
2920 
2921 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_BFREG_INDEX, true, qp);
2922 	process_vendor_flag(dev, &flags, MLX5_QP_FLAG_UAR_PAGE_INDEX, true, qp);
2923 
2924 	cond = qp->flags_en & ~(MLX5_QP_FLAG_TUNNEL_OFFLOADS |
2925 				MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
2926 				MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC);
2927 	if (attr->rwq_ind_tbl && cond) {
2928 		mlx5_ib_dbg(dev, "RSS RAW QP has unsupported flags 0x%X\n",
2929 			    cond);
2930 		return -EINVAL;
2931 	}
2932 
2933 	if (flags)
2934 		mlx5_ib_dbg(dev, "udata has unsupported flags 0x%X\n", flags);
2935 
2936 	return (flags) ? -EINVAL : 0;
2937 	}
2938 
2939 static void process_create_flag(struct mlx5_ib_dev *dev, int *flags, int flag,
2940 				bool cond, struct mlx5_ib_qp *qp)
2941 {
2942 	if (!(*flags & flag))
2943 		return;
2944 
2945 	if (cond) {
2946 		qp->flags |= flag;
2947 		*flags &= ~flag;
2948 		return;
2949 	}
2950 
2951 	if (flag == MLX5_IB_QP_CREATE_WC_TEST) {
2952 		/*
2953 		 * Special case, if condition didn't meet, it won't be error,
2954 		 * just different in-kernel flow.
2955 		 */
2956 		*flags &= ~MLX5_IB_QP_CREATE_WC_TEST;
2957 		return;
2958 	}
2959 	mlx5_ib_dbg(dev, "Verbs create QP flag 0x%X is not supported\n", flag);
2960 }
2961 
2962 static int process_create_flags(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2963 				struct ib_qp_init_attr *attr)
2964 {
2965 	enum ib_qp_type qp_type = qp->type;
2966 	struct mlx5_core_dev *mdev = dev->mdev;
2967 	int create_flags = attr->create_flags;
2968 	bool cond;
2969 
2970 	if (qp_type == MLX5_IB_QPT_DCT)
2971 		return (create_flags) ? -EINVAL : 0;
2972 
2973 	if (qp_type == IB_QPT_RAW_PACKET && attr->rwq_ind_tbl)
2974 		return (create_flags) ? -EINVAL : 0;
2975 
2976 	process_create_flag(dev, &create_flags, IB_QP_CREATE_NETIF_QP,
2977 			    mlx5_get_flow_namespace(dev->mdev,
2978 						    MLX5_FLOW_NAMESPACE_BYPASS),
2979 			    qp);
2980 	process_create_flag(dev, &create_flags,
2981 			    IB_QP_CREATE_INTEGRITY_EN,
2982 			    MLX5_CAP_GEN(mdev, sho), qp);
2983 	process_create_flag(dev, &create_flags,
2984 			    IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK,
2985 			    MLX5_CAP_GEN(mdev, block_lb_mc), qp);
2986 	process_create_flag(dev, &create_flags, IB_QP_CREATE_CROSS_CHANNEL,
2987 			    MLX5_CAP_GEN(mdev, cd), qp);
2988 	process_create_flag(dev, &create_flags, IB_QP_CREATE_MANAGED_SEND,
2989 			    MLX5_CAP_GEN(mdev, cd), qp);
2990 	process_create_flag(dev, &create_flags, IB_QP_CREATE_MANAGED_RECV,
2991 			    MLX5_CAP_GEN(mdev, cd), qp);
2992 
2993 	if (qp_type == IB_QPT_UD) {
2994 		process_create_flag(dev, &create_flags,
2995 				    IB_QP_CREATE_IPOIB_UD_LSO,
2996 				    MLX5_CAP_GEN(mdev, ipoib_basic_offloads),
2997 				    qp);
2998 		cond = MLX5_CAP_GEN(mdev, port_type) == MLX5_CAP_PORT_TYPE_IB;
2999 		process_create_flag(dev, &create_flags, IB_QP_CREATE_SOURCE_QPN,
3000 				    cond, qp);
3001 	}
3002 
3003 	if (qp_type == IB_QPT_RAW_PACKET) {
3004 		cond = MLX5_CAP_GEN(mdev, eth_net_offloads) &&
3005 		       MLX5_CAP_ETH(mdev, scatter_fcs);
3006 		process_create_flag(dev, &create_flags,
3007 				    IB_QP_CREATE_SCATTER_FCS, cond, qp);
3008 
3009 		cond = MLX5_CAP_GEN(mdev, eth_net_offloads) &&
3010 		       MLX5_CAP_ETH(mdev, vlan_cap);
3011 		process_create_flag(dev, &create_flags,
3012 				    IB_QP_CREATE_CVLAN_STRIPPING, cond, qp);
3013 	}
3014 
3015 	process_create_flag(dev, &create_flags,
3016 			    IB_QP_CREATE_PCI_WRITE_END_PADDING,
3017 			    MLX5_CAP_GEN(mdev, end_pad), qp);
3018 
3019 	process_create_flag(dev, &create_flags, MLX5_IB_QP_CREATE_WC_TEST,
3020 			    qp_type != MLX5_IB_QPT_REG_UMR, qp);
3021 	process_create_flag(dev, &create_flags, MLX5_IB_QP_CREATE_SQPN_QP1,
3022 			    true, qp);
3023 
3024 	if (create_flags) {
3025 		mlx5_ib_dbg(dev, "Create QP has unsupported flags 0x%X\n",
3026 			    create_flags);
3027 		return -EOPNOTSUPP;
3028 	}
3029 	return 0;
3030 }
3031 
3032 static int process_udata_size(struct mlx5_ib_dev *dev,
3033 			      struct mlx5_create_qp_params *params)
3034 {
3035 	size_t ucmd = sizeof(struct mlx5_ib_create_qp);
3036 	struct ib_udata *udata = params->udata;
3037 	size_t outlen = udata->outlen;
3038 	size_t inlen = udata->inlen;
3039 
3040 	params->outlen = min(outlen, sizeof(struct mlx5_ib_create_qp_resp));
3041 	params->ucmd_size = ucmd;
3042 	if (!params->is_rss_raw) {
3043 		/* User has old rdma-core, which doesn't support ECE */
3044 		size_t min_inlen =
3045 			offsetof(struct mlx5_ib_create_qp, ece_options);
3046 
3047 		/*
3048 		 * We will check in check_ucmd_data() that user
3049 		 * cleared everything after inlen.
3050 		 */
3051 		params->inlen = (inlen < min_inlen) ? 0 : min(inlen, ucmd);
3052 		goto out;
3053 	}
3054 
3055 	/* RSS RAW QP */
3056 	if (inlen < offsetofend(struct mlx5_ib_create_qp_rss, flags))
3057 		return -EINVAL;
3058 
3059 	if (outlen < offsetofend(struct mlx5_ib_create_qp_resp, bfreg_index))
3060 		return -EINVAL;
3061 
3062 	ucmd = sizeof(struct mlx5_ib_create_qp_rss);
3063 	params->ucmd_size = ucmd;
3064 	if (inlen > ucmd && !ib_is_udata_cleared(udata, ucmd, inlen - ucmd))
3065 		return -EINVAL;
3066 
3067 	params->inlen = min(ucmd, inlen);
3068 out:
3069 	if (!params->inlen)
3070 		mlx5_ib_dbg(dev, "udata is too small\n");
3071 
3072 	return (params->inlen) ? 0 : -EINVAL;
3073 }
3074 
3075 static int create_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
3076 		     struct mlx5_ib_qp *qp,
3077 		     struct mlx5_create_qp_params *params)
3078 {
3079 	int err;
3080 
3081 	if (params->is_rss_raw) {
3082 		err = create_rss_raw_qp_tir(dev, pd, qp, params);
3083 		goto out;
3084 	}
3085 
3086 	switch (qp->type) {
3087 	case MLX5_IB_QPT_DCT:
3088 		err = create_dct(dev, pd, qp, params);
3089 		rdma_restrack_no_track(&qp->ibqp.res);
3090 		break;
3091 	case MLX5_IB_QPT_DCI:
3092 		err = create_dci(dev, pd, qp, params);
3093 		break;
3094 	case IB_QPT_XRC_TGT:
3095 		err = create_xrc_tgt_qp(dev, qp, params);
3096 		break;
3097 	case IB_QPT_GSI:
3098 		err = mlx5_ib_create_gsi(pd, qp, params->attr);
3099 		break;
3100 	case MLX5_IB_QPT_HW_GSI:
3101 	case MLX5_IB_QPT_REG_UMR:
3102 		rdma_restrack_no_track(&qp->ibqp.res);
3103 		fallthrough;
3104 	default:
3105 		if (params->udata)
3106 			err = create_user_qp(dev, pd, qp, params);
3107 		else
3108 			err = create_kernel_qp(dev, pd, qp, params);
3109 	}
3110 
3111 out:
3112 	if (err) {
3113 		mlx5_ib_err(dev, "Create QP type %d failed\n", qp->type);
3114 		return err;
3115 	}
3116 
3117 	if (is_qp0(qp->type))
3118 		qp->ibqp.qp_num = 0;
3119 	else if (is_qp1(qp->type))
3120 		qp->ibqp.qp_num = 1;
3121 	else
3122 		qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
3123 
3124 	mlx5_ib_dbg(dev,
3125 		"QP type %d, ib qpn 0x%X, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x, ece 0x%x\n",
3126 		qp->type, qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
3127 		params->attr->recv_cq ? to_mcq(params->attr->recv_cq)->mcq.cqn :
3128 					-1,
3129 		params->attr->send_cq ? to_mcq(params->attr->send_cq)->mcq.cqn :
3130 					-1,
3131 		params->resp.ece_options);
3132 
3133 	return 0;
3134 }
3135 
3136 static int check_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3137 			 struct ib_qp_init_attr *attr)
3138 {
3139 	int ret = 0;
3140 
3141 	switch (qp->type) {
3142 	case MLX5_IB_QPT_DCT:
3143 		ret = (!attr->srq || !attr->recv_cq) ? -EINVAL : 0;
3144 		break;
3145 	case MLX5_IB_QPT_DCI:
3146 		ret = (attr->cap.max_recv_wr || attr->cap.max_recv_sge) ?
3147 			      -EINVAL :
3148 			      0;
3149 		break;
3150 	case IB_QPT_RAW_PACKET:
3151 		ret = (attr->rwq_ind_tbl && attr->send_cq) ? -EINVAL : 0;
3152 		break;
3153 	default:
3154 		break;
3155 	}
3156 
3157 	if (ret)
3158 		mlx5_ib_dbg(dev, "QP type %d has wrong attributes\n", qp->type);
3159 
3160 	return ret;
3161 }
3162 
3163 static int get_qp_uidx(struct mlx5_ib_qp *qp,
3164 		       struct mlx5_create_qp_params *params)
3165 {
3166 	struct mlx5_ib_create_qp *ucmd = params->ucmd;
3167 	struct ib_udata *udata = params->udata;
3168 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3169 		udata, struct mlx5_ib_ucontext, ibucontext);
3170 
3171 	if (params->is_rss_raw)
3172 		return 0;
3173 
3174 	return get_qp_user_index(ucontext, ucmd, sizeof(*ucmd), &params->uidx);
3175 }
3176 
3177 static int mlx5_ib_destroy_dct(struct mlx5_ib_qp *mqp)
3178 {
3179 	struct mlx5_ib_dev *dev = to_mdev(mqp->ibqp.device);
3180 
3181 	if (mqp->state == IB_QPS_RTR) {
3182 		int err;
3183 
3184 		err = mlx5_core_destroy_dct(dev, &mqp->dct.mdct);
3185 		if (err) {
3186 			mlx5_ib_warn(dev, "failed to destroy DCT %d\n", err);
3187 			return err;
3188 		}
3189 	}
3190 
3191 	kfree(mqp->dct.in);
3192 	return 0;
3193 }
3194 
3195 static int check_ucmd_data(struct mlx5_ib_dev *dev,
3196 			   struct mlx5_create_qp_params *params)
3197 {
3198 	struct ib_udata *udata = params->udata;
3199 	size_t size, last;
3200 	int ret;
3201 
3202 	if (params->is_rss_raw)
3203 		/*
3204 		 * These QPs don't have "reserved" field in their
3205 		 * create_qp input struct, so their data is always valid.
3206 		 */
3207 		last = sizeof(struct mlx5_ib_create_qp_rss);
3208 	else
3209 		last = offsetof(struct mlx5_ib_create_qp, reserved);
3210 
3211 	if (udata->inlen <= last)
3212 		return 0;
3213 
3214 	/*
3215 	 * User provides different create_qp structures based on the
3216 	 * flow and we need to know if he cleared memory after our
3217 	 * struct create_qp ends.
3218 	 */
3219 	size = udata->inlen - last;
3220 	ret = ib_is_udata_cleared(params->udata, last, size);
3221 	if (!ret)
3222 		mlx5_ib_dbg(
3223 			dev,
3224 			"udata is not cleared, inlen = %zu, ucmd = %zu, last = %zu, size = %zu\n",
3225 			udata->inlen, params->ucmd_size, last, size);
3226 	return ret ? 0 : -EINVAL;
3227 }
3228 
3229 int mlx5_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
3230 		      struct ib_udata *udata)
3231 {
3232 	struct mlx5_create_qp_params params = {};
3233 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3234 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3235 	struct ib_pd *pd = ibqp->pd;
3236 	enum ib_qp_type type;
3237 	int err;
3238 
3239 	err = check_qp_type(dev, attr, &type);
3240 	if (err)
3241 		return err;
3242 
3243 	err = check_valid_flow(dev, pd, attr, udata);
3244 	if (err)
3245 		return err;
3246 
3247 	params.udata = udata;
3248 	params.uidx = MLX5_IB_DEFAULT_UIDX;
3249 	params.attr = attr;
3250 	params.is_rss_raw = !!attr->rwq_ind_tbl;
3251 
3252 	if (udata) {
3253 		err = process_udata_size(dev, &params);
3254 		if (err)
3255 			return err;
3256 
3257 		err = check_ucmd_data(dev, &params);
3258 		if (err)
3259 			return err;
3260 
3261 		params.ucmd = kzalloc(params.ucmd_size, GFP_KERNEL);
3262 		if (!params.ucmd)
3263 			return -ENOMEM;
3264 
3265 		err = ib_copy_from_udata(params.ucmd, udata, params.inlen);
3266 		if (err)
3267 			goto free_ucmd;
3268 	}
3269 
3270 	mutex_init(&qp->mutex);
3271 	qp->type = type;
3272 	if (udata) {
3273 		err = process_vendor_flags(dev, qp, params.ucmd, attr);
3274 		if (err)
3275 			goto free_ucmd;
3276 
3277 		err = get_qp_uidx(qp, &params);
3278 		if (err)
3279 			goto free_ucmd;
3280 	}
3281 	err = process_create_flags(dev, qp, attr);
3282 	if (err)
3283 		goto free_ucmd;
3284 
3285 	err = check_qp_attr(dev, qp, attr);
3286 	if (err)
3287 		goto free_ucmd;
3288 
3289 	err = create_qp(dev, pd, qp, &params);
3290 	if (err)
3291 		goto free_ucmd;
3292 
3293 	kfree(params.ucmd);
3294 	params.ucmd = NULL;
3295 
3296 	if (udata)
3297 		/*
3298 		 * It is safe to copy response for all user create QP flows,
3299 		 * including MLX5_IB_QPT_DCT, which doesn't need it.
3300 		 * In that case, resp will be filled with zeros.
3301 		 */
3302 		err = ib_copy_to_udata(udata, &params.resp, params.outlen);
3303 	if (err)
3304 		goto destroy_qp;
3305 
3306 	return 0;
3307 
3308 destroy_qp:
3309 	switch (qp->type) {
3310 	case MLX5_IB_QPT_DCT:
3311 		mlx5_ib_destroy_dct(qp);
3312 		break;
3313 	case IB_QPT_GSI:
3314 		mlx5_ib_destroy_gsi(qp);
3315 		break;
3316 	default:
3317 		destroy_qp_common(dev, qp, udata);
3318 	}
3319 
3320 free_ucmd:
3321 	kfree(params.ucmd);
3322 	return err;
3323 }
3324 
3325 int mlx5_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
3326 {
3327 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3328 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3329 
3330 	if (mqp->type == IB_QPT_GSI)
3331 		return mlx5_ib_destroy_gsi(mqp);
3332 
3333 	if (mqp->type == MLX5_IB_QPT_DCT)
3334 		return mlx5_ib_destroy_dct(mqp);
3335 
3336 	destroy_qp_common(dev, mqp, udata);
3337 	return 0;
3338 }
3339 
3340 static int set_qpc_atomic_flags(struct mlx5_ib_qp *qp,
3341 				const struct ib_qp_attr *attr, int attr_mask,
3342 				void *qpc)
3343 {
3344 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
3345 	u8 dest_rd_atomic;
3346 	u32 access_flags;
3347 
3348 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
3349 		dest_rd_atomic = attr->max_dest_rd_atomic;
3350 	else
3351 		dest_rd_atomic = qp->trans_qp.resp_depth;
3352 
3353 	if (attr_mask & IB_QP_ACCESS_FLAGS)
3354 		access_flags = attr->qp_access_flags;
3355 	else
3356 		access_flags = qp->trans_qp.atomic_rd_en;
3357 
3358 	if (!dest_rd_atomic)
3359 		access_flags &= IB_ACCESS_REMOTE_WRITE;
3360 
3361 	MLX5_SET(qpc, qpc, rre, !!(access_flags & IB_ACCESS_REMOTE_READ));
3362 
3363 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC) {
3364 		int atomic_mode;
3365 
3366 		atomic_mode = get_atomic_mode(dev, qp->type);
3367 		if (atomic_mode < 0)
3368 			return -EOPNOTSUPP;
3369 
3370 		MLX5_SET(qpc, qpc, rae, 1);
3371 		MLX5_SET(qpc, qpc, atomic_mode, atomic_mode);
3372 	}
3373 
3374 	MLX5_SET(qpc, qpc, rwe, !!(access_flags & IB_ACCESS_REMOTE_WRITE));
3375 	return 0;
3376 }
3377 
3378 enum {
3379 	MLX5_PATH_FLAG_FL	= 1 << 0,
3380 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
3381 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
3382 };
3383 
3384 static int mlx5_to_ib_rate_map(u8 rate)
3385 {
3386 	static const int rates[] = { IB_RATE_PORT_CURRENT, IB_RATE_56_GBPS,
3387 				     IB_RATE_25_GBPS,	   IB_RATE_100_GBPS,
3388 				     IB_RATE_200_GBPS,	   IB_RATE_50_GBPS,
3389 				     IB_RATE_400_GBPS };
3390 
3391 	if (rate < ARRAY_SIZE(rates))
3392 		return rates[rate];
3393 
3394 	return rate - MLX5_STAT_RATE_OFFSET;
3395 }
3396 
3397 static int ib_to_mlx5_rate_map(u8 rate)
3398 {
3399 	switch (rate) {
3400 	case IB_RATE_PORT_CURRENT:
3401 		return 0;
3402 	case IB_RATE_56_GBPS:
3403 		return 1;
3404 	case IB_RATE_25_GBPS:
3405 		return 2;
3406 	case IB_RATE_100_GBPS:
3407 		return 3;
3408 	case IB_RATE_200_GBPS:
3409 		return 4;
3410 	case IB_RATE_50_GBPS:
3411 		return 5;
3412 	case IB_RATE_400_GBPS:
3413 		return 6;
3414 	default:
3415 		return rate + MLX5_STAT_RATE_OFFSET;
3416 	}
3417 
3418 	return 0;
3419 }
3420 
3421 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
3422 {
3423 	u32 stat_rate_support;
3424 
3425 	if (rate == IB_RATE_PORT_CURRENT)
3426 		return 0;
3427 
3428 	if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_600_GBPS)
3429 		return -EINVAL;
3430 
3431 	stat_rate_support = MLX5_CAP_GEN(dev->mdev, stat_rate_support);
3432 	while (rate != IB_RATE_PORT_CURRENT &&
3433 	       !(1 << ib_to_mlx5_rate_map(rate) & stat_rate_support))
3434 		--rate;
3435 
3436 	return ib_to_mlx5_rate_map(rate);
3437 }
3438 
3439 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
3440 				      struct mlx5_ib_sq *sq, u8 sl,
3441 				      struct ib_pd *pd)
3442 {
3443 	void *in;
3444 	void *tisc;
3445 	int inlen;
3446 	int err;
3447 
3448 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
3449 	in = kvzalloc(inlen, GFP_KERNEL);
3450 	if (!in)
3451 		return -ENOMEM;
3452 
3453 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
3454 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
3455 
3456 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
3457 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
3458 
3459 	err = mlx5_core_modify_tis(dev, sq->tisn, in);
3460 
3461 	kvfree(in);
3462 
3463 	return err;
3464 }
3465 
3466 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
3467 					 struct mlx5_ib_sq *sq, u8 tx_affinity,
3468 					 struct ib_pd *pd)
3469 {
3470 	void *in;
3471 	void *tisc;
3472 	int inlen;
3473 	int err;
3474 
3475 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
3476 	in = kvzalloc(inlen, GFP_KERNEL);
3477 	if (!in)
3478 		return -ENOMEM;
3479 
3480 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
3481 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
3482 
3483 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
3484 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
3485 
3486 	err = mlx5_core_modify_tis(dev, sq->tisn, in);
3487 
3488 	kvfree(in);
3489 
3490 	return err;
3491 }
3492 
3493 static void mlx5_set_path_udp_sport(void *path, const struct rdma_ah_attr *ah,
3494 				    u32 lqpn, u32 rqpn)
3495 
3496 {
3497 	u32 fl = ah->grh.flow_label;
3498 
3499 	if (!fl)
3500 		fl = rdma_calc_flow_label(lqpn, rqpn);
3501 
3502 	MLX5_SET(ads, path, udp_sport, rdma_flow_label_to_udp_sport(fl));
3503 }
3504 
3505 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3506 			 const struct rdma_ah_attr *ah, void *path, u8 port,
3507 			 int attr_mask, u32 path_flags,
3508 			 const struct ib_qp_attr *attr, bool alt)
3509 {
3510 	const struct ib_global_route *grh = rdma_ah_read_grh(ah);
3511 	int err;
3512 	enum ib_gid_type gid_type;
3513 	u8 ah_flags = rdma_ah_get_ah_flags(ah);
3514 	u8 sl = rdma_ah_get_sl(ah);
3515 
3516 	if (attr_mask & IB_QP_PKEY_INDEX)
3517 		MLX5_SET(ads, path, pkey_index,
3518 			 alt ? attr->alt_pkey_index : attr->pkey_index);
3519 
3520 	if (ah_flags & IB_AH_GRH) {
3521 		const struct ib_port_immutable *immutable;
3522 
3523 		immutable = ib_port_immutable_read(&dev->ib_dev, port);
3524 		if (grh->sgid_index >= immutable->gid_tbl_len) {
3525 			pr_err("sgid_index (%u) too large. max is %d\n",
3526 			       grh->sgid_index,
3527 			       immutable->gid_tbl_len);
3528 			return -EINVAL;
3529 		}
3530 	}
3531 
3532 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
3533 		if (!(ah_flags & IB_AH_GRH))
3534 			return -EINVAL;
3535 
3536 		ether_addr_copy(MLX5_ADDR_OF(ads, path, rmac_47_32),
3537 				ah->roce.dmac);
3538 		if ((qp->type == IB_QPT_RC ||
3539 		     qp->type == IB_QPT_UC ||
3540 		     qp->type == IB_QPT_XRC_INI ||
3541 		     qp->type == IB_QPT_XRC_TGT) &&
3542 		    (grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) &&
3543 		    (attr_mask & IB_QP_DEST_QPN))
3544 			mlx5_set_path_udp_sport(path, ah,
3545 						qp->ibqp.qp_num,
3546 						attr->dest_qp_num);
3547 		MLX5_SET(ads, path, eth_prio, sl & 0x7);
3548 		gid_type = ah->grh.sgid_attr->gid_type;
3549 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
3550 			MLX5_SET(ads, path, dscp, grh->traffic_class >> 2);
3551 	} else {
3552 		MLX5_SET(ads, path, fl, !!(path_flags & MLX5_PATH_FLAG_FL));
3553 		MLX5_SET(ads, path, free_ar,
3554 			 !!(path_flags & MLX5_PATH_FLAG_FREE_AR));
3555 		MLX5_SET(ads, path, rlid, rdma_ah_get_dlid(ah));
3556 		MLX5_SET(ads, path, mlid, rdma_ah_get_path_bits(ah));
3557 		MLX5_SET(ads, path, grh, !!(ah_flags & IB_AH_GRH));
3558 		MLX5_SET(ads, path, sl, sl);
3559 	}
3560 
3561 	if (ah_flags & IB_AH_GRH) {
3562 		MLX5_SET(ads, path, src_addr_index, grh->sgid_index);
3563 		MLX5_SET(ads, path, hop_limit, grh->hop_limit);
3564 		MLX5_SET(ads, path, tclass, grh->traffic_class);
3565 		MLX5_SET(ads, path, flow_label, grh->flow_label);
3566 		memcpy(MLX5_ADDR_OF(ads, path, rgid_rip), grh->dgid.raw,
3567 		       sizeof(grh->dgid.raw));
3568 	}
3569 
3570 	err = ib_rate_to_mlx5(dev, rdma_ah_get_static_rate(ah));
3571 	if (err < 0)
3572 		return err;
3573 	MLX5_SET(ads, path, stat_rate, err);
3574 	MLX5_SET(ads, path, vhca_port_num, port);
3575 
3576 	if (attr_mask & IB_QP_TIMEOUT)
3577 		MLX5_SET(ads, path, ack_timeout,
3578 			 alt ? attr->alt_timeout : attr->timeout);
3579 
3580 	if ((qp->type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
3581 		return modify_raw_packet_eth_prio(dev->mdev,
3582 						  &qp->raw_packet_qp.sq,
3583 						  sl & 0xf, qp->ibqp.pd);
3584 
3585 	return 0;
3586 }
3587 
3588 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
3589 	[MLX5_QP_STATE_INIT] = {
3590 		[MLX5_QP_STATE_INIT] = {
3591 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3592 					  MLX5_QP_OPTPAR_RAE		|
3593 					  MLX5_QP_OPTPAR_RWE		|
3594 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3595 					  MLX5_QP_OPTPAR_PRI_PORT	|
3596 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3597 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3598 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3599 					  MLX5_QP_OPTPAR_PRI_PORT	|
3600 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3601 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3602 					  MLX5_QP_OPTPAR_Q_KEY		|
3603 					  MLX5_QP_OPTPAR_PRI_PORT,
3604 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3605 					  MLX5_QP_OPTPAR_RAE		|
3606 					  MLX5_QP_OPTPAR_RWE		|
3607 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3608 					  MLX5_QP_OPTPAR_PRI_PORT	|
3609 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3610 		},
3611 		[MLX5_QP_STATE_RTR] = {
3612 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3613 					  MLX5_QP_OPTPAR_RRE            |
3614 					  MLX5_QP_OPTPAR_RAE            |
3615 					  MLX5_QP_OPTPAR_RWE            |
3616 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3617 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3618 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3619 					  MLX5_QP_OPTPAR_RWE            |
3620 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3621 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3622 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
3623 					  MLX5_QP_OPTPAR_Q_KEY,
3624 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3625 					   MLX5_QP_OPTPAR_Q_KEY,
3626 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
3627 					  MLX5_QP_OPTPAR_RRE            |
3628 					  MLX5_QP_OPTPAR_RAE            |
3629 					  MLX5_QP_OPTPAR_RWE            |
3630 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3631 					  MLX5_QP_OPTPAR_LAG_TX_AFF,
3632 		},
3633 	},
3634 	[MLX5_QP_STATE_RTR] = {
3635 		[MLX5_QP_STATE_RTS] = {
3636 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3637 					  MLX5_QP_OPTPAR_RRE		|
3638 					  MLX5_QP_OPTPAR_RAE		|
3639 					  MLX5_QP_OPTPAR_RWE		|
3640 					  MLX5_QP_OPTPAR_PM_STATE	|
3641 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3642 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3643 					  MLX5_QP_OPTPAR_RWE		|
3644 					  MLX5_QP_OPTPAR_PM_STATE,
3645 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
3646 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3647 					  MLX5_QP_OPTPAR_RRE		|
3648 					  MLX5_QP_OPTPAR_RAE		|
3649 					  MLX5_QP_OPTPAR_RWE		|
3650 					  MLX5_QP_OPTPAR_PM_STATE	|
3651 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3652 		},
3653 	},
3654 	[MLX5_QP_STATE_RTS] = {
3655 		[MLX5_QP_STATE_RTS] = {
3656 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3657 					  MLX5_QP_OPTPAR_RAE		|
3658 					  MLX5_QP_OPTPAR_RWE		|
3659 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3660 					  MLX5_QP_OPTPAR_PM_STATE	|
3661 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3662 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3663 					  MLX5_QP_OPTPAR_PM_STATE	|
3664 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3665 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
3666 					  MLX5_QP_OPTPAR_SRQN		|
3667 					  MLX5_QP_OPTPAR_CQN_RCV,
3668 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3669 					  MLX5_QP_OPTPAR_RAE		|
3670 					  MLX5_QP_OPTPAR_RWE		|
3671 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3672 					  MLX5_QP_OPTPAR_PM_STATE	|
3673 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3674 		},
3675 	},
3676 	[MLX5_QP_STATE_SQER] = {
3677 		[MLX5_QP_STATE_RTS] = {
3678 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
3679 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
3680 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
3681 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3682 					   MLX5_QP_OPTPAR_RWE		|
3683 					   MLX5_QP_OPTPAR_RAE		|
3684 					   MLX5_QP_OPTPAR_RRE,
3685 			[MLX5_QP_ST_XRC]  = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3686 					   MLX5_QP_OPTPAR_RWE		|
3687 					   MLX5_QP_OPTPAR_RAE		|
3688 					   MLX5_QP_OPTPAR_RRE,
3689 		},
3690 	},
3691 	[MLX5_QP_STATE_SQD] = {
3692 		[MLX5_QP_STATE_RTS] = {
3693 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
3694 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
3695 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE,
3696 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3697 					  MLX5_QP_OPTPAR_RWE		|
3698 					  MLX5_QP_OPTPAR_RAE		|
3699 					  MLX5_QP_OPTPAR_RRE,
3700 		},
3701 	},
3702 };
3703 
3704 static int ib_nr_to_mlx5_nr(int ib_mask)
3705 {
3706 	switch (ib_mask) {
3707 	case IB_QP_STATE:
3708 		return 0;
3709 	case IB_QP_CUR_STATE:
3710 		return 0;
3711 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
3712 		return 0;
3713 	case IB_QP_ACCESS_FLAGS:
3714 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
3715 			MLX5_QP_OPTPAR_RAE;
3716 	case IB_QP_PKEY_INDEX:
3717 		return MLX5_QP_OPTPAR_PKEY_INDEX;
3718 	case IB_QP_PORT:
3719 		return MLX5_QP_OPTPAR_PRI_PORT;
3720 	case IB_QP_QKEY:
3721 		return MLX5_QP_OPTPAR_Q_KEY;
3722 	case IB_QP_AV:
3723 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
3724 			MLX5_QP_OPTPAR_PRI_PORT;
3725 	case IB_QP_PATH_MTU:
3726 		return 0;
3727 	case IB_QP_TIMEOUT:
3728 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
3729 	case IB_QP_RETRY_CNT:
3730 		return MLX5_QP_OPTPAR_RETRY_COUNT;
3731 	case IB_QP_RNR_RETRY:
3732 		return MLX5_QP_OPTPAR_RNR_RETRY;
3733 	case IB_QP_RQ_PSN:
3734 		return 0;
3735 	case IB_QP_MAX_QP_RD_ATOMIC:
3736 		return MLX5_QP_OPTPAR_SRA_MAX;
3737 	case IB_QP_ALT_PATH:
3738 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
3739 	case IB_QP_MIN_RNR_TIMER:
3740 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
3741 	case IB_QP_SQ_PSN:
3742 		return 0;
3743 	case IB_QP_MAX_DEST_RD_ATOMIC:
3744 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
3745 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
3746 	case IB_QP_PATH_MIG_STATE:
3747 		return MLX5_QP_OPTPAR_PM_STATE;
3748 	case IB_QP_CAP:
3749 		return 0;
3750 	case IB_QP_DEST_QPN:
3751 		return 0;
3752 	}
3753 	return 0;
3754 }
3755 
3756 static int ib_mask_to_mlx5_opt(int ib_mask)
3757 {
3758 	int result = 0;
3759 	int i;
3760 
3761 	for (i = 0; i < 8 * sizeof(int); i++) {
3762 		if ((1 << i) & ib_mask)
3763 			result |= ib_nr_to_mlx5_nr(1 << i);
3764 	}
3765 
3766 	return result;
3767 }
3768 
3769 static int modify_raw_packet_qp_rq(
3770 	struct mlx5_ib_dev *dev, struct mlx5_ib_rq *rq, int new_state,
3771 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3772 {
3773 	void *in;
3774 	void *rqc;
3775 	int inlen;
3776 	int err;
3777 
3778 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
3779 	in = kvzalloc(inlen, GFP_KERNEL);
3780 	if (!in)
3781 		return -ENOMEM;
3782 
3783 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
3784 	MLX5_SET(modify_rq_in, in, uid, to_mpd(pd)->uid);
3785 
3786 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
3787 	MLX5_SET(rqc, rqc, state, new_state);
3788 
3789 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
3790 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
3791 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
3792 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
3793 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
3794 		} else
3795 			dev_info_once(
3796 				&dev->ib_dev.dev,
3797 				"RAW PACKET QP counters are not supported on current FW\n");
3798 	}
3799 
3800 	err = mlx5_core_modify_rq(dev->mdev, rq->base.mqp.qpn, in);
3801 	if (err)
3802 		goto out;
3803 
3804 	rq->state = new_state;
3805 
3806 out:
3807 	kvfree(in);
3808 	return err;
3809 }
3810 
3811 static int modify_raw_packet_qp_sq(
3812 	struct mlx5_core_dev *dev, struct mlx5_ib_sq *sq, int new_state,
3813 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3814 {
3815 	struct mlx5_ib_qp *ibqp = sq->base.container_mibqp;
3816 	struct mlx5_rate_limit old_rl = ibqp->rl;
3817 	struct mlx5_rate_limit new_rl = old_rl;
3818 	bool new_rate_added = false;
3819 	u16 rl_index = 0;
3820 	void *in;
3821 	void *sqc;
3822 	int inlen;
3823 	int err;
3824 
3825 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
3826 	in = kvzalloc(inlen, GFP_KERNEL);
3827 	if (!in)
3828 		return -ENOMEM;
3829 
3830 	MLX5_SET(modify_sq_in, in, uid, to_mpd(pd)->uid);
3831 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
3832 
3833 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
3834 	MLX5_SET(sqc, sqc, state, new_state);
3835 
3836 	if (raw_qp_param->set_mask & MLX5_RAW_QP_RATE_LIMIT) {
3837 		if (new_state != MLX5_SQC_STATE_RDY)
3838 			pr_warn("%s: Rate limit can only be changed when SQ is moving to RDY\n",
3839 				__func__);
3840 		else
3841 			new_rl = raw_qp_param->rl;
3842 	}
3843 
3844 	if (!mlx5_rl_are_equal(&old_rl, &new_rl)) {
3845 		if (new_rl.rate) {
3846 			err = mlx5_rl_add_rate(dev, &rl_index, &new_rl);
3847 			if (err) {
3848 				pr_err("Failed configuring rate limit(err %d): \
3849 				       rate %u, max_burst_sz %u, typical_pkt_sz %u\n",
3850 				       err, new_rl.rate, new_rl.max_burst_sz,
3851 				       new_rl.typical_pkt_sz);
3852 
3853 				goto out;
3854 			}
3855 			new_rate_added = true;
3856 		}
3857 
3858 		MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
3859 		/* index 0 means no limit */
3860 		MLX5_SET(sqc, sqc, packet_pacing_rate_limit_index, rl_index);
3861 	}
3862 
3863 	err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in);
3864 	if (err) {
3865 		/* Remove new rate from table if failed */
3866 		if (new_rate_added)
3867 			mlx5_rl_remove_rate(dev, &new_rl);
3868 		goto out;
3869 	}
3870 
3871 	/* Only remove the old rate after new rate was set */
3872 	if ((old_rl.rate && !mlx5_rl_are_equal(&old_rl, &new_rl)) ||
3873 	    (new_state != MLX5_SQC_STATE_RDY)) {
3874 		mlx5_rl_remove_rate(dev, &old_rl);
3875 		if (new_state != MLX5_SQC_STATE_RDY)
3876 			memset(&new_rl, 0, sizeof(new_rl));
3877 	}
3878 
3879 	ibqp->rl = new_rl;
3880 	sq->state = new_state;
3881 
3882 out:
3883 	kvfree(in);
3884 	return err;
3885 }
3886 
3887 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3888 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
3889 				u8 tx_affinity)
3890 {
3891 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
3892 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
3893 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
3894 	int modify_rq = !!qp->rq.wqe_cnt;
3895 	int modify_sq = !!qp->sq.wqe_cnt;
3896 	int rq_state;
3897 	int sq_state;
3898 	int err;
3899 
3900 	switch (raw_qp_param->operation) {
3901 	case MLX5_CMD_OP_RST2INIT_QP:
3902 		rq_state = MLX5_RQC_STATE_RDY;
3903 		sq_state = MLX5_SQC_STATE_RST;
3904 		break;
3905 	case MLX5_CMD_OP_2ERR_QP:
3906 		rq_state = MLX5_RQC_STATE_ERR;
3907 		sq_state = MLX5_SQC_STATE_ERR;
3908 		break;
3909 	case MLX5_CMD_OP_2RST_QP:
3910 		rq_state = MLX5_RQC_STATE_RST;
3911 		sq_state = MLX5_SQC_STATE_RST;
3912 		break;
3913 	case MLX5_CMD_OP_RTR2RTS_QP:
3914 	case MLX5_CMD_OP_RTS2RTS_QP:
3915 		if (raw_qp_param->set_mask & ~MLX5_RAW_QP_RATE_LIMIT)
3916 			return -EINVAL;
3917 
3918 		modify_rq = 0;
3919 		sq_state = MLX5_SQC_STATE_RDY;
3920 		break;
3921 	case MLX5_CMD_OP_INIT2INIT_QP:
3922 	case MLX5_CMD_OP_INIT2RTR_QP:
3923 		if (raw_qp_param->set_mask)
3924 			return -EINVAL;
3925 		else
3926 			return 0;
3927 	default:
3928 		WARN_ON(1);
3929 		return -EINVAL;
3930 	}
3931 
3932 	if (modify_rq) {
3933 		err =  modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param,
3934 					       qp->ibqp.pd);
3935 		if (err)
3936 			return err;
3937 	}
3938 
3939 	if (modify_sq) {
3940 		struct mlx5_flow_handle *flow_rule;
3941 
3942 		if (tx_affinity) {
3943 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
3944 							    tx_affinity,
3945 							    qp->ibqp.pd);
3946 			if (err)
3947 				return err;
3948 		}
3949 
3950 		flow_rule = create_flow_rule_vport_sq(dev, sq,
3951 						      raw_qp_param->port);
3952 		if (IS_ERR(flow_rule))
3953 			return PTR_ERR(flow_rule);
3954 
3955 		err = modify_raw_packet_qp_sq(dev->mdev, sq, sq_state,
3956 					      raw_qp_param, qp->ibqp.pd);
3957 		if (err) {
3958 			if (flow_rule)
3959 				mlx5_del_flow_rules(flow_rule);
3960 			return err;
3961 		}
3962 
3963 		if (flow_rule) {
3964 			destroy_flow_rule_vport_sq(sq);
3965 			sq->flow_rule = flow_rule;
3966 		}
3967 
3968 		return err;
3969 	}
3970 
3971 	return 0;
3972 }
3973 
3974 static unsigned int get_tx_affinity_rr(struct mlx5_ib_dev *dev,
3975 				       struct ib_udata *udata)
3976 {
3977 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3978 		udata, struct mlx5_ib_ucontext, ibucontext);
3979 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
3980 	atomic_t *tx_port_affinity;
3981 
3982 	if (ucontext)
3983 		tx_port_affinity = &ucontext->tx_port_affinity;
3984 	else
3985 		tx_port_affinity = &dev->port[port_num].roce.tx_port_affinity;
3986 
3987 	return (unsigned int)atomic_add_return(1, tx_port_affinity) %
3988 		(dev->lag_active ? dev->lag_ports : MLX5_CAP_GEN(dev->mdev, num_lag_ports)) + 1;
3989 }
3990 
3991 static bool qp_supports_affinity(struct mlx5_ib_qp *qp)
3992 {
3993 	if ((qp->type == IB_QPT_RC) || (qp->type == IB_QPT_UD) ||
3994 	    (qp->type == IB_QPT_UC) || (qp->type == IB_QPT_RAW_PACKET) ||
3995 	    (qp->type == IB_QPT_XRC_INI) || (qp->type == IB_QPT_XRC_TGT) ||
3996 	    (qp->type == MLX5_IB_QPT_DCI))
3997 		return true;
3998 	return false;
3999 }
4000 
4001 static unsigned int get_tx_affinity(struct ib_qp *qp,
4002 				    const struct ib_qp_attr *attr,
4003 				    int attr_mask, u8 init,
4004 				    struct ib_udata *udata)
4005 {
4006 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
4007 		udata, struct mlx5_ib_ucontext, ibucontext);
4008 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
4009 	struct mlx5_ib_qp *mqp = to_mqp(qp);
4010 	struct mlx5_ib_qp_base *qp_base;
4011 	unsigned int tx_affinity;
4012 
4013 	if (!(mlx5_ib_lag_should_assign_affinity(dev) &&
4014 	      qp_supports_affinity(mqp)))
4015 		return 0;
4016 
4017 	if (mqp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
4018 		tx_affinity = mqp->gsi_lag_port;
4019 	else if (init)
4020 		tx_affinity = get_tx_affinity_rr(dev, udata);
4021 	else if ((attr_mask & IB_QP_AV) && attr->xmit_slave)
4022 		tx_affinity =
4023 			mlx5_lag_get_slave_port(dev->mdev, attr->xmit_slave);
4024 	else
4025 		return 0;
4026 
4027 	qp_base = &mqp->trans_qp.base;
4028 	if (ucontext)
4029 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x ucontext %p\n",
4030 			    tx_affinity, qp_base->mqp.qpn, ucontext);
4031 	else
4032 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x\n",
4033 			    tx_affinity, qp_base->mqp.qpn);
4034 	return tx_affinity;
4035 }
4036 
4037 static int __mlx5_ib_qp_set_counter(struct ib_qp *qp,
4038 				    struct rdma_counter *counter)
4039 {
4040 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
4041 	u32 in[MLX5_ST_SZ_DW(rts2rts_qp_in)] = {};
4042 	struct mlx5_ib_qp *mqp = to_mqp(qp);
4043 	struct mlx5_ib_qp_base *base;
4044 	u32 set_id;
4045 	u32 *qpc;
4046 
4047 	if (counter)
4048 		set_id = counter->id;
4049 	else
4050 		set_id = mlx5_ib_get_counters_id(dev, mqp->port - 1);
4051 
4052 	base = &mqp->trans_qp.base;
4053 	MLX5_SET(rts2rts_qp_in, in, opcode, MLX5_CMD_OP_RTS2RTS_QP);
4054 	MLX5_SET(rts2rts_qp_in, in, qpn, base->mqp.qpn);
4055 	MLX5_SET(rts2rts_qp_in, in, uid, base->mqp.uid);
4056 	MLX5_SET(rts2rts_qp_in, in, opt_param_mask,
4057 		 MLX5_QP_OPTPAR_COUNTER_SET_ID);
4058 
4059 	qpc = MLX5_ADDR_OF(rts2rts_qp_in, in, qpc);
4060 	MLX5_SET(qpc, qpc, counter_set_id, set_id);
4061 	return mlx5_cmd_exec_in(dev->mdev, rts2rts_qp, in);
4062 }
4063 
4064 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
4065 			       const struct ib_qp_attr *attr, int attr_mask,
4066 			       enum ib_qp_state cur_state,
4067 			       enum ib_qp_state new_state,
4068 			       const struct mlx5_ib_modify_qp *ucmd,
4069 			       struct mlx5_ib_modify_qp_resp *resp,
4070 			       struct ib_udata *udata)
4071 {
4072 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
4073 		[MLX5_QP_STATE_RST] = {
4074 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4075 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4076 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
4077 		},
4078 		[MLX5_QP_STATE_INIT]  = {
4079 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4080 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4081 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
4082 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
4083 		},
4084 		[MLX5_QP_STATE_RTR]   = {
4085 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4086 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4087 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
4088 		},
4089 		[MLX5_QP_STATE_RTS]   = {
4090 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4091 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4092 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
4093 		},
4094 		[MLX5_QP_STATE_SQD] = {
4095 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4096 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4097 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQD_RTS_QP,
4098 		},
4099 		[MLX5_QP_STATE_SQER] = {
4100 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4101 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4102 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
4103 		},
4104 		[MLX5_QP_STATE_ERR] = {
4105 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
4106 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
4107 		}
4108 	};
4109 
4110 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4111 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4112 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
4113 	struct mlx5_ib_cq *send_cq, *recv_cq;
4114 	struct mlx5_ib_pd *pd;
4115 	enum mlx5_qp_state mlx5_cur, mlx5_new;
4116 	void *qpc, *pri_path, *alt_path;
4117 	enum mlx5_qp_optpar optpar = 0;
4118 	u32 set_id = 0;
4119 	int mlx5_st;
4120 	int err;
4121 	u16 op;
4122 	u8 tx_affinity = 0;
4123 
4124 	mlx5_st = to_mlx5_st(qp->type);
4125 	if (mlx5_st < 0)
4126 		return -EINVAL;
4127 
4128 	qpc = kzalloc(MLX5_ST_SZ_BYTES(qpc), GFP_KERNEL);
4129 	if (!qpc)
4130 		return -ENOMEM;
4131 
4132 	pd = to_mpd(qp->ibqp.pd);
4133 	MLX5_SET(qpc, qpc, st, mlx5_st);
4134 
4135 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
4136 		MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
4137 	} else {
4138 		switch (attr->path_mig_state) {
4139 		case IB_MIG_MIGRATED:
4140 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
4141 			break;
4142 		case IB_MIG_REARM:
4143 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_REARM);
4144 			break;
4145 		case IB_MIG_ARMED:
4146 			MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_ARMED);
4147 			break;
4148 		}
4149 	}
4150 
4151 	tx_affinity = get_tx_affinity(ibqp, attr, attr_mask,
4152 				      cur_state == IB_QPS_RESET &&
4153 				      new_state == IB_QPS_INIT, udata);
4154 
4155 	MLX5_SET(qpc, qpc, lag_tx_port_affinity, tx_affinity);
4156 	if (tx_affinity && new_state == IB_QPS_RTR &&
4157 	    MLX5_CAP_GEN(dev->mdev, init2_lag_tx_port_affinity))
4158 		optpar |= MLX5_QP_OPTPAR_LAG_TX_AFF;
4159 
4160 	if (is_sqp(qp->type)) {
4161 		MLX5_SET(qpc, qpc, mtu, IB_MTU_256);
4162 		MLX5_SET(qpc, qpc, log_msg_max, 8);
4163 	} else if ((qp->type == IB_QPT_UD &&
4164 		    !(qp->flags & IB_QP_CREATE_SOURCE_QPN)) ||
4165 		   qp->type == MLX5_IB_QPT_REG_UMR) {
4166 		MLX5_SET(qpc, qpc, mtu, IB_MTU_4096);
4167 		MLX5_SET(qpc, qpc, log_msg_max, 12);
4168 	} else if (attr_mask & IB_QP_PATH_MTU) {
4169 		if (attr->path_mtu < IB_MTU_256 ||
4170 		    attr->path_mtu > IB_MTU_4096) {
4171 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
4172 			err = -EINVAL;
4173 			goto out;
4174 		}
4175 		MLX5_SET(qpc, qpc, mtu, attr->path_mtu);
4176 		MLX5_SET(qpc, qpc, log_msg_max,
4177 			 MLX5_CAP_GEN(dev->mdev, log_max_msg));
4178 	}
4179 
4180 	if (attr_mask & IB_QP_DEST_QPN)
4181 		MLX5_SET(qpc, qpc, remote_qpn, attr->dest_qp_num);
4182 
4183 	pri_path = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
4184 	alt_path = MLX5_ADDR_OF(qpc, qpc, secondary_address_path);
4185 
4186 	if (attr_mask & IB_QP_PKEY_INDEX)
4187 		MLX5_SET(ads, pri_path, pkey_index, attr->pkey_index);
4188 
4189 	/* todo implement counter_index functionality */
4190 
4191 	if (is_sqp(qp->type))
4192 		MLX5_SET(ads, pri_path, vhca_port_num, qp->port);
4193 
4194 	if (attr_mask & IB_QP_PORT)
4195 		MLX5_SET(ads, pri_path, vhca_port_num, attr->port_num);
4196 
4197 	if (attr_mask & IB_QP_AV) {
4198 		err = mlx5_set_path(dev, qp, &attr->ah_attr, pri_path,
4199 				    attr_mask & IB_QP_PORT ? attr->port_num :
4200 							     qp->port,
4201 				    attr_mask, 0, attr, false);
4202 		if (err)
4203 			goto out;
4204 	}
4205 
4206 	if (attr_mask & IB_QP_TIMEOUT)
4207 		MLX5_SET(ads, pri_path, ack_timeout, attr->timeout);
4208 
4209 	if (attr_mask & IB_QP_ALT_PATH) {
4210 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr, alt_path,
4211 				    attr->alt_port_num,
4212 				    attr_mask | IB_QP_PKEY_INDEX |
4213 					    IB_QP_TIMEOUT,
4214 				    0, attr, true);
4215 		if (err)
4216 			goto out;
4217 	}
4218 
4219 	get_cqs(qp->type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
4220 		&send_cq, &recv_cq);
4221 
4222 	MLX5_SET(qpc, qpc, pd, pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
4223 	if (send_cq)
4224 		MLX5_SET(qpc, qpc, cqn_snd, send_cq->mcq.cqn);
4225 	if (recv_cq)
4226 		MLX5_SET(qpc, qpc, cqn_rcv, recv_cq->mcq.cqn);
4227 
4228 	MLX5_SET(qpc, qpc, log_ack_req_freq, MLX5_IB_ACK_REQ_FREQ);
4229 
4230 	if (attr_mask & IB_QP_RNR_RETRY)
4231 		MLX5_SET(qpc, qpc, rnr_retry, attr->rnr_retry);
4232 
4233 	if (attr_mask & IB_QP_RETRY_CNT)
4234 		MLX5_SET(qpc, qpc, retry_count, attr->retry_cnt);
4235 
4236 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC && attr->max_rd_atomic)
4237 		MLX5_SET(qpc, qpc, log_sra_max, ilog2(attr->max_rd_atomic));
4238 
4239 	if (attr_mask & IB_QP_SQ_PSN)
4240 		MLX5_SET(qpc, qpc, next_send_psn, attr->sq_psn);
4241 
4242 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC && attr->max_dest_rd_atomic)
4243 		MLX5_SET(qpc, qpc, log_rra_max,
4244 			 ilog2(attr->max_dest_rd_atomic));
4245 
4246 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
4247 		err = set_qpc_atomic_flags(qp, attr, attr_mask, qpc);
4248 		if (err)
4249 			goto out;
4250 	}
4251 
4252 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
4253 		MLX5_SET(qpc, qpc, min_rnr_nak, attr->min_rnr_timer);
4254 
4255 	if (attr_mask & IB_QP_RQ_PSN)
4256 		MLX5_SET(qpc, qpc, next_rcv_psn, attr->rq_psn);
4257 
4258 	if (attr_mask & IB_QP_QKEY)
4259 		MLX5_SET(qpc, qpc, q_key, attr->qkey);
4260 
4261 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
4262 		MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
4263 
4264 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4265 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
4266 			       qp->port) - 1;
4267 
4268 		/* Underlay port should be used - index 0 function per port */
4269 		if (qp->flags & IB_QP_CREATE_SOURCE_QPN)
4270 			port_num = 0;
4271 
4272 		if (ibqp->counter)
4273 			set_id = ibqp->counter->id;
4274 		else
4275 			set_id = mlx5_ib_get_counters_id(dev, port_num);
4276 		MLX5_SET(qpc, qpc, counter_set_id, set_id);
4277 	}
4278 
4279 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
4280 		MLX5_SET(qpc, qpc, rlky, 1);
4281 
4282 	if (qp->flags & MLX5_IB_QP_CREATE_SQPN_QP1)
4283 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
4284 
4285 	mlx5_cur = to_mlx5_state(cur_state);
4286 	mlx5_new = to_mlx5_state(new_state);
4287 
4288 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
4289 	    !optab[mlx5_cur][mlx5_new]) {
4290 		err = -EINVAL;
4291 		goto out;
4292 	}
4293 
4294 	op = optab[mlx5_cur][mlx5_new];
4295 	optpar |= ib_mask_to_mlx5_opt(attr_mask);
4296 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
4297 
4298 	if (qp->type == IB_QPT_RAW_PACKET ||
4299 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4300 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
4301 
4302 		raw_qp_param.operation = op;
4303 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4304 			raw_qp_param.rq_q_ctr_id = set_id;
4305 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
4306 		}
4307 
4308 		if (attr_mask & IB_QP_PORT)
4309 			raw_qp_param.port = attr->port_num;
4310 
4311 		if (attr_mask & IB_QP_RATE_LIMIT) {
4312 			raw_qp_param.rl.rate = attr->rate_limit;
4313 
4314 			if (ucmd->burst_info.max_burst_sz) {
4315 				if (attr->rate_limit &&
4316 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_burst_bound)) {
4317 					raw_qp_param.rl.max_burst_sz =
4318 						ucmd->burst_info.max_burst_sz;
4319 				} else {
4320 					err = -EINVAL;
4321 					goto out;
4322 				}
4323 			}
4324 
4325 			if (ucmd->burst_info.typical_pkt_sz) {
4326 				if (attr->rate_limit &&
4327 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_typical_size)) {
4328 					raw_qp_param.rl.typical_pkt_sz =
4329 						ucmd->burst_info.typical_pkt_sz;
4330 				} else {
4331 					err = -EINVAL;
4332 					goto out;
4333 				}
4334 			}
4335 
4336 			raw_qp_param.set_mask |= MLX5_RAW_QP_RATE_LIMIT;
4337 		}
4338 
4339 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, tx_affinity);
4340 	} else {
4341 		if (udata) {
4342 			/* For the kernel flows, the resp will stay zero */
4343 			resp->ece_options =
4344 				MLX5_CAP_GEN(dev->mdev, ece_support) ?
4345 					ucmd->ece_options : 0;
4346 			resp->response_length = sizeof(*resp);
4347 		}
4348 		err = mlx5_core_qp_modify(dev, op, optpar, qpc, &base->mqp,
4349 					  &resp->ece_options);
4350 	}
4351 
4352 	if (err)
4353 		goto out;
4354 
4355 	qp->state = new_state;
4356 
4357 	if (attr_mask & IB_QP_ACCESS_FLAGS)
4358 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
4359 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
4360 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
4361 	if (attr_mask & IB_QP_PORT)
4362 		qp->port = attr->port_num;
4363 	if (attr_mask & IB_QP_ALT_PATH)
4364 		qp->trans_qp.alt_port = attr->alt_port_num;
4365 
4366 	/*
4367 	 * If we moved a kernel QP to RESET, clean up all old CQ
4368 	 * entries and reinitialize the QP.
4369 	 */
4370 	if (new_state == IB_QPS_RESET &&
4371 	    !ibqp->uobject && qp->type != IB_QPT_XRC_TGT) {
4372 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
4373 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
4374 		if (send_cq != recv_cq)
4375 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
4376 
4377 		qp->rq.head = 0;
4378 		qp->rq.tail = 0;
4379 		qp->sq.head = 0;
4380 		qp->sq.tail = 0;
4381 		qp->sq.cur_post = 0;
4382 		if (qp->sq.wqe_cnt)
4383 			qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
4384 		qp->sq.last_poll = 0;
4385 		qp->db.db[MLX5_RCV_DBR] = 0;
4386 		qp->db.db[MLX5_SND_DBR] = 0;
4387 	}
4388 
4389 	if ((new_state == IB_QPS_RTS) && qp->counter_pending) {
4390 		err = __mlx5_ib_qp_set_counter(ibqp, ibqp->counter);
4391 		if (!err)
4392 			qp->counter_pending = 0;
4393 	}
4394 
4395 out:
4396 	kfree(qpc);
4397 	return err;
4398 }
4399 
4400 static inline bool is_valid_mask(int mask, int req, int opt)
4401 {
4402 	if ((mask & req) != req)
4403 		return false;
4404 
4405 	if (mask & ~(req | opt))
4406 		return false;
4407 
4408 	return true;
4409 }
4410 
4411 /* check valid transition for driver QP types
4412  * for now the only QP type that this function supports is DCI
4413  */
4414 static bool modify_dci_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state new_state,
4415 				enum ib_qp_attr_mask attr_mask)
4416 {
4417 	int req = IB_QP_STATE;
4418 	int opt = 0;
4419 
4420 	if (new_state == IB_QPS_RESET) {
4421 		return is_valid_mask(attr_mask, req, opt);
4422 	} else if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4423 		req |= IB_QP_PKEY_INDEX | IB_QP_PORT;
4424 		return is_valid_mask(attr_mask, req, opt);
4425 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
4426 		opt = IB_QP_PKEY_INDEX | IB_QP_PORT;
4427 		return is_valid_mask(attr_mask, req, opt);
4428 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
4429 		req |= IB_QP_PATH_MTU;
4430 		opt = IB_QP_PKEY_INDEX | IB_QP_AV;
4431 		return is_valid_mask(attr_mask, req, opt);
4432 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
4433 		req |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4434 		       IB_QP_MAX_QP_RD_ATOMIC | IB_QP_SQ_PSN;
4435 		opt = IB_QP_MIN_RNR_TIMER;
4436 		return is_valid_mask(attr_mask, req, opt);
4437 	} else if (cur_state == IB_QPS_RTS && new_state == IB_QPS_RTS) {
4438 		opt = IB_QP_MIN_RNR_TIMER;
4439 		return is_valid_mask(attr_mask, req, opt);
4440 	} else if (cur_state != IB_QPS_RESET && new_state == IB_QPS_ERR) {
4441 		return is_valid_mask(attr_mask, req, opt);
4442 	}
4443 	return false;
4444 }
4445 
4446 /* mlx5_ib_modify_dct: modify a DCT QP
4447  * valid transitions are:
4448  * RESET to INIT: must set access_flags, pkey_index and port
4449  * INIT  to RTR : must set min_rnr_timer, tclass, flow_label,
4450  *			   mtu, gid_index and hop_limit
4451  * Other transitions and attributes are illegal
4452  */
4453 static int mlx5_ib_modify_dct(struct ib_qp *ibqp, struct ib_qp_attr *attr,
4454 			      int attr_mask, struct mlx5_ib_modify_qp *ucmd,
4455 			      struct ib_udata *udata)
4456 {
4457 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4458 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4459 	enum ib_qp_state cur_state, new_state;
4460 	int required = IB_QP_STATE;
4461 	void *dctc;
4462 	int err;
4463 
4464 	if (!(attr_mask & IB_QP_STATE))
4465 		return -EINVAL;
4466 
4467 	cur_state = qp->state;
4468 	new_state = attr->qp_state;
4469 
4470 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
4471 	if (MLX5_CAP_GEN(dev->mdev, ece_support) && ucmd->ece_options)
4472 		/*
4473 		 * DCT doesn't initialize QP till modify command is executed,
4474 		 * so we need to overwrite previously set ECE field if user
4475 		 * provided any value except zero, which means not set/not
4476 		 * valid.
4477 		 */
4478 		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
4479 
4480 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
4481 		u16 set_id;
4482 
4483 		required |= IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
4484 		if (!is_valid_mask(attr_mask, required, 0))
4485 			return -EINVAL;
4486 
4487 		if (attr->port_num == 0 ||
4488 		    attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)) {
4489 			mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4490 				    attr->port_num, dev->num_ports);
4491 			return -EINVAL;
4492 		}
4493 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
4494 			MLX5_SET(dctc, dctc, rre, 1);
4495 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
4496 			MLX5_SET(dctc, dctc, rwe, 1);
4497 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC) {
4498 			int atomic_mode;
4499 
4500 			atomic_mode = get_atomic_mode(dev, MLX5_IB_QPT_DCT);
4501 			if (atomic_mode < 0)
4502 				return -EOPNOTSUPP;
4503 
4504 			MLX5_SET(dctc, dctc, atomic_mode, atomic_mode);
4505 			MLX5_SET(dctc, dctc, rae, 1);
4506 		}
4507 		MLX5_SET(dctc, dctc, pkey_index, attr->pkey_index);
4508 		if (mlx5_lag_is_active(dev->mdev))
4509 			MLX5_SET(dctc, dctc, port,
4510 				 get_tx_affinity_rr(dev, udata));
4511 		else
4512 			MLX5_SET(dctc, dctc, port, attr->port_num);
4513 
4514 		set_id = mlx5_ib_get_counters_id(dev, attr->port_num - 1);
4515 		MLX5_SET(dctc, dctc, counter_set_id, set_id);
4516 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
4517 		struct mlx5_ib_modify_qp_resp resp = {};
4518 		u32 out[MLX5_ST_SZ_DW(create_dct_out)] = {};
4519 		u32 min_resp_len = offsetofend(typeof(resp), dctn);
4520 
4521 		if (udata->outlen < min_resp_len)
4522 			return -EINVAL;
4523 		/*
4524 		 * If we don't have enough space for the ECE options,
4525 		 * simply indicate it with resp.response_length.
4526 		 */
4527 		resp.response_length = (udata->outlen < sizeof(resp)) ?
4528 					       min_resp_len :
4529 					       sizeof(resp);
4530 
4531 		required |= IB_QP_MIN_RNR_TIMER | IB_QP_AV | IB_QP_PATH_MTU;
4532 		if (!is_valid_mask(attr_mask, required, 0))
4533 			return -EINVAL;
4534 		MLX5_SET(dctc, dctc, min_rnr_nak, attr->min_rnr_timer);
4535 		MLX5_SET(dctc, dctc, tclass, attr->ah_attr.grh.traffic_class);
4536 		MLX5_SET(dctc, dctc, flow_label, attr->ah_attr.grh.flow_label);
4537 		MLX5_SET(dctc, dctc, mtu, attr->path_mtu);
4538 		MLX5_SET(dctc, dctc, my_addr_index, attr->ah_attr.grh.sgid_index);
4539 		MLX5_SET(dctc, dctc, hop_limit, attr->ah_attr.grh.hop_limit);
4540 		if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
4541 			MLX5_SET(dctc, dctc, eth_prio, attr->ah_attr.sl & 0x7);
4542 
4543 		err = mlx5_core_create_dct(dev, &qp->dct.mdct, qp->dct.in,
4544 					   MLX5_ST_SZ_BYTES(create_dct_in), out,
4545 					   sizeof(out));
4546 		err = mlx5_cmd_check(dev->mdev, err, qp->dct.in, out);
4547 		if (err)
4548 			return err;
4549 		resp.dctn = qp->dct.mdct.mqp.qpn;
4550 		if (MLX5_CAP_GEN(dev->mdev, ece_support))
4551 			resp.ece_options = MLX5_GET(create_dct_out, out, ece);
4552 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4553 		if (err) {
4554 			mlx5_core_destroy_dct(dev, &qp->dct.mdct);
4555 			return err;
4556 		}
4557 	} else {
4558 		mlx5_ib_warn(dev, "Modify DCT: Invalid transition from %d to %d\n", cur_state, new_state);
4559 		return -EINVAL;
4560 	}
4561 
4562 	qp->state = new_state;
4563 	return 0;
4564 }
4565 
4566 static bool mlx5_ib_modify_qp_allowed(struct mlx5_ib_dev *dev,
4567 				      struct mlx5_ib_qp *qp)
4568 {
4569 	if (dev->profile != &raw_eth_profile)
4570 		return true;
4571 
4572 	if (qp->type == IB_QPT_RAW_PACKET || qp->type == MLX5_IB_QPT_REG_UMR)
4573 		return true;
4574 
4575 	/* Internal QP used for wc testing, with NOPs in wq */
4576 	if (qp->flags & MLX5_IB_QP_CREATE_WC_TEST)
4577 		return true;
4578 
4579 	return false;
4580 }
4581 
4582 static int validate_rd_atomic(struct mlx5_ib_dev *dev, struct ib_qp_attr *attr,
4583 			      int attr_mask, enum ib_qp_type qp_type)
4584 {
4585 	int log_max_ra_res;
4586 	int log_max_ra_req;
4587 
4588 	if (qp_type == MLX5_IB_QPT_DCI) {
4589 		log_max_ra_res = 1 << MLX5_CAP_GEN(dev->mdev,
4590 						   log_max_ra_res_dc);
4591 		log_max_ra_req = 1 << MLX5_CAP_GEN(dev->mdev,
4592 						   log_max_ra_req_dc);
4593 	} else {
4594 		log_max_ra_res = 1 << MLX5_CAP_GEN(dev->mdev,
4595 						   log_max_ra_res_qp);
4596 		log_max_ra_req = 1 << MLX5_CAP_GEN(dev->mdev,
4597 						   log_max_ra_req_qp);
4598 	}
4599 
4600 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
4601 	    attr->max_rd_atomic > log_max_ra_res) {
4602 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
4603 			    attr->max_rd_atomic);
4604 		return false;
4605 	}
4606 
4607 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
4608 	    attr->max_dest_rd_atomic > log_max_ra_req) {
4609 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
4610 			    attr->max_dest_rd_atomic);
4611 		return false;
4612 	}
4613 	return true;
4614 }
4615 
4616 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
4617 		      int attr_mask, struct ib_udata *udata)
4618 {
4619 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4620 	struct mlx5_ib_modify_qp_resp resp = {};
4621 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
4622 	struct mlx5_ib_modify_qp ucmd = {};
4623 	enum ib_qp_type qp_type;
4624 	enum ib_qp_state cur_state, new_state;
4625 	int err = -EINVAL;
4626 
4627 	if (!mlx5_ib_modify_qp_allowed(dev, qp))
4628 		return -EOPNOTSUPP;
4629 
4630 	if (attr_mask & ~(IB_QP_ATTR_STANDARD_BITS | IB_QP_RATE_LIMIT))
4631 		return -EOPNOTSUPP;
4632 
4633 	if (ibqp->rwq_ind_tbl)
4634 		return -ENOSYS;
4635 
4636 	if (udata && udata->inlen) {
4637 		if (udata->inlen < offsetofend(typeof(ucmd), ece_options))
4638 			return -EINVAL;
4639 
4640 		if (udata->inlen > sizeof(ucmd) &&
4641 		    !ib_is_udata_cleared(udata, sizeof(ucmd),
4642 					 udata->inlen - sizeof(ucmd)))
4643 			return -EOPNOTSUPP;
4644 
4645 		if (ib_copy_from_udata(&ucmd, udata,
4646 				       min(udata->inlen, sizeof(ucmd))))
4647 			return -EFAULT;
4648 
4649 		if (ucmd.comp_mask ||
4650 		    memchr_inv(&ucmd.burst_info.reserved, 0,
4651 			       sizeof(ucmd.burst_info.reserved)))
4652 			return -EOPNOTSUPP;
4653 
4654 	}
4655 
4656 	if (qp->type == IB_QPT_GSI)
4657 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
4658 
4659 	qp_type = (qp->type == MLX5_IB_QPT_HW_GSI) ? IB_QPT_GSI : qp->type;
4660 
4661 	if (qp_type == MLX5_IB_QPT_DCT)
4662 		return mlx5_ib_modify_dct(ibqp, attr, attr_mask, &ucmd, udata);
4663 
4664 	mutex_lock(&qp->mutex);
4665 
4666 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
4667 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
4668 
4669 	if (qp->flags & IB_QP_CREATE_SOURCE_QPN) {
4670 		if (attr_mask & ~(IB_QP_STATE | IB_QP_CUR_STATE)) {
4671 			mlx5_ib_dbg(dev, "invalid attr_mask 0x%x when underlay QP is used\n",
4672 				    attr_mask);
4673 			goto out;
4674 		}
4675 	} else if (qp_type != MLX5_IB_QPT_REG_UMR &&
4676 		   qp_type != MLX5_IB_QPT_DCI &&
4677 		   !ib_modify_qp_is_ok(cur_state, new_state, qp_type,
4678 				       attr_mask)) {
4679 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4680 			    cur_state, new_state, qp->type, attr_mask);
4681 		goto out;
4682 	} else if (qp_type == MLX5_IB_QPT_DCI &&
4683 		   !modify_dci_qp_is_ok(cur_state, new_state, attr_mask)) {
4684 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4685 			    cur_state, new_state, qp_type, attr_mask);
4686 		goto out;
4687 	}
4688 
4689 	if ((attr_mask & IB_QP_PORT) &&
4690 	    (attr->port_num == 0 ||
4691 	     attr->port_num > dev->num_ports)) {
4692 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4693 			    attr->port_num, dev->num_ports);
4694 		goto out;
4695 	}
4696 
4697 	if ((attr_mask & IB_QP_PKEY_INDEX) &&
4698 	    attr->pkey_index >= dev->pkey_table_len) {
4699 		mlx5_ib_dbg(dev, "invalid pkey index %d\n", attr->pkey_index);
4700 		goto out;
4701 	}
4702 
4703 	if (!validate_rd_atomic(dev, attr, attr_mask, qp_type))
4704 		goto out;
4705 
4706 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
4707 		err = 0;
4708 		goto out;
4709 	}
4710 
4711 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state,
4712 				  new_state, &ucmd, &resp, udata);
4713 
4714 	/* resp.response_length is set in ECE supported flows only */
4715 	if (!err && resp.response_length &&
4716 	    udata->outlen >= resp.response_length)
4717 		/* Return -EFAULT to the user and expect him to destroy QP. */
4718 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
4719 
4720 out:
4721 	mutex_unlock(&qp->mutex);
4722 	return err;
4723 }
4724 
4725 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4726 {
4727 	switch (mlx5_state) {
4728 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
4729 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
4730 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
4731 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
4732 	case MLX5_QP_STATE_SQ_DRAINING:
4733 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
4734 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
4735 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
4736 	default:		     return -1;
4737 	}
4738 }
4739 
4740 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4741 {
4742 	switch (mlx5_mig_state) {
4743 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
4744 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
4745 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
4746 	default: return -1;
4747 	}
4748 }
4749 
4750 static void to_rdma_ah_attr(struct mlx5_ib_dev *ibdev,
4751 			    struct rdma_ah_attr *ah_attr, void *path)
4752 {
4753 	int port = MLX5_GET(ads, path, vhca_port_num);
4754 	int static_rate;
4755 
4756 	memset(ah_attr, 0, sizeof(*ah_attr));
4757 
4758 	if (!port || port > ibdev->num_ports)
4759 		return;
4760 
4761 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port);
4762 
4763 	rdma_ah_set_port_num(ah_attr, port);
4764 	rdma_ah_set_sl(ah_attr, MLX5_GET(ads, path, sl));
4765 
4766 	rdma_ah_set_dlid(ah_attr, MLX5_GET(ads, path, rlid));
4767 	rdma_ah_set_path_bits(ah_attr, MLX5_GET(ads, path, mlid));
4768 
4769 	static_rate = MLX5_GET(ads, path, stat_rate);
4770 	rdma_ah_set_static_rate(ah_attr, mlx5_to_ib_rate_map(static_rate));
4771 	if (MLX5_GET(ads, path, grh) ||
4772 	    ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
4773 		rdma_ah_set_grh(ah_attr, NULL, MLX5_GET(ads, path, flow_label),
4774 				MLX5_GET(ads, path, src_addr_index),
4775 				MLX5_GET(ads, path, hop_limit),
4776 				MLX5_GET(ads, path, tclass));
4777 		rdma_ah_set_dgid_raw(ah_attr, MLX5_ADDR_OF(ads, path, rgid_rip));
4778 	}
4779 }
4780 
4781 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4782 					struct mlx5_ib_sq *sq,
4783 					u8 *sq_state)
4784 {
4785 	int err;
4786 
4787 	err = mlx5_core_query_sq_state(dev->mdev, sq->base.mqp.qpn, sq_state);
4788 	if (err)
4789 		goto out;
4790 	sq->state = *sq_state;
4791 
4792 out:
4793 	return err;
4794 }
4795 
4796 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4797 					struct mlx5_ib_rq *rq,
4798 					u8 *rq_state)
4799 {
4800 	void *out;
4801 	void *rqc;
4802 	int inlen;
4803 	int err;
4804 
4805 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4806 	out = kvzalloc(inlen, GFP_KERNEL);
4807 	if (!out)
4808 		return -ENOMEM;
4809 
4810 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4811 	if (err)
4812 		goto out;
4813 
4814 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4815 	*rq_state = MLX5_GET(rqc, rqc, state);
4816 	rq->state = *rq_state;
4817 
4818 out:
4819 	kvfree(out);
4820 	return err;
4821 }
4822 
4823 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4824 				  struct mlx5_ib_qp *qp, u8 *qp_state)
4825 {
4826 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4827 		[MLX5_RQC_STATE_RST] = {
4828 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
4829 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4830 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
4831 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
4832 		},
4833 		[MLX5_RQC_STATE_RDY] = {
4834 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE,
4835 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4836 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
4837 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
4838 		},
4839 		[MLX5_RQC_STATE_ERR] = {
4840 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
4841 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
4842 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
4843 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
4844 		},
4845 		[MLX5_RQ_STATE_NA] = {
4846 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE,
4847 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
4848 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
4849 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
4850 		},
4851 	};
4852 
4853 	*qp_state = sqrq_trans[rq_state][sq_state];
4854 
4855 	if (*qp_state == MLX5_QP_STATE_BAD) {
4856 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4857 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4858 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4859 		return -EINVAL;
4860 	}
4861 
4862 	if (*qp_state == MLX5_QP_STATE)
4863 		*qp_state = qp->state;
4864 
4865 	return 0;
4866 }
4867 
4868 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4869 				     struct mlx5_ib_qp *qp,
4870 				     u8 *raw_packet_qp_state)
4871 {
4872 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4873 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4874 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4875 	int err;
4876 	u8 sq_state = MLX5_SQ_STATE_NA;
4877 	u8 rq_state = MLX5_RQ_STATE_NA;
4878 
4879 	if (qp->sq.wqe_cnt) {
4880 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4881 		if (err)
4882 			return err;
4883 	}
4884 
4885 	if (qp->rq.wqe_cnt) {
4886 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4887 		if (err)
4888 			return err;
4889 	}
4890 
4891 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4892 				      raw_packet_qp_state);
4893 }
4894 
4895 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4896 			 struct ib_qp_attr *qp_attr)
4897 {
4898 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
4899 	void *qpc, *pri_path, *alt_path;
4900 	u32 *outb;
4901 	int err;
4902 
4903 	outb = kzalloc(outlen, GFP_KERNEL);
4904 	if (!outb)
4905 		return -ENOMEM;
4906 
4907 	err = mlx5_core_qp_query(dev, &qp->trans_qp.base.mqp, outb, outlen,
4908 				 false);
4909 	if (err)
4910 		goto out;
4911 
4912 	qpc = MLX5_ADDR_OF(query_qp_out, outb, qpc);
4913 
4914 	qp->state = to_ib_qp_state(MLX5_GET(qpc, qpc, state));
4915 	if (MLX5_GET(qpc, qpc, state) == MLX5_QP_STATE_SQ_DRAINING)
4916 		qp_attr->sq_draining = 1;
4917 
4918 	qp_attr->path_mtu = MLX5_GET(qpc, qpc, mtu);
4919 	qp_attr->path_mig_state = to_ib_mig_state(MLX5_GET(qpc, qpc, pm_state));
4920 	qp_attr->qkey = MLX5_GET(qpc, qpc, q_key);
4921 	qp_attr->rq_psn = MLX5_GET(qpc, qpc, next_rcv_psn);
4922 	qp_attr->sq_psn = MLX5_GET(qpc, qpc, next_send_psn);
4923 	qp_attr->dest_qp_num = MLX5_GET(qpc, qpc, remote_qpn);
4924 
4925 	if (MLX5_GET(qpc, qpc, rre))
4926 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ;
4927 	if (MLX5_GET(qpc, qpc, rwe))
4928 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_WRITE;
4929 	if (MLX5_GET(qpc, qpc, rae))
4930 		qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_ATOMIC;
4931 
4932 	qp_attr->max_rd_atomic = 1 << MLX5_GET(qpc, qpc, log_sra_max);
4933 	qp_attr->max_dest_rd_atomic = 1 << MLX5_GET(qpc, qpc, log_rra_max);
4934 	qp_attr->min_rnr_timer = MLX5_GET(qpc, qpc, min_rnr_nak);
4935 	qp_attr->retry_cnt = MLX5_GET(qpc, qpc, retry_count);
4936 	qp_attr->rnr_retry = MLX5_GET(qpc, qpc, rnr_retry);
4937 
4938 	pri_path = MLX5_ADDR_OF(qpc, qpc, primary_address_path);
4939 	alt_path = MLX5_ADDR_OF(qpc, qpc, secondary_address_path);
4940 
4941 	if (qp->type == IB_QPT_RC || qp->type == IB_QPT_UC ||
4942 	    qp->type == IB_QPT_XRC_INI || qp->type == IB_QPT_XRC_TGT) {
4943 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, pri_path);
4944 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, alt_path);
4945 		qp_attr->alt_pkey_index = MLX5_GET(ads, alt_path, pkey_index);
4946 		qp_attr->alt_port_num = MLX5_GET(ads, alt_path, vhca_port_num);
4947 	}
4948 
4949 	qp_attr->pkey_index = MLX5_GET(ads, pri_path, pkey_index);
4950 	qp_attr->port_num = MLX5_GET(ads, pri_path, vhca_port_num);
4951 	qp_attr->timeout = MLX5_GET(ads, pri_path, ack_timeout);
4952 	qp_attr->alt_timeout = MLX5_GET(ads, alt_path, ack_timeout);
4953 
4954 out:
4955 	kfree(outb);
4956 	return err;
4957 }
4958 
4959 static int mlx5_ib_dct_query_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *mqp,
4960 				struct ib_qp_attr *qp_attr, int qp_attr_mask,
4961 				struct ib_qp_init_attr *qp_init_attr)
4962 {
4963 	struct mlx5_core_dct	*dct = &mqp->dct.mdct;
4964 	u32 *out;
4965 	u32 access_flags = 0;
4966 	int outlen = MLX5_ST_SZ_BYTES(query_dct_out);
4967 	void *dctc;
4968 	int err;
4969 	int supported_mask = IB_QP_STATE |
4970 			     IB_QP_ACCESS_FLAGS |
4971 			     IB_QP_PORT |
4972 			     IB_QP_MIN_RNR_TIMER |
4973 			     IB_QP_AV |
4974 			     IB_QP_PATH_MTU |
4975 			     IB_QP_PKEY_INDEX;
4976 
4977 	if (qp_attr_mask & ~supported_mask)
4978 		return -EINVAL;
4979 	if (mqp->state != IB_QPS_RTR)
4980 		return -EINVAL;
4981 
4982 	out = kzalloc(outlen, GFP_KERNEL);
4983 	if (!out)
4984 		return -ENOMEM;
4985 
4986 	err = mlx5_core_dct_query(dev, dct, out, outlen);
4987 	if (err)
4988 		goto out;
4989 
4990 	dctc = MLX5_ADDR_OF(query_dct_out, out, dct_context_entry);
4991 
4992 	if (qp_attr_mask & IB_QP_STATE)
4993 		qp_attr->qp_state = IB_QPS_RTR;
4994 
4995 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
4996 		if (MLX5_GET(dctc, dctc, rre))
4997 			access_flags |= IB_ACCESS_REMOTE_READ;
4998 		if (MLX5_GET(dctc, dctc, rwe))
4999 			access_flags |= IB_ACCESS_REMOTE_WRITE;
5000 		if (MLX5_GET(dctc, dctc, rae))
5001 			access_flags |= IB_ACCESS_REMOTE_ATOMIC;
5002 		qp_attr->qp_access_flags = access_flags;
5003 	}
5004 
5005 	if (qp_attr_mask & IB_QP_PORT)
5006 		qp_attr->port_num = MLX5_GET(dctc, dctc, port);
5007 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER)
5008 		qp_attr->min_rnr_timer = MLX5_GET(dctc, dctc, min_rnr_nak);
5009 	if (qp_attr_mask & IB_QP_AV) {
5010 		qp_attr->ah_attr.grh.traffic_class = MLX5_GET(dctc, dctc, tclass);
5011 		qp_attr->ah_attr.grh.flow_label = MLX5_GET(dctc, dctc, flow_label);
5012 		qp_attr->ah_attr.grh.sgid_index = MLX5_GET(dctc, dctc, my_addr_index);
5013 		qp_attr->ah_attr.grh.hop_limit = MLX5_GET(dctc, dctc, hop_limit);
5014 	}
5015 	if (qp_attr_mask & IB_QP_PATH_MTU)
5016 		qp_attr->path_mtu = MLX5_GET(dctc, dctc, mtu);
5017 	if (qp_attr_mask & IB_QP_PKEY_INDEX)
5018 		qp_attr->pkey_index = MLX5_GET(dctc, dctc, pkey_index);
5019 out:
5020 	kfree(out);
5021 	return err;
5022 }
5023 
5024 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
5025 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
5026 {
5027 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
5028 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
5029 	int err = 0;
5030 	u8 raw_packet_qp_state;
5031 
5032 	if (ibqp->rwq_ind_tbl)
5033 		return -ENOSYS;
5034 
5035 	if (qp->type == IB_QPT_GSI)
5036 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
5037 					    qp_init_attr);
5038 
5039 	/* Not all of output fields are applicable, make sure to zero them */
5040 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
5041 	memset(qp_attr, 0, sizeof(*qp_attr));
5042 
5043 	if (unlikely(qp->type == MLX5_IB_QPT_DCT))
5044 		return mlx5_ib_dct_query_qp(dev, qp, qp_attr,
5045 					    qp_attr_mask, qp_init_attr);
5046 
5047 	mutex_lock(&qp->mutex);
5048 
5049 	if (qp->type == IB_QPT_RAW_PACKET ||
5050 	    qp->flags & IB_QP_CREATE_SOURCE_QPN) {
5051 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
5052 		if (err)
5053 			goto out;
5054 		qp->state = raw_packet_qp_state;
5055 		qp_attr->port_num = 1;
5056 	} else {
5057 		err = query_qp_attr(dev, qp, qp_attr);
5058 		if (err)
5059 			goto out;
5060 	}
5061 
5062 	qp_attr->qp_state	     = qp->state;
5063 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
5064 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
5065 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
5066 
5067 	if (!ibqp->uobject) {
5068 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
5069 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
5070 		qp_init_attr->qp_context = ibqp->qp_context;
5071 	} else {
5072 		qp_attr->cap.max_send_wr  = 0;
5073 		qp_attr->cap.max_send_sge = 0;
5074 	}
5075 
5076 	qp_init_attr->qp_type = qp->type;
5077 	qp_init_attr->recv_cq = ibqp->recv_cq;
5078 	qp_init_attr->send_cq = ibqp->send_cq;
5079 	qp_init_attr->srq = ibqp->srq;
5080 	qp_attr->cap.max_inline_data = qp->max_inline_data;
5081 
5082 	qp_init_attr->cap	     = qp_attr->cap;
5083 
5084 	qp_init_attr->create_flags = qp->flags;
5085 
5086 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
5087 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
5088 
5089 out:
5090 	mutex_unlock(&qp->mutex);
5091 	return err;
5092 }
5093 
5094 int mlx5_ib_alloc_xrcd(struct ib_xrcd *ibxrcd, struct ib_udata *udata)
5095 {
5096 	struct mlx5_ib_dev *dev = to_mdev(ibxrcd->device);
5097 	struct mlx5_ib_xrcd *xrcd = to_mxrcd(ibxrcd);
5098 
5099 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
5100 		return -EOPNOTSUPP;
5101 
5102 	return mlx5_cmd_xrcd_alloc(dev->mdev, &xrcd->xrcdn, 0);
5103 }
5104 
5105 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
5106 {
5107 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
5108 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
5109 
5110 	return mlx5_cmd_xrcd_dealloc(dev->mdev, xrcdn, 0);
5111 }
5112 
5113 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
5114 {
5115 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
5116 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
5117 	struct ib_event event;
5118 
5119 	if (rwq->ibwq.event_handler) {
5120 		event.device     = rwq->ibwq.device;
5121 		event.element.wq = &rwq->ibwq;
5122 		switch (type) {
5123 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
5124 			event.event = IB_EVENT_WQ_FATAL;
5125 			break;
5126 		default:
5127 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
5128 			return;
5129 		}
5130 
5131 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
5132 	}
5133 }
5134 
5135 static int set_delay_drop(struct mlx5_ib_dev *dev)
5136 {
5137 	int err = 0;
5138 
5139 	mutex_lock(&dev->delay_drop.lock);
5140 	if (dev->delay_drop.activate)
5141 		goto out;
5142 
5143 	err = mlx5_core_set_delay_drop(dev, dev->delay_drop.timeout);
5144 	if (err)
5145 		goto out;
5146 
5147 	dev->delay_drop.activate = true;
5148 out:
5149 	mutex_unlock(&dev->delay_drop.lock);
5150 
5151 	if (!err)
5152 		atomic_inc(&dev->delay_drop.rqs_cnt);
5153 	return err;
5154 }
5155 
5156 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
5157 		      struct ib_wq_init_attr *init_attr)
5158 {
5159 	struct mlx5_ib_dev *dev;
5160 	int has_net_offloads;
5161 	__be64 *rq_pas0;
5162 	int ts_format;
5163 	void *in;
5164 	void *rqc;
5165 	void *wq;
5166 	int inlen;
5167 	int err;
5168 
5169 	dev = to_mdev(pd->device);
5170 
5171 	ts_format = get_rq_ts_format(dev, to_mcq(init_attr->cq));
5172 	if (ts_format < 0)
5173 		return ts_format;
5174 
5175 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
5176 	in = kvzalloc(inlen, GFP_KERNEL);
5177 	if (!in)
5178 		return -ENOMEM;
5179 
5180 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
5181 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
5182 	MLX5_SET(rqc,  rqc, mem_rq_type,
5183 		 MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
5184 	MLX5_SET(rqc, rqc, ts_format, ts_format);
5185 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
5186 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
5187 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
5188 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
5189 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
5190 	MLX5_SET(wq, wq, wq_type,
5191 		 rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ ?
5192 		 MLX5_WQ_TYPE_CYCLIC_STRIDING_RQ : MLX5_WQ_TYPE_CYCLIC);
5193 	if (init_attr->create_flags & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
5194 		if (!MLX5_CAP_GEN(dev->mdev, end_pad)) {
5195 			mlx5_ib_dbg(dev, "Scatter end padding is not supported\n");
5196 			err = -EOPNOTSUPP;
5197 			goto out;
5198 		} else {
5199 			MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
5200 		}
5201 	}
5202 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
5203 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ) {
5204 		/*
5205 		 * In Firmware number of strides in each WQE is:
5206 		 *   "512 * 2^single_wqe_log_num_of_strides"
5207 		 * Values 3 to 8 are accepted as 10 to 15, 9 to 18 are
5208 		 * accepted as 0 to 9
5209 		 */
5210 		static const u8 fw_map[] = { 10, 11, 12, 13, 14, 15, 0, 1,
5211 					     2,  3,  4,  5,  6,  7,  8, 9 };
5212 		MLX5_SET(wq, wq, two_byte_shift_en, rwq->two_byte_shift_en);
5213 		MLX5_SET(wq, wq, log_wqe_stride_size,
5214 			 rwq->single_stride_log_num_of_bytes -
5215 			 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES);
5216 		MLX5_SET(wq, wq, log_wqe_num_of_strides,
5217 			 fw_map[rwq->log_num_strides -
5218 				MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES]);
5219 	}
5220 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
5221 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
5222 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
5223 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
5224 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
5225 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
5226 	has_net_offloads = MLX5_CAP_GEN(dev->mdev, eth_net_offloads);
5227 	if (init_attr->create_flags & IB_WQ_FLAGS_CVLAN_STRIPPING) {
5228 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
5229 			mlx5_ib_dbg(dev, "VLAN offloads are not supported\n");
5230 			err = -EOPNOTSUPP;
5231 			goto out;
5232 		}
5233 	} else {
5234 		MLX5_SET(rqc, rqc, vsd, 1);
5235 	}
5236 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS) {
5237 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, scatter_fcs))) {
5238 			mlx5_ib_dbg(dev, "Scatter FCS is not supported\n");
5239 			err = -EOPNOTSUPP;
5240 			goto out;
5241 		}
5242 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
5243 	}
5244 	if (init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
5245 		if (!(dev->ib_dev.attrs.raw_packet_caps &
5246 		      IB_RAW_PACKET_CAP_DELAY_DROP)) {
5247 			mlx5_ib_dbg(dev, "Delay drop is not supported\n");
5248 			err = -EOPNOTSUPP;
5249 			goto out;
5250 		}
5251 		MLX5_SET(rqc, rqc, delay_drop_en, 1);
5252 	}
5253 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
5254 	mlx5_ib_populate_pas(rwq->umem, 1UL << rwq->page_shift, rq_pas0, 0);
5255 	err = mlx5_core_create_rq_tracked(dev, in, inlen, &rwq->core_qp);
5256 	if (!err && init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
5257 		err = set_delay_drop(dev);
5258 		if (err) {
5259 			mlx5_ib_warn(dev, "Failed to enable delay drop err=%d\n",
5260 				     err);
5261 			mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5262 		} else {
5263 			rwq->create_flags |= MLX5_IB_WQ_FLAGS_DELAY_DROP;
5264 		}
5265 	}
5266 out:
5267 	kvfree(in);
5268 	return err;
5269 }
5270 
5271 static int set_user_rq_size(struct mlx5_ib_dev *dev,
5272 			    struct ib_wq_init_attr *wq_init_attr,
5273 			    struct mlx5_ib_create_wq *ucmd,
5274 			    struct mlx5_ib_rwq *rwq)
5275 {
5276 	/* Sanity check RQ size before proceeding */
5277 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
5278 		return -EINVAL;
5279 
5280 	if (!ucmd->rq_wqe_count)
5281 		return -EINVAL;
5282 
5283 	rwq->wqe_count = ucmd->rq_wqe_count;
5284 	rwq->wqe_shift = ucmd->rq_wqe_shift;
5285 	if (check_shl_overflow(rwq->wqe_count, rwq->wqe_shift, &rwq->buf_size))
5286 		return -EINVAL;
5287 
5288 	rwq->log_rq_stride = rwq->wqe_shift;
5289 	rwq->log_rq_size = ilog2(rwq->wqe_count);
5290 	return 0;
5291 }
5292 
5293 static bool log_of_strides_valid(struct mlx5_ib_dev *dev, u32 log_num_strides)
5294 {
5295 	if ((log_num_strides > MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES) ||
5296 	    (log_num_strides < MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
5297 		return false;
5298 
5299 	if (!MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) &&
5300 	    (log_num_strides < MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
5301 		return false;
5302 
5303 	return true;
5304 }
5305 
5306 static int prepare_user_rq(struct ib_pd *pd,
5307 			   struct ib_wq_init_attr *init_attr,
5308 			   struct ib_udata *udata,
5309 			   struct mlx5_ib_rwq *rwq)
5310 {
5311 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
5312 	struct mlx5_ib_create_wq ucmd = {};
5313 	int err;
5314 	size_t required_cmd_sz;
5315 
5316 	required_cmd_sz = offsetofend(struct mlx5_ib_create_wq,
5317 				      single_stride_log_num_of_bytes);
5318 	if (udata->inlen < required_cmd_sz) {
5319 		mlx5_ib_dbg(dev, "invalid inlen\n");
5320 		return -EINVAL;
5321 	}
5322 
5323 	if (udata->inlen > sizeof(ucmd) &&
5324 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
5325 				 udata->inlen - sizeof(ucmd))) {
5326 		mlx5_ib_dbg(dev, "inlen is not supported\n");
5327 		return -EOPNOTSUPP;
5328 	}
5329 
5330 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
5331 		mlx5_ib_dbg(dev, "copy failed\n");
5332 		return -EFAULT;
5333 	}
5334 
5335 	if (ucmd.comp_mask & (~MLX5_IB_CREATE_WQ_STRIDING_RQ)) {
5336 		mlx5_ib_dbg(dev, "invalid comp mask\n");
5337 		return -EOPNOTSUPP;
5338 	} else if (ucmd.comp_mask & MLX5_IB_CREATE_WQ_STRIDING_RQ) {
5339 		if (!MLX5_CAP_GEN(dev->mdev, striding_rq)) {
5340 			mlx5_ib_dbg(dev, "Striding RQ is not supported\n");
5341 			return -EOPNOTSUPP;
5342 		}
5343 		if ((ucmd.single_stride_log_num_of_bytes <
5344 		    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES) ||
5345 		    (ucmd.single_stride_log_num_of_bytes >
5346 		     MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES)) {
5347 			mlx5_ib_dbg(dev, "Invalid log stride size (%u. Range is %u - %u)\n",
5348 				    ucmd.single_stride_log_num_of_bytes,
5349 				    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES,
5350 				    MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES);
5351 			return -EINVAL;
5352 		}
5353 		if (!log_of_strides_valid(dev,
5354 					  ucmd.single_wqe_log_num_of_strides)) {
5355 			mlx5_ib_dbg(
5356 				dev,
5357 				"Invalid log num strides (%u. Range is %u - %u)\n",
5358 				ucmd.single_wqe_log_num_of_strides,
5359 				MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) ?
5360 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES :
5361 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES,
5362 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES);
5363 			return -EINVAL;
5364 		}
5365 		rwq->single_stride_log_num_of_bytes =
5366 			ucmd.single_stride_log_num_of_bytes;
5367 		rwq->log_num_strides = ucmd.single_wqe_log_num_of_strides;
5368 		rwq->two_byte_shift_en = !!ucmd.two_byte_shift_en;
5369 		rwq->create_flags |= MLX5_IB_WQ_FLAGS_STRIDING_RQ;
5370 	}
5371 
5372 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
5373 	if (err) {
5374 		mlx5_ib_dbg(dev, "err %d\n", err);
5375 		return err;
5376 	}
5377 
5378 	err = create_user_rq(dev, pd, udata, rwq, &ucmd);
5379 	if (err) {
5380 		mlx5_ib_dbg(dev, "err %d\n", err);
5381 		return err;
5382 	}
5383 
5384 	rwq->user_index = ucmd.user_index;
5385 	return 0;
5386 }
5387 
5388 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
5389 				struct ib_wq_init_attr *init_attr,
5390 				struct ib_udata *udata)
5391 {
5392 	struct mlx5_ib_dev *dev;
5393 	struct mlx5_ib_rwq *rwq;
5394 	struct mlx5_ib_create_wq_resp resp = {};
5395 	size_t min_resp_len;
5396 	int err;
5397 
5398 	if (!udata)
5399 		return ERR_PTR(-ENOSYS);
5400 
5401 	min_resp_len = offsetofend(struct mlx5_ib_create_wq_resp, reserved);
5402 	if (udata->outlen && udata->outlen < min_resp_len)
5403 		return ERR_PTR(-EINVAL);
5404 
5405 	if (!capable(CAP_SYS_RAWIO) &&
5406 	    init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP)
5407 		return ERR_PTR(-EPERM);
5408 
5409 	dev = to_mdev(pd->device);
5410 	switch (init_attr->wq_type) {
5411 	case IB_WQT_RQ:
5412 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
5413 		if (!rwq)
5414 			return ERR_PTR(-ENOMEM);
5415 		err = prepare_user_rq(pd, init_attr, udata, rwq);
5416 		if (err)
5417 			goto err;
5418 		err = create_rq(rwq, pd, init_attr);
5419 		if (err)
5420 			goto err_user_rq;
5421 		break;
5422 	default:
5423 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
5424 			    init_attr->wq_type);
5425 		return ERR_PTR(-EINVAL);
5426 	}
5427 
5428 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
5429 	rwq->ibwq.state = IB_WQS_RESET;
5430 	if (udata->outlen) {
5431 		resp.response_length = offsetofend(
5432 			struct mlx5_ib_create_wq_resp, response_length);
5433 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
5434 		if (err)
5435 			goto err_copy;
5436 	}
5437 
5438 	rwq->core_qp.event = mlx5_ib_wq_event;
5439 	rwq->ibwq.event_handler = init_attr->event_handler;
5440 	return &rwq->ibwq;
5441 
5442 err_copy:
5443 	mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5444 err_user_rq:
5445 	destroy_user_rq(dev, pd, rwq, udata);
5446 err:
5447 	kfree(rwq);
5448 	return ERR_PTR(err);
5449 }
5450 
5451 int mlx5_ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
5452 {
5453 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
5454 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
5455 	int ret;
5456 
5457 	ret = mlx5_core_destroy_rq_tracked(dev, &rwq->core_qp);
5458 	if (ret)
5459 		return ret;
5460 	destroy_user_rq(dev, wq->pd, rwq, udata);
5461 	kfree(rwq);
5462 	return 0;
5463 }
5464 
5465 int mlx5_ib_create_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_table,
5466 				 struct ib_rwq_ind_table_init_attr *init_attr,
5467 				 struct ib_udata *udata)
5468 {
5469 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl =
5470 		to_mrwq_ind_table(ib_rwq_ind_table);
5471 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_table->device);
5472 	int sz = 1 << init_attr->log_ind_tbl_size;
5473 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
5474 	size_t min_resp_len;
5475 	int inlen;
5476 	int err;
5477 	int i;
5478 	u32 *in;
5479 	void *rqtc;
5480 
5481 	if (udata->inlen > 0 &&
5482 	    !ib_is_udata_cleared(udata, 0,
5483 				 udata->inlen))
5484 		return -EOPNOTSUPP;
5485 
5486 	if (init_attr->log_ind_tbl_size >
5487 	    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
5488 		mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
5489 			    init_attr->log_ind_tbl_size,
5490 			    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
5491 		return -EINVAL;
5492 	}
5493 
5494 	min_resp_len =
5495 		offsetofend(struct mlx5_ib_create_rwq_ind_tbl_resp, reserved);
5496 	if (udata->outlen && udata->outlen < min_resp_len)
5497 		return -EINVAL;
5498 
5499 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
5500 	in = kvzalloc(inlen, GFP_KERNEL);
5501 	if (!in)
5502 		return -ENOMEM;
5503 
5504 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
5505 
5506 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
5507 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
5508 
5509 	for (i = 0; i < sz; i++)
5510 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
5511 
5512 	rwq_ind_tbl->uid = to_mpd(init_attr->ind_tbl[0]->pd)->uid;
5513 	MLX5_SET(create_rqt_in, in, uid, rwq_ind_tbl->uid);
5514 
5515 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
5516 	kvfree(in);
5517 	if (err)
5518 		return err;
5519 
5520 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
5521 	if (udata->outlen) {
5522 		resp.response_length =
5523 			offsetofend(struct mlx5_ib_create_rwq_ind_tbl_resp,
5524 				    response_length);
5525 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
5526 		if (err)
5527 			goto err_copy;
5528 	}
5529 
5530 	return 0;
5531 
5532 err_copy:
5533 	mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
5534 	return err;
5535 }
5536 
5537 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
5538 {
5539 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
5540 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
5541 
5542 	return mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
5543 }
5544 
5545 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
5546 		      u32 wq_attr_mask, struct ib_udata *udata)
5547 {
5548 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
5549 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
5550 	struct mlx5_ib_modify_wq ucmd = {};
5551 	size_t required_cmd_sz;
5552 	int curr_wq_state;
5553 	int wq_state;
5554 	int inlen;
5555 	int err;
5556 	void *rqc;
5557 	void *in;
5558 
5559 	required_cmd_sz = offsetofend(struct mlx5_ib_modify_wq, reserved);
5560 	if (udata->inlen < required_cmd_sz)
5561 		return -EINVAL;
5562 
5563 	if (udata->inlen > sizeof(ucmd) &&
5564 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
5565 				 udata->inlen - sizeof(ucmd)))
5566 		return -EOPNOTSUPP;
5567 
5568 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
5569 		return -EFAULT;
5570 
5571 	if (ucmd.comp_mask || ucmd.reserved)
5572 		return -EOPNOTSUPP;
5573 
5574 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
5575 	in = kvzalloc(inlen, GFP_KERNEL);
5576 	if (!in)
5577 		return -ENOMEM;
5578 
5579 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
5580 
5581 	curr_wq_state = wq_attr->curr_wq_state;
5582 	wq_state = wq_attr->wq_state;
5583 	if (curr_wq_state == IB_WQS_ERR)
5584 		curr_wq_state = MLX5_RQC_STATE_ERR;
5585 	if (wq_state == IB_WQS_ERR)
5586 		wq_state = MLX5_RQC_STATE_ERR;
5587 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
5588 	MLX5_SET(modify_rq_in, in, uid, to_mpd(wq->pd)->uid);
5589 	MLX5_SET(rqc, rqc, state, wq_state);
5590 
5591 	if (wq_attr_mask & IB_WQ_FLAGS) {
5592 		if (wq_attr->flags_mask & IB_WQ_FLAGS_CVLAN_STRIPPING) {
5593 			if (!(MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
5594 			      MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
5595 				mlx5_ib_dbg(dev, "VLAN offloads are not "
5596 					    "supported\n");
5597 				err = -EOPNOTSUPP;
5598 				goto out;
5599 			}
5600 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
5601 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
5602 			MLX5_SET(rqc, rqc, vsd,
5603 				 (wq_attr->flags & IB_WQ_FLAGS_CVLAN_STRIPPING) ? 0 : 1);
5604 		}
5605 
5606 		if (wq_attr->flags_mask & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
5607 			mlx5_ib_dbg(dev, "Modifying scatter end padding is not supported\n");
5608 			err = -EOPNOTSUPP;
5609 			goto out;
5610 		}
5611 	}
5612 
5613 	if (curr_wq_state == IB_WQS_RESET && wq_state == IB_WQS_RDY) {
5614 		u16 set_id;
5615 
5616 		set_id = mlx5_ib_get_counters_id(dev, 0);
5617 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
5618 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
5619 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
5620 			MLX5_SET(rqc, rqc, counter_set_id, set_id);
5621 		} else
5622 			dev_info_once(
5623 				&dev->ib_dev.dev,
5624 				"Receive WQ counters are not supported on current FW\n");
5625 	}
5626 
5627 	err = mlx5_core_modify_rq(dev->mdev, rwq->core_qp.qpn, in);
5628 	if (!err)
5629 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
5630 
5631 out:
5632 	kvfree(in);
5633 	return err;
5634 }
5635 
5636 struct mlx5_ib_drain_cqe {
5637 	struct ib_cqe cqe;
5638 	struct completion done;
5639 };
5640 
5641 static void mlx5_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
5642 {
5643 	struct mlx5_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
5644 						     struct mlx5_ib_drain_cqe,
5645 						     cqe);
5646 
5647 	complete(&cqe->done);
5648 }
5649 
5650 /* This function returns only once the drained WR was completed */
5651 static void handle_drain_completion(struct ib_cq *cq,
5652 				    struct mlx5_ib_drain_cqe *sdrain,
5653 				    struct mlx5_ib_dev *dev)
5654 {
5655 	struct mlx5_core_dev *mdev = dev->mdev;
5656 
5657 	if (cq->poll_ctx == IB_POLL_DIRECT) {
5658 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
5659 			ib_process_cq_direct(cq, -1);
5660 		return;
5661 	}
5662 
5663 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5664 		struct mlx5_ib_cq *mcq = to_mcq(cq);
5665 		bool triggered = false;
5666 		unsigned long flags;
5667 
5668 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
5669 		/* Make sure that the CQ handler won't run if wasn't run yet */
5670 		if (!mcq->mcq.reset_notify_added)
5671 			mcq->mcq.reset_notify_added = 1;
5672 		else
5673 			triggered = true;
5674 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
5675 
5676 		if (triggered) {
5677 			/* Wait for any scheduled/running task to be ended */
5678 			switch (cq->poll_ctx) {
5679 			case IB_POLL_SOFTIRQ:
5680 				irq_poll_disable(&cq->iop);
5681 				irq_poll_enable(&cq->iop);
5682 				break;
5683 			case IB_POLL_WORKQUEUE:
5684 				cancel_work_sync(&cq->work);
5685 				break;
5686 			default:
5687 				WARN_ON_ONCE(1);
5688 			}
5689 		}
5690 
5691 		/* Run the CQ handler - this makes sure that the drain WR will
5692 		 * be processed if wasn't processed yet.
5693 		 */
5694 		mcq->mcq.comp(&mcq->mcq, NULL);
5695 	}
5696 
5697 	wait_for_completion(&sdrain->done);
5698 }
5699 
5700 void mlx5_ib_drain_sq(struct ib_qp *qp)
5701 {
5702 	struct ib_cq *cq = qp->send_cq;
5703 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
5704 	struct mlx5_ib_drain_cqe sdrain;
5705 	const struct ib_send_wr *bad_swr;
5706 	struct ib_rdma_wr swr = {
5707 		.wr = {
5708 			.next = NULL,
5709 			{ .wr_cqe	= &sdrain.cqe, },
5710 			.opcode	= IB_WR_RDMA_WRITE,
5711 		},
5712 	};
5713 	int ret;
5714 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5715 	struct mlx5_core_dev *mdev = dev->mdev;
5716 
5717 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
5718 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5719 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
5720 		return;
5721 	}
5722 
5723 	sdrain.cqe.done = mlx5_ib_drain_qp_done;
5724 	init_completion(&sdrain.done);
5725 
5726 	ret = mlx5_ib_post_send_drain(qp, &swr.wr, &bad_swr);
5727 	if (ret) {
5728 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
5729 		return;
5730 	}
5731 
5732 	handle_drain_completion(cq, &sdrain, dev);
5733 }
5734 
5735 void mlx5_ib_drain_rq(struct ib_qp *qp)
5736 {
5737 	struct ib_cq *cq = qp->recv_cq;
5738 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
5739 	struct mlx5_ib_drain_cqe rdrain;
5740 	struct ib_recv_wr rwr = {};
5741 	const struct ib_recv_wr *bad_rwr;
5742 	int ret;
5743 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5744 	struct mlx5_core_dev *mdev = dev->mdev;
5745 
5746 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
5747 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
5748 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
5749 		return;
5750 	}
5751 
5752 	rwr.wr_cqe = &rdrain.cqe;
5753 	rdrain.cqe.done = mlx5_ib_drain_qp_done;
5754 	init_completion(&rdrain.done);
5755 
5756 	ret = mlx5_ib_post_recv_drain(qp, &rwr, &bad_rwr);
5757 	if (ret) {
5758 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
5759 		return;
5760 	}
5761 
5762 	handle_drain_completion(cq, &rdrain, dev);
5763 }
5764 
5765 /*
5766  * Bind a qp to a counter. If @counter is NULL then bind the qp to
5767  * the default counter
5768  */
5769 int mlx5_ib_qp_set_counter(struct ib_qp *qp, struct rdma_counter *counter)
5770 {
5771 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
5772 	struct mlx5_ib_qp *mqp = to_mqp(qp);
5773 	int err = 0;
5774 
5775 	mutex_lock(&mqp->mutex);
5776 	if (mqp->state == IB_QPS_RESET) {
5777 		qp->counter = counter;
5778 		goto out;
5779 	}
5780 
5781 	if (!MLX5_CAP_GEN(dev->mdev, rts2rts_qp_counters_set_id)) {
5782 		err = -EOPNOTSUPP;
5783 		goto out;
5784 	}
5785 
5786 	if (mqp->state == IB_QPS_RTS) {
5787 		err = __mlx5_ib_qp_set_counter(qp, counter);
5788 		if (!err)
5789 			qp->counter = counter;
5790 
5791 		goto out;
5792 	}
5793 
5794 	mqp->counter_pending = 1;
5795 	qp->counter = counter;
5796 
5797 out:
5798 	mutex_unlock(&mqp->mutex);
5799 	return err;
5800 }
5801 
5802 int mlx5_ib_qp_event_init(void)
5803 {
5804 	mlx5_ib_qp_event_wq = alloc_ordered_workqueue("mlx5_ib_qp_event_wq", 0);
5805 	if (!mlx5_ib_qp_event_wq)
5806 		return -ENOMEM;
5807 
5808 	return 0;
5809 }
5810 
5811 void mlx5_ib_qp_event_cleanup(void)
5812 {
5813 	destroy_workqueue(mlx5_ib_qp_event_wq);
5814 }
5815