xref: /openbmc/linux/drivers/infiniband/hw/mlx5/qp.c (revision dc6a81c3)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/module.h>
34 #include <rdma/ib_umem.h>
35 #include <rdma/ib_cache.h>
36 #include <rdma/ib_user_verbs.h>
37 #include <rdma/rdma_counter.h>
38 #include <linux/mlx5/fs.h>
39 #include "mlx5_ib.h"
40 #include "ib_rep.h"
41 #include "cmd.h"
42 
43 /* not supported currently */
44 static int wq_signature;
45 
46 enum {
47 	MLX5_IB_ACK_REQ_FREQ	= 8,
48 };
49 
50 enum {
51 	MLX5_IB_DEFAULT_SCHED_QUEUE	= 0x83,
52 	MLX5_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
53 	MLX5_IB_LINK_TYPE_IB		= 0,
54 	MLX5_IB_LINK_TYPE_ETH		= 1
55 };
56 
57 enum {
58 	MLX5_IB_SQ_STRIDE	= 6,
59 	MLX5_IB_SQ_UMR_INLINE_THRESHOLD = 64,
60 };
61 
62 static const u32 mlx5_ib_opcode[] = {
63 	[IB_WR_SEND]				= MLX5_OPCODE_SEND,
64 	[IB_WR_LSO]				= MLX5_OPCODE_LSO,
65 	[IB_WR_SEND_WITH_IMM]			= MLX5_OPCODE_SEND_IMM,
66 	[IB_WR_RDMA_WRITE]			= MLX5_OPCODE_RDMA_WRITE,
67 	[IB_WR_RDMA_WRITE_WITH_IMM]		= MLX5_OPCODE_RDMA_WRITE_IMM,
68 	[IB_WR_RDMA_READ]			= MLX5_OPCODE_RDMA_READ,
69 	[IB_WR_ATOMIC_CMP_AND_SWP]		= MLX5_OPCODE_ATOMIC_CS,
70 	[IB_WR_ATOMIC_FETCH_AND_ADD]		= MLX5_OPCODE_ATOMIC_FA,
71 	[IB_WR_SEND_WITH_INV]			= MLX5_OPCODE_SEND_INVAL,
72 	[IB_WR_LOCAL_INV]			= MLX5_OPCODE_UMR,
73 	[IB_WR_REG_MR]				= MLX5_OPCODE_UMR,
74 	[IB_WR_MASKED_ATOMIC_CMP_AND_SWP]	= MLX5_OPCODE_ATOMIC_MASKED_CS,
75 	[IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]	= MLX5_OPCODE_ATOMIC_MASKED_FA,
76 	[MLX5_IB_WR_UMR]			= MLX5_OPCODE_UMR,
77 };
78 
79 struct mlx5_wqe_eth_pad {
80 	u8 rsvd0[16];
81 };
82 
83 enum raw_qp_set_mask_map {
84 	MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID		= 1UL << 0,
85 	MLX5_RAW_QP_RATE_LIMIT			= 1UL << 1,
86 };
87 
88 struct mlx5_modify_raw_qp_param {
89 	u16 operation;
90 
91 	u32 set_mask; /* raw_qp_set_mask_map */
92 
93 	struct mlx5_rate_limit rl;
94 
95 	u8 rq_q_ctr_id;
96 	u16 port;
97 };
98 
99 static void get_cqs(enum ib_qp_type qp_type,
100 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
101 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
102 
103 static int is_qp0(enum ib_qp_type qp_type)
104 {
105 	return qp_type == IB_QPT_SMI;
106 }
107 
108 static int is_sqp(enum ib_qp_type qp_type)
109 {
110 	return is_qp0(qp_type) || is_qp1(qp_type);
111 }
112 
113 /**
114  * mlx5_ib_read_user_wqe_common() - Copy a WQE (or part of) from user WQ
115  * to kernel buffer
116  *
117  * @umem: User space memory where the WQ is
118  * @buffer: buffer to copy to
119  * @buflen: buffer length
120  * @wqe_index: index of WQE to copy from
121  * @wq_offset: offset to start of WQ
122  * @wq_wqe_cnt: number of WQEs in WQ
123  * @wq_wqe_shift: log2 of WQE size
124  * @bcnt: number of bytes to copy
125  * @bytes_copied: number of bytes to copy (return value)
126  *
127  * Copies from start of WQE bcnt or less bytes.
128  * Does not gurantee to copy the entire WQE.
129  *
130  * Return: zero on success, or an error code.
131  */
132 static int mlx5_ib_read_user_wqe_common(struct ib_umem *umem, void *buffer,
133 					size_t buflen, int wqe_index,
134 					int wq_offset, int wq_wqe_cnt,
135 					int wq_wqe_shift, int bcnt,
136 					size_t *bytes_copied)
137 {
138 	size_t offset = wq_offset + ((wqe_index % wq_wqe_cnt) << wq_wqe_shift);
139 	size_t wq_end = wq_offset + (wq_wqe_cnt << wq_wqe_shift);
140 	size_t copy_length;
141 	int ret;
142 
143 	/* don't copy more than requested, more than buffer length or
144 	 * beyond WQ end
145 	 */
146 	copy_length = min_t(u32, buflen, wq_end - offset);
147 	copy_length = min_t(u32, copy_length, bcnt);
148 
149 	ret = ib_umem_copy_from(buffer, umem, offset, copy_length);
150 	if (ret)
151 		return ret;
152 
153 	if (!ret && bytes_copied)
154 		*bytes_copied = copy_length;
155 
156 	return 0;
157 }
158 
159 static int mlx5_ib_read_kernel_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index,
160 				      void *buffer, size_t buflen, size_t *bc)
161 {
162 	struct mlx5_wqe_ctrl_seg *ctrl;
163 	size_t bytes_copied = 0;
164 	size_t wqe_length;
165 	void *p;
166 	int ds;
167 
168 	wqe_index = wqe_index & qp->sq.fbc.sz_m1;
169 
170 	/* read the control segment first */
171 	p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, wqe_index);
172 	ctrl = p;
173 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
174 	wqe_length = ds * MLX5_WQE_DS_UNITS;
175 
176 	/* read rest of WQE if it spreads over more than one stride */
177 	while (bytes_copied < wqe_length) {
178 		size_t copy_length =
179 			min_t(size_t, buflen - bytes_copied, MLX5_SEND_WQE_BB);
180 
181 		if (!copy_length)
182 			break;
183 
184 		memcpy(buffer + bytes_copied, p, copy_length);
185 		bytes_copied += copy_length;
186 
187 		wqe_index = (wqe_index + 1) & qp->sq.fbc.sz_m1;
188 		p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, wqe_index);
189 	}
190 	*bc = bytes_copied;
191 	return 0;
192 }
193 
194 static int mlx5_ib_read_user_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index,
195 				    void *buffer, size_t buflen, size_t *bc)
196 {
197 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
198 	struct ib_umem *umem = base->ubuffer.umem;
199 	struct mlx5_ib_wq *wq = &qp->sq;
200 	struct mlx5_wqe_ctrl_seg *ctrl;
201 	size_t bytes_copied;
202 	size_t bytes_copied2;
203 	size_t wqe_length;
204 	int ret;
205 	int ds;
206 
207 	/* at first read as much as possible */
208 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index,
209 					   wq->offset, wq->wqe_cnt,
210 					   wq->wqe_shift, buflen,
211 					   &bytes_copied);
212 	if (ret)
213 		return ret;
214 
215 	/* we need at least control segment size to proceed */
216 	if (bytes_copied < sizeof(*ctrl))
217 		return -EINVAL;
218 
219 	ctrl = buffer;
220 	ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
221 	wqe_length = ds * MLX5_WQE_DS_UNITS;
222 
223 	/* if we copied enough then we are done */
224 	if (bytes_copied >= wqe_length) {
225 		*bc = bytes_copied;
226 		return 0;
227 	}
228 
229 	/* otherwise this a wrapped around wqe
230 	 * so read the remaining bytes starting
231 	 * from  wqe_index 0
232 	 */
233 	ret = mlx5_ib_read_user_wqe_common(umem, buffer + bytes_copied,
234 					   buflen - bytes_copied, 0, wq->offset,
235 					   wq->wqe_cnt, wq->wqe_shift,
236 					   wqe_length - bytes_copied,
237 					   &bytes_copied2);
238 
239 	if (ret)
240 		return ret;
241 	*bc = bytes_copied + bytes_copied2;
242 	return 0;
243 }
244 
245 int mlx5_ib_read_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
246 			size_t buflen, size_t *bc)
247 {
248 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
249 	struct ib_umem *umem = base->ubuffer.umem;
250 
251 	if (buflen < sizeof(struct mlx5_wqe_ctrl_seg))
252 		return -EINVAL;
253 
254 	if (!umem)
255 		return mlx5_ib_read_kernel_wqe_sq(qp, wqe_index, buffer,
256 						  buflen, bc);
257 
258 	return mlx5_ib_read_user_wqe_sq(qp, wqe_index, buffer, buflen, bc);
259 }
260 
261 static int mlx5_ib_read_user_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index,
262 				    void *buffer, size_t buflen, size_t *bc)
263 {
264 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
265 	struct ib_umem *umem = base->ubuffer.umem;
266 	struct mlx5_ib_wq *wq = &qp->rq;
267 	size_t bytes_copied;
268 	int ret;
269 
270 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index,
271 					   wq->offset, wq->wqe_cnt,
272 					   wq->wqe_shift, buflen,
273 					   &bytes_copied);
274 
275 	if (ret)
276 		return ret;
277 	*bc = bytes_copied;
278 	return 0;
279 }
280 
281 int mlx5_ib_read_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
282 			size_t buflen, size_t *bc)
283 {
284 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
285 	struct ib_umem *umem = base->ubuffer.umem;
286 	struct mlx5_ib_wq *wq = &qp->rq;
287 	size_t wqe_size = 1 << wq->wqe_shift;
288 
289 	if (buflen < wqe_size)
290 		return -EINVAL;
291 
292 	if (!umem)
293 		return -EOPNOTSUPP;
294 
295 	return mlx5_ib_read_user_wqe_rq(qp, wqe_index, buffer, buflen, bc);
296 }
297 
298 static int mlx5_ib_read_user_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index,
299 				     void *buffer, size_t buflen, size_t *bc)
300 {
301 	struct ib_umem *umem = srq->umem;
302 	size_t bytes_copied;
303 	int ret;
304 
305 	ret = mlx5_ib_read_user_wqe_common(umem, buffer, buflen, wqe_index, 0,
306 					   srq->msrq.max, srq->msrq.wqe_shift,
307 					   buflen, &bytes_copied);
308 
309 	if (ret)
310 		return ret;
311 	*bc = bytes_copied;
312 	return 0;
313 }
314 
315 int mlx5_ib_read_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index, void *buffer,
316 			 size_t buflen, size_t *bc)
317 {
318 	struct ib_umem *umem = srq->umem;
319 	size_t wqe_size = 1 << srq->msrq.wqe_shift;
320 
321 	if (buflen < wqe_size)
322 		return -EINVAL;
323 
324 	if (!umem)
325 		return -EOPNOTSUPP;
326 
327 	return mlx5_ib_read_user_wqe_srq(srq, wqe_index, buffer, buflen, bc);
328 }
329 
330 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
331 {
332 	struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
333 	struct ib_event event;
334 
335 	if (type == MLX5_EVENT_TYPE_PATH_MIG) {
336 		/* This event is only valid for trans_qps */
337 		to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
338 	}
339 
340 	if (ibqp->event_handler) {
341 		event.device     = ibqp->device;
342 		event.element.qp = ibqp;
343 		switch (type) {
344 		case MLX5_EVENT_TYPE_PATH_MIG:
345 			event.event = IB_EVENT_PATH_MIG;
346 			break;
347 		case MLX5_EVENT_TYPE_COMM_EST:
348 			event.event = IB_EVENT_COMM_EST;
349 			break;
350 		case MLX5_EVENT_TYPE_SQ_DRAINED:
351 			event.event = IB_EVENT_SQ_DRAINED;
352 			break;
353 		case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
354 			event.event = IB_EVENT_QP_LAST_WQE_REACHED;
355 			break;
356 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
357 			event.event = IB_EVENT_QP_FATAL;
358 			break;
359 		case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
360 			event.event = IB_EVENT_PATH_MIG_ERR;
361 			break;
362 		case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
363 			event.event = IB_EVENT_QP_REQ_ERR;
364 			break;
365 		case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
366 			event.event = IB_EVENT_QP_ACCESS_ERR;
367 			break;
368 		default:
369 			pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
370 			return;
371 		}
372 
373 		ibqp->event_handler(&event, ibqp->qp_context);
374 	}
375 }
376 
377 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
378 		       int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
379 {
380 	int wqe_size;
381 	int wq_size;
382 
383 	/* Sanity check RQ size before proceeding */
384 	if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
385 		return -EINVAL;
386 
387 	if (!has_rq) {
388 		qp->rq.max_gs = 0;
389 		qp->rq.wqe_cnt = 0;
390 		qp->rq.wqe_shift = 0;
391 		cap->max_recv_wr = 0;
392 		cap->max_recv_sge = 0;
393 	} else {
394 		if (ucmd) {
395 			qp->rq.wqe_cnt = ucmd->rq_wqe_count;
396 			if (ucmd->rq_wqe_shift > BITS_PER_BYTE * sizeof(ucmd->rq_wqe_shift))
397 				return -EINVAL;
398 			qp->rq.wqe_shift = ucmd->rq_wqe_shift;
399 			if ((1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) < qp->wq_sig)
400 				return -EINVAL;
401 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
402 			qp->rq.max_post = qp->rq.wqe_cnt;
403 		} else {
404 			wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
405 			wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
406 			wqe_size = roundup_pow_of_two(wqe_size);
407 			wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
408 			wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
409 			qp->rq.wqe_cnt = wq_size / wqe_size;
410 			if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
411 				mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
412 					    wqe_size,
413 					    MLX5_CAP_GEN(dev->mdev,
414 							 max_wqe_sz_rq));
415 				return -EINVAL;
416 			}
417 			qp->rq.wqe_shift = ilog2(wqe_size);
418 			qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
419 			qp->rq.max_post = qp->rq.wqe_cnt;
420 		}
421 	}
422 
423 	return 0;
424 }
425 
426 static int sq_overhead(struct ib_qp_init_attr *attr)
427 {
428 	int size = 0;
429 
430 	switch (attr->qp_type) {
431 	case IB_QPT_XRC_INI:
432 		size += sizeof(struct mlx5_wqe_xrc_seg);
433 		/* fall through */
434 	case IB_QPT_RC:
435 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
436 			max(sizeof(struct mlx5_wqe_atomic_seg) +
437 			    sizeof(struct mlx5_wqe_raddr_seg),
438 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
439 			    sizeof(struct mlx5_mkey_seg) +
440 			    MLX5_IB_SQ_UMR_INLINE_THRESHOLD /
441 			    MLX5_IB_UMR_OCTOWORD);
442 		break;
443 
444 	case IB_QPT_XRC_TGT:
445 		return 0;
446 
447 	case IB_QPT_UC:
448 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
449 			max(sizeof(struct mlx5_wqe_raddr_seg),
450 			    sizeof(struct mlx5_wqe_umr_ctrl_seg) +
451 			    sizeof(struct mlx5_mkey_seg));
452 		break;
453 
454 	case IB_QPT_UD:
455 		if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
456 			size += sizeof(struct mlx5_wqe_eth_pad) +
457 				sizeof(struct mlx5_wqe_eth_seg);
458 		/* fall through */
459 	case IB_QPT_SMI:
460 	case MLX5_IB_QPT_HW_GSI:
461 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
462 			sizeof(struct mlx5_wqe_datagram_seg);
463 		break;
464 
465 	case MLX5_IB_QPT_REG_UMR:
466 		size += sizeof(struct mlx5_wqe_ctrl_seg) +
467 			sizeof(struct mlx5_wqe_umr_ctrl_seg) +
468 			sizeof(struct mlx5_mkey_seg);
469 		break;
470 
471 	default:
472 		return -EINVAL;
473 	}
474 
475 	return size;
476 }
477 
478 static int calc_send_wqe(struct ib_qp_init_attr *attr)
479 {
480 	int inl_size = 0;
481 	int size;
482 
483 	size = sq_overhead(attr);
484 	if (size < 0)
485 		return size;
486 
487 	if (attr->cap.max_inline_data) {
488 		inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
489 			attr->cap.max_inline_data;
490 	}
491 
492 	size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
493 	if (attr->create_flags & IB_QP_CREATE_INTEGRITY_EN &&
494 	    ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
495 		return MLX5_SIG_WQE_SIZE;
496 	else
497 		return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
498 }
499 
500 static int get_send_sge(struct ib_qp_init_attr *attr, int wqe_size)
501 {
502 	int max_sge;
503 
504 	if (attr->qp_type == IB_QPT_RC)
505 		max_sge = (min_t(int, wqe_size, 512) -
506 			   sizeof(struct mlx5_wqe_ctrl_seg) -
507 			   sizeof(struct mlx5_wqe_raddr_seg)) /
508 			sizeof(struct mlx5_wqe_data_seg);
509 	else if (attr->qp_type == IB_QPT_XRC_INI)
510 		max_sge = (min_t(int, wqe_size, 512) -
511 			   sizeof(struct mlx5_wqe_ctrl_seg) -
512 			   sizeof(struct mlx5_wqe_xrc_seg) -
513 			   sizeof(struct mlx5_wqe_raddr_seg)) /
514 			sizeof(struct mlx5_wqe_data_seg);
515 	else
516 		max_sge = (wqe_size - sq_overhead(attr)) /
517 			sizeof(struct mlx5_wqe_data_seg);
518 
519 	return min_t(int, max_sge, wqe_size - sq_overhead(attr) /
520 		     sizeof(struct mlx5_wqe_data_seg));
521 }
522 
523 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
524 			struct mlx5_ib_qp *qp)
525 {
526 	int wqe_size;
527 	int wq_size;
528 
529 	if (!attr->cap.max_send_wr)
530 		return 0;
531 
532 	wqe_size = calc_send_wqe(attr);
533 	mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
534 	if (wqe_size < 0)
535 		return wqe_size;
536 
537 	if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
538 		mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
539 			    wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
540 		return -EINVAL;
541 	}
542 
543 	qp->max_inline_data = wqe_size - sq_overhead(attr) -
544 			      sizeof(struct mlx5_wqe_inline_seg);
545 	attr->cap.max_inline_data = qp->max_inline_data;
546 
547 	wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
548 	qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
549 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
550 		mlx5_ib_dbg(dev, "send queue size (%d * %d / %d -> %d) exceeds limits(%d)\n",
551 			    attr->cap.max_send_wr, wqe_size, MLX5_SEND_WQE_BB,
552 			    qp->sq.wqe_cnt,
553 			    1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
554 		return -ENOMEM;
555 	}
556 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
557 	qp->sq.max_gs = get_send_sge(attr, wqe_size);
558 	if (qp->sq.max_gs < attr->cap.max_send_sge)
559 		return -ENOMEM;
560 
561 	attr->cap.max_send_sge = qp->sq.max_gs;
562 	qp->sq.max_post = wq_size / wqe_size;
563 	attr->cap.max_send_wr = qp->sq.max_post;
564 
565 	return wq_size;
566 }
567 
568 static int set_user_buf_size(struct mlx5_ib_dev *dev,
569 			    struct mlx5_ib_qp *qp,
570 			    struct mlx5_ib_create_qp *ucmd,
571 			    struct mlx5_ib_qp_base *base,
572 			    struct ib_qp_init_attr *attr)
573 {
574 	int desc_sz = 1 << qp->sq.wqe_shift;
575 
576 	if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
577 		mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
578 			     desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
579 		return -EINVAL;
580 	}
581 
582 	if (ucmd->sq_wqe_count && !is_power_of_2(ucmd->sq_wqe_count)) {
583 		mlx5_ib_warn(dev, "sq_wqe_count %d is not a power of two\n",
584 			     ucmd->sq_wqe_count);
585 		return -EINVAL;
586 	}
587 
588 	qp->sq.wqe_cnt = ucmd->sq_wqe_count;
589 
590 	if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
591 		mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
592 			     qp->sq.wqe_cnt,
593 			     1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
594 		return -EINVAL;
595 	}
596 
597 	if (attr->qp_type == IB_QPT_RAW_PACKET ||
598 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
599 		base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
600 		qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
601 	} else {
602 		base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
603 					 (qp->sq.wqe_cnt << 6);
604 	}
605 
606 	return 0;
607 }
608 
609 static int qp_has_rq(struct ib_qp_init_attr *attr)
610 {
611 	if (attr->qp_type == IB_QPT_XRC_INI ||
612 	    attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
613 	    attr->qp_type == MLX5_IB_QPT_REG_UMR ||
614 	    !attr->cap.max_recv_wr)
615 		return 0;
616 
617 	return 1;
618 }
619 
620 enum {
621 	/* this is the first blue flame register in the array of bfregs assigned
622 	 * to a processes. Since we do not use it for blue flame but rather
623 	 * regular 64 bit doorbells, we do not need a lock for maintaiing
624 	 * "odd/even" order
625 	 */
626 	NUM_NON_BLUE_FLAME_BFREGS = 1,
627 };
628 
629 static int max_bfregs(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi)
630 {
631 	return get_num_static_uars(dev, bfregi) * MLX5_NON_FP_BFREGS_PER_UAR;
632 }
633 
634 static int num_med_bfreg(struct mlx5_ib_dev *dev,
635 			 struct mlx5_bfreg_info *bfregi)
636 {
637 	int n;
638 
639 	n = max_bfregs(dev, bfregi) - bfregi->num_low_latency_bfregs -
640 	    NUM_NON_BLUE_FLAME_BFREGS;
641 
642 	return n >= 0 ? n : 0;
643 }
644 
645 static int first_med_bfreg(struct mlx5_ib_dev *dev,
646 			   struct mlx5_bfreg_info *bfregi)
647 {
648 	return num_med_bfreg(dev, bfregi) ? 1 : -ENOMEM;
649 }
650 
651 static int first_hi_bfreg(struct mlx5_ib_dev *dev,
652 			  struct mlx5_bfreg_info *bfregi)
653 {
654 	int med;
655 
656 	med = num_med_bfreg(dev, bfregi);
657 	return ++med;
658 }
659 
660 static int alloc_high_class_bfreg(struct mlx5_ib_dev *dev,
661 				  struct mlx5_bfreg_info *bfregi)
662 {
663 	int i;
664 
665 	for (i = first_hi_bfreg(dev, bfregi); i < max_bfregs(dev, bfregi); i++) {
666 		if (!bfregi->count[i]) {
667 			bfregi->count[i]++;
668 			return i;
669 		}
670 	}
671 
672 	return -ENOMEM;
673 }
674 
675 static int alloc_med_class_bfreg(struct mlx5_ib_dev *dev,
676 				 struct mlx5_bfreg_info *bfregi)
677 {
678 	int minidx = first_med_bfreg(dev, bfregi);
679 	int i;
680 
681 	if (minidx < 0)
682 		return minidx;
683 
684 	for (i = minidx; i < first_hi_bfreg(dev, bfregi); i++) {
685 		if (bfregi->count[i] < bfregi->count[minidx])
686 			minidx = i;
687 		if (!bfregi->count[minidx])
688 			break;
689 	}
690 
691 	bfregi->count[minidx]++;
692 	return minidx;
693 }
694 
695 static int alloc_bfreg(struct mlx5_ib_dev *dev,
696 		       struct mlx5_bfreg_info *bfregi)
697 {
698 	int bfregn = -ENOMEM;
699 
700 	mutex_lock(&bfregi->lock);
701 	if (bfregi->ver >= 2) {
702 		bfregn = alloc_high_class_bfreg(dev, bfregi);
703 		if (bfregn < 0)
704 			bfregn = alloc_med_class_bfreg(dev, bfregi);
705 	}
706 
707 	if (bfregn < 0) {
708 		BUILD_BUG_ON(NUM_NON_BLUE_FLAME_BFREGS != 1);
709 		bfregn = 0;
710 		bfregi->count[bfregn]++;
711 	}
712 	mutex_unlock(&bfregi->lock);
713 
714 	return bfregn;
715 }
716 
717 void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
718 {
719 	mutex_lock(&bfregi->lock);
720 	bfregi->count[bfregn]--;
721 	mutex_unlock(&bfregi->lock);
722 }
723 
724 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
725 {
726 	switch (state) {
727 	case IB_QPS_RESET:	return MLX5_QP_STATE_RST;
728 	case IB_QPS_INIT:	return MLX5_QP_STATE_INIT;
729 	case IB_QPS_RTR:	return MLX5_QP_STATE_RTR;
730 	case IB_QPS_RTS:	return MLX5_QP_STATE_RTS;
731 	case IB_QPS_SQD:	return MLX5_QP_STATE_SQD;
732 	case IB_QPS_SQE:	return MLX5_QP_STATE_SQER;
733 	case IB_QPS_ERR:	return MLX5_QP_STATE_ERR;
734 	default:		return -1;
735 	}
736 }
737 
738 static int to_mlx5_st(enum ib_qp_type type)
739 {
740 	switch (type) {
741 	case IB_QPT_RC:			return MLX5_QP_ST_RC;
742 	case IB_QPT_UC:			return MLX5_QP_ST_UC;
743 	case IB_QPT_UD:			return MLX5_QP_ST_UD;
744 	case MLX5_IB_QPT_REG_UMR:	return MLX5_QP_ST_REG_UMR;
745 	case IB_QPT_XRC_INI:
746 	case IB_QPT_XRC_TGT:		return MLX5_QP_ST_XRC;
747 	case IB_QPT_SMI:		return MLX5_QP_ST_QP0;
748 	case MLX5_IB_QPT_HW_GSI:	return MLX5_QP_ST_QP1;
749 	case MLX5_IB_QPT_DCI:		return MLX5_QP_ST_DCI;
750 	case IB_QPT_RAW_IPV6:		return MLX5_QP_ST_RAW_IPV6;
751 	case IB_QPT_RAW_PACKET:
752 	case IB_QPT_RAW_ETHERTYPE:	return MLX5_QP_ST_RAW_ETHERTYPE;
753 	case IB_QPT_MAX:
754 	default:		return -EINVAL;
755 	}
756 }
757 
758 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
759 			     struct mlx5_ib_cq *recv_cq);
760 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
761 			       struct mlx5_ib_cq *recv_cq);
762 
763 int bfregn_to_uar_index(struct mlx5_ib_dev *dev,
764 			struct mlx5_bfreg_info *bfregi, u32 bfregn,
765 			bool dyn_bfreg)
766 {
767 	unsigned int bfregs_per_sys_page;
768 	u32 index_of_sys_page;
769 	u32 offset;
770 
771 	bfregs_per_sys_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k) *
772 				MLX5_NON_FP_BFREGS_PER_UAR;
773 	index_of_sys_page = bfregn / bfregs_per_sys_page;
774 
775 	if (dyn_bfreg) {
776 		index_of_sys_page += bfregi->num_static_sys_pages;
777 
778 		if (index_of_sys_page >= bfregi->num_sys_pages)
779 			return -EINVAL;
780 
781 		if (bfregn > bfregi->num_dyn_bfregs ||
782 		    bfregi->sys_pages[index_of_sys_page] == MLX5_IB_INVALID_UAR_INDEX) {
783 			mlx5_ib_dbg(dev, "Invalid dynamic uar index\n");
784 			return -EINVAL;
785 		}
786 	}
787 
788 	offset = bfregn % bfregs_per_sys_page / MLX5_NON_FP_BFREGS_PER_UAR;
789 	return bfregi->sys_pages[index_of_sys_page] + offset;
790 }
791 
792 static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev, struct ib_udata *udata,
793 			    unsigned long addr, size_t size,
794 			    struct ib_umem **umem, int *npages, int *page_shift,
795 			    int *ncont, u32 *offset)
796 {
797 	int err;
798 
799 	*umem = ib_umem_get(&dev->ib_dev, addr, size, 0);
800 	if (IS_ERR(*umem)) {
801 		mlx5_ib_dbg(dev, "umem_get failed\n");
802 		return PTR_ERR(*umem);
803 	}
804 
805 	mlx5_ib_cont_pages(*umem, addr, 0, npages, page_shift, ncont, NULL);
806 
807 	err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
808 	if (err) {
809 		mlx5_ib_warn(dev, "bad offset\n");
810 		goto err_umem;
811 	}
812 
813 	mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
814 		    addr, size, *npages, *page_shift, *ncont, *offset);
815 
816 	return 0;
817 
818 err_umem:
819 	ib_umem_release(*umem);
820 	*umem = NULL;
821 
822 	return err;
823 }
824 
825 static void destroy_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
826 			    struct mlx5_ib_rwq *rwq, struct ib_udata *udata)
827 {
828 	struct mlx5_ib_ucontext *context =
829 		rdma_udata_to_drv_context(
830 			udata,
831 			struct mlx5_ib_ucontext,
832 			ibucontext);
833 
834 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_DELAY_DROP)
835 		atomic_dec(&dev->delay_drop.rqs_cnt);
836 
837 	mlx5_ib_db_unmap_user(context, &rwq->db);
838 	ib_umem_release(rwq->umem);
839 }
840 
841 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
842 			  struct ib_udata *udata, struct mlx5_ib_rwq *rwq,
843 			  struct mlx5_ib_create_wq *ucmd)
844 {
845 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
846 		udata, struct mlx5_ib_ucontext, ibucontext);
847 	int page_shift = 0;
848 	int npages;
849 	u32 offset = 0;
850 	int ncont = 0;
851 	int err;
852 
853 	if (!ucmd->buf_addr)
854 		return -EINVAL;
855 
856 	rwq->umem = ib_umem_get(&dev->ib_dev, ucmd->buf_addr, rwq->buf_size, 0);
857 	if (IS_ERR(rwq->umem)) {
858 		mlx5_ib_dbg(dev, "umem_get failed\n");
859 		err = PTR_ERR(rwq->umem);
860 		return err;
861 	}
862 
863 	mlx5_ib_cont_pages(rwq->umem, ucmd->buf_addr, 0, &npages, &page_shift,
864 			   &ncont, NULL);
865 	err = mlx5_ib_get_buf_offset(ucmd->buf_addr, page_shift,
866 				     &rwq->rq_page_offset);
867 	if (err) {
868 		mlx5_ib_warn(dev, "bad offset\n");
869 		goto err_umem;
870 	}
871 
872 	rwq->rq_num_pas = ncont;
873 	rwq->page_shift = page_shift;
874 	rwq->log_page_size =  page_shift - MLX5_ADAPTER_PAGE_SHIFT;
875 	rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
876 
877 	mlx5_ib_dbg(dev, "addr 0x%llx, size %zd, npages %d, page_shift %d, ncont %d, offset %d\n",
878 		    (unsigned long long)ucmd->buf_addr, rwq->buf_size,
879 		    npages, page_shift, ncont, offset);
880 
881 	err = mlx5_ib_db_map_user(ucontext, udata, ucmd->db_addr, &rwq->db);
882 	if (err) {
883 		mlx5_ib_dbg(dev, "map failed\n");
884 		goto err_umem;
885 	}
886 
887 	rwq->create_type = MLX5_WQ_USER;
888 	return 0;
889 
890 err_umem:
891 	ib_umem_release(rwq->umem);
892 	return err;
893 }
894 
895 static int adjust_bfregn(struct mlx5_ib_dev *dev,
896 			 struct mlx5_bfreg_info *bfregi, int bfregn)
897 {
898 	return bfregn / MLX5_NON_FP_BFREGS_PER_UAR * MLX5_BFREGS_PER_UAR +
899 				bfregn % MLX5_NON_FP_BFREGS_PER_UAR;
900 }
901 
902 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
903 			  struct mlx5_ib_qp *qp, struct ib_udata *udata,
904 			  struct ib_qp_init_attr *attr,
905 			  u32 **in,
906 			  struct mlx5_ib_create_qp_resp *resp, int *inlen,
907 			  struct mlx5_ib_qp_base *base)
908 {
909 	struct mlx5_ib_ucontext *context;
910 	struct mlx5_ib_create_qp ucmd;
911 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
912 	int page_shift = 0;
913 	int uar_index = 0;
914 	int npages;
915 	u32 offset = 0;
916 	int bfregn;
917 	int ncont = 0;
918 	__be64 *pas;
919 	void *qpc;
920 	int err;
921 	u16 uid;
922 
923 	err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
924 	if (err) {
925 		mlx5_ib_dbg(dev, "copy failed\n");
926 		return err;
927 	}
928 
929 	context = rdma_udata_to_drv_context(udata, struct mlx5_ib_ucontext,
930 					    ibucontext);
931 	if (ucmd.flags & MLX5_QP_FLAG_BFREG_INDEX) {
932 		uar_index = bfregn_to_uar_index(dev, &context->bfregi,
933 						ucmd.bfreg_index, true);
934 		if (uar_index < 0)
935 			return uar_index;
936 
937 		bfregn = MLX5_IB_INVALID_BFREG;
938 	} else if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL) {
939 		/*
940 		 * TBD: should come from the verbs when we have the API
941 		 */
942 		/* In CROSS_CHANNEL CQ and QP must use the same UAR */
943 		bfregn = MLX5_CROSS_CHANNEL_BFREG;
944 	}
945 	else {
946 		bfregn = alloc_bfreg(dev, &context->bfregi);
947 		if (bfregn < 0)
948 			return bfregn;
949 	}
950 
951 	mlx5_ib_dbg(dev, "bfregn 0x%x, uar_index 0x%x\n", bfregn, uar_index);
952 	if (bfregn != MLX5_IB_INVALID_BFREG)
953 		uar_index = bfregn_to_uar_index(dev, &context->bfregi, bfregn,
954 						false);
955 
956 	qp->rq.offset = 0;
957 	qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
958 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
959 
960 	err = set_user_buf_size(dev, qp, &ucmd, base, attr);
961 	if (err)
962 		goto err_bfreg;
963 
964 	if (ucmd.buf_addr && ubuffer->buf_size) {
965 		ubuffer->buf_addr = ucmd.buf_addr;
966 		err = mlx5_ib_umem_get(dev, udata, ubuffer->buf_addr,
967 				       ubuffer->buf_size, &ubuffer->umem,
968 				       &npages, &page_shift, &ncont, &offset);
969 		if (err)
970 			goto err_bfreg;
971 	} else {
972 		ubuffer->umem = NULL;
973 	}
974 
975 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
976 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
977 	*in = kvzalloc(*inlen, GFP_KERNEL);
978 	if (!*in) {
979 		err = -ENOMEM;
980 		goto err_umem;
981 	}
982 
983 	uid = (attr->qp_type != IB_QPT_XRC_TGT &&
984 	       attr->qp_type != IB_QPT_XRC_INI) ? to_mpd(pd)->uid : 0;
985 	MLX5_SET(create_qp_in, *in, uid, uid);
986 	pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
987 	if (ubuffer->umem)
988 		mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift, pas, 0);
989 
990 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
991 
992 	MLX5_SET(qpc, qpc, log_page_size, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
993 	MLX5_SET(qpc, qpc, page_offset, offset);
994 
995 	MLX5_SET(qpc, qpc, uar_page, uar_index);
996 	if (bfregn != MLX5_IB_INVALID_BFREG)
997 		resp->bfreg_index = adjust_bfregn(dev, &context->bfregi, bfregn);
998 	else
999 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
1000 	qp->bfregn = bfregn;
1001 
1002 	err = mlx5_ib_db_map_user(context, udata, ucmd.db_addr, &qp->db);
1003 	if (err) {
1004 		mlx5_ib_dbg(dev, "map failed\n");
1005 		goto err_free;
1006 	}
1007 
1008 	err = ib_copy_to_udata(udata, resp, min(udata->outlen, sizeof(*resp)));
1009 	if (err) {
1010 		mlx5_ib_dbg(dev, "copy failed\n");
1011 		goto err_unmap;
1012 	}
1013 	qp->create_type = MLX5_QP_USER;
1014 
1015 	return 0;
1016 
1017 err_unmap:
1018 	mlx5_ib_db_unmap_user(context, &qp->db);
1019 
1020 err_free:
1021 	kvfree(*in);
1022 
1023 err_umem:
1024 	ib_umem_release(ubuffer->umem);
1025 
1026 err_bfreg:
1027 	if (bfregn != MLX5_IB_INVALID_BFREG)
1028 		mlx5_ib_free_bfreg(dev, &context->bfregi, bfregn);
1029 	return err;
1030 }
1031 
1032 static void destroy_qp_user(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1033 			    struct mlx5_ib_qp *qp, struct mlx5_ib_qp_base *base,
1034 			    struct ib_udata *udata)
1035 {
1036 	struct mlx5_ib_ucontext *context =
1037 		rdma_udata_to_drv_context(
1038 			udata,
1039 			struct mlx5_ib_ucontext,
1040 			ibucontext);
1041 
1042 	mlx5_ib_db_unmap_user(context, &qp->db);
1043 	ib_umem_release(base->ubuffer.umem);
1044 
1045 	/*
1046 	 * Free only the BFREGs which are handled by the kernel.
1047 	 * BFREGs of UARs allocated dynamically are handled by user.
1048 	 */
1049 	if (qp->bfregn != MLX5_IB_INVALID_BFREG)
1050 		mlx5_ib_free_bfreg(dev, &context->bfregi, qp->bfregn);
1051 }
1052 
1053 /* get_sq_edge - Get the next nearby edge.
1054  *
1055  * An 'edge' is defined as the first following address after the end
1056  * of the fragment or the SQ. Accordingly, during the WQE construction
1057  * which repetitively increases the pointer to write the next data, it
1058  * simply should check if it gets to an edge.
1059  *
1060  * @sq - SQ buffer.
1061  * @idx - Stride index in the SQ buffer.
1062  *
1063  * Return:
1064  *	The new edge.
1065  */
1066 static void *get_sq_edge(struct mlx5_ib_wq *sq, u32 idx)
1067 {
1068 	void *fragment_end;
1069 
1070 	fragment_end = mlx5_frag_buf_get_wqe
1071 		(&sq->fbc,
1072 		 mlx5_frag_buf_get_idx_last_contig_stride(&sq->fbc, idx));
1073 
1074 	return fragment_end + MLX5_SEND_WQE_BB;
1075 }
1076 
1077 static int create_kernel_qp(struct mlx5_ib_dev *dev,
1078 			    struct ib_qp_init_attr *init_attr,
1079 			    struct mlx5_ib_qp *qp,
1080 			    u32 **in, int *inlen,
1081 			    struct mlx5_ib_qp_base *base)
1082 {
1083 	int uar_index;
1084 	void *qpc;
1085 	int err;
1086 
1087 	if (init_attr->create_flags & ~(IB_QP_CREATE_INTEGRITY_EN |
1088 					IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
1089 					IB_QP_CREATE_IPOIB_UD_LSO |
1090 					IB_QP_CREATE_NETIF_QP |
1091 					MLX5_IB_QP_CREATE_SQPN_QP1 |
1092 					MLX5_IB_QP_CREATE_WC_TEST))
1093 		return -EINVAL;
1094 
1095 	if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
1096 		qp->bf.bfreg = &dev->fp_bfreg;
1097 	else if (init_attr->create_flags & MLX5_IB_QP_CREATE_WC_TEST)
1098 		qp->bf.bfreg = &dev->wc_bfreg;
1099 	else
1100 		qp->bf.bfreg = &dev->bfreg;
1101 
1102 	/* We need to divide by two since each register is comprised of
1103 	 * two buffers of identical size, namely odd and even
1104 	 */
1105 	qp->bf.buf_size = (1 << MLX5_CAP_GEN(dev->mdev, log_bf_reg_size)) / 2;
1106 	uar_index = qp->bf.bfreg->index;
1107 
1108 	err = calc_sq_size(dev, init_attr, qp);
1109 	if (err < 0) {
1110 		mlx5_ib_dbg(dev, "err %d\n", err);
1111 		return err;
1112 	}
1113 
1114 	qp->rq.offset = 0;
1115 	qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
1116 	base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
1117 
1118 	err = mlx5_frag_buf_alloc_node(dev->mdev, base->ubuffer.buf_size,
1119 				       &qp->buf, dev->mdev->priv.numa_node);
1120 	if (err) {
1121 		mlx5_ib_dbg(dev, "err %d\n", err);
1122 		return err;
1123 	}
1124 
1125 	if (qp->rq.wqe_cnt)
1126 		mlx5_init_fbc(qp->buf.frags, qp->rq.wqe_shift,
1127 			      ilog2(qp->rq.wqe_cnt), &qp->rq.fbc);
1128 
1129 	if (qp->sq.wqe_cnt) {
1130 		int sq_strides_offset = (qp->sq.offset  & (PAGE_SIZE - 1)) /
1131 					MLX5_SEND_WQE_BB;
1132 		mlx5_init_fbc_offset(qp->buf.frags +
1133 				     (qp->sq.offset / PAGE_SIZE),
1134 				     ilog2(MLX5_SEND_WQE_BB),
1135 				     ilog2(qp->sq.wqe_cnt),
1136 				     sq_strides_offset, &qp->sq.fbc);
1137 
1138 		qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
1139 	}
1140 
1141 	*inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
1142 		 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
1143 	*in = kvzalloc(*inlen, GFP_KERNEL);
1144 	if (!*in) {
1145 		err = -ENOMEM;
1146 		goto err_buf;
1147 	}
1148 
1149 	qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
1150 	MLX5_SET(qpc, qpc, uar_page, uar_index);
1151 	MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1152 
1153 	/* Set "fast registration enabled" for all kernel QPs */
1154 	MLX5_SET(qpc, qpc, fre, 1);
1155 	MLX5_SET(qpc, qpc, rlky, 1);
1156 
1157 	if (init_attr->create_flags & MLX5_IB_QP_CREATE_SQPN_QP1) {
1158 		MLX5_SET(qpc, qpc, deth_sqpn, 1);
1159 		qp->flags |= MLX5_IB_QP_SQPN_QP1;
1160 	}
1161 
1162 	mlx5_fill_page_frag_array(&qp->buf,
1163 				  (__be64 *)MLX5_ADDR_OF(create_qp_in,
1164 							 *in, pas));
1165 
1166 	err = mlx5_db_alloc(dev->mdev, &qp->db);
1167 	if (err) {
1168 		mlx5_ib_dbg(dev, "err %d\n", err);
1169 		goto err_free;
1170 	}
1171 
1172 	qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1173 				     sizeof(*qp->sq.wrid), GFP_KERNEL);
1174 	qp->sq.wr_data = kvmalloc_array(qp->sq.wqe_cnt,
1175 					sizeof(*qp->sq.wr_data), GFP_KERNEL);
1176 	qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1177 				     sizeof(*qp->rq.wrid), GFP_KERNEL);
1178 	qp->sq.w_list = kvmalloc_array(qp->sq.wqe_cnt,
1179 				       sizeof(*qp->sq.w_list), GFP_KERNEL);
1180 	qp->sq.wqe_head = kvmalloc_array(qp->sq.wqe_cnt,
1181 					 sizeof(*qp->sq.wqe_head), GFP_KERNEL);
1182 
1183 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
1184 	    !qp->sq.w_list || !qp->sq.wqe_head) {
1185 		err = -ENOMEM;
1186 		goto err_wrid;
1187 	}
1188 	qp->create_type = MLX5_QP_KERNEL;
1189 
1190 	return 0;
1191 
1192 err_wrid:
1193 	kvfree(qp->sq.wqe_head);
1194 	kvfree(qp->sq.w_list);
1195 	kvfree(qp->sq.wrid);
1196 	kvfree(qp->sq.wr_data);
1197 	kvfree(qp->rq.wrid);
1198 	mlx5_db_free(dev->mdev, &qp->db);
1199 
1200 err_free:
1201 	kvfree(*in);
1202 
1203 err_buf:
1204 	mlx5_frag_buf_free(dev->mdev, &qp->buf);
1205 	return err;
1206 }
1207 
1208 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1209 {
1210 	kvfree(qp->sq.wqe_head);
1211 	kvfree(qp->sq.w_list);
1212 	kvfree(qp->sq.wrid);
1213 	kvfree(qp->sq.wr_data);
1214 	kvfree(qp->rq.wrid);
1215 	mlx5_db_free(dev->mdev, &qp->db);
1216 	mlx5_frag_buf_free(dev->mdev, &qp->buf);
1217 }
1218 
1219 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
1220 {
1221 	if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
1222 	    (attr->qp_type == MLX5_IB_QPT_DCI) ||
1223 	    (attr->qp_type == IB_QPT_XRC_INI))
1224 		return MLX5_SRQ_RQ;
1225 	else if (!qp->has_rq)
1226 		return MLX5_ZERO_LEN_RQ;
1227 	else
1228 		return MLX5_NON_ZERO_RQ;
1229 }
1230 
1231 static int is_connected(enum ib_qp_type qp_type)
1232 {
1233 	if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC ||
1234 	    qp_type == MLX5_IB_QPT_DCI)
1235 		return 1;
1236 
1237 	return 0;
1238 }
1239 
1240 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1241 				    struct mlx5_ib_qp *qp,
1242 				    struct mlx5_ib_sq *sq, u32 tdn,
1243 				    struct ib_pd *pd)
1244 {
1245 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
1246 	void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1247 
1248 	MLX5_SET(create_tis_in, in, uid, to_mpd(pd)->uid);
1249 	MLX5_SET(tisc, tisc, transport_domain, tdn);
1250 	if (qp->flags & MLX5_IB_QP_UNDERLAY)
1251 		MLX5_SET(tisc, tisc, underlay_qpn, qp->underlay_qpn);
1252 
1253 	return mlx5_core_create_tis(dev->mdev, in, sizeof(in), &sq->tisn);
1254 }
1255 
1256 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1257 				      struct mlx5_ib_sq *sq, struct ib_pd *pd)
1258 {
1259 	mlx5_cmd_destroy_tis(dev->mdev, sq->tisn, to_mpd(pd)->uid);
1260 }
1261 
1262 static void destroy_flow_rule_vport_sq(struct mlx5_ib_sq *sq)
1263 {
1264 	if (sq->flow_rule)
1265 		mlx5_del_flow_rules(sq->flow_rule);
1266 	sq->flow_rule = NULL;
1267 }
1268 
1269 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1270 				   struct ib_udata *udata,
1271 				   struct mlx5_ib_sq *sq, void *qpin,
1272 				   struct ib_pd *pd)
1273 {
1274 	struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1275 	__be64 *pas;
1276 	void *in;
1277 	void *sqc;
1278 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1279 	void *wq;
1280 	int inlen;
1281 	int err;
1282 	int page_shift = 0;
1283 	int npages;
1284 	int ncont = 0;
1285 	u32 offset = 0;
1286 
1287 	err = mlx5_ib_umem_get(dev, udata, ubuffer->buf_addr, ubuffer->buf_size,
1288 			       &sq->ubuffer.umem, &npages, &page_shift, &ncont,
1289 			       &offset);
1290 	if (err)
1291 		return err;
1292 
1293 	inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
1294 	in = kvzalloc(inlen, GFP_KERNEL);
1295 	if (!in) {
1296 		err = -ENOMEM;
1297 		goto err_umem;
1298 	}
1299 
1300 	MLX5_SET(create_sq_in, in, uid, to_mpd(pd)->uid);
1301 	sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1302 	MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1303 	if (MLX5_CAP_ETH(dev->mdev, multi_pkt_send_wqe))
1304 		MLX5_SET(sqc, sqc, allow_multi_pkt_send_wqe, 1);
1305 	MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1306 	MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1307 	MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1308 	MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1309 	MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1310 	if (MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
1311 	    MLX5_CAP_ETH(dev->mdev, swp))
1312 		MLX5_SET(sqc, sqc, allow_swp, 1);
1313 
1314 	wq = MLX5_ADDR_OF(sqc, sqc, wq);
1315 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1316 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1317 	MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1318 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1319 	MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1320 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1321 	MLX5_SET(wq, wq, log_wq_pg_sz,  page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1322 	MLX5_SET(wq, wq, page_offset, offset);
1323 
1324 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1325 	mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
1326 
1327 	err = mlx5_core_create_sq_tracked(dev->mdev, in, inlen, &sq->base.mqp);
1328 
1329 	kvfree(in);
1330 
1331 	if (err)
1332 		goto err_umem;
1333 
1334 	return 0;
1335 
1336 err_umem:
1337 	ib_umem_release(sq->ubuffer.umem);
1338 	sq->ubuffer.umem = NULL;
1339 
1340 	return err;
1341 }
1342 
1343 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1344 				     struct mlx5_ib_sq *sq)
1345 {
1346 	destroy_flow_rule_vport_sq(sq);
1347 	mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1348 	ib_umem_release(sq->ubuffer.umem);
1349 }
1350 
1351 static size_t get_rq_pas_size(void *qpc)
1352 {
1353 	u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1354 	u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1355 	u32 log_rq_size   = MLX5_GET(qpc, qpc, log_rq_size);
1356 	u32 page_offset   = MLX5_GET(qpc, qpc, page_offset);
1357 	u32 po_quanta	  = 1 << (log_page_size - 6);
1358 	u32 rq_sz	  = 1 << (log_rq_size + 4 + log_rq_stride);
1359 	u32 page_size	  = 1 << log_page_size;
1360 	u32 rq_sz_po      = rq_sz + (page_offset * po_quanta);
1361 	u32 rq_num_pas	  = (rq_sz_po + page_size - 1) / page_size;
1362 
1363 	return rq_num_pas * sizeof(u64);
1364 }
1365 
1366 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1367 				   struct mlx5_ib_rq *rq, void *qpin,
1368 				   size_t qpinlen, struct ib_pd *pd)
1369 {
1370 	struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1371 	__be64 *pas;
1372 	__be64 *qp_pas;
1373 	void *in;
1374 	void *rqc;
1375 	void *wq;
1376 	void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1377 	size_t rq_pas_size = get_rq_pas_size(qpc);
1378 	size_t inlen;
1379 	int err;
1380 
1381 	if (qpinlen < rq_pas_size + MLX5_BYTE_OFF(create_qp_in, pas))
1382 		return -EINVAL;
1383 
1384 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1385 	in = kvzalloc(inlen, GFP_KERNEL);
1386 	if (!in)
1387 		return -ENOMEM;
1388 
1389 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
1390 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1391 	if (!(rq->flags & MLX5_IB_RQ_CVLAN_STRIPPING))
1392 		MLX5_SET(rqc, rqc, vsd, 1);
1393 	MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1394 	MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1395 	MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1396 	MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1397 	MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1398 
1399 	if (mqp->flags & MLX5_IB_QP_CAP_SCATTER_FCS)
1400 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
1401 
1402 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
1403 	MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1404 	if (rq->flags & MLX5_IB_RQ_PCI_WRITE_END_PADDING)
1405 		MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1406 	MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1407 	MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1408 	MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1409 	MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1410 	MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1411 	MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1412 
1413 	pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1414 	qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1415 	memcpy(pas, qp_pas, rq_pas_size);
1416 
1417 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rq->base.mqp);
1418 
1419 	kvfree(in);
1420 
1421 	return err;
1422 }
1423 
1424 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1425 				     struct mlx5_ib_rq *rq)
1426 {
1427 	mlx5_core_destroy_rq_tracked(dev->mdev, &rq->base.mqp);
1428 }
1429 
1430 static bool tunnel_offload_supported(struct mlx5_core_dev *dev)
1431 {
1432 	return  (MLX5_CAP_ETH(dev, tunnel_stateless_vxlan) ||
1433 		 MLX5_CAP_ETH(dev, tunnel_stateless_gre) ||
1434 		 MLX5_CAP_ETH(dev, tunnel_stateless_geneve_rx));
1435 }
1436 
1437 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1438 				      struct mlx5_ib_rq *rq,
1439 				      u32 qp_flags_en,
1440 				      struct ib_pd *pd)
1441 {
1442 	if (qp_flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1443 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1444 		mlx5_ib_disable_lb(dev, false, true);
1445 	mlx5_cmd_destroy_tir(dev->mdev, rq->tirn, to_mpd(pd)->uid);
1446 }
1447 
1448 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1449 				    struct mlx5_ib_rq *rq, u32 tdn,
1450 				    u32 *qp_flags_en,
1451 				    struct ib_pd *pd,
1452 				    u32 *out, int outlen)
1453 {
1454 	u8 lb_flag = 0;
1455 	u32 *in;
1456 	void *tirc;
1457 	int inlen;
1458 	int err;
1459 
1460 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1461 	in = kvzalloc(inlen, GFP_KERNEL);
1462 	if (!in)
1463 		return -ENOMEM;
1464 
1465 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1466 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1467 	MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1468 	MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1469 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1470 	if (*qp_flags_en & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1471 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1472 
1473 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC)
1474 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1475 
1476 	if (*qp_flags_en & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)
1477 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1478 
1479 	if (dev->is_rep) {
1480 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1481 		*qp_flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1482 	}
1483 
1484 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1485 
1486 	err = mlx5_core_create_tir_out(dev->mdev, in, inlen, out, outlen);
1487 
1488 	rq->tirn = MLX5_GET(create_tir_out, out, tirn);
1489 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1490 		err = mlx5_ib_enable_lb(dev, false, true);
1491 
1492 		if (err)
1493 			destroy_raw_packet_qp_tir(dev, rq, 0, pd);
1494 	}
1495 	kvfree(in);
1496 
1497 	return err;
1498 }
1499 
1500 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1501 				u32 *in, size_t inlen,
1502 				struct ib_pd *pd,
1503 				struct ib_udata *udata,
1504 				struct mlx5_ib_create_qp_resp *resp)
1505 {
1506 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1507 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1508 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1509 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1510 		udata, struct mlx5_ib_ucontext, ibucontext);
1511 	int err;
1512 	u32 tdn = mucontext->tdn;
1513 	u16 uid = to_mpd(pd)->uid;
1514 	u32 out[MLX5_ST_SZ_DW(create_tir_out)] = {};
1515 
1516 	if (qp->sq.wqe_cnt) {
1517 		err = create_raw_packet_qp_tis(dev, qp, sq, tdn, pd);
1518 		if (err)
1519 			return err;
1520 
1521 		err = create_raw_packet_qp_sq(dev, udata, sq, in, pd);
1522 		if (err)
1523 			goto err_destroy_tis;
1524 
1525 		if (uid) {
1526 			resp->tisn = sq->tisn;
1527 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TISN;
1528 			resp->sqn = sq->base.mqp.qpn;
1529 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_SQN;
1530 		}
1531 
1532 		sq->base.container_mibqp = qp;
1533 		sq->base.mqp.event = mlx5_ib_qp_event;
1534 	}
1535 
1536 	if (qp->rq.wqe_cnt) {
1537 		rq->base.container_mibqp = qp;
1538 
1539 		if (qp->flags & MLX5_IB_QP_CVLAN_STRIPPING)
1540 			rq->flags |= MLX5_IB_RQ_CVLAN_STRIPPING;
1541 		if (qp->flags & MLX5_IB_QP_PCI_WRITE_END_PADDING)
1542 			rq->flags |= MLX5_IB_RQ_PCI_WRITE_END_PADDING;
1543 		err = create_raw_packet_qp_rq(dev, rq, in, inlen, pd);
1544 		if (err)
1545 			goto err_destroy_sq;
1546 
1547 		err = create_raw_packet_qp_tir(
1548 			dev, rq, tdn, &qp->flags_en, pd, out,
1549 			MLX5_ST_SZ_BYTES(create_tir_out));
1550 		if (err)
1551 			goto err_destroy_rq;
1552 
1553 		if (uid) {
1554 			resp->rqn = rq->base.mqp.qpn;
1555 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_RQN;
1556 			resp->tirn = rq->tirn;
1557 			resp->comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1558 			if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner)) {
1559 				resp->tir_icm_addr = MLX5_GET(
1560 					create_tir_out, out, icm_address_31_0);
1561 				resp->tir_icm_addr |=
1562 					(u64)MLX5_GET(create_tir_out, out,
1563 						      icm_address_39_32)
1564 					<< 32;
1565 				resp->tir_icm_addr |=
1566 					(u64)MLX5_GET(create_tir_out, out,
1567 						      icm_address_63_40)
1568 					<< 40;
1569 				resp->comp_mask |=
1570 					MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1571 			}
1572 		}
1573 	}
1574 
1575 	qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1576 						     rq->base.mqp.qpn;
1577 	err = ib_copy_to_udata(udata, resp, min(udata->outlen, sizeof(*resp)));
1578 	if (err)
1579 		goto err_destroy_tir;
1580 
1581 	return 0;
1582 
1583 err_destroy_tir:
1584 	destroy_raw_packet_qp_tir(dev, rq, qp->flags_en, pd);
1585 err_destroy_rq:
1586 	destroy_raw_packet_qp_rq(dev, rq);
1587 err_destroy_sq:
1588 	if (!qp->sq.wqe_cnt)
1589 		return err;
1590 	destroy_raw_packet_qp_sq(dev, sq);
1591 err_destroy_tis:
1592 	destroy_raw_packet_qp_tis(dev, sq, pd);
1593 
1594 	return err;
1595 }
1596 
1597 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1598 				  struct mlx5_ib_qp *qp)
1599 {
1600 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1601 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1602 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1603 
1604 	if (qp->rq.wqe_cnt) {
1605 		destroy_raw_packet_qp_tir(dev, rq, qp->flags_en, qp->ibqp.pd);
1606 		destroy_raw_packet_qp_rq(dev, rq);
1607 	}
1608 
1609 	if (qp->sq.wqe_cnt) {
1610 		destroy_raw_packet_qp_sq(dev, sq);
1611 		destroy_raw_packet_qp_tis(dev, sq, qp->ibqp.pd);
1612 	}
1613 }
1614 
1615 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1616 				    struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1617 {
1618 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1619 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1620 
1621 	sq->sq = &qp->sq;
1622 	rq->rq = &qp->rq;
1623 	sq->doorbell = &qp->db;
1624 	rq->doorbell = &qp->db;
1625 }
1626 
1627 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1628 {
1629 	if (qp->flags_en & (MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1630 			    MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC))
1631 		mlx5_ib_disable_lb(dev, false, true);
1632 	mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1633 			     to_mpd(qp->ibqp.pd)->uid);
1634 }
1635 
1636 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1637 				 struct ib_pd *pd,
1638 				 struct ib_qp_init_attr *init_attr,
1639 				 struct ib_udata *udata)
1640 {
1641 	struct mlx5_ib_ucontext *mucontext = rdma_udata_to_drv_context(
1642 		udata, struct mlx5_ib_ucontext, ibucontext);
1643 	struct mlx5_ib_create_qp_resp resp = {};
1644 	int inlen;
1645 	int outlen;
1646 	int err;
1647 	u32 *in;
1648 	u32 *out;
1649 	void *tirc;
1650 	void *hfso;
1651 	u32 selected_fields = 0;
1652 	u32 outer_l4;
1653 	size_t min_resp_len;
1654 	u32 tdn = mucontext->tdn;
1655 	struct mlx5_ib_create_qp_rss ucmd = {};
1656 	size_t required_cmd_sz;
1657 	u8 lb_flag = 0;
1658 
1659 	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1660 		return -EOPNOTSUPP;
1661 
1662 	if (init_attr->create_flags || init_attr->send_cq)
1663 		return -EINVAL;
1664 
1665 	min_resp_len = offsetof(typeof(resp), bfreg_index) + sizeof(resp.bfreg_index);
1666 	if (udata->outlen < min_resp_len)
1667 		return -EINVAL;
1668 
1669 	required_cmd_sz = offsetof(typeof(ucmd), flags) + sizeof(ucmd.flags);
1670 	if (udata->inlen < required_cmd_sz) {
1671 		mlx5_ib_dbg(dev, "invalid inlen\n");
1672 		return -EINVAL;
1673 	}
1674 
1675 	if (udata->inlen > sizeof(ucmd) &&
1676 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
1677 				 udata->inlen - sizeof(ucmd))) {
1678 		mlx5_ib_dbg(dev, "inlen is not supported\n");
1679 		return -EOPNOTSUPP;
1680 	}
1681 
1682 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
1683 		mlx5_ib_dbg(dev, "copy failed\n");
1684 		return -EFAULT;
1685 	}
1686 
1687 	if (ucmd.comp_mask) {
1688 		mlx5_ib_dbg(dev, "invalid comp mask\n");
1689 		return -EOPNOTSUPP;
1690 	}
1691 
1692 	if (ucmd.flags & ~(MLX5_QP_FLAG_TUNNEL_OFFLOADS |
1693 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
1694 			   MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC)) {
1695 		mlx5_ib_dbg(dev, "invalid flags\n");
1696 		return -EOPNOTSUPP;
1697 	}
1698 
1699 	if (ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS &&
1700 	    !tunnel_offload_supported(dev->mdev)) {
1701 		mlx5_ib_dbg(dev, "tunnel offloads isn't supported\n");
1702 		return -EOPNOTSUPP;
1703 	}
1704 
1705 	if (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_INNER &&
1706 	    !(ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)) {
1707 		mlx5_ib_dbg(dev, "Tunnel offloads must be set for inner RSS\n");
1708 		return -EOPNOTSUPP;
1709 	}
1710 
1711 	if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC || dev->is_rep) {
1712 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
1713 		qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
1714 	}
1715 
1716 	if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC) {
1717 		lb_flag |= MLX5_TIRC_SELF_LB_BLOCK_BLOCK_MULTICAST;
1718 		qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC;
1719 	}
1720 
1721 	err = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
1722 	if (err) {
1723 		mlx5_ib_dbg(dev, "copy failed\n");
1724 		return -EINVAL;
1725 	}
1726 
1727 	inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1728 	outlen = MLX5_ST_SZ_BYTES(create_tir_out);
1729 	in = kvzalloc(inlen + outlen, GFP_KERNEL);
1730 	if (!in)
1731 		return -ENOMEM;
1732 
1733 	out = in + MLX5_ST_SZ_DW(create_tir_in);
1734 	MLX5_SET(create_tir_in, in, uid, to_mpd(pd)->uid);
1735 	tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1736 	MLX5_SET(tirc, tirc, disp_type,
1737 		 MLX5_TIRC_DISP_TYPE_INDIRECT);
1738 	MLX5_SET(tirc, tirc, indirect_table,
1739 		 init_attr->rwq_ind_tbl->ind_tbl_num);
1740 	MLX5_SET(tirc, tirc, transport_domain, tdn);
1741 
1742 	hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1743 
1744 	if (ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS)
1745 		MLX5_SET(tirc, tirc, tunneled_offload_en, 1);
1746 
1747 	MLX5_SET(tirc, tirc, self_lb_block, lb_flag);
1748 
1749 	if (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_INNER)
1750 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner);
1751 	else
1752 		hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1753 
1754 	switch (ucmd.rx_hash_function) {
1755 	case MLX5_RX_HASH_FUNC_TOEPLITZ:
1756 	{
1757 		void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1758 		size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1759 
1760 		if (len != ucmd.rx_key_len) {
1761 			err = -EINVAL;
1762 			goto err;
1763 		}
1764 
1765 		MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1766 		memcpy(rss_key, ucmd.rx_hash_key, len);
1767 		break;
1768 	}
1769 	default:
1770 		err = -EOPNOTSUPP;
1771 		goto err;
1772 	}
1773 
1774 	if (!ucmd.rx_hash_fields_mask) {
1775 		/* special case when this TIR serves as steering entry without hashing */
1776 		if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1777 			goto create_tir;
1778 		err = -EINVAL;
1779 		goto err;
1780 	}
1781 
1782 	if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1783 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1784 	     ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1785 	     (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1786 		err = -EINVAL;
1787 		goto err;
1788 	}
1789 
1790 	/* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1791 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1792 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1793 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1794 			 MLX5_L3_PROT_TYPE_IPV4);
1795 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1796 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1797 		MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1798 			 MLX5_L3_PROT_TYPE_IPV6);
1799 
1800 	outer_l4 = ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1801 		    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP)) << 0 |
1802 		   ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1803 		    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP)) << 1 |
1804 		   (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI) << 2;
1805 
1806 	/* Check that only one l4 protocol is set */
1807 	if (outer_l4 & (outer_l4 - 1)) {
1808 		err = -EINVAL;
1809 		goto err;
1810 	}
1811 
1812 	/* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1813 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1814 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1815 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1816 			 MLX5_L4_PROT_TYPE_TCP);
1817 	else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1818 		 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1819 		MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1820 			 MLX5_L4_PROT_TYPE_UDP);
1821 
1822 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1823 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1824 		selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1825 
1826 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1827 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1828 		selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1829 
1830 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1831 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1832 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1833 
1834 	if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1835 	    (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1836 		selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1837 
1838 	if (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_IPSEC_SPI)
1839 		selected_fields |= MLX5_HASH_FIELD_SEL_IPSEC_SPI;
1840 
1841 	MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1842 
1843 create_tir:
1844 	err = mlx5_core_create_tir_out(dev->mdev, in, inlen, out, outlen);
1845 
1846 	qp->rss_qp.tirn = MLX5_GET(create_tir_out, out, tirn);
1847 	if (!err && MLX5_GET(tirc, tirc, self_lb_block)) {
1848 		err = mlx5_ib_enable_lb(dev, false, true);
1849 
1850 		if (err)
1851 			mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn,
1852 					     to_mpd(pd)->uid);
1853 	}
1854 
1855 	if (err)
1856 		goto err;
1857 
1858 	if (mucontext->devx_uid) {
1859 		resp.comp_mask |= MLX5_IB_CREATE_QP_RESP_MASK_TIRN;
1860 		resp.tirn = qp->rss_qp.tirn;
1861 		if (MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, sw_owner)) {
1862 			resp.tir_icm_addr =
1863 				MLX5_GET(create_tir_out, out, icm_address_31_0);
1864 			resp.tir_icm_addr |= (u64)MLX5_GET(create_tir_out, out,
1865 							   icm_address_39_32)
1866 					     << 32;
1867 			resp.tir_icm_addr |= (u64)MLX5_GET(create_tir_out, out,
1868 							   icm_address_63_40)
1869 					     << 40;
1870 			resp.comp_mask |=
1871 				MLX5_IB_CREATE_QP_RESP_MASK_TIR_ICM_ADDR;
1872 		}
1873 	}
1874 
1875 	err = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
1876 	if (err)
1877 		goto err_copy;
1878 
1879 	kvfree(in);
1880 	/* qpn is reserved for that QP */
1881 	qp->trans_qp.base.mqp.qpn = 0;
1882 	qp->flags |= MLX5_IB_QP_RSS;
1883 	return 0;
1884 
1885 err_copy:
1886 	mlx5_cmd_destroy_tir(dev->mdev, qp->rss_qp.tirn, mucontext->devx_uid);
1887 err:
1888 	kvfree(in);
1889 	return err;
1890 }
1891 
1892 static void configure_responder_scat_cqe(struct ib_qp_init_attr *init_attr,
1893 					 void *qpc)
1894 {
1895 	int rcqe_sz;
1896 
1897 	if (init_attr->qp_type == MLX5_IB_QPT_DCI)
1898 		return;
1899 
1900 	rcqe_sz = mlx5_ib_get_cqe_size(init_attr->recv_cq);
1901 
1902 	if (init_attr->qp_type == MLX5_IB_QPT_DCT) {
1903 		if (rcqe_sz == 128)
1904 			MLX5_SET(dctc, qpc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
1905 
1906 		return;
1907 	}
1908 
1909 	MLX5_SET(qpc, qpc, cs_res,
1910 		 rcqe_sz == 128 ? MLX5_RES_SCAT_DATA64_CQE :
1911 				  MLX5_RES_SCAT_DATA32_CQE);
1912 }
1913 
1914 static void configure_requester_scat_cqe(struct mlx5_ib_dev *dev,
1915 					 struct ib_qp_init_attr *init_attr,
1916 					 struct mlx5_ib_create_qp *ucmd,
1917 					 void *qpc)
1918 {
1919 	enum ib_qp_type qpt = init_attr->qp_type;
1920 	int scqe_sz;
1921 	bool allow_scat_cqe = false;
1922 
1923 	if (qpt == IB_QPT_UC || qpt == IB_QPT_UD)
1924 		return;
1925 
1926 	if (ucmd)
1927 		allow_scat_cqe = ucmd->flags & MLX5_QP_FLAG_ALLOW_SCATTER_CQE;
1928 
1929 	if (!allow_scat_cqe && init_attr->sq_sig_type != IB_SIGNAL_ALL_WR)
1930 		return;
1931 
1932 	scqe_sz = mlx5_ib_get_cqe_size(init_attr->send_cq);
1933 	if (scqe_sz == 128) {
1934 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1935 		return;
1936 	}
1937 
1938 	if (init_attr->qp_type != MLX5_IB_QPT_DCI ||
1939 	    MLX5_CAP_GEN(dev->mdev, dc_req_scat_data_cqe))
1940 		MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1941 }
1942 
1943 static int atomic_size_to_mode(int size_mask)
1944 {
1945 	/* driver does not support atomic_size > 256B
1946 	 * and does not know how to translate bigger sizes
1947 	 */
1948 	int supported_size_mask = size_mask & 0x1ff;
1949 	int log_max_size;
1950 
1951 	if (!supported_size_mask)
1952 		return -EOPNOTSUPP;
1953 
1954 	log_max_size = __fls(supported_size_mask);
1955 
1956 	if (log_max_size > 3)
1957 		return log_max_size;
1958 
1959 	return MLX5_ATOMIC_MODE_8B;
1960 }
1961 
1962 static int get_atomic_mode(struct mlx5_ib_dev *dev,
1963 			   enum ib_qp_type qp_type)
1964 {
1965 	u8 atomic_operations = MLX5_CAP_ATOMIC(dev->mdev, atomic_operations);
1966 	u8 atomic = MLX5_CAP_GEN(dev->mdev, atomic);
1967 	int atomic_mode = -EOPNOTSUPP;
1968 	int atomic_size_mask;
1969 
1970 	if (!atomic)
1971 		return -EOPNOTSUPP;
1972 
1973 	if (qp_type == MLX5_IB_QPT_DCT)
1974 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_dc);
1975 	else
1976 		atomic_size_mask = MLX5_CAP_ATOMIC(dev->mdev, atomic_size_qp);
1977 
1978 	if ((atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_CMP_SWAP) ||
1979 	    (atomic_operations & MLX5_ATOMIC_OPS_EXTENDED_FETCH_ADD))
1980 		atomic_mode = atomic_size_to_mode(atomic_size_mask);
1981 
1982 	if (atomic_mode <= 0 &&
1983 	    (atomic_operations & MLX5_ATOMIC_OPS_CMP_SWAP &&
1984 	     atomic_operations & MLX5_ATOMIC_OPS_FETCH_ADD))
1985 		atomic_mode = MLX5_ATOMIC_MODE_IB_COMP;
1986 
1987 	return atomic_mode;
1988 }
1989 
1990 static inline bool check_flags_mask(uint64_t input, uint64_t supported)
1991 {
1992 	return (input & ~supported) == 0;
1993 }
1994 
1995 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1996 			    struct ib_qp_init_attr *init_attr,
1997 			    struct ib_udata *udata, struct mlx5_ib_qp *qp)
1998 {
1999 	struct mlx5_ib_resources *devr = &dev->devr;
2000 	int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
2001 	struct mlx5_core_dev *mdev = dev->mdev;
2002 	struct mlx5_ib_create_qp_resp resp = {};
2003 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2004 		udata, struct mlx5_ib_ucontext, ibucontext);
2005 	struct mlx5_ib_cq *send_cq;
2006 	struct mlx5_ib_cq *recv_cq;
2007 	unsigned long flags;
2008 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
2009 	struct mlx5_ib_create_qp ucmd;
2010 	struct mlx5_ib_qp_base *base;
2011 	int mlx5_st;
2012 	void *qpc;
2013 	u32 *in;
2014 	int err;
2015 
2016 	mutex_init(&qp->mutex);
2017 	spin_lock_init(&qp->sq.lock);
2018 	spin_lock_init(&qp->rq.lock);
2019 
2020 	mlx5_st = to_mlx5_st(init_attr->qp_type);
2021 	if (mlx5_st < 0)
2022 		return -EINVAL;
2023 
2024 	if (init_attr->rwq_ind_tbl) {
2025 		if (!udata)
2026 			return -ENOSYS;
2027 
2028 		err = create_rss_raw_qp_tir(dev, qp, pd, init_attr, udata);
2029 		return err;
2030 	}
2031 
2032 	if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
2033 		if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
2034 			mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
2035 			return -EINVAL;
2036 		} else {
2037 			qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
2038 		}
2039 	}
2040 
2041 	if (init_attr->create_flags &
2042 			(IB_QP_CREATE_CROSS_CHANNEL |
2043 			 IB_QP_CREATE_MANAGED_SEND |
2044 			 IB_QP_CREATE_MANAGED_RECV)) {
2045 		if (!MLX5_CAP_GEN(mdev, cd)) {
2046 			mlx5_ib_dbg(dev, "cross-channel isn't supported\n");
2047 			return -EINVAL;
2048 		}
2049 		if (init_attr->create_flags & IB_QP_CREATE_CROSS_CHANNEL)
2050 			qp->flags |= MLX5_IB_QP_CROSS_CHANNEL;
2051 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_SEND)
2052 			qp->flags |= MLX5_IB_QP_MANAGED_SEND;
2053 		if (init_attr->create_flags & IB_QP_CREATE_MANAGED_RECV)
2054 			qp->flags |= MLX5_IB_QP_MANAGED_RECV;
2055 	}
2056 
2057 	if (init_attr->qp_type == IB_QPT_UD &&
2058 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO))
2059 		if (!MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
2060 			mlx5_ib_dbg(dev, "ipoib UD lso qp isn't supported\n");
2061 			return -EOPNOTSUPP;
2062 		}
2063 
2064 	if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
2065 		if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2066 			mlx5_ib_dbg(dev, "Scatter FCS is supported only for Raw Packet QPs");
2067 			return -EOPNOTSUPP;
2068 		}
2069 		if (!MLX5_CAP_GEN(dev->mdev, eth_net_offloads) ||
2070 		    !MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
2071 			mlx5_ib_dbg(dev, "Scatter FCS isn't supported\n");
2072 			return -EOPNOTSUPP;
2073 		}
2074 		qp->flags |= MLX5_IB_QP_CAP_SCATTER_FCS;
2075 	}
2076 
2077 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
2078 		qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
2079 
2080 	if (init_attr->create_flags & IB_QP_CREATE_CVLAN_STRIPPING) {
2081 		if (!(MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
2082 		      MLX5_CAP_ETH(dev->mdev, vlan_cap)) ||
2083 		    (init_attr->qp_type != IB_QPT_RAW_PACKET))
2084 			return -EOPNOTSUPP;
2085 		qp->flags |= MLX5_IB_QP_CVLAN_STRIPPING;
2086 	}
2087 
2088 	if (udata) {
2089 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
2090 			mlx5_ib_dbg(dev, "copy failed\n");
2091 			return -EFAULT;
2092 		}
2093 
2094 		if (!check_flags_mask(ucmd.flags,
2095 				      MLX5_QP_FLAG_ALLOW_SCATTER_CQE |
2096 				      MLX5_QP_FLAG_BFREG_INDEX |
2097 				      MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE |
2098 				      MLX5_QP_FLAG_SCATTER_CQE |
2099 				      MLX5_QP_FLAG_SIGNATURE |
2100 				      MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC |
2101 				      MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC |
2102 				      MLX5_QP_FLAG_TUNNEL_OFFLOADS |
2103 				      MLX5_QP_FLAG_TYPE_DCI |
2104 				      MLX5_QP_FLAG_TYPE_DCT))
2105 			return -EINVAL;
2106 
2107 		err = get_qp_user_index(ucontext, &ucmd, udata->inlen, &uidx);
2108 		if (err)
2109 			return err;
2110 
2111 		qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
2112 		if (MLX5_CAP_GEN(dev->mdev, sctr_data_cqe))
2113 			qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
2114 		if (ucmd.flags & MLX5_QP_FLAG_TUNNEL_OFFLOADS) {
2115 			if (init_attr->qp_type != IB_QPT_RAW_PACKET ||
2116 			    !tunnel_offload_supported(mdev)) {
2117 				mlx5_ib_dbg(dev, "Tunnel offload isn't supported\n");
2118 				return -EOPNOTSUPP;
2119 			}
2120 			qp->flags_en |= MLX5_QP_FLAG_TUNNEL_OFFLOADS;
2121 		}
2122 
2123 		if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC) {
2124 			if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2125 				mlx5_ib_dbg(dev, "Self-LB UC isn't supported\n");
2126 				return -EOPNOTSUPP;
2127 			}
2128 			qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_UC;
2129 		}
2130 
2131 		if (ucmd.flags & MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC) {
2132 			if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2133 				mlx5_ib_dbg(dev, "Self-LB UM isn't supported\n");
2134 				return -EOPNOTSUPP;
2135 			}
2136 			qp->flags_en |= MLX5_QP_FLAG_TIR_ALLOW_SELF_LB_MC;
2137 		}
2138 
2139 		if (ucmd.flags & MLX5_QP_FLAG_PACKET_BASED_CREDIT_MODE) {
2140 			if (init_attr->qp_type != IB_QPT_RC ||
2141 				!MLX5_CAP_GEN(dev->mdev, qp_packet_based)) {
2142 				mlx5_ib_dbg(dev, "packet based credit mode isn't supported\n");
2143 				return -EOPNOTSUPP;
2144 			}
2145 			qp->flags |= MLX5_IB_QP_PACKET_BASED_CREDIT;
2146 		}
2147 
2148 		if (init_attr->create_flags & IB_QP_CREATE_SOURCE_QPN) {
2149 			if (init_attr->qp_type != IB_QPT_UD ||
2150 			    (MLX5_CAP_GEN(dev->mdev, port_type) !=
2151 			     MLX5_CAP_PORT_TYPE_IB) ||
2152 			    !mlx5_get_flow_namespace(dev->mdev, MLX5_FLOW_NAMESPACE_BYPASS)) {
2153 				mlx5_ib_dbg(dev, "Source QP option isn't supported\n");
2154 				return -EOPNOTSUPP;
2155 			}
2156 
2157 			qp->flags |= MLX5_IB_QP_UNDERLAY;
2158 			qp->underlay_qpn = init_attr->source_qpn;
2159 		}
2160 	} else {
2161 		qp->wq_sig = !!wq_signature;
2162 	}
2163 
2164 	base = (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2165 		qp->flags & MLX5_IB_QP_UNDERLAY) ?
2166 	       &qp->raw_packet_qp.rq.base :
2167 	       &qp->trans_qp.base;
2168 
2169 	qp->has_rq = qp_has_rq(init_attr);
2170 	err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
2171 			  qp, udata ? &ucmd : NULL);
2172 	if (err) {
2173 		mlx5_ib_dbg(dev, "err %d\n", err);
2174 		return err;
2175 	}
2176 
2177 	if (pd) {
2178 		if (udata) {
2179 			__u32 max_wqes =
2180 				1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
2181 			mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
2182 			if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
2183 			    ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
2184 				mlx5_ib_dbg(dev, "invalid rq params\n");
2185 				return -EINVAL;
2186 			}
2187 			if (ucmd.sq_wqe_count > max_wqes) {
2188 				mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
2189 					    ucmd.sq_wqe_count, max_wqes);
2190 				return -EINVAL;
2191 			}
2192 			if (init_attr->create_flags &
2193 			    MLX5_IB_QP_CREATE_SQPN_QP1) {
2194 				mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
2195 				return -EINVAL;
2196 			}
2197 			err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
2198 					     &resp, &inlen, base);
2199 			if (err)
2200 				mlx5_ib_dbg(dev, "err %d\n", err);
2201 		} else {
2202 			err = create_kernel_qp(dev, init_attr, qp, &in, &inlen,
2203 					       base);
2204 			if (err)
2205 				mlx5_ib_dbg(dev, "err %d\n", err);
2206 		}
2207 
2208 		if (err)
2209 			return err;
2210 	} else {
2211 		in = kvzalloc(inlen, GFP_KERNEL);
2212 		if (!in)
2213 			return -ENOMEM;
2214 
2215 		qp->create_type = MLX5_QP_EMPTY;
2216 	}
2217 
2218 	if (is_sqp(init_attr->qp_type))
2219 		qp->port = init_attr->port_num;
2220 
2221 	qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
2222 
2223 	MLX5_SET(qpc, qpc, st, mlx5_st);
2224 	MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
2225 
2226 	if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
2227 		MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
2228 	else
2229 		MLX5_SET(qpc, qpc, latency_sensitive, 1);
2230 
2231 
2232 	if (qp->wq_sig)
2233 		MLX5_SET(qpc, qpc, wq_signature, 1);
2234 
2235 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
2236 		MLX5_SET(qpc, qpc, block_lb_mc, 1);
2237 
2238 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
2239 		MLX5_SET(qpc, qpc, cd_master, 1);
2240 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
2241 		MLX5_SET(qpc, qpc, cd_slave_send, 1);
2242 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
2243 		MLX5_SET(qpc, qpc, cd_slave_receive, 1);
2244 	if (qp->flags & MLX5_IB_QP_PACKET_BASED_CREDIT)
2245 		MLX5_SET(qpc, qpc, req_e2e_credit_mode, 1);
2246 	if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
2247 		configure_responder_scat_cqe(init_attr, qpc);
2248 		configure_requester_scat_cqe(dev, init_attr,
2249 					     udata ? &ucmd : NULL,
2250 					     qpc);
2251 	}
2252 
2253 	if (qp->rq.wqe_cnt) {
2254 		MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
2255 		MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
2256 	}
2257 
2258 	MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
2259 
2260 	if (qp->sq.wqe_cnt) {
2261 		MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
2262 	} else {
2263 		MLX5_SET(qpc, qpc, no_sq, 1);
2264 		if (init_attr->srq &&
2265 		    init_attr->srq->srq_type == IB_SRQT_TM)
2266 			MLX5_SET(qpc, qpc, offload_type,
2267 				 MLX5_QPC_OFFLOAD_TYPE_RNDV);
2268 	}
2269 
2270 	/* Set default resources */
2271 	switch (init_attr->qp_type) {
2272 	case IB_QPT_XRC_TGT:
2273 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2274 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
2275 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2276 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(init_attr->xrcd)->xrcdn);
2277 		break;
2278 	case IB_QPT_XRC_INI:
2279 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
2280 		MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
2281 		MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
2282 		break;
2283 	default:
2284 		if (init_attr->srq) {
2285 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x0)->xrcdn);
2286 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(init_attr->srq)->msrq.srqn);
2287 		} else {
2288 			MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
2289 			MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s1)->msrq.srqn);
2290 		}
2291 	}
2292 
2293 	if (init_attr->send_cq)
2294 		MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
2295 
2296 	if (init_attr->recv_cq)
2297 		MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
2298 
2299 	MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
2300 
2301 	/* 0xffffff means we ask to work with cqe version 0 */
2302 	if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
2303 		MLX5_SET(qpc, qpc, user_index, uidx);
2304 
2305 	/* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
2306 	if (init_attr->qp_type == IB_QPT_UD &&
2307 	    (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)) {
2308 		MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
2309 		qp->flags |= MLX5_IB_QP_LSO;
2310 	}
2311 
2312 	if (init_attr->create_flags & IB_QP_CREATE_PCI_WRITE_END_PADDING) {
2313 		if (!MLX5_CAP_GEN(dev->mdev, end_pad)) {
2314 			mlx5_ib_dbg(dev, "scatter end padding is not supported\n");
2315 			err = -EOPNOTSUPP;
2316 			goto err;
2317 		} else if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
2318 			MLX5_SET(qpc, qpc, end_padding_mode,
2319 				 MLX5_WQ_END_PAD_MODE_ALIGN);
2320 		} else {
2321 			qp->flags |= MLX5_IB_QP_PCI_WRITE_END_PADDING;
2322 		}
2323 	}
2324 
2325 	if (inlen < 0) {
2326 		err = -EINVAL;
2327 		goto err;
2328 	}
2329 
2330 	if (init_attr->qp_type == IB_QPT_RAW_PACKET ||
2331 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
2332 		qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd.sq_buf_addr;
2333 		raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
2334 		err = create_raw_packet_qp(dev, qp, in, inlen, pd, udata,
2335 					   &resp);
2336 	} else {
2337 		err = mlx5_core_create_qp(dev->mdev, &base->mqp, in, inlen);
2338 	}
2339 
2340 	if (err) {
2341 		mlx5_ib_dbg(dev, "create qp failed\n");
2342 		goto err_create;
2343 	}
2344 
2345 	kvfree(in);
2346 
2347 	base->container_mibqp = qp;
2348 	base->mqp.event = mlx5_ib_qp_event;
2349 
2350 	get_cqs(init_attr->qp_type, init_attr->send_cq, init_attr->recv_cq,
2351 		&send_cq, &recv_cq);
2352 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2353 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2354 	/* Maintain device to QPs access, needed for further handling via reset
2355 	 * flow
2356 	 */
2357 	list_add_tail(&qp->qps_list, &dev->qp_list);
2358 	/* Maintain CQ to QPs access, needed for further handling via reset flow
2359 	 */
2360 	if (send_cq)
2361 		list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
2362 	if (recv_cq)
2363 		list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
2364 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2365 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2366 
2367 	return 0;
2368 
2369 err_create:
2370 	if (qp->create_type == MLX5_QP_USER)
2371 		destroy_qp_user(dev, pd, qp, base, udata);
2372 	else if (qp->create_type == MLX5_QP_KERNEL)
2373 		destroy_qp_kernel(dev, qp);
2374 
2375 err:
2376 	kvfree(in);
2377 	return err;
2378 }
2379 
2380 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2381 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
2382 {
2383 	if (send_cq) {
2384 		if (recv_cq) {
2385 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2386 				spin_lock(&send_cq->lock);
2387 				spin_lock_nested(&recv_cq->lock,
2388 						 SINGLE_DEPTH_NESTING);
2389 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2390 				spin_lock(&send_cq->lock);
2391 				__acquire(&recv_cq->lock);
2392 			} else {
2393 				spin_lock(&recv_cq->lock);
2394 				spin_lock_nested(&send_cq->lock,
2395 						 SINGLE_DEPTH_NESTING);
2396 			}
2397 		} else {
2398 			spin_lock(&send_cq->lock);
2399 			__acquire(&recv_cq->lock);
2400 		}
2401 	} else if (recv_cq) {
2402 		spin_lock(&recv_cq->lock);
2403 		__acquire(&send_cq->lock);
2404 	} else {
2405 		__acquire(&send_cq->lock);
2406 		__acquire(&recv_cq->lock);
2407 	}
2408 }
2409 
2410 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
2411 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
2412 {
2413 	if (send_cq) {
2414 		if (recv_cq) {
2415 			if (send_cq->mcq.cqn < recv_cq->mcq.cqn)  {
2416 				spin_unlock(&recv_cq->lock);
2417 				spin_unlock(&send_cq->lock);
2418 			} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
2419 				__release(&recv_cq->lock);
2420 				spin_unlock(&send_cq->lock);
2421 			} else {
2422 				spin_unlock(&send_cq->lock);
2423 				spin_unlock(&recv_cq->lock);
2424 			}
2425 		} else {
2426 			__release(&recv_cq->lock);
2427 			spin_unlock(&send_cq->lock);
2428 		}
2429 	} else if (recv_cq) {
2430 		__release(&send_cq->lock);
2431 		spin_unlock(&recv_cq->lock);
2432 	} else {
2433 		__release(&recv_cq->lock);
2434 		__release(&send_cq->lock);
2435 	}
2436 }
2437 
2438 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
2439 {
2440 	return to_mpd(qp->ibqp.pd);
2441 }
2442 
2443 static void get_cqs(enum ib_qp_type qp_type,
2444 		    struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
2445 		    struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
2446 {
2447 	switch (qp_type) {
2448 	case IB_QPT_XRC_TGT:
2449 		*send_cq = NULL;
2450 		*recv_cq = NULL;
2451 		break;
2452 	case MLX5_IB_QPT_REG_UMR:
2453 	case IB_QPT_XRC_INI:
2454 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2455 		*recv_cq = NULL;
2456 		break;
2457 
2458 	case IB_QPT_SMI:
2459 	case MLX5_IB_QPT_HW_GSI:
2460 	case IB_QPT_RC:
2461 	case IB_QPT_UC:
2462 	case IB_QPT_UD:
2463 	case IB_QPT_RAW_IPV6:
2464 	case IB_QPT_RAW_ETHERTYPE:
2465 	case IB_QPT_RAW_PACKET:
2466 		*send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
2467 		*recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
2468 		break;
2469 
2470 	case IB_QPT_MAX:
2471 	default:
2472 		*send_cq = NULL;
2473 		*recv_cq = NULL;
2474 		break;
2475 	}
2476 }
2477 
2478 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2479 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
2480 				u8 lag_tx_affinity);
2481 
2482 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2483 			      struct ib_udata *udata)
2484 {
2485 	struct mlx5_ib_cq *send_cq, *recv_cq;
2486 	struct mlx5_ib_qp_base *base;
2487 	unsigned long flags;
2488 	int err;
2489 
2490 	if (qp->ibqp.rwq_ind_tbl) {
2491 		destroy_rss_raw_qp_tir(dev, qp);
2492 		return;
2493 	}
2494 
2495 	base = (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
2496 		qp->flags & MLX5_IB_QP_UNDERLAY) ?
2497 	       &qp->raw_packet_qp.rq.base :
2498 	       &qp->trans_qp.base;
2499 
2500 	if (qp->state != IB_QPS_RESET) {
2501 		if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET &&
2502 		    !(qp->flags & MLX5_IB_QP_UNDERLAY)) {
2503 			err = mlx5_core_qp_modify(dev->mdev,
2504 						  MLX5_CMD_OP_2RST_QP, 0,
2505 						  NULL, &base->mqp);
2506 		} else {
2507 			struct mlx5_modify_raw_qp_param raw_qp_param = {
2508 				.operation = MLX5_CMD_OP_2RST_QP
2509 			};
2510 
2511 			err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2512 		}
2513 		if (err)
2514 			mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
2515 				     base->mqp.qpn);
2516 	}
2517 
2518 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2519 		&send_cq, &recv_cq);
2520 
2521 	spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
2522 	mlx5_ib_lock_cqs(send_cq, recv_cq);
2523 	/* del from lists under both locks above to protect reset flow paths */
2524 	list_del(&qp->qps_list);
2525 	if (send_cq)
2526 		list_del(&qp->cq_send_list);
2527 
2528 	if (recv_cq)
2529 		list_del(&qp->cq_recv_list);
2530 
2531 	if (qp->create_type == MLX5_QP_KERNEL) {
2532 		__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2533 				   qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
2534 		if (send_cq != recv_cq)
2535 			__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
2536 					   NULL);
2537 	}
2538 	mlx5_ib_unlock_cqs(send_cq, recv_cq);
2539 	spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
2540 
2541 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
2542 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
2543 		destroy_raw_packet_qp(dev, qp);
2544 	} else {
2545 		err = mlx5_core_destroy_qp(dev->mdev, &base->mqp);
2546 		if (err)
2547 			mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
2548 				     base->mqp.qpn);
2549 	}
2550 
2551 	if (qp->create_type == MLX5_QP_KERNEL)
2552 		destroy_qp_kernel(dev, qp);
2553 	else if (qp->create_type == MLX5_QP_USER)
2554 		destroy_qp_user(dev, &get_pd(qp)->ibpd, qp, base, udata);
2555 }
2556 
2557 static const char *ib_qp_type_str(enum ib_qp_type type)
2558 {
2559 	switch (type) {
2560 	case IB_QPT_SMI:
2561 		return "IB_QPT_SMI";
2562 	case IB_QPT_GSI:
2563 		return "IB_QPT_GSI";
2564 	case IB_QPT_RC:
2565 		return "IB_QPT_RC";
2566 	case IB_QPT_UC:
2567 		return "IB_QPT_UC";
2568 	case IB_QPT_UD:
2569 		return "IB_QPT_UD";
2570 	case IB_QPT_RAW_IPV6:
2571 		return "IB_QPT_RAW_IPV6";
2572 	case IB_QPT_RAW_ETHERTYPE:
2573 		return "IB_QPT_RAW_ETHERTYPE";
2574 	case IB_QPT_XRC_INI:
2575 		return "IB_QPT_XRC_INI";
2576 	case IB_QPT_XRC_TGT:
2577 		return "IB_QPT_XRC_TGT";
2578 	case IB_QPT_RAW_PACKET:
2579 		return "IB_QPT_RAW_PACKET";
2580 	case MLX5_IB_QPT_REG_UMR:
2581 		return "MLX5_IB_QPT_REG_UMR";
2582 	case IB_QPT_DRIVER:
2583 		return "IB_QPT_DRIVER";
2584 	case IB_QPT_MAX:
2585 	default:
2586 		return "Invalid QP type";
2587 	}
2588 }
2589 
2590 static struct ib_qp *mlx5_ib_create_dct(struct ib_pd *pd,
2591 					struct ib_qp_init_attr *attr,
2592 					struct mlx5_ib_create_qp *ucmd,
2593 					struct ib_udata *udata)
2594 {
2595 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2596 		udata, struct mlx5_ib_ucontext, ibucontext);
2597 	struct mlx5_ib_qp *qp;
2598 	int err = 0;
2599 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
2600 	void *dctc;
2601 
2602 	if (!attr->srq || !attr->recv_cq)
2603 		return ERR_PTR(-EINVAL);
2604 
2605 	err = get_qp_user_index(ucontext, ucmd, sizeof(*ucmd), &uidx);
2606 	if (err)
2607 		return ERR_PTR(err);
2608 
2609 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2610 	if (!qp)
2611 		return ERR_PTR(-ENOMEM);
2612 
2613 	qp->dct.in = kzalloc(MLX5_ST_SZ_BYTES(create_dct_in), GFP_KERNEL);
2614 	if (!qp->dct.in) {
2615 		err = -ENOMEM;
2616 		goto err_free;
2617 	}
2618 
2619 	MLX5_SET(create_dct_in, qp->dct.in, uid, to_mpd(pd)->uid);
2620 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
2621 	qp->qp_sub_type = MLX5_IB_QPT_DCT;
2622 	MLX5_SET(dctc, dctc, pd, to_mpd(pd)->pdn);
2623 	MLX5_SET(dctc, dctc, srqn_xrqn, to_msrq(attr->srq)->msrq.srqn);
2624 	MLX5_SET(dctc, dctc, cqn, to_mcq(attr->recv_cq)->mcq.cqn);
2625 	MLX5_SET64(dctc, dctc, dc_access_key, ucmd->access_key);
2626 	MLX5_SET(dctc, dctc, user_index, uidx);
2627 
2628 	if (ucmd->flags & MLX5_QP_FLAG_SCATTER_CQE)
2629 		configure_responder_scat_cqe(attr, dctc);
2630 
2631 	qp->state = IB_QPS_RESET;
2632 
2633 	return &qp->ibqp;
2634 err_free:
2635 	kfree(qp);
2636 	return ERR_PTR(err);
2637 }
2638 
2639 static int set_mlx_qp_type(struct mlx5_ib_dev *dev,
2640 			   struct ib_qp_init_attr *init_attr,
2641 			   struct mlx5_ib_create_qp *ucmd,
2642 			   struct ib_udata *udata)
2643 {
2644 	enum { MLX_QP_FLAGS = MLX5_QP_FLAG_TYPE_DCT | MLX5_QP_FLAG_TYPE_DCI };
2645 	int err;
2646 
2647 	if (!udata)
2648 		return -EINVAL;
2649 
2650 	if (udata->inlen < sizeof(*ucmd)) {
2651 		mlx5_ib_dbg(dev, "create_qp user command is smaller than expected\n");
2652 		return -EINVAL;
2653 	}
2654 	err = ib_copy_from_udata(ucmd, udata, sizeof(*ucmd));
2655 	if (err)
2656 		return err;
2657 
2658 	if ((ucmd->flags & MLX_QP_FLAGS) == MLX5_QP_FLAG_TYPE_DCI) {
2659 		init_attr->qp_type = MLX5_IB_QPT_DCI;
2660 	} else {
2661 		if ((ucmd->flags & MLX_QP_FLAGS) == MLX5_QP_FLAG_TYPE_DCT) {
2662 			init_attr->qp_type = MLX5_IB_QPT_DCT;
2663 		} else {
2664 			mlx5_ib_dbg(dev, "Invalid QP flags\n");
2665 			return -EINVAL;
2666 		}
2667 	}
2668 
2669 	if (!MLX5_CAP_GEN(dev->mdev, dct)) {
2670 		mlx5_ib_dbg(dev, "DC transport is not supported\n");
2671 		return -EOPNOTSUPP;
2672 	}
2673 
2674 	return 0;
2675 }
2676 
2677 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
2678 				struct ib_qp_init_attr *verbs_init_attr,
2679 				struct ib_udata *udata)
2680 {
2681 	struct mlx5_ib_dev *dev;
2682 	struct mlx5_ib_qp *qp;
2683 	u16 xrcdn = 0;
2684 	int err;
2685 	struct ib_qp_init_attr mlx_init_attr;
2686 	struct ib_qp_init_attr *init_attr = verbs_init_attr;
2687 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
2688 		udata, struct mlx5_ib_ucontext, ibucontext);
2689 
2690 	if (pd) {
2691 		dev = to_mdev(pd->device);
2692 
2693 		if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
2694 			if (!ucontext) {
2695 				mlx5_ib_dbg(dev, "Raw Packet QP is not supported for kernel consumers\n");
2696 				return ERR_PTR(-EINVAL);
2697 			} else if (!ucontext->cqe_version) {
2698 				mlx5_ib_dbg(dev, "Raw Packet QP is only supported for CQE version > 0\n");
2699 				return ERR_PTR(-EINVAL);
2700 			}
2701 		}
2702 	} else {
2703 		/* being cautious here */
2704 		if (init_attr->qp_type != IB_QPT_XRC_TGT &&
2705 		    init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
2706 			pr_warn("%s: no PD for transport %s\n", __func__,
2707 				ib_qp_type_str(init_attr->qp_type));
2708 			return ERR_PTR(-EINVAL);
2709 		}
2710 		dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
2711 	}
2712 
2713 	if (init_attr->qp_type == IB_QPT_DRIVER) {
2714 		struct mlx5_ib_create_qp ucmd;
2715 
2716 		init_attr = &mlx_init_attr;
2717 		memcpy(init_attr, verbs_init_attr, sizeof(*verbs_init_attr));
2718 		err = set_mlx_qp_type(dev, init_attr, &ucmd, udata);
2719 		if (err)
2720 			return ERR_PTR(err);
2721 
2722 		if (init_attr->qp_type == MLX5_IB_QPT_DCI) {
2723 			if (init_attr->cap.max_recv_wr ||
2724 			    init_attr->cap.max_recv_sge) {
2725 				mlx5_ib_dbg(dev, "DCI QP requires zero size receive queue\n");
2726 				return ERR_PTR(-EINVAL);
2727 			}
2728 		} else {
2729 			return mlx5_ib_create_dct(pd, init_attr, &ucmd, udata);
2730 		}
2731 	}
2732 
2733 	switch (init_attr->qp_type) {
2734 	case IB_QPT_XRC_TGT:
2735 	case IB_QPT_XRC_INI:
2736 		if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
2737 			mlx5_ib_dbg(dev, "XRC not supported\n");
2738 			return ERR_PTR(-ENOSYS);
2739 		}
2740 		init_attr->recv_cq = NULL;
2741 		if (init_attr->qp_type == IB_QPT_XRC_TGT) {
2742 			xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
2743 			init_attr->send_cq = NULL;
2744 		}
2745 
2746 		/* fall through */
2747 	case IB_QPT_RAW_PACKET:
2748 	case IB_QPT_RC:
2749 	case IB_QPT_UC:
2750 	case IB_QPT_UD:
2751 	case IB_QPT_SMI:
2752 	case MLX5_IB_QPT_HW_GSI:
2753 	case MLX5_IB_QPT_REG_UMR:
2754 	case MLX5_IB_QPT_DCI:
2755 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2756 		if (!qp)
2757 			return ERR_PTR(-ENOMEM);
2758 
2759 		err = create_qp_common(dev, pd, init_attr, udata, qp);
2760 		if (err) {
2761 			mlx5_ib_dbg(dev, "create_qp_common failed\n");
2762 			kfree(qp);
2763 			return ERR_PTR(err);
2764 		}
2765 
2766 		if (is_qp0(init_attr->qp_type))
2767 			qp->ibqp.qp_num = 0;
2768 		else if (is_qp1(init_attr->qp_type))
2769 			qp->ibqp.qp_num = 1;
2770 		else
2771 			qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
2772 
2773 		mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
2774 			    qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2775 			    init_attr->recv_cq ? to_mcq(init_attr->recv_cq)->mcq.cqn : -1,
2776 			    init_attr->send_cq ? to_mcq(init_attr->send_cq)->mcq.cqn : -1);
2777 
2778 		qp->trans_qp.xrcdn = xrcdn;
2779 
2780 		break;
2781 
2782 	case IB_QPT_GSI:
2783 		return mlx5_ib_gsi_create_qp(pd, init_attr);
2784 
2785 	case IB_QPT_RAW_IPV6:
2786 	case IB_QPT_RAW_ETHERTYPE:
2787 	case IB_QPT_MAX:
2788 	default:
2789 		mlx5_ib_dbg(dev, "unsupported qp type %d\n",
2790 			    init_attr->qp_type);
2791 		/* Don't support raw QPs */
2792 		return ERR_PTR(-EINVAL);
2793 	}
2794 
2795 	if (verbs_init_attr->qp_type == IB_QPT_DRIVER)
2796 		qp->qp_sub_type = init_attr->qp_type;
2797 
2798 	return &qp->ibqp;
2799 }
2800 
2801 static int mlx5_ib_destroy_dct(struct mlx5_ib_qp *mqp)
2802 {
2803 	struct mlx5_ib_dev *dev = to_mdev(mqp->ibqp.device);
2804 
2805 	if (mqp->state == IB_QPS_RTR) {
2806 		int err;
2807 
2808 		err = mlx5_core_destroy_dct(dev->mdev, &mqp->dct.mdct);
2809 		if (err) {
2810 			mlx5_ib_warn(dev, "failed to destroy DCT %d\n", err);
2811 			return err;
2812 		}
2813 	}
2814 
2815 	kfree(mqp->dct.in);
2816 	kfree(mqp);
2817 	return 0;
2818 }
2819 
2820 int mlx5_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
2821 {
2822 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
2823 	struct mlx5_ib_qp *mqp = to_mqp(qp);
2824 
2825 	if (unlikely(qp->qp_type == IB_QPT_GSI))
2826 		return mlx5_ib_gsi_destroy_qp(qp);
2827 
2828 	if (mqp->qp_sub_type == MLX5_IB_QPT_DCT)
2829 		return mlx5_ib_destroy_dct(mqp);
2830 
2831 	destroy_qp_common(dev, mqp, udata);
2832 
2833 	kfree(mqp);
2834 
2835 	return 0;
2836 }
2837 
2838 static int to_mlx5_access_flags(struct mlx5_ib_qp *qp,
2839 				const struct ib_qp_attr *attr,
2840 				int attr_mask, __be32 *hw_access_flags_be)
2841 {
2842 	u8 dest_rd_atomic;
2843 	u32 access_flags, hw_access_flags = 0;
2844 
2845 	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device);
2846 
2847 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2848 		dest_rd_atomic = attr->max_dest_rd_atomic;
2849 	else
2850 		dest_rd_atomic = qp->trans_qp.resp_depth;
2851 
2852 	if (attr_mask & IB_QP_ACCESS_FLAGS)
2853 		access_flags = attr->qp_access_flags;
2854 	else
2855 		access_flags = qp->trans_qp.atomic_rd_en;
2856 
2857 	if (!dest_rd_atomic)
2858 		access_flags &= IB_ACCESS_REMOTE_WRITE;
2859 
2860 	if (access_flags & IB_ACCESS_REMOTE_READ)
2861 		hw_access_flags |= MLX5_QP_BIT_RRE;
2862 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC) {
2863 		int atomic_mode;
2864 
2865 		atomic_mode = get_atomic_mode(dev, qp->ibqp.qp_type);
2866 		if (atomic_mode < 0)
2867 			return -EOPNOTSUPP;
2868 
2869 		hw_access_flags |= MLX5_QP_BIT_RAE;
2870 		hw_access_flags |= atomic_mode << MLX5_ATOMIC_MODE_OFFSET;
2871 	}
2872 
2873 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
2874 		hw_access_flags |= MLX5_QP_BIT_RWE;
2875 
2876 	*hw_access_flags_be = cpu_to_be32(hw_access_flags);
2877 
2878 	return 0;
2879 }
2880 
2881 enum {
2882 	MLX5_PATH_FLAG_FL	= 1 << 0,
2883 	MLX5_PATH_FLAG_FREE_AR	= 1 << 1,
2884 	MLX5_PATH_FLAG_COUNTER	= 1 << 2,
2885 };
2886 
2887 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
2888 {
2889 	if (rate == IB_RATE_PORT_CURRENT)
2890 		return 0;
2891 
2892 	if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_600_GBPS)
2893 		return -EINVAL;
2894 
2895 	while (rate != IB_RATE_PORT_CURRENT &&
2896 	       !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
2897 		 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
2898 		--rate;
2899 
2900 	return rate ? rate + MLX5_STAT_RATE_OFFSET : rate;
2901 }
2902 
2903 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
2904 				      struct mlx5_ib_sq *sq, u8 sl,
2905 				      struct ib_pd *pd)
2906 {
2907 	void *in;
2908 	void *tisc;
2909 	int inlen;
2910 	int err;
2911 
2912 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2913 	in = kvzalloc(inlen, GFP_KERNEL);
2914 	if (!in)
2915 		return -ENOMEM;
2916 
2917 	MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
2918 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
2919 
2920 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2921 	MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
2922 
2923 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2924 
2925 	kvfree(in);
2926 
2927 	return err;
2928 }
2929 
2930 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
2931 					 struct mlx5_ib_sq *sq, u8 tx_affinity,
2932 					 struct ib_pd *pd)
2933 {
2934 	void *in;
2935 	void *tisc;
2936 	int inlen;
2937 	int err;
2938 
2939 	inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2940 	in = kvzalloc(inlen, GFP_KERNEL);
2941 	if (!in)
2942 		return -ENOMEM;
2943 
2944 	MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
2945 	MLX5_SET(modify_tis_in, in, uid, to_mpd(pd)->uid);
2946 
2947 	tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2948 	MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
2949 
2950 	err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2951 
2952 	kvfree(in);
2953 
2954 	return err;
2955 }
2956 
2957 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2958 			 const struct rdma_ah_attr *ah,
2959 			 struct mlx5_qp_path *path, u8 port, int attr_mask,
2960 			 u32 path_flags, const struct ib_qp_attr *attr,
2961 			 bool alt)
2962 {
2963 	const struct ib_global_route *grh = rdma_ah_read_grh(ah);
2964 	int err;
2965 	enum ib_gid_type gid_type;
2966 	u8 ah_flags = rdma_ah_get_ah_flags(ah);
2967 	u8 sl = rdma_ah_get_sl(ah);
2968 
2969 	if (attr_mask & IB_QP_PKEY_INDEX)
2970 		path->pkey_index = cpu_to_be16(alt ? attr->alt_pkey_index :
2971 						     attr->pkey_index);
2972 
2973 	if (ah_flags & IB_AH_GRH) {
2974 		if (grh->sgid_index >=
2975 		    dev->mdev->port_caps[port - 1].gid_table_len) {
2976 			pr_err("sgid_index (%u) too large. max is %d\n",
2977 			       grh->sgid_index,
2978 			       dev->mdev->port_caps[port - 1].gid_table_len);
2979 			return -EINVAL;
2980 		}
2981 	}
2982 
2983 	if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
2984 		if (!(ah_flags & IB_AH_GRH))
2985 			return -EINVAL;
2986 
2987 		memcpy(path->rmac, ah->roce.dmac, sizeof(ah->roce.dmac));
2988 		if (qp->ibqp.qp_type == IB_QPT_RC ||
2989 		    qp->ibqp.qp_type == IB_QPT_UC ||
2990 		    qp->ibqp.qp_type == IB_QPT_XRC_INI ||
2991 		    qp->ibqp.qp_type == IB_QPT_XRC_TGT)
2992 			path->udp_sport =
2993 				mlx5_get_roce_udp_sport(dev, ah->grh.sgid_attr);
2994 		path->dci_cfi_prio_sl = (sl & 0x7) << 4;
2995 		gid_type = ah->grh.sgid_attr->gid_type;
2996 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
2997 			path->ecn_dscp = (grh->traffic_class >> 2) & 0x3f;
2998 	} else {
2999 		path->fl_free_ar = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
3000 		path->fl_free_ar |=
3001 			(path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x40 : 0;
3002 		path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
3003 		path->grh_mlid = rdma_ah_get_path_bits(ah) & 0x7f;
3004 		if (ah_flags & IB_AH_GRH)
3005 			path->grh_mlid	|= 1 << 7;
3006 		path->dci_cfi_prio_sl = sl & 0xf;
3007 	}
3008 
3009 	if (ah_flags & IB_AH_GRH) {
3010 		path->mgid_index = grh->sgid_index;
3011 		path->hop_limit  = grh->hop_limit;
3012 		path->tclass_flowlabel =
3013 			cpu_to_be32((grh->traffic_class << 20) |
3014 				    (grh->flow_label));
3015 		memcpy(path->rgid, grh->dgid.raw, 16);
3016 	}
3017 
3018 	err = ib_rate_to_mlx5(dev, rdma_ah_get_static_rate(ah));
3019 	if (err < 0)
3020 		return err;
3021 	path->static_rate = err;
3022 	path->port = port;
3023 
3024 	if (attr_mask & IB_QP_TIMEOUT)
3025 		path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) << 3;
3026 
3027 	if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
3028 		return modify_raw_packet_eth_prio(dev->mdev,
3029 						  &qp->raw_packet_qp.sq,
3030 						  sl & 0xf, qp->ibqp.pd);
3031 
3032 	return 0;
3033 }
3034 
3035 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
3036 	[MLX5_QP_STATE_INIT] = {
3037 		[MLX5_QP_STATE_INIT] = {
3038 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3039 					  MLX5_QP_OPTPAR_RAE		|
3040 					  MLX5_QP_OPTPAR_RWE		|
3041 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3042 					  MLX5_QP_OPTPAR_PRI_PORT,
3043 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3044 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3045 					  MLX5_QP_OPTPAR_PRI_PORT,
3046 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3047 					  MLX5_QP_OPTPAR_Q_KEY		|
3048 					  MLX5_QP_OPTPAR_PRI_PORT,
3049 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3050 					  MLX5_QP_OPTPAR_RAE		|
3051 					  MLX5_QP_OPTPAR_RWE		|
3052 					  MLX5_QP_OPTPAR_PKEY_INDEX	|
3053 					  MLX5_QP_OPTPAR_PRI_PORT,
3054 		},
3055 		[MLX5_QP_STATE_RTR] = {
3056 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3057 					  MLX5_QP_OPTPAR_RRE            |
3058 					  MLX5_QP_OPTPAR_RAE            |
3059 					  MLX5_QP_OPTPAR_RWE            |
3060 					  MLX5_QP_OPTPAR_PKEY_INDEX,
3061 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH  |
3062 					  MLX5_QP_OPTPAR_RWE            |
3063 					  MLX5_QP_OPTPAR_PKEY_INDEX,
3064 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX     |
3065 					  MLX5_QP_OPTPAR_Q_KEY,
3066 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX	|
3067 					   MLX5_QP_OPTPAR_Q_KEY,
3068 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
3069 					  MLX5_QP_OPTPAR_RRE            |
3070 					  MLX5_QP_OPTPAR_RAE            |
3071 					  MLX5_QP_OPTPAR_RWE            |
3072 					  MLX5_QP_OPTPAR_PKEY_INDEX,
3073 		},
3074 	},
3075 	[MLX5_QP_STATE_RTR] = {
3076 		[MLX5_QP_STATE_RTS] = {
3077 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3078 					  MLX5_QP_OPTPAR_RRE		|
3079 					  MLX5_QP_OPTPAR_RAE		|
3080 					  MLX5_QP_OPTPAR_RWE		|
3081 					  MLX5_QP_OPTPAR_PM_STATE	|
3082 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3083 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3084 					  MLX5_QP_OPTPAR_RWE		|
3085 					  MLX5_QP_OPTPAR_PM_STATE,
3086 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
3087 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH	|
3088 					  MLX5_QP_OPTPAR_RRE		|
3089 					  MLX5_QP_OPTPAR_RAE		|
3090 					  MLX5_QP_OPTPAR_RWE		|
3091 					  MLX5_QP_OPTPAR_PM_STATE	|
3092 					  MLX5_QP_OPTPAR_RNR_TIMEOUT,
3093 		},
3094 	},
3095 	[MLX5_QP_STATE_RTS] = {
3096 		[MLX5_QP_STATE_RTS] = {
3097 			[MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE		|
3098 					  MLX5_QP_OPTPAR_RAE		|
3099 					  MLX5_QP_OPTPAR_RWE		|
3100 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3101 					  MLX5_QP_OPTPAR_PM_STATE	|
3102 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3103 			[MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE		|
3104 					  MLX5_QP_OPTPAR_PM_STATE	|
3105 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3106 			[MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY		|
3107 					  MLX5_QP_OPTPAR_SRQN		|
3108 					  MLX5_QP_OPTPAR_CQN_RCV,
3109 			[MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_RRE		|
3110 					  MLX5_QP_OPTPAR_RAE		|
3111 					  MLX5_QP_OPTPAR_RWE		|
3112 					  MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3113 					  MLX5_QP_OPTPAR_PM_STATE	|
3114 					  MLX5_QP_OPTPAR_ALT_ADDR_PATH,
3115 		},
3116 	},
3117 	[MLX5_QP_STATE_SQER] = {
3118 		[MLX5_QP_STATE_RTS] = {
3119 			[MLX5_QP_ST_UD]	 = MLX5_QP_OPTPAR_Q_KEY,
3120 			[MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
3121 			[MLX5_QP_ST_UC]	 = MLX5_QP_OPTPAR_RWE,
3122 			[MLX5_QP_ST_RC]	 = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3123 					   MLX5_QP_OPTPAR_RWE		|
3124 					   MLX5_QP_OPTPAR_RAE		|
3125 					   MLX5_QP_OPTPAR_RRE,
3126 			[MLX5_QP_ST_XRC]  = MLX5_QP_OPTPAR_RNR_TIMEOUT	|
3127 					   MLX5_QP_OPTPAR_RWE		|
3128 					   MLX5_QP_OPTPAR_RAE		|
3129 					   MLX5_QP_OPTPAR_RRE,
3130 		},
3131 	},
3132 };
3133 
3134 static int ib_nr_to_mlx5_nr(int ib_mask)
3135 {
3136 	switch (ib_mask) {
3137 	case IB_QP_STATE:
3138 		return 0;
3139 	case IB_QP_CUR_STATE:
3140 		return 0;
3141 	case IB_QP_EN_SQD_ASYNC_NOTIFY:
3142 		return 0;
3143 	case IB_QP_ACCESS_FLAGS:
3144 		return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
3145 			MLX5_QP_OPTPAR_RAE;
3146 	case IB_QP_PKEY_INDEX:
3147 		return MLX5_QP_OPTPAR_PKEY_INDEX;
3148 	case IB_QP_PORT:
3149 		return MLX5_QP_OPTPAR_PRI_PORT;
3150 	case IB_QP_QKEY:
3151 		return MLX5_QP_OPTPAR_Q_KEY;
3152 	case IB_QP_AV:
3153 		return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
3154 			MLX5_QP_OPTPAR_PRI_PORT;
3155 	case IB_QP_PATH_MTU:
3156 		return 0;
3157 	case IB_QP_TIMEOUT:
3158 		return MLX5_QP_OPTPAR_ACK_TIMEOUT;
3159 	case IB_QP_RETRY_CNT:
3160 		return MLX5_QP_OPTPAR_RETRY_COUNT;
3161 	case IB_QP_RNR_RETRY:
3162 		return MLX5_QP_OPTPAR_RNR_RETRY;
3163 	case IB_QP_RQ_PSN:
3164 		return 0;
3165 	case IB_QP_MAX_QP_RD_ATOMIC:
3166 		return MLX5_QP_OPTPAR_SRA_MAX;
3167 	case IB_QP_ALT_PATH:
3168 		return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
3169 	case IB_QP_MIN_RNR_TIMER:
3170 		return MLX5_QP_OPTPAR_RNR_TIMEOUT;
3171 	case IB_QP_SQ_PSN:
3172 		return 0;
3173 	case IB_QP_MAX_DEST_RD_ATOMIC:
3174 		return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
3175 			MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
3176 	case IB_QP_PATH_MIG_STATE:
3177 		return MLX5_QP_OPTPAR_PM_STATE;
3178 	case IB_QP_CAP:
3179 		return 0;
3180 	case IB_QP_DEST_QPN:
3181 		return 0;
3182 	}
3183 	return 0;
3184 }
3185 
3186 static int ib_mask_to_mlx5_opt(int ib_mask)
3187 {
3188 	int result = 0;
3189 	int i;
3190 
3191 	for (i = 0; i < 8 * sizeof(int); i++) {
3192 		if ((1 << i) & ib_mask)
3193 			result |= ib_nr_to_mlx5_nr(1 << i);
3194 	}
3195 
3196 	return result;
3197 }
3198 
3199 static int modify_raw_packet_qp_rq(
3200 	struct mlx5_ib_dev *dev, struct mlx5_ib_rq *rq, int new_state,
3201 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3202 {
3203 	void *in;
3204 	void *rqc;
3205 	int inlen;
3206 	int err;
3207 
3208 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
3209 	in = kvzalloc(inlen, GFP_KERNEL);
3210 	if (!in)
3211 		return -ENOMEM;
3212 
3213 	MLX5_SET(modify_rq_in, in, rq_state, rq->state);
3214 	MLX5_SET(modify_rq_in, in, uid, to_mpd(pd)->uid);
3215 
3216 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
3217 	MLX5_SET(rqc, rqc, state, new_state);
3218 
3219 	if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
3220 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
3221 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
3222 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
3223 			MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
3224 		} else
3225 			dev_info_once(
3226 				&dev->ib_dev.dev,
3227 				"RAW PACKET QP counters are not supported on current FW\n");
3228 	}
3229 
3230 	err = mlx5_core_modify_rq(dev->mdev, rq->base.mqp.qpn, in, inlen);
3231 	if (err)
3232 		goto out;
3233 
3234 	rq->state = new_state;
3235 
3236 out:
3237 	kvfree(in);
3238 	return err;
3239 }
3240 
3241 static int modify_raw_packet_qp_sq(
3242 	struct mlx5_core_dev *dev, struct mlx5_ib_sq *sq, int new_state,
3243 	const struct mlx5_modify_raw_qp_param *raw_qp_param, struct ib_pd *pd)
3244 {
3245 	struct mlx5_ib_qp *ibqp = sq->base.container_mibqp;
3246 	struct mlx5_rate_limit old_rl = ibqp->rl;
3247 	struct mlx5_rate_limit new_rl = old_rl;
3248 	bool new_rate_added = false;
3249 	u16 rl_index = 0;
3250 	void *in;
3251 	void *sqc;
3252 	int inlen;
3253 	int err;
3254 
3255 	inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
3256 	in = kvzalloc(inlen, GFP_KERNEL);
3257 	if (!in)
3258 		return -ENOMEM;
3259 
3260 	MLX5_SET(modify_sq_in, in, uid, to_mpd(pd)->uid);
3261 	MLX5_SET(modify_sq_in, in, sq_state, sq->state);
3262 
3263 	sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
3264 	MLX5_SET(sqc, sqc, state, new_state);
3265 
3266 	if (raw_qp_param->set_mask & MLX5_RAW_QP_RATE_LIMIT) {
3267 		if (new_state != MLX5_SQC_STATE_RDY)
3268 			pr_warn("%s: Rate limit can only be changed when SQ is moving to RDY\n",
3269 				__func__);
3270 		else
3271 			new_rl = raw_qp_param->rl;
3272 	}
3273 
3274 	if (!mlx5_rl_are_equal(&old_rl, &new_rl)) {
3275 		if (new_rl.rate) {
3276 			err = mlx5_rl_add_rate(dev, &rl_index, &new_rl);
3277 			if (err) {
3278 				pr_err("Failed configuring rate limit(err %d): \
3279 				       rate %u, max_burst_sz %u, typical_pkt_sz %u\n",
3280 				       err, new_rl.rate, new_rl.max_burst_sz,
3281 				       new_rl.typical_pkt_sz);
3282 
3283 				goto out;
3284 			}
3285 			new_rate_added = true;
3286 		}
3287 
3288 		MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
3289 		/* index 0 means no limit */
3290 		MLX5_SET(sqc, sqc, packet_pacing_rate_limit_index, rl_index);
3291 	}
3292 
3293 	err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in, inlen);
3294 	if (err) {
3295 		/* Remove new rate from table if failed */
3296 		if (new_rate_added)
3297 			mlx5_rl_remove_rate(dev, &new_rl);
3298 		goto out;
3299 	}
3300 
3301 	/* Only remove the old rate after new rate was set */
3302 	if ((old_rl.rate && !mlx5_rl_are_equal(&old_rl, &new_rl)) ||
3303 	    (new_state != MLX5_SQC_STATE_RDY)) {
3304 		mlx5_rl_remove_rate(dev, &old_rl);
3305 		if (new_state != MLX5_SQC_STATE_RDY)
3306 			memset(&new_rl, 0, sizeof(new_rl));
3307 	}
3308 
3309 	ibqp->rl = new_rl;
3310 	sq->state = new_state;
3311 
3312 out:
3313 	kvfree(in);
3314 	return err;
3315 }
3316 
3317 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
3318 				const struct mlx5_modify_raw_qp_param *raw_qp_param,
3319 				u8 tx_affinity)
3320 {
3321 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
3322 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
3323 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
3324 	int modify_rq = !!qp->rq.wqe_cnt;
3325 	int modify_sq = !!qp->sq.wqe_cnt;
3326 	int rq_state;
3327 	int sq_state;
3328 	int err;
3329 
3330 	switch (raw_qp_param->operation) {
3331 	case MLX5_CMD_OP_RST2INIT_QP:
3332 		rq_state = MLX5_RQC_STATE_RDY;
3333 		sq_state = MLX5_SQC_STATE_RDY;
3334 		break;
3335 	case MLX5_CMD_OP_2ERR_QP:
3336 		rq_state = MLX5_RQC_STATE_ERR;
3337 		sq_state = MLX5_SQC_STATE_ERR;
3338 		break;
3339 	case MLX5_CMD_OP_2RST_QP:
3340 		rq_state = MLX5_RQC_STATE_RST;
3341 		sq_state = MLX5_SQC_STATE_RST;
3342 		break;
3343 	case MLX5_CMD_OP_RTR2RTS_QP:
3344 	case MLX5_CMD_OP_RTS2RTS_QP:
3345 		if (raw_qp_param->set_mask ==
3346 		    MLX5_RAW_QP_RATE_LIMIT) {
3347 			modify_rq = 0;
3348 			sq_state = sq->state;
3349 		} else {
3350 			return raw_qp_param->set_mask ? -EINVAL : 0;
3351 		}
3352 		break;
3353 	case MLX5_CMD_OP_INIT2INIT_QP:
3354 	case MLX5_CMD_OP_INIT2RTR_QP:
3355 		if (raw_qp_param->set_mask)
3356 			return -EINVAL;
3357 		else
3358 			return 0;
3359 	default:
3360 		WARN_ON(1);
3361 		return -EINVAL;
3362 	}
3363 
3364 	if (modify_rq) {
3365 		err =  modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param,
3366 					       qp->ibqp.pd);
3367 		if (err)
3368 			return err;
3369 	}
3370 
3371 	if (modify_sq) {
3372 		struct mlx5_flow_handle *flow_rule;
3373 
3374 		if (tx_affinity) {
3375 			err = modify_raw_packet_tx_affinity(dev->mdev, sq,
3376 							    tx_affinity,
3377 							    qp->ibqp.pd);
3378 			if (err)
3379 				return err;
3380 		}
3381 
3382 		flow_rule = create_flow_rule_vport_sq(dev, sq,
3383 						      raw_qp_param->port);
3384 		if (IS_ERR(flow_rule))
3385 			return PTR_ERR(flow_rule);
3386 
3387 		err = modify_raw_packet_qp_sq(dev->mdev, sq, sq_state,
3388 					      raw_qp_param, qp->ibqp.pd);
3389 		if (err) {
3390 			if (flow_rule)
3391 				mlx5_del_flow_rules(flow_rule);
3392 			return err;
3393 		}
3394 
3395 		if (flow_rule) {
3396 			destroy_flow_rule_vport_sq(sq);
3397 			sq->flow_rule = flow_rule;
3398 		}
3399 
3400 		return err;
3401 	}
3402 
3403 	return 0;
3404 }
3405 
3406 static unsigned int get_tx_affinity(struct mlx5_ib_dev *dev,
3407 				    struct mlx5_ib_pd *pd,
3408 				    struct mlx5_ib_qp_base *qp_base,
3409 				    u8 port_num, struct ib_udata *udata)
3410 {
3411 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
3412 		udata, struct mlx5_ib_ucontext, ibucontext);
3413 	unsigned int tx_port_affinity;
3414 
3415 	if (ucontext) {
3416 		tx_port_affinity = (unsigned int)atomic_add_return(
3417 					   1, &ucontext->tx_port_affinity) %
3418 					   MLX5_MAX_PORTS +
3419 				   1;
3420 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x ucontext %p\n",
3421 				tx_port_affinity, qp_base->mqp.qpn, ucontext);
3422 	} else {
3423 		tx_port_affinity =
3424 			(unsigned int)atomic_add_return(
3425 				1, &dev->port[port_num].roce.tx_port_affinity) %
3426 				MLX5_MAX_PORTS +
3427 			1;
3428 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x\n",
3429 				tx_port_affinity, qp_base->mqp.qpn);
3430 	}
3431 
3432 	return tx_port_affinity;
3433 }
3434 
3435 static int __mlx5_ib_qp_set_counter(struct ib_qp *qp,
3436 				    struct rdma_counter *counter)
3437 {
3438 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
3439 	struct mlx5_ib_qp *mqp = to_mqp(qp);
3440 	struct mlx5_qp_context context = {};
3441 	struct mlx5_ib_qp_base *base;
3442 	u32 set_id;
3443 
3444 	if (!MLX5_CAP_GEN(dev->mdev, rts2rts_qp_counters_set_id))
3445 		return 0;
3446 
3447 	if (counter)
3448 		set_id = counter->id;
3449 	else
3450 		set_id = mlx5_ib_get_counters_id(dev, mqp->port - 1);
3451 
3452 	base = &mqp->trans_qp.base;
3453 	context.qp_counter_set_usr_page &= cpu_to_be32(0xffffff);
3454 	context.qp_counter_set_usr_page |= cpu_to_be32(set_id << 24);
3455 	return mlx5_core_qp_modify(dev->mdev,
3456 				   MLX5_CMD_OP_RTS2RTS_QP,
3457 				   MLX5_QP_OPTPAR_COUNTER_SET_ID,
3458 				   &context, &base->mqp);
3459 }
3460 
3461 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
3462 			       const struct ib_qp_attr *attr, int attr_mask,
3463 			       enum ib_qp_state cur_state,
3464 			       enum ib_qp_state new_state,
3465 			       const struct mlx5_ib_modify_qp *ucmd,
3466 			       struct ib_udata *udata)
3467 {
3468 	static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
3469 		[MLX5_QP_STATE_RST] = {
3470 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3471 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3472 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_RST2INIT_QP,
3473 		},
3474 		[MLX5_QP_STATE_INIT]  = {
3475 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3476 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3477 			[MLX5_QP_STATE_INIT]	= MLX5_CMD_OP_INIT2INIT_QP,
3478 			[MLX5_QP_STATE_RTR]	= MLX5_CMD_OP_INIT2RTR_QP,
3479 		},
3480 		[MLX5_QP_STATE_RTR]   = {
3481 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3482 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3483 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTR2RTS_QP,
3484 		},
3485 		[MLX5_QP_STATE_RTS]   = {
3486 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3487 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3488 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_RTS2RTS_QP,
3489 		},
3490 		[MLX5_QP_STATE_SQD] = {
3491 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3492 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3493 		},
3494 		[MLX5_QP_STATE_SQER] = {
3495 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3496 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3497 			[MLX5_QP_STATE_RTS]	= MLX5_CMD_OP_SQERR2RTS_QP,
3498 		},
3499 		[MLX5_QP_STATE_ERR] = {
3500 			[MLX5_QP_STATE_RST]	= MLX5_CMD_OP_2RST_QP,
3501 			[MLX5_QP_STATE_ERR]	= MLX5_CMD_OP_2ERR_QP,
3502 		}
3503 	};
3504 
3505 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3506 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3507 	struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
3508 	struct mlx5_ib_cq *send_cq, *recv_cq;
3509 	struct mlx5_qp_context *context;
3510 	struct mlx5_ib_pd *pd;
3511 	enum mlx5_qp_state mlx5_cur, mlx5_new;
3512 	enum mlx5_qp_optpar optpar;
3513 	u32 set_id = 0;
3514 	int mlx5_st;
3515 	int err;
3516 	u16 op;
3517 	u8 tx_affinity = 0;
3518 
3519 	mlx5_st = to_mlx5_st(ibqp->qp_type == IB_QPT_DRIVER ?
3520 			     qp->qp_sub_type : ibqp->qp_type);
3521 	if (mlx5_st < 0)
3522 		return -EINVAL;
3523 
3524 	context = kzalloc(sizeof(*context), GFP_KERNEL);
3525 	if (!context)
3526 		return -ENOMEM;
3527 
3528 	pd = get_pd(qp);
3529 	context->flags = cpu_to_be32(mlx5_st << 16);
3530 
3531 	if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
3532 		context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
3533 	} else {
3534 		switch (attr->path_mig_state) {
3535 		case IB_MIG_MIGRATED:
3536 			context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
3537 			break;
3538 		case IB_MIG_REARM:
3539 			context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
3540 			break;
3541 		case IB_MIG_ARMED:
3542 			context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
3543 			break;
3544 		}
3545 	}
3546 
3547 	if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
3548 		if ((ibqp->qp_type == IB_QPT_RC) ||
3549 		    (ibqp->qp_type == IB_QPT_UD &&
3550 		     !(qp->flags & MLX5_IB_QP_SQPN_QP1)) ||
3551 		    (ibqp->qp_type == IB_QPT_UC) ||
3552 		    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
3553 		    (ibqp->qp_type == IB_QPT_XRC_INI) ||
3554 		    (ibqp->qp_type == IB_QPT_XRC_TGT)) {
3555 			if (dev->lag_active) {
3556 				u8 p = mlx5_core_native_port_num(dev->mdev) - 1;
3557 				tx_affinity = get_tx_affinity(dev, pd, base, p,
3558 							      udata);
3559 				context->flags |= cpu_to_be32(tx_affinity << 24);
3560 			}
3561 		}
3562 	}
3563 
3564 	if (is_sqp(ibqp->qp_type)) {
3565 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
3566 	} else if ((ibqp->qp_type == IB_QPT_UD &&
3567 		    !(qp->flags & MLX5_IB_QP_UNDERLAY)) ||
3568 		   ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
3569 		context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
3570 	} else if (attr_mask & IB_QP_PATH_MTU) {
3571 		if (attr->path_mtu < IB_MTU_256 ||
3572 		    attr->path_mtu > IB_MTU_4096) {
3573 			mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
3574 			err = -EINVAL;
3575 			goto out;
3576 		}
3577 		context->mtu_msgmax = (attr->path_mtu << 5) |
3578 				      (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
3579 	}
3580 
3581 	if (attr_mask & IB_QP_DEST_QPN)
3582 		context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
3583 
3584 	if (attr_mask & IB_QP_PKEY_INDEX)
3585 		context->pri_path.pkey_index = cpu_to_be16(attr->pkey_index);
3586 
3587 	/* todo implement counter_index functionality */
3588 
3589 	if (is_sqp(ibqp->qp_type))
3590 		context->pri_path.port = qp->port;
3591 
3592 	if (attr_mask & IB_QP_PORT)
3593 		context->pri_path.port = attr->port_num;
3594 
3595 	if (attr_mask & IB_QP_AV) {
3596 		err = mlx5_set_path(dev, qp, &attr->ah_attr, &context->pri_path,
3597 				    attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
3598 				    attr_mask, 0, attr, false);
3599 		if (err)
3600 			goto out;
3601 	}
3602 
3603 	if (attr_mask & IB_QP_TIMEOUT)
3604 		context->pri_path.ackto_lt |= attr->timeout << 3;
3605 
3606 	if (attr_mask & IB_QP_ALT_PATH) {
3607 		err = mlx5_set_path(dev, qp, &attr->alt_ah_attr,
3608 				    &context->alt_path,
3609 				    attr->alt_port_num,
3610 				    attr_mask | IB_QP_PKEY_INDEX | IB_QP_TIMEOUT,
3611 				    0, attr, true);
3612 		if (err)
3613 			goto out;
3614 	}
3615 
3616 	get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
3617 		&send_cq, &recv_cq);
3618 
3619 	context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
3620 	context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
3621 	context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
3622 	context->params1  = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
3623 
3624 	if (attr_mask & IB_QP_RNR_RETRY)
3625 		context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
3626 
3627 	if (attr_mask & IB_QP_RETRY_CNT)
3628 		context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
3629 
3630 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
3631 		if (attr->max_rd_atomic)
3632 			context->params1 |=
3633 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
3634 	}
3635 
3636 	if (attr_mask & IB_QP_SQ_PSN)
3637 		context->next_send_psn = cpu_to_be32(attr->sq_psn);
3638 
3639 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
3640 		if (attr->max_dest_rd_atomic)
3641 			context->params2 |=
3642 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
3643 	}
3644 
3645 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
3646 		__be32 access_flags;
3647 
3648 		err = to_mlx5_access_flags(qp, attr, attr_mask, &access_flags);
3649 		if (err)
3650 			goto out;
3651 
3652 		context->params2 |= access_flags;
3653 	}
3654 
3655 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
3656 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
3657 
3658 	if (attr_mask & IB_QP_RQ_PSN)
3659 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
3660 
3661 	if (attr_mask & IB_QP_QKEY)
3662 		context->qkey = cpu_to_be32(attr->qkey);
3663 
3664 	if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
3665 		context->db_rec_addr = cpu_to_be64(qp->db.dma);
3666 
3667 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3668 		u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
3669 			       qp->port) - 1;
3670 
3671 		/* Underlay port should be used - index 0 function per port */
3672 		if (qp->flags & MLX5_IB_QP_UNDERLAY)
3673 			port_num = 0;
3674 
3675 		if (ibqp->counter)
3676 			set_id = ibqp->counter->id;
3677 		else
3678 			set_id = mlx5_ib_get_counters_id(dev, port_num);
3679 		context->qp_counter_set_usr_page |=
3680 			cpu_to_be32(set_id << 24);
3681 	}
3682 
3683 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
3684 		context->sq_crq_size |= cpu_to_be16(1 << 4);
3685 
3686 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
3687 		context->deth_sqpn = cpu_to_be32(1);
3688 
3689 	mlx5_cur = to_mlx5_state(cur_state);
3690 	mlx5_new = to_mlx5_state(new_state);
3691 
3692 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
3693 	    !optab[mlx5_cur][mlx5_new]) {
3694 		err = -EINVAL;
3695 		goto out;
3696 	}
3697 
3698 	op = optab[mlx5_cur][mlx5_new];
3699 	optpar = ib_mask_to_mlx5_opt(attr_mask);
3700 	optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
3701 
3702 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
3703 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
3704 		struct mlx5_modify_raw_qp_param raw_qp_param = {};
3705 
3706 		raw_qp_param.operation = op;
3707 		if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3708 			raw_qp_param.rq_q_ctr_id = set_id;
3709 			raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
3710 		}
3711 
3712 		if (attr_mask & IB_QP_PORT)
3713 			raw_qp_param.port = attr->port_num;
3714 
3715 		if (attr_mask & IB_QP_RATE_LIMIT) {
3716 			raw_qp_param.rl.rate = attr->rate_limit;
3717 
3718 			if (ucmd->burst_info.max_burst_sz) {
3719 				if (attr->rate_limit &&
3720 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_burst_bound)) {
3721 					raw_qp_param.rl.max_burst_sz =
3722 						ucmd->burst_info.max_burst_sz;
3723 				} else {
3724 					err = -EINVAL;
3725 					goto out;
3726 				}
3727 			}
3728 
3729 			if (ucmd->burst_info.typical_pkt_sz) {
3730 				if (attr->rate_limit &&
3731 				    MLX5_CAP_QOS(dev->mdev, packet_pacing_typical_size)) {
3732 					raw_qp_param.rl.typical_pkt_sz =
3733 						ucmd->burst_info.typical_pkt_sz;
3734 				} else {
3735 					err = -EINVAL;
3736 					goto out;
3737 				}
3738 			}
3739 
3740 			raw_qp_param.set_mask |= MLX5_RAW_QP_RATE_LIMIT;
3741 		}
3742 
3743 		err = modify_raw_packet_qp(dev, qp, &raw_qp_param, tx_affinity);
3744 	} else {
3745 		err = mlx5_core_qp_modify(dev->mdev, op, optpar, context,
3746 					  &base->mqp);
3747 	}
3748 
3749 	if (err)
3750 		goto out;
3751 
3752 	qp->state = new_state;
3753 
3754 	if (attr_mask & IB_QP_ACCESS_FLAGS)
3755 		qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
3756 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
3757 		qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
3758 	if (attr_mask & IB_QP_PORT)
3759 		qp->port = attr->port_num;
3760 	if (attr_mask & IB_QP_ALT_PATH)
3761 		qp->trans_qp.alt_port = attr->alt_port_num;
3762 
3763 	/*
3764 	 * If we moved a kernel QP to RESET, clean up all old CQ
3765 	 * entries and reinitialize the QP.
3766 	 */
3767 	if (new_state == IB_QPS_RESET &&
3768 	    !ibqp->uobject && ibqp->qp_type != IB_QPT_XRC_TGT) {
3769 		mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
3770 				 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
3771 		if (send_cq != recv_cq)
3772 			mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
3773 
3774 		qp->rq.head = 0;
3775 		qp->rq.tail = 0;
3776 		qp->sq.head = 0;
3777 		qp->sq.tail = 0;
3778 		qp->sq.cur_post = 0;
3779 		if (qp->sq.wqe_cnt)
3780 			qp->sq.cur_edge = get_sq_edge(&qp->sq, 0);
3781 		qp->db.db[MLX5_RCV_DBR] = 0;
3782 		qp->db.db[MLX5_SND_DBR] = 0;
3783 	}
3784 
3785 	if ((new_state == IB_QPS_RTS) && qp->counter_pending) {
3786 		err = __mlx5_ib_qp_set_counter(ibqp, ibqp->counter);
3787 		if (!err)
3788 			qp->counter_pending = 0;
3789 	}
3790 
3791 out:
3792 	kfree(context);
3793 	return err;
3794 }
3795 
3796 static inline bool is_valid_mask(int mask, int req, int opt)
3797 {
3798 	if ((mask & req) != req)
3799 		return false;
3800 
3801 	if (mask & ~(req | opt))
3802 		return false;
3803 
3804 	return true;
3805 }
3806 
3807 /* check valid transition for driver QP types
3808  * for now the only QP type that this function supports is DCI
3809  */
3810 static bool modify_dci_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state new_state,
3811 				enum ib_qp_attr_mask attr_mask)
3812 {
3813 	int req = IB_QP_STATE;
3814 	int opt = 0;
3815 
3816 	if (new_state == IB_QPS_RESET) {
3817 		return is_valid_mask(attr_mask, req, opt);
3818 	} else if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3819 		req |= IB_QP_PKEY_INDEX | IB_QP_PORT;
3820 		return is_valid_mask(attr_mask, req, opt);
3821 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
3822 		opt = IB_QP_PKEY_INDEX | IB_QP_PORT;
3823 		return is_valid_mask(attr_mask, req, opt);
3824 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
3825 		req |= IB_QP_PATH_MTU;
3826 		opt = IB_QP_PKEY_INDEX | IB_QP_AV;
3827 		return is_valid_mask(attr_mask, req, opt);
3828 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
3829 		req |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
3830 		       IB_QP_MAX_QP_RD_ATOMIC | IB_QP_SQ_PSN;
3831 		opt = IB_QP_MIN_RNR_TIMER;
3832 		return is_valid_mask(attr_mask, req, opt);
3833 	} else if (cur_state == IB_QPS_RTS && new_state == IB_QPS_RTS) {
3834 		opt = IB_QP_MIN_RNR_TIMER;
3835 		return is_valid_mask(attr_mask, req, opt);
3836 	} else if (cur_state != IB_QPS_RESET && new_state == IB_QPS_ERR) {
3837 		return is_valid_mask(attr_mask, req, opt);
3838 	}
3839 	return false;
3840 }
3841 
3842 /* mlx5_ib_modify_dct: modify a DCT QP
3843  * valid transitions are:
3844  * RESET to INIT: must set access_flags, pkey_index and port
3845  * INIT  to RTR : must set min_rnr_timer, tclass, flow_label,
3846  *			   mtu, gid_index and hop_limit
3847  * Other transitions and attributes are illegal
3848  */
3849 static int mlx5_ib_modify_dct(struct ib_qp *ibqp, struct ib_qp_attr *attr,
3850 			      int attr_mask, struct ib_udata *udata)
3851 {
3852 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3853 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3854 	enum ib_qp_state cur_state, new_state;
3855 	int err = 0;
3856 	int required = IB_QP_STATE;
3857 	void *dctc;
3858 
3859 	if (!(attr_mask & IB_QP_STATE))
3860 		return -EINVAL;
3861 
3862 	cur_state = qp->state;
3863 	new_state = attr->qp_state;
3864 
3865 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
3866 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
3867 		u16 set_id;
3868 
3869 		required |= IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
3870 		if (!is_valid_mask(attr_mask, required, 0))
3871 			return -EINVAL;
3872 
3873 		if (attr->port_num == 0 ||
3874 		    attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)) {
3875 			mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
3876 				    attr->port_num, dev->num_ports);
3877 			return -EINVAL;
3878 		}
3879 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
3880 			MLX5_SET(dctc, dctc, rre, 1);
3881 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
3882 			MLX5_SET(dctc, dctc, rwe, 1);
3883 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC) {
3884 			int atomic_mode;
3885 
3886 			atomic_mode = get_atomic_mode(dev, MLX5_IB_QPT_DCT);
3887 			if (atomic_mode < 0)
3888 				return -EOPNOTSUPP;
3889 
3890 			MLX5_SET(dctc, dctc, atomic_mode, atomic_mode);
3891 			MLX5_SET(dctc, dctc, rae, 1);
3892 		}
3893 		MLX5_SET(dctc, dctc, pkey_index, attr->pkey_index);
3894 		MLX5_SET(dctc, dctc, port, attr->port_num);
3895 
3896 		set_id = mlx5_ib_get_counters_id(dev, attr->port_num - 1);
3897 		MLX5_SET(dctc, dctc, counter_set_id, set_id);
3898 
3899 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
3900 		struct mlx5_ib_modify_qp_resp resp = {};
3901 		u32 out[MLX5_ST_SZ_DW(create_dct_out)] = {0};
3902 		u32 min_resp_len = offsetof(typeof(resp), dctn) +
3903 				   sizeof(resp.dctn);
3904 
3905 		if (udata->outlen < min_resp_len)
3906 			return -EINVAL;
3907 		resp.response_length = min_resp_len;
3908 
3909 		required |= IB_QP_MIN_RNR_TIMER | IB_QP_AV | IB_QP_PATH_MTU;
3910 		if (!is_valid_mask(attr_mask, required, 0))
3911 			return -EINVAL;
3912 		MLX5_SET(dctc, dctc, min_rnr_nak, attr->min_rnr_timer);
3913 		MLX5_SET(dctc, dctc, tclass, attr->ah_attr.grh.traffic_class);
3914 		MLX5_SET(dctc, dctc, flow_label, attr->ah_attr.grh.flow_label);
3915 		MLX5_SET(dctc, dctc, mtu, attr->path_mtu);
3916 		MLX5_SET(dctc, dctc, my_addr_index, attr->ah_attr.grh.sgid_index);
3917 		MLX5_SET(dctc, dctc, hop_limit, attr->ah_attr.grh.hop_limit);
3918 
3919 		err = mlx5_core_create_dct(dev->mdev, &qp->dct.mdct, qp->dct.in,
3920 					   MLX5_ST_SZ_BYTES(create_dct_in), out,
3921 					   sizeof(out));
3922 		if (err)
3923 			return err;
3924 		resp.dctn = qp->dct.mdct.mqp.qpn;
3925 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
3926 		if (err) {
3927 			mlx5_core_destroy_dct(dev->mdev, &qp->dct.mdct);
3928 			return err;
3929 		}
3930 	} else {
3931 		mlx5_ib_warn(dev, "Modify DCT: Invalid transition from %d to %d\n", cur_state, new_state);
3932 		return -EINVAL;
3933 	}
3934 	if (err)
3935 		qp->state = IB_QPS_ERR;
3936 	else
3937 		qp->state = new_state;
3938 	return err;
3939 }
3940 
3941 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
3942 		      int attr_mask, struct ib_udata *udata)
3943 {
3944 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3945 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
3946 	struct mlx5_ib_modify_qp ucmd = {};
3947 	enum ib_qp_type qp_type;
3948 	enum ib_qp_state cur_state, new_state;
3949 	size_t required_cmd_sz;
3950 	int err = -EINVAL;
3951 	int port;
3952 
3953 	if (ibqp->rwq_ind_tbl)
3954 		return -ENOSYS;
3955 
3956 	if (udata && udata->inlen) {
3957 		required_cmd_sz = offsetof(typeof(ucmd), reserved) +
3958 			sizeof(ucmd.reserved);
3959 		if (udata->inlen < required_cmd_sz)
3960 			return -EINVAL;
3961 
3962 		if (udata->inlen > sizeof(ucmd) &&
3963 		    !ib_is_udata_cleared(udata, sizeof(ucmd),
3964 					 udata->inlen - sizeof(ucmd)))
3965 			return -EOPNOTSUPP;
3966 
3967 		if (ib_copy_from_udata(&ucmd, udata,
3968 				       min(udata->inlen, sizeof(ucmd))))
3969 			return -EFAULT;
3970 
3971 		if (ucmd.comp_mask ||
3972 		    memchr_inv(&ucmd.reserved, 0, sizeof(ucmd.reserved)) ||
3973 		    memchr_inv(&ucmd.burst_info.reserved, 0,
3974 			       sizeof(ucmd.burst_info.reserved)))
3975 			return -EOPNOTSUPP;
3976 	}
3977 
3978 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3979 		return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
3980 
3981 	if (ibqp->qp_type == IB_QPT_DRIVER)
3982 		qp_type = qp->qp_sub_type;
3983 	else
3984 		qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
3985 			IB_QPT_GSI : ibqp->qp_type;
3986 
3987 	if (qp_type == MLX5_IB_QPT_DCT)
3988 		return mlx5_ib_modify_dct(ibqp, attr, attr_mask, udata);
3989 
3990 	mutex_lock(&qp->mutex);
3991 
3992 	cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
3993 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
3994 
3995 	if (!(cur_state == new_state && cur_state == IB_QPS_RESET)) {
3996 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
3997 	}
3998 
3999 	if (qp->flags & MLX5_IB_QP_UNDERLAY) {
4000 		if (attr_mask & ~(IB_QP_STATE | IB_QP_CUR_STATE)) {
4001 			mlx5_ib_dbg(dev, "invalid attr_mask 0x%x when underlay QP is used\n",
4002 				    attr_mask);
4003 			goto out;
4004 		}
4005 	} else if (qp_type != MLX5_IB_QPT_REG_UMR &&
4006 		   qp_type != MLX5_IB_QPT_DCI &&
4007 		   !ib_modify_qp_is_ok(cur_state, new_state, qp_type,
4008 				       attr_mask)) {
4009 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4010 			    cur_state, new_state, ibqp->qp_type, attr_mask);
4011 		goto out;
4012 	} else if (qp_type == MLX5_IB_QPT_DCI &&
4013 		   !modify_dci_qp_is_ok(cur_state, new_state, attr_mask)) {
4014 		mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
4015 			    cur_state, new_state, qp_type, attr_mask);
4016 		goto out;
4017 	}
4018 
4019 	if ((attr_mask & IB_QP_PORT) &&
4020 	    (attr->port_num == 0 ||
4021 	     attr->port_num > dev->num_ports)) {
4022 		mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
4023 			    attr->port_num, dev->num_ports);
4024 		goto out;
4025 	}
4026 
4027 	if (attr_mask & IB_QP_PKEY_INDEX) {
4028 		port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
4029 		if (attr->pkey_index >=
4030 		    dev->mdev->port_caps[port - 1].pkey_table_len) {
4031 			mlx5_ib_dbg(dev, "invalid pkey index %d\n",
4032 				    attr->pkey_index);
4033 			goto out;
4034 		}
4035 	}
4036 
4037 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
4038 	    attr->max_rd_atomic >
4039 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
4040 		mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
4041 			    attr->max_rd_atomic);
4042 		goto out;
4043 	}
4044 
4045 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
4046 	    attr->max_dest_rd_atomic >
4047 	    (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
4048 		mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
4049 			    attr->max_dest_rd_atomic);
4050 		goto out;
4051 	}
4052 
4053 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
4054 		err = 0;
4055 		goto out;
4056 	}
4057 
4058 	err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state,
4059 				  new_state, &ucmd, udata);
4060 
4061 out:
4062 	mutex_unlock(&qp->mutex);
4063 	return err;
4064 }
4065 
4066 static void _handle_post_send_edge(struct mlx5_ib_wq *sq, void **seg,
4067 				   u32 wqe_sz, void **cur_edge)
4068 {
4069 	u32 idx;
4070 
4071 	idx = (sq->cur_post + (wqe_sz >> 2)) & (sq->wqe_cnt - 1);
4072 	*cur_edge = get_sq_edge(sq, idx);
4073 
4074 	*seg = mlx5_frag_buf_get_wqe(&sq->fbc, idx);
4075 }
4076 
4077 /* handle_post_send_edge - Check if we get to SQ edge. If yes, update to the
4078  * next nearby edge and get new address translation for current WQE position.
4079  * @sq - SQ buffer.
4080  * @seg: Current WQE position (16B aligned).
4081  * @wqe_sz: Total current WQE size [16B].
4082  * @cur_edge: Updated current edge.
4083  */
4084 static inline void handle_post_send_edge(struct mlx5_ib_wq *sq, void **seg,
4085 					 u32 wqe_sz, void **cur_edge)
4086 {
4087 	if (likely(*seg != *cur_edge))
4088 		return;
4089 
4090 	_handle_post_send_edge(sq, seg, wqe_sz, cur_edge);
4091 }
4092 
4093 /* memcpy_send_wqe - copy data from src to WQE and update the relevant WQ's
4094  * pointers. At the end @seg is aligned to 16B regardless the copied size.
4095  * @sq - SQ buffer.
4096  * @cur_edge: Updated current edge.
4097  * @seg: Current WQE position (16B aligned).
4098  * @wqe_sz: Total current WQE size [16B].
4099  * @src: Pointer to copy from.
4100  * @n: Number of bytes to copy.
4101  */
4102 static inline void memcpy_send_wqe(struct mlx5_ib_wq *sq, void **cur_edge,
4103 				   void **seg, u32 *wqe_sz, const void *src,
4104 				   size_t n)
4105 {
4106 	while (likely(n)) {
4107 		size_t leftlen = *cur_edge - *seg;
4108 		size_t copysz = min_t(size_t, leftlen, n);
4109 		size_t stride;
4110 
4111 		memcpy(*seg, src, copysz);
4112 
4113 		n -= copysz;
4114 		src += copysz;
4115 		stride = !n ? ALIGN(copysz, 16) : copysz;
4116 		*seg += stride;
4117 		*wqe_sz += stride >> 4;
4118 		handle_post_send_edge(sq, seg, *wqe_sz, cur_edge);
4119 	}
4120 }
4121 
4122 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
4123 {
4124 	struct mlx5_ib_cq *cq;
4125 	unsigned cur;
4126 
4127 	cur = wq->head - wq->tail;
4128 	if (likely(cur + nreq < wq->max_post))
4129 		return 0;
4130 
4131 	cq = to_mcq(ib_cq);
4132 	spin_lock(&cq->lock);
4133 	cur = wq->head - wq->tail;
4134 	spin_unlock(&cq->lock);
4135 
4136 	return cur + nreq >= wq->max_post;
4137 }
4138 
4139 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
4140 					  u64 remote_addr, u32 rkey)
4141 {
4142 	rseg->raddr    = cpu_to_be64(remote_addr);
4143 	rseg->rkey     = cpu_to_be32(rkey);
4144 	rseg->reserved = 0;
4145 }
4146 
4147 static void set_eth_seg(const struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
4148 			void **seg, int *size, void **cur_edge)
4149 {
4150 	struct mlx5_wqe_eth_seg *eseg = *seg;
4151 
4152 	memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
4153 
4154 	if (wr->send_flags & IB_SEND_IP_CSUM)
4155 		eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
4156 				 MLX5_ETH_WQE_L4_CSUM;
4157 
4158 	if (wr->opcode == IB_WR_LSO) {
4159 		struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
4160 		size_t left, copysz;
4161 		void *pdata = ud_wr->header;
4162 		size_t stride;
4163 
4164 		left = ud_wr->hlen;
4165 		eseg->mss = cpu_to_be16(ud_wr->mss);
4166 		eseg->inline_hdr.sz = cpu_to_be16(left);
4167 
4168 		/* memcpy_send_wqe should get a 16B align address. Hence, we
4169 		 * first copy up to the current edge and then, if needed,
4170 		 * fall-through to memcpy_send_wqe.
4171 		 */
4172 		copysz = min_t(u64, *cur_edge - (void *)eseg->inline_hdr.start,
4173 			       left);
4174 		memcpy(eseg->inline_hdr.start, pdata, copysz);
4175 		stride = ALIGN(sizeof(struct mlx5_wqe_eth_seg) -
4176 			       sizeof(eseg->inline_hdr.start) + copysz, 16);
4177 		*size += stride / 16;
4178 		*seg += stride;
4179 
4180 		if (copysz < left) {
4181 			handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4182 			left -= copysz;
4183 			pdata += copysz;
4184 			memcpy_send_wqe(&qp->sq, cur_edge, seg, size, pdata,
4185 					left);
4186 		}
4187 
4188 		return;
4189 	}
4190 
4191 	*seg += sizeof(struct mlx5_wqe_eth_seg);
4192 	*size += sizeof(struct mlx5_wqe_eth_seg) / 16;
4193 }
4194 
4195 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
4196 			     const struct ib_send_wr *wr)
4197 {
4198 	memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
4199 	dseg->av.dqp_dct = cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
4200 	dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
4201 }
4202 
4203 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
4204 {
4205 	dseg->byte_count = cpu_to_be32(sg->length);
4206 	dseg->lkey       = cpu_to_be32(sg->lkey);
4207 	dseg->addr       = cpu_to_be64(sg->addr);
4208 }
4209 
4210 static u64 get_xlt_octo(u64 bytes)
4211 {
4212 	return ALIGN(bytes, MLX5_IB_UMR_XLT_ALIGNMENT) /
4213 	       MLX5_IB_UMR_OCTOWORD;
4214 }
4215 
4216 static __be64 frwr_mkey_mask(bool atomic)
4217 {
4218 	u64 result;
4219 
4220 	result = MLX5_MKEY_MASK_LEN		|
4221 		MLX5_MKEY_MASK_PAGE_SIZE	|
4222 		MLX5_MKEY_MASK_START_ADDR	|
4223 		MLX5_MKEY_MASK_EN_RINVAL	|
4224 		MLX5_MKEY_MASK_KEY		|
4225 		MLX5_MKEY_MASK_LR		|
4226 		MLX5_MKEY_MASK_LW		|
4227 		MLX5_MKEY_MASK_RR		|
4228 		MLX5_MKEY_MASK_RW		|
4229 		MLX5_MKEY_MASK_SMALL_FENCE	|
4230 		MLX5_MKEY_MASK_FREE;
4231 
4232 	if (atomic)
4233 		result |= MLX5_MKEY_MASK_A;
4234 
4235 	return cpu_to_be64(result);
4236 }
4237 
4238 static __be64 sig_mkey_mask(void)
4239 {
4240 	u64 result;
4241 
4242 	result = MLX5_MKEY_MASK_LEN		|
4243 		MLX5_MKEY_MASK_PAGE_SIZE	|
4244 		MLX5_MKEY_MASK_START_ADDR	|
4245 		MLX5_MKEY_MASK_EN_SIGERR	|
4246 		MLX5_MKEY_MASK_EN_RINVAL	|
4247 		MLX5_MKEY_MASK_KEY		|
4248 		MLX5_MKEY_MASK_LR		|
4249 		MLX5_MKEY_MASK_LW		|
4250 		MLX5_MKEY_MASK_RR		|
4251 		MLX5_MKEY_MASK_RW		|
4252 		MLX5_MKEY_MASK_SMALL_FENCE	|
4253 		MLX5_MKEY_MASK_FREE		|
4254 		MLX5_MKEY_MASK_BSF_EN;
4255 
4256 	return cpu_to_be64(result);
4257 }
4258 
4259 static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
4260 			    struct mlx5_ib_mr *mr, u8 flags, bool atomic)
4261 {
4262 	int size = (mr->ndescs + mr->meta_ndescs) * mr->desc_size;
4263 
4264 	memset(umr, 0, sizeof(*umr));
4265 
4266 	umr->flags = flags;
4267 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(size));
4268 	umr->mkey_mask = frwr_mkey_mask(atomic);
4269 }
4270 
4271 static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
4272 {
4273 	memset(umr, 0, sizeof(*umr));
4274 	umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
4275 	umr->flags = MLX5_UMR_INLINE;
4276 }
4277 
4278 static __be64 get_umr_enable_mr_mask(void)
4279 {
4280 	u64 result;
4281 
4282 	result = MLX5_MKEY_MASK_KEY |
4283 		 MLX5_MKEY_MASK_FREE;
4284 
4285 	return cpu_to_be64(result);
4286 }
4287 
4288 static __be64 get_umr_disable_mr_mask(void)
4289 {
4290 	u64 result;
4291 
4292 	result = MLX5_MKEY_MASK_FREE;
4293 
4294 	return cpu_to_be64(result);
4295 }
4296 
4297 static __be64 get_umr_update_translation_mask(void)
4298 {
4299 	u64 result;
4300 
4301 	result = MLX5_MKEY_MASK_LEN |
4302 		 MLX5_MKEY_MASK_PAGE_SIZE |
4303 		 MLX5_MKEY_MASK_START_ADDR;
4304 
4305 	return cpu_to_be64(result);
4306 }
4307 
4308 static __be64 get_umr_update_access_mask(int atomic)
4309 {
4310 	u64 result;
4311 
4312 	result = MLX5_MKEY_MASK_LR |
4313 		 MLX5_MKEY_MASK_LW |
4314 		 MLX5_MKEY_MASK_RR |
4315 		 MLX5_MKEY_MASK_RW;
4316 
4317 	if (atomic)
4318 		result |= MLX5_MKEY_MASK_A;
4319 
4320 	return cpu_to_be64(result);
4321 }
4322 
4323 static __be64 get_umr_update_pd_mask(void)
4324 {
4325 	u64 result;
4326 
4327 	result = MLX5_MKEY_MASK_PD;
4328 
4329 	return cpu_to_be64(result);
4330 }
4331 
4332 static int umr_check_mkey_mask(struct mlx5_ib_dev *dev, u64 mask)
4333 {
4334 	if ((mask & MLX5_MKEY_MASK_PAGE_SIZE &&
4335 	     MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled)) ||
4336 	    (mask & MLX5_MKEY_MASK_A &&
4337 	     MLX5_CAP_GEN(dev->mdev, umr_modify_atomic_disabled)))
4338 		return -EPERM;
4339 	return 0;
4340 }
4341 
4342 static int set_reg_umr_segment(struct mlx5_ib_dev *dev,
4343 			       struct mlx5_wqe_umr_ctrl_seg *umr,
4344 			       const struct ib_send_wr *wr, int atomic)
4345 {
4346 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
4347 
4348 	memset(umr, 0, sizeof(*umr));
4349 
4350 	if (!umrwr->ignore_free_state) {
4351 		if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
4352 			 /* fail if free */
4353 			umr->flags = MLX5_UMR_CHECK_FREE;
4354 		else
4355 			/* fail if not free */
4356 			umr->flags = MLX5_UMR_CHECK_NOT_FREE;
4357 	}
4358 
4359 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(umrwr->xlt_size));
4360 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_XLT) {
4361 		u64 offset = get_xlt_octo(umrwr->offset);
4362 
4363 		umr->xlt_offset = cpu_to_be16(offset & 0xffff);
4364 		umr->xlt_offset_47_16 = cpu_to_be32(offset >> 16);
4365 		umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
4366 	}
4367 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
4368 		umr->mkey_mask |= get_umr_update_translation_mask();
4369 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS) {
4370 		umr->mkey_mask |= get_umr_update_access_mask(atomic);
4371 		umr->mkey_mask |= get_umr_update_pd_mask();
4372 	}
4373 	if (wr->send_flags & MLX5_IB_SEND_UMR_ENABLE_MR)
4374 		umr->mkey_mask |= get_umr_enable_mr_mask();
4375 	if (wr->send_flags & MLX5_IB_SEND_UMR_DISABLE_MR)
4376 		umr->mkey_mask |= get_umr_disable_mr_mask();
4377 
4378 	if (!wr->num_sge)
4379 		umr->flags |= MLX5_UMR_INLINE;
4380 
4381 	return umr_check_mkey_mask(dev, be64_to_cpu(umr->mkey_mask));
4382 }
4383 
4384 static u8 get_umr_flags(int acc)
4385 {
4386 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC       : 0) |
4387 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX5_PERM_REMOTE_WRITE : 0) |
4388 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX5_PERM_REMOTE_READ  : 0) |
4389 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX5_PERM_LOCAL_WRITE  : 0) |
4390 		MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
4391 }
4392 
4393 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
4394 			     struct mlx5_ib_mr *mr,
4395 			     u32 key, int access)
4396 {
4397 	int ndescs = ALIGN(mr->ndescs + mr->meta_ndescs, 8) >> 1;
4398 
4399 	memset(seg, 0, sizeof(*seg));
4400 
4401 	if (mr->access_mode == MLX5_MKC_ACCESS_MODE_MTT)
4402 		seg->log2_page_size = ilog2(mr->ibmr.page_size);
4403 	else if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
4404 		/* KLMs take twice the size of MTTs */
4405 		ndescs *= 2;
4406 
4407 	seg->flags = get_umr_flags(access) | mr->access_mode;
4408 	seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
4409 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
4410 	seg->start_addr = cpu_to_be64(mr->ibmr.iova);
4411 	seg->len = cpu_to_be64(mr->ibmr.length);
4412 	seg->xlt_oct_size = cpu_to_be32(ndescs);
4413 }
4414 
4415 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
4416 {
4417 	memset(seg, 0, sizeof(*seg));
4418 	seg->status = MLX5_MKEY_STATUS_FREE;
4419 }
4420 
4421 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg,
4422 				 const struct ib_send_wr *wr)
4423 {
4424 	const struct mlx5_umr_wr *umrwr = umr_wr(wr);
4425 
4426 	memset(seg, 0, sizeof(*seg));
4427 	if (wr->send_flags & MLX5_IB_SEND_UMR_DISABLE_MR)
4428 		seg->status = MLX5_MKEY_STATUS_FREE;
4429 
4430 	seg->flags = convert_access(umrwr->access_flags);
4431 	if (umrwr->pd)
4432 		seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
4433 	if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION &&
4434 	    !umrwr->length)
4435 		seg->flags_pd |= cpu_to_be32(MLX5_MKEY_LEN64);
4436 
4437 	seg->start_addr = cpu_to_be64(umrwr->virt_addr);
4438 	seg->len = cpu_to_be64(umrwr->length);
4439 	seg->log2_page_size = umrwr->page_shift;
4440 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
4441 				       mlx5_mkey_variant(umrwr->mkey));
4442 }
4443 
4444 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
4445 			     struct mlx5_ib_mr *mr,
4446 			     struct mlx5_ib_pd *pd)
4447 {
4448 	int bcount = mr->desc_size * (mr->ndescs + mr->meta_ndescs);
4449 
4450 	dseg->addr = cpu_to_be64(mr->desc_map);
4451 	dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
4452 	dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
4453 }
4454 
4455 static __be32 send_ieth(const struct ib_send_wr *wr)
4456 {
4457 	switch (wr->opcode) {
4458 	case IB_WR_SEND_WITH_IMM:
4459 	case IB_WR_RDMA_WRITE_WITH_IMM:
4460 		return wr->ex.imm_data;
4461 
4462 	case IB_WR_SEND_WITH_INV:
4463 		return cpu_to_be32(wr->ex.invalidate_rkey);
4464 
4465 	default:
4466 		return 0;
4467 	}
4468 }
4469 
4470 static u8 calc_sig(void *wqe, int size)
4471 {
4472 	u8 *p = wqe;
4473 	u8 res = 0;
4474 	int i;
4475 
4476 	for (i = 0; i < size; i++)
4477 		res ^= p[i];
4478 
4479 	return ~res;
4480 }
4481 
4482 static u8 wq_sig(void *wqe)
4483 {
4484 	return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
4485 }
4486 
4487 static int set_data_inl_seg(struct mlx5_ib_qp *qp, const struct ib_send_wr *wr,
4488 			    void **wqe, int *wqe_sz, void **cur_edge)
4489 {
4490 	struct mlx5_wqe_inline_seg *seg;
4491 	size_t offset;
4492 	int inl = 0;
4493 	int i;
4494 
4495 	seg = *wqe;
4496 	*wqe += sizeof(*seg);
4497 	offset = sizeof(*seg);
4498 
4499 	for (i = 0; i < wr->num_sge; i++) {
4500 		size_t len  = wr->sg_list[i].length;
4501 		void *addr = (void *)(unsigned long)(wr->sg_list[i].addr);
4502 
4503 		inl += len;
4504 
4505 		if (unlikely(inl > qp->max_inline_data))
4506 			return -ENOMEM;
4507 
4508 		while (likely(len)) {
4509 			size_t leftlen;
4510 			size_t copysz;
4511 
4512 			handle_post_send_edge(&qp->sq, wqe,
4513 					      *wqe_sz + (offset >> 4),
4514 					      cur_edge);
4515 
4516 			leftlen = *cur_edge - *wqe;
4517 			copysz = min_t(size_t, leftlen, len);
4518 
4519 			memcpy(*wqe, addr, copysz);
4520 			len -= copysz;
4521 			addr += copysz;
4522 			*wqe += copysz;
4523 			offset += copysz;
4524 		}
4525 	}
4526 
4527 	seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
4528 
4529 	*wqe_sz +=  ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
4530 
4531 	return 0;
4532 }
4533 
4534 static u16 prot_field_size(enum ib_signature_type type)
4535 {
4536 	switch (type) {
4537 	case IB_SIG_TYPE_T10_DIF:
4538 		return MLX5_DIF_SIZE;
4539 	default:
4540 		return 0;
4541 	}
4542 }
4543 
4544 static u8 bs_selector(int block_size)
4545 {
4546 	switch (block_size) {
4547 	case 512:	    return 0x1;
4548 	case 520:	    return 0x2;
4549 	case 4096:	    return 0x3;
4550 	case 4160:	    return 0x4;
4551 	case 1073741824:    return 0x5;
4552 	default:	    return 0;
4553 	}
4554 }
4555 
4556 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
4557 			      struct mlx5_bsf_inl *inl)
4558 {
4559 	/* Valid inline section and allow BSF refresh */
4560 	inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
4561 				       MLX5_BSF_REFRESH_DIF);
4562 	inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
4563 	inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
4564 	/* repeating block */
4565 	inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
4566 	inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
4567 			MLX5_DIF_CRC : MLX5_DIF_IPCS;
4568 
4569 	if (domain->sig.dif.ref_remap)
4570 		inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
4571 
4572 	if (domain->sig.dif.app_escape) {
4573 		if (domain->sig.dif.ref_escape)
4574 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
4575 		else
4576 			inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
4577 	}
4578 
4579 	inl->dif_app_bitmask_check =
4580 		cpu_to_be16(domain->sig.dif.apptag_check_mask);
4581 }
4582 
4583 static int mlx5_set_bsf(struct ib_mr *sig_mr,
4584 			struct ib_sig_attrs *sig_attrs,
4585 			struct mlx5_bsf *bsf, u32 data_size)
4586 {
4587 	struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
4588 	struct mlx5_bsf_basic *basic = &bsf->basic;
4589 	struct ib_sig_domain *mem = &sig_attrs->mem;
4590 	struct ib_sig_domain *wire = &sig_attrs->wire;
4591 
4592 	memset(bsf, 0, sizeof(*bsf));
4593 
4594 	/* Basic + Extended + Inline */
4595 	basic->bsf_size_sbs = 1 << 7;
4596 	/* Input domain check byte mask */
4597 	basic->check_byte_mask = sig_attrs->check_mask;
4598 	basic->raw_data_size = cpu_to_be32(data_size);
4599 
4600 	/* Memory domain */
4601 	switch (sig_attrs->mem.sig_type) {
4602 	case IB_SIG_TYPE_NONE:
4603 		break;
4604 	case IB_SIG_TYPE_T10_DIF:
4605 		basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
4606 		basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
4607 		mlx5_fill_inl_bsf(mem, &bsf->m_inl);
4608 		break;
4609 	default:
4610 		return -EINVAL;
4611 	}
4612 
4613 	/* Wire domain */
4614 	switch (sig_attrs->wire.sig_type) {
4615 	case IB_SIG_TYPE_NONE:
4616 		break;
4617 	case IB_SIG_TYPE_T10_DIF:
4618 		if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
4619 		    mem->sig_type == wire->sig_type) {
4620 			/* Same block structure */
4621 			basic->bsf_size_sbs |= 1 << 4;
4622 			if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
4623 				basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
4624 			if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
4625 				basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
4626 			if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
4627 				basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
4628 		} else
4629 			basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
4630 
4631 		basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
4632 		mlx5_fill_inl_bsf(wire, &bsf->w_inl);
4633 		break;
4634 	default:
4635 		return -EINVAL;
4636 	}
4637 
4638 	return 0;
4639 }
4640 
4641 static int set_sig_data_segment(const struct ib_send_wr *send_wr,
4642 				struct ib_mr *sig_mr,
4643 				struct ib_sig_attrs *sig_attrs,
4644 				struct mlx5_ib_qp *qp, void **seg, int *size,
4645 				void **cur_edge)
4646 {
4647 	struct mlx5_bsf *bsf;
4648 	u32 data_len;
4649 	u32 data_key;
4650 	u64 data_va;
4651 	u32 prot_len = 0;
4652 	u32 prot_key = 0;
4653 	u64 prot_va = 0;
4654 	bool prot = false;
4655 	int ret;
4656 	int wqe_size;
4657 	struct mlx5_ib_mr *mr = to_mmr(sig_mr);
4658 	struct mlx5_ib_mr *pi_mr = mr->pi_mr;
4659 
4660 	data_len = pi_mr->data_length;
4661 	data_key = pi_mr->ibmr.lkey;
4662 	data_va = pi_mr->data_iova;
4663 	if (pi_mr->meta_ndescs) {
4664 		prot_len = pi_mr->meta_length;
4665 		prot_key = pi_mr->ibmr.lkey;
4666 		prot_va = pi_mr->pi_iova;
4667 		prot = true;
4668 	}
4669 
4670 	if (!prot || (data_key == prot_key && data_va == prot_va &&
4671 		      data_len == prot_len)) {
4672 		/**
4673 		 * Source domain doesn't contain signature information
4674 		 * or data and protection are interleaved in memory.
4675 		 * So need construct:
4676 		 *                  ------------------
4677 		 *                 |     data_klm     |
4678 		 *                  ------------------
4679 		 *                 |       BSF        |
4680 		 *                  ------------------
4681 		 **/
4682 		struct mlx5_klm *data_klm = *seg;
4683 
4684 		data_klm->bcount = cpu_to_be32(data_len);
4685 		data_klm->key = cpu_to_be32(data_key);
4686 		data_klm->va = cpu_to_be64(data_va);
4687 		wqe_size = ALIGN(sizeof(*data_klm), 64);
4688 	} else {
4689 		/**
4690 		 * Source domain contains signature information
4691 		 * So need construct a strided block format:
4692 		 *               ---------------------------
4693 		 *              |     stride_block_ctrl     |
4694 		 *               ---------------------------
4695 		 *              |          data_klm         |
4696 		 *               ---------------------------
4697 		 *              |          prot_klm         |
4698 		 *               ---------------------------
4699 		 *              |             BSF           |
4700 		 *               ---------------------------
4701 		 **/
4702 		struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
4703 		struct mlx5_stride_block_entry *data_sentry;
4704 		struct mlx5_stride_block_entry *prot_sentry;
4705 		u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
4706 		int prot_size;
4707 
4708 		sblock_ctrl = *seg;
4709 		data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
4710 		prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
4711 
4712 		prot_size = prot_field_size(sig_attrs->mem.sig_type);
4713 		if (!prot_size) {
4714 			pr_err("Bad block size given: %u\n", block_size);
4715 			return -EINVAL;
4716 		}
4717 		sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
4718 							    prot_size);
4719 		sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
4720 		sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
4721 		sblock_ctrl->num_entries = cpu_to_be16(2);
4722 
4723 		data_sentry->bcount = cpu_to_be16(block_size);
4724 		data_sentry->key = cpu_to_be32(data_key);
4725 		data_sentry->va = cpu_to_be64(data_va);
4726 		data_sentry->stride = cpu_to_be16(block_size);
4727 
4728 		prot_sentry->bcount = cpu_to_be16(prot_size);
4729 		prot_sentry->key = cpu_to_be32(prot_key);
4730 		prot_sentry->va = cpu_to_be64(prot_va);
4731 		prot_sentry->stride = cpu_to_be16(prot_size);
4732 
4733 		wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
4734 				 sizeof(*prot_sentry), 64);
4735 	}
4736 
4737 	*seg += wqe_size;
4738 	*size += wqe_size / 16;
4739 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4740 
4741 	bsf = *seg;
4742 	ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
4743 	if (ret)
4744 		return -EINVAL;
4745 
4746 	*seg += sizeof(*bsf);
4747 	*size += sizeof(*bsf) / 16;
4748 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4749 
4750 	return 0;
4751 }
4752 
4753 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
4754 				 struct ib_mr *sig_mr, int access_flags,
4755 				 u32 size, u32 length, u32 pdn)
4756 {
4757 	u32 sig_key = sig_mr->rkey;
4758 	u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
4759 
4760 	memset(seg, 0, sizeof(*seg));
4761 
4762 	seg->flags = get_umr_flags(access_flags) | MLX5_MKC_ACCESS_MODE_KLMS;
4763 	seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
4764 	seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
4765 				    MLX5_MKEY_BSF_EN | pdn);
4766 	seg->len = cpu_to_be64(length);
4767 	seg->xlt_oct_size = cpu_to_be32(get_xlt_octo(size));
4768 	seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
4769 }
4770 
4771 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
4772 				u32 size)
4773 {
4774 	memset(umr, 0, sizeof(*umr));
4775 
4776 	umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
4777 	umr->xlt_octowords = cpu_to_be16(get_xlt_octo(size));
4778 	umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
4779 	umr->mkey_mask = sig_mkey_mask();
4780 }
4781 
4782 static int set_pi_umr_wr(const struct ib_send_wr *send_wr,
4783 			 struct mlx5_ib_qp *qp, void **seg, int *size,
4784 			 void **cur_edge)
4785 {
4786 	const struct ib_reg_wr *wr = reg_wr(send_wr);
4787 	struct mlx5_ib_mr *sig_mr = to_mmr(wr->mr);
4788 	struct mlx5_ib_mr *pi_mr = sig_mr->pi_mr;
4789 	struct ib_sig_attrs *sig_attrs = sig_mr->ibmr.sig_attrs;
4790 	u32 pdn = get_pd(qp)->pdn;
4791 	u32 xlt_size;
4792 	int region_len, ret;
4793 
4794 	if (unlikely(send_wr->num_sge != 0) ||
4795 	    unlikely(wr->access & IB_ACCESS_REMOTE_ATOMIC) ||
4796 	    unlikely(!sig_mr->sig) || unlikely(!qp->ibqp.integrity_en) ||
4797 	    unlikely(!sig_mr->sig->sig_status_checked))
4798 		return -EINVAL;
4799 
4800 	/* length of the protected region, data + protection */
4801 	region_len = pi_mr->ibmr.length;
4802 
4803 	/**
4804 	 * KLM octoword size - if protection was provided
4805 	 * then we use strided block format (3 octowords),
4806 	 * else we use single KLM (1 octoword)
4807 	 **/
4808 	if (sig_attrs->mem.sig_type != IB_SIG_TYPE_NONE)
4809 		xlt_size = 0x30;
4810 	else
4811 		xlt_size = sizeof(struct mlx5_klm);
4812 
4813 	set_sig_umr_segment(*seg, xlt_size);
4814 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4815 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4816 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4817 
4818 	set_sig_mkey_segment(*seg, wr->mr, wr->access, xlt_size, region_len,
4819 			     pdn);
4820 	*seg += sizeof(struct mlx5_mkey_seg);
4821 	*size += sizeof(struct mlx5_mkey_seg) / 16;
4822 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4823 
4824 	ret = set_sig_data_segment(send_wr, wr->mr, sig_attrs, qp, seg, size,
4825 				   cur_edge);
4826 	if (ret)
4827 		return ret;
4828 
4829 	sig_mr->sig->sig_status_checked = false;
4830 	return 0;
4831 }
4832 
4833 static int set_psv_wr(struct ib_sig_domain *domain,
4834 		      u32 psv_idx, void **seg, int *size)
4835 {
4836 	struct mlx5_seg_set_psv *psv_seg = *seg;
4837 
4838 	memset(psv_seg, 0, sizeof(*psv_seg));
4839 	psv_seg->psv_num = cpu_to_be32(psv_idx);
4840 	switch (domain->sig_type) {
4841 	case IB_SIG_TYPE_NONE:
4842 		break;
4843 	case IB_SIG_TYPE_T10_DIF:
4844 		psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
4845 						     domain->sig.dif.app_tag);
4846 		psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
4847 		break;
4848 	default:
4849 		pr_err("Bad signature type (%d) is given.\n",
4850 		       domain->sig_type);
4851 		return -EINVAL;
4852 	}
4853 
4854 	*seg += sizeof(*psv_seg);
4855 	*size += sizeof(*psv_seg) / 16;
4856 
4857 	return 0;
4858 }
4859 
4860 static int set_reg_wr(struct mlx5_ib_qp *qp,
4861 		      const struct ib_reg_wr *wr,
4862 		      void **seg, int *size, void **cur_edge,
4863 		      bool check_not_free)
4864 {
4865 	struct mlx5_ib_mr *mr = to_mmr(wr->mr);
4866 	struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
4867 	struct mlx5_ib_dev *dev = to_mdev(pd->ibpd.device);
4868 	int mr_list_size = (mr->ndescs + mr->meta_ndescs) * mr->desc_size;
4869 	bool umr_inline = mr_list_size <= MLX5_IB_SQ_UMR_INLINE_THRESHOLD;
4870 	bool atomic = wr->access & IB_ACCESS_REMOTE_ATOMIC;
4871 	u8 flags = 0;
4872 
4873 	if (!mlx5_ib_can_use_umr(dev, atomic, wr->access)) {
4874 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
4875 			     "Fast update of %s for MR is disabled\n",
4876 			     (MLX5_CAP_GEN(dev->mdev,
4877 					   umr_modify_entity_size_disabled)) ?
4878 				     "entity size" :
4879 				     "atomic access");
4880 		return -EINVAL;
4881 	}
4882 
4883 	if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
4884 		mlx5_ib_warn(to_mdev(qp->ibqp.device),
4885 			     "Invalid IB_SEND_INLINE send flag\n");
4886 		return -EINVAL;
4887 	}
4888 
4889 	if (check_not_free)
4890 		flags |= MLX5_UMR_CHECK_NOT_FREE;
4891 	if (umr_inline)
4892 		flags |= MLX5_UMR_INLINE;
4893 
4894 	set_reg_umr_seg(*seg, mr, flags, atomic);
4895 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4896 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4897 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4898 
4899 	set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
4900 	*seg += sizeof(struct mlx5_mkey_seg);
4901 	*size += sizeof(struct mlx5_mkey_seg) / 16;
4902 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4903 
4904 	if (umr_inline) {
4905 		memcpy_send_wqe(&qp->sq, cur_edge, seg, size, mr->descs,
4906 				mr_list_size);
4907 		*size = ALIGN(*size, MLX5_SEND_WQE_BB >> 4);
4908 	} else {
4909 		set_reg_data_seg(*seg, mr, pd);
4910 		*seg += sizeof(struct mlx5_wqe_data_seg);
4911 		*size += (sizeof(struct mlx5_wqe_data_seg) / 16);
4912 	}
4913 	return 0;
4914 }
4915 
4916 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size,
4917 			void **cur_edge)
4918 {
4919 	set_linv_umr_seg(*seg);
4920 	*seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4921 	*size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4922 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4923 	set_linv_mkey_seg(*seg);
4924 	*seg += sizeof(struct mlx5_mkey_seg);
4925 	*size += sizeof(struct mlx5_mkey_seg) / 16;
4926 	handle_post_send_edge(&qp->sq, seg, *size, cur_edge);
4927 }
4928 
4929 static void dump_wqe(struct mlx5_ib_qp *qp, u32 idx, int size_16)
4930 {
4931 	__be32 *p = NULL;
4932 	int i, j;
4933 
4934 	pr_debug("dump WQE index %u:\n", idx);
4935 	for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
4936 		if ((i & 0xf) == 0) {
4937 			p = mlx5_frag_buf_get_wqe(&qp->sq.fbc, idx);
4938 			pr_debug("WQBB at %p:\n", (void *)p);
4939 			j = 0;
4940 			idx = (idx + 1) & (qp->sq.wqe_cnt - 1);
4941 		}
4942 		pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
4943 			 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
4944 			 be32_to_cpu(p[j + 3]));
4945 	}
4946 }
4947 
4948 static int __begin_wqe(struct mlx5_ib_qp *qp, void **seg,
4949 		       struct mlx5_wqe_ctrl_seg **ctrl,
4950 		       const struct ib_send_wr *wr, unsigned int *idx,
4951 		       int *size, void **cur_edge, int nreq,
4952 		       bool send_signaled, bool solicited)
4953 {
4954 	if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)))
4955 		return -ENOMEM;
4956 
4957 	*idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
4958 	*seg = mlx5_frag_buf_get_wqe(&qp->sq.fbc, *idx);
4959 	*ctrl = *seg;
4960 	*(uint32_t *)(*seg + 8) = 0;
4961 	(*ctrl)->imm = send_ieth(wr);
4962 	(*ctrl)->fm_ce_se = qp->sq_signal_bits |
4963 		(send_signaled ? MLX5_WQE_CTRL_CQ_UPDATE : 0) |
4964 		(solicited ? MLX5_WQE_CTRL_SOLICITED : 0);
4965 
4966 	*seg += sizeof(**ctrl);
4967 	*size = sizeof(**ctrl) / 16;
4968 	*cur_edge = qp->sq.cur_edge;
4969 
4970 	return 0;
4971 }
4972 
4973 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
4974 		     struct mlx5_wqe_ctrl_seg **ctrl,
4975 		     const struct ib_send_wr *wr, unsigned *idx,
4976 		     int *size, void **cur_edge, int nreq)
4977 {
4978 	return __begin_wqe(qp, seg, ctrl, wr, idx, size, cur_edge, nreq,
4979 			   wr->send_flags & IB_SEND_SIGNALED,
4980 			   wr->send_flags & IB_SEND_SOLICITED);
4981 }
4982 
4983 static void finish_wqe(struct mlx5_ib_qp *qp,
4984 		       struct mlx5_wqe_ctrl_seg *ctrl,
4985 		       void *seg, u8 size, void *cur_edge,
4986 		       unsigned int idx, u64 wr_id, int nreq, u8 fence,
4987 		       u32 mlx5_opcode)
4988 {
4989 	u8 opmod = 0;
4990 
4991 	ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
4992 					     mlx5_opcode | ((u32)opmod << 24));
4993 	ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
4994 	ctrl->fm_ce_se |= fence;
4995 	if (unlikely(qp->wq_sig))
4996 		ctrl->signature = wq_sig(ctrl);
4997 
4998 	qp->sq.wrid[idx] = wr_id;
4999 	qp->sq.w_list[idx].opcode = mlx5_opcode;
5000 	qp->sq.wqe_head[idx] = qp->sq.head + nreq;
5001 	qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
5002 	qp->sq.w_list[idx].next = qp->sq.cur_post;
5003 
5004 	/* We save the edge which was possibly updated during the WQE
5005 	 * construction, into SQ's cache.
5006 	 */
5007 	seg = PTR_ALIGN(seg, MLX5_SEND_WQE_BB);
5008 	qp->sq.cur_edge = (unlikely(seg == cur_edge)) ?
5009 			  get_sq_edge(&qp->sq, qp->sq.cur_post &
5010 				      (qp->sq.wqe_cnt - 1)) :
5011 			  cur_edge;
5012 }
5013 
5014 static int _mlx5_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
5015 			      const struct ib_send_wr **bad_wr, bool drain)
5016 {
5017 	struct mlx5_wqe_ctrl_seg *ctrl = NULL;  /* compiler warning */
5018 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
5019 	struct mlx5_core_dev *mdev = dev->mdev;
5020 	struct ib_reg_wr reg_pi_wr;
5021 	struct mlx5_ib_qp *qp;
5022 	struct mlx5_ib_mr *mr;
5023 	struct mlx5_ib_mr *pi_mr;
5024 	struct mlx5_ib_mr pa_pi_mr;
5025 	struct ib_sig_attrs *sig_attrs;
5026 	struct mlx5_wqe_xrc_seg *xrc;
5027 	struct mlx5_bf *bf;
5028 	void *cur_edge;
5029 	int uninitialized_var(size);
5030 	unsigned long flags;
5031 	unsigned idx;
5032 	int err = 0;
5033 	int num_sge;
5034 	void *seg;
5035 	int nreq;
5036 	int i;
5037 	u8 next_fence = 0;
5038 	u8 fence;
5039 
5040 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR &&
5041 		     !drain)) {
5042 		*bad_wr = wr;
5043 		return -EIO;
5044 	}
5045 
5046 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
5047 		return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
5048 
5049 	qp = to_mqp(ibqp);
5050 	bf = &qp->bf;
5051 
5052 	spin_lock_irqsave(&qp->sq.lock, flags);
5053 
5054 	for (nreq = 0; wr; nreq++, wr = wr->next) {
5055 		if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
5056 			mlx5_ib_warn(dev, "\n");
5057 			err = -EINVAL;
5058 			*bad_wr = wr;
5059 			goto out;
5060 		}
5061 
5062 		num_sge = wr->num_sge;
5063 		if (unlikely(num_sge > qp->sq.max_gs)) {
5064 			mlx5_ib_warn(dev, "\n");
5065 			err = -EINVAL;
5066 			*bad_wr = wr;
5067 			goto out;
5068 		}
5069 
5070 		err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, &cur_edge,
5071 				nreq);
5072 		if (err) {
5073 			mlx5_ib_warn(dev, "\n");
5074 			err = -ENOMEM;
5075 			*bad_wr = wr;
5076 			goto out;
5077 		}
5078 
5079 		if (wr->opcode == IB_WR_REG_MR ||
5080 		    wr->opcode == IB_WR_REG_MR_INTEGRITY) {
5081 			fence = dev->umr_fence;
5082 			next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
5083 		} else  {
5084 			if (wr->send_flags & IB_SEND_FENCE) {
5085 				if (qp->next_fence)
5086 					fence = MLX5_FENCE_MODE_SMALL_AND_FENCE;
5087 				else
5088 					fence = MLX5_FENCE_MODE_FENCE;
5089 			} else {
5090 				fence = qp->next_fence;
5091 			}
5092 		}
5093 
5094 		switch (ibqp->qp_type) {
5095 		case IB_QPT_XRC_INI:
5096 			xrc = seg;
5097 			seg += sizeof(*xrc);
5098 			size += sizeof(*xrc) / 16;
5099 			/* fall through */
5100 		case IB_QPT_RC:
5101 			switch (wr->opcode) {
5102 			case IB_WR_RDMA_READ:
5103 			case IB_WR_RDMA_WRITE:
5104 			case IB_WR_RDMA_WRITE_WITH_IMM:
5105 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
5106 					      rdma_wr(wr)->rkey);
5107 				seg += sizeof(struct mlx5_wqe_raddr_seg);
5108 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
5109 				break;
5110 
5111 			case IB_WR_ATOMIC_CMP_AND_SWP:
5112 			case IB_WR_ATOMIC_FETCH_AND_ADD:
5113 			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
5114 				mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
5115 				err = -ENOSYS;
5116 				*bad_wr = wr;
5117 				goto out;
5118 
5119 			case IB_WR_LOCAL_INV:
5120 				qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
5121 				ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
5122 				set_linv_wr(qp, &seg, &size, &cur_edge);
5123 				num_sge = 0;
5124 				break;
5125 
5126 			case IB_WR_REG_MR:
5127 				qp->sq.wr_data[idx] = IB_WR_REG_MR;
5128 				ctrl->imm = cpu_to_be32(reg_wr(wr)->key);
5129 				err = set_reg_wr(qp, reg_wr(wr), &seg, &size,
5130 						 &cur_edge, true);
5131 				if (err) {
5132 					*bad_wr = wr;
5133 					goto out;
5134 				}
5135 				num_sge = 0;
5136 				break;
5137 
5138 			case IB_WR_REG_MR_INTEGRITY:
5139 				qp->sq.wr_data[idx] = IB_WR_REG_MR_INTEGRITY;
5140 
5141 				mr = to_mmr(reg_wr(wr)->mr);
5142 				pi_mr = mr->pi_mr;
5143 
5144 				if (pi_mr) {
5145 					memset(&reg_pi_wr, 0,
5146 					       sizeof(struct ib_reg_wr));
5147 
5148 					reg_pi_wr.mr = &pi_mr->ibmr;
5149 					reg_pi_wr.access = reg_wr(wr)->access;
5150 					reg_pi_wr.key = pi_mr->ibmr.rkey;
5151 
5152 					ctrl->imm = cpu_to_be32(reg_pi_wr.key);
5153 					/* UMR for data + prot registration */
5154 					err = set_reg_wr(qp, &reg_pi_wr, &seg,
5155 							 &size, &cur_edge,
5156 							 false);
5157 					if (err) {
5158 						*bad_wr = wr;
5159 						goto out;
5160 					}
5161 					finish_wqe(qp, ctrl, seg, size,
5162 						   cur_edge, idx, wr->wr_id,
5163 						   nreq, fence,
5164 						   MLX5_OPCODE_UMR);
5165 
5166 					err = begin_wqe(qp, &seg, &ctrl, wr,
5167 							&idx, &size, &cur_edge,
5168 							nreq);
5169 					if (err) {
5170 						mlx5_ib_warn(dev, "\n");
5171 						err = -ENOMEM;
5172 						*bad_wr = wr;
5173 						goto out;
5174 					}
5175 				} else {
5176 					memset(&pa_pi_mr, 0,
5177 					       sizeof(struct mlx5_ib_mr));
5178 					/* No UMR, use local_dma_lkey */
5179 					pa_pi_mr.ibmr.lkey =
5180 						mr->ibmr.pd->local_dma_lkey;
5181 
5182 					pa_pi_mr.ndescs = mr->ndescs;
5183 					pa_pi_mr.data_length = mr->data_length;
5184 					pa_pi_mr.data_iova = mr->data_iova;
5185 					if (mr->meta_ndescs) {
5186 						pa_pi_mr.meta_ndescs =
5187 							mr->meta_ndescs;
5188 						pa_pi_mr.meta_length =
5189 							mr->meta_length;
5190 						pa_pi_mr.pi_iova = mr->pi_iova;
5191 					}
5192 
5193 					pa_pi_mr.ibmr.length = mr->ibmr.length;
5194 					mr->pi_mr = &pa_pi_mr;
5195 				}
5196 				ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
5197 				/* UMR for sig MR */
5198 				err = set_pi_umr_wr(wr, qp, &seg, &size,
5199 						    &cur_edge);
5200 				if (err) {
5201 					mlx5_ib_warn(dev, "\n");
5202 					*bad_wr = wr;
5203 					goto out;
5204 				}
5205 				finish_wqe(qp, ctrl, seg, size, cur_edge, idx,
5206 					   wr->wr_id, nreq, fence,
5207 					   MLX5_OPCODE_UMR);
5208 
5209 				/*
5210 				 * SET_PSV WQEs are not signaled and solicited
5211 				 * on error
5212 				 */
5213 				sig_attrs = mr->ibmr.sig_attrs;
5214 				err = __begin_wqe(qp, &seg, &ctrl, wr, &idx,
5215 						  &size, &cur_edge, nreq, false,
5216 						  true);
5217 				if (err) {
5218 					mlx5_ib_warn(dev, "\n");
5219 					err = -ENOMEM;
5220 					*bad_wr = wr;
5221 					goto out;
5222 				}
5223 				err = set_psv_wr(&sig_attrs->mem,
5224 						 mr->sig->psv_memory.psv_idx,
5225 						 &seg, &size);
5226 				if (err) {
5227 					mlx5_ib_warn(dev, "\n");
5228 					*bad_wr = wr;
5229 					goto out;
5230 				}
5231 				finish_wqe(qp, ctrl, seg, size, cur_edge, idx,
5232 					   wr->wr_id, nreq, next_fence,
5233 					   MLX5_OPCODE_SET_PSV);
5234 
5235 				err = __begin_wqe(qp, &seg, &ctrl, wr, &idx,
5236 						  &size, &cur_edge, nreq, false,
5237 						  true);
5238 				if (err) {
5239 					mlx5_ib_warn(dev, "\n");
5240 					err = -ENOMEM;
5241 					*bad_wr = wr;
5242 					goto out;
5243 				}
5244 				err = set_psv_wr(&sig_attrs->wire,
5245 						 mr->sig->psv_wire.psv_idx,
5246 						 &seg, &size);
5247 				if (err) {
5248 					mlx5_ib_warn(dev, "\n");
5249 					*bad_wr = wr;
5250 					goto out;
5251 				}
5252 				finish_wqe(qp, ctrl, seg, size, cur_edge, idx,
5253 					   wr->wr_id, nreq, next_fence,
5254 					   MLX5_OPCODE_SET_PSV);
5255 
5256 				qp->next_fence =
5257 					MLX5_FENCE_MODE_INITIATOR_SMALL;
5258 				num_sge = 0;
5259 				goto skip_psv;
5260 
5261 			default:
5262 				break;
5263 			}
5264 			break;
5265 
5266 		case IB_QPT_UC:
5267 			switch (wr->opcode) {
5268 			case IB_WR_RDMA_WRITE:
5269 			case IB_WR_RDMA_WRITE_WITH_IMM:
5270 				set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
5271 					      rdma_wr(wr)->rkey);
5272 				seg  += sizeof(struct mlx5_wqe_raddr_seg);
5273 				size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
5274 				break;
5275 
5276 			default:
5277 				break;
5278 			}
5279 			break;
5280 
5281 		case IB_QPT_SMI:
5282 			if (unlikely(!mdev->port_caps[qp->port - 1].has_smi)) {
5283 				mlx5_ib_warn(dev, "Send SMP MADs is not allowed\n");
5284 				err = -EPERM;
5285 				*bad_wr = wr;
5286 				goto out;
5287 			}
5288 			/* fall through */
5289 		case MLX5_IB_QPT_HW_GSI:
5290 			set_datagram_seg(seg, wr);
5291 			seg += sizeof(struct mlx5_wqe_datagram_seg);
5292 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
5293 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5294 
5295 			break;
5296 		case IB_QPT_UD:
5297 			set_datagram_seg(seg, wr);
5298 			seg += sizeof(struct mlx5_wqe_datagram_seg);
5299 			size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
5300 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5301 
5302 			/* handle qp that supports ud offload */
5303 			if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
5304 				struct mlx5_wqe_eth_pad *pad;
5305 
5306 				pad = seg;
5307 				memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
5308 				seg += sizeof(struct mlx5_wqe_eth_pad);
5309 				size += sizeof(struct mlx5_wqe_eth_pad) / 16;
5310 				set_eth_seg(wr, qp, &seg, &size, &cur_edge);
5311 				handle_post_send_edge(&qp->sq, &seg, size,
5312 						      &cur_edge);
5313 			}
5314 			break;
5315 		case MLX5_IB_QPT_REG_UMR:
5316 			if (wr->opcode != MLX5_IB_WR_UMR) {
5317 				err = -EINVAL;
5318 				mlx5_ib_warn(dev, "bad opcode\n");
5319 				goto out;
5320 			}
5321 			qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
5322 			ctrl->imm = cpu_to_be32(umr_wr(wr)->mkey);
5323 			err = set_reg_umr_segment(dev, seg, wr, !!(MLX5_CAP_GEN(mdev, atomic)));
5324 			if (unlikely(err))
5325 				goto out;
5326 			seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
5327 			size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
5328 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5329 			set_reg_mkey_segment(seg, wr);
5330 			seg += sizeof(struct mlx5_mkey_seg);
5331 			size += sizeof(struct mlx5_mkey_seg) / 16;
5332 			handle_post_send_edge(&qp->sq, &seg, size, &cur_edge);
5333 			break;
5334 
5335 		default:
5336 			break;
5337 		}
5338 
5339 		if (wr->send_flags & IB_SEND_INLINE && num_sge) {
5340 			err = set_data_inl_seg(qp, wr, &seg, &size, &cur_edge);
5341 			if (unlikely(err)) {
5342 				mlx5_ib_warn(dev, "\n");
5343 				*bad_wr = wr;
5344 				goto out;
5345 			}
5346 		} else {
5347 			for (i = 0; i < num_sge; i++) {
5348 				handle_post_send_edge(&qp->sq, &seg, size,
5349 						      &cur_edge);
5350 				if (likely(wr->sg_list[i].length)) {
5351 					set_data_ptr_seg
5352 					((struct mlx5_wqe_data_seg *)seg,
5353 					 wr->sg_list + i);
5354 					size += sizeof(struct mlx5_wqe_data_seg) / 16;
5355 					seg += sizeof(struct mlx5_wqe_data_seg);
5356 				}
5357 			}
5358 		}
5359 
5360 		qp->next_fence = next_fence;
5361 		finish_wqe(qp, ctrl, seg, size, cur_edge, idx, wr->wr_id, nreq,
5362 			   fence, mlx5_ib_opcode[wr->opcode]);
5363 skip_psv:
5364 		if (0)
5365 			dump_wqe(qp, idx, size);
5366 	}
5367 
5368 out:
5369 	if (likely(nreq)) {
5370 		qp->sq.head += nreq;
5371 
5372 		/* Make sure that descriptors are written before
5373 		 * updating doorbell record and ringing the doorbell
5374 		 */
5375 		wmb();
5376 
5377 		qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
5378 
5379 		/* Make sure doorbell record is visible to the HCA before
5380 		 * we hit doorbell */
5381 		wmb();
5382 
5383 		mlx5_write64((__be32 *)ctrl, bf->bfreg->map + bf->offset);
5384 		/* Make sure doorbells don't leak out of SQ spinlock
5385 		 * and reach the HCA out of order.
5386 		 */
5387 		bf->offset ^= bf->buf_size;
5388 	}
5389 
5390 	spin_unlock_irqrestore(&qp->sq.lock, flags);
5391 
5392 	return err;
5393 }
5394 
5395 int mlx5_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
5396 		      const struct ib_send_wr **bad_wr)
5397 {
5398 	return _mlx5_ib_post_send(ibqp, wr, bad_wr, false);
5399 }
5400 
5401 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
5402 {
5403 	sig->signature = calc_sig(sig, size);
5404 }
5405 
5406 static int _mlx5_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
5407 		      const struct ib_recv_wr **bad_wr, bool drain)
5408 {
5409 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
5410 	struct mlx5_wqe_data_seg *scat;
5411 	struct mlx5_rwqe_sig *sig;
5412 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
5413 	struct mlx5_core_dev *mdev = dev->mdev;
5414 	unsigned long flags;
5415 	int err = 0;
5416 	int nreq;
5417 	int ind;
5418 	int i;
5419 
5420 	if (unlikely(mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR &&
5421 		     !drain)) {
5422 		*bad_wr = wr;
5423 		return -EIO;
5424 	}
5425 
5426 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
5427 		return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
5428 
5429 	spin_lock_irqsave(&qp->rq.lock, flags);
5430 
5431 	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
5432 
5433 	for (nreq = 0; wr; nreq++, wr = wr->next) {
5434 		if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
5435 			err = -ENOMEM;
5436 			*bad_wr = wr;
5437 			goto out;
5438 		}
5439 
5440 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
5441 			err = -EINVAL;
5442 			*bad_wr = wr;
5443 			goto out;
5444 		}
5445 
5446 		scat = mlx5_frag_buf_get_wqe(&qp->rq.fbc, ind);
5447 		if (qp->wq_sig)
5448 			scat++;
5449 
5450 		for (i = 0; i < wr->num_sge; i++)
5451 			set_data_ptr_seg(scat + i, wr->sg_list + i);
5452 
5453 		if (i < qp->rq.max_gs) {
5454 			scat[i].byte_count = 0;
5455 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
5456 			scat[i].addr       = 0;
5457 		}
5458 
5459 		if (qp->wq_sig) {
5460 			sig = (struct mlx5_rwqe_sig *)scat;
5461 			set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
5462 		}
5463 
5464 		qp->rq.wrid[ind] = wr->wr_id;
5465 
5466 		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
5467 	}
5468 
5469 out:
5470 	if (likely(nreq)) {
5471 		qp->rq.head += nreq;
5472 
5473 		/* Make sure that descriptors are written before
5474 		 * doorbell record.
5475 		 */
5476 		wmb();
5477 
5478 		*qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
5479 	}
5480 
5481 	spin_unlock_irqrestore(&qp->rq.lock, flags);
5482 
5483 	return err;
5484 }
5485 
5486 int mlx5_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
5487 		      const struct ib_recv_wr **bad_wr)
5488 {
5489 	return _mlx5_ib_post_recv(ibqp, wr, bad_wr, false);
5490 }
5491 
5492 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
5493 {
5494 	switch (mlx5_state) {
5495 	case MLX5_QP_STATE_RST:      return IB_QPS_RESET;
5496 	case MLX5_QP_STATE_INIT:     return IB_QPS_INIT;
5497 	case MLX5_QP_STATE_RTR:      return IB_QPS_RTR;
5498 	case MLX5_QP_STATE_RTS:      return IB_QPS_RTS;
5499 	case MLX5_QP_STATE_SQ_DRAINING:
5500 	case MLX5_QP_STATE_SQD:      return IB_QPS_SQD;
5501 	case MLX5_QP_STATE_SQER:     return IB_QPS_SQE;
5502 	case MLX5_QP_STATE_ERR:      return IB_QPS_ERR;
5503 	default:		     return -1;
5504 	}
5505 }
5506 
5507 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
5508 {
5509 	switch (mlx5_mig_state) {
5510 	case MLX5_QP_PM_ARMED:		return IB_MIG_ARMED;
5511 	case MLX5_QP_PM_REARM:		return IB_MIG_REARM;
5512 	case MLX5_QP_PM_MIGRATED:	return IB_MIG_MIGRATED;
5513 	default: return -1;
5514 	}
5515 }
5516 
5517 static int to_ib_qp_access_flags(int mlx5_flags)
5518 {
5519 	int ib_flags = 0;
5520 
5521 	if (mlx5_flags & MLX5_QP_BIT_RRE)
5522 		ib_flags |= IB_ACCESS_REMOTE_READ;
5523 	if (mlx5_flags & MLX5_QP_BIT_RWE)
5524 		ib_flags |= IB_ACCESS_REMOTE_WRITE;
5525 	if (mlx5_flags & MLX5_QP_BIT_RAE)
5526 		ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
5527 
5528 	return ib_flags;
5529 }
5530 
5531 static void to_rdma_ah_attr(struct mlx5_ib_dev *ibdev,
5532 			    struct rdma_ah_attr *ah_attr,
5533 			    struct mlx5_qp_path *path)
5534 {
5535 
5536 	memset(ah_attr, 0, sizeof(*ah_attr));
5537 
5538 	if (!path->port || path->port > ibdev->num_ports)
5539 		return;
5540 
5541 	ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, path->port);
5542 
5543 	rdma_ah_set_port_num(ah_attr, path->port);
5544 	rdma_ah_set_sl(ah_attr, path->dci_cfi_prio_sl & 0xf);
5545 
5546 	rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
5547 	rdma_ah_set_path_bits(ah_attr, path->grh_mlid & 0x7f);
5548 	rdma_ah_set_static_rate(ah_attr,
5549 				path->static_rate ? path->static_rate - 5 : 0);
5550 	if (path->grh_mlid & (1 << 7)) {
5551 		u32 tc_fl = be32_to_cpu(path->tclass_flowlabel);
5552 
5553 		rdma_ah_set_grh(ah_attr, NULL,
5554 				tc_fl & 0xfffff,
5555 				path->mgid_index,
5556 				path->hop_limit,
5557 				(tc_fl >> 20) & 0xff);
5558 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
5559 	}
5560 }
5561 
5562 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
5563 					struct mlx5_ib_sq *sq,
5564 					u8 *sq_state)
5565 {
5566 	int err;
5567 
5568 	err = mlx5_core_query_sq_state(dev->mdev, sq->base.mqp.qpn, sq_state);
5569 	if (err)
5570 		goto out;
5571 	sq->state = *sq_state;
5572 
5573 out:
5574 	return err;
5575 }
5576 
5577 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
5578 					struct mlx5_ib_rq *rq,
5579 					u8 *rq_state)
5580 {
5581 	void *out;
5582 	void *rqc;
5583 	int inlen;
5584 	int err;
5585 
5586 	inlen = MLX5_ST_SZ_BYTES(query_rq_out);
5587 	out = kvzalloc(inlen, GFP_KERNEL);
5588 	if (!out)
5589 		return -ENOMEM;
5590 
5591 	err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
5592 	if (err)
5593 		goto out;
5594 
5595 	rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
5596 	*rq_state = MLX5_GET(rqc, rqc, state);
5597 	rq->state = *rq_state;
5598 
5599 out:
5600 	kvfree(out);
5601 	return err;
5602 }
5603 
5604 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
5605 				  struct mlx5_ib_qp *qp, u8 *qp_state)
5606 {
5607 	static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
5608 		[MLX5_RQC_STATE_RST] = {
5609 			[MLX5_SQC_STATE_RST]	= IB_QPS_RESET,
5610 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
5611 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE_BAD,
5612 			[MLX5_SQ_STATE_NA]	= IB_QPS_RESET,
5613 		},
5614 		[MLX5_RQC_STATE_RDY] = {
5615 			[MLX5_SQC_STATE_RST]	= MLX5_QP_STATE_BAD,
5616 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
5617 			[MLX5_SQC_STATE_ERR]	= IB_QPS_SQE,
5618 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE,
5619 		},
5620 		[MLX5_RQC_STATE_ERR] = {
5621 			[MLX5_SQC_STATE_RST]    = MLX5_QP_STATE_BAD,
5622 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE_BAD,
5623 			[MLX5_SQC_STATE_ERR]	= IB_QPS_ERR,
5624 			[MLX5_SQ_STATE_NA]	= IB_QPS_ERR,
5625 		},
5626 		[MLX5_RQ_STATE_NA] = {
5627 			[MLX5_SQC_STATE_RST]    = IB_QPS_RESET,
5628 			[MLX5_SQC_STATE_RDY]	= MLX5_QP_STATE,
5629 			[MLX5_SQC_STATE_ERR]	= MLX5_QP_STATE,
5630 			[MLX5_SQ_STATE_NA]	= MLX5_QP_STATE_BAD,
5631 		},
5632 	};
5633 
5634 	*qp_state = sqrq_trans[rq_state][sq_state];
5635 
5636 	if (*qp_state == MLX5_QP_STATE_BAD) {
5637 		WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
5638 		     qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
5639 		     qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
5640 		return -EINVAL;
5641 	}
5642 
5643 	if (*qp_state == MLX5_QP_STATE)
5644 		*qp_state = qp->state;
5645 
5646 	return 0;
5647 }
5648 
5649 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
5650 				     struct mlx5_ib_qp *qp,
5651 				     u8 *raw_packet_qp_state)
5652 {
5653 	struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
5654 	struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
5655 	struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
5656 	int err;
5657 	u8 sq_state = MLX5_SQ_STATE_NA;
5658 	u8 rq_state = MLX5_RQ_STATE_NA;
5659 
5660 	if (qp->sq.wqe_cnt) {
5661 		err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
5662 		if (err)
5663 			return err;
5664 	}
5665 
5666 	if (qp->rq.wqe_cnt) {
5667 		err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
5668 		if (err)
5669 			return err;
5670 	}
5671 
5672 	return sqrq_state_to_qp_state(sq_state, rq_state, qp,
5673 				      raw_packet_qp_state);
5674 }
5675 
5676 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
5677 			 struct ib_qp_attr *qp_attr)
5678 {
5679 	int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
5680 	struct mlx5_qp_context *context;
5681 	int mlx5_state;
5682 	u32 *outb;
5683 	int err = 0;
5684 
5685 	outb = kzalloc(outlen, GFP_KERNEL);
5686 	if (!outb)
5687 		return -ENOMEM;
5688 
5689 	err = mlx5_core_qp_query(dev->mdev, &qp->trans_qp.base.mqp, outb,
5690 				 outlen);
5691 	if (err)
5692 		goto out;
5693 
5694 	/* FIXME: use MLX5_GET rather than mlx5_qp_context manual struct */
5695 	context = (struct mlx5_qp_context *)MLX5_ADDR_OF(query_qp_out, outb, qpc);
5696 
5697 	mlx5_state = be32_to_cpu(context->flags) >> 28;
5698 
5699 	qp->state		     = to_ib_qp_state(mlx5_state);
5700 	qp_attr->path_mtu	     = context->mtu_msgmax >> 5;
5701 	qp_attr->path_mig_state	     =
5702 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
5703 	qp_attr->qkey		     = be32_to_cpu(context->qkey);
5704 	qp_attr->rq_psn		     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
5705 	qp_attr->sq_psn		     = be32_to_cpu(context->next_send_psn) & 0xffffff;
5706 	qp_attr->dest_qp_num	     = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
5707 	qp_attr->qp_access_flags     =
5708 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
5709 
5710 	if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
5711 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
5712 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
5713 		qp_attr->alt_pkey_index =
5714 			be16_to_cpu(context->alt_path.pkey_index);
5715 		qp_attr->alt_port_num	=
5716 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
5717 	}
5718 
5719 	qp_attr->pkey_index = be16_to_cpu(context->pri_path.pkey_index);
5720 	qp_attr->port_num = context->pri_path.port;
5721 
5722 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
5723 	qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
5724 
5725 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
5726 
5727 	qp_attr->max_dest_rd_atomic =
5728 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
5729 	qp_attr->min_rnr_timer	    =
5730 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
5731 	qp_attr->timeout	    = context->pri_path.ackto_lt >> 3;
5732 	qp_attr->retry_cnt	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
5733 	qp_attr->rnr_retry	    = (be32_to_cpu(context->params1) >> 13) & 0x7;
5734 	qp_attr->alt_timeout	    = context->alt_path.ackto_lt >> 3;
5735 
5736 out:
5737 	kfree(outb);
5738 	return err;
5739 }
5740 
5741 static int mlx5_ib_dct_query_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *mqp,
5742 				struct ib_qp_attr *qp_attr, int qp_attr_mask,
5743 				struct ib_qp_init_attr *qp_init_attr)
5744 {
5745 	struct mlx5_core_dct	*dct = &mqp->dct.mdct;
5746 	u32 *out;
5747 	u32 access_flags = 0;
5748 	int outlen = MLX5_ST_SZ_BYTES(query_dct_out);
5749 	void *dctc;
5750 	int err;
5751 	int supported_mask = IB_QP_STATE |
5752 			     IB_QP_ACCESS_FLAGS |
5753 			     IB_QP_PORT |
5754 			     IB_QP_MIN_RNR_TIMER |
5755 			     IB_QP_AV |
5756 			     IB_QP_PATH_MTU |
5757 			     IB_QP_PKEY_INDEX;
5758 
5759 	if (qp_attr_mask & ~supported_mask)
5760 		return -EINVAL;
5761 	if (mqp->state != IB_QPS_RTR)
5762 		return -EINVAL;
5763 
5764 	out = kzalloc(outlen, GFP_KERNEL);
5765 	if (!out)
5766 		return -ENOMEM;
5767 
5768 	err = mlx5_core_dct_query(dev->mdev, dct, out, outlen);
5769 	if (err)
5770 		goto out;
5771 
5772 	dctc = MLX5_ADDR_OF(query_dct_out, out, dct_context_entry);
5773 
5774 	if (qp_attr_mask & IB_QP_STATE)
5775 		qp_attr->qp_state = IB_QPS_RTR;
5776 
5777 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
5778 		if (MLX5_GET(dctc, dctc, rre))
5779 			access_flags |= IB_ACCESS_REMOTE_READ;
5780 		if (MLX5_GET(dctc, dctc, rwe))
5781 			access_flags |= IB_ACCESS_REMOTE_WRITE;
5782 		if (MLX5_GET(dctc, dctc, rae))
5783 			access_flags |= IB_ACCESS_REMOTE_ATOMIC;
5784 		qp_attr->qp_access_flags = access_flags;
5785 	}
5786 
5787 	if (qp_attr_mask & IB_QP_PORT)
5788 		qp_attr->port_num = MLX5_GET(dctc, dctc, port);
5789 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER)
5790 		qp_attr->min_rnr_timer = MLX5_GET(dctc, dctc, min_rnr_nak);
5791 	if (qp_attr_mask & IB_QP_AV) {
5792 		qp_attr->ah_attr.grh.traffic_class = MLX5_GET(dctc, dctc, tclass);
5793 		qp_attr->ah_attr.grh.flow_label = MLX5_GET(dctc, dctc, flow_label);
5794 		qp_attr->ah_attr.grh.sgid_index = MLX5_GET(dctc, dctc, my_addr_index);
5795 		qp_attr->ah_attr.grh.hop_limit = MLX5_GET(dctc, dctc, hop_limit);
5796 	}
5797 	if (qp_attr_mask & IB_QP_PATH_MTU)
5798 		qp_attr->path_mtu = MLX5_GET(dctc, dctc, mtu);
5799 	if (qp_attr_mask & IB_QP_PKEY_INDEX)
5800 		qp_attr->pkey_index = MLX5_GET(dctc, dctc, pkey_index);
5801 out:
5802 	kfree(out);
5803 	return err;
5804 }
5805 
5806 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
5807 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
5808 {
5809 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
5810 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
5811 	int err = 0;
5812 	u8 raw_packet_qp_state;
5813 
5814 	if (ibqp->rwq_ind_tbl)
5815 		return -ENOSYS;
5816 
5817 	if (unlikely(ibqp->qp_type == IB_QPT_GSI))
5818 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
5819 					    qp_init_attr);
5820 
5821 	/* Not all of output fields are applicable, make sure to zero them */
5822 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
5823 	memset(qp_attr, 0, sizeof(*qp_attr));
5824 
5825 	if (unlikely(qp->qp_sub_type == MLX5_IB_QPT_DCT))
5826 		return mlx5_ib_dct_query_qp(dev, qp, qp_attr,
5827 					    qp_attr_mask, qp_init_attr);
5828 
5829 	mutex_lock(&qp->mutex);
5830 
5831 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET ||
5832 	    qp->flags & MLX5_IB_QP_UNDERLAY) {
5833 		err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
5834 		if (err)
5835 			goto out;
5836 		qp->state = raw_packet_qp_state;
5837 		qp_attr->port_num = 1;
5838 	} else {
5839 		err = query_qp_attr(dev, qp, qp_attr);
5840 		if (err)
5841 			goto out;
5842 	}
5843 
5844 	qp_attr->qp_state	     = qp->state;
5845 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
5846 	qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
5847 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
5848 
5849 	if (!ibqp->uobject) {
5850 		qp_attr->cap.max_send_wr  = qp->sq.max_post;
5851 		qp_attr->cap.max_send_sge = qp->sq.max_gs;
5852 		qp_init_attr->qp_context = ibqp->qp_context;
5853 	} else {
5854 		qp_attr->cap.max_send_wr  = 0;
5855 		qp_attr->cap.max_send_sge = 0;
5856 	}
5857 
5858 	qp_init_attr->qp_type = ibqp->qp_type;
5859 	qp_init_attr->recv_cq = ibqp->recv_cq;
5860 	qp_init_attr->send_cq = ibqp->send_cq;
5861 	qp_init_attr->srq = ibqp->srq;
5862 	qp_attr->cap.max_inline_data = qp->max_inline_data;
5863 
5864 	qp_init_attr->cap	     = qp_attr->cap;
5865 
5866 	qp_init_attr->create_flags = 0;
5867 	if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
5868 		qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
5869 
5870 	if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
5871 		qp_init_attr->create_flags |= IB_QP_CREATE_CROSS_CHANNEL;
5872 	if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
5873 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
5874 	if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
5875 		qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
5876 	if (qp->flags & MLX5_IB_QP_SQPN_QP1)
5877 		qp_init_attr->create_flags |= MLX5_IB_QP_CREATE_SQPN_QP1;
5878 
5879 	qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
5880 		IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
5881 
5882 out:
5883 	mutex_unlock(&qp->mutex);
5884 	return err;
5885 }
5886 
5887 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
5888 				   struct ib_udata *udata)
5889 {
5890 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
5891 	struct mlx5_ib_xrcd *xrcd;
5892 	int err;
5893 
5894 	if (!MLX5_CAP_GEN(dev->mdev, xrc))
5895 		return ERR_PTR(-ENOSYS);
5896 
5897 	xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
5898 	if (!xrcd)
5899 		return ERR_PTR(-ENOMEM);
5900 
5901 	err = mlx5_cmd_xrcd_alloc(dev->mdev, &xrcd->xrcdn, 0);
5902 	if (err) {
5903 		kfree(xrcd);
5904 		return ERR_PTR(-ENOMEM);
5905 	}
5906 
5907 	return &xrcd->ibxrcd;
5908 }
5909 
5910 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
5911 {
5912 	struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
5913 	u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
5914 	int err;
5915 
5916 	err = mlx5_cmd_xrcd_dealloc(dev->mdev, xrcdn, 0);
5917 	if (err)
5918 		mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
5919 
5920 	kfree(xrcd);
5921 	return 0;
5922 }
5923 
5924 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
5925 {
5926 	struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
5927 	struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
5928 	struct ib_event event;
5929 
5930 	if (rwq->ibwq.event_handler) {
5931 		event.device     = rwq->ibwq.device;
5932 		event.element.wq = &rwq->ibwq;
5933 		switch (type) {
5934 		case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
5935 			event.event = IB_EVENT_WQ_FATAL;
5936 			break;
5937 		default:
5938 			mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
5939 			return;
5940 		}
5941 
5942 		rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
5943 	}
5944 }
5945 
5946 static int set_delay_drop(struct mlx5_ib_dev *dev)
5947 {
5948 	int err = 0;
5949 
5950 	mutex_lock(&dev->delay_drop.lock);
5951 	if (dev->delay_drop.activate)
5952 		goto out;
5953 
5954 	err = mlx5_core_set_delay_drop(dev->mdev, dev->delay_drop.timeout);
5955 	if (err)
5956 		goto out;
5957 
5958 	dev->delay_drop.activate = true;
5959 out:
5960 	mutex_unlock(&dev->delay_drop.lock);
5961 
5962 	if (!err)
5963 		atomic_inc(&dev->delay_drop.rqs_cnt);
5964 	return err;
5965 }
5966 
5967 static int  create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
5968 		      struct ib_wq_init_attr *init_attr)
5969 {
5970 	struct mlx5_ib_dev *dev;
5971 	int has_net_offloads;
5972 	__be64 *rq_pas0;
5973 	void *in;
5974 	void *rqc;
5975 	void *wq;
5976 	int inlen;
5977 	int err;
5978 
5979 	dev = to_mdev(pd->device);
5980 
5981 	inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
5982 	in = kvzalloc(inlen, GFP_KERNEL);
5983 	if (!in)
5984 		return -ENOMEM;
5985 
5986 	MLX5_SET(create_rq_in, in, uid, to_mpd(pd)->uid);
5987 	rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
5988 	MLX5_SET(rqc,  rqc, mem_rq_type,
5989 		 MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
5990 	MLX5_SET(rqc, rqc, user_index, rwq->user_index);
5991 	MLX5_SET(rqc,  rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
5992 	MLX5_SET(rqc,  rqc, state, MLX5_RQC_STATE_RST);
5993 	MLX5_SET(rqc,  rqc, flush_in_error_en, 1);
5994 	wq = MLX5_ADDR_OF(rqc, rqc, wq);
5995 	MLX5_SET(wq, wq, wq_type,
5996 		 rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ ?
5997 		 MLX5_WQ_TYPE_CYCLIC_STRIDING_RQ : MLX5_WQ_TYPE_CYCLIC);
5998 	if (init_attr->create_flags & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
5999 		if (!MLX5_CAP_GEN(dev->mdev, end_pad)) {
6000 			mlx5_ib_dbg(dev, "Scatter end padding is not supported\n");
6001 			err = -EOPNOTSUPP;
6002 			goto out;
6003 		} else {
6004 			MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
6005 		}
6006 	}
6007 	MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
6008 	if (rwq->create_flags & MLX5_IB_WQ_FLAGS_STRIDING_RQ) {
6009 		/*
6010 		 * In Firmware number of strides in each WQE is:
6011 		 *   "512 * 2^single_wqe_log_num_of_strides"
6012 		 * Values 3 to 8 are accepted as 10 to 15, 9 to 18 are
6013 		 * accepted as 0 to 9
6014 		 */
6015 		static const u8 fw_map[] = { 10, 11, 12, 13, 14, 15, 0, 1,
6016 					     2,  3,  4,  5,  6,  7,  8, 9 };
6017 		MLX5_SET(wq, wq, two_byte_shift_en, rwq->two_byte_shift_en);
6018 		MLX5_SET(wq, wq, log_wqe_stride_size,
6019 			 rwq->single_stride_log_num_of_bytes -
6020 			 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES);
6021 		MLX5_SET(wq, wq, log_wqe_num_of_strides,
6022 			 fw_map[rwq->log_num_strides -
6023 				MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES]);
6024 	}
6025 	MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
6026 	MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
6027 	MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
6028 	MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
6029 	MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
6030 	MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
6031 	has_net_offloads = MLX5_CAP_GEN(dev->mdev, eth_net_offloads);
6032 	if (init_attr->create_flags & IB_WQ_FLAGS_CVLAN_STRIPPING) {
6033 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
6034 			mlx5_ib_dbg(dev, "VLAN offloads are not supported\n");
6035 			err = -EOPNOTSUPP;
6036 			goto out;
6037 		}
6038 	} else {
6039 		MLX5_SET(rqc, rqc, vsd, 1);
6040 	}
6041 	if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS) {
6042 		if (!(has_net_offloads && MLX5_CAP_ETH(dev->mdev, scatter_fcs))) {
6043 			mlx5_ib_dbg(dev, "Scatter FCS is not supported\n");
6044 			err = -EOPNOTSUPP;
6045 			goto out;
6046 		}
6047 		MLX5_SET(rqc, rqc, scatter_fcs, 1);
6048 	}
6049 	if (init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
6050 		if (!(dev->ib_dev.attrs.raw_packet_caps &
6051 		      IB_RAW_PACKET_CAP_DELAY_DROP)) {
6052 			mlx5_ib_dbg(dev, "Delay drop is not supported\n");
6053 			err = -EOPNOTSUPP;
6054 			goto out;
6055 		}
6056 		MLX5_SET(rqc, rqc, delay_drop_en, 1);
6057 	}
6058 	rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
6059 	mlx5_ib_populate_pas(dev, rwq->umem, rwq->page_shift, rq_pas0, 0);
6060 	err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rwq->core_qp);
6061 	if (!err && init_attr->create_flags & IB_WQ_FLAGS_DELAY_DROP) {
6062 		err = set_delay_drop(dev);
6063 		if (err) {
6064 			mlx5_ib_warn(dev, "Failed to enable delay drop err=%d\n",
6065 				     err);
6066 			mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
6067 		} else {
6068 			rwq->create_flags |= MLX5_IB_WQ_FLAGS_DELAY_DROP;
6069 		}
6070 	}
6071 out:
6072 	kvfree(in);
6073 	return err;
6074 }
6075 
6076 static int set_user_rq_size(struct mlx5_ib_dev *dev,
6077 			    struct ib_wq_init_attr *wq_init_attr,
6078 			    struct mlx5_ib_create_wq *ucmd,
6079 			    struct mlx5_ib_rwq *rwq)
6080 {
6081 	/* Sanity check RQ size before proceeding */
6082 	if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
6083 		return -EINVAL;
6084 
6085 	if (!ucmd->rq_wqe_count)
6086 		return -EINVAL;
6087 
6088 	rwq->wqe_count = ucmd->rq_wqe_count;
6089 	rwq->wqe_shift = ucmd->rq_wqe_shift;
6090 	if (check_shl_overflow(rwq->wqe_count, rwq->wqe_shift, &rwq->buf_size))
6091 		return -EINVAL;
6092 
6093 	rwq->log_rq_stride = rwq->wqe_shift;
6094 	rwq->log_rq_size = ilog2(rwq->wqe_count);
6095 	return 0;
6096 }
6097 
6098 static bool log_of_strides_valid(struct mlx5_ib_dev *dev, u32 log_num_strides)
6099 {
6100 	if ((log_num_strides > MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES) ||
6101 	    (log_num_strides < MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
6102 		return false;
6103 
6104 	if (!MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) &&
6105 	    (log_num_strides < MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES))
6106 		return false;
6107 
6108 	return true;
6109 }
6110 
6111 static int prepare_user_rq(struct ib_pd *pd,
6112 			   struct ib_wq_init_attr *init_attr,
6113 			   struct ib_udata *udata,
6114 			   struct mlx5_ib_rwq *rwq)
6115 {
6116 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
6117 	struct mlx5_ib_create_wq ucmd = {};
6118 	int err;
6119 	size_t required_cmd_sz;
6120 
6121 	required_cmd_sz = offsetof(typeof(ucmd), single_stride_log_num_of_bytes)
6122 		+ sizeof(ucmd.single_stride_log_num_of_bytes);
6123 	if (udata->inlen < required_cmd_sz) {
6124 		mlx5_ib_dbg(dev, "invalid inlen\n");
6125 		return -EINVAL;
6126 	}
6127 
6128 	if (udata->inlen > sizeof(ucmd) &&
6129 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
6130 				 udata->inlen - sizeof(ucmd))) {
6131 		mlx5_ib_dbg(dev, "inlen is not supported\n");
6132 		return -EOPNOTSUPP;
6133 	}
6134 
6135 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
6136 		mlx5_ib_dbg(dev, "copy failed\n");
6137 		return -EFAULT;
6138 	}
6139 
6140 	if (ucmd.comp_mask & (~MLX5_IB_CREATE_WQ_STRIDING_RQ)) {
6141 		mlx5_ib_dbg(dev, "invalid comp mask\n");
6142 		return -EOPNOTSUPP;
6143 	} else if (ucmd.comp_mask & MLX5_IB_CREATE_WQ_STRIDING_RQ) {
6144 		if (!MLX5_CAP_GEN(dev->mdev, striding_rq)) {
6145 			mlx5_ib_dbg(dev, "Striding RQ is not supported\n");
6146 			return -EOPNOTSUPP;
6147 		}
6148 		if ((ucmd.single_stride_log_num_of_bytes <
6149 		    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES) ||
6150 		    (ucmd.single_stride_log_num_of_bytes >
6151 		     MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES)) {
6152 			mlx5_ib_dbg(dev, "Invalid log stride size (%u. Range is %u - %u)\n",
6153 				    ucmd.single_stride_log_num_of_bytes,
6154 				    MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES,
6155 				    MLX5_MAX_SINGLE_STRIDE_LOG_NUM_BYTES);
6156 			return -EINVAL;
6157 		}
6158 		if (!log_of_strides_valid(dev,
6159 					  ucmd.single_wqe_log_num_of_strides)) {
6160 			mlx5_ib_dbg(
6161 				dev,
6162 				"Invalid log num strides (%u. Range is %u - %u)\n",
6163 				ucmd.single_wqe_log_num_of_strides,
6164 				MLX5_CAP_GEN(dev->mdev, ext_stride_num_range) ?
6165 					MLX5_EXT_MIN_SINGLE_WQE_LOG_NUM_STRIDES :
6166 					MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES,
6167 				MLX5_MAX_SINGLE_WQE_LOG_NUM_STRIDES);
6168 			return -EINVAL;
6169 		}
6170 		rwq->single_stride_log_num_of_bytes =
6171 			ucmd.single_stride_log_num_of_bytes;
6172 		rwq->log_num_strides = ucmd.single_wqe_log_num_of_strides;
6173 		rwq->two_byte_shift_en = !!ucmd.two_byte_shift_en;
6174 		rwq->create_flags |= MLX5_IB_WQ_FLAGS_STRIDING_RQ;
6175 	}
6176 
6177 	err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
6178 	if (err) {
6179 		mlx5_ib_dbg(dev, "err %d\n", err);
6180 		return err;
6181 	}
6182 
6183 	err = create_user_rq(dev, pd, udata, rwq, &ucmd);
6184 	if (err) {
6185 		mlx5_ib_dbg(dev, "err %d\n", err);
6186 		return err;
6187 	}
6188 
6189 	rwq->user_index = ucmd.user_index;
6190 	return 0;
6191 }
6192 
6193 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
6194 				struct ib_wq_init_attr *init_attr,
6195 				struct ib_udata *udata)
6196 {
6197 	struct mlx5_ib_dev *dev;
6198 	struct mlx5_ib_rwq *rwq;
6199 	struct mlx5_ib_create_wq_resp resp = {};
6200 	size_t min_resp_len;
6201 	int err;
6202 
6203 	if (!udata)
6204 		return ERR_PTR(-ENOSYS);
6205 
6206 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
6207 	if (udata->outlen && udata->outlen < min_resp_len)
6208 		return ERR_PTR(-EINVAL);
6209 
6210 	dev = to_mdev(pd->device);
6211 	switch (init_attr->wq_type) {
6212 	case IB_WQT_RQ:
6213 		rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
6214 		if (!rwq)
6215 			return ERR_PTR(-ENOMEM);
6216 		err = prepare_user_rq(pd, init_attr, udata, rwq);
6217 		if (err)
6218 			goto err;
6219 		err = create_rq(rwq, pd, init_attr);
6220 		if (err)
6221 			goto err_user_rq;
6222 		break;
6223 	default:
6224 		mlx5_ib_dbg(dev, "unsupported wq type %d\n",
6225 			    init_attr->wq_type);
6226 		return ERR_PTR(-EINVAL);
6227 	}
6228 
6229 	rwq->ibwq.wq_num = rwq->core_qp.qpn;
6230 	rwq->ibwq.state = IB_WQS_RESET;
6231 	if (udata->outlen) {
6232 		resp.response_length = offsetof(typeof(resp), response_length) +
6233 				sizeof(resp.response_length);
6234 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
6235 		if (err)
6236 			goto err_copy;
6237 	}
6238 
6239 	rwq->core_qp.event = mlx5_ib_wq_event;
6240 	rwq->ibwq.event_handler = init_attr->event_handler;
6241 	return &rwq->ibwq;
6242 
6243 err_copy:
6244 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
6245 err_user_rq:
6246 	destroy_user_rq(dev, pd, rwq, udata);
6247 err:
6248 	kfree(rwq);
6249 	return ERR_PTR(err);
6250 }
6251 
6252 void mlx5_ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
6253 {
6254 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
6255 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
6256 
6257 	mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
6258 	destroy_user_rq(dev, wq->pd, rwq, udata);
6259 	kfree(rwq);
6260 }
6261 
6262 struct ib_rwq_ind_table *mlx5_ib_create_rwq_ind_table(struct ib_device *device,
6263 						      struct ib_rwq_ind_table_init_attr *init_attr,
6264 						      struct ib_udata *udata)
6265 {
6266 	struct mlx5_ib_dev *dev = to_mdev(device);
6267 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl;
6268 	int sz = 1 << init_attr->log_ind_tbl_size;
6269 	struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
6270 	size_t min_resp_len;
6271 	int inlen;
6272 	int err;
6273 	int i;
6274 	u32 *in;
6275 	void *rqtc;
6276 
6277 	if (udata->inlen > 0 &&
6278 	    !ib_is_udata_cleared(udata, 0,
6279 				 udata->inlen))
6280 		return ERR_PTR(-EOPNOTSUPP);
6281 
6282 	if (init_attr->log_ind_tbl_size >
6283 	    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
6284 		mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
6285 			    init_attr->log_ind_tbl_size,
6286 			    MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
6287 		return ERR_PTR(-EINVAL);
6288 	}
6289 
6290 	min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
6291 	if (udata->outlen && udata->outlen < min_resp_len)
6292 		return ERR_PTR(-EINVAL);
6293 
6294 	rwq_ind_tbl = kzalloc(sizeof(*rwq_ind_tbl), GFP_KERNEL);
6295 	if (!rwq_ind_tbl)
6296 		return ERR_PTR(-ENOMEM);
6297 
6298 	inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
6299 	in = kvzalloc(inlen, GFP_KERNEL);
6300 	if (!in) {
6301 		err = -ENOMEM;
6302 		goto err;
6303 	}
6304 
6305 	rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
6306 
6307 	MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
6308 	MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
6309 
6310 	for (i = 0; i < sz; i++)
6311 		MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
6312 
6313 	rwq_ind_tbl->uid = to_mpd(init_attr->ind_tbl[0]->pd)->uid;
6314 	MLX5_SET(create_rqt_in, in, uid, rwq_ind_tbl->uid);
6315 
6316 	err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
6317 	kvfree(in);
6318 
6319 	if (err)
6320 		goto err;
6321 
6322 	rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
6323 	if (udata->outlen) {
6324 		resp.response_length = offsetof(typeof(resp), response_length) +
6325 					sizeof(resp.response_length);
6326 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
6327 		if (err)
6328 			goto err_copy;
6329 	}
6330 
6331 	return &rwq_ind_tbl->ib_rwq_ind_tbl;
6332 
6333 err_copy:
6334 	mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
6335 err:
6336 	kfree(rwq_ind_tbl);
6337 	return ERR_PTR(err);
6338 }
6339 
6340 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
6341 {
6342 	struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
6343 	struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
6344 
6345 	mlx5_cmd_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn, rwq_ind_tbl->uid);
6346 
6347 	kfree(rwq_ind_tbl);
6348 	return 0;
6349 }
6350 
6351 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
6352 		      u32 wq_attr_mask, struct ib_udata *udata)
6353 {
6354 	struct mlx5_ib_dev *dev = to_mdev(wq->device);
6355 	struct mlx5_ib_rwq *rwq = to_mrwq(wq);
6356 	struct mlx5_ib_modify_wq ucmd = {};
6357 	size_t required_cmd_sz;
6358 	int curr_wq_state;
6359 	int wq_state;
6360 	int inlen;
6361 	int err;
6362 	void *rqc;
6363 	void *in;
6364 
6365 	required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
6366 	if (udata->inlen < required_cmd_sz)
6367 		return -EINVAL;
6368 
6369 	if (udata->inlen > sizeof(ucmd) &&
6370 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
6371 				 udata->inlen - sizeof(ucmd)))
6372 		return -EOPNOTSUPP;
6373 
6374 	if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
6375 		return -EFAULT;
6376 
6377 	if (ucmd.comp_mask || ucmd.reserved)
6378 		return -EOPNOTSUPP;
6379 
6380 	inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
6381 	in = kvzalloc(inlen, GFP_KERNEL);
6382 	if (!in)
6383 		return -ENOMEM;
6384 
6385 	rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
6386 
6387 	curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
6388 		wq_attr->curr_wq_state : wq->state;
6389 	wq_state = (wq_attr_mask & IB_WQ_STATE) ?
6390 		wq_attr->wq_state : curr_wq_state;
6391 	if (curr_wq_state == IB_WQS_ERR)
6392 		curr_wq_state = MLX5_RQC_STATE_ERR;
6393 	if (wq_state == IB_WQS_ERR)
6394 		wq_state = MLX5_RQC_STATE_ERR;
6395 	MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
6396 	MLX5_SET(modify_rq_in, in, uid, to_mpd(wq->pd)->uid);
6397 	MLX5_SET(rqc, rqc, state, wq_state);
6398 
6399 	if (wq_attr_mask & IB_WQ_FLAGS) {
6400 		if (wq_attr->flags_mask & IB_WQ_FLAGS_CVLAN_STRIPPING) {
6401 			if (!(MLX5_CAP_GEN(dev->mdev, eth_net_offloads) &&
6402 			      MLX5_CAP_ETH(dev->mdev, vlan_cap))) {
6403 				mlx5_ib_dbg(dev, "VLAN offloads are not "
6404 					    "supported\n");
6405 				err = -EOPNOTSUPP;
6406 				goto out;
6407 			}
6408 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
6409 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
6410 			MLX5_SET(rqc, rqc, vsd,
6411 				 (wq_attr->flags & IB_WQ_FLAGS_CVLAN_STRIPPING) ? 0 : 1);
6412 		}
6413 
6414 		if (wq_attr->flags_mask & IB_WQ_FLAGS_PCI_WRITE_END_PADDING) {
6415 			mlx5_ib_dbg(dev, "Modifying scatter end padding is not supported\n");
6416 			err = -EOPNOTSUPP;
6417 			goto out;
6418 		}
6419 	}
6420 
6421 	if (curr_wq_state == IB_WQS_RESET && wq_state == IB_WQS_RDY) {
6422 		u16 set_id;
6423 
6424 		set_id = mlx5_ib_get_counters_id(dev, 0);
6425 		if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
6426 			MLX5_SET64(modify_rq_in, in, modify_bitmask,
6427 				   MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_RQ_COUNTER_SET_ID);
6428 			MLX5_SET(rqc, rqc, counter_set_id, set_id);
6429 		} else
6430 			dev_info_once(
6431 				&dev->ib_dev.dev,
6432 				"Receive WQ counters are not supported on current FW\n");
6433 	}
6434 
6435 	err = mlx5_core_modify_rq(dev->mdev, rwq->core_qp.qpn, in, inlen);
6436 	if (!err)
6437 		rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
6438 
6439 out:
6440 	kvfree(in);
6441 	return err;
6442 }
6443 
6444 struct mlx5_ib_drain_cqe {
6445 	struct ib_cqe cqe;
6446 	struct completion done;
6447 };
6448 
6449 static void mlx5_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
6450 {
6451 	struct mlx5_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
6452 						     struct mlx5_ib_drain_cqe,
6453 						     cqe);
6454 
6455 	complete(&cqe->done);
6456 }
6457 
6458 /* This function returns only once the drained WR was completed */
6459 static void handle_drain_completion(struct ib_cq *cq,
6460 				    struct mlx5_ib_drain_cqe *sdrain,
6461 				    struct mlx5_ib_dev *dev)
6462 {
6463 	struct mlx5_core_dev *mdev = dev->mdev;
6464 
6465 	if (cq->poll_ctx == IB_POLL_DIRECT) {
6466 		while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
6467 			ib_process_cq_direct(cq, -1);
6468 		return;
6469 	}
6470 
6471 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
6472 		struct mlx5_ib_cq *mcq = to_mcq(cq);
6473 		bool triggered = false;
6474 		unsigned long flags;
6475 
6476 		spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
6477 		/* Make sure that the CQ handler won't run if wasn't run yet */
6478 		if (!mcq->mcq.reset_notify_added)
6479 			mcq->mcq.reset_notify_added = 1;
6480 		else
6481 			triggered = true;
6482 		spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
6483 
6484 		if (triggered) {
6485 			/* Wait for any scheduled/running task to be ended */
6486 			switch (cq->poll_ctx) {
6487 			case IB_POLL_SOFTIRQ:
6488 				irq_poll_disable(&cq->iop);
6489 				irq_poll_enable(&cq->iop);
6490 				break;
6491 			case IB_POLL_WORKQUEUE:
6492 				cancel_work_sync(&cq->work);
6493 				break;
6494 			default:
6495 				WARN_ON_ONCE(1);
6496 			}
6497 		}
6498 
6499 		/* Run the CQ handler - this makes sure that the drain WR will
6500 		 * be processed if wasn't processed yet.
6501 		 */
6502 		mcq->mcq.comp(&mcq->mcq, NULL);
6503 	}
6504 
6505 	wait_for_completion(&sdrain->done);
6506 }
6507 
6508 void mlx5_ib_drain_sq(struct ib_qp *qp)
6509 {
6510 	struct ib_cq *cq = qp->send_cq;
6511 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
6512 	struct mlx5_ib_drain_cqe sdrain;
6513 	const struct ib_send_wr *bad_swr;
6514 	struct ib_rdma_wr swr = {
6515 		.wr = {
6516 			.next = NULL,
6517 			{ .wr_cqe	= &sdrain.cqe, },
6518 			.opcode	= IB_WR_RDMA_WRITE,
6519 		},
6520 	};
6521 	int ret;
6522 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
6523 	struct mlx5_core_dev *mdev = dev->mdev;
6524 
6525 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
6526 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
6527 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
6528 		return;
6529 	}
6530 
6531 	sdrain.cqe.done = mlx5_ib_drain_qp_done;
6532 	init_completion(&sdrain.done);
6533 
6534 	ret = _mlx5_ib_post_send(qp, &swr.wr, &bad_swr, true);
6535 	if (ret) {
6536 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
6537 		return;
6538 	}
6539 
6540 	handle_drain_completion(cq, &sdrain, dev);
6541 }
6542 
6543 void mlx5_ib_drain_rq(struct ib_qp *qp)
6544 {
6545 	struct ib_cq *cq = qp->recv_cq;
6546 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
6547 	struct mlx5_ib_drain_cqe rdrain;
6548 	struct ib_recv_wr rwr = {};
6549 	const struct ib_recv_wr *bad_rwr;
6550 	int ret;
6551 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
6552 	struct mlx5_core_dev *mdev = dev->mdev;
6553 
6554 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
6555 	if (ret && mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
6556 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
6557 		return;
6558 	}
6559 
6560 	rwr.wr_cqe = &rdrain.cqe;
6561 	rdrain.cqe.done = mlx5_ib_drain_qp_done;
6562 	init_completion(&rdrain.done);
6563 
6564 	ret = _mlx5_ib_post_recv(qp, &rwr, &bad_rwr, true);
6565 	if (ret) {
6566 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
6567 		return;
6568 	}
6569 
6570 	handle_drain_completion(cq, &rdrain, dev);
6571 }
6572 
6573 /**
6574  * Bind a qp to a counter. If @counter is NULL then bind the qp to
6575  * the default counter
6576  */
6577 int mlx5_ib_qp_set_counter(struct ib_qp *qp, struct rdma_counter *counter)
6578 {
6579 	struct mlx5_ib_qp *mqp = to_mqp(qp);
6580 	int err = 0;
6581 
6582 	mutex_lock(&mqp->mutex);
6583 	if (mqp->state == IB_QPS_RESET) {
6584 		qp->counter = counter;
6585 		goto out;
6586 	}
6587 
6588 	if (mqp->state == IB_QPS_RTS) {
6589 		err = __mlx5_ib_qp_set_counter(qp, counter);
6590 		if (!err)
6591 			qp->counter = counter;
6592 
6593 		goto out;
6594 	}
6595 
6596 	mqp->counter_pending = 1;
6597 	qp->counter = counter;
6598 
6599 out:
6600 	mutex_unlock(&mqp->mutex);
6601 	return err;
6602 }
6603